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 setOperationAction(ISD::CATCHRET , MVT::Other, Custom);
421 setOperationAction(ISD::CLEANUPRET , MVT::Other, Custom);
422 // NOTE: EH_SJLJ_SETJMP/_LONGJMP supported here is NOT intended to support
423 // SjLj exception handling but a light-weight setjmp/longjmp replacement to
424 // support continuation, user-level threading, and etc.. As a result, no
425 // other SjLj exception interfaces are implemented and please don't build
426 // your own exception handling based on them.
427 // LLVM/Clang supports zero-cost DWARF exception handling.
428 setOperationAction(ISD::EH_SJLJ_SETJMP, MVT::i32, Custom);
429 setOperationAction(ISD::EH_SJLJ_LONGJMP, MVT::Other, Custom);
432 setOperationAction(ISD::ConstantPool , MVT::i32 , Custom);
433 setOperationAction(ISD::JumpTable , MVT::i32 , Custom);
434 setOperationAction(ISD::GlobalAddress , MVT::i32 , Custom);
435 setOperationAction(ISD::GlobalTLSAddress, MVT::i32 , Custom);
436 if (Subtarget->is64Bit())
437 setOperationAction(ISD::GlobalTLSAddress, MVT::i64, Custom);
438 setOperationAction(ISD::ExternalSymbol , MVT::i32 , Custom);
439 setOperationAction(ISD::BlockAddress , MVT::i32 , Custom);
440 if (Subtarget->is64Bit()) {
441 setOperationAction(ISD::ConstantPool , MVT::i64 , Custom);
442 setOperationAction(ISD::JumpTable , MVT::i64 , Custom);
443 setOperationAction(ISD::GlobalAddress , MVT::i64 , Custom);
444 setOperationAction(ISD::ExternalSymbol, MVT::i64 , Custom);
445 setOperationAction(ISD::BlockAddress , MVT::i64 , Custom);
447 // 64-bit addm sub, shl, sra, srl (iff 32-bit x86)
448 setOperationAction(ISD::SHL_PARTS , MVT::i32 , Custom);
449 setOperationAction(ISD::SRA_PARTS , MVT::i32 , Custom);
450 setOperationAction(ISD::SRL_PARTS , MVT::i32 , Custom);
451 if (Subtarget->is64Bit()) {
452 setOperationAction(ISD::SHL_PARTS , MVT::i64 , Custom);
453 setOperationAction(ISD::SRA_PARTS , MVT::i64 , Custom);
454 setOperationAction(ISD::SRL_PARTS , MVT::i64 , Custom);
457 if (Subtarget->hasSSE1())
458 setOperationAction(ISD::PREFETCH , MVT::Other, Legal);
460 setOperationAction(ISD::ATOMIC_FENCE , MVT::Other, Custom);
462 // Expand certain atomics
463 for (unsigned i = 0; i != array_lengthof(IntVTs); ++i) {
465 setOperationAction(ISD::ATOMIC_CMP_SWAP_WITH_SUCCESS, VT, Custom);
466 setOperationAction(ISD::ATOMIC_LOAD_SUB, VT, Custom);
467 setOperationAction(ISD::ATOMIC_STORE, VT, Custom);
470 if (Subtarget->hasCmpxchg16b()) {
471 setOperationAction(ISD::ATOMIC_CMP_SWAP_WITH_SUCCESS, MVT::i128, Custom);
474 // FIXME - use subtarget debug flags
475 if (!Subtarget->isTargetDarwin() && !Subtarget->isTargetELF() &&
476 !Subtarget->isTargetCygMing() && !Subtarget->isTargetWin64()) {
477 setOperationAction(ISD::EH_LABEL, MVT::Other, Expand);
480 if (Subtarget->isTarget64BitLP64()) {
481 setExceptionPointerRegister(X86::RAX);
482 setExceptionSelectorRegister(X86::RDX);
484 setExceptionPointerRegister(X86::EAX);
485 setExceptionSelectorRegister(X86::EDX);
487 setOperationAction(ISD::FRAME_TO_ARGS_OFFSET, MVT::i32, Custom);
488 setOperationAction(ISD::FRAME_TO_ARGS_OFFSET, MVT::i64, Custom);
490 setOperationAction(ISD::INIT_TRAMPOLINE, MVT::Other, Custom);
491 setOperationAction(ISD::ADJUST_TRAMPOLINE, MVT::Other, Custom);
493 setOperationAction(ISD::TRAP, MVT::Other, Legal);
494 setOperationAction(ISD::DEBUGTRAP, MVT::Other, Legal);
496 // VASTART needs to be custom lowered to use the VarArgsFrameIndex
497 setOperationAction(ISD::VASTART , MVT::Other, Custom);
498 setOperationAction(ISD::VAEND , MVT::Other, Expand);
499 if (Subtarget->is64Bit()) {
500 setOperationAction(ISD::VAARG , MVT::Other, Custom);
501 setOperationAction(ISD::VACOPY , MVT::Other, Custom);
503 // TargetInfo::CharPtrBuiltinVaList
504 setOperationAction(ISD::VAARG , MVT::Other, Expand);
505 setOperationAction(ISD::VACOPY , MVT::Other, Expand);
508 setOperationAction(ISD::STACKSAVE, MVT::Other, Expand);
509 setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand);
511 setOperationAction(ISD::DYNAMIC_STACKALLOC, PtrVT, Custom);
513 // GC_TRANSITION_START and GC_TRANSITION_END need custom lowering.
514 setOperationAction(ISD::GC_TRANSITION_START, MVT::Other, Custom);
515 setOperationAction(ISD::GC_TRANSITION_END, MVT::Other, Custom);
517 if (!Subtarget->useSoftFloat() && X86ScalarSSEf64) {
518 // f32 and f64 use SSE.
519 // Set up the FP register classes.
520 addRegisterClass(MVT::f32, &X86::FR32RegClass);
521 addRegisterClass(MVT::f64, &X86::FR64RegClass);
523 // Use ANDPD to simulate FABS.
524 setOperationAction(ISD::FABS , MVT::f64, Custom);
525 setOperationAction(ISD::FABS , MVT::f32, Custom);
527 // Use XORP to simulate FNEG.
528 setOperationAction(ISD::FNEG , MVT::f64, Custom);
529 setOperationAction(ISD::FNEG , MVT::f32, Custom);
531 // Use ANDPD and ORPD to simulate FCOPYSIGN.
532 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Custom);
533 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
535 // Lower this to FGETSIGNx86 plus an AND.
536 setOperationAction(ISD::FGETSIGN, MVT::i64, Custom);
537 setOperationAction(ISD::FGETSIGN, MVT::i32, Custom);
539 // We don't support sin/cos/fmod
540 setOperationAction(ISD::FSIN , MVT::f64, Expand);
541 setOperationAction(ISD::FCOS , MVT::f64, Expand);
542 setOperationAction(ISD::FSINCOS, MVT::f64, Expand);
543 setOperationAction(ISD::FSIN , MVT::f32, Expand);
544 setOperationAction(ISD::FCOS , MVT::f32, Expand);
545 setOperationAction(ISD::FSINCOS, MVT::f32, Expand);
547 // Expand FP immediates into loads from the stack, except for the special
549 addLegalFPImmediate(APFloat(+0.0)); // xorpd
550 addLegalFPImmediate(APFloat(+0.0f)); // xorps
551 } else if (!Subtarget->useSoftFloat() && X86ScalarSSEf32) {
552 // Use SSE for f32, x87 for f64.
553 // Set up the FP register classes.
554 addRegisterClass(MVT::f32, &X86::FR32RegClass);
555 addRegisterClass(MVT::f64, &X86::RFP64RegClass);
557 // Use ANDPS to simulate FABS.
558 setOperationAction(ISD::FABS , MVT::f32, Custom);
560 // Use XORP to simulate FNEG.
561 setOperationAction(ISD::FNEG , MVT::f32, Custom);
563 setOperationAction(ISD::UNDEF, MVT::f64, Expand);
565 // Use ANDPS and ORPS to simulate FCOPYSIGN.
566 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
567 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
569 // We don't support sin/cos/fmod
570 setOperationAction(ISD::FSIN , MVT::f32, Expand);
571 setOperationAction(ISD::FCOS , MVT::f32, Expand);
572 setOperationAction(ISD::FSINCOS, MVT::f32, Expand);
574 // Special cases we handle for FP constants.
575 addLegalFPImmediate(APFloat(+0.0f)); // xorps
576 addLegalFPImmediate(APFloat(+0.0)); // FLD0
577 addLegalFPImmediate(APFloat(+1.0)); // FLD1
578 addLegalFPImmediate(APFloat(-0.0)); // FLD0/FCHS
579 addLegalFPImmediate(APFloat(-1.0)); // FLD1/FCHS
581 if (!TM.Options.UnsafeFPMath) {
582 setOperationAction(ISD::FSIN , MVT::f64, Expand);
583 setOperationAction(ISD::FCOS , MVT::f64, Expand);
584 setOperationAction(ISD::FSINCOS, MVT::f64, Expand);
586 } else if (!Subtarget->useSoftFloat()) {
587 // f32 and f64 in x87.
588 // Set up the FP register classes.
589 addRegisterClass(MVT::f64, &X86::RFP64RegClass);
590 addRegisterClass(MVT::f32, &X86::RFP32RegClass);
592 setOperationAction(ISD::UNDEF, MVT::f64, Expand);
593 setOperationAction(ISD::UNDEF, MVT::f32, Expand);
594 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
595 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Expand);
597 if (!TM.Options.UnsafeFPMath) {
598 setOperationAction(ISD::FSIN , MVT::f64, Expand);
599 setOperationAction(ISD::FSIN , MVT::f32, Expand);
600 setOperationAction(ISD::FCOS , MVT::f64, Expand);
601 setOperationAction(ISD::FCOS , MVT::f32, Expand);
602 setOperationAction(ISD::FSINCOS, MVT::f64, Expand);
603 setOperationAction(ISD::FSINCOS, MVT::f32, Expand);
605 addLegalFPImmediate(APFloat(+0.0)); // FLD0
606 addLegalFPImmediate(APFloat(+1.0)); // FLD1
607 addLegalFPImmediate(APFloat(-0.0)); // FLD0/FCHS
608 addLegalFPImmediate(APFloat(-1.0)); // FLD1/FCHS
609 addLegalFPImmediate(APFloat(+0.0f)); // FLD0
610 addLegalFPImmediate(APFloat(+1.0f)); // FLD1
611 addLegalFPImmediate(APFloat(-0.0f)); // FLD0/FCHS
612 addLegalFPImmediate(APFloat(-1.0f)); // FLD1/FCHS
615 // We don't support FMA.
616 setOperationAction(ISD::FMA, MVT::f64, Expand);
617 setOperationAction(ISD::FMA, MVT::f32, Expand);
619 // Long double always uses X87.
620 if (!Subtarget->useSoftFloat()) {
621 addRegisterClass(MVT::f80, &X86::RFP80RegClass);
622 setOperationAction(ISD::UNDEF, MVT::f80, Expand);
623 setOperationAction(ISD::FCOPYSIGN, MVT::f80, Expand);
625 APFloat TmpFlt = APFloat::getZero(APFloat::x87DoubleExtended);
626 addLegalFPImmediate(TmpFlt); // FLD0
628 addLegalFPImmediate(TmpFlt); // FLD0/FCHS
631 APFloat TmpFlt2(+1.0);
632 TmpFlt2.convert(APFloat::x87DoubleExtended, APFloat::rmNearestTiesToEven,
634 addLegalFPImmediate(TmpFlt2); // FLD1
635 TmpFlt2.changeSign();
636 addLegalFPImmediate(TmpFlt2); // FLD1/FCHS
639 if (!TM.Options.UnsafeFPMath) {
640 setOperationAction(ISD::FSIN , MVT::f80, Expand);
641 setOperationAction(ISD::FCOS , MVT::f80, Expand);
642 setOperationAction(ISD::FSINCOS, MVT::f80, Expand);
645 setOperationAction(ISD::FFLOOR, MVT::f80, Expand);
646 setOperationAction(ISD::FCEIL, MVT::f80, Expand);
647 setOperationAction(ISD::FTRUNC, MVT::f80, Expand);
648 setOperationAction(ISD::FRINT, MVT::f80, Expand);
649 setOperationAction(ISD::FNEARBYINT, MVT::f80, Expand);
650 setOperationAction(ISD::FMA, MVT::f80, Expand);
653 // Always use a library call for pow.
