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 setOperationAction(ISD::UINT_TO_FP , MVT::i32 , Promote);
168 setOperationAction(ISD::UINT_TO_FP , MVT::i64 , Custom);
169 } else if (!Subtarget->useSoftFloat()) {
170 // We have an algorithm for SSE2->double, and we turn this into a
171 // 64-bit FILD followed by conditional FADD for other targets.
172 setOperationAction(ISD::UINT_TO_FP , MVT::i64 , Custom);
173 // We have an algorithm for SSE2, and we turn this into a 64-bit
174 // FILD for other targets.
175 setOperationAction(ISD::UINT_TO_FP , MVT::i32 , Custom);
178 // Promote i1/i8 SINT_TO_FP to larger SINT_TO_FP's, as X86 doesn't have
180 setOperationAction(ISD::SINT_TO_FP , MVT::i1 , Promote);
181 setOperationAction(ISD::SINT_TO_FP , MVT::i8 , Promote);
183 if (!Subtarget->useSoftFloat()) {
184 // SSE has no i16 to fp conversion, only i32
185 if (X86ScalarSSEf32) {
186 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Promote);
187 // f32 and f64 cases are Legal, f80 case is not
188 setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Custom);
190 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Custom);
191 setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Custom);
194 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Promote);
195 setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Promote);
198 // In 32-bit mode these are custom lowered. In 64-bit mode F32 and F64
199 // are Legal, f80 is custom lowered.
200 setOperationAction(ISD::FP_TO_SINT , MVT::i64 , Custom);
201 setOperationAction(ISD::SINT_TO_FP , MVT::i64 , Custom);
203 // Promote i1/i8 FP_TO_SINT to larger FP_TO_SINTS's, as X86 doesn't have
205 setOperationAction(ISD::FP_TO_SINT , MVT::i1 , Promote);
206 setOperationAction(ISD::FP_TO_SINT , MVT::i8 , Promote);
208 if (X86ScalarSSEf32) {
209 setOperationAction(ISD::FP_TO_SINT , MVT::i16 , Promote);
210 // f32 and f64 cases are Legal, f80 case is not
211 setOperationAction(ISD::FP_TO_SINT , MVT::i32 , Custom);
213 setOperationAction(ISD::FP_TO_SINT , MVT::i16 , Custom);
214 setOperationAction(ISD::FP_TO_SINT , MVT::i32 , Custom);
217 // Handle FP_TO_UINT by promoting the destination to a larger signed
219 setOperationAction(ISD::FP_TO_UINT , MVT::i1 , Promote);
220 setOperationAction(ISD::FP_TO_UINT , MVT::i8 , Promote);
221 setOperationAction(ISD::FP_TO_UINT , MVT::i16 , Promote);
223 if (Subtarget->is64Bit()) {
224 if (!Subtarget->useSoftFloat() && Subtarget->hasAVX512()) {
225 // FP_TO_UINT-i32/i64 is legal for f32/f64, but custom for f80.
226 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Custom);
227 setOperationAction(ISD::FP_TO_UINT , MVT::i64 , Custom);
229 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Promote);
230 setOperationAction(ISD::FP_TO_UINT , MVT::i64 , Expand);
232 } else if (!Subtarget->useSoftFloat()) {
233 // Since AVX is a superset of SSE3, only check for SSE here.
234 if (Subtarget->hasSSE1() && !Subtarget->hasSSE3())
235 // Expand FP_TO_UINT into a select.
236 // FIXME: We would like to use a Custom expander here eventually to do
237 // the optimal thing for SSE vs. the default expansion in the legalizer.
238 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Expand);
240 // With AVX512 we can use vcvts[ds]2usi for f32/f64->i32, f80 is custom.
241 // With SSE3 we can use fisttpll to convert to a signed i64; without
242 // SSE, we're stuck with a fistpll.
243 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Custom);
245 setOperationAction(ISD::FP_TO_UINT , MVT::i64 , Custom);
248 // TODO: when we have SSE, these could be more efficient, by using movd/movq.
249 if (!X86ScalarSSEf64) {
250 setOperationAction(ISD::BITCAST , MVT::f32 , Expand);
251 setOperationAction(ISD::BITCAST , MVT::i32 , Expand);
252 if (Subtarget->is64Bit()) {
253 setOperationAction(ISD::BITCAST , MVT::f64 , Expand);
254 // Without SSE, i64->f64 goes through memory.
255 setOperationAction(ISD::BITCAST , MVT::i64 , Expand);
259 // Scalar integer divide and remainder are lowered to use operations that
260 // produce two results, to match the available instructions. This exposes
261 // the two-result form to trivial CSE, which is able to combine x/y and x%y
262 // into a single instruction.
264 // Scalar integer multiply-high is also lowered to use two-result
265 // operations, to match the available instructions. However, plain multiply
266 // (low) operations are left as Legal, as there are single-result
267 // instructions for this in x86. Using the two-result multiply instructions
268 // when both high and low results are needed must be arranged by dagcombine.
269 for (unsigned i = 0; i != array_lengthof(IntVTs); ++i) {
271 setOperationAction(ISD::MULHS, VT, Expand);
272 setOperationAction(ISD::MULHU, VT, Expand);
273 setOperationAction(ISD::SDIV, VT, Expand);
274 setOperationAction(ISD::UDIV, VT, Expand);
275 setOperationAction(ISD::SREM, VT, Expand);
276 setOperationAction(ISD::UREM, VT, Expand);
278 // Add/Sub overflow ops with MVT::Glues are lowered to EFLAGS dependences.
279 setOperationAction(ISD::ADDC, VT, Custom);
280 setOperationAction(ISD::ADDE, VT, Custom);
281 setOperationAction(ISD::SUBC, VT, Custom);
282 setOperationAction(ISD::SUBE, VT, Custom);
285 setOperationAction(ISD::BR_JT , MVT::Other, Expand);
286 setOperationAction(ISD::BRCOND , MVT::Other, Custom);
287 setOperationAction(ISD::BR_CC , MVT::f32, Expand);
288 setOperationAction(ISD::BR_CC , MVT::f64, Expand);
289 setOperationAction(ISD::BR_CC , MVT::f80, Expand);
290 setOperationAction(ISD::BR_CC , MVT::i8, Expand);
291 setOperationAction(ISD::BR_CC , MVT::i16, Expand);
292 setOperationAction(ISD::BR_CC , MVT::i32, Expand);
293 setOperationAction(ISD::BR_CC , MVT::i64, Expand);
294 setOperationAction(ISD::SELECT_CC , MVT::f32, Expand);
295 setOperationAction(ISD::SELECT_CC , MVT::f64, Expand);
296 setOperationAction(ISD::SELECT_CC , MVT::f80, Expand);
297 setOperationAction(ISD::SELECT_CC , MVT::i8, Expand);
298 setOperationAction(ISD::SELECT_CC , MVT::i16, Expand);
299 setOperationAction(ISD::SELECT_CC , MVT::i32, Expand);
300 setOperationAction(ISD::SELECT_CC , MVT::i64, Expand);
301 if (Subtarget->is64Bit())
302 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i32, Legal);
303 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16 , Legal);
304 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8 , Legal);
305 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1 , Expand);
306 setOperationAction(ISD::FP_ROUND_INREG , MVT::f32 , Expand);
308 if (Subtarget->is32Bit() && Subtarget->isTargetKnownWindowsMSVC()) {
309 // On 32 bit MSVC, `fmodf(f32)` is not defined - only `fmod(f64)`
310 // is. We should promote the value to 64-bits to solve this.
311 // This is what the CRT headers do - `fmodf` is an inline header
312 // function casting to f64 and calling `fmod`.
313 setOperationAction(ISD::FREM , MVT::f32 , Promote);
315 setOperationAction(ISD::FREM , MVT::f32 , Expand);
318 setOperationAction(ISD::FREM , MVT::f64 , Expand);
319 setOperationAction(ISD::FREM , MVT::f80 , Expand);
320 setOperationAction(ISD::FLT_ROUNDS_ , MVT::i32 , Custom);
322 // Promote the i8 variants and force them on up to i32 which has a shorter
324 setOperationAction(ISD::CTTZ , MVT::i8 , Promote);
325 AddPromotedToType (ISD::CTTZ , MVT::i8 , MVT::i32);
326 setOperationAction(ISD::CTTZ_ZERO_UNDEF , MVT::i8 , Promote);
327 AddPromotedToType (ISD::CTTZ_ZERO_UNDEF , MVT::i8 , MVT::i32);
328 if (Subtarget->hasBMI()) {
329 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::i16 , Expand);
330 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::i32 , Expand);
331 if (Subtarget->is64Bit())
332 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::i64, Expand);
334 setOperationAction(ISD::CTTZ , MVT::i16 , Custom);
335 setOperationAction(ISD::CTTZ , MVT::i32 , Custom);
336 if (Subtarget->is64Bit())
337 setOperationAction(ISD::CTTZ , MVT::i64 , Custom);
340 if (Subtarget->hasLZCNT()) {
341 // When promoting the i8 variants, force them to i32 for a shorter
343 setOperationAction(ISD::CTLZ , MVT::i8 , Promote);
344 AddPromotedToType (ISD::CTLZ , MVT::i8 , MVT::i32);
345 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i8 , Promote);
346 AddPromotedToType (ISD::CTLZ_ZERO_UNDEF, MVT::i8 , MVT::i32);
347 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i16 , Expand);
348 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i32 , Expand);
349 if (Subtarget->is64Bit())
350 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i64, Expand);
352 setOperationAction(ISD::CTLZ , MVT::i8 , Custom);
353 setOperationAction(ISD::CTLZ , MVT::i16 , Custom);
354 setOperationAction(ISD::CTLZ , MVT::i32 , Custom);
355 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i8 , Custom);
356 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i16 , Custom);
357 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i32 , Custom);
358 if (Subtarget->is64Bit()) {
359 setOperationAction(ISD::CTLZ , MVT::i64 , Custom);
360 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i64, Custom);
364 // Special handling for half-precision floating point conversions.
365 // If we don't have F16C support, then lower half float conversions
366 // into library calls.
367 if (Subtarget->useSoftFloat() || !Subtarget->hasF16C()) {
368 setOperationAction(ISD::FP16_TO_FP, MVT::f32, Expand);
369 setOperationAction(ISD::FP_TO_FP16, MVT::f32, Expand);
372 // There's never any support for operations beyond MVT::f32.
373 setOperationAction(ISD::FP16_TO_FP, MVT::f64, Expand);
374 setOperationAction(ISD::FP16_TO_FP, MVT::f80, Expand);
375 setOperationAction(ISD::FP_TO_FP16, MVT::f64, Expand);
376 setOperationAction(ISD::FP_TO_FP16, MVT::f80, Expand);
378 setLoadExtAction(ISD::EXTLOAD, MVT::f32, MVT::f16, Expand);
379 setLoadExtAction(ISD::EXTLOAD, MVT::f64, MVT::f16, Expand);
380 setLoadExtAction(ISD::EXTLOAD, MVT::f80, MVT::f16, Expand);
381 setTruncStoreAction(MVT::f32, MVT::f16, Expand);
382 setTruncStoreAction(MVT::f64, MVT::f16, Expand);
383 setTruncStoreAction(MVT::f80, MVT::f16, Expand);
385 if (Subtarget->hasPOPCNT()) {
386 setOperationAction(ISD::CTPOP , MVT::i8 , Promote);
388 setOperationAction(ISD::CTPOP , MVT::i8 , Expand);
389 setOperationAction(ISD::CTPOP , MVT::i16 , Expand);
390 setOperationAction(ISD::CTPOP , MVT::i32 , Expand);
391 if (Subtarget->is64Bit())
392 setOperationAction(ISD::CTPOP , MVT::i64 , Expand);
395 setOperationAction(ISD::READCYCLECOUNTER , MVT::i64 , Custom);
397 if (!Subtarget->hasMOVBE())
398 setOperationAction(ISD::BSWAP , MVT::i16 , Expand);
400 // These should be promoted to a larger select which is supported.
401 setOperationAction(ISD::SELECT , MVT::i1 , Promote);
402 // X86 wants to expand cmov itself.
403 setOperationAction(ISD::SELECT , MVT::i8 , Custom);
404 setOperationAction(ISD::SELECT , MVT::i16 , Custom);
405 setOperationAction(ISD::SELECT , MVT::i32 , Custom);
406 setOperationAction(ISD::SELECT , MVT::f32 , Custom);
407 setOperationAction(ISD::SELECT , MVT::f64 , Custom);
408 setOperationAction(ISD::SELECT , MVT::f80 , Custom);
409 setOperationAction(ISD::SETCC , MVT::i8 , Custom);
410 setOperationAction(ISD::SETCC , MVT::i16 , Custom);
411 setOperationAction(ISD::SETCC , MVT::i32 , Custom);
412 setOperationAction(ISD::SETCC , MVT::f32 , Custom);
413 setOperationAction(ISD::SETCC , MVT::f64 , Custom);
414 setOperationAction(ISD::SETCC , MVT::f80 , Custom);
415 if (Subtarget->is64Bit()) {
416 setOperationAction(ISD::SELECT , MVT::i64 , Custom);
417 setOperationAction(ISD::SETCC , MVT::i64 , Custom);
419 setOperationAction(ISD::EH_RETURN , MVT::Other, Custom);
420 // NOTE: EH_SJLJ_SETJMP/_LONGJMP supported here is NOT intended to support
421 // SjLj exception handling but a light-weight setjmp/longjmp replacement to
422 // support continuation, user-level threading, and etc.. As a result, no
423 // other SjLj exception interfaces are implemented and please don't build
424 // your own exception handling based on them.
425 // LLVM/Clang supports zero-cost DWARF exception handling.
426 setOperationAction(ISD::EH_SJLJ_SETJMP, MVT::i32, Custom);
427 setOperationAction(ISD::EH_SJLJ_LONGJMP, MVT::Other, Custom);
430 setOperationAction(ISD::ConstantPool , MVT::i32 , Custom);
431 setOperationAction(ISD::JumpTable , MVT::i32 , Custom);
432 setOperationAction(ISD::GlobalAddress , MVT::i32 , Custom);
433 setOperationAction(ISD::GlobalTLSAddress, MVT::i32 , Custom);
434 if (Subtarget->is64Bit())
435 setOperationAction(ISD::GlobalTLSAddress, MVT::i64, Custom);
436 setOperationAction(ISD::ExternalSymbol , MVT::i32 , Custom);
437 setOperationAction(ISD::BlockAddress , MVT::i32 , Custom);
438 if (Subtarget->is64Bit()) {
439 setOperationAction(ISD::ConstantPool , MVT::i64 , Custom);
440 setOperationAction(ISD::JumpTable , MVT::i64 , Custom);
441 setOperationAction(ISD::GlobalAddress , MVT::i64 , Custom);
442 setOperationAction(ISD::ExternalSymbol, MVT::i64 , Custom);
443 setOperationAction(ISD::BlockAddress , MVT::i64 , Custom);
445 // 64-bit addm sub, shl, sra, srl (iff 32-bit x86)
446 setOperationAction(ISD::SHL_PARTS , MVT::i32 , Custom);
447 setOperationAction(ISD::SRA_PARTS , MVT::i32 , Custom);
448 setOperationAction(ISD::SRL_PARTS , MVT::i32 , Custom);
449 if (Subtarget->is64Bit()) {
450 setOperationAction(ISD::SHL_PARTS , MVT::i64 , Custom);
451 setOperationAction(ISD::SRA_PARTS , MVT::i64 , Custom);
452 setOperationAction(ISD::SRL_PARTS , MVT::i64 , Custom);
455 if (Subtarget->hasSSE1())
456 setOperationAction(ISD::PREFETCH , MVT::Other, Legal);
458 setOperationAction(ISD::ATOMIC_FENCE , MVT::Other, Custom);
460 // Expand certain atomics
461 for (unsigned i = 0; i != array_lengthof(IntVTs); ++i) {
463 setOperationAction(ISD::ATOMIC_CMP_SWAP_WITH_SUCCESS, VT, Custom);
464 setOperationAction(ISD::ATOMIC_LOAD_SUB, VT, Custom);
465 setOperationAction(ISD::ATOMIC_STORE, VT, Custom);
468 if (Subtarget->hasCmpxchg16b()) {
469 setOperationAction(ISD::ATOMIC_CMP_SWAP_WITH_SUCCESS, MVT::i128, Custom);
472 // FIXME - use subtarget debug flags
473 if (!Subtarget->isTargetDarwin() && !Subtarget->isTargetELF() &&
474 !Subtarget->isTargetCygMing() && !Subtarget->isTargetWin64()) {
475 setOperationAction(ISD::EH_LABEL, MVT::Other, Expand);
478 if (Subtarget->isTarget64BitLP64()) {
479 setExceptionPointerRegister(X86::RAX);
480 setExceptionSelectorRegister(X86::RDX);
482 setExceptionPointerRegister(X86::EAX);
483 setExceptionSelectorRegister(X86::EDX);
485 setOperationAction(ISD::FRAME_TO_ARGS_OFFSET, MVT::i32, Custom);
486 setOperationAction(ISD::FRAME_TO_ARGS_OFFSET, MVT::i64, Custom);
488 setOperationAction(ISD::INIT_TRAMPOLINE, MVT::Other, Custom);
489 setOperationAction(ISD::ADJUST_TRAMPOLINE, MVT::Other, Custom);
491 setOperationAction(ISD::TRAP, MVT::Other, Legal);
492 setOperationAction(ISD::DEBUGTRAP, MVT::Other, Legal);
494 // VASTART needs to be custom lowered to use the VarArgsFrameIndex
495 setOperationAction(ISD::VASTART , MVT::Other, Custom);
496 setOperationAction(ISD::VAEND , MVT::Other, Expand);
497 if (Subtarget->is64Bit()) {
498 setOperationAction(ISD::VAARG , MVT::Other, Custom);
499 setOperationAction(ISD::VACOPY , MVT::Other, Custom);
501 // TargetInfo::CharPtrBuiltinVaList
502 setOperationAction(ISD::VAARG , MVT::Other, Expand);
503 setOperationAction(ISD::VACOPY , MVT::Other, Expand);
506 setOperationAction(ISD::STACKSAVE, MVT::Other, Expand);
507 setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand);
509 setOperationAction(ISD::DYNAMIC_STACKALLOC, PtrVT, Custom);
511 // GC_TRANSITION_START and GC_TRANSITION_END need custom lowering.
512 setOperationAction(ISD::GC_TRANSITION_START, MVT::Other, Custom);
513 setOperationAction(ISD::GC_TRANSITION_END, MVT::Other, Custom);
515 if (!Subtarget->useSoftFloat() && X86ScalarSSEf64) {
516 // f32 and f64 use SSE.
517 // Set up the FP register classes.
518 addRegisterClass(MVT::f32, &X86::FR32RegClass);
519 addRegisterClass(MVT::f64, &X86::FR64RegClass);
521 // Use ANDPD to simulate FABS.
522 setOperationAction(ISD::FABS , MVT::f64, Custom);
523 setOperationAction(ISD::FABS , MVT::f32, Custom);
525 // Use XORP to simulate FNEG.
526 setOperationAction(ISD::FNEG , MVT::f64, Custom);
527 setOperationAction(ISD::FNEG , MVT::f32, Custom);
529 // Use ANDPD and ORPD to simulate FCOPYSIGN.
530 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Custom);
531 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
533 // Lower this to FGETSIGNx86 plus an AND.
534 setOperationAction(ISD::FGETSIGN, MVT::i64, Custom);
535 setOperationAction(ISD::FGETSIGN, MVT::i32, Custom);
537 // We don't support sin/cos/fmod
538 setOperationAction(ISD::FSIN , MVT::f64, Expand);
539 setOperationAction(ISD::FCOS , MVT::f64, Expand);
540 setOperationAction(ISD::FSINCOS, MVT::f64, Expand);
541 setOperationAction(ISD::FSIN , MVT::f32, Expand);
542 setOperationAction(ISD::FCOS , MVT::f32, Expand);
543 setOperationAction(ISD::FSINCOS, MVT::f32, Expand);
545 // Expand FP immediates into loads from the stack, except for the special
547 addLegalFPImmediate(APFloat(+0.0)); // xorpd
548 addLegalFPImmediate(APFloat(+0.0f)); // xorps
549 } else if (!Subtarget->useSoftFloat() && X86ScalarSSEf32) {
550 // Use SSE for f32, x87 for f64.
551 // Set up the FP register classes.
552 addRegisterClass(MVT::f32, &X86::FR32RegClass);
553 addRegisterClass(MVT::f64, &X86::RFP64RegClass);
555 // Use ANDPS to simulate FABS.
556 setOperationAction(ISD::FABS , MVT::f32, Custom);
558 // Use XORP to simulate FNEG.
559 setOperationAction(ISD::FNEG , MVT::f32, Custom);
561 setOperationAction(ISD::UNDEF, MVT::f64, Expand);
563 // Use ANDPS and ORPS to simulate FCOPYSIGN.
564 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
565 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
567 // We don't support sin/cos/fmod
568 setOperationAction(ISD::FSIN , MVT::f32, Expand);
569 setOperationAction(ISD::FCOS , MVT::f32, Expand);
570 setOperationAction(ISD::FSINCOS, MVT::f32, Expand);
572 // Special cases we handle for FP constants.
573 addLegalFPImmediate(APFloat(+0.0f)); // xorps
574 addLegalFPImmediate(APFloat(+0.0)); // FLD0
575 addLegalFPImmediate(APFloat(+1.0)); // FLD1
576 addLegalFPImmediate(APFloat(-0.0)); // FLD0/FCHS
577 addLegalFPImmediate(APFloat(-1.0)); // FLD1/FCHS
579 if (!TM.Options.UnsafeFPMath) {
580 setOperationAction(ISD::FSIN , MVT::f64, Expand);
581 setOperationAction(ISD::FCOS , MVT::f64, Expand);
582 setOperationAction(ISD::FSINCOS, MVT::f64, Expand);
584 } else if (!Subtarget->useSoftFloat()) {
585 // f32 and f64 in x87.
586 // Set up the FP register classes.
587 addRegisterClass(MVT::f64, &X86::RFP64RegClass);
588 addRegisterClass(MVT::f32, &X86::RFP32RegClass);
590 setOperationAction(ISD::UNDEF, MVT::f64, Expand);
591 setOperationAction(ISD::UNDEF, MVT::f32, Expand);
592 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
593 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Expand);
595 if (!TM.Options.UnsafeFPMath) {
596 setOperationAction(ISD::FSIN , MVT::f64, Expand);
597 setOperationAction(ISD::FSIN , MVT::f32, Expand);
598 setOperationAction(ISD::FCOS , MVT::f64, Expand);
599 setOperationAction(ISD::FCOS , MVT::f32, Expand);
600 setOperationAction(ISD::FSINCOS, MVT::f64, Expand);
601 setOperationAction(ISD::FSINCOS, MVT::f32, Expand);
603 addLegalFPImmediate(APFloat(+0.0)); // FLD0
604 addLegalFPImmediate(APFloat(+1.0)); // FLD1
605 addLegalFPImmediate(APFloat(-0.0)); // FLD0/FCHS
606 addLegalFPImmediate(APFloat(-1.0)); // FLD1/FCHS
607 addLegalFPImmediate(APFloat(+0.0f)); // FLD0
608 addLegalFPImmediate(APFloat(+1.0f)); // FLD1
609 addLegalFPImmediate(APFloat(-0.0f)); // FLD0/FCHS
610 addLegalFPImmediate(APFloat(-1.0f)); // FLD1/FCHS
613 // We don't support FMA.
614 setOperationAction(ISD::FMA, MVT::f64, Expand);
615 setOperationAction(ISD::FMA, MVT::f32, Expand);
617 // Long double always uses X87.
618 if (!Subtarget->useSoftFloat()) {
619 addRegisterClass(MVT::f80, &X86::RFP80RegClass);
620 setOperationAction(ISD::UNDEF, MVT::f80, Expand);
621 setOperationAction(ISD::FCOPYSIGN, MVT::f80, Expand);
623 APFloat TmpFlt = APFloat::getZero(APFloat::x87DoubleExtended);
624 addLegalFPImmediate(TmpFlt); // FLD0
626 addLegalFPImmediate(TmpFlt); // FLD0/FCHS
629 APFloat TmpFlt2(+1.0);
630 TmpFlt2.convert(APFloat::x87DoubleExtended, APFloat::rmNearestTiesToEven,
632 addLegalFPImmediate(TmpFlt2); // FLD1
633 TmpFlt2.changeSign();
634 addLegalFPImmediate(TmpFlt2); // FLD1/FCHS
637 if (!TM.Options.UnsafeFPMath) {
638 setOperationAction(ISD::FSIN , MVT::f80, Expand);
639 setOperationAction(ISD::FCOS , MVT::f80, Expand);
640 setOperationAction(ISD::FSINCOS, MVT::f80, Expand);
643 setOperationAction(ISD::FFLOOR, MVT::f80, Expand);
644 setOperationAction(ISD::FCEIL, MVT::f80, Expand);
645 setOperationAction(ISD::FTRUNC, MVT::f80, Expand);
646 setOperationAction(ISD::FRINT, MVT::f80, Expand);
647 setOperationAction(ISD::FNEARBYINT, MVT::f80, Expand);
648 setOperationAction(ISD::FMA, MVT::f80, Expand);
651 // Always use a library call for pow.
652 setOperationAction(ISD::FPOW , MVT::f32 , Expand);
653 setOperationAction(ISD::FPOW , MVT::f64 , Expand);
654 setOperationAction(ISD::FPOW , MVT::f80 , Expand);
656 setOperationAction(ISD::FLOG, MVT::f80, Expand);
657 setOperationAction(ISD::FLOG2, MVT::f80, Expand);
658 setOperationAction(ISD::FLOG10, MVT::f80, Expand);
659 setOperationAction(ISD::FEXP, MVT::f80, Expand);
660 setOperationAction(ISD::FEXP2, MVT::f80, Expand);
661 setOperationAction(ISD::FMINNUM, MVT::f80, Expand);
662 setOperationAction(ISD::FMAXNUM, MVT::f80, Expand);
664 // First set operation action for all vector types to either promote
665 // (for widening) or expand (for scalarization). Then we will selectively
666 // turn on ones that can be effectively codegen'd.
667 for (MVT VT : MVT::vector_valuetypes()) {
668 setOperationAction(ISD::ADD , VT, Expand);
669 setOperationAction(ISD::SUB , VT, Expand);
670 setOperationAction(ISD::FADD, VT, Expand);
671 setOperationAction(ISD::FNEG, VT, Expand);
672 setOperationAction(ISD::FSUB, VT, Expand);
673 setOperationAction(ISD::MUL , VT, Expand);
674 setOperationAction(ISD::FMUL, VT, Expand);
675 setOperationAction(ISD::SDIV, VT, Expand);
676 setOperationAction(ISD::UDIV, VT, Expand);
677 setOperationAction(ISD::FDIV, VT, Expand);
678 setOperationAction(ISD::SREM, VT, Expand);
679 setOperationAction(ISD::UREM, VT, Expand);
680 setOperationAction(ISD::LOAD, VT, Expand);
681 setOperationAction(ISD::VECTOR_SHUFFLE, VT, Expand);
682 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT,Expand);
683 setOperationAction(ISD::INSERT_VECTOR_ELT, VT, Expand);
684 setOperationAction(ISD::EXTRACT_SUBVECTOR, VT,Expand);
685 setOperationAction(ISD::INSERT_SUBVECTOR, VT,Expand);
686 setOperationAction(ISD::FABS, VT, Expand);
687 setOperationAction(ISD::FSIN, VT, Expand);
688 setOperationAction(ISD::FSINCOS, VT, Expand);
689 setOperationAction(ISD::FCOS, VT, Expand);
690 setOperationAction(ISD::FSINCOS, VT, Expand);
691 setOperationAction(ISD::FREM, VT, Expand);
692 setOperationAction(ISD::FMA, VT, Expand);
693 setOperationAction(ISD::FPOWI, VT, Expand);
694 setOperationAction(ISD::FSQRT, VT, Expand);
695 setOperationAction(ISD::FCOPYSIGN, VT, Expand);
696 setOperationAction(ISD::FFLOOR, VT, Expand);
697 setOperationAction(ISD::FCEIL, VT, Expand);
698 setOperationAction(ISD::FTRUNC, VT, Expand);
699 setOperationAction(ISD::FRINT, VT, Expand);
700 setOperationAction(ISD::FNEARBYINT, VT, Expand);
701 setOperationAction(ISD::SMUL_LOHI, VT, Expand);
702 setOperationAction(ISD::MULHS, VT, Expand);
703 setOperationAction(ISD::UMUL_LOHI, VT, Expand);
704 setOperationAction(ISD::MULHU, VT, Expand);
705 setOperationAction(ISD::SDIVREM, VT, Expand);
706 setOperationAction(ISD::UDIVREM, VT, Expand);
707 setOperationAction(ISD::FPOW, VT, Expand);
708 setOperationAction(ISD::CTPOP, VT, Expand);
709 setOperationAction(ISD::CTTZ, VT, Expand);
710 setOperationAction(ISD::CTTZ_ZERO_UNDEF, VT, Expand);
711 setOperationAction(ISD::CTLZ, VT, Expand);
712 setOperationAction(ISD::CTLZ_ZERO_UNDEF, VT, Expand);
713 setOperationAction(ISD::SHL, VT, Expand);
714 setOperationAction(ISD::SRA, VT, Expand);
715 setOperationAction(ISD::SRL, VT, Expand);
716 setOperationAction(ISD::ROTL, VT, Expand);
717 setOperationAction(ISD::ROTR, VT, Expand);
718 setOperationAction(ISD::BSWAP, VT, Expand);
719 setOperationAction(ISD::SETCC, VT, Expand);
720 setOperationAction(ISD::FLOG, VT, Expand);
721 setOperationAction(ISD::FLOG2, VT, Expand);
722 setOperationAction(ISD::FLOG10, VT, Expand);
723 setOperationAction(ISD::FEXP, VT, Expand);
724 setOperationAction(ISD::FEXP2, VT, Expand);
725 setOperationAction(ISD::FP_TO_UINT, VT, Expand);
726 setOperationAction(ISD::FP_TO_SINT, VT, Expand);
727 setOperationAction(ISD::UINT_TO_FP, VT, Expand);
728 setOperationAction(ISD::SINT_TO_FP, VT, Expand);
729 setOperationAction(ISD::SIGN_EXTEND_INREG, VT,Expand);
730 setOperationAction(ISD::TRUNCATE, VT, Expand);
731 setOperationAction(ISD::SIGN_EXTEND, VT, Expand);
732 setOperationAction(ISD::ZERO_EXTEND, VT, Expand);
733 setOperationAction(ISD::ANY_EXTEND, VT, Expand);
734 setOperationAction(ISD::VSELECT, VT, Expand);
735 setOperationAction(ISD::SELECT_CC, VT, Expand);
736 for (MVT InnerVT : MVT::vector_valuetypes()) {
737 setTruncStoreAction(InnerVT, VT, Expand);
739 setLoadExtAction(ISD::SEXTLOAD, InnerVT, VT, Expand);
740 setLoadExtAction(ISD::ZEXTLOAD, InnerVT, VT, Expand);
742 // N.b. ISD::EXTLOAD legality is basically ignored except for i1-like
743 // types, we have to deal with them whether we ask for Expansion or not.
744 // Setting Expand causes its own optimisation problems though, so leave
746 if (VT.getVectorElementType() == MVT::i1)
747 setLoadExtAction(ISD::EXTLOAD, InnerVT, VT, Expand);
749 // EXTLOAD for MVT::f16 vectors is not legal because f16 vectors are
750 // split/scalarized right now.
751 if (VT.getVectorElementType() == MVT::f16)
752 setLoadExtAction(ISD::EXTLOAD, InnerVT, VT, Expand);
756 // FIXME: In order to prevent SSE instructions being expanded to MMX ones
757 // with -msoft-float, disable use of MMX as well.
758 if (!Subtarget->useSoftFloat() && Subtarget->hasMMX()) {
759 addRegisterClass(MVT::x86mmx, &X86::VR64RegClass);
760 // No operations on x86mmx supported, everything uses intrinsics.
763 // MMX-sized vectors (other than x86mmx) are expected to be expanded
764 // into smaller operations.
765 for (MVT MMXTy : {MVT::v8i8, MVT::v4i16, MVT::v2i32, MVT::v1i64}) {
766 setOperationAction(ISD::MULHS, MMXTy, Expand);
767 setOperationAction(ISD::AND, MMXTy, Expand);
768 setOperationAction(ISD::OR, MMXTy, Expand);
769 setOperationAction(ISD::XOR, MMXTy, Expand);
770 setOperationAction(ISD::SCALAR_TO_VECTOR, MMXTy, Expand);
771 setOperationAction(ISD::SELECT, MMXTy, Expand);
772 setOperationAction(ISD::BITCAST, MMXTy, Expand);
774 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v1i64, Expand);
776 if (!Subtarget->useSoftFloat() && Subtarget->hasSSE1()) {
777 addRegisterClass(MVT::v4f32, &X86::VR128RegClass);
779 setOperationAction(ISD::FADD, MVT::v4f32, Legal);
780 setOperationAction(ISD::FSUB, MVT::v4f32, Legal);
781 setOperationAction(ISD::FMUL, MVT::v4f32, Legal);
782 setOperationAction(ISD::FDIV, MVT::v4f32, Legal);
783 setOperationAction(ISD::FSQRT, MVT::v4f32, Legal);
784 setOperationAction(ISD::FNEG, MVT::v4f32, Custom);
785 setOperationAction(ISD::FABS, MVT::v4f32, Custom);
786 setOperationAction(ISD::LOAD, MVT::v4f32, Legal);
787 setOperationAction(ISD::BUILD_VECTOR, MVT::v4f32, Custom);
788 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4f32, Custom);
789 setOperationAction(ISD::VSELECT, MVT::v4f32, Custom);
790 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f32, Custom);
791 setOperationAction(ISD::SELECT, MVT::v4f32, Custom);
792 setOperationAction(ISD::UINT_TO_FP, MVT::v4i32, Custom);
795 if (!Subtarget->useSoftFloat() && Subtarget->hasSSE2()) {
796 addRegisterClass(MVT::v2f64, &X86::VR128RegClass);
798 // FIXME: Unfortunately, -soft-float and -no-implicit-float mean XMM
799 // registers cannot be used even for integer operations.
800 addRegisterClass(MVT::v16i8, &X86::VR128RegClass);
801 addRegisterClass(MVT::v8i16, &X86::VR128RegClass);
802 addRegisterClass(MVT::v4i32, &X86::VR128RegClass);
803 addRegisterClass(MVT::v2i64, &X86::VR128RegClass);
805 setOperationAction(ISD::ADD, MVT::v16i8, Legal);
806 setOperationAction(ISD::ADD, MVT::v8i16, Legal);
807 setOperationAction(ISD::ADD, MVT::v4i32, Legal);
808 setOperationAction(ISD::ADD, MVT::v2i64, Legal);
809 setOperationAction(ISD::MUL, MVT::v16i8, Custom);
810 setOperationAction(ISD::MUL, MVT::v4i32, Custom);
811 setOperationAction(ISD::MUL, MVT::v2i64, Custom);
812 setOperationAction(ISD::UMUL_LOHI, MVT::v4i32, Custom);
813 setOperationAction(ISD::SMUL_LOHI, MVT::v4i32, Custom);
814 setOperationAction(ISD::MULHU, MVT::v8i16, Legal);
815 setOperationAction(ISD::MULHS, MVT::v8i16, Legal);
816 setOperationAction(ISD::SUB, MVT::v16i8, Legal);
817 setOperationAction(ISD::SUB, MVT::v8i16, Legal);
818 setOperationAction(ISD::SUB, MVT::v4i32, Legal);
819 setOperationAction(ISD::SUB, MVT::v2i64, Legal);
820 setOperationAction(ISD::MUL, MVT::v8i16, Legal);
821 setOperationAction(ISD::FADD, MVT::v2f64, Legal);
822 setOperationAction(ISD::FSUB, MVT::v2f64, Legal);
823 setOperationAction(ISD::FMUL, MVT::v2f64, Legal);
824 setOperationAction(ISD::FDIV, MVT::v2f64, Legal);
825 setOperationAction(ISD::FSQRT, MVT::v2f64, Legal);
826 setOperationAction(ISD::FNEG, MVT::v2f64, Custom);
827 setOperationAction(ISD::FABS, MVT::v2f64, Custom);
829 setOperationAction(ISD::SMAX, MVT::v8i16, Legal);
830 setOperationAction(ISD::UMAX, MVT::v16i8, Legal);
831 setOperationAction(ISD::SMIN, MVT::v8i16, Legal);
832 setOperationAction(ISD::UMIN, MVT::v16i8, Legal);
834 setOperationAction(ISD::SETCC, MVT::v2i64, Custom);
835 setOperationAction(ISD::SETCC, MVT::v16i8, Custom);
836 setOperationAction(ISD::SETCC, MVT::v8i16, Custom);
837 setOperationAction(ISD::SETCC, MVT::v4i32, Custom);
839 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v16i8, Custom);
840 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v8i16, Custom);
841 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i16, Custom);
842 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i32, Custom);
843 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f32, Custom);
845 setOperationAction(ISD::CTPOP, MVT::v16i8, Custom);
846 setOperationAction(ISD::CTPOP, MVT::v8i16, Custom);
847 setOperationAction(ISD::CTPOP, MVT::v4i32, Custom);
848 setOperationAction(ISD::CTPOP, MVT::v2i64, Custom);
850 setOperationAction(ISD::CTTZ, MVT::v16i8, Custom);
851 setOperationAction(ISD::CTTZ, MVT::v8i16, Custom);
852 setOperationAction(ISD::CTTZ, MVT::v4i32, Custom);
853 // ISD::CTTZ v2i64 - scalarization is faster.
854 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::v16i8, Custom);
855 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::v8i16, Custom);
856 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::v4i32, Custom);
857 // ISD::CTTZ_ZERO_UNDEF v2i64 - scalarization is faster.
859 // Custom lower build_vector, vector_shuffle, and extract_vector_elt.
860 for (int i = MVT::v16i8; i != MVT::v2i64; ++i) {
861 MVT VT = (MVT::SimpleValueType)i;
862 // Do not attempt to custom lower non-power-of-2 vectors
863 if (!isPowerOf2_32(VT.getVectorNumElements()))
865 // Do not attempt to custom lower non-128-bit vectors
866 if (!VT.is128BitVector())
868 setOperationAction(ISD::BUILD_VECTOR, VT, Custom);
869 setOperationAction(ISD::VECTOR_SHUFFLE, VT, Custom);
870 setOperationAction(ISD::VSELECT, VT, Custom);
871 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Custom);
874 // We support custom legalizing of sext and anyext loads for specific
875 // memory vector types which we can load as a scalar (or sequence of
876 // scalars) and extend in-register to a legal 128-bit vector type. For sext
877 // loads these must work with a single scalar load.
878 for (MVT VT : MVT::integer_vector_valuetypes()) {
879 setLoadExtAction(ISD::SEXTLOAD, VT, MVT::v4i8, Custom);
880 setLoadExtAction(ISD::SEXTLOAD, VT, MVT::v4i16, Custom);
881 setLoadExtAction(ISD::SEXTLOAD, VT, MVT::v8i8, Custom);
882 setLoadExtAction(ISD::EXTLOAD, VT, MVT::v2i8, Custom);
883 setLoadExtAction(ISD::EXTLOAD, VT, MVT::v2i16, Custom);
884 setLoadExtAction(ISD::EXTLOAD, VT, MVT::v2i32, Custom);
885 setLoadExtAction(ISD::EXTLOAD, VT, MVT::v4i8, Custom);
886 setLoadExtAction(ISD::EXTLOAD, VT, MVT::v4i16, Custom);
887 setLoadExtAction(ISD::EXTLOAD, VT, MVT::v8i8, Custom);
890 setOperationAction(ISD::BUILD_VECTOR, MVT::v2f64, Custom);
891 setOperationAction(ISD::BUILD_VECTOR, MVT::v2i64, Custom);
892 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2f64, Custom);
893 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2i64, Custom);
894 setOperationAction(ISD::VSELECT, MVT::v2f64, Custom);
895 setOperationAction(ISD::VSELECT, MVT::v2i64, Custom);
896 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2f64, Custom);
897 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2f64, Custom);
899 if (Subtarget->is64Bit()) {
900 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2i64, Custom);
901 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2i64, Custom);
904 // Promote v16i8, v8i16, v4i32 load, select, and, or, xor to v2i64.
905 for (int i = MVT::v16i8; i != MVT::v2i64; ++i) {
906 MVT VT = (MVT::SimpleValueType)i;
908 // Do not attempt to promote non-128-bit vectors
909 if (!VT.is128BitVector())
912 setOperationAction(ISD::AND, VT, Promote);
913 AddPromotedToType (ISD::AND, VT, MVT::v2i64);
914 setOperationAction(ISD::OR, VT, Promote);
915 AddPromotedToType (ISD::OR, VT, MVT::v2i64);
916 setOperationAction(ISD::XOR, VT, Promote);
917 AddPromotedToType (ISD::XOR, VT, MVT::v2i64);
918 setOperationAction(ISD::LOAD, VT, Promote);
919 AddPromotedToType (ISD::LOAD, VT, MVT::v2i64);
920 setOperationAction(ISD::SELECT, VT, Promote);
921 AddPromotedToType (ISD::SELECT, VT, MVT::v2i64);
924 // Custom lower v2i64 and v2f64 selects.
925 setOperationAction(ISD::LOAD, MVT::v2f64, Legal);
926 setOperationAction(ISD::LOAD, MVT::v2i64, Legal);
927 setOperationAction(ISD::SELECT, MVT::v2f64, Custom);
928 setOperationAction(ISD::SELECT, MVT::v2i64, Custom);
930 setOperationAction(ISD::FP_TO_SINT, MVT::v4i32, Legal);
931 setOperationAction(ISD::SINT_TO_FP, MVT::v4i32, Legal);
933 setOperationAction(ISD::SINT_TO_FP, MVT::v2i32, Custom);
935 setOperationAction(ISD::UINT_TO_FP, MVT::v4i8, Custom);
936 setOperationAction(ISD::UINT_TO_FP, MVT::v4i16, Custom);
937 // As there is no 64-bit GPR available, we need build a special custom
938 // sequence to convert from v2i32 to v2f32.
939 if (!Subtarget->is64Bit())
940 setOperationAction(ISD::UINT_TO_FP, MVT::v2f32, Custom);
942 setOperationAction(ISD::FP_EXTEND, MVT::v2f32, Custom);
943 setOperationAction(ISD::FP_ROUND, MVT::v2f32, Custom);
945 for (MVT VT : MVT::fp_vector_valuetypes())
946 setLoadExtAction(ISD::EXTLOAD, VT, MVT::v2f32, Legal);
948 setOperationAction(ISD::BITCAST, MVT::v2i32, Custom);
949 setOperationAction(ISD::BITCAST, MVT::v4i16, Custom);
950 setOperationAction(ISD::BITCAST, MVT::v8i8, Custom);
953 if (!Subtarget->useSoftFloat() && Subtarget->hasSSE41()) {
954 for (MVT RoundedTy : {MVT::f32, MVT::f64, MVT::v4f32, MVT::v2f64}) {
955 setOperationAction(ISD::FFLOOR, RoundedTy, Legal);
956 setOperationAction(ISD::FCEIL, RoundedTy, Legal);
957 setOperationAction(ISD::FTRUNC, RoundedTy, Legal);
958 setOperationAction(ISD::FRINT, RoundedTy, Legal);
959 setOperationAction(ISD::FNEARBYINT, RoundedTy, Legal);
962 setOperationAction(ISD::SMAX, MVT::v16i8, Legal);
963 setOperationAction(ISD::SMAX, MVT::v4i32, Legal);
964 setOperationAction(ISD::UMAX, MVT::v8i16, Legal);
965 setOperationAction(ISD::UMAX, MVT::v4i32, Legal);
966 setOperationAction(ISD::SMIN, MVT::v16i8, Legal);
967 setOperationAction(ISD::SMIN, MVT::v4i32, Legal);
968 setOperationAction(ISD::UMIN, MVT::v8i16, Legal);
969 setOperationAction(ISD::UMIN, MVT::v4i32, Legal);
971 // FIXME: Do we need to handle scalar-to-vector here?
972 setOperationAction(ISD::MUL, MVT::v4i32, Legal);
974 // We directly match byte blends in the backend as they match the VSELECT
976 setOperationAction(ISD::VSELECT, MVT::v16i8, Legal);
978 // SSE41 brings specific instructions for doing vector sign extend even in
979 // cases where we don't have SRA.
980 for (MVT VT : MVT::integer_vector_valuetypes()) {
981 setLoadExtAction(ISD::SEXTLOAD, VT, MVT::v2i8, Custom);
982 setLoadExtAction(ISD::SEXTLOAD, VT, MVT::v2i16, Custom);
983 setLoadExtAction(ISD::SEXTLOAD, VT, MVT::v2i32, Custom);
986 // SSE41 also has vector sign/zero extending loads, PMOV[SZ]X
987 setLoadExtAction(ISD::SEXTLOAD, MVT::v8i16, MVT::v8i8, Legal);
988 setLoadExtAction(ISD::SEXTLOAD, MVT::v4i32, MVT::v4i8, Legal);
989 setLoadExtAction(ISD::SEXTLOAD, MVT::v2i64, MVT::v2i8, Legal);
990 setLoadExtAction(ISD::SEXTLOAD, MVT::v4i32, MVT::v4i16, Legal);
991 setLoadExtAction(ISD::SEXTLOAD, MVT::v2i64, MVT::v2i16, Legal);
992 setLoadExtAction(ISD::SEXTLOAD, MVT::v2i64, MVT::v2i32, Legal);
994 setLoadExtAction(ISD::ZEXTLOAD, MVT::v8i16, MVT::v8i8, Legal);
995 setLoadExtAction(ISD::ZEXTLOAD, MVT::v4i32, MVT::v4i8, Legal);
996 setLoadExtAction(ISD::ZEXTLOAD, MVT::v2i64, MVT::v2i8, Legal);
997 setLoadExtAction(ISD::ZEXTLOAD, MVT::v4i32, MVT::v4i16, Legal);
998 setLoadExtAction(ISD::ZEXTLOAD, MVT::v2i64, MVT::v2i16, Legal);
999 setLoadExtAction(ISD::ZEXTLOAD, MVT::v2i64, MVT::v2i32, Legal);
1001 // i8 and i16 vectors are custom because the source register and source
1002 // source memory operand types are not the same width. f32 vectors are
1003 // custom since the immediate controlling the insert encodes additional
1005 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v16i8, Custom);
1006 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i16, Custom);
1007 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i32, Custom);
1008 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f32, Custom);
1010 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v16i8, Custom);
1011 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v8i16, Custom);
1012 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4i32, Custom);
1013 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f32, Custom);
1015 // FIXME: these should be Legal, but that's only for the case where
1016 // the index is constant. For now custom expand to deal with that.
1017 if (Subtarget->is64Bit()) {
1018 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2i64, Custom);
1019 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2i64, Custom);
1023 if (Subtarget->hasSSE2()) {
1024 setOperationAction(ISD::SIGN_EXTEND_VECTOR_INREG, MVT::v2i64, Custom);
1025 setOperationAction(ISD::SIGN_EXTEND_VECTOR_INREG, MVT::v4i32, Custom);
1026 setOperationAction(ISD::SIGN_EXTEND_VECTOR_INREG, MVT::v8i16, Custom);
1028 setOperationAction(ISD::SRL, MVT::v8i16, Custom);
1029 setOperationAction(ISD::SRL, MVT::v16i8, Custom);
1031 setOperationAction(ISD::SHL, MVT::v8i16, Custom);
1032 setOperationAction(ISD::SHL, MVT::v16i8, Custom);
1034 setOperationAction(ISD::SRA, MVT::v8i16, Custom);
1035 setOperationAction(ISD::SRA, MVT::v16i8, Custom);
1037 // In the customized shift lowering, the legal cases in AVX2 will be
1039 setOperationAction(ISD::SRL, MVT::v2i64, Custom);
1040 setOperationAction(ISD::SRL, MVT::v4i32, Custom);
1042 setOperationAction(ISD::SHL, MVT::v2i64, Custom);
1043 setOperationAction(ISD::SHL, MVT::v4i32, Custom);
1045 setOperationAction(ISD::SRA, MVT::v2i64, Custom);
1046 setOperationAction(ISD::SRA, MVT::v4i32, Custom);
1049 if (!Subtarget->useSoftFloat() && Subtarget->hasFp256()) {
1050 addRegisterClass(MVT::v32i8, &X86::VR256RegClass);
1051 addRegisterClass(MVT::v16i16, &X86::VR256RegClass);
1052 addRegisterClass(MVT::v8i32, &X86::VR256RegClass);
1053 addRegisterClass(MVT::v8f32, &X86::VR256RegClass);
1054 addRegisterClass(MVT::v4i64, &X86::VR256RegClass);
1055 addRegisterClass(MVT::v4f64, &X86::VR256RegClass);
1057 setOperationAction(ISD::LOAD, MVT::v8f32, Legal);
1058 setOperationAction(ISD::LOAD, MVT::v4f64, Legal);
1059 setOperationAction(ISD::LOAD, MVT::v4i64, Legal);
1061 setOperationAction(ISD::FADD, MVT::v8f32, Legal);
1062 setOperationAction(ISD::FSUB, MVT::v8f32, Legal);
1063 setOperationAction(ISD::FMUL, MVT::v8f32, Legal);
1064 setOperationAction(ISD::FDIV, MVT::v8f32, Legal);
1065 setOperationAction(ISD::FSQRT, MVT::v8f32, Legal);
1066 setOperationAction(ISD::FFLOOR, MVT::v8f32, Legal);
1067 setOperationAction(ISD::FCEIL, MVT::v8f32, Legal);
1068 setOperationAction(ISD::FTRUNC, MVT::v8f32, Legal);
1069 setOperationAction(ISD::FRINT, MVT::v8f32, Legal);
1070 setOperationAction(ISD::FNEARBYINT, MVT::v8f32, Legal);
1071 setOperationAction(ISD::FNEG, MVT::v8f32, Custom);
1072 setOperationAction(ISD::FABS, MVT::v8f32, Custom);
1074 setOperationAction(ISD::FADD, MVT::v4f64, Legal);
1075 setOperationAction(ISD::FSUB, MVT::v4f64, Legal);
1076 setOperationAction(ISD::FMUL, MVT::v4f64, Legal);
1077 setOperationAction(ISD::FDIV, MVT::v4f64, Legal);
1078 setOperationAction(ISD::FSQRT, MVT::v4f64, Legal);
1079 setOperationAction(ISD::FFLOOR, MVT::v4f64, Legal);
1080 setOperationAction(ISD::FCEIL, MVT::v4f64, Legal);
1081 setOperationAction(ISD::FTRUNC, MVT::v4f64, Legal);
1082 setOperationAction(ISD::FRINT, MVT::v4f64, Legal);
1083 setOperationAction(ISD::FNEARBYINT, MVT::v4f64, Legal);
1084 setOperationAction(ISD::FNEG, MVT::v4f64, Custom);
1085 setOperationAction(ISD::FABS, MVT::v4f64, Custom);
1087 // (fp_to_int:v8i16 (v8f32 ..)) requires the result type to be promoted
1088 // even though v8i16 is a legal type.
1089 setOperationAction(ISD::FP_TO_SINT, MVT::v8i16, Promote);
1090 setOperationAction(ISD::FP_TO_UINT, MVT::v8i16, Promote);
1091 setOperationAction(ISD::FP_TO_SINT, MVT::v8i32, Legal);
1093 setOperationAction(ISD::SINT_TO_FP, MVT::v8i16, Promote);
1094 setOperationAction(ISD::SINT_TO_FP, MVT::v8i32, Legal);
1095 setOperationAction(ISD::FP_ROUND, MVT::v4f32, Legal);
1097 setOperationAction(ISD::UINT_TO_FP, MVT::v8i8, Custom);
1098 setOperationAction(ISD::UINT_TO_FP, MVT::v8i16, Custom);
1100 for (MVT VT : MVT::fp_vector_valuetypes())
1101 setLoadExtAction(ISD::EXTLOAD, VT, MVT::v4f32, Legal);
1103 setOperationAction(ISD::SRL, MVT::v16i16, Custom);
1104 setOperationAction(ISD::SRL, MVT::v32i8, Custom);
1106 setOperationAction(ISD::SHL, MVT::v16i16, Custom);
1107 setOperationAction(ISD::SHL, MVT::v32i8, Custom);
1109 setOperationAction(ISD::SRA, MVT::v16i16, Custom);
1110 setOperationAction(ISD::SRA, MVT::v32i8, Custom);
1112 setOperationAction(ISD::SETCC, MVT::v32i8, Custom);
1113 setOperationAction(ISD::SETCC, MVT::v16i16, Custom);
1114 setOperationAction(ISD::SETCC, MVT::v8i32, Custom);
1115 setOperationAction(ISD::SETCC, MVT::v4i64, Custom);
1117 setOperationAction(ISD::SELECT, MVT::v4f64, Custom);
1118 setOperationAction(ISD::SELECT, MVT::v4i64, Custom);
1119 setOperationAction(ISD::SELECT, MVT::v8f32, Custom);
1121 setOperationAction(ISD::SIGN_EXTEND, MVT::v4i64, Custom);
1122 setOperationAction(ISD::SIGN_EXTEND, MVT::v8i32, Custom);
1123 setOperationAction(ISD::SIGN_EXTEND, MVT::v16i16, Custom);
1124 setOperationAction(ISD::ZERO_EXTEND, MVT::v4i64, Custom);
1125 setOperationAction(ISD::ZERO_EXTEND, MVT::v8i32, Custom);
1126 setOperationAction(ISD::ZERO_EXTEND, MVT::v16i16, Custom);
1127 setOperationAction(ISD::ANY_EXTEND, MVT::v4i64, Custom);
1128 setOperationAction(ISD::ANY_EXTEND, MVT::v8i32, Custom);
1129 setOperationAction(ISD::ANY_EXTEND, MVT::v16i16, Custom);
1130 setOperationAction(ISD::TRUNCATE, MVT::v16i8, Custom);
1131 setOperationAction(ISD::TRUNCATE, MVT::v8i16, Custom);
1132 setOperationAction(ISD::TRUNCATE, MVT::v4i32, Custom);
1134 setOperationAction(ISD::CTPOP, MVT::v32i8, Custom);
1135 setOperationAction(ISD::CTPOP, MVT::v16i16, Custom);
1136 setOperationAction(ISD::CTPOP, MVT::v8i32, Custom);
1137 setOperationAction(ISD::CTPOP, MVT::v4i64, Custom);
1139 setOperationAction(ISD::CTTZ, MVT::v32i8, Custom);
1140 setOperationAction(ISD::CTTZ, MVT::v16i16, Custom);
1141 setOperationAction(ISD::CTTZ, MVT::v8i32, Custom);
1142 setOperationAction(ISD::CTTZ, MVT::v4i64, Custom);
1143 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::v32i8, Custom);
1144 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::v16i16, Custom);
1145 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::v8i32, Custom);
1146 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::v4i64, Custom);
1148 if (Subtarget->hasFMA() || Subtarget->hasFMA4() || Subtarget->hasAVX512()) {
1149 setOperationAction(ISD::FMA, MVT::v8f32, Legal);
1150 setOperationAction(ISD::FMA, MVT::v4f64, Legal);
1151 setOperationAction(ISD::FMA, MVT::v4f32, Legal);
1152 setOperationAction(ISD::FMA, MVT::v2f64, Legal);
1153 setOperationAction(ISD::FMA, MVT::f32, Legal);
1154 setOperationAction(ISD::FMA, MVT::f64, Legal);
1157 if (Subtarget->hasInt256()) {
1158 setOperationAction(ISD::ADD, MVT::v4i64, Legal);
1159 setOperationAction(ISD::ADD, MVT::v8i32, Legal);
1160 setOperationAction(ISD::ADD, MVT::v16i16, Legal);
1161 setOperationAction(ISD::ADD, MVT::v32i8, Legal);
1163 setOperationAction(ISD::SUB, MVT::v4i64, Legal);
1164 setOperationAction(ISD::SUB, MVT::v8i32, Legal);
1165 setOperationAction(ISD::SUB, MVT::v16i16, Legal);
1166 setOperationAction(ISD::SUB, MVT::v32i8, Legal);
1168 setOperationAction(ISD::MUL, MVT::v4i64, Custom);
1169 setOperationAction(ISD::MUL, MVT::v8i32, Legal);
1170 setOperationAction(ISD::MUL, MVT::v16i16, Legal);
1171 setOperationAction(ISD::MUL, MVT::v32i8, Custom);
1173 setOperationAction(ISD::UMUL_LOHI, MVT::v8i32, Custom);
1174 setOperationAction(ISD::SMUL_LOHI, MVT::v8i32, Custom);
1175 setOperationAction(ISD::MULHU, MVT::v16i16, Legal);
1176 setOperationAction(ISD::MULHS, MVT::v16i16, Legal);
1178 setOperationAction(ISD::SMAX, MVT::v32i8, Legal);
1179 setOperationAction(ISD::SMAX, MVT::v16i16, Legal);
1180 setOperationAction(ISD::SMAX, MVT::v8i32, Legal);
1181 setOperationAction(ISD::UMAX, MVT::v32i8, Legal);
1182 setOperationAction(ISD::UMAX, MVT::v16i16, Legal);
1183 setOperationAction(ISD::UMAX, MVT::v8i32, Legal);
1184 setOperationAction(ISD::SMIN, MVT::v32i8, Legal);
1185 setOperationAction(ISD::SMIN, MVT::v16i16, Legal);
1186 setOperationAction(ISD::SMIN, MVT::v8i32, Legal);
1187 setOperationAction(ISD::UMIN, MVT::v32i8, Legal);
1188 setOperationAction(ISD::UMIN, MVT::v16i16, Legal);
1189 setOperationAction(ISD::UMIN, MVT::v8i32, Legal);
1191 // The custom lowering for UINT_TO_FP for v8i32 becomes interesting
1192 // when we have a 256bit-wide blend with immediate.
1193 setOperationAction(ISD::UINT_TO_FP, MVT::v8i32, Custom);
1195 // AVX2 also has wider vector sign/zero extending loads, VPMOV[SZ]X
1196 setLoadExtAction(ISD::SEXTLOAD, MVT::v16i16, MVT::v16i8, Legal);
1197 setLoadExtAction(ISD::SEXTLOAD, MVT::v8i32, MVT::v8i8, Legal);
1198 setLoadExtAction(ISD::SEXTLOAD, MVT::v4i64, MVT::v4i8, Legal);
1199 setLoadExtAction(ISD::SEXTLOAD, MVT::v8i32, MVT::v8i16, Legal);
1200 setLoadExtAction(ISD::SEXTLOAD, MVT::v4i64, MVT::v4i16, Legal);
1201 setLoadExtAction(ISD::SEXTLOAD, MVT::v4i64, MVT::v4i32, Legal);
1203 setLoadExtAction(ISD::ZEXTLOAD, MVT::v16i16, MVT::v16i8, Legal);
1204 setLoadExtAction(ISD::ZEXTLOAD, MVT::v8i32, MVT::v8i8, Legal);
1205 setLoadExtAction(ISD::ZEXTLOAD, MVT::v4i64, MVT::v4i8, Legal);
1206 setLoadExtAction(ISD::ZEXTLOAD, MVT::v8i32, MVT::v8i16, Legal);
1207 setLoadExtAction(ISD::ZEXTLOAD, MVT::v4i64, MVT::v4i16, Legal);
1208 setLoadExtAction(ISD::ZEXTLOAD, MVT::v4i64, MVT::v4i32, Legal);
1210 setOperationAction(ISD::ADD, MVT::v4i64, Custom);
1211 setOperationAction(ISD::ADD, MVT::v8i32, Custom);
1212 setOperationAction(ISD::ADD, MVT::v16i16, Custom);
1213 setOperationAction(ISD::ADD, MVT::v32i8, Custom);
1215 setOperationAction(ISD::SUB, MVT::v4i64, Custom);
1216 setOperationAction(ISD::SUB, MVT::v8i32, Custom);
1217 setOperationAction(ISD::SUB, MVT::v16i16, Custom);
1218 setOperationAction(ISD::SUB, MVT::v32i8, Custom);
1220 setOperationAction(ISD::MUL, MVT::v4i64, Custom);
1221 setOperationAction(ISD::MUL, MVT::v8i32, Custom);
1222 setOperationAction(ISD::MUL, MVT::v16i16, Custom);
1223 setOperationAction(ISD::MUL, MVT::v32i8, Custom);
1225 setOperationAction(ISD::SMAX, MVT::v32i8, Custom);
1226 setOperationAction(ISD::SMAX, MVT::v16i16, Custom);
1227 setOperationAction(ISD::SMAX, MVT::v8i32, Custom);
1228 setOperationAction(ISD::UMAX, MVT::v32i8, Custom);
1229 setOperationAction(ISD::UMAX, MVT::v16i16, Custom);
1230 setOperationAction(ISD::UMAX, MVT::v8i32, Custom);
1231 setOperationAction(ISD::SMIN, MVT::v32i8, Custom);
1232 setOperationAction(ISD::SMIN, MVT::v16i16, Custom);
1233 setOperationAction(ISD::SMIN, MVT::v8i32, Custom);
1234 setOperationAction(ISD::UMIN, MVT::v32i8, Custom);
1235 setOperationAction(ISD::UMIN, MVT::v16i16, Custom);
1236 setOperationAction(ISD::UMIN, MVT::v8i32, Custom);
1239 // In the customized shift lowering, the legal cases in AVX2 will be
1241 setOperationAction(ISD::SRL, MVT::v4i64, Custom);
1242 setOperationAction(ISD::SRL, MVT::v8i32, Custom);
1244 setOperationAction(ISD::SHL, MVT::v4i64, Custom);
1245 setOperationAction(ISD::SHL, MVT::v8i32, Custom);
1247 setOperationAction(ISD::SRA, MVT::v4i64, Custom);
1248 setOperationAction(ISD::SRA, MVT::v8i32, Custom);
1250 // Custom lower several nodes for 256-bit types.
1251 for (MVT VT : MVT::vector_valuetypes()) {
1252 if (VT.getScalarSizeInBits() >= 32) {
1253 setOperationAction(ISD::MLOAD, VT, Legal);
1254 setOperationAction(ISD::MSTORE, VT, Legal);
1256 // Extract subvector is special because the value type
1257 // (result) is 128-bit but the source is 256-bit wide.
1258 if (VT.is128BitVector()) {
1259 setOperationAction(ISD::EXTRACT_SUBVECTOR, VT, Custom);
1261 // Do not attempt to custom lower other non-256-bit vectors
1262 if (!VT.is256BitVector())
1265 setOperationAction(ISD::BUILD_VECTOR, VT, Custom);
1266 setOperationAction(ISD::VECTOR_SHUFFLE, VT, Custom);
1267 setOperationAction(ISD::VSELECT, VT, Custom);
1268 setOperationAction(ISD::INSERT_VECTOR_ELT, VT, Custom);
1269 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Custom);
1270 setOperationAction(ISD::SCALAR_TO_VECTOR, VT, Custom);
1271 setOperationAction(ISD::INSERT_SUBVECTOR, VT, Custom);
1272 setOperationAction(ISD::CONCAT_VECTORS, VT, Custom);
1275 if (Subtarget->hasInt256())
1276 setOperationAction(ISD::VSELECT, MVT::v32i8, Legal);
1279 // Promote v32i8, v16i16, v8i32 select, and, or, xor to v4i64.
1280 for (int i = MVT::v32i8; i != MVT::v4i64; ++i) {
1281 MVT VT = (MVT::SimpleValueType)i;
1283 // Do not attempt to promote non-256-bit vectors
1284 if (!VT.is256BitVector())
1287 setOperationAction(ISD::AND, VT, Promote);
1288 AddPromotedToType (ISD::AND, VT, MVT::v4i64);
1289 setOperationAction(ISD::OR, VT, Promote);
1290 AddPromotedToType (ISD::OR, VT, MVT::v4i64);
1291 setOperationAction(ISD::XOR, VT, Promote);
1292 AddPromotedToType (ISD::XOR, VT, MVT::v4i64);
1293 setOperationAction(ISD::LOAD, VT, Promote);
1294 AddPromotedToType (ISD::LOAD, VT, MVT::v4i64);
1295 setOperationAction(ISD::SELECT, VT, Promote);
1296 AddPromotedToType (ISD::SELECT, VT, MVT::v4i64);
1300 if (!Subtarget->useSoftFloat() && Subtarget->hasAVX512()) {
1301 addRegisterClass(MVT::v16i32, &X86::VR512RegClass);
1302 addRegisterClass(MVT::v16f32, &X86::VR512RegClass);
1303 addRegisterClass(MVT::v8i64, &X86::VR512RegClass);
1304 addRegisterClass(MVT::v8f64, &X86::VR512RegClass);
1306 addRegisterClass(MVT::i1, &X86::VK1RegClass);
1307 addRegisterClass(MVT::v8i1, &X86::VK8RegClass);
1308 addRegisterClass(MVT::v16i1, &X86::VK16RegClass);
1310 for (MVT VT : MVT::fp_vector_valuetypes())
1311 setLoadExtAction(ISD::EXTLOAD, VT, MVT::v8f32, Legal);
1313 setLoadExtAction(ISD::ZEXTLOAD, MVT::v16i32, MVT::v16i8, Legal);
1314 setLoadExtAction(ISD::SEXTLOAD, MVT::v16i32, MVT::v16i8, Legal);
1315 setLoadExtAction(ISD::ZEXTLOAD, MVT::v16i32, MVT::v16i16, Legal);
1316 setLoadExtAction(ISD::SEXTLOAD, MVT::v16i32, MVT::v16i16, Legal);
1317 setLoadExtAction(ISD::ZEXTLOAD, MVT::v32i16, MVT::v32i8, Legal);
1318 setLoadExtAction(ISD::SEXTLOAD, MVT::v32i16, MVT::v32i8, Legal);
1319 setLoadExtAction(ISD::ZEXTLOAD, MVT::v8i64, MVT::v8i8, Legal);
1320 setLoadExtAction(ISD::SEXTLOAD, MVT::v8i64, MVT::v8i8, Legal);
1321 setLoadExtAction(ISD::ZEXTLOAD, MVT::v8i64, MVT::v8i16, Legal);
1322 setLoadExtAction(ISD::SEXTLOAD, MVT::v8i64, MVT::v8i16, Legal);
1323 setLoadExtAction(ISD::ZEXTLOAD, MVT::v8i64, MVT::v8i32, Legal);
1324 setLoadExtAction(ISD::SEXTLOAD, MVT::v8i64, MVT::v8i32, Legal);
1326 setOperationAction(ISD::BR_CC, MVT::i1, Expand);
1327 setOperationAction(ISD::SETCC, MVT::i1, Custom);
1328 setOperationAction(ISD::XOR, MVT::i1, Legal);
1329 setOperationAction(ISD::OR, MVT::i1, Legal);
1330 setOperationAction(ISD::AND, MVT::i1, Legal);
1331 setOperationAction(ISD::SUB, MVT::i1, Custom);
1332 setOperationAction(ISD::ADD, MVT::i1, Custom);
1333 setOperationAction(ISD::MUL, MVT::i1, Custom);
1334 setOperationAction(ISD::LOAD, MVT::v16f32, Legal);
1335 setOperationAction(ISD::LOAD, MVT::v8f64, Legal);
1336 setOperationAction(ISD::LOAD, MVT::v8i64, Legal);
1337 setOperationAction(ISD::LOAD, MVT::v16i32, Legal);
1338 setOperationAction(ISD::LOAD, MVT::v16i1, Legal);
1340 setOperationAction(ISD::FADD, MVT::v16f32, Legal);
1341 setOperationAction(ISD::FSUB, MVT::v16f32, Legal);
1342 setOperationAction(ISD::FMUL, MVT::v16f32, Legal);
1343 setOperationAction(ISD::FDIV, MVT::v16f32, Legal);
1344 setOperationAction(ISD::FSQRT, MVT::v16f32, Legal);
1345 setOperationAction(ISD::FNEG, MVT::v16f32, Custom);
1347 setOperationAction(ISD::FADD, MVT::v8f64, Legal);
1348 setOperationAction(ISD::FSUB, MVT::v8f64, Legal);
1349 setOperationAction(ISD::FMUL, MVT::v8f64, Legal);
1350 setOperationAction(ISD::FDIV, MVT::v8f64, Legal);
1351 setOperationAction(ISD::FSQRT, MVT::v8f64, Legal);
1352 setOperationAction(ISD::FNEG, MVT::v8f64, Custom);
1353 setOperationAction(ISD::FMA, MVT::v8f64, Legal);
1354 setOperationAction(ISD::FMA, MVT::v16f32, Legal);
1356 // FIXME: [US]INT_TO_FP are not legal for f80.
1357 setOperationAction(ISD::SINT_TO_FP, MVT::i32, Legal);
1358 setOperationAction(ISD::UINT_TO_FP, MVT::i32, Legal);
1359 if (Subtarget->is64Bit()) {
1360 setOperationAction(ISD::SINT_TO_FP, MVT::i64, Legal);
1361 setOperationAction(ISD::UINT_TO_FP, MVT::i64, Legal);
1363 setOperationAction(ISD::FP_TO_SINT, MVT::v16i32, Legal);
1364 setOperationAction(ISD::FP_TO_UINT, MVT::v16i32, Legal);
1365 setOperationAction(ISD::FP_TO_UINT, MVT::v8i32, Legal);
1366 setOperationAction(ISD::FP_TO_UINT, MVT::v4i32, Legal);
1367 setOperationAction(ISD::SINT_TO_FP, MVT::v16i32, Legal);
1368 setOperationAction(ISD::SINT_TO_FP, MVT::v8i1, Custom);
1369 setOperationAction(ISD::SINT_TO_FP, MVT::v16i1, Custom);
1370 setOperationAction(ISD::SINT_TO_FP, MVT::v16i8, Promote);
1371 setOperationAction(ISD::SINT_TO_FP, MVT::v16i16, Promote);
1372 setOperationAction(ISD::UINT_TO_FP, MVT::v16i32, Legal);
1373 setOperationAction(ISD::UINT_TO_FP, MVT::v8i32, Legal);
1374 setOperationAction(ISD::UINT_TO_FP, MVT::v4i32, Legal);
1375 setOperationAction(ISD::UINT_TO_FP, MVT::v16i8, Custom);
1376 setOperationAction(ISD::UINT_TO_FP, MVT::v16i16, Custom);
1377 setOperationAction(ISD::FP_ROUND, MVT::v8f32, Legal);
1378 setOperationAction(ISD::FP_EXTEND, MVT::v8f32, Legal);
1380 setTruncStoreAction(MVT::v8i64, MVT::v8i8, Legal);
1381 setTruncStoreAction(MVT::v8i64, MVT::v8i16, Legal);
1382 setTruncStoreAction(MVT::v8i64, MVT::v8i32, Legal);
1383 setTruncStoreAction(MVT::v16i32, MVT::v16i8, Legal);
1384 setTruncStoreAction(MVT::v16i32, MVT::v16i16, Legal);
1385 if (Subtarget->hasVLX()){
1386 setTruncStoreAction(MVT::v4i64, MVT::v4i8, Legal);
1387 setTruncStoreAction(MVT::v4i64, MVT::v4i16, Legal);
1388 setTruncStoreAction(MVT::v4i64, MVT::v4i32, Legal);
1389 setTruncStoreAction(MVT::v8i32, MVT::v8i8, Legal);
1390 setTruncStoreAction(MVT::v8i32, MVT::v8i16, Legal);
1392 setTruncStoreAction(MVT::v2i64, MVT::v2i8, Legal);
1393 setTruncStoreAction(MVT::v2i64, MVT::v2i16, Legal);
1394 setTruncStoreAction(MVT::v2i64, MVT::v2i32, Legal);
1395 setTruncStoreAction(MVT::v4i32, MVT::v4i8, Legal);
1396 setTruncStoreAction(MVT::v4i32, MVT::v4i16, Legal);
1398 setOperationAction(ISD::TRUNCATE, MVT::i1, Custom);
1399 setOperationAction(ISD::TRUNCATE, MVT::v16i8, Custom);
1400 setOperationAction(ISD::TRUNCATE, MVT::v8i32, Custom);
1401 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v8i1, Custom);
1402 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v16i1, Custom);
1403 if (Subtarget->hasDQI()) {
1404 setOperationAction(ISD::TRUNCATE, MVT::v2i1, Custom);
1405 setOperationAction(ISD::TRUNCATE, MVT::v4i1, Custom);
1407 setOperationAction(ISD::SINT_TO_FP, MVT::v8i64, Legal);
1408 setOperationAction(ISD::UINT_TO_FP, MVT::v8i64, Legal);
1409 setOperationAction(ISD::FP_TO_SINT, MVT::v8i64, Legal);
1410 setOperationAction(ISD::FP_TO_UINT, MVT::v8i64, Legal);
1411 if (Subtarget->hasVLX()) {
1412 setOperationAction(ISD::SINT_TO_FP, MVT::v4i64, Legal);
1413 setOperationAction(ISD::SINT_TO_FP, MVT::v2i64, Legal);
1414 setOperationAction(ISD::UINT_TO_FP, MVT::v4i64, Legal);
1415 setOperationAction(ISD::UINT_TO_FP, MVT::v2i64, Legal);
1416 setOperationAction(ISD::FP_TO_SINT, MVT::v4i64, Legal);
1417 setOperationAction(ISD::FP_TO_SINT, MVT::v2i64, Legal);
1418 setOperationAction(ISD::FP_TO_UINT, MVT::v4i64, Legal);
1419 setOperationAction(ISD::FP_TO_UINT, MVT::v2i64, Legal);
1422 if (Subtarget->hasVLX()) {
1423 setOperationAction(ISD::SINT_TO_FP, MVT::v8i32, Legal);
1424 setOperationAction(ISD::UINT_TO_FP, MVT::v8i32, Legal);
1425 setOperationAction(ISD::FP_TO_SINT, MVT::v8i32, Legal);
1426 setOperationAction(ISD::FP_TO_UINT, MVT::v8i32, Legal);
1427 setOperationAction(ISD::SINT_TO_FP, MVT::v4i32, Legal);
1428 setOperationAction(ISD::UINT_TO_FP, MVT::v4i32, Legal);
1429 setOperationAction(ISD::FP_TO_SINT, MVT::v4i32, Legal);
1430 setOperationAction(ISD::FP_TO_UINT, MVT::v4i32, Legal);
1432 setOperationAction(ISD::TRUNCATE, MVT::v8i1, Custom);
1433 setOperationAction(ISD::TRUNCATE, MVT::v16i1, Custom);
1434 setOperationAction(ISD::TRUNCATE, MVT::v16i16, Custom);
1435 setOperationAction(ISD::ZERO_EXTEND, MVT::v16i32, Custom);
1436 setOperationAction(ISD::ZERO_EXTEND, MVT::v8i64, Custom);
1437 setOperationAction(ISD::ANY_EXTEND, MVT::v16i32, Custom);
1438 setOperationAction(ISD::ANY_EXTEND, MVT::v8i64, Custom);
1439 setOperationAction(ISD::SIGN_EXTEND, MVT::v16i32, Custom);
1440 setOperationAction(ISD::SIGN_EXTEND, MVT::v8i64, Custom);
1441 setOperationAction(ISD::SIGN_EXTEND, MVT::v16i8, Custom);
1442 setOperationAction(ISD::SIGN_EXTEND, MVT::v8i16, Custom);
1443 setOperationAction(ISD::SIGN_EXTEND, MVT::v16i16, Custom);
1444 if (Subtarget->hasDQI()) {
1445 setOperationAction(ISD::SIGN_EXTEND, MVT::v4i32, Custom);
1446 setOperationAction(ISD::SIGN_EXTEND, MVT::v2i64, Custom);
1448 setOperationAction(ISD::FFLOOR, MVT::v16f32, Legal);
1449 setOperationAction(ISD::FFLOOR, MVT::v8f64, Legal);
1450 setOperationAction(ISD::FCEIL, MVT::v16f32, Legal);
1451 setOperationAction(ISD::FCEIL, MVT::v8f64, Legal);
1452 setOperationAction(ISD::FTRUNC, MVT::v16f32, Legal);
1453 setOperationAction(ISD::FTRUNC, MVT::v8f64, Legal);
1454 setOperationAction(ISD::FRINT, MVT::v16f32, Legal);
1455 setOperationAction(ISD::FRINT, MVT::v8f64, Legal);
1456 setOperationAction(ISD::FNEARBYINT, MVT::v16f32, Legal);
1457 setOperationAction(ISD::FNEARBYINT, MVT::v8f64, Legal);
1459 setOperationAction(ISD::CONCAT_VECTORS, MVT::v8f64, Custom);
1460 setOperationAction(ISD::CONCAT_VECTORS, MVT::v8i64, Custom);
1461 setOperationAction(ISD::CONCAT_VECTORS, MVT::v16f32, Custom);
1462 setOperationAction(ISD::CONCAT_VECTORS, MVT::v16i32, Custom);
1463 setOperationAction(ISD::CONCAT_VECTORS, MVT::v16i1, Legal);
1465 setOperationAction(ISD::SETCC, MVT::v16i1, Custom);
1466 setOperationAction(ISD::SETCC, MVT::v8i1, Custom);
1468 setOperationAction(ISD::MUL, MVT::v8i64, Custom);
1470 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v8i1, Custom);
1471 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v16i1, Custom);
1472 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v16i1, Custom);
1473 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i1, Custom);
1474 setOperationAction(ISD::BUILD_VECTOR, MVT::v8i1, Custom);
1475 setOperationAction(ISD::BUILD_VECTOR, MVT::v16i1, Custom);
1476 setOperationAction(ISD::SELECT, MVT::v8f64, Custom);
1477 setOperationAction(ISD::SELECT, MVT::v8i64, Custom);
1478 setOperationAction(ISD::SELECT, MVT::v16f32, Custom);
1479 setOperationAction(ISD::SELECT, MVT::v16i1, Custom);
1480 setOperationAction(ISD::SELECT, MVT::v8i1, Custom);
1482 setOperationAction(ISD::SMAX, MVT::v16i32, Legal);
1483 setOperationAction(ISD::SMAX, MVT::v8i64, Legal);
1484 setOperationAction(ISD::UMAX, MVT::v16i32, Legal);
1485 setOperationAction(ISD::UMAX, MVT::v8i64, Legal);
1486 setOperationAction(ISD::SMIN, MVT::v16i32, Legal);
1487 setOperationAction(ISD::SMIN, MVT::v8i64, Legal);
1488 setOperationAction(ISD::UMIN, MVT::v16i32, Legal);
1489 setOperationAction(ISD::UMIN, MVT::v8i64, Legal);
1491 setOperationAction(ISD::ADD, MVT::v8i64, Legal);
1492 setOperationAction(ISD::ADD, MVT::v16i32, Legal);
1494 setOperationAction(ISD::SUB, MVT::v8i64, Legal);
1495 setOperationAction(ISD::SUB, MVT::v16i32, Legal);
1497 setOperationAction(ISD::MUL, MVT::v16i32, Legal);
1499 setOperationAction(ISD::SRL, MVT::v8i64, Custom);
1500 setOperationAction(ISD::SRL, MVT::v16i32, Custom);
1502 setOperationAction(ISD::SHL, MVT::v8i64, Custom);
1503 setOperationAction(ISD::SHL, MVT::v16i32, Custom);
1505 setOperationAction(ISD::SRA, MVT::v8i64, Custom);
1506 setOperationAction(ISD::SRA, MVT::v16i32, Custom);
1508 setOperationAction(ISD::AND, MVT::v8i64, Legal);
1509 setOperationAction(ISD::OR, MVT::v8i64, Legal);
1510 setOperationAction(ISD::XOR, MVT::v8i64, Legal);
1511 setOperationAction(ISD::AND, MVT::v16i32, Legal);
1512 setOperationAction(ISD::OR, MVT::v16i32, Legal);
1513 setOperationAction(ISD::XOR, MVT::v16i32, Legal);
1515 if (Subtarget->hasCDI()) {
1516 setOperationAction(ISD::CTLZ, MVT::v8i64, Legal);
1517 setOperationAction(ISD::CTLZ, MVT::v16i32, Legal);
1518 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::v8i64, Legal);
1519 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::v16i32, Legal);
1521 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::v8i64, Custom);
1522 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::v16i32, Custom);
1524 if (Subtarget->hasVLX() && Subtarget->hasCDI()) {
1525 setOperationAction(ISD::CTLZ, MVT::v4i64, Legal);
1526 setOperationAction(ISD::CTLZ, MVT::v8i32, Legal);
1527 setOperationAction(ISD::CTLZ, MVT::v2i64, Legal);
1528 setOperationAction(ISD::CTLZ, MVT::v4i32, Legal);
1529 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::v4i64, Legal);
1530 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::v8i32, Legal);
1531 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::v2i64, Legal);
1532 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::v4i32, Legal);
1534 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::v4i64, Custom);
1535 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::v8i32, Custom);
1536 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::v2i64, Custom);
1537 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::v4i32, Custom);
1539 if (Subtarget->hasDQI()) {
1540 setOperationAction(ISD::MUL, MVT::v2i64, Legal);
1541 setOperationAction(ISD::MUL, MVT::v4i64, Legal);
1542 setOperationAction(ISD::MUL, MVT::v8i64, Legal);
1544 // Custom lower several nodes.
1545 for (MVT VT : MVT::vector_valuetypes()) {
1546 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
1548 setOperationAction(ISD::AND, VT, Legal);
1549 setOperationAction(ISD::OR, VT, Legal);
1550 setOperationAction(ISD::XOR, VT, Legal);
1552 if (EltSize >= 32 && VT.getSizeInBits() <= 512) {
1553 setOperationAction(ISD::MGATHER, VT, Custom);
1554 setOperationAction(ISD::MSCATTER, VT, Custom);
1556 // Extract subvector is special because the value type
1557 // (result) is 256/128-bit but the source is 512-bit wide.
1558 if (VT.is128BitVector() || VT.is256BitVector()) {
1559 setOperationAction(ISD::EXTRACT_SUBVECTOR, VT, Custom);
1561 if (VT.getVectorElementType() == MVT::i1)
1562 setOperationAction(ISD::EXTRACT_SUBVECTOR, VT, Legal);
1564 // Do not attempt to custom lower other non-512-bit vectors
1565 if (!VT.is512BitVector())
1568 if (EltSize >= 32) {
1569 setOperationAction(ISD::VECTOR_SHUFFLE, VT, Custom);
1570 setOperationAction(ISD::INSERT_VECTOR_ELT, VT, Custom);
1571 setOperationAction(ISD::BUILD_VECTOR, VT, Custom);
1572 setOperationAction(ISD::VSELECT, VT, Legal);
1573 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Custom);
1574 setOperationAction(ISD::SCALAR_TO_VECTOR, VT, Custom);
1575 setOperationAction(ISD::INSERT_SUBVECTOR, VT, Custom);
1576 setOperationAction(ISD::MLOAD, VT, Legal);
1577 setOperationAction(ISD::MSTORE, VT, Legal);
1580 for (int i = MVT::v32i8; i != MVT::v8i64; ++i) {
1581 MVT VT = (MVT::SimpleValueType)i;
1583 // Do not attempt to promote non-512-bit vectors.
1584 if (!VT.is512BitVector())
1587 setOperationAction(ISD::SELECT, VT, Promote);
1588 AddPromotedToType (ISD::SELECT, VT, MVT::v8i64);
1592 if (!Subtarget->useSoftFloat() && Subtarget->hasBWI()) {
1593 addRegisterClass(MVT::v32i16, &X86::VR512RegClass);
1594 addRegisterClass(MVT::v64i8, &X86::VR512RegClass);
1596 addRegisterClass(MVT::v32i1, &X86::VK32RegClass);
1597 addRegisterClass(MVT::v64i1, &X86::VK64RegClass);
1599 setOperationAction(ISD::LOAD, MVT::v32i16, Legal);
1600 setOperationAction(ISD::LOAD, MVT::v64i8, Legal);
1601 setOperationAction(ISD::SETCC, MVT::v32i1, Custom);
1602 setOperationAction(ISD::SETCC, MVT::v64i1, Custom);
1603 setOperationAction(ISD::ADD, MVT::v32i16, Legal);
1604 setOperationAction(ISD::ADD, MVT::v64i8, Legal);
1605 setOperationAction(ISD::SUB, MVT::v32i16, Legal);
1606 setOperationAction(ISD::SUB, MVT::v64i8, Legal);
1607 setOperationAction(ISD::MUL, MVT::v32i16, Legal);
1608 setOperationAction(ISD::MULHS, MVT::v32i16, Legal);
1609 setOperationAction(ISD::MULHU, MVT::v32i16, Legal);
1610 setOperationAction(ISD::CONCAT_VECTORS, MVT::v32i1, Legal);
1611 setOperationAction(ISD::CONCAT_VECTORS, MVT::v64i1, Legal);
1612 setOperationAction(ISD::INSERT_SUBVECTOR, MVT::v32i1, Custom);
1613 setOperationAction(ISD::INSERT_SUBVECTOR, MVT::v64i1, Custom);
1614 setOperationAction(ISD::SELECT, MVT::v32i1, Custom);
1615 setOperationAction(ISD::SELECT, MVT::v64i1, Custom);
1616 setOperationAction(ISD::SIGN_EXTEND, MVT::v32i8, Custom);
1617 setOperationAction(ISD::ZERO_EXTEND, MVT::v32i8, Custom);
1618 setOperationAction(ISD::SIGN_EXTEND, MVT::v32i16, Custom);
1619 setOperationAction(ISD::ZERO_EXTEND, MVT::v32i16, Custom);
1620 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v32i16, Custom);
1621 setOperationAction(ISD::SIGN_EXTEND, MVT::v64i8, Custom);
1622 setOperationAction(ISD::ZERO_EXTEND, MVT::v64i8, Custom);
1623 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v32i1, Custom);
1624 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v64i1, Custom);
1625 setOperationAction(ISD::VSELECT, MVT::v32i16, Legal);
1626 setOperationAction(ISD::VSELECT, MVT::v64i8, Legal);
1627 setOperationAction(ISD::TRUNCATE, MVT::v32i1, Custom);
1628 setOperationAction(ISD::TRUNCATE, MVT::v64i1, Custom);
1629 setOperationAction(ISD::TRUNCATE, MVT::v32i8, Custom);
1630 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v32i1, Custom);
1631 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v64i1, Custom);
1633 setOperationAction(ISD::SMAX, MVT::v64i8, Legal);
1634 setOperationAction(ISD::SMAX, MVT::v32i16, Legal);
1635 setOperationAction(ISD::UMAX, MVT::v64i8, Legal);
1636 setOperationAction(ISD::UMAX, MVT::v32i16, Legal);
1637 setOperationAction(ISD::SMIN, MVT::v64i8, Legal);
1638 setOperationAction(ISD::SMIN, MVT::v32i16, Legal);
1639 setOperationAction(ISD::UMIN, MVT::v64i8, Legal);
1640 setOperationAction(ISD::UMIN, MVT::v32i16, Legal);
1642 setTruncStoreAction(MVT::v32i16, MVT::v32i8, Legal);
1643 setTruncStoreAction(MVT::v16i16, MVT::v16i8, Legal);
1644 if (Subtarget->hasVLX())
1645 setTruncStoreAction(MVT::v8i16, MVT::v8i8, Legal);
1647 for (int i = MVT::v32i8; i != MVT::v8i64; ++i) {
1648 const MVT VT = (MVT::SimpleValueType)i;
1650 const unsigned EltSize = VT.getVectorElementType().getSizeInBits();
1652 // Do not attempt to promote non-512-bit vectors.
1653 if (!VT.is512BitVector())
1657 setOperationAction(ISD::BUILD_VECTOR, VT, Custom);
1658 setOperationAction(ISD::VSELECT, VT, Legal);
1663 if (!Subtarget->useSoftFloat() && Subtarget->hasVLX()) {
1664 addRegisterClass(MVT::v4i1, &X86::VK4RegClass);
1665 addRegisterClass(MVT::v2i1, &X86::VK2RegClass);
1667 setOperationAction(ISD::SETCC, MVT::v4i1, Custom);
1668 setOperationAction(ISD::SETCC, MVT::v2i1, Custom);
1669 setOperationAction(ISD::CONCAT_VECTORS, MVT::v4i1, Custom);
1670 setOperationAction(ISD::CONCAT_VECTORS, MVT::v8i1, Custom);
1671 setOperationAction(ISD::INSERT_SUBVECTOR, MVT::v8i1, Custom);
1672 setOperationAction(ISD::INSERT_SUBVECTOR, MVT::v4i1, Custom);
1673 setOperationAction(ISD::SELECT, MVT::v4i1, Custom);
1674 setOperationAction(ISD::SELECT, MVT::v2i1, Custom);
1675 setOperationAction(ISD::BUILD_VECTOR, MVT::v4i1, Custom);
1676 setOperationAction(ISD::BUILD_VECTOR, MVT::v2i1, Custom);
1677 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2i1, Custom);
1678 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4i1, Custom);
1680 setOperationAction(ISD::AND, MVT::v8i32, Legal);
1681 setOperationAction(ISD::OR, MVT::v8i32, Legal);
1682 setOperationAction(ISD::XOR, MVT::v8i32, Legal);
1683 setOperationAction(ISD::AND, MVT::v4i32, Legal);
1684 setOperationAction(ISD::OR, MVT::v4i32, Legal);
1685 setOperationAction(ISD::XOR, MVT::v4i32, Legal);
1686 setOperationAction(ISD::SRA, MVT::v2i64, Custom);
1687 setOperationAction(ISD::SRA, MVT::v4i64, Custom);
1689 setOperationAction(ISD::SMAX, MVT::v2i64, Legal);
1690 setOperationAction(ISD::SMAX, MVT::v4i64, Legal);
1691 setOperationAction(ISD::UMAX, MVT::v2i64, Legal);
1692 setOperationAction(ISD::UMAX, MVT::v4i64, Legal);
1693 setOperationAction(ISD::SMIN, MVT::v2i64, Legal);
1694 setOperationAction(ISD::SMIN, MVT::v4i64, Legal);
1695 setOperationAction(ISD::UMIN, MVT::v2i64, Legal);
1696 setOperationAction(ISD::UMIN, MVT::v4i64, Legal);
1699 // We want to custom lower some of our intrinsics.
1700 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
1701 setOperationAction(ISD::INTRINSIC_W_CHAIN, MVT::Other, Custom);
1702 setOperationAction(ISD::INTRINSIC_VOID, MVT::Other, Custom);
1703 if (!Subtarget->is64Bit())
1704 setOperationAction(ISD::INTRINSIC_W_CHAIN, MVT::i64, Custom);
1706 // Only custom-lower 64-bit SADDO and friends on 64-bit because we don't
1707 // handle type legalization for these operations here.
1709 // FIXME: We really should do custom legalization for addition and
1710 // subtraction on x86-32 once PR3203 is fixed. We really can't do much better
1711 // than generic legalization for 64-bit multiplication-with-overflow, though.
1712 for (unsigned i = 0, e = 3+Subtarget->is64Bit(); i != e; ++i) {
1713 // Add/Sub/Mul with overflow operations are custom lowered.
1715 setOperationAction(ISD::SADDO, VT, Custom);
1716 setOperationAction(ISD::UADDO, VT, Custom);
1717 setOperationAction(ISD::SSUBO, VT, Custom);
1718 setOperationAction(ISD::USUBO, VT, Custom);
1719 setOperationAction(ISD::SMULO, VT, Custom);
1720 setOperationAction(ISD::UMULO, VT, Custom);
1724 if (!Subtarget->is64Bit()) {
1725 // These libcalls are not available in 32-bit.
1726 setLibcallName(RTLIB::SHL_I128, nullptr);
1727 setLibcallName(RTLIB::SRL_I128, nullptr);
1728 setLibcallName(RTLIB::SRA_I128, nullptr);
1731 // Combine sin / cos into one node or libcall if possible.
1732 if (Subtarget->hasSinCos()) {
1733 setLibcallName(RTLIB::SINCOS_F32, "sincosf");
1734 setLibcallName(RTLIB::SINCOS_F64, "sincos");
1735 if (Subtarget->isTargetDarwin()) {
1736 // For MacOSX, we don't want the normal expansion of a libcall to sincos.
1737 // We want to issue a libcall to __sincos_stret to avoid memory traffic.
1738 setOperationAction(ISD::FSINCOS, MVT::f64, Custom);
1739 setOperationAction(ISD::FSINCOS, MVT::f32, Custom);
1743 if (Subtarget->isTargetWin64()) {
1744 setOperationAction(ISD::SDIV, MVT::i128, Custom);
1745 setOperationAction(ISD::UDIV, MVT::i128, Custom);
1746 setOperationAction(ISD::SREM, MVT::i128, Custom);
1747 setOperationAction(ISD::UREM, MVT::i128, Custom);
1748 setOperationAction(ISD::SDIVREM, MVT::i128, Custom);
1749 setOperationAction(ISD::UDIVREM, MVT::i128, Custom);
1752 // We have target-specific dag combine patterns for the following nodes:
1753 setTargetDAGCombine(ISD::VECTOR_SHUFFLE);
1754 setTargetDAGCombine(ISD::EXTRACT_VECTOR_ELT);
1755 setTargetDAGCombine(ISD::BITCAST);
1756 setTargetDAGCombine(ISD::VSELECT);
1757 setTargetDAGCombine(ISD::SELECT);
1758 setTargetDAGCombine(ISD::SHL);
1759 setTargetDAGCombine(ISD::SRA);
1760 setTargetDAGCombine(ISD::SRL);
1761 setTargetDAGCombine(ISD::OR);
1762 setTargetDAGCombine(ISD::AND);
1763 setTargetDAGCombine(ISD::ADD);
1764 setTargetDAGCombine(ISD::FADD);
1765 setTargetDAGCombine(ISD::FSUB);
1766 setTargetDAGCombine(ISD::FMA);
1767 setTargetDAGCombine(ISD::SUB);
1768 setTargetDAGCombine(ISD::LOAD);
1769 setTargetDAGCombine(ISD::MLOAD);
1770 setTargetDAGCombine(ISD::STORE);
1771 setTargetDAGCombine(ISD::MSTORE);
1772 setTargetDAGCombine(ISD::ZERO_EXTEND);
1773 setTargetDAGCombine(ISD::ANY_EXTEND);
1774 setTargetDAGCombine(ISD::SIGN_EXTEND);
1775 setTargetDAGCombine(ISD::SIGN_EXTEND_INREG);
1776 setTargetDAGCombine(ISD::SINT_TO_FP);
1777 setTargetDAGCombine(ISD::UINT_TO_FP);
1778 setTargetDAGCombine(ISD::SETCC);
1779 setTargetDAGCombine(ISD::BUILD_VECTOR);
1780 setTargetDAGCombine(ISD::MUL);
1781 setTargetDAGCombine(ISD::XOR);
1783 computeRegisterProperties(Subtarget->getRegisterInfo());
1785 MaxStoresPerMemset = 16; // For @llvm.memset -> sequence of stores
1786 MaxStoresPerMemsetOptSize = 8;
1787 MaxStoresPerMemcpy = 8; // For @llvm.memcpy -> sequence of stores
1788 MaxStoresPerMemcpyOptSize = 4;
1789 MaxStoresPerMemmove = 8; // For @llvm.memmove -> sequence of stores
1790 MaxStoresPerMemmoveOptSize = 4;
1791 setPrefLoopAlignment(4); // 2^4 bytes.
1793 // Predictable cmov don't hurt on atom because it's in-order.
1794 PredictableSelectIsExpensive = !Subtarget->isAtom();
1795 EnableExtLdPromotion = true;
1796 setPrefFunctionAlignment(4); // 2^4 bytes.
1798 verifyIntrinsicTables();
1801 // This has so far only been implemented for 64-bit MachO.
1802 bool X86TargetLowering::useLoadStackGuardNode() const {
1803 return Subtarget->isTargetMachO() && Subtarget->is64Bit();
1806 TargetLoweringBase::LegalizeTypeAction
1807 X86TargetLowering::getPreferredVectorAction(EVT VT) const {
1808 if (ExperimentalVectorWideningLegalization &&
1809 VT.getVectorNumElements() != 1 &&
1810 VT.getVectorElementType().getSimpleVT() != MVT::i1)
1811 return TypeWidenVector;
1813 return TargetLoweringBase::getPreferredVectorAction(VT);
1816 EVT X86TargetLowering::getSetCCResultType(const DataLayout &DL, LLVMContext &,
1819 return Subtarget->hasAVX512() ? MVT::i1: MVT::i8;
1821 const unsigned NumElts = VT.getVectorNumElements();
1822 const EVT EltVT = VT.getVectorElementType();
1823 if (VT.is512BitVector()) {
1824 if (Subtarget->hasAVX512())
1825 if (EltVT == MVT::i32 || EltVT == MVT::i64 ||
1826 EltVT == MVT::f32 || EltVT == MVT::f64)
1828 case 8: return MVT::v8i1;
1829 case 16: return MVT::v16i1;
1831 if (Subtarget->hasBWI())
1832 if (EltVT == MVT::i8 || EltVT == MVT::i16)
1834 case 32: return MVT::v32i1;
1835 case 64: return MVT::v64i1;
1839 if (VT.is256BitVector() || VT.is128BitVector()) {
1840 if (Subtarget->hasVLX())
1841 if (EltVT == MVT::i32 || EltVT == MVT::i64 ||
1842 EltVT == MVT::f32 || EltVT == MVT::f64)
1844 case 2: return MVT::v2i1;
1845 case 4: return MVT::v4i1;
1846 case 8: return MVT::v8i1;
1848 if (Subtarget->hasBWI() && Subtarget->hasVLX())
1849 if (EltVT == MVT::i8 || EltVT == MVT::i16)
1851 case 8: return MVT::v8i1;
1852 case 16: return MVT::v16i1;
1853 case 32: return MVT::v32i1;
1857 return VT.changeVectorElementTypeToInteger();
1860 /// Helper for getByValTypeAlignment to determine
1861 /// the desired ByVal argument alignment.
1862 static void getMaxByValAlign(Type *Ty, unsigned &MaxAlign) {
1865 if (VectorType *VTy = dyn_cast<VectorType>(Ty)) {
1866 if (VTy->getBitWidth() == 128)
1868 } else if (ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
1869 unsigned EltAlign = 0;
1870 getMaxByValAlign(ATy->getElementType(), EltAlign);
1871 if (EltAlign > MaxAlign)
1872 MaxAlign = EltAlign;
1873 } else if (StructType *STy = dyn_cast<StructType>(Ty)) {
1874 for (auto *EltTy : STy->elements()) {
1875 unsigned EltAlign = 0;
1876 getMaxByValAlign(EltTy, EltAlign);
1877 if (EltAlign > MaxAlign)
1878 MaxAlign = EltAlign;
1885 /// Return the desired alignment for ByVal aggregate
1886 /// function arguments in the caller parameter area. For X86, aggregates
1887 /// that contain SSE vectors are placed at 16-byte boundaries while the rest
1888 /// are at 4-byte boundaries.
1889 unsigned X86TargetLowering::getByValTypeAlignment(Type *Ty,
1890 const DataLayout &DL) const {
1891 if (Subtarget->is64Bit()) {
1892 // Max of 8 and alignment of type.
1893 unsigned TyAlign = DL.getABITypeAlignment(Ty);
1900 if (Subtarget->hasSSE1())
1901 getMaxByValAlign(Ty, Align);
1905 /// Returns the target specific optimal type for load
1906 /// and store operations as a result of memset, memcpy, and memmove
1907 /// lowering. If DstAlign is zero that means it's safe to destination
1908 /// alignment can satisfy any constraint. Similarly if SrcAlign is zero it
1909 /// means there isn't a need to check it against alignment requirement,
1910 /// probably because the source does not need to be loaded. If 'IsMemset' is
1911 /// true, that means it's expanding a memset. If 'ZeroMemset' is true, that
1912 /// means it's a memset of zero. 'MemcpyStrSrc' indicates whether the memcpy
1913 /// source is constant so it does not need to be loaded.
1914 /// It returns EVT::Other if the type should be determined using generic
1915 /// target-independent logic.
1917 X86TargetLowering::getOptimalMemOpType(uint64_t Size,
1918 unsigned DstAlign, unsigned SrcAlign,
1919 bool IsMemset, bool ZeroMemset,
1921 MachineFunction &MF) const {
1922 const Function *F = MF.getFunction();
1923 if ((!IsMemset || ZeroMemset) &&
1924 !F->hasFnAttribute(Attribute::NoImplicitFloat)) {
1926 (!Subtarget->isUnalignedMem16Slow() ||
1927 ((DstAlign == 0 || DstAlign >= 16) &&
1928 (SrcAlign == 0 || SrcAlign >= 16)))) {
1930 // FIXME: Check if unaligned 32-byte accesses are slow.
1931 if (Subtarget->hasInt256())
1933 if (Subtarget->hasFp256())
1936 if (Subtarget->hasSSE2())
1938 if (Subtarget->hasSSE1())
1940 } else if (!MemcpyStrSrc && Size >= 8 &&
1941 !Subtarget->is64Bit() &&
1942 Subtarget->hasSSE2()) {
1943 // Do not use f64 to lower memcpy if source is string constant. It's
1944 // better to use i32 to avoid the loads.
1948 // This is a compromise. If we reach here, unaligned accesses may be slow on
1949 // this target. However, creating smaller, aligned accesses could be even
1950 // slower and would certainly be a lot more code.
1951 if (Subtarget->is64Bit() && Size >= 8)
1956 bool X86TargetLowering::isSafeMemOpType(MVT VT) const {
1958 return X86ScalarSSEf32;
1959 else if (VT == MVT::f64)
1960 return X86ScalarSSEf64;
1965 X86TargetLowering::allowsMisalignedMemoryAccesses(EVT VT,
1970 switch (VT.getSizeInBits()) {
1972 // 8-byte and under are always assumed to be fast.
1976 *Fast = !Subtarget->isUnalignedMem16Slow();
1979 *Fast = !Subtarget->isUnalignedMem32Slow();
1981 // TODO: What about AVX-512 (512-bit) accesses?
1984 // Misaligned accesses of any size are always allowed.
1988 /// Return the entry encoding for a jump table in the
1989 /// current function. The returned value is a member of the
1990 /// MachineJumpTableInfo::JTEntryKind enum.
1991 unsigned X86TargetLowering::getJumpTableEncoding() const {
1992 // In GOT pic mode, each entry in the jump table is emitted as a @GOTOFF
1994 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
1995 Subtarget->isPICStyleGOT())
1996 return MachineJumpTableInfo::EK_Custom32;
1998 // Otherwise, use the normal jump table encoding heuristics.
1999 return TargetLowering::getJumpTableEncoding();
2002 bool X86TargetLowering::useSoftFloat() const {
2003 return Subtarget->useSoftFloat();
2007 X86TargetLowering::LowerCustomJumpTableEntry(const MachineJumpTableInfo *MJTI,
2008 const MachineBasicBlock *MBB,
2009 unsigned uid,MCContext &Ctx) const{
2010 assert(MBB->getParent()->getTarget().getRelocationModel() == Reloc::PIC_ &&
2011 Subtarget->isPICStyleGOT());
2012 // In 32-bit ELF systems, our jump table entries are formed with @GOTOFF
2014 return MCSymbolRefExpr::create(MBB->getSymbol(),
2015 MCSymbolRefExpr::VK_GOTOFF, Ctx);
2018 /// Returns relocation base for the given PIC jumptable.
2019 SDValue X86TargetLowering::getPICJumpTableRelocBase(SDValue Table,
2020 SelectionDAG &DAG) const {
2021 if (!Subtarget->is64Bit())
2022 // This doesn't have SDLoc associated with it, but is not really the
2023 // same as a Register.
2024 return DAG.getNode(X86ISD::GlobalBaseReg, SDLoc(),
2025 getPointerTy(DAG.getDataLayout()));
2029 /// This returns the relocation base for the given PIC jumptable,
2030 /// the same as getPICJumpTableRelocBase, but as an MCExpr.
2031 const MCExpr *X86TargetLowering::
2032 getPICJumpTableRelocBaseExpr(const MachineFunction *MF, unsigned JTI,
2033 MCContext &Ctx) const {
2034 // X86-64 uses RIP relative addressing based on the jump table label.
2035 if (Subtarget->isPICStyleRIPRel())
2036 return TargetLowering::getPICJumpTableRelocBaseExpr(MF, JTI, Ctx);
2038 // Otherwise, the reference is relative to the PIC base.
2039 return MCSymbolRefExpr::create(MF->getPICBaseSymbol(), Ctx);
2042 std::pair<const TargetRegisterClass *, uint8_t>
2043 X86TargetLowering::findRepresentativeClass(const TargetRegisterInfo *TRI,
2045 const TargetRegisterClass *RRC = nullptr;
2047 switch (VT.SimpleTy) {
2049 return TargetLowering::findRepresentativeClass(TRI, VT);
2050 case MVT::i8: case MVT::i16: case MVT::i32: case MVT::i64:
2051 RRC = Subtarget->is64Bit() ? &X86::GR64RegClass : &X86::GR32RegClass;
2054 RRC = &X86::VR64RegClass;
2056 case MVT::f32: case MVT::f64:
2057 case MVT::v16i8: case MVT::v8i16: case MVT::v4i32: case MVT::v2i64:
2058 case MVT::v4f32: case MVT::v2f64:
2059 case MVT::v32i8: case MVT::v8i32: case MVT::v4i64: case MVT::v8f32:
2061 RRC = &X86::VR128RegClass;
2064 return std::make_pair(RRC, Cost);
2067 bool X86TargetLowering::getStackCookieLocation(unsigned &AddressSpace,
2068 unsigned &Offset) const {
2069 if (!Subtarget->isTargetLinux())
2072 if (Subtarget->is64Bit()) {
2073 // %fs:0x28, unless we're using a Kernel code model, in which case it's %gs:
2075 if (getTargetMachine().getCodeModel() == CodeModel::Kernel)
2087 bool X86TargetLowering::isNoopAddrSpaceCast(unsigned SrcAS,
2088 unsigned DestAS) const {
2089 assert(SrcAS != DestAS && "Expected different address spaces!");
2091 return SrcAS < 256 && DestAS < 256;
2094 //===----------------------------------------------------------------------===//
2095 // Return Value Calling Convention Implementation
2096 //===----------------------------------------------------------------------===//
2098 #include "X86GenCallingConv.inc"
2101 X86TargetLowering::CanLowerReturn(CallingConv::ID CallConv,
2102 MachineFunction &MF, bool isVarArg,
2103 const SmallVectorImpl<ISD::OutputArg> &Outs,
2104 LLVMContext &Context) const {
2105 SmallVector<CCValAssign, 16> RVLocs;
2106 CCState CCInfo(CallConv, isVarArg, MF, RVLocs, Context);
2107 return CCInfo.CheckReturn(Outs, RetCC_X86);
2110 const MCPhysReg *X86TargetLowering::getScratchRegisters(CallingConv::ID) const {
2111 static const MCPhysReg ScratchRegs[] = { X86::R11, 0 };
2116 X86TargetLowering::LowerReturn(SDValue Chain,
2117 CallingConv::ID CallConv, bool isVarArg,
2118 const SmallVectorImpl<ISD::OutputArg> &Outs,
2119 const SmallVectorImpl<SDValue> &OutVals,
2120 SDLoc dl, SelectionDAG &DAG) const {
2121 MachineFunction &MF = DAG.getMachineFunction();
2122 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
2124 SmallVector<CCValAssign, 16> RVLocs;
2125 CCState CCInfo(CallConv, isVarArg, MF, RVLocs, *DAG.getContext());
2126 CCInfo.AnalyzeReturn(Outs, RetCC_X86);
2129 SmallVector<SDValue, 6> RetOps;
2130 RetOps.push_back(Chain); // Operand #0 = Chain (updated below)
2131 // Operand #1 = Bytes To Pop
2132 RetOps.push_back(DAG.getTargetConstant(FuncInfo->getBytesToPopOnReturn(), dl,
2135 // Copy the result values into the output registers.
2136 for (unsigned i = 0; i != RVLocs.size(); ++i) {
2137 CCValAssign &VA = RVLocs[i];
2138 assert(VA.isRegLoc() && "Can only return in registers!");
2139 SDValue ValToCopy = OutVals[i];
2140 EVT ValVT = ValToCopy.getValueType();
2142 // Promote values to the appropriate types.
2143 if (VA.getLocInfo() == CCValAssign::SExt)
2144 ValToCopy = DAG.getNode(ISD::SIGN_EXTEND, dl, VA.getLocVT(), ValToCopy);
2145 else if (VA.getLocInfo() == CCValAssign::ZExt)
2146 ValToCopy = DAG.getNode(ISD::ZERO_EXTEND, dl, VA.getLocVT(), ValToCopy);
2147 else if (VA.getLocInfo() == CCValAssign::AExt) {
2148 if (ValVT.isVector() && ValVT.getScalarType() == MVT::i1)
2149 ValToCopy = DAG.getNode(ISD::SIGN_EXTEND, dl, VA.getLocVT(), ValToCopy);
2151 ValToCopy = DAG.getNode(ISD::ANY_EXTEND, dl, VA.getLocVT(), ValToCopy);
2153 else if (VA.getLocInfo() == CCValAssign::BCvt)
2154 ValToCopy = DAG.getBitcast(VA.getLocVT(), ValToCopy);
2156 assert(VA.getLocInfo() != CCValAssign::FPExt &&
2157 "Unexpected FP-extend for return value.");
2159 // If this is x86-64, and we disabled SSE, we can't return FP values,
2160 // or SSE or MMX vectors.
2161 if ((ValVT == MVT::f32 || ValVT == MVT::f64 ||
2162 VA.getLocReg() == X86::XMM0 || VA.getLocReg() == X86::XMM1) &&
2163 (Subtarget->is64Bit() && !Subtarget->hasSSE1())) {
2164 report_fatal_error("SSE register return with SSE disabled");
2166 // Likewise we can't return F64 values with SSE1 only. gcc does so, but
2167 // llvm-gcc has never done it right and no one has noticed, so this
2168 // should be OK for now.
2169 if (ValVT == MVT::f64 &&
2170 (Subtarget->is64Bit() && !Subtarget->hasSSE2()))
2171 report_fatal_error("SSE2 register return with SSE2 disabled");
2173 // Returns in ST0/ST1 are handled specially: these are pushed as operands to
2174 // the RET instruction and handled by the FP Stackifier.
2175 if (VA.getLocReg() == X86::FP0 ||
2176 VA.getLocReg() == X86::FP1) {
2177 // If this is a copy from an xmm register to ST(0), use an FPExtend to
2178 // change the value to the FP stack register class.
2179 if (isScalarFPTypeInSSEReg(VA.getValVT()))
2180 ValToCopy = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f80, ValToCopy);
2181 RetOps.push_back(ValToCopy);
2182 // Don't emit a copytoreg.
2186 // 64-bit vector (MMX) values are returned in XMM0 / XMM1 except for v1i64
2187 // which is returned in RAX / RDX.
2188 if (Subtarget->is64Bit()) {
2189 if (ValVT == MVT::x86mmx) {
2190 if (VA.getLocReg() == X86::XMM0 || VA.getLocReg() == X86::XMM1) {
2191 ValToCopy = DAG.getBitcast(MVT::i64, ValToCopy);
2192 ValToCopy = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i64,
2194 // If we don't have SSE2 available, convert to v4f32 so the generated
2195 // register is legal.
2196 if (!Subtarget->hasSSE2())
2197 ValToCopy = DAG.getBitcast(MVT::v4f32, ValToCopy);
2202 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), ValToCopy, Flag);
2203 Flag = Chain.getValue(1);
2204 RetOps.push_back(DAG.getRegister(VA.getLocReg(), VA.getLocVT()));
2207 // All x86 ABIs require that for returning structs by value we copy
2208 // the sret argument into %rax/%eax (depending on ABI) for the return.
2209 // We saved the argument into a virtual register in the entry block,
2210 // so now we copy the value out and into %rax/%eax.
2212 // Checking Function.hasStructRetAttr() here is insufficient because the IR
2213 // may not have an explicit sret argument. If FuncInfo.CanLowerReturn is
2214 // false, then an sret argument may be implicitly inserted in the SelDAG. In
2215 // either case FuncInfo->setSRetReturnReg() will have been called.
2216 if (unsigned SRetReg = FuncInfo->getSRetReturnReg()) {
2217 SDValue Val = DAG.getCopyFromReg(Chain, dl, SRetReg,
2218 getPointerTy(MF.getDataLayout()));
2221 = (Subtarget->is64Bit() && !Subtarget->isTarget64BitILP32()) ?
2222 X86::RAX : X86::EAX;
2223 Chain = DAG.getCopyToReg(Chain, dl, RetValReg, Val, Flag);
2224 Flag = Chain.getValue(1);
2226 // RAX/EAX now acts like a return value.
2228 DAG.getRegister(RetValReg, getPointerTy(DAG.getDataLayout())));
2231 RetOps[0] = Chain; // Update chain.
2233 // Add the flag if we have it.
2235 RetOps.push_back(Flag);
2237 return DAG.getNode(X86ISD::RET_FLAG, dl, MVT::Other, RetOps);
2240 bool X86TargetLowering::isUsedByReturnOnly(SDNode *N, SDValue &Chain) const {
2241 if (N->getNumValues() != 1)
2243 if (!N->hasNUsesOfValue(1, 0))
2246 SDValue TCChain = Chain;
2247 SDNode *Copy = *N->use_begin();
2248 if (Copy->getOpcode() == ISD::CopyToReg) {
2249 // If the copy has a glue operand, we conservatively assume it isn't safe to
2250 // perform a tail call.
2251 if (Copy->getOperand(Copy->getNumOperands()-1).getValueType() == MVT::Glue)
2253 TCChain = Copy->getOperand(0);
2254 } else if (Copy->getOpcode() != ISD::FP_EXTEND)
2257 bool HasRet = false;
2258 for (SDNode::use_iterator UI = Copy->use_begin(), UE = Copy->use_end();
2260 if (UI->getOpcode() != X86ISD::RET_FLAG)
2262 // If we are returning more than one value, we can definitely
2263 // not make a tail call see PR19530
2264 if (UI->getNumOperands() > 4)
2266 if (UI->getNumOperands() == 4 &&
2267 UI->getOperand(UI->getNumOperands()-1).getValueType() != MVT::Glue)
2280 X86TargetLowering::getTypeForExtArgOrReturn(LLVMContext &Context, EVT VT,
2281 ISD::NodeType ExtendKind) const {
2283 // TODO: Is this also valid on 32-bit?
2284 if (Subtarget->is64Bit() && VT == MVT::i1 && ExtendKind == ISD::ZERO_EXTEND)
2285 ReturnMVT = MVT::i8;
2287 ReturnMVT = MVT::i32;
2289 EVT MinVT = getRegisterType(Context, ReturnMVT);
2290 return VT.bitsLT(MinVT) ? MinVT : VT;
2293 /// Lower the result values of a call into the
2294 /// appropriate copies out of appropriate physical registers.
2297 X86TargetLowering::LowerCallResult(SDValue Chain, SDValue InFlag,
2298 CallingConv::ID CallConv, bool isVarArg,
2299 const SmallVectorImpl<ISD::InputArg> &Ins,
2300 SDLoc dl, SelectionDAG &DAG,
2301 SmallVectorImpl<SDValue> &InVals) const {
2303 // Assign locations to each value returned by this call.
2304 SmallVector<CCValAssign, 16> RVLocs;
2305 bool Is64Bit = Subtarget->is64Bit();
2306 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(), RVLocs,
2308 CCInfo.AnalyzeCallResult(Ins, RetCC_X86);
2310 // Copy all of the result registers out of their specified physreg.
2311 for (unsigned i = 0, e = RVLocs.size(); i != e; ++i) {
2312 CCValAssign &VA = RVLocs[i];
2313 EVT CopyVT = VA.getLocVT();
2315 // If this is x86-64, and we disabled SSE, we can't return FP values
2316 if ((CopyVT == MVT::f32 || CopyVT == MVT::f64) &&
2317 ((Is64Bit || Ins[i].Flags.isInReg()) && !Subtarget->hasSSE1())) {
2318 report_fatal_error("SSE register return with SSE disabled");
2321 // If we prefer to use the value in xmm registers, copy it out as f80 and
2322 // use a truncate to move it from fp stack reg to xmm reg.
2323 bool RoundAfterCopy = false;
2324 if ((VA.getLocReg() == X86::FP0 || VA.getLocReg() == X86::FP1) &&
2325 isScalarFPTypeInSSEReg(VA.getValVT())) {
2327 RoundAfterCopy = (CopyVT != VA.getLocVT());
2330 Chain = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(),
2331 CopyVT, InFlag).getValue(1);
2332 SDValue Val = Chain.getValue(0);
2335 Val = DAG.getNode(ISD::FP_ROUND, dl, VA.getValVT(), Val,
2336 // This truncation won't change the value.
2337 DAG.getIntPtrConstant(1, dl));
2339 if (VA.isExtInLoc() && VA.getValVT().getScalarType() == MVT::i1)
2340 Val = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), Val);
2342 InFlag = Chain.getValue(2);
2343 InVals.push_back(Val);
2349 //===----------------------------------------------------------------------===//
2350 // C & StdCall & Fast Calling Convention implementation
2351 //===----------------------------------------------------------------------===//
2352 // StdCall calling convention seems to be standard for many Windows' API
2353 // routines and around. It differs from C calling convention just a little:
2354 // callee should clean up the stack, not caller. Symbols should be also
2355 // decorated in some fancy way :) It doesn't support any vector arguments.
2356 // For info on fast calling convention see Fast Calling Convention (tail call)
2357 // implementation LowerX86_32FastCCCallTo.
2359 /// CallIsStructReturn - Determines whether a call uses struct return
2361 enum StructReturnType {
2366 static StructReturnType
2367 callIsStructReturn(const SmallVectorImpl<ISD::OutputArg> &Outs) {
2369 return NotStructReturn;
2371 const ISD::ArgFlagsTy &Flags = Outs[0].Flags;
2372 if (!Flags.isSRet())
2373 return NotStructReturn;
2374 if (Flags.isInReg())
2375 return RegStructReturn;
2376 return StackStructReturn;
2379 /// Determines whether a function uses struct return semantics.
2380 static StructReturnType
2381 argsAreStructReturn(const SmallVectorImpl<ISD::InputArg> &Ins) {
2383 return NotStructReturn;
2385 const ISD::ArgFlagsTy &Flags = Ins[0].Flags;
2386 if (!Flags.isSRet())
2387 return NotStructReturn;
2388 if (Flags.isInReg())
2389 return RegStructReturn;
2390 return StackStructReturn;
2393 /// Make a copy of an aggregate at address specified by "Src" to address
2394 /// "Dst" with size and alignment information specified by the specific
2395 /// parameter attribute. The copy will be passed as a byval function parameter.
2397 CreateCopyOfByValArgument(SDValue Src, SDValue Dst, SDValue Chain,
2398 ISD::ArgFlagsTy Flags, SelectionDAG &DAG,
2400 SDValue SizeNode = DAG.getConstant(Flags.getByValSize(), dl, MVT::i32);
2402 return DAG.getMemcpy(Chain, dl, Dst, Src, SizeNode, Flags.getByValAlign(),
2403 /*isVolatile*/false, /*AlwaysInline=*/true,
2404 /*isTailCall*/false,
2405 MachinePointerInfo(), MachinePointerInfo());
2408 /// Return true if the calling convention is one that
2409 /// supports tail call optimization.
2410 static bool IsTailCallConvention(CallingConv::ID CC) {
2411 return (CC == CallingConv::Fast || CC == CallingConv::GHC ||
2412 CC == CallingConv::HiPE);
2415 /// \brief Return true if the calling convention is a C calling convention.
2416 static bool IsCCallConvention(CallingConv::ID CC) {
2417 return (CC == CallingConv::C || CC == CallingConv::X86_64_Win64 ||
2418 CC == CallingConv::X86_64_SysV);
2421 bool X86TargetLowering::mayBeEmittedAsTailCall(CallInst *CI) const {
2423 CI->getParent()->getParent()->getFnAttribute("disable-tail-calls");
2424 if (!CI->isTailCall() || Attr.getValueAsString() == "true")
2428 CallingConv::ID CalleeCC = CS.getCallingConv();
2429 if (!IsTailCallConvention(CalleeCC) && !IsCCallConvention(CalleeCC))
2435 /// Return true if the function is being made into
2436 /// a tailcall target by changing its ABI.
2437 static bool FuncIsMadeTailCallSafe(CallingConv::ID CC,
2438 bool GuaranteedTailCallOpt) {
2439 return GuaranteedTailCallOpt && IsTailCallConvention(CC);
2443 X86TargetLowering::LowerMemArgument(SDValue Chain,
2444 CallingConv::ID CallConv,
2445 const SmallVectorImpl<ISD::InputArg> &Ins,
2446 SDLoc dl, SelectionDAG &DAG,
2447 const CCValAssign &VA,
2448 MachineFrameInfo *MFI,
2450 // Create the nodes corresponding to a load from this parameter slot.
2451 ISD::ArgFlagsTy Flags = Ins[i].Flags;
2452 bool AlwaysUseMutable = FuncIsMadeTailCallSafe(
2453 CallConv, DAG.getTarget().Options.GuaranteedTailCallOpt);
2454 bool isImmutable = !AlwaysUseMutable && !Flags.isByVal();
2457 // If value is passed by pointer we have address passed instead of the value
2459 bool ExtendedInMem = VA.isExtInLoc() &&
2460 VA.getValVT().getScalarType() == MVT::i1;
2462 if (VA.getLocInfo() == CCValAssign::Indirect || ExtendedInMem)
2463 ValVT = VA.getLocVT();
2465 ValVT = VA.getValVT();
2467 // FIXME: For now, all byval parameter objects are marked mutable. This can be
2468 // changed with more analysis.
2469 // In case of tail call optimization mark all arguments mutable. Since they
2470 // could be overwritten by lowering of arguments in case of a tail call.
2471 if (Flags.isByVal()) {
2472 unsigned Bytes = Flags.getByValSize();
2473 if (Bytes == 0) Bytes = 1; // Don't create zero-sized stack objects.
2474 int FI = MFI->CreateFixedObject(Bytes, VA.getLocMemOffset(), isImmutable);
2475 return DAG.getFrameIndex(FI, getPointerTy(DAG.getDataLayout()));
2477 int FI = MFI->CreateFixedObject(ValVT.getSizeInBits()/8,
2478 VA.getLocMemOffset(), isImmutable);
2479 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy(DAG.getDataLayout()));
2480 SDValue Val = DAG.getLoad(
2481 ValVT, dl, Chain, FIN,
2482 MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), FI), false,
2484 return ExtendedInMem ?
2485 DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), Val) : Val;
2489 // FIXME: Get this from tablegen.
2490 static ArrayRef<MCPhysReg> get64BitArgumentGPRs(CallingConv::ID CallConv,
2491 const X86Subtarget *Subtarget) {
2492 assert(Subtarget->is64Bit());
2494 if (Subtarget->isCallingConvWin64(CallConv)) {
2495 static const MCPhysReg GPR64ArgRegsWin64[] = {
2496 X86::RCX, X86::RDX, X86::R8, X86::R9
2498 return makeArrayRef(std::begin(GPR64ArgRegsWin64), std::end(GPR64ArgRegsWin64));
2501 static const MCPhysReg GPR64ArgRegs64Bit[] = {
2502 X86::RDI, X86::RSI, X86::RDX, X86::RCX, X86::R8, X86::R9
2504 return makeArrayRef(std::begin(GPR64ArgRegs64Bit), std::end(GPR64ArgRegs64Bit));
2507 // FIXME: Get this from tablegen.
2508 static ArrayRef<MCPhysReg> get64BitArgumentXMMs(MachineFunction &MF,
2509 CallingConv::ID CallConv,
2510 const X86Subtarget *Subtarget) {
2511 assert(Subtarget->is64Bit());
2512 if (Subtarget->isCallingConvWin64(CallConv)) {
2513 // The XMM registers which might contain var arg parameters are shadowed
2514 // in their paired GPR. So we only need to save the GPR to their home
2516 // TODO: __vectorcall will change this.
2520 const Function *Fn = MF.getFunction();
2521 bool NoImplicitFloatOps = Fn->hasFnAttribute(Attribute::NoImplicitFloat);
2522 bool isSoftFloat = Subtarget->useSoftFloat();
2523 assert(!(isSoftFloat && NoImplicitFloatOps) &&
2524 "SSE register cannot be used when SSE is disabled!");
2525 if (isSoftFloat || NoImplicitFloatOps || !Subtarget->hasSSE1())
2526 // Kernel mode asks for SSE to be disabled, so there are no XMM argument
2530 static const MCPhysReg XMMArgRegs64Bit[] = {
2531 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
2532 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
2534 return makeArrayRef(std::begin(XMMArgRegs64Bit), std::end(XMMArgRegs64Bit));
2538 X86TargetLowering::LowerFormalArguments(SDValue Chain,
2539 CallingConv::ID CallConv,
2541 const SmallVectorImpl<ISD::InputArg> &Ins,
2544 SmallVectorImpl<SDValue> &InVals)
2546 MachineFunction &MF = DAG.getMachineFunction();
2547 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
2548 const TargetFrameLowering &TFI = *Subtarget->getFrameLowering();
2550 const Function* Fn = MF.getFunction();
2551 if (Fn->hasExternalLinkage() &&
2552 Subtarget->isTargetCygMing() &&
2553 Fn->getName() == "main")
2554 FuncInfo->setForceFramePointer(true);
2556 MachineFrameInfo *MFI = MF.getFrameInfo();
2557 bool Is64Bit = Subtarget->is64Bit();
2558 bool IsWin64 = Subtarget->isCallingConvWin64(CallConv);
2560 assert(!(isVarArg && IsTailCallConvention(CallConv)) &&
2561 "Var args not supported with calling convention fastcc, ghc or hipe");
2563 // Assign locations to all of the incoming arguments.
2564 SmallVector<CCValAssign, 16> ArgLocs;
2565 CCState CCInfo(CallConv, isVarArg, MF, ArgLocs, *DAG.getContext());
2567 // Allocate shadow area for Win64
2569 CCInfo.AllocateStack(32, 8);
2571 CCInfo.AnalyzeFormalArguments(Ins, CC_X86);
2573 unsigned LastVal = ~0U;
2575 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2576 CCValAssign &VA = ArgLocs[i];
2577 // TODO: If an arg is passed in two places (e.g. reg and stack), skip later
2579 assert(VA.getValNo() != LastVal &&
2580 "Don't support value assigned to multiple locs yet");
2582 LastVal = VA.getValNo();
2584 if (VA.isRegLoc()) {
2585 EVT RegVT = VA.getLocVT();
2586 const TargetRegisterClass *RC;
2587 if (RegVT == MVT::i32)
2588 RC = &X86::GR32RegClass;
2589 else if (Is64Bit && RegVT == MVT::i64)
2590 RC = &X86::GR64RegClass;
2591 else if (RegVT == MVT::f32)
2592 RC = &X86::FR32RegClass;
2593 else if (RegVT == MVT::f64)
2594 RC = &X86::FR64RegClass;
2595 else if (RegVT.is512BitVector())
2596 RC = &X86::VR512RegClass;
2597 else if (RegVT.is256BitVector())
2598 RC = &X86::VR256RegClass;
2599 else if (RegVT.is128BitVector())
2600 RC = &X86::VR128RegClass;
2601 else if (RegVT == MVT::x86mmx)
2602 RC = &X86::VR64RegClass;
2603 else if (RegVT == MVT::i1)
2604 RC = &X86::VK1RegClass;
2605 else if (RegVT == MVT::v8i1)
2606 RC = &X86::VK8RegClass;
2607 else if (RegVT == MVT::v16i1)
2608 RC = &X86::VK16RegClass;
2609 else if (RegVT == MVT::v32i1)
2610 RC = &X86::VK32RegClass;
2611 else if (RegVT == MVT::v64i1)
2612 RC = &X86::VK64RegClass;
2614 llvm_unreachable("Unknown argument type!");
2616 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
2617 ArgValue = DAG.getCopyFromReg(Chain, dl, Reg, RegVT);
2619 // If this is an 8 or 16-bit value, it is really passed promoted to 32
2620 // bits. Insert an assert[sz]ext to capture this, then truncate to the
2622 if (VA.getLocInfo() == CCValAssign::SExt)
2623 ArgValue = DAG.getNode(ISD::AssertSext, dl, RegVT, ArgValue,
2624 DAG.getValueType(VA.getValVT()));
2625 else if (VA.getLocInfo() == CCValAssign::ZExt)
2626 ArgValue = DAG.getNode(ISD::AssertZext, dl, RegVT, ArgValue,
2627 DAG.getValueType(VA.getValVT()));
2628 else if (VA.getLocInfo() == CCValAssign::BCvt)
2629 ArgValue = DAG.getBitcast(VA.getValVT(), ArgValue);
2631 if (VA.isExtInLoc()) {
2632 // Handle MMX values passed in XMM regs.
2633 if (RegVT.isVector() && VA.getValVT().getScalarType() != MVT::i1)
2634 ArgValue = DAG.getNode(X86ISD::MOVDQ2Q, dl, VA.getValVT(), ArgValue);
2636 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
2639 assert(VA.isMemLoc());
2640 ArgValue = LowerMemArgument(Chain, CallConv, Ins, dl, DAG, VA, MFI, i);
2643 // If value is passed via pointer - do a load.
2644 if (VA.getLocInfo() == CCValAssign::Indirect)
2645 ArgValue = DAG.getLoad(VA.getValVT(), dl, Chain, ArgValue,
2646 MachinePointerInfo(), false, false, false, 0);
2648 InVals.push_back(ArgValue);
2651 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2652 // All x86 ABIs require that for returning structs by value we copy the
2653 // sret argument into %rax/%eax (depending on ABI) for the return. Save
2654 // the argument into a virtual register so that we can access it from the
2656 if (Ins[i].Flags.isSRet()) {
2657 unsigned Reg = FuncInfo->getSRetReturnReg();
2659 MVT PtrTy = getPointerTy(DAG.getDataLayout());
2660 Reg = MF.getRegInfo().createVirtualRegister(getRegClassFor(PtrTy));
2661 FuncInfo->setSRetReturnReg(Reg);
2663 SDValue Copy = DAG.getCopyToReg(DAG.getEntryNode(), dl, Reg, InVals[i]);
2664 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Copy, Chain);
2669 unsigned StackSize = CCInfo.getNextStackOffset();
2670 // Align stack specially for tail calls.
2671 if (FuncIsMadeTailCallSafe(CallConv,
2672 MF.getTarget().Options.GuaranteedTailCallOpt))
2673 StackSize = GetAlignedArgumentStackSize(StackSize, DAG);
2675 // If the function takes variable number of arguments, make a frame index for
2676 // the start of the first vararg value... for expansion of llvm.va_start. We
2677 // can skip this if there are no va_start calls.
2678 if (MFI->hasVAStart() &&
2679 (Is64Bit || (CallConv != CallingConv::X86_FastCall &&
2680 CallConv != CallingConv::X86_ThisCall))) {
2681 FuncInfo->setVarArgsFrameIndex(
2682 MFI->CreateFixedObject(1, StackSize, true));
2685 MachineModuleInfo &MMI = MF.getMMI();
2686 const Function *WinEHParent = nullptr;
2687 if (MMI.hasWinEHFuncInfo(Fn))
2688 WinEHParent = MMI.getWinEHParent(Fn);
2689 bool IsWinEHParent = WinEHParent && WinEHParent == Fn;
2691 // Figure out if XMM registers are in use.
2692 assert(!(Subtarget->useSoftFloat() &&
2693 Fn->hasFnAttribute(Attribute::NoImplicitFloat)) &&
2694 "SSE register cannot be used when SSE is disabled!");
2696 // 64-bit calling conventions support varargs and register parameters, so we
2697 // have to do extra work to spill them in the prologue.
2698 if (Is64Bit && isVarArg && MFI->hasVAStart()) {
2699 // Find the first unallocated argument registers.
2700 ArrayRef<MCPhysReg> ArgGPRs = get64BitArgumentGPRs(CallConv, Subtarget);
2701 ArrayRef<MCPhysReg> ArgXMMs = get64BitArgumentXMMs(MF, CallConv, Subtarget);
2702 unsigned NumIntRegs = CCInfo.getFirstUnallocated(ArgGPRs);
2703 unsigned NumXMMRegs = CCInfo.getFirstUnallocated(ArgXMMs);
2704 assert(!(NumXMMRegs && !Subtarget->hasSSE1()) &&
2705 "SSE register cannot be used when SSE is disabled!");
2707 // Gather all the live in physical registers.
2708 SmallVector<SDValue, 6> LiveGPRs;
2709 SmallVector<SDValue, 8> LiveXMMRegs;
2711 for (MCPhysReg Reg : ArgGPRs.slice(NumIntRegs)) {
2712 unsigned GPR = MF.addLiveIn(Reg, &X86::GR64RegClass);
2714 DAG.getCopyFromReg(Chain, dl, GPR, MVT::i64));
2716 if (!ArgXMMs.empty()) {
2717 unsigned AL = MF.addLiveIn(X86::AL, &X86::GR8RegClass);
2718 ALVal = DAG.getCopyFromReg(Chain, dl, AL, MVT::i8);
2719 for (MCPhysReg Reg : ArgXMMs.slice(NumXMMRegs)) {
2720 unsigned XMMReg = MF.addLiveIn(Reg, &X86::VR128RegClass);
2721 LiveXMMRegs.push_back(
2722 DAG.getCopyFromReg(Chain, dl, XMMReg, MVT::v4f32));
2727 // Get to the caller-allocated home save location. Add 8 to account
2728 // for the return address.
2729 int HomeOffset = TFI.getOffsetOfLocalArea() + 8;
2730 FuncInfo->setRegSaveFrameIndex(
2731 MFI->CreateFixedObject(1, NumIntRegs * 8 + HomeOffset, false));
2732 // Fixup to set vararg frame on shadow area (4 x i64).
2734 FuncInfo->setVarArgsFrameIndex(FuncInfo->getRegSaveFrameIndex());
2736 // For X86-64, if there are vararg parameters that are passed via
2737 // registers, then we must store them to their spots on the stack so
2738 // they may be loaded by deferencing the result of va_next.
2739 FuncInfo->setVarArgsGPOffset(NumIntRegs * 8);
2740 FuncInfo->setVarArgsFPOffset(ArgGPRs.size() * 8 + NumXMMRegs * 16);
2741 FuncInfo->setRegSaveFrameIndex(MFI->CreateStackObject(
2742 ArgGPRs.size() * 8 + ArgXMMs.size() * 16, 16, false));
2745 // Store the integer parameter registers.
2746 SmallVector<SDValue, 8> MemOps;
2747 SDValue RSFIN = DAG.getFrameIndex(FuncInfo->getRegSaveFrameIndex(),
2748 getPointerTy(DAG.getDataLayout()));
2749 unsigned Offset = FuncInfo->getVarArgsGPOffset();
2750 for (SDValue Val : LiveGPRs) {
2751 SDValue FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(DAG.getDataLayout()),
2752 RSFIN, DAG.getIntPtrConstant(Offset, dl));
2754 DAG.getStore(Val.getValue(1), dl, Val, FIN,
2755 MachinePointerInfo::getFixedStack(
2756 DAG.getMachineFunction(),
2757 FuncInfo->getRegSaveFrameIndex(), Offset),
2759 MemOps.push_back(Store);
2763 if (!ArgXMMs.empty() && NumXMMRegs != ArgXMMs.size()) {
2764 // Now store the XMM (fp + vector) parameter registers.
2765 SmallVector<SDValue, 12> SaveXMMOps;
2766 SaveXMMOps.push_back(Chain);
2767 SaveXMMOps.push_back(ALVal);
2768 SaveXMMOps.push_back(DAG.getIntPtrConstant(
2769 FuncInfo->getRegSaveFrameIndex(), dl));
2770 SaveXMMOps.push_back(DAG.getIntPtrConstant(
2771 FuncInfo->getVarArgsFPOffset(), dl));
2772 SaveXMMOps.insert(SaveXMMOps.end(), LiveXMMRegs.begin(),
2774 MemOps.push_back(DAG.getNode(X86ISD::VASTART_SAVE_XMM_REGS, dl,
2775 MVT::Other, SaveXMMOps));
2778 if (!MemOps.empty())
2779 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, MemOps);
2782 if (isVarArg && MFI->hasMustTailInVarArgFunc()) {
2783 // Find the largest legal vector type.
2784 MVT VecVT = MVT::Other;
2785 // FIXME: Only some x86_32 calling conventions support AVX512.
2786 if (Subtarget->hasAVX512() &&
2787 (Is64Bit || (CallConv == CallingConv::X86_VectorCall ||
2788 CallConv == CallingConv::Intel_OCL_BI)))
2789 VecVT = MVT::v16f32;
2790 else if (Subtarget->hasAVX())
2792 else if (Subtarget->hasSSE2())
2795 // We forward some GPRs and some vector types.
2796 SmallVector<MVT, 2> RegParmTypes;
2797 MVT IntVT = Is64Bit ? MVT::i64 : MVT::i32;
2798 RegParmTypes.push_back(IntVT);
2799 if (VecVT != MVT::Other)
2800 RegParmTypes.push_back(VecVT);
2802 // Compute the set of forwarded registers. The rest are scratch.
2803 SmallVectorImpl<ForwardedRegister> &Forwards =
2804 FuncInfo->getForwardedMustTailRegParms();
2805 CCInfo.analyzeMustTailForwardedRegisters(Forwards, RegParmTypes, CC_X86);
2807 // Conservatively forward AL on x86_64, since it might be used for varargs.
2808 if (Is64Bit && !CCInfo.isAllocated(X86::AL)) {
2809 unsigned ALVReg = MF.addLiveIn(X86::AL, &X86::GR8RegClass);
2810 Forwards.push_back(ForwardedRegister(ALVReg, X86::AL, MVT::i8));
2813 // Copy all forwards from physical to virtual registers.
2814 for (ForwardedRegister &F : Forwards) {
2815 // FIXME: Can we use a less constrained schedule?
2816 SDValue RegVal = DAG.getCopyFromReg(Chain, dl, F.VReg, F.VT);
2817 F.VReg = MF.getRegInfo().createVirtualRegister(getRegClassFor(F.VT));
2818 Chain = DAG.getCopyToReg(Chain, dl, F.VReg, RegVal);
2822 // Some CCs need callee pop.
2823 if (X86::isCalleePop(CallConv, Is64Bit, isVarArg,
2824 MF.getTarget().Options.GuaranteedTailCallOpt)) {
2825 FuncInfo->setBytesToPopOnReturn(StackSize); // Callee pops everything.
2827 FuncInfo->setBytesToPopOnReturn(0); // Callee pops nothing.
2828 // If this is an sret function, the return should pop the hidden pointer.
2829 if (!Is64Bit && !IsTailCallConvention(CallConv) &&
2830 !Subtarget->getTargetTriple().isOSMSVCRT() &&
2831 argsAreStructReturn(Ins) == StackStructReturn)
2832 FuncInfo->setBytesToPopOnReturn(4);
2836 // RegSaveFrameIndex is X86-64 only.
2837 FuncInfo->setRegSaveFrameIndex(0xAAAAAAA);
2838 if (CallConv == CallingConv::X86_FastCall ||
2839 CallConv == CallingConv::X86_ThisCall)
2840 // fastcc functions can't have varargs.
2841 FuncInfo->setVarArgsFrameIndex(0xAAAAAAA);
2844 FuncInfo->setArgumentStackSize(StackSize);
2846 if (IsWinEHParent) {
2848 int UnwindHelpFI = MFI->CreateStackObject(8, 8, /*isSS=*/false);
2849 SDValue StackSlot = DAG.getFrameIndex(UnwindHelpFI, MVT::i64);
2850 MMI.getWinEHFuncInfo(MF.getFunction()).UnwindHelpFrameIdx = UnwindHelpFI;
2851 SDValue Neg2 = DAG.getConstant(-2, dl, MVT::i64);
2852 Chain = DAG.getStore(Chain, dl, Neg2, StackSlot,
2853 MachinePointerInfo::getFixedStack(
2854 DAG.getMachineFunction(), UnwindHelpFI),
2855 /*isVolatile=*/true,
2856 /*isNonTemporal=*/false, /*Alignment=*/0);
2858 // Functions using Win32 EH are considered to have opaque SP adjustments
2859 // to force local variables to be addressed from the frame or base
2861 MFI->setHasOpaqueSPAdjustment(true);
2869 X86TargetLowering::LowerMemOpCallTo(SDValue Chain,
2870 SDValue StackPtr, SDValue Arg,
2871 SDLoc dl, SelectionDAG &DAG,
2872 const CCValAssign &VA,
2873 ISD::ArgFlagsTy Flags) const {
2874 unsigned LocMemOffset = VA.getLocMemOffset();
2875 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset, dl);
2876 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(DAG.getDataLayout()),
2878 if (Flags.isByVal())
2879 return CreateCopyOfByValArgument(Arg, PtrOff, Chain, Flags, DAG, dl);
2881 return DAG.getStore(
2882 Chain, dl, Arg, PtrOff,
2883 MachinePointerInfo::getStack(DAG.getMachineFunction(), LocMemOffset),
2887 /// Emit a load of return address if tail call
2888 /// optimization is performed and it is required.
2890 X86TargetLowering::EmitTailCallLoadRetAddr(SelectionDAG &DAG,
2891 SDValue &OutRetAddr, SDValue Chain,
2892 bool IsTailCall, bool Is64Bit,
2893 int FPDiff, SDLoc dl) const {
2894 // Adjust the Return address stack slot.
2895 EVT VT = getPointerTy(DAG.getDataLayout());
2896 OutRetAddr = getReturnAddressFrameIndex(DAG);
2898 // Load the "old" Return address.
2899 OutRetAddr = DAG.getLoad(VT, dl, Chain, OutRetAddr, MachinePointerInfo(),
2900 false, false, false, 0);
2901 return SDValue(OutRetAddr.getNode(), 1);
2904 /// Emit a store of the return address if tail call
2905 /// optimization is performed and it is required (FPDiff!=0).
2906 static SDValue EmitTailCallStoreRetAddr(SelectionDAG &DAG, MachineFunction &MF,
2907 SDValue Chain, SDValue RetAddrFrIdx,
2908 EVT PtrVT, unsigned SlotSize,
2909 int FPDiff, SDLoc dl) {
2910 // Store the return address to the appropriate stack slot.
2911 if (!FPDiff) return Chain;
2912 // Calculate the new stack slot for the return address.
2913 int NewReturnAddrFI =
2914 MF.getFrameInfo()->CreateFixedObject(SlotSize, (int64_t)FPDiff - SlotSize,
2916 SDValue NewRetAddrFrIdx = DAG.getFrameIndex(NewReturnAddrFI, PtrVT);
2917 Chain = DAG.getStore(Chain, dl, RetAddrFrIdx, NewRetAddrFrIdx,
2918 MachinePointerInfo::getFixedStack(
2919 DAG.getMachineFunction(), NewReturnAddrFI),
2924 /// Returns a vector_shuffle mask for an movs{s|d}, movd
2925 /// operation of specified width.
2926 static SDValue getMOVL(SelectionDAG &DAG, SDLoc dl, EVT VT, SDValue V1,
2928 unsigned NumElems = VT.getVectorNumElements();
2929 SmallVector<int, 8> Mask;
2930 Mask.push_back(NumElems);
2931 for (unsigned i = 1; i != NumElems; ++i)
2933 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
2937 X86TargetLowering::LowerCall(TargetLowering::CallLoweringInfo &CLI,
2938 SmallVectorImpl<SDValue> &InVals) const {
2939 SelectionDAG &DAG = CLI.DAG;
2941 SmallVectorImpl<ISD::OutputArg> &Outs = CLI.Outs;
2942 SmallVectorImpl<SDValue> &OutVals = CLI.OutVals;
2943 SmallVectorImpl<ISD::InputArg> &Ins = CLI.Ins;
2944 SDValue Chain = CLI.Chain;
2945 SDValue Callee = CLI.Callee;
2946 CallingConv::ID CallConv = CLI.CallConv;
2947 bool &isTailCall = CLI.IsTailCall;
2948 bool isVarArg = CLI.IsVarArg;
2950 MachineFunction &MF = DAG.getMachineFunction();
2951 bool Is64Bit = Subtarget->is64Bit();
2952 bool IsWin64 = Subtarget->isCallingConvWin64(CallConv);
2953 StructReturnType SR = callIsStructReturn(Outs);
2954 bool IsSibcall = false;
2955 X86MachineFunctionInfo *X86Info = MF.getInfo<X86MachineFunctionInfo>();
2956 auto Attr = MF.getFunction()->getFnAttribute("disable-tail-calls");
2958 if (Attr.getValueAsString() == "true")
2961 if (Subtarget->isPICStyleGOT() &&
2962 !MF.getTarget().Options.GuaranteedTailCallOpt) {
2963 // If we are using a GOT, disable tail calls to external symbols with
2964 // default visibility. Tail calling such a symbol requires using a GOT
2965 // relocation, which forces early binding of the symbol. This breaks code
2966 // that require lazy function symbol resolution. Using musttail or
2967 // GuaranteedTailCallOpt will override this.
2968 GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee);
2969 if (!G || (!G->getGlobal()->hasLocalLinkage() &&
2970 G->getGlobal()->hasDefaultVisibility()))
2974 bool IsMustTail = CLI.CS && CLI.CS->isMustTailCall();
2976 // Force this to be a tail call. The verifier rules are enough to ensure
2977 // that we can lower this successfully without moving the return address
2980 } else if (isTailCall) {
2981 // Check if it's really possible to do a tail call.
2982 isTailCall = IsEligibleForTailCallOptimization(Callee, CallConv,
2983 isVarArg, SR != NotStructReturn,
2984 MF.getFunction()->hasStructRetAttr(), CLI.RetTy,
2985 Outs, OutVals, Ins, DAG);
2987 // Sibcalls are automatically detected tailcalls which do not require
2989 if (!MF.getTarget().Options.GuaranteedTailCallOpt && isTailCall)
2996 assert(!(isVarArg && IsTailCallConvention(CallConv)) &&
2997 "Var args not supported with calling convention fastcc, ghc or hipe");
2999 // Analyze operands of the call, assigning locations to each operand.
3000 SmallVector<CCValAssign, 16> ArgLocs;
3001 CCState CCInfo(CallConv, isVarArg, MF, ArgLocs, *DAG.getContext());
3003 // Allocate shadow area for Win64
3005 CCInfo.AllocateStack(32, 8);
3007 CCInfo.AnalyzeCallOperands(Outs, CC_X86);
3009 // Get a count of how many bytes are to be pushed on the stack.
3010 unsigned NumBytes = CCInfo.getNextStackOffset();
3012 // This is a sibcall. The memory operands are available in caller's
3013 // own caller's stack.
3015 else if (MF.getTarget().Options.GuaranteedTailCallOpt &&
3016 IsTailCallConvention(CallConv))
3017 NumBytes = GetAlignedArgumentStackSize(NumBytes, DAG);
3020 if (isTailCall && !IsSibcall && !IsMustTail) {
3021 // Lower arguments at fp - stackoffset + fpdiff.
3022 unsigned NumBytesCallerPushed = X86Info->getBytesToPopOnReturn();
3024 FPDiff = NumBytesCallerPushed - NumBytes;
3026 // Set the delta of movement of the returnaddr stackslot.
3027 // But only set if delta is greater than previous delta.
3028 if (FPDiff < X86Info->getTCReturnAddrDelta())
3029 X86Info->setTCReturnAddrDelta(FPDiff);
3032 unsigned NumBytesToPush = NumBytes;
3033 unsigned NumBytesToPop = NumBytes;
3035 // If we have an inalloca argument, all stack space has already been allocated
3036 // for us and be right at the top of the stack. We don't support multiple
3037 // arguments passed in memory when using inalloca.
3038 if (!Outs.empty() && Outs.back().Flags.isInAlloca()) {
3040 if (!ArgLocs.back().isMemLoc())
3041 report_fatal_error("cannot use inalloca attribute on a register "
3043 if (ArgLocs.back().getLocMemOffset() != 0)
3044 report_fatal_error("any parameter with the inalloca attribute must be "
3045 "the only memory argument");
3049 Chain = DAG.getCALLSEQ_START(
3050 Chain, DAG.getIntPtrConstant(NumBytesToPush, dl, true), dl);
3052 SDValue RetAddrFrIdx;
3053 // Load return address for tail calls.
3054 if (isTailCall && FPDiff)
3055 Chain = EmitTailCallLoadRetAddr(DAG, RetAddrFrIdx, Chain, isTailCall,
3056 Is64Bit, FPDiff, dl);
3058 SmallVector<std::pair<unsigned, SDValue>, 8> RegsToPass;
3059 SmallVector<SDValue, 8> MemOpChains;
3062 // Walk the register/memloc assignments, inserting copies/loads. In the case
3063 // of tail call optimization arguments are handle later.
3064 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
3065 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
3066 // Skip inalloca arguments, they have already been written.
3067 ISD::ArgFlagsTy Flags = Outs[i].Flags;
3068 if (Flags.isInAlloca())
3071 CCValAssign &VA = ArgLocs[i];
3072 EVT RegVT = VA.getLocVT();
3073 SDValue Arg = OutVals[i];
3074 bool isByVal = Flags.isByVal();
3076 // Promote the value if needed.
3077 switch (VA.getLocInfo()) {
3078 default: llvm_unreachable("Unknown loc info!");
3079 case CCValAssign::Full: break;
3080 case CCValAssign::SExt:
3081 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, RegVT, Arg);
3083 case CCValAssign::ZExt:
3084 Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, RegVT, Arg);
3086 case CCValAssign::AExt:
3087 if (Arg.getValueType().isVector() &&
3088 Arg.getValueType().getScalarType() == MVT::i1)
3089 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, RegVT, Arg);
3090 else if (RegVT.is128BitVector()) {
3091 // Special case: passing MMX values in XMM registers.
3092 Arg = DAG.getBitcast(MVT::i64, Arg);
3093 Arg = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i64, Arg);
3094 Arg = getMOVL(DAG, dl, MVT::v2i64, DAG.getUNDEF(MVT::v2i64), Arg);
3096 Arg = DAG.getNode(ISD::ANY_EXTEND, dl, RegVT, Arg);
3098 case CCValAssign::BCvt:
3099 Arg = DAG.getBitcast(RegVT, Arg);
3101 case CCValAssign::Indirect: {
3102 // Store the argument.
3103 SDValue SpillSlot = DAG.CreateStackTemporary(VA.getValVT());
3104 int FI = cast<FrameIndexSDNode>(SpillSlot)->getIndex();
3105 Chain = DAG.getStore(
3106 Chain, dl, Arg, SpillSlot,
3107 MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), FI),
3114 if (VA.isRegLoc()) {
3115 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
3116 if (isVarArg && IsWin64) {
3117 // Win64 ABI requires argument XMM reg to be copied to the corresponding
3118 // shadow reg if callee is a varargs function.
3119 unsigned ShadowReg = 0;
3120 switch (VA.getLocReg()) {
3121 case X86::XMM0: ShadowReg = X86::RCX; break;
3122 case X86::XMM1: ShadowReg = X86::RDX; break;
3123 case X86::XMM2: ShadowReg = X86::R8; break;
3124 case X86::XMM3: ShadowReg = X86::R9; break;
3127 RegsToPass.push_back(std::make_pair(ShadowReg, Arg));
3129 } else if (!IsSibcall && (!isTailCall || isByVal)) {
3130 assert(VA.isMemLoc());
3131 if (!StackPtr.getNode())
3132 StackPtr = DAG.getCopyFromReg(Chain, dl, RegInfo->getStackRegister(),
3133 getPointerTy(DAG.getDataLayout()));
3134 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Arg,
3135 dl, DAG, VA, Flags));
3139 if (!MemOpChains.empty())
3140 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, MemOpChains);
3142 if (Subtarget->isPICStyleGOT()) {
3143 // ELF / PIC requires GOT in the EBX register before function calls via PLT
3146 RegsToPass.push_back(std::make_pair(
3147 unsigned(X86::EBX), DAG.getNode(X86ISD::GlobalBaseReg, SDLoc(),
3148 getPointerTy(DAG.getDataLayout()))));
3150 // If we are tail calling and generating PIC/GOT style code load the
3151 // address of the callee into ECX. The value in ecx is used as target of
3152 // the tail jump. This is done to circumvent the ebx/callee-saved problem
3153 // for tail calls on PIC/GOT architectures. Normally we would just put the
3154 // address of GOT into ebx and then call target@PLT. But for tail calls
3155 // ebx would be restored (since ebx is callee saved) before jumping to the
3158 // Note: The actual moving to ECX is done further down.
3159 GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee);
3160 if (G && !G->getGlobal()->hasLocalLinkage() &&
3161 G->getGlobal()->hasDefaultVisibility())
3162 Callee = LowerGlobalAddress(Callee, DAG);
3163 else if (isa<ExternalSymbolSDNode>(Callee))
3164 Callee = LowerExternalSymbol(Callee, DAG);
3168 if (Is64Bit && isVarArg && !IsWin64 && !IsMustTail) {
3169 // From AMD64 ABI document:
3170 // For calls that may call functions that use varargs or stdargs
3171 // (prototype-less calls or calls to functions containing ellipsis (...) in
3172 // the declaration) %al is used as hidden argument to specify the number
3173 // of SSE registers used. The contents of %al do not need to match exactly
3174 // the number of registers, but must be an ubound on the number of SSE
3175 // registers used and is in the range 0 - 8 inclusive.
