1 //===-- X86ISelLowering.cpp - X86 DAG Lowering Implementation -------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file defines the interfaces that X86 uses to lower LLVM code into a
13 //===----------------------------------------------------------------------===//
15 #include "X86ISelLowering.h"
16 #include "Utils/X86ShuffleDecode.h"
17 #include "X86CallingConv.h"
18 #include "X86FrameLowering.h"
19 #include "X86InstrBuilder.h"
20 #include "X86MachineFunctionInfo.h"
21 #include "X86TargetMachine.h"
22 #include "X86TargetObjectFile.h"
23 #include "llvm/ADT/SmallBitVector.h"
24 #include "llvm/ADT/SmallSet.h"
25 #include "llvm/ADT/Statistic.h"
26 #include "llvm/ADT/StringExtras.h"
27 #include "llvm/ADT/StringSwitch.h"
28 #include "llvm/CodeGen/IntrinsicLowering.h"
29 #include "llvm/CodeGen/MachineFrameInfo.h"
30 #include "llvm/CodeGen/MachineFunction.h"
31 #include "llvm/CodeGen/MachineInstrBuilder.h"
32 #include "llvm/CodeGen/MachineJumpTableInfo.h"
33 #include "llvm/CodeGen/MachineModuleInfo.h"
34 #include "llvm/CodeGen/MachineRegisterInfo.h"
35 #include "llvm/CodeGen/WinEHFuncInfo.h"
36 #include "llvm/IR/CallSite.h"
37 #include "llvm/IR/CallingConv.h"
38 #include "llvm/IR/Constants.h"
39 #include "llvm/IR/DerivedTypes.h"
40 #include "llvm/IR/Function.h"
41 #include "llvm/IR/GlobalAlias.h"
42 #include "llvm/IR/GlobalVariable.h"
43 #include "llvm/IR/Instructions.h"
44 #include "llvm/IR/Intrinsics.h"
45 #include "llvm/MC/MCAsmInfo.h"
46 #include "llvm/MC/MCContext.h"
47 #include "llvm/MC/MCExpr.h"
48 #include "llvm/MC/MCSymbol.h"
49 #include "llvm/Support/CommandLine.h"
50 #include "llvm/Support/Debug.h"
51 #include "llvm/Support/ErrorHandling.h"
52 #include "llvm/Support/MathExtras.h"
53 #include "llvm/Target/TargetOptions.h"
54 #include "X86IntrinsicsInfo.h"
60 #define DEBUG_TYPE "x86-isel"
62 STATISTIC(NumTailCalls, "Number of tail calls");
64 static cl::opt<bool> ExperimentalVectorWideningLegalization(
65 "x86-experimental-vector-widening-legalization", cl::init(false),
66 cl::desc("Enable an experimental vector type legalization through widening "
67 "rather than promotion."),
70 X86TargetLowering::X86TargetLowering(const X86TargetMachine &TM,
71 const X86Subtarget &STI)
72 : TargetLowering(TM), Subtarget(&STI) {
73 X86ScalarSSEf64 = Subtarget->hasSSE2();
74 X86ScalarSSEf32 = Subtarget->hasSSE1();
75 MVT PtrVT = MVT::getIntegerVT(8 * TM.getPointerSize());
77 // Set up the TargetLowering object.
78 static const MVT IntVTs[] = { MVT::i8, MVT::i16, MVT::i32, MVT::i64 };
80 // X86 is weird. It always uses i8 for shift amounts and setcc results.
81 setBooleanContents(ZeroOrOneBooleanContent);
82 // X86-SSE is even stranger. It uses -1 or 0 for vector masks.
83 setBooleanVectorContents(ZeroOrNegativeOneBooleanContent);
85 // For 64-bit, since we have so many registers, use the ILP scheduler.
86 // For 32-bit, use the register pressure specific scheduling.
87 // For Atom, always use ILP scheduling.
88 if (Subtarget->isAtom())
89 setSchedulingPreference(Sched::ILP);
90 else if (Subtarget->is64Bit())
91 setSchedulingPreference(Sched::ILP);
93 setSchedulingPreference(Sched::RegPressure);
94 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
95 setStackPointerRegisterToSaveRestore(RegInfo->getStackRegister());
97 // Bypass expensive divides on Atom when compiling with O2.
98 if (TM.getOptLevel() >= CodeGenOpt::Default) {
99 if (Subtarget->hasSlowDivide32())
100 addBypassSlowDiv(32, 8);
101 if (Subtarget->hasSlowDivide64() && Subtarget->is64Bit())
102 addBypassSlowDiv(64, 16);
105 if (Subtarget->isTargetKnownWindowsMSVC()) {
106 // Setup Windows compiler runtime calls.
107 setLibcallName(RTLIB::SDIV_I64, "_alldiv");
108 setLibcallName(RTLIB::UDIV_I64, "_aulldiv");
109 setLibcallName(RTLIB::SREM_I64, "_allrem");
110 setLibcallName(RTLIB::UREM_I64, "_aullrem");
111 setLibcallName(RTLIB::MUL_I64, "_allmul");
112 setLibcallCallingConv(RTLIB::SDIV_I64, CallingConv::X86_StdCall);
113 setLibcallCallingConv(RTLIB::UDIV_I64, CallingConv::X86_StdCall);
114 setLibcallCallingConv(RTLIB::SREM_I64, CallingConv::X86_StdCall);
115 setLibcallCallingConv(RTLIB::UREM_I64, CallingConv::X86_StdCall);
116 setLibcallCallingConv(RTLIB::MUL_I64, CallingConv::X86_StdCall);
119 if (Subtarget->isTargetDarwin()) {
120 // Darwin should use _setjmp/_longjmp instead of setjmp/longjmp.
121 setUseUnderscoreSetJmp(false);
122 setUseUnderscoreLongJmp(false);
123 } else if (Subtarget->isTargetWindowsGNU()) {
124 // MS runtime is weird: it exports _setjmp, but longjmp!
125 setUseUnderscoreSetJmp(true);
126 setUseUnderscoreLongJmp(false);
128 setUseUnderscoreSetJmp(true);
129 setUseUnderscoreLongJmp(true);
132 // Set up the register classes.
133 addRegisterClass(MVT::i8, &X86::GR8RegClass);
134 addRegisterClass(MVT::i16, &X86::GR16RegClass);
135 addRegisterClass(MVT::i32, &X86::GR32RegClass);
136 if (Subtarget->is64Bit())
137 addRegisterClass(MVT::i64, &X86::GR64RegClass);
139 for (MVT VT : MVT::integer_valuetypes())
140 setLoadExtAction(ISD::SEXTLOAD, VT, MVT::i1, Promote);
142 // We don't accept any truncstore of integer registers.
143 setTruncStoreAction(MVT::i64, MVT::i32, Expand);
144 setTruncStoreAction(MVT::i64, MVT::i16, Expand);
145 setTruncStoreAction(MVT::i64, MVT::i8 , Expand);
146 setTruncStoreAction(MVT::i32, MVT::i16, Expand);
147 setTruncStoreAction(MVT::i32, MVT::i8 , Expand);
148 setTruncStoreAction(MVT::i16, MVT::i8, Expand);
150 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
152 // SETOEQ and SETUNE require checking two conditions.
153 setCondCodeAction(ISD::SETOEQ, MVT::f32, Expand);
154 setCondCodeAction(ISD::SETOEQ, MVT::f64, Expand);
155 setCondCodeAction(ISD::SETOEQ, MVT::f80, Expand);
156 setCondCodeAction(ISD::SETUNE, MVT::f32, Expand);
157 setCondCodeAction(ISD::SETUNE, MVT::f64, Expand);
158 setCondCodeAction(ISD::SETUNE, MVT::f80, Expand);
160 // Promote all UINT_TO_FP to larger SINT_TO_FP's, as X86 doesn't have this
162 setOperationAction(ISD::UINT_TO_FP , MVT::i1 , Promote);
163 setOperationAction(ISD::UINT_TO_FP , MVT::i8 , Promote);
164 setOperationAction(ISD::UINT_TO_FP , MVT::i16 , Promote);
166 if (Subtarget->is64Bit()) {
167 if (!Subtarget->useSoftFloat() && Subtarget->hasAVX512())
168 // f32/f64 are legal, f80 is custom.
169 setOperationAction(ISD::UINT_TO_FP , MVT::i32 , Custom);
171 setOperationAction(ISD::UINT_TO_FP , MVT::i32 , Promote);
172 setOperationAction(ISD::UINT_TO_FP , MVT::i64 , Custom);
173 } else if (!Subtarget->useSoftFloat()) {
174 // We have an algorithm for SSE2->double, and we turn this into a
175 // 64-bit FILD followed by conditional FADD for other targets.
176 setOperationAction(ISD::UINT_TO_FP , MVT::i64 , Custom);
177 // We have an algorithm for SSE2, and we turn this into a 64-bit
178 // FILD or VCVTUSI2SS/SD for other targets.
179 setOperationAction(ISD::UINT_TO_FP , MVT::i32 , Custom);
182 // Promote i1/i8 SINT_TO_FP to larger SINT_TO_FP's, as X86 doesn't have
184 setOperationAction(ISD::SINT_TO_FP , MVT::i1 , Promote);
185 setOperationAction(ISD::SINT_TO_FP , MVT::i8 , Promote);
187 if (!Subtarget->useSoftFloat()) {
188 // SSE has no i16 to fp conversion, only i32
189 if (X86ScalarSSEf32) {
190 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Promote);
191 // f32 and f64 cases are Legal, f80 case is not
192 setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Custom);
194 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Custom);
195 setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Custom);
198 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Promote);
199 setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Promote);
202 // In 32-bit mode these are custom lowered. In 64-bit mode F32 and F64
203 // are Legal, f80 is custom lowered.
204 setOperationAction(ISD::FP_TO_SINT , MVT::i64 , Custom);
205 setOperationAction(ISD::SINT_TO_FP , MVT::i64 , Custom);
207 // Promote i1/i8 FP_TO_SINT to larger FP_TO_SINTS's, as X86 doesn't have
209 setOperationAction(ISD::FP_TO_SINT , MVT::i1 , Promote);
210 setOperationAction(ISD::FP_TO_SINT , MVT::i8 , Promote);
212 if (X86ScalarSSEf32) {
213 setOperationAction(ISD::FP_TO_SINT , MVT::i16 , Promote);
214 // f32 and f64 cases are Legal, f80 case is not
215 setOperationAction(ISD::FP_TO_SINT , MVT::i32 , Custom);
217 setOperationAction(ISD::FP_TO_SINT , MVT::i16 , Custom);
218 setOperationAction(ISD::FP_TO_SINT , MVT::i32 , Custom);
221 // Handle FP_TO_UINT by promoting the destination to a larger signed
223 setOperationAction(ISD::FP_TO_UINT , MVT::i1 , Promote);
224 setOperationAction(ISD::FP_TO_UINT , MVT::i8 , Promote);
225 setOperationAction(ISD::FP_TO_UINT , MVT::i16 , Promote);
227 if (Subtarget->is64Bit()) {
228 if (!Subtarget->useSoftFloat() && Subtarget->hasAVX512()) {
229 // FP_TO_UINT-i32/i64 is legal for f32/f64, but custom for f80.
230 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Custom);
231 setOperationAction(ISD::FP_TO_UINT , MVT::i64 , Custom);
233 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Promote);
234 setOperationAction(ISD::FP_TO_UINT , MVT::i64 , Expand);
236 } else if (!Subtarget->useSoftFloat()) {
237 // Since AVX is a superset of SSE3, only check for SSE here.
238 if (Subtarget->hasSSE1() && !Subtarget->hasSSE3())
239 // Expand FP_TO_UINT into a select.
240 // FIXME: We would like to use a Custom expander here eventually to do
241 // the optimal thing for SSE vs. the default expansion in the legalizer.
242 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Expand);
244 // With AVX512 we can use vcvts[ds]2usi for f32/f64->i32, f80 is custom.
245 // With SSE3 we can use fisttpll to convert to a signed i64; without
246 // SSE, we're stuck with a fistpll.
247 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Custom);
249 setOperationAction(ISD::FP_TO_UINT , MVT::i64 , Custom);
252 // TODO: when we have SSE, these could be more efficient, by using movd/movq.
253 if (!X86ScalarSSEf64) {
254 setOperationAction(ISD::BITCAST , MVT::f32 , Expand);
255 setOperationAction(ISD::BITCAST , MVT::i32 , Expand);
256 if (Subtarget->is64Bit()) {
257 setOperationAction(ISD::BITCAST , MVT::f64 , Expand);
258 // Without SSE, i64->f64 goes through memory.
259 setOperationAction(ISD::BITCAST , MVT::i64 , Expand);
263 // Scalar integer divide and remainder are lowered to use operations that
264 // produce two results, to match the available instructions. This exposes
265 // the two-result form to trivial CSE, which is able to combine x/y and x%y
266 // into a single instruction.
268 // Scalar integer multiply-high is also lowered to use two-result
269 // operations, to match the available instructions. However, plain multiply
270 // (low) operations are left as Legal, as there are single-result
271 // instructions for this in x86. Using the two-result multiply instructions
272 // when both high and low results are needed must be arranged by dagcombine.
273 for (unsigned i = 0; i != array_lengthof(IntVTs); ++i) {
275 setOperationAction(ISD::MULHS, VT, Expand);
276 setOperationAction(ISD::MULHU, VT, Expand);
277 setOperationAction(ISD::SDIV, VT, Expand);
278 setOperationAction(ISD::UDIV, VT, Expand);
279 setOperationAction(ISD::SREM, VT, Expand);
280 setOperationAction(ISD::UREM, VT, Expand);
282 // Add/Sub overflow ops with MVT::Glues are lowered to EFLAGS dependences.
283 setOperationAction(ISD::ADDC, VT, Custom);
284 setOperationAction(ISD::ADDE, VT, Custom);
285 setOperationAction(ISD::SUBC, VT, Custom);
286 setOperationAction(ISD::SUBE, VT, Custom);
289 setOperationAction(ISD::BR_JT , MVT::Other, Expand);
290 setOperationAction(ISD::BRCOND , MVT::Other, Custom);
291 setOperationAction(ISD::BR_CC , MVT::f32, Expand);
292 setOperationAction(ISD::BR_CC , MVT::f64, Expand);
293 setOperationAction(ISD::BR_CC , MVT::f80, Expand);
294 setOperationAction(ISD::BR_CC , MVT::i8, Expand);
295 setOperationAction(ISD::BR_CC , MVT::i16, Expand);
296 setOperationAction(ISD::BR_CC , MVT::i32, Expand);
297 setOperationAction(ISD::BR_CC , MVT::i64, Expand);
298 setOperationAction(ISD::SELECT_CC , MVT::f32, Expand);
299 setOperationAction(ISD::SELECT_CC , MVT::f64, Expand);
300 setOperationAction(ISD::SELECT_CC , MVT::f80, Expand);
301 setOperationAction(ISD::SELECT_CC , MVT::i8, Expand);
302 setOperationAction(ISD::SELECT_CC , MVT::i16, Expand);
303 setOperationAction(ISD::SELECT_CC , MVT::i32, Expand);
304 setOperationAction(ISD::SELECT_CC , MVT::i64, Expand);
305 if (Subtarget->is64Bit())
306 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i32, Legal);
307 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16 , Legal);
308 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8 , Legal);
309 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1 , Expand);
310 setOperationAction(ISD::FP_ROUND_INREG , MVT::f32 , Expand);
312 if (Subtarget->is32Bit() && Subtarget->isTargetKnownWindowsMSVC()) {
313 // On 32 bit MSVC, `fmodf(f32)` is not defined - only `fmod(f64)`
314 // is. We should promote the value to 64-bits to solve this.
315 // This is what the CRT headers do - `fmodf` is an inline header
316 // function casting to f64 and calling `fmod`.
317 setOperationAction(ISD::FREM , MVT::f32 , Promote);
319 setOperationAction(ISD::FREM , MVT::f32 , Expand);
322 setOperationAction(ISD::FREM , MVT::f64 , Expand);
323 setOperationAction(ISD::FREM , MVT::f80 , Expand);
324 setOperationAction(ISD::FLT_ROUNDS_ , MVT::i32 , Custom);
326 // Promote the i8 variants and force them on up to i32 which has a shorter
328 setOperationAction(ISD::CTTZ , MVT::i8 , Promote);
329 AddPromotedToType (ISD::CTTZ , MVT::i8 , MVT::i32);
330 setOperationAction(ISD::CTTZ_ZERO_UNDEF , MVT::i8 , Promote);
331 AddPromotedToType (ISD::CTTZ_ZERO_UNDEF , MVT::i8 , MVT::i32);
332 if (Subtarget->hasBMI()) {
333 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::i16 , Expand);
334 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::i32 , Expand);
335 if (Subtarget->is64Bit())
336 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::i64, Expand);
338 setOperationAction(ISD::CTTZ , MVT::i16 , Custom);
339 setOperationAction(ISD::CTTZ , MVT::i32 , Custom);
340 if (Subtarget->is64Bit())
341 setOperationAction(ISD::CTTZ , MVT::i64 , Custom);
344 if (Subtarget->hasLZCNT()) {
345 // When promoting the i8 variants, force them to i32 for a shorter
347 setOperationAction(ISD::CTLZ , MVT::i8 , Promote);
348 AddPromotedToType (ISD::CTLZ , MVT::i8 , MVT::i32);
349 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i8 , Promote);
350 AddPromotedToType (ISD::CTLZ_ZERO_UNDEF, MVT::i8 , MVT::i32);
351 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i16 , Expand);
352 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i32 , Expand);
353 if (Subtarget->is64Bit())
354 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i64, Expand);
356 setOperationAction(ISD::CTLZ , MVT::i8 , Custom);
357 setOperationAction(ISD::CTLZ , MVT::i16 , Custom);
358 setOperationAction(ISD::CTLZ , MVT::i32 , Custom);
359 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i8 , Custom);
360 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i16 , Custom);
361 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i32 , Custom);
362 if (Subtarget->is64Bit()) {
363 setOperationAction(ISD::CTLZ , MVT::i64 , Custom);
364 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i64, Custom);
368 // Special handling for half-precision floating point conversions.
369 // If we don't have F16C support, then lower half float conversions
370 // into library calls.
371 if (Subtarget->useSoftFloat() || !Subtarget->hasF16C()) {
372 setOperationAction(ISD::FP16_TO_FP, MVT::f32, Expand);
373 setOperationAction(ISD::FP_TO_FP16, MVT::f32, Expand);
376 // There's never any support for operations beyond MVT::f32.
377 setOperationAction(ISD::FP16_TO_FP, MVT::f64, Expand);
378 setOperationAction(ISD::FP16_TO_FP, MVT::f80, Expand);
379 setOperationAction(ISD::FP_TO_FP16, MVT::f64, Expand);
380 setOperationAction(ISD::FP_TO_FP16, MVT::f80, Expand);
382 setLoadExtAction(ISD::EXTLOAD, MVT::f32, MVT::f16, Expand);
383 setLoadExtAction(ISD::EXTLOAD, MVT::f64, MVT::f16, Expand);
384 setLoadExtAction(ISD::EXTLOAD, MVT::f80, MVT::f16, Expand);
385 setTruncStoreAction(MVT::f32, MVT::f16, Expand);
386 setTruncStoreAction(MVT::f64, MVT::f16, Expand);
387 setTruncStoreAction(MVT::f80, MVT::f16, Expand);
389 if (Subtarget->hasPOPCNT()) {
390 setOperationAction(ISD::CTPOP , MVT::i8 , Promote);
392 setOperationAction(ISD::CTPOP , MVT::i8 , Expand);
393 setOperationAction(ISD::CTPOP , MVT::i16 , Expand);
394 setOperationAction(ISD::CTPOP , MVT::i32 , Expand);
395 if (Subtarget->is64Bit())
396 setOperationAction(ISD::CTPOP , MVT::i64 , Expand);
399 setOperationAction(ISD::READCYCLECOUNTER , MVT::i64 , Custom);
401 if (!Subtarget->hasMOVBE())
402 setOperationAction(ISD::BSWAP , MVT::i16 , Expand);
404 // These should be promoted to a larger select which is supported.
405 setOperationAction(ISD::SELECT , MVT::i1 , Promote);
406 // X86 wants to expand cmov itself.
407 setOperationAction(ISD::SELECT , MVT::i8 , Custom);
408 setOperationAction(ISD::SELECT , MVT::i16 , Custom);
409 setOperationAction(ISD::SELECT , MVT::i32 , Custom);
410 setOperationAction(ISD::SELECT , MVT::f32 , Custom);
411 setOperationAction(ISD::SELECT , MVT::f64 , Custom);
412 setOperationAction(ISD::SELECT , MVT::f80 , Custom);
413 setOperationAction(ISD::SETCC , MVT::i8 , Custom);
414 setOperationAction(ISD::SETCC , MVT::i16 , Custom);
415 setOperationAction(ISD::SETCC , MVT::i32 , Custom);
416 setOperationAction(ISD::SETCC , MVT::f32 , Custom);
417 setOperationAction(ISD::SETCC , MVT::f64 , Custom);
418 setOperationAction(ISD::SETCC , MVT::f80 , Custom);
419 if (Subtarget->is64Bit()) {
420 setOperationAction(ISD::SELECT , MVT::i64 , Custom);
421 setOperationAction(ISD::SETCC , MVT::i64 , Custom);
423 setOperationAction(ISD::EH_RETURN , MVT::Other, Custom);
424 // NOTE: EH_SJLJ_SETJMP/_LONGJMP supported here is NOT intended to support
425 // SjLj exception handling but a light-weight setjmp/longjmp replacement to
426 // support continuation, user-level threading, and etc.. As a result, no
427 // other SjLj exception interfaces are implemented and please don't build
428 // your own exception handling based on them.
429 // LLVM/Clang supports zero-cost DWARF exception handling.
430 setOperationAction(ISD::EH_SJLJ_SETJMP, MVT::i32, Custom);
431 setOperationAction(ISD::EH_SJLJ_LONGJMP, MVT::Other, Custom);
434 setOperationAction(ISD::ConstantPool , MVT::i32 , Custom);
435 setOperationAction(ISD::JumpTable , MVT::i32 , Custom);
436 setOperationAction(ISD::GlobalAddress , MVT::i32 , Custom);
437 setOperationAction(ISD::GlobalTLSAddress, MVT::i32 , Custom);
438 if (Subtarget->is64Bit())
439 setOperationAction(ISD::GlobalTLSAddress, MVT::i64, Custom);
440 setOperationAction(ISD::ExternalSymbol , MVT::i32 , Custom);
441 setOperationAction(ISD::BlockAddress , MVT::i32 , Custom);
442 if (Subtarget->is64Bit()) {
443 setOperationAction(ISD::ConstantPool , MVT::i64 , Custom);
444 setOperationAction(ISD::JumpTable , MVT::i64 , Custom);
445 setOperationAction(ISD::GlobalAddress , MVT::i64 , Custom);
446 setOperationAction(ISD::ExternalSymbol, MVT::i64 , Custom);
447 setOperationAction(ISD::BlockAddress , MVT::i64 , Custom);
449 // 64-bit addm sub, shl, sra, srl (iff 32-bit x86)
450 setOperationAction(ISD::SHL_PARTS , MVT::i32 , Custom);
451 setOperationAction(ISD::SRA_PARTS , MVT::i32 , Custom);
452 setOperationAction(ISD::SRL_PARTS , MVT::i32 , Custom);
453 if (Subtarget->is64Bit()) {
454 setOperationAction(ISD::SHL_PARTS , MVT::i64 , Custom);
455 setOperationAction(ISD::SRA_PARTS , MVT::i64 , Custom);
456 setOperationAction(ISD::SRL_PARTS , MVT::i64 , Custom);
459 if (Subtarget->hasSSE1())
460 setOperationAction(ISD::PREFETCH , MVT::Other, Legal);
462 setOperationAction(ISD::ATOMIC_FENCE , MVT::Other, Custom);
464 // Expand certain atomics
465 for (unsigned i = 0; i != array_lengthof(IntVTs); ++i) {
467 setOperationAction(ISD::ATOMIC_CMP_SWAP_WITH_SUCCESS, VT, Custom);
468 setOperationAction(ISD::ATOMIC_LOAD_SUB, VT, Custom);
469 setOperationAction(ISD::ATOMIC_STORE, VT, Custom);
472 if (Subtarget->hasCmpxchg16b()) {
473 setOperationAction(ISD::ATOMIC_CMP_SWAP_WITH_SUCCESS, MVT::i128, Custom);
476 // FIXME - use subtarget debug flags
477 if (!Subtarget->isTargetDarwin() && !Subtarget->isTargetELF() &&
478 !Subtarget->isTargetCygMing() && !Subtarget->isTargetWin64()) {
479 setOperationAction(ISD::EH_LABEL, MVT::Other, Expand);
482 if (Subtarget->isTarget64BitLP64()) {
483 setExceptionPointerRegister(X86::RAX);
484 setExceptionSelectorRegister(X86::RDX);
486 setExceptionPointerRegister(X86::EAX);
487 setExceptionSelectorRegister(X86::EDX);
489 setOperationAction(ISD::FRAME_TO_ARGS_OFFSET, MVT::i32, Custom);
490 setOperationAction(ISD::FRAME_TO_ARGS_OFFSET, MVT::i64, Custom);
492 setOperationAction(ISD::INIT_TRAMPOLINE, MVT::Other, Custom);
493 setOperationAction(ISD::ADJUST_TRAMPOLINE, MVT::Other, Custom);
495 setOperationAction(ISD::TRAP, MVT::Other, Legal);
496 setOperationAction(ISD::DEBUGTRAP, MVT::Other, Legal);
498 // VASTART needs to be custom lowered to use the VarArgsFrameIndex
499 setOperationAction(ISD::VASTART , MVT::Other, Custom);
500 setOperationAction(ISD::VAEND , MVT::Other, Expand);
501 if (Subtarget->is64Bit()) {
502 setOperationAction(ISD::VAARG , MVT::Other, Custom);
503 setOperationAction(ISD::VACOPY , MVT::Other, Custom);
505 // TargetInfo::CharPtrBuiltinVaList
506 setOperationAction(ISD::VAARG , MVT::Other, Expand);
507 setOperationAction(ISD::VACOPY , MVT::Other, Expand);
510 setOperationAction(ISD::STACKSAVE, MVT::Other, Expand);
511 setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand);
513 setOperationAction(ISD::DYNAMIC_STACKALLOC, PtrVT, Custom);
515 // GC_TRANSITION_START and GC_TRANSITION_END need custom lowering.
516 setOperationAction(ISD::GC_TRANSITION_START, MVT::Other, Custom);
517 setOperationAction(ISD::GC_TRANSITION_END, MVT::Other, Custom);
519 if (!Subtarget->useSoftFloat() && X86ScalarSSEf64) {
520 // f32 and f64 use SSE.
521 // Set up the FP register classes.
522 addRegisterClass(MVT::f32, &X86::FR32RegClass);
523 addRegisterClass(MVT::f64, &X86::FR64RegClass);
525 // Use ANDPD to simulate FABS.
526 setOperationAction(ISD::FABS , MVT::f64, Custom);
527 setOperationAction(ISD::FABS , MVT::f32, Custom);
529 // Use XORP to simulate FNEG.
530 setOperationAction(ISD::FNEG , MVT::f64, Custom);
531 setOperationAction(ISD::FNEG , MVT::f32, Custom);
533 // Use ANDPD and ORPD to simulate FCOPYSIGN.
534 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Custom);
535 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
537 // Lower this to FGETSIGNx86 plus an AND.
538 setOperationAction(ISD::FGETSIGN, MVT::i64, Custom);
539 setOperationAction(ISD::FGETSIGN, MVT::i32, Custom);
541 // We don't support sin/cos/fmod
542 setOperationAction(ISD::FSIN , MVT::f64, Expand);
543 setOperationAction(ISD::FCOS , MVT::f64, Expand);
544 setOperationAction(ISD::FSINCOS, MVT::f64, Expand);
545 setOperationAction(ISD::FSIN , MVT::f32, Expand);
546 setOperationAction(ISD::FCOS , MVT::f32, Expand);
547 setOperationAction(ISD::FSINCOS, MVT::f32, Expand);
549 // Expand FP immediates into loads from the stack, except for the special
551 addLegalFPImmediate(APFloat(+0.0)); // xorpd
552 addLegalFPImmediate(APFloat(+0.0f)); // xorps
553 } else if (!Subtarget->useSoftFloat() && X86ScalarSSEf32) {
554 // Use SSE for f32, x87 for f64.
555 // Set up the FP register classes.
556 addRegisterClass(MVT::f32, &X86::FR32RegClass);
557 addRegisterClass(MVT::f64, &X86::RFP64RegClass);
559 // Use ANDPS to simulate FABS.
560 setOperationAction(ISD::FABS , MVT::f32, Custom);
562 // Use XORP to simulate FNEG.
563 setOperationAction(ISD::FNEG , MVT::f32, Custom);
565 setOperationAction(ISD::UNDEF, MVT::f64, Expand);
567 // Use ANDPS and ORPS to simulate FCOPYSIGN.
568 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
569 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
571 // We don't support sin/cos/fmod
572 setOperationAction(ISD::FSIN , MVT::f32, Expand);
573 setOperationAction(ISD::FCOS , MVT::f32, Expand);
574 setOperationAction(ISD::FSINCOS, MVT::f32, Expand);
576 // Special cases we handle for FP constants.
577 addLegalFPImmediate(APFloat(+0.0f)); // xorps
578 addLegalFPImmediate(APFloat(+0.0)); // FLD0
579 addLegalFPImmediate(APFloat(+1.0)); // FLD1
580 addLegalFPImmediate(APFloat(-0.0)); // FLD0/FCHS
581 addLegalFPImmediate(APFloat(-1.0)); // FLD1/FCHS
583 if (!TM.Options.UnsafeFPMath) {
584 setOperationAction(ISD::FSIN , MVT::f64, Expand);
585 setOperationAction(ISD::FCOS , MVT::f64, Expand);
586 setOperationAction(ISD::FSINCOS, MVT::f64, Expand);
588 } else if (!Subtarget->useSoftFloat()) {
589 // f32 and f64 in x87.
590 // Set up the FP register classes.
591 addRegisterClass(MVT::f64, &X86::RFP64RegClass);
592 addRegisterClass(MVT::f32, &X86::RFP32RegClass);
594 setOperationAction(ISD::UNDEF, MVT::f64, Expand);
595 setOperationAction(ISD::UNDEF, MVT::f32, Expand);
596 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
597 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Expand);
599 if (!TM.Options.UnsafeFPMath) {
600 setOperationAction(ISD::FSIN , MVT::f64, Expand);
601 setOperationAction(ISD::FSIN , MVT::f32, Expand);
602 setOperationAction(ISD::FCOS , MVT::f64, Expand);
603 setOperationAction(ISD::FCOS , MVT::f32, Expand);
604 setOperationAction(ISD::FSINCOS, MVT::f64, Expand);
605 setOperationAction(ISD::FSINCOS, MVT::f32, Expand);
607 addLegalFPImmediate(APFloat(+0.0)); // FLD0
608 addLegalFPImmediate(APFloat(+1.0)); // FLD1
609 addLegalFPImmediate(APFloat(-0.0)); // FLD0/FCHS
610 addLegalFPImmediate(APFloat(-1.0)); // FLD1/FCHS
611 addLegalFPImmediate(APFloat(+0.0f)); // FLD0
612 addLegalFPImmediate(APFloat(+1.0f)); // FLD1
613 addLegalFPImmediate(APFloat(-0.0f)); // FLD0/FCHS
614 addLegalFPImmediate(APFloat(-1.0f)); // FLD1/FCHS
617 // We don't support FMA.
618 setOperationAction(ISD::FMA, MVT::f64, Expand);
619 setOperationAction(ISD::FMA, MVT::f32, Expand);
621 // Long double always uses X87.
622 if (!Subtarget->useSoftFloat()) {
623 addRegisterClass(MVT::f80, &X86::RFP80RegClass);
624 setOperationAction(ISD::UNDEF, MVT::f80, Expand);
625 setOperationAction(ISD::FCOPYSIGN, MVT::f80, Expand);
627 APFloat TmpFlt = APFloat::getZero(APFloat::x87DoubleExtended);
628 addLegalFPImmediate(TmpFlt); // FLD0
630 addLegalFPImmediate(TmpFlt); // FLD0/FCHS
633 APFloat TmpFlt2(+1.0);
634 TmpFlt2.convert(APFloat::x87DoubleExtended, APFloat::rmNearestTiesToEven,
636 addLegalFPImmediate(TmpFlt2); // FLD1
637 TmpFlt2.changeSign();
638 addLegalFPImmediate(TmpFlt2); // FLD1/FCHS
641 if (!TM.Options.UnsafeFPMath) {
642 setOperationAction(ISD::FSIN , MVT::f80, Expand);
643 setOperationAction(ISD::FCOS , MVT::f80, Expand);
644 setOperationAction(ISD::FSINCOS, MVT::f80, Expand);
647 setOperationAction(ISD::FFLOOR, MVT::f80, Expand);
648 setOperationAction(ISD::FCEIL, MVT::f80, Expand);
649 setOperationAction(ISD::FTRUNC, MVT::f80, Expand);
650 setOperationAction(ISD::FRINT, MVT::f80, Expand);
651 setOperationAction(ISD::FNEARBYINT, MVT::f80, Expand);
652 setOperationAction(ISD::FMA, MVT::f80, Expand);
655 // Always use a library call for pow.
656 setOperationAction(ISD::FPOW , MVT::f32 , Expand);
657 setOperationAction(ISD::FPOW , MVT::f64 , Expand);
658 setOperationAction(ISD::FPOW , MVT::f80 , Expand);
660 setOperationAction(ISD::FLOG, MVT::f80, Expand);
661 setOperationAction(ISD::FLOG2, MVT::f80, Expand);
662 setOperationAction(ISD::FLOG10, MVT::f80, Expand);
663 setOperationAction(ISD::FEXP, MVT::f80, Expand);
664 setOperationAction(ISD::FEXP2, MVT::f80, Expand);
665 setOperationAction(ISD::FMINNUM, MVT::f80, Expand);
666 setOperationAction(ISD::FMAXNUM, MVT::f80, Expand);
668 // First set operation action for all vector types to either promote
669 // (for widening) or expand (for scalarization). Then we will selectively
670 // turn on ones that can be effectively codegen'd.
671 for (MVT VT : MVT::vector_valuetypes()) {
672 setOperationAction(ISD::ADD , VT, Expand);
673 setOperationAction(ISD::SUB , VT, Expand);
674 setOperationAction(ISD::FADD, VT, Expand);
675 setOperationAction(ISD::FNEG, VT, Expand);
676 setOperationAction(ISD::FSUB, VT, Expand);
677 setOperationAction(ISD::MUL , VT, Expand);
678 setOperationAction(ISD::FMUL, VT, Expand);
679 setOperationAction(ISD::SDIV, VT, Expand);
680 setOperationAction(ISD::UDIV, VT, Expand);
681 setOperationAction(ISD::FDIV, VT, Expand);
682 setOperationAction(ISD::SREM, VT, Expand);
683 setOperationAction(ISD::UREM, VT, Expand);
684 setOperationAction(ISD::LOAD, VT, Expand);
685 setOperationAction(ISD::VECTOR_SHUFFLE, VT, Expand);
686 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT,Expand);
687 setOperationAction(ISD::INSERT_VECTOR_ELT, VT, Expand);
688 setOperationAction(ISD::EXTRACT_SUBVECTOR, VT,Expand);
689 setOperationAction(ISD::INSERT_SUBVECTOR, VT,Expand);
690 setOperationAction(ISD::FABS, VT, Expand);
691 setOperationAction(ISD::FSIN, VT, Expand);
692 setOperationAction(ISD::FSINCOS, VT, Expand);
693 setOperationAction(ISD::FCOS, VT, Expand);
694 setOperationAction(ISD::FSINCOS, VT, Expand);
695 setOperationAction(ISD::FREM, VT, Expand);
696 setOperationAction(ISD::FMA, VT, Expand);
697 setOperationAction(ISD::FPOWI, VT, Expand);
698 setOperationAction(ISD::FSQRT, VT, Expand);
699 setOperationAction(ISD::FCOPYSIGN, VT, Expand);
700 setOperationAction(ISD::FFLOOR, VT, Expand);
701 setOperationAction(ISD::FCEIL, VT, Expand);
702 setOperationAction(ISD::FTRUNC, VT, Expand);
703 setOperationAction(ISD::FRINT, VT, Expand);
704 setOperationAction(ISD::FNEARBYINT, VT, Expand);
705 setOperationAction(ISD::SMUL_LOHI, VT, Expand);
706 setOperationAction(ISD::MULHS, VT, Expand);
707 setOperationAction(ISD::UMUL_LOHI, VT, Expand);
708 setOperationAction(ISD::MULHU, VT, Expand);
709 setOperationAction(ISD::SDIVREM, VT, Expand);
710 setOperationAction(ISD::UDIVREM, VT, Expand);
711 setOperationAction(ISD::FPOW, VT, Expand);
712 setOperationAction(ISD::CTPOP, VT, Expand);
713 setOperationAction(ISD::CTTZ, VT, Expand);
714 setOperationAction(ISD::CTTZ_ZERO_UNDEF, VT, Expand);
715 setOperationAction(ISD::CTLZ, VT, Expand);
716 setOperationAction(ISD::CTLZ_ZERO_UNDEF, VT, Expand);
717 setOperationAction(ISD::SHL, VT, Expand);
718 setOperationAction(ISD::SRA, VT, Expand);
719 setOperationAction(ISD::SRL, VT, Expand);
720 setOperationAction(ISD::ROTL, VT, Expand);
721 setOperationAction(ISD::ROTR, VT, Expand);
722 setOperationAction(ISD::BSWAP, VT, Expand);
723 setOperationAction(ISD::SETCC, VT, Expand);
724 setOperationAction(ISD::FLOG, VT, Expand);
725 setOperationAction(ISD::FLOG2, VT, Expand);
726 setOperationAction(ISD::FLOG10, VT, Expand);
727 setOperationAction(ISD::FEXP, VT, Expand);
728 setOperationAction(ISD::FEXP2, VT, Expand);
729 setOperationAction(ISD::FP_TO_UINT, VT, Expand);
730 setOperationAction(ISD::FP_TO_SINT, VT, Expand);
731 setOperationAction(ISD::UINT_TO_FP, VT, Expand);
732 setOperationAction(ISD::SINT_TO_FP, VT, Expand);
733 setOperationAction(ISD::SIGN_EXTEND_INREG, VT,Expand);
734 setOperationAction(ISD::TRUNCATE, VT, Expand);
735 setOperationAction(ISD::SIGN_EXTEND, VT, Expand);
736 setOperationAction(ISD::ZERO_EXTEND, VT, Expand);
737 setOperationAction(ISD::ANY_EXTEND, VT, Expand);
738 setOperationAction(ISD::VSELECT, VT, Expand);
739 setOperationAction(ISD::SELECT_CC, VT, Expand);
740 for (MVT InnerVT : MVT::vector_valuetypes()) {
741 setTruncStoreAction(InnerVT, VT, Expand);
743 setLoadExtAction(ISD::SEXTLOAD, InnerVT, VT, Expand);
744 setLoadExtAction(ISD::ZEXTLOAD, InnerVT, VT, Expand);
746 // N.b. ISD::EXTLOAD legality is basically ignored except for i1-like
747 // types, we have to deal with them whether we ask for Expansion or not.
748 // Setting Expand causes its own optimisation problems though, so leave
750 if (VT.getVectorElementType() == MVT::i1)
751 setLoadExtAction(ISD::EXTLOAD, InnerVT, VT, Expand);
753 // EXTLOAD for MVT::f16 vectors is not legal because f16 vectors are
754 // split/scalarized right now.
755 if (VT.getVectorElementType() == MVT::f16)
756 setLoadExtAction(ISD::EXTLOAD, InnerVT, VT, Expand);
760 // FIXME: In order to prevent SSE instructions being expanded to MMX ones
761 // with -msoft-float, disable use of MMX as well.
762 if (!Subtarget->useSoftFloat() && Subtarget->hasMMX()) {
763 addRegisterClass(MVT::x86mmx, &X86::VR64RegClass);
764 // No operations on x86mmx supported, everything uses intrinsics.
767 // MMX-sized vectors (other than x86mmx) are expected to be expanded
768 // into smaller operations.
769 for (MVT MMXTy : {MVT::v8i8, MVT::v4i16, MVT::v2i32, MVT::v1i64}) {
770 setOperationAction(ISD::MULHS, MMXTy, Expand);
771 setOperationAction(ISD::AND, MMXTy, Expand);
772 setOperationAction(ISD::OR, MMXTy, Expand);
773 setOperationAction(ISD::XOR, MMXTy, Expand);
774 setOperationAction(ISD::SCALAR_TO_VECTOR, MMXTy, Expand);
775 setOperationAction(ISD::SELECT, MMXTy, Expand);
776 setOperationAction(ISD::BITCAST, MMXTy, Expand);
778 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v1i64, Expand);
780 if (!Subtarget->useSoftFloat() && Subtarget->hasSSE1()) {
781 addRegisterClass(MVT::v4f32, &X86::VR128RegClass);
783 setOperationAction(ISD::FADD, MVT::v4f32, Legal);
784 setOperationAction(ISD::FSUB, MVT::v4f32, Legal);
785 setOperationAction(ISD::FMUL, MVT::v4f32, Legal);
786 setOperationAction(ISD::FDIV, MVT::v4f32, Legal);
787 setOperationAction(ISD::FSQRT, MVT::v4f32, Legal);
788 setOperationAction(ISD::FNEG, MVT::v4f32, Custom);
789 setOperationAction(ISD::FABS, MVT::v4f32, Custom);
790 setOperationAction(ISD::LOAD, MVT::v4f32, Legal);
791 setOperationAction(ISD::BUILD_VECTOR, MVT::v4f32, Custom);
792 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4f32, Custom);
793 setOperationAction(ISD::VSELECT, MVT::v4f32, Custom);
794 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f32, Custom);
795 setOperationAction(ISD::SELECT, MVT::v4f32, Custom);
796 setOperationAction(ISD::UINT_TO_FP, MVT::v4i32, Custom);
799 if (!Subtarget->useSoftFloat() && Subtarget->hasSSE2()) {
800 addRegisterClass(MVT::v2f64, &X86::VR128RegClass);
802 // FIXME: Unfortunately, -soft-float and -no-implicit-float mean XMM
803 // registers cannot be used even for integer operations.
804 addRegisterClass(MVT::v16i8, &X86::VR128RegClass);
805 addRegisterClass(MVT::v8i16, &X86::VR128RegClass);
806 addRegisterClass(MVT::v4i32, &X86::VR128RegClass);
807 addRegisterClass(MVT::v2i64, &X86::VR128RegClass);
809 setOperationAction(ISD::ADD, MVT::v16i8, Legal);
810 setOperationAction(ISD::ADD, MVT::v8i16, Legal);
811 setOperationAction(ISD::ADD, MVT::v4i32, Legal);
812 setOperationAction(ISD::ADD, MVT::v2i64, Legal);
813 setOperationAction(ISD::MUL, MVT::v16i8, Custom);
814 setOperationAction(ISD::MUL, MVT::v4i32, Custom);
815 setOperationAction(ISD::MUL, MVT::v2i64, Custom);
816 setOperationAction(ISD::UMUL_LOHI, MVT::v4i32, Custom);
817 setOperationAction(ISD::SMUL_LOHI, MVT::v4i32, Custom);
818 setOperationAction(ISD::MULHU, MVT::v8i16, Legal);
819 setOperationAction(ISD::MULHS, MVT::v8i16, Legal);
820 setOperationAction(ISD::SUB, MVT::v16i8, Legal);
821 setOperationAction(ISD::SUB, MVT::v8i16, Legal);
822 setOperationAction(ISD::SUB, MVT::v4i32, Legal);
823 setOperationAction(ISD::SUB, MVT::v2i64, Legal);
824 setOperationAction(ISD::MUL, MVT::v8i16, Legal);
825 setOperationAction(ISD::FADD, MVT::v2f64, Legal);
826 setOperationAction(ISD::FSUB, MVT::v2f64, Legal);
827 setOperationAction(ISD::FMUL, MVT::v2f64, Legal);
828 setOperationAction(ISD::FDIV, MVT::v2f64, Legal);
829 setOperationAction(ISD::FSQRT, MVT::v2f64, Legal);
830 setOperationAction(ISD::FNEG, MVT::v2f64, Custom);
831 setOperationAction(ISD::FABS, MVT::v2f64, Custom);
833 setOperationAction(ISD::SMAX, MVT::v8i16, Legal);
834 setOperationAction(ISD::UMAX, MVT::v16i8, Legal);
835 setOperationAction(ISD::SMIN, MVT::v8i16, Legal);
836 setOperationAction(ISD::UMIN, MVT::v16i8, Legal);
838 setOperationAction(ISD::SETCC, MVT::v2i64, Custom);
839 setOperationAction(ISD::SETCC, MVT::v16i8, Custom);
840 setOperationAction(ISD::SETCC, MVT::v8i16, Custom);
841 setOperationAction(ISD::SETCC, MVT::v4i32, Custom);
843 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v16i8, Custom);
844 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v8i16, Custom);
845 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i16, Custom);
846 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i32, Custom);
847 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f32, Custom);
849 setOperationAction(ISD::CTPOP, MVT::v16i8, Custom);
850 setOperationAction(ISD::CTPOP, MVT::v8i16, Custom);
851 setOperationAction(ISD::CTPOP, MVT::v4i32, Custom);
852 setOperationAction(ISD::CTPOP, MVT::v2i64, Custom);
854 setOperationAction(ISD::CTTZ, MVT::v16i8, Custom);
855 setOperationAction(ISD::CTTZ, MVT::v8i16, Custom);
856 setOperationAction(ISD::CTTZ, MVT::v4i32, Custom);
857 // ISD::CTTZ v2i64 - scalarization is faster.
858 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::v16i8, Custom);
859 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::v8i16, Custom);
860 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::v4i32, Custom);
861 // ISD::CTTZ_ZERO_UNDEF v2i64 - scalarization is faster.
863 // Custom lower build_vector, vector_shuffle, and extract_vector_elt.
864 for (int i = MVT::v16i8; i != MVT::v2i64; ++i) {
865 MVT VT = (MVT::SimpleValueType)i;
866 // Do not attempt to custom lower non-power-of-2 vectors
867 if (!isPowerOf2_32(VT.getVectorNumElements()))
869 // Do not attempt to custom lower non-128-bit vectors
870 if (!VT.is128BitVector())
872 setOperationAction(ISD::BUILD_VECTOR, VT, Custom);
873 setOperationAction(ISD::VECTOR_SHUFFLE, VT, Custom);
874 setOperationAction(ISD::VSELECT, VT, Custom);
875 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Custom);
878 // We support custom legalizing of sext and anyext loads for specific
879 // memory vector types which we can load as a scalar (or sequence of
880 // scalars) and extend in-register to a legal 128-bit vector type. For sext
881 // loads these must work with a single scalar load.
882 for (MVT VT : MVT::integer_vector_valuetypes()) {
883 setLoadExtAction(ISD::SEXTLOAD, VT, MVT::v4i8, Custom);
884 setLoadExtAction(ISD::SEXTLOAD, VT, MVT::v4i16, Custom);
885 setLoadExtAction(ISD::SEXTLOAD, VT, MVT::v8i8, Custom);
886 setLoadExtAction(ISD::EXTLOAD, VT, MVT::v2i8, Custom);
887 setLoadExtAction(ISD::EXTLOAD, VT, MVT::v2i16, Custom);
888 setLoadExtAction(ISD::EXTLOAD, VT, MVT::v2i32, Custom);
889 setLoadExtAction(ISD::EXTLOAD, VT, MVT::v4i8, Custom);
890 setLoadExtAction(ISD::EXTLOAD, VT, MVT::v4i16, Custom);
891 setLoadExtAction(ISD::EXTLOAD, VT, MVT::v8i8, Custom);
894 setOperationAction(ISD::BUILD_VECTOR, MVT::v2f64, Custom);
895 setOperationAction(ISD::BUILD_VECTOR, MVT::v2i64, Custom);
896 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2f64, Custom);
897 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2i64, Custom);
898 setOperationAction(ISD::VSELECT, MVT::v2f64, Custom);
899 setOperationAction(ISD::VSELECT, MVT::v2i64, Custom);
900 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2f64, Custom);
901 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2f64, Custom);
903 if (Subtarget->is64Bit()) {
904 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2i64, Custom);
905 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2i64, Custom);
908 // Promote v16i8, v8i16, v4i32 load, select, and, or, xor to v2i64.
909 for (int i = MVT::v16i8; i != MVT::v2i64; ++i) {
910 MVT VT = (MVT::SimpleValueType)i;
912 // Do not attempt to promote non-128-bit vectors
913 if (!VT.is128BitVector())
916 setOperationAction(ISD::AND, VT, Promote);
917 AddPromotedToType (ISD::AND, VT, MVT::v2i64);
918 setOperationAction(ISD::OR, VT, Promote);
919 AddPromotedToType (ISD::OR, VT, MVT::v2i64);
920 setOperationAction(ISD::XOR, VT, Promote);
921 AddPromotedToType (ISD::XOR, VT, MVT::v2i64);
922 setOperationAction(ISD::LOAD, VT, Promote);
923 AddPromotedToType (ISD::LOAD, VT, MVT::v2i64);
924 setOperationAction(ISD::SELECT, VT, Promote);
925 AddPromotedToType (ISD::SELECT, VT, MVT::v2i64);
928 // Custom lower v2i64 and v2f64 selects.
929 setOperationAction(ISD::LOAD, MVT::v2f64, Legal);
930 setOperationAction(ISD::LOAD, MVT::v2i64, Legal);
931 setOperationAction(ISD::SELECT, MVT::v2f64, Custom);
932 setOperationAction(ISD::SELECT, MVT::v2i64, Custom);
934 setOperationAction(ISD::FP_TO_SINT, MVT::v4i32, Legal);
935 setOperationAction(ISD::SINT_TO_FP, MVT::v4i32, Legal);
937 setOperationAction(ISD::SINT_TO_FP, MVT::v2i32, Custom);
939 setOperationAction(ISD::UINT_TO_FP, MVT::v4i8, Custom);
940 setOperationAction(ISD::UINT_TO_FP, MVT::v4i16, Custom);
941 // As there is no 64-bit GPR available, we need build a special custom
942 // sequence to convert from v2i32 to v2f32.
943 if (!Subtarget->is64Bit())
944 setOperationAction(ISD::UINT_TO_FP, MVT::v2f32, Custom);
946 setOperationAction(ISD::FP_EXTEND, MVT::v2f32, Custom);
947 setOperationAction(ISD::FP_ROUND, MVT::v2f32, Custom);
949 for (MVT VT : MVT::fp_vector_valuetypes())
950 setLoadExtAction(ISD::EXTLOAD, VT, MVT::v2f32, Legal);
952 setOperationAction(ISD::BITCAST, MVT::v2i32, Custom);
953 setOperationAction(ISD::BITCAST, MVT::v4i16, Custom);
954 setOperationAction(ISD::BITCAST, MVT::v8i8, Custom);
957 if (!Subtarget->useSoftFloat() && Subtarget->hasSSE41()) {
958 for (MVT RoundedTy : {MVT::f32, MVT::f64, MVT::v4f32, MVT::v2f64}) {
959 setOperationAction(ISD::FFLOOR, RoundedTy, Legal);
960 setOperationAction(ISD::FCEIL, RoundedTy, Legal);
961 setOperationAction(ISD::FTRUNC, RoundedTy, Legal);
962 setOperationAction(ISD::FRINT, RoundedTy, Legal);
963 setOperationAction(ISD::FNEARBYINT, RoundedTy, Legal);
966 setOperationAction(ISD::SMAX, MVT::v16i8, Legal);
967 setOperationAction(ISD::SMAX, MVT::v4i32, Legal);
968 setOperationAction(ISD::UMAX, MVT::v8i16, Legal);
969 setOperationAction(ISD::UMAX, MVT::v4i32, Legal);
970 setOperationAction(ISD::SMIN, MVT::v16i8, Legal);
971 setOperationAction(ISD::SMIN, MVT::v4i32, Legal);
972 setOperationAction(ISD::UMIN, MVT::v8i16, Legal);
973 setOperationAction(ISD::UMIN, MVT::v4i32, Legal);
975 // FIXME: Do we need to handle scalar-to-vector here?
976 setOperationAction(ISD::MUL, MVT::v4i32, Legal);
978 // We directly match byte blends in the backend as they match the VSELECT
980 setOperationAction(ISD::VSELECT, MVT::v16i8, Legal);
982 // SSE41 brings specific instructions for doing vector sign extend even in
983 // cases where we don't have SRA.
984 for (MVT VT : MVT::integer_vector_valuetypes()) {
985 setLoadExtAction(ISD::SEXTLOAD, VT, MVT::v2i8, Custom);
986 setLoadExtAction(ISD::SEXTLOAD, VT, MVT::v2i16, Custom);
987 setLoadExtAction(ISD::SEXTLOAD, VT, MVT::v2i32, Custom);
990 // SSE41 also has vector sign/zero extending loads, PMOV[SZ]X
991 setLoadExtAction(ISD::SEXTLOAD, MVT::v8i16, MVT::v8i8, Legal);
992 setLoadExtAction(ISD::SEXTLOAD, MVT::v4i32, MVT::v4i8, Legal);
993 setLoadExtAction(ISD::SEXTLOAD, MVT::v2i64, MVT::v2i8, Legal);
994 setLoadExtAction(ISD::SEXTLOAD, MVT::v4i32, MVT::v4i16, Legal);
995 setLoadExtAction(ISD::SEXTLOAD, MVT::v2i64, MVT::v2i16, Legal);
996 setLoadExtAction(ISD::SEXTLOAD, MVT::v2i64, MVT::v2i32, Legal);
998 setLoadExtAction(ISD::ZEXTLOAD, MVT::v8i16, MVT::v8i8, Legal);
999 setLoadExtAction(ISD::ZEXTLOAD, MVT::v4i32, MVT::v4i8, Legal);
1000 setLoadExtAction(ISD::ZEXTLOAD, MVT::v2i64, MVT::v2i8, Legal);
1001 setLoadExtAction(ISD::ZEXTLOAD, MVT::v4i32, MVT::v4i16, Legal);
1002 setLoadExtAction(ISD::ZEXTLOAD, MVT::v2i64, MVT::v2i16, Legal);
1003 setLoadExtAction(ISD::ZEXTLOAD, MVT::v2i64, MVT::v2i32, Legal);
1005 // i8 and i16 vectors are custom because the source register and source
1006 // source memory operand types are not the same width. f32 vectors are
1007 // custom since the immediate controlling the insert encodes additional
1009 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v16i8, Custom);
1010 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i16, Custom);
1011 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i32, Custom);
1012 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f32, Custom);
1014 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v16i8, Custom);
1015 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v8i16, Custom);
1016 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4i32, Custom);
1017 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f32, Custom);
1019 // FIXME: these should be Legal, but that's only for the case where
1020 // the index is constant. For now custom expand to deal with that.
1021 if (Subtarget->is64Bit()) {
1022 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2i64, Custom);
1023 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2i64, Custom);
1027 if (Subtarget->hasSSE2()) {
1028 setOperationAction(ISD::SIGN_EXTEND_VECTOR_INREG, MVT::v2i64, Custom);
1029 setOperationAction(ISD::SIGN_EXTEND_VECTOR_INREG, MVT::v4i32, Custom);
1030 setOperationAction(ISD::SIGN_EXTEND_VECTOR_INREG, MVT::v8i16, Custom);
1032 setOperationAction(ISD::SRL, MVT::v8i16, Custom);
1033 setOperationAction(ISD::SRL, MVT::v16i8, Custom);
1035 setOperationAction(ISD::SHL, MVT::v8i16, Custom);
1036 setOperationAction(ISD::SHL, MVT::v16i8, Custom);
1038 setOperationAction(ISD::SRA, MVT::v8i16, Custom);
1039 setOperationAction(ISD::SRA, MVT::v16i8, Custom);
1041 // In the customized shift lowering, the legal cases in AVX2 will be
1043 setOperationAction(ISD::SRL, MVT::v2i64, Custom);
1044 setOperationAction(ISD::SRL, MVT::v4i32, Custom);
1046 setOperationAction(ISD::SHL, MVT::v2i64, Custom);
1047 setOperationAction(ISD::SHL, MVT::v4i32, Custom);
1049 setOperationAction(ISD::SRA, MVT::v2i64, Custom);
1050 setOperationAction(ISD::SRA, MVT::v4i32, Custom);
1053 if (Subtarget->hasXOP()) {
1054 setOperationAction(ISD::ROTL, MVT::v16i8, Custom);
1055 setOperationAction(ISD::ROTL, MVT::v8i16, Custom);
1056 setOperationAction(ISD::ROTL, MVT::v4i32, Custom);
1057 setOperationAction(ISD::ROTL, MVT::v2i64, Custom);
1058 setOperationAction(ISD::ROTL, MVT::v32i8, Custom);
1059 setOperationAction(ISD::ROTL, MVT::v16i16, Custom);
1060 setOperationAction(ISD::ROTL, MVT::v8i32, Custom);
1061 setOperationAction(ISD::ROTL, MVT::v4i64, Custom);
1064 if (!Subtarget->useSoftFloat() && Subtarget->hasFp256()) {
1065 addRegisterClass(MVT::v32i8, &X86::VR256RegClass);
1066 addRegisterClass(MVT::v16i16, &X86::VR256RegClass);
1067 addRegisterClass(MVT::v8i32, &X86::VR256RegClass);
1068 addRegisterClass(MVT::v8f32, &X86::VR256RegClass);
1069 addRegisterClass(MVT::v4i64, &X86::VR256RegClass);
1070 addRegisterClass(MVT::v4f64, &X86::VR256RegClass);
1072 setOperationAction(ISD::LOAD, MVT::v8f32, Legal);
1073 setOperationAction(ISD::LOAD, MVT::v4f64, Legal);
1074 setOperationAction(ISD::LOAD, MVT::v4i64, Legal);
1076 setOperationAction(ISD::FADD, MVT::v8f32, Legal);
1077 setOperationAction(ISD::FSUB, MVT::v8f32, Legal);
1078 setOperationAction(ISD::FMUL, MVT::v8f32, Legal);
1079 setOperationAction(ISD::FDIV, MVT::v8f32, Legal);
1080 setOperationAction(ISD::FSQRT, MVT::v8f32, Legal);
1081 setOperationAction(ISD::FFLOOR, MVT::v8f32, Legal);
1082 setOperationAction(ISD::FCEIL, MVT::v8f32, Legal);
1083 setOperationAction(ISD::FTRUNC, MVT::v8f32, Legal);
1084 setOperationAction(ISD::FRINT, MVT::v8f32, Legal);
1085 setOperationAction(ISD::FNEARBYINT, MVT::v8f32, Legal);
1086 setOperationAction(ISD::FNEG, MVT::v8f32, Custom);
1087 setOperationAction(ISD::FABS, MVT::v8f32, Custom);
1089 setOperationAction(ISD::FADD, MVT::v4f64, Legal);
1090 setOperationAction(ISD::FSUB, MVT::v4f64, Legal);
1091 setOperationAction(ISD::FMUL, MVT::v4f64, Legal);
1092 setOperationAction(ISD::FDIV, MVT::v4f64, Legal);
1093 setOperationAction(ISD::FSQRT, MVT::v4f64, Legal);
1094 setOperationAction(ISD::FFLOOR, MVT::v4f64, Legal);
1095 setOperationAction(ISD::FCEIL, MVT::v4f64, Legal);
1096 setOperationAction(ISD::FTRUNC, MVT::v4f64, Legal);
1097 setOperationAction(ISD::FRINT, MVT::v4f64, Legal);
1098 setOperationAction(ISD::FNEARBYINT, MVT::v4f64, Legal);
1099 setOperationAction(ISD::FNEG, MVT::v4f64, Custom);
1100 setOperationAction(ISD::FABS, MVT::v4f64, Custom);
1102 // (fp_to_int:v8i16 (v8f32 ..)) requires the result type to be promoted
1103 // even though v8i16 is a legal type.
1104 setOperationAction(ISD::FP_TO_SINT, MVT::v8i16, Promote);
1105 setOperationAction(ISD::FP_TO_UINT, MVT::v8i16, Promote);
1106 setOperationAction(ISD::FP_TO_SINT, MVT::v8i32, Legal);
1108 setOperationAction(ISD::SINT_TO_FP, MVT::v8i16, Promote);
1109 setOperationAction(ISD::SINT_TO_FP, MVT::v8i32, Legal);
1110 setOperationAction(ISD::FP_ROUND, MVT::v4f32, Legal);
1112 setOperationAction(ISD::UINT_TO_FP, MVT::v8i8, Custom);
1113 setOperationAction(ISD::UINT_TO_FP, MVT::v8i16, Custom);
1115 for (MVT VT : MVT::fp_vector_valuetypes())
1116 setLoadExtAction(ISD::EXTLOAD, VT, MVT::v4f32, Legal);
1118 setOperationAction(ISD::SRL, MVT::v16i16, Custom);
1119 setOperationAction(ISD::SRL, MVT::v32i8, Custom);
1121 setOperationAction(ISD::SHL, MVT::v16i16, Custom);
1122 setOperationAction(ISD::SHL, MVT::v32i8, Custom);
1124 setOperationAction(ISD::SRA, MVT::v16i16, Custom);
1125 setOperationAction(ISD::SRA, MVT::v32i8, Custom);
1127 setOperationAction(ISD::SETCC, MVT::v32i8, Custom);
1128 setOperationAction(ISD::SETCC, MVT::v16i16, Custom);
1129 setOperationAction(ISD::SETCC, MVT::v8i32, Custom);
1130 setOperationAction(ISD::SETCC, MVT::v4i64, Custom);
1132 setOperationAction(ISD::SELECT, MVT::v4f64, Custom);
1133 setOperationAction(ISD::SELECT, MVT::v4i64, Custom);
1134 setOperationAction(ISD::SELECT, MVT::v8f32, Custom);
1136 setOperationAction(ISD::SIGN_EXTEND, MVT::v4i64, Custom);
1137 setOperationAction(ISD::SIGN_EXTEND, MVT::v8i32, Custom);
1138 setOperationAction(ISD::SIGN_EXTEND, MVT::v16i16, Custom);
1139 setOperationAction(ISD::ZERO_EXTEND, MVT::v4i64, Custom);
1140 setOperationAction(ISD::ZERO_EXTEND, MVT::v8i32, Custom);
1141 setOperationAction(ISD::ZERO_EXTEND, MVT::v16i16, Custom);
1142 setOperationAction(ISD::ANY_EXTEND, MVT::v4i64, Custom);
1143 setOperationAction(ISD::ANY_EXTEND, MVT::v8i32, Custom);
1144 setOperationAction(ISD::ANY_EXTEND, MVT::v16i16, Custom);
1145 setOperationAction(ISD::TRUNCATE, MVT::v16i8, Custom);
1146 setOperationAction(ISD::TRUNCATE, MVT::v8i16, Custom);
1147 setOperationAction(ISD::TRUNCATE, MVT::v4i32, Custom);
1149 setOperationAction(ISD::CTPOP, MVT::v32i8, Custom);
1150 setOperationAction(ISD::CTPOP, MVT::v16i16, Custom);
1151 setOperationAction(ISD::CTPOP, MVT::v8i32, Custom);
1152 setOperationAction(ISD::CTPOP, MVT::v4i64, Custom);
1154 setOperationAction(ISD::CTTZ, MVT::v32i8, Custom);
1155 setOperationAction(ISD::CTTZ, MVT::v16i16, Custom);
1156 setOperationAction(ISD::CTTZ, MVT::v8i32, Custom);
1157 setOperationAction(ISD::CTTZ, MVT::v4i64, Custom);
1158 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::v32i8, Custom);
1159 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::v16i16, Custom);
1160 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::v8i32, Custom);
1161 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::v4i64, Custom);
1163 if (Subtarget->hasFMA() || Subtarget->hasFMA4() || Subtarget->hasAVX512()) {
1164 setOperationAction(ISD::FMA, MVT::v8f32, Legal);
1165 setOperationAction(ISD::FMA, MVT::v4f64, Legal);
1166 setOperationAction(ISD::FMA, MVT::v4f32, Legal);
1167 setOperationAction(ISD::FMA, MVT::v2f64, Legal);
1168 setOperationAction(ISD::FMA, MVT::f32, Legal);
1169 setOperationAction(ISD::FMA, MVT::f64, Legal);
1172 if (Subtarget->hasInt256()) {
1173 setOperationAction(ISD::ADD, MVT::v4i64, Legal);
1174 setOperationAction(ISD::ADD, MVT::v8i32, Legal);
1175 setOperationAction(ISD::ADD, MVT::v16i16, Legal);
1176 setOperationAction(ISD::ADD, MVT::v32i8, Legal);
1178 setOperationAction(ISD::SUB, MVT::v4i64, Legal);
1179 setOperationAction(ISD::SUB, MVT::v8i32, Legal);
1180 setOperationAction(ISD::SUB, MVT::v16i16, Legal);
1181 setOperationAction(ISD::SUB, MVT::v32i8, Legal);
1183 setOperationAction(ISD::MUL, MVT::v4i64, Custom);
1184 setOperationAction(ISD::MUL, MVT::v8i32, Legal);
1185 setOperationAction(ISD::MUL, MVT::v16i16, Legal);
1186 setOperationAction(ISD::MUL, MVT::v32i8, Custom);
1188 setOperationAction(ISD::UMUL_LOHI, MVT::v8i32, Custom);
1189 setOperationAction(ISD::SMUL_LOHI, MVT::v8i32, Custom);
1190 setOperationAction(ISD::MULHU, MVT::v16i16, Legal);
1191 setOperationAction(ISD::MULHS, MVT::v16i16, Legal);
1193 setOperationAction(ISD::SMAX, MVT::v32i8, Legal);
1194 setOperationAction(ISD::SMAX, MVT::v16i16, Legal);
1195 setOperationAction(ISD::SMAX, MVT::v8i32, Legal);
1196 setOperationAction(ISD::UMAX, MVT::v32i8, Legal);
1197 setOperationAction(ISD::UMAX, MVT::v16i16, Legal);
1198 setOperationAction(ISD::UMAX, MVT::v8i32, Legal);
1199 setOperationAction(ISD::SMIN, MVT::v32i8, Legal);
1200 setOperationAction(ISD::SMIN, MVT::v16i16, Legal);
1201 setOperationAction(ISD::SMIN, MVT::v8i32, Legal);
1202 setOperationAction(ISD::UMIN, MVT::v32i8, Legal);
1203 setOperationAction(ISD::UMIN, MVT::v16i16, Legal);
1204 setOperationAction(ISD::UMIN, MVT::v8i32, Legal);
1206 // The custom lowering for UINT_TO_FP for v8i32 becomes interesting
1207 // when we have a 256bit-wide blend with immediate.
1208 setOperationAction(ISD::UINT_TO_FP, MVT::v8i32, Custom);
1210 // AVX2 also has wider vector sign/zero extending loads, VPMOV[SZ]X
1211 setLoadExtAction(ISD::SEXTLOAD, MVT::v16i16, MVT::v16i8, Legal);
1212 setLoadExtAction(ISD::SEXTLOAD, MVT::v8i32, MVT::v8i8, Legal);
1213 setLoadExtAction(ISD::SEXTLOAD, MVT::v4i64, MVT::v4i8, Legal);
1214 setLoadExtAction(ISD::SEXTLOAD, MVT::v8i32, MVT::v8i16, Legal);
1215 setLoadExtAction(ISD::SEXTLOAD, MVT::v4i64, MVT::v4i16, Legal);
1216 setLoadExtAction(ISD::SEXTLOAD, MVT::v4i64, MVT::v4i32, Legal);
1218 setLoadExtAction(ISD::ZEXTLOAD, MVT::v16i16, MVT::v16i8, Legal);
1219 setLoadExtAction(ISD::ZEXTLOAD, MVT::v8i32, MVT::v8i8, Legal);
1220 setLoadExtAction(ISD::ZEXTLOAD, MVT::v4i64, MVT::v4i8, Legal);
1221 setLoadExtAction(ISD::ZEXTLOAD, MVT::v8i32, MVT::v8i16, Legal);
1222 setLoadExtAction(ISD::ZEXTLOAD, MVT::v4i64, MVT::v4i16, Legal);
1223 setLoadExtAction(ISD::ZEXTLOAD, MVT::v4i64, MVT::v4i32, Legal);
1225 setOperationAction(ISD::ADD, MVT::v4i64, Custom);
1226 setOperationAction(ISD::ADD, MVT::v8i32, Custom);
1227 setOperationAction(ISD::ADD, MVT::v16i16, Custom);
1228 setOperationAction(ISD::ADD, MVT::v32i8, Custom);
1230 setOperationAction(ISD::SUB, MVT::v4i64, Custom);
1231 setOperationAction(ISD::SUB, MVT::v8i32, Custom);
1232 setOperationAction(ISD::SUB, MVT::v16i16, Custom);
1233 setOperationAction(ISD::SUB, MVT::v32i8, Custom);
1235 setOperationAction(ISD::MUL, MVT::v4i64, Custom);
1236 setOperationAction(ISD::MUL, MVT::v8i32, Custom);
1237 setOperationAction(ISD::MUL, MVT::v16i16, Custom);
1238 setOperationAction(ISD::MUL, MVT::v32i8, Custom);
1240 setOperationAction(ISD::SMAX, MVT::v32i8, Custom);
1241 setOperationAction(ISD::SMAX, MVT::v16i16, Custom);
1242 setOperationAction(ISD::SMAX, MVT::v8i32, Custom);
1243 setOperationAction(ISD::UMAX, MVT::v32i8, Custom);
1244 setOperationAction(ISD::UMAX, MVT::v16i16, Custom);
1245 setOperationAction(ISD::UMAX, MVT::v8i32, Custom);
1246 setOperationAction(ISD::SMIN, MVT::v32i8, Custom);
1247 setOperationAction(ISD::SMIN, MVT::v16i16, Custom);
1248 setOperationAction(ISD::SMIN, MVT::v8i32, Custom);
1249 setOperationAction(ISD::UMIN, MVT::v32i8, Custom);
1250 setOperationAction(ISD::UMIN, MVT::v16i16, Custom);
1251 setOperationAction(ISD::UMIN, MVT::v8i32, Custom);
1254 // In the customized shift lowering, the legal cases in AVX2 will be
1256 setOperationAction(ISD::SRL, MVT::v4i64, Custom);
1257 setOperationAction(ISD::SRL, MVT::v8i32, Custom);
1259 setOperationAction(ISD::SHL, MVT::v4i64, Custom);
1260 setOperationAction(ISD::SHL, MVT::v8i32, Custom);
1262 setOperationAction(ISD::SRA, MVT::v4i64, Custom);
1263 setOperationAction(ISD::SRA, MVT::v8i32, Custom);
1265 // Custom lower several nodes for 256-bit types.
1266 for (MVT VT : MVT::vector_valuetypes()) {
1267 if (VT.getScalarSizeInBits() >= 32) {
1268 setOperationAction(ISD::MLOAD, VT, Legal);
1269 setOperationAction(ISD::MSTORE, VT, Legal);
1271 // Extract subvector is special because the value type
1272 // (result) is 128-bit but the source is 256-bit wide.
1273 if (VT.is128BitVector()) {
1274 setOperationAction(ISD::EXTRACT_SUBVECTOR, VT, Custom);
1276 // Do not attempt to custom lower other non-256-bit vectors
1277 if (!VT.is256BitVector())
1280 setOperationAction(ISD::BUILD_VECTOR, VT, Custom);
1281 setOperationAction(ISD::VECTOR_SHUFFLE, VT, Custom);
1282 setOperationAction(ISD::VSELECT, VT, Custom);
1283 setOperationAction(ISD::INSERT_VECTOR_ELT, VT, Custom);
1284 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Custom);
1285 setOperationAction(ISD::SCALAR_TO_VECTOR, VT, Custom);
1286 setOperationAction(ISD::INSERT_SUBVECTOR, VT, Custom);
1287 setOperationAction(ISD::CONCAT_VECTORS, VT, Custom);
1290 if (Subtarget->hasInt256())
1291 setOperationAction(ISD::VSELECT, MVT::v32i8, Legal);
1293 // Promote v32i8, v16i16, v8i32 select, and, or, xor to v4i64.
1294 for (int i = MVT::v32i8; i != MVT::v4i64; ++i) {
1295 MVT VT = (MVT::SimpleValueType)i;
1297 // Do not attempt to promote non-256-bit vectors
1298 if (!VT.is256BitVector())
1301 setOperationAction(ISD::AND, VT, Promote);
1302 AddPromotedToType (ISD::AND, VT, MVT::v4i64);
1303 setOperationAction(ISD::OR, VT, Promote);
1304 AddPromotedToType (ISD::OR, VT, MVT::v4i64);
1305 setOperationAction(ISD::XOR, VT, Promote);
1306 AddPromotedToType (ISD::XOR, VT, MVT::v4i64);
1307 setOperationAction(ISD::LOAD, VT, Promote);
1308 AddPromotedToType (ISD::LOAD, VT, MVT::v4i64);
1309 setOperationAction(ISD::SELECT, VT, Promote);
1310 AddPromotedToType (ISD::SELECT, VT, MVT::v4i64);
1314 if (!Subtarget->useSoftFloat() && Subtarget->hasAVX512()) {
1315 addRegisterClass(MVT::v16i32, &X86::VR512RegClass);
1316 addRegisterClass(MVT::v16f32, &X86::VR512RegClass);
1317 addRegisterClass(MVT::v8i64, &X86::VR512RegClass);
1318 addRegisterClass(MVT::v8f64, &X86::VR512RegClass);
1320 addRegisterClass(MVT::i1, &X86::VK1RegClass);
1321 addRegisterClass(MVT::v8i1, &X86::VK8RegClass);
1322 addRegisterClass(MVT::v16i1, &X86::VK16RegClass);
1324 for (MVT VT : MVT::fp_vector_valuetypes())
1325 setLoadExtAction(ISD::EXTLOAD, VT, MVT::v8f32, Legal);
1327 setLoadExtAction(ISD::ZEXTLOAD, MVT::v16i32, MVT::v16i8, Legal);
1328 setLoadExtAction(ISD::SEXTLOAD, MVT::v16i32, MVT::v16i8, Legal);
1329 setLoadExtAction(ISD::ZEXTLOAD, MVT::v16i32, MVT::v16i16, Legal);
1330 setLoadExtAction(ISD::SEXTLOAD, MVT::v16i32, MVT::v16i16, Legal);
1331 setLoadExtAction(ISD::ZEXTLOAD, MVT::v32i16, MVT::v32i8, Legal);
1332 setLoadExtAction(ISD::SEXTLOAD, MVT::v32i16, MVT::v32i8, Legal);
1333 setLoadExtAction(ISD::ZEXTLOAD, MVT::v8i64, MVT::v8i8, Legal);
1334 setLoadExtAction(ISD::SEXTLOAD, MVT::v8i64, MVT::v8i8, Legal);
1335 setLoadExtAction(ISD::ZEXTLOAD, MVT::v8i64, MVT::v8i16, Legal);
1336 setLoadExtAction(ISD::SEXTLOAD, MVT::v8i64, MVT::v8i16, Legal);
1337 setLoadExtAction(ISD::ZEXTLOAD, MVT::v8i64, MVT::v8i32, Legal);
1338 setLoadExtAction(ISD::SEXTLOAD, MVT::v8i64, MVT::v8i32, Legal);
1340 setOperationAction(ISD::BR_CC, MVT::i1, Expand);
1341 setOperationAction(ISD::SETCC, MVT::i1, Custom);
1342 setOperationAction(ISD::SELECT_CC, MVT::i1, Expand);
1343 setOperationAction(ISD::XOR, MVT::i1, Legal);
1344 setOperationAction(ISD::OR, MVT::i1, Legal);
1345 setOperationAction(ISD::AND, MVT::i1, Legal);
1346 setOperationAction(ISD::SUB, MVT::i1, Custom);
1347 setOperationAction(ISD::ADD, MVT::i1, Custom);
1348 setOperationAction(ISD::MUL, MVT::i1, Custom);
1349 setOperationAction(ISD::LOAD, MVT::v16f32, Legal);
1350 setOperationAction(ISD::LOAD, MVT::v8f64, Legal);
1351 setOperationAction(ISD::LOAD, MVT::v8i64, Legal);
1352 setOperationAction(ISD::LOAD, MVT::v16i32, Legal);
1353 setOperationAction(ISD::LOAD, MVT::v16i1, Legal);
1355 setOperationAction(ISD::FADD, MVT::v16f32, Legal);
1356 setOperationAction(ISD::FSUB, MVT::v16f32, Legal);
1357 setOperationAction(ISD::FMUL, MVT::v16f32, Legal);
1358 setOperationAction(ISD::FDIV, MVT::v16f32, Legal);
1359 setOperationAction(ISD::FSQRT, MVT::v16f32, Legal);
1360 setOperationAction(ISD::FNEG, MVT::v16f32, Custom);
1362 setOperationAction(ISD::FADD, MVT::v8f64, Legal);
1363 setOperationAction(ISD::FSUB, MVT::v8f64, Legal);
1364 setOperationAction(ISD::FMUL, MVT::v8f64, Legal);
1365 setOperationAction(ISD::FDIV, MVT::v8f64, Legal);
1366 setOperationAction(ISD::FSQRT, MVT::v8f64, Legal);
1367 setOperationAction(ISD::FNEG, MVT::v8f64, Custom);
1368 setOperationAction(ISD::FMA, MVT::v8f64, Legal);
1369 setOperationAction(ISD::FMA, MVT::v16f32, Legal);
1371 setOperationAction(ISD::FP_TO_SINT, MVT::v16i32, Legal);
1372 setOperationAction(ISD::FP_TO_UINT, MVT::v16i32, Legal);
1373 setOperationAction(ISD::FP_TO_UINT, MVT::v8i32, Legal);
1374 setOperationAction(ISD::FP_TO_UINT, MVT::v4i32, Legal);
1375 setOperationAction(ISD::SINT_TO_FP, MVT::v16i32, Legal);
1376 setOperationAction(ISD::SINT_TO_FP, MVT::v8i1, Custom);
1377 setOperationAction(ISD::SINT_TO_FP, MVT::v16i1, Custom);
1378 setOperationAction(ISD::SINT_TO_FP, MVT::v16i8, Promote);
1379 setOperationAction(ISD::SINT_TO_FP, MVT::v16i16, Promote);
1380 setOperationAction(ISD::UINT_TO_FP, MVT::v16i32, Legal);
1381 setOperationAction(ISD::UINT_TO_FP, MVT::v8i32, Legal);
1382 setOperationAction(ISD::UINT_TO_FP, MVT::v4i32, Legal);
1383 setOperationAction(ISD::UINT_TO_FP, MVT::v16i8, Custom);
1384 setOperationAction(ISD::UINT_TO_FP, MVT::v16i16, Custom);
1385 setOperationAction(ISD::FP_ROUND, MVT::v8f32, Legal);
1386 setOperationAction(ISD::FP_EXTEND, MVT::v8f32, Legal);
1388 setTruncStoreAction(MVT::v8i64, MVT::v8i8, Legal);
1389 setTruncStoreAction(MVT::v8i64, MVT::v8i16, Legal);
1390 setTruncStoreAction(MVT::v8i64, MVT::v8i32, Legal);
1391 setTruncStoreAction(MVT::v16i32, MVT::v16i8, Legal);
1392 setTruncStoreAction(MVT::v16i32, MVT::v16i16, Legal);
1393 if (Subtarget->hasVLX()){
1394 setTruncStoreAction(MVT::v4i64, MVT::v4i8, Legal);
1395 setTruncStoreAction(MVT::v4i64, MVT::v4i16, Legal);
1396 setTruncStoreAction(MVT::v4i64, MVT::v4i32, Legal);
1397 setTruncStoreAction(MVT::v8i32, MVT::v8i8, Legal);
1398 setTruncStoreAction(MVT::v8i32, MVT::v8i16, Legal);
1400 setTruncStoreAction(MVT::v2i64, MVT::v2i8, Legal);
1401 setTruncStoreAction(MVT::v2i64, MVT::v2i16, Legal);
1402 setTruncStoreAction(MVT::v2i64, MVT::v2i32, Legal);
1403 setTruncStoreAction(MVT::v4i32, MVT::v4i8, Legal);
1404 setTruncStoreAction(MVT::v4i32, MVT::v4i16, Legal);
1406 setOperationAction(ISD::TRUNCATE, MVT::i1, Custom);
1407 setOperationAction(ISD::TRUNCATE, MVT::v16i8, Custom);
1408 setOperationAction(ISD::TRUNCATE, MVT::v8i32, Custom);
1409 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v8i1, Custom);
1410 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v16i1, Custom);
1411 if (Subtarget->hasDQI()) {
1412 setOperationAction(ISD::TRUNCATE, MVT::v2i1, Custom);
1413 setOperationAction(ISD::TRUNCATE, MVT::v4i1, Custom);
1415 setOperationAction(ISD::SINT_TO_FP, MVT::v8i64, Legal);
1416 setOperationAction(ISD::UINT_TO_FP, MVT::v8i64, Legal);
1417 setOperationAction(ISD::FP_TO_SINT, MVT::v8i64, Legal);
1418 setOperationAction(ISD::FP_TO_UINT, MVT::v8i64, Legal);
1419 if (Subtarget->hasVLX()) {
1420 setOperationAction(ISD::SINT_TO_FP, MVT::v4i64, Legal);
1421 setOperationAction(ISD::SINT_TO_FP, MVT::v2i64, Legal);
1422 setOperationAction(ISD::UINT_TO_FP, MVT::v4i64, Legal);
1423 setOperationAction(ISD::UINT_TO_FP, MVT::v2i64, Legal);
1424 setOperationAction(ISD::FP_TO_SINT, MVT::v4i64, Legal);
1425 setOperationAction(ISD::FP_TO_SINT, MVT::v2i64, Legal);
1426 setOperationAction(ISD::FP_TO_UINT, MVT::v4i64, Legal);
1427 setOperationAction(ISD::FP_TO_UINT, MVT::v2i64, Legal);
1430 if (Subtarget->hasVLX()) {
1431 setOperationAction(ISD::SINT_TO_FP, MVT::v8i32, Legal);
1432 setOperationAction(ISD::UINT_TO_FP, MVT::v8i32, Legal);
1433 setOperationAction(ISD::FP_TO_SINT, MVT::v8i32, Legal);
1434 setOperationAction(ISD::FP_TO_UINT, MVT::v8i32, Legal);
1435 setOperationAction(ISD::SINT_TO_FP, MVT::v4i32, Legal);
1436 setOperationAction(ISD::UINT_TO_FP, MVT::v4i32, Legal);
1437 setOperationAction(ISD::FP_TO_SINT, MVT::v4i32, Legal);
1438 setOperationAction(ISD::FP_TO_UINT, MVT::v4i32, Legal);
1440 setOperationAction(ISD::TRUNCATE, MVT::v8i1, Custom);
1441 setOperationAction(ISD::TRUNCATE, MVT::v16i1, Custom);
1442 setOperationAction(ISD::TRUNCATE, MVT::v16i16, Custom);
1443 setOperationAction(ISD::ZERO_EXTEND, MVT::v16i32, Custom);
1444 setOperationAction(ISD::ZERO_EXTEND, MVT::v8i64, Custom);
1445 setOperationAction(ISD::ANY_EXTEND, MVT::v16i32, Custom);
1446 setOperationAction(ISD::ANY_EXTEND, MVT::v8i64, Custom);
1447 setOperationAction(ISD::SIGN_EXTEND, MVT::v16i32, Custom);
1448 setOperationAction(ISD::SIGN_EXTEND, MVT::v8i64, Custom);
1449 setOperationAction(ISD::SIGN_EXTEND, MVT::v16i8, Custom);
1450 setOperationAction(ISD::SIGN_EXTEND, MVT::v8i16, Custom);
1451 setOperationAction(ISD::SIGN_EXTEND, MVT::v16i16, Custom);
1452 if (Subtarget->hasDQI()) {
1453 setOperationAction(ISD::SIGN_EXTEND, MVT::v4i32, Custom);
1454 setOperationAction(ISD::SIGN_EXTEND, MVT::v2i64, Custom);
1456 setOperationAction(ISD::FFLOOR, MVT::v16f32, Legal);
1457 setOperationAction(ISD::FFLOOR, MVT::v8f64, Legal);
1458 setOperationAction(ISD::FCEIL, MVT::v16f32, Legal);
1459 setOperationAction(ISD::FCEIL, MVT::v8f64, Legal);
1460 setOperationAction(ISD::FTRUNC, MVT::v16f32, Legal);
1461 setOperationAction(ISD::FTRUNC, MVT::v8f64, Legal);
1462 setOperationAction(ISD::FRINT, MVT::v16f32, Legal);
1463 setOperationAction(ISD::FRINT, MVT::v8f64, Legal);
1464 setOperationAction(ISD::FNEARBYINT, MVT::v16f32, Legal);
1465 setOperationAction(ISD::FNEARBYINT, MVT::v8f64, Legal);
1467 setOperationAction(ISD::CONCAT_VECTORS, MVT::v8f64, Custom);
1468 setOperationAction(ISD::CONCAT_VECTORS, MVT::v8i64, Custom);
1469 setOperationAction(ISD::CONCAT_VECTORS, MVT::v16f32, Custom);
1470 setOperationAction(ISD::CONCAT_VECTORS, MVT::v16i32, Custom);
1471 setOperationAction(ISD::CONCAT_VECTORS, MVT::v16i1, Legal);
1473 setOperationAction(ISD::SETCC, MVT::v16i1, Custom);
1474 setOperationAction(ISD::SETCC, MVT::v8i1, Custom);
1476 setOperationAction(ISD::MUL, MVT::v8i64, Custom);
1478 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v8i1, Custom);
1479 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v16i1, Custom);
1480 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v16i1, Custom);
1481 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i1, Custom);
1482 setOperationAction(ISD::BUILD_VECTOR, MVT::v8i1, Custom);
1483 setOperationAction(ISD::BUILD_VECTOR, MVT::v16i1, Custom);
1484 setOperationAction(ISD::SELECT, MVT::v8f64, Custom);
1485 setOperationAction(ISD::SELECT, MVT::v8i64, Custom);
1486 setOperationAction(ISD::SELECT, MVT::v16f32, Custom);
1487 setOperationAction(ISD::SELECT, MVT::v16i1, Custom);
1488 setOperationAction(ISD::SELECT, MVT::v8i1, Custom);
1490 setOperationAction(ISD::SMAX, MVT::v16i32, Legal);
1491 setOperationAction(ISD::SMAX, MVT::v8i64, Legal);
1492 setOperationAction(ISD::UMAX, MVT::v16i32, Legal);
1493 setOperationAction(ISD::UMAX, MVT::v8i64, Legal);
1494 setOperationAction(ISD::SMIN, MVT::v16i32, Legal);
1495 setOperationAction(ISD::SMIN, MVT::v8i64, Legal);
1496 setOperationAction(ISD::UMIN, MVT::v16i32, Legal);
1497 setOperationAction(ISD::UMIN, MVT::v8i64, Legal);
1499 setOperationAction(ISD::ADD, MVT::v8i64, Legal);
1500 setOperationAction(ISD::ADD, MVT::v16i32, Legal);
1502 setOperationAction(ISD::SUB, MVT::v8i64, Legal);
1503 setOperationAction(ISD::SUB, MVT::v16i32, Legal);
1505 setOperationAction(ISD::MUL, MVT::v16i32, Legal);
1507 setOperationAction(ISD::SRL, MVT::v8i64, Custom);
1508 setOperationAction(ISD::SRL, MVT::v16i32, Custom);
1510 setOperationAction(ISD::SHL, MVT::v8i64, Custom);
1511 setOperationAction(ISD::SHL, MVT::v16i32, Custom);
1513 setOperationAction(ISD::SRA, MVT::v8i64, Custom);
1514 setOperationAction(ISD::SRA, MVT::v16i32, Custom);
1516 setOperationAction(ISD::AND, MVT::v8i64, Legal);
1517 setOperationAction(ISD::OR, MVT::v8i64, Legal);
1518 setOperationAction(ISD::XOR, MVT::v8i64, Legal);
1519 setOperationAction(ISD::AND, MVT::v16i32, Legal);
1520 setOperationAction(ISD::OR, MVT::v16i32, Legal);
1521 setOperationAction(ISD::XOR, MVT::v16i32, Legal);
1523 if (Subtarget->hasCDI()) {
1524 setOperationAction(ISD::CTLZ, MVT::v8i64, Legal);
1525 setOperationAction(ISD::CTLZ, MVT::v16i32, Legal);
1526 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::v8i64, Legal);
1527 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::v16i32, Legal);
1529 setOperationAction(ISD::CTLZ, MVT::v8i16, Custom);
1530 setOperationAction(ISD::CTLZ, MVT::v16i8, Custom);
1531 setOperationAction(ISD::CTLZ, MVT::v16i16, Custom);
1532 setOperationAction(ISD::CTLZ, MVT::v32i8, Custom);
1533 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::v8i16, Custom);
1534 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::v16i8, Custom);
1535 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::v16i16, Custom);
1536 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::v32i8, Custom);
1538 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::v8i64, Custom);
1539 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::v16i32, Custom);
1541 if (Subtarget->hasVLX()) {
1542 setOperationAction(ISD::CTLZ, MVT::v4i64, Legal);
1543 setOperationAction(ISD::CTLZ, MVT::v8i32, Legal);
1544 setOperationAction(ISD::CTLZ, MVT::v2i64, Legal);
1545 setOperationAction(ISD::CTLZ, MVT::v4i32, Legal);
1546 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::v4i64, Legal);
1547 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::v8i32, Legal);
1548 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::v2i64, Legal);
1549 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::v4i32, Legal);
1551 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::v4i64, Custom);
1552 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::v8i32, Custom);
1553 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::v2i64, Custom);
1554 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::v4i32, Custom);
1556 setOperationAction(ISD::CTLZ, MVT::v4i64, Custom);
1557 setOperationAction(ISD::CTLZ, MVT::v8i32, Custom);
1558 setOperationAction(ISD::CTLZ, MVT::v2i64, Custom);
1559 setOperationAction(ISD::CTLZ, MVT::v4i32, Custom);
1560 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::v4i64, Custom);
1561 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::v8i32, Custom);
1562 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::v2i64, Custom);
1563 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::v4i32, Custom);
1565 } // Subtarget->hasCDI()
1567 if (Subtarget->hasDQI()) {
1568 setOperationAction(ISD::MUL, MVT::v2i64, Legal);
1569 setOperationAction(ISD::MUL, MVT::v4i64, Legal);
1570 setOperationAction(ISD::MUL, MVT::v8i64, Legal);
1572 // Custom lower several nodes.
1573 for (MVT VT : MVT::vector_valuetypes()) {
1574 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
1576 setOperationAction(ISD::AND, VT, Legal);
1577 setOperationAction(ISD::OR, VT, Legal);
1578 setOperationAction(ISD::XOR, VT, Legal);
1580 if (EltSize >= 32 && VT.getSizeInBits() <= 512) {
1581 setOperationAction(ISD::MGATHER, VT, Custom);
1582 setOperationAction(ISD::MSCATTER, VT, Custom);
1584 // Extract subvector is special because the value type
1585 // (result) is 256/128-bit but the source is 512-bit wide.
1586 if (VT.is128BitVector() || VT.is256BitVector()) {
1587 setOperationAction(ISD::EXTRACT_SUBVECTOR, VT, Custom);
1589 if (VT.getVectorElementType() == MVT::i1)
1590 setOperationAction(ISD::EXTRACT_SUBVECTOR, VT, Legal);
1592 // Do not attempt to custom lower other non-512-bit vectors
1593 if (!VT.is512BitVector())
1596 if (EltSize >= 32) {
1597 setOperationAction(ISD::VECTOR_SHUFFLE, VT, Custom);
1598 setOperationAction(ISD::INSERT_VECTOR_ELT, VT, Custom);
1599 setOperationAction(ISD::BUILD_VECTOR, VT, Custom);
1600 setOperationAction(ISD::VSELECT, VT, Legal);
1601 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Custom);
1602 setOperationAction(ISD::SCALAR_TO_VECTOR, VT, Custom);
1603 setOperationAction(ISD::INSERT_SUBVECTOR, VT, Custom);
1604 setOperationAction(ISD::MLOAD, VT, Legal);
1605 setOperationAction(ISD::MSTORE, VT, Legal);
1608 for (int i = MVT::v32i8; i != MVT::v8i64; ++i) {
1609 MVT VT = (MVT::SimpleValueType)i;
1611 // Do not attempt to promote non-512-bit vectors.
1612 if (!VT.is512BitVector())
1615 setOperationAction(ISD::SELECT, VT, Promote);
1616 AddPromotedToType (ISD::SELECT, VT, MVT::v8i64);
1620 if (!Subtarget->useSoftFloat() && Subtarget->hasBWI()) {
1621 addRegisterClass(MVT::v32i16, &X86::VR512RegClass);
1622 addRegisterClass(MVT::v64i8, &X86::VR512RegClass);
1624 addRegisterClass(MVT::v32i1, &X86::VK32RegClass);
1625 addRegisterClass(MVT::v64i1, &X86::VK64RegClass);
1627 setOperationAction(ISD::LOAD, MVT::v32i16, Legal);
1628 setOperationAction(ISD::LOAD, MVT::v64i8, Legal);
1629 setOperationAction(ISD::SETCC, MVT::v32i1, Custom);
1630 setOperationAction(ISD::SETCC, MVT::v64i1, Custom);
1631 setOperationAction(ISD::ADD, MVT::v32i16, Legal);
1632 setOperationAction(ISD::ADD, MVT::v64i8, Legal);
1633 setOperationAction(ISD::SUB, MVT::v32i16, Legal);
1634 setOperationAction(ISD::SUB, MVT::v64i8, Legal);
1635 setOperationAction(ISD::MUL, MVT::v32i16, Legal);
1636 setOperationAction(ISD::MULHS, MVT::v32i16, Legal);
1637 setOperationAction(ISD::MULHU, MVT::v32i16, Legal);
1638 setOperationAction(ISD::CONCAT_VECTORS, MVT::v32i1, Legal);
1639 setOperationAction(ISD::CONCAT_VECTORS, MVT::v64i1, Legal);
1640 setOperationAction(ISD::CONCAT_VECTORS, MVT::v32i16, Custom);
1641 setOperationAction(ISD::CONCAT_VECTORS, MVT::v64i8, Custom);
1642 setOperationAction(ISD::INSERT_SUBVECTOR, MVT::v32i1, Custom);
1643 setOperationAction(ISD::INSERT_SUBVECTOR, MVT::v64i1, Custom);
1644 setOperationAction(ISD::INSERT_SUBVECTOR, MVT::v32i16, Custom);
1645 setOperationAction(ISD::INSERT_SUBVECTOR, MVT::v64i8, Custom);
1646 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v32i16, Custom);
1647 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v64i8, Custom);
1648 setOperationAction(ISD::SELECT, MVT::v32i1, Custom);
1649 setOperationAction(ISD::SELECT, MVT::v64i1, Custom);
1650 setOperationAction(ISD::SIGN_EXTEND, MVT::v32i8, Custom);
1651 setOperationAction(ISD::ZERO_EXTEND, MVT::v32i8, Custom);
1652 setOperationAction(ISD::SIGN_EXTEND, MVT::v32i16, Custom);
1653 setOperationAction(ISD::ZERO_EXTEND, MVT::v32i16, Custom);
1654 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v32i16, Custom);
1655 setOperationAction(ISD::SIGN_EXTEND, MVT::v64i8, Custom);
1656 setOperationAction(ISD::ZERO_EXTEND, MVT::v64i8, Custom);
1657 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v32i1, Custom);
1658 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v64i1, Custom);
1659 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v32i16, Custom);
1660 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v64i8, Custom);
1661 setOperationAction(ISD::VSELECT, MVT::v32i16, Legal);
1662 setOperationAction(ISD::VSELECT, MVT::v64i8, Legal);
1663 setOperationAction(ISD::TRUNCATE, MVT::v32i1, Custom);
1664 setOperationAction(ISD::TRUNCATE, MVT::v64i1, Custom);
1665 setOperationAction(ISD::TRUNCATE, MVT::v32i8, Custom);
1666 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v32i1, Custom);
1667 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v64i1, Custom);
1669 setOperationAction(ISD::SMAX, MVT::v64i8, Legal);
1670 setOperationAction(ISD::SMAX, MVT::v32i16, Legal);
1671 setOperationAction(ISD::UMAX, MVT::v64i8, Legal);
1672 setOperationAction(ISD::UMAX, MVT::v32i16, Legal);
1673 setOperationAction(ISD::SMIN, MVT::v64i8, Legal);
1674 setOperationAction(ISD::SMIN, MVT::v32i16, Legal);
1675 setOperationAction(ISD::UMIN, MVT::v64i8, Legal);
1676 setOperationAction(ISD::UMIN, MVT::v32i16, Legal);
1678 setTruncStoreAction(MVT::v32i16, MVT::v32i8, Legal);
1679 setTruncStoreAction(MVT::v16i16, MVT::v16i8, Legal);
1680 if (Subtarget->hasVLX())
1681 setTruncStoreAction(MVT::v8i16, MVT::v8i8, Legal);
1683 if (Subtarget->hasCDI()) {
1684 setOperationAction(ISD::CTLZ, MVT::v32i16, Custom);
1685 setOperationAction(ISD::CTLZ, MVT::v64i8, Custom);
1686 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::v32i16, Custom);
1687 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::v64i8, Custom);
1690 for (int i = MVT::v32i8; i != MVT::v8i64; ++i) {
1691 const MVT VT = (MVT::SimpleValueType)i;
1693 const unsigned EltSize = VT.getVectorElementType().getSizeInBits();
1695 // Do not attempt to promote non-512-bit vectors.
1696 if (!VT.is512BitVector())
1700 setOperationAction(ISD::BUILD_VECTOR, VT, Custom);
1701 setOperationAction(ISD::VSELECT, VT, Legal);
1706 if (!Subtarget->useSoftFloat() && Subtarget->hasVLX()) {
1707 addRegisterClass(MVT::v4i1, &X86::VK4RegClass);
1708 addRegisterClass(MVT::v2i1, &X86::VK2RegClass);
1710 setOperationAction(ISD::SETCC, MVT::v4i1, Custom);
1711 setOperationAction(ISD::SETCC, MVT::v2i1, Custom);
1712 setOperationAction(ISD::CONCAT_VECTORS, MVT::v4i1, Custom);
1713 setOperationAction(ISD::CONCAT_VECTORS, MVT::v8i1, Custom);
1714 setOperationAction(ISD::INSERT_SUBVECTOR, MVT::v8i1, Custom);
1715 setOperationAction(ISD::INSERT_SUBVECTOR, MVT::v4i1, Custom);
1716 setOperationAction(ISD::SELECT, MVT::v4i1, Custom);
1717 setOperationAction(ISD::SELECT, MVT::v2i1, Custom);
1718 setOperationAction(ISD::BUILD_VECTOR, MVT::v4i1, Custom);
1719 setOperationAction(ISD::BUILD_VECTOR, MVT::v2i1, Custom);
1720 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2i1, Custom);
1721 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4i1, Custom);
1723 setOperationAction(ISD::AND, MVT::v8i32, Legal);
1724 setOperationAction(ISD::OR, MVT::v8i32, Legal);
1725 setOperationAction(ISD::XOR, MVT::v8i32, Legal);
1726 setOperationAction(ISD::AND, MVT::v4i32, Legal);
1727 setOperationAction(ISD::OR, MVT::v4i32, Legal);
1728 setOperationAction(ISD::XOR, MVT::v4i32, Legal);
1729 setOperationAction(ISD::SRA, MVT::v2i64, Custom);
1730 setOperationAction(ISD::SRA, MVT::v4i64, Custom);
1732 setOperationAction(ISD::SMAX, MVT::v2i64, Legal);
1733 setOperationAction(ISD::SMAX, MVT::v4i64, Legal);
1734 setOperationAction(ISD::UMAX, MVT::v2i64, Legal);
1735 setOperationAction(ISD::UMAX, MVT::v4i64, Legal);
1736 setOperationAction(ISD::SMIN, MVT::v2i64, Legal);
1737 setOperationAction(ISD::SMIN, MVT::v4i64, Legal);
1738 setOperationAction(ISD::UMIN, MVT::v2i64, Legal);
1739 setOperationAction(ISD::UMIN, MVT::v4i64, Legal);
1742 // We want to custom lower some of our intrinsics.
1743 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
1744 setOperationAction(ISD::INTRINSIC_W_CHAIN, MVT::Other, Custom);
1745 setOperationAction(ISD::INTRINSIC_VOID, MVT::Other, Custom);
1746 if (!Subtarget->is64Bit())
1747 setOperationAction(ISD::INTRINSIC_W_CHAIN, MVT::i64, Custom);
1749 // Only custom-lower 64-bit SADDO and friends on 64-bit because we don't
1750 // handle type legalization for these operations here.
1752 // FIXME: We really should do custom legalization for addition and
1753 // subtraction on x86-32 once PR3203 is fixed. We really can't do much better
1754 // than generic legalization for 64-bit multiplication-with-overflow, though.
1755 for (unsigned i = 0, e = 3+Subtarget->is64Bit(); i != e; ++i) {
1756 // Add/Sub/Mul with overflow operations are custom lowered.
1758 setOperationAction(ISD::SADDO, VT, Custom);
1759 setOperationAction(ISD::UADDO, VT, Custom);
1760 setOperationAction(ISD::SSUBO, VT, Custom);
1761 setOperationAction(ISD::USUBO, VT, Custom);
1762 setOperationAction(ISD::SMULO, VT, Custom);
1763 setOperationAction(ISD::UMULO, VT, Custom);
1766 if (!Subtarget->is64Bit()) {
1767 // These libcalls are not available in 32-bit.
1768 setLibcallName(RTLIB::SHL_I128, nullptr);
1769 setLibcallName(RTLIB::SRL_I128, nullptr);
1770 setLibcallName(RTLIB::SRA_I128, nullptr);
1773 // Combine sin / cos into one node or libcall if possible.
1774 if (Subtarget->hasSinCos()) {
1775 setLibcallName(RTLIB::SINCOS_F32, "sincosf");
1776 setLibcallName(RTLIB::SINCOS_F64, "sincos");
1777 if (Subtarget->isTargetDarwin()) {
1778 // For MacOSX, we don't want the normal expansion of a libcall to sincos.
1779 // We want to issue a libcall to __sincos_stret to avoid memory traffic.
1780 setOperationAction(ISD::FSINCOS, MVT::f64, Custom);
1781 setOperationAction(ISD::FSINCOS, MVT::f32, Custom);
1785 if (Subtarget->isTargetWin64()) {
1786 setOperationAction(ISD::SDIV, MVT::i128, Custom);
1787 setOperationAction(ISD::UDIV, MVT::i128, Custom);
1788 setOperationAction(ISD::SREM, MVT::i128, Custom);
1789 setOperationAction(ISD::UREM, MVT::i128, Custom);
1790 setOperationAction(ISD::SDIVREM, MVT::i128, Custom);
1791 setOperationAction(ISD::UDIVREM, MVT::i128, Custom);
1794 // We have target-specific dag combine patterns for the following nodes:
1795 setTargetDAGCombine(ISD::VECTOR_SHUFFLE);
1796 setTargetDAGCombine(ISD::EXTRACT_VECTOR_ELT);
1797 setTargetDAGCombine(ISD::BITCAST);
1798 setTargetDAGCombine(ISD::VSELECT);
1799 setTargetDAGCombine(ISD::SELECT);
1800 setTargetDAGCombine(ISD::SHL);
1801 setTargetDAGCombine(ISD::SRA);
1802 setTargetDAGCombine(ISD::SRL);
1803 setTargetDAGCombine(ISD::OR);
1804 setTargetDAGCombine(ISD::AND);
1805 setTargetDAGCombine(ISD::ADD);
1806 setTargetDAGCombine(ISD::FADD);
1807 setTargetDAGCombine(ISD::FSUB);
1808 setTargetDAGCombine(ISD::FMA);
1809 setTargetDAGCombine(ISD::SUB);
1810 setTargetDAGCombine(ISD::LOAD);
1811 setTargetDAGCombine(ISD::MLOAD);
1812 setTargetDAGCombine(ISD::STORE);
1813 setTargetDAGCombine(ISD::MSTORE);
1814 setTargetDAGCombine(ISD::ZERO_EXTEND);
1815 setTargetDAGCombine(ISD::ANY_EXTEND);
1816 setTargetDAGCombine(ISD::SIGN_EXTEND);
1817 setTargetDAGCombine(ISD::SIGN_EXTEND_INREG);
1818 setTargetDAGCombine(ISD::SINT_TO_FP);
1819 setTargetDAGCombine(ISD::UINT_TO_FP);
1820 setTargetDAGCombine(ISD::SETCC);
1821 setTargetDAGCombine(ISD::BUILD_VECTOR);
1822 setTargetDAGCombine(ISD::MUL);
1823 setTargetDAGCombine(ISD::XOR);
1825 computeRegisterProperties(Subtarget->getRegisterInfo());
1827 MaxStoresPerMemset = 16; // For @llvm.memset -> sequence of stores
1828 MaxStoresPerMemsetOptSize = 8;
1829 MaxStoresPerMemcpy = 8; // For @llvm.memcpy -> sequence of stores
1830 MaxStoresPerMemcpyOptSize = 4;
1831 MaxStoresPerMemmove = 8; // For @llvm.memmove -> sequence of stores
1832 MaxStoresPerMemmoveOptSize = 4;
1833 setPrefLoopAlignment(4); // 2^4 bytes.
1835 // A predictable cmov does not hurt on an in-order CPU.
1836 // FIXME: Use a CPU attribute to trigger this, not a CPU model.
1837 PredictableSelectIsExpensive = !Subtarget->isAtom();
1838 EnableExtLdPromotion = true;
1839 setPrefFunctionAlignment(4); // 2^4 bytes.
1841 verifyIntrinsicTables();
1844 // This has so far only been implemented for 64-bit MachO.
1845 bool X86TargetLowering::useLoadStackGuardNode() const {
1846 return Subtarget->isTargetMachO() && Subtarget->is64Bit();
1849 TargetLoweringBase::LegalizeTypeAction
1850 X86TargetLowering::getPreferredVectorAction(EVT VT) const {
1851 if (ExperimentalVectorWideningLegalization &&
1852 VT.getVectorNumElements() != 1 &&
1853 VT.getVectorElementType().getSimpleVT() != MVT::i1)
1854 return TypeWidenVector;
1856 return TargetLoweringBase::getPreferredVectorAction(VT);
1859 EVT X86TargetLowering::getSetCCResultType(const DataLayout &DL, LLVMContext &,
1862 return Subtarget->hasAVX512() ? MVT::i1: MVT::i8;
1864 const unsigned NumElts = VT.getVectorNumElements();
1865 const EVT EltVT = VT.getVectorElementType();
1866 if (VT.is512BitVector()) {
1867 if (Subtarget->hasAVX512())
1868 if (EltVT == MVT::i32 || EltVT == MVT::i64 ||
1869 EltVT == MVT::f32 || EltVT == MVT::f64)
1871 case 8: return MVT::v8i1;
1872 case 16: return MVT::v16i1;
1874 if (Subtarget->hasBWI())
1875 if (EltVT == MVT::i8 || EltVT == MVT::i16)
1877 case 32: return MVT::v32i1;
1878 case 64: return MVT::v64i1;
1882 if (VT.is256BitVector() || VT.is128BitVector()) {
1883 if (Subtarget->hasVLX())
1884 if (EltVT == MVT::i32 || EltVT == MVT::i64 ||
1885 EltVT == MVT::f32 || EltVT == MVT::f64)
1887 case 2: return MVT::v2i1;
1888 case 4: return MVT::v4i1;
1889 case 8: return MVT::v8i1;
1891 if (Subtarget->hasBWI() && Subtarget->hasVLX())
1892 if (EltVT == MVT::i8 || EltVT == MVT::i16)
1894 case 8: return MVT::v8i1;
1895 case 16: return MVT::v16i1;
1896 case 32: return MVT::v32i1;
1900 return VT.changeVectorElementTypeToInteger();
1903 /// Helper for getByValTypeAlignment to determine
1904 /// the desired ByVal argument alignment.
1905 static void getMaxByValAlign(Type *Ty, unsigned &MaxAlign) {
1908 if (VectorType *VTy = dyn_cast<VectorType>(Ty)) {
1909 if (VTy->getBitWidth() == 128)
1911 } else if (ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
1912 unsigned EltAlign = 0;
1913 getMaxByValAlign(ATy->getElementType(), EltAlign);
1914 if (EltAlign > MaxAlign)
1915 MaxAlign = EltAlign;
1916 } else if (StructType *STy = dyn_cast<StructType>(Ty)) {
1917 for (auto *EltTy : STy->elements()) {
1918 unsigned EltAlign = 0;
1919 getMaxByValAlign(EltTy, EltAlign);
1920 if (EltAlign > MaxAlign)
1921 MaxAlign = EltAlign;
1928 /// Return the desired alignment for ByVal aggregate
1929 /// function arguments in the caller parameter area. For X86, aggregates
1930 /// that contain SSE vectors are placed at 16-byte boundaries while the rest
1931 /// are at 4-byte boundaries.
1932 unsigned X86TargetLowering::getByValTypeAlignment(Type *Ty,
1933 const DataLayout &DL) const {
1934 if (Subtarget->is64Bit()) {
1935 // Max of 8 and alignment of type.
1936 unsigned TyAlign = DL.getABITypeAlignment(Ty);
1943 if (Subtarget->hasSSE1())
1944 getMaxByValAlign(Ty, Align);
1948 /// Returns the target specific optimal type for load
1949 /// and store operations as a result of memset, memcpy, and memmove
1950 /// lowering. If DstAlign is zero that means it's safe to destination
1951 /// alignment can satisfy any constraint. Similarly if SrcAlign is zero it
1952 /// means there isn't a need to check it against alignment requirement,
1953 /// probably because the source does not need to be loaded. If 'IsMemset' is
1954 /// true, that means it's expanding a memset. If 'ZeroMemset' is true, that
1955 /// means it's a memset of zero. 'MemcpyStrSrc' indicates whether the memcpy
1956 /// source is constant so it does not need to be loaded.
1957 /// It returns EVT::Other if the type should be determined using generic
1958 /// target-independent logic.
1960 X86TargetLowering::getOptimalMemOpType(uint64_t Size,
1961 unsigned DstAlign, unsigned SrcAlign,
1962 bool IsMemset, bool ZeroMemset,
1964 MachineFunction &MF) const {
1965 const Function *F = MF.getFunction();
1966 if ((!IsMemset || ZeroMemset) &&
1967 !F->hasFnAttribute(Attribute::NoImplicitFloat)) {
1969 (!Subtarget->isUnalignedMem16Slow() ||
1970 ((DstAlign == 0 || DstAlign >= 16) &&
1971 (SrcAlign == 0 || SrcAlign >= 16)))) {
1973 // FIXME: Check if unaligned 32-byte accesses are slow.
1974 if (Subtarget->hasInt256())
1976 if (Subtarget->hasFp256())
1979 if (Subtarget->hasSSE2())
1981 if (Subtarget->hasSSE1())
1983 } else if (!MemcpyStrSrc && Size >= 8 &&
1984 !Subtarget->is64Bit() &&
1985 Subtarget->hasSSE2()) {
1986 // Do not use f64 to lower memcpy if source is string constant. It's
1987 // better to use i32 to avoid the loads.
1991 // This is a compromise. If we reach here, unaligned accesses may be slow on
1992 // this target. However, creating smaller, aligned accesses could be even
1993 // slower and would certainly be a lot more code.
1994 if (Subtarget->is64Bit() && Size >= 8)
1999 bool X86TargetLowering::isSafeMemOpType(MVT VT) const {
2001 return X86ScalarSSEf32;
2002 else if (VT == MVT::f64)
2003 return X86ScalarSSEf64;
2008 X86TargetLowering::allowsMisalignedMemoryAccesses(EVT VT,
2013 switch (VT.getSizeInBits()) {
2015 // 8-byte and under are always assumed to be fast.
2019 *Fast = !Subtarget->isUnalignedMem16Slow();
2022 *Fast = !Subtarget->isUnalignedMem32Slow();
2024 // TODO: What about AVX-512 (512-bit) accesses?
2027 // Misaligned accesses of any size are always allowed.
2031 /// Return the entry encoding for a jump table in the
2032 /// current function. The returned value is a member of the
2033 /// MachineJumpTableInfo::JTEntryKind enum.
2034 unsigned X86TargetLowering::getJumpTableEncoding() const {
2035 // In GOT pic mode, each entry in the jump table is emitted as a @GOTOFF
2037 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
2038 Subtarget->isPICStyleGOT())
2039 return MachineJumpTableInfo::EK_Custom32;
2041 // Otherwise, use the normal jump table encoding heuristics.
2042 return TargetLowering::getJumpTableEncoding();
2045 bool X86TargetLowering::useSoftFloat() const {
2046 return Subtarget->useSoftFloat();
2050 X86TargetLowering::LowerCustomJumpTableEntry(const MachineJumpTableInfo *MJTI,
2051 const MachineBasicBlock *MBB,
2052 unsigned uid,MCContext &Ctx) const{
2053 assert(MBB->getParent()->getTarget().getRelocationModel() == Reloc::PIC_ &&
2054 Subtarget->isPICStyleGOT());
2055 // In 32-bit ELF systems, our jump table entries are formed with @GOTOFF
2057 return MCSymbolRefExpr::create(MBB->getSymbol(),
2058 MCSymbolRefExpr::VK_GOTOFF, Ctx);
2061 /// Returns relocation base for the given PIC jumptable.
2062 SDValue X86TargetLowering::getPICJumpTableRelocBase(SDValue Table,
2063 SelectionDAG &DAG) const {
2064 if (!Subtarget->is64Bit())
2065 // This doesn't have SDLoc associated with it, but is not really the
2066 // same as a Register.
2067 return DAG.getNode(X86ISD::GlobalBaseReg, SDLoc(),
2068 getPointerTy(DAG.getDataLayout()));
2072 /// This returns the relocation base for the given PIC jumptable,
2073 /// the same as getPICJumpTableRelocBase, but as an MCExpr.
2074 const MCExpr *X86TargetLowering::
2075 getPICJumpTableRelocBaseExpr(const MachineFunction *MF, unsigned JTI,
2076 MCContext &Ctx) const {
2077 // X86-64 uses RIP relative addressing based on the jump table label.
2078 if (Subtarget->isPICStyleRIPRel())
2079 return TargetLowering::getPICJumpTableRelocBaseExpr(MF, JTI, Ctx);
2081 // Otherwise, the reference is relative to the PIC base.
2082 return MCSymbolRefExpr::create(MF->getPICBaseSymbol(), Ctx);
2085 std::pair<const TargetRegisterClass *, uint8_t>
2086 X86TargetLowering::findRepresentativeClass(const TargetRegisterInfo *TRI,
2088 const TargetRegisterClass *RRC = nullptr;
2090 switch (VT.SimpleTy) {
2092 return TargetLowering::findRepresentativeClass(TRI, VT);
2093 case MVT::i8: case MVT::i16: case MVT::i32: case MVT::i64:
2094 RRC = Subtarget->is64Bit() ? &X86::GR64RegClass : &X86::GR32RegClass;
2097 RRC = &X86::VR64RegClass;
2099 case MVT::f32: case MVT::f64:
2100 case MVT::v16i8: case MVT::v8i16: case MVT::v4i32: case MVT::v2i64:
2101 case MVT::v4f32: case MVT::v2f64:
2102 case MVT::v32i8: case MVT::v8i32: case MVT::v4i64: case MVT::v8f32:
2104 RRC = &X86::VR128RegClass;
2107 return std::make_pair(RRC, Cost);
2110 bool X86TargetLowering::getStackCookieLocation(unsigned &AddressSpace,
2111 unsigned &Offset) const {
2112 if (!Subtarget->isTargetLinux())
2115 if (Subtarget->is64Bit()) {
2116 // %fs:0x28, unless we're using a Kernel code model, in which case it's %gs:
2118 if (getTargetMachine().getCodeModel() == CodeModel::Kernel)
2130 /// Android provides a fixed TLS slot for the SafeStack pointer.
2131 /// See the definition of TLS_SLOT_SAFESTACK in
2132 /// https://android.googlesource.com/platform/bionic/+/master/libc/private/bionic_tls.h
2133 bool X86TargetLowering::getSafeStackPointerLocation(unsigned &AddressSpace,
2134 unsigned &Offset) const {
2135 if (!Subtarget->isTargetAndroid())
2138 if (Subtarget->is64Bit()) {
2139 // %fs:0x48, unless we're using a Kernel code model, in which case it's %gs:
2141 if (getTargetMachine().getCodeModel() == CodeModel::Kernel)
2153 bool X86TargetLowering::isNoopAddrSpaceCast(unsigned SrcAS,
2154 unsigned DestAS) const {
2155 assert(SrcAS != DestAS && "Expected different address spaces!");
2157 return SrcAS < 256 && DestAS < 256;
2160 //===----------------------------------------------------------------------===//
2161 // Return Value Calling Convention Implementation
2162 //===----------------------------------------------------------------------===//
2164 #include "X86GenCallingConv.inc"
2166 bool X86TargetLowering::CanLowerReturn(
2167 CallingConv::ID CallConv, MachineFunction &MF, bool isVarArg,
2168 const SmallVectorImpl<ISD::OutputArg> &Outs, LLVMContext &Context) const {
2169 SmallVector<CCValAssign, 16> RVLocs;
2170 CCState CCInfo(CallConv, isVarArg, MF, RVLocs, Context);
2171 return CCInfo.CheckReturn(Outs, RetCC_X86);
2174 const MCPhysReg *X86TargetLowering::getScratchRegisters(CallingConv::ID) const {
2175 static const MCPhysReg ScratchRegs[] = { X86::R11, 0 };
2180 X86TargetLowering::LowerReturn(SDValue Chain,
2181 CallingConv::ID CallConv, bool isVarArg,
2182 const SmallVectorImpl<ISD::OutputArg> &Outs,
2183 const SmallVectorImpl<SDValue> &OutVals,
2184 SDLoc dl, SelectionDAG &DAG) const {
2185 MachineFunction &MF = DAG.getMachineFunction();
2186 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
2188 SmallVector<CCValAssign, 16> RVLocs;
2189 CCState CCInfo(CallConv, isVarArg, MF, RVLocs, *DAG.getContext());
2190 CCInfo.AnalyzeReturn(Outs, RetCC_X86);
2193 SmallVector<SDValue, 6> RetOps;
2194 RetOps.push_back(Chain); // Operand #0 = Chain (updated below)
2195 // Operand #1 = Bytes To Pop
2196 RetOps.push_back(DAG.getTargetConstant(FuncInfo->getBytesToPopOnReturn(), dl,
2199 // Copy the result values into the output registers.
2200 for (unsigned i = 0; i != RVLocs.size(); ++i) {
2201 CCValAssign &VA = RVLocs[i];
2202 assert(VA.isRegLoc() && "Can only return in registers!");
2203 SDValue ValToCopy = OutVals[i];
2204 EVT ValVT = ValToCopy.getValueType();
2206 // Promote values to the appropriate types.
2207 if (VA.getLocInfo() == CCValAssign::SExt)
2208 ValToCopy = DAG.getNode(ISD::SIGN_EXTEND, dl, VA.getLocVT(), ValToCopy);
2209 else if (VA.getLocInfo() == CCValAssign::ZExt)
2210 ValToCopy = DAG.getNode(ISD::ZERO_EXTEND, dl, VA.getLocVT(), ValToCopy);
2211 else if (VA.getLocInfo() == CCValAssign::AExt) {
2212 if (ValVT.isVector() && ValVT.getScalarType() == MVT::i1)
2213 ValToCopy = DAG.getNode(ISD::SIGN_EXTEND, dl, VA.getLocVT(), ValToCopy);
2215 ValToCopy = DAG.getNode(ISD::ANY_EXTEND, dl, VA.getLocVT(), ValToCopy);
2217 else if (VA.getLocInfo() == CCValAssign::BCvt)
2218 ValToCopy = DAG.getBitcast(VA.getLocVT(), ValToCopy);
2220 assert(VA.getLocInfo() != CCValAssign::FPExt &&
2221 "Unexpected FP-extend for return value.");
2223 // If this is x86-64, and we disabled SSE, we can't return FP values,
2224 // or SSE or MMX vectors.
2225 if ((ValVT == MVT::f32 || ValVT == MVT::f64 ||
2226 VA.getLocReg() == X86::XMM0 || VA.getLocReg() == X86::XMM1) &&
2227 (Subtarget->is64Bit() && !Subtarget->hasSSE1())) {
2228 report_fatal_error("SSE register return with SSE disabled");
2230 // Likewise we can't return F64 values with SSE1 only. gcc does so, but
2231 // llvm-gcc has never done it right and no one has noticed, so this
2232 // should be OK for now.
2233 if (ValVT == MVT::f64 &&
2234 (Subtarget->is64Bit() && !Subtarget->hasSSE2()))
2235 report_fatal_error("SSE2 register return with SSE2 disabled");
2237 // Returns in ST0/ST1 are handled specially: these are pushed as operands to
2238 // the RET instruction and handled by the FP Stackifier.
2239 if (VA.getLocReg() == X86::FP0 ||
2240 VA.getLocReg() == X86::FP1) {
2241 // If this is a copy from an xmm register to ST(0), use an FPExtend to
2242 // change the value to the FP stack register class.
2243 if (isScalarFPTypeInSSEReg(VA.getValVT()))
2244 ValToCopy = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f80, ValToCopy);
2245 RetOps.push_back(ValToCopy);
2246 // Don't emit a copytoreg.
2250 // 64-bit vector (MMX) values are returned in XMM0 / XMM1 except for v1i64
2251 // which is returned in RAX / RDX.
2252 if (Subtarget->is64Bit()) {
2253 if (ValVT == MVT::x86mmx) {
2254 if (VA.getLocReg() == X86::XMM0 || VA.getLocReg() == X86::XMM1) {
2255 ValToCopy = DAG.getBitcast(MVT::i64, ValToCopy);
2256 ValToCopy = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i64,
2258 // If we don't have SSE2 available, convert to v4f32 so the generated
2259 // register is legal.
2260 if (!Subtarget->hasSSE2())
2261 ValToCopy = DAG.getBitcast(MVT::v4f32, ValToCopy);
2266 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), ValToCopy, Flag);
2267 Flag = Chain.getValue(1);
2268 RetOps.push_back(DAG.getRegister(VA.getLocReg(), VA.getLocVT()));
2271 // All x86 ABIs require that for returning structs by value we copy
2272 // the sret argument into %rax/%eax (depending on ABI) for the return.
2273 // We saved the argument into a virtual register in the entry block,
2274 // so now we copy the value out and into %rax/%eax.
2276 // Checking Function.hasStructRetAttr() here is insufficient because the IR
2277 // may not have an explicit sret argument. If FuncInfo.CanLowerReturn is
2278 // false, then an sret argument may be implicitly inserted in the SelDAG. In
2279 // either case FuncInfo->setSRetReturnReg() will have been called.
2280 if (unsigned SRetReg = FuncInfo->getSRetReturnReg()) {
2281 SDValue Val = DAG.getCopyFromReg(Chain, dl, SRetReg,
2282 getPointerTy(MF.getDataLayout()));
2285 = (Subtarget->is64Bit() && !Subtarget->isTarget64BitILP32()) ?
2286 X86::RAX : X86::EAX;
2287 Chain = DAG.getCopyToReg(Chain, dl, RetValReg, Val, Flag);
2288 Flag = Chain.getValue(1);
2290 // RAX/EAX now acts like a return value.
2292 DAG.getRegister(RetValReg, getPointerTy(DAG.getDataLayout())));
2295 RetOps[0] = Chain; // Update chain.
2297 // Add the flag if we have it.
2299 RetOps.push_back(Flag);
2301 return DAG.getNode(X86ISD::RET_FLAG, dl, MVT::Other, RetOps);
2304 bool X86TargetLowering::isUsedByReturnOnly(SDNode *N, SDValue &Chain) const {
2305 if (N->getNumValues() != 1)
2307 if (!N->hasNUsesOfValue(1, 0))
2310 SDValue TCChain = Chain;
2311 SDNode *Copy = *N->use_begin();
2312 if (Copy->getOpcode() == ISD::CopyToReg) {
2313 // If the copy has a glue operand, we conservatively assume it isn't safe to
2314 // perform a tail call.
2315 if (Copy->getOperand(Copy->getNumOperands()-1).getValueType() == MVT::Glue)
2317 TCChain = Copy->getOperand(0);
2318 } else if (Copy->getOpcode() != ISD::FP_EXTEND)
2321 bool HasRet = false;
2322 for (SDNode::use_iterator UI = Copy->use_begin(), UE = Copy->use_end();
2324 if (UI->getOpcode() != X86ISD::RET_FLAG)
2326 // If we are returning more than one value, we can definitely
2327 // not make a tail call see PR19530
2328 if (UI->getNumOperands() > 4)
2330 if (UI->getNumOperands() == 4 &&
2331 UI->getOperand(UI->getNumOperands()-1).getValueType() != MVT::Glue)
2344 X86TargetLowering::getTypeForExtArgOrReturn(LLVMContext &Context, EVT VT,
2345 ISD::NodeType ExtendKind) const {
2347 // TODO: Is this also valid on 32-bit?
2348 if (Subtarget->is64Bit() && VT == MVT::i1 && ExtendKind == ISD::ZERO_EXTEND)
2349 ReturnMVT = MVT::i8;
2351 ReturnMVT = MVT::i32;
2353 EVT MinVT = getRegisterType(Context, ReturnMVT);
2354 return VT.bitsLT(MinVT) ? MinVT : VT;
2357 /// Lower the result values of a call into the
2358 /// appropriate copies out of appropriate physical registers.
2361 X86TargetLowering::LowerCallResult(SDValue Chain, SDValue InFlag,
2362 CallingConv::ID CallConv, bool isVarArg,
2363 const SmallVectorImpl<ISD::InputArg> &Ins,
2364 SDLoc dl, SelectionDAG &DAG,
2365 SmallVectorImpl<SDValue> &InVals) const {
2367 // Assign locations to each value returned by this call.
2368 SmallVector<CCValAssign, 16> RVLocs;
2369 bool Is64Bit = Subtarget->is64Bit();
2370 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(), RVLocs,
2372 CCInfo.AnalyzeCallResult(Ins, RetCC_X86);
2374 // Copy all of the result registers out of their specified physreg.
2375 for (unsigned i = 0, e = RVLocs.size(); i != e; ++i) {
2376 CCValAssign &VA = RVLocs[i];
2377 EVT CopyVT = VA.getLocVT();
2379 // If this is x86-64, and we disabled SSE, we can't return FP values
2380 if ((CopyVT == MVT::f32 || CopyVT == MVT::f64) &&
2381 ((Is64Bit || Ins[i].Flags.isInReg()) && !Subtarget->hasSSE1())) {
2382 report_fatal_error("SSE register return with SSE disabled");
2385 // If we prefer to use the value in xmm registers, copy it out as f80 and
2386 // use a truncate to move it from fp stack reg to xmm reg.
2387 bool RoundAfterCopy = false;
2388 if ((VA.getLocReg() == X86::FP0 || VA.getLocReg() == X86::FP1) &&
2389 isScalarFPTypeInSSEReg(VA.getValVT())) {
2391 RoundAfterCopy = (CopyVT != VA.getLocVT());
2394 Chain = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(),
2395 CopyVT, InFlag).getValue(1);
2396 SDValue Val = Chain.getValue(0);
2399 Val = DAG.getNode(ISD::FP_ROUND, dl, VA.getValVT(), Val,
2400 // This truncation won't change the value.
2401 DAG.getIntPtrConstant(1, dl));
2403 if (VA.isExtInLoc() && VA.getValVT().getScalarType() == MVT::i1)
2404 Val = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), Val);
2406 InFlag = Chain.getValue(2);
2407 InVals.push_back(Val);
2413 //===----------------------------------------------------------------------===//
2414 // C & StdCall & Fast Calling Convention implementation
2415 //===----------------------------------------------------------------------===//
2416 // StdCall calling convention seems to be standard for many Windows' API
2417 // routines and around. It differs from C calling convention just a little:
2418 // callee should clean up the stack, not caller. Symbols should be also
2419 // decorated in some fancy way :) It doesn't support any vector arguments.
2420 // For info on fast calling convention see Fast Calling Convention (tail call)
2421 // implementation LowerX86_32FastCCCallTo.
2423 /// CallIsStructReturn - Determines whether a call uses struct return
2425 enum StructReturnType {
2430 static StructReturnType
2431 callIsStructReturn(const SmallVectorImpl<ISD::OutputArg> &Outs) {
2433 return NotStructReturn;
2435 const ISD::ArgFlagsTy &Flags = Outs[0].Flags;
2436 if (!Flags.isSRet())
2437 return NotStructReturn;
2438 if (Flags.isInReg())
2439 return RegStructReturn;
2440 return StackStructReturn;
2443 /// Determines whether a function uses struct return semantics.
2444 static StructReturnType
2445 argsAreStructReturn(const SmallVectorImpl<ISD::InputArg> &Ins) {
2447 return NotStructReturn;
2449 const ISD::ArgFlagsTy &Flags = Ins[0].Flags;
2450 if (!Flags.isSRet())
2451 return NotStructReturn;
2452 if (Flags.isInReg())
2453 return RegStructReturn;
2454 return StackStructReturn;
2457 /// Make a copy of an aggregate at address specified by "Src" to address
2458 /// "Dst" with size and alignment information specified by the specific
2459 /// parameter attribute. The copy will be passed as a byval function parameter.
2461 CreateCopyOfByValArgument(SDValue Src, SDValue Dst, SDValue Chain,
2462 ISD::ArgFlagsTy Flags, SelectionDAG &DAG,
2464 SDValue SizeNode = DAG.getConstant(Flags.getByValSize(), dl, MVT::i32);
2466 return DAG.getMemcpy(Chain, dl, Dst, Src, SizeNode, Flags.getByValAlign(),
2467 /*isVolatile*/false, /*AlwaysInline=*/true,
2468 /*isTailCall*/false,
2469 MachinePointerInfo(), MachinePointerInfo());
2472 /// Return true if the calling convention is one that we can guarantee TCO for.
2473 static bool canGuaranteeTCO(CallingConv::ID CC) {
2474 return (CC == CallingConv::Fast || CC == CallingConv::GHC ||
2475 CC == CallingConv::HiPE || CC == CallingConv::HHVM);
2478 /// Return true if we might ever do TCO for calls with this calling convention.
2479 static bool mayTailCallThisCC(CallingConv::ID CC) {
2481 // C calling conventions:
2482 case CallingConv::C:
2483 case CallingConv::X86_64_Win64:
2484 case CallingConv::X86_64_SysV:
2485 // Callee pop conventions:
2486 case CallingConv::X86_ThisCall:
2487 case CallingConv::X86_StdCall:
2488 case CallingConv::X86_VectorCall:
2489 case CallingConv::X86_FastCall:
2492 return canGuaranteeTCO(CC);
2496 /// Return true if the function is being made into a tailcall target by
2497 /// changing its ABI.
2498 static bool shouldGuaranteeTCO(CallingConv::ID CC, bool GuaranteedTailCallOpt) {
2499 return GuaranteedTailCallOpt && canGuaranteeTCO(CC);
2502 bool X86TargetLowering::mayBeEmittedAsTailCall(CallInst *CI) const {
2504 CI->getParent()->getParent()->getFnAttribute("disable-tail-calls");
2505 if (!CI->isTailCall() || Attr.getValueAsString() == "true")
2509 CallingConv::ID CalleeCC = CS.getCallingConv();
2510 if (!mayTailCallThisCC(CalleeCC))
2517 X86TargetLowering::LowerMemArgument(SDValue Chain,
2518 CallingConv::ID CallConv,
2519 const SmallVectorImpl<ISD::InputArg> &Ins,
2520 SDLoc dl, SelectionDAG &DAG,
2521 const CCValAssign &VA,
2522 MachineFrameInfo *MFI,
2524 // Create the nodes corresponding to a load from this parameter slot.
2525 ISD::ArgFlagsTy Flags = Ins[i].Flags;
2526 bool AlwaysUseMutable = shouldGuaranteeTCO(
2527 CallConv, DAG.getTarget().Options.GuaranteedTailCallOpt);
2528 bool isImmutable = !AlwaysUseMutable && !Flags.isByVal();
2531 // If value is passed by pointer we have address passed instead of the value
2533 bool ExtendedInMem = VA.isExtInLoc() &&
2534 VA.getValVT().getScalarType() == MVT::i1;
2536 if (VA.getLocInfo() == CCValAssign::Indirect || ExtendedInMem)
2537 ValVT = VA.getLocVT();
2539 ValVT = VA.getValVT();
2541 // FIXME: For now, all byval parameter objects are marked mutable. This can be
2542 // changed with more analysis.
2543 // In case of tail call optimization mark all arguments mutable. Since they
2544 // could be overwritten by lowering of arguments in case of a tail call.
2545 if (Flags.isByVal()) {
2546 unsigned Bytes = Flags.getByValSize();
2547 if (Bytes == 0) Bytes = 1; // Don't create zero-sized stack objects.
2548 int FI = MFI->CreateFixedObject(Bytes, VA.getLocMemOffset(), isImmutable);
2549 return DAG.getFrameIndex(FI, getPointerTy(DAG.getDataLayout()));
2551 int FI = MFI->CreateFixedObject(ValVT.getSizeInBits()/8,
2552 VA.getLocMemOffset(), isImmutable);
2553 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy(DAG.getDataLayout()));
2554 SDValue Val = DAG.getLoad(
2555 ValVT, dl, Chain, FIN,
2556 MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), FI), false,
2558 return ExtendedInMem ?
2559 DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), Val) : Val;
2563 // FIXME: Get this from tablegen.
2564 static ArrayRef<MCPhysReg> get64BitArgumentGPRs(CallingConv::ID CallConv,
2565 const X86Subtarget *Subtarget) {
2566 assert(Subtarget->is64Bit());
2568 if (Subtarget->isCallingConvWin64(CallConv)) {
2569 static const MCPhysReg GPR64ArgRegsWin64[] = {
2570 X86::RCX, X86::RDX, X86::R8, X86::R9
2572 return makeArrayRef(std::begin(GPR64ArgRegsWin64), std::end(GPR64ArgRegsWin64));
2575 static const MCPhysReg GPR64ArgRegs64Bit[] = {
2576 X86::RDI, X86::RSI, X86::RDX, X86::RCX, X86::R8, X86::R9
2578 return makeArrayRef(std::begin(GPR64ArgRegs64Bit), std::end(GPR64ArgRegs64Bit));
2581 // FIXME: Get this from tablegen.
2582 static ArrayRef<MCPhysReg> get64BitArgumentXMMs(MachineFunction &MF,
2583 CallingConv::ID CallConv,
2584 const X86Subtarget *Subtarget) {
2585 assert(Subtarget->is64Bit());
2586 if (Subtarget->isCallingConvWin64(CallConv)) {
2587 // The XMM registers which might contain var arg parameters are shadowed
2588 // in their paired GPR. So we only need to save the GPR to their home
2590 // TODO: __vectorcall will change this.
2594 const Function *Fn = MF.getFunction();
2595 bool NoImplicitFloatOps = Fn->hasFnAttribute(Attribute::NoImplicitFloat);
2596 bool isSoftFloat = Subtarget->useSoftFloat();
2597 assert(!(isSoftFloat && NoImplicitFloatOps) &&
2598 "SSE register cannot be used when SSE is disabled!");
2599 if (isSoftFloat || NoImplicitFloatOps || !Subtarget->hasSSE1())
2600 // Kernel mode asks for SSE to be disabled, so there are no XMM argument
2604 static const MCPhysReg XMMArgRegs64Bit[] = {
2605 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
2606 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
2608 return makeArrayRef(std::begin(XMMArgRegs64Bit), std::end(XMMArgRegs64Bit));
2611 SDValue X86TargetLowering::LowerFormalArguments(
2612 SDValue Chain, CallingConv::ID CallConv, bool isVarArg,
2613 const SmallVectorImpl<ISD::InputArg> &Ins, SDLoc dl, SelectionDAG &DAG,
2614 SmallVectorImpl<SDValue> &InVals) const {
2615 MachineFunction &MF = DAG.getMachineFunction();
2616 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
2617 const TargetFrameLowering &TFI = *Subtarget->getFrameLowering();
2619 const Function* Fn = MF.getFunction();
2620 if (Fn->hasExternalLinkage() &&
2621 Subtarget->isTargetCygMing() &&
2622 Fn->getName() == "main")
2623 FuncInfo->setForceFramePointer(true);
2625 MachineFrameInfo *MFI = MF.getFrameInfo();
2626 bool Is64Bit = Subtarget->is64Bit();
2627 bool IsWin64 = Subtarget->isCallingConvWin64(CallConv);
2629 assert(!(isVarArg && canGuaranteeTCO(CallConv)) &&
2630 "Var args not supported with calling convention fastcc, ghc or hipe");
2632 // Assign locations to all of the incoming arguments.
2633 SmallVector<CCValAssign, 16> ArgLocs;
2634 CCState CCInfo(CallConv, isVarArg, MF, ArgLocs, *DAG.getContext());
2636 // Allocate shadow area for Win64
2638 CCInfo.AllocateStack(32, 8);
2640 CCInfo.AnalyzeFormalArguments(Ins, CC_X86);
2642 unsigned LastVal = ~0U;
2644 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2645 CCValAssign &VA = ArgLocs[i];
2646 // TODO: If an arg is passed in two places (e.g. reg and stack), skip later
2648 assert(VA.getValNo() != LastVal &&
2649 "Don't support value assigned to multiple locs yet");
2651 LastVal = VA.getValNo();
2653 if (VA.isRegLoc()) {
2654 EVT RegVT = VA.getLocVT();
2655 const TargetRegisterClass *RC;
2656 if (RegVT == MVT::i32)
2657 RC = &X86::GR32RegClass;
2658 else if (Is64Bit && RegVT == MVT::i64)
2659 RC = &X86::GR64RegClass;
2660 else if (RegVT == MVT::f32)
2661 RC = &X86::FR32RegClass;
2662 else if (RegVT == MVT::f64)
2663 RC = &X86::FR64RegClass;
2664 else if (RegVT.is512BitVector())
2665 RC = &X86::VR512RegClass;
2666 else if (RegVT.is256BitVector())
2667 RC = &X86::VR256RegClass;
2668 else if (RegVT.is128BitVector())
2669 RC = &X86::VR128RegClass;
2670 else if (RegVT == MVT::x86mmx)
2671 RC = &X86::VR64RegClass;
2672 else if (RegVT == MVT::i1)
2673 RC = &X86::VK1RegClass;
2674 else if (RegVT == MVT::v8i1)
2675 RC = &X86::VK8RegClass;
2676 else if (RegVT == MVT::v16i1)
2677 RC = &X86::VK16RegClass;
2678 else if (RegVT == MVT::v32i1)
2679 RC = &X86::VK32RegClass;
2680 else if (RegVT == MVT::v64i1)
2681 RC = &X86::VK64RegClass;
2683 llvm_unreachable("Unknown argument type!");
2685 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
2686 ArgValue = DAG.getCopyFromReg(Chain, dl, Reg, RegVT);
2688 // If this is an 8 or 16-bit value, it is really passed promoted to 32
2689 // bits. Insert an assert[sz]ext to capture this, then truncate to the
2691 if (VA.getLocInfo() == CCValAssign::SExt)
2692 ArgValue = DAG.getNode(ISD::AssertSext, dl, RegVT, ArgValue,
2693 DAG.getValueType(VA.getValVT()));
2694 else if (VA.getLocInfo() == CCValAssign::ZExt)
2695 ArgValue = DAG.getNode(ISD::AssertZext, dl, RegVT, ArgValue,
2696 DAG.getValueType(VA.getValVT()));
2697 else if (VA.getLocInfo() == CCValAssign::BCvt)
2698 ArgValue = DAG.getBitcast(VA.getValVT(), ArgValue);
2700 if (VA.isExtInLoc()) {
2701 // Handle MMX values passed in XMM regs.
2702 if (RegVT.isVector() && VA.getValVT().getScalarType() != MVT::i1)
2703 ArgValue = DAG.getNode(X86ISD::MOVDQ2Q, dl, VA.getValVT(), ArgValue);
2705 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
2708 assert(VA.isMemLoc());
2709 ArgValue = LowerMemArgument(Chain, CallConv, Ins, dl, DAG, VA, MFI, i);
2712 // If value is passed via pointer - do a load.
2713 if (VA.getLocInfo() == CCValAssign::Indirect)
2714 ArgValue = DAG.getLoad(VA.getValVT(), dl, Chain, ArgValue,
2715 MachinePointerInfo(), false, false, false, 0);
2717 InVals.push_back(ArgValue);
2720 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2721 // All x86 ABIs require that for returning structs by value we copy the
2722 // sret argument into %rax/%eax (depending on ABI) for the return. Save
2723 // the argument into a virtual register so that we can access it from the
2725 if (Ins[i].Flags.isSRet()) {
2726 unsigned Reg = FuncInfo->getSRetReturnReg();
2728 MVT PtrTy = getPointerTy(DAG.getDataLayout());
2729 Reg = MF.getRegInfo().createVirtualRegister(getRegClassFor(PtrTy));
2730 FuncInfo->setSRetReturnReg(Reg);
2732 SDValue Copy = DAG.getCopyToReg(DAG.getEntryNode(), dl, Reg, InVals[i]);
2733 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Copy, Chain);
2738 unsigned StackSize = CCInfo.getNextStackOffset();
2739 // Align stack specially for tail calls.
2740 if (shouldGuaranteeTCO(CallConv,
2741 MF.getTarget().Options.GuaranteedTailCallOpt))
2742 StackSize = GetAlignedArgumentStackSize(StackSize, DAG);
2744 // If the function takes variable number of arguments, make a frame index for
2745 // the start of the first vararg value... for expansion of llvm.va_start. We
2746 // can skip this if there are no va_start calls.
2747 if (MFI->hasVAStart() &&
2748 (Is64Bit || (CallConv != CallingConv::X86_FastCall &&
2749 CallConv != CallingConv::X86_ThisCall))) {
2750 FuncInfo->setVarArgsFrameIndex(
2751 MFI->CreateFixedObject(1, StackSize, true));
2754 MachineModuleInfo &MMI = MF.getMMI();
2756 // Figure out if XMM registers are in use.
2757 assert(!(Subtarget->useSoftFloat() &&
2758 Fn->hasFnAttribute(Attribute::NoImplicitFloat)) &&
2759 "SSE register cannot be used when SSE is disabled!");
2761 // 64-bit calling conventions support varargs and register parameters, so we
2762 // have to do extra work to spill them in the prologue.
2763 if (Is64Bit && isVarArg && MFI->hasVAStart()) {
2764 // Find the first unallocated argument registers.
2765 ArrayRef<MCPhysReg> ArgGPRs = get64BitArgumentGPRs(CallConv, Subtarget);
2766 ArrayRef<MCPhysReg> ArgXMMs = get64BitArgumentXMMs(MF, CallConv, Subtarget);
2767 unsigned NumIntRegs = CCInfo.getFirstUnallocated(ArgGPRs);
2768 unsigned NumXMMRegs = CCInfo.getFirstUnallocated(ArgXMMs);
2769 assert(!(NumXMMRegs && !Subtarget->hasSSE1()) &&
2770 "SSE register cannot be used when SSE is disabled!");
2772 // Gather all the live in physical registers.
2773 SmallVector<SDValue, 6> LiveGPRs;
2774 SmallVector<SDValue, 8> LiveXMMRegs;
2776 for (MCPhysReg Reg : ArgGPRs.slice(NumIntRegs)) {
2777 unsigned GPR = MF.addLiveIn(Reg, &X86::GR64RegClass);
2779 DAG.getCopyFromReg(Chain, dl, GPR, MVT::i64));
2781 if (!ArgXMMs.empty()) {
2782 unsigned AL = MF.addLiveIn(X86::AL, &X86::GR8RegClass);
2783 ALVal = DAG.getCopyFromReg(Chain, dl, AL, MVT::i8);
2784 for (MCPhysReg Reg : ArgXMMs.slice(NumXMMRegs)) {
2785 unsigned XMMReg = MF.addLiveIn(Reg, &X86::VR128RegClass);
2786 LiveXMMRegs.push_back(
2787 DAG.getCopyFromReg(Chain, dl, XMMReg, MVT::v4f32));
2792 // Get to the caller-allocated home save location. Add 8 to account
2793 // for the return address.
2794 int HomeOffset = TFI.getOffsetOfLocalArea() + 8;
2795 FuncInfo->setRegSaveFrameIndex(
2796 MFI->CreateFixedObject(1, NumIntRegs * 8 + HomeOffset, false));
2797 // Fixup to set vararg frame on shadow area (4 x i64).
2799 FuncInfo->setVarArgsFrameIndex(FuncInfo->getRegSaveFrameIndex());
2801 // For X86-64, if there are vararg parameters that are passed via
2802 // registers, then we must store them to their spots on the stack so
2803 // they may be loaded by deferencing the result of va_next.
2804 FuncInfo->setVarArgsGPOffset(NumIntRegs * 8);
2805 FuncInfo->setVarArgsFPOffset(ArgGPRs.size() * 8 + NumXMMRegs * 16);
2806 FuncInfo->setRegSaveFrameIndex(MFI->CreateStackObject(
2807 ArgGPRs.size() * 8 + ArgXMMs.size() * 16, 16, false));
2810 // Store the integer parameter registers.
2811 SmallVector<SDValue, 8> MemOps;
2812 SDValue RSFIN = DAG.getFrameIndex(FuncInfo->getRegSaveFrameIndex(),
2813 getPointerTy(DAG.getDataLayout()));
2814 unsigned Offset = FuncInfo->getVarArgsGPOffset();
2815 for (SDValue Val : LiveGPRs) {
2816 SDValue FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(DAG.getDataLayout()),
2817 RSFIN, DAG.getIntPtrConstant(Offset, dl));
2819 DAG.getStore(Val.getValue(1), dl, Val, FIN,
2820 MachinePointerInfo::getFixedStack(
2821 DAG.getMachineFunction(),
2822 FuncInfo->getRegSaveFrameIndex(), Offset),
2824 MemOps.push_back(Store);
2828 if (!ArgXMMs.empty() && NumXMMRegs != ArgXMMs.size()) {
2829 // Now store the XMM (fp + vector) parameter registers.
2830 SmallVector<SDValue, 12> SaveXMMOps;
2831 SaveXMMOps.push_back(Chain);
2832 SaveXMMOps.push_back(ALVal);
2833 SaveXMMOps.push_back(DAG.getIntPtrConstant(
2834 FuncInfo->getRegSaveFrameIndex(), dl));
2835 SaveXMMOps.push_back(DAG.getIntPtrConstant(
2836 FuncInfo->getVarArgsFPOffset(), dl));
2837 SaveXMMOps.insert(SaveXMMOps.end(), LiveXMMRegs.begin(),
2839 MemOps.push_back(DAG.getNode(X86ISD::VASTART_SAVE_XMM_REGS, dl,
2840 MVT::Other, SaveXMMOps));
2843 if (!MemOps.empty())
2844 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, MemOps);
2847 if (isVarArg && MFI->hasMustTailInVarArgFunc()) {
2848 // Find the largest legal vector type.
2849 MVT VecVT = MVT::Other;
2850 // FIXME: Only some x86_32 calling conventions support AVX512.
2851 if (Subtarget->hasAVX512() &&
2852 (Is64Bit || (CallConv == CallingConv::X86_VectorCall ||
2853 CallConv == CallingConv::Intel_OCL_BI)))
2854 VecVT = MVT::v16f32;
2855 else if (Subtarget->hasAVX())
2857 else if (Subtarget->hasSSE2())
2860 // We forward some GPRs and some vector types.
2861 SmallVector<MVT, 2> RegParmTypes;
2862 MVT IntVT = Is64Bit ? MVT::i64 : MVT::i32;
2863 RegParmTypes.push_back(IntVT);
2864 if (VecVT != MVT::Other)
2865 RegParmTypes.push_back(VecVT);
2867 // Compute the set of forwarded registers. The rest are scratch.
2868 SmallVectorImpl<ForwardedRegister> &Forwards =
2869 FuncInfo->getForwardedMustTailRegParms();
2870 CCInfo.analyzeMustTailForwardedRegisters(Forwards, RegParmTypes, CC_X86);
2872 // Conservatively forward AL on x86_64, since it might be used for varargs.
2873 if (Is64Bit && !CCInfo.isAllocated(X86::AL)) {
2874 unsigned ALVReg = MF.addLiveIn(X86::AL, &X86::GR8RegClass);
2875 Forwards.push_back(ForwardedRegister(ALVReg, X86::AL, MVT::i8));
2878 // Copy all forwards from physical to virtual registers.
2879 for (ForwardedRegister &F : Forwards) {
2880 // FIXME: Can we use a less constrained schedule?
2881 SDValue RegVal = DAG.getCopyFromReg(Chain, dl, F.VReg, F.VT);
2882 F.VReg = MF.getRegInfo().createVirtualRegister(getRegClassFor(F.VT));
2883 Chain = DAG.getCopyToReg(Chain, dl, F.VReg, RegVal);
2887 // Some CCs need callee pop.
2888 if (X86::isCalleePop(CallConv, Is64Bit, isVarArg,
2889 MF.getTarget().Options.GuaranteedTailCallOpt)) {
2890 FuncInfo->setBytesToPopOnReturn(StackSize); // Callee pops everything.
2892 FuncInfo->setBytesToPopOnReturn(0); // Callee pops nothing.
2893 // If this is an sret function, the return should pop the hidden pointer.
2894 if (!Is64Bit && !canGuaranteeTCO(CallConv) &&
2895 !Subtarget->getTargetTriple().isOSMSVCRT() &&
2896 argsAreStructReturn(Ins) == StackStructReturn)
2897 FuncInfo->setBytesToPopOnReturn(4);
2901 // RegSaveFrameIndex is X86-64 only.
2902 FuncInfo->setRegSaveFrameIndex(0xAAAAAAA);
2903 if (CallConv == CallingConv::X86_FastCall ||
2904 CallConv == CallingConv::X86_ThisCall)
2905 // fastcc functions can't have varargs.
2906 FuncInfo->setVarArgsFrameIndex(0xAAAAAAA);
2909 FuncInfo->setArgumentStackSize(StackSize);
2911 if (MMI.hasWinEHFuncInfo(Fn)) {
2913 int UnwindHelpFI = MFI->CreateStackObject(8, 8, /*isSS=*/false);
2914 SDValue StackSlot = DAG.getFrameIndex(UnwindHelpFI, MVT::i64);
2915 MMI.getWinEHFuncInfo(MF.getFunction()).UnwindHelpFrameIdx = UnwindHelpFI;
2916 SDValue Neg2 = DAG.getConstant(-2, dl, MVT::i64);
2917 Chain = DAG.getStore(Chain, dl, Neg2, StackSlot,
2918 MachinePointerInfo::getFixedStack(
2919 DAG.getMachineFunction(), UnwindHelpFI),
2920 /*isVolatile=*/true,
2921 /*isNonTemporal=*/false, /*Alignment=*/0);
2923 // Functions using Win32 EH are considered to have opaque SP adjustments
2924 // to force local variables to be addressed from the frame or base
2926 MFI->setHasOpaqueSPAdjustment(true);
2934 X86TargetLowering::LowerMemOpCallTo(SDValue Chain,
2935 SDValue StackPtr, SDValue Arg,
2936 SDLoc dl, SelectionDAG &DAG,
2937 const CCValAssign &VA,
2938 ISD::ArgFlagsTy Flags) const {
2939 unsigned LocMemOffset = VA.getLocMemOffset();
2940 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset, dl);
2941 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(DAG.getDataLayout()),
2943 if (Flags.isByVal())
2944 return CreateCopyOfByValArgument(Arg, PtrOff, Chain, Flags, DAG, dl);
2946 return DAG.getStore(
2947 Chain, dl, Arg, PtrOff,
2948 MachinePointerInfo::getStack(DAG.getMachineFunction(), LocMemOffset),
2952 /// Emit a load of return address if tail call
2953 /// optimization is performed and it is required.
2955 X86TargetLowering::EmitTailCallLoadRetAddr(SelectionDAG &DAG,
2956 SDValue &OutRetAddr, SDValue Chain,
2957 bool IsTailCall, bool Is64Bit,
2958 int FPDiff, SDLoc dl) const {
2959 // Adjust the Return address stack slot.
2960 EVT VT = getPointerTy(DAG.getDataLayout());
2961 OutRetAddr = getReturnAddressFrameIndex(DAG);
2963 // Load the "old" Return address.
2964 OutRetAddr = DAG.getLoad(VT, dl, Chain, OutRetAddr, MachinePointerInfo(),
2965 false, false, false, 0);
2966 return SDValue(OutRetAddr.getNode(), 1);
2969 /// Emit a store of the return address if tail call
2970 /// optimization is performed and it is required (FPDiff!=0).
2971 static SDValue EmitTailCallStoreRetAddr(SelectionDAG &DAG, MachineFunction &MF,
2972 SDValue Chain, SDValue RetAddrFrIdx,
2973 EVT PtrVT, unsigned SlotSize,
2974 int FPDiff, SDLoc dl) {
2975 // Store the return address to the appropriate stack slot.
2976 if (!FPDiff) return Chain;
2977 // Calculate the new stack slot for the return address.
2978 int NewReturnAddrFI =
2979 MF.getFrameInfo()->CreateFixedObject(SlotSize, (int64_t)FPDiff - SlotSize,
2981 SDValue NewRetAddrFrIdx = DAG.getFrameIndex(NewReturnAddrFI, PtrVT);
2982 Chain = DAG.getStore(Chain, dl, RetAddrFrIdx, NewRetAddrFrIdx,
2983 MachinePointerInfo::getFixedStack(
2984 DAG.getMachineFunction(), NewReturnAddrFI),
2989 /// Returns a vector_shuffle mask for an movs{s|d}, movd
2990 /// operation of specified width.
2991 static SDValue getMOVL(SelectionDAG &DAG, SDLoc dl, EVT VT, SDValue V1,
2993 unsigned NumElems = VT.getVectorNumElements();
2994 SmallVector<int, 8> Mask;
2995 Mask.push_back(NumElems);
2996 for (unsigned i = 1; i != NumElems; ++i)
2998 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
3002 X86TargetLowering::LowerCall(TargetLowering::CallLoweringInfo &CLI,
3003 SmallVectorImpl<SDValue> &InVals) const {
3004 SelectionDAG &DAG = CLI.DAG;
3006 SmallVectorImpl<ISD::OutputArg> &Outs = CLI.Outs;
3007 SmallVectorImpl<SDValue> &OutVals = CLI.OutVals;
3008 SmallVectorImpl<ISD::InputArg> &Ins = CLI.Ins;
3009 SDValue Chain = CLI.Chain;
3010 SDValue Callee = CLI.Callee;
3011 CallingConv::ID CallConv = CLI.CallConv;
3012 bool &isTailCall = CLI.IsTailCall;
3013 bool isVarArg = CLI.IsVarArg;
3015 MachineFunction &MF = DAG.getMachineFunction();
3016 bool Is64Bit = Subtarget->is64Bit();
3017 bool IsWin64 = Subtarget->isCallingConvWin64(CallConv);
3018 StructReturnType SR = callIsStructReturn(Outs);
3019 bool IsSibcall = false;
3020 X86MachineFunctionInfo *X86Info = MF.getInfo<X86MachineFunctionInfo>();
3021 auto Attr = MF.getFunction()->getFnAttribute("disable-tail-calls");
3023 if (Attr.getValueAsString() == "true")
3026 if (Subtarget->isPICStyleGOT() &&
3027 !MF.getTarget().Options.GuaranteedTailCallOpt) {
3028 // If we are using a GOT, disable tail calls to external symbols with
3029 // default visibility. Tail calling such a symbol requires using a GOT
3030 // relocation, which forces early binding of the symbol. This breaks code
3031 // that require lazy function symbol resolution. Using musttail or
3032 // GuaranteedTailCallOpt will override this.
3033 GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee);
3034 if (!G || (!G->getGlobal()->hasLocalLinkage() &&
3035 G->getGlobal()->hasDefaultVisibility()))
3039 bool IsMustTail = CLI.CS && CLI.CS->isMustTailCall();
3041 // Force this to be a tail call. The verifier rules are enough to ensure
3042 // that we can lower this successfully without moving the return address
3045 } else if (isTailCall) {
3046 // Check if it's really possible to do a tail call.
3047 isTailCall = IsEligibleForTailCallOptimization(Callee, CallConv,
3048 isVarArg, SR != NotStructReturn,
3049 MF.getFunction()->hasStructRetAttr(), CLI.RetTy,
3050 Outs, OutVals, Ins, DAG);
3052 // Sibcalls are automatically detected tailcalls which do not require
3054 if (!MF.getTarget().Options.GuaranteedTailCallOpt && isTailCall)
3061 assert(!(isVarArg && canGuaranteeTCO(CallConv)) &&
3062 "Var args not supported with calling convention fastcc, ghc or hipe");
3064 // Analyze operands of the call, assigning locations to each operand.
3065 SmallVector<CCValAssign, 16> ArgLocs;
3066 CCState CCInfo(CallConv, isVarArg, MF, ArgLocs, *DAG.getContext());
3068 // Allocate shadow area for Win64
3070 CCInfo.AllocateStack(32, 8);
3072 CCInfo.AnalyzeCallOperands(Outs, CC_X86);
3074 // Get a count of how many bytes are to be pushed on the stack.
3075 unsigned NumBytes = CCInfo.getAlignedCallFrameSize();
3077 // This is a sibcall. The memory operands are available in caller's
3078 // own caller's stack.
3080 else if (MF.getTarget().Options.GuaranteedTailCallOpt &&
3081 canGuaranteeTCO(CallConv))
3082 NumBytes = GetAlignedArgumentStackSize(NumBytes, DAG);
3085 if (isTailCall && !IsSibcall && !IsMustTail) {
3086 // Lower arguments at fp - stackoffset + fpdiff.