654 setOperationAction(ISD::FPOW , MVT::f32 , Expand);
655 setOperationAction(ISD::FPOW , MVT::f64 , Expand);
656 setOperationAction(ISD::FPOW , MVT::f80 , Expand);
658 setOperationAction(ISD::FLOG, MVT::f80, Expand);
659 setOperationAction(ISD::FLOG2, MVT::f80, Expand);
660 setOperationAction(ISD::FLOG10, MVT::f80, Expand);
661 setOperationAction(ISD::FEXP, MVT::f80, Expand);
662 setOperationAction(ISD::FEXP2, MVT::f80, Expand);
663 setOperationAction(ISD::FMINNUM, MVT::f80, Expand);
664 setOperationAction(ISD::FMAXNUM, MVT::f80, Expand);
666 // First set operation action for all vector types to either promote
667 // (for widening) or expand (for scalarization). Then we will selectively
668 // turn on ones that can be effectively codegen'd.
669 for (MVT VT : MVT::vector_valuetypes()) {
670 setOperationAction(ISD::ADD , VT, Expand);
671 setOperationAction(ISD::SUB , VT, Expand);
672 setOperationAction(ISD::FADD, VT, Expand);
673 setOperationAction(ISD::FNEG, VT, Expand);
674 setOperationAction(ISD::FSUB, VT, Expand);
675 setOperationAction(ISD::MUL , VT, Expand);
676 setOperationAction(ISD::FMUL, VT, Expand);
677 setOperationAction(ISD::SDIV, VT, Expand);
678 setOperationAction(ISD::UDIV, VT, Expand);
679 setOperationAction(ISD::FDIV, VT, Expand);
680 setOperationAction(ISD::SREM, VT, Expand);
681 setOperationAction(ISD::UREM, VT, Expand);
682 setOperationAction(ISD::LOAD, VT, Expand);
683 setOperationAction(ISD::VECTOR_SHUFFLE, VT, Expand);
684 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT,Expand);
685 setOperationAction(ISD::INSERT_VECTOR_ELT, VT, Expand);
686 setOperationAction(ISD::EXTRACT_SUBVECTOR, VT,Expand);
687 setOperationAction(ISD::INSERT_SUBVECTOR, VT,Expand);
688 setOperationAction(ISD::FABS, VT, Expand);
689 setOperationAction(ISD::FSIN, VT, Expand);
690 setOperationAction(ISD::FSINCOS, VT, Expand);
691 setOperationAction(ISD::FCOS, VT, Expand);
692 setOperationAction(ISD::FSINCOS, VT, Expand);
693 setOperationAction(ISD::FREM, VT, Expand);
694 setOperationAction(ISD::FMA, VT, Expand);
695 setOperationAction(ISD::FPOWI, VT, Expand);
696 setOperationAction(ISD::FSQRT, VT, Expand);
697 setOperationAction(ISD::FCOPYSIGN, VT, Expand);
698 setOperationAction(ISD::FFLOOR, VT, Expand);
699 setOperationAction(ISD::FCEIL, VT, Expand);
700 setOperationAction(ISD::FTRUNC, VT, Expand);
701 setOperationAction(ISD::FRINT, VT, Expand);
702 setOperationAction(ISD::FNEARBYINT, VT, Expand);
703 setOperationAction(ISD::SMUL_LOHI, VT, Expand);
704 setOperationAction(ISD::MULHS, VT, Expand);
705 setOperationAction(ISD::UMUL_LOHI, VT, Expand);
706 setOperationAction(ISD::MULHU, VT, Expand);
707 setOperationAction(ISD::SDIVREM, VT, Expand);
708 setOperationAction(ISD::UDIVREM, VT, Expand);
709 setOperationAction(ISD::FPOW, VT, Expand);
710 setOperationAction(ISD::CTPOP, VT, Expand);
711 setOperationAction(ISD::CTTZ, VT, Expand);
712 setOperationAction(ISD::CTTZ_ZERO_UNDEF, VT, Expand);
713 setOperationAction(ISD::CTLZ, VT, Expand);
714 setOperationAction(ISD::CTLZ_ZERO_UNDEF, VT, Expand);
715 setOperationAction(ISD::SHL, VT, Expand);
716 setOperationAction(ISD::SRA, VT, Expand);
717 setOperationAction(ISD::SRL, VT, Expand);
718 setOperationAction(ISD::ROTL, VT, Expand);
719 setOperationAction(ISD::ROTR, VT, Expand);
720 setOperationAction(ISD::BSWAP, VT, Expand);
721 setOperationAction(ISD::SETCC, VT, Expand);
722 setOperationAction(ISD::FLOG, VT, Expand);
723 setOperationAction(ISD::FLOG2, VT, Expand);
724 setOperationAction(ISD::FLOG10, VT, Expand);
725 setOperationAction(ISD::FEXP, VT, Expand);
726 setOperationAction(ISD::FEXP2, VT, Expand);
727 setOperationAction(ISD::FP_TO_UINT, VT, Expand);
728 setOperationAction(ISD::FP_TO_SINT, VT, Expand);
729 setOperationAction(ISD::UINT_TO_FP, VT, Expand);
730 setOperationAction(ISD::SINT_TO_FP, VT, Expand);
731 setOperationAction(ISD::SIGN_EXTEND_INREG, VT,Expand);
732 setOperationAction(ISD::TRUNCATE, VT, Expand);
733 setOperationAction(ISD::SIGN_EXTEND, VT, Expand);
734 setOperationAction(ISD::ZERO_EXTEND, VT, Expand);
735 setOperationAction(ISD::ANY_EXTEND, VT, Expand);
736 setOperationAction(ISD::VSELECT, VT, Expand);
737 setOperationAction(ISD::SELECT_CC, VT, Expand);
738 for (MVT InnerVT : MVT::vector_valuetypes()) {
739 setTruncStoreAction(InnerVT, VT, Expand);
741 setLoadExtAction(ISD::SEXTLOAD, InnerVT, VT, Expand);
742 setLoadExtAction(ISD::ZEXTLOAD, InnerVT, VT, Expand);
744 // N.b. ISD::EXTLOAD legality is basically ignored except for i1-like
745 // types, we have to deal with them whether we ask for Expansion or not.
746 // Setting Expand causes its own optimisation problems though, so leave
748 if (VT.getVectorElementType() == MVT::i1)
749 setLoadExtAction(ISD::EXTLOAD, InnerVT, VT, Expand);
751 // EXTLOAD for MVT::f16 vectors is not legal because f16 vectors are
752 // split/scalarized right now.
753 if (VT.getVectorElementType() == MVT::f16)
754 setLoadExtAction(ISD::EXTLOAD, InnerVT, VT, Expand);
758 // FIXME: In order to prevent SSE instructions being expanded to MMX ones
759 // with -msoft-float, disable use of MMX as well.
760 if (!Subtarget->useSoftFloat() && Subtarget->hasMMX()) {
761 addRegisterClass(MVT::x86mmx, &X86::VR64RegClass);
762 // No operations on x86mmx supported, everything uses intrinsics.
765 // MMX-sized vectors (other than x86mmx) are expected to be expanded
766 // into smaller operations.
767 for (MVT MMXTy : {MVT::v8i8, MVT::v4i16, MVT::v2i32, MVT::v1i64}) {
768 setOperationAction(ISD::MULHS, MMXTy, Expand);
769 setOperationAction(ISD::AND, MMXTy, Expand);
770 setOperationAction(ISD::OR, MMXTy, Expand);
771 setOperationAction(ISD::XOR, MMXTy, Expand);
772 setOperationAction(ISD::SCALAR_TO_VECTOR, MMXTy, Expand);
773 setOperationAction(ISD::SELECT, MMXTy, Expand);
774 setOperationAction(ISD::BITCAST, MMXTy, Expand);
776 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v1i64, Expand);
778 if (!Subtarget->useSoftFloat() && Subtarget->hasSSE1()) {
779 addRegisterClass(MVT::v4f32, &X86::VR128RegClass);
781 setOperationAction(ISD::FADD, MVT::v4f32, Legal);
782 setOperationAction(ISD::FSUB, MVT::v4f32, Legal);
783 setOperationAction(ISD::FMUL, MVT::v4f32, Legal);
784 setOperationAction(ISD::FDIV, MVT::v4f32, Legal);
785 setOperationAction(ISD::FSQRT, MVT::v4f32, Legal);
786 setOperationAction(ISD::FNEG, MVT::v4f32, Custom);
787 setOperationAction(ISD::FABS, MVT::v4f32, Custom);
788 setOperationAction(ISD::LOAD, MVT::v4f32, Legal);
789 setOperationAction(ISD::BUILD_VECTOR, MVT::v4f32, Custom);
790 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4f32, Custom);
791 setOperationAction(ISD::VSELECT, MVT::v4f32, Custom);
792 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f32, Custom);
793 setOperationAction(ISD::SELECT, MVT::v4f32, Custom);
794 setOperationAction(ISD::UINT_TO_FP, MVT::v4i32, Custom);
797 if (!Subtarget->useSoftFloat() && Subtarget->hasSSE2()) {
798 addRegisterClass(MVT::v2f64, &X86::VR128RegClass);
800 // FIXME: Unfortunately, -soft-float and -no-implicit-float mean XMM
801 // registers cannot be used even for integer operations.
802 addRegisterClass(MVT::v16i8, &X86::VR128RegClass);
803 addRegisterClass(MVT::v8i16, &X86::VR128RegClass);
804 addRegisterClass(MVT::v4i32, &X86::VR128RegClass);
805 addRegisterClass(MVT::v2i64, &X86::VR128RegClass);
807 setOperationAction(ISD::ADD, MVT::v16i8, Legal);
808 setOperationAction(ISD::ADD, MVT::v8i16, Legal);
809 setOperationAction(ISD::ADD, MVT::v4i32, Legal);
810 setOperationAction(ISD::ADD, MVT::v2i64, Legal);
811 setOperationAction(ISD::MUL, MVT::v16i8, Custom);
812 setOperationAction(ISD::MUL, MVT::v4i32, Custom);
813 setOperationAction(ISD::MUL, MVT::v2i64, Custom);
814 setOperationAction(ISD::UMUL_LOHI, MVT::v4i32, Custom);
815 setOperationAction(ISD::SMUL_LOHI, MVT::v4i32, Custom);
816 setOperationAction(ISD::MULHU, MVT::v8i16, Legal);
817 setOperationAction(ISD::MULHS, MVT::v8i16, Legal);
818 setOperationAction(ISD::SUB, MVT::v16i8, Legal);
819 setOperationAction(ISD::SUB, MVT::v8i16, Legal);
820 setOperationAction(ISD::SUB, MVT::v4i32, Legal);
821 setOperationAction(ISD::SUB, MVT::v2i64, Legal);
822 setOperationAction(ISD::MUL, MVT::v8i16, Legal);
823 setOperationAction(ISD::FADD, MVT::v2f64, Legal);
824 setOperationAction(ISD::FSUB, MVT::v2f64, Legal);
825 setOperationAction(ISD::FMUL, MVT::v2f64, Legal);
826 setOperationAction(ISD::FDIV, MVT::v2f64, Legal);
827 setOperationAction(ISD::FSQRT, MVT::v2f64, Legal);
828 setOperationAction(ISD::FNEG, MVT::v2f64, Custom);
829 setOperationAction(ISD::FABS, MVT::v2f64, Custom);
831 setOperationAction(ISD::SMAX, MVT::v8i16, Legal);
832 setOperationAction(ISD::UMAX, MVT::v16i8, Legal);
833 setOperationAction(ISD::SMIN, MVT::v8i16, Legal);
834 setOperationAction(ISD::UMIN, MVT::v16i8, Legal);
836 setOperationAction(ISD::SETCC, MVT::v2i64, Custom);
837 setOperationAction(ISD::SETCC, MVT::v16i8, Custom);
838 setOperationAction(ISD::SETCC, MVT::v8i16, Custom);
839 setOperationAction(ISD::SETCC, MVT::v4i32, Custom);
841 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v16i8, Custom);
842 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v8i16, Custom);
843 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i16, Custom);
844 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i32, Custom);
845 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f32, Custom);
847 setOperationAction(ISD::CTPOP, MVT::v16i8, Custom);
848 setOperationAction(ISD::CTPOP, MVT::v8i16, Custom);
849 setOperationAction(ISD::CTPOP, MVT::v4i32, Custom);
850 setOperationAction(ISD::CTPOP, MVT::v2i64, Custom);
852 // Custom lower build_vector, vector_shuffle, and extract_vector_elt.