3177 // Count the number of XMM registers allocated.
3178 static const MCPhysReg XMMArgRegs[] = {
3179 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
3180 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
3182 unsigned NumXMMRegs = CCInfo.getFirstUnallocated(XMMArgRegs);
3183 assert((Subtarget->hasSSE1() || !NumXMMRegs)
3184 && "SSE registers cannot be used when SSE is disabled");
3186 RegsToPass.push_back(std::make_pair(unsigned(X86::AL),
3187 DAG.getConstant(NumXMMRegs, dl,
3191 if (isVarArg && IsMustTail) {
3192 const auto &Forwards = X86Info->getForwardedMustTailRegParms();
3193 for (const auto &F : Forwards) {
3194 SDValue Val = DAG.getCopyFromReg(Chain, dl, F.VReg, F.VT);
3195 RegsToPass.push_back(std::make_pair(unsigned(F.PReg), Val));
3199 // For tail calls lower the arguments to the 'real' stack slots. Sibcalls
3200 // don't need this because the eligibility check rejects calls that require
3201 // shuffling arguments passed in memory.
3202 if (!IsSibcall && isTailCall) {
3203 // Force all the incoming stack arguments to be loaded from the stack
3204 // before any new outgoing arguments are stored to the stack, because the
3205 // outgoing stack slots may alias the incoming argument stack slots, and
3206 // the alias isn't otherwise explicit. This is slightly more conservative
3207 // than necessary, because it means that each store effectively depends
3208 // on every argument instead of just those arguments it would clobber.
3209 SDValue ArgChain = DAG.getStackArgumentTokenFactor(Chain);
3211 SmallVector<SDValue, 8> MemOpChains2;
3214 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
3215 CCValAssign &VA = ArgLocs[i];
3218 assert(VA.isMemLoc());
3219 SDValue Arg = OutVals[i];
3220 ISD::ArgFlagsTy Flags = Outs[i].Flags;
3221 // Skip inalloca arguments. They don't require any work.
3222 if (Flags.isInAlloca())
3224 // Create frame index.
3225 int32_t Offset = VA.getLocMemOffset()+FPDiff;
3226 uint32_t OpSize = (VA.getLocVT().getSizeInBits()+7)/8;
3227 FI = MF.getFrameInfo()->CreateFixedObject(OpSize, Offset, true);
3228 FIN = DAG.getFrameIndex(FI, getPointerTy(DAG.getDataLayout()));
3230 if (Flags.isByVal()) {
3231 // Copy relative to framepointer.
3232 SDValue Source = DAG.getIntPtrConstant(VA.getLocMemOffset(), dl);
3233 if (!StackPtr.getNode())
3234 StackPtr = DAG.getCopyFromReg(Chain, dl, RegInfo->getStackRegister(),
3235 getPointerTy(DAG.getDataLayout()));
3236 Source = DAG.getNode(ISD::ADD, dl, getPointerTy(DAG.getDataLayout()),
3239 MemOpChains2.push_back(CreateCopyOfByValArgument(Source, FIN,
3243 // Store relative to framepointer.
3244 MemOpChains2.push_back(DAG.getStore(
3245 ArgChain, dl, Arg, FIN,
3246 MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), FI),
3251 if (!MemOpChains2.empty())
3252 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, MemOpChains2);
3254 // Store the return address to the appropriate stack slot.
3255 Chain = EmitTailCallStoreRetAddr(DAG, MF, Chain, RetAddrFrIdx,
3256 getPointerTy(DAG.getDataLayout()),
3257 RegInfo->getSlotSize(), FPDiff, dl);
3260 // Build a sequence of copy-to-reg nodes chained together with token chain
3261 // and flag operands which copy the outgoing args into registers.
3263 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
3264 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
3265 RegsToPass[i].second, InFlag);
3266 InFlag = Chain.getValue(1);
3269 if (DAG.getTarget().getCodeModel() == CodeModel::Large) {
3270 assert(Is64Bit && "Large code model is only legal in 64-bit mode.");
3271 // In the 64-bit large code model, we have to make all calls
3272 // through a register, since the call instruction's 32-bit
3273 // pc-relative offset may not be large enough to hold the whole
3275 } else if (Callee->getOpcode() == ISD::GlobalAddress) {
3276 // If the callee is a GlobalAddress node (quite common, every direct call
3277 // is) turn it into a TargetGlobalAddress node so that legalize doesn't hack
3279 GlobalAddressSDNode* G = cast<GlobalAddressSDNode>(Callee);
3281 // We should use extra load for direct calls to dllimported functions in
3283 const GlobalValue *GV = G->getGlobal();
3284 if (!GV->hasDLLImportStorageClass()) {
3285 unsigned char OpFlags = 0;
3286 bool ExtraLoad = false;
3287 unsigned WrapperKind = ISD::DELETED_NODE;
3289 // On ELF targets, in both X86-64 and X86-32 mode, direct calls to
3290 // external symbols most go through the PLT in PIC mode. If the symbol
3291 // has hidden or protected visibility, or if it is static or local, then
3292 // we don't need to use the PLT - we can directly call it.
3293 if (Subtarget->isTargetELF() &&
3294 DAG.getTarget().getRelocationModel() == Reloc::PIC_ &&
3295 GV->hasDefaultVisibility() && !GV->hasLocalLinkage()) {
3296 OpFlags = X86II::MO_PLT;
3297 } else if (Subtarget->isPICStyleStubAny() &&
3298 !GV->isStrongDefinitionForLinker() &&
3299 (!Subtarget->getTargetTriple().isMacOSX() ||
3300 Subtarget->getTargetTriple().isMacOSXVersionLT(10, 5))) {
3301 // PC-relative references to external symbols should go through $stub,
3302 // unless we're building with the leopard linker or later, which
3303 // automatically synthesizes these stubs.
3304 OpFlags = X86II::MO_DARWIN_STUB;
3305 } else if (Subtarget->isPICStyleRIPRel() && isa<Function>(GV) &&
3306 cast<Function>(GV)->hasFnAttribute(Attribute::NonLazyBind)) {
3307 // If the function is marked as non-lazy, generate an indirect call
3308 // which loads from the GOT directly. This avoids runtime overhead
3309 // at the cost of eager binding (and one extra byte of encoding).
3310 OpFlags = X86II::MO_GOTPCREL;
3311 WrapperKind = X86ISD::WrapperRIP;
3315 Callee = DAG.getTargetGlobalAddress(
3316 GV, dl, getPointerTy(DAG.getDataLayout()), G->getOffset(), OpFlags);
3318 // Add a wrapper if needed.
3319 if (WrapperKind != ISD::DELETED_NODE)
3320 Callee = DAG.getNode(X86ISD::WrapperRIP, dl,
3321 getPointerTy(DAG.getDataLayout()), Callee);
3322 // Add extra indirection if needed.
3324 Callee = DAG.getLoad(
3325 getPointerTy(DAG.getDataLayout()), dl, DAG.getEntryNode(), Callee,
3326 MachinePointerInfo::getGOT(DAG.getMachineFunction()), false, false,
3329 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
3330 unsigned char OpFlags = 0;
3332 // On ELF targets, in either X86-64 or X86-32 mode, direct calls to
3333 // external symbols should go through the PLT.
3334 if (Subtarget->isTargetELF() &&
3335 DAG.getTarget().getRelocationModel() == Reloc::PIC_) {
3336 OpFlags = X86II::MO_PLT;
3337 } else if (Subtarget->isPICStyleStubAny() &&
3338 (!Subtarget->getTargetTriple().isMacOSX() ||
3339 Subtarget->getTargetTriple().isMacOSXVersionLT(10, 5))) {
3340 // PC-relative references to external symbols should go through $stub,
3341 // unless we're building with the leopard linker or later, which
3342 // automatically synthesizes these stubs.
3343 OpFlags = X86II::MO_DARWIN_STUB;
3346 Callee = DAG.getTargetExternalSymbol(
3347 S->getSymbol(), getPointerTy(DAG.getDataLayout()), OpFlags);
3348 } else if (Subtarget->isTarget64BitILP32() &&
3349 Callee->getValueType(0) == MVT::i32) {
3350 // Zero-extend the 32-bit Callee address into a 64-bit according to x32 ABI
3351 Callee = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i64, Callee);
3354 // Returns a chain & a flag for retval copy to use.
3355 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
3356 SmallVector<SDValue, 8> Ops;
3358 if (!IsSibcall && isTailCall) {
3359 Chain = DAG.getCALLSEQ_END(Chain,
3360 DAG.getIntPtrConstant(NumBytesToPop, dl, true),
3361 DAG.getIntPtrConstant(0, dl, true), InFlag, dl);
3362 InFlag = Chain.getValue(1);
3365 Ops.push_back(Chain);
3366 Ops.push_back(Callee);
3369 Ops.push_back(DAG.getConstant(FPDiff, dl, MVT::i32));
3371 // Add argument registers to the end of the list so that they are known live
3373 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
3374 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
3375 RegsToPass[i].second.getValueType()));
3377 // Add a register mask operand representing the call-preserved registers.
3378 const uint32_t *Mask = RegInfo->getCallPreservedMask(MF, CallConv);
3379 assert(Mask && "Missing call preserved mask for calling convention");
3381 // If this is an invoke in a 32-bit function using an MSVC personality, assume
3382 // the function clobbers all registers. If an exception is thrown, the runtime
3383 // will not restore CSRs.
3384 // FIXME: Model this more precisely so that we can register allocate across
3385 // the normal edge and spill and fill across the exceptional edge.
3386 if (!Is64Bit && CLI.CS && CLI.CS->isInvoke()) {
3387 const Function *CallerFn = MF.getFunction();
3388 EHPersonality Pers =
3389 CallerFn->hasPersonalityFn()
3390 ? classifyEHPersonality(CallerFn->getPersonalityFn())
3391 : EHPersonality::Unknown;
3392 if (isMSVCEHPersonality(Pers))
3393 Mask = RegInfo->getNoPreservedMask();
3396 Ops.push_back(DAG.getRegisterMask(Mask));
3398 if (InFlag.getNode())
3399 Ops.push_back(InFlag);
3403 //// If this is the first return lowered for this function, add the regs
3404 //// to the liveout set for the function.
3405 // This isn't right, although it's probably harmless on x86; liveouts
3406 // should be computed from returns not tail calls. Consider a void
3407 // function making a tail call to a function returning int.
3408 MF.getFrameInfo()->setHasTailCall();
3409 return DAG.getNode(X86ISD::TC_RETURN, dl, NodeTys, Ops);
3412 Chain = DAG.getNode(X86ISD::CALL, dl, NodeTys, Ops);
3413 InFlag = Chain.getValue(1);
3415 // Create the CALLSEQ_END node.
3416 unsigned NumBytesForCalleeToPop;
3417 if (X86::isCalleePop(CallConv, Is64Bit, isVarArg,
3418 DAG.getTarget().Options.GuaranteedTailCallOpt))
3419 NumBytesForCalleeToPop = NumBytes; // Callee pops everything
3420 else if (!Is64Bit && !IsTailCallConvention(CallConv) &&
3421 !Subtarget->getTargetTriple().isOSMSVCRT() &&
3422 SR == StackStructReturn)
3423 // If this is a call to a struct-return function, the callee
3424 // pops the hidden struct pointer, so we have to push it back.
3425 // This is common for Darwin/X86, Linux & Mingw32 targets.
3426 // For MSVC Win32 targets, the caller pops the hidden struct pointer.
3427 NumBytesForCalleeToPop = 4;
3429 NumBytesForCalleeToPop = 0; // Callee pops nothing.
3431 // Returns a flag for retval copy to use.
3433 Chain = DAG.getCALLSEQ_END(Chain,
3434 DAG.getIntPtrConstant(NumBytesToPop, dl, true),
3435 DAG.getIntPtrConstant(NumBytesForCalleeToPop, dl,
3438 InFlag = Chain.getValue(1);
3441 // Handle result values, copying them out of physregs into vregs that we
3443 return LowerCallResult(Chain, InFlag, CallConv, isVarArg,
3444 Ins, dl, DAG, InVals);
3447 //===----------------------------------------------------------------------===//
3448 // Fast Calling Convention (tail call) implementation
3449 //===----------------------------------------------------------------------===//
3451 // Like std call, callee cleans arguments, convention except that ECX is
3452 // reserved for storing the tail called function address. Only 2 registers are
3453 // free for argument passing (inreg). Tail call optimization is performed
3455 // * tailcallopt is enabled
3456 // * caller/callee are fastcc
3457 // On X86_64 architecture with GOT-style position independent code only local
3458 // (within module) calls are supported at the moment.
3459 // To keep the stack aligned according to platform abi the function
3460 // GetAlignedArgumentStackSize ensures that argument delta is always multiples
3461 // of stack alignment. (Dynamic linkers need this - darwin's dyld for example)
3462 // If a tail called function callee has more arguments than the caller the
3463 // caller needs to make sure that there is room to move the RETADDR to. This is
3464 // achieved by reserving an area the size of the argument delta right after the
3465 // original RETADDR, but before the saved framepointer or the spilled registers
3466 // e.g. caller(arg1, arg2) calls callee(arg1, arg2,arg3,arg4)
3478 /// Make the stack size align e.g 16n + 12 aligned for a 16-byte align
3481 X86TargetLowering::GetAlignedArgumentStackSize(unsigned StackSize,
3482 SelectionDAG& DAG) const {
3483 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
3484 const TargetFrameLowering &TFI = *Subtarget->getFrameLowering();
3485 unsigned StackAlignment = TFI.getStackAlignment();
3486 uint64_t AlignMask = StackAlignment - 1;
3487 int64_t Offset = StackSize;
3488 unsigned SlotSize = RegInfo->getSlotSize();
3489 if ( (Offset & AlignMask) <= (StackAlignment - SlotSize) ) {
3490 // Number smaller than 12 so just add the difference.
3491 Offset += ((StackAlignment - SlotSize) - (Offset & AlignMask));
3493 // Mask out lower bits, add stackalignment once plus the 12 bytes.
3494 Offset = ((~AlignMask) & Offset) + StackAlignment +
3495 (StackAlignment-SlotSize);
3500 /// Return true if the given stack call argument is already available in the
3501 /// same position (relatively) of the caller's incoming argument stack.
3503 bool MatchingStackOffset(SDValue Arg, unsigned Offset, ISD::ArgFlagsTy Flags,
3504 MachineFrameInfo *MFI, const MachineRegisterInfo *MRI,
3505 const X86InstrInfo *TII) {
3506 unsigned Bytes = Arg.getValueType().getSizeInBits() / 8;
3508 if (Arg.getOpcode() == ISD::CopyFromReg) {
3509 unsigned VR = cast<RegisterSDNode>(Arg.getOperand(1))->getReg();
3510 if (!TargetRegisterInfo::isVirtualRegister(VR))
3512 MachineInstr *Def = MRI->getVRegDef(VR);
3515 if (!Flags.isByVal()) {
3516 if (!TII->isLoadFromStackSlot(Def, FI))
3519 unsigned Opcode = Def->getOpcode();
3520 if ((Opcode == X86::LEA32r || Opcode == X86::LEA64r ||
3521 Opcode == X86::LEA64_32r) &&
3522 Def->getOperand(1).isFI()) {
3523 FI = Def->getOperand(1).getIndex();
3524 Bytes = Flags.getByValSize();
3528 } else if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Arg)) {
3529 if (Flags.isByVal())
3530 // ByVal argument is passed in as a pointer but it's now being
3531 // dereferenced. e.g.
3532 // define @foo(%struct.X* %A) {
3533 // tail call @bar(%struct.X* byval %A)
3536 SDValue Ptr = Ld->getBasePtr();
3537 FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr);
3540 FI = FINode->getIndex();
3541 } else if (Arg.getOpcode() == ISD::FrameIndex && Flags.isByVal()) {
3542 FrameIndexSDNode *FINode = cast<FrameIndexSDNode>(Arg);
3543 FI = FINode->getIndex();
3544 Bytes = Flags.getByValSize();
3548 assert(FI != INT_MAX);
3549 if (!MFI->isFixedObjectIndex(FI))
3551 return Offset == MFI->getObjectOffset(FI) && Bytes == MFI->getObjectSize(FI);
3554 /// Check whether the call is eligible for tail call optimization. Targets
3555 /// that want to do tail call optimization should implement this function.
3557 X86TargetLowering::IsEligibleForTailCallOptimization(SDValue Callee,
3558 CallingConv::ID CalleeCC,
3560 bool isCalleeStructRet,
3561 bool isCallerStructRet,
3563 const SmallVectorImpl<ISD::OutputArg> &Outs,
3564 const SmallVectorImpl<SDValue> &OutVals,
3565 const SmallVectorImpl<ISD::InputArg> &Ins,
3566 SelectionDAG &DAG) const {
3567 if (!IsTailCallConvention(CalleeCC) && !IsCCallConvention(CalleeCC))
3570 // If -tailcallopt is specified, make fastcc functions tail-callable.
3571 const MachineFunction &MF = DAG.getMachineFunction();
3572 const Function *CallerF = MF.getFunction();
3574 // If the function return type is x86_fp80 and the callee return type is not,
3575 // then the FP_EXTEND of the call result is not a nop. It's not safe to
3576 // perform a tailcall optimization here.
3577 if (CallerF->getReturnType()->isX86_FP80Ty() && !RetTy->isX86_FP80Ty())
3580 CallingConv::ID CallerCC = CallerF->getCallingConv();
3581 bool CCMatch = CallerCC == CalleeCC;
3582 bool IsCalleeWin64 = Subtarget->isCallingConvWin64(CalleeCC);
3583 bool IsCallerWin64 = Subtarget->isCallingConvWin64(CallerCC);
3585 // Win64 functions have extra shadow space for argument homing. Don't do the
3586 // sibcall if the caller and callee have mismatched expectations for this
3588 if (IsCalleeWin64 != IsCallerWin64)
3591 if (DAG.getTarget().Options.GuaranteedTailCallOpt) {
3592 if (IsTailCallConvention(CalleeCC) && CCMatch)
3597 // Look for obvious safe cases to perform tail call optimization that do not
3598 // require ABI changes. This is what gcc calls sibcall.
3600 // Can't do sibcall if stack needs to be dynamically re-aligned. PEI needs to
3601 // emit a special epilogue.
3602 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
3603 if (RegInfo->needsStackRealignment(MF))
3606 // Also avoid sibcall optimization if either caller or callee uses struct
3607 // return semantics.
3608 if (isCalleeStructRet || isCallerStructRet)
3611 // An stdcall/thiscall caller is expected to clean up its arguments; the
3612 // callee isn't going to do that.
3613 // FIXME: this is more restrictive than needed. We could produce a tailcall
3614 // when the stack adjustment matches. For example, with a thiscall that takes
3615 // only one argument.
3616 if (!CCMatch && (CallerCC == CallingConv::X86_StdCall ||
3617 CallerCC == CallingConv::X86_ThisCall))
3620 // Do not sibcall optimize vararg calls unless all arguments are passed via
3622 if (isVarArg && !Outs.empty()) {
3624 // Optimizing for varargs on Win64 is unlikely to be safe without
3625 // additional testing.
3626 if (IsCalleeWin64 || IsCallerWin64)
3629 SmallVector<CCValAssign, 16> ArgLocs;
3630 CCState CCInfo(CalleeCC, isVarArg, DAG.getMachineFunction(), ArgLocs,
3633 CCInfo.AnalyzeCallOperands(Outs, CC_X86);
3634 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i)
3635 if (!ArgLocs[i].isRegLoc())
3639 // If the call result is in ST0 / ST1, it needs to be popped off the x87
3640 // stack. Therefore, if it's not used by the call it is not safe to optimize
3641 // this into a sibcall.
3642 bool Unused = false;
3643 for (unsigned i = 0, e = Ins.size(); i != e; ++i) {
3650 SmallVector<CCValAssign, 16> RVLocs;
3651 CCState CCInfo(CalleeCC, false, DAG.getMachineFunction(), RVLocs,
3653 CCInfo.AnalyzeCallResult(Ins, RetCC_X86);
3654 for (unsigned i = 0, e = RVLocs.size(); i != e; ++i) {
3655 CCValAssign &VA = RVLocs[i];
3656 if (VA.getLocReg() == X86::FP0 || VA.getLocReg() == X86::FP1)
3661 // If the calling conventions do not match, then we'd better make sure the
3662 // results are returned in the same way as what the caller expects.
3664 SmallVector<CCValAssign, 16> RVLocs1;
3665 CCState CCInfo1(CalleeCC, false, DAG.getMachineFunction(), RVLocs1,
3667 CCInfo1.AnalyzeCallResult(Ins, RetCC_X86);
3669 SmallVector<CCValAssign, 16> RVLocs2;
3670 CCState CCInfo2(CallerCC, false, DAG.getMachineFunction(), RVLocs2,
3672 CCInfo2.AnalyzeCallResult(Ins, RetCC_X86);
3674 if (RVLocs1.size() != RVLocs2.size())
3676 for (unsigned i = 0, e = RVLocs1.size(); i != e; ++i) {
3677 if (RVLocs1[i].isRegLoc() != RVLocs2[i].isRegLoc())
3679 if (RVLocs1[i].getLocInfo() != RVLocs2[i].getLocInfo())
3681 if (RVLocs1[i].isRegLoc()) {
3682 if (RVLocs1[i].getLocReg() != RVLocs2[i].getLocReg())
3685 if (RVLocs1[i].getLocMemOffset() != RVLocs2[i].getLocMemOffset())
3691 // If the callee takes no arguments then go on to check the results of the
3693 if (!Outs.empty()) {
3694 // Check if stack adjustment is needed. For now, do not do this if any
3695 // argument is passed on the stack.
3696 SmallVector<CCValAssign, 16> ArgLocs;
3697 CCState CCInfo(CalleeCC, isVarArg, DAG.getMachineFunction(), ArgLocs,
3700 // Allocate shadow area for Win64
3702 CCInfo.AllocateStack(32, 8);
3704 CCInfo.AnalyzeCallOperands(Outs, CC_X86);
3705 if (CCInfo.getNextStackOffset()) {
3706 MachineFunction &MF = DAG.getMachineFunction();
3707 if (MF.getInfo<X86MachineFunctionInfo>()->getBytesToPopOnReturn())
3710 // Check if the arguments are already laid out in the right way as
3711 // the caller's fixed stack objects.
3712 MachineFrameInfo *MFI = MF.getFrameInfo();
3713 const MachineRegisterInfo *MRI = &MF.getRegInfo();
3714 const X86InstrInfo *TII = Subtarget->getInstrInfo();
3715 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
3716 CCValAssign &VA = ArgLocs[i];
3717 SDValue Arg = OutVals[i];
3718 ISD::ArgFlagsTy Flags = Outs[i].Flags;
3719 if (VA.getLocInfo() == CCValAssign::Indirect)
3721 if (!VA.isRegLoc()) {
3722 if (!MatchingStackOffset(Arg, VA.getLocMemOffset(), Flags,
3729 // If the tailcall address may be in a register, then make sure it's
3730 // possible to register allocate for it. In 32-bit, the call address can
3731 // only target EAX, EDX, or ECX since the tail call must be scheduled after
3732 // callee-saved registers are restored. These happen to be the same
3733 // registers used to pass 'inreg' arguments so watch out for those.
3734 if (!Subtarget->is64Bit() &&
3735 ((!isa<GlobalAddressSDNode>(Callee) &&
3736 !isa<ExternalSymbolSDNode>(Callee)) ||
3737 DAG.getTarget().getRelocationModel() == Reloc::PIC_)) {
3738 unsigned NumInRegs = 0;
3739 // In PIC we need an extra register to formulate the address computation
3741 unsigned MaxInRegs =
3742 (DAG.getTarget().getRelocationModel() == Reloc::PIC_) ? 2 : 3;
3744 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
3745 CCValAssign &VA = ArgLocs[i];
3748 unsigned Reg = VA.getLocReg();
3751 case X86::EAX: case X86::EDX: case X86::ECX:
3752 if (++NumInRegs == MaxInRegs)
3764 X86TargetLowering::createFastISel(FunctionLoweringInfo &funcInfo,
3765 const TargetLibraryInfo *libInfo) const {
3766 return X86::createFastISel(funcInfo, libInfo);
3769 //===----------------------------------------------------------------------===//
3770 // Other Lowering Hooks
3771 //===----------------------------------------------------------------------===//
3773 static bool MayFoldLoad(SDValue Op) {
3774 return Op.hasOneUse() && ISD::isNormalLoad(Op.getNode());
3777 static bool MayFoldIntoStore(SDValue Op) {
3778 return Op.hasOneUse() && ISD::isNormalStore(*Op.getNode()->use_begin());
3781 static bool isTargetShuffle(unsigned Opcode) {
3783 default: return false;
3784 case X86ISD::BLENDI:
3785 case X86ISD::PSHUFB:
3786 case X86ISD::PSHUFD:
3787 case X86ISD::PSHUFHW:
3788 case X86ISD::PSHUFLW:
3790 case X86ISD::PALIGNR:
3791 case X86ISD::MOVLHPS:
3792 case X86ISD::MOVLHPD:
3793 case X86ISD::MOVHLPS:
3794 case X86ISD::MOVLPS:
3795 case X86ISD::MOVLPD:
3796 case X86ISD::MOVSHDUP:
3797 case X86ISD::MOVSLDUP:
3798 case X86ISD::MOVDDUP:
3801 case X86ISD::UNPCKL:
3802 case X86ISD::UNPCKH:
3803 case X86ISD::VPERMILPI:
3804 case X86ISD::VPERM2X128:
3805 case X86ISD::VPERMI:
3806 case X86ISD::VPERMV:
3807 case X86ISD::VPERMV3:
3812 static SDValue getTargetShuffleNode(unsigned Opc, SDLoc dl, EVT VT,
3813 SDValue V1, unsigned TargetMask,
3814 SelectionDAG &DAG) {
3816 default: llvm_unreachable("Unknown x86 shuffle node");
3817 case X86ISD::PSHUFD:
3818 case X86ISD::PSHUFHW:
3819 case X86ISD::PSHUFLW:
3820 case X86ISD::VPERMILPI:
3821 case X86ISD::VPERMI:
3822 return DAG.getNode(Opc, dl, VT, V1,
3823 DAG.getConstant(TargetMask, dl, MVT::i8));
3827 static SDValue getTargetShuffleNode(unsigned Opc, SDLoc dl, EVT VT,
3828 SDValue V1, SDValue V2, SelectionDAG &DAG) {
3830 default: llvm_unreachable("Unknown x86 shuffle node");
3831 case X86ISD::MOVLHPS:
3832 case X86ISD::MOVLHPD:
3833 case X86ISD::MOVHLPS:
3834 case X86ISD::MOVLPS:
3835 case X86ISD::MOVLPD:
3838 case X86ISD::UNPCKL:
3839 case X86ISD::UNPCKH:
3840 return DAG.getNode(Opc, dl, VT, V1, V2);
3844 SDValue X86TargetLowering::getReturnAddressFrameIndex(SelectionDAG &DAG) const {
3845 MachineFunction &MF = DAG.getMachineFunction();
3846 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
3847 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
3848 int ReturnAddrIndex = FuncInfo->getRAIndex();
3850 if (ReturnAddrIndex == 0) {
3851 // Set up a frame object for the return address.
3852 unsigned SlotSize = RegInfo->getSlotSize();
3853 ReturnAddrIndex = MF.getFrameInfo()->CreateFixedObject(SlotSize,
3856 FuncInfo->setRAIndex(ReturnAddrIndex);
3859 return DAG.getFrameIndex(ReturnAddrIndex, getPointerTy(DAG.getDataLayout()));
3862 bool X86::isOffsetSuitableForCodeModel(int64_t Offset, CodeModel::Model M,
3863 bool hasSymbolicDisplacement) {
3864 // Offset should fit into 32 bit immediate field.
3865 if (!isInt<32>(Offset))
3868 // If we don't have a symbolic displacement - we don't have any extra
3870 if (!hasSymbolicDisplacement)
3873 // FIXME: Some tweaks might be needed for medium code model.
3874 if (M != CodeModel::Small && M != CodeModel::Kernel)
3877 // For small code model we assume that latest object is 16MB before end of 31
3878 // bits boundary. We may also accept pretty large negative constants knowing
3879 // that all objects are in the positive half of address space.
3880 if (M == CodeModel::Small && Offset < 16*1024*1024)
3883 // For kernel code model we know that all object resist in the negative half
3884 // of 32bits address space. We may not accept negative offsets, since they may
3885 // be just off and we may accept pretty large positive ones.
3886 if (M == CodeModel::Kernel && Offset >= 0)
3892 /// Determines whether the callee is required to pop its own arguments.
3893 /// Callee pop is necessary to support tail calls.
3894 bool X86::isCalleePop(CallingConv::ID CallingConv,
3895 bool is64Bit, bool IsVarArg, bool TailCallOpt) {
3896 switch (CallingConv) {
3899 case CallingConv::X86_StdCall:
3900 case CallingConv::X86_FastCall:
3901 case CallingConv::X86_ThisCall:
3903 case CallingConv::Fast:
3904 case CallingConv::GHC:
3905 case CallingConv::HiPE:
3912 /// \brief Return true if the condition is an unsigned comparison operation.
3913 static bool isX86CCUnsigned(unsigned X86CC) {
3915 default: llvm_unreachable("Invalid integer condition!");
3916 case X86::COND_E: return true;
3917 case X86::COND_G: return false;
3918 case X86::COND_GE: return false;
3919 case X86::COND_L: return false;
3920 case X86::COND_LE: return false;
3921 case X86::COND_NE: return true;
3922 case X86::COND_B: return true;
3923 case X86::COND_A: return true;
3924 case X86::COND_BE: return true;
3925 case X86::COND_AE: return true;
3927 llvm_unreachable("covered switch fell through?!");
3930 /// Do a one-to-one translation of a ISD::CondCode to the X86-specific
3931 /// condition code, returning the condition code and the LHS/RHS of the
3932 /// comparison to make.
3933 static unsigned TranslateX86CC(ISD::CondCode SetCCOpcode, SDLoc DL, bool isFP,
3934 SDValue &LHS, SDValue &RHS, SelectionDAG &DAG) {
3936 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS)) {
3937 if (SetCCOpcode == ISD::SETGT && RHSC->isAllOnesValue()) {
3938 // X > -1 -> X == 0, jump !sign.
3939 RHS = DAG.getConstant(0, DL, RHS.getValueType());
3940 return X86::COND_NS;
3942 if (SetCCOpcode == ISD::SETLT && RHSC->isNullValue()) {
3943 // X < 0 -> X == 0, jump on sign.
3946 if (SetCCOpcode == ISD::SETLT && RHSC->getZExtValue() == 1) {
3948 RHS = DAG.getConstant(0, DL, RHS.getValueType());
3949 return X86::COND_LE;
3953 switch (SetCCOpcode) {
3954 default: llvm_unreachable("Invalid integer condition!");
3955 case ISD::SETEQ: return X86::COND_E;
3956 case ISD::SETGT: return X86::COND_G;
3957 case ISD::SETGE: return X86::COND_GE;
3958 case ISD::SETLT: return X86::COND_L;
3959 case ISD::SETLE: return X86::COND_LE;
3960 case ISD::SETNE: return X86::COND_NE;
3961 case ISD::SETULT: return X86::COND_B;
3962 case ISD::SETUGT: return X86::COND_A;
3963 case ISD::SETULE: return X86::COND_BE;
3964 case ISD::SETUGE: return X86::COND_AE;
3968 // First determine if it is required or is profitable to flip the operands.
3970 // If LHS is a foldable load, but RHS is not, flip the condition.
3971 if (ISD::isNON_EXTLoad(LHS.getNode()) &&
3972 !ISD::isNON_EXTLoad(RHS.getNode())) {
3973 SetCCOpcode = getSetCCSwappedOperands(SetCCOpcode);
3974 std::swap(LHS, RHS);
3977 switch (SetCCOpcode) {
3983 std::swap(LHS, RHS);
3987 // On a floating point condition, the flags are set as follows:
3989 // 0 | 0 | 0 | X > Y
3990 // 0 | 0 | 1 | X < Y
3991 // 1 | 0 | 0 | X == Y
3992 // 1 | 1 | 1 | unordered
3993 switch (SetCCOpcode) {
3994 default: llvm_unreachable("Condcode should be pre-legalized away");
3996 case ISD::SETEQ: return X86::COND_E;
3997 case ISD::SETOLT: // flipped
3999 case ISD::SETGT: return X86::COND_A;
4000 case ISD::SETOLE: // flipped
4002 case ISD::SETGE: return X86::COND_AE;
4003 case ISD::SETUGT: // flipped
4005 case ISD::SETLT: return X86::COND_B;
4006 case ISD::SETUGE: // flipped
4008 case ISD::SETLE: return X86::COND_BE;
4010 case ISD::SETNE: return X86::COND_NE;
4011 case ISD::SETUO: return X86::COND_P;
4012 case ISD::SETO: return X86::COND_NP;
4014 case ISD::SETUNE: return X86::COND_INVALID;
4018 /// Is there a floating point cmov for the specific X86 condition code?
4019 /// Current x86 isa includes the following FP cmov instructions:
4020 /// fcmovb, fcomvbe, fcomve, fcmovu, fcmovae, fcmova, fcmovne, fcmovnu.
4021 static bool hasFPCMov(unsigned X86CC) {
4037 /// Returns true if the target can instruction select the
4038 /// specified FP immediate natively. If false, the legalizer will
4039 /// materialize the FP immediate as a load from a constant pool.
4040 bool X86TargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
4041 for (unsigned i = 0, e = LegalFPImmediates.size(); i != e; ++i) {
4042 if (Imm.bitwiseIsEqual(LegalFPImmediates[i]))
4048 bool X86TargetLowering::shouldReduceLoadWidth(SDNode *Load,
4049 ISD::LoadExtType ExtTy,
4051 // "ELF Handling for Thread-Local Storage" specifies that R_X86_64_GOTTPOFF
4052 // relocation target a movq or addq instruction: don't let the load shrink.
4053 SDValue BasePtr = cast<LoadSDNode>(Load)->getBasePtr();
4054 if (BasePtr.getOpcode() == X86ISD::WrapperRIP)
4055 if (const auto *GA = dyn_cast<GlobalAddressSDNode>(BasePtr.getOperand(0)))
4056 return GA->getTargetFlags() != X86II::MO_GOTTPOFF;
4060 /// \brief Returns true if it is beneficial to convert a load of a constant
4061 /// to just the constant itself.
4062 bool X86TargetLowering::shouldConvertConstantLoadToIntImm(const APInt &Imm,
4064 assert(Ty->isIntegerTy());
4066 unsigned BitSize = Ty->getPrimitiveSizeInBits();
4067 if (BitSize == 0 || BitSize > 64)
4072 bool X86TargetLowering::isExtractSubvectorCheap(EVT ResVT,
4073 unsigned Index) const {
4074 if (!isOperationLegalOrCustom(ISD::EXTRACT_SUBVECTOR, ResVT))
4077 return (Index == 0 || Index == ResVT.getVectorNumElements());
4080 bool X86TargetLowering::isCheapToSpeculateCttz() const {
4081 // Speculate cttz only if we can directly use TZCNT.
4082 return Subtarget->hasBMI();
4085 bool X86TargetLowering::isCheapToSpeculateCtlz() const {
4086 // Speculate ctlz only if we can directly use LZCNT.
4087 return Subtarget->hasLZCNT();
4090 /// Return true if every element in Mask, beginning
4091 /// from position Pos and ending in Pos+Size is undef.
4092 static bool isUndefInRange(ArrayRef<int> Mask, unsigned Pos, unsigned Size) {
4093 for (unsigned i = Pos, e = Pos + Size; i != e; ++i)
4099 /// Return true if Val is undef or if its value falls within the
4100 /// specified range (L, H].
4101 static bool isUndefOrInRange(int Val, int Low, int Hi) {
4102 return (Val < 0) || (Val >= Low && Val < Hi);
4105 /// Val is either less than zero (undef) or equal to the specified value.
4106 static bool isUndefOrEqual(int Val, int CmpVal) {
4107 return (Val < 0 || Val == CmpVal);
4110 /// Return true if every element in Mask, beginning
4111 /// from position Pos and ending in Pos+Size, falls within the specified
4112 /// sequential range (Low, Low+Size]. or is undef.
4113 static bool isSequentialOrUndefInRange(ArrayRef<int> Mask,
4114 unsigned Pos, unsigned Size, int Low) {
4115 for (unsigned i = Pos, e = Pos+Size; i != e; ++i, ++Low)
4116 if (!isUndefOrEqual(Mask[i], Low))
4121 /// Return true if the specified EXTRACT_SUBVECTOR operand specifies a vector
4122 /// extract that is suitable for instruction that extract 128 or 256 bit vectors
4123 static bool isVEXTRACTIndex(SDNode *N, unsigned vecWidth) {
4124 assert((vecWidth == 128 || vecWidth == 256) && "Unexpected vector width");
4125 if (!isa<ConstantSDNode>(N->getOperand(1).getNode()))
4128 // The index should be aligned on a vecWidth-bit boundary.
4130 cast<ConstantSDNode>(N->getOperand(1).getNode())->getZExtValue();
4132 MVT VT = N->getSimpleValueType(0);
4133 unsigned ElSize = VT.getVectorElementType().getSizeInBits();
4134 bool Result = (Index * ElSize) % vecWidth == 0;
4139 /// Return true if the specified INSERT_SUBVECTOR
4140 /// operand specifies a subvector insert that is suitable for input to
4141 /// insertion of 128 or 256-bit subvectors
4142 static bool isVINSERTIndex(SDNode *N, unsigned vecWidth) {
4143 assert((vecWidth == 128 || vecWidth == 256) && "Unexpected vector width");
4144 if (!isa<ConstantSDNode>(N->getOperand(2).getNode()))
4146 // The index should be aligned on a vecWidth-bit boundary.
4148 cast<ConstantSDNode>(N->getOperand(2).getNode())->getZExtValue();
4150 MVT VT = N->getSimpleValueType(0);
4151 unsigned ElSize = VT.getVectorElementType().getSizeInBits();
4152 bool Result = (Index * ElSize) % vecWidth == 0;
4157 bool X86::isVINSERT128Index(SDNode *N) {
4158 return isVINSERTIndex(N, 128);
4161 bool X86::isVINSERT256Index(SDNode *N) {
4162 return isVINSERTIndex(N, 256);
4165 bool X86::isVEXTRACT128Index(SDNode *N) {
4166 return isVEXTRACTIndex(N, 128);
4169 bool X86::isVEXTRACT256Index(SDNode *N) {
4170 return isVEXTRACTIndex(N, 256);
4173 static unsigned getExtractVEXTRACTImmediate(SDNode *N, unsigned vecWidth) {
4174 assert((vecWidth == 128 || vecWidth == 256) && "Unsupported vector width");
4175 if (!isa<ConstantSDNode>(N->getOperand(1).getNode()))
4176 llvm_unreachable("Illegal extract subvector for VEXTRACT");
4179 cast<ConstantSDNode>(N->getOperand(1).getNode())->getZExtValue();
4181 MVT VecVT = N->getOperand(0).getSimpleValueType();
4182 MVT ElVT = VecVT.getVectorElementType();
4184 unsigned NumElemsPerChunk = vecWidth / ElVT.getSizeInBits();
4185 return Index / NumElemsPerChunk;
4188 static unsigned getInsertVINSERTImmediate(SDNode *N, unsigned vecWidth) {
4189 assert((vecWidth == 128 || vecWidth == 256) && "Unsupported vector width");
4190 if (!isa<ConstantSDNode>(N->getOperand(2).getNode()))
4191 llvm_unreachable("Illegal insert subvector for VINSERT");
4194 cast<ConstantSDNode>(N->getOperand(2).getNode())->getZExtValue();
4196 MVT VecVT = N->getSimpleValueType(0);
4197 MVT ElVT = VecVT.getVectorElementType();
4199 unsigned NumElemsPerChunk = vecWidth / ElVT.getSizeInBits();
4200 return Index / NumElemsPerChunk;
4203 /// Return the appropriate immediate to extract the specified
4204 /// EXTRACT_SUBVECTOR index with VEXTRACTF128 and VINSERTI128 instructions.
4205 unsigned X86::getExtractVEXTRACT128Immediate(SDNode *N) {
4206 return getExtractVEXTRACTImmediate(N, 128);
4209 /// Return the appropriate immediate to extract the specified
4210 /// EXTRACT_SUBVECTOR index with VEXTRACTF64x4 and VINSERTI64x4 instructions.
4211 unsigned X86::getExtractVEXTRACT256Immediate(SDNode *N) {
4212 return getExtractVEXTRACTImmediate(N, 256);
4215 /// Return the appropriate immediate to insert at the specified
4216 /// INSERT_SUBVECTOR index with VINSERTF128 and VINSERTI128 instructions.
4217 unsigned X86::getInsertVINSERT128Immediate(SDNode *N) {
4218 return getInsertVINSERTImmediate(N, 128);
4221 /// Return the appropriate immediate to insert at the specified
4222 /// INSERT_SUBVECTOR index with VINSERTF46x4 and VINSERTI64x4 instructions.
4223 unsigned X86::getInsertVINSERT256Immediate(SDNode *N) {
4224 return getInsertVINSERTImmediate(N, 256);
4227 /// Returns true if Elt is a constant integer zero
4228 static bool isZero(SDValue V) {
4229 ConstantSDNode *C = dyn_cast<ConstantSDNode>(V);
4230 return C && C->isNullValue();
4233 /// Returns true if Elt is a constant zero or a floating point constant +0.0.
4234 bool X86::isZeroNode(SDValue Elt) {
4237 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(Elt))
4238 return CFP->getValueAPF().isPosZero();
4242 /// Returns a vector of specified type with all zero elements.
4243 static SDValue getZeroVector(EVT VT, const X86Subtarget *Subtarget,
4244 SelectionDAG &DAG, SDLoc dl) {
4245 assert(VT.isVector() && "Expected a vector type");
4247 // Always build SSE zero vectors as <4 x i32> bitcasted
4248 // to their dest type. This ensures they get CSE'd.
4250 if (VT.is128BitVector()) { // SSE
4251 if (Subtarget->hasSSE2()) { // SSE2
4252 SDValue Cst = DAG.getConstant(0, dl, MVT::i32);
4253 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
4255 SDValue Cst = DAG.getConstantFP(+0.0, dl, MVT::f32);
4256 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4f32, Cst, Cst, Cst, Cst);
4258 } else if (VT.is256BitVector()) { // AVX
4259 if (Subtarget->hasInt256()) { // AVX2
4260 SDValue Cst = DAG.getConstant(0, dl, MVT::i32);
4261 SDValue Ops[] = { Cst, Cst, Cst, Cst, Cst, Cst, Cst, Cst };
4262 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v8i32, Ops);
4264 // 256-bit logic and arithmetic instructions in AVX are all
4265 // floating-point, no support for integer ops. Emit fp zeroed vectors.
4266 SDValue Cst = DAG.getConstantFP(+0.0, dl, MVT::f32);
4267 SDValue Ops[] = { Cst, Cst, Cst, Cst, Cst, Cst, Cst, Cst };
4268 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v8f32, Ops);
4270 } else if (VT.is512BitVector()) { // AVX-512
4271 SDValue Cst = DAG.getConstant(0, dl, MVT::i32);
4272 SDValue Ops[] = { Cst, Cst, Cst, Cst, Cst, Cst, Cst, Cst,
4273 Cst, Cst, Cst, Cst, Cst, Cst, Cst, Cst };
4274 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v16i32, Ops);
4275 } else if (VT.getScalarType() == MVT::i1) {
4277 assert((Subtarget->hasBWI() || VT.getVectorNumElements() <= 16)
4278 && "Unexpected vector type");
4279 assert((Subtarget->hasVLX() || VT.getVectorNumElements() >= 8)
4280 && "Unexpected vector type");
4281 SDValue Cst = DAG.getConstant(0, dl, MVT::i1);
4282 SmallVector<SDValue, 64> Ops(VT.getVectorNumElements(), Cst);
4283 return DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Ops);
4285 llvm_unreachable("Unexpected vector type");
4287 return DAG.getBitcast(VT, Vec);
4290 static SDValue ExtractSubVector(SDValue Vec, unsigned IdxVal,
4291 SelectionDAG &DAG, SDLoc dl,
4292 unsigned vectorWidth) {
4293 assert((vectorWidth == 128 || vectorWidth == 256) &&
4294 "Unsupported vector width");
4295 EVT VT = Vec.getValueType();
4296 EVT ElVT = VT.getVectorElementType();
4297 unsigned Factor = VT.getSizeInBits()/vectorWidth;
4298 EVT ResultVT = EVT::getVectorVT(*DAG.getContext(), ElVT,
4299 VT.getVectorNumElements()/Factor);
4301 // Extract from UNDEF is UNDEF.
4302 if (Vec.getOpcode() == ISD::UNDEF)
4303 return DAG.getUNDEF(ResultVT);
4305 // Extract the relevant vectorWidth bits. Generate an EXTRACT_SUBVECTOR
4306 unsigned ElemsPerChunk = vectorWidth / ElVT.getSizeInBits();
4308 // This is the index of the first element of the vectorWidth-bit chunk
4310 unsigned NormalizedIdxVal = (((IdxVal * ElVT.getSizeInBits()) / vectorWidth)
4313 // If the input is a buildvector just emit a smaller one.
4314 if (Vec.getOpcode() == ISD::BUILD_VECTOR)
4315 return DAG.getNode(ISD::BUILD_VECTOR, dl, ResultVT,
4316 makeArrayRef(Vec->op_begin() + NormalizedIdxVal,
4319 SDValue VecIdx = DAG.getIntPtrConstant(NormalizedIdxVal, dl);
4320 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, ResultVT, Vec, VecIdx);
4323 /// Generate a DAG to grab 128-bits from a vector > 128 bits. This
4324 /// sets things up to match to an AVX VEXTRACTF128 / VEXTRACTI128
4325 /// or AVX-512 VEXTRACTF32x4 / VEXTRACTI32x4
4326 /// instructions or a simple subregister reference. Idx is an index in the
4327 /// 128 bits we want. It need not be aligned to a 128-bit boundary. That makes
4328 /// lowering EXTRACT_VECTOR_ELT operations easier.
4329 static SDValue Extract128BitVector(SDValue Vec, unsigned IdxVal,
4330 SelectionDAG &DAG, SDLoc dl) {
4331 assert((Vec.getValueType().is256BitVector() ||
4332 Vec.getValueType().is512BitVector()) && "Unexpected vector size!");
4333 return ExtractSubVector(Vec, IdxVal, DAG, dl, 128);
4336 /// Generate a DAG to grab 256-bits from a 512-bit vector.
4337 static SDValue Extract256BitVector(SDValue Vec, unsigned IdxVal,
4338 SelectionDAG &DAG, SDLoc dl) {
4339 assert(Vec.getValueType().is512BitVector() && "Unexpected vector size!");
4340 return ExtractSubVector(Vec, IdxVal, DAG, dl, 256);
4343 static SDValue InsertSubVector(SDValue Result, SDValue Vec,
4344 unsigned IdxVal, SelectionDAG &DAG,
4345 SDLoc dl, unsigned vectorWidth) {
4346 assert((vectorWidth == 128 || vectorWidth == 256) &&
4347 "Unsupported vector width");
4348 // Inserting UNDEF is Result
4349 if (Vec.getOpcode() == ISD::UNDEF)
4351 EVT VT = Vec.getValueType();
4352 EVT ElVT = VT.getVectorElementType();
4353 EVT ResultVT = Result.getValueType();
4355 // Insert the relevant vectorWidth bits.
4356 unsigned ElemsPerChunk = vectorWidth/ElVT.getSizeInBits();
4358 // This is the index of the first element of the vectorWidth-bit chunk
4360 unsigned NormalizedIdxVal = (((IdxVal * ElVT.getSizeInBits())/vectorWidth)
4363 SDValue VecIdx = DAG.getIntPtrConstant(NormalizedIdxVal, dl);
4364 return DAG.getNode(ISD::INSERT_SUBVECTOR, dl, ResultVT, Result, Vec, VecIdx);
4367 /// Generate a DAG to put 128-bits into a vector > 128 bits. This
4368 /// sets things up to match to an AVX VINSERTF128/VINSERTI128 or
4369 /// AVX-512 VINSERTF32x4/VINSERTI32x4 instructions or a
4370 /// simple superregister reference. Idx is an index in the 128 bits
4371 /// we want. It need not be aligned to a 128-bit boundary. That makes
4372 /// lowering INSERT_VECTOR_ELT operations easier.
4373 static SDValue Insert128BitVector(SDValue Result, SDValue Vec, unsigned IdxVal,
4374 SelectionDAG &DAG, SDLoc dl) {
4375 assert(Vec.getValueType().is128BitVector() && "Unexpected vector size!");
4377 // For insertion into the zero index (low half) of a 256-bit vector, it is
4378 // more efficient to generate a blend with immediate instead of an insert*128.
4379 // We are still creating an INSERT_SUBVECTOR below with an undef node to
4380 // extend the subvector to the size of the result vector. Make sure that
4381 // we are not recursing on that node by checking for undef here.
4382 if (IdxVal == 0 && Result.getValueType().is256BitVector() &&
4383 Result.getOpcode() != ISD::UNDEF) {
4384 EVT ResultVT = Result.getValueType();
4385 SDValue ZeroIndex = DAG.getIntPtrConstant(0, dl);
4386 SDValue Undef = DAG.getUNDEF(ResultVT);
4387 SDValue Vec256 = DAG.getNode(ISD::INSERT_SUBVECTOR, dl, ResultVT, Undef,
4390 // The blend instruction, and therefore its mask, depend on the data type.
4391 MVT ScalarType = ResultVT.getScalarType().getSimpleVT();
4392 if (ScalarType.isFloatingPoint()) {
4393 // Choose either vblendps (float) or vblendpd (double).
4394 unsigned ScalarSize = ScalarType.getSizeInBits();
4395 assert((ScalarSize == 64 || ScalarSize == 32) && "Unknown float type");
4396 unsigned MaskVal = (ScalarSize == 64) ? 0x03 : 0x0f;
4397 SDValue Mask = DAG.getConstant(MaskVal, dl, MVT::i8);
4398 return DAG.getNode(X86ISD::BLENDI, dl, ResultVT, Result, Vec256, Mask);
4401 const X86Subtarget &Subtarget =
4402 static_cast<const X86Subtarget &>(DAG.getSubtarget());
4404 // AVX2 is needed for 256-bit integer blend support.
4405 // Integers must be cast to 32-bit because there is only vpblendd;
4406 // vpblendw can't be used for this because it has a handicapped mask.
4408 // If we don't have AVX2, then cast to float. Using a wrong domain blend
4409 // is still more efficient than using the wrong domain vinsertf128 that
4410 // will be created by InsertSubVector().
4411 MVT CastVT = Subtarget.hasAVX2() ? MVT::v8i32 : MVT::v8f32;
4413 SDValue Mask = DAG.getConstant(0x0f, dl, MVT::i8);
4414 Vec256 = DAG.getBitcast(CastVT, Vec256);
4415 Vec256 = DAG.getNode(X86ISD::BLENDI, dl, CastVT, Result, Vec256, Mask);
4416 return DAG.getBitcast(ResultVT, Vec256);
4419 return InsertSubVector(Result, Vec, IdxVal, DAG, dl, 128);
4422 static SDValue Insert256BitVector(SDValue Result, SDValue Vec, unsigned IdxVal,
4423 SelectionDAG &DAG, SDLoc dl) {
4424 assert(Vec.getValueType().is256BitVector() && "Unexpected vector size!");
4425 return InsertSubVector(Result, Vec, IdxVal, DAG, dl, 256);
4428 /// Concat two 128-bit vectors into a 256 bit vector using VINSERTF128
4429 /// instructions. This is used because creating CONCAT_VECTOR nodes of
4430 /// BUILD_VECTORS returns a larger BUILD_VECTOR while we're trying to lower
4431 /// large BUILD_VECTORS.
4432 static SDValue Concat128BitVectors(SDValue V1, SDValue V2, EVT VT,
4433 unsigned NumElems, SelectionDAG &DAG,
4435 SDValue V = Insert128BitVector(DAG.getUNDEF(VT), V1, 0, DAG, dl);
4436 return Insert128BitVector(V, V2, NumElems/2, DAG, dl);
4439 static SDValue Concat256BitVectors(SDValue V1, SDValue V2, EVT VT,
4440 unsigned NumElems, SelectionDAG &DAG,
4442 SDValue V = Insert256BitVector(DAG.getUNDEF(VT), V1, 0, DAG, dl);
4443 return Insert256BitVector(V, V2, NumElems/2, DAG, dl);
4446 /// Returns a vector of specified type with all bits set.
4447 /// Always build ones vectors as <4 x i32> or <8 x i32>. For 256-bit types with
4448 /// no AVX2 supprt, use two <4 x i32> inserted in a <8 x i32> appropriately.
4449 /// Then bitcast to their original type, ensuring they get CSE'd.
4450 static SDValue getOnesVector(EVT VT, const X86Subtarget *Subtarget,
4451 SelectionDAG &DAG, SDLoc dl) {
4452 assert(VT.isVector() && "Expected a vector type");
4454 SDValue Cst = DAG.getConstant(~0U, dl, MVT::i32);
4456 if (VT.is512BitVector()) {
4457 SDValue Ops[] = { Cst, Cst, Cst, Cst, Cst, Cst, Cst, Cst,
4458 Cst, Cst, Cst, Cst, Cst, Cst, Cst, Cst };
4459 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v16i32, Ops);
4460 } else if (VT.is256BitVector()) {
4461 if (Subtarget->hasInt256()) { // AVX2
4462 SDValue Ops[] = { Cst, Cst, Cst, Cst, Cst, Cst, Cst, Cst };
4463 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v8i32, Ops);
4465 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
4466 Vec = Concat128BitVectors(Vec, Vec, MVT::v8i32, 8, DAG, dl);
4468 } else if (VT.is128BitVector()) {
4469 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
4471 llvm_unreachable("Unexpected vector type");
4473 return DAG.getBitcast(VT, Vec);
4476 /// Returns a vector_shuffle node for an unpackl operation.
4477 static SDValue getUnpackl(SelectionDAG &DAG, SDLoc dl, MVT VT, SDValue V1,
4479 unsigned NumElems = VT.getVectorNumElements();
4480 SmallVector<int, 8> Mask;
4481 for (unsigned i = 0, e = NumElems/2; i != e; ++i) {
4483 Mask.push_back(i + NumElems);
4485 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
4488 /// Returns a vector_shuffle node for an unpackh operation.
4489 static SDValue getUnpackh(SelectionDAG &DAG, SDLoc dl, MVT VT, SDValue V1,
4491 unsigned NumElems = VT.getVectorNumElements();
4492 SmallVector<int, 8> Mask;
4493 for (unsigned i = 0, Half = NumElems/2; i != Half; ++i) {
4494 Mask.push_back(i + Half);
4495 Mask.push_back(i + NumElems + Half);
4497 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
4500 /// Return a vector_shuffle of the specified vector of zero or undef vector.
4501 /// This produces a shuffle where the low element of V2 is swizzled into the
4502 /// zero/undef vector, landing at element Idx.
4503 /// This produces a shuffle mask like 4,1,2,3 (idx=0) or 0,1,2,4 (idx=3).
4504 static SDValue getShuffleVectorZeroOrUndef(SDValue V2, unsigned Idx,
4506 const X86Subtarget *Subtarget,
4507 SelectionDAG &DAG) {
4508 MVT VT = V2.getSimpleValueType();
4510 ? getZeroVector(VT, Subtarget, DAG, SDLoc(V2)) : DAG.getUNDEF(VT);
4511 unsigned NumElems = VT.getVectorNumElements();
4512 SmallVector<int, 16> MaskVec;
4513 for (unsigned i = 0; i != NumElems; ++i)
4514 // If this is the insertion idx, put the low elt of V2 here.
4515 MaskVec.push_back(i == Idx ? NumElems : i);
4516 return DAG.getVectorShuffle(VT, SDLoc(V2), V1, V2, &MaskVec[0]);
4519 /// Calculates the shuffle mask corresponding to the target-specific opcode.
4520 /// Returns true if the Mask could be calculated. Sets IsUnary to true if only
4521 /// uses one source. Note that this will set IsUnary for shuffles which use a
4522 /// single input multiple times, and in those cases it will
4523 /// adjust the mask to only have indices within that single input.
4524 /// FIXME: Add support for Decode*Mask functions that return SM_SentinelZero.
4525 static bool getTargetShuffleMask(SDNode *N, MVT VT,
4526 SmallVectorImpl<int> &Mask, bool &IsUnary) {
4527 unsigned NumElems = VT.getVectorNumElements();
4531 bool IsFakeUnary = false;
4532 switch(N->getOpcode()) {
4533 case X86ISD::BLENDI:
4534 ImmN = N->getOperand(N->getNumOperands()-1);
4535 DecodeBLENDMask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
4538 ImmN = N->getOperand(N->getNumOperands()-1);
4539 DecodeSHUFPMask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
4540 IsUnary = IsFakeUnary = N->getOperand(0) == N->getOperand(1);
4542 case X86ISD::UNPCKH:
4543 DecodeUNPCKHMask(VT, Mask);
4544 IsUnary = IsFakeUnary = N->getOperand(0) == N->getOperand(1);
4546 case X86ISD::UNPCKL:
4547 DecodeUNPCKLMask(VT, Mask);
4548 IsUnary = IsFakeUnary = N->getOperand(0) == N->getOperand(1);
4550 case X86ISD::MOVHLPS:
4551 DecodeMOVHLPSMask(NumElems, Mask);
4552 IsUnary = IsFakeUnary = N->getOperand(0) == N->getOperand(1);
4554 case X86ISD::MOVLHPS:
4555 DecodeMOVLHPSMask(NumElems, Mask);
4556 IsUnary = IsFakeUnary = N->getOperand(0) == N->getOperand(1);
4558 case X86ISD::PALIGNR:
4559 ImmN = N->getOperand(N->getNumOperands()-1);
4560 DecodePALIGNRMask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
4562 case X86ISD::PSHUFD:
4563 case X86ISD::VPERMILPI:
4564 ImmN = N->getOperand(N->getNumOperands()-1);
4565 DecodePSHUFMask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
4568 case X86ISD::PSHUFHW:
4569 ImmN = N->getOperand(N->getNumOperands()-1);
4570 DecodePSHUFHWMask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
4573 case X86ISD::PSHUFLW:
4574 ImmN = N->getOperand(N->getNumOperands()-1);
4575 DecodePSHUFLWMask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
4578 case X86ISD::PSHUFB: {
4580 SDValue MaskNode = N->getOperand(1);
4581 while (MaskNode->getOpcode() == ISD::BITCAST)
4582 MaskNode = MaskNode->getOperand(0);
4584 if (MaskNode->getOpcode() == ISD::BUILD_VECTOR) {
4585 // If we have a build-vector, then things are easy.
4586 EVT VT = MaskNode.getValueType();
4587 assert(VT.isVector() &&
4588 "Can't produce a non-vector with a build_vector!");
4589 if (!VT.isInteger())
4592 int NumBytesPerElement = VT.getVectorElementType().getSizeInBits() / 8;
4594 SmallVector<uint64_t, 32> RawMask;
4595 for (int i = 0, e = MaskNode->getNumOperands(); i < e; ++i) {
4596 SDValue Op = MaskNode->getOperand(i);
4597 if (Op->getOpcode() == ISD::UNDEF) {
4598 RawMask.push_back((uint64_t)SM_SentinelUndef);
4601 auto *CN = dyn_cast<ConstantSDNode>(Op.getNode());
4604 APInt MaskElement = CN->getAPIntValue();
4606 // We now have to decode the element which could be any integer size and
4607 // extract each byte of it.
4608 for (int j = 0; j < NumBytesPerElement; ++j) {
4609 // Note that this is x86 and so always little endian: the low byte is
4610 // the first byte of the mask.
4611 RawMask.push_back(MaskElement.getLoBits(8).getZExtValue());
4612 MaskElement = MaskElement.lshr(8);
4615 DecodePSHUFBMask(RawMask, Mask);
4619 auto *MaskLoad = dyn_cast<LoadSDNode>(MaskNode);
4623 SDValue Ptr = MaskLoad->getBasePtr();
4624 if (Ptr->getOpcode() == X86ISD::Wrapper ||
4625 Ptr->getOpcode() == X86ISD::WrapperRIP)
4626 Ptr = Ptr->getOperand(0);
4628 auto *MaskCP = dyn_cast<ConstantPoolSDNode>(Ptr);
4629 if (!MaskCP || MaskCP->isMachineConstantPoolEntry())
4632 if (auto *C = dyn_cast<Constant>(MaskCP->getConstVal())) {
4633 DecodePSHUFBMask(C, Mask);
4641 case X86ISD::VPERMI:
4642 ImmN = N->getOperand(N->getNumOperands()-1);
4643 DecodeVPERMMask(cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
4648 DecodeScalarMoveMask(VT, /* IsLoad */ false, Mask);
4650 case X86ISD::VPERM2X128:
4651 ImmN = N->getOperand(N->getNumOperands()-1);
4652 DecodeVPERM2X128Mask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
4653 if (Mask.empty()) return false;
4654 // Mask only contains negative index if an element is zero.
4655 if (std::any_of(Mask.begin(), Mask.end(),
4656 [](int M){ return M == SM_SentinelZero; }))
4659 case X86ISD::MOVSLDUP:
4660 DecodeMOVSLDUPMask(VT, Mask);
4663 case X86ISD::MOVSHDUP:
4664 DecodeMOVSHDUPMask(VT, Mask);
4667 case X86ISD::MOVDDUP:
4668 DecodeMOVDDUPMask(VT, Mask);
4671 case X86ISD::MOVLHPD:
4672 case X86ISD::MOVLPD:
4673 case X86ISD::MOVLPS:
4674 // Not yet implemented
4676 case X86ISD::VPERMV: {
4678 SDValue MaskNode = N->getOperand(0);
4679 while (MaskNode->getOpcode() == ISD::BITCAST)
4680 MaskNode = MaskNode->getOperand(0);
4682 unsigned MaskLoBits = Log2_64(VT.getVectorNumElements());
4683 SmallVector<uint64_t, 32> RawMask;
4684 if (MaskNode->getOpcode() == ISD::BUILD_VECTOR) {
4685 // If we have a build-vector, then things are easy.
4686 assert(MaskNode.getValueType().isInteger() &&
4687 MaskNode.getValueType().getVectorNumElements() ==
4688 VT.getVectorNumElements());
4690 for (unsigned i = 0; i < MaskNode->getNumOperands(); ++i) {
4691 SDValue Op = MaskNode->getOperand(i);
4692 if (Op->getOpcode() == ISD::UNDEF)
4693 RawMask.push_back((uint64_t)SM_SentinelUndef);
4694 else if (isa<ConstantSDNode>(Op)) {
4695 APInt MaskElement = cast<ConstantSDNode>(Op)->getAPIntValue();
4696 RawMask.push_back(MaskElement.getLoBits(MaskLoBits).getZExtValue());
4700 DecodeVPERMVMask(RawMask, Mask);
4703 if (MaskNode->getOpcode() == X86ISD::VBROADCAST) {
4704 unsigned NumEltsInMask = MaskNode->getNumOperands();
4705 MaskNode = MaskNode->getOperand(0);
4706 auto *CN = dyn_cast<ConstantSDNode>(MaskNode);
4708 APInt MaskEltValue = CN->getAPIntValue();
4709 for (unsigned i = 0; i < NumEltsInMask; ++i)
4710 RawMask.push_back(MaskEltValue.getLoBits(MaskLoBits).getZExtValue());
4711 DecodeVPERMVMask(RawMask, Mask);
4714 // It may be a scalar load
4717 auto *MaskLoad = dyn_cast<LoadSDNode>(MaskNode);
4721 SDValue Ptr = MaskLoad->getBasePtr();
4722 if (Ptr->getOpcode() == X86ISD::Wrapper ||
4723 Ptr->getOpcode() == X86ISD::WrapperRIP)
4724 Ptr = Ptr->getOperand(0);
4726 auto *MaskCP = dyn_cast<ConstantPoolSDNode>(Ptr);
4727 if (!MaskCP || MaskCP->isMachineConstantPoolEntry())
4730 auto *C = dyn_cast<Constant>(MaskCP->getConstVal());
4732 DecodeVPERMVMask(C, VT, Mask);
4739 case X86ISD::VPERMV3: {
4741 SDValue MaskNode = N->getOperand(1);
4742 while (MaskNode->getOpcode() == ISD::BITCAST)
4743 MaskNode = MaskNode->getOperand(1);
4745 if (MaskNode->getOpcode() == ISD::BUILD_VECTOR) {
4746 // If we have a build-vector, then things are easy.
4747 assert(MaskNode.getValueType().isInteger() &&
4748 MaskNode.getValueType().getVectorNumElements() ==
4749 VT.getVectorNumElements());
4751 SmallVector<uint64_t, 32> RawMask;
4752 unsigned MaskLoBits = Log2_64(VT.getVectorNumElements()*2);
4754 for (unsigned i = 0; i < MaskNode->getNumOperands(); ++i) {
4755 SDValue Op = MaskNode->getOperand(i);
4756 if (Op->getOpcode() == ISD::UNDEF)
4757 RawMask.push_back((uint64_t)SM_SentinelUndef);
4759 auto *CN = dyn_cast<ConstantSDNode>(Op.getNode());
4762 APInt MaskElement = CN->getAPIntValue();
4763 RawMask.push_back(MaskElement.getLoBits(MaskLoBits).getZExtValue());
4766 DecodeVPERMV3Mask(RawMask, Mask);
4770 auto *MaskLoad = dyn_cast<LoadSDNode>(MaskNode);
4774 SDValue Ptr = MaskLoad->getBasePtr();
4775 if (Ptr->getOpcode() == X86ISD::Wrapper ||
4776 Ptr->getOpcode() == X86ISD::WrapperRIP)
4777 Ptr = Ptr->getOperand(0);
4779 auto *MaskCP = dyn_cast<ConstantPoolSDNode>(Ptr);
4780 if (!MaskCP || MaskCP->isMachineConstantPoolEntry())
4783 auto *C = dyn_cast<Constant>(MaskCP->getConstVal());
4785 DecodeVPERMV3Mask(C, VT, Mask);
4792 default: llvm_unreachable("unknown target shuffle node");
4795 // If we have a fake unary shuffle, the shuffle mask is spread across two
4796 // inputs that are actually the same node. Re-map the mask to always point
4797 // into the first input.
4800 if (M >= (int)Mask.size())
4806 /// Returns the scalar element that will make up the ith
4807 /// element of the result of the vector shuffle.
4808 static SDValue getShuffleScalarElt(SDNode *N, unsigned Index, SelectionDAG &DAG,
4811 return SDValue(); // Limit search depth.
4813 SDValue V = SDValue(N, 0);
4814 EVT VT = V.getValueType();
4815 unsigned Opcode = V.getOpcode();
4817 // Recurse into ISD::VECTOR_SHUFFLE node to find scalars.
4818 if (const ShuffleVectorSDNode *SV = dyn_cast<ShuffleVectorSDNode>(N)) {
4819 int Elt = SV->getMaskElt(Index);
4822 return DAG.getUNDEF(VT.getVectorElementType());
4824 unsigned NumElems = VT.getVectorNumElements();
4825 SDValue NewV = (Elt < (int)NumElems) ? SV->getOperand(0)
4826 : SV->getOperand(1);
4827 return getShuffleScalarElt(NewV.getNode(), Elt % NumElems, DAG, Depth+1);
4830 // Recurse into target specific vector shuffles to find scalars.
4831 if (isTargetShuffle(Opcode)) {
4832 MVT ShufVT = V.getSimpleValueType();
4833 unsigned NumElems = ShufVT.getVectorNumElements();
4834 SmallVector<int, 16> ShuffleMask;
4837 if (!getTargetShuffleMask(N, ShufVT, ShuffleMask, IsUnary))
4840 int Elt = ShuffleMask[Index];
4842 return DAG.getUNDEF(ShufVT.getVectorElementType());
4844 SDValue NewV = (Elt < (int)NumElems) ? N->getOperand(0)
4846 return getShuffleScalarElt(NewV.getNode(), Elt % NumElems, DAG,
4850 // Actual nodes that may contain scalar elements
4851 if (Opcode == ISD::BITCAST) {
4852 V = V.getOperand(0);
4853 EVT SrcVT = V.getValueType();
4854 unsigned NumElems = VT.getVectorNumElements();
4856 if (!SrcVT.isVector() || SrcVT.getVectorNumElements() != NumElems)
4860 if (V.getOpcode() == ISD::SCALAR_TO_VECTOR)
4861 return (Index == 0) ? V.getOperand(0)
4862 : DAG.getUNDEF(VT.getVectorElementType());
4864 if (V.getOpcode() == ISD::BUILD_VECTOR)
4865 return V.getOperand(Index);
4870 /// Custom lower build_vector of v16i8.
4871 static SDValue LowerBuildVectorv16i8(SDValue Op, unsigned NonZeros,
4872 unsigned NumNonZero, unsigned NumZero,
4874 const X86Subtarget* Subtarget,
4875 const TargetLowering &TLI) {
4883 // SSE4.1 - use PINSRB to insert each byte directly.
4884 if (Subtarget->hasSSE41()) {
4885 for (unsigned i = 0; i < 16; ++i) {
4886 bool isNonZero = (NonZeros & (1 << i)) != 0;
4890 V = getZeroVector(MVT::v16i8, Subtarget, DAG, dl);
4892 V = DAG.getUNDEF(MVT::v16i8);
4895 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl,
4896 MVT::v16i8, V, Op.getOperand(i),
4897 DAG.getIntPtrConstant(i, dl));
4904 // Pre-SSE4.1 - merge byte pairs and insert with PINSRW.
4905 for (unsigned i = 0; i < 16; ++i) {
4906 bool ThisIsNonZero = (NonZeros & (1 << i)) != 0;
4907 if (ThisIsNonZero && First) {
4909 V = getZeroVector(MVT::v8i16, Subtarget, DAG, dl);
4911 V = DAG.getUNDEF(MVT::v8i16);
4916 SDValue ThisElt, LastElt;
4917 bool LastIsNonZero = (NonZeros & (1 << (i-1))) != 0;
4918 if (LastIsNonZero) {
4919 LastElt = DAG.getNode(ISD::ZERO_EXTEND, dl,
4920 MVT::i16, Op.getOperand(i-1));
4922 if (ThisIsNonZero) {
4923 ThisElt = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i16, Op.getOperand(i));
4924 ThisElt = DAG.getNode(ISD::SHL, dl, MVT::i16,
4925 ThisElt, DAG.getConstant(8, dl, MVT::i8));
4927 ThisElt = DAG.getNode(ISD::OR, dl, MVT::i16, ThisElt, LastElt);
4931 if (ThisElt.getNode())
4932 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, V, ThisElt,
4933 DAG.getIntPtrConstant(i/2, dl));
4937 return DAG.getBitcast(MVT::v16i8, V);
4940 /// Custom lower build_vector of v8i16.
4941 static SDValue LowerBuildVectorv8i16(SDValue Op, unsigned NonZeros,
4942 unsigned NumNonZero, unsigned NumZero,
4944 const X86Subtarget* Subtarget,
4945 const TargetLowering &TLI) {
4952 for (unsigned i = 0; i < 8; ++i) {
4953 bool isNonZero = (NonZeros & (1 << i)) != 0;
4957 V = getZeroVector(MVT::v8i16, Subtarget, DAG, dl);
4959 V = DAG.getUNDEF(MVT::v8i16);
4962 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl,
4963 MVT::v8i16, V, Op.getOperand(i),
4964 DAG.getIntPtrConstant(i, dl));
4971 /// Custom lower build_vector of v4i32 or v4f32.
4972 static SDValue LowerBuildVectorv4x32(SDValue Op, SelectionDAG &DAG,
4973 const X86Subtarget *Subtarget,
4974 const TargetLowering &TLI) {
4975 // Find all zeroable elements.
4976 std::bitset<4> Zeroable;
4977 for (int i=0; i < 4; ++i) {
4978 SDValue Elt = Op->getOperand(i);
4979 Zeroable[i] = (Elt.getOpcode() == ISD::UNDEF || X86::isZeroNode(Elt));
4981 assert(Zeroable.size() - Zeroable.count() > 1 &&
4982 "We expect at least two non-zero elements!");
4984 // We only know how to deal with build_vector nodes where elements are either
4985 // zeroable or extract_vector_elt with constant index.
4986 SDValue FirstNonZero;
4987 unsigned FirstNonZeroIdx;
4988 for (unsigned i=0; i < 4; ++i) {
4991 SDValue Elt = Op->getOperand(i);
4992 if (Elt.getOpcode() != ISD::EXTRACT_VECTOR_ELT ||
4993 !isa<ConstantSDNode>(Elt.getOperand(1)))
4995 // Make sure that this node is extracting from a 128-bit vector.
4996 MVT VT = Elt.getOperand(0).getSimpleValueType();
4997 if (!VT.is128BitVector())
4999 if (!FirstNonZero.getNode()) {
5001 FirstNonZeroIdx = i;
5005 assert(FirstNonZero.getNode() && "Unexpected build vector of all zeros!");
5006 SDValue V1 = FirstNonZero.getOperand(0);
5007 MVT VT = V1.getSimpleValueType();
5009 // See if this build_vector can be lowered as a blend with zero.
5011 unsigned EltMaskIdx, EltIdx;
5013 for (EltIdx = 0; EltIdx < 4; ++EltIdx) {
5014 if (Zeroable[EltIdx]) {
5015 // The zero vector will be on the right hand side.
5016 Mask[EltIdx] = EltIdx+4;
5020 Elt = Op->getOperand(EltIdx);
5021 // By construction, Elt is a EXTRACT_VECTOR_ELT with constant index.
5022 EltMaskIdx = cast<ConstantSDNode>(Elt.getOperand(1))->getZExtValue();
5023 if (Elt.getOperand(0) != V1 || EltMaskIdx != EltIdx)
5025 Mask[EltIdx] = EltIdx;
5029 // Let the shuffle legalizer deal with blend operations.
5030 SDValue VZero = getZeroVector(VT, Subtarget, DAG, SDLoc(Op));
5031 if (V1.getSimpleValueType() != VT)
5032 V1 = DAG.getNode(ISD::BITCAST, SDLoc(V1), VT, V1);
5033 return DAG.getVectorShuffle(VT, SDLoc(V1), V1, VZero, &Mask[0]);
5036 // See if we can lower this build_vector to a INSERTPS.
5037 if (!Subtarget->hasSSE41())
5040 SDValue V2 = Elt.getOperand(0);
5041 if (Elt == FirstNonZero && EltIdx == FirstNonZeroIdx)
5044 bool CanFold = true;
5045 for (unsigned i = EltIdx + 1; i < 4 && CanFold; ++i) {
5049 SDValue Current = Op->getOperand(i);
5050 SDValue SrcVector = Current->getOperand(0);
5053 CanFold = SrcVector == V1 &&
5054 cast<ConstantSDNode>(Current.getOperand(1))->getZExtValue() == i;
5060 assert(V1.getNode() && "Expected at least two non-zero elements!");
5061 if (V1.getSimpleValueType() != MVT::v4f32)
5062 V1 = DAG.getNode(ISD::BITCAST, SDLoc(V1), MVT::v4f32, V1);
5063 if (V2.getSimpleValueType() != MVT::v4f32)
5064 V2 = DAG.getNode(ISD::BITCAST, SDLoc(V2), MVT::v4f32, V2);
5066 // Ok, we can emit an INSERTPS instruction.
5067 unsigned ZMask = Zeroable.to_ulong();
5069 unsigned InsertPSMask = EltMaskIdx << 6 | EltIdx << 4 | ZMask;
5070 assert((InsertPSMask & ~0xFFu) == 0 && "Invalid mask!");
5072 SDValue Result = DAG.getNode(X86ISD::INSERTPS, DL, MVT::v4f32, V1, V2,
5073 DAG.getIntPtrConstant(InsertPSMask, DL));
5074 return DAG.getBitcast(VT, Result);
5077 /// Return a vector logical shift node.
5078 static SDValue getVShift(bool isLeft, EVT VT, SDValue SrcOp,
5079 unsigned NumBits, SelectionDAG &DAG,
5080 const TargetLowering &TLI, SDLoc dl) {
5081 assert(VT.is128BitVector() && "Unknown type for VShift");
5082 MVT ShVT = MVT::v2i64;
5083 unsigned Opc = isLeft ? X86ISD::VSHLDQ : X86ISD::VSRLDQ;
5084 SrcOp = DAG.getBitcast(ShVT, SrcOp);
5085 MVT ScalarShiftTy = TLI.getScalarShiftAmountTy(DAG.getDataLayout(), VT);
5086 assert(NumBits % 8 == 0 && "Only support byte sized shifts");
5087 SDValue ShiftVal = DAG.getConstant(NumBits/8, dl, ScalarShiftTy);
5088 return DAG.getBitcast(VT, DAG.getNode(Opc, dl, ShVT, SrcOp, ShiftVal));
5092 LowerAsSplatVectorLoad(SDValue SrcOp, MVT VT, SDLoc dl, SelectionDAG &DAG) {
5094 // Check if the scalar load can be widened into a vector load. And if
5095 // the address is "base + cst" see if the cst can be "absorbed" into
5096 // the shuffle mask.
5097 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(SrcOp)) {
5098 SDValue Ptr = LD->getBasePtr();
5099 if (!ISD::isNormalLoad(LD) || LD->isVolatile())
5101 EVT PVT = LD->getValueType(0);
5102 if (PVT != MVT::i32 && PVT != MVT::f32)
5107 if (FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr)) {
5108 FI = FINode->getIndex();
5110 } else if (DAG.isBaseWithConstantOffset(Ptr) &&
5111 isa<FrameIndexSDNode>(Ptr.getOperand(0))) {
5112 FI = cast<FrameIndexSDNode>(Ptr.getOperand(0))->getIndex();
5113 Offset = Ptr.getConstantOperandVal(1);
5114 Ptr = Ptr.getOperand(0);
5119 // FIXME: 256-bit vector instructions don't require a strict alignment,
5120 // improve this code to support it better.
5121 unsigned RequiredAlign = VT.getSizeInBits()/8;
5122 SDValue Chain = LD->getChain();
5123 // Make sure the stack object alignment is at least 16 or 32.
5124 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
5125 if (DAG.InferPtrAlignment(Ptr) < RequiredAlign) {
5126 if (MFI->isFixedObjectIndex(FI)) {
5127 // Can't change the alignment. FIXME: It's possible to compute
5128 // the exact stack offset and reference FI + adjust offset instead.
5129 // If someone *really* cares about this. That's the way to implement it.
5132 MFI->setObjectAlignment(FI, RequiredAlign);
5136 // (Offset % 16 or 32) must be multiple of 4. Then address is then
5137 // Ptr + (Offset & ~15).
5140 if ((Offset % RequiredAlign) & 3)
5142 int64_t StartOffset = Offset & ~int64_t(RequiredAlign - 1);
5145 Ptr = DAG.getNode(ISD::ADD, DL, Ptr.getValueType(), Ptr,
5146 DAG.getConstant(StartOffset, DL, Ptr.getValueType()));
5149 int EltNo = (Offset - StartOffset) >> 2;
5150 unsigned NumElems = VT.getVectorNumElements();
5152 EVT NVT = EVT::getVectorVT(*DAG.getContext(), PVT, NumElems);
5153 SDValue V1 = DAG.getLoad(NVT, dl, Chain, Ptr,
5154 LD->getPointerInfo().getWithOffset(StartOffset),
5155 false, false, false, 0);
5157 SmallVector<int, 8> Mask(NumElems, EltNo);
5159 return DAG.getVectorShuffle(NVT, dl, V1, DAG.getUNDEF(NVT), &Mask[0]);
5165 /// Given the initializing elements 'Elts' of a vector of type 'VT', see if the
5166 /// elements can be replaced by a single large load which has the same value as
5167 /// a build_vector or insert_subvector whose loaded operands are 'Elts'.
5169 /// Example: <load i32 *a, load i32 *a+4, undef, undef> -> zextload a
5171 /// FIXME: we'd also like to handle the case where the last elements are zero
5172 /// rather than undef via VZEXT_LOAD, but we do not detect that case today.
5173 /// There's even a handy isZeroNode for that purpose.
5174 static SDValue EltsFromConsecutiveLoads(EVT VT, ArrayRef<SDValue> Elts,
5175 SDLoc &DL, SelectionDAG &DAG,
5176 bool isAfterLegalize) {
5177 unsigned NumElems = Elts.size();
5179 LoadSDNode *LDBase = nullptr;
5180 unsigned LastLoadedElt = -1U;
5182 // For each element in the initializer, see if we've found a load or an undef.
5183 // If we don't find an initial load element, or later load elements are
5184 // non-consecutive, bail out.
5185 for (unsigned i = 0; i < NumElems; ++i) {
5186 SDValue Elt = Elts[i];
5187 // Look through a bitcast.
5188 if (Elt.getNode() && Elt.getOpcode() == ISD::BITCAST)
5189 Elt = Elt.getOperand(0);
5190 if (!Elt.getNode() ||
5191 (Elt.getOpcode() != ISD::UNDEF && !ISD::isNON_EXTLoad(Elt.getNode())))
5194 if (Elt.getNode()->getOpcode() == ISD::UNDEF)
5196 LDBase = cast<LoadSDNode>(Elt.getNode());
5200 if (Elt.getOpcode() == ISD::UNDEF)
5203 LoadSDNode *LD = cast<LoadSDNode>(Elt);
5204 EVT LdVT = Elt.getValueType();
5205 // Each loaded element must be the correct fractional portion of the
5206 // requested vector load.
5207 if (LdVT.getSizeInBits() != VT.getSizeInBits() / NumElems)
5209 if (!DAG.isConsecutiveLoad(LD, LDBase, LdVT.getSizeInBits() / 8, i))
5214 // If we have found an entire vector of loads and undefs, then return a large
5215 // load of the entire vector width starting at the base pointer. If we found
5216 // consecutive loads for the low half, generate a vzext_load node.
5217 if (LastLoadedElt == NumElems - 1) {
5218 assert(LDBase && "Did not find base load for merging consecutive loads");
5219 EVT EltVT = LDBase->getValueType(0);
5220 // Ensure that the input vector size for the merged loads matches the
5221 // cumulative size of the input elements.
5222 if (VT.getSizeInBits() != EltVT.getSizeInBits() * NumElems)
5225 if (isAfterLegalize &&
5226 !DAG.getTargetLoweringInfo().isOperationLegal(ISD::LOAD, VT))
5229 SDValue NewLd = SDValue();
5231 NewLd = DAG.getLoad(VT, DL, LDBase->getChain(), LDBase->getBasePtr(),
5232 LDBase->getPointerInfo(), LDBase->isVolatile(),
5233 LDBase->isNonTemporal(), LDBase->isInvariant(),
5234 LDBase->getAlignment());
5236 if (LDBase->hasAnyUseOfValue(1)) {
5237 SDValue NewChain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other,
5239 SDValue(NewLd.getNode(), 1));
5240 DAG.ReplaceAllUsesOfValueWith(SDValue(LDBase, 1), NewChain);
5241 DAG.UpdateNodeOperands(NewChain.getNode(), SDValue(LDBase, 1),
5242 SDValue(NewLd.getNode(), 1));
5248 //TODO: The code below fires only for for loading the low v2i32 / v2f32
5249 //of a v4i32 / v4f32. It's probably worth generalizing.
5250 EVT EltVT = VT.getVectorElementType();
5251 if (NumElems == 4 && LastLoadedElt == 1 && (EltVT.getSizeInBits() == 32) &&
5252 DAG.getTargetLoweringInfo().isTypeLegal(MVT::v2i64)) {
5253 SDVTList Tys = DAG.getVTList(MVT::v2i64, MVT::Other);
5254 SDValue Ops[] = { LDBase->getChain(), LDBase->getBasePtr() };
5256 DAG.getMemIntrinsicNode(X86ISD::VZEXT_LOAD, DL, Tys, Ops, MVT::i64,
5257 LDBase->getPointerInfo(),
5258 LDBase->getAlignment(),
5259 false/*isVolatile*/, true/*ReadMem*/,
5262 // Make sure the newly-created LOAD is in the same position as LDBase in
5263 // terms of dependency. We create a TokenFactor for LDBase and ResNode, and
5264 // update uses of LDBase's output chain to use the TokenFactor.
5265 if (LDBase->hasAnyUseOfValue(1)) {
5266 SDValue NewChain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other,
5267 SDValue(LDBase, 1), SDValue(ResNode.getNode(), 1));
5268 DAG.ReplaceAllUsesOfValueWith(SDValue(LDBase, 1), NewChain);
5269 DAG.UpdateNodeOperands(NewChain.getNode(), SDValue(LDBase, 1),
5270 SDValue(ResNode.getNode(), 1));
5273 return DAG.getBitcast(VT, ResNode);
5278 /// LowerVectorBroadcast - Attempt to use the vbroadcast instruction
5279 /// to generate a splat value for the following cases:
5280 /// 1. A splat BUILD_VECTOR which uses a single scalar load, or a constant.
5281 /// 2. A splat shuffle which uses a scalar_to_vector node which comes from
5282 /// a scalar load, or a constant.
5283 /// The VBROADCAST node is returned when a pattern is found,
5284 /// or SDValue() otherwise.
5285 static SDValue LowerVectorBroadcast(SDValue Op, const X86Subtarget* Subtarget,
5286 SelectionDAG &DAG) {
5287 // VBROADCAST requires AVX.
5288 // TODO: Splats could be generated for non-AVX CPUs using SSE
5289 // instructions, but there's less potential gain for only 128-bit vectors.
5290 if (!Subtarget->hasAVX())
5293 MVT VT = Op.getSimpleValueType();
5296 assert((VT.is128BitVector() || VT.is256BitVector() || VT.is512BitVector()) &&
5297 "Unsupported vector type for broadcast.");
5302 switch (Op.getOpcode()) {
5304 // Unknown pattern found.
5307 case ISD::BUILD_VECTOR: {
5308 auto *BVOp = cast<BuildVectorSDNode>(Op.getNode());
5309 BitVector UndefElements;
5310 SDValue Splat = BVOp->getSplatValue(&UndefElements);
5312 // We need a splat of a single value to use broadcast, and it doesn't
5313 // make any sense if the value is only in one element of the vector.
5314 if (!Splat || (VT.getVectorNumElements() - UndefElements.count()) <= 1)
5318 ConstSplatVal = (Ld.getOpcode() == ISD::Constant ||
5319 Ld.getOpcode() == ISD::ConstantFP);
5321 // Make sure that all of the users of a non-constant load are from the
5322 // BUILD_VECTOR node.
5323 if (!ConstSplatVal && !BVOp->isOnlyUserOf(Ld.getNode()))
5328 case ISD::VECTOR_SHUFFLE: {
5329 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
5331 // Shuffles must have a splat mask where the first element is
5333 if ((!SVOp->isSplat()) || SVOp->getMaskElt(0) != 0)
5336 SDValue Sc = Op.getOperand(0);
5337 if (Sc.getOpcode() != ISD::SCALAR_TO_VECTOR &&
5338 Sc.getOpcode() != ISD::BUILD_VECTOR) {
5340 if (!Subtarget->hasInt256())
5343 // Use the register form of the broadcast instruction available on AVX2.
5344 if (VT.getSizeInBits() >= 256)
5345 Sc = Extract128BitVector(Sc, 0, DAG, dl);
5346 return DAG.getNode(X86ISD::VBROADCAST, dl, VT, Sc);
5349 Ld = Sc.getOperand(0);
5350 ConstSplatVal = (Ld.getOpcode() == ISD::Constant ||
5351 Ld.getOpcode() == ISD::ConstantFP);
5353 // The scalar_to_vector node and the suspected
5354 // load node must have exactly one user.
5355 // Constants may have multiple users.
5357 // AVX-512 has register version of the broadcast
5358 bool hasRegVer = Subtarget->hasAVX512() && VT.is512BitVector() &&
5359 Ld.getValueType().getSizeInBits() >= 32;
5360 if (!ConstSplatVal && ((!Sc.hasOneUse() || !Ld.hasOneUse()) &&
5367 unsigned ScalarSize = Ld.getValueType().getSizeInBits();
5368 bool IsGE256 = (VT.getSizeInBits() >= 256);
5370 // When optimizing for size, generate up to 5 extra bytes for a broadcast
5371 // instruction to save 8 or more bytes of constant pool data.
5372 // TODO: If multiple splats are generated to load the same constant,
5373 // it may be detrimental to overall size. There needs to be a way to detect
5374 // that condition to know if this is truly a size win.
5375 bool OptForSize = DAG.getMachineFunction().getFunction()->optForSize();
5377 // Handle broadcasting a single constant scalar from the constant pool
5379 // On Sandybridge (no AVX2), it is still better to load a constant vector
5380 // from the constant pool and not to broadcast it from a scalar.
5381 // But override that restriction when optimizing for size.
5382 // TODO: Check if splatting is recommended for other AVX-capable CPUs.
5383 if (ConstSplatVal && (Subtarget->hasAVX2() || OptForSize)) {
5384 EVT CVT = Ld.getValueType();
5385 assert(!CVT.isVector() && "Must not broadcast a vector type");
5387 // Splat f32, i32, v4f64, v4i64 in all cases with AVX2.
5388 // For size optimization, also splat v2f64 and v2i64, and for size opt
5389 // with AVX2, also splat i8 and i16.
5390 // With pattern matching, the VBROADCAST node may become a VMOVDDUP.
5391 if (ScalarSize == 32 || (IsGE256 && ScalarSize == 64) ||
5392 (OptForSize && (ScalarSize == 64 || Subtarget->hasAVX2()))) {
5393 const Constant *C = nullptr;
5394 if (ConstantSDNode *CI = dyn_cast<ConstantSDNode>(Ld))
5395 C = CI->getConstantIntValue();
5396 else if (ConstantFPSDNode *CF = dyn_cast<ConstantFPSDNode>(Ld))
5397 C = CF->getConstantFPValue();
5399 assert(C && "Invalid constant type");
5401 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
5403 DAG.getConstantPool(C, TLI.getPointerTy(DAG.getDataLayout()));
5404 unsigned Alignment = cast<ConstantPoolSDNode>(CP)->getAlignment();
5406 CVT, dl, DAG.getEntryNode(), CP,
5407 MachinePointerInfo::getConstantPool(DAG.getMachineFunction()), false,
5408 false, false, Alignment);
5410 return DAG.getNode(X86ISD::VBROADCAST, dl, VT, Ld);
5414 bool IsLoad = ISD::isNormalLoad(Ld.getNode());
5416 // Handle AVX2 in-register broadcasts.
5417 if (!IsLoad && Subtarget->hasInt256() &&
5418 (ScalarSize == 32 || (IsGE256 && ScalarSize == 64)))
5419 return DAG.getNode(X86ISD::VBROADCAST, dl, VT, Ld);
5421 // The scalar source must be a normal load.
5425 if (ScalarSize == 32 || (IsGE256 && ScalarSize == 64) ||
5426 (Subtarget->hasVLX() && ScalarSize == 64))
5427 return DAG.getNode(X86ISD::VBROADCAST, dl, VT, Ld);
5429 // The integer check is needed for the 64-bit into 128-bit so it doesn't match
5430 // double since there is no vbroadcastsd xmm
5431 if (Subtarget->hasInt256() && Ld.getValueType().isInteger()) {
5432 if (ScalarSize == 8 || ScalarSize == 16 || ScalarSize == 64)
5433 return DAG.getNode(X86ISD::VBROADCAST, dl, VT, Ld);
5436 // Unsupported broadcast.
5440 /// \brief For an EXTRACT_VECTOR_ELT with a constant index return the real
5441 /// underlying vector and index.
5443 /// Modifies \p ExtractedFromVec to the real vector and returns the real
5445 static int getUnderlyingExtractedFromVec(SDValue &ExtractedFromVec,
5447 int Idx = cast<ConstantSDNode>(ExtIdx)->getZExtValue();
5448 if (!isa<ShuffleVectorSDNode>(ExtractedFromVec))
5451 // For 256-bit vectors, LowerEXTRACT_VECTOR_ELT_SSE4 may have already
5453 // (extract_vector_elt (v8f32 %vreg1), Constant<6>)
5455 // (extract_vector_elt (vector_shuffle<2,u,u,u>
5456 // (extract_subvector (v8f32 %vreg0), Constant<4>),
5459 // In this case the vector is the extract_subvector expression and the index
5460 // is 2, as specified by the shuffle.
5461 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(ExtractedFromVec);
5462 SDValue ShuffleVec = SVOp->getOperand(0);
5463 MVT ShuffleVecVT = ShuffleVec.getSimpleValueType();
5464 assert(ShuffleVecVT.getVectorElementType() ==
5465 ExtractedFromVec.getSimpleValueType().getVectorElementType());
5467 int ShuffleIdx = SVOp->getMaskElt(Idx);
5468 if (isUndefOrInRange(ShuffleIdx, 0, ShuffleVecVT.getVectorNumElements())) {
5469 ExtractedFromVec = ShuffleVec;
5475 static SDValue buildFromShuffleMostly(SDValue Op, SelectionDAG &DAG) {
5476 MVT VT = Op.getSimpleValueType();
5478 // Skip if insert_vec_elt is not supported.
5479 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
5480 if (!TLI.isOperationLegalOrCustom(ISD::INSERT_VECTOR_ELT, VT))
5484 unsigned NumElems = Op.getNumOperands();
5488 SmallVector<unsigned, 4> InsertIndices;
5489 SmallVector<int, 8> Mask(NumElems, -1);
5491 for (unsigned i = 0; i != NumElems; ++i) {
5492 unsigned Opc = Op.getOperand(i).getOpcode();
5494 if (Opc == ISD::UNDEF)
5497 if (Opc != ISD::EXTRACT_VECTOR_ELT) {
5498 // Quit if more than 1 elements need inserting.
5499 if (InsertIndices.size() > 1)
5502 InsertIndices.push_back(i);
5506 SDValue ExtractedFromVec = Op.getOperand(i).getOperand(0);
5507 SDValue ExtIdx = Op.getOperand(i).getOperand(1);
5508 // Quit if non-constant index.
5509 if (!isa<ConstantSDNode>(ExtIdx))
5511 int Idx = getUnderlyingExtractedFromVec(ExtractedFromVec, ExtIdx);
5513 // Quit if extracted from vector of different type.
5514 if (ExtractedFromVec.getValueType() != VT)
5517 if (!VecIn1.getNode())
5518 VecIn1 = ExtractedFromVec;
5519 else if (VecIn1 != ExtractedFromVec) {
5520 if (!VecIn2.getNode())
5521 VecIn2 = ExtractedFromVec;
5522 else if (VecIn2 != ExtractedFromVec)
5523 // Quit if more than 2 vectors to shuffle
5527 if (ExtractedFromVec == VecIn1)
5529 else if (ExtractedFromVec == VecIn2)
5530 Mask[i] = Idx + NumElems;
5533 if (!VecIn1.getNode())
5536 VecIn2 = VecIn2.getNode() ? VecIn2 : DAG.getUNDEF(VT);
5537 SDValue NV = DAG.getVectorShuffle(VT, DL, VecIn1, VecIn2, &Mask[0]);
5538 for (unsigned i = 0, e = InsertIndices.size(); i != e; ++i) {
5539 unsigned Idx = InsertIndices[i];
5540 NV = DAG.getNode(ISD::INSERT_VECTOR_ELT, DL, VT, NV, Op.getOperand(Idx),
5541 DAG.getIntPtrConstant(Idx, DL));
5547 static SDValue ConvertI1VectorToInteger(SDValue Op, SelectionDAG &DAG) {
5548 assert(ISD::isBuildVectorOfConstantSDNodes(Op.getNode()) &&
5549 Op.getScalarValueSizeInBits() == 1 &&
5550 "Can not convert non-constant vector");
5551 uint64_t Immediate = 0;
5552 for (unsigned idx = 0, e = Op.getNumOperands(); idx < e; ++idx) {
5553 SDValue In = Op.getOperand(idx);
5554 if (In.getOpcode() != ISD::UNDEF)
5555 Immediate |= cast<ConstantSDNode>(In)->getZExtValue() << idx;
5559 MVT::getIntegerVT(std::max((int)Op.getValueType().getSizeInBits(), 8));
5560 return DAG.getConstant(Immediate, dl, VT);
5562 // Lower BUILD_VECTOR operation for v8i1 and v16i1 types.
5564 X86TargetLowering::LowerBUILD_VECTORvXi1(SDValue Op, SelectionDAG &DAG) const {
5566 MVT VT = Op.getSimpleValueType();
5567 assert((VT.getVectorElementType() == MVT::i1) &&
5568 "Unexpected type in LowerBUILD_VECTORvXi1!");
5571 if (ISD::isBuildVectorAllZeros(Op.getNode())) {
5572 SDValue Cst = DAG.getTargetConstant(0, dl, MVT::i1);
5573 SmallVector<SDValue, 16> Ops(VT.getVectorNumElements(), Cst);
5574 return DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Ops);
5577 if (ISD::isBuildVectorAllOnes(Op.getNode())) {
5578 SDValue Cst = DAG.getTargetConstant(1, dl, MVT::i1);
5579 SmallVector<SDValue, 16> Ops(VT.getVectorNumElements(), Cst);
5580 return DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Ops);
5583 if (ISD::isBuildVectorOfConstantSDNodes(Op.getNode())) {
5584 SDValue Imm = ConvertI1VectorToInteger(Op, DAG);
5585 if (Imm.getValueSizeInBits() == VT.getSizeInBits())
5586 return DAG.getBitcast(VT, Imm);
5587 SDValue ExtVec = DAG.getBitcast(MVT::v8i1, Imm);
5588 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT, ExtVec,
5589 DAG.getIntPtrConstant(0, dl));
5592 // Vector has one or more non-const elements
5593 uint64_t Immediate = 0;
5594 SmallVector<unsigned, 16> NonConstIdx;
5595 bool IsSplat = true;
5596 bool HasConstElts = false;
5598 for (unsigned idx = 0, e = Op.getNumOperands(); idx < e; ++idx) {
5599 SDValue In = Op.getOperand(idx);
5600 if (In.getOpcode() == ISD::UNDEF)
5602 if (!isa<ConstantSDNode>(In))
5603 NonConstIdx.push_back(idx);
5605 Immediate |= cast<ConstantSDNode>(In)->getZExtValue() << idx;
5606 HasConstElts = true;
5610 else if (In != Op.getOperand(SplatIdx))
5614 // for splat use " (select i1 splat_elt, all-ones, all-zeroes)"
5616 return DAG.getNode(ISD::SELECT, dl, VT, Op.getOperand(SplatIdx),
5617 DAG.getConstant(1, dl, VT),
5618 DAG.getConstant(0, dl, VT));
5620 // insert elements one by one
5624 MVT ImmVT = MVT::getIntegerVT(std::max((int)VT.getSizeInBits(), 8));
5625 Imm = DAG.getConstant(Immediate, dl, ImmVT);
5627 else if (HasConstElts)
5628 Imm = DAG.getConstant(0, dl, VT);
5630 Imm = DAG.getUNDEF(VT);
5631 if (Imm.getValueSizeInBits() == VT.getSizeInBits())
5632 DstVec = DAG.getBitcast(VT, Imm);
5634 SDValue ExtVec = DAG.getBitcast(MVT::v8i1, Imm);
5635 DstVec = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT, ExtVec,
5636 DAG.getIntPtrConstant(0, dl));
5639 for (unsigned i = 0; i < NonConstIdx.size(); ++i) {
5640 unsigned InsertIdx = NonConstIdx[i];
5641 DstVec = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, DstVec,
5642 Op.getOperand(InsertIdx),
5643 DAG.getIntPtrConstant(InsertIdx, dl));
5648 /// \brief Return true if \p N implements a horizontal binop and return the
5649 /// operands for the horizontal binop into V0 and V1.
5651 /// This is a helper function of LowerToHorizontalOp().
5652 /// This function checks that the build_vector \p N in input implements a
5653 /// horizontal operation. Parameter \p Opcode defines the kind of horizontal
5654 /// operation to match.
5655 /// For example, if \p Opcode is equal to ISD::ADD, then this function
5656 /// checks if \p N implements a horizontal arithmetic add; if instead \p Opcode
5657 /// is equal to ISD::SUB, then this function checks if this is a horizontal
5660 /// This function only analyzes elements of \p N whose indices are
5661 /// in range [BaseIdx, LastIdx).
5662 static bool isHorizontalBinOp(const BuildVectorSDNode *N, unsigned Opcode,
5664 unsigned BaseIdx, unsigned LastIdx,
5665 SDValue &V0, SDValue &V1) {
5666 EVT VT = N->getValueType(0);
5668 assert(BaseIdx * 2 <= LastIdx && "Invalid Indices in input!");
5669 assert(VT.isVector() && VT.getVectorNumElements() >= LastIdx &&
5670 "Invalid Vector in input!");
5672 bool IsCommutable = (Opcode == ISD::ADD || Opcode == ISD::FADD);
5673 bool CanFold = true;
5674 unsigned ExpectedVExtractIdx = BaseIdx;
5675 unsigned NumElts = LastIdx - BaseIdx;
5676 V0 = DAG.getUNDEF(VT);
5677 V1 = DAG.getUNDEF(VT);
5679 // Check if N implements a horizontal binop.
5680 for (unsigned i = 0, e = NumElts; i != e && CanFold; ++i) {
5681 SDValue Op = N->getOperand(i + BaseIdx);
5684 if (Op->getOpcode() == ISD::UNDEF) {
5685 // Update the expected vector extract index.
5686 if (i * 2 == NumElts)
5687 ExpectedVExtractIdx = BaseIdx;
5688 ExpectedVExtractIdx += 2;
5692 CanFold = Op->getOpcode() == Opcode && Op->hasOneUse();
5697 SDValue Op0 = Op.getOperand(0);
5698 SDValue Op1 = Op.getOperand(1);
5700 // Try to match the following pattern:
5701 // (BINOP (extract_vector_elt A, I), (extract_vector_elt A, I+1))
5702 CanFold = (Op0.getOpcode() == ISD::EXTRACT_VECTOR_ELT &&
5703 Op1.getOpcode() == ISD::EXTRACT_VECTOR_ELT &&
5704 Op0.getOperand(0) == Op1.getOperand(0) &&
5705 isa<ConstantSDNode>(Op0.getOperand(1)) &&
5706 isa<ConstantSDNode>(Op1.getOperand(1)));
5710 unsigned I0 = cast<ConstantSDNode>(Op0.getOperand(1))->getZExtValue();
5711 unsigned I1 = cast<ConstantSDNode>(Op1.getOperand(1))->getZExtValue();
5713 if (i * 2 < NumElts) {
5714 if (V0.getOpcode() == ISD::UNDEF) {
5715 V0 = Op0.getOperand(0);
5716 if (V0.getValueType() != VT)
5720 if (V1.getOpcode() == ISD::UNDEF) {
5721 V1 = Op0.getOperand(0);
5722 if (V1.getValueType() != VT)
5725 if (i * 2 == NumElts)
5726 ExpectedVExtractIdx = BaseIdx;
5729 SDValue Expected = (i * 2 < NumElts) ? V0 : V1;
5730 if (I0 == ExpectedVExtractIdx)
5731 CanFold = I1 == I0 + 1 && Op0.getOperand(0) == Expected;
5732 else if (IsCommutable && I1 == ExpectedVExtractIdx) {
5733 // Try to match the following dag sequence:
5734 // (BINOP (extract_vector_elt A, I+1), (extract_vector_elt A, I))
5735 CanFold = I0 == I1 + 1 && Op1.getOperand(0) == Expected;
5739 ExpectedVExtractIdx += 2;
5745 /// \brief Emit a sequence of two 128-bit horizontal add/sub followed by
5746 /// a concat_vector.
5748 /// This is a helper function of LowerToHorizontalOp().
5749 /// This function expects two 256-bit vectors called V0 and V1.
5750 /// At first, each vector is split into two separate 128-bit vectors.
5751 /// Then, the resulting 128-bit vectors are used to implement two
5752 /// horizontal binary operations.
5754 /// The kind of horizontal binary operation is defined by \p X86Opcode.
5756 /// \p Mode specifies how the 128-bit parts of V0 and V1 are passed in input to
5757 /// the two new horizontal binop.
5758 /// When Mode is set, the first horizontal binop dag node would take as input
5759 /// the lower 128-bit of V0 and the upper 128-bit of V0. The second
5760 /// horizontal binop dag node would take as input the lower 128-bit of V1
5761 /// and the upper 128-bit of V1.
5763 /// HADD V0_LO, V0_HI
5764 /// HADD V1_LO, V1_HI
5766 /// Otherwise, the first horizontal binop dag node takes as input the lower
5767 /// 128-bit of V0 and the lower 128-bit of V1, and the second horizontal binop
5768 /// dag node takes the upper 128-bit of V0 and the upper 128-bit of V1.
5770 /// HADD V0_LO, V1_LO
5771 /// HADD V0_HI, V1_HI
5773 /// If \p isUndefLO is set, then the algorithm propagates UNDEF to the lower
5774 /// 128-bits of the result. If \p isUndefHI is set, then UNDEF is propagated to
5775 /// the upper 128-bits of the result.
5776 static SDValue ExpandHorizontalBinOp(const SDValue &V0, const SDValue &V1,
5777 SDLoc DL, SelectionDAG &DAG,
5778 unsigned X86Opcode, bool Mode,
5779 bool isUndefLO, bool isUndefHI) {
5780 EVT VT = V0.getValueType();
5781 assert(VT.is256BitVector() && VT == V1.getValueType() &&
5782 "Invalid nodes in input!");
5784 unsigned NumElts = VT.getVectorNumElements();
5785 SDValue V0_LO = Extract128BitVector(V0, 0, DAG, DL);
5786 SDValue V0_HI = Extract128BitVector(V0, NumElts/2, DAG, DL);
5787 SDValue V1_LO = Extract128BitVector(V1, 0, DAG, DL);
5788 SDValue V1_HI = Extract128BitVector(V1, NumElts/2, DAG, DL);
5789 EVT NewVT = V0_LO.getValueType();
5791 SDValue LO = DAG.getUNDEF(NewVT);
5792 SDValue HI = DAG.getUNDEF(NewVT);
5795 // Don't emit a horizontal binop if the result is expected to be UNDEF.
5796 if (!isUndefLO && V0->getOpcode() != ISD::UNDEF)
5797 LO = DAG.getNode(X86Opcode, DL, NewVT, V0_LO, V0_HI);
5798 if (!isUndefHI && V1->getOpcode() != ISD::UNDEF)
5799 HI = DAG.getNode(X86Opcode, DL, NewVT, V1_LO, V1_HI);
5801 // Don't emit a horizontal binop if the result is expected to be UNDEF.
5802 if (!isUndefLO && (V0_LO->getOpcode() != ISD::UNDEF ||
5803 V1_LO->getOpcode() != ISD::UNDEF))
5804 LO = DAG.getNode(X86Opcode, DL, NewVT, V0_LO, V1_LO);
5806 if (!isUndefHI && (V0_HI->getOpcode() != ISD::UNDEF ||
5807 V1_HI->getOpcode() != ISD::UNDEF))
5808 HI = DAG.getNode(X86Opcode, DL, NewVT, V0_HI, V1_HI);
5811 return DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, LO, HI);
5814 /// Try to fold a build_vector that performs an 'addsub' to an X86ISD::ADDSUB
5816 static SDValue LowerToAddSub(const BuildVectorSDNode *BV,
5817 const X86Subtarget *Subtarget, SelectionDAG &DAG) {
5818 EVT VT = BV->getValueType(0);
5819 if ((!Subtarget->hasSSE3() || (VT != MVT::v4f32 && VT != MVT::v2f64)) &&
5820 (!Subtarget->hasAVX() || (VT != MVT::v8f32 && VT != MVT::v4f64)))
5824 unsigned NumElts = VT.getVectorNumElements();
5825 SDValue InVec0 = DAG.getUNDEF(VT);
5826 SDValue InVec1 = DAG.getUNDEF(VT);
5828 assert((VT == MVT::v8f32 || VT == MVT::v4f64 || VT == MVT::v4f32 ||
5829 VT == MVT::v2f64) && "build_vector with an invalid type found!");
5831 // Odd-numbered elements in the input build vector are obtained from
5832 // adding two integer/float elements.
5833 // Even-numbered elements in the input build vector are obtained from
5834 // subtracting two integer/float elements.
5835 unsigned ExpectedOpcode = ISD::FSUB;
5836 unsigned NextExpectedOpcode = ISD::FADD;
5837 bool AddFound = false;
5838 bool SubFound = false;
5840 for (unsigned i = 0, e = NumElts; i != e; ++i) {
5841 SDValue Op = BV->getOperand(i);
5843 // Skip 'undef' values.
5844 unsigned Opcode = Op.getOpcode();
5845 if (Opcode == ISD::UNDEF) {
5846 std::swap(ExpectedOpcode, NextExpectedOpcode);
5850 // Early exit if we found an unexpected opcode.
5851 if (Opcode != ExpectedOpcode)
5854 SDValue Op0 = Op.getOperand(0);
5855 SDValue Op1 = Op.getOperand(1);
5857 // Try to match the following pattern:
5858 // (BINOP (extract_vector_elt A, i), (extract_vector_elt B, i))
5859 // Early exit if we cannot match that sequence.
5860 if (Op0.getOpcode() != ISD::EXTRACT_VECTOR_ELT ||
5861 Op1.getOpcode() != ISD::EXTRACT_VECTOR_ELT ||
5862 !isa<ConstantSDNode>(Op0.getOperand(1)) ||
5863 !isa<ConstantSDNode>(Op1.getOperand(1)) ||
5864 Op0.getOperand(1) != Op1.getOperand(1))
5867 unsigned I0 = cast<ConstantSDNode>(Op0.getOperand(1))->getZExtValue();
5871 // We found a valid add/sub node. Update the information accordingly.
5877 // Update InVec0 and InVec1.
5878 if (InVec0.getOpcode() == ISD::UNDEF) {
5879 InVec0 = Op0.getOperand(0);
5880 if (InVec0.getValueType() != VT)
5883 if (InVec1.getOpcode() == ISD::UNDEF) {
5884 InVec1 = Op1.getOperand(0);
5885 if (InVec1.getValueType() != VT)
5889 // Make sure that operands in input to each add/sub node always
5890 // come from a same pair of vectors.
5891 if (InVec0 != Op0.getOperand(0)) {
5892 if (ExpectedOpcode == ISD::FSUB)
5895 // FADD is commutable. Try to commute the operands
5896 // and then test again.
5897 std::swap(Op0, Op1);
5898 if (InVec0 != Op0.getOperand(0))
5902 if (InVec1 != Op1.getOperand(0))
5905 // Update the pair of expected opcodes.
5906 std::swap(ExpectedOpcode, NextExpectedOpcode);
5909 // Don't try to fold this build_vector into an ADDSUB if the inputs are undef.
5910 if (AddFound && SubFound && InVec0.getOpcode() != ISD::UNDEF &&
5911 InVec1.getOpcode() != ISD::UNDEF)
5912 return DAG.getNode(X86ISD::ADDSUB, DL, VT, InVec0, InVec1);
5917 /// Lower BUILD_VECTOR to a horizontal add/sub operation if possible.
5918 static SDValue LowerToHorizontalOp(const BuildVectorSDNode *BV,
5919 const X86Subtarget *Subtarget,
5920 SelectionDAG &DAG) {
5921 EVT VT = BV->getValueType(0);
5922 unsigned NumElts = VT.getVectorNumElements();
5923 unsigned NumUndefsLO = 0;
5924 unsigned NumUndefsHI = 0;
5925 unsigned Half = NumElts/2;
5927 // Count the number of UNDEF operands in the build_vector in input.
5928 for (unsigned i = 0, e = Half; i != e; ++i)
5929 if (BV->getOperand(i)->getOpcode() == ISD::UNDEF)
5932 for (unsigned i = Half, e = NumElts; i != e; ++i)
5933 if (BV->getOperand(i)->getOpcode() == ISD::UNDEF)
5936 // Early exit if this is either a build_vector of all UNDEFs or all the
5937 // operands but one are UNDEF.
5938 if (NumUndefsLO + NumUndefsHI + 1 >= NumElts)
5942 SDValue InVec0, InVec1;
5943 if ((VT == MVT::v4f32 || VT == MVT::v2f64) && Subtarget->hasSSE3()) {
5944 // Try to match an SSE3 float HADD/HSUB.
5945 if (isHorizontalBinOp(BV, ISD::FADD, DAG, 0, NumElts, InVec0, InVec1))
5946 return DAG.getNode(X86ISD::FHADD, DL, VT, InVec0, InVec1);
5948 if (isHorizontalBinOp(BV, ISD::FSUB, DAG, 0, NumElts, InVec0, InVec1))
5949 return DAG.getNode(X86ISD::FHSUB, DL, VT, InVec0, InVec1);
5950 } else if ((VT == MVT::v4i32 || VT == MVT::v8i16) && Subtarget->hasSSSE3()) {
5951 // Try to match an SSSE3 integer HADD/HSUB.
5952 if (isHorizontalBinOp(BV, ISD::ADD, DAG, 0, NumElts, InVec0, InVec1))
5953 return DAG.getNode(X86ISD::HADD, DL, VT, InVec0, InVec1);
5955 if (isHorizontalBinOp(BV, ISD::SUB, DAG, 0, NumElts, InVec0, InVec1))
5956 return DAG.getNode(X86ISD::HSUB, DL, VT, InVec0, InVec1);
5959 if (!Subtarget->hasAVX())
5962 if ((VT == MVT::v8f32 || VT == MVT::v4f64)) {
5963 // Try to match an AVX horizontal add/sub of packed single/double
5964 // precision floating point values from 256-bit vectors.
5965 SDValue InVec2, InVec3;
5966 if (isHorizontalBinOp(BV, ISD::FADD, DAG, 0, Half, InVec0, InVec1) &&
5967 isHorizontalBinOp(BV, ISD::FADD, DAG, Half, NumElts, InVec2, InVec3) &&
5968 ((InVec0.getOpcode() == ISD::UNDEF ||
5969 InVec2.getOpcode() == ISD::UNDEF) || InVec0 == InVec2) &&
5970 ((InVec1.getOpcode() == ISD::UNDEF ||
5971 InVec3.getOpcode() == ISD::UNDEF) || InVec1 == InVec3))
5972 return DAG.getNode(X86ISD::FHADD, DL, VT, InVec0, InVec1);
5974 if (isHorizontalBinOp(BV, ISD::FSUB, DAG, 0, Half, InVec0, InVec1) &&
5975 isHorizontalBinOp(BV, ISD::FSUB, DAG, Half, NumElts, InVec2, InVec3) &&
5976 ((InVec0.getOpcode() == ISD::UNDEF ||
5977 InVec2.getOpcode() == ISD::UNDEF) || InVec0 == InVec2) &&
5978 ((InVec1.getOpcode() == ISD::UNDEF ||
5979 InVec3.getOpcode() == ISD::UNDEF) || InVec1 == InVec3))
5980 return DAG.getNode(X86ISD::FHSUB, DL, VT, InVec0, InVec1);
5981 } else if (VT == MVT::v8i32 || VT == MVT::v16i16) {
5982 // Try to match an AVX2 horizontal add/sub of signed integers.
5983 SDValue InVec2, InVec3;
5985 bool CanFold = true;
5987 if (isHorizontalBinOp(BV, ISD::ADD, DAG, 0, Half, InVec0, InVec1) &&
5988 isHorizontalBinOp(BV, ISD::ADD, DAG, Half, NumElts, InVec2, InVec3) &&
5989 ((InVec0.getOpcode() == ISD::UNDEF ||
5990 InVec2.getOpcode() == ISD::UNDEF) || InVec0 == InVec2) &&
5991 ((InVec1.getOpcode() == ISD::UNDEF ||
5992 InVec3.getOpcode() == ISD::UNDEF) || InVec1 == InVec3))
5993 X86Opcode = X86ISD::HADD;
5994 else if (isHorizontalBinOp(BV, ISD::SUB, DAG, 0, Half, InVec0, InVec1) &&
5995 isHorizontalBinOp(BV, ISD::SUB, DAG, Half, NumElts, InVec2, InVec3) &&
5996 ((InVec0.getOpcode() == ISD::UNDEF ||
5997 InVec2.getOpcode() == ISD::UNDEF) || InVec0 == InVec2) &&
5998 ((InVec1.getOpcode() == ISD::UNDEF ||
5999 InVec3.getOpcode() == ISD::UNDEF) || InVec1 == InVec3))
6000 X86Opcode = X86ISD::HSUB;
6005 // Fold this build_vector into a single horizontal add/sub.
6006 // Do this only if the target has AVX2.
6007 if (Subtarget->hasAVX2())
6008 return DAG.getNode(X86Opcode, DL, VT, InVec0, InVec1);
6010 // Do not try to expand this build_vector into a pair of horizontal
6011 // add/sub if we can emit a pair of scalar add/sub.
6012 if (NumUndefsLO + 1 == Half || NumUndefsHI + 1 == Half)
6015 // Convert this build_vector into a pair of horizontal binop followed by
6017 bool isUndefLO = NumUndefsLO == Half;
6018 bool isUndefHI = NumUndefsHI == Half;
6019 return ExpandHorizontalBinOp(InVec0, InVec1, DL, DAG, X86Opcode, false,
6020 isUndefLO, isUndefHI);
6024 if ((VT == MVT::v8f32 || VT == MVT::v4f64 || VT == MVT::v8i32 ||
6025 VT == MVT::v16i16) && Subtarget->hasAVX()) {
6027 if (isHorizontalBinOp(BV, ISD::ADD, DAG, 0, NumElts, InVec0, InVec1))
6028 X86Opcode = X86ISD::HADD;
6029 else if (isHorizontalBinOp(BV, ISD::SUB, DAG, 0, NumElts, InVec0, InVec1))
6030 X86Opcode = X86ISD::HSUB;
6031 else if (isHorizontalBinOp(BV, ISD::FADD, DAG, 0, NumElts, InVec0, InVec1))
6032 X86Opcode = X86ISD::FHADD;
6033 else if (isHorizontalBinOp(BV, ISD::FSUB, DAG, 0, NumElts, InVec0, InVec1))
6034 X86Opcode = X86ISD::FHSUB;
6038 // Don't try to expand this build_vector into a pair of horizontal add/sub
6039 // if we can simply emit a pair of scalar add/sub.
6040 if (NumUndefsLO + 1 == Half || NumUndefsHI + 1 == Half)
6043 // Convert this build_vector into two horizontal add/sub followed by
6045 bool isUndefLO = NumUndefsLO == Half;
6046 bool isUndefHI = NumUndefsHI == Half;
6047 return ExpandHorizontalBinOp(InVec0, InVec1, DL, DAG, X86Opcode, true,
6048 isUndefLO, isUndefHI);
6055 X86TargetLowering::LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG) const {
6058 MVT VT = Op.getSimpleValueType();
6059 MVT ExtVT = VT.getVectorElementType();
6060 unsigned NumElems = Op.getNumOperands();
6062 // Generate vectors for predicate vectors.
6063 if (VT.getScalarType() == MVT::i1 && Subtarget->hasAVX512())
6064 return LowerBUILD_VECTORvXi1(Op, DAG);
6066 // Vectors containing all zeros can be matched by pxor and xorps later
6067 if (ISD::isBuildVectorAllZeros(Op.getNode())) {
6068 // Canonicalize this to <4 x i32> to 1) ensure the zero vectors are CSE'd
6069 // and 2) ensure that i64 scalars are eliminated on x86-32 hosts.
6070 if (VT == MVT::v4i32 || VT == MVT::v8i32 || VT == MVT::v16i32)
6073 return getZeroVector(VT, Subtarget, DAG, dl);
6076 // Vectors containing all ones can be matched by pcmpeqd on 128-bit width
6077 // vectors or broken into v4i32 operations on 256-bit vectors. AVX2 can use
6078 // vpcmpeqd on 256-bit vectors.
6079 if (Subtarget->hasSSE2() && ISD::isBuildVectorAllOnes(Op.getNode())) {
6080 if (VT == MVT::v4i32 || (VT == MVT::v8i32 && Subtarget->hasInt256()))
6083 if (!VT.is512BitVector())
6084 return getOnesVector(VT, Subtarget, DAG, dl);
6087 BuildVectorSDNode *BV = cast<BuildVectorSDNode>(Op.getNode());
6088 if (SDValue AddSub = LowerToAddSub(BV, Subtarget, DAG))
6090 if (SDValue HorizontalOp = LowerToHorizontalOp(BV, Subtarget, DAG))
6091 return HorizontalOp;
6092 if (SDValue Broadcast = LowerVectorBroadcast(Op, Subtarget, DAG))
6095 unsigned EVTBits = ExtVT.getSizeInBits();
6097 unsigned NumZero = 0;
6098 unsigned NumNonZero = 0;
6099 unsigned NonZeros = 0;
6100 bool IsAllConstants = true;
6101 SmallSet<SDValue, 8> Values;
6102 for (unsigned i = 0; i < NumElems; ++i) {
6103 SDValue Elt = Op.getOperand(i);
6104 if (Elt.getOpcode() == ISD::UNDEF)
6107 if (Elt.getOpcode() != ISD::Constant &&
6108 Elt.getOpcode() != ISD::ConstantFP)
6109 IsAllConstants = false;
6110 if (X86::isZeroNode(Elt))
6113 NonZeros |= (1 << i);
6118 // All undef vector. Return an UNDEF. All zero vectors were handled above.
6119 if (NumNonZero == 0)
6120 return DAG.getUNDEF(VT);
6122 // Special case for single non-zero, non-undef, element.
6123 if (NumNonZero == 1) {
6124 unsigned Idx = countTrailingZeros(NonZeros);
6125 SDValue Item = Op.getOperand(Idx);
6127 // If this is an insertion of an i64 value on x86-32, and if the top bits of
6128 // the value are obviously zero, truncate the value to i32 and do the
6129 // insertion that way. Only do this if the value is non-constant or if the
6130 // value is a constant being inserted into element 0. It is cheaper to do
6131 // a constant pool load than it is to do a movd + shuffle.
6132 if (ExtVT == MVT::i64 && !Subtarget->is64Bit() &&
6133 (!IsAllConstants || Idx == 0)) {
6134 if (DAG.MaskedValueIsZero(Item, APInt::getBitsSet(64, 32, 64))) {
6136 assert(VT == MVT::v2i64 && "Expected an SSE value type!");
6137 EVT VecVT = MVT::v4i32;
6139 // Truncate the value (which may itself be a constant) to i32, and
6140 // convert it to a vector with movd (S2V+shuffle to zero extend).
6141 Item = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, Item);
6142 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VecVT, Item);
6143 return DAG.getBitcast(VT, getShuffleVectorZeroOrUndef(
6144 Item, Idx * 2, true, Subtarget, DAG));
6148 // If we have a constant or non-constant insertion into the low element of
6149 // a vector, we can do this with SCALAR_TO_VECTOR + shuffle of zero into
6150 // the rest of the elements. This will be matched as movd/movq/movss/movsd
6151 // depending on what the source datatype is.
6154 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
6156 if (ExtVT == MVT::i32 || ExtVT == MVT::f32 || ExtVT == MVT::f64 ||
6157 (ExtVT == MVT::i64 && Subtarget->is64Bit())) {
6158 if (VT.is512BitVector()) {
6159 SDValue ZeroVec = getZeroVector(VT, Subtarget, DAG, dl);
6160 return DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, ZeroVec,
6161 Item, DAG.getIntPtrConstant(0, dl));
6163 assert((VT.is128BitVector() || VT.is256BitVector()) &&
6164 "Expected an SSE value type!");
6165 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
6166 // Turn it into a MOVL (i.e. movss, movsd, or movd) to a zero vector.
6167 return getShuffleVectorZeroOrUndef(Item, 0, true, Subtarget, DAG);
6170 // We can't directly insert an i8 or i16 into a vector, so zero extend
6172 if (ExtVT == MVT::i16 || ExtVT == MVT::i8) {
6173 Item = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, Item);
6174 if (VT.is256BitVector()) {
6175 if (Subtarget->hasAVX()) {
6176 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v8i32, Item);
6177 Item = getShuffleVectorZeroOrUndef(Item, 0, true, Subtarget, DAG);
6179 // Without AVX, we need to extend to a 128-bit vector and then
6180 // insert into the 256-bit vector.
6181 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32, Item);
6182 SDValue ZeroVec = getZeroVector(MVT::v8i32, Subtarget, DAG, dl);
6183 Item = Insert128BitVector(ZeroVec, Item, 0, DAG, dl);
6186 assert(VT.is128BitVector() && "Expected an SSE value type!");
6187 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32, Item);
6188 Item = getShuffleVectorZeroOrUndef(Item, 0, true, Subtarget, DAG);
6190 return DAG.getBitcast(VT, Item);
6194 // Is it a vector logical left shift?
6195 if (NumElems == 2 && Idx == 1 &&
6196 X86::isZeroNode(Op.getOperand(0)) &&
6197 !X86::isZeroNode(Op.getOperand(1))) {
6198 unsigned NumBits = VT.getSizeInBits();
6199 return getVShift(true, VT,
6200 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
6201 VT, Op.getOperand(1)),
6202 NumBits/2, DAG, *this, dl);
6205 if (IsAllConstants) // Otherwise, it's better to do a constpool load.
6208 // Otherwise, if this is a vector with i32 or f32 elements, and the element
6209 // is a non-constant being inserted into an element other than the low one,
6210 // we can't use a constant pool load. Instead, use SCALAR_TO_VECTOR (aka
6211 // movd/movss) to move this into the low element, then shuffle it into
6213 if (EVTBits == 32) {
6214 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
6215 return getShuffleVectorZeroOrUndef(Item, Idx, NumZero > 0, Subtarget, DAG);
6219 // Splat is obviously ok. Let legalizer expand it to a shuffle.
6220 if (Values.size() == 1) {
6221 if (EVTBits == 32) {
6222 // Instead of a shuffle like this:
6223 // shuffle (scalar_to_vector (load (ptr + 4))), undef, <0, 0, 0, 0>
6224 // Check if it's possible to issue this instead.
6225 // shuffle (vload ptr)), undef, <1, 1, 1, 1>
6226 unsigned Idx = countTrailingZeros(NonZeros);
6227 SDValue Item = Op.getOperand(Idx);
6228 if (Op.getNode()->isOnlyUserOf(Item.getNode()))
6229 return LowerAsSplatVectorLoad(Item, VT, dl, DAG);
6234 // A vector full of immediates; various special cases are already
6235 // handled, so this is best done with a single constant-pool load.
6239 // For AVX-length vectors, see if we can use a vector load to get all of the
6240 // elements, otherwise build the individual 128-bit pieces and use
6241 // shuffles to put them in place.
6242 if (VT.is256BitVector() || VT.is512BitVector()) {
6243 SmallVector<SDValue, 64> V(Op->op_begin(), Op->op_begin() + NumElems);
6245 // Check for a build vector of consecutive loads.
6246 if (SDValue LD = EltsFromConsecutiveLoads(VT, V, dl, DAG, false))
6249 EVT HVT = EVT::getVectorVT(*DAG.getContext(), ExtVT, NumElems/2);
6251 // Build both the lower and upper subvector.
6252 SDValue Lower = DAG.getNode(ISD::BUILD_VECTOR, dl, HVT,
6253 makeArrayRef(&V[0], NumElems/2));
6254 SDValue Upper = DAG.getNode(ISD::BUILD_VECTOR, dl, HVT,
6255 makeArrayRef(&V[NumElems / 2], NumElems/2));
6257 // Recreate the wider vector with the lower and upper part.
6258 if (VT.is256BitVector())
6259 return Concat128BitVectors(Lower, Upper, VT, NumElems, DAG, dl);
6260 return Concat256BitVectors(Lower, Upper, VT, NumElems, DAG, dl);
6263 // Let legalizer expand 2-wide build_vectors.
6264 if (EVTBits == 64) {
6265 if (NumNonZero == 1) {
6266 // One half is zero or undef.
6267 unsigned Idx = countTrailingZeros(NonZeros);
6268 SDValue V2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT,
6269 Op.getOperand(Idx));
6270 return getShuffleVectorZeroOrUndef(V2, Idx, true, Subtarget, DAG);
6275 // If element VT is < 32 bits, convert it to inserts into a zero vector.
6276 if (EVTBits == 8 && NumElems == 16)
6277 if (SDValue V = LowerBuildVectorv16i8(Op, NonZeros,NumNonZero,NumZero, DAG,
6281 if (EVTBits == 16 && NumElems == 8)
6282 if (SDValue V = LowerBuildVectorv8i16(Op, NonZeros,NumNonZero,NumZero, DAG,
6286 // If element VT is == 32 bits and has 4 elems, try to generate an INSERTPS
6287 if (EVTBits == 32 && NumElems == 4)
6288 if (SDValue V = LowerBuildVectorv4x32(Op, DAG, Subtarget, *this))
6291 // If element VT is == 32 bits, turn it into a number of shuffles.
6292 SmallVector<SDValue, 8> V(NumElems);
6293 if (NumElems == 4 && NumZero > 0) {
6294 for (unsigned i = 0; i < 4; ++i) {
6295 bool isZero = !(NonZeros & (1 << i));
6297 V[i] = getZeroVector(VT, Subtarget, DAG, dl);
6299 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(i));
6302 for (unsigned i = 0; i < 2; ++i) {
6303 switch ((NonZeros & (0x3 << i*2)) >> (i*2)) {
6306 V[i] = V[i*2]; // Must be a zero vector.
6309 V[i] = getMOVL(DAG, dl, VT, V[i*2+1], V[i*2]);
6312 V[i] = getMOVL(DAG, dl, VT, V[i*2], V[i*2+1]);
6315 V[i] = getUnpackl(DAG, dl, VT, V[i*2], V[i*2+1]);
6320 bool Reverse1 = (NonZeros & 0x3) == 2;
6321 bool Reverse2 = ((NonZeros & (0x3 << 2)) >> 2) == 2;
6325 static_cast<int>(Reverse2 ? NumElems+1 : NumElems),
6326 static_cast<int>(Reverse2 ? NumElems : NumElems+1)
6328 return DAG.getVectorShuffle(VT, dl, V[0], V[1], &MaskVec[0]);
6331 if (Values.size() > 1 && VT.is128BitVector()) {
6332 // Check for a build vector of consecutive loads.
6333 for (unsigned i = 0; i < NumElems; ++i)
6334 V[i] = Op.getOperand(i);
6336 // Check for elements which are consecutive loads.
6337 if (SDValue LD = EltsFromConsecutiveLoads(VT, V, dl, DAG, false))
6340 // Check for a build vector from mostly shuffle plus few inserting.
6341 if (SDValue Sh = buildFromShuffleMostly(Op, DAG))
6344 // For SSE 4.1, use insertps to put the high elements into the low element.
6345 if (Subtarget->hasSSE41()) {
6347 if (Op.getOperand(0).getOpcode() != ISD::UNDEF)
6348 Result = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(0));
6350 Result = DAG.getUNDEF(VT);
6352 for (unsigned i = 1; i < NumElems; ++i) {
6353 if (Op.getOperand(i).getOpcode() == ISD::UNDEF) continue;
6354 Result = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, Result,
6355 Op.getOperand(i), DAG.getIntPtrConstant(i, dl));
6360 // Otherwise, expand into a number of unpckl*, start by extending each of
6361 // our (non-undef) elements to the full vector width with the element in the
6362 // bottom slot of the vector (which generates no code for SSE).
6363 for (unsigned i = 0; i < NumElems; ++i) {
6364 if (Op.getOperand(i).getOpcode() != ISD::UNDEF)
6365 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(i));
6367 V[i] = DAG.getUNDEF(VT);
6370 // Next, we iteratively mix elements, e.g. for v4f32:
6371 // Step 1: unpcklps 0, 2 ==> X: <?, ?, 2, 0>
6372 // : unpcklps 1, 3 ==> Y: <?, ?, 3, 1>
6373 // Step 2: unpcklps X, Y ==> <3, 2, 1, 0>
6374 unsigned EltStride = NumElems >> 1;
6375 while (EltStride != 0) {
6376 for (unsigned i = 0; i < EltStride; ++i) {
6377 // If V[i+EltStride] is undef and this is the first round of mixing,
6378 // then it is safe to just drop this shuffle: V[i] is already in the
6379 // right place, the one element (since it's the first round) being
6380 // inserted as undef can be dropped. This isn't safe for successive
6381 // rounds because they will permute elements within both vectors.
6382 if (V[i+EltStride].getOpcode() == ISD::UNDEF &&
6383 EltStride == NumElems/2)
6386 V[i] = getUnpackl(DAG, dl, VT, V[i], V[i + EltStride]);
6395 // 256-bit AVX can use the vinsertf128 instruction
6396 // to create 256-bit vectors from two other 128-bit ones.
6397 static SDValue LowerAVXCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) {
6399 MVT ResVT = Op.getSimpleValueType();
6401 assert((ResVT.is256BitVector() ||
6402 ResVT.is512BitVector()) && "Value type must be 256-/512-bit wide");
6404 SDValue V1 = Op.getOperand(0);
6405 SDValue V2 = Op.getOperand(1);
6406 unsigned NumElems = ResVT.getVectorNumElements();
6407 if (ResVT.is256BitVector())
6408 return Concat128BitVectors(V1, V2, ResVT, NumElems, DAG, dl);
6410 if (Op.getNumOperands() == 4) {
6411 MVT HalfVT = MVT::getVectorVT(ResVT.getScalarType(),
6412 ResVT.getVectorNumElements()/2);
6413 SDValue V3 = Op.getOperand(2);
6414 SDValue V4 = Op.getOperand(3);
6415 return Concat256BitVectors(Concat128BitVectors(V1, V2, HalfVT, NumElems/2, DAG, dl),
6416 Concat128BitVectors(V3, V4, HalfVT, NumElems/2, DAG, dl), ResVT, NumElems, DAG, dl);
6418 return Concat256BitVectors(V1, V2, ResVT, NumElems, DAG, dl);
6421 static SDValue LowerCONCAT_VECTORSvXi1(SDValue Op,
6422 const X86Subtarget *Subtarget,
6423 SelectionDAG & DAG) {
6425 MVT ResVT = Op.getSimpleValueType();
6426 unsigned NumOfOperands = Op.getNumOperands();
6428 assert(isPowerOf2_32(NumOfOperands) &&
6429 "Unexpected number of operands in CONCAT_VECTORS");
6431 if (NumOfOperands > 2) {
6432 MVT HalfVT = MVT::getVectorVT(ResVT.getScalarType(),
6433 ResVT.getVectorNumElements()/2);
6434 SmallVector<SDValue, 2> Ops;
6435 for (unsigned i = 0; i < NumOfOperands/2; i++)
6436 Ops.push_back(Op.getOperand(i));
6437 SDValue Lo = DAG.getNode(ISD::CONCAT_VECTORS, dl, HalfVT, Ops);
6439 for (unsigned i = NumOfOperands/2; i < NumOfOperands; i++)
6440 Ops.push_back(Op.getOperand(i));
6441 SDValue Hi = DAG.getNode(ISD::CONCAT_VECTORS, dl, HalfVT, Ops);
6442 return DAG.getNode(ISD::CONCAT_VECTORS, dl, ResVT, Lo, Hi);
6445 SDValue V1 = Op.getOperand(0);
6446 SDValue V2 = Op.getOperand(1);
6447 bool IsZeroV1 = ISD::isBuildVectorAllZeros(V1.getNode());
6448 bool IsZeroV2 = ISD::isBuildVectorAllZeros(V2.getNode());
6450 if (IsZeroV1 && IsZeroV2)
6451 return getZeroVector(ResVT, Subtarget, DAG, dl);
6453 SDValue ZeroIdx = DAG.getIntPtrConstant(0, dl);
6454 SDValue Undef = DAG.getUNDEF(ResVT);
6455 unsigned NumElems = ResVT.getVectorNumElements();
6456 SDValue ShiftBits = DAG.getConstant(NumElems/2, dl, MVT::i8);
6458 V2 = DAG.getNode(ISD::INSERT_SUBVECTOR, dl, ResVT, Undef, V2, ZeroIdx);
6459 V2 = DAG.getNode(X86ISD::VSHLI, dl, ResVT, V2, ShiftBits);
6463 V1 = DAG.getNode(ISD::INSERT_SUBVECTOR, dl, ResVT, Undef, V1, ZeroIdx);
6464 // Zero the upper bits of V1
6465 V1 = DAG.getNode(X86ISD::VSHLI, dl, ResVT, V1, ShiftBits);
6466 V1 = DAG.getNode(X86ISD::VSRLI, dl, ResVT, V1, ShiftBits);
6469 return DAG.getNode(ISD::OR, dl, ResVT, V1, V2);
6472 static SDValue LowerCONCAT_VECTORS(SDValue Op,
6473 const X86Subtarget *Subtarget,
6474 SelectionDAG &DAG) {
6475 MVT VT = Op.getSimpleValueType();
6476 if (VT.getVectorElementType() == MVT::i1)
6477 return LowerCONCAT_VECTORSvXi1(Op, Subtarget, DAG);
6479 assert((VT.is256BitVector() && Op.getNumOperands() == 2) ||
6480 (VT.is512BitVector() && (Op.getNumOperands() == 2 ||
6481 Op.getNumOperands() == 4)));
6483 // AVX can use the vinsertf128 instruction to create 256-bit vectors
6484 // from two other 128-bit ones.
6486 // 512-bit vector may contain 2 256-bit vectors or 4 128-bit vectors
6487 return LowerAVXCONCAT_VECTORS(Op, DAG);
6491 //===----------------------------------------------------------------------===//
6492 // Vector shuffle lowering
6494 // This is an experimental code path for lowering vector shuffles on x86. It is
6495 // designed to handle arbitrary vector shuffles and blends, gracefully
6496 // degrading performance as necessary. It works hard to recognize idiomatic
6497 // shuffles and lower them to optimal instruction patterns without leaving
6498 // a framework that allows reasonably efficient handling of all vector shuffle
6500 //===----------------------------------------------------------------------===//
6502 /// \brief Tiny helper function to identify a no-op mask.
6504 /// This is a somewhat boring predicate function. It checks whether the mask
6505 /// array input, which is assumed to be a single-input shuffle mask of the kind
6506 /// used by the X86 shuffle instructions (not a fully general
6507 /// ShuffleVectorSDNode mask) requires any shuffles to occur. Both undef and an
6508 /// in-place shuffle are 'no-op's.
6509 static bool isNoopShuffleMask(ArrayRef<int> Mask) {
6510 for (int i = 0, Size = Mask.size(); i < Size; ++i)
6511 if (Mask[i] != -1 && Mask[i] != i)
6516 /// \brief Helper function to classify a mask as a single-input mask.
6518 /// This isn't a generic single-input test because in the vector shuffle
6519 /// lowering we canonicalize single inputs to be the first input operand. This
6520 /// means we can more quickly test for a single input by only checking whether
6521 /// an input from the second operand exists. We also assume that the size of
6522 /// mask corresponds to the size of the input vectors which isn't true in the
6523 /// fully general case.
6524 static bool isSingleInputShuffleMask(ArrayRef<int> Mask) {
6526 if (M >= (int)Mask.size())
6531 /// \brief Test whether there are elements crossing 128-bit lanes in this
6534 /// X86 divides up its shuffles into in-lane and cross-lane shuffle operations
6535 /// and we routinely test for these.
6536 static bool is128BitLaneCrossingShuffleMask(MVT VT, ArrayRef<int> Mask) {
6537 int LaneSize = 128 / VT.getScalarSizeInBits();
6538 int Size = Mask.size();
6539 for (int i = 0; i < Size; ++i)
6540 if (Mask[i] >= 0 && (Mask[i] % Size) / LaneSize != i / LaneSize)
6545 /// \brief Test whether a shuffle mask is equivalent within each 128-bit lane.
6547 /// This checks a shuffle mask to see if it is performing the same
6548 /// 128-bit lane-relative shuffle in each 128-bit lane. This trivially implies
6549 /// that it is also not lane-crossing. It may however involve a blend from the
6550 /// same lane of a second vector.
6552 /// The specific repeated shuffle mask is populated in \p RepeatedMask, as it is
6553 /// non-trivial to compute in the face of undef lanes. The representation is
6554 /// *not* suitable for use with existing 128-bit shuffles as it will contain
6555 /// entries from both V1 and V2 inputs to the wider mask.
6557 is128BitLaneRepeatedShuffleMask(MVT VT, ArrayRef<int> Mask,
6558 SmallVectorImpl<int> &RepeatedMask) {
6559 int LaneSize = 128 / VT.getScalarSizeInBits();
6560 RepeatedMask.resize(LaneSize, -1);
6561 int Size = Mask.size();
6562 for (int i = 0; i < Size; ++i) {
6565 if ((Mask[i] % Size) / LaneSize != i / LaneSize)
6566 // This entry crosses lanes, so there is no way to model this shuffle.
6569 // Ok, handle the in-lane shuffles by detecting if and when they repeat.
6570 if (RepeatedMask[i % LaneSize] == -1)
6571 // This is the first non-undef entry in this slot of a 128-bit lane.
6572 RepeatedMask[i % LaneSize] =
6573 Mask[i] < Size ? Mask[i] % LaneSize : Mask[i] % LaneSize + Size;
6574 else if (RepeatedMask[i % LaneSize] + (i / LaneSize) * LaneSize != Mask[i])
6575 // Found a mismatch with the repeated mask.
6581 /// \brief Checks whether a shuffle mask is equivalent to an explicit list of
6584 /// This is a fast way to test a shuffle mask against a fixed pattern:
6586 /// if (isShuffleEquivalent(Mask, 3, 2, {1, 0})) { ... }
6588 /// It returns true if the mask is exactly as wide as the argument list, and
6589 /// each element of the mask is either -1 (signifying undef) or the value given
6590 /// in the argument.
6591 static bool isShuffleEquivalent(SDValue V1, SDValue V2, ArrayRef<int> Mask,
6592 ArrayRef<int> ExpectedMask) {
6593 if (Mask.size() != ExpectedMask.size())
6596 int Size = Mask.size();
6598 // If the values are build vectors, we can look through them to find
6599 // equivalent inputs that make the shuffles equivalent.
6600 auto *BV1 = dyn_cast<BuildVectorSDNode>(V1);
6601 auto *BV2 = dyn_cast<BuildVectorSDNode>(V2);
6603 for (int i = 0; i < Size; ++i)
6604 if (Mask[i] != -1 && Mask[i] != ExpectedMask[i]) {
6605 auto *MaskBV = Mask[i] < Size ? BV1 : BV2;
6606 auto *ExpectedBV = ExpectedMask[i] < Size ? BV1 : BV2;
6607 if (!MaskBV || !ExpectedBV ||
6608 MaskBV->getOperand(Mask[i] % Size) !=
6609 ExpectedBV->getOperand(ExpectedMask[i] % Size))
6616 /// \brief Get a 4-lane 8-bit shuffle immediate for a mask.
6618 /// This helper function produces an 8-bit shuffle immediate corresponding to
6619 /// the ubiquitous shuffle encoding scheme used in x86 instructions for
6620 /// shuffling 4 lanes. It can be used with most of the PSHUF instructions for
6623 /// NB: We rely heavily on "undef" masks preserving the input lane.
6624 static SDValue getV4X86ShuffleImm8ForMask(ArrayRef<int> Mask, SDLoc DL,
6625 SelectionDAG &DAG) {
6626 assert(Mask.size() == 4 && "Only 4-lane shuffle masks");
6627 assert(Mask[0] >= -1 && Mask[0] < 4 && "Out of bound mask element!");
6628 assert(Mask[1] >= -1 && Mask[1] < 4 && "Out of bound mask element!");
6629 assert(Mask[2] >= -1 && Mask[2] < 4 && "Out of bound mask element!");
6630 assert(Mask[3] >= -1 && Mask[3] < 4 && "Out of bound mask element!");
6633 Imm |= (Mask[0] == -1 ? 0 : Mask[0]) << 0;
6634 Imm |= (Mask[1] == -1 ? 1 : Mask[1]) << 2;
6635 Imm |= (Mask[2] == -1 ? 2 : Mask[2]) << 4;
6636 Imm |= (Mask[3] == -1 ? 3 : Mask[3]) << 6;
6637 return DAG.getConstant(Imm, DL, MVT::i8);
6640 /// \brief Compute whether each element of a shuffle is zeroable.
6642 /// A "zeroable" vector shuffle element is one which can be lowered to zero.
6643 /// Either it is an undef element in the shuffle mask, the element of the input
6644 /// referenced is undef, or the element of the input referenced is known to be
6645 /// zero. Many x86 shuffles can zero lanes cheaply and we often want to handle
6646 /// as many lanes with this technique as possible to simplify the remaining
6648 static SmallBitVector computeZeroableShuffleElements(ArrayRef<int> Mask,
6649 SDValue V1, SDValue V2) {
6650 SmallBitVector Zeroable(Mask.size(), false);
6652 while (V1.getOpcode() == ISD::BITCAST)
6653 V1 = V1->getOperand(0);
6654 while (V2.getOpcode() == ISD::BITCAST)
6655 V2 = V2->getOperand(0);
6657 bool V1IsZero = ISD::isBuildVectorAllZeros(V1.getNode());
6658 bool V2IsZero = ISD::isBuildVectorAllZeros(V2.getNode());
6660 for (int i = 0, Size = Mask.size(); i < Size; ++i) {
6662 // Handle the easy cases.
6663 if (M < 0 || (M >= 0 && M < Size && V1IsZero) || (M >= Size && V2IsZero)) {
6668 // If this is an index into a build_vector node (which has the same number
6669 // of elements), dig out the input value and use it.
6670 SDValue V = M < Size ? V1 : V2;
6671 if (V.getOpcode() != ISD::BUILD_VECTOR || Size != (int)V.getNumOperands())
6674 SDValue Input = V.getOperand(M % Size);
6675 // The UNDEF opcode check really should be dead code here, but not quite
6676 // worth asserting on (it isn't invalid, just unexpected).
6677 if (Input.getOpcode() == ISD::UNDEF || X86::isZeroNode(Input))
6684 // X86 has dedicated unpack instructions that can handle specific blend
6685 // operations: UNPCKH and UNPCKL.
6686 static SDValue lowerVectorShuffleWithUNPCK(SDLoc DL, MVT VT, ArrayRef<int> Mask,
6687 SDValue V1, SDValue V2,
6688 SelectionDAG &DAG) {
6689 int NumElts = VT.getVectorNumElements();
6692 bool UnpcklSwapped = true;
6693 bool UnpckhSwapped = true;
6694 int NumEltsInLane = 128 / VT.getScalarSizeInBits();
6696 for (int i = 0; i < NumElts; ++i) {
6697 unsigned LaneStart = (i / NumEltsInLane) * NumEltsInLane;
6699 int LoPos = (i % NumEltsInLane) / 2 + LaneStart + NumElts * (i % 2);
6700 int HiPos = LoPos + NumEltsInLane / 2;
6701 int LoPosSwapped = (LoPos + NumElts) % (NumElts * 2);
6702 int HiPosSwapped = (HiPos + NumElts) % (NumElts * 2);
6706 if (Mask[i] != LoPos)
6708 if (Mask[i] != HiPos)
6710 if (Mask[i] != LoPosSwapped)
6711 UnpcklSwapped = false;
6712 if (Mask[i] != HiPosSwapped)
6713 UnpckhSwapped = false;
6714 if (!Unpckl && !Unpckh && !UnpcklSwapped && !UnpckhSwapped)
6718 return DAG.getNode(X86ISD::UNPCKL, DL, VT, V1, V2);
6720 return DAG.getNode(X86ISD::UNPCKH, DL, VT, V1, V2);
6722 return DAG.getNode(X86ISD::UNPCKL, DL, VT, V2, V1);
6724 return DAG.getNode(X86ISD::UNPCKH, DL, VT, V2, V1);
6726 llvm_unreachable("Unexpected result of UNPCK mask analysis");
6730 /// \brief Try to emit a bitmask instruction for a shuffle.
6732 /// This handles cases where we can model a blend exactly as a bitmask due to
6733 /// one of the inputs being zeroable.
6734 static SDValue lowerVectorShuffleAsBitMask(SDLoc DL, MVT VT, SDValue V1,
6735 SDValue V2, ArrayRef<int> Mask,
6736 SelectionDAG &DAG) {
6737 MVT EltVT = VT.getScalarType();
6738 int NumEltBits = EltVT.getSizeInBits();
6739 MVT IntEltVT = MVT::getIntegerVT(NumEltBits);
6740 SDValue Zero = DAG.getConstant(0, DL, IntEltVT);
6741 SDValue AllOnes = DAG.getConstant(APInt::getAllOnesValue(NumEltBits), DL,
6743 if (EltVT.isFloatingPoint()) {
6744 Zero = DAG.getBitcast(EltVT, Zero);
6745 AllOnes = DAG.getBitcast(EltVT, AllOnes);
6747 SmallVector<SDValue, 16> VMaskOps(Mask.size(), Zero);
6748 SmallBitVector Zeroable = computeZeroableShuffleElements(Mask, V1, V2);
6750 for (int i = 0, Size = Mask.size(); i < Size; ++i) {
6753 if (Mask[i] % Size != i)
6754 return SDValue(); // Not a blend.
6756 V = Mask[i] < Size ? V1 : V2;
6757 else if (V != (Mask[i] < Size ? V1 : V2))
6758 return SDValue(); // Can only let one input through the mask.
6760 VMaskOps[i] = AllOnes;
6763 return SDValue(); // No non-zeroable elements!
6765 SDValue VMask = DAG.getNode(ISD::BUILD_VECTOR, DL, VT, VMaskOps);
6766 V = DAG.getNode(VT.isFloatingPoint()
6767 ? (unsigned) X86ISD::FAND : (unsigned) ISD::AND,
6772 /// \brief Try to emit a blend instruction for a shuffle using bit math.
6774 /// This is used as a fallback approach when first class blend instructions are
6775 /// unavailable. Currently it is only suitable for integer vectors, but could
6776 /// be generalized for floating point vectors if desirable.
6777 static SDValue lowerVectorShuffleAsBitBlend(SDLoc DL, MVT VT, SDValue V1,
6778 SDValue V2, ArrayRef<int> Mask,
6779 SelectionDAG &DAG) {
6780 assert(VT.isInteger() && "Only supports integer vector types!");
6781 MVT EltVT = VT.getScalarType();
6782 int NumEltBits = EltVT.getSizeInBits();
6783 SDValue Zero = DAG.getConstant(0, DL, EltVT);
6784 SDValue AllOnes = DAG.getConstant(APInt::getAllOnesValue(NumEltBits), DL,
6786 SmallVector<SDValue, 16> MaskOps;
6787 for (int i = 0, Size = Mask.size(); i < Size; ++i) {
6788 if (Mask[i] != -1 && Mask[i] != i && Mask[i] != i + Size)
6789 return SDValue(); // Shuffled input!
6790 MaskOps.push_back(Mask[i] < Size ? AllOnes : Zero);
6793 SDValue V1Mask = DAG.getNode(ISD::BUILD_VECTOR, DL, VT, MaskOps);
6794 V1 = DAG.getNode(ISD::AND, DL, VT, V1, V1Mask);
6795 // We have to cast V2 around.
6796 MVT MaskVT = MVT::getVectorVT(MVT::i64, VT.getSizeInBits() / 64);
6797 V2 = DAG.getBitcast(VT, DAG.getNode(X86ISD::ANDNP, DL, MaskVT,
6798 DAG.getBitcast(MaskVT, V1Mask),
6799 DAG.getBitcast(MaskVT, V2)));
6800 return DAG.getNode(ISD::OR, DL, VT, V1, V2);
6803 /// \brief Try to emit a blend instruction for a shuffle.
6805 /// This doesn't do any checks for the availability of instructions for blending
6806 /// these values. It relies on the availability of the X86ISD::BLENDI pattern to
6807 /// be matched in the backend with the type given. What it does check for is
6808 /// that the shuffle mask is in fact a blend.
6809 static SDValue lowerVectorShuffleAsBlend(SDLoc DL, MVT VT, SDValue V1,
6810 SDValue V2, ArrayRef<int> Mask,
6811 const X86Subtarget *Subtarget,
6812 SelectionDAG &DAG) {
6813 unsigned BlendMask = 0;
6814 for (int i = 0, Size = Mask.size(); i < Size; ++i) {
6815 if (Mask[i] >= Size) {
6816 if (Mask[i] != i + Size)
6817 return SDValue(); // Shuffled V2 input!
6818 BlendMask |= 1u << i;
6821 if (Mask[i] >= 0 && Mask[i] != i)
6822 return SDValue(); // Shuffled V1 input!
6824 switch (VT.SimpleTy) {
6829 return DAG.getNode(X86ISD::BLENDI, DL, VT, V1, V2,
6830 DAG.getConstant(BlendMask, DL, MVT::i8));
6834 assert(Subtarget->hasAVX2() && "256-bit integer blends require AVX2!");
6838 // If we have AVX2 it is faster to use VPBLENDD when the shuffle fits into
6839 // that instruction.
6840 if (Subtarget->hasAVX2()) {
6841 // Scale the blend by the number of 32-bit dwords per element.