3087 unsigned NumBytesCallerPushed = X86Info->getBytesToPopOnReturn();
3089 FPDiff = NumBytesCallerPushed - NumBytes;
3091 // Set the delta of movement of the returnaddr stackslot.
3092 // But only set if delta is greater than previous delta.
3093 if (FPDiff < X86Info->getTCReturnAddrDelta())
3094 X86Info->setTCReturnAddrDelta(FPDiff);
3097 unsigned NumBytesToPush = NumBytes;
3098 unsigned NumBytesToPop = NumBytes;
3100 // If we have an inalloca argument, all stack space has already been allocated
3101 // for us and be right at the top of the stack. We don't support multiple
3102 // arguments passed in memory when using inalloca.
3103 if (!Outs.empty() && Outs.back().Flags.isInAlloca()) {
3105 if (!ArgLocs.back().isMemLoc())
3106 report_fatal_error("cannot use inalloca attribute on a register "
3108 if (ArgLocs.back().getLocMemOffset() != 0)
3109 report_fatal_error("any parameter with the inalloca attribute must be "
3110 "the only memory argument");
3114 Chain = DAG.getCALLSEQ_START(
3115 Chain, DAG.getIntPtrConstant(NumBytesToPush, dl, true), dl);
3117 SDValue RetAddrFrIdx;
3118 // Load return address for tail calls.
3119 if (isTailCall && FPDiff)
3120 Chain = EmitTailCallLoadRetAddr(DAG, RetAddrFrIdx, Chain, isTailCall,
3121 Is64Bit, FPDiff, dl);
3123 SmallVector<std::pair<unsigned, SDValue>, 8> RegsToPass;
3124 SmallVector<SDValue, 8> MemOpChains;
3127 // Walk the register/memloc assignments, inserting copies/loads. In the case
3128 // of tail call optimization arguments are handle later.
3129 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
3130 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
3131 // Skip inalloca arguments, they have already been written.
3132 ISD::ArgFlagsTy Flags = Outs[i].Flags;
3133 if (Flags.isInAlloca())
3136 CCValAssign &VA = ArgLocs[i];
3137 EVT RegVT = VA.getLocVT();
3138 SDValue Arg = OutVals[i];
3139 bool isByVal = Flags.isByVal();
3141 // Promote the value if needed.
3142 switch (VA.getLocInfo()) {
3143 default: llvm_unreachable("Unknown loc info!");
3144 case CCValAssign::Full: break;
3145 case CCValAssign::SExt:
3146 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, RegVT, Arg);
3148 case CCValAssign::ZExt:
3149 Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, RegVT, Arg);
3151 case CCValAssign::AExt:
3152 if (Arg.getValueType().isVector() &&
3153 Arg.getValueType().getScalarType() == MVT::i1)
3154 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, RegVT, Arg);
3155 else if (RegVT.is128BitVector()) {
3156 // Special case: passing MMX values in XMM registers.
3157 Arg = DAG.getBitcast(MVT::i64, Arg);
3158 Arg = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i64, Arg);
3159 Arg = getMOVL(DAG, dl, MVT::v2i64, DAG.getUNDEF(MVT::v2i64), Arg);
3161 Arg = DAG.getNode(ISD::ANY_EXTEND, dl, RegVT, Arg);
3163 case CCValAssign::BCvt:
3164 Arg = DAG.getBitcast(RegVT, Arg);
3166 case CCValAssign::Indirect: {
3167 // Store the argument.
3168 SDValue SpillSlot = DAG.CreateStackTemporary(VA.getValVT());
3169 int FI = cast<FrameIndexSDNode>(SpillSlot)->getIndex();
3170 Chain = DAG.getStore(
3171 Chain, dl, Arg, SpillSlot,
3172 MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), FI),
3179 if (VA.isRegLoc()) {
3180 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
3181 if (isVarArg && IsWin64) {
3182 // Win64 ABI requires argument XMM reg to be copied to the corresponding
3183 // shadow reg if callee is a varargs function.
3184 unsigned ShadowReg = 0;
3185 switch (VA.getLocReg()) {
3186 case X86::XMM0: ShadowReg = X86::RCX; break;
3187 case X86::XMM1: ShadowReg = X86::RDX; break;
3188 case X86::XMM2: ShadowReg = X86::R8; break;
3189 case X86::XMM3: ShadowReg = X86::R9; break;
3192 RegsToPass.push_back(std::make_pair(ShadowReg, Arg));
3194 } else if (!IsSibcall && (!isTailCall || isByVal)) {
3195 assert(VA.isMemLoc());
3196 if (!StackPtr.getNode())
3197 StackPtr = DAG.getCopyFromReg(Chain, dl, RegInfo->getStackRegister(),
3198 getPointerTy(DAG.getDataLayout()));
3199 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Arg,
3200 dl, DAG, VA, Flags));
3204 if (!MemOpChains.empty())
3205 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, MemOpChains);
3207 if (Subtarget->isPICStyleGOT()) {
3208 // ELF / PIC requires GOT in the EBX register before function calls via PLT
3211 RegsToPass.push_back(std::make_pair(
3212 unsigned(X86::EBX), DAG.getNode(X86ISD::GlobalBaseReg, SDLoc(),
3213 getPointerTy(DAG.getDataLayout()))));
3215 // If we are tail calling and generating PIC/GOT style code load the
3216 // address of the callee into ECX. The value in ecx is used as target of
3217 // the tail jump. This is done to circumvent the ebx/callee-saved problem
3218 // for tail calls on PIC/GOT architectures. Normally we would just put the
3219 // address of GOT into ebx and then call target@PLT. But for tail calls
3220 // ebx would be restored (since ebx is callee saved) before jumping to the
3223 // Note: The actual moving to ECX is done further down.
3224 GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee);
3225 if (G && !G->getGlobal()->hasLocalLinkage() &&
3226 G->getGlobal()->hasDefaultVisibility())
3227 Callee = LowerGlobalAddress(Callee, DAG);
3228 else if (isa<ExternalSymbolSDNode>(Callee))
3229 Callee = LowerExternalSymbol(Callee, DAG);
3233 if (Is64Bit && isVarArg && !IsWin64 && !IsMustTail) {
3234 // From AMD64 ABI document:
3235 // For calls that may call functions that use varargs or stdargs
3236 // (prototype-less calls or calls to functions containing ellipsis (...) in
3237 // the declaration) %al is used as hidden argument to specify the number
3238 // of SSE registers used. The contents of %al do not need to match exactly
3239 // the number of registers, but must be an ubound on the number of SSE
3240 // registers used and is in the range 0 - 8 inclusive.
3242 // Count the number of XMM registers allocated.
3243 static const MCPhysReg XMMArgRegs[] = {
3244 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
3245 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
3247 unsigned NumXMMRegs = CCInfo.getFirstUnallocated(XMMArgRegs);
3248 assert((Subtarget->hasSSE1() || !NumXMMRegs)
3249 && "SSE registers cannot be used when SSE is disabled");
3251 RegsToPass.push_back(std::make_pair(unsigned(X86::AL),
3252 DAG.getConstant(NumXMMRegs, dl,
3256 if (isVarArg && IsMustTail) {
3257 const auto &Forwards = X86Info->getForwardedMustTailRegParms();
3258 for (const auto &F : Forwards) {
3259 SDValue Val = DAG.getCopyFromReg(Chain, dl, F.VReg, F.VT);
3260 RegsToPass.push_back(std::make_pair(unsigned(F.PReg), Val));
3264 // For tail calls lower the arguments to the 'real' stack slots. Sibcalls
3265 // don't need this because the eligibility check rejects calls that require
3266 // shuffling arguments passed in memory.
3267 if (!IsSibcall && isTailCall) {
3268 // Force all the incoming stack arguments to be loaded from the stack
3269 // before any new outgoing arguments are stored to the stack, because the
3270 // outgoing stack slots may alias the incoming argument stack slots, and
3271 // the alias isn't otherwise explicit. This is slightly more conservative
3272 // than necessary, because it means that each store effectively depends
3273 // on every argument instead of just those arguments it would clobber.
3274 SDValue ArgChain = DAG.getStackArgumentTokenFactor(Chain);
3276 SmallVector<SDValue, 8> MemOpChains2;
3279 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
3280 CCValAssign &VA = ArgLocs[i];
3283 assert(VA.isMemLoc());
3284 SDValue Arg = OutVals[i];
3285 ISD::ArgFlagsTy Flags = Outs[i].Flags;
3286 // Skip inalloca arguments. They don't require any work.
3287 if (Flags.isInAlloca())
3289 // Create frame index.
3290 int32_t Offset = VA.getLocMemOffset()+FPDiff;
3291 uint32_t OpSize = (VA.getLocVT().getSizeInBits()+7)/8;
3292 FI = MF.getFrameInfo()->CreateFixedObject(OpSize, Offset, true);
3293 FIN = DAG.getFrameIndex(FI, getPointerTy(DAG.getDataLayout()));
3295 if (Flags.isByVal()) {
3296 // Copy relative to framepointer.
3297 SDValue Source = DAG.getIntPtrConstant(VA.getLocMemOffset(), dl);
3298 if (!StackPtr.getNode())
3299 StackPtr = DAG.getCopyFromReg(Chain, dl, RegInfo->getStackRegister(),
3300 getPointerTy(DAG.getDataLayout()));
3301 Source = DAG.getNode(ISD::ADD, dl, getPointerTy(DAG.getDataLayout()),
3304 MemOpChains2.push_back(CreateCopyOfByValArgument(Source, FIN,
3308 // Store relative to framepointer.
3309 MemOpChains2.push_back(DAG.getStore(
3310 ArgChain, dl, Arg, FIN,
3311 MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), FI),
3316 if (!MemOpChains2.empty())
3317 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, MemOpChains2);
3319 // Store the return address to the appropriate stack slot.
3320 Chain = EmitTailCallStoreRetAddr(DAG, MF, Chain, RetAddrFrIdx,
3321 getPointerTy(DAG.getDataLayout()),
3322 RegInfo->getSlotSize(), FPDiff, dl);
3325 // Build a sequence of copy-to-reg nodes chained together with token chain
3326 // and flag operands which copy the outgoing args into registers.
3328 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
3329 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
3330 RegsToPass[i].second, InFlag);
3331 InFlag = Chain.getValue(1);
3334 if (DAG.getTarget().getCodeModel() == CodeModel::Large) {
3335 assert(Is64Bit && "Large code model is only legal in 64-bit mode.");
3336 // In the 64-bit large code model, we have to make all calls
3337 // through a register, since the call instruction's 32-bit
3338 // pc-relative offset may not be large enough to hold the whole
3340 } else if (Callee->getOpcode() == ISD::GlobalAddress) {
3341 // If the callee is a GlobalAddress node (quite common, every direct call
3342 // is) turn it into a TargetGlobalAddress node so that legalize doesn't hack
3344 GlobalAddressSDNode* G = cast<GlobalAddressSDNode>(Callee);
3346 // We should use extra load for direct calls to dllimported functions in
3348 const GlobalValue *GV = G->getGlobal();
3349 if (!GV->hasDLLImportStorageClass()) {
3350 unsigned char OpFlags = 0;
3351 bool ExtraLoad = false;
3352 unsigned WrapperKind = ISD::DELETED_NODE;
3354 // On ELF targets, in both X86-64 and X86-32 mode, direct calls to
3355 // external symbols most go through the PLT in PIC mode. If the symbol
3356 // has hidden or protected visibility, or if it is static or local, then
3357 // we don't need to use the PLT - we can directly call it.
3358 if (Subtarget->isTargetELF() &&
3359 DAG.getTarget().getRelocationModel() == Reloc::PIC_ &&
3360 GV->hasDefaultVisibility() && !GV->hasLocalLinkage()) {
3361 OpFlags = X86II::MO_PLT;
3362 } else if (Subtarget->isPICStyleStubAny() &&
3363 !GV->isStrongDefinitionForLinker() &&
3364 (!Subtarget->getTargetTriple().isMacOSX() ||
3365 Subtarget->getTargetTriple().isMacOSXVersionLT(10, 5))) {
3366 // PC-relative references to external symbols should go through $stub,
3367 // unless we're building with the leopard linker or later, which
3368 // automatically synthesizes these stubs.
3369 OpFlags = X86II::MO_DARWIN_STUB;
3370 } else if (Subtarget->isPICStyleRIPRel() && isa<Function>(GV) &&
3371 cast<Function>(GV)->hasFnAttribute(Attribute::NonLazyBind)) {
3372 // If the function is marked as non-lazy, generate an indirect call
3373 // which loads from the GOT directly. This avoids runtime overhead
3374 // at the cost of eager binding (and one extra byte of encoding).
3375 OpFlags = X86II::MO_GOTPCREL;
3376 WrapperKind = X86ISD::WrapperRIP;
3380 Callee = DAG.getTargetGlobalAddress(
3381 GV, dl, getPointerTy(DAG.getDataLayout()), G->getOffset(), OpFlags);
3383 // Add a wrapper if needed.
3384 if (WrapperKind != ISD::DELETED_NODE)
3385 Callee = DAG.getNode(X86ISD::WrapperRIP, dl,
3386 getPointerTy(DAG.getDataLayout()), Callee);
3387 // Add extra indirection if needed.
3389 Callee = DAG.getLoad(
3390 getPointerTy(DAG.getDataLayout()), dl, DAG.getEntryNode(), Callee,
3391 MachinePointerInfo::getGOT(DAG.getMachineFunction()), false, false,
3394 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
3395 unsigned char OpFlags = 0;
3397 // On ELF targets, in either X86-64 or X86-32 mode, direct calls to
3398 // external symbols should go through the PLT.
3399 if (Subtarget->isTargetELF() &&
3400 DAG.getTarget().getRelocationModel() == Reloc::PIC_) {
3401 OpFlags = X86II::MO_PLT;
3402 } else if (Subtarget->isPICStyleStubAny() &&
3403 (!Subtarget->getTargetTriple().isMacOSX() ||
3404 Subtarget->getTargetTriple().isMacOSXVersionLT(10, 5))) {
3405 // PC-relative references to external symbols should go through $stub,
3406 // unless we're building with the leopard linker or later, which
3407 // automatically synthesizes these stubs.
3408 OpFlags = X86II::MO_DARWIN_STUB;
3411 Callee = DAG.getTargetExternalSymbol(
3412 S->getSymbol(), getPointerTy(DAG.getDataLayout()), OpFlags);
3413 } else if (Subtarget->isTarget64BitILP32() &&
3414 Callee->getValueType(0) == MVT::i32) {
3415 // Zero-extend the 32-bit Callee address into a 64-bit according to x32 ABI
3416 Callee = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i64, Callee);
3419 // Returns a chain & a flag for retval copy to use.
3420 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
3421 SmallVector<SDValue, 8> Ops;
3423 if (!IsSibcall && isTailCall) {
3424 Chain = DAG.getCALLSEQ_END(Chain,
3425 DAG.getIntPtrConstant(NumBytesToPop, dl, true),
3426 DAG.getIntPtrConstant(0, dl, true), InFlag, dl);
3427 InFlag = Chain.getValue(1);
3430 Ops.push_back(Chain);
3431 Ops.push_back(Callee);
3434 Ops.push_back(DAG.getConstant(FPDiff, dl, MVT::i32));
3436 // Add argument registers to the end of the list so that they are known live
3438 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
3439 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
3440 RegsToPass[i].second.getValueType()));
3442 // Add a register mask operand representing the call-preserved registers.
3443 const uint32_t *Mask = RegInfo->getCallPreservedMask(MF, CallConv);
3444 assert(Mask && "Missing call preserved mask for calling convention");
3446 // If this is an invoke in a 32-bit function using a funclet-based
3447 // personality, assume the function clobbers all registers. If an exception
3448 // is thrown, the runtime will not restore CSRs.
3449 // FIXME: Model this more precisely so that we can register allocate across
3450 // the normal edge and spill and fill across the exceptional edge.
3451 if (!Is64Bit && CLI.CS && CLI.CS->isInvoke()) {
3452 const Function *CallerFn = MF.getFunction();
3453 EHPersonality Pers =
3454 CallerFn->hasPersonalityFn()
3455 ? classifyEHPersonality(CallerFn->getPersonalityFn())
3456 : EHPersonality::Unknown;
3457 if (isFuncletEHPersonality(Pers))
3458 Mask = RegInfo->getNoPreservedMask();
3461 Ops.push_back(DAG.getRegisterMask(Mask));
3463 if (InFlag.getNode())
3464 Ops.push_back(InFlag);
3468 //// If this is the first return lowered for this function, add the regs
3469 //// to the liveout set for the function.
3470 // This isn't right, although it's probably harmless on x86; liveouts
3471 // should be computed from returns not tail calls. Consider a void
3472 // function making a tail call to a function returning int.
3473 MF.getFrameInfo()->setHasTailCall();
3474 return DAG.getNode(X86ISD::TC_RETURN, dl, NodeTys, Ops);
3477 Chain = DAG.getNode(X86ISD::CALL, dl, NodeTys, Ops);
3478 InFlag = Chain.getValue(1);
3480 // Create the CALLSEQ_END node.
3481 unsigned NumBytesForCalleeToPop;
3482 if (X86::isCalleePop(CallConv, Is64Bit, isVarArg,
3483 DAG.getTarget().Options.GuaranteedTailCallOpt))
3484 NumBytesForCalleeToPop = NumBytes; // Callee pops everything
3485 else if (!Is64Bit && !canGuaranteeTCO(CallConv) &&
3486 !Subtarget->getTargetTriple().isOSMSVCRT() &&
3487 SR == StackStructReturn)
3488 // If this is a call to a struct-return function, the callee
3489 // pops the hidden struct pointer, so we have to push it back.
3490 // This is common for Darwin/X86, Linux & Mingw32 targets.
3491 // For MSVC Win32 targets, the caller pops the hidden struct pointer.
3492 NumBytesForCalleeToPop = 4;
3494 NumBytesForCalleeToPop = 0; // Callee pops nothing.
3496 // Returns a flag for retval copy to use.
3498 Chain = DAG.getCALLSEQ_END(Chain,
3499 DAG.getIntPtrConstant(NumBytesToPop, dl, true),
3500 DAG.getIntPtrConstant(NumBytesForCalleeToPop, dl,
3503 InFlag = Chain.getValue(1);
3506 // Handle result values, copying them out of physregs into vregs that we
3508 return LowerCallResult(Chain, InFlag, CallConv, isVarArg,
3509 Ins, dl, DAG, InVals);
3512 //===----------------------------------------------------------------------===//
3513 // Fast Calling Convention (tail call) implementation
3514 //===----------------------------------------------------------------------===//
3516 // Like std call, callee cleans arguments, convention except that ECX is
3517 // reserved for storing the tail called function address. Only 2 registers are
3518 // free for argument passing (inreg). Tail call optimization is performed
3520 // * tailcallopt is enabled
3521 // * caller/callee are fastcc
3522 // On X86_64 architecture with GOT-style position independent code only local
3523 // (within module) calls are supported at the moment.
3524 // To keep the stack aligned according to platform abi the function
3525 // GetAlignedArgumentStackSize ensures that argument delta is always multiples
3526 // of stack alignment. (Dynamic linkers need this - darwin's dyld for example)
3527 // If a tail called function callee has more arguments than the caller the
3528 // caller needs to make sure that there is room to move the RETADDR to. This is
3529 // achieved by reserving an area the size of the argument delta right after the
3530 // original RETADDR, but before the saved framepointer or the spilled registers
3531 // e.g. caller(arg1, arg2) calls callee(arg1, arg2,arg3,arg4)
3543 /// Make the stack size align e.g 16n + 12 aligned for a 16-byte align
3546 X86TargetLowering::GetAlignedArgumentStackSize(unsigned StackSize,
3547 SelectionDAG& DAG) const {
3548 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
3549 const TargetFrameLowering &TFI = *Subtarget->getFrameLowering();
3550 unsigned StackAlignment = TFI.getStackAlignment();
3551 uint64_t AlignMask = StackAlignment - 1;
3552 int64_t Offset = StackSize;
3553 unsigned SlotSize = RegInfo->getSlotSize();
3554 if ( (Offset & AlignMask) <= (StackAlignment - SlotSize) ) {
3555 // Number smaller than 12 so just add the difference.
3556 Offset += ((StackAlignment - SlotSize) - (Offset & AlignMask));
3558 // Mask out lower bits, add stackalignment once plus the 12 bytes.
3559 Offset = ((~AlignMask) & Offset) + StackAlignment +
3560 (StackAlignment-SlotSize);
3565 /// Return true if the given stack call argument is already available in the
3566 /// same position (relatively) of the caller's incoming argument stack.
3568 bool MatchingStackOffset(SDValue Arg, unsigned Offset, ISD::ArgFlagsTy Flags,
3569 MachineFrameInfo *MFI, const MachineRegisterInfo *MRI,
3570 const X86InstrInfo *TII) {
3571 unsigned Bytes = Arg.getValueType().getSizeInBits() / 8;
3573 if (Arg.getOpcode() == ISD::CopyFromReg) {
3574 unsigned VR = cast<RegisterSDNode>(Arg.getOperand(1))->getReg();
3575 if (!TargetRegisterInfo::isVirtualRegister(VR))
3577 MachineInstr *Def = MRI->getVRegDef(VR);
3580 if (!Flags.isByVal()) {
3581 if (!TII->isLoadFromStackSlot(Def, FI))
3584 unsigned Opcode = Def->getOpcode();
3585 if ((Opcode == X86::LEA32r || Opcode == X86::LEA64r ||
3586 Opcode == X86::LEA64_32r) &&
3587 Def->getOperand(1).isFI()) {
3588 FI = Def->getOperand(1).getIndex();
3589 Bytes = Flags.getByValSize();
3593 } else if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Arg)) {
3594 if (Flags.isByVal())
3595 // ByVal argument is passed in as a pointer but it's now being
3596 // dereferenced. e.g.
3597 // define @foo(%struct.X* %A) {
3598 // tail call @bar(%struct.X* byval %A)
3601 SDValue Ptr = Ld->getBasePtr();
3602 FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr);
3605 FI = FINode->getIndex();
3606 } else if (Arg.getOpcode() == ISD::FrameIndex && Flags.isByVal()) {
3607 FrameIndexSDNode *FINode = cast<FrameIndexSDNode>(Arg);
3608 FI = FINode->getIndex();
3609 Bytes = Flags.getByValSize();
3613 assert(FI != INT_MAX);
3614 if (!MFI->isFixedObjectIndex(FI))
3616 return Offset == MFI->getObjectOffset(FI) && Bytes == MFI->getObjectSize(FI);
3619 /// Check whether the call is eligible for tail call optimization. Targets
3620 /// that want to do tail call optimization should implement this function.
3621 bool X86TargetLowering::IsEligibleForTailCallOptimization(
3622 SDValue Callee, CallingConv::ID CalleeCC, bool isVarArg,
3623 bool isCalleeStructRet, bool isCallerStructRet, Type *RetTy,
3624 const SmallVectorImpl<ISD::OutputArg> &Outs,
3625 const SmallVectorImpl<SDValue> &OutVals,
3626 const SmallVectorImpl<ISD::InputArg> &Ins, SelectionDAG &DAG) const {
3627 if (!mayTailCallThisCC(CalleeCC))
3630 // If -tailcallopt is specified, make fastcc functions tail-callable.
3631 MachineFunction &MF = DAG.getMachineFunction();
3632 const Function *CallerF = MF.getFunction();
3634 // If the function return type is x86_fp80 and the callee return type is not,
3635 // then the FP_EXTEND of the call result is not a nop. It's not safe to
3636 // perform a tailcall optimization here.
3637 if (CallerF->getReturnType()->isX86_FP80Ty() && !RetTy->isX86_FP80Ty())
3640 CallingConv::ID CallerCC = CallerF->getCallingConv();
3641 bool CCMatch = CallerCC == CalleeCC;
3642 bool IsCalleeWin64 = Subtarget->isCallingConvWin64(CalleeCC);
3643 bool IsCallerWin64 = Subtarget->isCallingConvWin64(CallerCC);
3645 // Win64 functions have extra shadow space for argument homing. Don't do the
3646 // sibcall if the caller and callee have mismatched expectations for this
3648 if (IsCalleeWin64 != IsCallerWin64)
3651 if (DAG.getTarget().Options.GuaranteedTailCallOpt) {
3652 if (canGuaranteeTCO(CalleeCC) && CCMatch)
3657 // Look for obvious safe cases to perform tail call optimization that do not
3658 // require ABI changes. This is what gcc calls sibcall.
3660 // Can't do sibcall if stack needs to be dynamically re-aligned. PEI needs to
3661 // emit a special epilogue.
3662 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
3663 if (RegInfo->needsStackRealignment(MF))
3666 // Also avoid sibcall optimization if either caller or callee uses struct
3667 // return semantics.
3668 if (isCalleeStructRet || isCallerStructRet)
3671 // Do not sibcall optimize vararg calls unless all arguments are passed via
3673 if (isVarArg && !Outs.empty()) {
3674 // Optimizing for varargs on Win64 is unlikely to be safe without
3675 // additional testing.
3676 if (IsCalleeWin64 || IsCallerWin64)
3679 SmallVector<CCValAssign, 16> ArgLocs;
3680 CCState CCInfo(CalleeCC, isVarArg, DAG.getMachineFunction(), ArgLocs,
3683 CCInfo.AnalyzeCallOperands(Outs, CC_X86);
3684 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i)
3685 if (!ArgLocs[i].isRegLoc())
3689 // If the call result is in ST0 / ST1, it needs to be popped off the x87
3690 // stack. Therefore, if it's not used by the call it is not safe to optimize
3691 // this into a sibcall.
3692 bool Unused = false;
3693 for (unsigned i = 0, e = Ins.size(); i != e; ++i) {
3700 SmallVector<CCValAssign, 16> RVLocs;
3701 CCState CCInfo(CalleeCC, false, DAG.getMachineFunction(), RVLocs,
3703 CCInfo.AnalyzeCallResult(Ins, RetCC_X86);
3704 for (unsigned i = 0, e = RVLocs.size(); i != e; ++i) {
3705 CCValAssign &VA = RVLocs[i];
3706 if (VA.getLocReg() == X86::FP0 || VA.getLocReg() == X86::FP1)
3711 // If the calling conventions do not match, then we'd better make sure the
3712 // results are returned in the same way as what the caller expects.
3714 SmallVector<CCValAssign, 16> RVLocs1;
3715 CCState CCInfo1(CalleeCC, false, DAG.getMachineFunction(), RVLocs1,
3717 CCInfo1.AnalyzeCallResult(Ins, RetCC_X86);
3719 SmallVector<CCValAssign, 16> RVLocs2;
3720 CCState CCInfo2(CallerCC, false, DAG.getMachineFunction(), RVLocs2,
3722 CCInfo2.AnalyzeCallResult(Ins, RetCC_X86);
3724 if (RVLocs1.size() != RVLocs2.size())
3726 for (unsigned i = 0, e = RVLocs1.size(); i != e; ++i) {
3727 if (RVLocs1[i].isRegLoc() != RVLocs2[i].isRegLoc())
3729 if (RVLocs1[i].getLocInfo() != RVLocs2[i].getLocInfo())
3731 if (RVLocs1[i].isRegLoc()) {
3732 if (RVLocs1[i].getLocReg() != RVLocs2[i].getLocReg())
3735 if (RVLocs1[i].getLocMemOffset() != RVLocs2[i].getLocMemOffset())
3741 unsigned StackArgsSize = 0;
3743 // If the callee takes no arguments then go on to check the results of the
3745 if (!Outs.empty()) {
3746 // Check if stack adjustment is needed. For now, do not do this if any
3747 // argument is passed on the stack.
3748 SmallVector<CCValAssign, 16> ArgLocs;
3749 CCState CCInfo(CalleeCC, isVarArg, DAG.getMachineFunction(), ArgLocs,
3752 // Allocate shadow area for Win64
3754 CCInfo.AllocateStack(32, 8);
3756 CCInfo.AnalyzeCallOperands(Outs, CC_X86);
3757 StackArgsSize = CCInfo.getNextStackOffset();
3759 if (CCInfo.getNextStackOffset()) {
3760 // Check if the arguments are already laid out in the right way as
3761 // the caller's fixed stack objects.
3762 MachineFrameInfo *MFI = MF.getFrameInfo();
3763 const MachineRegisterInfo *MRI = &MF.getRegInfo();
3764 const X86InstrInfo *TII = Subtarget->getInstrInfo();
3765 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
3766 CCValAssign &VA = ArgLocs[i];
3767 SDValue Arg = OutVals[i];
3768 ISD::ArgFlagsTy Flags = Outs[i].Flags;
3769 if (VA.getLocInfo() == CCValAssign::Indirect)
3771 if (!VA.isRegLoc()) {
3772 if (!MatchingStackOffset(Arg, VA.getLocMemOffset(), Flags,
3779 // If the tailcall address may be in a register, then make sure it's
3780 // possible to register allocate for it. In 32-bit, the call address can
3781 // only target EAX, EDX, or ECX since the tail call must be scheduled after
3782 // callee-saved registers are restored. These happen to be the same
3783 // registers used to pass 'inreg' arguments so watch out for those.
3784 if (!Subtarget->is64Bit() &&
3785 ((!isa<GlobalAddressSDNode>(Callee) &&
3786 !isa<ExternalSymbolSDNode>(Callee)) ||
3787 DAG.getTarget().getRelocationModel() == Reloc::PIC_)) {
3788 unsigned NumInRegs = 0;
3789 // In PIC we need an extra register to formulate the address computation
3791 unsigned MaxInRegs =
3792 (DAG.getTarget().getRelocationModel() == Reloc::PIC_) ? 2 : 3;
3794 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
3795 CCValAssign &VA = ArgLocs[i];
3798 unsigned Reg = VA.getLocReg();
3801 case X86::EAX: case X86::EDX: case X86::ECX:
3802 if (++NumInRegs == MaxInRegs)
3810 bool CalleeWillPop =
3811 X86::isCalleePop(CalleeCC, Subtarget->is64Bit(), isVarArg,
3812 MF.getTarget().Options.GuaranteedTailCallOpt);
3814 if (unsigned BytesToPop =
3815 MF.getInfo<X86MachineFunctionInfo>()->getBytesToPopOnReturn()) {
3816 // If we have bytes to pop, the callee must pop them.
3817 bool CalleePopMatches = CalleeWillPop && BytesToPop == StackArgsSize;
3818 if (!CalleePopMatches)
3820 } else if (CalleeWillPop && StackArgsSize > 0) {
3821 // If we don't have bytes to pop, make sure the callee doesn't pop any.
3829 X86TargetLowering::createFastISel(FunctionLoweringInfo &funcInfo,
3830 const TargetLibraryInfo *libInfo) const {
3831 return X86::createFastISel(funcInfo, libInfo);
3834 //===----------------------------------------------------------------------===//
3835 // Other Lowering Hooks
3836 //===----------------------------------------------------------------------===//
3838 static bool MayFoldLoad(SDValue Op) {
3839 return Op.hasOneUse() && ISD::isNormalLoad(Op.getNode());
3842 static bool MayFoldIntoStore(SDValue Op) {
3843 return Op.hasOneUse() && ISD::isNormalStore(*Op.getNode()->use_begin());
3846 static bool isTargetShuffle(unsigned Opcode) {
3848 default: return false;
3849 case X86ISD::BLENDI:
3850 case X86ISD::PSHUFB:
3851 case X86ISD::PSHUFD:
3852 case X86ISD::PSHUFHW:
3853 case X86ISD::PSHUFLW:
3855 case X86ISD::PALIGNR:
3856 case X86ISD::MOVLHPS:
3857 case X86ISD::MOVLHPD:
3858 case X86ISD::MOVHLPS:
3859 case X86ISD::MOVLPS:
3860 case X86ISD::MOVLPD:
3861 case X86ISD::MOVSHDUP:
3862 case X86ISD::MOVSLDUP:
3863 case X86ISD::MOVDDUP:
3866 case X86ISD::UNPCKL:
3867 case X86ISD::UNPCKH:
3868 case X86ISD::VPERMILPI:
3869 case X86ISD::VPERM2X128:
3870 case X86ISD::VPERMI:
3871 case X86ISD::VPERMV:
3872 case X86ISD::VPERMV3:
3877 static SDValue getTargetShuffleNode(unsigned Opc, SDLoc dl, EVT VT,
3878 SDValue V1, unsigned TargetMask,
3879 SelectionDAG &DAG) {
3881 default: llvm_unreachable("Unknown x86 shuffle node");
3882 case X86ISD::PSHUFD:
3883 case X86ISD::PSHUFHW:
3884 case X86ISD::PSHUFLW:
3885 case X86ISD::VPERMILPI:
3886 case X86ISD::VPERMI:
3887 return DAG.getNode(Opc, dl, VT, V1,
3888 DAG.getConstant(TargetMask, dl, MVT::i8));
3892 static SDValue getTargetShuffleNode(unsigned Opc, SDLoc dl, EVT VT,
3893 SDValue V1, SDValue V2, SelectionDAG &DAG) {
3895 default: llvm_unreachable("Unknown x86 shuffle node");
3896 case X86ISD::MOVLHPS:
3897 case X86ISD::MOVLHPD:
3898 case X86ISD::MOVHLPS:
3899 case X86ISD::MOVLPS:
3900 case X86ISD::MOVLPD:
3903 case X86ISD::UNPCKL:
3904 case X86ISD::UNPCKH:
3905 return DAG.getNode(Opc, dl, VT, V1, V2);
3909 SDValue X86TargetLowering::getReturnAddressFrameIndex(SelectionDAG &DAG) const {
3910 MachineFunction &MF = DAG.getMachineFunction();
3911 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
3912 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
3913 int ReturnAddrIndex = FuncInfo->getRAIndex();
3915 if (ReturnAddrIndex == 0) {
3916 // Set up a frame object for the return address.
3917 unsigned SlotSize = RegInfo->getSlotSize();
3918 ReturnAddrIndex = MF.getFrameInfo()->CreateFixedObject(SlotSize,
3921 FuncInfo->setRAIndex(ReturnAddrIndex);
3924 return DAG.getFrameIndex(ReturnAddrIndex, getPointerTy(DAG.getDataLayout()));
3927 bool X86::isOffsetSuitableForCodeModel(int64_t Offset, CodeModel::Model M,
3928 bool hasSymbolicDisplacement) {
3929 // Offset should fit into 32 bit immediate field.
3930 if (!isInt<32>(Offset))
3933 // If we don't have a symbolic displacement - we don't have any extra
3935 if (!hasSymbolicDisplacement)
3938 // FIXME: Some tweaks might be needed for medium code model.
3939 if (M != CodeModel::Small && M != CodeModel::Kernel)
3942 // For small code model we assume that latest object is 16MB before end of 31
3943 // bits boundary. We may also accept pretty large negative constants knowing
3944 // that all objects are in the positive half of address space.
3945 if (M == CodeModel::Small && Offset < 16*1024*1024)
3948 // For kernel code model we know that all object resist in the negative half
3949 // of 32bits address space. We may not accept negative offsets, since they may
3950 // be just off and we may accept pretty large positive ones.
3951 if (M == CodeModel::Kernel && Offset >= 0)
3957 /// Determines whether the callee is required to pop its own arguments.
3958 /// Callee pop is necessary to support tail calls.
3959 bool X86::isCalleePop(CallingConv::ID CallingConv,
3960 bool is64Bit, bool IsVarArg, bool GuaranteeTCO) {
3961 // If GuaranteeTCO is true, we force some calls to be callee pop so that we
3962 // can guarantee TCO.
3963 if (!IsVarArg && shouldGuaranteeTCO(CallingConv, GuaranteeTCO))
3966 switch (CallingConv) {
3969 case CallingConv::X86_StdCall:
3970 case CallingConv::X86_FastCall:
3971 case CallingConv::X86_ThisCall:
3972 case CallingConv::X86_VectorCall:
3977 /// \brief Return true if the condition is an unsigned comparison operation.
3978 static bool isX86CCUnsigned(unsigned X86CC) {
3980 default: llvm_unreachable("Invalid integer condition!");
3981 case X86::COND_E: return true;
3982 case X86::COND_G: return false;
3983 case X86::COND_GE: return false;
3984 case X86::COND_L: return false;
3985 case X86::COND_LE: return false;
3986 case X86::COND_NE: return true;
3987 case X86::COND_B: return true;
3988 case X86::COND_A: return true;
3989 case X86::COND_BE: return true;
3990 case X86::COND_AE: return true;
3992 llvm_unreachable("covered switch fell through?!");
3995 /// Do a one-to-one translation of a ISD::CondCode to the X86-specific
3996 /// condition code, returning the condition code and the LHS/RHS of the
3997 /// comparison to make.
3998 static unsigned TranslateX86CC(ISD::CondCode SetCCOpcode, SDLoc DL, bool isFP,
3999 SDValue &LHS, SDValue &RHS, SelectionDAG &DAG) {
4001 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS)) {
4002 if (SetCCOpcode == ISD::SETGT && RHSC->isAllOnesValue()) {
4003 // X > -1 -> X == 0, jump !sign.
4004 RHS = DAG.getConstant(0, DL, RHS.getValueType());
4005 return X86::COND_NS;
4007 if (SetCCOpcode == ISD::SETLT && RHSC->isNullValue()) {
4008 // X < 0 -> X == 0, jump on sign.
4011 if (SetCCOpcode == ISD::SETLT && RHSC->getZExtValue() == 1) {
4013 RHS = DAG.getConstant(0, DL, RHS.getValueType());
4014 return X86::COND_LE;
4018 switch (SetCCOpcode) {
4019 default: llvm_unreachable("Invalid integer condition!");
4020 case ISD::SETEQ: return X86::COND_E;
4021 case ISD::SETGT: return X86::COND_G;
4022 case ISD::SETGE: return X86::COND_GE;
4023 case ISD::SETLT: return X86::COND_L;
4024 case ISD::SETLE: return X86::COND_LE;
4025 case ISD::SETNE: return X86::COND_NE;
4026 case ISD::SETULT: return X86::COND_B;
4027 case ISD::SETUGT: return X86::COND_A;
4028 case ISD::SETULE: return X86::COND_BE;
4029 case ISD::SETUGE: return X86::COND_AE;
4033 // First determine if it is required or is profitable to flip the operands.
4035 // If LHS is a foldable load, but RHS is not, flip the condition.
4036 if (ISD::isNON_EXTLoad(LHS.getNode()) &&
4037 !ISD::isNON_EXTLoad(RHS.getNode())) {
4038 SetCCOpcode = getSetCCSwappedOperands(SetCCOpcode);
4039 std::swap(LHS, RHS);
4042 switch (SetCCOpcode) {
4048 std::swap(LHS, RHS);
4052 // On a floating point condition, the flags are set as follows:
4054 // 0 | 0 | 0 | X > Y
4055 // 0 | 0 | 1 | X < Y
4056 // 1 | 0 | 0 | X == Y
4057 // 1 | 1 | 1 | unordered
4058 switch (SetCCOpcode) {
4059 default: llvm_unreachable("Condcode should be pre-legalized away");
4061 case ISD::SETEQ: return X86::COND_E;
4062 case ISD::SETOLT: // flipped
4064 case ISD::SETGT: return X86::COND_A;
4065 case ISD::SETOLE: // flipped
4067 case ISD::SETGE: return X86::COND_AE;
4068 case ISD::SETUGT: // flipped
4070 case ISD::SETLT: return X86::COND_B;
4071 case ISD::SETUGE: // flipped
4073 case ISD::SETLE: return X86::COND_BE;
4075 case ISD::SETNE: return X86::COND_NE;
4076 case ISD::SETUO: return X86::COND_P;
4077 case ISD::SETO: return X86::COND_NP;
4079 case ISD::SETUNE: return X86::COND_INVALID;
4083 /// Is there a floating point cmov for the specific X86 condition code?
4084 /// Current x86 isa includes the following FP cmov instructions:
4085 /// fcmovb, fcomvbe, fcomve, fcmovu, fcmovae, fcmova, fcmovne, fcmovnu.
4086 static bool hasFPCMov(unsigned X86CC) {
4102 /// Returns true if the target can instruction select the
4103 /// specified FP immediate natively. If false, the legalizer will
4104 /// materialize the FP immediate as a load from a constant pool.
4105 bool X86TargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
4106 for (unsigned i = 0, e = LegalFPImmediates.size(); i != e; ++i) {
4107 if (Imm.bitwiseIsEqual(LegalFPImmediates[i]))
4113 bool X86TargetLowering::shouldReduceLoadWidth(SDNode *Load,
4114 ISD::LoadExtType ExtTy,
4116 // "ELF Handling for Thread-Local Storage" specifies that R_X86_64_GOTTPOFF
4117 // relocation target a movq or addq instruction: don't let the load shrink.
4118 SDValue BasePtr = cast<LoadSDNode>(Load)->getBasePtr();
4119 if (BasePtr.getOpcode() == X86ISD::WrapperRIP)
4120 if (const auto *GA = dyn_cast<GlobalAddressSDNode>(BasePtr.getOperand(0)))
4121 return GA->getTargetFlags() != X86II::MO_GOTTPOFF;
4125 /// \brief Returns true if it is beneficial to convert a load of a constant
4126 /// to just the constant itself.
4127 bool X86TargetLowering::shouldConvertConstantLoadToIntImm(const APInt &Imm,
4129 assert(Ty->isIntegerTy());
4131 unsigned BitSize = Ty->getPrimitiveSizeInBits();
4132 if (BitSize == 0 || BitSize > 64)
4137 bool X86TargetLowering::isExtractSubvectorCheap(EVT ResVT,
4138 unsigned Index) const {
4139 if (!isOperationLegalOrCustom(ISD::EXTRACT_SUBVECTOR, ResVT))
4142 return (Index == 0 || Index == ResVT.getVectorNumElements());
4145 bool X86TargetLowering::isCheapToSpeculateCttz() const {
4146 // Speculate cttz only if we can directly use TZCNT.
4147 return Subtarget->hasBMI();
4150 bool X86TargetLowering::isCheapToSpeculateCtlz() const {
4151 // Speculate ctlz only if we can directly use LZCNT.
4152 return Subtarget->hasLZCNT();
4155 /// Return true if every element in Mask, beginning
4156 /// from position Pos and ending in Pos+Size is undef.
4157 static bool isUndefInRange(ArrayRef<int> Mask, unsigned Pos, unsigned Size) {
4158 for (unsigned i = Pos, e = Pos + Size; i != e; ++i)
4164 /// Return true if Val is undef or if its value falls within the
4165 /// specified range (L, H].
4166 static bool isUndefOrInRange(int Val, int Low, int Hi) {
4167 return (Val < 0) || (Val >= Low && Val < Hi);
4170 /// Val is either less than zero (undef) or equal to the specified value.
4171 static bool isUndefOrEqual(int Val, int CmpVal) {
4172 return (Val < 0 || Val == CmpVal);
4175 /// Return true if every element in Mask, beginning
4176 /// from position Pos and ending in Pos+Size, falls within the specified
4177 /// sequential range (Low, Low+Size]. or is undef.
4178 static bool isSequentialOrUndefInRange(ArrayRef<int> Mask,
4179 unsigned Pos, unsigned Size, int Low) {
4180 for (unsigned i = Pos, e = Pos+Size; i != e; ++i, ++Low)
4181 if (!isUndefOrEqual(Mask[i], Low))
4186 /// Return true if the specified EXTRACT_SUBVECTOR operand specifies a vector
4187 /// extract that is suitable for instruction that extract 128 or 256 bit vectors
4188 static bool isVEXTRACTIndex(SDNode *N, unsigned vecWidth) {
4189 assert((vecWidth == 128 || vecWidth == 256) && "Unexpected vector width");
4190 if (!isa<ConstantSDNode>(N->getOperand(1).getNode()))
4193 // The index should be aligned on a vecWidth-bit boundary.
4195 cast<ConstantSDNode>(N->getOperand(1).getNode())->getZExtValue();
4197 MVT VT = N->getSimpleValueType(0);
4198 unsigned ElSize = VT.getVectorElementType().getSizeInBits();
4199 bool Result = (Index * ElSize) % vecWidth == 0;
4204 /// Return true if the specified INSERT_SUBVECTOR
4205 /// operand specifies a subvector insert that is suitable for input to
4206 /// insertion of 128 or 256-bit subvectors
4207 static bool isVINSERTIndex(SDNode *N, unsigned vecWidth) {
4208 assert((vecWidth == 128 || vecWidth == 256) && "Unexpected vector width");
4209 if (!isa<ConstantSDNode>(N->getOperand(2).getNode()))
4211 // The index should be aligned on a vecWidth-bit boundary.
4213 cast<ConstantSDNode>(N->getOperand(2).getNode())->getZExtValue();
4215 MVT VT = N->getSimpleValueType(0);
4216 unsigned ElSize = VT.getVectorElementType().getSizeInBits();
4217 bool Result = (Index * ElSize) % vecWidth == 0;
4222 bool X86::isVINSERT128Index(SDNode *N) {
4223 return isVINSERTIndex(N, 128);
4226 bool X86::isVINSERT256Index(SDNode *N) {
4227 return isVINSERTIndex(N, 256);
4230 bool X86::isVEXTRACT128Index(SDNode *N) {
4231 return isVEXTRACTIndex(N, 128);
4234 bool X86::isVEXTRACT256Index(SDNode *N) {
4235 return isVEXTRACTIndex(N, 256);
4238 static unsigned getExtractVEXTRACTImmediate(SDNode *N, unsigned vecWidth) {
4239 assert((vecWidth == 128 || vecWidth == 256) && "Unsupported vector width");
4240 if (!isa<ConstantSDNode>(N->getOperand(1).getNode()))
4241 llvm_unreachable("Illegal extract subvector for VEXTRACT");
4244 cast<ConstantSDNode>(N->getOperand(1).getNode())->getZExtValue();
4246 MVT VecVT = N->getOperand(0).getSimpleValueType();
4247 MVT ElVT = VecVT.getVectorElementType();
4249 unsigned NumElemsPerChunk = vecWidth / ElVT.getSizeInBits();
4250 return Index / NumElemsPerChunk;
4253 static unsigned getInsertVINSERTImmediate(SDNode *N, unsigned vecWidth) {
4254 assert((vecWidth == 128 || vecWidth == 256) && "Unsupported vector width");
4255 if (!isa<ConstantSDNode>(N->getOperand(2).getNode()))
4256 llvm_unreachable("Illegal insert subvector for VINSERT");
4259 cast<ConstantSDNode>(N->getOperand(2).getNode())->getZExtValue();
4261 MVT VecVT = N->getSimpleValueType(0);
4262 MVT ElVT = VecVT.getVectorElementType();
4264 unsigned NumElemsPerChunk = vecWidth / ElVT.getSizeInBits();
4265 return Index / NumElemsPerChunk;
4268 /// Return the appropriate immediate to extract the specified
4269 /// EXTRACT_SUBVECTOR index with VEXTRACTF128 and VINSERTI128 instructions.
4270 unsigned X86::getExtractVEXTRACT128Immediate(SDNode *N) {
4271 return getExtractVEXTRACTImmediate(N, 128);
4274 /// Return the appropriate immediate to extract the specified
4275 /// EXTRACT_SUBVECTOR index with VEXTRACTF64x4 and VINSERTI64x4 instructions.
4276 unsigned X86::getExtractVEXTRACT256Immediate(SDNode *N) {
4277 return getExtractVEXTRACTImmediate(N, 256);
4280 /// Return the appropriate immediate to insert at the specified
4281 /// INSERT_SUBVECTOR index with VINSERTF128 and VINSERTI128 instructions.
4282 unsigned X86::getInsertVINSERT128Immediate(SDNode *N) {
4283 return getInsertVINSERTImmediate(N, 128);
4286 /// Return the appropriate immediate to insert at the specified
4287 /// INSERT_SUBVECTOR index with VINSERTF46x4 and VINSERTI64x4 instructions.
4288 unsigned X86::getInsertVINSERT256Immediate(SDNode *N) {
4289 return getInsertVINSERTImmediate(N, 256);
4292 /// Returns true if V is a constant integer zero.
4293 static bool isZero(SDValue V) {
4294 ConstantSDNode *C = dyn_cast<ConstantSDNode>(V);
4295 return C && C->isNullValue();
4298 /// Returns true if Elt is a constant zero or a floating point constant +0.0.
4299 bool X86::isZeroNode(SDValue Elt) {
4302 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(Elt))
4303 return CFP->getValueAPF().isPosZero();
4307 // Build a vector of constants
4308 // Use an UNDEF node if MaskElt == -1.
4309 // Spilt 64-bit constants in the 32-bit mode.
4310 static SDValue getConstVector(ArrayRef<int> Values, EVT VT,
4312 SDLoc dl, bool IsMask = false) {
4314 SmallVector<SDValue, 32> Ops;
4317 EVT ConstVecVT = VT;
4318 unsigned NumElts = VT.getVectorNumElements();
4319 bool In64BitMode = DAG.getTargetLoweringInfo().isTypeLegal(MVT::i64);
4320 if (!In64BitMode && VT.getScalarType() == MVT::i64) {
4321 ConstVecVT = MVT::getVectorVT(MVT::i32, NumElts * 2);
4325 EVT EltVT = ConstVecVT.getScalarType();
4326 for (unsigned i = 0; i < NumElts; ++i) {
4327 bool IsUndef = Values[i] < 0 && IsMask;
4328 SDValue OpNode = IsUndef ? DAG.getUNDEF(EltVT) :
4329 DAG.getConstant(Values[i], dl, EltVT);
4330 Ops.push_back(OpNode);
4332 Ops.push_back(IsUndef ? DAG.getUNDEF(EltVT) :
4333 DAG.getConstant(0, dl, EltVT));
4335 SDValue ConstsNode = DAG.getNode(ISD::BUILD_VECTOR, dl, ConstVecVT, Ops);
4337 ConstsNode = DAG.getBitcast(VT, ConstsNode);
4341 /// Returns a vector of specified type with all zero elements.
4342 static SDValue getZeroVector(EVT VT, const X86Subtarget *Subtarget,
4343 SelectionDAG &DAG, SDLoc dl) {
4344 assert(VT.isVector() && "Expected a vector type");
4346 // Always build SSE zero vectors as <4 x i32> bitcasted
4347 // to their dest type. This ensures they get CSE'd.
4349 if (VT.is128BitVector()) { // SSE
4350 if (Subtarget->hasSSE2()) { // SSE2
4351 SDValue Cst = DAG.getConstant(0, dl, MVT::i32);
4352 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
4354 SDValue Cst = DAG.getConstantFP(+0.0, dl, MVT::f32);
4355 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4f32, Cst, Cst, Cst, Cst);
4357 } else if (VT.is256BitVector()) { // AVX
4358 if (Subtarget->hasInt256()) { // AVX2
4359 SDValue Cst = DAG.getConstant(0, dl, MVT::i32);
4360 SDValue Ops[] = { Cst, Cst, Cst, Cst, Cst, Cst, Cst, Cst };
4361 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v8i32, Ops);
4363 // 256-bit logic and arithmetic instructions in AVX are all
4364 // floating-point, no support for integer ops. Emit fp zeroed vectors.
4365 SDValue Cst = DAG.getConstantFP(+0.0, dl, MVT::f32);
4366 SDValue Ops[] = { Cst, Cst, Cst, Cst, Cst, Cst, Cst, Cst };
4367 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v8f32, Ops);
4369 } else if (VT.is512BitVector()) { // AVX-512
4370 SDValue Cst = DAG.getConstant(0, dl, MVT::i32);
4371 SDValue Ops[] = { Cst, Cst, Cst, Cst, Cst, Cst, Cst, Cst,
4372 Cst, Cst, Cst, Cst, Cst, Cst, Cst, Cst };
4373 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v16i32, Ops);
4374 } else if (VT.getScalarType() == MVT::i1) {
4376 assert((Subtarget->hasBWI() || VT.getVectorNumElements() <= 16)
4377 && "Unexpected vector type");
4378 assert((Subtarget->hasVLX() || VT.getVectorNumElements() >= 8)
4379 && "Unexpected vector type");
4380 SDValue Cst = DAG.getConstant(0, dl, MVT::i1);
4381 SmallVector<SDValue, 64> Ops(VT.getVectorNumElements(), Cst);
4382 return DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Ops);
4384 llvm_unreachable("Unexpected vector type");
4386 return DAG.getBitcast(VT, Vec);
4389 static SDValue ExtractSubVector(SDValue Vec, unsigned IdxVal,
4390 SelectionDAG &DAG, SDLoc dl,
4391 unsigned vectorWidth) {
4392 assert((vectorWidth == 128 || vectorWidth == 256) &&
4393 "Unsupported vector width");
4394 EVT VT = Vec.getValueType();
4395 EVT ElVT = VT.getVectorElementType();
4396 unsigned Factor = VT.getSizeInBits()/vectorWidth;
4397 EVT ResultVT = EVT::getVectorVT(*DAG.getContext(), ElVT,
4398 VT.getVectorNumElements()/Factor);
4400 // Extract from UNDEF is UNDEF.
4401 if (Vec.getOpcode() == ISD::UNDEF)
4402 return DAG.getUNDEF(ResultVT);
4404 // Extract the relevant vectorWidth bits. Generate an EXTRACT_SUBVECTOR
4405 unsigned ElemsPerChunk = vectorWidth / ElVT.getSizeInBits();
4407 // This is the index of the first element of the vectorWidth-bit chunk
4409 unsigned NormalizedIdxVal = (((IdxVal * ElVT.getSizeInBits()) / vectorWidth)
4412 // If the input is a buildvector just emit a smaller one.
4413 if (Vec.getOpcode() == ISD::BUILD_VECTOR)
4414 return DAG.getNode(ISD::BUILD_VECTOR, dl, ResultVT,
4415 makeArrayRef(Vec->op_begin() + NormalizedIdxVal,
4418 SDValue VecIdx = DAG.getIntPtrConstant(NormalizedIdxVal, dl);
4419 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, ResultVT, Vec, VecIdx);
4422 /// Generate a DAG to grab 128-bits from a vector > 128 bits. This
4423 /// sets things up to match to an AVX VEXTRACTF128 / VEXTRACTI128
4424 /// or AVX-512 VEXTRACTF32x4 / VEXTRACTI32x4
4425 /// instructions or a simple subregister reference. Idx is an index in the
4426 /// 128 bits we want. It need not be aligned to a 128-bit boundary. That makes
4427 /// lowering EXTRACT_VECTOR_ELT operations easier.
4428 static SDValue Extract128BitVector(SDValue Vec, unsigned IdxVal,
4429 SelectionDAG &DAG, SDLoc dl) {
4430 assert((Vec.getValueType().is256BitVector() ||
4431 Vec.getValueType().is512BitVector()) && "Unexpected vector size!");
4432 return ExtractSubVector(Vec, IdxVal, DAG, dl, 128);
4435 /// Generate a DAG to grab 256-bits from a 512-bit vector.
4436 static SDValue Extract256BitVector(SDValue Vec, unsigned IdxVal,
4437 SelectionDAG &DAG, SDLoc dl) {
4438 assert(Vec.getValueType().is512BitVector() && "Unexpected vector size!");
4439 return ExtractSubVector(Vec, IdxVal, DAG, dl, 256);
4442 static SDValue InsertSubVector(SDValue Result, SDValue Vec,
4443 unsigned IdxVal, SelectionDAG &DAG,
4444 SDLoc dl, unsigned vectorWidth) {
4445 assert((vectorWidth == 128 || vectorWidth == 256) &&
4446 "Unsupported vector width");
4447 // Inserting UNDEF is Result
4448 if (Vec.getOpcode() == ISD::UNDEF)
4450 EVT VT = Vec.getValueType();
4451 EVT ElVT = VT.getVectorElementType();
4452 EVT ResultVT = Result.getValueType();
4454 // Insert the relevant vectorWidth bits.
4455 unsigned ElemsPerChunk = vectorWidth/ElVT.getSizeInBits();
4457 // This is the index of the first element of the vectorWidth-bit chunk
4459 unsigned NormalizedIdxVal = (((IdxVal * ElVT.getSizeInBits())/vectorWidth)
4462 SDValue VecIdx = DAG.getIntPtrConstant(NormalizedIdxVal, dl);
4463 return DAG.getNode(ISD::INSERT_SUBVECTOR, dl, ResultVT, Result, Vec, VecIdx);
4466 /// Generate a DAG to put 128-bits into a vector > 128 bits. This
4467 /// sets things up to match to an AVX VINSERTF128/VINSERTI128 or
4468 /// AVX-512 VINSERTF32x4/VINSERTI32x4 instructions or a
4469 /// simple superregister reference. Idx is an index in the 128 bits
4470 /// we want. It need not be aligned to a 128-bit boundary. That makes
4471 /// lowering INSERT_VECTOR_ELT operations easier.
4472 static SDValue Insert128BitVector(SDValue Result, SDValue Vec, unsigned IdxVal,
4473 SelectionDAG &DAG, SDLoc dl) {
4474 assert(Vec.getValueType().is128BitVector() && "Unexpected vector size!");
4476 // For insertion into the zero index (low half) of a 256-bit vector, it is
4477 // more efficient to generate a blend with immediate instead of an insert*128.
4478 // We are still creating an INSERT_SUBVECTOR below with an undef node to
4479 // extend the subvector to the size of the result vector. Make sure that
4480 // we are not recursing on that node by checking for undef here.
4481 if (IdxVal == 0 && Result.getValueType().is256BitVector() &&
4482 Result.getOpcode() != ISD::UNDEF) {
4483 EVT ResultVT = Result.getValueType();
4484 SDValue ZeroIndex = DAG.getIntPtrConstant(0, dl);
4485 SDValue Undef = DAG.getUNDEF(ResultVT);
4486 SDValue Vec256 = DAG.getNode(ISD::INSERT_SUBVECTOR, dl, ResultVT, Undef,
4489 // The blend instruction, and therefore its mask, depend on the data type.
4490 MVT ScalarType = ResultVT.getScalarType().getSimpleVT();
4491 if (ScalarType.isFloatingPoint()) {
4492 // Choose either vblendps (float) or vblendpd (double).
4493 unsigned ScalarSize = ScalarType.getSizeInBits();
4494 assert((ScalarSize == 64 || ScalarSize == 32) && "Unknown float type");
4495 unsigned MaskVal = (ScalarSize == 64) ? 0x03 : 0x0f;
4496 SDValue Mask = DAG.getConstant(MaskVal, dl, MVT::i8);
4497 return DAG.getNode(X86ISD::BLENDI, dl, ResultVT, Result, Vec256, Mask);
4500 const X86Subtarget &Subtarget =
4501 static_cast<const X86Subtarget &>(DAG.getSubtarget());
4503 // AVX2 is needed for 256-bit integer blend support.
4504 // Integers must be cast to 32-bit because there is only vpblendd;
4505 // vpblendw can't be used for this because it has a handicapped mask.
4507 // If we don't have AVX2, then cast to float. Using a wrong domain blend
4508 // is still more efficient than using the wrong domain vinsertf128 that
4509 // will be created by InsertSubVector().
4510 MVT CastVT = Subtarget.hasAVX2() ? MVT::v8i32 : MVT::v8f32;
4512 SDValue Mask = DAG.getConstant(0x0f, dl, MVT::i8);
4513 Vec256 = DAG.getBitcast(CastVT, Vec256);
4514 Vec256 = DAG.getNode(X86ISD::BLENDI, dl, CastVT, Result, Vec256, Mask);
4515 return DAG.getBitcast(ResultVT, Vec256);
4518 return InsertSubVector(Result, Vec, IdxVal, DAG, dl, 128);
4521 static SDValue Insert256BitVector(SDValue Result, SDValue Vec, unsigned IdxVal,
4522 SelectionDAG &DAG, SDLoc dl) {
4523 assert(Vec.getValueType().is256BitVector() && "Unexpected vector size!");
4524 return InsertSubVector(Result, Vec, IdxVal, DAG, dl, 256);
4527 /// Concat two 128-bit vectors into a 256 bit vector using VINSERTF128
4528 /// instructions. This is used because creating CONCAT_VECTOR nodes of
4529 /// BUILD_VECTORS returns a larger BUILD_VECTOR while we're trying to lower
4530 /// large BUILD_VECTORS.
4531 static SDValue Concat128BitVectors(SDValue V1, SDValue V2, EVT VT,
4532 unsigned NumElems, SelectionDAG &DAG,
4534 SDValue V = Insert128BitVector(DAG.getUNDEF(VT), V1, 0, DAG, dl);
4535 return Insert128BitVector(V, V2, NumElems/2, DAG, dl);
4538 static SDValue Concat256BitVectors(SDValue V1, SDValue V2, EVT VT,
4539 unsigned NumElems, SelectionDAG &DAG,
4541 SDValue V = Insert256BitVector(DAG.getUNDEF(VT), V1, 0, DAG, dl);
4542 return Insert256BitVector(V, V2, NumElems/2, DAG, dl);
4545 /// Returns a vector of specified type with all bits set.
4546 /// Always build ones vectors as <4 x i32> or <8 x i32>. For 256-bit types with
4547 /// no AVX2 supprt, use two <4 x i32> inserted in a <8 x i32> appropriately.
4548 /// Then bitcast to their original type, ensuring they get CSE'd.
4549 static SDValue getOnesVector(EVT VT, const X86Subtarget *Subtarget,
4550 SelectionDAG &DAG, SDLoc dl) {
4551 assert(VT.isVector() && "Expected a vector type");
4553 SDValue Cst = DAG.getConstant(~0U, dl, MVT::i32);
4555 if (VT.is512BitVector()) {
4556 SDValue Ops[] = { Cst, Cst, Cst, Cst, Cst, Cst, Cst, Cst,
4557 Cst, Cst, Cst, Cst, Cst, Cst, Cst, Cst };
4558 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v16i32, Ops);
4559 } else if (VT.is256BitVector()) {
4560 if (Subtarget->hasInt256()) { // AVX2
4561 SDValue Ops[] = { Cst, Cst, Cst, Cst, Cst, Cst, Cst, Cst };
4562 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v8i32, Ops);
4564 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
4565 Vec = Concat128BitVectors(Vec, Vec, MVT::v8i32, 8, DAG, dl);
4567 } else if (VT.is128BitVector()) {
4568 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
4570 llvm_unreachable("Unexpected vector type");
4572 return DAG.getBitcast(VT, Vec);
4575 /// Returns a vector_shuffle node for an unpackl operation.
4576 static SDValue getUnpackl(SelectionDAG &DAG, SDLoc dl, MVT VT, SDValue V1,
4578 unsigned NumElems = VT.getVectorNumElements();
4579 SmallVector<int, 8> Mask;
4580 for (unsigned i = 0, e = NumElems/2; i != e; ++i) {
4582 Mask.push_back(i + NumElems);
4584 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
4587 /// Returns a vector_shuffle node for an unpackh operation.
4588 static SDValue getUnpackh(SelectionDAG &DAG, SDLoc dl, MVT VT, SDValue V1,
4590 unsigned NumElems = VT.getVectorNumElements();
4591 SmallVector<int, 8> Mask;
4592 for (unsigned i = 0, Half = NumElems/2; i != Half; ++i) {
4593 Mask.push_back(i + Half);
4594 Mask.push_back(i + NumElems + Half);
4596 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
4599 /// Return a vector_shuffle of the specified vector of zero or undef vector.
4600 /// This produces a shuffle where the low element of V2 is swizzled into the
4601 /// zero/undef vector, landing at element Idx.
4602 /// This produces a shuffle mask like 4,1,2,3 (idx=0) or 0,1,2,4 (idx=3).
4603 static SDValue getShuffleVectorZeroOrUndef(SDValue V2, unsigned Idx,
4605 const X86Subtarget *Subtarget,
4606 SelectionDAG &DAG) {
4607 MVT VT = V2.getSimpleValueType();
4609 ? getZeroVector(VT, Subtarget, DAG, SDLoc(V2)) : DAG.getUNDEF(VT);
4610 unsigned NumElems = VT.getVectorNumElements();
4611 SmallVector<int, 16> MaskVec;
4612 for (unsigned i = 0; i != NumElems; ++i)
4613 // If this is the insertion idx, put the low elt of V2 here.
4614 MaskVec.push_back(i == Idx ? NumElems : i);
4615 return DAG.getVectorShuffle(VT, SDLoc(V2), V1, V2, &MaskVec[0]);
4618 /// Calculates the shuffle mask corresponding to the target-specific opcode.
4619 /// Returns true if the Mask could be calculated. Sets IsUnary to true if only
4620 /// uses one source. Note that this will set IsUnary for shuffles which use a
4621 /// single input multiple times, and in those cases it will
4622 /// adjust the mask to only have indices within that single input.
4623 /// FIXME: Add support for Decode*Mask functions that return SM_SentinelZero.
4624 static bool getTargetShuffleMask(SDNode *N, MVT VT,
4625 SmallVectorImpl<int> &Mask, bool &IsUnary) {
4626 unsigned NumElems = VT.getVectorNumElements();
4630 bool IsFakeUnary = false;
4631 switch(N->getOpcode()) {
4632 case X86ISD::BLENDI:
4633 ImmN = N->getOperand(N->getNumOperands()-1);
4634 DecodeBLENDMask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
4637 ImmN = N->getOperand(N->getNumOperands()-1);
4638 DecodeSHUFPMask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
4639 IsUnary = IsFakeUnary = N->getOperand(0) == N->getOperand(1);
4641 case X86ISD::UNPCKH:
4642 DecodeUNPCKHMask(VT, Mask);
4643 IsUnary = IsFakeUnary = N->getOperand(0) == N->getOperand(1);
4645 case X86ISD::UNPCKL:
4646 DecodeUNPCKLMask(VT, Mask);
4647 IsUnary = IsFakeUnary = N->getOperand(0) == N->getOperand(1);
4649 case X86ISD::MOVHLPS:
4650 DecodeMOVHLPSMask(NumElems, Mask);
4651 IsUnary = IsFakeUnary = N->getOperand(0) == N->getOperand(1);
4653 case X86ISD::MOVLHPS:
4654 DecodeMOVLHPSMask(NumElems, Mask);
4655 IsUnary = IsFakeUnary = N->getOperand(0) == N->getOperand(1);
4657 case X86ISD::PALIGNR:
4658 ImmN = N->getOperand(N->getNumOperands()-1);
4659 DecodePALIGNRMask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
4661 case X86ISD::PSHUFD:
4662 case X86ISD::VPERMILPI:
4663 ImmN = N->getOperand(N->getNumOperands()-1);
4664 DecodePSHUFMask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
4667 case X86ISD::PSHUFHW:
4668 ImmN = N->getOperand(N->getNumOperands()-1);
4669 DecodePSHUFHWMask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
4672 case X86ISD::PSHUFLW:
4673 ImmN = N->getOperand(N->getNumOperands()-1);
4674 DecodePSHUFLWMask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
4677 case X86ISD::PSHUFB: {
4679 SDValue MaskNode = N->getOperand(1);
4680 while (MaskNode->getOpcode() == ISD::BITCAST)
4681 MaskNode = MaskNode->getOperand(0);
4683 if (MaskNode->getOpcode() == ISD::BUILD_VECTOR) {
4684 // If we have a build-vector, then things are easy.
4685 EVT VT = MaskNode.getValueType();
4686 assert(VT.isVector() &&
4687 "Can't produce a non-vector with a build_vector!");
4688 if (!VT.isInteger())
4691 int NumBytesPerElement = VT.getVectorElementType().getSizeInBits() / 8;
4693 SmallVector<uint64_t, 32> RawMask;
4694 for (int i = 0, e = MaskNode->getNumOperands(); i < e; ++i) {
4695 SDValue Op = MaskNode->getOperand(i);
4696 if (Op->getOpcode() == ISD::UNDEF) {
4697 RawMask.push_back((uint64_t)SM_SentinelUndef);
4700 auto *CN = dyn_cast<ConstantSDNode>(Op.getNode());
4703 APInt MaskElement = CN->getAPIntValue();
4705 // We now have to decode the element which could be any integer size and
4706 // extract each byte of it.
4707 for (int j = 0; j < NumBytesPerElement; ++j) {
4708 // Note that this is x86 and so always little endian: the low byte is
4709 // the first byte of the mask.
4710 RawMask.push_back(MaskElement.getLoBits(8).getZExtValue());
4711 MaskElement = MaskElement.lshr(8);
4714 DecodePSHUFBMask(RawMask, Mask);
4718 auto *MaskLoad = dyn_cast<LoadSDNode>(MaskNode);
4722 SDValue Ptr = MaskLoad->getBasePtr();
4723 if (Ptr->getOpcode() == X86ISD::Wrapper ||
4724 Ptr->getOpcode() == X86ISD::WrapperRIP)
4725 Ptr = Ptr->getOperand(0);
4727 auto *MaskCP = dyn_cast<ConstantPoolSDNode>(Ptr);
4728 if (!MaskCP || MaskCP->isMachineConstantPoolEntry())
4731 if (auto *C = dyn_cast<Constant>(MaskCP->getConstVal())) {
4732 DecodePSHUFBMask(C, Mask);
4740 case X86ISD::VPERMI:
4741 ImmN = N->getOperand(N->getNumOperands()-1);
4742 DecodeVPERMMask(cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
4747 DecodeScalarMoveMask(VT, /* IsLoad */ false, Mask);
4749 case X86ISD::VPERM2X128:
4750 ImmN = N->getOperand(N->getNumOperands()-1);
4751 DecodeVPERM2X128Mask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
4752 if (Mask.empty()) return false;
4753 // Mask only contains negative index if an element is zero.
4754 if (std::any_of(Mask.begin(), Mask.end(),
4755 [](int M){ return M == SM_SentinelZero; }))
4758 case X86ISD::MOVSLDUP:
4759 DecodeMOVSLDUPMask(VT, Mask);
4762 case X86ISD::MOVSHDUP:
4763 DecodeMOVSHDUPMask(VT, Mask);
4766 case X86ISD::MOVDDUP:
4767 DecodeMOVDDUPMask(VT, Mask);
4770 case X86ISD::MOVLHPD:
4771 case X86ISD::MOVLPD:
4772 case X86ISD::MOVLPS:
4773 // Not yet implemented
4775 case X86ISD::VPERMV: {
4777 SDValue MaskNode = N->getOperand(0);
4778 while (MaskNode->getOpcode() == ISD::BITCAST)
4779 MaskNode = MaskNode->getOperand(0);
4781 unsigned MaskLoBits = Log2_64(VT.getVectorNumElements());
4782 SmallVector<uint64_t, 32> RawMask;
4783 if (MaskNode->getOpcode() == ISD::BUILD_VECTOR) {
4784 // If we have a build-vector, then things are easy.
4785 assert(MaskNode.getValueType().isInteger() &&
4786 MaskNode.getValueType().getVectorNumElements() ==
4787 VT.getVectorNumElements());
4789 for (unsigned i = 0; i < MaskNode->getNumOperands(); ++i) {
4790 SDValue Op = MaskNode->getOperand(i);
4791 if (Op->getOpcode() == ISD::UNDEF)
4792 RawMask.push_back((uint64_t)SM_SentinelUndef);
4793 else if (isa<ConstantSDNode>(Op)) {
4794 APInt MaskElement = cast<ConstantSDNode>(Op)->getAPIntValue();
4795 RawMask.push_back(MaskElement.getLoBits(MaskLoBits).getZExtValue());
4799 DecodeVPERMVMask(RawMask, Mask);
4802 if (MaskNode->getOpcode() == X86ISD::VBROADCAST) {
4803 unsigned NumEltsInMask = MaskNode->getNumOperands();
4804 MaskNode = MaskNode->getOperand(0);
4805 auto *CN = dyn_cast<ConstantSDNode>(MaskNode);
4807 APInt MaskEltValue = CN->getAPIntValue();
4808 for (unsigned i = 0; i < NumEltsInMask; ++i)
4809 RawMask.push_back(MaskEltValue.getLoBits(MaskLoBits).getZExtValue());
4810 DecodeVPERMVMask(RawMask, Mask);
4813 // It may be a scalar load
4816 auto *MaskLoad = dyn_cast<LoadSDNode>(MaskNode);
4820 SDValue Ptr = MaskLoad->getBasePtr();
4821 if (Ptr->getOpcode() == X86ISD::Wrapper ||
4822 Ptr->getOpcode() == X86ISD::WrapperRIP)
4823 Ptr = Ptr->getOperand(0);
4825 auto *MaskCP = dyn_cast<ConstantPoolSDNode>(Ptr);
4826 if (!MaskCP || MaskCP->isMachineConstantPoolEntry())
4829 auto *C = dyn_cast<Constant>(MaskCP->getConstVal());
4831 DecodeVPERMVMask(C, VT, Mask);
4838 case X86ISD::VPERMV3: {
4840 SDValue MaskNode = N->getOperand(1);
4841 while (MaskNode->getOpcode() == ISD::BITCAST)
4842 MaskNode = MaskNode->getOperand(1);
4844 if (MaskNode->getOpcode() == ISD::BUILD_VECTOR) {
4845 // If we have a build-vector, then things are easy.
4846 assert(MaskNode.getValueType().isInteger() &&
4847 MaskNode.getValueType().getVectorNumElements() ==
4848 VT.getVectorNumElements());
4850 SmallVector<uint64_t, 32> RawMask;
4851 unsigned MaskLoBits = Log2_64(VT.getVectorNumElements()*2);
4853 for (unsigned i = 0; i < MaskNode->getNumOperands(); ++i) {
4854 SDValue Op = MaskNode->getOperand(i);
4855 if (Op->getOpcode() == ISD::UNDEF)
4856 RawMask.push_back((uint64_t)SM_SentinelUndef);
4858 auto *CN = dyn_cast<ConstantSDNode>(Op.getNode());
4861 APInt MaskElement = CN->getAPIntValue();
4862 RawMask.push_back(MaskElement.getLoBits(MaskLoBits).getZExtValue());
4865 DecodeVPERMV3Mask(RawMask, Mask);
4869 auto *MaskLoad = dyn_cast<LoadSDNode>(MaskNode);
4873 SDValue Ptr = MaskLoad->getBasePtr();
4874 if (Ptr->getOpcode() == X86ISD::Wrapper ||
4875 Ptr->getOpcode() == X86ISD::WrapperRIP)
4876 Ptr = Ptr->getOperand(0);
4878 auto *MaskCP = dyn_cast<ConstantPoolSDNode>(Ptr);
4879 if (!MaskCP || MaskCP->isMachineConstantPoolEntry())
4882 auto *C = dyn_cast<Constant>(MaskCP->getConstVal());
4884 DecodeVPERMV3Mask(C, VT, Mask);
4891 default: llvm_unreachable("unknown target shuffle node");
4894 // If we have a fake unary shuffle, the shuffle mask is spread across two
4895 // inputs that are actually the same node. Re-map the mask to always point
4896 // into the first input.
4899 if (M >= (int)Mask.size())
4905 /// Returns the scalar element that will make up the ith
4906 /// element of the result of the vector shuffle.
4907 static SDValue getShuffleScalarElt(SDNode *N, unsigned Index, SelectionDAG &DAG,
4910 return SDValue(); // Limit search depth.
4912 SDValue V = SDValue(N, 0);
4913 EVT VT = V.getValueType();
4914 unsigned Opcode = V.getOpcode();
4916 // Recurse into ISD::VECTOR_SHUFFLE node to find scalars.
4917 if (const ShuffleVectorSDNode *SV = dyn_cast<ShuffleVectorSDNode>(N)) {
4918 int Elt = SV->getMaskElt(Index);
4921 return DAG.getUNDEF(VT.getVectorElementType());
4923 unsigned NumElems = VT.getVectorNumElements();
4924 SDValue NewV = (Elt < (int)NumElems) ? SV->getOperand(0)
4925 : SV->getOperand(1);
4926 return getShuffleScalarElt(NewV.getNode(), Elt % NumElems, DAG, Depth+1);
4929 // Recurse into target specific vector shuffles to find scalars.
4930 if (isTargetShuffle(Opcode)) {
4931 MVT ShufVT = V.getSimpleValueType();
4932 unsigned NumElems = ShufVT.getVectorNumElements();
4933 SmallVector<int, 16> ShuffleMask;
4936 if (!getTargetShuffleMask(N, ShufVT, ShuffleMask, IsUnary))
4939 int Elt = ShuffleMask[Index];
4941 return DAG.getUNDEF(ShufVT.getVectorElementType());
4943 SDValue NewV = (Elt < (int)NumElems) ? N->getOperand(0)
4945 return getShuffleScalarElt(NewV.getNode(), Elt % NumElems, DAG,
4949 // Actual nodes that may contain scalar elements
4950 if (Opcode == ISD::BITCAST) {
4951 V = V.getOperand(0);
4952 EVT SrcVT = V.getValueType();
4953 unsigned NumElems = VT.getVectorNumElements();
4955 if (!SrcVT.isVector() || SrcVT.getVectorNumElements() != NumElems)
4959 if (V.getOpcode() == ISD::SCALAR_TO_VECTOR)
4960 return (Index == 0) ? V.getOperand(0)
4961 : DAG.getUNDEF(VT.getVectorElementType());
4963 if (V.getOpcode() == ISD::BUILD_VECTOR)
4964 return V.getOperand(Index);
4969 /// Custom lower build_vector of v16i8.
4970 static SDValue LowerBuildVectorv16i8(SDValue Op, unsigned NonZeros,
4971 unsigned NumNonZero, unsigned NumZero,
4973 const X86Subtarget* Subtarget,
4974 const TargetLowering &TLI) {
4982 // SSE4.1 - use PINSRB to insert each byte directly.
4983 if (Subtarget->hasSSE41()) {
4984 for (unsigned i = 0; i < 16; ++i) {
4985 bool isNonZero = (NonZeros & (1 << i)) != 0;
4989 V = getZeroVector(MVT::v16i8, Subtarget, DAG, dl);
4991 V = DAG.getUNDEF(MVT::v16i8);
4994 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl,
4995 MVT::v16i8, V, Op.getOperand(i),
4996 DAG.getIntPtrConstant(i, dl));
5003 // Pre-SSE4.1 - merge byte pairs and insert with PINSRW.
5004 for (unsigned i = 0; i < 16; ++i) {
5005 bool ThisIsNonZero = (NonZeros & (1 << i)) != 0;
5006 if (ThisIsNonZero && First) {
5008 V = getZeroVector(MVT::v8i16, Subtarget, DAG, dl);
5010 V = DAG.getUNDEF(MVT::v8i16);
5015 SDValue ThisElt, LastElt;
5016 bool LastIsNonZero = (NonZeros & (1 << (i-1))) != 0;
5017 if (LastIsNonZero) {
5018 LastElt = DAG.getNode(ISD::ZERO_EXTEND, dl,
5019 MVT::i16, Op.getOperand(i-1));
5021 if (ThisIsNonZero) {
5022 ThisElt = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i16, Op.getOperand(i));
5023 ThisElt = DAG.getNode(ISD::SHL, dl, MVT::i16,
5024 ThisElt, DAG.getConstant(8, dl, MVT::i8));
5026 ThisElt = DAG.getNode(ISD::OR, dl, MVT::i16, ThisElt, LastElt);
5030 if (ThisElt.getNode())
5031 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, V, ThisElt,
5032 DAG.getIntPtrConstant(i/2, dl));
5036 return DAG.getBitcast(MVT::v16i8, V);
5039 /// Custom lower build_vector of v8i16.
5040 static SDValue LowerBuildVectorv8i16(SDValue Op, unsigned NonZeros,
5041 unsigned NumNonZero, unsigned NumZero,
5043 const X86Subtarget* Subtarget,
5044 const TargetLowering &TLI) {
5051 for (unsigned i = 0; i < 8; ++i) {
5052 bool isNonZero = (NonZeros & (1 << i)) != 0;
5056 V = getZeroVector(MVT::v8i16, Subtarget, DAG, dl);
5058 V = DAG.getUNDEF(MVT::v8i16);
5061 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl,
5062 MVT::v8i16, V, Op.getOperand(i),
5063 DAG.getIntPtrConstant(i, dl));
5070 /// Custom lower build_vector of v4i32 or v4f32.
5071 static SDValue LowerBuildVectorv4x32(SDValue Op, SelectionDAG &DAG,
5072 const X86Subtarget *Subtarget,
5073 const TargetLowering &TLI) {
5074 // Find all zeroable elements.
5075 std::bitset<4> Zeroable;
5076 for (int i=0; i < 4; ++i) {
5077 SDValue Elt = Op->getOperand(i);
5078 Zeroable[i] = (Elt.getOpcode() == ISD::UNDEF || X86::isZeroNode(Elt));
5080 assert(Zeroable.size() - Zeroable.count() > 1 &&
5081 "We expect at least two non-zero elements!");
5083 // We only know how to deal with build_vector nodes where elements are either
5084 // zeroable or extract_vector_elt with constant index.
5085 SDValue FirstNonZero;
5086 unsigned FirstNonZeroIdx;
5087 for (unsigned i=0; i < 4; ++i) {
5090 SDValue Elt = Op->getOperand(i);
5091 if (Elt.getOpcode() != ISD::EXTRACT_VECTOR_ELT ||
5092 !isa<ConstantSDNode>(Elt.getOperand(1)))
5094 // Make sure that this node is extracting from a 128-bit vector.
5095 MVT VT = Elt.getOperand(0).getSimpleValueType();
5096 if (!VT.is128BitVector())
5098 if (!FirstNonZero.getNode()) {
5100 FirstNonZeroIdx = i;
5104 assert(FirstNonZero.getNode() && "Unexpected build vector of all zeros!");
5105 SDValue V1 = FirstNonZero.getOperand(0);
5106 MVT VT = V1.getSimpleValueType();
5108 // See if this build_vector can be lowered as a blend with zero.
5110 unsigned EltMaskIdx, EltIdx;
5112 for (EltIdx = 0; EltIdx < 4; ++EltIdx) {
5113 if (Zeroable[EltIdx]) {
5114 // The zero vector will be on the right hand side.
5115 Mask[EltIdx] = EltIdx+4;
5119 Elt = Op->getOperand(EltIdx);
5120 // By construction, Elt is a EXTRACT_VECTOR_ELT with constant index.
5121 EltMaskIdx = cast<ConstantSDNode>(Elt.getOperand(1))->getZExtValue();
5122 if (Elt.getOperand(0) != V1 || EltMaskIdx != EltIdx)
5124 Mask[EltIdx] = EltIdx;
5128 // Let the shuffle legalizer deal with blend operations.
5129 SDValue VZero = getZeroVector(VT, Subtarget, DAG, SDLoc(Op));
5130 if (V1.getSimpleValueType() != VT)
5131 V1 = DAG.getNode(ISD::BITCAST, SDLoc(V1), VT, V1);
5132 return DAG.getVectorShuffle(VT, SDLoc(V1), V1, VZero, &Mask[0]);
5135 // See if we can lower this build_vector to a INSERTPS.
5136 if (!Subtarget->hasSSE41())
5139 SDValue V2 = Elt.getOperand(0);
5140 if (Elt == FirstNonZero && EltIdx == FirstNonZeroIdx)
5143 bool CanFold = true;
5144 for (unsigned i = EltIdx + 1; i < 4 && CanFold; ++i) {
5148 SDValue Current = Op->getOperand(i);
5149 SDValue SrcVector = Current->getOperand(0);
5152 CanFold = SrcVector == V1 &&
5153 cast<ConstantSDNode>(Current.getOperand(1))->getZExtValue() == i;
5159 assert(V1.getNode() && "Expected at least two non-zero elements!");
5160 if (V1.getSimpleValueType() != MVT::v4f32)
5161 V1 = DAG.getNode(ISD::BITCAST, SDLoc(V1), MVT::v4f32, V1);
5162 if (V2.getSimpleValueType() != MVT::v4f32)
5163 V2 = DAG.getNode(ISD::BITCAST, SDLoc(V2), MVT::v4f32, V2);
5165 // Ok, we can emit an INSERTPS instruction.
5166 unsigned ZMask = Zeroable.to_ulong();
5168 unsigned InsertPSMask = EltMaskIdx << 6 | EltIdx << 4 | ZMask;
5169 assert((InsertPSMask & ~0xFFu) == 0 && "Invalid mask!");
5171 SDValue Result = DAG.getNode(X86ISD::INSERTPS, DL, MVT::v4f32, V1, V2,
5172 DAG.getIntPtrConstant(InsertPSMask, DL));
5173 return DAG.getBitcast(VT, Result);
5176 /// Return a vector logical shift node.
5177 static SDValue getVShift(bool isLeft, EVT VT, SDValue SrcOp,
5178 unsigned NumBits, SelectionDAG &DAG,
5179 const TargetLowering &TLI, SDLoc dl) {
5180 assert(VT.is128BitVector() && "Unknown type for VShift");
5181 MVT ShVT = MVT::v2i64;
5182 unsigned Opc = isLeft ? X86ISD::VSHLDQ : X86ISD::VSRLDQ;
5183 SrcOp = DAG.getBitcast(ShVT, SrcOp);
5184 MVT ScalarShiftTy = TLI.getScalarShiftAmountTy(DAG.getDataLayout(), VT);
5185 assert(NumBits % 8 == 0 && "Only support byte sized shifts");
5186 SDValue ShiftVal = DAG.getConstant(NumBits/8, dl, ScalarShiftTy);
5187 return DAG.getBitcast(VT, DAG.getNode(Opc, dl, ShVT, SrcOp, ShiftVal));
5191 LowerAsSplatVectorLoad(SDValue SrcOp, MVT VT, SDLoc dl, SelectionDAG &DAG) {
5193 // Check if the scalar load can be widened into a vector load. And if
5194 // the address is "base + cst" see if the cst can be "absorbed" into
5195 // the shuffle mask.
5196 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(SrcOp)) {
5197 SDValue Ptr = LD->getBasePtr();
5198 if (!ISD::isNormalLoad(LD) || LD->isVolatile())
5200 EVT PVT = LD->getValueType(0);
5201 if (PVT != MVT::i32 && PVT != MVT::f32)
5206 if (FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr)) {
5207 FI = FINode->getIndex();
5209 } else if (DAG.isBaseWithConstantOffset(Ptr) &&
5210 isa<FrameIndexSDNode>(Ptr.getOperand(0))) {
5211 FI = cast<FrameIndexSDNode>(Ptr.getOperand(0))->getIndex();
5212 Offset = Ptr.getConstantOperandVal(1);
5213 Ptr = Ptr.getOperand(0);
5218 // FIXME: 256-bit vector instructions don't require a strict alignment,
5219 // improve this code to support it better.
5220 unsigned RequiredAlign = VT.getSizeInBits()/8;
5221 SDValue Chain = LD->getChain();
5222 // Make sure the stack object alignment is at least 16 or 32.
5223 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
5224 if (DAG.InferPtrAlignment(Ptr) < RequiredAlign) {
5225 if (MFI->isFixedObjectIndex(FI)) {
5226 // Can't change the alignment. FIXME: It's possible to compute
5227 // the exact stack offset and reference FI + adjust offset instead.
5228 // If someone *really* cares about this. That's the way to implement it.
5231 MFI->setObjectAlignment(FI, RequiredAlign);
5235 // (Offset % 16 or 32) must be multiple of 4. Then address is then
5236 // Ptr + (Offset & ~15).
5239 if ((Offset % RequiredAlign) & 3)
5241 int64_t StartOffset = Offset & ~int64_t(RequiredAlign - 1);
5244 Ptr = DAG.getNode(ISD::ADD, DL, Ptr.getValueType(), Ptr,
5245 DAG.getConstant(StartOffset, DL, Ptr.getValueType()));
5248 int EltNo = (Offset - StartOffset) >> 2;
5249 unsigned NumElems = VT.getVectorNumElements();
5251 EVT NVT = EVT::getVectorVT(*DAG.getContext(), PVT, NumElems);
5252 SDValue V1 = DAG.getLoad(NVT, dl, Chain, Ptr,
5253 LD->getPointerInfo().getWithOffset(StartOffset),
5254 false, false, false, 0);
5256 SmallVector<int, 8> Mask(NumElems, EltNo);
5258 return DAG.getVectorShuffle(NVT, dl, V1, DAG.getUNDEF(NVT), &Mask[0]);
5264 /// Given the initializing elements 'Elts' of a vector of type 'VT', see if the
5265 /// elements can be replaced by a single large load which has the same value as
5266 /// a build_vector or insert_subvector whose loaded operands are 'Elts'.
5268 /// Example: <load i32 *a, load i32 *a+4, undef, undef> -> zextload a
5270 /// FIXME: we'd also like to handle the case where the last elements are zero
5271 /// rather than undef via VZEXT_LOAD, but we do not detect that case today.
5272 /// There's even a handy isZeroNode for that purpose.
5273 static SDValue EltsFromConsecutiveLoads(EVT VT, ArrayRef<SDValue> Elts,
5274 SDLoc &DL, SelectionDAG &DAG,
5275 bool isAfterLegalize) {
5276 unsigned NumElems = Elts.size();
5278 LoadSDNode *LDBase = nullptr;
5279 unsigned LastLoadedElt = -1U;
5281 // For each element in the initializer, see if we've found a load or an undef.
5282 // If we don't find an initial load element, or later load elements are
5283 // non-consecutive, bail out.
5284 for (unsigned i = 0; i < NumElems; ++i) {
5285 SDValue Elt = Elts[i];
5286 // Look through a bitcast.
5287 if (Elt.getNode() && Elt.getOpcode() == ISD::BITCAST)
5288 Elt = Elt.getOperand(0);
5289 if (!Elt.getNode() ||
5290 (Elt.getOpcode() != ISD::UNDEF && !ISD::isNON_EXTLoad(Elt.getNode())))
5293 if (Elt.getNode()->getOpcode() == ISD::UNDEF)
5295 LDBase = cast<LoadSDNode>(Elt.getNode());
5299 if (Elt.getOpcode() == ISD::UNDEF)
5302 LoadSDNode *LD = cast<LoadSDNode>(Elt);
5303 EVT LdVT = Elt.getValueType();
5304 // Each loaded element must be the correct fractional portion of the
5305 // requested vector load.
5306 if (LdVT.getSizeInBits() != VT.getSizeInBits() / NumElems)
5308 if (!DAG.isConsecutiveLoad(LD, LDBase, LdVT.getSizeInBits() / 8, i))
5313 // If we have found an entire vector of loads and undefs, then return a large
5314 // load of the entire vector width starting at the base pointer. If we found
5315 // consecutive loads for the low half, generate a vzext_load node.
5316 if (LastLoadedElt == NumElems - 1) {
5317 assert(LDBase && "Did not find base load for merging consecutive loads");
5318 EVT EltVT = LDBase->getValueType(0);
5319 // Ensure that the input vector size for the merged loads matches the
5320 // cumulative size of the input elements.
5321 if (VT.getSizeInBits() != EltVT.getSizeInBits() * NumElems)
5324 if (isAfterLegalize &&
5325 !DAG.getTargetLoweringInfo().isOperationLegal(ISD::LOAD, VT))
5328 SDValue NewLd = SDValue();
5330 NewLd = DAG.getLoad(VT, DL, LDBase->getChain(), LDBase->getBasePtr(),
5331 LDBase->getPointerInfo(), LDBase->isVolatile(),
5332 LDBase->isNonTemporal(), LDBase->isInvariant(),
5333 LDBase->getAlignment());
5335 if (LDBase->hasAnyUseOfValue(1)) {
5336 SDValue NewChain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other,
5338 SDValue(NewLd.getNode(), 1));
5339 DAG.ReplaceAllUsesOfValueWith(SDValue(LDBase, 1), NewChain);
5340 DAG.UpdateNodeOperands(NewChain.getNode(), SDValue(LDBase, 1),
5341 SDValue(NewLd.getNode(), 1));
5347 //TODO: The code below fires only for for loading the low v2i32 / v2f32
5348 //of a v4i32 / v4f32. It's probably worth generalizing.
5349 EVT EltVT = VT.getVectorElementType();
5350 if (NumElems == 4 && LastLoadedElt == 1 && (EltVT.getSizeInBits() == 32) &&
5351 DAG.getTargetLoweringInfo().isTypeLegal(MVT::v2i64)) {
5352 SDVTList Tys = DAG.getVTList(MVT::v2i64, MVT::Other);
5353 SDValue Ops[] = { LDBase->getChain(), LDBase->getBasePtr() };
5355 DAG.getMemIntrinsicNode(X86ISD::VZEXT_LOAD, DL, Tys, Ops, MVT::i64,
5356 LDBase->getPointerInfo(),
5357 LDBase->getAlignment(),
5358 false/*isVolatile*/, true/*ReadMem*/,
5361 // Make sure the newly-created LOAD is in the same position as LDBase in
5362 // terms of dependency. We create a TokenFactor for LDBase and ResNode, and
5363 // update uses of LDBase's output chain to use the TokenFactor.
5364 if (LDBase->hasAnyUseOfValue(1)) {
5365 SDValue NewChain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other,
5366 SDValue(LDBase, 1), SDValue(ResNode.getNode(), 1));
5367 DAG.ReplaceAllUsesOfValueWith(SDValue(LDBase, 1), NewChain);
5368 DAG.UpdateNodeOperands(NewChain.getNode(), SDValue(LDBase, 1),
5369 SDValue(ResNode.getNode(), 1));
5372 return DAG.getBitcast(VT, ResNode);
5377 /// LowerVectorBroadcast - Attempt to use the vbroadcast instruction
5378 /// to generate a splat value for the following cases:
5379 /// 1. A splat BUILD_VECTOR which uses a single scalar load, or a constant.
5380 /// 2. A splat shuffle which uses a scalar_to_vector node which comes from
5381 /// a scalar load, or a constant.
5382 /// The VBROADCAST node is returned when a pattern is found,
5383 /// or SDValue() otherwise.
5384 static SDValue LowerVectorBroadcast(SDValue Op, const X86Subtarget* Subtarget,
5385 SelectionDAG &DAG) {
5386 // VBROADCAST requires AVX.
5387 // TODO: Splats could be generated for non-AVX CPUs using SSE
5388 // instructions, but there's less potential gain for only 128-bit vectors.
5389 if (!Subtarget->hasAVX())
5392 MVT VT = Op.getSimpleValueType();
5395 assert((VT.is128BitVector() || VT.is256BitVector() || VT.is512BitVector()) &&
5396 "Unsupported vector type for broadcast.");
5401 switch (Op.getOpcode()) {
5403 // Unknown pattern found.
5406 case ISD::BUILD_VECTOR: {
5407 auto *BVOp = cast<BuildVectorSDNode>(Op.getNode());
5408 BitVector UndefElements;
5409 SDValue Splat = BVOp->getSplatValue(&UndefElements);
5411 // We need a splat of a single value to use broadcast, and it doesn't
5412 // make any sense if the value is only in one element of the vector.
5413 if (!Splat || (VT.getVectorNumElements() - UndefElements.count()) <= 1)
5417 ConstSplatVal = (Ld.getOpcode() == ISD::Constant ||
5418 Ld.getOpcode() == ISD::ConstantFP);
5420 // Make sure that all of the users of a non-constant load are from the
5421 // BUILD_VECTOR node.
5422 if (!ConstSplatVal && !BVOp->isOnlyUserOf(Ld.getNode()))
5427 case ISD::VECTOR_SHUFFLE: {
5428 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
5430 // Shuffles must have a splat mask where the first element is
5432 if ((!SVOp->isSplat()) || SVOp->getMaskElt(0) != 0)
5435 SDValue Sc = Op.getOperand(0);
5436 if (Sc.getOpcode() != ISD::SCALAR_TO_VECTOR &&
5437 Sc.getOpcode() != ISD::BUILD_VECTOR) {
5439 if (!Subtarget->hasInt256())
5442 // Use the register form of the broadcast instruction available on AVX2.
5443 if (VT.getSizeInBits() >= 256)
5444 Sc = Extract128BitVector(Sc, 0, DAG, dl);
5445 return DAG.getNode(X86ISD::VBROADCAST, dl, VT, Sc);
5448 Ld = Sc.getOperand(0);
5449 ConstSplatVal = (Ld.getOpcode() == ISD::Constant ||
5450 Ld.getOpcode() == ISD::ConstantFP);
5452 // The scalar_to_vector node and the suspected
5453 // load node must have exactly one user.
5454 // Constants may have multiple users.
5456 // AVX-512 has register version of the broadcast
5457 bool hasRegVer = Subtarget->hasAVX512() && VT.is512BitVector() &&
5458 Ld.getValueType().getSizeInBits() >= 32;
5459 if (!ConstSplatVal && ((!Sc.hasOneUse() || !Ld.hasOneUse()) &&
5466 unsigned ScalarSize = Ld.getValueType().getSizeInBits();
5467 bool IsGE256 = (VT.getSizeInBits() >= 256);
5469 // When optimizing for size, generate up to 5 extra bytes for a broadcast
5470 // instruction to save 8 or more bytes of constant pool data.
5471 // TODO: If multiple splats are generated to load the same constant,
5472 // it may be detrimental to overall size. There needs to be a way to detect
5473 // that condition to know if this is truly a size win.
5474 bool OptForSize = DAG.getMachineFunction().getFunction()->optForSize();
5476 // Handle broadcasting a single constant scalar from the constant pool
5478 // On Sandybridge (no AVX2), it is still better to load a constant vector
5479 // from the constant pool and not to broadcast it from a scalar.
5480 // But override that restriction when optimizing for size.
5481 // TODO: Check if splatting is recommended for other AVX-capable CPUs.
5482 if (ConstSplatVal && (Subtarget->hasAVX2() || OptForSize)) {
5483 EVT CVT = Ld.getValueType();
5484 assert(!CVT.isVector() && "Must not broadcast a vector type");
5486 // Splat f32, i32, v4f64, v4i64 in all cases with AVX2.
5487 // For size optimization, also splat v2f64 and v2i64, and for size opt
5488 // with AVX2, also splat i8 and i16.
5489 // With pattern matching, the VBROADCAST node may become a VMOVDDUP.
5490 if (ScalarSize == 32 || (IsGE256 && ScalarSize == 64) ||
5491 (OptForSize && (ScalarSize == 64 || Subtarget->hasAVX2()))) {
5492 const Constant *C = nullptr;
5493 if (ConstantSDNode *CI = dyn_cast<ConstantSDNode>(Ld))
5494 C = CI->getConstantIntValue();
5495 else if (ConstantFPSDNode *CF = dyn_cast<ConstantFPSDNode>(Ld))
5496 C = CF->getConstantFPValue();
5498 assert(C && "Invalid constant type");
5500 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
5502 DAG.getConstantPool(C, TLI.getPointerTy(DAG.getDataLayout()));
5503 unsigned Alignment = cast<ConstantPoolSDNode>(CP)->getAlignment();
5505 CVT, dl, DAG.getEntryNode(), CP,
5506 MachinePointerInfo::getConstantPool(DAG.getMachineFunction()), false,
5507 false, false, Alignment);
5509 return DAG.getNode(X86ISD::VBROADCAST, dl, VT, Ld);
5513 bool IsLoad = ISD::isNormalLoad(Ld.getNode());
5515 // Handle AVX2 in-register broadcasts.
5516 if (!IsLoad && Subtarget->hasInt256() &&
5517 (ScalarSize == 32 || (IsGE256 && ScalarSize == 64)))
5518 return DAG.getNode(X86ISD::VBROADCAST, dl, VT, Ld);
5520 // The scalar source must be a normal load.
5524 if (ScalarSize == 32 || (IsGE256 && ScalarSize == 64) ||
5525 (Subtarget->hasVLX() && ScalarSize == 64))
5526 return DAG.getNode(X86ISD::VBROADCAST, dl, VT, Ld);
5528 // The integer check is needed for the 64-bit into 128-bit so it doesn't match
5529 // double since there is no vbroadcastsd xmm
5530 if (Subtarget->hasInt256() && Ld.getValueType().isInteger()) {
5531 if (ScalarSize == 8 || ScalarSize == 16 || ScalarSize == 64)
5532 return DAG.getNode(X86ISD::VBROADCAST, dl, VT, Ld);
5535 // Unsupported broadcast.
5539 /// \brief For an EXTRACT_VECTOR_ELT with a constant index return the real
5540 /// underlying vector and index.
5542 /// Modifies \p ExtractedFromVec to the real vector and returns the real
5544 static int getUnderlyingExtractedFromVec(SDValue &ExtractedFromVec,
5546 int Idx = cast<ConstantSDNode>(ExtIdx)->getZExtValue();
5547 if (!isa<ShuffleVectorSDNode>(ExtractedFromVec))
5550 // For 256-bit vectors, LowerEXTRACT_VECTOR_ELT_SSE4 may have already
5552 // (extract_vector_elt (v8f32 %vreg1), Constant<6>)
5554 // (extract_vector_elt (vector_shuffle<2,u,u,u>
5555 // (extract_subvector (v8f32 %vreg0), Constant<4>),
5558 // In this case the vector is the extract_subvector expression and the index
5559 // is 2, as specified by the shuffle.
5560 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(ExtractedFromVec);
5561 SDValue ShuffleVec = SVOp->getOperand(0);
5562 MVT ShuffleVecVT = ShuffleVec.getSimpleValueType();
5563 assert(ShuffleVecVT.getVectorElementType() ==
5564 ExtractedFromVec.getSimpleValueType().getVectorElementType());
5566 int ShuffleIdx = SVOp->getMaskElt(Idx);
5567 if (isUndefOrInRange(ShuffleIdx, 0, ShuffleVecVT.getVectorNumElements())) {
5568 ExtractedFromVec = ShuffleVec;
5574 static SDValue buildFromShuffleMostly(SDValue Op, SelectionDAG &DAG) {
5575 MVT VT = Op.getSimpleValueType();
5577 // Skip if insert_vec_elt is not supported.
5578 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
5579 if (!TLI.isOperationLegalOrCustom(ISD::INSERT_VECTOR_ELT, VT))
5583 unsigned NumElems = Op.getNumOperands();
5587 SmallVector<unsigned, 4> InsertIndices;
5588 SmallVector<int, 8> Mask(NumElems, -1);
5590 for (unsigned i = 0; i != NumElems; ++i) {
5591 unsigned Opc = Op.getOperand(i).getOpcode();
5593 if (Opc == ISD::UNDEF)
5596 if (Opc != ISD::EXTRACT_VECTOR_ELT) {
5597 // Quit if more than 1 elements need inserting.
5598 if (InsertIndices.size() > 1)
5601 InsertIndices.push_back(i);
5605 SDValue ExtractedFromVec = Op.getOperand(i).getOperand(0);
5606 SDValue ExtIdx = Op.getOperand(i).getOperand(1);
5607 // Quit if non-constant index.
5608 if (!isa<ConstantSDNode>(ExtIdx))
5610 int Idx = getUnderlyingExtractedFromVec(ExtractedFromVec, ExtIdx);
5612 // Quit if extracted from vector of different type.
5613 if (ExtractedFromVec.getValueType() != VT)
5616 if (!VecIn1.getNode())
5617 VecIn1 = ExtractedFromVec;
5618 else if (VecIn1 != ExtractedFromVec) {
5619 if (!VecIn2.getNode())
5620 VecIn2 = ExtractedFromVec;
5621 else if (VecIn2 != ExtractedFromVec)
5622 // Quit if more than 2 vectors to shuffle
5626 if (ExtractedFromVec == VecIn1)
5628 else if (ExtractedFromVec == VecIn2)
5629 Mask[i] = Idx + NumElems;
5632 if (!VecIn1.getNode())
5635 VecIn2 = VecIn2.getNode() ? VecIn2 : DAG.getUNDEF(VT);
5636 SDValue NV = DAG.getVectorShuffle(VT, DL, VecIn1, VecIn2, &Mask[0]);
5637 for (unsigned i = 0, e = InsertIndices.size(); i != e; ++i) {
5638 unsigned Idx = InsertIndices[i];
5639 NV = DAG.getNode(ISD::INSERT_VECTOR_ELT, DL, VT, NV, Op.getOperand(Idx),
5640 DAG.getIntPtrConstant(Idx, DL));
5646 static SDValue ConvertI1VectorToInteger(SDValue Op, SelectionDAG &DAG) {
5647 assert(ISD::isBuildVectorOfConstantSDNodes(Op.getNode()) &&
5648 Op.getScalarValueSizeInBits() == 1 &&
5649 "Can not convert non-constant vector");
5650 uint64_t Immediate = 0;
5651 for (unsigned idx = 0, e = Op.getNumOperands(); idx < e; ++idx) {
5652 SDValue In = Op.getOperand(idx);
5653 if (In.getOpcode() != ISD::UNDEF)
5654 Immediate |= cast<ConstantSDNode>(In)->getZExtValue() << idx;
5658 MVT::getIntegerVT(std::max((int)Op.getValueType().getSizeInBits(), 8));
5659 return DAG.getConstant(Immediate, dl, VT);
5661 // Lower BUILD_VECTOR operation for v8i1 and v16i1 types.
5663 X86TargetLowering::LowerBUILD_VECTORvXi1(SDValue Op, SelectionDAG &DAG) const {
5665 MVT VT = Op.getSimpleValueType();
5666 assert((VT.getVectorElementType() == MVT::i1) &&
5667 "Unexpected type in LowerBUILD_VECTORvXi1!");
5670 if (ISD::isBuildVectorAllZeros(Op.getNode())) {
5671 SDValue Cst = DAG.getTargetConstant(0, dl, MVT::i1);
5672 SmallVector<SDValue, 16> Ops(VT.getVectorNumElements(), Cst);
5673 return DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Ops);
5676 if (ISD::isBuildVectorAllOnes(Op.getNode())) {
5677 SDValue Cst = DAG.getTargetConstant(1, dl, MVT::i1);
5678 SmallVector<SDValue, 16> Ops(VT.getVectorNumElements(), Cst);
5679 return DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Ops);
5682 if (ISD::isBuildVectorOfConstantSDNodes(Op.getNode())) {
5683 SDValue Imm = ConvertI1VectorToInteger(Op, DAG);
5684 if (Imm.getValueSizeInBits() == VT.getSizeInBits())
5685 return DAG.getBitcast(VT, Imm);
5686 SDValue ExtVec = DAG.getBitcast(MVT::v8i1, Imm);
5687 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT, ExtVec,
5688 DAG.getIntPtrConstant(0, dl));
5691 // Vector has one or more non-const elements
5692 uint64_t Immediate = 0;
5693 SmallVector<unsigned, 16> NonConstIdx;
5694 bool IsSplat = true;
5695 bool HasConstElts = false;
5697 for (unsigned idx = 0, e = Op.getNumOperands(); idx < e; ++idx) {
5698 SDValue In = Op.getOperand(idx);
5699 if (In.getOpcode() == ISD::UNDEF)
5701 if (!isa<ConstantSDNode>(In))
5702 NonConstIdx.push_back(idx);
5704 Immediate |= cast<ConstantSDNode>(In)->getZExtValue() << idx;
5705 HasConstElts = true;
5709 else if (In != Op.getOperand(SplatIdx))
5713 // for splat use " (select i1 splat_elt, all-ones, all-zeroes)"
5715 return DAG.getNode(ISD::SELECT, dl, VT, Op.getOperand(SplatIdx),
5716 DAG.getConstant(1, dl, VT),
5717 DAG.getConstant(0, dl, VT));
5719 // insert elements one by one
5723 MVT ImmVT = MVT::getIntegerVT(std::max((int)VT.getSizeInBits(), 8));
5724 Imm = DAG.getConstant(Immediate, dl, ImmVT);
5726 else if (HasConstElts)
5727 Imm = DAG.getConstant(0, dl, VT);
5729 Imm = DAG.getUNDEF(VT);
5730 if (Imm.getValueSizeInBits() == VT.getSizeInBits())
5731 DstVec = DAG.getBitcast(VT, Imm);
5733 SDValue ExtVec = DAG.getBitcast(MVT::v8i1, Imm);
5734 DstVec = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT, ExtVec,
5735 DAG.getIntPtrConstant(0, dl));
5738 for (unsigned i = 0; i < NonConstIdx.size(); ++i) {
5739 unsigned InsertIdx = NonConstIdx[i];
5740 DstVec = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, DstVec,
5741 Op.getOperand(InsertIdx),
5742 DAG.getIntPtrConstant(InsertIdx, dl));
5747 /// \brief Return true if \p N implements a horizontal binop and return the
5748 /// operands for the horizontal binop into V0 and V1.
5750 /// This is a helper function of LowerToHorizontalOp().
5751 /// This function checks that the build_vector \p N in input implements a
5752 /// horizontal operation. Parameter \p Opcode defines the kind of horizontal
5753 /// operation to match.
5754 /// For example, if \p Opcode is equal to ISD::ADD, then this function
5755 /// checks if \p N implements a horizontal arithmetic add; if instead \p Opcode
5756 /// is equal to ISD::SUB, then this function checks if this is a horizontal
5759 /// This function only analyzes elements of \p N whose indices are
5760 /// in range [BaseIdx, LastIdx).
5761 static bool isHorizontalBinOp(const BuildVectorSDNode *N, unsigned Opcode,
5763 unsigned BaseIdx, unsigned LastIdx,
5764 SDValue &V0, SDValue &V1) {
5765 EVT VT = N->getValueType(0);
5767 assert(BaseIdx * 2 <= LastIdx && "Invalid Indices in input!");
5768 assert(VT.isVector() && VT.getVectorNumElements() >= LastIdx &&
5769 "Invalid Vector in input!");
5771 bool IsCommutable = (Opcode == ISD::ADD || Opcode == ISD::FADD);
5772 bool CanFold = true;
5773 unsigned ExpectedVExtractIdx = BaseIdx;
5774 unsigned NumElts = LastIdx - BaseIdx;
5775 V0 = DAG.getUNDEF(VT);
5776 V1 = DAG.getUNDEF(VT);
5778 // Check if N implements a horizontal binop.
5779 for (unsigned i = 0, e = NumElts; i != e && CanFold; ++i) {
5780 SDValue Op = N->getOperand(i + BaseIdx);
5783 if (Op->getOpcode() == ISD::UNDEF) {
5784 // Update the expected vector extract index.
5785 if (i * 2 == NumElts)
5786 ExpectedVExtractIdx = BaseIdx;
5787 ExpectedVExtractIdx += 2;
5791 CanFold = Op->getOpcode() == Opcode && Op->hasOneUse();
5796 SDValue Op0 = Op.getOperand(0);
5797 SDValue Op1 = Op.getOperand(1);
5799 // Try to match the following pattern:
5800 // (BINOP (extract_vector_elt A, I), (extract_vector_elt A, I+1))
5801 CanFold = (Op0.getOpcode() == ISD::EXTRACT_VECTOR_ELT &&
5802 Op1.getOpcode() == ISD::EXTRACT_VECTOR_ELT &&
5803 Op0.getOperand(0) == Op1.getOperand(0) &&
5804 isa<ConstantSDNode>(Op0.getOperand(1)) &&
5805 isa<ConstantSDNode>(Op1.getOperand(1)));
5809 unsigned I0 = cast<ConstantSDNode>(Op0.getOperand(1))->getZExtValue();
5810 unsigned I1 = cast<ConstantSDNode>(Op1.getOperand(1))->getZExtValue();
5812 if (i * 2 < NumElts) {
5813 if (V0.getOpcode() == ISD::UNDEF) {
5814 V0 = Op0.getOperand(0);
5815 if (V0.getValueType() != VT)
5819 if (V1.getOpcode() == ISD::UNDEF) {
5820 V1 = Op0.getOperand(0);
5821 if (V1.getValueType() != VT)
5824 if (i * 2 == NumElts)
5825 ExpectedVExtractIdx = BaseIdx;
5828 SDValue Expected = (i * 2 < NumElts) ? V0 : V1;
5829 if (I0 == ExpectedVExtractIdx)
5830 CanFold = I1 == I0 + 1 && Op0.getOperand(0) == Expected;
5831 else if (IsCommutable && I1 == ExpectedVExtractIdx) {
5832 // Try to match the following dag sequence:
5833 // (BINOP (extract_vector_elt A, I+1), (extract_vector_elt A, I))
5834 CanFold = I0 == I1 + 1 && Op1.getOperand(0) == Expected;
5838 ExpectedVExtractIdx += 2;
5844 /// \brief Emit a sequence of two 128-bit horizontal add/sub followed by
5845 /// a concat_vector.
5847 /// This is a helper function of LowerToHorizontalOp().
5848 /// This function expects two 256-bit vectors called V0 and V1.
5849 /// At first, each vector is split into two separate 128-bit vectors.
5850 /// Then, the resulting 128-bit vectors are used to implement two
5851 /// horizontal binary operations.
5853 /// The kind of horizontal binary operation is defined by \p X86Opcode.
5855 /// \p Mode specifies how the 128-bit parts of V0 and V1 are passed in input to
5856 /// the two new horizontal binop.
5857 /// When Mode is set, the first horizontal binop dag node would take as input
5858 /// the lower 128-bit of V0 and the upper 128-bit of V0. The second
5859 /// horizontal binop dag node would take as input the lower 128-bit of V1
5860 /// and the upper 128-bit of V1.
5862 /// HADD V0_LO, V0_HI
5863 /// HADD V1_LO, V1_HI
5865 /// Otherwise, the first horizontal binop dag node takes as input the lower
5866 /// 128-bit of V0 and the lower 128-bit of V1, and the second horizontal binop
5867 /// dag node takes the upper 128-bit of V0 and the upper 128-bit of V1.
5869 /// HADD V0_LO, V1_LO
5870 /// HADD V0_HI, V1_HI
5872 /// If \p isUndefLO is set, then the algorithm propagates UNDEF to the lower
5873 /// 128-bits of the result. If \p isUndefHI is set, then UNDEF is propagated to
5874 /// the upper 128-bits of the result.
5875 static SDValue ExpandHorizontalBinOp(const SDValue &V0, const SDValue &V1,
5876 SDLoc DL, SelectionDAG &DAG,
5877 unsigned X86Opcode, bool Mode,
5878 bool isUndefLO, bool isUndefHI) {
5879 EVT VT = V0.getValueType();
5880 assert(VT.is256BitVector() && VT == V1.getValueType() &&
5881 "Invalid nodes in input!");
5883 unsigned NumElts = VT.getVectorNumElements();
5884 SDValue V0_LO = Extract128BitVector(V0, 0, DAG, DL);
5885 SDValue V0_HI = Extract128BitVector(V0, NumElts/2, DAG, DL);
5886 SDValue V1_LO = Extract128BitVector(V1, 0, DAG, DL);
5887 SDValue V1_HI = Extract128BitVector(V1, NumElts/2, DAG, DL);
5888 EVT NewVT = V0_LO.getValueType();
5890 SDValue LO = DAG.getUNDEF(NewVT);
5891 SDValue HI = DAG.getUNDEF(NewVT);
5894 // Don't emit a horizontal binop if the result is expected to be UNDEF.
5895 if (!isUndefLO && V0->getOpcode() != ISD::UNDEF)
5896 LO = DAG.getNode(X86Opcode, DL, NewVT, V0_LO, V0_HI);
5897 if (!isUndefHI && V1->getOpcode() != ISD::UNDEF)
5898 HI = DAG.getNode(X86Opcode, DL, NewVT, V1_LO, V1_HI);
5900 // Don't emit a horizontal binop if the result is expected to be UNDEF.
5901 if (!isUndefLO && (V0_LO->getOpcode() != ISD::UNDEF ||
5902 V1_LO->getOpcode() != ISD::UNDEF))
5903 LO = DAG.getNode(X86Opcode, DL, NewVT, V0_LO, V1_LO);
5905 if (!isUndefHI && (V0_HI->getOpcode() != ISD::UNDEF ||
5906 V1_HI->getOpcode() != ISD::UNDEF))
5907 HI = DAG.getNode(X86Opcode, DL, NewVT, V0_HI, V1_HI);
5910 return DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, LO, HI);
5913 /// Try to fold a build_vector that performs an 'addsub' to an X86ISD::ADDSUB
5915 static SDValue LowerToAddSub(const BuildVectorSDNode *BV,
5916 const X86Subtarget *Subtarget, SelectionDAG &DAG) {
5917 EVT VT = BV->getValueType(0);
5918 if ((!Subtarget->hasSSE3() || (VT != MVT::v4f32 && VT != MVT::v2f64)) &&
5919 (!Subtarget->hasAVX() || (VT != MVT::v8f32 && VT != MVT::v4f64)))
5923 unsigned NumElts = VT.getVectorNumElements();
5924 SDValue InVec0 = DAG.getUNDEF(VT);
5925 SDValue InVec1 = DAG.getUNDEF(VT);
5927 assert((VT == MVT::v8f32 || VT == MVT::v4f64 || VT == MVT::v4f32 ||
5928 VT == MVT::v2f64) && "build_vector with an invalid type found!");
5930 // Odd-numbered elements in the input build vector are obtained from
5931 // adding two integer/float elements.
5932 // Even-numbered elements in the input build vector are obtained from
5933 // subtracting two integer/float elements.
5934 unsigned ExpectedOpcode = ISD::FSUB;
5935 unsigned NextExpectedOpcode = ISD::FADD;
5936 bool AddFound = false;
5937 bool SubFound = false;
5939 for (unsigned i = 0, e = NumElts; i != e; ++i) {
5940 SDValue Op = BV->getOperand(i);
5942 // Skip 'undef' values.
5943 unsigned Opcode = Op.getOpcode();
5944 if (Opcode == ISD::UNDEF) {
5945 std::swap(ExpectedOpcode, NextExpectedOpcode);
5949 // Early exit if we found an unexpected opcode.
5950 if (Opcode != ExpectedOpcode)
5953 SDValue Op0 = Op.getOperand(0);
5954 SDValue Op1 = Op.getOperand(1);
5956 // Try to match the following pattern:
5957 // (BINOP (extract_vector_elt A, i), (extract_vector_elt B, i))
5958 // Early exit if we cannot match that sequence.
5959 if (Op0.getOpcode() != ISD::EXTRACT_VECTOR_ELT ||
5960 Op1.getOpcode() != ISD::EXTRACT_VECTOR_ELT ||
5961 !isa<ConstantSDNode>(Op0.getOperand(1)) ||
5962 !isa<ConstantSDNode>(Op1.getOperand(1)) ||
5963 Op0.getOperand(1) != Op1.getOperand(1))
5966 unsigned I0 = cast<ConstantSDNode>(Op0.getOperand(1))->getZExtValue();
5970 // We found a valid add/sub node. Update the information accordingly.
5976 // Update InVec0 and InVec1.
5977 if (InVec0.getOpcode() == ISD::UNDEF) {
5978 InVec0 = Op0.getOperand(0);
5979 if (InVec0.getValueType() != VT)
5982 if (InVec1.getOpcode() == ISD::UNDEF) {
5983 InVec1 = Op1.getOperand(0);
5984 if (InVec1.getValueType() != VT)
5988 // Make sure that operands in input to each add/sub node always
5989 // come from a same pair of vectors.
5990 if (InVec0 != Op0.getOperand(0)) {
5991 if (ExpectedOpcode == ISD::FSUB)
5994 // FADD is commutable. Try to commute the operands
5995 // and then test again.
5996 std::swap(Op0, Op1);
5997 if (InVec0 != Op0.getOperand(0))
6001 if (InVec1 != Op1.getOperand(0))
6004 // Update the pair of expected opcodes.
6005 std::swap(ExpectedOpcode, NextExpectedOpcode);
6008 // Don't try to fold this build_vector into an ADDSUB if the inputs are undef.
6009 if (AddFound && SubFound && InVec0.getOpcode() != ISD::UNDEF &&
6010 InVec1.getOpcode() != ISD::UNDEF)
6011 return DAG.getNode(X86ISD::ADDSUB, DL, VT, InVec0, InVec1);
6016 /// Lower BUILD_VECTOR to a horizontal add/sub operation if possible.
6017 static SDValue LowerToHorizontalOp(const BuildVectorSDNode *BV,
6018 const X86Subtarget *Subtarget,
6019 SelectionDAG &DAG) {
6020 EVT VT = BV->getValueType(0);
6021 unsigned NumElts = VT.getVectorNumElements();
6022 unsigned NumUndefsLO = 0;
6023 unsigned NumUndefsHI = 0;
6024 unsigned Half = NumElts/2;
6026 // Count the number of UNDEF operands in the build_vector in input.
6027 for (unsigned i = 0, e = Half; i != e; ++i)
6028 if (BV->getOperand(i)->getOpcode() == ISD::UNDEF)
6031 for (unsigned i = Half, e = NumElts; i != e; ++i)
6032 if (BV->getOperand(i)->getOpcode() == ISD::UNDEF)
6035 // Early exit if this is either a build_vector of all UNDEFs or all the
6036 // operands but one are UNDEF.
6037 if (NumUndefsLO + NumUndefsHI + 1 >= NumElts)
6041 SDValue InVec0, InVec1;
6042 if ((VT == MVT::v4f32 || VT == MVT::v2f64) && Subtarget->hasSSE3()) {
6043 // Try to match an SSE3 float HADD/HSUB.
6044 if (isHorizontalBinOp(BV, ISD::FADD, DAG, 0, NumElts, InVec0, InVec1))
6045 return DAG.getNode(X86ISD::FHADD, DL, VT, InVec0, InVec1);
6047 if (isHorizontalBinOp(BV, ISD::FSUB, DAG, 0, NumElts, InVec0, InVec1))
6048 return DAG.getNode(X86ISD::FHSUB, DL, VT, InVec0, InVec1);
6049 } else if ((VT == MVT::v4i32 || VT == MVT::v8i16) && Subtarget->hasSSSE3()) {
6050 // Try to match an SSSE3 integer HADD/HSUB.
6051 if (isHorizontalBinOp(BV, ISD::ADD, DAG, 0, NumElts, InVec0, InVec1))
6052 return DAG.getNode(X86ISD::HADD, DL, VT, InVec0, InVec1);
6054 if (isHorizontalBinOp(BV, ISD::SUB, DAG, 0, NumElts, InVec0, InVec1))
6055 return DAG.getNode(X86ISD::HSUB, DL, VT, InVec0, InVec1);
6058 if (!Subtarget->hasAVX())
6061 if ((VT == MVT::v8f32 || VT == MVT::v4f64)) {
6062 // Try to match an AVX horizontal add/sub of packed single/double
6063 // precision floating point values from 256-bit vectors.
6064 SDValue InVec2, InVec3;
6065 if (isHorizontalBinOp(BV, ISD::FADD, DAG, 0, Half, InVec0, InVec1) &&
6066 isHorizontalBinOp(BV, ISD::FADD, DAG, Half, NumElts, InVec2, InVec3) &&
6067 ((InVec0.getOpcode() == ISD::UNDEF ||
6068 InVec2.getOpcode() == ISD::UNDEF) || InVec0 == InVec2) &&
6069 ((InVec1.getOpcode() == ISD::UNDEF ||
6070 InVec3.getOpcode() == ISD::UNDEF) || InVec1 == InVec3))
6071 return DAG.getNode(X86ISD::FHADD, DL, VT, InVec0, InVec1);
6073 if (isHorizontalBinOp(BV, ISD::FSUB, DAG, 0, Half, InVec0, InVec1) &&
6074 isHorizontalBinOp(BV, ISD::FSUB, DAG, Half, NumElts, InVec2, InVec3) &&
6075 ((InVec0.getOpcode() == ISD::UNDEF ||
6076 InVec2.getOpcode() == ISD::UNDEF) || InVec0 == InVec2) &&
6077 ((InVec1.getOpcode() == ISD::UNDEF ||
6078 InVec3.getOpcode() == ISD::UNDEF) || InVec1 == InVec3))
6079 return DAG.getNode(X86ISD::FHSUB, DL, VT, InVec0, InVec1);
6080 } else if (VT == MVT::v8i32 || VT == MVT::v16i16) {
6081 // Try to match an AVX2 horizontal add/sub of signed integers.
6082 SDValue InVec2, InVec3;
6084 bool CanFold = true;
6086 if (isHorizontalBinOp(BV, ISD::ADD, DAG, 0, Half, InVec0, InVec1) &&
6087 isHorizontalBinOp(BV, ISD::ADD, DAG, Half, NumElts, InVec2, InVec3) &&
6088 ((InVec0.getOpcode() == ISD::UNDEF ||
6089 InVec2.getOpcode() == ISD::UNDEF) || InVec0 == InVec2) &&
6090 ((InVec1.getOpcode() == ISD::UNDEF ||
6091 InVec3.getOpcode() == ISD::UNDEF) || InVec1 == InVec3))
6092 X86Opcode = X86ISD::HADD;
6093 else if (isHorizontalBinOp(BV, ISD::SUB, DAG, 0, Half, InVec0, InVec1) &&
6094 isHorizontalBinOp(BV, ISD::SUB, DAG, Half, NumElts, InVec2, InVec3) &&
6095 ((InVec0.getOpcode() == ISD::UNDEF ||
6096 InVec2.getOpcode() == ISD::UNDEF) || InVec0 == InVec2) &&
6097 ((InVec1.getOpcode() == ISD::UNDEF ||
6098 InVec3.getOpcode() == ISD::UNDEF) || InVec1 == InVec3))
6099 X86Opcode = X86ISD::HSUB;
6104 // Fold this build_vector into a single horizontal add/sub.
6105 // Do this only if the target has AVX2.
6106 if (Subtarget->hasAVX2())
6107 return DAG.getNode(X86Opcode, DL, VT, InVec0, InVec1);
6109 // Do not try to expand this build_vector into a pair of horizontal
6110 // add/sub if we can emit a pair of scalar add/sub.
6111 if (NumUndefsLO + 1 == Half || NumUndefsHI + 1 == Half)
6114 // Convert this build_vector into a pair of horizontal binop followed by
6116 bool isUndefLO = NumUndefsLO == Half;
6117 bool isUndefHI = NumUndefsHI == Half;
6118 return ExpandHorizontalBinOp(InVec0, InVec1, DL, DAG, X86Opcode, false,
6119 isUndefLO, isUndefHI);
6123 if ((VT == MVT::v8f32 || VT == MVT::v4f64 || VT == MVT::v8i32 ||
6124 VT == MVT::v16i16) && Subtarget->hasAVX()) {
6126 if (isHorizontalBinOp(BV, ISD::ADD, DAG, 0, NumElts, InVec0, InVec1))
6127 X86Opcode = X86ISD::HADD;
6128 else if (isHorizontalBinOp(BV, ISD::SUB, DAG, 0, NumElts, InVec0, InVec1))
6129 X86Opcode = X86ISD::HSUB;
6130 else if (isHorizontalBinOp(BV, ISD::FADD, DAG, 0, NumElts, InVec0, InVec1))
6131 X86Opcode = X86ISD::FHADD;
6132 else if (isHorizontalBinOp(BV, ISD::FSUB, DAG, 0, NumElts, InVec0, InVec1))
6133 X86Opcode = X86ISD::FHSUB;
6137 // Don't try to expand this build_vector into a pair of horizontal add/sub
6138 // if we can simply emit a pair of scalar add/sub.
6139 if (NumUndefsLO + 1 == Half || NumUndefsHI + 1 == Half)
6142 // Convert this build_vector into two horizontal add/sub followed by
6144 bool isUndefLO = NumUndefsLO == Half;
6145 bool isUndefHI = NumUndefsHI == Half;
6146 return ExpandHorizontalBinOp(InVec0, InVec1, DL, DAG, X86Opcode, true,
6147 isUndefLO, isUndefHI);
6154 X86TargetLowering::LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG) const {
6157 MVT VT = Op.getSimpleValueType();
6158 MVT ExtVT = VT.getVectorElementType();
6159 unsigned NumElems = Op.getNumOperands();
6161 // Generate vectors for predicate vectors.
6162 if (VT.getScalarType() == MVT::i1 && Subtarget->hasAVX512())
6163 return LowerBUILD_VECTORvXi1(Op, DAG);
6165 // Vectors containing all zeros can be matched by pxor and xorps later
6166 if (ISD::isBuildVectorAllZeros(Op.getNode())) {
6167 // Canonicalize this to <4 x i32> to 1) ensure the zero vectors are CSE'd
6168 // and 2) ensure that i64 scalars are eliminated on x86-32 hosts.
6169 if (VT == MVT::v4i32 || VT == MVT::v8i32 || VT == MVT::v16i32)
6172 return getZeroVector(VT, Subtarget, DAG, dl);
6175 // Vectors containing all ones can be matched by pcmpeqd on 128-bit width
6176 // vectors or broken into v4i32 operations on 256-bit vectors. AVX2 can use
6177 // vpcmpeqd on 256-bit vectors.
6178 if (Subtarget->hasSSE2() && ISD::isBuildVectorAllOnes(Op.getNode())) {
6179 if (VT == MVT::v4i32 || (VT == MVT::v8i32 && Subtarget->hasInt256()))
6182 if (!VT.is512BitVector())
6183 return getOnesVector(VT, Subtarget, DAG, dl);
6186 BuildVectorSDNode *BV = cast<BuildVectorSDNode>(Op.getNode());
6187 if (SDValue AddSub = LowerToAddSub(BV, Subtarget, DAG))
6189 if (SDValue HorizontalOp = LowerToHorizontalOp(BV, Subtarget, DAG))
6190 return HorizontalOp;
6191 if (SDValue Broadcast = LowerVectorBroadcast(Op, Subtarget, DAG))
6194 unsigned EVTBits = ExtVT.getSizeInBits();
6196 unsigned NumZero = 0;
6197 unsigned NumNonZero = 0;
6198 unsigned NonZeros = 0;
6199 bool IsAllConstants = true;
6200 SmallSet<SDValue, 8> Values;
6201 for (unsigned i = 0; i < NumElems; ++i) {
6202 SDValue Elt = Op.getOperand(i);
6203 if (Elt.getOpcode() == ISD::UNDEF)
6206 if (Elt.getOpcode() != ISD::Constant &&
6207 Elt.getOpcode() != ISD::ConstantFP)
6208 IsAllConstants = false;
6209 if (X86::isZeroNode(Elt))
6212 NonZeros |= (1 << i);
6217 // All undef vector. Return an UNDEF. All zero vectors were handled above.
6218 if (NumNonZero == 0)
6219 return DAG.getUNDEF(VT);
6221 // Special case for single non-zero, non-undef, element.
6222 if (NumNonZero == 1) {
6223 unsigned Idx = countTrailingZeros(NonZeros);
6224 SDValue Item = Op.getOperand(Idx);
6226 // If this is an insertion of an i64 value on x86-32, and if the top bits of
6227 // the value are obviously zero, truncate the value to i32 and do the
6228 // insertion that way. Only do this if the value is non-constant or if the
6229 // value is a constant being inserted into element 0. It is cheaper to do
6230 // a constant pool load than it is to do a movd + shuffle.
6231 if (ExtVT == MVT::i64 && !Subtarget->is64Bit() &&
6232 (!IsAllConstants || Idx == 0)) {
6233 if (DAG.MaskedValueIsZero(Item, APInt::getBitsSet(64, 32, 64))) {
6235 assert(VT == MVT::v2i64 && "Expected an SSE value type!");
6236 EVT VecVT = MVT::v4i32;
6238 // Truncate the value (which may itself be a constant) to i32, and
6239 // convert it to a vector with movd (S2V+shuffle to zero extend).
6240 Item = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, Item);
6241 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VecVT, Item);
6242 return DAG.getBitcast(VT, getShuffleVectorZeroOrUndef(
6243 Item, Idx * 2, true, Subtarget, DAG));
6247 // If we have a constant or non-constant insertion into the low element of
6248 // a vector, we can do this with SCALAR_TO_VECTOR + shuffle of zero into
6249 // the rest of the elements. This will be matched as movd/movq/movss/movsd
6250 // depending on what the source datatype is.
6253 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
6255 if (ExtVT == MVT::i32 || ExtVT == MVT::f32 || ExtVT == MVT::f64 ||
6256 (ExtVT == MVT::i64 && Subtarget->is64Bit())) {
6257 if (VT.is512BitVector()) {
6258 SDValue ZeroVec = getZeroVector(VT, Subtarget, DAG, dl);
6259 return DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, ZeroVec,
6260 Item, DAG.getIntPtrConstant(0, dl));
6262 assert((VT.is128BitVector() || VT.is256BitVector()) &&
6263 "Expected an SSE value type!");
6264 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
6265 // Turn it into a MOVL (i.e. movss, movsd, or movd) to a zero vector.
6266 return getShuffleVectorZeroOrUndef(Item, 0, true, Subtarget, DAG);
6269 // We can't directly insert an i8 or i16 into a vector, so zero extend
6271 if (ExtVT == MVT::i16 || ExtVT == MVT::i8) {
6272 Item = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, Item);
6273 if (VT.is256BitVector()) {
6274 if (Subtarget->hasAVX()) {
6275 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v8i32, Item);
6276 Item = getShuffleVectorZeroOrUndef(Item, 0, true, Subtarget, DAG);
6278 // Without AVX, we need to extend to a 128-bit vector and then
6279 // insert into the 256-bit vector.
6280 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32, Item);
6281 SDValue ZeroVec = getZeroVector(MVT::v8i32, Subtarget, DAG, dl);
6282 Item = Insert128BitVector(ZeroVec, Item, 0, DAG, dl);
6285 assert(VT.is128BitVector() && "Expected an SSE value type!");
6286 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32, Item);
6287 Item = getShuffleVectorZeroOrUndef(Item, 0, true, Subtarget, DAG);
6289 return DAG.getBitcast(VT, Item);
6293 // Is it a vector logical left shift?
6294 if (NumElems == 2 && Idx == 1 &&
6295 X86::isZeroNode(Op.getOperand(0)) &&
6296 !X86::isZeroNode(Op.getOperand(1))) {
6297 unsigned NumBits = VT.getSizeInBits();
6298 return getVShift(true, VT,
6299 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
6300 VT, Op.getOperand(1)),
6301 NumBits/2, DAG, *this, dl);
6304 if (IsAllConstants) // Otherwise, it's better to do a constpool load.
6307 // Otherwise, if this is a vector with i32 or f32 elements, and the element
6308 // is a non-constant being inserted into an element other than the low one,
6309 // we can't use a constant pool load. Instead, use SCALAR_TO_VECTOR (aka
6310 // movd/movss) to move this into the low element, then shuffle it into
6312 if (EVTBits == 32) {
6313 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
6314 return getShuffleVectorZeroOrUndef(Item, Idx, NumZero > 0, Subtarget, DAG);
6318 // Splat is obviously ok. Let legalizer expand it to a shuffle.
6319 if (Values.size() == 1) {
6320 if (EVTBits == 32) {
6321 // Instead of a shuffle like this:
6322 // shuffle (scalar_to_vector (load (ptr + 4))), undef, <0, 0, 0, 0>
6323 // Check if it's possible to issue this instead.
6324 // shuffle (vload ptr)), undef, <1, 1, 1, 1>
6325 unsigned Idx = countTrailingZeros(NonZeros);
6326 SDValue Item = Op.getOperand(Idx);
6327 if (Op.getNode()->isOnlyUserOf(Item.getNode()))
6328 return LowerAsSplatVectorLoad(Item, VT, dl, DAG);
6333 // A vector full of immediates; various special cases are already
6334 // handled, so this is best done with a single constant-pool load.
6338 // For AVX-length vectors, see if we can use a vector load to get all of the
6339 // elements, otherwise build the individual 128-bit pieces and use
6340 // shuffles to put them in place.
6341 if (VT.is256BitVector() || VT.is512BitVector()) {
6342 SmallVector<SDValue, 64> V(Op->op_begin(), Op->op_begin() + NumElems);
6344 // Check for a build vector of consecutive loads.
6345 if (SDValue LD = EltsFromConsecutiveLoads(VT, V, dl, DAG, false))
6348 EVT HVT = EVT::getVectorVT(*DAG.getContext(), ExtVT, NumElems/2);
6350 // Build both the lower and upper subvector.
6351 SDValue Lower = DAG.getNode(ISD::BUILD_VECTOR, dl, HVT,
6352 makeArrayRef(&V[0], NumElems/2));
6353 SDValue Upper = DAG.getNode(ISD::BUILD_VECTOR, dl, HVT,
6354 makeArrayRef(&V[NumElems / 2], NumElems/2));
6356 // Recreate the wider vector with the lower and upper part.
6357 if (VT.is256BitVector())
6358 return Concat128BitVectors(Lower, Upper, VT, NumElems, DAG, dl);
6359 return Concat256BitVectors(Lower, Upper, VT, NumElems, DAG, dl);
6362 // Let legalizer expand 2-wide build_vectors.
6363 if (EVTBits == 64) {
6364 if (NumNonZero == 1) {
6365 // One half is zero or undef.
6366 unsigned Idx = countTrailingZeros(NonZeros);
6367 SDValue V2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT,
6368 Op.getOperand(Idx));
6369 return getShuffleVectorZeroOrUndef(V2, Idx, true, Subtarget, DAG);
6374 // If element VT is < 32 bits, convert it to inserts into a zero vector.
6375 if (EVTBits == 8 && NumElems == 16)
6376 if (SDValue V = LowerBuildVectorv16i8(Op, NonZeros,NumNonZero,NumZero, DAG,
6380 if (EVTBits == 16 && NumElems == 8)
6381 if (SDValue V = LowerBuildVectorv8i16(Op, NonZeros,NumNonZero,NumZero, DAG,
6385 // If element VT is == 32 bits and has 4 elems, try to generate an INSERTPS
6386 if (EVTBits == 32 && NumElems == 4)
6387 if (SDValue V = LowerBuildVectorv4x32(Op, DAG, Subtarget, *this))
6390 // If element VT is == 32 bits, turn it into a number of shuffles.
6391 SmallVector<SDValue, 8> V(NumElems);
6392 if (NumElems == 4 && NumZero > 0) {
6393 for (unsigned i = 0; i < 4; ++i) {
6394 bool isZero = !(NonZeros & (1 << i));
6396 V[i] = getZeroVector(VT, Subtarget, DAG, dl);
6398 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(i));
6401 for (unsigned i = 0; i < 2; ++i) {
6402 switch ((NonZeros & (0x3 << i*2)) >> (i*2)) {
6405 V[i] = V[i*2]; // Must be a zero vector.
6408 V[i] = getMOVL(DAG, dl, VT, V[i*2+1], V[i*2]);
6411 V[i] = getMOVL(DAG, dl, VT, V[i*2], V[i*2+1]);
6414 V[i] = getUnpackl(DAG, dl, VT, V[i*2], V[i*2+1]);
6419 bool Reverse1 = (NonZeros & 0x3) == 2;
6420 bool Reverse2 = ((NonZeros & (0x3 << 2)) >> 2) == 2;
6424 static_cast<int>(Reverse2 ? NumElems+1 : NumElems),
6425 static_cast<int>(Reverse2 ? NumElems : NumElems+1)
6427 return DAG.getVectorShuffle(VT, dl, V[0], V[1], &MaskVec[0]);
6430 if (Values.size() > 1 && VT.is128BitVector()) {
6431 // Check for a build vector of consecutive loads.
6432 for (unsigned i = 0; i < NumElems; ++i)
6433 V[i] = Op.getOperand(i);
6435 // Check for elements which are consecutive loads.
6436 if (SDValue LD = EltsFromConsecutiveLoads(VT, V, dl, DAG, false))
6439 // Check for a build vector from mostly shuffle plus few inserting.
6440 if (SDValue Sh = buildFromShuffleMostly(Op, DAG))
6443 // For SSE 4.1, use insertps to put the high elements into the low element.
6444 if (Subtarget->hasSSE41()) {
6446 if (Op.getOperand(0).getOpcode() != ISD::UNDEF)
6447 Result = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(0));
6449 Result = DAG.getUNDEF(VT);
6451 for (unsigned i = 1; i < NumElems; ++i) {
6452 if (Op.getOperand(i).getOpcode() == ISD::UNDEF) continue;
6453 Result = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, Result,
6454 Op.getOperand(i), DAG.getIntPtrConstant(i, dl));
6459 // Otherwise, expand into a number of unpckl*, start by extending each of
6460 // our (non-undef) elements to the full vector width with the element in the
6461 // bottom slot of the vector (which generates no code for SSE).
6462 for (unsigned i = 0; i < NumElems; ++i) {
6463 if (Op.getOperand(i).getOpcode() != ISD::UNDEF)
6464 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(i));
6466 V[i] = DAG.getUNDEF(VT);
6469 // Next, we iteratively mix elements, e.g. for v4f32:
6470 // Step 1: unpcklps 0, 2 ==> X: <?, ?, 2, 0>
6471 // : unpcklps 1, 3 ==> Y: <?, ?, 3, 1>
6472 // Step 2: unpcklps X, Y ==> <3, 2, 1, 0>
6473 unsigned EltStride = NumElems >> 1;
6474 while (EltStride != 0) {
6475 for (unsigned i = 0; i < EltStride; ++i) {
6476 // If V[i+EltStride] is undef and this is the first round of mixing,
6477 // then it is safe to just drop this shuffle: V[i] is already in the
6478 // right place, the one element (since it's the first round) being
6479 // inserted as undef can be dropped. This isn't safe for successive
6480 // rounds because they will permute elements within both vectors.
6481 if (V[i+EltStride].getOpcode() == ISD::UNDEF &&
6482 EltStride == NumElems/2)
6485 V[i] = getUnpackl(DAG, dl, VT, V[i], V[i + EltStride]);
6494 // 256-bit AVX can use the vinsertf128 instruction
6495 // to create 256-bit vectors from two other 128-bit ones.
6496 static SDValue LowerAVXCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) {
6498 MVT ResVT = Op.getSimpleValueType();
6500 assert((ResVT.is256BitVector() ||
6501 ResVT.is512BitVector()) && "Value type must be 256-/512-bit wide");
6503 SDValue V1 = Op.getOperand(0);
6504 SDValue V2 = Op.getOperand(1);
6505 unsigned NumElems = ResVT.getVectorNumElements();
6506 if (ResVT.is256BitVector())
6507 return Concat128BitVectors(V1, V2, ResVT, NumElems, DAG, dl);
6509 if (Op.getNumOperands() == 4) {
6510 MVT HalfVT = MVT::getVectorVT(ResVT.getScalarType(),
6511 ResVT.getVectorNumElements()/2);
6512 SDValue V3 = Op.getOperand(2);
6513 SDValue V4 = Op.getOperand(3);
6514 return Concat256BitVectors(Concat128BitVectors(V1, V2, HalfVT, NumElems/2, DAG, dl),
6515 Concat128BitVectors(V3, V4, HalfVT, NumElems/2, DAG, dl), ResVT, NumElems, DAG, dl);
6517 return Concat256BitVectors(V1, V2, ResVT, NumElems, DAG, dl);
6520 static SDValue LowerCONCAT_VECTORSvXi1(SDValue Op,
6521 const X86Subtarget *Subtarget,
6522 SelectionDAG & DAG) {
6524 MVT ResVT = Op.getSimpleValueType();
6525 unsigned NumOfOperands = Op.getNumOperands();
6527 assert(isPowerOf2_32(NumOfOperands) &&
6528 "Unexpected number of operands in CONCAT_VECTORS");
6530 if (NumOfOperands > 2) {
6531 MVT HalfVT = MVT::getVectorVT(ResVT.getScalarType(),
6532 ResVT.getVectorNumElements()/2);
6533 SmallVector<SDValue, 2> Ops;
6534 for (unsigned i = 0; i < NumOfOperands/2; i++)
6535 Ops.push_back(Op.getOperand(i));
6536 SDValue Lo = DAG.getNode(ISD::CONCAT_VECTORS, dl, HalfVT, Ops);
6538 for (unsigned i = NumOfOperands/2; i < NumOfOperands; i++)
6539 Ops.push_back(Op.getOperand(i));
6540 SDValue Hi = DAG.getNode(ISD::CONCAT_VECTORS, dl, HalfVT, Ops);
6541 return DAG.getNode(ISD::CONCAT_VECTORS, dl, ResVT, Lo, Hi);
6544 SDValue V1 = Op.getOperand(0);
6545 SDValue V2 = Op.getOperand(1);
6546 bool IsZeroV1 = ISD::isBuildVectorAllZeros(V1.getNode());
6547 bool IsZeroV2 = ISD::isBuildVectorAllZeros(V2.getNode());
6549 if (IsZeroV1 && IsZeroV2)
6550 return getZeroVector(ResVT, Subtarget, DAG, dl);
6552 SDValue ZeroIdx = DAG.getIntPtrConstant(0, dl);
6553 SDValue Undef = DAG.getUNDEF(ResVT);
6554 unsigned NumElems = ResVT.getVectorNumElements();
6555 SDValue ShiftBits = DAG.getConstant(NumElems/2, dl, MVT::i8);
6557 V2 = DAG.getNode(ISD::INSERT_SUBVECTOR, dl, ResVT, Undef, V2, ZeroIdx);
6558 V2 = DAG.getNode(X86ISD::VSHLI, dl, ResVT, V2, ShiftBits);
6562 V1 = DAG.getNode(ISD::INSERT_SUBVECTOR, dl, ResVT, Undef, V1, ZeroIdx);
6563 // Zero the upper bits of V1
6564 V1 = DAG.getNode(X86ISD::VSHLI, dl, ResVT, V1, ShiftBits);
6565 V1 = DAG.getNode(X86ISD::VSRLI, dl, ResVT, V1, ShiftBits);
6568 return DAG.getNode(ISD::OR, dl, ResVT, V1, V2);
6571 static SDValue LowerCONCAT_VECTORS(SDValue Op,
6572 const X86Subtarget *Subtarget,
6573 SelectionDAG &DAG) {
6574 MVT VT = Op.getSimpleValueType();
6575 if (VT.getVectorElementType() == MVT::i1)
6576 return LowerCONCAT_VECTORSvXi1(Op, Subtarget, DAG);
6578 assert((VT.is256BitVector() && Op.getNumOperands() == 2) ||
6579 (VT.is512BitVector() && (Op.getNumOperands() == 2 ||
6580 Op.getNumOperands() == 4)));
6582 // AVX can use the vinsertf128 instruction to create 256-bit vectors
6583 // from two other 128-bit ones.
6585 // 512-bit vector may contain 2 256-bit vectors or 4 128-bit vectors
6586 return LowerAVXCONCAT_VECTORS(Op, DAG);
6589 //===----------------------------------------------------------------------===//
6590 // Vector shuffle lowering
6592 // This is an experimental code path for lowering vector shuffles on x86. It is
6593 // designed to handle arbitrary vector shuffles and blends, gracefully
6594 // degrading performance as necessary. It works hard to recognize idiomatic
6595 // shuffles and lower them to optimal instruction patterns without leaving
6596 // a framework that allows reasonably efficient handling of all vector shuffle
6598 //===----------------------------------------------------------------------===//
6600 /// \brief Tiny helper function to identify a no-op mask.
6602 /// This is a somewhat boring predicate function. It checks whether the mask
6603 /// array input, which is assumed to be a single-input shuffle mask of the kind
6604 /// used by the X86 shuffle instructions (not a fully general
6605 /// ShuffleVectorSDNode mask) requires any shuffles to occur. Both undef and an
6606 /// in-place shuffle are 'no-op's.
6607 static bool isNoopShuffleMask(ArrayRef<int> Mask) {
6608 for (int i = 0, Size = Mask.size(); i < Size; ++i)
6609 if (Mask[i] != -1 && Mask[i] != i)
6614 /// \brief Helper function to classify a mask as a single-input mask.
6616 /// This isn't a generic single-input test because in the vector shuffle
6617 /// lowering we canonicalize single inputs to be the first input operand. This
6618 /// means we can more quickly test for a single input by only checking whether
6619 /// an input from the second operand exists. We also assume that the size of
6620 /// mask corresponds to the size of the input vectors which isn't true in the
6621 /// fully general case.
6622 static bool isSingleInputShuffleMask(ArrayRef<int> Mask) {
6624 if (M >= (int)Mask.size())
6629 /// \brief Test whether there are elements crossing 128-bit lanes in this
6632 /// X86 divides up its shuffles into in-lane and cross-lane shuffle operations
6633 /// and we routinely test for these.
6634 static bool is128BitLaneCrossingShuffleMask(MVT VT, ArrayRef<int> Mask) {
6635 int LaneSize = 128 / VT.getScalarSizeInBits();
6636 int Size = Mask.size();
6637 for (int i = 0; i < Size; ++i)
6638 if (Mask[i] >= 0 && (Mask[i] % Size) / LaneSize != i / LaneSize)
6643 /// \brief Test whether a shuffle mask is equivalent within each 128-bit lane.
6645 /// This checks a shuffle mask to see if it is performing the same
6646 /// 128-bit lane-relative shuffle in each 128-bit lane. This trivially implies
6647 /// that it is also not lane-crossing. It may however involve a blend from the
6648 /// same lane of a second vector.
6650 /// The specific repeated shuffle mask is populated in \p RepeatedMask, as it is
6651 /// non-trivial to compute in the face of undef lanes. The representation is
6652 /// *not* suitable for use with existing 128-bit shuffles as it will contain
6653 /// entries from both V1 and V2 inputs to the wider mask.
6655 is128BitLaneRepeatedShuffleMask(MVT VT, ArrayRef<int> Mask,
6656 SmallVectorImpl<int> &RepeatedMask) {
6657 int LaneSize = 128 / VT.getScalarSizeInBits();
6658 RepeatedMask.resize(LaneSize, -1);
6659 int Size = Mask.size();
6660 for (int i = 0; i < Size; ++i) {
6663 if ((Mask[i] % Size) / LaneSize != i / LaneSize)
6664 // This entry crosses lanes, so there is no way to model this shuffle.
6667 // Ok, handle the in-lane shuffles by detecting if and when they repeat.
6668 if (RepeatedMask[i % LaneSize] == -1)
6669 // This is the first non-undef entry in this slot of a 128-bit lane.
6670 RepeatedMask[i % LaneSize] =
6671 Mask[i] < Size ? Mask[i] % LaneSize : Mask[i] % LaneSize + Size;
6672 else if (RepeatedMask[i % LaneSize] + (i / LaneSize) * LaneSize != Mask[i])
6673 // Found a mismatch with the repeated mask.
6679 /// \brief Checks whether a shuffle mask is equivalent to an explicit list of
6682 /// This is a fast way to test a shuffle mask against a fixed pattern:
6684 /// if (isShuffleEquivalent(Mask, 3, 2, {1, 0})) { ... }
6686 /// It returns true if the mask is exactly as wide as the argument list, and
6687 /// each element of the mask is either -1 (signifying undef) or the value given
6688 /// in the argument.
6689 static bool isShuffleEquivalent(SDValue V1, SDValue V2, ArrayRef<int> Mask,
6690 ArrayRef<int> ExpectedMask) {
6691 if (Mask.size() != ExpectedMask.size())
6694 int Size = Mask.size();
6696 // If the values are build vectors, we can look through them to find
6697 // equivalent inputs that make the shuffles equivalent.
6698 auto *BV1 = dyn_cast<BuildVectorSDNode>(V1);
6699 auto *BV2 = dyn_cast<BuildVectorSDNode>(V2);
6701 for (int i = 0; i < Size; ++i)
6702 if (Mask[i] != -1 && Mask[i] != ExpectedMask[i]) {
6703 auto *MaskBV = Mask[i] < Size ? BV1 : BV2;
6704 auto *ExpectedBV = ExpectedMask[i] < Size ? BV1 : BV2;
6705 if (!MaskBV || !ExpectedBV ||
6706 MaskBV->getOperand(Mask[i] % Size) !=
6707 ExpectedBV->getOperand(ExpectedMask[i] % Size))
6714 /// \brief Get a 4-lane 8-bit shuffle immediate for a mask.
6716 /// This helper function produces an 8-bit shuffle immediate corresponding to
6717 /// the ubiquitous shuffle encoding scheme used in x86 instructions for
6718 /// shuffling 4 lanes. It can be used with most of the PSHUF instructions for
6721 /// NB: We rely heavily on "undef" masks preserving the input lane.
6722 static SDValue getV4X86ShuffleImm8ForMask(ArrayRef<int> Mask, SDLoc DL,
6723 SelectionDAG &DAG) {
6724 assert(Mask.size() == 4 && "Only 4-lane shuffle masks");
6725 assert(Mask[0] >= -1 && Mask[0] < 4 && "Out of bound mask element!");
6726 assert(Mask[1] >= -1 && Mask[1] < 4 && "Out of bound mask element!");
6727 assert(Mask[2] >= -1 && Mask[2] < 4 && "Out of bound mask element!");
6728 assert(Mask[3] >= -1 && Mask[3] < 4 && "Out of bound mask element!");
6731 Imm |= (Mask[0] == -1 ? 0 : Mask[0]) << 0;
6732 Imm |= (Mask[1] == -1 ? 1 : Mask[1]) << 2;
6733 Imm |= (Mask[2] == -1 ? 2 : Mask[2]) << 4;
6734 Imm |= (Mask[3] == -1 ? 3 : Mask[3]) << 6;
6735 return DAG.getConstant(Imm, DL, MVT::i8);
6738 /// \brief Compute whether each element of a shuffle is zeroable.
6740 /// A "zeroable" vector shuffle element is one which can be lowered to zero.
6741 /// Either it is an undef element in the shuffle mask, the element of the input
6742 /// referenced is undef, or the element of the input referenced is known to be
6743 /// zero. Many x86 shuffles can zero lanes cheaply and we often want to handle
6744 /// as many lanes with this technique as possible to simplify the remaining
6746 static SmallBitVector computeZeroableShuffleElements(ArrayRef<int> Mask,
6747 SDValue V1, SDValue V2) {
6748 SmallBitVector Zeroable(Mask.size(), false);
6750 while (V1.getOpcode() == ISD::BITCAST)
6751 V1 = V1->getOperand(0);
6752 while (V2.getOpcode() == ISD::BITCAST)
6753 V2 = V2->getOperand(0);
6755 bool V1IsZero = ISD::isBuildVectorAllZeros(V1.getNode());
6756 bool V2IsZero = ISD::isBuildVectorAllZeros(V2.getNode());
6758 for (int i = 0, Size = Mask.size(); i < Size; ++i) {
6760 // Handle the easy cases.
6761 if (M < 0 || (M >= 0 && M < Size && V1IsZero) || (M >= Size && V2IsZero)) {
6766 // If this is an index into a build_vector node (which has the same number
6767 // of elements), dig out the input value and use it.
6768 SDValue V = M < Size ? V1 : V2;
6769 if (V.getOpcode() != ISD::BUILD_VECTOR || Size != (int)V.getNumOperands())
6772 SDValue Input = V.getOperand(M % Size);
6773 // The UNDEF opcode check really should be dead code here, but not quite
6774 // worth asserting on (it isn't invalid, just unexpected).
6775 if (Input.getOpcode() == ISD::UNDEF || X86::isZeroNode(Input))
6782 // X86 has dedicated unpack instructions that can handle specific blend
6783 // operations: UNPCKH and UNPCKL.
6784 static SDValue lowerVectorShuffleWithUNPCK(SDLoc DL, MVT VT, ArrayRef<int> Mask,
6785 SDValue V1, SDValue V2,
6786 SelectionDAG &DAG) {
6787 int NumElts = VT.getVectorNumElements();
6790 bool UnpcklSwapped = true;
6791 bool UnpckhSwapped = true;
6792 int NumEltsInLane = 128 / VT.getScalarSizeInBits();
6794 for (int i = 0; i < NumElts; ++i) {
6795 unsigned LaneStart = (i / NumEltsInLane) * NumEltsInLane;
6797 int LoPos = (i % NumEltsInLane) / 2 + LaneStart + NumElts * (i % 2);
6798 int HiPos = LoPos + NumEltsInLane / 2;
6799 int LoPosSwapped = (LoPos + NumElts) % (NumElts * 2);
6800 int HiPosSwapped = (HiPos + NumElts) % (NumElts * 2);
6804 if (Mask[i] != LoPos)
6806 if (Mask[i] != HiPos)
6808 if (Mask[i] != LoPosSwapped)
6809 UnpcklSwapped = false;
6810 if (Mask[i] != HiPosSwapped)
6811 UnpckhSwapped = false;
6812 if (!Unpckl && !Unpckh && !UnpcklSwapped && !UnpckhSwapped)
6816 return DAG.getNode(X86ISD::UNPCKL, DL, VT, V1, V2);
6818 return DAG.getNode(X86ISD::UNPCKH, DL, VT, V1, V2);
6820 return DAG.getNode(X86ISD::UNPCKL, DL, VT, V2, V1);
6822 return DAG.getNode(X86ISD::UNPCKH, DL, VT, V2, V1);
6824 llvm_unreachable("Unexpected result of UNPCK mask analysis");
6828 /// \brief Try to emit a bitmask instruction for a shuffle.
6830 /// This handles cases where we can model a blend exactly as a bitmask due to
6831 /// one of the inputs being zeroable.
6832 static SDValue lowerVectorShuffleAsBitMask(SDLoc DL, MVT VT, SDValue V1,
6833 SDValue V2, ArrayRef<int> Mask,
6834 SelectionDAG &DAG) {
6835 MVT EltVT = VT.getScalarType();
6836 int NumEltBits = EltVT.getSizeInBits();
6837 MVT IntEltVT = MVT::getIntegerVT(NumEltBits);
6838 SDValue Zero = DAG.getConstant(0, DL, IntEltVT);
6839 SDValue AllOnes = DAG.getConstant(APInt::getAllOnesValue(NumEltBits), DL,
6841 if (EltVT.isFloatingPoint()) {
6842 Zero = DAG.getBitcast(EltVT, Zero);
6843 AllOnes = DAG.getBitcast(EltVT, AllOnes);
6845 SmallVector<SDValue, 16> VMaskOps(Mask.size(), Zero);
6846 SmallBitVector Zeroable = computeZeroableShuffleElements(Mask, V1, V2);
6848 for (int i = 0, Size = Mask.size(); i < Size; ++i) {
6851 if (Mask[i] % Size != i)
6852 return SDValue(); // Not a blend.
6854 V = Mask[i] < Size ? V1 : V2;
6855 else if (V != (Mask[i] < Size ? V1 : V2))
6856 return SDValue(); // Can only let one input through the mask.
6858 VMaskOps[i] = AllOnes;
6861 return SDValue(); // No non-zeroable elements!
6863 SDValue VMask = DAG.getNode(ISD::BUILD_VECTOR, DL, VT, VMaskOps);
6864 V = DAG.getNode(VT.isFloatingPoint()
6865 ? (unsigned) X86ISD::FAND : (unsigned) ISD::AND,
6870 /// \brief Try to emit a blend instruction for a shuffle using bit math.
6872 /// This is used as a fallback approach when first class blend instructions are
6873 /// unavailable. Currently it is only suitable for integer vectors, but could
6874 /// be generalized for floating point vectors if desirable.