853 for (int i = MVT::v16i8; i != MVT::v2i64; ++i) {
854 MVT VT = (MVT::SimpleValueType)i;
855 // Do not attempt to custom lower non-power-of-2 vectors
856 if (!isPowerOf2_32(VT.getVectorNumElements()))
858 // Do not attempt to custom lower non-128-bit vectors
859 if (!VT.is128BitVector())
861 setOperationAction(ISD::BUILD_VECTOR, VT, Custom);
862 setOperationAction(ISD::VECTOR_SHUFFLE, VT, Custom);
863 setOperationAction(ISD::VSELECT, VT, Custom);
864 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Custom);
867 // We support custom legalizing of sext and anyext loads for specific
868 // memory vector types which we can load as a scalar (or sequence of
869 // scalars) and extend in-register to a legal 128-bit vector type. For sext
870 // loads these must work with a single scalar load.
871 for (MVT VT : MVT::integer_vector_valuetypes()) {
872 setLoadExtAction(ISD::SEXTLOAD, VT, MVT::v4i8, Custom);
873 setLoadExtAction(ISD::SEXTLOAD, VT, MVT::v4i16, Custom);
874 setLoadExtAction(ISD::SEXTLOAD, VT, MVT::v8i8, Custom);
875 setLoadExtAction(ISD::EXTLOAD, VT, MVT::v2i8, Custom);
876 setLoadExtAction(ISD::EXTLOAD, VT, MVT::v2i16, Custom);
877 setLoadExtAction(ISD::EXTLOAD, VT, MVT::v2i32, Custom);
878 setLoadExtAction(ISD::EXTLOAD, VT, MVT::v4i8, Custom);
879 setLoadExtAction(ISD::EXTLOAD, VT, MVT::v4i16, Custom);
880 setLoadExtAction(ISD::EXTLOAD, VT, MVT::v8i8, Custom);
883 setOperationAction(ISD::BUILD_VECTOR, MVT::v2f64, Custom);
884 setOperationAction(ISD::BUILD_VECTOR, MVT::v2i64, Custom);
885 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2f64, Custom);
886 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2i64, Custom);
887 setOperationAction(ISD::VSELECT, MVT::v2f64, Custom);
888 setOperationAction(ISD::VSELECT, MVT::v2i64, Custom);
889 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2f64, Custom);
890 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2f64, Custom);
892 if (Subtarget->is64Bit()) {
893 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2i64, Custom);
894 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2i64, Custom);
897 // Promote v16i8, v8i16, v4i32 load, select, and, or, xor to v2i64.
898 for (int i = MVT::v16i8; i != MVT::v2i64; ++i) {
899 MVT VT = (MVT::SimpleValueType)i;
901 // Do not attempt to promote non-128-bit vectors
902 if (!VT.is128BitVector())
905 setOperationAction(ISD::AND, VT, Promote);
906 AddPromotedToType (ISD::AND, VT, MVT::v2i64);
907 setOperationAction(ISD::OR, VT, Promote);
908 AddPromotedToType (ISD::OR, VT, MVT::v2i64);
909 setOperationAction(ISD::XOR, VT, Promote);
910 AddPromotedToType (ISD::XOR, VT, MVT::v2i64);
911 setOperationAction(ISD::LOAD, VT, Promote);
912 AddPromotedToType (ISD::LOAD, VT, MVT::v2i64);
913 setOperationAction(ISD::SELECT, VT, Promote);
914 AddPromotedToType (ISD::SELECT, VT, MVT::v2i64);
917 // Custom lower v2i64 and v2f64 selects.
918 setOperationAction(ISD::LOAD, MVT::v2f64, Legal);
919 setOperationAction(ISD::LOAD, MVT::v2i64, Legal);
920 setOperationAction(ISD::SELECT, MVT::v2f64, Custom);
921 setOperationAction(ISD::SELECT, MVT::v2i64, Custom);
923 setOperationAction(ISD::FP_TO_SINT, MVT::v4i32, Legal);
924 setOperationAction(ISD::SINT_TO_FP, MVT::v4i32, Legal);
926 setOperationAction(ISD::SINT_TO_FP, MVT::v2i32, Custom);
928 setOperationAction(ISD::UINT_TO_FP, MVT::v4i8, Custom);
929 setOperationAction(ISD::UINT_TO_FP, MVT::v4i16, Custom);
930 // As there is no 64-bit GPR available, we need build a special custom
931 // sequence to convert from v2i32 to v2f32.
932 if (!Subtarget->is64Bit())
933 setOperationAction(ISD::UINT_TO_FP, MVT::v2f32, Custom);
935 setOperationAction(ISD::FP_EXTEND, MVT::v2f32, Custom);
936 setOperationAction(ISD::FP_ROUND, MVT::v2f32, Custom);
938 for (MVT VT : MVT::fp_vector_valuetypes())
939 setLoadExtAction(ISD::EXTLOAD, VT, MVT::v2f32, Legal);
941 setOperationAction(ISD::BITCAST, MVT::v2i32, Custom);
942 setOperationAction(ISD::BITCAST, MVT::v4i16, Custom);
943 setOperationAction(ISD::BITCAST, MVT::v8i8, Custom);
946 if (!Subtarget->useSoftFloat() && Subtarget->hasSSE41()) {
947 for (MVT RoundedTy : {MVT::f32, MVT::f64, MVT::v4f32, MVT::v2f64}) {
948 setOperationAction(ISD::FFLOOR, RoundedTy, Legal);
949 setOperationAction(ISD::FCEIL, RoundedTy, Legal);
950 setOperationAction(ISD::FTRUNC, RoundedTy, Legal);
951 setOperationAction(ISD::FRINT, RoundedTy, Legal);
952 setOperationAction(ISD::FNEARBYINT, RoundedTy, Legal);
955 setOperationAction(ISD::SMAX, MVT::v16i8, Legal);
956 setOperationAction(ISD::SMAX, MVT::v4i32, Legal);
957 setOperationAction(ISD::UMAX, MVT::v8i16, Legal);
958 setOperationAction(ISD::UMAX, MVT::v4i32, Legal);
959 setOperationAction(ISD::SMIN, MVT::v16i8, Legal);
960 setOperationAction(ISD::SMIN, MVT::v4i32, Legal);
961 setOperationAction(ISD::UMIN, MVT::v8i16, Legal);
962 setOperationAction(ISD::UMIN, MVT::v4i32, Legal);
964 // FIXME: Do we need to handle scalar-to-vector here?
965 setOperationAction(ISD::MUL, MVT::v4i32, Legal);
967 // We directly match byte blends in the backend as they match the VSELECT
969 setOperationAction(ISD::VSELECT, MVT::v16i8, Legal);
971 // SSE41 brings specific instructions for doing vector sign extend even in
972 // cases where we don't have SRA.
973 for (MVT VT : MVT::integer_vector_valuetypes()) {
974 setLoadExtAction(ISD::SEXTLOAD, VT, MVT::v2i8, Custom);
975 setLoadExtAction(ISD::SEXTLOAD, VT, MVT::v2i16, Custom);
976 setLoadExtAction(ISD::SEXTLOAD, VT, MVT::v2i32, Custom);
979 // SSE41 also has vector sign/zero extending loads, PMOV[SZ]X
980 setLoadExtAction(ISD::SEXTLOAD, MVT::v8i16, MVT::v8i8, Legal);
981 setLoadExtAction(ISD::SEXTLOAD, MVT::v4i32, MVT::v4i8, Legal);
982 setLoadExtAction(ISD::SEXTLOAD, MVT::v2i64, MVT::v2i8, Legal);
983 setLoadExtAction(ISD::SEXTLOAD, MVT::v4i32, MVT::v4i16, Legal);
984 setLoadExtAction(ISD::SEXTLOAD, MVT::v2i64, MVT::v2i16, Legal);
985 setLoadExtAction(ISD::SEXTLOAD, MVT::v2i64, MVT::v2i32, Legal);
987 setLoadExtAction(ISD::ZEXTLOAD, MVT::v8i16, MVT::v8i8, Legal);
988 setLoadExtAction(ISD::ZEXTLOAD, MVT::v4i32, MVT::v4i8, Legal);
989 setLoadExtAction(ISD::ZEXTLOAD, MVT::v2i64, MVT::v2i8, Legal);
990 setLoadExtAction(ISD::ZEXTLOAD, MVT::v4i32, MVT::v4i16, Legal);
991 setLoadExtAction(ISD::ZEXTLOAD, MVT::v2i64, MVT::v2i16, Legal);
992 setLoadExtAction(ISD::ZEXTLOAD, MVT::v2i64, MVT::v2i32, Legal);
994 // i8 and i16 vectors are custom because the source register and source
995 // source memory operand types are not the same width. f32 vectors are
996 // custom since the immediate controlling the insert encodes additional
998 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v16i8, Custom);
999 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i16, Custom);
1000 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i32, Custom);
1001 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f32, Custom);
1003 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v16i8, Custom);
1004 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v8i16, Custom);
1005 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4i32, Custom);
1006 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f32, Custom);
1008 // FIXME: these should be Legal, but that's only for the case where
1009 // the index is constant. For now custom expand to deal with that.
1010 if (Subtarget->is64Bit()) {
1011 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2i64, Custom);
1012 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2i64, Custom);
1016 if (Subtarget->hasSSE2()) {
1017 setOperationAction(ISD::SIGN_EXTEND_VECTOR_INREG, MVT::v2i64, Custom);
1018 setOperationAction(ISD::SIGN_EXTEND_VECTOR_INREG, MVT::v4i32, Custom);
1019 setOperationAction(ISD::SIGN_EXTEND_VECTOR_INREG, MVT::v8i16, Custom);
1021 setOperationAction(ISD::SRL, MVT::v8i16, Custom);
1022 setOperationAction(ISD::SRL, MVT::v16i8, Custom);
1024 setOperationAction(ISD::SHL, MVT::v8i16, Custom);
1025 setOperationAction(ISD::SHL, MVT::v16i8, Custom);
1027 setOperationAction(ISD::SRA, MVT::v8i16, Custom);
1028 setOperationAction(ISD::SRA, MVT::v16i8, Custom);
1030 // In the customized shift lowering, the legal cases in AVX2 will be
1032 setOperationAction(ISD::SRL, MVT::v2i64, Custom);
1033 setOperationAction(ISD::SRL, MVT::v4i32, Custom);
1035 setOperationAction(ISD::SHL, MVT::v2i64, Custom);
1036 setOperationAction(ISD::SHL, MVT::v4i32, Custom);
1038 setOperationAction(ISD::SRA, MVT::v2i64, Custom);
1039 setOperationAction(ISD::SRA, MVT::v4i32, Custom);
1042 if (!Subtarget->useSoftFloat() && Subtarget->hasFp256()) {
1043 addRegisterClass(MVT::v32i8, &X86::VR256RegClass);
1044 addRegisterClass(MVT::v16i16, &X86::VR256RegClass);
1045 addRegisterClass(MVT::v8i32, &X86::VR256RegClass);
1046 addRegisterClass(MVT::v8f32, &X86::VR256RegClass);
1047 addRegisterClass(MVT::v4i64, &X86::VR256RegClass);
1048 addRegisterClass(MVT::v4f64, &X86::VR256RegClass);
1050 setOperationAction(ISD::LOAD, MVT::v8f32, Legal);
1051 setOperationAction(ISD::LOAD, MVT::v4f64, Legal);
1052 setOperationAction(ISD::LOAD, MVT::v4i64, Legal);
1054 setOperationAction(ISD::FADD, MVT::v8f32, Legal);
1055 setOperationAction(ISD::FSUB, MVT::v8f32, Legal);
1056 setOperationAction(ISD::FMUL, MVT::v8f32, Legal);
1057 setOperationAction(ISD::FDIV, MVT::v8f32, Legal);
1058 setOperationAction(ISD::FSQRT, MVT::v8f32, Legal);
1059 setOperationAction(ISD::FFLOOR, MVT::v8f32, Legal);
1060 setOperationAction(ISD::FCEIL, MVT::v8f32, Legal);
1061 setOperationAction(ISD::FTRUNC, MVT::v8f32, Legal);
1062 setOperationAction(ISD::FRINT, MVT::v8f32, Legal);
1063 setOperationAction(ISD::FNEARBYINT, MVT::v8f32, Legal);
1064 setOperationAction(ISD::FNEG, MVT::v8f32, Custom);
1065 setOperationAction(ISD::FABS, MVT::v8f32, Custom);
1067 setOperationAction(ISD::FADD, MVT::v4f64, Legal);
1068 setOperationAction(ISD::FSUB, MVT::v4f64, Legal);
1069 setOperationAction(ISD::FMUL, MVT::v4f64, Legal);
1070 setOperationAction(ISD::FDIV, MVT::v4f64, Legal);
1071 setOperationAction(ISD::FSQRT, MVT::v4f64, Legal);
1072 setOperationAction(ISD::FFLOOR, MVT::v4f64, Legal);
1073 setOperationAction(ISD::FCEIL, MVT::v4f64, Legal);
1074 setOperationAction(ISD::FTRUNC, MVT::v4f64, Legal);
1075 setOperationAction(ISD::FRINT, MVT::v4f64, Legal);
1076 setOperationAction(ISD::FNEARBYINT, MVT::v4f64, Legal);
1077 setOperationAction(ISD::FNEG, MVT::v4f64, Custom);
1078 setOperationAction(ISD::FABS, MVT::v4f64, Custom);
1080 // (fp_to_int:v8i16 (v8f32 ..)) requires the result type to be promoted
1081 // even though v8i16 is a legal type.