6842 int Scale = VT.getScalarSizeInBits() / 32;
6844 for (int i = 0, Size = Mask.size(); i < Size; ++i)
6845 if (Mask[i] >= Size)
6846 for (int j = 0; j < Scale; ++j)
6847 BlendMask |= 1u << (i * Scale + j);
6849 MVT BlendVT = VT.getSizeInBits() > 128 ? MVT::v8i32 : MVT::v4i32;
6850 V1 = DAG.getBitcast(BlendVT, V1);
6851 V2 = DAG.getBitcast(BlendVT, V2);
6852 return DAG.getBitcast(
6853 VT, DAG.getNode(X86ISD::BLENDI, DL, BlendVT, V1, V2,
6854 DAG.getConstant(BlendMask, DL, MVT::i8)));
6858 // For integer shuffles we need to expand the mask and cast the inputs to
6859 // v8i16s prior to blending.
6860 int Scale = 8 / VT.getVectorNumElements();
6862 for (int i = 0, Size = Mask.size(); i < Size; ++i)
6863 if (Mask[i] >= Size)
6864 for (int j = 0; j < Scale; ++j)
6865 BlendMask |= 1u << (i * Scale + j);
6867 V1 = DAG.getBitcast(MVT::v8i16, V1);
6868 V2 = DAG.getBitcast(MVT::v8i16, V2);
6869 return DAG.getBitcast(VT,
6870 DAG.getNode(X86ISD::BLENDI, DL, MVT::v8i16, V1, V2,
6871 DAG.getConstant(BlendMask, DL, MVT::i8)));
6875 assert(Subtarget->hasAVX2() && "256-bit integer blends require AVX2!");
6876 SmallVector<int, 8> RepeatedMask;
6877 if (is128BitLaneRepeatedShuffleMask(MVT::v16i16, Mask, RepeatedMask)) {
6878 // We can lower these with PBLENDW which is mirrored across 128-bit lanes.
6879 assert(RepeatedMask.size() == 8 && "Repeated mask size doesn't match!");
6881 for (int i = 0; i < 8; ++i)
6882 if (RepeatedMask[i] >= 16)
6883 BlendMask |= 1u << i;
6884 return DAG.getNode(X86ISD::BLENDI, DL, MVT::v16i16, V1, V2,
6885 DAG.getConstant(BlendMask, DL, MVT::i8));
6891 assert((VT.getSizeInBits() == 128 || Subtarget->hasAVX2()) &&
6892 "256-bit byte-blends require AVX2 support!");
6894 // Attempt to lower to a bitmask if we can. VPAND is faster than VPBLENDVB.
6895 if (SDValue Masked = lowerVectorShuffleAsBitMask(DL, VT, V1, V2, Mask, DAG))
6898 // Scale the blend by the number of bytes per element.
6899 int Scale = VT.getScalarSizeInBits() / 8;
6901 // This form of blend is always done on bytes. Compute the byte vector
6903 MVT BlendVT = MVT::getVectorVT(MVT::i8, VT.getSizeInBits() / 8);
6905 // Compute the VSELECT mask. Note that VSELECT is really confusing in the
6906 // mix of LLVM's code generator and the x86 backend. We tell the code
6907 // generator that boolean values in the elements of an x86 vector register
6908 // are -1 for true and 0 for false. We then use the LLVM semantics of 'true'
6909 // mapping a select to operand #1, and 'false' mapping to operand #2. The
6910 // reality in x86 is that vector masks (pre-AVX-512) use only the high bit
6911 // of the element (the remaining are ignored) and 0 in that high bit would
6912 // mean operand #1 while 1 in the high bit would mean operand #2. So while
6913 // the LLVM model for boolean values in vector elements gets the relevant
6914 // bit set, it is set backwards and over constrained relative to x86's
6916 SmallVector<SDValue, 32> VSELECTMask;
6917 for (int i = 0, Size = Mask.size(); i < Size; ++i)
6918 for (int j = 0; j < Scale; ++j)
6919 VSELECTMask.push_back(
6920 Mask[i] < 0 ? DAG.getUNDEF(MVT::i8)
6921 : DAG.getConstant(Mask[i] < Size ? -1 : 0, DL,
6924 V1 = DAG.getBitcast(BlendVT, V1);
6925 V2 = DAG.getBitcast(BlendVT, V2);
6926 return DAG.getBitcast(VT, DAG.getNode(ISD::VSELECT, DL, BlendVT,
6927 DAG.getNode(ISD::BUILD_VECTOR, DL,
6928 BlendVT, VSELECTMask),
6933 llvm_unreachable("Not a supported integer vector type!");
6937 /// \brief Try to lower as a blend of elements from two inputs followed by
6938 /// a single-input permutation.
6940 /// This matches the pattern where we can blend elements from two inputs and
6941 /// then reduce the shuffle to a single-input permutation.
6942 static SDValue lowerVectorShuffleAsBlendAndPermute(SDLoc DL, MVT VT, SDValue V1,
6945 SelectionDAG &DAG) {
6946 // We build up the blend mask while checking whether a blend is a viable way
6947 // to reduce the shuffle.
6948 SmallVector<int, 32> BlendMask(Mask.size(), -1);
6949 SmallVector<int, 32> PermuteMask(Mask.size(), -1);
6951 for (int i = 0, Size = Mask.size(); i < Size; ++i) {
6955 assert(Mask[i] < Size * 2 && "Shuffle input is out of bounds.");
6957 if (BlendMask[Mask[i] % Size] == -1)
6958 BlendMask[Mask[i] % Size] = Mask[i];
6959 else if (BlendMask[Mask[i] % Size] != Mask[i])
6960 return SDValue(); // Can't blend in the needed input!
6962 PermuteMask[i] = Mask[i] % Size;
6965 SDValue V = DAG.getVectorShuffle(VT, DL, V1, V2, BlendMask);
6966 return DAG.getVectorShuffle(VT, DL, V, DAG.getUNDEF(VT), PermuteMask);
6969 /// \brief Generic routine to decompose a shuffle and blend into indepndent
6970 /// blends and permutes.
6972 /// This matches the extremely common pattern for handling combined
6973 /// shuffle+blend operations on newer X86 ISAs where we have very fast blend
6974 /// operations. It will try to pick the best arrangement of shuffles and
6976 static SDValue lowerVectorShuffleAsDecomposedShuffleBlend(SDLoc DL, MVT VT,
6980 SelectionDAG &DAG) {
6981 // Shuffle the input elements into the desired positions in V1 and V2 and
6982 // blend them together.
6983 SmallVector<int, 32> V1Mask(Mask.size(), -1);
6984 SmallVector<int, 32> V2Mask(Mask.size(), -1);
6985 SmallVector<int, 32> BlendMask(Mask.size(), -1);
6986 for (int i = 0, Size = Mask.size(); i < Size; ++i)
6987 if (Mask[i] >= 0 && Mask[i] < Size) {
6988 V1Mask[i] = Mask[i];
6990 } else if (Mask[i] >= Size) {
6991 V2Mask[i] = Mask[i] - Size;
6992 BlendMask[i] = i + Size;
6995 // Try to lower with the simpler initial blend strategy unless one of the
6996 // input shuffles would be a no-op. We prefer to shuffle inputs as the
6997 // shuffle may be able to fold with a load or other benefit. However, when
6998 // we'll have to do 2x as many shuffles in order to achieve this, blending
6999 // first is a better strategy.
7000 if (!isNoopShuffleMask(V1Mask) && !isNoopShuffleMask(V2Mask))
7001 if (SDValue BlendPerm =
7002 lowerVectorShuffleAsBlendAndPermute(DL, VT, V1, V2, Mask, DAG))
7005 V1 = DAG.getVectorShuffle(VT, DL, V1, DAG.getUNDEF(VT), V1Mask);
7006 V2 = DAG.getVectorShuffle(VT, DL, V2, DAG.getUNDEF(VT), V2Mask);
7007 return DAG.getVectorShuffle(VT, DL, V1, V2, BlendMask);
7010 /// \brief Try to lower a vector shuffle as a byte rotation.
7012 /// SSSE3 has a generic PALIGNR instruction in x86 that will do an arbitrary
7013 /// byte-rotation of the concatenation of two vectors; pre-SSSE3 can use
7014 /// a PSRLDQ/PSLLDQ/POR pattern to get a similar effect. This routine will
7015 /// try to generically lower a vector shuffle through such an pattern. It
7016 /// does not check for the profitability of lowering either as PALIGNR or
7017 /// PSRLDQ/PSLLDQ/POR, only whether the mask is valid to lower in that form.
7018 /// This matches shuffle vectors that look like:
7020 /// v8i16 [11, 12, 13, 14, 15, 0, 1, 2]
7022 /// Essentially it concatenates V1 and V2, shifts right by some number of
7023 /// elements, and takes the low elements as the result. Note that while this is
7024 /// specified as a *right shift* because x86 is little-endian, it is a *left
7025 /// rotate* of the vector lanes.
7026 static SDValue lowerVectorShuffleAsByteRotate(SDLoc DL, MVT VT, SDValue V1,
7029 const X86Subtarget *Subtarget,
7030 SelectionDAG &DAG) {
7031 assert(!isNoopShuffleMask(Mask) && "We shouldn't lower no-op shuffles!");
7033 int NumElts = Mask.size();
7034 int NumLanes = VT.getSizeInBits() / 128;
7035 int NumLaneElts = NumElts / NumLanes;
7037 // We need to detect various ways of spelling a rotation:
7038 // [11, 12, 13, 14, 15, 0, 1, 2]
7039 // [-1, 12, 13, 14, -1, -1, 1, -1]
7040 // [-1, -1, -1, -1, -1, -1, 1, 2]
7041 // [ 3, 4, 5, 6, 7, 8, 9, 10]
7042 // [-1, 4, 5, 6, -1, -1, 9, -1]
7043 // [-1, 4, 5, 6, -1, -1, -1, -1]
7046 for (int l = 0; l < NumElts; l += NumLaneElts) {
7047 for (int i = 0; i < NumLaneElts; ++i) {
7048 if (Mask[l + i] == -1)
7050 assert(Mask[l + i] >= 0 && "Only -1 is a valid negative mask element!");
7052 // Get the mod-Size index and lane correct it.
7053 int LaneIdx = (Mask[l + i] % NumElts) - l;
7054 // Make sure it was in this lane.
7055 if (LaneIdx < 0 || LaneIdx >= NumLaneElts)
7058 // Determine where a rotated vector would have started.
7059 int StartIdx = i - LaneIdx;
7061 // The identity rotation isn't interesting, stop.
7064 // If we found the tail of a vector the rotation must be the missing
7065 // front. If we found the head of a vector, it must be how much of the
7067 int CandidateRotation = StartIdx < 0 ? -StartIdx : NumLaneElts - StartIdx;
7070 Rotation = CandidateRotation;
7071 else if (Rotation != CandidateRotation)
7072 // The rotations don't match, so we can't match this mask.
7075 // Compute which value this mask is pointing at.
7076 SDValue MaskV = Mask[l + i] < NumElts ? V1 : V2;
7078 // Compute which of the two target values this index should be assigned
7079 // to. This reflects whether the high elements are remaining or the low
7080 // elements are remaining.
7081 SDValue &TargetV = StartIdx < 0 ? Hi : Lo;
7083 // Either set up this value if we've not encountered it before, or check
7084 // that it remains consistent.
7087 else if (TargetV != MaskV)
7088 // This may be a rotation, but it pulls from the inputs in some
7089 // unsupported interleaving.
7094 // Check that we successfully analyzed the mask, and normalize the results.
7095 assert(Rotation != 0 && "Failed to locate a viable rotation!");
7096 assert((Lo || Hi) && "Failed to find a rotated input vector!");
7102 // The actual rotate instruction rotates bytes, so we need to scale the
7103 // rotation based on how many bytes are in the vector lane.
7104 int Scale = 16 / NumLaneElts;
7106 // SSSE3 targets can use the palignr instruction.
7107 if (Subtarget->hasSSSE3()) {
7108 // Cast the inputs to i8 vector of correct length to match PALIGNR.
7109 MVT AlignVT = MVT::getVectorVT(MVT::i8, 16 * NumLanes);
7110 Lo = DAG.getBitcast(AlignVT, Lo);
7111 Hi = DAG.getBitcast(AlignVT, Hi);
7113 return DAG.getBitcast(
7114 VT, DAG.getNode(X86ISD::PALIGNR, DL, AlignVT, Lo, Hi,
7115 DAG.getConstant(Rotation * Scale, DL, MVT::i8)));
7118 assert(VT.getSizeInBits() == 128 &&
7119 "Rotate-based lowering only supports 128-bit lowering!");
7120 assert(Mask.size() <= 16 &&
7121 "Can shuffle at most 16 bytes in a 128-bit vector!");
7123 // Default SSE2 implementation
7124 int LoByteShift = 16 - Rotation * Scale;
7125 int HiByteShift = Rotation * Scale;
7127 // Cast the inputs to v2i64 to match PSLLDQ/PSRLDQ.
7128 Lo = DAG.getBitcast(MVT::v2i64, Lo);
7129 Hi = DAG.getBitcast(MVT::v2i64, Hi);
7131 SDValue LoShift = DAG.getNode(X86ISD::VSHLDQ, DL, MVT::v2i64, Lo,
7132 DAG.getConstant(LoByteShift, DL, MVT::i8));
7133 SDValue HiShift = DAG.getNode(X86ISD::VSRLDQ, DL, MVT::v2i64, Hi,
7134 DAG.getConstant(HiByteShift, DL, MVT::i8));
7135 return DAG.getBitcast(VT,
7136 DAG.getNode(ISD::OR, DL, MVT::v2i64, LoShift, HiShift));
7139 /// \brief Try to lower a vector shuffle as a bit shift (shifts in zeros).
7141 /// Attempts to match a shuffle mask against the PSLL(W/D/Q/DQ) and
7142 /// PSRL(W/D/Q/DQ) SSE2 and AVX2 logical bit-shift instructions. The function
7143 /// matches elements from one of the input vectors shuffled to the left or
7144 /// right with zeroable elements 'shifted in'. It handles both the strictly
7145 /// bit-wise element shifts and the byte shift across an entire 128-bit double
7148 /// PSHL : (little-endian) left bit shift.
7149 /// [ zz, 0, zz, 2 ]
7150 /// [ -1, 4, zz, -1 ]
7151 /// PSRL : (little-endian) right bit shift.
7153 /// [ -1, -1, 7, zz]
7154 /// PSLLDQ : (little-endian) left byte shift
7155 /// [ zz, 0, 1, 2, 3, 4, 5, 6]
7156 /// [ zz, zz, -1, -1, 2, 3, 4, -1]
7157 /// [ zz, zz, zz, zz, zz, zz, -1, 1]
7158 /// PSRLDQ : (little-endian) right byte shift
7159 /// [ 5, 6, 7, zz, zz, zz, zz, zz]
7160 /// [ -1, 5, 6, 7, zz, zz, zz, zz]
7161 /// [ 1, 2, -1, -1, -1, -1, zz, zz]
7162 static SDValue lowerVectorShuffleAsShift(SDLoc DL, MVT VT, SDValue V1,
7163 SDValue V2, ArrayRef<int> Mask,
7164 SelectionDAG &DAG) {
7165 SmallBitVector Zeroable = computeZeroableShuffleElements(Mask, V1, V2);
7167 int Size = Mask.size();
7168 assert(Size == (int)VT.getVectorNumElements() && "Unexpected mask size");
7170 auto CheckZeros = [&](int Shift, int Scale, bool Left) {
7171 for (int i = 0; i < Size; i += Scale)
7172 for (int j = 0; j < Shift; ++j)
7173 if (!Zeroable[i + j + (Left ? 0 : (Scale - Shift))])
7179 auto MatchShift = [&](int Shift, int Scale, bool Left, SDValue V) {
7180 for (int i = 0; i != Size; i += Scale) {
7181 unsigned Pos = Left ? i + Shift : i;
7182 unsigned Low = Left ? i : i + Shift;
7183 unsigned Len = Scale - Shift;
7184 if (!isSequentialOrUndefInRange(Mask, Pos, Len,
7185 Low + (V == V1 ? 0 : Size)))
7189 int ShiftEltBits = VT.getScalarSizeInBits() * Scale;
7190 bool ByteShift = ShiftEltBits > 64;
7191 unsigned OpCode = Left ? (ByteShift ? X86ISD::VSHLDQ : X86ISD::VSHLI)
7192 : (ByteShift ? X86ISD::VSRLDQ : X86ISD::VSRLI);
7193 int ShiftAmt = Shift * VT.getScalarSizeInBits() / (ByteShift ? 8 : 1);
7195 // Normalize the scale for byte shifts to still produce an i64 element
7197 Scale = ByteShift ? Scale / 2 : Scale;
7199 // We need to round trip through the appropriate type for the shift.
7200 MVT ShiftSVT = MVT::getIntegerVT(VT.getScalarSizeInBits() * Scale);
7201 MVT ShiftVT = MVT::getVectorVT(ShiftSVT, Size / Scale);
7202 assert(DAG.getTargetLoweringInfo().isTypeLegal(ShiftVT) &&
7203 "Illegal integer vector type");
7204 V = DAG.getBitcast(ShiftVT, V);
7206 V = DAG.getNode(OpCode, DL, ShiftVT, V,
7207 DAG.getConstant(ShiftAmt, DL, MVT::i8));
7208 return DAG.getBitcast(VT, V);
7211 // SSE/AVX supports logical shifts up to 64-bit integers - so we can just
7212 // keep doubling the size of the integer elements up to that. We can
7213 // then shift the elements of the integer vector by whole multiples of
7214 // their width within the elements of the larger integer vector. Test each
7215 // multiple to see if we can find a match with the moved element indices
7216 // and that the shifted in elements are all zeroable.
7217 for (int Scale = 2; Scale * VT.getScalarSizeInBits() <= 128; Scale *= 2)
7218 for (int Shift = 1; Shift != Scale; ++Shift)
7219 for (bool Left : {true, false})
7220 if (CheckZeros(Shift, Scale, Left))
7221 for (SDValue V : {V1, V2})
7222 if (SDValue Match = MatchShift(Shift, Scale, Left, V))
7229 /// \brief Try to lower a vector shuffle using SSE4a EXTRQ/INSERTQ.
7230 static SDValue lowerVectorShuffleWithSSE4A(SDLoc DL, MVT VT, SDValue V1,
7231 SDValue V2, ArrayRef<int> Mask,
7232 SelectionDAG &DAG) {
7233 SmallBitVector Zeroable = computeZeroableShuffleElements(Mask, V1, V2);
7234 assert(!Zeroable.all() && "Fully zeroable shuffle mask");
7236 int Size = Mask.size();
7237 int HalfSize = Size / 2;
7238 assert(Size == (int)VT.getVectorNumElements() && "Unexpected mask size");
7240 // Upper half must be undefined.
7241 if (!isUndefInRange(Mask, HalfSize, HalfSize))
7244 // EXTRQ: Extract Len elements from lower half of source, starting at Idx.
7245 // Remainder of lower half result is zero and upper half is all undef.
7246 auto LowerAsEXTRQ = [&]() {
7247 // Determine the extraction length from the part of the
7248 // lower half that isn't zeroable.
7250 for (; Len >= 0; --Len)
7251 if (!Zeroable[Len - 1])
7253 assert(Len > 0 && "Zeroable shuffle mask");
7255 // Attempt to match first Len sequential elements from the lower half.
7258 for (int i = 0; i != Len; ++i) {
7262 SDValue &V = (M < Size ? V1 : V2);
7265 // All mask elements must be in the lower half.
7269 if (Idx < 0 || (Src == V && Idx == (M - i))) {
7280 assert((Idx + Len) <= HalfSize && "Illegal extraction mask");
7281 int BitLen = (Len * VT.getScalarSizeInBits()) & 0x3f;
7282 int BitIdx = (Idx * VT.getScalarSizeInBits()) & 0x3f;
7283 return DAG.getNode(X86ISD::EXTRQI, DL, VT, Src,
7284 DAG.getConstant(BitLen, DL, MVT::i8),
7285 DAG.getConstant(BitIdx, DL, MVT::i8));
7288 if (SDValue ExtrQ = LowerAsEXTRQ())
7291 // INSERTQ: Extract lowest Len elements from lower half of second source and
7292 // insert over first source, starting at Idx.
7293 // { A[0], .., A[Idx-1], B[0], .., B[Len-1], A[Idx+Len], .., UNDEF, ... }
7294 auto LowerAsInsertQ = [&]() {
7295 for (int Idx = 0; Idx != HalfSize; ++Idx) {
7298 // Attempt to match first source from mask before insertion point.
7299 if (isUndefInRange(Mask, 0, Idx)) {
7301 } else if (isSequentialOrUndefInRange(Mask, 0, Idx, 0)) {
7303 } else if (isSequentialOrUndefInRange(Mask, 0, Idx, Size)) {
7309 // Extend the extraction length looking to match both the insertion of
7310 // the second source and the remaining elements of the first.
7311 for (int Hi = Idx + 1; Hi <= HalfSize; ++Hi) {
7316 if (isSequentialOrUndefInRange(Mask, Idx, Len, 0)) {
7318 } else if (isSequentialOrUndefInRange(Mask, Idx, Len, Size)) {
7324 // Match the remaining elements of the lower half.
7325 if (isUndefInRange(Mask, Hi, HalfSize - Hi)) {
7327 } else if ((!Base || (Base == V1)) &&
7328 isSequentialOrUndefInRange(Mask, Hi, HalfSize - Hi, Hi)) {
7330 } else if ((!Base || (Base == V2)) &&
7331 isSequentialOrUndefInRange(Mask, Hi, HalfSize - Hi,
7338 // We may not have a base (first source) - this can safely be undefined.
7340 Base = DAG.getUNDEF(VT);
7342 int BitLen = (Len * VT.getScalarSizeInBits()) & 0x3f;
7343 int BitIdx = (Idx * VT.getScalarSizeInBits()) & 0x3f;
7344 return DAG.getNode(X86ISD::INSERTQI, DL, VT, Base, Insert,
7345 DAG.getConstant(BitLen, DL, MVT::i8),
7346 DAG.getConstant(BitIdx, DL, MVT::i8));
7353 if (SDValue InsertQ = LowerAsInsertQ())
7359 /// \brief Lower a vector shuffle as a zero or any extension.
7361 /// Given a specific number of elements, element bit width, and extension
7362 /// stride, produce either a zero or any extension based on the available
7363 /// features of the subtarget.
7364 static SDValue lowerVectorShuffleAsSpecificZeroOrAnyExtend(
7365 SDLoc DL, MVT VT, int Scale, bool AnyExt, SDValue InputV,
7366 ArrayRef<int> Mask, const X86Subtarget *Subtarget, SelectionDAG &DAG) {
7367 assert(Scale > 1 && "Need a scale to extend.");
7368 int NumElements = VT.getVectorNumElements();
7369 int EltBits = VT.getScalarSizeInBits();
7370 assert((EltBits == 8 || EltBits == 16 || EltBits == 32) &&
7371 "Only 8, 16, and 32 bit elements can be extended.");
7372 assert(Scale * EltBits <= 64 && "Cannot zero extend past 64 bits.");
7374 // Found a valid zext mask! Try various lowering strategies based on the
7375 // input type and available ISA extensions.
7376 if (Subtarget->hasSSE41()) {
7377 MVT ExtVT = MVT::getVectorVT(MVT::getIntegerVT(EltBits * Scale),
7378 NumElements / Scale);
7379 return DAG.getBitcast(VT, DAG.getNode(X86ISD::VZEXT, DL, ExtVT, InputV));
7382 // For any extends we can cheat for larger element sizes and use shuffle
7383 // instructions that can fold with a load and/or copy.
7384 if (AnyExt && EltBits == 32) {
7385 int PSHUFDMask[4] = {0, -1, 1, -1};
7386 return DAG.getBitcast(
7387 VT, DAG.getNode(X86ISD::PSHUFD, DL, MVT::v4i32,
7388 DAG.getBitcast(MVT::v4i32, InputV),
7389 getV4X86ShuffleImm8ForMask(PSHUFDMask, DL, DAG)));
7391 if (AnyExt && EltBits == 16 && Scale > 2) {
7392 int PSHUFDMask[4] = {0, -1, 0, -1};
7393 InputV = DAG.getNode(X86ISD::PSHUFD, DL, MVT::v4i32,
7394 DAG.getBitcast(MVT::v4i32, InputV),
7395 getV4X86ShuffleImm8ForMask(PSHUFDMask, DL, DAG));
7396 int PSHUFHWMask[4] = {1, -1, -1, -1};
7397 return DAG.getBitcast(
7398 VT, DAG.getNode(X86ISD::PSHUFHW, DL, MVT::v8i16,
7399 DAG.getBitcast(MVT::v8i16, InputV),
7400 getV4X86ShuffleImm8ForMask(PSHUFHWMask, DL, DAG)));
7403 // The SSE4A EXTRQ instruction can efficiently extend the first 2 lanes
7405 if ((Scale * EltBits) == 64 && EltBits < 32 && Subtarget->hasSSE4A()) {
7406 assert(NumElements == (int)Mask.size() && "Unexpected shuffle mask size!");
7407 assert(VT.getSizeInBits() == 128 && "Unexpected vector width!");
7409 SDValue Lo = DAG.getNode(ISD::BITCAST, DL, MVT::v2i64,
7410 DAG.getNode(X86ISD::EXTRQI, DL, VT, InputV,
7411 DAG.getConstant(EltBits, DL, MVT::i8),
7412 DAG.getConstant(0, DL, MVT::i8)));
7413 if (isUndefInRange(Mask, NumElements/2, NumElements/2))
7414 return DAG.getNode(ISD::BITCAST, DL, VT, Lo);
7417 DAG.getNode(ISD::BITCAST, DL, MVT::v2i64,
7418 DAG.getNode(X86ISD::EXTRQI, DL, VT, InputV,
7419 DAG.getConstant(EltBits, DL, MVT::i8),
7420 DAG.getConstant(EltBits, DL, MVT::i8)));
7421 return DAG.getNode(ISD::BITCAST, DL, VT,
7422 DAG.getNode(X86ISD::UNPCKL, DL, MVT::v2i64, Lo, Hi));
7425 // If this would require more than 2 unpack instructions to expand, use
7426 // pshufb when available. We can only use more than 2 unpack instructions
7427 // when zero extending i8 elements which also makes it easier to use pshufb.
7428 if (Scale > 4 && EltBits == 8 && Subtarget->hasSSSE3()) {
7429 assert(NumElements == 16 && "Unexpected byte vector width!");
7430 SDValue PSHUFBMask[16];
7431 for (int i = 0; i < 16; ++i)
7433 DAG.getConstant((i % Scale == 0) ? i / Scale : 0x80, DL, MVT::i8);
7434 InputV = DAG.getBitcast(MVT::v16i8, InputV);
7435 return DAG.getBitcast(VT,
7436 DAG.getNode(X86ISD::PSHUFB, DL, MVT::v16i8, InputV,
7437 DAG.getNode(ISD::BUILD_VECTOR, DL,
7438 MVT::v16i8, PSHUFBMask)));
7441 // Otherwise emit a sequence of unpacks.
7443 MVT InputVT = MVT::getVectorVT(MVT::getIntegerVT(EltBits), NumElements);
7444 SDValue Ext = AnyExt ? DAG.getUNDEF(InputVT)
7445 : getZeroVector(InputVT, Subtarget, DAG, DL);
7446 InputV = DAG.getBitcast(InputVT, InputV);
7447 InputV = DAG.getNode(X86ISD::UNPCKL, DL, InputVT, InputV, Ext);
7451 } while (Scale > 1);
7452 return DAG.getBitcast(VT, InputV);
7455 /// \brief Try to lower a vector shuffle as a zero extension on any microarch.
7457 /// This routine will try to do everything in its power to cleverly lower
7458 /// a shuffle which happens to match the pattern of a zero extend. It doesn't
7459 /// check for the profitability of this lowering, it tries to aggressively
7460 /// match this pattern. It will use all of the micro-architectural details it
7461 /// can to emit an efficient lowering. It handles both blends with all-zero
7462 /// inputs to explicitly zero-extend and undef-lanes (sometimes undef due to
7463 /// masking out later).
7465 /// The reason we have dedicated lowering for zext-style shuffles is that they
7466 /// are both incredibly common and often quite performance sensitive.
7467 static SDValue lowerVectorShuffleAsZeroOrAnyExtend(
7468 SDLoc DL, MVT VT, SDValue V1, SDValue V2, ArrayRef<int> Mask,
7469 const X86Subtarget *Subtarget, SelectionDAG &DAG) {
7470 SmallBitVector Zeroable = computeZeroableShuffleElements(Mask, V1, V2);
7472 int Bits = VT.getSizeInBits();
7473 int NumElements = VT.getVectorNumElements();
7474 assert(VT.getScalarSizeInBits() <= 32 &&
7475 "Exceeds 32-bit integer zero extension limit");
7476 assert((int)Mask.size() == NumElements && "Unexpected shuffle mask size");
7478 // Define a helper function to check a particular ext-scale and lower to it if
7480 auto Lower = [&](int Scale) -> SDValue {
7483 for (int i = 0; i < NumElements; ++i) {
7485 continue; // Valid anywhere but doesn't tell us anything.
7486 if (i % Scale != 0) {
7487 // Each of the extended elements need to be zeroable.
7491 // We no longer are in the anyext case.
7496 // Each of the base elements needs to be consecutive indices into the
7497 // same input vector.
7498 SDValue V = Mask[i] < NumElements ? V1 : V2;
7501 else if (InputV != V)
7502 return SDValue(); // Flip-flopping inputs.
7504 if (Mask[i] % NumElements != i / Scale)
7505 return SDValue(); // Non-consecutive strided elements.
7508 // If we fail to find an input, we have a zero-shuffle which should always
7509 // have already been handled.
7510 // FIXME: Maybe handle this here in case during blending we end up with one?
7514 return lowerVectorShuffleAsSpecificZeroOrAnyExtend(
7515 DL, VT, Scale, AnyExt, InputV, Mask, Subtarget, DAG);
7518 // The widest scale possible for extending is to a 64-bit integer.
7519 assert(Bits % 64 == 0 &&
7520 "The number of bits in a vector must be divisible by 64 on x86!");
7521 int NumExtElements = Bits / 64;
7523 // Each iteration, try extending the elements half as much, but into twice as
7525 for (; NumExtElements < NumElements; NumExtElements *= 2) {
7526 assert(NumElements % NumExtElements == 0 &&
7527 "The input vector size must be divisible by the extended size.");
7528 if (SDValue V = Lower(NumElements / NumExtElements))
7532 // General extends failed, but 128-bit vectors may be able to use MOVQ.
7536 // Returns one of the source operands if the shuffle can be reduced to a
7537 // MOVQ, copying the lower 64-bits and zero-extending to the upper 64-bits.
7538 auto CanZExtLowHalf = [&]() {
7539 for (int i = NumElements / 2; i != NumElements; ++i)
7542 if (isSequentialOrUndefInRange(Mask, 0, NumElements / 2, 0))
7544 if (isSequentialOrUndefInRange(Mask, 0, NumElements / 2, NumElements))
7549 if (SDValue V = CanZExtLowHalf()) {
7550 V = DAG.getBitcast(MVT::v2i64, V);
7551 V = DAG.getNode(X86ISD::VZEXT_MOVL, DL, MVT::v2i64, V);
7552 return DAG.getBitcast(VT, V);
7555 // No viable ext lowering found.
7559 /// \brief Try to get a scalar value for a specific element of a vector.
7561 /// Looks through BUILD_VECTOR and SCALAR_TO_VECTOR nodes to find a scalar.
7562 static SDValue getScalarValueForVectorElement(SDValue V, int Idx,
7563 SelectionDAG &DAG) {
7564 MVT VT = V.getSimpleValueType();
7565 MVT EltVT = VT.getVectorElementType();
7566 while (V.getOpcode() == ISD::BITCAST)
7567 V = V.getOperand(0);
7568 // If the bitcasts shift the element size, we can't extract an equivalent
7570 MVT NewVT = V.getSimpleValueType();
7571 if (!NewVT.isVector() || NewVT.getScalarSizeInBits() != VT.getScalarSizeInBits())
7574 if (V.getOpcode() == ISD::BUILD_VECTOR ||
7575 (Idx == 0 && V.getOpcode() == ISD::SCALAR_TO_VECTOR)) {
7576 // Ensure the scalar operand is the same size as the destination.
7577 // FIXME: Add support for scalar truncation where possible.
7578 SDValue S = V.getOperand(Idx);
7579 if (EltVT.getSizeInBits() == S.getSimpleValueType().getSizeInBits())
7580 return DAG.getNode(ISD::BITCAST, SDLoc(V), EltVT, S);
7586 /// \brief Helper to test for a load that can be folded with x86 shuffles.
7588 /// This is particularly important because the set of instructions varies
7589 /// significantly based on whether the operand is a load or not.
7590 static bool isShuffleFoldableLoad(SDValue V) {
7591 while (V.getOpcode() == ISD::BITCAST)
7592 V = V.getOperand(0);
7594 return ISD::isNON_EXTLoad(V.getNode());
7597 /// \brief Try to lower insertion of a single element into a zero vector.
7599 /// This is a common pattern that we have especially efficient patterns to lower
7600 /// across all subtarget feature sets.
7601 static SDValue lowerVectorShuffleAsElementInsertion(
7602 SDLoc DL, MVT VT, SDValue V1, SDValue V2, ArrayRef<int> Mask,
7603 const X86Subtarget *Subtarget, SelectionDAG &DAG) {
7604 SmallBitVector Zeroable = computeZeroableShuffleElements(Mask, V1, V2);
7606 MVT EltVT = VT.getVectorElementType();
7608 int V2Index = std::find_if(Mask.begin(), Mask.end(),
7609 [&Mask](int M) { return M >= (int)Mask.size(); }) -
7611 bool IsV1Zeroable = true;
7612 for (int i = 0, Size = Mask.size(); i < Size; ++i)
7613 if (i != V2Index && !Zeroable[i]) {
7614 IsV1Zeroable = false;
7618 // Check for a single input from a SCALAR_TO_VECTOR node.
7619 // FIXME: All of this should be canonicalized into INSERT_VECTOR_ELT and
7620 // all the smarts here sunk into that routine. However, the current
7621 // lowering of BUILD_VECTOR makes that nearly impossible until the old
7622 // vector shuffle lowering is dead.
7623 SDValue V2S = getScalarValueForVectorElement(V2, Mask[V2Index] - Mask.size(),
7625 if (V2S && DAG.getTargetLoweringInfo().isTypeLegal(V2S.getValueType())) {
7626 // We need to zext the scalar if it is smaller than an i32.
7627 V2S = DAG.getBitcast(EltVT, V2S);
7628 if (EltVT == MVT::i8 || EltVT == MVT::i16) {
7629 // Using zext to expand a narrow element won't work for non-zero
7634 // Zero-extend directly to i32.
7636 V2S = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::i32, V2S);
7638 V2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, ExtVT, V2S);
7639 } else if (Mask[V2Index] != (int)Mask.size() || EltVT == MVT::i8 ||
7640 EltVT == MVT::i16) {
7641 // Either not inserting from the low element of the input or the input
7642 // element size is too small to use VZEXT_MOVL to clear the high bits.
7646 if (!IsV1Zeroable) {
7647 // If V1 can't be treated as a zero vector we have fewer options to lower
7648 // this. We can't support integer vectors or non-zero targets cheaply, and
7649 // the V1 elements can't be permuted in any way.
7650 assert(VT == ExtVT && "Cannot change extended type when non-zeroable!");
7651 if (!VT.isFloatingPoint() || V2Index != 0)
7653 SmallVector<int, 8> V1Mask(Mask.begin(), Mask.end());
7654 V1Mask[V2Index] = -1;
7655 if (!isNoopShuffleMask(V1Mask))
7657 // This is essentially a special case blend operation, but if we have
7658 // general purpose blend operations, they are always faster. Bail and let
7659 // the rest of the lowering handle these as blends.
7660 if (Subtarget->hasSSE41())
7663 // Otherwise, use MOVSD or MOVSS.
7664 assert((EltVT == MVT::f32 || EltVT == MVT::f64) &&
7665 "Only two types of floating point element types to handle!");
7666 return DAG.getNode(EltVT == MVT::f32 ? X86ISD::MOVSS : X86ISD::MOVSD, DL,
7670 // This lowering only works for the low element with floating point vectors.
7671 if (VT.isFloatingPoint() && V2Index != 0)
7674 V2 = DAG.getNode(X86ISD::VZEXT_MOVL, DL, ExtVT, V2);
7676 V2 = DAG.getBitcast(VT, V2);
7679 // If we have 4 or fewer lanes we can cheaply shuffle the element into
7680 // the desired position. Otherwise it is more efficient to do a vector
7681 // shift left. We know that we can do a vector shift left because all
7682 // the inputs are zero.
7683 if (VT.isFloatingPoint() || VT.getVectorNumElements() <= 4) {
7684 SmallVector<int, 4> V2Shuffle(Mask.size(), 1);
7685 V2Shuffle[V2Index] = 0;
7686 V2 = DAG.getVectorShuffle(VT, DL, V2, DAG.getUNDEF(VT), V2Shuffle);
7688 V2 = DAG.getBitcast(MVT::v2i64, V2);
7690 X86ISD::VSHLDQ, DL, MVT::v2i64, V2,
7691 DAG.getConstant(V2Index * EltVT.getSizeInBits() / 8, DL,
7692 DAG.getTargetLoweringInfo().getScalarShiftAmountTy(
7693 DAG.getDataLayout(), VT)));
7694 V2 = DAG.getBitcast(VT, V2);
7700 /// \brief Try to lower broadcast of a single element.
7702 /// For convenience, this code also bundles all of the subtarget feature set
7703 /// filtering. While a little annoying to re-dispatch on type here, there isn't
7704 /// a convenient way to factor it out.
7705 static SDValue lowerVectorShuffleAsBroadcast(SDLoc DL, MVT VT, SDValue V,
7707 const X86Subtarget *Subtarget,
7708 SelectionDAG &DAG) {
7709 if (!Subtarget->hasAVX())
7711 if (VT.isInteger() && !Subtarget->hasAVX2())
7714 // Check that the mask is a broadcast.
7715 int BroadcastIdx = -1;
7717 if (M >= 0 && BroadcastIdx == -1)
7719 else if (M >= 0 && M != BroadcastIdx)
7722 assert(BroadcastIdx < (int)Mask.size() && "We only expect to be called with "
7723 "a sorted mask where the broadcast "
7726 // Go up the chain of (vector) values to find a scalar load that we can
7727 // combine with the broadcast.
7729 switch (V.getOpcode()) {
7730 case ISD::CONCAT_VECTORS: {
7731 int OperandSize = Mask.size() / V.getNumOperands();
7732 V = V.getOperand(BroadcastIdx / OperandSize);
7733 BroadcastIdx %= OperandSize;
7737 case ISD::INSERT_SUBVECTOR: {
7738 SDValue VOuter = V.getOperand(0), VInner = V.getOperand(1);
7739 auto ConstantIdx = dyn_cast<ConstantSDNode>(V.getOperand(2));
7743 int BeginIdx = (int)ConstantIdx->getZExtValue();
7745 BeginIdx + (int)VInner.getValueType().getVectorNumElements();
7746 if (BroadcastIdx >= BeginIdx && BroadcastIdx < EndIdx) {
7747 BroadcastIdx -= BeginIdx;
7758 // Check if this is a broadcast of a scalar. We special case lowering
7759 // for scalars so that we can more effectively fold with loads.
7760 // First, look through bitcast: if the original value has a larger element
7761 // type than the shuffle, the broadcast element is in essence truncated.
7762 // Make that explicit to ease folding.
7763 if (V.getOpcode() == ISD::BITCAST && VT.isInteger()) {
7764 EVT EltVT = VT.getVectorElementType();
7765 SDValue V0 = V.getOperand(0);
7766 EVT V0VT = V0.getValueType();
7768 if (V0VT.isInteger() && V0VT.getVectorElementType().bitsGT(EltVT) &&
7769 ((V0.getOpcode() == ISD::BUILD_VECTOR ||
7770 (V0.getOpcode() == ISD::SCALAR_TO_VECTOR && BroadcastIdx == 0)))) {
7771 V = DAG.getNode(ISD::TRUNCATE, DL, EltVT, V0.getOperand(BroadcastIdx));
7776 // Also check the simpler case, where we can directly reuse the scalar.
7777 if (V.getOpcode() == ISD::BUILD_VECTOR ||
7778 (V.getOpcode() == ISD::SCALAR_TO_VECTOR && BroadcastIdx == 0)) {
7779 V = V.getOperand(BroadcastIdx);
7781 // If the scalar isn't a load, we can't broadcast from it in AVX1.
7782 // Only AVX2 has register broadcasts.
7783 if (!Subtarget->hasAVX2() && !isShuffleFoldableLoad(V))
7785 } else if (BroadcastIdx != 0 || !Subtarget->hasAVX2()) {
7786 // We can't broadcast from a vector register without AVX2, and we can only
7787 // broadcast from the zero-element of a vector register.
7791 return DAG.getNode(X86ISD::VBROADCAST, DL, VT, V);
7794 // Check for whether we can use INSERTPS to perform the shuffle. We only use
7795 // INSERTPS when the V1 elements are already in the correct locations
7796 // because otherwise we can just always use two SHUFPS instructions which
7797 // are much smaller to encode than a SHUFPS and an INSERTPS. We can also
7798 // perform INSERTPS if a single V1 element is out of place and all V2
7799 // elements are zeroable.
7800 static SDValue lowerVectorShuffleAsInsertPS(SDValue Op, SDValue V1, SDValue V2,
7802 SelectionDAG &DAG) {
7803 assert(Op.getSimpleValueType() == MVT::v4f32 && "Bad shuffle type!");
7804 assert(V1.getSimpleValueType() == MVT::v4f32 && "Bad operand type!");
7805 assert(V2.getSimpleValueType() == MVT::v4f32 && "Bad operand type!");
7806 assert(Mask.size() == 4 && "Unexpected mask size for v4 shuffle!");
7808 SmallBitVector Zeroable = computeZeroableShuffleElements(Mask, V1, V2);
7811 int V1DstIndex = -1;
7812 int V2DstIndex = -1;
7813 bool V1UsedInPlace = false;
7815 for (int i = 0; i < 4; ++i) {
7816 // Synthesize a zero mask from the zeroable elements (includes undefs).
7822 // Flag if we use any V1 inputs in place.
7824 V1UsedInPlace = true;
7828 // We can only insert a single non-zeroable element.
7829 if (V1DstIndex != -1 || V2DstIndex != -1)
7833 // V1 input out of place for insertion.
7836 // V2 input for insertion.
7841 // Don't bother if we have no (non-zeroable) element for insertion.
7842 if (V1DstIndex == -1 && V2DstIndex == -1)
7845 // Determine element insertion src/dst indices. The src index is from the
7846 // start of the inserted vector, not the start of the concatenated vector.
7847 unsigned V2SrcIndex = 0;
7848 if (V1DstIndex != -1) {
7849 // If we have a V1 input out of place, we use V1 as the V2 element insertion
7850 // and don't use the original V2 at all.
7851 V2SrcIndex = Mask[V1DstIndex];
7852 V2DstIndex = V1DstIndex;
7855 V2SrcIndex = Mask[V2DstIndex] - 4;
7858 // If no V1 inputs are used in place, then the result is created only from
7859 // the zero mask and the V2 insertion - so remove V1 dependency.
7861 V1 = DAG.getUNDEF(MVT::v4f32);
7863 unsigned InsertPSMask = V2SrcIndex << 6 | V2DstIndex << 4 | ZMask;
7864 assert((InsertPSMask & ~0xFFu) == 0 && "Invalid mask!");
7866 // Insert the V2 element into the desired position.
7868 return DAG.getNode(X86ISD::INSERTPS, DL, MVT::v4f32, V1, V2,
7869 DAG.getConstant(InsertPSMask, DL, MVT::i8));
7872 /// \brief Try to lower a shuffle as a permute of the inputs followed by an
7873 /// UNPCK instruction.
7875 /// This specifically targets cases where we end up with alternating between
7876 /// the two inputs, and so can permute them into something that feeds a single
7877 /// UNPCK instruction. Note that this routine only targets integer vectors
7878 /// because for floating point vectors we have a generalized SHUFPS lowering
7879 /// strategy that handles everything that doesn't *exactly* match an unpack,
7880 /// making this clever lowering unnecessary.
7881 static SDValue lowerVectorShuffleAsPermuteAndUnpack(SDLoc DL, MVT VT,
7882 SDValue V1, SDValue V2,
7884 SelectionDAG &DAG) {
7885 assert(!VT.isFloatingPoint() &&
7886 "This routine only supports integer vectors.");
7887 assert(!isSingleInputShuffleMask(Mask) &&
7888 "This routine should only be used when blending two inputs.");
7889 assert(Mask.size() >= 2 && "Single element masks are invalid.");
7891 int Size = Mask.size();
7893 int NumLoInputs = std::count_if(Mask.begin(), Mask.end(), [Size](int M) {
7894 return M >= 0 && M % Size < Size / 2;
7896 int NumHiInputs = std::count_if(
7897 Mask.begin(), Mask.end(), [Size](int M) { return M % Size >= Size / 2; });
7899 bool UnpackLo = NumLoInputs >= NumHiInputs;
7901 auto TryUnpack = [&](MVT UnpackVT, int Scale) {
7902 SmallVector<int, 32> V1Mask(Mask.size(), -1);
7903 SmallVector<int, 32> V2Mask(Mask.size(), -1);
7905 for (int i = 0; i < Size; ++i) {
7909 // Each element of the unpack contains Scale elements from this mask.
7910 int UnpackIdx = i / Scale;
7912 // We only handle the case where V1 feeds the first slots of the unpack.
7913 // We rely on canonicalization to ensure this is the case.
7914 if ((UnpackIdx % 2 == 0) != (Mask[i] < Size))
7917 // Setup the mask for this input. The indexing is tricky as we have to
7918 // handle the unpack stride.
7919 SmallVectorImpl<int> &VMask = (UnpackIdx % 2 == 0) ? V1Mask : V2Mask;
7920 VMask[(UnpackIdx / 2) * Scale + i % Scale + (UnpackLo ? 0 : Size / 2)] =
7924 // If we will have to shuffle both inputs to use the unpack, check whether
7925 // we can just unpack first and shuffle the result. If so, skip this unpack.
7926 if ((NumLoInputs == 0 || NumHiInputs == 0) && !isNoopShuffleMask(V1Mask) &&
7927 !isNoopShuffleMask(V2Mask))
7930 // Shuffle the inputs into place.
7931 V1 = DAG.getVectorShuffle(VT, DL, V1, DAG.getUNDEF(VT), V1Mask);
7932 V2 = DAG.getVectorShuffle(VT, DL, V2, DAG.getUNDEF(VT), V2Mask);
7934 // Cast the inputs to the type we will use to unpack them.
7935 V1 = DAG.getBitcast(UnpackVT, V1);
7936 V2 = DAG.getBitcast(UnpackVT, V2);
7938 // Unpack the inputs and cast the result back to the desired type.
7939 return DAG.getBitcast(
7940 VT, DAG.getNode(UnpackLo ? X86ISD::UNPCKL : X86ISD::UNPCKH, DL,
7944 // We try each unpack from the largest to the smallest to try and find one
7945 // that fits this mask.
7946 int OrigNumElements = VT.getVectorNumElements();
7947 int OrigScalarSize = VT.getScalarSizeInBits();
7948 for (int ScalarSize = 64; ScalarSize >= OrigScalarSize; ScalarSize /= 2) {
7949 int Scale = ScalarSize / OrigScalarSize;
7950 int NumElements = OrigNumElements / Scale;
7951 MVT UnpackVT = MVT::getVectorVT(MVT::getIntegerVT(ScalarSize), NumElements);
7952 if (SDValue Unpack = TryUnpack(UnpackVT, Scale))
7956 // If none of the unpack-rooted lowerings worked (or were profitable) try an
7958 if (NumLoInputs == 0 || NumHiInputs == 0) {
7959 assert((NumLoInputs > 0 || NumHiInputs > 0) &&
7960 "We have to have *some* inputs!");
7961 int HalfOffset = NumLoInputs == 0 ? Size / 2 : 0;
7963 // FIXME: We could consider the total complexity of the permute of each
7964 // possible unpacking. Or at the least we should consider how many
7965 // half-crossings are created.
7966 // FIXME: We could consider commuting the unpacks.
7968 SmallVector<int, 32> PermMask;
7969 PermMask.assign(Size, -1);
7970 for (int i = 0; i < Size; ++i) {
7974 assert(Mask[i] % Size >= HalfOffset && "Found input from wrong half!");
7977 2 * ((Mask[i] % Size) - HalfOffset) + (Mask[i] < Size ? 0 : 1);
7979 return DAG.getVectorShuffle(
7980 VT, DL, DAG.getNode(NumLoInputs == 0 ? X86ISD::UNPCKH : X86ISD::UNPCKL,
7982 DAG.getUNDEF(VT), PermMask);
7988 /// \brief Handle lowering of 2-lane 64-bit floating point shuffles.
7990 /// This is the basis function for the 2-lane 64-bit shuffles as we have full
7991 /// support for floating point shuffles but not integer shuffles. These
7992 /// instructions will incur a domain crossing penalty on some chips though so
7993 /// it is better to avoid lowering through this for integer vectors where
7995 static SDValue lowerV2F64VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
7996 const X86Subtarget *Subtarget,
7997 SelectionDAG &DAG) {
7999 assert(Op.getSimpleValueType() == MVT::v2f64 && "Bad shuffle type!");
8000 assert(V1.getSimpleValueType() == MVT::v2f64 && "Bad operand type!");
8001 assert(V2.getSimpleValueType() == MVT::v2f64 && "Bad operand type!");
8002 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
8003 ArrayRef<int> Mask = SVOp->getMask();
8004 assert(Mask.size() == 2 && "Unexpected mask size for v2 shuffle!");
8006 if (isSingleInputShuffleMask(Mask)) {
8007 // Use low duplicate instructions for masks that match their pattern.
8008 if (Subtarget->hasSSE3())
8009 if (isShuffleEquivalent(V1, V2, Mask, {0, 0}))
8010 return DAG.getNode(X86ISD::MOVDDUP, DL, MVT::v2f64, V1);
8012 // Straight shuffle of a single input vector. Simulate this by using the
8013 // single input as both of the "inputs" to this instruction..
8014 unsigned SHUFPDMask = (Mask[0] == 1) | ((Mask[1] == 1) << 1);
8016 if (Subtarget->hasAVX()) {
8017 // If we have AVX, we can use VPERMILPS which will allow folding a load
8018 // into the shuffle.
8019 return DAG.getNode(X86ISD::VPERMILPI, DL, MVT::v2f64, V1,
8020 DAG.getConstant(SHUFPDMask, DL, MVT::i8));
8023 return DAG.getNode(X86ISD::SHUFP, DL, MVT::v2f64, V1, V1,
8024 DAG.getConstant(SHUFPDMask, DL, MVT::i8));
8026 assert(Mask[0] >= 0 && Mask[0] < 2 && "Non-canonicalized blend!");
8027 assert(Mask[1] >= 2 && "Non-canonicalized blend!");
8029 // If we have a single input, insert that into V1 if we can do so cheaply.
8030 if ((Mask[0] >= 2) + (Mask[1] >= 2) == 1) {
8031 if (SDValue Insertion = lowerVectorShuffleAsElementInsertion(
8032 DL, MVT::v2f64, V1, V2, Mask, Subtarget, DAG))
8034 // Try inverting the insertion since for v2 masks it is easy to do and we
8035 // can't reliably sort the mask one way or the other.
8036 int InverseMask[2] = {Mask[0] < 0 ? -1 : (Mask[0] ^ 2),
8037 Mask[1] < 0 ? -1 : (Mask[1] ^ 2)};
8038 if (SDValue Insertion = lowerVectorShuffleAsElementInsertion(
8039 DL, MVT::v2f64, V2, V1, InverseMask, Subtarget, DAG))
8043 // Try to use one of the special instruction patterns to handle two common
8044 // blend patterns if a zero-blend above didn't work.
8045 if (isShuffleEquivalent(V1, V2, Mask, {0, 3}) ||
8046 isShuffleEquivalent(V1, V2, Mask, {1, 3}))
8047 if (SDValue V1S = getScalarValueForVectorElement(V1, Mask[0], DAG))
8048 // We can either use a special instruction to load over the low double or
8049 // to move just the low double.
8051 isShuffleFoldableLoad(V1S) ? X86ISD::MOVLPD : X86ISD::MOVSD,
8053 DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, MVT::v2f64, V1S));
8055 if (Subtarget->hasSSE41())
8056 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v2f64, V1, V2, Mask,
8060 // Use dedicated unpack instructions for masks that match their pattern.
8061 if (isShuffleEquivalent(V1, V2, Mask, {0, 2}))
8062 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v2f64, V1, V2);
8063 if (isShuffleEquivalent(V1, V2, Mask, {1, 3}))
8064 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v2f64, V1, V2);
8066 unsigned SHUFPDMask = (Mask[0] == 1) | (((Mask[1] - 2) == 1) << 1);
8067 return DAG.getNode(X86ISD::SHUFP, DL, MVT::v2f64, V1, V2,
8068 DAG.getConstant(SHUFPDMask, DL, MVT::i8));
8071 /// \brief Handle lowering of 2-lane 64-bit integer shuffles.
8073 /// Tries to lower a 2-lane 64-bit shuffle using shuffle operations provided by
8074 /// the integer unit to minimize domain crossing penalties. However, for blends
8075 /// it falls back to the floating point shuffle operation with appropriate bit
8077 static SDValue lowerV2I64VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
8078 const X86Subtarget *Subtarget,
8079 SelectionDAG &DAG) {
8081 assert(Op.getSimpleValueType() == MVT::v2i64 && "Bad shuffle type!");
8082 assert(V1.getSimpleValueType() == MVT::v2i64 && "Bad operand type!");
8083 assert(V2.getSimpleValueType() == MVT::v2i64 && "Bad operand type!");
8084 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
8085 ArrayRef<int> Mask = SVOp->getMask();
8086 assert(Mask.size() == 2 && "Unexpected mask size for v2 shuffle!");
8088 if (isSingleInputShuffleMask(Mask)) {
8089 // Check for being able to broadcast a single element.
8090 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v2i64, V1,
8091 Mask, Subtarget, DAG))
8094 // Straight shuffle of a single input vector. For everything from SSE2
8095 // onward this has a single fast instruction with no scary immediates.
8096 // We have to map the mask as it is actually a v4i32 shuffle instruction.
8097 V1 = DAG.getBitcast(MVT::v4i32, V1);
8098 int WidenedMask[4] = {
8099 std::max(Mask[0], 0) * 2, std::max(Mask[0], 0) * 2 + 1,
8100 std::max(Mask[1], 0) * 2, std::max(Mask[1], 0) * 2 + 1};
8101 return DAG.getBitcast(
8103 DAG.getNode(X86ISD::PSHUFD, DL, MVT::v4i32, V1,
8104 getV4X86ShuffleImm8ForMask(WidenedMask, DL, DAG)));
8106 assert(Mask[0] != -1 && "No undef lanes in multi-input v2 shuffles!");
8107 assert(Mask[1] != -1 && "No undef lanes in multi-input v2 shuffles!");
8108 assert(Mask[0] < 2 && "We sort V1 to be the first input.");
8109 assert(Mask[1] >= 2 && "We sort V2 to be the second input.");
8111 // If we have a blend of two PACKUS operations an the blend aligns with the
8112 // low and half halves, we can just merge the PACKUS operations. This is
8113 // particularly important as it lets us merge shuffles that this routine itself
8115 auto GetPackNode = [](SDValue V) {
8116 while (V.getOpcode() == ISD::BITCAST)
8117 V = V.getOperand(0);
8119 return V.getOpcode() == X86ISD::PACKUS ? V : SDValue();
8121 if (SDValue V1Pack = GetPackNode(V1))
8122 if (SDValue V2Pack = GetPackNode(V2))
8123 return DAG.getBitcast(MVT::v2i64,
8124 DAG.getNode(X86ISD::PACKUS, DL, MVT::v16i8,
8125 Mask[0] == 0 ? V1Pack.getOperand(0)
8126 : V1Pack.getOperand(1),
8127 Mask[1] == 2 ? V2Pack.getOperand(0)
8128 : V2Pack.getOperand(1)));
8130 // Try to use shift instructions.
8132 lowerVectorShuffleAsShift(DL, MVT::v2i64, V1, V2, Mask, DAG))
8135 // When loading a scalar and then shuffling it into a vector we can often do
8136 // the insertion cheaply.
8137 if (SDValue Insertion = lowerVectorShuffleAsElementInsertion(
8138 DL, MVT::v2i64, V1, V2, Mask, Subtarget, DAG))
8140 // Try inverting the insertion since for v2 masks it is easy to do and we
8141 // can't reliably sort the mask one way or the other.
8142 int InverseMask[2] = {Mask[0] ^ 2, Mask[1] ^ 2};
8143 if (SDValue Insertion = lowerVectorShuffleAsElementInsertion(
8144 DL, MVT::v2i64, V2, V1, InverseMask, Subtarget, DAG))
8147 // We have different paths for blend lowering, but they all must use the
8148 // *exact* same predicate.
8149 bool IsBlendSupported = Subtarget->hasSSE41();
8150 if (IsBlendSupported)
8151 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v2i64, V1, V2, Mask,
8155 // Use dedicated unpack instructions for masks that match their pattern.
8156 if (isShuffleEquivalent(V1, V2, Mask, {0, 2}))
8157 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v2i64, V1, V2);
8158 if (isShuffleEquivalent(V1, V2, Mask, {1, 3}))
8159 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v2i64, V1, V2);
8161 // Try to use byte rotation instructions.
8162 // Its more profitable for pre-SSSE3 to use shuffles/unpacks.
8163 if (Subtarget->hasSSSE3())
8164 if (SDValue Rotate = lowerVectorShuffleAsByteRotate(
8165 DL, MVT::v2i64, V1, V2, Mask, Subtarget, DAG))
8168 // If we have direct support for blends, we should lower by decomposing into
8169 // a permute. That will be faster than the domain cross.
8170 if (IsBlendSupported)
8171 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v2i64, V1, V2,
8174 // We implement this with SHUFPD which is pretty lame because it will likely
8175 // incur 2 cycles of stall for integer vectors on Nehalem and older chips.
8176 // However, all the alternatives are still more cycles and newer chips don't
8177 // have this problem. It would be really nice if x86 had better shuffles here.
8178 V1 = DAG.getBitcast(MVT::v2f64, V1);
8179 V2 = DAG.getBitcast(MVT::v2f64, V2);
8180 return DAG.getBitcast(MVT::v2i64,
8181 DAG.getVectorShuffle(MVT::v2f64, DL, V1, V2, Mask));
8184 /// \brief Test whether this can be lowered with a single SHUFPS instruction.
8186 /// This is used to disable more specialized lowerings when the shufps lowering
8187 /// will happen to be efficient.
8188 static bool isSingleSHUFPSMask(ArrayRef<int> Mask) {
8189 // This routine only handles 128-bit shufps.
8190 assert(Mask.size() == 4 && "Unsupported mask size!");
8192 // To lower with a single SHUFPS we need to have the low half and high half
8193 // each requiring a single input.
8194 if (Mask[0] != -1 && Mask[1] != -1 && (Mask[0] < 4) != (Mask[1] < 4))
8196 if (Mask[2] != -1 && Mask[3] != -1 && (Mask[2] < 4) != (Mask[3] < 4))
8202 /// \brief Lower a vector shuffle using the SHUFPS instruction.
8204 /// This is a helper routine dedicated to lowering vector shuffles using SHUFPS.
8205 /// It makes no assumptions about whether this is the *best* lowering, it simply
8207 static SDValue lowerVectorShuffleWithSHUFPS(SDLoc DL, MVT VT,
8208 ArrayRef<int> Mask, SDValue V1,
8209 SDValue V2, SelectionDAG &DAG) {
8210 SDValue LowV = V1, HighV = V2;
8211 int NewMask[4] = {Mask[0], Mask[1], Mask[2], Mask[3]};
8214 std::count_if(Mask.begin(), Mask.end(), [](int M) { return M >= 4; });
8216 if (NumV2Elements == 1) {
8218 std::find_if(Mask.begin(), Mask.end(), [](int M) { return M >= 4; }) -
8221 // Compute the index adjacent to V2Index and in the same half by toggling
8223 int V2AdjIndex = V2Index ^ 1;
8225 if (Mask[V2AdjIndex] == -1) {
8226 // Handles all the cases where we have a single V2 element and an undef.
8227 // This will only ever happen in the high lanes because we commute the
8228 // vector otherwise.
8230 std::swap(LowV, HighV);
8231 NewMask[V2Index] -= 4;
8233 // Handle the case where the V2 element ends up adjacent to a V1 element.
8234 // To make this work, blend them together as the first step.
8235 int V1Index = V2AdjIndex;
8236 int BlendMask[4] = {Mask[V2Index] - 4, 0, Mask[V1Index], 0};
8237 V2 = DAG.getNode(X86ISD::SHUFP, DL, VT, V2, V1,
8238 getV4X86ShuffleImm8ForMask(BlendMask, DL, DAG));
8240 // Now proceed to reconstruct the final blend as we have the necessary
8241 // high or low half formed.
8248 NewMask[V1Index] = 2; // We put the V1 element in V2[2].
8249 NewMask[V2Index] = 0; // We shifted the V2 element into V2[0].
8251 } else if (NumV2Elements == 2) {
8252 if (Mask[0] < 4 && Mask[1] < 4) {
8253 // Handle the easy case where we have V1 in the low lanes and V2 in the
8257 } else if (Mask[2] < 4 && Mask[3] < 4) {
8258 // We also handle the reversed case because this utility may get called
8259 // when we detect a SHUFPS pattern but can't easily commute the shuffle to
8260 // arrange things in the right direction.
8266 // We have a mixture of V1 and V2 in both low and high lanes. Rather than
8267 // trying to place elements directly, just blend them and set up the final
8268 // shuffle to place them.
8270 // The first two blend mask elements are for V1, the second two are for
8272 int BlendMask[4] = {Mask[0] < 4 ? Mask[0] : Mask[1],
8273 Mask[2] < 4 ? Mask[2] : Mask[3],
8274 (Mask[0] >= 4 ? Mask[0] : Mask[1]) - 4,
8275 (Mask[2] >= 4 ? Mask[2] : Mask[3]) - 4};
8276 V1 = DAG.getNode(X86ISD::SHUFP, DL, VT, V1, V2,
8277 getV4X86ShuffleImm8ForMask(BlendMask, DL, DAG));
8279 // Now we do a normal shuffle of V1 by giving V1 as both operands to
8282 NewMask[0] = Mask[0] < 4 ? 0 : 2;
8283 NewMask[1] = Mask[0] < 4 ? 2 : 0;
8284 NewMask[2] = Mask[2] < 4 ? 1 : 3;
8285 NewMask[3] = Mask[2] < 4 ? 3 : 1;
8288 return DAG.getNode(X86ISD::SHUFP, DL, VT, LowV, HighV,
8289 getV4X86ShuffleImm8ForMask(NewMask, DL, DAG));
8292 /// \brief Lower 4-lane 32-bit floating point shuffles.
8294 /// Uses instructions exclusively from the floating point unit to minimize
8295 /// domain crossing penalties, as these are sufficient to implement all v4f32
8297 static SDValue lowerV4F32VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
8298 const X86Subtarget *Subtarget,
8299 SelectionDAG &DAG) {
8301 assert(Op.getSimpleValueType() == MVT::v4f32 && "Bad shuffle type!");
8302 assert(V1.getSimpleValueType() == MVT::v4f32 && "Bad operand type!");
8303 assert(V2.getSimpleValueType() == MVT::v4f32 && "Bad operand type!");
8304 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
8305 ArrayRef<int> Mask = SVOp->getMask();
8306 assert(Mask.size() == 4 && "Unexpected mask size for v4 shuffle!");
8309 std::count_if(Mask.begin(), Mask.end(), [](int M) { return M >= 4; });
8311 if (NumV2Elements == 0) {
8312 // Check for being able to broadcast a single element.
8313 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v4f32, V1,
8314 Mask, Subtarget, DAG))
8317 // Use even/odd duplicate instructions for masks that match their pattern.
8318 if (Subtarget->hasSSE3()) {
8319 if (isShuffleEquivalent(V1, V2, Mask, {0, 0, 2, 2}))
8320 return DAG.getNode(X86ISD::MOVSLDUP, DL, MVT::v4f32, V1);
8321 if (isShuffleEquivalent(V1, V2, Mask, {1, 1, 3, 3}))
8322 return DAG.getNode(X86ISD::MOVSHDUP, DL, MVT::v4f32, V1);
8325 if (Subtarget->hasAVX()) {
8326 // If we have AVX, we can use VPERMILPS which will allow folding a load
8327 // into the shuffle.
8328 return DAG.getNode(X86ISD::VPERMILPI, DL, MVT::v4f32, V1,
8329 getV4X86ShuffleImm8ForMask(Mask, DL, DAG));
8332 // Otherwise, use a straight shuffle of a single input vector. We pass the
8333 // input vector to both operands to simulate this with a SHUFPS.
8334 return DAG.getNode(X86ISD::SHUFP, DL, MVT::v4f32, V1, V1,
8335 getV4X86ShuffleImm8ForMask(Mask, DL, DAG));
8338 // There are special ways we can lower some single-element blends. However, we
8339 // have custom ways we can lower more complex single-element blends below that
8340 // we defer to if both this and BLENDPS fail to match, so restrict this to
8341 // when the V2 input is targeting element 0 of the mask -- that is the fast
8343 if (NumV2Elements == 1 && Mask[0] >= 4)
8344 if (SDValue V = lowerVectorShuffleAsElementInsertion(DL, MVT::v4f32, V1, V2,
8345 Mask, Subtarget, DAG))
8348 if (Subtarget->hasSSE41()) {
8349 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v4f32, V1, V2, Mask,
8353 // Use INSERTPS if we can complete the shuffle efficiently.
8354 if (SDValue V = lowerVectorShuffleAsInsertPS(Op, V1, V2, Mask, DAG))
8357 if (!isSingleSHUFPSMask(Mask))
8358 if (SDValue BlendPerm = lowerVectorShuffleAsBlendAndPermute(
8359 DL, MVT::v4f32, V1, V2, Mask, DAG))
8363 // Use dedicated unpack instructions for masks that match their pattern.
8364 if (isShuffleEquivalent(V1, V2, Mask, {0, 4, 1, 5}))
8365 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v4f32, V1, V2);
8366 if (isShuffleEquivalent(V1, V2, Mask, {2, 6, 3, 7}))
8367 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v4f32, V1, V2);
8368 if (isShuffleEquivalent(V1, V2, Mask, {4, 0, 5, 1}))
8369 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v4f32, V2, V1);
8370 if (isShuffleEquivalent(V1, V2, Mask, {6, 2, 7, 3}))
8371 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v4f32, V2, V1);
8373 // Otherwise fall back to a SHUFPS lowering strategy.
8374 return lowerVectorShuffleWithSHUFPS(DL, MVT::v4f32, Mask, V1, V2, DAG);
8377 /// \brief Lower 4-lane i32 vector shuffles.
8379 /// We try to handle these with integer-domain shuffles where we can, but for
8380 /// blends we use the floating point domain blend instructions.
8381 static SDValue lowerV4I32VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
8382 const X86Subtarget *Subtarget,
8383 SelectionDAG &DAG) {
8385 assert(Op.getSimpleValueType() == MVT::v4i32 && "Bad shuffle type!");
8386 assert(V1.getSimpleValueType() == MVT::v4i32 && "Bad operand type!");
8387 assert(V2.getSimpleValueType() == MVT::v4i32 && "Bad operand type!");
8388 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
8389 ArrayRef<int> Mask = SVOp->getMask();
8390 assert(Mask.size() == 4 && "Unexpected mask size for v4 shuffle!");
8392 // Whenever we can lower this as a zext, that instruction is strictly faster
8393 // than any alternative. It also allows us to fold memory operands into the
8394 // shuffle in many cases.
8395 if (SDValue ZExt = lowerVectorShuffleAsZeroOrAnyExtend(DL, MVT::v4i32, V1, V2,
8396 Mask, Subtarget, DAG))
8400 std::count_if(Mask.begin(), Mask.end(), [](int M) { return M >= 4; });
8402 if (NumV2Elements == 0) {
8403 // Check for being able to broadcast a single element.
8404 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v4i32, V1,
8405 Mask, Subtarget, DAG))
8408 // Straight shuffle of a single input vector. For everything from SSE2
8409 // onward this has a single fast instruction with no scary immediates.
8410 // We coerce the shuffle pattern to be compatible with UNPCK instructions
8411 // but we aren't actually going to use the UNPCK instruction because doing
8412 // so prevents folding a load into this instruction or making a copy.
8413 const int UnpackLoMask[] = {0, 0, 1, 1};
8414 const int UnpackHiMask[] = {2, 2, 3, 3};
8415 if (isShuffleEquivalent(V1, V2, Mask, {0, 0, 1, 1}))
8416 Mask = UnpackLoMask;
8417 else if (isShuffleEquivalent(V1, V2, Mask, {2, 2, 3, 3}))
8418 Mask = UnpackHiMask;
8420 return DAG.getNode(X86ISD::PSHUFD, DL, MVT::v4i32, V1,
8421 getV4X86ShuffleImm8ForMask(Mask, DL, DAG));
8424 // Try to use shift instructions.
8426 lowerVectorShuffleAsShift(DL, MVT::v4i32, V1, V2, Mask, DAG))
8429 // There are special ways we can lower some single-element blends.
8430 if (NumV2Elements == 1)
8431 if (SDValue V = lowerVectorShuffleAsElementInsertion(DL, MVT::v4i32, V1, V2,
8432 Mask, Subtarget, DAG))
8435 // We have different paths for blend lowering, but they all must use the
8436 // *exact* same predicate.
8437 bool IsBlendSupported = Subtarget->hasSSE41();
8438 if (IsBlendSupported)
8439 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v4i32, V1, V2, Mask,
8443 if (SDValue Masked =
8444 lowerVectorShuffleAsBitMask(DL, MVT::v4i32, V1, V2, Mask, DAG))
8447 // Use dedicated unpack instructions for masks that match their pattern.
8448 if (isShuffleEquivalent(V1, V2, Mask, {0, 4, 1, 5}))
8449 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v4i32, V1, V2);
8450 if (isShuffleEquivalent(V1, V2, Mask, {2, 6, 3, 7}))
8451 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v4i32, V1, V2);
8452 if (isShuffleEquivalent(V1, V2, Mask, {4, 0, 5, 1}))
8453 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v4i32, V2, V1);
8454 if (isShuffleEquivalent(V1, V2, Mask, {6, 2, 7, 3}))
8455 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v4i32, V2, V1);
8457 // Try to use byte rotation instructions.
8458 // Its more profitable for pre-SSSE3 to use shuffles/unpacks.
8459 if (Subtarget->hasSSSE3())
8460 if (SDValue Rotate = lowerVectorShuffleAsByteRotate(
8461 DL, MVT::v4i32, V1, V2, Mask, Subtarget, DAG))
8464 // If we have direct support for blends, we should lower by decomposing into
8465 // a permute. That will be faster than the domain cross.
8466 if (IsBlendSupported)
8467 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v4i32, V1, V2,
8470 // Try to lower by permuting the inputs into an unpack instruction.
8471 if (SDValue Unpack = lowerVectorShuffleAsPermuteAndUnpack(DL, MVT::v4i32, V1,
8475 // We implement this with SHUFPS because it can blend from two vectors.
8476 // Because we're going to eventually use SHUFPS, we use SHUFPS even to build
8477 // up the inputs, bypassing domain shift penalties that we would encur if we
8478 // directly used PSHUFD on Nehalem and older. For newer chips, this isn't
8480 return DAG.getBitcast(
8482 DAG.getVectorShuffle(MVT::v4f32, DL, DAG.getBitcast(MVT::v4f32, V1),
8483 DAG.getBitcast(MVT::v4f32, V2), Mask));
8486 /// \brief Lowering of single-input v8i16 shuffles is the cornerstone of SSE2
8487 /// shuffle lowering, and the most complex part.
8489 /// The lowering strategy is to try to form pairs of input lanes which are
8490 /// targeted at the same half of the final vector, and then use a dword shuffle
8491 /// to place them onto the right half, and finally unpack the paired lanes into
8492 /// their final position.
8494 /// The exact breakdown of how to form these dword pairs and align them on the
8495 /// correct sides is really tricky. See the comments within the function for
8496 /// more of the details.
8498 /// This code also handles repeated 128-bit lanes of v8i16 shuffles, but each
8499 /// lane must shuffle the *exact* same way. In fact, you must pass a v8 Mask to
8500 /// this routine for it to work correctly. To shuffle a 256-bit or 512-bit i16
8501 /// vector, form the analogous 128-bit 8-element Mask.
8502 static SDValue lowerV8I16GeneralSingleInputVectorShuffle(
8503 SDLoc DL, MVT VT, SDValue V, MutableArrayRef<int> Mask,
8504 const X86Subtarget *Subtarget, SelectionDAG &DAG) {
8505 assert(VT.getScalarType() == MVT::i16 && "Bad input type!");
8506 MVT PSHUFDVT = MVT::getVectorVT(MVT::i32, VT.getVectorNumElements() / 2);
8508 assert(Mask.size() == 8 && "Shuffle mask length doen't match!");
8509 MutableArrayRef<int> LoMask = Mask.slice(0, 4);
8510 MutableArrayRef<int> HiMask = Mask.slice(4, 4);
8512 SmallVector<int, 4> LoInputs;
8513 std::copy_if(LoMask.begin(), LoMask.end(), std::back_inserter(LoInputs),
8514 [](int M) { return M >= 0; });
8515 std::sort(LoInputs.begin(), LoInputs.end());
8516 LoInputs.erase(std::unique(LoInputs.begin(), LoInputs.end()), LoInputs.end());
8517 SmallVector<int, 4> HiInputs;
8518 std::copy_if(HiMask.begin(), HiMask.end(), std::back_inserter(HiInputs),
8519 [](int M) { return M >= 0; });
8520 std::sort(HiInputs.begin(), HiInputs.end());
8521 HiInputs.erase(std::unique(HiInputs.begin(), HiInputs.end()), HiInputs.end());
8523 std::lower_bound(LoInputs.begin(), LoInputs.end(), 4) - LoInputs.begin();
8524 int NumHToL = LoInputs.size() - NumLToL;
8526 std::lower_bound(HiInputs.begin(), HiInputs.end(), 4) - HiInputs.begin();
8527 int NumHToH = HiInputs.size() - NumLToH;
8528 MutableArrayRef<int> LToLInputs(LoInputs.data(), NumLToL);
8529 MutableArrayRef<int> LToHInputs(HiInputs.data(), NumLToH);
8530 MutableArrayRef<int> HToLInputs(LoInputs.data() + NumLToL, NumHToL);
8531 MutableArrayRef<int> HToHInputs(HiInputs.data() + NumLToH, NumHToH);
8533 // Simplify the 1-into-3 and 3-into-1 cases with a single pshufd. For all
8534 // such inputs we can swap two of the dwords across the half mark and end up
8535 // with <=2 inputs to each half in each half. Once there, we can fall through
8536 // to the generic code below. For example:
8538 // Input: [a, b, c, d, e, f, g, h] -PSHUFD[0,2,1,3]-> [a, b, e, f, c, d, g, h]
8539 // Mask: [0, 1, 2, 7, 4, 5, 6, 3] -----------------> [0, 1, 4, 7, 2, 3, 6, 5]
8541 // However in some very rare cases we have a 1-into-3 or 3-into-1 on one half
8542 // and an existing 2-into-2 on the other half. In this case we may have to
8543 // pre-shuffle the 2-into-2 half to avoid turning it into a 3-into-1 or
8544 // 1-into-3 which could cause us to cycle endlessly fixing each side in turn.
8545 // Fortunately, we don't have to handle anything but a 2-into-2 pattern
8546 // because any other situation (including a 3-into-1 or 1-into-3 in the other
8547 // half than the one we target for fixing) will be fixed when we re-enter this
8548 // path. We will also combine away any sequence of PSHUFD instructions that
8549 // result into a single instruction. Here is an example of the tricky case:
8551 // Input: [a, b, c, d, e, f, g, h] -PSHUFD[0,2,1,3]-> [a, b, e, f, c, d, g, h]
8552 // Mask: [3, 7, 1, 0, 2, 7, 3, 5] -THIS-IS-BAD!!!!-> [5, 7, 1, 0, 4, 7, 5, 3]
8554 // This now has a 1-into-3 in the high half! Instead, we do two shuffles:
8556 // Input: [a, b, c, d, e, f, g, h] PSHUFHW[0,2,1,3]-> [a, b, c, d, e, g, f, h]
8557 // Mask: [3, 7, 1, 0, 2, 7, 3, 5] -----------------> [3, 7, 1, 0, 2, 7, 3, 6]
8559 // Input: [a, b, c, d, e, g, f, h] -PSHUFD[0,2,1,3]-> [a, b, e, g, c, d, f, h]
8560 // Mask: [3, 7, 1, 0, 2, 7, 3, 6] -----------------> [5, 7, 1, 0, 4, 7, 5, 6]
8562 // The result is fine to be handled by the generic logic.
8563 auto balanceSides = [&](ArrayRef<int> AToAInputs, ArrayRef<int> BToAInputs,
8564 ArrayRef<int> BToBInputs, ArrayRef<int> AToBInputs,
8565 int AOffset, int BOffset) {
8566 assert((AToAInputs.size() == 3 || AToAInputs.size() == 1) &&
8567 "Must call this with A having 3 or 1 inputs from the A half.");
8568 assert((BToAInputs.size() == 1 || BToAInputs.size() == 3) &&
8569 "Must call this with B having 1 or 3 inputs from the B half.");
8570 assert(AToAInputs.size() + BToAInputs.size() == 4 &&
8571 "Must call this with either 3:1 or 1:3 inputs (summing to 4).");
8573 bool ThreeAInputs = AToAInputs.size() == 3;
8575 // Compute the index of dword with only one word among the three inputs in
8576 // a half by taking the sum of the half with three inputs and subtracting
8577 // the sum of the actual three inputs. The difference is the remaining
8580 int &TripleDWord = ThreeAInputs ? ADWord : BDWord;
8581 int &OneInputDWord = ThreeAInputs ? BDWord : ADWord;
8582 int TripleInputOffset = ThreeAInputs ? AOffset : BOffset;
8583 ArrayRef<int> TripleInputs = ThreeAInputs ? AToAInputs : BToAInputs;
8584 int OneInput = ThreeAInputs ? BToAInputs[0] : AToAInputs[0];
8585 int TripleInputSum = 0 + 1 + 2 + 3 + (4 * TripleInputOffset);
8586 int TripleNonInputIdx =
8587 TripleInputSum - std::accumulate(TripleInputs.begin(), TripleInputs.end(), 0);
8588 TripleDWord = TripleNonInputIdx / 2;
8590 // We use xor with one to compute the adjacent DWord to whichever one the
8592 OneInputDWord = (OneInput / 2) ^ 1;
8594 // Check for one tricky case: We're fixing a 3<-1 or a 1<-3 shuffle for AToA
8595 // and BToA inputs. If there is also such a problem with the BToB and AToB
8596 // inputs, we don't try to fix it necessarily -- we'll recurse and see it in
8597 // the next pass. However, if we have a 2<-2 in the BToB and AToB inputs, it
8598 // is essential that we don't *create* a 3<-1 as then we might oscillate.
8599 if (BToBInputs.size() == 2 && AToBInputs.size() == 2) {
8600 // Compute how many inputs will be flipped by swapping these DWords. We
8602 // to balance this to ensure we don't form a 3-1 shuffle in the other
8604 int NumFlippedAToBInputs =
8605 std::count(AToBInputs.begin(), AToBInputs.end(), 2 * ADWord) +
8606 std::count(AToBInputs.begin(), AToBInputs.end(), 2 * ADWord + 1);
8607 int NumFlippedBToBInputs =
8608 std::count(BToBInputs.begin(), BToBInputs.end(), 2 * BDWord) +
8609 std::count(BToBInputs.begin(), BToBInputs.end(), 2 * BDWord + 1);
8610 if ((NumFlippedAToBInputs == 1 &&
8611 (NumFlippedBToBInputs == 0 || NumFlippedBToBInputs == 2)) ||
8612 (NumFlippedBToBInputs == 1 &&
8613 (NumFlippedAToBInputs == 0 || NumFlippedAToBInputs == 2))) {
8614 // We choose whether to fix the A half or B half based on whether that
8615 // half has zero flipped inputs. At zero, we may not be able to fix it
8616 // with that half. We also bias towards fixing the B half because that
8617 // will more commonly be the high half, and we have to bias one way.
8618 auto FixFlippedInputs = [&V, &DL, &Mask, &DAG](int PinnedIdx, int DWord,
8619 ArrayRef<int> Inputs) {
8620 int FixIdx = PinnedIdx ^ 1; // The adjacent slot to the pinned slot.
8621 bool IsFixIdxInput = std::find(Inputs.begin(), Inputs.end(),
8622 PinnedIdx ^ 1) != Inputs.end();
8623 // Determine whether the free index is in the flipped dword or the
8624 // unflipped dword based on where the pinned index is. We use this bit
8625 // in an xor to conditionally select the adjacent dword.
8626 int FixFreeIdx = 2 * (DWord ^ (PinnedIdx / 2 == DWord));
8627 bool IsFixFreeIdxInput = std::find(Inputs.begin(), Inputs.end(),
8628 FixFreeIdx) != Inputs.end();
8629 if (IsFixIdxInput == IsFixFreeIdxInput)
8631 IsFixFreeIdxInput = std::find(Inputs.begin(), Inputs.end(),
8632 FixFreeIdx) != Inputs.end();
8633 assert(IsFixIdxInput != IsFixFreeIdxInput &&
8634 "We need to be changing the number of flipped inputs!");
8635 int PSHUFHalfMask[] = {0, 1, 2, 3};
8636 std::swap(PSHUFHalfMask[FixFreeIdx % 4], PSHUFHalfMask[FixIdx % 4]);
8637 V = DAG.getNode(FixIdx < 4 ? X86ISD::PSHUFLW : X86ISD::PSHUFHW, DL,
8639 getV4X86ShuffleImm8ForMask(PSHUFHalfMask, DL, DAG));
8642 if (M != -1 && M == FixIdx)
8644 else if (M != -1 && M == FixFreeIdx)
8647 if (NumFlippedBToBInputs != 0) {
8649 BToAInputs.size() == 3 ? TripleNonInputIdx : OneInput;
8650 FixFlippedInputs(BPinnedIdx, BDWord, BToBInputs);
8652 assert(NumFlippedAToBInputs != 0 && "Impossible given predicates!");
8653 int APinnedIdx = ThreeAInputs ? TripleNonInputIdx : OneInput;
8654 FixFlippedInputs(APinnedIdx, ADWord, AToBInputs);
8659 int PSHUFDMask[] = {0, 1, 2, 3};
8660 PSHUFDMask[ADWord] = BDWord;
8661 PSHUFDMask[BDWord] = ADWord;
8664 DAG.getNode(X86ISD::PSHUFD, DL, PSHUFDVT, DAG.getBitcast(PSHUFDVT, V),
8665 getV4X86ShuffleImm8ForMask(PSHUFDMask, DL, DAG)));
8667 // Adjust the mask to match the new locations of A and B.
8669 if (M != -1 && M/2 == ADWord)
8670 M = 2 * BDWord + M % 2;
8671 else if (M != -1 && M/2 == BDWord)
8672 M = 2 * ADWord + M % 2;
8674 // Recurse back into this routine to re-compute state now that this isn't
8675 // a 3 and 1 problem.
8676 return lowerV8I16GeneralSingleInputVectorShuffle(DL, VT, V, Mask, Subtarget,
8679 if ((NumLToL == 3 && NumHToL == 1) || (NumLToL == 1 && NumHToL == 3))
8680 return balanceSides(LToLInputs, HToLInputs, HToHInputs, LToHInputs, 0, 4);
8681 else if ((NumHToH == 3 && NumLToH == 1) || (NumHToH == 1 && NumLToH == 3))
8682 return balanceSides(HToHInputs, LToHInputs, LToLInputs, HToLInputs, 4, 0);
8684 // At this point there are at most two inputs to the low and high halves from
8685 // each half. That means the inputs can always be grouped into dwords and
8686 // those dwords can then be moved to the correct half with a dword shuffle.
8687 // We use at most one low and one high word shuffle to collect these paired
8688 // inputs into dwords, and finally a dword shuffle to place them.
8689 int PSHUFLMask[4] = {-1, -1, -1, -1};
8690 int PSHUFHMask[4] = {-1, -1, -1, -1};
8691 int PSHUFDMask[4] = {-1, -1, -1, -1};
8693 // First fix the masks for all the inputs that are staying in their
8694 // original halves. This will then dictate the targets of the cross-half
8696 auto fixInPlaceInputs =
8697 [&PSHUFDMask](ArrayRef<int> InPlaceInputs, ArrayRef<int> IncomingInputs,
8698 MutableArrayRef<int> SourceHalfMask,
8699 MutableArrayRef<int> HalfMask, int HalfOffset) {
8700 if (InPlaceInputs.empty())
8702 if (InPlaceInputs.size() == 1) {
8703 SourceHalfMask[InPlaceInputs[0] - HalfOffset] =
8704 InPlaceInputs[0] - HalfOffset;
8705 PSHUFDMask[InPlaceInputs[0] / 2] = InPlaceInputs[0] / 2;
8708 if (IncomingInputs.empty()) {
8709 // Just fix all of the in place inputs.
8710 for (int Input : InPlaceInputs) {
8711 SourceHalfMask[Input - HalfOffset] = Input - HalfOffset;
8712 PSHUFDMask[Input / 2] = Input / 2;
8717 assert(InPlaceInputs.size() == 2 && "Cannot handle 3 or 4 inputs!");
8718 SourceHalfMask[InPlaceInputs[0] - HalfOffset] =
8719 InPlaceInputs[0] - HalfOffset;
8720 // Put the second input next to the first so that they are packed into
8721 // a dword. We find the adjacent index by toggling the low bit.
8722 int AdjIndex = InPlaceInputs[0] ^ 1;
8723 SourceHalfMask[AdjIndex - HalfOffset] = InPlaceInputs[1] - HalfOffset;
8724 std::replace(HalfMask.begin(), HalfMask.end(), InPlaceInputs[1], AdjIndex);
8725 PSHUFDMask[AdjIndex / 2] = AdjIndex / 2;
8727 fixInPlaceInputs(LToLInputs, HToLInputs, PSHUFLMask, LoMask, 0);
8728 fixInPlaceInputs(HToHInputs, LToHInputs, PSHUFHMask, HiMask, 4);
8730 // Now gather the cross-half inputs and place them into a free dword of
8731 // their target half.
8732 // FIXME: This operation could almost certainly be simplified dramatically to
8733 // look more like the 3-1 fixing operation.
8734 auto moveInputsToRightHalf = [&PSHUFDMask](
8735 MutableArrayRef<int> IncomingInputs, ArrayRef<int> ExistingInputs,
8736 MutableArrayRef<int> SourceHalfMask, MutableArrayRef<int> HalfMask,
8737 MutableArrayRef<int> FinalSourceHalfMask, int SourceOffset,
8739 auto isWordClobbered = [](ArrayRef<int> SourceHalfMask, int Word) {
8740 return SourceHalfMask[Word] != -1 && SourceHalfMask[Word] != Word;
8742 auto isDWordClobbered = [&isWordClobbered](ArrayRef<int> SourceHalfMask,
8744 int LowWord = Word & ~1;
8745 int HighWord = Word | 1;
8746 return isWordClobbered(SourceHalfMask, LowWord) ||
8747 isWordClobbered(SourceHalfMask, HighWord);
8750 if (IncomingInputs.empty())
8753 if (ExistingInputs.empty()) {
8754 // Map any dwords with inputs from them into the right half.
8755 for (int Input : IncomingInputs) {
8756 // If the source half mask maps over the inputs, turn those into
8757 // swaps and use the swapped lane.
8758 if (isWordClobbered(SourceHalfMask, Input - SourceOffset)) {
8759 if (SourceHalfMask[SourceHalfMask[Input - SourceOffset]] == -1) {
8760 SourceHalfMask[SourceHalfMask[Input - SourceOffset]] =
8761 Input - SourceOffset;
8762 // We have to swap the uses in our half mask in one sweep.
8763 for (int &M : HalfMask)
8764 if (M == SourceHalfMask[Input - SourceOffset] + SourceOffset)
8766 else if (M == Input)
8767 M = SourceHalfMask[Input - SourceOffset] + SourceOffset;
8769 assert(SourceHalfMask[SourceHalfMask[Input - SourceOffset]] ==
8770 Input - SourceOffset &&
8771 "Previous placement doesn't match!");
8773 // Note that this correctly re-maps both when we do a swap and when
8774 // we observe the other side of the swap above. We rely on that to
8775 // avoid swapping the members of the input list directly.
8776 Input = SourceHalfMask[Input - SourceOffset] + SourceOffset;
8779 // Map the input's dword into the correct half.
8780 if (PSHUFDMask[(Input - SourceOffset + DestOffset) / 2] == -1)
8781 PSHUFDMask[(Input - SourceOffset + DestOffset) / 2] = Input / 2;
8783 assert(PSHUFDMask[(Input - SourceOffset + DestOffset) / 2] ==
8785 "Previous placement doesn't match!");
8788 // And just directly shift any other-half mask elements to be same-half
8789 // as we will have mirrored the dword containing the element into the
8790 // same position within that half.
8791 for (int &M : HalfMask)
8792 if (M >= SourceOffset && M < SourceOffset + 4) {
8793 M = M - SourceOffset + DestOffset;
8794 assert(M >= 0 && "This should never wrap below zero!");
8799 // Ensure we have the input in a viable dword of its current half. This
8800 // is particularly tricky because the original position may be clobbered
8801 // by inputs being moved and *staying* in that half.
8802 if (IncomingInputs.size() == 1) {
8803 if (isWordClobbered(SourceHalfMask, IncomingInputs[0] - SourceOffset)) {
8804 int InputFixed = std::find(std::begin(SourceHalfMask),
8805 std::end(SourceHalfMask), -1) -
8806 std::begin(SourceHalfMask) + SourceOffset;
8807 SourceHalfMask[InputFixed - SourceOffset] =
8808 IncomingInputs[0] - SourceOffset;
8809 std::replace(HalfMask.begin(), HalfMask.end(), IncomingInputs[0],
8811 IncomingInputs[0] = InputFixed;
8813 } else if (IncomingInputs.size() == 2) {
8814 if (IncomingInputs[0] / 2 != IncomingInputs[1] / 2 ||
8815 isDWordClobbered(SourceHalfMask, IncomingInputs[0] - SourceOffset)) {
8816 // We have two non-adjacent or clobbered inputs we need to extract from
8817 // the source half. To do this, we need to map them into some adjacent
8818 // dword slot in the source mask.
8819 int InputsFixed[2] = {IncomingInputs[0] - SourceOffset,
8820 IncomingInputs[1] - SourceOffset};
8822 // If there is a free slot in the source half mask adjacent to one of
8823 // the inputs, place the other input in it. We use (Index XOR 1) to
8824 // compute an adjacent index.
8825 if (!isWordClobbered(SourceHalfMask, InputsFixed[0]) &&
8826 SourceHalfMask[InputsFixed[0] ^ 1] == -1) {
8827 SourceHalfMask[InputsFixed[0]] = InputsFixed[0];
8828 SourceHalfMask[InputsFixed[0] ^ 1] = InputsFixed[1];
8829 InputsFixed[1] = InputsFixed[0] ^ 1;
8830 } else if (!isWordClobbered(SourceHalfMask, InputsFixed[1]) &&
8831 SourceHalfMask[InputsFixed[1] ^ 1] == -1) {
8832 SourceHalfMask[InputsFixed[1]] = InputsFixed[1];
8833 SourceHalfMask[InputsFixed[1] ^ 1] = InputsFixed[0];
8834 InputsFixed[0] = InputsFixed[1] ^ 1;
8835 } else if (SourceHalfMask[2 * ((InputsFixed[0] / 2) ^ 1)] == -1 &&
8836 SourceHalfMask[2 * ((InputsFixed[0] / 2) ^ 1) + 1] == -1) {
8837 // The two inputs are in the same DWord but it is clobbered and the
8838 // adjacent DWord isn't used at all. Move both inputs to the free
8840 SourceHalfMask[2 * ((InputsFixed[0] / 2) ^ 1)] = InputsFixed[0];
8841 SourceHalfMask[2 * ((InputsFixed[0] / 2) ^ 1) + 1] = InputsFixed[1];
8842 InputsFixed[0] = 2 * ((InputsFixed[0] / 2) ^ 1);
8843 InputsFixed[1] = 2 * ((InputsFixed[0] / 2) ^ 1) + 1;
8845 // The only way we hit this point is if there is no clobbering
8846 // (because there are no off-half inputs to this half) and there is no
8847 // free slot adjacent to one of the inputs. In this case, we have to
8848 // swap an input with a non-input.
8849 for (int i = 0; i < 4; ++i)
8850 assert((SourceHalfMask[i] == -1 || SourceHalfMask[i] == i) &&
8851 "We can't handle any clobbers here!");
8852 assert(InputsFixed[1] != (InputsFixed[0] ^ 1) &&
8853 "Cannot have adjacent inputs here!");
8855 SourceHalfMask[InputsFixed[0] ^ 1] = InputsFixed[1];
8856 SourceHalfMask[InputsFixed[1]] = InputsFixed[0] ^ 1;
8858 // We also have to update the final source mask in this case because
8859 // it may need to undo the above swap.
8860 for (int &M : FinalSourceHalfMask)
8861 if (M == (InputsFixed[0] ^ 1) + SourceOffset)
8862 M = InputsFixed[1] + SourceOffset;
8863 else if (M == InputsFixed[1] + SourceOffset)
8864 M = (InputsFixed[0] ^ 1) + SourceOffset;
8866 InputsFixed[1] = InputsFixed[0] ^ 1;
8869 // Point everything at the fixed inputs.
8870 for (int &M : HalfMask)
8871 if (M == IncomingInputs[0])
8872 M = InputsFixed[0] + SourceOffset;
8873 else if (M == IncomingInputs[1])
8874 M = InputsFixed[1] + SourceOffset;
8876 IncomingInputs[0] = InputsFixed[0] + SourceOffset;
8877 IncomingInputs[1] = InputsFixed[1] + SourceOffset;
8880 llvm_unreachable("Unhandled input size!");
8883 // Now hoist the DWord down to the right half.
8884 int FreeDWord = (PSHUFDMask[DestOffset / 2] == -1 ? 0 : 1) + DestOffset / 2;
8885 assert(PSHUFDMask[FreeDWord] == -1 && "DWord not free");
8886 PSHUFDMask[FreeDWord] = IncomingInputs[0] / 2;
8887 for (int &M : HalfMask)
8888 for (int Input : IncomingInputs)
8890 M = FreeDWord * 2 + Input % 2;
8892 moveInputsToRightHalf(HToLInputs, LToLInputs, PSHUFHMask, LoMask, HiMask,
8893 /*SourceOffset*/ 4, /*DestOffset*/ 0);
8894 moveInputsToRightHalf(LToHInputs, HToHInputs, PSHUFLMask, HiMask, LoMask,
8895 /*SourceOffset*/ 0, /*DestOffset*/ 4);
8897 // Now enact all the shuffles we've computed to move the inputs into their
8899 if (!isNoopShuffleMask(PSHUFLMask))
8900 V = DAG.getNode(X86ISD::PSHUFLW, DL, VT, V,
8901 getV4X86ShuffleImm8ForMask(PSHUFLMask, DL, DAG));
8902 if (!isNoopShuffleMask(PSHUFHMask))
8903 V = DAG.getNode(X86ISD::PSHUFHW, DL, VT, V,
8904 getV4X86ShuffleImm8ForMask(PSHUFHMask, DL, DAG));
8905 if (!isNoopShuffleMask(PSHUFDMask))
8908 DAG.getNode(X86ISD::PSHUFD, DL, PSHUFDVT, DAG.getBitcast(PSHUFDVT, V),
8909 getV4X86ShuffleImm8ForMask(PSHUFDMask, DL, DAG)));
8911 // At this point, each half should contain all its inputs, and we can then
8912 // just shuffle them into their final position.
8913 assert(std::count_if(LoMask.begin(), LoMask.end(),
8914 [](int M) { return M >= 4; }) == 0 &&
8915 "Failed to lift all the high half inputs to the low mask!");
8916 assert(std::count_if(HiMask.begin(), HiMask.end(),
8917 [](int M) { return M >= 0 && M < 4; }) == 0 &&
8918 "Failed to lift all the low half inputs to the high mask!");
8920 // Do a half shuffle for the low mask.
8921 if (!isNoopShuffleMask(LoMask))
8922 V = DAG.getNode(X86ISD::PSHUFLW, DL, VT, V,
8923 getV4X86ShuffleImm8ForMask(LoMask, DL, DAG));
8925 // Do a half shuffle with the high mask after shifting its values down.
8926 for (int &M : HiMask)
8929 if (!isNoopShuffleMask(HiMask))
8930 V = DAG.getNode(X86ISD::PSHUFHW, DL, VT, V,
8931 getV4X86ShuffleImm8ForMask(HiMask, DL, DAG));
8936 /// \brief Helper to form a PSHUFB-based shuffle+blend.
8937 static SDValue lowerVectorShuffleAsPSHUFB(SDLoc DL, MVT VT, SDValue V1,
8938 SDValue V2, ArrayRef<int> Mask,
8939 SelectionDAG &DAG, bool &V1InUse,
8941 SmallBitVector Zeroable = computeZeroableShuffleElements(Mask, V1, V2);
8947 int Size = Mask.size();
8948 int Scale = 16 / Size;
8949 for (int i = 0; i < 16; ++i) {
8950 if (Mask[i / Scale] == -1) {
8951 V1Mask[i] = V2Mask[i] = DAG.getUNDEF(MVT::i8);
8953 const int ZeroMask = 0x80;
8954 int V1Idx = Mask[i / Scale] < Size ? Mask[i / Scale] * Scale + i % Scale
8956 int V2Idx = Mask[i / Scale] < Size
8958 : (Mask[i / Scale] - Size) * Scale + i % Scale;
8959 if (Zeroable[i / Scale])
8960 V1Idx = V2Idx = ZeroMask;
8961 V1Mask[i] = DAG.getConstant(V1Idx, DL, MVT::i8);
8962 V2Mask[i] = DAG.getConstant(V2Idx, DL, MVT::i8);
8963 V1InUse |= (ZeroMask != V1Idx);
8964 V2InUse |= (ZeroMask != V2Idx);
8969 V1 = DAG.getNode(X86ISD::PSHUFB, DL, MVT::v16i8,
8970 DAG.getBitcast(MVT::v16i8, V1),
8971 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v16i8, V1Mask));
8973 V2 = DAG.getNode(X86ISD::PSHUFB, DL, MVT::v16i8,
8974 DAG.getBitcast(MVT::v16i8, V2),
8975 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v16i8, V2Mask));
8977 // If we need shuffled inputs from both, blend the two.
8979 if (V1InUse && V2InUse)
8980 V = DAG.getNode(ISD::OR, DL, MVT::v16i8, V1, V2);
8982 V = V1InUse ? V1 : V2;
8984 // Cast the result back to the correct type.
8985 return DAG.getBitcast(VT, V);
8988 /// \brief Generic lowering of 8-lane i16 shuffles.
8990 /// This handles both single-input shuffles and combined shuffle/blends with
8991 /// two inputs. The single input shuffles are immediately delegated to
8992 /// a dedicated lowering routine.
8994 /// The blends are lowered in one of three fundamental ways. If there are few
8995 /// enough inputs, it delegates to a basic UNPCK-based strategy. If the shuffle
8996 /// of the input is significantly cheaper when lowered as an interleaving of
8997 /// the two inputs, try to interleave them. Otherwise, blend the low and high
8998 /// halves of the inputs separately (making them have relatively few inputs)
8999 /// and then concatenate them.
9000 static SDValue lowerV8I16VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
9001 const X86Subtarget *Subtarget,
9002 SelectionDAG &DAG) {
9004 assert(Op.getSimpleValueType() == MVT::v8i16 && "Bad shuffle type!");
9005 assert(V1.getSimpleValueType() == MVT::v8i16 && "Bad operand type!");
9006 assert(V2.getSimpleValueType() == MVT::v8i16 && "Bad operand type!");
9007 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
9008 ArrayRef<int> OrigMask = SVOp->getMask();
9009 int MaskStorage[8] = {OrigMask[0], OrigMask[1], OrigMask[2], OrigMask[3],
9010 OrigMask[4], OrigMask[5], OrigMask[6], OrigMask[7]};
9011 MutableArrayRef<int> Mask(MaskStorage);
9013 assert(Mask.size() == 8 && "Unexpected mask size for v8 shuffle!");
9015 // Whenever we can lower this as a zext, that instruction is strictly faster
9016 // than any alternative.
9017 if (SDValue ZExt = lowerVectorShuffleAsZeroOrAnyExtend(
9018 DL, MVT::v8i16, V1, V2, OrigMask, Subtarget, DAG))
9021 auto isV1 = [](int M) { return M >= 0 && M < 8; };
9023 auto isV2 = [](int M) { return M >= 8; };
9025 int NumV2Inputs = std::count_if(Mask.begin(), Mask.end(), isV2);
9027 if (NumV2Inputs == 0) {
9028 // Check for being able to broadcast a single element.
9029 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v8i16, V1,
9030 Mask, Subtarget, DAG))
9033 // Try to use shift instructions.
9035 lowerVectorShuffleAsShift(DL, MVT::v8i16, V1, V1, Mask, DAG))
9038 // Use dedicated unpack instructions for masks that match their pattern.
9039 if (isShuffleEquivalent(V1, V1, Mask, {0, 0, 1, 1, 2, 2, 3, 3}))
9040 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v8i16, V1, V1);
9041 if (isShuffleEquivalent(V1, V1, Mask, {4, 4, 5, 5, 6, 6, 7, 7}))
9042 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v8i16, V1, V1);
9044 // Try to use byte rotation instructions.
9045 if (SDValue Rotate = lowerVectorShuffleAsByteRotate(DL, MVT::v8i16, V1, V1,
9046 Mask, Subtarget, DAG))
9049 return lowerV8I16GeneralSingleInputVectorShuffle(DL, MVT::v8i16, V1, Mask,
9053 assert(std::any_of(Mask.begin(), Mask.end(), isV1) &&
9054 "All single-input shuffles should be canonicalized to be V1-input "
9057 // Try to use shift instructions.
9059 lowerVectorShuffleAsShift(DL, MVT::v8i16, V1, V2, Mask, DAG))
9062 // See if we can use SSE4A Extraction / Insertion.
9063 if (Subtarget->hasSSE4A())
9064 if (SDValue V = lowerVectorShuffleWithSSE4A(DL, MVT::v8i16, V1, V2, Mask, DAG))
9067 // There are special ways we can lower some single-element blends.
9068 if (NumV2Inputs == 1)
9069 if (SDValue V = lowerVectorShuffleAsElementInsertion(DL, MVT::v8i16, V1, V2,
9070 Mask, Subtarget, DAG))
9073 // We have different paths for blend lowering, but they all must use the
9074 // *exact* same predicate.
9075 bool IsBlendSupported = Subtarget->hasSSE41();
9076 if (IsBlendSupported)
9077 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v8i16, V1, V2, Mask,
9081 if (SDValue Masked =
9082 lowerVectorShuffleAsBitMask(DL, MVT::v8i16, V1, V2, Mask, DAG))
9085 // Use dedicated unpack instructions for masks that match their pattern.
9086 if (isShuffleEquivalent(V1, V2, Mask, {0, 8, 1, 9, 2, 10, 3, 11}))
9087 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v8i16, V1, V2);
9088 if (isShuffleEquivalent(V1, V2, Mask, {4, 12, 5, 13, 6, 14, 7, 15}))
9089 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v8i16, V1, V2);
9091 // Try to use byte rotation instructions.
9092 if (SDValue Rotate = lowerVectorShuffleAsByteRotate(
9093 DL, MVT::v8i16, V1, V2, Mask, Subtarget, DAG))
9096 if (SDValue BitBlend =
9097 lowerVectorShuffleAsBitBlend(DL, MVT::v8i16, V1, V2, Mask, DAG))
9100 if (SDValue Unpack = lowerVectorShuffleAsPermuteAndUnpack(DL, MVT::v8i16, V1,
9104 // If we can't directly blend but can use PSHUFB, that will be better as it
9105 // can both shuffle and set up the inefficient blend.
9106 if (!IsBlendSupported && Subtarget->hasSSSE3()) {
9107 bool V1InUse, V2InUse;
9108 return lowerVectorShuffleAsPSHUFB(DL, MVT::v8i16, V1, V2, Mask, DAG,
9112 // We can always bit-blend if we have to so the fallback strategy is to
9113 // decompose into single-input permutes and blends.
9114 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v8i16, V1, V2,
9118 /// \brief Check whether a compaction lowering can be done by dropping even
9119 /// elements and compute how many times even elements must be dropped.
9121 /// This handles shuffles which take every Nth element where N is a power of
9122 /// two. Example shuffle masks:
9124 /// N = 1: 0, 2, 4, 6, 8, 10, 12, 14, 0, 2, 4, 6, 8, 10, 12, 14
9125 /// N = 1: 0, 2, 4, 6, 8, 10, 12, 14, 16, 18, 20, 22, 24, 26, 28, 30
9126 /// N = 2: 0, 4, 8, 12, 0, 4, 8, 12, 0, 4, 8, 12, 0, 4, 8, 12
9127 /// N = 2: 0, 4, 8, 12, 16, 20, 24, 28, 0, 4, 8, 12, 16, 20, 24, 28
9128 /// N = 3: 0, 8, 0, 8, 0, 8, 0, 8, 0, 8, 0, 8, 0, 8, 0, 8
9129 /// N = 3: 0, 8, 16, 24, 0, 8, 16, 24, 0, 8, 16, 24, 0, 8, 16, 24
9131 /// Any of these lanes can of course be undef.
9133 /// This routine only supports N <= 3.
9134 /// FIXME: Evaluate whether either AVX or AVX-512 have any opportunities here
9137 /// \returns N above, or the number of times even elements must be dropped if
9138 /// there is such a number. Otherwise returns zero.
9139 static int canLowerByDroppingEvenElements(ArrayRef<int> Mask) {
9140 // Figure out whether we're looping over two inputs or just one.
9141 bool IsSingleInput = isSingleInputShuffleMask(Mask);
9143 // The modulus for the shuffle vector entries is based on whether this is
9144 // a single input or not.
9145 int ShuffleModulus = Mask.size() * (IsSingleInput ? 1 : 2);
9146 assert(isPowerOf2_32((uint32_t)ShuffleModulus) &&
9147 "We should only be called with masks with a power-of-2 size!");
9149 uint64_t ModMask = (uint64_t)ShuffleModulus - 1;
9151 // We track whether the input is viable for all power-of-2 strides 2^1, 2^2,
9152 // and 2^3 simultaneously. This is because we may have ambiguity with
9153 // partially undef inputs.
9154 bool ViableForN[3] = {true, true, true};
9156 for (int i = 0, e = Mask.size(); i < e; ++i) {
9157 // Ignore undef lanes, we'll optimistically collapse them to the pattern we
9162 bool IsAnyViable = false;
9163 for (unsigned j = 0; j != array_lengthof(ViableForN); ++j)
9164 if (ViableForN[j]) {
9167 // The shuffle mask must be equal to (i * 2^N) % M.
9168 if ((uint64_t)Mask[i] == (((uint64_t)i << N) & ModMask))
9171 ViableForN[j] = false;
9173 // Early exit if we exhaust the possible powers of two.
9178 for (unsigned j = 0; j != array_lengthof(ViableForN); ++j)
9182 // Return 0 as there is no viable power of two.
9186 /// \brief Generic lowering of v16i8 shuffles.
9188 /// This is a hybrid strategy to lower v16i8 vectors. It first attempts to
9189 /// detect any complexity reducing interleaving. If that doesn't help, it uses
9190 /// UNPCK to spread the i8 elements across two i16-element vectors, and uses
9191 /// the existing lowering for v8i16 blends on each half, finally PACK-ing them
9193 static SDValue lowerV16I8VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
9194 const X86Subtarget *Subtarget,
9195 SelectionDAG &DAG) {
9197 assert(Op.getSimpleValueType() == MVT::v16i8 && "Bad shuffle type!");
9198 assert(V1.getSimpleValueType() == MVT::v16i8 && "Bad operand type!");
9199 assert(V2.getSimpleValueType() == MVT::v16i8 && "Bad operand type!");
9200 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
9201 ArrayRef<int> Mask = SVOp->getMask();
9202 assert(Mask.size() == 16 && "Unexpected mask size for v16 shuffle!");
9204 // Try to use shift instructions.
9206 lowerVectorShuffleAsShift(DL, MVT::v16i8, V1, V2, Mask, DAG))
9209 // Try to use byte rotation instructions.
9210 if (SDValue Rotate = lowerVectorShuffleAsByteRotate(
9211 DL, MVT::v16i8, V1, V2, Mask, Subtarget, DAG))
9214 // Try to use a zext lowering.
9215 if (SDValue ZExt = lowerVectorShuffleAsZeroOrAnyExtend(
9216 DL, MVT::v16i8, V1, V2, Mask, Subtarget, DAG))
9219 // See if we can use SSE4A Extraction / Insertion.
9220 if (Subtarget->hasSSE4A())
9221 if (SDValue V = lowerVectorShuffleWithSSE4A(DL, MVT::v16i8, V1, V2, Mask, DAG))
9225 std::count_if(Mask.begin(), Mask.end(), [](int M) { return M >= 16; });
9227 // For single-input shuffles, there are some nicer lowering tricks we can use.
9228 if (NumV2Elements == 0) {
9229 // Check for being able to broadcast a single element.
9230 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v16i8, V1,
9231 Mask, Subtarget, DAG))
9234 // Check whether we can widen this to an i16 shuffle by duplicating bytes.
9235 // Notably, this handles splat and partial-splat shuffles more efficiently.
9236 // However, it only makes sense if the pre-duplication shuffle simplifies
9237 // things significantly. Currently, this means we need to be able to
9238 // express the pre-duplication shuffle as an i16 shuffle.
9240 // FIXME: We should check for other patterns which can be widened into an
9241 // i16 shuffle as well.
9242 auto canWidenViaDuplication = [](ArrayRef<int> Mask) {
9243 for (int i = 0; i < 16; i += 2)
9244 if (Mask[i] != -1 && Mask[i + 1] != -1 && Mask[i] != Mask[i + 1])
9249 auto tryToWidenViaDuplication = [&]() -> SDValue {
9250 if (!canWidenViaDuplication(Mask))
9252 SmallVector<int, 4> LoInputs;
9253 std::copy_if(Mask.begin(), Mask.end(), std::back_inserter(LoInputs),
9254 [](int M) { return M >= 0 && M < 8; });
9255 std::sort(LoInputs.begin(), LoInputs.end());
9256 LoInputs.erase(std::unique(LoInputs.begin(), LoInputs.end()),
9258 SmallVector<int, 4> HiInputs;
9259 std::copy_if(Mask.begin(), Mask.end(), std::back_inserter(HiInputs),
9260 [](int M) { return M >= 8; });
9261 std::sort(HiInputs.begin(), HiInputs.end());
9262 HiInputs.erase(std::unique(HiInputs.begin(), HiInputs.end()),
9265 bool TargetLo = LoInputs.size() >= HiInputs.size();
9266 ArrayRef<int> InPlaceInputs = TargetLo ? LoInputs : HiInputs;
9267 ArrayRef<int> MovingInputs = TargetLo ? HiInputs : LoInputs;
9269 int PreDupI16Shuffle[] = {-1, -1, -1, -1, -1, -1, -1, -1};
9270 SmallDenseMap<int, int, 8> LaneMap;
9271 for (int I : InPlaceInputs) {
9272 PreDupI16Shuffle[I/2] = I/2;
9275 int j = TargetLo ? 0 : 4, je = j + 4;
9276 for (int i = 0, ie = MovingInputs.size(); i < ie; ++i) {
9277 // Check if j is already a shuffle of this input. This happens when
9278 // there are two adjacent bytes after we move the low one.
9279 if (PreDupI16Shuffle[j] != MovingInputs[i] / 2) {
9280 // If we haven't yet mapped the input, search for a slot into which
9282 while (j < je && PreDupI16Shuffle[j] != -1)
9286 // We can't place the inputs into a single half with a simple i16 shuffle, so bail.
9289 // Map this input with the i16 shuffle.
9290 PreDupI16Shuffle[j] = MovingInputs[i] / 2;
9293 // Update the lane map based on the mapping we ended up with.
9294 LaneMap[MovingInputs[i]] = 2 * j + MovingInputs[i] % 2;
9296 V1 = DAG.getBitcast(
9298 DAG.getVectorShuffle(MVT::v8i16, DL, DAG.getBitcast(MVT::v8i16, V1),
9299 DAG.getUNDEF(MVT::v8i16), PreDupI16Shuffle));
9301 // Unpack the bytes to form the i16s that will be shuffled into place.
9302 V1 = DAG.getNode(TargetLo ? X86ISD::UNPCKL : X86ISD::UNPCKH, DL,
9303 MVT::v16i8, V1, V1);
9305 int PostDupI16Shuffle[8] = {-1, -1, -1, -1, -1, -1, -1, -1};
9306 for (int i = 0; i < 16; ++i)
9307 if (Mask[i] != -1) {
9308 int MappedMask = LaneMap[Mask[i]] - (TargetLo ? 0 : 8);
9309 assert(MappedMask < 8 && "Invalid v8 shuffle mask!");
9310 if (PostDupI16Shuffle[i / 2] == -1)
9311 PostDupI16Shuffle[i / 2] = MappedMask;
9313 assert(PostDupI16Shuffle[i / 2] == MappedMask &&
9314 "Conflicting entrties in the original shuffle!");
9316 return DAG.getBitcast(
9318 DAG.getVectorShuffle(MVT::v8i16, DL, DAG.getBitcast(MVT::v8i16, V1),
9319 DAG.getUNDEF(MVT::v8i16), PostDupI16Shuffle));
9321 if (SDValue V = tryToWidenViaDuplication())
9325 if (SDValue Masked =
9326 lowerVectorShuffleAsBitMask(DL, MVT::v16i8, V1, V2, Mask, DAG))
9329 // Use dedicated unpack instructions for masks that match their pattern.
9330 if (isShuffleEquivalent(V1, V2, Mask, {// Low half.
9331 0, 16, 1, 17, 2, 18, 3, 19,
9333 4, 20, 5, 21, 6, 22, 7, 23}))
9334 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v16i8, V1, V2);
9335 if (isShuffleEquivalent(V1, V2, Mask, {// Low half.
9336 8, 24, 9, 25, 10, 26, 11, 27,
9338 12, 28, 13, 29, 14, 30, 15, 31}))
9339 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v16i8, V1, V2);
9341 // Check for SSSE3 which lets us lower all v16i8 shuffles much more directly
9342 // with PSHUFB. It is important to do this before we attempt to generate any
9343 // blends but after all of the single-input lowerings. If the single input
9344 // lowerings can find an instruction sequence that is faster than a PSHUFB, we
9345 // want to preserve that and we can DAG combine any longer sequences into
9346 // a PSHUFB in the end. But once we start blending from multiple inputs,
9347 // the complexity of DAG combining bad patterns back into PSHUFB is too high,
9348 // and there are *very* few patterns that would actually be faster than the
9349 // PSHUFB approach because of its ability to zero lanes.
9351 // FIXME: The only exceptions to the above are blends which are exact
9352 // interleavings with direct instructions supporting them. We currently don't
9353 // handle those well here.
9354 if (Subtarget->hasSSSE3()) {
9355 bool V1InUse = false;
9356 bool V2InUse = false;
9358 SDValue PSHUFB = lowerVectorShuffleAsPSHUFB(DL, MVT::v16i8, V1, V2, Mask,
9359 DAG, V1InUse, V2InUse);
9361 // If both V1 and V2 are in use and we can use a direct blend or an unpack,
9362 // do so. This avoids using them to handle blends-with-zero which is
9363 // important as a single pshufb is significantly faster for that.
9364 if (V1InUse && V2InUse) {
9365 if (Subtarget->hasSSE41())
9366 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v16i8, V1, V2,
9367 Mask, Subtarget, DAG))
9370 // We can use an unpack to do the blending rather than an or in some
9371 // cases. Even though the or may be (very minorly) more efficient, we
9372 // preference this lowering because there are common cases where part of
9373 // the complexity of the shuffles goes away when we do the final blend as
9375 // FIXME: It might be worth trying to detect if the unpack-feeding
9376 // shuffles will both be pshufb, in which case we shouldn't bother with
9378 if (SDValue Unpack = lowerVectorShuffleAsPermuteAndUnpack(
9379 DL, MVT::v16i8, V1, V2, Mask, DAG))
9386 // There are special ways we can lower some single-element blends.
9387 if (NumV2Elements == 1)
9388 if (SDValue V = lowerVectorShuffleAsElementInsertion(DL, MVT::v16i8, V1, V2,
9389 Mask, Subtarget, DAG))
9392 if (SDValue BitBlend =
9393 lowerVectorShuffleAsBitBlend(DL, MVT::v16i8, V1, V2, Mask, DAG))
9396 // Check whether a compaction lowering can be done. This handles shuffles
9397 // which take every Nth element for some even N. See the helper function for
9400 // We special case these as they can be particularly efficiently handled with
9401 // the PACKUSB instruction on x86 and they show up in common patterns of
9402 // rearranging bytes to truncate wide elements.
9403 if (int NumEvenDrops = canLowerByDroppingEvenElements(Mask)) {
9404 // NumEvenDrops is the power of two stride of the elements. Another way of
9405 // thinking about it is that we need to drop the even elements this many
9406 // times to get the original input.
9407 bool IsSingleInput = isSingleInputShuffleMask(Mask);
9409 // First we need to zero all the dropped bytes.
9410 assert(NumEvenDrops <= 3 &&
9411 "No support for dropping even elements more than 3 times.");
9412 // We use the mask type to pick which bytes are preserved based on how many
9413 // elements are dropped.
9414 MVT MaskVTs[] = { MVT::v8i16, MVT::v4i32, MVT::v2i64 };
9415 SDValue ByteClearMask = DAG.getBitcast(
9416 MVT::v16i8, DAG.getConstant(0xFF, DL, MaskVTs[NumEvenDrops - 1]));
9417 V1 = DAG.getNode(ISD::AND, DL, MVT::v16i8, V1, ByteClearMask);
9419 V2 = DAG.getNode(ISD::AND, DL, MVT::v16i8, V2, ByteClearMask);
9421 // Now pack things back together.
9422 V1 = DAG.getBitcast(MVT::v8i16, V1);
9423 V2 = IsSingleInput ? V1 : DAG.getBitcast(MVT::v8i16, V2);
9424 SDValue Result = DAG.getNode(X86ISD::PACKUS, DL, MVT::v16i8, V1, V2);
9425 for (int i = 1; i < NumEvenDrops; ++i) {
9426 Result = DAG.getBitcast(MVT::v8i16, Result);
9427 Result = DAG.getNode(X86ISD::PACKUS, DL, MVT::v16i8, Result, Result);
9433 // Handle multi-input cases by blending single-input shuffles.
9434 if (NumV2Elements > 0)
9435 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v16i8, V1, V2,
9438 // The fallback path for single-input shuffles widens this into two v8i16
9439 // vectors with unpacks, shuffles those, and then pulls them back together
9443 int LoBlendMask[8] = {-1, -1, -1, -1, -1, -1, -1, -1};
9444 int HiBlendMask[8] = {-1, -1, -1, -1, -1, -1, -1, -1};
9445 for (int i = 0; i < 16; ++i)
9447 (i < 8 ? LoBlendMask[i] : HiBlendMask[i % 8]) = Mask[i];
9449 SDValue Zero = getZeroVector(MVT::v8i16, Subtarget, DAG, DL);
9451 SDValue VLoHalf, VHiHalf;
9452 // Check if any of the odd lanes in the v16i8 are used. If not, we can mask
9453 // them out and avoid using UNPCK{L,H} to extract the elements of V as
9455 if (std::none_of(std::begin(LoBlendMask), std::end(LoBlendMask),
9456 [](int M) { return M >= 0 && M % 2 == 1; }) &&
9457 std::none_of(std::begin(HiBlendMask), std::end(HiBlendMask),
9458 [](int M) { return M >= 0 && M % 2 == 1; })) {
9459 // Use a mask to drop the high bytes.
9460 VLoHalf = DAG.getBitcast(MVT::v8i16, V);
9461 VLoHalf = DAG.getNode(ISD::AND, DL, MVT::v8i16, VLoHalf,
9462 DAG.getConstant(0x00FF, DL, MVT::v8i16));
9464 // This will be a single vector shuffle instead of a blend so nuke VHiHalf.
9465 VHiHalf = DAG.getUNDEF(MVT::v8i16);
9467 // Squash the masks to point directly into VLoHalf.
9468 for (int &M : LoBlendMask)
9471 for (int &M : HiBlendMask)
9475 // Otherwise just unpack the low half of V into VLoHalf and the high half into
9476 // VHiHalf so that we can blend them as i16s.
9477 VLoHalf = DAG.getBitcast(
9478 MVT::v8i16, DAG.getNode(X86ISD::UNPCKL, DL, MVT::v16i8, V, Zero));
9479 VHiHalf = DAG.getBitcast(
9480 MVT::v8i16, DAG.getNode(X86ISD::UNPCKH, DL, MVT::v16i8, V, Zero));
9483 SDValue LoV = DAG.getVectorShuffle(MVT::v8i16, DL, VLoHalf, VHiHalf, LoBlendMask);
9484 SDValue HiV = DAG.getVectorShuffle(MVT::v8i16, DL, VLoHalf, VHiHalf, HiBlendMask);
9486 return DAG.getNode(X86ISD::PACKUS, DL, MVT::v16i8, LoV, HiV);
9489 /// \brief Dispatching routine to lower various 128-bit x86 vector shuffles.
9491 /// This routine breaks down the specific type of 128-bit shuffle and
9492 /// dispatches to the lowering routines accordingly.
9493 static SDValue lower128BitVectorShuffle(SDValue Op, SDValue V1, SDValue V2,
9494 MVT VT, const X86Subtarget *Subtarget,
9495 SelectionDAG &DAG) {
9496 switch (VT.SimpleTy) {
9498 return lowerV2I64VectorShuffle(Op, V1, V2, Subtarget, DAG);
9500 return lowerV2F64VectorShuffle(Op, V1, V2, Subtarget, DAG);
9502 return lowerV4I32VectorShuffle(Op, V1, V2, Subtarget, DAG);
9504 return lowerV4F32VectorShuffle(Op, V1, V2, Subtarget, DAG);
9506 return lowerV8I16VectorShuffle(Op, V1, V2, Subtarget, DAG);
9508 return lowerV16I8VectorShuffle(Op, V1, V2, Subtarget, DAG);
9511 llvm_unreachable("Unimplemented!");
9515 /// \brief Helper function to test whether a shuffle mask could be
9516 /// simplified by widening the elements being shuffled.
9518 /// Appends the mask for wider elements in WidenedMask if valid. Otherwise
9519 /// leaves it in an unspecified state.
9521 /// NOTE: This must handle normal vector shuffle masks and *target* vector
9522 /// shuffle masks. The latter have the special property of a '-2' representing
9523 /// a zero-ed lane of a vector.
9524 static bool canWidenShuffleElements(ArrayRef<int> Mask,
9525 SmallVectorImpl<int> &WidenedMask) {
9526 for (int i = 0, Size = Mask.size(); i < Size; i += 2) {
9527 // If both elements are undef, its trivial.
9528 if (Mask[i] == SM_SentinelUndef && Mask[i + 1] == SM_SentinelUndef) {
9529 WidenedMask.push_back(SM_SentinelUndef);
9533 // Check for an undef mask and a mask value properly aligned to fit with
9534 // a pair of values. If we find such a case, use the non-undef mask's value.
9535 if (Mask[i] == SM_SentinelUndef && Mask[i + 1] >= 0 && Mask[i + 1] % 2 == 1) {
9536 WidenedMask.push_back(Mask[i + 1] / 2);
9539 if (Mask[i + 1] == SM_SentinelUndef && Mask[i] >= 0 && Mask[i] % 2 == 0) {
9540 WidenedMask.push_back(Mask[i] / 2);
9544 // When zeroing, we need to spread the zeroing across both lanes to widen.
9545 if (Mask[i] == SM_SentinelZero || Mask[i + 1] == SM_SentinelZero) {
9546 if ((Mask[i] == SM_SentinelZero || Mask[i] == SM_SentinelUndef) &&
9547 (Mask[i + 1] == SM_SentinelZero || Mask[i + 1] == SM_SentinelUndef)) {
9548 WidenedMask.push_back(SM_SentinelZero);
9554 // Finally check if the two mask values are adjacent and aligned with
9556 if (Mask[i] != SM_SentinelUndef && Mask[i] % 2 == 0 && Mask[i] + 1 == Mask[i + 1]) {
9557 WidenedMask.push_back(Mask[i] / 2);
9561 // Otherwise we can't safely widen the elements used in this shuffle.
9564 assert(WidenedMask.size() == Mask.size() / 2 &&
9565 "Incorrect size of mask after widening the elements!");
9570 /// \brief Generic routine to split vector shuffle into half-sized shuffles.
9572 /// This routine just extracts two subvectors, shuffles them independently, and
9573 /// then concatenates them back together. This should work effectively with all
9574 /// AVX vector shuffle types.
9575 static SDValue splitAndLowerVectorShuffle(SDLoc DL, MVT VT, SDValue V1,
9576 SDValue V2, ArrayRef<int> Mask,
9577 SelectionDAG &DAG) {
9578 assert(VT.getSizeInBits() >= 256 &&
9579 "Only for 256-bit or wider vector shuffles!");
9580 assert(V1.getSimpleValueType() == VT && "Bad operand type!");
9581 assert(V2.getSimpleValueType() == VT && "Bad operand type!");
9583 ArrayRef<int> LoMask = Mask.slice(0, Mask.size() / 2);
9584 ArrayRef<int> HiMask = Mask.slice(Mask.size() / 2);
9586 int NumElements = VT.getVectorNumElements();
9587 int SplitNumElements = NumElements / 2;
9588 MVT ScalarVT = VT.getScalarType();
9589 MVT SplitVT = MVT::getVectorVT(ScalarVT, NumElements / 2);
9591 // Rather than splitting build-vectors, just build two narrower build
9592 // vectors. This helps shuffling with splats and zeros.
9593 auto SplitVector = [&](SDValue V) {
9594 while (V.getOpcode() == ISD::BITCAST)
9595 V = V->getOperand(0);
9597 MVT OrigVT = V.getSimpleValueType();
9598 int OrigNumElements = OrigVT.getVectorNumElements();
9599 int OrigSplitNumElements = OrigNumElements / 2;
9600 MVT OrigScalarVT = OrigVT.getScalarType();
9601 MVT OrigSplitVT = MVT::getVectorVT(OrigScalarVT, OrigNumElements / 2);
9605 auto *BV = dyn_cast<BuildVectorSDNode>(V);
9607 LoV = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, OrigSplitVT, V,
9608 DAG.getIntPtrConstant(0, DL));
9609 HiV = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, OrigSplitVT, V,
9610 DAG.getIntPtrConstant(OrigSplitNumElements, DL));
9613 SmallVector<SDValue, 16> LoOps, HiOps;
9614 for (int i = 0; i < OrigSplitNumElements; ++i) {
9615 LoOps.push_back(BV->getOperand(i));
9616 HiOps.push_back(BV->getOperand(i + OrigSplitNumElements));
9618 LoV = DAG.getNode(ISD::BUILD_VECTOR, DL, OrigSplitVT, LoOps);
9619 HiV = DAG.getNode(ISD::BUILD_VECTOR, DL, OrigSplitVT, HiOps);
9621 return std::make_pair(DAG.getBitcast(SplitVT, LoV),
9622 DAG.getBitcast(SplitVT, HiV));
9625 SDValue LoV1, HiV1, LoV2, HiV2;
9626 std::tie(LoV1, HiV1) = SplitVector(V1);
9627 std::tie(LoV2, HiV2) = SplitVector(V2);
9629 // Now create two 4-way blends of these half-width vectors.
9630 auto HalfBlend = [&](ArrayRef<int> HalfMask) {
9631 bool UseLoV1 = false, UseHiV1 = false, UseLoV2 = false, UseHiV2 = false;
9632 SmallVector<int, 32> V1BlendMask, V2BlendMask, BlendMask;
9633 for (int i = 0; i < SplitNumElements; ++i) {
9634 int M = HalfMask[i];
9635 if (M >= NumElements) {
9636 if (M >= NumElements + SplitNumElements)
9640 V2BlendMask.push_back(M - NumElements);
9641 V1BlendMask.push_back(-1);
9642 BlendMask.push_back(SplitNumElements + i);
9643 } else if (M >= 0) {
9644 if (M >= SplitNumElements)
9648 V2BlendMask.push_back(-1);
9649 V1BlendMask.push_back(M);
9650 BlendMask.push_back(i);
9652 V2BlendMask.push_back(-1);
9653 V1BlendMask.push_back(-1);
9654 BlendMask.push_back(-1);
9658 // Because the lowering happens after all combining takes place, we need to
9659 // manually combine these blend masks as much as possible so that we create
9660 // a minimal number of high-level vector shuffle nodes.
9662 // First try just blending the halves of V1 or V2.
9663 if (!UseLoV1 && !UseHiV1 && !UseLoV2 && !UseHiV2)
9664 return DAG.getUNDEF(SplitVT);
9665 if (!UseLoV2 && !UseHiV2)
9666 return DAG.getVectorShuffle(SplitVT, DL, LoV1, HiV1, V1BlendMask);
9667 if (!UseLoV1 && !UseHiV1)
9668 return DAG.getVectorShuffle(SplitVT, DL, LoV2, HiV2, V2BlendMask);
9670 SDValue V1Blend, V2Blend;
9671 if (UseLoV1 && UseHiV1) {
9673 DAG.getVectorShuffle(SplitVT, DL, LoV1, HiV1, V1BlendMask);
9675 // We only use half of V1 so map the usage down into the final blend mask.
9676 V1Blend = UseLoV1 ? LoV1 : HiV1;
9677 for (int i = 0; i < SplitNumElements; ++i)
9678 if (BlendMask[i] >= 0 && BlendMask[i] < SplitNumElements)
9679 BlendMask[i] = V1BlendMask[i] - (UseLoV1 ? 0 : SplitNumElements);
9681 if (UseLoV2 && UseHiV2) {
9683 DAG.getVectorShuffle(SplitVT, DL, LoV2, HiV2, V2BlendMask);
9685 // We only use half of V2 so map the usage down into the final blend mask.
9686 V2Blend = UseLoV2 ? LoV2 : HiV2;
9687 for (int i = 0; i < SplitNumElements; ++i)
9688 if (BlendMask[i] >= SplitNumElements)
9689 BlendMask[i] = V2BlendMask[i] + (UseLoV2 ? SplitNumElements : 0);
9691 return DAG.getVectorShuffle(SplitVT, DL, V1Blend, V2Blend, BlendMask);
9693 SDValue Lo = HalfBlend(LoMask);
9694 SDValue Hi = HalfBlend(HiMask);
9695 return DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, Lo, Hi);
9698 /// \brief Either split a vector in halves or decompose the shuffles and the
9701 /// This is provided as a good fallback for many lowerings of non-single-input
9702 /// shuffles with more than one 128-bit lane. In those cases, we want to select
9703 /// between splitting the shuffle into 128-bit components and stitching those
9704 /// back together vs. extracting the single-input shuffles and blending those
9706 static SDValue lowerVectorShuffleAsSplitOrBlend(SDLoc DL, MVT VT, SDValue V1,
9707 SDValue V2, ArrayRef<int> Mask,
9708 SelectionDAG &DAG) {
9709 assert(!isSingleInputShuffleMask(Mask) && "This routine must not be used to "
9710 "lower single-input shuffles as it "
9711 "could then recurse on itself.");
9712 int Size = Mask.size();
9714 // If this can be modeled as a broadcast of two elements followed by a blend,
9715 // prefer that lowering. This is especially important because broadcasts can
9716 // often fold with memory operands.
9717 auto DoBothBroadcast = [&] {
9718 int V1BroadcastIdx = -1, V2BroadcastIdx = -1;
9721 if (V2BroadcastIdx == -1)
9722 V2BroadcastIdx = M - Size;
9723 else if (M - Size != V2BroadcastIdx)
9725 } else if (M >= 0) {
9726 if (V1BroadcastIdx == -1)
9728 else if (M != V1BroadcastIdx)
9733 if (DoBothBroadcast())
9734 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, VT, V1, V2, Mask,
9737 // If the inputs all stem from a single 128-bit lane of each input, then we
9738 // split them rather than blending because the split will decompose to
9739 // unusually few instructions.
9740 int LaneCount = VT.getSizeInBits() / 128;
9741 int LaneSize = Size / LaneCount;
9742 SmallBitVector LaneInputs[2];
9743 LaneInputs[0].resize(LaneCount, false);
9744 LaneInputs[1].resize(LaneCount, false);
9745 for (int i = 0; i < Size; ++i)
9747 LaneInputs[Mask[i] / Size][(Mask[i] % Size) / LaneSize] = true;
9748 if (LaneInputs[0].count() <= 1 && LaneInputs[1].count() <= 1)
9749 return splitAndLowerVectorShuffle(DL, VT, V1, V2, Mask, DAG);
9751 // Otherwise, just fall back to decomposed shuffles and a blend. This requires
9752 // that the decomposed single-input shuffles don't end up here.
9753 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, VT, V1, V2, Mask, DAG);
9756 /// \brief Lower a vector shuffle crossing multiple 128-bit lanes as
9757 /// a permutation and blend of those lanes.
9759 /// This essentially blends the out-of-lane inputs to each lane into the lane
9760 /// from a permuted copy of the vector. This lowering strategy results in four
9761 /// instructions in the worst case for a single-input cross lane shuffle which
9762 /// is lower than any other fully general cross-lane shuffle strategy I'm aware
9763 /// of. Special cases for each particular shuffle pattern should be handled
9764 /// prior to trying this lowering.
9765 static SDValue lowerVectorShuffleAsLanePermuteAndBlend(SDLoc DL, MVT VT,
9766 SDValue V1, SDValue V2,
9768 SelectionDAG &DAG) {
9769 // FIXME: This should probably be generalized for 512-bit vectors as well.
9770 assert(VT.getSizeInBits() == 256 && "Only for 256-bit vector shuffles!");
9771 int LaneSize = Mask.size() / 2;
9773 // If there are only inputs from one 128-bit lane, splitting will in fact be
9774 // less expensive. The flags track whether the given lane contains an element
9775 // that crosses to another lane.
9776 bool LaneCrossing[2] = {false, false};
9777 for (int i = 0, Size = Mask.size(); i < Size; ++i)
9778 if (Mask[i] >= 0 && (Mask[i] % Size) / LaneSize != i / LaneSize)
9779 LaneCrossing[(Mask[i] % Size) / LaneSize] = true;
9780 if (!LaneCrossing[0] || !LaneCrossing[1])
9781 return splitAndLowerVectorShuffle(DL, VT, V1, V2, Mask, DAG);
9783 if (isSingleInputShuffleMask(Mask)) {
9784 SmallVector<int, 32> FlippedBlendMask;
9785 for (int i = 0, Size = Mask.size(); i < Size; ++i)
9786 FlippedBlendMask.push_back(
9787 Mask[i] < 0 ? -1 : (((Mask[i] % Size) / LaneSize == i / LaneSize)
9789 : Mask[i] % LaneSize +
9790 (i / LaneSize) * LaneSize + Size));
9792 // Flip the vector, and blend the results which should now be in-lane. The
9793 // VPERM2X128 mask uses the low 2 bits for the low source and bits 4 and
9794 // 5 for the high source. The value 3 selects the high half of source 2 and
9795 // the value 2 selects the low half of source 2. We only use source 2 to
9796 // allow folding it into a memory operand.
9797 unsigned PERMMask = 3 | 2 << 4;
9798 SDValue Flipped = DAG.getNode(X86ISD::VPERM2X128, DL, VT, DAG.getUNDEF(VT),
9799 V1, DAG.getConstant(PERMMask, DL, MVT::i8));
9800 return DAG.getVectorShuffle(VT, DL, V1, Flipped, FlippedBlendMask);
9803 // This now reduces to two single-input shuffles of V1 and V2 which at worst
9804 // will be handled by the above logic and a blend of the results, much like
9805 // other patterns in AVX.
9806 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, VT, V1, V2, Mask, DAG);
9809 /// \brief Handle lowering 2-lane 128-bit shuffles.
9810 static SDValue lowerV2X128VectorShuffle(SDLoc DL, MVT VT, SDValue V1,
9811 SDValue V2, ArrayRef<int> Mask,
9812 const X86Subtarget *Subtarget,
9813 SelectionDAG &DAG) {
9814 // TODO: If minimizing size and one of the inputs is a zero vector and the
9815 // the zero vector has only one use, we could use a VPERM2X128 to save the
9816 // instruction bytes needed to explicitly generate the zero vector.
9818 // Blends are faster and handle all the non-lane-crossing cases.
9819 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, VT, V1, V2, Mask,
9823 bool IsV1Zero = ISD::isBuildVectorAllZeros(V1.getNode());
9824 bool IsV2Zero = ISD::isBuildVectorAllZeros(V2.getNode());
9826 // If either input operand is a zero vector, use VPERM2X128 because its mask
9827 // allows us to replace the zero input with an implicit zero.
9828 if (!IsV1Zero && !IsV2Zero) {
9829 // Check for patterns which can be matched with a single insert of a 128-bit
9831 bool OnlyUsesV1 = isShuffleEquivalent(V1, V2, Mask, {0, 1, 0, 1});
9832 if (OnlyUsesV1 || isShuffleEquivalent(V1, V2, Mask, {0, 1, 4, 5})) {
9833 MVT SubVT = MVT::getVectorVT(VT.getVectorElementType(),
9834 VT.getVectorNumElements() / 2);
9835 SDValue LoV = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, SubVT, V1,
9836 DAG.getIntPtrConstant(0, DL));
9837 SDValue HiV = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, SubVT,
9838 OnlyUsesV1 ? V1 : V2,
9839 DAG.getIntPtrConstant(0, DL));
9840 return DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, LoV, HiV);
9844 // Otherwise form a 128-bit permutation. After accounting for undefs,
9845 // convert the 64-bit shuffle mask selection values into 128-bit
9846 // selection bits by dividing the indexes by 2 and shifting into positions
9847 // defined by a vperm2*128 instruction's immediate control byte.
9849 // The immediate permute control byte looks like this:
9850 // [1:0] - select 128 bits from sources for low half of destination
9852 // [3] - zero low half of destination
9853 // [5:4] - select 128 bits from sources for high half of destination
9855 // [7] - zero high half of destination
9857 int MaskLO = Mask[0];
9858 if (MaskLO == SM_SentinelUndef)
9859 MaskLO = Mask[1] == SM_SentinelUndef ? 0 : Mask[1];
9861 int MaskHI = Mask[2];
9862 if (MaskHI == SM_SentinelUndef)
9863 MaskHI = Mask[3] == SM_SentinelUndef ? 0 : Mask[3];
9865 unsigned PermMask = MaskLO / 2 | (MaskHI / 2) << 4;
9867 // If either input is a zero vector, replace it with an undef input.
9868 // Shuffle mask values < 4 are selecting elements of V1.
9869 // Shuffle mask values >= 4 are selecting elements of V2.
9870 // Adjust each half of the permute mask by clearing the half that was
9871 // selecting the zero vector and setting the zero mask bit.
9873 V1 = DAG.getUNDEF(VT);
9875 PermMask = (PermMask & 0xf0) | 0x08;
9877 PermMask = (PermMask & 0x0f) | 0x80;
9880 V2 = DAG.getUNDEF(VT);
9882 PermMask = (PermMask & 0xf0) | 0x08;
9884 PermMask = (PermMask & 0x0f) | 0x80;
9887 return DAG.getNode(X86ISD::VPERM2X128, DL, VT, V1, V2,
9888 DAG.getConstant(PermMask, DL, MVT::i8));
9891 /// \brief Lower a vector shuffle by first fixing the 128-bit lanes and then
9892 /// shuffling each lane.
9894 /// This will only succeed when the result of fixing the 128-bit lanes results
9895 /// in a single-input non-lane-crossing shuffle with a repeating shuffle mask in
9896 /// each 128-bit lanes. This handles many cases where we can quickly blend away
9897 /// the lane crosses early and then use simpler shuffles within each lane.
9899 /// FIXME: It might be worthwhile at some point to support this without
9900 /// requiring the 128-bit lane-relative shuffles to be repeating, but currently
9901 /// in x86 only floating point has interesting non-repeating shuffles, and even
9902 /// those are still *marginally* more expensive.
9903 static SDValue lowerVectorShuffleByMerging128BitLanes(
9904 SDLoc DL, MVT VT, SDValue V1, SDValue V2, ArrayRef<int> Mask,
9905 const X86Subtarget *Subtarget, SelectionDAG &DAG) {
9906 assert(!isSingleInputShuffleMask(Mask) &&
9907 "This is only useful with multiple inputs.");
9909 int Size = Mask.size();
9910 int LaneSize = 128 / VT.getScalarSizeInBits();
9911 int NumLanes = Size / LaneSize;
9912 assert(NumLanes > 1 && "Only handles 256-bit and wider shuffles.");
9914 // See if we can build a hypothetical 128-bit lane-fixing shuffle mask. Also
9915 // check whether the in-128-bit lane shuffles share a repeating pattern.
9916 SmallVector<int, 4> Lanes;
9917 Lanes.resize(NumLanes, -1);
9918 SmallVector<int, 4> InLaneMask;
9919 InLaneMask.resize(LaneSize, -1);
9920 for (int i = 0; i < Size; ++i) {
9924 int j = i / LaneSize;
9927 // First entry we've seen for this lane.
9928 Lanes[j] = Mask[i] / LaneSize;
9929 } else if (Lanes[j] != Mask[i] / LaneSize) {
9930 // This doesn't match the lane selected previously!
9934 // Check that within each lane we have a consistent shuffle mask.
9935 int k = i % LaneSize;
9936 if (InLaneMask[k] < 0) {
9937 InLaneMask[k] = Mask[i] % LaneSize;
9938 } else if (InLaneMask[k] != Mask[i] % LaneSize) {
9939 // This doesn't fit a repeating in-lane mask.
9944 // First shuffle the lanes into place.
9945 MVT LaneVT = MVT::getVectorVT(VT.isFloatingPoint() ? MVT::f64 : MVT::i64,
9946 VT.getSizeInBits() / 64);
9947 SmallVector<int, 8> LaneMask;
9948 LaneMask.resize(NumLanes * 2, -1);
9949 for (int i = 0; i < NumLanes; ++i)
9950 if (Lanes[i] >= 0) {
9951 LaneMask[2 * i + 0] = 2*Lanes[i] + 0;
9952 LaneMask[2 * i + 1] = 2*Lanes[i] + 1;
9955 V1 = DAG.getBitcast(LaneVT, V1);
9956 V2 = DAG.getBitcast(LaneVT, V2);
9957 SDValue LaneShuffle = DAG.getVectorShuffle(LaneVT, DL, V1, V2, LaneMask);
9959 // Cast it back to the type we actually want.
9960 LaneShuffle = DAG.getBitcast(VT, LaneShuffle);
9962 // Now do a simple shuffle that isn't lane crossing.
9963 SmallVector<int, 8> NewMask;
9964 NewMask.resize(Size, -1);
9965 for (int i = 0; i < Size; ++i)
9967 NewMask[i] = (i / LaneSize) * LaneSize + Mask[i] % LaneSize;
9968 assert(!is128BitLaneCrossingShuffleMask(VT, NewMask) &&
9969 "Must not introduce lane crosses at this point!");
9971 return DAG.getVectorShuffle(VT, DL, LaneShuffle, DAG.getUNDEF(VT), NewMask);
9974 /// \brief Test whether the specified input (0 or 1) is in-place blended by the
9977 /// This returns true if the elements from a particular input are already in the
9978 /// slot required by the given mask and require no permutation.
9979 static bool isShuffleMaskInputInPlace(int Input, ArrayRef<int> Mask) {
9980 assert((Input == 0 || Input == 1) && "Only two inputs to shuffles.");
9981 int Size = Mask.size();
9982 for (int i = 0; i < Size; ++i)
9983 if (Mask[i] >= 0 && Mask[i] / Size == Input && Mask[i] % Size != i)
9989 static SDValue lowerVectorShuffleWithSHUFPD(SDLoc DL, MVT VT,
9990 ArrayRef<int> Mask, SDValue V1,
9991 SDValue V2, SelectionDAG &DAG) {
9993 // Mask for V8F64: 0/1, 8/9, 2/3, 10/11, 4/5, ..
9994 // Mask for V4F64; 0/1, 4/5, 2/3, 6/7..
9995 assert(VT.getScalarSizeInBits() == 64 && "Unexpected data type for VSHUFPD");
9996 int NumElts = VT.getVectorNumElements();
9997 bool ShufpdMask = true;
9998 bool CommutableMask = true;
9999 unsigned Immediate = 0;
10000 for (int i = 0; i < NumElts; ++i) {
10003 int Val = (i & 6) + NumElts * (i & 1);
10004 int CommutVal = (i & 0xe) + NumElts * ((i & 1)^1);
10005 if (Mask[i] < Val || Mask[i] > Val + 1)
10006 ShufpdMask = false;
10007 if (Mask[i] < CommutVal || Mask[i] > CommutVal + 1)
10008 CommutableMask = false;
10009 Immediate |= (Mask[i] % 2) << i;
10012 return DAG.getNode(X86ISD::SHUFP, DL, VT, V1, V2,
10013 DAG.getConstant(Immediate, DL, MVT::i8));
10014 if (CommutableMask)
10015 return DAG.getNode(X86ISD::SHUFP, DL, VT, V2, V1,
10016 DAG.getConstant(Immediate, DL, MVT::i8));
10020 /// \brief Handle lowering of 4-lane 64-bit floating point shuffles.
10022 /// Also ends up handling lowering of 4-lane 64-bit integer shuffles when AVX2
10023 /// isn't available.
10024 static SDValue lowerV4F64VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10025 const X86Subtarget *Subtarget,
10026 SelectionDAG &DAG) {
10028 assert(V1.getSimpleValueType() == MVT::v4f64 && "Bad operand type!");
10029 assert(V2.getSimpleValueType() == MVT::v4f64 && "Bad operand type!");
10030 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10031 ArrayRef<int> Mask = SVOp->getMask();
10032 assert(Mask.size() == 4 && "Unexpected mask size for v4 shuffle!");
10034 SmallVector<int, 4> WidenedMask;
10035 if (canWidenShuffleElements(Mask, WidenedMask))
10036 return lowerV2X128VectorShuffle(DL, MVT::v4f64, V1, V2, Mask, Subtarget,
10039 if (isSingleInputShuffleMask(Mask)) {
10040 // Check for being able to broadcast a single element.
10041 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v4f64, V1,
10042 Mask, Subtarget, DAG))
10045 // Use low duplicate instructions for masks that match their pattern.
10046 if (isShuffleEquivalent(V1, V2, Mask, {0, 0, 2, 2}))
10047 return DAG.getNode(X86ISD::MOVDDUP, DL, MVT::v4f64, V1);
10049 if (!is128BitLaneCrossingShuffleMask(MVT::v4f64, Mask)) {
10050 // Non-half-crossing single input shuffles can be lowerid with an
10051 // interleaved permutation.
10052 unsigned VPERMILPMask = (Mask[0] == 1) | ((Mask[1] == 1) << 1) |
10053 ((Mask[2] == 3) << 2) | ((Mask[3] == 3) << 3);
10054 return DAG.getNode(X86ISD::VPERMILPI, DL, MVT::v4f64, V1,
10055 DAG.getConstant(VPERMILPMask, DL, MVT::i8));
10058 // With AVX2 we have direct support for this permutation.
10059 if (Subtarget->hasAVX2())
10060 return DAG.getNode(X86ISD::VPERMI, DL, MVT::v4f64, V1,
10061 getV4X86ShuffleImm8ForMask(Mask, DL, DAG));
10063 // Otherwise, fall back.
10064 return lowerVectorShuffleAsLanePermuteAndBlend(DL, MVT::v4f64, V1, V2, Mask,
10068 // X86 has dedicated unpack instructions that can handle specific blend
10069 // operations: UNPCKH and UNPCKL.
10070 if (isShuffleEquivalent(V1, V2, Mask, {0, 4, 2, 6}))
10071 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v4f64, V1, V2);
10072 if (isShuffleEquivalent(V1, V2, Mask, {1, 5, 3, 7}))
10073 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v4f64, V1, V2);
10074 if (isShuffleEquivalent(V1, V2, Mask, {4, 0, 6, 2}))
10075 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v4f64, V2, V1);
10076 if (isShuffleEquivalent(V1, V2, Mask, {5, 1, 7, 3}))
10077 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v4f64, V2, V1);
10079 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v4f64, V1, V2, Mask,
10083 // Check if the blend happens to exactly fit that of SHUFPD.
10085 lowerVectorShuffleWithSHUFPD(DL, MVT::v4f64, Mask, V1, V2, DAG))
10088 // Try to simplify this by merging 128-bit lanes to enable a lane-based
10089 // shuffle. However, if we have AVX2 and either inputs are already in place,
10090 // we will be able to shuffle even across lanes the other input in a single
10091 // instruction so skip this pattern.
10092 if (!(Subtarget->hasAVX2() && (isShuffleMaskInputInPlace(0, Mask) ||
10093 isShuffleMaskInputInPlace(1, Mask))))
10094 if (SDValue Result = lowerVectorShuffleByMerging128BitLanes(
10095 DL, MVT::v4f64, V1, V2, Mask, Subtarget, DAG))
10098 // If we have AVX2 then we always want to lower with a blend because an v4 we
10099 // can fully permute the elements.
10100 if (Subtarget->hasAVX2())
10101 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v4f64, V1, V2,
10104 // Otherwise fall back on generic lowering.
10105 return lowerVectorShuffleAsSplitOrBlend(DL, MVT::v4f64, V1, V2, Mask, DAG);
10108 /// \brief Handle lowering of 4-lane 64-bit integer shuffles.
10110 /// This routine is only called when we have AVX2 and thus a reasonable
10111 /// instruction set for v4i64 shuffling..
10112 static SDValue lowerV4I64VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10113 const X86Subtarget *Subtarget,
10114 SelectionDAG &DAG) {
10116 assert(V1.getSimpleValueType() == MVT::v4i64 && "Bad operand type!");
10117 assert(V2.getSimpleValueType() == MVT::v4i64 && "Bad operand type!");
10118 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10119 ArrayRef<int> Mask = SVOp->getMask();
10120 assert(Mask.size() == 4 && "Unexpected mask size for v4 shuffle!");
10121 assert(Subtarget->hasAVX2() && "We can only lower v4i64 with AVX2!");
10123 SmallVector<int, 4> WidenedMask;
10124 if (canWidenShuffleElements(Mask, WidenedMask))
10125 return lowerV2X128VectorShuffle(DL, MVT::v4i64, V1, V2, Mask, Subtarget,
10128 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v4i64, V1, V2, Mask,
10132 // Check for being able to broadcast a single element.
10133 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v4i64, V1,
10134 Mask, Subtarget, DAG))
10137 // When the shuffle is mirrored between the 128-bit lanes of the unit, we can
10138 // use lower latency instructions that will operate on both 128-bit lanes.
10139 SmallVector<int, 2> RepeatedMask;
10140 if (is128BitLaneRepeatedShuffleMask(MVT::v4i64, Mask, RepeatedMask)) {
10141 if (isSingleInputShuffleMask(Mask)) {
10142 int PSHUFDMask[] = {-1, -1, -1, -1};
10143 for (int i = 0; i < 2; ++i)
10144 if (RepeatedMask[i] >= 0) {
10145 PSHUFDMask[2 * i] = 2 * RepeatedMask[i];
10146 PSHUFDMask[2 * i + 1] = 2 * RepeatedMask[i] + 1;
10148 return DAG.getBitcast(
10150 DAG.getNode(X86ISD::PSHUFD, DL, MVT::v8i32,
10151 DAG.getBitcast(MVT::v8i32, V1),
10152 getV4X86ShuffleImm8ForMask(PSHUFDMask, DL, DAG)));
10156 // AVX2 provides a direct instruction for permuting a single input across
10158 if (isSingleInputShuffleMask(Mask))
10159 return DAG.getNode(X86ISD::VPERMI, DL, MVT::v4i64, V1,
10160 getV4X86ShuffleImm8ForMask(Mask, DL, DAG));
10162 // Try to use shift instructions.
10163 if (SDValue Shift =
10164 lowerVectorShuffleAsShift(DL, MVT::v4i64, V1, V2, Mask, DAG))
10167 // Use dedicated unpack instructions for masks that match their pattern.
10168 if (isShuffleEquivalent(V1, V2, Mask, {0, 4, 2, 6}))
10169 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v4i64, V1, V2);
10170 if (isShuffleEquivalent(V1, V2, Mask, {1, 5, 3, 7}))
10171 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v4i64, V1, V2);
10172 if (isShuffleEquivalent(V1, V2, Mask, {4, 0, 6, 2}))
10173 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v4i64, V2, V1);
10174 if (isShuffleEquivalent(V1, V2, Mask, {5, 1, 7, 3}))
10175 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v4i64, V2, V1);
10177 // Try to simplify this by merging 128-bit lanes to enable a lane-based
10178 // shuffle. However, if we have AVX2 and either inputs are already in place,
10179 // we will be able to shuffle even across lanes the other input in a single
10180 // instruction so skip this pattern.