6875 static SDValue lowerVectorShuffleAsBitBlend(SDLoc DL, MVT VT, SDValue V1,
6876 SDValue V2, ArrayRef<int> Mask,
6877 SelectionDAG &DAG) {
6878 assert(VT.isInteger() && "Only supports integer vector types!");
6879 MVT EltVT = VT.getScalarType();
6880 int NumEltBits = EltVT.getSizeInBits();
6881 SDValue Zero = DAG.getConstant(0, DL, EltVT);
6882 SDValue AllOnes = DAG.getConstant(APInt::getAllOnesValue(NumEltBits), DL,
6884 SmallVector<SDValue, 16> MaskOps;
6885 for (int i = 0, Size = Mask.size(); i < Size; ++i) {
6886 if (Mask[i] != -1 && Mask[i] != i && Mask[i] != i + Size)
6887 return SDValue(); // Shuffled input!
6888 MaskOps.push_back(Mask[i] < Size ? AllOnes : Zero);
6891 SDValue V1Mask = DAG.getNode(ISD::BUILD_VECTOR, DL, VT, MaskOps);
6892 V1 = DAG.getNode(ISD::AND, DL, VT, V1, V1Mask);
6893 // We have to cast V2 around.
6894 MVT MaskVT = MVT::getVectorVT(MVT::i64, VT.getSizeInBits() / 64);
6895 V2 = DAG.getBitcast(VT, DAG.getNode(X86ISD::ANDNP, DL, MaskVT,
6896 DAG.getBitcast(MaskVT, V1Mask),
6897 DAG.getBitcast(MaskVT, V2)));
6898 return DAG.getNode(ISD::OR, DL, VT, V1, V2);
6901 /// \brief Try to emit a blend instruction for a shuffle.
6903 /// This doesn't do any checks for the availability of instructions for blending
6904 /// these values. It relies on the availability of the X86ISD::BLENDI pattern to
6905 /// be matched in the backend with the type given. What it does check for is
6906 /// that the shuffle mask is in fact a blend.
6907 static SDValue lowerVectorShuffleAsBlend(SDLoc DL, MVT VT, SDValue V1,
6908 SDValue V2, ArrayRef<int> Mask,
6909 const X86Subtarget *Subtarget,
6910 SelectionDAG &DAG) {
6911 unsigned BlendMask = 0;
6912 for (int i = 0, Size = Mask.size(); i < Size; ++i) {
6913 if (Mask[i] >= Size) {
6914 if (Mask[i] != i + Size)
6915 return SDValue(); // Shuffled V2 input!
6916 BlendMask |= 1u << i;
6919 if (Mask[i] >= 0 && Mask[i] != i)
6920 return SDValue(); // Shuffled V1 input!
6922 switch (VT.SimpleTy) {
6927 return DAG.getNode(X86ISD::BLENDI, DL, VT, V1, V2,
6928 DAG.getConstant(BlendMask, DL, MVT::i8));
6932 assert(Subtarget->hasAVX2() && "256-bit integer blends require AVX2!");
6936 // If we have AVX2 it is faster to use VPBLENDD when the shuffle fits into
6937 // that instruction.
6938 if (Subtarget->hasAVX2()) {
6939 // Scale the blend by the number of 32-bit dwords per element.
6940 int Scale = VT.getScalarSizeInBits() / 32;
6942 for (int i = 0, Size = Mask.size(); i < Size; ++i)
6943 if (Mask[i] >= Size)
6944 for (int j = 0; j < Scale; ++j)
6945 BlendMask |= 1u << (i * Scale + j);
6947 MVT BlendVT = VT.getSizeInBits() > 128 ? MVT::v8i32 : MVT::v4i32;
6948 V1 = DAG.getBitcast(BlendVT, V1);
6949 V2 = DAG.getBitcast(BlendVT, V2);
6950 return DAG.getBitcast(
6951 VT, DAG.getNode(X86ISD::BLENDI, DL, BlendVT, V1, V2,
6952 DAG.getConstant(BlendMask, DL, MVT::i8)));
6956 // For integer shuffles we need to expand the mask and cast the inputs to
6957 // v8i16s prior to blending.
6958 int Scale = 8 / VT.getVectorNumElements();
6960 for (int i = 0, Size = Mask.size(); i < Size; ++i)
6961 if (Mask[i] >= Size)
6962 for (int j = 0; j < Scale; ++j)
6963 BlendMask |= 1u << (i * Scale + j);
6965 V1 = DAG.getBitcast(MVT::v8i16, V1);
6966 V2 = DAG.getBitcast(MVT::v8i16, V2);
6967 return DAG.getBitcast(VT,
6968 DAG.getNode(X86ISD::BLENDI, DL, MVT::v8i16, V1, V2,
6969 DAG.getConstant(BlendMask, DL, MVT::i8)));
6973 assert(Subtarget->hasAVX2() && "256-bit integer blends require AVX2!");
6974 SmallVector<int, 8> RepeatedMask;
6975 if (is128BitLaneRepeatedShuffleMask(MVT::v16i16, Mask, RepeatedMask)) {
6976 // We can lower these with PBLENDW which is mirrored across 128-bit lanes.
6977 assert(RepeatedMask.size() == 8 && "Repeated mask size doesn't match!");
6979 for (int i = 0; i < 8; ++i)
6980 if (RepeatedMask[i] >= 16)
6981 BlendMask |= 1u << i;
6982 return DAG.getNode(X86ISD::BLENDI, DL, MVT::v16i16, V1, V2,
6983 DAG.getConstant(BlendMask, DL, MVT::i8));
6989 assert((VT.getSizeInBits() == 128 || Subtarget->hasAVX2()) &&
6990 "256-bit byte-blends require AVX2 support!");
6992 // Attempt to lower to a bitmask if we can. VPAND is faster than VPBLENDVB.
6993 if (SDValue Masked = lowerVectorShuffleAsBitMask(DL, VT, V1, V2, Mask, DAG))
6996 // Scale the blend by the number of bytes per element.
6997 int Scale = VT.getScalarSizeInBits() / 8;
6999 // This form of blend is always done on bytes. Compute the byte vector
7001 MVT BlendVT = MVT::getVectorVT(MVT::i8, VT.getSizeInBits() / 8);
7003 // Compute the VSELECT mask. Note that VSELECT is really confusing in the
7004 // mix of LLVM's code generator and the x86 backend. We tell the code
7005 // generator that boolean values in the elements of an x86 vector register
7006 // are -1 for true and 0 for false. We then use the LLVM semantics of 'true'
7007 // mapping a select to operand #1, and 'false' mapping to operand #2. The
7008 // reality in x86 is that vector masks (pre-AVX-512) use only the high bit
7009 // of the element (the remaining are ignored) and 0 in that high bit would
7010 // mean operand #1 while 1 in the high bit would mean operand #2. So while
7011 // the LLVM model for boolean values in vector elements gets the relevant
7012 // bit set, it is set backwards and over constrained relative to x86's
7014 SmallVector<SDValue, 32> VSELECTMask;
7015 for (int i = 0, Size = Mask.size(); i < Size; ++i)
7016 for (int j = 0; j < Scale; ++j)
7017 VSELECTMask.push_back(
7018 Mask[i] < 0 ? DAG.getUNDEF(MVT::i8)
7019 : DAG.getConstant(Mask[i] < Size ? -1 : 0, DL,
7022 V1 = DAG.getBitcast(BlendVT, V1);
7023 V2 = DAG.getBitcast(BlendVT, V2);
7024 return DAG.getBitcast(VT, DAG.getNode(ISD::VSELECT, DL, BlendVT,
7025 DAG.getNode(ISD::BUILD_VECTOR, DL,
7026 BlendVT, VSELECTMask),
7031 llvm_unreachable("Not a supported integer vector type!");
7035 /// \brief Try to lower as a blend of elements from two inputs followed by
7036 /// a single-input permutation.
7038 /// This matches the pattern where we can blend elements from two inputs and
7039 /// then reduce the shuffle to a single-input permutation.
7040 static SDValue lowerVectorShuffleAsBlendAndPermute(SDLoc DL, MVT VT, SDValue V1,
7043 SelectionDAG &DAG) {
7044 // We build up the blend mask while checking whether a blend is a viable way
7045 // to reduce the shuffle.
7046 SmallVector<int, 32> BlendMask(Mask.size(), -1);
7047 SmallVector<int, 32> PermuteMask(Mask.size(), -1);
7049 for (int i = 0, Size = Mask.size(); i < Size; ++i) {
7053 assert(Mask[i] < Size * 2 && "Shuffle input is out of bounds.");
7055 if (BlendMask[Mask[i] % Size] == -1)
7056 BlendMask[Mask[i] % Size] = Mask[i];
7057 else if (BlendMask[Mask[i] % Size] != Mask[i])
7058 return SDValue(); // Can't blend in the needed input!
7060 PermuteMask[i] = Mask[i] % Size;
7063 SDValue V = DAG.getVectorShuffle(VT, DL, V1, V2, BlendMask);
7064 return DAG.getVectorShuffle(VT, DL, V, DAG.getUNDEF(VT), PermuteMask);
7067 /// \brief Generic routine to decompose a shuffle and blend into indepndent
7068 /// blends and permutes.
7070 /// This matches the extremely common pattern for handling combined
7071 /// shuffle+blend operations on newer X86 ISAs where we have very fast blend
7072 /// operations. It will try to pick the best arrangement of shuffles and
7074 static SDValue lowerVectorShuffleAsDecomposedShuffleBlend(SDLoc DL, MVT VT,
7078 SelectionDAG &DAG) {
7079 // Shuffle the input elements into the desired positions in V1 and V2 and
7080 // blend them together.
7081 SmallVector<int, 32> V1Mask(Mask.size(), -1);
7082 SmallVector<int, 32> V2Mask(Mask.size(), -1);
7083 SmallVector<int, 32> BlendMask(Mask.size(), -1);
7084 for (int i = 0, Size = Mask.size(); i < Size; ++i)
7085 if (Mask[i] >= 0 && Mask[i] < Size) {
7086 V1Mask[i] = Mask[i];
7088 } else if (Mask[i] >= Size) {
7089 V2Mask[i] = Mask[i] - Size;
7090 BlendMask[i] = i + Size;
7093 // Try to lower with the simpler initial blend strategy unless one of the
7094 // input shuffles would be a no-op. We prefer to shuffle inputs as the
7095 // shuffle may be able to fold with a load or other benefit. However, when
7096 // we'll have to do 2x as many shuffles in order to achieve this, blending
7097 // first is a better strategy.
7098 if (!isNoopShuffleMask(V1Mask) && !isNoopShuffleMask(V2Mask))
7099 if (SDValue BlendPerm =
7100 lowerVectorShuffleAsBlendAndPermute(DL, VT, V1, V2, Mask, DAG))
7103 V1 = DAG.getVectorShuffle(VT, DL, V1, DAG.getUNDEF(VT), V1Mask);
7104 V2 = DAG.getVectorShuffle(VT, DL, V2, DAG.getUNDEF(VT), V2Mask);
7105 return DAG.getVectorShuffle(VT, DL, V1, V2, BlendMask);
7108 /// \brief Try to lower a vector shuffle as a byte rotation.
7110 /// SSSE3 has a generic PALIGNR instruction in x86 that will do an arbitrary
7111 /// byte-rotation of the concatenation of two vectors; pre-SSSE3 can use
7112 /// a PSRLDQ/PSLLDQ/POR pattern to get a similar effect. This routine will
7113 /// try to generically lower a vector shuffle through such an pattern. It
7114 /// does not check for the profitability of lowering either as PALIGNR or
7115 /// PSRLDQ/PSLLDQ/POR, only whether the mask is valid to lower in that form.
7116 /// This matches shuffle vectors that look like:
7118 /// v8i16 [11, 12, 13, 14, 15, 0, 1, 2]
7120 /// Essentially it concatenates V1 and V2, shifts right by some number of
7121 /// elements, and takes the low elements as the result. Note that while this is
7122 /// specified as a *right shift* because x86 is little-endian, it is a *left
7123 /// rotate* of the vector lanes.
7124 static SDValue lowerVectorShuffleAsByteRotate(SDLoc DL, MVT VT, SDValue V1,
7127 const X86Subtarget *Subtarget,
7128 SelectionDAG &DAG) {
7129 assert(!isNoopShuffleMask(Mask) && "We shouldn't lower no-op shuffles!");
7131 int NumElts = Mask.size();
7132 int NumLanes = VT.getSizeInBits() / 128;
7133 int NumLaneElts = NumElts / NumLanes;
7135 // We need to detect various ways of spelling a rotation:
7136 // [11, 12, 13, 14, 15, 0, 1, 2]
7137 // [-1, 12, 13, 14, -1, -1, 1, -1]
7138 // [-1, -1, -1, -1, -1, -1, 1, 2]
7139 // [ 3, 4, 5, 6, 7, 8, 9, 10]
7140 // [-1, 4, 5, 6, -1, -1, 9, -1]
7141 // [-1, 4, 5, 6, -1, -1, -1, -1]
7144 for (int l = 0; l < NumElts; l += NumLaneElts) {
7145 for (int i = 0; i < NumLaneElts; ++i) {
7146 if (Mask[l + i] == -1)
7148 assert(Mask[l + i] >= 0 && "Only -1 is a valid negative mask element!");
7150 // Get the mod-Size index and lane correct it.
7151 int LaneIdx = (Mask[l + i] % NumElts) - l;
7152 // Make sure it was in this lane.
7153 if (LaneIdx < 0 || LaneIdx >= NumLaneElts)
7156 // Determine where a rotated vector would have started.
7157 int StartIdx = i - LaneIdx;
7159 // The identity rotation isn't interesting, stop.
7162 // If we found the tail of a vector the rotation must be the missing
7163 // front. If we found the head of a vector, it must be how much of the
7165 int CandidateRotation = StartIdx < 0 ? -StartIdx : NumLaneElts - StartIdx;
7168 Rotation = CandidateRotation;
7169 else if (Rotation != CandidateRotation)
7170 // The rotations don't match, so we can't match this mask.
7173 // Compute which value this mask is pointing at.
7174 SDValue MaskV = Mask[l + i] < NumElts ? V1 : V2;
7176 // Compute which of the two target values this index should be assigned
7177 // to. This reflects whether the high elements are remaining or the low
7178 // elements are remaining.
7179 SDValue &TargetV = StartIdx < 0 ? Hi : Lo;
7181 // Either set up this value if we've not encountered it before, or check
7182 // that it remains consistent.
7185 else if (TargetV != MaskV)
7186 // This may be a rotation, but it pulls from the inputs in some
7187 // unsupported interleaving.
7192 // Check that we successfully analyzed the mask, and normalize the results.
7193 assert(Rotation != 0 && "Failed to locate a viable rotation!");
7194 assert((Lo || Hi) && "Failed to find a rotated input vector!");
7200 // The actual rotate instruction rotates bytes, so we need to scale the
7201 // rotation based on how many bytes are in the vector lane.
7202 int Scale = 16 / NumLaneElts;
7204 // SSSE3 targets can use the palignr instruction.
7205 if (Subtarget->hasSSSE3()) {
7206 // Cast the inputs to i8 vector of correct length to match PALIGNR.
7207 MVT AlignVT = MVT::getVectorVT(MVT::i8, 16 * NumLanes);
7208 Lo = DAG.getBitcast(AlignVT, Lo);
7209 Hi = DAG.getBitcast(AlignVT, Hi);
7211 return DAG.getBitcast(
7212 VT, DAG.getNode(X86ISD::PALIGNR, DL, AlignVT, Lo, Hi,
7213 DAG.getConstant(Rotation * Scale, DL, MVT::i8)));
7216 assert(VT.getSizeInBits() == 128 &&
7217 "Rotate-based lowering only supports 128-bit lowering!");
7218 assert(Mask.size() <= 16 &&
7219 "Can shuffle at most 16 bytes in a 128-bit vector!");
7221 // Default SSE2 implementation
7222 int LoByteShift = 16 - Rotation * Scale;
7223 int HiByteShift = Rotation * Scale;
7225 // Cast the inputs to v2i64 to match PSLLDQ/PSRLDQ.
7226 Lo = DAG.getBitcast(MVT::v2i64, Lo);
7227 Hi = DAG.getBitcast(MVT::v2i64, Hi);
7229 SDValue LoShift = DAG.getNode(X86ISD::VSHLDQ, DL, MVT::v2i64, Lo,
7230 DAG.getConstant(LoByteShift, DL, MVT::i8));
7231 SDValue HiShift = DAG.getNode(X86ISD::VSRLDQ, DL, MVT::v2i64, Hi,
7232 DAG.getConstant(HiByteShift, DL, MVT::i8));
7233 return DAG.getBitcast(VT,
7234 DAG.getNode(ISD::OR, DL, MVT::v2i64, LoShift, HiShift));
7237 /// \brief Try to lower a vector shuffle as a bit shift (shifts in zeros).
7239 /// Attempts to match a shuffle mask against the PSLL(W/D/Q/DQ) and
7240 /// PSRL(W/D/Q/DQ) SSE2 and AVX2 logical bit-shift instructions. The function
7241 /// matches elements from one of the input vectors shuffled to the left or
7242 /// right with zeroable elements 'shifted in'. It handles both the strictly
7243 /// bit-wise element shifts and the byte shift across an entire 128-bit double
7246 /// PSHL : (little-endian) left bit shift.
7247 /// [ zz, 0, zz, 2 ]
7248 /// [ -1, 4, zz, -1 ]
7249 /// PSRL : (little-endian) right bit shift.
7251 /// [ -1, -1, 7, zz]
7252 /// PSLLDQ : (little-endian) left byte shift
7253 /// [ zz, 0, 1, 2, 3, 4, 5, 6]
7254 /// [ zz, zz, -1, -1, 2, 3, 4, -1]
7255 /// [ zz, zz, zz, zz, zz, zz, -1, 1]
7256 /// PSRLDQ : (little-endian) right byte shift
7257 /// [ 5, 6, 7, zz, zz, zz, zz, zz]
7258 /// [ -1, 5, 6, 7, zz, zz, zz, zz]
7259 /// [ 1, 2, -1, -1, -1, -1, zz, zz]
7260 static SDValue lowerVectorShuffleAsShift(SDLoc DL, MVT VT, SDValue V1,
7261 SDValue V2, ArrayRef<int> Mask,
7262 SelectionDAG &DAG) {
7263 SmallBitVector Zeroable = computeZeroableShuffleElements(Mask, V1, V2);
7265 int Size = Mask.size();
7266 assert(Size == (int)VT.getVectorNumElements() && "Unexpected mask size");
7268 auto CheckZeros = [&](int Shift, int Scale, bool Left) {
7269 for (int i = 0; i < Size; i += Scale)
7270 for (int j = 0; j < Shift; ++j)
7271 if (!Zeroable[i + j + (Left ? 0 : (Scale - Shift))])
7277 auto MatchShift = [&](int Shift, int Scale, bool Left, SDValue V) {
7278 for (int i = 0; i != Size; i += Scale) {
7279 unsigned Pos = Left ? i + Shift : i;
7280 unsigned Low = Left ? i : i + Shift;
7281 unsigned Len = Scale - Shift;
7282 if (!isSequentialOrUndefInRange(Mask, Pos, Len,
7283 Low + (V == V1 ? 0 : Size)))
7287 int ShiftEltBits = VT.getScalarSizeInBits() * Scale;
7288 bool ByteShift = ShiftEltBits > 64;
7289 unsigned OpCode = Left ? (ByteShift ? X86ISD::VSHLDQ : X86ISD::VSHLI)
7290 : (ByteShift ? X86ISD::VSRLDQ : X86ISD::VSRLI);
7291 int ShiftAmt = Shift * VT.getScalarSizeInBits() / (ByteShift ? 8 : 1);
7293 // Normalize the scale for byte shifts to still produce an i64 element
7295 Scale = ByteShift ? Scale / 2 : Scale;
7297 // We need to round trip through the appropriate type for the shift.
7298 MVT ShiftSVT = MVT::getIntegerVT(VT.getScalarSizeInBits() * Scale);
7299 MVT ShiftVT = MVT::getVectorVT(ShiftSVT, Size / Scale);
7300 assert(DAG.getTargetLoweringInfo().isTypeLegal(ShiftVT) &&
7301 "Illegal integer vector type");
7302 V = DAG.getBitcast(ShiftVT, V);
7304 V = DAG.getNode(OpCode, DL, ShiftVT, V,
7305 DAG.getConstant(ShiftAmt, DL, MVT::i8));
7306 return DAG.getBitcast(VT, V);
7309 // SSE/AVX supports logical shifts up to 64-bit integers - so we can just
7310 // keep doubling the size of the integer elements up to that. We can
7311 // then shift the elements of the integer vector by whole multiples of
7312 // their width within the elements of the larger integer vector. Test each
7313 // multiple to see if we can find a match with the moved element indices
7314 // and that the shifted in elements are all zeroable.
7315 for (int Scale = 2; Scale * VT.getScalarSizeInBits() <= 128; Scale *= 2)
7316 for (int Shift = 1; Shift != Scale; ++Shift)
7317 for (bool Left : {true, false})
7318 if (CheckZeros(Shift, Scale, Left))
7319 for (SDValue V : {V1, V2})
7320 if (SDValue Match = MatchShift(Shift, Scale, Left, V))
7327 /// \brief Try to lower a vector shuffle using SSE4a EXTRQ/INSERTQ.
7328 static SDValue lowerVectorShuffleWithSSE4A(SDLoc DL, MVT VT, SDValue V1,
7329 SDValue V2, ArrayRef<int> Mask,
7330 SelectionDAG &DAG) {
7331 SmallBitVector Zeroable = computeZeroableShuffleElements(Mask, V1, V2);
7332 assert(!Zeroable.all() && "Fully zeroable shuffle mask");
7334 int Size = Mask.size();
7335 int HalfSize = Size / 2;
7336 assert(Size == (int)VT.getVectorNumElements() && "Unexpected mask size");
7338 // Upper half must be undefined.
7339 if (!isUndefInRange(Mask, HalfSize, HalfSize))
7342 // EXTRQ: Extract Len elements from lower half of source, starting at Idx.
7343 // Remainder of lower half result is zero and upper half is all undef.
7344 auto LowerAsEXTRQ = [&]() {
7345 // Determine the extraction length from the part of the
7346 // lower half that isn't zeroable.
7348 for (; Len >= 0; --Len)
7349 if (!Zeroable[Len - 1])
7351 assert(Len > 0 && "Zeroable shuffle mask");
7353 // Attempt to match first Len sequential elements from the lower half.
7356 for (int i = 0; i != Len; ++i) {
7360 SDValue &V = (M < Size ? V1 : V2);
7363 // All mask elements must be in the lower half.
7367 if (Idx < 0 || (Src == V && Idx == (M - i))) {
7378 assert((Idx + Len) <= HalfSize && "Illegal extraction mask");
7379 int BitLen = (Len * VT.getScalarSizeInBits()) & 0x3f;
7380 int BitIdx = (Idx * VT.getScalarSizeInBits()) & 0x3f;
7381 return DAG.getNode(X86ISD::EXTRQI, DL, VT, Src,
7382 DAG.getConstant(BitLen, DL, MVT::i8),
7383 DAG.getConstant(BitIdx, DL, MVT::i8));
7386 if (SDValue ExtrQ = LowerAsEXTRQ())
7389 // INSERTQ: Extract lowest Len elements from lower half of second source and
7390 // insert over first source, starting at Idx.
7391 // { A[0], .., A[Idx-1], B[0], .., B[Len-1], A[Idx+Len], .., UNDEF, ... }
7392 auto LowerAsInsertQ = [&]() {
7393 for (int Idx = 0; Idx != HalfSize; ++Idx) {
7396 // Attempt to match first source from mask before insertion point.
7397 if (isUndefInRange(Mask, 0, Idx)) {
7399 } else if (isSequentialOrUndefInRange(Mask, 0, Idx, 0)) {
7401 } else if (isSequentialOrUndefInRange(Mask, 0, Idx, Size)) {
7407 // Extend the extraction length looking to match both the insertion of
7408 // the second source and the remaining elements of the first.
7409 for (int Hi = Idx + 1; Hi <= HalfSize; ++Hi) {
7414 if (isSequentialOrUndefInRange(Mask, Idx, Len, 0)) {
7416 } else if (isSequentialOrUndefInRange(Mask, Idx, Len, Size)) {
7422 // Match the remaining elements of the lower half.
7423 if (isUndefInRange(Mask, Hi, HalfSize - Hi)) {
7425 } else if ((!Base || (Base == V1)) &&
7426 isSequentialOrUndefInRange(Mask, Hi, HalfSize - Hi, Hi)) {
7428 } else if ((!Base || (Base == V2)) &&
7429 isSequentialOrUndefInRange(Mask, Hi, HalfSize - Hi,
7436 // We may not have a base (first source) - this can safely be undefined.
7438 Base = DAG.getUNDEF(VT);
7440 int BitLen = (Len * VT.getScalarSizeInBits()) & 0x3f;
7441 int BitIdx = (Idx * VT.getScalarSizeInBits()) & 0x3f;
7442 return DAG.getNode(X86ISD::INSERTQI, DL, VT, Base, Insert,
7443 DAG.getConstant(BitLen, DL, MVT::i8),
7444 DAG.getConstant(BitIdx, DL, MVT::i8));
7451 if (SDValue InsertQ = LowerAsInsertQ())
7457 /// \brief Lower a vector shuffle as a zero or any extension.
7459 /// Given a specific number of elements, element bit width, and extension
7460 /// stride, produce either a zero or any extension based on the available
7461 /// features of the subtarget. The extended elements are consecutive and
7462 /// begin and can start from an offseted element index in the input; to
7463 /// avoid excess shuffling the offset must either being in the bottom lane
7464 /// or at the start of a higher lane. All extended elements must be from
7466 static SDValue lowerVectorShuffleAsSpecificZeroOrAnyExtend(
7467 SDLoc DL, MVT VT, int Scale, int Offset, bool AnyExt, SDValue InputV,
7468 ArrayRef<int> Mask, const X86Subtarget *Subtarget, SelectionDAG &DAG) {
7469 assert(Scale > 1 && "Need a scale to extend.");
7470 int EltBits = VT.getScalarSizeInBits();
7471 int NumElements = VT.getVectorNumElements();
7472 int NumEltsPerLane = 128 / EltBits;
7473 int OffsetLane = Offset / NumEltsPerLane;
7474 assert((EltBits == 8 || EltBits == 16 || EltBits == 32) &&
7475 "Only 8, 16, and 32 bit elements can be extended.");
7476 assert(Scale * EltBits <= 64 && "Cannot zero extend past 64 bits.");
7477 assert(0 <= Offset && "Extension offset must be positive.");
7478 assert((Offset < NumEltsPerLane || Offset % NumEltsPerLane == 0) &&
7479 "Extension offset must be in the first lane or start an upper lane.");
7481 // Check that an index is in same lane as the base offset.
7482 auto SafeOffset = [&](int Idx) {
7483 return OffsetLane == (Idx / NumEltsPerLane);
7486 // Shift along an input so that the offset base moves to the first element.
7487 auto ShuffleOffset = [&](SDValue V) {
7491 SmallVector<int, 8> ShMask((unsigned)NumElements, -1);
7492 for (int i = 0; i * Scale < NumElements; ++i) {
7493 int SrcIdx = i + Offset;
7494 ShMask[i] = SafeOffset(SrcIdx) ? SrcIdx : -1;
7496 return DAG.getVectorShuffle(VT, DL, V, DAG.getUNDEF(VT), ShMask);
7499 // Found a valid zext mask! Try various lowering strategies based on the
7500 // input type and available ISA extensions.
7501 if (Subtarget->hasSSE41()) {
7502 // Not worth offseting 128-bit vectors if scale == 2, a pattern using
7503 // PUNPCK will catch this in a later shuffle match.
7504 if (Offset && Scale == 2 && VT.getSizeInBits() == 128)
7506 MVT ExtVT = MVT::getVectorVT(MVT::getIntegerVT(EltBits * Scale),
7507 NumElements / Scale);
7508 InputV = DAG.getNode(X86ISD::VZEXT, DL, ExtVT, ShuffleOffset(InputV));
7509 return DAG.getBitcast(VT, InputV);
7512 assert(VT.getSizeInBits() == 128 && "Only 128-bit vectors can be extended.");
7514 // For any extends we can cheat for larger element sizes and use shuffle
7515 // instructions that can fold with a load and/or copy.
7516 if (AnyExt && EltBits == 32) {
7517 int PSHUFDMask[4] = {Offset, -1, SafeOffset(Offset + 1) ? Offset + 1 : -1,
7519 return DAG.getBitcast(
7520 VT, DAG.getNode(X86ISD::PSHUFD, DL, MVT::v4i32,
7521 DAG.getBitcast(MVT::v4i32, InputV),
7522 getV4X86ShuffleImm8ForMask(PSHUFDMask, DL, DAG)));
7524 if (AnyExt && EltBits == 16 && Scale > 2) {
7525 int PSHUFDMask[4] = {Offset / 2, -1,
7526 SafeOffset(Offset + 1) ? (Offset + 1) / 2 : -1, -1};
7527 InputV = DAG.getNode(X86ISD::PSHUFD, DL, MVT::v4i32,
7528 DAG.getBitcast(MVT::v4i32, InputV),
7529 getV4X86ShuffleImm8ForMask(PSHUFDMask, DL, DAG));
7530 int PSHUFWMask[4] = {1, -1, -1, -1};
7531 unsigned OddEvenOp = (Offset & 1 ? X86ISD::PSHUFLW : X86ISD::PSHUFHW);
7532 return DAG.getBitcast(
7533 VT, DAG.getNode(OddEvenOp, DL, MVT::v8i16,
7534 DAG.getBitcast(MVT::v8i16, InputV),
7535 getV4X86ShuffleImm8ForMask(PSHUFWMask, DL, DAG)));
7538 // The SSE4A EXTRQ instruction can efficiently extend the first 2 lanes
7540 if ((Scale * EltBits) == 64 && EltBits < 32 && Subtarget->hasSSE4A()) {
7541 assert(NumElements == (int)Mask.size() && "Unexpected shuffle mask size!");
7542 assert(VT.getSizeInBits() == 128 && "Unexpected vector width!");
7544 int LoIdx = Offset * EltBits;
7545 SDValue Lo = DAG.getNode(ISD::BITCAST, DL, MVT::v2i64,
7546 DAG.getNode(X86ISD::EXTRQI, DL, VT, InputV,
7547 DAG.getConstant(EltBits, DL, MVT::i8),
7548 DAG.getConstant(LoIdx, DL, MVT::i8)));
7550 if (isUndefInRange(Mask, NumElements / 2, NumElements / 2) ||
7551 !SafeOffset(Offset + 1))
7552 return DAG.getNode(ISD::BITCAST, DL, VT, Lo);
7554 int HiIdx = (Offset + 1) * EltBits;
7555 SDValue Hi = DAG.getNode(ISD::BITCAST, DL, MVT::v2i64,
7556 DAG.getNode(X86ISD::EXTRQI, DL, VT, InputV,
7557 DAG.getConstant(EltBits, DL, MVT::i8),
7558 DAG.getConstant(HiIdx, DL, MVT::i8)));
7559 return DAG.getNode(ISD::BITCAST, DL, VT,
7560 DAG.getNode(X86ISD::UNPCKL, DL, MVT::v2i64, Lo, Hi));
7563 // If this would require more than 2 unpack instructions to expand, use
7564 // pshufb when available. We can only use more than 2 unpack instructions
7565 // when zero extending i8 elements which also makes it easier to use pshufb.
7566 if (Scale > 4 && EltBits == 8 && Subtarget->hasSSSE3()) {
7567 assert(NumElements == 16 && "Unexpected byte vector width!");
7568 SDValue PSHUFBMask[16];
7569 for (int i = 0; i < 16; ++i) {
7570 int Idx = Offset + (i / Scale);
7571 PSHUFBMask[i] = DAG.getConstant(
7572 (i % Scale == 0 && SafeOffset(Idx)) ? Idx : 0x80, DL, MVT::i8);
7574 InputV = DAG.getBitcast(MVT::v16i8, InputV);
7575 return DAG.getBitcast(VT,
7576 DAG.getNode(X86ISD::PSHUFB, DL, MVT::v16i8, InputV,
7577 DAG.getNode(ISD::BUILD_VECTOR, DL,
7578 MVT::v16i8, PSHUFBMask)));
7581 // If we are extending from an offset, ensure we start on a boundary that
7582 // we can unpack from.
7583 int AlignToUnpack = Offset % (NumElements / Scale);
7584 if (AlignToUnpack) {
7585 SmallVector<int, 8> ShMask((unsigned)NumElements, -1);
7586 for (int i = AlignToUnpack; i < NumElements; ++i)
7587 ShMask[i - AlignToUnpack] = i;
7588 InputV = DAG.getVectorShuffle(VT, DL, InputV, DAG.getUNDEF(VT), ShMask);
7589 Offset -= AlignToUnpack;
7592 // Otherwise emit a sequence of unpacks.
7594 unsigned UnpackLoHi = X86ISD::UNPCKL;
7595 if (Offset >= (NumElements / 2)) {
7596 UnpackLoHi = X86ISD::UNPCKH;
7597 Offset -= (NumElements / 2);
7600 MVT InputVT = MVT::getVectorVT(MVT::getIntegerVT(EltBits), NumElements);
7601 SDValue Ext = AnyExt ? DAG.getUNDEF(InputVT)
7602 : getZeroVector(InputVT, Subtarget, DAG, DL);
7603 InputV = DAG.getBitcast(InputVT, InputV);
7604 InputV = DAG.getNode(UnpackLoHi, DL, InputVT, InputV, Ext);
7608 } while (Scale > 1);
7609 return DAG.getBitcast(VT, InputV);
7612 /// \brief Try to lower a vector shuffle as a zero extension on any microarch.
7614 /// This routine will try to do everything in its power to cleverly lower
7615 /// a shuffle which happens to match the pattern of a zero extend. It doesn't
7616 /// check for the profitability of this lowering, it tries to aggressively
7617 /// match this pattern. It will use all of the micro-architectural details it
7618 /// can to emit an efficient lowering. It handles both blends with all-zero
7619 /// inputs to explicitly zero-extend and undef-lanes (sometimes undef due to
7620 /// masking out later).
7622 /// The reason we have dedicated lowering for zext-style shuffles is that they
7623 /// are both incredibly common and often quite performance sensitive.
7624 static SDValue lowerVectorShuffleAsZeroOrAnyExtend(
7625 SDLoc DL, MVT VT, SDValue V1, SDValue V2, ArrayRef<int> Mask,
7626 const X86Subtarget *Subtarget, SelectionDAG &DAG) {
7627 SmallBitVector Zeroable = computeZeroableShuffleElements(Mask, V1, V2);
7629 int Bits = VT.getSizeInBits();
7630 int NumLanes = Bits / 128;
7631 int NumElements = VT.getVectorNumElements();
7632 int NumEltsPerLane = NumElements / NumLanes;
7633 assert(VT.getScalarSizeInBits() <= 32 &&
7634 "Exceeds 32-bit integer zero extension limit");
7635 assert((int)Mask.size() == NumElements && "Unexpected shuffle mask size");
7637 // Define a helper function to check a particular ext-scale and lower to it if
7639 auto Lower = [&](int Scale) -> SDValue {
7644 for (int i = 0; i < NumElements; ++i) {
7647 continue; // Valid anywhere but doesn't tell us anything.
7648 if (i % Scale != 0) {
7649 // Each of the extended elements need to be zeroable.
7653 // We no longer are in the anyext case.
7658 // Each of the base elements needs to be consecutive indices into the
7659 // same input vector.
7660 SDValue V = M < NumElements ? V1 : V2;
7661 M = M % NumElements;
7664 Offset = M - (i / Scale);
7665 } else if (InputV != V)
7666 return SDValue(); // Flip-flopping inputs.
7668 // Offset must start in the lowest 128-bit lane or at the start of an
7670 // FIXME: Is it ever worth allowing a negative base offset?
7671 if (!((0 <= Offset && Offset < NumEltsPerLane) ||
7672 (Offset % NumEltsPerLane) == 0))
7675 // If we are offsetting, all referenced entries must come from the same
7677 if (Offset && (Offset / NumEltsPerLane) != (M / NumEltsPerLane))
7680 if ((M % NumElements) != (Offset + (i / Scale)))
7681 return SDValue(); // Non-consecutive strided elements.
7685 // If we fail to find an input, we have a zero-shuffle which should always
7686 // have already been handled.
7687 // FIXME: Maybe handle this here in case during blending we end up with one?
7691 // If we are offsetting, don't extend if we only match a single input, we
7692 // can always do better by using a basic PSHUF or PUNPCK.
7693 if (Offset != 0 && Matches < 2)
7696 return lowerVectorShuffleAsSpecificZeroOrAnyExtend(
7697 DL, VT, Scale, Offset, AnyExt, InputV, Mask, Subtarget, DAG);
7700 // The widest scale possible for extending is to a 64-bit integer.
7701 assert(Bits % 64 == 0 &&
7702 "The number of bits in a vector must be divisible by 64 on x86!");
7703 int NumExtElements = Bits / 64;
7705 // Each iteration, try extending the elements half as much, but into twice as
7707 for (; NumExtElements < NumElements; NumExtElements *= 2) {
7708 assert(NumElements % NumExtElements == 0 &&
7709 "The input vector size must be divisible by the extended size.");
7710 if (SDValue V = Lower(NumElements / NumExtElements))
7714 // General extends failed, but 128-bit vectors may be able to use MOVQ.
7718 // Returns one of the source operands if the shuffle can be reduced to a
7719 // MOVQ, copying the lower 64-bits and zero-extending to the upper 64-bits.
7720 auto CanZExtLowHalf = [&]() {
7721 for (int i = NumElements / 2; i != NumElements; ++i)
7724 if (isSequentialOrUndefInRange(Mask, 0, NumElements / 2, 0))
7726 if (isSequentialOrUndefInRange(Mask, 0, NumElements / 2, NumElements))
7731 if (SDValue V = CanZExtLowHalf()) {
7732 V = DAG.getBitcast(MVT::v2i64, V);
7733 V = DAG.getNode(X86ISD::VZEXT_MOVL, DL, MVT::v2i64, V);
7734 return DAG.getBitcast(VT, V);
7737 // No viable ext lowering found.
7741 /// \brief Try to get a scalar value for a specific element of a vector.
7743 /// Looks through BUILD_VECTOR and SCALAR_TO_VECTOR nodes to find a scalar.
7744 static SDValue getScalarValueForVectorElement(SDValue V, int Idx,
7745 SelectionDAG &DAG) {
7746 MVT VT = V.getSimpleValueType();
7747 MVT EltVT = VT.getVectorElementType();
7748 while (V.getOpcode() == ISD::BITCAST)
7749 V = V.getOperand(0);
7750 // If the bitcasts shift the element size, we can't extract an equivalent
7752 MVT NewVT = V.getSimpleValueType();
7753 if (!NewVT.isVector() || NewVT.getScalarSizeInBits() != VT.getScalarSizeInBits())
7756 if (V.getOpcode() == ISD::BUILD_VECTOR ||
7757 (Idx == 0 && V.getOpcode() == ISD::SCALAR_TO_VECTOR)) {
7758 // Ensure the scalar operand is the same size as the destination.
7759 // FIXME: Add support for scalar truncation where possible.
7760 SDValue S = V.getOperand(Idx);
7761 if (EltVT.getSizeInBits() == S.getSimpleValueType().getSizeInBits())
7762 return DAG.getNode(ISD::BITCAST, SDLoc(V), EltVT, S);
7768 /// \brief Helper to test for a load that can be folded with x86 shuffles.
7770 /// This is particularly important because the set of instructions varies
7771 /// significantly based on whether the operand is a load or not.
7772 static bool isShuffleFoldableLoad(SDValue V) {
7773 while (V.getOpcode() == ISD::BITCAST)
7774 V = V.getOperand(0);
7776 return ISD::isNON_EXTLoad(V.getNode());
7779 /// \brief Try to lower insertion of a single element into a zero vector.
7781 /// This is a common pattern that we have especially efficient patterns to lower
7782 /// across all subtarget feature sets.
7783 static SDValue lowerVectorShuffleAsElementInsertion(
7784 SDLoc DL, MVT VT, SDValue V1, SDValue V2, ArrayRef<int> Mask,
7785 const X86Subtarget *Subtarget, SelectionDAG &DAG) {
7786 SmallBitVector Zeroable = computeZeroableShuffleElements(Mask, V1, V2);
7788 MVT EltVT = VT.getVectorElementType();
7790 int V2Index = std::find_if(Mask.begin(), Mask.end(),
7791 [&Mask](int M) { return M >= (int)Mask.size(); }) -
7793 bool IsV1Zeroable = true;
7794 for (int i = 0, Size = Mask.size(); i < Size; ++i)
7795 if (i != V2Index && !Zeroable[i]) {
7796 IsV1Zeroable = false;
7800 // Check for a single input from a SCALAR_TO_VECTOR node.
7801 // FIXME: All of this should be canonicalized into INSERT_VECTOR_ELT and
7802 // all the smarts here sunk into that routine. However, the current
7803 // lowering of BUILD_VECTOR makes that nearly impossible until the old
7804 // vector shuffle lowering is dead.
7805 SDValue V2S = getScalarValueForVectorElement(V2, Mask[V2Index] - Mask.size(),
7807 if (V2S && DAG.getTargetLoweringInfo().isTypeLegal(V2S.getValueType())) {
7808 // We need to zext the scalar if it is smaller than an i32.
7809 V2S = DAG.getBitcast(EltVT, V2S);
7810 if (EltVT == MVT::i8 || EltVT == MVT::i16) {
7811 // Using zext to expand a narrow element won't work for non-zero
7816 // Zero-extend directly to i32.
7818 V2S = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::i32, V2S);
7820 V2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, ExtVT, V2S);
7821 } else if (Mask[V2Index] != (int)Mask.size() || EltVT == MVT::i8 ||
7822 EltVT == MVT::i16) {
7823 // Either not inserting from the low element of the input or the input
7824 // element size is too small to use VZEXT_MOVL to clear the high bits.
7828 if (!IsV1Zeroable) {
7829 // If V1 can't be treated as a zero vector we have fewer options to lower
7830 // this. We can't support integer vectors or non-zero targets cheaply, and
7831 // the V1 elements can't be permuted in any way.
7832 assert(VT == ExtVT && "Cannot change extended type when non-zeroable!");
7833 if (!VT.isFloatingPoint() || V2Index != 0)
7835 SmallVector<int, 8> V1Mask(Mask.begin(), Mask.end());
7836 V1Mask[V2Index] = -1;
7837 if (!isNoopShuffleMask(V1Mask))
7839 // This is essentially a special case blend operation, but if we have
7840 // general purpose blend operations, they are always faster. Bail and let
7841 // the rest of the lowering handle these as blends.
7842 if (Subtarget->hasSSE41())
7845 // Otherwise, use MOVSD or MOVSS.
7846 assert((EltVT == MVT::f32 || EltVT == MVT::f64) &&
7847 "Only two types of floating point element types to handle!");
7848 return DAG.getNode(EltVT == MVT::f32 ? X86ISD::MOVSS : X86ISD::MOVSD, DL,
7852 // This lowering only works for the low element with floating point vectors.
7853 if (VT.isFloatingPoint() && V2Index != 0)
7856 V2 = DAG.getNode(X86ISD::VZEXT_MOVL, DL, ExtVT, V2);
7858 V2 = DAG.getBitcast(VT, V2);
7861 // If we have 4 or fewer lanes we can cheaply shuffle the element into
7862 // the desired position. Otherwise it is more efficient to do a vector
7863 // shift left. We know that we can do a vector shift left because all
7864 // the inputs are zero.
7865 if (VT.isFloatingPoint() || VT.getVectorNumElements() <= 4) {
7866 SmallVector<int, 4> V2Shuffle(Mask.size(), 1);
7867 V2Shuffle[V2Index] = 0;
7868 V2 = DAG.getVectorShuffle(VT, DL, V2, DAG.getUNDEF(VT), V2Shuffle);
7870 V2 = DAG.getBitcast(MVT::v2i64, V2);
7872 X86ISD::VSHLDQ, DL, MVT::v2i64, V2,
7873 DAG.getConstant(V2Index * EltVT.getSizeInBits() / 8, DL,
7874 DAG.getTargetLoweringInfo().getScalarShiftAmountTy(
7875 DAG.getDataLayout(), VT)));
7876 V2 = DAG.getBitcast(VT, V2);
7882 /// \brief Try to lower broadcast of a single element.
7884 /// For convenience, this code also bundles all of the subtarget feature set
7885 /// filtering. While a little annoying to re-dispatch on type here, there isn't
7886 /// a convenient way to factor it out.
7887 static SDValue lowerVectorShuffleAsBroadcast(SDLoc DL, MVT VT, SDValue V,
7889 const X86Subtarget *Subtarget,
7890 SelectionDAG &DAG) {
7891 if (!Subtarget->hasAVX())
7893 if (VT.isInteger() && !Subtarget->hasAVX2())
7896 // Check that the mask is a broadcast.
7897 int BroadcastIdx = -1;
7899 if (M >= 0 && BroadcastIdx == -1)
7901 else if (M >= 0 && M != BroadcastIdx)
7904 assert(BroadcastIdx < (int)Mask.size() && "We only expect to be called with "
7905 "a sorted mask where the broadcast "
7908 // Go up the chain of (vector) values to find a scalar load that we can
7909 // combine with the broadcast.
7911 switch (V.getOpcode()) {
7912 case ISD::CONCAT_VECTORS: {
7913 int OperandSize = Mask.size() / V.getNumOperands();
7914 V = V.getOperand(BroadcastIdx / OperandSize);
7915 BroadcastIdx %= OperandSize;
7919 case ISD::INSERT_SUBVECTOR: {
7920 SDValue VOuter = V.getOperand(0), VInner = V.getOperand(1);
7921 auto ConstantIdx = dyn_cast<ConstantSDNode>(V.getOperand(2));
7925 int BeginIdx = (int)ConstantIdx->getZExtValue();
7927 BeginIdx + (int)VInner.getValueType().getVectorNumElements();
7928 if (BroadcastIdx >= BeginIdx && BroadcastIdx < EndIdx) {
7929 BroadcastIdx -= BeginIdx;
7940 // Check if this is a broadcast of a scalar. We special case lowering
7941 // for scalars so that we can more effectively fold with loads.
7942 // First, look through bitcast: if the original value has a larger element
7943 // type than the shuffle, the broadcast element is in essence truncated.
7944 // Make that explicit to ease folding.
7945 if (V.getOpcode() == ISD::BITCAST && VT.isInteger()) {
7946 EVT EltVT = VT.getVectorElementType();
7947 SDValue V0 = V.getOperand(0);
7948 EVT V0VT = V0.getValueType();
7950 if (V0VT.isInteger() && V0VT.getVectorElementType().bitsGT(EltVT) &&
7951 ((V0.getOpcode() == ISD::BUILD_VECTOR ||
7952 (V0.getOpcode() == ISD::SCALAR_TO_VECTOR && BroadcastIdx == 0)))) {
7953 V = DAG.getNode(ISD::TRUNCATE, DL, EltVT, V0.getOperand(BroadcastIdx));
7958 // Also check the simpler case, where we can directly reuse the scalar.
7959 if (V.getOpcode() == ISD::BUILD_VECTOR ||
7960 (V.getOpcode() == ISD::SCALAR_TO_VECTOR && BroadcastIdx == 0)) {
7961 V = V.getOperand(BroadcastIdx);
7963 // If the scalar isn't a load, we can't broadcast from it in AVX1.
7964 // Only AVX2 has register broadcasts.
7965 if (!Subtarget->hasAVX2() && !isShuffleFoldableLoad(V))
7967 } else if (BroadcastIdx != 0 || !Subtarget->hasAVX2()) {
7968 // We can't broadcast from a vector register without AVX2, and we can only
7969 // broadcast from the zero-element of a vector register.
7973 return DAG.getNode(X86ISD::VBROADCAST, DL, VT, V);
7976 // Check for whether we can use INSERTPS to perform the shuffle. We only use
7977 // INSERTPS when the V1 elements are already in the correct locations
7978 // because otherwise we can just always use two SHUFPS instructions which
7979 // are much smaller to encode than a SHUFPS and an INSERTPS. We can also
7980 // perform INSERTPS if a single V1 element is out of place and all V2
7981 // elements are zeroable.
7982 static SDValue lowerVectorShuffleAsInsertPS(SDValue Op, SDValue V1, SDValue V2,
7984 SelectionDAG &DAG) {
7985 assert(Op.getSimpleValueType() == MVT::v4f32 && "Bad shuffle type!");
7986 assert(V1.getSimpleValueType() == MVT::v4f32 && "Bad operand type!");
7987 assert(V2.getSimpleValueType() == MVT::v4f32 && "Bad operand type!");
7988 assert(Mask.size() == 4 && "Unexpected mask size for v4 shuffle!");
7990 SmallBitVector Zeroable = computeZeroableShuffleElements(Mask, V1, V2);
7993 int V1DstIndex = -1;
7994 int V2DstIndex = -1;
7995 bool V1UsedInPlace = false;
7997 for (int i = 0; i < 4; ++i) {
7998 // Synthesize a zero mask from the zeroable elements (includes undefs).
8004 // Flag if we use any V1 inputs in place.
8006 V1UsedInPlace = true;
8010 // We can only insert a single non-zeroable element.
8011 if (V1DstIndex != -1 || V2DstIndex != -1)
8015 // V1 input out of place for insertion.
8018 // V2 input for insertion.
8023 // Don't bother if we have no (non-zeroable) element for insertion.
8024 if (V1DstIndex == -1 && V2DstIndex == -1)
8027 // Determine element insertion src/dst indices. The src index is from the
8028 // start of the inserted vector, not the start of the concatenated vector.
8029 unsigned V2SrcIndex = 0;
8030 if (V1DstIndex != -1) {
8031 // If we have a V1 input out of place, we use V1 as the V2 element insertion
8032 // and don't use the original V2 at all.
8033 V2SrcIndex = Mask[V1DstIndex];
8034 V2DstIndex = V1DstIndex;
8037 V2SrcIndex = Mask[V2DstIndex] - 4;
8040 // If no V1 inputs are used in place, then the result is created only from
8041 // the zero mask and the V2 insertion - so remove V1 dependency.
8043 V1 = DAG.getUNDEF(MVT::v4f32);
8045 unsigned InsertPSMask = V2SrcIndex << 6 | V2DstIndex << 4 | ZMask;
8046 assert((InsertPSMask & ~0xFFu) == 0 && "Invalid mask!");
8048 // Insert the V2 element into the desired position.
8050 return DAG.getNode(X86ISD::INSERTPS, DL, MVT::v4f32, V1, V2,
8051 DAG.getConstant(InsertPSMask, DL, MVT::i8));
8054 /// \brief Try to lower a shuffle as a permute of the inputs followed by an
8055 /// UNPCK instruction.
8057 /// This specifically targets cases where we end up with alternating between
8058 /// the two inputs, and so can permute them into something that feeds a single
8059 /// UNPCK instruction. Note that this routine only targets integer vectors
8060 /// because for floating point vectors we have a generalized SHUFPS lowering
8061 /// strategy that handles everything that doesn't *exactly* match an unpack,
8062 /// making this clever lowering unnecessary.
8063 static SDValue lowerVectorShuffleAsPermuteAndUnpack(SDLoc DL, MVT VT,
8064 SDValue V1, SDValue V2,
8066 SelectionDAG &DAG) {
8067 assert(!VT.isFloatingPoint() &&
8068 "This routine only supports integer vectors.");
8069 assert(!isSingleInputShuffleMask(Mask) &&
8070 "This routine should only be used when blending two inputs.");
8071 assert(Mask.size() >= 2 && "Single element masks are invalid.");
8073 int Size = Mask.size();
8075 int NumLoInputs = std::count_if(Mask.begin(), Mask.end(), [Size](int M) {
8076 return M >= 0 && M % Size < Size / 2;
8078 int NumHiInputs = std::count_if(
8079 Mask.begin(), Mask.end(), [Size](int M) { return M % Size >= Size / 2; });
8081 bool UnpackLo = NumLoInputs >= NumHiInputs;
8083 auto TryUnpack = [&](MVT UnpackVT, int Scale) {
8084 SmallVector<int, 32> V1Mask(Mask.size(), -1);
8085 SmallVector<int, 32> V2Mask(Mask.size(), -1);
8087 for (int i = 0; i < Size; ++i) {
8091 // Each element of the unpack contains Scale elements from this mask.
8092 int UnpackIdx = i / Scale;
8094 // We only handle the case where V1 feeds the first slots of the unpack.
8095 // We rely on canonicalization to ensure this is the case.
8096 if ((UnpackIdx % 2 == 0) != (Mask[i] < Size))
8099 // Setup the mask for this input. The indexing is tricky as we have to
8100 // handle the unpack stride.
8101 SmallVectorImpl<int> &VMask = (UnpackIdx % 2 == 0) ? V1Mask : V2Mask;
8102 VMask[(UnpackIdx / 2) * Scale + i % Scale + (UnpackLo ? 0 : Size / 2)] =
8106 // If we will have to shuffle both inputs to use the unpack, check whether
8107 // we can just unpack first and shuffle the result. If so, skip this unpack.
8108 if ((NumLoInputs == 0 || NumHiInputs == 0) && !isNoopShuffleMask(V1Mask) &&
8109 !isNoopShuffleMask(V2Mask))
8112 // Shuffle the inputs into place.
8113 V1 = DAG.getVectorShuffle(VT, DL, V1, DAG.getUNDEF(VT), V1Mask);
8114 V2 = DAG.getVectorShuffle(VT, DL, V2, DAG.getUNDEF(VT), V2Mask);
8116 // Cast the inputs to the type we will use to unpack them.
8117 V1 = DAG.getBitcast(UnpackVT, V1);
8118 V2 = DAG.getBitcast(UnpackVT, V2);
8120 // Unpack the inputs and cast the result back to the desired type.
8121 return DAG.getBitcast(
8122 VT, DAG.getNode(UnpackLo ? X86ISD::UNPCKL : X86ISD::UNPCKH, DL,
8126 // We try each unpack from the largest to the smallest to try and find one
8127 // that fits this mask.
8128 int OrigNumElements = VT.getVectorNumElements();
8129 int OrigScalarSize = VT.getScalarSizeInBits();
8130 for (int ScalarSize = 64; ScalarSize >= OrigScalarSize; ScalarSize /= 2) {
8131 int Scale = ScalarSize / OrigScalarSize;
8132 int NumElements = OrigNumElements / Scale;
8133 MVT UnpackVT = MVT::getVectorVT(MVT::getIntegerVT(ScalarSize), NumElements);
8134 if (SDValue Unpack = TryUnpack(UnpackVT, Scale))
8138 // If none of the unpack-rooted lowerings worked (or were profitable) try an
8140 if (NumLoInputs == 0 || NumHiInputs == 0) {
8141 assert((NumLoInputs > 0 || NumHiInputs > 0) &&
8142 "We have to have *some* inputs!");
8143 int HalfOffset = NumLoInputs == 0 ? Size / 2 : 0;
8145 // FIXME: We could consider the total complexity of the permute of each
8146 // possible unpacking. Or at the least we should consider how many
8147 // half-crossings are created.
8148 // FIXME: We could consider commuting the unpacks.
8150 SmallVector<int, 32> PermMask;
8151 PermMask.assign(Size, -1);
8152 for (int i = 0; i < Size; ++i) {
8156 assert(Mask[i] % Size >= HalfOffset && "Found input from wrong half!");
8159 2 * ((Mask[i] % Size) - HalfOffset) + (Mask[i] < Size ? 0 : 1);
8161 return DAG.getVectorShuffle(
8162 VT, DL, DAG.getNode(NumLoInputs == 0 ? X86ISD::UNPCKH : X86ISD::UNPCKL,
8164 DAG.getUNDEF(VT), PermMask);
8170 /// \brief Handle lowering of 2-lane 64-bit floating point shuffles.
8172 /// This is the basis function for the 2-lane 64-bit shuffles as we have full
8173 /// support for floating point shuffles but not integer shuffles. These
8174 /// instructions will incur a domain crossing penalty on some chips though so
8175 /// it is better to avoid lowering through this for integer vectors where
8177 static SDValue lowerV2F64VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
8178 const X86Subtarget *Subtarget,
8179 SelectionDAG &DAG) {
8181 assert(Op.getSimpleValueType() == MVT::v2f64 && "Bad shuffle type!");
8182 assert(V1.getSimpleValueType() == MVT::v2f64 && "Bad operand type!");
8183 assert(V2.getSimpleValueType() == MVT::v2f64 && "Bad operand type!");
8184 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
8185 ArrayRef<int> Mask = SVOp->getMask();
8186 assert(Mask.size() == 2 && "Unexpected mask size for v2 shuffle!");
8188 if (isSingleInputShuffleMask(Mask)) {
8189 // Use low duplicate instructions for masks that match their pattern.
8190 if (Subtarget->hasSSE3())
8191 if (isShuffleEquivalent(V1, V2, Mask, {0, 0}))
8192 return DAG.getNode(X86ISD::MOVDDUP, DL, MVT::v2f64, V1);
8194 // Straight shuffle of a single input vector. Simulate this by using the
8195 // single input as both of the "inputs" to this instruction..
8196 unsigned SHUFPDMask = (Mask[0] == 1) | ((Mask[1] == 1) << 1);
8198 if (Subtarget->hasAVX()) {
8199 // If we have AVX, we can use VPERMILPS which will allow folding a load
8200 // into the shuffle.
8201 return DAG.getNode(X86ISD::VPERMILPI, DL, MVT::v2f64, V1,
8202 DAG.getConstant(SHUFPDMask, DL, MVT::i8));
8205 return DAG.getNode(X86ISD::SHUFP, DL, MVT::v2f64, V1, V1,
8206 DAG.getConstant(SHUFPDMask, DL, MVT::i8));
8208 assert(Mask[0] >= 0 && Mask[0] < 2 && "Non-canonicalized blend!");
8209 assert(Mask[1] >= 2 && "Non-canonicalized blend!");
8211 // If we have a single input, insert that into V1 if we can do so cheaply.
8212 if ((Mask[0] >= 2) + (Mask[1] >= 2) == 1) {
8213 if (SDValue Insertion = lowerVectorShuffleAsElementInsertion(
8214 DL, MVT::v2f64, V1, V2, Mask, Subtarget, DAG))
8216 // Try inverting the insertion since for v2 masks it is easy to do and we
8217 // can't reliably sort the mask one way or the other.
8218 int InverseMask[2] = {Mask[0] < 0 ? -1 : (Mask[0] ^ 2),
8219 Mask[1] < 0 ? -1 : (Mask[1] ^ 2)};
8220 if (SDValue Insertion = lowerVectorShuffleAsElementInsertion(
8221 DL, MVT::v2f64, V2, V1, InverseMask, Subtarget, DAG))
8225 // Try to use one of the special instruction patterns to handle two common
8226 // blend patterns if a zero-blend above didn't work.
8227 if (isShuffleEquivalent(V1, V2, Mask, {0, 3}) ||
8228 isShuffleEquivalent(V1, V2, Mask, {1, 3}))
8229 if (SDValue V1S = getScalarValueForVectorElement(V1, Mask[0], DAG))
8230 // We can either use a special instruction to load over the low double or
8231 // to move just the low double.
8233 isShuffleFoldableLoad(V1S) ? X86ISD::MOVLPD : X86ISD::MOVSD,
8235 DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, MVT::v2f64, V1S));
8237 if (Subtarget->hasSSE41())
8238 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v2f64, V1, V2, Mask,
8242 // Use dedicated unpack instructions for masks that match their pattern.
8243 if (isShuffleEquivalent(V1, V2, Mask, {0, 2}))
8244 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v2f64, V1, V2);
8245 if (isShuffleEquivalent(V1, V2, Mask, {1, 3}))
8246 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v2f64, V1, V2);
8248 unsigned SHUFPDMask = (Mask[0] == 1) | (((Mask[1] - 2) == 1) << 1);
8249 return DAG.getNode(X86ISD::SHUFP, DL, MVT::v2f64, V1, V2,
8250 DAG.getConstant(SHUFPDMask, DL, MVT::i8));
8253 /// \brief Handle lowering of 2-lane 64-bit integer shuffles.
8255 /// Tries to lower a 2-lane 64-bit shuffle using shuffle operations provided by
8256 /// the integer unit to minimize domain crossing penalties. However, for blends
8257 /// it falls back to the floating point shuffle operation with appropriate bit
8259 static SDValue lowerV2I64VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
8260 const X86Subtarget *Subtarget,
8261 SelectionDAG &DAG) {
8263 assert(Op.getSimpleValueType() == MVT::v2i64 && "Bad shuffle type!");
8264 assert(V1.getSimpleValueType() == MVT::v2i64 && "Bad operand type!");
8265 assert(V2.getSimpleValueType() == MVT::v2i64 && "Bad operand type!");
8266 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
8267 ArrayRef<int> Mask = SVOp->getMask();
8268 assert(Mask.size() == 2 && "Unexpected mask size for v2 shuffle!");
8270 if (isSingleInputShuffleMask(Mask)) {
8271 // Check for being able to broadcast a single element.
8272 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v2i64, V1,
8273 Mask, Subtarget, DAG))
8276 // Straight shuffle of a single input vector. For everything from SSE2
8277 // onward this has a single fast instruction with no scary immediates.
8278 // We have to map the mask as it is actually a v4i32 shuffle instruction.
8279 V1 = DAG.getBitcast(MVT::v4i32, V1);
8280 int WidenedMask[4] = {
8281 std::max(Mask[0], 0) * 2, std::max(Mask[0], 0) * 2 + 1,
8282 std::max(Mask[1], 0) * 2, std::max(Mask[1], 0) * 2 + 1};
8283 return DAG.getBitcast(
8285 DAG.getNode(X86ISD::PSHUFD, DL, MVT::v4i32, V1,
8286 getV4X86ShuffleImm8ForMask(WidenedMask, DL, DAG)));
8288 assert(Mask[0] != -1 && "No undef lanes in multi-input v2 shuffles!");
8289 assert(Mask[1] != -1 && "No undef lanes in multi-input v2 shuffles!");
8290 assert(Mask[0] < 2 && "We sort V1 to be the first input.");
8291 assert(Mask[1] >= 2 && "We sort V2 to be the second input.");
8293 // If we have a blend of two PACKUS operations an the blend aligns with the
8294 // low and half halves, we can just merge the PACKUS operations. This is
8295 // particularly important as it lets us merge shuffles that this routine itself
8297 auto GetPackNode = [](SDValue V) {
8298 while (V.getOpcode() == ISD::BITCAST)
8299 V = V.getOperand(0);
8301 return V.getOpcode() == X86ISD::PACKUS ? V : SDValue();
8303 if (SDValue V1Pack = GetPackNode(V1))
8304 if (SDValue V2Pack = GetPackNode(V2))
8305 return DAG.getBitcast(MVT::v2i64,
8306 DAG.getNode(X86ISD::PACKUS, DL, MVT::v16i8,
8307 Mask[0] == 0 ? V1Pack.getOperand(0)
8308 : V1Pack.getOperand(1),
8309 Mask[1] == 2 ? V2Pack.getOperand(0)
8310 : V2Pack.getOperand(1)));
8312 // Try to use shift instructions.
8314 lowerVectorShuffleAsShift(DL, MVT::v2i64, V1, V2, Mask, DAG))
8317 // When loading a scalar and then shuffling it into a vector we can often do
8318 // the insertion cheaply.
8319 if (SDValue Insertion = lowerVectorShuffleAsElementInsertion(
8320 DL, MVT::v2i64, V1, V2, Mask, Subtarget, DAG))
8322 // Try inverting the insertion since for v2 masks it is easy to do and we
8323 // can't reliably sort the mask one way or the other.
8324 int InverseMask[2] = {Mask[0] ^ 2, Mask[1] ^ 2};
8325 if (SDValue Insertion = lowerVectorShuffleAsElementInsertion(
8326 DL, MVT::v2i64, V2, V1, InverseMask, Subtarget, DAG))
8329 // We have different paths for blend lowering, but they all must use the
8330 // *exact* same predicate.
8331 bool IsBlendSupported = Subtarget->hasSSE41();
8332 if (IsBlendSupported)
8333 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v2i64, V1, V2, Mask,
8337 // Use dedicated unpack instructions for masks that match their pattern.
8338 if (isShuffleEquivalent(V1, V2, Mask, {0, 2}))
8339 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v2i64, V1, V2);
8340 if (isShuffleEquivalent(V1, V2, Mask, {1, 3}))
8341 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v2i64, V1, V2);
8343 // Try to use byte rotation instructions.
8344 // Its more profitable for pre-SSSE3 to use shuffles/unpacks.
8345 if (Subtarget->hasSSSE3())
8346 if (SDValue Rotate = lowerVectorShuffleAsByteRotate(
8347 DL, MVT::v2i64, V1, V2, Mask, Subtarget, DAG))
8350 // If we have direct support for blends, we should lower by decomposing into
8351 // a permute. That will be faster than the domain cross.
8352 if (IsBlendSupported)
8353 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v2i64, V1, V2,
8356 // We implement this with SHUFPD which is pretty lame because it will likely
8357 // incur 2 cycles of stall for integer vectors on Nehalem and older chips.
8358 // However, all the alternatives are still more cycles and newer chips don't
8359 // have this problem. It would be really nice if x86 had better shuffles here.
8360 V1 = DAG.getBitcast(MVT::v2f64, V1);
8361 V2 = DAG.getBitcast(MVT::v2f64, V2);
8362 return DAG.getBitcast(MVT::v2i64,
8363 DAG.getVectorShuffle(MVT::v2f64, DL, V1, V2, Mask));
8366 /// \brief Test whether this can be lowered with a single SHUFPS instruction.
8368 /// This is used to disable more specialized lowerings when the shufps lowering
8369 /// will happen to be efficient.
8370 static bool isSingleSHUFPSMask(ArrayRef<int> Mask) {
8371 // This routine only handles 128-bit shufps.
8372 assert(Mask.size() == 4 && "Unsupported mask size!");
8374 // To lower with a single SHUFPS we need to have the low half and high half
8375 // each requiring a single input.
8376 if (Mask[0] != -1 && Mask[1] != -1 && (Mask[0] < 4) != (Mask[1] < 4))
8378 if (Mask[2] != -1 && Mask[3] != -1 && (Mask[2] < 4) != (Mask[3] < 4))
8384 /// \brief Lower a vector shuffle using the SHUFPS instruction.
8386 /// This is a helper routine dedicated to lowering vector shuffles using SHUFPS.
8387 /// It makes no assumptions about whether this is the *best* lowering, it simply
8389 static SDValue lowerVectorShuffleWithSHUFPS(SDLoc DL, MVT VT,
8390 ArrayRef<int> Mask, SDValue V1,
8391 SDValue V2, SelectionDAG &DAG) {
8392 SDValue LowV = V1, HighV = V2;
8393 int NewMask[4] = {Mask[0], Mask[1], Mask[2], Mask[3]};
8396 std::count_if(Mask.begin(), Mask.end(), [](int M) { return M >= 4; });
8398 if (NumV2Elements == 1) {
8400 std::find_if(Mask.begin(), Mask.end(), [](int M) { return M >= 4; }) -
8403 // Compute the index adjacent to V2Index and in the same half by toggling
8405 int V2AdjIndex = V2Index ^ 1;
8407 if (Mask[V2AdjIndex] == -1) {
8408 // Handles all the cases where we have a single V2 element and an undef.
8409 // This will only ever happen in the high lanes because we commute the
8410 // vector otherwise.
8412 std::swap(LowV, HighV);
8413 NewMask[V2Index] -= 4;
8415 // Handle the case where the V2 element ends up adjacent to a V1 element.
8416 // To make this work, blend them together as the first step.
8417 int V1Index = V2AdjIndex;
8418 int BlendMask[4] = {Mask[V2Index] - 4, 0, Mask[V1Index], 0};
8419 V2 = DAG.getNode(X86ISD::SHUFP, DL, VT, V2, V1,
8420 getV4X86ShuffleImm8ForMask(BlendMask, DL, DAG));
8422 // Now proceed to reconstruct the final blend as we have the necessary
8423 // high or low half formed.
8430 NewMask[V1Index] = 2; // We put the V1 element in V2[2].
8431 NewMask[V2Index] = 0; // We shifted the V2 element into V2[0].
8433 } else if (NumV2Elements == 2) {
8434 if (Mask[0] < 4 && Mask[1] < 4) {
8435 // Handle the easy case where we have V1 in the low lanes and V2 in the
8439 } else if (Mask[2] < 4 && Mask[3] < 4) {
8440 // We also handle the reversed case because this utility may get called
8441 // when we detect a SHUFPS pattern but can't easily commute the shuffle to
8442 // arrange things in the right direction.
8448 // We have a mixture of V1 and V2 in both low and high lanes. Rather than
8449 // trying to place elements directly, just blend them and set up the final
8450 // shuffle to place them.
8452 // The first two blend mask elements are for V1, the second two are for
8454 int BlendMask[4] = {Mask[0] < 4 ? Mask[0] : Mask[1],
8455 Mask[2] < 4 ? Mask[2] : Mask[3],
8456 (Mask[0] >= 4 ? Mask[0] : Mask[1]) - 4,
8457 (Mask[2] >= 4 ? Mask[2] : Mask[3]) - 4};
8458 V1 = DAG.getNode(X86ISD::SHUFP, DL, VT, V1, V2,
8459 getV4X86ShuffleImm8ForMask(BlendMask, DL, DAG));
8461 // Now we do a normal shuffle of V1 by giving V1 as both operands to
8464 NewMask[0] = Mask[0] < 4 ? 0 : 2;
8465 NewMask[1] = Mask[0] < 4 ? 2 : 0;
8466 NewMask[2] = Mask[2] < 4 ? 1 : 3;
8467 NewMask[3] = Mask[2] < 4 ? 3 : 1;
8470 return DAG.getNode(X86ISD::SHUFP, DL, VT, LowV, HighV,
8471 getV4X86ShuffleImm8ForMask(NewMask, DL, DAG));
8474 /// \brief Lower 4-lane 32-bit floating point shuffles.
8476 /// Uses instructions exclusively from the floating point unit to minimize
8477 /// domain crossing penalties, as these are sufficient to implement all v4f32
8479 static SDValue lowerV4F32VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
8480 const X86Subtarget *Subtarget,
8481 SelectionDAG &DAG) {
8483 assert(Op.getSimpleValueType() == MVT::v4f32 && "Bad shuffle type!");
8484 assert(V1.getSimpleValueType() == MVT::v4f32 && "Bad operand type!");
8485 assert(V2.getSimpleValueType() == MVT::v4f32 && "Bad operand type!");
8486 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
8487 ArrayRef<int> Mask = SVOp->getMask();
8488 assert(Mask.size() == 4 && "Unexpected mask size for v4 shuffle!");
8491 std::count_if(Mask.begin(), Mask.end(), [](int M) { return M >= 4; });
8493 if (NumV2Elements == 0) {
8494 // Check for being able to broadcast a single element.
8495 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v4f32, V1,
8496 Mask, Subtarget, DAG))
8499 // Use even/odd duplicate instructions for masks that match their pattern.
8500 if (Subtarget->hasSSE3()) {
8501 if (isShuffleEquivalent(V1, V2, Mask, {0, 0, 2, 2}))
8502 return DAG.getNode(X86ISD::MOVSLDUP, DL, MVT::v4f32, V1);
8503 if (isShuffleEquivalent(V1, V2, Mask, {1, 1, 3, 3}))
8504 return DAG.getNode(X86ISD::MOVSHDUP, DL, MVT::v4f32, V1);
8507 if (Subtarget->hasAVX()) {
8508 // If we have AVX, we can use VPERMILPS which will allow folding a load
8509 // into the shuffle.
8510 return DAG.getNode(X86ISD::VPERMILPI, DL, MVT::v4f32, V1,
8511 getV4X86ShuffleImm8ForMask(Mask, DL, DAG));
8514 // Otherwise, use a straight shuffle of a single input vector. We pass the
8515 // input vector to both operands to simulate this with a SHUFPS.
8516 return DAG.getNode(X86ISD::SHUFP, DL, MVT::v4f32, V1, V1,
8517 getV4X86ShuffleImm8ForMask(Mask, DL, DAG));
8520 // There are special ways we can lower some single-element blends. However, we
8521 // have custom ways we can lower more complex single-element blends below that
8522 // we defer to if both this and BLENDPS fail to match, so restrict this to
8523 // when the V2 input is targeting element 0 of the mask -- that is the fast
8525 if (NumV2Elements == 1 && Mask[0] >= 4)
8526 if (SDValue V = lowerVectorShuffleAsElementInsertion(DL, MVT::v4f32, V1, V2,
8527 Mask, Subtarget, DAG))
8530 if (Subtarget->hasSSE41()) {
8531 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v4f32, V1, V2, Mask,
8535 // Use INSERTPS if we can complete the shuffle efficiently.
8536 if (SDValue V = lowerVectorShuffleAsInsertPS(Op, V1, V2, Mask, DAG))
8539 if (!isSingleSHUFPSMask(Mask))
8540 if (SDValue BlendPerm = lowerVectorShuffleAsBlendAndPermute(
8541 DL, MVT::v4f32, V1, V2, Mask, DAG))
8545 // Use dedicated unpack instructions for masks that match their pattern.
8546 if (isShuffleEquivalent(V1, V2, Mask, {0, 4, 1, 5}))
8547 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v4f32, V1, V2);
8548 if (isShuffleEquivalent(V1, V2, Mask, {2, 6, 3, 7}))
8549 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v4f32, V1, V2);
8550 if (isShuffleEquivalent(V1, V2, Mask, {4, 0, 5, 1}))
8551 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v4f32, V2, V1);
8552 if (isShuffleEquivalent(V1, V2, Mask, {6, 2, 7, 3}))
8553 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v4f32, V2, V1);
8555 // Otherwise fall back to a SHUFPS lowering strategy.
8556 return lowerVectorShuffleWithSHUFPS(DL, MVT::v4f32, Mask, V1, V2, DAG);
8559 /// \brief Lower 4-lane i32 vector shuffles.
8561 /// We try to handle these with integer-domain shuffles where we can, but for
8562 /// blends we use the floating point domain blend instructions.
8563 static SDValue lowerV4I32VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
8564 const X86Subtarget *Subtarget,
8565 SelectionDAG &DAG) {
8567 assert(Op.getSimpleValueType() == MVT::v4i32 && "Bad shuffle type!");
8568 assert(V1.getSimpleValueType() == MVT::v4i32 && "Bad operand type!");
8569 assert(V2.getSimpleValueType() == MVT::v4i32 && "Bad operand type!");
8570 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
8571 ArrayRef<int> Mask = SVOp->getMask();
8572 assert(Mask.size() == 4 && "Unexpected mask size for v4 shuffle!");
8574 // Whenever we can lower this as a zext, that instruction is strictly faster
8575 // than any alternative. It also allows us to fold memory operands into the
8576 // shuffle in many cases.
8577 if (SDValue ZExt = lowerVectorShuffleAsZeroOrAnyExtend(DL, MVT::v4i32, V1, V2,
8578 Mask, Subtarget, DAG))
8582 std::count_if(Mask.begin(), Mask.end(), [](int M) { return M >= 4; });
8584 if (NumV2Elements == 0) {
8585 // Check for being able to broadcast a single element.
8586 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v4i32, V1,
8587 Mask, Subtarget, DAG))
8590 // Straight shuffle of a single input vector. For everything from SSE2
8591 // onward this has a single fast instruction with no scary immediates.
8592 // We coerce the shuffle pattern to be compatible with UNPCK instructions
8593 // but we aren't actually going to use the UNPCK instruction because doing
8594 // so prevents folding a load into this instruction or making a copy.
8595 const int UnpackLoMask[] = {0, 0, 1, 1};
8596 const int UnpackHiMask[] = {2, 2, 3, 3};
8597 if (isShuffleEquivalent(V1, V2, Mask, {0, 0, 1, 1}))
8598 Mask = UnpackLoMask;
8599 else if (isShuffleEquivalent(V1, V2, Mask, {2, 2, 3, 3}))
8600 Mask = UnpackHiMask;
8602 return DAG.getNode(X86ISD::PSHUFD, DL, MVT::v4i32, V1,
8603 getV4X86ShuffleImm8ForMask(Mask, DL, DAG));
8606 // Try to use shift instructions.
8608 lowerVectorShuffleAsShift(DL, MVT::v4i32, V1, V2, Mask, DAG))
8611 // There are special ways we can lower some single-element blends.
8612 if (NumV2Elements == 1)
8613 if (SDValue V = lowerVectorShuffleAsElementInsertion(DL, MVT::v4i32, V1, V2,
8614 Mask, Subtarget, DAG))
8617 // We have different paths for blend lowering, but they all must use the
8618 // *exact* same predicate.
8619 bool IsBlendSupported = Subtarget->hasSSE41();
8620 if (IsBlendSupported)
8621 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v4i32, V1, V2, Mask,
8625 if (SDValue Masked =
8626 lowerVectorShuffleAsBitMask(DL, MVT::v4i32, V1, V2, Mask, DAG))
8629 // Use dedicated unpack instructions for masks that match their pattern.
8630 if (isShuffleEquivalent(V1, V2, Mask, {0, 4, 1, 5}))
8631 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v4i32, V1, V2);
8632 if (isShuffleEquivalent(V1, V2, Mask, {2, 6, 3, 7}))
8633 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v4i32, V1, V2);
8634 if (isShuffleEquivalent(V1, V2, Mask, {4, 0, 5, 1}))
8635 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v4i32, V2, V1);
8636 if (isShuffleEquivalent(V1, V2, Mask, {6, 2, 7, 3}))
8637 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v4i32, V2, V1);
8639 // Try to use byte rotation instructions.
8640 // Its more profitable for pre-SSSE3 to use shuffles/unpacks.
8641 if (Subtarget->hasSSSE3())
8642 if (SDValue Rotate = lowerVectorShuffleAsByteRotate(
8643 DL, MVT::v4i32, V1, V2, Mask, Subtarget, DAG))
8646 // If we have direct support for blends, we should lower by decomposing into
8647 // a permute. That will be faster than the domain cross.
8648 if (IsBlendSupported)
8649 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v4i32, V1, V2,
8652 // Try to lower by permuting the inputs into an unpack instruction.
8653 if (SDValue Unpack = lowerVectorShuffleAsPermuteAndUnpack(DL, MVT::v4i32, V1,
8657 // We implement this with SHUFPS because it can blend from two vectors.
8658 // Because we're going to eventually use SHUFPS, we use SHUFPS even to build
8659 // up the inputs, bypassing domain shift penalties that we would encur if we
8660 // directly used PSHUFD on Nehalem and older. For newer chips, this isn't
8662 return DAG.getBitcast(
8664 DAG.getVectorShuffle(MVT::v4f32, DL, DAG.getBitcast(MVT::v4f32, V1),
8665 DAG.getBitcast(MVT::v4f32, V2), Mask));
8668 /// \brief Lowering of single-input v8i16 shuffles is the cornerstone of SSE2
8669 /// shuffle lowering, and the most complex part.
8671 /// The lowering strategy is to try to form pairs of input lanes which are
8672 /// targeted at the same half of the final vector, and then use a dword shuffle
8673 /// to place them onto the right half, and finally unpack the paired lanes into
8674 /// their final position.
8676 /// The exact breakdown of how to form these dword pairs and align them on the
8677 /// correct sides is really tricky. See the comments within the function for
8678 /// more of the details.
8680 /// This code also handles repeated 128-bit lanes of v8i16 shuffles, but each
8681 /// lane must shuffle the *exact* same way. In fact, you must pass a v8 Mask to
8682 /// this routine for it to work correctly. To shuffle a 256-bit or 512-bit i16
8683 /// vector, form the analogous 128-bit 8-element Mask.
8684 static SDValue lowerV8I16GeneralSingleInputVectorShuffle(
8685 SDLoc DL, MVT VT, SDValue V, MutableArrayRef<int> Mask,
8686 const X86Subtarget *Subtarget, SelectionDAG &DAG) {
8687 assert(VT.getScalarType() == MVT::i16 && "Bad input type!");
8688 MVT PSHUFDVT = MVT::getVectorVT(MVT::i32, VT.getVectorNumElements() / 2);
8690 assert(Mask.size() == 8 && "Shuffle mask length doen't match!");
8691 MutableArrayRef<int> LoMask = Mask.slice(0, 4);
8692 MutableArrayRef<int> HiMask = Mask.slice(4, 4);
8694 SmallVector<int, 4> LoInputs;
8695 std::copy_if(LoMask.begin(), LoMask.end(), std::back_inserter(LoInputs),
8696 [](int M) { return M >= 0; });
8697 std::sort(LoInputs.begin(), LoInputs.end());
8698 LoInputs.erase(std::unique(LoInputs.begin(), LoInputs.end()), LoInputs.end());
8699 SmallVector<int, 4> HiInputs;
8700 std::copy_if(HiMask.begin(), HiMask.end(), std::back_inserter(HiInputs),
8701 [](int M) { return M >= 0; });
8702 std::sort(HiInputs.begin(), HiInputs.end());
8703 HiInputs.erase(std::unique(HiInputs.begin(), HiInputs.end()), HiInputs.end());
8705 std::lower_bound(LoInputs.begin(), LoInputs.end(), 4) - LoInputs.begin();
8706 int NumHToL = LoInputs.size() - NumLToL;
8708 std::lower_bound(HiInputs.begin(), HiInputs.end(), 4) - HiInputs.begin();
8709 int NumHToH = HiInputs.size() - NumLToH;
8710 MutableArrayRef<int> LToLInputs(LoInputs.data(), NumLToL);
8711 MutableArrayRef<int> LToHInputs(HiInputs.data(), NumLToH);
8712 MutableArrayRef<int> HToLInputs(LoInputs.data() + NumLToL, NumHToL);
8713 MutableArrayRef<int> HToHInputs(HiInputs.data() + NumLToH, NumHToH);
8715 // Simplify the 1-into-3 and 3-into-1 cases with a single pshufd. For all
8716 // such inputs we can swap two of the dwords across the half mark and end up
8717 // with <=2 inputs to each half in each half. Once there, we can fall through
8718 // to the generic code below. For example:
8720 // Input: [a, b, c, d, e, f, g, h] -PSHUFD[0,2,1,3]-> [a, b, e, f, c, d, g, h]
8721 // Mask: [0, 1, 2, 7, 4, 5, 6, 3] -----------------> [0, 1, 4, 7, 2, 3, 6, 5]
8723 // However in some very rare cases we have a 1-into-3 or 3-into-1 on one half
8724 // and an existing 2-into-2 on the other half. In this case we may have to
8725 // pre-shuffle the 2-into-2 half to avoid turning it into a 3-into-1 or
8726 // 1-into-3 which could cause us to cycle endlessly fixing each side in turn.
8727 // Fortunately, we don't have to handle anything but a 2-into-2 pattern
8728 // because any other situation (including a 3-into-1 or 1-into-3 in the other
8729 // half than the one we target for fixing) will be fixed when we re-enter this
8730 // path. We will also combine away any sequence of PSHUFD instructions that
8731 // result into a single instruction. Here is an example of the tricky case:
8733 // Input: [a, b, c, d, e, f, g, h] -PSHUFD[0,2,1,3]-> [a, b, e, f, c, d, g, h]
8734 // Mask: [3, 7, 1, 0, 2, 7, 3, 5] -THIS-IS-BAD!!!!-> [5, 7, 1, 0, 4, 7, 5, 3]
8736 // This now has a 1-into-3 in the high half! Instead, we do two shuffles:
8738 // Input: [a, b, c, d, e, f, g, h] PSHUFHW[0,2,1,3]-> [a, b, c, d, e, g, f, h]
8739 // Mask: [3, 7, 1, 0, 2, 7, 3, 5] -----------------> [3, 7, 1, 0, 2, 7, 3, 6]
8741 // Input: [a, b, c, d, e, g, f, h] -PSHUFD[0,2,1,3]-> [a, b, e, g, c, d, f, h]
8742 // Mask: [3, 7, 1, 0, 2, 7, 3, 6] -----------------> [5, 7, 1, 0, 4, 7, 5, 6]
8744 // The result is fine to be handled by the generic logic.
8745 auto balanceSides = [&](ArrayRef<int> AToAInputs, ArrayRef<int> BToAInputs,
8746 ArrayRef<int> BToBInputs, ArrayRef<int> AToBInputs,
8747 int AOffset, int BOffset) {
8748 assert((AToAInputs.size() == 3 || AToAInputs.size() == 1) &&
8749 "Must call this with A having 3 or 1 inputs from the A half.");
8750 assert((BToAInputs.size() == 1 || BToAInputs.size() == 3) &&
8751 "Must call this with B having 1 or 3 inputs from the B half.");
8752 assert(AToAInputs.size() + BToAInputs.size() == 4 &&
8753 "Must call this with either 3:1 or 1:3 inputs (summing to 4).");
8755 bool ThreeAInputs = AToAInputs.size() == 3;
8757 // Compute the index of dword with only one word among the three inputs in
8758 // a half by taking the sum of the half with three inputs and subtracting
8759 // the sum of the actual three inputs. The difference is the remaining
8762 int &TripleDWord = ThreeAInputs ? ADWord : BDWord;
8763 int &OneInputDWord = ThreeAInputs ? BDWord : ADWord;
8764 int TripleInputOffset = ThreeAInputs ? AOffset : BOffset;
8765 ArrayRef<int> TripleInputs = ThreeAInputs ? AToAInputs : BToAInputs;
8766 int OneInput = ThreeAInputs ? BToAInputs[0] : AToAInputs[0];
8767 int TripleInputSum = 0 + 1 + 2 + 3 + (4 * TripleInputOffset);
8768 int TripleNonInputIdx =
8769 TripleInputSum - std::accumulate(TripleInputs.begin(), TripleInputs.end(), 0);
8770 TripleDWord = TripleNonInputIdx / 2;
8772 // We use xor with one to compute the adjacent DWord to whichever one the
8774 OneInputDWord = (OneInput / 2) ^ 1;
8776 // Check for one tricky case: We're fixing a 3<-1 or a 1<-3 shuffle for AToA
8777 // and BToA inputs. If there is also such a problem with the BToB and AToB
8778 // inputs, we don't try to fix it necessarily -- we'll recurse and see it in
8779 // the next pass. However, if we have a 2<-2 in the BToB and AToB inputs, it
8780 // is essential that we don't *create* a 3<-1 as then we might oscillate.
8781 if (BToBInputs.size() == 2 && AToBInputs.size() == 2) {
8782 // Compute how many inputs will be flipped by swapping these DWords. We
8784 // to balance this to ensure we don't form a 3-1 shuffle in the other
8786 int NumFlippedAToBInputs =
8787 std::count(AToBInputs.begin(), AToBInputs.end(), 2 * ADWord) +
8788 std::count(AToBInputs.begin(), AToBInputs.end(), 2 * ADWord + 1);
8789 int NumFlippedBToBInputs =
8790 std::count(BToBInputs.begin(), BToBInputs.end(), 2 * BDWord) +
8791 std::count(BToBInputs.begin(), BToBInputs.end(), 2 * BDWord + 1);
8792 if ((NumFlippedAToBInputs == 1 &&
8793 (NumFlippedBToBInputs == 0 || NumFlippedBToBInputs == 2)) ||
8794 (NumFlippedBToBInputs == 1 &&
8795 (NumFlippedAToBInputs == 0 || NumFlippedAToBInputs == 2))) {
8796 // We choose whether to fix the A half or B half based on whether that
8797 // half has zero flipped inputs. At zero, we may not be able to fix it
8798 // with that half. We also bias towards fixing the B half because that
8799 // will more commonly be the high half, and we have to bias one way.
8800 auto FixFlippedInputs = [&V, &DL, &Mask, &DAG](int PinnedIdx, int DWord,
8801 ArrayRef<int> Inputs) {
8802 int FixIdx = PinnedIdx ^ 1; // The adjacent slot to the pinned slot.
8803 bool IsFixIdxInput = std::find(Inputs.begin(), Inputs.end(),
8804 PinnedIdx ^ 1) != Inputs.end();
8805 // Determine whether the free index is in the flipped dword or the
8806 // unflipped dword based on where the pinned index is. We use this bit
8807 // in an xor to conditionally select the adjacent dword.
8808 int FixFreeIdx = 2 * (DWord ^ (PinnedIdx / 2 == DWord));
8809 bool IsFixFreeIdxInput = std::find(Inputs.begin(), Inputs.end(),
8810 FixFreeIdx) != Inputs.end();
8811 if (IsFixIdxInput == IsFixFreeIdxInput)
8813 IsFixFreeIdxInput = std::find(Inputs.begin(), Inputs.end(),
8814 FixFreeIdx) != Inputs.end();
8815 assert(IsFixIdxInput != IsFixFreeIdxInput &&
8816 "We need to be changing the number of flipped inputs!");
8817 int PSHUFHalfMask[] = {0, 1, 2, 3};
8818 std::swap(PSHUFHalfMask[FixFreeIdx % 4], PSHUFHalfMask[FixIdx % 4]);
8819 V = DAG.getNode(FixIdx < 4 ? X86ISD::PSHUFLW : X86ISD::PSHUFHW, DL,
8821 getV4X86ShuffleImm8ForMask(PSHUFHalfMask, DL, DAG));
8824 if (M != -1 && M == FixIdx)
8826 else if (M != -1 && M == FixFreeIdx)
8829 if (NumFlippedBToBInputs != 0) {
8831 BToAInputs.size() == 3 ? TripleNonInputIdx : OneInput;
8832 FixFlippedInputs(BPinnedIdx, BDWord, BToBInputs);
8834 assert(NumFlippedAToBInputs != 0 && "Impossible given predicates!");
8835 int APinnedIdx = ThreeAInputs ? TripleNonInputIdx : OneInput;
8836 FixFlippedInputs(APinnedIdx, ADWord, AToBInputs);
8841 int PSHUFDMask[] = {0, 1, 2, 3};
8842 PSHUFDMask[ADWord] = BDWord;
8843 PSHUFDMask[BDWord] = ADWord;
8846 DAG.getNode(X86ISD::PSHUFD, DL, PSHUFDVT, DAG.getBitcast(PSHUFDVT, V),
8847 getV4X86ShuffleImm8ForMask(PSHUFDMask, DL, DAG)));
8849 // Adjust the mask to match the new locations of A and B.
8851 if (M != -1 && M/2 == ADWord)
8852 M = 2 * BDWord + M % 2;
8853 else if (M != -1 && M/2 == BDWord)
8854 M = 2 * ADWord + M % 2;
8856 // Recurse back into this routine to re-compute state now that this isn't
8857 // a 3 and 1 problem.
8858 return lowerV8I16GeneralSingleInputVectorShuffle(DL, VT, V, Mask, Subtarget,
8861 if ((NumLToL == 3 && NumHToL == 1) || (NumLToL == 1 && NumHToL == 3))
8862 return balanceSides(LToLInputs, HToLInputs, HToHInputs, LToHInputs, 0, 4);
8863 else if ((NumHToH == 3 && NumLToH == 1) || (NumHToH == 1 && NumLToH == 3))
8864 return balanceSides(HToHInputs, LToHInputs, LToLInputs, HToLInputs, 4, 0);
8866 // At this point there are at most two inputs to the low and high halves from
8867 // each half. That means the inputs can always be grouped into dwords and
8868 // those dwords can then be moved to the correct half with a dword shuffle.
8869 // We use at most one low and one high word shuffle to collect these paired
8870 // inputs into dwords, and finally a dword shuffle to place them.
8871 int PSHUFLMask[4] = {-1, -1, -1, -1};
8872 int PSHUFHMask[4] = {-1, -1, -1, -1};
8873 int PSHUFDMask[4] = {-1, -1, -1, -1};
8875 // First fix the masks for all the inputs that are staying in their
8876 // original halves. This will then dictate the targets of the cross-half
8878 auto fixInPlaceInputs =
8879 [&PSHUFDMask](ArrayRef<int> InPlaceInputs, ArrayRef<int> IncomingInputs,
8880 MutableArrayRef<int> SourceHalfMask,
8881 MutableArrayRef<int> HalfMask, int HalfOffset) {
8882 if (InPlaceInputs.empty())
8884 if (InPlaceInputs.size() == 1) {
8885 SourceHalfMask[InPlaceInputs[0] - HalfOffset] =
8886 InPlaceInputs[0] - HalfOffset;
8887 PSHUFDMask[InPlaceInputs[0] / 2] = InPlaceInputs[0] / 2;
8890 if (IncomingInputs.empty()) {
8891 // Just fix all of the in place inputs.
8892 for (int Input : InPlaceInputs) {
8893 SourceHalfMask[Input - HalfOffset] = Input - HalfOffset;
8894 PSHUFDMask[Input / 2] = Input / 2;
8899 assert(InPlaceInputs.size() == 2 && "Cannot handle 3 or 4 inputs!");
8900 SourceHalfMask[InPlaceInputs[0] - HalfOffset] =
8901 InPlaceInputs[0] - HalfOffset;
8902 // Put the second input next to the first so that they are packed into
8903 // a dword. We find the adjacent index by toggling the low bit.
8904 int AdjIndex = InPlaceInputs[0] ^ 1;
8905 SourceHalfMask[AdjIndex - HalfOffset] = InPlaceInputs[1] - HalfOffset;
8906 std::replace(HalfMask.begin(), HalfMask.end(), InPlaceInputs[1], AdjIndex);
8907 PSHUFDMask[AdjIndex / 2] = AdjIndex / 2;
8909 fixInPlaceInputs(LToLInputs, HToLInputs, PSHUFLMask, LoMask, 0);
8910 fixInPlaceInputs(HToHInputs, LToHInputs, PSHUFHMask, HiMask, 4);
8912 // Now gather the cross-half inputs and place them into a free dword of
8913 // their target half.
8914 // FIXME: This operation could almost certainly be simplified dramatically to
8915 // look more like the 3-1 fixing operation.
8916 auto moveInputsToRightHalf = [&PSHUFDMask](
8917 MutableArrayRef<int> IncomingInputs, ArrayRef<int> ExistingInputs,
8918 MutableArrayRef<int> SourceHalfMask, MutableArrayRef<int> HalfMask,
8919 MutableArrayRef<int> FinalSourceHalfMask, int SourceOffset,
8921 auto isWordClobbered = [](ArrayRef<int> SourceHalfMask, int Word) {
8922 return SourceHalfMask[Word] != -1 && SourceHalfMask[Word] != Word;
8924 auto isDWordClobbered = [&isWordClobbered](ArrayRef<int> SourceHalfMask,
8926 int LowWord = Word & ~1;
8927 int HighWord = Word | 1;
8928 return isWordClobbered(SourceHalfMask, LowWord) ||
8929 isWordClobbered(SourceHalfMask, HighWord);
8932 if (IncomingInputs.empty())
8935 if (ExistingInputs.empty()) {
8936 // Map any dwords with inputs from them into the right half.
8937 for (int Input : IncomingInputs) {
8938 // If the source half mask maps over the inputs, turn those into
8939 // swaps and use the swapped lane.
8940 if (isWordClobbered(SourceHalfMask, Input - SourceOffset)) {
8941 if (SourceHalfMask[SourceHalfMask[Input - SourceOffset]] == -1) {
8942 SourceHalfMask[SourceHalfMask[Input - SourceOffset]] =
8943 Input - SourceOffset;
8944 // We have to swap the uses in our half mask in one sweep.
8945 for (int &M : HalfMask)
8946 if (M == SourceHalfMask[Input - SourceOffset] + SourceOffset)
8948 else if (M == Input)
8949 M = SourceHalfMask[Input - SourceOffset] + SourceOffset;
8951 assert(SourceHalfMask[SourceHalfMask[Input - SourceOffset]] ==
8952 Input - SourceOffset &&
8953 "Previous placement doesn't match!");
8955 // Note that this correctly re-maps both when we do a swap and when
8956 // we observe the other side of the swap above. We rely on that to
8957 // avoid swapping the members of the input list directly.
8958 Input = SourceHalfMask[Input - SourceOffset] + SourceOffset;
8961 // Map the input's dword into the correct half.
8962 if (PSHUFDMask[(Input - SourceOffset + DestOffset) / 2] == -1)
8963 PSHUFDMask[(Input - SourceOffset + DestOffset) / 2] = Input / 2;
8965 assert(PSHUFDMask[(Input - SourceOffset + DestOffset) / 2] ==
8967 "Previous placement doesn't match!");
8970 // And just directly shift any other-half mask elements to be same-half
8971 // as we will have mirrored the dword containing the element into the
8972 // same position within that half.
8973 for (int &M : HalfMask)
8974 if (M >= SourceOffset && M < SourceOffset + 4) {
8975 M = M - SourceOffset + DestOffset;
8976 assert(M >= 0 && "This should never wrap below zero!");
8981 // Ensure we have the input in a viable dword of its current half. This
8982 // is particularly tricky because the original position may be clobbered
8983 // by inputs being moved and *staying* in that half.
8984 if (IncomingInputs.size() == 1) {
8985 if (isWordClobbered(SourceHalfMask, IncomingInputs[0] - SourceOffset)) {
8986 int InputFixed = std::find(std::begin(SourceHalfMask),
8987 std::end(SourceHalfMask), -1) -
8988 std::begin(SourceHalfMask) + SourceOffset;
8989 SourceHalfMask[InputFixed - SourceOffset] =
8990 IncomingInputs[0] - SourceOffset;
8991 std::replace(HalfMask.begin(), HalfMask.end(), IncomingInputs[0],
8993 IncomingInputs[0] = InputFixed;
8995 } else if (IncomingInputs.size() == 2) {
8996 if (IncomingInputs[0] / 2 != IncomingInputs[1] / 2 ||
8997 isDWordClobbered(SourceHalfMask, IncomingInputs[0] - SourceOffset)) {
8998 // We have two non-adjacent or clobbered inputs we need to extract from
8999 // the source half. To do this, we need to map them into some adjacent
9000 // dword slot in the source mask.
9001 int InputsFixed[2] = {IncomingInputs[0] - SourceOffset,
9002 IncomingInputs[1] - SourceOffset};
9004 // If there is a free slot in the source half mask adjacent to one of
9005 // the inputs, place the other input in it. We use (Index XOR 1) to
9006 // compute an adjacent index.
9007 if (!isWordClobbered(SourceHalfMask, InputsFixed[0]) &&
9008 SourceHalfMask[InputsFixed[0] ^ 1] == -1) {
9009 SourceHalfMask[InputsFixed[0]] = InputsFixed[0];
9010 SourceHalfMask[InputsFixed[0] ^ 1] = InputsFixed[1];
9011 InputsFixed[1] = InputsFixed[0] ^ 1;
9012 } else if (!isWordClobbered(SourceHalfMask, InputsFixed[1]) &&
9013 SourceHalfMask[InputsFixed[1] ^ 1] == -1) {
9014 SourceHalfMask[InputsFixed[1]] = InputsFixed[1];
9015 SourceHalfMask[InputsFixed[1] ^ 1] = InputsFixed[0];
9016 InputsFixed[0] = InputsFixed[1] ^ 1;
9017 } else if (SourceHalfMask[2 * ((InputsFixed[0] / 2) ^ 1)] == -1 &&
9018 SourceHalfMask[2 * ((InputsFixed[0] / 2) ^ 1) + 1] == -1) {
9019 // The two inputs are in the same DWord but it is clobbered and the
9020 // adjacent DWord isn't used at all. Move both inputs to the free
9022 SourceHalfMask[2 * ((InputsFixed[0] / 2) ^ 1)] = InputsFixed[0];
9023 SourceHalfMask[2 * ((InputsFixed[0] / 2) ^ 1) + 1] = InputsFixed[1];
9024 InputsFixed[0] = 2 * ((InputsFixed[0] / 2) ^ 1);
9025 InputsFixed[1] = 2 * ((InputsFixed[0] / 2) ^ 1) + 1;
9027 // The only way we hit this point is if there is no clobbering
9028 // (because there are no off-half inputs to this half) and there is no
9029 // free slot adjacent to one of the inputs. In this case, we have to
9030 // swap an input with a non-input.
9031 for (int i = 0; i < 4; ++i)
9032 assert((SourceHalfMask[i] == -1 || SourceHalfMask[i] == i) &&
9033 "We can't handle any clobbers here!");
9034 assert(InputsFixed[1] != (InputsFixed[0] ^ 1) &&
9035 "Cannot have adjacent inputs here!");
9037 SourceHalfMask[InputsFixed[0] ^ 1] = InputsFixed[1];
9038 SourceHalfMask[InputsFixed[1]] = InputsFixed[0] ^ 1;
9040 // We also have to update the final source mask in this case because
9041 // it may need to undo the above swap.
9042 for (int &M : FinalSourceHalfMask)
9043 if (M == (InputsFixed[0] ^ 1) + SourceOffset)
9044 M = InputsFixed[1] + SourceOffset;
9045 else if (M == InputsFixed[1] + SourceOffset)
9046 M = (InputsFixed[0] ^ 1) + SourceOffset;
9048 InputsFixed[1] = InputsFixed[0] ^ 1;
9051 // Point everything at the fixed inputs.
9052 for (int &M : HalfMask)
9053 if (M == IncomingInputs[0])
9054 M = InputsFixed[0] + SourceOffset;
9055 else if (M == IncomingInputs[1])
9056 M = InputsFixed[1] + SourceOffset;
9058 IncomingInputs[0] = InputsFixed[0] + SourceOffset;
9059 IncomingInputs[1] = InputsFixed[1] + SourceOffset;
9062 llvm_unreachable("Unhandled input size!");
9065 // Now hoist the DWord down to the right half.
9066 int FreeDWord = (PSHUFDMask[DestOffset / 2] == -1 ? 0 : 1) + DestOffset / 2;
9067 assert(PSHUFDMask[FreeDWord] == -1 && "DWord not free");
9068 PSHUFDMask[FreeDWord] = IncomingInputs[0] / 2;
9069 for (int &M : HalfMask)
9070 for (int Input : IncomingInputs)
9072 M = FreeDWord * 2 + Input % 2;
9074 moveInputsToRightHalf(HToLInputs, LToLInputs, PSHUFHMask, LoMask, HiMask,
9075 /*SourceOffset*/ 4, /*DestOffset*/ 0);
9076 moveInputsToRightHalf(LToHInputs, HToHInputs, PSHUFLMask, HiMask, LoMask,
9077 /*SourceOffset*/ 0, /*DestOffset*/ 4);
9079 // Now enact all the shuffles we've computed to move the inputs into their
9081 if (!isNoopShuffleMask(PSHUFLMask))
9082 V = DAG.getNode(X86ISD::PSHUFLW, DL, VT, V,
9083 getV4X86ShuffleImm8ForMask(PSHUFLMask, DL, DAG));
9084 if (!isNoopShuffleMask(PSHUFHMask))
9085 V = DAG.getNode(X86ISD::PSHUFHW, DL, VT, V,
9086 getV4X86ShuffleImm8ForMask(PSHUFHMask, DL, DAG));
9087 if (!isNoopShuffleMask(PSHUFDMask))
9090 DAG.getNode(X86ISD::PSHUFD, DL, PSHUFDVT, DAG.getBitcast(PSHUFDVT, V),
9091 getV4X86ShuffleImm8ForMask(PSHUFDMask, DL, DAG)));
9093 // At this point, each half should contain all its inputs, and we can then
9094 // just shuffle them into their final position.
9095 assert(std::count_if(LoMask.begin(), LoMask.end(),
9096 [](int M) { return M >= 4; }) == 0 &&
9097 "Failed to lift all the high half inputs to the low mask!");
9098 assert(std::count_if(HiMask.begin(), HiMask.end(),
9099 [](int M) { return M >= 0 && M < 4; }) == 0 &&
9100 "Failed to lift all the low half inputs to the high mask!");
9102 // Do a half shuffle for the low mask.
9103 if (!isNoopShuffleMask(LoMask))
9104 V = DAG.getNode(X86ISD::PSHUFLW, DL, VT, V,
9105 getV4X86ShuffleImm8ForMask(LoMask, DL, DAG));
9107 // Do a half shuffle with the high mask after shifting its values down.
9108 for (int &M : HiMask)
9111 if (!isNoopShuffleMask(HiMask))
9112 V = DAG.getNode(X86ISD::PSHUFHW, DL, VT, V,
9113 getV4X86ShuffleImm8ForMask(HiMask, DL, DAG));
9118 /// \brief Helper to form a PSHUFB-based shuffle+blend.
9119 static SDValue lowerVectorShuffleAsPSHUFB(SDLoc DL, MVT VT, SDValue V1,
9120 SDValue V2, ArrayRef<int> Mask,
9121 SelectionDAG &DAG, bool &V1InUse,
9123 SmallBitVector Zeroable = computeZeroableShuffleElements(Mask, V1, V2);
9129 int Size = Mask.size();
9130 int Scale = 16 / Size;
9131 for (int i = 0; i < 16; ++i) {
9132 if (Mask[i / Scale] == -1) {
9133 V1Mask[i] = V2Mask[i] = DAG.getUNDEF(MVT::i8);
9135 const int ZeroMask = 0x80;
9136 int V1Idx = Mask[i / Scale] < Size ? Mask[i / Scale] * Scale + i % Scale
9138 int V2Idx = Mask[i / Scale] < Size
9140 : (Mask[i / Scale] - Size) * Scale + i % Scale;
9141 if (Zeroable[i / Scale])
9142 V1Idx = V2Idx = ZeroMask;
9143 V1Mask[i] = DAG.getConstant(V1Idx, DL, MVT::i8);
9144 V2Mask[i] = DAG.getConstant(V2Idx, DL, MVT::i8);
9145 V1InUse |= (ZeroMask != V1Idx);
9146 V2InUse |= (ZeroMask != V2Idx);
9151 V1 = DAG.getNode(X86ISD::PSHUFB, DL, MVT::v16i8,
9152 DAG.getBitcast(MVT::v16i8, V1),
9153 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v16i8, V1Mask));
9155 V2 = DAG.getNode(X86ISD::PSHUFB, DL, MVT::v16i8,
9156 DAG.getBitcast(MVT::v16i8, V2),
9157 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v16i8, V2Mask));
9159 // If we need shuffled inputs from both, blend the two.
9161 if (V1InUse && V2InUse)
9162 V = DAG.getNode(ISD::OR, DL, MVT::v16i8, V1, V2);
9164 V = V1InUse ? V1 : V2;
9166 // Cast the result back to the correct type.
9167 return DAG.getBitcast(VT, V);
9170 /// \brief Generic lowering of 8-lane i16 shuffles.
9172 /// This handles both single-input shuffles and combined shuffle/blends with
9173 /// two inputs. The single input shuffles are immediately delegated to
9174 /// a dedicated lowering routine.
9176 /// The blends are lowered in one of three fundamental ways. If there are few
9177 /// enough inputs, it delegates to a basic UNPCK-based strategy. If the shuffle
9178 /// of the input is significantly cheaper when lowered as an interleaving of
9179 /// the two inputs, try to interleave them. Otherwise, blend the low and high
9180 /// halves of the inputs separately (making them have relatively few inputs)
9181 /// and then concatenate them.
9182 static SDValue lowerV8I16VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
9183 const X86Subtarget *Subtarget,
9184 SelectionDAG &DAG) {
9186 assert(Op.getSimpleValueType() == MVT::v8i16 && "Bad shuffle type!");
9187 assert(V1.getSimpleValueType() == MVT::v8i16 && "Bad operand type!");
9188 assert(V2.getSimpleValueType() == MVT::v8i16 && "Bad operand type!");
9189 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
9190 ArrayRef<int> OrigMask = SVOp->getMask();
9191 int MaskStorage[8] = {OrigMask[0], OrigMask[1], OrigMask[2], OrigMask[3],
9192 OrigMask[4], OrigMask[5], OrigMask[6], OrigMask[7]};
9193 MutableArrayRef<int> Mask(MaskStorage);
9195 assert(Mask.size() == 8 && "Unexpected mask size for v8 shuffle!");
9197 // Whenever we can lower this as a zext, that instruction is strictly faster
9198 // than any alternative.
9199 if (SDValue ZExt = lowerVectorShuffleAsZeroOrAnyExtend(
9200 DL, MVT::v8i16, V1, V2, OrigMask, Subtarget, DAG))
9203 auto isV1 = [](int M) { return M >= 0 && M < 8; };
9205 auto isV2 = [](int M) { return M >= 8; };
9207 int NumV2Inputs = std::count_if(Mask.begin(), Mask.end(), isV2);
9209 if (NumV2Inputs == 0) {
9210 // Check for being able to broadcast a single element.
9211 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v8i16, V1,
9212 Mask, Subtarget, DAG))
9215 // Try to use shift instructions.
9217 lowerVectorShuffleAsShift(DL, MVT::v8i16, V1, V1, Mask, DAG))
9220 // Use dedicated unpack instructions for masks that match their pattern.
9221 if (isShuffleEquivalent(V1, V1, Mask, {0, 0, 1, 1, 2, 2, 3, 3}))
9222 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v8i16, V1, V1);
9223 if (isShuffleEquivalent(V1, V1, Mask, {4, 4, 5, 5, 6, 6, 7, 7}))
9224 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v8i16, V1, V1);
9226 // Try to use byte rotation instructions.
9227 if (SDValue Rotate = lowerVectorShuffleAsByteRotate(DL, MVT::v8i16, V1, V1,
9228 Mask, Subtarget, DAG))
9231 return lowerV8I16GeneralSingleInputVectorShuffle(DL, MVT::v8i16, V1, Mask,
9235 assert(std::any_of(Mask.begin(), Mask.end(), isV1) &&
9236 "All single-input shuffles should be canonicalized to be V1-input "
9239 // Try to use shift instructions.
9241 lowerVectorShuffleAsShift(DL, MVT::v8i16, V1, V2, Mask, DAG))
9244 // See if we can use SSE4A Extraction / Insertion.
9245 if (Subtarget->hasSSE4A())
9246 if (SDValue V = lowerVectorShuffleWithSSE4A(DL, MVT::v8i16, V1, V2, Mask, DAG))
9249 // There are special ways we can lower some single-element blends.
9250 if (NumV2Inputs == 1)
9251 if (SDValue V = lowerVectorShuffleAsElementInsertion(DL, MVT::v8i16, V1, V2,
9252 Mask, Subtarget, DAG))
9255 // We have different paths for blend lowering, but they all must use the
9256 // *exact* same predicate.
9257 bool IsBlendSupported = Subtarget->hasSSE41();
9258 if (IsBlendSupported)
9259 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v8i16, V1, V2, Mask,
9263 if (SDValue Masked =
9264 lowerVectorShuffleAsBitMask(DL, MVT::v8i16, V1, V2, Mask, DAG))
9267 // Use dedicated unpack instructions for masks that match their pattern.
9268 if (isShuffleEquivalent(V1, V2, Mask, {0, 8, 1, 9, 2, 10, 3, 11}))
9269 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v8i16, V1, V2);
9270 if (isShuffleEquivalent(V1, V2, Mask, {4, 12, 5, 13, 6, 14, 7, 15}))
9271 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v8i16, V1, V2);
9273 // Try to use byte rotation instructions.
9274 if (SDValue Rotate = lowerVectorShuffleAsByteRotate(
9275 DL, MVT::v8i16, V1, V2, Mask, Subtarget, DAG))
9278 if (SDValue BitBlend =
9279 lowerVectorShuffleAsBitBlend(DL, MVT::v8i16, V1, V2, Mask, DAG))
9282 if (SDValue Unpack = lowerVectorShuffleAsPermuteAndUnpack(DL, MVT::v8i16, V1,
9286 // If we can't directly blend but can use PSHUFB, that will be better as it
9287 // can both shuffle and set up the inefficient blend.
9288 if (!IsBlendSupported && Subtarget->hasSSSE3()) {
9289 bool V1InUse, V2InUse;
9290 return lowerVectorShuffleAsPSHUFB(DL, MVT::v8i16, V1, V2, Mask, DAG,
9294 // We can always bit-blend if we have to so the fallback strategy is to
9295 // decompose into single-input permutes and blends.
9296 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v8i16, V1, V2,
9300 /// \brief Check whether a compaction lowering can be done by dropping even
9301 /// elements and compute how many times even elements must be dropped.
9303 /// This handles shuffles which take every Nth element where N is a power of
9304 /// two. Example shuffle masks:
9306 /// N = 1: 0, 2, 4, 6, 8, 10, 12, 14, 0, 2, 4, 6, 8, 10, 12, 14
9307 /// N = 1: 0, 2, 4, 6, 8, 10, 12, 14, 16, 18, 20, 22, 24, 26, 28, 30
9308 /// N = 2: 0, 4, 8, 12, 0, 4, 8, 12, 0, 4, 8, 12, 0, 4, 8, 12
9309 /// N = 2: 0, 4, 8, 12, 16, 20, 24, 28, 0, 4, 8, 12, 16, 20, 24, 28
9310 /// N = 3: 0, 8, 0, 8, 0, 8, 0, 8, 0, 8, 0, 8, 0, 8, 0, 8
9311 /// N = 3: 0, 8, 16, 24, 0, 8, 16, 24, 0, 8, 16, 24, 0, 8, 16, 24
9313 /// Any of these lanes can of course be undef.
9315 /// This routine only supports N <= 3.
9316 /// FIXME: Evaluate whether either AVX or AVX-512 have any opportunities here
9319 /// \returns N above, or the number of times even elements must be dropped if
9320 /// there is such a number. Otherwise returns zero.
9321 static int canLowerByDroppingEvenElements(ArrayRef<int> Mask) {
9322 // Figure out whether we're looping over two inputs or just one.
9323 bool IsSingleInput = isSingleInputShuffleMask(Mask);
9325 // The modulus for the shuffle vector entries is based on whether this is
9326 // a single input or not.
9327 int ShuffleModulus = Mask.size() * (IsSingleInput ? 1 : 2);
9328 assert(isPowerOf2_32((uint32_t)ShuffleModulus) &&
9329 "We should only be called with masks with a power-of-2 size!");
9331 uint64_t ModMask = (uint64_t)ShuffleModulus - 1;
9333 // We track whether the input is viable for all power-of-2 strides 2^1, 2^2,
9334 // and 2^3 simultaneously. This is because we may have ambiguity with
9335 // partially undef inputs.
9336 bool ViableForN[3] = {true, true, true};
9338 for (int i = 0, e = Mask.size(); i < e; ++i) {
9339 // Ignore undef lanes, we'll optimistically collapse them to the pattern we
9344 bool IsAnyViable = false;
9345 for (unsigned j = 0; j != array_lengthof(ViableForN); ++j)
9346 if (ViableForN[j]) {
9349 // The shuffle mask must be equal to (i * 2^N) % M.
9350 if ((uint64_t)Mask[i] == (((uint64_t)i << N) & ModMask))
9353 ViableForN[j] = false;
9355 // Early exit if we exhaust the possible powers of two.
9360 for (unsigned j = 0; j != array_lengthof(ViableForN); ++j)
9364 // Return 0 as there is no viable power of two.
9368 /// \brief Generic lowering of v16i8 shuffles.
9370 /// This is a hybrid strategy to lower v16i8 vectors. It first attempts to
9371 /// detect any complexity reducing interleaving. If that doesn't help, it uses
9372 /// UNPCK to spread the i8 elements across two i16-element vectors, and uses
9373 /// the existing lowering for v8i16 blends on each half, finally PACK-ing them
9375 static SDValue lowerV16I8VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
9376 const X86Subtarget *Subtarget,
9377 SelectionDAG &DAG) {
9379 assert(Op.getSimpleValueType() == MVT::v16i8 && "Bad shuffle type!");
9380 assert(V1.getSimpleValueType() == MVT::v16i8 && "Bad operand type!");
9381 assert(V2.getSimpleValueType() == MVT::v16i8 && "Bad operand type!");
9382 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
9383 ArrayRef<int> Mask = SVOp->getMask();
9384 assert(Mask.size() == 16 && "Unexpected mask size for v16 shuffle!");
9386 // Try to use shift instructions.
9388 lowerVectorShuffleAsShift(DL, MVT::v16i8, V1, V2, Mask, DAG))
9391 // Try to use byte rotation instructions.
9392 if (SDValue Rotate = lowerVectorShuffleAsByteRotate(
9393 DL, MVT::v16i8, V1, V2, Mask, Subtarget, DAG))
9396 // Try to use a zext lowering.
9397 if (SDValue ZExt = lowerVectorShuffleAsZeroOrAnyExtend(
9398 DL, MVT::v16i8, V1, V2, Mask, Subtarget, DAG))
9401 // See if we can use SSE4A Extraction / Insertion.
9402 if (Subtarget->hasSSE4A())
9403 if (SDValue V = lowerVectorShuffleWithSSE4A(DL, MVT::v16i8, V1, V2, Mask, DAG))
9407 std::count_if(Mask.begin(), Mask.end(), [](int M) { return M >= 16; });
9409 // For single-input shuffles, there are some nicer lowering tricks we can use.
9410 if (NumV2Elements == 0) {
9411 // Check for being able to broadcast a single element.
9412 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v16i8, V1,
9413 Mask, Subtarget, DAG))
9416 // Check whether we can widen this to an i16 shuffle by duplicating bytes.
9417 // Notably, this handles splat and partial-splat shuffles more efficiently.
9418 // However, it only makes sense if the pre-duplication shuffle simplifies
9419 // things significantly. Currently, this means we need to be able to
9420 // express the pre-duplication shuffle as an i16 shuffle.
9422 // FIXME: We should check for other patterns which can be widened into an
9423 // i16 shuffle as well.
9424 auto canWidenViaDuplication = [](ArrayRef<int> Mask) {
9425 for (int i = 0; i < 16; i += 2)
9426 if (Mask[i] != -1 && Mask[i + 1] != -1 && Mask[i] != Mask[i + 1])
9431 auto tryToWidenViaDuplication = [&]() -> SDValue {
9432 if (!canWidenViaDuplication(Mask))
9434 SmallVector<int, 4> LoInputs;
9435 std::copy_if(Mask.begin(), Mask.end(), std::back_inserter(LoInputs),
9436 [](int M) { return M >= 0 && M < 8; });
9437 std::sort(LoInputs.begin(), LoInputs.end());
9438 LoInputs.erase(std::unique(LoInputs.begin(), LoInputs.end()),
9440 SmallVector<int, 4> HiInputs;
9441 std::copy_if(Mask.begin(), Mask.end(), std::back_inserter(HiInputs),
9442 [](int M) { return M >= 8; });
9443 std::sort(HiInputs.begin(), HiInputs.end());
9444 HiInputs.erase(std::unique(HiInputs.begin(), HiInputs.end()),
9447 bool TargetLo = LoInputs.size() >= HiInputs.size();
9448 ArrayRef<int> InPlaceInputs = TargetLo ? LoInputs : HiInputs;
9449 ArrayRef<int> MovingInputs = TargetLo ? HiInputs : LoInputs;
9451 int PreDupI16Shuffle[] = {-1, -1, -1, -1, -1, -1, -1, -1};
9452 SmallDenseMap<int, int, 8> LaneMap;
9453 for (int I : InPlaceInputs) {
9454 PreDupI16Shuffle[I/2] = I/2;
9457 int j = TargetLo ? 0 : 4, je = j + 4;
9458 for (int i = 0, ie = MovingInputs.size(); i < ie; ++i) {
9459 // Check if j is already a shuffle of this input. This happens when
9460 // there are two adjacent bytes after we move the low one.
9461 if (PreDupI16Shuffle[j] != MovingInputs[i] / 2) {
9462 // If we haven't yet mapped the input, search for a slot into which
9464 while (j < je && PreDupI16Shuffle[j] != -1)
9468 // We can't place the inputs into a single half with a simple i16 shuffle, so bail.
9471 // Map this input with the i16 shuffle.
9472 PreDupI16Shuffle[j] = MovingInputs[i] / 2;
9475 // Update the lane map based on the mapping we ended up with.
9476 LaneMap[MovingInputs[i]] = 2 * j + MovingInputs[i] % 2;
9478 V1 = DAG.getBitcast(
9480 DAG.getVectorShuffle(MVT::v8i16, DL, DAG.getBitcast(MVT::v8i16, V1),
9481 DAG.getUNDEF(MVT::v8i16), PreDupI16Shuffle));
9483 // Unpack the bytes to form the i16s that will be shuffled into place.
9484 V1 = DAG.getNode(TargetLo ? X86ISD::UNPCKL : X86ISD::UNPCKH, DL,
9485 MVT::v16i8, V1, V1);
9487 int PostDupI16Shuffle[8] = {-1, -1, -1, -1, -1, -1, -1, -1};
9488 for (int i = 0; i < 16; ++i)
9489 if (Mask[i] != -1) {
9490 int MappedMask = LaneMap[Mask[i]] - (TargetLo ? 0 : 8);
9491 assert(MappedMask < 8 && "Invalid v8 shuffle mask!");
9492 if (PostDupI16Shuffle[i / 2] == -1)
9493 PostDupI16Shuffle[i / 2] = MappedMask;
9495 assert(PostDupI16Shuffle[i / 2] == MappedMask &&
9496 "Conflicting entrties in the original shuffle!");
9498 return DAG.getBitcast(
9500 DAG.getVectorShuffle(MVT::v8i16, DL, DAG.getBitcast(MVT::v8i16, V1),
9501 DAG.getUNDEF(MVT::v8i16), PostDupI16Shuffle));
9503 if (SDValue V = tryToWidenViaDuplication())
9507 if (SDValue Masked =
9508 lowerVectorShuffleAsBitMask(DL, MVT::v16i8, V1, V2, Mask, DAG))
9511 // Use dedicated unpack instructions for masks that match their pattern.
9512 if (isShuffleEquivalent(V1, V2, Mask, {// Low half.
9513 0, 16, 1, 17, 2, 18, 3, 19,
9515 4, 20, 5, 21, 6, 22, 7, 23}))
9516 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v16i8, V1, V2);
9517 if (isShuffleEquivalent(V1, V2, Mask, {// Low half.
9518 8, 24, 9, 25, 10, 26, 11, 27,
9520 12, 28, 13, 29, 14, 30, 15, 31}))
9521 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v16i8, V1, V2);
9523 // Check for SSSE3 which lets us lower all v16i8 shuffles much more directly
9524 // with PSHUFB. It is important to do this before we attempt to generate any
9525 // blends but after all of the single-input lowerings. If the single input
9526 // lowerings can find an instruction sequence that is faster than a PSHUFB, we
9527 // want to preserve that and we can DAG combine any longer sequences into
9528 // a PSHUFB in the end. But once we start blending from multiple inputs,
9529 // the complexity of DAG combining bad patterns back into PSHUFB is too high,
9530 // and there are *very* few patterns that would actually be faster than the
9531 // PSHUFB approach because of its ability to zero lanes.
9533 // FIXME: The only exceptions to the above are blends which are exact
9534 // interleavings with direct instructions supporting them. We currently don't
9535 // handle those well here.
9536 if (Subtarget->hasSSSE3()) {
9537 bool V1InUse = false;
9538 bool V2InUse = false;
9540 SDValue PSHUFB = lowerVectorShuffleAsPSHUFB(DL, MVT::v16i8, V1, V2, Mask,
9541 DAG, V1InUse, V2InUse);
9543 // If both V1 and V2 are in use and we can use a direct blend or an unpack,
9544 // do so. This avoids using them to handle blends-with-zero which is
9545 // important as a single pshufb is significantly faster for that.
9546 if (V1InUse && V2InUse) {
9547 if (Subtarget->hasSSE41())
9548 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v16i8, V1, V2,
9549 Mask, Subtarget, DAG))
9552 // We can use an unpack to do the blending rather than an or in some
9553 // cases. Even though the or may be (very minorly) more efficient, we
9554 // preference this lowering because there are common cases where part of
9555 // the complexity of the shuffles goes away when we do the final blend as
9557 // FIXME: It might be worth trying to detect if the unpack-feeding
9558 // shuffles will both be pshufb, in which case we shouldn't bother with
9560 if (SDValue Unpack = lowerVectorShuffleAsPermuteAndUnpack(
9561 DL, MVT::v16i8, V1, V2, Mask, DAG))
9568 // There are special ways we can lower some single-element blends.
9569 if (NumV2Elements == 1)
9570 if (SDValue V = lowerVectorShuffleAsElementInsertion(DL, MVT::v16i8, V1, V2,
9571 Mask, Subtarget, DAG))
9574 if (SDValue BitBlend =
9575 lowerVectorShuffleAsBitBlend(DL, MVT::v16i8, V1, V2, Mask, DAG))
9578 // Check whether a compaction lowering can be done. This handles shuffles
9579 // which take every Nth element for some even N. See the helper function for
9582 // We special case these as they can be particularly efficiently handled with
9583 // the PACKUSB instruction on x86 and they show up in common patterns of
9584 // rearranging bytes to truncate wide elements.
9585 if (int NumEvenDrops = canLowerByDroppingEvenElements(Mask)) {
9586 // NumEvenDrops is the power of two stride of the elements. Another way of
9587 // thinking about it is that we need to drop the even elements this many
9588 // times to get the original input.
9589 bool IsSingleInput = isSingleInputShuffleMask(Mask);
9591 // First we need to zero all the dropped bytes.
9592 assert(NumEvenDrops <= 3 &&
9593 "No support for dropping even elements more than 3 times.");
9594 // We use the mask type to pick which bytes are preserved based on how many
9595 // elements are dropped.
9596 MVT MaskVTs[] = { MVT::v8i16, MVT::v4i32, MVT::v2i64 };
9597 SDValue ByteClearMask = DAG.getBitcast(
9598 MVT::v16i8, DAG.getConstant(0xFF, DL, MaskVTs[NumEvenDrops - 1]));
9599 V1 = DAG.getNode(ISD::AND, DL, MVT::v16i8, V1, ByteClearMask);
9601 V2 = DAG.getNode(ISD::AND, DL, MVT::v16i8, V2, ByteClearMask);
9603 // Now pack things back together.
9604 V1 = DAG.getBitcast(MVT::v8i16, V1);
9605 V2 = IsSingleInput ? V1 : DAG.getBitcast(MVT::v8i16, V2);
9606 SDValue Result = DAG.getNode(X86ISD::PACKUS, DL, MVT::v16i8, V1, V2);
9607 for (int i = 1; i < NumEvenDrops; ++i) {
9608 Result = DAG.getBitcast(MVT::v8i16, Result);
9609 Result = DAG.getNode(X86ISD::PACKUS, DL, MVT::v16i8, Result, Result);
9615 // Handle multi-input cases by blending single-input shuffles.
9616 if (NumV2Elements > 0)
9617 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v16i8, V1, V2,
9620 // The fallback path for single-input shuffles widens this into two v8i16
9621 // vectors with unpacks, shuffles those, and then pulls them back together
9625 int LoBlendMask[8] = {-1, -1, -1, -1, -1, -1, -1, -1};
9626 int HiBlendMask[8] = {-1, -1, -1, -1, -1, -1, -1, -1};
9627 for (int i = 0; i < 16; ++i)
9629 (i < 8 ? LoBlendMask[i] : HiBlendMask[i % 8]) = Mask[i];
9631 SDValue Zero = getZeroVector(MVT::v8i16, Subtarget, DAG, DL);
9633 SDValue VLoHalf, VHiHalf;
9634 // Check if any of the odd lanes in the v16i8 are used. If not, we can mask
9635 // them out and avoid using UNPCK{L,H} to extract the elements of V as
9637 if (std::none_of(std::begin(LoBlendMask), std::end(LoBlendMask),
9638 [](int M) { return M >= 0 && M % 2 == 1; }) &&
9639 std::none_of(std::begin(HiBlendMask), std::end(HiBlendMask),
9640 [](int M) { return M >= 0 && M % 2 == 1; })) {
9641 // Use a mask to drop the high bytes.
9642 VLoHalf = DAG.getBitcast(MVT::v8i16, V);
9643 VLoHalf = DAG.getNode(ISD::AND, DL, MVT::v8i16, VLoHalf,
9644 DAG.getConstant(0x00FF, DL, MVT::v8i16));
9646 // This will be a single vector shuffle instead of a blend so nuke VHiHalf.
9647 VHiHalf = DAG.getUNDEF(MVT::v8i16);
9649 // Squash the masks to point directly into VLoHalf.
9650 for (int &M : LoBlendMask)
9653 for (int &M : HiBlendMask)
9657 // Otherwise just unpack the low half of V into VLoHalf and the high half into
9658 // VHiHalf so that we can blend them as i16s.
9659 VLoHalf = DAG.getBitcast(
9660 MVT::v8i16, DAG.getNode(X86ISD::UNPCKL, DL, MVT::v16i8, V, Zero));
9661 VHiHalf = DAG.getBitcast(
9662 MVT::v8i16, DAG.getNode(X86ISD::UNPCKH, DL, MVT::v16i8, V, Zero));
9665 SDValue LoV = DAG.getVectorShuffle(MVT::v8i16, DL, VLoHalf, VHiHalf, LoBlendMask);
9666 SDValue HiV = DAG.getVectorShuffle(MVT::v8i16, DL, VLoHalf, VHiHalf, HiBlendMask);
9668 return DAG.getNode(X86ISD::PACKUS, DL, MVT::v16i8, LoV, HiV);
9671 /// \brief Dispatching routine to lower various 128-bit x86 vector shuffles.
9673 /// This routine breaks down the specific type of 128-bit shuffle and
9674 /// dispatches to the lowering routines accordingly.
9675 static SDValue lower128BitVectorShuffle(SDValue Op, SDValue V1, SDValue V2,
9676 MVT VT, const X86Subtarget *Subtarget,
9677 SelectionDAG &DAG) {
9678 switch (VT.SimpleTy) {
9680 return lowerV2I64VectorShuffle(Op, V1, V2, Subtarget, DAG);
9682 return lowerV2F64VectorShuffle(Op, V1, V2, Subtarget, DAG);
9684 return lowerV4I32VectorShuffle(Op, V1, V2, Subtarget, DAG);
9686 return lowerV4F32VectorShuffle(Op, V1, V2, Subtarget, DAG);
9688 return lowerV8I16VectorShuffle(Op, V1, V2, Subtarget, DAG);
9690 return lowerV16I8VectorShuffle(Op, V1, V2, Subtarget, DAG);
9693 llvm_unreachable("Unimplemented!");
9697 /// \brief Helper function to test whether a shuffle mask could be
9698 /// simplified by widening the elements being shuffled.
9700 /// Appends the mask for wider elements in WidenedMask if valid. Otherwise
9701 /// leaves it in an unspecified state.
9703 /// NOTE: This must handle normal vector shuffle masks and *target* vector
9704 /// shuffle masks. The latter have the special property of a '-2' representing
9705 /// a zero-ed lane of a vector.
9706 static bool canWidenShuffleElements(ArrayRef<int> Mask,
9707 SmallVectorImpl<int> &WidenedMask) {
9708 for (int i = 0, Size = Mask.size(); i < Size; i += 2) {
9709 // If both elements are undef, its trivial.
9710 if (Mask[i] == SM_SentinelUndef && Mask[i + 1] == SM_SentinelUndef) {
9711 WidenedMask.push_back(SM_SentinelUndef);
9715 // Check for an undef mask and a mask value properly aligned to fit with
9716 // a pair of values. If we find such a case, use the non-undef mask's value.
9717 if (Mask[i] == SM_SentinelUndef && Mask[i + 1] >= 0 && Mask[i + 1] % 2 == 1) {
9718 WidenedMask.push_back(Mask[i + 1] / 2);
9721 if (Mask[i + 1] == SM_SentinelUndef && Mask[i] >= 0 && Mask[i] % 2 == 0) {
9722 WidenedMask.push_back(Mask[i] / 2);
9726 // When zeroing, we need to spread the zeroing across both lanes to widen.
9727 if (Mask[i] == SM_SentinelZero || Mask[i + 1] == SM_SentinelZero) {
9728 if ((Mask[i] == SM_SentinelZero || Mask[i] == SM_SentinelUndef) &&
9729 (Mask[i + 1] == SM_SentinelZero || Mask[i + 1] == SM_SentinelUndef)) {
9730 WidenedMask.push_back(SM_SentinelZero);
9736 // Finally check if the two mask values are adjacent and aligned with
9738 if (Mask[i] != SM_SentinelUndef && Mask[i] % 2 == 0 && Mask[i] + 1 == Mask[i + 1]) {
9739 WidenedMask.push_back(Mask[i] / 2);
9743 // Otherwise we can't safely widen the elements used in this shuffle.
9746 assert(WidenedMask.size() == Mask.size() / 2 &&
9747 "Incorrect size of mask after widening the elements!");
9752 /// \brief Generic routine to split vector shuffle into half-sized shuffles.
9754 /// This routine just extracts two subvectors, shuffles them independently, and
9755 /// then concatenates them back together. This should work effectively with all
9756 /// AVX vector shuffle types.
9757 static SDValue splitAndLowerVectorShuffle(SDLoc DL, MVT VT, SDValue V1,
9758 SDValue V2, ArrayRef<int> Mask,
9759 SelectionDAG &DAG) {
9760 assert(VT.getSizeInBits() >= 256 &&
9761 "Only for 256-bit or wider vector shuffles!");
9762 assert(V1.getSimpleValueType() == VT && "Bad operand type!");
9763 assert(V2.getSimpleValueType() == VT && "Bad operand type!");
9765 ArrayRef<int> LoMask = Mask.slice(0, Mask.size() / 2);
9766 ArrayRef<int> HiMask = Mask.slice(Mask.size() / 2);
9768 int NumElements = VT.getVectorNumElements();
9769 int SplitNumElements = NumElements / 2;
9770 MVT ScalarVT = VT.getScalarType();
9771 MVT SplitVT = MVT::getVectorVT(ScalarVT, NumElements / 2);
9773 // Rather than splitting build-vectors, just build two narrower build
9774 // vectors. This helps shuffling with splats and zeros.
9775 auto SplitVector = [&](SDValue V) {
9776 while (V.getOpcode() == ISD::BITCAST)
9777 V = V->getOperand(0);
9779 MVT OrigVT = V.getSimpleValueType();
9780 int OrigNumElements = OrigVT.getVectorNumElements();
9781 int OrigSplitNumElements = OrigNumElements / 2;
9782 MVT OrigScalarVT = OrigVT.getScalarType();
9783 MVT OrigSplitVT = MVT::getVectorVT(OrigScalarVT, OrigNumElements / 2);
9787 auto *BV = dyn_cast<BuildVectorSDNode>(V);
9789 LoV = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, OrigSplitVT, V,
9790 DAG.getIntPtrConstant(0, DL));
9791 HiV = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, OrigSplitVT, V,
9792 DAG.getIntPtrConstant(OrigSplitNumElements, DL));
9795 SmallVector<SDValue, 16> LoOps, HiOps;
9796 for (int i = 0; i < OrigSplitNumElements; ++i) {
9797 LoOps.push_back(BV->getOperand(i));
9798 HiOps.push_back(BV->getOperand(i + OrigSplitNumElements));
9800 LoV = DAG.getNode(ISD::BUILD_VECTOR, DL, OrigSplitVT, LoOps);
9801 HiV = DAG.getNode(ISD::BUILD_VECTOR, DL, OrigSplitVT, HiOps);
9803 return std::make_pair(DAG.getBitcast(SplitVT, LoV),
9804 DAG.getBitcast(SplitVT, HiV));
9807 SDValue LoV1, HiV1, LoV2, HiV2;
9808 std::tie(LoV1, HiV1) = SplitVector(V1);
9809 std::tie(LoV2, HiV2) = SplitVector(V2);
9811 // Now create two 4-way blends of these half-width vectors.
9812 auto HalfBlend = [&](ArrayRef<int> HalfMask) {
9813 bool UseLoV1 = false, UseHiV1 = false, UseLoV2 = false, UseHiV2 = false;
9814 SmallVector<int, 32> V1BlendMask, V2BlendMask, BlendMask;
9815 for (int i = 0; i < SplitNumElements; ++i) {
9816 int M = HalfMask[i];
9817 if (M >= NumElements) {
9818 if (M >= NumElements + SplitNumElements)
9822 V2BlendMask.push_back(M - NumElements);
9823 V1BlendMask.push_back(-1);
9824 BlendMask.push_back(SplitNumElements + i);
9825 } else if (M >= 0) {
9826 if (M >= SplitNumElements)
9830 V2BlendMask.push_back(-1);
9831 V1BlendMask.push_back(M);
9832 BlendMask.push_back(i);
9834 V2BlendMask.push_back(-1);
9835 V1BlendMask.push_back(-1);
9836 BlendMask.push_back(-1);
9840 // Because the lowering happens after all combining takes place, we need to
9841 // manually combine these blend masks as much as possible so that we create
9842 // a minimal number of high-level vector shuffle nodes.
9844 // First try just blending the halves of V1 or V2.
9845 if (!UseLoV1 && !UseHiV1 && !UseLoV2 && !UseHiV2)
9846 return DAG.getUNDEF(SplitVT);
9847 if (!UseLoV2 && !UseHiV2)
9848 return DAG.getVectorShuffle(SplitVT, DL, LoV1, HiV1, V1BlendMask);
9849 if (!UseLoV1 && !UseHiV1)
9850 return DAG.getVectorShuffle(SplitVT, DL, LoV2, HiV2, V2BlendMask);
9852 SDValue V1Blend, V2Blend;
9853 if (UseLoV1 && UseHiV1) {
9855 DAG.getVectorShuffle(SplitVT, DL, LoV1, HiV1, V1BlendMask);
9857 // We only use half of V1 so map the usage down into the final blend mask.
9858 V1Blend = UseLoV1 ? LoV1 : HiV1;
9859 for (int i = 0; i < SplitNumElements; ++i)
9860 if (BlendMask[i] >= 0 && BlendMask[i] < SplitNumElements)
9861 BlendMask[i] = V1BlendMask[i] - (UseLoV1 ? 0 : SplitNumElements);
9863 if (UseLoV2 && UseHiV2) {
9865 DAG.getVectorShuffle(SplitVT, DL, LoV2, HiV2, V2BlendMask);
9867 // We only use half of V2 so map the usage down into the final blend mask.
9868 V2Blend = UseLoV2 ? LoV2 : HiV2;
9869 for (int i = 0; i < SplitNumElements; ++i)
9870 if (BlendMask[i] >= SplitNumElements)
9871 BlendMask[i] = V2BlendMask[i] + (UseLoV2 ? SplitNumElements : 0);
9873 return DAG.getVectorShuffle(SplitVT, DL, V1Blend, V2Blend, BlendMask);
9875 SDValue Lo = HalfBlend(LoMask);
9876 SDValue Hi = HalfBlend(HiMask);
9877 return DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, Lo, Hi);
9880 /// \brief Either split a vector in halves or decompose the shuffles and the
9883 /// This is provided as a good fallback for many lowerings of non-single-input
9884 /// shuffles with more than one 128-bit lane. In those cases, we want to select
9885 /// between splitting the shuffle into 128-bit components and stitching those
9886 /// back together vs. extracting the single-input shuffles and blending those
9888 static SDValue lowerVectorShuffleAsSplitOrBlend(SDLoc DL, MVT VT, SDValue V1,
9889 SDValue V2, ArrayRef<int> Mask,
9890 SelectionDAG &DAG) {
9891 assert(!isSingleInputShuffleMask(Mask) && "This routine must not be used to "
9892 "lower single-input shuffles as it "
9893 "could then recurse on itself.");
9894 int Size = Mask.size();
9896 // If this can be modeled as a broadcast of two elements followed by a blend,
9897 // prefer that lowering. This is especially important because broadcasts can
9898 // often fold with memory operands.
9899 auto DoBothBroadcast = [&] {
9900 int V1BroadcastIdx = -1, V2BroadcastIdx = -1;
9903 if (V2BroadcastIdx == -1)
9904 V2BroadcastIdx = M - Size;
9905 else if (M - Size != V2BroadcastIdx)
9907 } else if (M >= 0) {
9908 if (V1BroadcastIdx == -1)
9910 else if (M != V1BroadcastIdx)
9915 if (DoBothBroadcast())
9916 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, VT, V1, V2, Mask,
9919 // If the inputs all stem from a single 128-bit lane of each input, then we
9920 // split them rather than blending because the split will decompose to
9921 // unusually few instructions.
9922 int LaneCount = VT.getSizeInBits() / 128;
9923 int LaneSize = Size / LaneCount;
9924 SmallBitVector LaneInputs[2];
9925 LaneInputs[0].resize(LaneCount, false);
9926 LaneInputs[1].resize(LaneCount, false);
9927 for (int i = 0; i < Size; ++i)
9929 LaneInputs[Mask[i] / Size][(Mask[i] % Size) / LaneSize] = true;
9930 if (LaneInputs[0].count() <= 1 && LaneInputs[1].count() <= 1)
9931 return splitAndLowerVectorShuffle(DL, VT, V1, V2, Mask, DAG);
9933 // Otherwise, just fall back to decomposed shuffles and a blend. This requires
9934 // that the decomposed single-input shuffles don't end up here.
9935 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, VT, V1, V2, Mask, DAG);
9938 /// \brief Lower a vector shuffle crossing multiple 128-bit lanes as
9939 /// a permutation and blend of those lanes.
9941 /// This essentially blends the out-of-lane inputs to each lane into the lane
9942 /// from a permuted copy of the vector. This lowering strategy results in four
9943 /// instructions in the worst case for a single-input cross lane shuffle which
9944 /// is lower than any other fully general cross-lane shuffle strategy I'm aware
9945 /// of. Special cases for each particular shuffle pattern should be handled
9946 /// prior to trying this lowering.
9947 static SDValue lowerVectorShuffleAsLanePermuteAndBlend(SDLoc DL, MVT VT,
9948 SDValue V1, SDValue V2,
9950 SelectionDAG &DAG) {
9951 // FIXME: This should probably be generalized for 512-bit vectors as well.
9952 assert(VT.getSizeInBits() == 256 && "Only for 256-bit vector shuffles!");
9953 int LaneSize = Mask.size() / 2;
9955 // If there are only inputs from one 128-bit lane, splitting will in fact be
9956 // less expensive. The flags track whether the given lane contains an element
9957 // that crosses to another lane.
9958 bool LaneCrossing[2] = {false, false};
9959 for (int i = 0, Size = Mask.size(); i < Size; ++i)
9960 if (Mask[i] >= 0 && (Mask[i] % Size) / LaneSize != i / LaneSize)
9961 LaneCrossing[(Mask[i] % Size) / LaneSize] = true;
9962 if (!LaneCrossing[0] || !LaneCrossing[1])
9963 return splitAndLowerVectorShuffle(DL, VT, V1, V2, Mask, DAG);
9965 if (isSingleInputShuffleMask(Mask)) {
9966 SmallVector<int, 32> FlippedBlendMask;
9967 for (int i = 0, Size = Mask.size(); i < Size; ++i)
9968 FlippedBlendMask.push_back(
9969 Mask[i] < 0 ? -1 : (((Mask[i] % Size) / LaneSize == i / LaneSize)
9971 : Mask[i] % LaneSize +
9972 (i / LaneSize) * LaneSize + Size));
9974 // Flip the vector, and blend the results which should now be in-lane. The
9975 // VPERM2X128 mask uses the low 2 bits for the low source and bits 4 and
9976 // 5 for the high source. The value 3 selects the high half of source 2 and
9977 // the value 2 selects the low half of source 2. We only use source 2 to
9978 // allow folding it into a memory operand.
9979 unsigned PERMMask = 3 | 2 << 4;
9980 SDValue Flipped = DAG.getNode(X86ISD::VPERM2X128, DL, VT, DAG.getUNDEF(VT),
9981 V1, DAG.getConstant(PERMMask, DL, MVT::i8));
9982 return DAG.getVectorShuffle(VT, DL, V1, Flipped, FlippedBlendMask);
9985 // This now reduces to two single-input shuffles of V1 and V2 which at worst
9986 // will be handled by the above logic and a blend of the results, much like
9987 // other patterns in AVX.
9988 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, VT, V1, V2, Mask, DAG);
9991 /// \brief Handle lowering 2-lane 128-bit shuffles.
9992 static SDValue lowerV2X128VectorShuffle(SDLoc DL, MVT VT, SDValue V1,
9993 SDValue V2, ArrayRef<int> Mask,
9994 const X86Subtarget *Subtarget,
9995 SelectionDAG &DAG) {
9996 // TODO: If minimizing size and one of the inputs is a zero vector and the
9997 // the zero vector has only one use, we could use a VPERM2X128 to save the
9998 // instruction bytes needed to explicitly generate the zero vector.
10000 // Blends are faster and handle all the non-lane-crossing cases.
10001 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, VT, V1, V2, Mask,
10005 bool IsV1Zero = ISD::isBuildVectorAllZeros(V1.getNode());
10006 bool IsV2Zero = ISD::isBuildVectorAllZeros(V2.getNode());
10008 // If either input operand is a zero vector, use VPERM2X128 because its mask
10009 // allows us to replace the zero input with an implicit zero.
10010 if (!IsV1Zero && !IsV2Zero) {
10011 // Check for patterns which can be matched with a single insert of a 128-bit
10013 bool OnlyUsesV1 = isShuffleEquivalent(V1, V2, Mask, {0, 1, 0, 1});
10014 if (OnlyUsesV1 || isShuffleEquivalent(V1, V2, Mask, {0, 1, 4, 5})) {
10015 MVT SubVT = MVT::getVectorVT(VT.getVectorElementType(),
10016 VT.getVectorNumElements() / 2);
10017 SDValue LoV = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, SubVT, V1,
10018 DAG.getIntPtrConstant(0, DL));
10019 SDValue HiV = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, SubVT,
10020 OnlyUsesV1 ? V1 : V2,
10021 DAG.getIntPtrConstant(0, DL));
10022 return DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, LoV, HiV);
10026 // Otherwise form a 128-bit permutation. After accounting for undefs,
10027 // convert the 64-bit shuffle mask selection values into 128-bit
10028 // selection bits by dividing the indexes by 2 and shifting into positions
10029 // defined by a vperm2*128 instruction's immediate control byte.
10031 // The immediate permute control byte looks like this:
10032 // [1:0] - select 128 bits from sources for low half of destination
10034 // [3] - zero low half of destination
10035 // [5:4] - select 128 bits from sources for high half of destination
10037 // [7] - zero high half of destination
10039 int MaskLO = Mask[0];
10040 if (MaskLO == SM_SentinelUndef)
10041 MaskLO = Mask[1] == SM_SentinelUndef ? 0 : Mask[1];
10043 int MaskHI = Mask[2];
10044 if (MaskHI == SM_SentinelUndef)
10045 MaskHI = Mask[3] == SM_SentinelUndef ? 0 : Mask[3];
10047 unsigned PermMask = MaskLO / 2 | (MaskHI / 2) << 4;
10049 // If either input is a zero vector, replace it with an undef input.
10050 // Shuffle mask values < 4 are selecting elements of V1.
10051 // Shuffle mask values >= 4 are selecting elements of V2.
10052 // Adjust each half of the permute mask by clearing the half that was
10053 // selecting the zero vector and setting the zero mask bit.
10055 V1 = DAG.getUNDEF(VT);
10057 PermMask = (PermMask & 0xf0) | 0x08;
10059 PermMask = (PermMask & 0x0f) | 0x80;
10062 V2 = DAG.getUNDEF(VT);
10064 PermMask = (PermMask & 0xf0) | 0x08;
10066 PermMask = (PermMask & 0x0f) | 0x80;
10069 return DAG.getNode(X86ISD::VPERM2X128, DL, VT, V1, V2,
10070 DAG.getConstant(PermMask, DL, MVT::i8));
10073 /// \brief Lower a vector shuffle by first fixing the 128-bit lanes and then
10074 /// shuffling each lane.
10076 /// This will only succeed when the result of fixing the 128-bit lanes results
10077 /// in a single-input non-lane-crossing shuffle with a repeating shuffle mask in
10078 /// each 128-bit lanes. This handles many cases where we can quickly blend away
10079 /// the lane crosses early and then use simpler shuffles within each lane.
10081 /// FIXME: It might be worthwhile at some point to support this without
10082 /// requiring the 128-bit lane-relative shuffles to be repeating, but currently
10083 /// in x86 only floating point has interesting non-repeating shuffles, and even
10084 /// those are still *marginally* more expensive.
10085 static SDValue lowerVectorShuffleByMerging128BitLanes(
10086 SDLoc DL, MVT VT, SDValue V1, SDValue V2, ArrayRef<int> Mask,
10087 const X86Subtarget *Subtarget, SelectionDAG &DAG) {
10088 assert(!isSingleInputShuffleMask(Mask) &&
10089 "This is only useful with multiple inputs.");
10091 int Size = Mask.size();
10092 int LaneSize = 128 / VT.getScalarSizeInBits();
10093 int NumLanes = Size / LaneSize;
10094 assert(NumLanes > 1 && "Only handles 256-bit and wider shuffles.");
10096 // See if we can build a hypothetical 128-bit lane-fixing shuffle mask. Also
10097 // check whether the in-128-bit lane shuffles share a repeating pattern.
10098 SmallVector<int, 4> Lanes;
10099 Lanes.resize(NumLanes, -1);
10100 SmallVector<int, 4> InLaneMask;
10101 InLaneMask.resize(LaneSize, -1);
10102 for (int i = 0; i < Size; ++i) {
10106 int j = i / LaneSize;
10108 if (Lanes[j] < 0) {
10109 // First entry we've seen for this lane.
10110 Lanes[j] = Mask[i] / LaneSize;
10111 } else if (Lanes[j] != Mask[i] / LaneSize) {
10112 // This doesn't match the lane selected previously!
10116 // Check that within each lane we have a consistent shuffle mask.
10117 int k = i % LaneSize;
10118 if (InLaneMask[k] < 0) {
10119 InLaneMask[k] = Mask[i] % LaneSize;
10120 } else if (InLaneMask[k] != Mask[i] % LaneSize) {
10121 // This doesn't fit a repeating in-lane mask.
10126 // First shuffle the lanes into place.
10127 MVT LaneVT = MVT::getVectorVT(VT.isFloatingPoint() ? MVT::f64 : MVT::i64,
10128 VT.getSizeInBits() / 64);
10129 SmallVector<int, 8> LaneMask;
10130 LaneMask.resize(NumLanes * 2, -1);
10131 for (int i = 0; i < NumLanes; ++i)
10132 if (Lanes[i] >= 0) {
10133 LaneMask[2 * i + 0] = 2*Lanes[i] + 0;
10134 LaneMask[2 * i + 1] = 2*Lanes[i] + 1;
10137 V1 = DAG.getBitcast(LaneVT, V1);
10138 V2 = DAG.getBitcast(LaneVT, V2);
10139 SDValue LaneShuffle = DAG.getVectorShuffle(LaneVT, DL, V1, V2, LaneMask);
10141 // Cast it back to the type we actually want.
10142 LaneShuffle = DAG.getBitcast(VT, LaneShuffle);
10144 // Now do a simple shuffle that isn't lane crossing.
10145 SmallVector<int, 8> NewMask;
10146 NewMask.resize(Size, -1);
10147 for (int i = 0; i < Size; ++i)
10149 NewMask[i] = (i / LaneSize) * LaneSize + Mask[i] % LaneSize;
10150 assert(!is128BitLaneCrossingShuffleMask(VT, NewMask) &&
10151 "Must not introduce lane crosses at this point!");
10153 return DAG.getVectorShuffle(VT, DL, LaneShuffle, DAG.getUNDEF(VT), NewMask);
10156 /// \brief Test whether the specified input (0 or 1) is in-place blended by the
10159 /// This returns true if the elements from a particular input are already in the
10160 /// slot required by the given mask and require no permutation.
10161 static bool isShuffleMaskInputInPlace(int Input, ArrayRef<int> Mask) {
10162 assert((Input == 0 || Input == 1) && "Only two inputs to shuffles.");
10163 int Size = Mask.size();
10164 for (int i = 0; i < Size; ++i)
10165 if (Mask[i] >= 0 && Mask[i] / Size == Input && Mask[i] % Size != i)
10171 static SDValue lowerVectorShuffleWithSHUFPD(SDLoc DL, MVT VT,
10172 ArrayRef<int> Mask, SDValue V1,
10173 SDValue V2, SelectionDAG &DAG) {
10175 // Mask for V8F64: 0/1, 8/9, 2/3, 10/11, 4/5, ..
10176 // Mask for V4F64; 0/1, 4/5, 2/3, 6/7..
10177 assert(VT.getScalarSizeInBits() == 64 && "Unexpected data type for VSHUFPD");
10178 int NumElts = VT.getVectorNumElements();
10179 bool ShufpdMask = true;
10180 bool CommutableMask = true;
10181 unsigned Immediate = 0;
10182 for (int i = 0; i < NumElts; ++i) {
10185 int Val = (i & 6) + NumElts * (i & 1);
10186 int CommutVal = (i & 0xe) + NumElts * ((i & 1)^1);
10187 if (Mask[i] < Val || Mask[i] > Val + 1)
10188 ShufpdMask = false;
10189 if (Mask[i] < CommutVal || Mask[i] > CommutVal + 1)
10190 CommutableMask = false;
10191 Immediate |= (Mask[i] % 2) << i;
10194 return DAG.getNode(X86ISD::SHUFP, DL, VT, V1, V2,
10195 DAG.getConstant(Immediate, DL, MVT::i8));
10196 if (CommutableMask)
10197 return DAG.getNode(X86ISD::SHUFP, DL, VT, V2, V1,
10198 DAG.getConstant(Immediate, DL, MVT::i8));
10202 /// \brief Handle lowering of 4-lane 64-bit floating point shuffles.
10204 /// Also ends up handling lowering of 4-lane 64-bit integer shuffles when AVX2
10205 /// isn't available.
10206 static SDValue lowerV4F64VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10207 const X86Subtarget *Subtarget,
10208 SelectionDAG &DAG) {
10210 assert(V1.getSimpleValueType() == MVT::v4f64 && "Bad operand type!");
10211 assert(V2.getSimpleValueType() == MVT::v4f64 && "Bad operand type!");
10212 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10213 ArrayRef<int> Mask = SVOp->getMask();
10214 assert(Mask.size() == 4 && "Unexpected mask size for v4 shuffle!");
10216 SmallVector<int, 4> WidenedMask;
10217 if (canWidenShuffleElements(Mask, WidenedMask))
10218 return lowerV2X128VectorShuffle(DL, MVT::v4f64, V1, V2, Mask, Subtarget,
10221 if (isSingleInputShuffleMask(Mask)) {
10222 // Check for being able to broadcast a single element.
10223 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v4f64, V1,
10224 Mask, Subtarget, DAG))
10227 // Use low duplicate instructions for masks that match their pattern.
10228 if (isShuffleEquivalent(V1, V2, Mask, {0, 0, 2, 2}))
10229 return DAG.getNode(X86ISD::MOVDDUP, DL, MVT::v4f64, V1);
10231 if (!is128BitLaneCrossingShuffleMask(MVT::v4f64, Mask)) {
10232 // Non-half-crossing single input shuffles can be lowerid with an
10233 // interleaved permutation.
10234 unsigned VPERMILPMask = (Mask[0] == 1) | ((Mask[1] == 1) << 1) |
10235 ((Mask[2] == 3) << 2) | ((Mask[3] == 3) << 3);
10236 return DAG.getNode(X86ISD::VPERMILPI, DL, MVT::v4f64, V1,
10237 DAG.getConstant(VPERMILPMask, DL, MVT::i8));
10240 // With AVX2 we have direct support for this permutation.
10241 if (Subtarget->hasAVX2())
10242 return DAG.getNode(X86ISD::VPERMI, DL, MVT::v4f64, V1,
10243 getV4X86ShuffleImm8ForMask(Mask, DL, DAG));
10245 // Otherwise, fall back.
10246 return lowerVectorShuffleAsLanePermuteAndBlend(DL, MVT::v4f64, V1, V2, Mask,
10250 // X86 has dedicated unpack instructions that can handle specific blend
10251 // operations: UNPCKH and UNPCKL.
10252 if (isShuffleEquivalent(V1, V2, Mask, {0, 4, 2, 6}))
10253 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v4f64, V1, V2);
10254 if (isShuffleEquivalent(V1, V2, Mask, {1, 5, 3, 7}))
10255 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v4f64, V1, V2);
10256 if (isShuffleEquivalent(V1, V2, Mask, {4, 0, 6, 2}))
10257 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v4f64, V2, V1);
10258 if (isShuffleEquivalent(V1, V2, Mask, {5, 1, 7, 3}))
10259 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v4f64, V2, V1);
10261 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v4f64, V1, V2, Mask,
10265 // Check if the blend happens to exactly fit that of SHUFPD.
10267 lowerVectorShuffleWithSHUFPD(DL, MVT::v4f64, Mask, V1, V2, DAG))
10270 // Try to simplify this by merging 128-bit lanes to enable a lane-based
10271 // shuffle. However, if we have AVX2 and either inputs are already in place,
10272 // we will be able to shuffle even across lanes the other input in a single
10273 // instruction so skip this pattern.
10274 if (!(Subtarget->hasAVX2() && (isShuffleMaskInputInPlace(0, Mask) ||
10275 isShuffleMaskInputInPlace(1, Mask))))
10276 if (SDValue Result = lowerVectorShuffleByMerging128BitLanes(
10277 DL, MVT::v4f64, V1, V2, Mask, Subtarget, DAG))
10280 // If we have AVX2 then we always want to lower with a blend because an v4 we
10281 // can fully permute the elements.
10282 if (Subtarget->hasAVX2())
10283 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v4f64, V1, V2,
10286 // Otherwise fall back on generic lowering.
10287 return lowerVectorShuffleAsSplitOrBlend(DL, MVT::v4f64, V1, V2, Mask, DAG);
10290 /// \brief Handle lowering of 4-lane 64-bit integer shuffles.
10292 /// This routine is only called when we have AVX2 and thus a reasonable
10293 /// instruction set for v4i64 shuffling..
10294 static SDValue lowerV4I64VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10295 const X86Subtarget *Subtarget,
10296 SelectionDAG &DAG) {
10298 assert(V1.getSimpleValueType() == MVT::v4i64 && "Bad operand type!");
10299 assert(V2.getSimpleValueType() == MVT::v4i64 && "Bad operand type!");
10300 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10301 ArrayRef<int> Mask = SVOp->getMask();
10302 assert(Mask.size() == 4 && "Unexpected mask size for v4 shuffle!");
10303 assert(Subtarget->hasAVX2() && "We can only lower v4i64 with AVX2!");
10305 SmallVector<int, 4> WidenedMask;
10306 if (canWidenShuffleElements(Mask, WidenedMask))
10307 return lowerV2X128VectorShuffle(DL, MVT::v4i64, V1, V2, Mask, Subtarget,
10310 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v4i64, V1, V2, Mask,
10314 // Check for being able to broadcast a single element.