1082 setOperationAction(ISD::FP_TO_SINT, MVT::v8i16, Promote);
1083 setOperationAction(ISD::FP_TO_UINT, MVT::v8i16, Promote);
1084 setOperationAction(ISD::FP_TO_SINT, MVT::v8i32, Legal);
1086 setOperationAction(ISD::SINT_TO_FP, MVT::v8i16, Promote);
1087 setOperationAction(ISD::SINT_TO_FP, MVT::v8i32, Legal);
1088 setOperationAction(ISD::FP_ROUND, MVT::v4f32, Legal);
1090 setOperationAction(ISD::UINT_TO_FP, MVT::v8i8, Custom);
1091 setOperationAction(ISD::UINT_TO_FP, MVT::v8i16, Custom);
1093 for (MVT VT : MVT::fp_vector_valuetypes())
1094 setLoadExtAction(ISD::EXTLOAD, VT, MVT::v4f32, Legal);
1096 setOperationAction(ISD::SRL, MVT::v16i16, Custom);
1097 setOperationAction(ISD::SRL, MVT::v32i8, Custom);
1099 setOperationAction(ISD::SHL, MVT::v16i16, Custom);
1100 setOperationAction(ISD::SHL, MVT::v32i8, Custom);
1102 setOperationAction(ISD::SRA, MVT::v16i16, Custom);
1103 setOperationAction(ISD::SRA, MVT::v32i8, Custom);
1105 setOperationAction(ISD::SETCC, MVT::v32i8, Custom);
1106 setOperationAction(ISD::SETCC, MVT::v16i16, Custom);
1107 setOperationAction(ISD::SETCC, MVT::v8i32, Custom);
1108 setOperationAction(ISD::SETCC, MVT::v4i64, Custom);
1110 setOperationAction(ISD::SELECT, MVT::v4f64, Custom);
1111 setOperationAction(ISD::SELECT, MVT::v4i64, Custom);
1112 setOperationAction(ISD::SELECT, MVT::v8f32, Custom);
1114 setOperationAction(ISD::SIGN_EXTEND, MVT::v4i64, Custom);
1115 setOperationAction(ISD::SIGN_EXTEND, MVT::v8i32, Custom);
1116 setOperationAction(ISD::SIGN_EXTEND, MVT::v16i16, Custom);
1117 setOperationAction(ISD::ZERO_EXTEND, MVT::v4i64, Custom);
1118 setOperationAction(ISD::ZERO_EXTEND, MVT::v8i32, Custom);
1119 setOperationAction(ISD::ZERO_EXTEND, MVT::v16i16, Custom);
1120 setOperationAction(ISD::ANY_EXTEND, MVT::v4i64, Custom);
1121 setOperationAction(ISD::ANY_EXTEND, MVT::v8i32, Custom);
1122 setOperationAction(ISD::ANY_EXTEND, MVT::v16i16, Custom);
1123 setOperationAction(ISD::TRUNCATE, MVT::v16i8, Custom);
1124 setOperationAction(ISD::TRUNCATE, MVT::v8i16, Custom);
1125 setOperationAction(ISD::TRUNCATE, MVT::v4i32, Custom);
1127 setOperationAction(ISD::CTPOP, MVT::v32i8, Custom);
1128 setOperationAction(ISD::CTPOP, MVT::v16i16, Custom);
1129 setOperationAction(ISD::CTPOP, MVT::v8i32, Custom);
1130 setOperationAction(ISD::CTPOP, MVT::v4i64, Custom);
1132 if (Subtarget->hasFMA() || Subtarget->hasFMA4() || Subtarget->hasAVX512()) {
1133 setOperationAction(ISD::FMA, MVT::v8f32, Legal);
1134 setOperationAction(ISD::FMA, MVT::v4f64, Legal);
1135 setOperationAction(ISD::FMA, MVT::v4f32, Legal);
1136 setOperationAction(ISD::FMA, MVT::v2f64, Legal);
1137 setOperationAction(ISD::FMA, MVT::f32, Legal);
1138 setOperationAction(ISD::FMA, MVT::f64, Legal);
1141 if (Subtarget->hasInt256()) {
1142 setOperationAction(ISD::ADD, MVT::v4i64, Legal);
1143 setOperationAction(ISD::ADD, MVT::v8i32, Legal);
1144 setOperationAction(ISD::ADD, MVT::v16i16, Legal);
1145 setOperationAction(ISD::ADD, MVT::v32i8, Legal);
1147 setOperationAction(ISD::SUB, MVT::v4i64, Legal);
1148 setOperationAction(ISD::SUB, MVT::v8i32, Legal);
1149 setOperationAction(ISD::SUB, MVT::v16i16, Legal);
1150 setOperationAction(ISD::SUB, MVT::v32i8, Legal);
1152 setOperationAction(ISD::MUL, MVT::v4i64, Custom);
1153 setOperationAction(ISD::MUL, MVT::v8i32, Legal);
1154 setOperationAction(ISD::MUL, MVT::v16i16, Legal);
1155 setOperationAction(ISD::MUL, MVT::v32i8, Custom);
1157 setOperationAction(ISD::UMUL_LOHI, MVT::v8i32, Custom);
1158 setOperationAction(ISD::SMUL_LOHI, MVT::v8i32, Custom);
1159 setOperationAction(ISD::MULHU, MVT::v16i16, Legal);
1160 setOperationAction(ISD::MULHS, MVT::v16i16, Legal);
1162 setOperationAction(ISD::SMAX, MVT::v32i8, Legal);
1163 setOperationAction(ISD::SMAX, MVT::v16i16, Legal);
1164 setOperationAction(ISD::SMAX, MVT::v8i32, Legal);
1165 setOperationAction(ISD::UMAX, MVT::v32i8, Legal);
1166 setOperationAction(ISD::UMAX, MVT::v16i16, Legal);
1167 setOperationAction(ISD::UMAX, MVT::v8i32, Legal);
1168 setOperationAction(ISD::SMIN, MVT::v32i8, Legal);
1169 setOperationAction(ISD::SMIN, MVT::v16i16, Legal);
1170 setOperationAction(ISD::SMIN, MVT::v8i32, Legal);
1171 setOperationAction(ISD::UMIN, MVT::v32i8, Legal);
1172 setOperationAction(ISD::UMIN, MVT::v16i16, Legal);
1173 setOperationAction(ISD::UMIN, MVT::v8i32, Legal);
1175 // The custom lowering for UINT_TO_FP for v8i32 becomes interesting
1176 // when we have a 256bit-wide blend with immediate.
1177 setOperationAction(ISD::UINT_TO_FP, MVT::v8i32, Custom);
1179 // AVX2 also has wider vector sign/zero extending loads, VPMOV[SZ]X
1180 setLoadExtAction(ISD::SEXTLOAD, MVT::v16i16, MVT::v16i8, Legal);
1181 setLoadExtAction(ISD::SEXTLOAD, MVT::v8i32, MVT::v8i8, Legal);
1182 setLoadExtAction(ISD::SEXTLOAD, MVT::v4i64, MVT::v4i8, Legal);
1183 setLoadExtAction(ISD::SEXTLOAD, MVT::v8i32, MVT::v8i16, Legal);
1184 setLoadExtAction(ISD::SEXTLOAD, MVT::v4i64, MVT::v4i16, Legal);
1185 setLoadExtAction(ISD::SEXTLOAD, MVT::v4i64, MVT::v4i32, Legal);
1187 setLoadExtAction(ISD::ZEXTLOAD, MVT::v16i16, MVT::v16i8, Legal);
1188 setLoadExtAction(ISD::ZEXTLOAD, MVT::v8i32, MVT::v8i8, Legal);
1189 setLoadExtAction(ISD::ZEXTLOAD, MVT::v4i64, MVT::v4i8, Legal);
1190 setLoadExtAction(ISD::ZEXTLOAD, MVT::v8i32, MVT::v8i16, Legal);
1191 setLoadExtAction(ISD::ZEXTLOAD, MVT::v4i64, MVT::v4i16, Legal);
1192 setLoadExtAction(ISD::ZEXTLOAD, MVT::v4i64, MVT::v4i32, Legal);
1194 setOperationAction(ISD::ADD, MVT::v4i64, Custom);
1195 setOperationAction(ISD::ADD, MVT::v8i32, Custom);
1196 setOperationAction(ISD::ADD, MVT::v16i16, Custom);
1197 setOperationAction(ISD::ADD, MVT::v32i8, Custom);
1199 setOperationAction(ISD::SUB, MVT::v4i64, Custom);
1200 setOperationAction(ISD::SUB, MVT::v8i32, Custom);
1201 setOperationAction(ISD::SUB, MVT::v16i16, Custom);
1202 setOperationAction(ISD::SUB, MVT::v32i8, Custom);
1204 setOperationAction(ISD::MUL, MVT::v4i64, Custom);
1205 setOperationAction(ISD::MUL, MVT::v8i32, Custom);
1206 setOperationAction(ISD::MUL, MVT::v16i16, Custom);
1207 setOperationAction(ISD::MUL, MVT::v32i8, Custom);
1209 setOperationAction(ISD::SMAX, MVT::v32i8, Custom);
1210 setOperationAction(ISD::SMAX, MVT::v16i16, Custom);
1211 setOperationAction(ISD::SMAX, MVT::v8i32, Custom);
1212 setOperationAction(ISD::UMAX, MVT::v32i8, Custom);
1213 setOperationAction(ISD::UMAX, MVT::v16i16, Custom);
1214 setOperationAction(ISD::UMAX, MVT::v8i32, Custom);
1215 setOperationAction(ISD::SMIN, MVT::v32i8, Custom);
1216 setOperationAction(ISD::SMIN, MVT::v16i16, Custom);
1217 setOperationAction(ISD::SMIN, MVT::v8i32, Custom);
1218 setOperationAction(ISD::UMIN, MVT::v32i8, Custom);
1219 setOperationAction(ISD::UMIN, MVT::v16i16, Custom);
1220 setOperationAction(ISD::UMIN, MVT::v8i32, Custom);
1223 // In the customized shift lowering, the legal cases in AVX2 will be
1225 setOperationAction(ISD::SRL, MVT::v4i64, Custom);
1226 setOperationAction(ISD::SRL, MVT::v8i32, Custom);
1228 setOperationAction(ISD::SHL, MVT::v4i64, Custom);
1229 setOperationAction(ISD::SHL, MVT::v8i32, Custom);
1231 setOperationAction(ISD::SRA, MVT::v4i64, Custom);
1232 setOperationAction(ISD::SRA, MVT::v8i32, Custom);
1234 // Custom lower several nodes for 256-bit types.