10181 if (!(Subtarget->hasAVX2() && (isShuffleMaskInputInPlace(0, Mask) ||
10182 isShuffleMaskInputInPlace(1, Mask))))
10183 if (SDValue Result = lowerVectorShuffleByMerging128BitLanes(
10184 DL, MVT::v4i64, V1, V2, Mask, Subtarget, DAG))
10187 // Otherwise fall back on generic blend lowering.
10188 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v4i64, V1, V2,
10192 /// \brief Handle lowering of 8-lane 32-bit floating point shuffles.
10194 /// Also ends up handling lowering of 8-lane 32-bit integer shuffles when AVX2
10195 /// isn't available.
10196 static SDValue lowerV8F32VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10197 const X86Subtarget *Subtarget,
10198 SelectionDAG &DAG) {
10200 assert(V1.getSimpleValueType() == MVT::v8f32 && "Bad operand type!");
10201 assert(V2.getSimpleValueType() == MVT::v8f32 && "Bad operand type!");
10202 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10203 ArrayRef<int> Mask = SVOp->getMask();
10204 assert(Mask.size() == 8 && "Unexpected mask size for v8 shuffle!");
10206 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v8f32, V1, V2, Mask,
10210 // Check for being able to broadcast a single element.
10211 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v8f32, V1,
10212 Mask, Subtarget, DAG))
10215 // If the shuffle mask is repeated in each 128-bit lane, we have many more
10216 // options to efficiently lower the shuffle.
10217 SmallVector<int, 4> RepeatedMask;
10218 if (is128BitLaneRepeatedShuffleMask(MVT::v8f32, Mask, RepeatedMask)) {
10219 assert(RepeatedMask.size() == 4 &&
10220 "Repeated masks must be half the mask width!");
10222 // Use even/odd duplicate instructions for masks that match their pattern.
10223 if (isShuffleEquivalent(V1, V2, Mask, {0, 0, 2, 2, 4, 4, 6, 6}))
10224 return DAG.getNode(X86ISD::MOVSLDUP, DL, MVT::v8f32, V1);
10225 if (isShuffleEquivalent(V1, V2, Mask, {1, 1, 3, 3, 5, 5, 7, 7}))
10226 return DAG.getNode(X86ISD::MOVSHDUP, DL, MVT::v8f32, V1);
10228 if (isSingleInputShuffleMask(Mask))
10229 return DAG.getNode(X86ISD::VPERMILPI, DL, MVT::v8f32, V1,
10230 getV4X86ShuffleImm8ForMask(RepeatedMask, DL, DAG));
10232 // Use dedicated unpack instructions for masks that match their pattern.
10233 if (isShuffleEquivalent(V1, V2, Mask, {0, 8, 1, 9, 4, 12, 5, 13}))
10234 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v8f32, V1, V2);
10235 if (isShuffleEquivalent(V1, V2, Mask, {2, 10, 3, 11, 6, 14, 7, 15}))
10236 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v8f32, V1, V2);
10237 if (isShuffleEquivalent(V1, V2, Mask, {8, 0, 9, 1, 12, 4, 13, 5}))
10238 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v8f32, V2, V1);
10239 if (isShuffleEquivalent(V1, V2, Mask, {10, 2, 11, 3, 14, 6, 15, 7}))
10240 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v8f32, V2, V1);
10242 // Otherwise, fall back to a SHUFPS sequence. Here it is important that we
10243 // have already handled any direct blends. We also need to squash the
10244 // repeated mask into a simulated v4f32 mask.
10245 for (int i = 0; i < 4; ++i)
10246 if (RepeatedMask[i] >= 8)
10247 RepeatedMask[i] -= 4;
10248 return lowerVectorShuffleWithSHUFPS(DL, MVT::v8f32, RepeatedMask, V1, V2, DAG);
10251 // If we have a single input shuffle with different shuffle patterns in the
10252 // two 128-bit lanes use the variable mask to VPERMILPS.
10253 if (isSingleInputShuffleMask(Mask)) {
10254 SDValue VPermMask[8];
10255 for (int i = 0; i < 8; ++i)
10256 VPermMask[i] = Mask[i] < 0 ? DAG.getUNDEF(MVT::i32)
10257 : DAG.getConstant(Mask[i], DL, MVT::i32);
10258 if (!is128BitLaneCrossingShuffleMask(MVT::v8f32, Mask))
10259 return DAG.getNode(
10260 X86ISD::VPERMILPV, DL, MVT::v8f32, V1,
10261 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v8i32, VPermMask));
10263 if (Subtarget->hasAVX2())
10264 return DAG.getNode(
10265 X86ISD::VPERMV, DL, MVT::v8f32,
10266 DAG.getBitcast(MVT::v8f32, DAG.getNode(ISD::BUILD_VECTOR, DL,
10267 MVT::v8i32, VPermMask)),
10270 // Otherwise, fall back.
10271 return lowerVectorShuffleAsLanePermuteAndBlend(DL, MVT::v8f32, V1, V2, Mask,
10275 // Try to simplify this by merging 128-bit lanes to enable a lane-based
10277 if (SDValue Result = lowerVectorShuffleByMerging128BitLanes(
10278 DL, MVT::v8f32, V1, V2, Mask, Subtarget, DAG))
10281 // If we have AVX2 then we always want to lower with a blend because at v8 we
10282 // can fully permute the elements.
10283 if (Subtarget->hasAVX2())
10284 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v8f32, V1, V2,
10287 // Otherwise fall back on generic lowering.
10288 return lowerVectorShuffleAsSplitOrBlend(DL, MVT::v8f32, V1, V2, Mask, DAG);
10291 /// \brief Handle lowering of 8-lane 32-bit integer shuffles.
10293 /// This routine is only called when we have AVX2 and thus a reasonable
10294 /// instruction set for v8i32 shuffling..
10295 static SDValue lowerV8I32VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10296 const X86Subtarget *Subtarget,
10297 SelectionDAG &DAG) {
10299 assert(V1.getSimpleValueType() == MVT::v8i32 && "Bad operand type!");
10300 assert(V2.getSimpleValueType() == MVT::v8i32 && "Bad operand type!");
10301 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10302 ArrayRef<int> Mask = SVOp->getMask();
10303 assert(Mask.size() == 8 && "Unexpected mask size for v8 shuffle!");
10304 assert(Subtarget->hasAVX2() && "We can only lower v8i32 with AVX2!");
10306 // Whenever we can lower this as a zext, that instruction is strictly faster
10307 // than any alternative. It also allows us to fold memory operands into the
10308 // shuffle in many cases.
10309 if (SDValue ZExt = lowerVectorShuffleAsZeroOrAnyExtend(DL, MVT::v8i32, V1, V2,
10310 Mask, Subtarget, DAG))
10313 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v8i32, V1, V2, Mask,
10317 // Check for being able to broadcast a single element.
10318 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v8i32, V1,
10319 Mask, Subtarget, DAG))
10322 // If the shuffle mask is repeated in each 128-bit lane we can use more
10323 // efficient instructions that mirror the shuffles across the two 128-bit
10325 SmallVector<int, 4> RepeatedMask;
10326 if (is128BitLaneRepeatedShuffleMask(MVT::v8i32, Mask, RepeatedMask)) {
10327 assert(RepeatedMask.size() == 4 && "Unexpected repeated mask size!");
10328 if (isSingleInputShuffleMask(Mask))
10329 return DAG.getNode(X86ISD::PSHUFD, DL, MVT::v8i32, V1,
10330 getV4X86ShuffleImm8ForMask(RepeatedMask, DL, DAG));
10332 // Use dedicated unpack instructions for masks that match their pattern.
10333 if (isShuffleEquivalent(V1, V2, Mask, {0, 8, 1, 9, 4, 12, 5, 13}))
10334 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v8i32, V1, V2);
10335 if (isShuffleEquivalent(V1, V2, Mask, {2, 10, 3, 11, 6, 14, 7, 15}))
10336 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v8i32, V1, V2);
10337 if (isShuffleEquivalent(V1, V2, Mask, {8, 0, 9, 1, 12, 4, 13, 5}))
10338 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v8i32, V2, V1);
10339 if (isShuffleEquivalent(V1, V2, Mask, {10, 2, 11, 3, 14, 6, 15, 7}))
10340 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v8i32, V2, V1);
10343 // Try to use shift instructions.
10344 if (SDValue Shift =
10345 lowerVectorShuffleAsShift(DL, MVT::v8i32, V1, V2, Mask, DAG))
10348 if (SDValue Rotate = lowerVectorShuffleAsByteRotate(
10349 DL, MVT::v8i32, V1, V2, Mask, Subtarget, DAG))
10352 // If the shuffle patterns aren't repeated but it is a single input, directly
10353 // generate a cross-lane VPERMD instruction.
10354 if (isSingleInputShuffleMask(Mask)) {
10355 SDValue VPermMask[8];
10356 for (int i = 0; i < 8; ++i)
10357 VPermMask[i] = Mask[i] < 0 ? DAG.getUNDEF(MVT::i32)
10358 : DAG.getConstant(Mask[i], DL, MVT::i32);
10359 return DAG.getNode(
10360 X86ISD::VPERMV, DL, MVT::v8i32,
10361 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v8i32, VPermMask), V1);
10364 // Try to simplify this by merging 128-bit lanes to enable a lane-based
10366 if (SDValue Result = lowerVectorShuffleByMerging128BitLanes(
10367 DL, MVT::v8i32, V1, V2, Mask, Subtarget, DAG))
10370 // Otherwise fall back on generic blend lowering.
10371 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v8i32, V1, V2,
10375 /// \brief Handle lowering of 16-lane 16-bit integer shuffles.
10377 /// This routine is only called when we have AVX2 and thus a reasonable
10378 /// instruction set for v16i16 shuffling..
10379 static SDValue lowerV16I16VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10380 const X86Subtarget *Subtarget,
10381 SelectionDAG &DAG) {
10383 assert(V1.getSimpleValueType() == MVT::v16i16 && "Bad operand type!");
10384 assert(V2.getSimpleValueType() == MVT::v16i16 && "Bad operand type!");
10385 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10386 ArrayRef<int> Mask = SVOp->getMask();
10387 assert(Mask.size() == 16 && "Unexpected mask size for v16 shuffle!");
10388 assert(Subtarget->hasAVX2() && "We can only lower v16i16 with AVX2!");
10390 // Whenever we can lower this as a zext, that instruction is strictly faster
10391 // than any alternative. It also allows us to fold memory operands into the
10392 // shuffle in many cases.
10393 if (SDValue ZExt = lowerVectorShuffleAsZeroOrAnyExtend(DL, MVT::v16i16, V1, V2,
10394 Mask, Subtarget, DAG))
10397 // Check for being able to broadcast a single element.
10398 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v16i16, V1,
10399 Mask, Subtarget, DAG))
10402 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v16i16, V1, V2, Mask,
10406 // Use dedicated unpack instructions for masks that match their pattern.
10407 if (isShuffleEquivalent(V1, V2, Mask,
10408 {// First 128-bit lane:
10409 0, 16, 1, 17, 2, 18, 3, 19,
10410 // Second 128-bit lane:
10411 8, 24, 9, 25, 10, 26, 11, 27}))
10412 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v16i16, V1, V2);
10413 if (isShuffleEquivalent(V1, V2, Mask,
10414 {// First 128-bit lane:
10415 4, 20, 5, 21, 6, 22, 7, 23,
10416 // Second 128-bit lane:
10417 12, 28, 13, 29, 14, 30, 15, 31}))
10418 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v16i16, V1, V2);
10420 // Try to use shift instructions.
10421 if (SDValue Shift =
10422 lowerVectorShuffleAsShift(DL, MVT::v16i16, V1, V2, Mask, DAG))
10425 // Try to use byte rotation instructions.
10426 if (SDValue Rotate = lowerVectorShuffleAsByteRotate(
10427 DL, MVT::v16i16, V1, V2, Mask, Subtarget, DAG))
10430 if (isSingleInputShuffleMask(Mask)) {
10431 // There are no generalized cross-lane shuffle operations available on i16
10433 if (is128BitLaneCrossingShuffleMask(MVT::v16i16, Mask))
10434 return lowerVectorShuffleAsLanePermuteAndBlend(DL, MVT::v16i16, V1, V2,
10437 SmallVector<int, 8> RepeatedMask;
10438 if (is128BitLaneRepeatedShuffleMask(MVT::v16i16, Mask, RepeatedMask)) {
10439 // As this is a single-input shuffle, the repeated mask should be
10440 // a strictly valid v8i16 mask that we can pass through to the v8i16
10441 // lowering to handle even the v16 case.
10442 return lowerV8I16GeneralSingleInputVectorShuffle(
10443 DL, MVT::v16i16, V1, RepeatedMask, Subtarget, DAG);
10446 SDValue PSHUFBMask[32];
10447 for (int i = 0; i < 16; ++i) {
10448 if (Mask[i] == -1) {
10449 PSHUFBMask[2 * i] = PSHUFBMask[2 * i + 1] = DAG.getUNDEF(MVT::i8);
10453 int M = i < 8 ? Mask[i] : Mask[i] - 8;
10454 assert(M >= 0 && M < 8 && "Invalid single-input mask!");
10455 PSHUFBMask[2 * i] = DAG.getConstant(2 * M, DL, MVT::i8);
10456 PSHUFBMask[2 * i + 1] = DAG.getConstant(2 * M + 1, DL, MVT::i8);
10458 return DAG.getBitcast(MVT::v16i16,
10459 DAG.getNode(X86ISD::PSHUFB, DL, MVT::v32i8,
10460 DAG.getBitcast(MVT::v32i8, V1),
10461 DAG.getNode(ISD::BUILD_VECTOR, DL,
10462 MVT::v32i8, PSHUFBMask)));
10465 // Try to simplify this by merging 128-bit lanes to enable a lane-based
10467 if (SDValue Result = lowerVectorShuffleByMerging128BitLanes(
10468 DL, MVT::v16i16, V1, V2, Mask, Subtarget, DAG))
10471 // Otherwise fall back on generic lowering.
10472 return lowerVectorShuffleAsSplitOrBlend(DL, MVT::v16i16, V1, V2, Mask, DAG);
10475 /// \brief Handle lowering of 32-lane 8-bit integer shuffles.
10477 /// This routine is only called when we have AVX2 and thus a reasonable
10478 /// instruction set for v32i8 shuffling..
10479 static SDValue lowerV32I8VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10480 const X86Subtarget *Subtarget,
10481 SelectionDAG &DAG) {
10483 assert(V1.getSimpleValueType() == MVT::v32i8 && "Bad operand type!");
10484 assert(V2.getSimpleValueType() == MVT::v32i8 && "Bad operand type!");
10485 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10486 ArrayRef<int> Mask = SVOp->getMask();
10487 assert(Mask.size() == 32 && "Unexpected mask size for v32 shuffle!");
10488 assert(Subtarget->hasAVX2() && "We can only lower v32i8 with AVX2!");
10490 // Whenever we can lower this as a zext, that instruction is strictly faster
10491 // than any alternative. It also allows us to fold memory operands into the
10492 // shuffle in many cases.
10493 if (SDValue ZExt = lowerVectorShuffleAsZeroOrAnyExtend(DL, MVT::v32i8, V1, V2,
10494 Mask, Subtarget, DAG))
10497 // Check for being able to broadcast a single element.
10498 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v32i8, V1,
10499 Mask, Subtarget, DAG))
10502 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v32i8, V1, V2, Mask,
10506 // Use dedicated unpack instructions for masks that match their pattern.
10507 // Note that these are repeated 128-bit lane unpacks, not unpacks across all
10509 if (isShuffleEquivalent(
10511 {// First 128-bit lane:
10512 0, 32, 1, 33, 2, 34, 3, 35, 4, 36, 5, 37, 6, 38, 7, 39,
10513 // Second 128-bit lane:
10514 16, 48, 17, 49, 18, 50, 19, 51, 20, 52, 21, 53, 22, 54, 23, 55}))
10515 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v32i8, V1, V2);
10516 if (isShuffleEquivalent(
10518 {// First 128-bit lane:
10519 8, 40, 9, 41, 10, 42, 11, 43, 12, 44, 13, 45, 14, 46, 15, 47,
10520 // Second 128-bit lane:
10521 24, 56, 25, 57, 26, 58, 27, 59, 28, 60, 29, 61, 30, 62, 31, 63}))
10522 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v32i8, V1, V2);
10524 // Try to use shift instructions.
10525 if (SDValue Shift =
10526 lowerVectorShuffleAsShift(DL, MVT::v32i8, V1, V2, Mask, DAG))
10529 // Try to use byte rotation instructions.
10530 if (SDValue Rotate = lowerVectorShuffleAsByteRotate(
10531 DL, MVT::v32i8, V1, V2, Mask, Subtarget, DAG))
10534 if (isSingleInputShuffleMask(Mask)) {
10535 // There are no generalized cross-lane shuffle operations available on i8
10537 if (is128BitLaneCrossingShuffleMask(MVT::v32i8, Mask))
10538 return lowerVectorShuffleAsLanePermuteAndBlend(DL, MVT::v32i8, V1, V2,
10541 SDValue PSHUFBMask[32];
10542 for (int i = 0; i < 32; ++i)
10545 ? DAG.getUNDEF(MVT::i8)
10546 : DAG.getConstant(Mask[i] < 16 ? Mask[i] : Mask[i] - 16, DL,
10549 return DAG.getNode(
10550 X86ISD::PSHUFB, DL, MVT::v32i8, V1,
10551 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v32i8, PSHUFBMask));
10554 // Try to simplify this by merging 128-bit lanes to enable a lane-based
10556 if (SDValue Result = lowerVectorShuffleByMerging128BitLanes(
10557 DL, MVT::v32i8, V1, V2, Mask, Subtarget, DAG))
10560 // Otherwise fall back on generic lowering.
10561 return lowerVectorShuffleAsSplitOrBlend(DL, MVT::v32i8, V1, V2, Mask, DAG);
10564 /// \brief High-level routine to lower various 256-bit x86 vector shuffles.
10566 /// This routine either breaks down the specific type of a 256-bit x86 vector
10567 /// shuffle or splits it into two 128-bit shuffles and fuses the results back
10568 /// together based on the available instructions.
10569 static SDValue lower256BitVectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10570 MVT VT, const X86Subtarget *Subtarget,
10571 SelectionDAG &DAG) {
10573 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10574 ArrayRef<int> Mask = SVOp->getMask();
10576 // If we have a single input to the zero element, insert that into V1 if we
10577 // can do so cheaply.
10578 int NumElts = VT.getVectorNumElements();
10579 int NumV2Elements = std::count_if(Mask.begin(), Mask.end(), [NumElts](int M) {
10580 return M >= NumElts;
10583 if (NumV2Elements == 1 && Mask[0] >= NumElts)
10584 if (SDValue Insertion = lowerVectorShuffleAsElementInsertion(
10585 DL, VT, V1, V2, Mask, Subtarget, DAG))
10588 // There is a really nice hard cut-over between AVX1 and AVX2 that means we can
10589 // check for those subtargets here and avoid much of the subtarget querying in
10590 // the per-vector-type lowering routines. With AVX1 we have essentially *zero*
10591 // ability to manipulate a 256-bit vector with integer types. Since we'll use
10592 // floating point types there eventually, just immediately cast everything to
10593 // a float and operate entirely in that domain.
10594 if (VT.isInteger() && !Subtarget->hasAVX2()) {
10595 int ElementBits = VT.getScalarSizeInBits();
10596 if (ElementBits < 32)
10597 // No floating point type available, decompose into 128-bit vectors.
10598 return splitAndLowerVectorShuffle(DL, VT, V1, V2, Mask, DAG);
10600 MVT FpVT = MVT::getVectorVT(MVT::getFloatingPointVT(ElementBits),
10601 VT.getVectorNumElements());
10602 V1 = DAG.getBitcast(FpVT, V1);
10603 V2 = DAG.getBitcast(FpVT, V2);
10604 return DAG.getBitcast(VT, DAG.getVectorShuffle(FpVT, DL, V1, V2, Mask));
10607 switch (VT.SimpleTy) {
10609 return lowerV4F64VectorShuffle(Op, V1, V2, Subtarget, DAG);
10611 return lowerV4I64VectorShuffle(Op, V1, V2, Subtarget, DAG);
10613 return lowerV8F32VectorShuffle(Op, V1, V2, Subtarget, DAG);
10615 return lowerV8I32VectorShuffle(Op, V1, V2, Subtarget, DAG);
10617 return lowerV16I16VectorShuffle(Op, V1, V2, Subtarget, DAG);
10619 return lowerV32I8VectorShuffle(Op, V1, V2, Subtarget, DAG);
10622 llvm_unreachable("Not a valid 256-bit x86 vector type!");
10626 static SDValue lowerVectorShuffleWithPERMV(SDLoc DL, MVT VT,
10627 ArrayRef<int> Mask, SDValue V1,
10628 SDValue V2, SelectionDAG &DAG) {
10630 assert(VT.getScalarSizeInBits() >= 16 && "Unexpected data type for PERMV");
10632 MVT MaskEltVT = MVT::getIntegerVT(VT.getScalarSizeInBits());
10633 MVT MaskVecVT = MVT::getVectorVT(MaskEltVT, VT.getVectorNumElements());
10635 SmallVector<SDValue, 32> VPermMask;
10636 for (unsigned i = 0; i < VT.getVectorNumElements(); ++i)
10637 VPermMask.push_back(Mask[i] < 0 ? DAG.getUNDEF(MaskEltVT) :
10638 DAG.getConstant(Mask[i], DL, MaskEltVT));
10639 SDValue MaskNode = DAG.getNode(ISD::BUILD_VECTOR, DL, MaskVecVT,
10641 if (isSingleInputShuffleMask(Mask))
10642 return DAG.getNode(X86ISD::VPERMV, DL, VT, MaskNode, V1);
10644 return DAG.getNode(X86ISD::VPERMV3, DL, VT, V1, MaskNode, V2);
10647 /// \brief Handle lowering of 8-lane 64-bit floating point shuffles.
10648 static SDValue lowerV8F64VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10649 const X86Subtarget *Subtarget,
10650 SelectionDAG &DAG) {
10652 assert(V1.getSimpleValueType() == MVT::v8f64 && "Bad operand type!");
10653 assert(V2.getSimpleValueType() == MVT::v8f64 && "Bad operand type!");
10654 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10655 ArrayRef<int> Mask = SVOp->getMask();
10656 assert(Mask.size() == 8 && "Unexpected mask size for v8 shuffle!");
10658 if (SDValue Unpck =
10659 lowerVectorShuffleWithUNPCK(DL, MVT::v8f64, Mask, V1, V2, DAG))
10662 return lowerVectorShuffleWithPERMV(DL, MVT::v8f64, Mask, V1, V2, DAG);
10665 /// \brief Handle lowering of 16-lane 32-bit floating point shuffles.
10666 static SDValue lowerV16F32VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10667 const X86Subtarget *Subtarget,
10668 SelectionDAG &DAG) {
10670 assert(V1.getSimpleValueType() == MVT::v16f32 && "Bad operand type!");
10671 assert(V2.getSimpleValueType() == MVT::v16f32 && "Bad operand type!");
10672 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10673 ArrayRef<int> Mask = SVOp->getMask();
10674 assert(Mask.size() == 16 && "Unexpected mask size for v16 shuffle!");
10676 if (SDValue Unpck =
10677 lowerVectorShuffleWithUNPCK(DL, MVT::v16f32, Mask, V1, V2, DAG))
10680 return lowerVectorShuffleWithPERMV(DL, MVT::v16f32, Mask, V1, V2, DAG);
10683 /// \brief Handle lowering of 8-lane 64-bit integer shuffles.
10684 static SDValue lowerV8I64VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10685 const X86Subtarget *Subtarget,
10686 SelectionDAG &DAG) {
10688 assert(V1.getSimpleValueType() == MVT::v8i64 && "Bad operand type!");
10689 assert(V2.getSimpleValueType() == MVT::v8i64 && "Bad operand type!");
10690 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10691 ArrayRef<int> Mask = SVOp->getMask();
10692 assert(Mask.size() == 8 && "Unexpected mask size for v8 shuffle!");
10694 if (SDValue Unpck =
10695 lowerVectorShuffleWithUNPCK(DL, MVT::v8i64, Mask, V1, V2, DAG))
10698 return lowerVectorShuffleWithPERMV(DL, MVT::v8i64, Mask, V1, V2, DAG);
10701 /// \brief Handle lowering of 16-lane 32-bit integer shuffles.
10702 static SDValue lowerV16I32VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10703 const X86Subtarget *Subtarget,
10704 SelectionDAG &DAG) {
10706 assert(V1.getSimpleValueType() == MVT::v16i32 && "Bad operand type!");
10707 assert(V2.getSimpleValueType() == MVT::v16i32 && "Bad operand type!");
10708 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10709 ArrayRef<int> Mask = SVOp->getMask();
10710 assert(Mask.size() == 16 && "Unexpected mask size for v16 shuffle!");
10712 if (SDValue Unpck =
10713 lowerVectorShuffleWithUNPCK(DL, MVT::v16i32, Mask, V1, V2, DAG))
10716 return lowerVectorShuffleWithPERMV(DL, MVT::v16i32, Mask, V1, V2, DAG);
10719 /// \brief Handle lowering of 32-lane 16-bit integer shuffles.
10720 static SDValue lowerV32I16VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10721 const X86Subtarget *Subtarget,
10722 SelectionDAG &DAG) {
10724 assert(V1.getSimpleValueType() == MVT::v32i16 && "Bad operand type!");
10725 assert(V2.getSimpleValueType() == MVT::v32i16 && "Bad operand type!");
10726 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10727 ArrayRef<int> Mask = SVOp->getMask();
10728 assert(Mask.size() == 32 && "Unexpected mask size for v32 shuffle!");
10729 assert(Subtarget->hasBWI() && "We can only lower v32i16 with AVX-512-BWI!");
10731 return lowerVectorShuffleWithPERMV(DL, MVT::v32i16, Mask, V1, V2, DAG);
10734 /// \brief Handle lowering of 64-lane 8-bit integer shuffles.
10735 static SDValue lowerV64I8VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10736 const X86Subtarget *Subtarget,
10737 SelectionDAG &DAG) {
10739 assert(V1.getSimpleValueType() == MVT::v64i8 && "Bad operand type!");
10740 assert(V2.getSimpleValueType() == MVT::v64i8 && "Bad operand type!");
10741 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10742 ArrayRef<int> Mask = SVOp->getMask();
10743 assert(Mask.size() == 64 && "Unexpected mask size for v64 shuffle!");
10744 assert(Subtarget->hasBWI() && "We can only lower v64i8 with AVX-512-BWI!");
10746 // FIXME: Implement direct support for this type!
10747 return splitAndLowerVectorShuffle(DL, MVT::v64i8, V1, V2, Mask, DAG);
10750 /// \brief High-level routine to lower various 512-bit x86 vector shuffles.
10752 /// This routine either breaks down the specific type of a 512-bit x86 vector
10753 /// shuffle or splits it into two 256-bit shuffles and fuses the results back
10754 /// together based on the available instructions.
10755 static SDValue lower512BitVectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10756 MVT VT, const X86Subtarget *Subtarget,
10757 SelectionDAG &DAG) {
10759 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10760 ArrayRef<int> Mask = SVOp->getMask();
10761 assert(Subtarget->hasAVX512() &&
10762 "Cannot lower 512-bit vectors w/ basic ISA!");
10764 // Check for being able to broadcast a single element.
10765 if (SDValue Broadcast =
10766 lowerVectorShuffleAsBroadcast(DL, VT, V1, Mask, Subtarget, DAG))
10769 // Dispatch to each element type for lowering. If we don't have supprot for
10770 // specific element type shuffles at 512 bits, immediately split them and
10771 // lower them. Each lowering routine of a given type is allowed to assume that
10772 // the requisite ISA extensions for that element type are available.
10773 switch (VT.SimpleTy) {
10775 return lowerV8F64VectorShuffle(Op, V1, V2, Subtarget, DAG);
10777 return lowerV16F32VectorShuffle(Op, V1, V2, Subtarget, DAG);
10779 return lowerV8I64VectorShuffle(Op, V1, V2, Subtarget, DAG);
10781 return lowerV16I32VectorShuffle(Op, V1, V2, Subtarget, DAG);
10783 if (Subtarget->hasBWI())
10784 return lowerV32I16VectorShuffle(Op, V1, V2, Subtarget, DAG);
10787 if (Subtarget->hasBWI())
10788 return lowerV64I8VectorShuffle(Op, V1, V2, Subtarget, DAG);
10792 llvm_unreachable("Not a valid 512-bit x86 vector type!");
10795 // Otherwise fall back on splitting.
10796 return splitAndLowerVectorShuffle(DL, VT, V1, V2, Mask, DAG);
10799 // Lower vXi1 vector shuffles.
10800 // There is no a dedicated instruction on AVX-512 that shuffles the masks.
10801 // The only way to shuffle bits is to sign-extend the mask vector to SIMD
10802 // vector, shuffle and then truncate it back.
10803 static SDValue lower1BitVectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10804 MVT VT, const X86Subtarget *Subtarget,
10805 SelectionDAG &DAG) {
10807 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10808 ArrayRef<int> Mask = SVOp->getMask();
10809 assert(Subtarget->hasAVX512() &&
10810 "Cannot lower 512-bit vectors w/o basic ISA!");
10812 switch (VT.SimpleTy) {
10814 assert(false && "Expected a vector of i1 elements");
10817 ExtVT = MVT::v2i64;
10820 ExtVT = MVT::v4i32;
10823 ExtVT = MVT::v8i64; // Take 512-bit type, more shuffles on KNL
10826 ExtVT = MVT::v16i32;
10829 ExtVT = MVT::v32i16;
10832 ExtVT = MVT::v64i8;
10836 if (ISD::isBuildVectorAllZeros(V1.getNode()))
10837 V1 = getZeroVector(ExtVT, Subtarget, DAG, DL);
10838 else if (ISD::isBuildVectorAllOnes(V1.getNode()))
10839 V1 = getOnesVector(ExtVT, Subtarget, DAG, DL);
10841 V1 = DAG.getNode(ISD::SIGN_EXTEND, DL, ExtVT, V1);
10844 V2 = DAG.getUNDEF(ExtVT);
10845 else if (ISD::isBuildVectorAllZeros(V2.getNode()))
10846 V2 = getZeroVector(ExtVT, Subtarget, DAG, DL);
10847 else if (ISD::isBuildVectorAllOnes(V2.getNode()))
10848 V2 = getOnesVector(ExtVT, Subtarget, DAG, DL);
10850 V2 = DAG.getNode(ISD::SIGN_EXTEND, DL, ExtVT, V2);
10851 return DAG.getNode(ISD::TRUNCATE, DL, VT,
10852 DAG.getVectorShuffle(ExtVT, DL, V1, V2, Mask));
10854 /// \brief Top-level lowering for x86 vector shuffles.
10856 /// This handles decomposition, canonicalization, and lowering of all x86
10857 /// vector shuffles. Most of the specific lowering strategies are encapsulated
10858 /// above in helper routines. The canonicalization attempts to widen shuffles
10859 /// to involve fewer lanes of wider elements, consolidate symmetric patterns
10860 /// s.t. only one of the two inputs needs to be tested, etc.
10861 static SDValue lowerVectorShuffle(SDValue Op, const X86Subtarget *Subtarget,
10862 SelectionDAG &DAG) {
10863 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10864 ArrayRef<int> Mask = SVOp->getMask();
10865 SDValue V1 = Op.getOperand(0);
10866 SDValue V2 = Op.getOperand(1);
10867 MVT VT = Op.getSimpleValueType();
10868 int NumElements = VT.getVectorNumElements();
10870 bool Is1BitVector = (VT.getScalarType() == MVT::i1);
10872 assert((VT.getSizeInBits() != 64 || Is1BitVector) &&
10873 "Can't lower MMX shuffles");
10875 bool V1IsUndef = V1.getOpcode() == ISD::UNDEF;
10876 bool V2IsUndef = V2.getOpcode() == ISD::UNDEF;
10877 if (V1IsUndef && V2IsUndef)
10878 return DAG.getUNDEF(VT);
10880 // When we create a shuffle node we put the UNDEF node to second operand,
10881 // but in some cases the first operand may be transformed to UNDEF.
10882 // In this case we should just commute the node.
10884 return DAG.getCommutedVectorShuffle(*SVOp);
10886 // Check for non-undef masks pointing at an undef vector and make the masks
10887 // undef as well. This makes it easier to match the shuffle based solely on
10891 if (M >= NumElements) {
10892 SmallVector<int, 8> NewMask(Mask.begin(), Mask.end());
10893 for (int &M : NewMask)
10894 if (M >= NumElements)
10896 return DAG.getVectorShuffle(VT, dl, V1, V2, NewMask);
10899 // We actually see shuffles that are entirely re-arrangements of a set of
10900 // zero inputs. This mostly happens while decomposing complex shuffles into
10901 // simple ones. Directly lower these as a buildvector of zeros.
10902 SmallBitVector Zeroable = computeZeroableShuffleElements(Mask, V1, V2);
10903 if (Zeroable.all())
10904 return getZeroVector(VT, Subtarget, DAG, dl);
10906 // Try to collapse shuffles into using a vector type with fewer elements but
10907 // wider element types. We cap this to not form integers or floating point
10908 // elements wider than 64 bits, but it might be interesting to form i128
10909 // integers to handle flipping the low and high halves of AVX 256-bit vectors.
10910 SmallVector<int, 16> WidenedMask;
10911 if (VT.getScalarSizeInBits() < 64 && !Is1BitVector &&
10912 canWidenShuffleElements(Mask, WidenedMask)) {
10913 MVT NewEltVT = VT.isFloatingPoint()
10914 ? MVT::getFloatingPointVT(VT.getScalarSizeInBits() * 2)
10915 : MVT::getIntegerVT(VT.getScalarSizeInBits() * 2);
10916 MVT NewVT = MVT::getVectorVT(NewEltVT, VT.getVectorNumElements() / 2);
10917 // Make sure that the new vector type is legal. For example, v2f64 isn't
10919 if (DAG.getTargetLoweringInfo().isTypeLegal(NewVT)) {
10920 V1 = DAG.getBitcast(NewVT, V1);
10921 V2 = DAG.getBitcast(NewVT, V2);
10922 return DAG.getBitcast(
10923 VT, DAG.getVectorShuffle(NewVT, dl, V1, V2, WidenedMask));
10927 int NumV1Elements = 0, NumUndefElements = 0, NumV2Elements = 0;
10928 for (int M : SVOp->getMask())
10930 ++NumUndefElements;
10931 else if (M < NumElements)
10936 // Commute the shuffle as needed such that more elements come from V1 than
10937 // V2. This allows us to match the shuffle pattern strictly on how many
10938 // elements come from V1 without handling the symmetric cases.
10939 if (NumV2Elements > NumV1Elements)
10940 return DAG.getCommutedVectorShuffle(*SVOp);
10942 // When the number of V1 and V2 elements are the same, try to minimize the
10943 // number of uses of V2 in the low half of the vector. When that is tied,
10944 // ensure that the sum of indices for V1 is equal to or lower than the sum
10945 // indices for V2. When those are equal, try to ensure that the number of odd
10946 // indices for V1 is lower than the number of odd indices for V2.
10947 if (NumV1Elements == NumV2Elements) {
10948 int LowV1Elements = 0, LowV2Elements = 0;
10949 for (int M : SVOp->getMask().slice(0, NumElements / 2))
10950 if (M >= NumElements)
10954 if (LowV2Elements > LowV1Elements) {
10955 return DAG.getCommutedVectorShuffle(*SVOp);
10956 } else if (LowV2Elements == LowV1Elements) {
10957 int SumV1Indices = 0, SumV2Indices = 0;
10958 for (int i = 0, Size = SVOp->getMask().size(); i < Size; ++i)
10959 if (SVOp->getMask()[i] >= NumElements)
10961 else if (SVOp->getMask()[i] >= 0)
10963 if (SumV2Indices < SumV1Indices) {
10964 return DAG.getCommutedVectorShuffle(*SVOp);
10965 } else if (SumV2Indices == SumV1Indices) {
10966 int NumV1OddIndices = 0, NumV2OddIndices = 0;
10967 for (int i = 0, Size = SVOp->getMask().size(); i < Size; ++i)
10968 if (SVOp->getMask()[i] >= NumElements)
10969 NumV2OddIndices += i % 2;
10970 else if (SVOp->getMask()[i] >= 0)
10971 NumV1OddIndices += i % 2;
10972 if (NumV2OddIndices < NumV1OddIndices)
10973 return DAG.getCommutedVectorShuffle(*SVOp);
10978 // For each vector width, delegate to a specialized lowering routine.
10979 if (VT.getSizeInBits() == 128)
10980 return lower128BitVectorShuffle(Op, V1, V2, VT, Subtarget, DAG);
10982 if (VT.getSizeInBits() == 256)
10983 return lower256BitVectorShuffle(Op, V1, V2, VT, Subtarget, DAG);
10985 if (VT.getSizeInBits() == 512)
10986 return lower512BitVectorShuffle(Op, V1, V2, VT, Subtarget, DAG);
10989 return lower1BitVectorShuffle(Op, V1, V2, VT, Subtarget, DAG);
10990 llvm_unreachable("Unimplemented!");
10993 // This function assumes its argument is a BUILD_VECTOR of constants or
10994 // undef SDNodes. i.e: ISD::isBuildVectorOfConstantSDNodes(BuildVector) is
10996 static bool BUILD_VECTORtoBlendMask(BuildVectorSDNode *BuildVector,
10997 unsigned &MaskValue) {
10999 unsigned NumElems = BuildVector->getNumOperands();
11000 // There are 2 lanes if (NumElems > 8), and 1 lane otherwise.
11001 unsigned NumLanes = (NumElems - 1) / 8 + 1;
11002 unsigned NumElemsInLane = NumElems / NumLanes;
11004 // Blend for v16i16 should be symmetric for the both lanes.
11005 for (unsigned i = 0; i < NumElemsInLane; ++i) {
11006 SDValue EltCond = BuildVector->getOperand(i);
11007 SDValue SndLaneEltCond =
11008 (NumLanes == 2) ? BuildVector->getOperand(i + NumElemsInLane) : EltCond;
11010 int Lane1Cond = -1, Lane2Cond = -1;
11011 if (isa<ConstantSDNode>(EltCond))
11012 Lane1Cond = !isZero(EltCond);
11013 if (isa<ConstantSDNode>(SndLaneEltCond))
11014 Lane2Cond = !isZero(SndLaneEltCond);
11016 if (Lane1Cond == Lane2Cond || Lane2Cond < 0)
11017 // Lane1Cond != 0, means we want the first argument.
11018 // Lane1Cond == 0, means we want the second argument.
11019 // The encoding of this argument is 0 for the first argument, 1
11020 // for the second. Therefore, invert the condition.
11021 MaskValue |= !Lane1Cond << i;
11022 else if (Lane1Cond < 0)
11023 MaskValue |= !Lane2Cond << i;
11030 /// \brief Try to lower a VSELECT instruction to a vector shuffle.
11031 static SDValue lowerVSELECTtoVectorShuffle(SDValue Op,
11032 const X86Subtarget *Subtarget,
11033 SelectionDAG &DAG) {
11034 SDValue Cond = Op.getOperand(0);
11035 SDValue LHS = Op.getOperand(1);
11036 SDValue RHS = Op.getOperand(2);
11038 MVT VT = Op.getSimpleValueType();
11040 if (!ISD::isBuildVectorOfConstantSDNodes(Cond.getNode()))
11042 auto *CondBV = cast<BuildVectorSDNode>(Cond);
11044 // Only non-legal VSELECTs reach this lowering, convert those into generic
11045 // shuffles and re-use the shuffle lowering path for blends.
11046 SmallVector<int, 32> Mask;
11047 for (int i = 0, Size = VT.getVectorNumElements(); i < Size; ++i) {
11048 SDValue CondElt = CondBV->getOperand(i);
11050 isa<ConstantSDNode>(CondElt) ? i + (isZero(CondElt) ? Size : 0) : -1);
11052 return DAG.getVectorShuffle(VT, dl, LHS, RHS, Mask);
11055 SDValue X86TargetLowering::LowerVSELECT(SDValue Op, SelectionDAG &DAG) const {
11056 // A vselect where all conditions and data are constants can be optimized into
11057 // a single vector load by SelectionDAGLegalize::ExpandBUILD_VECTOR().
11058 if (ISD::isBuildVectorOfConstantSDNodes(Op.getOperand(0).getNode()) &&
11059 ISD::isBuildVectorOfConstantSDNodes(Op.getOperand(1).getNode()) &&
11060 ISD::isBuildVectorOfConstantSDNodes(Op.getOperand(2).getNode()))
11063 // Try to lower this to a blend-style vector shuffle. This can handle all
11064 // constant condition cases.
11065 if (SDValue BlendOp = lowerVSELECTtoVectorShuffle(Op, Subtarget, DAG))
11068 // Variable blends are only legal from SSE4.1 onward.
11069 if (!Subtarget->hasSSE41())
11072 // Only some types will be legal on some subtargets. If we can emit a legal
11073 // VSELECT-matching blend, return Op, and but if we need to expand, return
11075 switch (Op.getSimpleValueType().SimpleTy) {
11077 // Most of the vector types have blends past SSE4.1.
11081 // The byte blends for AVX vectors were introduced only in AVX2.
11082 if (Subtarget->hasAVX2())
11089 // AVX-512 BWI and VLX features support VSELECT with i16 elements.
11090 if (Subtarget->hasBWI() && Subtarget->hasVLX())
11093 // FIXME: We should custom lower this by fixing the condition and using i8
11099 static SDValue LowerEXTRACT_VECTOR_ELT_SSE4(SDValue Op, SelectionDAG &DAG) {
11100 MVT VT = Op.getSimpleValueType();
11103 if (!Op.getOperand(0).getSimpleValueType().is128BitVector())
11106 if (VT.getSizeInBits() == 8) {
11107 SDValue Extract = DAG.getNode(X86ISD::PEXTRB, dl, MVT::i32,
11108 Op.getOperand(0), Op.getOperand(1));
11109 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Extract,
11110 DAG.getValueType(VT));
11111 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
11114 if (VT.getSizeInBits() == 16) {
11115 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
11116 // If Idx is 0, it's cheaper to do a move instead of a pextrw.
11118 return DAG.getNode(
11119 ISD::TRUNCATE, dl, MVT::i16,
11120 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
11121 DAG.getBitcast(MVT::v4i32, Op.getOperand(0)),
11122 Op.getOperand(1)));
11123 SDValue Extract = DAG.getNode(X86ISD::PEXTRW, dl, MVT::i32,
11124 Op.getOperand(0), Op.getOperand(1));
11125 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Extract,
11126 DAG.getValueType(VT));
11127 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
11130 if (VT == MVT::f32) {
11131 // EXTRACTPS outputs to a GPR32 register which will require a movd to copy
11132 // the result back to FR32 register. It's only worth matching if the
11133 // result has a single use which is a store or a bitcast to i32. And in
11134 // the case of a store, it's not worth it if the index is a constant 0,
11135 // because a MOVSSmr can be used instead, which is smaller and faster.
11136 if (!Op.hasOneUse())
11138 SDNode *User = *Op.getNode()->use_begin();
11139 if ((User->getOpcode() != ISD::STORE ||
11140 (isa<ConstantSDNode>(Op.getOperand(1)) &&
11141 cast<ConstantSDNode>(Op.getOperand(1))->isNullValue())) &&
11142 (User->getOpcode() != ISD::BITCAST ||
11143 User->getValueType(0) != MVT::i32))
11145 SDValue Extract = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
11146 DAG.getBitcast(MVT::v4i32, Op.getOperand(0)),
11148 return DAG.getBitcast(MVT::f32, Extract);
11151 if (VT == MVT::i32 || VT == MVT::i64) {
11152 // ExtractPS/pextrq works with constant index.
11153 if (isa<ConstantSDNode>(Op.getOperand(1)))
11159 /// Extract one bit from mask vector, like v16i1 or v8i1.
11160 /// AVX-512 feature.
11162 X86TargetLowering::ExtractBitFromMaskVector(SDValue Op, SelectionDAG &DAG) const {
11163 SDValue Vec = Op.getOperand(0);
11165 MVT VecVT = Vec.getSimpleValueType();
11166 SDValue Idx = Op.getOperand(1);
11167 MVT EltVT = Op.getSimpleValueType();
11169 assert((EltVT == MVT::i1) && "Unexpected operands in ExtractBitFromMaskVector");
11170 assert((VecVT.getVectorNumElements() <= 16 || Subtarget->hasBWI()) &&
11171 "Unexpected vector type in ExtractBitFromMaskVector");
11173 // variable index can't be handled in mask registers,
11174 // extend vector to VR512
11175 if (!isa<ConstantSDNode>(Idx)) {
11176 MVT ExtVT = (VecVT == MVT::v8i1 ? MVT::v8i64 : MVT::v16i32);
11177 SDValue Ext = DAG.getNode(ISD::ZERO_EXTEND, dl, ExtVT, Vec);
11178 SDValue Elt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl,
11179 ExtVT.getVectorElementType(), Ext, Idx);
11180 return DAG.getNode(ISD::TRUNCATE, dl, EltVT, Elt);
11183 unsigned IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
11184 const TargetRegisterClass* rc = getRegClassFor(VecVT);
11185 if (!Subtarget->hasDQI() && (VecVT.getVectorNumElements() <= 8))
11186 rc = getRegClassFor(MVT::v16i1);
11187 unsigned MaxSift = rc->getSize()*8 - 1;
11188 Vec = DAG.getNode(X86ISD::VSHLI, dl, VecVT, Vec,
11189 DAG.getConstant(MaxSift - IdxVal, dl, MVT::i8));
11190 Vec = DAG.getNode(X86ISD::VSRLI, dl, VecVT, Vec,
11191 DAG.getConstant(MaxSift, dl, MVT::i8));
11192 return DAG.getNode(X86ISD::VEXTRACT, dl, MVT::i1, Vec,
11193 DAG.getIntPtrConstant(0, dl));
11197 X86TargetLowering::LowerEXTRACT_VECTOR_ELT(SDValue Op,
11198 SelectionDAG &DAG) const {
11200 SDValue Vec = Op.getOperand(0);
11201 MVT VecVT = Vec.getSimpleValueType();
11202 SDValue Idx = Op.getOperand(1);
11204 if (Op.getSimpleValueType() == MVT::i1)
11205 return ExtractBitFromMaskVector(Op, DAG);
11207 if (!isa<ConstantSDNode>(Idx)) {
11208 if (VecVT.is512BitVector() ||
11209 (VecVT.is256BitVector() && Subtarget->hasInt256() &&
11210 VecVT.getVectorElementType().getSizeInBits() == 32)) {
11213 MVT::getIntegerVT(VecVT.getVectorElementType().getSizeInBits());
11214 MVT MaskVT = MVT::getVectorVT(MaskEltVT, VecVT.getSizeInBits() /
11215 MaskEltVT.getSizeInBits());
11217 Idx = DAG.getZExtOrTrunc(Idx, dl, MaskEltVT);
11218 auto PtrVT = getPointerTy(DAG.getDataLayout());
11219 SDValue Mask = DAG.getNode(X86ISD::VINSERT, dl, MaskVT,
11220 getZeroVector(MaskVT, Subtarget, DAG, dl), Idx,
11221 DAG.getConstant(0, dl, PtrVT));
11222 SDValue Perm = DAG.getNode(X86ISD::VPERMV, dl, VecVT, Mask, Vec);
11223 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, Op.getValueType(), Perm,
11224 DAG.getConstant(0, dl, PtrVT));
11229 // If this is a 256-bit vector result, first extract the 128-bit vector and
11230 // then extract the element from the 128-bit vector.
11231 if (VecVT.is256BitVector() || VecVT.is512BitVector()) {
11233 unsigned IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
11234 // Get the 128-bit vector.
11235 Vec = Extract128BitVector(Vec, IdxVal, DAG, dl);
11236 MVT EltVT = VecVT.getVectorElementType();
11238 unsigned ElemsPerChunk = 128 / EltVT.getSizeInBits();
11240 //if (IdxVal >= NumElems/2)
11241 // IdxVal -= NumElems/2;
11242 IdxVal -= (IdxVal/ElemsPerChunk)*ElemsPerChunk;
11243 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, Op.getValueType(), Vec,
11244 DAG.getConstant(IdxVal, dl, MVT::i32));
11247 assert(VecVT.is128BitVector() && "Unexpected vector length");
11249 if (Subtarget->hasSSE41())
11250 if (SDValue Res = LowerEXTRACT_VECTOR_ELT_SSE4(Op, DAG))
11253 MVT VT = Op.getSimpleValueType();
11254 // TODO: handle v16i8.
11255 if (VT.getSizeInBits() == 16) {
11256 SDValue Vec = Op.getOperand(0);
11257 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
11259 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i16,
11260 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
11261 DAG.getBitcast(MVT::v4i32, Vec),
11262 Op.getOperand(1)));
11263 // Transform it so it match pextrw which produces a 32-bit result.
11264 MVT EltVT = MVT::i32;
11265 SDValue Extract = DAG.getNode(X86ISD::PEXTRW, dl, EltVT,
11266 Op.getOperand(0), Op.getOperand(1));
11267 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, EltVT, Extract,
11268 DAG.getValueType(VT));
11269 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
11272 if (VT.getSizeInBits() == 32) {
11273 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
11277 // SHUFPS the element to the lowest double word, then movss.
11278 int Mask[4] = { static_cast<int>(Idx), -1, -1, -1 };
11279 MVT VVT = Op.getOperand(0).getSimpleValueType();
11280 SDValue Vec = DAG.getVectorShuffle(VVT, dl, Op.getOperand(0),
11281 DAG.getUNDEF(VVT), Mask);
11282 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, Vec,
11283 DAG.getIntPtrConstant(0, dl));
11286 if (VT.getSizeInBits() == 64) {
11287 // FIXME: .td only matches this for <2 x f64>, not <2 x i64> on 32b
11288 // FIXME: seems like this should be unnecessary if mov{h,l}pd were taught
11289 // to match extract_elt for f64.
11290 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
11294 // UNPCKHPD the element to the lowest double word, then movsd.
11295 // Note if the lower 64 bits of the result of the UNPCKHPD is then stored
11296 // to a f64mem, the whole operation is folded into a single MOVHPDmr.
11297 int Mask[2] = { 1, -1 };
11298 MVT VVT = Op.getOperand(0).getSimpleValueType();
11299 SDValue Vec = DAG.getVectorShuffle(VVT, dl, Op.getOperand(0),
11300 DAG.getUNDEF(VVT), Mask);
11301 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, Vec,
11302 DAG.getIntPtrConstant(0, dl));
11308 /// Insert one bit to mask vector, like v16i1 or v8i1.
11309 /// AVX-512 feature.
11311 X86TargetLowering::InsertBitToMaskVector(SDValue Op, SelectionDAG &DAG) const {
11313 SDValue Vec = Op.getOperand(0);
11314 SDValue Elt = Op.getOperand(1);
11315 SDValue Idx = Op.getOperand(2);
11316 MVT VecVT = Vec.getSimpleValueType();
11318 if (!isa<ConstantSDNode>(Idx)) {
11319 // Non constant index. Extend source and destination,
11320 // insert element and then truncate the result.
11321 MVT ExtVecVT = (VecVT == MVT::v8i1 ? MVT::v8i64 : MVT::v16i32);
11322 MVT ExtEltVT = (VecVT == MVT::v8i1 ? MVT::i64 : MVT::i32);
11323 SDValue ExtOp = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, ExtVecVT,
11324 DAG.getNode(ISD::ZERO_EXTEND, dl, ExtVecVT, Vec),
11325 DAG.getNode(ISD::ZERO_EXTEND, dl, ExtEltVT, Elt), Idx);
11326 return DAG.getNode(ISD::TRUNCATE, dl, VecVT, ExtOp);
11329 unsigned IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
11330 SDValue EltInVec = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VecVT, Elt);
11332 EltInVec = DAG.getNode(X86ISD::VSHLI, dl, VecVT, EltInVec,
11333 DAG.getConstant(IdxVal, dl, MVT::i8));
11334 if (Vec.getOpcode() == ISD::UNDEF)
11336 return DAG.getNode(ISD::OR, dl, VecVT, Vec, EltInVec);
11339 SDValue X86TargetLowering::LowerINSERT_VECTOR_ELT(SDValue Op,
11340 SelectionDAG &DAG) const {
11341 MVT VT = Op.getSimpleValueType();
11342 MVT EltVT = VT.getVectorElementType();
11344 if (EltVT == MVT::i1)
11345 return InsertBitToMaskVector(Op, DAG);
11348 SDValue N0 = Op.getOperand(0);
11349 SDValue N1 = Op.getOperand(1);
11350 SDValue N2 = Op.getOperand(2);
11351 if (!isa<ConstantSDNode>(N2))
11353 auto *N2C = cast<ConstantSDNode>(N2);
11354 unsigned IdxVal = N2C->getZExtValue();
11356 // If the vector is wider than 128 bits, extract the 128-bit subvector, insert
11357 // into that, and then insert the subvector back into the result.
11358 if (VT.is256BitVector() || VT.is512BitVector()) {
11359 // With a 256-bit vector, we can insert into the zero element efficiently
11360 // using a blend if we have AVX or AVX2 and the right data type.
11361 if (VT.is256BitVector() && IdxVal == 0) {
11362 // TODO: It is worthwhile to cast integer to floating point and back
11363 // and incur a domain crossing penalty if that's what we'll end up
11364 // doing anyway after extracting to a 128-bit vector.
11365 if ((Subtarget->hasAVX() && (EltVT == MVT::f64 || EltVT == MVT::f32)) ||
11366 (Subtarget->hasAVX2() && EltVT == MVT::i32)) {
11367 SDValue N1Vec = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, N1);
11368 N2 = DAG.getIntPtrConstant(1, dl);
11369 return DAG.getNode(X86ISD::BLENDI, dl, VT, N0, N1Vec, N2);
11373 // Get the desired 128-bit vector chunk.
11374 SDValue V = Extract128BitVector(N0, IdxVal, DAG, dl);
11376 // Insert the element into the desired chunk.
11377 unsigned NumEltsIn128 = 128 / EltVT.getSizeInBits();
11378 unsigned IdxIn128 = IdxVal - (IdxVal / NumEltsIn128) * NumEltsIn128;
11380 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, V.getValueType(), V, N1,
11381 DAG.getConstant(IdxIn128, dl, MVT::i32));
11383 // Insert the changed part back into the bigger vector
11384 return Insert128BitVector(N0, V, IdxVal, DAG, dl);
11386 assert(VT.is128BitVector() && "Only 128-bit vector types should be left!");
11388 if (Subtarget->hasSSE41()) {
11389 if (EltVT.getSizeInBits() == 8 || EltVT.getSizeInBits() == 16) {
11391 if (VT == MVT::v8i16) {
11392 Opc = X86ISD::PINSRW;
11394 assert(VT == MVT::v16i8);
11395 Opc = X86ISD::PINSRB;
11398 // Transform it so it match pinsr{b,w} which expects a GR32 as its second
11400 if (N1.getValueType() != MVT::i32)
11401 N1 = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, N1);
11402 if (N2.getValueType() != MVT::i32)
11403 N2 = DAG.getIntPtrConstant(IdxVal, dl);
11404 return DAG.getNode(Opc, dl, VT, N0, N1, N2);
11407 if (EltVT == MVT::f32) {
11408 // Bits [7:6] of the constant are the source select. This will always be
11409 // zero here. The DAG Combiner may combine an extract_elt index into
11410 // these bits. For example (insert (extract, 3), 2) could be matched by
11411 // putting the '3' into bits [7:6] of X86ISD::INSERTPS.
11412 // Bits [5:4] of the constant are the destination select. This is the
11413 // value of the incoming immediate.
11414 // Bits [3:0] of the constant are the zero mask. The DAG Combiner may
11415 // combine either bitwise AND or insert of float 0.0 to set these bits.
11417 bool MinSize = DAG.getMachineFunction().getFunction()->optForMinSize();
11418 if (IdxVal == 0 && (!MinSize || !MayFoldLoad(N1))) {
11419 // If this is an insertion of 32-bits into the low 32-bits of
11420 // a vector, we prefer to generate a blend with immediate rather
11421 // than an insertps. Blends are simpler operations in hardware and so
11422 // will always have equal or better performance than insertps.
11423 // But if optimizing for size and there's a load folding opportunity,
11424 // generate insertps because blendps does not have a 32-bit memory
11426 N2 = DAG.getIntPtrConstant(1, dl);
11427 N1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4f32, N1);
11428 return DAG.getNode(X86ISD::BLENDI, dl, VT, N0, N1, N2);
11430 N2 = DAG.getIntPtrConstant(IdxVal << 4, dl);
11431 // Create this as a scalar to vector..
11432 N1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4f32, N1);
11433 return DAG.getNode(X86ISD::INSERTPS, dl, VT, N0, N1, N2);
11436 if (EltVT == MVT::i32 || EltVT == MVT::i64) {
11437 // PINSR* works with constant index.
11442 if (EltVT == MVT::i8)
11445 if (EltVT.getSizeInBits() == 16) {
11446 // Transform it so it match pinsrw which expects a 16-bit value in a GR32
11447 // as its second argument.
11448 if (N1.getValueType() != MVT::i32)
11449 N1 = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, N1);
11450 if (N2.getValueType() != MVT::i32)
11451 N2 = DAG.getIntPtrConstant(IdxVal, dl);
11452 return DAG.getNode(X86ISD::PINSRW, dl, VT, N0, N1, N2);
11457 static SDValue LowerSCALAR_TO_VECTOR(SDValue Op, SelectionDAG &DAG) {
11459 MVT OpVT = Op.getSimpleValueType();
11461 // If this is a 256-bit vector result, first insert into a 128-bit
11462 // vector and then insert into the 256-bit vector.
11463 if (!OpVT.is128BitVector()) {
11464 // Insert into a 128-bit vector.
11465 unsigned SizeFactor = OpVT.getSizeInBits()/128;
11466 MVT VT128 = MVT::getVectorVT(OpVT.getVectorElementType(),
11467 OpVT.getVectorNumElements() / SizeFactor);
11469 Op = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT128, Op.getOperand(0));
11471 // Insert the 128-bit vector.
11472 return Insert128BitVector(DAG.getUNDEF(OpVT), Op, 0, DAG, dl);
11475 if (OpVT == MVT::v1i64 &&
11476 Op.getOperand(0).getValueType() == MVT::i64)
11477 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v1i64, Op.getOperand(0));
11479 SDValue AnyExt = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, Op.getOperand(0));
11480 assert(OpVT.is128BitVector() && "Expected an SSE type!");
11481 return DAG.getBitcast(
11482 OpVT, DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32, AnyExt));
11485 // Lower a node with an EXTRACT_SUBVECTOR opcode. This may result in
11486 // a simple subregister reference or explicit instructions to grab
11487 // upper bits of a vector.
11488 static SDValue LowerEXTRACT_SUBVECTOR(SDValue Op, const X86Subtarget *Subtarget,
11489 SelectionDAG &DAG) {
11491 SDValue In = Op.getOperand(0);
11492 SDValue Idx = Op.getOperand(1);
11493 unsigned IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
11494 MVT ResVT = Op.getSimpleValueType();
11495 MVT InVT = In.getSimpleValueType();
11497 if (Subtarget->hasFp256()) {
11498 if (ResVT.is128BitVector() &&
11499 (InVT.is256BitVector() || InVT.is512BitVector()) &&
11500 isa<ConstantSDNode>(Idx)) {
11501 return Extract128BitVector(In, IdxVal, DAG, dl);
11503 if (ResVT.is256BitVector() && InVT.is512BitVector() &&
11504 isa<ConstantSDNode>(Idx)) {
11505 return Extract256BitVector(In, IdxVal, DAG, dl);
11511 // Lower a node with an INSERT_SUBVECTOR opcode. This may result in a
11512 // simple superregister reference or explicit instructions to insert
11513 // the upper bits of a vector.
11514 static SDValue LowerINSERT_SUBVECTOR(SDValue Op, const X86Subtarget *Subtarget,
11515 SelectionDAG &DAG) {
11516 if (!Subtarget->hasAVX())
11520 SDValue Vec = Op.getOperand(0);
11521 SDValue SubVec = Op.getOperand(1);
11522 SDValue Idx = Op.getOperand(2);
11524 if (!isa<ConstantSDNode>(Idx))
11527 unsigned IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
11528 MVT OpVT = Op.getSimpleValueType();
11529 MVT SubVecVT = SubVec.getSimpleValueType();
11531 // Fold two 16-byte subvector loads into one 32-byte load:
11532 // (insert_subvector (insert_subvector undef, (load addr), 0),
11533 // (load addr + 16), Elts/2)
11535 if ((IdxVal == OpVT.getVectorNumElements() / 2) &&
11536 Vec.getOpcode() == ISD::INSERT_SUBVECTOR &&
11537 OpVT.is256BitVector() && SubVecVT.is128BitVector()) {
11538 auto *Idx2 = dyn_cast<ConstantSDNode>(Vec.getOperand(2));
11539 if (Idx2 && Idx2->getZExtValue() == 0) {
11540 SDValue SubVec2 = Vec.getOperand(1);
11541 // If needed, look through a bitcast to get to the load.
11542 if (SubVec2.getNode() && SubVec2.getOpcode() == ISD::BITCAST)
11543 SubVec2 = SubVec2.getOperand(0);
11545 if (auto *FirstLd = dyn_cast<LoadSDNode>(SubVec2)) {
11547 unsigned Alignment = FirstLd->getAlignment();
11548 unsigned AS = FirstLd->getAddressSpace();
11549 const X86TargetLowering *TLI = Subtarget->getTargetLowering();
11550 if (TLI->allowsMemoryAccess(*DAG.getContext(), DAG.getDataLayout(),
11551 OpVT, AS, Alignment, &Fast) && Fast) {
11552 SDValue Ops[] = { SubVec2, SubVec };
11553 if (SDValue Ld = EltsFromConsecutiveLoads(OpVT, Ops, dl, DAG, false))
11560 if ((OpVT.is256BitVector() || OpVT.is512BitVector()) &&
11561 SubVecVT.is128BitVector())
11562 return Insert128BitVector(Vec, SubVec, IdxVal, DAG, dl);
11564 if (OpVT.is512BitVector() && SubVecVT.is256BitVector())
11565 return Insert256BitVector(Vec, SubVec, IdxVal, DAG, dl);
11567 if (OpVT.getVectorElementType() == MVT::i1) {
11568 if (IdxVal == 0 && Vec.getOpcode() == ISD::UNDEF) // the operation is legal
11570 SDValue ZeroIdx = DAG.getIntPtrConstant(0, dl);
11571 SDValue Undef = DAG.getUNDEF(OpVT);
11572 unsigned NumElems = OpVT.getVectorNumElements();
11573 SDValue ShiftBits = DAG.getConstant(NumElems/2, dl, MVT::i8);
11575 if (IdxVal == OpVT.getVectorNumElements() / 2) {
11576 // Zero upper bits of the Vec
11577 Vec = DAG.getNode(X86ISD::VSHLI, dl, OpVT, Vec, ShiftBits);
11578 Vec = DAG.getNode(X86ISD::VSRLI, dl, OpVT, Vec, ShiftBits);
11580 SDValue Vec2 = DAG.getNode(ISD::INSERT_SUBVECTOR, dl, OpVT, Undef,
11582 Vec2 = DAG.getNode(X86ISD::VSHLI, dl, OpVT, Vec2, ShiftBits);
11583 return DAG.getNode(ISD::OR, dl, OpVT, Vec, Vec2);
11586 SDValue Vec2 = DAG.getNode(ISD::INSERT_SUBVECTOR, dl, OpVT, Undef,
11588 // Zero upper bits of the Vec2
11589 Vec2 = DAG.getNode(X86ISD::VSHLI, dl, OpVT, Vec2, ShiftBits);
11590 Vec2 = DAG.getNode(X86ISD::VSRLI, dl, OpVT, Vec2, ShiftBits);
11591 // Zero lower bits of the Vec
11592 Vec = DAG.getNode(X86ISD::VSRLI, dl, OpVT, Vec, ShiftBits);
11593 Vec = DAG.getNode(X86ISD::VSHLI, dl, OpVT, Vec, ShiftBits);
11594 // Merge them together
11595 return DAG.getNode(ISD::OR, dl, OpVT, Vec, Vec2);
11601 // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
11602 // their target countpart wrapped in the X86ISD::Wrapper node. Suppose N is
11603 // one of the above mentioned nodes. It has to be wrapped because otherwise
11604 // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
11605 // be used to form addressing mode. These wrapped nodes will be selected
11608 X86TargetLowering::LowerConstantPool(SDValue Op, SelectionDAG &DAG) const {
11609 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
11611 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
11612 // global base reg.
11613 unsigned char OpFlag = 0;
11614 unsigned WrapperKind = X86ISD::Wrapper;
11615 CodeModel::Model M = DAG.getTarget().getCodeModel();
11617 if (Subtarget->isPICStyleRIPRel() &&
11618 (M == CodeModel::Small || M == CodeModel::Kernel))
11619 WrapperKind = X86ISD::WrapperRIP;
11620 else if (Subtarget->isPICStyleGOT())
11621 OpFlag = X86II::MO_GOTOFF;
11622 else if (Subtarget->isPICStyleStubPIC())
11623 OpFlag = X86II::MO_PIC_BASE_OFFSET;
11625 auto PtrVT = getPointerTy(DAG.getDataLayout());
11626 SDValue Result = DAG.getTargetConstantPool(
11627 CP->getConstVal(), PtrVT, CP->getAlignment(), CP->getOffset(), OpFlag);
11629 Result = DAG.getNode(WrapperKind, DL, PtrVT, Result);
11630 // With PIC, the address is actually $g + Offset.
11633 DAG.getNode(ISD::ADD, DL, PtrVT,
11634 DAG.getNode(X86ISD::GlobalBaseReg, SDLoc(), PtrVT), Result);
11640 SDValue X86TargetLowering::LowerJumpTable(SDValue Op, SelectionDAG &DAG) const {
11641 JumpTableSDNode *JT = cast<JumpTableSDNode>(Op);
11643 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
11644 // global base reg.
11645 unsigned char OpFlag = 0;
11646 unsigned WrapperKind = X86ISD::Wrapper;
11647 CodeModel::Model M = DAG.getTarget().getCodeModel();
11649 if (Subtarget->isPICStyleRIPRel() &&
11650 (M == CodeModel::Small || M == CodeModel::Kernel))
11651 WrapperKind = X86ISD::WrapperRIP;
11652 else if (Subtarget->isPICStyleGOT())
11653 OpFlag = X86II::MO_GOTOFF;
11654 else if (Subtarget->isPICStyleStubPIC())
11655 OpFlag = X86II::MO_PIC_BASE_OFFSET;
11657 auto PtrVT = getPointerTy(DAG.getDataLayout());
11658 SDValue Result = DAG.getTargetJumpTable(JT->getIndex(), PtrVT, OpFlag);
11660 Result = DAG.getNode(WrapperKind, DL, PtrVT, Result);
11662 // With PIC, the address is actually $g + Offset.
11665 DAG.getNode(ISD::ADD, DL, PtrVT,
11666 DAG.getNode(X86ISD::GlobalBaseReg, SDLoc(), PtrVT), Result);
11672 X86TargetLowering::LowerExternalSymbol(SDValue Op, SelectionDAG &DAG) const {
11673 const char *Sym = cast<ExternalSymbolSDNode>(Op)->getSymbol();
11675 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
11676 // global base reg.
11677 unsigned char OpFlag = 0;
11678 unsigned WrapperKind = X86ISD::Wrapper;
11679 CodeModel::Model M = DAG.getTarget().getCodeModel();
11681 if (Subtarget->isPICStyleRIPRel() &&
11682 (M == CodeModel::Small || M == CodeModel::Kernel)) {
11683 if (Subtarget->isTargetDarwin() || Subtarget->isTargetELF())
11684 OpFlag = X86II::MO_GOTPCREL;
11685 WrapperKind = X86ISD::WrapperRIP;
11686 } else if (Subtarget->isPICStyleGOT()) {
11687 OpFlag = X86II::MO_GOT;
11688 } else if (Subtarget->isPICStyleStubPIC()) {
11689 OpFlag = X86II::MO_DARWIN_NONLAZY_PIC_BASE;
11690 } else if (Subtarget->isPICStyleStubNoDynamic()) {
11691 OpFlag = X86II::MO_DARWIN_NONLAZY;
11694 auto PtrVT = getPointerTy(DAG.getDataLayout());
11695 SDValue Result = DAG.getTargetExternalSymbol(Sym, PtrVT, OpFlag);
11698 Result = DAG.getNode(WrapperKind, DL, PtrVT, Result);
11700 // With PIC, the address is actually $g + Offset.
11701 if (DAG.getTarget().getRelocationModel() == Reloc::PIC_ &&
11702 !Subtarget->is64Bit()) {
11704 DAG.getNode(ISD::ADD, DL, PtrVT,
11705 DAG.getNode(X86ISD::GlobalBaseReg, SDLoc(), PtrVT), Result);
11708 // For symbols that require a load from a stub to get the address, emit the
11710 if (isGlobalStubReference(OpFlag))
11711 Result = DAG.getLoad(PtrVT, DL, DAG.getEntryNode(), Result,
11712 MachinePointerInfo::getGOT(DAG.getMachineFunction()),
11713 false, false, false, 0);
11719 X86TargetLowering::LowerBlockAddress(SDValue Op, SelectionDAG &DAG) const {
11720 // Create the TargetBlockAddressAddress node.
11721 unsigned char OpFlags =
11722 Subtarget->ClassifyBlockAddressReference();
11723 CodeModel::Model M = DAG.getTarget().getCodeModel();
11724 const BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress();
11725 int64_t Offset = cast<BlockAddressSDNode>(Op)->getOffset();
11727 auto PtrVT = getPointerTy(DAG.getDataLayout());
11728 SDValue Result = DAG.getTargetBlockAddress(BA, PtrVT, Offset, OpFlags);
11730 if (Subtarget->isPICStyleRIPRel() &&
11731 (M == CodeModel::Small || M == CodeModel::Kernel))
11732 Result = DAG.getNode(X86ISD::WrapperRIP, dl, PtrVT, Result);
11734 Result = DAG.getNode(X86ISD::Wrapper, dl, PtrVT, Result);
11736 // With PIC, the address is actually $g + Offset.
11737 if (isGlobalRelativeToPICBase(OpFlags)) {
11738 Result = DAG.getNode(ISD::ADD, dl, PtrVT,
11739 DAG.getNode(X86ISD::GlobalBaseReg, dl, PtrVT), Result);
11746 X86TargetLowering::LowerGlobalAddress(const GlobalValue *GV, SDLoc dl,
11747 int64_t Offset, SelectionDAG &DAG) const {
11748 // Create the TargetGlobalAddress node, folding in the constant
11749 // offset if it is legal.
11750 unsigned char OpFlags =
11751 Subtarget->ClassifyGlobalReference(GV, DAG.getTarget());
11752 CodeModel::Model M = DAG.getTarget().getCodeModel();
11753 auto PtrVT = getPointerTy(DAG.getDataLayout());
11755 if (OpFlags == X86II::MO_NO_FLAG &&
11756 X86::isOffsetSuitableForCodeModel(Offset, M)) {
11757 // A direct static reference to a global.
11758 Result = DAG.getTargetGlobalAddress(GV, dl, PtrVT, Offset);
11761 Result = DAG.getTargetGlobalAddress(GV, dl, PtrVT, 0, OpFlags);
11764 if (Subtarget->isPICStyleRIPRel() &&
11765 (M == CodeModel::Small || M == CodeModel::Kernel))
11766 Result = DAG.getNode(X86ISD::WrapperRIP, dl, PtrVT, Result);
11768 Result = DAG.getNode(X86ISD::Wrapper, dl, PtrVT, Result);
11770 // With PIC, the address is actually $g + Offset.
11771 if (isGlobalRelativeToPICBase(OpFlags)) {
11772 Result = DAG.getNode(ISD::ADD, dl, PtrVT,
11773 DAG.getNode(X86ISD::GlobalBaseReg, dl, PtrVT), Result);
11776 // For globals that require a load from a stub to get the address, emit the
11778 if (isGlobalStubReference(OpFlags))
11779 Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Result,
11780 MachinePointerInfo::getGOT(DAG.getMachineFunction()),
11781 false, false, false, 0);
11783 // If there was a non-zero offset that we didn't fold, create an explicit
11784 // addition for it.
11786 Result = DAG.getNode(ISD::ADD, dl, PtrVT, Result,
11787 DAG.getConstant(Offset, dl, PtrVT));
11793 X86TargetLowering::LowerGlobalAddress(SDValue Op, SelectionDAG &DAG) const {
11794 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
11795 int64_t Offset = cast<GlobalAddressSDNode>(Op)->getOffset();
11796 return LowerGlobalAddress(GV, SDLoc(Op), Offset, DAG);
11800 GetTLSADDR(SelectionDAG &DAG, SDValue Chain, GlobalAddressSDNode *GA,
11801 SDValue *InFlag, const EVT PtrVT, unsigned ReturnReg,
11802 unsigned char OperandFlags, bool LocalDynamic = false) {
11803 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
11804 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
11806 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(), dl,
11807 GA->getValueType(0),
11811 X86ISD::NodeType CallType = LocalDynamic ? X86ISD::TLSBASEADDR
11815 SDValue Ops[] = { Chain, TGA, *InFlag };
11816 Chain = DAG.getNode(CallType, dl, NodeTys, Ops);
11818 SDValue Ops[] = { Chain, TGA };
11819 Chain = DAG.getNode(CallType, dl, NodeTys, Ops);
11822 // TLSADDR will be codegen'ed as call. Inform MFI that function has calls.
11823 MFI->setAdjustsStack(true);
11824 MFI->setHasCalls(true);
11826 SDValue Flag = Chain.getValue(1);
11827 return DAG.getCopyFromReg(Chain, dl, ReturnReg, PtrVT, Flag);
11830 // Lower ISD::GlobalTLSAddress using the "general dynamic" model, 32 bit
11832 LowerToTLSGeneralDynamicModel32(GlobalAddressSDNode *GA, SelectionDAG &DAG,
11835 SDLoc dl(GA); // ? function entry point might be better
11836 SDValue Chain = DAG.getCopyToReg(DAG.getEntryNode(), dl, X86::EBX,
11837 DAG.getNode(X86ISD::GlobalBaseReg,
11838 SDLoc(), PtrVT), InFlag);
11839 InFlag = Chain.getValue(1);
11841 return GetTLSADDR(DAG, Chain, GA, &InFlag, PtrVT, X86::EAX, X86II::MO_TLSGD);
11844 // Lower ISD::GlobalTLSAddress using the "general dynamic" model, 64 bit
11846 LowerToTLSGeneralDynamicModel64(GlobalAddressSDNode *GA, SelectionDAG &DAG,
11848 return GetTLSADDR(DAG, DAG.getEntryNode(), GA, nullptr, PtrVT,
11849 X86::RAX, X86II::MO_TLSGD);
11852 static SDValue LowerToTLSLocalDynamicModel(GlobalAddressSDNode *GA,
11858 // Get the start address of the TLS block for this module.
11859 X86MachineFunctionInfo* MFI = DAG.getMachineFunction()
11860 .getInfo<X86MachineFunctionInfo>();
11861 MFI->incNumLocalDynamicTLSAccesses();
11865 Base = GetTLSADDR(DAG, DAG.getEntryNode(), GA, nullptr, PtrVT, X86::RAX,
11866 X86II::MO_TLSLD, /*LocalDynamic=*/true);
11869 SDValue Chain = DAG.getCopyToReg(DAG.getEntryNode(), dl, X86::EBX,
11870 DAG.getNode(X86ISD::GlobalBaseReg, SDLoc(), PtrVT), InFlag);
11871 InFlag = Chain.getValue(1);
11872 Base = GetTLSADDR(DAG, Chain, GA, &InFlag, PtrVT, X86::EAX,
11873 X86II::MO_TLSLDM, /*LocalDynamic=*/true);
11876 // Note: the CleanupLocalDynamicTLSPass will remove redundant computations
11880 unsigned char OperandFlags = X86II::MO_DTPOFF;
11881 unsigned WrapperKind = X86ISD::Wrapper;
11882 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(), dl,
11883 GA->getValueType(0),
11884 GA->getOffset(), OperandFlags);
11885 SDValue Offset = DAG.getNode(WrapperKind, dl, PtrVT, TGA);
11887 // Add x@dtpoff with the base.
11888 return DAG.getNode(ISD::ADD, dl, PtrVT, Offset, Base);
11891 // Lower ISD::GlobalTLSAddress using the "initial exec" or "local exec" model.
11892 static SDValue LowerToTLSExecModel(GlobalAddressSDNode *GA, SelectionDAG &DAG,
11893 const EVT PtrVT, TLSModel::Model model,
11894 bool is64Bit, bool isPIC) {
11897 // Get the Thread Pointer, which is %gs:0 (32-bit) or %fs:0 (64-bit).
11898 Value *Ptr = Constant::getNullValue(Type::getInt8PtrTy(*DAG.getContext(),
11899 is64Bit ? 257 : 256));
11901 SDValue ThreadPointer =
11902 DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), DAG.getIntPtrConstant(0, dl),
11903 MachinePointerInfo(Ptr), false, false, false, 0);
11905 unsigned char OperandFlags = 0;
11906 // Most TLS accesses are not RIP relative, even on x86-64. One exception is
11908 unsigned WrapperKind = X86ISD::Wrapper;
11909 if (model == TLSModel::LocalExec) {
11910 OperandFlags = is64Bit ? X86II::MO_TPOFF : X86II::MO_NTPOFF;
11911 } else if (model == TLSModel::InitialExec) {
11913 OperandFlags = X86II::MO_GOTTPOFF;
11914 WrapperKind = X86ISD::WrapperRIP;
11916 OperandFlags = isPIC ? X86II::MO_GOTNTPOFF : X86II::MO_INDNTPOFF;
11919 llvm_unreachable("Unexpected model");
11922 // emit "addl x@ntpoff,%eax" (local exec)
11923 // or "addl x@indntpoff,%eax" (initial exec)
11924 // or "addl x@gotntpoff(%ebx) ,%eax" (initial exec, 32-bit pic)
11926 DAG.getTargetGlobalAddress(GA->getGlobal(), dl, GA->getValueType(0),
11927 GA->getOffset(), OperandFlags);
11928 SDValue Offset = DAG.getNode(WrapperKind, dl, PtrVT, TGA);
11930 if (model == TLSModel::InitialExec) {
11931 if (isPIC && !is64Bit) {
11932 Offset = DAG.getNode(ISD::ADD, dl, PtrVT,
11933 DAG.getNode(X86ISD::GlobalBaseReg, SDLoc(), PtrVT),
11937 Offset = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Offset,
11938 MachinePointerInfo::getGOT(DAG.getMachineFunction()),
11939 false, false, false, 0);
11942 // The address of the thread local variable is the add of the thread
11943 // pointer with the offset of the variable.
11944 return DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, Offset);
11948 X86TargetLowering::LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const {
11950 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
11951 const GlobalValue *GV = GA->getGlobal();
11952 auto PtrVT = getPointerTy(DAG.getDataLayout());
11954 if (Subtarget->isTargetELF()) {
11955 if (DAG.getTarget().Options.EmulatedTLS)
11956 return LowerToTLSEmulatedModel(GA, DAG);
11957 TLSModel::Model model = DAG.getTarget().getTLSModel(GV);
11959 case TLSModel::GeneralDynamic:
11960 if (Subtarget->is64Bit())
11961 return LowerToTLSGeneralDynamicModel64(GA, DAG, PtrVT);
11962 return LowerToTLSGeneralDynamicModel32(GA, DAG, PtrVT);
11963 case TLSModel::LocalDynamic:
11964 return LowerToTLSLocalDynamicModel(GA, DAG, PtrVT,
11965 Subtarget->is64Bit());
11966 case TLSModel::InitialExec:
11967 case TLSModel::LocalExec:
11968 return LowerToTLSExecModel(GA, DAG, PtrVT, model, Subtarget->is64Bit(),
11969 DAG.getTarget().getRelocationModel() ==
11972 llvm_unreachable("Unknown TLS model.");
11975 if (Subtarget->isTargetDarwin()) {
11976 // Darwin only has one model of TLS. Lower to that.
11977 unsigned char OpFlag = 0;
11978 unsigned WrapperKind = Subtarget->isPICStyleRIPRel() ?
11979 X86ISD::WrapperRIP : X86ISD::Wrapper;
11981 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
11982 // global base reg.
11983 bool PIC32 = (DAG.getTarget().getRelocationModel() == Reloc::PIC_) &&
11984 !Subtarget->is64Bit();
11986 OpFlag = X86II::MO_TLVP_PIC_BASE;
11988 OpFlag = X86II::MO_TLVP;
11990 SDValue Result = DAG.getTargetGlobalAddress(GA->getGlobal(), DL,
11991 GA->getValueType(0),
11992 GA->getOffset(), OpFlag);
11993 SDValue Offset = DAG.getNode(WrapperKind, DL, PtrVT, Result);
11995 // With PIC32, the address is actually $g + Offset.
11997 Offset = DAG.getNode(ISD::ADD, DL, PtrVT,
11998 DAG.getNode(X86ISD::GlobalBaseReg, SDLoc(), PtrVT),
12001 // Lowering the machine isd will make sure everything is in the right
12003 SDValue Chain = DAG.getEntryNode();
12004 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
12005 SDValue Args[] = { Chain, Offset };
12006 Chain = DAG.getNode(X86ISD::TLSCALL, DL, NodeTys, Args);
12008 // TLSCALL will be codegen'ed as call. Inform MFI that function has calls.
12009 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
12010 MFI->setAdjustsStack(true);
12012 // And our return value (tls address) is in the standard call return value
12014 unsigned Reg = Subtarget->is64Bit() ? X86::RAX : X86::EAX;
12015 return DAG.getCopyFromReg(Chain, DL, Reg, PtrVT, Chain.getValue(1));
12018 if (Subtarget->isTargetKnownWindowsMSVC() ||
12019 Subtarget->isTargetWindowsGNU()) {
12020 // Just use the implicit TLS architecture
12021 // Need to generate someting similar to:
12022 // mov rdx, qword [gs:abs 58H]; Load pointer to ThreadLocalStorage
12024 // mov ecx, dword [rel _tls_index]: Load index (from C runtime)
12025 // mov rcx, qword [rdx+rcx*8]
12026 // mov eax, .tls$:tlsvar
12027 // [rax+rcx] contains the address
12028 // Windows 64bit: gs:0x58
12029 // Windows 32bit: fs:__tls_array
12032 SDValue Chain = DAG.getEntryNode();
12034 // Get the Thread Pointer, which is %fs:__tls_array (32-bit) or
12035 // %gs:0x58 (64-bit). On MinGW, __tls_array is not available, so directly
12036 // use its literal value of 0x2C.
12037 Value *Ptr = Constant::getNullValue(Subtarget->is64Bit()
12038 ? Type::getInt8PtrTy(*DAG.getContext(),
12040 : Type::getInt32PtrTy(*DAG.getContext(),
12043 SDValue TlsArray = Subtarget->is64Bit()
12044 ? DAG.getIntPtrConstant(0x58, dl)
12045 : (Subtarget->isTargetWindowsGNU()
12046 ? DAG.getIntPtrConstant(0x2C, dl)
12047 : DAG.getExternalSymbol("_tls_array", PtrVT));
12049 SDValue ThreadPointer =
12050 DAG.getLoad(PtrVT, dl, Chain, TlsArray, MachinePointerInfo(Ptr), false,
12054 if (GV->getThreadLocalMode() == GlobalVariable::LocalExecTLSModel) {
12055 res = ThreadPointer;
12057 // Load the _tls_index variable
12058 SDValue IDX = DAG.getExternalSymbol("_tls_index", PtrVT);
12059 if (Subtarget->is64Bit())
12060 IDX = DAG.getExtLoad(ISD::ZEXTLOAD, dl, PtrVT, Chain, IDX,
12061 MachinePointerInfo(), MVT::i32, false, false,
12064 IDX = DAG.getLoad(PtrVT, dl, Chain, IDX, MachinePointerInfo(), false,
12067 auto &DL = DAG.getDataLayout();
12069 DAG.getConstant(Log2_64_Ceil(DL.getPointerSize()), dl, PtrVT);
12070 IDX = DAG.getNode(ISD::SHL, dl, PtrVT, IDX, Scale);
12072 res = DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, IDX);
12075 res = DAG.getLoad(PtrVT, dl, Chain, res, MachinePointerInfo(), false, false,
12078 // Get the offset of start of .tls section
12079 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(), dl,
12080 GA->getValueType(0),
12081 GA->getOffset(), X86II::MO_SECREL);
12082 SDValue Offset = DAG.getNode(X86ISD::Wrapper, dl, PtrVT, TGA);
12084 // The address of the thread local variable is the add of the thread
12085 // pointer with the offset of the variable.
12086 return DAG.getNode(ISD::ADD, dl, PtrVT, res, Offset);
12089 llvm_unreachable("TLS not implemented for this target.");
12092 /// LowerShiftParts - Lower SRA_PARTS and friends, which return two i32 values
12093 /// and take a 2 x i32 value to shift plus a shift amount.
12094 static SDValue LowerShiftParts(SDValue Op, SelectionDAG &DAG) {
12095 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
12096 MVT VT = Op.getSimpleValueType();
12097 unsigned VTBits = VT.getSizeInBits();
12099 bool isSRA = Op.getOpcode() == ISD::SRA_PARTS;
12100 SDValue ShOpLo = Op.getOperand(0);
12101 SDValue ShOpHi = Op.getOperand(1);
12102 SDValue ShAmt = Op.getOperand(2);
12103 // X86ISD::SHLD and X86ISD::SHRD have defined overflow behavior but the
12104 // generic ISD nodes haven't. Insert an AND to be safe, it's optimized away
12106 SDValue SafeShAmt = DAG.getNode(ISD::AND, dl, MVT::i8, ShAmt,
12107 DAG.getConstant(VTBits - 1, dl, MVT::i8));
12108 SDValue Tmp1 = isSRA ? DAG.getNode(ISD::SRA, dl, VT, ShOpHi,
12109 DAG.getConstant(VTBits - 1, dl, MVT::i8))
12110 : DAG.getConstant(0, dl, VT);
12112 SDValue Tmp2, Tmp3;
12113 if (Op.getOpcode() == ISD::SHL_PARTS) {
12114 Tmp2 = DAG.getNode(X86ISD::SHLD, dl, VT, ShOpHi, ShOpLo, ShAmt);
12115 Tmp3 = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, SafeShAmt);
12117 Tmp2 = DAG.getNode(X86ISD::SHRD, dl, VT, ShOpLo, ShOpHi, ShAmt);
12118 Tmp3 = DAG.getNode(isSRA ? ISD::SRA : ISD::SRL, dl, VT, ShOpHi, SafeShAmt);
12121 // If the shift amount is larger or equal than the width of a part we can't
12122 // rely on the results of shld/shrd. Insert a test and select the appropriate
12123 // values for large shift amounts.
12124 SDValue AndNode = DAG.getNode(ISD::AND, dl, MVT::i8, ShAmt,
12125 DAG.getConstant(VTBits, dl, MVT::i8));
12126 SDValue Cond = DAG.getNode(X86ISD::CMP, dl, MVT::i32,
12127 AndNode, DAG.getConstant(0, dl, MVT::i8));
12130 SDValue CC = DAG.getConstant(X86::COND_NE, dl, MVT::i8);
12131 SDValue Ops0[4] = { Tmp2, Tmp3, CC, Cond };
12132 SDValue Ops1[4] = { Tmp3, Tmp1, CC, Cond };
12134 if (Op.getOpcode() == ISD::SHL_PARTS) {
12135 Hi = DAG.getNode(X86ISD::CMOV, dl, VT, Ops0);
12136 Lo = DAG.getNode(X86ISD::CMOV, dl, VT, Ops1);
12138 Lo = DAG.getNode(X86ISD::CMOV, dl, VT, Ops0);
12139 Hi = DAG.getNode(X86ISD::CMOV, dl, VT, Ops1);
12142 SDValue Ops[2] = { Lo, Hi };
12143 return DAG.getMergeValues(Ops, dl);
12146 SDValue X86TargetLowering::LowerSINT_TO_FP(SDValue Op,
12147 SelectionDAG &DAG) const {
12148 SDValue Src = Op.getOperand(0);
12149 MVT SrcVT = Src.getSimpleValueType();
12150 MVT VT = Op.getSimpleValueType();
12153 if (SrcVT.isVector()) {
12154 if (SrcVT == MVT::v2i32 && VT == MVT::v2f64) {
12155 return DAG.getNode(X86ISD::CVTDQ2PD, dl, VT,
12156 DAG.getNode(ISD::CONCAT_VECTORS, dl, MVT::v4i32, Src,
12157 DAG.getUNDEF(SrcVT)));
12159 if (SrcVT.getVectorElementType() == MVT::i1) {
12160 MVT IntegerVT = MVT::getVectorVT(MVT::i32, SrcVT.getVectorNumElements());
12161 return DAG.getNode(ISD::SINT_TO_FP, dl, Op.getValueType(),
12162 DAG.getNode(ISD::SIGN_EXTEND, dl, IntegerVT, Src));
12167 assert(SrcVT <= MVT::i64 && SrcVT >= MVT::i16 &&
12168 "Unknown SINT_TO_FP to lower!");
12170 // These are really Legal; return the operand so the caller accepts it as
12172 if (SrcVT == MVT::i32 && isScalarFPTypeInSSEReg(Op.getValueType()))
12174 if (SrcVT == MVT::i64 && isScalarFPTypeInSSEReg(Op.getValueType()) &&
12175 Subtarget->is64Bit()) {
12179 unsigned Size = SrcVT.getSizeInBits()/8;
12180 MachineFunction &MF = DAG.getMachineFunction();
12181 auto PtrVT = getPointerTy(MF.getDataLayout());
12182 int SSFI = MF.getFrameInfo()->CreateStackObject(Size, Size, false);
12183 SDValue StackSlot = DAG.getFrameIndex(SSFI, PtrVT);
12184 SDValue Chain = DAG.getStore(
12185 DAG.getEntryNode(), dl, Op.getOperand(0), StackSlot,
12186 MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), SSFI), false,
12188 return BuildFILD(Op, SrcVT, Chain, StackSlot, DAG);
12191 SDValue X86TargetLowering::BuildFILD(SDValue Op, EVT SrcVT, SDValue Chain,
12193 SelectionDAG &DAG) const {
12197 bool useSSE = isScalarFPTypeInSSEReg(Op.getValueType());
12199 Tys = DAG.getVTList(MVT::f64, MVT::Other, MVT::Glue);
12201 Tys = DAG.getVTList(Op.getValueType(), MVT::Other);
12203 unsigned ByteSize = SrcVT.getSizeInBits()/8;
12205 FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(StackSlot);
12206 MachineMemOperand *MMO;
12208 int SSFI = FI->getIndex();
12209 MMO = DAG.getMachineFunction().getMachineMemOperand(
12210 MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), SSFI),
12211 MachineMemOperand::MOLoad, ByteSize, ByteSize);
12213 MMO = cast<LoadSDNode>(StackSlot)->getMemOperand();
12214 StackSlot = StackSlot.getOperand(1);
12216 SDValue Ops[] = { Chain, StackSlot, DAG.getValueType(SrcVT) };
12217 SDValue Result = DAG.getMemIntrinsicNode(useSSE ? X86ISD::FILD_FLAG :
12219 Tys, Ops, SrcVT, MMO);
12222 Chain = Result.getValue(1);
12223 SDValue InFlag = Result.getValue(2);
12225 // FIXME: Currently the FST is flagged to the FILD_FLAG. This
12226 // shouldn't be necessary except that RFP cannot be live across
12227 // multiple blocks. When stackifier is fixed, they can be uncoupled.
12228 MachineFunction &MF = DAG.getMachineFunction();
12229 unsigned SSFISize = Op.getValueType().getSizeInBits()/8;
12230 int SSFI = MF.getFrameInfo()->CreateStackObject(SSFISize, SSFISize, false);
12231 auto PtrVT = getPointerTy(MF.getDataLayout());
12232 SDValue StackSlot = DAG.getFrameIndex(SSFI, PtrVT);
12233 Tys = DAG.getVTList(MVT::Other);
12235 Chain, Result, StackSlot, DAG.getValueType(Op.getValueType()), InFlag
12237 MachineMemOperand *MMO = DAG.getMachineFunction().getMachineMemOperand(
12238 MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), SSFI),
12239 MachineMemOperand::MOStore, SSFISize, SSFISize);
12241 Chain = DAG.getMemIntrinsicNode(X86ISD::FST, DL, Tys,
12242 Ops, Op.getValueType(), MMO);
12243 Result = DAG.getLoad(
12244 Op.getValueType(), DL, Chain, StackSlot,
12245 MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), SSFI),
12246 false, false, false, 0);
12252 // LowerUINT_TO_FP_i64 - 64-bit unsigned integer to double expansion.
12253 SDValue X86TargetLowering::LowerUINT_TO_FP_i64(SDValue Op,
12254 SelectionDAG &DAG) const {
12255 // This algorithm is not obvious. Here it is what we're trying to output:
12258 punpckldq (c0), %xmm0 // c0: (uint4){ 0x43300000U, 0x45300000U, 0U, 0U }
12259 subpd (c1), %xmm0 // c1: (double2){ 0x1.0p52, 0x1.0p52 * 0x1.0p32 }
12261 haddpd %xmm0, %xmm0
12263 pshufd $0x4e, %xmm0, %xmm1
12269 LLVMContext *Context = DAG.getContext();
12271 // Build some magic constants.
12272 static const uint32_t CV0[] = { 0x43300000, 0x45300000, 0, 0 };
12273 Constant *C0 = ConstantDataVector::get(*Context, CV0);
12274 auto PtrVT = getPointerTy(DAG.getDataLayout());
12275 SDValue CPIdx0 = DAG.getConstantPool(C0, PtrVT, 16);
12277 SmallVector<Constant*,2> CV1;
12279 ConstantFP::get(*Context, APFloat(APFloat::IEEEdouble,
12280 APInt(64, 0x4330000000000000ULL))));
12282 ConstantFP::get(*Context, APFloat(APFloat::IEEEdouble,
12283 APInt(64, 0x4530000000000000ULL))));
12284 Constant *C1 = ConstantVector::get(CV1);
12285 SDValue CPIdx1 = DAG.getConstantPool(C1, PtrVT, 16);
12287 // Load the 64-bit value into an XMM register.
12288 SDValue XR1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i64,
12291 DAG.getLoad(MVT::v4i32, dl, DAG.getEntryNode(), CPIdx0,
12292 MachinePointerInfo::getConstantPool(DAG.getMachineFunction()),
12293 false, false, false, 16);
12295 getUnpackl(DAG, dl, MVT::v4i32, DAG.getBitcast(MVT::v4i32, XR1), CLod0);
12298 DAG.getLoad(MVT::v2f64, dl, CLod0.getValue(1), CPIdx1,
12299 MachinePointerInfo::getConstantPool(DAG.getMachineFunction()),
12300 false, false, false, 16);
12301 SDValue XR2F = DAG.getBitcast(MVT::v2f64, Unpck1);
12302 // TODO: Are there any fast-math-flags to propagate here?
12303 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::v2f64, XR2F, CLod1);
12306 if (Subtarget->hasSSE3()) {
12307 // FIXME: The 'haddpd' instruction may be slower than 'movhlps + addsd'.
12308 Result = DAG.getNode(X86ISD::FHADD, dl, MVT::v2f64, Sub, Sub);
12310 SDValue S2F = DAG.getBitcast(MVT::v4i32, Sub);
12311 SDValue Shuffle = getTargetShuffleNode(X86ISD::PSHUFD, dl, MVT::v4i32,
12313 Result = DAG.getNode(ISD::FADD, dl, MVT::v2f64,
12314 DAG.getBitcast(MVT::v2f64, Shuffle), Sub);
12317 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Result,
12318 DAG.getIntPtrConstant(0, dl));
12321 // LowerUINT_TO_FP_i32 - 32-bit unsigned integer to float expansion.
12322 SDValue X86TargetLowering::LowerUINT_TO_FP_i32(SDValue Op,
12323 SelectionDAG &DAG) const {
12325 // FP constant to bias correct the final result.
12326 SDValue Bias = DAG.getConstantFP(BitsToDouble(0x4330000000000000ULL), dl,
12329 // Load the 32-bit value into an XMM register.
12330 SDValue Load = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
12333 // Zero out the upper parts of the register.
12334 Load = getShuffleVectorZeroOrUndef(Load, 0, true, Subtarget, DAG);
12336 Load = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
12337 DAG.getBitcast(MVT::v2f64, Load),
12338 DAG.getIntPtrConstant(0, dl));
12340 // Or the load with the bias.
12341 SDValue Or = DAG.getNode(
12342 ISD::OR, dl, MVT::v2i64,
12343 DAG.getBitcast(MVT::v2i64,
12344 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2f64, Load)),
12345 DAG.getBitcast(MVT::v2i64,
12346 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2f64, Bias)));
12348 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
12349 DAG.getBitcast(MVT::v2f64, Or), DAG.getIntPtrConstant(0, dl));
12351 // Subtract the bias.
12352 // TODO: Are there any fast-math-flags to propagate here?
12353 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::f64, Or, Bias);
12355 // Handle final rounding.
12356 EVT DestVT = Op.getValueType();
12358 if (DestVT.bitsLT(MVT::f64))
12359 return DAG.getNode(ISD::FP_ROUND, dl, DestVT, Sub,
12360 DAG.getIntPtrConstant(0, dl));
12361 if (DestVT.bitsGT(MVT::f64))
12362 return DAG.getNode(ISD::FP_EXTEND, dl, DestVT, Sub);
12364 // Handle final rounding.
12368 static SDValue lowerUINT_TO_FP_vXi32(SDValue Op, SelectionDAG &DAG,
12369 const X86Subtarget &Subtarget) {
12370 // The algorithm is the following:
12371 // #ifdef __SSE4_1__
12372 // uint4 lo = _mm_blend_epi16( v, (uint4) 0x4b000000, 0xaa);
12373 // uint4 hi = _mm_blend_epi16( _mm_srli_epi32(v,16),
12374 // (uint4) 0x53000000, 0xaa);
12376 // uint4 lo = (v & (uint4) 0xffff) | (uint4) 0x4b000000;
12377 // uint4 hi = (v >> 16) | (uint4) 0x53000000;
12379 // float4 fhi = (float4) hi - (0x1.0p39f + 0x1.0p23f);
12380 // return (float4) lo + fhi;
12383 SDValue V = Op->getOperand(0);
12384 EVT VecIntVT = V.getValueType();
12385 bool Is128 = VecIntVT == MVT::v4i32;
12386 EVT VecFloatVT = Is128 ? MVT::v4f32 : MVT::v8f32;
12387 // If we convert to something else than the supported type, e.g., to v4f64,
12389 if (VecFloatVT != Op->getValueType(0))
12392 unsigned NumElts = VecIntVT.getVectorNumElements();
12393 assert((VecIntVT == MVT::v4i32 || VecIntVT == MVT::v8i32) &&
12394 "Unsupported custom type");
12395 assert(NumElts <= 8 && "The size of the constant array must be fixed");
12397 // In the #idef/#else code, we have in common:
12398 // - The vector of constants:
12404 // Create the splat vector for 0x4b000000.
12405 SDValue CstLow = DAG.getConstant(0x4b000000, DL, MVT::i32);
12406 SDValue CstLowArray[] = {CstLow, CstLow, CstLow, CstLow,
12407 CstLow, CstLow, CstLow, CstLow};
12408 SDValue VecCstLow = DAG.getNode(ISD::BUILD_VECTOR, DL, VecIntVT,
12409 makeArrayRef(&CstLowArray[0], NumElts));
12410 // Create the splat vector for 0x53000000.
12411 SDValue CstHigh = DAG.getConstant(0x53000000, DL, MVT::i32);
12412 SDValue CstHighArray[] = {CstHigh, CstHigh, CstHigh, CstHigh,
12413 CstHigh, CstHigh, CstHigh, CstHigh};
12414 SDValue VecCstHigh = DAG.getNode(ISD::BUILD_VECTOR, DL, VecIntVT,
12415 makeArrayRef(&CstHighArray[0], NumElts));
12417 // Create the right shift.
12418 SDValue CstShift = DAG.getConstant(16, DL, MVT::i32);
12419 SDValue CstShiftArray[] = {CstShift, CstShift, CstShift, CstShift,
12420 CstShift, CstShift, CstShift, CstShift};
12421 SDValue VecCstShift = DAG.getNode(ISD::BUILD_VECTOR, DL, VecIntVT,
12422 makeArrayRef(&CstShiftArray[0], NumElts));
12423 SDValue HighShift = DAG.getNode(ISD::SRL, DL, VecIntVT, V, VecCstShift);
12426 if (Subtarget.hasSSE41()) {
12427 EVT VecI16VT = Is128 ? MVT::v8i16 : MVT::v16i16;
12428 // uint4 lo = _mm_blend_epi16( v, (uint4) 0x4b000000, 0xaa);
12429 SDValue VecCstLowBitcast = DAG.getBitcast(VecI16VT, VecCstLow);
12430 SDValue VecBitcast = DAG.getBitcast(VecI16VT, V);
12431 // Low will be bitcasted right away, so do not bother bitcasting back to its
12433 Low = DAG.getNode(X86ISD::BLENDI, DL, VecI16VT, VecBitcast,
12434 VecCstLowBitcast, DAG.getConstant(0xaa, DL, MVT::i32));
12435 // uint4 hi = _mm_blend_epi16( _mm_srli_epi32(v,16),
12436 // (uint4) 0x53000000, 0xaa);
12437 SDValue VecCstHighBitcast = DAG.getBitcast(VecI16VT, VecCstHigh);
12438 SDValue VecShiftBitcast = DAG.getBitcast(VecI16VT, HighShift);
12439 // High will be bitcasted right away, so do not bother bitcasting back to
12440 // its original type.
12441 High = DAG.getNode(X86ISD::BLENDI, DL, VecI16VT, VecShiftBitcast,
12442 VecCstHighBitcast, DAG.getConstant(0xaa, DL, MVT::i32));
12444 SDValue CstMask = DAG.getConstant(0xffff, DL, MVT::i32);
12445 SDValue VecCstMask = DAG.getNode(ISD::BUILD_VECTOR, DL, VecIntVT, CstMask,
12446 CstMask, CstMask, CstMask);
12447 // uint4 lo = (v & (uint4) 0xffff) | (uint4) 0x4b000000;
12448 SDValue LowAnd = DAG.getNode(ISD::AND, DL, VecIntVT, V, VecCstMask);
12449 Low = DAG.getNode(ISD::OR, DL, VecIntVT, LowAnd, VecCstLow);
12451 // uint4 hi = (v >> 16) | (uint4) 0x53000000;
12452 High = DAG.getNode(ISD::OR, DL, VecIntVT, HighShift, VecCstHigh);
12455 // Create the vector constant for -(0x1.0p39f + 0x1.0p23f).
12456 SDValue CstFAdd = DAG.getConstantFP(
12457 APFloat(APFloat::IEEEsingle, APInt(32, 0xD3000080)), DL, MVT::f32);
12458 SDValue CstFAddArray[] = {CstFAdd, CstFAdd, CstFAdd, CstFAdd,
12459 CstFAdd, CstFAdd, CstFAdd, CstFAdd};
12460 SDValue VecCstFAdd = DAG.getNode(ISD::BUILD_VECTOR, DL, VecFloatVT,
12461 makeArrayRef(&CstFAddArray[0], NumElts));
12463 // float4 fhi = (float4) hi - (0x1.0p39f + 0x1.0p23f);
12464 SDValue HighBitcast = DAG.getBitcast(VecFloatVT, High);
12465 // TODO: Are there any fast-math-flags to propagate here?
12467 DAG.getNode(ISD::FADD, DL, VecFloatVT, HighBitcast, VecCstFAdd);
12468 // return (float4) lo + fhi;
12469 SDValue LowBitcast = DAG.getBitcast(VecFloatVT, Low);
12470 return DAG.getNode(ISD::FADD, DL, VecFloatVT, LowBitcast, FHigh);
12473 SDValue X86TargetLowering::lowerUINT_TO_FP_vec(SDValue Op,
12474 SelectionDAG &DAG) const {
12475 SDValue N0 = Op.getOperand(0);
12476 MVT SVT = N0.getSimpleValueType();
12479 switch (SVT.SimpleTy) {
12481 llvm_unreachable("Custom UINT_TO_FP is not supported!");
12486 MVT NVT = MVT::getVectorVT(MVT::i32, SVT.getVectorNumElements());
12487 return DAG.getNode(ISD::SINT_TO_FP, dl, Op.getValueType(),
12488 DAG.getNode(ISD::ZERO_EXTEND, dl, NVT, N0));
12492 return lowerUINT_TO_FP_vXi32(Op, DAG, *Subtarget);
12495 if (Subtarget->hasAVX512())
12496 return DAG.getNode(ISD::UINT_TO_FP, dl, Op.getValueType(),
12497 DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v16i32, N0));
12499 llvm_unreachable(nullptr);
12502 SDValue X86TargetLowering::LowerUINT_TO_FP(SDValue Op,
12503 SelectionDAG &DAG) const {
12504 SDValue N0 = Op.getOperand(0);
12506 auto PtrVT = getPointerTy(DAG.getDataLayout());
12508 if (Op.getValueType().isVector())
12509 return lowerUINT_TO_FP_vec(Op, DAG);
12511 // Since UINT_TO_FP is legal (it's marked custom), dag combiner won't
12512 // optimize it to a SINT_TO_FP when the sign bit is known zero. Perform
12513 // the optimization here.
12514 if (DAG.SignBitIsZero(N0))
12515 return DAG.getNode(ISD::SINT_TO_FP, dl, Op.getValueType(), N0);
12517 MVT SrcVT = N0.getSimpleValueType();
12518 MVT DstVT = Op.getSimpleValueType();
12519 if (SrcVT == MVT::i64 && DstVT == MVT::f64 && X86ScalarSSEf64)
12520 return LowerUINT_TO_FP_i64(Op, DAG);
12521 if (SrcVT == MVT::i32 && X86ScalarSSEf64)
12522 return LowerUINT_TO_FP_i32(Op, DAG);
12523 if (Subtarget->is64Bit() && SrcVT == MVT::i64 && DstVT == MVT::f32)
12526 // Make a 64-bit buffer, and use it to build an FILD.
12527 SDValue StackSlot = DAG.CreateStackTemporary(MVT::i64);
12528 if (SrcVT == MVT::i32) {
12529 SDValue WordOff = DAG.getConstant(4, dl, PtrVT);
12530 SDValue OffsetSlot = DAG.getNode(ISD::ADD, dl, PtrVT, StackSlot, WordOff);
12531 SDValue Store1 = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
12532 StackSlot, MachinePointerInfo(),
12534 SDValue Store2 = DAG.getStore(Store1, dl, DAG.getConstant(0, dl, MVT::i32),
12535 OffsetSlot, MachinePointerInfo(),
12537 SDValue Fild = BuildFILD(Op, MVT::i64, Store2, StackSlot, DAG);
12541 assert(SrcVT == MVT::i64 && "Unexpected type in UINT_TO_FP");
12542 SDValue Store = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
12543 StackSlot, MachinePointerInfo(),
12545 // For i64 source, we need to add the appropriate power of 2 if the input
12546 // was negative. This is the same as the optimization in
12547 // DAGTypeLegalizer::ExpandIntOp_UNIT_TO_FP, and for it to be safe here,
12548 // we must be careful to do the computation in x87 extended precision, not
12549 // in SSE. (The generic code can't know it's OK to do this, or how to.)
12550 int SSFI = cast<FrameIndexSDNode>(StackSlot)->getIndex();
12551 MachineMemOperand *MMO = DAG.getMachineFunction().getMachineMemOperand(
12552 MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), SSFI),
12553 MachineMemOperand::MOLoad, 8, 8);
12555 SDVTList Tys = DAG.getVTList(MVT::f80, MVT::Other);
12556 SDValue Ops[] = { Store, StackSlot, DAG.getValueType(MVT::i64) };
12557 SDValue Fild = DAG.getMemIntrinsicNode(X86ISD::FILD, dl, Tys, Ops,
12560 APInt FF(32, 0x5F800000ULL);
12562 // Check whether the sign bit is set.
12563 SDValue SignSet = DAG.getSetCC(
12564 dl, getSetCCResultType(DAG.getDataLayout(), *DAG.getContext(), MVT::i64),
12565 Op.getOperand(0), DAG.getConstant(0, dl, MVT::i64), ISD::SETLT);
12567 // Build a 64 bit pair (0, FF) in the constant pool, with FF in the lo bits.
12568 SDValue FudgePtr = DAG.getConstantPool(
12569 ConstantInt::get(*DAG.getContext(), FF.zext(64)), PtrVT);
12571 // Get a pointer to FF if the sign bit was set, or to 0 otherwise.
12572 SDValue Zero = DAG.getIntPtrConstant(0, dl);
12573 SDValue Four = DAG.getIntPtrConstant(4, dl);
12574 SDValue Offset = DAG.getNode(ISD::SELECT, dl, Zero.getValueType(), SignSet,
12576 FudgePtr = DAG.getNode(ISD::ADD, dl, PtrVT, FudgePtr, Offset);
12578 // Load the value out, extending it from f32 to f80.
12579 // FIXME: Avoid the extend by constructing the right constant pool?
12580 SDValue Fudge = DAG.getExtLoad(
12581 ISD::EXTLOAD, dl, MVT::f80, DAG.getEntryNode(), FudgePtr,
12582 MachinePointerInfo::getConstantPool(DAG.getMachineFunction()), MVT::f32,
12583 false, false, false, 4);
12584 // Extend everything to 80 bits to force it to be done on x87.
12585 // TODO: Are there any fast-math-flags to propagate here?
12586 SDValue Add = DAG.getNode(ISD::FADD, dl, MVT::f80, Fild, Fudge);
12587 return DAG.getNode(ISD::FP_ROUND, dl, DstVT, Add,
12588 DAG.getIntPtrConstant(0, dl));
12591 // If the given FP_TO_SINT (IsSigned) or FP_TO_UINT (!IsSigned) operation
12592 // is legal, or has an f16 source (which needs to be promoted to f32),
12593 // just return an <SDValue(), SDValue()> pair.
12594 // Otherwise it is assumed to be a conversion from one of f32, f64 or f80
12595 // to i16, i32 or i64, and we lower it to a legal sequence.
12596 // If lowered to the final integer result we return a <result, SDValue()> pair.
12597 // Otherwise we lower it to a sequence ending with a FIST, return a
12598 // <FIST, StackSlot> pair, and the caller is responsible for loading
12599 // the final integer result from StackSlot.
12600 std::pair<SDValue,SDValue>
12601 X86TargetLowering::FP_TO_INTHelper(SDValue Op, SelectionDAG &DAG,
12602 bool IsSigned, bool IsReplace) const {
12605 EVT DstTy = Op.getValueType();
12606 EVT TheVT = Op.getOperand(0).getValueType();
12607 auto PtrVT = getPointerTy(DAG.getDataLayout());
12609 if (TheVT == MVT::f16)
12610 // We need to promote the f16 to f32 before using the lowering
12611 // in this routine.
12612 return std::make_pair(SDValue(), SDValue());
12614 assert((TheVT == MVT::f32 ||
12615 TheVT == MVT::f64 ||
12616 TheVT == MVT::f80) &&
12617 "Unexpected FP operand type in FP_TO_INTHelper");
12619 // If using FIST to compute an unsigned i64, we'll need some fixup
12620 // to handle values above the maximum signed i64. A FIST is always
12621 // used for the 32-bit subtarget, but also for f80 on a 64-bit target.
12622 bool UnsignedFixup = !IsSigned &&
12623 DstTy == MVT::i64 &&
12624 (!Subtarget->is64Bit() ||
12625 !isScalarFPTypeInSSEReg(TheVT));
12627 if (!IsSigned && DstTy != MVT::i64 && !Subtarget->hasAVX512()) {
12628 // Replace the fp-to-uint32 operation with an fp-to-sint64 FIST.
12629 // The low 32 bits of the fist result will have the correct uint32 result.
12630 assert(DstTy == MVT::i32 && "Unexpected FP_TO_UINT");
12634 assert(DstTy.getSimpleVT() <= MVT::i64 &&
12635 DstTy.getSimpleVT() >= MVT::i16 &&
12636 "Unknown FP_TO_INT to lower!");
12638 // These are really Legal.
12639 if (DstTy == MVT::i32 &&
12640 isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType()))
12641 return std::make_pair(SDValue(), SDValue());
12642 if (Subtarget->is64Bit() &&
12643 DstTy == MVT::i64 &&
12644 isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType()))
12645 return std::make_pair(SDValue(), SDValue());
12647 // We lower FP->int64 into FISTP64 followed by a load from a temporary
12649 MachineFunction &MF = DAG.getMachineFunction();
12650 unsigned MemSize = DstTy.getSizeInBits()/8;
12651 int SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize, false);
12652 SDValue StackSlot = DAG.getFrameIndex(SSFI, PtrVT);
12655 switch (DstTy.getSimpleVT().SimpleTy) {
12656 default: llvm_unreachable("Invalid FP_TO_SINT to lower!");
12657 case MVT::i16: Opc = X86ISD::FP_TO_INT16_IN_MEM; break;
12658 case MVT::i32: Opc = X86ISD::FP_TO_INT32_IN_MEM; break;
12659 case MVT::i64: Opc = X86ISD::FP_TO_INT64_IN_MEM; break;
12662 SDValue Chain = DAG.getEntryNode();
12663 SDValue Value = Op.getOperand(0);
12664 SDValue Adjust; // 0x0 or 0x80000000, for result sign bit adjustment.
12666 if (UnsignedFixup) {
12668 // Conversion to unsigned i64 is implemented with a select,
12669 // depending on whether the source value fits in the range
12670 // of a signed i64. Let Thresh be the FP equivalent of
12671 // 0x8000000000000000ULL.
12673 // Adjust i32 = (Value < Thresh) ? 0 : 0x80000000;
12674 // FistSrc = (Value < Thresh) ? Value : (Value - Thresh);
12675 // Fist-to-mem64 FistSrc
12676 // Add 0 or 0x800...0ULL to the 64-bit result, which is equivalent
12677 // to XOR'ing the high 32 bits with Adjust.
12679 // Being a power of 2, Thresh is exactly representable in all FP formats.
12680 // For X87 we'd like to use the smallest FP type for this constant, but
12681 // for DAG type consistency we have to match the FP operand type.
12683 APFloat Thresh(APFloat::IEEEsingle, APInt(32, 0x5f000000));
12684 APFloat::opStatus Status = APFloat::opOK;
12685 bool LosesInfo = false;
12686 if (TheVT == MVT::f64)
12687 // The rounding mode is irrelevant as the conversion should be exact.
12688 Status = Thresh.convert(APFloat::IEEEdouble, APFloat::rmNearestTiesToEven,
12690 else if (TheVT == MVT::f80)
12691 Status = Thresh.convert(APFloat::x87DoubleExtended,
12692 APFloat::rmNearestTiesToEven, &LosesInfo);
12694 assert(Status == APFloat::opOK && !LosesInfo &&
12695 "FP conversion should have been exact");
12697 SDValue ThreshVal = DAG.getConstantFP(Thresh, DL, TheVT);
12699 SDValue Cmp = DAG.getSetCC(DL,
12700 getSetCCResultType(DAG.getDataLayout(),
12701 *DAG.getContext(), TheVT),
12702 Value, ThreshVal, ISD::SETLT);
12703 Adjust = DAG.getSelect(DL, MVT::i32, Cmp,
12704 DAG.getConstant(0, DL, MVT::i32),
12705 DAG.getConstant(0x80000000, DL, MVT::i32));
12706 SDValue Sub = DAG.getNode(ISD::FSUB, DL, TheVT, Value, ThreshVal);
12707 Cmp = DAG.getSetCC(DL, getSetCCResultType(DAG.getDataLayout(),
12708 *DAG.getContext(), TheVT),
12709 Value, ThreshVal, ISD::SETLT);
12710 Value = DAG.getSelect(DL, TheVT, Cmp, Value, Sub);
12713 // FIXME This causes a redundant load/store if the SSE-class value is already
12714 // in memory, such as if it is on the callstack.
12715 if (isScalarFPTypeInSSEReg(TheVT)) {
12716 assert(DstTy == MVT::i64 && "Invalid FP_TO_SINT to lower!");
12717 Chain = DAG.getStore(Chain, DL, Value, StackSlot,
12718 MachinePointerInfo::getFixedStack(MF, SSFI), false,
12720 SDVTList Tys = DAG.getVTList(Op.getOperand(0).getValueType(), MVT::Other);
12722 Chain, StackSlot, DAG.getValueType(TheVT)
12725 MachineMemOperand *MMO =
12726 MF.getMachineMemOperand(MachinePointerInfo::getFixedStack(MF, SSFI),
12727 MachineMemOperand::MOLoad, MemSize, MemSize);
12728 Value = DAG.getMemIntrinsicNode(X86ISD::FLD, DL, Tys, Ops, DstTy, MMO);
12729 Chain = Value.getValue(1);
12730 SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize, false);
12731 StackSlot = DAG.getFrameIndex(SSFI, PtrVT);
12734 MachineMemOperand *MMO =
12735 MF.getMachineMemOperand(MachinePointerInfo::getFixedStack(MF, SSFI),
12736 MachineMemOperand::MOStore, MemSize, MemSize);
12738 if (UnsignedFixup) {
12740 // Insert the FIST, load its result as two i32's,
12741 // and XOR the high i32 with Adjust.
12743 SDValue FistOps[] = { Chain, Value, StackSlot };
12744 SDValue FIST = DAG.getMemIntrinsicNode(Opc, DL, DAG.getVTList(MVT::Other),
12745 FistOps, DstTy, MMO);
12747 SDValue Low32 = DAG.getLoad(MVT::i32, DL, FIST, StackSlot,
12748 MachinePointerInfo(),
12749 false, false, false, 0);
12750 SDValue HighAddr = DAG.getNode(ISD::ADD, DL, PtrVT, StackSlot,
12751 DAG.getConstant(4, DL, PtrVT));
12753 SDValue High32 = DAG.getLoad(MVT::i32, DL, FIST, HighAddr,
12754 MachinePointerInfo(),
12755 false, false, false, 0);
12756 High32 = DAG.getNode(ISD::XOR, DL, MVT::i32, High32, Adjust);
12758 if (Subtarget->is64Bit()) {
12759 // Join High32 and Low32 into a 64-bit result.
12760 // (High32 << 32) | Low32
12761 Low32 = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::i64, Low32);
12762 High32 = DAG.getNode(ISD::ANY_EXTEND, DL, MVT::i64, High32);
12763 High32 = DAG.getNode(ISD::SHL, DL, MVT::i64, High32,
12764 DAG.getConstant(32, DL, MVT::i8));
12765 SDValue Result = DAG.getNode(ISD::OR, DL, MVT::i64, High32, Low32);
12766 return std::make_pair(Result, SDValue());
12769 SDValue ResultOps[] = { Low32, High32 };
12771 SDValue pair = IsReplace
12772 ? DAG.getNode(ISD::BUILD_PAIR, DL, MVT::i64, ResultOps)
12773 : DAG.getMergeValues(ResultOps, DL);
12774 return std::make_pair(pair, SDValue());
12776 // Build the FP_TO_INT*_IN_MEM
12777 SDValue Ops[] = { Chain, Value, StackSlot };
12778 SDValue FIST = DAG.getMemIntrinsicNode(Opc, DL, DAG.getVTList(MVT::Other),
12780 return std::make_pair(FIST, StackSlot);
12784 static SDValue LowerAVXExtend(SDValue Op, SelectionDAG &DAG,
12785 const X86Subtarget *Subtarget) {
12786 MVT VT = Op->getSimpleValueType(0);
12787 SDValue In = Op->getOperand(0);
12788 MVT InVT = In.getSimpleValueType();
12791 if (VT.is512BitVector() || InVT.getScalarType() == MVT::i1)
12792 return DAG.getNode(ISD::ZERO_EXTEND, dl, VT, In);
12794 // Optimize vectors in AVX mode:
12797 // Use vpunpcklwd for 4 lower elements v8i16 -> v4i32.
12798 // Use vpunpckhwd for 4 upper elements v8i16 -> v4i32.
12799 // Concat upper and lower parts.
12802 // Use vpunpckldq for 4 lower elements v4i32 -> v2i64.
12803 // Use vpunpckhdq for 4 upper elements v4i32 -> v2i64.
12804 // Concat upper and lower parts.
12807 if (((VT != MVT::v16i16) || (InVT != MVT::v16i8)) &&
12808 ((VT != MVT::v8i32) || (InVT != MVT::v8i16)) &&
12809 ((VT != MVT::v4i64) || (InVT != MVT::v4i32)))
12812 if (Subtarget->hasInt256())
12813 return DAG.getNode(X86ISD::VZEXT, dl, VT, In);
12815 SDValue ZeroVec = getZeroVector(InVT, Subtarget, DAG, dl);
12816 SDValue Undef = DAG.getUNDEF(InVT);
12817 bool NeedZero = Op.getOpcode() == ISD::ZERO_EXTEND;
12818 SDValue OpLo = getUnpackl(DAG, dl, InVT, In, NeedZero ? ZeroVec : Undef);
12819 SDValue OpHi = getUnpackh(DAG, dl, InVT, In, NeedZero ? ZeroVec : Undef);
12821 MVT HVT = MVT::getVectorVT(VT.getVectorElementType(),
12822 VT.getVectorNumElements()/2);
12824 OpLo = DAG.getBitcast(HVT, OpLo);
12825 OpHi = DAG.getBitcast(HVT, OpHi);
12827 return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT, OpLo, OpHi);
12830 static SDValue LowerZERO_EXTEND_AVX512(SDValue Op,
12831 const X86Subtarget *Subtarget, SelectionDAG &DAG) {
12832 MVT VT = Op->getSimpleValueType(0);
12833 SDValue In = Op->getOperand(0);
12834 MVT InVT = In.getSimpleValueType();
12836 unsigned int NumElts = VT.getVectorNumElements();
12837 if (NumElts != 8 && NumElts != 16 && !Subtarget->hasBWI())
12840 if (VT.is512BitVector() && InVT.getVectorElementType() != MVT::i1)
12841 return DAG.getNode(X86ISD::VZEXT, DL, VT, In);
12843 assert(InVT.getVectorElementType() == MVT::i1);
12844 MVT ExtVT = NumElts == 8 ? MVT::v8i64 : MVT::v16i32;
12846 DAG.getConstant(APInt(ExtVT.getScalarSizeInBits(), 1), DL, ExtVT);
12848 DAG.getConstant(APInt::getNullValue(ExtVT.getScalarSizeInBits()), DL, ExtVT);
12850 SDValue V = DAG.getNode(ISD::VSELECT, DL, ExtVT, In, One, Zero);
12851 if (VT.is512BitVector())
12853 return DAG.getNode(X86ISD::VTRUNC, DL, VT, V);
12856 static SDValue LowerANY_EXTEND(SDValue Op, const X86Subtarget *Subtarget,
12857 SelectionDAG &DAG) {
12858 if (Subtarget->hasFp256())
12859 if (SDValue Res = LowerAVXExtend(Op, DAG, Subtarget))
12865 static SDValue LowerZERO_EXTEND(SDValue Op, const X86Subtarget *Subtarget,
12866 SelectionDAG &DAG) {
12868 MVT VT = Op.getSimpleValueType();
12869 SDValue In = Op.getOperand(0);
12870 MVT SVT = In.getSimpleValueType();
12872 if (VT.is512BitVector() || SVT.getVectorElementType() == MVT::i1)
12873 return LowerZERO_EXTEND_AVX512(Op, Subtarget, DAG);
12875 if (Subtarget->hasFp256())
12876 if (SDValue Res = LowerAVXExtend(Op, DAG, Subtarget))
12879 assert(!VT.is256BitVector() || !SVT.is128BitVector() ||
12880 VT.getVectorNumElements() != SVT.getVectorNumElements());
12884 SDValue X86TargetLowering::LowerTRUNCATE(SDValue Op, SelectionDAG &DAG) const {
12886 MVT VT = Op.getSimpleValueType();
12887 SDValue In = Op.getOperand(0);
12888 MVT InVT = In.getSimpleValueType();
12890 if (VT == MVT::i1) {
12891 assert((InVT.isInteger() && (InVT.getSizeInBits() <= 64)) &&
12892 "Invalid scalar TRUNCATE operation");
12893 if (InVT.getSizeInBits() >= 32)
12895 In = DAG.getNode(ISD::ANY_EXTEND, DL, MVT::i32, In);
12896 return DAG.getNode(ISD::TRUNCATE, DL, VT, In);
12898 assert(VT.getVectorNumElements() == InVT.getVectorNumElements() &&
12899 "Invalid TRUNCATE operation");
12901 // move vector to mask - truncate solution for SKX
12902 if (VT.getVectorElementType() == MVT::i1) {
12903 if (InVT.is512BitVector() && InVT.getScalarSizeInBits() <= 16 &&
12904 Subtarget->hasBWI())
12905 return Op; // legal, will go to VPMOVB2M, VPMOVW2M
12906 if ((InVT.is256BitVector() || InVT.is128BitVector())
12907 && InVT.getScalarSizeInBits() <= 16 &&
12908 Subtarget->hasBWI() && Subtarget->hasVLX())
12909 return Op; // legal, will go to VPMOVB2M, VPMOVW2M
12910 if (InVT.is512BitVector() && InVT.getScalarSizeInBits() >= 32 &&
12911 Subtarget->hasDQI())
12912 return Op; // legal, will go to VPMOVD2M, VPMOVQ2M
12913 if ((InVT.is256BitVector() || InVT.is128BitVector())
12914 && InVT.getScalarSizeInBits() >= 32 &&
12915 Subtarget->hasDQI() && Subtarget->hasVLX())
12916 return Op; // legal, will go to VPMOVB2M, VPMOVQ2M
12919 if (VT.getVectorElementType() == MVT::i1) {
12920 assert(VT.getVectorElementType() == MVT::i1 && "Unexpected vector type");
12921 unsigned NumElts = InVT.getVectorNumElements();
12922 assert ((NumElts == 8 || NumElts == 16) && "Unexpected vector type");
12923 if (InVT.getSizeInBits() < 512) {
12924 MVT ExtVT = (NumElts == 16)? MVT::v16i32 : MVT::v8i64;
12925 In = DAG.getNode(ISD::SIGN_EXTEND, DL, ExtVT, In);
12930 DAG.getConstant(APInt::getSignBit(InVT.getScalarSizeInBits()), DL, InVT);
12931 SDValue And = DAG.getNode(ISD::AND, DL, InVT, OneV, In);
12932 return DAG.getNode(X86ISD::TESTM, DL, VT, And, And);
12935 // vpmovqb/w/d, vpmovdb/w, vpmovwb
12936 if (((!InVT.is512BitVector() && Subtarget->hasVLX()) || InVT.is512BitVector()) &&
12937 (InVT.getVectorElementType() != MVT::i16 || Subtarget->hasBWI()))
12938 return DAG.getNode(X86ISD::VTRUNC, DL, VT, In);
12940 if ((VT == MVT::v4i32) && (InVT == MVT::v4i64)) {
12941 // On AVX2, v4i64 -> v4i32 becomes VPERMD.
12942 if (Subtarget->hasInt256()) {
12943 static const int ShufMask[] = {0, 2, 4, 6, -1, -1, -1, -1};
12944 In = DAG.getBitcast(MVT::v8i32, In);
12945 In = DAG.getVectorShuffle(MVT::v8i32, DL, In, DAG.getUNDEF(MVT::v8i32),
12947 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, VT, In,
12948 DAG.getIntPtrConstant(0, DL));
12951 SDValue OpLo = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v2i64, In,
12952 DAG.getIntPtrConstant(0, DL));
12953 SDValue OpHi = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v2i64, In,
12954 DAG.getIntPtrConstant(2, DL));
12955 OpLo = DAG.getBitcast(MVT::v4i32, OpLo);
12956 OpHi = DAG.getBitcast(MVT::v4i32, OpHi);
12957 static const int ShufMask[] = {0, 2, 4, 6};
12958 return DAG.getVectorShuffle(VT, DL, OpLo, OpHi, ShufMask);
12961 if ((VT == MVT::v8i16) && (InVT == MVT::v8i32)) {
12962 // On AVX2, v8i32 -> v8i16 becomed PSHUFB.
12963 if (Subtarget->hasInt256()) {
12964 In = DAG.getBitcast(MVT::v32i8, In);
12966 SmallVector<SDValue,32> pshufbMask;
12967 for (unsigned i = 0; i < 2; ++i) {
12968 pshufbMask.push_back(DAG.getConstant(0x0, DL, MVT::i8));
12969 pshufbMask.push_back(DAG.getConstant(0x1, DL, MVT::i8));
12970 pshufbMask.push_back(DAG.getConstant(0x4, DL, MVT::i8));
12971 pshufbMask.push_back(DAG.getConstant(0x5, DL, MVT::i8));
12972 pshufbMask.push_back(DAG.getConstant(0x8, DL, MVT::i8));
12973 pshufbMask.push_back(DAG.getConstant(0x9, DL, MVT::i8));
12974 pshufbMask.push_back(DAG.getConstant(0xc, DL, MVT::i8));
12975 pshufbMask.push_back(DAG.getConstant(0xd, DL, MVT::i8));
12976 for (unsigned j = 0; j < 8; ++j)
12977 pshufbMask.push_back(DAG.getConstant(0x80, DL, MVT::i8));
12979 SDValue BV = DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v32i8, pshufbMask);
12980 In = DAG.getNode(X86ISD::PSHUFB, DL, MVT::v32i8, In, BV);
12981 In = DAG.getBitcast(MVT::v4i64, In);
12983 static const int ShufMask[] = {0, 2, -1, -1};
12984 In = DAG.getVectorShuffle(MVT::v4i64, DL, In, DAG.getUNDEF(MVT::v4i64),
12986 In = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v2i64, In,
12987 DAG.getIntPtrConstant(0, DL));
12988 return DAG.getBitcast(VT, In);
12991 SDValue OpLo = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v4i32, In,
12992 DAG.getIntPtrConstant(0, DL));
12994 SDValue OpHi = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v4i32, In,
12995 DAG.getIntPtrConstant(4, DL));
12997 OpLo = DAG.getBitcast(MVT::v16i8, OpLo);
12998 OpHi = DAG.getBitcast(MVT::v16i8, OpHi);
13000 // The PSHUFB mask:
13001 static const int ShufMask1[] = {0, 1, 4, 5, 8, 9, 12, 13,
13002 -1, -1, -1, -1, -1, -1, -1, -1};
13004 SDValue Undef = DAG.getUNDEF(MVT::v16i8);
13005 OpLo = DAG.getVectorShuffle(MVT::v16i8, DL, OpLo, Undef, ShufMask1);
13006 OpHi = DAG.getVectorShuffle(MVT::v16i8, DL, OpHi, Undef, ShufMask1);
13008 OpLo = DAG.getBitcast(MVT::v4i32, OpLo);
13009 OpHi = DAG.getBitcast(MVT::v4i32, OpHi);
13011 // The MOVLHPS Mask:
13012 static const int ShufMask2[] = {0, 1, 4, 5};
13013 SDValue res = DAG.getVectorShuffle(MVT::v4i32, DL, OpLo, OpHi, ShufMask2);
13014 return DAG.getBitcast(MVT::v8i16, res);
13017 // Handle truncation of V256 to V128 using shuffles.
13018 if (!VT.is128BitVector() || !InVT.is256BitVector())
13021 assert(Subtarget->hasFp256() && "256-bit vector without AVX!");
13023 unsigned NumElems = VT.getVectorNumElements();
13024 MVT NVT = MVT::getVectorVT(VT.getVectorElementType(), NumElems * 2);
13026 SmallVector<int, 16> MaskVec(NumElems * 2, -1);
13027 // Prepare truncation shuffle mask
13028 for (unsigned i = 0; i != NumElems; ++i)
13029 MaskVec[i] = i * 2;
13030 SDValue V = DAG.getVectorShuffle(NVT, DL, DAG.getBitcast(NVT, In),
13031 DAG.getUNDEF(NVT), &MaskVec[0]);
13032 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, VT, V,
13033 DAG.getIntPtrConstant(0, DL));
13036 SDValue X86TargetLowering::LowerFP_TO_SINT(SDValue Op,
13037 SelectionDAG &DAG) const {
13038 assert(!Op.getSimpleValueType().isVector());
13040 std::pair<SDValue,SDValue> Vals = FP_TO_INTHelper(Op, DAG,
13041 /*IsSigned=*/ true, /*IsReplace=*/ false);
13042 SDValue FIST = Vals.first, StackSlot = Vals.second;
13043 // If FP_TO_INTHelper failed, the node is actually supposed to be Legal.
13044 if (!FIST.getNode())
13047 if (StackSlot.getNode())
13048 // Load the result.
13049 return DAG.getLoad(Op.getValueType(), SDLoc(Op),
13050 FIST, StackSlot, MachinePointerInfo(),
13051 false, false, false, 0);
13053 // The node is the result.
13057 SDValue X86TargetLowering::LowerFP_TO_UINT(SDValue Op,
13058 SelectionDAG &DAG) const {
13059 std::pair<SDValue,SDValue> Vals = FP_TO_INTHelper(Op, DAG,
13060 /*IsSigned=*/ false, /*IsReplace=*/ false);
13061 SDValue FIST = Vals.first, StackSlot = Vals.second;
13062 // If FP_TO_INTHelper failed, the node is actually supposed to be Legal.
13063 if (!FIST.getNode())
13066 if (StackSlot.getNode())
13067 // Load the result.
13068 return DAG.getLoad(Op.getValueType(), SDLoc(Op),
13069 FIST, StackSlot, MachinePointerInfo(),
13070 false, false, false, 0);
13072 // The node is the result.
13076 static SDValue LowerFP_EXTEND(SDValue Op, SelectionDAG &DAG) {
13078 MVT VT = Op.getSimpleValueType();
13079 SDValue In = Op.getOperand(0);
13080 MVT SVT = In.getSimpleValueType();
13082 assert(SVT == MVT::v2f32 && "Only customize MVT::v2f32 type legalization!");
13084 return DAG.getNode(X86ISD::VFPEXT, DL, VT,
13085 DAG.getNode(ISD::CONCAT_VECTORS, DL, MVT::v4f32,
13086 In, DAG.getUNDEF(SVT)));
13089 /// The only differences between FABS and FNEG are the mask and the logic op.
13090 /// FNEG also has a folding opportunity for FNEG(FABS(x)).
13091 static SDValue LowerFABSorFNEG(SDValue Op, SelectionDAG &DAG) {
13092 assert((Op.getOpcode() == ISD::FABS || Op.getOpcode() == ISD::FNEG) &&
13093 "Wrong opcode for lowering FABS or FNEG.");
13095 bool IsFABS = (Op.getOpcode() == ISD::FABS);
13097 // If this is a FABS and it has an FNEG user, bail out to fold the combination
13098 // into an FNABS. We'll lower the FABS after that if it is still in use.
13100 for (SDNode *User : Op->uses())
13101 if (User->getOpcode() == ISD::FNEG)
13105 MVT VT = Op.getSimpleValueType();
13107 // FIXME: Use function attribute "OptimizeForSize" and/or CodeGenOpt::Level to
13108 // decide if we should generate a 16-byte constant mask when we only need 4 or
13109 // 8 bytes for the scalar case.
13115 if (VT.isVector()) {
13117 EltVT = VT.getVectorElementType();
13118 NumElts = VT.getVectorNumElements();
13120 // There are no scalar bitwise logical SSE/AVX instructions, so we
13121 // generate a 16-byte vector constant and logic op even for the scalar case.
13122 // Using a 16-byte mask allows folding the load of the mask with
13123 // the logic op, so it can save (~4 bytes) on code size.
13124 LogicVT = (VT == MVT::f64) ? MVT::v2f64 : MVT::v4f32;
13126 NumElts = (VT == MVT::f64) ? 2 : 4;
13129 unsigned EltBits = EltVT.getSizeInBits();
13130 LLVMContext *Context = DAG.getContext();
13131 // For FABS, mask is 0x7f...; for FNEG, mask is 0x80...
13133 IsFABS ? APInt::getSignedMaxValue(EltBits) : APInt::getSignBit(EltBits);
13134 Constant *C = ConstantInt::get(*Context, MaskElt);
13135 C = ConstantVector::getSplat(NumElts, C);
13136 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
13137 SDValue CPIdx = DAG.getConstantPool(C, TLI.getPointerTy(DAG.getDataLayout()));
13138 unsigned Alignment = cast<ConstantPoolSDNode>(CPIdx)->getAlignment();
13140 DAG.getLoad(LogicVT, dl, DAG.getEntryNode(), CPIdx,
13141 MachinePointerInfo::getConstantPool(DAG.getMachineFunction()),
13142 false, false, false, Alignment);
13144 SDValue Op0 = Op.getOperand(0);
13145 bool IsFNABS = !IsFABS && (Op0.getOpcode() == ISD::FABS);
13147 IsFABS ? X86ISD::FAND : IsFNABS ? X86ISD::FOR : X86ISD::FXOR;
13148 SDValue Operand = IsFNABS ? Op0.getOperand(0) : Op0;
13151 return DAG.getNode(LogicOp, dl, LogicVT, Operand, Mask);
13153 // For the scalar case extend to a 128-bit vector, perform the logic op,
13154 // and extract the scalar result back out.
13155 Operand = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, LogicVT, Operand);
13156 SDValue LogicNode = DAG.getNode(LogicOp, dl, LogicVT, Operand, Mask);
13157 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, LogicNode,
13158 DAG.getIntPtrConstant(0, dl));
13161 static SDValue LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) {
13162 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
13163 LLVMContext *Context = DAG.getContext();
13164 SDValue Op0 = Op.getOperand(0);
13165 SDValue Op1 = Op.getOperand(1);
13167 MVT VT = Op.getSimpleValueType();
13168 MVT SrcVT = Op1.getSimpleValueType();
13170 // If second operand is smaller, extend it first.
13171 if (SrcVT.bitsLT(VT)) {
13172 Op1 = DAG.getNode(ISD::FP_EXTEND, dl, VT, Op1);
13175 // And if it is bigger, shrink it first.
13176 if (SrcVT.bitsGT(VT)) {
13177 Op1 = DAG.getNode(ISD::FP_ROUND, dl, VT, Op1, DAG.getIntPtrConstant(1, dl));
13181 // At this point the operands and the result should have the same
13182 // type, and that won't be f80 since that is not custom lowered.
13184 const fltSemantics &Sem =
13185 VT == MVT::f64 ? APFloat::IEEEdouble : APFloat::IEEEsingle;
13186 const unsigned SizeInBits = VT.getSizeInBits();
13188 SmallVector<Constant *, 4> CV(
13189 VT == MVT::f64 ? 2 : 4,
13190 ConstantFP::get(*Context, APFloat(Sem, APInt(SizeInBits, 0))));
13192 // First, clear all bits but the sign bit from the second operand (sign).
13193 CV[0] = ConstantFP::get(*Context,
13194 APFloat(Sem, APInt::getHighBitsSet(SizeInBits, 1)));
13195 Constant *C = ConstantVector::get(CV);
13196 auto PtrVT = TLI.getPointerTy(DAG.getDataLayout());
13197 SDValue CPIdx = DAG.getConstantPool(C, PtrVT, 16);
13199 // Perform all logic operations as 16-byte vectors because there are no
13200 // scalar FP logic instructions in SSE. This allows load folding of the
13201 // constants into the logic instructions.
13202 MVT LogicVT = (VT == MVT::f64) ? MVT::v2f64 : MVT::v4f32;
13204 DAG.getLoad(LogicVT, dl, DAG.getEntryNode(), CPIdx,
13205 MachinePointerInfo::getConstantPool(DAG.getMachineFunction()),
13206 false, false, false, 16);
13207 Op1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, LogicVT, Op1);
13208 SDValue SignBit = DAG.getNode(X86ISD::FAND, dl, LogicVT, Op1, Mask1);
13210 // Next, clear the sign bit from the first operand (magnitude).
13211 // If it's a constant, we can clear it here.
13212 if (ConstantFPSDNode *Op0CN = dyn_cast<ConstantFPSDNode>(Op0)) {
13213 APFloat APF = Op0CN->getValueAPF();
13214 // If the magnitude is a positive zero, the sign bit alone is enough.
13215 if (APF.isPosZero())
13216 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, SrcVT, SignBit,
13217 DAG.getIntPtrConstant(0, dl));
13219 CV[0] = ConstantFP::get(*Context, APF);
13221 CV[0] = ConstantFP::get(
13223 APFloat(Sem, APInt::getLowBitsSet(SizeInBits, SizeInBits - 1)));
13225 C = ConstantVector::get(CV);
13226 CPIdx = DAG.getConstantPool(C, PtrVT, 16);
13228 DAG.getLoad(LogicVT, dl, DAG.getEntryNode(), CPIdx,
13229 MachinePointerInfo::getConstantPool(DAG.getMachineFunction()),
13230 false, false, false, 16);
13231 // If the magnitude operand wasn't a constant, we need to AND out the sign.
13232 if (!isa<ConstantFPSDNode>(Op0)) {
13233 Op0 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, LogicVT, Op0);
13234 Val = DAG.getNode(X86ISD::FAND, dl, LogicVT, Op0, Val);
13236 // OR the magnitude value with the sign bit.
13237 Val = DAG.getNode(X86ISD::FOR, dl, LogicVT, Val, SignBit);
13238 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, SrcVT, Val,
13239 DAG.getIntPtrConstant(0, dl));
13242 static SDValue LowerFGETSIGN(SDValue Op, SelectionDAG &DAG) {
13243 SDValue N0 = Op.getOperand(0);
13245 MVT VT = Op.getSimpleValueType();
13247 // Lower ISD::FGETSIGN to (AND (X86ISD::FGETSIGNx86 ...) 1).
13248 SDValue xFGETSIGN = DAG.getNode(X86ISD::FGETSIGNx86, dl, VT, N0,
13249 DAG.getConstant(1, dl, VT));
13250 return DAG.getNode(ISD::AND, dl, VT, xFGETSIGN, DAG.getConstant(1, dl, VT));
13253 // Check whether an OR'd tree is PTEST-able.
13254 static SDValue LowerVectorAllZeroTest(SDValue Op, const X86Subtarget *Subtarget,
13255 SelectionDAG &DAG) {
13256 assert(Op.getOpcode() == ISD::OR && "Only check OR'd tree.");
13258 if (!Subtarget->hasSSE41())
13261 if (!Op->hasOneUse())
13264 SDNode *N = Op.getNode();
13267 SmallVector<SDValue, 8> Opnds;
13268 DenseMap<SDValue, unsigned> VecInMap;
13269 SmallVector<SDValue, 8> VecIns;
13270 EVT VT = MVT::Other;
13272 // Recognize a special case where a vector is casted into wide integer to
13274 Opnds.push_back(N->getOperand(0));
13275 Opnds.push_back(N->getOperand(1));
13277 for (unsigned Slot = 0, e = Opnds.size(); Slot < e; ++Slot) {
13278 SmallVectorImpl<SDValue>::const_iterator I = Opnds.begin() + Slot;
13279 // BFS traverse all OR'd operands.
13280 if (I->getOpcode() == ISD::OR) {
13281 Opnds.push_back(I->getOperand(0));
13282 Opnds.push_back(I->getOperand(1));
13283 // Re-evaluate the number of nodes to be traversed.
13284 e += 2; // 2 more nodes (LHS and RHS) are pushed.
13288 // Quit if a non-EXTRACT_VECTOR_ELT
13289 if (I->getOpcode() != ISD::EXTRACT_VECTOR_ELT)
13292 // Quit if without a constant index.
13293 SDValue Idx = I->getOperand(1);
13294 if (!isa<ConstantSDNode>(Idx))
13297 SDValue ExtractedFromVec = I->getOperand(0);
13298 DenseMap<SDValue, unsigned>::iterator M = VecInMap.find(ExtractedFromVec);
13299 if (M == VecInMap.end()) {
13300 VT = ExtractedFromVec.getValueType();
13301 // Quit if not 128/256-bit vector.
13302 if (!VT.is128BitVector() && !VT.is256BitVector())
13304 // Quit if not the same type.
13305 if (VecInMap.begin() != VecInMap.end() &&
13306 VT != VecInMap.begin()->first.getValueType())
13308 M = VecInMap.insert(std::make_pair(ExtractedFromVec, 0)).first;
13309 VecIns.push_back(ExtractedFromVec);
13311 M->second |= 1U << cast<ConstantSDNode>(Idx)->getZExtValue();
13314 assert((VT.is128BitVector() || VT.is256BitVector()) &&
13315 "Not extracted from 128-/256-bit vector.");
13317 unsigned FullMask = (1U << VT.getVectorNumElements()) - 1U;
13319 for (DenseMap<SDValue, unsigned>::const_iterator
13320 I = VecInMap.begin(), E = VecInMap.end(); I != E; ++I) {
13321 // Quit if not all elements are used.
13322 if (I->second != FullMask)
13326 EVT TestVT = VT.is128BitVector() ? MVT::v2i64 : MVT::v4i64;
13328 // Cast all vectors into TestVT for PTEST.
13329 for (unsigned i = 0, e = VecIns.size(); i < e; ++i)
13330 VecIns[i] = DAG.getBitcast(TestVT, VecIns[i]);
13332 // If more than one full vectors are evaluated, OR them first before PTEST.
13333 for (unsigned Slot = 0, e = VecIns.size(); e - Slot > 1; Slot += 2, e += 1) {
13334 // Each iteration will OR 2 nodes and append the result until there is only
13335 // 1 node left, i.e. the final OR'd value of all vectors.
13336 SDValue LHS = VecIns[Slot];
13337 SDValue RHS = VecIns[Slot + 1];
13338 VecIns.push_back(DAG.getNode(ISD::OR, DL, TestVT, LHS, RHS));
13341 return DAG.getNode(X86ISD::PTEST, DL, MVT::i32,
13342 VecIns.back(), VecIns.back());
13345 /// \brief return true if \c Op has a use that doesn't just read flags.
13346 static bool hasNonFlagsUse(SDValue Op) {
13347 for (SDNode::use_iterator UI = Op->use_begin(), UE = Op->use_end(); UI != UE;
13349 SDNode *User = *UI;
13350 unsigned UOpNo = UI.getOperandNo();
13351 if (User->getOpcode() == ISD::TRUNCATE && User->hasOneUse()) {
13352 // Look pass truncate.
13353 UOpNo = User->use_begin().getOperandNo();
13354 User = *User->use_begin();
13357 if (User->getOpcode() != ISD::BRCOND && User->getOpcode() != ISD::SETCC &&
13358 !(User->getOpcode() == ISD::SELECT && UOpNo == 0))
13364 /// Emit nodes that will be selected as "test Op0,Op0", or something
13366 SDValue X86TargetLowering::EmitTest(SDValue Op, unsigned X86CC, SDLoc dl,
13367 SelectionDAG &DAG) const {
13368 if (Op.getValueType() == MVT::i1) {
13369 SDValue ExtOp = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i8, Op);
13370 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, ExtOp,
13371 DAG.getConstant(0, dl, MVT::i8));
13373 // CF and OF aren't always set the way we want. Determine which
13374 // of these we need.
13375 bool NeedCF = false;
13376 bool NeedOF = false;
13379 case X86::COND_A: case X86::COND_AE:
13380 case X86::COND_B: case X86::COND_BE:
13383 case X86::COND_G: case X86::COND_GE:
13384 case X86::COND_L: case X86::COND_LE:
13385 case X86::COND_O: case X86::COND_NO: {
13386 // Check if we really need to set the
13387 // Overflow flag. If NoSignedWrap is present
13388 // that is not actually needed.
13389 switch (Op->getOpcode()) {
13394 const auto *BinNode = cast<BinaryWithFlagsSDNode>(Op.getNode());
13395 if (BinNode->Flags.hasNoSignedWrap())
13405 // See if we can use the EFLAGS value from the operand instead of
13406 // doing a separate TEST. TEST always sets OF and CF to 0, so unless
13407 // we prove that the arithmetic won't overflow, we can't use OF or CF.
13408 if (Op.getResNo() != 0 || NeedOF || NeedCF) {
13409 // Emit a CMP with 0, which is the TEST pattern.
13410 //if (Op.getValueType() == MVT::i1)
13411 // return DAG.getNode(X86ISD::CMP, dl, MVT::i1, Op,
13412 // DAG.getConstant(0, MVT::i1));
13413 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op,
13414 DAG.getConstant(0, dl, Op.getValueType()));
13416 unsigned Opcode = 0;
13417 unsigned NumOperands = 0;
13419 // Truncate operations may prevent the merge of the SETCC instruction
13420 // and the arithmetic instruction before it. Attempt to truncate the operands
13421 // of the arithmetic instruction and use a reduced bit-width instruction.
13422 bool NeedTruncation = false;
13423 SDValue ArithOp = Op;
13424 if (Op->getOpcode() == ISD::TRUNCATE && Op->hasOneUse()) {
13425 SDValue Arith = Op->getOperand(0);
13426 // Both the trunc and the arithmetic op need to have one user each.
13427 if (Arith->hasOneUse())
13428 switch (Arith.getOpcode()) {
13435 NeedTruncation = true;
13441 // NOTICE: In the code below we use ArithOp to hold the arithmetic operation
13442 // which may be the result of a CAST. We use the variable 'Op', which is the
13443 // non-casted variable when we check for possible users.
13444 switch (ArithOp.getOpcode()) {
13446 // Due to an isel shortcoming, be conservative if this add is likely to be
13447 // selected as part of a load-modify-store instruction. When the root node
13448 // in a match is a store, isel doesn't know how to remap non-chain non-flag
13449 // uses of other nodes in the match, such as the ADD in this case. This
13450 // leads to the ADD being left around and reselected, with the result being
13451 // two adds in the output. Alas, even if none our users are stores, that
13452 // doesn't prove we're O.K. Ergo, if we have any parents that aren't
13453 // CopyToReg or SETCC, eschew INC/DEC. A better fix seems to require
13454 // climbing the DAG back to the root, and it doesn't seem to be worth the
13456 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
13457 UE = Op.getNode()->use_end(); UI != UE; ++UI)
13458 if (UI->getOpcode() != ISD::CopyToReg &&
13459 UI->getOpcode() != ISD::SETCC &&
13460 UI->getOpcode() != ISD::STORE)
13463 if (ConstantSDNode *C =
13464 dyn_cast<ConstantSDNode>(ArithOp.getNode()->getOperand(1))) {
13465 // An add of one will be selected as an INC.
13466 if (C->getAPIntValue() == 1 && !Subtarget->slowIncDec()) {
13467 Opcode = X86ISD::INC;
13472 // An add of negative one (subtract of one) will be selected as a DEC.
13473 if (C->getAPIntValue().isAllOnesValue() && !Subtarget->slowIncDec()) {
13474 Opcode = X86ISD::DEC;
13480 // Otherwise use a regular EFLAGS-setting add.
13481 Opcode = X86ISD::ADD;
13486 // If we have a constant logical shift that's only used in a comparison
13487 // against zero turn it into an equivalent AND. This allows turning it into
13488 // a TEST instruction later.
13489 if ((X86CC == X86::COND_E || X86CC == X86::COND_NE) && Op->hasOneUse() &&
13490 isa<ConstantSDNode>(Op->getOperand(1)) && !hasNonFlagsUse(Op)) {
13491 EVT VT = Op.getValueType();
13492 unsigned BitWidth = VT.getSizeInBits();
13493 unsigned ShAmt = Op->getConstantOperandVal(1);
13494 if (ShAmt >= BitWidth) // Avoid undefined shifts.
13496 APInt Mask = ArithOp.getOpcode() == ISD::SRL
13497 ? APInt::getHighBitsSet(BitWidth, BitWidth - ShAmt)
13498 : APInt::getLowBitsSet(BitWidth, BitWidth - ShAmt);
13499 if (!Mask.isSignedIntN(32)) // Avoid large immediates.
13501 SDValue New = DAG.getNode(ISD::AND, dl, VT, Op->getOperand(0),
13502 DAG.getConstant(Mask, dl, VT));
13503 DAG.ReplaceAllUsesWith(Op, New);
13509 // If the primary and result isn't used, don't bother using X86ISD::AND,
13510 // because a TEST instruction will be better.
13511 if (!hasNonFlagsUse(Op))
13517 // Due to the ISEL shortcoming noted above, be conservative if this op is
13518 // likely to be selected as part of a load-modify-store instruction.
13519 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
13520 UE = Op.getNode()->use_end(); UI != UE; ++UI)
13521 if (UI->getOpcode() == ISD::STORE)
13524 // Otherwise use a regular EFLAGS-setting instruction.