10315 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v4i64, V1,
10316 Mask, Subtarget, DAG))
10319 // When the shuffle is mirrored between the 128-bit lanes of the unit, we can
10320 // use lower latency instructions that will operate on both 128-bit lanes.
10321 SmallVector<int, 2> RepeatedMask;
10322 if (is128BitLaneRepeatedShuffleMask(MVT::v4i64, Mask, RepeatedMask)) {
10323 if (isSingleInputShuffleMask(Mask)) {
10324 int PSHUFDMask[] = {-1, -1, -1, -1};
10325 for (int i = 0; i < 2; ++i)
10326 if (RepeatedMask[i] >= 0) {
10327 PSHUFDMask[2 * i] = 2 * RepeatedMask[i];
10328 PSHUFDMask[2 * i + 1] = 2 * RepeatedMask[i] + 1;
10330 return DAG.getBitcast(
10332 DAG.getNode(X86ISD::PSHUFD, DL, MVT::v8i32,
10333 DAG.getBitcast(MVT::v8i32, V1),
10334 getV4X86ShuffleImm8ForMask(PSHUFDMask, DL, DAG)));
10338 // AVX2 provides a direct instruction for permuting a single input across
10340 if (isSingleInputShuffleMask(Mask))
10341 return DAG.getNode(X86ISD::VPERMI, DL, MVT::v4i64, V1,
10342 getV4X86ShuffleImm8ForMask(Mask, DL, DAG));
10344 // Try to use shift instructions.
10345 if (SDValue Shift =
10346 lowerVectorShuffleAsShift(DL, MVT::v4i64, V1, V2, Mask, DAG))
10349 // Use dedicated unpack instructions for masks that match their pattern.
10350 if (isShuffleEquivalent(V1, V2, Mask, {0, 4, 2, 6}))
10351 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v4i64, V1, V2);
10352 if (isShuffleEquivalent(V1, V2, Mask, {1, 5, 3, 7}))
10353 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v4i64, V1, V2);
10354 if (isShuffleEquivalent(V1, V2, Mask, {4, 0, 6, 2}))
10355 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v4i64, V2, V1);
10356 if (isShuffleEquivalent(V1, V2, Mask, {5, 1, 7, 3}))
10357 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v4i64, V2, V1);
10359 // Try to simplify this by merging 128-bit lanes to enable a lane-based
10360 // shuffle. However, if we have AVX2 and either inputs are already in place,
10361 // we will be able to shuffle even across lanes the other input in a single
10362 // instruction so skip this pattern.
10363 if (!(Subtarget->hasAVX2() && (isShuffleMaskInputInPlace(0, Mask) ||
10364 isShuffleMaskInputInPlace(1, Mask))))
10365 if (SDValue Result = lowerVectorShuffleByMerging128BitLanes(
10366 DL, MVT::v4i64, V1, V2, Mask, Subtarget, DAG))
10369 // Otherwise fall back on generic blend lowering.
10370 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v4i64, V1, V2,
10374 /// \brief Handle lowering of 8-lane 32-bit floating point shuffles.
10376 /// Also ends up handling lowering of 8-lane 32-bit integer shuffles when AVX2
10377 /// isn't available.
10378 static SDValue lowerV8F32VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10379 const X86Subtarget *Subtarget,
10380 SelectionDAG &DAG) {
10382 assert(V1.getSimpleValueType() == MVT::v8f32 && "Bad operand type!");
10383 assert(V2.getSimpleValueType() == MVT::v8f32 && "Bad operand type!");
10384 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10385 ArrayRef<int> Mask = SVOp->getMask();
10386 assert(Mask.size() == 8 && "Unexpected mask size for v8 shuffle!");
10388 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v8f32, V1, V2, Mask,
10392 // Check for being able to broadcast a single element.
10393 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v8f32, V1,
10394 Mask, Subtarget, DAG))
10397 // If the shuffle mask is repeated in each 128-bit lane, we have many more
10398 // options to efficiently lower the shuffle.
10399 SmallVector<int, 4> RepeatedMask;
10400 if (is128BitLaneRepeatedShuffleMask(MVT::v8f32, Mask, RepeatedMask)) {
10401 assert(RepeatedMask.size() == 4 &&
10402 "Repeated masks must be half the mask width!");
10404 // Use even/odd duplicate instructions for masks that match their pattern.
10405 if (isShuffleEquivalent(V1, V2, Mask, {0, 0, 2, 2, 4, 4, 6, 6}))
10406 return DAG.getNode(X86ISD::MOVSLDUP, DL, MVT::v8f32, V1);
10407 if (isShuffleEquivalent(V1, V2, Mask, {1, 1, 3, 3, 5, 5, 7, 7}))
10408 return DAG.getNode(X86ISD::MOVSHDUP, DL, MVT::v8f32, V1);
10410 if (isSingleInputShuffleMask(Mask))
10411 return DAG.getNode(X86ISD::VPERMILPI, DL, MVT::v8f32, V1,
10412 getV4X86ShuffleImm8ForMask(RepeatedMask, DL, DAG));
10414 // Use dedicated unpack instructions for masks that match their pattern.
10415 if (isShuffleEquivalent(V1, V2, Mask, {0, 8, 1, 9, 4, 12, 5, 13}))
10416 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v8f32, V1, V2);
10417 if (isShuffleEquivalent(V1, V2, Mask, {2, 10, 3, 11, 6, 14, 7, 15}))
10418 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v8f32, V1, V2);
10419 if (isShuffleEquivalent(V1, V2, Mask, {8, 0, 9, 1, 12, 4, 13, 5}))
10420 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v8f32, V2, V1);
10421 if (isShuffleEquivalent(V1, V2, Mask, {10, 2, 11, 3, 14, 6, 15, 7}))
10422 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v8f32, V2, V1);
10424 // Otherwise, fall back to a SHUFPS sequence. Here it is important that we
10425 // have already handled any direct blends. We also need to squash the
10426 // repeated mask into a simulated v4f32 mask.
10427 for (int i = 0; i < 4; ++i)
10428 if (RepeatedMask[i] >= 8)
10429 RepeatedMask[i] -= 4;
10430 return lowerVectorShuffleWithSHUFPS(DL, MVT::v8f32, RepeatedMask, V1, V2, DAG);
10433 // If we have a single input shuffle with different shuffle patterns in the
10434 // two 128-bit lanes use the variable mask to VPERMILPS.
10435 if (isSingleInputShuffleMask(Mask)) {
10436 SDValue VPermMask[8];
10437 for (int i = 0; i < 8; ++i)
10438 VPermMask[i] = Mask[i] < 0 ? DAG.getUNDEF(MVT::i32)
10439 : DAG.getConstant(Mask[i], DL, MVT::i32);
10440 if (!is128BitLaneCrossingShuffleMask(MVT::v8f32, Mask))
10441 return DAG.getNode(
10442 X86ISD::VPERMILPV, DL, MVT::v8f32, V1,
10443 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v8i32, VPermMask));
10445 if (Subtarget->hasAVX2())
10446 return DAG.getNode(
10447 X86ISD::VPERMV, DL, MVT::v8f32,
10448 DAG.getBitcast(MVT::v8f32, DAG.getNode(ISD::BUILD_VECTOR, DL,
10449 MVT::v8i32, VPermMask)),
10452 // Otherwise, fall back.
10453 return lowerVectorShuffleAsLanePermuteAndBlend(DL, MVT::v8f32, V1, V2, Mask,
10457 // Try to simplify this by merging 128-bit lanes to enable a lane-based
10459 if (SDValue Result = lowerVectorShuffleByMerging128BitLanes(
10460 DL, MVT::v8f32, V1, V2, Mask, Subtarget, DAG))
10463 // If we have AVX2 then we always want to lower with a blend because at v8 we
10464 // can fully permute the elements.
10465 if (Subtarget->hasAVX2())
10466 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v8f32, V1, V2,
10469 // Otherwise fall back on generic lowering.
10470 return lowerVectorShuffleAsSplitOrBlend(DL, MVT::v8f32, V1, V2, Mask, DAG);
10473 /// \brief Handle lowering of 8-lane 32-bit integer shuffles.
10475 /// This routine is only called when we have AVX2 and thus a reasonable
10476 /// instruction set for v8i32 shuffling..
10477 static SDValue lowerV8I32VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10478 const X86Subtarget *Subtarget,
10479 SelectionDAG &DAG) {
10481 assert(V1.getSimpleValueType() == MVT::v8i32 && "Bad operand type!");
10482 assert(V2.getSimpleValueType() == MVT::v8i32 && "Bad operand type!");
10483 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10484 ArrayRef<int> Mask = SVOp->getMask();
10485 assert(Mask.size() == 8 && "Unexpected mask size for v8 shuffle!");
10486 assert(Subtarget->hasAVX2() && "We can only lower v8i32 with AVX2!");
10488 // Whenever we can lower this as a zext, that instruction is strictly faster
10489 // than any alternative. It also allows us to fold memory operands into the
10490 // shuffle in many cases.
10491 if (SDValue ZExt = lowerVectorShuffleAsZeroOrAnyExtend(DL, MVT::v8i32, V1, V2,
10492 Mask, Subtarget, DAG))
10495 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v8i32, V1, V2, Mask,
10499 // Check for being able to broadcast a single element.
10500 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v8i32, V1,
10501 Mask, Subtarget, DAG))
10504 // If the shuffle mask is repeated in each 128-bit lane we can use more
10505 // efficient instructions that mirror the shuffles across the two 128-bit
10507 SmallVector<int, 4> RepeatedMask;
10508 if (is128BitLaneRepeatedShuffleMask(MVT::v8i32, Mask, RepeatedMask)) {
10509 assert(RepeatedMask.size() == 4 && "Unexpected repeated mask size!");
10510 if (isSingleInputShuffleMask(Mask))
10511 return DAG.getNode(X86ISD::PSHUFD, DL, MVT::v8i32, V1,
10512 getV4X86ShuffleImm8ForMask(RepeatedMask, DL, DAG));
10514 // Use dedicated unpack instructions for masks that match their pattern.
10515 if (isShuffleEquivalent(V1, V2, Mask, {0, 8, 1, 9, 4, 12, 5, 13}))
10516 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v8i32, V1, V2);
10517 if (isShuffleEquivalent(V1, V2, Mask, {2, 10, 3, 11, 6, 14, 7, 15}))
10518 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v8i32, V1, V2);
10519 if (isShuffleEquivalent(V1, V2, Mask, {8, 0, 9, 1, 12, 4, 13, 5}))
10520 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v8i32, V2, V1);
10521 if (isShuffleEquivalent(V1, V2, Mask, {10, 2, 11, 3, 14, 6, 15, 7}))
10522 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v8i32, V2, V1);
10525 // Try to use shift instructions.
10526 if (SDValue Shift =
10527 lowerVectorShuffleAsShift(DL, MVT::v8i32, V1, V2, Mask, DAG))
10530 if (SDValue Rotate = lowerVectorShuffleAsByteRotate(
10531 DL, MVT::v8i32, V1, V2, Mask, Subtarget, DAG))
10534 // If the shuffle patterns aren't repeated but it is a single input, directly
10535 // generate a cross-lane VPERMD instruction.
10536 if (isSingleInputShuffleMask(Mask)) {
10537 SDValue VPermMask[8];
10538 for (int i = 0; i < 8; ++i)
10539 VPermMask[i] = Mask[i] < 0 ? DAG.getUNDEF(MVT::i32)
10540 : DAG.getConstant(Mask[i], DL, MVT::i32);
10541 return DAG.getNode(
10542 X86ISD::VPERMV, DL, MVT::v8i32,
10543 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v8i32, VPermMask), V1);
10546 // Try to simplify this by merging 128-bit lanes to enable a lane-based
10548 if (SDValue Result = lowerVectorShuffleByMerging128BitLanes(
10549 DL, MVT::v8i32, V1, V2, Mask, Subtarget, DAG))
10552 // Otherwise fall back on generic blend lowering.
10553 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v8i32, V1, V2,
10557 /// \brief Handle lowering of 16-lane 16-bit integer shuffles.
10559 /// This routine is only called when we have AVX2 and thus a reasonable
10560 /// instruction set for v16i16 shuffling..
10561 static SDValue lowerV16I16VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10562 const X86Subtarget *Subtarget,
10563 SelectionDAG &DAG) {
10565 assert(V1.getSimpleValueType() == MVT::v16i16 && "Bad operand type!");
10566 assert(V2.getSimpleValueType() == MVT::v16i16 && "Bad operand type!");
10567 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10568 ArrayRef<int> Mask = SVOp->getMask();
10569 assert(Mask.size() == 16 && "Unexpected mask size for v16 shuffle!");
10570 assert(Subtarget->hasAVX2() && "We can only lower v16i16 with AVX2!");
10572 // Whenever we can lower this as a zext, that instruction is strictly faster
10573 // than any alternative. It also allows us to fold memory operands into the
10574 // shuffle in many cases.
10575 if (SDValue ZExt = lowerVectorShuffleAsZeroOrAnyExtend(DL, MVT::v16i16, V1, V2,
10576 Mask, Subtarget, DAG))
10579 // Check for being able to broadcast a single element.
10580 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v16i16, V1,
10581 Mask, Subtarget, DAG))
10584 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v16i16, V1, V2, Mask,
10588 // Use dedicated unpack instructions for masks that match their pattern.
10589 if (isShuffleEquivalent(V1, V2, Mask,
10590 {// First 128-bit lane:
10591 0, 16, 1, 17, 2, 18, 3, 19,
10592 // Second 128-bit lane:
10593 8, 24, 9, 25, 10, 26, 11, 27}))
10594 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v16i16, V1, V2);
10595 if (isShuffleEquivalent(V1, V2, Mask,
10596 {// First 128-bit lane:
10597 4, 20, 5, 21, 6, 22, 7, 23,
10598 // Second 128-bit lane:
10599 12, 28, 13, 29, 14, 30, 15, 31}))
10600 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v16i16, V1, V2);
10602 // Try to use shift instructions.
10603 if (SDValue Shift =
10604 lowerVectorShuffleAsShift(DL, MVT::v16i16, V1, V2, Mask, DAG))
10607 // Try to use byte rotation instructions.
10608 if (SDValue Rotate = lowerVectorShuffleAsByteRotate(
10609 DL, MVT::v16i16, V1, V2, Mask, Subtarget, DAG))
10612 if (isSingleInputShuffleMask(Mask)) {
10613 // There are no generalized cross-lane shuffle operations available on i16
10615 if (is128BitLaneCrossingShuffleMask(MVT::v16i16, Mask))
10616 return lowerVectorShuffleAsLanePermuteAndBlend(DL, MVT::v16i16, V1, V2,
10619 SmallVector<int, 8> RepeatedMask;
10620 if (is128BitLaneRepeatedShuffleMask(MVT::v16i16, Mask, RepeatedMask)) {
10621 // As this is a single-input shuffle, the repeated mask should be
10622 // a strictly valid v8i16 mask that we can pass through to the v8i16
10623 // lowering to handle even the v16 case.
10624 return lowerV8I16GeneralSingleInputVectorShuffle(
10625 DL, MVT::v16i16, V1, RepeatedMask, Subtarget, DAG);
10628 SDValue PSHUFBMask[32];
10629 for (int i = 0; i < 16; ++i) {
10630 if (Mask[i] == -1) {
10631 PSHUFBMask[2 * i] = PSHUFBMask[2 * i + 1] = DAG.getUNDEF(MVT::i8);
10635 int M = i < 8 ? Mask[i] : Mask[i] - 8;
10636 assert(M >= 0 && M < 8 && "Invalid single-input mask!");
10637 PSHUFBMask[2 * i] = DAG.getConstant(2 * M, DL, MVT::i8);
10638 PSHUFBMask[2 * i + 1] = DAG.getConstant(2 * M + 1, DL, MVT::i8);
10640 return DAG.getBitcast(MVT::v16i16,
10641 DAG.getNode(X86ISD::PSHUFB, DL, MVT::v32i8,
10642 DAG.getBitcast(MVT::v32i8, V1),
10643 DAG.getNode(ISD::BUILD_VECTOR, DL,
10644 MVT::v32i8, PSHUFBMask)));
10647 // Try to simplify this by merging 128-bit lanes to enable a lane-based
10649 if (SDValue Result = lowerVectorShuffleByMerging128BitLanes(
10650 DL, MVT::v16i16, V1, V2, Mask, Subtarget, DAG))
10653 // Otherwise fall back on generic lowering.
10654 return lowerVectorShuffleAsSplitOrBlend(DL, MVT::v16i16, V1, V2, Mask, DAG);
10657 /// \brief Handle lowering of 32-lane 8-bit integer shuffles.
10659 /// This routine is only called when we have AVX2 and thus a reasonable
10660 /// instruction set for v32i8 shuffling..
10661 static SDValue lowerV32I8VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10662 const X86Subtarget *Subtarget,
10663 SelectionDAG &DAG) {
10665 assert(V1.getSimpleValueType() == MVT::v32i8 && "Bad operand type!");
10666 assert(V2.getSimpleValueType() == MVT::v32i8 && "Bad operand type!");
10667 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10668 ArrayRef<int> Mask = SVOp->getMask();
10669 assert(Mask.size() == 32 && "Unexpected mask size for v32 shuffle!");
10670 assert(Subtarget->hasAVX2() && "We can only lower v32i8 with AVX2!");
10672 // Whenever we can lower this as a zext, that instruction is strictly faster
10673 // than any alternative. It also allows us to fold memory operands into the
10674 // shuffle in many cases.
10675 if (SDValue ZExt = lowerVectorShuffleAsZeroOrAnyExtend(DL, MVT::v32i8, V1, V2,
10676 Mask, Subtarget, DAG))
10679 // Check for being able to broadcast a single element.
10680 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v32i8, V1,
10681 Mask, Subtarget, DAG))
10684 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v32i8, V1, V2, Mask,
10688 // Use dedicated unpack instructions for masks that match their pattern.
10689 // Note that these are repeated 128-bit lane unpacks, not unpacks across all
10691 if (isShuffleEquivalent(
10693 {// First 128-bit lane:
10694 0, 32, 1, 33, 2, 34, 3, 35, 4, 36, 5, 37, 6, 38, 7, 39,
10695 // Second 128-bit lane:
10696 16, 48, 17, 49, 18, 50, 19, 51, 20, 52, 21, 53, 22, 54, 23, 55}))
10697 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v32i8, V1, V2);
10698 if (isShuffleEquivalent(
10700 {// First 128-bit lane:
10701 8, 40, 9, 41, 10, 42, 11, 43, 12, 44, 13, 45, 14, 46, 15, 47,
10702 // Second 128-bit lane:
10703 24, 56, 25, 57, 26, 58, 27, 59, 28, 60, 29, 61, 30, 62, 31, 63}))
10704 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v32i8, V1, V2);
10706 // Try to use shift instructions.
10707 if (SDValue Shift =
10708 lowerVectorShuffleAsShift(DL, MVT::v32i8, V1, V2, Mask, DAG))
10711 // Try to use byte rotation instructions.
10712 if (SDValue Rotate = lowerVectorShuffleAsByteRotate(
10713 DL, MVT::v32i8, V1, V2, Mask, Subtarget, DAG))
10716 if (isSingleInputShuffleMask(Mask)) {
10717 // There are no generalized cross-lane shuffle operations available on i8
10719 if (is128BitLaneCrossingShuffleMask(MVT::v32i8, Mask))
10720 return lowerVectorShuffleAsLanePermuteAndBlend(DL, MVT::v32i8, V1, V2,
10723 SDValue PSHUFBMask[32];
10724 for (int i = 0; i < 32; ++i)
10727 ? DAG.getUNDEF(MVT::i8)
10728 : DAG.getConstant(Mask[i] < 16 ? Mask[i] : Mask[i] - 16, DL,
10731 return DAG.getNode(
10732 X86ISD::PSHUFB, DL, MVT::v32i8, V1,
10733 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v32i8, PSHUFBMask));
10736 // Try to simplify this by merging 128-bit lanes to enable a lane-based
10738 if (SDValue Result = lowerVectorShuffleByMerging128BitLanes(
10739 DL, MVT::v32i8, V1, V2, Mask, Subtarget, DAG))
10742 // Otherwise fall back on generic lowering.
10743 return lowerVectorShuffleAsSplitOrBlend(DL, MVT::v32i8, V1, V2, Mask, DAG);
10746 /// \brief High-level routine to lower various 256-bit x86 vector shuffles.
10748 /// This routine either breaks down the specific type of a 256-bit x86 vector
10749 /// shuffle or splits it into two 128-bit shuffles and fuses the results back
10750 /// together based on the available instructions.
10751 static SDValue lower256BitVectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10752 MVT VT, const X86Subtarget *Subtarget,
10753 SelectionDAG &DAG) {
10755 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10756 ArrayRef<int> Mask = SVOp->getMask();
10758 // If we have a single input to the zero element, insert that into V1 if we
10759 // can do so cheaply.
10760 int NumElts = VT.getVectorNumElements();
10761 int NumV2Elements = std::count_if(Mask.begin(), Mask.end(), [NumElts](int M) {
10762 return M >= NumElts;
10765 if (NumV2Elements == 1 && Mask[0] >= NumElts)
10766 if (SDValue Insertion = lowerVectorShuffleAsElementInsertion(
10767 DL, VT, V1, V2, Mask, Subtarget, DAG))
10770 // There is a really nice hard cut-over between AVX1 and AVX2 that means we
10771 // can check for those subtargets here and avoid much of the subtarget
10772 // querying in the per-vector-type lowering routines. With AVX1 we have
10773 // essentially *zero* ability to manipulate a 256-bit vector with integer
10774 // types. Since we'll use floating point types there eventually, just
10775 // immediately cast everything to a float and operate entirely in that domain.
10776 if (VT.isInteger() && !Subtarget->hasAVX2()) {
10777 int ElementBits = VT.getScalarSizeInBits();
10778 if (ElementBits < 32)
10779 // No floating point type available, decompose into 128-bit vectors.
10780 return splitAndLowerVectorShuffle(DL, VT, V1, V2, Mask, DAG);
10782 MVT FpVT = MVT::getVectorVT(MVT::getFloatingPointVT(ElementBits),
10783 VT.getVectorNumElements());
10784 V1 = DAG.getBitcast(FpVT, V1);
10785 V2 = DAG.getBitcast(FpVT, V2);
10786 return DAG.getBitcast(VT, DAG.getVectorShuffle(FpVT, DL, V1, V2, Mask));
10789 switch (VT.SimpleTy) {
10791 return lowerV4F64VectorShuffle(Op, V1, V2, Subtarget, DAG);
10793 return lowerV4I64VectorShuffle(Op, V1, V2, Subtarget, DAG);
10795 return lowerV8F32VectorShuffle(Op, V1, V2, Subtarget, DAG);
10797 return lowerV8I32VectorShuffle(Op, V1, V2, Subtarget, DAG);
10799 return lowerV16I16VectorShuffle(Op, V1, V2, Subtarget, DAG);
10801 return lowerV32I8VectorShuffle(Op, V1, V2, Subtarget, DAG);
10804 llvm_unreachable("Not a valid 256-bit x86 vector type!");
10808 /// \brief Try to lower a vector shuffle as a 128-bit shuffles.
10809 static SDValue lowerV4X128VectorShuffle(SDLoc DL, MVT VT,
10810 ArrayRef<int> Mask,
10811 SDValue V1, SDValue V2,
10812 SelectionDAG &DAG) {
10813 assert(VT.getScalarSizeInBits() == 64 &&
10814 "Unexpected element type size for 128bit shuffle.");
10816 // To handle 256 bit vector requires VLX and most probably
10817 // function lowerV2X128VectorShuffle() is better solution.
10818 assert(VT.getSizeInBits() == 512 &&
10819 "Unexpected vector size for 128bit shuffle.");
10821 SmallVector<int, 4> WidenedMask;
10822 if (!canWidenShuffleElements(Mask, WidenedMask))
10825 // Form a 128-bit permutation.
10826 // Convert the 64-bit shuffle mask selection values into 128-bit selection
10827 // bits defined by a vshuf64x2 instruction's immediate control byte.
10828 unsigned PermMask = 0, Imm = 0;
10829 unsigned ControlBitsNum = WidenedMask.size() / 2;
10831 for (int i = 0, Size = WidenedMask.size(); i < Size; ++i) {
10832 if (WidenedMask[i] == SM_SentinelZero)
10835 // Use first element in place of undef mask.
10836 Imm = (WidenedMask[i] == SM_SentinelUndef) ? 0 : WidenedMask[i];
10837 PermMask |= (Imm % WidenedMask.size()) << (i * ControlBitsNum);
10840 return DAG.getNode(X86ISD::SHUF128, DL, VT, V1, V2,
10841 DAG.getConstant(PermMask, DL, MVT::i8));
10844 static SDValue lowerVectorShuffleWithPERMV(SDLoc DL, MVT VT,
10845 ArrayRef<int> Mask, SDValue V1,
10846 SDValue V2, SelectionDAG &DAG) {
10848 assert(VT.getScalarSizeInBits() >= 16 && "Unexpected data type for PERMV");
10850 MVT MaskEltVT = MVT::getIntegerVT(VT.getScalarSizeInBits());
10851 MVT MaskVecVT = MVT::getVectorVT(MaskEltVT, VT.getVectorNumElements());
10853 SDValue MaskNode = getConstVector(Mask, MaskVecVT, DAG, DL, true);
10854 if (isSingleInputShuffleMask(Mask))
10855 return DAG.getNode(X86ISD::VPERMV, DL, VT, MaskNode, V1);
10857 return DAG.getNode(X86ISD::VPERMV3, DL, VT, V1, MaskNode, V2);
10860 /// \brief Handle lowering of 8-lane 64-bit floating point shuffles.
10861 static SDValue lowerV8F64VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10862 const X86Subtarget *Subtarget,
10863 SelectionDAG &DAG) {
10865 assert(V1.getSimpleValueType() == MVT::v8f64 && "Bad operand type!");
10866 assert(V2.getSimpleValueType() == MVT::v8f64 && "Bad operand type!");
10867 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10868 ArrayRef<int> Mask = SVOp->getMask();
10869 assert(Mask.size() == 8 && "Unexpected mask size for v8 shuffle!");
10871 if (SDValue Shuf128 =
10872 lowerV4X128VectorShuffle(DL, MVT::v8f64, Mask, V1, V2, DAG))
10875 if (SDValue Unpck =
10876 lowerVectorShuffleWithUNPCK(DL, MVT::v8f64, Mask, V1, V2, DAG))
10879 return lowerVectorShuffleWithPERMV(DL, MVT::v8f64, Mask, V1, V2, DAG);
10882 /// \brief Handle lowering of 16-lane 32-bit floating point shuffles.
10883 static SDValue lowerV16F32VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10884 const X86Subtarget *Subtarget,
10885 SelectionDAG &DAG) {
10887 assert(V1.getSimpleValueType() == MVT::v16f32 && "Bad operand type!");
10888 assert(V2.getSimpleValueType() == MVT::v16f32 && "Bad operand type!");
10889 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10890 ArrayRef<int> Mask = SVOp->getMask();
10891 assert(Mask.size() == 16 && "Unexpected mask size for v16 shuffle!");
10893 if (SDValue Unpck =
10894 lowerVectorShuffleWithUNPCK(DL, MVT::v16f32, Mask, V1, V2, DAG))
10897 return lowerVectorShuffleWithPERMV(DL, MVT::v16f32, Mask, V1, V2, DAG);
10900 /// \brief Handle lowering of 8-lane 64-bit integer shuffles.
10901 static SDValue lowerV8I64VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10902 const X86Subtarget *Subtarget,
10903 SelectionDAG &DAG) {
10905 assert(V1.getSimpleValueType() == MVT::v8i64 && "Bad operand type!");
10906 assert(V2.getSimpleValueType() == MVT::v8i64 && "Bad operand type!");
10907 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10908 ArrayRef<int> Mask = SVOp->getMask();
10909 assert(Mask.size() == 8 && "Unexpected mask size for v8 shuffle!");
10911 if (SDValue Shuf128 =
10912 lowerV4X128VectorShuffle(DL, MVT::v8i64, Mask, V1, V2, DAG))
10915 if (SDValue Unpck =
10916 lowerVectorShuffleWithUNPCK(DL, MVT::v8i64, Mask, V1, V2, DAG))
10919 return lowerVectorShuffleWithPERMV(DL, MVT::v8i64, Mask, V1, V2, DAG);
10922 /// \brief Handle lowering of 16-lane 32-bit integer shuffles.
10923 static SDValue lowerV16I32VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10924 const X86Subtarget *Subtarget,
10925 SelectionDAG &DAG) {
10927 assert(V1.getSimpleValueType() == MVT::v16i32 && "Bad operand type!");
10928 assert(V2.getSimpleValueType() == MVT::v16i32 && "Bad operand type!");
10929 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10930 ArrayRef<int> Mask = SVOp->getMask();
10931 assert(Mask.size() == 16 && "Unexpected mask size for v16 shuffle!");
10933 if (SDValue Unpck =
10934 lowerVectorShuffleWithUNPCK(DL, MVT::v16i32, Mask, V1, V2, DAG))
10937 return lowerVectorShuffleWithPERMV(DL, MVT::v16i32, Mask, V1, V2, DAG);
10940 /// \brief Handle lowering of 32-lane 16-bit integer shuffles.
10941 static SDValue lowerV32I16VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10942 const X86Subtarget *Subtarget,
10943 SelectionDAG &DAG) {
10945 assert(V1.getSimpleValueType() == MVT::v32i16 && "Bad operand type!");
10946 assert(V2.getSimpleValueType() == MVT::v32i16 && "Bad operand type!");
10947 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10948 ArrayRef<int> Mask = SVOp->getMask();
10949 assert(Mask.size() == 32 && "Unexpected mask size for v32 shuffle!");
10950 assert(Subtarget->hasBWI() && "We can only lower v32i16 with AVX-512-BWI!");
10952 return lowerVectorShuffleWithPERMV(DL, MVT::v32i16, Mask, V1, V2, DAG);
10955 /// \brief Handle lowering of 64-lane 8-bit integer shuffles.
10956 static SDValue lowerV64I8VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10957 const X86Subtarget *Subtarget,
10958 SelectionDAG &DAG) {
10960 assert(V1.getSimpleValueType() == MVT::v64i8 && "Bad operand type!");
10961 assert(V2.getSimpleValueType() == MVT::v64i8 && "Bad operand type!");
10962 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10963 ArrayRef<int> Mask = SVOp->getMask();
10964 assert(Mask.size() == 64 && "Unexpected mask size for v64 shuffle!");
10965 assert(Subtarget->hasBWI() && "We can only lower v64i8 with AVX-512-BWI!");
10967 // FIXME: Implement direct support for this type!
10968 return splitAndLowerVectorShuffle(DL, MVT::v64i8, V1, V2, Mask, DAG);
10971 /// \brief High-level routine to lower various 512-bit x86 vector shuffles.
10973 /// This routine either breaks down the specific type of a 512-bit x86 vector
10974 /// shuffle or splits it into two 256-bit shuffles and fuses the results back
10975 /// together based on the available instructions.
10976 static SDValue lower512BitVectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10977 MVT VT, const X86Subtarget *Subtarget,
10978 SelectionDAG &DAG) {
10980 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10981 ArrayRef<int> Mask = SVOp->getMask();
10982 assert(Subtarget->hasAVX512() &&
10983 "Cannot lower 512-bit vectors w/ basic ISA!");
10985 // Check for being able to broadcast a single element.
10986 if (SDValue Broadcast =
10987 lowerVectorShuffleAsBroadcast(DL, VT, V1, Mask, Subtarget, DAG))
10990 // Dispatch to each element type for lowering. If we don't have supprot for
10991 // specific element type shuffles at 512 bits, immediately split them and
10992 // lower them. Each lowering routine of a given type is allowed to assume that
10993 // the requisite ISA extensions for that element type are available.
10994 switch (VT.SimpleTy) {
10996 return lowerV8F64VectorShuffle(Op, V1, V2, Subtarget, DAG);
10998 return lowerV16F32VectorShuffle(Op, V1, V2, Subtarget, DAG);
11000 return lowerV8I64VectorShuffle(Op, V1, V2, Subtarget, DAG);
11002 return lowerV16I32VectorShuffle(Op, V1, V2, Subtarget, DAG);
11004 if (Subtarget->hasBWI())
11005 return lowerV32I16VectorShuffle(Op, V1, V2, Subtarget, DAG);
11008 if (Subtarget->hasBWI())
11009 return lowerV64I8VectorShuffle(Op, V1, V2, Subtarget, DAG);
11013 llvm_unreachable("Not a valid 512-bit x86 vector type!");
11016 // Otherwise fall back on splitting.
11017 return splitAndLowerVectorShuffle(DL, VT, V1, V2, Mask, DAG);
11020 // Lower vXi1 vector shuffles.
11021 // There is no a dedicated instruction on AVX-512 that shuffles the masks.
11022 // The only way to shuffle bits is to sign-extend the mask vector to SIMD
11023 // vector, shuffle and then truncate it back.
11024 static SDValue lower1BitVectorShuffle(SDValue Op, SDValue V1, SDValue V2,
11025 MVT VT, const X86Subtarget *Subtarget,
11026 SelectionDAG &DAG) {
11028 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
11029 ArrayRef<int> Mask = SVOp->getMask();
11030 assert(Subtarget->hasAVX512() &&
11031 "Cannot lower 512-bit vectors w/o basic ISA!");
11033 switch (VT.SimpleTy) {
11035 assert(false && "Expected a vector of i1 elements");
11038 ExtVT = MVT::v2i64;
11041 ExtVT = MVT::v4i32;
11044 ExtVT = MVT::v8i64; // Take 512-bit type, more shuffles on KNL
11047 ExtVT = MVT::v16i32;
11050 ExtVT = MVT::v32i16;
11053 ExtVT = MVT::v64i8;
11057 if (ISD::isBuildVectorAllZeros(V1.getNode()))
11058 V1 = getZeroVector(ExtVT, Subtarget, DAG, DL);
11059 else if (ISD::isBuildVectorAllOnes(V1.getNode()))
11060 V1 = getOnesVector(ExtVT, Subtarget, DAG, DL);
11062 V1 = DAG.getNode(ISD::SIGN_EXTEND, DL, ExtVT, V1);
11065 V2 = DAG.getUNDEF(ExtVT);
11066 else if (ISD::isBuildVectorAllZeros(V2.getNode()))
11067 V2 = getZeroVector(ExtVT, Subtarget, DAG, DL);
11068 else if (ISD::isBuildVectorAllOnes(V2.getNode()))
11069 V2 = getOnesVector(ExtVT, Subtarget, DAG, DL);
11071 V2 = DAG.getNode(ISD::SIGN_EXTEND, DL, ExtVT, V2);
11072 return DAG.getNode(ISD::TRUNCATE, DL, VT,
11073 DAG.getVectorShuffle(ExtVT, DL, V1, V2, Mask));
11075 /// \brief Top-level lowering for x86 vector shuffles.
11077 /// This handles decomposition, canonicalization, and lowering of all x86
11078 /// vector shuffles. Most of the specific lowering strategies are encapsulated
11079 /// above in helper routines. The canonicalization attempts to widen shuffles
11080 /// to involve fewer lanes of wider elements, consolidate symmetric patterns
11081 /// s.t. only one of the two inputs needs to be tested, etc.
11082 static SDValue lowerVectorShuffle(SDValue Op, const X86Subtarget *Subtarget,
11083 SelectionDAG &DAG) {
11084 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
11085 ArrayRef<int> Mask = SVOp->getMask();
11086 SDValue V1 = Op.getOperand(0);
11087 SDValue V2 = Op.getOperand(1);
11088 MVT VT = Op.getSimpleValueType();
11089 int NumElements = VT.getVectorNumElements();
11091 bool Is1BitVector = (VT.getScalarType() == MVT::i1);
11093 assert((VT.getSizeInBits() != 64 || Is1BitVector) &&
11094 "Can't lower MMX shuffles");
11096 bool V1IsUndef = V1.getOpcode() == ISD::UNDEF;
11097 bool V2IsUndef = V2.getOpcode() == ISD::UNDEF;
11098 if (V1IsUndef && V2IsUndef)
11099 return DAG.getUNDEF(VT);
11101 // When we create a shuffle node we put the UNDEF node to second operand,
11102 // but in some cases the first operand may be transformed to UNDEF.
11103 // In this case we should just commute the node.
11105 return DAG.getCommutedVectorShuffle(*SVOp);
11107 // Check for non-undef masks pointing at an undef vector and make the masks
11108 // undef as well. This makes it easier to match the shuffle based solely on
11112 if (M >= NumElements) {
11113 SmallVector<int, 8> NewMask(Mask.begin(), Mask.end());
11114 for (int &M : NewMask)
11115 if (M >= NumElements)
11117 return DAG.getVectorShuffle(VT, dl, V1, V2, NewMask);
11120 // We actually see shuffles that are entirely re-arrangements of a set of
11121 // zero inputs. This mostly happens while decomposing complex shuffles into
11122 // simple ones. Directly lower these as a buildvector of zeros.
11123 SmallBitVector Zeroable = computeZeroableShuffleElements(Mask, V1, V2);
11124 if (Zeroable.all())
11125 return getZeroVector(VT, Subtarget, DAG, dl);
11127 // Try to collapse shuffles into using a vector type with fewer elements but
11128 // wider element types. We cap this to not form integers or floating point
11129 // elements wider than 64 bits, but it might be interesting to form i128
11130 // integers to handle flipping the low and high halves of AVX 256-bit vectors.
11131 SmallVector<int, 16> WidenedMask;
11132 if (VT.getScalarSizeInBits() < 64 && !Is1BitVector &&
11133 canWidenShuffleElements(Mask, WidenedMask)) {
11134 MVT NewEltVT = VT.isFloatingPoint()
11135 ? MVT::getFloatingPointVT(VT.getScalarSizeInBits() * 2)
11136 : MVT::getIntegerVT(VT.getScalarSizeInBits() * 2);
11137 MVT NewVT = MVT::getVectorVT(NewEltVT, VT.getVectorNumElements() / 2);
11138 // Make sure that the new vector type is legal. For example, v2f64 isn't
11140 if (DAG.getTargetLoweringInfo().isTypeLegal(NewVT)) {
11141 V1 = DAG.getBitcast(NewVT, V1);
11142 V2 = DAG.getBitcast(NewVT, V2);
11143 return DAG.getBitcast(
11144 VT, DAG.getVectorShuffle(NewVT, dl, V1, V2, WidenedMask));
11148 int NumV1Elements = 0, NumUndefElements = 0, NumV2Elements = 0;
11149 for (int M : SVOp->getMask())
11151 ++NumUndefElements;
11152 else if (M < NumElements)
11157 // Commute the shuffle as needed such that more elements come from V1 than
11158 // V2. This allows us to match the shuffle pattern strictly on how many
11159 // elements come from V1 without handling the symmetric cases.
11160 if (NumV2Elements > NumV1Elements)
11161 return DAG.getCommutedVectorShuffle(*SVOp);
11163 // When the number of V1 and V2 elements are the same, try to minimize the
11164 // number of uses of V2 in the low half of the vector. When that is tied,
11165 // ensure that the sum of indices for V1 is equal to or lower than the sum
11166 // indices for V2. When those are equal, try to ensure that the number of odd
11167 // indices for V1 is lower than the number of odd indices for V2.
11168 if (NumV1Elements == NumV2Elements) {
11169 int LowV1Elements = 0, LowV2Elements = 0;
11170 for (int M : SVOp->getMask().slice(0, NumElements / 2))
11171 if (M >= NumElements)
11175 if (LowV2Elements > LowV1Elements) {
11176 return DAG.getCommutedVectorShuffle(*SVOp);
11177 } else if (LowV2Elements == LowV1Elements) {
11178 int SumV1Indices = 0, SumV2Indices = 0;
11179 for (int i = 0, Size = SVOp->getMask().size(); i < Size; ++i)
11180 if (SVOp->getMask()[i] >= NumElements)
11182 else if (SVOp->getMask()[i] >= 0)
11184 if (SumV2Indices < SumV1Indices) {
11185 return DAG.getCommutedVectorShuffle(*SVOp);
11186 } else if (SumV2Indices == SumV1Indices) {
11187 int NumV1OddIndices = 0, NumV2OddIndices = 0;
11188 for (int i = 0, Size = SVOp->getMask().size(); i < Size; ++i)
11189 if (SVOp->getMask()[i] >= NumElements)
11190 NumV2OddIndices += i % 2;
11191 else if (SVOp->getMask()[i] >= 0)
11192 NumV1OddIndices += i % 2;
11193 if (NumV2OddIndices < NumV1OddIndices)
11194 return DAG.getCommutedVectorShuffle(*SVOp);
11199 // For each vector width, delegate to a specialized lowering routine.
11200 if (VT.getSizeInBits() == 128)
11201 return lower128BitVectorShuffle(Op, V1, V2, VT, Subtarget, DAG);
11203 if (VT.getSizeInBits() == 256)
11204 return lower256BitVectorShuffle(Op, V1, V2, VT, Subtarget, DAG);
11206 if (VT.getSizeInBits() == 512)
11207 return lower512BitVectorShuffle(Op, V1, V2, VT, Subtarget, DAG);
11210 return lower1BitVectorShuffle(Op, V1, V2, VT, Subtarget, DAG);
11211 llvm_unreachable("Unimplemented!");
11214 // This function assumes its argument is a BUILD_VECTOR of constants or
11215 // undef SDNodes. i.e: ISD::isBuildVectorOfConstantSDNodes(BuildVector) is
11217 static bool BUILD_VECTORtoBlendMask(BuildVectorSDNode *BuildVector,
11218 unsigned &MaskValue) {
11220 unsigned NumElems = BuildVector->getNumOperands();
11222 // There are 2 lanes if (NumElems > 8), and 1 lane otherwise.
11223 // We don't handle the >2 lanes case right now.
11224 unsigned NumLanes = (NumElems - 1) / 8 + 1;
11228 unsigned NumElemsInLane = NumElems / NumLanes;
11230 // Blend for v16i16 should be symmetric for the both lanes.
11231 for (unsigned i = 0; i < NumElemsInLane; ++i) {
11232 SDValue EltCond = BuildVector->getOperand(i);
11233 SDValue SndLaneEltCond =
11234 (NumLanes == 2) ? BuildVector->getOperand(i + NumElemsInLane) : EltCond;
11236 int Lane1Cond = -1, Lane2Cond = -1;
11237 if (isa<ConstantSDNode>(EltCond))
11238 Lane1Cond = !isZero(EltCond);
11239 if (isa<ConstantSDNode>(SndLaneEltCond))
11240 Lane2Cond = !isZero(SndLaneEltCond);
11242 unsigned LaneMask = 0;
11243 if (Lane1Cond == Lane2Cond || Lane2Cond < 0)
11244 // Lane1Cond != 0, means we want the first argument.
11245 // Lane1Cond == 0, means we want the second argument.
11246 // The encoding of this argument is 0 for the first argument, 1
11247 // for the second. Therefore, invert the condition.
11248 LaneMask = !Lane1Cond << i;
11249 else if (Lane1Cond < 0)
11250 LaneMask = !Lane2Cond << i;
11254 MaskValue |= LaneMask;
11256 MaskValue |= LaneMask << NumElemsInLane;
11261 /// \brief Try to lower a VSELECT instruction to a vector shuffle.
11262 static SDValue lowerVSELECTtoVectorShuffle(SDValue Op,
11263 const X86Subtarget *Subtarget,
11264 SelectionDAG &DAG) {
11265 SDValue Cond = Op.getOperand(0);
11266 SDValue LHS = Op.getOperand(1);
11267 SDValue RHS = Op.getOperand(2);
11269 MVT VT = Op.getSimpleValueType();
11271 if (!ISD::isBuildVectorOfConstantSDNodes(Cond.getNode()))
11273 auto *CondBV = cast<BuildVectorSDNode>(Cond);
11275 // Only non-legal VSELECTs reach this lowering, convert those into generic
11276 // shuffles and re-use the shuffle lowering path for blends.
11277 SmallVector<int, 32> Mask;
11278 for (int i = 0, Size = VT.getVectorNumElements(); i < Size; ++i) {
11279 SDValue CondElt = CondBV->getOperand(i);
11281 isa<ConstantSDNode>(CondElt) ? i + (isZero(CondElt) ? Size : 0) : -1);
11283 return DAG.getVectorShuffle(VT, dl, LHS, RHS, Mask);
11286 SDValue X86TargetLowering::LowerVSELECT(SDValue Op, SelectionDAG &DAG) const {
11287 // A vselect where all conditions and data are constants can be optimized into
11288 // a single vector load by SelectionDAGLegalize::ExpandBUILD_VECTOR().
11289 if (ISD::isBuildVectorOfConstantSDNodes(Op.getOperand(0).getNode()) &&
11290 ISD::isBuildVectorOfConstantSDNodes(Op.getOperand(1).getNode()) &&
11291 ISD::isBuildVectorOfConstantSDNodes(Op.getOperand(2).getNode()))
11294 // Try to lower this to a blend-style vector shuffle. This can handle all
11295 // constant condition cases.
11296 if (SDValue BlendOp = lowerVSELECTtoVectorShuffle(Op, Subtarget, DAG))
11299 // Variable blends are only legal from SSE4.1 onward.
11300 if (!Subtarget->hasSSE41())
11303 // Only some types will be legal on some subtargets. If we can emit a legal
11304 // VSELECT-matching blend, return Op, and but if we need to expand, return
11306 switch (Op.getSimpleValueType().SimpleTy) {
11308 // Most of the vector types have blends past SSE4.1.
11312 // The byte blends for AVX vectors were introduced only in AVX2.
11313 if (Subtarget->hasAVX2())
11320 // AVX-512 BWI and VLX features support VSELECT with i16 elements.
11321 if (Subtarget->hasBWI() && Subtarget->hasVLX())
11324 // FIXME: We should custom lower this by fixing the condition and using i8
11330 static SDValue LowerEXTRACT_VECTOR_ELT_SSE4(SDValue Op, SelectionDAG &DAG) {
11331 MVT VT = Op.getSimpleValueType();
11334 if (!Op.getOperand(0).getSimpleValueType().is128BitVector())
11337 if (VT.getSizeInBits() == 8) {
11338 SDValue Extract = DAG.getNode(X86ISD::PEXTRB, dl, MVT::i32,
11339 Op.getOperand(0), Op.getOperand(1));
11340 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Extract,
11341 DAG.getValueType(VT));
11342 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
11345 if (VT.getSizeInBits() == 16) {
11346 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
11347 // If Idx is 0, it's cheaper to do a move instead of a pextrw.
11349 return DAG.getNode(
11350 ISD::TRUNCATE, dl, MVT::i16,
11351 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
11352 DAG.getBitcast(MVT::v4i32, Op.getOperand(0)),
11353 Op.getOperand(1)));
11354 SDValue Extract = DAG.getNode(X86ISD::PEXTRW, dl, MVT::i32,
11355 Op.getOperand(0), Op.getOperand(1));
11356 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Extract,
11357 DAG.getValueType(VT));
11358 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
11361 if (VT == MVT::f32) {
11362 // EXTRACTPS outputs to a GPR32 register which will require a movd to copy
11363 // the result back to FR32 register. It's only worth matching if the
11364 // result has a single use which is a store or a bitcast to i32. And in
11365 // the case of a store, it's not worth it if the index is a constant 0,
11366 // because a MOVSSmr can be used instead, which is smaller and faster.
11367 if (!Op.hasOneUse())
11369 SDNode *User = *Op.getNode()->use_begin();
11370 if ((User->getOpcode() != ISD::STORE ||
11371 (isa<ConstantSDNode>(Op.getOperand(1)) &&
11372 cast<ConstantSDNode>(Op.getOperand(1))->isNullValue())) &&
11373 (User->getOpcode() != ISD::BITCAST ||
11374 User->getValueType(0) != MVT::i32))
11376 SDValue Extract = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
11377 DAG.getBitcast(MVT::v4i32, Op.getOperand(0)),
11379 return DAG.getBitcast(MVT::f32, Extract);
11382 if (VT == MVT::i32 || VT == MVT::i64) {
11383 // ExtractPS/pextrq works with constant index.
11384 if (isa<ConstantSDNode>(Op.getOperand(1)))
11390 /// Extract one bit from mask vector, like v16i1 or v8i1.
11391 /// AVX-512 feature.
11393 X86TargetLowering::ExtractBitFromMaskVector(SDValue Op, SelectionDAG &DAG) const {
11394 SDValue Vec = Op.getOperand(0);
11396 MVT VecVT = Vec.getSimpleValueType();
11397 SDValue Idx = Op.getOperand(1);
11398 MVT EltVT = Op.getSimpleValueType();
11400 assert((EltVT == MVT::i1) && "Unexpected operands in ExtractBitFromMaskVector");
11401 assert((VecVT.getVectorNumElements() <= 16 || Subtarget->hasBWI()) &&
11402 "Unexpected vector type in ExtractBitFromMaskVector");
11404 // variable index can't be handled in mask registers,
11405 // extend vector to VR512
11406 if (!isa<ConstantSDNode>(Idx)) {
11407 MVT ExtVT = (VecVT == MVT::v8i1 ? MVT::v8i64 : MVT::v16i32);
11408 SDValue Ext = DAG.getNode(ISD::ZERO_EXTEND, dl, ExtVT, Vec);
11409 SDValue Elt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl,
11410 ExtVT.getVectorElementType(), Ext, Idx);
11411 return DAG.getNode(ISD::TRUNCATE, dl, EltVT, Elt);
11414 unsigned IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
11415 const TargetRegisterClass* rc = getRegClassFor(VecVT);
11416 if (!Subtarget->hasDQI() && (VecVT.getVectorNumElements() <= 8))
11417 rc = getRegClassFor(MVT::v16i1);
11418 unsigned MaxSift = rc->getSize()*8 - 1;
11419 Vec = DAG.getNode(X86ISD::VSHLI, dl, VecVT, Vec,
11420 DAG.getConstant(MaxSift - IdxVal, dl, MVT::i8));
11421 Vec = DAG.getNode(X86ISD::VSRLI, dl, VecVT, Vec,
11422 DAG.getConstant(MaxSift, dl, MVT::i8));
11423 return DAG.getNode(X86ISD::VEXTRACT, dl, MVT::i1, Vec,
11424 DAG.getIntPtrConstant(0, dl));
11428 X86TargetLowering::LowerEXTRACT_VECTOR_ELT(SDValue Op,
11429 SelectionDAG &DAG) const {
11431 SDValue Vec = Op.getOperand(0);
11432 MVT VecVT = Vec.getSimpleValueType();
11433 SDValue Idx = Op.getOperand(1);
11435 if (Op.getSimpleValueType() == MVT::i1)
11436 return ExtractBitFromMaskVector(Op, DAG);
11438 if (!isa<ConstantSDNode>(Idx)) {
11439 if (VecVT.is512BitVector() ||
11440 (VecVT.is256BitVector() && Subtarget->hasInt256() &&
11441 VecVT.getVectorElementType().getSizeInBits() == 32)) {
11444 MVT::getIntegerVT(VecVT.getVectorElementType().getSizeInBits());
11445 MVT MaskVT = MVT::getVectorVT(MaskEltVT, VecVT.getSizeInBits() /
11446 MaskEltVT.getSizeInBits());
11448 Idx = DAG.getZExtOrTrunc(Idx, dl, MaskEltVT);
11449 auto PtrVT = getPointerTy(DAG.getDataLayout());
11450 SDValue Mask = DAG.getNode(X86ISD::VINSERT, dl, MaskVT,
11451 getZeroVector(MaskVT, Subtarget, DAG, dl), Idx,
11452 DAG.getConstant(0, dl, PtrVT));
11453 SDValue Perm = DAG.getNode(X86ISD::VPERMV, dl, VecVT, Mask, Vec);
11454 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, Op.getValueType(), Perm,
11455 DAG.getConstant(0, dl, PtrVT));
11460 // If this is a 256-bit vector result, first extract the 128-bit vector and
11461 // then extract the element from the 128-bit vector.
11462 if (VecVT.is256BitVector() || VecVT.is512BitVector()) {
11464 unsigned IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
11465 // Get the 128-bit vector.
11466 Vec = Extract128BitVector(Vec, IdxVal, DAG, dl);
11467 MVT EltVT = VecVT.getVectorElementType();
11469 unsigned ElemsPerChunk = 128 / EltVT.getSizeInBits();
11471 //if (IdxVal >= NumElems/2)
11472 // IdxVal -= NumElems/2;
11473 IdxVal -= (IdxVal/ElemsPerChunk)*ElemsPerChunk;
11474 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, Op.getValueType(), Vec,
11475 DAG.getConstant(IdxVal, dl, MVT::i32));
11478 assert(VecVT.is128BitVector() && "Unexpected vector length");
11480 if (Subtarget->hasSSE41())
11481 if (SDValue Res = LowerEXTRACT_VECTOR_ELT_SSE4(Op, DAG))
11484 MVT VT = Op.getSimpleValueType();
11485 // TODO: handle v16i8.
11486 if (VT.getSizeInBits() == 16) {
11487 SDValue Vec = Op.getOperand(0);
11488 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
11490 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i16,
11491 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
11492 DAG.getBitcast(MVT::v4i32, Vec),
11493 Op.getOperand(1)));
11494 // Transform it so it match pextrw which produces a 32-bit result.
11495 MVT EltVT = MVT::i32;
11496 SDValue Extract = DAG.getNode(X86ISD::PEXTRW, dl, EltVT,
11497 Op.getOperand(0), Op.getOperand(1));
11498 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, EltVT, Extract,
11499 DAG.getValueType(VT));
11500 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
11503 if (VT.getSizeInBits() == 32) {
11504 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
11508 // SHUFPS the element to the lowest double word, then movss.
11509 int Mask[4] = { static_cast<int>(Idx), -1, -1, -1 };
11510 MVT VVT = Op.getOperand(0).getSimpleValueType();
11511 SDValue Vec = DAG.getVectorShuffle(VVT, dl, Op.getOperand(0),
11512 DAG.getUNDEF(VVT), Mask);
11513 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, Vec,
11514 DAG.getIntPtrConstant(0, dl));
11517 if (VT.getSizeInBits() == 64) {
11518 // FIXME: .td only matches this for <2 x f64>, not <2 x i64> on 32b
11519 // FIXME: seems like this should be unnecessary if mov{h,l}pd were taught
11520 // to match extract_elt for f64.
11521 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
11525 // UNPCKHPD the element to the lowest double word, then movsd.
11526 // Note if the lower 64 bits of the result of the UNPCKHPD is then stored
11527 // to a f64mem, the whole operation is folded into a single MOVHPDmr.
11528 int Mask[2] = { 1, -1 };
11529 MVT VVT = Op.getOperand(0).getSimpleValueType();
11530 SDValue Vec = DAG.getVectorShuffle(VVT, dl, Op.getOperand(0),
11531 DAG.getUNDEF(VVT), Mask);
11532 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, Vec,
11533 DAG.getIntPtrConstant(0, dl));
11539 /// Insert one bit to mask vector, like v16i1 or v8i1.
11540 /// AVX-512 feature.
11542 X86TargetLowering::InsertBitToMaskVector(SDValue Op, SelectionDAG &DAG) const {
11544 SDValue Vec = Op.getOperand(0);
11545 SDValue Elt = Op.getOperand(1);
11546 SDValue Idx = Op.getOperand(2);
11547 MVT VecVT = Vec.getSimpleValueType();
11549 if (!isa<ConstantSDNode>(Idx)) {
11550 // Non constant index. Extend source and destination,
11551 // insert element and then truncate the result.
11552 MVT ExtVecVT = (VecVT == MVT::v8i1 ? MVT::v8i64 : MVT::v16i32);
11553 MVT ExtEltVT = (VecVT == MVT::v8i1 ? MVT::i64 : MVT::i32);
11554 SDValue ExtOp = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, ExtVecVT,
11555 DAG.getNode(ISD::ZERO_EXTEND, dl, ExtVecVT, Vec),
11556 DAG.getNode(ISD::ZERO_EXTEND, dl, ExtEltVT, Elt), Idx);
11557 return DAG.getNode(ISD::TRUNCATE, dl, VecVT, ExtOp);
11560 unsigned IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
11561 SDValue EltInVec = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VecVT, Elt);
11563 EltInVec = DAG.getNode(X86ISD::VSHLI, dl, VecVT, EltInVec,
11564 DAG.getConstant(IdxVal, dl, MVT::i8));
11565 if (Vec.getOpcode() == ISD::UNDEF)
11567 return DAG.getNode(ISD::OR, dl, VecVT, Vec, EltInVec);
11570 SDValue X86TargetLowering::LowerINSERT_VECTOR_ELT(SDValue Op,
11571 SelectionDAG &DAG) const {
11572 MVT VT = Op.getSimpleValueType();
11573 MVT EltVT = VT.getVectorElementType();
11575 if (EltVT == MVT::i1)
11576 return InsertBitToMaskVector(Op, DAG);
11579 SDValue N0 = Op.getOperand(0);
11580 SDValue N1 = Op.getOperand(1);
11581 SDValue N2 = Op.getOperand(2);
11582 if (!isa<ConstantSDNode>(N2))
11584 auto *N2C = cast<ConstantSDNode>(N2);
11585 unsigned IdxVal = N2C->getZExtValue();
11587 // If the vector is wider than 128 bits, extract the 128-bit subvector, insert
11588 // into that, and then insert the subvector back into the result.
11589 if (VT.is256BitVector() || VT.is512BitVector()) {
11590 // With a 256-bit vector, we can insert into the zero element efficiently
11591 // using a blend if we have AVX or AVX2 and the right data type.
11592 if (VT.is256BitVector() && IdxVal == 0) {
11593 // TODO: It is worthwhile to cast integer to floating point and back
11594 // and incur a domain crossing penalty if that's what we'll end up
11595 // doing anyway after extracting to a 128-bit vector.
11596 if ((Subtarget->hasAVX() && (EltVT == MVT::f64 || EltVT == MVT::f32)) ||
11597 (Subtarget->hasAVX2() && EltVT == MVT::i32)) {
11598 SDValue N1Vec = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, N1);
11599 N2 = DAG.getIntPtrConstant(1, dl);
11600 return DAG.getNode(X86ISD::BLENDI, dl, VT, N0, N1Vec, N2);
11604 // Get the desired 128-bit vector chunk.
11605 SDValue V = Extract128BitVector(N0, IdxVal, DAG, dl);
11607 // Insert the element into the desired chunk.
11608 unsigned NumEltsIn128 = 128 / EltVT.getSizeInBits();
11609 unsigned IdxIn128 = IdxVal - (IdxVal / NumEltsIn128) * NumEltsIn128;
11611 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, V.getValueType(), V, N1,
11612 DAG.getConstant(IdxIn128, dl, MVT::i32));
11614 // Insert the changed part back into the bigger vector
11615 return Insert128BitVector(N0, V, IdxVal, DAG, dl);
11617 assert(VT.is128BitVector() && "Only 128-bit vector types should be left!");
11619 if (Subtarget->hasSSE41()) {
11620 if (EltVT.getSizeInBits() == 8 || EltVT.getSizeInBits() == 16) {
11622 if (VT == MVT::v8i16) {
11623 Opc = X86ISD::PINSRW;
11625 assert(VT == MVT::v16i8);
11626 Opc = X86ISD::PINSRB;
11629 // Transform it so it match pinsr{b,w} which expects a GR32 as its second
11631 if (N1.getValueType() != MVT::i32)
11632 N1 = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, N1);
11633 if (N2.getValueType() != MVT::i32)
11634 N2 = DAG.getIntPtrConstant(IdxVal, dl);
11635 return DAG.getNode(Opc, dl, VT, N0, N1, N2);
11638 if (EltVT == MVT::f32) {
11639 // Bits [7:6] of the constant are the source select. This will always be
11640 // zero here. The DAG Combiner may combine an extract_elt index into
11641 // these bits. For example (insert (extract, 3), 2) could be matched by
11642 // putting the '3' into bits [7:6] of X86ISD::INSERTPS.
11643 // Bits [5:4] of the constant are the destination select. This is the
11644 // value of the incoming immediate.
11645 // Bits [3:0] of the constant are the zero mask. The DAG Combiner may
11646 // combine either bitwise AND or insert of float 0.0 to set these bits.
11648 bool MinSize = DAG.getMachineFunction().getFunction()->optForMinSize();
11649 if (IdxVal == 0 && (!MinSize || !MayFoldLoad(N1))) {
11650 // If this is an insertion of 32-bits into the low 32-bits of
11651 // a vector, we prefer to generate a blend with immediate rather
11652 // than an insertps. Blends are simpler operations in hardware and so
11653 // will always have equal or better performance than insertps.
11654 // But if optimizing for size and there's a load folding opportunity,
11655 // generate insertps because blendps does not have a 32-bit memory
11657 N2 = DAG.getIntPtrConstant(1, dl);
11658 N1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4f32, N1);
11659 return DAG.getNode(X86ISD::BLENDI, dl, VT, N0, N1, N2);
11661 N2 = DAG.getIntPtrConstant(IdxVal << 4, dl);
11662 // Create this as a scalar to vector..
11663 N1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4f32, N1);
11664 return DAG.getNode(X86ISD::INSERTPS, dl, VT, N0, N1, N2);
11667 if (EltVT == MVT::i32 || EltVT == MVT::i64) {
11668 // PINSR* works with constant index.
11673 if (EltVT == MVT::i8)
11676 if (EltVT.getSizeInBits() == 16) {
11677 // Transform it so it match pinsrw which expects a 16-bit value in a GR32
11678 // as its second argument.
11679 if (N1.getValueType() != MVT::i32)
11680 N1 = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, N1);
11681 if (N2.getValueType() != MVT::i32)
11682 N2 = DAG.getIntPtrConstant(IdxVal, dl);
11683 return DAG.getNode(X86ISD::PINSRW, dl, VT, N0, N1, N2);
11688 static SDValue LowerSCALAR_TO_VECTOR(SDValue Op, SelectionDAG &DAG) {
11690 MVT OpVT = Op.getSimpleValueType();
11692 // If this is a 256-bit vector result, first insert into a 128-bit
11693 // vector and then insert into the 256-bit vector.
11694 if (!OpVT.is128BitVector()) {
11695 // Insert into a 128-bit vector.
11696 unsigned SizeFactor = OpVT.getSizeInBits()/128;
11697 MVT VT128 = MVT::getVectorVT(OpVT.getVectorElementType(),
11698 OpVT.getVectorNumElements() / SizeFactor);
11700 Op = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT128, Op.getOperand(0));
11702 // Insert the 128-bit vector.
11703 return Insert128BitVector(DAG.getUNDEF(OpVT), Op, 0, DAG, dl);
11706 if (OpVT == MVT::v1i64 &&
11707 Op.getOperand(0).getValueType() == MVT::i64)
11708 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v1i64, Op.getOperand(0));
11710 SDValue AnyExt = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, Op.getOperand(0));
11711 assert(OpVT.is128BitVector() && "Expected an SSE type!");
11712 return DAG.getBitcast(
11713 OpVT, DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32, AnyExt));
11716 // Lower a node with an EXTRACT_SUBVECTOR opcode. This may result in
11717 // a simple subregister reference or explicit instructions to grab
11718 // upper bits of a vector.
11719 static SDValue LowerEXTRACT_SUBVECTOR(SDValue Op, const X86Subtarget *Subtarget,
11720 SelectionDAG &DAG) {
11722 SDValue In = Op.getOperand(0);
11723 SDValue Idx = Op.getOperand(1);
11724 unsigned IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
11725 MVT ResVT = Op.getSimpleValueType();
11726 MVT InVT = In.getSimpleValueType();
11728 if (Subtarget->hasFp256()) {
11729 if (ResVT.is128BitVector() &&
11730 (InVT.is256BitVector() || InVT.is512BitVector()) &&
11731 isa<ConstantSDNode>(Idx)) {
11732 return Extract128BitVector(In, IdxVal, DAG, dl);
11734 if (ResVT.is256BitVector() && InVT.is512BitVector() &&
11735 isa<ConstantSDNode>(Idx)) {
11736 return Extract256BitVector(In, IdxVal, DAG, dl);
11742 // Lower a node with an INSERT_SUBVECTOR opcode. This may result in a
11743 // simple superregister reference or explicit instructions to insert
11744 // the upper bits of a vector.
11745 static SDValue LowerINSERT_SUBVECTOR(SDValue Op, const X86Subtarget *Subtarget,
11746 SelectionDAG &DAG) {
11747 if (!Subtarget->hasAVX())
11751 SDValue Vec = Op.getOperand(0);
11752 SDValue SubVec = Op.getOperand(1);
11753 SDValue Idx = Op.getOperand(2);
11755 if (!isa<ConstantSDNode>(Idx))
11758 unsigned IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
11759 MVT OpVT = Op.getSimpleValueType();
11760 MVT SubVecVT = SubVec.getSimpleValueType();
11762 // Fold two 16-byte subvector loads into one 32-byte load:
11763 // (insert_subvector (insert_subvector undef, (load addr), 0),
11764 // (load addr + 16), Elts/2)
11766 if ((IdxVal == OpVT.getVectorNumElements() / 2) &&
11767 Vec.getOpcode() == ISD::INSERT_SUBVECTOR &&
11768 OpVT.is256BitVector() && SubVecVT.is128BitVector()) {
11769 auto *Idx2 = dyn_cast<ConstantSDNode>(Vec.getOperand(2));
11770 if (Idx2 && Idx2->getZExtValue() == 0) {
11771 SDValue SubVec2 = Vec.getOperand(1);
11772 // If needed, look through a bitcast to get to the load.
11773 if (SubVec2.getNode() && SubVec2.getOpcode() == ISD::BITCAST)
11774 SubVec2 = SubVec2.getOperand(0);
11776 if (auto *FirstLd = dyn_cast<LoadSDNode>(SubVec2)) {
11778 unsigned Alignment = FirstLd->getAlignment();
11779 unsigned AS = FirstLd->getAddressSpace();
11780 const X86TargetLowering *TLI = Subtarget->getTargetLowering();
11781 if (TLI->allowsMemoryAccess(*DAG.getContext(), DAG.getDataLayout(),
11782 OpVT, AS, Alignment, &Fast) && Fast) {
11783 SDValue Ops[] = { SubVec2, SubVec };
11784 if (SDValue Ld = EltsFromConsecutiveLoads(OpVT, Ops, dl, DAG, false))
11791 if ((OpVT.is256BitVector() || OpVT.is512BitVector()) &&
11792 SubVecVT.is128BitVector())
11793 return Insert128BitVector(Vec, SubVec, IdxVal, DAG, dl);
11795 if (OpVT.is512BitVector() && SubVecVT.is256BitVector())
11796 return Insert256BitVector(Vec, SubVec, IdxVal, DAG, dl);
11798 if (OpVT.getVectorElementType() == MVT::i1) {
11799 if (IdxVal == 0 && Vec.getOpcode() == ISD::UNDEF) // the operation is legal
11801 SDValue ZeroIdx = DAG.getIntPtrConstant(0, dl);
11802 SDValue Undef = DAG.getUNDEF(OpVT);
11803 unsigned NumElems = OpVT.getVectorNumElements();
11804 SDValue ShiftBits = DAG.getConstant(NumElems/2, dl, MVT::i8);
11806 if (IdxVal == OpVT.getVectorNumElements() / 2) {
11807 // Zero upper bits of the Vec
11808 Vec = DAG.getNode(X86ISD::VSHLI, dl, OpVT, Vec, ShiftBits);
11809 Vec = DAG.getNode(X86ISD::VSRLI, dl, OpVT, Vec, ShiftBits);
11811 SDValue Vec2 = DAG.getNode(ISD::INSERT_SUBVECTOR, dl, OpVT, Undef,
11813 Vec2 = DAG.getNode(X86ISD::VSHLI, dl, OpVT, Vec2, ShiftBits);
11814 return DAG.getNode(ISD::OR, dl, OpVT, Vec, Vec2);
11817 SDValue Vec2 = DAG.getNode(ISD::INSERT_SUBVECTOR, dl, OpVT, Undef,
11819 // Zero upper bits of the Vec2
11820 Vec2 = DAG.getNode(X86ISD::VSHLI, dl, OpVT, Vec2, ShiftBits);
11821 Vec2 = DAG.getNode(X86ISD::VSRLI, dl, OpVT, Vec2, ShiftBits);
11822 // Zero lower bits of the Vec
11823 Vec = DAG.getNode(X86ISD::VSRLI, dl, OpVT, Vec, ShiftBits);
11824 Vec = DAG.getNode(X86ISD::VSHLI, dl, OpVT, Vec, ShiftBits);
11825 // Merge them together
11826 return DAG.getNode(ISD::OR, dl, OpVT, Vec, Vec2);
11832 // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
11833 // their target countpart wrapped in the X86ISD::Wrapper node. Suppose N is
11834 // one of the above mentioned nodes. It has to be wrapped because otherwise
11835 // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
11836 // be used to form addressing mode. These wrapped nodes will be selected
11839 X86TargetLowering::LowerConstantPool(SDValue Op, SelectionDAG &DAG) const {
11840 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
11842 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
11843 // global base reg.
11844 unsigned char OpFlag = 0;
11845 unsigned WrapperKind = X86ISD::Wrapper;
11846 CodeModel::Model M = DAG.getTarget().getCodeModel();
11848 if (Subtarget->isPICStyleRIPRel() &&
11849 (M == CodeModel::Small || M == CodeModel::Kernel))
11850 WrapperKind = X86ISD::WrapperRIP;
11851 else if (Subtarget->isPICStyleGOT())
11852 OpFlag = X86II::MO_GOTOFF;
11853 else if (Subtarget->isPICStyleStubPIC())
11854 OpFlag = X86II::MO_PIC_BASE_OFFSET;
11856 auto PtrVT = getPointerTy(DAG.getDataLayout());
11857 SDValue Result = DAG.getTargetConstantPool(
11858 CP->getConstVal(), PtrVT, CP->getAlignment(), CP->getOffset(), OpFlag);
11860 Result = DAG.getNode(WrapperKind, DL, PtrVT, Result);
11861 // With PIC, the address is actually $g + Offset.
11864 DAG.getNode(ISD::ADD, DL, PtrVT,
11865 DAG.getNode(X86ISD::GlobalBaseReg, SDLoc(), PtrVT), Result);
11871 SDValue X86TargetLowering::LowerJumpTable(SDValue Op, SelectionDAG &DAG) const {
11872 JumpTableSDNode *JT = cast<JumpTableSDNode>(Op);
11874 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
11875 // global base reg.
11876 unsigned char OpFlag = 0;
11877 unsigned WrapperKind = X86ISD::Wrapper;
11878 CodeModel::Model M = DAG.getTarget().getCodeModel();
11880 if (Subtarget->isPICStyleRIPRel() &&
11881 (M == CodeModel::Small || M == CodeModel::Kernel))
11882 WrapperKind = X86ISD::WrapperRIP;
11883 else if (Subtarget->isPICStyleGOT())
11884 OpFlag = X86II::MO_GOTOFF;
11885 else if (Subtarget->isPICStyleStubPIC())
11886 OpFlag = X86II::MO_PIC_BASE_OFFSET;
11888 auto PtrVT = getPointerTy(DAG.getDataLayout());
11889 SDValue Result = DAG.getTargetJumpTable(JT->getIndex(), PtrVT, OpFlag);
11891 Result = DAG.getNode(WrapperKind, DL, PtrVT, Result);
11893 // With PIC, the address is actually $g + Offset.
11896 DAG.getNode(ISD::ADD, DL, PtrVT,
11897 DAG.getNode(X86ISD::GlobalBaseReg, SDLoc(), PtrVT), Result);
11903 X86TargetLowering::LowerExternalSymbol(SDValue Op, SelectionDAG &DAG) const {
11904 const char *Sym = cast<ExternalSymbolSDNode>(Op)->getSymbol();
11906 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
11907 // global base reg.
11908 unsigned char OpFlag = 0;
11909 unsigned WrapperKind = X86ISD::Wrapper;
11910 CodeModel::Model M = DAG.getTarget().getCodeModel();
11912 if (Subtarget->isPICStyleRIPRel() &&
11913 (M == CodeModel::Small || M == CodeModel::Kernel)) {
11914 if (Subtarget->isTargetDarwin() || Subtarget->isTargetELF())
11915 OpFlag = X86II::MO_GOTPCREL;
11916 WrapperKind = X86ISD::WrapperRIP;
11917 } else if (Subtarget->isPICStyleGOT()) {
11918 OpFlag = X86II::MO_GOT;
11919 } else if (Subtarget->isPICStyleStubPIC()) {
11920 OpFlag = X86II::MO_DARWIN_NONLAZY_PIC_BASE;
11921 } else if (Subtarget->isPICStyleStubNoDynamic()) {
11922 OpFlag = X86II::MO_DARWIN_NONLAZY;
11925 auto PtrVT = getPointerTy(DAG.getDataLayout());
11926 SDValue Result = DAG.getTargetExternalSymbol(Sym, PtrVT, OpFlag);
11929 Result = DAG.getNode(WrapperKind, DL, PtrVT, Result);
11931 // With PIC, the address is actually $g + Offset.
11932 if (DAG.getTarget().getRelocationModel() == Reloc::PIC_ &&
11933 !Subtarget->is64Bit()) {
11935 DAG.getNode(ISD::ADD, DL, PtrVT,
11936 DAG.getNode(X86ISD::GlobalBaseReg, SDLoc(), PtrVT), Result);
11939 // For symbols that require a load from a stub to get the address, emit the
11941 if (isGlobalStubReference(OpFlag))
11942 Result = DAG.getLoad(PtrVT, DL, DAG.getEntryNode(), Result,
11943 MachinePointerInfo::getGOT(DAG.getMachineFunction()),
11944 false, false, false, 0);
11950 X86TargetLowering::LowerBlockAddress(SDValue Op, SelectionDAG &DAG) const {
11951 // Create the TargetBlockAddressAddress node.
11952 unsigned char OpFlags =
11953 Subtarget->ClassifyBlockAddressReference();
11954 CodeModel::Model M = DAG.getTarget().getCodeModel();
11955 const BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress();
11956 int64_t Offset = cast<BlockAddressSDNode>(Op)->getOffset();
11958 auto PtrVT = getPointerTy(DAG.getDataLayout());
11959 SDValue Result = DAG.getTargetBlockAddress(BA, PtrVT, Offset, OpFlags);
11961 if (Subtarget->isPICStyleRIPRel() &&
11962 (M == CodeModel::Small || M == CodeModel::Kernel))
11963 Result = DAG.getNode(X86ISD::WrapperRIP, dl, PtrVT, Result);
11965 Result = DAG.getNode(X86ISD::Wrapper, dl, PtrVT, Result);
11967 // With PIC, the address is actually $g + Offset.
11968 if (isGlobalRelativeToPICBase(OpFlags)) {
11969 Result = DAG.getNode(ISD::ADD, dl, PtrVT,
11970 DAG.getNode(X86ISD::GlobalBaseReg, dl, PtrVT), Result);
11977 X86TargetLowering::LowerGlobalAddress(const GlobalValue *GV, SDLoc dl,
11978 int64_t Offset, SelectionDAG &DAG) const {
11979 // Create the TargetGlobalAddress node, folding in the constant
11980 // offset if it is legal.
11981 unsigned char OpFlags =
11982 Subtarget->ClassifyGlobalReference(GV, DAG.getTarget());
11983 CodeModel::Model M = DAG.getTarget().getCodeModel();
11984 auto PtrVT = getPointerTy(DAG.getDataLayout());
11986 if (OpFlags == X86II::MO_NO_FLAG &&
11987 X86::isOffsetSuitableForCodeModel(Offset, M)) {
11988 // A direct static reference to a global.
11989 Result = DAG.getTargetGlobalAddress(GV, dl, PtrVT, Offset);
11992 Result = DAG.getTargetGlobalAddress(GV, dl, PtrVT, 0, OpFlags);
11995 if (Subtarget->isPICStyleRIPRel() &&
11996 (M == CodeModel::Small || M == CodeModel::Kernel))
11997 Result = DAG.getNode(X86ISD::WrapperRIP, dl, PtrVT, Result);
11999 Result = DAG.getNode(X86ISD::Wrapper, dl, PtrVT, Result);
12001 // With PIC, the address is actually $g + Offset.
12002 if (isGlobalRelativeToPICBase(OpFlags)) {
12003 Result = DAG.getNode(ISD::ADD, dl, PtrVT,
12004 DAG.getNode(X86ISD::GlobalBaseReg, dl, PtrVT), Result);
12007 // For globals that require a load from a stub to get the address, emit the
12009 if (isGlobalStubReference(OpFlags))
12010 Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Result,
12011 MachinePointerInfo::getGOT(DAG.getMachineFunction()),
12012 false, false, false, 0);
12014 // If there was a non-zero offset that we didn't fold, create an explicit
12015 // addition for it.
12017 Result = DAG.getNode(ISD::ADD, dl, PtrVT, Result,
12018 DAG.getConstant(Offset, dl, PtrVT));
12024 X86TargetLowering::LowerGlobalAddress(SDValue Op, SelectionDAG &DAG) const {
12025 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
12026 int64_t Offset = cast<GlobalAddressSDNode>(Op)->getOffset();
12027 return LowerGlobalAddress(GV, SDLoc(Op), Offset, DAG);
12031 GetTLSADDR(SelectionDAG &DAG, SDValue Chain, GlobalAddressSDNode *GA,
12032 SDValue *InFlag, const EVT PtrVT, unsigned ReturnReg,
12033 unsigned char OperandFlags, bool LocalDynamic = false) {
12034 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
12035 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
12037 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(), dl,
12038 GA->getValueType(0),
12042 X86ISD::NodeType CallType = LocalDynamic ? X86ISD::TLSBASEADDR
12046 SDValue Ops[] = { Chain, TGA, *InFlag };
12047 Chain = DAG.getNode(CallType, dl, NodeTys, Ops);
12049 SDValue Ops[] = { Chain, TGA };
12050 Chain = DAG.getNode(CallType, dl, NodeTys, Ops);
12053 // TLSADDR will be codegen'ed as call. Inform MFI that function has calls.
12054 MFI->setAdjustsStack(true);
12055 MFI->setHasCalls(true);
12057 SDValue Flag = Chain.getValue(1);
12058 return DAG.getCopyFromReg(Chain, dl, ReturnReg, PtrVT, Flag);
12061 // Lower ISD::GlobalTLSAddress using the "general dynamic" model, 32 bit
12063 LowerToTLSGeneralDynamicModel32(GlobalAddressSDNode *GA, SelectionDAG &DAG,
12066 SDLoc dl(GA); // ? function entry point might be better
12067 SDValue Chain = DAG.getCopyToReg(DAG.getEntryNode(), dl, X86::EBX,
12068 DAG.getNode(X86ISD::GlobalBaseReg,
12069 SDLoc(), PtrVT), InFlag);
12070 InFlag = Chain.getValue(1);
12072 return GetTLSADDR(DAG, Chain, GA, &InFlag, PtrVT, X86::EAX, X86II::MO_TLSGD);
12075 // Lower ISD::GlobalTLSAddress using the "general dynamic" model, 64 bit
12077 LowerToTLSGeneralDynamicModel64(GlobalAddressSDNode *GA, SelectionDAG &DAG,
12079 return GetTLSADDR(DAG, DAG.getEntryNode(), GA, nullptr, PtrVT,
12080 X86::RAX, X86II::MO_TLSGD);
12083 static SDValue LowerToTLSLocalDynamicModel(GlobalAddressSDNode *GA,
12089 // Get the start address of the TLS block for this module.
12090 X86MachineFunctionInfo* MFI = DAG.getMachineFunction()
12091 .getInfo<X86MachineFunctionInfo>();
12092 MFI->incNumLocalDynamicTLSAccesses();
12096 Base = GetTLSADDR(DAG, DAG.getEntryNode(), GA, nullptr, PtrVT, X86::RAX,
12097 X86II::MO_TLSLD, /*LocalDynamic=*/true);
12100 SDValue Chain = DAG.getCopyToReg(DAG.getEntryNode(), dl, X86::EBX,
12101 DAG.getNode(X86ISD::GlobalBaseReg, SDLoc(), PtrVT), InFlag);
12102 InFlag = Chain.getValue(1);
12103 Base = GetTLSADDR(DAG, Chain, GA, &InFlag, PtrVT, X86::EAX,
12104 X86II::MO_TLSLDM, /*LocalDynamic=*/true);
12107 // Note: the CleanupLocalDynamicTLSPass will remove redundant computations
12111 unsigned char OperandFlags = X86II::MO_DTPOFF;
12112 unsigned WrapperKind = X86ISD::Wrapper;
12113 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(), dl,
12114 GA->getValueType(0),
12115 GA->getOffset(), OperandFlags);
12116 SDValue Offset = DAG.getNode(WrapperKind, dl, PtrVT, TGA);
12118 // Add x@dtpoff with the base.
12119 return DAG.getNode(ISD::ADD, dl, PtrVT, Offset, Base);
12122 // Lower ISD::GlobalTLSAddress using the "initial exec" or "local exec" model.
12123 static SDValue LowerToTLSExecModel(GlobalAddressSDNode *GA, SelectionDAG &DAG,
12124 const EVT PtrVT, TLSModel::Model model,
12125 bool is64Bit, bool isPIC) {
12128 // Get the Thread Pointer, which is %gs:0 (32-bit) or %fs:0 (64-bit).
12129 Value *Ptr = Constant::getNullValue(Type::getInt8PtrTy(*DAG.getContext(),
12130 is64Bit ? 257 : 256));
12132 SDValue ThreadPointer =
12133 DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), DAG.getIntPtrConstant(0, dl),
12134 MachinePointerInfo(Ptr), false, false, false, 0);
12136 unsigned char OperandFlags = 0;
12137 // Most TLS accesses are not RIP relative, even on x86-64. One exception is
12139 unsigned WrapperKind = X86ISD::Wrapper;
12140 if (model == TLSModel::LocalExec) {
12141 OperandFlags = is64Bit ? X86II::MO_TPOFF : X86II::MO_NTPOFF;
12142 } else if (model == TLSModel::InitialExec) {
12144 OperandFlags = X86II::MO_GOTTPOFF;
12145 WrapperKind = X86ISD::WrapperRIP;
12147 OperandFlags = isPIC ? X86II::MO_GOTNTPOFF : X86II::MO_INDNTPOFF;
12150 llvm_unreachable("Unexpected model");
12153 // emit "addl x@ntpoff,%eax" (local exec)
12154 // or "addl x@indntpoff,%eax" (initial exec)
12155 // or "addl x@gotntpoff(%ebx) ,%eax" (initial exec, 32-bit pic)
12157 DAG.getTargetGlobalAddress(GA->getGlobal(), dl, GA->getValueType(0),
12158 GA->getOffset(), OperandFlags);
12159 SDValue Offset = DAG.getNode(WrapperKind, dl, PtrVT, TGA);
12161 if (model == TLSModel::InitialExec) {
12162 if (isPIC && !is64Bit) {
12163 Offset = DAG.getNode(ISD::ADD, dl, PtrVT,
12164 DAG.getNode(X86ISD::GlobalBaseReg, SDLoc(), PtrVT),
12168 Offset = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Offset,
12169 MachinePointerInfo::getGOT(DAG.getMachineFunction()),
12170 false, false, false, 0);
12173 // The address of the thread local variable is the add of the thread
12174 // pointer with the offset of the variable.
12175 return DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, Offset);
12179 X86TargetLowering::LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const {
12181 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
12182 const GlobalValue *GV = GA->getGlobal();
12183 auto PtrVT = getPointerTy(DAG.getDataLayout());
12185 if (Subtarget->isTargetELF()) {
12186 if (DAG.getTarget().Options.EmulatedTLS)
12187 return LowerToTLSEmulatedModel(GA, DAG);
12188 TLSModel::Model model = DAG.getTarget().getTLSModel(GV);
12190 case TLSModel::GeneralDynamic:
12191 if (Subtarget->is64Bit())
12192 return LowerToTLSGeneralDynamicModel64(GA, DAG, PtrVT);
12193 return LowerToTLSGeneralDynamicModel32(GA, DAG, PtrVT);
12194 case TLSModel::LocalDynamic:
12195 return LowerToTLSLocalDynamicModel(GA, DAG, PtrVT,
12196 Subtarget->is64Bit());
12197 case TLSModel::InitialExec:
12198 case TLSModel::LocalExec:
12199 return LowerToTLSExecModel(GA, DAG, PtrVT, model, Subtarget->is64Bit(),
12200 DAG.getTarget().getRelocationModel() ==
12203 llvm_unreachable("Unknown TLS model.");
12206 if (Subtarget->isTargetDarwin()) {
12207 // Darwin only has one model of TLS. Lower to that.
12208 unsigned char OpFlag = 0;
12209 unsigned WrapperKind = Subtarget->isPICStyleRIPRel() ?
12210 X86ISD::WrapperRIP : X86ISD::Wrapper;
12212 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
12213 // global base reg.
12214 bool PIC32 = (DAG.getTarget().getRelocationModel() == Reloc::PIC_) &&
12215 !Subtarget->is64Bit();
12217 OpFlag = X86II::MO_TLVP_PIC_BASE;
12219 OpFlag = X86II::MO_TLVP;
12221 SDValue Result = DAG.getTargetGlobalAddress(GA->getGlobal(), DL,
12222 GA->getValueType(0),
12223 GA->getOffset(), OpFlag);
12224 SDValue Offset = DAG.getNode(WrapperKind, DL, PtrVT, Result);
12226 // With PIC32, the address is actually $g + Offset.
12228 Offset = DAG.getNode(ISD::ADD, DL, PtrVT,
12229 DAG.getNode(X86ISD::GlobalBaseReg, SDLoc(), PtrVT),
12232 // Lowering the machine isd will make sure everything is in the right
12234 SDValue Chain = DAG.getEntryNode();
12235 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
12236 SDValue Args[] = { Chain, Offset };
12237 Chain = DAG.getNode(X86ISD::TLSCALL, DL, NodeTys, Args);
12239 // TLSCALL will be codegen'ed as call. Inform MFI that function has calls.
12240 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
12241 MFI->setAdjustsStack(true);
12243 // And our return value (tls address) is in the standard call return value
12245 unsigned Reg = Subtarget->is64Bit() ? X86::RAX : X86::EAX;
12246 return DAG.getCopyFromReg(Chain, DL, Reg, PtrVT, Chain.getValue(1));
12249 if (Subtarget->isTargetKnownWindowsMSVC() ||
12250 Subtarget->isTargetWindowsGNU()) {
12251 // Just use the implicit TLS architecture
12252 // Need to generate someting similar to:
12253 // mov rdx, qword [gs:abs 58H]; Load pointer to ThreadLocalStorage
12255 // mov ecx, dword [rel _tls_index]: Load index (from C runtime)
12256 // mov rcx, qword [rdx+rcx*8]
12257 // mov eax, .tls$:tlsvar
12258 // [rax+rcx] contains the address
12259 // Windows 64bit: gs:0x58
12260 // Windows 32bit: fs:__tls_array
12263 SDValue Chain = DAG.getEntryNode();
12265 // Get the Thread Pointer, which is %fs:__tls_array (32-bit) or
12266 // %gs:0x58 (64-bit). On MinGW, __tls_array is not available, so directly
12267 // use its literal value of 0x2C.
12268 Value *Ptr = Constant::getNullValue(Subtarget->is64Bit()
12269 ? Type::getInt8PtrTy(*DAG.getContext(),
12271 : Type::getInt32PtrTy(*DAG.getContext(),
12274 SDValue TlsArray = Subtarget->is64Bit()
12275 ? DAG.getIntPtrConstant(0x58, dl)
12276 : (Subtarget->isTargetWindowsGNU()
12277 ? DAG.getIntPtrConstant(0x2C, dl)
12278 : DAG.getExternalSymbol("_tls_array", PtrVT));
12280 SDValue ThreadPointer =
12281 DAG.getLoad(PtrVT, dl, Chain, TlsArray, MachinePointerInfo(Ptr), false,
12285 if (GV->getThreadLocalMode() == GlobalVariable::LocalExecTLSModel) {
12286 res = ThreadPointer;
12288 // Load the _tls_index variable
12289 SDValue IDX = DAG.getExternalSymbol("_tls_index", PtrVT);
12290 if (Subtarget->is64Bit())
12291 IDX = DAG.getExtLoad(ISD::ZEXTLOAD, dl, PtrVT, Chain, IDX,
12292 MachinePointerInfo(), MVT::i32, false, false,
12295 IDX = DAG.getLoad(PtrVT, dl, Chain, IDX, MachinePointerInfo(), false,
12298 auto &DL = DAG.getDataLayout();
12300 DAG.getConstant(Log2_64_Ceil(DL.getPointerSize()), dl, PtrVT);
12301 IDX = DAG.getNode(ISD::SHL, dl, PtrVT, IDX, Scale);
12303 res = DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, IDX);
12306 res = DAG.getLoad(PtrVT, dl, Chain, res, MachinePointerInfo(), false, false,
12309 // Get the offset of start of .tls section
12310 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(), dl,
12311 GA->getValueType(0),
12312 GA->getOffset(), X86II::MO_SECREL);
12313 SDValue Offset = DAG.getNode(X86ISD::Wrapper, dl, PtrVT, TGA);
12315 // The address of the thread local variable is the add of the thread
12316 // pointer with the offset of the variable.
12317 return DAG.getNode(ISD::ADD, dl, PtrVT, res, Offset);
12320 llvm_unreachable("TLS not implemented for this target.");
12323 /// LowerShiftParts - Lower SRA_PARTS and friends, which return two i32 values
12324 /// and take a 2 x i32 value to shift plus a shift amount.
12325 static SDValue LowerShiftParts(SDValue Op, SelectionDAG &DAG) {
12326 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
12327 MVT VT = Op.getSimpleValueType();
12328 unsigned VTBits = VT.getSizeInBits();
12330 bool isSRA = Op.getOpcode() == ISD::SRA_PARTS;
12331 SDValue ShOpLo = Op.getOperand(0);
12332 SDValue ShOpHi = Op.getOperand(1);
12333 SDValue ShAmt = Op.getOperand(2);
12334 // X86ISD::SHLD and X86ISD::SHRD have defined overflow behavior but the
12335 // generic ISD nodes haven't. Insert an AND to be safe, it's optimized away
12337 SDValue SafeShAmt = DAG.getNode(ISD::AND, dl, MVT::i8, ShAmt,
12338 DAG.getConstant(VTBits - 1, dl, MVT::i8));
12339 SDValue Tmp1 = isSRA ? DAG.getNode(ISD::SRA, dl, VT, ShOpHi,
12340 DAG.getConstant(VTBits - 1, dl, MVT::i8))
12341 : DAG.getConstant(0, dl, VT);
12343 SDValue Tmp2, Tmp3;
12344 if (Op.getOpcode() == ISD::SHL_PARTS) {
12345 Tmp2 = DAG.getNode(X86ISD::SHLD, dl, VT, ShOpHi, ShOpLo, ShAmt);
12346 Tmp3 = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, SafeShAmt);
12348 Tmp2 = DAG.getNode(X86ISD::SHRD, dl, VT, ShOpLo, ShOpHi, ShAmt);
12349 Tmp3 = DAG.getNode(isSRA ? ISD::SRA : ISD::SRL, dl, VT, ShOpHi, SafeShAmt);
12352 // If the shift amount is larger or equal than the width of a part we can't
12353 // rely on the results of shld/shrd. Insert a test and select the appropriate
12354 // values for large shift amounts.
12355 SDValue AndNode = DAG.getNode(ISD::AND, dl, MVT::i8, ShAmt,
12356 DAG.getConstant(VTBits, dl, MVT::i8));
12357 SDValue Cond = DAG.getNode(X86ISD::CMP, dl, MVT::i32,
12358 AndNode, DAG.getConstant(0, dl, MVT::i8));
12361 SDValue CC = DAG.getConstant(X86::COND_NE, dl, MVT::i8);
12362 SDValue Ops0[4] = { Tmp2, Tmp3, CC, Cond };
12363 SDValue Ops1[4] = { Tmp3, Tmp1, CC, Cond };
12365 if (Op.getOpcode() == ISD::SHL_PARTS) {
12366 Hi = DAG.getNode(X86ISD::CMOV, dl, VT, Ops0);
12367 Lo = DAG.getNode(X86ISD::CMOV, dl, VT, Ops1);
12369 Lo = DAG.getNode(X86ISD::CMOV, dl, VT, Ops0);
12370 Hi = DAG.getNode(X86ISD::CMOV, dl, VT, Ops1);
12373 SDValue Ops[2] = { Lo, Hi };
12374 return DAG.getMergeValues(Ops, dl);
12377 SDValue X86TargetLowering::LowerSINT_TO_FP(SDValue Op,
12378 SelectionDAG &DAG) const {
12379 SDValue Src = Op.getOperand(0);
12380 MVT SrcVT = Src.getSimpleValueType();
12381 MVT VT = Op.getSimpleValueType();
12384 if (SrcVT.isVector()) {
12385 if (SrcVT == MVT::v2i32 && VT == MVT::v2f64) {
12386 return DAG.getNode(X86ISD::CVTDQ2PD, dl, VT,
12387 DAG.getNode(ISD::CONCAT_VECTORS, dl, MVT::v4i32, Src,
12388 DAG.getUNDEF(SrcVT)));
12390 if (SrcVT.getVectorElementType() == MVT::i1) {
12391 MVT IntegerVT = MVT::getVectorVT(MVT::i32, SrcVT.getVectorNumElements());
12392 return DAG.getNode(ISD::SINT_TO_FP, dl, Op.getValueType(),
12393 DAG.getNode(ISD::SIGN_EXTEND, dl, IntegerVT, Src));
12398 assert(SrcVT <= MVT::i64 && SrcVT >= MVT::i16 &&
12399 "Unknown SINT_TO_FP to lower!");
12401 // These are really Legal; return the operand so the caller accepts it as
12403 if (SrcVT == MVT::i32 && isScalarFPTypeInSSEReg(Op.getValueType()))
12405 if (SrcVT == MVT::i64 && isScalarFPTypeInSSEReg(Op.getValueType()) &&
12406 Subtarget->is64Bit()) {
12410 unsigned Size = SrcVT.getSizeInBits()/8;
12411 MachineFunction &MF = DAG.getMachineFunction();
12412 auto PtrVT = getPointerTy(MF.getDataLayout());
12413 int SSFI = MF.getFrameInfo()->CreateStackObject(Size, Size, false);
12414 SDValue StackSlot = DAG.getFrameIndex(SSFI, PtrVT);
12415 SDValue Chain = DAG.getStore(
12416 DAG.getEntryNode(), dl, Op.getOperand(0), StackSlot,
12417 MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), SSFI), false,
12419 return BuildFILD(Op, SrcVT, Chain, StackSlot, DAG);
12422 SDValue X86TargetLowering::BuildFILD(SDValue Op, EVT SrcVT, SDValue Chain,
12424 SelectionDAG &DAG) const {
12428 bool useSSE = isScalarFPTypeInSSEReg(Op.getValueType());
12430 Tys = DAG.getVTList(MVT::f64, MVT::Other, MVT::Glue);
12432 Tys = DAG.getVTList(Op.getValueType(), MVT::Other);
12434 unsigned ByteSize = SrcVT.getSizeInBits()/8;
12436 FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(StackSlot);
12437 MachineMemOperand *MMO;
12439 int SSFI = FI->getIndex();
12440 MMO = DAG.getMachineFunction().getMachineMemOperand(
12441 MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), SSFI),
12442 MachineMemOperand::MOLoad, ByteSize, ByteSize);
12444 MMO = cast<LoadSDNode>(StackSlot)->getMemOperand();
12445 StackSlot = StackSlot.getOperand(1);
12447 SDValue Ops[] = { Chain, StackSlot, DAG.getValueType(SrcVT) };
12448 SDValue Result = DAG.getMemIntrinsicNode(useSSE ? X86ISD::FILD_FLAG :
12450 Tys, Ops, SrcVT, MMO);
12453 Chain = Result.getValue(1);
12454 SDValue InFlag = Result.getValue(2);
12456 // FIXME: Currently the FST is flagged to the FILD_FLAG. This
12457 // shouldn't be necessary except that RFP cannot be live across
12458 // multiple blocks. When stackifier is fixed, they can be uncoupled.
12459 MachineFunction &MF = DAG.getMachineFunction();
12460 unsigned SSFISize = Op.getValueType().getSizeInBits()/8;
12461 int SSFI = MF.getFrameInfo()->CreateStackObject(SSFISize, SSFISize, false);
12462 auto PtrVT = getPointerTy(MF.getDataLayout());
12463 SDValue StackSlot = DAG.getFrameIndex(SSFI, PtrVT);
12464 Tys = DAG.getVTList(MVT::Other);
12466 Chain, Result, StackSlot, DAG.getValueType(Op.getValueType()), InFlag
12468 MachineMemOperand *MMO = DAG.getMachineFunction().getMachineMemOperand(
12469 MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), SSFI),
12470 MachineMemOperand::MOStore, SSFISize, SSFISize);
12472 Chain = DAG.getMemIntrinsicNode(X86ISD::FST, DL, Tys,
12473 Ops, Op.getValueType(), MMO);
12474 Result = DAG.getLoad(
12475 Op.getValueType(), DL, Chain, StackSlot,
12476 MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), SSFI),
12477 false, false, false, 0);
12483 // LowerUINT_TO_FP_i64 - 64-bit unsigned integer to double expansion.
12484 SDValue X86TargetLowering::LowerUINT_TO_FP_i64(SDValue Op,
12485 SelectionDAG &DAG) const {
12486 // This algorithm is not obvious. Here it is what we're trying to output:
12489 punpckldq (c0), %xmm0 // c0: (uint4){ 0x43300000U, 0x45300000U, 0U, 0U }
12490 subpd (c1), %xmm0 // c1: (double2){ 0x1.0p52, 0x1.0p52 * 0x1.0p32 }
12492 haddpd %xmm0, %xmm0
12494 pshufd $0x4e, %xmm0, %xmm1
12500 LLVMContext *Context = DAG.getContext();
12502 // Build some magic constants.
12503 static const uint32_t CV0[] = { 0x43300000, 0x45300000, 0, 0 };
12504 Constant *C0 = ConstantDataVector::get(*Context, CV0);
12505 auto PtrVT = getPointerTy(DAG.getDataLayout());
12506 SDValue CPIdx0 = DAG.getConstantPool(C0, PtrVT, 16);
12508 SmallVector<Constant*,2> CV1;
12510 ConstantFP::get(*Context, APFloat(APFloat::IEEEdouble,
12511 APInt(64, 0x4330000000000000ULL))));
12513 ConstantFP::get(*Context, APFloat(APFloat::IEEEdouble,
12514 APInt(64, 0x4530000000000000ULL))));
12515 Constant *C1 = ConstantVector::get(CV1);
12516 SDValue CPIdx1 = DAG.getConstantPool(C1, PtrVT, 16);
12518 // Load the 64-bit value into an XMM register.
12519 SDValue XR1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i64,
12522 DAG.getLoad(MVT::v4i32, dl, DAG.getEntryNode(), CPIdx0,
12523 MachinePointerInfo::getConstantPool(DAG.getMachineFunction()),
12524 false, false, false, 16);
12526 getUnpackl(DAG, dl, MVT::v4i32, DAG.getBitcast(MVT::v4i32, XR1), CLod0);
12529 DAG.getLoad(MVT::v2f64, dl, CLod0.getValue(1), CPIdx1,
12530 MachinePointerInfo::getConstantPool(DAG.getMachineFunction()),
12531 false, false, false, 16);
12532 SDValue XR2F = DAG.getBitcast(MVT::v2f64, Unpck1);
12533 // TODO: Are there any fast-math-flags to propagate here?
12534 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::v2f64, XR2F, CLod1);
12537 if (Subtarget->hasSSE3()) {
12538 // FIXME: The 'haddpd' instruction may be slower than 'movhlps + addsd'.
12539 Result = DAG.getNode(X86ISD::FHADD, dl, MVT::v2f64, Sub, Sub);
12541 SDValue S2F = DAG.getBitcast(MVT::v4i32, Sub);
12542 SDValue Shuffle = getTargetShuffleNode(X86ISD::PSHUFD, dl, MVT::v4i32,
12544 Result = DAG.getNode(ISD::FADD, dl, MVT::v2f64,
12545 DAG.getBitcast(MVT::v2f64, Shuffle), Sub);
12548 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Result,
12549 DAG.getIntPtrConstant(0, dl));
12552 // LowerUINT_TO_FP_i32 - 32-bit unsigned integer to float expansion.
12553 SDValue X86TargetLowering::LowerUINT_TO_FP_i32(SDValue Op,
12554 SelectionDAG &DAG) const {
12556 // FP constant to bias correct the final result.
12557 SDValue Bias = DAG.getConstantFP(BitsToDouble(0x4330000000000000ULL), dl,
12560 // Load the 32-bit value into an XMM register.
12561 SDValue Load = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
12564 // Zero out the upper parts of the register.
12565 Load = getShuffleVectorZeroOrUndef(Load, 0, true, Subtarget, DAG);
12567 Load = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
12568 DAG.getBitcast(MVT::v2f64, Load),
12569 DAG.getIntPtrConstant(0, dl));
12571 // Or the load with the bias.
12572 SDValue Or = DAG.getNode(
12573 ISD::OR, dl, MVT::v2i64,
12574 DAG.getBitcast(MVT::v2i64,
12575 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2f64, Load)),
12576 DAG.getBitcast(MVT::v2i64,
12577 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2f64, Bias)));
12579 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
12580 DAG.getBitcast(MVT::v2f64, Or), DAG.getIntPtrConstant(0, dl));
12582 // Subtract the bias.
12583 // TODO: Are there any fast-math-flags to propagate here?
12584 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::f64, Or, Bias);
12586 // Handle final rounding.
12587 EVT DestVT = Op.getValueType();
12589 if (DestVT.bitsLT(MVT::f64))
12590 return DAG.getNode(ISD::FP_ROUND, dl, DestVT, Sub,
12591 DAG.getIntPtrConstant(0, dl));
12592 if (DestVT.bitsGT(MVT::f64))
12593 return DAG.getNode(ISD::FP_EXTEND, dl, DestVT, Sub);
12595 // Handle final rounding.
12599 static SDValue lowerUINT_TO_FP_vXi32(SDValue Op, SelectionDAG &DAG,
12600 const X86Subtarget &Subtarget) {
12601 // The algorithm is the following:
12602 // #ifdef __SSE4_1__
12603 // uint4 lo = _mm_blend_epi16( v, (uint4) 0x4b000000, 0xaa);
12604 // uint4 hi = _mm_blend_epi16( _mm_srli_epi32(v,16),
12605 // (uint4) 0x53000000, 0xaa);
12607 // uint4 lo = (v & (uint4) 0xffff) | (uint4) 0x4b000000;
12608 // uint4 hi = (v >> 16) | (uint4) 0x53000000;
12610 // float4 fhi = (float4) hi - (0x1.0p39f + 0x1.0p23f);
12611 // return (float4) lo + fhi;
12613 // We shouldn't use it when unsafe-fp-math is enabled though: we might later
12614 // reassociate the two FADDs, and if we do that, the algorithm fails
12615 // spectacularly (PR24512).
12616 // FIXME: If we ever have some kind of Machine FMF, this should be marked
12617 // as non-fast and always be enabled. Why isn't SDAG FMF enough? Because
12618 // there's also the MachineCombiner reassociations happening on Machine IR.
12619 if (DAG.getTarget().Options.UnsafeFPMath)
12623 SDValue V = Op->getOperand(0);
12624 EVT VecIntVT = V.getValueType();
12625 bool Is128 = VecIntVT == MVT::v4i32;
12626 EVT VecFloatVT = Is128 ? MVT::v4f32 : MVT::v8f32;
12627 // If we convert to something else than the supported type, e.g., to v4f64,
12629 if (VecFloatVT != Op->getValueType(0))
12632 unsigned NumElts = VecIntVT.getVectorNumElements();
12633 assert((VecIntVT == MVT::v4i32 || VecIntVT == MVT::v8i32) &&
12634 "Unsupported custom type");
12635 assert(NumElts <= 8 && "The size of the constant array must be fixed");
12637 // In the #idef/#else code, we have in common:
12638 // - The vector of constants:
12644 // Create the splat vector for 0x4b000000.
12645 SDValue CstLow = DAG.getConstant(0x4b000000, DL, MVT::i32);
12646 SDValue CstLowArray[] = {CstLow, CstLow, CstLow, CstLow,
12647 CstLow, CstLow, CstLow, CstLow};
12648 SDValue VecCstLow = DAG.getNode(ISD::BUILD_VECTOR, DL, VecIntVT,
12649 makeArrayRef(&CstLowArray[0], NumElts));
12650 // Create the splat vector for 0x53000000.
12651 SDValue CstHigh = DAG.getConstant(0x53000000, DL, MVT::i32);
12652 SDValue CstHighArray[] = {CstHigh, CstHigh, CstHigh, CstHigh,
12653 CstHigh, CstHigh, CstHigh, CstHigh};
12654 SDValue VecCstHigh = DAG.getNode(ISD::BUILD_VECTOR, DL, VecIntVT,
12655 makeArrayRef(&CstHighArray[0], NumElts));
12657 // Create the right shift.
12658 SDValue CstShift = DAG.getConstant(16, DL, MVT::i32);
12659 SDValue CstShiftArray[] = {CstShift, CstShift, CstShift, CstShift,
12660 CstShift, CstShift, CstShift, CstShift};
12661 SDValue VecCstShift = DAG.getNode(ISD::BUILD_VECTOR, DL, VecIntVT,
12662 makeArrayRef(&CstShiftArray[0], NumElts));
12663 SDValue HighShift = DAG.getNode(ISD::SRL, DL, VecIntVT, V, VecCstShift);
12666 if (Subtarget.hasSSE41()) {
12667 EVT VecI16VT = Is128 ? MVT::v8i16 : MVT::v16i16;
12668 // uint4 lo = _mm_blend_epi16( v, (uint4) 0x4b000000, 0xaa);
12669 SDValue VecCstLowBitcast = DAG.getBitcast(VecI16VT, VecCstLow);
12670 SDValue VecBitcast = DAG.getBitcast(VecI16VT, V);
12671 // Low will be bitcasted right away, so do not bother bitcasting back to its
12673 Low = DAG.getNode(X86ISD::BLENDI, DL, VecI16VT, VecBitcast,
12674 VecCstLowBitcast, DAG.getConstant(0xaa, DL, MVT::i32));
12675 // uint4 hi = _mm_blend_epi16( _mm_srli_epi32(v,16),
12676 // (uint4) 0x53000000, 0xaa);
12677 SDValue VecCstHighBitcast = DAG.getBitcast(VecI16VT, VecCstHigh);
12678 SDValue VecShiftBitcast = DAG.getBitcast(VecI16VT, HighShift);
12679 // High will be bitcasted right away, so do not bother bitcasting back to
12680 // its original type.
12681 High = DAG.getNode(X86ISD::BLENDI, DL, VecI16VT, VecShiftBitcast,
12682 VecCstHighBitcast, DAG.getConstant(0xaa, DL, MVT::i32));
12684 SDValue CstMask = DAG.getConstant(0xffff, DL, MVT::i32);
12685 SDValue VecCstMask = DAG.getNode(ISD::BUILD_VECTOR, DL, VecIntVT, CstMask,
12686 CstMask, CstMask, CstMask);
12687 // uint4 lo = (v & (uint4) 0xffff) | (uint4) 0x4b000000;
12688 SDValue LowAnd = DAG.getNode(ISD::AND, DL, VecIntVT, V, VecCstMask);
12689 Low = DAG.getNode(ISD::OR, DL, VecIntVT, LowAnd, VecCstLow);
12691 // uint4 hi = (v >> 16) | (uint4) 0x53000000;
12692 High = DAG.getNode(ISD::OR, DL, VecIntVT, HighShift, VecCstHigh);
12695 // Create the vector constant for -(0x1.0p39f + 0x1.0p23f).
12696 SDValue CstFAdd = DAG.getConstantFP(
12697 APFloat(APFloat::IEEEsingle, APInt(32, 0xD3000080)), DL, MVT::f32);
12698 SDValue CstFAddArray[] = {CstFAdd, CstFAdd, CstFAdd, CstFAdd,
12699 CstFAdd, CstFAdd, CstFAdd, CstFAdd};
12700 SDValue VecCstFAdd = DAG.getNode(ISD::BUILD_VECTOR, DL, VecFloatVT,
12701 makeArrayRef(&CstFAddArray[0], NumElts));
12703 // float4 fhi = (float4) hi - (0x1.0p39f + 0x1.0p23f);
12704 SDValue HighBitcast = DAG.getBitcast(VecFloatVT, High);
12705 // TODO: Are there any fast-math-flags to propagate here?
12707 DAG.getNode(ISD::FADD, DL, VecFloatVT, HighBitcast, VecCstFAdd);
12708 // return (float4) lo + fhi;
12709 SDValue LowBitcast = DAG.getBitcast(VecFloatVT, Low);
12710 return DAG.getNode(ISD::FADD, DL, VecFloatVT, LowBitcast, FHigh);
12713 SDValue X86TargetLowering::lowerUINT_TO_FP_vec(SDValue Op,
12714 SelectionDAG &DAG) const {
12715 SDValue N0 = Op.getOperand(0);
12716 MVT SVT = N0.getSimpleValueType();
12719 switch (SVT.SimpleTy) {
12721 llvm_unreachable("Custom UINT_TO_FP is not supported!");
12726 MVT NVT = MVT::getVectorVT(MVT::i32, SVT.getVectorNumElements());
12727 return DAG.getNode(ISD::SINT_TO_FP, dl, Op.getValueType(),
12728 DAG.getNode(ISD::ZERO_EXTEND, dl, NVT, N0));
12732 return lowerUINT_TO_FP_vXi32(Op, DAG, *Subtarget);
12735 if (Subtarget->hasAVX512())
12736 return DAG.getNode(ISD::UINT_TO_FP, dl, Op.getValueType(),
12737 DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v16i32, N0));
12739 llvm_unreachable(nullptr);
12742 SDValue X86TargetLowering::LowerUINT_TO_FP(SDValue Op,
12743 SelectionDAG &DAG) const {
12744 SDValue N0 = Op.getOperand(0);
12746 auto PtrVT = getPointerTy(DAG.getDataLayout());
12748 if (Op.getValueType().isVector())
12749 return lowerUINT_TO_FP_vec(Op, DAG);
12751 // Since UINT_TO_FP is legal (it's marked custom), dag combiner won't
12752 // optimize it to a SINT_TO_FP when the sign bit is known zero. Perform
12753 // the optimization here.
12754 if (DAG.SignBitIsZero(N0))
12755 return DAG.getNode(ISD::SINT_TO_FP, dl, Op.getValueType(), N0);
12757 MVT SrcVT = N0.getSimpleValueType();
12758 MVT DstVT = Op.getSimpleValueType();
12760 if (Subtarget->hasAVX512() && isScalarFPTypeInSSEReg(DstVT) &&
12761 (SrcVT == MVT::i32 || (SrcVT == MVT::i64 && Subtarget->is64Bit()))) {
12762 // Conversions from unsigned i32 to f32/f64 are legal,
12763 // using VCVTUSI2SS/SD. Same for i64 in 64-bit mode.
12767 if (SrcVT == MVT::i64 && DstVT == MVT::f64 && X86ScalarSSEf64)
12768 return LowerUINT_TO_FP_i64(Op, DAG);
12769 if (SrcVT == MVT::i32 && X86ScalarSSEf64)
12770 return LowerUINT_TO_FP_i32(Op, DAG);
12771 if (Subtarget->is64Bit() && SrcVT == MVT::i64 && DstVT == MVT::f32)
12774 // Make a 64-bit buffer, and use it to build an FILD.
12775 SDValue StackSlot = DAG.CreateStackTemporary(MVT::i64);
12776 if (SrcVT == MVT::i32) {
12777 SDValue WordOff = DAG.getConstant(4, dl, PtrVT);
12778 SDValue OffsetSlot = DAG.getNode(ISD::ADD, dl, PtrVT, StackSlot, WordOff);
12779 SDValue Store1 = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
12780 StackSlot, MachinePointerInfo(),
12782 SDValue Store2 = DAG.getStore(Store1, dl, DAG.getConstant(0, dl, MVT::i32),
12783 OffsetSlot, MachinePointerInfo(),
12785 SDValue Fild = BuildFILD(Op, MVT::i64, Store2, StackSlot, DAG);
12789 assert(SrcVT == MVT::i64 && "Unexpected type in UINT_TO_FP");
12790 SDValue Store = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
12791 StackSlot, MachinePointerInfo(),
12793 // For i64 source, we need to add the appropriate power of 2 if the input
12794 // was negative. This is the same as the optimization in
12795 // DAGTypeLegalizer::ExpandIntOp_UNIT_TO_FP, and for it to be safe here,
12796 // we must be careful to do the computation in x87 extended precision, not
12797 // in SSE. (The generic code can't know it's OK to do this, or how to.)
12798 int SSFI = cast<FrameIndexSDNode>(StackSlot)->getIndex();
12799 MachineMemOperand *MMO = DAG.getMachineFunction().getMachineMemOperand(
12800 MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), SSFI),
12801 MachineMemOperand::MOLoad, 8, 8);
12803 SDVTList Tys = DAG.getVTList(MVT::f80, MVT::Other);
12804 SDValue Ops[] = { Store, StackSlot, DAG.getValueType(MVT::i64) };
12805 SDValue Fild = DAG.getMemIntrinsicNode(X86ISD::FILD, dl, Tys, Ops,
12808 APInt FF(32, 0x5F800000ULL);
12810 // Check whether the sign bit is set.
12811 SDValue SignSet = DAG.getSetCC(
12812 dl, getSetCCResultType(DAG.getDataLayout(), *DAG.getContext(), MVT::i64),
12813 Op.getOperand(0), DAG.getConstant(0, dl, MVT::i64), ISD::SETLT);
12815 // Build a 64 bit pair (0, FF) in the constant pool, with FF in the lo bits.
12816 SDValue FudgePtr = DAG.getConstantPool(
12817 ConstantInt::get(*DAG.getContext(), FF.zext(64)), PtrVT);
12819 // Get a pointer to FF if the sign bit was set, or to 0 otherwise.
12820 SDValue Zero = DAG.getIntPtrConstant(0, dl);
12821 SDValue Four = DAG.getIntPtrConstant(4, dl);
12822 SDValue Offset = DAG.getNode(ISD::SELECT, dl, Zero.getValueType(), SignSet,
12824 FudgePtr = DAG.getNode(ISD::ADD, dl, PtrVT, FudgePtr, Offset);
12826 // Load the value out, extending it from f32 to f80.
12827 // FIXME: Avoid the extend by constructing the right constant pool?
12828 SDValue Fudge = DAG.getExtLoad(
12829 ISD::EXTLOAD, dl, MVT::f80, DAG.getEntryNode(), FudgePtr,
12830 MachinePointerInfo::getConstantPool(DAG.getMachineFunction()), MVT::f32,
12831 false, false, false, 4);
12832 // Extend everything to 80 bits to force it to be done on x87.
12833 // TODO: Are there any fast-math-flags to propagate here?
12834 SDValue Add = DAG.getNode(ISD::FADD, dl, MVT::f80, Fild, Fudge);
12835 return DAG.getNode(ISD::FP_ROUND, dl, DstVT, Add,
12836 DAG.getIntPtrConstant(0, dl));
12839 // If the given FP_TO_SINT (IsSigned) or FP_TO_UINT (!IsSigned) operation
12840 // is legal, or has an fp128 or f16 source (which needs to be promoted to f32),
12841 // just return an <SDValue(), SDValue()> pair.
12842 // Otherwise it is assumed to be a conversion from one of f32, f64 or f80
12843 // to i16, i32 or i64, and we lower it to a legal sequence.
12844 // If lowered to the final integer result we return a <result, SDValue()> pair.
12845 // Otherwise we lower it to a sequence ending with a FIST, return a
12846 // <FIST, StackSlot> pair, and the caller is responsible for loading
12847 // the final integer result from StackSlot.
12848 std::pair<SDValue,SDValue>
12849 X86TargetLowering::FP_TO_INTHelper(SDValue Op, SelectionDAG &DAG,
12850 bool IsSigned, bool IsReplace) const {
12853 EVT DstTy = Op.getValueType();
12854 EVT TheVT = Op.getOperand(0).getValueType();
12855 auto PtrVT = getPointerTy(DAG.getDataLayout());
12857 if (TheVT != MVT::f32 && TheVT != MVT::f64 && TheVT != MVT::f80) {
12858 // f16 must be promoted before using the lowering in this routine.
12859 // fp128 does not use this lowering.
12860 return std::make_pair(SDValue(), SDValue());
12863 // If using FIST to compute an unsigned i64, we'll need some fixup
12864 // to handle values above the maximum signed i64. A FIST is always
12865 // used for the 32-bit subtarget, but also for f80 on a 64-bit target.
12866 bool UnsignedFixup = !IsSigned &&
12867 DstTy == MVT::i64 &&
12868 (!Subtarget->is64Bit() ||
12869 !isScalarFPTypeInSSEReg(TheVT));
12871 if (!IsSigned && DstTy != MVT::i64 && !Subtarget->hasAVX512()) {
12872 // Replace the fp-to-uint32 operation with an fp-to-sint64 FIST.
12873 // The low 32 bits of the fist result will have the correct uint32 result.
12874 assert(DstTy == MVT::i32 && "Unexpected FP_TO_UINT");
12878 assert(DstTy.getSimpleVT() <= MVT::i64 &&
12879 DstTy.getSimpleVT() >= MVT::i16 &&
12880 "Unknown FP_TO_INT to lower!");
12882 // These are really Legal.
12883 if (DstTy == MVT::i32 &&
12884 isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType()))
12885 return std::make_pair(SDValue(), SDValue());
12886 if (Subtarget->is64Bit() &&
12887 DstTy == MVT::i64 &&
12888 isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType()))
12889 return std::make_pair(SDValue(), SDValue());
12891 // We lower FP->int64 into FISTP64 followed by a load from a temporary
12893 MachineFunction &MF = DAG.getMachineFunction();
12894 unsigned MemSize = DstTy.getSizeInBits()/8;
12895 int SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize, false);
12896 SDValue StackSlot = DAG.getFrameIndex(SSFI, PtrVT);
12899 switch (DstTy.getSimpleVT().SimpleTy) {
12900 default: llvm_unreachable("Invalid FP_TO_SINT to lower!");
12901 case MVT::i16: Opc = X86ISD::FP_TO_INT16_IN_MEM; break;
12902 case MVT::i32: Opc = X86ISD::FP_TO_INT32_IN_MEM; break;
12903 case MVT::i64: Opc = X86ISD::FP_TO_INT64_IN_MEM; break;
12906 SDValue Chain = DAG.getEntryNode();
12907 SDValue Value = Op.getOperand(0);
12908 SDValue Adjust; // 0x0 or 0x80000000, for result sign bit adjustment.
12910 if (UnsignedFixup) {
12912 // Conversion to unsigned i64 is implemented with a select,
12913 // depending on whether the source value fits in the range
12914 // of a signed i64. Let Thresh be the FP equivalent of
12915 // 0x8000000000000000ULL.
12917 // Adjust i32 = (Value < Thresh) ? 0 : 0x80000000;
12918 // FistSrc = (Value < Thresh) ? Value : (Value - Thresh);
12919 // Fist-to-mem64 FistSrc
12920 // Add 0 or 0x800...0ULL to the 64-bit result, which is equivalent
12921 // to XOR'ing the high 32 bits with Adjust.
12923 // Being a power of 2, Thresh is exactly representable in all FP formats.
12924 // For X87 we'd like to use the smallest FP type for this constant, but
12925 // for DAG type consistency we have to match the FP operand type.
12927 APFloat Thresh(APFloat::IEEEsingle, APInt(32, 0x5f000000));
12928 LLVM_ATTRIBUTE_UNUSED APFloat::opStatus Status = APFloat::opOK;
12929 bool LosesInfo = false;
12930 if (TheVT == MVT::f64)
12931 // The rounding mode is irrelevant as the conversion should be exact.
12932 Status = Thresh.convert(APFloat::IEEEdouble, APFloat::rmNearestTiesToEven,
12934 else if (TheVT == MVT::f80)
12935 Status = Thresh.convert(APFloat::x87DoubleExtended,
12936 APFloat::rmNearestTiesToEven, &LosesInfo);
12938 assert(Status == APFloat::opOK && !LosesInfo &&
12939 "FP conversion should have been exact");
12941 SDValue ThreshVal = DAG.getConstantFP(Thresh, DL, TheVT);
12943 SDValue Cmp = DAG.getSetCC(DL,
12944 getSetCCResultType(DAG.getDataLayout(),
12945 *DAG.getContext(), TheVT),
12946 Value, ThreshVal, ISD::SETLT);
12947 Adjust = DAG.getSelect(DL, MVT::i32, Cmp,
12948 DAG.getConstant(0, DL, MVT::i32),
12949 DAG.getConstant(0x80000000, DL, MVT::i32));
12950 SDValue Sub = DAG.getNode(ISD::FSUB, DL, TheVT, Value, ThreshVal);
12951 Cmp = DAG.getSetCC(DL, getSetCCResultType(DAG.getDataLayout(),
12952 *DAG.getContext(), TheVT),
12953 Value, ThreshVal, ISD::SETLT);
12954 Value = DAG.getSelect(DL, TheVT, Cmp, Value, Sub);
12957 // FIXME This causes a redundant load/store if the SSE-class value is already
12958 // in memory, such as if it is on the callstack.
12959 if (isScalarFPTypeInSSEReg(TheVT)) {
12960 assert(DstTy == MVT::i64 && "Invalid FP_TO_SINT to lower!");
12961 Chain = DAG.getStore(Chain, DL, Value, StackSlot,
12962 MachinePointerInfo::getFixedStack(MF, SSFI), false,
12964 SDVTList Tys = DAG.getVTList(Op.getOperand(0).getValueType(), MVT::Other);
12966 Chain, StackSlot, DAG.getValueType(TheVT)
12969 MachineMemOperand *MMO =
12970 MF.getMachineMemOperand(MachinePointerInfo::getFixedStack(MF, SSFI),
12971 MachineMemOperand::MOLoad, MemSize, MemSize);
12972 Value = DAG.getMemIntrinsicNode(X86ISD::FLD, DL, Tys, Ops, DstTy, MMO);
12973 Chain = Value.getValue(1);
12974 SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize, false);
12975 StackSlot = DAG.getFrameIndex(SSFI, PtrVT);
12978 MachineMemOperand *MMO =
12979 MF.getMachineMemOperand(MachinePointerInfo::getFixedStack(MF, SSFI),
12980 MachineMemOperand::MOStore, MemSize, MemSize);
12982 if (UnsignedFixup) {
12984 // Insert the FIST, load its result as two i32's,
12985 // and XOR the high i32 with Adjust.
12987 SDValue FistOps[] = { Chain, Value, StackSlot };
12988 SDValue FIST = DAG.getMemIntrinsicNode(Opc, DL, DAG.getVTList(MVT::Other),
12989 FistOps, DstTy, MMO);
12991 SDValue Low32 = DAG.getLoad(MVT::i32, DL, FIST, StackSlot,
12992 MachinePointerInfo(),
12993 false, false, false, 0);
12994 SDValue HighAddr = DAG.getNode(ISD::ADD, DL, PtrVT, StackSlot,
12995 DAG.getConstant(4, DL, PtrVT));
12997 SDValue High32 = DAG.getLoad(MVT::i32, DL, FIST, HighAddr,
12998 MachinePointerInfo(),
12999 false, false, false, 0);
13000 High32 = DAG.getNode(ISD::XOR, DL, MVT::i32, High32, Adjust);
13002 if (Subtarget->is64Bit()) {
13003 // Join High32 and Low32 into a 64-bit result.
13004 // (High32 << 32) | Low32
13005 Low32 = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::i64, Low32);
13006 High32 = DAG.getNode(ISD::ANY_EXTEND, DL, MVT::i64, High32);
13007 High32 = DAG.getNode(ISD::SHL, DL, MVT::i64, High32,
13008 DAG.getConstant(32, DL, MVT::i8));
13009 SDValue Result = DAG.getNode(ISD::OR, DL, MVT::i64, High32, Low32);
13010 return std::make_pair(Result, SDValue());
13013 SDValue ResultOps[] = { Low32, High32 };
13015 SDValue pair = IsReplace
13016 ? DAG.getNode(ISD::BUILD_PAIR, DL, MVT::i64, ResultOps)
13017 : DAG.getMergeValues(ResultOps, DL);
13018 return std::make_pair(pair, SDValue());
13020 // Build the FP_TO_INT*_IN_MEM
13021 SDValue Ops[] = { Chain, Value, StackSlot };
13022 SDValue FIST = DAG.getMemIntrinsicNode(Opc, DL, DAG.getVTList(MVT::Other),
13024 return std::make_pair(FIST, StackSlot);
13028 static SDValue LowerAVXExtend(SDValue Op, SelectionDAG &DAG,
13029 const X86Subtarget *Subtarget) {
13030 MVT VT = Op->getSimpleValueType(0);
13031 SDValue In = Op->getOperand(0);
13032 MVT InVT = In.getSimpleValueType();
13035 if (VT.is512BitVector() || InVT.getScalarType() == MVT::i1)
13036 return DAG.getNode(ISD::ZERO_EXTEND, dl, VT, In);
13038 // Optimize vectors in AVX mode:
13041 // Use vpunpcklwd for 4 lower elements v8i16 -> v4i32.
13042 // Use vpunpckhwd for 4 upper elements v8i16 -> v4i32.
13043 // Concat upper and lower parts.
13046 // Use vpunpckldq for 4 lower elements v4i32 -> v2i64.
13047 // Use vpunpckhdq for 4 upper elements v4i32 -> v2i64.
13048 // Concat upper and lower parts.
13051 if (((VT != MVT::v16i16) || (InVT != MVT::v16i8)) &&
13052 ((VT != MVT::v8i32) || (InVT != MVT::v8i16)) &&
13053 ((VT != MVT::v4i64) || (InVT != MVT::v4i32)))
13056 if (Subtarget->hasInt256())
13057 return DAG.getNode(X86ISD::VZEXT, dl, VT, In);
13059 SDValue ZeroVec = getZeroVector(InVT, Subtarget, DAG, dl);
13060 SDValue Undef = DAG.getUNDEF(InVT);
13061 bool NeedZero = Op.getOpcode() == ISD::ZERO_EXTEND;
13062 SDValue OpLo = getUnpackl(DAG, dl, InVT, In, NeedZero ? ZeroVec : Undef);
13063 SDValue OpHi = getUnpackh(DAG, dl, InVT, In, NeedZero ? ZeroVec : Undef);
13065 MVT HVT = MVT::getVectorVT(VT.getVectorElementType(),
13066 VT.getVectorNumElements()/2);
13068 OpLo = DAG.getBitcast(HVT, OpLo);
13069 OpHi = DAG.getBitcast(HVT, OpHi);
13071 return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT, OpLo, OpHi);
13074 static SDValue LowerZERO_EXTEND_AVX512(SDValue Op,
13075 const X86Subtarget *Subtarget, SelectionDAG &DAG) {
13076 MVT VT = Op->getSimpleValueType(0);
13077 SDValue In = Op->getOperand(0);
13078 MVT InVT = In.getSimpleValueType();
13080 unsigned int NumElts = VT.getVectorNumElements();
13081 if (NumElts != 8 && NumElts != 16 && !Subtarget->hasBWI())
13084 if (VT.is512BitVector() && InVT.getVectorElementType() != MVT::i1)
13085 return DAG.getNode(X86ISD::VZEXT, DL, VT, In);
13087 assert(InVT.getVectorElementType() == MVT::i1);
13088 MVT ExtVT = NumElts == 8 ? MVT::v8i64 : MVT::v16i32;
13090 DAG.getConstant(APInt(ExtVT.getScalarSizeInBits(), 1), DL, ExtVT);
13092 DAG.getConstant(APInt::getNullValue(ExtVT.getScalarSizeInBits()), DL, ExtVT);
13094 SDValue V = DAG.getNode(ISD::VSELECT, DL, ExtVT, In, One, Zero);
13095 if (VT.is512BitVector())
13097 return DAG.getNode(X86ISD::VTRUNC, DL, VT, V);
13100 static SDValue LowerANY_EXTEND(SDValue Op, const X86Subtarget *Subtarget,
13101 SelectionDAG &DAG) {
13102 if (Subtarget->hasFp256())
13103 if (SDValue Res = LowerAVXExtend(Op, DAG, Subtarget))
13109 static SDValue LowerZERO_EXTEND(SDValue Op, const X86Subtarget *Subtarget,
13110 SelectionDAG &DAG) {
13112 MVT VT = Op.getSimpleValueType();
13113 SDValue In = Op.getOperand(0);
13114 MVT SVT = In.getSimpleValueType();
13116 if (VT.is512BitVector() || SVT.getVectorElementType() == MVT::i1)
13117 return LowerZERO_EXTEND_AVX512(Op, Subtarget, DAG);
13119 if (Subtarget->hasFp256())
13120 if (SDValue Res = LowerAVXExtend(Op, DAG, Subtarget))
13123 assert(!VT.is256BitVector() || !SVT.is128BitVector() ||
13124 VT.getVectorNumElements() != SVT.getVectorNumElements());
13128 SDValue X86TargetLowering::LowerTRUNCATE(SDValue Op, SelectionDAG &DAG) const {
13130 MVT VT = Op.getSimpleValueType();
13131 SDValue In = Op.getOperand(0);
13132 MVT InVT = In.getSimpleValueType();
13134 if (VT == MVT::i1) {
13135 assert((InVT.isInteger() && (InVT.getSizeInBits() <= 64)) &&
13136 "Invalid scalar TRUNCATE operation");
13137 if (InVT.getSizeInBits() >= 32)
13139 In = DAG.getNode(ISD::ANY_EXTEND, DL, MVT::i32, In);
13140 return DAG.getNode(ISD::TRUNCATE, DL, VT, In);
13142 assert(VT.getVectorNumElements() == InVT.getVectorNumElements() &&
13143 "Invalid TRUNCATE operation");
13145 // move vector to mask - truncate solution for SKX
13146 if (VT.getVectorElementType() == MVT::i1) {
13147 if (InVT.is512BitVector() && InVT.getScalarSizeInBits() <= 16 &&
13148 Subtarget->hasBWI())
13149 return Op; // legal, will go to VPMOVB2M, VPMOVW2M
13150 if ((InVT.is256BitVector() || InVT.is128BitVector())
13151 && InVT.getScalarSizeInBits() <= 16 &&
13152 Subtarget->hasBWI() && Subtarget->hasVLX())
13153 return Op; // legal, will go to VPMOVB2M, VPMOVW2M
13154 if (InVT.is512BitVector() && InVT.getScalarSizeInBits() >= 32 &&
13155 Subtarget->hasDQI())
13156 return Op; // legal, will go to VPMOVD2M, VPMOVQ2M
13157 if ((InVT.is256BitVector() || InVT.is128BitVector())
13158 && InVT.getScalarSizeInBits() >= 32 &&
13159 Subtarget->hasDQI() && Subtarget->hasVLX())
13160 return Op; // legal, will go to VPMOVB2M, VPMOVQ2M
13163 if (VT.getVectorElementType() == MVT::i1) {
13164 assert(VT.getVectorElementType() == MVT::i1 && "Unexpected vector type");
13165 unsigned NumElts = InVT.getVectorNumElements();
13166 assert ((NumElts == 8 || NumElts == 16) && "Unexpected vector type");
13167 if (InVT.getSizeInBits() < 512) {
13168 MVT ExtVT = (NumElts == 16)? MVT::v16i32 : MVT::v8i64;
13169 In = DAG.getNode(ISD::SIGN_EXTEND, DL, ExtVT, In);
13174 DAG.getConstant(APInt::getSignBit(InVT.getScalarSizeInBits()), DL, InVT);
13175 SDValue And = DAG.getNode(ISD::AND, DL, InVT, OneV, In);
13176 return DAG.getNode(X86ISD::TESTM, DL, VT, And, And);
13179 // vpmovqb/w/d, vpmovdb/w, vpmovwb
13180 if (((!InVT.is512BitVector() && Subtarget->hasVLX()) || InVT.is512BitVector()) &&
13181 (InVT.getVectorElementType() != MVT::i16 || Subtarget->hasBWI()))
13182 return DAG.getNode(X86ISD::VTRUNC, DL, VT, In);
13184 if ((VT == MVT::v4i32) && (InVT == MVT::v4i64)) {
13185 // On AVX2, v4i64 -> v4i32 becomes VPERMD.
13186 if (Subtarget->hasInt256()) {
13187 static const int ShufMask[] = {0, 2, 4, 6, -1, -1, -1, -1};
13188 In = DAG.getBitcast(MVT::v8i32, In);
13189 In = DAG.getVectorShuffle(MVT::v8i32, DL, In, DAG.getUNDEF(MVT::v8i32),
13191 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, VT, In,
13192 DAG.getIntPtrConstant(0, DL));
13195 SDValue OpLo = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v2i64, In,
13196 DAG.getIntPtrConstant(0, DL));
13197 SDValue OpHi = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v2i64, In,
13198 DAG.getIntPtrConstant(2, DL));
13199 OpLo = DAG.getBitcast(MVT::v4i32, OpLo);
13200 OpHi = DAG.getBitcast(MVT::v4i32, OpHi);
13201 static const int ShufMask[] = {0, 2, 4, 6};
13202 return DAG.getVectorShuffle(VT, DL, OpLo, OpHi, ShufMask);
13205 if ((VT == MVT::v8i16) && (InVT == MVT::v8i32)) {
13206 // On AVX2, v8i32 -> v8i16 becomed PSHUFB.
13207 if (Subtarget->hasInt256()) {
13208 In = DAG.getBitcast(MVT::v32i8, In);
13210 SmallVector<SDValue,32> pshufbMask;
13211 for (unsigned i = 0; i < 2; ++i) {
13212 pshufbMask.push_back(DAG.getConstant(0x0, DL, MVT::i8));
13213 pshufbMask.push_back(DAG.getConstant(0x1, DL, MVT::i8));
13214 pshufbMask.push_back(DAG.getConstant(0x4, DL, MVT::i8));
13215 pshufbMask.push_back(DAG.getConstant(0x5, DL, MVT::i8));
13216 pshufbMask.push_back(DAG.getConstant(0x8, DL, MVT::i8));
13217 pshufbMask.push_back(DAG.getConstant(0x9, DL, MVT::i8));
13218 pshufbMask.push_back(DAG.getConstant(0xc, DL, MVT::i8));
13219 pshufbMask.push_back(DAG.getConstant(0xd, DL, MVT::i8));
13220 for (unsigned j = 0; j < 8; ++j)
13221 pshufbMask.push_back(DAG.getConstant(0x80, DL, MVT::i8));
13223 SDValue BV = DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v32i8, pshufbMask);
13224 In = DAG.getNode(X86ISD::PSHUFB, DL, MVT::v32i8, In, BV);
13225 In = DAG.getBitcast(MVT::v4i64, In);
13227 static const int ShufMask[] = {0, 2, -1, -1};
13228 In = DAG.getVectorShuffle(MVT::v4i64, DL, In, DAG.getUNDEF(MVT::v4i64),
13230 In = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v2i64, In,
13231 DAG.getIntPtrConstant(0, DL));
13232 return DAG.getBitcast(VT, In);
13235 SDValue OpLo = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v4i32, In,
13236 DAG.getIntPtrConstant(0, DL));
13238 SDValue OpHi = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v4i32, In,
13239 DAG.getIntPtrConstant(4, DL));
13241 OpLo = DAG.getBitcast(MVT::v16i8, OpLo);
13242 OpHi = DAG.getBitcast(MVT::v16i8, OpHi);
13244 // The PSHUFB mask:
13245 static const int ShufMask1[] = {0, 1, 4, 5, 8, 9, 12, 13,
13246 -1, -1, -1, -1, -1, -1, -1, -1};
13248 SDValue Undef = DAG.getUNDEF(MVT::v16i8);
13249 OpLo = DAG.getVectorShuffle(MVT::v16i8, DL, OpLo, Undef, ShufMask1);
13250 OpHi = DAG.getVectorShuffle(MVT::v16i8, DL, OpHi, Undef, ShufMask1);
13252 OpLo = DAG.getBitcast(MVT::v4i32, OpLo);
13253 OpHi = DAG.getBitcast(MVT::v4i32, OpHi);
13255 // The MOVLHPS Mask:
13256 static const int ShufMask2[] = {0, 1, 4, 5};
13257 SDValue res = DAG.getVectorShuffle(MVT::v4i32, DL, OpLo, OpHi, ShufMask2);
13258 return DAG.getBitcast(MVT::v8i16, res);
13261 // Handle truncation of V256 to V128 using shuffles.
13262 if (!VT.is128BitVector() || !InVT.is256BitVector())
13265 assert(Subtarget->hasFp256() && "256-bit vector without AVX!");
13267 unsigned NumElems = VT.getVectorNumElements();
13268 MVT NVT = MVT::getVectorVT(VT.getVectorElementType(), NumElems * 2);
13270 SmallVector<int, 16> MaskVec(NumElems * 2, -1);
13271 // Prepare truncation shuffle mask
13272 for (unsigned i = 0; i != NumElems; ++i)
13273 MaskVec[i] = i * 2;
13274 SDValue V = DAG.getVectorShuffle(NVT, DL, DAG.getBitcast(NVT, In),
13275 DAG.getUNDEF(NVT), &MaskVec[0]);
13276 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, VT, V,
13277 DAG.getIntPtrConstant(0, DL));
13280 SDValue X86TargetLowering::LowerFP_TO_SINT(SDValue Op,
13281 SelectionDAG &DAG) const {
13282 assert(!Op.getSimpleValueType().isVector());
13284 std::pair<SDValue,SDValue> Vals = FP_TO_INTHelper(Op, DAG,
13285 /*IsSigned=*/ true, /*IsReplace=*/ false);
13286 SDValue FIST = Vals.first, StackSlot = Vals.second;
13287 // If FP_TO_INTHelper failed, the node is actually supposed to be Legal.
13288 if (!FIST.getNode())
13291 if (StackSlot.getNode())
13292 // Load the result.
13293 return DAG.getLoad(Op.getValueType(), SDLoc(Op),
13294 FIST, StackSlot, MachinePointerInfo(),
13295 false, false, false, 0);
13297 // The node is the result.
13301 SDValue X86TargetLowering::LowerFP_TO_UINT(SDValue Op,
13302 SelectionDAG &DAG) const {
13303 std::pair<SDValue,SDValue> Vals = FP_TO_INTHelper(Op, DAG,
13304 /*IsSigned=*/ false, /*IsReplace=*/ false);
13305 SDValue FIST = Vals.first, StackSlot = Vals.second;
13306 // If FP_TO_INTHelper failed, the node is actually supposed to be Legal.
13307 if (!FIST.getNode())
13310 if (StackSlot.getNode())
13311 // Load the result.
13312 return DAG.getLoad(Op.getValueType(), SDLoc(Op),
13313 FIST, StackSlot, MachinePointerInfo(),
13314 false, false, false, 0);
13316 // The node is the result.
13320 static SDValue LowerFP_EXTEND(SDValue Op, SelectionDAG &DAG) {
13322 MVT VT = Op.getSimpleValueType();
13323 SDValue In = Op.getOperand(0);
13324 MVT SVT = In.getSimpleValueType();
13326 assert(SVT == MVT::v2f32 && "Only customize MVT::v2f32 type legalization!");
13328 return DAG.getNode(X86ISD::VFPEXT, DL, VT,
13329 DAG.getNode(ISD::CONCAT_VECTORS, DL, MVT::v4f32,
13330 In, DAG.getUNDEF(SVT)));
13333 /// The only differences between FABS and FNEG are the mask and the logic op.
13334 /// FNEG also has a folding opportunity for FNEG(FABS(x)).
13335 static SDValue LowerFABSorFNEG(SDValue Op, SelectionDAG &DAG) {
13336 assert((Op.getOpcode() == ISD::FABS || Op.getOpcode() == ISD::FNEG) &&
13337 "Wrong opcode for lowering FABS or FNEG.");
13339 bool IsFABS = (Op.getOpcode() == ISD::FABS);
13341 // If this is a FABS and it has an FNEG user, bail out to fold the combination
13342 // into an FNABS. We'll lower the FABS after that if it is still in use.
13344 for (SDNode *User : Op->uses())
13345 if (User->getOpcode() == ISD::FNEG)
13349 MVT VT = Op.getSimpleValueType();
13351 // FIXME: Use function attribute "OptimizeForSize" and/or CodeGenOpt::Level to
13352 // decide if we should generate a 16-byte constant mask when we only need 4 or
13353 // 8 bytes for the scalar case.
13359 if (VT.isVector()) {
13361 EltVT = VT.getVectorElementType();
13362 NumElts = VT.getVectorNumElements();
13364 // There are no scalar bitwise logical SSE/AVX instructions, so we
13365 // generate a 16-byte vector constant and logic op even for the scalar case.
13366 // Using a 16-byte mask allows folding the load of the mask with
13367 // the logic op, so it can save (~4 bytes) on code size.
13368 LogicVT = (VT == MVT::f64) ? MVT::v2f64 : MVT::v4f32;
13370 NumElts = (VT == MVT::f64) ? 2 : 4;
13373 unsigned EltBits = EltVT.getSizeInBits();
13374 LLVMContext *Context = DAG.getContext();
13375 // For FABS, mask is 0x7f...; for FNEG, mask is 0x80...
13377 IsFABS ? APInt::getSignedMaxValue(EltBits) : APInt::getSignBit(EltBits);
13378 Constant *C = ConstantInt::get(*Context, MaskElt);
13379 C = ConstantVector::getSplat(NumElts, C);
13380 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
13381 SDValue CPIdx = DAG.getConstantPool(C, TLI.getPointerTy(DAG.getDataLayout()));
13382 unsigned Alignment = cast<ConstantPoolSDNode>(CPIdx)->getAlignment();
13384 DAG.getLoad(LogicVT, dl, DAG.getEntryNode(), CPIdx,
13385 MachinePointerInfo::getConstantPool(DAG.getMachineFunction()),
13386 false, false, false, Alignment);
13388 SDValue Op0 = Op.getOperand(0);
13389 bool IsFNABS = !IsFABS && (Op0.getOpcode() == ISD::FABS);
13391 IsFABS ? X86ISD::FAND : IsFNABS ? X86ISD::FOR : X86ISD::FXOR;
13392 SDValue Operand = IsFNABS ? Op0.getOperand(0) : Op0;
13395 return DAG.getNode(LogicOp, dl, LogicVT, Operand, Mask);
13397 // For the scalar case extend to a 128-bit vector, perform the logic op,
13398 // and extract the scalar result back out.
13399 Operand = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, LogicVT, Operand);
13400 SDValue LogicNode = DAG.getNode(LogicOp, dl, LogicVT, Operand, Mask);
13401 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, LogicNode,
13402 DAG.getIntPtrConstant(0, dl));
13405 static SDValue LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) {
13406 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
13407 LLVMContext *Context = DAG.getContext();
13408 SDValue Op0 = Op.getOperand(0);
13409 SDValue Op1 = Op.getOperand(1);
13411 MVT VT = Op.getSimpleValueType();
13412 MVT SrcVT = Op1.getSimpleValueType();
13414 // If second operand is smaller, extend it first.
13415 if (SrcVT.bitsLT(VT)) {
13416 Op1 = DAG.getNode(ISD::FP_EXTEND, dl, VT, Op1);
13419 // And if it is bigger, shrink it first.
13420 if (SrcVT.bitsGT(VT)) {
13421 Op1 = DAG.getNode(ISD::FP_ROUND, dl, VT, Op1, DAG.getIntPtrConstant(1, dl));
13425 // At this point the operands and the result should have the same
13426 // type, and that won't be f80 since that is not custom lowered.
13428 const fltSemantics &Sem =
13429 VT == MVT::f64 ? APFloat::IEEEdouble : APFloat::IEEEsingle;
13430 const unsigned SizeInBits = VT.getSizeInBits();
13432 SmallVector<Constant *, 4> CV(
13433 VT == MVT::f64 ? 2 : 4,
13434 ConstantFP::get(*Context, APFloat(Sem, APInt(SizeInBits, 0))));
13436 // First, clear all bits but the sign bit from the second operand (sign).
13437 CV[0] = ConstantFP::get(*Context,
13438 APFloat(Sem, APInt::getHighBitsSet(SizeInBits, 1)));
13439 Constant *C = ConstantVector::get(CV);
13440 auto PtrVT = TLI.getPointerTy(DAG.getDataLayout());
13441 SDValue CPIdx = DAG.getConstantPool(C, PtrVT, 16);
13443 // Perform all logic operations as 16-byte vectors because there are no
13444 // scalar FP logic instructions in SSE. This allows load folding of the
13445 // constants into the logic instructions.
13446 MVT LogicVT = (VT == MVT::f64) ? MVT::v2f64 : MVT::v4f32;
13448 DAG.getLoad(LogicVT, dl, DAG.getEntryNode(), CPIdx,
13449 MachinePointerInfo::getConstantPool(DAG.getMachineFunction()),
13450 false, false, false, 16);
13451 Op1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, LogicVT, Op1);
13452 SDValue SignBit = DAG.getNode(X86ISD::FAND, dl, LogicVT, Op1, Mask1);
13454 // Next, clear the sign bit from the first operand (magnitude).
13455 // If it's a constant, we can clear it here.
13456 if (ConstantFPSDNode *Op0CN = dyn_cast<ConstantFPSDNode>(Op0)) {
13457 APFloat APF = Op0CN->getValueAPF();
13458 // If the magnitude is a positive zero, the sign bit alone is enough.
13459 if (APF.isPosZero())
13460 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, SrcVT, SignBit,
13461 DAG.getIntPtrConstant(0, dl));
13463 CV[0] = ConstantFP::get(*Context, APF);
13465 CV[0] = ConstantFP::get(
13467 APFloat(Sem, APInt::getLowBitsSet(SizeInBits, SizeInBits - 1)));
13469 C = ConstantVector::get(CV);
13470 CPIdx = DAG.getConstantPool(C, PtrVT, 16);
13472 DAG.getLoad(LogicVT, dl, DAG.getEntryNode(), CPIdx,
13473 MachinePointerInfo::getConstantPool(DAG.getMachineFunction()),
13474 false, false, false, 16);
13475 // If the magnitude operand wasn't a constant, we need to AND out the sign.
13476 if (!isa<ConstantFPSDNode>(Op0)) {
13477 Op0 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, LogicVT, Op0);
13478 Val = DAG.getNode(X86ISD::FAND, dl, LogicVT, Op0, Val);
13480 // OR the magnitude value with the sign bit.
13481 Val = DAG.getNode(X86ISD::FOR, dl, LogicVT, Val, SignBit);
13482 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, SrcVT, Val,
13483 DAG.getIntPtrConstant(0, dl));
13486 static SDValue LowerFGETSIGN(SDValue Op, SelectionDAG &DAG) {
13487 SDValue N0 = Op.getOperand(0);
13489 MVT VT = Op.getSimpleValueType();
13491 // Lower ISD::FGETSIGN to (AND (X86ISD::FGETSIGNx86 ...) 1).
13492 SDValue xFGETSIGN = DAG.getNode(X86ISD::FGETSIGNx86, dl, VT, N0,
13493 DAG.getConstant(1, dl, VT));
13494 return DAG.getNode(ISD::AND, dl, VT, xFGETSIGN, DAG.getConstant(1, dl, VT));
13497 // Check whether an OR'd tree is PTEST-able.
13498 static SDValue LowerVectorAllZeroTest(SDValue Op, const X86Subtarget *Subtarget,
13499 SelectionDAG &DAG) {
13500 assert(Op.getOpcode() == ISD::OR && "Only check OR'd tree.");
13502 if (!Subtarget->hasSSE41())
13505 if (!Op->hasOneUse())
13508 SDNode *N = Op.getNode();
13511 SmallVector<SDValue, 8> Opnds;
13512 DenseMap<SDValue, unsigned> VecInMap;
13513 SmallVector<SDValue, 8> VecIns;
13514 EVT VT = MVT::Other;
13516 // Recognize a special case where a vector is casted into wide integer to
13518 Opnds.push_back(N->getOperand(0));
13519 Opnds.push_back(N->getOperand(1));
13521 for (unsigned Slot = 0, e = Opnds.size(); Slot < e; ++Slot) {
13522 SmallVectorImpl<SDValue>::const_iterator I = Opnds.begin() + Slot;
13523 // BFS traverse all OR'd operands.
13524 if (I->getOpcode() == ISD::OR) {
13525 Opnds.push_back(I->getOperand(0));
13526 Opnds.push_back(I->getOperand(1));
13527 // Re-evaluate the number of nodes to be traversed.
13528 e += 2; // 2 more nodes (LHS and RHS) are pushed.
13532 // Quit if a non-EXTRACT_VECTOR_ELT
13533 if (I->getOpcode() != ISD::EXTRACT_VECTOR_ELT)
13536 // Quit if without a constant index.
13537 SDValue Idx = I->getOperand(1);
13538 if (!isa<ConstantSDNode>(Idx))
13541 SDValue ExtractedFromVec = I->getOperand(0);
13542 DenseMap<SDValue, unsigned>::iterator M = VecInMap.find(ExtractedFromVec);
13543 if (M == VecInMap.end()) {
13544 VT = ExtractedFromVec.getValueType();
13545 // Quit if not 128/256-bit vector.
13546 if (!VT.is128BitVector() && !VT.is256BitVector())
13548 // Quit if not the same type.
13549 if (VecInMap.begin() != VecInMap.end() &&
13550 VT != VecInMap.begin()->first.getValueType())
13552 M = VecInMap.insert(std::make_pair(ExtractedFromVec, 0)).first;
13553 VecIns.push_back(ExtractedFromVec);
13555 M->second |= 1U << cast<ConstantSDNode>(Idx)->getZExtValue();
13558 assert((VT.is128BitVector() || VT.is256BitVector()) &&
13559 "Not extracted from 128-/256-bit vector.");
13561 unsigned FullMask = (1U << VT.getVectorNumElements()) - 1U;
13563 for (DenseMap<SDValue, unsigned>::const_iterator
13564 I = VecInMap.begin(), E = VecInMap.end(); I != E; ++I) {
13565 // Quit if not all elements are used.
13566 if (I->second != FullMask)
13570 EVT TestVT = VT.is128BitVector() ? MVT::v2i64 : MVT::v4i64;
13572 // Cast all vectors into TestVT for PTEST.
13573 for (unsigned i = 0, e = VecIns.size(); i < e; ++i)
13574 VecIns[i] = DAG.getBitcast(TestVT, VecIns[i]);
13576 // If more than one full vectors are evaluated, OR them first before PTEST.
13577 for (unsigned Slot = 0, e = VecIns.size(); e - Slot > 1; Slot += 2, e += 1) {
13578 // Each iteration will OR 2 nodes and append the result until there is only
13579 // 1 node left, i.e. the final OR'd value of all vectors.
13580 SDValue LHS = VecIns[Slot];
13581 SDValue RHS = VecIns[Slot + 1];
13582 VecIns.push_back(DAG.getNode(ISD::OR, DL, TestVT, LHS, RHS));
13585 return DAG.getNode(X86ISD::PTEST, DL, MVT::i32,
13586 VecIns.back(), VecIns.back());
13589 /// \brief return true if \c Op has a use that doesn't just read flags.
13590 static bool hasNonFlagsUse(SDValue Op) {
13591 for (SDNode::use_iterator UI = Op->use_begin(), UE = Op->use_end(); UI != UE;
13593 SDNode *User = *UI;
13594 unsigned UOpNo = UI.getOperandNo();
13595 if (User->getOpcode() == ISD::TRUNCATE && User->hasOneUse()) {
13596 // Look pass truncate.
13597 UOpNo = User->use_begin().getOperandNo();
13598 User = *User->use_begin();
13601 if (User->getOpcode() != ISD::BRCOND && User->getOpcode() != ISD::SETCC &&
13602 !(User->getOpcode() == ISD::SELECT && UOpNo == 0))
13608 /// Emit nodes that will be selected as "test Op0,Op0", or something
13610 SDValue X86TargetLowering::EmitTest(SDValue Op, unsigned X86CC, SDLoc dl,
13611 SelectionDAG &DAG) const {
13612 if (Op.getValueType() == MVT::i1) {
13613 SDValue ExtOp = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i8, Op);
13614 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, ExtOp,
13615 DAG.getConstant(0, dl, MVT::i8));
13617 // CF and OF aren't always set the way we want. Determine which
13618 // of these we need.
13619 bool NeedCF = false;
13620 bool NeedOF = false;
13623 case X86::COND_A: case X86::COND_AE:
13624 case X86::COND_B: case X86::COND_BE:
13627 case X86::COND_G: case X86::COND_GE:
13628 case X86::COND_L: case X86::COND_LE:
13629 case X86::COND_O: case X86::COND_NO: {
13630 // Check if we really need to set the
13631 // Overflow flag. If NoSignedWrap is present
13632 // that is not actually needed.
13633 switch (Op->getOpcode()) {
13638 const auto *BinNode = cast<BinaryWithFlagsSDNode>(Op.getNode());
13639 if (BinNode->Flags.hasNoSignedWrap())
13649 // See if we can use the EFLAGS value from the operand instead of
13650 // doing a separate TEST. TEST always sets OF and CF to 0, so unless
13651 // we prove that the arithmetic won't overflow, we can't use OF or CF.
13652 if (Op.getResNo() != 0 || NeedOF || NeedCF) {
13653 // Emit a CMP with 0, which is the TEST pattern.
13654 //if (Op.getValueType() == MVT::i1)
13655 // return DAG.getNode(X86ISD::CMP, dl, MVT::i1, Op,
13656 // DAG.getConstant(0, MVT::i1));
13657 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op,
13658 DAG.getConstant(0, dl, Op.getValueType()));
13660 unsigned Opcode = 0;
13661 unsigned NumOperands = 0;
13663 // Truncate operations may prevent the merge of the SETCC instruction
13664 // and the arithmetic instruction before it. Attempt to truncate the operands
13665 // of the arithmetic instruction and use a reduced bit-width instruction.
13666 bool NeedTruncation = false;
13667 SDValue ArithOp = Op;
13668 if (Op->getOpcode() == ISD::TRUNCATE && Op->hasOneUse()) {
13669 SDValue Arith = Op->getOperand(0);
13670 // Both the trunc and the arithmetic op need to have one user each.
13671 if (Arith->hasOneUse())
13672 switch (Arith.getOpcode()) {
13679 NeedTruncation = true;
13685 // NOTICE: In the code below we use ArithOp to hold the arithmetic operation
13686 // which may be the result of a CAST. We use the variable 'Op', which is the
13687 // non-casted variable when we check for possible users.
13688 switch (ArithOp.getOpcode()) {
13690 // Due to an isel shortcoming, be conservative if this add is likely to be
13691 // selected as part of a load-modify-store instruction. When the root node
13692 // in a match is a store, isel doesn't know how to remap non-chain non-flag
13693 // uses of other nodes in the match, such as the ADD in this case. This
13694 // leads to the ADD being left around and reselected, with the result being
13695 // two adds in the output. Alas, even if none our users are stores, that
13696 // doesn't prove we're O.K. Ergo, if we have any parents that aren't
13697 // CopyToReg or SETCC, eschew INC/DEC. A better fix seems to require
13698 // climbing the DAG back to the root, and it doesn't seem to be worth the
13700 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
13701 UE = Op.getNode()->use_end(); UI != UE; ++UI)
13702 if (UI->getOpcode() != ISD::CopyToReg &&
13703 UI->getOpcode() != ISD::SETCC &&
13704 UI->getOpcode() != ISD::STORE)
13707 if (ConstantSDNode *C =
13708 dyn_cast<ConstantSDNode>(ArithOp.getNode()->getOperand(1))) {
13709 // An add of one will be selected as an INC.
13710 if (C->getAPIntValue() == 1 && !Subtarget->slowIncDec()) {
13711 Opcode = X86ISD::INC;
13716 // An add of negative one (subtract of one) will be selected as a DEC.
13717 if (C->getAPIntValue().isAllOnesValue() && !Subtarget->slowIncDec()) {
13718 Opcode = X86ISD::DEC;
13724 // Otherwise use a regular EFLAGS-setting add.
13725 Opcode = X86ISD::ADD;
13730 // If we have a constant logical shift that's only used in a comparison
13731 // against zero turn it into an equivalent AND. This allows turning it into
13732 // a TEST instruction later.
13733 if ((X86CC == X86::COND_E || X86CC == X86::COND_NE) && Op->hasOneUse() &&
13734 isa<ConstantSDNode>(Op->getOperand(1)) && !hasNonFlagsUse(Op)) {
13735 EVT VT = Op.getValueType();
13736 unsigned BitWidth = VT.getSizeInBits();
13737 unsigned ShAmt = Op->getConstantOperandVal(1);
13738 if (ShAmt >= BitWidth) // Avoid undefined shifts.
13740 APInt Mask = ArithOp.getOpcode() == ISD::SRL
13741 ? APInt::getHighBitsSet(BitWidth, BitWidth - ShAmt)
13742 : APInt::getLowBitsSet(BitWidth, BitWidth - ShAmt);
13743 if (!Mask.isSignedIntN(32)) // Avoid large immediates.
13745 SDValue New = DAG.getNode(ISD::AND, dl, VT, Op->getOperand(0),
13746 DAG.getConstant(Mask, dl, VT));
13747 DAG.ReplaceAllUsesWith(Op, New);
13753 // If the primary and result isn't used, don't bother using X86ISD::AND,
13754 // because a TEST instruction will be better.
13755 if (!hasNonFlagsUse(Op))
13761 // Due to the ISEL shortcoming noted above, be conservative if this op is
13762 // likely to be selected as part of a load-modify-store instruction.
13763 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
13764 UE = Op.getNode()->use_end(); UI != UE; ++UI)
13765 if (UI->getOpcode() == ISD::STORE)
13768 // Otherwise use a regular EFLAGS-setting instruction.