1235 for (MVT VT : MVT::vector_valuetypes()) {
1236 if (VT.getScalarSizeInBits() >= 32) {
1237 setOperationAction(ISD::MLOAD, VT, Legal);
1238 setOperationAction(ISD::MSTORE, VT, Legal);
1240 // Extract subvector is special because the value type
1241 // (result) is 128-bit but the source is 256-bit wide.
1242 if (VT.is128BitVector()) {
1243 setOperationAction(ISD::EXTRACT_SUBVECTOR, VT, Custom);
1245 // Do not attempt to custom lower other non-256-bit vectors
1246 if (!VT.is256BitVector())
1249 setOperationAction(ISD::BUILD_VECTOR, VT, Custom);
1250 setOperationAction(ISD::VECTOR_SHUFFLE, VT, Custom);
1251 setOperationAction(ISD::VSELECT, VT, Custom);
1252 setOperationAction(ISD::INSERT_VECTOR_ELT, VT, Custom);
1253 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Custom);
1254 setOperationAction(ISD::SCALAR_TO_VECTOR, VT, Custom);
1255 setOperationAction(ISD::INSERT_SUBVECTOR, VT, Custom);
1256 setOperationAction(ISD::CONCAT_VECTORS, VT, Custom);
1259 if (Subtarget->hasInt256())
1260 setOperationAction(ISD::VSELECT, MVT::v32i8, Legal);
1263 // Promote v32i8, v16i16, v8i32 select, and, or, xor to v4i64.
1264 for (int i = MVT::v32i8; i != MVT::v4i64; ++i) {
1265 MVT VT = (MVT::SimpleValueType)i;
1267 // Do not attempt to promote non-256-bit vectors
1268 if (!VT.is256BitVector())
1271 setOperationAction(ISD::AND, VT, Promote);
1272 AddPromotedToType (ISD::AND, VT, MVT::v4i64);
1273 setOperationAction(ISD::OR, VT, Promote);
1274 AddPromotedToType (ISD::OR, VT, MVT::v4i64);
1275 setOperationAction(ISD::XOR, VT, Promote);
1276 AddPromotedToType (ISD::XOR, VT, MVT::v4i64);
1277 setOperationAction(ISD::LOAD, VT, Promote);
1278 AddPromotedToType (ISD::LOAD, VT, MVT::v4i64);
1279 setOperationAction(ISD::SELECT, VT, Promote);
1280 AddPromotedToType (ISD::SELECT, VT, MVT::v4i64);
1284 if (!Subtarget->useSoftFloat() && Subtarget->hasAVX512()) {
1285 addRegisterClass(MVT::v16i32, &X86::VR512RegClass);
1286 addRegisterClass(MVT::v16f32, &X86::VR512RegClass);
1287 addRegisterClass(MVT::v8i64, &X86::VR512RegClass);
1288 addRegisterClass(MVT::v8f64, &X86::VR512RegClass);
1290 addRegisterClass(MVT::i1, &X86::VK1RegClass);
1291 addRegisterClass(MVT::v8i1, &X86::VK8RegClass);
1292 addRegisterClass(MVT::v16i1, &X86::VK16RegClass);
1294 for (MVT VT : MVT::fp_vector_valuetypes())
1295 setLoadExtAction(ISD::EXTLOAD, VT, MVT::v8f32, Legal);
1297 setLoadExtAction(ISD::ZEXTLOAD, MVT::v16i32, MVT::v16i8, Legal);
1298 setLoadExtAction(ISD::SEXTLOAD, MVT::v16i32, MVT::v16i8, Legal);
1299 setLoadExtAction(ISD::ZEXTLOAD, MVT::v16i32, MVT::v16i16, Legal);
1300 setLoadExtAction(ISD::SEXTLOAD, MVT::v16i32, MVT::v16i16, Legal);
1301 setLoadExtAction(ISD::ZEXTLOAD, MVT::v32i16, MVT::v32i8, Legal);
1302 setLoadExtAction(ISD::SEXTLOAD, MVT::v32i16, MVT::v32i8, Legal);
1303 setLoadExtAction(ISD::ZEXTLOAD, MVT::v8i64, MVT::v8i8, Legal);
1304 setLoadExtAction(ISD::SEXTLOAD, MVT::v8i64, MVT::v8i8, Legal);
1305 setLoadExtAction(ISD::ZEXTLOAD, MVT::v8i64, MVT::v8i16, Legal);
1306 setLoadExtAction(ISD::SEXTLOAD, MVT::v8i64, MVT::v8i16, Legal);
1307 setLoadExtAction(ISD::ZEXTLOAD, MVT::v8i64, MVT::v8i32, Legal);
1308 setLoadExtAction(ISD::SEXTLOAD, MVT::v8i64, MVT::v8i32, Legal);
1310 setOperationAction(ISD::BR_CC, MVT::i1, Expand);
1311 setOperationAction(ISD::SETCC, MVT::i1, Custom);
1312 setOperationAction(ISD::XOR, MVT::i1, Legal);
1313 setOperationAction(ISD::OR, MVT::i1, Legal);
1314 setOperationAction(ISD::AND, MVT::i1, Legal);
1315 setOperationAction(ISD::SUB, MVT::i1, Custom);
1316 setOperationAction(ISD::ADD, MVT::i1, Custom);
1317 setOperationAction(ISD::MUL, MVT::i1, Custom);
1318 setOperationAction(ISD::LOAD, MVT::v16f32, Legal);
1319 setOperationAction(ISD::LOAD, MVT::v8f64, Legal);
1320 setOperationAction(ISD::LOAD, MVT::v8i64, Legal);
1321 setOperationAction(ISD::LOAD, MVT::v16i32, Legal);
1322 setOperationAction(ISD::LOAD, MVT::v16i1, Legal);
1324 setOperationAction(ISD::FADD, MVT::v16f32, Legal);
1325 setOperationAction(ISD::FSUB, MVT::v16f32, Legal);
1326 setOperationAction(ISD::FMUL, MVT::v16f32, Legal);
1327 setOperationAction(ISD::FDIV, MVT::v16f32, Legal);
1328 setOperationAction(ISD::FSQRT, MVT::v16f32, Legal);
1329 setOperationAction(ISD::FNEG, MVT::v16f32, Custom);
1331 setOperationAction(ISD::FADD, MVT::v8f64, Legal);
1332 setOperationAction(ISD::FSUB, MVT::v8f64, Legal);
1333 setOperationAction(ISD::FMUL, MVT::v8f64, Legal);
1334 setOperationAction(ISD::FDIV, MVT::v8f64, Legal);
1335 setOperationAction(ISD::FSQRT, MVT::v8f64, Legal);
1336 setOperationAction(ISD::FNEG, MVT::v8f64, Custom);
1337 setOperationAction(ISD::FMA, MVT::v8f64, Legal);
1338 setOperationAction(ISD::FMA, MVT::v16f32, Legal);
1340 // FIXME: [US]INT_TO_FP are not legal for f80.
1341 setOperationAction(ISD::SINT_TO_FP, MVT::i32, Legal);
1342 setOperationAction(ISD::UINT_TO_FP, MVT::i32, Legal);
1343 if (Subtarget->is64Bit()) {
1344 setOperationAction(ISD::SINT_TO_FP, MVT::i64, Legal);
1345 setOperationAction(ISD::UINT_TO_FP, MVT::i64, Legal);
1347 setOperationAction(ISD::FP_TO_SINT, MVT::v16i32, Legal);
1348 setOperationAction(ISD::FP_TO_UINT, MVT::v16i32, Legal);
1349 setOperationAction(ISD::FP_TO_UINT, MVT::v8i32, Legal);
1350 setOperationAction(ISD::FP_TO_UINT, MVT::v4i32, Legal);
1351 setOperationAction(ISD::SINT_TO_FP, MVT::v16i32, Legal);
1352 setOperationAction(ISD::SINT_TO_FP, MVT::v8i1, Custom);
1353 setOperationAction(ISD::SINT_TO_FP, MVT::v16i1, Custom);
1354 setOperationAction(ISD::SINT_TO_FP, MVT::v16i8, Promote);
1355 setOperationAction(ISD::SINT_TO_FP, MVT::v16i16, Promote);
1356 setOperationAction(ISD::UINT_TO_FP, MVT::v16i32, Legal);
1357 setOperationAction(ISD::UINT_TO_FP, MVT::v8i32, Legal);
1358 setOperationAction(ISD::UINT_TO_FP, MVT::v4i32, Legal);
1359 setOperationAction(ISD::UINT_TO_FP, MVT::v16i8, Custom);
1360 setOperationAction(ISD::UINT_TO_FP, MVT::v16i16, Custom);
1361 setOperationAction(ISD::FP_ROUND, MVT::v8f32, Legal);
1362 setOperationAction(ISD::FP_EXTEND, MVT::v8f32, Legal);
1364 setTruncStoreAction(MVT::v8i64, MVT::v8i8, Legal);
1365 setTruncStoreAction(MVT::v8i64, MVT::v8i16, Legal);
1366 setTruncStoreAction(MVT::v8i64, MVT::v8i32, Legal);
1367 setTruncStoreAction(MVT::v16i32, MVT::v16i8, Legal);
1368 setTruncStoreAction(MVT::v16i32, MVT::v16i16, Legal);
1369 if (Subtarget->hasVLX()){
1370 setTruncStoreAction(MVT::v4i64, MVT::v4i8, Legal);
1371 setTruncStoreAction(MVT::v4i64, MVT::v4i16, Legal);
1372 setTruncStoreAction(MVT::v4i64, MVT::v4i32, Legal);
1373 setTruncStoreAction(MVT::v8i32, MVT::v8i8, Legal);
1374 setTruncStoreAction(MVT::v8i32, MVT::v8i16, Legal);
1376 setTruncStoreAction(MVT::v2i64, MVT::v2i8, Legal);
1377 setTruncStoreAction(MVT::v2i64, MVT::v2i16, Legal);
1378 setTruncStoreAction(MVT::v2i64, MVT::v2i32, Legal);
1379 setTruncStoreAction(MVT::v4i32, MVT::v4i8, Legal);
1380 setTruncStoreAction(MVT::v4i32, MVT::v4i16, Legal);
1382 setOperationAction(ISD::TRUNCATE, MVT::i1, Custom);
1383 setOperationAction(ISD::TRUNCATE, MVT::v16i8, Custom);
1384 setOperationAction(ISD::TRUNCATE, MVT::v8i32, Custom);
1385 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v8i1, Custom);
1386 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v16i1, Custom);
1387 if (Subtarget->hasDQI()) {
1388 setOperationAction(ISD::TRUNCATE, MVT::v2i1, Custom);
1389 setOperationAction(ISD::TRUNCATE, MVT::v4i1, Custom);
1391 setOperationAction(ISD::SINT_TO_FP, MVT::v8i64, Legal);
1392 setOperationAction(ISD::UINT_TO_FP, MVT::v8i64, Legal);
1393 setOperationAction(ISD::FP_TO_SINT, MVT::v8i64, Legal);
1394 setOperationAction(ISD::FP_TO_UINT, MVT::v8i64, Legal);
1395 if (Subtarget->hasVLX()) {
1396 setOperationAction(ISD::SINT_TO_FP, MVT::v4i64, Legal);
1397 setOperationAction(ISD::SINT_TO_FP, MVT::v2i64, Legal);
1398 setOperationAction(ISD::UINT_TO_FP, MVT::v4i64, Legal);
1399 setOperationAction(ISD::UINT_TO_FP, MVT::v2i64, Legal);
1400 setOperationAction(ISD::FP_TO_SINT, MVT::v4i64, Legal);
1401 setOperationAction(ISD::FP_TO_SINT, MVT::v2i64, Legal);
1402 setOperationAction(ISD::FP_TO_UINT, MVT::v4i64, Legal);
1403 setOperationAction(ISD::FP_TO_UINT, MVT::v2i64, Legal);
1406 if (Subtarget->hasVLX()) {
1407 setOperationAction(ISD::SINT_TO_FP, MVT::v8i32, Legal);
1408 setOperationAction(ISD::UINT_TO_FP, MVT::v8i32, Legal);
1409 setOperationAction(ISD::FP_TO_SINT, MVT::v8i32, Legal);
1410 setOperationAction(ISD::FP_TO_UINT, MVT::v8i32, Legal);
1411 setOperationAction(ISD::SINT_TO_FP, MVT::v4i32, Legal);
1412 setOperationAction(ISD::UINT_TO_FP, MVT::v4i32, Legal);
1413 setOperationAction(ISD::FP_TO_SINT, MVT::v4i32, Legal);
1414 setOperationAction(ISD::FP_TO_UINT, MVT::v4i32, Legal);
1416 setOperationAction(ISD::TRUNCATE, MVT::v8i1, Custom);
1417 setOperationAction(ISD::TRUNCATE, MVT::v16i1, Custom);
1418 setOperationAction(ISD::TRUNCATE, MVT::v16i16, Custom);
1419 setOperationAction(ISD::ZERO_EXTEND, MVT::v16i32, Custom);
1420 setOperationAction(ISD::ZERO_EXTEND, MVT::v8i64, Custom);
1421 setOperationAction(ISD::ANY_EXTEND, MVT::v16i32, Custom);
1422 setOperationAction(ISD::ANY_EXTEND, MVT::v8i64, Custom);
1423 setOperationAction(ISD::SIGN_EXTEND, MVT::v16i32, Custom);
1424 setOperationAction(ISD::SIGN_EXTEND, MVT::v8i64, Custom);
1425 setOperationAction(ISD::SIGN_EXTEND, MVT::v16i8, Custom);
1426 setOperationAction(ISD::SIGN_EXTEND, MVT::v8i16, Custom);