13525 switch (ArithOp.getOpcode()) {
13526 default: llvm_unreachable("unexpected operator!");
13527 case ISD::SUB: Opcode = X86ISD::SUB; break;
13528 case ISD::XOR: Opcode = X86ISD::XOR; break;
13529 case ISD::AND: Opcode = X86ISD::AND; break;
13531 if (!NeedTruncation && (X86CC == X86::COND_E || X86CC == X86::COND_NE)) {
13532 SDValue EFLAGS = LowerVectorAllZeroTest(Op, Subtarget, DAG);
13533 if (EFLAGS.getNode())
13536 Opcode = X86ISD::OR;
13550 return SDValue(Op.getNode(), 1);
13556 // If we found that truncation is beneficial, perform the truncation and
13558 if (NeedTruncation) {
13559 EVT VT = Op.getValueType();
13560 SDValue WideVal = Op->getOperand(0);
13561 EVT WideVT = WideVal.getValueType();
13562 unsigned ConvertedOp = 0;
13563 // Use a target machine opcode to prevent further DAGCombine
13564 // optimizations that may separate the arithmetic operations
13565 // from the setcc node.
13566 switch (WideVal.getOpcode()) {
13568 case ISD::ADD: ConvertedOp = X86ISD::ADD; break;
13569 case ISD::SUB: ConvertedOp = X86ISD::SUB; break;
13570 case ISD::AND: ConvertedOp = X86ISD::AND; break;
13571 case ISD::OR: ConvertedOp = X86ISD::OR; break;
13572 case ISD::XOR: ConvertedOp = X86ISD::XOR; break;
13576 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
13577 if (TLI.isOperationLegal(WideVal.getOpcode(), WideVT)) {
13578 SDValue V0 = DAG.getNode(ISD::TRUNCATE, dl, VT, WideVal.getOperand(0));
13579 SDValue V1 = DAG.getNode(ISD::TRUNCATE, dl, VT, WideVal.getOperand(1));
13580 Op = DAG.getNode(ConvertedOp, dl, VT, V0, V1);
13586 // Emit a CMP with 0, which is the TEST pattern.
13587 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op,
13588 DAG.getConstant(0, dl, Op.getValueType()));
13590 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::i32);
13591 SmallVector<SDValue, 4> Ops(Op->op_begin(), Op->op_begin() + NumOperands);
13593 SDValue New = DAG.getNode(Opcode, dl, VTs, Ops);
13594 DAG.ReplaceAllUsesWith(Op, New);
13595 return SDValue(New.getNode(), 1);
13598 /// Emit nodes that will be selected as "cmp Op0,Op1", or something
13600 SDValue X86TargetLowering::EmitCmp(SDValue Op0, SDValue Op1, unsigned X86CC,
13601 SDLoc dl, SelectionDAG &DAG) const {
13602 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op1)) {
13603 if (C->getAPIntValue() == 0)
13604 return EmitTest(Op0, X86CC, dl, DAG);
13606 if (Op0.getValueType() == MVT::i1)
13607 llvm_unreachable("Unexpected comparison operation for MVT::i1 operands");
13610 if ((Op0.getValueType() == MVT::i8 || Op0.getValueType() == MVT::i16 ||
13611 Op0.getValueType() == MVT::i32 || Op0.getValueType() == MVT::i64)) {
13612 // Do the comparison at i32 if it's smaller, besides the Atom case.
13613 // This avoids subregister aliasing issues. Keep the smaller reference
13614 // if we're optimizing for size, however, as that'll allow better folding
13615 // of memory operations.
13616 if (Op0.getValueType() != MVT::i32 && Op0.getValueType() != MVT::i64 &&
13617 !DAG.getMachineFunction().getFunction()->optForMinSize() &&
13618 !Subtarget->isAtom()) {
13619 unsigned ExtendOp =
13620 isX86CCUnsigned(X86CC) ? ISD::ZERO_EXTEND : ISD::SIGN_EXTEND;
13621 Op0 = DAG.getNode(ExtendOp, dl, MVT::i32, Op0);
13622 Op1 = DAG.getNode(ExtendOp, dl, MVT::i32, Op1);
13624 // Use SUB instead of CMP to enable CSE between SUB and CMP.
13625 SDVTList VTs = DAG.getVTList(Op0.getValueType(), MVT::i32);
13626 SDValue Sub = DAG.getNode(X86ISD::SUB, dl, VTs,
13628 return SDValue(Sub.getNode(), 1);
13630 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op0, Op1);
13633 /// Convert a comparison if required by the subtarget.
13634 SDValue X86TargetLowering::ConvertCmpIfNecessary(SDValue Cmp,
13635 SelectionDAG &DAG) const {
13636 // If the subtarget does not support the FUCOMI instruction, floating-point
13637 // comparisons have to be converted.
13638 if (Subtarget->hasCMov() ||
13639 Cmp.getOpcode() != X86ISD::CMP ||
13640 !Cmp.getOperand(0).getValueType().isFloatingPoint() ||
13641 !Cmp.getOperand(1).getValueType().isFloatingPoint())
13644 // The instruction selector will select an FUCOM instruction instead of
13645 // FUCOMI, which writes the comparison result to FPSW instead of EFLAGS. Hence
13646 // build an SDNode sequence that transfers the result from FPSW into EFLAGS:
13647 // (X86sahf (trunc (srl (X86fp_stsw (trunc (X86cmp ...)), 8))))
13649 SDValue TruncFPSW = DAG.getNode(ISD::TRUNCATE, dl, MVT::i16, Cmp);
13650 SDValue FNStSW = DAG.getNode(X86ISD::FNSTSW16r, dl, MVT::i16, TruncFPSW);
13651 SDValue Srl = DAG.getNode(ISD::SRL, dl, MVT::i16, FNStSW,
13652 DAG.getConstant(8, dl, MVT::i8));
13653 SDValue TruncSrl = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Srl);
13654 return DAG.getNode(X86ISD::SAHF, dl, MVT::i32, TruncSrl);
13657 /// The minimum architected relative accuracy is 2^-12. We need one
13658 /// Newton-Raphson step to have a good float result (24 bits of precision).
13659 SDValue X86TargetLowering::getRsqrtEstimate(SDValue Op,
13660 DAGCombinerInfo &DCI,
13661 unsigned &RefinementSteps,
13662 bool &UseOneConstNR) const {
13663 EVT VT = Op.getValueType();
13664 const char *RecipOp;
13666 // SSE1 has rsqrtss and rsqrtps. AVX adds a 256-bit variant for rsqrtps.
13667 // TODO: Add support for AVX512 (v16f32).
13668 // It is likely not profitable to do this for f64 because a double-precision
13669 // rsqrt estimate with refinement on x86 prior to FMA requires at least 16
13670 // instructions: convert to single, rsqrtss, convert back to double, refine
13671 // (3 steps = at least 13 insts). If an 'rsqrtsd' variant was added to the ISA
13672 // along with FMA, this could be a throughput win.
13673 if (VT == MVT::f32 && Subtarget->hasSSE1())
13675 else if ((VT == MVT::v4f32 && Subtarget->hasSSE1()) ||
13676 (VT == MVT::v8f32 && Subtarget->hasAVX()))
13677 RecipOp = "vec-sqrtf";
13681 TargetRecip Recips = DCI.DAG.getTarget().Options.Reciprocals;
13682 if (!Recips.isEnabled(RecipOp))
13685 RefinementSteps = Recips.getRefinementSteps(RecipOp);
13686 UseOneConstNR = false;
13687 return DCI.DAG.getNode(X86ISD::FRSQRT, SDLoc(Op), VT, Op);
13690 /// The minimum architected relative accuracy is 2^-12. We need one
13691 /// Newton-Raphson step to have a good float result (24 bits of precision).
13692 SDValue X86TargetLowering::getRecipEstimate(SDValue Op,
13693 DAGCombinerInfo &DCI,
13694 unsigned &RefinementSteps) const {
13695 EVT VT = Op.getValueType();
13696 const char *RecipOp;
13698 // SSE1 has rcpss and rcpps. AVX adds a 256-bit variant for rcpps.
13699 // TODO: Add support for AVX512 (v16f32).
13700 // It is likely not profitable to do this for f64 because a double-precision
13701 // reciprocal estimate with refinement on x86 prior to FMA requires
13702 // 15 instructions: convert to single, rcpss, convert back to double, refine
13703 // (3 steps = 12 insts). If an 'rcpsd' variant was added to the ISA
13704 // along with FMA, this could be a throughput win.
13705 if (VT == MVT::f32 && Subtarget->hasSSE1())
13707 else if ((VT == MVT::v4f32 && Subtarget->hasSSE1()) ||
13708 (VT == MVT::v8f32 && Subtarget->hasAVX()))
13709 RecipOp = "vec-divf";
13713 TargetRecip Recips = DCI.DAG.getTarget().Options.Reciprocals;
13714 if (!Recips.isEnabled(RecipOp))
13717 RefinementSteps = Recips.getRefinementSteps(RecipOp);
13718 return DCI.DAG.getNode(X86ISD::FRCP, SDLoc(Op), VT, Op);
13721 /// If we have at least two divisions that use the same divisor, convert to
13722 /// multplication by a reciprocal. This may need to be adjusted for a given
13723 /// CPU if a division's cost is not at least twice the cost of a multiplication.
13724 /// This is because we still need one division to calculate the reciprocal and
13725 /// then we need two multiplies by that reciprocal as replacements for the
13726 /// original divisions.
13727 unsigned X86TargetLowering::combineRepeatedFPDivisors() const {
13731 static bool isAllOnes(SDValue V) {
13732 ConstantSDNode *C = dyn_cast<ConstantSDNode>(V);
13733 return C && C->isAllOnesValue();
13736 /// LowerToBT - Result of 'and' is compared against zero. Turn it into a BT node
13737 /// if it's possible.
13738 SDValue X86TargetLowering::LowerToBT(SDValue And, ISD::CondCode CC,
13739 SDLoc dl, SelectionDAG &DAG) const {
13740 SDValue Op0 = And.getOperand(0);
13741 SDValue Op1 = And.getOperand(1);
13742 if (Op0.getOpcode() == ISD::TRUNCATE)
13743 Op0 = Op0.getOperand(0);
13744 if (Op1.getOpcode() == ISD::TRUNCATE)
13745 Op1 = Op1.getOperand(0);
13748 if (Op1.getOpcode() == ISD::SHL)
13749 std::swap(Op0, Op1);
13750 if (Op0.getOpcode() == ISD::SHL) {
13751 if (ConstantSDNode *And00C = dyn_cast<ConstantSDNode>(Op0.getOperand(0)))
13752 if (And00C->getZExtValue() == 1) {
13753 // If we looked past a truncate, check that it's only truncating away
13755 unsigned BitWidth = Op0.getValueSizeInBits();
13756 unsigned AndBitWidth = And.getValueSizeInBits();
13757 if (BitWidth > AndBitWidth) {
13759 DAG.computeKnownBits(Op0, Zeros, Ones);
13760 if (Zeros.countLeadingOnes() < BitWidth - AndBitWidth)
13764 RHS = Op0.getOperand(1);
13766 } else if (Op1.getOpcode() == ISD::Constant) {
13767 ConstantSDNode *AndRHS = cast<ConstantSDNode>(Op1);
13768 uint64_t AndRHSVal = AndRHS->getZExtValue();
13769 SDValue AndLHS = Op0;
13771 if (AndRHSVal == 1 && AndLHS.getOpcode() == ISD::SRL) {
13772 LHS = AndLHS.getOperand(0);
13773 RHS = AndLHS.getOperand(1);
13776 // Use BT if the immediate can't be encoded in a TEST instruction.
13777 if (!isUInt<32>(AndRHSVal) && isPowerOf2_64(AndRHSVal)) {
13779 RHS = DAG.getConstant(Log2_64_Ceil(AndRHSVal), dl, LHS.getValueType());
13783 if (LHS.getNode()) {
13784 // If LHS is i8, promote it to i32 with any_extend. There is no i8 BT
13785 // instruction. Since the shift amount is in-range-or-undefined, we know
13786 // that doing a bittest on the i32 value is ok. We extend to i32 because
13787 // the encoding for the i16 version is larger than the i32 version.
13788 // Also promote i16 to i32 for performance / code size reason.
13789 if (LHS.getValueType() == MVT::i8 ||
13790 LHS.getValueType() == MVT::i16)
13791 LHS = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, LHS);
13793 // If the operand types disagree, extend the shift amount to match. Since
13794 // BT ignores high bits (like shifts) we can use anyextend.
13795 if (LHS.getValueType() != RHS.getValueType())
13796 RHS = DAG.getNode(ISD::ANY_EXTEND, dl, LHS.getValueType(), RHS);
13798 SDValue BT = DAG.getNode(X86ISD::BT, dl, MVT::i32, LHS, RHS);
13799 X86::CondCode Cond = CC == ISD::SETEQ ? X86::COND_AE : X86::COND_B;
13800 return DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
13801 DAG.getConstant(Cond, dl, MVT::i8), BT);
13807 /// \brief - Turns an ISD::CondCode into a value suitable for SSE floating point
13809 static int translateX86FSETCC(ISD::CondCode SetCCOpcode, SDValue &Op0,
13814 // SSE Condition code mapping:
13823 switch (SetCCOpcode) {
13824 default: llvm_unreachable("Unexpected SETCC condition");
13826 case ISD::SETEQ: SSECC = 0; break;
13828 case ISD::SETGT: Swap = true; // Fallthrough
13830 case ISD::SETOLT: SSECC = 1; break;
13832 case ISD::SETGE: Swap = true; // Fallthrough
13834 case ISD::SETOLE: SSECC = 2; break;
13835 case ISD::SETUO: SSECC = 3; break;
13837 case ISD::SETNE: SSECC = 4; break;
13838 case ISD::SETULE: Swap = true; // Fallthrough
13839 case ISD::SETUGE: SSECC = 5; break;
13840 case ISD::SETULT: Swap = true; // Fallthrough
13841 case ISD::SETUGT: SSECC = 6; break;
13842 case ISD::SETO: SSECC = 7; break;
13844 case ISD::SETONE: SSECC = 8; break;
13847 std::swap(Op0, Op1);
13852 // Lower256IntVSETCC - Break a VSETCC 256-bit integer VSETCC into two new 128
13853 // ones, and then concatenate the result back.
13854 static SDValue Lower256IntVSETCC(SDValue Op, SelectionDAG &DAG) {
13855 MVT VT = Op.getSimpleValueType();
13857 assert(VT.is256BitVector() && Op.getOpcode() == ISD::SETCC &&
13858 "Unsupported value type for operation");
13860 unsigned NumElems = VT.getVectorNumElements();
13862 SDValue CC = Op.getOperand(2);
13864 // Extract the LHS vectors
13865 SDValue LHS = Op.getOperand(0);
13866 SDValue LHS1 = Extract128BitVector(LHS, 0, DAG, dl);
13867 SDValue LHS2 = Extract128BitVector(LHS, NumElems/2, DAG, dl);
13869 // Extract the RHS vectors
13870 SDValue RHS = Op.getOperand(1);
13871 SDValue RHS1 = Extract128BitVector(RHS, 0, DAG, dl);
13872 SDValue RHS2 = Extract128BitVector(RHS, NumElems/2, DAG, dl);
13874 // Issue the operation on the smaller types and concatenate the result back
13875 MVT EltVT = VT.getVectorElementType();
13876 MVT NewVT = MVT::getVectorVT(EltVT, NumElems/2);
13877 return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT,
13878 DAG.getNode(Op.getOpcode(), dl, NewVT, LHS1, RHS1, CC),
13879 DAG.getNode(Op.getOpcode(), dl, NewVT, LHS2, RHS2, CC));
13882 static SDValue LowerBoolVSETCC_AVX512(SDValue Op, SelectionDAG &DAG) {
13883 SDValue Op0 = Op.getOperand(0);
13884 SDValue Op1 = Op.getOperand(1);
13885 SDValue CC = Op.getOperand(2);
13886 MVT VT = Op.getSimpleValueType();
13889 assert(Op0.getValueType().getVectorElementType() == MVT::i1 &&
13890 "Unexpected type for boolean compare operation");
13891 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
13892 SDValue NotOp0 = DAG.getNode(ISD::XOR, dl, VT, Op0,
13893 DAG.getConstant(-1, dl, VT));
13894 SDValue NotOp1 = DAG.getNode(ISD::XOR, dl, VT, Op1,
13895 DAG.getConstant(-1, dl, VT));
13896 switch (SetCCOpcode) {
13897 default: llvm_unreachable("Unexpected SETCC condition");
13899 // (x == y) -> ~(x ^ y)
13900 return DAG.getNode(ISD::XOR, dl, VT,
13901 DAG.getNode(ISD::XOR, dl, VT, Op0, Op1),
13902 DAG.getConstant(-1, dl, VT));
13904 // (x != y) -> (x ^ y)
13905 return DAG.getNode(ISD::XOR, dl, VT, Op0, Op1);
13908 // (x > y) -> (x & ~y)
13909 return DAG.getNode(ISD::AND, dl, VT, Op0, NotOp1);
13912 // (x < y) -> (~x & y)
13913 return DAG.getNode(ISD::AND, dl, VT, NotOp0, Op1);
13916 // (x <= y) -> (~x | y)
13917 return DAG.getNode(ISD::OR, dl, VT, NotOp0, Op1);
13920 // (x >=y) -> (x | ~y)
13921 return DAG.getNode(ISD::OR, dl, VT, Op0, NotOp1);
13925 static SDValue LowerIntVSETCC_AVX512(SDValue Op, SelectionDAG &DAG,
13926 const X86Subtarget *Subtarget) {
13927 SDValue Op0 = Op.getOperand(0);
13928 SDValue Op1 = Op.getOperand(1);
13929 SDValue CC = Op.getOperand(2);
13930 MVT VT = Op.getSimpleValueType();
13933 assert(Op0.getValueType().getVectorElementType().getSizeInBits() >= 8 &&
13934 Op.getValueType().getScalarType() == MVT::i1 &&
13935 "Cannot set masked compare for this operation");
13937 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
13939 bool Unsigned = false;
13942 switch (SetCCOpcode) {
13943 default: llvm_unreachable("Unexpected SETCC condition");
13944 case ISD::SETNE: SSECC = 4; break;
13945 case ISD::SETEQ: Opc = X86ISD::PCMPEQM; break;
13946 case ISD::SETUGT: SSECC = 6; Unsigned = true; break;
13947 case ISD::SETLT: Swap = true; //fall-through
13948 case ISD::SETGT: Opc = X86ISD::PCMPGTM; break;
13949 case ISD::SETULT: SSECC = 1; Unsigned = true; break;
13950 case ISD::SETUGE: SSECC = 5; Unsigned = true; break; //NLT
13951 case ISD::SETGE: Swap = true; SSECC = 2; break; // LE + swap
13952 case ISD::SETULE: Unsigned = true; //fall-through
13953 case ISD::SETLE: SSECC = 2; break;
13957 std::swap(Op0, Op1);
13959 return DAG.getNode(Opc, dl, VT, Op0, Op1);
13960 Opc = Unsigned ? X86ISD::CMPMU: X86ISD::CMPM;
13961 return DAG.getNode(Opc, dl, VT, Op0, Op1,
13962 DAG.getConstant(SSECC, dl, MVT::i8));
13965 /// \brief Try to turn a VSETULT into a VSETULE by modifying its second
13966 /// operand \p Op1. If non-trivial (for example because it's not constant)
13967 /// return an empty value.
13968 static SDValue ChangeVSETULTtoVSETULE(SDLoc dl, SDValue Op1, SelectionDAG &DAG)
13970 BuildVectorSDNode *BV = dyn_cast<BuildVectorSDNode>(Op1.getNode());
13974 MVT VT = Op1.getSimpleValueType();
13975 MVT EVT = VT.getVectorElementType();
13976 unsigned n = VT.getVectorNumElements();
13977 SmallVector<SDValue, 8> ULTOp1;
13979 for (unsigned i = 0; i < n; ++i) {
13980 ConstantSDNode *Elt = dyn_cast<ConstantSDNode>(BV->getOperand(i));
13981 if (!Elt || Elt->isOpaque() || Elt->getValueType(0) != EVT)
13984 // Avoid underflow.
13985 APInt Val = Elt->getAPIntValue();
13989 ULTOp1.push_back(DAG.getConstant(Val - 1, dl, EVT));
13992 return DAG.getNode(ISD::BUILD_VECTOR, dl, VT, ULTOp1);
13995 static SDValue LowerVSETCC(SDValue Op, const X86Subtarget *Subtarget,
13996 SelectionDAG &DAG) {
13997 SDValue Op0 = Op.getOperand(0);
13998 SDValue Op1 = Op.getOperand(1);
13999 SDValue CC = Op.getOperand(2);
14000 MVT VT = Op.getSimpleValueType();
14001 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
14002 bool isFP = Op.getOperand(1).getSimpleValueType().isFloatingPoint();
14007 MVT EltVT = Op0.getSimpleValueType().getVectorElementType();
14008 assert(EltVT == MVT::f32 || EltVT == MVT::f64);
14011 unsigned SSECC = translateX86FSETCC(SetCCOpcode, Op0, Op1);
14012 unsigned Opc = X86ISD::CMPP;
14013 if (Subtarget->hasAVX512() && VT.getVectorElementType() == MVT::i1) {
14014 assert(VT.getVectorNumElements() <= 16);
14015 Opc = X86ISD::CMPM;
14017 // In the two special cases we can't handle, emit two comparisons.
14020 unsigned CombineOpc;
14021 if (SetCCOpcode == ISD::SETUEQ) {
14022 CC0 = 3; CC1 = 0; CombineOpc = ISD::OR;
14024 assert(SetCCOpcode == ISD::SETONE);
14025 CC0 = 7; CC1 = 4; CombineOpc = ISD::AND;
14028 SDValue Cmp0 = DAG.getNode(Opc, dl, VT, Op0, Op1,
14029 DAG.getConstant(CC0, dl, MVT::i8));
14030 SDValue Cmp1 = DAG.getNode(Opc, dl, VT, Op0, Op1,
14031 DAG.getConstant(CC1, dl, MVT::i8));
14032 return DAG.getNode(CombineOpc, dl, VT, Cmp0, Cmp1);
14034 // Handle all other FP comparisons here.
14035 return DAG.getNode(Opc, dl, VT, Op0, Op1,
14036 DAG.getConstant(SSECC, dl, MVT::i8));
14039 // Break 256-bit integer vector compare into smaller ones.
14040 if (VT.is256BitVector() && !Subtarget->hasInt256())
14041 return Lower256IntVSETCC(Op, DAG);
14043 EVT OpVT = Op1.getValueType();
14044 if (OpVT.getVectorElementType() == MVT::i1)
14045 return LowerBoolVSETCC_AVX512(Op, DAG);
14047 bool MaskResult = (VT.getVectorElementType() == MVT::i1);
14048 if (Subtarget->hasAVX512()) {
14049 if (Op1.getValueType().is512BitVector() ||
14050 (Subtarget->hasBWI() && Subtarget->hasVLX()) ||
14051 (MaskResult && OpVT.getVectorElementType().getSizeInBits() >= 32))
14052 return LowerIntVSETCC_AVX512(Op, DAG, Subtarget);
14054 // In AVX-512 architecture setcc returns mask with i1 elements,
14055 // But there is no compare instruction for i8 and i16 elements in KNL.
14056 // We are not talking about 512-bit operands in this case, these
14057 // types are illegal.
14059 (OpVT.getVectorElementType().getSizeInBits() < 32 &&
14060 OpVT.getVectorElementType().getSizeInBits() >= 8))
14061 return DAG.getNode(ISD::TRUNCATE, dl, VT,
14062 DAG.getNode(ISD::SETCC, dl, OpVT, Op0, Op1, CC));
14065 // We are handling one of the integer comparisons here. Since SSE only has
14066 // GT and EQ comparisons for integer, swapping operands and multiple
14067 // operations may be required for some comparisons.
14069 bool Swap = false, Invert = false, FlipSigns = false, MinMax = false;
14070 bool Subus = false;
14072 switch (SetCCOpcode) {
14073 default: llvm_unreachable("Unexpected SETCC condition");
14074 case ISD::SETNE: Invert = true;
14075 case ISD::SETEQ: Opc = X86ISD::PCMPEQ; break;
14076 case ISD::SETLT: Swap = true;
14077 case ISD::SETGT: Opc = X86ISD::PCMPGT; break;
14078 case ISD::SETGE: Swap = true;
14079 case ISD::SETLE: Opc = X86ISD::PCMPGT;
14080 Invert = true; break;
14081 case ISD::SETULT: Swap = true;
14082 case ISD::SETUGT: Opc = X86ISD::PCMPGT;
14083 FlipSigns = true; break;
14084 case ISD::SETUGE: Swap = true;
14085 case ISD::SETULE: Opc = X86ISD::PCMPGT;
14086 FlipSigns = true; Invert = true; break;
14089 // Special case: Use min/max operations for SETULE/SETUGE
14090 MVT VET = VT.getVectorElementType();
14092 (Subtarget->hasSSE41() && (VET >= MVT::i8 && VET <= MVT::i32))
14093 || (Subtarget->hasSSE2() && (VET == MVT::i8));
14096 switch (SetCCOpcode) {
14098 case ISD::SETULE: Opc = ISD::UMIN; MinMax = true; break;
14099 case ISD::SETUGE: Opc = ISD::UMAX; MinMax = true; break;
14102 if (MinMax) { Swap = false; Invert = false; FlipSigns = false; }
14105 bool hasSubus = Subtarget->hasSSE2() && (VET == MVT::i8 || VET == MVT::i16);
14106 if (!MinMax && hasSubus) {
14107 // As another special case, use PSUBUS[BW] when it's profitable. E.g. for
14109 // t = psubus Op0, Op1
14110 // pcmpeq t, <0..0>
14111 switch (SetCCOpcode) {
14113 case ISD::SETULT: {
14114 // If the comparison is against a constant we can turn this into a
14115 // setule. With psubus, setule does not require a swap. This is
14116 // beneficial because the constant in the register is no longer
14117 // destructed as the destination so it can be hoisted out of a loop.
14118 // Only do this pre-AVX since vpcmp* is no longer destructive.
14119 if (Subtarget->hasAVX())
14121 SDValue ULEOp1 = ChangeVSETULTtoVSETULE(dl, Op1, DAG);
14122 if (ULEOp1.getNode()) {
14124 Subus = true; Invert = false; Swap = false;
14128 // Psubus is better than flip-sign because it requires no inversion.
14129 case ISD::SETUGE: Subus = true; Invert = false; Swap = true; break;
14130 case ISD::SETULE: Subus = true; Invert = false; Swap = false; break;
14134 Opc = X86ISD::SUBUS;
14140 std::swap(Op0, Op1);
14142 // Check that the operation in question is available (most are plain SSE2,
14143 // but PCMPGTQ and PCMPEQQ have different requirements).
14144 if (VT == MVT::v2i64) {
14145 if (Opc == X86ISD::PCMPGT && !Subtarget->hasSSE42()) {
14146 assert(Subtarget->hasSSE2() && "Don't know how to lower!");
14148 // First cast everything to the right type.
14149 Op0 = DAG.getBitcast(MVT::v4i32, Op0);
14150 Op1 = DAG.getBitcast(MVT::v4i32, Op1);
14152 // Since SSE has no unsigned integer comparisons, we need to flip the sign
14153 // bits of the inputs before performing those operations. The lower
14154 // compare is always unsigned.
14157 SB = DAG.getConstant(0x80000000U, dl, MVT::v4i32);
14159 SDValue Sign = DAG.getConstant(0x80000000U, dl, MVT::i32);
14160 SDValue Zero = DAG.getConstant(0x00000000U, dl, MVT::i32);
14161 SB = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32,
14162 Sign, Zero, Sign, Zero);
14164 Op0 = DAG.getNode(ISD::XOR, dl, MVT::v4i32, Op0, SB);
14165 Op1 = DAG.getNode(ISD::XOR, dl, MVT::v4i32, Op1, SB);
14167 // Emulate PCMPGTQ with (hi1 > hi2) | ((hi1 == hi2) & (lo1 > lo2))
14168 SDValue GT = DAG.getNode(X86ISD::PCMPGT, dl, MVT::v4i32, Op0, Op1);
14169 SDValue EQ = DAG.getNode(X86ISD::PCMPEQ, dl, MVT::v4i32, Op0, Op1);
14171 // Create masks for only the low parts/high parts of the 64 bit integers.
14172 static const int MaskHi[] = { 1, 1, 3, 3 };
14173 static const int MaskLo[] = { 0, 0, 2, 2 };
14174 SDValue EQHi = DAG.getVectorShuffle(MVT::v4i32, dl, EQ, EQ, MaskHi);
14175 SDValue GTLo = DAG.getVectorShuffle(MVT::v4i32, dl, GT, GT, MaskLo);
14176 SDValue GTHi = DAG.getVectorShuffle(MVT::v4i32, dl, GT, GT, MaskHi);
14178 SDValue Result = DAG.getNode(ISD::AND, dl, MVT::v4i32, EQHi, GTLo);
14179 Result = DAG.getNode(ISD::OR, dl, MVT::v4i32, Result, GTHi);
14182 Result = DAG.getNOT(dl, Result, MVT::v4i32);
14184 return DAG.getBitcast(VT, Result);
14187 if (Opc == X86ISD::PCMPEQ && !Subtarget->hasSSE41()) {
14188 // If pcmpeqq is missing but pcmpeqd is available synthesize pcmpeqq with
14189 // pcmpeqd + pshufd + pand.
14190 assert(Subtarget->hasSSE2() && !FlipSigns && "Don't know how to lower!");
14192 // First cast everything to the right type.
14193 Op0 = DAG.getBitcast(MVT::v4i32, Op0);
14194 Op1 = DAG.getBitcast(MVT::v4i32, Op1);
14197 SDValue Result = DAG.getNode(Opc, dl, MVT::v4i32, Op0, Op1);
14199 // Make sure the lower and upper halves are both all-ones.
14200 static const int Mask[] = { 1, 0, 3, 2 };
14201 SDValue Shuf = DAG.getVectorShuffle(MVT::v4i32, dl, Result, Result, Mask);
14202 Result = DAG.getNode(ISD::AND, dl, MVT::v4i32, Result, Shuf);
14205 Result = DAG.getNOT(dl, Result, MVT::v4i32);
14207 return DAG.getBitcast(VT, Result);
14211 // Since SSE has no unsigned integer comparisons, we need to flip the sign
14212 // bits of the inputs before performing those operations.
14214 EVT EltVT = VT.getVectorElementType();
14215 SDValue SB = DAG.getConstant(APInt::getSignBit(EltVT.getSizeInBits()), dl,
14217 Op0 = DAG.getNode(ISD::XOR, dl, VT, Op0, SB);
14218 Op1 = DAG.getNode(ISD::XOR, dl, VT, Op1, SB);
14221 SDValue Result = DAG.getNode(Opc, dl, VT, Op0, Op1);
14223 // If the logical-not of the result is required, perform that now.
14225 Result = DAG.getNOT(dl, Result, VT);
14228 Result = DAG.getNode(X86ISD::PCMPEQ, dl, VT, Op0, Result);
14231 Result = DAG.getNode(X86ISD::PCMPEQ, dl, VT, Result,
14232 getZeroVector(VT, Subtarget, DAG, dl));
14237 SDValue X86TargetLowering::LowerSETCC(SDValue Op, SelectionDAG &DAG) const {
14239 MVT VT = Op.getSimpleValueType();
14241 if (VT.isVector()) return LowerVSETCC(Op, Subtarget, DAG);
14243 assert(((!Subtarget->hasAVX512() && VT == MVT::i8) || (VT == MVT::i1))
14244 && "SetCC type must be 8-bit or 1-bit integer");
14245 SDValue Op0 = Op.getOperand(0);
14246 SDValue Op1 = Op.getOperand(1);
14248 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
14250 // Optimize to BT if possible.
14251 // Lower (X & (1 << N)) == 0 to BT(X, N).
14252 // Lower ((X >>u N) & 1) != 0 to BT(X, N).
14253 // Lower ((X >>s N) & 1) != 0 to BT(X, N).
14254 if (Op0.getOpcode() == ISD::AND && Op0.hasOneUse() &&
14255 Op1.getOpcode() == ISD::Constant &&
14256 cast<ConstantSDNode>(Op1)->isNullValue() &&
14257 (CC == ISD::SETEQ || CC == ISD::SETNE)) {
14258 SDValue NewSetCC = LowerToBT(Op0, CC, dl, DAG);
14259 if (NewSetCC.getNode()) {
14261 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i1, NewSetCC);
14266 // Look for X == 0, X == 1, X != 0, or X != 1. We can simplify some forms of
14268 if (Op1.getOpcode() == ISD::Constant &&
14269 (cast<ConstantSDNode>(Op1)->getZExtValue() == 1 ||
14270 cast<ConstantSDNode>(Op1)->isNullValue()) &&
14271 (CC == ISD::SETEQ || CC == ISD::SETNE)) {
14273 // If the input is a setcc, then reuse the input setcc or use a new one with
14274 // the inverted condition.
14275 if (Op0.getOpcode() == X86ISD::SETCC) {
14276 X86::CondCode CCode = (X86::CondCode)Op0.getConstantOperandVal(0);
14277 bool Invert = (CC == ISD::SETNE) ^
14278 cast<ConstantSDNode>(Op1)->isNullValue();
14282 CCode = X86::GetOppositeBranchCondition(CCode);
14283 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
14284 DAG.getConstant(CCode, dl, MVT::i8),
14285 Op0.getOperand(1));
14287 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i1, SetCC);
14291 if ((Op0.getValueType() == MVT::i1) && (Op1.getOpcode() == ISD::Constant) &&
14292 (cast<ConstantSDNode>(Op1)->getZExtValue() == 1) &&
14293 (CC == ISD::SETEQ || CC == ISD::SETNE)) {
14295 ISD::CondCode NewCC = ISD::getSetCCInverse(CC, true);
14296 return DAG.getSetCC(dl, VT, Op0, DAG.getConstant(0, dl, MVT::i1), NewCC);
14299 bool isFP = Op1.getSimpleValueType().isFloatingPoint();
14300 unsigned X86CC = TranslateX86CC(CC, dl, isFP, Op0, Op1, DAG);
14301 if (X86CC == X86::COND_INVALID)
14304 SDValue EFLAGS = EmitCmp(Op0, Op1, X86CC, dl, DAG);
14305 EFLAGS = ConvertCmpIfNecessary(EFLAGS, DAG);
14306 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
14307 DAG.getConstant(X86CC, dl, MVT::i8), EFLAGS);
14309 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i1, SetCC);
14313 // isX86LogicalCmp - Return true if opcode is a X86 logical comparison.
14314 static bool isX86LogicalCmp(SDValue Op) {
14315 unsigned Opc = Op.getNode()->getOpcode();
14316 if (Opc == X86ISD::CMP || Opc == X86ISD::COMI || Opc == X86ISD::UCOMI ||
14317 Opc == X86ISD::SAHF)
14319 if (Op.getResNo() == 1 &&
14320 (Opc == X86ISD::ADD ||
14321 Opc == X86ISD::SUB ||
14322 Opc == X86ISD::ADC ||
14323 Opc == X86ISD::SBB ||
14324 Opc == X86ISD::SMUL ||
14325 Opc == X86ISD::UMUL ||
14326 Opc == X86ISD::INC ||
14327 Opc == X86ISD::DEC ||
14328 Opc == X86ISD::OR ||
14329 Opc == X86ISD::XOR ||
14330 Opc == X86ISD::AND))
14333 if (Op.getResNo() == 2 && Opc == X86ISD::UMUL)
14339 static bool isTruncWithZeroHighBitsInput(SDValue V, SelectionDAG &DAG) {
14340 if (V.getOpcode() != ISD::TRUNCATE)
14343 SDValue VOp0 = V.getOperand(0);
14344 unsigned InBits = VOp0.getValueSizeInBits();
14345 unsigned Bits = V.getValueSizeInBits();
14346 return DAG.MaskedValueIsZero(VOp0, APInt::getHighBitsSet(InBits,InBits-Bits));
14349 SDValue X86TargetLowering::LowerSELECT(SDValue Op, SelectionDAG &DAG) const {
14350 bool addTest = true;
14351 SDValue Cond = Op.getOperand(0);
14352 SDValue Op1 = Op.getOperand(1);
14353 SDValue Op2 = Op.getOperand(2);
14355 EVT VT = Op1.getValueType();
14358 // Lower FP selects into a CMP/AND/ANDN/OR sequence when the necessary SSE ops
14359 // are available or VBLENDV if AVX is available.
14360 // Otherwise FP cmovs get lowered into a less efficient branch sequence later.
14361 if (Cond.getOpcode() == ISD::SETCC &&
14362 ((Subtarget->hasSSE2() && (VT == MVT::f32 || VT == MVT::f64)) ||
14363 (Subtarget->hasSSE1() && VT == MVT::f32)) &&
14364 VT == Cond.getOperand(0).getValueType() && Cond->hasOneUse()) {
14365 SDValue CondOp0 = Cond.getOperand(0), CondOp1 = Cond.getOperand(1);
14366 int SSECC = translateX86FSETCC(
14367 cast<CondCodeSDNode>(Cond.getOperand(2))->get(), CondOp0, CondOp1);
14370 if (Subtarget->hasAVX512()) {
14371 SDValue Cmp = DAG.getNode(X86ISD::FSETCC, DL, MVT::i1, CondOp0, CondOp1,
14372 DAG.getConstant(SSECC, DL, MVT::i8));
14373 return DAG.getNode(X86ISD::SELECT, DL, VT, Cmp, Op1, Op2);
14376 SDValue Cmp = DAG.getNode(X86ISD::FSETCC, DL, VT, CondOp0, CondOp1,
14377 DAG.getConstant(SSECC, DL, MVT::i8));
14379 // If we have AVX, we can use a variable vector select (VBLENDV) instead
14380 // of 3 logic instructions for size savings and potentially speed.
14381 // Unfortunately, there is no scalar form of VBLENDV.
14383 // If either operand is a constant, don't try this. We can expect to
14384 // optimize away at least one of the logic instructions later in that
14385 // case, so that sequence would be faster than a variable blend.
14387 // BLENDV was introduced with SSE 4.1, but the 2 register form implicitly
14388 // uses XMM0 as the selection register. That may need just as many
14389 // instructions as the AND/ANDN/OR sequence due to register moves, so
14392 if (Subtarget->hasAVX() &&
14393 !isa<ConstantFPSDNode>(Op1) && !isa<ConstantFPSDNode>(Op2)) {
14395 // Convert to vectors, do a VSELECT, and convert back to scalar.
14396 // All of the conversions should be optimized away.
14398 EVT VecVT = VT == MVT::f32 ? MVT::v4f32 : MVT::v2f64;
14399 SDValue VOp1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, VecVT, Op1);
14400 SDValue VOp2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, VecVT, Op2);
14401 SDValue VCmp = DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, VecVT, Cmp);
14403 EVT VCmpVT = VT == MVT::f32 ? MVT::v4i32 : MVT::v2i64;
14404 VCmp = DAG.getBitcast(VCmpVT, VCmp);
14406 SDValue VSel = DAG.getNode(ISD::VSELECT, DL, VecVT, VCmp, VOp1, VOp2);
14408 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, VT,
14409 VSel, DAG.getIntPtrConstant(0, DL));
14411 SDValue AndN = DAG.getNode(X86ISD::FANDN, DL, VT, Cmp, Op2);
14412 SDValue And = DAG.getNode(X86ISD::FAND, DL, VT, Cmp, Op1);
14413 return DAG.getNode(X86ISD::FOR, DL, VT, AndN, And);
14417 if (VT.isVector() && VT.getScalarType() == MVT::i1) {
14419 if (ISD::isBuildVectorOfConstantSDNodes(Op1.getNode()))
14420 Op1Scalar = ConvertI1VectorToInteger(Op1, DAG);
14421 else if (Op1.getOpcode() == ISD::BITCAST && Op1.getOperand(0))
14422 Op1Scalar = Op1.getOperand(0);
14424 if (ISD::isBuildVectorOfConstantSDNodes(Op2.getNode()))
14425 Op2Scalar = ConvertI1VectorToInteger(Op2, DAG);
14426 else if (Op2.getOpcode() == ISD::BITCAST && Op2.getOperand(0))
14427 Op2Scalar = Op2.getOperand(0);
14428 if (Op1Scalar.getNode() && Op2Scalar.getNode()) {
14429 SDValue newSelect = DAG.getNode(ISD::SELECT, DL,
14430 Op1Scalar.getValueType(),
14431 Cond, Op1Scalar, Op2Scalar);
14432 if (newSelect.getValueSizeInBits() == VT.getSizeInBits())
14433 return DAG.getBitcast(VT, newSelect);
14434 SDValue ExtVec = DAG.getBitcast(MVT::v8i1, newSelect);
14435 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, VT, ExtVec,
14436 DAG.getIntPtrConstant(0, DL));
14440 if (VT == MVT::v4i1 || VT == MVT::v2i1) {
14441 SDValue zeroConst = DAG.getIntPtrConstant(0, DL);
14442 Op1 = DAG.getNode(ISD::INSERT_SUBVECTOR, DL, MVT::v8i1,
14443 DAG.getUNDEF(MVT::v8i1), Op1, zeroConst);
14444 Op2 = DAG.getNode(ISD::INSERT_SUBVECTOR, DL, MVT::v8i1,
14445 DAG.getUNDEF(MVT::v8i1), Op2, zeroConst);
14446 SDValue newSelect = DAG.getNode(ISD::SELECT, DL, MVT::v8i1,
14448 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, VT, newSelect, zeroConst);
14451 if (Cond.getOpcode() == ISD::SETCC) {
14452 SDValue NewCond = LowerSETCC(Cond, DAG);
14453 if (NewCond.getNode())
14457 // (select (x == 0), -1, y) -> (sign_bit (x - 1)) | y
14458 // (select (x == 0), y, -1) -> ~(sign_bit (x - 1)) | y
14459 // (select (x != 0), y, -1) -> (sign_bit (x - 1)) | y
14460 // (select (x != 0), -1, y) -> ~(sign_bit (x - 1)) | y
14461 if (Cond.getOpcode() == X86ISD::SETCC &&
14462 Cond.getOperand(1).getOpcode() == X86ISD::CMP &&
14463 isZero(Cond.getOperand(1).getOperand(1))) {
14464 SDValue Cmp = Cond.getOperand(1);
14466 unsigned CondCode =cast<ConstantSDNode>(Cond.getOperand(0))->getZExtValue();
14468 if ((isAllOnes(Op1) || isAllOnes(Op2)) &&
14469 (CondCode == X86::COND_E || CondCode == X86::COND_NE)) {
14470 SDValue Y = isAllOnes(Op2) ? Op1 : Op2;
14472 SDValue CmpOp0 = Cmp.getOperand(0);
14473 // Apply further optimizations for special cases
14474 // (select (x != 0), -1, 0) -> neg & sbb
14475 // (select (x == 0), 0, -1) -> neg & sbb
14476 if (ConstantSDNode *YC = dyn_cast<ConstantSDNode>(Y))
14477 if (YC->isNullValue() &&
14478 (isAllOnes(Op1) == (CondCode == X86::COND_NE))) {
14479 SDVTList VTs = DAG.getVTList(CmpOp0.getValueType(), MVT::i32);
14480 SDValue Neg = DAG.getNode(X86ISD::SUB, DL, VTs,
14481 DAG.getConstant(0, DL,
14482 CmpOp0.getValueType()),
14484 SDValue Res = DAG.getNode(X86ISD::SETCC_CARRY, DL, Op.getValueType(),
14485 DAG.getConstant(X86::COND_B, DL, MVT::i8),
14486 SDValue(Neg.getNode(), 1));
14490 Cmp = DAG.getNode(X86ISD::CMP, DL, MVT::i32,
14491 CmpOp0, DAG.getConstant(1, DL, CmpOp0.getValueType()));
14492 Cmp = ConvertCmpIfNecessary(Cmp, DAG);
14494 SDValue Res = // Res = 0 or -1.
14495 DAG.getNode(X86ISD::SETCC_CARRY, DL, Op.getValueType(),
14496 DAG.getConstant(X86::COND_B, DL, MVT::i8), Cmp);
14498 if (isAllOnes(Op1) != (CondCode == X86::COND_E))
14499 Res = DAG.getNOT(DL, Res, Res.getValueType());
14501 ConstantSDNode *N2C = dyn_cast<ConstantSDNode>(Op2);
14502 if (!N2C || !N2C->isNullValue())
14503 Res = DAG.getNode(ISD::OR, DL, Res.getValueType(), Res, Y);
14508 // Look past (and (setcc_carry (cmp ...)), 1).
14509 if (Cond.getOpcode() == ISD::AND &&
14510 Cond.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY) {
14511 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Cond.getOperand(1));
14512 if (C && C->getAPIntValue() == 1)
14513 Cond = Cond.getOperand(0);
14516 // If condition flag is set by a X86ISD::CMP, then use it as the condition
14517 // setting operand in place of the X86ISD::SETCC.
14518 unsigned CondOpcode = Cond.getOpcode();
14519 if (CondOpcode == X86ISD::SETCC ||
14520 CondOpcode == X86ISD::SETCC_CARRY) {
14521 CC = Cond.getOperand(0);
14523 SDValue Cmp = Cond.getOperand(1);
14524 unsigned Opc = Cmp.getOpcode();
14525 MVT VT = Op.getSimpleValueType();
14527 bool IllegalFPCMov = false;
14528 if (VT.isFloatingPoint() && !VT.isVector() &&
14529 !isScalarFPTypeInSSEReg(VT)) // FPStack?
14530 IllegalFPCMov = !hasFPCMov(cast<ConstantSDNode>(CC)->getSExtValue());
14532 if ((isX86LogicalCmp(Cmp) && !IllegalFPCMov) ||
14533 Opc == X86ISD::BT) { // FIXME
14537 } else if (CondOpcode == ISD::USUBO || CondOpcode == ISD::SSUBO ||
14538 CondOpcode == ISD::UADDO || CondOpcode == ISD::SADDO ||
14539 ((CondOpcode == ISD::UMULO || CondOpcode == ISD::SMULO) &&
14540 Cond.getOperand(0).getValueType() != MVT::i8)) {
14541 SDValue LHS = Cond.getOperand(0);
14542 SDValue RHS = Cond.getOperand(1);
14543 unsigned X86Opcode;
14546 switch (CondOpcode) {
14547 case ISD::UADDO: X86Opcode = X86ISD::ADD; X86Cond = X86::COND_B; break;
14548 case ISD::SADDO: X86Opcode = X86ISD::ADD; X86Cond = X86::COND_O; break;
14549 case ISD::USUBO: X86Opcode = X86ISD::SUB; X86Cond = X86::COND_B; break;
14550 case ISD::SSUBO: X86Opcode = X86ISD::SUB; X86Cond = X86::COND_O; break;
14551 case ISD::UMULO: X86Opcode = X86ISD::UMUL; X86Cond = X86::COND_O; break;
14552 case ISD::SMULO: X86Opcode = X86ISD::SMUL; X86Cond = X86::COND_O; break;
14553 default: llvm_unreachable("unexpected overflowing operator");
14555 if (CondOpcode == ISD::UMULO)
14556 VTs = DAG.getVTList(LHS.getValueType(), LHS.getValueType(),
14559 VTs = DAG.getVTList(LHS.getValueType(), MVT::i32);
14561 SDValue X86Op = DAG.getNode(X86Opcode, DL, VTs, LHS, RHS);
14563 if (CondOpcode == ISD::UMULO)
14564 Cond = X86Op.getValue(2);
14566 Cond = X86Op.getValue(1);
14568 CC = DAG.getConstant(X86Cond, DL, MVT::i8);
14573 // Look past the truncate if the high bits are known zero.
14574 if (isTruncWithZeroHighBitsInput(Cond, DAG))
14575 Cond = Cond.getOperand(0);
14577 // We know the result of AND is compared against zero. Try to match
14579 if (Cond.getOpcode() == ISD::AND && Cond.hasOneUse()) {
14580 SDValue NewSetCC = LowerToBT(Cond, ISD::SETNE, DL, DAG);
14581 if (NewSetCC.getNode()) {
14582 CC = NewSetCC.getOperand(0);
14583 Cond = NewSetCC.getOperand(1);
14590 CC = DAG.getConstant(X86::COND_NE, DL, MVT::i8);
14591 Cond = EmitTest(Cond, X86::COND_NE, DL, DAG);
14594 // a < b ? -1 : 0 -> RES = ~setcc_carry
14595 // a < b ? 0 : -1 -> RES = setcc_carry
14596 // a >= b ? -1 : 0 -> RES = setcc_carry
14597 // a >= b ? 0 : -1 -> RES = ~setcc_carry
14598 if (Cond.getOpcode() == X86ISD::SUB) {
14599 Cond = ConvertCmpIfNecessary(Cond, DAG);
14600 unsigned CondCode = cast<ConstantSDNode>(CC)->getZExtValue();
14602 if ((CondCode == X86::COND_AE || CondCode == X86::COND_B) &&
14603 (isAllOnes(Op1) || isAllOnes(Op2)) && (isZero(Op1) || isZero(Op2))) {
14604 SDValue Res = DAG.getNode(X86ISD::SETCC_CARRY, DL, Op.getValueType(),
14605 DAG.getConstant(X86::COND_B, DL, MVT::i8),
14607 if (isAllOnes(Op1) != (CondCode == X86::COND_B))
14608 return DAG.getNOT(DL, Res, Res.getValueType());
14613 // X86 doesn't have an i8 cmov. If both operands are the result of a truncate
14614 // widen the cmov and push the truncate through. This avoids introducing a new
14615 // branch during isel and doesn't add any extensions.
14616 if (Op.getValueType() == MVT::i8 &&
14617 Op1.getOpcode() == ISD::TRUNCATE && Op2.getOpcode() == ISD::TRUNCATE) {
14618 SDValue T1 = Op1.getOperand(0), T2 = Op2.getOperand(0);
14619 if (T1.getValueType() == T2.getValueType() &&
14620 // Blacklist CopyFromReg to avoid partial register stalls.
14621 T1.getOpcode() != ISD::CopyFromReg && T2.getOpcode()!=ISD::CopyFromReg){
14622 SDVTList VTs = DAG.getVTList(T1.getValueType(), MVT::Glue);
14623 SDValue Cmov = DAG.getNode(X86ISD::CMOV, DL, VTs, T2, T1, CC, Cond);
14624 return DAG.getNode(ISD::TRUNCATE, DL, Op.getValueType(), Cmov);
14628 // X86ISD::CMOV means set the result (which is operand 1) to the RHS if
14629 // condition is true.
14630 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::Glue);
14631 SDValue Ops[] = { Op2, Op1, CC, Cond };
14632 return DAG.getNode(X86ISD::CMOV, DL, VTs, Ops);
14635 static SDValue LowerSIGN_EXTEND_AVX512(SDValue Op,
14636 const X86Subtarget *Subtarget,
14637 SelectionDAG &DAG) {
14638 MVT VT = Op->getSimpleValueType(0);
14639 SDValue In = Op->getOperand(0);
14640 MVT InVT = In.getSimpleValueType();
14641 MVT VTElt = VT.getVectorElementType();
14642 MVT InVTElt = InVT.getVectorElementType();
14646 if ((InVTElt == MVT::i1) &&
14647 (((Subtarget->hasBWI() && Subtarget->hasVLX() &&
14648 VT.getSizeInBits() <= 256 && VTElt.getSizeInBits() <= 16)) ||
14650 ((Subtarget->hasBWI() && VT.is512BitVector() &&
14651 VTElt.getSizeInBits() <= 16)) ||
14653 ((Subtarget->hasDQI() && Subtarget->hasVLX() &&
14654 VT.getSizeInBits() <= 256 && VTElt.getSizeInBits() >= 32)) ||
14656 ((Subtarget->hasDQI() && VT.is512BitVector() &&
14657 VTElt.getSizeInBits() >= 32))))
14658 return DAG.getNode(X86ISD::VSEXT, dl, VT, In);
14660 unsigned int NumElts = VT.getVectorNumElements();
14662 if (NumElts != 8 && NumElts != 16 && !Subtarget->hasBWI())
14665 if (VT.is512BitVector() && InVT.getVectorElementType() != MVT::i1) {
14666 if (In.getOpcode() == X86ISD::VSEXT || In.getOpcode() == X86ISD::VZEXT)
14667 return DAG.getNode(In.getOpcode(), dl, VT, In.getOperand(0));
14668 return DAG.getNode(X86ISD::VSEXT, dl, VT, In);
14671 assert (InVT.getVectorElementType() == MVT::i1 && "Unexpected vector type");
14672 MVT ExtVT = NumElts == 8 ? MVT::v8i64 : MVT::v16i32;
14674 DAG.getConstant(APInt::getAllOnesValue(ExtVT.getScalarSizeInBits()), dl,
14677 DAG.getConstant(APInt::getNullValue(ExtVT.getScalarSizeInBits()), dl, ExtVT);
14679 SDValue V = DAG.getNode(ISD::VSELECT, dl, ExtVT, In, NegOne, Zero);
14680 if (VT.is512BitVector())
14682 return DAG.getNode(X86ISD::VTRUNC, dl, VT, V);
14685 static SDValue LowerSIGN_EXTEND_VECTOR_INREG(SDValue Op,
14686 const X86Subtarget *Subtarget,
14687 SelectionDAG &DAG) {
14688 SDValue In = Op->getOperand(0);
14689 MVT VT = Op->getSimpleValueType(0);
14690 MVT InVT = In.getSimpleValueType();
14691 assert(VT.getSizeInBits() == InVT.getSizeInBits());
14693 MVT InSVT = InVT.getScalarType();
14694 assert(VT.getScalarType().getScalarSizeInBits() > InSVT.getScalarSizeInBits());
14696 if (VT != MVT::v2i64 && VT != MVT::v4i32 && VT != MVT::v8i16)
14698 if (InSVT != MVT::i32 && InSVT != MVT::i16 && InSVT != MVT::i8)
14703 // SSE41 targets can use the pmovsx* instructions directly.
14704 if (Subtarget->hasSSE41())
14705 return DAG.getNode(X86ISD::VSEXT, dl, VT, In);
14707 // pre-SSE41 targets unpack lower lanes and then sign-extend using SRAI.
14711 // As SRAI is only available on i16/i32 types, we expand only up to i32
14712 // and handle i64 separately.
14713 while (CurrVT != VT && CurrVT.getScalarType() != MVT::i32) {
14714 Curr = DAG.getNode(X86ISD::UNPCKL, dl, CurrVT, DAG.getUNDEF(CurrVT), Curr);
14715 MVT CurrSVT = MVT::getIntegerVT(CurrVT.getScalarSizeInBits() * 2);
14716 CurrVT = MVT::getVectorVT(CurrSVT, CurrVT.getVectorNumElements() / 2);
14717 Curr = DAG.getBitcast(CurrVT, Curr);
14720 SDValue SignExt = Curr;
14721 if (CurrVT != InVT) {
14722 unsigned SignExtShift =
14723 CurrVT.getScalarSizeInBits() - InSVT.getScalarSizeInBits();
14724 SignExt = DAG.getNode(X86ISD::VSRAI, dl, CurrVT, Curr,
14725 DAG.getConstant(SignExtShift, dl, MVT::i8));
14731 if (VT == MVT::v2i64 && CurrVT == MVT::v4i32) {
14732 SDValue Sign = DAG.getNode(X86ISD::VSRAI, dl, CurrVT, Curr,
14733 DAG.getConstant(31, dl, MVT::i8));
14734 SDValue Ext = DAG.getVectorShuffle(CurrVT, dl, SignExt, Sign, {0, 4, 1, 5});
14735 return DAG.getBitcast(VT, Ext);
14741 static SDValue LowerSIGN_EXTEND(SDValue Op, const X86Subtarget *Subtarget,
14742 SelectionDAG &DAG) {
14743 MVT VT = Op->getSimpleValueType(0);
14744 SDValue In = Op->getOperand(0);
14745 MVT InVT = In.getSimpleValueType();
14748 if (VT.is512BitVector() || InVT.getVectorElementType() == MVT::i1)
14749 return LowerSIGN_EXTEND_AVX512(Op, Subtarget, DAG);
14751 if ((VT != MVT::v4i64 || InVT != MVT::v4i32) &&
14752 (VT != MVT::v8i32 || InVT != MVT::v8i16) &&
14753 (VT != MVT::v16i16 || InVT != MVT::v16i8))
14756 if (Subtarget->hasInt256())
14757 return DAG.getNode(X86ISD::VSEXT, dl, VT, In);
14759 // Optimize vectors in AVX mode
14760 // Sign extend v8i16 to v8i32 and
14763 // Divide input vector into two parts
14764 // for v4i32 the shuffle mask will be { 0, 1, -1, -1} {2, 3, -1, -1}
14765 // use vpmovsx instruction to extend v4i32 -> v2i64; v8i16 -> v4i32
14766 // concat the vectors to original VT
14768 unsigned NumElems = InVT.getVectorNumElements();
14769 SDValue Undef = DAG.getUNDEF(InVT);
14771 SmallVector<int,8> ShufMask1(NumElems, -1);
14772 for (unsigned i = 0; i != NumElems/2; ++i)
14775 SDValue OpLo = DAG.getVectorShuffle(InVT, dl, In, Undef, &ShufMask1[0]);
14777 SmallVector<int,8> ShufMask2(NumElems, -1);
14778 for (unsigned i = 0; i != NumElems/2; ++i)
14779 ShufMask2[i] = i + NumElems/2;
14781 SDValue OpHi = DAG.getVectorShuffle(InVT, dl, In, Undef, &ShufMask2[0]);
14783 MVT HalfVT = MVT::getVectorVT(VT.getScalarType(),
14784 VT.getVectorNumElements()/2);
14786 OpLo = DAG.getNode(X86ISD::VSEXT, dl, HalfVT, OpLo);
14787 OpHi = DAG.getNode(X86ISD::VSEXT, dl, HalfVT, OpHi);
14789 return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT, OpLo, OpHi);
14792 // Lower vector extended loads using a shuffle. If SSSE3 is not available we
14793 // may emit an illegal shuffle but the expansion is still better than scalar
14794 // code. We generate X86ISD::VSEXT for SEXTLOADs if it's available, otherwise
14795 // we'll emit a shuffle and a arithmetic shift.
14796 // FIXME: Is the expansion actually better than scalar code? It doesn't seem so.
14797 // TODO: It is possible to support ZExt by zeroing the undef values during
14798 // the shuffle phase or after the shuffle.
14799 static SDValue LowerExtendedLoad(SDValue Op, const X86Subtarget *Subtarget,
14800 SelectionDAG &DAG) {
14801 MVT RegVT = Op.getSimpleValueType();
14802 assert(RegVT.isVector() && "We only custom lower vector sext loads.");
14803 assert(RegVT.isInteger() &&
14804 "We only custom lower integer vector sext loads.");
14806 // Nothing useful we can do without SSE2 shuffles.
14807 assert(Subtarget->hasSSE2() && "We only custom lower sext loads with SSE2.");
14809 LoadSDNode *Ld = cast<LoadSDNode>(Op.getNode());
14811 EVT MemVT = Ld->getMemoryVT();
14812 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
14813 unsigned RegSz = RegVT.getSizeInBits();
14815 ISD::LoadExtType Ext = Ld->getExtensionType();
14817 assert((Ext == ISD::EXTLOAD || Ext == ISD::SEXTLOAD)
14818 && "Only anyext and sext are currently implemented.");
14819 assert(MemVT != RegVT && "Cannot extend to the same type");
14820 assert(MemVT.isVector() && "Must load a vector from memory");
14822 unsigned NumElems = RegVT.getVectorNumElements();
14823 unsigned MemSz = MemVT.getSizeInBits();
14824 assert(RegSz > MemSz && "Register size must be greater than the mem size");
14826 if (Ext == ISD::SEXTLOAD && RegSz == 256 && !Subtarget->hasInt256()) {
14827 // The only way in which we have a legal 256-bit vector result but not the
14828 // integer 256-bit operations needed to directly lower a sextload is if we
14829 // have AVX1 but not AVX2. In that case, we can always emit a sextload to
14830 // a 128-bit vector and a normal sign_extend to 256-bits that should get
14831 // correctly legalized. We do this late to allow the canonical form of
14832 // sextload to persist throughout the rest of the DAG combiner -- it wants
14833 // to fold together any extensions it can, and so will fuse a sign_extend
14834 // of an sextload into a sextload targeting a wider value.
14836 if (MemSz == 128) {
14837 // Just switch this to a normal load.
14838 assert(TLI.isTypeLegal(MemVT) && "If the memory type is a 128-bit type, "
14839 "it must be a legal 128-bit vector "
14841 Load = DAG.getLoad(MemVT, dl, Ld->getChain(), Ld->getBasePtr(),
14842 Ld->getPointerInfo(), Ld->isVolatile(), Ld->isNonTemporal(),
14843 Ld->isInvariant(), Ld->getAlignment());
14845 assert(MemSz < 128 &&
14846 "Can't extend a type wider than 128 bits to a 256 bit vector!");
14847 // Do an sext load to a 128-bit vector type. We want to use the same
14848 // number of elements, but elements half as wide. This will end up being
14849 // recursively lowered by this routine, but will succeed as we definitely
14850 // have all the necessary features if we're using AVX1.
14852 EVT::getIntegerVT(*DAG.getContext(), RegVT.getScalarSizeInBits() / 2);
14853 EVT HalfVecVT = EVT::getVectorVT(*DAG.getContext(), HalfEltVT, NumElems);
14855 DAG.getExtLoad(Ext, dl, HalfVecVT, Ld->getChain(), Ld->getBasePtr(),
14856 Ld->getPointerInfo(), MemVT, Ld->isVolatile(),
14857 Ld->isNonTemporal(), Ld->isInvariant(),
14858 Ld->getAlignment());
14861 // Replace chain users with the new chain.
14862 assert(Load->getNumValues() == 2 && "Loads must carry a chain!");
14863 DAG.ReplaceAllUsesOfValueWith(SDValue(Ld, 1), Load.getValue(1));
14865 // Finally, do a normal sign-extend to the desired register.
14866 return DAG.getSExtOrTrunc(Load, dl, RegVT);
14869 // All sizes must be a power of two.
14870 assert(isPowerOf2_32(RegSz * MemSz * NumElems) &&
14871 "Non-power-of-two elements are not custom lowered!");
14873 // Attempt to load the original value using scalar loads.
14874 // Find the largest scalar type that divides the total loaded size.
14875 MVT SclrLoadTy = MVT::i8;
14876 for (MVT Tp : MVT::integer_valuetypes()) {
14877 if (TLI.isTypeLegal(Tp) && ((MemSz % Tp.getSizeInBits()) == 0)) {
14882 // On 32bit systems, we can't save 64bit integers. Try bitcasting to F64.
14883 if (TLI.isTypeLegal(MVT::f64) && SclrLoadTy.getSizeInBits() < 64 &&
14885 SclrLoadTy = MVT::f64;
14887 // Calculate the number of scalar loads that we need to perform
14888 // in order to load our vector from memory.
14889 unsigned NumLoads = MemSz / SclrLoadTy.getSizeInBits();
14891 assert((Ext != ISD::SEXTLOAD || NumLoads == 1) &&
14892 "Can only lower sext loads with a single scalar load!");
14894 unsigned loadRegZize = RegSz;
14895 if (Ext == ISD::SEXTLOAD && RegSz >= 256)
14898 // Represent our vector as a sequence of elements which are the
14899 // largest scalar that we can load.
14900 EVT LoadUnitVecVT = EVT::getVectorVT(
14901 *DAG.getContext(), SclrLoadTy, loadRegZize / SclrLoadTy.getSizeInBits());
14903 // Represent the data using the same element type that is stored in
14904 // memory. In practice, we ''widen'' MemVT.
14906 EVT::getVectorVT(*DAG.getContext(), MemVT.getScalarType(),
14907 loadRegZize / MemVT.getScalarType().getSizeInBits());
14909 assert(WideVecVT.getSizeInBits() == LoadUnitVecVT.getSizeInBits() &&
14910 "Invalid vector type");
14912 // We can't shuffle using an illegal type.
14913 assert(TLI.isTypeLegal(WideVecVT) &&
14914 "We only lower types that form legal widened vector types");
14916 SmallVector<SDValue, 8> Chains;
14917 SDValue Ptr = Ld->getBasePtr();
14918 SDValue Increment = DAG.getConstant(SclrLoadTy.getSizeInBits() / 8, dl,
14919 TLI.getPointerTy(DAG.getDataLayout()));
14920 SDValue Res = DAG.getUNDEF(LoadUnitVecVT);
14922 for (unsigned i = 0; i < NumLoads; ++i) {
14923 // Perform a single load.
14924 SDValue ScalarLoad =
14925 DAG.getLoad(SclrLoadTy, dl, Ld->getChain(), Ptr, Ld->getPointerInfo(),
14926 Ld->isVolatile(), Ld->isNonTemporal(), Ld->isInvariant(),
14927 Ld->getAlignment());
14928 Chains.push_back(ScalarLoad.getValue(1));
14929 // Create the first element type using SCALAR_TO_VECTOR in order to avoid
14930 // another round of DAGCombining.
14932 Res = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, LoadUnitVecVT, ScalarLoad);
14934 Res = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, LoadUnitVecVT, Res,
14935 ScalarLoad, DAG.getIntPtrConstant(i, dl));
14937 Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr, Increment);
14940 SDValue TF = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Chains);
14942 // Bitcast the loaded value to a vector of the original element type, in
14943 // the size of the target vector type.
14944 SDValue SlicedVec = DAG.getBitcast(WideVecVT, Res);
14945 unsigned SizeRatio = RegSz / MemSz;
14947 if (Ext == ISD::SEXTLOAD) {
14948 // If we have SSE4.1, we can directly emit a VSEXT node.
14949 if (Subtarget->hasSSE41()) {
14950 SDValue Sext = DAG.getNode(X86ISD::VSEXT, dl, RegVT, SlicedVec);
14951 DAG.ReplaceAllUsesOfValueWith(SDValue(Ld, 1), TF);
14955 // Otherwise we'll shuffle the small elements in the high bits of the
14956 // larger type and perform an arithmetic shift. If the shift is not legal
14957 // it's better to scalarize.
14958 assert(TLI.isOperationLegalOrCustom(ISD::SRA, RegVT) &&
14959 "We can't implement a sext load without an arithmetic right shift!");
14961 // Redistribute the loaded elements into the different locations.
14962 SmallVector<int, 16> ShuffleVec(NumElems * SizeRatio, -1);
14963 for (unsigned i = 0; i != NumElems; ++i)
14964 ShuffleVec[i * SizeRatio + SizeRatio - 1] = i;
14966 SDValue Shuff = DAG.getVectorShuffle(
14967 WideVecVT, dl, SlicedVec, DAG.getUNDEF(WideVecVT), &ShuffleVec[0]);
14969 Shuff = DAG.getBitcast(RegVT, Shuff);
14971 // Build the arithmetic shift.
14972 unsigned Amt = RegVT.getVectorElementType().getSizeInBits() -
14973 MemVT.getVectorElementType().getSizeInBits();
14975 DAG.getNode(ISD::SRA, dl, RegVT, Shuff,
14976 DAG.getConstant(Amt, dl, RegVT));
14978 DAG.ReplaceAllUsesOfValueWith(SDValue(Ld, 1), TF);
14982 // Redistribute the loaded elements into the different locations.
14983 SmallVector<int, 16> ShuffleVec(NumElems * SizeRatio, -1);
14984 for (unsigned i = 0; i != NumElems; ++i)
14985 ShuffleVec[i * SizeRatio] = i;
14987 SDValue Shuff = DAG.getVectorShuffle(WideVecVT, dl, SlicedVec,
14988 DAG.getUNDEF(WideVecVT), &ShuffleVec[0]);
14990 // Bitcast to the requested type.
14991 Shuff = DAG.getBitcast(RegVT, Shuff);
14992 DAG.ReplaceAllUsesOfValueWith(SDValue(Ld, 1), TF);
14996 // isAndOrOfSingleUseSetCCs - Return true if node is an ISD::AND or
14997 // ISD::OR of two X86ISD::SETCC nodes each of which has no other use apart
14998 // from the AND / OR.
14999 static bool isAndOrOfSetCCs(SDValue Op, unsigned &Opc) {
15000 Opc = Op.getOpcode();
15001 if (Opc != ISD::OR && Opc != ISD::AND)
15003 return (Op.getOperand(0).getOpcode() == X86ISD::SETCC &&
15004 Op.getOperand(0).hasOneUse() &&
15005 Op.getOperand(1).getOpcode() == X86ISD::SETCC &&
15006 Op.getOperand(1).hasOneUse());
15009 // isXor1OfSetCC - Return true if node is an ISD::XOR of a X86ISD::SETCC and
15010 // 1 and that the SETCC node has a single use.
15011 static bool isXor1OfSetCC(SDValue Op) {
15012 if (Op.getOpcode() != ISD::XOR)
15014 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(Op.getOperand(1));
15015 if (N1C && N1C->getAPIntValue() == 1) {
15016 return Op.getOperand(0).getOpcode() == X86ISD::SETCC &&
15017 Op.getOperand(0).hasOneUse();
15022 SDValue X86TargetLowering::LowerBRCOND(SDValue Op, SelectionDAG &DAG) const {
15023 bool addTest = true;
15024 SDValue Chain = Op.getOperand(0);
15025 SDValue Cond = Op.getOperand(1);
15026 SDValue Dest = Op.getOperand(2);
15029 bool Inverted = false;
15031 if (Cond.getOpcode() == ISD::SETCC) {
15032 // Check for setcc([su]{add,sub,mul}o == 0).
15033 if (cast<CondCodeSDNode>(Cond.getOperand(2))->get() == ISD::SETEQ &&
15034 isa<ConstantSDNode>(Cond.getOperand(1)) &&
15035 cast<ConstantSDNode>(Cond.getOperand(1))->isNullValue() &&
15036 Cond.getOperand(0).getResNo() == 1 &&
15037 (Cond.getOperand(0).getOpcode() == ISD::SADDO ||
15038 Cond.getOperand(0).getOpcode() == ISD::UADDO ||
15039 Cond.getOperand(0).getOpcode() == ISD::SSUBO ||
15040 Cond.getOperand(0).getOpcode() == ISD::USUBO ||
15041 Cond.getOperand(0).getOpcode() == ISD::SMULO ||
15042 Cond.getOperand(0).getOpcode() == ISD::UMULO)) {
15044 Cond = Cond.getOperand(0);
15046 SDValue NewCond = LowerSETCC(Cond, DAG);
15047 if (NewCond.getNode())
15052 // FIXME: LowerXALUO doesn't handle these!!
15053 else if (Cond.getOpcode() == X86ISD::ADD ||
15054 Cond.getOpcode() == X86ISD::SUB ||
15055 Cond.getOpcode() == X86ISD::SMUL ||
15056 Cond.getOpcode() == X86ISD::UMUL)
15057 Cond = LowerXALUO(Cond, DAG);
15060 // Look pass (and (setcc_carry (cmp ...)), 1).
15061 if (Cond.getOpcode() == ISD::AND &&
15062 Cond.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY) {
15063 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Cond.getOperand(1));
15064 if (C && C->getAPIntValue() == 1)
15065 Cond = Cond.getOperand(0);
15068 // If condition flag is set by a X86ISD::CMP, then use it as the condition
15069 // setting operand in place of the X86ISD::SETCC.
15070 unsigned CondOpcode = Cond.getOpcode();
15071 if (CondOpcode == X86ISD::SETCC ||
15072 CondOpcode == X86ISD::SETCC_CARRY) {
15073 CC = Cond.getOperand(0);
15075 SDValue Cmp = Cond.getOperand(1);
15076 unsigned Opc = Cmp.getOpcode();
15077 // FIXME: WHY THE SPECIAL CASING OF LogicalCmp??
15078 if (isX86LogicalCmp(Cmp) || Opc == X86ISD::BT) {
15082 switch (cast<ConstantSDNode>(CC)->getZExtValue()) {
15086 // These can only come from an arithmetic instruction with overflow,
15087 // e.g. SADDO, UADDO.
15088 Cond = Cond.getNode()->getOperand(1);
15094 CondOpcode = Cond.getOpcode();
15095 if (CondOpcode == ISD::UADDO || CondOpcode == ISD::SADDO ||
15096 CondOpcode == ISD::USUBO || CondOpcode == ISD::SSUBO ||
15097 ((CondOpcode == ISD::UMULO || CondOpcode == ISD::SMULO) &&
15098 Cond.getOperand(0).getValueType() != MVT::i8)) {
15099 SDValue LHS = Cond.getOperand(0);
15100 SDValue RHS = Cond.getOperand(1);
15101 unsigned X86Opcode;
15104 // Keep this in sync with LowerXALUO, otherwise we might create redundant
15105 // instructions that can't be removed afterwards (i.e. X86ISD::ADD and
15107 switch (CondOpcode) {
15108 case ISD::UADDO: X86Opcode = X86ISD::ADD; X86Cond = X86::COND_B; break;
15110 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(RHS))
15112 X86Opcode = X86ISD::INC; X86Cond = X86::COND_O;
15115 X86Opcode = X86ISD::ADD; X86Cond = X86::COND_O; break;
15116 case ISD::USUBO: X86Opcode = X86ISD::SUB; X86Cond = X86::COND_B; break;
15118 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(RHS))
15120 X86Opcode = X86ISD::DEC; X86Cond = X86::COND_O;
15123 X86Opcode = X86ISD::SUB; X86Cond = X86::COND_O; break;
15124 case ISD::UMULO: X86Opcode = X86ISD::UMUL; X86Cond = X86::COND_O; break;
15125 case ISD::SMULO: X86Opcode = X86ISD::SMUL; X86Cond = X86::COND_O; break;
15126 default: llvm_unreachable("unexpected overflowing operator");
15129 X86Cond = X86::GetOppositeBranchCondition((X86::CondCode)X86Cond);
15130 if (CondOpcode == ISD::UMULO)
15131 VTs = DAG.getVTList(LHS.getValueType(), LHS.getValueType(),
15134 VTs = DAG.getVTList(LHS.getValueType(), MVT::i32);
15136 SDValue X86Op = DAG.getNode(X86Opcode, dl, VTs, LHS, RHS);
15138 if (CondOpcode == ISD::UMULO)
15139 Cond = X86Op.getValue(2);
15141 Cond = X86Op.getValue(1);
15143 CC = DAG.getConstant(X86Cond, dl, MVT::i8);
15147 if (Cond.hasOneUse() && isAndOrOfSetCCs(Cond, CondOpc)) {
15148 SDValue Cmp = Cond.getOperand(0).getOperand(1);
15149 if (CondOpc == ISD::OR) {
15150 // Also, recognize the pattern generated by an FCMP_UNE. We can emit
15151 // two branches instead of an explicit OR instruction with a
15153 if (Cmp == Cond.getOperand(1).getOperand(1) &&
15154 isX86LogicalCmp(Cmp)) {
15155 CC = Cond.getOperand(0).getOperand(0);
15156 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
15157 Chain, Dest, CC, Cmp);
15158 CC = Cond.getOperand(1).getOperand(0);
15162 } else { // ISD::AND
15163 // Also, recognize the pattern generated by an FCMP_OEQ. We can emit
15164 // two branches instead of an explicit AND instruction with a
15165 // separate test. However, we only do this if this block doesn't
15166 // have a fall-through edge, because this requires an explicit
15167 // jmp when the condition is false.
15168 if (Cmp == Cond.getOperand(1).getOperand(1) &&
15169 isX86LogicalCmp(Cmp) &&
15170 Op.getNode()->hasOneUse()) {
15171 X86::CondCode CCode =
15172 (X86::CondCode)Cond.getOperand(0).getConstantOperandVal(0);
15173 CCode = X86::GetOppositeBranchCondition(CCode);
15174 CC = DAG.getConstant(CCode, dl, MVT::i8);
15175 SDNode *User = *Op.getNode()->use_begin();
15176 // Look for an unconditional branch following this conditional branch.
15177 // We need this because we need to reverse the successors in order
15178 // to implement FCMP_OEQ.
15179 if (User->getOpcode() == ISD::BR) {
15180 SDValue FalseBB = User->getOperand(1);
15182 DAG.UpdateNodeOperands(User, User->getOperand(0), Dest);
15183 assert(NewBR == User);
15187 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
15188 Chain, Dest, CC, Cmp);
15189 X86::CondCode CCode =
15190 (X86::CondCode)Cond.getOperand(1).getConstantOperandVal(0);
15191 CCode = X86::GetOppositeBranchCondition(CCode);
15192 CC = DAG.getConstant(CCode, dl, MVT::i8);
15198 } else if (Cond.hasOneUse() && isXor1OfSetCC(Cond)) {
15199 // Recognize for xorb (setcc), 1 patterns. The xor inverts the condition.
15200 // It should be transformed during dag combiner except when the condition
15201 // is set by a arithmetics with overflow node.
15202 X86::CondCode CCode =
15203 (X86::CondCode)Cond.getOperand(0).getConstantOperandVal(0);
15204 CCode = X86::GetOppositeBranchCondition(CCode);
15205 CC = DAG.getConstant(CCode, dl, MVT::i8);
15206 Cond = Cond.getOperand(0).getOperand(1);
15208 } else if (Cond.getOpcode() == ISD::SETCC &&
15209 cast<CondCodeSDNode>(Cond.getOperand(2))->get() == ISD::SETOEQ) {
15210 // For FCMP_OEQ, we can emit
15211 // two branches instead of an explicit AND instruction with a
15212 // separate test. However, we only do this if this block doesn't
15213 // have a fall-through edge, because this requires an explicit
15214 // jmp when the condition is false.
15215 if (Op.getNode()->hasOneUse()) {
15216 SDNode *User = *Op.getNode()->use_begin();
15217 // Look for an unconditional branch following this conditional branch.
15218 // We need this because we need to reverse the successors in order
15219 // to implement FCMP_OEQ.
15220 if (User->getOpcode() == ISD::BR) {
15221 SDValue FalseBB = User->getOperand(1);
15223 DAG.UpdateNodeOperands(User, User->getOperand(0), Dest);
15224 assert(NewBR == User);
15228 SDValue Cmp = DAG.getNode(X86ISD::CMP, dl, MVT::i32,
15229 Cond.getOperand(0), Cond.getOperand(1));
15230 Cmp = ConvertCmpIfNecessary(Cmp, DAG);
15231 CC = DAG.getConstant(X86::COND_NE, dl, MVT::i8);
15232 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
15233 Chain, Dest, CC, Cmp);
15234 CC = DAG.getConstant(X86::COND_P, dl, MVT::i8);
15239 } else if (Cond.getOpcode() == ISD::SETCC &&
15240 cast<CondCodeSDNode>(Cond.getOperand(2))->get() == ISD::SETUNE) {
15241 // For FCMP_UNE, we can emit
15242 // two branches instead of an explicit AND instruction with a
15243 // separate test. However, we only do this if this block doesn't
15244 // have a fall-through edge, because this requires an explicit
15245 // jmp when the condition is false.
15246 if (Op.getNode()->hasOneUse()) {
15247 SDNode *User = *Op.getNode()->use_begin();
15248 // Look for an unconditional branch following this conditional branch.
15249 // We need this because we need to reverse the successors in order
15250 // to implement FCMP_UNE.
15251 if (User->getOpcode() == ISD::BR) {
15252 SDValue FalseBB = User->getOperand(1);
15254 DAG.UpdateNodeOperands(User, User->getOperand(0), Dest);
15255 assert(NewBR == User);
15258 SDValue Cmp = DAG.getNode(X86ISD::CMP, dl, MVT::i32,
15259 Cond.getOperand(0), Cond.getOperand(1));
15260 Cmp = ConvertCmpIfNecessary(Cmp, DAG);
15261 CC = DAG.getConstant(X86::COND_NE, dl, MVT::i8);
15262 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
15263 Chain, Dest, CC, Cmp);
15264 CC = DAG.getConstant(X86::COND_NP, dl, MVT::i8);
15274 // Look pass the truncate if the high bits are known zero.
15275 if (isTruncWithZeroHighBitsInput(Cond, DAG))
15276 Cond = Cond.getOperand(0);
15278 // We know the result of AND is compared against zero. Try to match
15280 if (Cond.getOpcode() == ISD::AND && Cond.hasOneUse()) {
15281 SDValue NewSetCC = LowerToBT(Cond, ISD::SETNE, dl, DAG);
15282 if (NewSetCC.getNode()) {
15283 CC = NewSetCC.getOperand(0);
15284 Cond = NewSetCC.getOperand(1);
15291 X86::CondCode X86Cond = Inverted ? X86::COND_E : X86::COND_NE;
15292 CC = DAG.getConstant(X86Cond, dl, MVT::i8);
15293 Cond = EmitTest(Cond, X86Cond, dl, DAG);
15295 Cond = ConvertCmpIfNecessary(Cond, DAG);
15296 return DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
15297 Chain, Dest, CC, Cond);
15300 // Lower dynamic stack allocation to _alloca call for Cygwin/Mingw targets.
15301 // Calls to _alloca are needed to probe the stack when allocating more than 4k
15302 // bytes in one go. Touching the stack at 4K increments is necessary to ensure
15303 // that the guard pages used by the OS virtual memory manager are allocated in
15304 // correct sequence.
15306 X86TargetLowering::LowerDYNAMIC_STACKALLOC(SDValue Op,
15307 SelectionDAG &DAG) const {
15308 MachineFunction &MF = DAG.getMachineFunction();
15309 bool SplitStack = MF.shouldSplitStack();
15310 bool Lower = (Subtarget->isOSWindows() && !Subtarget->isTargetMachO()) ||
15315 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
15316 SDNode* Node = Op.getNode();
15318 unsigned SPReg = TLI.getStackPointerRegisterToSaveRestore();
15319 assert(SPReg && "Target cannot require DYNAMIC_STACKALLOC expansion and"
15320 " not tell us which reg is the stack pointer!");
15321 EVT VT = Node->getValueType(0);
15322 SDValue Tmp1 = SDValue(Node, 0);
15323 SDValue Tmp2 = SDValue(Node, 1);
15324 SDValue Tmp3 = Node->getOperand(2);
15325 SDValue Chain = Tmp1.getOperand(0);
15327 // Chain the dynamic stack allocation so that it doesn't modify the stack
15328 // pointer when other instructions are using the stack.
15329 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(0, dl, true),
15332 SDValue Size = Tmp2.getOperand(1);
15333 SDValue SP = DAG.getCopyFromReg(Chain, dl, SPReg, VT);
15334 Chain = SP.getValue(1);
15335 unsigned Align = cast<ConstantSDNode>(Tmp3)->getZExtValue();
15336 const TargetFrameLowering &TFI = *Subtarget->getFrameLowering();
15337 unsigned StackAlign = TFI.getStackAlignment();
15338 Tmp1 = DAG.getNode(ISD::SUB, dl, VT, SP, Size); // Value
15339 if (Align > StackAlign)
15340 Tmp1 = DAG.getNode(ISD::AND, dl, VT, Tmp1,
15341 DAG.getConstant(-(uint64_t)Align, dl, VT));
15342 Chain = DAG.getCopyToReg(Chain, dl, SPReg, Tmp1); // Output chain
15344 Tmp2 = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(0, dl, true),
15345 DAG.getIntPtrConstant(0, dl, true), SDValue(),
15348 SDValue Ops[2] = { Tmp1, Tmp2 };
15349 return DAG.getMergeValues(Ops, dl);
15353 SDValue Chain = Op.getOperand(0);
15354 SDValue Size = Op.getOperand(1);
15355 unsigned Align = cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue();
15356 EVT VT = Op.getNode()->getValueType(0);
15358 bool Is64Bit = Subtarget->is64Bit();
15359 MVT SPTy = getPointerTy(DAG.getDataLayout());
15362 MachineRegisterInfo &MRI = MF.getRegInfo();
15365 // The 64 bit implementation of segmented stacks needs to clobber both r10
15366 // r11. This makes it impossible to use it along with nested parameters.
15367 const Function *F = MF.getFunction();
15369 for (Function::const_arg_iterator I = F->arg_begin(), E = F->arg_end();
15371 if (I->hasNestAttr())
15372 report_fatal_error("Cannot use segmented stacks with functions that "
15373 "have nested arguments.");
15376 const TargetRegisterClass *AddrRegClass = getRegClassFor(SPTy);
15377 unsigned Vreg = MRI.createVirtualRegister(AddrRegClass);
15378 Chain = DAG.getCopyToReg(Chain, dl, Vreg, Size);
15379 SDValue Value = DAG.getNode(X86ISD::SEG_ALLOCA, dl, SPTy, Chain,
15380 DAG.getRegister(Vreg, SPTy));
15381 SDValue Ops1[2] = { Value, Chain };
15382 return DAG.getMergeValues(Ops1, dl);
15385 const unsigned Reg = (Subtarget->isTarget64BitLP64() ? X86::RAX : X86::EAX);
15387 Chain = DAG.getCopyToReg(Chain, dl, Reg, Size, Flag);
15388 Flag = Chain.getValue(1);
15389 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
15391 Chain = DAG.getNode(X86ISD::WIN_ALLOCA, dl, NodeTys, Chain, Flag);
15393 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
15394 unsigned SPReg = RegInfo->getStackRegister();
15395 SDValue SP = DAG.getCopyFromReg(Chain, dl, SPReg, SPTy);
15396 Chain = SP.getValue(1);
15399 SP = DAG.getNode(ISD::AND, dl, VT, SP.getValue(0),
15400 DAG.getConstant(-(uint64_t)Align, dl, VT));
15401 Chain = DAG.getCopyToReg(Chain, dl, SPReg, SP);
15404 SDValue Ops1[2] = { SP, Chain };
15405 return DAG.getMergeValues(Ops1, dl);
15409 SDValue X86TargetLowering::LowerVASTART(SDValue Op, SelectionDAG &DAG) const {
15410 MachineFunction &MF = DAG.getMachineFunction();
15411 auto PtrVT = getPointerTy(MF.getDataLayout());
15412 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
15414 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
15417 if (!Subtarget->is64Bit() ||
15418 Subtarget->isCallingConvWin64(MF.getFunction()->getCallingConv())) {
15419 // vastart just stores the address of the VarArgsFrameIndex slot into the
15420 // memory location argument.
15421 SDValue FR = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(), PtrVT);
15422 return DAG.getStore(Op.getOperand(0), DL, FR, Op.getOperand(1),
15423 MachinePointerInfo(SV), false, false, 0);
15427 // gp_offset (0 - 6 * 8)
15428 // fp_offset (48 - 48 + 8 * 16)
15429 // overflow_arg_area (point to parameters coming in memory).
15431 SmallVector<SDValue, 8> MemOps;
15432 SDValue FIN = Op.getOperand(1);
15434 SDValue Store = DAG.getStore(Op.getOperand(0), DL,
15435 DAG.getConstant(FuncInfo->getVarArgsGPOffset(),
15437 FIN, MachinePointerInfo(SV), false, false, 0);
15438 MemOps.push_back(Store);
15441 FIN = DAG.getNode(ISD::ADD, DL, PtrVT, FIN, DAG.getIntPtrConstant(4, DL));
15442 Store = DAG.getStore(Op.getOperand(0), DL,
15443 DAG.getConstant(FuncInfo->getVarArgsFPOffset(), DL,
15445 FIN, MachinePointerInfo(SV, 4), false, false, 0);
15446 MemOps.push_back(Store);
15448 // Store ptr to overflow_arg_area
15449 FIN = DAG.getNode(ISD::ADD, DL, PtrVT, FIN, DAG.getIntPtrConstant(4, DL));
15450 SDValue OVFIN = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(), PtrVT);
15451 Store = DAG.getStore(Op.getOperand(0), DL, OVFIN, FIN,
15452 MachinePointerInfo(SV, 8),
15454 MemOps.push_back(Store);
15456 // Store ptr to reg_save_area.
15457 FIN = DAG.getNode(ISD::ADD, DL, PtrVT, FIN, DAG.getIntPtrConstant(
15458 Subtarget->isTarget64BitLP64() ? 8 : 4, DL));
15459 SDValue RSFIN = DAG.getFrameIndex(FuncInfo->getRegSaveFrameIndex(), PtrVT);
15460 Store = DAG.getStore(Op.getOperand(0), DL, RSFIN, FIN, MachinePointerInfo(
15461 SV, Subtarget->isTarget64BitLP64() ? 16 : 12), false, false, 0);
15462 MemOps.push_back(Store);
15463 return DAG.getNode(ISD::TokenFactor, DL, MVT::Other, MemOps);
15466 SDValue X86TargetLowering::LowerVAARG(SDValue Op, SelectionDAG &DAG) const {
15467 assert(Subtarget->is64Bit() &&
15468 "LowerVAARG only handles 64-bit va_arg!");
15469 assert(Op.getNode()->getNumOperands() == 4);
15471 MachineFunction &MF = DAG.getMachineFunction();
15472 if (Subtarget->isCallingConvWin64(MF.getFunction()->getCallingConv()))
15473 // The Win64 ABI uses char* instead of a structure.
15474 return DAG.expandVAArg(Op.getNode());
15476 SDValue Chain = Op.getOperand(0);
15477 SDValue SrcPtr = Op.getOperand(1);
15478 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
15479 unsigned Align = Op.getConstantOperandVal(3);
15482 EVT ArgVT = Op.getNode()->getValueType(0);
15483 Type *ArgTy = ArgVT.getTypeForEVT(*DAG.getContext());
15484 uint32_t ArgSize = DAG.getDataLayout().getTypeAllocSize(ArgTy);
15487 // Decide which area this value should be read from.
15488 // TODO: Implement the AMD64 ABI in its entirety. This simple
15489 // selection mechanism works only for the basic types.
15490 if (ArgVT == MVT::f80) {
15491 llvm_unreachable("va_arg for f80 not yet implemented");
15492 } else if (ArgVT.isFloatingPoint() && ArgSize <= 16 /*bytes*/) {
15493 ArgMode = 2; // Argument passed in XMM register. Use fp_offset.
15494 } else if (ArgVT.isInteger() && ArgSize <= 32 /*bytes*/) {
15495 ArgMode = 1; // Argument passed in GPR64 register(s). Use gp_offset.
15497 llvm_unreachable("Unhandled argument type in LowerVAARG");
15500 if (ArgMode == 2) {
15501 // Sanity Check: Make sure using fp_offset makes sense.
15502 assert(!Subtarget->useSoftFloat() &&
15503 !(MF.getFunction()->hasFnAttribute(Attribute::NoImplicitFloat)) &&
15504 Subtarget->hasSSE1());
15507 // Insert VAARG_64 node into the DAG
15508 // VAARG_64 returns two values: Variable Argument Address, Chain
15509 SDValue InstOps[] = {Chain, SrcPtr, DAG.getConstant(ArgSize, dl, MVT::i32),
15510 DAG.getConstant(ArgMode, dl, MVT::i8),
15511 DAG.getConstant(Align, dl, MVT::i32)};
15512 SDVTList VTs = DAG.getVTList(getPointerTy(DAG.getDataLayout()), MVT::Other);
15513 SDValue VAARG = DAG.getMemIntrinsicNode(X86ISD::VAARG_64, dl,
15514 VTs, InstOps, MVT::i64,
15515 MachinePointerInfo(SV),
15517 /*Volatile=*/false,
15519 /*WriteMem=*/true);
15520 Chain = VAARG.getValue(1);
15522 // Load the next argument and return it
15523 return DAG.getLoad(ArgVT, dl,
15526 MachinePointerInfo(),
15527 false, false, false, 0);
15530 static SDValue LowerVACOPY(SDValue Op, const X86Subtarget *Subtarget,
15531 SelectionDAG &DAG) {
15532 // X86-64 va_list is a struct { i32, i32, i8*, i8* }, except on Windows,
15533 // where a va_list is still an i8*.
15534 assert(Subtarget->is64Bit() && "This code only handles 64-bit va_copy!");
15535 if (Subtarget->isCallingConvWin64(
15536 DAG.getMachineFunction().getFunction()->getCallingConv()))
15537 // Probably a Win64 va_copy.
15538 return DAG.expandVACopy(Op.getNode());
15540 SDValue Chain = Op.getOperand(0);
15541 SDValue DstPtr = Op.getOperand(1);
15542 SDValue SrcPtr = Op.getOperand(2);
15543 const Value *DstSV = cast<SrcValueSDNode>(Op.getOperand(3))->getValue();
15544 const Value *SrcSV = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
15547 return DAG.getMemcpy(Chain, DL, DstPtr, SrcPtr,
15548 DAG.getIntPtrConstant(24, DL), 8, /*isVolatile*/false,
15550 MachinePointerInfo(DstSV), MachinePointerInfo(SrcSV));
15553 // getTargetVShiftByConstNode - Handle vector element shifts where the shift
15554 // amount is a constant. Takes immediate version of shift as input.
15555 static SDValue getTargetVShiftByConstNode(unsigned Opc, SDLoc dl, MVT VT,
15556 SDValue SrcOp, uint64_t ShiftAmt,
15557 SelectionDAG &DAG) {
15558 MVT ElementType = VT.getVectorElementType();
15560 // Fold this packed shift into its first operand if ShiftAmt is 0.
15564 // Check for ShiftAmt >= element width
15565 if (ShiftAmt >= ElementType.getSizeInBits()) {
15566 if (Opc == X86ISD::VSRAI)
15567 ShiftAmt = ElementType.getSizeInBits() - 1;
15569 return DAG.getConstant(0, dl, VT);
15572 assert((Opc == X86ISD::VSHLI || Opc == X86ISD::VSRLI || Opc == X86ISD::VSRAI)
15573 && "Unknown target vector shift-by-constant node");
15575 // Fold this packed vector shift into a build vector if SrcOp is a
15576 // vector of Constants or UNDEFs, and SrcOp valuetype is the same as VT.
15577 if (VT == SrcOp.getSimpleValueType() &&
15578 ISD::isBuildVectorOfConstantSDNodes(SrcOp.getNode())) {
15579 SmallVector<SDValue, 8> Elts;
15580 unsigned NumElts = SrcOp->getNumOperands();
15581 ConstantSDNode *ND;
15584 default: llvm_unreachable(nullptr);
15585 case X86ISD::VSHLI:
15586 for (unsigned i=0; i!=NumElts; ++i) {
15587 SDValue CurrentOp = SrcOp->getOperand(i);
15588 if (CurrentOp->getOpcode() == ISD::UNDEF) {
15589 Elts.push_back(CurrentOp);
15592 ND = cast<ConstantSDNode>(CurrentOp);
15593 const APInt &C = ND->getAPIntValue();
15594 Elts.push_back(DAG.getConstant(C.shl(ShiftAmt), dl, ElementType));
15597 case X86ISD::VSRLI:
15598 for (unsigned i=0; i!=NumElts; ++i) {
15599 SDValue CurrentOp = SrcOp->getOperand(i);
15600 if (CurrentOp->getOpcode() == ISD::UNDEF) {
15601 Elts.push_back(CurrentOp);
15604 ND = cast<ConstantSDNode>(CurrentOp);
15605 const APInt &C = ND->getAPIntValue();
15606 Elts.push_back(DAG.getConstant(C.lshr(ShiftAmt), dl, ElementType));
15609 case X86ISD::VSRAI:
15610 for (unsigned i=0; i!=NumElts; ++i) {
15611 SDValue CurrentOp = SrcOp->getOperand(i);
15612 if (CurrentOp->getOpcode() == ISD::UNDEF) {
15613 Elts.push_back(CurrentOp);
15616 ND = cast<ConstantSDNode>(CurrentOp);
15617 const APInt &C = ND->getAPIntValue();
15618 Elts.push_back(DAG.getConstant(C.ashr(ShiftAmt), dl, ElementType));
15623 return DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Elts);
15626 return DAG.getNode(Opc, dl, VT, SrcOp,
15627 DAG.getConstant(ShiftAmt, dl, MVT::i8));
15630 // getTargetVShiftNode - Handle vector element shifts where the shift amount
15631 // may or may not be a constant. Takes immediate version of shift as input.
15632 static SDValue getTargetVShiftNode(unsigned Opc, SDLoc dl, MVT VT,
15633 SDValue SrcOp, SDValue ShAmt,
15634 SelectionDAG &DAG) {
15635 MVT SVT = ShAmt.getSimpleValueType();
15636 assert((SVT == MVT::i32 || SVT == MVT::i64) && "Unexpected value type!");
15638 // Catch shift-by-constant.
15639 if (ConstantSDNode *CShAmt = dyn_cast<ConstantSDNode>(ShAmt))
15640 return getTargetVShiftByConstNode(Opc, dl, VT, SrcOp,
15641 CShAmt->getZExtValue(), DAG);
15643 // Change opcode to non-immediate version
15645 default: llvm_unreachable("Unknown target vector shift node");
15646 case X86ISD::VSHLI: Opc = X86ISD::VSHL; break;
15647 case X86ISD::VSRLI: Opc = X86ISD::VSRL; break;
15648 case X86ISD::VSRAI: Opc = X86ISD::VSRA; break;
15651 const X86Subtarget &Subtarget =
15652 static_cast<const X86Subtarget &>(DAG.getSubtarget());
15653 if (Subtarget.hasSSE41() && ShAmt.getOpcode() == ISD::ZERO_EXTEND &&
15654 ShAmt.getOperand(0).getSimpleValueType() == MVT::i16) {
15655 // Let the shuffle legalizer expand this shift amount node.
15656 SDValue Op0 = ShAmt.getOperand(0);
15657 Op0 = DAG.getNode(ISD::SCALAR_TO_VECTOR, SDLoc(Op0), MVT::v8i16, Op0);
15658 ShAmt = getShuffleVectorZeroOrUndef(Op0, 0, true, &Subtarget, DAG);
15660 // Need to build a vector containing shift amount.
15661 // SSE/AVX packed shifts only use the lower 64-bit of the shift count.
15662 SmallVector<SDValue, 4> ShOps;
15663 ShOps.push_back(ShAmt);
15664 if (SVT == MVT::i32) {
15665 ShOps.push_back(DAG.getConstant(0, dl, SVT));
15666 ShOps.push_back(DAG.getUNDEF(SVT));
15668 ShOps.push_back(DAG.getUNDEF(SVT));
15670 MVT BVT = SVT == MVT::i32 ? MVT::v4i32 : MVT::v2i64;
15671 ShAmt = DAG.getNode(ISD::BUILD_VECTOR, dl, BVT, ShOps);
15674 // The return type has to be a 128-bit type with the same element
15675 // type as the input type.
15676 MVT EltVT = VT.getVectorElementType();
15677 EVT ShVT = MVT::getVectorVT(EltVT, 128/EltVT.getSizeInBits());
15679 ShAmt = DAG.getBitcast(ShVT, ShAmt);
15680 return DAG.getNode(Opc, dl, VT, SrcOp, ShAmt);
15683 /// \brief Return (and \p Op, \p Mask) for compare instructions or
15684 /// (vselect \p Mask, \p Op, \p PreservedSrc) for others along with the
15685 /// necessary casting or extending for \p Mask when lowering masking intrinsics
15686 static SDValue getVectorMaskingNode(SDValue Op, SDValue Mask,
15687 SDValue PreservedSrc,
15688 const X86Subtarget *Subtarget,
15689 SelectionDAG &DAG) {
15690 EVT VT = Op.getValueType();
15691 EVT MaskVT = EVT::getVectorVT(*DAG.getContext(),
15692 MVT::i1, VT.getVectorNumElements());
15693 SDValue VMask = SDValue();
15694 unsigned OpcodeSelect = ISD::VSELECT;
15697 assert(MaskVT.isSimple() && "invalid mask type");
15699 if (isAllOnes(Mask))
15702 if (MaskVT.bitsGT(Mask.getValueType())) {
15703 EVT newMaskVT = EVT::getIntegerVT(*DAG.getContext(),
15704 MaskVT.getSizeInBits());
15705 VMask = DAG.getBitcast(MaskVT,
15706 DAG.getNode(ISD::ANY_EXTEND, dl, newMaskVT, Mask));
15708 EVT BitcastVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
15709 Mask.getValueType().getSizeInBits());
15710 // In case when MaskVT equals v2i1 or v4i1, low 2 or 4 elements
15711 // are extracted by EXTRACT_SUBVECTOR.
15712 VMask = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MaskVT,
15713 DAG.getBitcast(BitcastVT, Mask),
15714 DAG.getIntPtrConstant(0, dl));
15717 switch (Op.getOpcode()) {
15719 case X86ISD::PCMPEQM:
15720 case X86ISD::PCMPGTM:
15722 case X86ISD::CMPMU:
15723 return DAG.getNode(ISD::AND, dl, VT, Op, VMask);
15724 case X86ISD::VTRUNC:
15725 case X86ISD::VTRUNCS:
15726 case X86ISD::VTRUNCUS:
15727 // We can't use ISD::VSELECT here because it is not always "Legal"
15728 // for the destination type. For example vpmovqb require only AVX512
15729 // and vselect that can operate on byte element type require BWI
15730 OpcodeSelect = X86ISD::SELECT;
15733 if (PreservedSrc.getOpcode() == ISD::UNDEF)
15734 PreservedSrc = getZeroVector(VT, Subtarget, DAG, dl);
15735 return DAG.getNode(OpcodeSelect, dl, VT, VMask, Op, PreservedSrc);
15738 /// \brief Creates an SDNode for a predicated scalar operation.
15739 /// \returns (X86vselect \p Mask, \p Op, \p PreservedSrc).
15740 /// The mask is coming as MVT::i8 and it should be truncated
15741 /// to MVT::i1 while lowering masking intrinsics.
15742 /// The main difference between ScalarMaskingNode and VectorMaskingNode is using
15743 /// "X86select" instead of "vselect". We just can't create the "vselect" node
15744 /// for a scalar instruction.
15745 static SDValue getScalarMaskingNode(SDValue Op, SDValue Mask,
15746 SDValue PreservedSrc,
15747 const X86Subtarget *Subtarget,
15748 SelectionDAG &DAG) {
15749 if (isAllOnes(Mask))
15752 EVT VT = Op.getValueType();
15754 // The mask should be of type MVT::i1
15755 SDValue IMask = DAG.getNode(ISD::TRUNCATE, dl, MVT::i1, Mask);
15757 if (PreservedSrc.getOpcode() == ISD::UNDEF)
15758 PreservedSrc = getZeroVector(VT, Subtarget, DAG, dl);
15759 return DAG.getNode(X86ISD::SELECT, dl, VT, IMask, Op, PreservedSrc);
15762 static int getSEHRegistrationNodeSize(const Function *Fn) {
15763 if (!Fn->hasPersonalityFn())
15764 report_fatal_error(
15765 "querying registration node size for function without personality");
15766 // The RegNodeSize is 6 32-bit words for SEH and 4 for C++ EH. See
15767 // WinEHStatePass for the full struct definition.
15768 switch (classifyEHPersonality(Fn->getPersonalityFn())) {
15769 case EHPersonality::MSVC_X86SEH: return 24;
15770 case EHPersonality::MSVC_CXX: return 16;
15773 report_fatal_error("can only recover FP for MSVC EH personality functions");
15776 /// When the 32-bit MSVC runtime transfers control to us, either to an outlined
15777 /// function or when returning to a parent frame after catching an exception, we
15778 /// recover the parent frame pointer by doing arithmetic on the incoming EBP.
15779 /// Here's the math:
15780 /// RegNodeBase = EntryEBP - RegNodeSize
15781 /// ParentFP = RegNodeBase - RegNodeFrameOffset
15782 /// Subtracting RegNodeSize takes us to the offset of the registration node, and
15783 /// subtracting the offset (negative on x86) takes us back to the parent FP.
15784 static SDValue recoverFramePointer(SelectionDAG &DAG, const Function *Fn,
15785 SDValue EntryEBP) {
15786 MachineFunction &MF = DAG.getMachineFunction();
15789 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
15790 MVT PtrVT = TLI.getPointerTy(DAG.getDataLayout());
15792 // It's possible that the parent function no longer has a personality function
15793 // if the exceptional code was optimized away, in which case we just return
15794 // the incoming EBP.
15795 if (!Fn->hasPersonalityFn())
15798 int RegNodeSize = getSEHRegistrationNodeSize(Fn);
15800 // Get an MCSymbol that will ultimately resolve to the frame offset of the EH
15802 MCSymbol *OffsetSym =
15803 MF.getMMI().getContext().getOrCreateParentFrameOffsetSymbol(
15804 GlobalValue::getRealLinkageName(Fn->getName()));
15805 SDValue OffsetSymVal = DAG.getMCSymbol(OffsetSym, PtrVT);
15806 SDValue RegNodeFrameOffset =
15807 DAG.getNode(ISD::LOCAL_RECOVER, dl, PtrVT, OffsetSymVal);
15809 // RegNodeBase = EntryEBP - RegNodeSize
15810 // ParentFP = RegNodeBase - RegNodeFrameOffset
15811 SDValue RegNodeBase = DAG.getNode(ISD::SUB, dl, PtrVT, EntryEBP,
15812 DAG.getConstant(RegNodeSize, dl, PtrVT));
15813 return DAG.getNode(ISD::SUB, dl, PtrVT, RegNodeBase, RegNodeFrameOffset);
15816 static SDValue LowerINTRINSIC_WO_CHAIN(SDValue Op, const X86Subtarget *Subtarget,
15817 SelectionDAG &DAG) {
15819 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
15820 EVT VT = Op.getValueType();
15821 const IntrinsicData* IntrData = getIntrinsicWithoutChain(IntNo);
15823 switch(IntrData->Type) {
15824 case INTR_TYPE_1OP:
15825 return DAG.getNode(IntrData->Opc0, dl, Op.getValueType(), Op.getOperand(1));
15826 case INTR_TYPE_2OP:
15827 return DAG.getNode(IntrData->Opc0, dl, Op.getValueType(), Op.getOperand(1),
15829 case INTR_TYPE_2OP_IMM8:
15830 return DAG.getNode(IntrData->Opc0, dl, Op.getValueType(), Op.getOperand(1),
15831 DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op.getOperand(2)));
15832 case INTR_TYPE_3OP:
15833 return DAG.getNode(IntrData->Opc0, dl, Op.getValueType(), Op.getOperand(1),
15834 Op.getOperand(2), Op.getOperand(3));
15835 case INTR_TYPE_4OP:
15836 return DAG.getNode(IntrData->Opc0, dl, Op.getValueType(), Op.getOperand(1),
15837 Op.getOperand(2), Op.getOperand(3), Op.getOperand(4));
15838 case INTR_TYPE_1OP_MASK_RM: {
15839 SDValue Src = Op.getOperand(1);
15840 SDValue PassThru = Op.getOperand(2);
15841 SDValue Mask = Op.getOperand(3);
15842 SDValue RoundingMode;
15843 // We allways add rounding mode to the Node.
15844 // If the rounding mode is not specified, we add the
15845 // "current direction" mode.
15846 if (Op.getNumOperands() == 4)
15848 DAG.getConstant(X86::STATIC_ROUNDING::CUR_DIRECTION, dl, MVT::i32);
15850 RoundingMode = Op.getOperand(4);
15851 unsigned IntrWithRoundingModeOpcode = IntrData->Opc1;
15852 if (IntrWithRoundingModeOpcode != 0)
15853 if (cast<ConstantSDNode>(RoundingMode)->getZExtValue() !=
15854 X86::STATIC_ROUNDING::CUR_DIRECTION)
15855 return getVectorMaskingNode(DAG.getNode(IntrWithRoundingModeOpcode,
15856 dl, Op.getValueType(), Src, RoundingMode),
15857 Mask, PassThru, Subtarget, DAG);
15858 return getVectorMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT, Src,
15860 Mask, PassThru, Subtarget, DAG);
15862 case INTR_TYPE_1OP_MASK: {
15863 SDValue Src = Op.getOperand(1);
15864 SDValue PassThru = Op.getOperand(2);
15865 SDValue Mask = Op.getOperand(3);
15866 // We add rounding mode to the Node when
15867 // - RM Opcode is specified and
15868 // - RM is not "current direction".
15869 unsigned IntrWithRoundingModeOpcode = IntrData->Opc1;
15870 if (IntrWithRoundingModeOpcode != 0) {
15871 SDValue Rnd = Op.getOperand(4);
15872 unsigned Round = cast<ConstantSDNode>(Rnd)->getZExtValue();
15873 if (Round != X86::STATIC_ROUNDING::CUR_DIRECTION) {
15874 return getVectorMaskingNode(DAG.getNode(IntrWithRoundingModeOpcode,
15875 dl, Op.getValueType(),
15877 Mask, PassThru, Subtarget, DAG);
15880 return getVectorMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT, Src),
15881 Mask, PassThru, Subtarget, DAG);
15883 case INTR_TYPE_SCALAR_MASK_RM: {
15884 SDValue Src1 = Op.getOperand(1);
15885 SDValue Src2 = Op.getOperand(2);
15886 SDValue Src0 = Op.getOperand(3);
15887 SDValue Mask = Op.getOperand(4);
15888 // There are 2 kinds of intrinsics in this group:
15889 // (1) With suppress-all-exceptions (sae) or rounding mode- 6 operands
15890 // (2) With rounding mode and sae - 7 operands.
15891 if (Op.getNumOperands() == 6) {
15892 SDValue Sae = Op.getOperand(5);
15893 unsigned Opc = IntrData->Opc1 ? IntrData->Opc1 : IntrData->Opc0;
15894 return getScalarMaskingNode(DAG.getNode(Opc, dl, VT, Src1, Src2,
15896 Mask, Src0, Subtarget, DAG);
15898 assert(Op.getNumOperands() == 7 && "Unexpected intrinsic form");
15899 SDValue RoundingMode = Op.getOperand(5);
15900 SDValue Sae = Op.getOperand(6);
15901 return getScalarMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT, Src1, Src2,
15902 RoundingMode, Sae),
15903 Mask, Src0, Subtarget, DAG);
15905 case INTR_TYPE_2OP_MASK: {
15906 SDValue Src1 = Op.getOperand(1);
15907 SDValue Src2 = Op.getOperand(2);
15908 SDValue PassThru = Op.getOperand(3);
15909 SDValue Mask = Op.getOperand(4);
15910 // We specify 2 possible opcodes for intrinsics with rounding modes.
15911 // First, we check if the intrinsic may have non-default rounding mode,
15912 // (IntrData->Opc1 != 0), then we check the rounding mode operand.
15913 unsigned IntrWithRoundingModeOpcode = IntrData->Opc1;
15914 if (IntrWithRoundingModeOpcode != 0) {
15915 SDValue Rnd = Op.getOperand(5);
15916 unsigned Round = cast<ConstantSDNode>(Rnd)->getZExtValue();
15917 if (Round != X86::STATIC_ROUNDING::CUR_DIRECTION) {
15918 return getVectorMaskingNode(DAG.getNode(IntrWithRoundingModeOpcode,
15919 dl, Op.getValueType(),
15921 Mask, PassThru, Subtarget, DAG);
15924 // TODO: Intrinsics should have fast-math-flags to propagate.
15925 return getVectorMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT,Src1,Src2),
15926 Mask, PassThru, Subtarget, DAG);
15928 case INTR_TYPE_2OP_MASK_RM: {
15929 SDValue Src1 = Op.getOperand(1);
15930 SDValue Src2 = Op.getOperand(2);
15931 SDValue PassThru = Op.getOperand(3);
15932 SDValue Mask = Op.getOperand(4);
15933 // We specify 2 possible modes for intrinsics, with/without rounding modes.
15934 // First, we check if the intrinsic have rounding mode (6 operands),
15935 // if not, we set rounding mode to "current".
15937 if (Op.getNumOperands() == 6)
15938 Rnd = Op.getOperand(5);
15940 Rnd = DAG.getConstant(X86::STATIC_ROUNDING::CUR_DIRECTION, dl, MVT::i32);
15941 return getVectorMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT,
15943 Mask, PassThru, Subtarget, DAG);
15945 case INTR_TYPE_3OP_SCALAR_MASK_RM: {
15946 SDValue Src1 = Op.getOperand(1);
15947 SDValue Src2 = Op.getOperand(2);
15948 SDValue Src3 = Op.getOperand(3);
15949 SDValue PassThru = Op.getOperand(4);
15950 SDValue Mask = Op.getOperand(5);
15951 SDValue Sae = Op.getOperand(6);
15953 return getScalarMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT, Src1,
15955 Mask, PassThru, Subtarget, DAG);
15957 case INTR_TYPE_3OP_MASK_RM: {
15958 SDValue Src1 = Op.getOperand(1);
15959 SDValue Src2 = Op.getOperand(2);
15960 SDValue Imm = Op.getOperand(3);
15961 SDValue PassThru = Op.getOperand(4);
15962 SDValue Mask = Op.getOperand(5);
15963 // We specify 2 possible modes for intrinsics, with/without rounding modes.
15964 // First, we check if the intrinsic have rounding mode (7 operands),
15965 // if not, we set rounding mode to "current".
15967 if (Op.getNumOperands() == 7)
15968 Rnd = Op.getOperand(6);
15970 Rnd = DAG.getConstant(X86::STATIC_ROUNDING::CUR_DIRECTION, dl, MVT::i32);
15971 return getVectorMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT,
15972 Src1, Src2, Imm, Rnd),
15973 Mask, PassThru, Subtarget, DAG);
15975 case INTR_TYPE_3OP_IMM8_MASK:
15976 case INTR_TYPE_3OP_MASK: {
15977 SDValue Src1 = Op.getOperand(1);
15978 SDValue Src2 = Op.getOperand(2);
15979 SDValue Src3 = Op.getOperand(3);
15980 SDValue PassThru = Op.getOperand(4);
15981 SDValue Mask = Op.getOperand(5);
15983 if (IntrData->Type == INTR_TYPE_3OP_IMM8_MASK)
15984 Src3 = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Src3);
15985 // We specify 2 possible opcodes for intrinsics with rounding modes.
15986 // First, we check if the intrinsic may have non-default rounding mode,
15987 // (IntrData->Opc1 != 0), then we check the rounding mode operand.
15988 unsigned IntrWithRoundingModeOpcode = IntrData->Opc1;
15989 if (IntrWithRoundingModeOpcode != 0) {
15990 SDValue Rnd = Op.getOperand(6);
15991 unsigned Round = cast<ConstantSDNode>(Rnd)->getZExtValue();
15992 if (Round != X86::STATIC_ROUNDING::CUR_DIRECTION) {
15993 return getVectorMaskingNode(DAG.getNode(IntrWithRoundingModeOpcode,
15994 dl, Op.getValueType(),
15995 Src1, Src2, Src3, Rnd),
15996 Mask, PassThru, Subtarget, DAG);
15999 return getVectorMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT,
16001 Mask, PassThru, Subtarget, DAG);
16003 case VPERM_3OP_MASKZ:
16004 case VPERM_3OP_MASK:
16007 case FMA_OP_MASK: {
16008 SDValue Src1 = Op.getOperand(1);
16009 SDValue Src2 = Op.getOperand(2);
16010 SDValue Src3 = Op.getOperand(3);
16011 SDValue Mask = Op.getOperand(4);
16012 EVT VT = Op.getValueType();
16013 SDValue PassThru = SDValue();
16015 // set PassThru element
16016 if (IntrData->Type == VPERM_3OP_MASKZ || IntrData->Type == FMA_OP_MASKZ)
16017 PassThru = getZeroVector(VT, Subtarget, DAG, dl);
16018 else if (IntrData->Type == FMA_OP_MASK3)
16023 // We specify 2 possible opcodes for intrinsics with rounding modes.
16024 // First, we check if the intrinsic may have non-default rounding mode,
16025 // (IntrData->Opc1 != 0), then we check the rounding mode operand.
16026 unsigned IntrWithRoundingModeOpcode = IntrData->Opc1;
16027 if (IntrWithRoundingModeOpcode != 0) {
16028 SDValue Rnd = Op.getOperand(5);
16029 if (cast<ConstantSDNode>(Rnd)->getZExtValue() !=
16030 X86::STATIC_ROUNDING::CUR_DIRECTION)
16031 return getVectorMaskingNode(DAG.getNode(IntrWithRoundingModeOpcode,
16032 dl, Op.getValueType(),
16033 Src1, Src2, Src3, Rnd),
16034 Mask, PassThru, Subtarget, DAG);
16036 return getVectorMaskingNode(DAG.getNode(IntrData->Opc0,
16037 dl, Op.getValueType(),
16039 Mask, PassThru, Subtarget, DAG);
16042 case CMP_MASK_CC: {
16043 // Comparison intrinsics with masks.
16044 // Example of transformation:
16045 // (i8 (int_x86_avx512_mask_pcmpeq_q_128
16046 // (v2i64 %a), (v2i64 %b), (i8 %mask))) ->
16048 // (v8i1 (insert_subvector undef,
16049 // (v2i1 (and (PCMPEQM %a, %b),
16050 // (extract_subvector
16051 // (v8i1 (bitcast %mask)), 0))), 0))))
16052 EVT VT = Op.getOperand(1).getValueType();
16053 EVT MaskVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
16054 VT.getVectorNumElements());
16055 SDValue Mask = Op.getOperand((IntrData->Type == CMP_MASK_CC) ? 4 : 3);
16056 EVT BitcastVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
16057 Mask.getValueType().getSizeInBits());
16059 if (IntrData->Type == CMP_MASK_CC) {
16060 SDValue CC = Op.getOperand(3);
16061 CC = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, CC);
16062 // We specify 2 possible opcodes for intrinsics with rounding modes.
16063 // First, we check if the intrinsic may have non-default rounding mode,
16064 // (IntrData->Opc1 != 0), then we check the rounding mode operand.
16065 if (IntrData->Opc1 != 0) {
16066 SDValue Rnd = Op.getOperand(5);
16067 if (cast<ConstantSDNode>(Rnd)->getZExtValue() !=
16068 X86::STATIC_ROUNDING::CUR_DIRECTION)
16069 Cmp = DAG.getNode(IntrData->Opc1, dl, MaskVT, Op.getOperand(1),
16070 Op.getOperand(2), CC, Rnd);
16072 //default rounding mode
16074 Cmp = DAG.getNode(IntrData->Opc0, dl, MaskVT, Op.getOperand(1),
16075 Op.getOperand(2), CC);
16078 assert(IntrData->Type == CMP_MASK && "Unexpected intrinsic type!");
16079 Cmp = DAG.getNode(IntrData->Opc0, dl, MaskVT, Op.getOperand(1),
16082 SDValue CmpMask = getVectorMaskingNode(Cmp, Mask,
16083 DAG.getTargetConstant(0, dl,
16086 SDValue Res = DAG.getNode(ISD::INSERT_SUBVECTOR, dl, BitcastVT,
16087 DAG.getUNDEF(BitcastVT), CmpMask,
16088 DAG.getIntPtrConstant(0, dl));
16089 return DAG.getBitcast(Op.getValueType(), Res);
16091 case COMI: { // Comparison intrinsics
16092 ISD::CondCode CC = (ISD::CondCode)IntrData->Opc1;
16093 SDValue LHS = Op.getOperand(1);
16094 SDValue RHS = Op.getOperand(2);
16095 unsigned X86CC = TranslateX86CC(CC, dl, true, LHS, RHS, DAG);
16096 assert(X86CC != X86::COND_INVALID && "Unexpected illegal condition!");
16097 SDValue Cond = DAG.getNode(IntrData->Opc0, dl, MVT::i32, LHS, RHS);
16098 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
16099 DAG.getConstant(X86CC, dl, MVT::i8), Cond);
16100 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
16103 return getTargetVShiftNode(IntrData->Opc0, dl, Op.getSimpleValueType(),
16104 Op.getOperand(1), Op.getOperand(2), DAG);
16106 return getVectorMaskingNode(getTargetVShiftNode(IntrData->Opc0, dl,
16107 Op.getSimpleValueType(),
16109 Op.getOperand(2), DAG),
16110 Op.getOperand(4), Op.getOperand(3), Subtarget,
16112 case COMPRESS_EXPAND_IN_REG: {
16113 SDValue Mask = Op.getOperand(3);
16114 SDValue DataToCompress = Op.getOperand(1);
16115 SDValue PassThru = Op.getOperand(2);
16116 if (isAllOnes(Mask)) // return data as is
16117 return Op.getOperand(1);
16119 return getVectorMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT,
16121 Mask, PassThru, Subtarget, DAG);
16124 SDValue Mask = Op.getOperand(3);
16125 EVT VT = Op.getValueType();
16126 EVT MaskVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
16127 VT.getVectorNumElements());
16128 EVT BitcastVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
16129 Mask.getValueType().getSizeInBits());
16131 SDValue VMask = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MaskVT,
16132 DAG.getBitcast(BitcastVT, Mask),
16133 DAG.getIntPtrConstant(0, dl));
16134 return DAG.getNode(IntrData->Opc0, dl, VT, VMask, Op.getOperand(1),
16143 default: return SDValue(); // Don't custom lower most intrinsics.
16145 case Intrinsic::x86_avx2_permd:
16146 case Intrinsic::x86_avx2_permps:
16147 // Operands intentionally swapped. Mask is last operand to intrinsic,
16148 // but second operand for node/instruction.
16149 return DAG.getNode(X86ISD::VPERMV, dl, Op.getValueType(),
16150 Op.getOperand(2), Op.getOperand(1));
16152 // ptest and testp intrinsics. The intrinsic these come from are designed to
16153 // return an integer value, not just an instruction so lower it to the ptest
16154 // or testp pattern and a setcc for the result.
16155 case Intrinsic::x86_sse41_ptestz:
16156 case Intrinsic::x86_sse41_ptestc:
16157 case Intrinsic::x86_sse41_ptestnzc:
16158 case Intrinsic::x86_avx_ptestz_256:
16159 case Intrinsic::x86_avx_ptestc_256:
16160 case Intrinsic::x86_avx_ptestnzc_256:
16161 case Intrinsic::x86_avx_vtestz_ps:
16162 case Intrinsic::x86_avx_vtestc_ps:
16163 case Intrinsic::x86_avx_vtestnzc_ps:
16164 case Intrinsic::x86_avx_vtestz_pd:
16165 case Intrinsic::x86_avx_vtestc_pd:
16166 case Intrinsic::x86_avx_vtestnzc_pd:
16167 case Intrinsic::x86_avx_vtestz_ps_256:
16168 case Intrinsic::x86_avx_vtestc_ps_256:
16169 case Intrinsic::x86_avx_vtestnzc_ps_256:
16170 case Intrinsic::x86_avx_vtestz_pd_256:
16171 case Intrinsic::x86_avx_vtestc_pd_256:
16172 case Intrinsic::x86_avx_vtestnzc_pd_256: {
16173 bool IsTestPacked = false;
16176 default: llvm_unreachable("Bad fallthrough in Intrinsic lowering.");
16177 case Intrinsic::x86_avx_vtestz_ps:
16178 case Intrinsic::x86_avx_vtestz_pd:
16179 case Intrinsic::x86_avx_vtestz_ps_256:
16180 case Intrinsic::x86_avx_vtestz_pd_256:
16181 IsTestPacked = true; // Fallthrough
16182 case Intrinsic::x86_sse41_ptestz:
16183 case Intrinsic::x86_avx_ptestz_256:
16185 X86CC = X86::COND_E;
16187 case Intrinsic::x86_avx_vtestc_ps:
16188 case Intrinsic::x86_avx_vtestc_pd:
16189 case Intrinsic::x86_avx_vtestc_ps_256:
16190 case Intrinsic::x86_avx_vtestc_pd_256:
16191 IsTestPacked = true; // Fallthrough
16192 case Intrinsic::x86_sse41_ptestc:
16193 case Intrinsic::x86_avx_ptestc_256:
16195 X86CC = X86::COND_B;
16197 case Intrinsic::x86_avx_vtestnzc_ps:
16198 case Intrinsic::x86_avx_vtestnzc_pd:
16199 case Intrinsic::x86_avx_vtestnzc_ps_256:
16200 case Intrinsic::x86_avx_vtestnzc_pd_256:
16201 IsTestPacked = true; // Fallthrough
16202 case Intrinsic::x86_sse41_ptestnzc:
16203 case Intrinsic::x86_avx_ptestnzc_256:
16205 X86CC = X86::COND_A;
16209 SDValue LHS = Op.getOperand(1);
16210 SDValue RHS = Op.getOperand(2);
16211 unsigned TestOpc = IsTestPacked ? X86ISD::TESTP : X86ISD::PTEST;
16212 SDValue Test = DAG.getNode(TestOpc, dl, MVT::i32, LHS, RHS);
16213 SDValue CC = DAG.getConstant(X86CC, dl, MVT::i8);
16214 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8, CC, Test);
16215 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
16217 case Intrinsic::x86_avx512_kortestz_w:
16218 case Intrinsic::x86_avx512_kortestc_w: {
16219 unsigned X86CC = (IntNo == Intrinsic::x86_avx512_kortestz_w)? X86::COND_E: X86::COND_B;
16220 SDValue LHS = DAG.getBitcast(MVT::v16i1, Op.getOperand(1));
16221 SDValue RHS = DAG.getBitcast(MVT::v16i1, Op.getOperand(2));
16222 SDValue CC = DAG.getConstant(X86CC, dl, MVT::i8);
16223 SDValue Test = DAG.getNode(X86ISD::KORTEST, dl, MVT::i32, LHS, RHS);
16224 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i1, CC, Test);
16225 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
16228 case Intrinsic::x86_sse42_pcmpistria128:
16229 case Intrinsic::x86_sse42_pcmpestria128:
16230 case Intrinsic::x86_sse42_pcmpistric128:
16231 case Intrinsic::x86_sse42_pcmpestric128:
16232 case Intrinsic::x86_sse42_pcmpistrio128:
16233 case Intrinsic::x86_sse42_pcmpestrio128:
16234 case Intrinsic::x86_sse42_pcmpistris128:
16235 case Intrinsic::x86_sse42_pcmpestris128:
16236 case Intrinsic::x86_sse42_pcmpistriz128:
16237 case Intrinsic::x86_sse42_pcmpestriz128: {
16241 default: llvm_unreachable("Impossible intrinsic"); // Can't reach here.
16242 case Intrinsic::x86_sse42_pcmpistria128:
16243 Opcode = X86ISD::PCMPISTRI;
16244 X86CC = X86::COND_A;
16246 case Intrinsic::x86_sse42_pcmpestria128:
16247 Opcode = X86ISD::PCMPESTRI;
16248 X86CC = X86::COND_A;
16250 case Intrinsic::x86_sse42_pcmpistric128:
16251 Opcode = X86ISD::PCMPISTRI;
16252 X86CC = X86::COND_B;
16254 case Intrinsic::x86_sse42_pcmpestric128:
16255 Opcode = X86ISD::PCMPESTRI;
16256 X86CC = X86::COND_B;
16258 case Intrinsic::x86_sse42_pcmpistrio128:
16259 Opcode = X86ISD::PCMPISTRI;
16260 X86CC = X86::COND_O;
16262 case Intrinsic::x86_sse42_pcmpestrio128:
16263 Opcode = X86ISD::PCMPESTRI;
16264 X86CC = X86::COND_O;
16266 case Intrinsic::x86_sse42_pcmpistris128:
16267 Opcode = X86ISD::PCMPISTRI;
16268 X86CC = X86::COND_S;
16270 case Intrinsic::x86_sse42_pcmpestris128:
16271 Opcode = X86ISD::PCMPESTRI;
16272 X86CC = X86::COND_S;
16274 case Intrinsic::x86_sse42_pcmpistriz128:
16275 Opcode = X86ISD::PCMPISTRI;
16276 X86CC = X86::COND_E;
16278 case Intrinsic::x86_sse42_pcmpestriz128:
16279 Opcode = X86ISD::PCMPESTRI;
16280 X86CC = X86::COND_E;
16283 SmallVector<SDValue, 5> NewOps(Op->op_begin()+1, Op->op_end());
16284 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::i32);
16285 SDValue PCMP = DAG.getNode(Opcode, dl, VTs, NewOps);
16286 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
16287 DAG.getConstant(X86CC, dl, MVT::i8),
16288 SDValue(PCMP.getNode(), 1));
16289 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
16292 case Intrinsic::x86_sse42_pcmpistri128:
16293 case Intrinsic::x86_sse42_pcmpestri128: {
16295 if (IntNo == Intrinsic::x86_sse42_pcmpistri128)
16296 Opcode = X86ISD::PCMPISTRI;
16298 Opcode = X86ISD::PCMPESTRI;
16300 SmallVector<SDValue, 5> NewOps(Op->op_begin()+1, Op->op_end());
16301 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::i32);
16302 return DAG.getNode(Opcode, dl, VTs, NewOps);
16305 case Intrinsic::x86_seh_lsda: {
16306 // Compute the symbol for the LSDA. We know it'll get emitted later.
16307 MachineFunction &MF = DAG.getMachineFunction();
16308 SDValue Op1 = Op.getOperand(1);
16309 auto *Fn = cast<Function>(cast<GlobalAddressSDNode>(Op1)->getGlobal());
16310 MCSymbol *LSDASym = MF.getMMI().getContext().getOrCreateLSDASymbol(
16311 GlobalValue::getRealLinkageName(Fn->getName()));
16313 // Generate a simple absolute symbol reference. This intrinsic is only
16314 // supported on 32-bit Windows, which isn't PIC.
16315 SDValue Result = DAG.getMCSymbol(LSDASym, VT);
16316 return DAG.getNode(X86ISD::Wrapper, dl, VT, Result);
16319 case Intrinsic::x86_seh_recoverfp: {
16320 SDValue FnOp = Op.getOperand(1);
16321 SDValue IncomingFPOp = Op.getOperand(2);
16322 GlobalAddressSDNode *GSD = dyn_cast<GlobalAddressSDNode>(FnOp);
16323 auto *Fn = dyn_cast_or_null<Function>(GSD ? GSD->getGlobal() : nullptr);
16325 report_fatal_error(
16326 "llvm.x86.seh.recoverfp must take a function as the first argument");
16327 return recoverFramePointer(DAG, Fn, IncomingFPOp);
16330 case Intrinsic::localaddress: {
16331 // Returns one of the stack, base, or frame pointer registers, depending on
16332 // which is used to reference local variables.
16333 MachineFunction &MF = DAG.getMachineFunction();
16334 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
16336 if (RegInfo->hasBasePointer(MF))
16337 Reg = RegInfo->getBaseRegister();
16338 else // This function handles the SP or FP case.
16339 Reg = RegInfo->getPtrSizedFrameRegister(MF);
16340 return DAG.getCopyFromReg(DAG.getEntryNode(), dl, Reg, VT);
16345 static SDValue getGatherNode(unsigned Opc, SDValue Op, SelectionDAG &DAG,
16346 SDValue Src, SDValue Mask, SDValue Base,
16347 SDValue Index, SDValue ScaleOp, SDValue Chain,
16348 const X86Subtarget * Subtarget) {
16350 ConstantSDNode *C = dyn_cast<ConstantSDNode>(ScaleOp);
16352 llvm_unreachable("Invalid scale type");
16353 unsigned ScaleVal = C->getZExtValue();
16354 if (ScaleVal > 2 && ScaleVal != 4 && ScaleVal != 8)
16355 llvm_unreachable("Valid scale values are 1, 2, 4, 8");
16357 SDValue Scale = DAG.getTargetConstant(C->getZExtValue(), dl, MVT::i8);
16358 EVT MaskVT = MVT::getVectorVT(MVT::i1,
16359 Index.getSimpleValueType().getVectorNumElements());
16361 ConstantSDNode *MaskC = dyn_cast<ConstantSDNode>(Mask);
16363 MaskInReg = DAG.getTargetConstant(MaskC->getSExtValue(), dl, MaskVT);
16365 EVT BitcastVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
16366 Mask.getValueType().getSizeInBits());
16368 // In case when MaskVT equals v2i1 or v4i1, low 2 or 4 elements
16369 // are extracted by EXTRACT_SUBVECTOR.
16370 MaskInReg = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MaskVT,
16371 DAG.getBitcast(BitcastVT, Mask),
16372 DAG.getIntPtrConstant(0, dl));
16374 SDVTList VTs = DAG.getVTList(Op.getValueType(), MaskVT, MVT::Other);
16375 SDValue Disp = DAG.getTargetConstant(0, dl, MVT::i32);
16376 SDValue Segment = DAG.getRegister(0, MVT::i32);
16377 if (Src.getOpcode() == ISD::UNDEF)
16378 Src = getZeroVector(Op.getValueType(), Subtarget, DAG, dl);
16379 SDValue Ops[] = {Src, MaskInReg, Base, Scale, Index, Disp, Segment, Chain};
16380 SDNode *Res = DAG.getMachineNode(Opc, dl, VTs, Ops);
16381 SDValue RetOps[] = { SDValue(Res, 0), SDValue(Res, 2) };
16382 return DAG.getMergeValues(RetOps, dl);
16385 static SDValue getScatterNode(unsigned Opc, SDValue Op, SelectionDAG &DAG,
16386 SDValue Src, SDValue Mask, SDValue Base,
16387 SDValue Index, SDValue ScaleOp, SDValue Chain) {
16389 ConstantSDNode *C = dyn_cast<ConstantSDNode>(ScaleOp);
16391 llvm_unreachable("Invalid scale type");
16392 unsigned ScaleVal = C->getZExtValue();
16393 if (ScaleVal > 2 && ScaleVal != 4 && ScaleVal != 8)
16394 llvm_unreachable("Valid scale values are 1, 2, 4, 8");
16396 SDValue Scale = DAG.getTargetConstant(C->getZExtValue(), dl, MVT::i8);
16397 SDValue Disp = DAG.getTargetConstant(0, dl, MVT::i32);
16398 SDValue Segment = DAG.getRegister(0, MVT::i32);
16399 EVT MaskVT = MVT::getVectorVT(MVT::i1,
16400 Index.getSimpleValueType().getVectorNumElements());
16402 ConstantSDNode *MaskC = dyn_cast<ConstantSDNode>(Mask);
16404 MaskInReg = DAG.getTargetConstant(MaskC->getSExtValue(), dl, MaskVT);
16406 EVT BitcastVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
16407 Mask.getValueType().getSizeInBits());
16409 // In case when MaskVT equals v2i1 or v4i1, low 2 or 4 elements
16410 // are extracted by EXTRACT_SUBVECTOR.
16411 MaskInReg = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MaskVT,
16412 DAG.getBitcast(BitcastVT, Mask),
16413 DAG.getIntPtrConstant(0, dl));
16415 SDVTList VTs = DAG.getVTList(MaskVT, MVT::Other);
16416 SDValue Ops[] = {Base, Scale, Index, Disp, Segment, MaskInReg, Src, Chain};
16417 SDNode *Res = DAG.getMachineNode(Opc, dl, VTs, Ops);
16418 return SDValue(Res, 1);
16421 static SDValue getPrefetchNode(unsigned Opc, SDValue Op, SelectionDAG &DAG,
16422 SDValue Mask, SDValue Base, SDValue Index,
16423 SDValue ScaleOp, SDValue Chain) {
16425 ConstantSDNode *C = dyn_cast<ConstantSDNode>(ScaleOp);
16426 assert(C && "Invalid scale type");
16427 SDValue Scale = DAG.getTargetConstant(C->getZExtValue(), dl, MVT::i8);
16428 SDValue Disp = DAG.getTargetConstant(0, dl, MVT::i32);
16429 SDValue Segment = DAG.getRegister(0, MVT::i32);
16431 MVT::getVectorVT(MVT::i1, Index.getSimpleValueType().getVectorNumElements());
16433 ConstantSDNode *MaskC = dyn_cast<ConstantSDNode>(Mask);
16435 MaskInReg = DAG.getTargetConstant(MaskC->getSExtValue(), dl, MaskVT);
16437 MaskInReg = DAG.getBitcast(MaskVT, Mask);
16438 //SDVTList VTs = DAG.getVTList(MVT::Other);
16439 SDValue Ops[] = {MaskInReg, Base, Scale, Index, Disp, Segment, Chain};
16440 SDNode *Res = DAG.getMachineNode(Opc, dl, MVT::Other, Ops);
16441 return SDValue(Res, 0);
16444 // getReadPerformanceCounter - Handles the lowering of builtin intrinsics that
16445 // read performance monitor counters (x86_rdpmc).
16446 static void getReadPerformanceCounter(SDNode *N, SDLoc DL,
16447 SelectionDAG &DAG, const X86Subtarget *Subtarget,
16448 SmallVectorImpl<SDValue> &Results) {
16449 assert(N->getNumOperands() == 3 && "Unexpected number of operands!");
16450 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Glue);
16453 // The ECX register is used to select the index of the performance counter
16455 SDValue Chain = DAG.getCopyToReg(N->getOperand(0), DL, X86::ECX,
16457 SDValue rd = DAG.getNode(X86ISD::RDPMC_DAG, DL, Tys, Chain);
16459 // Reads the content of a 64-bit performance counter and returns it in the
16460 // registers EDX:EAX.
16461 if (Subtarget->is64Bit()) {
16462 LO = DAG.getCopyFromReg(rd, DL, X86::RAX, MVT::i64, rd.getValue(1));
16463 HI = DAG.getCopyFromReg(LO.getValue(1), DL, X86::RDX, MVT::i64,
16466 LO = DAG.getCopyFromReg(rd, DL, X86::EAX, MVT::i32, rd.getValue(1));
16467 HI = DAG.getCopyFromReg(LO.getValue(1), DL, X86::EDX, MVT::i32,
16470 Chain = HI.getValue(1);
16472 if (Subtarget->is64Bit()) {
16473 // The EAX register is loaded with the low-order 32 bits. The EDX register
16474 // is loaded with the supported high-order bits of the counter.
16475 SDValue Tmp = DAG.getNode(ISD::SHL, DL, MVT::i64, HI,
16476 DAG.getConstant(32, DL, MVT::i8));
16477 Results.push_back(DAG.getNode(ISD::OR, DL, MVT::i64, LO, Tmp));
16478 Results.push_back(Chain);
16482 // Use a buildpair to merge the two 32-bit values into a 64-bit one.
16483 SDValue Ops[] = { LO, HI };
16484 SDValue Pair = DAG.getNode(ISD::BUILD_PAIR, DL, MVT::i64, Ops);
16485 Results.push_back(Pair);
16486 Results.push_back(Chain);
16489 // getReadTimeStampCounter - Handles the lowering of builtin intrinsics that
16490 // read the time stamp counter (x86_rdtsc and x86_rdtscp). This function is
16491 // also used to custom lower READCYCLECOUNTER nodes.
16492 static void getReadTimeStampCounter(SDNode *N, SDLoc DL, unsigned Opcode,
16493 SelectionDAG &DAG, const X86Subtarget *Subtarget,
16494 SmallVectorImpl<SDValue> &Results) {
16495 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Glue);
16496 SDValue rd = DAG.getNode(Opcode, DL, Tys, N->getOperand(0));
16499 // The processor's time-stamp counter (a 64-bit MSR) is stored into the
16500 // EDX:EAX registers. EDX is loaded with the high-order 32 bits of the MSR
16501 // and the EAX register is loaded with the low-order 32 bits.
16502 if (Subtarget->is64Bit()) {
16503 LO = DAG.getCopyFromReg(rd, DL, X86::RAX, MVT::i64, rd.getValue(1));
16504 HI = DAG.getCopyFromReg(LO.getValue(1), DL, X86::RDX, MVT::i64,
16507 LO = DAG.getCopyFromReg(rd, DL, X86::EAX, MVT::i32, rd.getValue(1));
16508 HI = DAG.getCopyFromReg(LO.getValue(1), DL, X86::EDX, MVT::i32,
16511 SDValue Chain = HI.getValue(1);
16513 if (Opcode == X86ISD::RDTSCP_DAG) {
16514 assert(N->getNumOperands() == 3 && "Unexpected number of operands!");
16516 // Instruction RDTSCP loads the IA32:TSC_AUX_MSR (address C000_0103H) into
16517 // the ECX register. Add 'ecx' explicitly to the chain.
16518 SDValue ecx = DAG.getCopyFromReg(Chain, DL, X86::ECX, MVT::i32,
16520 // Explicitly store the content of ECX at the location passed in input
16521 // to the 'rdtscp' intrinsic.
16522 Chain = DAG.getStore(ecx.getValue(1), DL, ecx, N->getOperand(2),
16523 MachinePointerInfo(), false, false, 0);
16526 if (Subtarget->is64Bit()) {
16527 // The EDX register is loaded with the high-order 32 bits of the MSR, and
16528 // the EAX register is loaded with the low-order 32 bits.
16529 SDValue Tmp = DAG.getNode(ISD::SHL, DL, MVT::i64, HI,
16530 DAG.getConstant(32, DL, MVT::i8));
16531 Results.push_back(DAG.getNode(ISD::OR, DL, MVT::i64, LO, Tmp));
16532 Results.push_back(Chain);
16536 // Use a buildpair to merge the two 32-bit values into a 64-bit one.
16537 SDValue Ops[] = { LO, HI };
16538 SDValue Pair = DAG.getNode(ISD::BUILD_PAIR, DL, MVT::i64, Ops);
16539 Results.push_back(Pair);
16540 Results.push_back(Chain);
16543 static SDValue LowerREADCYCLECOUNTER(SDValue Op, const X86Subtarget *Subtarget,
16544 SelectionDAG &DAG) {
16545 SmallVector<SDValue, 2> Results;
16547 getReadTimeStampCounter(Op.getNode(), DL, X86ISD::RDTSC_DAG, DAG, Subtarget,
16549 return DAG.getMergeValues(Results, DL);
16552 static SDValue LowerSEHRESTOREFRAME(SDValue Op, const X86Subtarget *Subtarget,
16553 SelectionDAG &DAG) {
16554 MachineFunction &MF = DAG.getMachineFunction();
16555 const Function *Fn = MF.getFunction();
16557 SDValue Chain = Op.getOperand(0);
16559 assert(Subtarget->getFrameLowering()->hasFP(MF) &&
16560 "using llvm.x86.seh.restoreframe requires a frame pointer");
16562 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
16563 MVT VT = TLI.getPointerTy(DAG.getDataLayout());
16565 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
16566 unsigned FrameReg =
16567 RegInfo->getPtrSizedFrameRegister(DAG.getMachineFunction());
16568 unsigned SPReg = RegInfo->getStackRegister();
16569 unsigned SlotSize = RegInfo->getSlotSize();
16571 // Get incoming EBP.
16572 SDValue IncomingEBP =
16573 DAG.getCopyFromReg(Chain, dl, FrameReg, VT);
16575 // SP is saved in the first field of every registration node, so load
16576 // [EBP-RegNodeSize] into SP.
16577 int RegNodeSize = getSEHRegistrationNodeSize(Fn);
16578 SDValue SPAddr = DAG.getNode(ISD::ADD, dl, VT, IncomingEBP,
16579 DAG.getConstant(-RegNodeSize, dl, VT));
16581 DAG.getLoad(VT, dl, Chain, SPAddr, MachinePointerInfo(), false, false,
16582 false, VT.getScalarSizeInBits() / 8);
16583 Chain = DAG.getCopyToReg(Chain, dl, SPReg, NewSP);
16585 if (!RegInfo->needsStackRealignment(MF)) {
16586 // Adjust EBP to point back to the original frame position.
16587 SDValue NewFP = recoverFramePointer(DAG, Fn, IncomingEBP);
16588 Chain = DAG.getCopyToReg(Chain, dl, FrameReg, NewFP);
16590 assert(RegInfo->hasBasePointer(MF) &&
16591 "functions with Win32 EH must use frame or base pointer register");
16593 // Reload the base pointer (ESI) with the adjusted incoming EBP.
16594 SDValue NewBP = recoverFramePointer(DAG, Fn, IncomingEBP);
16595 Chain = DAG.getCopyToReg(Chain, dl, RegInfo->getBaseRegister(), NewBP);
16597 // Reload the spilled EBP value, now that the stack and base pointers are
16599 X86MachineFunctionInfo *X86FI = MF.getInfo<X86MachineFunctionInfo>();
16600 X86FI->setHasSEHFramePtrSave(true);
16601 int FI = MF.getFrameInfo()->CreateSpillStackObject(SlotSize, SlotSize);
16602 X86FI->setSEHFramePtrSaveIndex(FI);
16603 SDValue NewFP = DAG.getLoad(VT, dl, Chain, DAG.getFrameIndex(FI, VT),
16604 MachinePointerInfo(), false, false, false,
16605 VT.getScalarSizeInBits() / 8);
16606 Chain = DAG.getCopyToReg(NewFP, dl, FrameReg, NewFP);
16612 /// \brief Lower intrinsics for TRUNCATE_TO_MEM case
16613 /// return truncate Store/MaskedStore Node
16614 static SDValue LowerINTRINSIC_TRUNCATE_TO_MEM(const SDValue & Op,
16618 SDValue Mask = Op.getOperand(4);
16619 SDValue DataToTruncate = Op.getOperand(3);
16620 SDValue Addr = Op.getOperand(2);
16621 SDValue Chain = Op.getOperand(0);
16623 EVT VT = DataToTruncate.getValueType();
16624 EVT SVT = EVT::getVectorVT(*DAG.getContext(),
16625 ElementType, VT.getVectorNumElements());
16627 if (isAllOnes(Mask)) // return just a truncate store
16628 return DAG.getTruncStore(Chain, dl, DataToTruncate, Addr,
16629 MachinePointerInfo(), SVT, false, false,
16630 SVT.getScalarSizeInBits()/8);
16632 EVT MaskVT = EVT::getVectorVT(*DAG.getContext(),
16633 MVT::i1, VT.getVectorNumElements());
16634 EVT BitcastVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
16635 Mask.getValueType().getSizeInBits());
16636 // In case when MaskVT equals v2i1 or v4i1, low 2 or 4 elements
16637 // are extracted by EXTRACT_SUBVECTOR.
16638 SDValue VMask = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MaskVT,
16639 DAG.getBitcast(BitcastVT, Mask),
16640 DAG.getIntPtrConstant(0, dl));
16642 MachineMemOperand *MMO = DAG.getMachineFunction().
16643 getMachineMemOperand(MachinePointerInfo(),
16644 MachineMemOperand::MOStore, SVT.getStoreSize(),
16645 SVT.getScalarSizeInBits()/8);
16647 return DAG.getMaskedStore(Chain, dl, DataToTruncate, Addr,
16648 VMask, SVT, MMO, true);
16651 static SDValue LowerINTRINSIC_W_CHAIN(SDValue Op, const X86Subtarget *Subtarget,
16652 SelectionDAG &DAG) {
16653 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
16655 const IntrinsicData* IntrData = getIntrinsicWithChain(IntNo);
16657 if (IntNo == llvm::Intrinsic::x86_seh_restoreframe)
16658 return LowerSEHRESTOREFRAME(Op, Subtarget, DAG);
16663 switch(IntrData->Type) {
16665 llvm_unreachable("Unknown Intrinsic Type");
16669 // Emit the node with the right value type.
16670 SDVTList VTs = DAG.getVTList(Op->getValueType(0), MVT::Glue, MVT::Other);
16671 SDValue Result = DAG.getNode(IntrData->Opc0, dl, VTs, Op.getOperand(0));
16673 // If the value returned by RDRAND/RDSEED was valid (CF=1), return 1.
16674 // Otherwise return the value from Rand, which is always 0, casted to i32.
16675 SDValue Ops[] = { DAG.getZExtOrTrunc(Result, dl, Op->getValueType(1)),
16676 DAG.getConstant(1, dl, Op->getValueType(1)),
16677 DAG.getConstant(X86::COND_B, dl, MVT::i32),
16678 SDValue(Result.getNode(), 1) };
16679 SDValue isValid = DAG.getNode(X86ISD::CMOV, dl,
16680 DAG.getVTList(Op->getValueType(1), MVT::Glue),
16683 // Return { result, isValid, chain }.
16684 return DAG.getNode(ISD::MERGE_VALUES, dl, Op->getVTList(), Result, isValid,
16685 SDValue(Result.getNode(), 2));
16688 //gather(v1, mask, index, base, scale);
16689 SDValue Chain = Op.getOperand(0);
16690 SDValue Src = Op.getOperand(2);
16691 SDValue Base = Op.getOperand(3);
16692 SDValue Index = Op.getOperand(4);
16693 SDValue Mask = Op.getOperand(5);
16694 SDValue Scale = Op.getOperand(6);
16695 return getGatherNode(IntrData->Opc0, Op, DAG, Src, Mask, Base, Index, Scale,
16699 //scatter(base, mask, index, v1, scale);
16700 SDValue Chain = Op.getOperand(0);
16701 SDValue Base = Op.getOperand(2);
16702 SDValue Mask = Op.getOperand(3);
16703 SDValue Index = Op.getOperand(4);
16704 SDValue Src = Op.getOperand(5);
16705 SDValue Scale = Op.getOperand(6);
16706 return getScatterNode(IntrData->Opc0, Op, DAG, Src, Mask, Base, Index,
16710 SDValue Hint = Op.getOperand(6);
16711 unsigned HintVal = cast<ConstantSDNode>(Hint)->getZExtValue();
16712 assert(HintVal < 2 && "Wrong prefetch hint in intrinsic: should be 0 or 1");
16713 unsigned Opcode = (HintVal ? IntrData->Opc1 : IntrData->Opc0);
16714 SDValue Chain = Op.getOperand(0);
16715 SDValue Mask = Op.getOperand(2);
16716 SDValue Index = Op.getOperand(3);
16717 SDValue Base = Op.getOperand(4);
16718 SDValue Scale = Op.getOperand(5);
16719 return getPrefetchNode(Opcode, Op, DAG, Mask, Base, Index, Scale, Chain);
16721 // Read Time Stamp Counter (RDTSC) and Processor ID (RDTSCP).
16723 SmallVector<SDValue, 2> Results;
16724 getReadTimeStampCounter(Op.getNode(), dl, IntrData->Opc0, DAG, Subtarget,
16726 return DAG.getMergeValues(Results, dl);
16728 // Read Performance Monitoring Counters.
16730 SmallVector<SDValue, 2> Results;
16731 getReadPerformanceCounter(Op.getNode(), dl, DAG, Subtarget, Results);
16732 return DAG.getMergeValues(Results, dl);
16734 // XTEST intrinsics.
16736 SDVTList VTs = DAG.getVTList(Op->getValueType(0), MVT::Other);
16737 SDValue InTrans = DAG.getNode(IntrData->Opc0, dl, VTs, Op.getOperand(0));
16738 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
16739 DAG.getConstant(X86::COND_NE, dl, MVT::i8),
16741 SDValue Ret = DAG.getNode(ISD::ZERO_EXTEND, dl, Op->getValueType(0), SetCC);
16742 return DAG.getNode(ISD::MERGE_VALUES, dl, Op->getVTList(),
16743 Ret, SDValue(InTrans.getNode(), 1));
16747 SmallVector<SDValue, 2> Results;
16748 SDVTList CFVTs = DAG.getVTList(Op->getValueType(0), MVT::Other);
16749 SDVTList VTs = DAG.getVTList(Op.getOperand(3)->getValueType(0), MVT::Other);
16750 SDValue GenCF = DAG.getNode(X86ISD::ADD, dl, CFVTs, Op.getOperand(2),
16751 DAG.getConstant(-1, dl, MVT::i8));
16752 SDValue Res = DAG.getNode(IntrData->Opc0, dl, VTs, Op.getOperand(3),
16753 Op.getOperand(4), GenCF.getValue(1));
16754 SDValue Store = DAG.getStore(Op.getOperand(0), dl, Res.getValue(0),
16755 Op.getOperand(5), MachinePointerInfo(),
16757 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
16758 DAG.getConstant(X86::COND_B, dl, MVT::i8),
16760 Results.push_back(SetCC);
16761 Results.push_back(Store);
16762 return DAG.getMergeValues(Results, dl);
16764 case COMPRESS_TO_MEM: {
16766 SDValue Mask = Op.getOperand(4);
16767 SDValue DataToCompress = Op.getOperand(3);
16768 SDValue Addr = Op.getOperand(2);
16769 SDValue Chain = Op.getOperand(0);
16771 EVT VT = DataToCompress.getValueType();
16772 if (isAllOnes(Mask)) // return just a store
16773 return DAG.getStore(Chain, dl, DataToCompress, Addr,
16774 MachinePointerInfo(), false, false,
16775 VT.getScalarSizeInBits()/8);
16777 SDValue Compressed =
16778 getVectorMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT, DataToCompress),
16779 Mask, DAG.getUNDEF(VT), Subtarget, DAG);
16780 return DAG.getStore(Chain, dl, Compressed, Addr,
16781 MachinePointerInfo(), false, false,
16782 VT.getScalarSizeInBits()/8);
16784 case TRUNCATE_TO_MEM_VI8:
16785 return LowerINTRINSIC_TRUNCATE_TO_MEM(Op, DAG, MVT::i8);
16786 case TRUNCATE_TO_MEM_VI16:
16787 return LowerINTRINSIC_TRUNCATE_TO_MEM(Op, DAG, MVT::i16);
16788 case TRUNCATE_TO_MEM_VI32:
16789 return LowerINTRINSIC_TRUNCATE_TO_MEM(Op, DAG, MVT::i32);
16790 case EXPAND_FROM_MEM: {
16792 SDValue Mask = Op.getOperand(4);
16793 SDValue PassThru = Op.getOperand(3);
16794 SDValue Addr = Op.getOperand(2);
16795 SDValue Chain = Op.getOperand(0);
16796 EVT VT = Op.getValueType();
16798 if (isAllOnes(Mask)) // return just a load
16799 return DAG.getLoad(VT, dl, Chain, Addr, MachinePointerInfo(), false, false,
16800 false, VT.getScalarSizeInBits()/8);
16802 SDValue DataToExpand = DAG.getLoad(VT, dl, Chain, Addr, MachinePointerInfo(),
16803 false, false, false,
16804 VT.getScalarSizeInBits()/8);
16806 SDValue Results[] = {
16807 getVectorMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT, DataToExpand),
16808 Mask, PassThru, Subtarget, DAG), Chain};
16809 return DAG.getMergeValues(Results, dl);
16814 SDValue X86TargetLowering::LowerRETURNADDR(SDValue Op,
16815 SelectionDAG &DAG) const {
16816 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
16817 MFI->setReturnAddressIsTaken(true);
16819 if (verifyReturnAddressArgumentIsConstant(Op, DAG))
16822 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
16824 EVT PtrVT = getPointerTy(DAG.getDataLayout());
16827 SDValue FrameAddr = LowerFRAMEADDR(Op, DAG);
16828 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
16829 SDValue Offset = DAG.getConstant(RegInfo->getSlotSize(), dl, PtrVT);
16830 return DAG.getLoad(PtrVT, dl, DAG.getEntryNode(),
16831 DAG.getNode(ISD::ADD, dl, PtrVT,
16832 FrameAddr, Offset),
16833 MachinePointerInfo(), false, false, false, 0);
16836 // Just load the return address.