13769 switch (ArithOp.getOpcode()) {
13770 default: llvm_unreachable("unexpected operator!");
13771 case ISD::SUB: Opcode = X86ISD::SUB; break;
13772 case ISD::XOR: Opcode = X86ISD::XOR; break;
13773 case ISD::AND: Opcode = X86ISD::AND; break;
13775 if (!NeedTruncation && (X86CC == X86::COND_E || X86CC == X86::COND_NE)) {
13776 SDValue EFLAGS = LowerVectorAllZeroTest(Op, Subtarget, DAG);
13777 if (EFLAGS.getNode())
13780 Opcode = X86ISD::OR;
13794 return SDValue(Op.getNode(), 1);
13800 // If we found that truncation is beneficial, perform the truncation and
13802 if (NeedTruncation) {
13803 EVT VT = Op.getValueType();
13804 SDValue WideVal = Op->getOperand(0);
13805 EVT WideVT = WideVal.getValueType();
13806 unsigned ConvertedOp = 0;
13807 // Use a target machine opcode to prevent further DAGCombine
13808 // optimizations that may separate the arithmetic operations
13809 // from the setcc node.
13810 switch (WideVal.getOpcode()) {
13812 case ISD::ADD: ConvertedOp = X86ISD::ADD; break;
13813 case ISD::SUB: ConvertedOp = X86ISD::SUB; break;
13814 case ISD::AND: ConvertedOp = X86ISD::AND; break;
13815 case ISD::OR: ConvertedOp = X86ISD::OR; break;
13816 case ISD::XOR: ConvertedOp = X86ISD::XOR; break;
13820 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
13821 if (TLI.isOperationLegal(WideVal.getOpcode(), WideVT)) {
13822 SDValue V0 = DAG.getNode(ISD::TRUNCATE, dl, VT, WideVal.getOperand(0));
13823 SDValue V1 = DAG.getNode(ISD::TRUNCATE, dl, VT, WideVal.getOperand(1));
13824 Op = DAG.getNode(ConvertedOp, dl, VT, V0, V1);
13830 // Emit a CMP with 0, which is the TEST pattern.
13831 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op,
13832 DAG.getConstant(0, dl, Op.getValueType()));
13834 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::i32);
13835 SmallVector<SDValue, 4> Ops(Op->op_begin(), Op->op_begin() + NumOperands);
13837 SDValue New = DAG.getNode(Opcode, dl, VTs, Ops);
13838 DAG.ReplaceAllUsesWith(Op, New);
13839 return SDValue(New.getNode(), 1);
13842 /// Emit nodes that will be selected as "cmp Op0,Op1", or something
13844 SDValue X86TargetLowering::EmitCmp(SDValue Op0, SDValue Op1, unsigned X86CC,
13845 SDLoc dl, SelectionDAG &DAG) const {
13846 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op1)) {
13847 if (C->getAPIntValue() == 0)
13848 return EmitTest(Op0, X86CC, dl, DAG);
13850 if (Op0.getValueType() == MVT::i1)
13851 llvm_unreachable("Unexpected comparison operation for MVT::i1 operands");
13854 if ((Op0.getValueType() == MVT::i8 || Op0.getValueType() == MVT::i16 ||
13855 Op0.getValueType() == MVT::i32 || Op0.getValueType() == MVT::i64)) {
13856 // Do the comparison at i32 if it's smaller, besides the Atom case.
13857 // This avoids subregister aliasing issues. Keep the smaller reference
13858 // if we're optimizing for size, however, as that'll allow better folding
13859 // of memory operations.
13860 if (Op0.getValueType() != MVT::i32 && Op0.getValueType() != MVT::i64 &&
13861 !DAG.getMachineFunction().getFunction()->optForMinSize() &&
13862 !Subtarget->isAtom()) {
13863 unsigned ExtendOp =
13864 isX86CCUnsigned(X86CC) ? ISD::ZERO_EXTEND : ISD::SIGN_EXTEND;
13865 Op0 = DAG.getNode(ExtendOp, dl, MVT::i32, Op0);
13866 Op1 = DAG.getNode(ExtendOp, dl, MVT::i32, Op1);
13868 // Use SUB instead of CMP to enable CSE between SUB and CMP.
13869 SDVTList VTs = DAG.getVTList(Op0.getValueType(), MVT::i32);
13870 SDValue Sub = DAG.getNode(X86ISD::SUB, dl, VTs,
13872 return SDValue(Sub.getNode(), 1);
13874 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op0, Op1);
13877 /// Convert a comparison if required by the subtarget.
13878 SDValue X86TargetLowering::ConvertCmpIfNecessary(SDValue Cmp,
13879 SelectionDAG &DAG) const {
13880 // If the subtarget does not support the FUCOMI instruction, floating-point
13881 // comparisons have to be converted.
13882 if (Subtarget->hasCMov() ||
13883 Cmp.getOpcode() != X86ISD::CMP ||
13884 !Cmp.getOperand(0).getValueType().isFloatingPoint() ||
13885 !Cmp.getOperand(1).getValueType().isFloatingPoint())
13888 // The instruction selector will select an FUCOM instruction instead of
13889 // FUCOMI, which writes the comparison result to FPSW instead of EFLAGS. Hence
13890 // build an SDNode sequence that transfers the result from FPSW into EFLAGS:
13891 // (X86sahf (trunc (srl (X86fp_stsw (trunc (X86cmp ...)), 8))))
13893 SDValue TruncFPSW = DAG.getNode(ISD::TRUNCATE, dl, MVT::i16, Cmp);
13894 SDValue FNStSW = DAG.getNode(X86ISD::FNSTSW16r, dl, MVT::i16, TruncFPSW);
13895 SDValue Srl = DAG.getNode(ISD::SRL, dl, MVT::i16, FNStSW,
13896 DAG.getConstant(8, dl, MVT::i8));
13897 SDValue TruncSrl = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Srl);
13898 return DAG.getNode(X86ISD::SAHF, dl, MVT::i32, TruncSrl);
13901 /// The minimum architected relative accuracy is 2^-12. We need one
13902 /// Newton-Raphson step to have a good float result (24 bits of precision).
13903 SDValue X86TargetLowering::getRsqrtEstimate(SDValue Op,
13904 DAGCombinerInfo &DCI,
13905 unsigned &RefinementSteps,
13906 bool &UseOneConstNR) const {
13907 EVT VT = Op.getValueType();
13908 const char *RecipOp;
13910 // SSE1 has rsqrtss and rsqrtps. AVX adds a 256-bit variant for rsqrtps.
13911 // TODO: Add support for AVX512 (v16f32).
13912 // It is likely not profitable to do this for f64 because a double-precision
13913 // rsqrt estimate with refinement on x86 prior to FMA requires at least 16
13914 // instructions: convert to single, rsqrtss, convert back to double, refine
13915 // (3 steps = at least 13 insts). If an 'rsqrtsd' variant was added to the ISA
13916 // along with FMA, this could be a throughput win.
13917 if (VT == MVT::f32 && Subtarget->hasSSE1())
13919 else if ((VT == MVT::v4f32 && Subtarget->hasSSE1()) ||
13920 (VT == MVT::v8f32 && Subtarget->hasAVX()))
13921 RecipOp = "vec-sqrtf";
13925 TargetRecip Recips = DCI.DAG.getTarget().Options.Reciprocals;
13926 if (!Recips.isEnabled(RecipOp))
13929 RefinementSteps = Recips.getRefinementSteps(RecipOp);
13930 UseOneConstNR = false;
13931 return DCI.DAG.getNode(X86ISD::FRSQRT, SDLoc(Op), VT, Op);
13934 /// The minimum architected relative accuracy is 2^-12. We need one
13935 /// Newton-Raphson step to have a good float result (24 bits of precision).
13936 SDValue X86TargetLowering::getRecipEstimate(SDValue Op,
13937 DAGCombinerInfo &DCI,
13938 unsigned &RefinementSteps) const {
13939 EVT VT = Op.getValueType();
13940 const char *RecipOp;
13942 // SSE1 has rcpss and rcpps. AVX adds a 256-bit variant for rcpps.
13943 // TODO: Add support for AVX512 (v16f32).
13944 // It is likely not profitable to do this for f64 because a double-precision
13945 // reciprocal estimate with refinement on x86 prior to FMA requires
13946 // 15 instructions: convert to single, rcpss, convert back to double, refine
13947 // (3 steps = 12 insts). If an 'rcpsd' variant was added to the ISA
13948 // along with FMA, this could be a throughput win.
13949 if (VT == MVT::f32 && Subtarget->hasSSE1())
13951 else if ((VT == MVT::v4f32 && Subtarget->hasSSE1()) ||
13952 (VT == MVT::v8f32 && Subtarget->hasAVX()))
13953 RecipOp = "vec-divf";
13957 TargetRecip Recips = DCI.DAG.getTarget().Options.Reciprocals;
13958 if (!Recips.isEnabled(RecipOp))
13961 RefinementSteps = Recips.getRefinementSteps(RecipOp);
13962 return DCI.DAG.getNode(X86ISD::FRCP, SDLoc(Op), VT, Op);
13965 /// If we have at least two divisions that use the same divisor, convert to
13966 /// multplication by a reciprocal. This may need to be adjusted for a given
13967 /// CPU if a division's cost is not at least twice the cost of a multiplication.
13968 /// This is because we still need one division to calculate the reciprocal and
13969 /// then we need two multiplies by that reciprocal as replacements for the
13970 /// original divisions.
13971 unsigned X86TargetLowering::combineRepeatedFPDivisors() const {
13975 static bool isAllOnes(SDValue V) {
13976 ConstantSDNode *C = dyn_cast<ConstantSDNode>(V);
13977 return C && C->isAllOnesValue();
13980 /// LowerToBT - Result of 'and' is compared against zero. Turn it into a BT node
13981 /// if it's possible.
13982 SDValue X86TargetLowering::LowerToBT(SDValue And, ISD::CondCode CC,
13983 SDLoc dl, SelectionDAG &DAG) const {
13984 SDValue Op0 = And.getOperand(0);
13985 SDValue Op1 = And.getOperand(1);
13986 if (Op0.getOpcode() == ISD::TRUNCATE)
13987 Op0 = Op0.getOperand(0);
13988 if (Op1.getOpcode() == ISD::TRUNCATE)
13989 Op1 = Op1.getOperand(0);
13992 if (Op1.getOpcode() == ISD::SHL)
13993 std::swap(Op0, Op1);
13994 if (Op0.getOpcode() == ISD::SHL) {
13995 if (ConstantSDNode *And00C = dyn_cast<ConstantSDNode>(Op0.getOperand(0)))
13996 if (And00C->getZExtValue() == 1) {
13997 // If we looked past a truncate, check that it's only truncating away
13999 unsigned BitWidth = Op0.getValueSizeInBits();
14000 unsigned AndBitWidth = And.getValueSizeInBits();
14001 if (BitWidth > AndBitWidth) {
14003 DAG.computeKnownBits(Op0, Zeros, Ones);
14004 if (Zeros.countLeadingOnes() < BitWidth - AndBitWidth)
14008 RHS = Op0.getOperand(1);
14010 } else if (Op1.getOpcode() == ISD::Constant) {
14011 ConstantSDNode *AndRHS = cast<ConstantSDNode>(Op1);
14012 uint64_t AndRHSVal = AndRHS->getZExtValue();
14013 SDValue AndLHS = Op0;
14015 if (AndRHSVal == 1 && AndLHS.getOpcode() == ISD::SRL) {
14016 LHS = AndLHS.getOperand(0);
14017 RHS = AndLHS.getOperand(1);
14020 // Use BT if the immediate can't be encoded in a TEST instruction.
14021 if (!isUInt<32>(AndRHSVal) && isPowerOf2_64(AndRHSVal)) {
14023 RHS = DAG.getConstant(Log2_64_Ceil(AndRHSVal), dl, LHS.getValueType());
14027 if (LHS.getNode()) {
14028 // If LHS is i8, promote it to i32 with any_extend. There is no i8 BT
14029 // instruction. Since the shift amount is in-range-or-undefined, we know
14030 // that doing a bittest on the i32 value is ok. We extend to i32 because
14031 // the encoding for the i16 version is larger than the i32 version.
14032 // Also promote i16 to i32 for performance / code size reason.
14033 if (LHS.getValueType() == MVT::i8 ||
14034 LHS.getValueType() == MVT::i16)
14035 LHS = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, LHS);
14037 // If the operand types disagree, extend the shift amount to match. Since
14038 // BT ignores high bits (like shifts) we can use anyextend.
14039 if (LHS.getValueType() != RHS.getValueType())
14040 RHS = DAG.getNode(ISD::ANY_EXTEND, dl, LHS.getValueType(), RHS);
14042 SDValue BT = DAG.getNode(X86ISD::BT, dl, MVT::i32, LHS, RHS);
14043 X86::CondCode Cond = CC == ISD::SETEQ ? X86::COND_AE : X86::COND_B;
14044 return DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
14045 DAG.getConstant(Cond, dl, MVT::i8), BT);
14051 /// \brief - Turns an ISD::CondCode into a value suitable for SSE floating point
14053 static int translateX86FSETCC(ISD::CondCode SetCCOpcode, SDValue &Op0,
14058 // SSE Condition code mapping:
14067 switch (SetCCOpcode) {
14068 default: llvm_unreachable("Unexpected SETCC condition");
14070 case ISD::SETEQ: SSECC = 0; break;
14072 case ISD::SETGT: Swap = true; // Fallthrough
14074 case ISD::SETOLT: SSECC = 1; break;
14076 case ISD::SETGE: Swap = true; // Fallthrough
14078 case ISD::SETOLE: SSECC = 2; break;
14079 case ISD::SETUO: SSECC = 3; break;
14081 case ISD::SETNE: SSECC = 4; break;
14082 case ISD::SETULE: Swap = true; // Fallthrough
14083 case ISD::SETUGE: SSECC = 5; break;
14084 case ISD::SETULT: Swap = true; // Fallthrough
14085 case ISD::SETUGT: SSECC = 6; break;
14086 case ISD::SETO: SSECC = 7; break;
14088 case ISD::SETONE: SSECC = 8; break;
14091 std::swap(Op0, Op1);
14096 // Lower256IntVSETCC - Break a VSETCC 256-bit integer VSETCC into two new 128
14097 // ones, and then concatenate the result back.
14098 static SDValue Lower256IntVSETCC(SDValue Op, SelectionDAG &DAG) {
14099 MVT VT = Op.getSimpleValueType();
14101 assert(VT.is256BitVector() && Op.getOpcode() == ISD::SETCC &&
14102 "Unsupported value type for operation");
14104 unsigned NumElems = VT.getVectorNumElements();
14106 SDValue CC = Op.getOperand(2);
14108 // Extract the LHS vectors
14109 SDValue LHS = Op.getOperand(0);
14110 SDValue LHS1 = Extract128BitVector(LHS, 0, DAG, dl);
14111 SDValue LHS2 = Extract128BitVector(LHS, NumElems/2, DAG, dl);
14113 // Extract the RHS vectors
14114 SDValue RHS = Op.getOperand(1);
14115 SDValue RHS1 = Extract128BitVector(RHS, 0, DAG, dl);
14116 SDValue RHS2 = Extract128BitVector(RHS, NumElems/2, DAG, dl);
14118 // Issue the operation on the smaller types and concatenate the result back
14119 MVT EltVT = VT.getVectorElementType();
14120 MVT NewVT = MVT::getVectorVT(EltVT, NumElems/2);
14121 return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT,
14122 DAG.getNode(Op.getOpcode(), dl, NewVT, LHS1, RHS1, CC),
14123 DAG.getNode(Op.getOpcode(), dl, NewVT, LHS2, RHS2, CC));
14126 static SDValue LowerBoolVSETCC_AVX512(SDValue Op, SelectionDAG &DAG) {
14127 SDValue Op0 = Op.getOperand(0);
14128 SDValue Op1 = Op.getOperand(1);
14129 SDValue CC = Op.getOperand(2);
14130 MVT VT = Op.getSimpleValueType();
14133 assert(Op0.getValueType().getVectorElementType() == MVT::i1 &&
14134 "Unexpected type for boolean compare operation");
14135 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
14136 SDValue NotOp0 = DAG.getNode(ISD::XOR, dl, VT, Op0,
14137 DAG.getConstant(-1, dl, VT));
14138 SDValue NotOp1 = DAG.getNode(ISD::XOR, dl, VT, Op1,
14139 DAG.getConstant(-1, dl, VT));
14140 switch (SetCCOpcode) {
14141 default: llvm_unreachable("Unexpected SETCC condition");
14143 // (x == y) -> ~(x ^ y)
14144 return DAG.getNode(ISD::XOR, dl, VT,
14145 DAG.getNode(ISD::XOR, dl, VT, Op0, Op1),
14146 DAG.getConstant(-1, dl, VT));
14148 // (x != y) -> (x ^ y)
14149 return DAG.getNode(ISD::XOR, dl, VT, Op0, Op1);
14152 // (x > y) -> (x & ~y)
14153 return DAG.getNode(ISD::AND, dl, VT, Op0, NotOp1);
14156 // (x < y) -> (~x & y)
14157 return DAG.getNode(ISD::AND, dl, VT, NotOp0, Op1);
14160 // (x <= y) -> (~x | y)
14161 return DAG.getNode(ISD::OR, dl, VT, NotOp0, Op1);
14164 // (x >=y) -> (x | ~y)
14165 return DAG.getNode(ISD::OR, dl, VT, Op0, NotOp1);
14169 static SDValue LowerIntVSETCC_AVX512(SDValue Op, SelectionDAG &DAG,
14170 const X86Subtarget *Subtarget) {
14171 SDValue Op0 = Op.getOperand(0);
14172 SDValue Op1 = Op.getOperand(1);
14173 SDValue CC = Op.getOperand(2);
14174 MVT VT = Op.getSimpleValueType();
14177 assert(Op0.getValueType().getVectorElementType().getSizeInBits() >= 8 &&
14178 Op.getValueType().getScalarType() == MVT::i1 &&
14179 "Cannot set masked compare for this operation");
14181 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
14183 bool Unsigned = false;
14186 switch (SetCCOpcode) {
14187 default: llvm_unreachable("Unexpected SETCC condition");
14188 case ISD::SETNE: SSECC = 4; break;
14189 case ISD::SETEQ: Opc = X86ISD::PCMPEQM; break;
14190 case ISD::SETUGT: SSECC = 6; Unsigned = true; break;
14191 case ISD::SETLT: Swap = true; //fall-through
14192 case ISD::SETGT: Opc = X86ISD::PCMPGTM; break;
14193 case ISD::SETULT: SSECC = 1; Unsigned = true; break;
14194 case ISD::SETUGE: SSECC = 5; Unsigned = true; break; //NLT
14195 case ISD::SETGE: Swap = true; SSECC = 2; break; // LE + swap
14196 case ISD::SETULE: Unsigned = true; //fall-through
14197 case ISD::SETLE: SSECC = 2; break;
14201 std::swap(Op0, Op1);
14203 return DAG.getNode(Opc, dl, VT, Op0, Op1);
14204 Opc = Unsigned ? X86ISD::CMPMU: X86ISD::CMPM;
14205 return DAG.getNode(Opc, dl, VT, Op0, Op1,
14206 DAG.getConstant(SSECC, dl, MVT::i8));
14209 /// \brief Try to turn a VSETULT into a VSETULE by modifying its second
14210 /// operand \p Op1. If non-trivial (for example because it's not constant)
14211 /// return an empty value.
14212 static SDValue ChangeVSETULTtoVSETULE(SDLoc dl, SDValue Op1, SelectionDAG &DAG)
14214 BuildVectorSDNode *BV = dyn_cast<BuildVectorSDNode>(Op1.getNode());
14218 MVT VT = Op1.getSimpleValueType();
14219 MVT EVT = VT.getVectorElementType();
14220 unsigned n = VT.getVectorNumElements();
14221 SmallVector<SDValue, 8> ULTOp1;
14223 for (unsigned i = 0; i < n; ++i) {
14224 ConstantSDNode *Elt = dyn_cast<ConstantSDNode>(BV->getOperand(i));
14225 if (!Elt || Elt->isOpaque() || Elt->getValueType(0) != EVT)
14228 // Avoid underflow.
14229 APInt Val = Elt->getAPIntValue();
14233 ULTOp1.push_back(DAG.getConstant(Val - 1, dl, EVT));
14236 return DAG.getNode(ISD::BUILD_VECTOR, dl, VT, ULTOp1);
14239 static SDValue LowerVSETCC(SDValue Op, const X86Subtarget *Subtarget,
14240 SelectionDAG &DAG) {
14241 SDValue Op0 = Op.getOperand(0);
14242 SDValue Op1 = Op.getOperand(1);
14243 SDValue CC = Op.getOperand(2);
14244 MVT VT = Op.getSimpleValueType();
14245 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
14246 bool isFP = Op.getOperand(1).getSimpleValueType().isFloatingPoint();
14251 MVT EltVT = Op0.getSimpleValueType().getVectorElementType();
14252 assert(EltVT == MVT::f32 || EltVT == MVT::f64);
14255 unsigned SSECC = translateX86FSETCC(SetCCOpcode, Op0, Op1);
14256 unsigned Opc = X86ISD::CMPP;
14257 if (Subtarget->hasAVX512() && VT.getVectorElementType() == MVT::i1) {
14258 assert(VT.getVectorNumElements() <= 16);
14259 Opc = X86ISD::CMPM;
14261 // In the two special cases we can't handle, emit two comparisons.
14264 unsigned CombineOpc;
14265 if (SetCCOpcode == ISD::SETUEQ) {
14266 CC0 = 3; CC1 = 0; CombineOpc = ISD::OR;
14268 assert(SetCCOpcode == ISD::SETONE);
14269 CC0 = 7; CC1 = 4; CombineOpc = ISD::AND;
14272 SDValue Cmp0 = DAG.getNode(Opc, dl, VT, Op0, Op1,
14273 DAG.getConstant(CC0, dl, MVT::i8));
14274 SDValue Cmp1 = DAG.getNode(Opc, dl, VT, Op0, Op1,
14275 DAG.getConstant(CC1, dl, MVT::i8));
14276 return DAG.getNode(CombineOpc, dl, VT, Cmp0, Cmp1);
14278 // Handle all other FP comparisons here.
14279 return DAG.getNode(Opc, dl, VT, Op0, Op1,
14280 DAG.getConstant(SSECC, dl, MVT::i8));
14283 MVT VTOp0 = Op0.getSimpleValueType();
14284 assert(VTOp0 == Op1.getSimpleValueType() &&
14285 "Expected operands with same type!");
14286 assert(VT.getVectorNumElements() == VTOp0.getVectorNumElements() &&
14287 "Invalid number of packed elements for source and destination!");
14289 if (VT.is128BitVector() && VTOp0.is256BitVector()) {
14290 // On non-AVX512 targets, a vector of MVT::i1 is promoted by the type
14291 // legalizer to a wider vector type. In the case of 'vsetcc' nodes, the
14292 // legalizer firstly checks if the first operand in input to the setcc has
14293 // a legal type. If so, then it promotes the return type to that same type.
14294 // Otherwise, the return type is promoted to the 'next legal type' which,
14295 // for a vector of MVT::i1 is always a 128-bit integer vector type.
14297 // We reach this code only if the following two conditions are met:
14298 // 1. Both return type and operand type have been promoted to wider types
14299 // by the type legalizer.
14300 // 2. The original operand type has been promoted to a 256-bit vector.
14302 // Note that condition 2. only applies for AVX targets.
14303 SDValue NewOp = DAG.getSetCC(dl, VTOp0, Op0, Op1, SetCCOpcode);
14304 return DAG.getZExtOrTrunc(NewOp, dl, VT);
14307 // The non-AVX512 code below works under the assumption that source and
14308 // destination types are the same.
14309 assert((Subtarget->hasAVX512() || (VT == VTOp0)) &&
14310 "Value types for source and destination must be the same!");
14312 // Break 256-bit integer vector compare into smaller ones.
14313 if (VT.is256BitVector() && !Subtarget->hasInt256())
14314 return Lower256IntVSETCC(Op, DAG);
14316 EVT OpVT = Op1.getValueType();
14317 if (OpVT.getVectorElementType() == MVT::i1)
14318 return LowerBoolVSETCC_AVX512(Op, DAG);
14320 bool MaskResult = (VT.getVectorElementType() == MVT::i1);
14321 if (Subtarget->hasAVX512()) {
14322 if (Op1.getValueType().is512BitVector() ||
14323 (Subtarget->hasBWI() && Subtarget->hasVLX()) ||
14324 (MaskResult && OpVT.getVectorElementType().getSizeInBits() >= 32))
14325 return LowerIntVSETCC_AVX512(Op, DAG, Subtarget);
14327 // In AVX-512 architecture setcc returns mask with i1 elements,
14328 // But there is no compare instruction for i8 and i16 elements in KNL.
14329 // We are not talking about 512-bit operands in this case, these
14330 // types are illegal.
14332 (OpVT.getVectorElementType().getSizeInBits() < 32 &&
14333 OpVT.getVectorElementType().getSizeInBits() >= 8))
14334 return DAG.getNode(ISD::TRUNCATE, dl, VT,
14335 DAG.getNode(ISD::SETCC, dl, OpVT, Op0, Op1, CC));
14338 // Lower using XOP integer comparisons.
14339 if ((VT == MVT::v16i8 || VT == MVT::v8i16 ||
14340 VT == MVT::v4i32 || VT == MVT::v2i64) && Subtarget->hasXOP()) {
14341 // Translate compare code to XOP PCOM compare mode.
14342 unsigned CmpMode = 0;
14343 switch (SetCCOpcode) {
14344 default: llvm_unreachable("Unexpected SETCC condition");
14346 case ISD::SETLT: CmpMode = 0x00; break;
14348 case ISD::SETLE: CmpMode = 0x01; break;
14350 case ISD::SETGT: CmpMode = 0x02; break;
14352 case ISD::SETGE: CmpMode = 0x03; break;
14353 case ISD::SETEQ: CmpMode = 0x04; break;
14354 case ISD::SETNE: CmpMode = 0x05; break;
14357 // Are we comparing unsigned or signed integers?
14358 unsigned Opc = ISD::isUnsignedIntSetCC(SetCCOpcode)
14359 ? X86ISD::VPCOMU : X86ISD::VPCOM;
14361 return DAG.getNode(Opc, dl, VT, Op0, Op1,
14362 DAG.getConstant(CmpMode, dl, MVT::i8));
14365 // We are handling one of the integer comparisons here. Since SSE only has
14366 // GT and EQ comparisons for integer, swapping operands and multiple
14367 // operations may be required for some comparisons.
14369 bool Swap = false, Invert = false, FlipSigns = false, MinMax = false;
14370 bool Subus = false;
14372 switch (SetCCOpcode) {
14373 default: llvm_unreachable("Unexpected SETCC condition");
14374 case ISD::SETNE: Invert = true;
14375 case ISD::SETEQ: Opc = X86ISD::PCMPEQ; break;
14376 case ISD::SETLT: Swap = true;
14377 case ISD::SETGT: Opc = X86ISD::PCMPGT; break;
14378 case ISD::SETGE: Swap = true;
14379 case ISD::SETLE: Opc = X86ISD::PCMPGT;
14380 Invert = true; break;
14381 case ISD::SETULT: Swap = true;
14382 case ISD::SETUGT: Opc = X86ISD::PCMPGT;
14383 FlipSigns = true; break;
14384 case ISD::SETUGE: Swap = true;
14385 case ISD::SETULE: Opc = X86ISD::PCMPGT;
14386 FlipSigns = true; Invert = true; break;
14389 // Special case: Use min/max operations for SETULE/SETUGE
14390 MVT VET = VT.getVectorElementType();
14392 (Subtarget->hasSSE41() && (VET >= MVT::i8 && VET <= MVT::i32))
14393 || (Subtarget->hasSSE2() && (VET == MVT::i8));
14396 switch (SetCCOpcode) {
14398 case ISD::SETULE: Opc = ISD::UMIN; MinMax = true; break;
14399 case ISD::SETUGE: Opc = ISD::UMAX; MinMax = true; break;
14402 if (MinMax) { Swap = false; Invert = false; FlipSigns = false; }
14405 bool hasSubus = Subtarget->hasSSE2() && (VET == MVT::i8 || VET == MVT::i16);
14406 if (!MinMax && hasSubus) {
14407 // As another special case, use PSUBUS[BW] when it's profitable. E.g. for
14409 // t = psubus Op0, Op1
14410 // pcmpeq t, <0..0>
14411 switch (SetCCOpcode) {
14413 case ISD::SETULT: {
14414 // If the comparison is against a constant we can turn this into a
14415 // setule. With psubus, setule does not require a swap. This is
14416 // beneficial because the constant in the register is no longer
14417 // destructed as the destination so it can be hoisted out of a loop.
14418 // Only do this pre-AVX since vpcmp* is no longer destructive.
14419 if (Subtarget->hasAVX())
14421 SDValue ULEOp1 = ChangeVSETULTtoVSETULE(dl, Op1, DAG);
14422 if (ULEOp1.getNode()) {
14424 Subus = true; Invert = false; Swap = false;
14428 // Psubus is better than flip-sign because it requires no inversion.
14429 case ISD::SETUGE: Subus = true; Invert = false; Swap = true; break;
14430 case ISD::SETULE: Subus = true; Invert = false; Swap = false; break;
14434 Opc = X86ISD::SUBUS;
14440 std::swap(Op0, Op1);
14442 // Check that the operation in question is available (most are plain SSE2,
14443 // but PCMPGTQ and PCMPEQQ have different requirements).
14444 if (VT == MVT::v2i64) {
14445 if (Opc == X86ISD::PCMPGT && !Subtarget->hasSSE42()) {
14446 assert(Subtarget->hasSSE2() && "Don't know how to lower!");
14448 // First cast everything to the right type.
14449 Op0 = DAG.getBitcast(MVT::v4i32, Op0);
14450 Op1 = DAG.getBitcast(MVT::v4i32, Op1);
14452 // Since SSE has no unsigned integer comparisons, we need to flip the sign
14453 // bits of the inputs before performing those operations. The lower
14454 // compare is always unsigned.
14457 SB = DAG.getConstant(0x80000000U, dl, MVT::v4i32);
14459 SDValue Sign = DAG.getConstant(0x80000000U, dl, MVT::i32);
14460 SDValue Zero = DAG.getConstant(0x00000000U, dl, MVT::i32);
14461 SB = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32,
14462 Sign, Zero, Sign, Zero);
14464 Op0 = DAG.getNode(ISD::XOR, dl, MVT::v4i32, Op0, SB);
14465 Op1 = DAG.getNode(ISD::XOR, dl, MVT::v4i32, Op1, SB);
14467 // Emulate PCMPGTQ with (hi1 > hi2) | ((hi1 == hi2) & (lo1 > lo2))
14468 SDValue GT = DAG.getNode(X86ISD::PCMPGT, dl, MVT::v4i32, Op0, Op1);
14469 SDValue EQ = DAG.getNode(X86ISD::PCMPEQ, dl, MVT::v4i32, Op0, Op1);
14471 // Create masks for only the low parts/high parts of the 64 bit integers.
14472 static const int MaskHi[] = { 1, 1, 3, 3 };
14473 static const int MaskLo[] = { 0, 0, 2, 2 };
14474 SDValue EQHi = DAG.getVectorShuffle(MVT::v4i32, dl, EQ, EQ, MaskHi);
14475 SDValue GTLo = DAG.getVectorShuffle(MVT::v4i32, dl, GT, GT, MaskLo);
14476 SDValue GTHi = DAG.getVectorShuffle(MVT::v4i32, dl, GT, GT, MaskHi);
14478 SDValue Result = DAG.getNode(ISD::AND, dl, MVT::v4i32, EQHi, GTLo);
14479 Result = DAG.getNode(ISD::OR, dl, MVT::v4i32, Result, GTHi);
14482 Result = DAG.getNOT(dl, Result, MVT::v4i32);
14484 return DAG.getBitcast(VT, Result);
14487 if (Opc == X86ISD::PCMPEQ && !Subtarget->hasSSE41()) {
14488 // If pcmpeqq is missing but pcmpeqd is available synthesize pcmpeqq with
14489 // pcmpeqd + pshufd + pand.
14490 assert(Subtarget->hasSSE2() && !FlipSigns && "Don't know how to lower!");
14492 // First cast everything to the right type.
14493 Op0 = DAG.getBitcast(MVT::v4i32, Op0);
14494 Op1 = DAG.getBitcast(MVT::v4i32, Op1);
14497 SDValue Result = DAG.getNode(Opc, dl, MVT::v4i32, Op0, Op1);
14499 // Make sure the lower and upper halves are both all-ones.
14500 static const int Mask[] = { 1, 0, 3, 2 };
14501 SDValue Shuf = DAG.getVectorShuffle(MVT::v4i32, dl, Result, Result, Mask);
14502 Result = DAG.getNode(ISD::AND, dl, MVT::v4i32, Result, Shuf);
14505 Result = DAG.getNOT(dl, Result, MVT::v4i32);
14507 return DAG.getBitcast(VT, Result);
14511 // Since SSE has no unsigned integer comparisons, we need to flip the sign
14512 // bits of the inputs before performing those operations.
14514 EVT EltVT = VT.getVectorElementType();
14515 SDValue SB = DAG.getConstant(APInt::getSignBit(EltVT.getSizeInBits()), dl,
14517 Op0 = DAG.getNode(ISD::XOR, dl, VT, Op0, SB);
14518 Op1 = DAG.getNode(ISD::XOR, dl, VT, Op1, SB);
14521 SDValue Result = DAG.getNode(Opc, dl, VT, Op0, Op1);
14523 // If the logical-not of the result is required, perform that now.
14525 Result = DAG.getNOT(dl, Result, VT);
14528 Result = DAG.getNode(X86ISD::PCMPEQ, dl, VT, Op0, Result);
14531 Result = DAG.getNode(X86ISD::PCMPEQ, dl, VT, Result,
14532 getZeroVector(VT, Subtarget, DAG, dl));
14537 SDValue X86TargetLowering::LowerSETCC(SDValue Op, SelectionDAG &DAG) const {
14539 MVT VT = Op.getSimpleValueType();
14541 if (VT.isVector()) return LowerVSETCC(Op, Subtarget, DAG);
14543 assert(((!Subtarget->hasAVX512() && VT == MVT::i8) || (VT == MVT::i1))
14544 && "SetCC type must be 8-bit or 1-bit integer");
14545 SDValue Op0 = Op.getOperand(0);
14546 SDValue Op1 = Op.getOperand(1);
14548 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
14550 // Optimize to BT if possible.
14551 // Lower (X & (1 << N)) == 0 to BT(X, N).
14552 // Lower ((X >>u N) & 1) != 0 to BT(X, N).
14553 // Lower ((X >>s N) & 1) != 0 to BT(X, N).
14554 if (Op0.getOpcode() == ISD::AND && Op0.hasOneUse() &&
14555 Op1.getOpcode() == ISD::Constant &&
14556 cast<ConstantSDNode>(Op1)->isNullValue() &&
14557 (CC == ISD::SETEQ || CC == ISD::SETNE)) {
14558 SDValue NewSetCC = LowerToBT(Op0, CC, dl, DAG);
14559 if (NewSetCC.getNode()) {
14561 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i1, NewSetCC);
14566 // Look for X == 0, X == 1, X != 0, or X != 1. We can simplify some forms of
14568 if (Op1.getOpcode() == ISD::Constant &&
14569 (cast<ConstantSDNode>(Op1)->getZExtValue() == 1 ||
14570 cast<ConstantSDNode>(Op1)->isNullValue()) &&
14571 (CC == ISD::SETEQ || CC == ISD::SETNE)) {
14573 // If the input is a setcc, then reuse the input setcc or use a new one with
14574 // the inverted condition.
14575 if (Op0.getOpcode() == X86ISD::SETCC) {
14576 X86::CondCode CCode = (X86::CondCode)Op0.getConstantOperandVal(0);
14577 bool Invert = (CC == ISD::SETNE) ^
14578 cast<ConstantSDNode>(Op1)->isNullValue();
14582 CCode = X86::GetOppositeBranchCondition(CCode);
14583 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
14584 DAG.getConstant(CCode, dl, MVT::i8),
14585 Op0.getOperand(1));
14587 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i1, SetCC);
14591 if ((Op0.getValueType() == MVT::i1) && (Op1.getOpcode() == ISD::Constant) &&
14592 (cast<ConstantSDNode>(Op1)->getZExtValue() == 1) &&
14593 (CC == ISD::SETEQ || CC == ISD::SETNE)) {
14595 ISD::CondCode NewCC = ISD::getSetCCInverse(CC, true);
14596 return DAG.getSetCC(dl, VT, Op0, DAG.getConstant(0, dl, MVT::i1), NewCC);
14599 bool isFP = Op1.getSimpleValueType().isFloatingPoint();
14600 unsigned X86CC = TranslateX86CC(CC, dl, isFP, Op0, Op1, DAG);
14601 if (X86CC == X86::COND_INVALID)
14604 SDValue EFLAGS = EmitCmp(Op0, Op1, X86CC, dl, DAG);
14605 EFLAGS = ConvertCmpIfNecessary(EFLAGS, DAG);
14606 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
14607 DAG.getConstant(X86CC, dl, MVT::i8), EFLAGS);
14609 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i1, SetCC);
14613 // isX86LogicalCmp - Return true if opcode is a X86 logical comparison.
14614 static bool isX86LogicalCmp(SDValue Op) {
14615 unsigned Opc = Op.getNode()->getOpcode();
14616 if (Opc == X86ISD::CMP || Opc == X86ISD::COMI || Opc == X86ISD::UCOMI ||
14617 Opc == X86ISD::SAHF)
14619 if (Op.getResNo() == 1 &&
14620 (Opc == X86ISD::ADD ||
14621 Opc == X86ISD::SUB ||
14622 Opc == X86ISD::ADC ||
14623 Opc == X86ISD::SBB ||
14624 Opc == X86ISD::SMUL ||
14625 Opc == X86ISD::UMUL ||
14626 Opc == X86ISD::INC ||
14627 Opc == X86ISD::DEC ||
14628 Opc == X86ISD::OR ||
14629 Opc == X86ISD::XOR ||
14630 Opc == X86ISD::AND))
14633 if (Op.getResNo() == 2 && Opc == X86ISD::UMUL)
14639 static bool isTruncWithZeroHighBitsInput(SDValue V, SelectionDAG &DAG) {
14640 if (V.getOpcode() != ISD::TRUNCATE)
14643 SDValue VOp0 = V.getOperand(0);
14644 unsigned InBits = VOp0.getValueSizeInBits();
14645 unsigned Bits = V.getValueSizeInBits();
14646 return DAG.MaskedValueIsZero(VOp0, APInt::getHighBitsSet(InBits,InBits-Bits));
14649 SDValue X86TargetLowering::LowerSELECT(SDValue Op, SelectionDAG &DAG) const {
14650 bool addTest = true;
14651 SDValue Cond = Op.getOperand(0);
14652 SDValue Op1 = Op.getOperand(1);
14653 SDValue Op2 = Op.getOperand(2);
14655 EVT VT = Op1.getValueType();
14658 // Lower FP selects into a CMP/AND/ANDN/OR sequence when the necessary SSE ops
14659 // are available or VBLENDV if AVX is available.
14660 // Otherwise FP cmovs get lowered into a less efficient branch sequence later.
14661 if (Cond.getOpcode() == ISD::SETCC &&
14662 ((Subtarget->hasSSE2() && (VT == MVT::f32 || VT == MVT::f64)) ||
14663 (Subtarget->hasSSE1() && VT == MVT::f32)) &&
14664 VT == Cond.getOperand(0).getValueType() && Cond->hasOneUse()) {
14665 SDValue CondOp0 = Cond.getOperand(0), CondOp1 = Cond.getOperand(1);
14666 int SSECC = translateX86FSETCC(
14667 cast<CondCodeSDNode>(Cond.getOperand(2))->get(), CondOp0, CondOp1);
14670 if (Subtarget->hasAVX512()) {
14671 SDValue Cmp = DAG.getNode(X86ISD::FSETCC, DL, MVT::i1, CondOp0, CondOp1,
14672 DAG.getConstant(SSECC, DL, MVT::i8));
14673 return DAG.getNode(X86ISD::SELECT, DL, VT, Cmp, Op1, Op2);
14676 SDValue Cmp = DAG.getNode(X86ISD::FSETCC, DL, VT, CondOp0, CondOp1,
14677 DAG.getConstant(SSECC, DL, MVT::i8));
14679 // If we have AVX, we can use a variable vector select (VBLENDV) instead
14680 // of 3 logic instructions for size savings and potentially speed.
14681 // Unfortunately, there is no scalar form of VBLENDV.
14683 // If either operand is a constant, don't try this. We can expect to
14684 // optimize away at least one of the logic instructions later in that
14685 // case, so that sequence would be faster than a variable blend.
14687 // BLENDV was introduced with SSE 4.1, but the 2 register form implicitly
14688 // uses XMM0 as the selection register. That may need just as many
14689 // instructions as the AND/ANDN/OR sequence due to register moves, so
14692 if (Subtarget->hasAVX() &&
14693 !isa<ConstantFPSDNode>(Op1) && !isa<ConstantFPSDNode>(Op2)) {
14695 // Convert to vectors, do a VSELECT, and convert back to scalar.
14696 // All of the conversions should be optimized away.
14698 EVT VecVT = VT == MVT::f32 ? MVT::v4f32 : MVT::v2f64;
14699 SDValue VOp1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, VecVT, Op1);
14700 SDValue VOp2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, VecVT, Op2);
14701 SDValue VCmp = DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, VecVT, Cmp);
14703 EVT VCmpVT = VT == MVT::f32 ? MVT::v4i32 : MVT::v2i64;
14704 VCmp = DAG.getBitcast(VCmpVT, VCmp);
14706 SDValue VSel = DAG.getNode(ISD::VSELECT, DL, VecVT, VCmp, VOp1, VOp2);
14708 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, VT,
14709 VSel, DAG.getIntPtrConstant(0, DL));
14711 SDValue AndN = DAG.getNode(X86ISD::FANDN, DL, VT, Cmp, Op2);
14712 SDValue And = DAG.getNode(X86ISD::FAND, DL, VT, Cmp, Op1);
14713 return DAG.getNode(X86ISD::FOR, DL, VT, AndN, And);
14717 if (VT.isVector() && VT.getScalarType() == MVT::i1) {
14719 if (ISD::isBuildVectorOfConstantSDNodes(Op1.getNode()))
14720 Op1Scalar = ConvertI1VectorToInteger(Op1, DAG);
14721 else if (Op1.getOpcode() == ISD::BITCAST && Op1.getOperand(0))
14722 Op1Scalar = Op1.getOperand(0);
14724 if (ISD::isBuildVectorOfConstantSDNodes(Op2.getNode()))
14725 Op2Scalar = ConvertI1VectorToInteger(Op2, DAG);
14726 else if (Op2.getOpcode() == ISD::BITCAST && Op2.getOperand(0))
14727 Op2Scalar = Op2.getOperand(0);
14728 if (Op1Scalar.getNode() && Op2Scalar.getNode()) {
14729 SDValue newSelect = DAG.getNode(ISD::SELECT, DL,
14730 Op1Scalar.getValueType(),
14731 Cond, Op1Scalar, Op2Scalar);
14732 if (newSelect.getValueSizeInBits() == VT.getSizeInBits())
14733 return DAG.getBitcast(VT, newSelect);
14734 SDValue ExtVec = DAG.getBitcast(MVT::v8i1, newSelect);
14735 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, VT, ExtVec,
14736 DAG.getIntPtrConstant(0, DL));
14740 if (VT == MVT::v4i1 || VT == MVT::v2i1) {
14741 SDValue zeroConst = DAG.getIntPtrConstant(0, DL);
14742 Op1 = DAG.getNode(ISD::INSERT_SUBVECTOR, DL, MVT::v8i1,
14743 DAG.getUNDEF(MVT::v8i1), Op1, zeroConst);
14744 Op2 = DAG.getNode(ISD::INSERT_SUBVECTOR, DL, MVT::v8i1,
14745 DAG.getUNDEF(MVT::v8i1), Op2, zeroConst);
14746 SDValue newSelect = DAG.getNode(ISD::SELECT, DL, MVT::v8i1,
14748 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, VT, newSelect, zeroConst);
14751 if (Cond.getOpcode() == ISD::SETCC) {
14752 SDValue NewCond = LowerSETCC(Cond, DAG);
14753 if (NewCond.getNode())
14757 // (select (x == 0), -1, y) -> (sign_bit (x - 1)) | y
14758 // (select (x == 0), y, -1) -> ~(sign_bit (x - 1)) | y
14759 // (select (x != 0), y, -1) -> (sign_bit (x - 1)) | y
14760 // (select (x != 0), -1, y) -> ~(sign_bit (x - 1)) | y
14761 if (Cond.getOpcode() == X86ISD::SETCC &&
14762 Cond.getOperand(1).getOpcode() == X86ISD::CMP &&
14763 isZero(Cond.getOperand(1).getOperand(1))) {
14764 SDValue Cmp = Cond.getOperand(1);
14766 unsigned CondCode =cast<ConstantSDNode>(Cond.getOperand(0))->getZExtValue();
14768 if ((isAllOnes(Op1) || isAllOnes(Op2)) &&
14769 (CondCode == X86::COND_E || CondCode == X86::COND_NE)) {
14770 SDValue Y = isAllOnes(Op2) ? Op1 : Op2;
14772 SDValue CmpOp0 = Cmp.getOperand(0);
14773 // Apply further optimizations for special cases
14774 // (select (x != 0), -1, 0) -> neg & sbb
14775 // (select (x == 0), 0, -1) -> neg & sbb
14776 if (ConstantSDNode *YC = dyn_cast<ConstantSDNode>(Y))
14777 if (YC->isNullValue() &&
14778 (isAllOnes(Op1) == (CondCode == X86::COND_NE))) {
14779 SDVTList VTs = DAG.getVTList(CmpOp0.getValueType(), MVT::i32);
14780 SDValue Neg = DAG.getNode(X86ISD::SUB, DL, VTs,
14781 DAG.getConstant(0, DL,
14782 CmpOp0.getValueType()),
14784 SDValue Res = DAG.getNode(X86ISD::SETCC_CARRY, DL, Op.getValueType(),
14785 DAG.getConstant(X86::COND_B, DL, MVT::i8),
14786 SDValue(Neg.getNode(), 1));
14790 Cmp = DAG.getNode(X86ISD::CMP, DL, MVT::i32,
14791 CmpOp0, DAG.getConstant(1, DL, CmpOp0.getValueType()));
14792 Cmp = ConvertCmpIfNecessary(Cmp, DAG);
14794 SDValue Res = // Res = 0 or -1.
14795 DAG.getNode(X86ISD::SETCC_CARRY, DL, Op.getValueType(),
14796 DAG.getConstant(X86::COND_B, DL, MVT::i8), Cmp);
14798 if (isAllOnes(Op1) != (CondCode == X86::COND_E))
14799 Res = DAG.getNOT(DL, Res, Res.getValueType());
14801 ConstantSDNode *N2C = dyn_cast<ConstantSDNode>(Op2);
14802 if (!N2C || !N2C->isNullValue())
14803 Res = DAG.getNode(ISD::OR, DL, Res.getValueType(), Res, Y);
14808 // Look past (and (setcc_carry (cmp ...)), 1).
14809 if (Cond.getOpcode() == ISD::AND &&
14810 Cond.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY) {
14811 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Cond.getOperand(1));
14812 if (C && C->getAPIntValue() == 1)
14813 Cond = Cond.getOperand(0);
14816 // If condition flag is set by a X86ISD::CMP, then use it as the condition
14817 // setting operand in place of the X86ISD::SETCC.
14818 unsigned CondOpcode = Cond.getOpcode();
14819 if (CondOpcode == X86ISD::SETCC ||
14820 CondOpcode == X86ISD::SETCC_CARRY) {
14821 CC = Cond.getOperand(0);
14823 SDValue Cmp = Cond.getOperand(1);
14824 unsigned Opc = Cmp.getOpcode();
14825 MVT VT = Op.getSimpleValueType();
14827 bool IllegalFPCMov = false;
14828 if (VT.isFloatingPoint() && !VT.isVector() &&
14829 !isScalarFPTypeInSSEReg(VT)) // FPStack?
14830 IllegalFPCMov = !hasFPCMov(cast<ConstantSDNode>(CC)->getSExtValue());
14832 if ((isX86LogicalCmp(Cmp) && !IllegalFPCMov) ||
14833 Opc == X86ISD::BT) { // FIXME
14837 } else if (CondOpcode == ISD::USUBO || CondOpcode == ISD::SSUBO ||
14838 CondOpcode == ISD::UADDO || CondOpcode == ISD::SADDO ||
14839 ((CondOpcode == ISD::UMULO || CondOpcode == ISD::SMULO) &&
14840 Cond.getOperand(0).getValueType() != MVT::i8)) {
14841 SDValue LHS = Cond.getOperand(0);
14842 SDValue RHS = Cond.getOperand(1);
14843 unsigned X86Opcode;
14846 switch (CondOpcode) {
14847 case ISD::UADDO: X86Opcode = X86ISD::ADD; X86Cond = X86::COND_B; break;
14848 case ISD::SADDO: X86Opcode = X86ISD::ADD; X86Cond = X86::COND_O; break;
14849 case ISD::USUBO: X86Opcode = X86ISD::SUB; X86Cond = X86::COND_B; break;
14850 case ISD::SSUBO: X86Opcode = X86ISD::SUB; X86Cond = X86::COND_O; break;
14851 case ISD::UMULO: X86Opcode = X86ISD::UMUL; X86Cond = X86::COND_O; break;
14852 case ISD::SMULO: X86Opcode = X86ISD::SMUL; X86Cond = X86::COND_O; break;
14853 default: llvm_unreachable("unexpected overflowing operator");
14855 if (CondOpcode == ISD::UMULO)
14856 VTs = DAG.getVTList(LHS.getValueType(), LHS.getValueType(),
14859 VTs = DAG.getVTList(LHS.getValueType(), MVT::i32);
14861 SDValue X86Op = DAG.getNode(X86Opcode, DL, VTs, LHS, RHS);
14863 if (CondOpcode == ISD::UMULO)
14864 Cond = X86Op.getValue(2);
14866 Cond = X86Op.getValue(1);
14868 CC = DAG.getConstant(X86Cond, DL, MVT::i8);
14873 // Look past the truncate if the high bits are known zero.
14874 if (isTruncWithZeroHighBitsInput(Cond, DAG))
14875 Cond = Cond.getOperand(0);
14877 // We know the result of AND is compared against zero. Try to match
14879 if (Cond.getOpcode() == ISD::AND && Cond.hasOneUse()) {
14880 SDValue NewSetCC = LowerToBT(Cond, ISD::SETNE, DL, DAG);
14881 if (NewSetCC.getNode()) {
14882 CC = NewSetCC.getOperand(0);
14883 Cond = NewSetCC.getOperand(1);
14890 CC = DAG.getConstant(X86::COND_NE, DL, MVT::i8);
14891 Cond = EmitTest(Cond, X86::COND_NE, DL, DAG);
14894 // a < b ? -1 : 0 -> RES = ~setcc_carry
14895 // a < b ? 0 : -1 -> RES = setcc_carry
14896 // a >= b ? -1 : 0 -> RES = setcc_carry
14897 // a >= b ? 0 : -1 -> RES = ~setcc_carry
14898 if (Cond.getOpcode() == X86ISD::SUB) {
14899 Cond = ConvertCmpIfNecessary(Cond, DAG);
14900 unsigned CondCode = cast<ConstantSDNode>(CC)->getZExtValue();
14902 if ((CondCode == X86::COND_AE || CondCode == X86::COND_B) &&
14903 (isAllOnes(Op1) || isAllOnes(Op2)) && (isZero(Op1) || isZero(Op2))) {
14904 SDValue Res = DAG.getNode(X86ISD::SETCC_CARRY, DL, Op.getValueType(),
14905 DAG.getConstant(X86::COND_B, DL, MVT::i8),
14907 if (isAllOnes(Op1) != (CondCode == X86::COND_B))
14908 return DAG.getNOT(DL, Res, Res.getValueType());
14913 // X86 doesn't have an i8 cmov. If both operands are the result of a truncate
14914 // widen the cmov and push the truncate through. This avoids introducing a new
14915 // branch during isel and doesn't add any extensions.
14916 if (Op.getValueType() == MVT::i8 &&
14917 Op1.getOpcode() == ISD::TRUNCATE && Op2.getOpcode() == ISD::TRUNCATE) {
14918 SDValue T1 = Op1.getOperand(0), T2 = Op2.getOperand(0);
14919 if (T1.getValueType() == T2.getValueType() &&
14920 // Blacklist CopyFromReg to avoid partial register stalls.
14921 T1.getOpcode() != ISD::CopyFromReg && T2.getOpcode()!=ISD::CopyFromReg){
14922 SDVTList VTs = DAG.getVTList(T1.getValueType(), MVT::Glue);
14923 SDValue Cmov = DAG.getNode(X86ISD::CMOV, DL, VTs, T2, T1, CC, Cond);
14924 return DAG.getNode(ISD::TRUNCATE, DL, Op.getValueType(), Cmov);
14928 // X86ISD::CMOV means set the result (which is operand 1) to the RHS if
14929 // condition is true.
14930 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::Glue);
14931 SDValue Ops[] = { Op2, Op1, CC, Cond };
14932 return DAG.getNode(X86ISD::CMOV, DL, VTs, Ops);
14935 static SDValue LowerSIGN_EXTEND_AVX512(SDValue Op,
14936 const X86Subtarget *Subtarget,
14937 SelectionDAG &DAG) {
14938 MVT VT = Op->getSimpleValueType(0);
14939 SDValue In = Op->getOperand(0);
14940 MVT InVT = In.getSimpleValueType();
14941 MVT VTElt = VT.getVectorElementType();
14942 MVT InVTElt = InVT.getVectorElementType();
14946 if ((InVTElt == MVT::i1) &&
14947 (((Subtarget->hasBWI() && Subtarget->hasVLX() &&
14948 VT.getSizeInBits() <= 256 && VTElt.getSizeInBits() <= 16)) ||
14950 ((Subtarget->hasBWI() && VT.is512BitVector() &&
14951 VTElt.getSizeInBits() <= 16)) ||
14953 ((Subtarget->hasDQI() && Subtarget->hasVLX() &&
14954 VT.getSizeInBits() <= 256 && VTElt.getSizeInBits() >= 32)) ||
14956 ((Subtarget->hasDQI() && VT.is512BitVector() &&
14957 VTElt.getSizeInBits() >= 32))))
14958 return DAG.getNode(X86ISD::VSEXT, dl, VT, In);
14960 unsigned int NumElts = VT.getVectorNumElements();
14962 if (NumElts != 8 && NumElts != 16 && !Subtarget->hasBWI())
14965 if (VT.is512BitVector() && InVT.getVectorElementType() != MVT::i1) {
14966 if (In.getOpcode() == X86ISD::VSEXT || In.getOpcode() == X86ISD::VZEXT)
14967 return DAG.getNode(In.getOpcode(), dl, VT, In.getOperand(0));
14968 return DAG.getNode(X86ISD::VSEXT, dl, VT, In);
14971 assert (InVT.getVectorElementType() == MVT::i1 && "Unexpected vector type");
14972 MVT ExtVT = NumElts == 8 ? MVT::v8i64 : MVT::v16i32;
14974 DAG.getConstant(APInt::getAllOnesValue(ExtVT.getScalarSizeInBits()), dl,
14977 DAG.getConstant(APInt::getNullValue(ExtVT.getScalarSizeInBits()), dl, ExtVT);
14979 SDValue V = DAG.getNode(ISD::VSELECT, dl, ExtVT, In, NegOne, Zero);
14980 if (VT.is512BitVector())
14982 return DAG.getNode(X86ISD::VTRUNC, dl, VT, V);
14985 static SDValue LowerSIGN_EXTEND_VECTOR_INREG(SDValue Op,
14986 const X86Subtarget *Subtarget,
14987 SelectionDAG &DAG) {
14988 SDValue In = Op->getOperand(0);
14989 MVT VT = Op->getSimpleValueType(0);
14990 MVT InVT = In.getSimpleValueType();
14991 assert(VT.getSizeInBits() == InVT.getSizeInBits());
14993 MVT InSVT = InVT.getScalarType();
14994 assert(VT.getScalarType().getScalarSizeInBits() > InSVT.getScalarSizeInBits());
14996 if (VT != MVT::v2i64 && VT != MVT::v4i32 && VT != MVT::v8i16)
14998 if (InSVT != MVT::i32 && InSVT != MVT::i16 && InSVT != MVT::i8)
15003 // SSE41 targets can use the pmovsx* instructions directly.
15004 if (Subtarget->hasSSE41())
15005 return DAG.getNode(X86ISD::VSEXT, dl, VT, In);
15007 // pre-SSE41 targets unpack lower lanes and then sign-extend using SRAI.
15011 // As SRAI is only available on i16/i32 types, we expand only up to i32
15012 // and handle i64 separately.
15013 while (CurrVT != VT && CurrVT.getScalarType() != MVT::i32) {
15014 Curr = DAG.getNode(X86ISD::UNPCKL, dl, CurrVT, DAG.getUNDEF(CurrVT), Curr);
15015 MVT CurrSVT = MVT::getIntegerVT(CurrVT.getScalarSizeInBits() * 2);
15016 CurrVT = MVT::getVectorVT(CurrSVT, CurrVT.getVectorNumElements() / 2);
15017 Curr = DAG.getBitcast(CurrVT, Curr);
15020 SDValue SignExt = Curr;
15021 if (CurrVT != InVT) {
15022 unsigned SignExtShift =
15023 CurrVT.getScalarSizeInBits() - InSVT.getScalarSizeInBits();
15024 SignExt = DAG.getNode(X86ISD::VSRAI, dl, CurrVT, Curr,
15025 DAG.getConstant(SignExtShift, dl, MVT::i8));
15031 if (VT == MVT::v2i64 && CurrVT == MVT::v4i32) {
15032 SDValue Sign = DAG.getNode(X86ISD::VSRAI, dl, CurrVT, Curr,
15033 DAG.getConstant(31, dl, MVT::i8));
15034 SDValue Ext = DAG.getVectorShuffle(CurrVT, dl, SignExt, Sign, {0, 4, 1, 5});
15035 return DAG.getBitcast(VT, Ext);
15041 static SDValue LowerSIGN_EXTEND(SDValue Op, const X86Subtarget *Subtarget,
15042 SelectionDAG &DAG) {
15043 MVT VT = Op->getSimpleValueType(0);
15044 SDValue In = Op->getOperand(0);
15045 MVT InVT = In.getSimpleValueType();
15048 if (VT.is512BitVector() || InVT.getVectorElementType() == MVT::i1)
15049 return LowerSIGN_EXTEND_AVX512(Op, Subtarget, DAG);
15051 if ((VT != MVT::v4i64 || InVT != MVT::v4i32) &&
15052 (VT != MVT::v8i32 || InVT != MVT::v8i16) &&
15053 (VT != MVT::v16i16 || InVT != MVT::v16i8))
15056 if (Subtarget->hasInt256())
15057 return DAG.getNode(X86ISD::VSEXT, dl, VT, In);
15059 // Optimize vectors in AVX mode
15060 // Sign extend v8i16 to v8i32 and
15063 // Divide input vector into two parts
15064 // for v4i32 the shuffle mask will be { 0, 1, -1, -1} {2, 3, -1, -1}
15065 // use vpmovsx instruction to extend v4i32 -> v2i64; v8i16 -> v4i32
15066 // concat the vectors to original VT
15068 unsigned NumElems = InVT.getVectorNumElements();
15069 SDValue Undef = DAG.getUNDEF(InVT);
15071 SmallVector<int,8> ShufMask1(NumElems, -1);
15072 for (unsigned i = 0; i != NumElems/2; ++i)
15075 SDValue OpLo = DAG.getVectorShuffle(InVT, dl, In, Undef, &ShufMask1[0]);
15077 SmallVector<int,8> ShufMask2(NumElems, -1);
15078 for (unsigned i = 0; i != NumElems/2; ++i)
15079 ShufMask2[i] = i + NumElems/2;
15081 SDValue OpHi = DAG.getVectorShuffle(InVT, dl, In, Undef, &ShufMask2[0]);
15083 MVT HalfVT = MVT::getVectorVT(VT.getScalarType(),
15084 VT.getVectorNumElements()/2);
15086 OpLo = DAG.getNode(X86ISD::VSEXT, dl, HalfVT, OpLo);
15087 OpHi = DAG.getNode(X86ISD::VSEXT, dl, HalfVT, OpHi);
15089 return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT, OpLo, OpHi);
15092 // Lower vector extended loads using a shuffle. If SSSE3 is not available we
15093 // may emit an illegal shuffle but the expansion is still better than scalar
15094 // code. We generate X86ISD::VSEXT for SEXTLOADs if it's available, otherwise
15095 // we'll emit a shuffle and a arithmetic shift.
15096 // FIXME: Is the expansion actually better than scalar code? It doesn't seem so.
15097 // TODO: It is possible to support ZExt by zeroing the undef values during
15098 // the shuffle phase or after the shuffle.
15099 static SDValue LowerExtendedLoad(SDValue Op, const X86Subtarget *Subtarget,
15100 SelectionDAG &DAG) {
15101 MVT RegVT = Op.getSimpleValueType();
15102 assert(RegVT.isVector() && "We only custom lower vector sext loads.");
15103 assert(RegVT.isInteger() &&
15104 "We only custom lower integer vector sext loads.");
15106 // Nothing useful we can do without SSE2 shuffles.
15107 assert(Subtarget->hasSSE2() && "We only custom lower sext loads with SSE2.");
15109 LoadSDNode *Ld = cast<LoadSDNode>(Op.getNode());
15111 EVT MemVT = Ld->getMemoryVT();
15112 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
15113 unsigned RegSz = RegVT.getSizeInBits();
15115 ISD::LoadExtType Ext = Ld->getExtensionType();
15117 assert((Ext == ISD::EXTLOAD || Ext == ISD::SEXTLOAD)
15118 && "Only anyext and sext are currently implemented.");
15119 assert(MemVT != RegVT && "Cannot extend to the same type");
15120 assert(MemVT.isVector() && "Must load a vector from memory");
15122 unsigned NumElems = RegVT.getVectorNumElements();
15123 unsigned MemSz = MemVT.getSizeInBits();
15124 assert(RegSz > MemSz && "Register size must be greater than the mem size");
15126 if (Ext == ISD::SEXTLOAD && RegSz == 256 && !Subtarget->hasInt256()) {
15127 // The only way in which we have a legal 256-bit vector result but not the
15128 // integer 256-bit operations needed to directly lower a sextload is if we
15129 // have AVX1 but not AVX2. In that case, we can always emit a sextload to
15130 // a 128-bit vector and a normal sign_extend to 256-bits that should get
15131 // correctly legalized. We do this late to allow the canonical form of
15132 // sextload to persist throughout the rest of the DAG combiner -- it wants
15133 // to fold together any extensions it can, and so will fuse a sign_extend
15134 // of an sextload into a sextload targeting a wider value.
15136 if (MemSz == 128) {
15137 // Just switch this to a normal load.
15138 assert(TLI.isTypeLegal(MemVT) && "If the memory type is a 128-bit type, "
15139 "it must be a legal 128-bit vector "
15141 Load = DAG.getLoad(MemVT, dl, Ld->getChain(), Ld->getBasePtr(),
15142 Ld->getPointerInfo(), Ld->isVolatile(), Ld->isNonTemporal(),
15143 Ld->isInvariant(), Ld->getAlignment());
15145 assert(MemSz < 128 &&
15146 "Can't extend a type wider than 128 bits to a 256 bit vector!");
15147 // Do an sext load to a 128-bit vector type. We want to use the same
15148 // number of elements, but elements half as wide. This will end up being
15149 // recursively lowered by this routine, but will succeed as we definitely
15150 // have all the necessary features if we're using AVX1.
15152 EVT::getIntegerVT(*DAG.getContext(), RegVT.getScalarSizeInBits() / 2);
15153 EVT HalfVecVT = EVT::getVectorVT(*DAG.getContext(), HalfEltVT, NumElems);
15155 DAG.getExtLoad(Ext, dl, HalfVecVT, Ld->getChain(), Ld->getBasePtr(),
15156 Ld->getPointerInfo(), MemVT, Ld->isVolatile(),
15157 Ld->isNonTemporal(), Ld->isInvariant(),
15158 Ld->getAlignment());
15161 // Replace chain users with the new chain.
15162 assert(Load->getNumValues() == 2 && "Loads must carry a chain!");
15163 DAG.ReplaceAllUsesOfValueWith(SDValue(Ld, 1), Load.getValue(1));
15165 // Finally, do a normal sign-extend to the desired register.
15166 return DAG.getSExtOrTrunc(Load, dl, RegVT);
15169 // All sizes must be a power of two.
15170 assert(isPowerOf2_32(RegSz * MemSz * NumElems) &&
15171 "Non-power-of-two elements are not custom lowered!");
15173 // Attempt to load the original value using scalar loads.
15174 // Find the largest scalar type that divides the total loaded size.
15175 MVT SclrLoadTy = MVT::i8;
15176 for (MVT Tp : MVT::integer_valuetypes()) {
15177 if (TLI.isTypeLegal(Tp) && ((MemSz % Tp.getSizeInBits()) == 0)) {
15182 // On 32bit systems, we can't save 64bit integers. Try bitcasting to F64.
15183 if (TLI.isTypeLegal(MVT::f64) && SclrLoadTy.getSizeInBits() < 64 &&
15185 SclrLoadTy = MVT::f64;
15187 // Calculate the number of scalar loads that we need to perform
15188 // in order to load our vector from memory.
15189 unsigned NumLoads = MemSz / SclrLoadTy.getSizeInBits();
15191 assert((Ext != ISD::SEXTLOAD || NumLoads == 1) &&
15192 "Can only lower sext loads with a single scalar load!");
15194 unsigned loadRegZize = RegSz;
15195 if (Ext == ISD::SEXTLOAD && RegSz >= 256)
15198 // Represent our vector as a sequence of elements which are the
15199 // largest scalar that we can load.
15200 EVT LoadUnitVecVT = EVT::getVectorVT(
15201 *DAG.getContext(), SclrLoadTy, loadRegZize / SclrLoadTy.getSizeInBits());
15203 // Represent the data using the same element type that is stored in
15204 // memory. In practice, we ''widen'' MemVT.
15206 EVT::getVectorVT(*DAG.getContext(), MemVT.getScalarType(),
15207 loadRegZize / MemVT.getScalarType().getSizeInBits());
15209 assert(WideVecVT.getSizeInBits() == LoadUnitVecVT.getSizeInBits() &&
15210 "Invalid vector type");
15212 // We can't shuffle using an illegal type.
15213 assert(TLI.isTypeLegal(WideVecVT) &&
15214 "We only lower types that form legal widened vector types");
15216 SmallVector<SDValue, 8> Chains;
15217 SDValue Ptr = Ld->getBasePtr();
15218 SDValue Increment = DAG.getConstant(SclrLoadTy.getSizeInBits() / 8, dl,
15219 TLI.getPointerTy(DAG.getDataLayout()));
15220 SDValue Res = DAG.getUNDEF(LoadUnitVecVT);
15222 for (unsigned i = 0; i < NumLoads; ++i) {
15223 // Perform a single load.
15224 SDValue ScalarLoad =
15225 DAG.getLoad(SclrLoadTy, dl, Ld->getChain(), Ptr, Ld->getPointerInfo(),
15226 Ld->isVolatile(), Ld->isNonTemporal(), Ld->isInvariant(),
15227 Ld->getAlignment());
15228 Chains.push_back(ScalarLoad.getValue(1));
15229 // Create the first element type using SCALAR_TO_VECTOR in order to avoid
15230 // another round of DAGCombining.
15232 Res = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, LoadUnitVecVT, ScalarLoad);
15234 Res = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, LoadUnitVecVT, Res,
15235 ScalarLoad, DAG.getIntPtrConstant(i, dl));
15237 Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr, Increment);
15240 SDValue TF = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Chains);
15242 // Bitcast the loaded value to a vector of the original element type, in
15243 // the size of the target vector type.
15244 SDValue SlicedVec = DAG.getBitcast(WideVecVT, Res);
15245 unsigned SizeRatio = RegSz / MemSz;
15247 if (Ext == ISD::SEXTLOAD) {
15248 // If we have SSE4.1, we can directly emit a VSEXT node.
15249 if (Subtarget->hasSSE41()) {
15250 SDValue Sext = DAG.getNode(X86ISD::VSEXT, dl, RegVT, SlicedVec);
15251 DAG.ReplaceAllUsesOfValueWith(SDValue(Ld, 1), TF);
15255 // Otherwise we'll use SIGN_EXTEND_VECTOR_INREG to sign extend the lowest
15257 assert(TLI.isOperationLegalOrCustom(ISD::SIGN_EXTEND_VECTOR_INREG, RegVT) &&
15258 "We can't implement a sext load without SIGN_EXTEND_VECTOR_INREG!");
15260 SDValue Shuff = DAG.getSignExtendVectorInReg(SlicedVec, dl, RegVT);
15261 DAG.ReplaceAllUsesOfValueWith(SDValue(Ld, 1), TF);
15265 // Redistribute the loaded elements into the different locations.
15266 SmallVector<int, 16> ShuffleVec(NumElems * SizeRatio, -1);
15267 for (unsigned i = 0; i != NumElems; ++i)
15268 ShuffleVec[i * SizeRatio] = i;
15270 SDValue Shuff = DAG.getVectorShuffle(WideVecVT, dl, SlicedVec,
15271 DAG.getUNDEF(WideVecVT), &ShuffleVec[0]);
15273 // Bitcast to the requested type.
15274 Shuff = DAG.getBitcast(RegVT, Shuff);
15275 DAG.ReplaceAllUsesOfValueWith(SDValue(Ld, 1), TF);
15279 // isAndOrOfSingleUseSetCCs - Return true if node is an ISD::AND or
15280 // ISD::OR of two X86ISD::SETCC nodes each of which has no other use apart
15281 // from the AND / OR.
15282 static bool isAndOrOfSetCCs(SDValue Op, unsigned &Opc) {
15283 Opc = Op.getOpcode();
15284 if (Opc != ISD::OR && Opc != ISD::AND)
15286 return (Op.getOperand(0).getOpcode() == X86ISD::SETCC &&
15287 Op.getOperand(0).hasOneUse() &&
15288 Op.getOperand(1).getOpcode() == X86ISD::SETCC &&
15289 Op.getOperand(1).hasOneUse());
15292 // isXor1OfSetCC - Return true if node is an ISD::XOR of a X86ISD::SETCC and
15293 // 1 and that the SETCC node has a single use.
15294 static bool isXor1OfSetCC(SDValue Op) {
15295 if (Op.getOpcode() != ISD::XOR)
15297 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(Op.getOperand(1));
15298 if (N1C && N1C->getAPIntValue() == 1) {
15299 return Op.getOperand(0).getOpcode() == X86ISD::SETCC &&
15300 Op.getOperand(0).hasOneUse();
15305 SDValue X86TargetLowering::LowerBRCOND(SDValue Op, SelectionDAG &DAG) const {
15306 bool addTest = true;
15307 SDValue Chain = Op.getOperand(0);
15308 SDValue Cond = Op.getOperand(1);
15309 SDValue Dest = Op.getOperand(2);
15312 bool Inverted = false;
15314 if (Cond.getOpcode() == ISD::SETCC) {
15315 // Check for setcc([su]{add,sub,mul}o == 0).
15316 if (cast<CondCodeSDNode>(Cond.getOperand(2))->get() == ISD::SETEQ &&
15317 isa<ConstantSDNode>(Cond.getOperand(1)) &&
15318 cast<ConstantSDNode>(Cond.getOperand(1))->isNullValue() &&
15319 Cond.getOperand(0).getResNo() == 1 &&
15320 (Cond.getOperand(0).getOpcode() == ISD::SADDO ||
15321 Cond.getOperand(0).getOpcode() == ISD::UADDO ||
15322 Cond.getOperand(0).getOpcode() == ISD::SSUBO ||
15323 Cond.getOperand(0).getOpcode() == ISD::USUBO ||
15324 Cond.getOperand(0).getOpcode() == ISD::SMULO ||
15325 Cond.getOperand(0).getOpcode() == ISD::UMULO)) {
15327 Cond = Cond.getOperand(0);
15329 SDValue NewCond = LowerSETCC(Cond, DAG);
15330 if (NewCond.getNode())
15335 // FIXME: LowerXALUO doesn't handle these!!
15336 else if (Cond.getOpcode() == X86ISD::ADD ||
15337 Cond.getOpcode() == X86ISD::SUB ||
15338 Cond.getOpcode() == X86ISD::SMUL ||
15339 Cond.getOpcode() == X86ISD::UMUL)
15340 Cond = LowerXALUO(Cond, DAG);
15343 // Look pass (and (setcc_carry (cmp ...)), 1).
15344 if (Cond.getOpcode() == ISD::AND &&
15345 Cond.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY) {
15346 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Cond.getOperand(1));
15347 if (C && C->getAPIntValue() == 1)
15348 Cond = Cond.getOperand(0);
15351 // If condition flag is set by a X86ISD::CMP, then use it as the condition
15352 // setting operand in place of the X86ISD::SETCC.
15353 unsigned CondOpcode = Cond.getOpcode();
15354 if (CondOpcode == X86ISD::SETCC ||
15355 CondOpcode == X86ISD::SETCC_CARRY) {
15356 CC = Cond.getOperand(0);
15358 SDValue Cmp = Cond.getOperand(1);
15359 unsigned Opc = Cmp.getOpcode();
15360 // FIXME: WHY THE SPECIAL CASING OF LogicalCmp??
15361 if (isX86LogicalCmp(Cmp) || Opc == X86ISD::BT) {
15365 switch (cast<ConstantSDNode>(CC)->getZExtValue()) {
15369 // These can only come from an arithmetic instruction with overflow,
15370 // e.g. SADDO, UADDO.
15371 Cond = Cond.getNode()->getOperand(1);
15377 CondOpcode = Cond.getOpcode();
15378 if (CondOpcode == ISD::UADDO || CondOpcode == ISD::SADDO ||
15379 CondOpcode == ISD::USUBO || CondOpcode == ISD::SSUBO ||
15380 ((CondOpcode == ISD::UMULO || CondOpcode == ISD::SMULO) &&
15381 Cond.getOperand(0).getValueType() != MVT::i8)) {
15382 SDValue LHS = Cond.getOperand(0);
15383 SDValue RHS = Cond.getOperand(1);
15384 unsigned X86Opcode;
15387 // Keep this in sync with LowerXALUO, otherwise we might create redundant
15388 // instructions that can't be removed afterwards (i.e. X86ISD::ADD and
15390 switch (CondOpcode) {
15391 case ISD::UADDO: X86Opcode = X86ISD::ADD; X86Cond = X86::COND_B; break;
15393 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(RHS))
15395 X86Opcode = X86ISD::INC; X86Cond = X86::COND_O;
15398 X86Opcode = X86ISD::ADD; X86Cond = X86::COND_O; break;
15399 case ISD::USUBO: X86Opcode = X86ISD::SUB; X86Cond = X86::COND_B; break;
15401 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(RHS))
15403 X86Opcode = X86ISD::DEC; X86Cond = X86::COND_O;
15406 X86Opcode = X86ISD::SUB; X86Cond = X86::COND_O; break;
15407 case ISD::UMULO: X86Opcode = X86ISD::UMUL; X86Cond = X86::COND_O; break;
15408 case ISD::SMULO: X86Opcode = X86ISD::SMUL; X86Cond = X86::COND_O; break;
15409 default: llvm_unreachable("unexpected overflowing operator");
15412 X86Cond = X86::GetOppositeBranchCondition((X86::CondCode)X86Cond);
15413 if (CondOpcode == ISD::UMULO)
15414 VTs = DAG.getVTList(LHS.getValueType(), LHS.getValueType(),
15417 VTs = DAG.getVTList(LHS.getValueType(), MVT::i32);
15419 SDValue X86Op = DAG.getNode(X86Opcode, dl, VTs, LHS, RHS);
15421 if (CondOpcode == ISD::UMULO)
15422 Cond = X86Op.getValue(2);
15424 Cond = X86Op.getValue(1);
15426 CC = DAG.getConstant(X86Cond, dl, MVT::i8);
15430 if (Cond.hasOneUse() && isAndOrOfSetCCs(Cond, CondOpc)) {
15431 SDValue Cmp = Cond.getOperand(0).getOperand(1);
15432 if (CondOpc == ISD::OR) {
15433 // Also, recognize the pattern generated by an FCMP_UNE. We can emit
15434 // two branches instead of an explicit OR instruction with a
15436 if (Cmp == Cond.getOperand(1).getOperand(1) &&
15437 isX86LogicalCmp(Cmp)) {
15438 CC = Cond.getOperand(0).getOperand(0);
15439 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
15440 Chain, Dest, CC, Cmp);
15441 CC = Cond.getOperand(1).getOperand(0);
15445 } else { // ISD::AND
15446 // Also, recognize the pattern generated by an FCMP_OEQ. We can emit
15447 // two branches instead of an explicit AND instruction with a
15448 // separate test. However, we only do this if this block doesn't
15449 // have a fall-through edge, because this requires an explicit
15450 // jmp when the condition is false.
15451 if (Cmp == Cond.getOperand(1).getOperand(1) &&
15452 isX86LogicalCmp(Cmp) &&
15453 Op.getNode()->hasOneUse()) {
15454 X86::CondCode CCode =
15455 (X86::CondCode)Cond.getOperand(0).getConstantOperandVal(0);
15456 CCode = X86::GetOppositeBranchCondition(CCode);
15457 CC = DAG.getConstant(CCode, dl, MVT::i8);
15458 SDNode *User = *Op.getNode()->use_begin();
15459 // Look for an unconditional branch following this conditional branch.
15460 // We need this because we need to reverse the successors in order
15461 // to implement FCMP_OEQ.
15462 if (User->getOpcode() == ISD::BR) {
15463 SDValue FalseBB = User->getOperand(1);
15465 DAG.UpdateNodeOperands(User, User->getOperand(0), Dest);
15466 assert(NewBR == User);
15470 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
15471 Chain, Dest, CC, Cmp);
15472 X86::CondCode CCode =
15473 (X86::CondCode)Cond.getOperand(1).getConstantOperandVal(0);
15474 CCode = X86::GetOppositeBranchCondition(CCode);
15475 CC = DAG.getConstant(CCode, dl, MVT::i8);
15481 } else if (Cond.hasOneUse() && isXor1OfSetCC(Cond)) {
15482 // Recognize for xorb (setcc), 1 patterns. The xor inverts the condition.
15483 // It should be transformed during dag combiner except when the condition
15484 // is set by a arithmetics with overflow node.
15485 X86::CondCode CCode =
15486 (X86::CondCode)Cond.getOperand(0).getConstantOperandVal(0);
15487 CCode = X86::GetOppositeBranchCondition(CCode);
15488 CC = DAG.getConstant(CCode, dl, MVT::i8);
15489 Cond = Cond.getOperand(0).getOperand(1);
15491 } else if (Cond.getOpcode() == ISD::SETCC &&
15492 cast<CondCodeSDNode>(Cond.getOperand(2))->get() == ISD::SETOEQ) {
15493 // For FCMP_OEQ, we can emit
15494 // two branches instead of an explicit AND instruction with a
15495 // separate test. However, we only do this if this block doesn't
15496 // have a fall-through edge, because this requires an explicit
15497 // jmp when the condition is false.
15498 if (Op.getNode()->hasOneUse()) {
15499 SDNode *User = *Op.getNode()->use_begin();
15500 // Look for an unconditional branch following this conditional branch.
15501 // We need this because we need to reverse the successors in order
15502 // to implement FCMP_OEQ.
15503 if (User->getOpcode() == ISD::BR) {
15504 SDValue FalseBB = User->getOperand(1);
15506 DAG.UpdateNodeOperands(User, User->getOperand(0), Dest);
15507 assert(NewBR == User);
15511 SDValue Cmp = DAG.getNode(X86ISD::CMP, dl, MVT::i32,
15512 Cond.getOperand(0), Cond.getOperand(1));
15513 Cmp = ConvertCmpIfNecessary(Cmp, DAG);
15514 CC = DAG.getConstant(X86::COND_NE, dl, MVT::i8);
15515 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
15516 Chain, Dest, CC, Cmp);
15517 CC = DAG.getConstant(X86::COND_P, dl, MVT::i8);
15522 } else if (Cond.getOpcode() == ISD::SETCC &&
15523 cast<CondCodeSDNode>(Cond.getOperand(2))->get() == ISD::SETUNE) {
15524 // For FCMP_UNE, we can emit
15525 // two branches instead of an explicit AND instruction with a
15526 // separate test. However, we only do this if this block doesn't
15527 // have a fall-through edge, because this requires an explicit
15528 // jmp when the condition is false.
15529 if (Op.getNode()->hasOneUse()) {
15530 SDNode *User = *Op.getNode()->use_begin();
15531 // Look for an unconditional branch following this conditional branch.
15532 // We need this because we need to reverse the successors in order
15533 // to implement FCMP_UNE.
15534 if (User->getOpcode() == ISD::BR) {
15535 SDValue FalseBB = User->getOperand(1);
15537 DAG.UpdateNodeOperands(User, User->getOperand(0), Dest);
15538 assert(NewBR == User);
15541 SDValue Cmp = DAG.getNode(X86ISD::CMP, dl, MVT::i32,
15542 Cond.getOperand(0), Cond.getOperand(1));
15543 Cmp = ConvertCmpIfNecessary(Cmp, DAG);
15544 CC = DAG.getConstant(X86::COND_NE, dl, MVT::i8);
15545 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
15546 Chain, Dest, CC, Cmp);
15547 CC = DAG.getConstant(X86::COND_NP, dl, MVT::i8);
15557 // Look pass the truncate if the high bits are known zero.
15558 if (isTruncWithZeroHighBitsInput(Cond, DAG))
15559 Cond = Cond.getOperand(0);
15561 // We know the result of AND is compared against zero. Try to match
15563 if (Cond.getOpcode() == ISD::AND && Cond.hasOneUse()) {
15564 SDValue NewSetCC = LowerToBT(Cond, ISD::SETNE, dl, DAG);
15565 if (NewSetCC.getNode()) {
15566 CC = NewSetCC.getOperand(0);
15567 Cond = NewSetCC.getOperand(1);
15574 X86::CondCode X86Cond = Inverted ? X86::COND_E : X86::COND_NE;
15575 CC = DAG.getConstant(X86Cond, dl, MVT::i8);
15576 Cond = EmitTest(Cond, X86Cond, dl, DAG);
15578 Cond = ConvertCmpIfNecessary(Cond, DAG);
15579 return DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
15580 Chain, Dest, CC, Cond);
15583 // Lower dynamic stack allocation to _alloca call for Cygwin/Mingw targets.
15584 // Calls to _alloca are needed to probe the stack when allocating more than 4k
15585 // bytes in one go. Touching the stack at 4K increments is necessary to ensure
15586 // that the guard pages used by the OS virtual memory manager are allocated in
15587 // correct sequence.
15589 X86TargetLowering::LowerDYNAMIC_STACKALLOC(SDValue Op,
15590 SelectionDAG &DAG) const {
15591 MachineFunction &MF = DAG.getMachineFunction();
15592 bool SplitStack = MF.shouldSplitStack();
15593 bool Lower = (Subtarget->isOSWindows() && !Subtarget->isTargetMachO()) ||
15598 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
15599 SDNode* Node = Op.getNode();
15601 unsigned SPReg = TLI.getStackPointerRegisterToSaveRestore();
15602 assert(SPReg && "Target cannot require DYNAMIC_STACKALLOC expansion and"
15603 " not tell us which reg is the stack pointer!");
15604 EVT VT = Node->getValueType(0);
15605 SDValue Tmp1 = SDValue(Node, 0);
15606 SDValue Tmp2 = SDValue(Node, 1);
15607 SDValue Tmp3 = Node->getOperand(2);
15608 SDValue Chain = Tmp1.getOperand(0);
15610 // Chain the dynamic stack allocation so that it doesn't modify the stack
15611 // pointer when other instructions are using the stack.
15612 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(0, dl, true),
15615 SDValue Size = Tmp2.getOperand(1);
15616 SDValue SP = DAG.getCopyFromReg(Chain, dl, SPReg, VT);
15617 Chain = SP.getValue(1);
15618 unsigned Align = cast<ConstantSDNode>(Tmp3)->getZExtValue();
15619 const TargetFrameLowering &TFI = *Subtarget->getFrameLowering();
15620 unsigned StackAlign = TFI.getStackAlignment();
15621 Tmp1 = DAG.getNode(ISD::SUB, dl, VT, SP, Size); // Value
15622 if (Align > StackAlign)
15623 Tmp1 = DAG.getNode(ISD::AND, dl, VT, Tmp1,
15624 DAG.getConstant(-(uint64_t)Align, dl, VT));
15625 Chain = DAG.getCopyToReg(Chain, dl, SPReg, Tmp1); // Output chain
15627 Tmp2 = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(0, dl, true),
15628 DAG.getIntPtrConstant(0, dl, true), SDValue(),
15631 SDValue Ops[2] = { Tmp1, Tmp2 };
15632 return DAG.getMergeValues(Ops, dl);
15636 SDValue Chain = Op.getOperand(0);
15637 SDValue Size = Op.getOperand(1);
15638 unsigned Align = cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue();
15639 EVT VT = Op.getNode()->getValueType(0);
15641 bool Is64Bit = Subtarget->is64Bit();
15642 MVT SPTy = getPointerTy(DAG.getDataLayout());
15645 MachineRegisterInfo &MRI = MF.getRegInfo();
15648 // The 64 bit implementation of segmented stacks needs to clobber both r10
15649 // r11. This makes it impossible to use it along with nested parameters.
15650 const Function *F = MF.getFunction();
15652 for (Function::const_arg_iterator I = F->arg_begin(), E = F->arg_end();
15654 if (I->hasNestAttr())
15655 report_fatal_error("Cannot use segmented stacks with functions that "
15656 "have nested arguments.");
15659 const TargetRegisterClass *AddrRegClass = getRegClassFor(SPTy);
15660 unsigned Vreg = MRI.createVirtualRegister(AddrRegClass);
15661 Chain = DAG.getCopyToReg(Chain, dl, Vreg, Size);
15662 SDValue Value = DAG.getNode(X86ISD::SEG_ALLOCA, dl, SPTy, Chain,
15663 DAG.getRegister(Vreg, SPTy));
15664 SDValue Ops1[2] = { Value, Chain };
15665 return DAG.getMergeValues(Ops1, dl);
15668 const unsigned Reg = (Subtarget->isTarget64BitLP64() ? X86::RAX : X86::EAX);
15670 Chain = DAG.getCopyToReg(Chain, dl, Reg, Size, Flag);
15671 Flag = Chain.getValue(1);
15672 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
15674 Chain = DAG.getNode(X86ISD::WIN_ALLOCA, dl, NodeTys, Chain, Flag);
15676 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
15677 unsigned SPReg = RegInfo->getStackRegister();
15678 SDValue SP = DAG.getCopyFromReg(Chain, dl, SPReg, SPTy);
15679 Chain = SP.getValue(1);
15682 SP = DAG.getNode(ISD::AND, dl, VT, SP.getValue(0),
15683 DAG.getConstant(-(uint64_t)Align, dl, VT));
15684 Chain = DAG.getCopyToReg(Chain, dl, SPReg, SP);
15687 SDValue Ops1[2] = { SP, Chain };
15688 return DAG.getMergeValues(Ops1, dl);
15692 SDValue X86TargetLowering::LowerVASTART(SDValue Op, SelectionDAG &DAG) const {
15693 MachineFunction &MF = DAG.getMachineFunction();
15694 auto PtrVT = getPointerTy(MF.getDataLayout());
15695 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
15697 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
15700 if (!Subtarget->is64Bit() ||
15701 Subtarget->isCallingConvWin64(MF.getFunction()->getCallingConv())) {
15702 // vastart just stores the address of the VarArgsFrameIndex slot into the
15703 // memory location argument.
15704 SDValue FR = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(), PtrVT);
15705 return DAG.getStore(Op.getOperand(0), DL, FR, Op.getOperand(1),
15706 MachinePointerInfo(SV), false, false, 0);
15710 // gp_offset (0 - 6 * 8)
15711 // fp_offset (48 - 48 + 8 * 16)
15712 // overflow_arg_area (point to parameters coming in memory).
15714 SmallVector<SDValue, 8> MemOps;
15715 SDValue FIN = Op.getOperand(1);
15717 SDValue Store = DAG.getStore(Op.getOperand(0), DL,
15718 DAG.getConstant(FuncInfo->getVarArgsGPOffset(),
15720 FIN, MachinePointerInfo(SV), false, false, 0);
15721 MemOps.push_back(Store);
15724 FIN = DAG.getNode(ISD::ADD, DL, PtrVT, FIN, DAG.getIntPtrConstant(4, DL));
15725 Store = DAG.getStore(Op.getOperand(0), DL,
15726 DAG.getConstant(FuncInfo->getVarArgsFPOffset(), DL,
15728 FIN, MachinePointerInfo(SV, 4), false, false, 0);
15729 MemOps.push_back(Store);
15731 // Store ptr to overflow_arg_area
15732 FIN = DAG.getNode(ISD::ADD, DL, PtrVT, FIN, DAG.getIntPtrConstant(4, DL));
15733 SDValue OVFIN = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(), PtrVT);
15734 Store = DAG.getStore(Op.getOperand(0), DL, OVFIN, FIN,
15735 MachinePointerInfo(SV, 8),
15737 MemOps.push_back(Store);
15739 // Store ptr to reg_save_area.
15740 FIN = DAG.getNode(ISD::ADD, DL, PtrVT, FIN, DAG.getIntPtrConstant(
15741 Subtarget->isTarget64BitLP64() ? 8 : 4, DL));
15742 SDValue RSFIN = DAG.getFrameIndex(FuncInfo->getRegSaveFrameIndex(), PtrVT);
15743 Store = DAG.getStore(Op.getOperand(0), DL, RSFIN, FIN, MachinePointerInfo(
15744 SV, Subtarget->isTarget64BitLP64() ? 16 : 12), false, false, 0);
15745 MemOps.push_back(Store);
15746 return DAG.getNode(ISD::TokenFactor, DL, MVT::Other, MemOps);
15749 SDValue X86TargetLowering::LowerVAARG(SDValue Op, SelectionDAG &DAG) const {
15750 assert(Subtarget->is64Bit() &&
15751 "LowerVAARG only handles 64-bit va_arg!");
15752 assert(Op.getNode()->getNumOperands() == 4);
15754 MachineFunction &MF = DAG.getMachineFunction();
15755 if (Subtarget->isCallingConvWin64(MF.getFunction()->getCallingConv()))
15756 // The Win64 ABI uses char* instead of a structure.
15757 return DAG.expandVAArg(Op.getNode());
15759 SDValue Chain = Op.getOperand(0);
15760 SDValue SrcPtr = Op.getOperand(1);
15761 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
15762 unsigned Align = Op.getConstantOperandVal(3);
15765 EVT ArgVT = Op.getNode()->getValueType(0);
15766 Type *ArgTy = ArgVT.getTypeForEVT(*DAG.getContext());
15767 uint32_t ArgSize = DAG.getDataLayout().getTypeAllocSize(ArgTy);
15770 // Decide which area this value should be read from.
15771 // TODO: Implement the AMD64 ABI in its entirety. This simple
15772 // selection mechanism works only for the basic types.
15773 if (ArgVT == MVT::f80) {
15774 llvm_unreachable("va_arg for f80 not yet implemented");
15775 } else if (ArgVT.isFloatingPoint() && ArgSize <= 16 /*bytes*/) {
15776 ArgMode = 2; // Argument passed in XMM register. Use fp_offset.
15777 } else if (ArgVT.isInteger() && ArgSize <= 32 /*bytes*/) {
15778 ArgMode = 1; // Argument passed in GPR64 register(s). Use gp_offset.
15780 llvm_unreachable("Unhandled argument type in LowerVAARG");
15783 if (ArgMode == 2) {
15784 // Sanity Check: Make sure using fp_offset makes sense.
15785 assert(!Subtarget->useSoftFloat() &&
15786 !(MF.getFunction()->hasFnAttribute(Attribute::NoImplicitFloat)) &&
15787 Subtarget->hasSSE1());
15790 // Insert VAARG_64 node into the DAG
15791 // VAARG_64 returns two values: Variable Argument Address, Chain
15792 SDValue InstOps[] = {Chain, SrcPtr, DAG.getConstant(ArgSize, dl, MVT::i32),
15793 DAG.getConstant(ArgMode, dl, MVT::i8),
15794 DAG.getConstant(Align, dl, MVT::i32)};
15795 SDVTList VTs = DAG.getVTList(getPointerTy(DAG.getDataLayout()), MVT::Other);
15796 SDValue VAARG = DAG.getMemIntrinsicNode(X86ISD::VAARG_64, dl,
15797 VTs, InstOps, MVT::i64,
15798 MachinePointerInfo(SV),
15800 /*Volatile=*/false,
15802 /*WriteMem=*/true);
15803 Chain = VAARG.getValue(1);
15805 // Load the next argument and return it
15806 return DAG.getLoad(ArgVT, dl,
15809 MachinePointerInfo(),
15810 false, false, false, 0);
15813 static SDValue LowerVACOPY(SDValue Op, const X86Subtarget *Subtarget,
15814 SelectionDAG &DAG) {
15815 // X86-64 va_list is a struct { i32, i32, i8*, i8* }, except on Windows,
15816 // where a va_list is still an i8*.
15817 assert(Subtarget->is64Bit() && "This code only handles 64-bit va_copy!");
15818 if (Subtarget->isCallingConvWin64(
15819 DAG.getMachineFunction().getFunction()->getCallingConv()))
15820 // Probably a Win64 va_copy.
15821 return DAG.expandVACopy(Op.getNode());
15823 SDValue Chain = Op.getOperand(0);
15824 SDValue DstPtr = Op.getOperand(1);
15825 SDValue SrcPtr = Op.getOperand(2);
15826 const Value *DstSV = cast<SrcValueSDNode>(Op.getOperand(3))->getValue();
15827 const Value *SrcSV = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
15830 return DAG.getMemcpy(Chain, DL, DstPtr, SrcPtr,
15831 DAG.getIntPtrConstant(24, DL), 8, /*isVolatile*/false,
15833 MachinePointerInfo(DstSV), MachinePointerInfo(SrcSV));
15836 // getTargetVShiftByConstNode - Handle vector element shifts where the shift
15837 // amount is a constant. Takes immediate version of shift as input.
15838 static SDValue getTargetVShiftByConstNode(unsigned Opc, SDLoc dl, MVT VT,
15839 SDValue SrcOp, uint64_t ShiftAmt,
15840 SelectionDAG &DAG) {
15841 MVT ElementType = VT.getVectorElementType();
15843 // Fold this packed shift into its first operand if ShiftAmt is 0.
15847 // Check for ShiftAmt >= element width
15848 if (ShiftAmt >= ElementType.getSizeInBits()) {
15849 if (Opc == X86ISD::VSRAI)
15850 ShiftAmt = ElementType.getSizeInBits() - 1;
15852 return DAG.getConstant(0, dl, VT);
15855 assert((Opc == X86ISD::VSHLI || Opc == X86ISD::VSRLI || Opc == X86ISD::VSRAI)
15856 && "Unknown target vector shift-by-constant node");
15858 // Fold this packed vector shift into a build vector if SrcOp is a
15859 // vector of Constants or UNDEFs, and SrcOp valuetype is the same as VT.
15860 if (VT == SrcOp.getSimpleValueType() &&
15861 ISD::isBuildVectorOfConstantSDNodes(SrcOp.getNode())) {
15862 SmallVector<SDValue, 8> Elts;
15863 unsigned NumElts = SrcOp->getNumOperands();
15864 ConstantSDNode *ND;
15867 default: llvm_unreachable(nullptr);
15868 case X86ISD::VSHLI:
15869 for (unsigned i=0; i!=NumElts; ++i) {
15870 SDValue CurrentOp = SrcOp->getOperand(i);
15871 if (CurrentOp->getOpcode() == ISD::UNDEF) {
15872 Elts.push_back(CurrentOp);
15875 ND = cast<ConstantSDNode>(CurrentOp);
15876 const APInt &C = ND->getAPIntValue();
15877 Elts.push_back(DAG.getConstant(C.shl(ShiftAmt), dl, ElementType));
15880 case X86ISD::VSRLI:
15881 for (unsigned i=0; i!=NumElts; ++i) {
15882 SDValue CurrentOp = SrcOp->getOperand(i);
15883 if (CurrentOp->getOpcode() == ISD::UNDEF) {
15884 Elts.push_back(CurrentOp);
15887 ND = cast<ConstantSDNode>(CurrentOp);
15888 const APInt &C = ND->getAPIntValue();
15889 Elts.push_back(DAG.getConstant(C.lshr(ShiftAmt), dl, ElementType));
15892 case X86ISD::VSRAI:
15893 for (unsigned i=0; i!=NumElts; ++i) {
15894 SDValue CurrentOp = SrcOp->getOperand(i);
15895 if (CurrentOp->getOpcode() == ISD::UNDEF) {
15896 Elts.push_back(CurrentOp);
15899 ND = cast<ConstantSDNode>(CurrentOp);
15900 const APInt &C = ND->getAPIntValue();
15901 Elts.push_back(DAG.getConstant(C.ashr(ShiftAmt), dl, ElementType));
15906 return DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Elts);
15909 return DAG.getNode(Opc, dl, VT, SrcOp,
15910 DAG.getConstant(ShiftAmt, dl, MVT::i8));
15913 // getTargetVShiftNode - Handle vector element shifts where the shift amount
15914 // may or may not be a constant. Takes immediate version of shift as input.
15915 static SDValue getTargetVShiftNode(unsigned Opc, SDLoc dl, MVT VT,
15916 SDValue SrcOp, SDValue ShAmt,
15917 SelectionDAG &DAG) {
15918 MVT SVT = ShAmt.getSimpleValueType();
15919 assert((SVT == MVT::i32 || SVT == MVT::i64) && "Unexpected value type!");
15921 // Catch shift-by-constant.
15922 if (ConstantSDNode *CShAmt = dyn_cast<ConstantSDNode>(ShAmt))
15923 return getTargetVShiftByConstNode(Opc, dl, VT, SrcOp,
15924 CShAmt->getZExtValue(), DAG);
15926 // Change opcode to non-immediate version
15928 default: llvm_unreachable("Unknown target vector shift node");
15929 case X86ISD::VSHLI: Opc = X86ISD::VSHL; break;
15930 case X86ISD::VSRLI: Opc = X86ISD::VSRL; break;
15931 case X86ISD::VSRAI: Opc = X86ISD::VSRA; break;
15934 const X86Subtarget &Subtarget =
15935 static_cast<const X86Subtarget &>(DAG.getSubtarget());
15936 if (Subtarget.hasSSE41() && ShAmt.getOpcode() == ISD::ZERO_EXTEND &&
15937 ShAmt.getOperand(0).getSimpleValueType() == MVT::i16) {
15938 // Let the shuffle legalizer expand this shift amount node.
15939 SDValue Op0 = ShAmt.getOperand(0);
15940 Op0 = DAG.getNode(ISD::SCALAR_TO_VECTOR, SDLoc(Op0), MVT::v8i16, Op0);
15941 ShAmt = getShuffleVectorZeroOrUndef(Op0, 0, true, &Subtarget, DAG);
15943 // Need to build a vector containing shift amount.
15944 // SSE/AVX packed shifts only use the lower 64-bit of the shift count.
15945 SmallVector<SDValue, 4> ShOps;
15946 ShOps.push_back(ShAmt);
15947 if (SVT == MVT::i32) {
15948 ShOps.push_back(DAG.getConstant(0, dl, SVT));
15949 ShOps.push_back(DAG.getUNDEF(SVT));
15951 ShOps.push_back(DAG.getUNDEF(SVT));
15953 MVT BVT = SVT == MVT::i32 ? MVT::v4i32 : MVT::v2i64;
15954 ShAmt = DAG.getNode(ISD::BUILD_VECTOR, dl, BVT, ShOps);
15957 // The return type has to be a 128-bit type with the same element
15958 // type as the input type.
15959 MVT EltVT = VT.getVectorElementType();
15960 EVT ShVT = MVT::getVectorVT(EltVT, 128/EltVT.getSizeInBits());
15962 ShAmt = DAG.getBitcast(ShVT, ShAmt);
15963 return DAG.getNode(Opc, dl, VT, SrcOp, ShAmt);
15966 /// \brief Return (and \p Op, \p Mask) for compare instructions or
15967 /// (vselect \p Mask, \p Op, \p PreservedSrc) for others along with the
15968 /// necessary casting or extending for \p Mask when lowering masking intrinsics
15969 static SDValue getVectorMaskingNode(SDValue Op, SDValue Mask,
15970 SDValue PreservedSrc,
15971 const X86Subtarget *Subtarget,
15972 SelectionDAG &DAG) {
15973 EVT VT = Op.getValueType();
15974 EVT MaskVT = EVT::getVectorVT(*DAG.getContext(),
15975 MVT::i1, VT.getVectorNumElements());
15976 SDValue VMask = SDValue();
15977 unsigned OpcodeSelect = ISD::VSELECT;
15980 assert(MaskVT.isSimple() && "invalid mask type");
15982 if (isAllOnes(Mask))
15985 if (MaskVT.bitsGT(Mask.getValueType())) {
15986 EVT newMaskVT = EVT::getIntegerVT(*DAG.getContext(),
15987 MaskVT.getSizeInBits());
15988 VMask = DAG.getBitcast(MaskVT,
15989 DAG.getNode(ISD::ANY_EXTEND, dl, newMaskVT, Mask));
15991 EVT BitcastVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
15992 Mask.getValueType().getSizeInBits());
15993 // In case when MaskVT equals v2i1 or v4i1, low 2 or 4 elements
15994 // are extracted by EXTRACT_SUBVECTOR.
15995 VMask = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MaskVT,
15996 DAG.getBitcast(BitcastVT, Mask),
15997 DAG.getIntPtrConstant(0, dl));
16000 switch (Op.getOpcode()) {
16002 case X86ISD::PCMPEQM:
16003 case X86ISD::PCMPGTM:
16005 case X86ISD::CMPMU:
16006 return DAG.getNode(ISD::AND, dl, VT, Op, VMask);
16007 case X86ISD::VFPCLASS:
16008 return DAG.getNode(ISD::OR, dl, VT, Op, VMask);
16009 case X86ISD::VTRUNC:
16010 case X86ISD::VTRUNCS:
16011 case X86ISD::VTRUNCUS:
16012 // We can't use ISD::VSELECT here because it is not always "Legal"
16013 // for the destination type. For example vpmovqb require only AVX512
16014 // and vselect that can operate on byte element type require BWI
16015 OpcodeSelect = X86ISD::SELECT;
16018 if (PreservedSrc.getOpcode() == ISD::UNDEF)
16019 PreservedSrc = getZeroVector(VT, Subtarget, DAG, dl);
16020 return DAG.getNode(OpcodeSelect, dl, VT, VMask, Op, PreservedSrc);
16023 /// \brief Creates an SDNode for a predicated scalar operation.
16024 /// \returns (X86vselect \p Mask, \p Op, \p PreservedSrc).
16025 /// The mask is coming as MVT::i8 and it should be truncated
16026 /// to MVT::i1 while lowering masking intrinsics.
16027 /// The main difference between ScalarMaskingNode and VectorMaskingNode is using
16028 /// "X86select" instead of "vselect". We just can't create the "vselect" node
16029 /// for a scalar instruction.
16030 static SDValue getScalarMaskingNode(SDValue Op, SDValue Mask,
16031 SDValue PreservedSrc,
16032 const X86Subtarget *Subtarget,
16033 SelectionDAG &DAG) {
16034 if (isAllOnes(Mask))
16037 EVT VT = Op.getValueType();
16039 // The mask should be of type MVT::i1
16040 SDValue IMask = DAG.getNode(ISD::TRUNCATE, dl, MVT::i1, Mask);
16042 if (Op.getOpcode() == X86ISD::FSETCC)
16043 return DAG.getNode(ISD::AND, dl, VT, Op, IMask);
16044 if (Op.getOpcode() == X86ISD::VFPCLASS)
16045 return DAG.getNode(ISD::OR, dl, VT, Op, IMask);
16047 if (PreservedSrc.getOpcode() == ISD::UNDEF)
16048 PreservedSrc = getZeroVector(VT, Subtarget, DAG, dl);
16049 return DAG.getNode(X86ISD::SELECT, dl, VT, IMask, Op, PreservedSrc);
16052 static int getSEHRegistrationNodeSize(const Function *Fn) {
16053 if (!Fn->hasPersonalityFn())
16054 report_fatal_error(
16055 "querying registration node size for function without personality");
16056 // The RegNodeSize is 6 32-bit words for SEH and 4 for C++ EH. See
16057 // WinEHStatePass for the full struct definition.
16058 switch (classifyEHPersonality(Fn->getPersonalityFn())) {
16059 case EHPersonality::MSVC_X86SEH: return 24;
16060 case EHPersonality::MSVC_CXX: return 16;
16063 report_fatal_error("can only recover FP for MSVC EH personality functions");
16066 /// When the 32-bit MSVC runtime transfers control to us, either to an outlined
16067 /// function or when returning to a parent frame after catching an exception, we
16068 /// recover the parent frame pointer by doing arithmetic on the incoming EBP.
16069 /// Here's the math:
16070 /// RegNodeBase = EntryEBP - RegNodeSize
16071 /// ParentFP = RegNodeBase - RegNodeFrameOffset
16072 /// Subtracting RegNodeSize takes us to the offset of the registration node, and
16073 /// subtracting the offset (negative on x86) takes us back to the parent FP.
16074 static SDValue recoverFramePointer(SelectionDAG &DAG, const Function *Fn,
16075 SDValue EntryEBP) {
16076 MachineFunction &MF = DAG.getMachineFunction();
16079 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
16080 MVT PtrVT = TLI.getPointerTy(DAG.getDataLayout());
16082 // It's possible that the parent function no longer has a personality function
16083 // if the exceptional code was optimized away, in which case we just return
16084 // the incoming EBP.
16085 if (!Fn->hasPersonalityFn())
16088 int RegNodeSize = getSEHRegistrationNodeSize(Fn);
16090 // Get an MCSymbol that will ultimately resolve to the frame offset of the EH
16092 MCSymbol *OffsetSym =
16093 MF.getMMI().getContext().getOrCreateParentFrameOffsetSymbol(
16094 GlobalValue::getRealLinkageName(Fn->getName()));
16095 SDValue OffsetSymVal = DAG.getMCSymbol(OffsetSym, PtrVT);
16096 SDValue RegNodeFrameOffset =
16097 DAG.getNode(ISD::LOCAL_RECOVER, dl, PtrVT, OffsetSymVal);
16099 // RegNodeBase = EntryEBP - RegNodeSize
16100 // ParentFP = RegNodeBase - RegNodeFrameOffset
16101 SDValue RegNodeBase = DAG.getNode(ISD::SUB, dl, PtrVT, EntryEBP,
16102 DAG.getConstant(RegNodeSize, dl, PtrVT));
16103 return DAG.getNode(ISD::SUB, dl, PtrVT, RegNodeBase, RegNodeFrameOffset);
16106 static SDValue LowerINTRINSIC_WO_CHAIN(SDValue Op, const X86Subtarget *Subtarget,
16107 SelectionDAG &DAG) {
16109 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
16110 EVT VT = Op.getValueType();
16111 const IntrinsicData* IntrData = getIntrinsicWithoutChain(IntNo);
16113 switch(IntrData->Type) {
16114 case INTR_TYPE_1OP:
16115 return DAG.getNode(IntrData->Opc0, dl, Op.getValueType(), Op.getOperand(1));
16116 case INTR_TYPE_2OP:
16117 return DAG.getNode(IntrData->Opc0, dl, Op.getValueType(), Op.getOperand(1),
16119 case INTR_TYPE_2OP_IMM8:
16120 return DAG.getNode(IntrData->Opc0, dl, Op.getValueType(), Op.getOperand(1),
16121 DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op.getOperand(2)));
16122 case INTR_TYPE_3OP:
16123 return DAG.getNode(IntrData->Opc0, dl, Op.getValueType(), Op.getOperand(1),
16124 Op.getOperand(2), Op.getOperand(3));
16125 case INTR_TYPE_4OP:
16126 return DAG.getNode(IntrData->Opc0, dl, Op.getValueType(), Op.getOperand(1),
16127 Op.getOperand(2), Op.getOperand(3), Op.getOperand(4));
16128 case INTR_TYPE_1OP_MASK_RM: {
16129 SDValue Src = Op.getOperand(1);
16130 SDValue PassThru = Op.getOperand(2);
16131 SDValue Mask = Op.getOperand(3);
16132 SDValue RoundingMode;
16133 // We allways add rounding mode to the Node.
16134 // If the rounding mode is not specified, we add the
16135 // "current direction" mode.
16136 if (Op.getNumOperands() == 4)
16138 DAG.getConstant(X86::STATIC_ROUNDING::CUR_DIRECTION, dl, MVT::i32);
16140 RoundingMode = Op.getOperand(4);
16141 unsigned IntrWithRoundingModeOpcode = IntrData->Opc1;
16142 if (IntrWithRoundingModeOpcode != 0)
16143 if (cast<ConstantSDNode>(RoundingMode)->getZExtValue() !=
16144 X86::STATIC_ROUNDING::CUR_DIRECTION)
16145 return getVectorMaskingNode(DAG.getNode(IntrWithRoundingModeOpcode,
16146 dl, Op.getValueType(), Src, RoundingMode),
16147 Mask, PassThru, Subtarget, DAG);
16148 return getVectorMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT, Src,
16150 Mask, PassThru, Subtarget, DAG);
16152 case INTR_TYPE_1OP_MASK: {
16153 SDValue Src = Op.getOperand(1);
16154 SDValue PassThru = Op.getOperand(2);
16155 SDValue Mask = Op.getOperand(3);
16156 // We add rounding mode to the Node when
16157 // - RM Opcode is specified and
16158 // - RM is not "current direction".
16159 unsigned IntrWithRoundingModeOpcode = IntrData->Opc1;
16160 if (IntrWithRoundingModeOpcode != 0) {
16161 SDValue Rnd = Op.getOperand(4);
16162 unsigned Round = cast<ConstantSDNode>(Rnd)->getZExtValue();
16163 if (Round != X86::STATIC_ROUNDING::CUR_DIRECTION) {
16164 return getVectorMaskingNode(DAG.getNode(IntrWithRoundingModeOpcode,
16165 dl, Op.getValueType(),
16167 Mask, PassThru, Subtarget, DAG);
16170 return getVectorMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT, Src),
16171 Mask, PassThru, Subtarget, DAG);
16173 case INTR_TYPE_SCALAR_MASK: {
16174 SDValue Src1 = Op.getOperand(1);
16175 SDValue Src2 = Op.getOperand(2);
16176 SDValue passThru = Op.getOperand(3);
16177 SDValue Mask = Op.getOperand(4);
16178 return getScalarMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT, Src1, Src2),
16179 Mask, passThru, Subtarget, DAG);
16181 case INTR_TYPE_SCALAR_MASK_RM: {
16182 SDValue Src1 = Op.getOperand(1);
16183 SDValue Src2 = Op.getOperand(2);
16184 SDValue Src0 = Op.getOperand(3);
16185 SDValue Mask = Op.getOperand(4);
16186 // There are 2 kinds of intrinsics in this group:
16187 // (1) With suppress-all-exceptions (sae) or rounding mode- 6 operands
16188 // (2) With rounding mode and sae - 7 operands.
16189 if (Op.getNumOperands() == 6) {
16190 SDValue Sae = Op.getOperand(5);
16191 unsigned Opc = IntrData->Opc1 ? IntrData->Opc1 : IntrData->Opc0;
16192 return getScalarMaskingNode(DAG.getNode(Opc, dl, VT, Src1, Src2,
16194 Mask, Src0, Subtarget, DAG);
16196 assert(Op.getNumOperands() == 7 && "Unexpected intrinsic form");
16197 SDValue RoundingMode = Op.getOperand(5);
16198 SDValue Sae = Op.getOperand(6);
16199 return getScalarMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT, Src1, Src2,
16200 RoundingMode, Sae),
16201 Mask, Src0, Subtarget, DAG);
16203 case INTR_TYPE_2OP_MASK:
16204 case INTR_TYPE_2OP_IMM8_MASK: {
16205 SDValue Src1 = Op.getOperand(1);
16206 SDValue Src2 = Op.getOperand(2);
16207 SDValue PassThru = Op.getOperand(3);
16208 SDValue Mask = Op.getOperand(4);
16210 if (IntrData->Type == INTR_TYPE_2OP_IMM8_MASK)
16211 Src2 = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Src2);
16213 // We specify 2 possible opcodes for intrinsics with rounding modes.
16214 // First, we check if the intrinsic may have non-default rounding mode,
16215 // (IntrData->Opc1 != 0), then we check the rounding mode operand.
16216 unsigned IntrWithRoundingModeOpcode = IntrData->Opc1;
16217 if (IntrWithRoundingModeOpcode != 0) {
16218 SDValue Rnd = Op.getOperand(5);
16219 unsigned Round = cast<ConstantSDNode>(Rnd)->getZExtValue();
16220 if (Round != X86::STATIC_ROUNDING::CUR_DIRECTION) {
16221 return getVectorMaskingNode(DAG.getNode(IntrWithRoundingModeOpcode,
16222 dl, Op.getValueType(),
16224 Mask, PassThru, Subtarget, DAG);
16227 // TODO: Intrinsics should have fast-math-flags to propagate.
16228 return getVectorMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT,Src1,Src2),
16229 Mask, PassThru, Subtarget, DAG);
16231 case INTR_TYPE_2OP_MASK_RM: {
16232 SDValue Src1 = Op.getOperand(1);
16233 SDValue Src2 = Op.getOperand(2);
16234 SDValue PassThru = Op.getOperand(3);
16235 SDValue Mask = Op.getOperand(4);
16236 // We specify 2 possible modes for intrinsics, with/without rounding
16238 // First, we check if the intrinsic have rounding mode (6 operands),
16239 // if not, we set rounding mode to "current".
16241 if (Op.getNumOperands() == 6)
16242 Rnd = Op.getOperand(5);
16244 Rnd = DAG.getConstant(X86::STATIC_ROUNDING::CUR_DIRECTION, dl, MVT::i32);
16245 return getVectorMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT,
16247 Mask, PassThru, Subtarget, DAG);
16249 case INTR_TYPE_3OP_SCALAR_MASK_RM: {
16250 SDValue Src1 = Op.getOperand(1);
16251 SDValue Src2 = Op.getOperand(2);
16252 SDValue Src3 = Op.getOperand(3);
16253 SDValue PassThru = Op.getOperand(4);
16254 SDValue Mask = Op.getOperand(5);
16255 SDValue Sae = Op.getOperand(6);
16257 return getScalarMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT, Src1,
16259 Mask, PassThru, Subtarget, DAG);
16261 case INTR_TYPE_3OP_MASK_RM: {
16262 SDValue Src1 = Op.getOperand(1);
16263 SDValue Src2 = Op.getOperand(2);
16264 SDValue Imm = Op.getOperand(3);
16265 SDValue PassThru = Op.getOperand(4);
16266 SDValue Mask = Op.getOperand(5);
16267 // We specify 2 possible modes for intrinsics, with/without rounding
16269 // First, we check if the intrinsic have rounding mode (7 operands),
16270 // if not, we set rounding mode to "current".
16272 if (Op.getNumOperands() == 7)
16273 Rnd = Op.getOperand(6);
16275 Rnd = DAG.getConstant(X86::STATIC_ROUNDING::CUR_DIRECTION, dl, MVT::i32);
16276 return getVectorMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT,
16277 Src1, Src2, Imm, Rnd),
16278 Mask, PassThru, Subtarget, DAG);
16280 case INTR_TYPE_3OP_IMM8_MASK:
16281 case INTR_TYPE_3OP_MASK:
16282 case INSERT_SUBVEC: {
16283 SDValue Src1 = Op.getOperand(1);
16284 SDValue Src2 = Op.getOperand(2);
16285 SDValue Src3 = Op.getOperand(3);
16286 SDValue PassThru = Op.getOperand(4);
16287 SDValue Mask = Op.getOperand(5);
16289 if (IntrData->Type == INTR_TYPE_3OP_IMM8_MASK)
16290 Src3 = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Src3);
16291 else if (IntrData->Type == INSERT_SUBVEC) {
16292 // imm should be adapted to ISD::INSERT_SUBVECTOR behavior
16293 assert(isa<ConstantSDNode>(Src3) && "Expected a ConstantSDNode here!");
16294 unsigned Imm = cast<ConstantSDNode>(Src3)->getZExtValue();
16295 Imm *= Src2.getValueType().getVectorNumElements();
16296 Src3 = DAG.getTargetConstant(Imm, dl, MVT::i32);
16299 // We specify 2 possible opcodes for intrinsics with rounding modes.
16300 // First, we check if the intrinsic may have non-default rounding mode,
16301 // (IntrData->Opc1 != 0), then we check the rounding mode operand.
16302 unsigned IntrWithRoundingModeOpcode = IntrData->Opc1;
16303 if (IntrWithRoundingModeOpcode != 0) {
16304 SDValue Rnd = Op.getOperand(6);
16305 unsigned Round = cast<ConstantSDNode>(Rnd)->getZExtValue();
16306 if (Round != X86::STATIC_ROUNDING::CUR_DIRECTION) {
16307 return getVectorMaskingNode(DAG.getNode(IntrWithRoundingModeOpcode,
16308 dl, Op.getValueType(),
16309 Src1, Src2, Src3, Rnd),
16310 Mask, PassThru, Subtarget, DAG);
16313 return getVectorMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT,
16315 Mask, PassThru, Subtarget, DAG);
16317 case VPERM_3OP_MASKZ:
16318 case VPERM_3OP_MASK:
16321 case FMA_OP_MASK: {
16322 SDValue Src1 = Op.getOperand(1);
16323 SDValue Src2 = Op.getOperand(2);
16324 SDValue Src3 = Op.getOperand(3);
16325 SDValue Mask = Op.getOperand(4);
16326 EVT VT = Op.getValueType();
16327 SDValue PassThru = SDValue();
16329 // set PassThru element
16330 if (IntrData->Type == VPERM_3OP_MASKZ || IntrData->Type == FMA_OP_MASKZ)
16331 PassThru = getZeroVector(VT, Subtarget, DAG, dl);
16332 else if (IntrData->Type == FMA_OP_MASK3)
16337 // We specify 2 possible opcodes for intrinsics with rounding modes.
16338 // First, we check if the intrinsic may have non-default rounding mode,
16339 // (IntrData->Opc1 != 0), then we check the rounding mode operand.
16340 unsigned IntrWithRoundingModeOpcode = IntrData->Opc1;
16341 if (IntrWithRoundingModeOpcode != 0) {
16342 SDValue Rnd = Op.getOperand(5);
16343 if (cast<ConstantSDNode>(Rnd)->getZExtValue() !=
16344 X86::STATIC_ROUNDING::CUR_DIRECTION)
16345 return getVectorMaskingNode(DAG.getNode(IntrWithRoundingModeOpcode,
16346 dl, Op.getValueType(),
16347 Src1, Src2, Src3, Rnd),
16348 Mask, PassThru, Subtarget, DAG);
16350 return getVectorMaskingNode(DAG.getNode(IntrData->Opc0,
16351 dl, Op.getValueType(),
16353 Mask, PassThru, Subtarget, DAG);
16355 case TERLOG_OP_MASK:
16356 case TERLOG_OP_MASKZ: {
16357 SDValue Src1 = Op.getOperand(1);
16358 SDValue Src2 = Op.getOperand(2);
16359 SDValue Src3 = Op.getOperand(3);
16360 SDValue Src4 = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op.getOperand(4));
16361 SDValue Mask = Op.getOperand(5);
16362 EVT VT = Op.getValueType();
16363 SDValue PassThru = Src1;
16364 // Set PassThru element.
16365 if (IntrData->Type == TERLOG_OP_MASKZ)
16366 PassThru = getZeroVector(VT, Subtarget, DAG, dl);
16368 return getVectorMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT,
16369 Src1, Src2, Src3, Src4),
16370 Mask, PassThru, Subtarget, DAG);
16373 // FPclass intrinsics with mask
16374 SDValue Src1 = Op.getOperand(1);
16375 EVT VT = Src1.getValueType();
16376 EVT MaskVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
16377 VT.getVectorNumElements());
16378 SDValue Imm = Op.getOperand(2);
16379 SDValue Mask = Op.getOperand(3);
16380 EVT BitcastVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
16381 Mask.getValueType().getSizeInBits());
16382 SDValue FPclass = DAG.getNode(IntrData->Opc0, dl, MaskVT, Src1, Imm);
16383 SDValue FPclassMask = getVectorMaskingNode(FPclass, Mask,
16384 DAG.getTargetConstant(0, dl, MaskVT),
16386 SDValue Res = DAG.getNode(ISD::INSERT_SUBVECTOR, dl, BitcastVT,
16387 DAG.getUNDEF(BitcastVT), FPclassMask,
16388 DAG.getIntPtrConstant(0, dl));
16389 return DAG.getBitcast(Op.getValueType(), Res);
16392 SDValue Src1 = Op.getOperand(1);
16393 SDValue Imm = Op.getOperand(2);
16394 SDValue Mask = Op.getOperand(3);
16395 SDValue FPclass = DAG.getNode(IntrData->Opc0, dl, MVT::i1, Src1, Imm);
16396 SDValue FPclassMask = getScalarMaskingNode(FPclass, Mask,
16397 DAG.getTargetConstant(0, dl, MVT::i1), Subtarget, DAG);
16398 return DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::i8, FPclassMask);
16401 case CMP_MASK_CC: {
16402 // Comparison intrinsics with masks.
16403 // Example of transformation:
16404 // (i8 (int_x86_avx512_mask_pcmpeq_q_128
16405 // (v2i64 %a), (v2i64 %b), (i8 %mask))) ->
16407 // (v8i1 (insert_subvector undef,
16408 // (v2i1 (and (PCMPEQM %a, %b),
16409 // (extract_subvector
16410 // (v8i1 (bitcast %mask)), 0))), 0))))
16411 EVT VT = Op.getOperand(1).getValueType();
16412 EVT MaskVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
16413 VT.getVectorNumElements());
16414 SDValue Mask = Op.getOperand((IntrData->Type == CMP_MASK_CC) ? 4 : 3);
16415 EVT BitcastVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
16416 Mask.getValueType().getSizeInBits());
16418 if (IntrData->Type == CMP_MASK_CC) {
16419 SDValue CC = Op.getOperand(3);
16420 CC = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, CC);
16421 // We specify 2 possible opcodes for intrinsics with rounding modes.
16422 // First, we check if the intrinsic may have non-default rounding mode,
16423 // (IntrData->Opc1 != 0), then we check the rounding mode operand.
16424 if (IntrData->Opc1 != 0) {
16425 SDValue Rnd = Op.getOperand(5);
16426 if (cast<ConstantSDNode>(Rnd)->getZExtValue() !=
16427 X86::STATIC_ROUNDING::CUR_DIRECTION)
16428 Cmp = DAG.getNode(IntrData->Opc1, dl, MaskVT, Op.getOperand(1),
16429 Op.getOperand(2), CC, Rnd);
16431 //default rounding mode
16433 Cmp = DAG.getNode(IntrData->Opc0, dl, MaskVT, Op.getOperand(1),
16434 Op.getOperand(2), CC);
16437 assert(IntrData->Type == CMP_MASK && "Unexpected intrinsic type!");
16438 Cmp = DAG.getNode(IntrData->Opc0, dl, MaskVT, Op.getOperand(1),
16441 SDValue CmpMask = getVectorMaskingNode(Cmp, Mask,
16442 DAG.getTargetConstant(0, dl,
16445 SDValue Res = DAG.getNode(ISD::INSERT_SUBVECTOR, dl, BitcastVT,
16446 DAG.getUNDEF(BitcastVT), CmpMask,
16447 DAG.getIntPtrConstant(0, dl));
16448 return DAG.getBitcast(Op.getValueType(), Res);
16450 case CMP_MASK_SCALAR_CC: {
16451 SDValue Src1 = Op.getOperand(1);
16452 SDValue Src2 = Op.getOperand(2);
16453 SDValue CC = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op.getOperand(3));
16454 SDValue Mask = Op.getOperand(4);
16457 if (IntrData->Opc1 != 0) {
16458 SDValue Rnd = Op.getOperand(5);
16459 if (cast<ConstantSDNode>(Rnd)->getZExtValue() !=
16460 X86::STATIC_ROUNDING::CUR_DIRECTION)
16461 Cmp = DAG.getNode(IntrData->Opc1, dl, MVT::i1, Src1, Src2, CC, Rnd);
16463 //default rounding mode
16465 Cmp = DAG.getNode(IntrData->Opc0, dl, MVT::i1, Src1, Src2, CC);
16467 SDValue CmpMask = getScalarMaskingNode(Cmp, Mask,
16468 DAG.getTargetConstant(0, dl,
16472 return DAG.getNode(ISD::SIGN_EXTEND_INREG, dl, MVT::i8,
16473 DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i8, CmpMask),
16474 DAG.getValueType(MVT::i1));
16476 case COMI: { // Comparison intrinsics
16477 ISD::CondCode CC = (ISD::CondCode)IntrData->Opc1;
16478 SDValue LHS = Op.getOperand(1);
16479 SDValue RHS = Op.getOperand(2);
16480 unsigned X86CC = TranslateX86CC(CC, dl, true, LHS, RHS, DAG);
16481 assert(X86CC != X86::COND_INVALID && "Unexpected illegal condition!");
16482 SDValue Cond = DAG.getNode(IntrData->Opc0, dl, MVT::i32, LHS, RHS);
16483 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
16484 DAG.getConstant(X86CC, dl, MVT::i8), Cond);
16485 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
16488 return getTargetVShiftNode(IntrData->Opc0, dl, Op.getSimpleValueType(),
16489 Op.getOperand(1), Op.getOperand(2), DAG);
16491 return getVectorMaskingNode(getTargetVShiftNode(IntrData->Opc0, dl,
16492 Op.getSimpleValueType(),
16494 Op.getOperand(2), DAG),
16495 Op.getOperand(4), Op.getOperand(3), Subtarget,
16497 case COMPRESS_EXPAND_IN_REG: {
16498 SDValue Mask = Op.getOperand(3);
16499 SDValue DataToCompress = Op.getOperand(1);
16500 SDValue PassThru = Op.getOperand(2);
16501 if (isAllOnes(Mask)) // return data as is
16502 return Op.getOperand(1);
16504 return getVectorMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT,
16506 Mask, PassThru, Subtarget, DAG);
16509 SDValue Mask = Op.getOperand(3);
16510 EVT VT = Op.getValueType();
16511 EVT MaskVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
16512 VT.getVectorNumElements());
16513 EVT BitcastVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
16514 Mask.getValueType().getSizeInBits());
16516 SDValue VMask = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MaskVT,
16517 DAG.getBitcast(BitcastVT, Mask),
16518 DAG.getIntPtrConstant(0, dl));
16519 return DAG.getNode(IntrData->Opc0, dl, VT, VMask, Op.getOperand(1),
16528 default: return SDValue(); // Don't custom lower most intrinsics.
16530 case Intrinsic::x86_avx2_permd:
16531 case Intrinsic::x86_avx2_permps:
16532 // Operands intentionally swapped. Mask is last operand to intrinsic,
16533 // but second operand for node/instruction.
16534 return DAG.getNode(X86ISD::VPERMV, dl, Op.getValueType(),
16535 Op.getOperand(2), Op.getOperand(1));
16537 // ptest and testp intrinsics. The intrinsic these come from are designed to
16538 // return an integer value, not just an instruction so lower it to the ptest
16539 // or testp pattern and a setcc for the result.
16540 case Intrinsic::x86_sse41_ptestz:
16541 case Intrinsic::x86_sse41_ptestc:
16542 case Intrinsic::x86_sse41_ptestnzc:
16543 case Intrinsic::x86_avx_ptestz_256:
16544 case Intrinsic::x86_avx_ptestc_256:
16545 case Intrinsic::x86_avx_ptestnzc_256:
16546 case Intrinsic::x86_avx_vtestz_ps:
16547 case Intrinsic::x86_avx_vtestc_ps:
16548 case Intrinsic::x86_avx_vtestnzc_ps:
16549 case Intrinsic::x86_avx_vtestz_pd:
16550 case Intrinsic::x86_avx_vtestc_pd:
16551 case Intrinsic::x86_avx_vtestnzc_pd:
16552 case Intrinsic::x86_avx_vtestz_ps_256:
16553 case Intrinsic::x86_avx_vtestc_ps_256:
16554 case Intrinsic::x86_avx_vtestnzc_ps_256:
16555 case Intrinsic::x86_avx_vtestz_pd_256:
16556 case Intrinsic::x86_avx_vtestc_pd_256:
16557 case Intrinsic::x86_avx_vtestnzc_pd_256: {
16558 bool IsTestPacked = false;
16561 default: llvm_unreachable("Bad fallthrough in Intrinsic lowering.");
16562 case Intrinsic::x86_avx_vtestz_ps:
16563 case Intrinsic::x86_avx_vtestz_pd:
16564 case Intrinsic::x86_avx_vtestz_ps_256:
16565 case Intrinsic::x86_avx_vtestz_pd_256:
16566 IsTestPacked = true; // Fallthrough
16567 case Intrinsic::x86_sse41_ptestz:
16568 case Intrinsic::x86_avx_ptestz_256:
16570 X86CC = X86::COND_E;
16572 case Intrinsic::x86_avx_vtestc_ps:
16573 case Intrinsic::x86_avx_vtestc_pd:
16574 case Intrinsic::x86_avx_vtestc_ps_256:
16575 case Intrinsic::x86_avx_vtestc_pd_256:
16576 IsTestPacked = true; // Fallthrough
16577 case Intrinsic::x86_sse41_ptestc:
16578 case Intrinsic::x86_avx_ptestc_256:
16580 X86CC = X86::COND_B;
16582 case Intrinsic::x86_avx_vtestnzc_ps:
16583 case Intrinsic::x86_avx_vtestnzc_pd:
16584 case Intrinsic::x86_avx_vtestnzc_ps_256:
16585 case Intrinsic::x86_avx_vtestnzc_pd_256:
16586 IsTestPacked = true; // Fallthrough
16587 case Intrinsic::x86_sse41_ptestnzc:
16588 case Intrinsic::x86_avx_ptestnzc_256:
16590 X86CC = X86::COND_A;
16594 SDValue LHS = Op.getOperand(1);
16595 SDValue RHS = Op.getOperand(2);
16596 unsigned TestOpc = IsTestPacked ? X86ISD::TESTP : X86ISD::PTEST;
16597 SDValue Test = DAG.getNode(TestOpc, dl, MVT::i32, LHS, RHS);
16598 SDValue CC = DAG.getConstant(X86CC, dl, MVT::i8);
16599 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8, CC, Test);
16600 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
16602 case Intrinsic::x86_avx512_kortestz_w:
16603 case Intrinsic::x86_avx512_kortestc_w: {
16604 unsigned X86CC = (IntNo == Intrinsic::x86_avx512_kortestz_w)? X86::COND_E: X86::COND_B;
16605 SDValue LHS = DAG.getBitcast(MVT::v16i1, Op.getOperand(1));
16606 SDValue RHS = DAG.getBitcast(MVT::v16i1, Op.getOperand(2));
16607 SDValue CC = DAG.getConstant(X86CC, dl, MVT::i8);
16608 SDValue Test = DAG.getNode(X86ISD::KORTEST, dl, MVT::i32, LHS, RHS);
16609 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i1, CC, Test);
16610 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
16613 case Intrinsic::x86_sse42_pcmpistria128:
16614 case Intrinsic::x86_sse42_pcmpestria128:
16615 case Intrinsic::x86_sse42_pcmpistric128:
16616 case Intrinsic::x86_sse42_pcmpestric128:
16617 case Intrinsic::x86_sse42_pcmpistrio128:
16618 case Intrinsic::x86_sse42_pcmpestrio128:
16619 case Intrinsic::x86_sse42_pcmpistris128:
16620 case Intrinsic::x86_sse42_pcmpestris128:
16621 case Intrinsic::x86_sse42_pcmpistriz128:
16622 case Intrinsic::x86_sse42_pcmpestriz128: {
16626 default: llvm_unreachable("Impossible intrinsic"); // Can't reach here.
16627 case Intrinsic::x86_sse42_pcmpistria128:
16628 Opcode = X86ISD::PCMPISTRI;
16629 X86CC = X86::COND_A;
16631 case Intrinsic::x86_sse42_pcmpestria128:
16632 Opcode = X86ISD::PCMPESTRI;
16633 X86CC = X86::COND_A;
16635 case Intrinsic::x86_sse42_pcmpistric128:
16636 Opcode = X86ISD::PCMPISTRI;
16637 X86CC = X86::COND_B;
16639 case Intrinsic::x86_sse42_pcmpestric128:
16640 Opcode = X86ISD::PCMPESTRI;
16641 X86CC = X86::COND_B;
16643 case Intrinsic::x86_sse42_pcmpistrio128:
16644 Opcode = X86ISD::PCMPISTRI;
16645 X86CC = X86::COND_O;
16647 case Intrinsic::x86_sse42_pcmpestrio128:
16648 Opcode = X86ISD::PCMPESTRI;
16649 X86CC = X86::COND_O;
16651 case Intrinsic::x86_sse42_pcmpistris128:
16652 Opcode = X86ISD::PCMPISTRI;
16653 X86CC = X86::COND_S;
16655 case Intrinsic::x86_sse42_pcmpestris128:
16656 Opcode = X86ISD::PCMPESTRI;
16657 X86CC = X86::COND_S;
16659 case Intrinsic::x86_sse42_pcmpistriz128:
16660 Opcode = X86ISD::PCMPISTRI;
16661 X86CC = X86::COND_E;
16663 case Intrinsic::x86_sse42_pcmpestriz128:
16664 Opcode = X86ISD::PCMPESTRI;
16665 X86CC = X86::COND_E;
16668 SmallVector<SDValue, 5> NewOps(Op->op_begin()+1, Op->op_end());
16669 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::i32);
16670 SDValue PCMP = DAG.getNode(Opcode, dl, VTs, NewOps);
16671 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
16672 DAG.getConstant(X86CC, dl, MVT::i8),
16673 SDValue(PCMP.getNode(), 1));
16674 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
16677 case Intrinsic::x86_sse42_pcmpistri128:
16678 case Intrinsic::x86_sse42_pcmpestri128: {
16680 if (IntNo == Intrinsic::x86_sse42_pcmpistri128)
16681 Opcode = X86ISD::PCMPISTRI;
16683 Opcode = X86ISD::PCMPESTRI;
16685 SmallVector<SDValue, 5> NewOps(Op->op_begin()+1, Op->op_end());
16686 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::i32);
16687 return DAG.getNode(Opcode, dl, VTs, NewOps);
16690 case Intrinsic::x86_seh_lsda: {
16691 // Compute the symbol for the LSDA. We know it'll get emitted later.
16692 MachineFunction &MF = DAG.getMachineFunction();
16693 SDValue Op1 = Op.getOperand(1);
16694 auto *Fn = cast<Function>(cast<GlobalAddressSDNode>(Op1)->getGlobal());
16695 MCSymbol *LSDASym = MF.getMMI().getContext().getOrCreateLSDASymbol(
16696 GlobalValue::getRealLinkageName(Fn->getName()));
16698 // Generate a simple absolute symbol reference. This intrinsic is only
16699 // supported on 32-bit Windows, which isn't PIC.
16700 SDValue Result = DAG.getMCSymbol(LSDASym, VT);
16701 return DAG.getNode(X86ISD::Wrapper, dl, VT, Result);
16704 case Intrinsic::x86_seh_recoverfp: {
16705 SDValue FnOp = Op.getOperand(1);
16706 SDValue IncomingFPOp = Op.getOperand(2);
16707 GlobalAddressSDNode *GSD = dyn_cast<GlobalAddressSDNode>(FnOp);
16708 auto *Fn = dyn_cast_or_null<Function>(GSD ? GSD->getGlobal() : nullptr);
16710 report_fatal_error(
16711 "llvm.x86.seh.recoverfp must take a function as the first argument");
16712 return recoverFramePointer(DAG, Fn, IncomingFPOp);
16715 case Intrinsic::localaddress: {
16716 // Returns one of the stack, base, or frame pointer registers, depending on
16717 // which is used to reference local variables.
16718 MachineFunction &MF = DAG.getMachineFunction();
16719 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
16721 if (RegInfo->hasBasePointer(MF))
16722 Reg = RegInfo->getBaseRegister();
16723 else // This function handles the SP or FP case.
16724 Reg = RegInfo->getPtrSizedFrameRegister(MF);
16725 return DAG.getCopyFromReg(DAG.getEntryNode(), dl, Reg, VT);
16730 static SDValue getGatherNode(unsigned Opc, SDValue Op, SelectionDAG &DAG,
16731 SDValue Src, SDValue Mask, SDValue Base,
16732 SDValue Index, SDValue ScaleOp, SDValue Chain,
16733 const X86Subtarget * Subtarget) {
16735 ConstantSDNode *C = dyn_cast<ConstantSDNode>(ScaleOp);
16737 llvm_unreachable("Invalid scale type");
16738 unsigned ScaleVal = C->getZExtValue();
16739 if (ScaleVal > 2 && ScaleVal != 4 && ScaleVal != 8)
16740 llvm_unreachable("Valid scale values are 1, 2, 4, 8");
16742 SDValue Scale = DAG.getTargetConstant(C->getZExtValue(), dl, MVT::i8);
16743 EVT MaskVT = MVT::getVectorVT(MVT::i1,
16744 Index.getSimpleValueType().getVectorNumElements());
16746 ConstantSDNode *MaskC = dyn_cast<ConstantSDNode>(Mask);
16748 MaskInReg = DAG.getTargetConstant(MaskC->getSExtValue(), dl, MaskVT);
16750 EVT BitcastVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
16751 Mask.getValueType().getSizeInBits());
16753 // In case when MaskVT equals v2i1 or v4i1, low 2 or 4 elements
16754 // are extracted by EXTRACT_SUBVECTOR.
16755 MaskInReg = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MaskVT,
16756 DAG.getBitcast(BitcastVT, Mask),
16757 DAG.getIntPtrConstant(0, dl));
16759 SDVTList VTs = DAG.getVTList(Op.getValueType(), MaskVT, MVT::Other);
16760 SDValue Disp = DAG.getTargetConstant(0, dl, MVT::i32);
16761 SDValue Segment = DAG.getRegister(0, MVT::i32);
16762 if (Src.getOpcode() == ISD::UNDEF)
16763 Src = getZeroVector(Op.getValueType(), Subtarget, DAG, dl);
16764 SDValue Ops[] = {Src, MaskInReg, Base, Scale, Index, Disp, Segment, Chain};
16765 SDNode *Res = DAG.getMachineNode(Opc, dl, VTs, Ops);
16766 SDValue RetOps[] = { SDValue(Res, 0), SDValue(Res, 2) };
16767 return DAG.getMergeValues(RetOps, dl);
16770 static SDValue getScatterNode(unsigned Opc, SDValue Op, SelectionDAG &DAG,
16771 SDValue Src, SDValue Mask, SDValue Base,
16772 SDValue Index, SDValue ScaleOp, SDValue Chain) {
16774 ConstantSDNode *C = dyn_cast<ConstantSDNode>(ScaleOp);
16776 llvm_unreachable("Invalid scale type");
16777 unsigned ScaleVal = C->getZExtValue();
16778 if (ScaleVal > 2 && ScaleVal != 4 && ScaleVal != 8)
16779 llvm_unreachable("Valid scale values are 1, 2, 4, 8");
16781 SDValue Scale = DAG.getTargetConstant(C->getZExtValue(), dl, MVT::i8);
16782 SDValue Disp = DAG.getTargetConstant(0, dl, MVT::i32);
16783 SDValue Segment = DAG.getRegister(0, MVT::i32);
16784 EVT MaskVT = MVT::getVectorVT(MVT::i1,
16785 Index.getSimpleValueType().getVectorNumElements());
16787 ConstantSDNode *MaskC = dyn_cast<ConstantSDNode>(Mask);
16789 MaskInReg = DAG.getTargetConstant(MaskC->getSExtValue(), dl, MaskVT);
16791 EVT BitcastVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
16792 Mask.getValueType().getSizeInBits());
16794 // In case when MaskVT equals v2i1 or v4i1, low 2 or 4 elements
16795 // are extracted by EXTRACT_SUBVECTOR.
16796 MaskInReg = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MaskVT,
16797 DAG.getBitcast(BitcastVT, Mask),
16798 DAG.getIntPtrConstant(0, dl));
16800 SDVTList VTs = DAG.getVTList(MaskVT, MVT::Other);
16801 SDValue Ops[] = {Base, Scale, Index, Disp, Segment, MaskInReg, Src, Chain};
16802 SDNode *Res = DAG.getMachineNode(Opc, dl, VTs, Ops);
16803 return SDValue(Res, 1);
16806 static SDValue getPrefetchNode(unsigned Opc, SDValue Op, SelectionDAG &DAG,
16807 SDValue Mask, SDValue Base, SDValue Index,
16808 SDValue ScaleOp, SDValue Chain) {
16810 ConstantSDNode *C = dyn_cast<ConstantSDNode>(ScaleOp);
16811 assert(C && "Invalid scale type");
16812 SDValue Scale = DAG.getTargetConstant(C->getZExtValue(), dl, MVT::i8);
16813 SDValue Disp = DAG.getTargetConstant(0, dl, MVT::i32);
16814 SDValue Segment = DAG.getRegister(0, MVT::i32);
16816 MVT::getVectorVT(MVT::i1, Index.getSimpleValueType().getVectorNumElements());
16818 ConstantSDNode *MaskC = dyn_cast<ConstantSDNode>(Mask);
16820 MaskInReg = DAG.getTargetConstant(MaskC->getSExtValue(), dl, MaskVT);
16822 MaskInReg = DAG.getBitcast(MaskVT, Mask);
16823 //SDVTList VTs = DAG.getVTList(MVT::Other);
16824 SDValue Ops[] = {MaskInReg, Base, Scale, Index, Disp, Segment, Chain};
16825 SDNode *Res = DAG.getMachineNode(Opc, dl, MVT::Other, Ops);
16826 return SDValue(Res, 0);
16829 // getReadPerformanceCounter - Handles the lowering of builtin intrinsics that
16830 // read performance monitor counters (x86_rdpmc).
16831 static void getReadPerformanceCounter(SDNode *N, SDLoc DL,
16832 SelectionDAG &DAG, const X86Subtarget *Subtarget,
16833 SmallVectorImpl<SDValue> &Results) {
16834 assert(N->getNumOperands() == 3 && "Unexpected number of operands!");
16835 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Glue);
16838 // The ECX register is used to select the index of the performance counter
16840 SDValue Chain = DAG.getCopyToReg(N->getOperand(0), DL, X86::ECX,
16842 SDValue rd = DAG.getNode(X86ISD::RDPMC_DAG, DL, Tys, Chain);
16844 // Reads the content of a 64-bit performance counter and returns it in the
16845 // registers EDX:EAX.
16846 if (Subtarget->is64Bit()) {
16847 LO = DAG.getCopyFromReg(rd, DL, X86::RAX, MVT::i64, rd.getValue(1));
16848 HI = DAG.getCopyFromReg(LO.getValue(1), DL, X86::RDX, MVT::i64,
16851 LO = DAG.getCopyFromReg(rd, DL, X86::EAX, MVT::i32, rd.getValue(1));
16852 HI = DAG.getCopyFromReg(LO.getValue(1), DL, X86::EDX, MVT::i32,
16855 Chain = HI.getValue(1);
16857 if (Subtarget->is64Bit()) {
16858 // The EAX register is loaded with the low-order 32 bits. The EDX register
16859 // is loaded with the supported high-order bits of the counter.
16860 SDValue Tmp = DAG.getNode(ISD::SHL, DL, MVT::i64, HI,
16861 DAG.getConstant(32, DL, MVT::i8));
16862 Results.push_back(DAG.getNode(ISD::OR, DL, MVT::i64, LO, Tmp));
16863 Results.push_back(Chain);
16867 // Use a buildpair to merge the two 32-bit values into a 64-bit one.
16868 SDValue Ops[] = { LO, HI };
16869 SDValue Pair = DAG.getNode(ISD::BUILD_PAIR, DL, MVT::i64, Ops);
16870 Results.push_back(Pair);
16871 Results.push_back(Chain);
16874 // getReadTimeStampCounter - Handles the lowering of builtin intrinsics that
16875 // read the time stamp counter (x86_rdtsc and x86_rdtscp). This function is
16876 // also used to custom lower READCYCLECOUNTER nodes.
16877 static void getReadTimeStampCounter(SDNode *N, SDLoc DL, unsigned Opcode,
16878 SelectionDAG &DAG, const X86Subtarget *Subtarget,
16879 SmallVectorImpl<SDValue> &Results) {
16880 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Glue);
16881 SDValue rd = DAG.getNode(Opcode, DL, Tys, N->getOperand(0));
16884 // The processor's time-stamp counter (a 64-bit MSR) is stored into the
16885 // EDX:EAX registers. EDX is loaded with the high-order 32 bits of the MSR
16886 // and the EAX register is loaded with the low-order 32 bits.
16887 if (Subtarget->is64Bit()) {
16888 LO = DAG.getCopyFromReg(rd, DL, X86::RAX, MVT::i64, rd.getValue(1));
16889 HI = DAG.getCopyFromReg(LO.getValue(1), DL, X86::RDX, MVT::i64,
16892 LO = DAG.getCopyFromReg(rd, DL, X86::EAX, MVT::i32, rd.getValue(1));
16893 HI = DAG.getCopyFromReg(LO.getValue(1), DL, X86::EDX, MVT::i32,
16896 SDValue Chain = HI.getValue(1);
16898 if (Opcode == X86ISD::RDTSCP_DAG) {
16899 assert(N->getNumOperands() == 3 && "Unexpected number of operands!");
16901 // Instruction RDTSCP loads the IA32:TSC_AUX_MSR (address C000_0103H) into
16902 // the ECX register. Add 'ecx' explicitly to the chain.
16903 SDValue ecx = DAG.getCopyFromReg(Chain, DL, X86::ECX, MVT::i32,
16905 // Explicitly store the content of ECX at the location passed in input
16906 // to the 'rdtscp' intrinsic.
16907 Chain = DAG.getStore(ecx.getValue(1), DL, ecx, N->getOperand(2),
16908 MachinePointerInfo(), false, false, 0);
16911 if (Subtarget->is64Bit()) {
16912 // The EDX register is loaded with the high-order 32 bits of the MSR, and
16913 // the EAX register is loaded with the low-order 32 bits.
16914 SDValue Tmp = DAG.getNode(ISD::SHL, DL, MVT::i64, HI,
16915 DAG.getConstant(32, DL, MVT::i8));
16916 Results.push_back(DAG.getNode(ISD::OR, DL, MVT::i64, LO, Tmp));
16917 Results.push_back(Chain);
16921 // Use a buildpair to merge the two 32-bit values into a 64-bit one.
16922 SDValue Ops[] = { LO, HI };
16923 SDValue Pair = DAG.getNode(ISD::BUILD_PAIR, DL, MVT::i64, Ops);
16924 Results.push_back(Pair);
16925 Results.push_back(Chain);
16928 static SDValue LowerREADCYCLECOUNTER(SDValue Op, const X86Subtarget *Subtarget,
16929 SelectionDAG &DAG) {
16930 SmallVector<SDValue, 2> Results;
16932 getReadTimeStampCounter(Op.getNode(), DL, X86ISD::RDTSC_DAG, DAG, Subtarget,
16934 return DAG.getMergeValues(Results, DL);
16937 static SDValue LowerSEHRESTOREFRAME(SDValue Op, const X86Subtarget *Subtarget,
16938 SelectionDAG &DAG) {
16939 MachineFunction &MF = DAG.getMachineFunction();
16940 const Function *Fn = MF.getFunction();
16942 SDValue Chain = Op.getOperand(0);
16944 assert(Subtarget->getFrameLowering()->hasFP(MF) &&
16945 "using llvm.x86.seh.restoreframe requires a frame pointer");
16947 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
16948 MVT VT = TLI.getPointerTy(DAG.getDataLayout());
16950 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
16951 unsigned FrameReg =
16952 RegInfo->getPtrSizedFrameRegister(DAG.getMachineFunction());
16953 unsigned SPReg = RegInfo->getStackRegister();
16954 unsigned SlotSize = RegInfo->getSlotSize();
16956 // Get incoming EBP.
16957 SDValue IncomingEBP =
16958 DAG.getCopyFromReg(Chain, dl, FrameReg, VT);
16960 // SP is saved in the first field of every registration node, so load
16961 // [EBP-RegNodeSize] into SP.
16962 int RegNodeSize = getSEHRegistrationNodeSize(Fn);
16963 SDValue SPAddr = DAG.getNode(ISD::ADD, dl, VT, IncomingEBP,
16964 DAG.getConstant(-RegNodeSize, dl, VT));
16966 DAG.getLoad(VT, dl, Chain, SPAddr, MachinePointerInfo(), false, false,
16967 false, VT.getScalarSizeInBits() / 8);
16968 Chain = DAG.getCopyToReg(Chain, dl, SPReg, NewSP);
16970 if (!RegInfo->needsStackRealignment(MF)) {
16971 // Adjust EBP to point back to the original frame position.
16972 SDValue NewFP = recoverFramePointer(DAG, Fn, IncomingEBP);
16973 Chain = DAG.getCopyToReg(Chain, dl, FrameReg, NewFP);
16975 assert(RegInfo->hasBasePointer(MF) &&
16976 "functions with Win32 EH must use frame or base pointer register");
16978 // Reload the base pointer (ESI) with the adjusted incoming EBP.
16979 SDValue NewBP = recoverFramePointer(DAG, Fn, IncomingEBP);
16980 Chain = DAG.getCopyToReg(Chain, dl, RegInfo->getBaseRegister(), NewBP);
16982 // Reload the spilled EBP value, now that the stack and base pointers are
16984 X86MachineFunctionInfo *X86FI = MF.getInfo<X86MachineFunctionInfo>();
16985 X86FI->setHasSEHFramePtrSave(true);
16986 int FI = MF.getFrameInfo()->CreateSpillStackObject(SlotSize, SlotSize);
16987 X86FI->setSEHFramePtrSaveIndex(FI);
16988 SDValue NewFP = DAG.getLoad(VT, dl, Chain, DAG.getFrameIndex(FI, VT),
16989 MachinePointerInfo(), false, false, false,
16990 VT.getScalarSizeInBits() / 8);
16991 Chain = DAG.getCopyToReg(NewFP, dl, FrameReg, NewFP);
16997 /// \brief Lower intrinsics for TRUNCATE_TO_MEM case
16998 /// return truncate Store/MaskedStore Node
16999 static SDValue LowerINTRINSIC_TRUNCATE_TO_MEM(const SDValue & Op,
17003 SDValue Mask = Op.getOperand(4);
17004 SDValue DataToTruncate = Op.getOperand(3);
17005 SDValue Addr = Op.getOperand(2);
17006 SDValue Chain = Op.getOperand(0);
17008 EVT VT = DataToTruncate.getValueType();
17009 EVT SVT = EVT::getVectorVT(*DAG.getContext(),
17010 ElementType, VT.getVectorNumElements());
17012 if (isAllOnes(Mask)) // return just a truncate store
17013 return DAG.getTruncStore(Chain, dl, DataToTruncate, Addr,
17014 MachinePointerInfo(), SVT, false, false,
17015 SVT.getScalarSizeInBits()/8);
17017 EVT MaskVT = EVT::getVectorVT(*DAG.getContext(),
17018 MVT::i1, VT.getVectorNumElements());
17019 EVT BitcastVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
17020 Mask.getValueType().getSizeInBits());
17021 // In case when MaskVT equals v2i1 or v4i1, low 2 or 4 elements
17022 // are extracted by EXTRACT_SUBVECTOR.
17023 SDValue VMask = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MaskVT,
17024 DAG.getBitcast(BitcastVT, Mask),
17025 DAG.getIntPtrConstant(0, dl));
17027 MachineMemOperand *MMO = DAG.getMachineFunction().
17028 getMachineMemOperand(MachinePointerInfo(),
17029 MachineMemOperand::MOStore, SVT.getStoreSize(),
17030 SVT.getScalarSizeInBits()/8);
17032 return DAG.getMaskedStore(Chain, dl, DataToTruncate, Addr,
17033 VMask, SVT, MMO, true);
17036 static SDValue LowerINTRINSIC_W_CHAIN(SDValue Op, const X86Subtarget *Subtarget,
17037 SelectionDAG &DAG) {
17038 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
17040 const IntrinsicData* IntrData = getIntrinsicWithChain(IntNo);
17042 if (IntNo == llvm::Intrinsic::x86_seh_restoreframe)
17043 return LowerSEHRESTOREFRAME(Op, Subtarget, DAG);
17048 switch(IntrData->Type) {
17050 llvm_unreachable("Unknown Intrinsic Type");
17054 // Emit the node with the right value type.
17055 SDVTList VTs = DAG.getVTList(Op->getValueType(0), MVT::Glue, MVT::Other);
17056 SDValue Result = DAG.getNode(IntrData->Opc0, dl, VTs, Op.getOperand(0));
17058 // If the value returned by RDRAND/RDSEED was valid (CF=1), return 1.
17059 // Otherwise return the value from Rand, which is always 0, casted to i32.
17060 SDValue Ops[] = { DAG.getZExtOrTrunc(Result, dl, Op->getValueType(1)),
17061 DAG.getConstant(1, dl, Op->getValueType(1)),
17062 DAG.getConstant(X86::COND_B, dl, MVT::i32),
17063 SDValue(Result.getNode(), 1) };
17064 SDValue isValid = DAG.getNode(X86ISD::CMOV, dl,
17065 DAG.getVTList(Op->getValueType(1), MVT::Glue),
17068 // Return { result, isValid, chain }.
17069 return DAG.getNode(ISD::MERGE_VALUES, dl, Op->getVTList(), Result, isValid,
17070 SDValue(Result.getNode(), 2));
17073 //gather(v1, mask, index, base, scale);
17074 SDValue Chain = Op.getOperand(0);
17075 SDValue Src = Op.getOperand(2);
17076 SDValue Base = Op.getOperand(3);
17077 SDValue Index = Op.getOperand(4);
17078 SDValue Mask = Op.getOperand(5);
17079 SDValue Scale = Op.getOperand(6);
17080 return getGatherNode(IntrData->Opc0, Op, DAG, Src, Mask, Base, Index, Scale,
17084 //scatter(base, mask, index, v1, scale);
17085 SDValue Chain = Op.getOperand(0);
17086 SDValue Base = Op.getOperand(2);
17087 SDValue Mask = Op.getOperand(3);
17088 SDValue Index = Op.getOperand(4);
17089 SDValue Src = Op.getOperand(5);
17090 SDValue Scale = Op.getOperand(6);
17091 return getScatterNode(IntrData->Opc0, Op, DAG, Src, Mask, Base, Index,
17095 SDValue Hint = Op.getOperand(6);
17096 unsigned HintVal = cast<ConstantSDNode>(Hint)->getZExtValue();
17097 assert(HintVal < 2 && "Wrong prefetch hint in intrinsic: should be 0 or 1");
17098 unsigned Opcode = (HintVal ? IntrData->Opc1 : IntrData->Opc0);
17099 SDValue Chain = Op.getOperand(0);
17100 SDValue Mask = Op.getOperand(2);
17101 SDValue Index = Op.getOperand(3);
17102 SDValue Base = Op.getOperand(4);
17103 SDValue Scale = Op.getOperand(5);
17104 return getPrefetchNode(Opcode, Op, DAG, Mask, Base, Index, Scale, Chain);
17106 // Read Time Stamp Counter (RDTSC) and Processor ID (RDTSCP).
17108 SmallVector<SDValue, 2> Results;
17109 getReadTimeStampCounter(Op.getNode(), dl, IntrData->Opc0, DAG, Subtarget,
17111 return DAG.getMergeValues(Results, dl);
17113 // Read Performance Monitoring Counters.
17115 SmallVector<SDValue, 2> Results;
17116 getReadPerformanceCounter(Op.getNode(), dl, DAG, Subtarget, Results);
17117 return DAG.getMergeValues(Results, dl);
17119 // XTEST intrinsics.