1427 setOperationAction(ISD::SIGN_EXTEND, MVT::v16i16, Custom);
1428 if (Subtarget->hasDQI()) {
1429 setOperationAction(ISD::SIGN_EXTEND, MVT::v4i32, Custom);
1430 setOperationAction(ISD::SIGN_EXTEND, MVT::v2i64, Custom);
1432 setOperationAction(ISD::FFLOOR, MVT::v16f32, Legal);
1433 setOperationAction(ISD::FFLOOR, MVT::v8f64, Legal);
1434 setOperationAction(ISD::FCEIL, MVT::v16f32, Legal);
1435 setOperationAction(ISD::FCEIL, MVT::v8f64, Legal);
1436 setOperationAction(ISD::FTRUNC, MVT::v16f32, Legal);
1437 setOperationAction(ISD::FTRUNC, MVT::v8f64, Legal);
1438 setOperationAction(ISD::FRINT, MVT::v16f32, Legal);
1439 setOperationAction(ISD::FRINT, MVT::v8f64, Legal);
1440 setOperationAction(ISD::FNEARBYINT, MVT::v16f32, Legal);
1441 setOperationAction(ISD::FNEARBYINT, MVT::v8f64, Legal);
1443 setOperationAction(ISD::CONCAT_VECTORS, MVT::v8f64, Custom);
1444 setOperationAction(ISD::CONCAT_VECTORS, MVT::v8i64, Custom);
1445 setOperationAction(ISD::CONCAT_VECTORS, MVT::v16f32, Custom);
1446 setOperationAction(ISD::CONCAT_VECTORS, MVT::v16i32, Custom);
1447 setOperationAction(ISD::CONCAT_VECTORS, MVT::v16i1, Legal);
1449 setOperationAction(ISD::SETCC, MVT::v16i1, Custom);
1450 setOperationAction(ISD::SETCC, MVT::v8i1, Custom);
1452 setOperationAction(ISD::MUL, MVT::v8i64, Custom);
1454 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v8i1, Custom);
1455 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v16i1, Custom);
1456 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v16i1, Custom);
1457 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i1, Custom);
1458 setOperationAction(ISD::BUILD_VECTOR, MVT::v8i1, Custom);
1459 setOperationAction(ISD::BUILD_VECTOR, MVT::v16i1, Custom);
1460 setOperationAction(ISD::SELECT, MVT::v8f64, Custom);
1461 setOperationAction(ISD::SELECT, MVT::v8i64, Custom);
1462 setOperationAction(ISD::SELECT, MVT::v16f32, Custom);
1463 setOperationAction(ISD::SELECT, MVT::v16i1, Custom);
1464 setOperationAction(ISD::SELECT, MVT::v8i1, Custom);
1466 setOperationAction(ISD::SMAX, MVT::v16i32, Legal);
1467 setOperationAction(ISD::SMAX, MVT::v8i64, Legal);
1468 setOperationAction(ISD::UMAX, MVT::v16i32, Legal);
1469 setOperationAction(ISD::UMAX, MVT::v8i64, Legal);
1470 setOperationAction(ISD::SMIN, MVT::v16i32, Legal);
1471 setOperationAction(ISD::SMIN, MVT::v8i64, Legal);
1472 setOperationAction(ISD::UMIN, MVT::v16i32, Legal);
1473 setOperationAction(ISD::UMIN, MVT::v8i64, Legal);
1475 setOperationAction(ISD::ADD, MVT::v8i64, Legal);
1476 setOperationAction(ISD::ADD, MVT::v16i32, Legal);
1478 setOperationAction(ISD::SUB, MVT::v8i64, Legal);
1479 setOperationAction(ISD::SUB, MVT::v16i32, Legal);
1481 setOperationAction(ISD::MUL, MVT::v16i32, Legal);
1483 setOperationAction(ISD::SRL, MVT::v8i64, Custom);
1484 setOperationAction(ISD::SRL, MVT::v16i32, Custom);
1486 setOperationAction(ISD::SHL, MVT::v8i64, Custom);
1487 setOperationAction(ISD::SHL, MVT::v16i32, Custom);
1489 setOperationAction(ISD::SRA, MVT::v8i64, Custom);
1490 setOperationAction(ISD::SRA, MVT::v16i32, Custom);
1492 setOperationAction(ISD::AND, MVT::v8i64, Legal);
1493 setOperationAction(ISD::OR, MVT::v8i64, Legal);
1494 setOperationAction(ISD::XOR, MVT::v8i64, Legal);
1495 setOperationAction(ISD::AND, MVT::v16i32, Legal);
1496 setOperationAction(ISD::OR, MVT::v16i32, Legal);
1497 setOperationAction(ISD::XOR, MVT::v16i32, Legal);
1499 if (Subtarget->hasCDI()) {
1500 setOperationAction(ISD::CTLZ, MVT::v8i64, Legal);
1501 setOperationAction(ISD::CTLZ, MVT::v16i32, Legal);
1502 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::v8i64, Legal);
1503 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::v16i32, Legal);
1505 if (Subtarget->hasVLX() && Subtarget->hasCDI()) {
1506 setOperationAction(ISD::CTLZ, MVT::v4i64, Legal);
1507 setOperationAction(ISD::CTLZ, MVT::v8i32, Legal);
1508 setOperationAction(ISD::CTLZ, MVT::v2i64, Legal);
1509 setOperationAction(ISD::CTLZ, MVT::v4i32, Legal);
1510 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::v4i64, Legal);
1511 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::v8i32, Legal);
1512 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::v2i64, Legal);
1513 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::v4i32, Legal);
1515 if (Subtarget->hasDQI()) {
1516 setOperationAction(ISD::MUL, MVT::v2i64, Legal);
1517 setOperationAction(ISD::MUL, MVT::v4i64, Legal);
1518 setOperationAction(ISD::MUL, MVT::v8i64, Legal);
1520 // Custom lower several nodes.
1521 for (MVT VT : MVT::vector_valuetypes()) {
1522 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
1524 setOperationAction(ISD::AND, VT, Legal);
1525 setOperationAction(ISD::OR, VT, Legal);
1526 setOperationAction(ISD::XOR, VT, Legal);
1528 if (EltSize >= 32 && VT.getSizeInBits() <= 512) {
1529 setOperationAction(ISD::MGATHER, VT, Custom);
1530 setOperationAction(ISD::MSCATTER, VT, Custom);
1532 // Extract subvector is special because the value type
1533 // (result) is 256/128-bit but the source is 512-bit wide.
1534 if (VT.is128BitVector() || VT.is256BitVector()) {
1535 setOperationAction(ISD::EXTRACT_SUBVECTOR, VT, Custom);
1537 if (VT.getVectorElementType() == MVT::i1)
1538 setOperationAction(ISD::EXTRACT_SUBVECTOR, VT, Legal);
1540 // Do not attempt to custom lower other non-512-bit vectors
1541 if (!VT.is512BitVector())
1544 if (EltSize >= 32) {
1545 setOperationAction(ISD::VECTOR_SHUFFLE, VT, Custom);
1546 setOperationAction(ISD::INSERT_VECTOR_ELT, VT, Custom);
1547 setOperationAction(ISD::BUILD_VECTOR, VT, Custom);
1548 setOperationAction(ISD::VSELECT, VT, Legal);
1549 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Custom);
1550 setOperationAction(ISD::SCALAR_TO_VECTOR, VT, Custom);
1551 setOperationAction(ISD::INSERT_SUBVECTOR, VT, Custom);
1552 setOperationAction(ISD::MLOAD, VT, Legal);
1553 setOperationAction(ISD::MSTORE, VT, Legal);
1556 for (int i = MVT::v32i8; i != MVT::v8i64; ++i) {
1557 MVT VT = (MVT::SimpleValueType)i;
1559 // Do not attempt to promote non-512-bit vectors.
1560 if (!VT.is512BitVector())
1563 setOperationAction(ISD::SELECT, VT, Promote);
1564 AddPromotedToType (ISD::SELECT, VT, MVT::v8i64);
1568 if (!Subtarget->useSoftFloat() && Subtarget->hasBWI()) {
1569 addRegisterClass(MVT::v32i16, &X86::VR512RegClass);
1570 addRegisterClass(MVT::v64i8, &X86::VR512RegClass);
1572 addRegisterClass(MVT::v32i1, &X86::VK32RegClass);
1573 addRegisterClass(MVT::v64i1, &X86::VK64RegClass);
1575 setOperationAction(ISD::LOAD, MVT::v32i16, Legal);
1576 setOperationAction(ISD::LOAD, MVT::v64i8, Legal);
1577 setOperationAction(ISD::SETCC, MVT::v32i1, Custom);
1578 setOperationAction(ISD::SETCC, MVT::v64i1, Custom);
1579 setOperationAction(ISD::ADD, MVT::v32i16, Legal);
1580 setOperationAction(ISD::ADD, MVT::v64i8, Legal);
1581 setOperationAction(ISD::SUB, MVT::v32i16, Legal);
1582 setOperationAction(ISD::SUB, MVT::v64i8, Legal);
1583 setOperationAction(ISD::MUL, MVT::v32i16, Legal);
1584 setOperationAction(ISD::MULHS, MVT::v32i16, Legal);
1585 setOperationAction(ISD::MULHU, MVT::v32i16, Legal);
1586 setOperationAction(ISD::CONCAT_VECTORS, MVT::v32i1, Legal);
1587 setOperationAction(ISD::CONCAT_VECTORS, MVT::v64i1, Legal);
1588 setOperationAction(ISD::INSERT_SUBVECTOR, MVT::v32i1, Custom);
1589 setOperationAction(ISD::INSERT_SUBVECTOR, MVT::v64i1, Custom);
1590 setOperationAction(ISD::SELECT, MVT::v32i1, Custom);
1591 setOperationAction(ISD::SELECT, MVT::v64i1, Custom);
1592 setOperationAction(ISD::SIGN_EXTEND, MVT::v32i8, Custom);
1593 setOperationAction(ISD::ZERO_EXTEND, MVT::v32i8, Custom);
1594 setOperationAction(ISD::SIGN_EXTEND, MVT::v32i16, Custom);
1595 setOperationAction(ISD::ZERO_EXTEND, MVT::v32i16, Custom);
1596 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v32i16, Custom);
1597 setOperationAction(ISD::SIGN_EXTEND, MVT::v64i8, Custom);
1598 setOperationAction(ISD::ZERO_EXTEND, MVT::v64i8, Custom);
1599 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v32i1, Custom);
1600 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v64i1, Custom);
1601 setOperationAction(ISD::VSELECT, MVT::v32i16, Legal);
1602 setOperationAction(ISD::VSELECT, MVT::v64i8, Legal);
1603 setOperationAction(ISD::TRUNCATE, MVT::v32i1, Custom);
1604 setOperationAction(ISD::TRUNCATE, MVT::v64i1, Custom);
1605 setOperationAction(ISD::TRUNCATE, MVT::v32i8, Custom);
1606 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v32i1, Custom);
1607 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v64i1, Custom);
1609 setOperationAction(ISD::SMAX, MVT::v64i8, Legal);
1610 setOperationAction(ISD::SMAX, MVT::v32i16, Legal);
1611 setOperationAction(ISD::UMAX, MVT::v64i8, Legal);
1612 setOperationAction(ISD::UMAX, MVT::v32i16, Legal);
1613 setOperationAction(ISD::SMIN, MVT::v64i8, Legal);
1614 setOperationAction(ISD::SMIN, MVT::v32i16, Legal);
1615 setOperationAction(ISD::UMIN, MVT::v64i8, Legal);
1616 setOperationAction(ISD::UMIN, MVT::v32i16, Legal);
1618 setTruncStoreAction(MVT::v32i16, MVT::v32i8, Legal);
1619 setTruncStoreAction(MVT::v16i16, MVT::v16i8, Legal);
1620 if (Subtarget->hasVLX())
1621 setTruncStoreAction(MVT::v8i16, MVT::v8i8, Legal);
1623 for (int i = MVT::v32i8; i != MVT::v8i64; ++i) {
1624 const MVT VT = (MVT::SimpleValueType)i;
1626 const unsigned EltSize = VT.getVectorElementType().getSizeInBits();
1628 // Do not attempt to promote non-512-bit vectors.