16837 SDValue RetAddrFI = getReturnAddressFrameIndex(DAG);
16838 return DAG.getLoad(PtrVT, dl, DAG.getEntryNode(),
16839 RetAddrFI, MachinePointerInfo(), false, false, false, 0);
16842 SDValue X86TargetLowering::LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const {
16843 MachineFunction &MF = DAG.getMachineFunction();
16844 MachineFrameInfo *MFI = MF.getFrameInfo();
16845 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
16846 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
16847 EVT VT = Op.getValueType();
16849 MFI->setFrameAddressIsTaken(true);
16851 if (MF.getTarget().getMCAsmInfo()->usesWindowsCFI()) {
16852 // Depth > 0 makes no sense on targets which use Windows unwind codes. It
16853 // is not possible to crawl up the stack without looking at the unwind codes
16855 int FrameAddrIndex = FuncInfo->getFAIndex();
16856 if (!FrameAddrIndex) {
16857 // Set up a frame object for the return address.
16858 unsigned SlotSize = RegInfo->getSlotSize();
16859 FrameAddrIndex = MF.getFrameInfo()->CreateFixedObject(
16860 SlotSize, /*Offset=*/0, /*IsImmutable=*/false);
16861 FuncInfo->setFAIndex(FrameAddrIndex);
16863 return DAG.getFrameIndex(FrameAddrIndex, VT);
16866 unsigned FrameReg =
16867 RegInfo->getPtrSizedFrameRegister(DAG.getMachineFunction());
16868 SDLoc dl(Op); // FIXME probably not meaningful
16869 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
16870 assert(((FrameReg == X86::RBP && VT == MVT::i64) ||
16871 (FrameReg == X86::EBP && VT == MVT::i32)) &&
16872 "Invalid Frame Register!");
16873 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, VT);
16875 FrameAddr = DAG.getLoad(VT, dl, DAG.getEntryNode(), FrameAddr,
16876 MachinePointerInfo(),
16877 false, false, false, 0);
16881 // FIXME? Maybe this could be a TableGen attribute on some registers and
16882 // this table could be generated automatically from RegInfo.
16883 unsigned X86TargetLowering::getRegisterByName(const char* RegName, EVT VT,
16884 SelectionDAG &DAG) const {
16885 const TargetFrameLowering &TFI = *Subtarget->getFrameLowering();
16886 const MachineFunction &MF = DAG.getMachineFunction();
16888 unsigned Reg = StringSwitch<unsigned>(RegName)
16889 .Case("esp", X86::ESP)
16890 .Case("rsp", X86::RSP)
16891 .Case("ebp", X86::EBP)
16892 .Case("rbp", X86::RBP)
16895 if (Reg == X86::EBP || Reg == X86::RBP) {
16896 if (!TFI.hasFP(MF))
16897 report_fatal_error("register " + StringRef(RegName) +
16898 " is allocatable: function has no frame pointer");
16901 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
16902 unsigned FrameReg =
16903 RegInfo->getPtrSizedFrameRegister(DAG.getMachineFunction());
16904 assert((FrameReg == X86::EBP || FrameReg == X86::RBP) &&
16905 "Invalid Frame Register!");
16913 report_fatal_error("Invalid register name global variable");
16916 SDValue X86TargetLowering::LowerFRAME_TO_ARGS_OFFSET(SDValue Op,
16917 SelectionDAG &DAG) const {
16918 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
16919 return DAG.getIntPtrConstant(2 * RegInfo->getSlotSize(), SDLoc(Op));
16922 SDValue X86TargetLowering::LowerEH_RETURN(SDValue Op, SelectionDAG &DAG) const {
16923 SDValue Chain = Op.getOperand(0);
16924 SDValue Offset = Op.getOperand(1);
16925 SDValue Handler = Op.getOperand(2);
16928 EVT PtrVT = getPointerTy(DAG.getDataLayout());
16929 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
16930 unsigned FrameReg = RegInfo->getFrameRegister(DAG.getMachineFunction());
16931 assert(((FrameReg == X86::RBP && PtrVT == MVT::i64) ||
16932 (FrameReg == X86::EBP && PtrVT == MVT::i32)) &&
16933 "Invalid Frame Register!");
16934 SDValue Frame = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, PtrVT);
16935 unsigned StoreAddrReg = (PtrVT == MVT::i64) ? X86::RCX : X86::ECX;
16937 SDValue StoreAddr = DAG.getNode(ISD::ADD, dl, PtrVT, Frame,
16938 DAG.getIntPtrConstant(RegInfo->getSlotSize(),
16940 StoreAddr = DAG.getNode(ISD::ADD, dl, PtrVT, StoreAddr, Offset);
16941 Chain = DAG.getStore(Chain, dl, Handler, StoreAddr, MachinePointerInfo(),
16943 Chain = DAG.getCopyToReg(Chain, dl, StoreAddrReg, StoreAddr);
16945 return DAG.getNode(X86ISD::EH_RETURN, dl, MVT::Other, Chain,
16946 DAG.getRegister(StoreAddrReg, PtrVT));
16949 SDValue X86TargetLowering::lowerEH_SJLJ_SETJMP(SDValue Op,
16950 SelectionDAG &DAG) const {
16952 return DAG.getNode(X86ISD::EH_SJLJ_SETJMP, DL,
16953 DAG.getVTList(MVT::i32, MVT::Other),
16954 Op.getOperand(0), Op.getOperand(1));
16957 SDValue X86TargetLowering::lowerEH_SJLJ_LONGJMP(SDValue Op,
16958 SelectionDAG &DAG) const {
16960 return DAG.getNode(X86ISD::EH_SJLJ_LONGJMP, DL, MVT::Other,
16961 Op.getOperand(0), Op.getOperand(1));
16964 static SDValue LowerADJUST_TRAMPOLINE(SDValue Op, SelectionDAG &DAG) {
16965 return Op.getOperand(0);
16968 SDValue X86TargetLowering::LowerINIT_TRAMPOLINE(SDValue Op,
16969 SelectionDAG &DAG) const {
16970 SDValue Root = Op.getOperand(0);
16971 SDValue Trmp = Op.getOperand(1); // trampoline
16972 SDValue FPtr = Op.getOperand(2); // nested function
16973 SDValue Nest = Op.getOperand(3); // 'nest' parameter value
16976 const Value *TrmpAddr = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
16977 const TargetRegisterInfo *TRI = Subtarget->getRegisterInfo();
16979 if (Subtarget->is64Bit()) {
16980 SDValue OutChains[6];
16982 // Large code-model.
16983 const unsigned char JMP64r = 0xFF; // 64-bit jmp through register opcode.
16984 const unsigned char MOV64ri = 0xB8; // X86::MOV64ri opcode.
16986 const unsigned char N86R10 = TRI->getEncodingValue(X86::R10) & 0x7;
16987 const unsigned char N86R11 = TRI->getEncodingValue(X86::R11) & 0x7;
16989 const unsigned char REX_WB = 0x40 | 0x08 | 0x01; // REX prefix
16991 // Load the pointer to the nested function into R11.
16992 unsigned OpCode = ((MOV64ri | N86R11) << 8) | REX_WB; // movabsq r11
16993 SDValue Addr = Trmp;
16994 OutChains[0] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, dl, MVT::i16),
16995 Addr, MachinePointerInfo(TrmpAddr),
16998 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
16999 DAG.getConstant(2, dl, MVT::i64));
17000 OutChains[1] = DAG.getStore(Root, dl, FPtr, Addr,
17001 MachinePointerInfo(TrmpAddr, 2),
17004 // Load the 'nest' parameter value into R10.
17005 // R10 is specified in X86CallingConv.td
17006 OpCode = ((MOV64ri | N86R10) << 8) | REX_WB; // movabsq r10
17007 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
17008 DAG.getConstant(10, dl, MVT::i64));
17009 OutChains[2] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, dl, MVT::i16),
17010 Addr, MachinePointerInfo(TrmpAddr, 10),
17013 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
17014 DAG.getConstant(12, dl, MVT::i64));
17015 OutChains[3] = DAG.getStore(Root, dl, Nest, Addr,
17016 MachinePointerInfo(TrmpAddr, 12),
17019 // Jump to the nested function.
17020 OpCode = (JMP64r << 8) | REX_WB; // jmpq *...
17021 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
17022 DAG.getConstant(20, dl, MVT::i64));
17023 OutChains[4] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, dl, MVT::i16),
17024 Addr, MachinePointerInfo(TrmpAddr, 20),
17027 unsigned char ModRM = N86R11 | (4 << 3) | (3 << 6); // ...r11
17028 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
17029 DAG.getConstant(22, dl, MVT::i64));
17030 OutChains[5] = DAG.getStore(Root, dl, DAG.getConstant(ModRM, dl, MVT::i8),
17031 Addr, MachinePointerInfo(TrmpAddr, 22),
17034 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other, OutChains);
17036 const Function *Func =
17037 cast<Function>(cast<SrcValueSDNode>(Op.getOperand(5))->getValue());
17038 CallingConv::ID CC = Func->getCallingConv();
17043 llvm_unreachable("Unsupported calling convention");
17044 case CallingConv::C:
17045 case CallingConv::X86_StdCall: {
17046 // Pass 'nest' parameter in ECX.
17047 // Must be kept in sync with X86CallingConv.td
17048 NestReg = X86::ECX;
17050 // Check that ECX wasn't needed by an 'inreg' parameter.
17051 FunctionType *FTy = Func->getFunctionType();
17052 const AttributeSet &Attrs = Func->getAttributes();
17054 if (!Attrs.isEmpty() && !Func->isVarArg()) {
17055 unsigned InRegCount = 0;
17058 for (FunctionType::param_iterator I = FTy->param_begin(),
17059 E = FTy->param_end(); I != E; ++I, ++Idx)
17060 if (Attrs.hasAttribute(Idx, Attribute::InReg)) {
17061 auto &DL = DAG.getDataLayout();
17062 // FIXME: should only count parameters that are lowered to integers.
17063 InRegCount += (DL.getTypeSizeInBits(*I) + 31) / 32;
17066 if (InRegCount > 2) {
17067 report_fatal_error("Nest register in use - reduce number of inreg"
17073 case CallingConv::X86_FastCall:
17074 case CallingConv::X86_ThisCall:
17075 case CallingConv::Fast:
17076 // Pass 'nest' parameter in EAX.
17077 // Must be kept in sync with X86CallingConv.td
17078 NestReg = X86::EAX;
17082 SDValue OutChains[4];
17083 SDValue Addr, Disp;
17085 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
17086 DAG.getConstant(10, dl, MVT::i32));
17087 Disp = DAG.getNode(ISD::SUB, dl, MVT::i32, FPtr, Addr);
17089 // This is storing the opcode for MOV32ri.
17090 const unsigned char MOV32ri = 0xB8; // X86::MOV32ri's opcode byte.
17091 const unsigned char N86Reg = TRI->getEncodingValue(NestReg) & 0x7;
17092 OutChains[0] = DAG.getStore(Root, dl,
17093 DAG.getConstant(MOV32ri|N86Reg, dl, MVT::i8),
17094 Trmp, MachinePointerInfo(TrmpAddr),
17097 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
17098 DAG.getConstant(1, dl, MVT::i32));
17099 OutChains[1] = DAG.getStore(Root, dl, Nest, Addr,
17100 MachinePointerInfo(TrmpAddr, 1),
17103 const unsigned char JMP = 0xE9; // jmp <32bit dst> opcode.
17104 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
17105 DAG.getConstant(5, dl, MVT::i32));
17106 OutChains[2] = DAG.getStore(Root, dl, DAG.getConstant(JMP, dl, MVT::i8),
17107 Addr, MachinePointerInfo(TrmpAddr, 5),
17110 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
17111 DAG.getConstant(6, dl, MVT::i32));
17112 OutChains[3] = DAG.getStore(Root, dl, Disp, Addr,
17113 MachinePointerInfo(TrmpAddr, 6),
17116 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other, OutChains);
17120 SDValue X86TargetLowering::LowerFLT_ROUNDS_(SDValue Op,
17121 SelectionDAG &DAG) const {
17123 The rounding mode is in bits 11:10 of FPSR, and has the following
17125 00 Round to nearest
17130 FLT_ROUNDS, on the other hand, expects the following:
17137 To perform the conversion, we do:
17138 (((((FPSR & 0x800) >> 11) | ((FPSR & 0x400) >> 9)) + 1) & 3)
17141 MachineFunction &MF = DAG.getMachineFunction();
17142 const TargetFrameLowering &TFI = *Subtarget->getFrameLowering();
17143 unsigned StackAlignment = TFI.getStackAlignment();
17144 MVT VT = Op.getSimpleValueType();
17147 // Save FP Control Word to stack slot
17148 int SSFI = MF.getFrameInfo()->CreateStackObject(2, StackAlignment, false);
17149 SDValue StackSlot =
17150 DAG.getFrameIndex(SSFI, getPointerTy(DAG.getDataLayout()));
17152 MachineMemOperand *MMO =
17153 MF.getMachineMemOperand(MachinePointerInfo::getFixedStack(MF, SSFI),
17154 MachineMemOperand::MOStore, 2, 2);
17156 SDValue Ops[] = { DAG.getEntryNode(), StackSlot };
17157 SDValue Chain = DAG.getMemIntrinsicNode(X86ISD::FNSTCW16m, DL,
17158 DAG.getVTList(MVT::Other),
17159 Ops, MVT::i16, MMO);
17161 // Load FP Control Word from stack slot
17162 SDValue CWD = DAG.getLoad(MVT::i16, DL, Chain, StackSlot,
17163 MachinePointerInfo(), false, false, false, 0);
17165 // Transform as necessary
17167 DAG.getNode(ISD::SRL, DL, MVT::i16,
17168 DAG.getNode(ISD::AND, DL, MVT::i16,
17169 CWD, DAG.getConstant(0x800, DL, MVT::i16)),
17170 DAG.getConstant(11, DL, MVT::i8));
17172 DAG.getNode(ISD::SRL, DL, MVT::i16,
17173 DAG.getNode(ISD::AND, DL, MVT::i16,
17174 CWD, DAG.getConstant(0x400, DL, MVT::i16)),
17175 DAG.getConstant(9, DL, MVT::i8));
17178 DAG.getNode(ISD::AND, DL, MVT::i16,
17179 DAG.getNode(ISD::ADD, DL, MVT::i16,
17180 DAG.getNode(ISD::OR, DL, MVT::i16, CWD1, CWD2),
17181 DAG.getConstant(1, DL, MVT::i16)),
17182 DAG.getConstant(3, DL, MVT::i16));
17184 return DAG.getNode((VT.getSizeInBits() < 16 ?
17185 ISD::TRUNCATE : ISD::ZERO_EXTEND), DL, VT, RetVal);
17188 static SDValue LowerCTLZ(SDValue Op, SelectionDAG &DAG) {
17189 MVT VT = Op.getSimpleValueType();
17191 unsigned NumBits = VT.getSizeInBits();
17194 Op = Op.getOperand(0);
17195 if (VT == MVT::i8) {
17196 // Zero extend to i32 since there is not an i8 bsr.
17198 Op = DAG.getNode(ISD::ZERO_EXTEND, dl, OpVT, Op);
17201 // Issue a bsr (scan bits in reverse) which also sets EFLAGS.
17202 SDVTList VTs = DAG.getVTList(OpVT, MVT::i32);
17203 Op = DAG.getNode(X86ISD::BSR, dl, VTs, Op);
17205 // If src is zero (i.e. bsr sets ZF), returns NumBits.
17208 DAG.getConstant(NumBits + NumBits - 1, dl, OpVT),
17209 DAG.getConstant(X86::COND_E, dl, MVT::i8),
17212 Op = DAG.getNode(X86ISD::CMOV, dl, OpVT, Ops);
17214 // Finally xor with NumBits-1.
17215 Op = DAG.getNode(ISD::XOR, dl, OpVT, Op,
17216 DAG.getConstant(NumBits - 1, dl, OpVT));
17219 Op = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op);
17223 static SDValue LowerCTLZ_ZERO_UNDEF(SDValue Op, SelectionDAG &DAG) {
17224 MVT VT = Op.getSimpleValueType();
17226 unsigned NumBits = VT.getSizeInBits();
17229 Op = Op.getOperand(0);
17230 if (VT == MVT::i8) {
17231 // Zero extend to i32 since there is not an i8 bsr.
17233 Op = DAG.getNode(ISD::ZERO_EXTEND, dl, OpVT, Op);
17236 // Issue a bsr (scan bits in reverse).
17237 SDVTList VTs = DAG.getVTList(OpVT, MVT::i32);
17238 Op = DAG.getNode(X86ISD::BSR, dl, VTs, Op);
17240 // And xor with NumBits-1.
17241 Op = DAG.getNode(ISD::XOR, dl, OpVT, Op,
17242 DAG.getConstant(NumBits - 1, dl, OpVT));
17245 Op = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op);
17249 static SDValue LowerCTTZ(SDValue Op, SelectionDAG &DAG) {
17250 MVT VT = Op.getSimpleValueType();
17251 unsigned NumBits = VT.getScalarSizeInBits();
17254 if (VT.isVector()) {
17255 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
17257 SDValue N0 = Op.getOperand(0);
17258 SDValue Zero = DAG.getConstant(0, dl, VT);
17260 // lsb(x) = (x & -x)
17261 SDValue LSB = DAG.getNode(ISD::AND, dl, VT, N0,
17262 DAG.getNode(ISD::SUB, dl, VT, Zero, N0));
17264 // cttz_undef(x) = (width - 1) - ctlz(lsb)
17265 if (Op.getOpcode() == ISD::CTTZ_ZERO_UNDEF &&
17266 TLI.isOperationLegal(ISD::CTLZ, VT)) {
17267 SDValue WidthMinusOne = DAG.getConstant(NumBits - 1, dl, VT);
17268 return DAG.getNode(ISD::SUB, dl, VT, WidthMinusOne,
17269 DAG.getNode(ISD::CTLZ, dl, VT, LSB));
17272 // cttz(x) = ctpop(lsb - 1)
17273 SDValue One = DAG.getConstant(1, dl, VT);
17274 return DAG.getNode(ISD::CTPOP, dl, VT,
17275 DAG.getNode(ISD::SUB, dl, VT, LSB, One));
17278 assert(Op.getOpcode() == ISD::CTTZ &&
17279 "Only scalar CTTZ requires custom lowering");
17281 // Issue a bsf (scan bits forward) which also sets EFLAGS.
17282 SDVTList VTs = DAG.getVTList(VT, MVT::i32);
17283 Op = DAG.getNode(X86ISD::BSF, dl, VTs, Op.getOperand(0));
17285 // If src is zero (i.e. bsf sets ZF), returns NumBits.
17288 DAG.getConstant(NumBits, dl, VT),
17289 DAG.getConstant(X86::COND_E, dl, MVT::i8),
17292 return DAG.getNode(X86ISD::CMOV, dl, VT, Ops);
17295 // Lower256IntArith - Break a 256-bit integer operation into two new 128-bit
17296 // ones, and then concatenate the result back.
17297 static SDValue Lower256IntArith(SDValue Op, SelectionDAG &DAG) {
17298 MVT VT = Op.getSimpleValueType();
17300 assert(VT.is256BitVector() && VT.isInteger() &&
17301 "Unsupported value type for operation");
17303 unsigned NumElems = VT.getVectorNumElements();
17306 // Extract the LHS vectors
17307 SDValue LHS = Op.getOperand(0);
17308 SDValue LHS1 = Extract128BitVector(LHS, 0, DAG, dl);
17309 SDValue LHS2 = Extract128BitVector(LHS, NumElems/2, DAG, dl);
17311 // Extract the RHS vectors
17312 SDValue RHS = Op.getOperand(1);
17313 SDValue RHS1 = Extract128BitVector(RHS, 0, DAG, dl);
17314 SDValue RHS2 = Extract128BitVector(RHS, NumElems/2, DAG, dl);
17316 MVT EltVT = VT.getVectorElementType();
17317 MVT NewVT = MVT::getVectorVT(EltVT, NumElems/2);
17319 return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT,
17320 DAG.getNode(Op.getOpcode(), dl, NewVT, LHS1, RHS1),
17321 DAG.getNode(Op.getOpcode(), dl, NewVT, LHS2, RHS2));
17324 static SDValue LowerADD(SDValue Op, SelectionDAG &DAG) {
17325 if (Op.getValueType() == MVT::i1)
17326 return DAG.getNode(ISD::XOR, SDLoc(Op), Op.getValueType(),
17327 Op.getOperand(0), Op.getOperand(1));
17328 assert(Op.getSimpleValueType().is256BitVector() &&
17329 Op.getSimpleValueType().isInteger() &&
17330 "Only handle AVX 256-bit vector integer operation");
17331 return Lower256IntArith(Op, DAG);
17334 static SDValue LowerSUB(SDValue Op, SelectionDAG &DAG) {
17335 if (Op.getValueType() == MVT::i1)
17336 return DAG.getNode(ISD::XOR, SDLoc(Op), Op.getValueType(),
17337 Op.getOperand(0), Op.getOperand(1));
17338 assert(Op.getSimpleValueType().is256BitVector() &&
17339 Op.getSimpleValueType().isInteger() &&
17340 "Only handle AVX 256-bit vector integer operation");
17341 return Lower256IntArith(Op, DAG);
17344 static SDValue LowerMINMAX(SDValue Op, SelectionDAG &DAG) {
17345 assert(Op.getSimpleValueType().is256BitVector() &&
17346 Op.getSimpleValueType().isInteger() &&
17347 "Only handle AVX 256-bit vector integer operation");
17348 return Lower256IntArith(Op, DAG);
17351 static SDValue LowerMUL(SDValue Op, const X86Subtarget *Subtarget,
17352 SelectionDAG &DAG) {
17354 MVT VT = Op.getSimpleValueType();
17357 return DAG.getNode(ISD::AND, dl, VT, Op.getOperand(0), Op.getOperand(1));
17359 // Decompose 256-bit ops into smaller 128-bit ops.
17360 if (VT.is256BitVector() && !Subtarget->hasInt256())
17361 return Lower256IntArith(Op, DAG);
17363 SDValue A = Op.getOperand(0);
17364 SDValue B = Op.getOperand(1);
17366 // Lower v16i8/v32i8 mul as promotion to v8i16/v16i16 vector
17367 // pairs, multiply and truncate.
17368 if (VT == MVT::v16i8 || VT == MVT::v32i8) {
17369 if (Subtarget->hasInt256()) {
17370 if (VT == MVT::v32i8) {
17371 MVT SubVT = MVT::getVectorVT(MVT::i8, VT.getVectorNumElements() / 2);
17372 SDValue Lo = DAG.getIntPtrConstant(0, dl);
17373 SDValue Hi = DAG.getIntPtrConstant(VT.getVectorNumElements() / 2, dl);
17374 SDValue ALo = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, SubVT, A, Lo);
17375 SDValue BLo = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, SubVT, B, Lo);
17376 SDValue AHi = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, SubVT, A, Hi);
17377 SDValue BHi = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, SubVT, B, Hi);
17378 return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT,
17379 DAG.getNode(ISD::MUL, dl, SubVT, ALo, BLo),
17380 DAG.getNode(ISD::MUL, dl, SubVT, AHi, BHi));
17383 MVT ExVT = MVT::getVectorVT(MVT::i16, VT.getVectorNumElements());
17384 return DAG.getNode(
17385 ISD::TRUNCATE, dl, VT,
17386 DAG.getNode(ISD::MUL, dl, ExVT,
17387 DAG.getNode(ISD::SIGN_EXTEND, dl, ExVT, A),
17388 DAG.getNode(ISD::SIGN_EXTEND, dl, ExVT, B)));
17391 assert(VT == MVT::v16i8 &&
17392 "Pre-AVX2 support only supports v16i8 multiplication");
17393 MVT ExVT = MVT::v8i16;
17395 // Extract the lo parts and sign extend to i16
17397 if (Subtarget->hasSSE41()) {
17398 ALo = DAG.getNode(X86ISD::VSEXT, dl, ExVT, A);
17399 BLo = DAG.getNode(X86ISD::VSEXT, dl, ExVT, B);
17401 const int ShufMask[] = {-1, 0, -1, 1, -1, 2, -1, 3,
17402 -1, 4, -1, 5, -1, 6, -1, 7};
17403 ALo = DAG.getVectorShuffle(VT, dl, A, A, ShufMask);
17404 BLo = DAG.getVectorShuffle(VT, dl, B, B, ShufMask);
17405 ALo = DAG.getBitcast(ExVT, ALo);
17406 BLo = DAG.getBitcast(ExVT, BLo);
17407 ALo = DAG.getNode(ISD::SRA, dl, ExVT, ALo, DAG.getConstant(8, dl, ExVT));
17408 BLo = DAG.getNode(ISD::SRA, dl, ExVT, BLo, DAG.getConstant(8, dl, ExVT));
17411 // Extract the hi parts and sign extend to i16
17413 if (Subtarget->hasSSE41()) {
17414 const int ShufMask[] = {8, 9, 10, 11, 12, 13, 14, 15,
17415 -1, -1, -1, -1, -1, -1, -1, -1};
17416 AHi = DAG.getVectorShuffle(VT, dl, A, A, ShufMask);
17417 BHi = DAG.getVectorShuffle(VT, dl, B, B, ShufMask);
17418 AHi = DAG.getNode(X86ISD::VSEXT, dl, ExVT, AHi);
17419 BHi = DAG.getNode(X86ISD::VSEXT, dl, ExVT, BHi);
17421 const int ShufMask[] = {-1, 8, -1, 9, -1, 10, -1, 11,
17422 -1, 12, -1, 13, -1, 14, -1, 15};
17423 AHi = DAG.getVectorShuffle(VT, dl, A, A, ShufMask);
17424 BHi = DAG.getVectorShuffle(VT, dl, B, B, ShufMask);
17425 AHi = DAG.getBitcast(ExVT, AHi);
17426 BHi = DAG.getBitcast(ExVT, BHi);
17427 AHi = DAG.getNode(ISD::SRA, dl, ExVT, AHi, DAG.getConstant(8, dl, ExVT));
17428 BHi = DAG.getNode(ISD::SRA, dl, ExVT, BHi, DAG.getConstant(8, dl, ExVT));
17431 // Multiply, mask the lower 8bits of the lo/hi results and pack
17432 SDValue RLo = DAG.getNode(ISD::MUL, dl, ExVT, ALo, BLo);
17433 SDValue RHi = DAG.getNode(ISD::MUL, dl, ExVT, AHi, BHi);
17434 RLo = DAG.getNode(ISD::AND, dl, ExVT, RLo, DAG.getConstant(255, dl, ExVT));
17435 RHi = DAG.getNode(ISD::AND, dl, ExVT, RHi, DAG.getConstant(255, dl, ExVT));
17436 return DAG.getNode(X86ISD::PACKUS, dl, VT, RLo, RHi);
17439 // Lower v4i32 mul as 2x shuffle, 2x pmuludq, 2x shuffle.
17440 if (VT == MVT::v4i32) {
17441 assert(Subtarget->hasSSE2() && !Subtarget->hasSSE41() &&
17442 "Should not custom lower when pmuldq is available!");
17444 // Extract the odd parts.
17445 static const int UnpackMask[] = { 1, -1, 3, -1 };
17446 SDValue Aodds = DAG.getVectorShuffle(VT, dl, A, A, UnpackMask);
17447 SDValue Bodds = DAG.getVectorShuffle(VT, dl, B, B, UnpackMask);
17449 // Multiply the even parts.
17450 SDValue Evens = DAG.getNode(X86ISD::PMULUDQ, dl, MVT::v2i64, A, B);
17451 // Now multiply odd parts.
17452 SDValue Odds = DAG.getNode(X86ISD::PMULUDQ, dl, MVT::v2i64, Aodds, Bodds);
17454 Evens = DAG.getBitcast(VT, Evens);
17455 Odds = DAG.getBitcast(VT, Odds);
17457 // Merge the two vectors back together with a shuffle. This expands into 2
17459 static const int ShufMask[] = { 0, 4, 2, 6 };
17460 return DAG.getVectorShuffle(VT, dl, Evens, Odds, ShufMask);
17463 assert((VT == MVT::v2i64 || VT == MVT::v4i64 || VT == MVT::v8i64) &&
17464 "Only know how to lower V2I64/V4I64/V8I64 multiply");
17466 // Ahi = psrlqi(a, 32);
17467 // Bhi = psrlqi(b, 32);
17469 // AloBlo = pmuludq(a, b);
17470 // AloBhi = pmuludq(a, Bhi);
17471 // AhiBlo = pmuludq(Ahi, b);
17473 // AloBhi = psllqi(AloBhi, 32);
17474 // AhiBlo = psllqi(AhiBlo, 32);
17475 // return AloBlo + AloBhi + AhiBlo;
17477 SDValue Ahi = getTargetVShiftByConstNode(X86ISD::VSRLI, dl, VT, A, 32, DAG);
17478 SDValue Bhi = getTargetVShiftByConstNode(X86ISD::VSRLI, dl, VT, B, 32, DAG);
17480 SDValue AhiBlo = Ahi;
17481 SDValue AloBhi = Bhi;
17482 // Bit cast to 32-bit vectors for MULUDQ
17483 EVT MulVT = (VT == MVT::v2i64) ? MVT::v4i32 :
17484 (VT == MVT::v4i64) ? MVT::v8i32 : MVT::v16i32;
17485 A = DAG.getBitcast(MulVT, A);
17486 B = DAG.getBitcast(MulVT, B);
17487 Ahi = DAG.getBitcast(MulVT, Ahi);
17488 Bhi = DAG.getBitcast(MulVT, Bhi);
17490 SDValue AloBlo = DAG.getNode(X86ISD::PMULUDQ, dl, VT, A, B);
17491 // After shifting right const values the result may be all-zero.
17492 if (!ISD::isBuildVectorAllZeros(Ahi.getNode())) {
17493 AhiBlo = DAG.getNode(X86ISD::PMULUDQ, dl, VT, Ahi, B);
17494 AhiBlo = getTargetVShiftByConstNode(X86ISD::VSHLI, dl, VT, AhiBlo, 32, DAG);
17496 if (!ISD::isBuildVectorAllZeros(Bhi.getNode())) {
17497 AloBhi = DAG.getNode(X86ISD::PMULUDQ, dl, VT, A, Bhi);
17498 AloBhi = getTargetVShiftByConstNode(X86ISD::VSHLI, dl, VT, AloBhi, 32, DAG);
17501 SDValue Res = DAG.getNode(ISD::ADD, dl, VT, AloBlo, AloBhi);
17502 return DAG.getNode(ISD::ADD, dl, VT, Res, AhiBlo);
17505 SDValue X86TargetLowering::LowerWin64_i128OP(SDValue Op, SelectionDAG &DAG) const {
17506 assert(Subtarget->isTargetWin64() && "Unexpected target");
17507 EVT VT = Op.getValueType();
17508 assert(VT.isInteger() && VT.getSizeInBits() == 128 &&
17509 "Unexpected return type for lowering");
17513 switch (Op->getOpcode()) {
17514 default: llvm_unreachable("Unexpected request for libcall!");
17515 case ISD::SDIV: isSigned = true; LC = RTLIB::SDIV_I128; break;
17516 case ISD::UDIV: isSigned = false; LC = RTLIB::UDIV_I128; break;
17517 case ISD::SREM: isSigned = true; LC = RTLIB::SREM_I128; break;
17518 case ISD::UREM: isSigned = false; LC = RTLIB::UREM_I128; break;
17519 case ISD::SDIVREM: isSigned = true; LC = RTLIB::SDIVREM_I128; break;
17520 case ISD::UDIVREM: isSigned = false; LC = RTLIB::UDIVREM_I128; break;
17524 SDValue InChain = DAG.getEntryNode();
17526 TargetLowering::ArgListTy Args;
17527 TargetLowering::ArgListEntry Entry;
17528 for (unsigned i = 0, e = Op->getNumOperands(); i != e; ++i) {
17529 EVT ArgVT = Op->getOperand(i).getValueType();
17530 assert(ArgVT.isInteger() && ArgVT.getSizeInBits() == 128 &&
17531 "Unexpected argument type for lowering");
17532 SDValue StackPtr = DAG.CreateStackTemporary(ArgVT, 16);
17533 Entry.Node = StackPtr;
17534 InChain = DAG.getStore(InChain, dl, Op->getOperand(i), StackPtr, MachinePointerInfo(),
17536 Type *ArgTy = ArgVT.getTypeForEVT(*DAG.getContext());
17537 Entry.Ty = PointerType::get(ArgTy,0);
17538 Entry.isSExt = false;
17539 Entry.isZExt = false;
17540 Args.push_back(Entry);
17543 SDValue Callee = DAG.getExternalSymbol(getLibcallName(LC),
17544 getPointerTy(DAG.getDataLayout()));
17546 TargetLowering::CallLoweringInfo CLI(DAG);
17547 CLI.setDebugLoc(dl).setChain(InChain)
17548 .setCallee(getLibcallCallingConv(LC),
17549 static_cast<EVT>(MVT::v2i64).getTypeForEVT(*DAG.getContext()),
17550 Callee, std::move(Args), 0)
17551 .setInRegister().setSExtResult(isSigned).setZExtResult(!isSigned);
17553 std::pair<SDValue, SDValue> CallInfo = LowerCallTo(CLI);
17554 return DAG.getBitcast(VT, CallInfo.first);
17557 static SDValue LowerMUL_LOHI(SDValue Op, const X86Subtarget *Subtarget,
17558 SelectionDAG &DAG) {
17559 SDValue Op0 = Op.getOperand(0), Op1 = Op.getOperand(1);
17560 EVT VT = Op0.getValueType();
17563 assert((VT == MVT::v4i32 && Subtarget->hasSSE2()) ||
17564 (VT == MVT::v8i32 && Subtarget->hasInt256()));
17566 // PMULxD operations multiply each even value (starting at 0) of LHS with
17567 // the related value of RHS and produce a widen result.
17568 // E.g., PMULUDQ <4 x i32> <a|b|c|d>, <4 x i32> <e|f|g|h>
17569 // => <2 x i64> <ae|cg>
17571 // In other word, to have all the results, we need to perform two PMULxD:
17572 // 1. one with the even values.
17573 // 2. one with the odd values.
17574 // To achieve #2, with need to place the odd values at an even position.
17576 // Place the odd value at an even position (basically, shift all values 1
17577 // step to the left):
17578 const int Mask[] = {1, -1, 3, -1, 5, -1, 7, -1};
17579 // <a|b|c|d> => <b|undef|d|undef>
17580 SDValue Odd0 = DAG.getVectorShuffle(VT, dl, Op0, Op0, Mask);
17581 // <e|f|g|h> => <f|undef|h|undef>
17582 SDValue Odd1 = DAG.getVectorShuffle(VT, dl, Op1, Op1, Mask);
17584 // Emit two multiplies, one for the lower 2 ints and one for the higher 2
17586 MVT MulVT = VT == MVT::v4i32 ? MVT::v2i64 : MVT::v4i64;
17587 bool IsSigned = Op->getOpcode() == ISD::SMUL_LOHI;
17589 (!IsSigned || !Subtarget->hasSSE41()) ? X86ISD::PMULUDQ : X86ISD::PMULDQ;
17590 // PMULUDQ <4 x i32> <a|b|c|d>, <4 x i32> <e|f|g|h>
17591 // => <2 x i64> <ae|cg>
17592 SDValue Mul1 = DAG.getBitcast(VT, DAG.getNode(Opcode, dl, MulVT, Op0, Op1));
17593 // PMULUDQ <4 x i32> <b|undef|d|undef>, <4 x i32> <f|undef|h|undef>
17594 // => <2 x i64> <bf|dh>
17595 SDValue Mul2 = DAG.getBitcast(VT, DAG.getNode(Opcode, dl, MulVT, Odd0, Odd1));
17597 // Shuffle it back into the right order.
17598 SDValue Highs, Lows;
17599 if (VT == MVT::v8i32) {
17600 const int HighMask[] = {1, 9, 3, 11, 5, 13, 7, 15};
17601 Highs = DAG.getVectorShuffle(VT, dl, Mul1, Mul2, HighMask);
17602 const int LowMask[] = {0, 8, 2, 10, 4, 12, 6, 14};
17603 Lows = DAG.getVectorShuffle(VT, dl, Mul1, Mul2, LowMask);
17605 const int HighMask[] = {1, 5, 3, 7};
17606 Highs = DAG.getVectorShuffle(VT, dl, Mul1, Mul2, HighMask);
17607 const int LowMask[] = {0, 4, 2, 6};
17608 Lows = DAG.getVectorShuffle(VT, dl, Mul1, Mul2, LowMask);
17611 // If we have a signed multiply but no PMULDQ fix up the high parts of a
17612 // unsigned multiply.
17613 if (IsSigned && !Subtarget->hasSSE41()) {
17614 SDValue ShAmt = DAG.getConstant(
17616 DAG.getTargetLoweringInfo().getShiftAmountTy(VT, DAG.getDataLayout()));
17617 SDValue T1 = DAG.getNode(ISD::AND, dl, VT,
17618 DAG.getNode(ISD::SRA, dl, VT, Op0, ShAmt), Op1);
17619 SDValue T2 = DAG.getNode(ISD::AND, dl, VT,
17620 DAG.getNode(ISD::SRA, dl, VT, Op1, ShAmt), Op0);
17622 SDValue Fixup = DAG.getNode(ISD::ADD, dl, VT, T1, T2);
17623 Highs = DAG.getNode(ISD::SUB, dl, VT, Highs, Fixup);
17626 // The first result of MUL_LOHI is actually the low value, followed by the
17628 SDValue Ops[] = {Lows, Highs};
17629 return DAG.getMergeValues(Ops, dl);
17632 // Return true if the required (according to Opcode) shift-imm form is natively
17633 // supported by the Subtarget
17634 static bool SupportedVectorShiftWithImm(MVT VT, const X86Subtarget *Subtarget,
17636 if (VT.getScalarSizeInBits() < 16)
17639 if (VT.is512BitVector() &&
17640 (VT.getScalarSizeInBits() > 16 || Subtarget->hasBWI()))
17643 bool LShift = VT.is128BitVector() ||
17644 (VT.is256BitVector() && Subtarget->hasInt256());
17646 bool AShift = LShift && (Subtarget->hasVLX() ||
17647 (VT != MVT::v2i64 && VT != MVT::v4i64));
17648 return (Opcode == ISD::SRA) ? AShift : LShift;
17651 // The shift amount is a variable, but it is the same for all vector lanes.
17652 // These instructions are defined together with shift-immediate.
17654 bool SupportedVectorShiftWithBaseAmnt(MVT VT, const X86Subtarget *Subtarget,
17656 return SupportedVectorShiftWithImm(VT, Subtarget, Opcode);
17659 // Return true if the required (according to Opcode) variable-shift form is
17660 // natively supported by the Subtarget
17661 static bool SupportedVectorVarShift(MVT VT, const X86Subtarget *Subtarget,
17664 if (!Subtarget->hasInt256() || VT.getScalarSizeInBits() < 16)
17667 // vXi16 supported only on AVX-512, BWI
17668 if (VT.getScalarSizeInBits() == 16 && !Subtarget->hasBWI())
17671 if (VT.is512BitVector() || Subtarget->hasVLX())
17674 bool LShift = VT.is128BitVector() || VT.is256BitVector();
17675 bool AShift = LShift && VT != MVT::v2i64 && VT != MVT::v4i64;
17676 return (Opcode == ISD::SRA) ? AShift : LShift;
17679 static SDValue LowerScalarImmediateShift(SDValue Op, SelectionDAG &DAG,
17680 const X86Subtarget *Subtarget) {
17681 MVT VT = Op.getSimpleValueType();
17683 SDValue R = Op.getOperand(0);
17684 SDValue Amt = Op.getOperand(1);
17686 unsigned X86Opc = (Op.getOpcode() == ISD::SHL) ? X86ISD::VSHLI :
17687 (Op.getOpcode() == ISD::SRL) ? X86ISD::VSRLI : X86ISD::VSRAI;
17689 auto ArithmeticShiftRight64 = [&](uint64_t ShiftAmt) {
17690 assert((VT == MVT::v2i64 || VT == MVT::v4i64) && "Unexpected SRA type");
17691 MVT ExVT = MVT::getVectorVT(MVT::i32, VT.getVectorNumElements() * 2);
17692 SDValue Ex = DAG.getBitcast(ExVT, R);
17694 if (ShiftAmt >= 32) {
17695 // Splat sign to upper i32 dst, and SRA upper i32 src to lower i32.
17697 getTargetVShiftByConstNode(X86ISD::VSRAI, dl, ExVT, Ex, 31, DAG);
17698 SDValue Lower = getTargetVShiftByConstNode(X86ISD::VSRAI, dl, ExVT, Ex,
17699 ShiftAmt - 32, DAG);
17700 if (VT == MVT::v2i64)
17701 Ex = DAG.getVectorShuffle(ExVT, dl, Upper, Lower, {5, 1, 7, 3});
17702 if (VT == MVT::v4i64)
17703 Ex = DAG.getVectorShuffle(ExVT, dl, Upper, Lower,
17704 {9, 1, 11, 3, 13, 5, 15, 7});
17706 // SRA upper i32, SHL whole i64 and select lower i32.
17707 SDValue Upper = getTargetVShiftByConstNode(X86ISD::VSRAI, dl, ExVT, Ex,
17710 getTargetVShiftByConstNode(X86ISD::VSRLI, dl, VT, R, ShiftAmt, DAG);
17711 Lower = DAG.getBitcast(ExVT, Lower);
17712 if (VT == MVT::v2i64)
17713 Ex = DAG.getVectorShuffle(ExVT, dl, Upper, Lower, {4, 1, 6, 3});
17714 if (VT == MVT::v4i64)
17715 Ex = DAG.getVectorShuffle(ExVT, dl, Upper, Lower,
17716 {8, 1, 10, 3, 12, 5, 14, 7});
17718 return DAG.getBitcast(VT, Ex);
17721 // Optimize shl/srl/sra with constant shift amount.
17722 if (auto *BVAmt = dyn_cast<BuildVectorSDNode>(Amt)) {
17723 if (auto *ShiftConst = BVAmt->getConstantSplatNode()) {
17724 uint64_t ShiftAmt = ShiftConst->getZExtValue();
17726 if (SupportedVectorShiftWithImm(VT, Subtarget, Op.getOpcode()))
17727 return getTargetVShiftByConstNode(X86Opc, dl, VT, R, ShiftAmt, DAG);
17729 // i64 SRA needs to be performed as partial shifts.
17730 if ((VT == MVT::v2i64 || (Subtarget->hasInt256() && VT == MVT::v4i64)) &&
17731 Op.getOpcode() == ISD::SRA)
17732 return ArithmeticShiftRight64(ShiftAmt);
17734 if (VT == MVT::v16i8 || (Subtarget->hasInt256() && VT == MVT::v32i8)) {
17735 unsigned NumElts = VT.getVectorNumElements();
17736 MVT ShiftVT = MVT::getVectorVT(MVT::i16, NumElts / 2);
17738 if (Op.getOpcode() == ISD::SHL) {
17739 // Simple i8 add case
17741 return DAG.getNode(ISD::ADD, dl, VT, R, R);
17743 // Make a large shift.
17744 SDValue SHL = getTargetVShiftByConstNode(X86ISD::VSHLI, dl, ShiftVT,
17746 SHL = DAG.getBitcast(VT, SHL);
17747 // Zero out the rightmost bits.
17748 SmallVector<SDValue, 32> V(
17749 NumElts, DAG.getConstant(uint8_t(-1U << ShiftAmt), dl, MVT::i8));
17750 return DAG.getNode(ISD::AND, dl, VT, SHL,
17751 DAG.getNode(ISD::BUILD_VECTOR, dl, VT, V));
17753 if (Op.getOpcode() == ISD::SRL) {
17754 // Make a large shift.
17755 SDValue SRL = getTargetVShiftByConstNode(X86ISD::VSRLI, dl, ShiftVT,
17757 SRL = DAG.getBitcast(VT, SRL);
17758 // Zero out the leftmost bits.
17759 SmallVector<SDValue, 32> V(
17760 NumElts, DAG.getConstant(uint8_t(-1U) >> ShiftAmt, dl, MVT::i8));
17761 return DAG.getNode(ISD::AND, dl, VT, SRL,
17762 DAG.getNode(ISD::BUILD_VECTOR, dl, VT, V));
17764 if (Op.getOpcode() == ISD::SRA) {
17765 if (ShiftAmt == 7) {
17766 // ashr(R, 7) === cmp_slt(R, 0)
17767 SDValue Zeros = getZeroVector(VT, Subtarget, DAG, dl);
17768 return DAG.getNode(X86ISD::PCMPGT, dl, VT, Zeros, R);
17771 // ashr(R, Amt) === sub(xor(lshr(R, Amt), Mask), Mask)
17772 SDValue Res = DAG.getNode(ISD::SRL, dl, VT, R, Amt);
17773 SmallVector<SDValue, 32> V(NumElts,
17774 DAG.getConstant(128 >> ShiftAmt, dl,
17776 SDValue Mask = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, V);
17777 Res = DAG.getNode(ISD::XOR, dl, VT, Res, Mask);
17778 Res = DAG.getNode(ISD::SUB, dl, VT, Res, Mask);
17781 llvm_unreachable("Unknown shift opcode.");
17786 // Special case in 32-bit mode, where i64 is expanded into high and low parts.
17787 if (!Subtarget->is64Bit() &&
17788 (VT == MVT::v2i64 || (Subtarget->hasInt256() && VT == MVT::v4i64))) {
17790 // Peek through any splat that was introduced for i64 shift vectorization.
17791 int SplatIndex = -1;
17792 if (ShuffleVectorSDNode *SVN = dyn_cast<ShuffleVectorSDNode>(Amt.getNode()))
17793 if (SVN->isSplat()) {
17794 SplatIndex = SVN->getSplatIndex();
17795 Amt = Amt.getOperand(0);
17796 assert(SplatIndex < (int)VT.getVectorNumElements() &&
17797 "Splat shuffle referencing second operand");
17800 if (Amt.getOpcode() != ISD::BITCAST ||
17801 Amt.getOperand(0).getOpcode() != ISD::BUILD_VECTOR)
17804 Amt = Amt.getOperand(0);
17805 unsigned Ratio = Amt.getSimpleValueType().getVectorNumElements() /
17806 VT.getVectorNumElements();
17807 unsigned RatioInLog2 = Log2_32_Ceil(Ratio);
17808 uint64_t ShiftAmt = 0;
17809 unsigned BaseOp = (SplatIndex < 0 ? 0 : SplatIndex * Ratio);
17810 for (unsigned i = 0; i != Ratio; ++i) {
17811 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Amt.getOperand(i + BaseOp));
17815 ShiftAmt |= C->getZExtValue() << (i * (1 << (6 - RatioInLog2)));
17818 // Check remaining shift amounts (if not a splat).
17819 if (SplatIndex < 0) {
17820 for (unsigned i = Ratio; i != Amt.getNumOperands(); i += Ratio) {
17821 uint64_t ShAmt = 0;
17822 for (unsigned j = 0; j != Ratio; ++j) {
17823 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Amt.getOperand(i + j));
17827 ShAmt |= C->getZExtValue() << (j * (1 << (6 - RatioInLog2)));
17829 if (ShAmt != ShiftAmt)
17834 if (SupportedVectorShiftWithImm(VT, Subtarget, Op.getOpcode()))
17835 return getTargetVShiftByConstNode(X86Opc, dl, VT, R, ShiftAmt, DAG);
17837 if (Op.getOpcode() == ISD::SRA)
17838 return ArithmeticShiftRight64(ShiftAmt);
17844 static SDValue LowerScalarVariableShift(SDValue Op, SelectionDAG &DAG,
17845 const X86Subtarget* Subtarget) {
17846 MVT VT = Op.getSimpleValueType();
17848 SDValue R = Op.getOperand(0);
17849 SDValue Amt = Op.getOperand(1);
17851 unsigned X86OpcI = (Op.getOpcode() == ISD::SHL) ? X86ISD::VSHLI :
17852 (Op.getOpcode() == ISD::SRL) ? X86ISD::VSRLI : X86ISD::VSRAI;
17854 unsigned X86OpcV = (Op.getOpcode() == ISD::SHL) ? X86ISD::VSHL :
17855 (Op.getOpcode() == ISD::SRL) ? X86ISD::VSRL : X86ISD::VSRA;
17857 if (SupportedVectorShiftWithBaseAmnt(VT, Subtarget, Op.getOpcode())) {
17859 EVT EltVT = VT.getVectorElementType();
17861 if (BuildVectorSDNode *BV = dyn_cast<BuildVectorSDNode>(Amt)) {
17862 // Check if this build_vector node is doing a splat.
17863 // If so, then set BaseShAmt equal to the splat value.
17864 BaseShAmt = BV->getSplatValue();
17865 if (BaseShAmt && BaseShAmt.getOpcode() == ISD::UNDEF)
17866 BaseShAmt = SDValue();
17868 if (Amt.getOpcode() == ISD::EXTRACT_SUBVECTOR)
17869 Amt = Amt.getOperand(0);
17871 ShuffleVectorSDNode *SVN = dyn_cast<ShuffleVectorSDNode>(Amt);
17872 if (SVN && SVN->isSplat()) {
17873 unsigned SplatIdx = (unsigned)SVN->getSplatIndex();
17874 SDValue InVec = Amt.getOperand(0);
17875 if (InVec.getOpcode() == ISD::BUILD_VECTOR) {
17876 assert((SplatIdx < InVec.getValueType().getVectorNumElements()) &&
17877 "Unexpected shuffle index found!");
17878 BaseShAmt = InVec.getOperand(SplatIdx);
17879 } else if (InVec.getOpcode() == ISD::INSERT_VECTOR_ELT) {
17880 if (ConstantSDNode *C =
17881 dyn_cast<ConstantSDNode>(InVec.getOperand(2))) {
17882 if (C->getZExtValue() == SplatIdx)
17883 BaseShAmt = InVec.getOperand(1);
17888 // Avoid introducing an extract element from a shuffle.
17889 BaseShAmt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, EltVT, InVec,
17890 DAG.getIntPtrConstant(SplatIdx, dl));
17894 if (BaseShAmt.getNode()) {
17895 assert(EltVT.bitsLE(MVT::i64) && "Unexpected element type!");
17896 if (EltVT != MVT::i64 && EltVT.bitsGT(MVT::i32))
17897 BaseShAmt = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i64, BaseShAmt);
17898 else if (EltVT.bitsLT(MVT::i32))
17899 BaseShAmt = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, BaseShAmt);
17901 return getTargetVShiftNode(X86OpcI, dl, VT, R, BaseShAmt, DAG);
17905 // Special case in 32-bit mode, where i64 is expanded into high and low parts.
17906 if (!Subtarget->is64Bit() && VT == MVT::v2i64 &&
17907 Amt.getOpcode() == ISD::BITCAST &&
17908 Amt.getOperand(0).getOpcode() == ISD::BUILD_VECTOR) {
17909 Amt = Amt.getOperand(0);
17910 unsigned Ratio = Amt.getSimpleValueType().getVectorNumElements() /
17911 VT.getVectorNumElements();
17912 std::vector<SDValue> Vals(Ratio);
17913 for (unsigned i = 0; i != Ratio; ++i)
17914 Vals[i] = Amt.getOperand(i);
17915 for (unsigned i = Ratio; i != Amt.getNumOperands(); i += Ratio) {
17916 for (unsigned j = 0; j != Ratio; ++j)
17917 if (Vals[j] != Amt.getOperand(i + j))
17921 if (SupportedVectorShiftWithBaseAmnt(VT, Subtarget, Op.getOpcode()))
17922 return DAG.getNode(X86OpcV, dl, VT, R, Op.getOperand(1));
17927 static SDValue LowerShift(SDValue Op, const X86Subtarget* Subtarget,
17928 SelectionDAG &DAG) {
17929 MVT VT = Op.getSimpleValueType();
17931 SDValue R = Op.getOperand(0);
17932 SDValue Amt = Op.getOperand(1);
17934 assert(VT.isVector() && "Custom lowering only for vector shifts!");
17935 assert(Subtarget->hasSSE2() && "Only custom lower when we have SSE2!");
17937 if (SDValue V = LowerScalarImmediateShift(Op, DAG, Subtarget))
17940 if (SDValue V = LowerScalarVariableShift(Op, DAG, Subtarget))
17943 if (SupportedVectorVarShift(VT, Subtarget, Op.getOpcode()))
17946 // 2i64 vector logical shifts can efficiently avoid scalarization - do the
17947 // shifts per-lane and then shuffle the partial results back together.
17948 if (VT == MVT::v2i64 && Op.getOpcode() != ISD::SRA) {
17949 // Splat the shift amounts so the scalar shifts above will catch it.
17950 SDValue Amt0 = DAG.getVectorShuffle(VT, dl, Amt, Amt, {0, 0});
17951 SDValue Amt1 = DAG.getVectorShuffle(VT, dl, Amt, Amt, {1, 1});
17952 SDValue R0 = DAG.getNode(Op->getOpcode(), dl, VT, R, Amt0);
17953 SDValue R1 = DAG.getNode(Op->getOpcode(), dl, VT, R, Amt1);
17954 return DAG.getVectorShuffle(VT, dl, R0, R1, {0, 3});
17957 // i64 vector arithmetic shift can be emulated with the transform:
17958 // M = lshr(SIGN_BIT, Amt)
17959 // ashr(R, Amt) === sub(xor(lshr(R, Amt), M), M)
17960 if ((VT == MVT::v2i64 || (VT == MVT::v4i64 && Subtarget->hasInt256())) &&
17961 Op.getOpcode() == ISD::SRA) {
17962 SDValue S = DAG.getConstant(APInt::getSignBit(64), dl, VT);
17963 SDValue M = DAG.getNode(ISD::SRL, dl, VT, S, Amt);
17964 R = DAG.getNode(ISD::SRL, dl, VT, R, Amt);
17965 R = DAG.getNode(ISD::XOR, dl, VT, R, M);
17966 R = DAG.getNode(ISD::SUB, dl, VT, R, M);
17970 // If possible, lower this packed shift into a vector multiply instead of
17971 // expanding it into a sequence of scalar shifts.
17972 // Do this only if the vector shift count is a constant build_vector.
17973 if (Op.getOpcode() == ISD::SHL &&
17974 (VT == MVT::v8i16 || VT == MVT::v4i32 ||
17975 (Subtarget->hasInt256() && VT == MVT::v16i16)) &&
17976 ISD::isBuildVectorOfConstantSDNodes(Amt.getNode())) {
17977 SmallVector<SDValue, 8> Elts;
17978 EVT SVT = VT.getScalarType();
17979 unsigned SVTBits = SVT.getSizeInBits();
17980 const APInt &One = APInt(SVTBits, 1);
17981 unsigned NumElems = VT.getVectorNumElements();
17983 for (unsigned i=0; i !=NumElems; ++i) {
17984 SDValue Op = Amt->getOperand(i);
17985 if (Op->getOpcode() == ISD::UNDEF) {
17986 Elts.push_back(Op);
17990 ConstantSDNode *ND = cast<ConstantSDNode>(Op);
17991 const APInt &C = APInt(SVTBits, ND->getAPIntValue().getZExtValue());
17992 uint64_t ShAmt = C.getZExtValue();
17993 if (ShAmt >= SVTBits) {
17994 Elts.push_back(DAG.getUNDEF(SVT));
17997 Elts.push_back(DAG.getConstant(One.shl(ShAmt), dl, SVT));
17999 SDValue BV = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Elts);
18000 return DAG.getNode(ISD::MUL, dl, VT, R, BV);
18003 // Lower SHL with variable shift amount.
18004 if (VT == MVT::v4i32 && Op->getOpcode() == ISD::SHL) {
18005 Op = DAG.getNode(ISD::SHL, dl, VT, Amt, DAG.getConstant(23, dl, VT));
18007 Op = DAG.getNode(ISD::ADD, dl, VT, Op,
18008 DAG.getConstant(0x3f800000U, dl, VT));
18009 Op = DAG.getBitcast(MVT::v4f32, Op);
18010 Op = DAG.getNode(ISD::FP_TO_SINT, dl, VT, Op);
18011 return DAG.getNode(ISD::MUL, dl, VT, Op, R);
18014 // If possible, lower this shift as a sequence of two shifts by
18015 // constant plus a MOVSS/MOVSD instead of scalarizing it.
18017 // (v4i32 (srl A, (build_vector < X, Y, Y, Y>)))
18019 // Could be rewritten as:
18020 // (v4i32 (MOVSS (srl A, <Y,Y,Y,Y>), (srl A, <X,X,X,X>)))
18022 // The advantage is that the two shifts from the example would be
18023 // lowered as X86ISD::VSRLI nodes. This would be cheaper than scalarizing
18024 // the vector shift into four scalar shifts plus four pairs of vector
18026 if ((VT == MVT::v8i16 || VT == MVT::v4i32) &&
18027 ISD::isBuildVectorOfConstantSDNodes(Amt.getNode())) {
18028 unsigned TargetOpcode = X86ISD::MOVSS;
18029 bool CanBeSimplified;
18030 // The splat value for the first packed shift (the 'X' from the example).
18031 SDValue Amt1 = Amt->getOperand(0);
18032 // The splat value for the second packed shift (the 'Y' from the example).
18033 SDValue Amt2 = (VT == MVT::v4i32) ? Amt->getOperand(1) :
18034 Amt->getOperand(2);
18036 // See if it is possible to replace this node with a sequence of
18037 // two shifts followed by a MOVSS/MOVSD
18038 if (VT == MVT::v4i32) {
18039 // Check if it is legal to use a MOVSS.
18040 CanBeSimplified = Amt2 == Amt->getOperand(2) &&
18041 Amt2 == Amt->getOperand(3);
18042 if (!CanBeSimplified) {
18043 // Otherwise, check if we can still simplify this node using a MOVSD.
18044 CanBeSimplified = Amt1 == Amt->getOperand(1) &&
18045 Amt->getOperand(2) == Amt->getOperand(3);
18046 TargetOpcode = X86ISD::MOVSD;
18047 Amt2 = Amt->getOperand(2);
18050 // Do similar checks for the case where the machine value type
18052 CanBeSimplified = Amt1 == Amt->getOperand(1);
18053 for (unsigned i=3; i != 8 && CanBeSimplified; ++i)
18054 CanBeSimplified = Amt2 == Amt->getOperand(i);
18056 if (!CanBeSimplified) {
18057 TargetOpcode = X86ISD::MOVSD;
18058 CanBeSimplified = true;
18059 Amt2 = Amt->getOperand(4);
18060 for (unsigned i=0; i != 4 && CanBeSimplified; ++i)
18061 CanBeSimplified = Amt1 == Amt->getOperand(i);
18062 for (unsigned j=4; j != 8 && CanBeSimplified; ++j)
18063 CanBeSimplified = Amt2 == Amt->getOperand(j);
18067 if (CanBeSimplified && isa<ConstantSDNode>(Amt1) &&
18068 isa<ConstantSDNode>(Amt2)) {
18069 // Replace this node with two shifts followed by a MOVSS/MOVSD.
18070 EVT CastVT = MVT::v4i32;
18072 DAG.getConstant(cast<ConstantSDNode>(Amt1)->getAPIntValue(), dl, VT);
18073 SDValue Shift1 = DAG.getNode(Op->getOpcode(), dl, VT, R, Splat1);
18075 DAG.getConstant(cast<ConstantSDNode>(Amt2)->getAPIntValue(), dl, VT);
18076 SDValue Shift2 = DAG.getNode(Op->getOpcode(), dl, VT, R, Splat2);
18077 if (TargetOpcode == X86ISD::MOVSD)
18078 CastVT = MVT::v2i64;
18079 SDValue BitCast1 = DAG.getBitcast(CastVT, Shift1);
18080 SDValue BitCast2 = DAG.getBitcast(CastVT, Shift2);
18081 SDValue Result = getTargetShuffleNode(TargetOpcode, dl, CastVT, BitCast2,
18083 return DAG.getBitcast(VT, Result);
18087 // v4i32 Non Uniform Shifts.
18088 // If the shift amount is constant we can shift each lane using the SSE2
18089 // immediate shifts, else we need to zero-extend each lane to the lower i64
18090 // and shift using the SSE2 variable shifts.
18091 // The separate results can then be blended together.
18092 if (VT == MVT::v4i32) {
18093 unsigned Opc = Op.getOpcode();
18094 SDValue Amt0, Amt1, Amt2, Amt3;
18095 if (ISD::isBuildVectorOfConstantSDNodes(Amt.getNode())) {
18096 Amt0 = DAG.getVectorShuffle(VT, dl, Amt, DAG.getUNDEF(VT), {0, 0, 0, 0});
18097 Amt1 = DAG.getVectorShuffle(VT, dl, Amt, DAG.getUNDEF(VT), {1, 1, 1, 1});
18098 Amt2 = DAG.getVectorShuffle(VT, dl, Amt, DAG.getUNDEF(VT), {2, 2, 2, 2});
18099 Amt3 = DAG.getVectorShuffle(VT, dl, Amt, DAG.getUNDEF(VT), {3, 3, 3, 3});
18101 // ISD::SHL is handled above but we include it here for completeness.
18104 llvm_unreachable("Unknown target vector shift node");
18106 Opc = X86ISD::VSHL;
18109 Opc = X86ISD::VSRL;
18112 Opc = X86ISD::VSRA;
18115 // The SSE2 shifts use the lower i64 as the same shift amount for
18116 // all lanes and the upper i64 is ignored. These shuffle masks
18117 // optimally zero-extend each lanes on SSE2/SSE41/AVX targets.
18118 SDValue Z = getZeroVector(VT, Subtarget, DAG, dl);
18119 Amt0 = DAG.getVectorShuffle(VT, dl, Amt, Z, {0, 4, -1, -1});
18120 Amt1 = DAG.getVectorShuffle(VT, dl, Amt, Z, {1, 5, -1, -1});
18121 Amt2 = DAG.getVectorShuffle(VT, dl, Amt, Z, {2, 6, -1, -1});
18122 Amt3 = DAG.getVectorShuffle(VT, dl, Amt, Z, {3, 7, -1, -1});
18125 SDValue R0 = DAG.getNode(Opc, dl, VT, R, Amt0);
18126 SDValue R1 = DAG.getNode(Opc, dl, VT, R, Amt1);
18127 SDValue R2 = DAG.getNode(Opc, dl, VT, R, Amt2);
18128 SDValue R3 = DAG.getNode(Opc, dl, VT, R, Amt3);
18129 SDValue R02 = DAG.getVectorShuffle(VT, dl, R0, R2, {0, -1, 6, -1});
18130 SDValue R13 = DAG.getVectorShuffle(VT, dl, R1, R3, {-1, 1, -1, 7});
18131 return DAG.getVectorShuffle(VT, dl, R02, R13, {0, 5, 2, 7});
18134 if (VT == MVT::v16i8 || (VT == MVT::v32i8 && Subtarget->hasInt256())) {
18135 MVT ExtVT = MVT::getVectorVT(MVT::i16, VT.getVectorNumElements() / 2);
18136 unsigned ShiftOpcode = Op->getOpcode();
18138 auto SignBitSelect = [&](MVT SelVT, SDValue Sel, SDValue V0, SDValue V1) {
18139 // On SSE41 targets we make use of the fact that VSELECT lowers
18140 // to PBLENDVB which selects bytes based just on the sign bit.
18141 if (Subtarget->hasSSE41()) {
18142 V0 = DAG.getBitcast(VT, V0);
18143 V1 = DAG.getBitcast(VT, V1);
18144 Sel = DAG.getBitcast(VT, Sel);
18145 return DAG.getBitcast(SelVT,
18146 DAG.getNode(ISD::VSELECT, dl, VT, Sel, V0, V1));
18148 // On pre-SSE41 targets we test for the sign bit by comparing to
18149 // zero - a negative value will set all bits of the lanes to true
18150 // and VSELECT uses that in its OR(AND(V0,C),AND(V1,~C)) lowering.
18151 SDValue Z = getZeroVector(SelVT, Subtarget, DAG, dl);
18152 SDValue C = DAG.getNode(X86ISD::PCMPGT, dl, SelVT, Z, Sel);
18153 return DAG.getNode(ISD::VSELECT, dl, SelVT, C, V0, V1);
18156 // Turn 'a' into a mask suitable for VSELECT: a = a << 5;
18157 // We can safely do this using i16 shifts as we're only interested in
18158 // the 3 lower bits of each byte.
18159 Amt = DAG.getBitcast(ExtVT, Amt);
18160 Amt = DAG.getNode(ISD::SHL, dl, ExtVT, Amt, DAG.getConstant(5, dl, ExtVT));
18161 Amt = DAG.getBitcast(VT, Amt);
18163 if (Op->getOpcode() == ISD::SHL || Op->getOpcode() == ISD::SRL) {
18164 // r = VSELECT(r, shift(r, 4), a);
18166 DAG.getNode(ShiftOpcode, dl, VT, R, DAG.getConstant(4, dl, VT));
18167 R = SignBitSelect(VT, Amt, M, R);
18170 Amt = DAG.getNode(ISD::ADD, dl, VT, Amt, Amt);
18172 // r = VSELECT(r, shift(r, 2), a);
18173 M = DAG.getNode(ShiftOpcode, dl, VT, R, DAG.getConstant(2, dl, VT));
18174 R = SignBitSelect(VT, Amt, M, R);
18177 Amt = DAG.getNode(ISD::ADD, dl, VT, Amt, Amt);
18179 // return VSELECT(r, shift(r, 1), a);
18180 M = DAG.getNode(ShiftOpcode, dl, VT, R, DAG.getConstant(1, dl, VT));
18181 R = SignBitSelect(VT, Amt, M, R);
18185 if (Op->getOpcode() == ISD::SRA) {
18186 // For SRA we need to unpack each byte to the higher byte of a i16 vector
18187 // so we can correctly sign extend. We don't care what happens to the
18189 SDValue ALo = DAG.getNode(X86ISD::UNPCKL, dl, VT, DAG.getUNDEF(VT), Amt);
18190 SDValue AHi = DAG.getNode(X86ISD::UNPCKH, dl, VT, DAG.getUNDEF(VT), Amt);
18191 SDValue RLo = DAG.getNode(X86ISD::UNPCKL, dl, VT, DAG.getUNDEF(VT), R);
18192 SDValue RHi = DAG.getNode(X86ISD::UNPCKH, dl, VT, DAG.getUNDEF(VT), R);
18193 ALo = DAG.getBitcast(ExtVT, ALo);
18194 AHi = DAG.getBitcast(ExtVT, AHi);
18195 RLo = DAG.getBitcast(ExtVT, RLo);
18196 RHi = DAG.getBitcast(ExtVT, RHi);
18198 // r = VSELECT(r, shift(r, 4), a);
18199 SDValue MLo = DAG.getNode(ShiftOpcode, dl, ExtVT, RLo,
18200 DAG.getConstant(4, dl, ExtVT));
18201 SDValue MHi = DAG.getNode(ShiftOpcode, dl, ExtVT, RHi,
18202 DAG.getConstant(4, dl, ExtVT));
18203 RLo = SignBitSelect(ExtVT, ALo, MLo, RLo);
18204 RHi = SignBitSelect(ExtVT, AHi, MHi, RHi);
18207 ALo = DAG.getNode(ISD::ADD, dl, ExtVT, ALo, ALo);
18208 AHi = DAG.getNode(ISD::ADD, dl, ExtVT, AHi, AHi);
18210 // r = VSELECT(r, shift(r, 2), a);
18211 MLo = DAG.getNode(ShiftOpcode, dl, ExtVT, RLo,
18212 DAG.getConstant(2, dl, ExtVT));
18213 MHi = DAG.getNode(ShiftOpcode, dl, ExtVT, RHi,
18214 DAG.getConstant(2, dl, ExtVT));
18215 RLo = SignBitSelect(ExtVT, ALo, MLo, RLo);
18216 RHi = SignBitSelect(ExtVT, AHi, MHi, RHi);
18219 ALo = DAG.getNode(ISD::ADD, dl, ExtVT, ALo, ALo);
18220 AHi = DAG.getNode(ISD::ADD, dl, ExtVT, AHi, AHi);
18222 // r = VSELECT(r, shift(r, 1), a);
18223 MLo = DAG.getNode(ShiftOpcode, dl, ExtVT, RLo,
18224 DAG.getConstant(1, dl, ExtVT));
18225 MHi = DAG.getNode(ShiftOpcode, dl, ExtVT, RHi,
18226 DAG.getConstant(1, dl, ExtVT));
18227 RLo = SignBitSelect(ExtVT, ALo, MLo, RLo);
18228 RHi = SignBitSelect(ExtVT, AHi, MHi, RHi);
18230 // Logical shift the result back to the lower byte, leaving a zero upper
18232 // meaning that we can safely pack with PACKUSWB.
18234 DAG.getNode(ISD::SRL, dl, ExtVT, RLo, DAG.getConstant(8, dl, ExtVT));
18236 DAG.getNode(ISD::SRL, dl, ExtVT, RHi, DAG.getConstant(8, dl, ExtVT));
18237 return DAG.getNode(X86ISD::PACKUS, dl, VT, RLo, RHi);
18241 // It's worth extending once and using the v8i32 shifts for 16-bit types, but
18242 // the extra overheads to get from v16i8 to v8i32 make the existing SSE
18243 // solution better.
18244 if (Subtarget->hasInt256() && VT == MVT::v8i16) {
18245 MVT ExtVT = MVT::v8i32;
18247 Op.getOpcode() == ISD::SRA ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND;
18248 R = DAG.getNode(ExtOpc, dl, ExtVT, R);
18249 Amt = DAG.getNode(ISD::ANY_EXTEND, dl, ExtVT, Amt);
18250 return DAG.getNode(ISD::TRUNCATE, dl, VT,
18251 DAG.getNode(Op.getOpcode(), dl, ExtVT, R, Amt));
18254 if (Subtarget->hasInt256() && VT == MVT::v16i16) {
18255 MVT ExtVT = MVT::v8i32;
18256 SDValue Z = getZeroVector(VT, Subtarget, DAG, dl);
18257 SDValue ALo = DAG.getNode(X86ISD::UNPCKL, dl, VT, Amt, Z);
18258 SDValue AHi = DAG.getNode(X86ISD::UNPCKH, dl, VT, Amt, Z);
18259 SDValue RLo = DAG.getNode(X86ISD::UNPCKL, dl, VT, R, R);
18260 SDValue RHi = DAG.getNode(X86ISD::UNPCKH, dl, VT, R, R);
18261 ALo = DAG.getBitcast(ExtVT, ALo);
18262 AHi = DAG.getBitcast(ExtVT, AHi);
18263 RLo = DAG.getBitcast(ExtVT, RLo);
18264 RHi = DAG.getBitcast(ExtVT, RHi);
18265 SDValue Lo = DAG.getNode(Op.getOpcode(), dl, ExtVT, RLo, ALo);
18266 SDValue Hi = DAG.getNode(Op.getOpcode(), dl, ExtVT, RHi, AHi);
18267 Lo = DAG.getNode(ISD::SRL, dl, ExtVT, Lo, DAG.getConstant(16, dl, ExtVT));
18268 Hi = DAG.getNode(ISD::SRL, dl, ExtVT, Hi, DAG.getConstant(16, dl, ExtVT));
18269 return DAG.getNode(X86ISD::PACKUS, dl, VT, Lo, Hi);
18272 if (VT == MVT::v8i16) {
18273 unsigned ShiftOpcode = Op->getOpcode();
18275 auto SignBitSelect = [&](SDValue Sel, SDValue V0, SDValue V1) {
18276 // On SSE41 targets we make use of the fact that VSELECT lowers
18277 // to PBLENDVB which selects bytes based just on the sign bit.
18278 if (Subtarget->hasSSE41()) {
18279 MVT ExtVT = MVT::getVectorVT(MVT::i8, VT.getVectorNumElements() * 2);
18280 V0 = DAG.getBitcast(ExtVT, V0);
18281 V1 = DAG.getBitcast(ExtVT, V1);
18282 Sel = DAG.getBitcast(ExtVT, Sel);
18283 return DAG.getBitcast(
18284 VT, DAG.getNode(ISD::VSELECT, dl, ExtVT, Sel, V0, V1));
18286 // On pre-SSE41 targets we splat the sign bit - a negative value will
18287 // set all bits of the lanes to true and VSELECT uses that in
18288 // its OR(AND(V0,C),AND(V1,~C)) lowering.
18290 DAG.getNode(ISD::SRA, dl, VT, Sel, DAG.getConstant(15, dl, VT));
18291 return DAG.getNode(ISD::VSELECT, dl, VT, C, V0, V1);
18294 // Turn 'a' into a mask suitable for VSELECT: a = a << 12;
18295 if (Subtarget->hasSSE41()) {
18296 // On SSE41 targets we need to replicate the shift mask in both
18297 // bytes for PBLENDVB.
18300 DAG.getNode(ISD::SHL, dl, VT, Amt, DAG.getConstant(4, dl, VT)),
18301 DAG.getNode(ISD::SHL, dl, VT, Amt, DAG.getConstant(12, dl, VT)));
18303 Amt = DAG.getNode(ISD::SHL, dl, VT, Amt, DAG.getConstant(12, dl, VT));
18306 // r = VSELECT(r, shift(r, 8), a);
18307 SDValue M = DAG.getNode(ShiftOpcode, dl, VT, R, DAG.getConstant(8, dl, VT));
18308 R = SignBitSelect(Amt, M, R);
18311 Amt = DAG.getNode(ISD::ADD, dl, VT, Amt, Amt);
18313 // r = VSELECT(r, shift(r, 4), a);
18314 M = DAG.getNode(ShiftOpcode, dl, VT, R, DAG.getConstant(4, dl, VT));
18315 R = SignBitSelect(Amt, M, R);
18318 Amt = DAG.getNode(ISD::ADD, dl, VT, Amt, Amt);
18320 // r = VSELECT(r, shift(r, 2), a);
18321 M = DAG.getNode(ShiftOpcode, dl, VT, R, DAG.getConstant(2, dl, VT));
18322 R = SignBitSelect(Amt, M, R);
18325 Amt = DAG.getNode(ISD::ADD, dl, VT, Amt, Amt);
18327 // return VSELECT(r, shift(r, 1), a);
18328 M = DAG.getNode(ShiftOpcode, dl, VT, R, DAG.getConstant(1, dl, VT));
18329 R = SignBitSelect(Amt, M, R);
18333 // Decompose 256-bit shifts into smaller 128-bit shifts.
18334 if (VT.is256BitVector()) {
18335 unsigned NumElems = VT.getVectorNumElements();
18336 MVT EltVT = VT.getVectorElementType();
18337 EVT NewVT = MVT::getVectorVT(EltVT, NumElems/2);
18339 // Extract the two vectors
18340 SDValue V1 = Extract128BitVector(R, 0, DAG, dl);
18341 SDValue V2 = Extract128BitVector(R, NumElems/2, DAG, dl);
18343 // Recreate the shift amount vectors
18344 SDValue Amt1, Amt2;
18345 if (Amt.getOpcode() == ISD::BUILD_VECTOR) {
18346 // Constant shift amount
18347 SmallVector<SDValue, 8> Ops(Amt->op_begin(), Amt->op_begin() + NumElems);
18348 ArrayRef<SDValue> Amt1Csts = makeArrayRef(Ops).slice(0, NumElems / 2);
18349 ArrayRef<SDValue> Amt2Csts = makeArrayRef(Ops).slice(NumElems / 2);
18351 Amt1 = DAG.getNode(ISD::BUILD_VECTOR, dl, NewVT, Amt1Csts);
18352 Amt2 = DAG.getNode(ISD::BUILD_VECTOR, dl, NewVT, Amt2Csts);
18354 // Variable shift amount
18355 Amt1 = Extract128BitVector(Amt, 0, DAG, dl);
18356 Amt2 = Extract128BitVector(Amt, NumElems/2, DAG, dl);
18359 // Issue new vector shifts for the smaller types
18360 V1 = DAG.getNode(Op.getOpcode(), dl, NewVT, V1, Amt1);
18361 V2 = DAG.getNode(Op.getOpcode(), dl, NewVT, V2, Amt2);
18363 // Concatenate the result back
18364 return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT, V1, V2);
18370 static SDValue LowerXALUO(SDValue Op, SelectionDAG &DAG) {
18371 // Lower the "add/sub/mul with overflow" instruction into a regular ins plus
18372 // a "setcc" instruction that checks the overflow flag. The "brcond" lowering
18373 // looks for this combo and may remove the "setcc" instruction if the "setcc"
18374 // has only one use.
18375 SDNode *N = Op.getNode();
18376 SDValue LHS = N->getOperand(0);
18377 SDValue RHS = N->getOperand(1);
18378 unsigned BaseOp = 0;
18381 switch (Op.getOpcode()) {
18382 default: llvm_unreachable("Unknown ovf instruction!");
18384 // A subtract of one will be selected as a INC. Note that INC doesn't
18385 // set CF, so we can't do this for UADDO.
18386 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(RHS))
18388 BaseOp = X86ISD::INC;
18389 Cond = X86::COND_O;
18392 BaseOp = X86ISD::ADD;
18393 Cond = X86::COND_O;
18396 BaseOp = X86ISD::ADD;
18397 Cond = X86::COND_B;
18400 // A subtract of one will be selected as a DEC. Note that DEC doesn't
18401 // set CF, so we can't do this for USUBO.
18402 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(RHS))
18404 BaseOp = X86ISD::DEC;
18405 Cond = X86::COND_O;
18408 BaseOp = X86ISD::SUB;
18409 Cond = X86::COND_O;
18412 BaseOp = X86ISD::SUB;
18413 Cond = X86::COND_B;
18416 BaseOp = N->getValueType(0) == MVT::i8 ? X86ISD::SMUL8 : X86ISD::SMUL;
18417 Cond = X86::COND_O;
18419 case ISD::UMULO: { // i64, i8 = umulo lhs, rhs --> i64, i64, i32 umul lhs,rhs
18420 if (N->getValueType(0) == MVT::i8) {
18421 BaseOp = X86ISD::UMUL8;
18422 Cond = X86::COND_O;
18425 SDVTList VTs = DAG.getVTList(N->getValueType(0), N->getValueType(0),
18427 SDValue Sum = DAG.getNode(X86ISD::UMUL, DL, VTs, LHS, RHS);
18430 DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
18431 DAG.getConstant(X86::COND_O, DL, MVT::i32),
18432 SDValue(Sum.getNode(), 2));
18434 return DAG.getNode(ISD::MERGE_VALUES, DL, N->getVTList(), Sum, SetCC);
18438 // Also sets EFLAGS.
18439 SDVTList VTs = DAG.getVTList(N->getValueType(0), MVT::i32);
18440 SDValue Sum = DAG.getNode(BaseOp, DL, VTs, LHS, RHS);
18443 DAG.getNode(X86ISD::SETCC, DL, N->getValueType(1),
18444 DAG.getConstant(Cond, DL, MVT::i32),
18445 SDValue(Sum.getNode(), 1));
18447 return DAG.getNode(ISD::MERGE_VALUES, DL, N->getVTList(), Sum, SetCC);
18450 /// Returns true if the operand type is exactly twice the native width, and
18451 /// the corresponding cmpxchg8b or cmpxchg16b instruction is available.
18452 /// Used to know whether to use cmpxchg8/16b when expanding atomic operations
18453 /// (otherwise we leave them alone to become __sync_fetch_and_... calls).
18454 bool X86TargetLowering::needsCmpXchgNb(Type *MemType) const {
18455 unsigned OpWidth = MemType->getPrimitiveSizeInBits();
18458 return !Subtarget->is64Bit(); // FIXME this should be Subtarget.hasCmpxchg8b
18459 else if (OpWidth == 128)
18460 return Subtarget->hasCmpxchg16b();
18465 bool X86TargetLowering::shouldExpandAtomicStoreInIR(StoreInst *SI) const {
18466 return needsCmpXchgNb(SI->getValueOperand()->getType());
18469 // Note: this turns large loads into lock cmpxchg8b/16b.
18470 // FIXME: On 32 bits x86, fild/movq might be faster than lock cmpxchg8b.
18471 TargetLowering::AtomicExpansionKind
18472 X86TargetLowering::shouldExpandAtomicLoadInIR(LoadInst *LI) const {
18473 auto PTy = cast<PointerType>(LI->getPointerOperand()->getType());
18474 return needsCmpXchgNb(PTy->getElementType()) ? AtomicExpansionKind::CmpXChg
18475 : AtomicExpansionKind::None;
18478 TargetLowering::AtomicExpansionKind
18479 X86TargetLowering::shouldExpandAtomicRMWInIR(AtomicRMWInst *AI) const {
18480 unsigned NativeWidth = Subtarget->is64Bit() ? 64 : 32;
18481 Type *MemType = AI->getType();
18483 // If the operand is too big, we must see if cmpxchg8/16b is available
18484 // and default to library calls otherwise.
18485 if (MemType->getPrimitiveSizeInBits() > NativeWidth) {
18486 return needsCmpXchgNb(MemType) ? AtomicExpansionKind::CmpXChg
18487 : AtomicExpansionKind::None;
18490 AtomicRMWInst::BinOp Op = AI->getOperation();
18493 llvm_unreachable("Unknown atomic operation");
18494 case AtomicRMWInst::Xchg:
18495 case AtomicRMWInst::Add:
18496 case AtomicRMWInst::Sub:
18497 // It's better to use xadd, xsub or xchg for these in all cases.
18498 return AtomicExpansionKind::None;
18499 case AtomicRMWInst::Or:
18500 case AtomicRMWInst::And:
18501 case AtomicRMWInst::Xor:
18502 // If the atomicrmw's result isn't actually used, we can just add a "lock"
18503 // prefix to a normal instruction for these operations.
18504 return !AI->use_empty() ? AtomicExpansionKind::CmpXChg
18505 : AtomicExpansionKind::None;
18506 case AtomicRMWInst::Nand:
18507 case AtomicRMWInst::Max:
18508 case AtomicRMWInst::Min:
18509 case AtomicRMWInst::UMax:
18510 case AtomicRMWInst::UMin:
18511 // These always require a non-trivial set of data operations on x86. We must
18512 // use a cmpxchg loop.
18513 return AtomicExpansionKind::CmpXChg;
18517 static bool hasMFENCE(const X86Subtarget& Subtarget) {
18518 // Use mfence if we have SSE2 or we're on x86-64 (even if we asked for
18519 // no-sse2). There isn't any reason to disable it if the target processor
18521 return Subtarget.hasSSE2() || Subtarget.is64Bit();
18525 X86TargetLowering::lowerIdempotentRMWIntoFencedLoad(AtomicRMWInst *AI) const {
18526 unsigned NativeWidth = Subtarget->is64Bit() ? 64 : 32;
18527 Type *MemType = AI->getType();
18528 // Accesses larger than the native width are turned into cmpxchg/libcalls, so
18529 // there is no benefit in turning such RMWs into loads, and it is actually
18530 // harmful as it introduces a mfence.
18531 if (MemType->getPrimitiveSizeInBits() > NativeWidth)
18534 auto Builder = IRBuilder<>(AI);
18535 Module *M = Builder.GetInsertBlock()->getParent()->getParent();
18536 auto SynchScope = AI->getSynchScope();
18537 // We must restrict the ordering to avoid generating loads with Release or
18538 // ReleaseAcquire orderings.
18539 auto Order = AtomicCmpXchgInst::getStrongestFailureOrdering(AI->getOrdering());
18540 auto Ptr = AI->getPointerOperand();
18542 // Before the load we need a fence. Here is an example lifted from
18543 // http://www.hpl.hp.com/techreports/2012/HPL-2012-68.pdf showing why a fence
18546 // x.store(1, relaxed);
18547 // r1 = y.fetch_add(0, release);
18549 // y.fetch_add(42, acquire);
18550 // r2 = x.load(relaxed);
18551 // r1 = r2 = 0 is impossible, but becomes possible if the idempotent rmw is
18552 // lowered to just a load without a fence. A mfence flushes the store buffer,
18553 // making the optimization clearly correct.
18554 // FIXME: it is required if isAtLeastRelease(Order) but it is not clear
18555 // otherwise, we might be able to be more aggressive on relaxed idempotent
18556 // rmw. In practice, they do not look useful, so we don't try to be
18557 // especially clever.
18558 if (SynchScope == SingleThread)
18559 // FIXME: we could just insert an X86ISD::MEMBARRIER here, except we are at
18560 // the IR level, so we must wrap it in an intrinsic.
18563 if (!hasMFENCE(*Subtarget))
18564 // FIXME: it might make sense to use a locked operation here but on a
18565 // different cache-line to prevent cache-line bouncing. In practice it
18566 // is probably a small win, and x86 processors without mfence are rare
18567 // enough that we do not bother.
18571 llvm::Intrinsic::getDeclaration(M, Intrinsic::x86_sse2_mfence);
18572 Builder.CreateCall(MFence, {});
18574 // Finally we can emit the atomic load.
18575 LoadInst *Loaded = Builder.CreateAlignedLoad(Ptr,
18576 AI->getType()->getPrimitiveSizeInBits());
18577 Loaded->setAtomic(Order, SynchScope);
18578 AI->replaceAllUsesWith(Loaded);
18579 AI->eraseFromParent();
18583 static SDValue LowerATOMIC_FENCE(SDValue Op, const X86Subtarget *Subtarget,
18584 SelectionDAG &DAG) {
18586 AtomicOrdering FenceOrdering = static_cast<AtomicOrdering>(
18587 cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue());
18588 SynchronizationScope FenceScope = static_cast<SynchronizationScope>(
18589 cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue());
18591 // The only fence that needs an instruction is a sequentially-consistent
18592 // cross-thread fence.
18593 if (FenceOrdering == SequentiallyConsistent && FenceScope == CrossThread) {
18594 if (hasMFENCE(*Subtarget))
18595 return DAG.getNode(X86ISD::MFENCE, dl, MVT::Other, Op.getOperand(0));
18597 SDValue Chain = Op.getOperand(0);
18598 SDValue Zero = DAG.getConstant(0, dl, MVT::i32);
18600 DAG.getRegister(X86::ESP, MVT::i32), // Base
18601 DAG.getTargetConstant(1, dl, MVT::i8), // Scale
18602 DAG.getRegister(0, MVT::i32), // Index
18603 DAG.getTargetConstant(0, dl, MVT::i32), // Disp
18604 DAG.getRegister(0, MVT::i32), // Segment.
18608 SDNode *Res = DAG.getMachineNode(X86::OR32mrLocked, dl, MVT::Other, Ops);
18609 return SDValue(Res, 0);
18612 // MEMBARRIER is a compiler barrier; it codegens to a no-op.
18613 return DAG.getNode(X86ISD::MEMBARRIER, dl, MVT::Other, Op.getOperand(0));
18616 static SDValue LowerCMP_SWAP(SDValue Op, const X86Subtarget *Subtarget,
18617 SelectionDAG &DAG) {
18618 MVT T = Op.getSimpleValueType();
18622 switch(T.SimpleTy) {
18623 default: llvm_unreachable("Invalid value type!");
18624 case MVT::i8: Reg = X86::AL; size = 1; break;
18625 case MVT::i16: Reg = X86::AX; size = 2; break;
18626 case MVT::i32: Reg = X86::EAX; size = 4; break;
18628 assert(Subtarget->is64Bit() && "Node not type legal!");
18629 Reg = X86::RAX; size = 8;
18632 SDValue cpIn = DAG.getCopyToReg(Op.getOperand(0), DL, Reg,
18633 Op.getOperand(2), SDValue());
18634 SDValue Ops[] = { cpIn.getValue(0),
18637 DAG.getTargetConstant(size, DL, MVT::i8),
18638 cpIn.getValue(1) };
18639 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Glue);
18640 MachineMemOperand *MMO = cast<AtomicSDNode>(Op)->getMemOperand();
18641 SDValue Result = DAG.getMemIntrinsicNode(X86ISD::LCMPXCHG_DAG, DL, Tys,
18645 DAG.getCopyFromReg(Result.getValue(0), DL, Reg, T, Result.getValue(1));
18646 SDValue EFLAGS = DAG.getCopyFromReg(cpOut.getValue(1), DL, X86::EFLAGS,
18647 MVT::i32, cpOut.getValue(2));
18648 SDValue Success = DAG.getNode(X86ISD::SETCC, DL, Op->getValueType(1),
18649 DAG.getConstant(X86::COND_E, DL, MVT::i8),
18652 DAG.ReplaceAllUsesOfValueWith(Op.getValue(0), cpOut);
18653 DAG.ReplaceAllUsesOfValueWith(Op.getValue(1), Success);
18654 DAG.ReplaceAllUsesOfValueWith(Op.getValue(2), EFLAGS.getValue(1));
18658 static SDValue LowerBITCAST(SDValue Op, const X86Subtarget *Subtarget,
18659 SelectionDAG &DAG) {
18660 MVT SrcVT = Op.getOperand(0).getSimpleValueType();
18661 MVT DstVT = Op.getSimpleValueType();
18663 if (SrcVT == MVT::v2i32 || SrcVT == MVT::v4i16 || SrcVT == MVT::v8i8) {
18664 assert(Subtarget->hasSSE2() && "Requires at least SSE2!");
18665 if (DstVT != MVT::f64)
18666 // This conversion needs to be expanded.
18669 SDValue InVec = Op->getOperand(0);
18671 unsigned NumElts = SrcVT.getVectorNumElements();
18672 EVT SVT = SrcVT.getVectorElementType();
18674 // Widen the vector in input in the case of MVT::v2i32.
18675 // Example: from MVT::v2i32 to MVT::v4i32.
18676 SmallVector<SDValue, 16> Elts;
18677 for (unsigned i = 0, e = NumElts; i != e; ++i)
18678 Elts.push_back(DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, SVT, InVec,
18679 DAG.getIntPtrConstant(i, dl)));
18681 // Explicitly mark the extra elements as Undef.
18682 Elts.append(NumElts, DAG.getUNDEF(SVT));
18684 EVT NewVT = EVT::getVectorVT(*DAG.getContext(), SVT, NumElts * 2);
18685 SDValue BV = DAG.getNode(ISD::BUILD_VECTOR, dl, NewVT, Elts);
18686 SDValue ToV2F64 = DAG.getBitcast(MVT::v2f64, BV);
18687 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, ToV2F64,
18688 DAG.getIntPtrConstant(0, dl));
18691 assert(Subtarget->is64Bit() && !Subtarget->hasSSE2() &&
18692 Subtarget->hasMMX() && "Unexpected custom BITCAST");
18693 assert((DstVT == MVT::i64 ||
18694 (DstVT.isVector() && DstVT.getSizeInBits()==64)) &&
18695 "Unexpected custom BITCAST");
18696 // i64 <=> MMX conversions are Legal.
18697 if (SrcVT==MVT::i64 && DstVT.isVector())
18699 if (DstVT==MVT::i64 && SrcVT.isVector())
18701 // MMX <=> MMX conversions are Legal.
18702 if (SrcVT.isVector() && DstVT.isVector())
18704 // All other conversions need to be expanded.
18708 /// Compute the horizontal sum of bytes in V for the elements of VT.
18710 /// Requires V to be a byte vector and VT to be an integer vector type with
18711 /// wider elements than V's type. The width of the elements of VT determines
18712 /// how many bytes of V are summed horizontally to produce each element of the
18714 static SDValue LowerHorizontalByteSum(SDValue V, MVT VT,
18715 const X86Subtarget *Subtarget,
18716 SelectionDAG &DAG) {
18718 MVT ByteVecVT = V.getSimpleValueType();
18719 MVT EltVT = VT.getVectorElementType();
18720 int NumElts = VT.getVectorNumElements();
18721 assert(ByteVecVT.getVectorElementType() == MVT::i8 &&
18722 "Expected value to have byte element type.");
18723 assert(EltVT != MVT::i8 &&
18724 "Horizontal byte sum only makes sense for wider elements!");
18725 unsigned VecSize = VT.getSizeInBits();
18726 assert(ByteVecVT.getSizeInBits() == VecSize && "Cannot change vector size!");
18728 // PSADBW instruction horizontally add all bytes and leave the result in i64
18729 // chunks, thus directly computes the pop count for v2i64 and v4i64.
18730 if (EltVT == MVT::i64) {
18731 SDValue Zeros = getZeroVector(ByteVecVT, Subtarget, DAG, DL);
18732 V = DAG.getNode(X86ISD::PSADBW, DL, ByteVecVT, V, Zeros);
18733 return DAG.getBitcast(VT, V);
18736 if (EltVT == MVT::i32) {
18737 // We unpack the low half and high half into i32s interleaved with zeros so
18738 // that we can use PSADBW to horizontally sum them. The most useful part of
18739 // this is that it lines up the results of two PSADBW instructions to be
18740 // two v2i64 vectors which concatenated are the 4 population counts. We can
18741 // then use PACKUSWB to shrink and concatenate them into a v4i32 again.
18742 SDValue Zeros = getZeroVector(VT, Subtarget, DAG, DL);
18743 SDValue Low = DAG.getNode(X86ISD::UNPCKL, DL, VT, V, Zeros);
18744 SDValue High = DAG.getNode(X86ISD::UNPCKH, DL, VT, V, Zeros);
18746 // Do the horizontal sums into two v2i64s.
18747 Zeros = getZeroVector(ByteVecVT, Subtarget, DAG, DL);
18748 Low = DAG.getNode(X86ISD::PSADBW, DL, ByteVecVT,
18749 DAG.getBitcast(ByteVecVT, Low), Zeros);
18750 High = DAG.getNode(X86ISD::PSADBW, DL, ByteVecVT,
18751 DAG.getBitcast(ByteVecVT, High), Zeros);
18753 // Merge them together.
18754 MVT ShortVecVT = MVT::getVectorVT(MVT::i16, VecSize / 16);
18755 V = DAG.getNode(X86ISD::PACKUS, DL, ByteVecVT,
18756 DAG.getBitcast(ShortVecVT, Low),
18757 DAG.getBitcast(ShortVecVT, High));
18759 return DAG.getBitcast(VT, V);
18762 // The only element type left is i16.
18763 assert(EltVT == MVT::i16 && "Unknown how to handle type");
18765 // To obtain pop count for each i16 element starting from the pop count for
18766 // i8 elements, shift the i16s left by 8, sum as i8s, and then shift as i16s
18767 // right by 8. It is important to shift as i16s as i8 vector shift isn't
18768 // directly supported.
18769 SmallVector<SDValue, 16> Shifters(NumElts, DAG.getConstant(8, DL, EltVT));
18770 SDValue Shifter = DAG.getNode(ISD::BUILD_VECTOR, DL, VT, Shifters);
18771 SDValue Shl = DAG.getNode(ISD::SHL, DL, VT, DAG.getBitcast(VT, V), Shifter);
18772 V = DAG.getNode(ISD::ADD, DL, ByteVecVT, DAG.getBitcast(ByteVecVT, Shl),
18773 DAG.getBitcast(ByteVecVT, V));
18774 return DAG.getNode(ISD::SRL, DL, VT, DAG.getBitcast(VT, V), Shifter);
18777 static SDValue LowerVectorCTPOPInRegLUT(SDValue Op, SDLoc DL,
18778 const X86Subtarget *Subtarget,
18779 SelectionDAG &DAG) {
18780 MVT VT = Op.getSimpleValueType();
18781 MVT EltVT = VT.getVectorElementType();
18782 unsigned VecSize = VT.getSizeInBits();
18784 // Implement a lookup table in register by using an algorithm based on:
18785 // http://wm.ite.pl/articles/sse-popcount.html
18787 // The general idea is that every lower byte nibble in the input vector is an
18788 // index into a in-register pre-computed pop count table. We then split up the
18789 // input vector in two new ones: (1) a vector with only the shifted-right
18790 // higher nibbles for each byte and (2) a vector with the lower nibbles (and
18791 // masked out higher ones) for each byte. PSHUB is used separately with both
18792 // to index the in-register table. Next, both are added and the result is a
18793 // i8 vector where each element contains the pop count for input byte.
18795 // To obtain the pop count for elements != i8, we follow up with the same
18796 // approach and use additional tricks as described below.
18798 const int LUT[16] = {/* 0 */ 0, /* 1 */ 1, /* 2 */ 1, /* 3 */ 2,
18799 /* 4 */ 1, /* 5 */ 2, /* 6 */ 2, /* 7 */ 3,
18800 /* 8 */ 1, /* 9 */ 2, /* a */ 2, /* b */ 3,
18801 /* c */ 2, /* d */ 3, /* e */ 3, /* f */ 4};
18803 int NumByteElts = VecSize / 8;
18804 MVT ByteVecVT = MVT::getVectorVT(MVT::i8, NumByteElts);
18805 SDValue In = DAG.getBitcast(ByteVecVT, Op);
18806 SmallVector<SDValue, 16> LUTVec;
18807 for (int i = 0; i < NumByteElts; ++i)
18808 LUTVec.push_back(DAG.getConstant(LUT[i % 16], DL, MVT::i8));
18809 SDValue InRegLUT = DAG.getNode(ISD::BUILD_VECTOR, DL, ByteVecVT, LUTVec);
18810 SmallVector<SDValue, 16> Mask0F(NumByteElts,
18811 DAG.getConstant(0x0F, DL, MVT::i8));
18812 SDValue M0F = DAG.getNode(ISD::BUILD_VECTOR, DL, ByteVecVT, Mask0F);
18815 SmallVector<SDValue, 16> Four(NumByteElts, DAG.getConstant(4, DL, MVT::i8));
18816 SDValue FourV = DAG.getNode(ISD::BUILD_VECTOR, DL, ByteVecVT, Four);
18817 SDValue HighNibbles = DAG.getNode(ISD::SRL, DL, ByteVecVT, In, FourV);
18820 SDValue LowNibbles = DAG.getNode(ISD::AND, DL, ByteVecVT, In, M0F);
18822 // The input vector is used as the shuffle mask that index elements into the
18823 // LUT. After counting low and high nibbles, add the vector to obtain the
18824 // final pop count per i8 element.
18825 SDValue HighPopCnt =
18826 DAG.getNode(X86ISD::PSHUFB, DL, ByteVecVT, InRegLUT, HighNibbles);
18827 SDValue LowPopCnt =
18828 DAG.getNode(X86ISD::PSHUFB, DL, ByteVecVT, InRegLUT, LowNibbles);
18829 SDValue PopCnt = DAG.getNode(ISD::ADD, DL, ByteVecVT, HighPopCnt, LowPopCnt);
18831 if (EltVT == MVT::i8)
18834 return LowerHorizontalByteSum(PopCnt, VT, Subtarget, DAG);
18837 static SDValue LowerVectorCTPOPBitmath(SDValue Op, SDLoc DL,
18838 const X86Subtarget *Subtarget,
18839 SelectionDAG &DAG) {
18840 MVT VT = Op.getSimpleValueType();
18841 assert(VT.is128BitVector() &&
18842 "Only 128-bit vector bitmath lowering supported.");
18844 int VecSize = VT.getSizeInBits();
18845 MVT EltVT = VT.getVectorElementType();
18846 int Len = EltVT.getSizeInBits();
18848 // This is the vectorized version of the "best" algorithm from
18849 // http://graphics.stanford.edu/~seander/bithacks.html#CountBitsSetParallel
18850 // with a minor tweak to use a series of adds + shifts instead of vector
18851 // multiplications. Implemented for all integer vector types. We only use
18852 // this when we don't have SSSE3 which allows a LUT-based lowering that is
18853 // much faster, even faster than using native popcnt instructions.
18855 auto GetShift = [&](unsigned OpCode, SDValue V, int Shifter) {
18856 MVT VT = V.getSimpleValueType();
18857 SmallVector<SDValue, 32> Shifters(
18858 VT.getVectorNumElements(),
18859 DAG.getConstant(Shifter, DL, VT.getVectorElementType()));
18860 return DAG.getNode(OpCode, DL, VT, V,
18861 DAG.getNode(ISD::BUILD_VECTOR, DL, VT, Shifters));
18863 auto GetMask = [&](SDValue V, APInt Mask) {
18864 MVT VT = V.getSimpleValueType();
18865 SmallVector<SDValue, 32> Masks(
18866 VT.getVectorNumElements(),
18867 DAG.getConstant(Mask, DL, VT.getVectorElementType()));
18868 return DAG.getNode(ISD::AND, DL, VT, V,
18869 DAG.getNode(ISD::BUILD_VECTOR, DL, VT, Masks));
18872 // We don't want to incur the implicit masks required to SRL vNi8 vectors on
18873 // x86, so set the SRL type to have elements at least i16 wide. This is
18874 // correct because all of our SRLs are followed immediately by a mask anyways
18875 // that handles any bits that sneak into the high bits of the byte elements.
18876 MVT SrlVT = Len > 8 ? VT : MVT::getVectorVT(MVT::i16, VecSize / 16);
18880 // v = v - ((v >> 1) & 0x55555555...)
18882 DAG.getBitcast(VT, GetShift(ISD::SRL, DAG.getBitcast(SrlVT, V), 1));
18883 SDValue And = GetMask(Srl, APInt::getSplat(Len, APInt(8, 0x55)));
18884 V = DAG.getNode(ISD::SUB, DL, VT, V, And);
18886 // v = (v & 0x33333333...) + ((v >> 2) & 0x33333333...)
18887 SDValue AndLHS = GetMask(V, APInt::getSplat(Len, APInt(8, 0x33)));
18888 Srl = DAG.getBitcast(VT, GetShift(ISD::SRL, DAG.getBitcast(SrlVT, V), 2));
18889 SDValue AndRHS = GetMask(Srl, APInt::getSplat(Len, APInt(8, 0x33)));
18890 V = DAG.getNode(ISD::ADD, DL, VT, AndLHS, AndRHS);
18892 // v = (v + (v >> 4)) & 0x0F0F0F0F...
18893 Srl = DAG.getBitcast(VT, GetShift(ISD::SRL, DAG.getBitcast(SrlVT, V), 4));
18894 SDValue Add = DAG.getNode(ISD::ADD, DL, VT, V, Srl);
18895 V = GetMask(Add, APInt::getSplat(Len, APInt(8, 0x0F)));
18897 // At this point, V contains the byte-wise population count, and we are
18898 // merely doing a horizontal sum if necessary to get the wider element
18900 if (EltVT == MVT::i8)
18903 return LowerHorizontalByteSum(
18904 DAG.getBitcast(MVT::getVectorVT(MVT::i8, VecSize / 8), V), VT, Subtarget,
18908 static SDValue LowerVectorCTPOP(SDValue Op, const X86Subtarget *Subtarget,
18909 SelectionDAG &DAG) {
18910 MVT VT = Op.getSimpleValueType();
18911 // FIXME: Need to add AVX-512 support here!
18912 assert((VT.is256BitVector() || VT.is128BitVector()) &&
18913 "Unknown CTPOP type to handle");
18914 SDLoc DL(Op.getNode());
18915 SDValue Op0 = Op.getOperand(0);
18917 if (!Subtarget->hasSSSE3()) {
18918 // We can't use the fast LUT approach, so fall back on vectorized bitmath.
18919 assert(VT.is128BitVector() && "Only 128-bit vectors supported in SSE!");
18920 return LowerVectorCTPOPBitmath(Op0, DL, Subtarget, DAG);
18923 if (VT.is256BitVector() && !Subtarget->hasInt256()) {
18924 unsigned NumElems = VT.getVectorNumElements();
18926 // Extract each 128-bit vector, compute pop count and concat the result.
18927 SDValue LHS = Extract128BitVector(Op0, 0, DAG, DL);
18928 SDValue RHS = Extract128BitVector(Op0, NumElems/2, DAG, DL);
18930 return DAG.getNode(ISD::CONCAT_VECTORS, DL, VT,
18931 LowerVectorCTPOPInRegLUT(LHS, DL, Subtarget, DAG),
18932 LowerVectorCTPOPInRegLUT(RHS, DL, Subtarget, DAG));
18935 return LowerVectorCTPOPInRegLUT(Op0, DL, Subtarget, DAG);
18938 static SDValue LowerCTPOP(SDValue Op, const X86Subtarget *Subtarget,
18939 SelectionDAG &DAG) {
18940 assert(Op.getValueType().isVector() &&
18941 "We only do custom lowering for vector population count.");
18942 return LowerVectorCTPOP(Op, Subtarget, DAG);
18945 static SDValue LowerLOAD_SUB(SDValue Op, SelectionDAG &DAG) {
18946 SDNode *Node = Op.getNode();
18948 EVT T = Node->getValueType(0);
18949 SDValue negOp = DAG.getNode(ISD::SUB, dl, T,
18950 DAG.getConstant(0, dl, T), Node->getOperand(2));
18951 return DAG.getAtomic(ISD::ATOMIC_LOAD_ADD, dl,
18952 cast<AtomicSDNode>(Node)->getMemoryVT(),
18953 Node->getOperand(0),
18954 Node->getOperand(1), negOp,
18955 cast<AtomicSDNode>(Node)->getMemOperand(),
18956 cast<AtomicSDNode>(Node)->getOrdering(),
18957 cast<AtomicSDNode>(Node)->getSynchScope());
18960 static SDValue LowerATOMIC_STORE(SDValue Op, SelectionDAG &DAG) {
18961 SDNode *Node = Op.getNode();
18963 EVT VT = cast<AtomicSDNode>(Node)->getMemoryVT();
18965 // Convert seq_cst store -> xchg
18966 // Convert wide store -> swap (-> cmpxchg8b/cmpxchg16b)
18967 // FIXME: On 32-bit, store -> fist or movq would be more efficient
18968 // (The only way to get a 16-byte store is cmpxchg16b)
18969 // FIXME: 16-byte ATOMIC_SWAP isn't actually hooked up at the moment.
18970 if (cast<AtomicSDNode>(Node)->getOrdering() == SequentiallyConsistent ||
18971 !DAG.getTargetLoweringInfo().isTypeLegal(VT)) {
18972 SDValue Swap = DAG.getAtomic(ISD::ATOMIC_SWAP, dl,
18973 cast<AtomicSDNode>(Node)->getMemoryVT(),
18974 Node->getOperand(0),
18975 Node->getOperand(1), Node->getOperand(2),
18976 cast<AtomicSDNode>(Node)->getMemOperand(),
18977 cast<AtomicSDNode>(Node)->getOrdering(),
18978 cast<AtomicSDNode>(Node)->getSynchScope());
18979 return Swap.getValue(1);
18981 // Other atomic stores have a simple pattern.
18985 static SDValue LowerADDC_ADDE_SUBC_SUBE(SDValue Op, SelectionDAG &DAG) {
18986 EVT VT = Op.getNode()->getSimpleValueType(0);
18988 // Let legalize expand this if it isn't a legal type yet.
18989 if (!DAG.getTargetLoweringInfo().isTypeLegal(VT))
18992 SDVTList VTs = DAG.getVTList(VT, MVT::i32);
18995 bool ExtraOp = false;
18996 switch (Op.getOpcode()) {
18997 default: llvm_unreachable("Invalid code");
18998 case ISD::ADDC: Opc = X86ISD::ADD; break;
18999 case ISD::ADDE: Opc = X86ISD::ADC; ExtraOp = true; break;
19000 case ISD::SUBC: Opc = X86ISD::SUB; break;
19001 case ISD::SUBE: Opc = X86ISD::SBB; ExtraOp = true; break;
19005 return DAG.getNode(Opc, SDLoc(Op), VTs, Op.getOperand(0),
19007 return DAG.getNode(Opc, SDLoc(Op), VTs, Op.getOperand(0),
19008 Op.getOperand(1), Op.getOperand(2));
19011 static SDValue LowerFSINCOS(SDValue Op, const X86Subtarget *Subtarget,
19012 SelectionDAG &DAG) {
19013 assert(Subtarget->isTargetDarwin() && Subtarget->is64Bit());
19015 // For MacOSX, we want to call an alternative entry point: __sincos_stret,
19016 // which returns the values as { float, float } (in XMM0) or
19017 // { double, double } (which is returned in XMM0, XMM1).
19019 SDValue Arg = Op.getOperand(0);
19020 EVT ArgVT = Arg.getValueType();
19021 Type *ArgTy = ArgVT.getTypeForEVT(*DAG.getContext());
19023 TargetLowering::ArgListTy Args;
19024 TargetLowering::ArgListEntry Entry;
19028 Entry.isSExt = false;
19029 Entry.isZExt = false;
19030 Args.push_back(Entry);
19032 bool isF64 = ArgVT == MVT::f64;
19033 // Only optimize x86_64 for now. i386 is a bit messy. For f32,
19034 // the small struct {f32, f32} is returned in (eax, edx). For f64,
19035 // the results are returned via SRet in memory.
19036 const char *LibcallName = isF64 ? "__sincos_stret" : "__sincosf_stret";
19037 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
19039 DAG.getExternalSymbol(LibcallName, TLI.getPointerTy(DAG.getDataLayout()));
19041 Type *RetTy = isF64
19042 ? (Type*)StructType::get(ArgTy, ArgTy, nullptr)
19043 : (Type*)VectorType::get(ArgTy, 4);
19045 TargetLowering::CallLoweringInfo CLI(DAG);
19046 CLI.setDebugLoc(dl).setChain(DAG.getEntryNode())
19047 .setCallee(CallingConv::C, RetTy, Callee, std::move(Args), 0);
19049 std::pair<SDValue, SDValue> CallResult = TLI.LowerCallTo(CLI);
19052 // Returned in xmm0 and xmm1.
19053 return CallResult.first;
19055 // Returned in bits 0:31 and 32:64 xmm0.
19056 SDValue SinVal = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, ArgVT,
19057 CallResult.first, DAG.getIntPtrConstant(0, dl));
19058 SDValue CosVal = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, ArgVT,
19059 CallResult.first, DAG.getIntPtrConstant(1, dl));
19060 SDVTList Tys = DAG.getVTList(ArgVT, ArgVT);
19061 return DAG.getNode(ISD::MERGE_VALUES, dl, Tys, SinVal, CosVal);
19064 static SDValue LowerMSCATTER(SDValue Op, const X86Subtarget *Subtarget,
19065 SelectionDAG &DAG) {
19066 assert(Subtarget->hasAVX512() &&
19067 "MGATHER/MSCATTER are supported on AVX-512 arch only");
19069 MaskedScatterSDNode *N = cast<MaskedScatterSDNode>(Op.getNode());
19070 EVT VT = N->getValue().getValueType();
19071 assert(VT.getScalarSizeInBits() >= 32 && "Unsupported scatter op");
19074 // X86 scatter kills mask register, so its type should be added to
19075 // the list of return values
19076 if (N->getNumValues() == 1) {
19077 SDValue Index = N->getIndex();
19078 if (!Subtarget->hasVLX() && !VT.is512BitVector() &&
19079 !Index.getValueType().is512BitVector())
19080 Index = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v8i64, Index);
19082 SDVTList VTs = DAG.getVTList(N->getMask().getValueType(), MVT::Other);
19083 SDValue Ops[] = { N->getOperand(0), N->getOperand(1), N->getOperand(2),
19084 N->getOperand(3), Index };
19086 SDValue NewScatter = DAG.getMaskedScatter(VTs, VT, dl, Ops, N->getMemOperand());
19087 DAG.ReplaceAllUsesWith(Op, SDValue(NewScatter.getNode(), 1));
19088 return SDValue(NewScatter.getNode(), 0);
19093 static SDValue LowerMGATHER(SDValue Op, const X86Subtarget *Subtarget,
19094 SelectionDAG &DAG) {
19095 assert(Subtarget->hasAVX512() &&
19096 "MGATHER/MSCATTER are supported on AVX-512 arch only");
19098 MaskedGatherSDNode *N = cast<MaskedGatherSDNode>(Op.getNode());
19099 EVT VT = Op.getValueType();
19100 assert(VT.getScalarSizeInBits() >= 32 && "Unsupported gather op");
19103 SDValue Index = N->getIndex();
19104 if (!Subtarget->hasVLX() && !VT.is512BitVector() &&
19105 !Index.getValueType().is512BitVector()) {
19106 Index = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v8i64, Index);
19107 SDValue Ops[] = { N->getOperand(0), N->getOperand(1), N->getOperand(2),
19108 N->getOperand(3), Index };
19109 DAG.UpdateNodeOperands(N, Ops);
19114 SDValue X86TargetLowering::LowerGC_TRANSITION_START(SDValue Op,
19115 SelectionDAG &DAG) const {
19116 // TODO: Eventually, the lowering of these nodes should be informed by or
19117 // deferred to the GC strategy for the function in which they appear. For
19118 // now, however, they must be lowered to something. Since they are logically
19119 // no-ops in the case of a null GC strategy (or a GC strategy which does not
19120 // require special handling for these nodes), lower them as literal NOOPs for
19122 SmallVector<SDValue, 2> Ops;
19124 Ops.push_back(Op.getOperand(0));
19125 if (Op->getGluedNode())
19126 Ops.push_back(Op->getOperand(Op->getNumOperands() - 1));
19129 SDVTList VTs = DAG.getVTList(MVT::Other, MVT::Glue);
19130 SDValue NOOP(DAG.getMachineNode(X86::NOOP, SDLoc(Op), VTs, Ops), 0);
19135 SDValue X86TargetLowering::LowerGC_TRANSITION_END(SDValue Op,
19136 SelectionDAG &DAG) const {
19137 // TODO: Eventually, the lowering of these nodes should be informed by or
19138 // deferred to the GC strategy for the function in which they appear. For
19139 // now, however, they must be lowered to something. Since they are logically
19140 // no-ops in the case of a null GC strategy (or a GC strategy which does not
19141 // require special handling for these nodes), lower them as literal NOOPs for
19143 SmallVector<SDValue, 2> Ops;
19145 Ops.push_back(Op.getOperand(0));
19146 if (Op->getGluedNode())
19147 Ops.push_back(Op->getOperand(Op->getNumOperands() - 1));
19150 SDVTList VTs = DAG.getVTList(MVT::Other, MVT::Glue);
19151 SDValue NOOP(DAG.getMachineNode(X86::NOOP, SDLoc(Op), VTs, Ops), 0);
19156 /// LowerOperation - Provide custom lowering hooks for some operations.
19158 SDValue X86TargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
19159 switch (Op.getOpcode()) {
19160 default: llvm_unreachable("Should not custom lower this!");
19161 case ISD::ATOMIC_FENCE: return LowerATOMIC_FENCE(Op, Subtarget, DAG);
19162 case ISD::ATOMIC_CMP_SWAP_WITH_SUCCESS:
19163 return LowerCMP_SWAP(Op, Subtarget, DAG);
19164 case ISD::CTPOP: return LowerCTPOP(Op, Subtarget, DAG);
19165 case ISD::ATOMIC_LOAD_SUB: return LowerLOAD_SUB(Op,DAG);
19166 case ISD::ATOMIC_STORE: return LowerATOMIC_STORE(Op,DAG);
19167 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG);
19168 case ISD::CONCAT_VECTORS: return LowerCONCAT_VECTORS(Op, Subtarget, DAG);
19169 case ISD::VECTOR_SHUFFLE: return lowerVectorShuffle(Op, Subtarget, DAG);
19170 case ISD::VSELECT: return LowerVSELECT(Op, DAG);
19171 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
19172 case ISD::INSERT_VECTOR_ELT: return LowerINSERT_VECTOR_ELT(Op, DAG);
19173 case ISD::EXTRACT_SUBVECTOR: return LowerEXTRACT_SUBVECTOR(Op,Subtarget,DAG);
19174 case ISD::INSERT_SUBVECTOR: return LowerINSERT_SUBVECTOR(Op, Subtarget,DAG);
19175 case ISD::SCALAR_TO_VECTOR: return LowerSCALAR_TO_VECTOR(Op, DAG);
19176 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
19177 case ISD::GlobalAddress: return LowerGlobalAddress(Op, DAG);
19178 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
19179 case ISD::ExternalSymbol: return LowerExternalSymbol(Op, DAG);
19180 case ISD::BlockAddress: return LowerBlockAddress(Op, DAG);
19181 case ISD::SHL_PARTS:
19182 case ISD::SRA_PARTS:
19183 case ISD::SRL_PARTS: return LowerShiftParts(Op, DAG);
19184 case ISD::SINT_TO_FP: return LowerSINT_TO_FP(Op, DAG);
19185 case ISD::UINT_TO_FP: return LowerUINT_TO_FP(Op, DAG);
19186 case ISD::TRUNCATE: return LowerTRUNCATE(Op, DAG);
19187 case ISD::ZERO_EXTEND: return LowerZERO_EXTEND(Op, Subtarget, DAG);
19188 case ISD::SIGN_EXTEND: return LowerSIGN_EXTEND(Op, Subtarget, DAG);
19189 case ISD::ANY_EXTEND: return LowerANY_EXTEND(Op, Subtarget, DAG);
19190 case ISD::SIGN_EXTEND_VECTOR_INREG:
19191 return LowerSIGN_EXTEND_VECTOR_INREG(Op, Subtarget, DAG);
19192 case ISD::FP_TO_SINT: return LowerFP_TO_SINT(Op, DAG);
19193 case ISD::FP_TO_UINT: return LowerFP_TO_UINT(Op, DAG);
19194 case ISD::FP_EXTEND: return LowerFP_EXTEND(Op, DAG);
19195 case ISD::LOAD: return LowerExtendedLoad(Op, Subtarget, DAG);
19197 case ISD::FNEG: return LowerFABSorFNEG(Op, DAG);
19198 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
19199 case ISD::FGETSIGN: return LowerFGETSIGN(Op, DAG);
19200 case ISD::SETCC: return LowerSETCC(Op, DAG);
19201 case ISD::SELECT: return LowerSELECT(Op, DAG);
19202 case ISD::BRCOND: return LowerBRCOND(Op, DAG);
19203 case ISD::JumpTable: return LowerJumpTable(Op, DAG);
19204 case ISD::VASTART: return LowerVASTART(Op, DAG);
19205 case ISD::VAARG: return LowerVAARG(Op, DAG);
19206 case ISD::VACOPY: return LowerVACOPY(Op, Subtarget, DAG);
19207 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, Subtarget, DAG);
19208 case ISD::INTRINSIC_VOID:
19209 case ISD::INTRINSIC_W_CHAIN: return LowerINTRINSIC_W_CHAIN(Op, Subtarget, DAG);
19210 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
19211 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
19212 case ISD::FRAME_TO_ARGS_OFFSET:
19213 return LowerFRAME_TO_ARGS_OFFSET(Op, DAG);
19214 case ISD::DYNAMIC_STACKALLOC: return LowerDYNAMIC_STACKALLOC(Op, DAG);
19215 case ISD::EH_RETURN: return LowerEH_RETURN(Op, DAG);
19216 case ISD::EH_SJLJ_SETJMP: return lowerEH_SJLJ_SETJMP(Op, DAG);
19217 case ISD::EH_SJLJ_LONGJMP: return lowerEH_SJLJ_LONGJMP(Op, DAG);
19218 case ISD::INIT_TRAMPOLINE: return LowerINIT_TRAMPOLINE(Op, DAG);
19219 case ISD::ADJUST_TRAMPOLINE: return LowerADJUST_TRAMPOLINE(Op, DAG);
19220 case ISD::FLT_ROUNDS_: return LowerFLT_ROUNDS_(Op, DAG);
19221 case ISD::CTLZ: return LowerCTLZ(Op, DAG);
19222 case ISD::CTLZ_ZERO_UNDEF: return LowerCTLZ_ZERO_UNDEF(Op, DAG);
19224 case ISD::CTTZ_ZERO_UNDEF: return LowerCTTZ(Op, DAG);
19225 case ISD::MUL: return LowerMUL(Op, Subtarget, DAG);
19226 case ISD::UMUL_LOHI:
19227 case ISD::SMUL_LOHI: return LowerMUL_LOHI(Op, Subtarget, DAG);
19230 case ISD::SHL: return LowerShift(Op, Subtarget, DAG);
19236 case ISD::UMULO: return LowerXALUO(Op, DAG);
19237 case ISD::READCYCLECOUNTER: return LowerREADCYCLECOUNTER(Op, Subtarget,DAG);
19238 case ISD::BITCAST: return LowerBITCAST(Op, Subtarget, DAG);
19242 case ISD::SUBE: return LowerADDC_ADDE_SUBC_SUBE(Op, DAG);
19243 case ISD::ADD: return LowerADD(Op, DAG);
19244 case ISD::SUB: return LowerSUB(Op, DAG);
19248 case ISD::UMIN: return LowerMINMAX(Op, DAG);
19249 case ISD::FSINCOS: return LowerFSINCOS(Op, Subtarget, DAG);
19250 case ISD::MGATHER: return LowerMGATHER(Op, Subtarget, DAG);
19251 case ISD::MSCATTER: return LowerMSCATTER(Op, Subtarget, DAG);
19252 case ISD::GC_TRANSITION_START:
19253 return LowerGC_TRANSITION_START(Op, DAG);
19254 case ISD::GC_TRANSITION_END: return LowerGC_TRANSITION_END(Op, DAG);
19258 /// ReplaceNodeResults - Replace a node with an illegal result type
19259 /// with a new node built out of custom code.
19260 void X86TargetLowering::ReplaceNodeResults(SDNode *N,
19261 SmallVectorImpl<SDValue>&Results,
19262 SelectionDAG &DAG) const {
19264 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
19265 switch (N->getOpcode()) {
19267 llvm_unreachable("Do not know how to custom type legalize this operation!");
19268 // We might have generated v2f32 FMIN/FMAX operations. Widen them to v4f32.
19269 case X86ISD::FMINC:
19271 case X86ISD::FMAXC:
19272 case X86ISD::FMAX: {
19273 EVT VT = N->getValueType(0);
19274 if (VT != MVT::v2f32)
19275 llvm_unreachable("Unexpected type (!= v2f32) on FMIN/FMAX.");
19276 SDValue UNDEF = DAG.getUNDEF(VT);
19277 SDValue LHS = DAG.getNode(ISD::CONCAT_VECTORS, dl, MVT::v4f32,
19278 N->getOperand(0), UNDEF);
19279 SDValue RHS = DAG.getNode(ISD::CONCAT_VECTORS, dl, MVT::v4f32,
19280 N->getOperand(1), UNDEF);
19281 Results.push_back(DAG.getNode(N->getOpcode(), dl, MVT::v4f32, LHS, RHS));
19284 case ISD::SIGN_EXTEND_INREG:
19289 // We don't want to expand or promote these.
19296 case ISD::UDIVREM: {
19297 SDValue V = LowerWin64_i128OP(SDValue(N,0), DAG);
19298 Results.push_back(V);
19301 case ISD::FP_TO_SINT:
19302 case ISD::FP_TO_UINT: {
19303 bool IsSigned = N->getOpcode() == ISD::FP_TO_SINT;
19305 std::pair<SDValue,SDValue> Vals =
19306 FP_TO_INTHelper(SDValue(N, 0), DAG, IsSigned, /*IsReplace=*/ true);
19307 SDValue FIST = Vals.first, StackSlot = Vals.second;
19308 if (FIST.getNode()) {
19309 EVT VT = N->getValueType(0);
19310 // Return a load from the stack slot.
19311 if (StackSlot.getNode())
19312 Results.push_back(DAG.getLoad(VT, dl, FIST, StackSlot,
19313 MachinePointerInfo(),
19314 false, false, false, 0));
19316 Results.push_back(FIST);
19320 case ISD::UINT_TO_FP: {
19321 assert(Subtarget->hasSSE2() && "Requires at least SSE2!");
19322 if (N->getOperand(0).getValueType() != MVT::v2i32 ||
19323 N->getValueType(0) != MVT::v2f32)
19325 SDValue ZExtIn = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v2i64,
19327 SDValue Bias = DAG.getConstantFP(BitsToDouble(0x4330000000000000ULL), dl,
19329 SDValue VBias = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v2f64, Bias, Bias);
19330 SDValue Or = DAG.getNode(ISD::OR, dl, MVT::v2i64, ZExtIn,
19331 DAG.getBitcast(MVT::v2i64, VBias));
19332 Or = DAG.getBitcast(MVT::v2f64, Or);
19333 // TODO: Are there any fast-math-flags to propagate here?
19334 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::v2f64, Or, VBias);
19335 Results.push_back(DAG.getNode(X86ISD::VFPROUND, dl, MVT::v4f32, Sub));
19338 case ISD::FP_ROUND: {
19339 if (!TLI.isTypeLegal(N->getOperand(0).getValueType()))
19341 SDValue V = DAG.getNode(X86ISD::VFPROUND, dl, MVT::v4f32, N->getOperand(0));
19342 Results.push_back(V);
19345 case ISD::FP_EXTEND: {
19346 // Right now, only MVT::v2f32 has OperationAction for FP_EXTEND.
19347 // No other ValueType for FP_EXTEND should reach this point.
19348 assert(N->getValueType(0) == MVT::v2f32 &&
19349 "Do not know how to legalize this Node");
19352 case ISD::INTRINSIC_W_CHAIN: {
19353 unsigned IntNo = cast<ConstantSDNode>(N->getOperand(1))->getZExtValue();
19355 default : llvm_unreachable("Do not know how to custom type "
19356 "legalize this intrinsic operation!");
19357 case Intrinsic::x86_rdtsc:
19358 return getReadTimeStampCounter(N, dl, X86ISD::RDTSC_DAG, DAG, Subtarget,
19360 case Intrinsic::x86_rdtscp:
19361 return getReadTimeStampCounter(N, dl, X86ISD::RDTSCP_DAG, DAG, Subtarget,
19363 case Intrinsic::x86_rdpmc:
19364 return getReadPerformanceCounter(N, dl, DAG, Subtarget, Results);
19367 case ISD::READCYCLECOUNTER: {
19368 return getReadTimeStampCounter(N, dl, X86ISD::RDTSC_DAG, DAG, Subtarget,
19371 case ISD::ATOMIC_CMP_SWAP_WITH_SUCCESS: {
19372 EVT T = N->getValueType(0);
19373 assert((T == MVT::i64 || T == MVT::i128) && "can only expand cmpxchg pair");
19374 bool Regs64bit = T == MVT::i128;
19375 EVT HalfT = Regs64bit ? MVT::i64 : MVT::i32;
19376 SDValue cpInL, cpInH;
19377 cpInL = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, HalfT, N->getOperand(2),
19378 DAG.getConstant(0, dl, HalfT));
19379 cpInH = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, HalfT, N->getOperand(2),
19380 DAG.getConstant(1, dl, HalfT));
19381 cpInL = DAG.getCopyToReg(N->getOperand(0), dl,
19382 Regs64bit ? X86::RAX : X86::EAX,
19384 cpInH = DAG.getCopyToReg(cpInL.getValue(0), dl,
19385 Regs64bit ? X86::RDX : X86::EDX,
19386 cpInH, cpInL.getValue(1));
19387 SDValue swapInL, swapInH;
19388 swapInL = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, HalfT, N->getOperand(3),
19389 DAG.getConstant(0, dl, HalfT));
19390 swapInH = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, HalfT, N->getOperand(3),
19391 DAG.getConstant(1, dl, HalfT));
19392 swapInL = DAG.getCopyToReg(cpInH.getValue(0), dl,
19393 Regs64bit ? X86::RBX : X86::EBX,
19394 swapInL, cpInH.getValue(1));
19395 swapInH = DAG.getCopyToReg(swapInL.getValue(0), dl,
19396 Regs64bit ? X86::RCX : X86::ECX,
19397 swapInH, swapInL.getValue(1));
19398 SDValue Ops[] = { swapInH.getValue(0),
19400 swapInH.getValue(1) };
19401 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Glue);
19402 MachineMemOperand *MMO = cast<AtomicSDNode>(N)->getMemOperand();
19403 unsigned Opcode = Regs64bit ? X86ISD::LCMPXCHG16_DAG :
19404 X86ISD::LCMPXCHG8_DAG;
19405 SDValue Result = DAG.getMemIntrinsicNode(Opcode, dl, Tys, Ops, T, MMO);
19406 SDValue cpOutL = DAG.getCopyFromReg(Result.getValue(0), dl,
19407 Regs64bit ? X86::RAX : X86::EAX,
19408 HalfT, Result.getValue(1));
19409 SDValue cpOutH = DAG.getCopyFromReg(cpOutL.getValue(1), dl,
19410 Regs64bit ? X86::RDX : X86::EDX,
19411 HalfT, cpOutL.getValue(2));
19412 SDValue OpsF[] = { cpOutL.getValue(0), cpOutH.getValue(0)};
19414 SDValue EFLAGS = DAG.getCopyFromReg(cpOutH.getValue(1), dl, X86::EFLAGS,
19415 MVT::i32, cpOutH.getValue(2));
19417 DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
19418 DAG.getConstant(X86::COND_E, dl, MVT::i8), EFLAGS);
19419 Success = DAG.getZExtOrTrunc(Success, dl, N->getValueType(1));
19421 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, T, OpsF));
19422 Results.push_back(Success);
19423 Results.push_back(EFLAGS.getValue(1));
19426 case ISD::ATOMIC_SWAP:
19427 case ISD::ATOMIC_LOAD_ADD:
19428 case ISD::ATOMIC_LOAD_SUB:
19429 case ISD::ATOMIC_LOAD_AND:
19430 case ISD::ATOMIC_LOAD_OR:
19431 case ISD::ATOMIC_LOAD_XOR:
19432 case ISD::ATOMIC_LOAD_NAND:
19433 case ISD::ATOMIC_LOAD_MIN:
19434 case ISD::ATOMIC_LOAD_MAX:
19435 case ISD::ATOMIC_LOAD_UMIN:
19436 case ISD::ATOMIC_LOAD_UMAX:
19437 case ISD::ATOMIC_LOAD: {
19438 // Delegate to generic TypeLegalization. Situations we can really handle
19439 // should have already been dealt with by AtomicExpandPass.cpp.
19442 case ISD::BITCAST: {
19443 assert(Subtarget->hasSSE2() && "Requires at least SSE2!");
19444 EVT DstVT = N->getValueType(0);
19445 EVT SrcVT = N->getOperand(0)->getValueType(0);
19447 if (SrcVT != MVT::f64 ||
19448 (DstVT != MVT::v2i32 && DstVT != MVT::v4i16 && DstVT != MVT::v8i8))
19451 unsigned NumElts = DstVT.getVectorNumElements();
19452 EVT SVT = DstVT.getVectorElementType();
19453 EVT WiderVT = EVT::getVectorVT(*DAG.getContext(), SVT, NumElts * 2);
19454 SDValue Expanded = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
19455 MVT::v2f64, N->getOperand(0));
19456 SDValue ToVecInt = DAG.getBitcast(WiderVT, Expanded);
19458 if (ExperimentalVectorWideningLegalization) {
19459 // If we are legalizing vectors by widening, we already have the desired
19460 // legal vector type, just return it.
19461 Results.push_back(ToVecInt);
19465 SmallVector<SDValue, 8> Elts;
19466 for (unsigned i = 0, e = NumElts; i != e; ++i)
19467 Elts.push_back(DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, SVT,
19468 ToVecInt, DAG.getIntPtrConstant(i, dl)));
19470 Results.push_back(DAG.getNode(ISD::BUILD_VECTOR, dl, DstVT, Elts));
19475 const char *X86TargetLowering::getTargetNodeName(unsigned Opcode) const {
19476 switch ((X86ISD::NodeType)Opcode) {
19477 case X86ISD::FIRST_NUMBER: break;
19478 case X86ISD::BSF: return "X86ISD::BSF";
19479 case X86ISD::BSR: return "X86ISD::BSR";
19480 case X86ISD::SHLD: return "X86ISD::SHLD";
19481 case X86ISD::SHRD: return "X86ISD::SHRD";
19482 case X86ISD::FAND: return "X86ISD::FAND";
19483 case X86ISD::FANDN: return "X86ISD::FANDN";
19484 case X86ISD::FOR: return "X86ISD::FOR";
19485 case X86ISD::FXOR: return "X86ISD::FXOR";
19486 case X86ISD::FILD: return "X86ISD::FILD";
19487 case X86ISD::FILD_FLAG: return "X86ISD::FILD_FLAG";
19488 case X86ISD::FP_TO_INT16_IN_MEM: return "X86ISD::FP_TO_INT16_IN_MEM";
19489 case X86ISD::FP_TO_INT32_IN_MEM: return "X86ISD::FP_TO_INT32_IN_MEM";
19490 case X86ISD::FP_TO_INT64_IN_MEM: return "X86ISD::FP_TO_INT64_IN_MEM";
19491 case X86ISD::FLD: return "X86ISD::FLD";
19492 case X86ISD::FST: return "X86ISD::FST";
19493 case X86ISD::CALL: return "X86ISD::CALL";
19494 case X86ISD::RDTSC_DAG: return "X86ISD::RDTSC_DAG";
19495 case X86ISD::RDTSCP_DAG: return "X86ISD::RDTSCP_DAG";
19496 case X86ISD::RDPMC_DAG: return "X86ISD::RDPMC_DAG";
19497 case X86ISD::BT: return "X86ISD::BT";
19498 case X86ISD::CMP: return "X86ISD::CMP";
19499 case X86ISD::COMI: return "X86ISD::COMI";
19500 case X86ISD::UCOMI: return "X86ISD::UCOMI";
19501 case X86ISD::CMPM: return "X86ISD::CMPM";
19502 case X86ISD::CMPMU: return "X86ISD::CMPMU";
19503 case X86ISD::CMPM_RND: return "X86ISD::CMPM_RND";
19504 case X86ISD::SETCC: return "X86ISD::SETCC";
19505 case X86ISD::SETCC_CARRY: return "X86ISD::SETCC_CARRY";
19506 case X86ISD::FSETCC: return "X86ISD::FSETCC";
19507 case X86ISD::FGETSIGNx86: return "X86ISD::FGETSIGNx86";
19508 case X86ISD::CMOV: return "X86ISD::CMOV";
19509 case X86ISD::BRCOND: return "X86ISD::BRCOND";
19510 case X86ISD::RET_FLAG: return "X86ISD::RET_FLAG";
19511 case X86ISD::REP_STOS: return "X86ISD::REP_STOS";
19512 case X86ISD::REP_MOVS: return "X86ISD::REP_MOVS";
19513 case X86ISD::GlobalBaseReg: return "X86ISD::GlobalBaseReg";
19514 case X86ISD::Wrapper: return "X86ISD::Wrapper";
19515 case X86ISD::WrapperRIP: return "X86ISD::WrapperRIP";
19516 case X86ISD::MOVDQ2Q: return "X86ISD::MOVDQ2Q";
19517 case X86ISD::MMX_MOVD2W: return "X86ISD::MMX_MOVD2W";
19518 case X86ISD::MMX_MOVW2D: return "X86ISD::MMX_MOVW2D";
19519 case X86ISD::PEXTRB: return "X86ISD::PEXTRB";
19520 case X86ISD::PEXTRW: return "X86ISD::PEXTRW";
19521 case X86ISD::INSERTPS: return "X86ISD::INSERTPS";
19522 case X86ISD::PINSRB: return "X86ISD::PINSRB";
19523 case X86ISD::PINSRW: return "X86ISD::PINSRW";
19524 case X86ISD::MMX_PINSRW: return "X86ISD::MMX_PINSRW";
19525 case X86ISD::PSHUFB: return "X86ISD::PSHUFB";
19526 case X86ISD::ANDNP: return "X86ISD::ANDNP";
19527 case X86ISD::PSIGN: return "X86ISD::PSIGN";
19528 case X86ISD::BLENDI: return "X86ISD::BLENDI";
19529 case X86ISD::SHRUNKBLEND: return "X86ISD::SHRUNKBLEND";
19530 case X86ISD::ADDUS: return "X86ISD::ADDUS";
19531 case X86ISD::SUBUS: return "X86ISD::SUBUS";
19532 case X86ISD::HADD: return "X86ISD::HADD";
19533 case X86ISD::HSUB: return "X86ISD::HSUB";
19534 case X86ISD::FHADD: return "X86ISD::FHADD";
19535 case X86ISD::FHSUB: return "X86ISD::FHSUB";
19536 case X86ISD::ABS: return "X86ISD::ABS";
19537 case X86ISD::CONFLICT: return "X86ISD::CONFLICT";
19538 case X86ISD::FMAX: return "X86ISD::FMAX";
19539 case X86ISD::FMAX_RND: return "X86ISD::FMAX_RND";
19540 case X86ISD::FMIN: return "X86ISD::FMIN";
19541 case X86ISD::FMIN_RND: return "X86ISD::FMIN_RND";
19542 case X86ISD::FMAXC: return "X86ISD::FMAXC";
19543 case X86ISD::FMINC: return "X86ISD::FMINC";
19544 case X86ISD::FRSQRT: return "X86ISD::FRSQRT";
19545 case X86ISD::FRCP: return "X86ISD::FRCP";
19546 case X86ISD::EXTRQI: return "X86ISD::EXTRQI";
19547 case X86ISD::INSERTQI: return "X86ISD::INSERTQI";
19548 case X86ISD::TLSADDR: return "X86ISD::TLSADDR";
19549 case X86ISD::TLSBASEADDR: return "X86ISD::TLSBASEADDR";
19550 case X86ISD::TLSCALL: return "X86ISD::TLSCALL";
19551 case X86ISD::EH_SJLJ_SETJMP: return "X86ISD::EH_SJLJ_SETJMP";
19552 case X86ISD::EH_SJLJ_LONGJMP: return "X86ISD::EH_SJLJ_LONGJMP";
19553 case X86ISD::EH_RETURN: return "X86ISD::EH_RETURN";
19554 case X86ISD::TC_RETURN: return "X86ISD::TC_RETURN";
19555 case X86ISD::FNSTCW16m: return "X86ISD::FNSTCW16m";
19556 case X86ISD::FNSTSW16r: return "X86ISD::FNSTSW16r";
19557 case X86ISD::LCMPXCHG_DAG: return "X86ISD::LCMPXCHG_DAG";
19558 case X86ISD::LCMPXCHG8_DAG: return "X86ISD::LCMPXCHG8_DAG";
19559 case X86ISD::LCMPXCHG16_DAG: return "X86ISD::LCMPXCHG16_DAG";
19560 case X86ISD::VZEXT_MOVL: return "X86ISD::VZEXT_MOVL";
19561 case X86ISD::VZEXT_LOAD: return "X86ISD::VZEXT_LOAD";
19562 case X86ISD::VZEXT: return "X86ISD::VZEXT";
19563 case X86ISD::VSEXT: return "X86ISD::VSEXT";
19564 case X86ISD::VTRUNC: return "X86ISD::VTRUNC";
19565 case X86ISD::VTRUNCS: return "X86ISD::VTRUNCS";
19566 case X86ISD::VTRUNCUS: return "X86ISD::VTRUNCUS";
19567 case X86ISD::VINSERT: return "X86ISD::VINSERT";
19568 case X86ISD::VFPEXT: return "X86ISD::VFPEXT";
19569 case X86ISD::VFPROUND: return "X86ISD::VFPROUND";
19570 case X86ISD::CVTDQ2PD: return "X86ISD::CVTDQ2PD";
19571 case X86ISD::CVTUDQ2PD: return "X86ISD::CVTUDQ2PD";
19572 case X86ISD::VSHLDQ: return "X86ISD::VSHLDQ";
19573 case X86ISD::VSRLDQ: return "X86ISD::VSRLDQ";
19574 case X86ISD::VSHL: return "X86ISD::VSHL";
19575 case X86ISD::VSRL: return "X86ISD::VSRL";
19576 case X86ISD::VSRA: return "X86ISD::VSRA";
19577 case X86ISD::VSHLI: return "X86ISD::VSHLI";
19578 case X86ISD::VSRLI: return "X86ISD::VSRLI";
19579 case X86ISD::VSRAI: return "X86ISD::VSRAI";
19580 case X86ISD::CMPP: return "X86ISD::CMPP";
19581 case X86ISD::PCMPEQ: return "X86ISD::PCMPEQ";
19582 case X86ISD::PCMPGT: return "X86ISD::PCMPGT";
19583 case X86ISD::PCMPEQM: return "X86ISD::PCMPEQM";
19584 case X86ISD::PCMPGTM: return "X86ISD::PCMPGTM";
19585 case X86ISD::ADD: return "X86ISD::ADD";
19586 case X86ISD::SUB: return "X86ISD::SUB";
19587 case X86ISD::ADC: return "X86ISD::ADC";
19588 case X86ISD::SBB: return "X86ISD::SBB";
19589 case X86ISD::SMUL: return "X86ISD::SMUL";
19590 case X86ISD::UMUL: return "X86ISD::UMUL";
19591 case X86ISD::SMUL8: return "X86ISD::SMUL8";
19592 case X86ISD::UMUL8: return "X86ISD::UMUL8";
19593 case X86ISD::SDIVREM8_SEXT_HREG: return "X86ISD::SDIVREM8_SEXT_HREG";
19594 case X86ISD::UDIVREM8_ZEXT_HREG: return "X86ISD::UDIVREM8_ZEXT_HREG";
19595 case X86ISD::INC: return "X86ISD::INC";
19596 case X86ISD::DEC: return "X86ISD::DEC";
19597 case X86ISD::OR: return "X86ISD::OR";
19598 case X86ISD::XOR: return "X86ISD::XOR";
19599 case X86ISD::AND: return "X86ISD::AND";
19600 case X86ISD::BEXTR: return "X86ISD::BEXTR";
19601 case X86ISD::MUL_IMM: return "X86ISD::MUL_IMM";
19602 case X86ISD::PTEST: return "X86ISD::PTEST";
19603 case X86ISD::TESTP: return "X86ISD::TESTP";
19604 case X86ISD::TESTM: return "X86ISD::TESTM";
19605 case X86ISD::TESTNM: return "X86ISD::TESTNM";
19606 case X86ISD::KORTEST: return "X86ISD::KORTEST";
19607 case X86ISD::KTEST: return "X86ISD::KTEST";
19608 case X86ISD::PACKSS: return "X86ISD::PACKSS";
19609 case X86ISD::PACKUS: return "X86ISD::PACKUS";
19610 case X86ISD::PALIGNR: return "X86ISD::PALIGNR";
19611 case X86ISD::VALIGN: return "X86ISD::VALIGN";
19612 case X86ISD::PSHUFD: return "X86ISD::PSHUFD";
19613 case X86ISD::PSHUFHW: return "X86ISD::PSHUFHW";
19614 case X86ISD::PSHUFLW: return "X86ISD::PSHUFLW";
19615 case X86ISD::SHUFP: return "X86ISD::SHUFP";
19616 case X86ISD::SHUF128: return "X86ISD::SHUF128";
19617 case X86ISD::MOVLHPS: return "X86ISD::MOVLHPS";
19618 case X86ISD::MOVLHPD: return "X86ISD::MOVLHPD";
19619 case X86ISD::MOVHLPS: return "X86ISD::MOVHLPS";
19620 case X86ISD::MOVLPS: return "X86ISD::MOVLPS";
19621 case X86ISD::MOVLPD: return "X86ISD::MOVLPD";
19622 case X86ISD::MOVDDUP: return "X86ISD::MOVDDUP";
19623 case X86ISD::MOVSHDUP: return "X86ISD::MOVSHDUP";
19624 case X86ISD::MOVSLDUP: return "X86ISD::MOVSLDUP";
19625 case X86ISD::MOVSD: return "X86ISD::MOVSD";
19626 case X86ISD::MOVSS: return "X86ISD::MOVSS";
19627 case X86ISD::UNPCKL: return "X86ISD::UNPCKL";
19628 case X86ISD::UNPCKH: return "X86ISD::UNPCKH";
19629 case X86ISD::VBROADCAST: return "X86ISD::VBROADCAST";
19630 case X86ISD::SUBV_BROADCAST: return "X86ISD::SUBV_BROADCAST";
19631 case X86ISD::VEXTRACT: return "X86ISD::VEXTRACT";
19632 case X86ISD::VPERMILPV: return "X86ISD::VPERMILPV";
19633 case X86ISD::VPERMILPI: return "X86ISD::VPERMILPI";
19634 case X86ISD::VPERM2X128: return "X86ISD::VPERM2X128";
19635 case X86ISD::VPERMV: return "X86ISD::VPERMV";
19636 case X86ISD::VPERMV3: return "X86ISD::VPERMV3";
19637 case X86ISD::VPERMIV3: return "X86ISD::VPERMIV3";
19638 case X86ISD::VPERMI: return "X86ISD::VPERMI";
19639 case X86ISD::VFIXUPIMM: return "X86ISD::VFIXUPIMM";
19640 case X86ISD::VRANGE: return "X86ISD::VRANGE";
19641 case X86ISD::PMULUDQ: return "X86ISD::PMULUDQ";
19642 case X86ISD::PMULDQ: return "X86ISD::PMULDQ";
19643 case X86ISD::PSADBW: return "X86ISD::PSADBW";
19644 case X86ISD::DBPSADBW: return "X86ISD::DBPSADBW";
19645 case X86ISD::VASTART_SAVE_XMM_REGS: return "X86ISD::VASTART_SAVE_XMM_REGS";
19646 case X86ISD::VAARG_64: return "X86ISD::VAARG_64";
19647 case X86ISD::WIN_ALLOCA: return "X86ISD::WIN_ALLOCA";
19648 case X86ISD::MEMBARRIER: return "X86ISD::MEMBARRIER";
19649 case X86ISD::MFENCE: return "X86ISD::MFENCE";
19650 case X86ISD::SFENCE: return "X86ISD::SFENCE";
19651 case X86ISD::LFENCE: return "X86ISD::LFENCE";
19652 case X86ISD::SEG_ALLOCA: return "X86ISD::SEG_ALLOCA";
19653 case X86ISD::SAHF: return "X86ISD::SAHF";
19654 case X86ISD::RDRAND: return "X86ISD::RDRAND";
19655 case X86ISD::RDSEED: return "X86ISD::RDSEED";
19656 case X86ISD::VPMADDUBSW: return "X86ISD::VPMADDUBSW";
19657 case X86ISD::VPMADDWD: return "X86ISD::VPMADDWD";
19658 case X86ISD::FMADD: return "X86ISD::FMADD";
19659 case X86ISD::FMSUB: return "X86ISD::FMSUB";
19660 case X86ISD::FNMADD: return "X86ISD::FNMADD";
19661 case X86ISD::FNMSUB: return "X86ISD::FNMSUB";
19662 case X86ISD::FMADDSUB: return "X86ISD::FMADDSUB";
19663 case X86ISD::FMSUBADD: return "X86ISD::FMSUBADD";
19664 case X86ISD::FMADD_RND: return "X86ISD::FMADD_RND";
19665 case X86ISD::FNMADD_RND: return "X86ISD::FNMADD_RND";
19666 case X86ISD::FMSUB_RND: return "X86ISD::FMSUB_RND";
19667 case X86ISD::FNMSUB_RND: return "X86ISD::FNMSUB_RND";
19668 case X86ISD::FMADDSUB_RND: return "X86ISD::FMADDSUB_RND";
19669 case X86ISD::FMSUBADD_RND: return "X86ISD::FMSUBADD_RND";
19670 case X86ISD::VRNDSCALE: return "X86ISD::VRNDSCALE";
19671 case X86ISD::VREDUCE: return "X86ISD::VREDUCE";
19672 case X86ISD::VGETMANT: return "X86ISD::VGETMANT";
19673 case X86ISD::PCMPESTRI: return "X86ISD::PCMPESTRI";
19674 case X86ISD::PCMPISTRI: return "X86ISD::PCMPISTRI";
19675 case X86ISD::XTEST: return "X86ISD::XTEST";
19676 case X86ISD::COMPRESS: return "X86ISD::COMPRESS";
19677 case X86ISD::EXPAND: return "X86ISD::EXPAND";
19678 case X86ISD::SELECT: return "X86ISD::SELECT";
19679 case X86ISD::ADDSUB: return "X86ISD::ADDSUB";
19680 case X86ISD::RCP28: return "X86ISD::RCP28";
19681 case X86ISD::EXP2: return "X86ISD::EXP2";
19682 case X86ISD::RSQRT28: return "X86ISD::RSQRT28";
19683 case X86ISD::FADD_RND: return "X86ISD::FADD_RND";
19684 case X86ISD::FSUB_RND: return "X86ISD::FSUB_RND";
19685 case X86ISD::FMUL_RND: return "X86ISD::FMUL_RND";
19686 case X86ISD::FDIV_RND: return "X86ISD::FDIV_RND";
19687 case X86ISD::FSQRT_RND: return "X86ISD::FSQRT_RND";
19688 case X86ISD::FGETEXP_RND: return "X86ISD::FGETEXP_RND";
19689 case X86ISD::SCALEF: return "X86ISD::SCALEF";
19690 case X86ISD::ADDS: return "X86ISD::ADDS";
19691 case X86ISD::SUBS: return "X86ISD::SUBS";
19692 case X86ISD::AVG: return "X86ISD::AVG";
19693 case X86ISD::MULHRS: return "X86ISD::MULHRS";
19694 case X86ISD::SINT_TO_FP_RND: return "X86ISD::SINT_TO_FP_RND";
19695 case X86ISD::UINT_TO_FP_RND: return "X86ISD::UINT_TO_FP_RND";
19696 case X86ISD::FP_TO_SINT_RND: return "X86ISD::FP_TO_SINT_RND";
19697 case X86ISD::FP_TO_UINT_RND: return "X86ISD::FP_TO_UINT_RND";
19702 // isLegalAddressingMode - Return true if the addressing mode represented
19703 // by AM is legal for this target, for a load/store of the specified type.
19704 bool X86TargetLowering::isLegalAddressingMode(const DataLayout &DL,
19705 const AddrMode &AM, Type *Ty,
19706 unsigned AS) const {
19707 // X86 supports extremely general addressing modes.
19708 CodeModel::Model M = getTargetMachine().getCodeModel();
19709 Reloc::Model R = getTargetMachine().getRelocationModel();
19711 // X86 allows a sign-extended 32-bit immediate field as a displacement.
19712 if (!X86::isOffsetSuitableForCodeModel(AM.BaseOffs, M, AM.BaseGV != nullptr))
19717 Subtarget->ClassifyGlobalReference(AM.BaseGV, getTargetMachine());
19719 // If a reference to this global requires an extra load, we can't fold it.
19720 if (isGlobalStubReference(GVFlags))
19723 // If BaseGV requires a register for the PIC base, we cannot also have a
19724 // BaseReg specified.
19725 if (AM.HasBaseReg && isGlobalRelativeToPICBase(GVFlags))
19728 // If lower 4G is not available, then we must use rip-relative addressing.
19729 if ((M != CodeModel::Small || R != Reloc::Static) &&
19730 Subtarget->is64Bit() && (AM.BaseOffs || AM.Scale > 1))
19734 switch (AM.Scale) {
19740 // These scales always work.
19745 // These scales are formed with basereg+scalereg. Only accept if there is
19750 default: // Other stuff never works.
19757 bool X86TargetLowering::isVectorShiftByScalarCheap(Type *Ty) const {
19758 unsigned Bits = Ty->getScalarSizeInBits();
19760 // 8-bit shifts are always expensive, but versions with a scalar amount aren't
19761 // particularly cheaper than those without.
19765 // On AVX2 there are new vpsllv[dq] instructions (and other shifts), that make
19766 // variable shifts just as cheap as scalar ones.
19767 if (Subtarget->hasInt256() && (Bits == 32 || Bits == 64))
19770 // Otherwise, it's significantly cheaper to shift by a scalar amount than by a
19771 // fully general vector.
19775 bool X86TargetLowering::isTruncateFree(Type *Ty1, Type *Ty2) const {
19776 if (!Ty1->isIntegerTy() || !Ty2->isIntegerTy())
19778 unsigned NumBits1 = Ty1->getPrimitiveSizeInBits();
19779 unsigned NumBits2 = Ty2->getPrimitiveSizeInBits();
19780 return NumBits1 > NumBits2;
19783 bool X86TargetLowering::allowTruncateForTailCall(Type *Ty1, Type *Ty2) const {
19784 if (!Ty1->isIntegerTy() || !Ty2->isIntegerTy())
19787 if (!isTypeLegal(EVT::getEVT(Ty1)))
19790 assert(Ty1->getPrimitiveSizeInBits() <= 64 && "i128 is probably not a noop");
19792 // Assuming the caller doesn't have a zeroext or signext return parameter,
19793 // truncation all the way down to i1 is valid.
19797 bool X86TargetLowering::isLegalICmpImmediate(int64_t Imm) const {
19798 return isInt<32>(Imm);
19801 bool X86TargetLowering::isLegalAddImmediate(int64_t Imm) const {
19802 // Can also use sub to handle negated immediates.
19803 return isInt<32>(Imm);
19806 bool X86TargetLowering::isTruncateFree(EVT VT1, EVT VT2) const {
19807 if (!VT1.isInteger() || !VT2.isInteger())
19809 unsigned NumBits1 = VT1.getSizeInBits();
19810 unsigned NumBits2 = VT2.getSizeInBits();
19811 return NumBits1 > NumBits2;
19814 bool X86TargetLowering::isZExtFree(Type *Ty1, Type *Ty2) const {
19815 // x86-64 implicitly zero-extends 32-bit results in 64-bit registers.