17121 SDVTList VTs = DAG.getVTList(Op->getValueType(0), MVT::Other);
17122 SDValue InTrans = DAG.getNode(IntrData->Opc0, dl, VTs, Op.getOperand(0));
17123 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
17124 DAG.getConstant(X86::COND_NE, dl, MVT::i8),
17126 SDValue Ret = DAG.getNode(ISD::ZERO_EXTEND, dl, Op->getValueType(0), SetCC);
17127 return DAG.getNode(ISD::MERGE_VALUES, dl, Op->getVTList(),
17128 Ret, SDValue(InTrans.getNode(), 1));
17132 SmallVector<SDValue, 2> Results;
17133 SDVTList CFVTs = DAG.getVTList(Op->getValueType(0), MVT::Other);
17134 SDVTList VTs = DAG.getVTList(Op.getOperand(3)->getValueType(0), MVT::Other);
17135 SDValue GenCF = DAG.getNode(X86ISD::ADD, dl, CFVTs, Op.getOperand(2),
17136 DAG.getConstant(-1, dl, MVT::i8));
17137 SDValue Res = DAG.getNode(IntrData->Opc0, dl, VTs, Op.getOperand(3),
17138 Op.getOperand(4), GenCF.getValue(1));
17139 SDValue Store = DAG.getStore(Op.getOperand(0), dl, Res.getValue(0),
17140 Op.getOperand(5), MachinePointerInfo(),
17142 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
17143 DAG.getConstant(X86::COND_B, dl, MVT::i8),
17145 Results.push_back(SetCC);
17146 Results.push_back(Store);
17147 return DAG.getMergeValues(Results, dl);
17149 case COMPRESS_TO_MEM: {
17151 SDValue Mask = Op.getOperand(4);
17152 SDValue DataToCompress = Op.getOperand(3);
17153 SDValue Addr = Op.getOperand(2);
17154 SDValue Chain = Op.getOperand(0);
17156 EVT VT = DataToCompress.getValueType();
17157 if (isAllOnes(Mask)) // return just a store
17158 return DAG.getStore(Chain, dl, DataToCompress, Addr,
17159 MachinePointerInfo(), false, false,
17160 VT.getScalarSizeInBits()/8);
17162 SDValue Compressed =
17163 getVectorMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT, DataToCompress),
17164 Mask, DAG.getUNDEF(VT), Subtarget, DAG);
17165 return DAG.getStore(Chain, dl, Compressed, Addr,
17166 MachinePointerInfo(), false, false,
17167 VT.getScalarSizeInBits()/8);
17169 case TRUNCATE_TO_MEM_VI8:
17170 return LowerINTRINSIC_TRUNCATE_TO_MEM(Op, DAG, MVT::i8);
17171 case TRUNCATE_TO_MEM_VI16:
17172 return LowerINTRINSIC_TRUNCATE_TO_MEM(Op, DAG, MVT::i16);
17173 case TRUNCATE_TO_MEM_VI32:
17174 return LowerINTRINSIC_TRUNCATE_TO_MEM(Op, DAG, MVT::i32);
17175 case EXPAND_FROM_MEM: {
17177 SDValue Mask = Op.getOperand(4);
17178 SDValue PassThru = Op.getOperand(3);
17179 SDValue Addr = Op.getOperand(2);
17180 SDValue Chain = Op.getOperand(0);
17181 EVT VT = Op.getValueType();
17183 if (isAllOnes(Mask)) // return just a load
17184 return DAG.getLoad(VT, dl, Chain, Addr, MachinePointerInfo(), false, false,
17185 false, VT.getScalarSizeInBits()/8);
17187 SDValue DataToExpand = DAG.getLoad(VT, dl, Chain, Addr, MachinePointerInfo(),
17188 false, false, false,
17189 VT.getScalarSizeInBits()/8);
17191 SDValue Results[] = {
17192 getVectorMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT, DataToExpand),
17193 Mask, PassThru, Subtarget, DAG), Chain};
17194 return DAG.getMergeValues(Results, dl);
17199 SDValue X86TargetLowering::LowerRETURNADDR(SDValue Op,
17200 SelectionDAG &DAG) const {
17201 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
17202 MFI->setReturnAddressIsTaken(true);
17204 if (verifyReturnAddressArgumentIsConstant(Op, DAG))
17207 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
17209 EVT PtrVT = getPointerTy(DAG.getDataLayout());
17212 SDValue FrameAddr = LowerFRAMEADDR(Op, DAG);
17213 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
17214 SDValue Offset = DAG.getConstant(RegInfo->getSlotSize(), dl, PtrVT);
17215 return DAG.getLoad(PtrVT, dl, DAG.getEntryNode(),
17216 DAG.getNode(ISD::ADD, dl, PtrVT,
17217 FrameAddr, Offset),
17218 MachinePointerInfo(), false, false, false, 0);
17221 // Just load the return address.
17222 SDValue RetAddrFI = getReturnAddressFrameIndex(DAG);
17223 return DAG.getLoad(PtrVT, dl, DAG.getEntryNode(),
17224 RetAddrFI, MachinePointerInfo(), false, false, false, 0);
17227 SDValue X86TargetLowering::LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const {
17228 MachineFunction &MF = DAG.getMachineFunction();
17229 MachineFrameInfo *MFI = MF.getFrameInfo();
17230 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
17231 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
17232 EVT VT = Op.getValueType();
17234 MFI->setFrameAddressIsTaken(true);
17236 if (MF.getTarget().getMCAsmInfo()->usesWindowsCFI()) {
17237 // Depth > 0 makes no sense on targets which use Windows unwind codes. It
17238 // is not possible to crawl up the stack without looking at the unwind codes
17240 int FrameAddrIndex = FuncInfo->getFAIndex();
17241 if (!FrameAddrIndex) {
17242 // Set up a frame object for the return address.
17243 unsigned SlotSize = RegInfo->getSlotSize();
17244 FrameAddrIndex = MF.getFrameInfo()->CreateFixedObject(
17245 SlotSize, /*Offset=*/0, /*IsImmutable=*/false);
17246 FuncInfo->setFAIndex(FrameAddrIndex);
17248 return DAG.getFrameIndex(FrameAddrIndex, VT);
17251 unsigned FrameReg =
17252 RegInfo->getPtrSizedFrameRegister(DAG.getMachineFunction());
17253 SDLoc dl(Op); // FIXME probably not meaningful
17254 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
17255 assert(((FrameReg == X86::RBP && VT == MVT::i64) ||
17256 (FrameReg == X86::EBP && VT == MVT::i32)) &&
17257 "Invalid Frame Register!");
17258 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, VT);
17260 FrameAddr = DAG.getLoad(VT, dl, DAG.getEntryNode(), FrameAddr,
17261 MachinePointerInfo(),
17262 false, false, false, 0);
17266 // FIXME? Maybe this could be a TableGen attribute on some registers and
17267 // this table could be generated automatically from RegInfo.
17268 unsigned X86TargetLowering::getRegisterByName(const char* RegName, EVT VT,
17269 SelectionDAG &DAG) const {
17270 const TargetFrameLowering &TFI = *Subtarget->getFrameLowering();
17271 const MachineFunction &MF = DAG.getMachineFunction();
17273 unsigned Reg = StringSwitch<unsigned>(RegName)
17274 .Case("esp", X86::ESP)
17275 .Case("rsp", X86::RSP)
17276 .Case("ebp", X86::EBP)
17277 .Case("rbp", X86::RBP)
17280 if (Reg == X86::EBP || Reg == X86::RBP) {
17281 if (!TFI.hasFP(MF))
17282 report_fatal_error("register " + StringRef(RegName) +
17283 " is allocatable: function has no frame pointer");
17286 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
17287 unsigned FrameReg =
17288 RegInfo->getPtrSizedFrameRegister(DAG.getMachineFunction());
17289 assert((FrameReg == X86::EBP || FrameReg == X86::RBP) &&
17290 "Invalid Frame Register!");
17298 report_fatal_error("Invalid register name global variable");
17301 SDValue X86TargetLowering::LowerFRAME_TO_ARGS_OFFSET(SDValue Op,
17302 SelectionDAG &DAG) const {
17303 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
17304 return DAG.getIntPtrConstant(2 * RegInfo->getSlotSize(), SDLoc(Op));
17307 SDValue X86TargetLowering::LowerEH_RETURN(SDValue Op, SelectionDAG &DAG) const {
17308 SDValue Chain = Op.getOperand(0);
17309 SDValue Offset = Op.getOperand(1);
17310 SDValue Handler = Op.getOperand(2);
17313 EVT PtrVT = getPointerTy(DAG.getDataLayout());
17314 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
17315 unsigned FrameReg = RegInfo->getFrameRegister(DAG.getMachineFunction());
17316 assert(((FrameReg == X86::RBP && PtrVT == MVT::i64) ||
17317 (FrameReg == X86::EBP && PtrVT == MVT::i32)) &&
17318 "Invalid Frame Register!");
17319 SDValue Frame = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, PtrVT);
17320 unsigned StoreAddrReg = (PtrVT == MVT::i64) ? X86::RCX : X86::ECX;
17322 SDValue StoreAddr = DAG.getNode(ISD::ADD, dl, PtrVT, Frame,
17323 DAG.getIntPtrConstant(RegInfo->getSlotSize(),
17325 StoreAddr = DAG.getNode(ISD::ADD, dl, PtrVT, StoreAddr, Offset);
17326 Chain = DAG.getStore(Chain, dl, Handler, StoreAddr, MachinePointerInfo(),
17328 Chain = DAG.getCopyToReg(Chain, dl, StoreAddrReg, StoreAddr);
17330 return DAG.getNode(X86ISD::EH_RETURN, dl, MVT::Other, Chain,
17331 DAG.getRegister(StoreAddrReg, PtrVT));
17334 SDValue X86TargetLowering::lowerEH_SJLJ_SETJMP(SDValue Op,
17335 SelectionDAG &DAG) const {
17337 return DAG.getNode(X86ISD::EH_SJLJ_SETJMP, DL,
17338 DAG.getVTList(MVT::i32, MVT::Other),
17339 Op.getOperand(0), Op.getOperand(1));
17342 SDValue X86TargetLowering::lowerEH_SJLJ_LONGJMP(SDValue Op,
17343 SelectionDAG &DAG) const {
17345 return DAG.getNode(X86ISD::EH_SJLJ_LONGJMP, DL, MVT::Other,
17346 Op.getOperand(0), Op.getOperand(1));
17349 static SDValue LowerADJUST_TRAMPOLINE(SDValue Op, SelectionDAG &DAG) {
17350 return Op.getOperand(0);
17353 SDValue X86TargetLowering::LowerINIT_TRAMPOLINE(SDValue Op,
17354 SelectionDAG &DAG) const {
17355 SDValue Root = Op.getOperand(0);
17356 SDValue Trmp = Op.getOperand(1); // trampoline
17357 SDValue FPtr = Op.getOperand(2); // nested function
17358 SDValue Nest = Op.getOperand(3); // 'nest' parameter value
17361 const Value *TrmpAddr = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
17362 const TargetRegisterInfo *TRI = Subtarget->getRegisterInfo();
17364 if (Subtarget->is64Bit()) {
17365 SDValue OutChains[6];
17367 // Large code-model.
17368 const unsigned char JMP64r = 0xFF; // 64-bit jmp through register opcode.
17369 const unsigned char MOV64ri = 0xB8; // X86::MOV64ri opcode.
17371 const unsigned char N86R10 = TRI->getEncodingValue(X86::R10) & 0x7;
17372 const unsigned char N86R11 = TRI->getEncodingValue(X86::R11) & 0x7;
17374 const unsigned char REX_WB = 0x40 | 0x08 | 0x01; // REX prefix
17376 // Load the pointer to the nested function into R11.
17377 unsigned OpCode = ((MOV64ri | N86R11) << 8) | REX_WB; // movabsq r11
17378 SDValue Addr = Trmp;
17379 OutChains[0] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, dl, MVT::i16),
17380 Addr, MachinePointerInfo(TrmpAddr),
17383 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
17384 DAG.getConstant(2, dl, MVT::i64));
17385 OutChains[1] = DAG.getStore(Root, dl, FPtr, Addr,
17386 MachinePointerInfo(TrmpAddr, 2),
17389 // Load the 'nest' parameter value into R10.
17390 // R10 is specified in X86CallingConv.td
17391 OpCode = ((MOV64ri | N86R10) << 8) | REX_WB; // movabsq r10
17392 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
17393 DAG.getConstant(10, dl, MVT::i64));
17394 OutChains[2] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, dl, MVT::i16),
17395 Addr, MachinePointerInfo(TrmpAddr, 10),
17398 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
17399 DAG.getConstant(12, dl, MVT::i64));
17400 OutChains[3] = DAG.getStore(Root, dl, Nest, Addr,
17401 MachinePointerInfo(TrmpAddr, 12),
17404 // Jump to the nested function.
17405 OpCode = (JMP64r << 8) | REX_WB; // jmpq *...
17406 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
17407 DAG.getConstant(20, dl, MVT::i64));
17408 OutChains[4] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, dl, MVT::i16),
17409 Addr, MachinePointerInfo(TrmpAddr, 20),
17412 unsigned char ModRM = N86R11 | (4 << 3) | (3 << 6); // ...r11
17413 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
17414 DAG.getConstant(22, dl, MVT::i64));
17415 OutChains[5] = DAG.getStore(Root, dl, DAG.getConstant(ModRM, dl, MVT::i8),
17416 Addr, MachinePointerInfo(TrmpAddr, 22),
17419 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other, OutChains);
17421 const Function *Func =
17422 cast<Function>(cast<SrcValueSDNode>(Op.getOperand(5))->getValue());
17423 CallingConv::ID CC = Func->getCallingConv();
17428 llvm_unreachable("Unsupported calling convention");
17429 case CallingConv::C:
17430 case CallingConv::X86_StdCall: {
17431 // Pass 'nest' parameter in ECX.
17432 // Must be kept in sync with X86CallingConv.td
17433 NestReg = X86::ECX;
17435 // Check that ECX wasn't needed by an 'inreg' parameter.
17436 FunctionType *FTy = Func->getFunctionType();
17437 const AttributeSet &Attrs = Func->getAttributes();
17439 if (!Attrs.isEmpty() && !Func->isVarArg()) {
17440 unsigned InRegCount = 0;
17443 for (FunctionType::param_iterator I = FTy->param_begin(),
17444 E = FTy->param_end(); I != E; ++I, ++Idx)
17445 if (Attrs.hasAttribute(Idx, Attribute::InReg)) {
17446 auto &DL = DAG.getDataLayout();
17447 // FIXME: should only count parameters that are lowered to integers.
17448 InRegCount += (DL.getTypeSizeInBits(*I) + 31) / 32;
17451 if (InRegCount > 2) {
17452 report_fatal_error("Nest register in use - reduce number of inreg"
17458 case CallingConv::X86_FastCall:
17459 case CallingConv::X86_ThisCall:
17460 case CallingConv::Fast:
17461 // Pass 'nest' parameter in EAX.
17462 // Must be kept in sync with X86CallingConv.td
17463 NestReg = X86::EAX;
17467 SDValue OutChains[4];
17468 SDValue Addr, Disp;
17470 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
17471 DAG.getConstant(10, dl, MVT::i32));
17472 Disp = DAG.getNode(ISD::SUB, dl, MVT::i32, FPtr, Addr);
17474 // This is storing the opcode for MOV32ri.
17475 const unsigned char MOV32ri = 0xB8; // X86::MOV32ri's opcode byte.
17476 const unsigned char N86Reg = TRI->getEncodingValue(NestReg) & 0x7;
17477 OutChains[0] = DAG.getStore(Root, dl,
17478 DAG.getConstant(MOV32ri|N86Reg, dl, MVT::i8),
17479 Trmp, MachinePointerInfo(TrmpAddr),
17482 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
17483 DAG.getConstant(1, dl, MVT::i32));
17484 OutChains[1] = DAG.getStore(Root, dl, Nest, Addr,
17485 MachinePointerInfo(TrmpAddr, 1),
17488 const unsigned char JMP = 0xE9; // jmp <32bit dst> opcode.
17489 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
17490 DAG.getConstant(5, dl, MVT::i32));
17491 OutChains[2] = DAG.getStore(Root, dl, DAG.getConstant(JMP, dl, MVT::i8),
17492 Addr, MachinePointerInfo(TrmpAddr, 5),
17495 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
17496 DAG.getConstant(6, dl, MVT::i32));
17497 OutChains[3] = DAG.getStore(Root, dl, Disp, Addr,
17498 MachinePointerInfo(TrmpAddr, 6),
17501 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other, OutChains);
17505 SDValue X86TargetLowering::LowerFLT_ROUNDS_(SDValue Op,
17506 SelectionDAG &DAG) const {
17508 The rounding mode is in bits 11:10 of FPSR, and has the following
17510 00 Round to nearest
17515 FLT_ROUNDS, on the other hand, expects the following:
17522 To perform the conversion, we do:
17523 (((((FPSR & 0x800) >> 11) | ((FPSR & 0x400) >> 9)) + 1) & 3)
17526 MachineFunction &MF = DAG.getMachineFunction();
17527 const TargetFrameLowering &TFI = *Subtarget->getFrameLowering();
17528 unsigned StackAlignment = TFI.getStackAlignment();
17529 MVT VT = Op.getSimpleValueType();
17532 // Save FP Control Word to stack slot
17533 int SSFI = MF.getFrameInfo()->CreateStackObject(2, StackAlignment, false);
17534 SDValue StackSlot =
17535 DAG.getFrameIndex(SSFI, getPointerTy(DAG.getDataLayout()));
17537 MachineMemOperand *MMO =
17538 MF.getMachineMemOperand(MachinePointerInfo::getFixedStack(MF, SSFI),
17539 MachineMemOperand::MOStore, 2, 2);
17541 SDValue Ops[] = { DAG.getEntryNode(), StackSlot };
17542 SDValue Chain = DAG.getMemIntrinsicNode(X86ISD::FNSTCW16m, DL,
17543 DAG.getVTList(MVT::Other),
17544 Ops, MVT::i16, MMO);
17546 // Load FP Control Word from stack slot
17547 SDValue CWD = DAG.getLoad(MVT::i16, DL, Chain, StackSlot,
17548 MachinePointerInfo(), false, false, false, 0);
17550 // Transform as necessary
17552 DAG.getNode(ISD::SRL, DL, MVT::i16,
17553 DAG.getNode(ISD::AND, DL, MVT::i16,
17554 CWD, DAG.getConstant(0x800, DL, MVT::i16)),
17555 DAG.getConstant(11, DL, MVT::i8));
17557 DAG.getNode(ISD::SRL, DL, MVT::i16,
17558 DAG.getNode(ISD::AND, DL, MVT::i16,
17559 CWD, DAG.getConstant(0x400, DL, MVT::i16)),
17560 DAG.getConstant(9, DL, MVT::i8));
17563 DAG.getNode(ISD::AND, DL, MVT::i16,
17564 DAG.getNode(ISD::ADD, DL, MVT::i16,
17565 DAG.getNode(ISD::OR, DL, MVT::i16, CWD1, CWD2),
17566 DAG.getConstant(1, DL, MVT::i16)),
17567 DAG.getConstant(3, DL, MVT::i16));
17569 return DAG.getNode((VT.getSizeInBits() < 16 ?
17570 ISD::TRUNCATE : ISD::ZERO_EXTEND), DL, VT, RetVal);
17573 /// \brief Lower a vector CTLZ using native supported vector CTLZ instruction.
17575 // 1. i32/i64 128/256-bit vector (native support require VLX) are expended
17576 // to 512-bit vector.
17577 // 2. i8/i16 vector implemented using dword LZCNT vector instruction
17578 // ( sub(trunc(lzcnt(zext32(x)))) ). In case zext32(x) is illegal,
17579 // split the vector, perform operation on it's Lo a Hi part and
17580 // concatenate the results.
17581 static SDValue LowerVectorCTLZ_AVX512(SDValue Op, SelectionDAG &DAG) {
17583 MVT VT = Op.getSimpleValueType();
17584 MVT EltVT = VT.getVectorElementType();
17585 unsigned NumElems = VT.getVectorNumElements();
17587 if (EltVT == MVT::i64 || EltVT == MVT::i32) {
17588 // Extend to 512 bit vector.
17589 assert((VT.is256BitVector() || VT.is128BitVector()) &&
17590 "Unsupported value type for operation");
17592 MVT NewVT = MVT::getVectorVT(EltVT, 512 / VT.getScalarSizeInBits());
17593 SDValue Vec512 = DAG.getNode(ISD::INSERT_SUBVECTOR, dl, NewVT,
17594 DAG.getUNDEF(NewVT),
17596 DAG.getIntPtrConstant(0, dl));
17597 SDValue CtlzNode = DAG.getNode(ISD::CTLZ, dl, NewVT, Vec512);
17599 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT, CtlzNode,
17600 DAG.getIntPtrConstant(0, dl));
17603 assert((EltVT == MVT::i8 || EltVT == MVT::i16) &&
17604 "Unsupported element type");
17606 if (16 < NumElems) {
17607 // Split vector, it's Lo and Hi parts will be handled in next iteration.
17609 std::tie(Lo, Hi) = DAG.SplitVector(Op.getOperand(0), dl);
17610 MVT OutVT = MVT::getVectorVT(EltVT, NumElems/2);
17612 Lo = DAG.getNode(Op.getOpcode(), dl, OutVT, Lo);
17613 Hi = DAG.getNode(Op.getOpcode(), dl, OutVT, Hi);
17615 return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT, Lo, Hi);
17618 MVT NewVT = MVT::getVectorVT(MVT::i32, NumElems);
17620 assert((NewVT.is256BitVector() || NewVT.is512BitVector()) &&
17621 "Unsupported value type for operation");
17623 // Use native supported vector instruction vplzcntd.
17624 Op = DAG.getNode(ISD::ZERO_EXTEND, dl, NewVT, Op.getOperand(0));
17625 SDValue CtlzNode = DAG.getNode(ISD::CTLZ, dl, NewVT, Op);
17626 SDValue TruncNode = DAG.getNode(ISD::TRUNCATE, dl, VT, CtlzNode);
17627 SDValue Delta = DAG.getConstant(32 - EltVT.getSizeInBits(), dl, VT);
17629 return DAG.getNode(ISD::SUB, dl, VT, TruncNode, Delta);
17632 static SDValue LowerCTLZ(SDValue Op, const X86Subtarget *Subtarget,
17633 SelectionDAG &DAG) {
17634 MVT VT = Op.getSimpleValueType();
17636 unsigned NumBits = VT.getSizeInBits();
17639 if (VT.isVector() && Subtarget->hasAVX512())
17640 return LowerVectorCTLZ_AVX512(Op, DAG);
17642 Op = Op.getOperand(0);
17643 if (VT == MVT::i8) {
17644 // Zero extend to i32 since there is not an i8 bsr.
17646 Op = DAG.getNode(ISD::ZERO_EXTEND, dl, OpVT, Op);
17649 // Issue a bsr (scan bits in reverse) which also sets EFLAGS.
17650 SDVTList VTs = DAG.getVTList(OpVT, MVT::i32);
17651 Op = DAG.getNode(X86ISD::BSR, dl, VTs, Op);
17653 // If src is zero (i.e. bsr sets ZF), returns NumBits.
17656 DAG.getConstant(NumBits + NumBits - 1, dl, OpVT),
17657 DAG.getConstant(X86::COND_E, dl, MVT::i8),
17660 Op = DAG.getNode(X86ISD::CMOV, dl, OpVT, Ops);
17662 // Finally xor with NumBits-1.
17663 Op = DAG.getNode(ISD::XOR, dl, OpVT, Op,
17664 DAG.getConstant(NumBits - 1, dl, OpVT));
17667 Op = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op);
17671 static SDValue LowerCTLZ_ZERO_UNDEF(SDValue Op, const X86Subtarget *Subtarget,
17672 SelectionDAG &DAG) {
17673 MVT VT = Op.getSimpleValueType();
17675 unsigned NumBits = VT.getSizeInBits();
17678 if (VT.isVector() && Subtarget->hasAVX512())
17679 return LowerVectorCTLZ_AVX512(Op, DAG);
17681 Op = Op.getOperand(0);
17682 if (VT == MVT::i8) {
17683 // Zero extend to i32 since there is not an i8 bsr.
17685 Op = DAG.getNode(ISD::ZERO_EXTEND, dl, OpVT, Op);
17688 // Issue a bsr (scan bits in reverse).
17689 SDVTList VTs = DAG.getVTList(OpVT, MVT::i32);
17690 Op = DAG.getNode(X86ISD::BSR, dl, VTs, Op);
17692 // And xor with NumBits-1.
17693 Op = DAG.getNode(ISD::XOR, dl, OpVT, Op,
17694 DAG.getConstant(NumBits - 1, dl, OpVT));
17697 Op = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op);
17701 static SDValue LowerCTTZ(SDValue Op, SelectionDAG &DAG) {
17702 MVT VT = Op.getSimpleValueType();
17703 unsigned NumBits = VT.getScalarSizeInBits();
17706 if (VT.isVector()) {
17707 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
17709 SDValue N0 = Op.getOperand(0);
17710 SDValue Zero = DAG.getConstant(0, dl, VT);
17712 // lsb(x) = (x & -x)
17713 SDValue LSB = DAG.getNode(ISD::AND, dl, VT, N0,
17714 DAG.getNode(ISD::SUB, dl, VT, Zero, N0));
17716 // cttz_undef(x) = (width - 1) - ctlz(lsb)
17717 if (Op.getOpcode() == ISD::CTTZ_ZERO_UNDEF &&
17718 TLI.isOperationLegal(ISD::CTLZ, VT)) {
17719 SDValue WidthMinusOne = DAG.getConstant(NumBits - 1, dl, VT);
17720 return DAG.getNode(ISD::SUB, dl, VT, WidthMinusOne,
17721 DAG.getNode(ISD::CTLZ, dl, VT, LSB));
17724 // cttz(x) = ctpop(lsb - 1)
17725 SDValue One = DAG.getConstant(1, dl, VT);
17726 return DAG.getNode(ISD::CTPOP, dl, VT,
17727 DAG.getNode(ISD::SUB, dl, VT, LSB, One));
17730 assert(Op.getOpcode() == ISD::CTTZ &&
17731 "Only scalar CTTZ requires custom lowering");
17733 // Issue a bsf (scan bits forward) which also sets EFLAGS.
17734 SDVTList VTs = DAG.getVTList(VT, MVT::i32);
17735 Op = DAG.getNode(X86ISD::BSF, dl, VTs, Op.getOperand(0));
17737 // If src is zero (i.e. bsf sets ZF), returns NumBits.
17740 DAG.getConstant(NumBits, dl, VT),
17741 DAG.getConstant(X86::COND_E, dl, MVT::i8),
17744 return DAG.getNode(X86ISD::CMOV, dl, VT, Ops);
17747 // Lower256IntArith - Break a 256-bit integer operation into two new 128-bit
17748 // ones, and then concatenate the result back.
17749 static SDValue Lower256IntArith(SDValue Op, SelectionDAG &DAG) {
17750 MVT VT = Op.getSimpleValueType();
17752 assert(VT.is256BitVector() && VT.isInteger() &&
17753 "Unsupported value type for operation");
17755 unsigned NumElems = VT.getVectorNumElements();
17758 // Extract the LHS vectors
17759 SDValue LHS = Op.getOperand(0);
17760 SDValue LHS1 = Extract128BitVector(LHS, 0, DAG, dl);
17761 SDValue LHS2 = Extract128BitVector(LHS, NumElems/2, DAG, dl);
17763 // Extract the RHS vectors
17764 SDValue RHS = Op.getOperand(1);
17765 SDValue RHS1 = Extract128BitVector(RHS, 0, DAG, dl);
17766 SDValue RHS2 = Extract128BitVector(RHS, NumElems/2, DAG, dl);
17768 MVT EltVT = VT.getVectorElementType();
17769 MVT NewVT = MVT::getVectorVT(EltVT, NumElems/2);
17771 return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT,
17772 DAG.getNode(Op.getOpcode(), dl, NewVT, LHS1, RHS1),
17773 DAG.getNode(Op.getOpcode(), dl, NewVT, LHS2, RHS2));
17776 static SDValue LowerADD(SDValue Op, SelectionDAG &DAG) {
17777 if (Op.getValueType() == MVT::i1)
17778 return DAG.getNode(ISD::XOR, SDLoc(Op), Op.getValueType(),
17779 Op.getOperand(0), Op.getOperand(1));
17780 assert(Op.getSimpleValueType().is256BitVector() &&
17781 Op.getSimpleValueType().isInteger() &&
17782 "Only handle AVX 256-bit vector integer operation");
17783 return Lower256IntArith(Op, DAG);
17786 static SDValue LowerSUB(SDValue Op, SelectionDAG &DAG) {
17787 if (Op.getValueType() == MVT::i1)
17788 return DAG.getNode(ISD::XOR, SDLoc(Op), Op.getValueType(),
17789 Op.getOperand(0), Op.getOperand(1));
17790 assert(Op.getSimpleValueType().is256BitVector() &&
17791 Op.getSimpleValueType().isInteger() &&
17792 "Only handle AVX 256-bit vector integer operation");
17793 return Lower256IntArith(Op, DAG);
17796 static SDValue LowerMINMAX(SDValue Op, SelectionDAG &DAG) {
17797 assert(Op.getSimpleValueType().is256BitVector() &&
17798 Op.getSimpleValueType().isInteger() &&
17799 "Only handle AVX 256-bit vector integer operation");
17800 return Lower256IntArith(Op, DAG);
17803 static SDValue LowerMUL(SDValue Op, const X86Subtarget *Subtarget,
17804 SelectionDAG &DAG) {
17806 MVT VT = Op.getSimpleValueType();
17809 return DAG.getNode(ISD::AND, dl, VT, Op.getOperand(0), Op.getOperand(1));
17811 // Decompose 256-bit ops into smaller 128-bit ops.
17812 if (VT.is256BitVector() && !Subtarget->hasInt256())
17813 return Lower256IntArith(Op, DAG);
17815 SDValue A = Op.getOperand(0);
17816 SDValue B = Op.getOperand(1);
17818 // Lower v16i8/v32i8 mul as promotion to v8i16/v16i16 vector
17819 // pairs, multiply and truncate.
17820 if (VT == MVT::v16i8 || VT == MVT::v32i8) {
17821 if (Subtarget->hasInt256()) {
17822 if (VT == MVT::v32i8) {
17823 MVT SubVT = MVT::getVectorVT(MVT::i8, VT.getVectorNumElements() / 2);
17824 SDValue Lo = DAG.getIntPtrConstant(0, dl);
17825 SDValue Hi = DAG.getIntPtrConstant(VT.getVectorNumElements() / 2, dl);
17826 SDValue ALo = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, SubVT, A, Lo);
17827 SDValue BLo = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, SubVT, B, Lo);
17828 SDValue AHi = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, SubVT, A, Hi);
17829 SDValue BHi = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, SubVT, B, Hi);
17830 return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT,
17831 DAG.getNode(ISD::MUL, dl, SubVT, ALo, BLo),
17832 DAG.getNode(ISD::MUL, dl, SubVT, AHi, BHi));
17835 MVT ExVT = MVT::getVectorVT(MVT::i16, VT.getVectorNumElements());
17836 return DAG.getNode(
17837 ISD::TRUNCATE, dl, VT,
17838 DAG.getNode(ISD::MUL, dl, ExVT,
17839 DAG.getNode(ISD::SIGN_EXTEND, dl, ExVT, A),
17840 DAG.getNode(ISD::SIGN_EXTEND, dl, ExVT, B)));
17843 assert(VT == MVT::v16i8 &&
17844 "Pre-AVX2 support only supports v16i8 multiplication");
17845 MVT ExVT = MVT::v8i16;
17847 // Extract the lo parts and sign extend to i16
17849 if (Subtarget->hasSSE41()) {
17850 ALo = DAG.getNode(X86ISD::VSEXT, dl, ExVT, A);
17851 BLo = DAG.getNode(X86ISD::VSEXT, dl, ExVT, B);
17853 const int ShufMask[] = {-1, 0, -1, 1, -1, 2, -1, 3,
17854 -1, 4, -1, 5, -1, 6, -1, 7};
17855 ALo = DAG.getVectorShuffle(VT, dl, A, A, ShufMask);
17856 BLo = DAG.getVectorShuffle(VT, dl, B, B, ShufMask);
17857 ALo = DAG.getBitcast(ExVT, ALo);
17858 BLo = DAG.getBitcast(ExVT, BLo);
17859 ALo = DAG.getNode(ISD::SRA, dl, ExVT, ALo, DAG.getConstant(8, dl, ExVT));
17860 BLo = DAG.getNode(ISD::SRA, dl, ExVT, BLo, DAG.getConstant(8, dl, ExVT));
17863 // Extract the hi parts and sign extend to i16
17865 if (Subtarget->hasSSE41()) {
17866 const int ShufMask[] = {8, 9, 10, 11, 12, 13, 14, 15,
17867 -1, -1, -1, -1, -1, -1, -1, -1};
17868 AHi = DAG.getVectorShuffle(VT, dl, A, A, ShufMask);
17869 BHi = DAG.getVectorShuffle(VT, dl, B, B, ShufMask);
17870 AHi = DAG.getNode(X86ISD::VSEXT, dl, ExVT, AHi);
17871 BHi = DAG.getNode(X86ISD::VSEXT, dl, ExVT, BHi);
17873 const int ShufMask[] = {-1, 8, -1, 9, -1, 10, -1, 11,
17874 -1, 12, -1, 13, -1, 14, -1, 15};
17875 AHi = DAG.getVectorShuffle(VT, dl, A, A, ShufMask);
17876 BHi = DAG.getVectorShuffle(VT, dl, B, B, ShufMask);
17877 AHi = DAG.getBitcast(ExVT, AHi);
17878 BHi = DAG.getBitcast(ExVT, BHi);
17879 AHi = DAG.getNode(ISD::SRA, dl, ExVT, AHi, DAG.getConstant(8, dl, ExVT));
17880 BHi = DAG.getNode(ISD::SRA, dl, ExVT, BHi, DAG.getConstant(8, dl, ExVT));
17883 // Multiply, mask the lower 8bits of the lo/hi results and pack
17884 SDValue RLo = DAG.getNode(ISD::MUL, dl, ExVT, ALo, BLo);
17885 SDValue RHi = DAG.getNode(ISD::MUL, dl, ExVT, AHi, BHi);
17886 RLo = DAG.getNode(ISD::AND, dl, ExVT, RLo, DAG.getConstant(255, dl, ExVT));
17887 RHi = DAG.getNode(ISD::AND, dl, ExVT, RHi, DAG.getConstant(255, dl, ExVT));
17888 return DAG.getNode(X86ISD::PACKUS, dl, VT, RLo, RHi);
17891 // Lower v4i32 mul as 2x shuffle, 2x pmuludq, 2x shuffle.
17892 if (VT == MVT::v4i32) {
17893 assert(Subtarget->hasSSE2() && !Subtarget->hasSSE41() &&
17894 "Should not custom lower when pmuldq is available!");
17896 // Extract the odd parts.
17897 static const int UnpackMask[] = { 1, -1, 3, -1 };
17898 SDValue Aodds = DAG.getVectorShuffle(VT, dl, A, A, UnpackMask);
17899 SDValue Bodds = DAG.getVectorShuffle(VT, dl, B, B, UnpackMask);
17901 // Multiply the even parts.
17902 SDValue Evens = DAG.getNode(X86ISD::PMULUDQ, dl, MVT::v2i64, A, B);
17903 // Now multiply odd parts.
17904 SDValue Odds = DAG.getNode(X86ISD::PMULUDQ, dl, MVT::v2i64, Aodds, Bodds);
17906 Evens = DAG.getBitcast(VT, Evens);
17907 Odds = DAG.getBitcast(VT, Odds);
17909 // Merge the two vectors back together with a shuffle. This expands into 2
17911 static const int ShufMask[] = { 0, 4, 2, 6 };
17912 return DAG.getVectorShuffle(VT, dl, Evens, Odds, ShufMask);
17915 assert((VT == MVT::v2i64 || VT == MVT::v4i64 || VT == MVT::v8i64) &&
17916 "Only know how to lower V2I64/V4I64/V8I64 multiply");
17918 // Ahi = psrlqi(a, 32);
17919 // Bhi = psrlqi(b, 32);
17921 // AloBlo = pmuludq(a, b);
17922 // AloBhi = pmuludq(a, Bhi);
17923 // AhiBlo = pmuludq(Ahi, b);
17925 // AloBhi = psllqi(AloBhi, 32);
17926 // AhiBlo = psllqi(AhiBlo, 32);
17927 // return AloBlo + AloBhi + AhiBlo;
17929 SDValue Ahi = getTargetVShiftByConstNode(X86ISD::VSRLI, dl, VT, A, 32, DAG);
17930 SDValue Bhi = getTargetVShiftByConstNode(X86ISD::VSRLI, dl, VT, B, 32, DAG);
17932 SDValue AhiBlo = Ahi;
17933 SDValue AloBhi = Bhi;
17934 // Bit cast to 32-bit vectors for MULUDQ
17935 EVT MulVT = (VT == MVT::v2i64) ? MVT::v4i32 :
17936 (VT == MVT::v4i64) ? MVT::v8i32 : MVT::v16i32;
17937 A = DAG.getBitcast(MulVT, A);
17938 B = DAG.getBitcast(MulVT, B);
17939 Ahi = DAG.getBitcast(MulVT, Ahi);
17940 Bhi = DAG.getBitcast(MulVT, Bhi);
17942 SDValue AloBlo = DAG.getNode(X86ISD::PMULUDQ, dl, VT, A, B);
17943 // After shifting right const values the result may be all-zero.
17944 if (!ISD::isBuildVectorAllZeros(Ahi.getNode())) {
17945 AhiBlo = DAG.getNode(X86ISD::PMULUDQ, dl, VT, Ahi, B);
17946 AhiBlo = getTargetVShiftByConstNode(X86ISD::VSHLI, dl, VT, AhiBlo, 32, DAG);
17948 if (!ISD::isBuildVectorAllZeros(Bhi.getNode())) {
17949 AloBhi = DAG.getNode(X86ISD::PMULUDQ, dl, VT, A, Bhi);
17950 AloBhi = getTargetVShiftByConstNode(X86ISD::VSHLI, dl, VT, AloBhi, 32, DAG);
17953 SDValue Res = DAG.getNode(ISD::ADD, dl, VT, AloBlo, AloBhi);
17954 return DAG.getNode(ISD::ADD, dl, VT, Res, AhiBlo);
17957 SDValue X86TargetLowering::LowerWin64_i128OP(SDValue Op, SelectionDAG &DAG) const {
17958 assert(Subtarget->isTargetWin64() && "Unexpected target");
17959 EVT VT = Op.getValueType();
17960 assert(VT.isInteger() && VT.getSizeInBits() == 128 &&
17961 "Unexpected return type for lowering");
17965 switch (Op->getOpcode()) {
17966 default: llvm_unreachable("Unexpected request for libcall!");
17967 case ISD::SDIV: isSigned = true; LC = RTLIB::SDIV_I128; break;
17968 case ISD::UDIV: isSigned = false; LC = RTLIB::UDIV_I128; break;
17969 case ISD::SREM: isSigned = true; LC = RTLIB::SREM_I128; break;
17970 case ISD::UREM: isSigned = false; LC = RTLIB::UREM_I128; break;
17971 case ISD::SDIVREM: isSigned = true; LC = RTLIB::SDIVREM_I128; break;
17972 case ISD::UDIVREM: isSigned = false; LC = RTLIB::UDIVREM_I128; break;
17976 SDValue InChain = DAG.getEntryNode();
17978 TargetLowering::ArgListTy Args;
17979 TargetLowering::ArgListEntry Entry;
17980 for (unsigned i = 0, e = Op->getNumOperands(); i != e; ++i) {
17981 EVT ArgVT = Op->getOperand(i).getValueType();
17982 assert(ArgVT.isInteger() && ArgVT.getSizeInBits() == 128 &&
17983 "Unexpected argument type for lowering");
17984 SDValue StackPtr = DAG.CreateStackTemporary(ArgVT, 16);
17985 Entry.Node = StackPtr;
17986 InChain = DAG.getStore(InChain, dl, Op->getOperand(i), StackPtr, MachinePointerInfo(),
17988 Type *ArgTy = ArgVT.getTypeForEVT(*DAG.getContext());
17989 Entry.Ty = PointerType::get(ArgTy,0);
17990 Entry.isSExt = false;
17991 Entry.isZExt = false;
17992 Args.push_back(Entry);
17995 SDValue Callee = DAG.getExternalSymbol(getLibcallName(LC),
17996 getPointerTy(DAG.getDataLayout()));
17998 TargetLowering::CallLoweringInfo CLI(DAG);
17999 CLI.setDebugLoc(dl).setChain(InChain)
18000 .setCallee(getLibcallCallingConv(LC),
18001 static_cast<EVT>(MVT::v2i64).getTypeForEVT(*DAG.getContext()),
18002 Callee, std::move(Args), 0)
18003 .setInRegister().setSExtResult(isSigned).setZExtResult(!isSigned);
18005 std::pair<SDValue, SDValue> CallInfo = LowerCallTo(CLI);
18006 return DAG.getBitcast(VT, CallInfo.first);
18009 static SDValue LowerMUL_LOHI(SDValue Op, const X86Subtarget *Subtarget,
18010 SelectionDAG &DAG) {
18011 SDValue Op0 = Op.getOperand(0), Op1 = Op.getOperand(1);
18012 EVT VT = Op0.getValueType();
18015 assert((VT == MVT::v4i32 && Subtarget->hasSSE2()) ||
18016 (VT == MVT::v8i32 && Subtarget->hasInt256()));
18018 // PMULxD operations multiply each even value (starting at 0) of LHS with
18019 // the related value of RHS and produce a widen result.
18020 // E.g., PMULUDQ <4 x i32> <a|b|c|d>, <4 x i32> <e|f|g|h>
18021 // => <2 x i64> <ae|cg>
18023 // In other word, to have all the results, we need to perform two PMULxD:
18024 // 1. one with the even values.
18025 // 2. one with the odd values.
18026 // To achieve #2, with need to place the odd values at an even position.
18028 // Place the odd value at an even position (basically, shift all values 1
18029 // step to the left):
18030 const int Mask[] = {1, -1, 3, -1, 5, -1, 7, -1};
18031 // <a|b|c|d> => <b|undef|d|undef>
18032 SDValue Odd0 = DAG.getVectorShuffle(VT, dl, Op0, Op0, Mask);
18033 // <e|f|g|h> => <f|undef|h|undef>
18034 SDValue Odd1 = DAG.getVectorShuffle(VT, dl, Op1, Op1, Mask);
18036 // Emit two multiplies, one for the lower 2 ints and one for the higher 2
18038 MVT MulVT = VT == MVT::v4i32 ? MVT::v2i64 : MVT::v4i64;
18039 bool IsSigned = Op->getOpcode() == ISD::SMUL_LOHI;
18041 (!IsSigned || !Subtarget->hasSSE41()) ? X86ISD::PMULUDQ : X86ISD::PMULDQ;
18042 // PMULUDQ <4 x i32> <a|b|c|d>, <4 x i32> <e|f|g|h>
18043 // => <2 x i64> <ae|cg>
18044 SDValue Mul1 = DAG.getBitcast(VT, DAG.getNode(Opcode, dl, MulVT, Op0, Op1));
18045 // PMULUDQ <4 x i32> <b|undef|d|undef>, <4 x i32> <f|undef|h|undef>
18046 // => <2 x i64> <bf|dh>
18047 SDValue Mul2 = DAG.getBitcast(VT, DAG.getNode(Opcode, dl, MulVT, Odd0, Odd1));
18049 // Shuffle it back into the right order.
18050 SDValue Highs, Lows;
18051 if (VT == MVT::v8i32) {
18052 const int HighMask[] = {1, 9, 3, 11, 5, 13, 7, 15};
18053 Highs = DAG.getVectorShuffle(VT, dl, Mul1, Mul2, HighMask);
18054 const int LowMask[] = {0, 8, 2, 10, 4, 12, 6, 14};
18055 Lows = DAG.getVectorShuffle(VT, dl, Mul1, Mul2, LowMask);
18057 const int HighMask[] = {1, 5, 3, 7};
18058 Highs = DAG.getVectorShuffle(VT, dl, Mul1, Mul2, HighMask);
18059 const int LowMask[] = {0, 4, 2, 6};
18060 Lows = DAG.getVectorShuffle(VT, dl, Mul1, Mul2, LowMask);
18063 // If we have a signed multiply but no PMULDQ fix up the high parts of a
18064 // unsigned multiply.
18065 if (IsSigned && !Subtarget->hasSSE41()) {
18066 SDValue ShAmt = DAG.getConstant(
18068 DAG.getTargetLoweringInfo().getShiftAmountTy(VT, DAG.getDataLayout()));
18069 SDValue T1 = DAG.getNode(ISD::AND, dl, VT,
18070 DAG.getNode(ISD::SRA, dl, VT, Op0, ShAmt), Op1);
18071 SDValue T2 = DAG.getNode(ISD::AND, dl, VT,
18072 DAG.getNode(ISD::SRA, dl, VT, Op1, ShAmt), Op0);
18074 SDValue Fixup = DAG.getNode(ISD::ADD, dl, VT, T1, T2);
18075 Highs = DAG.getNode(ISD::SUB, dl, VT, Highs, Fixup);
18078 // The first result of MUL_LOHI is actually the low value, followed by the
18080 SDValue Ops[] = {Lows, Highs};
18081 return DAG.getMergeValues(Ops, dl);
18084 // Return true if the required (according to Opcode) shift-imm form is natively
18085 // supported by the Subtarget
18086 static bool SupportedVectorShiftWithImm(MVT VT, const X86Subtarget *Subtarget,
18088 if (VT.getScalarSizeInBits() < 16)
18091 if (VT.is512BitVector() &&
18092 (VT.getScalarSizeInBits() > 16 || Subtarget->hasBWI()))
18095 bool LShift = VT.is128BitVector() ||
18096 (VT.is256BitVector() && Subtarget->hasInt256());
18098 bool AShift = LShift && (Subtarget->hasVLX() ||
18099 (VT != MVT::v2i64 && VT != MVT::v4i64));
18100 return (Opcode == ISD::SRA) ? AShift : LShift;
18103 // The shift amount is a variable, but it is the same for all vector lanes.
18104 // These instructions are defined together with shift-immediate.
18106 bool SupportedVectorShiftWithBaseAmnt(MVT VT, const X86Subtarget *Subtarget,
18108 return SupportedVectorShiftWithImm(VT, Subtarget, Opcode);
18111 // Return true if the required (according to Opcode) variable-shift form is
18112 // natively supported by the Subtarget
18113 static bool SupportedVectorVarShift(MVT VT, const X86Subtarget *Subtarget,
18116 if (!Subtarget->hasInt256() || VT.getScalarSizeInBits() < 16)
18119 // vXi16 supported only on AVX-512, BWI
18120 if (VT.getScalarSizeInBits() == 16 && !Subtarget->hasBWI())
18123 if (VT.is512BitVector() || Subtarget->hasVLX())
18126 bool LShift = VT.is128BitVector() || VT.is256BitVector();
18127 bool AShift = LShift && VT != MVT::v2i64 && VT != MVT::v4i64;
18128 return (Opcode == ISD::SRA) ? AShift : LShift;
18131 static SDValue LowerScalarImmediateShift(SDValue Op, SelectionDAG &DAG,
18132 const X86Subtarget *Subtarget) {
18133 MVT VT = Op.getSimpleValueType();
18135 SDValue R = Op.getOperand(0);
18136 SDValue Amt = Op.getOperand(1);
18138 unsigned X86Opc = (Op.getOpcode() == ISD::SHL) ? X86ISD::VSHLI :
18139 (Op.getOpcode() == ISD::SRL) ? X86ISD::VSRLI : X86ISD::VSRAI;
18141 auto ArithmeticShiftRight64 = [&](uint64_t ShiftAmt) {
18142 assert((VT == MVT::v2i64 || VT == MVT::v4i64) && "Unexpected SRA type");
18143 MVT ExVT = MVT::getVectorVT(MVT::i32, VT.getVectorNumElements() * 2);
18144 SDValue Ex = DAG.getBitcast(ExVT, R);
18146 if (ShiftAmt >= 32) {
18147 // Splat sign to upper i32 dst, and SRA upper i32 src to lower i32.
18149 getTargetVShiftByConstNode(X86ISD::VSRAI, dl, ExVT, Ex, 31, DAG);
18150 SDValue Lower = getTargetVShiftByConstNode(X86ISD::VSRAI, dl, ExVT, Ex,
18151 ShiftAmt - 32, DAG);
18152 if (VT == MVT::v2i64)
18153 Ex = DAG.getVectorShuffle(ExVT, dl, Upper, Lower, {5, 1, 7, 3});
18154 if (VT == MVT::v4i64)
18155 Ex = DAG.getVectorShuffle(ExVT, dl, Upper, Lower,
18156 {9, 1, 11, 3, 13, 5, 15, 7});
18158 // SRA upper i32, SHL whole i64 and select lower i32.
18159 SDValue Upper = getTargetVShiftByConstNode(X86ISD::VSRAI, dl, ExVT, Ex,
18162 getTargetVShiftByConstNode(X86ISD::VSRLI, dl, VT, R, ShiftAmt, DAG);
18163 Lower = DAG.getBitcast(ExVT, Lower);
18164 if (VT == MVT::v2i64)
18165 Ex = DAG.getVectorShuffle(ExVT, dl, Upper, Lower, {4, 1, 6, 3});
18166 if (VT == MVT::v4i64)
18167 Ex = DAG.getVectorShuffle(ExVT, dl, Upper, Lower,
18168 {8, 1, 10, 3, 12, 5, 14, 7});
18170 return DAG.getBitcast(VT, Ex);
18173 // Optimize shl/srl/sra with constant shift amount.
18174 if (auto *BVAmt = dyn_cast<BuildVectorSDNode>(Amt)) {
18175 if (auto *ShiftConst = BVAmt->getConstantSplatNode()) {
18176 uint64_t ShiftAmt = ShiftConst->getZExtValue();
18178 if (SupportedVectorShiftWithImm(VT, Subtarget, Op.getOpcode()))
18179 return getTargetVShiftByConstNode(X86Opc, dl, VT, R, ShiftAmt, DAG);
18181 // i64 SRA needs to be performed as partial shifts.
18182 if ((VT == MVT::v2i64 || (Subtarget->hasInt256() && VT == MVT::v4i64)) &&
18183 Op.getOpcode() == ISD::SRA && !Subtarget->hasXOP())
18184 return ArithmeticShiftRight64(ShiftAmt);
18186 if (VT == MVT::v16i8 || (Subtarget->hasInt256() && VT == MVT::v32i8)) {
18187 unsigned NumElts = VT.getVectorNumElements();
18188 MVT ShiftVT = MVT::getVectorVT(MVT::i16, NumElts / 2);
18190 // Simple i8 add case
18191 if (Op.getOpcode() == ISD::SHL && ShiftAmt == 1)
18192 return DAG.getNode(ISD::ADD, dl, VT, R, R);
18194 // ashr(R, 7) === cmp_slt(R, 0)
18195 if (Op.getOpcode() == ISD::SRA && ShiftAmt == 7) {
18196 SDValue Zeros = getZeroVector(VT, Subtarget, DAG, dl);
18197 return DAG.getNode(X86ISD::PCMPGT, dl, VT, Zeros, R);
18200 // XOP can shift v16i8 directly instead of as shift v8i16 + mask.
18201 if (VT == MVT::v16i8 && Subtarget->hasXOP())
18204 if (Op.getOpcode() == ISD::SHL) {
18205 // Make a large shift.
18206 SDValue SHL = getTargetVShiftByConstNode(X86ISD::VSHLI, dl, ShiftVT,
18208 SHL = DAG.getBitcast(VT, SHL);
18209 // Zero out the rightmost bits.
18210 SmallVector<SDValue, 32> V(
18211 NumElts, DAG.getConstant(uint8_t(-1U << ShiftAmt), dl, MVT::i8));
18212 return DAG.getNode(ISD::AND, dl, VT, SHL,
18213 DAG.getNode(ISD::BUILD_VECTOR, dl, VT, V));
18215 if (Op.getOpcode() == ISD::SRL) {
18216 // Make a large shift.
18217 SDValue SRL = getTargetVShiftByConstNode(X86ISD::VSRLI, dl, ShiftVT,
18219 SRL = DAG.getBitcast(VT, SRL);
18220 // Zero out the leftmost bits.
18221 SmallVector<SDValue, 32> V(
18222 NumElts, DAG.getConstant(uint8_t(-1U) >> ShiftAmt, dl, MVT::i8));
18223 return DAG.getNode(ISD::AND, dl, VT, SRL,
18224 DAG.getNode(ISD::BUILD_VECTOR, dl, VT, V));
18226 if (Op.getOpcode() == ISD::SRA) {
18227 // ashr(R, Amt) === sub(xor(lshr(R, Amt), Mask), Mask)
18228 SDValue Res = DAG.getNode(ISD::SRL, dl, VT, R, Amt);
18229 SmallVector<SDValue, 32> V(NumElts,
18230 DAG.getConstant(128 >> ShiftAmt, dl,
18232 SDValue Mask = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, V);
18233 Res = DAG.getNode(ISD::XOR, dl, VT, Res, Mask);
18234 Res = DAG.getNode(ISD::SUB, dl, VT, Res, Mask);
18237 llvm_unreachable("Unknown shift opcode.");
18242 // Special case in 32-bit mode, where i64 is expanded into high and low parts.
18243 if (!Subtarget->is64Bit() && !Subtarget->hasXOP() &&
18244 (VT == MVT::v2i64 || (Subtarget->hasInt256() && VT == MVT::v4i64))) {
18246 // Peek through any splat that was introduced for i64 shift vectorization.
18247 int SplatIndex = -1;
18248 if (ShuffleVectorSDNode *SVN = dyn_cast<ShuffleVectorSDNode>(Amt.getNode()))
18249 if (SVN->isSplat()) {
18250 SplatIndex = SVN->getSplatIndex();
18251 Amt = Amt.getOperand(0);
18252 assert(SplatIndex < (int)VT.getVectorNumElements() &&
18253 "Splat shuffle referencing second operand");
18256 if (Amt.getOpcode() != ISD::BITCAST ||
18257 Amt.getOperand(0).getOpcode() != ISD::BUILD_VECTOR)
18260 Amt = Amt.getOperand(0);
18261 unsigned Ratio = Amt.getSimpleValueType().getVectorNumElements() /
18262 VT.getVectorNumElements();
18263 unsigned RatioInLog2 = Log2_32_Ceil(Ratio);
18264 uint64_t ShiftAmt = 0;
18265 unsigned BaseOp = (SplatIndex < 0 ? 0 : SplatIndex * Ratio);
18266 for (unsigned i = 0; i != Ratio; ++i) {
18267 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Amt.getOperand(i + BaseOp));
18271 ShiftAmt |= C->getZExtValue() << (i * (1 << (6 - RatioInLog2)));
18274 // Check remaining shift amounts (if not a splat).
18275 if (SplatIndex < 0) {
18276 for (unsigned i = Ratio; i != Amt.getNumOperands(); i += Ratio) {
18277 uint64_t ShAmt = 0;
18278 for (unsigned j = 0; j != Ratio; ++j) {
18279 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Amt.getOperand(i + j));
18283 ShAmt |= C->getZExtValue() << (j * (1 << (6 - RatioInLog2)));
18285 if (ShAmt != ShiftAmt)
18290 if (SupportedVectorShiftWithImm(VT, Subtarget, Op.getOpcode()))
18291 return getTargetVShiftByConstNode(X86Opc, dl, VT, R, ShiftAmt, DAG);
18293 if (Op.getOpcode() == ISD::SRA)
18294 return ArithmeticShiftRight64(ShiftAmt);
18300 static SDValue LowerScalarVariableShift(SDValue Op, SelectionDAG &DAG,
18301 const X86Subtarget* Subtarget) {
18302 MVT VT = Op.getSimpleValueType();
18304 SDValue R = Op.getOperand(0);
18305 SDValue Amt = Op.getOperand(1);
18307 unsigned X86OpcI = (Op.getOpcode() == ISD::SHL) ? X86ISD::VSHLI :
18308 (Op.getOpcode() == ISD::SRL) ? X86ISD::VSRLI : X86ISD::VSRAI;
18310 unsigned X86OpcV = (Op.getOpcode() == ISD::SHL) ? X86ISD::VSHL :
18311 (Op.getOpcode() == ISD::SRL) ? X86ISD::VSRL : X86ISD::VSRA;
18313 if (SupportedVectorShiftWithBaseAmnt(VT, Subtarget, Op.getOpcode())) {
18315 EVT EltVT = VT.getVectorElementType();
18317 if (BuildVectorSDNode *BV = dyn_cast<BuildVectorSDNode>(Amt)) {
18318 // Check if this build_vector node is doing a splat.
18319 // If so, then set BaseShAmt equal to the splat value.
18320 BaseShAmt = BV->getSplatValue();
18321 if (BaseShAmt && BaseShAmt.getOpcode() == ISD::UNDEF)
18322 BaseShAmt = SDValue();
18324 if (Amt.getOpcode() == ISD::EXTRACT_SUBVECTOR)
18325 Amt = Amt.getOperand(0);
18327 ShuffleVectorSDNode *SVN = dyn_cast<ShuffleVectorSDNode>(Amt);
18328 if (SVN && SVN->isSplat()) {
18329 unsigned SplatIdx = (unsigned)SVN->getSplatIndex();
18330 SDValue InVec = Amt.getOperand(0);
18331 if (InVec.getOpcode() == ISD::BUILD_VECTOR) {
18332 assert((SplatIdx < InVec.getValueType().getVectorNumElements()) &&
18333 "Unexpected shuffle index found!");
18334 BaseShAmt = InVec.getOperand(SplatIdx);
18335 } else if (InVec.getOpcode() == ISD::INSERT_VECTOR_ELT) {
18336 if (ConstantSDNode *C =
18337 dyn_cast<ConstantSDNode>(InVec.getOperand(2))) {
18338 if (C->getZExtValue() == SplatIdx)
18339 BaseShAmt = InVec.getOperand(1);
18344 // Avoid introducing an extract element from a shuffle.
18345 BaseShAmt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, EltVT, InVec,
18346 DAG.getIntPtrConstant(SplatIdx, dl));
18350 if (BaseShAmt.getNode()) {
18351 assert(EltVT.bitsLE(MVT::i64) && "Unexpected element type!");
18352 if (EltVT != MVT::i64 && EltVT.bitsGT(MVT::i32))
18353 BaseShAmt = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i64, BaseShAmt);
18354 else if (EltVT.bitsLT(MVT::i32))
18355 BaseShAmt = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, BaseShAmt);
18357 return getTargetVShiftNode(X86OpcI, dl, VT, R, BaseShAmt, DAG);
18361 // Special case in 32-bit mode, where i64 is expanded into high and low parts.
18362 if (!Subtarget->is64Bit() && VT == MVT::v2i64 &&
18363 Amt.getOpcode() == ISD::BITCAST &&
18364 Amt.getOperand(0).getOpcode() == ISD::BUILD_VECTOR) {
18365 Amt = Amt.getOperand(0);
18366 unsigned Ratio = Amt.getSimpleValueType().getVectorNumElements() /
18367 VT.getVectorNumElements();
18368 std::vector<SDValue> Vals(Ratio);
18369 for (unsigned i = 0; i != Ratio; ++i)
18370 Vals[i] = Amt.getOperand(i);
18371 for (unsigned i = Ratio; i != Amt.getNumOperands(); i += Ratio) {
18372 for (unsigned j = 0; j != Ratio; ++j)
18373 if (Vals[j] != Amt.getOperand(i + j))
18377 if (SupportedVectorShiftWithBaseAmnt(VT, Subtarget, Op.getOpcode()))
18378 return DAG.getNode(X86OpcV, dl, VT, R, Op.getOperand(1));
18383 static SDValue LowerShift(SDValue Op, const X86Subtarget* Subtarget,
18384 SelectionDAG &DAG) {
18385 MVT VT = Op.getSimpleValueType();
18387 SDValue R = Op.getOperand(0);
18388 SDValue Amt = Op.getOperand(1);
18390 assert(VT.isVector() && "Custom lowering only for vector shifts!");
18391 assert(Subtarget->hasSSE2() && "Only custom lower when we have SSE2!");
18393 if (SDValue V = LowerScalarImmediateShift(Op, DAG, Subtarget))
18396 if (SDValue V = LowerScalarVariableShift(Op, DAG, Subtarget))
18399 if (SupportedVectorVarShift(VT, Subtarget, Op.getOpcode()))
18402 // XOP has 128-bit variable logical/arithmetic shifts.
18403 // +ve/-ve Amt = shift left/right.
18404 if (Subtarget->hasXOP() &&
18405 (VT == MVT::v2i64 || VT == MVT::v4i32 ||
18406 VT == MVT::v8i16 || VT == MVT::v16i8)) {
18407 if (Op.getOpcode() == ISD::SRL || Op.getOpcode() == ISD::SRA) {
18408 SDValue Zero = getZeroVector(VT, Subtarget, DAG, dl);
18409 Amt = DAG.getNode(ISD::SUB, dl, VT, Zero, Amt);
18411 if (Op.getOpcode() == ISD::SHL || Op.getOpcode() == ISD::SRL)
18412 return DAG.getNode(X86ISD::VPSHL, dl, VT, R, Amt);
18413 if (Op.getOpcode() == ISD::SRA)
18414 return DAG.getNode(X86ISD::VPSHA, dl, VT, R, Amt);
18417 // 2i64 vector logical shifts can efficiently avoid scalarization - do the
18418 // shifts per-lane and then shuffle the partial results back together.
18419 if (VT == MVT::v2i64 && Op.getOpcode() != ISD::SRA) {
18420 // Splat the shift amounts so the scalar shifts above will catch it.
18421 SDValue Amt0 = DAG.getVectorShuffle(VT, dl, Amt, Amt, {0, 0});
18422 SDValue Amt1 = DAG.getVectorShuffle(VT, dl, Amt, Amt, {1, 1});
18423 SDValue R0 = DAG.getNode(Op->getOpcode(), dl, VT, R, Amt0);
18424 SDValue R1 = DAG.getNode(Op->getOpcode(), dl, VT, R, Amt1);
18425 return DAG.getVectorShuffle(VT, dl, R0, R1, {0, 3});
18428 // i64 vector arithmetic shift can be emulated with the transform:
18429 // M = lshr(SIGN_BIT, Amt)
18430 // ashr(R, Amt) === sub(xor(lshr(R, Amt), M), M)
18431 if ((VT == MVT::v2i64 || (VT == MVT::v4i64 && Subtarget->hasInt256())) &&
18432 Op.getOpcode() == ISD::SRA) {
18433 SDValue S = DAG.getConstant(APInt::getSignBit(64), dl, VT);
18434 SDValue M = DAG.getNode(ISD::SRL, dl, VT, S, Amt);
18435 R = DAG.getNode(ISD::SRL, dl, VT, R, Amt);
18436 R = DAG.getNode(ISD::XOR, dl, VT, R, M);
18437 R = DAG.getNode(ISD::SUB, dl, VT, R, M);
18441 // If possible, lower this packed shift into a vector multiply instead of
18442 // expanding it into a sequence of scalar shifts.
18443 // Do this only if the vector shift count is a constant build_vector.
18444 if (Op.getOpcode() == ISD::SHL &&
18445 (VT == MVT::v8i16 || VT == MVT::v4i32 ||
18446 (Subtarget->hasInt256() && VT == MVT::v16i16)) &&
18447 ISD::isBuildVectorOfConstantSDNodes(Amt.getNode())) {
18448 SmallVector<SDValue, 8> Elts;
18449 EVT SVT = VT.getScalarType();
18450 unsigned SVTBits = SVT.getSizeInBits();
18451 const APInt &One = APInt(SVTBits, 1);
18452 unsigned NumElems = VT.getVectorNumElements();
18454 for (unsigned i=0; i !=NumElems; ++i) {
18455 SDValue Op = Amt->getOperand(i);
18456 if (Op->getOpcode() == ISD::UNDEF) {
18457 Elts.push_back(Op);
18461 ConstantSDNode *ND = cast<ConstantSDNode>(Op);
18462 const APInt &C = APInt(SVTBits, ND->getAPIntValue().getZExtValue());
18463 uint64_t ShAmt = C.getZExtValue();
18464 if (ShAmt >= SVTBits) {
18465 Elts.push_back(DAG.getUNDEF(SVT));
18468 Elts.push_back(DAG.getConstant(One.shl(ShAmt), dl, SVT));
18470 SDValue BV = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Elts);
18471 return DAG.getNode(ISD::MUL, dl, VT, R, BV);
18474 // Lower SHL with variable shift amount.
18475 if (VT == MVT::v4i32 && Op->getOpcode() == ISD::SHL) {
18476 Op = DAG.getNode(ISD::SHL, dl, VT, Amt, DAG.getConstant(23, dl, VT));
18478 Op = DAG.getNode(ISD::ADD, dl, VT, Op,
18479 DAG.getConstant(0x3f800000U, dl, VT));
18480 Op = DAG.getBitcast(MVT::v4f32, Op);
18481 Op = DAG.getNode(ISD::FP_TO_SINT, dl, VT, Op);
18482 return DAG.getNode(ISD::MUL, dl, VT, Op, R);
18485 // If possible, lower this shift as a sequence of two shifts by
18486 // constant plus a MOVSS/MOVSD instead of scalarizing it.
18488 // (v4i32 (srl A, (build_vector < X, Y, Y, Y>)))
18490 // Could be rewritten as:
18491 // (v4i32 (MOVSS (srl A, <Y,Y,Y,Y>), (srl A, <X,X,X,X>)))
18493 // The advantage is that the two shifts from the example would be
18494 // lowered as X86ISD::VSRLI nodes. This would be cheaper than scalarizing
18495 // the vector shift into four scalar shifts plus four pairs of vector
18497 if ((VT == MVT::v8i16 || VT == MVT::v4i32) &&
18498 ISD::isBuildVectorOfConstantSDNodes(Amt.getNode())) {
18499 unsigned TargetOpcode = X86ISD::MOVSS;
18500 bool CanBeSimplified;
18501 // The splat value for the first packed shift (the 'X' from the example).
18502 SDValue Amt1 = Amt->getOperand(0);
18503 // The splat value for the second packed shift (the 'Y' from the example).
18504 SDValue Amt2 = (VT == MVT::v4i32) ? Amt->getOperand(1) :
18505 Amt->getOperand(2);
18507 // See if it is possible to replace this node with a sequence of
18508 // two shifts followed by a MOVSS/MOVSD
18509 if (VT == MVT::v4i32) {
18510 // Check if it is legal to use a MOVSS.
18511 CanBeSimplified = Amt2 == Amt->getOperand(2) &&
18512 Amt2 == Amt->getOperand(3);
18513 if (!CanBeSimplified) {
18514 // Otherwise, check if we can still simplify this node using a MOVSD.
18515 CanBeSimplified = Amt1 == Amt->getOperand(1) &&
18516 Amt->getOperand(2) == Amt->getOperand(3);
18517 TargetOpcode = X86ISD::MOVSD;
18518 Amt2 = Amt->getOperand(2);
18521 // Do similar checks for the case where the machine value type
18523 CanBeSimplified = Amt1 == Amt->getOperand(1);
18524 for (unsigned i=3; i != 8 && CanBeSimplified; ++i)
18525 CanBeSimplified = Amt2 == Amt->getOperand(i);
18527 if (!CanBeSimplified) {
18528 TargetOpcode = X86ISD::MOVSD;
18529 CanBeSimplified = true;
18530 Amt2 = Amt->getOperand(4);
18531 for (unsigned i=0; i != 4 && CanBeSimplified; ++i)
18532 CanBeSimplified = Amt1 == Amt->getOperand(i);
18533 for (unsigned j=4; j != 8 && CanBeSimplified; ++j)
18534 CanBeSimplified = Amt2 == Amt->getOperand(j);
18538 if (CanBeSimplified && isa<ConstantSDNode>(Amt1) &&
18539 isa<ConstantSDNode>(Amt2)) {
18540 // Replace this node with two shifts followed by a MOVSS/MOVSD.
18541 EVT CastVT = MVT::v4i32;
18543 DAG.getConstant(cast<ConstantSDNode>(Amt1)->getAPIntValue(), dl, VT);
18544 SDValue Shift1 = DAG.getNode(Op->getOpcode(), dl, VT, R, Splat1);
18546 DAG.getConstant(cast<ConstantSDNode>(Amt2)->getAPIntValue(), dl, VT);
18547 SDValue Shift2 = DAG.getNode(Op->getOpcode(), dl, VT, R, Splat2);
18548 if (TargetOpcode == X86ISD::MOVSD)
18549 CastVT = MVT::v2i64;
18550 SDValue BitCast1 = DAG.getBitcast(CastVT, Shift1);
18551 SDValue BitCast2 = DAG.getBitcast(CastVT, Shift2);
18552 SDValue Result = getTargetShuffleNode(TargetOpcode, dl, CastVT, BitCast2,
18554 return DAG.getBitcast(VT, Result);
18558 // v4i32 Non Uniform Shifts.
18559 // If the shift amount is constant we can shift each lane using the SSE2
18560 // immediate shifts, else we need to zero-extend each lane to the lower i64
18561 // and shift using the SSE2 variable shifts.
18562 // The separate results can then be blended together.
18563 if (VT == MVT::v4i32) {
18564 unsigned Opc = Op.getOpcode();
18565 SDValue Amt0, Amt1, Amt2, Amt3;
18566 if (ISD::isBuildVectorOfConstantSDNodes(Amt.getNode())) {
18567 Amt0 = DAG.getVectorShuffle(VT, dl, Amt, DAG.getUNDEF(VT), {0, 0, 0, 0});
18568 Amt1 = DAG.getVectorShuffle(VT, dl, Amt, DAG.getUNDEF(VT), {1, 1, 1, 1});
18569 Amt2 = DAG.getVectorShuffle(VT, dl, Amt, DAG.getUNDEF(VT), {2, 2, 2, 2});
18570 Amt3 = DAG.getVectorShuffle(VT, dl, Amt, DAG.getUNDEF(VT), {3, 3, 3, 3});
18572 // ISD::SHL is handled above but we include it here for completeness.
18575 llvm_unreachable("Unknown target vector shift node");
18577 Opc = X86ISD::VSHL;
18580 Opc = X86ISD::VSRL;
18583 Opc = X86ISD::VSRA;
18586 // The SSE2 shifts use the lower i64 as the same shift amount for
18587 // all lanes and the upper i64 is ignored. These shuffle masks
18588 // optimally zero-extend each lanes on SSE2/SSE41/AVX targets.
18589 SDValue Z = getZeroVector(VT, Subtarget, DAG, dl);
18590 Amt0 = DAG.getVectorShuffle(VT, dl, Amt, Z, {0, 4, -1, -1});
18591 Amt1 = DAG.getVectorShuffle(VT, dl, Amt, Z, {1, 5, -1, -1});
18592 Amt2 = DAG.getVectorShuffle(VT, dl, Amt, Z, {2, 6, -1, -1});
18593 Amt3 = DAG.getVectorShuffle(VT, dl, Amt, Z, {3, 7, -1, -1});
18596 SDValue R0 = DAG.getNode(Opc, dl, VT, R, Amt0);
18597 SDValue R1 = DAG.getNode(Opc, dl, VT, R, Amt1);
18598 SDValue R2 = DAG.getNode(Opc, dl, VT, R, Amt2);
18599 SDValue R3 = DAG.getNode(Opc, dl, VT, R, Amt3);
18600 SDValue R02 = DAG.getVectorShuffle(VT, dl, R0, R2, {0, -1, 6, -1});
18601 SDValue R13 = DAG.getVectorShuffle(VT, dl, R1, R3, {-1, 1, -1, 7});
18602 return DAG.getVectorShuffle(VT, dl, R02, R13, {0, 5, 2, 7});
18605 if (VT == MVT::v16i8 ||
18606 (VT == MVT::v32i8 && Subtarget->hasInt256() && !Subtarget->hasXOP())) {
18607 MVT ExtVT = MVT::getVectorVT(MVT::i16, VT.getVectorNumElements() / 2);
18608 unsigned ShiftOpcode = Op->getOpcode();
18610 auto SignBitSelect = [&](MVT SelVT, SDValue Sel, SDValue V0, SDValue V1) {
18611 // On SSE41 targets we make use of the fact that VSELECT lowers
18612 // to PBLENDVB which selects bytes based just on the sign bit.
18613 if (Subtarget->hasSSE41()) {
18614 V0 = DAG.getBitcast(VT, V0);
18615 V1 = DAG.getBitcast(VT, V1);
18616 Sel = DAG.getBitcast(VT, Sel);
18617 return DAG.getBitcast(SelVT,
18618 DAG.getNode(ISD::VSELECT, dl, VT, Sel, V0, V1));
18620 // On pre-SSE41 targets we test for the sign bit by comparing to
18621 // zero - a negative value will set all bits of the lanes to true
18622 // and VSELECT uses that in its OR(AND(V0,C),AND(V1,~C)) lowering.
18623 SDValue Z = getZeroVector(SelVT, Subtarget, DAG, dl);
18624 SDValue C = DAG.getNode(X86ISD::PCMPGT, dl, SelVT, Z, Sel);
18625 return DAG.getNode(ISD::VSELECT, dl, SelVT, C, V0, V1);
18628 // Turn 'a' into a mask suitable for VSELECT: a = a << 5;
18629 // We can safely do this using i16 shifts as we're only interested in
18630 // the 3 lower bits of each byte.
18631 Amt = DAG.getBitcast(ExtVT, Amt);
18632 Amt = DAG.getNode(ISD::SHL, dl, ExtVT, Amt, DAG.getConstant(5, dl, ExtVT));
18633 Amt = DAG.getBitcast(VT, Amt);
18635 if (Op->getOpcode() == ISD::SHL || Op->getOpcode() == ISD::SRL) {
18636 // r = VSELECT(r, shift(r, 4), a);
18638 DAG.getNode(ShiftOpcode, dl, VT, R, DAG.getConstant(4, dl, VT));
18639 R = SignBitSelect(VT, Amt, M, R);
18642 Amt = DAG.getNode(ISD::ADD, dl, VT, Amt, Amt);
18644 // r = VSELECT(r, shift(r, 2), a);
18645 M = DAG.getNode(ShiftOpcode, dl, VT, R, DAG.getConstant(2, dl, VT));
18646 R = SignBitSelect(VT, Amt, M, R);
18649 Amt = DAG.getNode(ISD::ADD, dl, VT, Amt, Amt);
18651 // return VSELECT(r, shift(r, 1), a);
18652 M = DAG.getNode(ShiftOpcode, dl, VT, R, DAG.getConstant(1, dl, VT));
18653 R = SignBitSelect(VT, Amt, M, R);
18657 if (Op->getOpcode() == ISD::SRA) {
18658 // For SRA we need to unpack each byte to the higher byte of a i16 vector
18659 // so we can correctly sign extend. We don't care what happens to the
18661 SDValue ALo = DAG.getNode(X86ISD::UNPCKL, dl, VT, DAG.getUNDEF(VT), Amt);
18662 SDValue AHi = DAG.getNode(X86ISD::UNPCKH, dl, VT, DAG.getUNDEF(VT), Amt);
18663 SDValue RLo = DAG.getNode(X86ISD::UNPCKL, dl, VT, DAG.getUNDEF(VT), R);
18664 SDValue RHi = DAG.getNode(X86ISD::UNPCKH, dl, VT, DAG.getUNDEF(VT), R);
18665 ALo = DAG.getBitcast(ExtVT, ALo);
18666 AHi = DAG.getBitcast(ExtVT, AHi);
18667 RLo = DAG.getBitcast(ExtVT, RLo);
18668 RHi = DAG.getBitcast(ExtVT, RHi);
18670 // r = VSELECT(r, shift(r, 4), a);
18671 SDValue MLo = DAG.getNode(ShiftOpcode, dl, ExtVT, RLo,
18672 DAG.getConstant(4, dl, ExtVT));
18673 SDValue MHi = DAG.getNode(ShiftOpcode, dl, ExtVT, RHi,
18674 DAG.getConstant(4, dl, ExtVT));
18675 RLo = SignBitSelect(ExtVT, ALo, MLo, RLo);
18676 RHi = SignBitSelect(ExtVT, AHi, MHi, RHi);
18679 ALo = DAG.getNode(ISD::ADD, dl, ExtVT, ALo, ALo);
18680 AHi = DAG.getNode(ISD::ADD, dl, ExtVT, AHi, AHi);
18682 // r = VSELECT(r, shift(r, 2), a);
18683 MLo = DAG.getNode(ShiftOpcode, dl, ExtVT, RLo,
18684 DAG.getConstant(2, dl, ExtVT));
18685 MHi = DAG.getNode(ShiftOpcode, dl, ExtVT, RHi,
18686 DAG.getConstant(2, dl, ExtVT));
18687 RLo = SignBitSelect(ExtVT, ALo, MLo, RLo);
18688 RHi = SignBitSelect(ExtVT, AHi, MHi, RHi);
18691 ALo = DAG.getNode(ISD::ADD, dl, ExtVT, ALo, ALo);
18692 AHi = DAG.getNode(ISD::ADD, dl, ExtVT, AHi, AHi);
18694 // r = VSELECT(r, shift(r, 1), a);
18695 MLo = DAG.getNode(ShiftOpcode, dl, ExtVT, RLo,
18696 DAG.getConstant(1, dl, ExtVT));
18697 MHi = DAG.getNode(ShiftOpcode, dl, ExtVT, RHi,
18698 DAG.getConstant(1, dl, ExtVT));
18699 RLo = SignBitSelect(ExtVT, ALo, MLo, RLo);
18700 RHi = SignBitSelect(ExtVT, AHi, MHi, RHi);
18702 // Logical shift the result back to the lower byte, leaving a zero upper
18704 // meaning that we can safely pack with PACKUSWB.
18706 DAG.getNode(ISD::SRL, dl, ExtVT, RLo, DAG.getConstant(8, dl, ExtVT));
18708 DAG.getNode(ISD::SRL, dl, ExtVT, RHi, DAG.getConstant(8, dl, ExtVT));
18709 return DAG.getNode(X86ISD::PACKUS, dl, VT, RLo, RHi);
18713 // It's worth extending once and using the v8i32 shifts for 16-bit types, but
18714 // the extra overheads to get from v16i8 to v8i32 make the existing SSE
18715 // solution better.
18716 if (Subtarget->hasInt256() && VT == MVT::v8i16) {
18717 MVT ExtVT = MVT::v8i32;
18719 Op.getOpcode() == ISD::SRA ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND;
18720 R = DAG.getNode(ExtOpc, dl, ExtVT, R);
18721 Amt = DAG.getNode(ISD::ANY_EXTEND, dl, ExtVT, Amt);
18722 return DAG.getNode(ISD::TRUNCATE, dl, VT,
18723 DAG.getNode(Op.getOpcode(), dl, ExtVT, R, Amt));
18726 if (Subtarget->hasInt256() && !Subtarget->hasXOP() && VT == MVT::v16i16) {
18727 MVT ExtVT = MVT::v8i32;
18728 SDValue Z = getZeroVector(VT, Subtarget, DAG, dl);
18729 SDValue ALo = DAG.getNode(X86ISD::UNPCKL, dl, VT, Amt, Z);
18730 SDValue AHi = DAG.getNode(X86ISD::UNPCKH, dl, VT, Amt, Z);
18731 SDValue RLo = DAG.getNode(X86ISD::UNPCKL, dl, VT, R, R);
18732 SDValue RHi = DAG.getNode(X86ISD::UNPCKH, dl, VT, R, R);
18733 ALo = DAG.getBitcast(ExtVT, ALo);
18734 AHi = DAG.getBitcast(ExtVT, AHi);
18735 RLo = DAG.getBitcast(ExtVT, RLo);
18736 RHi = DAG.getBitcast(ExtVT, RHi);
18737 SDValue Lo = DAG.getNode(Op.getOpcode(), dl, ExtVT, RLo, ALo);
18738 SDValue Hi = DAG.getNode(Op.getOpcode(), dl, ExtVT, RHi, AHi);
18739 Lo = DAG.getNode(ISD::SRL, dl, ExtVT, Lo, DAG.getConstant(16, dl, ExtVT));
18740 Hi = DAG.getNode(ISD::SRL, dl, ExtVT, Hi, DAG.getConstant(16, dl, ExtVT));
18741 return DAG.getNode(X86ISD::PACKUS, dl, VT, Lo, Hi);
18744 if (VT == MVT::v8i16) {
18745 unsigned ShiftOpcode = Op->getOpcode();
18747 auto SignBitSelect = [&](SDValue Sel, SDValue V0, SDValue V1) {
18748 // On SSE41 targets we make use of the fact that VSELECT lowers
18749 // to PBLENDVB which selects bytes based just on the sign bit.
18750 if (Subtarget->hasSSE41()) {
18751 MVT ExtVT = MVT::getVectorVT(MVT::i8, VT.getVectorNumElements() * 2);
18752 V0 = DAG.getBitcast(ExtVT, V0);
18753 V1 = DAG.getBitcast(ExtVT, V1);
18754 Sel = DAG.getBitcast(ExtVT, Sel);
18755 return DAG.getBitcast(
18756 VT, DAG.getNode(ISD::VSELECT, dl, ExtVT, Sel, V0, V1));
18758 // On pre-SSE41 targets we splat the sign bit - a negative value will
18759 // set all bits of the lanes to true and VSELECT uses that in
18760 // its OR(AND(V0,C),AND(V1,~C)) lowering.
18762 DAG.getNode(ISD::SRA, dl, VT, Sel, DAG.getConstant(15, dl, VT));
18763 return DAG.getNode(ISD::VSELECT, dl, VT, C, V0, V1);
18766 // Turn 'a' into a mask suitable for VSELECT: a = a << 12;
18767 if (Subtarget->hasSSE41()) {
18768 // On SSE41 targets we need to replicate the shift mask in both
18769 // bytes for PBLENDVB.
18772 DAG.getNode(ISD::SHL, dl, VT, Amt, DAG.getConstant(4, dl, VT)),
18773 DAG.getNode(ISD::SHL, dl, VT, Amt, DAG.getConstant(12, dl, VT)));
18775 Amt = DAG.getNode(ISD::SHL, dl, VT, Amt, DAG.getConstant(12, dl, VT));
18778 // r = VSELECT(r, shift(r, 8), a);
18779 SDValue M = DAG.getNode(ShiftOpcode, dl, VT, R, DAG.getConstant(8, dl, VT));
18780 R = SignBitSelect(Amt, M, R);
18783 Amt = DAG.getNode(ISD::ADD, dl, VT, Amt, Amt);
18785 // r = VSELECT(r, shift(r, 4), a);
18786 M = DAG.getNode(ShiftOpcode, dl, VT, R, DAG.getConstant(4, dl, VT));
18787 R = SignBitSelect(Amt, M, R);
18790 Amt = DAG.getNode(ISD::ADD, dl, VT, Amt, Amt);
18792 // r = VSELECT(r, shift(r, 2), a);
18793 M = DAG.getNode(ShiftOpcode, dl, VT, R, DAG.getConstant(2, dl, VT));
18794 R = SignBitSelect(Amt, M, R);
18797 Amt = DAG.getNode(ISD::ADD, dl, VT, Amt, Amt);
18799 // return VSELECT(r, shift(r, 1), a);
18800 M = DAG.getNode(ShiftOpcode, dl, VT, R, DAG.getConstant(1, dl, VT));
18801 R = SignBitSelect(Amt, M, R);
18805 // Decompose 256-bit shifts into smaller 128-bit shifts.
18806 if (VT.is256BitVector()) {
18807 unsigned NumElems = VT.getVectorNumElements();
18808 MVT EltVT = VT.getVectorElementType();
18809 EVT NewVT = MVT::getVectorVT(EltVT, NumElems/2);
18811 // Extract the two vectors
18812 SDValue V1 = Extract128BitVector(R, 0, DAG, dl);
18813 SDValue V2 = Extract128BitVector(R, NumElems/2, DAG, dl);
18815 // Recreate the shift amount vectors
18816 SDValue Amt1, Amt2;
18817 if (Amt.getOpcode() == ISD::BUILD_VECTOR) {
18818 // Constant shift amount
18819 SmallVector<SDValue, 8> Ops(Amt->op_begin(), Amt->op_begin() + NumElems);
18820 ArrayRef<SDValue> Amt1Csts = makeArrayRef(Ops).slice(0, NumElems / 2);
18821 ArrayRef<SDValue> Amt2Csts = makeArrayRef(Ops).slice(NumElems / 2);
18823 Amt1 = DAG.getNode(ISD::BUILD_VECTOR, dl, NewVT, Amt1Csts);
18824 Amt2 = DAG.getNode(ISD::BUILD_VECTOR, dl, NewVT, Amt2Csts);
18826 // Variable shift amount
18827 Amt1 = Extract128BitVector(Amt, 0, DAG, dl);
18828 Amt2 = Extract128BitVector(Amt, NumElems/2, DAG, dl);
18831 // Issue new vector shifts for the smaller types
18832 V1 = DAG.getNode(Op.getOpcode(), dl, NewVT, V1, Amt1);
18833 V2 = DAG.getNode(Op.getOpcode(), dl, NewVT, V2, Amt2);
18835 // Concatenate the result back
18836 return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT, V1, V2);
18842 static SDValue LowerRotate(SDValue Op, const X86Subtarget *Subtarget,
18843 SelectionDAG &DAG) {
18844 MVT VT = Op.getSimpleValueType();
18846 SDValue R = Op.getOperand(0);
18847 SDValue Amt = Op.getOperand(1);
18849 assert(VT.isVector() && "Custom lowering only for vector rotates!");
18850 assert(Subtarget->hasXOP() && "XOP support required for vector rotates!");
18851 assert((Op.getOpcode() == ISD::ROTL) && "Only ROTL supported");
18853 // XOP has 128-bit vector variable + immediate rotates.
18854 // +ve/-ve Amt = rotate left/right.
18856 // Split 256-bit integers.
18857 if (VT.getSizeInBits() == 256)
18858 return Lower256IntArith(Op, DAG);
18860 assert(VT.getSizeInBits() == 128 && "Only rotate 128-bit vectors!");
18862 // Attempt to rotate by immediate.
18863 if (auto *BVAmt = dyn_cast<BuildVectorSDNode>(Amt)) {
18864 if (auto *RotateConst = BVAmt->getConstantSplatNode()) {
18865 uint64_t RotateAmt = RotateConst->getAPIntValue().getZExtValue();
18866 assert(RotateAmt < VT.getScalarSizeInBits() && "Rotation out of range");
18867 return DAG.getNode(X86ISD::VPROTI, DL, VT, R,
18868 DAG.getConstant(RotateAmt, DL, MVT::i8));
18872 // Use general rotate by variable (per-element).
18873 return DAG.getNode(X86ISD::VPROT, DL, VT, R, Amt);
18876 static SDValue LowerXALUO(SDValue Op, SelectionDAG &DAG) {
18877 // Lower the "add/sub/mul with overflow" instruction into a regular ins plus
18878 // a "setcc" instruction that checks the overflow flag. The "brcond" lowering
18879 // looks for this combo and may remove the "setcc" instruction if the "setcc"
18880 // has only one use.
18881 SDNode *N = Op.getNode();
18882 SDValue LHS = N->getOperand(0);
18883 SDValue RHS = N->getOperand(1);
18884 unsigned BaseOp = 0;
18887 switch (Op.getOpcode()) {
18888 default: llvm_unreachable("Unknown ovf instruction!");
18890 // A subtract of one will be selected as a INC. Note that INC doesn't
18891 // set CF, so we can't do this for UADDO.
18892 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(RHS))
18894 BaseOp = X86ISD::INC;
18895 Cond = X86::COND_O;
18898 BaseOp = X86ISD::ADD;
18899 Cond = X86::COND_O;
18902 BaseOp = X86ISD::ADD;
18903 Cond = X86::COND_B;
18906 // A subtract of one will be selected as a DEC. Note that DEC doesn't
18907 // set CF, so we can't do this for USUBO.
18908 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(RHS))
18910 BaseOp = X86ISD::DEC;
18911 Cond = X86::COND_O;
18914 BaseOp = X86ISD::SUB;
18915 Cond = X86::COND_O;
18918 BaseOp = X86ISD::SUB;
18919 Cond = X86::COND_B;
18922 BaseOp = N->getValueType(0) == MVT::i8 ? X86ISD::SMUL8 : X86ISD::SMUL;
18923 Cond = X86::COND_O;
18925 case ISD::UMULO: { // i64, i8 = umulo lhs, rhs --> i64, i64, i32 umul lhs,rhs
18926 if (N->getValueType(0) == MVT::i8) {
18927 BaseOp = X86ISD::UMUL8;
18928 Cond = X86::COND_O;
18931 SDVTList VTs = DAG.getVTList(N->getValueType(0), N->getValueType(0),
18933 SDValue Sum = DAG.getNode(X86ISD::UMUL, DL, VTs, LHS, RHS);
18936 DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
18937 DAG.getConstant(X86::COND_O, DL, MVT::i32),
18938 SDValue(Sum.getNode(), 2));
18940 return DAG.getNode(ISD::MERGE_VALUES, DL, N->getVTList(), Sum, SetCC);
18944 // Also sets EFLAGS.
18945 SDVTList VTs = DAG.getVTList(N->getValueType(0), MVT::i32);
18946 SDValue Sum = DAG.getNode(BaseOp, DL, VTs, LHS, RHS);
18949 DAG.getNode(X86ISD::SETCC, DL, N->getValueType(1),
18950 DAG.getConstant(Cond, DL, MVT::i32),
18951 SDValue(Sum.getNode(), 1));
18953 return DAG.getNode(ISD::MERGE_VALUES, DL, N->getVTList(), Sum, SetCC);
18956 /// Returns true if the operand type is exactly twice the native width, and
18957 /// the corresponding cmpxchg8b or cmpxchg16b instruction is available.
18958 /// Used to know whether to use cmpxchg8/16b when expanding atomic operations
18959 /// (otherwise we leave them alone to become __sync_fetch_and_... calls).
18960 bool X86TargetLowering::needsCmpXchgNb(Type *MemType) const {
18961 unsigned OpWidth = MemType->getPrimitiveSizeInBits();
18964 return !Subtarget->is64Bit(); // FIXME this should be Subtarget.hasCmpxchg8b
18965 else if (OpWidth == 128)
18966 return Subtarget->hasCmpxchg16b();
18971 bool X86TargetLowering::shouldExpandAtomicStoreInIR(StoreInst *SI) const {
18972 return needsCmpXchgNb(SI->getValueOperand()->getType());
18975 // Note: this turns large loads into lock cmpxchg8b/16b.
18976 // FIXME: On 32 bits x86, fild/movq might be faster than lock cmpxchg8b.
18977 TargetLowering::AtomicExpansionKind
18978 X86TargetLowering::shouldExpandAtomicLoadInIR(LoadInst *LI) const {
18979 auto PTy = cast<PointerType>(LI->getPointerOperand()->getType());
18980 return needsCmpXchgNb(PTy->getElementType()) ? AtomicExpansionKind::CmpXChg
18981 : AtomicExpansionKind::None;
18984 TargetLowering::AtomicExpansionKind
18985 X86TargetLowering::shouldExpandAtomicRMWInIR(AtomicRMWInst *AI) const {
18986 unsigned NativeWidth = Subtarget->is64Bit() ? 64 : 32;
18987 Type *MemType = AI->getType();
18989 // If the operand is too big, we must see if cmpxchg8/16b is available
18990 // and default to library calls otherwise.
18991 if (MemType->getPrimitiveSizeInBits() > NativeWidth) {
18992 return needsCmpXchgNb(MemType) ? AtomicExpansionKind::CmpXChg
18993 : AtomicExpansionKind::None;
18996 AtomicRMWInst::BinOp Op = AI->getOperation();
18999 llvm_unreachable("Unknown atomic operation");
19000 case AtomicRMWInst::Xchg:
19001 case AtomicRMWInst::Add:
19002 case AtomicRMWInst::Sub:
19003 // It's better to use xadd, xsub or xchg for these in all cases.
19004 return AtomicExpansionKind::None;
19005 case AtomicRMWInst::Or:
19006 case AtomicRMWInst::And:
19007 case AtomicRMWInst::Xor:
19008 // If the atomicrmw's result isn't actually used, we can just add a "lock"
19009 // prefix to a normal instruction for these operations.
19010 return !AI->use_empty() ? AtomicExpansionKind::CmpXChg
19011 : AtomicExpansionKind::None;
19012 case AtomicRMWInst::Nand:
19013 case AtomicRMWInst::Max:
19014 case AtomicRMWInst::Min:
19015 case AtomicRMWInst::UMax:
19016 case AtomicRMWInst::UMin:
19017 // These always require a non-trivial set of data operations on x86. We must
19018 // use a cmpxchg loop.
19019 return AtomicExpansionKind::CmpXChg;
19023 static bool hasMFENCE(const X86Subtarget& Subtarget) {
19024 // Use mfence if we have SSE2 or we're on x86-64 (even if we asked for
19025 // no-sse2). There isn't any reason to disable it if the target processor
19027 return Subtarget.hasSSE2() || Subtarget.is64Bit();
19031 X86TargetLowering::lowerIdempotentRMWIntoFencedLoad(AtomicRMWInst *AI) const {
19032 unsigned NativeWidth = Subtarget->is64Bit() ? 64 : 32;
19033 Type *MemType = AI->getType();
19034 // Accesses larger than the native width are turned into cmpxchg/libcalls, so
19035 // there is no benefit in turning such RMWs into loads, and it is actually
19036 // harmful as it introduces a mfence.
19037 if (MemType->getPrimitiveSizeInBits() > NativeWidth)
19040 auto Builder = IRBuilder<>(AI);
19041 Module *M = Builder.GetInsertBlock()->getParent()->getParent();
19042 auto SynchScope = AI->getSynchScope();
19043 // We must restrict the ordering to avoid generating loads with Release or
19044 // ReleaseAcquire orderings.
19045 auto Order = AtomicCmpXchgInst::getStrongestFailureOrdering(AI->getOrdering());
19046 auto Ptr = AI->getPointerOperand();
19048 // Before the load we need a fence. Here is an example lifted from
19049 // http://www.hpl.hp.com/techreports/2012/HPL-2012-68.pdf showing why a fence
19052 // x.store(1, relaxed);
19053 // r1 = y.fetch_add(0, release);
19055 // y.fetch_add(42, acquire);
19056 // r2 = x.load(relaxed);
19057 // r1 = r2 = 0 is impossible, but becomes possible if the idempotent rmw is
19058 // lowered to just a load without a fence. A mfence flushes the store buffer,
19059 // making the optimization clearly correct.
19060 // FIXME: it is required if isAtLeastRelease(Order) but it is not clear
19061 // otherwise, we might be able to be more aggressive on relaxed idempotent
19062 // rmw. In practice, they do not look useful, so we don't try to be
19063 // especially clever.
19064 if (SynchScope == SingleThread)
19065 // FIXME: we could just insert an X86ISD::MEMBARRIER here, except we are at
19066 // the IR level, so we must wrap it in an intrinsic.
19069 if (!hasMFENCE(*Subtarget))
19070 // FIXME: it might make sense to use a locked operation here but on a
19071 // different cache-line to prevent cache-line bouncing. In practice it
19072 // is probably a small win, and x86 processors without mfence are rare
19073 // enough that we do not bother.
19077 llvm::Intrinsic::getDeclaration(M, Intrinsic::x86_sse2_mfence);
19078 Builder.CreateCall(MFence, {});
19080 // Finally we can emit the atomic load.
19081 LoadInst *Loaded = Builder.CreateAlignedLoad(Ptr,
19082 AI->getType()->getPrimitiveSizeInBits());
19083 Loaded->setAtomic(Order, SynchScope);
19084 AI->replaceAllUsesWith(Loaded);
19085 AI->eraseFromParent();
19089 static SDValue LowerATOMIC_FENCE(SDValue Op, const X86Subtarget *Subtarget,
19090 SelectionDAG &DAG) {
19092 AtomicOrdering FenceOrdering = static_cast<AtomicOrdering>(
19093 cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue());
19094 SynchronizationScope FenceScope = static_cast<SynchronizationScope>(
19095 cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue());
19097 // The only fence that needs an instruction is a sequentially-consistent
19098 // cross-thread fence.
19099 if (FenceOrdering == SequentiallyConsistent && FenceScope == CrossThread) {
19100 if (hasMFENCE(*Subtarget))
19101 return DAG.getNode(X86ISD::MFENCE, dl, MVT::Other, Op.getOperand(0));
19103 SDValue Chain = Op.getOperand(0);
19104 SDValue Zero = DAG.getConstant(0, dl, MVT::i32);
19106 DAG.getRegister(X86::ESP, MVT::i32), // Base
19107 DAG.getTargetConstant(1, dl, MVT::i8), // Scale
19108 DAG.getRegister(0, MVT::i32), // Index
19109 DAG.getTargetConstant(0, dl, MVT::i32), // Disp
19110 DAG.getRegister(0, MVT::i32), // Segment.
19114 SDNode *Res = DAG.getMachineNode(X86::OR32mrLocked, dl, MVT::Other, Ops);
19115 return SDValue(Res, 0);
19118 // MEMBARRIER is a compiler barrier; it codegens to a no-op.
19119 return DAG.getNode(X86ISD::MEMBARRIER, dl, MVT::Other, Op.getOperand(0));
19122 static SDValue LowerCMP_SWAP(SDValue Op, const X86Subtarget *Subtarget,
19123 SelectionDAG &DAG) {
19124 MVT T = Op.getSimpleValueType();
19128 switch(T.SimpleTy) {
19129 default: llvm_unreachable("Invalid value type!");
19130 case MVT::i8: Reg = X86::AL; size = 1; break;
19131 case MVT::i16: Reg = X86::AX; size = 2; break;
19132 case MVT::i32: Reg = X86::EAX; size = 4; break;
19134 assert(Subtarget->is64Bit() && "Node not type legal!");
19135 Reg = X86::RAX; size = 8;
19138 SDValue cpIn = DAG.getCopyToReg(Op.getOperand(0), DL, Reg,
19139 Op.getOperand(2), SDValue());
19140 SDValue Ops[] = { cpIn.getValue(0),
19143 DAG.getTargetConstant(size, DL, MVT::i8),
19144 cpIn.getValue(1) };
19145 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Glue);
19146 MachineMemOperand *MMO = cast<AtomicSDNode>(Op)->getMemOperand();
19147 SDValue Result = DAG.getMemIntrinsicNode(X86ISD::LCMPXCHG_DAG, DL, Tys,
19151 DAG.getCopyFromReg(Result.getValue(0), DL, Reg, T, Result.getValue(1));
19152 SDValue EFLAGS = DAG.getCopyFromReg(cpOut.getValue(1), DL, X86::EFLAGS,
19153 MVT::i32, cpOut.getValue(2));
19154 SDValue Success = DAG.getNode(X86ISD::SETCC, DL, Op->getValueType(1),
19155 DAG.getConstant(X86::COND_E, DL, MVT::i8),
19158 DAG.ReplaceAllUsesOfValueWith(Op.getValue(0), cpOut);
19159 DAG.ReplaceAllUsesOfValueWith(Op.getValue(1), Success);
19160 DAG.ReplaceAllUsesOfValueWith(Op.getValue(2), EFLAGS.getValue(1));
19164 static SDValue LowerBITCAST(SDValue Op, const X86Subtarget *Subtarget,
19165 SelectionDAG &DAG) {
19166 MVT SrcVT = Op.getOperand(0).getSimpleValueType();
19167 MVT DstVT = Op.getSimpleValueType();
19169 if (SrcVT == MVT::v2i32 || SrcVT == MVT::v4i16 || SrcVT == MVT::v8i8) {
19170 assert(Subtarget->hasSSE2() && "Requires at least SSE2!");
19171 if (DstVT != MVT::f64)
19172 // This conversion needs to be expanded.
19175 SDValue InVec = Op->getOperand(0);
19177 unsigned NumElts = SrcVT.getVectorNumElements();
19178 EVT SVT = SrcVT.getVectorElementType();
19180 // Widen the vector in input in the case of MVT::v2i32.
19181 // Example: from MVT::v2i32 to MVT::v4i32.
19182 SmallVector<SDValue, 16> Elts;
19183 for (unsigned i = 0, e = NumElts; i != e; ++i)
19184 Elts.push_back(DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, SVT, InVec,
19185 DAG.getIntPtrConstant(i, dl)));
19187 // Explicitly mark the extra elements as Undef.
19188 Elts.append(NumElts, DAG.getUNDEF(SVT));
19190 EVT NewVT = EVT::getVectorVT(*DAG.getContext(), SVT, NumElts * 2);
19191 SDValue BV = DAG.getNode(ISD::BUILD_VECTOR, dl, NewVT, Elts);
19192 SDValue ToV2F64 = DAG.getBitcast(MVT::v2f64, BV);
19193 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, ToV2F64,
19194 DAG.getIntPtrConstant(0, dl));
19197 assert(Subtarget->is64Bit() && !Subtarget->hasSSE2() &&
19198 Subtarget->hasMMX() && "Unexpected custom BITCAST");
19199 assert((DstVT == MVT::i64 ||
19200 (DstVT.isVector() && DstVT.getSizeInBits()==64)) &&
19201 "Unexpected custom BITCAST");
19202 // i64 <=> MMX conversions are Legal.
19203 if (SrcVT==MVT::i64 && DstVT.isVector())
19205 if (DstVT==MVT::i64 && SrcVT.isVector())
19207 // MMX <=> MMX conversions are Legal.
19208 if (SrcVT.isVector() && DstVT.isVector())
19210 // All other conversions need to be expanded.
19214 /// Compute the horizontal sum of bytes in V for the elements of VT.
19216 /// Requires V to be a byte vector and VT to be an integer vector type with
19217 /// wider elements than V's type. The width of the elements of VT determines
19218 /// how many bytes of V are summed horizontally to produce each element of the
19220 static SDValue LowerHorizontalByteSum(SDValue V, MVT VT,
19221 const X86Subtarget *Subtarget,
19222 SelectionDAG &DAG) {
19224 MVT ByteVecVT = V.getSimpleValueType();
19225 MVT EltVT = VT.getVectorElementType();
19226 int NumElts = VT.getVectorNumElements();
19227 assert(ByteVecVT.getVectorElementType() == MVT::i8 &&
19228 "Expected value to have byte element type.");
19229 assert(EltVT != MVT::i8 &&
19230 "Horizontal byte sum only makes sense for wider elements!");
19231 unsigned VecSize = VT.getSizeInBits();
19232 assert(ByteVecVT.getSizeInBits() == VecSize && "Cannot change vector size!");
19234 // PSADBW instruction horizontally add all bytes and leave the result in i64
19235 // chunks, thus directly computes the pop count for v2i64 and v4i64.
19236 if (EltVT == MVT::i64) {
19237 SDValue Zeros = getZeroVector(ByteVecVT, Subtarget, DAG, DL);
19238 V = DAG.getNode(X86ISD::PSADBW, DL, ByteVecVT, V, Zeros);
19239 return DAG.getBitcast(VT, V);
19242 if (EltVT == MVT::i32) {
19243 // We unpack the low half and high half into i32s interleaved with zeros so
19244 // that we can use PSADBW to horizontally sum them. The most useful part of
19245 // this is that it lines up the results of two PSADBW instructions to be
19246 // two v2i64 vectors which concatenated are the 4 population counts. We can
19247 // then use PACKUSWB to shrink and concatenate them into a v4i32 again.
19248 SDValue Zeros = getZeroVector(VT, Subtarget, DAG, DL);
19249 SDValue Low = DAG.getNode(X86ISD::UNPCKL, DL, VT, V, Zeros);
19250 SDValue High = DAG.getNode(X86ISD::UNPCKH, DL, VT, V, Zeros);
19252 // Do the horizontal sums into two v2i64s.
19253 Zeros = getZeroVector(ByteVecVT, Subtarget, DAG, DL);
19254 Low = DAG.getNode(X86ISD::PSADBW, DL, ByteVecVT,
19255 DAG.getBitcast(ByteVecVT, Low), Zeros);
19256 High = DAG.getNode(X86ISD::PSADBW, DL, ByteVecVT,
19257 DAG.getBitcast(ByteVecVT, High), Zeros);
19259 // Merge them together.
19260 MVT ShortVecVT = MVT::getVectorVT(MVT::i16, VecSize / 16);
19261 V = DAG.getNode(X86ISD::PACKUS, DL, ByteVecVT,
19262 DAG.getBitcast(ShortVecVT, Low),
19263 DAG.getBitcast(ShortVecVT, High));
19265 return DAG.getBitcast(VT, V);
19268 // The only element type left is i16.
19269 assert(EltVT == MVT::i16 && "Unknown how to handle type");
19271 // To obtain pop count for each i16 element starting from the pop count for
19272 // i8 elements, shift the i16s left by 8, sum as i8s, and then shift as i16s
19273 // right by 8. It is important to shift as i16s as i8 vector shift isn't
19274 // directly supported.
19275 SmallVector<SDValue, 16> Shifters(NumElts, DAG.getConstant(8, DL, EltVT));
19276 SDValue Shifter = DAG.getNode(ISD::BUILD_VECTOR, DL, VT, Shifters);
19277 SDValue Shl = DAG.getNode(ISD::SHL, DL, VT, DAG.getBitcast(VT, V), Shifter);
19278 V = DAG.getNode(ISD::ADD, DL, ByteVecVT, DAG.getBitcast(ByteVecVT, Shl),
19279 DAG.getBitcast(ByteVecVT, V));
19280 return DAG.getNode(ISD::SRL, DL, VT, DAG.getBitcast(VT, V), Shifter);
19283 static SDValue LowerVectorCTPOPInRegLUT(SDValue Op, SDLoc DL,
19284 const X86Subtarget *Subtarget,
19285 SelectionDAG &DAG) {
19286 MVT VT = Op.getSimpleValueType();
19287 MVT EltVT = VT.getVectorElementType();
19288 unsigned VecSize = VT.getSizeInBits();
19290 // Implement a lookup table in register by using an algorithm based on:
19291 // http://wm.ite.pl/articles/sse-popcount.html
19293 // The general idea is that every lower byte nibble in the input vector is an
19294 // index into a in-register pre-computed pop count table. We then split up the
19295 // input vector in two new ones: (1) a vector with only the shifted-right
19296 // higher nibbles for each byte and (2) a vector with the lower nibbles (and
19297 // masked out higher ones) for each byte. PSHUB is used separately with both
19298 // to index the in-register table. Next, both are added and the result is a
19299 // i8 vector where each element contains the pop count for input byte.
19301 // To obtain the pop count for elements != i8, we follow up with the same
19302 // approach and use additional tricks as described below.
19304 const int LUT[16] = {/* 0 */ 0, /* 1 */ 1, /* 2 */ 1, /* 3 */ 2,
19305 /* 4 */ 1, /* 5 */ 2, /* 6 */ 2, /* 7 */ 3,
19306 /* 8 */ 1, /* 9 */ 2, /* a */ 2, /* b */ 3,
19307 /* c */ 2, /* d */ 3, /* e */ 3, /* f */ 4};
19309 int NumByteElts = VecSize / 8;
19310 MVT ByteVecVT = MVT::getVectorVT(MVT::i8, NumByteElts);
19311 SDValue In = DAG.getBitcast(ByteVecVT, Op);
19312 SmallVector<SDValue, 16> LUTVec;
19313 for (int i = 0; i < NumByteElts; ++i)
19314 LUTVec.push_back(DAG.getConstant(LUT[i % 16], DL, MVT::i8));
19315 SDValue InRegLUT = DAG.getNode(ISD::BUILD_VECTOR, DL, ByteVecVT, LUTVec);
19316 SmallVector<SDValue, 16> Mask0F(NumByteElts,
19317 DAG.getConstant(0x0F, DL, MVT::i8));
19318 SDValue M0F = DAG.getNode(ISD::BUILD_VECTOR, DL, ByteVecVT, Mask0F);
19321 SmallVector<SDValue, 16> Four(NumByteElts, DAG.getConstant(4, DL, MVT::i8));
19322 SDValue FourV = DAG.getNode(ISD::BUILD_VECTOR, DL, ByteVecVT, Four);
19323 SDValue HighNibbles = DAG.getNode(ISD::SRL, DL, ByteVecVT, In, FourV);
19326 SDValue LowNibbles = DAG.getNode(ISD::AND, DL, ByteVecVT, In, M0F);
19328 // The input vector is used as the shuffle mask that index elements into the
19329 // LUT. After counting low and high nibbles, add the vector to obtain the
19330 // final pop count per i8 element.
19331 SDValue HighPopCnt =
19332 DAG.getNode(X86ISD::PSHUFB, DL, ByteVecVT, InRegLUT, HighNibbles);
19333 SDValue LowPopCnt =
19334 DAG.getNode(X86ISD::PSHUFB, DL, ByteVecVT, InRegLUT, LowNibbles);
19335 SDValue PopCnt = DAG.getNode(ISD::ADD, DL, ByteVecVT, HighPopCnt, LowPopCnt);
19337 if (EltVT == MVT::i8)
19340 return LowerHorizontalByteSum(PopCnt, VT, Subtarget, DAG);
19343 static SDValue LowerVectorCTPOPBitmath(SDValue Op, SDLoc DL,
19344 const X86Subtarget *Subtarget,
19345 SelectionDAG &DAG) {
19346 MVT VT = Op.getSimpleValueType();
19347 assert(VT.is128BitVector() &&
19348 "Only 128-bit vector bitmath lowering supported.");
19350 int VecSize = VT.getSizeInBits();
19351 MVT EltVT = VT.getVectorElementType();
19352 int Len = EltVT.getSizeInBits();
19354 // This is the vectorized version of the "best" algorithm from
19355 // http://graphics.stanford.edu/~seander/bithacks.html#CountBitsSetParallel
19356 // with a minor tweak to use a series of adds + shifts instead of vector
19357 // multiplications. Implemented for all integer vector types. We only use
19358 // this when we don't have SSSE3 which allows a LUT-based lowering that is
19359 // much faster, even faster than using native popcnt instructions.
19361 auto GetShift = [&](unsigned OpCode, SDValue V, int Shifter) {
19362 MVT VT = V.getSimpleValueType();
19363 SmallVector<SDValue, 32> Shifters(
19364 VT.getVectorNumElements(),
19365 DAG.getConstant(Shifter, DL, VT.getVectorElementType()));
19366 return DAG.getNode(OpCode, DL, VT, V,
19367 DAG.getNode(ISD::BUILD_VECTOR, DL, VT, Shifters));
19369 auto GetMask = [&](SDValue V, APInt Mask) {
19370 MVT VT = V.getSimpleValueType();
19371 SmallVector<SDValue, 32> Masks(
19372 VT.getVectorNumElements(),
19373 DAG.getConstant(Mask, DL, VT.getVectorElementType()));
19374 return DAG.getNode(ISD::AND, DL, VT, V,
19375 DAG.getNode(ISD::BUILD_VECTOR, DL, VT, Masks));
19378 // We don't want to incur the implicit masks required to SRL vNi8 vectors on
19379 // x86, so set the SRL type to have elements at least i16 wide. This is
19380 // correct because all of our SRLs are followed immediately by a mask anyways
19381 // that handles any bits that sneak into the high bits of the byte elements.
19382 MVT SrlVT = Len > 8 ? VT : MVT::getVectorVT(MVT::i16, VecSize / 16);
19386 // v = v - ((v >> 1) & 0x55555555...)
19388 DAG.getBitcast(VT, GetShift(ISD::SRL, DAG.getBitcast(SrlVT, V), 1));
19389 SDValue And = GetMask(Srl, APInt::getSplat(Len, APInt(8, 0x55)));
19390 V = DAG.getNode(ISD::SUB, DL, VT, V, And);
19392 // v = (v & 0x33333333...) + ((v >> 2) & 0x33333333...)
19393 SDValue AndLHS = GetMask(V, APInt::getSplat(Len, APInt(8, 0x33)));
19394 Srl = DAG.getBitcast(VT, GetShift(ISD::SRL, DAG.getBitcast(SrlVT, V), 2));
19395 SDValue AndRHS = GetMask(Srl, APInt::getSplat(Len, APInt(8, 0x33)));
19396 V = DAG.getNode(ISD::ADD, DL, VT, AndLHS, AndRHS);
19398 // v = (v + (v >> 4)) & 0x0F0F0F0F...
19399 Srl = DAG.getBitcast(VT, GetShift(ISD::SRL, DAG.getBitcast(SrlVT, V), 4));
19400 SDValue Add = DAG.getNode(ISD::ADD, DL, VT, V, Srl);
19401 V = GetMask(Add, APInt::getSplat(Len, APInt(8, 0x0F)));
19403 // At this point, V contains the byte-wise population count, and we are
19404 // merely doing a horizontal sum if necessary to get the wider element
19406 if (EltVT == MVT::i8)
19409 return LowerHorizontalByteSum(
19410 DAG.getBitcast(MVT::getVectorVT(MVT::i8, VecSize / 8), V), VT, Subtarget,
19414 static SDValue LowerVectorCTPOP(SDValue Op, const X86Subtarget *Subtarget,
19415 SelectionDAG &DAG) {
19416 MVT VT = Op.getSimpleValueType();
19417 // FIXME: Need to add AVX-512 support here!
19418 assert((VT.is256BitVector() || VT.is128BitVector()) &&
19419 "Unknown CTPOP type to handle");
19420 SDLoc DL(Op.getNode());
19421 SDValue Op0 = Op.getOperand(0);
19423 if (!Subtarget->hasSSSE3()) {
19424 // We can't use the fast LUT approach, so fall back on vectorized bitmath.
19425 assert(VT.is128BitVector() && "Only 128-bit vectors supported in SSE!");
19426 return LowerVectorCTPOPBitmath(Op0, DL, Subtarget, DAG);
19429 if (VT.is256BitVector() && !Subtarget->hasInt256()) {
19430 unsigned NumElems = VT.getVectorNumElements();
19432 // Extract each 128-bit vector, compute pop count and concat the result.
19433 SDValue LHS = Extract128BitVector(Op0, 0, DAG, DL);
19434 SDValue RHS = Extract128BitVector(Op0, NumElems/2, DAG, DL);
19436 return DAG.getNode(ISD::CONCAT_VECTORS, DL, VT,
19437 LowerVectorCTPOPInRegLUT(LHS, DL, Subtarget, DAG),
19438 LowerVectorCTPOPInRegLUT(RHS, DL, Subtarget, DAG));
19441 return LowerVectorCTPOPInRegLUT(Op0, DL, Subtarget, DAG);
19444 static SDValue LowerCTPOP(SDValue Op, const X86Subtarget *Subtarget,
19445 SelectionDAG &DAG) {
19446 assert(Op.getValueType().isVector() &&
19447 "We only do custom lowering for vector population count.");
19448 return LowerVectorCTPOP(Op, Subtarget, DAG);
19451 static SDValue LowerLOAD_SUB(SDValue Op, SelectionDAG &DAG) {
19452 SDNode *Node = Op.getNode();
19454 EVT T = Node->getValueType(0);
19455 SDValue negOp = DAG.getNode(ISD::SUB, dl, T,
19456 DAG.getConstant(0, dl, T), Node->getOperand(2));
19457 return DAG.getAtomic(ISD::ATOMIC_LOAD_ADD, dl,
19458 cast<AtomicSDNode>(Node)->getMemoryVT(),
19459 Node->getOperand(0),
19460 Node->getOperand(1), negOp,
19461 cast<AtomicSDNode>(Node)->getMemOperand(),
19462 cast<AtomicSDNode>(Node)->getOrdering(),
19463 cast<AtomicSDNode>(Node)->getSynchScope());
19466 static SDValue LowerATOMIC_STORE(SDValue Op, SelectionDAG &DAG) {
19467 SDNode *Node = Op.getNode();
19469 EVT VT = cast<AtomicSDNode>(Node)->getMemoryVT();
19471 // Convert seq_cst store -> xchg
19472 // Convert wide store -> swap (-> cmpxchg8b/cmpxchg16b)
19473 // FIXME: On 32-bit, store -> fist or movq would be more efficient
19474 // (The only way to get a 16-byte store is cmpxchg16b)
19475 // FIXME: 16-byte ATOMIC_SWAP isn't actually hooked up at the moment.
19476 if (cast<AtomicSDNode>(Node)->getOrdering() == SequentiallyConsistent ||
19477 !DAG.getTargetLoweringInfo().isTypeLegal(VT)) {
19478 SDValue Swap = DAG.getAtomic(ISD::ATOMIC_SWAP, dl,
19479 cast<AtomicSDNode>(Node)->getMemoryVT(),
19480 Node->getOperand(0),
19481 Node->getOperand(1), Node->getOperand(2),
19482 cast<AtomicSDNode>(Node)->getMemOperand(),
19483 cast<AtomicSDNode>(Node)->getOrdering(),
19484 cast<AtomicSDNode>(Node)->getSynchScope());
19485 return Swap.getValue(1);
19487 // Other atomic stores have a simple pattern.
19491 static SDValue LowerADDC_ADDE_SUBC_SUBE(SDValue Op, SelectionDAG &DAG) {
19492 EVT VT = Op.getNode()->getSimpleValueType(0);
19494 // Let legalize expand this if it isn't a legal type yet.
19495 if (!DAG.getTargetLoweringInfo().isTypeLegal(VT))
19498 SDVTList VTs = DAG.getVTList(VT, MVT::i32);
19501 bool ExtraOp = false;
19502 switch (Op.getOpcode()) {
19503 default: llvm_unreachable("Invalid code");
19504 case ISD::ADDC: Opc = X86ISD::ADD; break;
19505 case ISD::ADDE: Opc = X86ISD::ADC; ExtraOp = true; break;
19506 case ISD::SUBC: Opc = X86ISD::SUB; break;
19507 case ISD::SUBE: Opc = X86ISD::SBB; ExtraOp = true; break;
19511 return DAG.getNode(Opc, SDLoc(Op), VTs, Op.getOperand(0),
19513 return DAG.getNode(Opc, SDLoc(Op), VTs, Op.getOperand(0),
19514 Op.getOperand(1), Op.getOperand(2));
19517 static SDValue LowerFSINCOS(SDValue Op, const X86Subtarget *Subtarget,
19518 SelectionDAG &DAG) {
19519 assert(Subtarget->isTargetDarwin() && Subtarget->is64Bit());
19521 // For MacOSX, we want to call an alternative entry point: __sincos_stret,
19522 // which returns the values as { float, float } (in XMM0) or
19523 // { double, double } (which is returned in XMM0, XMM1).
19525 SDValue Arg = Op.getOperand(0);
19526 EVT ArgVT = Arg.getValueType();
19527 Type *ArgTy = ArgVT.getTypeForEVT(*DAG.getContext());
19529 TargetLowering::ArgListTy Args;
19530 TargetLowering::ArgListEntry Entry;
19534 Entry.isSExt = false;
19535 Entry.isZExt = false;
19536 Args.push_back(Entry);
19538 bool isF64 = ArgVT == MVT::f64;
19539 // Only optimize x86_64 for now. i386 is a bit messy. For f32,
19540 // the small struct {f32, f32} is returned in (eax, edx). For f64,
19541 // the results are returned via SRet in memory.
19542 const char *LibcallName = isF64 ? "__sincos_stret" : "__sincosf_stret";
19543 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
19545 DAG.getExternalSymbol(LibcallName, TLI.getPointerTy(DAG.getDataLayout()));
19547 Type *RetTy = isF64
19548 ? (Type*)StructType::get(ArgTy, ArgTy, nullptr)
19549 : (Type*)VectorType::get(ArgTy, 4);
19551 TargetLowering::CallLoweringInfo CLI(DAG);
19552 CLI.setDebugLoc(dl).setChain(DAG.getEntryNode())
19553 .setCallee(CallingConv::C, RetTy, Callee, std::move(Args), 0);
19555 std::pair<SDValue, SDValue> CallResult = TLI.LowerCallTo(CLI);
19558 // Returned in xmm0 and xmm1.
19559 return CallResult.first;
19561 // Returned in bits 0:31 and 32:64 xmm0.
19562 SDValue SinVal = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, ArgVT,
19563 CallResult.first, DAG.getIntPtrConstant(0, dl));
19564 SDValue CosVal = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, ArgVT,
19565 CallResult.first, DAG.getIntPtrConstant(1, dl));
19566 SDVTList Tys = DAG.getVTList(ArgVT, ArgVT);
19567 return DAG.getNode(ISD::MERGE_VALUES, dl, Tys, SinVal, CosVal);
19570 static SDValue LowerMSCATTER(SDValue Op, const X86Subtarget *Subtarget,
19571 SelectionDAG &DAG) {
19572 assert(Subtarget->hasAVX512() &&
19573 "MGATHER/MSCATTER are supported on AVX-512 arch only");
19575 MaskedScatterSDNode *N = cast<MaskedScatterSDNode>(Op.getNode());
19576 EVT VT = N->getValue().getValueType();
19577 assert(VT.getScalarSizeInBits() >= 32 && "Unsupported scatter op");
19580 // X86 scatter kills mask register, so its type should be added to
19581 // the list of return values
19582 if (N->getNumValues() == 1) {
19583 SDValue Index = N->getIndex();
19584 if (!Subtarget->hasVLX() && !VT.is512BitVector() &&
19585 !Index.getValueType().is512BitVector())
19586 Index = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v8i64, Index);
19588 SDVTList VTs = DAG.getVTList(N->getMask().getValueType(), MVT::Other);
19589 SDValue Ops[] = { N->getOperand(0), N->getOperand(1), N->getOperand(2),
19590 N->getOperand(3), Index };
19592 SDValue NewScatter = DAG.getMaskedScatter(VTs, VT, dl, Ops, N->getMemOperand());
19593 DAG.ReplaceAllUsesWith(Op, SDValue(NewScatter.getNode(), 1));
19594 return SDValue(NewScatter.getNode(), 0);
19599 static SDValue LowerMGATHER(SDValue Op, const X86Subtarget *Subtarget,
19600 SelectionDAG &DAG) {
19601 assert(Subtarget->hasAVX512() &&
19602 "MGATHER/MSCATTER are supported on AVX-512 arch only");
19604 MaskedGatherSDNode *N = cast<MaskedGatherSDNode>(Op.getNode());
19605 EVT VT = Op.getValueType();
19606 assert(VT.getScalarSizeInBits() >= 32 && "Unsupported gather op");
19609 SDValue Index = N->getIndex();
19610 if (!Subtarget->hasVLX() && !VT.is512BitVector() &&
19611 !Index.getValueType().is512BitVector()) {
19612 Index = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v8i64, Index);
19613 SDValue Ops[] = { N->getOperand(0), N->getOperand(1), N->getOperand(2),
19614 N->getOperand(3), Index };
19615 DAG.UpdateNodeOperands(N, Ops);
19620 SDValue X86TargetLowering::LowerGC_TRANSITION_START(SDValue Op,
19621 SelectionDAG &DAG) const {
19622 // TODO: Eventually, the lowering of these nodes should be informed by or
19623 // deferred to the GC strategy for the function in which they appear. For
19624 // now, however, they must be lowered to something. Since they are logically
19625 // no-ops in the case of a null GC strategy (or a GC strategy which does not
19626 // require special handling for these nodes), lower them as literal NOOPs for
19628 SmallVector<SDValue, 2> Ops;
19630 Ops.push_back(Op.getOperand(0));
19631 if (Op->getGluedNode())
19632 Ops.push_back(Op->getOperand(Op->getNumOperands() - 1));
19635 SDVTList VTs = DAG.getVTList(MVT::Other, MVT::Glue);
19636 SDValue NOOP(DAG.getMachineNode(X86::NOOP, SDLoc(Op), VTs, Ops), 0);
19641 SDValue X86TargetLowering::LowerGC_TRANSITION_END(SDValue Op,
19642 SelectionDAG &DAG) const {
19643 // TODO: Eventually, the lowering of these nodes should be informed by or
19644 // deferred to the GC strategy for the function in which they appear. For
19645 // now, however, they must be lowered to something. Since they are logically
19646 // no-ops in the case of a null GC strategy (or a GC strategy which does not
19647 // require special handling for these nodes), lower them as literal NOOPs for
19649 SmallVector<SDValue, 2> Ops;
19651 Ops.push_back(Op.getOperand(0));
19652 if (Op->getGluedNode())
19653 Ops.push_back(Op->getOperand(Op->getNumOperands() - 1));
19656 SDVTList VTs = DAG.getVTList(MVT::Other, MVT::Glue);
19657 SDValue NOOP(DAG.getMachineNode(X86::NOOP, SDLoc(Op), VTs, Ops), 0);
19662 /// LowerOperation - Provide custom lowering hooks for some operations.
19664 SDValue X86TargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
19665 switch (Op.getOpcode()) {
19666 default: llvm_unreachable("Should not custom lower this!");
19667 case ISD::ATOMIC_FENCE: return LowerATOMIC_FENCE(Op, Subtarget, DAG);
19668 case ISD::ATOMIC_CMP_SWAP_WITH_SUCCESS:
19669 return LowerCMP_SWAP(Op, Subtarget, DAG);
19670 case ISD::CTPOP: return LowerCTPOP(Op, Subtarget, DAG);
19671 case ISD::ATOMIC_LOAD_SUB: return LowerLOAD_SUB(Op,DAG);
19672 case ISD::ATOMIC_STORE: return LowerATOMIC_STORE(Op,DAG);
19673 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG);
19674 case ISD::CONCAT_VECTORS: return LowerCONCAT_VECTORS(Op, Subtarget, DAG);
19675 case ISD::VECTOR_SHUFFLE: return lowerVectorShuffle(Op, Subtarget, DAG);
19676 case ISD::VSELECT: return LowerVSELECT(Op, DAG);
19677 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
19678 case ISD::INSERT_VECTOR_ELT: return LowerINSERT_VECTOR_ELT(Op, DAG);
19679 case ISD::EXTRACT_SUBVECTOR: return LowerEXTRACT_SUBVECTOR(Op,Subtarget,DAG);
19680 case ISD::INSERT_SUBVECTOR: return LowerINSERT_SUBVECTOR(Op, Subtarget,DAG);
19681 case ISD::SCALAR_TO_VECTOR: return LowerSCALAR_TO_VECTOR(Op, DAG);
19682 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
19683 case ISD::GlobalAddress: return LowerGlobalAddress(Op, DAG);
19684 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
19685 case ISD::ExternalSymbol: return LowerExternalSymbol(Op, DAG);
19686 case ISD::BlockAddress: return LowerBlockAddress(Op, DAG);
19687 case ISD::SHL_PARTS:
19688 case ISD::SRA_PARTS:
19689 case ISD::SRL_PARTS: return LowerShiftParts(Op, DAG);
19690 case ISD::SINT_TO_FP: return LowerSINT_TO_FP(Op, DAG);
19691 case ISD::UINT_TO_FP: return LowerUINT_TO_FP(Op, DAG);
19692 case ISD::TRUNCATE: return LowerTRUNCATE(Op, DAG);
19693 case ISD::ZERO_EXTEND: return LowerZERO_EXTEND(Op, Subtarget, DAG);
19694 case ISD::SIGN_EXTEND: return LowerSIGN_EXTEND(Op, Subtarget, DAG);
19695 case ISD::ANY_EXTEND: return LowerANY_EXTEND(Op, Subtarget, DAG);
19696 case ISD::SIGN_EXTEND_VECTOR_INREG:
19697 return LowerSIGN_EXTEND_VECTOR_INREG(Op, Subtarget, DAG);
19698 case ISD::FP_TO_SINT: return LowerFP_TO_SINT(Op, DAG);
19699 case ISD::FP_TO_UINT: return LowerFP_TO_UINT(Op, DAG);
19700 case ISD::FP_EXTEND: return LowerFP_EXTEND(Op, DAG);
19701 case ISD::LOAD: return LowerExtendedLoad(Op, Subtarget, DAG);
19703 case ISD::FNEG: return LowerFABSorFNEG(Op, DAG);
19704 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
19705 case ISD::FGETSIGN: return LowerFGETSIGN(Op, DAG);
19706 case ISD::SETCC: return LowerSETCC(Op, DAG);
19707 case ISD::SELECT: return LowerSELECT(Op, DAG);
19708 case ISD::BRCOND: return LowerBRCOND(Op, DAG);
19709 case ISD::JumpTable: return LowerJumpTable(Op, DAG);
19710 case ISD::VASTART: return LowerVASTART(Op, DAG);
19711 case ISD::VAARG: return LowerVAARG(Op, DAG);
19712 case ISD::VACOPY: return LowerVACOPY(Op, Subtarget, DAG);
19713 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, Subtarget, DAG);
19714 case ISD::INTRINSIC_VOID:
19715 case ISD::INTRINSIC_W_CHAIN: return LowerINTRINSIC_W_CHAIN(Op, Subtarget, DAG);
19716 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
19717 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
19718 case ISD::FRAME_TO_ARGS_OFFSET:
19719 return LowerFRAME_TO_ARGS_OFFSET(Op, DAG);
19720 case ISD::DYNAMIC_STACKALLOC: return LowerDYNAMIC_STACKALLOC(Op, DAG);
19721 case ISD::EH_RETURN: return LowerEH_RETURN(Op, DAG);
19722 case ISD::EH_SJLJ_SETJMP: return lowerEH_SJLJ_SETJMP(Op, DAG);
19723 case ISD::EH_SJLJ_LONGJMP: return lowerEH_SJLJ_LONGJMP(Op, DAG);
19724 case ISD::INIT_TRAMPOLINE: return LowerINIT_TRAMPOLINE(Op, DAG);
19725 case ISD::ADJUST_TRAMPOLINE: return LowerADJUST_TRAMPOLINE(Op, DAG);
19726 case ISD::FLT_ROUNDS_: return LowerFLT_ROUNDS_(Op, DAG);
19727 case ISD::CTLZ: return LowerCTLZ(Op, Subtarget, DAG);
19728 case ISD::CTLZ_ZERO_UNDEF: return LowerCTLZ_ZERO_UNDEF(Op, Subtarget, DAG);
19730 case ISD::CTTZ_ZERO_UNDEF: return LowerCTTZ(Op, DAG);
19731 case ISD::MUL: return LowerMUL(Op, Subtarget, DAG);
19732 case ISD::UMUL_LOHI:
19733 case ISD::SMUL_LOHI: return LowerMUL_LOHI(Op, Subtarget, DAG);
19734 case ISD::ROTL: return LowerRotate(Op, Subtarget, DAG);
19737 case ISD::SHL: return LowerShift(Op, Subtarget, DAG);
19743 case ISD::UMULO: return LowerXALUO(Op, DAG);
19744 case ISD::READCYCLECOUNTER: return LowerREADCYCLECOUNTER(Op, Subtarget,DAG);
19745 case ISD::BITCAST: return LowerBITCAST(Op, Subtarget, DAG);
19749 case ISD::SUBE: return LowerADDC_ADDE_SUBC_SUBE(Op, DAG);
19750 case ISD::ADD: return LowerADD(Op, DAG);
19751 case ISD::SUB: return LowerSUB(Op, DAG);
19755 case ISD::UMIN: return LowerMINMAX(Op, DAG);
19756 case ISD::FSINCOS: return LowerFSINCOS(Op, Subtarget, DAG);
19757 case ISD::MGATHER: return LowerMGATHER(Op, Subtarget, DAG);
19758 case ISD::MSCATTER: return LowerMSCATTER(Op, Subtarget, DAG);
19759 case ISD::GC_TRANSITION_START:
19760 return LowerGC_TRANSITION_START(Op, DAG);
19761 case ISD::GC_TRANSITION_END: return LowerGC_TRANSITION_END(Op, DAG);
19765 /// ReplaceNodeResults - Replace a node with an illegal result type
19766 /// with a new node built out of custom code.
19767 void X86TargetLowering::ReplaceNodeResults(SDNode *N,
19768 SmallVectorImpl<SDValue>&Results,
19769 SelectionDAG &DAG) const {
19771 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
19772 switch (N->getOpcode()) {
19774 llvm_unreachable("Do not know how to custom type legalize this operation!");
19775 // We might have generated v2f32 FMIN/FMAX operations. Widen them to v4f32.
19776 case X86ISD::FMINC:
19778 case X86ISD::FMAXC:
19779 case X86ISD::FMAX: {
19780 EVT VT = N->getValueType(0);
19781 if (VT != MVT::v2f32)
19782 llvm_unreachable("Unexpected type (!= v2f32) on FMIN/FMAX.");
19783 SDValue UNDEF = DAG.getUNDEF(VT);
19784 SDValue LHS = DAG.getNode(ISD::CONCAT_VECTORS, dl, MVT::v4f32,
19785 N->getOperand(0), UNDEF);
19786 SDValue RHS = DAG.getNode(ISD::CONCAT_VECTORS, dl, MVT::v4f32,
19787 N->getOperand(1), UNDEF);
19788 Results.push_back(DAG.getNode(N->getOpcode(), dl, MVT::v4f32, LHS, RHS));
19791 case ISD::SIGN_EXTEND_INREG:
19796 // We don't want to expand or promote these.
19803 case ISD::UDIVREM: {
19804 SDValue V = LowerWin64_i128OP(SDValue(N,0), DAG);
19805 Results.push_back(V);
19808 case ISD::FP_TO_SINT:
19809 case ISD::FP_TO_UINT: {
19810 bool IsSigned = N->getOpcode() == ISD::FP_TO_SINT;
19812 std::pair<SDValue,SDValue> Vals =
19813 FP_TO_INTHelper(SDValue(N, 0), DAG, IsSigned, /*IsReplace=*/ true);
19814 SDValue FIST = Vals.first, StackSlot = Vals.second;
19815 if (FIST.getNode()) {
19816 EVT VT = N->getValueType(0);
19817 // Return a load from the stack slot.
19818 if (StackSlot.getNode())
19819 Results.push_back(DAG.getLoad(VT, dl, FIST, StackSlot,
19820 MachinePointerInfo(),
19821 false, false, false, 0));
19823 Results.push_back(FIST);
19827 case ISD::UINT_TO_FP: {
19828 assert(Subtarget->hasSSE2() && "Requires at least SSE2!");
19829 if (N->getOperand(0).getValueType() != MVT::v2i32 ||
19830 N->getValueType(0) != MVT::v2f32)
19832 SDValue ZExtIn = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v2i64,
19834 SDValue Bias = DAG.getConstantFP(BitsToDouble(0x4330000000000000ULL), dl,
19836 SDValue VBias = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v2f64, Bias, Bias);
19837 SDValue Or = DAG.getNode(ISD::OR, dl, MVT::v2i64, ZExtIn,
19838 DAG.getBitcast(MVT::v2i64, VBias));
19839 Or = DAG.getBitcast(MVT::v2f64, Or);
19840 // TODO: Are there any fast-math-flags to propagate here?
19841 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::v2f64, Or, VBias);
19842 Results.push_back(DAG.getNode(X86ISD::VFPROUND, dl, MVT::v4f32, Sub));
19845 case ISD::FP_ROUND: {
19846 if (!TLI.isTypeLegal(N->getOperand(0).getValueType()))
19848 SDValue V = DAG.getNode(X86ISD::VFPROUND, dl, MVT::v4f32, N->getOperand(0));
19849 Results.push_back(V);
19852 case ISD::FP_EXTEND: {
19853 // Right now, only MVT::v2f32 has OperationAction for FP_EXTEND.
19854 // No other ValueType for FP_EXTEND should reach this point.
19855 assert(N->getValueType(0) == MVT::v2f32 &&
19856 "Do not know how to legalize this Node");
19859 case ISD::INTRINSIC_W_CHAIN: {
19860 unsigned IntNo = cast<ConstantSDNode>(N->getOperand(1))->getZExtValue();
19862 default : llvm_unreachable("Do not know how to custom type "
19863 "legalize this intrinsic operation!");
19864 case Intrinsic::x86_rdtsc:
19865 return getReadTimeStampCounter(N, dl, X86ISD::RDTSC_DAG, DAG, Subtarget,
19867 case Intrinsic::x86_rdtscp:
19868 return getReadTimeStampCounter(N, dl, X86ISD::RDTSCP_DAG, DAG, Subtarget,
19870 case Intrinsic::x86_rdpmc:
19871 return getReadPerformanceCounter(N, dl, DAG, Subtarget, Results);
19874 case ISD::READCYCLECOUNTER: {
19875 return getReadTimeStampCounter(N, dl, X86ISD::RDTSC_DAG, DAG, Subtarget,
19878 case ISD::ATOMIC_CMP_SWAP_WITH_SUCCESS: {
19879 EVT T = N->getValueType(0);
19880 assert((T == MVT::i64 || T == MVT::i128) && "can only expand cmpxchg pair");
19881 bool Regs64bit = T == MVT::i128;
19882 EVT HalfT = Regs64bit ? MVT::i64 : MVT::i32;
19883 SDValue cpInL, cpInH;
19884 cpInL = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, HalfT, N->getOperand(2),
19885 DAG.getConstant(0, dl, HalfT));
19886 cpInH = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, HalfT, N->getOperand(2),
19887 DAG.getConstant(1, dl, HalfT));
19888 cpInL = DAG.getCopyToReg(N->getOperand(0), dl,
19889 Regs64bit ? X86::RAX : X86::EAX,
19891 cpInH = DAG.getCopyToReg(cpInL.getValue(0), dl,
19892 Regs64bit ? X86::RDX : X86::EDX,
19893 cpInH, cpInL.getValue(1));
19894 SDValue swapInL, swapInH;
19895 swapInL = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, HalfT, N->getOperand(3),
19896 DAG.getConstant(0, dl, HalfT));
19897 swapInH = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, HalfT, N->getOperand(3),
19898 DAG.getConstant(1, dl, HalfT));
19899 swapInL = DAG.getCopyToReg(cpInH.getValue(0), dl,
19900 Regs64bit ? X86::RBX : X86::EBX,
19901 swapInL, cpInH.getValue(1));
19902 swapInH = DAG.getCopyToReg(swapInL.getValue(0), dl,
19903 Regs64bit ? X86::RCX : X86::ECX,
19904 swapInH, swapInL.getValue(1));
19905 SDValue Ops[] = { swapInH.getValue(0),
19907 swapInH.getValue(1) };
19908 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Glue);
19909 MachineMemOperand *MMO = cast<AtomicSDNode>(N)->getMemOperand();
19910 unsigned Opcode = Regs64bit ? X86ISD::LCMPXCHG16_DAG :
19911 X86ISD::LCMPXCHG8_DAG;
19912 SDValue Result = DAG.getMemIntrinsicNode(Opcode, dl, Tys, Ops, T, MMO);
19913 SDValue cpOutL = DAG.getCopyFromReg(Result.getValue(0), dl,
19914 Regs64bit ? X86::RAX : X86::EAX,
19915 HalfT, Result.getValue(1));
19916 SDValue cpOutH = DAG.getCopyFromReg(cpOutL.getValue(1), dl,
19917 Regs64bit ? X86::RDX : X86::EDX,
19918 HalfT, cpOutL.getValue(2));
19919 SDValue OpsF[] = { cpOutL.getValue(0), cpOutH.getValue(0)};
19921 SDValue EFLAGS = DAG.getCopyFromReg(cpOutH.getValue(1), dl, X86::EFLAGS,
19922 MVT::i32, cpOutH.getValue(2));
19924 DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
19925 DAG.getConstant(X86::COND_E, dl, MVT::i8), EFLAGS);
19926 Success = DAG.getZExtOrTrunc(Success, dl, N->getValueType(1));
19928 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, T, OpsF));
19929 Results.push_back(Success);
19930 Results.push_back(EFLAGS.getValue(1));
19933 case ISD::ATOMIC_SWAP:
19934 case ISD::ATOMIC_LOAD_ADD:
19935 case ISD::ATOMIC_LOAD_SUB:
19936 case ISD::ATOMIC_LOAD_AND:
19937 case ISD::ATOMIC_LOAD_OR:
19938 case ISD::ATOMIC_LOAD_XOR:
19939 case ISD::ATOMIC_LOAD_NAND:
19940 case ISD::ATOMIC_LOAD_MIN:
19941 case ISD::ATOMIC_LOAD_MAX:
19942 case ISD::ATOMIC_LOAD_UMIN:
19943 case ISD::ATOMIC_LOAD_UMAX:
19944 case ISD::ATOMIC_LOAD: {
19945 // Delegate to generic TypeLegalization. Situations we can really handle
19946 // should have already been dealt with by AtomicExpandPass.cpp.
19949 case ISD::BITCAST: {
19950 assert(Subtarget->hasSSE2() && "Requires at least SSE2!");
19951 EVT DstVT = N->getValueType(0);
19952 EVT SrcVT = N->getOperand(0)->getValueType(0);
19954 if (SrcVT != MVT::f64 ||
19955 (DstVT != MVT::v2i32 && DstVT != MVT::v4i16 && DstVT != MVT::v8i8))
19958 unsigned NumElts = DstVT.getVectorNumElements();
19959 EVT SVT = DstVT.getVectorElementType();
19960 EVT WiderVT = EVT::getVectorVT(*DAG.getContext(), SVT, NumElts * 2);
19961 SDValue Expanded = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
19962 MVT::v2f64, N->getOperand(0));
19963 SDValue ToVecInt = DAG.getBitcast(WiderVT, Expanded);
19965 if (ExperimentalVectorWideningLegalization) {
19966 // If we are legalizing vectors by widening, we already have the desired
19967 // legal vector type, just return it.
19968 Results.push_back(ToVecInt);
19972 SmallVector<SDValue, 8> Elts;
19973 for (unsigned i = 0, e = NumElts; i != e; ++i)
19974 Elts.push_back(DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, SVT,
19975 ToVecInt, DAG.getIntPtrConstant(i, dl)));
19977 Results.push_back(DAG.getNode(ISD::BUILD_VECTOR, dl, DstVT, Elts));
19982 const char *X86TargetLowering::getTargetNodeName(unsigned Opcode) const {
19983 switch ((X86ISD::NodeType)Opcode) {
19984 case X86ISD::FIRST_NUMBER: break;
19985 case X86ISD::BSF: return "X86ISD::BSF";
19986 case X86ISD::BSR: return "X86ISD::BSR";
19987 case X86ISD::SHLD: return "X86ISD::SHLD";
19988 case X86ISD::SHRD: return "X86ISD::SHRD";
19989 case X86ISD::FAND: return "X86ISD::FAND";
19990 case X86ISD::FANDN: return "X86ISD::FANDN";
19991 case X86ISD::FOR: return "X86ISD::FOR";
19992 case X86ISD::FXOR: return "X86ISD::FXOR";
19993 case X86ISD::FILD: return "X86ISD::FILD";
19994 case X86ISD::FILD_FLAG: return "X86ISD::FILD_FLAG";
19995 case X86ISD::FP_TO_INT16_IN_MEM: return "X86ISD::FP_TO_INT16_IN_MEM";
19996 case X86ISD::FP_TO_INT32_IN_MEM: return "X86ISD::FP_TO_INT32_IN_MEM";
19997 case X86ISD::FP_TO_INT64_IN_MEM: return "X86ISD::FP_TO_INT64_IN_MEM";
19998 case X86ISD::FLD: return "X86ISD::FLD";
19999 case X86ISD::FST: return "X86ISD::FST";
20000 case X86ISD::CALL: return "X86ISD::CALL";
20001 case X86ISD::RDTSC_DAG: return "X86ISD::RDTSC_DAG";
20002 case X86ISD::RDTSCP_DAG: return "X86ISD::RDTSCP_DAG";
20003 case X86ISD::RDPMC_DAG: return "X86ISD::RDPMC_DAG";
20004 case X86ISD::BT: return "X86ISD::BT";
20005 case X86ISD::CMP: return "X86ISD::CMP";
20006 case X86ISD::COMI: return "X86ISD::COMI";
20007 case X86ISD::UCOMI: return "X86ISD::UCOMI";
20008 case X86ISD::CMPM: return "X86ISD::CMPM";
20009 case X86ISD::CMPMU: return "X86ISD::CMPMU";
20010 case X86ISD::CMPM_RND: return "X86ISD::CMPM_RND";
20011 case X86ISD::SETCC: return "X86ISD::SETCC";
20012 case X86ISD::SETCC_CARRY: return "X86ISD::SETCC_CARRY";
20013 case X86ISD::FSETCC: return "X86ISD::FSETCC";
20014 case X86ISD::FGETSIGNx86: return "X86ISD::FGETSIGNx86";
20015 case X86ISD::CMOV: return "X86ISD::CMOV";
20016 case X86ISD::BRCOND: return "X86ISD::BRCOND";
20017 case X86ISD::RET_FLAG: return "X86ISD::RET_FLAG";
20018 case X86ISD::REP_STOS: return "X86ISD::REP_STOS";
20019 case X86ISD::REP_MOVS: return "X86ISD::REP_MOVS";
20020 case X86ISD::GlobalBaseReg: return "X86ISD::GlobalBaseReg";
20021 case X86ISD::Wrapper: return "X86ISD::Wrapper";
20022 case X86ISD::WrapperRIP: return "X86ISD::WrapperRIP";
20023 case X86ISD::MOVDQ2Q: return "X86ISD::MOVDQ2Q";
20024 case X86ISD::MMX_MOVD2W: return "X86ISD::MMX_MOVD2W";
20025 case X86ISD::MMX_MOVW2D: return "X86ISD::MMX_MOVW2D";
20026 case X86ISD::PEXTRB: return "X86ISD::PEXTRB";
20027 case X86ISD::PEXTRW: return "X86ISD::PEXTRW";
20028 case X86ISD::INSERTPS: return "X86ISD::INSERTPS";
20029 case X86ISD::PINSRB: return "X86ISD::PINSRB";
20030 case X86ISD::PINSRW: return "X86ISD::PINSRW";
20031 case X86ISD::MMX_PINSRW: return "X86ISD::MMX_PINSRW";
20032 case X86ISD::PSHUFB: return "X86ISD::PSHUFB";
20033 case X86ISD::ANDNP: return "X86ISD::ANDNP";
20034 case X86ISD::PSIGN: return "X86ISD::PSIGN";
20035 case X86ISD::BLENDI: return "X86ISD::BLENDI";
20036 case X86ISD::SHRUNKBLEND: return "X86ISD::SHRUNKBLEND";
20037 case X86ISD::ADDUS: return "X86ISD::ADDUS";
20038 case X86ISD::SUBUS: return "X86ISD::SUBUS";
20039 case X86ISD::HADD: return "X86ISD::HADD";
20040 case X86ISD::HSUB: return "X86ISD::HSUB";
20041 case X86ISD::FHADD: return "X86ISD::FHADD";
20042 case X86ISD::FHSUB: return "X86ISD::FHSUB";
20043 case X86ISD::ABS: return "X86ISD::ABS";
20044 case X86ISD::CONFLICT: return "X86ISD::CONFLICT";
20045 case X86ISD::FMAX: return "X86ISD::FMAX";
20046 case X86ISD::FMAX_RND: return "X86ISD::FMAX_RND";
20047 case X86ISD::FMIN: return "X86ISD::FMIN";
20048 case X86ISD::FMIN_RND: return "X86ISD::FMIN_RND";
20049 case X86ISD::FMAXC: return "X86ISD::FMAXC";
20050 case X86ISD::FMINC: return "X86ISD::FMINC";
20051 case X86ISD::FRSQRT: return "X86ISD::FRSQRT";
20052 case X86ISD::FRCP: return "X86ISD::FRCP";
20053 case X86ISD::EXTRQI: return "X86ISD::EXTRQI";
20054 case X86ISD::INSERTQI: return "X86ISD::INSERTQI";
20055 case X86ISD::TLSADDR: return "X86ISD::TLSADDR";
20056 case X86ISD::TLSBASEADDR: return "X86ISD::TLSBASEADDR";
20057 case X86ISD::TLSCALL: return "X86ISD::TLSCALL";
20058 case X86ISD::EH_SJLJ_SETJMP: return "X86ISD::EH_SJLJ_SETJMP";
20059 case X86ISD::EH_SJLJ_LONGJMP: return "X86ISD::EH_SJLJ_LONGJMP";
20060 case X86ISD::EH_RETURN: return "X86ISD::EH_RETURN";
20061 case X86ISD::TC_RETURN: return "X86ISD::TC_RETURN";
20062 case X86ISD::FNSTCW16m: return "X86ISD::FNSTCW16m";
20063 case X86ISD::FNSTSW16r: return "X86ISD::FNSTSW16r";
20064 case X86ISD::LCMPXCHG_DAG: return "X86ISD::LCMPXCHG_DAG";
20065 case X86ISD::LCMPXCHG8_DAG: return "X86ISD::LCMPXCHG8_DAG";
20066 case X86ISD::LCMPXCHG16_DAG: return "X86ISD::LCMPXCHG16_DAG";
20067 case X86ISD::VZEXT_MOVL: return "X86ISD::VZEXT_MOVL";
20068 case X86ISD::VZEXT_LOAD: return "X86ISD::VZEXT_LOAD";
20069 case X86ISD::VZEXT: return "X86ISD::VZEXT";
20070 case X86ISD::VSEXT: return "X86ISD::VSEXT";
20071 case X86ISD::VTRUNC: return "X86ISD::VTRUNC";
20072 case X86ISD::VTRUNCS: return "X86ISD::VTRUNCS";
20073 case X86ISD::VTRUNCUS: return "X86ISD::VTRUNCUS";
20074 case X86ISD::VINSERT: return "X86ISD::VINSERT";
20075 case X86ISD::VFPEXT: return "X86ISD::VFPEXT";
20076 case X86ISD::VFPROUND: return "X86ISD::VFPROUND";
20077 case X86ISD::CVTDQ2PD: return "X86ISD::CVTDQ2PD";
20078 case X86ISD::CVTUDQ2PD: return "X86ISD::CVTUDQ2PD";
20079 case X86ISD::VSHLDQ: return "X86ISD::VSHLDQ";
20080 case X86ISD::VSRLDQ: return "X86ISD::VSRLDQ";
20081 case X86ISD::VSHL: return "X86ISD::VSHL";
20082 case X86ISD::VSRL: return "X86ISD::VSRL";
20083 case X86ISD::VSRA: return "X86ISD::VSRA";
20084 case X86ISD::VSHLI: return "X86ISD::VSHLI";
20085 case X86ISD::VSRLI: return "X86ISD::VSRLI";
20086 case X86ISD::VSRAI: return "X86ISD::VSRAI";
20087 case X86ISD::CMPP: return "X86ISD::CMPP";
20088 case X86ISD::PCMPEQ: return "X86ISD::PCMPEQ";
20089 case X86ISD::PCMPGT: return "X86ISD::PCMPGT";
20090 case X86ISD::PCMPEQM: return "X86ISD::PCMPEQM";
20091 case X86ISD::PCMPGTM: return "X86ISD::PCMPGTM";
20092 case X86ISD::ADD: return "X86ISD::ADD";
20093 case X86ISD::SUB: return "X86ISD::SUB";
20094 case X86ISD::ADC: return "X86ISD::ADC";
20095 case X86ISD::SBB: return "X86ISD::SBB";
20096 case X86ISD::SMUL: return "X86ISD::SMUL";
20097 case X86ISD::UMUL: return "X86ISD::UMUL";
20098 case X86ISD::SMUL8: return "X86ISD::SMUL8";
20099 case X86ISD::UMUL8: return "X86ISD::UMUL8";
20100 case X86ISD::SDIVREM8_SEXT_HREG: return "X86ISD::SDIVREM8_SEXT_HREG";
20101 case X86ISD::UDIVREM8_ZEXT_HREG: return "X86ISD::UDIVREM8_ZEXT_HREG";
20102 case X86ISD::INC: return "X86ISD::INC";
20103 case X86ISD::DEC: return "X86ISD::DEC";
20104 case X86ISD::OR: return "X86ISD::OR";
20105 case X86ISD::XOR: return "X86ISD::XOR";
20106 case X86ISD::AND: return "X86ISD::AND";
20107 case X86ISD::BEXTR: return "X86ISD::BEXTR";
20108 case X86ISD::MUL_IMM: return "X86ISD::MUL_IMM";
20109 case X86ISD::PTEST: return "X86ISD::PTEST";
20110 case X86ISD::TESTP: return "X86ISD::TESTP";
20111 case X86ISD::TESTM: return "X86ISD::TESTM";
20112 case X86ISD::TESTNM: return "X86ISD::TESTNM";
20113 case X86ISD::KORTEST: return "X86ISD::KORTEST";
20114 case X86ISD::KTEST: return "X86ISD::KTEST";
20115 case X86ISD::PACKSS: return "X86ISD::PACKSS";
20116 case X86ISD::PACKUS: return "X86ISD::PACKUS";
20117 case X86ISD::PALIGNR: return "X86ISD::PALIGNR";
20118 case X86ISD::VALIGN: return "X86ISD::VALIGN";
20119 case X86ISD::PSHUFD: return "X86ISD::PSHUFD";
20120 case X86ISD::PSHUFHW: return "X86ISD::PSHUFHW";
20121 case X86ISD::PSHUFLW: return "X86ISD::PSHUFLW";
20122 case X86ISD::SHUFP: return "X86ISD::SHUFP";
20123 case X86ISD::SHUF128: return "X86ISD::SHUF128";
20124 case X86ISD::MOVLHPS: return "X86ISD::MOVLHPS";
20125 case X86ISD::MOVLHPD: return "X86ISD::MOVLHPD";
20126 case X86ISD::MOVHLPS: return "X86ISD::MOVHLPS";
20127 case X86ISD::MOVLPS: return "X86ISD::MOVLPS";
20128 case X86ISD::MOVLPD: return "X86ISD::MOVLPD";
20129 case X86ISD::MOVDDUP: return "X86ISD::MOVDDUP";
20130 case X86ISD::MOVSHDUP: return "X86ISD::MOVSHDUP";
20131 case X86ISD::MOVSLDUP: return "X86ISD::MOVSLDUP";
20132 case X86ISD::MOVSD: return "X86ISD::MOVSD";
20133 case X86ISD::MOVSS: return "X86ISD::MOVSS";
20134 case X86ISD::UNPCKL: return "X86ISD::UNPCKL";
20135 case X86ISD::UNPCKH: return "X86ISD::UNPCKH";
20136 case X86ISD::VBROADCAST: return "X86ISD::VBROADCAST";
20137 case X86ISD::SUBV_BROADCAST: return "X86ISD::SUBV_BROADCAST";
20138 case X86ISD::VEXTRACT: return "X86ISD::VEXTRACT";
20139 case X86ISD::VPERMILPV: return "X86ISD::VPERMILPV";
20140 case X86ISD::VPERMILPI: return "X86ISD::VPERMILPI";
20141 case X86ISD::VPERM2X128: return "X86ISD::VPERM2X128";
20142 case X86ISD::VPERMV: return "X86ISD::VPERMV";
20143 case X86ISD::VPERMV3: return "X86ISD::VPERMV3";
20144 case X86ISD::VPERMIV3: return "X86ISD::VPERMIV3";
20145 case X86ISD::VPERMI: return "X86ISD::VPERMI";
20146 case X86ISD::VPTERNLOG: return "X86ISD::VPTERNLOG";
20147 case X86ISD::VFIXUPIMM: return "X86ISD::VFIXUPIMM";
20148 case X86ISD::VRANGE: return "X86ISD::VRANGE";
20149 case X86ISD::PMULUDQ: return "X86ISD::PMULUDQ";
20150 case X86ISD::PMULDQ: return "X86ISD::PMULDQ";
20151 case X86ISD::PSADBW: return "X86ISD::PSADBW";
20152 case X86ISD::DBPSADBW: return "X86ISD::DBPSADBW";
20153 case X86ISD::VASTART_SAVE_XMM_REGS: return "X86ISD::VASTART_SAVE_XMM_REGS";
20154 case X86ISD::VAARG_64: return "X86ISD::VAARG_64";
20155 case X86ISD::WIN_ALLOCA: return "X86ISD::WIN_ALLOCA";
20156 case X86ISD::MEMBARRIER: return "X86ISD::MEMBARRIER";
20157 case X86ISD::MFENCE: return "X86ISD::MFENCE";
20158 case X86ISD::SFENCE: return "X86ISD::SFENCE";
20159 case X86ISD::LFENCE: return "X86ISD::LFENCE";
20160 case X86ISD::SEG_ALLOCA: return "X86ISD::SEG_ALLOCA";
20161 case X86ISD::SAHF: return "X86ISD::SAHF";
20162 case X86ISD::RDRAND: return "X86ISD::RDRAND";
20163 case X86ISD::RDSEED: return "X86ISD::RDSEED";
20164 case X86ISD::VPMADDUBSW: return "X86ISD::VPMADDUBSW";
20165 case X86ISD::VPMADDWD: return "X86ISD::VPMADDWD";
20166 case X86ISD::VPROT: return "X86ISD::VPROT";
20167 case X86ISD::VPROTI: return "X86ISD::VPROTI";
20168 case X86ISD::VPSHA: return "X86ISD::VPSHA";
20169 case X86ISD::VPSHL: return "X86ISD::VPSHL";
20170 case X86ISD::VPCOM: return "X86ISD::VPCOM";
20171 case X86ISD::VPCOMU: return "X86ISD::VPCOMU";
20172 case X86ISD::FMADD: return "X86ISD::FMADD";
20173 case X86ISD::FMSUB: return "X86ISD::FMSUB";
20174 case X86ISD::FNMADD: return "X86ISD::FNMADD";
20175 case X86ISD::FNMSUB: return "X86ISD::FNMSUB";
20176 case X86ISD::FMADDSUB: return "X86ISD::FMADDSUB";
20177 case X86ISD::FMSUBADD: return "X86ISD::FMSUBADD";
20178 case X86ISD::FMADD_RND: return "X86ISD::FMADD_RND";
20179 case X86ISD::FNMADD_RND: return "X86ISD::FNMADD_RND";
20180 case X86ISD::FMSUB_RND: return "X86ISD::FMSUB_RND";
20181 case X86ISD::FNMSUB_RND: return "X86ISD::FNMSUB_RND";
20182 case X86ISD::FMADDSUB_RND: return "X86ISD::FMADDSUB_RND";
20183 case X86ISD::FMSUBADD_RND: return "X86ISD::FMSUBADD_RND";
20184 case X86ISD::VRNDSCALE: return "X86ISD::VRNDSCALE";
20185 case X86ISD::VREDUCE: return "X86ISD::VREDUCE";
20186 case X86ISD::VGETMANT: return "X86ISD::VGETMANT";
20187 case X86ISD::PCMPESTRI: return "X86ISD::PCMPESTRI";
20188 case X86ISD::PCMPISTRI: return "X86ISD::PCMPISTRI";
20189 case X86ISD::XTEST: return "X86ISD::XTEST";
20190 case X86ISD::COMPRESS: return "X86ISD::COMPRESS";
20191 case X86ISD::EXPAND: return "X86ISD::EXPAND";
20192 case X86ISD::SELECT: return "X86ISD::SELECT";
20193 case X86ISD::ADDSUB: return "X86ISD::ADDSUB";
20194 case X86ISD::RCP28: return "X86ISD::RCP28";
20195 case X86ISD::EXP2: return "X86ISD::EXP2";
20196 case X86ISD::RSQRT28: return "X86ISD::RSQRT28";
20197 case X86ISD::FADD_RND: return "X86ISD::FADD_RND";
20198 case X86ISD::FSUB_RND: return "X86ISD::FSUB_RND";
20199 case X86ISD::FMUL_RND: return "X86ISD::FMUL_RND";
20200 case X86ISD::FDIV_RND: return "X86ISD::FDIV_RND";
20201 case X86ISD::FSQRT_RND: return "X86ISD::FSQRT_RND";
20202 case X86ISD::FGETEXP_RND: return "X86ISD::FGETEXP_RND";
20203 case X86ISD::SCALEF: return "X86ISD::SCALEF";
20204 case X86ISD::ADDS: return "X86ISD::ADDS";
20205 case X86ISD::SUBS: return "X86ISD::SUBS";
20206 case X86ISD::AVG: return "X86ISD::AVG";
20207 case X86ISD::MULHRS: return "X86ISD::MULHRS";
20208 case X86ISD::SINT_TO_FP_RND: return "X86ISD::SINT_TO_FP_RND";
20209 case X86ISD::UINT_TO_FP_RND: return "X86ISD::UINT_TO_FP_RND";
20210 case X86ISD::FP_TO_SINT_RND: return "X86ISD::FP_TO_SINT_RND";
20211 case X86ISD::FP_TO_UINT_RND: return "X86ISD::FP_TO_UINT_RND";
20212 case X86ISD::VFPCLASS: return "X86ISD::VFPCLASS";
20217 // isLegalAddressingMode - Return true if the addressing mode represented
20218 // by AM is legal for this target, for a load/store of the specified type.