1629 if (!VT.is512BitVector())
1633 setOperationAction(ISD::BUILD_VECTOR, VT, Custom);
1634 setOperationAction(ISD::VSELECT, VT, Legal);
1639 if (!Subtarget->useSoftFloat() && Subtarget->hasVLX()) {
1640 addRegisterClass(MVT::v4i1, &X86::VK4RegClass);
1641 addRegisterClass(MVT::v2i1, &X86::VK2RegClass);
1643 setOperationAction(ISD::SETCC, MVT::v4i1, Custom);
1644 setOperationAction(ISD::SETCC, MVT::v2i1, Custom);
1645 setOperationAction(ISD::CONCAT_VECTORS, MVT::v4i1, Custom);
1646 setOperationAction(ISD::CONCAT_VECTORS, MVT::v8i1, Custom);
1647 setOperationAction(ISD::INSERT_SUBVECTOR, MVT::v8i1, Custom);
1648 setOperationAction(ISD::INSERT_SUBVECTOR, MVT::v4i1, Custom);
1649 setOperationAction(ISD::SELECT, MVT::v4i1, Custom);
1650 setOperationAction(ISD::SELECT, MVT::v2i1, Custom);
1651 setOperationAction(ISD::BUILD_VECTOR, MVT::v4i1, Custom);
1652 setOperationAction(ISD::BUILD_VECTOR, MVT::v2i1, Custom);
1653 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2i1, Custom);
1654 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4i1, Custom);
1656 setOperationAction(ISD::AND, MVT::v8i32, Legal);
1657 setOperationAction(ISD::OR, MVT::v8i32, Legal);
1658 setOperationAction(ISD::XOR, MVT::v8i32, Legal);
1659 setOperationAction(ISD::AND, MVT::v4i32, Legal);
1660 setOperationAction(ISD::OR, MVT::v4i32, Legal);
1661 setOperationAction(ISD::XOR, MVT::v4i32, Legal);
1662 setOperationAction(ISD::SRA, MVT::v2i64, Custom);
1663 setOperationAction(ISD::SRA, MVT::v4i64, Custom);
1665 setOperationAction(ISD::SMAX, MVT::v2i64, Legal);
1666 setOperationAction(ISD::SMAX, MVT::v4i64, Legal);
1667 setOperationAction(ISD::UMAX, MVT::v2i64, Legal);
1668 setOperationAction(ISD::UMAX, MVT::v4i64, Legal);
1669 setOperationAction(ISD::SMIN, MVT::v2i64, Legal);
1670 setOperationAction(ISD::SMIN, MVT::v4i64, Legal);
1671 setOperationAction(ISD::UMIN, MVT::v2i64, Legal);
1672 setOperationAction(ISD::UMIN, MVT::v4i64, Legal);
1675 // We want to custom lower some of our intrinsics.
1676 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
1677 setOperationAction(ISD::INTRINSIC_W_CHAIN, MVT::Other, Custom);
1678 setOperationAction(ISD::INTRINSIC_VOID, MVT::Other, Custom);
1679 if (!Subtarget->is64Bit())
1680 setOperationAction(ISD::INTRINSIC_W_CHAIN, MVT::i64, Custom);
1682 // Only custom-lower 64-bit SADDO and friends on 64-bit because we don't
1683 // handle type legalization for these operations here.
1685 // FIXME: We really should do custom legalization for addition and
1686 // subtraction on x86-32 once PR3203 is fixed. We really can't do much better
1687 // than generic legalization for 64-bit multiplication-with-overflow, though.
1688 for (unsigned i = 0, e = 3+Subtarget->is64Bit(); i != e; ++i) {
1689 // Add/Sub/Mul with overflow operations are custom lowered.
1691 setOperationAction(ISD::SADDO, VT, Custom);
1692 setOperationAction(ISD::UADDO, VT, Custom);
1693 setOperationAction(ISD::SSUBO, VT, Custom);
1694 setOperationAction(ISD::USUBO, VT, Custom);
1695 setOperationAction(ISD::SMULO, VT, Custom);
1696 setOperationAction(ISD::UMULO, VT, Custom);
1700 if (!Subtarget->is64Bit()) {
1701 // These libcalls are not available in 32-bit.
1702 setLibcallName(RTLIB::SHL_I128, nullptr);
1703 setLibcallName(RTLIB::SRL_I128, nullptr);
1704 setLibcallName(RTLIB::SRA_I128, nullptr);
1707 // Combine sin / cos into one node or libcall if possible.
1708 if (Subtarget->hasSinCos()) {
1709 setLibcallName(RTLIB::SINCOS_F32, "sincosf");
1710 setLibcallName(RTLIB::SINCOS_F64, "sincos");
1711 if (Subtarget->isTargetDarwin()) {
1712 // For MacOSX, we don't want the normal expansion of a libcall to sincos.
1713 // We want to issue a libcall to __sincos_stret to avoid memory traffic.
1714 setOperationAction(ISD::FSINCOS, MVT::f64, Custom);
1715 setOperationAction(ISD::FSINCOS, MVT::f32, Custom);
1719 if (Subtarget->isTargetWin64()) {
1720 setOperationAction(ISD::SDIV, MVT::i128, Custom);
1721 setOperationAction(ISD::UDIV, MVT::i128, Custom);
1722 setOperationAction(ISD::SREM, MVT::i128, Custom);
1723 setOperationAction(ISD::UREM, MVT::i128, Custom);
1724 setOperationAction(ISD::SDIVREM, MVT::i128, Custom);
1725 setOperationAction(ISD::UDIVREM, MVT::i128, Custom);
1728 // We have target-specific dag combine patterns for the following nodes:
1729 setTargetDAGCombine(ISD::VECTOR_SHUFFLE);
1730 setTargetDAGCombine(ISD::EXTRACT_VECTOR_ELT);
1731 setTargetDAGCombine(ISD::BITCAST);
1732 setTargetDAGCombine(ISD::VSELECT);
1733 setTargetDAGCombine(ISD::SELECT);
1734 setTargetDAGCombine(ISD::SHL);
1735 setTargetDAGCombine(ISD::SRA);
1736 setTargetDAGCombine(ISD::SRL);
1737 setTargetDAGCombine(ISD::OR);
1738 setTargetDAGCombine(ISD::AND);
1739 setTargetDAGCombine(ISD::ADD);
1740 setTargetDAGCombine(ISD::FADD);
1741 setTargetDAGCombine(ISD::FSUB);
1742 setTargetDAGCombine(ISD::FMA);
1743 setTargetDAGCombine(ISD::SUB);
1744 setTargetDAGCombine(ISD::LOAD);
1745 setTargetDAGCombine(ISD::MLOAD);
1746 setTargetDAGCombine(ISD::STORE);
1747 setTargetDAGCombine(ISD::MSTORE);
1748 setTargetDAGCombine(ISD::ZERO_EXTEND);
1749 setTargetDAGCombine(ISD::ANY_EXTEND);
1750 setTargetDAGCombine(ISD::SIGN_EXTEND);
1751 setTargetDAGCombine(ISD::SIGN_EXTEND_INREG);
1752 setTargetDAGCombine(ISD::SINT_TO_FP);
1753 setTargetDAGCombine(ISD::UINT_TO_FP);
1754 setTargetDAGCombine(ISD::SETCC);
1755 setTargetDAGCombine(ISD::BUILD_VECTOR);
1756 setTargetDAGCombine(ISD::MUL);
1757 setTargetDAGCombine(ISD::XOR);
1759 computeRegisterProperties(Subtarget->getRegisterInfo());
1761 MaxStoresPerMemset = 16; // For @llvm.memset -> sequence of stores
1762 MaxStoresPerMemsetOptSize = 8;
1763 MaxStoresPerMemcpy = 8; // For @llvm.memcpy -> sequence of stores
1764 MaxStoresPerMemcpyOptSize = 4;
1765 MaxStoresPerMemmove = 8; // For @llvm.memmove -> sequence of stores
1766 MaxStoresPerMemmoveOptSize = 4;
1767 setPrefLoopAlignment(4); // 2^4 bytes.
1769 // Predictable cmov don't hurt on atom because it's in-order.
1770 PredictableSelectIsExpensive = !Subtarget->isAtom();
1771 EnableExtLdPromotion = true;
1772 setPrefFunctionAlignment(4); // 2^4 bytes.
1774 verifyIntrinsicTables();
1777 // This has so far only been implemented for 64-bit MachO.
1778 bool X86TargetLowering::useLoadStackGuardNode() const {
1779 return Subtarget->isTargetMachO() && Subtarget->is64Bit();
1782 TargetLoweringBase::LegalizeTypeAction
1783 X86TargetLowering::getPreferredVectorAction(EVT VT) const {
1784 if (ExperimentalVectorWideningLegalization &&
1785 VT.getVectorNumElements() != 1 &&
1786 VT.getVectorElementType().getSimpleVT() != MVT::i1)
1787 return TypeWidenVector;
1789 return TargetLoweringBase::getPreferredVectorAction(VT);
1792 EVT X86TargetLowering::getSetCCResultType(const DataLayout &DL, LLVMContext &,
1795 return Subtarget->hasAVX512() ? MVT::i1: MVT::i8;
1797 const unsigned NumElts = VT.getVectorNumElements();
1798 const EVT EltVT = VT.getVectorElementType();
1799 if (VT.is512BitVector()) {
1800 if (Subtarget->hasAVX512())
1801 if (EltVT == MVT::i32 || EltVT == MVT::i64 ||
1802 EltVT == MVT::f32 || EltVT == MVT::f64)
1804 case 8: return MVT::v8i1;
1805 case 16: return MVT::v16i1;
1807 if (Subtarget->hasBWI())
1808 if (EltVT == MVT::i8 || EltVT == MVT::i16)
1810 case 32: return MVT::v32i1;
1811 case 64: return MVT::v64i1;
1815 if (VT.is256BitVector() || VT.is128BitVector()) {
1816 if (Subtarget->hasVLX())
1817 if (EltVT == MVT::i32 || EltVT == MVT::i64 ||
1818 EltVT == MVT::f32 || EltVT == MVT::f64)
1820 case 2: return MVT::v2i1;
1821 case 4: return MVT::v4i1;
1822 case 8: return MVT::v8i1;
1824 if (Subtarget->hasBWI() && Subtarget->hasVLX())
1825 if (EltVT == MVT::i8 || EltVT == MVT::i16)
1827 case 8: return MVT::v8i1;
1828 case 16: return MVT::v16i1;
1829 case 32: return MVT::v32i1;
1833 return VT.changeVectorElementTypeToInteger();
1836 /// Helper for getByValTypeAlignment to determine
1837 /// the desired ByVal argument alignment.
1838 static void getMaxByValAlign(Type *Ty, unsigned &MaxAlign) {
1841 if (VectorType *VTy = dyn_cast<VectorType>(Ty)) {
1842 if (VTy->getBitWidth() == 128)
1844 } else if (ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
1845 unsigned EltAlign = 0;
1846 getMaxByValAlign(ATy->getElementType(), EltAlign);
1847 if (EltAlign > MaxAlign)
1848 MaxAlign = EltAlign;
1849 } else if (StructType *STy = dyn_cast<StructType>(Ty)) {
1850 for (auto *EltTy : STy->elements()) {
1851 unsigned EltAlign = 0;
1852 getMaxByValAlign(EltTy, EltAlign);
1853 if (EltAlign > MaxAlign)
1854 MaxAlign = EltAlign;
1861 /// Return the desired alignment for ByVal aggregate
1862 /// function arguments in the caller parameter area. For X86, aggregates
1863 /// that contain SSE vectors are placed at 16-byte boundaries while the rest
1864 /// are at 4-byte boundaries.