19816 return Ty1->isIntegerTy(32) && Ty2->isIntegerTy(64) && Subtarget->is64Bit();
19819 bool X86TargetLowering::isZExtFree(EVT VT1, EVT VT2) const {
19820 // x86-64 implicitly zero-extends 32-bit results in 64-bit registers.
19821 return VT1 == MVT::i32 && VT2 == MVT::i64 && Subtarget->is64Bit();
19824 bool X86TargetLowering::isZExtFree(SDValue Val, EVT VT2) const {
19825 EVT VT1 = Val.getValueType();
19826 if (isZExtFree(VT1, VT2))
19829 if (Val.getOpcode() != ISD::LOAD)
19832 if (!VT1.isSimple() || !VT1.isInteger() ||
19833 !VT2.isSimple() || !VT2.isInteger())
19836 switch (VT1.getSimpleVT().SimpleTy) {
19841 // X86 has 8, 16, and 32-bit zero-extending loads.
19848 bool X86TargetLowering::isVectorLoadExtDesirable(SDValue) const { return true; }
19851 X86TargetLowering::isFMAFasterThanFMulAndFAdd(EVT VT) const {
19852 if (!(Subtarget->hasFMA() || Subtarget->hasFMA4() || Subtarget->hasAVX512()))
19855 VT = VT.getScalarType();
19857 if (!VT.isSimple())
19860 switch (VT.getSimpleVT().SimpleTy) {
19871 bool X86TargetLowering::isNarrowingProfitable(EVT VT1, EVT VT2) const {
19872 // i16 instructions are longer (0x66 prefix) and potentially slower.
19873 return !(VT1 == MVT::i32 && VT2 == MVT::i16);
19876 /// isShuffleMaskLegal - Targets can use this to indicate that they only
19877 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
19878 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
19879 /// are assumed to be legal.
19881 X86TargetLowering::isShuffleMaskLegal(const SmallVectorImpl<int> &M,
19883 if (!VT.isSimple())
19886 // Not for i1 vectors
19887 if (VT.getScalarType() == MVT::i1)
19890 // Very little shuffling can be done for 64-bit vectors right now.
19891 if (VT.getSizeInBits() == 64)
19894 // We only care that the types being shuffled are legal. The lowering can
19895 // handle any possible shuffle mask that results.
19896 return isTypeLegal(VT.getSimpleVT());
19900 X86TargetLowering::isVectorClearMaskLegal(const SmallVectorImpl<int> &Mask,
19902 // Just delegate to the generic legality, clear masks aren't special.
19903 return isShuffleMaskLegal(Mask, VT);
19906 //===----------------------------------------------------------------------===//
19907 // X86 Scheduler Hooks
19908 //===----------------------------------------------------------------------===//
19910 /// Utility function to emit xbegin specifying the start of an RTM region.
19911 static MachineBasicBlock *EmitXBegin(MachineInstr *MI, MachineBasicBlock *MBB,
19912 const TargetInstrInfo *TII) {
19913 DebugLoc DL = MI->getDebugLoc();
19915 const BasicBlock *BB = MBB->getBasicBlock();
19916 MachineFunction::iterator I = MBB;
19919 // For the v = xbegin(), we generate
19930 MachineBasicBlock *thisMBB = MBB;
19931 MachineFunction *MF = MBB->getParent();
19932 MachineBasicBlock *mainMBB = MF->CreateMachineBasicBlock(BB);
19933 MachineBasicBlock *sinkMBB = MF->CreateMachineBasicBlock(BB);
19934 MF->insert(I, mainMBB);
19935 MF->insert(I, sinkMBB);
19937 // Transfer the remainder of BB and its successor edges to sinkMBB.
19938 sinkMBB->splice(sinkMBB->begin(), MBB,
19939 std::next(MachineBasicBlock::iterator(MI)), MBB->end());
19940 sinkMBB->transferSuccessorsAndUpdatePHIs(MBB);
19944 // # fallthrough to mainMBB
19945 // # abortion to sinkMBB
19946 BuildMI(thisMBB, DL, TII->get(X86::XBEGIN_4)).addMBB(sinkMBB);
19947 thisMBB->addSuccessor(mainMBB);
19948 thisMBB->addSuccessor(sinkMBB);
19952 BuildMI(mainMBB, DL, TII->get(X86::MOV32ri), X86::EAX).addImm(-1);
19953 mainMBB->addSuccessor(sinkMBB);
19956 // EAX is live into the sinkMBB
19957 sinkMBB->addLiveIn(X86::EAX);
19958 BuildMI(*sinkMBB, sinkMBB->begin(), DL,
19959 TII->get(TargetOpcode::COPY), MI->getOperand(0).getReg())
19962 MI->eraseFromParent();
19966 // FIXME: When we get size specific XMM0 registers, i.e. XMM0_V16I8
19967 // or XMM0_V32I8 in AVX all of this code can be replaced with that
19968 // in the .td file.
19969 static MachineBasicBlock *EmitPCMPSTRM(MachineInstr *MI, MachineBasicBlock *BB,
19970 const TargetInstrInfo *TII) {
19972 switch (MI->getOpcode()) {
19973 default: llvm_unreachable("illegal opcode!");
19974 case X86::PCMPISTRM128REG: Opc = X86::PCMPISTRM128rr; break;
19975 case X86::VPCMPISTRM128REG: Opc = X86::VPCMPISTRM128rr; break;
19976 case X86::PCMPISTRM128MEM: Opc = X86::PCMPISTRM128rm; break;
19977 case X86::VPCMPISTRM128MEM: Opc = X86::VPCMPISTRM128rm; break;
19978 case X86::PCMPESTRM128REG: Opc = X86::PCMPESTRM128rr; break;
19979 case X86::VPCMPESTRM128REG: Opc = X86::VPCMPESTRM128rr; break;
19980 case X86::PCMPESTRM128MEM: Opc = X86::PCMPESTRM128rm; break;
19981 case X86::VPCMPESTRM128MEM: Opc = X86::VPCMPESTRM128rm; break;
19984 DebugLoc dl = MI->getDebugLoc();
19985 MachineInstrBuilder MIB = BuildMI(*BB, MI, dl, TII->get(Opc));
19987 unsigned NumArgs = MI->getNumOperands();
19988 for (unsigned i = 1; i < NumArgs; ++i) {
19989 MachineOperand &Op = MI->getOperand(i);
19990 if (!(Op.isReg() && Op.isImplicit()))
19991 MIB.addOperand(Op);
19993 if (MI->hasOneMemOperand())
19994 MIB->setMemRefs(MI->memoperands_begin(), MI->memoperands_end());
19996 BuildMI(*BB, MI, dl,
19997 TII->get(TargetOpcode::COPY), MI->getOperand(0).getReg())
19998 .addReg(X86::XMM0);
20000 MI->eraseFromParent();
20004 // FIXME: Custom handling because TableGen doesn't support multiple implicit
20005 // defs in an instruction pattern
20006 static MachineBasicBlock *EmitPCMPSTRI(MachineInstr *MI, MachineBasicBlock *BB,
20007 const TargetInstrInfo *TII) {
20009 switch (MI->getOpcode()) {
20010 default: llvm_unreachable("illegal opcode!");
20011 case X86::PCMPISTRIREG: Opc = X86::PCMPISTRIrr; break;
20012 case X86::VPCMPISTRIREG: Opc = X86::VPCMPISTRIrr; break;
20013 case X86::PCMPISTRIMEM: Opc = X86::PCMPISTRIrm; break;
20014 case X86::VPCMPISTRIMEM: Opc = X86::VPCMPISTRIrm; break;
20015 case X86::PCMPESTRIREG: Opc = X86::PCMPESTRIrr; break;
20016 case X86::VPCMPESTRIREG: Opc = X86::VPCMPESTRIrr; break;
20017 case X86::PCMPESTRIMEM: Opc = X86::PCMPESTRIrm; break;
20018 case X86::VPCMPESTRIMEM: Opc = X86::VPCMPESTRIrm; break;
20021 DebugLoc dl = MI->getDebugLoc();
20022 MachineInstrBuilder MIB = BuildMI(*BB, MI, dl, TII->get(Opc));
20024 unsigned NumArgs = MI->getNumOperands(); // remove the results
20025 for (unsigned i = 1; i < NumArgs; ++i) {
20026 MachineOperand &Op = MI->getOperand(i);
20027 if (!(Op.isReg() && Op.isImplicit()))
20028 MIB.addOperand(Op);
20030 if (MI->hasOneMemOperand())
20031 MIB->setMemRefs(MI->memoperands_begin(), MI->memoperands_end());
20033 BuildMI(*BB, MI, dl,
20034 TII->get(TargetOpcode::COPY), MI->getOperand(0).getReg())
20037 MI->eraseFromParent();
20041 static MachineBasicBlock *EmitMonitor(MachineInstr *MI, MachineBasicBlock *BB,
20042 const X86Subtarget *Subtarget) {
20043 DebugLoc dl = MI->getDebugLoc();
20044 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
20045 // Address into RAX/EAX, other two args into ECX, EDX.
20046 unsigned MemOpc = Subtarget->is64Bit() ? X86::LEA64r : X86::LEA32r;
20047 unsigned MemReg = Subtarget->is64Bit() ? X86::RAX : X86::EAX;
20048 MachineInstrBuilder MIB = BuildMI(*BB, MI, dl, TII->get(MemOpc), MemReg);
20049 for (int i = 0; i < X86::AddrNumOperands; ++i)
20050 MIB.addOperand(MI->getOperand(i));
20052 unsigned ValOps = X86::AddrNumOperands;
20053 BuildMI(*BB, MI, dl, TII->get(TargetOpcode::COPY), X86::ECX)
20054 .addReg(MI->getOperand(ValOps).getReg());
20055 BuildMI(*BB, MI, dl, TII->get(TargetOpcode::COPY), X86::EDX)
20056 .addReg(MI->getOperand(ValOps+1).getReg());
20058 // The instruction doesn't actually take any operands though.
20059 BuildMI(*BB, MI, dl, TII->get(X86::MONITORrrr));
20061 MI->eraseFromParent(); // The pseudo is gone now.
20065 MachineBasicBlock *
20066 X86TargetLowering::EmitVAARG64WithCustomInserter(MachineInstr *MI,
20067 MachineBasicBlock *MBB) const {
20068 // Emit va_arg instruction on X86-64.
20070 // Operands to this pseudo-instruction:
20071 // 0 ) Output : destination address (reg)
20072 // 1-5) Input : va_list address (addr, i64mem)
20073 // 6 ) ArgSize : Size (in bytes) of vararg type
20074 // 7 ) ArgMode : 0=overflow only, 1=use gp_offset, 2=use fp_offset
20075 // 8 ) Align : Alignment of type
20076 // 9 ) EFLAGS (implicit-def)
20078 assert(MI->getNumOperands() == 10 && "VAARG_64 should have 10 operands!");
20079 static_assert(X86::AddrNumOperands == 5,
20080 "VAARG_64 assumes 5 address operands");
20082 unsigned DestReg = MI->getOperand(0).getReg();
20083 MachineOperand &Base = MI->getOperand(1);
20084 MachineOperand &Scale = MI->getOperand(2);
20085 MachineOperand &Index = MI->getOperand(3);
20086 MachineOperand &Disp = MI->getOperand(4);
20087 MachineOperand &Segment = MI->getOperand(5);
20088 unsigned ArgSize = MI->getOperand(6).getImm();
20089 unsigned ArgMode = MI->getOperand(7).getImm();
20090 unsigned Align = MI->getOperand(8).getImm();
20092 // Memory Reference
20093 assert(MI->hasOneMemOperand() && "Expected VAARG_64 to have one memoperand");
20094 MachineInstr::mmo_iterator MMOBegin = MI->memoperands_begin();
20095 MachineInstr::mmo_iterator MMOEnd = MI->memoperands_end();
20097 // Machine Information
20098 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
20099 MachineRegisterInfo &MRI = MBB->getParent()->getRegInfo();
20100 const TargetRegisterClass *AddrRegClass = getRegClassFor(MVT::i64);
20101 const TargetRegisterClass *OffsetRegClass = getRegClassFor(MVT::i32);
20102 DebugLoc DL = MI->getDebugLoc();
20104 // struct va_list {
20107 // i64 overflow_area (address)
20108 // i64 reg_save_area (address)
20110 // sizeof(va_list) = 24
20111 // alignment(va_list) = 8
20113 unsigned TotalNumIntRegs = 6;
20114 unsigned TotalNumXMMRegs = 8;
20115 bool UseGPOffset = (ArgMode == 1);
20116 bool UseFPOffset = (ArgMode == 2);
20117 unsigned MaxOffset = TotalNumIntRegs * 8 +
20118 (UseFPOffset ? TotalNumXMMRegs * 16 : 0);
20120 /* Align ArgSize to a multiple of 8 */
20121 unsigned ArgSizeA8 = (ArgSize + 7) & ~7;
20122 bool NeedsAlign = (Align > 8);
20124 MachineBasicBlock *thisMBB = MBB;
20125 MachineBasicBlock *overflowMBB;
20126 MachineBasicBlock *offsetMBB;
20127 MachineBasicBlock *endMBB;
20129 unsigned OffsetDestReg = 0; // Argument address computed by offsetMBB
20130 unsigned OverflowDestReg = 0; // Argument address computed by overflowMBB
20131 unsigned OffsetReg = 0;
20133 if (!UseGPOffset && !UseFPOffset) {
20134 // If we only pull from the overflow region, we don't create a branch.
20135 // We don't need to alter control flow.
20136 OffsetDestReg = 0; // unused
20137 OverflowDestReg = DestReg;
20139 offsetMBB = nullptr;
20140 overflowMBB = thisMBB;
20143 // First emit code to check if gp_offset (or fp_offset) is below the bound.
20144 // If so, pull the argument from reg_save_area. (branch to offsetMBB)
20145 // If not, pull from overflow_area. (branch to overflowMBB)
20150 // offsetMBB overflowMBB
20155 // Registers for the PHI in endMBB
20156 OffsetDestReg = MRI.createVirtualRegister(AddrRegClass);
20157 OverflowDestReg = MRI.createVirtualRegister(AddrRegClass);
20159 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
20160 MachineFunction *MF = MBB->getParent();
20161 overflowMBB = MF->CreateMachineBasicBlock(LLVM_BB);
20162 offsetMBB = MF->CreateMachineBasicBlock(LLVM_BB);
20163 endMBB = MF->CreateMachineBasicBlock(LLVM_BB);
20165 MachineFunction::iterator MBBIter = MBB;
20168 // Insert the new basic blocks
20169 MF->insert(MBBIter, offsetMBB);
20170 MF->insert(MBBIter, overflowMBB);
20171 MF->insert(MBBIter, endMBB);
20173 // Transfer the remainder of MBB and its successor edges to endMBB.
20174 endMBB->splice(endMBB->begin(), thisMBB,
20175 std::next(MachineBasicBlock::iterator(MI)), thisMBB->end());
20176 endMBB->transferSuccessorsAndUpdatePHIs(thisMBB);
20178 // Make offsetMBB and overflowMBB successors of thisMBB
20179 thisMBB->addSuccessor(offsetMBB);
20180 thisMBB->addSuccessor(overflowMBB);
20182 // endMBB is a successor of both offsetMBB and overflowMBB
20183 offsetMBB->addSuccessor(endMBB);
20184 overflowMBB->addSuccessor(endMBB);
20186 // Load the offset value into a register
20187 OffsetReg = MRI.createVirtualRegister(OffsetRegClass);
20188 BuildMI(thisMBB, DL, TII->get(X86::MOV32rm), OffsetReg)
20192 .addDisp(Disp, UseFPOffset ? 4 : 0)
20193 .addOperand(Segment)
20194 .setMemRefs(MMOBegin, MMOEnd);
20196 // Check if there is enough room left to pull this argument.
20197 BuildMI(thisMBB, DL, TII->get(X86::CMP32ri))
20199 .addImm(MaxOffset + 8 - ArgSizeA8);
20201 // Branch to "overflowMBB" if offset >= max
20202 // Fall through to "offsetMBB" otherwise
20203 BuildMI(thisMBB, DL, TII->get(X86::GetCondBranchFromCond(X86::COND_AE)))
20204 .addMBB(overflowMBB);
20207 // In offsetMBB, emit code to use the reg_save_area.
20209 assert(OffsetReg != 0);
20211 // Read the reg_save_area address.
20212 unsigned RegSaveReg = MRI.createVirtualRegister(AddrRegClass);
20213 BuildMI(offsetMBB, DL, TII->get(X86::MOV64rm), RegSaveReg)
20218 .addOperand(Segment)
20219 .setMemRefs(MMOBegin, MMOEnd);
20221 // Zero-extend the offset
20222 unsigned OffsetReg64 = MRI.createVirtualRegister(AddrRegClass);
20223 BuildMI(offsetMBB, DL, TII->get(X86::SUBREG_TO_REG), OffsetReg64)
20226 .addImm(X86::sub_32bit);
20228 // Add the offset to the reg_save_area to get the final address.
20229 BuildMI(offsetMBB, DL, TII->get(X86::ADD64rr), OffsetDestReg)
20230 .addReg(OffsetReg64)
20231 .addReg(RegSaveReg);
20233 // Compute the offset for the next argument
20234 unsigned NextOffsetReg = MRI.createVirtualRegister(OffsetRegClass);
20235 BuildMI(offsetMBB, DL, TII->get(X86::ADD32ri), NextOffsetReg)
20237 .addImm(UseFPOffset ? 16 : 8);
20239 // Store it back into the va_list.
20240 BuildMI(offsetMBB, DL, TII->get(X86::MOV32mr))
20244 .addDisp(Disp, UseFPOffset ? 4 : 0)
20245 .addOperand(Segment)
20246 .addReg(NextOffsetReg)
20247 .setMemRefs(MMOBegin, MMOEnd);
20250 BuildMI(offsetMBB, DL, TII->get(X86::JMP_1))
20255 // Emit code to use overflow area
20258 // Load the overflow_area address into a register.
20259 unsigned OverflowAddrReg = MRI.createVirtualRegister(AddrRegClass);
20260 BuildMI(overflowMBB, DL, TII->get(X86::MOV64rm), OverflowAddrReg)
20265 .addOperand(Segment)
20266 .setMemRefs(MMOBegin, MMOEnd);
20268 // If we need to align it, do so. Otherwise, just copy the address
20269 // to OverflowDestReg.
20271 // Align the overflow address
20272 assert((Align & (Align-1)) == 0 && "Alignment must be a power of 2");
20273 unsigned TmpReg = MRI.createVirtualRegister(AddrRegClass);
20275 // aligned_addr = (addr + (align-1)) & ~(align-1)
20276 BuildMI(overflowMBB, DL, TII->get(X86::ADD64ri32), TmpReg)
20277 .addReg(OverflowAddrReg)
20280 BuildMI(overflowMBB, DL, TII->get(X86::AND64ri32), OverflowDestReg)
20282 .addImm(~(uint64_t)(Align-1));
20284 BuildMI(overflowMBB, DL, TII->get(TargetOpcode::COPY), OverflowDestReg)
20285 .addReg(OverflowAddrReg);
20288 // Compute the next overflow address after this argument.
20289 // (the overflow address should be kept 8-byte aligned)
20290 unsigned NextAddrReg = MRI.createVirtualRegister(AddrRegClass);
20291 BuildMI(overflowMBB, DL, TII->get(X86::ADD64ri32), NextAddrReg)
20292 .addReg(OverflowDestReg)
20293 .addImm(ArgSizeA8);
20295 // Store the new overflow address.
20296 BuildMI(overflowMBB, DL, TII->get(X86::MOV64mr))
20301 .addOperand(Segment)
20302 .addReg(NextAddrReg)
20303 .setMemRefs(MMOBegin, MMOEnd);
20305 // If we branched, emit the PHI to the front of endMBB.
20307 BuildMI(*endMBB, endMBB->begin(), DL,
20308 TII->get(X86::PHI), DestReg)
20309 .addReg(OffsetDestReg).addMBB(offsetMBB)
20310 .addReg(OverflowDestReg).addMBB(overflowMBB);
20313 // Erase the pseudo instruction
20314 MI->eraseFromParent();
20319 MachineBasicBlock *
20320 X86TargetLowering::EmitVAStartSaveXMMRegsWithCustomInserter(
20322 MachineBasicBlock *MBB) const {
20323 // Emit code to save XMM registers to the stack. The ABI says that the
20324 // number of registers to save is given in %al, so it's theoretically
20325 // possible to do an indirect jump trick to avoid saving all of them,
20326 // however this code takes a simpler approach and just executes all
20327 // of the stores if %al is non-zero. It's less code, and it's probably
20328 // easier on the hardware branch predictor, and stores aren't all that
20329 // expensive anyway.
20331 // Create the new basic blocks. One block contains all the XMM stores,
20332 // and one block is the final destination regardless of whether any
20333 // stores were performed.
20334 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
20335 MachineFunction *F = MBB->getParent();
20336 MachineFunction::iterator MBBIter = MBB;
20338 MachineBasicBlock *XMMSaveMBB = F->CreateMachineBasicBlock(LLVM_BB);
20339 MachineBasicBlock *EndMBB = F->CreateMachineBasicBlock(LLVM_BB);
20340 F->insert(MBBIter, XMMSaveMBB);
20341 F->insert(MBBIter, EndMBB);
20343 // Transfer the remainder of MBB and its successor edges to EndMBB.
20344 EndMBB->splice(EndMBB->begin(), MBB,
20345 std::next(MachineBasicBlock::iterator(MI)), MBB->end());
20346 EndMBB->transferSuccessorsAndUpdatePHIs(MBB);
20348 // The original block will now fall through to the XMM save block.
20349 MBB->addSuccessor(XMMSaveMBB);
20350 // The XMMSaveMBB will fall through to the end block.
20351 XMMSaveMBB->addSuccessor(EndMBB);
20353 // Now add the instructions.
20354 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
20355 DebugLoc DL = MI->getDebugLoc();
20357 unsigned CountReg = MI->getOperand(0).getReg();
20358 int64_t RegSaveFrameIndex = MI->getOperand(1).getImm();
20359 int64_t VarArgsFPOffset = MI->getOperand(2).getImm();
20361 if (!Subtarget->isCallingConvWin64(F->getFunction()->getCallingConv())) {
20362 // If %al is 0, branch around the XMM save block.
20363 BuildMI(MBB, DL, TII->get(X86::TEST8rr)).addReg(CountReg).addReg(CountReg);
20364 BuildMI(MBB, DL, TII->get(X86::JE_1)).addMBB(EndMBB);
20365 MBB->addSuccessor(EndMBB);
20368 // Make sure the last operand is EFLAGS, which gets clobbered by the branch
20369 // that was just emitted, but clearly shouldn't be "saved".
20370 assert((MI->getNumOperands() <= 3 ||
20371 !MI->getOperand(MI->getNumOperands() - 1).isReg() ||
20372 MI->getOperand(MI->getNumOperands() - 1).getReg() == X86::EFLAGS)
20373 && "Expected last argument to be EFLAGS");
20374 unsigned MOVOpc = Subtarget->hasFp256() ? X86::VMOVAPSmr : X86::MOVAPSmr;
20375 // In the XMM save block, save all the XMM argument registers.
20376 for (int i = 3, e = MI->getNumOperands() - 1; i != e; ++i) {
20377 int64_t Offset = (i - 3) * 16 + VarArgsFPOffset;
20378 MachineMemOperand *MMO = F->getMachineMemOperand(
20379 MachinePointerInfo::getFixedStack(*F, RegSaveFrameIndex, Offset),
20380 MachineMemOperand::MOStore,
20381 /*Size=*/16, /*Align=*/16);
20382 BuildMI(XMMSaveMBB, DL, TII->get(MOVOpc))
20383 .addFrameIndex(RegSaveFrameIndex)
20384 .addImm(/*Scale=*/1)
20385 .addReg(/*IndexReg=*/0)
20386 .addImm(/*Disp=*/Offset)
20387 .addReg(/*Segment=*/0)
20388 .addReg(MI->getOperand(i).getReg())
20389 .addMemOperand(MMO);
20392 MI->eraseFromParent(); // The pseudo instruction is gone now.
20397 // The EFLAGS operand of SelectItr might be missing a kill marker
20398 // because there were multiple uses of EFLAGS, and ISel didn't know
20399 // which to mark. Figure out whether SelectItr should have had a
20400 // kill marker, and set it if it should. Returns the correct kill
20402 static bool checkAndUpdateEFLAGSKill(MachineBasicBlock::iterator SelectItr,
20403 MachineBasicBlock* BB,
20404 const TargetRegisterInfo* TRI) {
20405 // Scan forward through BB for a use/def of EFLAGS.
20406 MachineBasicBlock::iterator miI(std::next(SelectItr));
20407 for (MachineBasicBlock::iterator miE = BB->end(); miI != miE; ++miI) {
20408 const MachineInstr& mi = *miI;
20409 if (mi.readsRegister(X86::EFLAGS))
20411 if (mi.definesRegister(X86::EFLAGS))
20412 break; // Should have kill-flag - update below.
20415 // If we hit the end of the block, check whether EFLAGS is live into a
20417 if (miI == BB->end()) {
20418 for (MachineBasicBlock::succ_iterator sItr = BB->succ_begin(),
20419 sEnd = BB->succ_end();
20420 sItr != sEnd; ++sItr) {
20421 MachineBasicBlock* succ = *sItr;
20422 if (succ->isLiveIn(X86::EFLAGS))
20427 // We found a def, or hit the end of the basic block and EFLAGS wasn't live
20428 // out. SelectMI should have a kill flag on EFLAGS.
20429 SelectItr->addRegisterKilled(X86::EFLAGS, TRI);
20433 // Return true if it is OK for this CMOV pseudo-opcode to be cascaded
20434 // together with other CMOV pseudo-opcodes into a single basic-block with
20435 // conditional jump around it.
20436 static bool isCMOVPseudo(MachineInstr *MI) {
20437 switch (MI->getOpcode()) {
20438 case X86::CMOV_FR32:
20439 case X86::CMOV_FR64:
20440 case X86::CMOV_GR8:
20441 case X86::CMOV_GR16:
20442 case X86::CMOV_GR32:
20443 case X86::CMOV_RFP32:
20444 case X86::CMOV_RFP64:
20445 case X86::CMOV_RFP80:
20446 case X86::CMOV_V2F64:
20447 case X86::CMOV_V2I64:
20448 case X86::CMOV_V4F32:
20449 case X86::CMOV_V4F64:
20450 case X86::CMOV_V4I64:
20451 case X86::CMOV_V16F32:
20452 case X86::CMOV_V8F32:
20453 case X86::CMOV_V8F64:
20454 case X86::CMOV_V8I64:
20455 case X86::CMOV_V8I1:
20456 case X86::CMOV_V16I1:
20457 case X86::CMOV_V32I1:
20458 case X86::CMOV_V64I1:
20466 MachineBasicBlock *
20467 X86TargetLowering::EmitLoweredSelect(MachineInstr *MI,
20468 MachineBasicBlock *BB) const {
20469 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
20470 DebugLoc DL = MI->getDebugLoc();
20472 // To "insert" a SELECT_CC instruction, we actually have to insert the
20473 // diamond control-flow pattern. The incoming instruction knows the
20474 // destination vreg to set, the condition code register to branch on, the
20475 // true/false values to select between, and a branch opcode to use.
20476 const BasicBlock *LLVM_BB = BB->getBasicBlock();
20477 MachineFunction::iterator It = BB;
20483 // cmpTY ccX, r1, r2
20485 // fallthrough --> copy0MBB
20486 MachineBasicBlock *thisMBB = BB;
20487 MachineFunction *F = BB->getParent();
20489 // This code lowers all pseudo-CMOV instructions. Generally it lowers these
20490 // as described above, by inserting a BB, and then making a PHI at the join
20491 // point to select the true and false operands of the CMOV in the PHI.
20493 // The code also handles two different cases of multiple CMOV opcodes
20497 // In this case, there are multiple CMOVs in a row, all which are based on
20498 // the same condition setting (or the exact opposite condition setting).
20499 // In this case we can lower all the CMOVs using a single inserted BB, and
20500 // then make a number of PHIs at the join point to model the CMOVs. The only
20501 // trickiness here, is that in a case like:
20503 // t2 = CMOV cond1 t1, f1
20504 // t3 = CMOV cond1 t2, f2
20506 // when rewriting this into PHIs, we have to perform some renaming on the
20507 // temps since you cannot have a PHI operand refer to a PHI result earlier
20508 // in the same block. The "simple" but wrong lowering would be:
20510 // t2 = PHI t1(BB1), f1(BB2)
20511 // t3 = PHI t2(BB1), f2(BB2)
20513 // but clearly t2 is not defined in BB1, so that is incorrect. The proper
20514 // renaming is to note that on the path through BB1, t2 is really just a
20515 // copy of t1, and do that renaming, properly generating:
20517 // t2 = PHI t1(BB1), f1(BB2)
20518 // t3 = PHI t1(BB1), f2(BB2)
20520 // Case 2, we lower cascaded CMOVs such as
20522 // (CMOV (CMOV F, T, cc1), T, cc2)
20524 // to two successives branches. For that, we look for another CMOV as the
20525 // following instruction.
20527 // Without this, we would add a PHI between the two jumps, which ends up
20528 // creating a few copies all around. For instance, for
20530 // (sitofp (zext (fcmp une)))
20532 // we would generate:
20534 // ucomiss %xmm1, %xmm0
20535 // movss <1.0f>, %xmm0
20536 // movaps %xmm0, %xmm1
20538 // xorps %xmm1, %xmm1
20541 // movaps %xmm1, %xmm0
20545 // because this custom-inserter would have generated:
20557 // A: X = ...; Y = ...
20559 // C: Z = PHI [X, A], [Y, B]
20561 // E: PHI [X, C], [Z, D]
20563 // If we lower both CMOVs in a single step, we can instead generate:
20575 // A: X = ...; Y = ...
20577 // E: PHI [X, A], [X, C], [Y, D]
20579 // Which, in our sitofp/fcmp example, gives us something like:
20581 // ucomiss %xmm1, %xmm0
20582 // movss <1.0f>, %xmm0
20585 // xorps %xmm0, %xmm0
20589 MachineInstr *CascadedCMOV = nullptr;
20590 MachineInstr *LastCMOV = MI;
20591 X86::CondCode CC = X86::CondCode(MI->getOperand(3).getImm());
20592 X86::CondCode OppCC = X86::GetOppositeBranchCondition(CC);
20593 MachineBasicBlock::iterator NextMIIt =
20594 std::next(MachineBasicBlock::iterator(MI));
20596 // Check for case 1, where there are multiple CMOVs with the same condition
20597 // first. Of the two cases of multiple CMOV lowerings, case 1 reduces the
20598 // number of jumps the most.
20600 if (isCMOVPseudo(MI)) {
20601 // See if we have a string of CMOVS with the same condition.
20602 while (NextMIIt != BB->end() &&
20603 isCMOVPseudo(NextMIIt) &&
20604 (NextMIIt->getOperand(3).getImm() == CC ||
20605 NextMIIt->getOperand(3).getImm() == OppCC)) {
20606 LastCMOV = &*NextMIIt;
20611 // This checks for case 2, but only do this if we didn't already find
20612 // case 1, as indicated by LastCMOV == MI.
20613 if (LastCMOV == MI &&
20614 NextMIIt != BB->end() && NextMIIt->getOpcode() == MI->getOpcode() &&
20615 NextMIIt->getOperand(2).getReg() == MI->getOperand(2).getReg() &&
20616 NextMIIt->getOperand(1).getReg() == MI->getOperand(0).getReg()) {
20617 CascadedCMOV = &*NextMIIt;
20620 MachineBasicBlock *jcc1MBB = nullptr;
20622 // If we have a cascaded CMOV, we lower it to two successive branches to
20623 // the same block. EFLAGS is used by both, so mark it as live in the second.
20624 if (CascadedCMOV) {
20625 jcc1MBB = F->CreateMachineBasicBlock(LLVM_BB);
20626 F->insert(It, jcc1MBB);
20627 jcc1MBB->addLiveIn(X86::EFLAGS);
20630 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
20631 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
20632 F->insert(It, copy0MBB);
20633 F->insert(It, sinkMBB);
20635 // If the EFLAGS register isn't dead in the terminator, then claim that it's
20636 // live into the sink and copy blocks.
20637 const TargetRegisterInfo *TRI = Subtarget->getRegisterInfo();
20639 MachineInstr *LastEFLAGSUser = CascadedCMOV ? CascadedCMOV : LastCMOV;
20640 if (!LastEFLAGSUser->killsRegister(X86::EFLAGS) &&
20641 !checkAndUpdateEFLAGSKill(LastEFLAGSUser, BB, TRI)) {
20642 copy0MBB->addLiveIn(X86::EFLAGS);
20643 sinkMBB->addLiveIn(X86::EFLAGS);
20646 // Transfer the remainder of BB and its successor edges to sinkMBB.
20647 sinkMBB->splice(sinkMBB->begin(), BB,
20648 std::next(MachineBasicBlock::iterator(LastCMOV)), BB->end());
20649 sinkMBB->transferSuccessorsAndUpdatePHIs(BB);
20651 // Add the true and fallthrough blocks as its successors.
20652 if (CascadedCMOV) {
20653 // The fallthrough block may be jcc1MBB, if we have a cascaded CMOV.
20654 BB->addSuccessor(jcc1MBB);
20656 // In that case, jcc1MBB will itself fallthrough the copy0MBB, and
20657 // jump to the sinkMBB.
20658 jcc1MBB->addSuccessor(copy0MBB);
20659 jcc1MBB->addSuccessor(sinkMBB);
20661 BB->addSuccessor(copy0MBB);
20664 // The true block target of the first (or only) branch is always sinkMBB.
20665 BB->addSuccessor(sinkMBB);
20667 // Create the conditional branch instruction.
20668 unsigned Opc = X86::GetCondBranchFromCond(CC);
20669 BuildMI(BB, DL, TII->get(Opc)).addMBB(sinkMBB);
20671 if (CascadedCMOV) {
20672 unsigned Opc2 = X86::GetCondBranchFromCond(
20673 (X86::CondCode)CascadedCMOV->getOperand(3).getImm());
20674 BuildMI(jcc1MBB, DL, TII->get(Opc2)).addMBB(sinkMBB);
20678 // %FalseValue = ...
20679 // # fallthrough to sinkMBB
20680 copy0MBB->addSuccessor(sinkMBB);
20683 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
20685 MachineBasicBlock::iterator MIItBegin = MachineBasicBlock::iterator(MI);
20686 MachineBasicBlock::iterator MIItEnd =
20687 std::next(MachineBasicBlock::iterator(LastCMOV));
20688 MachineBasicBlock::iterator SinkInsertionPoint = sinkMBB->begin();
20689 DenseMap<unsigned, std::pair<unsigned, unsigned>> RegRewriteTable;
20690 MachineInstrBuilder MIB;
20692 // As we are creating the PHIs, we have to be careful if there is more than
20693 // one. Later CMOVs may reference the results of earlier CMOVs, but later
20694 // PHIs have to reference the individual true/false inputs from earlier PHIs.
20695 // That also means that PHI construction must work forward from earlier to
20696 // later, and that the code must maintain a mapping from earlier PHI's
20697 // destination registers, and the registers that went into the PHI.
20699 for (MachineBasicBlock::iterator MIIt = MIItBegin; MIIt != MIItEnd; ++MIIt) {
20700 unsigned DestReg = MIIt->getOperand(0).getReg();
20701 unsigned Op1Reg = MIIt->getOperand(1).getReg();
20702 unsigned Op2Reg = MIIt->getOperand(2).getReg();
20704 // If this CMOV we are generating is the opposite condition from
20705 // the jump we generated, then we have to swap the operands for the
20706 // PHI that is going to be generated.
20707 if (MIIt->getOperand(3).getImm() == OppCC)
20708 std::swap(Op1Reg, Op2Reg);
20710 if (RegRewriteTable.find(Op1Reg) != RegRewriteTable.end())
20711 Op1Reg = RegRewriteTable[Op1Reg].first;
20713 if (RegRewriteTable.find(Op2Reg) != RegRewriteTable.end())
20714 Op2Reg = RegRewriteTable[Op2Reg].second;
20716 MIB = BuildMI(*sinkMBB, SinkInsertionPoint, DL,
20717 TII->get(X86::PHI), DestReg)
20718 .addReg(Op1Reg).addMBB(copy0MBB)
20719 .addReg(Op2Reg).addMBB(thisMBB);
20721 // Add this PHI to the rewrite table.
20722 RegRewriteTable[DestReg] = std::make_pair(Op1Reg, Op2Reg);
20725 // If we have a cascaded CMOV, the second Jcc provides the same incoming
20726 // value as the first Jcc (the True operand of the SELECT_CC/CMOV nodes).
20727 if (CascadedCMOV) {
20728 MIB.addReg(MI->getOperand(2).getReg()).addMBB(jcc1MBB);
20729 // Copy the PHI result to the register defined by the second CMOV.
20730 BuildMI(*sinkMBB, std::next(MachineBasicBlock::iterator(MIB.getInstr())),
20731 DL, TII->get(TargetOpcode::COPY),
20732 CascadedCMOV->getOperand(0).getReg())
20733 .addReg(MI->getOperand(0).getReg());
20734 CascadedCMOV->eraseFromParent();
20737 // Now remove the CMOV(s).
20738 for (MachineBasicBlock::iterator MIIt = MIItBegin; MIIt != MIItEnd; )
20739 (MIIt++)->eraseFromParent();
20744 MachineBasicBlock *
20745 X86TargetLowering::EmitLoweredAtomicFP(MachineInstr *MI,
20746 MachineBasicBlock *BB) const {
20747 // Combine the following atomic floating-point modification pattern:
20748 // a.store(reg OP a.load(acquire), release)
20749 // Transform them into:
20750 // OPss (%gpr), %xmm
20751 // movss %xmm, (%gpr)
20752 // Or sd equivalent for 64-bit operations.
20754 switch (MI->getOpcode()) {
20755 default: llvm_unreachable("unexpected instr type for EmitLoweredAtomicFP");
20756 case X86::RELEASE_FADD32mr: MOp = X86::MOVSSmr; FOp = X86::ADDSSrm; break;
20757 case X86::RELEASE_FADD64mr: MOp = X86::MOVSDmr; FOp = X86::ADDSDrm; break;
20759 const X86InstrInfo *TII = Subtarget->getInstrInfo();
20760 DebugLoc DL = MI->getDebugLoc();
20761 MachineRegisterInfo &MRI = BB->getParent()->getRegInfo();
20762 unsigned MSrc = MI->getOperand(0).getReg();
20763 unsigned VSrc = MI->getOperand(5).getReg();
20764 MachineInstrBuilder MIM = BuildMI(*BB, MI, DL, TII->get(MOp))
20765 .addReg(/*Base=*/MSrc)
20766 .addImm(/*Scale=*/1)
20767 .addReg(/*Index=*/0)
20770 MachineInstr *MIO = BuildMI(*BB, (MachineInstr *)MIM, DL, TII->get(FOp),
20771 MRI.createVirtualRegister(MRI.getRegClass(VSrc)))
20773 .addReg(/*Base=*/MSrc)
20774 .addImm(/*Scale=*/1)
20775 .addReg(/*Index=*/0)
20776 .addImm(/*Disp=*/0)
20777 .addReg(/*Segment=*/0);
20778 MIM.addReg(MIO->getOperand(0).getReg(), RegState::Kill);
20779 MI->eraseFromParent(); // The pseudo instruction is gone now.
20783 MachineBasicBlock *
20784 X86TargetLowering::EmitLoweredSegAlloca(MachineInstr *MI,
20785 MachineBasicBlock *BB) const {
20786 MachineFunction *MF = BB->getParent();
20787 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
20788 DebugLoc DL = MI->getDebugLoc();
20789 const BasicBlock *LLVM_BB = BB->getBasicBlock();
20791 assert(MF->shouldSplitStack());
20793 const bool Is64Bit = Subtarget->is64Bit();
20794 const bool IsLP64 = Subtarget->isTarget64BitLP64();
20796 const unsigned TlsReg = Is64Bit ? X86::FS : X86::GS;
20797 const unsigned TlsOffset = IsLP64 ? 0x70 : Is64Bit ? 0x40 : 0x30;
20800 // ... [Till the alloca]
20801 // If stacklet is not large enough, jump to mallocMBB
20804 // Allocate by subtracting from RSP
20805 // Jump to continueMBB
20808 // Allocate by call to runtime
20812 // [rest of original BB]
20815 MachineBasicBlock *mallocMBB = MF->CreateMachineBasicBlock(LLVM_BB);
20816 MachineBasicBlock *bumpMBB = MF->CreateMachineBasicBlock(LLVM_BB);
20817 MachineBasicBlock *continueMBB = MF->CreateMachineBasicBlock(LLVM_BB);
20819 MachineRegisterInfo &MRI = MF->getRegInfo();
20820 const TargetRegisterClass *AddrRegClass =
20821 getRegClassFor(getPointerTy(MF->getDataLayout()));
20823 unsigned mallocPtrVReg = MRI.createVirtualRegister(AddrRegClass),
20824 bumpSPPtrVReg = MRI.createVirtualRegister(AddrRegClass),
20825 tmpSPVReg = MRI.createVirtualRegister(AddrRegClass),
20826 SPLimitVReg = MRI.createVirtualRegister(AddrRegClass),
20827 sizeVReg = MI->getOperand(1).getReg(),
20828 physSPReg = IsLP64 || Subtarget->isTargetNaCl64() ? X86::RSP : X86::ESP;
20830 MachineFunction::iterator MBBIter = BB;
20833 MF->insert(MBBIter, bumpMBB);
20834 MF->insert(MBBIter, mallocMBB);
20835 MF->insert(MBBIter, continueMBB);
20837 continueMBB->splice(continueMBB->begin(), BB,
20838 std::next(MachineBasicBlock::iterator(MI)), BB->end());
20839 continueMBB->transferSuccessorsAndUpdatePHIs(BB);
20841 // Add code to the main basic block to check if the stack limit has been hit,
20842 // and if so, jump to mallocMBB otherwise to bumpMBB.
20843 BuildMI(BB, DL, TII->get(TargetOpcode::COPY), tmpSPVReg).addReg(physSPReg);
20844 BuildMI(BB, DL, TII->get(IsLP64 ? X86::SUB64rr:X86::SUB32rr), SPLimitVReg)
20845 .addReg(tmpSPVReg).addReg(sizeVReg);
20846 BuildMI(BB, DL, TII->get(IsLP64 ? X86::CMP64mr:X86::CMP32mr))
20847 .addReg(0).addImm(1).addReg(0).addImm(TlsOffset).addReg(TlsReg)
20848 .addReg(SPLimitVReg);
20849 BuildMI(BB, DL, TII->get(X86::JG_1)).addMBB(mallocMBB);
20851 // bumpMBB simply decreases the stack pointer, since we know the current
20852 // stacklet has enough space.
20853 BuildMI(bumpMBB, DL, TII->get(TargetOpcode::COPY), physSPReg)
20854 .addReg(SPLimitVReg);
20855 BuildMI(bumpMBB, DL, TII->get(TargetOpcode::COPY), bumpSPPtrVReg)
20856 .addReg(SPLimitVReg);
20857 BuildMI(bumpMBB, DL, TII->get(X86::JMP_1)).addMBB(continueMBB);
20859 // Calls into a routine in libgcc to allocate more space from the heap.
20860 const uint32_t *RegMask =
20861 Subtarget->getRegisterInfo()->getCallPreservedMask(*MF, CallingConv::C);
20863 BuildMI(mallocMBB, DL, TII->get(X86::MOV64rr), X86::RDI)
20865 BuildMI(mallocMBB, DL, TII->get(X86::CALL64pcrel32))
20866 .addExternalSymbol("__morestack_allocate_stack_space")
20867 .addRegMask(RegMask)
20868 .addReg(X86::RDI, RegState::Implicit)
20869 .addReg(X86::RAX, RegState::ImplicitDefine);
20870 } else if (Is64Bit) {
20871 BuildMI(mallocMBB, DL, TII->get(X86::MOV32rr), X86::EDI)
20873 BuildMI(mallocMBB, DL, TII->get(X86::CALL64pcrel32))
20874 .addExternalSymbol("__morestack_allocate_stack_space")
20875 .addRegMask(RegMask)
20876 .addReg(X86::EDI, RegState::Implicit)
20877 .addReg(X86::EAX, RegState::ImplicitDefine);
20879 BuildMI(mallocMBB, DL, TII->get(X86::SUB32ri), physSPReg).addReg(physSPReg)
20881 BuildMI(mallocMBB, DL, TII->get(X86::PUSH32r)).addReg(sizeVReg);
20882 BuildMI(mallocMBB, DL, TII->get(X86::CALLpcrel32))
20883 .addExternalSymbol("__morestack_allocate_stack_space")
20884 .addRegMask(RegMask)
20885 .addReg(X86::EAX, RegState::ImplicitDefine);
20889 BuildMI(mallocMBB, DL, TII->get(X86::ADD32ri), physSPReg).addReg(physSPReg)
20892 BuildMI(mallocMBB, DL, TII->get(TargetOpcode::COPY), mallocPtrVReg)
20893 .addReg(IsLP64 ? X86::RAX : X86::EAX);
20894 BuildMI(mallocMBB, DL, TII->get(X86::JMP_1)).addMBB(continueMBB);
20896 // Set up the CFG correctly.
20897 BB->addSuccessor(bumpMBB);
20898 BB->addSuccessor(mallocMBB);
20899 mallocMBB->addSuccessor(continueMBB);
20900 bumpMBB->addSuccessor(continueMBB);
20902 // Take care of the PHI nodes.
20903 BuildMI(*continueMBB, continueMBB->begin(), DL, TII->get(X86::PHI),
20904 MI->getOperand(0).getReg())
20905 .addReg(mallocPtrVReg).addMBB(mallocMBB)
20906 .addReg(bumpSPPtrVReg).addMBB(bumpMBB);
20908 // Delete the original pseudo instruction.
20909 MI->eraseFromParent();
20912 return continueMBB;
20915 MachineBasicBlock *
20916 X86TargetLowering::EmitLoweredWinAlloca(MachineInstr *MI,
20917 MachineBasicBlock *BB) const {
20918 DebugLoc DL = MI->getDebugLoc();
20920 assert(!Subtarget->isTargetMachO());
20922 Subtarget->getFrameLowering()->emitStackProbeCall(*BB->getParent(), *BB, MI,
20925 MI->eraseFromParent(); // The pseudo instruction is gone now.
20929 MachineBasicBlock *
20930 X86TargetLowering::EmitLoweredTLSCall(MachineInstr *MI,
20931 MachineBasicBlock *BB) const {
20932 // This is pretty easy. We're taking the value that we received from
20933 // our load from the relocation, sticking it in either RDI (x86-64)
20934 // or EAX and doing an indirect call. The return value will then
20935 // be in the normal return register.
20936 MachineFunction *F = BB->getParent();
20937 const X86InstrInfo *TII = Subtarget->getInstrInfo();
20938 DebugLoc DL = MI->getDebugLoc();
20940 assert(Subtarget->isTargetDarwin() && "Darwin only instr emitted?");
20941 assert(MI->getOperand(3).isGlobal() && "This should be a global");
20943 // Get a register mask for the lowered call.
20944 // FIXME: The 32-bit calls have non-standard calling conventions. Use a
20945 // proper register mask.
20946 const uint32_t *RegMask =
20947 Subtarget->getRegisterInfo()->getCallPreservedMask(*F, CallingConv::C);
20948 if (Subtarget->is64Bit()) {
20949 MachineInstrBuilder MIB = BuildMI(*BB, MI, DL,
20950 TII->get(X86::MOV64rm), X86::RDI)
20952 .addImm(0).addReg(0)
20953 .addGlobalAddress(MI->getOperand(3).getGlobal(), 0,
20954 MI->getOperand(3).getTargetFlags())
20956 MIB = BuildMI(*BB, MI, DL, TII->get(X86::CALL64m));
20957 addDirectMem(MIB, X86::RDI);
20958 MIB.addReg(X86::RAX, RegState::ImplicitDefine).addRegMask(RegMask);
20959 } else if (F->getTarget().getRelocationModel() != Reloc::PIC_) {
20960 MachineInstrBuilder MIB = BuildMI(*BB, MI, DL,
20961 TII->get(X86::MOV32rm), X86::EAX)
20963 .addImm(0).addReg(0)
20964 .addGlobalAddress(MI->getOperand(3).getGlobal(), 0,
20965 MI->getOperand(3).getTargetFlags())
20967 MIB = BuildMI(*BB, MI, DL, TII->get(X86::CALL32m));
20968 addDirectMem(MIB, X86::EAX);
20969 MIB.addReg(X86::EAX, RegState::ImplicitDefine).addRegMask(RegMask);
20971 MachineInstrBuilder MIB = BuildMI(*BB, MI, DL,
20972 TII->get(X86::MOV32rm), X86::EAX)
20973 .addReg(TII->getGlobalBaseReg(F))
20974 .addImm(0).addReg(0)
20975 .addGlobalAddress(MI->getOperand(3).getGlobal(), 0,
20976 MI->getOperand(3).getTargetFlags())
20978 MIB = BuildMI(*BB, MI, DL, TII->get(X86::CALL32m));
20979 addDirectMem(MIB, X86::EAX);
20980 MIB.addReg(X86::EAX, RegState::ImplicitDefine).addRegMask(RegMask);
20983 MI->eraseFromParent(); // The pseudo instruction is gone now.
20987 MachineBasicBlock *
20988 X86TargetLowering::emitEHSjLjSetJmp(MachineInstr *MI,
20989 MachineBasicBlock *MBB) const {
20990 DebugLoc DL = MI->getDebugLoc();
20991 MachineFunction *MF = MBB->getParent();
20992 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
20993 MachineRegisterInfo &MRI = MF->getRegInfo();
20995 const BasicBlock *BB = MBB->getBasicBlock();
20996 MachineFunction::iterator I = MBB;
20999 // Memory Reference
21000 MachineInstr::mmo_iterator MMOBegin = MI->memoperands_begin();
21001 MachineInstr::mmo_iterator MMOEnd = MI->memoperands_end();
21004 unsigned MemOpndSlot = 0;
21006 unsigned CurOp = 0;
21008 DstReg = MI->getOperand(CurOp++).getReg();
21009 const TargetRegisterClass *RC = MRI.getRegClass(DstReg);
21010 assert(RC->hasType(MVT::i32) && "Invalid destination!");
21011 unsigned mainDstReg = MRI.createVirtualRegister(RC);
21012 unsigned restoreDstReg = MRI.createVirtualRegister(RC);
21014 MemOpndSlot = CurOp;
21016 MVT PVT = getPointerTy(MF->getDataLayout());
21017 assert((PVT == MVT::i64 || PVT == MVT::i32) &&
21018 "Invalid Pointer Size!");
21020 // For v = setjmp(buf), we generate
21023 // buf[LabelOffset] = restoreMBB
21024 // SjLjSetup restoreMBB
21030 // v = phi(main, restore)
21033 // if base pointer being used, load it from frame
21036 MachineBasicBlock *thisMBB = MBB;
21037 MachineBasicBlock *mainMBB = MF->CreateMachineBasicBlock(BB);
21038 MachineBasicBlock *sinkMBB = MF->CreateMachineBasicBlock(BB);
21039 MachineBasicBlock *restoreMBB = MF->CreateMachineBasicBlock(BB);
21040 MF->insert(I, mainMBB);
21041 MF->insert(I, sinkMBB);
21042 MF->push_back(restoreMBB);
21044 MachineInstrBuilder MIB;
21046 // Transfer the remainder of BB and its successor edges to sinkMBB.
21047 sinkMBB->splice(sinkMBB->begin(), MBB,
21048 std::next(MachineBasicBlock::iterator(MI)), MBB->end());
21049 sinkMBB->transferSuccessorsAndUpdatePHIs(MBB);
21052 unsigned PtrStoreOpc = 0;
21053 unsigned LabelReg = 0;
21054 const int64_t LabelOffset = 1 * PVT.getStoreSize();
21055 Reloc::Model RM = MF->getTarget().getRelocationModel();
21056 bool UseImmLabel = (MF->getTarget().getCodeModel() == CodeModel::Small) &&
21057 (RM == Reloc::Static || RM == Reloc::DynamicNoPIC);
21059 // Prepare IP either in reg or imm.
21060 if (!UseImmLabel) {
21061 PtrStoreOpc = (PVT == MVT::i64) ? X86::MOV64mr : X86::MOV32mr;
21062 const TargetRegisterClass *PtrRC = getRegClassFor(PVT);
21063 LabelReg = MRI.createVirtualRegister(PtrRC);
21064 if (Subtarget->is64Bit()) {
21065 MIB = BuildMI(*thisMBB, MI, DL, TII->get(X86::LEA64r), LabelReg)
21069 .addMBB(restoreMBB)
21072 const X86InstrInfo *XII = static_cast<const X86InstrInfo*>(TII);
21073 MIB = BuildMI(*thisMBB, MI, DL, TII->get(X86::LEA32r), LabelReg)
21074 .addReg(XII->getGlobalBaseReg(MF))
21077 .addMBB(restoreMBB, Subtarget->ClassifyBlockAddressReference())
21081 PtrStoreOpc = (PVT == MVT::i64) ? X86::MOV64mi32 : X86::MOV32mi;
21083 MIB = BuildMI(*thisMBB, MI, DL, TII->get(PtrStoreOpc));
21084 for (unsigned i = 0; i < X86::AddrNumOperands; ++i) {
21085 if (i == X86::AddrDisp)
21086 MIB.addDisp(MI->getOperand(MemOpndSlot + i), LabelOffset);
21088 MIB.addOperand(MI->getOperand(MemOpndSlot + i));
21091 MIB.addReg(LabelReg);
21093 MIB.addMBB(restoreMBB);
21094 MIB.setMemRefs(MMOBegin, MMOEnd);
21096 MIB = BuildMI(*thisMBB, MI, DL, TII->get(X86::EH_SjLj_Setup))
21097 .addMBB(restoreMBB);
21099 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
21100 MIB.addRegMask(RegInfo->getNoPreservedMask());
21101 thisMBB->addSuccessor(mainMBB);
21102 thisMBB->addSuccessor(restoreMBB);
21106 BuildMI(mainMBB, DL, TII->get(X86::MOV32r0), mainDstReg);
21107 mainMBB->addSuccessor(sinkMBB);
21110 BuildMI(*sinkMBB, sinkMBB->begin(), DL,
21111 TII->get(X86::PHI), DstReg)
21112 .addReg(mainDstReg).addMBB(mainMBB)
21113 .addReg(restoreDstReg).addMBB(restoreMBB);
21116 if (RegInfo->hasBasePointer(*MF)) {
21117 const bool Uses64BitFramePtr =
21118 Subtarget->isTarget64BitLP64() || Subtarget->isTargetNaCl64();
21119 X86MachineFunctionInfo *X86FI = MF->getInfo<X86MachineFunctionInfo>();
21120 X86FI->setRestoreBasePointer(MF);
21121 unsigned FramePtr = RegInfo->getFrameRegister(*MF);
21122 unsigned BasePtr = RegInfo->getBaseRegister();
21123 unsigned Opm = Uses64BitFramePtr ? X86::MOV64rm : X86::MOV32rm;
21124 addRegOffset(BuildMI(restoreMBB, DL, TII->get(Opm), BasePtr),
21125 FramePtr, true, X86FI->getRestoreBasePointerOffset())
21126 .setMIFlag(MachineInstr::FrameSetup);
21128 BuildMI(restoreMBB, DL, TII->get(X86::MOV32ri), restoreDstReg).addImm(1);
21129 BuildMI(restoreMBB, DL, TII->get(X86::JMP_1)).addMBB(sinkMBB);
21130 restoreMBB->addSuccessor(sinkMBB);
21132 MI->eraseFromParent();
21136 MachineBasicBlock *
21137 X86TargetLowering::emitEHSjLjLongJmp(MachineInstr *MI,
21138 MachineBasicBlock *MBB) const {
21139 DebugLoc DL = MI->getDebugLoc();
21140 MachineFunction *MF = MBB->getParent();
21141 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
21142 MachineRegisterInfo &MRI = MF->getRegInfo();
21144 // Memory Reference
21145 MachineInstr::mmo_iterator MMOBegin = MI->memoperands_begin();
21146 MachineInstr::mmo_iterator MMOEnd = MI->memoperands_end();
21148 MVT PVT = getPointerTy(MF->getDataLayout());
21149 assert((PVT == MVT::i64 || PVT == MVT::i32) &&
21150 "Invalid Pointer Size!");
21152 const TargetRegisterClass *RC =
21153 (PVT == MVT::i64) ? &X86::GR64RegClass : &X86::GR32RegClass;
21154 unsigned Tmp = MRI.createVirtualRegister(RC);
21155 // Since FP is only updated here but NOT referenced, it's treated as GPR.
21156 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
21157 unsigned FP = (PVT == MVT::i64) ? X86::RBP : X86::EBP;
21158 unsigned SP = RegInfo->getStackRegister();
21160 MachineInstrBuilder MIB;
21162 const int64_t LabelOffset = 1 * PVT.getStoreSize();
21163 const int64_t SPOffset = 2 * PVT.getStoreSize();
21165 unsigned PtrLoadOpc = (PVT == MVT::i64) ? X86::MOV64rm : X86::MOV32rm;
21166 unsigned IJmpOpc = (PVT == MVT::i64) ? X86::JMP64r : X86::JMP32r;
21169 MIB = BuildMI(*MBB, MI, DL, TII->get(PtrLoadOpc), FP);
21170 for (unsigned i = 0; i < X86::AddrNumOperands; ++i)
21171 MIB.addOperand(MI->getOperand(i));
21172 MIB.setMemRefs(MMOBegin, MMOEnd);
21174 MIB = BuildMI(*MBB, MI, DL, TII->get(PtrLoadOpc), Tmp);
21175 for (unsigned i = 0; i < X86::AddrNumOperands; ++i) {
21176 if (i == X86::AddrDisp)
21177 MIB.addDisp(MI->getOperand(i), LabelOffset);
21179 MIB.addOperand(MI->getOperand(i));
21181 MIB.setMemRefs(MMOBegin, MMOEnd);
21183 MIB = BuildMI(*MBB, MI, DL, TII->get(PtrLoadOpc), SP);
21184 for (unsigned i = 0; i < X86::AddrNumOperands; ++i) {
21185 if (i == X86::AddrDisp)
21186 MIB.addDisp(MI->getOperand(i), SPOffset);
21188 MIB.addOperand(MI->getOperand(i));
21190 MIB.setMemRefs(MMOBegin, MMOEnd);
21192 BuildMI(*MBB, MI, DL, TII->get(IJmpOpc)).addReg(Tmp);
21194 MI->eraseFromParent();
21198 // Replace 213-type (isel default) FMA3 instructions with 231-type for
21199 // accumulator loops. Writing back to the accumulator allows the coalescer
21200 // to remove extra copies in the loop.
21201 // FIXME: Do this on AVX512. We don't support 231 variants yet (PR23937).
21202 MachineBasicBlock *
21203 X86TargetLowering::emitFMA3Instr(MachineInstr *MI,
21204 MachineBasicBlock *MBB) const {
21205 MachineOperand &AddendOp = MI->getOperand(3);
21207 // Bail out early if the addend isn't a register - we can't switch these.
21208 if (!AddendOp.isReg())
21211 MachineFunction &MF = *MBB->getParent();
21212 MachineRegisterInfo &MRI = MF.getRegInfo();
21214 // Check whether the addend is defined by a PHI:
21215 assert(MRI.hasOneDef(AddendOp.getReg()) && "Multiple defs in SSA?");
21216 MachineInstr &AddendDef = *MRI.def_instr_begin(AddendOp.getReg());
21217 if (!AddendDef.isPHI())
21220 // Look for the following pattern:
21222 // %addend = phi [%entry, 0], [%loop, %result]
21224 // %result<tied1> = FMA213 %m2<tied0>, %m1, %addend
21228 // %addend = phi [%entry, 0], [%loop, %result]
21230 // %result<tied1> = FMA231 %addend<tied0>, %m1, %m2
21232 for (unsigned i = 1, e = AddendDef.getNumOperands(); i < e; i += 2) {
21233 assert(AddendDef.getOperand(i).isReg());
21234 MachineOperand PHISrcOp = AddendDef.getOperand(i);
21235 MachineInstr &PHISrcInst = *MRI.def_instr_begin(PHISrcOp.getReg());
21236 if (&PHISrcInst == MI) {
21237 // Found a matching instruction.
21238 unsigned NewFMAOpc = 0;
21239 switch (MI->getOpcode()) {
21240 case X86::VFMADDPDr213r: NewFMAOpc = X86::VFMADDPDr231r; break;
21241 case X86::VFMADDPSr213r: NewFMAOpc = X86::VFMADDPSr231r; break;
21242 case X86::VFMADDSDr213r: NewFMAOpc = X86::VFMADDSDr231r; break;
21243 case X86::VFMADDSSr213r: NewFMAOpc = X86::VFMADDSSr231r; break;
21244 case X86::VFMSUBPDr213r: NewFMAOpc = X86::VFMSUBPDr231r; break;
21245 case X86::VFMSUBPSr213r: NewFMAOpc = X86::VFMSUBPSr231r; break;
21246 case X86::VFMSUBSDr213r: NewFMAOpc = X86::VFMSUBSDr231r; break;
21247 case X86::VFMSUBSSr213r: NewFMAOpc = X86::VFMSUBSSr231r; break;
21248 case X86::VFNMADDPDr213r: NewFMAOpc = X86::VFNMADDPDr231r; break;
21249 case X86::VFNMADDPSr213r: NewFMAOpc = X86::VFNMADDPSr231r; break;
21250 case X86::VFNMADDSDr213r: NewFMAOpc = X86::VFNMADDSDr231r; break;
21251 case X86::VFNMADDSSr213r: NewFMAOpc = X86::VFNMADDSSr231r; break;
21252 case X86::VFNMSUBPDr213r: NewFMAOpc = X86::VFNMSUBPDr231r; break;
21253 case X86::VFNMSUBPSr213r: NewFMAOpc = X86::VFNMSUBPSr231r; break;
21254 case X86::VFNMSUBSDr213r: NewFMAOpc = X86::VFNMSUBSDr231r; break;
21255 case X86::VFNMSUBSSr213r: NewFMAOpc = X86::VFNMSUBSSr231r; break;
21256 case X86::VFMADDSUBPDr213r: NewFMAOpc = X86::VFMADDSUBPDr231r; break;
21257 case X86::VFMADDSUBPSr213r: NewFMAOpc = X86::VFMADDSUBPSr231r; break;
21258 case X86::VFMSUBADDPDr213r: NewFMAOpc = X86::VFMSUBADDPDr231r; break;
21259 case X86::VFMSUBADDPSr213r: NewFMAOpc = X86::VFMSUBADDPSr231r; break;
21261 case X86::VFMADDPDr213rY: NewFMAOpc = X86::VFMADDPDr231rY; break;
21262 case X86::VFMADDPSr213rY: NewFMAOpc = X86::VFMADDPSr231rY; break;
21263 case X86::VFMSUBPDr213rY: NewFMAOpc = X86::VFMSUBPDr231rY; break;
21264 case X86::VFMSUBPSr213rY: NewFMAOpc = X86::VFMSUBPSr231rY; break;
21265 case X86::VFNMADDPDr213rY: NewFMAOpc = X86::VFNMADDPDr231rY; break;
21266 case X86::VFNMADDPSr213rY: NewFMAOpc = X86::VFNMADDPSr231rY; break;
21267 case X86::VFNMSUBPDr213rY: NewFMAOpc = X86::VFNMSUBPDr231rY; break;
21268 case X86::VFNMSUBPSr213rY: NewFMAOpc = X86::VFNMSUBPSr231rY; break;
21269 case X86::VFMADDSUBPDr213rY: NewFMAOpc = X86::VFMADDSUBPDr231rY; break;
21270 case X86::VFMADDSUBPSr213rY: NewFMAOpc = X86::VFMADDSUBPSr231rY; break;
21271 case X86::VFMSUBADDPDr213rY: NewFMAOpc = X86::VFMSUBADDPDr231rY; break;
21272 case X86::VFMSUBADDPSr213rY: NewFMAOpc = X86::VFMSUBADDPSr231rY; break;
21273 default: llvm_unreachable("Unrecognized FMA variant.");
21276 const TargetInstrInfo &TII = *Subtarget->getInstrInfo();
21277 MachineInstrBuilder MIB =
21278 BuildMI(MF, MI->getDebugLoc(), TII.get(NewFMAOpc))
21279 .addOperand(MI->getOperand(0))
21280 .addOperand(MI->getOperand(3))
21281 .addOperand(MI->getOperand(2))
21282 .addOperand(MI->getOperand(1));
21283 MBB->insert(MachineBasicBlock::iterator(MI), MIB);
21284 MI->eraseFromParent();
21291 MachineBasicBlock *
21292 X86TargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
21293 MachineBasicBlock *BB) const {
21294 switch (MI->getOpcode()) {
21295 default: llvm_unreachable("Unexpected instr type to insert");
21296 case X86::TAILJMPd64:
21297 case X86::TAILJMPr64:
21298 case X86::TAILJMPm64:
21299 case X86::TAILJMPd64_REX:
21300 case X86::TAILJMPr64_REX:
21301 case X86::TAILJMPm64_REX:
21302 llvm_unreachable("TAILJMP64 would not be touched here.");
21303 case X86::TCRETURNdi64:
21304 case X86::TCRETURNri64:
21305 case X86::TCRETURNmi64:
21307 case X86::WIN_ALLOCA:
21308 return EmitLoweredWinAlloca(MI, BB);
21309 case X86::SEG_ALLOCA_32:
21310 case X86::SEG_ALLOCA_64:
21311 return EmitLoweredSegAlloca(MI, BB);
21312 case X86::TLSCall_32:
21313 case X86::TLSCall_64:
21314 return EmitLoweredTLSCall(MI, BB);
21315 case X86::CMOV_FR32:
21316 case X86::CMOV_FR64:
21317 case X86::CMOV_GR8:
21318 case X86::CMOV_GR16:
21319 case X86::CMOV_GR32:
21320 case X86::CMOV_RFP32:
21321 case X86::CMOV_RFP64:
21322 case X86::CMOV_RFP80:
21323 case X86::CMOV_V2F64:
21324 case X86::CMOV_V2I64:
21325 case X86::CMOV_V4F32:
21326 case X86::CMOV_V4F64:
21327 case X86::CMOV_V4I64:
21328 case X86::CMOV_V16F32:
21329 case X86::CMOV_V8F32:
21330 case X86::CMOV_V8F64:
21331 case X86::CMOV_V8I64:
21332 case X86::CMOV_V8I1:
21333 case X86::CMOV_V16I1:
21334 case X86::CMOV_V32I1:
21335 case X86::CMOV_V64I1:
21336 return EmitLoweredSelect(MI, BB);
21338 case X86::RELEASE_FADD32mr:
21339 case X86::RELEASE_FADD64mr:
21340 return EmitLoweredAtomicFP(MI, BB);
21342 case X86::FP32_TO_INT16_IN_MEM:
21343 case X86::FP32_TO_INT32_IN_MEM:
21344 case X86::FP32_TO_INT64_IN_MEM:
21345 case X86::FP64_TO_INT16_IN_MEM:
21346 case X86::FP64_TO_INT32_IN_MEM:
21347 case X86::FP64_TO_INT64_IN_MEM:
21348 case X86::FP80_TO_INT16_IN_MEM:
21349 case X86::FP80_TO_INT32_IN_MEM:
21350 case X86::FP80_TO_INT64_IN_MEM: {
21351 MachineFunction *F = BB->getParent();
21352 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
21353 DebugLoc DL = MI->getDebugLoc();
21355 // Change the floating point control register to use "round towards zero"
21356 // mode when truncating to an integer value.
21357 int CWFrameIdx = F->getFrameInfo()->CreateStackObject(2, 2, false);
21358 addFrameReference(BuildMI(*BB, MI, DL,
21359 TII->get(X86::FNSTCW16m)), CWFrameIdx);
21361 // Load the old value of the high byte of the control word...
21363 F->getRegInfo().createVirtualRegister(&X86::GR16RegClass);
21364 addFrameReference(BuildMI(*BB, MI, DL, TII->get(X86::MOV16rm), OldCW),
21367 // Set the high part to be round to zero...
21368 addFrameReference(BuildMI(*BB, MI, DL, TII->get(X86::MOV16mi)), CWFrameIdx)
21371 // Reload the modified control word now...
21372 addFrameReference(BuildMI(*BB, MI, DL,
21373 TII->get(X86::FLDCW16m)), CWFrameIdx);
21375 // Restore the memory image of control word to original value
21376 addFrameReference(BuildMI(*BB, MI, DL, TII->get(X86::MOV16mr)), CWFrameIdx)
21379 // Get the X86 opcode to use.
21381 switch (MI->getOpcode()) {
21382 default: llvm_unreachable("illegal opcode!");
21383 case X86::FP32_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m32; break;
21384 case X86::FP32_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m32; break;
21385 case X86::FP32_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m32; break;
21386 case X86::FP64_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m64; break;
21387 case X86::FP64_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m64; break;
21388 case X86::FP64_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m64; break;
21389 case X86::FP80_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m80; break;
21390 case X86::FP80_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m80; break;
21391 case X86::FP80_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m80; break;
21395 MachineOperand &Op = MI->getOperand(0);
21397 AM.BaseType = X86AddressMode::RegBase;
21398 AM.Base.Reg = Op.getReg();
21400 AM.BaseType = X86AddressMode::FrameIndexBase;
21401 AM.Base.FrameIndex = Op.getIndex();
21403 Op = MI->getOperand(1);
21405 AM.Scale = Op.getImm();
21406 Op = MI->getOperand(2);
21408 AM.IndexReg = Op.getImm();
21409 Op = MI->getOperand(3);
21410 if (Op.isGlobal()) {
21411 AM.GV = Op.getGlobal();
21413 AM.Disp = Op.getImm();
21415 addFullAddress(BuildMI(*BB, MI, DL, TII->get(Opc)), AM)
21416 .addReg(MI->getOperand(X86::AddrNumOperands).getReg());
21418 // Reload the original control word now.
21419 addFrameReference(BuildMI(*BB, MI, DL,
21420 TII->get(X86::FLDCW16m)), CWFrameIdx);
21422 MI->eraseFromParent(); // The pseudo instruction is gone now.
21425 // String/text processing lowering.
21426 case X86::PCMPISTRM128REG:
21427 case X86::VPCMPISTRM128REG:
21428 case X86::PCMPISTRM128MEM:
21429 case X86::VPCMPISTRM128MEM:
21430 case X86::PCMPESTRM128REG:
21431 case X86::VPCMPESTRM128REG:
21432 case X86::PCMPESTRM128MEM:
21433 case X86::VPCMPESTRM128MEM:
21434 assert(Subtarget->hasSSE42() &&
21435 "Target must have SSE4.2 or AVX features enabled");
21436 return EmitPCMPSTRM(MI, BB, Subtarget->getInstrInfo());
21438 // String/text processing lowering.
21439 case X86::PCMPISTRIREG:
21440 case X86::VPCMPISTRIREG:
21441 case X86::PCMPISTRIMEM:
21442 case X86::VPCMPISTRIMEM:
21443 case X86::PCMPESTRIREG:
21444 case X86::VPCMPESTRIREG:
21445 case X86::PCMPESTRIMEM:
21446 case X86::VPCMPESTRIMEM:
21447 assert(Subtarget->hasSSE42() &&
21448 "Target must have SSE4.2 or AVX features enabled");
21449 return EmitPCMPSTRI(MI, BB, Subtarget->getInstrInfo());
21451 // Thread synchronization.
21453 return EmitMonitor(MI, BB, Subtarget);
21457 return EmitXBegin(MI, BB, Subtarget->getInstrInfo());
21459 case X86::VASTART_SAVE_XMM_REGS:
21460 return EmitVAStartSaveXMMRegsWithCustomInserter(MI, BB);
21462 case X86::VAARG_64:
21463 return EmitVAARG64WithCustomInserter(MI, BB);
21465 case X86::EH_SjLj_SetJmp32:
21466 case X86::EH_SjLj_SetJmp64:
21467 return emitEHSjLjSetJmp(MI, BB);
21469 case X86::EH_SjLj_LongJmp32:
21470 case X86::EH_SjLj_LongJmp64:
21471 return emitEHSjLjLongJmp(MI, BB);
21473 case TargetOpcode::STATEPOINT:
21474 // As an implementation detail, STATEPOINT shares the STACKMAP format at
21475 // this point in the process. We diverge later.
21476 return emitPatchPoint(MI, BB);
21478 case TargetOpcode::STACKMAP:
21479 case TargetOpcode::PATCHPOINT:
21480 return emitPatchPoint(MI, BB);
21482 case X86::VFMADDPDr213r:
21483 case X86::VFMADDPSr213r:
21484 case X86::VFMADDSDr213r:
21485 case X86::VFMADDSSr213r:
21486 case X86::VFMSUBPDr213r:
21487 case X86::VFMSUBPSr213r:
21488 case X86::VFMSUBSDr213r:
21489 case X86::VFMSUBSSr213r:
21490 case X86::VFNMADDPDr213r:
21491 case X86::VFNMADDPSr213r:
21492 case X86::VFNMADDSDr213r:
21493 case X86::VFNMADDSSr213r:
21494 case X86::VFNMSUBPDr213r:
21495 case X86::VFNMSUBPSr213r:
21496 case X86::VFNMSUBSDr213r:
21497 case X86::VFNMSUBSSr213r:
21498 case X86::VFMADDSUBPDr213r:
21499 case X86::VFMADDSUBPSr213r:
21500 case X86::VFMSUBADDPDr213r:
21501 case X86::VFMSUBADDPSr213r:
21502 case X86::VFMADDPDr213rY:
21503 case X86::VFMADDPSr213rY:
21504 case X86::VFMSUBPDr213rY:
21505 case X86::VFMSUBPSr213rY:
21506 case X86::VFNMADDPDr213rY:
21507 case X86::VFNMADDPSr213rY:
21508 case X86::VFNMSUBPDr213rY:
21509 case X86::VFNMSUBPSr213rY:
21510 case X86::VFMADDSUBPDr213rY:
21511 case X86::VFMADDSUBPSr213rY:
21512 case X86::VFMSUBADDPDr213rY:
21513 case X86::VFMSUBADDPSr213rY:
21514 return emitFMA3Instr(MI, BB);
21518 //===----------------------------------------------------------------------===//
21519 // X86 Optimization Hooks
21520 //===----------------------------------------------------------------------===//
21522 void X86TargetLowering::computeKnownBitsForTargetNode(const SDValue Op,
21525 const SelectionDAG &DAG,
21526 unsigned Depth) const {
21527 unsigned BitWidth = KnownZero.getBitWidth();
21528 unsigned Opc = Op.getOpcode();
21529 assert((Opc >= ISD::BUILTIN_OP_END ||
21530 Opc == ISD::INTRINSIC_WO_CHAIN ||
21531 Opc == ISD::INTRINSIC_W_CHAIN ||
21532 Opc == ISD::INTRINSIC_VOID) &&
21533 "Should use MaskedValueIsZero if you don't know whether Op"
21534 " is a target node!");
21536 KnownZero = KnownOne = APInt(BitWidth, 0); // Don't know anything.
21550 // These nodes' second result is a boolean.
21551 if (Op.getResNo() == 0)
21554 case X86ISD::SETCC:
21555 KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - 1);
21557 case ISD::INTRINSIC_WO_CHAIN: {
21558 unsigned IntId = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
21559 unsigned NumLoBits = 0;
21562 case Intrinsic::x86_sse_movmsk_ps:
21563 case Intrinsic::x86_avx_movmsk_ps_256:
21564 case Intrinsic::x86_sse2_movmsk_pd:
21565 case Intrinsic::x86_avx_movmsk_pd_256:
21566 case Intrinsic::x86_mmx_pmovmskb:
21567 case Intrinsic::x86_sse2_pmovmskb_128:
21568 case Intrinsic::x86_avx2_pmovmskb: {
21569 // High bits of movmskp{s|d}, pmovmskb are known zero.
21571 default: llvm_unreachable("Impossible intrinsic"); // Can't reach here.
21572 case Intrinsic::x86_sse_movmsk_ps: NumLoBits = 4; break;
21573 case Intrinsic::x86_avx_movmsk_ps_256: NumLoBits = 8; break;
21574 case Intrinsic::x86_sse2_movmsk_pd: NumLoBits = 2; break;
21575 case Intrinsic::x86_avx_movmsk_pd_256: NumLoBits = 4; break;
21576 case Intrinsic::x86_mmx_pmovmskb: NumLoBits = 8; break;
21577 case Intrinsic::x86_sse2_pmovmskb_128: NumLoBits = 16; break;
21578 case Intrinsic::x86_avx2_pmovmskb: NumLoBits = 32; break;
21580 KnownZero = APInt::getHighBitsSet(BitWidth, BitWidth - NumLoBits);
21589 unsigned X86TargetLowering::ComputeNumSignBitsForTargetNode(
21591 const SelectionDAG &,
21592 unsigned Depth) const {
21593 // SETCC_CARRY sets the dest to ~0 for true or 0 for false.
21594 if (Op.getOpcode() == X86ISD::SETCC_CARRY)
21595 return Op.getValueType().getScalarType().getSizeInBits();
21601 /// isGAPlusOffset - Returns true (and the GlobalValue and the offset) if the
21602 /// node is a GlobalAddress + offset.
21603 bool X86TargetLowering::isGAPlusOffset(SDNode *N,
21604 const GlobalValue* &GA,
21605 int64_t &Offset) const {
21606 if (N->getOpcode() == X86ISD::Wrapper) {
21607 if (isa<GlobalAddressSDNode>(N->getOperand(0))) {
21608 GA = cast<GlobalAddressSDNode>(N->getOperand(0))->getGlobal();
21609 Offset = cast<GlobalAddressSDNode>(N->getOperand(0))->getOffset();
21613 return TargetLowering::isGAPlusOffset(N, GA, Offset);
21616 /// isShuffleHigh128VectorInsertLow - Checks whether the shuffle node is the
21617 /// same as extracting the high 128-bit part of 256-bit vector and then
21618 /// inserting the result into the low part of a new 256-bit vector
21619 static bool isShuffleHigh128VectorInsertLow(ShuffleVectorSDNode *SVOp) {
21620 EVT VT = SVOp->getValueType(0);
21621 unsigned NumElems = VT.getVectorNumElements();
21623 // vector_shuffle <4, 5, 6, 7, u, u, u, u> or <2, 3, u, u>
21624 for (unsigned i = 0, j = NumElems/2; i != NumElems/2; ++i, ++j)
21625 if (!isUndefOrEqual(SVOp->getMaskElt(i), j) ||
21626 SVOp->getMaskElt(j) >= 0)
21632 /// isShuffleLow128VectorInsertHigh - Checks whether the shuffle node is the
21633 /// same as extracting the low 128-bit part of 256-bit vector and then
21634 /// inserting the result into the high part of a new 256-bit vector
21635 static bool isShuffleLow128VectorInsertHigh(ShuffleVectorSDNode *SVOp) {
21636 EVT VT = SVOp->getValueType(0);
21637 unsigned NumElems = VT.getVectorNumElements();
21639 // vector_shuffle <u, u, u, u, 0, 1, 2, 3> or <u, u, 0, 1>
21640 for (unsigned i = NumElems/2, j = 0; i != NumElems; ++i, ++j)
21641 if (!isUndefOrEqual(SVOp->getMaskElt(i), j) ||
21642 SVOp->getMaskElt(j) >= 0)
21648 /// PerformShuffleCombine256 - Performs shuffle combines for 256-bit vectors.
21649 static SDValue PerformShuffleCombine256(SDNode *N, SelectionDAG &DAG,
21650 TargetLowering::DAGCombinerInfo &DCI,
21651 const X86Subtarget* Subtarget) {
21653 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
21654 SDValue V1 = SVOp->getOperand(0);
21655 SDValue V2 = SVOp->getOperand(1);
21656 EVT VT = SVOp->getValueType(0);
21657 unsigned NumElems = VT.getVectorNumElements();
21659 if (V1.getOpcode() == ISD::CONCAT_VECTORS &&
21660 V2.getOpcode() == ISD::CONCAT_VECTORS) {
21664 // V UNDEF BUILD_VECTOR UNDEF
21666 // CONCAT_VECTOR CONCAT_VECTOR
21669 // RESULT: V + zero extended
21671 if (V2.getOperand(0).getOpcode() != ISD::BUILD_VECTOR ||
21672 V2.getOperand(1).getOpcode() != ISD::UNDEF ||
21673 V1.getOperand(1).getOpcode() != ISD::UNDEF)
21676 if (!ISD::isBuildVectorAllZeros(V2.getOperand(0).getNode()))
21679 // To match the shuffle mask, the first half of the mask should
21680 // be exactly the first vector, and all the rest a splat with the
21681 // first element of the second one.
21682 for (unsigned i = 0; i != NumElems/2; ++i)
21683 if (!isUndefOrEqual(SVOp->getMaskElt(i), i) ||
21684 !isUndefOrEqual(SVOp->getMaskElt(i+NumElems/2), NumElems))
21687 // If V1 is coming from a vector load then just fold to a VZEXT_LOAD.
21688 if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(V1.getOperand(0))) {
21689 if (Ld->hasNUsesOfValue(1, 0)) {
21690 SDVTList Tys = DAG.getVTList(MVT::v4i64, MVT::Other);
21691 SDValue Ops[] = { Ld->getChain(), Ld->getBasePtr() };
21693 DAG.getMemIntrinsicNode(X86ISD::VZEXT_LOAD, dl, Tys, Ops,
21695 Ld->getPointerInfo(),
21696 Ld->getAlignment(),
21697 false/*isVolatile*/, true/*ReadMem*/,
21698 false/*WriteMem*/);
21700 // Make sure the newly-created LOAD is in the same position as Ld in
21701 // terms of dependency. We create a TokenFactor for Ld and ResNode,
21702 // and update uses of Ld's output chain to use the TokenFactor.
21703 if (Ld->hasAnyUseOfValue(1)) {
21704 SDValue NewChain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
21705 SDValue(Ld, 1), SDValue(ResNode.getNode(), 1));
21706 DAG.ReplaceAllUsesOfValueWith(SDValue(Ld, 1), NewChain);
21707 DAG.UpdateNodeOperands(NewChain.getNode(), SDValue(Ld, 1),
21708 SDValue(ResNode.getNode(), 1));
21711 return DAG.getBitcast(VT, ResNode);
21715 // Emit a zeroed vector and insert the desired subvector on its
21717 SDValue Zeros = getZeroVector(VT, Subtarget, DAG, dl);
21718 SDValue InsV = Insert128BitVector(Zeros, V1.getOperand(0), 0, DAG, dl);
21719 return DCI.CombineTo(N, InsV);
21722 //===--------------------------------------------------------------------===//
21723 // Combine some shuffles into subvector extracts and inserts:
21726 // vector_shuffle <4, 5, 6, 7, u, u, u, u> or <2, 3, u, u>
21727 if (isShuffleHigh128VectorInsertLow(SVOp)) {
21728 SDValue V = Extract128BitVector(V1, NumElems/2, DAG, dl);
21729 SDValue InsV = Insert128BitVector(DAG.getUNDEF(VT), V, 0, DAG, dl);
21730 return DCI.CombineTo(N, InsV);
21733 // vector_shuffle <u, u, u, u, 0, 1, 2, 3> or <u, u, 0, 1>
21734 if (isShuffleLow128VectorInsertHigh(SVOp)) {
21735 SDValue V = Extract128BitVector(V1, 0, DAG, dl);
21736 SDValue InsV = Insert128BitVector(DAG.getUNDEF(VT), V, NumElems/2, DAG, dl);
21737 return DCI.CombineTo(N, InsV);
21743 /// \brief Combine an arbitrary chain of shuffles into a single instruction if
21746 /// This is the leaf of the recursive combinine below. When we have found some
21747 /// chain of single-use x86 shuffle instructions and accumulated the combined
21748 /// shuffle mask represented by them, this will try to pattern match that mask
21749 /// into either a single instruction if there is a special purpose instruction
21750 /// for this operation, or into a PSHUFB instruction which is a fully general
21751 /// instruction but should only be used to replace chains over a certain depth.
21752 static bool combineX86ShuffleChain(SDValue Op, SDValue Root, ArrayRef<int> Mask,
21753 int Depth, bool HasPSHUFB, SelectionDAG &DAG,
21754 TargetLowering::DAGCombinerInfo &DCI,
21755 const X86Subtarget *Subtarget) {
21756 assert(!Mask.empty() && "Cannot combine an empty shuffle mask!");
21758 // Find the operand that enters the chain. Note that multiple uses are OK
21759 // here, we're not going to remove the operand we find.
21760 SDValue Input = Op.getOperand(0);
21761 while (Input.getOpcode() == ISD::BITCAST)
21762 Input = Input.getOperand(0);
21764 MVT VT = Input.getSimpleValueType();
21765 MVT RootVT = Root.getSimpleValueType();
21768 // Just remove no-op shuffle masks.
21769 if (Mask.size() == 1) {
21770 DCI.CombineTo(Root.getNode(), DAG.getBitcast(RootVT, Input),
21775 // Use the float domain if the operand type is a floating point type.
21776 bool FloatDomain = VT.isFloatingPoint();
21778 // For floating point shuffles, we don't have free copies in the shuffle
21779 // instructions or the ability to load as part of the instruction, so
21780 // canonicalize their shuffles to UNPCK or MOV variants.
21782 // Note that even with AVX we prefer the PSHUFD form of shuffle for integer
21783 // vectors because it can have a load folded into it that UNPCK cannot. This
21784 // doesn't preclude something switching to the shorter encoding post-RA.
21786 // FIXME: Should teach these routines about AVX vector widths.
21787 if (FloatDomain && VT.getSizeInBits() == 128) {
21788 if (Mask.equals({0, 0}) || Mask.equals({1, 1})) {
21789 bool Lo = Mask.equals({0, 0});
21792 // Check if we have SSE3 which will let us use MOVDDUP. That instruction
21793 // is no slower than UNPCKLPD but has the option to fold the input operand
21794 // into even an unaligned memory load.
21795 if (Lo && Subtarget->hasSSE3()) {
21796 Shuffle = X86ISD::MOVDDUP;
21797 ShuffleVT = MVT::v2f64;
21799 // We have MOVLHPS and MOVHLPS throughout SSE and they encode smaller
21800 // than the UNPCK variants.
21801 Shuffle = Lo ? X86ISD::MOVLHPS : X86ISD::MOVHLPS;
21802 ShuffleVT = MVT::v4f32;
21804 if (Depth == 1 && Root->getOpcode() == Shuffle)
21805 return false; // Nothing to do!
21806 Op = DAG.getBitcast(ShuffleVT, Input);
21807 DCI.AddToWorklist(Op.getNode());
21808 if (Shuffle == X86ISD::MOVDDUP)
21809 Op = DAG.getNode(Shuffle, DL, ShuffleVT, Op);
21811 Op = DAG.getNode(Shuffle, DL, ShuffleVT, Op, Op);
21812 DCI.AddToWorklist(Op.getNode());
21813 DCI.CombineTo(Root.getNode(), DAG.getBitcast(RootVT, Op),
21817 if (Subtarget->hasSSE3() &&
21818 (Mask.equals({0, 0, 2, 2}) || Mask.equals({1, 1, 3, 3}))) {
21819 bool Lo = Mask.equals({0, 0, 2, 2});
21820 unsigned Shuffle = Lo ? X86ISD::MOVSLDUP : X86ISD::MOVSHDUP;
21821 MVT ShuffleVT = MVT::v4f32;
21822 if (Depth == 1 && Root->getOpcode() == Shuffle)
21823 return false; // Nothing to do!
21824 Op = DAG.getBitcast(ShuffleVT, Input);
21825 DCI.AddToWorklist(Op.getNode());
21826 Op = DAG.getNode(Shuffle, DL, ShuffleVT, Op);
21827 DCI.AddToWorklist(Op.getNode());
21828 DCI.CombineTo(Root.getNode(), DAG.getBitcast(RootVT, Op),
21832 if (Mask.equals({0, 0, 1, 1}) || Mask.equals({2, 2, 3, 3})) {
21833 bool Lo = Mask.equals({0, 0, 1, 1});
21834 unsigned Shuffle = Lo ? X86ISD::UNPCKL : X86ISD::UNPCKH;
21835 MVT ShuffleVT = MVT::v4f32;
21836 if (Depth == 1 && Root->getOpcode() == Shuffle)
21837 return false; // Nothing to do!
21838 Op = DAG.getBitcast(ShuffleVT, Input);
21839 DCI.AddToWorklist(Op.getNode());
21840 Op = DAG.getNode(Shuffle, DL, ShuffleVT, Op, Op);
21841 DCI.AddToWorklist(Op.getNode());
21842 DCI.CombineTo(Root.getNode(), DAG.getBitcast(RootVT, Op),
21848 // We always canonicalize the 8 x i16 and 16 x i8 shuffles into their UNPCK
21849 // variants as none of these have single-instruction variants that are
21850 // superior to the UNPCK formulation.
21851 if (!FloatDomain && VT.getSizeInBits() == 128 &&
21852 (Mask.equals({0, 0, 1, 1, 2, 2, 3, 3}) ||
21853 Mask.equals({4, 4, 5, 5, 6, 6, 7, 7}) ||
21854 Mask.equals({0, 0, 1, 1, 2, 2, 3, 3, 4, 4, 5, 5, 6, 6, 7, 7}) ||
21856 {8, 8, 9, 9, 10, 10, 11, 11, 12, 12, 13, 13, 14, 14, 15, 15}))) {
21857 bool Lo = Mask[0] == 0;
21858 unsigned Shuffle = Lo ? X86ISD::UNPCKL : X86ISD::UNPCKH;
21859 if (Depth == 1 && Root->getOpcode() == Shuffle)
21860 return false; // Nothing to do!
21862 switch (Mask.size()) {
21864 ShuffleVT = MVT::v8i16;
21867 ShuffleVT = MVT::v16i8;
21870 llvm_unreachable("Impossible mask size!");
21872 Op = DAG.getBitcast(ShuffleVT, Input);
21873 DCI.AddToWorklist(Op.getNode());
21874 Op = DAG.getNode(Shuffle, DL, ShuffleVT, Op, Op);
21875 DCI.AddToWorklist(Op.getNode());
21876 DCI.CombineTo(Root.getNode(), DAG.getBitcast(RootVT, Op),
21881 // Don't try to re-form single instruction chains under any circumstances now
21882 // that we've done encoding canonicalization for them.
21886 // If we have 3 or more shuffle instructions or a chain involving PSHUFB, we
21887 // can replace them with a single PSHUFB instruction profitably. Intel's
21888 // manuals suggest only using PSHUFB if doing so replacing 5 instructions, but
21889 // in practice PSHUFB tends to be *very* fast so we're more aggressive.
21890 if ((Depth >= 3 || HasPSHUFB) && Subtarget->hasSSSE3()) {
21891 SmallVector<SDValue, 16> PSHUFBMask;
21892 int NumBytes = VT.getSizeInBits() / 8;
21893 int Ratio = NumBytes / Mask.size();
21894 for (int i = 0; i < NumBytes; ++i) {
21895 if (Mask[i / Ratio] == SM_SentinelUndef) {
21896 PSHUFBMask.push_back(DAG.getUNDEF(MVT::i8));
21899 int M = Mask[i / Ratio] != SM_SentinelZero
21900 ? Ratio * Mask[i / Ratio] + i % Ratio
21902 PSHUFBMask.push_back(DAG.getConstant(M, DL, MVT::i8));
21904 MVT ByteVT = MVT::getVectorVT(MVT::i8, NumBytes);
21905 Op = DAG.getBitcast(ByteVT, Input);
21906 DCI.AddToWorklist(Op.getNode());
21907 SDValue PSHUFBMaskOp =
21908 DAG.getNode(ISD::BUILD_VECTOR, DL, ByteVT, PSHUFBMask);
21909 DCI.AddToWorklist(PSHUFBMaskOp.getNode());
21910 Op = DAG.getNode(X86ISD::PSHUFB, DL, ByteVT, Op, PSHUFBMaskOp);
21911 DCI.AddToWorklist(Op.getNode());
21912 DCI.CombineTo(Root.getNode(), DAG.getBitcast(RootVT, Op),
21917 // Failed to find any combines.
21921 /// \brief Fully generic combining of x86 shuffle instructions.
21923 /// This should be the last combine run over the x86 shuffle instructions. Once
21924 /// they have been fully optimized, this will recursively consider all chains
21925 /// of single-use shuffle instructions, build a generic model of the cumulative
21926 /// shuffle operation, and check for simpler instructions which implement this
21927 /// operation. We use this primarily for two purposes:
21929 /// 1) Collapse generic shuffles to specialized single instructions when
21930 /// equivalent. In most cases, this is just an encoding size win, but
21931 /// sometimes we will collapse multiple generic shuffles into a single
21932 /// special-purpose shuffle.
21933 /// 2) Look for sequences of shuffle instructions with 3 or more total
21934 /// instructions, and replace them with the slightly more expensive SSSE3
21935 /// PSHUFB instruction if available. We do this as the last combining step
21936 /// to ensure we avoid using PSHUFB if we can implement the shuffle with
21937 /// a suitable short sequence of other instructions. The PHUFB will either
21938 /// use a register or have to read from memory and so is slightly (but only
21939 /// slightly) more expensive than the other shuffle instructions.
21941 /// Because this is inherently a quadratic operation (for each shuffle in
21942 /// a chain, we recurse up the chain), the depth is limited to 8 instructions.
21943 /// This should never be an issue in practice as the shuffle lowering doesn't
21944 /// produce sequences of more than 8 instructions.
21946 /// FIXME: We will currently miss some cases where the redundant shuffling
21947 /// would simplify under the threshold for PSHUFB formation because of
21948 /// combine-ordering. To fix this, we should do the redundant instruction
21949 /// combining in this recursive walk.
21950 static bool combineX86ShufflesRecursively(SDValue Op, SDValue Root,
21951 ArrayRef<int> RootMask,
21952 int Depth, bool HasPSHUFB,
21954 TargetLowering::DAGCombinerInfo &DCI,
21955 const X86Subtarget *Subtarget) {
21956 // Bound the depth of our recursive combine because this is ultimately
21957 // quadratic in nature.
21961 // Directly rip through bitcasts to find the underlying operand.
21962 while (Op.getOpcode() == ISD::BITCAST && Op.getOperand(0).hasOneUse())
21963 Op = Op.getOperand(0);
21965 MVT VT = Op.getSimpleValueType();
21966 if (!VT.isVector())
21967 return false; // Bail if we hit a non-vector.
21969 assert(Root.getSimpleValueType().isVector() &&
21970 "Shuffles operate on vector types!");
21971 assert(VT.getSizeInBits() == Root.getSimpleValueType().getSizeInBits() &&
21972 "Can only combine shuffles of the same vector register size.");
21974 if (!isTargetShuffle(Op.getOpcode()))
21976 SmallVector<int, 16> OpMask;
21978 bool HaveMask = getTargetShuffleMask(Op.getNode(), VT, OpMask, IsUnary);
21979 // We only can combine unary shuffles which we can decode the mask for.
21980 if (!HaveMask || !IsUnary)
21983 assert(VT.getVectorNumElements() == OpMask.size() &&
21984 "Different mask size from vector size!");
21985 assert(((RootMask.size() > OpMask.size() &&
21986 RootMask.size() % OpMask.size() == 0) ||
21987 (OpMask.size() > RootMask.size() &&
21988 OpMask.size() % RootMask.size() == 0) ||
21989 OpMask.size() == RootMask.size()) &&
21990 "The smaller number of elements must divide the larger.");
21991 int RootRatio = std::max<int>(1, OpMask.size() / RootMask.size());
21992 int OpRatio = std::max<int>(1, RootMask.size() / OpMask.size());
21993 assert(((RootRatio == 1 && OpRatio == 1) ||
21994 (RootRatio == 1) != (OpRatio == 1)) &&
21995 "Must not have a ratio for both incoming and op masks!");
21997 SmallVector<int, 16> Mask;
21998 Mask.reserve(std::max(OpMask.size(), RootMask.size()));
22000 // Merge this shuffle operation's mask into our accumulated mask. Note that
22001 // this shuffle's mask will be the first applied to the input, followed by the
22002 // root mask to get us all the way to the root value arrangement. The reason
22003 // for this order is that we are recursing up the operation chain.
22004 for (int i = 0, e = std::max(OpMask.size(), RootMask.size()); i < e; ++i) {
22005 int RootIdx = i / RootRatio;
22006 if (RootMask[RootIdx] < 0) {
22007 // This is a zero or undef lane, we're done.
22008 Mask.push_back(RootMask[RootIdx]);
22012 int RootMaskedIdx = RootMask[RootIdx] * RootRatio + i % RootRatio;
22013 int OpIdx = RootMaskedIdx / OpRatio;
22014 if (OpMask[OpIdx] < 0) {
22015 // The incoming lanes are zero or undef, it doesn't matter which ones we
22017 Mask.push_back(OpMask[OpIdx]);
22021 // Ok, we have non-zero lanes, map them through.
22022 Mask.push_back(OpMask[OpIdx] * OpRatio +
22023 RootMaskedIdx % OpRatio);
22026 // See if we can recurse into the operand to combine more things.
22027 switch (Op.getOpcode()) {
22028 case X86ISD::PSHUFB:
22030 case X86ISD::PSHUFD:
22031 case X86ISD::PSHUFHW:
22032 case X86ISD::PSHUFLW:
22033 if (Op.getOperand(0).hasOneUse() &&
22034 combineX86ShufflesRecursively(Op.getOperand(0), Root, Mask, Depth + 1,
22035 HasPSHUFB, DAG, DCI, Subtarget))
22039 case X86ISD::UNPCKL:
22040 case X86ISD::UNPCKH:
22041 assert(Op.getOperand(0) == Op.getOperand(1) && "We only combine unary shuffles!");
22042 // We can't check for single use, we have to check that this shuffle is the only user.
22043 if (Op->isOnlyUserOf(Op.getOperand(0).getNode()) &&
22044 combineX86ShufflesRecursively(Op.getOperand(0), Root, Mask, Depth + 1,
22045 HasPSHUFB, DAG, DCI, Subtarget))
22050 // Minor canonicalization of the accumulated shuffle mask to make it easier
22051 // to match below. All this does is detect masks with squential pairs of
22052 // elements, and shrink them to the half-width mask. It does this in a loop
22053 // so it will reduce the size of the mask to the minimal width mask which
22054 // performs an equivalent shuffle.
22055 SmallVector<int, 16> WidenedMask;
22056 while (Mask.size() > 1 && canWidenShuffleElements(Mask, WidenedMask)) {
22057 Mask = std::move(WidenedMask);
22058 WidenedMask.clear();
22061 return combineX86ShuffleChain(Op, Root, Mask, Depth, HasPSHUFB, DAG, DCI,
22065 /// \brief Get the PSHUF-style mask from PSHUF node.
22067 /// This is a very minor wrapper around getTargetShuffleMask to easy forming v4
22068 /// PSHUF-style masks that can be reused with such instructions.
22069 static SmallVector<int, 4> getPSHUFShuffleMask(SDValue N) {
22070 MVT VT = N.getSimpleValueType();
22071 SmallVector<int, 4> Mask;
22073 bool HaveMask = getTargetShuffleMask(N.getNode(), VT, Mask, IsUnary);
22077 // If we have more than 128-bits, only the low 128-bits of shuffle mask
22078 // matter. Check that the upper masks are repeats and remove them.
22079 if (VT.getSizeInBits() > 128) {
22080 int LaneElts = 128 / VT.getScalarSizeInBits();
22082 for (int i = 1, NumLanes = VT.getSizeInBits() / 128; i < NumLanes; ++i)
22083 for (int j = 0; j < LaneElts; ++j)
22084 assert(Mask[j] == Mask[i * LaneElts + j] - (LaneElts * i) &&
22085 "Mask doesn't repeat in high 128-bit lanes!");
22087 Mask.resize(LaneElts);
22090 switch (N.getOpcode()) {
22091 case X86ISD::PSHUFD:
22093 case X86ISD::PSHUFLW:
22096 case X86ISD::PSHUFHW:
22097 Mask.erase(Mask.begin(), Mask.begin() + 4);
22098 for (int &M : Mask)
22102 llvm_unreachable("No valid shuffle instruction found!");
22106 /// \brief Search for a combinable shuffle across a chain ending in pshufd.
22108 /// We walk up the chain and look for a combinable shuffle, skipping over
22109 /// shuffles that we could hoist this shuffle's transformation past without
22110 /// altering anything.
22112 combineRedundantDWordShuffle(SDValue N, MutableArrayRef<int> Mask,
22114 TargetLowering::DAGCombinerInfo &DCI) {
22115 assert(N.getOpcode() == X86ISD::PSHUFD &&
22116 "Called with something other than an x86 128-bit half shuffle!");
22119 // Walk up a single-use chain looking for a combinable shuffle. Keep a stack
22120 // of the shuffles in the chain so that we can form a fresh chain to replace
22122 SmallVector<SDValue, 8> Chain;
22123 SDValue V = N.getOperand(0);
22124 for (; V.hasOneUse(); V = V.getOperand(0)) {
22125 switch (V.getOpcode()) {
22127 return SDValue(); // Nothing combined!
22130 // Skip bitcasts as we always know the type for the target specific
22134 case X86ISD::PSHUFD:
22135 // Found another dword shuffle.
22138 case X86ISD::PSHUFLW:
22139 // Check that the low words (being shuffled) are the identity in the
22140 // dword shuffle, and the high words are self-contained.
22141 if (Mask[0] != 0 || Mask[1] != 1 ||
22142 !(Mask[2] >= 2 && Mask[2] < 4 && Mask[3] >= 2 && Mask[3] < 4))
22145 Chain.push_back(V);
22148 case X86ISD::PSHUFHW:
22149 // Check that the high words (being shuffled) are the identity in the
22150 // dword shuffle, and the low words are self-contained.
22151 if (Mask[2] != 2 || Mask[3] != 3 ||
22152 !(Mask[0] >= 0 && Mask[0] < 2 && Mask[1] >= 0 && Mask[1] < 2))
22155 Chain.push_back(V);
22158 case X86ISD::UNPCKL:
22159 case X86ISD::UNPCKH:
22160 // For either i8 -> i16 or i16 -> i32 unpacks, we can combine a dword
22161 // shuffle into a preceding word shuffle.
22162 if (V.getSimpleValueType().getScalarType() != MVT::i8 &&
22163 V.getSimpleValueType().getScalarType() != MVT::i16)
22166 // Search for a half-shuffle which we can combine with.
22167 unsigned CombineOp =
22168 V.getOpcode() == X86ISD::UNPCKL ? X86ISD::PSHUFLW : X86ISD::PSHUFHW;
22169 if (V.getOperand(0) != V.getOperand(1) ||
22170 !V->isOnlyUserOf(V.getOperand(0).getNode()))
22172 Chain.push_back(V);
22173 V = V.getOperand(0);
22175 switch (V.getOpcode()) {
22177 return SDValue(); // Nothing to combine.
22179 case X86ISD::PSHUFLW:
22180 case X86ISD::PSHUFHW:
22181 if (V.getOpcode() == CombineOp)
22184 Chain.push_back(V);
22188 V = V.getOperand(0);
22192 } while (V.hasOneUse());
22195 // Break out of the loop if we break out of the switch.
22199 if (!V.hasOneUse())
22200 // We fell out of the loop without finding a viable combining instruction.
22203 // Merge this node's mask and our incoming mask.
22204 SmallVector<int, 4> VMask = getPSHUFShuffleMask(V);
22205 for (int &M : Mask)
22207 V = DAG.getNode(V.getOpcode(), DL, V.getValueType(), V.getOperand(0),
22208 getV4X86ShuffleImm8ForMask(Mask, DL, DAG));
22210 // Rebuild the chain around this new shuffle.
22211 while (!Chain.empty()) {
22212 SDValue W = Chain.pop_back_val();
22214 if (V.getValueType() != W.getOperand(0).getValueType())
22215 V = DAG.getBitcast(W.getOperand(0).getValueType(), V);
22217 switch (W.getOpcode()) {
22219 llvm_unreachable("Only PSHUF and UNPCK instructions get here!");
22221 case X86ISD::UNPCKL:
22222 case X86ISD::UNPCKH:
22223 V = DAG.getNode(W.getOpcode(), DL, W.getValueType(), V, V);
22226 case X86ISD::PSHUFD:
22227 case X86ISD::PSHUFLW:
22228 case X86ISD::PSHUFHW:
22229 V = DAG.getNode(W.getOpcode(), DL, W.getValueType(), V, W.getOperand(1));
22233 if (V.getValueType() != N.getValueType())
22234 V = DAG.getBitcast(N.getValueType(), V);
22236 // Return the new chain to replace N.
22240 /// \brief Search for a combinable shuffle across a chain ending in pshuflw or pshufhw.
22242 /// We walk up the chain, skipping shuffles of the other half and looking
22243 /// through shuffles which switch halves trying to find a shuffle of the same
22244 /// pair of dwords.
22245 static bool combineRedundantHalfShuffle(SDValue N, MutableArrayRef<int> Mask,
22247 TargetLowering::DAGCombinerInfo &DCI) {
22249 (N.getOpcode() == X86ISD::PSHUFLW || N.getOpcode() == X86ISD::PSHUFHW) &&
22250 "Called with something other than an x86 128-bit half shuffle!");
22252 unsigned CombineOpcode = N.getOpcode();
22254 // Walk up a single-use chain looking for a combinable shuffle.
22255 SDValue V = N.getOperand(0);
22256 for (; V.hasOneUse(); V = V.getOperand(0)) {
22257 switch (V.getOpcode()) {
22259 return false; // Nothing combined!
22262 // Skip bitcasts as we always know the type for the target specific
22266 case X86ISD::PSHUFLW:
22267 case X86ISD::PSHUFHW:
22268 if (V.getOpcode() == CombineOpcode)
22271 // Other-half shuffles are no-ops.
22274 // Break out of the loop if we break out of the switch.
22278 if (!V.hasOneUse())
22279 // We fell out of the loop without finding a viable combining instruction.
22282 // Combine away the bottom node as its shuffle will be accumulated into
22283 // a preceding shuffle.
22284 DCI.CombineTo(N.getNode(), N.getOperand(0), /*AddTo*/ true);
22286 // Record the old value.
22289 // Merge this node's mask and our incoming mask (adjusted to account for all
22290 // the pshufd instructions encountered).
22291 SmallVector<int, 4> VMask = getPSHUFShuffleMask(V);
22292 for (int &M : Mask)
22294 V = DAG.getNode(V.getOpcode(), DL, MVT::v8i16, V.getOperand(0),
22295 getV4X86ShuffleImm8ForMask(Mask, DL, DAG));
22297 // Check that the shuffles didn't cancel each other out. If not, we need to
22298 // combine to the new one.
22300 // Replace the combinable shuffle with the combined one, updating all users
22301 // so that we re-evaluate the chain here.
22302 DCI.CombineTo(Old.getNode(), V, /*AddTo*/ true);
22307 /// \brief Try to combine x86 target specific shuffles.
22308 static SDValue PerformTargetShuffleCombine(SDValue N, SelectionDAG &DAG,
22309 TargetLowering::DAGCombinerInfo &DCI,
22310 const X86Subtarget *Subtarget) {
22312 MVT VT = N.getSimpleValueType();
22313 SmallVector<int, 4> Mask;
22315 switch (N.getOpcode()) {
22316 case X86ISD::PSHUFD:
22317 case X86ISD::PSHUFLW:
22318 case X86ISD::PSHUFHW:
22319 Mask = getPSHUFShuffleMask(N);
22320 assert(Mask.size() == 4);
22326 // Nuke no-op shuffles that show up after combining.
22327 if (isNoopShuffleMask(Mask))
22328 return DCI.CombineTo(N.getNode(), N.getOperand(0), /*AddTo*/ true);
22330 // Look for simplifications involving one or two shuffle instructions.
22331 SDValue V = N.getOperand(0);
22332 switch (N.getOpcode()) {
22335 case X86ISD::PSHUFLW:
22336 case X86ISD::PSHUFHW:
22337 assert(VT.getScalarType() == MVT::i16 && "Bad word shuffle type!");
22339 if (combineRedundantHalfShuffle(N, Mask, DAG, DCI))
22340 return SDValue(); // We combined away this shuffle, so we're done.
22342 // See if this reduces to a PSHUFD which is no more expensive and can
22343 // combine with more operations. Note that it has to at least flip the
22344 // dwords as otherwise it would have been removed as a no-op.
22345 if (makeArrayRef(Mask).equals({2, 3, 0, 1})) {
22346 int DMask[] = {0, 1, 2, 3};
22347 int DOffset = N.getOpcode() == X86ISD::PSHUFLW ? 0 : 2;
22348 DMask[DOffset + 0] = DOffset + 1;
22349 DMask[DOffset + 1] = DOffset + 0;
22350 MVT DVT = MVT::getVectorVT(MVT::i32, VT.getVectorNumElements() / 2);
22351 V = DAG.getBitcast(DVT, V);
22352 DCI.AddToWorklist(V.getNode());
22353 V = DAG.getNode(X86ISD::PSHUFD, DL, DVT, V,
22354 getV4X86ShuffleImm8ForMask(DMask, DL, DAG));
22355 DCI.AddToWorklist(V.getNode());
22356 return DAG.getBitcast(VT, V);
22359 // Look for shuffle patterns which can be implemented as a single unpack.
22360 // FIXME: This doesn't handle the location of the PSHUFD generically, and
22361 // only works when we have a PSHUFD followed by two half-shuffles.
22362 if (Mask[0] == Mask[1] && Mask[2] == Mask[3] &&
22363 (V.getOpcode() == X86ISD::PSHUFLW ||
22364 V.getOpcode() == X86ISD::PSHUFHW) &&
22365 V.getOpcode() != N.getOpcode() &&
22367 SDValue D = V.getOperand(0);
22368 while (D.getOpcode() == ISD::BITCAST && D.hasOneUse())
22369 D = D.getOperand(0);
22370 if (D.getOpcode() == X86ISD::PSHUFD && D.hasOneUse()) {
22371 SmallVector<int, 4> VMask = getPSHUFShuffleMask(V);
22372 SmallVector<int, 4> DMask = getPSHUFShuffleMask(D);
22373 int NOffset = N.getOpcode() == X86ISD::PSHUFLW ? 0 : 4;
22374 int VOffset = V.getOpcode() == X86ISD::PSHUFLW ? 0 : 4;
22376 for (int i = 0; i < 4; ++i) {
22377 WordMask[i + NOffset] = Mask[i] + NOffset;
22378 WordMask[i + VOffset] = VMask[i] + VOffset;
22380 // Map the word mask through the DWord mask.
22382 for (int i = 0; i < 8; ++i)
22383 MappedMask[i] = 2 * DMask[WordMask[i] / 2] + WordMask[i] % 2;
22384 if (makeArrayRef(MappedMask).equals({0, 0, 1, 1, 2, 2, 3, 3}) ||
22385 makeArrayRef(MappedMask).equals({4, 4, 5, 5, 6, 6, 7, 7})) {
22386 // We can replace all three shuffles with an unpack.
22387 V = DAG.getBitcast(VT, D.getOperand(0));
22388 DCI.AddToWorklist(V.getNode());
22389 return DAG.getNode(MappedMask[0] == 0 ? X86ISD::UNPCKL
22398 case X86ISD::PSHUFD:
22399 if (SDValue NewN = combineRedundantDWordShuffle(N, Mask, DAG, DCI))
22408 /// \brief Try to combine a shuffle into a target-specific add-sub node.
22410 /// We combine this directly on the abstract vector shuffle nodes so it is
22411 /// easier to generically match. We also insert dummy vector shuffle nodes for
22412 /// the operands which explicitly discard the lanes which are unused by this
22413 /// operation to try to flow through the rest of the combiner the fact that
22414 /// they're unused.
22415 static SDValue combineShuffleToAddSub(SDNode *N, SelectionDAG &DAG) {
22417 EVT VT = N->getValueType(0);
22419 // We only handle target-independent shuffles.
22420 // FIXME: It would be easy and harmless to use the target shuffle mask
22421 // extraction tool to support more.
22422 if (N->getOpcode() != ISD::VECTOR_SHUFFLE)
22425 auto *SVN = cast<ShuffleVectorSDNode>(N);
22426 ArrayRef<int> Mask = SVN->getMask();
22427 SDValue V1 = N->getOperand(0);
22428 SDValue V2 = N->getOperand(1);
22430 // We require the first shuffle operand to be the SUB node, and the second to
22431 // be the ADD node.
22432 // FIXME: We should support the commuted patterns.
22433 if (V1->getOpcode() != ISD::FSUB || V2->getOpcode() != ISD::FADD)
22436 // If there are other uses of these operations we can't fold them.
22437 if (!V1->hasOneUse() || !V2->hasOneUse())
22440 // Ensure that both operations have the same operands. Note that we can
22441 // commute the FADD operands.
22442 SDValue LHS = V1->getOperand(0), RHS = V1->getOperand(1);
22443 if ((V2->getOperand(0) != LHS || V2->getOperand(1) != RHS) &&
22444 (V2->getOperand(0) != RHS || V2->getOperand(1) != LHS))
22447 // We're looking for blends between FADD and FSUB nodes. We insist on these
22448 // nodes being lined up in a specific expected pattern.
22449 if (!(isShuffleEquivalent(V1, V2, Mask, {0, 3}) ||
22450 isShuffleEquivalent(V1, V2, Mask, {0, 5, 2, 7}) ||
22451 isShuffleEquivalent(V1, V2, Mask, {0, 9, 2, 11, 4, 13, 6, 15})))
22454 // Only specific types are legal at this point, assert so we notice if and
22455 // when these change.
22456 assert((VT == MVT::v4f32 || VT == MVT::v2f64 || VT == MVT::v8f32 ||
22457 VT == MVT::v4f64) &&
22458 "Unknown vector type encountered!");
22460 return DAG.getNode(X86ISD::ADDSUB, DL, VT, LHS, RHS);
22463 /// PerformShuffleCombine - Performs several different shuffle combines.
22464 static SDValue PerformShuffleCombine(SDNode *N, SelectionDAG &DAG,
22465 TargetLowering::DAGCombinerInfo &DCI,
22466 const X86Subtarget *Subtarget) {
22468 SDValue N0 = N->getOperand(0);
22469 SDValue N1 = N->getOperand(1);
22470 EVT VT = N->getValueType(0);
22472 // Don't create instructions with illegal types after legalize types has run.
22473 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
22474 if (!DCI.isBeforeLegalize() && !TLI.isTypeLegal(VT.getVectorElementType()))
22477 // If we have legalized the vector types, look for blends of FADD and FSUB
22478 // nodes that we can fuse into an ADDSUB node.
22479 if (TLI.isTypeLegal(VT) && Subtarget->hasSSE3())
22480 if (SDValue AddSub = combineShuffleToAddSub(N, DAG))
22483 // Combine 256-bit vector shuffles. This is only profitable when in AVX mode
22484 if (Subtarget->hasFp256() && VT.is256BitVector() &&
22485 N->getOpcode() == ISD::VECTOR_SHUFFLE)
22486 return PerformShuffleCombine256(N, DAG, DCI, Subtarget);
22488 // During Type Legalization, when promoting illegal vector types,
22489 // the backend might introduce new shuffle dag nodes and bitcasts.
22491 // This code performs the following transformation:
22492 // fold: (shuffle (bitcast (BINOP A, B)), Undef, <Mask>) ->
22493 // (shuffle (BINOP (bitcast A), (bitcast B)), Undef, <Mask>)
22495 // We do this only if both the bitcast and the BINOP dag nodes have
22496 // one use. Also, perform this transformation only if the new binary
22497 // operation is legal. This is to avoid introducing dag nodes that
22498 // potentially need to be further expanded (or custom lowered) into a
22499 // less optimal sequence of dag nodes.
22500 if (!DCI.isBeforeLegalize() && DCI.isBeforeLegalizeOps() &&
22501 N1.getOpcode() == ISD::UNDEF && N0.hasOneUse() &&
22502 N0.getOpcode() == ISD::BITCAST) {
22503 SDValue BC0 = N0.getOperand(0);
22504 EVT SVT = BC0.getValueType();
22505 unsigned Opcode = BC0.getOpcode();
22506 unsigned NumElts = VT.getVectorNumElements();
22508 if (BC0.hasOneUse() && SVT.isVector() &&
22509 SVT.getVectorNumElements() * 2 == NumElts &&
22510 TLI.isOperationLegal(Opcode, VT)) {
22511 bool CanFold = false;
22523 unsigned SVTNumElts = SVT.getVectorNumElements();
22524 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
22525 for (unsigned i = 0, e = SVTNumElts; i != e && CanFold; ++i)
22526 CanFold = SVOp->getMaskElt(i) == (int)(i * 2);
22527 for (unsigned i = SVTNumElts, e = NumElts; i != e && CanFold; ++i)
22528 CanFold = SVOp->getMaskElt(i) < 0;
22531 SDValue BC00 = DAG.getBitcast(VT, BC0.getOperand(0));
22532 SDValue BC01 = DAG.getBitcast(VT, BC0.getOperand(1));
22533 SDValue NewBinOp = DAG.getNode(BC0.getOpcode(), dl, VT, BC00, BC01);
22534 return DAG.getVectorShuffle(VT, dl, NewBinOp, N1, &SVOp->getMask()[0]);
22539 // Combine a vector_shuffle that is equal to build_vector load1, load2, load3,
22540 // load4, <0, 1, 2, 3> into a 128-bit load if the load addresses are
22541 // consecutive, non-overlapping, and in the right order.
22542 SmallVector<SDValue, 16> Elts;
22543 for (unsigned i = 0, e = VT.getVectorNumElements(); i != e; ++i)
22544 Elts.push_back(getShuffleScalarElt(N, i, DAG, 0));
22546 if (SDValue LD = EltsFromConsecutiveLoads(VT, Elts, dl, DAG, true))
22549 if (isTargetShuffle(N->getOpcode())) {
22551 PerformTargetShuffleCombine(SDValue(N, 0), DAG, DCI, Subtarget);
22552 if (Shuffle.getNode())
22555 // Try recursively combining arbitrary sequences of x86 shuffle
22556 // instructions into higher-order shuffles. We do this after combining
22557 // specific PSHUF instruction sequences into their minimal form so that we
22558 // can evaluate how many specialized shuffle instructions are involved in
22559 // a particular chain.
22560 SmallVector<int, 1> NonceMask; // Just a placeholder.
22561 NonceMask.push_back(0);
22562 if (combineX86ShufflesRecursively(SDValue(N, 0), SDValue(N, 0), NonceMask,
22563 /*Depth*/ 1, /*HasPSHUFB*/ false, DAG,
22565 return SDValue(); // This routine will use CombineTo to replace N.
22571 /// XFormVExtractWithShuffleIntoLoad - Check if a vector extract from a target
22572 /// specific shuffle of a load can be folded into a single element load.
22573 /// Similar handling for VECTOR_SHUFFLE is performed by DAGCombiner, but
22574 /// shuffles have been custom lowered so we need to handle those here.
22575 static SDValue XFormVExtractWithShuffleIntoLoad(SDNode *N, SelectionDAG &DAG,
22576 TargetLowering::DAGCombinerInfo &DCI) {
22577 if (DCI.isBeforeLegalizeOps())
22580 SDValue InVec = N->getOperand(0);
22581 SDValue EltNo = N->getOperand(1);
22583 if (!isa<ConstantSDNode>(EltNo))
22586 EVT OriginalVT = InVec.getValueType();
22588 if (InVec.getOpcode() == ISD::BITCAST) {
22589 // Don't duplicate a load with other uses.
22590 if (!InVec.hasOneUse())
22592 EVT BCVT = InVec.getOperand(0).getValueType();
22593 if (!BCVT.isVector() ||
22594 BCVT.getVectorNumElements() != OriginalVT.getVectorNumElements())
22596 InVec = InVec.getOperand(0);
22599 EVT CurrentVT = InVec.getValueType();
22601 if (!isTargetShuffle(InVec.getOpcode()))
22604 // Don't duplicate a load with other uses.
22605 if (!InVec.hasOneUse())
22608 SmallVector<int, 16> ShuffleMask;
22610 if (!getTargetShuffleMask(InVec.getNode(), CurrentVT.getSimpleVT(),
22611 ShuffleMask, UnaryShuffle))
22614 // Select the input vector, guarding against out of range extract vector.
22615 unsigned NumElems = CurrentVT.getVectorNumElements();
22616 int Elt = cast<ConstantSDNode>(EltNo)->getZExtValue();
22617 int Idx = (Elt > (int)NumElems) ? -1 : ShuffleMask[Elt];
22618 SDValue LdNode = (Idx < (int)NumElems) ? InVec.getOperand(0)
22619 : InVec.getOperand(1);
22621 // If inputs to shuffle are the same for both ops, then allow 2 uses
22622 unsigned AllowedUses = InVec.getNumOperands() > 1 &&
22623 InVec.getOperand(0) == InVec.getOperand(1) ? 2 : 1;
22625 if (LdNode.getOpcode() == ISD::BITCAST) {
22626 // Don't duplicate a load with other uses.
22627 if (!LdNode.getNode()->hasNUsesOfValue(AllowedUses, 0))
22630 AllowedUses = 1; // only allow 1 load use if we have a bitcast
22631 LdNode = LdNode.getOperand(0);
22634 if (!ISD::isNormalLoad(LdNode.getNode()))
22637 LoadSDNode *LN0 = cast<LoadSDNode>(LdNode);
22639 if (!LN0 ||!LN0->hasNUsesOfValue(AllowedUses, 0) || LN0->isVolatile())
22642 EVT EltVT = N->getValueType(0);
22643 // If there's a bitcast before the shuffle, check if the load type and
22644 // alignment is valid.
22645 unsigned Align = LN0->getAlignment();
22646 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
22647 unsigned NewAlign = DAG.getDataLayout().getABITypeAlignment(
22648 EltVT.getTypeForEVT(*DAG.getContext()));
22650 if (NewAlign > Align || !TLI.isOperationLegalOrCustom(ISD::LOAD, EltVT))
22653 // All checks match so transform back to vector_shuffle so that DAG combiner
22654 // can finish the job
22657 // Create shuffle node taking into account the case that its a unary shuffle
22658 SDValue Shuffle = (UnaryShuffle) ? DAG.getUNDEF(CurrentVT)
22659 : InVec.getOperand(1);
22660 Shuffle = DAG.getVectorShuffle(CurrentVT, dl,
22661 InVec.getOperand(0), Shuffle,
22663 Shuffle = DAG.getBitcast(OriginalVT, Shuffle);
22664 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, N->getValueType(0), Shuffle,
22668 /// \brief Detect bitcasts between i32 to x86mmx low word. Since MMX types are
22669 /// special and don't usually play with other vector types, it's better to
22670 /// handle them early to be sure we emit efficient code by avoiding
22671 /// store-load conversions.
22672 static SDValue PerformBITCASTCombine(SDNode *N, SelectionDAG &DAG) {
22673 if (N->getValueType(0) != MVT::x86mmx ||
22674 N->getOperand(0)->getOpcode() != ISD::BUILD_VECTOR ||
22675 N->getOperand(0)->getValueType(0) != MVT::v2i32)
22678 SDValue V = N->getOperand(0);
22679 ConstantSDNode *C = dyn_cast<ConstantSDNode>(V.getOperand(1));
22680 if (C && C->getZExtValue() == 0 && V.getOperand(0).getValueType() == MVT::i32)
22681 return DAG.getNode(X86ISD::MMX_MOVW2D, SDLoc(V.getOperand(0)),
22682 N->getValueType(0), V.getOperand(0));
22687 /// PerformEXTRACT_VECTOR_ELTCombine - Detect vector gather/scatter index
22688 /// generation and convert it from being a bunch of shuffles and extracts
22689 /// into a somewhat faster sequence. For i686, the best sequence is apparently
22690 /// storing the value and loading scalars back, while for x64 we should
22691 /// use 64-bit extracts and shifts.
22692 static SDValue PerformEXTRACT_VECTOR_ELTCombine(SDNode *N, SelectionDAG &DAG,
22693 TargetLowering::DAGCombinerInfo &DCI) {
22694 if (SDValue NewOp = XFormVExtractWithShuffleIntoLoad(N, DAG, DCI))
22697 SDValue InputVector = N->getOperand(0);
22698 SDLoc dl(InputVector);
22699 // Detect mmx to i32 conversion through a v2i32 elt extract.
22700 if (InputVector.getOpcode() == ISD::BITCAST && InputVector.hasOneUse() &&
22701 N->getValueType(0) == MVT::i32 &&
22702 InputVector.getValueType() == MVT::v2i32) {
22704 // The bitcast source is a direct mmx result.
22705 SDValue MMXSrc = InputVector.getNode()->getOperand(0);
22706 if (MMXSrc.getValueType() == MVT::x86mmx)
22707 return DAG.getNode(X86ISD::MMX_MOVD2W, SDLoc(InputVector),
22708 N->getValueType(0),
22709 InputVector.getNode()->getOperand(0));
22711 // The mmx is indirect: (i64 extract_elt (v1i64 bitcast (x86mmx ...))).
22712 SDValue MMXSrcOp = MMXSrc.getOperand(0);
22713 if (MMXSrc.getOpcode() == ISD::EXTRACT_VECTOR_ELT && MMXSrc.hasOneUse() &&
22714 MMXSrc.getValueType() == MVT::i64 && MMXSrcOp.hasOneUse() &&
22715 MMXSrcOp.getOpcode() == ISD::BITCAST &&
22716 MMXSrcOp.getValueType() == MVT::v1i64 &&
22717 MMXSrcOp.getOperand(0).getValueType() == MVT::x86mmx)
22718 return DAG.getNode(X86ISD::MMX_MOVD2W, SDLoc(InputVector),
22719 N->getValueType(0),
22720 MMXSrcOp.getOperand(0));
22723 EVT VT = N->getValueType(0);
22725 if (VT == MVT::i1 && dyn_cast<ConstantSDNode>(N->getOperand(1)) &&
22726 InputVector.getOpcode() == ISD::BITCAST &&
22727 dyn_cast<ConstantSDNode>(InputVector.getOperand(0))) {
22728 uint64_t ExtractedElt =
22729 cast<ConstantSDNode>(N->getOperand(1))->getZExtValue();
22730 uint64_t InputValue =
22731 cast<ConstantSDNode>(InputVector.getOperand(0))->getZExtValue();
22732 uint64_t Res = (InputValue >> ExtractedElt) & 1;
22733 return DAG.getConstant(Res, dl, MVT::i1);
22735 // Only operate on vectors of 4 elements, where the alternative shuffling
22736 // gets to be more expensive.
22737 if (InputVector.getValueType() != MVT::v4i32)
22740 // Check whether every use of InputVector is an EXTRACT_VECTOR_ELT with a
22741 // single use which is a sign-extend or zero-extend, and all elements are
22743 SmallVector<SDNode *, 4> Uses;
22744 unsigned ExtractedElements = 0;
22745 for (SDNode::use_iterator UI = InputVector.getNode()->use_begin(),
22746 UE = InputVector.getNode()->use_end(); UI != UE; ++UI) {
22747 if (UI.getUse().getResNo() != InputVector.getResNo())
22750 SDNode *Extract = *UI;
22751 if (Extract->getOpcode() != ISD::EXTRACT_VECTOR_ELT)
22754 if (Extract->getValueType(0) != MVT::i32)
22756 if (!Extract->hasOneUse())
22758 if (Extract->use_begin()->getOpcode() != ISD::SIGN_EXTEND &&
22759 Extract->use_begin()->getOpcode() != ISD::ZERO_EXTEND)
22761 if (!isa<ConstantSDNode>(Extract->getOperand(1)))
22764 // Record which element was extracted.
22765 ExtractedElements |=
22766 1 << cast<ConstantSDNode>(Extract->getOperand(1))->getZExtValue();
22768 Uses.push_back(Extract);
22771 // If not all the elements were used, this may not be worthwhile.
22772 if (ExtractedElements != 15)
22775 // Ok, we've now decided to do the transformation.
22776 // If 64-bit shifts are legal, use the extract-shift sequence,
22777 // otherwise bounce the vector off the cache.
22778 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
22781 if (TLI.isOperationLegal(ISD::SRA, MVT::i64)) {
22782 SDValue Cst = DAG.getBitcast(MVT::v2i64, InputVector);
22783 auto &DL = DAG.getDataLayout();
22784 EVT VecIdxTy = DAG.getTargetLoweringInfo().getVectorIdxTy(DL);
22785 SDValue BottomHalf = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i64, Cst,
22786 DAG.getConstant(0, dl, VecIdxTy));
22787 SDValue TopHalf = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i64, Cst,
22788 DAG.getConstant(1, dl, VecIdxTy));
22790 SDValue ShAmt = DAG.getConstant(
22791 32, dl, DAG.getTargetLoweringInfo().getShiftAmountTy(MVT::i64, DL));
22792 Vals[0] = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, BottomHalf);
22793 Vals[1] = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32,
22794 DAG.getNode(ISD::SRA, dl, MVT::i64, BottomHalf, ShAmt));
22795 Vals[2] = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, TopHalf);
22796 Vals[3] = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32,
22797 DAG.getNode(ISD::SRA, dl, MVT::i64, TopHalf, ShAmt));
22799 // Store the value to a temporary stack slot.
22800 SDValue StackPtr = DAG.CreateStackTemporary(InputVector.getValueType());
22801 SDValue Ch = DAG.getStore(DAG.getEntryNode(), dl, InputVector, StackPtr,
22802 MachinePointerInfo(), false, false, 0);
22804 EVT ElementType = InputVector.getValueType().getVectorElementType();
22805 unsigned EltSize = ElementType.getSizeInBits() / 8;
22807 // Replace each use (extract) with a load of the appropriate element.
22808 for (unsigned i = 0; i < 4; ++i) {
22809 uint64_t Offset = EltSize * i;
22810 auto PtrVT = TLI.getPointerTy(DAG.getDataLayout());
22811 SDValue OffsetVal = DAG.getConstant(Offset, dl, PtrVT);
22813 SDValue ScalarAddr =
22814 DAG.getNode(ISD::ADD, dl, PtrVT, StackPtr, OffsetVal);
22816 // Load the scalar.
22817 Vals[i] = DAG.getLoad(ElementType, dl, Ch,
22818 ScalarAddr, MachinePointerInfo(),
22819 false, false, false, 0);
22824 // Replace the extracts
22825 for (SmallVectorImpl<SDNode *>::iterator UI = Uses.begin(),
22826 UE = Uses.end(); UI != UE; ++UI) {
22827 SDNode *Extract = *UI;
22829 SDValue Idx = Extract->getOperand(1);
22830 uint64_t IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
22831 DAG.ReplaceAllUsesOfValueWith(SDValue(Extract, 0), Vals[IdxVal]);
22834 // The replacement was made in place; don't return anything.
22839 transformVSELECTtoBlendVECTOR_SHUFFLE(SDNode *N, SelectionDAG &DAG,
22840 const X86Subtarget *Subtarget) {
22842 SDValue Cond = N->getOperand(0);
22843 SDValue LHS = N->getOperand(1);
22844 SDValue RHS = N->getOperand(2);
22846 if (Cond.getOpcode() == ISD::SIGN_EXTEND) {
22847 SDValue CondSrc = Cond->getOperand(0);
22848 if (CondSrc->getOpcode() == ISD::SIGN_EXTEND_INREG)
22849 Cond = CondSrc->getOperand(0);
22852 if (!ISD::isBuildVectorOfConstantSDNodes(Cond.getNode()))
22855 // A vselect where all conditions and data are constants can be optimized into
22856 // a single vector load by SelectionDAGLegalize::ExpandBUILD_VECTOR().
22857 if (ISD::isBuildVectorOfConstantSDNodes(LHS.getNode()) &&
22858 ISD::isBuildVectorOfConstantSDNodes(RHS.getNode()))
22861 unsigned MaskValue = 0;
22862 if (!BUILD_VECTORtoBlendMask(cast<BuildVectorSDNode>(Cond), MaskValue))
22865 MVT VT = N->getSimpleValueType(0);
22866 unsigned NumElems = VT.getVectorNumElements();
22867 SmallVector<int, 8> ShuffleMask(NumElems, -1);
22868 for (unsigned i = 0; i < NumElems; ++i) {
22869 // Be sure we emit undef where we can.
22870 if (Cond.getOperand(i)->getOpcode() == ISD::UNDEF)
22871 ShuffleMask[i] = -1;
22873 ShuffleMask[i] = i + NumElems * ((MaskValue >> i) & 1);
22876 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
22877 if (!TLI.isShuffleMaskLegal(ShuffleMask, VT))
22879 return DAG.getVectorShuffle(VT, dl, LHS, RHS, &ShuffleMask[0]);
22882 /// PerformSELECTCombine - Do target-specific dag combines on SELECT and VSELECT
22884 static SDValue PerformSELECTCombine(SDNode *N, SelectionDAG &DAG,
22885 TargetLowering::DAGCombinerInfo &DCI,
22886 const X86Subtarget *Subtarget) {
22888 SDValue Cond = N->getOperand(0);
22889 // Get the LHS/RHS of the select.
22890 SDValue LHS = N->getOperand(1);
22891 SDValue RHS = N->getOperand(2);
22892 EVT VT = LHS.getValueType();
22893 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
22895 // If we have SSE[12] support, try to form min/max nodes. SSE min/max
22896 // instructions match the semantics of the common C idiom x<y?x:y but not
22897 // x<=y?x:y, because of how they handle negative zero (which can be
22898 // ignored in unsafe-math mode).
22899 // We also try to create v2f32 min/max nodes, which we later widen to v4f32.
22900 if (Cond.getOpcode() == ISD::SETCC && VT.isFloatingPoint() &&
22901 VT != MVT::f80 && (TLI.isTypeLegal(VT) || VT == MVT::v2f32) &&
22902 (Subtarget->hasSSE2() ||
22903 (Subtarget->hasSSE1() && VT.getScalarType() == MVT::f32))) {
22904 ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
22906 unsigned Opcode = 0;
22907 // Check for x CC y ? x : y.
22908 if (DAG.isEqualTo(LHS, Cond.getOperand(0)) &&
22909 DAG.isEqualTo(RHS, Cond.getOperand(1))) {
22913 // Converting this to a min would handle NaNs incorrectly, and swapping
22914 // the operands would cause it to handle comparisons between positive
22915 // and negative zero incorrectly.
22916 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)) {
22917 if (!DAG.getTarget().Options.UnsafeFPMath &&
22918 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
22920 std::swap(LHS, RHS);
22922 Opcode = X86ISD::FMIN;
22925 // Converting this to a min would handle comparisons between positive
22926 // and negative zero incorrectly.
22927 if (!DAG.getTarget().Options.UnsafeFPMath &&
22928 !DAG.isKnownNeverZero(LHS) && !DAG.isKnownNeverZero(RHS))
22930 Opcode = X86ISD::FMIN;
22933 // Converting this to a min would handle both negative zeros and NaNs
22934 // incorrectly, but we can swap the operands to fix both.
22935 std::swap(LHS, RHS);
22939 Opcode = X86ISD::FMIN;
22943 // Converting this to a max would handle comparisons between positive
22944 // and negative zero incorrectly.
22945 if (!DAG.getTarget().Options.UnsafeFPMath &&
22946 !DAG.isKnownNeverZero(LHS) && !DAG.isKnownNeverZero(RHS))
22948 Opcode = X86ISD::FMAX;
22951 // Converting this to a max would handle NaNs incorrectly, and swapping
22952 // the operands would cause it to handle comparisons between positive
22953 // and negative zero incorrectly.
22954 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)) {
22955 if (!DAG.getTarget().Options.UnsafeFPMath &&
22956 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
22958 std::swap(LHS, RHS);
22960 Opcode = X86ISD::FMAX;
22963 // Converting this to a max would handle both negative zeros and NaNs
22964 // incorrectly, but we can swap the operands to fix both.
22965 std::swap(LHS, RHS);
22969 Opcode = X86ISD::FMAX;
22972 // Check for x CC y ? y : x -- a min/max with reversed arms.
22973 } else if (DAG.isEqualTo(LHS, Cond.getOperand(1)) &&
22974 DAG.isEqualTo(RHS, Cond.getOperand(0))) {
22978 // Converting this to a min would handle comparisons between positive
22979 // and negative zero incorrectly, and swapping the operands would
22980 // cause it to handle NaNs incorrectly.
22981 if (!DAG.getTarget().Options.UnsafeFPMath &&
22982 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS))) {
22983 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS))
22985 std::swap(LHS, RHS);
22987 Opcode = X86ISD::FMIN;
22990 // Converting this to a min would handle NaNs incorrectly.
22991 if (!DAG.getTarget().Options.UnsafeFPMath &&
22992 (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)))
22994 Opcode = X86ISD::FMIN;
22997 // Converting this to a min would handle both negative zeros and NaNs
22998 // incorrectly, but we can swap the operands to fix both.
22999 std::swap(LHS, RHS);
23003 Opcode = X86ISD::FMIN;
23007 // Converting this to a max would handle NaNs incorrectly.
23008 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS))
23010 Opcode = X86ISD::FMAX;
23013 // Converting this to a max would handle comparisons between positive
23014 // and negative zero incorrectly, and swapping the operands would
23015 // cause it to handle NaNs incorrectly.
23016 if (!DAG.getTarget().Options.UnsafeFPMath &&
23017 !DAG.isKnownNeverZero(LHS) && !DAG.isKnownNeverZero(RHS)) {
23018 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS))
23020 std::swap(LHS, RHS);
23022 Opcode = X86ISD::FMAX;
23025 // Converting this to a max would handle both negative zeros and NaNs
23026 // incorrectly, but we can swap the operands to fix both.
23027 std::swap(LHS, RHS);
23031 Opcode = X86ISD::FMAX;
23037 return DAG.getNode(Opcode, DL, N->getValueType(0), LHS, RHS);
23040 EVT CondVT = Cond.getValueType();
23041 if (Subtarget->hasAVX512() && VT.isVector() && CondVT.isVector() &&
23042 CondVT.getVectorElementType() == MVT::i1) {
23043 // v16i8 (select v16i1, v16i8, v16i8) does not have a proper
23044 // lowering on KNL. In this case we convert it to
23045 // v16i8 (select v16i8, v16i8, v16i8) and use AVX instruction.
23046 // The same situation for all 128 and 256-bit vectors of i8 and i16.
23047 // Since SKX these selects have a proper lowering.
23048 EVT OpVT = LHS.getValueType();
23049 if ((OpVT.is128BitVector() || OpVT.is256BitVector()) &&
23050 (OpVT.getVectorElementType() == MVT::i8 ||
23051 OpVT.getVectorElementType() == MVT::i16) &&
23052 !(Subtarget->hasBWI() && Subtarget->hasVLX())) {
23053 Cond = DAG.getNode(ISD::SIGN_EXTEND, DL, OpVT, Cond);
23054 DCI.AddToWorklist(Cond.getNode());
23055 return DAG.getNode(N->getOpcode(), DL, OpVT, Cond, LHS, RHS);
23058 // If this is a select between two integer constants, try to do some
23060 if (ConstantSDNode *TrueC = dyn_cast<ConstantSDNode>(LHS)) {
23061 if (ConstantSDNode *FalseC = dyn_cast<ConstantSDNode>(RHS))
23062 // Don't do this for crazy integer types.
23063 if (DAG.getTargetLoweringInfo().isTypeLegal(LHS.getValueType())) {
23064 // If this is efficiently invertible, canonicalize the LHSC/RHSC values
23065 // so that TrueC (the true value) is larger than FalseC.
23066 bool NeedsCondInvert = false;
23068 if (TrueC->getAPIntValue().ult(FalseC->getAPIntValue()) &&
23069 // Efficiently invertible.
23070 (Cond.getOpcode() == ISD::SETCC || // setcc -> invertible.
23071 (Cond.getOpcode() == ISD::XOR && // xor(X, C) -> invertible.
23072 isa<ConstantSDNode>(Cond.getOperand(1))))) {
23073 NeedsCondInvert = true;
23074 std::swap(TrueC, FalseC);
23077 // Optimize C ? 8 : 0 -> zext(C) << 3. Likewise for any pow2/0.
23078 if (FalseC->getAPIntValue() == 0 &&
23079 TrueC->getAPIntValue().isPowerOf2()) {
23080 if (NeedsCondInvert) // Invert the condition if needed.
23081 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
23082 DAG.getConstant(1, DL, Cond.getValueType()));
23084 // Zero extend the condition if needed.
23085 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, LHS.getValueType(), Cond);
23087 unsigned ShAmt = TrueC->getAPIntValue().logBase2();
23088 return DAG.getNode(ISD::SHL, DL, LHS.getValueType(), Cond,
23089 DAG.getConstant(ShAmt, DL, MVT::i8));
23092 // Optimize Cond ? cst+1 : cst -> zext(setcc(C)+cst.
23093 if (FalseC->getAPIntValue()+1 == TrueC->getAPIntValue()) {
23094 if (NeedsCondInvert) // Invert the condition if needed.
23095 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
23096 DAG.getConstant(1, DL, Cond.getValueType()));
23098 // Zero extend the condition if needed.
23099 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL,
23100 FalseC->getValueType(0), Cond);
23101 return DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
23102 SDValue(FalseC, 0));
23105 // Optimize cases that will turn into an LEA instruction. This requires
23106 // an i32 or i64 and an efficient multiplier (1, 2, 3, 4, 5, 8, 9).
23107 if (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i64) {
23108 uint64_t Diff = TrueC->getZExtValue()-FalseC->getZExtValue();
23109 if (N->getValueType(0) == MVT::i32) Diff = (unsigned)Diff;
23111 bool isFastMultiplier = false;
23113 switch ((unsigned char)Diff) {
23115 case 1: // result = add base, cond
23116 case 2: // result = lea base( , cond*2)
23117 case 3: // result = lea base(cond, cond*2)
23118 case 4: // result = lea base( , cond*4)
23119 case 5: // result = lea base(cond, cond*4)
23120 case 8: // result = lea base( , cond*8)
23121 case 9: // result = lea base(cond, cond*8)
23122 isFastMultiplier = true;
23127 if (isFastMultiplier) {
23128 APInt Diff = TrueC->getAPIntValue()-FalseC->getAPIntValue();
23129 if (NeedsCondInvert) // Invert the condition if needed.
23130 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
23131 DAG.getConstant(1, DL, Cond.getValueType()));
23133 // Zero extend the condition if needed.
23134 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, FalseC->getValueType(0),
23136 // Scale the condition by the difference.
23138 Cond = DAG.getNode(ISD::MUL, DL, Cond.getValueType(), Cond,
23139 DAG.getConstant(Diff, DL,
23140 Cond.getValueType()));
23142 // Add the base if non-zero.
23143 if (FalseC->getAPIntValue() != 0)
23144 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
23145 SDValue(FalseC, 0));
23152 // Canonicalize max and min:
23153 // (x > y) ? x : y -> (x >= y) ? x : y
23154 // (x < y) ? x : y -> (x <= y) ? x : y
23155 // This allows use of COND_S / COND_NS (see TranslateX86CC) which eliminates
23156 // the need for an extra compare
23157 // against zero. e.g.
23158 // (x - y) > 0 : (x - y) ? 0 -> (x - y) >= 0 : (x - y) ? 0
23160 // testl %edi, %edi
23162 // cmovgl %edi, %eax
23166 // cmovsl %eax, %edi
23167 if (N->getOpcode() == ISD::SELECT && Cond.getOpcode() == ISD::SETCC &&
23168 DAG.isEqualTo(LHS, Cond.getOperand(0)) &&
23169 DAG.isEqualTo(RHS, Cond.getOperand(1))) {
23170 ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
23175 ISD::CondCode NewCC = (CC == ISD::SETLT) ? ISD::SETLE : ISD::SETGE;
23176 Cond = DAG.getSetCC(SDLoc(Cond), Cond.getValueType(),
23177 Cond.getOperand(0), Cond.getOperand(1), NewCC);
23178 return DAG.getNode(ISD::SELECT, DL, VT, Cond, LHS, RHS);
23183 // Early exit check
23184 if (!TLI.isTypeLegal(VT))
23187 // Match VSELECTs into subs with unsigned saturation.
23188 if (N->getOpcode() == ISD::VSELECT && Cond.getOpcode() == ISD::SETCC &&
23189 // psubus is available in SSE2 and AVX2 for i8 and i16 vectors.
23190 ((Subtarget->hasSSE2() && (VT == MVT::v16i8 || VT == MVT::v8i16)) ||
23191 (Subtarget->hasAVX2() && (VT == MVT::v32i8 || VT == MVT::v16i16)))) {
23192 ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
23194 // Check if one of the arms of the VSELECT is a zero vector. If it's on the
23195 // left side invert the predicate to simplify logic below.
23197 if (ISD::isBuildVectorAllZeros(LHS.getNode())) {
23199 CC = ISD::getSetCCInverse(CC, true);
23200 } else if (ISD::isBuildVectorAllZeros(RHS.getNode())) {
23204 if (Other.getNode() && Other->getNumOperands() == 2 &&
23205 DAG.isEqualTo(Other->getOperand(0), Cond.getOperand(0))) {
23206 SDValue OpLHS = Other->getOperand(0), OpRHS = Other->getOperand(1);
23207 SDValue CondRHS = Cond->getOperand(1);
23209 // Look for a general sub with unsigned saturation first.
23210 // x >= y ? x-y : 0 --> subus x, y
23211 // x > y ? x-y : 0 --> subus x, y
23212 if ((CC == ISD::SETUGE || CC == ISD::SETUGT) &&
23213 Other->getOpcode() == ISD::SUB && DAG.isEqualTo(OpRHS, CondRHS))
23214 return DAG.getNode(X86ISD::SUBUS, DL, VT, OpLHS, OpRHS);
23216 if (auto *OpRHSBV = dyn_cast<BuildVectorSDNode>(OpRHS))
23217 if (auto *OpRHSConst = OpRHSBV->getConstantSplatNode()) {
23218 if (auto *CondRHSBV = dyn_cast<BuildVectorSDNode>(CondRHS))
23219 if (auto *CondRHSConst = CondRHSBV->getConstantSplatNode())
23220 // If the RHS is a constant we have to reverse the const
23221 // canonicalization.
23222 // x > C-1 ? x+-C : 0 --> subus x, C
23223 if (CC == ISD::SETUGT && Other->getOpcode() == ISD::ADD &&
23224 CondRHSConst->getAPIntValue() ==
23225 (-OpRHSConst->getAPIntValue() - 1))
23226 return DAG.getNode(
23227 X86ISD::SUBUS, DL, VT, OpLHS,
23228 DAG.getConstant(-OpRHSConst->getAPIntValue(), DL, VT));
23230 // Another special case: If C was a sign bit, the sub has been
23231 // canonicalized into a xor.
23232 // FIXME: Would it be better to use computeKnownBits to determine
23233 // whether it's safe to decanonicalize the xor?
23234 // x s< 0 ? x^C : 0 --> subus x, C
23235 if (CC == ISD::SETLT && Other->getOpcode() == ISD::XOR &&
23236 ISD::isBuildVectorAllZeros(CondRHS.getNode()) &&
23237 OpRHSConst->getAPIntValue().isSignBit())
23238 // Note that we have to rebuild the RHS constant here to ensure we
23239 // don't rely on particular values of undef lanes.
23240 return DAG.getNode(
23241 X86ISD::SUBUS, DL, VT, OpLHS,
23242 DAG.getConstant(OpRHSConst->getAPIntValue(), DL, VT));
23247 // Simplify vector selection if condition value type matches vselect
23249 if (N->getOpcode() == ISD::VSELECT && CondVT == VT) {
23250 assert(Cond.getValueType().isVector() &&
23251 "vector select expects a vector selector!");
23253 bool TValIsAllOnes = ISD::isBuildVectorAllOnes(LHS.getNode());
23254 bool FValIsAllZeros = ISD::isBuildVectorAllZeros(RHS.getNode());
23256 // Try invert the condition if true value is not all 1s and false value
23258 if (!TValIsAllOnes && !FValIsAllZeros &&
23259 // Check if the selector will be produced by CMPP*/PCMP*
23260 Cond.getOpcode() == ISD::SETCC &&
23261 // Check if SETCC has already been promoted
23262 TLI.getSetCCResultType(DAG.getDataLayout(), *DAG.getContext(), VT) ==
23264 bool TValIsAllZeros = ISD::isBuildVectorAllZeros(LHS.getNode());
23265 bool FValIsAllOnes = ISD::isBuildVectorAllOnes(RHS.getNode());
23267 if (TValIsAllZeros || FValIsAllOnes) {
23268 SDValue CC = Cond.getOperand(2);
23269 ISD::CondCode NewCC =
23270 ISD::getSetCCInverse(cast<CondCodeSDNode>(CC)->get(),
23271 Cond.getOperand(0).getValueType().isInteger());
23272 Cond = DAG.getSetCC(DL, CondVT, Cond.getOperand(0), Cond.getOperand(1), NewCC);
23273 std::swap(LHS, RHS);
23274 TValIsAllOnes = FValIsAllOnes;
23275 FValIsAllZeros = TValIsAllZeros;
23279 if (TValIsAllOnes || FValIsAllZeros) {
23282 if (TValIsAllOnes && FValIsAllZeros)
23284 else if (TValIsAllOnes)
23286 DAG.getNode(ISD::OR, DL, CondVT, Cond, DAG.getBitcast(CondVT, RHS));
23287 else if (FValIsAllZeros)
23288 Ret = DAG.getNode(ISD::AND, DL, CondVT, Cond,
23289 DAG.getBitcast(CondVT, LHS));
23291 return DAG.getBitcast(VT, Ret);
23295 // We should generate an X86ISD::BLENDI from a vselect if its argument
23296 // is a sign_extend_inreg of an any_extend of a BUILD_VECTOR of
23297 // constants. This specific pattern gets generated when we split a
23298 // selector for a 512 bit vector in a machine without AVX512 (but with
23299 // 256-bit vectors), during legalization:
23301 // (vselect (sign_extend (any_extend (BUILD_VECTOR)) i1) LHS RHS)
23303 // Iff we find this pattern and the build_vectors are built from
23304 // constants, we translate the vselect into a shuffle_vector that we
23305 // know will be matched by LowerVECTOR_SHUFFLEtoBlend.
23306 if ((N->getOpcode() == ISD::VSELECT ||
23307 N->getOpcode() == X86ISD::SHRUNKBLEND) &&
23308 !DCI.isBeforeLegalize() && !VT.is512BitVector()) {
23309 SDValue Shuffle = transformVSELECTtoBlendVECTOR_SHUFFLE(N, DAG, Subtarget);
23310 if (Shuffle.getNode())
23314 // If this is a *dynamic* select (non-constant condition) and we can match
23315 // this node with one of the variable blend instructions, restructure the
23316 // condition so that the blends can use the high bit of each element and use
23317 // SimplifyDemandedBits to simplify the condition operand.
23318 if (N->getOpcode() == ISD::VSELECT && DCI.isBeforeLegalizeOps() &&
23319 !DCI.isBeforeLegalize() &&
23320 !ISD::isBuildVectorOfConstantSDNodes(Cond.getNode())) {
23321 unsigned BitWidth = Cond.getValueType().getScalarType().getSizeInBits();
23323 // Don't optimize vector selects that map to mask-registers.
23327 // We can only handle the cases where VSELECT is directly legal on the
23328 // subtarget. We custom lower VSELECT nodes with constant conditions and
23329 // this makes it hard to see whether a dynamic VSELECT will correctly
23330 // lower, so we both check the operation's status and explicitly handle the
23331 // cases where a *dynamic* blend will fail even though a constant-condition
23332 // blend could be custom lowered.
23333 // FIXME: We should find a better way to handle this class of problems.
23334 // Potentially, we should combine constant-condition vselect nodes
23335 // pre-legalization into shuffles and not mark as many types as custom
23337 if (!TLI.isOperationLegalOrCustom(ISD::VSELECT, VT))
23339 // FIXME: We don't support i16-element blends currently. We could and
23340 // should support them by making *all* the bits in the condition be set
23341 // rather than just the high bit and using an i8-element blend.
23342 if (VT.getScalarType() == MVT::i16)
23344 // Dynamic blending was only available from SSE4.1 onward.
23345 if (VT.getSizeInBits() == 128 && !Subtarget->hasSSE41())
23347 // Byte blends are only available in AVX2
23348 if (VT.getSizeInBits() == 256 && VT.getScalarType() == MVT::i8 &&
23349 !Subtarget->hasAVX2())
23352 assert(BitWidth >= 8 && BitWidth <= 64 && "Invalid mask size");
23353 APInt DemandedMask = APInt::getHighBitsSet(BitWidth, 1);
23355 APInt KnownZero, KnownOne;
23356 TargetLowering::TargetLoweringOpt TLO(DAG, DCI.isBeforeLegalize(),
23357 DCI.isBeforeLegalizeOps());
23358 if (TLO.ShrinkDemandedConstant(Cond, DemandedMask) ||
23359 TLI.SimplifyDemandedBits(Cond, DemandedMask, KnownZero, KnownOne,
23361 // If we changed the computation somewhere in the DAG, this change
23362 // will affect all users of Cond.