20219 bool X86TargetLowering::isLegalAddressingMode(const DataLayout &DL,
20220 const AddrMode &AM, Type *Ty,
20221 unsigned AS) const {
20222 // X86 supports extremely general addressing modes.
20223 CodeModel::Model M = getTargetMachine().getCodeModel();
20224 Reloc::Model R = getTargetMachine().getRelocationModel();
20226 // X86 allows a sign-extended 32-bit immediate field as a displacement.
20227 if (!X86::isOffsetSuitableForCodeModel(AM.BaseOffs, M, AM.BaseGV != nullptr))
20232 Subtarget->ClassifyGlobalReference(AM.BaseGV, getTargetMachine());
20234 // If a reference to this global requires an extra load, we can't fold it.
20235 if (isGlobalStubReference(GVFlags))
20238 // If BaseGV requires a register for the PIC base, we cannot also have a
20239 // BaseReg specified.
20240 if (AM.HasBaseReg && isGlobalRelativeToPICBase(GVFlags))
20243 // If lower 4G is not available, then we must use rip-relative addressing.
20244 if ((M != CodeModel::Small || R != Reloc::Static) &&
20245 Subtarget->is64Bit() && (AM.BaseOffs || AM.Scale > 1))
20249 switch (AM.Scale) {
20255 // These scales always work.
20260 // These scales are formed with basereg+scalereg. Only accept if there is
20265 default: // Other stuff never works.
20272 bool X86TargetLowering::isVectorShiftByScalarCheap(Type *Ty) const {
20273 unsigned Bits = Ty->getScalarSizeInBits();
20275 // 8-bit shifts are always expensive, but versions with a scalar amount aren't
20276 // particularly cheaper than those without.
20280 // On AVX2 there are new vpsllv[dq] instructions (and other shifts), that make
20281 // variable shifts just as cheap as scalar ones.
20282 if (Subtarget->hasInt256() && (Bits == 32 || Bits == 64))
20285 // Otherwise, it's significantly cheaper to shift by a scalar amount than by a
20286 // fully general vector.
20290 bool X86TargetLowering::isTruncateFree(Type *Ty1, Type *Ty2) const {
20291 if (!Ty1->isIntegerTy() || !Ty2->isIntegerTy())
20293 unsigned NumBits1 = Ty1->getPrimitiveSizeInBits();
20294 unsigned NumBits2 = Ty2->getPrimitiveSizeInBits();
20295 return NumBits1 > NumBits2;
20298 bool X86TargetLowering::allowTruncateForTailCall(Type *Ty1, Type *Ty2) const {
20299 if (!Ty1->isIntegerTy() || !Ty2->isIntegerTy())
20302 if (!isTypeLegal(EVT::getEVT(Ty1)))
20305 assert(Ty1->getPrimitiveSizeInBits() <= 64 && "i128 is probably not a noop");
20307 // Assuming the caller doesn't have a zeroext or signext return parameter,
20308 // truncation all the way down to i1 is valid.
20312 bool X86TargetLowering::isLegalICmpImmediate(int64_t Imm) const {
20313 return isInt<32>(Imm);
20316 bool X86TargetLowering::isLegalAddImmediate(int64_t Imm) const {
20317 // Can also use sub to handle negated immediates.
20318 return isInt<32>(Imm);
20321 bool X86TargetLowering::isTruncateFree(EVT VT1, EVT VT2) const {
20322 if (!VT1.isInteger() || !VT2.isInteger())
20324 unsigned NumBits1 = VT1.getSizeInBits();
20325 unsigned NumBits2 = VT2.getSizeInBits();
20326 return NumBits1 > NumBits2;
20329 bool X86TargetLowering::isZExtFree(Type *Ty1, Type *Ty2) const {
20330 // x86-64 implicitly zero-extends 32-bit results in 64-bit registers.
20331 return Ty1->isIntegerTy(32) && Ty2->isIntegerTy(64) && Subtarget->is64Bit();
20334 bool X86TargetLowering::isZExtFree(EVT VT1, EVT VT2) const {
20335 // x86-64 implicitly zero-extends 32-bit results in 64-bit registers.
20336 return VT1 == MVT::i32 && VT2 == MVT::i64 && Subtarget->is64Bit();
20339 bool X86TargetLowering::isZExtFree(SDValue Val, EVT VT2) const {
20340 EVT VT1 = Val.getValueType();
20341 if (isZExtFree(VT1, VT2))
20344 if (Val.getOpcode() != ISD::LOAD)
20347 if (!VT1.isSimple() || !VT1.isInteger() ||
20348 !VT2.isSimple() || !VT2.isInteger())
20351 switch (VT1.getSimpleVT().SimpleTy) {
20356 // X86 has 8, 16, and 32-bit zero-extending loads.
20363 bool X86TargetLowering::isVectorLoadExtDesirable(SDValue) const { return true; }
20366 X86TargetLowering::isFMAFasterThanFMulAndFAdd(EVT VT) const {
20367 if (!(Subtarget->hasFMA() || Subtarget->hasFMA4() || Subtarget->hasAVX512()))
20370 VT = VT.getScalarType();
20372 if (!VT.isSimple())
20375 switch (VT.getSimpleVT().SimpleTy) {
20386 bool X86TargetLowering::isNarrowingProfitable(EVT VT1, EVT VT2) const {
20387 // i16 instructions are longer (0x66 prefix) and potentially slower.
20388 return !(VT1 == MVT::i32 && VT2 == MVT::i16);
20391 /// isShuffleMaskLegal - Targets can use this to indicate that they only
20392 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
20393 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
20394 /// are assumed to be legal.
20396 X86TargetLowering::isShuffleMaskLegal(const SmallVectorImpl<int> &M,
20398 if (!VT.isSimple())
20401 // Not for i1 vectors
20402 if (VT.getScalarType() == MVT::i1)
20405 // Very little shuffling can be done for 64-bit vectors right now.
20406 if (VT.getSizeInBits() == 64)
20409 // We only care that the types being shuffled are legal. The lowering can
20410 // handle any possible shuffle mask that results.
20411 return isTypeLegal(VT.getSimpleVT());
20415 X86TargetLowering::isVectorClearMaskLegal(const SmallVectorImpl<int> &Mask,
20417 // Just delegate to the generic legality, clear masks aren't special.
20418 return isShuffleMaskLegal(Mask, VT);
20421 //===----------------------------------------------------------------------===//
20422 // X86 Scheduler Hooks
20423 //===----------------------------------------------------------------------===//
20425 /// Utility function to emit xbegin specifying the start of an RTM region.
20426 static MachineBasicBlock *EmitXBegin(MachineInstr *MI, MachineBasicBlock *MBB,
20427 const TargetInstrInfo *TII) {
20428 DebugLoc DL = MI->getDebugLoc();
20430 const BasicBlock *BB = MBB->getBasicBlock();
20431 MachineFunction::iterator I = ++MBB->getIterator();
20433 // For the v = xbegin(), we generate
20444 MachineBasicBlock *thisMBB = MBB;
20445 MachineFunction *MF = MBB->getParent();
20446 MachineBasicBlock *mainMBB = MF->CreateMachineBasicBlock(BB);
20447 MachineBasicBlock *sinkMBB = MF->CreateMachineBasicBlock(BB);
20448 MF->insert(I, mainMBB);
20449 MF->insert(I, sinkMBB);
20451 // Transfer the remainder of BB and its successor edges to sinkMBB.
20452 sinkMBB->splice(sinkMBB->begin(), MBB,
20453 std::next(MachineBasicBlock::iterator(MI)), MBB->end());
20454 sinkMBB->transferSuccessorsAndUpdatePHIs(MBB);
20458 // # fallthrough to mainMBB
20459 // # abortion to sinkMBB
20460 BuildMI(thisMBB, DL, TII->get(X86::XBEGIN_4)).addMBB(sinkMBB);
20461 thisMBB->addSuccessor(mainMBB);
20462 thisMBB->addSuccessor(sinkMBB);
20466 BuildMI(mainMBB, DL, TII->get(X86::MOV32ri), X86::EAX).addImm(-1);
20467 mainMBB->addSuccessor(sinkMBB);
20470 // EAX is live into the sinkMBB
20471 sinkMBB->addLiveIn(X86::EAX);
20472 BuildMI(*sinkMBB, sinkMBB->begin(), DL,
20473 TII->get(TargetOpcode::COPY), MI->getOperand(0).getReg())
20476 MI->eraseFromParent();
20480 // FIXME: When we get size specific XMM0 registers, i.e. XMM0_V16I8
20481 // or XMM0_V32I8 in AVX all of this code can be replaced with that
20482 // in the .td file.
20483 static MachineBasicBlock *EmitPCMPSTRM(MachineInstr *MI, MachineBasicBlock *BB,
20484 const TargetInstrInfo *TII) {
20486 switch (MI->getOpcode()) {
20487 default: llvm_unreachable("illegal opcode!");
20488 case X86::PCMPISTRM128REG: Opc = X86::PCMPISTRM128rr; break;
20489 case X86::VPCMPISTRM128REG: Opc = X86::VPCMPISTRM128rr; break;
20490 case X86::PCMPISTRM128MEM: Opc = X86::PCMPISTRM128rm; break;
20491 case X86::VPCMPISTRM128MEM: Opc = X86::VPCMPISTRM128rm; break;
20492 case X86::PCMPESTRM128REG: Opc = X86::PCMPESTRM128rr; break;
20493 case X86::VPCMPESTRM128REG: Opc = X86::VPCMPESTRM128rr; break;
20494 case X86::PCMPESTRM128MEM: Opc = X86::PCMPESTRM128rm; break;
20495 case X86::VPCMPESTRM128MEM: Opc = X86::VPCMPESTRM128rm; break;
20498 DebugLoc dl = MI->getDebugLoc();
20499 MachineInstrBuilder MIB = BuildMI(*BB, MI, dl, TII->get(Opc));
20501 unsigned NumArgs = MI->getNumOperands();
20502 for (unsigned i = 1; i < NumArgs; ++i) {
20503 MachineOperand &Op = MI->getOperand(i);
20504 if (!(Op.isReg() && Op.isImplicit()))
20505 MIB.addOperand(Op);
20507 if (MI->hasOneMemOperand())
20508 MIB->setMemRefs(MI->memoperands_begin(), MI->memoperands_end());
20510 BuildMI(*BB, MI, dl,
20511 TII->get(TargetOpcode::COPY), MI->getOperand(0).getReg())
20512 .addReg(X86::XMM0);
20514 MI->eraseFromParent();
20518 // FIXME: Custom handling because TableGen doesn't support multiple implicit
20519 // defs in an instruction pattern
20520 static MachineBasicBlock *EmitPCMPSTRI(MachineInstr *MI, MachineBasicBlock *BB,
20521 const TargetInstrInfo *TII) {
20523 switch (MI->getOpcode()) {
20524 default: llvm_unreachable("illegal opcode!");
20525 case X86::PCMPISTRIREG: Opc = X86::PCMPISTRIrr; break;
20526 case X86::VPCMPISTRIREG: Opc = X86::VPCMPISTRIrr; break;
20527 case X86::PCMPISTRIMEM: Opc = X86::PCMPISTRIrm; break;
20528 case X86::VPCMPISTRIMEM: Opc = X86::VPCMPISTRIrm; break;
20529 case X86::PCMPESTRIREG: Opc = X86::PCMPESTRIrr; break;
20530 case X86::VPCMPESTRIREG: Opc = X86::VPCMPESTRIrr; break;
20531 case X86::PCMPESTRIMEM: Opc = X86::PCMPESTRIrm; break;
20532 case X86::VPCMPESTRIMEM: Opc = X86::VPCMPESTRIrm; break;
20535 DebugLoc dl = MI->getDebugLoc();
20536 MachineInstrBuilder MIB = BuildMI(*BB, MI, dl, TII->get(Opc));
20538 unsigned NumArgs = MI->getNumOperands(); // remove the results
20539 for (unsigned i = 1; i < NumArgs; ++i) {
20540 MachineOperand &Op = MI->getOperand(i);
20541 if (!(Op.isReg() && Op.isImplicit()))
20542 MIB.addOperand(Op);
20544 if (MI->hasOneMemOperand())
20545 MIB->setMemRefs(MI->memoperands_begin(), MI->memoperands_end());
20547 BuildMI(*BB, MI, dl,
20548 TII->get(TargetOpcode::COPY), MI->getOperand(0).getReg())
20551 MI->eraseFromParent();
20555 static MachineBasicBlock *EmitMonitor(MachineInstr *MI, MachineBasicBlock *BB,
20556 const X86Subtarget *Subtarget) {
20557 DebugLoc dl = MI->getDebugLoc();
20558 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
20559 // Address into RAX/EAX, other two args into ECX, EDX.
20560 unsigned MemOpc = Subtarget->is64Bit() ? X86::LEA64r : X86::LEA32r;
20561 unsigned MemReg = Subtarget->is64Bit() ? X86::RAX : X86::EAX;
20562 MachineInstrBuilder MIB = BuildMI(*BB, MI, dl, TII->get(MemOpc), MemReg);
20563 for (int i = 0; i < X86::AddrNumOperands; ++i)
20564 MIB.addOperand(MI->getOperand(i));
20566 unsigned ValOps = X86::AddrNumOperands;
20567 BuildMI(*BB, MI, dl, TII->get(TargetOpcode::COPY), X86::ECX)
20568 .addReg(MI->getOperand(ValOps).getReg());
20569 BuildMI(*BB, MI, dl, TII->get(TargetOpcode::COPY), X86::EDX)
20570 .addReg(MI->getOperand(ValOps+1).getReg());
20572 // The instruction doesn't actually take any operands though.
20573 BuildMI(*BB, MI, dl, TII->get(X86::MONITORrrr));
20575 MI->eraseFromParent(); // The pseudo is gone now.
20579 MachineBasicBlock *
20580 X86TargetLowering::EmitVAARG64WithCustomInserter(MachineInstr *MI,
20581 MachineBasicBlock *MBB) const {
20582 // Emit va_arg instruction on X86-64.
20584 // Operands to this pseudo-instruction:
20585 // 0 ) Output : destination address (reg)
20586 // 1-5) Input : va_list address (addr, i64mem)
20587 // 6 ) ArgSize : Size (in bytes) of vararg type
20588 // 7 ) ArgMode : 0=overflow only, 1=use gp_offset, 2=use fp_offset
20589 // 8 ) Align : Alignment of type
20590 // 9 ) EFLAGS (implicit-def)
20592 assert(MI->getNumOperands() == 10 && "VAARG_64 should have 10 operands!");
20593 static_assert(X86::AddrNumOperands == 5,
20594 "VAARG_64 assumes 5 address operands");
20596 unsigned DestReg = MI->getOperand(0).getReg();
20597 MachineOperand &Base = MI->getOperand(1);
20598 MachineOperand &Scale = MI->getOperand(2);
20599 MachineOperand &Index = MI->getOperand(3);
20600 MachineOperand &Disp = MI->getOperand(4);
20601 MachineOperand &Segment = MI->getOperand(5);
20602 unsigned ArgSize = MI->getOperand(6).getImm();
20603 unsigned ArgMode = MI->getOperand(7).getImm();
20604 unsigned Align = MI->getOperand(8).getImm();
20606 // Memory Reference
20607 assert(MI->hasOneMemOperand() && "Expected VAARG_64 to have one memoperand");
20608 MachineInstr::mmo_iterator MMOBegin = MI->memoperands_begin();
20609 MachineInstr::mmo_iterator MMOEnd = MI->memoperands_end();
20611 // Machine Information
20612 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
20613 MachineRegisterInfo &MRI = MBB->getParent()->getRegInfo();
20614 const TargetRegisterClass *AddrRegClass = getRegClassFor(MVT::i64);
20615 const TargetRegisterClass *OffsetRegClass = getRegClassFor(MVT::i32);
20616 DebugLoc DL = MI->getDebugLoc();
20618 // struct va_list {
20621 // i64 overflow_area (address)
20622 // i64 reg_save_area (address)
20624 // sizeof(va_list) = 24
20625 // alignment(va_list) = 8
20627 unsigned TotalNumIntRegs = 6;
20628 unsigned TotalNumXMMRegs = 8;
20629 bool UseGPOffset = (ArgMode == 1);
20630 bool UseFPOffset = (ArgMode == 2);
20631 unsigned MaxOffset = TotalNumIntRegs * 8 +
20632 (UseFPOffset ? TotalNumXMMRegs * 16 : 0);
20634 /* Align ArgSize to a multiple of 8 */
20635 unsigned ArgSizeA8 = (ArgSize + 7) & ~7;
20636 bool NeedsAlign = (Align > 8);
20638 MachineBasicBlock *thisMBB = MBB;
20639 MachineBasicBlock *overflowMBB;
20640 MachineBasicBlock *offsetMBB;
20641 MachineBasicBlock *endMBB;
20643 unsigned OffsetDestReg = 0; // Argument address computed by offsetMBB
20644 unsigned OverflowDestReg = 0; // Argument address computed by overflowMBB
20645 unsigned OffsetReg = 0;
20647 if (!UseGPOffset && !UseFPOffset) {
20648 // If we only pull from the overflow region, we don't create a branch.
20649 // We don't need to alter control flow.
20650 OffsetDestReg = 0; // unused
20651 OverflowDestReg = DestReg;
20653 offsetMBB = nullptr;
20654 overflowMBB = thisMBB;
20657 // First emit code to check if gp_offset (or fp_offset) is below the bound.
20658 // If so, pull the argument from reg_save_area. (branch to offsetMBB)
20659 // If not, pull from overflow_area. (branch to overflowMBB)
20664 // offsetMBB overflowMBB
20669 // Registers for the PHI in endMBB
20670 OffsetDestReg = MRI.createVirtualRegister(AddrRegClass);
20671 OverflowDestReg = MRI.createVirtualRegister(AddrRegClass);
20673 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
20674 MachineFunction *MF = MBB->getParent();
20675 overflowMBB = MF->CreateMachineBasicBlock(LLVM_BB);
20676 offsetMBB = MF->CreateMachineBasicBlock(LLVM_BB);
20677 endMBB = MF->CreateMachineBasicBlock(LLVM_BB);
20679 MachineFunction::iterator MBBIter = ++MBB->getIterator();
20681 // Insert the new basic blocks
20682 MF->insert(MBBIter, offsetMBB);
20683 MF->insert(MBBIter, overflowMBB);
20684 MF->insert(MBBIter, endMBB);
20686 // Transfer the remainder of MBB and its successor edges to endMBB.
20687 endMBB->splice(endMBB->begin(), thisMBB,
20688 std::next(MachineBasicBlock::iterator(MI)), thisMBB->end());
20689 endMBB->transferSuccessorsAndUpdatePHIs(thisMBB);
20691 // Make offsetMBB and overflowMBB successors of thisMBB
20692 thisMBB->addSuccessor(offsetMBB);
20693 thisMBB->addSuccessor(overflowMBB);
20695 // endMBB is a successor of both offsetMBB and overflowMBB
20696 offsetMBB->addSuccessor(endMBB);
20697 overflowMBB->addSuccessor(endMBB);
20699 // Load the offset value into a register
20700 OffsetReg = MRI.createVirtualRegister(OffsetRegClass);
20701 BuildMI(thisMBB, DL, TII->get(X86::MOV32rm), OffsetReg)
20705 .addDisp(Disp, UseFPOffset ? 4 : 0)
20706 .addOperand(Segment)
20707 .setMemRefs(MMOBegin, MMOEnd);
20709 // Check if there is enough room left to pull this argument.
20710 BuildMI(thisMBB, DL, TII->get(X86::CMP32ri))
20712 .addImm(MaxOffset + 8 - ArgSizeA8);
20714 // Branch to "overflowMBB" if offset >= max
20715 // Fall through to "offsetMBB" otherwise
20716 BuildMI(thisMBB, DL, TII->get(X86::GetCondBranchFromCond(X86::COND_AE)))
20717 .addMBB(overflowMBB);
20720 // In offsetMBB, emit code to use the reg_save_area.
20722 assert(OffsetReg != 0);
20724 // Read the reg_save_area address.
20725 unsigned RegSaveReg = MRI.createVirtualRegister(AddrRegClass);
20726 BuildMI(offsetMBB, DL, TII->get(X86::MOV64rm), RegSaveReg)
20731 .addOperand(Segment)
20732 .setMemRefs(MMOBegin, MMOEnd);
20734 // Zero-extend the offset
20735 unsigned OffsetReg64 = MRI.createVirtualRegister(AddrRegClass);
20736 BuildMI(offsetMBB, DL, TII->get(X86::SUBREG_TO_REG), OffsetReg64)
20739 .addImm(X86::sub_32bit);
20741 // Add the offset to the reg_save_area to get the final address.
20742 BuildMI(offsetMBB, DL, TII->get(X86::ADD64rr), OffsetDestReg)
20743 .addReg(OffsetReg64)
20744 .addReg(RegSaveReg);
20746 // Compute the offset for the next argument
20747 unsigned NextOffsetReg = MRI.createVirtualRegister(OffsetRegClass);
20748 BuildMI(offsetMBB, DL, TII->get(X86::ADD32ri), NextOffsetReg)
20750 .addImm(UseFPOffset ? 16 : 8);
20752 // Store it back into the va_list.
20753 BuildMI(offsetMBB, DL, TII->get(X86::MOV32mr))
20757 .addDisp(Disp, UseFPOffset ? 4 : 0)
20758 .addOperand(Segment)
20759 .addReg(NextOffsetReg)
20760 .setMemRefs(MMOBegin, MMOEnd);
20763 BuildMI(offsetMBB, DL, TII->get(X86::JMP_1))
20768 // Emit code to use overflow area
20771 // Load the overflow_area address into a register.
20772 unsigned OverflowAddrReg = MRI.createVirtualRegister(AddrRegClass);
20773 BuildMI(overflowMBB, DL, TII->get(X86::MOV64rm), OverflowAddrReg)
20778 .addOperand(Segment)
20779 .setMemRefs(MMOBegin, MMOEnd);
20781 // If we need to align it, do so. Otherwise, just copy the address
20782 // to OverflowDestReg.
20784 // Align the overflow address
20785 assert((Align & (Align-1)) == 0 && "Alignment must be a power of 2");
20786 unsigned TmpReg = MRI.createVirtualRegister(AddrRegClass);
20788 // aligned_addr = (addr + (align-1)) & ~(align-1)
20789 BuildMI(overflowMBB, DL, TII->get(X86::ADD64ri32), TmpReg)
20790 .addReg(OverflowAddrReg)
20793 BuildMI(overflowMBB, DL, TII->get(X86::AND64ri32), OverflowDestReg)
20795 .addImm(~(uint64_t)(Align-1));
20797 BuildMI(overflowMBB, DL, TII->get(TargetOpcode::COPY), OverflowDestReg)
20798 .addReg(OverflowAddrReg);
20801 // Compute the next overflow address after this argument.
20802 // (the overflow address should be kept 8-byte aligned)
20803 unsigned NextAddrReg = MRI.createVirtualRegister(AddrRegClass);
20804 BuildMI(overflowMBB, DL, TII->get(X86::ADD64ri32), NextAddrReg)
20805 .addReg(OverflowDestReg)
20806 .addImm(ArgSizeA8);
20808 // Store the new overflow address.
20809 BuildMI(overflowMBB, DL, TII->get(X86::MOV64mr))
20814 .addOperand(Segment)
20815 .addReg(NextAddrReg)
20816 .setMemRefs(MMOBegin, MMOEnd);
20818 // If we branched, emit the PHI to the front of endMBB.
20820 BuildMI(*endMBB, endMBB->begin(), DL,
20821 TII->get(X86::PHI), DestReg)
20822 .addReg(OffsetDestReg).addMBB(offsetMBB)
20823 .addReg(OverflowDestReg).addMBB(overflowMBB);
20826 // Erase the pseudo instruction
20827 MI->eraseFromParent();
20832 MachineBasicBlock *
20833 X86TargetLowering::EmitVAStartSaveXMMRegsWithCustomInserter(
20835 MachineBasicBlock *MBB) const {
20836 // Emit code to save XMM registers to the stack. The ABI says that the
20837 // number of registers to save is given in %al, so it's theoretically
20838 // possible to do an indirect jump trick to avoid saving all of them,
20839 // however this code takes a simpler approach and just executes all
20840 // of the stores if %al is non-zero. It's less code, and it's probably
20841 // easier on the hardware branch predictor, and stores aren't all that
20842 // expensive anyway.
20844 // Create the new basic blocks. One block contains all the XMM stores,
20845 // and one block is the final destination regardless of whether any
20846 // stores were performed.
20847 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
20848 MachineFunction *F = MBB->getParent();
20849 MachineFunction::iterator MBBIter = ++MBB->getIterator();
20850 MachineBasicBlock *XMMSaveMBB = F->CreateMachineBasicBlock(LLVM_BB);
20851 MachineBasicBlock *EndMBB = F->CreateMachineBasicBlock(LLVM_BB);
20852 F->insert(MBBIter, XMMSaveMBB);
20853 F->insert(MBBIter, EndMBB);
20855 // Transfer the remainder of MBB and its successor edges to EndMBB.
20856 EndMBB->splice(EndMBB->begin(), MBB,
20857 std::next(MachineBasicBlock::iterator(MI)), MBB->end());
20858 EndMBB->transferSuccessorsAndUpdatePHIs(MBB);
20860 // The original block will now fall through to the XMM save block.
20861 MBB->addSuccessor(XMMSaveMBB);
20862 // The XMMSaveMBB will fall through to the end block.
20863 XMMSaveMBB->addSuccessor(EndMBB);
20865 // Now add the instructions.
20866 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
20867 DebugLoc DL = MI->getDebugLoc();
20869 unsigned CountReg = MI->getOperand(0).getReg();
20870 int64_t RegSaveFrameIndex = MI->getOperand(1).getImm();
20871 int64_t VarArgsFPOffset = MI->getOperand(2).getImm();
20873 if (!Subtarget->isCallingConvWin64(F->getFunction()->getCallingConv())) {
20874 // If %al is 0, branch around the XMM save block.
20875 BuildMI(MBB, DL, TII->get(X86::TEST8rr)).addReg(CountReg).addReg(CountReg);
20876 BuildMI(MBB, DL, TII->get(X86::JE_1)).addMBB(EndMBB);
20877 MBB->addSuccessor(EndMBB);
20880 // Make sure the last operand is EFLAGS, which gets clobbered by the branch
20881 // that was just emitted, but clearly shouldn't be "saved".
20882 assert((MI->getNumOperands() <= 3 ||
20883 !MI->getOperand(MI->getNumOperands() - 1).isReg() ||
20884 MI->getOperand(MI->getNumOperands() - 1).getReg() == X86::EFLAGS)
20885 && "Expected last argument to be EFLAGS");
20886 unsigned MOVOpc = Subtarget->hasFp256() ? X86::VMOVAPSmr : X86::MOVAPSmr;
20887 // In the XMM save block, save all the XMM argument registers.
20888 for (int i = 3, e = MI->getNumOperands() - 1; i != e; ++i) {
20889 int64_t Offset = (i - 3) * 16 + VarArgsFPOffset;
20890 MachineMemOperand *MMO = F->getMachineMemOperand(
20891 MachinePointerInfo::getFixedStack(*F, RegSaveFrameIndex, Offset),
20892 MachineMemOperand::MOStore,
20893 /*Size=*/16, /*Align=*/16);
20894 BuildMI(XMMSaveMBB, DL, TII->get(MOVOpc))
20895 .addFrameIndex(RegSaveFrameIndex)
20896 .addImm(/*Scale=*/1)
20897 .addReg(/*IndexReg=*/0)
20898 .addImm(/*Disp=*/Offset)
20899 .addReg(/*Segment=*/0)
20900 .addReg(MI->getOperand(i).getReg())
20901 .addMemOperand(MMO);
20904 MI->eraseFromParent(); // The pseudo instruction is gone now.
20909 // The EFLAGS operand of SelectItr might be missing a kill marker
20910 // because there were multiple uses of EFLAGS, and ISel didn't know
20911 // which to mark. Figure out whether SelectItr should have had a
20912 // kill marker, and set it if it should. Returns the correct kill
20914 static bool checkAndUpdateEFLAGSKill(MachineBasicBlock::iterator SelectItr,
20915 MachineBasicBlock* BB,
20916 const TargetRegisterInfo* TRI) {
20917 // Scan forward through BB for a use/def of EFLAGS.
20918 MachineBasicBlock::iterator miI(std::next(SelectItr));
20919 for (MachineBasicBlock::iterator miE = BB->end(); miI != miE; ++miI) {
20920 const MachineInstr& mi = *miI;
20921 if (mi.readsRegister(X86::EFLAGS))
20923 if (mi.definesRegister(X86::EFLAGS))
20924 break; // Should have kill-flag - update below.
20927 // If we hit the end of the block, check whether EFLAGS is live into a
20929 if (miI == BB->end()) {
20930 for (MachineBasicBlock::succ_iterator sItr = BB->succ_begin(),
20931 sEnd = BB->succ_end();
20932 sItr != sEnd; ++sItr) {
20933 MachineBasicBlock* succ = *sItr;
20934 if (succ->isLiveIn(X86::EFLAGS))
20939 // We found a def, or hit the end of the basic block and EFLAGS wasn't live
20940 // out. SelectMI should have a kill flag on EFLAGS.
20941 SelectItr->addRegisterKilled(X86::EFLAGS, TRI);
20945 // Return true if it is OK for this CMOV pseudo-opcode to be cascaded
20946 // together with other CMOV pseudo-opcodes into a single basic-block with
20947 // conditional jump around it.
20948 static bool isCMOVPseudo(MachineInstr *MI) {
20949 switch (MI->getOpcode()) {
20950 case X86::CMOV_FR32:
20951 case X86::CMOV_FR64:
20952 case X86::CMOV_GR8:
20953 case X86::CMOV_GR16:
20954 case X86::CMOV_GR32:
20955 case X86::CMOV_RFP32:
20956 case X86::CMOV_RFP64:
20957 case X86::CMOV_RFP80:
20958 case X86::CMOV_V2F64:
20959 case X86::CMOV_V2I64:
20960 case X86::CMOV_V4F32:
20961 case X86::CMOV_V4F64:
20962 case X86::CMOV_V4I64:
20963 case X86::CMOV_V16F32:
20964 case X86::CMOV_V8F32:
20965 case X86::CMOV_V8F64:
20966 case X86::CMOV_V8I64:
20967 case X86::CMOV_V8I1:
20968 case X86::CMOV_V16I1:
20969 case X86::CMOV_V32I1:
20970 case X86::CMOV_V64I1:
20978 MachineBasicBlock *
20979 X86TargetLowering::EmitLoweredSelect(MachineInstr *MI,
20980 MachineBasicBlock *BB) const {
20981 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
20982 DebugLoc DL = MI->getDebugLoc();
20984 // To "insert" a SELECT_CC instruction, we actually have to insert the
20985 // diamond control-flow pattern. The incoming instruction knows the
20986 // destination vreg to set, the condition code register to branch on, the
20987 // true/false values to select between, and a branch opcode to use.
20988 const BasicBlock *LLVM_BB = BB->getBasicBlock();
20989 MachineFunction::iterator It = ++BB->getIterator();
20994 // cmpTY ccX, r1, r2
20996 // fallthrough --> copy0MBB
20997 MachineBasicBlock *thisMBB = BB;
20998 MachineFunction *F = BB->getParent();
21000 // This code lowers all pseudo-CMOV instructions. Generally it lowers these
21001 // as described above, by inserting a BB, and then making a PHI at the join
21002 // point to select the true and false operands of the CMOV in the PHI.
21004 // The code also handles two different cases of multiple CMOV opcodes
21008 // In this case, there are multiple CMOVs in a row, all which are based on
21009 // the same condition setting (or the exact opposite condition setting).
21010 // In this case we can lower all the CMOVs using a single inserted BB, and
21011 // then make a number of PHIs at the join point to model the CMOVs. The only
21012 // trickiness here, is that in a case like:
21014 // t2 = CMOV cond1 t1, f1
21015 // t3 = CMOV cond1 t2, f2
21017 // when rewriting this into PHIs, we have to perform some renaming on the
21018 // temps since you cannot have a PHI operand refer to a PHI result earlier
21019 // in the same block. The "simple" but wrong lowering would be:
21021 // t2 = PHI t1(BB1), f1(BB2)
21022 // t3 = PHI t2(BB1), f2(BB2)
21024 // but clearly t2 is not defined in BB1, so that is incorrect. The proper
21025 // renaming is to note that on the path through BB1, t2 is really just a
21026 // copy of t1, and do that renaming, properly generating:
21028 // t2 = PHI t1(BB1), f1(BB2)
21029 // t3 = PHI t1(BB1), f2(BB2)
21031 // Case 2, we lower cascaded CMOVs such as
21033 // (CMOV (CMOV F, T, cc1), T, cc2)
21035 // to two successives branches. For that, we look for another CMOV as the
21036 // following instruction.
21038 // Without this, we would add a PHI between the two jumps, which ends up
21039 // creating a few copies all around. For instance, for
21041 // (sitofp (zext (fcmp une)))
21043 // we would generate:
21045 // ucomiss %xmm1, %xmm0
21046 // movss <1.0f>, %xmm0
21047 // movaps %xmm0, %xmm1
21049 // xorps %xmm1, %xmm1
21052 // movaps %xmm1, %xmm0
21056 // because this custom-inserter would have generated:
21068 // A: X = ...; Y = ...
21070 // C: Z = PHI [X, A], [Y, B]
21072 // E: PHI [X, C], [Z, D]
21074 // If we lower both CMOVs in a single step, we can instead generate:
21086 // A: X = ...; Y = ...
21088 // E: PHI [X, A], [X, C], [Y, D]
21090 // Which, in our sitofp/fcmp example, gives us something like:
21092 // ucomiss %xmm1, %xmm0
21093 // movss <1.0f>, %xmm0
21096 // xorps %xmm0, %xmm0
21100 MachineInstr *CascadedCMOV = nullptr;
21101 MachineInstr *LastCMOV = MI;
21102 X86::CondCode CC = X86::CondCode(MI->getOperand(3).getImm());
21103 X86::CondCode OppCC = X86::GetOppositeBranchCondition(CC);
21104 MachineBasicBlock::iterator NextMIIt =
21105 std::next(MachineBasicBlock::iterator(MI));
21107 // Check for case 1, where there are multiple CMOVs with the same condition
21108 // first. Of the two cases of multiple CMOV lowerings, case 1 reduces the
21109 // number of jumps the most.
21111 if (isCMOVPseudo(MI)) {
21112 // See if we have a string of CMOVS with the same condition.
21113 while (NextMIIt != BB->end() &&
21114 isCMOVPseudo(NextMIIt) &&
21115 (NextMIIt->getOperand(3).getImm() == CC ||
21116 NextMIIt->getOperand(3).getImm() == OppCC)) {
21117 LastCMOV = &*NextMIIt;
21122 // This checks for case 2, but only do this if we didn't already find
21123 // case 1, as indicated by LastCMOV == MI.
21124 if (LastCMOV == MI &&
21125 NextMIIt != BB->end() && NextMIIt->getOpcode() == MI->getOpcode() &&
21126 NextMIIt->getOperand(2).getReg() == MI->getOperand(2).getReg() &&
21127 NextMIIt->getOperand(1).getReg() == MI->getOperand(0).getReg()) {
21128 CascadedCMOV = &*NextMIIt;
21131 MachineBasicBlock *jcc1MBB = nullptr;
21133 // If we have a cascaded CMOV, we lower it to two successive branches to
21134 // the same block. EFLAGS is used by both, so mark it as live in the second.
21135 if (CascadedCMOV) {
21136 jcc1MBB = F->CreateMachineBasicBlock(LLVM_BB);
21137 F->insert(It, jcc1MBB);
21138 jcc1MBB->addLiveIn(X86::EFLAGS);
21141 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
21142 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
21143 F->insert(It, copy0MBB);
21144 F->insert(It, sinkMBB);
21146 // If the EFLAGS register isn't dead in the terminator, then claim that it's
21147 // live into the sink and copy blocks.
21148 const TargetRegisterInfo *TRI = Subtarget->getRegisterInfo();
21150 MachineInstr *LastEFLAGSUser = CascadedCMOV ? CascadedCMOV : LastCMOV;
21151 if (!LastEFLAGSUser->killsRegister(X86::EFLAGS) &&
21152 !checkAndUpdateEFLAGSKill(LastEFLAGSUser, BB, TRI)) {
21153 copy0MBB->addLiveIn(X86::EFLAGS);
21154 sinkMBB->addLiveIn(X86::EFLAGS);
21157 // Transfer the remainder of BB and its successor edges to sinkMBB.
21158 sinkMBB->splice(sinkMBB->begin(), BB,
21159 std::next(MachineBasicBlock::iterator(LastCMOV)), BB->end());
21160 sinkMBB->transferSuccessorsAndUpdatePHIs(BB);
21162 // Add the true and fallthrough blocks as its successors.
21163 if (CascadedCMOV) {
21164 // The fallthrough block may be jcc1MBB, if we have a cascaded CMOV.
21165 BB->addSuccessor(jcc1MBB);
21167 // In that case, jcc1MBB will itself fallthrough the copy0MBB, and
21168 // jump to the sinkMBB.
21169 jcc1MBB->addSuccessor(copy0MBB);
21170 jcc1MBB->addSuccessor(sinkMBB);
21172 BB->addSuccessor(copy0MBB);
21175 // The true block target of the first (or only) branch is always sinkMBB.
21176 BB->addSuccessor(sinkMBB);
21178 // Create the conditional branch instruction.
21179 unsigned Opc = X86::GetCondBranchFromCond(CC);
21180 BuildMI(BB, DL, TII->get(Opc)).addMBB(sinkMBB);
21182 if (CascadedCMOV) {
21183 unsigned Opc2 = X86::GetCondBranchFromCond(
21184 (X86::CondCode)CascadedCMOV->getOperand(3).getImm());
21185 BuildMI(jcc1MBB, DL, TII->get(Opc2)).addMBB(sinkMBB);
21189 // %FalseValue = ...
21190 // # fallthrough to sinkMBB
21191 copy0MBB->addSuccessor(sinkMBB);
21194 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
21196 MachineBasicBlock::iterator MIItBegin = MachineBasicBlock::iterator(MI);
21197 MachineBasicBlock::iterator MIItEnd =
21198 std::next(MachineBasicBlock::iterator(LastCMOV));
21199 MachineBasicBlock::iterator SinkInsertionPoint = sinkMBB->begin();
21200 DenseMap<unsigned, std::pair<unsigned, unsigned>> RegRewriteTable;
21201 MachineInstrBuilder MIB;
21203 // As we are creating the PHIs, we have to be careful if there is more than
21204 // one. Later CMOVs may reference the results of earlier CMOVs, but later
21205 // PHIs have to reference the individual true/false inputs from earlier PHIs.
21206 // That also means that PHI construction must work forward from earlier to
21207 // later, and that the code must maintain a mapping from earlier PHI's
21208 // destination registers, and the registers that went into the PHI.
21210 for (MachineBasicBlock::iterator MIIt = MIItBegin; MIIt != MIItEnd; ++MIIt) {
21211 unsigned DestReg = MIIt->getOperand(0).getReg();
21212 unsigned Op1Reg = MIIt->getOperand(1).getReg();
21213 unsigned Op2Reg = MIIt->getOperand(2).getReg();
21215 // If this CMOV we are generating is the opposite condition from
21216 // the jump we generated, then we have to swap the operands for the
21217 // PHI that is going to be generated.
21218 if (MIIt->getOperand(3).getImm() == OppCC)
21219 std::swap(Op1Reg, Op2Reg);
21221 if (RegRewriteTable.find(Op1Reg) != RegRewriteTable.end())
21222 Op1Reg = RegRewriteTable[Op1Reg].first;
21224 if (RegRewriteTable.find(Op2Reg) != RegRewriteTable.end())
21225 Op2Reg = RegRewriteTable[Op2Reg].second;
21227 MIB = BuildMI(*sinkMBB, SinkInsertionPoint, DL,
21228 TII->get(X86::PHI), DestReg)
21229 .addReg(Op1Reg).addMBB(copy0MBB)
21230 .addReg(Op2Reg).addMBB(thisMBB);
21232 // Add this PHI to the rewrite table.
21233 RegRewriteTable[DestReg] = std::make_pair(Op1Reg, Op2Reg);
21236 // If we have a cascaded CMOV, the second Jcc provides the same incoming
21237 // value as the first Jcc (the True operand of the SELECT_CC/CMOV nodes).
21238 if (CascadedCMOV) {
21239 MIB.addReg(MI->getOperand(2).getReg()).addMBB(jcc1MBB);
21240 // Copy the PHI result to the register defined by the second CMOV.
21241 BuildMI(*sinkMBB, std::next(MachineBasicBlock::iterator(MIB.getInstr())),
21242 DL, TII->get(TargetOpcode::COPY),
21243 CascadedCMOV->getOperand(0).getReg())
21244 .addReg(MI->getOperand(0).getReg());
21245 CascadedCMOV->eraseFromParent();
21248 // Now remove the CMOV(s).
21249 for (MachineBasicBlock::iterator MIIt = MIItBegin; MIIt != MIItEnd; )
21250 (MIIt++)->eraseFromParent();
21255 MachineBasicBlock *
21256 X86TargetLowering::EmitLoweredAtomicFP(MachineInstr *MI,
21257 MachineBasicBlock *BB) const {
21258 // Combine the following atomic floating-point modification pattern:
21259 // a.store(reg OP a.load(acquire), release)
21260 // Transform them into:
21261 // OPss (%gpr), %xmm
21262 // movss %xmm, (%gpr)
21263 // Or sd equivalent for 64-bit operations.
21265 switch (MI->getOpcode()) {
21266 default: llvm_unreachable("unexpected instr type for EmitLoweredAtomicFP");
21267 case X86::RELEASE_FADD32mr: MOp = X86::MOVSSmr; FOp = X86::ADDSSrm; break;
21268 case X86::RELEASE_FADD64mr: MOp = X86::MOVSDmr; FOp = X86::ADDSDrm; break;
21270 const X86InstrInfo *TII = Subtarget->getInstrInfo();
21271 DebugLoc DL = MI->getDebugLoc();
21272 MachineRegisterInfo &MRI = BB->getParent()->getRegInfo();
21273 MachineOperand MSrc = MI->getOperand(0);
21274 unsigned VSrc = MI->getOperand(5).getReg();
21275 const MachineOperand &Disp = MI->getOperand(3);
21276 MachineOperand ZeroDisp = MachineOperand::CreateImm(0);
21277 bool hasDisp = Disp.isGlobal() || Disp.isImm();
21278 if (hasDisp && MSrc.isReg())
21279 MSrc.setIsKill(false);
21280 MachineInstrBuilder MIM = BuildMI(*BB, MI, DL, TII->get(MOp))
21281 .addOperand(/*Base=*/MSrc)
21282 .addImm(/*Scale=*/1)
21283 .addReg(/*Index=*/0)
21284 .addDisp(hasDisp ? Disp : ZeroDisp, /*off=*/0)
21286 MachineInstr *MIO = BuildMI(*BB, (MachineInstr *)MIM, DL, TII->get(FOp),
21287 MRI.createVirtualRegister(MRI.getRegClass(VSrc)))
21289 .addOperand(/*Base=*/MSrc)
21290 .addImm(/*Scale=*/1)
21291 .addReg(/*Index=*/0)
21292 .addDisp(hasDisp ? Disp : ZeroDisp, /*off=*/0)
21293 .addReg(/*Segment=*/0);
21294 MIM.addReg(MIO->getOperand(0).getReg(), RegState::Kill);
21295 MI->eraseFromParent(); // The pseudo instruction is gone now.
21299 MachineBasicBlock *
21300 X86TargetLowering::EmitLoweredSegAlloca(MachineInstr *MI,
21301 MachineBasicBlock *BB) const {
21302 MachineFunction *MF = BB->getParent();
21303 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
21304 DebugLoc DL = MI->getDebugLoc();
21305 const BasicBlock *LLVM_BB = BB->getBasicBlock();
21307 assert(MF->shouldSplitStack());
21309 const bool Is64Bit = Subtarget->is64Bit();
21310 const bool IsLP64 = Subtarget->isTarget64BitLP64();
21312 const unsigned TlsReg = Is64Bit ? X86::FS : X86::GS;
21313 const unsigned TlsOffset = IsLP64 ? 0x70 : Is64Bit ? 0x40 : 0x30;
21316 // ... [Till the alloca]
21317 // If stacklet is not large enough, jump to mallocMBB
21320 // Allocate by subtracting from RSP
21321 // Jump to continueMBB
21324 // Allocate by call to runtime
21328 // [rest of original BB]
21331 MachineBasicBlock *mallocMBB = MF->CreateMachineBasicBlock(LLVM_BB);
21332 MachineBasicBlock *bumpMBB = MF->CreateMachineBasicBlock(LLVM_BB);
21333 MachineBasicBlock *continueMBB = MF->CreateMachineBasicBlock(LLVM_BB);
21335 MachineRegisterInfo &MRI = MF->getRegInfo();
21336 const TargetRegisterClass *AddrRegClass =
21337 getRegClassFor(getPointerTy(MF->getDataLayout()));
21339 unsigned mallocPtrVReg = MRI.createVirtualRegister(AddrRegClass),
21340 bumpSPPtrVReg = MRI.createVirtualRegister(AddrRegClass),
21341 tmpSPVReg = MRI.createVirtualRegister(AddrRegClass),
21342 SPLimitVReg = MRI.createVirtualRegister(AddrRegClass),
21343 sizeVReg = MI->getOperand(1).getReg(),
21344 physSPReg = IsLP64 || Subtarget->isTargetNaCl64() ? X86::RSP : X86::ESP;
21346 MachineFunction::iterator MBBIter = ++BB->getIterator();
21348 MF->insert(MBBIter, bumpMBB);
21349 MF->insert(MBBIter, mallocMBB);
21350 MF->insert(MBBIter, continueMBB);
21352 continueMBB->splice(continueMBB->begin(), BB,
21353 std::next(MachineBasicBlock::iterator(MI)), BB->end());
21354 continueMBB->transferSuccessorsAndUpdatePHIs(BB);
21356 // Add code to the main basic block to check if the stack limit has been hit,
21357 // and if so, jump to mallocMBB otherwise to bumpMBB.
21358 BuildMI(BB, DL, TII->get(TargetOpcode::COPY), tmpSPVReg).addReg(physSPReg);
21359 BuildMI(BB, DL, TII->get(IsLP64 ? X86::SUB64rr:X86::SUB32rr), SPLimitVReg)
21360 .addReg(tmpSPVReg).addReg(sizeVReg);
21361 BuildMI(BB, DL, TII->get(IsLP64 ? X86::CMP64mr:X86::CMP32mr))
21362 .addReg(0).addImm(1).addReg(0).addImm(TlsOffset).addReg(TlsReg)
21363 .addReg(SPLimitVReg);
21364 BuildMI(BB, DL, TII->get(X86::JG_1)).addMBB(mallocMBB);
21366 // bumpMBB simply decreases the stack pointer, since we know the current
21367 // stacklet has enough space.
21368 BuildMI(bumpMBB, DL, TII->get(TargetOpcode::COPY), physSPReg)
21369 .addReg(SPLimitVReg);
21370 BuildMI(bumpMBB, DL, TII->get(TargetOpcode::COPY), bumpSPPtrVReg)
21371 .addReg(SPLimitVReg);
21372 BuildMI(bumpMBB, DL, TII->get(X86::JMP_1)).addMBB(continueMBB);
21374 // Calls into a routine in libgcc to allocate more space from the heap.
21375 const uint32_t *RegMask =
21376 Subtarget->getRegisterInfo()->getCallPreservedMask(*MF, CallingConv::C);
21378 BuildMI(mallocMBB, DL, TII->get(X86::MOV64rr), X86::RDI)
21380 BuildMI(mallocMBB, DL, TII->get(X86::CALL64pcrel32))
21381 .addExternalSymbol("__morestack_allocate_stack_space")
21382 .addRegMask(RegMask)
21383 .addReg(X86::RDI, RegState::Implicit)
21384 .addReg(X86::RAX, RegState::ImplicitDefine);
21385 } else if (Is64Bit) {
21386 BuildMI(mallocMBB, DL, TII->get(X86::MOV32rr), X86::EDI)
21388 BuildMI(mallocMBB, DL, TII->get(X86::CALL64pcrel32))
21389 .addExternalSymbol("__morestack_allocate_stack_space")
21390 .addRegMask(RegMask)
21391 .addReg(X86::EDI, RegState::Implicit)
21392 .addReg(X86::EAX, RegState::ImplicitDefine);
21394 BuildMI(mallocMBB, DL, TII->get(X86::SUB32ri), physSPReg).addReg(physSPReg)
21396 BuildMI(mallocMBB, DL, TII->get(X86::PUSH32r)).addReg(sizeVReg);
21397 BuildMI(mallocMBB, DL, TII->get(X86::CALLpcrel32))
21398 .addExternalSymbol("__morestack_allocate_stack_space")
21399 .addRegMask(RegMask)
21400 .addReg(X86::EAX, RegState::ImplicitDefine);
21404 BuildMI(mallocMBB, DL, TII->get(X86::ADD32ri), physSPReg).addReg(physSPReg)
21407 BuildMI(mallocMBB, DL, TII->get(TargetOpcode::COPY), mallocPtrVReg)
21408 .addReg(IsLP64 ? X86::RAX : X86::EAX);
21409 BuildMI(mallocMBB, DL, TII->get(X86::JMP_1)).addMBB(continueMBB);
21411 // Set up the CFG correctly.
21412 BB->addSuccessor(bumpMBB);
21413 BB->addSuccessor(mallocMBB);
21414 mallocMBB->addSuccessor(continueMBB);
21415 bumpMBB->addSuccessor(continueMBB);
21417 // Take care of the PHI nodes.
21418 BuildMI(*continueMBB, continueMBB->begin(), DL, TII->get(X86::PHI),
21419 MI->getOperand(0).getReg())
21420 .addReg(mallocPtrVReg).addMBB(mallocMBB)
21421 .addReg(bumpSPPtrVReg).addMBB(bumpMBB);
21423 // Delete the original pseudo instruction.
21424 MI->eraseFromParent();
21427 return continueMBB;
21430 MachineBasicBlock *
21431 X86TargetLowering::EmitLoweredWinAlloca(MachineInstr *MI,
21432 MachineBasicBlock *BB) const {
21433 DebugLoc DL = MI->getDebugLoc();
21435 assert(!Subtarget->isTargetMachO());
21437 Subtarget->getFrameLowering()->emitStackProbeCall(*BB->getParent(), *BB, MI,
21440 MI->eraseFromParent(); // The pseudo instruction is gone now.
21444 MachineBasicBlock *
21445 X86TargetLowering::EmitLoweredTLSCall(MachineInstr *MI,
21446 MachineBasicBlock *BB) const {
21447 // This is pretty easy. We're taking the value that we received from
21448 // our load from the relocation, sticking it in either RDI (x86-64)
21449 // or EAX and doing an indirect call. The return value will then
21450 // be in the normal return register.
21451 MachineFunction *F = BB->getParent();
21452 const X86InstrInfo *TII = Subtarget->getInstrInfo();
21453 DebugLoc DL = MI->getDebugLoc();
21455 assert(Subtarget->isTargetDarwin() && "Darwin only instr emitted?");
21456 assert(MI->getOperand(3).isGlobal() && "This should be a global");
21458 // Get a register mask for the lowered call.
21459 // FIXME: The 32-bit calls have non-standard calling conventions. Use a
21460 // proper register mask.
21461 const uint32_t *RegMask =
21462 Subtarget->getRegisterInfo()->getCallPreservedMask(*F, CallingConv::C);
21463 if (Subtarget->is64Bit()) {
21464 MachineInstrBuilder MIB = BuildMI(*BB, MI, DL,
21465 TII->get(X86::MOV64rm), X86::RDI)
21467 .addImm(0).addReg(0)
21468 .addGlobalAddress(MI->getOperand(3).getGlobal(), 0,
21469 MI->getOperand(3).getTargetFlags())
21471 MIB = BuildMI(*BB, MI, DL, TII->get(X86::CALL64m));
21472 addDirectMem(MIB, X86::RDI);
21473 MIB.addReg(X86::RAX, RegState::ImplicitDefine).addRegMask(RegMask);
21474 } else if (F->getTarget().getRelocationModel() != Reloc::PIC_) {
21475 MachineInstrBuilder MIB = BuildMI(*BB, MI, DL,
21476 TII->get(X86::MOV32rm), X86::EAX)
21478 .addImm(0).addReg(0)
21479 .addGlobalAddress(MI->getOperand(3).getGlobal(), 0,
21480 MI->getOperand(3).getTargetFlags())
21482 MIB = BuildMI(*BB, MI, DL, TII->get(X86::CALL32m));
21483 addDirectMem(MIB, X86::EAX);
21484 MIB.addReg(X86::EAX, RegState::ImplicitDefine).addRegMask(RegMask);
21486 MachineInstrBuilder MIB = BuildMI(*BB, MI, DL,
21487 TII->get(X86::MOV32rm), X86::EAX)
21488 .addReg(TII->getGlobalBaseReg(F))
21489 .addImm(0).addReg(0)
21490 .addGlobalAddress(MI->getOperand(3).getGlobal(), 0,
21491 MI->getOperand(3).getTargetFlags())
21493 MIB = BuildMI(*BB, MI, DL, TII->get(X86::CALL32m));
21494 addDirectMem(MIB, X86::EAX);
21495 MIB.addReg(X86::EAX, RegState::ImplicitDefine).addRegMask(RegMask);
21498 MI->eraseFromParent(); // The pseudo instruction is gone now.
21502 MachineBasicBlock *
21503 X86TargetLowering::emitEHSjLjSetJmp(MachineInstr *MI,
21504 MachineBasicBlock *MBB) const {
21505 DebugLoc DL = MI->getDebugLoc();
21506 MachineFunction *MF = MBB->getParent();
21507 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
21508 MachineRegisterInfo &MRI = MF->getRegInfo();
21510 const BasicBlock *BB = MBB->getBasicBlock();
21511 MachineFunction::iterator I = ++MBB->getIterator();
21513 // Memory Reference
21514 MachineInstr::mmo_iterator MMOBegin = MI->memoperands_begin();
21515 MachineInstr::mmo_iterator MMOEnd = MI->memoperands_end();
21518 unsigned MemOpndSlot = 0;
21520 unsigned CurOp = 0;
21522 DstReg = MI->getOperand(CurOp++).getReg();
21523 const TargetRegisterClass *RC = MRI.getRegClass(DstReg);
21524 assert(RC->hasType(MVT::i32) && "Invalid destination!");
21525 unsigned mainDstReg = MRI.createVirtualRegister(RC);
21526 unsigned restoreDstReg = MRI.createVirtualRegister(RC);
21528 MemOpndSlot = CurOp;
21530 MVT PVT = getPointerTy(MF->getDataLayout());
21531 assert((PVT == MVT::i64 || PVT == MVT::i32) &&
21532 "Invalid Pointer Size!");
21534 // For v = setjmp(buf), we generate
21537 // buf[LabelOffset] = restoreMBB <-- takes address of restoreMBB
21538 // SjLjSetup restoreMBB
21544 // v = phi(main, restore)
21547 // if base pointer being used, load it from frame
21550 MachineBasicBlock *thisMBB = MBB;
21551 MachineBasicBlock *mainMBB = MF->CreateMachineBasicBlock(BB);
21552 MachineBasicBlock *sinkMBB = MF->CreateMachineBasicBlock(BB);
21553 MachineBasicBlock *restoreMBB = MF->CreateMachineBasicBlock(BB);
21554 MF->insert(I, mainMBB);
21555 MF->insert(I, sinkMBB);
21556 MF->push_back(restoreMBB);
21557 restoreMBB->setHasAddressTaken();
21559 MachineInstrBuilder MIB;
21561 // Transfer the remainder of BB and its successor edges to sinkMBB.
21562 sinkMBB->splice(sinkMBB->begin(), MBB,
21563 std::next(MachineBasicBlock::iterator(MI)), MBB->end());
21564 sinkMBB->transferSuccessorsAndUpdatePHIs(MBB);
21567 unsigned PtrStoreOpc = 0;
21568 unsigned LabelReg = 0;
21569 const int64_t LabelOffset = 1 * PVT.getStoreSize();
21570 Reloc::Model RM = MF->getTarget().getRelocationModel();
21571 bool UseImmLabel = (MF->getTarget().getCodeModel() == CodeModel::Small) &&
21572 (RM == Reloc::Static || RM == Reloc::DynamicNoPIC);
21574 // Prepare IP either in reg or imm.
21575 if (!UseImmLabel) {
21576 PtrStoreOpc = (PVT == MVT::i64) ? X86::MOV64mr : X86::MOV32mr;
21577 const TargetRegisterClass *PtrRC = getRegClassFor(PVT);
21578 LabelReg = MRI.createVirtualRegister(PtrRC);
21579 if (Subtarget->is64Bit()) {
21580 MIB = BuildMI(*thisMBB, MI, DL, TII->get(X86::LEA64r), LabelReg)
21584 .addMBB(restoreMBB)
21587 const X86InstrInfo *XII = static_cast<const X86InstrInfo*>(TII);
21588 MIB = BuildMI(*thisMBB, MI, DL, TII->get(X86::LEA32r), LabelReg)
21589 .addReg(XII->getGlobalBaseReg(MF))
21592 .addMBB(restoreMBB, Subtarget->ClassifyBlockAddressReference())
21596 PtrStoreOpc = (PVT == MVT::i64) ? X86::MOV64mi32 : X86::MOV32mi;
21598 MIB = BuildMI(*thisMBB, MI, DL, TII->get(PtrStoreOpc));
21599 for (unsigned i = 0; i < X86::AddrNumOperands; ++i) {
21600 if (i == X86::AddrDisp)
21601 MIB.addDisp(MI->getOperand(MemOpndSlot + i), LabelOffset);
21603 MIB.addOperand(MI->getOperand(MemOpndSlot + i));
21606 MIB.addReg(LabelReg);
21608 MIB.addMBB(restoreMBB);
21609 MIB.setMemRefs(MMOBegin, MMOEnd);
21611 MIB = BuildMI(*thisMBB, MI, DL, TII->get(X86::EH_SjLj_Setup))
21612 .addMBB(restoreMBB);
21614 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
21615 MIB.addRegMask(RegInfo->getNoPreservedMask());
21616 thisMBB->addSuccessor(mainMBB);
21617 thisMBB->addSuccessor(restoreMBB);
21621 BuildMI(mainMBB, DL, TII->get(X86::MOV32r0), mainDstReg);
21622 mainMBB->addSuccessor(sinkMBB);
21625 BuildMI(*sinkMBB, sinkMBB->begin(), DL,
21626 TII->get(X86::PHI), DstReg)
21627 .addReg(mainDstReg).addMBB(mainMBB)
21628 .addReg(restoreDstReg).addMBB(restoreMBB);
21631 if (RegInfo->hasBasePointer(*MF)) {
21632 const bool Uses64BitFramePtr =
21633 Subtarget->isTarget64BitLP64() || Subtarget->isTargetNaCl64();
21634 X86MachineFunctionInfo *X86FI = MF->getInfo<X86MachineFunctionInfo>();
21635 X86FI->setRestoreBasePointer(MF);
21636 unsigned FramePtr = RegInfo->getFrameRegister(*MF);
21637 unsigned BasePtr = RegInfo->getBaseRegister();
21638 unsigned Opm = Uses64BitFramePtr ? X86::MOV64rm : X86::MOV32rm;
21639 addRegOffset(BuildMI(restoreMBB, DL, TII->get(Opm), BasePtr),
21640 FramePtr, true, X86FI->getRestoreBasePointerOffset())
21641 .setMIFlag(MachineInstr::FrameSetup);
21643 BuildMI(restoreMBB, DL, TII->get(X86::MOV32ri), restoreDstReg).addImm(1);
21644 BuildMI(restoreMBB, DL, TII->get(X86::JMP_1)).addMBB(sinkMBB);
21645 restoreMBB->addSuccessor(sinkMBB);
21647 MI->eraseFromParent();
21651 MachineBasicBlock *
21652 X86TargetLowering::emitEHSjLjLongJmp(MachineInstr *MI,
21653 MachineBasicBlock *MBB) const {
21654 DebugLoc DL = MI->getDebugLoc();
21655 MachineFunction *MF = MBB->getParent();
21656 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
21657 MachineRegisterInfo &MRI = MF->getRegInfo();
21659 // Memory Reference
21660 MachineInstr::mmo_iterator MMOBegin = MI->memoperands_begin();
21661 MachineInstr::mmo_iterator MMOEnd = MI->memoperands_end();
21663 MVT PVT = getPointerTy(MF->getDataLayout());
21664 assert((PVT == MVT::i64 || PVT == MVT::i32) &&
21665 "Invalid Pointer Size!");
21667 const TargetRegisterClass *RC =
21668 (PVT == MVT::i64) ? &X86::GR64RegClass : &X86::GR32RegClass;
21669 unsigned Tmp = MRI.createVirtualRegister(RC);
21670 // Since FP is only updated here but NOT referenced, it's treated as GPR.
21671 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
21672 unsigned FP = (PVT == MVT::i64) ? X86::RBP : X86::EBP;
21673 unsigned SP = RegInfo->getStackRegister();
21675 MachineInstrBuilder MIB;
21677 const int64_t LabelOffset = 1 * PVT.getStoreSize();
21678 const int64_t SPOffset = 2 * PVT.getStoreSize();
21680 unsigned PtrLoadOpc = (PVT == MVT::i64) ? X86::MOV64rm : X86::MOV32rm;
21681 unsigned IJmpOpc = (PVT == MVT::i64) ? X86::JMP64r : X86::JMP32r;
21684 MIB = BuildMI(*MBB, MI, DL, TII->get(PtrLoadOpc), FP);
21685 for (unsigned i = 0; i < X86::AddrNumOperands; ++i)
21686 MIB.addOperand(MI->getOperand(i));
21687 MIB.setMemRefs(MMOBegin, MMOEnd);
21689 MIB = BuildMI(*MBB, MI, DL, TII->get(PtrLoadOpc), Tmp);
21690 for (unsigned i = 0; i < X86::AddrNumOperands; ++i) {
21691 if (i == X86::AddrDisp)
21692 MIB.addDisp(MI->getOperand(i), LabelOffset);
21694 MIB.addOperand(MI->getOperand(i));
21696 MIB.setMemRefs(MMOBegin, MMOEnd);
21698 MIB = BuildMI(*MBB, MI, DL, TII->get(PtrLoadOpc), SP);
21699 for (unsigned i = 0; i < X86::AddrNumOperands; ++i) {
21700 if (i == X86::AddrDisp)
21701 MIB.addDisp(MI->getOperand(i), SPOffset);
21703 MIB.addOperand(MI->getOperand(i));
21705 MIB.setMemRefs(MMOBegin, MMOEnd);
21707 BuildMI(*MBB, MI, DL, TII->get(IJmpOpc)).addReg(Tmp);
21709 MI->eraseFromParent();
21713 // Replace 213-type (isel default) FMA3 instructions with 231-type for
21714 // accumulator loops. Writing back to the accumulator allows the coalescer
21715 // to remove extra copies in the loop.
21716 // FIXME: Do this on AVX512. We don't support 231 variants yet (PR23937).
21717 MachineBasicBlock *
21718 X86TargetLowering::emitFMA3Instr(MachineInstr *MI,
21719 MachineBasicBlock *MBB) const {
21720 MachineOperand &AddendOp = MI->getOperand(3);
21722 // Bail out early if the addend isn't a register - we can't switch these.
21723 if (!AddendOp.isReg())
21726 MachineFunction &MF = *MBB->getParent();
21727 MachineRegisterInfo &MRI = MF.getRegInfo();
21729 // Check whether the addend is defined by a PHI:
21730 assert(MRI.hasOneDef(AddendOp.getReg()) && "Multiple defs in SSA?");
21731 MachineInstr &AddendDef = *MRI.def_instr_begin(AddendOp.getReg());
21732 if (!AddendDef.isPHI())
21735 // Look for the following pattern:
21737 // %addend = phi [%entry, 0], [%loop, %result]
21739 // %result<tied1> = FMA213 %m2<tied0>, %m1, %addend
21743 // %addend = phi [%entry, 0], [%loop, %result]
21745 // %result<tied1> = FMA231 %addend<tied0>, %m1, %m2
21747 for (unsigned i = 1, e = AddendDef.getNumOperands(); i < e; i += 2) {
21748 assert(AddendDef.getOperand(i).isReg());
21749 MachineOperand PHISrcOp = AddendDef.getOperand(i);
21750 MachineInstr &PHISrcInst = *MRI.def_instr_begin(PHISrcOp.getReg());
21751 if (&PHISrcInst == MI) {
21752 // Found a matching instruction.
21753 unsigned NewFMAOpc = 0;
21754 switch (MI->getOpcode()) {
21755 case X86::VFMADDPDr213r: NewFMAOpc = X86::VFMADDPDr231r; break;
21756 case X86::VFMADDPSr213r: NewFMAOpc = X86::VFMADDPSr231r; break;
21757 case X86::VFMADDSDr213r: NewFMAOpc = X86::VFMADDSDr231r; break;
21758 case X86::VFMADDSSr213r: NewFMAOpc = X86::VFMADDSSr231r; break;
21759 case X86::VFMSUBPDr213r: NewFMAOpc = X86::VFMSUBPDr231r; break;
21760 case X86::VFMSUBPSr213r: NewFMAOpc = X86::VFMSUBPSr231r; break;
21761 case X86::VFMSUBSDr213r: NewFMAOpc = X86::VFMSUBSDr231r; break;
21762 case X86::VFMSUBSSr213r: NewFMAOpc = X86::VFMSUBSSr231r; break;
21763 case X86::VFNMADDPDr213r: NewFMAOpc = X86::VFNMADDPDr231r; break;
21764 case X86::VFNMADDPSr213r: NewFMAOpc = X86::VFNMADDPSr231r; break;
21765 case X86::VFNMADDSDr213r: NewFMAOpc = X86::VFNMADDSDr231r; break;
21766 case X86::VFNMADDSSr213r: NewFMAOpc = X86::VFNMADDSSr231r; break;
21767 case X86::VFNMSUBPDr213r: NewFMAOpc = X86::VFNMSUBPDr231r; break;
21768 case X86::VFNMSUBPSr213r: NewFMAOpc = X86::VFNMSUBPSr231r; break;
21769 case X86::VFNMSUBSDr213r: NewFMAOpc = X86::VFNMSUBSDr231r; break;
21770 case X86::VFNMSUBSSr213r: NewFMAOpc = X86::VFNMSUBSSr231r; break;
21771 case X86::VFMADDSUBPDr213r: NewFMAOpc = X86::VFMADDSUBPDr231r; break;
21772 case X86::VFMADDSUBPSr213r: NewFMAOpc = X86::VFMADDSUBPSr231r; break;
21773 case X86::VFMSUBADDPDr213r: NewFMAOpc = X86::VFMSUBADDPDr231r; break;
21774 case X86::VFMSUBADDPSr213r: NewFMAOpc = X86::VFMSUBADDPSr231r; break;
21776 case X86::VFMADDPDr213rY: NewFMAOpc = X86::VFMADDPDr231rY; break;
21777 case X86::VFMADDPSr213rY: NewFMAOpc = X86::VFMADDPSr231rY; break;
21778 case X86::VFMSUBPDr213rY: NewFMAOpc = X86::VFMSUBPDr231rY; break;
21779 case X86::VFMSUBPSr213rY: NewFMAOpc = X86::VFMSUBPSr231rY; break;
21780 case X86::VFNMADDPDr213rY: NewFMAOpc = X86::VFNMADDPDr231rY; break;
21781 case X86::VFNMADDPSr213rY: NewFMAOpc = X86::VFNMADDPSr231rY; break;
21782 case X86::VFNMSUBPDr213rY: NewFMAOpc = X86::VFNMSUBPDr231rY; break;
21783 case X86::VFNMSUBPSr213rY: NewFMAOpc = X86::VFNMSUBPSr231rY; break;
21784 case X86::VFMADDSUBPDr213rY: NewFMAOpc = X86::VFMADDSUBPDr231rY; break;
21785 case X86::VFMADDSUBPSr213rY: NewFMAOpc = X86::VFMADDSUBPSr231rY; break;
21786 case X86::VFMSUBADDPDr213rY: NewFMAOpc = X86::VFMSUBADDPDr231rY; break;
21787 case X86::VFMSUBADDPSr213rY: NewFMAOpc = X86::VFMSUBADDPSr231rY; break;
21788 default: llvm_unreachable("Unrecognized FMA variant.");
21791 const TargetInstrInfo &TII = *Subtarget->getInstrInfo();
21792 MachineInstrBuilder MIB =
21793 BuildMI(MF, MI->getDebugLoc(), TII.get(NewFMAOpc))
21794 .addOperand(MI->getOperand(0))
21795 .addOperand(MI->getOperand(3))
21796 .addOperand(MI->getOperand(2))
21797 .addOperand(MI->getOperand(1));
21798 MBB->insert(MachineBasicBlock::iterator(MI), MIB);
21799 MI->eraseFromParent();
21806 MachineBasicBlock *
21807 X86TargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
21808 MachineBasicBlock *BB) const {
21809 switch (MI->getOpcode()) {
21810 default: llvm_unreachable("Unexpected instr type to insert");
21811 case X86::TAILJMPd64:
21812 case X86::TAILJMPr64:
21813 case X86::TAILJMPm64:
21814 case X86::TAILJMPd64_REX:
21815 case X86::TAILJMPr64_REX:
21816 case X86::TAILJMPm64_REX:
21817 llvm_unreachable("TAILJMP64 would not be touched here.");
21818 case X86::TCRETURNdi64:
21819 case X86::TCRETURNri64:
21820 case X86::TCRETURNmi64:
21822 case X86::WIN_ALLOCA:
21823 return EmitLoweredWinAlloca(MI, BB);
21824 case X86::SEG_ALLOCA_32:
21825 case X86::SEG_ALLOCA_64:
21826 return EmitLoweredSegAlloca(MI, BB);
21827 case X86::TLSCall_32:
21828 case X86::TLSCall_64:
21829 return EmitLoweredTLSCall(MI, BB);
21830 case X86::CMOV_FR32:
21831 case X86::CMOV_FR64:
21832 case X86::CMOV_GR8:
21833 case X86::CMOV_GR16:
21834 case X86::CMOV_GR32:
21835 case X86::CMOV_RFP32:
21836 case X86::CMOV_RFP64:
21837 case X86::CMOV_RFP80:
21838 case X86::CMOV_V2F64:
21839 case X86::CMOV_V2I64:
21840 case X86::CMOV_V4F32:
21841 case X86::CMOV_V4F64:
21842 case X86::CMOV_V4I64:
21843 case X86::CMOV_V16F32:
21844 case X86::CMOV_V8F32:
21845 case X86::CMOV_V8F64:
21846 case X86::CMOV_V8I64:
21847 case X86::CMOV_V8I1:
21848 case X86::CMOV_V16I1:
21849 case X86::CMOV_V32I1:
21850 case X86::CMOV_V64I1:
21851 return EmitLoweredSelect(MI, BB);
21853 case X86::RELEASE_FADD32mr:
21854 case X86::RELEASE_FADD64mr:
21855 return EmitLoweredAtomicFP(MI, BB);
21857 case X86::FP32_TO_INT16_IN_MEM:
21858 case X86::FP32_TO_INT32_IN_MEM:
21859 case X86::FP32_TO_INT64_IN_MEM:
21860 case X86::FP64_TO_INT16_IN_MEM:
21861 case X86::FP64_TO_INT32_IN_MEM:
21862 case X86::FP64_TO_INT64_IN_MEM:
21863 case X86::FP80_TO_INT16_IN_MEM:
21864 case X86::FP80_TO_INT32_IN_MEM:
21865 case X86::FP80_TO_INT64_IN_MEM: {
21866 MachineFunction *F = BB->getParent();
21867 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
21868 DebugLoc DL = MI->getDebugLoc();
21870 // Change the floating point control register to use "round towards zero"
21871 // mode when truncating to an integer value.
21872 int CWFrameIdx = F->getFrameInfo()->CreateStackObject(2, 2, false);
21873 addFrameReference(BuildMI(*BB, MI, DL,
21874 TII->get(X86::FNSTCW16m)), CWFrameIdx);
21876 // Load the old value of the high byte of the control word...
21878 F->getRegInfo().createVirtualRegister(&X86::GR16RegClass);
21879 addFrameReference(BuildMI(*BB, MI, DL, TII->get(X86::MOV16rm), OldCW),
21882 // Set the high part to be round to zero...
21883 addFrameReference(BuildMI(*BB, MI, DL, TII->get(X86::MOV16mi)), CWFrameIdx)
21886 // Reload the modified control word now...
21887 addFrameReference(BuildMI(*BB, MI, DL,
21888 TII->get(X86::FLDCW16m)), CWFrameIdx);
21890 // Restore the memory image of control word to original value
21891 addFrameReference(BuildMI(*BB, MI, DL, TII->get(X86::MOV16mr)), CWFrameIdx)
21894 // Get the X86 opcode to use.
21896 switch (MI->getOpcode()) {
21897 default: llvm_unreachable("illegal opcode!");
21898 case X86::FP32_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m32; break;
21899 case X86::FP32_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m32; break;
21900 case X86::FP32_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m32; break;
21901 case X86::FP64_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m64; break;
21902 case X86::FP64_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m64; break;
21903 case X86::FP64_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m64; break;
21904 case X86::FP80_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m80; break;
21905 case X86::FP80_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m80; break;
21906 case X86::FP80_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m80; break;
21910 MachineOperand &Op = MI->getOperand(0);
21912 AM.BaseType = X86AddressMode::RegBase;
21913 AM.Base.Reg = Op.getReg();
21915 AM.BaseType = X86AddressMode::FrameIndexBase;
21916 AM.Base.FrameIndex = Op.getIndex();
21918 Op = MI->getOperand(1);
21920 AM.Scale = Op.getImm();
21921 Op = MI->getOperand(2);
21923 AM.IndexReg = Op.getImm();
21924 Op = MI->getOperand(3);
21925 if (Op.isGlobal()) {
21926 AM.GV = Op.getGlobal();
21928 AM.Disp = Op.getImm();
21930 addFullAddress(BuildMI(*BB, MI, DL, TII->get(Opc)), AM)
21931 .addReg(MI->getOperand(X86::AddrNumOperands).getReg());
21933 // Reload the original control word now.
21934 addFrameReference(BuildMI(*BB, MI, DL,
21935 TII->get(X86::FLDCW16m)), CWFrameIdx);
21937 MI->eraseFromParent(); // The pseudo instruction is gone now.
21940 // String/text processing lowering.
21941 case X86::PCMPISTRM128REG:
21942 case X86::VPCMPISTRM128REG:
21943 case X86::PCMPISTRM128MEM:
21944 case X86::VPCMPISTRM128MEM:
21945 case X86::PCMPESTRM128REG:
21946 case X86::VPCMPESTRM128REG:
21947 case X86::PCMPESTRM128MEM:
21948 case X86::VPCMPESTRM128MEM:
21949 assert(Subtarget->hasSSE42() &&
21950 "Target must have SSE4.2 or AVX features enabled");
21951 return EmitPCMPSTRM(MI, BB, Subtarget->getInstrInfo());
21953 // String/text processing lowering.
21954 case X86::PCMPISTRIREG:
21955 case X86::VPCMPISTRIREG:
21956 case X86::PCMPISTRIMEM:
21957 case X86::VPCMPISTRIMEM:
21958 case X86::PCMPESTRIREG:
21959 case X86::VPCMPESTRIREG:
21960 case X86::PCMPESTRIMEM:
21961 case X86::VPCMPESTRIMEM:
21962 assert(Subtarget->hasSSE42() &&
21963 "Target must have SSE4.2 or AVX features enabled");
21964 return EmitPCMPSTRI(MI, BB, Subtarget->getInstrInfo());
21966 // Thread synchronization.
21968 return EmitMonitor(MI, BB, Subtarget);
21972 return EmitXBegin(MI, BB, Subtarget->getInstrInfo());
21974 case X86::VASTART_SAVE_XMM_REGS:
21975 return EmitVAStartSaveXMMRegsWithCustomInserter(MI, BB);
21977 case X86::VAARG_64:
21978 return EmitVAARG64WithCustomInserter(MI, BB);
21980 case X86::EH_SjLj_SetJmp32:
21981 case X86::EH_SjLj_SetJmp64:
21982 return emitEHSjLjSetJmp(MI, BB);
21984 case X86::EH_SjLj_LongJmp32:
21985 case X86::EH_SjLj_LongJmp64:
21986 return emitEHSjLjLongJmp(MI, BB);
21988 case TargetOpcode::STATEPOINT:
21989 // As an implementation detail, STATEPOINT shares the STACKMAP format at
21990 // this point in the process. We diverge later.
21991 return emitPatchPoint(MI, BB);
21993 case TargetOpcode::STACKMAP:
21994 case TargetOpcode::PATCHPOINT:
21995 return emitPatchPoint(MI, BB);
21997 case X86::VFMADDPDr213r:
21998 case X86::VFMADDPSr213r:
21999 case X86::VFMADDSDr213r:
22000 case X86::VFMADDSSr213r:
22001 case X86::VFMSUBPDr213r:
22002 case X86::VFMSUBPSr213r:
22003 case X86::VFMSUBSDr213r:
22004 case X86::VFMSUBSSr213r:
22005 case X86::VFNMADDPDr213r:
22006 case X86::VFNMADDPSr213r:
22007 case X86::VFNMADDSDr213r:
22008 case X86::VFNMADDSSr213r:
22009 case X86::VFNMSUBPDr213r:
22010 case X86::VFNMSUBPSr213r:
22011 case X86::VFNMSUBSDr213r:
22012 case X86::VFNMSUBSSr213r:
22013 case X86::VFMADDSUBPDr213r:
22014 case X86::VFMADDSUBPSr213r:
22015 case X86::VFMSUBADDPDr213r:
22016 case X86::VFMSUBADDPSr213r:
22017 case X86::VFMADDPDr213rY:
22018 case X86::VFMADDPSr213rY:
22019 case X86::VFMSUBPDr213rY:
22020 case X86::VFMSUBPSr213rY:
22021 case X86::VFNMADDPDr213rY:
22022 case X86::VFNMADDPSr213rY:
22023 case X86::VFNMSUBPDr213rY:
22024 case X86::VFNMSUBPSr213rY:
22025 case X86::VFMADDSUBPDr213rY:
22026 case X86::VFMADDSUBPSr213rY:
22027 case X86::VFMSUBADDPDr213rY:
22028 case X86::VFMSUBADDPSr213rY:
22029 return emitFMA3Instr(MI, BB);
22033 //===----------------------------------------------------------------------===//
22034 // X86 Optimization Hooks
22035 //===----------------------------------------------------------------------===//
22037 void X86TargetLowering::computeKnownBitsForTargetNode(const SDValue Op,
22040 const SelectionDAG &DAG,
22041 unsigned Depth) const {
22042 unsigned BitWidth = KnownZero.getBitWidth();
22043 unsigned Opc = Op.getOpcode();
22044 assert((Opc >= ISD::BUILTIN_OP_END ||
22045 Opc == ISD::INTRINSIC_WO_CHAIN ||
22046 Opc == ISD::INTRINSIC_W_CHAIN ||
22047 Opc == ISD::INTRINSIC_VOID) &&
22048 "Should use MaskedValueIsZero if you don't know whether Op"
22049 " is a target node!");
22051 KnownZero = KnownOne = APInt(BitWidth, 0); // Don't know anything.
22065 // These nodes' second result is a boolean.
22066 if (Op.getResNo() == 0)
22069 case X86ISD::SETCC:
22070 KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - 1);
22072 case ISD::INTRINSIC_WO_CHAIN: {
22073 unsigned IntId = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
22074 unsigned NumLoBits = 0;
22077 case Intrinsic::x86_sse_movmsk_ps:
22078 case Intrinsic::x86_avx_movmsk_ps_256:
22079 case Intrinsic::x86_sse2_movmsk_pd:
22080 case Intrinsic::x86_avx_movmsk_pd_256:
22081 case Intrinsic::x86_mmx_pmovmskb:
22082 case Intrinsic::x86_sse2_pmovmskb_128:
22083 case Intrinsic::x86_avx2_pmovmskb: {
22084 // High bits of movmskp{s|d}, pmovmskb are known zero.
22086 default: llvm_unreachable("Impossible intrinsic"); // Can't reach here.
22087 case Intrinsic::x86_sse_movmsk_ps: NumLoBits = 4; break;
22088 case Intrinsic::x86_avx_movmsk_ps_256: NumLoBits = 8; break;
22089 case Intrinsic::x86_sse2_movmsk_pd: NumLoBits = 2; break;
22090 case Intrinsic::x86_avx_movmsk_pd_256: NumLoBits = 4; break;
22091 case Intrinsic::x86_mmx_pmovmskb: NumLoBits = 8; break;
22092 case Intrinsic::x86_sse2_pmovmskb_128: NumLoBits = 16; break;
22093 case Intrinsic::x86_avx2_pmovmskb: NumLoBits = 32; break;
22095 KnownZero = APInt::getHighBitsSet(BitWidth, BitWidth - NumLoBits);
22104 unsigned X86TargetLowering::ComputeNumSignBitsForTargetNode(
22106 const SelectionDAG &,
22107 unsigned Depth) const {
22108 // SETCC_CARRY sets the dest to ~0 for true or 0 for false.
22109 if (Op.getOpcode() == X86ISD::SETCC_CARRY)
22110 return Op.getValueType().getScalarType().getSizeInBits();
22116 /// isGAPlusOffset - Returns true (and the GlobalValue and the offset) if the
22117 /// node is a GlobalAddress + offset.
22118 bool X86TargetLowering::isGAPlusOffset(SDNode *N,
22119 const GlobalValue* &GA,
22120 int64_t &Offset) const {
22121 if (N->getOpcode() == X86ISD::Wrapper) {
22122 if (isa<GlobalAddressSDNode>(N->getOperand(0))) {
22123 GA = cast<GlobalAddressSDNode>(N->getOperand(0))->getGlobal();
22124 Offset = cast<GlobalAddressSDNode>(N->getOperand(0))->getOffset();
22128 return TargetLowering::isGAPlusOffset(N, GA, Offset);
22131 /// isShuffleHigh128VectorInsertLow - Checks whether the shuffle node is the
22132 /// same as extracting the high 128-bit part of 256-bit vector and then
22133 /// inserting the result into the low part of a new 256-bit vector
22134 static bool isShuffleHigh128VectorInsertLow(ShuffleVectorSDNode *SVOp) {
22135 EVT VT = SVOp->getValueType(0);
22136 unsigned NumElems = VT.getVectorNumElements();
22138 // vector_shuffle <4, 5, 6, 7, u, u, u, u> or <2, 3, u, u>
22139 for (unsigned i = 0, j = NumElems/2; i != NumElems/2; ++i, ++j)
22140 if (!isUndefOrEqual(SVOp->getMaskElt(i), j) ||
22141 SVOp->getMaskElt(j) >= 0)
22147 /// isShuffleLow128VectorInsertHigh - Checks whether the shuffle node is the
22148 /// same as extracting the low 128-bit part of 256-bit vector and then
22149 /// inserting the result into the high part of a new 256-bit vector
22150 static bool isShuffleLow128VectorInsertHigh(ShuffleVectorSDNode *SVOp) {
22151 EVT VT = SVOp->getValueType(0);
22152 unsigned NumElems = VT.getVectorNumElements();
22154 // vector_shuffle <u, u, u, u, 0, 1, 2, 3> or <u, u, 0, 1>
22155 for (unsigned i = NumElems/2, j = 0; i != NumElems; ++i, ++j)
22156 if (!isUndefOrEqual(SVOp->getMaskElt(i), j) ||
22157 SVOp->getMaskElt(j) >= 0)
22163 /// PerformShuffleCombine256 - Performs shuffle combines for 256-bit vectors.
22164 static SDValue PerformShuffleCombine256(SDNode *N, SelectionDAG &DAG,
22165 TargetLowering::DAGCombinerInfo &DCI,
22166 const X86Subtarget* Subtarget) {
22168 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
22169 SDValue V1 = SVOp->getOperand(0);
22170 SDValue V2 = SVOp->getOperand(1);
22171 EVT VT = SVOp->getValueType(0);
22172 unsigned NumElems = VT.getVectorNumElements();
22174 if (V1.getOpcode() == ISD::CONCAT_VECTORS &&
22175 V2.getOpcode() == ISD::CONCAT_VECTORS) {
22179 // V UNDEF BUILD_VECTOR UNDEF
22181 // CONCAT_VECTOR CONCAT_VECTOR
22184 // RESULT: V + zero extended
22186 if (V2.getOperand(0).getOpcode() != ISD::BUILD_VECTOR ||
22187 V2.getOperand(1).getOpcode() != ISD::UNDEF ||
22188 V1.getOperand(1).getOpcode() != ISD::UNDEF)
22191 if (!ISD::isBuildVectorAllZeros(V2.getOperand(0).getNode()))
22194 // To match the shuffle mask, the first half of the mask should
22195 // be exactly the first vector, and all the rest a splat with the
22196 // first element of the second one.
22197 for (unsigned i = 0; i != NumElems/2; ++i)
22198 if (!isUndefOrEqual(SVOp->getMaskElt(i), i) ||
22199 !isUndefOrEqual(SVOp->getMaskElt(i+NumElems/2), NumElems))
22202 // If V1 is coming from a vector load then just fold to a VZEXT_LOAD.
22203 if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(V1.getOperand(0))) {
22204 if (Ld->hasNUsesOfValue(1, 0)) {
22205 SDVTList Tys = DAG.getVTList(MVT::v4i64, MVT::Other);
22206 SDValue Ops[] = { Ld->getChain(), Ld->getBasePtr() };
22208 DAG.getMemIntrinsicNode(X86ISD::VZEXT_LOAD, dl, Tys, Ops,
22210 Ld->getPointerInfo(),
22211 Ld->getAlignment(),
22212 false/*isVolatile*/, true/*ReadMem*/,
22213 false/*WriteMem*/);
22215 // Make sure the newly-created LOAD is in the same position as Ld in
22216 // terms of dependency. We create a TokenFactor for Ld and ResNode,
22217 // and update uses of Ld's output chain to use the TokenFactor.
22218 if (Ld->hasAnyUseOfValue(1)) {
22219 SDValue NewChain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
22220 SDValue(Ld, 1), SDValue(ResNode.getNode(), 1));
22221 DAG.ReplaceAllUsesOfValueWith(SDValue(Ld, 1), NewChain);
22222 DAG.UpdateNodeOperands(NewChain.getNode(), SDValue(Ld, 1),
22223 SDValue(ResNode.getNode(), 1));
22226 return DAG.getBitcast(VT, ResNode);
22230 // Emit a zeroed vector and insert the desired subvector on its
22232 SDValue Zeros = getZeroVector(VT, Subtarget, DAG, dl);
22233 SDValue InsV = Insert128BitVector(Zeros, V1.getOperand(0), 0, DAG, dl);
22234 return DCI.CombineTo(N, InsV);
22237 //===--------------------------------------------------------------------===//
22238 // Combine some shuffles into subvector extracts and inserts:
22241 // vector_shuffle <4, 5, 6, 7, u, u, u, u> or <2, 3, u, u>
22242 if (isShuffleHigh128VectorInsertLow(SVOp)) {
22243 SDValue V = Extract128BitVector(V1, NumElems/2, DAG, dl);
22244 SDValue InsV = Insert128BitVector(DAG.getUNDEF(VT), V, 0, DAG, dl);
22245 return DCI.CombineTo(N, InsV);
22248 // vector_shuffle <u, u, u, u, 0, 1, 2, 3> or <u, u, 0, 1>
22249 if (isShuffleLow128VectorInsertHigh(SVOp)) {
22250 SDValue V = Extract128BitVector(V1, 0, DAG, dl);
22251 SDValue InsV = Insert128BitVector(DAG.getUNDEF(VT), V, NumElems/2, DAG, dl);
22252 return DCI.CombineTo(N, InsV);
22258 /// \brief Combine an arbitrary chain of shuffles into a single instruction if
22261 /// This is the leaf of the recursive combinine below. When we have found some
22262 /// chain of single-use x86 shuffle instructions and accumulated the combined
22263 /// shuffle mask represented by them, this will try to pattern match that mask
22264 /// into either a single instruction if there is a special purpose instruction
22265 /// for this operation, or into a PSHUFB instruction which is a fully general
22266 /// instruction but should only be used to replace chains over a certain depth.
22267 static bool combineX86ShuffleChain(SDValue Op, SDValue Root, ArrayRef<int> Mask,
22268 int Depth, bool HasPSHUFB, SelectionDAG &DAG,
22269 TargetLowering::DAGCombinerInfo &DCI,
22270 const X86Subtarget *Subtarget) {
22271 assert(!Mask.empty() && "Cannot combine an empty shuffle mask!");
22273 // Find the operand that enters the chain. Note that multiple uses are OK
22274 // here, we're not going to remove the operand we find.
22275 SDValue Input = Op.getOperand(0);
22276 while (Input.getOpcode() == ISD::BITCAST)
22277 Input = Input.getOperand(0);
22279 MVT VT = Input.getSimpleValueType();
22280 MVT RootVT = Root.getSimpleValueType();
22283 if (Mask.size() == 1) {
22284 int Index = Mask[0];
22285 assert((Index >= 0 || Index == SM_SentinelUndef ||
22286 Index == SM_SentinelZero) &&
22287 "Invalid shuffle index found!");
22289 // We may end up with an accumulated mask of size 1 as a result of
22290 // widening of shuffle operands (see function canWidenShuffleElements).
22291 // If the only shuffle index is equal to SM_SentinelZero then propagate
22292 // a zero vector. Otherwise, the combine shuffle mask is a no-op shuffle
22293 // mask, and therefore the entire chain of shuffles can be folded away.
22294 if (Index == SM_SentinelZero)
22295 DCI.CombineTo(Root.getNode(), getZeroVector(RootVT, Subtarget, DAG, DL));
22297 DCI.CombineTo(Root.getNode(), DAG.getBitcast(RootVT, Input),
22302 // Use the float domain if the operand type is a floating point type.
22303 bool FloatDomain = VT.isFloatingPoint();
22305 // For floating point shuffles, we don't have free copies in the shuffle
22306 // instructions or the ability to load as part of the instruction, so
22307 // canonicalize their shuffles to UNPCK or MOV variants.
22309 // Note that even with AVX we prefer the PSHUFD form of shuffle for integer
22310 // vectors because it can have a load folded into it that UNPCK cannot. This
22311 // doesn't preclude something switching to the shorter encoding post-RA.
22313 // FIXME: Should teach these routines about AVX vector widths.
22314 if (FloatDomain && VT.getSizeInBits() == 128) {
22315 if (Mask.equals({0, 0}) || Mask.equals({1, 1})) {
22316 bool Lo = Mask.equals({0, 0});
22319 // Check if we have SSE3 which will let us use MOVDDUP. That instruction
22320 // is no slower than UNPCKLPD but has the option to fold the input operand
22321 // into even an unaligned memory load.
22322 if (Lo && Subtarget->hasSSE3()) {
22323 Shuffle = X86ISD::MOVDDUP;
22324 ShuffleVT = MVT::v2f64;
22326 // We have MOVLHPS and MOVHLPS throughout SSE and they encode smaller
22327 // than the UNPCK variants.
22328 Shuffle = Lo ? X86ISD::MOVLHPS : X86ISD::MOVHLPS;
22329 ShuffleVT = MVT::v4f32;
22331 if (Depth == 1 && Root->getOpcode() == Shuffle)
22332 return false; // Nothing to do!
22333 Op = DAG.getBitcast(ShuffleVT, Input);
22334 DCI.AddToWorklist(Op.getNode());
22335 if (Shuffle == X86ISD::MOVDDUP)
22336 Op = DAG.getNode(Shuffle, DL, ShuffleVT, Op);
22338 Op = DAG.getNode(Shuffle, DL, ShuffleVT, Op, Op);
22339 DCI.AddToWorklist(Op.getNode());
22340 DCI.CombineTo(Root.getNode(), DAG.getBitcast(RootVT, Op),
22344 if (Subtarget->hasSSE3() &&
22345 (Mask.equals({0, 0, 2, 2}) || Mask.equals({1, 1, 3, 3}))) {
22346 bool Lo = Mask.equals({0, 0, 2, 2});
22347 unsigned Shuffle = Lo ? X86ISD::MOVSLDUP : X86ISD::MOVSHDUP;
22348 MVT ShuffleVT = MVT::v4f32;
22349 if (Depth == 1 && Root->getOpcode() == Shuffle)
22350 return false; // Nothing to do!
22351 Op = DAG.getBitcast(ShuffleVT, Input);
22352 DCI.AddToWorklist(Op.getNode());
22353 Op = DAG.getNode(Shuffle, DL, ShuffleVT, Op);
22354 DCI.AddToWorklist(Op.getNode());
22355 DCI.CombineTo(Root.getNode(), DAG.getBitcast(RootVT, Op),
22359 if (Mask.equals({0, 0, 1, 1}) || Mask.equals({2, 2, 3, 3})) {
22360 bool Lo = Mask.equals({0, 0, 1, 1});
22361 unsigned Shuffle = Lo ? X86ISD::UNPCKL : X86ISD::UNPCKH;
22362 MVT ShuffleVT = MVT::v4f32;
22363 if (Depth == 1 && Root->getOpcode() == Shuffle)
22364 return false; // Nothing to do!
22365 Op = DAG.getBitcast(ShuffleVT, Input);
22366 DCI.AddToWorklist(Op.getNode());
22367 Op = DAG.getNode(Shuffle, DL, ShuffleVT, Op, Op);
22368 DCI.AddToWorklist(Op.getNode());
22369 DCI.CombineTo(Root.getNode(), DAG.getBitcast(RootVT, Op),
22375 // We always canonicalize the 8 x i16 and 16 x i8 shuffles into their UNPCK
22376 // variants as none of these have single-instruction variants that are
22377 // superior to the UNPCK formulation.
22378 if (!FloatDomain && VT.getSizeInBits() == 128 &&
22379 (Mask.equals({0, 0, 1, 1, 2, 2, 3, 3}) ||
22380 Mask.equals({4, 4, 5, 5, 6, 6, 7, 7}) ||
22381 Mask.equals({0, 0, 1, 1, 2, 2, 3, 3, 4, 4, 5, 5, 6, 6, 7, 7}) ||
22383 {8, 8, 9, 9, 10, 10, 11, 11, 12, 12, 13, 13, 14, 14, 15, 15}))) {
22384 bool Lo = Mask[0] == 0;
22385 unsigned Shuffle = Lo ? X86ISD::UNPCKL : X86ISD::UNPCKH;
22386 if (Depth == 1 && Root->getOpcode() == Shuffle)
22387 return false; // Nothing to do!
22389 switch (Mask.size()) {
22391 ShuffleVT = MVT::v8i16;
22394 ShuffleVT = MVT::v16i8;
22397 llvm_unreachable("Impossible mask size!");
22399 Op = DAG.getBitcast(ShuffleVT, Input);
22400 DCI.AddToWorklist(Op.getNode());
22401 Op = DAG.getNode(Shuffle, DL, ShuffleVT, Op, Op);
22402 DCI.AddToWorklist(Op.getNode());
22403 DCI.CombineTo(Root.getNode(), DAG.getBitcast(RootVT, Op),
22408 // Don't try to re-form single instruction chains under any circumstances now
22409 // that we've done encoding canonicalization for them.
22413 // If we have 3 or more shuffle instructions or a chain involving PSHUFB, we
22414 // can replace them with a single PSHUFB instruction profitably. Intel's
22415 // manuals suggest only using PSHUFB if doing so replacing 5 instructions, but
22416 // in practice PSHUFB tends to be *very* fast so we're more aggressive.
22417 if ((Depth >= 3 || HasPSHUFB) && Subtarget->hasSSSE3()) {
22418 SmallVector<SDValue, 16> PSHUFBMask;
22419 int NumBytes = VT.getSizeInBits() / 8;
22420 int Ratio = NumBytes / Mask.size();
22421 for (int i = 0; i < NumBytes; ++i) {
22422 if (Mask[i / Ratio] == SM_SentinelUndef) {
22423 PSHUFBMask.push_back(DAG.getUNDEF(MVT::i8));
22426 int M = Mask[i / Ratio] != SM_SentinelZero
22427 ? Ratio * Mask[i / Ratio] + i % Ratio
22429 PSHUFBMask.push_back(DAG.getConstant(M, DL, MVT::i8));
22431 MVT ByteVT = MVT::getVectorVT(MVT::i8, NumBytes);
22432 Op = DAG.getBitcast(ByteVT, Input);
22433 DCI.AddToWorklist(Op.getNode());
22434 SDValue PSHUFBMaskOp =
22435 DAG.getNode(ISD::BUILD_VECTOR, DL, ByteVT, PSHUFBMask);
22436 DCI.AddToWorklist(PSHUFBMaskOp.getNode());
22437 Op = DAG.getNode(X86ISD::PSHUFB, DL, ByteVT, Op, PSHUFBMaskOp);
22438 DCI.AddToWorklist(Op.getNode());
22439 DCI.CombineTo(Root.getNode(), DAG.getBitcast(RootVT, Op),
22444 // Failed to find any combines.
22448 /// \brief Fully generic combining of x86 shuffle instructions.
22450 /// This should be the last combine run over the x86 shuffle instructions. Once
22451 /// they have been fully optimized, this will recursively consider all chains
22452 /// of single-use shuffle instructions, build a generic model of the cumulative
22453 /// shuffle operation, and check for simpler instructions which implement this
22454 /// operation. We use this primarily for two purposes:
22456 /// 1) Collapse generic shuffles to specialized single instructions when
22457 /// equivalent. In most cases, this is just an encoding size win, but
22458 /// sometimes we will collapse multiple generic shuffles into a single
22459 /// special-purpose shuffle.
22460 /// 2) Look for sequences of shuffle instructions with 3 or more total
22461 /// instructions, and replace them with the slightly more expensive SSSE3
22462 /// PSHUFB instruction if available. We do this as the last combining step
22463 /// to ensure we avoid using PSHUFB if we can implement the shuffle with
22464 /// a suitable short sequence of other instructions. The PHUFB will either
22465 /// use a register or have to read from memory and so is slightly (but only
22466 /// slightly) more expensive than the other shuffle instructions.
22468 /// Because this is inherently a quadratic operation (for each shuffle in
22469 /// a chain, we recurse up the chain), the depth is limited to 8 instructions.
22470 /// This should never be an issue in practice as the shuffle lowering doesn't
22471 /// produce sequences of more than 8 instructions.
22473 /// FIXME: We will currently miss some cases where the redundant shuffling
22474 /// would simplify under the threshold for PSHUFB formation because of
22475 /// combine-ordering. To fix this, we should do the redundant instruction
22476 /// combining in this recursive walk.
22477 static bool combineX86ShufflesRecursively(SDValue Op, SDValue Root,
22478 ArrayRef<int> RootMask,
22479 int Depth, bool HasPSHUFB,
22481 TargetLowering::DAGCombinerInfo &DCI,
22482 const X86Subtarget *Subtarget) {
22483 // Bound the depth of our recursive combine because this is ultimately
22484 // quadratic in nature.
22488 // Directly rip through bitcasts to find the underlying operand.
22489 while (Op.getOpcode() == ISD::BITCAST && Op.getOperand(0).hasOneUse())
22490 Op = Op.getOperand(0);
22492 MVT VT = Op.getSimpleValueType();
22493 if (!VT.isVector())
22494 return false; // Bail if we hit a non-vector.
22496 assert(Root.getSimpleValueType().isVector() &&
22497 "Shuffles operate on vector types!");
22498 assert(VT.getSizeInBits() == Root.getSimpleValueType().getSizeInBits() &&
22499 "Can only combine shuffles of the same vector register size.");
22501 if (!isTargetShuffle(Op.getOpcode()))
22503 SmallVector<int, 16> OpMask;
22505 bool HaveMask = getTargetShuffleMask(Op.getNode(), VT, OpMask, IsUnary);
22506 // We only can combine unary shuffles which we can decode the mask for.
22507 if (!HaveMask || !IsUnary)
22510 assert(VT.getVectorNumElements() == OpMask.size() &&
22511 "Different mask size from vector size!");
22512 assert(((RootMask.size() > OpMask.size() &&
22513 RootMask.size() % OpMask.size() == 0) ||
22514 (OpMask.size() > RootMask.size() &&
22515 OpMask.size() % RootMask.size() == 0) ||
22516 OpMask.size() == RootMask.size()) &&
22517 "The smaller number of elements must divide the larger.");
22518 int RootRatio = std::max<int>(1, OpMask.size() / RootMask.size());
22519 int OpRatio = std::max<int>(1, RootMask.size() / OpMask.size());
22520 assert(((RootRatio == 1 && OpRatio == 1) ||
22521 (RootRatio == 1) != (OpRatio == 1)) &&
22522 "Must not have a ratio for both incoming and op masks!");
22524 SmallVector<int, 16> Mask;
22525 Mask.reserve(std::max(OpMask.size(), RootMask.size()));
22527 // Merge this shuffle operation's mask into our accumulated mask. Note that
22528 // this shuffle's mask will be the first applied to the input, followed by the
22529 // root mask to get us all the way to the root value arrangement. The reason
22530 // for this order is that we are recursing up the operation chain.
22531 for (int i = 0, e = std::max(OpMask.size(), RootMask.size()); i < e; ++i) {
22532 int RootIdx = i / RootRatio;
22533 if (RootMask[RootIdx] < 0) {
22534 // This is a zero or undef lane, we're done.
22535 Mask.push_back(RootMask[RootIdx]);
22539 int RootMaskedIdx = RootMask[RootIdx] * RootRatio + i % RootRatio;
22540 int OpIdx = RootMaskedIdx / OpRatio;
22541 if (OpMask[OpIdx] < 0) {
22542 // The incoming lanes are zero or undef, it doesn't matter which ones we
22544 Mask.push_back(OpMask[OpIdx]);
22548 // Ok, we have non-zero lanes, map them through.
22549 Mask.push_back(OpMask[OpIdx] * OpRatio +
22550 RootMaskedIdx % OpRatio);
22553 // See if we can recurse into the operand to combine more things.
22554 switch (Op.getOpcode()) {
22555 case X86ISD::PSHUFB:
22557 case X86ISD::PSHUFD:
22558 case X86ISD::PSHUFHW:
22559 case X86ISD::PSHUFLW:
22560 if (Op.getOperand(0).hasOneUse() &&
22561 combineX86ShufflesRecursively(Op.getOperand(0), Root, Mask, Depth + 1,
22562 HasPSHUFB, DAG, DCI, Subtarget))
22566 case X86ISD::UNPCKL:
22567 case X86ISD::UNPCKH:
22568 assert(Op.getOperand(0) == Op.getOperand(1) &&
22569 "We only combine unary shuffles!");
22570 // We can't check for single use, we have to check that this shuffle is the
22572 if (Op->isOnlyUserOf(Op.getOperand(0).getNode()) &&
22573 combineX86ShufflesRecursively(Op.getOperand(0), Root, Mask, Depth + 1,
22574 HasPSHUFB, DAG, DCI, Subtarget))
22579 // Minor canonicalization of the accumulated shuffle mask to make it easier
22580 // to match below. All this does is detect masks with squential pairs of
22581 // elements, and shrink them to the half-width mask. It does this in a loop
22582 // so it will reduce the size of the mask to the minimal width mask which
22583 // performs an equivalent shuffle.
22584 SmallVector<int, 16> WidenedMask;
22585 while (Mask.size() > 1 && canWidenShuffleElements(Mask, WidenedMask)) {
22586 Mask = std::move(WidenedMask);
22587 WidenedMask.clear();
22590 return combineX86ShuffleChain(Op, Root, Mask, Depth, HasPSHUFB, DAG, DCI,
22594 /// \brief Get the PSHUF-style mask from PSHUF node.
22596 /// This is a very minor wrapper around getTargetShuffleMask to easy forming v4
22597 /// PSHUF-style masks that can be reused with such instructions.
22598 static SmallVector<int, 4> getPSHUFShuffleMask(SDValue N) {
22599 MVT VT = N.getSimpleValueType();
22600 SmallVector<int, 4> Mask;
22602 bool HaveMask = getTargetShuffleMask(N.getNode(), VT, Mask, IsUnary);
22606 // If we have more than 128-bits, only the low 128-bits of shuffle mask
22607 // matter. Check that the upper masks are repeats and remove them.
22608 if (VT.getSizeInBits() > 128) {
22609 int LaneElts = 128 / VT.getScalarSizeInBits();
22611 for (int i = 1, NumLanes = VT.getSizeInBits() / 128; i < NumLanes; ++i)
22612 for (int j = 0; j < LaneElts; ++j)
22613 assert(Mask[j] == Mask[i * LaneElts + j] - (LaneElts * i) &&
22614 "Mask doesn't repeat in high 128-bit lanes!");
22616 Mask.resize(LaneElts);
22619 switch (N.getOpcode()) {
22620 case X86ISD::PSHUFD:
22622 case X86ISD::PSHUFLW:
22625 case X86ISD::PSHUFHW:
22626 Mask.erase(Mask.begin(), Mask.begin() + 4);
22627 for (int &M : Mask)
22631 llvm_unreachable("No valid shuffle instruction found!");
22635 /// \brief Search for a combinable shuffle across a chain ending in pshufd.
22637 /// We walk up the chain and look for a combinable shuffle, skipping over
22638 /// shuffles that we could hoist this shuffle's transformation past without
22639 /// altering anything.
22641 combineRedundantDWordShuffle(SDValue N, MutableArrayRef<int> Mask,
22643 TargetLowering::DAGCombinerInfo &DCI) {
22644 assert(N.getOpcode() == X86ISD::PSHUFD &&
22645 "Called with something other than an x86 128-bit half shuffle!");
22648 // Walk up a single-use chain looking for a combinable shuffle. Keep a stack
22649 // of the shuffles in the chain so that we can form a fresh chain to replace
22651 SmallVector<SDValue, 8> Chain;
22652 SDValue V = N.getOperand(0);
22653 for (; V.hasOneUse(); V = V.getOperand(0)) {
22654 switch (V.getOpcode()) {
22656 return SDValue(); // Nothing combined!
22659 // Skip bitcasts as we always know the type for the target specific
22663 case X86ISD::PSHUFD:
22664 // Found another dword shuffle.
22667 case X86ISD::PSHUFLW:
22668 // Check that the low words (being shuffled) are the identity in the
22669 // dword shuffle, and the high words are self-contained.
22670 if (Mask[0] != 0 || Mask[1] != 1 ||
22671 !(Mask[2] >= 2 && Mask[2] < 4 && Mask[3] >= 2 && Mask[3] < 4))
22674 Chain.push_back(V);
22677 case X86ISD::PSHUFHW:
22678 // Check that the high words (being shuffled) are the identity in the
22679 // dword shuffle, and the low words are self-contained.
22680 if (Mask[2] != 2 || Mask[3] != 3 ||
22681 !(Mask[0] >= 0 && Mask[0] < 2 && Mask[1] >= 0 && Mask[1] < 2))
22684 Chain.push_back(V);
22687 case X86ISD::UNPCKL:
22688 case X86ISD::UNPCKH:
22689 // For either i8 -> i16 or i16 -> i32 unpacks, we can combine a dword
22690 // shuffle into a preceding word shuffle.
22691 if (V.getSimpleValueType().getScalarType() != MVT::i8 &&
22692 V.getSimpleValueType().getScalarType() != MVT::i16)
22695 // Search for a half-shuffle which we can combine with.
22696 unsigned CombineOp =
22697 V.getOpcode() == X86ISD::UNPCKL ? X86ISD::PSHUFLW : X86ISD::PSHUFHW;
22698 if (V.getOperand(0) != V.getOperand(1) ||
22699 !V->isOnlyUserOf(V.getOperand(0).getNode()))
22701 Chain.push_back(V);
22702 V = V.getOperand(0);
22704 switch (V.getOpcode()) {
22706 return SDValue(); // Nothing to combine.
22708 case X86ISD::PSHUFLW:
22709 case X86ISD::PSHUFHW:
22710 if (V.getOpcode() == CombineOp)
22713 Chain.push_back(V);
22717 V = V.getOperand(0);
22721 } while (V.hasOneUse());
22724 // Break out of the loop if we break out of the switch.
22728 if (!V.hasOneUse())
22729 // We fell out of the loop without finding a viable combining instruction.
22732 // Merge this node's mask and our incoming mask.
22733 SmallVector<int, 4> VMask = getPSHUFShuffleMask(V);
22734 for (int &M : Mask)
22736 V = DAG.getNode(V.getOpcode(), DL, V.getValueType(), V.getOperand(0),
22737 getV4X86ShuffleImm8ForMask(Mask, DL, DAG));
22739 // Rebuild the chain around this new shuffle.
22740 while (!Chain.empty()) {
22741 SDValue W = Chain.pop_back_val();
22743 if (V.getValueType() != W.getOperand(0).getValueType())
22744 V = DAG.getBitcast(W.getOperand(0).getValueType(), V);
22746 switch (W.getOpcode()) {
22748 llvm_unreachable("Only PSHUF and UNPCK instructions get here!");
22750 case X86ISD::UNPCKL:
22751 case X86ISD::UNPCKH:
22752 V = DAG.getNode(W.getOpcode(), DL, W.getValueType(), V, V);
22755 case X86ISD::PSHUFD:
22756 case X86ISD::PSHUFLW:
22757 case X86ISD::PSHUFHW:
22758 V = DAG.getNode(W.getOpcode(), DL, W.getValueType(), V, W.getOperand(1));
22762 if (V.getValueType() != N.getValueType())
22763 V = DAG.getBitcast(N.getValueType(), V);
22765 // Return the new chain to replace N.
22769 /// \brief Search for a combinable shuffle across a chain ending in pshuflw or
22772 /// We walk up the chain, skipping shuffles of the other half and looking
22773 /// through shuffles which switch halves trying to find a shuffle of the same
22774 /// pair of dwords.
22775 static bool combineRedundantHalfShuffle(SDValue N, MutableArrayRef<int> Mask,
22777 TargetLowering::DAGCombinerInfo &DCI) {
22779 (N.getOpcode() == X86ISD::PSHUFLW || N.getOpcode() == X86ISD::PSHUFHW) &&
22780 "Called with something other than an x86 128-bit half shuffle!");
22782 unsigned CombineOpcode = N.getOpcode();
22784 // Walk up a single-use chain looking for a combinable shuffle.
22785 SDValue V = N.getOperand(0);
22786 for (; V.hasOneUse(); V = V.getOperand(0)) {
22787 switch (V.getOpcode()) {
22789 return false; // Nothing combined!
22792 // Skip bitcasts as we always know the type for the target specific
22796 case X86ISD::PSHUFLW:
22797 case X86ISD::PSHUFHW:
22798 if (V.getOpcode() == CombineOpcode)
22801 // Other-half shuffles are no-ops.
22804 // Break out of the loop if we break out of the switch.
22808 if (!V.hasOneUse())
22809 // We fell out of the loop without finding a viable combining instruction.
22812 // Combine away the bottom node as its shuffle will be accumulated into
22813 // a preceding shuffle.
22814 DCI.CombineTo(N.getNode(), N.getOperand(0), /*AddTo*/ true);
22816 // Record the old value.
22819 // Merge this node's mask and our incoming mask (adjusted to account for all
22820 // the pshufd instructions encountered).
22821 SmallVector<int, 4> VMask = getPSHUFShuffleMask(V);
22822 for (int &M : Mask)
22824 V = DAG.getNode(V.getOpcode(), DL, MVT::v8i16, V.getOperand(0),
22825 getV4X86ShuffleImm8ForMask(Mask, DL, DAG));
22827 // Check that the shuffles didn't cancel each other out. If not, we need to
22828 // combine to the new one.
22830 // Replace the combinable shuffle with the combined one, updating all users
22831 // so that we re-evaluate the chain here.
22832 DCI.CombineTo(Old.getNode(), V, /*AddTo*/ true);
22837 /// \brief Try to combine x86 target specific shuffles.
22838 static SDValue PerformTargetShuffleCombine(SDValue N, SelectionDAG &DAG,
22839 TargetLowering::DAGCombinerInfo &DCI,
22840 const X86Subtarget *Subtarget) {
22842 MVT VT = N.getSimpleValueType();
22843 SmallVector<int, 4> Mask;
22845 switch (N.getOpcode()) {
22846 case X86ISD::PSHUFD:
22847 case X86ISD::PSHUFLW:
22848 case X86ISD::PSHUFHW:
22849 Mask = getPSHUFShuffleMask(N);
22850 assert(Mask.size() == 4);
22856 // Nuke no-op shuffles that show up after combining.
22857 if (isNoopShuffleMask(Mask))
22858 return DCI.CombineTo(N.getNode(), N.getOperand(0), /*AddTo*/ true);
22860 // Look for simplifications involving one or two shuffle instructions.
22861 SDValue V = N.getOperand(0);
22862 switch (N.getOpcode()) {
22865 case X86ISD::PSHUFLW:
22866 case X86ISD::PSHUFHW:
22867 assert(VT.getScalarType() == MVT::i16 && "Bad word shuffle type!");
22869 if (combineRedundantHalfShuffle(N, Mask, DAG, DCI))
22870 return SDValue(); // We combined away this shuffle, so we're done.
22872 // See if this reduces to a PSHUFD which is no more expensive and can
22873 // combine with more operations. Note that it has to at least flip the
22874 // dwords as otherwise it would have been removed as a no-op.
22875 if (makeArrayRef(Mask).equals({2, 3, 0, 1})) {
22876 int DMask[] = {0, 1, 2, 3};
22877 int DOffset = N.getOpcode() == X86ISD::PSHUFLW ? 0 : 2;
22878 DMask[DOffset + 0] = DOffset + 1;
22879 DMask[DOffset + 1] = DOffset + 0;
22880 MVT DVT = MVT::getVectorVT(MVT::i32, VT.getVectorNumElements() / 2);
22881 V = DAG.getBitcast(DVT, V);
22882 DCI.AddToWorklist(V.getNode());
22883 V = DAG.getNode(X86ISD::PSHUFD, DL, DVT, V,
22884 getV4X86ShuffleImm8ForMask(DMask, DL, DAG));
22885 DCI.AddToWorklist(V.getNode());
22886 return DAG.getBitcast(VT, V);
22889 // Look for shuffle patterns which can be implemented as a single unpack.
22890 // FIXME: This doesn't handle the location of the PSHUFD generically, and
22891 // only works when we have a PSHUFD followed by two half-shuffles.
22892 if (Mask[0] == Mask[1] && Mask[2] == Mask[3] &&
22893 (V.getOpcode() == X86ISD::PSHUFLW ||
22894 V.getOpcode() == X86ISD::PSHUFHW) &&
22895 V.getOpcode() != N.getOpcode() &&
22897 SDValue D = V.getOperand(0);
22898 while (D.getOpcode() == ISD::BITCAST && D.hasOneUse())
22899 D = D.getOperand(0);
22900 if (D.getOpcode() == X86ISD::PSHUFD && D.hasOneUse()) {
22901 SmallVector<int, 4> VMask = getPSHUFShuffleMask(V);
22902 SmallVector<int, 4> DMask = getPSHUFShuffleMask(D);
22903 int NOffset = N.getOpcode() == X86ISD::PSHUFLW ? 0 : 4;
22904 int VOffset = V.getOpcode() == X86ISD::PSHUFLW ? 0 : 4;
22906 for (int i = 0; i < 4; ++i) {
22907 WordMask[i + NOffset] = Mask[i] + NOffset;
22908 WordMask[i + VOffset] = VMask[i] + VOffset;
22910 // Map the word mask through the DWord mask.
22912 for (int i = 0; i < 8; ++i)
22913 MappedMask[i] = 2 * DMask[WordMask[i] / 2] + WordMask[i] % 2;
22914 if (makeArrayRef(MappedMask).equals({0, 0, 1, 1, 2, 2, 3, 3}) ||
22915 makeArrayRef(MappedMask).equals({4, 4, 5, 5, 6, 6, 7, 7})) {
22916 // We can replace all three shuffles with an unpack.
22917 V = DAG.getBitcast(VT, D.getOperand(0));
22918 DCI.AddToWorklist(V.getNode());
22919 return DAG.getNode(MappedMask[0] == 0 ? X86ISD::UNPCKL
22928 case X86ISD::PSHUFD:
22929 if (SDValue NewN = combineRedundantDWordShuffle(N, Mask, DAG, DCI))
22938 /// \brief Try to combine a shuffle into a target-specific add-sub node.
22940 /// We combine this directly on the abstract vector shuffle nodes so it is
22941 /// easier to generically match. We also insert dummy vector shuffle nodes for
22942 /// the operands which explicitly discard the lanes which are unused by this
22943 /// operation to try to flow through the rest of the combiner the fact that
22944 /// they're unused.
22945 static SDValue combineShuffleToAddSub(SDNode *N, SelectionDAG &DAG) {
22947 EVT VT = N->getValueType(0);
22949 // We only handle target-independent shuffles.
22950 // FIXME: It would be easy and harmless to use the target shuffle mask
22951 // extraction tool to support more.
22952 if (N->getOpcode() != ISD::VECTOR_SHUFFLE)
22955 auto *SVN = cast<ShuffleVectorSDNode>(N);
22956 ArrayRef<int> Mask = SVN->getMask();
22957 SDValue V1 = N->getOperand(0);
22958 SDValue V2 = N->getOperand(1);
22960 // We require the first shuffle operand to be the SUB node, and the second to
22961 // be the ADD node.
22962 // FIXME: We should support the commuted patterns.
22963 if (V1->getOpcode() != ISD::FSUB || V2->getOpcode() != ISD::FADD)
22966 // If there are other uses of these operations we can't fold them.
22967 if (!V1->hasOneUse() || !V2->hasOneUse())
22970 // Ensure that both operations have the same operands. Note that we can
22971 // commute the FADD operands.
22972 SDValue LHS = V1->getOperand(0), RHS = V1->getOperand(1);
22973 if ((V2->getOperand(0) != LHS || V2->getOperand(1) != RHS) &&
22974 (V2->getOperand(0) != RHS || V2->getOperand(1) != LHS))
22977 // We're looking for blends between FADD and FSUB nodes. We insist on these
22978 // nodes being lined up in a specific expected pattern.
22979 if (!(isShuffleEquivalent(V1, V2, Mask, {0, 3}) ||
22980 isShuffleEquivalent(V1, V2, Mask, {0, 5, 2, 7}) ||
22981 isShuffleEquivalent(V1, V2, Mask, {0, 9, 2, 11, 4, 13, 6, 15})))
22984 // Only specific types are legal at this point, assert so we notice if and
22985 // when these change.
22986 assert((VT == MVT::v4f32 || VT == MVT::v2f64 || VT == MVT::v8f32 ||
22987 VT == MVT::v4f64) &&
22988 "Unknown vector type encountered!");
22990 return DAG.getNode(X86ISD::ADDSUB, DL, VT, LHS, RHS);
22993 /// PerformShuffleCombine - Performs several different shuffle combines.
22994 static SDValue PerformShuffleCombine(SDNode *N, SelectionDAG &DAG,
22995 TargetLowering::DAGCombinerInfo &DCI,
22996 const X86Subtarget *Subtarget) {
22998 SDValue N0 = N->getOperand(0);
22999 SDValue N1 = N->getOperand(1);
23000 EVT VT = N->getValueType(0);
23002 // Don't create instructions with illegal types after legalize types has run.
23003 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
23004 if (!DCI.isBeforeLegalize() && !TLI.isTypeLegal(VT.getVectorElementType()))
23007 // If we have legalized the vector types, look for blends of FADD and FSUB
23008 // nodes that we can fuse into an ADDSUB node.
23009 if (TLI.isTypeLegal(VT) && Subtarget->hasSSE3())
23010 if (SDValue AddSub = combineShuffleToAddSub(N, DAG))
23013 // Combine 256-bit vector shuffles. This is only profitable when in AVX mode
23014 if (Subtarget->hasFp256() && VT.is256BitVector() &&
23015 N->getOpcode() == ISD::VECTOR_SHUFFLE)
23016 return PerformShuffleCombine256(N, DAG, DCI, Subtarget);
23018 // During Type Legalization, when promoting illegal vector types,
23019 // the backend might introduce new shuffle dag nodes and bitcasts.
23021 // This code performs the following transformation:
23022 // fold: (shuffle (bitcast (BINOP A, B)), Undef, <Mask>) ->
23023 // (shuffle (BINOP (bitcast A), (bitcast B)), Undef, <Mask>)
23025 // We do this only if both the bitcast and the BINOP dag nodes have
23026 // one use. Also, perform this transformation only if the new binary
23027 // operation is legal. This is to avoid introducing dag nodes that
23028 // potentially need to be further expanded (or custom lowered) into a
23029 // less optimal sequence of dag nodes.
23030 if (!DCI.isBeforeLegalize() && DCI.isBeforeLegalizeOps() &&
23031 N1.getOpcode() == ISD::UNDEF && N0.hasOneUse() &&
23032 N0.getOpcode() == ISD::BITCAST) {
23033 SDValue BC0 = N0.getOperand(0);
23034 EVT SVT = BC0.getValueType();
23035 unsigned Opcode = BC0.getOpcode();
23036 unsigned NumElts = VT.getVectorNumElements();
23038 if (BC0.hasOneUse() && SVT.isVector() &&
23039 SVT.getVectorNumElements() * 2 == NumElts &&
23040 TLI.isOperationLegal(Opcode, VT)) {
23041 bool CanFold = false;
23053 unsigned SVTNumElts = SVT.getVectorNumElements();
23054 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
23055 for (unsigned i = 0, e = SVTNumElts; i != e && CanFold; ++i)
23056 CanFold = SVOp->getMaskElt(i) == (int)(i * 2);
23057 for (unsigned i = SVTNumElts, e = NumElts; i != e && CanFold; ++i)
23058 CanFold = SVOp->getMaskElt(i) < 0;
23061 SDValue BC00 = DAG.getBitcast(VT, BC0.getOperand(0));
23062 SDValue BC01 = DAG.getBitcast(VT, BC0.getOperand(1));
23063 SDValue NewBinOp = DAG.getNode(BC0.getOpcode(), dl, VT, BC00, BC01);
23064 return DAG.getVectorShuffle(VT, dl, NewBinOp, N1, &SVOp->getMask()[0]);
23069 // Combine a vector_shuffle that is equal to build_vector load1, load2, load3,
23070 // load4, <0, 1, 2, 3> into a 128-bit load if the load addresses are
23071 // consecutive, non-overlapping, and in the right order.
23072 SmallVector<SDValue, 16> Elts;
23073 for (unsigned i = 0, e = VT.getVectorNumElements(); i != e; ++i)
23074 Elts.push_back(getShuffleScalarElt(N, i, DAG, 0));
23076 if (SDValue LD = EltsFromConsecutiveLoads(VT, Elts, dl, DAG, true))
23079 if (isTargetShuffle(N->getOpcode())) {
23081 PerformTargetShuffleCombine(SDValue(N, 0), DAG, DCI, Subtarget);
23082 if (Shuffle.getNode())
23085 // Try recursively combining arbitrary sequences of x86 shuffle
23086 // instructions into higher-order shuffles. We do this after combining
23087 // specific PSHUF instruction sequences into their minimal form so that we
23088 // can evaluate how many specialized shuffle instructions are involved in
23089 // a particular chain.
23090 SmallVector<int, 1> NonceMask; // Just a placeholder.
23091 NonceMask.push_back(0);
23092 if (combineX86ShufflesRecursively(SDValue(N, 0), SDValue(N, 0), NonceMask,
23093 /*Depth*/ 1, /*HasPSHUFB*/ false, DAG,
23095 return SDValue(); // This routine will use CombineTo to replace N.
23101 /// XFormVExtractWithShuffleIntoLoad - Check if a vector extract from a target
23102 /// specific shuffle of a load can be folded into a single element load.
23103 /// Similar handling for VECTOR_SHUFFLE is performed by DAGCombiner, but
23104 /// shuffles have been custom lowered so we need to handle those here.
23105 static SDValue XFormVExtractWithShuffleIntoLoad(SDNode *N, SelectionDAG &DAG,
23106 TargetLowering::DAGCombinerInfo &DCI) {
23107 if (DCI.isBeforeLegalizeOps())
23110 SDValue InVec = N->getOperand(0);
23111 SDValue EltNo = N->getOperand(1);
23113 if (!isa<ConstantSDNode>(EltNo))
23116 EVT OriginalVT = InVec.getValueType();
23118 if (InVec.getOpcode() == ISD::BITCAST) {
23119 // Don't duplicate a load with other uses.
23120 if (!InVec.hasOneUse())
23122 EVT BCVT = InVec.getOperand(0).getValueType();
23123 if (!BCVT.isVector() ||
23124 BCVT.getVectorNumElements() != OriginalVT.getVectorNumElements())
23126 InVec = InVec.getOperand(0);
23129 EVT CurrentVT = InVec.getValueType();
23131 if (!isTargetShuffle(InVec.getOpcode()))
23134 // Don't duplicate a load with other uses.
23135 if (!InVec.hasOneUse())
23138 SmallVector<int, 16> ShuffleMask;
23140 if (!getTargetShuffleMask(InVec.getNode(), CurrentVT.getSimpleVT(),
23141 ShuffleMask, UnaryShuffle))
23144 // Select the input vector, guarding against out of range extract vector.
23145 unsigned NumElems = CurrentVT.getVectorNumElements();
23146 int Elt = cast<ConstantSDNode>(EltNo)->getZExtValue();
23147 int Idx = (Elt > (int)NumElems) ? -1 : ShuffleMask[Elt];
23148 SDValue LdNode = (Idx < (int)NumElems) ? InVec.getOperand(0)
23149 : InVec.getOperand(1);
23151 // If inputs to shuffle are the same for both ops, then allow 2 uses
23152 unsigned AllowedUses = InVec.getNumOperands() > 1 &&
23153 InVec.getOperand(0) == InVec.getOperand(1) ? 2 : 1;
23155 if (LdNode.getOpcode() == ISD::BITCAST) {
23156 // Don't duplicate a load with other uses.
23157 if (!LdNode.getNode()->hasNUsesOfValue(AllowedUses, 0))
23160 AllowedUses = 1; // only allow 1 load use if we have a bitcast
23161 LdNode = LdNode.getOperand(0);
23164 if (!ISD::isNormalLoad(LdNode.getNode()))
23167 LoadSDNode *LN0 = cast<LoadSDNode>(LdNode);
23169 if (!LN0 ||!LN0->hasNUsesOfValue(AllowedUses, 0) || LN0->isVolatile())
23172 EVT EltVT = N->getValueType(0);
23173 // If there's a bitcast before the shuffle, check if the load type and
23174 // alignment is valid.
23175 unsigned Align = LN0->getAlignment();
23176 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
23177 unsigned NewAlign = DAG.getDataLayout().getABITypeAlignment(
23178 EltVT.getTypeForEVT(*DAG.getContext()));
23180 if (NewAlign > Align || !TLI.isOperationLegalOrCustom(ISD::LOAD, EltVT))
23183 // All checks match so transform back to vector_shuffle so that DAG combiner
23184 // can finish the job
23187 // Create shuffle node taking into account the case that its a unary shuffle
23188 SDValue Shuffle = (UnaryShuffle) ? DAG.getUNDEF(CurrentVT)
23189 : InVec.getOperand(1);
23190 Shuffle = DAG.getVectorShuffle(CurrentVT, dl,
23191 InVec.getOperand(0), Shuffle,
23193 Shuffle = DAG.getBitcast(OriginalVT, Shuffle);
23194 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, N->getValueType(0), Shuffle,
23198 /// \brief Detect bitcasts between i32 to x86mmx low word. Since MMX types are
23199 /// special and don't usually play with other vector types, it's better to
23200 /// handle them early to be sure we emit efficient code by avoiding
23201 /// store-load conversions.
23202 static SDValue PerformBITCASTCombine(SDNode *N, SelectionDAG &DAG) {
23203 if (N->getValueType(0) != MVT::x86mmx ||
23204 N->getOperand(0)->getOpcode() != ISD::BUILD_VECTOR ||
23205 N->getOperand(0)->getValueType(0) != MVT::v2i32)
23208 SDValue V = N->getOperand(0);
23209 ConstantSDNode *C = dyn_cast<ConstantSDNode>(V.getOperand(1));
23210 if (C && C->getZExtValue() == 0 && V.getOperand(0).getValueType() == MVT::i32)
23211 return DAG.getNode(X86ISD::MMX_MOVW2D, SDLoc(V.getOperand(0)),
23212 N->getValueType(0), V.getOperand(0));
23217 /// PerformEXTRACT_VECTOR_ELTCombine - Detect vector gather/scatter index
23218 /// generation and convert it from being a bunch of shuffles and extracts
23219 /// into a somewhat faster sequence. For i686, the best sequence is apparently
23220 /// storing the value and loading scalars back, while for x64 we should
23221 /// use 64-bit extracts and shifts.
23222 static SDValue PerformEXTRACT_VECTOR_ELTCombine(SDNode *N, SelectionDAG &DAG,
23223 TargetLowering::DAGCombinerInfo &DCI) {
23224 if (SDValue NewOp = XFormVExtractWithShuffleIntoLoad(N, DAG, DCI))
23227 SDValue InputVector = N->getOperand(0);
23228 SDLoc dl(InputVector);
23229 // Detect mmx to i32 conversion through a v2i32 elt extract.
23230 if (InputVector.getOpcode() == ISD::BITCAST && InputVector.hasOneUse() &&
23231 N->getValueType(0) == MVT::i32 &&
23232 InputVector.getValueType() == MVT::v2i32) {
23234 // The bitcast source is a direct mmx result.
23235 SDValue MMXSrc = InputVector.getNode()->getOperand(0);
23236 if (MMXSrc.getValueType() == MVT::x86mmx)
23237 return DAG.getNode(X86ISD::MMX_MOVD2W, SDLoc(InputVector),
23238 N->getValueType(0),
23239 InputVector.getNode()->getOperand(0));
23241 // The mmx is indirect: (i64 extract_elt (v1i64 bitcast (x86mmx ...))).
23242 if (MMXSrc.getOpcode() == ISD::EXTRACT_VECTOR_ELT && MMXSrc.hasOneUse() &&
23243 MMXSrc.getValueType() == MVT::i64) {
23244 SDValue MMXSrcOp = MMXSrc.getOperand(0);
23245 if (MMXSrcOp.hasOneUse() && MMXSrcOp.getOpcode() == ISD::BITCAST &&
23246 MMXSrcOp.getValueType() == MVT::v1i64 &&
23247 MMXSrcOp.getOperand(0).getValueType() == MVT::x86mmx)
23248 return DAG.getNode(X86ISD::MMX_MOVD2W, SDLoc(InputVector),
23249 N->getValueType(0), MMXSrcOp.getOperand(0));
23253 EVT VT = N->getValueType(0);
23255 if (VT == MVT::i1 && dyn_cast<ConstantSDNode>(N->getOperand(1)) &&
23256 InputVector.getOpcode() == ISD::BITCAST &&
23257 dyn_cast<ConstantSDNode>(InputVector.getOperand(0))) {
23258 uint64_t ExtractedElt =
23259 cast<ConstantSDNode>(N->getOperand(1))->getZExtValue();
23260 uint64_t InputValue =
23261 cast<ConstantSDNode>(InputVector.getOperand(0))->getZExtValue();
23262 uint64_t Res = (InputValue >> ExtractedElt) & 1;
23263 return DAG.getConstant(Res, dl, MVT::i1);
23265 // Only operate on vectors of 4 elements, where the alternative shuffling
23266 // gets to be more expensive.
23267 if (InputVector.getValueType() != MVT::v4i32)
23270 // Check whether every use of InputVector is an EXTRACT_VECTOR_ELT with a
23271 // single use which is a sign-extend or zero-extend, and all elements are
23273 SmallVector<SDNode *, 4> Uses;
23274 unsigned ExtractedElements = 0;
23275 for (SDNode::use_iterator UI = InputVector.getNode()->use_begin(),
23276 UE = InputVector.getNode()->use_end(); UI != UE; ++UI) {
23277 if (UI.getUse().getResNo() != InputVector.getResNo())
23280 SDNode *Extract = *UI;
23281 if (Extract->getOpcode() != ISD::EXTRACT_VECTOR_ELT)
23284 if (Extract->getValueType(0) != MVT::i32)
23286 if (!Extract->hasOneUse())
23288 if (Extract->use_begin()->getOpcode() != ISD::SIGN_EXTEND &&
23289 Extract->use_begin()->getOpcode() != ISD::ZERO_EXTEND)
23291 if (!isa<ConstantSDNode>(Extract->getOperand(1)))
23294 // Record which element was extracted.
23295 ExtractedElements |=
23296 1 << cast<ConstantSDNode>(Extract->getOperand(1))->getZExtValue();
23298 Uses.push_back(Extract);
23301 // If not all the elements were used, this may not be worthwhile.
23302 if (ExtractedElements != 15)
23305 // Ok, we've now decided to do the transformation.
23306 // If 64-bit shifts are legal, use the extract-shift sequence,
23307 // otherwise bounce the vector off the cache.
23308 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
23311 if (TLI.isOperationLegal(ISD::SRA, MVT::i64)) {
23312 SDValue Cst = DAG.getBitcast(MVT::v2i64, InputVector);
23313 auto &DL = DAG.getDataLayout();
23314 EVT VecIdxTy = DAG.getTargetLoweringInfo().getVectorIdxTy(DL);
23315 SDValue BottomHalf = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i64, Cst,
23316 DAG.getConstant(0, dl, VecIdxTy));
23317 SDValue TopHalf = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i64, Cst,
23318 DAG.getConstant(1, dl, VecIdxTy));
23320 SDValue ShAmt = DAG.getConstant(
23321 32, dl, DAG.getTargetLoweringInfo().getShiftAmountTy(MVT::i64, DL));
23322 Vals[0] = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, BottomHalf);
23323 Vals[1] = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32,
23324 DAG.getNode(ISD::SRA, dl, MVT::i64, BottomHalf, ShAmt));
23325 Vals[2] = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, TopHalf);
23326 Vals[3] = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32,
23327 DAG.getNode(ISD::SRA, dl, MVT::i64, TopHalf, ShAmt));
23329 // Store the value to a temporary stack slot.
23330 SDValue StackPtr = DAG.CreateStackTemporary(InputVector.getValueType());
23331 SDValue Ch = DAG.getStore(DAG.getEntryNode(), dl, InputVector, StackPtr,
23332 MachinePointerInfo(), false, false, 0);
23334 EVT ElementType = InputVector.getValueType().getVectorElementType();
23335 unsigned EltSize = ElementType.getSizeInBits() / 8;
23337 // Replace each use (extract) with a load of the appropriate element.
23338 for (unsigned i = 0; i < 4; ++i) {
23339 uint64_t Offset = EltSize * i;
23340 auto PtrVT = TLI.getPointerTy(DAG.getDataLayout());
23341 SDValue OffsetVal = DAG.getConstant(Offset, dl, PtrVT);
23343 SDValue ScalarAddr =
23344 DAG.getNode(ISD::ADD, dl, PtrVT, StackPtr, OffsetVal);
23346 // Load the scalar.
23347 Vals[i] = DAG.getLoad(ElementType, dl, Ch,
23348 ScalarAddr, MachinePointerInfo(),
23349 false, false, false, 0);
23354 // Replace the extracts
23355 for (SmallVectorImpl<SDNode *>::iterator UI = Uses.begin(),
23356 UE = Uses.end(); UI != UE; ++UI) {
23357 SDNode *Extract = *UI;
23359 SDValue Idx = Extract->getOperand(1);
23360 uint64_t IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
23361 DAG.ReplaceAllUsesOfValueWith(SDValue(Extract, 0), Vals[IdxVal]);
23364 // The replacement was made in place; don't return anything.
23369 transformVSELECTtoBlendVECTOR_SHUFFLE(SDNode *N, SelectionDAG &DAG,
23370 const X86Subtarget *Subtarget) {
23372 SDValue Cond = N->getOperand(0);
23373 SDValue LHS = N->getOperand(1);
23374 SDValue RHS = N->getOperand(2);
23376 if (Cond.getOpcode() == ISD::SIGN_EXTEND) {
23377 SDValue CondSrc = Cond->getOperand(0);
23378 if (CondSrc->getOpcode() == ISD::SIGN_EXTEND_INREG)
23379 Cond = CondSrc->getOperand(0);
23382 if (!ISD::isBuildVectorOfConstantSDNodes(Cond.getNode()))
23385 // A vselect where all conditions and data are constants can be optimized into
23386 // a single vector load by SelectionDAGLegalize::ExpandBUILD_VECTOR().
23387 if (ISD::isBuildVectorOfConstantSDNodes(LHS.getNode()) &&
23388 ISD::isBuildVectorOfConstantSDNodes(RHS.getNode()))
23391 unsigned MaskValue = 0;
23392 if (!BUILD_VECTORtoBlendMask(cast<BuildVectorSDNode>(Cond), MaskValue))
23395 MVT VT = N->getSimpleValueType(0);
23396 unsigned NumElems = VT.getVectorNumElements();
23397 SmallVector<int, 8> ShuffleMask(NumElems, -1);
23398 for (unsigned i = 0; i < NumElems; ++i) {
23399 // Be sure we emit undef where we can.
23400 if (Cond.getOperand(i)->getOpcode() == ISD::UNDEF)
23401 ShuffleMask[i] = -1;
23403 ShuffleMask[i] = i + NumElems * ((MaskValue >> i) & 1);
23406 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
23407 if (!TLI.isShuffleMaskLegal(ShuffleMask, VT))
23409 return DAG.getVectorShuffle(VT, dl, LHS, RHS, &ShuffleMask[0]);
23412 /// PerformSELECTCombine - Do target-specific dag combines on SELECT and VSELECT
23414 static SDValue PerformSELECTCombine(SDNode *N, SelectionDAG &DAG,
23415 TargetLowering::DAGCombinerInfo &DCI,
23416 const X86Subtarget *Subtarget) {
23418 SDValue Cond = N->getOperand(0);
23419 // Get the LHS/RHS of the select.
23420 SDValue LHS = N->getOperand(1);
23421 SDValue RHS = N->getOperand(2);
23422 EVT VT = LHS.getValueType();
23423 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
23425 // If we have SSE[12] support, try to form min/max nodes. SSE min/max
23426 // instructions match the semantics of the common C idiom x<y?x:y but not
23427 // x<=y?x:y, because of how they handle negative zero (which can be
23428 // ignored in unsafe-math mode).
23429 // We also try to create v2f32 min/max nodes, which we later widen to v4f32.
23430 if (Cond.getOpcode() == ISD::SETCC && VT.isFloatingPoint() &&
23431 VT != MVT::f80 && (TLI.isTypeLegal(VT) || VT == MVT::v2f32) &&
23432 (Subtarget->hasSSE2() ||
23433 (Subtarget->hasSSE1() && VT.getScalarType() == MVT::f32))) {
23434 ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
23436 unsigned Opcode = 0;
23437 // Check for x CC y ? x : y.
23438 if (DAG.isEqualTo(LHS, Cond.getOperand(0)) &&
23439 DAG.isEqualTo(RHS, Cond.getOperand(1))) {
23443 // Converting this to a min would handle NaNs incorrectly, and swapping
23444 // the operands would cause it to handle comparisons between positive
23445 // and negative zero incorrectly.
23446 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)) {
23447 if (!DAG.getTarget().Options.UnsafeFPMath &&
23448 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
23450 std::swap(LHS, RHS);
23452 Opcode = X86ISD::FMIN;
23455 // Converting this to a min would handle comparisons between positive
23456 // and negative zero incorrectly.
23457 if (!DAG.getTarget().Options.UnsafeFPMath &&
23458 !DAG.isKnownNeverZero(LHS) && !DAG.isKnownNeverZero(RHS))
23460 Opcode = X86ISD::FMIN;
23463 // Converting this to a min would handle both negative zeros and NaNs
23464 // incorrectly, but we can swap the operands to fix both.
23465 std::swap(LHS, RHS);
23469 Opcode = X86ISD::FMIN;
23473 // Converting this to a max would handle comparisons between positive
23474 // and negative zero incorrectly.
23475 if (!DAG.getTarget().Options.UnsafeFPMath &&
23476 !DAG.isKnownNeverZero(LHS) && !DAG.isKnownNeverZero(RHS))
23478 Opcode = X86ISD::FMAX;
23481 // Converting this to a max would handle NaNs incorrectly, and swapping
23482 // the operands would cause it to handle comparisons between positive
23483 // and negative zero incorrectly.
23484 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)) {
23485 if (!DAG.getTarget().Options.UnsafeFPMath &&
23486 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
23488 std::swap(LHS, RHS);
23490 Opcode = X86ISD::FMAX;
23493 // Converting this to a max would handle both negative zeros and NaNs
23494 // incorrectly, but we can swap the operands to fix both.
23495 std::swap(LHS, RHS);
23499 Opcode = X86ISD::FMAX;
23502 // Check for x CC y ? y : x -- a min/max with reversed arms.
23503 } else if (DAG.isEqualTo(LHS, Cond.getOperand(1)) &&
23504 DAG.isEqualTo(RHS, Cond.getOperand(0))) {
23508 // Converting this to a min would handle comparisons between positive
23509 // and negative zero incorrectly, and swapping the operands would
23510 // cause it to handle NaNs incorrectly.
23511 if (!DAG.getTarget().Options.UnsafeFPMath &&
23512 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS))) {
23513 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS))
23515 std::swap(LHS, RHS);
23517 Opcode = X86ISD::FMIN;
23520 // Converting this to a min would handle NaNs incorrectly.
23521 if (!DAG.getTarget().Options.UnsafeFPMath &&
23522 (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)))
23524 Opcode = X86ISD::FMIN;
23527 // Converting this to a min would handle both negative zeros and NaNs
23528 // incorrectly, but we can swap the operands to fix both.
23529 std::swap(LHS, RHS);
23533 Opcode = X86ISD::FMIN;
23537 // Converting this to a max would handle NaNs incorrectly.
23538 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS))
23540 Opcode = X86ISD::FMAX;
23543 // Converting this to a max would handle comparisons between positive
23544 // and negative zero incorrectly, and swapping the operands would
23545 // cause it to handle NaNs incorrectly.
23546 if (!DAG.getTarget().Options.UnsafeFPMath &&
23547 !DAG.isKnownNeverZero(LHS) && !DAG.isKnownNeverZero(RHS)) {
23548 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS))
23550 std::swap(LHS, RHS);
23552 Opcode = X86ISD::FMAX;
23555 // Converting this to a max would handle both negative zeros and NaNs
23556 // incorrectly, but we can swap the operands to fix both.
23557 std::swap(LHS, RHS);
23561 Opcode = X86ISD::FMAX;
23567 return DAG.getNode(Opcode, DL, N->getValueType(0), LHS, RHS);
23570 EVT CondVT = Cond.getValueType();
23571 if (Subtarget->hasAVX512() && VT.isVector() && CondVT.isVector() &&
23572 CondVT.getVectorElementType() == MVT::i1) {
23573 // v16i8 (select v16i1, v16i8, v16i8) does not have a proper
23574 // lowering on KNL. In this case we convert it to
23575 // v16i8 (select v16i8, v16i8, v16i8) and use AVX instruction.
23576 // The same situation for all 128 and 256-bit vectors of i8 and i16.
23577 // Since SKX these selects have a proper lowering.
23578 EVT OpVT = LHS.getValueType();
23579 if ((OpVT.is128BitVector() || OpVT.is256BitVector()) &&
23580 (OpVT.getVectorElementType() == MVT::i8 ||
23581 OpVT.getVectorElementType() == MVT::i16) &&
23582 !(Subtarget->hasBWI() && Subtarget->hasVLX())) {
23583 Cond = DAG.getNode(ISD::SIGN_EXTEND, DL, OpVT, Cond);
23584 DCI.AddToWorklist(Cond.getNode());
23585 return DAG.getNode(N->getOpcode(), DL, OpVT, Cond, LHS, RHS);
23588 // If this is a select between two integer constants, try to do some
23590 if (ConstantSDNode *TrueC = dyn_cast<ConstantSDNode>(LHS)) {
23591 if (ConstantSDNode *FalseC = dyn_cast<ConstantSDNode>(RHS))
23592 // Don't do this for crazy integer types.
23593 if (DAG.getTargetLoweringInfo().isTypeLegal(LHS.getValueType())) {
23594 // If this is efficiently invertible, canonicalize the LHSC/RHSC values
23595 // so that TrueC (the true value) is larger than FalseC.
23596 bool NeedsCondInvert = false;
23598 if (TrueC->getAPIntValue().ult(FalseC->getAPIntValue()) &&
23599 // Efficiently invertible.
23600 (Cond.getOpcode() == ISD::SETCC || // setcc -> invertible.
23601 (Cond.getOpcode() == ISD::XOR && // xor(X, C) -> invertible.
23602 isa<ConstantSDNode>(Cond.getOperand(1))))) {
23603 NeedsCondInvert = true;
23604 std::swap(TrueC, FalseC);
23607 // Optimize C ? 8 : 0 -> zext(C) << 3. Likewise for any pow2/0.
23608 if (FalseC->getAPIntValue() == 0 &&
23609 TrueC->getAPIntValue().isPowerOf2()) {
23610 if (NeedsCondInvert) // Invert the condition if needed.
23611 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
23612 DAG.getConstant(1, DL, Cond.getValueType()));
23614 // Zero extend the condition if needed.
23615 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, LHS.getValueType(), Cond);
23617 unsigned ShAmt = TrueC->getAPIntValue().logBase2();
23618 return DAG.getNode(ISD::SHL, DL, LHS.getValueType(), Cond,
23619 DAG.getConstant(ShAmt, DL, MVT::i8));
23622 // Optimize Cond ? cst+1 : cst -> zext(setcc(C)+cst.
23623 if (FalseC->getAPIntValue()+1 == TrueC->getAPIntValue()) {
23624 if (NeedsCondInvert) // Invert the condition if needed.
23625 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
23626 DAG.getConstant(1, DL, Cond.getValueType()));
23628 // Zero extend the condition if needed.
23629 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL,
23630 FalseC->getValueType(0), Cond);
23631 return DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
23632 SDValue(FalseC, 0));
23635 // Optimize cases that will turn into an LEA instruction. This requires
23636 // an i32 or i64 and an efficient multiplier (1, 2, 3, 4, 5, 8, 9).
23637 if (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i64) {
23638 uint64_t Diff = TrueC->getZExtValue()-FalseC->getZExtValue();
23639 if (N->getValueType(0) == MVT::i32) Diff = (unsigned)Diff;
23641 bool isFastMultiplier = false;
23643 switch ((unsigned char)Diff) {
23645 case 1: // result = add base, cond
23646 case 2: // result = lea base( , cond*2)
23647 case 3: // result = lea base(cond, cond*2)
23648 case 4: // result = lea base( , cond*4)
23649 case 5: // result = lea base(cond, cond*4)
23650 case 8: // result = lea base( , cond*8)
23651 case 9: // result = lea base(cond, cond*8)
23652 isFastMultiplier = true;
23657 if (isFastMultiplier) {
23658 APInt Diff = TrueC->getAPIntValue()-FalseC->getAPIntValue();
23659 if (NeedsCondInvert) // Invert the condition if needed.
23660 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
23661 DAG.getConstant(1, DL, Cond.getValueType()));
23663 // Zero extend the condition if needed.
23664 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, FalseC->getValueType(0),
23666 // Scale the condition by the difference.
23668 Cond = DAG.getNode(ISD::MUL, DL, Cond.getValueType(), Cond,
23669 DAG.getConstant(Diff, DL,
23670 Cond.getValueType()));
23672 // Add the base if non-zero.
23673 if (FalseC->getAPIntValue() != 0)
23674 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
23675 SDValue(FalseC, 0));
23682 // Canonicalize max and min:
23683 // (x > y) ? x : y -> (x >= y) ? x : y
23684 // (x < y) ? x : y -> (x <= y) ? x : y
23685 // This allows use of COND_S / COND_NS (see TranslateX86CC) which eliminates
23686 // the need for an extra compare
23687 // against zero. e.g.
23688 // (x - y) > 0 : (x - y) ? 0 -> (x - y) >= 0 : (x - y) ? 0
23690 // testl %edi, %edi
23692 // cmovgl %edi, %eax
23696 // cmovsl %eax, %edi
23697 if (N->getOpcode() == ISD::SELECT && Cond.getOpcode() == ISD::SETCC &&
23698 DAG.isEqualTo(LHS, Cond.getOperand(0)) &&
23699 DAG.isEqualTo(RHS, Cond.getOperand(1))) {
23700 ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
23705 ISD::CondCode NewCC = (CC == ISD::SETLT) ? ISD::SETLE : ISD::SETGE;
23706 Cond = DAG.getSetCC(SDLoc(Cond), Cond.getValueType(),
23707 Cond.getOperand(0), Cond.getOperand(1), NewCC);
23708 return DAG.getNode(ISD::SELECT, DL, VT, Cond, LHS, RHS);
23713 // Early exit check
23714 if (!TLI.isTypeLegal(VT))
23717 // Match VSELECTs into subs with unsigned saturation.
23718 if (N->getOpcode() == ISD::VSELECT && Cond.getOpcode() == ISD::SETCC &&
23719 // psubus is available in SSE2 and AVX2 for i8 and i16 vectors.
23720 ((Subtarget->hasSSE2() && (VT == MVT::v16i8 || VT == MVT::v8i16)) ||
23721 (Subtarget->hasAVX2() && (VT == MVT::v32i8 || VT == MVT::v16i16)))) {
23722 ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
23724 // Check if one of the arms of the VSELECT is a zero vector. If it's on the
23725 // left side invert the predicate to simplify logic below.
23727 if (ISD::isBuildVectorAllZeros(LHS.getNode())) {
23729 CC = ISD::getSetCCInverse(CC, true);
23730 } else if (ISD::isBuildVectorAllZeros(RHS.getNode())) {
23734 if (Other.getNode() && Other->getNumOperands() == 2 &&
23735 DAG.isEqualTo(Other->getOperand(0), Cond.getOperand(0))) {
23736 SDValue OpLHS = Other->getOperand(0), OpRHS = Other->getOperand(1);
23737 SDValue CondRHS = Cond->getOperand(1);
23739 // Look for a general sub with unsigned saturation first.
23740 // x >= y ? x-y : 0 --> subus x, y
23741 // x > y ? x-y : 0 --> subus x, y
23742 if ((CC == ISD::SETUGE || CC == ISD::SETUGT) &&
23743 Other->getOpcode() == ISD::SUB && DAG.isEqualTo(OpRHS, CondRHS))
23744 return DAG.getNode(X86ISD::SUBUS, DL, VT, OpLHS, OpRHS);
23746 if (auto *OpRHSBV = dyn_cast<BuildVectorSDNode>(OpRHS))
23747 if (auto *OpRHSConst = OpRHSBV->getConstantSplatNode()) {
23748 if (auto *CondRHSBV = dyn_cast<BuildVectorSDNode>(CondRHS))
23749 if (auto *CondRHSConst = CondRHSBV->getConstantSplatNode())
23750 // If the RHS is a constant we have to reverse the const
23751 // canonicalization.
23752 // x > C-1 ? x+-C : 0 --> subus x, C
23753 if (CC == ISD::SETUGT && Other->getOpcode() == ISD::ADD &&
23754 CondRHSConst->getAPIntValue() ==
23755 (-OpRHSConst->getAPIntValue() - 1))
23756 return DAG.getNode(
23757 X86ISD::SUBUS, DL, VT, OpLHS,
23758 DAG.getConstant(-OpRHSConst->getAPIntValue(), DL, VT));
23760 // Another special case: If C was a sign bit, the sub has been
23761 // canonicalized into a xor.
23762 // FIXME: Would it be better to use computeKnownBits to determine
23763 // whether it's safe to decanonicalize the xor?
23764 // x s< 0 ? x^C : 0 --> subus x, C
23765 if (CC == ISD::SETLT && Other->getOpcode() == ISD::XOR &&
23766 ISD::isBuildVectorAllZeros(CondRHS.getNode()) &&
23767 OpRHSConst->getAPIntValue().isSignBit())
23768 // Note that we have to rebuild the RHS constant here to ensure we
23769 // don't rely on particular values of undef lanes.
23770 return DAG.getNode(
23771 X86ISD::SUBUS, DL, VT, OpLHS,
23772 DAG.getConstant(OpRHSConst->getAPIntValue(), DL, VT));
23777 // Simplify vector selection if condition value type matches vselect
23779 if (N->getOpcode() == ISD::VSELECT && CondVT == VT) {
23780 assert(Cond.getValueType().isVector() &&
23781 "vector select expects a vector selector!");
23783 bool TValIsAllOnes = ISD::isBuildVectorAllOnes(LHS.getNode());
23784 bool FValIsAllZeros = ISD::isBuildVectorAllZeros(RHS.getNode());
23786 // Try invert the condition if true value is not all 1s and false value
23788 if (!TValIsAllOnes && !FValIsAllZeros &&
23789 // Check if the selector will be produced by CMPP*/PCMP*
23790 Cond.getOpcode() == ISD::SETCC &&
23791 // Check if SETCC has already been promoted
23792 TLI.getSetCCResultType(DAG.getDataLayout(), *DAG.getContext(), VT) ==
23794 bool TValIsAllZeros = ISD::isBuildVectorAllZeros(LHS.getNode());
23795 bool FValIsAllOnes = ISD::isBuildVectorAllOnes(RHS.getNode());
23797 if (TValIsAllZeros || FValIsAllOnes) {
23798 SDValue CC = Cond.getOperand(2);
23799 ISD::CondCode NewCC =
23800 ISD::getSetCCInverse(cast<CondCodeSDNode>(CC)->get(),
23801 Cond.getOperand(0).getValueType().isInteger());
23802 Cond = DAG.getSetCC(DL, CondVT, Cond.getOperand(0), Cond.getOperand(1), NewCC);
23803 std::swap(LHS, RHS);
23804 TValIsAllOnes = FValIsAllOnes;
23805 FValIsAllZeros = TValIsAllZeros;
23809 if (TValIsAllOnes || FValIsAllZeros) {
23812 if (TValIsAllOnes && FValIsAllZeros)
23814 else if (TValIsAllOnes)
23816 DAG.getNode(ISD::OR, DL, CondVT, Cond, DAG.getBitcast(CondVT, RHS));
23817 else if (FValIsAllZeros)
23818 Ret = DAG.getNode(ISD::AND, DL, CondVT, Cond,
23819 DAG.getBitcast(CondVT, LHS));
23821 return DAG.getBitcast(VT, Ret);
23825 // We should generate an X86ISD::BLENDI from a vselect if its argument
23826 // is a sign_extend_inreg of an any_extend of a BUILD_VECTOR of
23827 // constants. This specific pattern gets generated when we split a
23828 // selector for a 512 bit vector in a machine without AVX512 (but with
23829 // 256-bit vectors), during legalization:
23831 // (vselect (sign_extend (any_extend (BUILD_VECTOR)) i1) LHS RHS)
23833 // Iff we find this pattern and the build_vectors are built from
23834 // constants, we translate the vselect into a shuffle_vector that we
23835 // know will be matched by LowerVECTOR_SHUFFLEtoBlend.
23836 if ((N->getOpcode() == ISD::VSELECT ||
23837 N->getOpcode() == X86ISD::SHRUNKBLEND) &&
23838 !DCI.isBeforeLegalize() && !VT.is512BitVector()) {
23839 SDValue Shuffle = transformVSELECTtoBlendVECTOR_SHUFFLE(N, DAG, Subtarget);
23840 if (Shuffle.getNode())
23844 // If this is a *dynamic* select (non-constant condition) and we can match
23845 // this node with one of the variable blend instructions, restructure the
23846 // condition so that the blends can use the high bit of each element and use
23847 // SimplifyDemandedBits to simplify the condition operand.
23848 if (N->getOpcode() == ISD::VSELECT && DCI.isBeforeLegalizeOps() &&
23849 !DCI.isBeforeLegalize() &&
23850 !ISD::isBuildVectorOfConstantSDNodes(Cond.getNode())) {
23851 unsigned BitWidth = Cond.getValueType().getScalarType().getSizeInBits();
23853 // Don't optimize vector selects that map to mask-registers.
23857 // We can only handle the cases where VSELECT is directly legal on the
23858 // subtarget. We custom lower VSELECT nodes with constant conditions and
23859 // this makes it hard to see whether a dynamic VSELECT will correctly
23860 // lower, so we both check the operation's status and explicitly handle the
23861 // cases where a *dynamic* blend will fail even though a constant-condition
23862 // blend could be custom lowered.
23863 // FIXME: We should find a better way to handle this class of problems.
23864 // Potentially, we should combine constant-condition vselect nodes
23865 // pre-legalization into shuffles and not mark as many types as custom
23867 if (!TLI.isOperationLegalOrCustom(ISD::VSELECT, VT))
23869 // FIXME: We don't support i16-element blends currently. We could and
23870 // should support them by making *all* the bits in the condition be set
23871 // rather than just the high bit and using an i8-element blend.
23872 if (VT.getScalarType() == MVT::i16)
23874 // Dynamic blending was only available from SSE4.1 onward.
23875 if (VT.getSizeInBits() == 128 && !Subtarget->hasSSE41())
23877 // Byte blends are only available in AVX2
23878 if (VT.getSizeInBits() == 256 && VT.getScalarType() == MVT::i8 &&
23879 !Subtarget->hasAVX2())
23882 assert(BitWidth >= 8 && BitWidth <= 64 && "Invalid mask size");
23883 APInt DemandedMask = APInt::getHighBitsSet(BitWidth, 1);
23885 APInt KnownZero, KnownOne;
23886 TargetLowering::TargetLoweringOpt TLO(DAG, DCI.isBeforeLegalize(),
23887 DCI.isBeforeLegalizeOps());
23888 if (TLO.ShrinkDemandedConstant(Cond, DemandedMask) ||
23889 TLI.SimplifyDemandedBits(Cond, DemandedMask, KnownZero, KnownOne,
23891 // If we changed the computation somewhere in the DAG, this change
23892 // will affect all users of Cond.