23363 // Make sure it is fine and update all the nodes so that we do not
23364 // use the generic VSELECT anymore. Otherwise, we may perform
23365 // wrong optimizations as we messed up with the actual expectation
23366 // for the vector boolean values.
23367 if (Cond != TLO.Old) {
23368 // Check all uses of that condition operand to check whether it will be
23369 // consumed by non-BLEND instructions, which may depend on all bits are
23371 for (SDNode::use_iterator I = Cond->use_begin(), E = Cond->use_end();
23373 if (I->getOpcode() != ISD::VSELECT)
23374 // TODO: Add other opcodes eventually lowered into BLEND.
23377 // Update all the users of the condition, before committing the change,
23378 // so that the VSELECT optimizations that expect the correct vector
23379 // boolean value will not be triggered.
23380 for (SDNode::use_iterator I = Cond->use_begin(), E = Cond->use_end();
23382 DAG.ReplaceAllUsesOfValueWith(
23384 DAG.getNode(X86ISD::SHRUNKBLEND, SDLoc(*I), I->getValueType(0),
23385 Cond, I->getOperand(1), I->getOperand(2)));
23386 DCI.CommitTargetLoweringOpt(TLO);
23389 // At this point, only Cond is changed. Change the condition
23390 // just for N to keep the opportunity to optimize all other
23391 // users their own way.
23392 DAG.ReplaceAllUsesOfValueWith(
23394 DAG.getNode(X86ISD::SHRUNKBLEND, SDLoc(N), N->getValueType(0),
23395 TLO.New, N->getOperand(1), N->getOperand(2)));
23403 // Check whether a boolean test is testing a boolean value generated by
23404 // X86ISD::SETCC. If so, return the operand of that SETCC and proper condition
23407 // Simplify the following patterns:
23408 // (Op (CMP (SETCC Cond EFLAGS) 1) EQ) or
23409 // (Op (CMP (SETCC Cond EFLAGS) 0) NEQ)
23410 // to (Op EFLAGS Cond)
23412 // (Op (CMP (SETCC Cond EFLAGS) 0) EQ) or
23413 // (Op (CMP (SETCC Cond EFLAGS) 1) NEQ)
23414 // to (Op EFLAGS !Cond)
23416 // where Op could be BRCOND or CMOV.
23418 static SDValue checkBoolTestSetCCCombine(SDValue Cmp, X86::CondCode &CC) {
23419 // Quit if not CMP and SUB with its value result used.
23420 if (Cmp.getOpcode() != X86ISD::CMP &&
23421 (Cmp.getOpcode() != X86ISD::SUB || Cmp.getNode()->hasAnyUseOfValue(0)))
23424 // Quit if not used as a boolean value.
23425 if (CC != X86::COND_E && CC != X86::COND_NE)
23428 // Check CMP operands. One of them should be 0 or 1 and the other should be
23429 // an SetCC or extended from it.
23430 SDValue Op1 = Cmp.getOperand(0);
23431 SDValue Op2 = Cmp.getOperand(1);
23434 const ConstantSDNode* C = nullptr;
23435 bool needOppositeCond = (CC == X86::COND_E);
23436 bool checkAgainstTrue = false; // Is it a comparison against 1?
23438 if ((C = dyn_cast<ConstantSDNode>(Op1)))
23440 else if ((C = dyn_cast<ConstantSDNode>(Op2)))
23442 else // Quit if all operands are not constants.
23445 if (C->getZExtValue() == 1) {
23446 needOppositeCond = !needOppositeCond;
23447 checkAgainstTrue = true;
23448 } else if (C->getZExtValue() != 0)
23449 // Quit if the constant is neither 0 or 1.
23452 bool truncatedToBoolWithAnd = false;
23453 // Skip (zext $x), (trunc $x), or (and $x, 1) node.
23454 while (SetCC.getOpcode() == ISD::ZERO_EXTEND ||
23455 SetCC.getOpcode() == ISD::TRUNCATE ||
23456 SetCC.getOpcode() == ISD::AND) {
23457 if (SetCC.getOpcode() == ISD::AND) {
23459 ConstantSDNode *CS;
23460 if ((CS = dyn_cast<ConstantSDNode>(SetCC.getOperand(0))) &&
23461 CS->getZExtValue() == 1)
23463 if ((CS = dyn_cast<ConstantSDNode>(SetCC.getOperand(1))) &&
23464 CS->getZExtValue() == 1)
23468 SetCC = SetCC.getOperand(OpIdx);
23469 truncatedToBoolWithAnd = true;
23471 SetCC = SetCC.getOperand(0);
23474 switch (SetCC.getOpcode()) {
23475 case X86ISD::SETCC_CARRY:
23476 // Since SETCC_CARRY gives output based on R = CF ? ~0 : 0, it's unsafe to
23477 // simplify it if the result of SETCC_CARRY is not canonicalized to 0 or 1,
23478 // i.e. it's a comparison against true but the result of SETCC_CARRY is not
23479 // truncated to i1 using 'and'.
23480 if (checkAgainstTrue && !truncatedToBoolWithAnd)
23482 assert(X86::CondCode(SetCC.getConstantOperandVal(0)) == X86::COND_B &&
23483 "Invalid use of SETCC_CARRY!");
23485 case X86ISD::SETCC:
23486 // Set the condition code or opposite one if necessary.
23487 CC = X86::CondCode(SetCC.getConstantOperandVal(0));
23488 if (needOppositeCond)
23489 CC = X86::GetOppositeBranchCondition(CC);
23490 return SetCC.getOperand(1);
23491 case X86ISD::CMOV: {
23492 // Check whether false/true value has canonical one, i.e. 0 or 1.
23493 ConstantSDNode *FVal = dyn_cast<ConstantSDNode>(SetCC.getOperand(0));
23494 ConstantSDNode *TVal = dyn_cast<ConstantSDNode>(SetCC.getOperand(1));
23495 // Quit if true value is not a constant.
23498 // Quit if false value is not a constant.
23500 SDValue Op = SetCC.getOperand(0);
23501 // Skip 'zext' or 'trunc' node.
23502 if (Op.getOpcode() == ISD::ZERO_EXTEND ||
23503 Op.getOpcode() == ISD::TRUNCATE)
23504 Op = Op.getOperand(0);
23505 // A special case for rdrand/rdseed, where 0 is set if false cond is
23507 if ((Op.getOpcode() != X86ISD::RDRAND &&
23508 Op.getOpcode() != X86ISD::RDSEED) || Op.getResNo() != 0)
23511 // Quit if false value is not the constant 0 or 1.
23512 bool FValIsFalse = true;
23513 if (FVal && FVal->getZExtValue() != 0) {
23514 if (FVal->getZExtValue() != 1)
23516 // If FVal is 1, opposite cond is needed.
23517 needOppositeCond = !needOppositeCond;
23518 FValIsFalse = false;
23520 // Quit if TVal is not the constant opposite of FVal.
23521 if (FValIsFalse && TVal->getZExtValue() != 1)
23523 if (!FValIsFalse && TVal->getZExtValue() != 0)
23525 CC = X86::CondCode(SetCC.getConstantOperandVal(2));
23526 if (needOppositeCond)
23527 CC = X86::GetOppositeBranchCondition(CC);
23528 return SetCC.getOperand(3);
23535 /// Check whether Cond is an AND/OR of SETCCs off of the same EFLAGS.
23537 /// (X86or (X86setcc) (X86setcc))
23538 /// (X86cmp (and (X86setcc) (X86setcc)), 0)
23539 static bool checkBoolTestAndOrSetCCCombine(SDValue Cond, X86::CondCode &CC0,
23540 X86::CondCode &CC1, SDValue &Flags,
23542 if (Cond->getOpcode() == X86ISD::CMP) {
23543 ConstantSDNode *CondOp1C = dyn_cast<ConstantSDNode>(Cond->getOperand(1));
23544 if (!CondOp1C || !CondOp1C->isNullValue())
23547 Cond = Cond->getOperand(0);
23552 SDValue SetCC0, SetCC1;
23553 switch (Cond->getOpcode()) {
23554 default: return false;
23561 SetCC0 = Cond->getOperand(0);
23562 SetCC1 = Cond->getOperand(1);
23566 // Make sure we have SETCC nodes, using the same flags value.
23567 if (SetCC0.getOpcode() != X86ISD::SETCC ||
23568 SetCC1.getOpcode() != X86ISD::SETCC ||
23569 SetCC0->getOperand(1) != SetCC1->getOperand(1))
23572 CC0 = (X86::CondCode)SetCC0->getConstantOperandVal(0);
23573 CC1 = (X86::CondCode)SetCC1->getConstantOperandVal(0);
23574 Flags = SetCC0->getOperand(1);
23578 /// Optimize X86ISD::CMOV [LHS, RHS, CONDCODE (e.g. X86::COND_NE), CONDVAL]
23579 static SDValue PerformCMOVCombine(SDNode *N, SelectionDAG &DAG,
23580 TargetLowering::DAGCombinerInfo &DCI,
23581 const X86Subtarget *Subtarget) {
23584 // If the flag operand isn't dead, don't touch this CMOV.
23585 if (N->getNumValues() == 2 && !SDValue(N, 1).use_empty())
23588 SDValue FalseOp = N->getOperand(0);
23589 SDValue TrueOp = N->getOperand(1);
23590 X86::CondCode CC = (X86::CondCode)N->getConstantOperandVal(2);
23591 SDValue Cond = N->getOperand(3);
23593 if (CC == X86::COND_E || CC == X86::COND_NE) {
23594 switch (Cond.getOpcode()) {
23598 // If operand of BSR / BSF are proven never zero, then ZF cannot be set.
23599 if (DAG.isKnownNeverZero(Cond.getOperand(0)))
23600 return (CC == X86::COND_E) ? FalseOp : TrueOp;
23606 Flags = checkBoolTestSetCCCombine(Cond, CC);
23607 if (Flags.getNode() &&
23608 // Extra check as FCMOV only supports a subset of X86 cond.
23609 (FalseOp.getValueType() != MVT::f80 || hasFPCMov(CC))) {
23610 SDValue Ops[] = { FalseOp, TrueOp,
23611 DAG.getConstant(CC, DL, MVT::i8), Flags };
23612 return DAG.getNode(X86ISD::CMOV, DL, N->getVTList(), Ops);
23615 // If this is a select between two integer constants, try to do some
23616 // optimizations. Note that the operands are ordered the opposite of SELECT
23618 if (ConstantSDNode *TrueC = dyn_cast<ConstantSDNode>(TrueOp)) {
23619 if (ConstantSDNode *FalseC = dyn_cast<ConstantSDNode>(FalseOp)) {
23620 // Canonicalize the TrueC/FalseC values so that TrueC (the true value) is
23621 // larger than FalseC (the false value).
23622 if (TrueC->getAPIntValue().ult(FalseC->getAPIntValue())) {
23623 CC = X86::GetOppositeBranchCondition(CC);
23624 std::swap(TrueC, FalseC);
23625 std::swap(TrueOp, FalseOp);
23628 // Optimize C ? 8 : 0 -> zext(setcc(C)) << 3. Likewise for any pow2/0.
23629 // This is efficient for any integer data type (including i8/i16) and
23631 if (FalseC->getAPIntValue() == 0 && TrueC->getAPIntValue().isPowerOf2()) {
23632 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
23633 DAG.getConstant(CC, DL, MVT::i8), Cond);
23635 // Zero extend the condition if needed.
23636 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, TrueC->getValueType(0), Cond);
23638 unsigned ShAmt = TrueC->getAPIntValue().logBase2();
23639 Cond = DAG.getNode(ISD::SHL, DL, Cond.getValueType(), Cond,
23640 DAG.getConstant(ShAmt, DL, MVT::i8));
23641 if (N->getNumValues() == 2) // Dead flag value?
23642 return DCI.CombineTo(N, Cond, SDValue());
23646 // Optimize Cond ? cst+1 : cst -> zext(setcc(C)+cst. This is efficient
23647 // for any integer data type, including i8/i16.
23648 if (FalseC->getAPIntValue()+1 == TrueC->getAPIntValue()) {
23649 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
23650 DAG.getConstant(CC, DL, MVT::i8), Cond);
23652 // Zero extend the condition if needed.
23653 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL,
23654 FalseC->getValueType(0), Cond);
23655 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
23656 SDValue(FalseC, 0));
23658 if (N->getNumValues() == 2) // Dead flag value?
23659 return DCI.CombineTo(N, Cond, SDValue());
23663 // Optimize cases that will turn into an LEA instruction. This requires
23664 // an i32 or i64 and an efficient multiplier (1, 2, 3, 4, 5, 8, 9).
23665 if (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i64) {
23666 uint64_t Diff = TrueC->getZExtValue()-FalseC->getZExtValue();
23667 if (N->getValueType(0) == MVT::i32) Diff = (unsigned)Diff;
23669 bool isFastMultiplier = false;
23671 switch ((unsigned char)Diff) {
23673 case 1: // result = add base, cond
23674 case 2: // result = lea base( , cond*2)
23675 case 3: // result = lea base(cond, cond*2)
23676 case 4: // result = lea base( , cond*4)
23677 case 5: // result = lea base(cond, cond*4)
23678 case 8: // result = lea base( , cond*8)
23679 case 9: // result = lea base(cond, cond*8)
23680 isFastMultiplier = true;
23685 if (isFastMultiplier) {
23686 APInt Diff = TrueC->getAPIntValue()-FalseC->getAPIntValue();
23687 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
23688 DAG.getConstant(CC, DL, MVT::i8), Cond);
23689 // Zero extend the condition if needed.
23690 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, FalseC->getValueType(0),
23692 // Scale the condition by the difference.
23694 Cond = DAG.getNode(ISD::MUL, DL, Cond.getValueType(), Cond,
23695 DAG.getConstant(Diff, DL, Cond.getValueType()));
23697 // Add the base if non-zero.
23698 if (FalseC->getAPIntValue() != 0)
23699 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
23700 SDValue(FalseC, 0));
23701 if (N->getNumValues() == 2) // Dead flag value?
23702 return DCI.CombineTo(N, Cond, SDValue());
23709 // Handle these cases:
23710 // (select (x != c), e, c) -> select (x != c), e, x),
23711 // (select (x == c), c, e) -> select (x == c), x, e)
23712 // where the c is an integer constant, and the "select" is the combination
23713 // of CMOV and CMP.
23715 // The rationale for this change is that the conditional-move from a constant
23716 // needs two instructions, however, conditional-move from a register needs
23717 // only one instruction.
23719 // CAVEAT: By replacing a constant with a symbolic value, it may obscure
23720 // some instruction-combining opportunities. This opt needs to be
23721 // postponed as late as possible.
23723 if (!DCI.isBeforeLegalize() && !DCI.isBeforeLegalizeOps()) {
23724 // the DCI.xxxx conditions are provided to postpone the optimization as
23725 // late as possible.
23727 ConstantSDNode *CmpAgainst = nullptr;
23728 if ((Cond.getOpcode() == X86ISD::CMP || Cond.getOpcode() == X86ISD::SUB) &&
23729 (CmpAgainst = dyn_cast<ConstantSDNode>(Cond.getOperand(1))) &&
23730 !isa<ConstantSDNode>(Cond.getOperand(0))) {
23732 if (CC == X86::COND_NE &&
23733 CmpAgainst == dyn_cast<ConstantSDNode>(FalseOp)) {
23734 CC = X86::GetOppositeBranchCondition(CC);
23735 std::swap(TrueOp, FalseOp);
23738 if (CC == X86::COND_E &&
23739 CmpAgainst == dyn_cast<ConstantSDNode>(TrueOp)) {
23740 SDValue Ops[] = { FalseOp, Cond.getOperand(0),
23741 DAG.getConstant(CC, DL, MVT::i8), Cond };
23742 return DAG.getNode(X86ISD::CMOV, DL, N->getVTList (), Ops);
23747 // Fold and/or of setcc's to double CMOV:
23748 // (CMOV F, T, ((cc1 | cc2) != 0)) -> (CMOV (CMOV F, T, cc1), T, cc2)
23749 // (CMOV F, T, ((cc1 & cc2) != 0)) -> (CMOV (CMOV T, F, !cc1), F, !cc2)
23751 // This combine lets us generate:
23752 // cmovcc1 (jcc1 if we don't have CMOV)
23758 // cmovne (jne if we don't have CMOV)
23759 // When we can't use the CMOV instruction, it might increase branch
23761 // When we can use CMOV, or when there is no mispredict, this improves
23762 // throughput and reduces register pressure.
23764 if (CC == X86::COND_NE) {
23766 X86::CondCode CC0, CC1;
23768 if (checkBoolTestAndOrSetCCCombine(Cond, CC0, CC1, Flags, isAndSetCC)) {
23770 std::swap(FalseOp, TrueOp);
23771 CC0 = X86::GetOppositeBranchCondition(CC0);
23772 CC1 = X86::GetOppositeBranchCondition(CC1);
23775 SDValue LOps[] = {FalseOp, TrueOp, DAG.getConstant(CC0, DL, MVT::i8),
23777 SDValue LCMOV = DAG.getNode(X86ISD::CMOV, DL, N->getVTList(), LOps);
23778 SDValue Ops[] = {LCMOV, TrueOp, DAG.getConstant(CC1, DL, MVT::i8), Flags};
23779 SDValue CMOV = DAG.getNode(X86ISD::CMOV, DL, N->getVTList(), Ops);
23780 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), SDValue(CMOV.getNode(), 1));
23788 /// PerformMulCombine - Optimize a single multiply with constant into two
23789 /// in order to implement it with two cheaper instructions, e.g.
23790 /// LEA + SHL, LEA + LEA.
23791 static SDValue PerformMulCombine(SDNode *N, SelectionDAG &DAG,
23792 TargetLowering::DAGCombinerInfo &DCI) {
23793 // An imul is usually smaller than the alternative sequence.
23794 if (DAG.getMachineFunction().getFunction()->optForMinSize())
23797 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
23800 EVT VT = N->getValueType(0);
23801 if (VT != MVT::i64 && VT != MVT::i32)
23804 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1));
23807 uint64_t MulAmt = C->getZExtValue();
23808 if (isPowerOf2_64(MulAmt) || MulAmt == 3 || MulAmt == 5 || MulAmt == 9)
23811 uint64_t MulAmt1 = 0;
23812 uint64_t MulAmt2 = 0;
23813 if ((MulAmt % 9) == 0) {
23815 MulAmt2 = MulAmt / 9;
23816 } else if ((MulAmt % 5) == 0) {
23818 MulAmt2 = MulAmt / 5;
23819 } else if ((MulAmt % 3) == 0) {
23821 MulAmt2 = MulAmt / 3;
23824 (isPowerOf2_64(MulAmt2) || MulAmt2 == 3 || MulAmt2 == 5 || MulAmt2 == 9)){
23827 if (isPowerOf2_64(MulAmt2) &&
23828 !(N->hasOneUse() && N->use_begin()->getOpcode() == ISD::ADD))
23829 // If second multiplifer is pow2, issue it first. We want the multiply by
23830 // 3, 5, or 9 to be folded into the addressing mode unless the lone use
23832 std::swap(MulAmt1, MulAmt2);
23835 if (isPowerOf2_64(MulAmt1))
23836 NewMul = DAG.getNode(ISD::SHL, DL, VT, N->getOperand(0),
23837 DAG.getConstant(Log2_64(MulAmt1), DL, MVT::i8));
23839 NewMul = DAG.getNode(X86ISD::MUL_IMM, DL, VT, N->getOperand(0),
23840 DAG.getConstant(MulAmt1, DL, VT));
23842 if (isPowerOf2_64(MulAmt2))
23843 NewMul = DAG.getNode(ISD::SHL, DL, VT, NewMul,
23844 DAG.getConstant(Log2_64(MulAmt2), DL, MVT::i8));
23846 NewMul = DAG.getNode(X86ISD::MUL_IMM, DL, VT, NewMul,
23847 DAG.getConstant(MulAmt2, DL, VT));
23849 // Do not add new nodes to DAG combiner worklist.
23850 DCI.CombineTo(N, NewMul, false);
23855 static SDValue PerformSHLCombine(SDNode *N, SelectionDAG &DAG) {
23856 SDValue N0 = N->getOperand(0);
23857 SDValue N1 = N->getOperand(1);
23858 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
23859 EVT VT = N0.getValueType();
23861 // fold (shl (and (setcc_c), c1), c2) -> (and setcc_c, (c1 << c2))
23862 // since the result of setcc_c is all zero's or all ones.
23863 if (VT.isInteger() && !VT.isVector() &&
23864 N1C && N0.getOpcode() == ISD::AND &&
23865 N0.getOperand(1).getOpcode() == ISD::Constant) {
23866 SDValue N00 = N0.getOperand(0);
23867 APInt Mask = cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
23868 APInt ShAmt = N1C->getAPIntValue();
23869 Mask = Mask.shl(ShAmt);
23870 bool MaskOK = false;
23871 // We can handle cases concerning bit-widening nodes containing setcc_c if
23872 // we carefully interrogate the mask to make sure we are semantics
23874 // The transform is not safe if the result of C1 << C2 exceeds the bitwidth
23875 // of the underlying setcc_c operation if the setcc_c was zero extended.
23876 // Consider the following example:
23877 // zext(setcc_c) -> i32 0x0000FFFF
23878 // c1 -> i32 0x0000FFFF
23879 // c2 -> i32 0x00000001
23880 // (shl (and (setcc_c), c1), c2) -> i32 0x0001FFFE
23881 // (and setcc_c, (c1 << c2)) -> i32 0x0000FFFE
23882 if (N00.getOpcode() == X86ISD::SETCC_CARRY) {
23884 } else if (N00.getOpcode() == ISD::SIGN_EXTEND &&
23885 N00.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY) {
23887 } else if ((N00.getOpcode() == ISD::ZERO_EXTEND ||
23888 N00.getOpcode() == ISD::ANY_EXTEND) &&
23889 N00.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY) {
23890 MaskOK = Mask.isIntN(N00.getOperand(0).getValueSizeInBits());
23892 if (MaskOK && Mask != 0) {
23894 return DAG.getNode(ISD::AND, DL, VT, N00, DAG.getConstant(Mask, DL, VT));
23898 // Hardware support for vector shifts is sparse which makes us scalarize the
23899 // vector operations in many cases. Also, on sandybridge ADD is faster than
23901 // (shl V, 1) -> add V,V
23902 if (auto *N1BV = dyn_cast<BuildVectorSDNode>(N1))
23903 if (auto *N1SplatC = N1BV->getConstantSplatNode()) {
23904 assert(N0.getValueType().isVector() && "Invalid vector shift type");
23905 // We shift all of the values by one. In many cases we do not have
23906 // hardware support for this operation. This is better expressed as an ADD
23908 if (N1SplatC->getAPIntValue() == 1)
23909 return DAG.getNode(ISD::ADD, SDLoc(N), VT, N0, N0);
23915 /// \brief Returns a vector of 0s if the node in input is a vector logical
23916 /// shift by a constant amount which is known to be bigger than or equal
23917 /// to the vector element size in bits.
23918 static SDValue performShiftToAllZeros(SDNode *N, SelectionDAG &DAG,
23919 const X86Subtarget *Subtarget) {
23920 EVT VT = N->getValueType(0);
23922 if (VT != MVT::v2i64 && VT != MVT::v4i32 && VT != MVT::v8i16 &&
23923 (!Subtarget->hasInt256() ||
23924 (VT != MVT::v4i64 && VT != MVT::v8i32 && VT != MVT::v16i16)))
23927 SDValue Amt = N->getOperand(1);
23929 if (auto *AmtBV = dyn_cast<BuildVectorSDNode>(Amt))
23930 if (auto *AmtSplat = AmtBV->getConstantSplatNode()) {
23931 APInt ShiftAmt = AmtSplat->getAPIntValue();
23932 unsigned MaxAmount = VT.getVectorElementType().getSizeInBits();
23934 // SSE2/AVX2 logical shifts always return a vector of 0s
23935 // if the shift amount is bigger than or equal to
23936 // the element size. The constant shift amount will be
23937 // encoded as a 8-bit immediate.
23938 if (ShiftAmt.trunc(8).uge(MaxAmount))
23939 return getZeroVector(VT, Subtarget, DAG, DL);
23945 /// PerformShiftCombine - Combine shifts.
23946 static SDValue PerformShiftCombine(SDNode* N, SelectionDAG &DAG,
23947 TargetLowering::DAGCombinerInfo &DCI,
23948 const X86Subtarget *Subtarget) {
23949 if (N->getOpcode() == ISD::SHL)
23950 if (SDValue V = PerformSHLCombine(N, DAG))
23953 // Try to fold this logical shift into a zero vector.
23954 if (N->getOpcode() != ISD::SRA)
23955 if (SDValue V = performShiftToAllZeros(N, DAG, Subtarget))
23961 // CMPEQCombine - Recognize the distinctive (AND (setcc ...) (setcc ..))
23962 // where both setccs reference the same FP CMP, and rewrite for CMPEQSS
23963 // and friends. Likewise for OR -> CMPNEQSS.
23964 static SDValue CMPEQCombine(SDNode *N, SelectionDAG &DAG,
23965 TargetLowering::DAGCombinerInfo &DCI,
23966 const X86Subtarget *Subtarget) {
23969 // SSE1 supports CMP{eq|ne}SS, and SSE2 added CMP{eq|ne}SD, but
23970 // we're requiring SSE2 for both.
23971 if (Subtarget->hasSSE2() && isAndOrOfSetCCs(SDValue(N, 0U), opcode)) {
23972 SDValue N0 = N->getOperand(0);
23973 SDValue N1 = N->getOperand(1);
23974 SDValue CMP0 = N0->getOperand(1);
23975 SDValue CMP1 = N1->getOperand(1);
23978 // The SETCCs should both refer to the same CMP.
23979 if (CMP0.getOpcode() != X86ISD::CMP || CMP0 != CMP1)
23982 SDValue CMP00 = CMP0->getOperand(0);
23983 SDValue CMP01 = CMP0->getOperand(1);
23984 EVT VT = CMP00.getValueType();
23986 if (VT == MVT::f32 || VT == MVT::f64) {
23987 bool ExpectingFlags = false;
23988 // Check for any users that want flags:
23989 for (SDNode::use_iterator UI = N->use_begin(), UE = N->use_end();
23990 !ExpectingFlags && UI != UE; ++UI)
23991 switch (UI->getOpcode()) {
23996 ExpectingFlags = true;
23998 case ISD::CopyToReg:
23999 case ISD::SIGN_EXTEND:
24000 case ISD::ZERO_EXTEND:
24001 case ISD::ANY_EXTEND:
24005 if (!ExpectingFlags) {
24006 enum X86::CondCode cc0 = (enum X86::CondCode)N0.getConstantOperandVal(0);
24007 enum X86::CondCode cc1 = (enum X86::CondCode)N1.getConstantOperandVal(0);
24009 if (cc1 == X86::COND_E || cc1 == X86::COND_NE) {
24010 X86::CondCode tmp = cc0;
24015 if ((cc0 == X86::COND_E && cc1 == X86::COND_NP) ||
24016 (cc0 == X86::COND_NE && cc1 == X86::COND_P)) {
24017 // FIXME: need symbolic constants for these magic numbers.
24018 // See X86ATTInstPrinter.cpp:printSSECC().
24019 unsigned x86cc = (cc0 == X86::COND_E) ? 0 : 4;
24020 if (Subtarget->hasAVX512()) {
24021 SDValue FSetCC = DAG.getNode(X86ISD::FSETCC, DL, MVT::i1, CMP00,
24023 DAG.getConstant(x86cc, DL, MVT::i8));
24024 if (N->getValueType(0) != MVT::i1)
24025 return DAG.getNode(ISD::ZERO_EXTEND, DL, N->getValueType(0),
24029 SDValue OnesOrZeroesF = DAG.getNode(X86ISD::FSETCC, DL,
24030 CMP00.getValueType(), CMP00, CMP01,
24031 DAG.getConstant(x86cc, DL,
24034 bool is64BitFP = (CMP00.getValueType() == MVT::f64);
24035 MVT IntVT = is64BitFP ? MVT::i64 : MVT::i32;
24037 if (is64BitFP && !Subtarget->is64Bit()) {
24038 // On a 32-bit target, we cannot bitcast the 64-bit float to a
24039 // 64-bit integer, since that's not a legal type. Since
24040 // OnesOrZeroesF is all ones of all zeroes, we don't need all the
24041 // bits, but can do this little dance to extract the lowest 32 bits
24042 // and work with those going forward.
24043 SDValue Vector64 = DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, MVT::v2f64,
24045 SDValue Vector32 = DAG.getBitcast(MVT::v4f32, Vector64);
24046 OnesOrZeroesF = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::f32,
24047 Vector32, DAG.getIntPtrConstant(0, DL));
24051 SDValue OnesOrZeroesI = DAG.getBitcast(IntVT, OnesOrZeroesF);
24052 SDValue ANDed = DAG.getNode(ISD::AND, DL, IntVT, OnesOrZeroesI,
24053 DAG.getConstant(1, DL, IntVT));
24054 SDValue OneBitOfTruth = DAG.getNode(ISD::TRUNCATE, DL, MVT::i8,
24056 return OneBitOfTruth;
24064 /// CanFoldXORWithAllOnes - Test whether the XOR operand is a AllOnes vector
24065 /// so it can be folded inside ANDNP.
24066 static bool CanFoldXORWithAllOnes(const SDNode *N) {
24067 EVT VT = N->getValueType(0);
24069 // Match direct AllOnes for 128 and 256-bit vectors
24070 if (ISD::isBuildVectorAllOnes(N))
24073 // Look through a bit convert.
24074 if (N->getOpcode() == ISD::BITCAST)
24075 N = N->getOperand(0).getNode();
24077 // Sometimes the operand may come from a insert_subvector building a 256-bit
24079 if (VT.is256BitVector() &&
24080 N->getOpcode() == ISD::INSERT_SUBVECTOR) {
24081 SDValue V1 = N->getOperand(0);
24082 SDValue V2 = N->getOperand(1);
24084 if (V1.getOpcode() == ISD::INSERT_SUBVECTOR &&
24085 V1.getOperand(0).getOpcode() == ISD::UNDEF &&
24086 ISD::isBuildVectorAllOnes(V1.getOperand(1).getNode()) &&
24087 ISD::isBuildVectorAllOnes(V2.getNode()))
24094 // On AVX/AVX2 the type v8i1 is legalized to v8i16, which is an XMM sized
24095 // register. In most cases we actually compare or select YMM-sized registers
24096 // and mixing the two types creates horrible code. This method optimizes
24097 // some of the transition sequences.
24098 static SDValue WidenMaskArithmetic(SDNode *N, SelectionDAG &DAG,
24099 TargetLowering::DAGCombinerInfo &DCI,
24100 const X86Subtarget *Subtarget) {
24101 EVT VT = N->getValueType(0);
24102 if (!VT.is256BitVector())
24105 assert((N->getOpcode() == ISD::ANY_EXTEND ||
24106 N->getOpcode() == ISD::ZERO_EXTEND ||
24107 N->getOpcode() == ISD::SIGN_EXTEND) && "Invalid Node");
24109 SDValue Narrow = N->getOperand(0);
24110 EVT NarrowVT = Narrow->getValueType(0);
24111 if (!NarrowVT.is128BitVector())
24114 if (Narrow->getOpcode() != ISD::XOR &&
24115 Narrow->getOpcode() != ISD::AND &&
24116 Narrow->getOpcode() != ISD::OR)
24119 SDValue N0 = Narrow->getOperand(0);
24120 SDValue N1 = Narrow->getOperand(1);
24123 // The Left side has to be a trunc.
24124 if (N0.getOpcode() != ISD::TRUNCATE)
24127 // The type of the truncated inputs.
24128 EVT WideVT = N0->getOperand(0)->getValueType(0);
24132 // The right side has to be a 'trunc' or a constant vector.
24133 bool RHSTrunc = N1.getOpcode() == ISD::TRUNCATE;
24134 ConstantSDNode *RHSConstSplat = nullptr;
24135 if (auto *RHSBV = dyn_cast<BuildVectorSDNode>(N1))
24136 RHSConstSplat = RHSBV->getConstantSplatNode();
24137 if (!RHSTrunc && !RHSConstSplat)
24140 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
24142 if (!TLI.isOperationLegalOrPromote(Narrow->getOpcode(), WideVT))
24145 // Set N0 and N1 to hold the inputs to the new wide operation.
24146 N0 = N0->getOperand(0);
24147 if (RHSConstSplat) {
24148 N1 = DAG.getNode(ISD::ZERO_EXTEND, DL, WideVT.getScalarType(),
24149 SDValue(RHSConstSplat, 0));
24150 SmallVector<SDValue, 8> C(WideVT.getVectorNumElements(), N1);
24151 N1 = DAG.getNode(ISD::BUILD_VECTOR, DL, WideVT, C);
24152 } else if (RHSTrunc) {
24153 N1 = N1->getOperand(0);
24156 // Generate the wide operation.
24157 SDValue Op = DAG.getNode(Narrow->getOpcode(), DL, WideVT, N0, N1);
24158 unsigned Opcode = N->getOpcode();
24160 case ISD::ANY_EXTEND:
24162 case ISD::ZERO_EXTEND: {
24163 unsigned InBits = NarrowVT.getScalarType().getSizeInBits();
24164 APInt Mask = APInt::getAllOnesValue(InBits);
24165 Mask = Mask.zext(VT.getScalarType().getSizeInBits());
24166 return DAG.getNode(ISD::AND, DL, VT,
24167 Op, DAG.getConstant(Mask, DL, VT));
24169 case ISD::SIGN_EXTEND:
24170 return DAG.getNode(ISD::SIGN_EXTEND_INREG, DL, VT,
24171 Op, DAG.getValueType(NarrowVT));
24173 llvm_unreachable("Unexpected opcode");
24177 static SDValue VectorZextCombine(SDNode *N, SelectionDAG &DAG,
24178 TargetLowering::DAGCombinerInfo &DCI,
24179 const X86Subtarget *Subtarget) {
24180 SDValue N0 = N->getOperand(0);
24181 SDValue N1 = N->getOperand(1);
24184 // A vector zext_in_reg may be represented as a shuffle,
24185 // feeding into a bitcast (this represents anyext) feeding into
24186 // an and with a mask.
24187 // We'd like to try to combine that into a shuffle with zero
24188 // plus a bitcast, removing the and.
24189 if (N0.getOpcode() != ISD::BITCAST ||
24190 N0.getOperand(0).getOpcode() != ISD::VECTOR_SHUFFLE)
24193 // The other side of the AND should be a splat of 2^C, where C
24194 // is the number of bits in the source type.
24195 if (N1.getOpcode() == ISD::BITCAST)
24196 N1 = N1.getOperand(0);
24197 if (N1.getOpcode() != ISD::BUILD_VECTOR)
24199 BuildVectorSDNode *Vector = cast<BuildVectorSDNode>(N1);
24201 ShuffleVectorSDNode *Shuffle = cast<ShuffleVectorSDNode>(N0.getOperand(0));
24202 EVT SrcType = Shuffle->getValueType(0);
24204 // We expect a single-source shuffle
24205 if (Shuffle->getOperand(1)->getOpcode() != ISD::UNDEF)
24208 unsigned SrcSize = SrcType.getScalarSizeInBits();
24210 APInt SplatValue, SplatUndef;
24211 unsigned SplatBitSize;
24213 if (!Vector->isConstantSplat(SplatValue, SplatUndef,
24214 SplatBitSize, HasAnyUndefs))
24217 unsigned ResSize = N1.getValueType().getScalarSizeInBits();
24218 // Make sure the splat matches the mask we expect
24219 if (SplatBitSize > ResSize ||
24220 (SplatValue + 1).exactLogBase2() != (int)SrcSize)
24223 // Make sure the input and output size make sense
24224 if (SrcSize >= ResSize || ResSize % SrcSize)
24227 // We expect a shuffle of the form <0, u, u, u, 1, u, u, u...>
24228 // The number of u's between each two values depends on the ratio between
24229 // the source and dest type.
24230 unsigned ZextRatio = ResSize / SrcSize;
24231 bool IsZext = true;
24232 for (unsigned i = 0; i < SrcType.getVectorNumElements(); ++i) {
24233 if (i % ZextRatio) {
24234 if (Shuffle->getMaskElt(i) > 0) {
24240 if (Shuffle->getMaskElt(i) != (int)(i / ZextRatio)) {
24241 // Expected element number
24251 // Ok, perform the transformation - replace the shuffle with
24252 // a shuffle of the form <0, k, k, k, 1, k, k, k> with zero
24253 // (instead of undef) where the k elements come from the zero vector.
24254 SmallVector<int, 8> Mask;
24255 unsigned NumElems = SrcType.getVectorNumElements();
24256 for (unsigned i = 0; i < NumElems; ++i)
24258 Mask.push_back(NumElems);
24260 Mask.push_back(i / ZextRatio);
24262 SDValue NewShuffle = DAG.getVectorShuffle(Shuffle->getValueType(0), DL,
24263 Shuffle->getOperand(0), DAG.getConstant(0, DL, SrcType), Mask);
24264 return DAG.getBitcast(N0.getValueType(), NewShuffle);
24267 static SDValue PerformAndCombine(SDNode *N, SelectionDAG &DAG,
24268 TargetLowering::DAGCombinerInfo &DCI,
24269 const X86Subtarget *Subtarget) {
24270 if (DCI.isBeforeLegalizeOps())
24273 if (SDValue Zext = VectorZextCombine(N, DAG, DCI, Subtarget))
24276 if (SDValue R = CMPEQCombine(N, DAG, DCI, Subtarget))
24279 EVT VT = N->getValueType(0);
24280 SDValue N0 = N->getOperand(0);
24281 SDValue N1 = N->getOperand(1);
24284 // Create BEXTR instructions
24285 // BEXTR is ((X >> imm) & (2**size-1))
24286 if (VT == MVT::i32 || VT == MVT::i64) {
24287 // Check for BEXTR.
24288 if ((Subtarget->hasBMI() || Subtarget->hasTBM()) &&
24289 (N0.getOpcode() == ISD::SRA || N0.getOpcode() == ISD::SRL)) {
24290 ConstantSDNode *MaskNode = dyn_cast<ConstantSDNode>(N1);
24291 ConstantSDNode *ShiftNode = dyn_cast<ConstantSDNode>(N0.getOperand(1));
24292 if (MaskNode && ShiftNode) {
24293 uint64_t Mask = MaskNode->getZExtValue();
24294 uint64_t Shift = ShiftNode->getZExtValue();
24295 if (isMask_64(Mask)) {
24296 uint64_t MaskSize = countPopulation(Mask);
24297 if (Shift + MaskSize <= VT.getSizeInBits())
24298 return DAG.getNode(X86ISD::BEXTR, DL, VT, N0.getOperand(0),
24299 DAG.getConstant(Shift | (MaskSize << 8), DL,
24308 // Want to form ANDNP nodes:
24309 // 1) In the hopes of then easily combining them with OR and AND nodes
24310 // to form PBLEND/PSIGN.
24311 // 2) To match ANDN packed intrinsics
24312 if (VT != MVT::v2i64 && VT != MVT::v4i64)
24315 // Check LHS for vnot
24316 if (N0.getOpcode() == ISD::XOR &&
24317 //ISD::isBuildVectorAllOnes(N0.getOperand(1).getNode()))
24318 CanFoldXORWithAllOnes(N0.getOperand(1).getNode()))
24319 return DAG.getNode(X86ISD::ANDNP, DL, VT, N0.getOperand(0), N1);
24321 // Check RHS for vnot
24322 if (N1.getOpcode() == ISD::XOR &&
24323 //ISD::isBuildVectorAllOnes(N1.getOperand(1).getNode()))
24324 CanFoldXORWithAllOnes(N1.getOperand(1).getNode()))
24325 return DAG.getNode(X86ISD::ANDNP, DL, VT, N1.getOperand(0), N0);
24330 static SDValue PerformOrCombine(SDNode *N, SelectionDAG &DAG,
24331 TargetLowering::DAGCombinerInfo &DCI,
24332 const X86Subtarget *Subtarget) {
24333 if (DCI.isBeforeLegalizeOps())
24336 if (SDValue R = CMPEQCombine(N, DAG, DCI, Subtarget))
24339 SDValue N0 = N->getOperand(0);
24340 SDValue N1 = N->getOperand(1);
24341 EVT VT = N->getValueType(0);
24343 // look for psign/blend
24344 if (VT == MVT::v2i64 || VT == MVT::v4i64) {
24345 if (!Subtarget->hasSSSE3() ||
24346 (VT == MVT::v4i64 && !Subtarget->hasInt256()))
24349 // Canonicalize pandn to RHS
24350 if (N0.getOpcode() == X86ISD::ANDNP)
24352 // or (and (m, y), (pandn m, x))
24353 if (N0.getOpcode() == ISD::AND && N1.getOpcode() == X86ISD::ANDNP) {
24354 SDValue Mask = N1.getOperand(0);
24355 SDValue X = N1.getOperand(1);
24357 if (N0.getOperand(0) == Mask)
24358 Y = N0.getOperand(1);
24359 if (N0.getOperand(1) == Mask)
24360 Y = N0.getOperand(0);
24362 // Check to see if the mask appeared in both the AND and ANDNP and
24366 // Validate that X, Y, and Mask are BIT_CONVERTS, and see through them.
24367 // Look through mask bitcast.
24368 if (Mask.getOpcode() == ISD::BITCAST)
24369 Mask = Mask.getOperand(0);
24370 if (X.getOpcode() == ISD::BITCAST)
24371 X = X.getOperand(0);
24372 if (Y.getOpcode() == ISD::BITCAST)
24373 Y = Y.getOperand(0);
24375 EVT MaskVT = Mask.getValueType();
24377 // Validate that the Mask operand is a vector sra node.
24378 // FIXME: what to do for bytes, since there is a psignb/pblendvb, but
24379 // there is no psrai.b
24380 unsigned EltBits = MaskVT.getVectorElementType().getSizeInBits();
24381 unsigned SraAmt = ~0;
24382 if (Mask.getOpcode() == ISD::SRA) {
24383 if (auto *AmtBV = dyn_cast<BuildVectorSDNode>(Mask.getOperand(1)))
24384 if (auto *AmtConst = AmtBV->getConstantSplatNode())
24385 SraAmt = AmtConst->getZExtValue();
24386 } else if (Mask.getOpcode() == X86ISD::VSRAI) {
24387 SDValue SraC = Mask.getOperand(1);
24388 SraAmt = cast<ConstantSDNode>(SraC)->getZExtValue();
24390 if ((SraAmt + 1) != EltBits)
24395 // Now we know we at least have a plendvb with the mask val. See if
24396 // we can form a psignb/w/d.
24397 // psign = x.type == y.type == mask.type && y = sub(0, x);
24398 if (Y.getOpcode() == ISD::SUB && Y.getOperand(1) == X &&
24399 ISD::isBuildVectorAllZeros(Y.getOperand(0).getNode()) &&
24400 X.getValueType() == MaskVT && Y.getValueType() == MaskVT) {
24401 assert((EltBits == 8 || EltBits == 16 || EltBits == 32) &&
24402 "Unsupported VT for PSIGN");
24403 Mask = DAG.getNode(X86ISD::PSIGN, DL, MaskVT, X, Mask.getOperand(0));
24404 return DAG.getBitcast(VT, Mask);
24406 // PBLENDVB only available on SSE 4.1
24407 if (!Subtarget->hasSSE41())
24410 EVT BlendVT = (VT == MVT::v4i64) ? MVT::v32i8 : MVT::v16i8;
24412 X = DAG.getBitcast(BlendVT, X);
24413 Y = DAG.getBitcast(BlendVT, Y);
24414 Mask = DAG.getBitcast(BlendVT, Mask);
24415 Mask = DAG.getNode(ISD::VSELECT, DL, BlendVT, Mask, Y, X);
24416 return DAG.getBitcast(VT, Mask);
24420 if (VT != MVT::i16 && VT != MVT::i32 && VT != MVT::i64)
24423 // fold (or (x << c) | (y >> (64 - c))) ==> (shld64 x, y, c)
24424 bool OptForSize = DAG.getMachineFunction().getFunction()->optForSize();
24426 // SHLD/SHRD instructions have lower register pressure, but on some
24427 // platforms they have higher latency than the equivalent
24428 // series of shifts/or that would otherwise be generated.
24429 // Don't fold (or (x << c) | (y >> (64 - c))) if SHLD/SHRD instructions
24430 // have higher latencies and we are not optimizing for size.
24431 if (!OptForSize && Subtarget->isSHLDSlow())
24434 if (N0.getOpcode() == ISD::SRL && N1.getOpcode() == ISD::SHL)
24436 if (N0.getOpcode() != ISD::SHL || N1.getOpcode() != ISD::SRL)
24438 if (!N0.hasOneUse() || !N1.hasOneUse())
24441 SDValue ShAmt0 = N0.getOperand(1);
24442 if (ShAmt0.getValueType() != MVT::i8)
24444 SDValue ShAmt1 = N1.getOperand(1);
24445 if (ShAmt1.getValueType() != MVT::i8)
24447 if (ShAmt0.getOpcode() == ISD::TRUNCATE)
24448 ShAmt0 = ShAmt0.getOperand(0);
24449 if (ShAmt1.getOpcode() == ISD::TRUNCATE)
24450 ShAmt1 = ShAmt1.getOperand(0);
24453 unsigned Opc = X86ISD::SHLD;
24454 SDValue Op0 = N0.getOperand(0);
24455 SDValue Op1 = N1.getOperand(0);
24456 if (ShAmt0.getOpcode() == ISD::SUB) {
24457 Opc = X86ISD::SHRD;
24458 std::swap(Op0, Op1);
24459 std::swap(ShAmt0, ShAmt1);
24462 unsigned Bits = VT.getSizeInBits();
24463 if (ShAmt1.getOpcode() == ISD::SUB) {
24464 SDValue Sum = ShAmt1.getOperand(0);
24465 if (ConstantSDNode *SumC = dyn_cast<ConstantSDNode>(Sum)) {
24466 SDValue ShAmt1Op1 = ShAmt1.getOperand(1);
24467 if (ShAmt1Op1.getNode()->getOpcode() == ISD::TRUNCATE)
24468 ShAmt1Op1 = ShAmt1Op1.getOperand(0);
24469 if (SumC->getSExtValue() == Bits && ShAmt1Op1 == ShAmt0)
24470 return DAG.getNode(Opc, DL, VT,
24472 DAG.getNode(ISD::TRUNCATE, DL,
24475 } else if (ConstantSDNode *ShAmt1C = dyn_cast<ConstantSDNode>(ShAmt1)) {
24476 ConstantSDNode *ShAmt0C = dyn_cast<ConstantSDNode>(ShAmt0);
24478 ShAmt0C->getSExtValue() + ShAmt1C->getSExtValue() == Bits)
24479 return DAG.getNode(Opc, DL, VT,
24480 N0.getOperand(0), N1.getOperand(0),
24481 DAG.getNode(ISD::TRUNCATE, DL,
24488 // Generate NEG and CMOV for integer abs.
24489 static SDValue performIntegerAbsCombine(SDNode *N, SelectionDAG &DAG) {
24490 EVT VT = N->getValueType(0);
24492 // Since X86 does not have CMOV for 8-bit integer, we don't convert
24493 // 8-bit integer abs to NEG and CMOV.
24494 if (VT.isInteger() && VT.getSizeInBits() == 8)
24497 SDValue N0 = N->getOperand(0);
24498 SDValue N1 = N->getOperand(1);
24501 // Check pattern of XOR(ADD(X,Y), Y) where Y is SRA(X, size(X)-1)
24502 // and change it to SUB and CMOV.
24503 if (VT.isInteger() && N->getOpcode() == ISD::XOR &&
24504 N0.getOpcode() == ISD::ADD &&
24505 N0.getOperand(1) == N1 &&
24506 N1.getOpcode() == ISD::SRA &&
24507 N1.getOperand(0) == N0.getOperand(0))
24508 if (ConstantSDNode *Y1C = dyn_cast<ConstantSDNode>(N1.getOperand(1)))
24509 if (Y1C->getAPIntValue() == VT.getSizeInBits()-1) {
24510 // Generate SUB & CMOV.
24511 SDValue Neg = DAG.getNode(X86ISD::SUB, DL, DAG.getVTList(VT, MVT::i32),
24512 DAG.getConstant(0, DL, VT), N0.getOperand(0));
24514 SDValue Ops[] = { N0.getOperand(0), Neg,
24515 DAG.getConstant(X86::COND_GE, DL, MVT::i8),
24516 SDValue(Neg.getNode(), 1) };
24517 return DAG.getNode(X86ISD::CMOV, DL, DAG.getVTList(VT, MVT::Glue), Ops);
24522 // Try to turn tests against the signbit in the form of:
24523 // XOR(TRUNCATE(SRL(X, size(X)-1)), 1)
24526 static SDValue foldXorTruncShiftIntoCmp(SDNode *N, SelectionDAG &DAG) {
24527 // This is only worth doing if the output type is i8.
24528 if (N->getValueType(0) != MVT::i8)
24531 SDValue N0 = N->getOperand(0);
24532 SDValue N1 = N->getOperand(1);
24534 // We should be performing an xor against a truncated shift.
24535 if (N0.getOpcode() != ISD::TRUNCATE || !N0.hasOneUse())
24538 // Make sure we are performing an xor against one.
24539 if (!isa<ConstantSDNode>(N1) || !cast<ConstantSDNode>(N1)->isOne())
24542 // SetCC on x86 zero extends so only act on this if it's a logical shift.
24543 SDValue Shift = N0.getOperand(0);
24544 if (Shift.getOpcode() != ISD::SRL || !Shift.hasOneUse())
24547 // Make sure we are truncating from one of i16, i32 or i64.
24548 EVT ShiftTy = Shift.getValueType();
24549 if (ShiftTy != MVT::i16 && ShiftTy != MVT::i32 && ShiftTy != MVT::i64)
24552 // Make sure the shift amount extracts the sign bit.
24553 if (!isa<ConstantSDNode>(Shift.getOperand(1)) ||
24554 Shift.getConstantOperandVal(1) != ShiftTy.getSizeInBits() - 1)
24557 // Create a greater-than comparison against -1.
24558 // N.B. Using SETGE against 0 works but we want a canonical looking
24559 // comparison, using SETGT matches up with what TranslateX86CC.
24561 SDValue ShiftOp = Shift.getOperand(0);
24562 EVT ShiftOpTy = ShiftOp.getValueType();
24563 SDValue Cond = DAG.getSetCC(DL, MVT::i8, ShiftOp,
24564 DAG.getConstant(-1, DL, ShiftOpTy), ISD::SETGT);
24568 static SDValue PerformXorCombine(SDNode *N, SelectionDAG &DAG,
24569 TargetLowering::DAGCombinerInfo &DCI,
24570 const X86Subtarget *Subtarget) {
24571 if (DCI.isBeforeLegalizeOps())
24574 if (SDValue RV = foldXorTruncShiftIntoCmp(N, DAG))
24577 if (Subtarget->hasCMov())
24578 if (SDValue RV = performIntegerAbsCombine(N, DAG))
24584 /// PerformLOADCombine - Do target-specific dag combines on LOAD nodes.
24585 static SDValue PerformLOADCombine(SDNode *N, SelectionDAG &DAG,
24586 TargetLowering::DAGCombinerInfo &DCI,
24587 const X86Subtarget *Subtarget) {
24588 LoadSDNode *Ld = cast<LoadSDNode>(N);
24589 EVT RegVT = Ld->getValueType(0);
24590 EVT MemVT = Ld->getMemoryVT();
24592 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
24594 // For chips with slow 32-byte unaligned loads, break the 32-byte operation
24595 // into two 16-byte operations.
24596 ISD::LoadExtType Ext = Ld->getExtensionType();
24598 unsigned AddressSpace = Ld->getAddressSpace();
24599 unsigned Alignment = Ld->getAlignment();
24600 if (RegVT.is256BitVector() && !DCI.isBeforeLegalizeOps() &&
24601 Ext == ISD::NON_EXTLOAD &&
24602 TLI.allowsMemoryAccess(*DAG.getContext(), DAG.getDataLayout(), RegVT,
24603 AddressSpace, Alignment, &Fast) && !Fast) {
24604 unsigned NumElems = RegVT.getVectorNumElements();
24608 SDValue Ptr = Ld->getBasePtr();
24609 SDValue Increment =
24610 DAG.getConstant(16, dl, TLI.getPointerTy(DAG.getDataLayout()));
24612 EVT HalfVT = EVT::getVectorVT(*DAG.getContext(), MemVT.getScalarType(),
24614 SDValue Load1 = DAG.getLoad(HalfVT, dl, Ld->getChain(), Ptr,
24615 Ld->getPointerInfo(), Ld->isVolatile(),
24616 Ld->isNonTemporal(), Ld->isInvariant(),
24618 Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr, Increment);
24619 SDValue Load2 = DAG.getLoad(HalfVT, dl, Ld->getChain(), Ptr,
24620 Ld->getPointerInfo(), Ld->isVolatile(),
24621 Ld->isNonTemporal(), Ld->isInvariant(),
24622 std::min(16U, Alignment));
24623 SDValue TF = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
24625 Load2.getValue(1));
24627 SDValue NewVec = DAG.getUNDEF(RegVT);
24628 NewVec = Insert128BitVector(NewVec, Load1, 0, DAG, dl);
24629 NewVec = Insert128BitVector(NewVec, Load2, NumElems/2, DAG, dl);
24630 return DCI.CombineTo(N, NewVec, TF, true);
24636 /// PerformMLOADCombine - Resolve extending loads
24637 static SDValue PerformMLOADCombine(SDNode *N, SelectionDAG &DAG,
24638 TargetLowering::DAGCombinerInfo &DCI,
24639 const X86Subtarget *Subtarget) {
24640 MaskedLoadSDNode *Mld = cast<MaskedLoadSDNode>(N);
24641 if (Mld->getExtensionType() != ISD::SEXTLOAD)
24644 EVT VT = Mld->getValueType(0);
24645 unsigned NumElems = VT.getVectorNumElements();
24646 EVT LdVT = Mld->getMemoryVT();
24649 assert(LdVT != VT && "Cannot extend to the same type");
24650 unsigned ToSz = VT.getVectorElementType().getSizeInBits();
24651 unsigned FromSz = LdVT.getVectorElementType().getSizeInBits();
24652 // From, To sizes and ElemCount must be pow of two
24653 assert (isPowerOf2_32(NumElems * FromSz * ToSz) &&
24654 "Unexpected size for extending masked load");
24656 unsigned SizeRatio = ToSz / FromSz;
24657 assert(SizeRatio * NumElems * FromSz == VT.getSizeInBits());
24659 // Create a type on which we perform the shuffle
24660 EVT WideVecVT = EVT::getVectorVT(*DAG.getContext(),
24661 LdVT.getScalarType(), NumElems*SizeRatio);
24662 assert(WideVecVT.getSizeInBits() == VT.getSizeInBits());
24664 // Convert Src0 value
24665 SDValue WideSrc0 = DAG.getBitcast(WideVecVT, Mld->getSrc0());
24666 if (Mld->getSrc0().getOpcode() != ISD::UNDEF) {
24667 SmallVector<int, 16> ShuffleVec(NumElems * SizeRatio, -1);
24668 for (unsigned i = 0; i != NumElems; ++i)
24669 ShuffleVec[i] = i * SizeRatio;
24671 // Can't shuffle using an illegal type.
24672 assert (DAG.getTargetLoweringInfo().isTypeLegal(WideVecVT)
24673 && "WideVecVT should be legal");
24674 WideSrc0 = DAG.getVectorShuffle(WideVecVT, dl, WideSrc0,
24675 DAG.getUNDEF(WideVecVT), &ShuffleVec[0]);
24677 // Prepare the new mask
24679 SDValue Mask = Mld->getMask();
24680 if (Mask.getValueType() == VT) {
24681 // Mask and original value have the same type
24682 NewMask = DAG.getBitcast(WideVecVT, Mask);
24683 SmallVector<int, 16> ShuffleVec(NumElems * SizeRatio, -1);
24684 for (unsigned i = 0; i != NumElems; ++i)
24685 ShuffleVec[i] = i * SizeRatio;
24686 for (unsigned i = NumElems; i != NumElems*SizeRatio; ++i)
24687 ShuffleVec[i] = NumElems*SizeRatio;
24688 NewMask = DAG.getVectorShuffle(WideVecVT, dl, NewMask,
24689 DAG.getConstant(0, dl, WideVecVT),
24693 assert(Mask.getValueType().getVectorElementType() == MVT::i1);
24694 unsigned WidenNumElts = NumElems*SizeRatio;
24695 unsigned MaskNumElts = VT.getVectorNumElements();
24696 EVT NewMaskVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
24699 unsigned NumConcat = WidenNumElts / MaskNumElts;
24700 SmallVector<SDValue, 16> Ops(NumConcat);
24701 SDValue ZeroVal = DAG.getConstant(0, dl, Mask.getValueType());
24703 for (unsigned i = 1; i != NumConcat; ++i)
24706 NewMask = DAG.getNode(ISD::CONCAT_VECTORS, dl, NewMaskVT, Ops);
24709 SDValue WideLd = DAG.getMaskedLoad(WideVecVT, dl, Mld->getChain(),
24710 Mld->getBasePtr(), NewMask, WideSrc0,
24711 Mld->getMemoryVT(), Mld->getMemOperand(),
24713 SDValue NewVec = DAG.getNode(X86ISD::VSEXT, dl, VT, WideLd);
24714 return DCI.CombineTo(N, NewVec, WideLd.getValue(1), true);
24717 /// PerformMSTORECombine - Resolve truncating stores
24718 static SDValue PerformMSTORECombine(SDNode *N, SelectionDAG &DAG,
24719 const X86Subtarget *Subtarget) {
24720 MaskedStoreSDNode *Mst = cast<MaskedStoreSDNode>(N);
24721 if (!Mst->isTruncatingStore())
24724 EVT VT = Mst->getValue().getValueType();
24725 unsigned NumElems = VT.getVectorNumElements();
24726 EVT StVT = Mst->getMemoryVT();
24729 assert(StVT != VT && "Cannot truncate to the same type");
24730 unsigned FromSz = VT.getVectorElementType().getSizeInBits();
24731 unsigned ToSz = StVT.getVectorElementType().getSizeInBits();
24733 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
24735 // The truncating store is legal in some cases. For example
24736 // vpmovqb, vpmovqw, vpmovqd, vpmovdb, vpmovdw
24737 // are designated for truncate store.
24738 // In this case we don't need any further transformations.
24739 if (TLI.isTruncStoreLegal(VT, StVT))
24742 // From, To sizes and ElemCount must be pow of two
24743 assert (isPowerOf2_32(NumElems * FromSz * ToSz) &&
24744 "Unexpected size for truncating masked store");
24745 // We are going to use the original vector elt for storing.
24746 // Accumulated smaller vector elements must be a multiple of the store size.
24747 assert (((NumElems * FromSz) % ToSz) == 0 &&
24748 "Unexpected ratio for truncating masked store");
24750 unsigned SizeRatio = FromSz / ToSz;
24751 assert(SizeRatio * NumElems * ToSz == VT.getSizeInBits());
24753 // Create a type on which we perform the shuffle
24754 EVT WideVecVT = EVT::getVectorVT(*DAG.getContext(),
24755 StVT.getScalarType(), NumElems*SizeRatio);
24757 assert(WideVecVT.getSizeInBits() == VT.getSizeInBits());
24759 SDValue WideVec = DAG.getBitcast(WideVecVT, Mst->getValue());
24760 SmallVector<int, 16> ShuffleVec(NumElems * SizeRatio, -1);
24761 for (unsigned i = 0; i != NumElems; ++i)
24762 ShuffleVec[i] = i * SizeRatio;
24764 // Can't shuffle using an illegal type.
24765 assert (DAG.getTargetLoweringInfo().isTypeLegal(WideVecVT)
24766 && "WideVecVT should be legal");
24768 SDValue TruncatedVal = DAG.getVectorShuffle(WideVecVT, dl, WideVec,
24769 DAG.getUNDEF(WideVecVT),
24773 SDValue Mask = Mst->getMask();
24774 if (Mask.getValueType() == VT) {
24775 // Mask and original value have the same type
24776 NewMask = DAG.getBitcast(WideVecVT, Mask);
24777 for (unsigned i = 0; i != NumElems; ++i)
24778 ShuffleVec[i] = i * SizeRatio;
24779 for (unsigned i = NumElems; i != NumElems*SizeRatio; ++i)
24780 ShuffleVec[i] = NumElems*SizeRatio;
24781 NewMask = DAG.getVectorShuffle(WideVecVT, dl, NewMask,
24782 DAG.getConstant(0, dl, WideVecVT),
24786 assert(Mask.getValueType().getVectorElementType() == MVT::i1);
24787 unsigned WidenNumElts = NumElems*SizeRatio;
24788 unsigned MaskNumElts = VT.getVectorNumElements();
24789 EVT NewMaskVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
24792 unsigned NumConcat = WidenNumElts / MaskNumElts;
24793 SmallVector<SDValue, 16> Ops(NumConcat);
24794 SDValue ZeroVal = DAG.getConstant(0, dl, Mask.getValueType());
24796 for (unsigned i = 1; i != NumConcat; ++i)
24799 NewMask = DAG.getNode(ISD::CONCAT_VECTORS, dl, NewMaskVT, Ops);
24802 return DAG.getMaskedStore(Mst->getChain(), dl, TruncatedVal, Mst->getBasePtr(),
24803 NewMask, StVT, Mst->getMemOperand(), false);
24805 /// PerformSTORECombine - Do target-specific dag combines on STORE nodes.
24806 static SDValue PerformSTORECombine(SDNode *N, SelectionDAG &DAG,
24807 const X86Subtarget *Subtarget) {
24808 StoreSDNode *St = cast<StoreSDNode>(N);
24809 EVT VT = St->getValue().getValueType();
24810 EVT StVT = St->getMemoryVT();
24812 SDValue StoredVal = St->getOperand(1);
24813 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
24815 // If we are saving a concatenation of two XMM registers and 32-byte stores
24816 // are slow, such as on Sandy Bridge, perform two 16-byte stores.
24818 unsigned AddressSpace = St->getAddressSpace();
24819 unsigned Alignment = St->getAlignment();
24820 if (VT.is256BitVector() && StVT == VT &&
24821 TLI.allowsMemoryAccess(*DAG.getContext(), DAG.getDataLayout(), VT,
24822 AddressSpace, Alignment, &Fast) && !Fast) {
24823 unsigned NumElems = VT.getVectorNumElements();
24827 SDValue Value0 = Extract128BitVector(StoredVal, 0, DAG, dl);
24828 SDValue Value1 = Extract128BitVector(StoredVal, NumElems/2, DAG, dl);
24831 DAG.getConstant(16, dl, TLI.getPointerTy(DAG.getDataLayout()));
24832 SDValue Ptr0 = St->getBasePtr();
24833 SDValue Ptr1 = DAG.getNode(ISD::ADD, dl, Ptr0.getValueType(), Ptr0, Stride);
24835 SDValue Ch0 = DAG.getStore(St->getChain(), dl, Value0, Ptr0,
24836 St->getPointerInfo(), St->isVolatile(),
24837 St->isNonTemporal(), Alignment);
24838 SDValue Ch1 = DAG.getStore(St->getChain(), dl, Value1, Ptr1,
24839 St->getPointerInfo(), St->isVolatile(),
24840 St->isNonTemporal(),
24841 std::min(16U, Alignment));
24842 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Ch0, Ch1);
24845 // Optimize trunc store (of multiple scalars) to shuffle and store.
24846 // First, pack all of the elements in one place. Next, store to memory
24847 // in fewer chunks.
24848 if (St->isTruncatingStore() && VT.isVector()) {
24849 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
24850 unsigned NumElems = VT.getVectorNumElements();
24851 assert(StVT != VT && "Cannot truncate to the same type");
24852 unsigned FromSz = VT.getVectorElementType().getSizeInBits();
24853 unsigned ToSz = StVT.getVectorElementType().getSizeInBits();
24855 // The truncating store is legal in some cases. For example
24856 // vpmovqb, vpmovqw, vpmovqd, vpmovdb, vpmovdw
24857 // are designated for truncate store.
24858 // In this case we don't need any further transformations.
24859 if (TLI.isTruncStoreLegal(VT, StVT))
24862 // From, To sizes and ElemCount must be pow of two
24863 if (!isPowerOf2_32(NumElems * FromSz * ToSz)) return SDValue();
24864 // We are going to use the original vector elt for storing.
24865 // Accumulated smaller vector elements must be a multiple of the store size.
24866 if (0 != (NumElems * FromSz) % ToSz) return SDValue();
24868 unsigned SizeRatio = FromSz / ToSz;
24870 assert(SizeRatio * NumElems * ToSz == VT.getSizeInBits());
24872 // Create a type on which we perform the shuffle
24873 EVT WideVecVT = EVT::getVectorVT(*DAG.getContext(),
24874 StVT.getScalarType(), NumElems*SizeRatio);
24876 assert(WideVecVT.getSizeInBits() == VT.getSizeInBits());
24878 SDValue WideVec = DAG.getBitcast(WideVecVT, St->getValue());
24879 SmallVector<int, 8> ShuffleVec(NumElems * SizeRatio, -1);
24880 for (unsigned i = 0; i != NumElems; ++i)
24881 ShuffleVec[i] = i * SizeRatio;
24883 // Can't shuffle using an illegal type.
24884 if (!TLI.isTypeLegal(WideVecVT))
24887 SDValue Shuff = DAG.getVectorShuffle(WideVecVT, dl, WideVec,
24888 DAG.getUNDEF(WideVecVT),
24890 // At this point all of the data is stored at the bottom of the
24891 // register. We now need to save it to mem.
24893 // Find the largest store unit
24894 MVT StoreType = MVT::i8;
24895 for (MVT Tp : MVT::integer_valuetypes()) {
24896 if (TLI.isTypeLegal(Tp) && Tp.getSizeInBits() <= NumElems * ToSz)
24900 // On 32bit systems, we can't save 64bit integers. Try bitcasting to F64.
24901 if (TLI.isTypeLegal(MVT::f64) && StoreType.getSizeInBits() < 64 &&
24902 (64 <= NumElems * ToSz))
24903 StoreType = MVT::f64;
24905 // Bitcast the original vector into a vector of store-size units
24906 EVT StoreVecVT = EVT::getVectorVT(*DAG.getContext(),
24907 StoreType, VT.getSizeInBits()/StoreType.getSizeInBits());
24908 assert(StoreVecVT.getSizeInBits() == VT.getSizeInBits());
24909 SDValue ShuffWide = DAG.getBitcast(StoreVecVT, Shuff);
24910 SmallVector<SDValue, 8> Chains;
24911 SDValue Increment = DAG.getConstant(StoreType.getSizeInBits() / 8, dl,
24912 TLI.getPointerTy(DAG.getDataLayout()));
24913 SDValue Ptr = St->getBasePtr();
24915 // Perform one or more big stores into memory.
24916 for (unsigned i=0, e=(ToSz*NumElems)/StoreType.getSizeInBits(); i!=e; ++i) {
24917 SDValue SubVec = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl,
24918 StoreType, ShuffWide,
24919 DAG.getIntPtrConstant(i, dl));
24920 SDValue Ch = DAG.getStore(St->getChain(), dl, SubVec, Ptr,
24921 St->getPointerInfo(), St->isVolatile(),
24922 St->isNonTemporal(), St->getAlignment());
24923 Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr, Increment);
24924 Chains.push_back(Ch);
24927 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Chains);
24930 // Turn load->store of MMX types into GPR load/stores. This avoids clobbering
24931 // the FP state in cases where an emms may be missing.
24932 // A preferable solution to the general problem is to figure out the right
24933 // places to insert EMMS. This qualifies as a quick hack.
24935 // Similarly, turn load->store of i64 into double load/stores in 32-bit mode.
24936 if (VT.getSizeInBits() != 64)
24939 const Function *F = DAG.getMachineFunction().getFunction();
24940 bool NoImplicitFloatOps = F->hasFnAttribute(Attribute::NoImplicitFloat);
24942 !Subtarget->useSoftFloat() && !NoImplicitFloatOps && Subtarget->hasSSE2();
24943 if ((VT.isVector() ||
24944 (VT == MVT::i64 && F64IsLegal && !Subtarget->is64Bit())) &&
24945 isa<LoadSDNode>(St->getValue()) &&
24946 !cast<LoadSDNode>(St->getValue())->isVolatile() &&
24947 St->getChain().hasOneUse() && !St->isVolatile()) {
24948 SDNode* LdVal = St->getValue().getNode();
24949 LoadSDNode *Ld = nullptr;
24950 int TokenFactorIndex = -1;
24951 SmallVector<SDValue, 8> Ops;
24952 SDNode* ChainVal = St->getChain().getNode();
24953 // Must be a store of a load. We currently handle two cases: the load
24954 // is a direct child, and it's under an intervening TokenFactor. It is
24955 // possible to dig deeper under nested TokenFactors.
24956 if (ChainVal == LdVal)
24957 Ld = cast<LoadSDNode>(St->getChain());
24958 else if (St->getValue().hasOneUse() &&
24959 ChainVal->getOpcode() == ISD::TokenFactor) {
24960 for (unsigned i = 0, e = ChainVal->getNumOperands(); i != e; ++i) {
24961 if (ChainVal->getOperand(i).getNode() == LdVal) {
24962 TokenFactorIndex = i;
24963 Ld = cast<LoadSDNode>(St->getValue());
24965 Ops.push_back(ChainVal->getOperand(i));
24969 if (!Ld || !ISD::isNormalLoad(Ld))
24972 // If this is not the MMX case, i.e. we are just turning i64 load/store
24973 // into f64 load/store, avoid the transformation if there are multiple
24974 // uses of the loaded value.
24975 if (!VT.isVector() && !Ld->hasNUsesOfValue(1, 0))
24980 // If we are a 64-bit capable x86, lower to a single movq load/store pair.
24981 // Otherwise, if it's legal to use f64 SSE instructions, use f64 load/store
24983 if (Subtarget->is64Bit() || F64IsLegal) {
24984 EVT LdVT = Subtarget->is64Bit() ? MVT::i64 : MVT::f64;
24985 SDValue NewLd = DAG.getLoad(LdVT, LdDL, Ld->getChain(), Ld->getBasePtr(),
24986 Ld->getPointerInfo(), Ld->isVolatile(),
24987 Ld->isNonTemporal(), Ld->isInvariant(),
24988 Ld->getAlignment());
24989 SDValue NewChain = NewLd.getValue(1);
24990 if (TokenFactorIndex != -1) {
24991 Ops.push_back(NewChain);
24992 NewChain = DAG.getNode(ISD::TokenFactor, LdDL, MVT::Other, Ops);
24994 return DAG.getStore(NewChain, StDL, NewLd, St->getBasePtr(),
24995 St->getPointerInfo(),
24996 St->isVolatile(), St->isNonTemporal(),
24997 St->getAlignment());
25000 // Otherwise, lower to two pairs of 32-bit loads / stores.
25001 SDValue LoAddr = Ld->getBasePtr();
25002 SDValue HiAddr = DAG.getNode(ISD::ADD, LdDL, MVT::i32, LoAddr,
25003 DAG.getConstant(4, LdDL, MVT::i32));
25005 SDValue LoLd = DAG.getLoad(MVT::i32, LdDL, Ld->getChain(), LoAddr,
25006 Ld->getPointerInfo(),
25007 Ld->isVolatile(), Ld->isNonTemporal(),
25008 Ld->isInvariant(), Ld->getAlignment());
25009 SDValue HiLd = DAG.getLoad(MVT::i32, LdDL, Ld->getChain(), HiAddr,
25010 Ld->getPointerInfo().getWithOffset(4),
25011 Ld->isVolatile(), Ld->isNonTemporal(),
25013 MinAlign(Ld->getAlignment(), 4));
25015 SDValue NewChain = LoLd.getValue(1);
25016 if (TokenFactorIndex != -1) {
25017 Ops.push_back(LoLd);
25018 Ops.push_back(HiLd);
25019 NewChain = DAG.getNode(ISD::TokenFactor, LdDL, MVT::Other, Ops);
25022 LoAddr = St->getBasePtr();
25023 HiAddr = DAG.getNode(ISD::ADD, StDL, MVT::i32, LoAddr,
25024 DAG.getConstant(4, StDL, MVT::i32));
25026 SDValue LoSt = DAG.getStore(NewChain, StDL, LoLd, LoAddr,
25027 St->getPointerInfo(),
25028 St->isVolatile(), St->isNonTemporal(),
25029 St->getAlignment());
25030 SDValue HiSt = DAG.getStore(NewChain, StDL, HiLd, HiAddr,
25031 St->getPointerInfo().getWithOffset(4),
25033 St->isNonTemporal(),
25034 MinAlign(St->getAlignment(), 4));
25035 return DAG.getNode(ISD::TokenFactor, StDL, MVT::Other, LoSt, HiSt);
25038 // This is similar to the above case, but here we handle a scalar 64-bit
25039 // integer store that is extracted from a vector on a 32-bit target.
25040 // If we have SSE2, then we can treat it like a floating-point double
25041 // to get past legalization. The execution dependencies fixup pass will
25042 // choose the optimal machine instruction for the store if this really is
25043 // an integer or v2f32 rather than an f64.
25044 if (VT == MVT::i64 && F64IsLegal && !Subtarget->is64Bit() &&
25045 St->getOperand(1).getOpcode() == ISD::EXTRACT_VECTOR_ELT) {
25046 SDValue OldExtract = St->getOperand(1);
25047 SDValue ExtOp0 = OldExtract.getOperand(0);
25048 unsigned VecSize = ExtOp0.getValueSizeInBits();
25049 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), MVT::f64, VecSize / 64);
25050 SDValue BitCast = DAG.getBitcast(VecVT, ExtOp0);
25051 SDValue NewExtract = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
25052 BitCast, OldExtract.getOperand(1));
25053 return DAG.getStore(St->getChain(), dl, NewExtract, St->getBasePtr(),
25054 St->getPointerInfo(), St->isVolatile(),
25055 St->isNonTemporal(), St->getAlignment());
25061 /// Return 'true' if this vector operation is "horizontal"
25062 /// and return the operands for the horizontal operation in LHS and RHS. A
25063 /// horizontal operation performs the binary operation on successive elements
25064 /// of its first operand, then on successive elements of its second operand,
25065 /// returning the resulting values in a vector. For example, if
25066 /// A = < float a0, float a1, float a2, float a3 >
25068 /// B = < float b0, float b1, float b2, float b3 >
25069 /// then the result of doing a horizontal operation on A and B is
25070 /// A horizontal-op B = < a0 op a1, a2 op a3, b0 op b1, b2 op b3 >.
25071 /// In short, LHS and RHS are inspected to see if LHS op RHS is of the form
25072 /// A horizontal-op B, for some already available A and B, and if so then LHS is
25073 /// set to A, RHS to B, and the routine returns 'true'.
25074 /// Note that the binary operation should have the property that if one of the
25075 /// operands is UNDEF then the result is UNDEF.
25076 static bool isHorizontalBinOp(SDValue &LHS, SDValue &RHS, bool IsCommutative) {
25077 // Look for the following pattern: if
25078 // A = < float a0, float a1, float a2, float a3 >
25079 // B = < float b0, float b1, float b2, float b3 >
25081 // LHS = VECTOR_SHUFFLE A, B, <0, 2, 4, 6>
25082 // RHS = VECTOR_SHUFFLE A, B, <1, 3, 5, 7>
25083 // then LHS op RHS = < a0 op a1, a2 op a3, b0 op b1, b2 op b3 >
25084 // which is A horizontal-op B.
25086 // At least one of the operands should be a vector shuffle.
25087 if (LHS.getOpcode() != ISD::VECTOR_SHUFFLE &&
25088 RHS.getOpcode() != ISD::VECTOR_SHUFFLE)
25091 MVT VT = LHS.getSimpleValueType();
25093 assert((VT.is128BitVector() || VT.is256BitVector()) &&
25094 "Unsupported vector type for horizontal add/sub");
25096 // Handle 128 and 256-bit vector lengths. AVX defines horizontal add/sub to
25097 // operate independently on 128-bit lanes.
25098 unsigned NumElts = VT.getVectorNumElements();
25099 unsigned NumLanes = VT.getSizeInBits()/128;
25100 unsigned NumLaneElts = NumElts / NumLanes;
25101 assert((NumLaneElts % 2 == 0) &&
25102 "Vector type should have an even number of elements in each lane");
25103 unsigned HalfLaneElts = NumLaneElts/2;
25105 // View LHS in the form
25106 // LHS = VECTOR_SHUFFLE A, B, LMask
25107 // If LHS is not a shuffle then pretend it is the shuffle
25108 // LHS = VECTOR_SHUFFLE LHS, undef, <0, 1, ..., N-1>
25109 // NOTE: in what follows a default initialized SDValue represents an UNDEF of
25112 SmallVector<int, 16> LMask(NumElts);
25113 if (LHS.getOpcode() == ISD::VECTOR_SHUFFLE) {
25114 if (LHS.getOperand(0).getOpcode() != ISD::UNDEF)
25115 A = LHS.getOperand(0);
25116 if (LHS.getOperand(1).getOpcode() != ISD::UNDEF)
25117 B = LHS.getOperand(1);
25118 ArrayRef<int> Mask = cast<ShuffleVectorSDNode>(LHS.getNode())->getMask();
25119 std::copy(Mask.begin(), Mask.end(), LMask.begin());
25121 if (LHS.getOpcode() != ISD::UNDEF)
25123 for (unsigned i = 0; i != NumElts; ++i)
25127 // Likewise, view RHS in the form
25128 // RHS = VECTOR_SHUFFLE C, D, RMask
25130 SmallVector<int, 16> RMask(NumElts);
25131 if (RHS.getOpcode() == ISD::VECTOR_SHUFFLE) {
25132 if (RHS.getOperand(0).getOpcode() != ISD::UNDEF)
25133 C = RHS.getOperand(0);
25134 if (RHS.getOperand(1).getOpcode() != ISD::UNDEF)
25135 D = RHS.getOperand(1);
25136 ArrayRef<int> Mask = cast<ShuffleVectorSDNode>(RHS.getNode())->getMask();
25137 std::copy(Mask.begin(), Mask.end(), RMask.begin());
25139 if (RHS.getOpcode() != ISD::UNDEF)
25141 for (unsigned i = 0; i != NumElts; ++i)
25145 // Check that the shuffles are both shuffling the same vectors.
25146 if (!(A == C && B == D) && !(A == D && B == C))
25149 // If everything is UNDEF then bail out: it would be better to fold to UNDEF.
25150 if (!A.getNode() && !B.getNode())
25153 // If A and B occur in reverse order in RHS, then "swap" them (which means
25154 // rewriting the mask).
25156 ShuffleVectorSDNode::commuteMask(RMask);
25158 // At this point LHS and RHS are equivalent to
25159 // LHS = VECTOR_SHUFFLE A, B, LMask
25160 // RHS = VECTOR_SHUFFLE A, B, RMask
25161 // Check that the masks correspond to performing a horizontal operation.
25162 for (unsigned l = 0; l != NumElts; l += NumLaneElts) {
25163 for (unsigned i = 0; i != NumLaneElts; ++i) {
25164 int LIdx = LMask[i+l], RIdx = RMask[i+l];
25166 // Ignore any UNDEF components.
25167 if (LIdx < 0 || RIdx < 0 ||
25168 (!A.getNode() && (LIdx < (int)NumElts || RIdx < (int)NumElts)) ||
25169 (!B.getNode() && (LIdx >= (int)NumElts || RIdx >= (int)NumElts)))
25172 // Check that successive elements are being operated on. If not, this is
25173 // not a horizontal operation.
25174 unsigned Src = (i/HalfLaneElts); // each lane is split between srcs
25175 int Index = 2*(i%HalfLaneElts) + NumElts*Src + l;
25176 if (!(LIdx == Index && RIdx == Index + 1) &&
25177 !(IsCommutative && LIdx == Index + 1 && RIdx == Index))
25182 LHS = A.getNode() ? A : B; // If A is 'UNDEF', use B for it.
25183 RHS = B.getNode() ? B : A; // If B is 'UNDEF', use A for it.
25187 /// Do target-specific dag combines on floating point adds.
25188 static SDValue PerformFADDCombine(SDNode *N, SelectionDAG &DAG,
25189 const X86Subtarget *Subtarget) {
25190 EVT VT = N->getValueType(0);
25191 SDValue LHS = N->getOperand(0);
25192 SDValue RHS = N->getOperand(1);
25194 // Try to synthesize horizontal adds from adds of shuffles.
25195 if (((Subtarget->hasSSE3() && (VT == MVT::v4f32 || VT == MVT::v2f64)) ||
25196 (Subtarget->hasFp256() && (VT == MVT::v8f32 || VT == MVT::v4f64))) &&
25197 isHorizontalBinOp(LHS, RHS, true))
25198 return DAG.getNode(X86ISD::FHADD, SDLoc(N), VT, LHS, RHS);
25202 /// Do target-specific dag combines on floating point subs.
25203 static SDValue PerformFSUBCombine(SDNode *N, SelectionDAG &DAG,
25204 const X86Subtarget *Subtarget) {
25205 EVT VT = N->getValueType(0);
25206 SDValue LHS = N->getOperand(0);
25207 SDValue RHS = N->getOperand(1);
25209 // Try to synthesize horizontal subs from subs of shuffles.
25210 if (((Subtarget->hasSSE3() && (VT == MVT::v4f32 || VT == MVT::v2f64)) ||
25211 (Subtarget->hasFp256() && (VT == MVT::v8f32 || VT == MVT::v4f64))) &&
25212 isHorizontalBinOp(LHS, RHS, false))
25213 return DAG.getNode(X86ISD::FHSUB, SDLoc(N), VT, LHS, RHS);
25217 /// Do target-specific dag combines on X86ISD::FOR and X86ISD::FXOR nodes.
25218 static SDValue PerformFORCombine(SDNode *N, SelectionDAG &DAG,
25219 const X86Subtarget *Subtarget) {
25220 assert(N->getOpcode() == X86ISD::FOR || N->getOpcode() == X86ISD::FXOR);
25222 // F[X]OR(0.0, x) -> x
25223 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
25224 if (C->getValueAPF().isPosZero())
25225 return N->getOperand(1);
25227 // F[X]OR(x, 0.0) -> x
25228 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
25229 if (C->getValueAPF().isPosZero())
25230 return N->getOperand(0);
25232 EVT VT = N->getValueType(0);
25233 if (VT.is512BitVector() && !Subtarget->hasDQI()) {
25235 MVT IntScalar = MVT::getIntegerVT(VT.getScalarSizeInBits());
25236 MVT IntVT = MVT::getVectorVT(IntScalar, VT.getVectorNumElements());
25238 SDValue Op0 = DAG.getNode(ISD::BITCAST, dl, IntVT, N->getOperand(0));
25239 SDValue Op1 = DAG.getNode(ISD::BITCAST, dl, IntVT, N->getOperand(1));
25240 unsigned IntOpcode = (N->getOpcode() == X86ISD::FOR) ? ISD::OR : ISD::XOR;
25241 SDValue IntOp = DAG.getNode(IntOpcode, dl, IntVT, Op0, Op1);
25242 return DAG.getNode(ISD::BITCAST, dl, VT, IntOp);
25247 /// Do target-specific dag combines on X86ISD::FMIN and X86ISD::FMAX nodes.
25248 static SDValue PerformFMinFMaxCombine(SDNode *N, SelectionDAG &DAG) {
25249 assert(N->getOpcode() == X86ISD::FMIN || N->getOpcode() == X86ISD::FMAX);
25251 // Only perform optimizations if UnsafeMath is used.
25252 if (!DAG.getTarget().Options.UnsafeFPMath)
25255 // If we run in unsafe-math mode, then convert the FMAX and FMIN nodes
25256 // into FMINC and FMAXC, which are Commutative operations.
25257 unsigned NewOp = 0;
25258 switch (N->getOpcode()) {
25259 default: llvm_unreachable("unknown opcode");
25260 case X86ISD::FMIN: NewOp = X86ISD::FMINC; break;
25261 case X86ISD::FMAX: NewOp = X86ISD::FMAXC; break;
25264 return DAG.getNode(NewOp, SDLoc(N), N->getValueType(0),
25265 N->getOperand(0), N->getOperand(1));
25268 /// Do target-specific dag combines on X86ISD::FAND nodes.
25269 static SDValue PerformFANDCombine(SDNode *N, SelectionDAG &DAG) {
25270 // FAND(0.0, x) -> 0.0
25271 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
25272 if (C->getValueAPF().isPosZero())
25273 return N->getOperand(0);
25275 // FAND(x, 0.0) -> 0.0
25276 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
25277 if (C->getValueAPF().isPosZero())
25278 return N->getOperand(1);
25283 /// Do target-specific dag combines on X86ISD::FANDN nodes
25284 static SDValue PerformFANDNCombine(SDNode *N, SelectionDAG &DAG) {
25285 // FANDN(0.0, x) -> x
25286 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
25287 if (C->getValueAPF().isPosZero())
25288 return N->getOperand(1);
25290 // FANDN(x, 0.0) -> 0.0
25291 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
25292 if (C->getValueAPF().isPosZero())
25293 return N->getOperand(1);
25298 static SDValue PerformBTCombine(SDNode *N,
25300 TargetLowering::DAGCombinerInfo &DCI) {
25301 // BT ignores high bits in the bit index operand.
25302 SDValue Op1 = N->getOperand(1);
25303 if (Op1.hasOneUse()) {
25304 unsigned BitWidth = Op1.getValueSizeInBits();
25305 APInt DemandedMask = APInt::getLowBitsSet(BitWidth, Log2_32(BitWidth));
25306 APInt KnownZero, KnownOne;
25307 TargetLowering::TargetLoweringOpt TLO(DAG, !DCI.isBeforeLegalize(),
25308 !DCI.isBeforeLegalizeOps());
25309 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
25310 if (TLO.ShrinkDemandedConstant(Op1, DemandedMask) ||
25311 TLI.SimplifyDemandedBits(Op1, DemandedMask, KnownZero, KnownOne, TLO))
25312 DCI.CommitTargetLoweringOpt(TLO);
25317 static SDValue PerformVZEXT_MOVLCombine(SDNode *N, SelectionDAG &DAG) {
25318 SDValue Op = N->getOperand(0);
25319 if (Op.getOpcode() == ISD::BITCAST)
25320 Op = Op.getOperand(0);
25321 EVT VT = N->getValueType(0), OpVT = Op.getValueType();
25322 if (Op.getOpcode() == X86ISD::VZEXT_LOAD &&
25323 VT.getVectorElementType().getSizeInBits() ==
25324 OpVT.getVectorElementType().getSizeInBits()) {
25325 return DAG.getNode(ISD::BITCAST, SDLoc(N), VT, Op);
25330 static SDValue PerformSIGN_EXTEND_INREGCombine(SDNode *N, SelectionDAG &DAG,
25331 const X86Subtarget *Subtarget) {
25332 EVT VT = N->getValueType(0);
25333 if (!VT.isVector())
25336 SDValue N0 = N->getOperand(0);
25337 SDValue N1 = N->getOperand(1);
25338 EVT ExtraVT = cast<VTSDNode>(N1)->getVT();
25341 // The SIGN_EXTEND_INREG to v4i64 is expensive operation on the
25342 // both SSE and AVX2 since there is no sign-extended shift right
25343 // operation on a vector with 64-bit elements.
25344 //(sext_in_reg (v4i64 anyext (v4i32 x )), ExtraVT) ->
25345 // (v4i64 sext (v4i32 sext_in_reg (v4i32 x , ExtraVT)))
25346 if (VT == MVT::v4i64 && (N0.getOpcode() == ISD::ANY_EXTEND ||
25347 N0.getOpcode() == ISD::SIGN_EXTEND)) {
25348 SDValue N00 = N0.getOperand(0);
25350 // EXTLOAD has a better solution on AVX2,
25351 // it may be replaced with X86ISD::VSEXT node.
25352 if (N00.getOpcode() == ISD::LOAD && Subtarget->hasInt256())
25353 if (!ISD::isNormalLoad(N00.getNode()))
25356 if (N00.getValueType() == MVT::v4i32 && ExtraVT.getSizeInBits() < 128) {
25357 SDValue Tmp = DAG.getNode(ISD::SIGN_EXTEND_INREG, dl, MVT::v4i32,
25359 return DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i64, Tmp);
25365 static SDValue PerformSExtCombine(SDNode *N, SelectionDAG &DAG,
25366 TargetLowering::DAGCombinerInfo &DCI,
25367 const X86Subtarget *Subtarget) {
25368 SDValue N0 = N->getOperand(0);
25369 EVT VT = N->getValueType(0);
25370 EVT SVT = VT.getScalarType();
25371 EVT InVT = N0.getValueType();
25372 EVT InSVT = InVT.getScalarType();
25375 // (i8,i32 sext (sdivrem (i8 x, i8 y)) ->
25376 // (i8,i32 (sdivrem_sext_hreg (i8 x, i8 y)
25377 // This exposes the sext to the sdivrem lowering, so that it directly extends
25378 // from AH (which we otherwise need to do contortions to access).
25379 if (N0.getOpcode() == ISD::SDIVREM && N0.getResNo() == 1 &&
25380 InVT == MVT::i8 && VT == MVT::i32) {
25381 SDVTList NodeTys = DAG.getVTList(MVT::i8, VT);
25382 SDValue R = DAG.getNode(X86ISD::SDIVREM8_SEXT_HREG, DL, NodeTys,
25383 N0.getOperand(0), N0.getOperand(1));
25384 DAG.ReplaceAllUsesOfValueWith(N0.getValue(0), R.getValue(0));
25385 return R.getValue(1);
25388 if (!DCI.isBeforeLegalizeOps()) {
25389 if (InVT == MVT::i1) {
25390 SDValue Zero = DAG.getConstant(0, DL, VT);
25392 DAG.getConstant(APInt::getAllOnesValue(VT.getSizeInBits()), DL, VT);
25393 return DAG.getNode(ISD::SELECT, DL, VT, N0, AllOnes, Zero);
25398 if (VT.isVector() && Subtarget->hasSSE2()) {
25399 auto ExtendVecSize = [&DAG](SDLoc DL, SDValue N, unsigned Size) {
25400 EVT InVT = N.getValueType();
25401 EVT OutVT = EVT::getVectorVT(*DAG.getContext(), InVT.getScalarType(),
25402 Size / InVT.getScalarSizeInBits());
25403 SmallVector<SDValue, 8> Opnds(Size / InVT.getSizeInBits(),
25404 DAG.getUNDEF(InVT));
25406 return DAG.getNode(ISD::CONCAT_VECTORS, DL, OutVT, Opnds);
25409 // If target-size is less than 128-bits, extend to a type that would extend
25410 // to 128 bits, extend that and extract the original target vector.
25411 if (VT.getSizeInBits() < 128 && !(128 % VT.getSizeInBits()) &&
25412 (SVT == MVT::i64 || SVT == MVT::i32 || SVT == MVT::i16) &&
25413 (InSVT == MVT::i32 || InSVT == MVT::i16 || InSVT == MVT::i8)) {
25414 unsigned Scale = 128 / VT.getSizeInBits();
25416 EVT::getVectorVT(*DAG.getContext(), SVT, 128 / SVT.getSizeInBits());
25417 SDValue Ex = ExtendVecSize(DL, N0, Scale * InVT.getSizeInBits());
25418 SDValue SExt = DAG.getNode(ISD::SIGN_EXTEND, DL, ExVT, Ex);
25419 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, VT, SExt,
25420 DAG.getIntPtrConstant(0, DL));
25423 // If target-size is 128-bits, then convert to ISD::SIGN_EXTEND_VECTOR_INREG
25424 // which ensures lowering to X86ISD::VSEXT (pmovsx*).
25425 if (VT.getSizeInBits() == 128 &&
25426 (SVT == MVT::i64 || SVT == MVT::i32 || SVT == MVT::i16) &&
25427 (InSVT == MVT::i32 || InSVT == MVT::i16 || InSVT == MVT::i8)) {
25428 SDValue ExOp = ExtendVecSize(DL, N0, 128);
25429 return DAG.getSignExtendVectorInReg(ExOp, DL, VT);
25432 // On pre-AVX2 targets, split into 128-bit nodes of
25433 // ISD::SIGN_EXTEND_VECTOR_INREG.
25434 if (!Subtarget->hasInt256() && !(VT.getSizeInBits() % 128) &&
25435 (SVT == MVT::i64 || SVT == MVT::i32 || SVT == MVT::i16) &&
25436 (InSVT == MVT::i32 || InSVT == MVT::i16 || InSVT == MVT::i8)) {
25437 unsigned NumVecs = VT.getSizeInBits() / 128;
25438 unsigned NumSubElts = 128 / SVT.getSizeInBits();
25439 EVT SubVT = EVT::getVectorVT(*DAG.getContext(), SVT, NumSubElts);
25440 EVT InSubVT = EVT::getVectorVT(*DAG.getContext(), InSVT, NumSubElts);
25442 SmallVector<SDValue, 8> Opnds;
25443 for (unsigned i = 0, Offset = 0; i != NumVecs;
25444 ++i, Offset += NumSubElts) {
25445 SDValue SrcVec = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, InSubVT, N0,
25446 DAG.getIntPtrConstant(Offset, DL));
25447 SrcVec = ExtendVecSize(DL, SrcVec, 128);
25448 SrcVec = DAG.getSignExtendVectorInReg(SrcVec, DL, SubVT);
25449 Opnds.push_back(SrcVec);
25451 return DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, Opnds);
25455 if (!Subtarget->hasFp256())
25458 if (VT.isVector() && VT.getSizeInBits() == 256)
25459 if (SDValue R = WidenMaskArithmetic(N, DAG, DCI, Subtarget))
25465 static SDValue PerformFMACombine(SDNode *N, SelectionDAG &DAG,
25466 const X86Subtarget* Subtarget) {
25468 EVT VT = N->getValueType(0);
25470 // Let legalize expand this if it isn't a legal type yet.
25471 if (!DAG.getTargetLoweringInfo().isTypeLegal(VT))
25474 EVT ScalarVT = VT.getScalarType();
25475 if ((ScalarVT != MVT::f32 && ScalarVT != MVT::f64) ||
25476 (!Subtarget->hasFMA() && !Subtarget->hasFMA4() &&
25477 !Subtarget->hasAVX512()))
25480 SDValue A = N->getOperand(0);
25481 SDValue B = N->getOperand(1);
25482 SDValue C = N->getOperand(2);
25484 bool NegA = (A.getOpcode() == ISD::FNEG);
25485 bool NegB = (B.getOpcode() == ISD::FNEG);
25486 bool NegC = (C.getOpcode() == ISD::FNEG);
25488 // Negative multiplication when NegA xor NegB
25489 bool NegMul = (NegA != NegB);
25491 A = A.getOperand(0);
25493 B = B.getOperand(0);
25495 C = C.getOperand(0);
25499 Opcode = (!NegC) ? X86ISD::FMADD : X86ISD::FMSUB;
25501 Opcode = (!NegC) ? X86ISD::FNMADD : X86ISD::FNMSUB;
25503 return DAG.getNode(Opcode, dl, VT, A, B, C);
25506 static SDValue PerformZExtCombine(SDNode *N, SelectionDAG &DAG,
25507 TargetLowering::DAGCombinerInfo &DCI,
25508 const X86Subtarget *Subtarget) {
25509 // (i32 zext (and (i8 x86isd::setcc_carry), 1)) ->
25510 // (and (i32 x86isd::setcc_carry), 1)
25511 // This eliminates the zext. This transformation is necessary because
25512 // ISD::SETCC is always legalized to i8.
25514 SDValue N0 = N->getOperand(0);
25515 EVT VT = N->getValueType(0);
25517 if (N0.getOpcode() == ISD::AND &&
25519 N0.getOperand(0).hasOneUse()) {
25520 SDValue N00 = N0.getOperand(0);
25521 if (N00.getOpcode() == X86ISD::SETCC_CARRY) {
25522 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
25523 if (!C || C->getZExtValue() != 1)
25525 return DAG.getNode(ISD::AND, dl, VT,
25526 DAG.getNode(X86ISD::SETCC_CARRY, dl, VT,
25527 N00.getOperand(0), N00.getOperand(1)),
25528 DAG.getConstant(1, dl, VT));
25532 if (N0.getOpcode() == ISD::TRUNCATE &&
25534 N0.getOperand(0).hasOneUse()) {
25535 SDValue N00 = N0.getOperand(0);
25536 if (N00.getOpcode() == X86ISD::SETCC_CARRY) {
25537 return DAG.getNode(ISD::AND, dl, VT,
25538 DAG.getNode(X86ISD::SETCC_CARRY, dl, VT,
25539 N00.getOperand(0), N00.getOperand(1)),
25540 DAG.getConstant(1, dl, VT));
25544 if (VT.is256BitVector())
25545 if (SDValue R = WidenMaskArithmetic(N, DAG, DCI, Subtarget))
25548 // (i8,i32 zext (udivrem (i8 x, i8 y)) ->
25549 // (i8,i32 (udivrem_zext_hreg (i8 x, i8 y)
25550 // This exposes the zext to the udivrem lowering, so that it directly extends
25551 // from AH (which we otherwise need to do contortions to access).
25552 if (N0.getOpcode() == ISD::UDIVREM &&
25553 N0.getResNo() == 1 && N0.getValueType() == MVT::i8 &&
25554 (VT == MVT::i32 || VT == MVT::i64)) {
25555 SDVTList NodeTys = DAG.getVTList(MVT::i8, VT);
25556 SDValue R = DAG.getNode(X86ISD::UDIVREM8_ZEXT_HREG, dl, NodeTys,
25557 N0.getOperand(0), N0.getOperand(1));
25558 DAG.ReplaceAllUsesOfValueWith(N0.getValue(0), R.getValue(0));
25559 return R.getValue(1);
25565 // Optimize x == -y --> x+y == 0
25566 // x != -y --> x+y != 0
25567 static SDValue PerformISDSETCCCombine(SDNode *N, SelectionDAG &DAG,
25568 const X86Subtarget* Subtarget) {
25569 ISD::CondCode CC = cast<CondCodeSDNode>(N->getOperand(2))->get();
25570 SDValue LHS = N->getOperand(0);
25571 SDValue RHS = N->getOperand(1);
25572 EVT VT = N->getValueType(0);
25575 if ((CC == ISD::SETNE || CC == ISD::SETEQ) && LHS.getOpcode() == ISD::SUB)
25576 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(LHS.getOperand(0)))
25577 if (C->getAPIntValue() == 0 && LHS.hasOneUse()) {
25578 SDValue addV = DAG.getNode(ISD::ADD, DL, LHS.getValueType(), RHS,
25579 LHS.getOperand(1));
25580 return DAG.getSetCC(DL, N->getValueType(0), addV,
25581 DAG.getConstant(0, DL, addV.getValueType()), CC);
25583 if ((CC == ISD::SETNE || CC == ISD::SETEQ) && RHS.getOpcode() == ISD::SUB)
25584 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(RHS.getOperand(0)))
25585 if (C->getAPIntValue() == 0 && RHS.hasOneUse()) {
25586 SDValue addV = DAG.getNode(ISD::ADD, DL, RHS.getValueType(), LHS,
25587 RHS.getOperand(1));
25588 return DAG.getSetCC(DL, N->getValueType(0), addV,
25589 DAG.getConstant(0, DL, addV.getValueType()), CC);
25592 if (VT.getScalarType() == MVT::i1 &&
25593 (CC == ISD::SETNE || CC == ISD::SETEQ || ISD::isSignedIntSetCC(CC))) {
25595 (LHS.getOpcode() == ISD::SIGN_EXTEND) &&
25596 (LHS.getOperand(0).getValueType().getScalarType() == MVT::i1);
25597 bool IsVZero1 = ISD::isBuildVectorAllZeros(RHS.getNode());
25599 if (!IsSEXT0 || !IsVZero1) {
25600 // Swap the operands and update the condition code.
25601 std::swap(LHS, RHS);
25602 CC = ISD::getSetCCSwappedOperands(CC);
25604 IsSEXT0 = (LHS.getOpcode() == ISD::SIGN_EXTEND) &&
25605 (LHS.getOperand(0).getValueType().getScalarType() == MVT::i1);
25606 IsVZero1 = ISD::isBuildVectorAllZeros(RHS.getNode());
25609 if (IsSEXT0 && IsVZero1) {
25610 assert(VT == LHS.getOperand(0).getValueType() &&
25611 "Uexpected operand type");
25612 if (CC == ISD::SETGT)
25613 return DAG.getConstant(0, DL, VT);
25614 if (CC == ISD::SETLE)
25615 return DAG.getConstant(1, DL, VT);
25616 if (CC == ISD::SETEQ || CC == ISD::SETGE)
25617 return DAG.getNOT(DL, LHS.getOperand(0), VT);
25619 assert((CC == ISD::SETNE || CC == ISD::SETLT) &&
25620 "Unexpected condition code!");
25621 return LHS.getOperand(0);
25628 static SDValue NarrowVectorLoadToElement(LoadSDNode *Load, unsigned Index,
25629 SelectionDAG &DAG) {
25631 MVT VT = Load->getSimpleValueType(0);
25632 MVT EVT = VT.getVectorElementType();
25633 SDValue Addr = Load->getOperand(1);
25634 SDValue NewAddr = DAG.getNode(
25635 ISD::ADD, dl, Addr.getSimpleValueType(), Addr,
25636 DAG.getConstant(Index * EVT.getStoreSize(), dl,
25637 Addr.getSimpleValueType()));
25640 DAG.getLoad(EVT, dl, Load->getChain(), NewAddr,
25641 DAG.getMachineFunction().getMachineMemOperand(
25642 Load->getMemOperand(), 0, EVT.getStoreSize()));
25646 static SDValue PerformINSERTPSCombine(SDNode *N, SelectionDAG &DAG,
25647 const X86Subtarget *Subtarget) {
25649 MVT VT = N->getOperand(1)->getSimpleValueType(0);
25650 assert((VT == MVT::v4f32 || VT == MVT::v4i32) &&
25651 "X86insertps is only defined for v4x32");
25653 SDValue Ld = N->getOperand(1);
25654 if (MayFoldLoad(Ld)) {
25655 // Extract the countS bits from the immediate so we can get the proper
25656 // address when narrowing the vector load to a specific element.
25657 // When the second source op is a memory address, insertps doesn't use
25658 // countS and just gets an f32 from that address.
25659 unsigned DestIndex =
25660 cast<ConstantSDNode>(N->getOperand(2))->getZExtValue() >> 6;
25662 Ld = NarrowVectorLoadToElement(cast<LoadSDNode>(Ld), DestIndex, DAG);
25664 // Create this as a scalar to vector to match the instruction pattern.
25665 SDValue LoadScalarToVector = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Ld);
25666 // countS bits are ignored when loading from memory on insertps, which
25667 // means we don't need to explicitly set them to 0.
25668 return DAG.getNode(X86ISD::INSERTPS, dl, VT, N->getOperand(0),
25669 LoadScalarToVector, N->getOperand(2));
25674 static SDValue PerformBLENDICombine(SDNode *N, SelectionDAG &DAG) {
25675 SDValue V0 = N->getOperand(0);
25676 SDValue V1 = N->getOperand(1);
25678 EVT VT = N->getValueType(0);
25680 // Canonicalize a v2f64 blend with a mask of 2 by swapping the vector
25681 // operands and changing the mask to 1. This saves us a bunch of
25682 // pattern-matching possibilities related to scalar math ops in SSE/AVX.
25683 // x86InstrInfo knows how to commute this back after instruction selection
25684 // if it would help register allocation.
25686 // TODO: If optimizing for size or a processor that doesn't suffer from
25687 // partial register update stalls, this should be transformed into a MOVSD
25688 // instruction because a MOVSD is 1-2 bytes smaller than a BLENDPD.
25690 if (VT == MVT::v2f64)
25691 if (auto *Mask = dyn_cast<ConstantSDNode>(N->getOperand(2)))
25692 if (Mask->getZExtValue() == 2 && !isShuffleFoldableLoad(V0)) {
25693 SDValue NewMask = DAG.getConstant(1, DL, MVT::i8);
25694 return DAG.getNode(X86ISD::BLENDI, DL, VT, V1, V0, NewMask);
25700 // Helper function of PerformSETCCCombine. It is to materialize "setb reg"
25701 // as "sbb reg,reg", since it can be extended without zext and produces
25702 // an all-ones bit which is more useful than 0/1 in some cases.
25703 static SDValue MaterializeSETB(SDLoc DL, SDValue EFLAGS, SelectionDAG &DAG,
25706 return DAG.getNode(ISD::AND, DL, VT,
25707 DAG.getNode(X86ISD::SETCC_CARRY, DL, MVT::i8,
25708 DAG.getConstant(X86::COND_B, DL, MVT::i8),
25710 DAG.getConstant(1, DL, VT));
25711 assert (VT == MVT::i1 && "Unexpected type for SECCC node");
25712 return DAG.getNode(ISD::TRUNCATE, DL, MVT::i1,
25713 DAG.getNode(X86ISD::SETCC_CARRY, DL, MVT::i8,
25714 DAG.getConstant(X86::COND_B, DL, MVT::i8),
25718 // Optimize RES = X86ISD::SETCC CONDCODE, EFLAG_INPUT
25719 static SDValue PerformSETCCCombine(SDNode *N, SelectionDAG &DAG,
25720 TargetLowering::DAGCombinerInfo &DCI,
25721 const X86Subtarget *Subtarget) {
25723 X86::CondCode CC = X86::CondCode(N->getConstantOperandVal(0));
25724 SDValue EFLAGS = N->getOperand(1);
25726 if (CC == X86::COND_A) {
25727 // Try to convert COND_A into COND_B in an attempt to facilitate
25728 // materializing "setb reg".
25730 // Do not flip "e > c", where "c" is a constant, because Cmp instruction
25731 // cannot take an immediate as its first operand.
25733 if (EFLAGS.getOpcode() == X86ISD::SUB && EFLAGS.hasOneUse() &&
25734 EFLAGS.getValueType().isInteger() &&
25735 !isa<ConstantSDNode>(EFLAGS.getOperand(1))) {
25736 SDValue NewSub = DAG.getNode(X86ISD::SUB, SDLoc(EFLAGS),
25737 EFLAGS.getNode()->getVTList(),
25738 EFLAGS.getOperand(1), EFLAGS.getOperand(0));
25739 SDValue NewEFLAGS = SDValue(NewSub.getNode(), EFLAGS.getResNo());
25740 return MaterializeSETB(DL, NewEFLAGS, DAG, N->getSimpleValueType(0));
25744 // Materialize "setb reg" as "sbb reg,reg", since it can be extended without
25745 // a zext and produces an all-ones bit which is more useful than 0/1 in some
25747 if (CC == X86::COND_B)
25748 return MaterializeSETB(DL, EFLAGS, DAG, N->getSimpleValueType(0));
25750 if (SDValue Flags = checkBoolTestSetCCCombine(EFLAGS, CC)) {
25751 SDValue Cond = DAG.getConstant(CC, DL, MVT::i8);
25752 return DAG.getNode(X86ISD::SETCC, DL, N->getVTList(), Cond, Flags);
25758 // Optimize branch condition evaluation.
25760 static SDValue PerformBrCondCombine(SDNode *N, SelectionDAG &DAG,
25761 TargetLowering::DAGCombinerInfo &DCI,
25762 const X86Subtarget *Subtarget) {
25764 SDValue Chain = N->getOperand(0);
25765 SDValue Dest = N->getOperand(1);
25766 SDValue EFLAGS = N->getOperand(3);
25767 X86::CondCode CC = X86::CondCode(N->getConstantOperandVal(2));
25769 if (SDValue Flags = checkBoolTestSetCCCombine(EFLAGS, CC)) {
25770 SDValue Cond = DAG.getConstant(CC, DL, MVT::i8);
25771 return DAG.getNode(X86ISD::BRCOND, DL, N->getVTList(), Chain, Dest, Cond,
25778 static SDValue performVectorCompareAndMaskUnaryOpCombine(SDNode *N,
25779 SelectionDAG &DAG) {
25780 // Take advantage of vector comparisons producing 0 or -1 in each lane to
25781 // optimize away operation when it's from a constant.
25783 // The general transformation is:
25784 // UNARYOP(AND(VECTOR_CMP(x,y), constant)) -->
25785 // AND(VECTOR_CMP(x,y), constant2)
25786 // constant2 = UNARYOP(constant)
25788 // Early exit if this isn't a vector operation, the operand of the
25789 // unary operation isn't a bitwise AND, or if the sizes of the operations
25790 // aren't the same.
25791 EVT VT = N->getValueType(0);
25792 if (!VT.isVector() || N->getOperand(0)->getOpcode() != ISD::AND ||
25793 N->getOperand(0)->getOperand(0)->getOpcode() != ISD::SETCC ||
25794 VT.getSizeInBits() != N->getOperand(0)->getValueType(0).getSizeInBits())
25797 // Now check that the other operand of the AND is a constant. We could
25798 // make the transformation for non-constant splats as well, but it's unclear
25799 // that would be a benefit as it would not eliminate any operations, just
25800 // perform one more step in scalar code before moving to the vector unit.
25801 if (BuildVectorSDNode *BV =
25802 dyn_cast<BuildVectorSDNode>(N->getOperand(0)->getOperand(1))) {
25803 // Bail out if the vector isn't a constant.
25804 if (!BV->isConstant())
25807 // Everything checks out. Build up the new and improved node.
25809 EVT IntVT = BV->getValueType(0);
25810 // Create a new constant of the appropriate type for the transformed
25812 SDValue SourceConst = DAG.getNode(N->getOpcode(), DL, VT, SDValue(BV, 0));
25813 // The AND node needs bitcasts to/from an integer vector type around it.
25814 SDValue MaskConst = DAG.getBitcast(IntVT, SourceConst);
25815 SDValue NewAnd = DAG.getNode(ISD::AND, DL, IntVT,
25816 N->getOperand(0)->getOperand(0), MaskConst);
25817 SDValue Res = DAG.getBitcast(VT, NewAnd);
25824 static SDValue PerformUINT_TO_FPCombine(SDNode *N, SelectionDAG &DAG,
25825 const X86Subtarget *Subtarget) {
25826 SDValue Op0 = N->getOperand(0);
25827 EVT VT = N->getValueType(0);
25828 EVT InVT = Op0.getValueType();
25829 EVT InSVT = InVT.getScalarType();
25830 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
25832 // UINT_TO_FP(vXi8) -> SINT_TO_FP(ZEXT(vXi8 to vXi32))
25833 // UINT_TO_FP(vXi16) -> SINT_TO_FP(ZEXT(vXi16 to vXi32))
25834 if (InVT.isVector() && (InSVT == MVT::i8 || InSVT == MVT::i16)) {
25836 EVT DstVT = EVT::getVectorVT(*DAG.getContext(), MVT::i32,
25837 InVT.getVectorNumElements());
25838 SDValue P = DAG.getNode(ISD::ZERO_EXTEND, dl, DstVT, Op0);
25840 if (TLI.isOperationLegal(ISD::UINT_TO_FP, DstVT))
25841 return DAG.getNode(ISD::UINT_TO_FP, dl, VT, P);
25843 return DAG.getNode(ISD::SINT_TO_FP, dl, VT, P);
25849 static SDValue PerformSINT_TO_FPCombine(SDNode *N, SelectionDAG &DAG,
25850 const X86Subtarget *Subtarget) {
25851 // First try to optimize away the conversion entirely when it's
25852 // conditionally from a constant. Vectors only.
25853 if (SDValue Res = performVectorCompareAndMaskUnaryOpCombine(N, DAG))
25856 // Now move on to more general possibilities.
25857 SDValue Op0 = N->getOperand(0);
25858 EVT VT = N->getValueType(0);
25859 EVT InVT = Op0.getValueType();
25860 EVT InSVT = InVT.getScalarType();
25862 // SINT_TO_FP(vXi8) -> SINT_TO_FP(SEXT(vXi8 to vXi32))
25863 // SINT_TO_FP(vXi16) -> SINT_TO_FP(SEXT(vXi16 to vXi32))
25864 if (InVT.isVector() && (InSVT == MVT::i8 || InSVT == MVT::i16)) {
25866 EVT DstVT = EVT::getVectorVT(*DAG.getContext(), MVT::i32,
25867 InVT.getVectorNumElements());
25868 SDValue P = DAG.getNode(ISD::SIGN_EXTEND, dl, DstVT, Op0);
25869 return DAG.getNode(ISD::SINT_TO_FP, dl, VT, P);
25872 // Transform (SINT_TO_FP (i64 ...)) into an x87 operation if we have
25873 // a 32-bit target where SSE doesn't support i64->FP operations.
25874 if (Op0.getOpcode() == ISD::LOAD) {
25875 LoadSDNode *Ld = cast<LoadSDNode>(Op0.getNode());
25876 EVT LdVT = Ld->getValueType(0);
25878 // This transformation is not supported if the result type is f16
25879 if (VT == MVT::f16)
25882 if (!Ld->isVolatile() && !VT.isVector() &&
25883 ISD::isNON_EXTLoad(Op0.getNode()) && Op0.hasOneUse() &&
25884 !Subtarget->is64Bit() && LdVT == MVT::i64) {
25885 SDValue FILDChain = Subtarget->getTargetLowering()->BuildFILD(
25886 SDValue(N, 0), LdVT, Ld->getChain(), Op0, DAG);
25887 DAG.ReplaceAllUsesOfValueWith(Op0.getValue(1), FILDChain.getValue(1));
25894 // Optimize RES, EFLAGS = X86ISD::ADC LHS, RHS, EFLAGS
25895 static SDValue PerformADCCombine(SDNode *N, SelectionDAG &DAG,
25896 X86TargetLowering::DAGCombinerInfo &DCI) {
25897 // If the LHS and RHS of the ADC node are zero, then it can't overflow and
25898 // the result is either zero or one (depending on the input carry bit).
25899 // Strength reduce this down to a "set on carry" aka SETCC_CARRY&1.
25900 if (X86::isZeroNode(N->getOperand(0)) &&
25901 X86::isZeroNode(N->getOperand(1)) &&
25902 // We don't have a good way to replace an EFLAGS use, so only do this when
25904 SDValue(N, 1).use_empty()) {
25906 EVT VT = N->getValueType(0);
25907 SDValue CarryOut = DAG.getConstant(0, DL, N->getValueType(1));
25908 SDValue Res1 = DAG.getNode(ISD::AND, DL, VT,
25909 DAG.getNode(X86ISD::SETCC_CARRY, DL, VT,
25910 DAG.getConstant(X86::COND_B, DL,
25913 DAG.getConstant(1, DL, VT));
25914 return DCI.CombineTo(N, Res1, CarryOut);
25920 // fold (add Y, (sete X, 0)) -> adc 0, Y
25921 // (add Y, (setne X, 0)) -> sbb -1, Y
25922 // (sub (sete X, 0), Y) -> sbb 0, Y
25923 // (sub (setne X, 0), Y) -> adc -1, Y
25924 static SDValue OptimizeConditionalInDecrement(SDNode *N, SelectionDAG &DAG) {
25927 // Look through ZExts.
25928 SDValue Ext = N->getOperand(N->getOpcode() == ISD::SUB ? 1 : 0);
25929 if (Ext.getOpcode() != ISD::ZERO_EXTEND || !Ext.hasOneUse())
25932 SDValue SetCC = Ext.getOperand(0);
25933 if (SetCC.getOpcode() != X86ISD::SETCC || !SetCC.hasOneUse())
25936 X86::CondCode CC = (X86::CondCode)SetCC.getConstantOperandVal(0);
25937 if (CC != X86::COND_E && CC != X86::COND_NE)
25940 SDValue Cmp = SetCC.getOperand(1);
25941 if (Cmp.getOpcode() != X86ISD::CMP || !Cmp.hasOneUse() ||
25942 !X86::isZeroNode(Cmp.getOperand(1)) ||
25943 !Cmp.getOperand(0).getValueType().isInteger())
25946 SDValue CmpOp0 = Cmp.getOperand(0);
25947 SDValue NewCmp = DAG.getNode(X86ISD::CMP, DL, MVT::i32, CmpOp0,
25948 DAG.getConstant(1, DL, CmpOp0.getValueType()));
25950 SDValue OtherVal = N->getOperand(N->getOpcode() == ISD::SUB ? 0 : 1);
25951 if (CC == X86::COND_NE)
25952 return DAG.getNode(N->getOpcode() == ISD::SUB ? X86ISD::ADC : X86ISD::SBB,
25953 DL, OtherVal.getValueType(), OtherVal,
25954 DAG.getConstant(-1ULL, DL, OtherVal.getValueType()),
25956 return DAG.getNode(N->getOpcode() == ISD::SUB ? X86ISD::SBB : X86ISD::ADC,
25957 DL, OtherVal.getValueType(), OtherVal,
25958 DAG.getConstant(0, DL, OtherVal.getValueType()), NewCmp);
25961 /// PerformADDCombine - Do target-specific dag combines on integer adds.
25962 static SDValue PerformAddCombine(SDNode *N, SelectionDAG &DAG,
25963 const X86Subtarget *Subtarget) {
25964 EVT VT = N->getValueType(0);
25965 SDValue Op0 = N->getOperand(0);
25966 SDValue Op1 = N->getOperand(1);
25968 // Try to synthesize horizontal adds from adds of shuffles.
25969 if (((Subtarget->hasSSSE3() && (VT == MVT::v8i16 || VT == MVT::v4i32)) ||
25970 (Subtarget->hasInt256() && (VT == MVT::v16i16 || VT == MVT::v8i32))) &&
25971 isHorizontalBinOp(Op0, Op1, true))
25972 return DAG.getNode(X86ISD::HADD, SDLoc(N), VT, Op0, Op1);
25974 return OptimizeConditionalInDecrement(N, DAG);
25977 static SDValue PerformSubCombine(SDNode *N, SelectionDAG &DAG,
25978 const X86Subtarget *Subtarget) {
25979 SDValue Op0 = N->getOperand(0);
25980 SDValue Op1 = N->getOperand(1);
25982 // X86 can't encode an immediate LHS of a sub. See if we can push the
25983 // negation into a preceding instruction.
25984 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op0)) {
25985 // If the RHS of the sub is a XOR with one use and a constant, invert the
25986 // immediate. Then add one to the LHS of the sub so we can turn
25987 // X-Y -> X+~Y+1, saving one register.
25988 if (Op1->hasOneUse() && Op1.getOpcode() == ISD::XOR &&
25989 isa<ConstantSDNode>(Op1.getOperand(1))) {
25990 APInt XorC = cast<ConstantSDNode>(Op1.getOperand(1))->getAPIntValue();
25991 EVT VT = Op0.getValueType();
25992 SDValue NewXor = DAG.getNode(ISD::XOR, SDLoc(Op1), VT,
25994 DAG.getConstant(~XorC, SDLoc(Op1), VT));
25995 return DAG.getNode(ISD::ADD, SDLoc(N), VT, NewXor,
25996 DAG.getConstant(C->getAPIntValue() + 1, SDLoc(N), VT));
26000 // Try to synthesize horizontal adds from adds of shuffles.
26001 EVT VT = N->getValueType(0);
26002 if (((Subtarget->hasSSSE3() && (VT == MVT::v8i16 || VT == MVT::v4i32)) ||
26003 (Subtarget->hasInt256() && (VT == MVT::v16i16 || VT == MVT::v8i32))) &&
26004 isHorizontalBinOp(Op0, Op1, true))
26005 return DAG.getNode(X86ISD::HSUB, SDLoc(N), VT, Op0, Op1);
26007 return OptimizeConditionalInDecrement(N, DAG);
26010 /// performVZEXTCombine - Performs build vector combines
26011 static SDValue performVZEXTCombine(SDNode *N, SelectionDAG &DAG,
26012 TargetLowering::DAGCombinerInfo &DCI,
26013 const X86Subtarget *Subtarget) {
26015 MVT VT = N->getSimpleValueType(0);
26016 SDValue Op = N->getOperand(0);
26017 MVT OpVT = Op.getSimpleValueType();
26018 MVT OpEltVT = OpVT.getVectorElementType();
26019 unsigned InputBits = OpEltVT.getSizeInBits() * VT.getVectorNumElements();
26021 // (vzext (bitcast (vzext (x)) -> (vzext x)
26023 while (V.getOpcode() == ISD::BITCAST)
26024 V = V.getOperand(0);
26026 if (V != Op && V.getOpcode() == X86ISD::VZEXT) {
26027 MVT InnerVT = V.getSimpleValueType();
26028 MVT InnerEltVT = InnerVT.getVectorElementType();
26030 // If the element sizes match exactly, we can just do one larger vzext. This
26031 // is always an exact type match as vzext operates on integer types.
26032 if (OpEltVT == InnerEltVT) {
26033 assert(OpVT == InnerVT && "Types must match for vzext!");
26034 return DAG.getNode(X86ISD::VZEXT, DL, VT, V.getOperand(0));
26037 // The only other way we can combine them is if only a single element of the
26038 // inner vzext is used in the input to the outer vzext.
26039 if (InnerEltVT.getSizeInBits() < InputBits)
26042 // In this case, the inner vzext is completely dead because we're going to
26043 // only look at bits inside of the low element. Just do the outer vzext on
26044 // a bitcast of the input to the inner.
26045 return DAG.getNode(X86ISD::VZEXT, DL, VT, DAG.getBitcast(OpVT, V));
26048 // Check if we can bypass extracting and re-inserting an element of an input
26049 // vector. Essentially:
26050 // (bitcast (sclr2vec (ext_vec_elt x))) -> (bitcast x)
26051 if (V.getOpcode() == ISD::SCALAR_TO_VECTOR &&
26052 V.getOperand(0).getOpcode() == ISD::EXTRACT_VECTOR_ELT &&
26053 V.getOperand(0).getSimpleValueType().getSizeInBits() == InputBits) {
26054 SDValue ExtractedV = V.getOperand(0);
26055 SDValue OrigV = ExtractedV.getOperand(0);
26056 if (auto *ExtractIdx = dyn_cast<ConstantSDNode>(ExtractedV.getOperand(1)))
26057 if (ExtractIdx->getZExtValue() == 0) {
26058 MVT OrigVT = OrigV.getSimpleValueType();
26059 // Extract a subvector if necessary...
26060 if (OrigVT.getSizeInBits() > OpVT.getSizeInBits()) {
26061 int Ratio = OrigVT.getSizeInBits() / OpVT.getSizeInBits();
26062 OrigVT = MVT::getVectorVT(OrigVT.getVectorElementType(),
26063 OrigVT.getVectorNumElements() / Ratio);
26064 OrigV = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, OrigVT, OrigV,
26065 DAG.getIntPtrConstant(0, DL));
26067 Op = DAG.getBitcast(OpVT, OrigV);
26068 return DAG.getNode(X86ISD::VZEXT, DL, VT, Op);
26075 SDValue X86TargetLowering::PerformDAGCombine(SDNode *N,
26076 DAGCombinerInfo &DCI) const {
26077 SelectionDAG &DAG = DCI.DAG;
26078 switch (N->getOpcode()) {
26080 case ISD::EXTRACT_VECTOR_ELT:
26081 return PerformEXTRACT_VECTOR_ELTCombine(N, DAG, DCI);
26084 case X86ISD::SHRUNKBLEND:
26085 return PerformSELECTCombine(N, DAG, DCI, Subtarget);
26086 case ISD::BITCAST: return PerformBITCASTCombine(N, DAG);
26087 case X86ISD::CMOV: return PerformCMOVCombine(N, DAG, DCI, Subtarget);
26088 case ISD::ADD: return PerformAddCombine(N, DAG, Subtarget);
26089 case ISD::SUB: return PerformSubCombine(N, DAG, Subtarget);
26090 case X86ISD::ADC: return PerformADCCombine(N, DAG, DCI);
26091 case ISD::MUL: return PerformMulCombine(N, DAG, DCI);
26094 case ISD::SRL: return PerformShiftCombine(N, DAG, DCI, Subtarget);
26095 case ISD::AND: return PerformAndCombine(N, DAG, DCI, Subtarget);
26096 case ISD::OR: return PerformOrCombine(N, DAG, DCI, Subtarget);
26097 case ISD::XOR: return PerformXorCombine(N, DAG, DCI, Subtarget);
26098 case ISD::LOAD: return PerformLOADCombine(N, DAG, DCI, Subtarget);
26099 case ISD::MLOAD: return PerformMLOADCombine(N, DAG, DCI, Subtarget);
26100 case ISD::STORE: return PerformSTORECombine(N, DAG, Subtarget);
26101 case ISD::MSTORE: return PerformMSTORECombine(N, DAG, Subtarget);
26102 case ISD::SINT_TO_FP: return PerformSINT_TO_FPCombine(N, DAG, Subtarget);
26103 case ISD::UINT_TO_FP: return PerformUINT_TO_FPCombine(N, DAG, Subtarget);
26104 case ISD::FADD: return PerformFADDCombine(N, DAG, Subtarget);
26105 case ISD::FSUB: return PerformFSUBCombine(N, DAG, Subtarget);
26107 case X86ISD::FOR: return PerformFORCombine(N, DAG, Subtarget);
26109 case X86ISD::FMAX: return PerformFMinFMaxCombine(N, DAG);
26110 case X86ISD::FAND: return PerformFANDCombine(N, DAG);
26111 case X86ISD::FANDN: return PerformFANDNCombine(N, DAG);
26112 case X86ISD::BT: return PerformBTCombine(N, DAG, DCI);
26113 case X86ISD::VZEXT_MOVL: return PerformVZEXT_MOVLCombine(N, DAG);
26114 case ISD::ANY_EXTEND:
26115 case ISD::ZERO_EXTEND: return PerformZExtCombine(N, DAG, DCI, Subtarget);
26116 case ISD::SIGN_EXTEND: return PerformSExtCombine(N, DAG, DCI, Subtarget);
26117 case ISD::SIGN_EXTEND_INREG:
26118 return PerformSIGN_EXTEND_INREGCombine(N, DAG, Subtarget);
26119 case ISD::SETCC: return PerformISDSETCCCombine(N, DAG, Subtarget);
26120 case X86ISD::SETCC: return PerformSETCCCombine(N, DAG, DCI, Subtarget);
26121 case X86ISD::BRCOND: return PerformBrCondCombine(N, DAG, DCI, Subtarget);
26122 case X86ISD::VZEXT: return performVZEXTCombine(N, DAG, DCI, Subtarget);
26123 case X86ISD::SHUFP: // Handle all target specific shuffles
26124 case X86ISD::PALIGNR:
26125 case X86ISD::UNPCKH:
26126 case X86ISD::UNPCKL:
26127 case X86ISD::MOVHLPS:
26128 case X86ISD::MOVLHPS:
26129 case X86ISD::PSHUFB:
26130 case X86ISD::PSHUFD:
26131 case X86ISD::PSHUFHW:
26132 case X86ISD::PSHUFLW:
26133 case X86ISD::MOVSS:
26134 case X86ISD::MOVSD:
26135 case X86ISD::VPERMILPI:
26136 case X86ISD::VPERM2X128:
26137 case ISD::VECTOR_SHUFFLE: return PerformShuffleCombine(N, DAG, DCI,Subtarget);
26138 case ISD::FMA: return PerformFMACombine(N, DAG, Subtarget);
26139 case X86ISD::INSERTPS: {
26140 if (getTargetMachine().getOptLevel() > CodeGenOpt::None)
26141 return PerformINSERTPSCombine(N, DAG, Subtarget);
26144 case X86ISD::BLENDI: return PerformBLENDICombine(N, DAG);
26150 /// isTypeDesirableForOp - Return true if the target has native support for
26151 /// the specified value type and it is 'desirable' to use the type for the
26152 /// given node type. e.g. On x86 i16 is legal, but undesirable since i16
26153 /// instruction encodings are longer and some i16 instructions are slow.
26154 bool X86TargetLowering::isTypeDesirableForOp(unsigned Opc, EVT VT) const {
26155 if (!isTypeLegal(VT))
26157 if (VT != MVT::i16)
26164 case ISD::SIGN_EXTEND:
26165 case ISD::ZERO_EXTEND:
26166 case ISD::ANY_EXTEND:
26179 /// IsDesirableToPromoteOp - This method query the target whether it is
26180 /// beneficial for dag combiner to promote the specified node. If true, it
26181 /// should return the desired promotion type by reference.
26182 bool X86TargetLowering::IsDesirableToPromoteOp(SDValue Op, EVT &PVT) const {
26183 EVT VT = Op.getValueType();
26184 if (VT != MVT::i16)
26187 bool Promote = false;
26188 bool Commute = false;
26189 switch (Op.getOpcode()) {
26192 LoadSDNode *LD = cast<LoadSDNode>(Op);
26193 // If the non-extending load has a single use and it's not live out, then it
26194 // might be folded.
26195 if (LD->getExtensionType() == ISD::NON_EXTLOAD /*&&
26196 Op.hasOneUse()*/) {
26197 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
26198 UE = Op.getNode()->use_end(); UI != UE; ++UI) {
26199 // The only case where we'd want to promote LOAD (rather then it being
26200 // promoted as an operand is when it's only use is liveout.
26201 if (UI->getOpcode() != ISD::CopyToReg)
26208 case ISD::SIGN_EXTEND:
26209 case ISD::ZERO_EXTEND:
26210 case ISD::ANY_EXTEND:
26215 SDValue N0 = Op.getOperand(0);
26216 // Look out for (store (shl (load), x)).
26217 if (MayFoldLoad(N0) && MayFoldIntoStore(Op))
26230 SDValue N0 = Op.getOperand(0);
26231 SDValue N1 = Op.getOperand(1);
26232 if (!Commute && MayFoldLoad(N1))
26234 // Avoid disabling potential load folding opportunities.
26235 if (MayFoldLoad(N0) && (!isa<ConstantSDNode>(N1) || MayFoldIntoStore(Op)))
26237 if (MayFoldLoad(N1) && (!isa<ConstantSDNode>(N0) || MayFoldIntoStore(Op)))
26247 //===----------------------------------------------------------------------===//
26248 // X86 Inline Assembly Support
26249 //===----------------------------------------------------------------------===//
26251 // Helper to match a string separated by whitespace.
26252 static bool matchAsm(StringRef S, ArrayRef<const char *> Pieces) {
26253 S = S.substr(S.find_first_not_of(" \t")); // Skip leading whitespace.
26255 for (StringRef Piece : Pieces) {
26256 if (!S.startswith(Piece)) // Check if the piece matches.
26259 S = S.substr(Piece.size());
26260 StringRef::size_type Pos = S.find_first_not_of(" \t");
26261 if (Pos == 0) // We matched a prefix.
26270 static bool clobbersFlagRegisters(const SmallVector<StringRef, 4> &AsmPieces) {
26272 if (AsmPieces.size() == 3 || AsmPieces.size() == 4) {
26273 if (std::count(AsmPieces.begin(), AsmPieces.end(), "~{cc}") &&
26274 std::count(AsmPieces.begin(), AsmPieces.end(), "~{flags}") &&
26275 std::count(AsmPieces.begin(), AsmPieces.end(), "~{fpsr}")) {
26277 if (AsmPieces.size() == 3)
26279 else if (std::count(AsmPieces.begin(), AsmPieces.end(), "~{dirflag}"))
26286 bool X86TargetLowering::ExpandInlineAsm(CallInst *CI) const {
26287 InlineAsm *IA = cast<InlineAsm>(CI->getCalledValue());
26289 std::string AsmStr = IA->getAsmString();
26291 IntegerType *Ty = dyn_cast<IntegerType>(CI->getType());
26292 if (!Ty || Ty->getBitWidth() % 16 != 0)
26295 // TODO: should remove alternatives from the asmstring: "foo {a|b}" -> "foo a"
26296 SmallVector<StringRef, 4> AsmPieces;
26297 SplitString(AsmStr, AsmPieces, ";\n");
26299 switch (AsmPieces.size()) {
26300 default: return false;
26302 // FIXME: this should verify that we are targeting a 486 or better. If not,
26303 // we will turn this bswap into something that will be lowered to logical
26304 // ops instead of emitting the bswap asm. For now, we don't support 486 or
26305 // lower so don't worry about this.
26307 if (matchAsm(AsmPieces[0], {"bswap", "$0"}) ||
26308 matchAsm(AsmPieces[0], {"bswapl", "$0"}) ||
26309 matchAsm(AsmPieces[0], {"bswapq", "$0"}) ||
26310 matchAsm(AsmPieces[0], {"bswap", "${0:q}"}) ||
26311 matchAsm(AsmPieces[0], {"bswapl", "${0:q}"}) ||
26312 matchAsm(AsmPieces[0], {"bswapq", "${0:q}"})) {
26313 // No need to check constraints, nothing other than the equivalent of
26314 // "=r,0" would be valid here.
26315 return IntrinsicLowering::LowerToByteSwap(CI);
26318 // rorw $$8, ${0:w} --> llvm.bswap.i16
26319 if (CI->getType()->isIntegerTy(16) &&
26320 IA->getConstraintString().compare(0, 5, "=r,0,") == 0 &&
26321 (matchAsm(AsmPieces[0], {"rorw", "$$8,", "${0:w}"}) ||
26322 matchAsm(AsmPieces[0], {"rolw", "$$8,", "${0:w}"}))) {
26324 StringRef ConstraintsStr = IA->getConstraintString();
26325 SplitString(StringRef(ConstraintsStr).substr(5), AsmPieces, ",");
26326 array_pod_sort(AsmPieces.begin(), AsmPieces.end());
26327 if (clobbersFlagRegisters(AsmPieces))
26328 return IntrinsicLowering::LowerToByteSwap(CI);
26332 if (CI->getType()->isIntegerTy(32) &&
26333 IA->getConstraintString().compare(0, 5, "=r,0,") == 0 &&
26334 matchAsm(AsmPieces[0], {"rorw", "$$8,", "${0:w}"}) &&
26335 matchAsm(AsmPieces[1], {"rorl", "$$16,", "$0"}) &&
26336 matchAsm(AsmPieces[2], {"rorw", "$$8,", "${0:w}"})) {
26338 StringRef ConstraintsStr = IA->getConstraintString();
26339 SplitString(StringRef(ConstraintsStr).substr(5), AsmPieces, ",");
26340 array_pod_sort(AsmPieces.begin(), AsmPieces.end());
26341 if (clobbersFlagRegisters(AsmPieces))
26342 return IntrinsicLowering::LowerToByteSwap(CI);
26345 if (CI->getType()->isIntegerTy(64)) {
26346 InlineAsm::ConstraintInfoVector Constraints = IA->ParseConstraints();
26347 if (Constraints.size() >= 2 &&
26348 Constraints[0].Codes.size() == 1 && Constraints[0].Codes[0] == "A" &&
26349 Constraints[1].Codes.size() == 1 && Constraints[1].Codes[0] == "0") {
26350 // bswap %eax / bswap %edx / xchgl %eax, %edx -> llvm.bswap.i64
26351 if (matchAsm(AsmPieces[0], {"bswap", "%eax"}) &&
26352 matchAsm(AsmPieces[1], {"bswap", "%edx"}) &&
26353 matchAsm(AsmPieces[2], {"xchgl", "%eax,", "%edx"}))
26354 return IntrinsicLowering::LowerToByteSwap(CI);
26362 /// getConstraintType - Given a constraint letter, return the type of
26363 /// constraint it is for this target.
26364 X86TargetLowering::ConstraintType
26365 X86TargetLowering::getConstraintType(StringRef Constraint) const {
26366 if (Constraint.size() == 1) {
26367 switch (Constraint[0]) {
26378 return C_RegisterClass;
26402 return TargetLowering::getConstraintType(Constraint);
26405 /// Examine constraint type and operand type and determine a weight value.
26406 /// This object must already have been set up with the operand type
26407 /// and the current alternative constraint selected.
26408 TargetLowering::ConstraintWeight
26409 X86TargetLowering::getSingleConstraintMatchWeight(
26410 AsmOperandInfo &info, const char *constraint) const {
26411 ConstraintWeight weight = CW_Invalid;
26412 Value *CallOperandVal = info.CallOperandVal;
26413 // If we don't have a value, we can't do a match,
26414 // but allow it at the lowest weight.
26415 if (!CallOperandVal)
26417 Type *type = CallOperandVal->getType();
26418 // Look at the constraint type.
26419 switch (*constraint) {
26421 weight = TargetLowering::getSingleConstraintMatchWeight(info, constraint);
26432 if (CallOperandVal->getType()->isIntegerTy())
26433 weight = CW_SpecificReg;
26438 if (type->isFloatingPointTy())
26439 weight = CW_SpecificReg;
26442 if (type->isX86_MMXTy() && Subtarget->hasMMX())
26443 weight = CW_SpecificReg;
26447 if (((type->getPrimitiveSizeInBits() == 128) && Subtarget->hasSSE1()) ||
26448 ((type->getPrimitiveSizeInBits() == 256) && Subtarget->hasFp256()))
26449 weight = CW_Register;
26452 if (ConstantInt *C = dyn_cast<ConstantInt>(info.CallOperandVal)) {
26453 if (C->getZExtValue() <= 31)
26454 weight = CW_Constant;
26458 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
26459 if (C->getZExtValue() <= 63)
26460 weight = CW_Constant;
26464 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
26465 if ((C->getSExtValue() >= -0x80) && (C->getSExtValue() <= 0x7f))
26466 weight = CW_Constant;
26470 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
26471 if ((C->getZExtValue() == 0xff) || (C->getZExtValue() == 0xffff))
26472 weight = CW_Constant;
26476 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
26477 if (C->getZExtValue() <= 3)
26478 weight = CW_Constant;
26482 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
26483 if (C->getZExtValue() <= 0xff)
26484 weight = CW_Constant;
26489 if (isa<ConstantFP>(CallOperandVal)) {
26490 weight = CW_Constant;
26494 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
26495 if ((C->getSExtValue() >= -0x80000000LL) &&
26496 (C->getSExtValue() <= 0x7fffffffLL))
26497 weight = CW_Constant;
26501 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
26502 if (C->getZExtValue() <= 0xffffffff)
26503 weight = CW_Constant;
26510 /// LowerXConstraint - try to replace an X constraint, which matches anything,
26511 /// with another that has more specific requirements based on the type of the
26512 /// corresponding operand.
26513 const char *X86TargetLowering::
26514 LowerXConstraint(EVT ConstraintVT) const {
26515 // FP X constraints get lowered to SSE1/2 registers if available, otherwise
26516 // 'f' like normal targets.
26517 if (ConstraintVT.isFloatingPoint()) {
26518 if (Subtarget->hasSSE2())
26520 if (Subtarget->hasSSE1())
26524 return TargetLowering::LowerXConstraint(ConstraintVT);
26527 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
26528 /// vector. If it is invalid, don't add anything to Ops.
26529 void X86TargetLowering::LowerAsmOperandForConstraint(SDValue Op,
26530 std::string &Constraint,
26531 std::vector<SDValue>&Ops,
26532 SelectionDAG &DAG) const {
26535 // Only support length 1 constraints for now.
26536 if (Constraint.length() > 1) return;
26538 char ConstraintLetter = Constraint[0];
26539 switch (ConstraintLetter) {
26542 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
26543 if (C->getZExtValue() <= 31) {
26544 Result = DAG.getTargetConstant(C->getZExtValue(), SDLoc(Op),
26545 Op.getValueType());
26551 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
26552 if (C->getZExtValue() <= 63) {
26553 Result = DAG.getTargetConstant(C->getZExtValue(), SDLoc(Op),
26554 Op.getValueType());
26560 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
26561 if (isInt<8>(C->getSExtValue())) {
26562 Result = DAG.getTargetConstant(C->getZExtValue(), SDLoc(Op),
26563 Op.getValueType());
26569 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
26570 if (C->getZExtValue() == 0xff || C->getZExtValue() == 0xffff ||
26571 (Subtarget->is64Bit() && C->getZExtValue() == 0xffffffff)) {
26572 Result = DAG.getTargetConstant(C->getSExtValue(), SDLoc(Op),
26573 Op.getValueType());
26579 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
26580 if (C->getZExtValue() <= 3) {
26581 Result = DAG.getTargetConstant(C->getZExtValue(), SDLoc(Op),
26582 Op.getValueType());
26588 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
26589 if (C->getZExtValue() <= 255) {
26590 Result = DAG.getTargetConstant(C->getZExtValue(), SDLoc(Op),
26591 Op.getValueType());
26597 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
26598 if (C->getZExtValue() <= 127) {
26599 Result = DAG.getTargetConstant(C->getZExtValue(), SDLoc(Op),
26600 Op.getValueType());
26606 // 32-bit signed value
26607 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
26608 if (ConstantInt::isValueValidForType(Type::getInt32Ty(*DAG.getContext()),
26609 C->getSExtValue())) {
26610 // Widen to 64 bits here to get it sign extended.
26611 Result = DAG.getTargetConstant(C->getSExtValue(), SDLoc(Op), MVT::i64);
26614 // FIXME gcc accepts some relocatable values here too, but only in certain
26615 // memory models; it's complicated.
26620 // 32-bit unsigned value
26621 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
26622 if (ConstantInt::isValueValidForType(Type::getInt32Ty(*DAG.getContext()),
26623 C->getZExtValue())) {
26624 Result = DAG.getTargetConstant(C->getZExtValue(), SDLoc(Op),
26625 Op.getValueType());
26629 // FIXME gcc accepts some relocatable values here too, but only in certain
26630 // memory models; it's complicated.
26634 // Literal immediates are always ok.
26635 if (ConstantSDNode *CST = dyn_cast<ConstantSDNode>(Op)) {
26636 // Widen to 64 bits here to get it sign extended.
26637 Result = DAG.getTargetConstant(CST->getSExtValue(), SDLoc(Op), MVT::i64);
26641 // In any sort of PIC mode addresses need to be computed at runtime by
26642 // adding in a register or some sort of table lookup. These can't
26643 // be used as immediates.
26644 if (Subtarget->isPICStyleGOT() || Subtarget->isPICStyleStubPIC())
26647 // If we are in non-pic codegen mode, we allow the address of a global (with
26648 // an optional displacement) to be used with 'i'.
26649 GlobalAddressSDNode *GA = nullptr;
26650 int64_t Offset = 0;
26652 // Match either (GA), (GA+C), (GA+C1+C2), etc.
26654 if ((GA = dyn_cast<GlobalAddressSDNode>(Op))) {
26655 Offset += GA->getOffset();
26657 } else if (Op.getOpcode() == ISD::ADD) {
26658 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
26659 Offset += C->getZExtValue();
26660 Op = Op.getOperand(0);
26663 } else if (Op.getOpcode() == ISD::SUB) {
26664 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
26665 Offset += -C->getZExtValue();
26666 Op = Op.getOperand(0);
26671 // Otherwise, this isn't something we can handle, reject it.
26675 const GlobalValue *GV = GA->getGlobal();
26676 // If we require an extra load to get this address, as in PIC mode, we
26677 // can't accept it.
26678 if (isGlobalStubReference(
26679 Subtarget->ClassifyGlobalReference(GV, DAG.getTarget())))
26682 Result = DAG.getTargetGlobalAddress(GV, SDLoc(Op),
26683 GA->getValueType(0), Offset);
26688 if (Result.getNode()) {
26689 Ops.push_back(Result);
26692 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG);
26695 std::pair<unsigned, const TargetRegisterClass *>
26696 X86TargetLowering::getRegForInlineAsmConstraint(const TargetRegisterInfo *TRI,
26697 StringRef Constraint,
26699 // First, see if this is a constraint that directly corresponds to an LLVM
26701 if (Constraint.size() == 1) {
26702 // GCC Constraint Letters
26703 switch (Constraint[0]) {
26705 // TODO: Slight differences here in allocation order and leaving
26706 // RIP in the class. Do they matter any more here than they do
26707 // in the normal allocation?
26708 case 'q': // GENERAL_REGS in 64-bit mode, Q_REGS in 32-bit mode.
26709 if (Subtarget->is64Bit()) {
26710 if (VT == MVT::i32 || VT == MVT::f32)
26711 return std::make_pair(0U, &X86::GR32RegClass);
26712 if (VT == MVT::i16)
26713 return std::make_pair(0U, &X86::GR16RegClass);
26714 if (VT == MVT::i8 || VT == MVT::i1)
26715 return std::make_pair(0U, &X86::GR8RegClass);
26716 if (VT == MVT::i64 || VT == MVT::f64)
26717 return std::make_pair(0U, &X86::GR64RegClass);
26720 // 32-bit fallthrough
26721 case 'Q': // Q_REGS
26722 if (VT == MVT::i32 || VT == MVT::f32)
26723 return std::make_pair(0U, &X86::GR32_ABCDRegClass);
26724 if (VT == MVT::i16)
26725 return std::make_pair(0U, &X86::GR16_ABCDRegClass);
26726 if (VT == MVT::i8 || VT == MVT::i1)
26727 return std::make_pair(0U, &X86::GR8_ABCD_LRegClass);
26728 if (VT == MVT::i64)
26729 return std::make_pair(0U, &X86::GR64_ABCDRegClass);
26731 case 'r': // GENERAL_REGS
26732 case 'l': // INDEX_REGS
26733 if (VT == MVT::i8 || VT == MVT::i1)
26734 return std::make_pair(0U, &X86::GR8RegClass);
26735 if (VT == MVT::i16)
26736 return std::make_pair(0U, &X86::GR16RegClass);
26737 if (VT == MVT::i32 || VT == MVT::f32 || !Subtarget->is64Bit())
26738 return std::make_pair(0U, &X86::GR32RegClass);
26739 return std::make_pair(0U, &X86::GR64RegClass);
26740 case 'R': // LEGACY_REGS
26741 if (VT == MVT::i8 || VT == MVT::i1)
26742 return std::make_pair(0U, &X86::GR8_NOREXRegClass);
26743 if (VT == MVT::i16)
26744 return std::make_pair(0U, &X86::GR16_NOREXRegClass);
26745 if (VT == MVT::i32 || !Subtarget->is64Bit())
26746 return std::make_pair(0U, &X86::GR32_NOREXRegClass);
26747 return std::make_pair(0U, &X86::GR64_NOREXRegClass);
26748 case 'f': // FP Stack registers.
26749 // If SSE is enabled for this VT, use f80 to ensure the isel moves the
26750 // value to the correct fpstack register class.
26751 if (VT == MVT::f32 && !isScalarFPTypeInSSEReg(VT))
26752 return std::make_pair(0U, &X86::RFP32RegClass);
26753 if (VT == MVT::f64 && !isScalarFPTypeInSSEReg(VT))
26754 return std::make_pair(0U, &X86::RFP64RegClass);
26755 return std::make_pair(0U, &X86::RFP80RegClass);
26756 case 'y': // MMX_REGS if MMX allowed.
26757 if (!Subtarget->hasMMX()) break;
26758 return std::make_pair(0U, &X86::VR64RegClass);
26759 case 'Y': // SSE_REGS if SSE2 allowed
26760 if (!Subtarget->hasSSE2()) break;
26762 case 'x': // SSE_REGS if SSE1 allowed or AVX_REGS if AVX allowed
26763 if (!Subtarget->hasSSE1()) break;
26765 switch (VT.SimpleTy) {
26767 // Scalar SSE types.
26770 return std::make_pair(0U, &X86::FR32RegClass);
26773 return std::make_pair(0U, &X86::FR64RegClass);
26781 return std::make_pair(0U, &X86::VR128RegClass);
26789 return std::make_pair(0U, &X86::VR256RegClass);
26794 return std::make_pair(0U, &X86::VR512RegClass);
26800 // Use the default implementation in TargetLowering to convert the register
26801 // constraint into a member of a register class.
26802 std::pair<unsigned, const TargetRegisterClass*> Res;
26803 Res = TargetLowering::getRegForInlineAsmConstraint(TRI, Constraint, VT);
26805 // Not found as a standard register?
26807 // Map st(0) -> st(7) -> ST0
26808 if (Constraint.size() == 7 && Constraint[0] == '{' &&
26809 tolower(Constraint[1]) == 's' &&
26810 tolower(Constraint[2]) == 't' &&
26811 Constraint[3] == '(' &&
26812 (Constraint[4] >= '0' && Constraint[4] <= '7') &&
26813 Constraint[5] == ')' &&
26814 Constraint[6] == '}') {
26816 Res.first = X86::FP0+Constraint[4]-'0';
26817 Res.second = &X86::RFP80RegClass;
26821 // GCC allows "st(0)" to be called just plain "st".
26822 if (StringRef("{st}").equals_lower(Constraint)) {
26823 Res.first = X86::FP0;
26824 Res.second = &X86::RFP80RegClass;
26829 if (StringRef("{flags}").equals_lower(Constraint)) {
26830 Res.first = X86::EFLAGS;
26831 Res.second = &X86::CCRRegClass;
26835 // 'A' means EAX + EDX.
26836 if (Constraint == "A") {
26837 Res.first = X86::EAX;
26838 Res.second = &X86::GR32_ADRegClass;
26844 // Otherwise, check to see if this is a register class of the wrong value
26845 // type. For example, we want to map "{ax},i32" -> {eax}, we don't want it to
26846 // turn into {ax},{dx}.
26847 // MVT::Other is used to specify clobber names.
26848 if (Res.second->hasType(VT) || VT == MVT::Other)
26849 return Res; // Correct type already, nothing to do.
26851 // Get a matching integer of the correct size. i.e. "ax" with MVT::32 should
26852 // return "eax". This should even work for things like getting 64bit integer
26853 // registers when given an f64 type.
26854 const TargetRegisterClass *Class = Res.second;
26855 if (Class == &X86::GR8RegClass || Class == &X86::GR16RegClass ||
26856 Class == &X86::GR32RegClass || Class == &X86::GR64RegClass) {
26857 unsigned Size = VT.getSizeInBits();
26858 MVT::SimpleValueType SimpleTy = Size == 1 || Size == 8 ? MVT::i8
26859 : Size == 16 ? MVT::i16
26860 : Size == 32 ? MVT::i32
26861 : Size == 64 ? MVT::i64
26863 unsigned DestReg = getX86SubSuperRegisterOrZero(Res.first, SimpleTy);
26865 Res.first = DestReg;
26866 Res.second = SimpleTy == MVT::i8 ? &X86::GR8RegClass
26867 : SimpleTy == MVT::i16 ? &X86::GR16RegClass
26868 : SimpleTy == MVT::i32 ? &X86::GR32RegClass
26869 : &X86::GR64RegClass;
26870 assert(Res.second->contains(Res.first) && "Register in register class");
26872 // No register found/type mismatch.
26874 Res.second = nullptr;
26876 } else if (Class == &X86::FR32RegClass || Class == &X86::FR64RegClass ||
26877 Class == &X86::VR128RegClass || Class == &X86::VR256RegClass ||
26878 Class == &X86::FR32XRegClass || Class == &X86::FR64XRegClass ||
26879 Class == &X86::VR128XRegClass || Class == &X86::VR256XRegClass ||
26880 Class == &X86::VR512RegClass) {
26881 // Handle references to XMM physical registers that got mapped into the
26882 // wrong class. This can happen with constraints like {xmm0} where the
26883 // target independent register mapper will just pick the first match it can
26884 // find, ignoring the required type.
26886 if (VT == MVT::f32 || VT == MVT::i32)
26887 Res.second = &X86::FR32RegClass;
26888 else if (VT == MVT::f64 || VT == MVT::i64)
26889 Res.second = &X86::FR64RegClass;
26890 else if (X86::VR128RegClass.hasType(VT))
26891 Res.second = &X86::VR128RegClass;
26892 else if (X86::VR256RegClass.hasType(VT))
26893 Res.second = &X86::VR256RegClass;
26894 else if (X86::VR512RegClass.hasType(VT))
26895 Res.second = &X86::VR512RegClass;
26897 // Type mismatch and not a clobber: Return an error;
26899 Res.second = nullptr;
26906 int X86TargetLowering::getScalingFactorCost(const DataLayout &DL,
26907 const AddrMode &AM, Type *Ty,
26908 unsigned AS) const {
26909 // Scaling factors are not free at all.
26910 // An indexed folded instruction, i.e., inst (reg1, reg2, scale),
26911 // will take 2 allocations in the out of order engine instead of 1
26912 // for plain addressing mode, i.e. inst (reg1).
26914 // vaddps (%rsi,%drx), %ymm0, %ymm1
26915 // Requires two allocations (one for the load, one for the computation)
26917 // vaddps (%rsi), %ymm0, %ymm1
26918 // Requires just 1 allocation, i.e., freeing allocations for other operations
26919 // and having less micro operations to execute.
26921 // For some X86 architectures, this is even worse because for instance for
26922 // stores, the complex addressing mode forces the instruction to use the
26923 // "load" ports instead of the dedicated "store" port.
26924 // E.g., on Haswell:
26925 // vmovaps %ymm1, (%r8, %rdi) can use port 2 or 3.
26926 // vmovaps %ymm1, (%r8) can use port 2, 3, or 7.
26927 if (isLegalAddressingMode(DL, AM, Ty, AS))
26928 // Scale represents reg2 * scale, thus account for 1
26929 // as soon as we use a second register.
26930 return AM.Scale != 0;
26934 bool X86TargetLowering::isIntDivCheap(EVT VT, AttributeSet Attr) const {
26935 // Integer division on x86 is expensive. However, when aggressively optimizing
26936 // for code size, we prefer to use a div instruction, as it is usually smaller
26937 // than the alternative sequence.
26938 // The exception to this is vector division. Since x86 doesn't have vector
26939 // integer division, leaving the division as-is is a loss even in terms of
26940 // size, because it will have to be scalarized, while the alternative code
26941 // sequence can be performed in vector form.
26942 bool OptSize = Attr.hasAttribute(AttributeSet::FunctionIndex,
26943 Attribute::MinSize);
26944 return OptSize && !VT.isVector();