23893 // Make sure it is fine and update all the nodes so that we do not
23894 // use the generic VSELECT anymore. Otherwise, we may perform
23895 // wrong optimizations as we messed up with the actual expectation
23896 // for the vector boolean values.
23897 if (Cond != TLO.Old) {
23898 // Check all uses of that condition operand to check whether it will be
23899 // consumed by non-BLEND instructions, which may depend on all bits are
23901 for (SDNode::use_iterator I = Cond->use_begin(), E = Cond->use_end();
23903 if (I->getOpcode() != ISD::VSELECT)
23904 // TODO: Add other opcodes eventually lowered into BLEND.
23907 // Update all the users of the condition, before committing the change,
23908 // so that the VSELECT optimizations that expect the correct vector
23909 // boolean value will not be triggered.
23910 for (SDNode::use_iterator I = Cond->use_begin(), E = Cond->use_end();
23912 DAG.ReplaceAllUsesOfValueWith(
23914 DAG.getNode(X86ISD::SHRUNKBLEND, SDLoc(*I), I->getValueType(0),
23915 Cond, I->getOperand(1), I->getOperand(2)));
23916 DCI.CommitTargetLoweringOpt(TLO);
23919 // At this point, only Cond is changed. Change the condition
23920 // just for N to keep the opportunity to optimize all other
23921 // users their own way.
23922 DAG.ReplaceAllUsesOfValueWith(
23924 DAG.getNode(X86ISD::SHRUNKBLEND, SDLoc(N), N->getValueType(0),
23925 TLO.New, N->getOperand(1), N->getOperand(2)));
23933 // Check whether a boolean test is testing a boolean value generated by
23934 // X86ISD::SETCC. If so, return the operand of that SETCC and proper condition
23937 // Simplify the following patterns:
23938 // (Op (CMP (SETCC Cond EFLAGS) 1) EQ) or
23939 // (Op (CMP (SETCC Cond EFLAGS) 0) NEQ)
23940 // to (Op EFLAGS Cond)
23942 // (Op (CMP (SETCC Cond EFLAGS) 0) EQ) or
23943 // (Op (CMP (SETCC Cond EFLAGS) 1) NEQ)
23944 // to (Op EFLAGS !Cond)
23946 // where Op could be BRCOND or CMOV.
23948 static SDValue checkBoolTestSetCCCombine(SDValue Cmp, X86::CondCode &CC) {
23949 // Quit if not CMP and SUB with its value result used.
23950 if (Cmp.getOpcode() != X86ISD::CMP &&
23951 (Cmp.getOpcode() != X86ISD::SUB || Cmp.getNode()->hasAnyUseOfValue(0)))
23954 // Quit if not used as a boolean value.
23955 if (CC != X86::COND_E && CC != X86::COND_NE)
23958 // Check CMP operands. One of them should be 0 or 1 and the other should be
23959 // an SetCC or extended from it.
23960 SDValue Op1 = Cmp.getOperand(0);
23961 SDValue Op2 = Cmp.getOperand(1);
23964 const ConstantSDNode* C = nullptr;
23965 bool needOppositeCond = (CC == X86::COND_E);
23966 bool checkAgainstTrue = false; // Is it a comparison against 1?
23968 if ((C = dyn_cast<ConstantSDNode>(Op1)))
23970 else if ((C = dyn_cast<ConstantSDNode>(Op2)))
23972 else // Quit if all operands are not constants.
23975 if (C->getZExtValue() == 1) {
23976 needOppositeCond = !needOppositeCond;
23977 checkAgainstTrue = true;
23978 } else if (C->getZExtValue() != 0)
23979 // Quit if the constant is neither 0 or 1.
23982 bool truncatedToBoolWithAnd = false;
23983 // Skip (zext $x), (trunc $x), or (and $x, 1) node.
23984 while (SetCC.getOpcode() == ISD::ZERO_EXTEND ||
23985 SetCC.getOpcode() == ISD::TRUNCATE ||
23986 SetCC.getOpcode() == ISD::AND) {
23987 if (SetCC.getOpcode() == ISD::AND) {
23989 ConstantSDNode *CS;
23990 if ((CS = dyn_cast<ConstantSDNode>(SetCC.getOperand(0))) &&
23991 CS->getZExtValue() == 1)
23993 if ((CS = dyn_cast<ConstantSDNode>(SetCC.getOperand(1))) &&
23994 CS->getZExtValue() == 1)
23998 SetCC = SetCC.getOperand(OpIdx);
23999 truncatedToBoolWithAnd = true;
24001 SetCC = SetCC.getOperand(0);
24004 switch (SetCC.getOpcode()) {
24005 case X86ISD::SETCC_CARRY:
24006 // Since SETCC_CARRY gives output based on R = CF ? ~0 : 0, it's unsafe to
24007 // simplify it if the result of SETCC_CARRY is not canonicalized to 0 or 1,
24008 // i.e. it's a comparison against true but the result of SETCC_CARRY is not
24009 // truncated to i1 using 'and'.
24010 if (checkAgainstTrue && !truncatedToBoolWithAnd)
24012 assert(X86::CondCode(SetCC.getConstantOperandVal(0)) == X86::COND_B &&
24013 "Invalid use of SETCC_CARRY!");
24015 case X86ISD::SETCC:
24016 // Set the condition code or opposite one if necessary.
24017 CC = X86::CondCode(SetCC.getConstantOperandVal(0));
24018 if (needOppositeCond)
24019 CC = X86::GetOppositeBranchCondition(CC);
24020 return SetCC.getOperand(1);
24021 case X86ISD::CMOV: {
24022 // Check whether false/true value has canonical one, i.e. 0 or 1.
24023 ConstantSDNode *FVal = dyn_cast<ConstantSDNode>(SetCC.getOperand(0));
24024 ConstantSDNode *TVal = dyn_cast<ConstantSDNode>(SetCC.getOperand(1));
24025 // Quit if true value is not a constant.
24028 // Quit if false value is not a constant.
24030 SDValue Op = SetCC.getOperand(0);
24031 // Skip 'zext' or 'trunc' node.
24032 if (Op.getOpcode() == ISD::ZERO_EXTEND ||
24033 Op.getOpcode() == ISD::TRUNCATE)
24034 Op = Op.getOperand(0);
24035 // A special case for rdrand/rdseed, where 0 is set if false cond is
24037 if ((Op.getOpcode() != X86ISD::RDRAND &&
24038 Op.getOpcode() != X86ISD::RDSEED) || Op.getResNo() != 0)
24041 // Quit if false value is not the constant 0 or 1.
24042 bool FValIsFalse = true;
24043 if (FVal && FVal->getZExtValue() != 0) {
24044 if (FVal->getZExtValue() != 1)
24046 // If FVal is 1, opposite cond is needed.
24047 needOppositeCond = !needOppositeCond;
24048 FValIsFalse = false;
24050 // Quit if TVal is not the constant opposite of FVal.
24051 if (FValIsFalse && TVal->getZExtValue() != 1)
24053 if (!FValIsFalse && TVal->getZExtValue() != 0)
24055 CC = X86::CondCode(SetCC.getConstantOperandVal(2));
24056 if (needOppositeCond)
24057 CC = X86::GetOppositeBranchCondition(CC);
24058 return SetCC.getOperand(3);
24065 /// Check whether Cond is an AND/OR of SETCCs off of the same EFLAGS.
24067 /// (X86or (X86setcc) (X86setcc))
24068 /// (X86cmp (and (X86setcc) (X86setcc)), 0)
24069 static bool checkBoolTestAndOrSetCCCombine(SDValue Cond, X86::CondCode &CC0,
24070 X86::CondCode &CC1, SDValue &Flags,
24072 if (Cond->getOpcode() == X86ISD::CMP) {
24073 ConstantSDNode *CondOp1C = dyn_cast<ConstantSDNode>(Cond->getOperand(1));
24074 if (!CondOp1C || !CondOp1C->isNullValue())
24077 Cond = Cond->getOperand(0);
24082 SDValue SetCC0, SetCC1;
24083 switch (Cond->getOpcode()) {
24084 default: return false;
24091 SetCC0 = Cond->getOperand(0);
24092 SetCC1 = Cond->getOperand(1);
24096 // Make sure we have SETCC nodes, using the same flags value.
24097 if (SetCC0.getOpcode() != X86ISD::SETCC ||
24098 SetCC1.getOpcode() != X86ISD::SETCC ||
24099 SetCC0->getOperand(1) != SetCC1->getOperand(1))
24102 CC0 = (X86::CondCode)SetCC0->getConstantOperandVal(0);
24103 CC1 = (X86::CondCode)SetCC1->getConstantOperandVal(0);
24104 Flags = SetCC0->getOperand(1);
24108 /// Optimize X86ISD::CMOV [LHS, RHS, CONDCODE (e.g. X86::COND_NE), CONDVAL]
24109 static SDValue PerformCMOVCombine(SDNode *N, SelectionDAG &DAG,
24110 TargetLowering::DAGCombinerInfo &DCI,
24111 const X86Subtarget *Subtarget) {
24114 // If the flag operand isn't dead, don't touch this CMOV.
24115 if (N->getNumValues() == 2 && !SDValue(N, 1).use_empty())
24118 SDValue FalseOp = N->getOperand(0);
24119 SDValue TrueOp = N->getOperand(1);
24120 X86::CondCode CC = (X86::CondCode)N->getConstantOperandVal(2);
24121 SDValue Cond = N->getOperand(3);
24123 if (CC == X86::COND_E || CC == X86::COND_NE) {
24124 switch (Cond.getOpcode()) {
24128 // If operand of BSR / BSF are proven never zero, then ZF cannot be set.
24129 if (DAG.isKnownNeverZero(Cond.getOperand(0)))
24130 return (CC == X86::COND_E) ? FalseOp : TrueOp;
24136 Flags = checkBoolTestSetCCCombine(Cond, CC);
24137 if (Flags.getNode() &&
24138 // Extra check as FCMOV only supports a subset of X86 cond.
24139 (FalseOp.getValueType() != MVT::f80 || hasFPCMov(CC))) {
24140 SDValue Ops[] = { FalseOp, TrueOp,
24141 DAG.getConstant(CC, DL, MVT::i8), Flags };
24142 return DAG.getNode(X86ISD::CMOV, DL, N->getVTList(), Ops);
24145 // If this is a select between two integer constants, try to do some
24146 // optimizations. Note that the operands are ordered the opposite of SELECT
24148 if (ConstantSDNode *TrueC = dyn_cast<ConstantSDNode>(TrueOp)) {
24149 if (ConstantSDNode *FalseC = dyn_cast<ConstantSDNode>(FalseOp)) {
24150 // Canonicalize the TrueC/FalseC values so that TrueC (the true value) is
24151 // larger than FalseC (the false value).
24152 if (TrueC->getAPIntValue().ult(FalseC->getAPIntValue())) {
24153 CC = X86::GetOppositeBranchCondition(CC);
24154 std::swap(TrueC, FalseC);
24155 std::swap(TrueOp, FalseOp);
24158 // Optimize C ? 8 : 0 -> zext(setcc(C)) << 3. Likewise for any pow2/0.
24159 // This is efficient for any integer data type (including i8/i16) and
24161 if (FalseC->getAPIntValue() == 0 && TrueC->getAPIntValue().isPowerOf2()) {
24162 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
24163 DAG.getConstant(CC, DL, MVT::i8), Cond);
24165 // Zero extend the condition if needed.
24166 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, TrueC->getValueType(0), Cond);
24168 unsigned ShAmt = TrueC->getAPIntValue().logBase2();
24169 Cond = DAG.getNode(ISD::SHL, DL, Cond.getValueType(), Cond,
24170 DAG.getConstant(ShAmt, DL, MVT::i8));
24171 if (N->getNumValues() == 2) // Dead flag value?
24172 return DCI.CombineTo(N, Cond, SDValue());
24176 // Optimize Cond ? cst+1 : cst -> zext(setcc(C)+cst. This is efficient
24177 // for any integer data type, including i8/i16.
24178 if (FalseC->getAPIntValue()+1 == TrueC->getAPIntValue()) {
24179 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
24180 DAG.getConstant(CC, DL, MVT::i8), Cond);
24182 // Zero extend the condition if needed.
24183 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL,
24184 FalseC->getValueType(0), Cond);
24185 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
24186 SDValue(FalseC, 0));
24188 if (N->getNumValues() == 2) // Dead flag value?
24189 return DCI.CombineTo(N, Cond, SDValue());
24193 // Optimize cases that will turn into an LEA instruction. This requires
24194 // an i32 or i64 and an efficient multiplier (1, 2, 3, 4, 5, 8, 9).
24195 if (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i64) {
24196 uint64_t Diff = TrueC->getZExtValue()-FalseC->getZExtValue();
24197 if (N->getValueType(0) == MVT::i32) Diff = (unsigned)Diff;
24199 bool isFastMultiplier = false;
24201 switch ((unsigned char)Diff) {
24203 case 1: // result = add base, cond
24204 case 2: // result = lea base( , cond*2)
24205 case 3: // result = lea base(cond, cond*2)
24206 case 4: // result = lea base( , cond*4)
24207 case 5: // result = lea base(cond, cond*4)
24208 case 8: // result = lea base( , cond*8)
24209 case 9: // result = lea base(cond, cond*8)
24210 isFastMultiplier = true;
24215 if (isFastMultiplier) {
24216 APInt Diff = TrueC->getAPIntValue()-FalseC->getAPIntValue();
24217 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
24218 DAG.getConstant(CC, DL, MVT::i8), Cond);
24219 // Zero extend the condition if needed.
24220 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, FalseC->getValueType(0),
24222 // Scale the condition by the difference.
24224 Cond = DAG.getNode(ISD::MUL, DL, Cond.getValueType(), Cond,
24225 DAG.getConstant(Diff, DL, Cond.getValueType()));
24227 // Add the base if non-zero.
24228 if (FalseC->getAPIntValue() != 0)
24229 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
24230 SDValue(FalseC, 0));
24231 if (N->getNumValues() == 2) // Dead flag value?
24232 return DCI.CombineTo(N, Cond, SDValue());
24239 // Handle these cases:
24240 // (select (x != c), e, c) -> select (x != c), e, x),
24241 // (select (x == c), c, e) -> select (x == c), x, e)
24242 // where the c is an integer constant, and the "select" is the combination
24243 // of CMOV and CMP.
24245 // The rationale for this change is that the conditional-move from a constant
24246 // needs two instructions, however, conditional-move from a register needs
24247 // only one instruction.
24249 // CAVEAT: By replacing a constant with a symbolic value, it may obscure
24250 // some instruction-combining opportunities. This opt needs to be
24251 // postponed as late as possible.
24253 if (!DCI.isBeforeLegalize() && !DCI.isBeforeLegalizeOps()) {
24254 // the DCI.xxxx conditions are provided to postpone the optimization as
24255 // late as possible.
24257 ConstantSDNode *CmpAgainst = nullptr;
24258 if ((Cond.getOpcode() == X86ISD::CMP || Cond.getOpcode() == X86ISD::SUB) &&
24259 (CmpAgainst = dyn_cast<ConstantSDNode>(Cond.getOperand(1))) &&
24260 !isa<ConstantSDNode>(Cond.getOperand(0))) {
24262 if (CC == X86::COND_NE &&
24263 CmpAgainst == dyn_cast<ConstantSDNode>(FalseOp)) {
24264 CC = X86::GetOppositeBranchCondition(CC);
24265 std::swap(TrueOp, FalseOp);
24268 if (CC == X86::COND_E &&
24269 CmpAgainst == dyn_cast<ConstantSDNode>(TrueOp)) {
24270 SDValue Ops[] = { FalseOp, Cond.getOperand(0),
24271 DAG.getConstant(CC, DL, MVT::i8), Cond };
24272 return DAG.getNode(X86ISD::CMOV, DL, N->getVTList (), Ops);
24277 // Fold and/or of setcc's to double CMOV:
24278 // (CMOV F, T, ((cc1 | cc2) != 0)) -> (CMOV (CMOV F, T, cc1), T, cc2)
24279 // (CMOV F, T, ((cc1 & cc2) != 0)) -> (CMOV (CMOV T, F, !cc1), F, !cc2)
24281 // This combine lets us generate:
24282 // cmovcc1 (jcc1 if we don't have CMOV)
24288 // cmovne (jne if we don't have CMOV)
24289 // When we can't use the CMOV instruction, it might increase branch
24291 // When we can use CMOV, or when there is no mispredict, this improves
24292 // throughput and reduces register pressure.
24294 if (CC == X86::COND_NE) {
24296 X86::CondCode CC0, CC1;
24298 if (checkBoolTestAndOrSetCCCombine(Cond, CC0, CC1, Flags, isAndSetCC)) {
24300 std::swap(FalseOp, TrueOp);
24301 CC0 = X86::GetOppositeBranchCondition(CC0);
24302 CC1 = X86::GetOppositeBranchCondition(CC1);
24305 SDValue LOps[] = {FalseOp, TrueOp, DAG.getConstant(CC0, DL, MVT::i8),
24307 SDValue LCMOV = DAG.getNode(X86ISD::CMOV, DL, N->getVTList(), LOps);
24308 SDValue Ops[] = {LCMOV, TrueOp, DAG.getConstant(CC1, DL, MVT::i8), Flags};
24309 SDValue CMOV = DAG.getNode(X86ISD::CMOV, DL, N->getVTList(), Ops);
24310 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), SDValue(CMOV.getNode(), 1));
24318 /// PerformMulCombine - Optimize a single multiply with constant into two
24319 /// in order to implement it with two cheaper instructions, e.g.
24320 /// LEA + SHL, LEA + LEA.
24321 static SDValue PerformMulCombine(SDNode *N, SelectionDAG &DAG,
24322 TargetLowering::DAGCombinerInfo &DCI) {
24323 // An imul is usually smaller than the alternative sequence.
24324 if (DAG.getMachineFunction().getFunction()->optForMinSize())
24327 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
24330 EVT VT = N->getValueType(0);
24331 if (VT != MVT::i64 && VT != MVT::i32)
24334 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1));
24337 uint64_t MulAmt = C->getZExtValue();
24338 if (isPowerOf2_64(MulAmt) || MulAmt == 3 || MulAmt == 5 || MulAmt == 9)
24341 uint64_t MulAmt1 = 0;
24342 uint64_t MulAmt2 = 0;
24343 if ((MulAmt % 9) == 0) {
24345 MulAmt2 = MulAmt / 9;
24346 } else if ((MulAmt % 5) == 0) {
24348 MulAmt2 = MulAmt / 5;
24349 } else if ((MulAmt % 3) == 0) {
24351 MulAmt2 = MulAmt / 3;
24354 (isPowerOf2_64(MulAmt2) || MulAmt2 == 3 || MulAmt2 == 5 || MulAmt2 == 9)){
24357 if (isPowerOf2_64(MulAmt2) &&
24358 !(N->hasOneUse() && N->use_begin()->getOpcode() == ISD::ADD))
24359 // If second multiplifer is pow2, issue it first. We want the multiply by
24360 // 3, 5, or 9 to be folded into the addressing mode unless the lone use
24362 std::swap(MulAmt1, MulAmt2);
24365 if (isPowerOf2_64(MulAmt1))
24366 NewMul = DAG.getNode(ISD::SHL, DL, VT, N->getOperand(0),
24367 DAG.getConstant(Log2_64(MulAmt1), DL, MVT::i8));
24369 NewMul = DAG.getNode(X86ISD::MUL_IMM, DL, VT, N->getOperand(0),
24370 DAG.getConstant(MulAmt1, DL, VT));
24372 if (isPowerOf2_64(MulAmt2))
24373 NewMul = DAG.getNode(ISD::SHL, DL, VT, NewMul,
24374 DAG.getConstant(Log2_64(MulAmt2), DL, MVT::i8));
24376 NewMul = DAG.getNode(X86ISD::MUL_IMM, DL, VT, NewMul,
24377 DAG.getConstant(MulAmt2, DL, VT));
24379 // Do not add new nodes to DAG combiner worklist.
24380 DCI.CombineTo(N, NewMul, false);
24385 static SDValue PerformSHLCombine(SDNode *N, SelectionDAG &DAG) {
24386 SDValue N0 = N->getOperand(0);
24387 SDValue N1 = N->getOperand(1);
24388 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
24389 EVT VT = N0.getValueType();
24391 // fold (shl (and (setcc_c), c1), c2) -> (and setcc_c, (c1 << c2))
24392 // since the result of setcc_c is all zero's or all ones.
24393 if (VT.isInteger() && !VT.isVector() &&
24394 N1C && N0.getOpcode() == ISD::AND &&
24395 N0.getOperand(1).getOpcode() == ISD::Constant) {
24396 SDValue N00 = N0.getOperand(0);
24397 APInt Mask = cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
24398 APInt ShAmt = N1C->getAPIntValue();
24399 Mask = Mask.shl(ShAmt);
24400 bool MaskOK = false;
24401 // We can handle cases concerning bit-widening nodes containing setcc_c if
24402 // we carefully interrogate the mask to make sure we are semantics
24404 // The transform is not safe if the result of C1 << C2 exceeds the bitwidth
24405 // of the underlying setcc_c operation if the setcc_c was zero extended.
24406 // Consider the following example:
24407 // zext(setcc_c) -> i32 0x0000FFFF
24408 // c1 -> i32 0x0000FFFF
24409 // c2 -> i32 0x00000001
24410 // (shl (and (setcc_c), c1), c2) -> i32 0x0001FFFE
24411 // (and setcc_c, (c1 << c2)) -> i32 0x0000FFFE
24412 if (N00.getOpcode() == X86ISD::SETCC_CARRY) {
24414 } else if (N00.getOpcode() == ISD::SIGN_EXTEND &&
24415 N00.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY) {
24417 } else if ((N00.getOpcode() == ISD::ZERO_EXTEND ||
24418 N00.getOpcode() == ISD::ANY_EXTEND) &&
24419 N00.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY) {
24420 MaskOK = Mask.isIntN(N00.getOperand(0).getValueSizeInBits());
24422 if (MaskOK && Mask != 0) {
24424 return DAG.getNode(ISD::AND, DL, VT, N00, DAG.getConstant(Mask, DL, VT));
24428 // Hardware support for vector shifts is sparse which makes us scalarize the
24429 // vector operations in many cases. Also, on sandybridge ADD is faster than
24431 // (shl V, 1) -> add V,V
24432 if (auto *N1BV = dyn_cast<BuildVectorSDNode>(N1))
24433 if (auto *N1SplatC = N1BV->getConstantSplatNode()) {
24434 assert(N0.getValueType().isVector() && "Invalid vector shift type");
24435 // We shift all of the values by one. In many cases we do not have
24436 // hardware support for this operation. This is better expressed as an ADD
24438 if (N1SplatC->getAPIntValue() == 1)
24439 return DAG.getNode(ISD::ADD, SDLoc(N), VT, N0, N0);
24445 /// \brief Returns a vector of 0s if the node in input is a vector logical
24446 /// shift by a constant amount which is known to be bigger than or equal
24447 /// to the vector element size in bits.
24448 static SDValue performShiftToAllZeros(SDNode *N, SelectionDAG &DAG,
24449 const X86Subtarget *Subtarget) {
24450 EVT VT = N->getValueType(0);
24452 if (VT != MVT::v2i64 && VT != MVT::v4i32 && VT != MVT::v8i16 &&
24453 (!Subtarget->hasInt256() ||
24454 (VT != MVT::v4i64 && VT != MVT::v8i32 && VT != MVT::v16i16)))
24457 SDValue Amt = N->getOperand(1);
24459 if (auto *AmtBV = dyn_cast<BuildVectorSDNode>(Amt))
24460 if (auto *AmtSplat = AmtBV->getConstantSplatNode()) {
24461 APInt ShiftAmt = AmtSplat->getAPIntValue();
24462 unsigned MaxAmount = VT.getVectorElementType().getSizeInBits();
24464 // SSE2/AVX2 logical shifts always return a vector of 0s
24465 // if the shift amount is bigger than or equal to
24466 // the element size. The constant shift amount will be
24467 // encoded as a 8-bit immediate.
24468 if (ShiftAmt.trunc(8).uge(MaxAmount))
24469 return getZeroVector(VT, Subtarget, DAG, DL);
24475 /// PerformShiftCombine - Combine shifts.
24476 static SDValue PerformShiftCombine(SDNode* N, SelectionDAG &DAG,
24477 TargetLowering::DAGCombinerInfo &DCI,
24478 const X86Subtarget *Subtarget) {
24479 if (N->getOpcode() == ISD::SHL)
24480 if (SDValue V = PerformSHLCombine(N, DAG))
24483 // Try to fold this logical shift into a zero vector.
24484 if (N->getOpcode() != ISD::SRA)
24485 if (SDValue V = performShiftToAllZeros(N, DAG, Subtarget))
24491 // CMPEQCombine - Recognize the distinctive (AND (setcc ...) (setcc ..))
24492 // where both setccs reference the same FP CMP, and rewrite for CMPEQSS
24493 // and friends. Likewise for OR -> CMPNEQSS.
24494 static SDValue CMPEQCombine(SDNode *N, SelectionDAG &DAG,
24495 TargetLowering::DAGCombinerInfo &DCI,
24496 const X86Subtarget *Subtarget) {
24499 // SSE1 supports CMP{eq|ne}SS, and SSE2 added CMP{eq|ne}SD, but
24500 // we're requiring SSE2 for both.
24501 if (Subtarget->hasSSE2() && isAndOrOfSetCCs(SDValue(N, 0U), opcode)) {
24502 SDValue N0 = N->getOperand(0);
24503 SDValue N1 = N->getOperand(1);
24504 SDValue CMP0 = N0->getOperand(1);
24505 SDValue CMP1 = N1->getOperand(1);
24508 // The SETCCs should both refer to the same CMP.
24509 if (CMP0.getOpcode() != X86ISD::CMP || CMP0 != CMP1)
24512 SDValue CMP00 = CMP0->getOperand(0);
24513 SDValue CMP01 = CMP0->getOperand(1);
24514 EVT VT = CMP00.getValueType();
24516 if (VT == MVT::f32 || VT == MVT::f64) {
24517 bool ExpectingFlags = false;
24518 // Check for any users that want flags:
24519 for (SDNode::use_iterator UI = N->use_begin(), UE = N->use_end();
24520 !ExpectingFlags && UI != UE; ++UI)
24521 switch (UI->getOpcode()) {
24526 ExpectingFlags = true;
24528 case ISD::CopyToReg:
24529 case ISD::SIGN_EXTEND:
24530 case ISD::ZERO_EXTEND:
24531 case ISD::ANY_EXTEND:
24535 if (!ExpectingFlags) {
24536 enum X86::CondCode cc0 = (enum X86::CondCode)N0.getConstantOperandVal(0);
24537 enum X86::CondCode cc1 = (enum X86::CondCode)N1.getConstantOperandVal(0);
24539 if (cc1 == X86::COND_E || cc1 == X86::COND_NE) {
24540 X86::CondCode tmp = cc0;
24545 if ((cc0 == X86::COND_E && cc1 == X86::COND_NP) ||
24546 (cc0 == X86::COND_NE && cc1 == X86::COND_P)) {
24547 // FIXME: need symbolic constants for these magic numbers.
24548 // See X86ATTInstPrinter.cpp:printSSECC().
24549 unsigned x86cc = (cc0 == X86::COND_E) ? 0 : 4;
24550 if (Subtarget->hasAVX512()) {
24551 SDValue FSetCC = DAG.getNode(X86ISD::FSETCC, DL, MVT::i1, CMP00,
24553 DAG.getConstant(x86cc, DL, MVT::i8));
24554 if (N->getValueType(0) != MVT::i1)
24555 return DAG.getNode(ISD::ZERO_EXTEND, DL, N->getValueType(0),
24559 SDValue OnesOrZeroesF = DAG.getNode(X86ISD::FSETCC, DL,
24560 CMP00.getValueType(), CMP00, CMP01,
24561 DAG.getConstant(x86cc, DL,
24564 bool is64BitFP = (CMP00.getValueType() == MVT::f64);
24565 MVT IntVT = is64BitFP ? MVT::i64 : MVT::i32;
24567 if (is64BitFP && !Subtarget->is64Bit()) {
24568 // On a 32-bit target, we cannot bitcast the 64-bit float to a
24569 // 64-bit integer, since that's not a legal type. Since
24570 // OnesOrZeroesF is all ones of all zeroes, we don't need all the
24571 // bits, but can do this little dance to extract the lowest 32 bits
24572 // and work with those going forward.
24573 SDValue Vector64 = DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, MVT::v2f64,
24575 SDValue Vector32 = DAG.getBitcast(MVT::v4f32, Vector64);
24576 OnesOrZeroesF = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::f32,
24577 Vector32, DAG.getIntPtrConstant(0, DL));
24581 SDValue OnesOrZeroesI = DAG.getBitcast(IntVT, OnesOrZeroesF);
24582 SDValue ANDed = DAG.getNode(ISD::AND, DL, IntVT, OnesOrZeroesI,
24583 DAG.getConstant(1, DL, IntVT));
24584 SDValue OneBitOfTruth = DAG.getNode(ISD::TRUNCATE, DL, MVT::i8,
24586 return OneBitOfTruth;
24594 /// CanFoldXORWithAllOnes - Test whether the XOR operand is a AllOnes vector
24595 /// so it can be folded inside ANDNP.
24596 static bool CanFoldXORWithAllOnes(const SDNode *N) {
24597 EVT VT = N->getValueType(0);
24599 // Match direct AllOnes for 128 and 256-bit vectors
24600 if (ISD::isBuildVectorAllOnes(N))
24603 // Look through a bit convert.
24604 if (N->getOpcode() == ISD::BITCAST)
24605 N = N->getOperand(0).getNode();
24607 // Sometimes the operand may come from a insert_subvector building a 256-bit
24609 if (VT.is256BitVector() &&
24610 N->getOpcode() == ISD::INSERT_SUBVECTOR) {
24611 SDValue V1 = N->getOperand(0);
24612 SDValue V2 = N->getOperand(1);
24614 if (V1.getOpcode() == ISD::INSERT_SUBVECTOR &&
24615 V1.getOperand(0).getOpcode() == ISD::UNDEF &&
24616 ISD::isBuildVectorAllOnes(V1.getOperand(1).getNode()) &&
24617 ISD::isBuildVectorAllOnes(V2.getNode()))
24624 // On AVX/AVX2 the type v8i1 is legalized to v8i16, which is an XMM sized
24625 // register. In most cases we actually compare or select YMM-sized registers
24626 // and mixing the two types creates horrible code. This method optimizes
24627 // some of the transition sequences.
24628 static SDValue WidenMaskArithmetic(SDNode *N, SelectionDAG &DAG,
24629 TargetLowering::DAGCombinerInfo &DCI,
24630 const X86Subtarget *Subtarget) {
24631 EVT VT = N->getValueType(0);
24632 if (!VT.is256BitVector())
24635 assert((N->getOpcode() == ISD::ANY_EXTEND ||
24636 N->getOpcode() == ISD::ZERO_EXTEND ||
24637 N->getOpcode() == ISD::SIGN_EXTEND) && "Invalid Node");
24639 SDValue Narrow = N->getOperand(0);
24640 EVT NarrowVT = Narrow->getValueType(0);
24641 if (!NarrowVT.is128BitVector())
24644 if (Narrow->getOpcode() != ISD::XOR &&
24645 Narrow->getOpcode() != ISD::AND &&
24646 Narrow->getOpcode() != ISD::OR)
24649 SDValue N0 = Narrow->getOperand(0);
24650 SDValue N1 = Narrow->getOperand(1);
24653 // The Left side has to be a trunc.
24654 if (N0.getOpcode() != ISD::TRUNCATE)
24657 // The type of the truncated inputs.
24658 EVT WideVT = N0->getOperand(0)->getValueType(0);
24662 // The right side has to be a 'trunc' or a constant vector.
24663 bool RHSTrunc = N1.getOpcode() == ISD::TRUNCATE;
24664 ConstantSDNode *RHSConstSplat = nullptr;
24665 if (auto *RHSBV = dyn_cast<BuildVectorSDNode>(N1))
24666 RHSConstSplat = RHSBV->getConstantSplatNode();
24667 if (!RHSTrunc && !RHSConstSplat)
24670 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
24672 if (!TLI.isOperationLegalOrPromote(Narrow->getOpcode(), WideVT))
24675 // Set N0 and N1 to hold the inputs to the new wide operation.
24676 N0 = N0->getOperand(0);
24677 if (RHSConstSplat) {
24678 N1 = DAG.getNode(ISD::ZERO_EXTEND, DL, WideVT.getScalarType(),
24679 SDValue(RHSConstSplat, 0));
24680 SmallVector<SDValue, 8> C(WideVT.getVectorNumElements(), N1);
24681 N1 = DAG.getNode(ISD::BUILD_VECTOR, DL, WideVT, C);
24682 } else if (RHSTrunc) {
24683 N1 = N1->getOperand(0);
24686 // Generate the wide operation.
24687 SDValue Op = DAG.getNode(Narrow->getOpcode(), DL, WideVT, N0, N1);
24688 unsigned Opcode = N->getOpcode();
24690 case ISD::ANY_EXTEND:
24692 case ISD::ZERO_EXTEND: {
24693 unsigned InBits = NarrowVT.getScalarType().getSizeInBits();
24694 APInt Mask = APInt::getAllOnesValue(InBits);
24695 Mask = Mask.zext(VT.getScalarType().getSizeInBits());
24696 return DAG.getNode(ISD::AND, DL, VT,
24697 Op, DAG.getConstant(Mask, DL, VT));
24699 case ISD::SIGN_EXTEND:
24700 return DAG.getNode(ISD::SIGN_EXTEND_INREG, DL, VT,
24701 Op, DAG.getValueType(NarrowVT));
24703 llvm_unreachable("Unexpected opcode");
24707 static SDValue VectorZextCombine(SDNode *N, SelectionDAG &DAG,
24708 TargetLowering::DAGCombinerInfo &DCI,
24709 const X86Subtarget *Subtarget) {
24710 SDValue N0 = N->getOperand(0);
24711 SDValue N1 = N->getOperand(1);
24714 // A vector zext_in_reg may be represented as a shuffle,
24715 // feeding into a bitcast (this represents anyext) feeding into
24716 // an and with a mask.
24717 // We'd like to try to combine that into a shuffle with zero
24718 // plus a bitcast, removing the and.
24719 if (N0.getOpcode() != ISD::BITCAST ||
24720 N0.getOperand(0).getOpcode() != ISD::VECTOR_SHUFFLE)
24723 // The other side of the AND should be a splat of 2^C, where C
24724 // is the number of bits in the source type.
24725 if (N1.getOpcode() == ISD::BITCAST)
24726 N1 = N1.getOperand(0);
24727 if (N1.getOpcode() != ISD::BUILD_VECTOR)
24729 BuildVectorSDNode *Vector = cast<BuildVectorSDNode>(N1);
24731 ShuffleVectorSDNode *Shuffle = cast<ShuffleVectorSDNode>(N0.getOperand(0));
24732 EVT SrcType = Shuffle->getValueType(0);
24734 // We expect a single-source shuffle
24735 if (Shuffle->getOperand(1)->getOpcode() != ISD::UNDEF)
24738 unsigned SrcSize = SrcType.getScalarSizeInBits();
24740 APInt SplatValue, SplatUndef;
24741 unsigned SplatBitSize;
24743 if (!Vector->isConstantSplat(SplatValue, SplatUndef,
24744 SplatBitSize, HasAnyUndefs))
24747 unsigned ResSize = N1.getValueType().getScalarSizeInBits();
24748 // Make sure the splat matches the mask we expect
24749 if (SplatBitSize > ResSize ||
24750 (SplatValue + 1).exactLogBase2() != (int)SrcSize)
24753 // Make sure the input and output size make sense
24754 if (SrcSize >= ResSize || ResSize % SrcSize)
24757 // We expect a shuffle of the form <0, u, u, u, 1, u, u, u...>
24758 // The number of u's between each two values depends on the ratio between
24759 // the source and dest type.
24760 unsigned ZextRatio = ResSize / SrcSize;
24761 bool IsZext = true;
24762 for (unsigned i = 0; i < SrcType.getVectorNumElements(); ++i) {
24763 if (i % ZextRatio) {
24764 if (Shuffle->getMaskElt(i) > 0) {
24770 if (Shuffle->getMaskElt(i) != (int)(i / ZextRatio)) {
24771 // Expected element number
24781 // Ok, perform the transformation - replace the shuffle with
24782 // a shuffle of the form <0, k, k, k, 1, k, k, k> with zero
24783 // (instead of undef) where the k elements come from the zero vector.
24784 SmallVector<int, 8> Mask;
24785 unsigned NumElems = SrcType.getVectorNumElements();
24786 for (unsigned i = 0; i < NumElems; ++i)
24788 Mask.push_back(NumElems);
24790 Mask.push_back(i / ZextRatio);
24792 SDValue NewShuffle = DAG.getVectorShuffle(Shuffle->getValueType(0), DL,
24793 Shuffle->getOperand(0), DAG.getConstant(0, DL, SrcType), Mask);
24794 return DAG.getBitcast(N0.getValueType(), NewShuffle);
24797 /// If both input operands of a logic op are being cast from floating point
24798 /// types, try to convert this into a floating point logic node to avoid
24799 /// unnecessary moves from SSE to integer registers.
24800 static SDValue convertIntLogicToFPLogic(SDNode *N, SelectionDAG &DAG,
24801 const X86Subtarget *Subtarget) {
24802 unsigned FPOpcode = ISD::DELETED_NODE;
24803 if (N->getOpcode() == ISD::AND)
24804 FPOpcode = X86ISD::FAND;
24805 else if (N->getOpcode() == ISD::OR)
24806 FPOpcode = X86ISD::FOR;
24807 else if (N->getOpcode() == ISD::XOR)
24808 FPOpcode = X86ISD::FXOR;
24810 assert(FPOpcode != ISD::DELETED_NODE &&
24811 "Unexpected input node for FP logic conversion");
24813 EVT VT = N->getValueType(0);
24814 SDValue N0 = N->getOperand(0);
24815 SDValue N1 = N->getOperand(1);
24817 if (N0.getOpcode() == ISD::BITCAST && N1.getOpcode() == ISD::BITCAST &&
24818 ((Subtarget->hasSSE1() && VT == MVT::i32) ||
24819 (Subtarget->hasSSE2() && VT == MVT::i64))) {
24820 SDValue N00 = N0.getOperand(0);
24821 SDValue N10 = N1.getOperand(0);
24822 EVT N00Type = N00.getValueType();
24823 EVT N10Type = N10.getValueType();
24824 if (N00Type.isFloatingPoint() && N10Type.isFloatingPoint()) {
24825 SDValue FPLogic = DAG.getNode(FPOpcode, DL, N00Type, N00, N10);
24826 return DAG.getBitcast(VT, FPLogic);
24832 static SDValue PerformAndCombine(SDNode *N, SelectionDAG &DAG,
24833 TargetLowering::DAGCombinerInfo &DCI,
24834 const X86Subtarget *Subtarget) {
24835 if (DCI.isBeforeLegalizeOps())
24838 if (SDValue Zext = VectorZextCombine(N, DAG, DCI, Subtarget))
24841 if (SDValue R = CMPEQCombine(N, DAG, DCI, Subtarget))
24844 if (SDValue FPLogic = convertIntLogicToFPLogic(N, DAG, Subtarget))
24847 EVT VT = N->getValueType(0);
24848 SDValue N0 = N->getOperand(0);
24849 SDValue N1 = N->getOperand(1);
24852 // Create BEXTR instructions
24853 // BEXTR is ((X >> imm) & (2**size-1))
24854 if (VT == MVT::i32 || VT == MVT::i64) {
24855 // Check for BEXTR.
24856 if ((Subtarget->hasBMI() || Subtarget->hasTBM()) &&
24857 (N0.getOpcode() == ISD::SRA || N0.getOpcode() == ISD::SRL)) {
24858 ConstantSDNode *MaskNode = dyn_cast<ConstantSDNode>(N1);
24859 ConstantSDNode *ShiftNode = dyn_cast<ConstantSDNode>(N0.getOperand(1));
24860 if (MaskNode && ShiftNode) {
24861 uint64_t Mask = MaskNode->getZExtValue();
24862 uint64_t Shift = ShiftNode->getZExtValue();
24863 if (isMask_64(Mask)) {
24864 uint64_t MaskSize = countPopulation(Mask);
24865 if (Shift + MaskSize <= VT.getSizeInBits())
24866 return DAG.getNode(X86ISD::BEXTR, DL, VT, N0.getOperand(0),
24867 DAG.getConstant(Shift | (MaskSize << 8), DL,
24876 // Want to form ANDNP nodes:
24877 // 1) In the hopes of then easily combining them with OR and AND nodes
24878 // to form PBLEND/PSIGN.
24879 // 2) To match ANDN packed intrinsics
24880 if (VT != MVT::v2i64 && VT != MVT::v4i64)
24883 // Check LHS for vnot
24884 if (N0.getOpcode() == ISD::XOR &&
24885 //ISD::isBuildVectorAllOnes(N0.getOperand(1).getNode()))
24886 CanFoldXORWithAllOnes(N0.getOperand(1).getNode()))
24887 return DAG.getNode(X86ISD::ANDNP, DL, VT, N0.getOperand(0), N1);
24889 // Check RHS for vnot
24890 if (N1.getOpcode() == ISD::XOR &&
24891 //ISD::isBuildVectorAllOnes(N1.getOperand(1).getNode()))
24892 CanFoldXORWithAllOnes(N1.getOperand(1).getNode()))
24893 return DAG.getNode(X86ISD::ANDNP, DL, VT, N1.getOperand(0), N0);
24898 static SDValue PerformOrCombine(SDNode *N, SelectionDAG &DAG,
24899 TargetLowering::DAGCombinerInfo &DCI,
24900 const X86Subtarget *Subtarget) {
24901 if (DCI.isBeforeLegalizeOps())
24904 if (SDValue R = CMPEQCombine(N, DAG, DCI, Subtarget))
24907 if (SDValue FPLogic = convertIntLogicToFPLogic(N, DAG, Subtarget))
24910 SDValue N0 = N->getOperand(0);
24911 SDValue N1 = N->getOperand(1);
24912 EVT VT = N->getValueType(0);
24914 // look for psign/blend
24915 if (VT == MVT::v2i64 || VT == MVT::v4i64) {
24916 if (!Subtarget->hasSSSE3() ||
24917 (VT == MVT::v4i64 && !Subtarget->hasInt256()))
24920 // Canonicalize pandn to RHS
24921 if (N0.getOpcode() == X86ISD::ANDNP)
24923 // or (and (m, y), (pandn m, x))
24924 if (N0.getOpcode() == ISD::AND && N1.getOpcode() == X86ISD::ANDNP) {
24925 SDValue Mask = N1.getOperand(0);
24926 SDValue X = N1.getOperand(1);
24928 if (N0.getOperand(0) == Mask)
24929 Y = N0.getOperand(1);
24930 if (N0.getOperand(1) == Mask)
24931 Y = N0.getOperand(0);
24933 // Check to see if the mask appeared in both the AND and ANDNP and
24937 // Validate that X, Y, and Mask are BIT_CONVERTS, and see through them.
24938 // Look through mask bitcast.
24939 if (Mask.getOpcode() == ISD::BITCAST)
24940 Mask = Mask.getOperand(0);
24941 if (X.getOpcode() == ISD::BITCAST)
24942 X = X.getOperand(0);
24943 if (Y.getOpcode() == ISD::BITCAST)
24944 Y = Y.getOperand(0);
24946 EVT MaskVT = Mask.getValueType();
24948 // Validate that the Mask operand is a vector sra node.
24949 // FIXME: what to do for bytes, since there is a psignb/pblendvb, but
24950 // there is no psrai.b
24951 unsigned EltBits = MaskVT.getVectorElementType().getSizeInBits();
24952 unsigned SraAmt = ~0;
24953 if (Mask.getOpcode() == ISD::SRA) {
24954 if (auto *AmtBV = dyn_cast<BuildVectorSDNode>(Mask.getOperand(1)))
24955 if (auto *AmtConst = AmtBV->getConstantSplatNode())
24956 SraAmt = AmtConst->getZExtValue();
24957 } else if (Mask.getOpcode() == X86ISD::VSRAI) {
24958 SDValue SraC = Mask.getOperand(1);
24959 SraAmt = cast<ConstantSDNode>(SraC)->getZExtValue();
24961 if ((SraAmt + 1) != EltBits)
24966 // Now we know we at least have a plendvb with the mask val. See if
24967 // we can form a psignb/w/d.
24968 // psign = x.type == y.type == mask.type && y = sub(0, x);
24969 if (Y.getOpcode() == ISD::SUB && Y.getOperand(1) == X &&
24970 ISD::isBuildVectorAllZeros(Y.getOperand(0).getNode()) &&
24971 X.getValueType() == MaskVT && Y.getValueType() == MaskVT) {
24972 assert((EltBits == 8 || EltBits == 16 || EltBits == 32) &&
24973 "Unsupported VT for PSIGN");
24974 Mask = DAG.getNode(X86ISD::PSIGN, DL, MaskVT, X, Mask.getOperand(0));
24975 return DAG.getBitcast(VT, Mask);
24977 // PBLENDVB only available on SSE 4.1
24978 if (!Subtarget->hasSSE41())
24981 EVT BlendVT = (VT == MVT::v4i64) ? MVT::v32i8 : MVT::v16i8;
24983 X = DAG.getBitcast(BlendVT, X);
24984 Y = DAG.getBitcast(BlendVT, Y);
24985 Mask = DAG.getBitcast(BlendVT, Mask);
24986 Mask = DAG.getNode(ISD::VSELECT, DL, BlendVT, Mask, Y, X);
24987 return DAG.getBitcast(VT, Mask);
24991 if (VT != MVT::i16 && VT != MVT::i32 && VT != MVT::i64)
24994 // fold (or (x << c) | (y >> (64 - c))) ==> (shld64 x, y, c)
24995 bool OptForSize = DAG.getMachineFunction().getFunction()->optForSize();
24997 // SHLD/SHRD instructions have lower register pressure, but on some
24998 // platforms they have higher latency than the equivalent
24999 // series of shifts/or that would otherwise be generated.
25000 // Don't fold (or (x << c) | (y >> (64 - c))) if SHLD/SHRD instructions
25001 // have higher latencies and we are not optimizing for size.
25002 if (!OptForSize && Subtarget->isSHLDSlow())
25005 if (N0.getOpcode() == ISD::SRL && N1.getOpcode() == ISD::SHL)
25007 if (N0.getOpcode() != ISD::SHL || N1.getOpcode() != ISD::SRL)
25009 if (!N0.hasOneUse() || !N1.hasOneUse())
25012 SDValue ShAmt0 = N0.getOperand(1);
25013 if (ShAmt0.getValueType() != MVT::i8)
25015 SDValue ShAmt1 = N1.getOperand(1);
25016 if (ShAmt1.getValueType() != MVT::i8)
25018 if (ShAmt0.getOpcode() == ISD::TRUNCATE)
25019 ShAmt0 = ShAmt0.getOperand(0);
25020 if (ShAmt1.getOpcode() == ISD::TRUNCATE)
25021 ShAmt1 = ShAmt1.getOperand(0);
25024 unsigned Opc = X86ISD::SHLD;
25025 SDValue Op0 = N0.getOperand(0);
25026 SDValue Op1 = N1.getOperand(0);
25027 if (ShAmt0.getOpcode() == ISD::SUB) {
25028 Opc = X86ISD::SHRD;
25029 std::swap(Op0, Op1);
25030 std::swap(ShAmt0, ShAmt1);
25033 unsigned Bits = VT.getSizeInBits();
25034 if (ShAmt1.getOpcode() == ISD::SUB) {
25035 SDValue Sum = ShAmt1.getOperand(0);
25036 if (ConstantSDNode *SumC = dyn_cast<ConstantSDNode>(Sum)) {
25037 SDValue ShAmt1Op1 = ShAmt1.getOperand(1);
25038 if (ShAmt1Op1.getNode()->getOpcode() == ISD::TRUNCATE)
25039 ShAmt1Op1 = ShAmt1Op1.getOperand(0);
25040 if (SumC->getSExtValue() == Bits && ShAmt1Op1 == ShAmt0)
25041 return DAG.getNode(Opc, DL, VT,
25043 DAG.getNode(ISD::TRUNCATE, DL,
25046 } else if (ConstantSDNode *ShAmt1C = dyn_cast<ConstantSDNode>(ShAmt1)) {
25047 ConstantSDNode *ShAmt0C = dyn_cast<ConstantSDNode>(ShAmt0);
25049 ShAmt0C->getSExtValue() + ShAmt1C->getSExtValue() == Bits)
25050 return DAG.getNode(Opc, DL, VT,
25051 N0.getOperand(0), N1.getOperand(0),
25052 DAG.getNode(ISD::TRUNCATE, DL,
25059 // Generate NEG and CMOV for integer abs.
25060 static SDValue performIntegerAbsCombine(SDNode *N, SelectionDAG &DAG) {
25061 EVT VT = N->getValueType(0);
25063 // Since X86 does not have CMOV for 8-bit integer, we don't convert
25064 // 8-bit integer abs to NEG and CMOV.
25065 if (VT.isInteger() && VT.getSizeInBits() == 8)
25068 SDValue N0 = N->getOperand(0);
25069 SDValue N1 = N->getOperand(1);
25072 // Check pattern of XOR(ADD(X,Y), Y) where Y is SRA(X, size(X)-1)
25073 // and change it to SUB and CMOV.
25074 if (VT.isInteger() && N->getOpcode() == ISD::XOR &&
25075 N0.getOpcode() == ISD::ADD &&
25076 N0.getOperand(1) == N1 &&
25077 N1.getOpcode() == ISD::SRA &&
25078 N1.getOperand(0) == N0.getOperand(0))
25079 if (ConstantSDNode *Y1C = dyn_cast<ConstantSDNode>(N1.getOperand(1)))
25080 if (Y1C->getAPIntValue() == VT.getSizeInBits()-1) {
25081 // Generate SUB & CMOV.
25082 SDValue Neg = DAG.getNode(X86ISD::SUB, DL, DAG.getVTList(VT, MVT::i32),
25083 DAG.getConstant(0, DL, VT), N0.getOperand(0));
25085 SDValue Ops[] = { N0.getOperand(0), Neg,
25086 DAG.getConstant(X86::COND_GE, DL, MVT::i8),
25087 SDValue(Neg.getNode(), 1) };
25088 return DAG.getNode(X86ISD::CMOV, DL, DAG.getVTList(VT, MVT::Glue), Ops);
25093 // Try to turn tests against the signbit in the form of:
25094 // XOR(TRUNCATE(SRL(X, size(X)-1)), 1)
25097 static SDValue foldXorTruncShiftIntoCmp(SDNode *N, SelectionDAG &DAG) {
25098 // This is only worth doing if the output type is i8.
25099 if (N->getValueType(0) != MVT::i8)
25102 SDValue N0 = N->getOperand(0);
25103 SDValue N1 = N->getOperand(1);
25105 // We should be performing an xor against a truncated shift.
25106 if (N0.getOpcode() != ISD::TRUNCATE || !N0.hasOneUse())
25109 // Make sure we are performing an xor against one.
25110 if (!isa<ConstantSDNode>(N1) || !cast<ConstantSDNode>(N1)->isOne())
25113 // SetCC on x86 zero extends so only act on this if it's a logical shift.
25114 SDValue Shift = N0.getOperand(0);
25115 if (Shift.getOpcode() != ISD::SRL || !Shift.hasOneUse())
25118 // Make sure we are truncating from one of i16, i32 or i64.
25119 EVT ShiftTy = Shift.getValueType();
25120 if (ShiftTy != MVT::i16 && ShiftTy != MVT::i32 && ShiftTy != MVT::i64)
25123 // Make sure the shift amount extracts the sign bit.
25124 if (!isa<ConstantSDNode>(Shift.getOperand(1)) ||
25125 Shift.getConstantOperandVal(1) != ShiftTy.getSizeInBits() - 1)
25128 // Create a greater-than comparison against -1.
25129 // N.B. Using SETGE against 0 works but we want a canonical looking
25130 // comparison, using SETGT matches up with what TranslateX86CC.
25132 SDValue ShiftOp = Shift.getOperand(0);
25133 EVT ShiftOpTy = ShiftOp.getValueType();
25134 SDValue Cond = DAG.getSetCC(DL, MVT::i8, ShiftOp,
25135 DAG.getConstant(-1, DL, ShiftOpTy), ISD::SETGT);
25139 static SDValue PerformXorCombine(SDNode *N, SelectionDAG &DAG,
25140 TargetLowering::DAGCombinerInfo &DCI,
25141 const X86Subtarget *Subtarget) {
25142 if (DCI.isBeforeLegalizeOps())
25145 if (SDValue RV = foldXorTruncShiftIntoCmp(N, DAG))
25148 if (Subtarget->hasCMov())
25149 if (SDValue RV = performIntegerAbsCombine(N, DAG))
25152 if (SDValue FPLogic = convertIntLogicToFPLogic(N, DAG, Subtarget))
25158 /// PerformLOADCombine - Do target-specific dag combines on LOAD nodes.
25159 static SDValue PerformLOADCombine(SDNode *N, SelectionDAG &DAG,
25160 TargetLowering::DAGCombinerInfo &DCI,
25161 const X86Subtarget *Subtarget) {
25162 LoadSDNode *Ld = cast<LoadSDNode>(N);
25163 EVT RegVT = Ld->getValueType(0);
25164 EVT MemVT = Ld->getMemoryVT();
25166 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
25168 // For chips with slow 32-byte unaligned loads, break the 32-byte operation
25169 // into two 16-byte operations.
25170 ISD::LoadExtType Ext = Ld->getExtensionType();
25172 unsigned AddressSpace = Ld->getAddressSpace();
25173 unsigned Alignment = Ld->getAlignment();
25174 if (RegVT.is256BitVector() && !DCI.isBeforeLegalizeOps() &&
25175 Ext == ISD::NON_EXTLOAD &&
25176 TLI.allowsMemoryAccess(*DAG.getContext(), DAG.getDataLayout(), RegVT,
25177 AddressSpace, Alignment, &Fast) && !Fast) {
25178 unsigned NumElems = RegVT.getVectorNumElements();
25182 SDValue Ptr = Ld->getBasePtr();
25183 SDValue Increment =
25184 DAG.getConstant(16, dl, TLI.getPointerTy(DAG.getDataLayout()));
25186 EVT HalfVT = EVT::getVectorVT(*DAG.getContext(), MemVT.getScalarType(),
25188 SDValue Load1 = DAG.getLoad(HalfVT, dl, Ld->getChain(), Ptr,
25189 Ld->getPointerInfo(), Ld->isVolatile(),
25190 Ld->isNonTemporal(), Ld->isInvariant(),
25192 Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr, Increment);
25193 SDValue Load2 = DAG.getLoad(HalfVT, dl, Ld->getChain(), Ptr,
25194 Ld->getPointerInfo(), Ld->isVolatile(),
25195 Ld->isNonTemporal(), Ld->isInvariant(),
25196 std::min(16U, Alignment));
25197 SDValue TF = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
25199 Load2.getValue(1));
25201 SDValue NewVec = DAG.getUNDEF(RegVT);
25202 NewVec = Insert128BitVector(NewVec, Load1, 0, DAG, dl);
25203 NewVec = Insert128BitVector(NewVec, Load2, NumElems/2, DAG, dl);
25204 return DCI.CombineTo(N, NewVec, TF, true);
25210 /// PerformMLOADCombine - Resolve extending loads
25211 static SDValue PerformMLOADCombine(SDNode *N, SelectionDAG &DAG,
25212 TargetLowering::DAGCombinerInfo &DCI,
25213 const X86Subtarget *Subtarget) {
25214 MaskedLoadSDNode *Mld = cast<MaskedLoadSDNode>(N);
25215 if (Mld->getExtensionType() != ISD::SEXTLOAD)
25218 EVT VT = Mld->getValueType(0);
25219 unsigned NumElems = VT.getVectorNumElements();
25220 EVT LdVT = Mld->getMemoryVT();
25223 assert(LdVT != VT && "Cannot extend to the same type");
25224 unsigned ToSz = VT.getVectorElementType().getSizeInBits();
25225 unsigned FromSz = LdVT.getVectorElementType().getSizeInBits();
25226 // From, To sizes and ElemCount must be pow of two
25227 assert (isPowerOf2_32(NumElems * FromSz * ToSz) &&
25228 "Unexpected size for extending masked load");
25230 unsigned SizeRatio = ToSz / FromSz;
25231 assert(SizeRatio * NumElems * FromSz == VT.getSizeInBits());
25233 // Create a type on which we perform the shuffle
25234 EVT WideVecVT = EVT::getVectorVT(*DAG.getContext(),
25235 LdVT.getScalarType(), NumElems*SizeRatio);
25236 assert(WideVecVT.getSizeInBits() == VT.getSizeInBits());
25238 // Convert Src0 value
25239 SDValue WideSrc0 = DAG.getBitcast(WideVecVT, Mld->getSrc0());
25240 if (Mld->getSrc0().getOpcode() != ISD::UNDEF) {
25241 SmallVector<int, 16> ShuffleVec(NumElems * SizeRatio, -1);
25242 for (unsigned i = 0; i != NumElems; ++i)
25243 ShuffleVec[i] = i * SizeRatio;
25245 // Can't shuffle using an illegal type.
25246 assert(DAG.getTargetLoweringInfo().isTypeLegal(WideVecVT) &&
25247 "WideVecVT should be legal");
25248 WideSrc0 = DAG.getVectorShuffle(WideVecVT, dl, WideSrc0,
25249 DAG.getUNDEF(WideVecVT), &ShuffleVec[0]);
25251 // Prepare the new mask
25253 SDValue Mask = Mld->getMask();
25254 if (Mask.getValueType() == VT) {
25255 // Mask and original value have the same type
25256 NewMask = DAG.getBitcast(WideVecVT, Mask);
25257 SmallVector<int, 16> ShuffleVec(NumElems * SizeRatio, -1);
25258 for (unsigned i = 0; i != NumElems; ++i)
25259 ShuffleVec[i] = i * SizeRatio;
25260 for (unsigned i = NumElems; i != NumElems*SizeRatio; ++i)
25261 ShuffleVec[i] = NumElems*SizeRatio;
25262 NewMask = DAG.getVectorShuffle(WideVecVT, dl, NewMask,
25263 DAG.getConstant(0, dl, WideVecVT),
25267 assert(Mask.getValueType().getVectorElementType() == MVT::i1);
25268 unsigned WidenNumElts = NumElems*SizeRatio;
25269 unsigned MaskNumElts = VT.getVectorNumElements();
25270 EVT NewMaskVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
25273 unsigned NumConcat = WidenNumElts / MaskNumElts;
25274 SmallVector<SDValue, 16> Ops(NumConcat);
25275 SDValue ZeroVal = DAG.getConstant(0, dl, Mask.getValueType());
25277 for (unsigned i = 1; i != NumConcat; ++i)
25280 NewMask = DAG.getNode(ISD::CONCAT_VECTORS, dl, NewMaskVT, Ops);
25283 SDValue WideLd = DAG.getMaskedLoad(WideVecVT, dl, Mld->getChain(),
25284 Mld->getBasePtr(), NewMask, WideSrc0,
25285 Mld->getMemoryVT(), Mld->getMemOperand(),
25287 SDValue NewVec = DAG.getNode(X86ISD::VSEXT, dl, VT, WideLd);
25288 return DCI.CombineTo(N, NewVec, WideLd.getValue(1), true);
25290 /// PerformMSTORECombine - Resolve truncating stores
25291 static SDValue PerformMSTORECombine(SDNode *N, SelectionDAG &DAG,
25292 const X86Subtarget *Subtarget) {
25293 MaskedStoreSDNode *Mst = cast<MaskedStoreSDNode>(N);
25294 if (!Mst->isTruncatingStore())
25297 EVT VT = Mst->getValue().getValueType();
25298 unsigned NumElems = VT.getVectorNumElements();
25299 EVT StVT = Mst->getMemoryVT();
25302 assert(StVT != VT && "Cannot truncate to the same type");
25303 unsigned FromSz = VT.getVectorElementType().getSizeInBits();
25304 unsigned ToSz = StVT.getVectorElementType().getSizeInBits();
25306 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
25308 // The truncating store is legal in some cases. For example
25309 // vpmovqb, vpmovqw, vpmovqd, vpmovdb, vpmovdw
25310 // are designated for truncate store.
25311 // In this case we don't need any further transformations.
25312 if (TLI.isTruncStoreLegal(VT, StVT))
25315 // From, To sizes and ElemCount must be pow of two
25316 assert (isPowerOf2_32(NumElems * FromSz * ToSz) &&
25317 "Unexpected size for truncating masked store");
25318 // We are going to use the original vector elt for storing.
25319 // Accumulated smaller vector elements must be a multiple of the store size.
25320 assert (((NumElems * FromSz) % ToSz) == 0 &&
25321 "Unexpected ratio for truncating masked store");
25323 unsigned SizeRatio = FromSz / ToSz;
25324 assert(SizeRatio * NumElems * ToSz == VT.getSizeInBits());
25326 // Create a type on which we perform the shuffle
25327 EVT WideVecVT = EVT::getVectorVT(*DAG.getContext(),
25328 StVT.getScalarType(), NumElems*SizeRatio);
25330 assert(WideVecVT.getSizeInBits() == VT.getSizeInBits());
25332 SDValue WideVec = DAG.getBitcast(WideVecVT, Mst->getValue());
25333 SmallVector<int, 16> ShuffleVec(NumElems * SizeRatio, -1);
25334 for (unsigned i = 0; i != NumElems; ++i)
25335 ShuffleVec[i] = i * SizeRatio;
25337 // Can't shuffle using an illegal type.
25338 assert(DAG.getTargetLoweringInfo().isTypeLegal(WideVecVT) &&
25339 "WideVecVT should be legal");
25341 SDValue TruncatedVal = DAG.getVectorShuffle(WideVecVT, dl, WideVec,
25342 DAG.getUNDEF(WideVecVT),
25346 SDValue Mask = Mst->getMask();
25347 if (Mask.getValueType() == VT) {
25348 // Mask and original value have the same type
25349 NewMask = DAG.getBitcast(WideVecVT, Mask);
25350 for (unsigned i = 0; i != NumElems; ++i)
25351 ShuffleVec[i] = i * SizeRatio;
25352 for (unsigned i = NumElems; i != NumElems*SizeRatio; ++i)
25353 ShuffleVec[i] = NumElems*SizeRatio;
25354 NewMask = DAG.getVectorShuffle(WideVecVT, dl, NewMask,
25355 DAG.getConstant(0, dl, WideVecVT),
25359 assert(Mask.getValueType().getVectorElementType() == MVT::i1);
25360 unsigned WidenNumElts = NumElems*SizeRatio;
25361 unsigned MaskNumElts = VT.getVectorNumElements();
25362 EVT NewMaskVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
25365 unsigned NumConcat = WidenNumElts / MaskNumElts;
25366 SmallVector<SDValue, 16> Ops(NumConcat);
25367 SDValue ZeroVal = DAG.getConstant(0, dl, Mask.getValueType());
25369 for (unsigned i = 1; i != NumConcat; ++i)
25372 NewMask = DAG.getNode(ISD::CONCAT_VECTORS, dl, NewMaskVT, Ops);
25375 return DAG.getMaskedStore(Mst->getChain(), dl, TruncatedVal, Mst->getBasePtr(),
25376 NewMask, StVT, Mst->getMemOperand(), false);
25378 /// PerformSTORECombine - Do target-specific dag combines on STORE nodes.
25379 static SDValue PerformSTORECombine(SDNode *N, SelectionDAG &DAG,
25380 const X86Subtarget *Subtarget) {
25381 StoreSDNode *St = cast<StoreSDNode>(N);
25382 EVT VT = St->getValue().getValueType();
25383 EVT StVT = St->getMemoryVT();
25385 SDValue StoredVal = St->getOperand(1);
25386 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
25388 // If we are saving a concatenation of two XMM registers and 32-byte stores
25389 // are slow, such as on Sandy Bridge, perform two 16-byte stores.
25391 unsigned AddressSpace = St->getAddressSpace();
25392 unsigned Alignment = St->getAlignment();
25393 if (VT.is256BitVector() && StVT == VT &&
25394 TLI.allowsMemoryAccess(*DAG.getContext(), DAG.getDataLayout(), VT,
25395 AddressSpace, Alignment, &Fast) && !Fast) {
25396 unsigned NumElems = VT.getVectorNumElements();
25400 SDValue Value0 = Extract128BitVector(StoredVal, 0, DAG, dl);
25401 SDValue Value1 = Extract128BitVector(StoredVal, NumElems/2, DAG, dl);
25404 DAG.getConstant(16, dl, TLI.getPointerTy(DAG.getDataLayout()));
25405 SDValue Ptr0 = St->getBasePtr();
25406 SDValue Ptr1 = DAG.getNode(ISD::ADD, dl, Ptr0.getValueType(), Ptr0, Stride);
25408 SDValue Ch0 = DAG.getStore(St->getChain(), dl, Value0, Ptr0,
25409 St->getPointerInfo(), St->isVolatile(),
25410 St->isNonTemporal(), Alignment);
25411 SDValue Ch1 = DAG.getStore(St->getChain(), dl, Value1, Ptr1,
25412 St->getPointerInfo(), St->isVolatile(),
25413 St->isNonTemporal(),
25414 std::min(16U, Alignment));
25415 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Ch0, Ch1);
25418 // Optimize trunc store (of multiple scalars) to shuffle and store.
25419 // First, pack all of the elements in one place. Next, store to memory
25420 // in fewer chunks.
25421 if (St->isTruncatingStore() && VT.isVector()) {
25422 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
25423 unsigned NumElems = VT.getVectorNumElements();
25424 assert(StVT != VT && "Cannot truncate to the same type");
25425 unsigned FromSz = VT.getVectorElementType().getSizeInBits();
25426 unsigned ToSz = StVT.getVectorElementType().getSizeInBits();
25428 // The truncating store is legal in some cases. For example
25429 // vpmovqb, vpmovqw, vpmovqd, vpmovdb, vpmovdw
25430 // are designated for truncate store.
25431 // In this case we don't need any further transformations.
25432 if (TLI.isTruncStoreLegal(VT, StVT))
25435 // From, To sizes and ElemCount must be pow of two
25436 if (!isPowerOf2_32(NumElems * FromSz * ToSz)) return SDValue();
25437 // We are going to use the original vector elt for storing.
25438 // Accumulated smaller vector elements must be a multiple of the store size.
25439 if (0 != (NumElems * FromSz) % ToSz) return SDValue();
25441 unsigned SizeRatio = FromSz / ToSz;
25443 assert(SizeRatio * NumElems * ToSz == VT.getSizeInBits());
25445 // Create a type on which we perform the shuffle
25446 EVT WideVecVT = EVT::getVectorVT(*DAG.getContext(),
25447 StVT.getScalarType(), NumElems*SizeRatio);
25449 assert(WideVecVT.getSizeInBits() == VT.getSizeInBits());
25451 SDValue WideVec = DAG.getBitcast(WideVecVT, St->getValue());
25452 SmallVector<int, 8> ShuffleVec(NumElems * SizeRatio, -1);
25453 for (unsigned i = 0; i != NumElems; ++i)
25454 ShuffleVec[i] = i * SizeRatio;
25456 // Can't shuffle using an illegal type.
25457 if (!TLI.isTypeLegal(WideVecVT))
25460 SDValue Shuff = DAG.getVectorShuffle(WideVecVT, dl, WideVec,
25461 DAG.getUNDEF(WideVecVT),
25463 // At this point all of the data is stored at the bottom of the
25464 // register. We now need to save it to mem.
25466 // Find the largest store unit
25467 MVT StoreType = MVT::i8;
25468 for (MVT Tp : MVT::integer_valuetypes()) {
25469 if (TLI.isTypeLegal(Tp) && Tp.getSizeInBits() <= NumElems * ToSz)
25473 // On 32bit systems, we can't save 64bit integers. Try bitcasting to F64.
25474 if (TLI.isTypeLegal(MVT::f64) && StoreType.getSizeInBits() < 64 &&
25475 (64 <= NumElems * ToSz))
25476 StoreType = MVT::f64;
25478 // Bitcast the original vector into a vector of store-size units
25479 EVT StoreVecVT = EVT::getVectorVT(*DAG.getContext(),
25480 StoreType, VT.getSizeInBits()/StoreType.getSizeInBits());
25481 assert(StoreVecVT.getSizeInBits() == VT.getSizeInBits());
25482 SDValue ShuffWide = DAG.getBitcast(StoreVecVT, Shuff);
25483 SmallVector<SDValue, 8> Chains;
25484 SDValue Increment = DAG.getConstant(StoreType.getSizeInBits() / 8, dl,
25485 TLI.getPointerTy(DAG.getDataLayout()));
25486 SDValue Ptr = St->getBasePtr();
25488 // Perform one or more big stores into memory.
25489 for (unsigned i=0, e=(ToSz*NumElems)/StoreType.getSizeInBits(); i!=e; ++i) {
25490 SDValue SubVec = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl,
25491 StoreType, ShuffWide,
25492 DAG.getIntPtrConstant(i, dl));
25493 SDValue Ch = DAG.getStore(St->getChain(), dl, SubVec, Ptr,
25494 St->getPointerInfo(), St->isVolatile(),
25495 St->isNonTemporal(), St->getAlignment());
25496 Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr, Increment);
25497 Chains.push_back(Ch);
25500 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Chains);
25503 // Turn load->store of MMX types into GPR load/stores. This avoids clobbering
25504 // the FP state in cases where an emms may be missing.
25505 // A preferable solution to the general problem is to figure out the right
25506 // places to insert EMMS. This qualifies as a quick hack.
25508 // Similarly, turn load->store of i64 into double load/stores in 32-bit mode.
25509 if (VT.getSizeInBits() != 64)
25512 const Function *F = DAG.getMachineFunction().getFunction();
25513 bool NoImplicitFloatOps = F->hasFnAttribute(Attribute::NoImplicitFloat);
25515 !Subtarget->useSoftFloat() && !NoImplicitFloatOps && Subtarget->hasSSE2();
25516 if ((VT.isVector() ||
25517 (VT == MVT::i64 && F64IsLegal && !Subtarget->is64Bit())) &&
25518 isa<LoadSDNode>(St->getValue()) &&
25519 !cast<LoadSDNode>(St->getValue())->isVolatile() &&
25520 St->getChain().hasOneUse() && !St->isVolatile()) {
25521 SDNode* LdVal = St->getValue().getNode();
25522 LoadSDNode *Ld = nullptr;
25523 int TokenFactorIndex = -1;
25524 SmallVector<SDValue, 8> Ops;
25525 SDNode* ChainVal = St->getChain().getNode();
25526 // Must be a store of a load. We currently handle two cases: the load
25527 // is a direct child, and it's under an intervening TokenFactor. It is
25528 // possible to dig deeper under nested TokenFactors.
25529 if (ChainVal == LdVal)
25530 Ld = cast<LoadSDNode>(St->getChain());
25531 else if (St->getValue().hasOneUse() &&
25532 ChainVal->getOpcode() == ISD::TokenFactor) {
25533 for (unsigned i = 0, e = ChainVal->getNumOperands(); i != e; ++i) {
25534 if (ChainVal->getOperand(i).getNode() == LdVal) {
25535 TokenFactorIndex = i;
25536 Ld = cast<LoadSDNode>(St->getValue());
25538 Ops.push_back(ChainVal->getOperand(i));
25542 if (!Ld || !ISD::isNormalLoad(Ld))
25545 // If this is not the MMX case, i.e. we are just turning i64 load/store
25546 // into f64 load/store, avoid the transformation if there are multiple
25547 // uses of the loaded value.
25548 if (!VT.isVector() && !Ld->hasNUsesOfValue(1, 0))
25553 // If we are a 64-bit capable x86, lower to a single movq load/store pair.
25554 // Otherwise, if it's legal to use f64 SSE instructions, use f64 load/store
25556 if (Subtarget->is64Bit() || F64IsLegal) {
25557 EVT LdVT = Subtarget->is64Bit() ? MVT::i64 : MVT::f64;
25558 SDValue NewLd = DAG.getLoad(LdVT, LdDL, Ld->getChain(), Ld->getBasePtr(),
25559 Ld->getPointerInfo(), Ld->isVolatile(),
25560 Ld->isNonTemporal(), Ld->isInvariant(),
25561 Ld->getAlignment());
25562 SDValue NewChain = NewLd.getValue(1);
25563 if (TokenFactorIndex != -1) {
25564 Ops.push_back(NewChain);
25565 NewChain = DAG.getNode(ISD::TokenFactor, LdDL, MVT::Other, Ops);
25567 return DAG.getStore(NewChain, StDL, NewLd, St->getBasePtr(),
25568 St->getPointerInfo(),
25569 St->isVolatile(), St->isNonTemporal(),
25570 St->getAlignment());
25573 // Otherwise, lower to two pairs of 32-bit loads / stores.
25574 SDValue LoAddr = Ld->getBasePtr();
25575 SDValue HiAddr = DAG.getNode(ISD::ADD, LdDL, MVT::i32, LoAddr,
25576 DAG.getConstant(4, LdDL, MVT::i32));
25578 SDValue LoLd = DAG.getLoad(MVT::i32, LdDL, Ld->getChain(), LoAddr,
25579 Ld->getPointerInfo(),
25580 Ld->isVolatile(), Ld->isNonTemporal(),
25581 Ld->isInvariant(), Ld->getAlignment());
25582 SDValue HiLd = DAG.getLoad(MVT::i32, LdDL, Ld->getChain(), HiAddr,
25583 Ld->getPointerInfo().getWithOffset(4),
25584 Ld->isVolatile(), Ld->isNonTemporal(),
25586 MinAlign(Ld->getAlignment(), 4));
25588 SDValue NewChain = LoLd.getValue(1);
25589 if (TokenFactorIndex != -1) {
25590 Ops.push_back(LoLd);
25591 Ops.push_back(HiLd);
25592 NewChain = DAG.getNode(ISD::TokenFactor, LdDL, MVT::Other, Ops);
25595 LoAddr = St->getBasePtr();
25596 HiAddr = DAG.getNode(ISD::ADD, StDL, MVT::i32, LoAddr,
25597 DAG.getConstant(4, StDL, MVT::i32));
25599 SDValue LoSt = DAG.getStore(NewChain, StDL, LoLd, LoAddr,
25600 St->getPointerInfo(),
25601 St->isVolatile(), St->isNonTemporal(),
25602 St->getAlignment());
25603 SDValue HiSt = DAG.getStore(NewChain, StDL, HiLd, HiAddr,
25604 St->getPointerInfo().getWithOffset(4),
25606 St->isNonTemporal(),
25607 MinAlign(St->getAlignment(), 4));
25608 return DAG.getNode(ISD::TokenFactor, StDL, MVT::Other, LoSt, HiSt);
25611 // This is similar to the above case, but here we handle a scalar 64-bit
25612 // integer store that is extracted from a vector on a 32-bit target.
25613 // If we have SSE2, then we can treat it like a floating-point double
25614 // to get past legalization. The execution dependencies fixup pass will
25615 // choose the optimal machine instruction for the store if this really is
25616 // an integer or v2f32 rather than an f64.
25617 if (VT == MVT::i64 && F64IsLegal && !Subtarget->is64Bit() &&
25618 St->getOperand(1).getOpcode() == ISD::EXTRACT_VECTOR_ELT) {
25619 SDValue OldExtract = St->getOperand(1);
25620 SDValue ExtOp0 = OldExtract.getOperand(0);
25621 unsigned VecSize = ExtOp0.getValueSizeInBits();
25622 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), MVT::f64, VecSize / 64);
25623 SDValue BitCast = DAG.getBitcast(VecVT, ExtOp0);
25624 SDValue NewExtract = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
25625 BitCast, OldExtract.getOperand(1));
25626 return DAG.getStore(St->getChain(), dl, NewExtract, St->getBasePtr(),
25627 St->getPointerInfo(), St->isVolatile(),
25628 St->isNonTemporal(), St->getAlignment());
25634 /// Return 'true' if this vector operation is "horizontal"
25635 /// and return the operands for the horizontal operation in LHS and RHS. A
25636 /// horizontal operation performs the binary operation on successive elements
25637 /// of its first operand, then on successive elements of its second operand,
25638 /// returning the resulting values in a vector. For example, if
25639 /// A = < float a0, float a1, float a2, float a3 >
25641 /// B = < float b0, float b1, float b2, float b3 >
25642 /// then the result of doing a horizontal operation on A and B is
25643 /// A horizontal-op B = < a0 op a1, a2 op a3, b0 op b1, b2 op b3 >.
25644 /// In short, LHS and RHS are inspected to see if LHS op RHS is of the form
25645 /// A horizontal-op B, for some already available A and B, and if so then LHS is
25646 /// set to A, RHS to B, and the routine returns 'true'.
25647 /// Note that the binary operation should have the property that if one of the
25648 /// operands is UNDEF then the result is UNDEF.
25649 static bool isHorizontalBinOp(SDValue &LHS, SDValue &RHS, bool IsCommutative) {
25650 // Look for the following pattern: if
25651 // A = < float a0, float a1, float a2, float a3 >
25652 // B = < float b0, float b1, float b2, float b3 >
25654 // LHS = VECTOR_SHUFFLE A, B, <0, 2, 4, 6>
25655 // RHS = VECTOR_SHUFFLE A, B, <1, 3, 5, 7>
25656 // then LHS op RHS = < a0 op a1, a2 op a3, b0 op b1, b2 op b3 >
25657 // which is A horizontal-op B.
25659 // At least one of the operands should be a vector shuffle.
25660 if (LHS.getOpcode() != ISD::VECTOR_SHUFFLE &&
25661 RHS.getOpcode() != ISD::VECTOR_SHUFFLE)
25664 MVT VT = LHS.getSimpleValueType();
25666 assert((VT.is128BitVector() || VT.is256BitVector()) &&
25667 "Unsupported vector type for horizontal add/sub");
25669 // Handle 128 and 256-bit vector lengths. AVX defines horizontal add/sub to
25670 // operate independently on 128-bit lanes.
25671 unsigned NumElts = VT.getVectorNumElements();
25672 unsigned NumLanes = VT.getSizeInBits()/128;
25673 unsigned NumLaneElts = NumElts / NumLanes;
25674 assert((NumLaneElts % 2 == 0) &&
25675 "Vector type should have an even number of elements in each lane");
25676 unsigned HalfLaneElts = NumLaneElts/2;
25678 // View LHS in the form
25679 // LHS = VECTOR_SHUFFLE A, B, LMask
25680 // If LHS is not a shuffle then pretend it is the shuffle
25681 // LHS = VECTOR_SHUFFLE LHS, undef, <0, 1, ..., N-1>
25682 // NOTE: in what follows a default initialized SDValue represents an UNDEF of
25685 SmallVector<int, 16> LMask(NumElts);
25686 if (LHS.getOpcode() == ISD::VECTOR_SHUFFLE) {
25687 if (LHS.getOperand(0).getOpcode() != ISD::UNDEF)
25688 A = LHS.getOperand(0);
25689 if (LHS.getOperand(1).getOpcode() != ISD::UNDEF)
25690 B = LHS.getOperand(1);
25691 ArrayRef<int> Mask = cast<ShuffleVectorSDNode>(LHS.getNode())->getMask();
25692 std::copy(Mask.begin(), Mask.end(), LMask.begin());
25694 if (LHS.getOpcode() != ISD::UNDEF)
25696 for (unsigned i = 0; i != NumElts; ++i)
25700 // Likewise, view RHS in the form
25701 // RHS = VECTOR_SHUFFLE C, D, RMask
25703 SmallVector<int, 16> RMask(NumElts);
25704 if (RHS.getOpcode() == ISD::VECTOR_SHUFFLE) {
25705 if (RHS.getOperand(0).getOpcode() != ISD::UNDEF)
25706 C = RHS.getOperand(0);
25707 if (RHS.getOperand(1).getOpcode() != ISD::UNDEF)
25708 D = RHS.getOperand(1);
25709 ArrayRef<int> Mask = cast<ShuffleVectorSDNode>(RHS.getNode())->getMask();
25710 std::copy(Mask.begin(), Mask.end(), RMask.begin());
25712 if (RHS.getOpcode() != ISD::UNDEF)
25714 for (unsigned i = 0; i != NumElts; ++i)
25718 // Check that the shuffles are both shuffling the same vectors.
25719 if (!(A == C && B == D) && !(A == D && B == C))
25722 // If everything is UNDEF then bail out: it would be better to fold to UNDEF.
25723 if (!A.getNode() && !B.getNode())
25726 // If A and B occur in reverse order in RHS, then "swap" them (which means
25727 // rewriting the mask).
25729 ShuffleVectorSDNode::commuteMask(RMask);
25731 // At this point LHS and RHS are equivalent to
25732 // LHS = VECTOR_SHUFFLE A, B, LMask
25733 // RHS = VECTOR_SHUFFLE A, B, RMask
25734 // Check that the masks correspond to performing a horizontal operation.
25735 for (unsigned l = 0; l != NumElts; l += NumLaneElts) {
25736 for (unsigned i = 0; i != NumLaneElts; ++i) {
25737 int LIdx = LMask[i+l], RIdx = RMask[i+l];
25739 // Ignore any UNDEF components.
25740 if (LIdx < 0 || RIdx < 0 ||
25741 (!A.getNode() && (LIdx < (int)NumElts || RIdx < (int)NumElts)) ||
25742 (!B.getNode() && (LIdx >= (int)NumElts || RIdx >= (int)NumElts)))
25745 // Check that successive elements are being operated on. If not, this is
25746 // not a horizontal operation.
25747 unsigned Src = (i/HalfLaneElts); // each lane is split between srcs
25748 int Index = 2*(i%HalfLaneElts) + NumElts*Src + l;
25749 if (!(LIdx == Index && RIdx == Index + 1) &&
25750 !(IsCommutative && LIdx == Index + 1 && RIdx == Index))
25755 LHS = A.getNode() ? A : B; // If A is 'UNDEF', use B for it.
25756 RHS = B.getNode() ? B : A; // If B is 'UNDEF', use A for it.
25760 /// Do target-specific dag combines on floating point adds.
25761 static SDValue PerformFADDCombine(SDNode *N, SelectionDAG &DAG,
25762 const X86Subtarget *Subtarget) {
25763 EVT VT = N->getValueType(0);
25764 SDValue LHS = N->getOperand(0);
25765 SDValue RHS = N->getOperand(1);
25767 // Try to synthesize horizontal adds from adds of shuffles.
25768 if (((Subtarget->hasSSE3() && (VT == MVT::v4f32 || VT == MVT::v2f64)) ||
25769 (Subtarget->hasFp256() && (VT == MVT::v8f32 || VT == MVT::v4f64))) &&
25770 isHorizontalBinOp(LHS, RHS, true))
25771 return DAG.getNode(X86ISD::FHADD, SDLoc(N), VT, LHS, RHS);
25775 /// Do target-specific dag combines on floating point subs.
25776 static SDValue PerformFSUBCombine(SDNode *N, SelectionDAG &DAG,
25777 const X86Subtarget *Subtarget) {
25778 EVT VT = N->getValueType(0);
25779 SDValue LHS = N->getOperand(0);
25780 SDValue RHS = N->getOperand(1);
25782 // Try to synthesize horizontal subs from subs of shuffles.
25783 if (((Subtarget->hasSSE3() && (VT == MVT::v4f32 || VT == MVT::v2f64)) ||
25784 (Subtarget->hasFp256() && (VT == MVT::v8f32 || VT == MVT::v4f64))) &&
25785 isHorizontalBinOp(LHS, RHS, false))
25786 return DAG.getNode(X86ISD::FHSUB, SDLoc(N), VT, LHS, RHS);
25790 /// Do target-specific dag combines on X86ISD::FOR and X86ISD::FXOR nodes.
25791 static SDValue PerformFORCombine(SDNode *N, SelectionDAG &DAG,
25792 const X86Subtarget *Subtarget) {
25793 assert(N->getOpcode() == X86ISD::FOR || N->getOpcode() == X86ISD::FXOR);
25795 // F[X]OR(0.0, x) -> x
25796 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
25797 if (C->getValueAPF().isPosZero())
25798 return N->getOperand(1);
25800 // F[X]OR(x, 0.0) -> x
25801 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
25802 if (C->getValueAPF().isPosZero())
25803 return N->getOperand(0);
25805 EVT VT = N->getValueType(0);
25806 if (VT.is512BitVector() && !Subtarget->hasDQI()) {
25808 MVT IntScalar = MVT::getIntegerVT(VT.getScalarSizeInBits());
25809 MVT IntVT = MVT::getVectorVT(IntScalar, VT.getVectorNumElements());
25811 SDValue Op0 = DAG.getNode(ISD::BITCAST, dl, IntVT, N->getOperand(0));
25812 SDValue Op1 = DAG.getNode(ISD::BITCAST, dl, IntVT, N->getOperand(1));
25813 unsigned IntOpcode = (N->getOpcode() == X86ISD::FOR) ? ISD::OR : ISD::XOR;
25814 SDValue IntOp = DAG.getNode(IntOpcode, dl, IntVT, Op0, Op1);
25815 return DAG.getNode(ISD::BITCAST, dl, VT, IntOp);
25820 /// Do target-specific dag combines on X86ISD::FMIN and X86ISD::FMAX nodes.
25821 static SDValue PerformFMinFMaxCombine(SDNode *N, SelectionDAG &DAG) {
25822 assert(N->getOpcode() == X86ISD::FMIN || N->getOpcode() == X86ISD::FMAX);
25824 // Only perform optimizations if UnsafeMath is used.
25825 if (!DAG.getTarget().Options.UnsafeFPMath)
25828 // If we run in unsafe-math mode, then convert the FMAX and FMIN nodes
25829 // into FMINC and FMAXC, which are Commutative operations.
25830 unsigned NewOp = 0;
25831 switch (N->getOpcode()) {
25832 default: llvm_unreachable("unknown opcode");
25833 case X86ISD::FMIN: NewOp = X86ISD::FMINC; break;
25834 case X86ISD::FMAX: NewOp = X86ISD::FMAXC; break;
25837 return DAG.getNode(NewOp, SDLoc(N), N->getValueType(0),
25838 N->getOperand(0), N->getOperand(1));
25841 /// Do target-specific dag combines on X86ISD::FAND nodes.
25842 static SDValue PerformFANDCombine(SDNode *N, SelectionDAG &DAG) {
25843 // FAND(0.0, x) -> 0.0
25844 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
25845 if (C->getValueAPF().isPosZero())
25846 return N->getOperand(0);
25848 // FAND(x, 0.0) -> 0.0
25849 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
25850 if (C->getValueAPF().isPosZero())
25851 return N->getOperand(1);
25856 /// Do target-specific dag combines on X86ISD::FANDN nodes
25857 static SDValue PerformFANDNCombine(SDNode *N, SelectionDAG &DAG) {
25858 // FANDN(0.0, x) -> x
25859 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
25860 if (C->getValueAPF().isPosZero())
25861 return N->getOperand(1);
25863 // FANDN(x, 0.0) -> 0.0
25864 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
25865 if (C->getValueAPF().isPosZero())
25866 return N->getOperand(1);
25871 static SDValue PerformBTCombine(SDNode *N,
25873 TargetLowering::DAGCombinerInfo &DCI) {
25874 // BT ignores high bits in the bit index operand.
25875 SDValue Op1 = N->getOperand(1);
25876 if (Op1.hasOneUse()) {
25877 unsigned BitWidth = Op1.getValueSizeInBits();
25878 APInt DemandedMask = APInt::getLowBitsSet(BitWidth, Log2_32(BitWidth));
25879 APInt KnownZero, KnownOne;
25880 TargetLowering::TargetLoweringOpt TLO(DAG, !DCI.isBeforeLegalize(),
25881 !DCI.isBeforeLegalizeOps());
25882 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
25883 if (TLO.ShrinkDemandedConstant(Op1, DemandedMask) ||
25884 TLI.SimplifyDemandedBits(Op1, DemandedMask, KnownZero, KnownOne, TLO))
25885 DCI.CommitTargetLoweringOpt(TLO);
25890 static SDValue PerformVZEXT_MOVLCombine(SDNode *N, SelectionDAG &DAG) {
25891 SDValue Op = N->getOperand(0);
25892 if (Op.getOpcode() == ISD::BITCAST)
25893 Op = Op.getOperand(0);
25894 EVT VT = N->getValueType(0), OpVT = Op.getValueType();
25895 if (Op.getOpcode() == X86ISD::VZEXT_LOAD &&
25896 VT.getVectorElementType().getSizeInBits() ==
25897 OpVT.getVectorElementType().getSizeInBits()) {
25898 return DAG.getNode(ISD::BITCAST, SDLoc(N), VT, Op);
25903 static SDValue PerformSIGN_EXTEND_INREGCombine(SDNode *N, SelectionDAG &DAG,
25904 const X86Subtarget *Subtarget) {
25905 EVT VT = N->getValueType(0);
25906 if (!VT.isVector())
25909 SDValue N0 = N->getOperand(0);
25910 SDValue N1 = N->getOperand(1);
25911 EVT ExtraVT = cast<VTSDNode>(N1)->getVT();
25914 // The SIGN_EXTEND_INREG to v4i64 is expensive operation on the
25915 // both SSE and AVX2 since there is no sign-extended shift right
25916 // operation on a vector with 64-bit elements.
25917 //(sext_in_reg (v4i64 anyext (v4i32 x )), ExtraVT) ->
25918 // (v4i64 sext (v4i32 sext_in_reg (v4i32 x , ExtraVT)))
25919 if (VT == MVT::v4i64 && (N0.getOpcode() == ISD::ANY_EXTEND ||
25920 N0.getOpcode() == ISD::SIGN_EXTEND)) {
25921 SDValue N00 = N0.getOperand(0);
25923 // EXTLOAD has a better solution on AVX2,
25924 // it may be replaced with X86ISD::VSEXT node.
25925 if (N00.getOpcode() == ISD::LOAD && Subtarget->hasInt256())
25926 if (!ISD::isNormalLoad(N00.getNode()))
25929 if (N00.getValueType() == MVT::v4i32 && ExtraVT.getSizeInBits() < 128) {
25930 SDValue Tmp = DAG.getNode(ISD::SIGN_EXTEND_INREG, dl, MVT::v4i32,
25932 return DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i64, Tmp);
25938 /// sext(add_nsw(x, C)) --> add(sext(x), C_sext)
25939 /// Promoting a sign extension ahead of an 'add nsw' exposes opportunities
25940 /// to combine math ops, use an LEA, or use a complex addressing mode. This can
25941 /// eliminate extend, add, and shift instructions.
25942 static SDValue promoteSextBeforeAddNSW(SDNode *Sext, SelectionDAG &DAG,
25943 const X86Subtarget *Subtarget) {
25944 // TODO: This should be valid for other integer types.
25945 EVT VT = Sext->getValueType(0);
25946 if (VT != MVT::i64)
25949 // We need an 'add nsw' feeding into the 'sext'.
25950 SDValue Add = Sext->getOperand(0);
25951 if (Add.getOpcode() != ISD::ADD || !Add->getFlags()->hasNoSignedWrap())
25954 // Having a constant operand to the 'add' ensures that we are not increasing
25955 // the instruction count because the constant is extended for free below.
25956 // A constant operand can also become the displacement field of an LEA.
25957 auto *AddOp1 = dyn_cast<ConstantSDNode>(Add.getOperand(1));
25961 // Don't make the 'add' bigger if there's no hope of combining it with some
25962 // other 'add' or 'shl' instruction.
25963 // TODO: It may be profitable to generate simpler LEA instructions in place
25964 // of single 'add' instructions, but the cost model for selecting an LEA
25965 // currently has a high threshold.
25966 bool HasLEAPotential = false;
25967 for (auto *User : Sext->uses()) {
25968 if (User->getOpcode() == ISD::ADD || User->getOpcode() == ISD::SHL) {
25969 HasLEAPotential = true;
25973 if (!HasLEAPotential)
25976 // Everything looks good, so pull the 'sext' ahead of the 'add'.
25977 int64_t AddConstant = AddOp1->getSExtValue();
25978 SDValue AddOp0 = Add.getOperand(0);
25979 SDValue NewSext = DAG.getNode(ISD::SIGN_EXTEND, SDLoc(Sext), VT, AddOp0);
25980 SDValue NewConstant = DAG.getConstant(AddConstant, SDLoc(Add), VT);
25982 // The wider add is guaranteed to not wrap because both operands are
25985 Flags.setNoSignedWrap(true);
25986 return DAG.getNode(ISD::ADD, SDLoc(Add), VT, NewSext, NewConstant, &Flags);
25989 static SDValue PerformSExtCombine(SDNode *N, SelectionDAG &DAG,
25990 TargetLowering::DAGCombinerInfo &DCI,
25991 const X86Subtarget *Subtarget) {
25992 SDValue N0 = N->getOperand(0);
25993 EVT VT = N->getValueType(0);
25994 EVT SVT = VT.getScalarType();
25995 EVT InVT = N0.getValueType();
25996 EVT InSVT = InVT.getScalarType();
25999 // (i8,i32 sext (sdivrem (i8 x, i8 y)) ->
26000 // (i8,i32 (sdivrem_sext_hreg (i8 x, i8 y)
26001 // This exposes the sext to the sdivrem lowering, so that it directly extends
26002 // from AH (which we otherwise need to do contortions to access).
26003 if (N0.getOpcode() == ISD::SDIVREM && N0.getResNo() == 1 &&
26004 InVT == MVT::i8 && VT == MVT::i32) {
26005 SDVTList NodeTys = DAG.getVTList(MVT::i8, VT);
26006 SDValue R = DAG.getNode(X86ISD::SDIVREM8_SEXT_HREG, DL, NodeTys,
26007 N0.getOperand(0), N0.getOperand(1));
26008 DAG.ReplaceAllUsesOfValueWith(N0.getValue(0), R.getValue(0));
26009 return R.getValue(1);
26012 if (!DCI.isBeforeLegalizeOps()) {
26013 if (InVT == MVT::i1) {
26014 SDValue Zero = DAG.getConstant(0, DL, VT);
26016 DAG.getConstant(APInt::getAllOnesValue(VT.getSizeInBits()), DL, VT);
26017 return DAG.getNode(ISD::SELECT, DL, VT, N0, AllOnes, Zero);
26022 if (VT.isVector() && Subtarget->hasSSE2()) {
26023 auto ExtendVecSize = [&DAG](SDLoc DL, SDValue N, unsigned Size) {
26024 EVT InVT = N.getValueType();
26025 EVT OutVT = EVT::getVectorVT(*DAG.getContext(), InVT.getScalarType(),
26026 Size / InVT.getScalarSizeInBits());
26027 SmallVector<SDValue, 8> Opnds(Size / InVT.getSizeInBits(),
26028 DAG.getUNDEF(InVT));
26030 return DAG.getNode(ISD::CONCAT_VECTORS, DL, OutVT, Opnds);
26033 // If target-size is less than 128-bits, extend to a type that would extend
26034 // to 128 bits, extend that and extract the original target vector.
26035 if (VT.getSizeInBits() < 128 && !(128 % VT.getSizeInBits()) &&
26036 (SVT == MVT::i64 || SVT == MVT::i32 || SVT == MVT::i16) &&
26037 (InSVT == MVT::i32 || InSVT == MVT::i16 || InSVT == MVT::i8)) {
26038 unsigned Scale = 128 / VT.getSizeInBits();
26040 EVT::getVectorVT(*DAG.getContext(), SVT, 128 / SVT.getSizeInBits());
26041 SDValue Ex = ExtendVecSize(DL, N0, Scale * InVT.getSizeInBits());
26042 SDValue SExt = DAG.getNode(ISD::SIGN_EXTEND, DL, ExVT, Ex);
26043 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, VT, SExt,
26044 DAG.getIntPtrConstant(0, DL));
26047 // If target-size is 128-bits, then convert to ISD::SIGN_EXTEND_VECTOR_INREG
26048 // which ensures lowering to X86ISD::VSEXT (pmovsx*).
26049 if (VT.getSizeInBits() == 128 &&
26050 (SVT == MVT::i64 || SVT == MVT::i32 || SVT == MVT::i16) &&
26051 (InSVT == MVT::i32 || InSVT == MVT::i16 || InSVT == MVT::i8)) {
26052 SDValue ExOp = ExtendVecSize(DL, N0, 128);
26053 return DAG.getSignExtendVectorInReg(ExOp, DL, VT);
26056 // On pre-AVX2 targets, split into 128-bit nodes of
26057 // ISD::SIGN_EXTEND_VECTOR_INREG.
26058 if (!Subtarget->hasInt256() && !(VT.getSizeInBits() % 128) &&
26059 (SVT == MVT::i64 || SVT == MVT::i32 || SVT == MVT::i16) &&
26060 (InSVT == MVT::i32 || InSVT == MVT::i16 || InSVT == MVT::i8)) {
26061 unsigned NumVecs = VT.getSizeInBits() / 128;
26062 unsigned NumSubElts = 128 / SVT.getSizeInBits();
26063 EVT SubVT = EVT::getVectorVT(*DAG.getContext(), SVT, NumSubElts);
26064 EVT InSubVT = EVT::getVectorVT(*DAG.getContext(), InSVT, NumSubElts);
26066 SmallVector<SDValue, 8> Opnds;
26067 for (unsigned i = 0, Offset = 0; i != NumVecs;
26068 ++i, Offset += NumSubElts) {
26069 SDValue SrcVec = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, InSubVT, N0,
26070 DAG.getIntPtrConstant(Offset, DL));
26071 SrcVec = ExtendVecSize(DL, SrcVec, 128);
26072 SrcVec = DAG.getSignExtendVectorInReg(SrcVec, DL, SubVT);
26073 Opnds.push_back(SrcVec);
26075 return DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, Opnds);
26079 if (Subtarget->hasAVX() && VT.isVector() && VT.getSizeInBits() == 256)
26080 if (SDValue R = WidenMaskArithmetic(N, DAG, DCI, Subtarget))
26083 if (SDValue NewAdd = promoteSextBeforeAddNSW(N, DAG, Subtarget))
26089 static SDValue PerformFMACombine(SDNode *N, SelectionDAG &DAG,
26090 const X86Subtarget* Subtarget) {
26092 EVT VT = N->getValueType(0);
26094 // Let legalize expand this if it isn't a legal type yet.
26095 if (!DAG.getTargetLoweringInfo().isTypeLegal(VT))
26098 EVT ScalarVT = VT.getScalarType();
26099 if ((ScalarVT != MVT::f32 && ScalarVT != MVT::f64) ||
26100 (!Subtarget->hasFMA() && !Subtarget->hasFMA4() &&
26101 !Subtarget->hasAVX512()))
26104 SDValue A = N->getOperand(0);
26105 SDValue B = N->getOperand(1);
26106 SDValue C = N->getOperand(2);
26108 bool NegA = (A.getOpcode() == ISD::FNEG);
26109 bool NegB = (B.getOpcode() == ISD::FNEG);
26110 bool NegC = (C.getOpcode() == ISD::FNEG);
26112 // Negative multiplication when NegA xor NegB
26113 bool NegMul = (NegA != NegB);
26115 A = A.getOperand(0);
26117 B = B.getOperand(0);
26119 C = C.getOperand(0);
26123 Opcode = (!NegC) ? X86ISD::FMADD : X86ISD::FMSUB;
26125 Opcode = (!NegC) ? X86ISD::FNMADD : X86ISD::FNMSUB;
26127 return DAG.getNode(Opcode, dl, VT, A, B, C);
26130 static SDValue PerformZExtCombine(SDNode *N, SelectionDAG &DAG,
26131 TargetLowering::DAGCombinerInfo &DCI,
26132 const X86Subtarget *Subtarget) {
26133 // (i32 zext (and (i8 x86isd::setcc_carry), 1)) ->
26134 // (and (i32 x86isd::setcc_carry), 1)
26135 // This eliminates the zext. This transformation is necessary because
26136 // ISD::SETCC is always legalized to i8.
26138 SDValue N0 = N->getOperand(0);
26139 EVT VT = N->getValueType(0);
26141 if (N0.getOpcode() == ISD::AND &&
26143 N0.getOperand(0).hasOneUse()) {
26144 SDValue N00 = N0.getOperand(0);
26145 if (N00.getOpcode() == X86ISD::SETCC_CARRY) {
26146 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
26147 if (!C || C->getZExtValue() != 1)
26149 return DAG.getNode(ISD::AND, dl, VT,
26150 DAG.getNode(X86ISD::SETCC_CARRY, dl, VT,
26151 N00.getOperand(0), N00.getOperand(1)),
26152 DAG.getConstant(1, dl, VT));
26156 if (N0.getOpcode() == ISD::TRUNCATE &&
26158 N0.getOperand(0).hasOneUse()) {
26159 SDValue N00 = N0.getOperand(0);
26160 if (N00.getOpcode() == X86ISD::SETCC_CARRY) {
26161 return DAG.getNode(ISD::AND, dl, VT,
26162 DAG.getNode(X86ISD::SETCC_CARRY, dl, VT,
26163 N00.getOperand(0), N00.getOperand(1)),
26164 DAG.getConstant(1, dl, VT));
26168 if (VT.is256BitVector())
26169 if (SDValue R = WidenMaskArithmetic(N, DAG, DCI, Subtarget))
26172 // (i8,i32 zext (udivrem (i8 x, i8 y)) ->
26173 // (i8,i32 (udivrem_zext_hreg (i8 x, i8 y)
26174 // This exposes the zext to the udivrem lowering, so that it directly extends
26175 // from AH (which we otherwise need to do contortions to access).
26176 if (N0.getOpcode() == ISD::UDIVREM &&
26177 N0.getResNo() == 1 && N0.getValueType() == MVT::i8 &&
26178 (VT == MVT::i32 || VT == MVT::i64)) {
26179 SDVTList NodeTys = DAG.getVTList(MVT::i8, VT);
26180 SDValue R = DAG.getNode(X86ISD::UDIVREM8_ZEXT_HREG, dl, NodeTys,
26181 N0.getOperand(0), N0.getOperand(1));
26182 DAG.ReplaceAllUsesOfValueWith(N0.getValue(0), R.getValue(0));
26183 return R.getValue(1);
26189 // Optimize x == -y --> x+y == 0
26190 // x != -y --> x+y != 0
26191 static SDValue PerformISDSETCCCombine(SDNode *N, SelectionDAG &DAG,
26192 const X86Subtarget* Subtarget) {
26193 ISD::CondCode CC = cast<CondCodeSDNode>(N->getOperand(2))->get();
26194 SDValue LHS = N->getOperand(0);
26195 SDValue RHS = N->getOperand(1);
26196 EVT VT = N->getValueType(0);
26199 if ((CC == ISD::SETNE || CC == ISD::SETEQ) && LHS.getOpcode() == ISD::SUB)
26200 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(LHS.getOperand(0)))
26201 if (C->getAPIntValue() == 0 && LHS.hasOneUse()) {
26202 SDValue addV = DAG.getNode(ISD::ADD, DL, LHS.getValueType(), RHS,
26203 LHS.getOperand(1));
26204 return DAG.getSetCC(DL, N->getValueType(0), addV,
26205 DAG.getConstant(0, DL, addV.getValueType()), CC);
26207 if ((CC == ISD::SETNE || CC == ISD::SETEQ) && RHS.getOpcode() == ISD::SUB)
26208 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(RHS.getOperand(0)))
26209 if (C->getAPIntValue() == 0 && RHS.hasOneUse()) {
26210 SDValue addV = DAG.getNode(ISD::ADD, DL, RHS.getValueType(), LHS,
26211 RHS.getOperand(1));
26212 return DAG.getSetCC(DL, N->getValueType(0), addV,
26213 DAG.getConstant(0, DL, addV.getValueType()), CC);
26216 if (VT.getScalarType() == MVT::i1 &&
26217 (CC == ISD::SETNE || CC == ISD::SETEQ || ISD::isSignedIntSetCC(CC))) {
26219 (LHS.getOpcode() == ISD::SIGN_EXTEND) &&
26220 (LHS.getOperand(0).getValueType().getScalarType() == MVT::i1);
26221 bool IsVZero1 = ISD::isBuildVectorAllZeros(RHS.getNode());
26223 if (!IsSEXT0 || !IsVZero1) {
26224 // Swap the operands and update the condition code.
26225 std::swap(LHS, RHS);
26226 CC = ISD::getSetCCSwappedOperands(CC);
26228 IsSEXT0 = (LHS.getOpcode() == ISD::SIGN_EXTEND) &&
26229 (LHS.getOperand(0).getValueType().getScalarType() == MVT::i1);
26230 IsVZero1 = ISD::isBuildVectorAllZeros(RHS.getNode());
26233 if (IsSEXT0 && IsVZero1) {
26234 assert(VT == LHS.getOperand(0).getValueType() &&
26235 "Uexpected operand type");
26236 if (CC == ISD::SETGT)
26237 return DAG.getConstant(0, DL, VT);
26238 if (CC == ISD::SETLE)
26239 return DAG.getConstant(1, DL, VT);
26240 if (CC == ISD::SETEQ || CC == ISD::SETGE)
26241 return DAG.getNOT(DL, LHS.getOperand(0), VT);
26243 assert((CC == ISD::SETNE || CC == ISD::SETLT) &&
26244 "Unexpected condition code!");
26245 return LHS.getOperand(0);
26252 static SDValue NarrowVectorLoadToElement(LoadSDNode *Load, unsigned Index,
26253 SelectionDAG &DAG) {
26255 MVT VT = Load->getSimpleValueType(0);
26256 MVT EVT = VT.getVectorElementType();
26257 SDValue Addr = Load->getOperand(1);
26258 SDValue NewAddr = DAG.getNode(
26259 ISD::ADD, dl, Addr.getSimpleValueType(), Addr,
26260 DAG.getConstant(Index * EVT.getStoreSize(), dl,
26261 Addr.getSimpleValueType()));
26264 DAG.getLoad(EVT, dl, Load->getChain(), NewAddr,
26265 DAG.getMachineFunction().getMachineMemOperand(
26266 Load->getMemOperand(), 0, EVT.getStoreSize()));
26270 static SDValue PerformINSERTPSCombine(SDNode *N, SelectionDAG &DAG,
26271 const X86Subtarget *Subtarget) {
26273 MVT VT = N->getOperand(1)->getSimpleValueType(0);
26274 assert((VT == MVT::v4f32 || VT == MVT::v4i32) &&
26275 "X86insertps is only defined for v4x32");
26277 SDValue Ld = N->getOperand(1);
26278 if (MayFoldLoad(Ld)) {
26279 // Extract the countS bits from the immediate so we can get the proper
26280 // address when narrowing the vector load to a specific element.
26281 // When the second source op is a memory address, insertps doesn't use
26282 // countS and just gets an f32 from that address.
26283 unsigned DestIndex =
26284 cast<ConstantSDNode>(N->getOperand(2))->getZExtValue() >> 6;
26286 Ld = NarrowVectorLoadToElement(cast<LoadSDNode>(Ld), DestIndex, DAG);
26288 // Create this as a scalar to vector to match the instruction pattern.
26289 SDValue LoadScalarToVector = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Ld);
26290 // countS bits are ignored when loading from memory on insertps, which
26291 // means we don't need to explicitly set them to 0.
26292 return DAG.getNode(X86ISD::INSERTPS, dl, VT, N->getOperand(0),
26293 LoadScalarToVector, N->getOperand(2));
26298 static SDValue PerformBLENDICombine(SDNode *N, SelectionDAG &DAG) {
26299 SDValue V0 = N->getOperand(0);
26300 SDValue V1 = N->getOperand(1);
26302 EVT VT = N->getValueType(0);
26304 // Canonicalize a v2f64 blend with a mask of 2 by swapping the vector
26305 // operands and changing the mask to 1. This saves us a bunch of
26306 // pattern-matching possibilities related to scalar math ops in SSE/AVX.
26307 // x86InstrInfo knows how to commute this back after instruction selection
26308 // if it would help register allocation.
26310 // TODO: If optimizing for size or a processor that doesn't suffer from
26311 // partial register update stalls, this should be transformed into a MOVSD
26312 // instruction because a MOVSD is 1-2 bytes smaller than a BLENDPD.
26314 if (VT == MVT::v2f64)
26315 if (auto *Mask = dyn_cast<ConstantSDNode>(N->getOperand(2)))
26316 if (Mask->getZExtValue() == 2 && !isShuffleFoldableLoad(V0)) {
26317 SDValue NewMask = DAG.getConstant(1, DL, MVT::i8);
26318 return DAG.getNode(X86ISD::BLENDI, DL, VT, V1, V0, NewMask);
26324 // Helper function of PerformSETCCCombine. It is to materialize "setb reg"
26325 // as "sbb reg,reg", since it can be extended without zext and produces
26326 // an all-ones bit which is more useful than 0/1 in some cases.
26327 static SDValue MaterializeSETB(SDLoc DL, SDValue EFLAGS, SelectionDAG &DAG,
26330 return DAG.getNode(ISD::AND, DL, VT,
26331 DAG.getNode(X86ISD::SETCC_CARRY, DL, MVT::i8,
26332 DAG.getConstant(X86::COND_B, DL, MVT::i8),
26334 DAG.getConstant(1, DL, VT));
26335 assert (VT == MVT::i1 && "Unexpected type for SECCC node");
26336 return DAG.getNode(ISD::TRUNCATE, DL, MVT::i1,
26337 DAG.getNode(X86ISD::SETCC_CARRY, DL, MVT::i8,
26338 DAG.getConstant(X86::COND_B, DL, MVT::i8),
26342 // Optimize RES = X86ISD::SETCC CONDCODE, EFLAG_INPUT
26343 static SDValue PerformSETCCCombine(SDNode *N, SelectionDAG &DAG,
26344 TargetLowering::DAGCombinerInfo &DCI,
26345 const X86Subtarget *Subtarget) {
26347 X86::CondCode CC = X86::CondCode(N->getConstantOperandVal(0));
26348 SDValue EFLAGS = N->getOperand(1);
26350 if (CC == X86::COND_A) {
26351 // Try to convert COND_A into COND_B in an attempt to facilitate
26352 // materializing "setb reg".
26354 // Do not flip "e > c", where "c" is a constant, because Cmp instruction
26355 // cannot take an immediate as its first operand.
26357 if (EFLAGS.getOpcode() == X86ISD::SUB && EFLAGS.hasOneUse() &&
26358 EFLAGS.getValueType().isInteger() &&
26359 !isa<ConstantSDNode>(EFLAGS.getOperand(1))) {
26360 SDValue NewSub = DAG.getNode(X86ISD::SUB, SDLoc(EFLAGS),
26361 EFLAGS.getNode()->getVTList(),
26362 EFLAGS.getOperand(1), EFLAGS.getOperand(0));
26363 SDValue NewEFLAGS = SDValue(NewSub.getNode(), EFLAGS.getResNo());
26364 return MaterializeSETB(DL, NewEFLAGS, DAG, N->getSimpleValueType(0));
26368 // Materialize "setb reg" as "sbb reg,reg", since it can be extended without
26369 // a zext and produces an all-ones bit which is more useful than 0/1 in some
26371 if (CC == X86::COND_B)
26372 return MaterializeSETB(DL, EFLAGS, DAG, N->getSimpleValueType(0));
26374 if (SDValue Flags = checkBoolTestSetCCCombine(EFLAGS, CC)) {
26375 SDValue Cond = DAG.getConstant(CC, DL, MVT::i8);
26376 return DAG.getNode(X86ISD::SETCC, DL, N->getVTList(), Cond, Flags);
26382 // Optimize branch condition evaluation.
26384 static SDValue PerformBrCondCombine(SDNode *N, SelectionDAG &DAG,
26385 TargetLowering::DAGCombinerInfo &DCI,
26386 const X86Subtarget *Subtarget) {
26388 SDValue Chain = N->getOperand(0);
26389 SDValue Dest = N->getOperand(1);
26390 SDValue EFLAGS = N->getOperand(3);
26391 X86::CondCode CC = X86::CondCode(N->getConstantOperandVal(2));
26393 if (SDValue Flags = checkBoolTestSetCCCombine(EFLAGS, CC)) {
26394 SDValue Cond = DAG.getConstant(CC, DL, MVT::i8);
26395 return DAG.getNode(X86ISD::BRCOND, DL, N->getVTList(), Chain, Dest, Cond,
26402 static SDValue performVectorCompareAndMaskUnaryOpCombine(SDNode *N,
26403 SelectionDAG &DAG) {
26404 // Take advantage of vector comparisons producing 0 or -1 in each lane to
26405 // optimize away operation when it's from a constant.
26407 // The general transformation is:
26408 // UNARYOP(AND(VECTOR_CMP(x,y), constant)) -->
26409 // AND(VECTOR_CMP(x,y), constant2)
26410 // constant2 = UNARYOP(constant)
26412 // Early exit if this isn't a vector operation, the operand of the
26413 // unary operation isn't a bitwise AND, or if the sizes of the operations
26414 // aren't the same.
26415 EVT VT = N->getValueType(0);
26416 if (!VT.isVector() || N->getOperand(0)->getOpcode() != ISD::AND ||
26417 N->getOperand(0)->getOperand(0)->getOpcode() != ISD::SETCC ||
26418 VT.getSizeInBits() != N->getOperand(0)->getValueType(0).getSizeInBits())
26421 // Now check that the other operand of the AND is a constant. We could
26422 // make the transformation for non-constant splats as well, but it's unclear
26423 // that would be a benefit as it would not eliminate any operations, just
26424 // perform one more step in scalar code before moving to the vector unit.
26425 if (BuildVectorSDNode *BV =
26426 dyn_cast<BuildVectorSDNode>(N->getOperand(0)->getOperand(1))) {
26427 // Bail out if the vector isn't a constant.
26428 if (!BV->isConstant())
26431 // Everything checks out. Build up the new and improved node.
26433 EVT IntVT = BV->getValueType(0);
26434 // Create a new constant of the appropriate type for the transformed
26436 SDValue SourceConst = DAG.getNode(N->getOpcode(), DL, VT, SDValue(BV, 0));
26437 // The AND node needs bitcasts to/from an integer vector type around it.
26438 SDValue MaskConst = DAG.getBitcast(IntVT, SourceConst);
26439 SDValue NewAnd = DAG.getNode(ISD::AND, DL, IntVT,
26440 N->getOperand(0)->getOperand(0), MaskConst);
26441 SDValue Res = DAG.getBitcast(VT, NewAnd);
26448 static SDValue PerformUINT_TO_FPCombine(SDNode *N, SelectionDAG &DAG,
26449 const X86Subtarget *Subtarget) {
26450 SDValue Op0 = N->getOperand(0);
26451 EVT VT = N->getValueType(0);
26452 EVT InVT = Op0.getValueType();
26453 EVT InSVT = InVT.getScalarType();
26454 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
26456 // UINT_TO_FP(vXi8) -> SINT_TO_FP(ZEXT(vXi8 to vXi32))
26457 // UINT_TO_FP(vXi16) -> SINT_TO_FP(ZEXT(vXi16 to vXi32))
26458 if (InVT.isVector() && (InSVT == MVT::i8 || InSVT == MVT::i16)) {
26460 EVT DstVT = EVT::getVectorVT(*DAG.getContext(), MVT::i32,
26461 InVT.getVectorNumElements());
26462 SDValue P = DAG.getNode(ISD::ZERO_EXTEND, dl, DstVT, Op0);
26464 if (TLI.isOperationLegal(ISD::UINT_TO_FP, DstVT))
26465 return DAG.getNode(ISD::UINT_TO_FP, dl, VT, P);
26467 return DAG.getNode(ISD::SINT_TO_FP, dl, VT, P);
26473 static SDValue PerformSINT_TO_FPCombine(SDNode *N, SelectionDAG &DAG,
26474 const X86Subtarget *Subtarget) {
26475 // First try to optimize away the conversion entirely when it's
26476 // conditionally from a constant. Vectors only.
26477 if (SDValue Res = performVectorCompareAndMaskUnaryOpCombine(N, DAG))
26480 // Now move on to more general possibilities.
26481 SDValue Op0 = N->getOperand(0);
26482 EVT VT = N->getValueType(0);
26483 EVT InVT = Op0.getValueType();
26484 EVT InSVT = InVT.getScalarType();
26486 // SINT_TO_FP(vXi8) -> SINT_TO_FP(SEXT(vXi8 to vXi32))
26487 // SINT_TO_FP(vXi16) -> SINT_TO_FP(SEXT(vXi16 to vXi32))
26488 if (InVT.isVector() && (InSVT == MVT::i8 || InSVT == MVT::i16)) {
26490 EVT DstVT = EVT::getVectorVT(*DAG.getContext(), MVT::i32,
26491 InVT.getVectorNumElements());
26492 SDValue P = DAG.getNode(ISD::SIGN_EXTEND, dl, DstVT, Op0);
26493 return DAG.getNode(ISD::SINT_TO_FP, dl, VT, P);
26496 // Transform (SINT_TO_FP (i64 ...)) into an x87 operation if we have
26497 // a 32-bit target where SSE doesn't support i64->FP operations.
26498 if (Op0.getOpcode() == ISD::LOAD) {
26499 LoadSDNode *Ld = cast<LoadSDNode>(Op0.getNode());
26500 EVT LdVT = Ld->getValueType(0);
26502 // This transformation is not supported if the result type is f16
26503 if (VT == MVT::f16)
26506 if (!Ld->isVolatile() && !VT.isVector() &&
26507 ISD::isNON_EXTLoad(Op0.getNode()) && Op0.hasOneUse() &&
26508 !Subtarget->is64Bit() && LdVT == MVT::i64) {
26509 SDValue FILDChain = Subtarget->getTargetLowering()->BuildFILD(
26510 SDValue(N, 0), LdVT, Ld->getChain(), Op0, DAG);
26511 DAG.ReplaceAllUsesOfValueWith(Op0.getValue(1), FILDChain.getValue(1));
26518 // Optimize RES, EFLAGS = X86ISD::ADC LHS, RHS, EFLAGS
26519 static SDValue PerformADCCombine(SDNode *N, SelectionDAG &DAG,
26520 X86TargetLowering::DAGCombinerInfo &DCI) {
26521 // If the LHS and RHS of the ADC node are zero, then it can't overflow and
26522 // the result is either zero or one (depending on the input carry bit).
26523 // Strength reduce this down to a "set on carry" aka SETCC_CARRY&1.
26524 if (X86::isZeroNode(N->getOperand(0)) &&
26525 X86::isZeroNode(N->getOperand(1)) &&
26526 // We don't have a good way to replace an EFLAGS use, so only do this when
26528 SDValue(N, 1).use_empty()) {
26530 EVT VT = N->getValueType(0);
26531 SDValue CarryOut = DAG.getConstant(0, DL, N->getValueType(1));
26532 SDValue Res1 = DAG.getNode(ISD::AND, DL, VT,
26533 DAG.getNode(X86ISD::SETCC_CARRY, DL, VT,
26534 DAG.getConstant(X86::COND_B, DL,
26537 DAG.getConstant(1, DL, VT));
26538 return DCI.CombineTo(N, Res1, CarryOut);
26544 // fold (add Y, (sete X, 0)) -> adc 0, Y
26545 // (add Y, (setne X, 0)) -> sbb -1, Y
26546 // (sub (sete X, 0), Y) -> sbb 0, Y
26547 // (sub (setne X, 0), Y) -> adc -1, Y
26548 static SDValue OptimizeConditionalInDecrement(SDNode *N, SelectionDAG &DAG) {
26551 // Look through ZExts.
26552 SDValue Ext = N->getOperand(N->getOpcode() == ISD::SUB ? 1 : 0);
26553 if (Ext.getOpcode() != ISD::ZERO_EXTEND || !Ext.hasOneUse())
26556 SDValue SetCC = Ext.getOperand(0);
26557 if (SetCC.getOpcode() != X86ISD::SETCC || !SetCC.hasOneUse())
26560 X86::CondCode CC = (X86::CondCode)SetCC.getConstantOperandVal(0);
26561 if (CC != X86::COND_E && CC != X86::COND_NE)
26564 SDValue Cmp = SetCC.getOperand(1);
26565 if (Cmp.getOpcode() != X86ISD::CMP || !Cmp.hasOneUse() ||
26566 !X86::isZeroNode(Cmp.getOperand(1)) ||
26567 !Cmp.getOperand(0).getValueType().isInteger())
26570 SDValue CmpOp0 = Cmp.getOperand(0);
26571 SDValue NewCmp = DAG.getNode(X86ISD::CMP, DL, MVT::i32, CmpOp0,
26572 DAG.getConstant(1, DL, CmpOp0.getValueType()));
26574 SDValue OtherVal = N->getOperand(N->getOpcode() == ISD::SUB ? 0 : 1);
26575 if (CC == X86::COND_NE)
26576 return DAG.getNode(N->getOpcode() == ISD::SUB ? X86ISD::ADC : X86ISD::SBB,
26577 DL, OtherVal.getValueType(), OtherVal,
26578 DAG.getConstant(-1ULL, DL, OtherVal.getValueType()),
26580 return DAG.getNode(N->getOpcode() == ISD::SUB ? X86ISD::SBB : X86ISD::ADC,
26581 DL, OtherVal.getValueType(), OtherVal,
26582 DAG.getConstant(0, DL, OtherVal.getValueType()), NewCmp);
26585 /// PerformADDCombine - Do target-specific dag combines on integer adds.
26586 static SDValue PerformAddCombine(SDNode *N, SelectionDAG &DAG,
26587 const X86Subtarget *Subtarget) {
26588 EVT VT = N->getValueType(0);
26589 SDValue Op0 = N->getOperand(0);
26590 SDValue Op1 = N->getOperand(1);
26592 // Try to synthesize horizontal adds from adds of shuffles.
26593 if (((Subtarget->hasSSSE3() && (VT == MVT::v8i16 || VT == MVT::v4i32)) ||
26594 (Subtarget->hasInt256() && (VT == MVT::v16i16 || VT == MVT::v8i32))) &&
26595 isHorizontalBinOp(Op0, Op1, true))
26596 return DAG.getNode(X86ISD::HADD, SDLoc(N), VT, Op0, Op1);
26598 return OptimizeConditionalInDecrement(N, DAG);
26601 static SDValue PerformSubCombine(SDNode *N, SelectionDAG &DAG,
26602 const X86Subtarget *Subtarget) {
26603 SDValue Op0 = N->getOperand(0);
26604 SDValue Op1 = N->getOperand(1);
26606 // X86 can't encode an immediate LHS of a sub. See if we can push the
26607 // negation into a preceding instruction.
26608 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op0)) {
26609 // If the RHS of the sub is a XOR with one use and a constant, invert the
26610 // immediate. Then add one to the LHS of the sub so we can turn
26611 // X-Y -> X+~Y+1, saving one register.
26612 if (Op1->hasOneUse() && Op1.getOpcode() == ISD::XOR &&
26613 isa<ConstantSDNode>(Op1.getOperand(1))) {
26614 APInt XorC = cast<ConstantSDNode>(Op1.getOperand(1))->getAPIntValue();
26615 EVT VT = Op0.getValueType();
26616 SDValue NewXor = DAG.getNode(ISD::XOR, SDLoc(Op1), VT,
26618 DAG.getConstant(~XorC, SDLoc(Op1), VT));
26619 return DAG.getNode(ISD::ADD, SDLoc(N), VT, NewXor,
26620 DAG.getConstant(C->getAPIntValue() + 1, SDLoc(N), VT));
26624 // Try to synthesize horizontal adds from adds of shuffles.
26625 EVT VT = N->getValueType(0);
26626 if (((Subtarget->hasSSSE3() && (VT == MVT::v8i16 || VT == MVT::v4i32)) ||
26627 (Subtarget->hasInt256() && (VT == MVT::v16i16 || VT == MVT::v8i32))) &&
26628 isHorizontalBinOp(Op0, Op1, true))
26629 return DAG.getNode(X86ISD::HSUB, SDLoc(N), VT, Op0, Op1);
26631 return OptimizeConditionalInDecrement(N, DAG);
26634 /// performVZEXTCombine - Performs build vector combines
26635 static SDValue performVZEXTCombine(SDNode *N, SelectionDAG &DAG,
26636 TargetLowering::DAGCombinerInfo &DCI,
26637 const X86Subtarget *Subtarget) {
26639 MVT VT = N->getSimpleValueType(0);
26640 SDValue Op = N->getOperand(0);
26641 MVT OpVT = Op.getSimpleValueType();
26642 MVT OpEltVT = OpVT.getVectorElementType();
26643 unsigned InputBits = OpEltVT.getSizeInBits() * VT.getVectorNumElements();
26645 // (vzext (bitcast (vzext (x)) -> (vzext x)
26647 while (V.getOpcode() == ISD::BITCAST)
26648 V = V.getOperand(0);
26650 if (V != Op && V.getOpcode() == X86ISD::VZEXT) {
26651 MVT InnerVT = V.getSimpleValueType();
26652 MVT InnerEltVT = InnerVT.getVectorElementType();
26654 // If the element sizes match exactly, we can just do one larger vzext. This
26655 // is always an exact type match as vzext operates on integer types.
26656 if (OpEltVT == InnerEltVT) {
26657 assert(OpVT == InnerVT && "Types must match for vzext!");
26658 return DAG.getNode(X86ISD::VZEXT, DL, VT, V.getOperand(0));
26661 // The only other way we can combine them is if only a single element of the
26662 // inner vzext is used in the input to the outer vzext.
26663 if (InnerEltVT.getSizeInBits() < InputBits)
26666 // In this case, the inner vzext is completely dead because we're going to
26667 // only look at bits inside of the low element. Just do the outer vzext on
26668 // a bitcast of the input to the inner.
26669 return DAG.getNode(X86ISD::VZEXT, DL, VT, DAG.getBitcast(OpVT, V));
26672 // Check if we can bypass extracting and re-inserting an element of an input
26673 // vector. Essentially:
26674 // (bitcast (sclr2vec (ext_vec_elt x))) -> (bitcast x)
26675 if (V.getOpcode() == ISD::SCALAR_TO_VECTOR &&
26676 V.getOperand(0).getOpcode() == ISD::EXTRACT_VECTOR_ELT &&
26677 V.getOperand(0).getSimpleValueType().getSizeInBits() == InputBits) {
26678 SDValue ExtractedV = V.getOperand(0);
26679 SDValue OrigV = ExtractedV.getOperand(0);
26680 if (auto *ExtractIdx = dyn_cast<ConstantSDNode>(ExtractedV.getOperand(1)))
26681 if (ExtractIdx->getZExtValue() == 0) {
26682 MVT OrigVT = OrigV.getSimpleValueType();
26683 // Extract a subvector if necessary...
26684 if (OrigVT.getSizeInBits() > OpVT.getSizeInBits()) {
26685 int Ratio = OrigVT.getSizeInBits() / OpVT.getSizeInBits();
26686 OrigVT = MVT::getVectorVT(OrigVT.getVectorElementType(),
26687 OrigVT.getVectorNumElements() / Ratio);
26688 OrigV = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, OrigVT, OrigV,
26689 DAG.getIntPtrConstant(0, DL));
26691 Op = DAG.getBitcast(OpVT, OrigV);
26692 return DAG.getNode(X86ISD::VZEXT, DL, VT, Op);
26699 SDValue X86TargetLowering::PerformDAGCombine(SDNode *N,
26700 DAGCombinerInfo &DCI) const {
26701 SelectionDAG &DAG = DCI.DAG;
26702 switch (N->getOpcode()) {
26704 case ISD::EXTRACT_VECTOR_ELT:
26705 return PerformEXTRACT_VECTOR_ELTCombine(N, DAG, DCI);
26708 case X86ISD::SHRUNKBLEND:
26709 return PerformSELECTCombine(N, DAG, DCI, Subtarget);
26710 case ISD::BITCAST: return PerformBITCASTCombine(N, DAG);
26711 case X86ISD::CMOV: return PerformCMOVCombine(N, DAG, DCI, Subtarget);
26712 case ISD::ADD: return PerformAddCombine(N, DAG, Subtarget);
26713 case ISD::SUB: return PerformSubCombine(N, DAG, Subtarget);
26714 case X86ISD::ADC: return PerformADCCombine(N, DAG, DCI);
26715 case ISD::MUL: return PerformMulCombine(N, DAG, DCI);
26718 case ISD::SRL: return PerformShiftCombine(N, DAG, DCI, Subtarget);
26719 case ISD::AND: return PerformAndCombine(N, DAG, DCI, Subtarget);
26720 case ISD::OR: return PerformOrCombine(N, DAG, DCI, Subtarget);
26721 case ISD::XOR: return PerformXorCombine(N, DAG, DCI, Subtarget);
26722 case ISD::LOAD: return PerformLOADCombine(N, DAG, DCI, Subtarget);
26723 case ISD::MLOAD: return PerformMLOADCombine(N, DAG, DCI, Subtarget);
26724 case ISD::STORE: return PerformSTORECombine(N, DAG, Subtarget);
26725 case ISD::MSTORE: return PerformMSTORECombine(N, DAG, Subtarget);
26726 case ISD::SINT_TO_FP: return PerformSINT_TO_FPCombine(N, DAG, Subtarget);
26727 case ISD::UINT_TO_FP: return PerformUINT_TO_FPCombine(N, DAG, Subtarget);
26728 case ISD::FADD: return PerformFADDCombine(N, DAG, Subtarget);
26729 case ISD::FSUB: return PerformFSUBCombine(N, DAG, Subtarget);
26731 case X86ISD::FOR: return PerformFORCombine(N, DAG, Subtarget);
26733 case X86ISD::FMAX: return PerformFMinFMaxCombine(N, DAG);
26734 case X86ISD::FAND: return PerformFANDCombine(N, DAG);
26735 case X86ISD::FANDN: return PerformFANDNCombine(N, DAG);
26736 case X86ISD::BT: return PerformBTCombine(N, DAG, DCI);
26737 case X86ISD::VZEXT_MOVL: return PerformVZEXT_MOVLCombine(N, DAG);
26738 case ISD::ANY_EXTEND:
26739 case ISD::ZERO_EXTEND: return PerformZExtCombine(N, DAG, DCI, Subtarget);
26740 case ISD::SIGN_EXTEND: return PerformSExtCombine(N, DAG, DCI, Subtarget);
26741 case ISD::SIGN_EXTEND_INREG:
26742 return PerformSIGN_EXTEND_INREGCombine(N, DAG, Subtarget);
26743 case ISD::SETCC: return PerformISDSETCCCombine(N, DAG, Subtarget);
26744 case X86ISD::SETCC: return PerformSETCCCombine(N, DAG, DCI, Subtarget);
26745 case X86ISD::BRCOND: return PerformBrCondCombine(N, DAG, DCI, Subtarget);
26746 case X86ISD::VZEXT: return performVZEXTCombine(N, DAG, DCI, Subtarget);
26747 case X86ISD::SHUFP: // Handle all target specific shuffles
26748 case X86ISD::PALIGNR:
26749 case X86ISD::UNPCKH:
26750 case X86ISD::UNPCKL:
26751 case X86ISD::MOVHLPS:
26752 case X86ISD::MOVLHPS:
26753 case X86ISD::PSHUFB:
26754 case X86ISD::PSHUFD:
26755 case X86ISD::PSHUFHW:
26756 case X86ISD::PSHUFLW:
26757 case X86ISD::MOVSS:
26758 case X86ISD::MOVSD:
26759 case X86ISD::VPERMILPI:
26760 case X86ISD::VPERM2X128:
26761 case ISD::VECTOR_SHUFFLE: return PerformShuffleCombine(N, DAG, DCI,Subtarget);
26762 case ISD::FMA: return PerformFMACombine(N, DAG, Subtarget);
26763 case X86ISD::INSERTPS: {
26764 if (getTargetMachine().getOptLevel() > CodeGenOpt::None)
26765 return PerformINSERTPSCombine(N, DAG, Subtarget);
26768 case X86ISD::BLENDI: return PerformBLENDICombine(N, DAG);
26774 /// isTypeDesirableForOp - Return true if the target has native support for
26775 /// the specified value type and it is 'desirable' to use the type for the
26776 /// given node type. e.g. On x86 i16 is legal, but undesirable since i16
26777 /// instruction encodings are longer and some i16 instructions are slow.
26778 bool X86TargetLowering::isTypeDesirableForOp(unsigned Opc, EVT VT) const {
26779 if (!isTypeLegal(VT))
26781 if (VT != MVT::i16)
26788 case ISD::SIGN_EXTEND:
26789 case ISD::ZERO_EXTEND:
26790 case ISD::ANY_EXTEND:
26803 /// IsDesirableToPromoteOp - This method query the target whether it is
26804 /// beneficial for dag combiner to promote the specified node. If true, it
26805 /// should return the desired promotion type by reference.
26806 bool X86TargetLowering::IsDesirableToPromoteOp(SDValue Op, EVT &PVT) const {
26807 EVT VT = Op.getValueType();
26808 if (VT != MVT::i16)
26811 bool Promote = false;
26812 bool Commute = false;
26813 switch (Op.getOpcode()) {
26816 LoadSDNode *LD = cast<LoadSDNode>(Op);
26817 // If the non-extending load has a single use and it's not live out, then it
26818 // might be folded.
26819 if (LD->getExtensionType() == ISD::NON_EXTLOAD /*&&
26820 Op.hasOneUse()*/) {
26821 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
26822 UE = Op.getNode()->use_end(); UI != UE; ++UI) {
26823 // The only case where we'd want to promote LOAD (rather then it being
26824 // promoted as an operand is when it's only use is liveout.
26825 if (UI->getOpcode() != ISD::CopyToReg)
26832 case ISD::SIGN_EXTEND:
26833 case ISD::ZERO_EXTEND:
26834 case ISD::ANY_EXTEND:
26839 SDValue N0 = Op.getOperand(0);
26840 // Look out for (store (shl (load), x)).
26841 if (MayFoldLoad(N0) && MayFoldIntoStore(Op))
26854 SDValue N0 = Op.getOperand(0);
26855 SDValue N1 = Op.getOperand(1);
26856 if (!Commute && MayFoldLoad(N1))
26858 // Avoid disabling potential load folding opportunities.
26859 if (MayFoldLoad(N0) && (!isa<ConstantSDNode>(N1) || MayFoldIntoStore(Op)))
26861 if (MayFoldLoad(N1) && (!isa<ConstantSDNode>(N0) || MayFoldIntoStore(Op)))
26871 //===----------------------------------------------------------------------===//
26872 // X86 Inline Assembly Support
26873 //===----------------------------------------------------------------------===//
26875 // Helper to match a string separated by whitespace.
26876 static bool matchAsm(StringRef S, ArrayRef<const char *> Pieces) {
26877 S = S.substr(S.find_first_not_of(" \t")); // Skip leading whitespace.
26879 for (StringRef Piece : Pieces) {
26880 if (!S.startswith(Piece)) // Check if the piece matches.
26883 S = S.substr(Piece.size());
26884 StringRef::size_type Pos = S.find_first_not_of(" \t");
26885 if (Pos == 0) // We matched a prefix.
26894 static bool clobbersFlagRegisters(const SmallVector<StringRef, 4> &AsmPieces) {
26896 if (AsmPieces.size() == 3 || AsmPieces.size() == 4) {
26897 if (std::count(AsmPieces.begin(), AsmPieces.end(), "~{cc}") &&
26898 std::count(AsmPieces.begin(), AsmPieces.end(), "~{flags}") &&
26899 std::count(AsmPieces.begin(), AsmPieces.end(), "~{fpsr}")) {
26901 if (AsmPieces.size() == 3)
26903 else if (std::count(AsmPieces.begin(), AsmPieces.end(), "~{dirflag}"))
26910 bool X86TargetLowering::ExpandInlineAsm(CallInst *CI) const {
26911 InlineAsm *IA = cast<InlineAsm>(CI->getCalledValue());
26913 std::string AsmStr = IA->getAsmString();
26915 IntegerType *Ty = dyn_cast<IntegerType>(CI->getType());
26916 if (!Ty || Ty->getBitWidth() % 16 != 0)
26919 // TODO: should remove alternatives from the asmstring: "foo {a|b}" -> "foo a"
26920 SmallVector<StringRef, 4> AsmPieces;
26921 SplitString(AsmStr, AsmPieces, ";\n");
26923 switch (AsmPieces.size()) {
26924 default: return false;
26926 // FIXME: this should verify that we are targeting a 486 or better. If not,
26927 // we will turn this bswap into something that will be lowered to logical
26928 // ops instead of emitting the bswap asm. For now, we don't support 486 or
26929 // lower so don't worry about this.
26931 if (matchAsm(AsmPieces[0], {"bswap", "$0"}) ||
26932 matchAsm(AsmPieces[0], {"bswapl", "$0"}) ||
26933 matchAsm(AsmPieces[0], {"bswapq", "$0"}) ||
26934 matchAsm(AsmPieces[0], {"bswap", "${0:q}"}) ||
26935 matchAsm(AsmPieces[0], {"bswapl", "${0:q}"}) ||
26936 matchAsm(AsmPieces[0], {"bswapq", "${0:q}"})) {
26937 // No need to check constraints, nothing other than the equivalent of
26938 // "=r,0" would be valid here.
26939 return IntrinsicLowering::LowerToByteSwap(CI);
26942 // rorw $$8, ${0:w} --> llvm.bswap.i16
26943 if (CI->getType()->isIntegerTy(16) &&
26944 IA->getConstraintString().compare(0, 5, "=r,0,") == 0 &&
26945 (matchAsm(AsmPieces[0], {"rorw", "$$8,", "${0:w}"}) ||
26946 matchAsm(AsmPieces[0], {"rolw", "$$8,", "${0:w}"}))) {
26948 StringRef ConstraintsStr = IA->getConstraintString();
26949 SplitString(StringRef(ConstraintsStr).substr(5), AsmPieces, ",");
26950 array_pod_sort(AsmPieces.begin(), AsmPieces.end());
26951 if (clobbersFlagRegisters(AsmPieces))
26952 return IntrinsicLowering::LowerToByteSwap(CI);
26956 if (CI->getType()->isIntegerTy(32) &&
26957 IA->getConstraintString().compare(0, 5, "=r,0,") == 0 &&
26958 matchAsm(AsmPieces[0], {"rorw", "$$8,", "${0:w}"}) &&
26959 matchAsm(AsmPieces[1], {"rorl", "$$16,", "$0"}) &&
26960 matchAsm(AsmPieces[2], {"rorw", "$$8,", "${0:w}"})) {
26962 StringRef ConstraintsStr = IA->getConstraintString();
26963 SplitString(StringRef(ConstraintsStr).substr(5), AsmPieces, ",");
26964 array_pod_sort(AsmPieces.begin(), AsmPieces.end());
26965 if (clobbersFlagRegisters(AsmPieces))
26966 return IntrinsicLowering::LowerToByteSwap(CI);
26969 if (CI->getType()->isIntegerTy(64)) {
26970 InlineAsm::ConstraintInfoVector Constraints = IA->ParseConstraints();
26971 if (Constraints.size() >= 2 &&
26972 Constraints[0].Codes.size() == 1 && Constraints[0].Codes[0] == "A" &&
26973 Constraints[1].Codes.size() == 1 && Constraints[1].Codes[0] == "0") {
26974 // bswap %eax / bswap %edx / xchgl %eax, %edx -> llvm.bswap.i64
26975 if (matchAsm(AsmPieces[0], {"bswap", "%eax"}) &&
26976 matchAsm(AsmPieces[1], {"bswap", "%edx"}) &&
26977 matchAsm(AsmPieces[2], {"xchgl", "%eax,", "%edx"}))
26978 return IntrinsicLowering::LowerToByteSwap(CI);
26986 /// getConstraintType - Given a constraint letter, return the type of
26987 /// constraint it is for this target.
26988 X86TargetLowering::ConstraintType
26989 X86TargetLowering::getConstraintType(StringRef Constraint) const {
26990 if (Constraint.size() == 1) {
26991 switch (Constraint[0]) {
27002 return C_RegisterClass;
27026 return TargetLowering::getConstraintType(Constraint);
27029 /// Examine constraint type and operand type and determine a weight value.
27030 /// This object must already have been set up with the operand type
27031 /// and the current alternative constraint selected.
27032 TargetLowering::ConstraintWeight
27033 X86TargetLowering::getSingleConstraintMatchWeight(
27034 AsmOperandInfo &info, const char *constraint) const {
27035 ConstraintWeight weight = CW_Invalid;
27036 Value *CallOperandVal = info.CallOperandVal;
27037 // If we don't have a value, we can't do a match,
27038 // but allow it at the lowest weight.
27039 if (!CallOperandVal)
27041 Type *type = CallOperandVal->getType();
27042 // Look at the constraint type.
27043 switch (*constraint) {
27045 weight = TargetLowering::getSingleConstraintMatchWeight(info, constraint);
27056 if (CallOperandVal->getType()->isIntegerTy())
27057 weight = CW_SpecificReg;
27062 if (type->isFloatingPointTy())
27063 weight = CW_SpecificReg;
27066 if (type->isX86_MMXTy() && Subtarget->hasMMX())
27067 weight = CW_SpecificReg;
27071 if (((type->getPrimitiveSizeInBits() == 128) && Subtarget->hasSSE1()) ||
27072 ((type->getPrimitiveSizeInBits() == 256) && Subtarget->hasFp256()))
27073 weight = CW_Register;
27076 if (ConstantInt *C = dyn_cast<ConstantInt>(info.CallOperandVal)) {
27077 if (C->getZExtValue() <= 31)
27078 weight = CW_Constant;
27082 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
27083 if (C->getZExtValue() <= 63)
27084 weight = CW_Constant;
27088 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
27089 if ((C->getSExtValue() >= -0x80) && (C->getSExtValue() <= 0x7f))
27090 weight = CW_Constant;
27094 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
27095 if ((C->getZExtValue() == 0xff) || (C->getZExtValue() == 0xffff))
27096 weight = CW_Constant;
27100 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
27101 if (C->getZExtValue() <= 3)
27102 weight = CW_Constant;
27106 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
27107 if (C->getZExtValue() <= 0xff)
27108 weight = CW_Constant;
27113 if (isa<ConstantFP>(CallOperandVal)) {
27114 weight = CW_Constant;
27118 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
27119 if ((C->getSExtValue() >= -0x80000000LL) &&
27120 (C->getSExtValue() <= 0x7fffffffLL))
27121 weight = CW_Constant;
27125 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
27126 if (C->getZExtValue() <= 0xffffffff)
27127 weight = CW_Constant;
27134 /// LowerXConstraint - try to replace an X constraint, which matches anything,
27135 /// with another that has more specific requirements based on the type of the
27136 /// corresponding operand.
27137 const char *X86TargetLowering::
27138 LowerXConstraint(EVT ConstraintVT) const {
27139 // FP X constraints get lowered to SSE1/2 registers if available, otherwise
27140 // 'f' like normal targets.
27141 if (ConstraintVT.isFloatingPoint()) {
27142 if (Subtarget->hasSSE2())
27144 if (Subtarget->hasSSE1())
27148 return TargetLowering::LowerXConstraint(ConstraintVT);
27151 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
27152 /// vector. If it is invalid, don't add anything to Ops.
27153 void X86TargetLowering::LowerAsmOperandForConstraint(SDValue Op,
27154 std::string &Constraint,
27155 std::vector<SDValue>&Ops,
27156 SelectionDAG &DAG) const {
27159 // Only support length 1 constraints for now.
27160 if (Constraint.length() > 1) return;
27162 char ConstraintLetter = Constraint[0];
27163 switch (ConstraintLetter) {
27166 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
27167 if (C->getZExtValue() <= 31) {
27168 Result = DAG.getTargetConstant(C->getZExtValue(), SDLoc(Op),
27169 Op.getValueType());
27175 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
27176 if (C->getZExtValue() <= 63) {
27177 Result = DAG.getTargetConstant(C->getZExtValue(), SDLoc(Op),
27178 Op.getValueType());
27184 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
27185 if (isInt<8>(C->getSExtValue())) {
27186 Result = DAG.getTargetConstant(C->getZExtValue(), SDLoc(Op),
27187 Op.getValueType());
27193 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
27194 if (C->getZExtValue() == 0xff || C->getZExtValue() == 0xffff ||
27195 (Subtarget->is64Bit() && C->getZExtValue() == 0xffffffff)) {
27196 Result = DAG.getTargetConstant(C->getSExtValue(), SDLoc(Op),
27197 Op.getValueType());
27203 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
27204 if (C->getZExtValue() <= 3) {
27205 Result = DAG.getTargetConstant(C->getZExtValue(), SDLoc(Op),
27206 Op.getValueType());
27212 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
27213 if (C->getZExtValue() <= 255) {
27214 Result = DAG.getTargetConstant(C->getZExtValue(), SDLoc(Op),
27215 Op.getValueType());
27221 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
27222 if (C->getZExtValue() <= 127) {
27223 Result = DAG.getTargetConstant(C->getZExtValue(), SDLoc(Op),
27224 Op.getValueType());
27230 // 32-bit signed value
27231 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
27232 if (ConstantInt::isValueValidForType(Type::getInt32Ty(*DAG.getContext()),
27233 C->getSExtValue())) {
27234 // Widen to 64 bits here to get it sign extended.
27235 Result = DAG.getTargetConstant(C->getSExtValue(), SDLoc(Op), MVT::i64);
27238 // FIXME gcc accepts some relocatable values here too, but only in certain
27239 // memory models; it's complicated.
27244 // 32-bit unsigned value
27245 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
27246 if (ConstantInt::isValueValidForType(Type::getInt32Ty(*DAG.getContext()),
27247 C->getZExtValue())) {
27248 Result = DAG.getTargetConstant(C->getZExtValue(), SDLoc(Op),
27249 Op.getValueType());
27253 // FIXME gcc accepts some relocatable values here too, but only in certain
27254 // memory models; it's complicated.
27258 // Literal immediates are always ok.
27259 if (ConstantSDNode *CST = dyn_cast<ConstantSDNode>(Op)) {
27260 // Widen to 64 bits here to get it sign extended.
27261 Result = DAG.getTargetConstant(CST->getSExtValue(), SDLoc(Op), MVT::i64);
27265 // In any sort of PIC mode addresses need to be computed at runtime by
27266 // adding in a register or some sort of table lookup. These can't
27267 // be used as immediates.
27268 if (Subtarget->isPICStyleGOT() || Subtarget->isPICStyleStubPIC())
27271 // If we are in non-pic codegen mode, we allow the address of a global (with
27272 // an optional displacement) to be used with 'i'.
27273 GlobalAddressSDNode *GA = nullptr;
27274 int64_t Offset = 0;
27276 // Match either (GA), (GA+C), (GA+C1+C2), etc.
27278 if ((GA = dyn_cast<GlobalAddressSDNode>(Op))) {
27279 Offset += GA->getOffset();
27281 } else if (Op.getOpcode() == ISD::ADD) {
27282 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
27283 Offset += C->getZExtValue();
27284 Op = Op.getOperand(0);
27287 } else if (Op.getOpcode() == ISD::SUB) {
27288 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
27289 Offset += -C->getZExtValue();
27290 Op = Op.getOperand(0);
27295 // Otherwise, this isn't something we can handle, reject it.
27299 const GlobalValue *GV = GA->getGlobal();
27300 // If we require an extra load to get this address, as in PIC mode, we
27301 // can't accept it.
27302 if (isGlobalStubReference(
27303 Subtarget->ClassifyGlobalReference(GV, DAG.getTarget())))
27306 Result = DAG.getTargetGlobalAddress(GV, SDLoc(Op),
27307 GA->getValueType(0), Offset);
27312 if (Result.getNode()) {
27313 Ops.push_back(Result);
27316 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG);
27319 std::pair<unsigned, const TargetRegisterClass *>
27320 X86TargetLowering::getRegForInlineAsmConstraint(const TargetRegisterInfo *TRI,
27321 StringRef Constraint,
27323 // First, see if this is a constraint that directly corresponds to an LLVM
27325 if (Constraint.size() == 1) {
27326 // GCC Constraint Letters
27327 switch (Constraint[0]) {
27329 // TODO: Slight differences here in allocation order and leaving
27330 // RIP in the class. Do they matter any more here than they do
27331 // in the normal allocation?
27332 case 'q': // GENERAL_REGS in 64-bit mode, Q_REGS in 32-bit mode.
27333 if (Subtarget->is64Bit()) {
27334 if (VT == MVT::i32 || VT == MVT::f32)
27335 return std::make_pair(0U, &X86::GR32RegClass);
27336 if (VT == MVT::i16)
27337 return std::make_pair(0U, &X86::GR16RegClass);
27338 if (VT == MVT::i8 || VT == MVT::i1)
27339 return std::make_pair(0U, &X86::GR8RegClass);
27340 if (VT == MVT::i64 || VT == MVT::f64)
27341 return std::make_pair(0U, &X86::GR64RegClass);
27344 // 32-bit fallthrough
27345 case 'Q': // Q_REGS
27346 if (VT == MVT::i32 || VT == MVT::f32)
27347 return std::make_pair(0U, &X86::GR32_ABCDRegClass);
27348 if (VT == MVT::i16)
27349 return std::make_pair(0U, &X86::GR16_ABCDRegClass);
27350 if (VT == MVT::i8 || VT == MVT::i1)
27351 return std::make_pair(0U, &X86::GR8_ABCD_LRegClass);
27352 if (VT == MVT::i64)
27353 return std::make_pair(0U, &X86::GR64_ABCDRegClass);
27355 case 'r': // GENERAL_REGS
27356 case 'l': // INDEX_REGS
27357 if (VT == MVT::i8 || VT == MVT::i1)
27358 return std::make_pair(0U, &X86::GR8RegClass);
27359 if (VT == MVT::i16)
27360 return std::make_pair(0U, &X86::GR16RegClass);
27361 if (VT == MVT::i32 || VT == MVT::f32 || !Subtarget->is64Bit())
27362 return std::make_pair(0U, &X86::GR32RegClass);
27363 return std::make_pair(0U, &X86::GR64RegClass);
27364 case 'R': // LEGACY_REGS
27365 if (VT == MVT::i8 || VT == MVT::i1)
27366 return std::make_pair(0U, &X86::GR8_NOREXRegClass);
27367 if (VT == MVT::i16)
27368 return std::make_pair(0U, &X86::GR16_NOREXRegClass);
27369 if (VT == MVT::i32 || !Subtarget->is64Bit())
27370 return std::make_pair(0U, &X86::GR32_NOREXRegClass);
27371 return std::make_pair(0U, &X86::GR64_NOREXRegClass);
27372 case 'f': // FP Stack registers.
27373 // If SSE is enabled for this VT, use f80 to ensure the isel moves the
27374 // value to the correct fpstack register class.
27375 if (VT == MVT::f32 && !isScalarFPTypeInSSEReg(VT))
27376 return std::make_pair(0U, &X86::RFP32RegClass);
27377 if (VT == MVT::f64 && !isScalarFPTypeInSSEReg(VT))
27378 return std::make_pair(0U, &X86::RFP64RegClass);
27379 return std::make_pair(0U, &X86::RFP80RegClass);
27380 case 'y': // MMX_REGS if MMX allowed.
27381 if (!Subtarget->hasMMX()) break;
27382 return std::make_pair(0U, &X86::VR64RegClass);
27383 case 'Y': // SSE_REGS if SSE2 allowed
27384 if (!Subtarget->hasSSE2()) break;
27386 case 'x': // SSE_REGS if SSE1 allowed or AVX_REGS if AVX allowed
27387 if (!Subtarget->hasSSE1()) break;
27389 switch (VT.SimpleTy) {
27391 // Scalar SSE types.
27394 return std::make_pair(0U, &X86::FR32RegClass);
27397 return std::make_pair(0U, &X86::FR64RegClass);
27405 return std::make_pair(0U, &X86::VR128RegClass);
27413 return std::make_pair(0U, &X86::VR256RegClass);
27418 return std::make_pair(0U, &X86::VR512RegClass);
27424 // Use the default implementation in TargetLowering to convert the register
27425 // constraint into a member of a register class.
27426 std::pair<unsigned, const TargetRegisterClass*> Res;
27427 Res = TargetLowering::getRegForInlineAsmConstraint(TRI, Constraint, VT);
27429 // Not found as a standard register?
27431 // Map st(0) -> st(7) -> ST0
27432 if (Constraint.size() == 7 && Constraint[0] == '{' &&
27433 tolower(Constraint[1]) == 's' &&
27434 tolower(Constraint[2]) == 't' &&
27435 Constraint[3] == '(' &&
27436 (Constraint[4] >= '0' && Constraint[4] <= '7') &&
27437 Constraint[5] == ')' &&
27438 Constraint[6] == '}') {
27440 Res.first = X86::FP0+Constraint[4]-'0';
27441 Res.second = &X86::RFP80RegClass;
27445 // GCC allows "st(0)" to be called just plain "st".
27446 if (StringRef("{st}").equals_lower(Constraint)) {
27447 Res.first = X86::FP0;
27448 Res.second = &X86::RFP80RegClass;
27453 if (StringRef("{flags}").equals_lower(Constraint)) {
27454 Res.first = X86::EFLAGS;
27455 Res.second = &X86::CCRRegClass;
27459 // 'A' means EAX + EDX.
27460 if (Constraint == "A") {
27461 Res.first = X86::EAX;
27462 Res.second = &X86::GR32_ADRegClass;
27468 // Otherwise, check to see if this is a register class of the wrong value
27469 // type. For example, we want to map "{ax},i32" -> {eax}, we don't want it to
27470 // turn into {ax},{dx}.
27471 // MVT::Other is used to specify clobber names.
27472 if (Res.second->hasType(VT) || VT == MVT::Other)
27473 return Res; // Correct type already, nothing to do.
27475 // Get a matching integer of the correct size. i.e. "ax" with MVT::32 should
27476 // return "eax". This should even work for things like getting 64bit integer
27477 // registers when given an f64 type.
27478 const TargetRegisterClass *Class = Res.second;
27479 if (Class == &X86::GR8RegClass || Class == &X86::GR16RegClass ||
27480 Class == &X86::GR32RegClass || Class == &X86::GR64RegClass) {
27481 unsigned Size = VT.getSizeInBits();
27482 MVT::SimpleValueType SimpleTy = Size == 1 || Size == 8 ? MVT::i8
27483 : Size == 16 ? MVT::i16
27484 : Size == 32 ? MVT::i32
27485 : Size == 64 ? MVT::i64
27487 unsigned DestReg = getX86SubSuperRegisterOrZero(Res.first, SimpleTy);
27489 Res.first = DestReg;
27490 Res.second = SimpleTy == MVT::i8 ? &X86::GR8RegClass
27491 : SimpleTy == MVT::i16 ? &X86::GR16RegClass
27492 : SimpleTy == MVT::i32 ? &X86::GR32RegClass
27493 : &X86::GR64RegClass;
27494 assert(Res.second->contains(Res.first) && "Register in register class");
27496 // No register found/type mismatch.
27498 Res.second = nullptr;
27500 } else if (Class == &X86::FR32RegClass || Class == &X86::FR64RegClass ||
27501 Class == &X86::VR128RegClass || Class == &X86::VR256RegClass ||
27502 Class == &X86::FR32XRegClass || Class == &X86::FR64XRegClass ||
27503 Class == &X86::VR128XRegClass || Class == &X86::VR256XRegClass ||
27504 Class == &X86::VR512RegClass) {
27505 // Handle references to XMM physical registers that got mapped into the
27506 // wrong class. This can happen with constraints like {xmm0} where the
27507 // target independent register mapper will just pick the first match it can
27508 // find, ignoring the required type.
27510 if (VT == MVT::f32 || VT == MVT::i32)
27511 Res.second = &X86::FR32RegClass;
27512 else if (VT == MVT::f64 || VT == MVT::i64)
27513 Res.second = &X86::FR64RegClass;
27514 else if (X86::VR128RegClass.hasType(VT))
27515 Res.second = &X86::VR128RegClass;
27516 else if (X86::VR256RegClass.hasType(VT))
27517 Res.second = &X86::VR256RegClass;
27518 else if (X86::VR512RegClass.hasType(VT))
27519 Res.second = &X86::VR512RegClass;
27521 // Type mismatch and not a clobber: Return an error;
27523 Res.second = nullptr;
27530 int X86TargetLowering::getScalingFactorCost(const DataLayout &DL,
27531 const AddrMode &AM, Type *Ty,
27532 unsigned AS) const {
27533 // Scaling factors are not free at all.
27534 // An indexed folded instruction, i.e., inst (reg1, reg2, scale),
27535 // will take 2 allocations in the out of order engine instead of 1
27536 // for plain addressing mode, i.e. inst (reg1).
27538 // vaddps (%rsi,%drx), %ymm0, %ymm1
27539 // Requires two allocations (one for the load, one for the computation)
27541 // vaddps (%rsi), %ymm0, %ymm1
27542 // Requires just 1 allocation, i.e., freeing allocations for other operations
27543 // and having less micro operations to execute.
27545 // For some X86 architectures, this is even worse because for instance for
27546 // stores, the complex addressing mode forces the instruction to use the
27547 // "load" ports instead of the dedicated "store" port.
27548 // E.g., on Haswell:
27549 // vmovaps %ymm1, (%r8, %rdi) can use port 2 or 3.
27550 // vmovaps %ymm1, (%r8) can use port 2, 3, or 7.
27551 if (isLegalAddressingMode(DL, AM, Ty, AS))
27552 // Scale represents reg2 * scale, thus account for 1
27553 // as soon as we use a second register.
27554 return AM.Scale != 0;
27558 bool X86TargetLowering::isIntDivCheap(EVT VT, AttributeSet Attr) const {
27559 // Integer division on x86 is expensive. However, when aggressively optimizing
27560 // for code size, we prefer to use a div instruction, as it is usually smaller
27561 // than the alternative sequence.
27562 // The exception to this is vector division. Since x86 doesn't have vector
27563 // integer division, leaving the division as-is is a loss even in terms of
27564 // size, because it will have to be scalarized, while the alternative code
27565 // sequence can be performed in vector form.
27566 bool OptSize = Attr.hasAttribute(AttributeSet::FunctionIndex,
27567 Attribute::MinSize);
27568 return OptSize && !VT.isVector();