1865 unsigned X86TargetLowering::getByValTypeAlignment(Type *Ty,
1866 const DataLayout &DL) const {
1867 if (Subtarget->is64Bit()) {
1868 // Max of 8 and alignment of type.
1869 unsigned TyAlign = DL.getABITypeAlignment(Ty);
1876 if (Subtarget->hasSSE1())
1877 getMaxByValAlign(Ty, Align);
1881 /// Returns the target specific optimal type for load
1882 /// and store operations as a result of memset, memcpy, and memmove
1883 /// lowering. If DstAlign is zero that means it's safe to destination
1884 /// alignment can satisfy any constraint. Similarly if SrcAlign is zero it
1885 /// means there isn't a need to check it against alignment requirement,
1886 /// probably because the source does not need to be loaded. If 'IsMemset' is
1887 /// true, that means it's expanding a memset. If 'ZeroMemset' is true, that
1888 /// means it's a memset of zero. 'MemcpyStrSrc' indicates whether the memcpy
1889 /// source is constant so it does not need to be loaded.
1890 /// It returns EVT::Other if the type should be determined using generic
1891 /// target-independent logic.
1893 X86TargetLowering::getOptimalMemOpType(uint64_t Size,
1894 unsigned DstAlign, unsigned SrcAlign,
1895 bool IsMemset, bool ZeroMemset,
1897 MachineFunction &MF) const {
1898 const Function *F = MF.getFunction();
1899 if ((!IsMemset || ZeroMemset) &&
1900 !F->hasFnAttribute(Attribute::NoImplicitFloat)) {
1902 (!Subtarget->isUnalignedMem16Slow() ||
1903 ((DstAlign == 0 || DstAlign >= 16) &&
1904 (SrcAlign == 0 || SrcAlign >= 16)))) {
1906 // FIXME: Check if unaligned 32-byte accesses are slow.
1907 if (Subtarget->hasInt256())
1909 if (Subtarget->hasFp256())
1912 if (Subtarget->hasSSE2())
1914 if (Subtarget->hasSSE1())
1916 } else if (!MemcpyStrSrc && Size >= 8 &&
1917 !Subtarget->is64Bit() &&
1918 Subtarget->hasSSE2()) {
1919 // Do not use f64 to lower memcpy if source is string constant. It's
1920 // better to use i32 to avoid the loads.
1924 // This is a compromise. If we reach here, unaligned accesses may be slow on
1925 // this target. However, creating smaller, aligned accesses could be even
1926 // slower and would certainly be a lot more code.
1927 if (Subtarget->is64Bit() && Size >= 8)
1932 bool X86TargetLowering::isSafeMemOpType(MVT VT) const {
1934 return X86ScalarSSEf32;
1935 else if (VT == MVT::f64)
1936 return X86ScalarSSEf64;
1941 X86TargetLowering::allowsMisalignedMemoryAccesses(EVT VT,
1946 switch (VT.getSizeInBits()) {
1948 // 8-byte and under are always assumed to be fast.
1952 *Fast = !Subtarget->isUnalignedMem16Slow();
1955 *Fast = !Subtarget->isUnalignedMem32Slow();
1957 // TODO: What about AVX-512 (512-bit) accesses?
1960 // Misaligned accesses of any size are always allowed.
1964 /// Return the entry encoding for a jump table in the
1965 /// current function. The returned value is a member of the
1966 /// MachineJumpTableInfo::JTEntryKind enum.
1967 unsigned X86TargetLowering::getJumpTableEncoding() const {
1968 // In GOT pic mode, each entry in the jump table is emitted as a @GOTOFF
1970 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
1971 Subtarget->isPICStyleGOT())
1972 return MachineJumpTableInfo::EK_Custom32;
1974 // Otherwise, use the normal jump table encoding heuristics.
1975 return TargetLowering::getJumpTableEncoding();
1978 bool X86TargetLowering::useSoftFloat() const {
1979 return Subtarget->useSoftFloat();
1983 X86TargetLowering::LowerCustomJumpTableEntry(const MachineJumpTableInfo *MJTI,
1984 const MachineBasicBlock *MBB,
1985 unsigned uid,MCContext &Ctx) const{
1986 assert(MBB->getParent()->getTarget().getRelocationModel() == Reloc::PIC_ &&
1987 Subtarget->isPICStyleGOT());
1988 // In 32-bit ELF systems, our jump table entries are formed with @GOTOFF
1990 return MCSymbolRefExpr::create(MBB->getSymbol(),
1991 MCSymbolRefExpr::VK_GOTOFF, Ctx);
1994 /// Returns relocation base for the given PIC jumptable.
1995 SDValue X86TargetLowering::getPICJumpTableRelocBase(SDValue Table,
1996 SelectionDAG &DAG) const {
1997 if (!Subtarget->is64Bit())
1998 // This doesn't have SDLoc associated with it, but is not really the
1999 // same as a Register.
2000 return DAG.getNode(X86ISD::GlobalBaseReg, SDLoc(),
2001 getPointerTy(DAG.getDataLayout()));
2005 /// This returns the relocation base for the given PIC jumptable,
2006 /// the same as getPICJumpTableRelocBase, but as an MCExpr.
2007 const MCExpr *X86TargetLowering::
2008 getPICJumpTableRelocBaseExpr(const MachineFunction *MF, unsigned JTI,
2009 MCContext &Ctx) const {
2010 // X86-64 uses RIP relative addressing based on the jump table label.
2011 if (Subtarget->isPICStyleRIPRel())
2012 return TargetLowering::getPICJumpTableRelocBaseExpr(MF, JTI, Ctx);
2014 // Otherwise, the reference is relative to the PIC base.
2015 return MCSymbolRefExpr::create(MF->getPICBaseSymbol(), Ctx);
2018 std::pair<const TargetRegisterClass *, uint8_t>
2019 X86TargetLowering::findRepresentativeClass(const TargetRegisterInfo *TRI,
2021 const TargetRegisterClass *RRC = nullptr;
2023 switch (VT.SimpleTy) {
2025 return TargetLowering::findRepresentativeClass(TRI, VT);
2026 case MVT::i8: case MVT::i16: case MVT::i32: case MVT::i64:
2027 RRC = Subtarget->is64Bit() ? &X86::GR64RegClass : &X86::GR32RegClass;
2030 RRC = &X86::VR64RegClass;
2032 case MVT::f32: case MVT::f64:
2033 case MVT::v16i8: case MVT::v8i16: case MVT::v4i32: case MVT::v2i64:
2034 case MVT::v4f32: case MVT::v2f64:
2035 case MVT::v32i8: case MVT::v8i32: case MVT::v4i64: case MVT::v8f32:
2037 RRC = &X86::VR128RegClass;
2040 return std::make_pair(RRC, Cost);
2043 bool X86TargetLowering::getStackCookieLocation(unsigned &AddressSpace,
2044 unsigned &Offset) const {
2045 if (!Subtarget->isTargetLinux())
2048 if (Subtarget->is64Bit()) {
2049 // %fs:0x28, unless we're using a Kernel code model, in which case it's %gs:
2051 if (getTargetMachine().getCodeModel() == CodeModel::Kernel)
2063 bool X86TargetLowering::isNoopAddrSpaceCast(unsigned SrcAS,
2064 unsigned DestAS) const {
2065 assert(SrcAS != DestAS && "Expected different address spaces!");
2067 return SrcAS < 256 && DestAS < 256;
2070 //===----------------------------------------------------------------------===//
2071 // Return Value Calling Convention Implementation
2072 //===----------------------------------------------------------------------===//
2074 #include "X86GenCallingConv.inc"
2077 X86TargetLowering::CanLowerReturn(CallingConv::ID CallConv,
2078 MachineFunction &MF, bool isVarArg,
2079 const SmallVectorImpl<ISD::OutputArg> &Outs,
2080 LLVMContext &Context) const {
2081 SmallVector<CCValAssign, 16> RVLocs;
2082 CCState CCInfo(CallConv, isVarArg, MF, RVLocs, Context);
2083 return CCInfo.CheckReturn(Outs, RetCC_X86);
2086 const MCPhysReg *X86TargetLowering::getScratchRegisters(CallingConv::ID) const {
2087 static const MCPhysReg ScratchRegs[] = { X86::R11, 0 };
2092 X86TargetLowering::LowerReturn(SDValue Chain,
2093 CallingConv::ID CallConv, bool isVarArg,
2094 const SmallVectorImpl<ISD::OutputArg> &Outs,
2095 const SmallVectorImpl<SDValue> &OutVals,
2096 SDLoc dl, SelectionDAG &DAG) const {
2097 MachineFunction &MF = DAG.getMachineFunction();
2098 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
2100 SmallVector<CCValAssign, 16> RVLocs;
2101 CCState CCInfo(CallConv, isVarArg, MF, RVLocs, *DAG.getContext());
2102 CCInfo.AnalyzeReturn(Outs, RetCC_X86);
2105 SmallVector<SDValue, 6> RetOps;
2106 RetOps.push_back(Chain); // Operand #0 = Chain (updated below)
2107 // Operand #1 = Bytes To Pop
2108 RetOps.push_back(DAG.getTargetConstant(FuncInfo->getBytesToPopOnReturn(), dl,
2111 // Copy the result values into the output registers.
2112 for (unsigned i = 0; i != RVLocs.size(); ++i) {
2113 CCValAssign &VA = RVLocs[i];
2114 assert(VA.isRegLoc() && "Can only return in registers!");
2115 SDValue ValToCopy = OutVals[i];
2116 EVT ValVT = ValToCopy.getValueType();
2118 // Promote values to the appropriate types.
2119 if (VA.getLocInfo() == CCValAssign::SExt)
2120 ValToCopy = DAG.getNode(ISD::SIGN_EXTEND, dl, VA.getLocVT(), ValToCopy);
2121 else if (VA.getLocInfo() == CCValAssign::ZExt)
2122 ValToCopy = DAG.getNode(ISD::ZERO_EXTEND, dl, VA.getLocVT(), ValToCopy);
2123 else if (VA.getLocInfo() == CCValAssign::AExt) {
2124 if (ValVT.isVector() && ValVT.getScalarType() == MVT::i1)
2125 ValToCopy = DAG.getNode(ISD::SIGN_EXTEND, dl, VA.getLocVT(), ValToCopy);
2127 ValToCopy = DAG.getNode(ISD::ANY_EXTEND, dl, VA.getLocVT(), ValToCopy);
2129 else if (VA.getLocInfo() == CCValAssign::BCvt)
2130 ValToCopy = DAG.getBitcast(VA.getLocVT(), ValToCopy);
2132 assert(VA.getLocInfo() != CCValAssign::FPExt &&
2133 "Unexpected FP-extend for return value.");
2135 // If this is x86-64, and we disabled SSE, we can't return FP values,
2136 // or SSE or MMX vectors.
2137 if ((ValVT == MVT::f32 || ValVT == MVT::f64 ||
2138 VA.getLocReg() == X86::XMM0 || VA.getLocReg() == X86::XMM1) &&
2139 (Subtarget->is64Bit() && !Subtarget->hasSSE1())) {
2140 report_fatal_error("SSE register return with SSE disabled");
2142 // Likewise we can't return F64 values with SSE1 only. gcc does so, but
2143 // llvm-gcc has never done it right and no one has noticed, so this
2144 // should be OK for now.
2145 if (ValVT == MVT::f64 &&
2146 (Subtarget->is64Bit() && !Subtarget->hasSSE2()))
2147 report_fatal_error("SSE2 register return with SSE2 disabled");
2149 // Returns in ST0/ST1 are handled specially: these are pushed as operands to
2150 // the RET instruction and handled by the FP Stackifier.
2151 if (VA.getLocReg() == X86::FP0 ||
2152 VA.getLocReg() == X86::FP1) {
2153 // If this is a copy from an xmm register to ST(0), use an FPExtend to
2154 // change the value to the FP stack register class.
2155 if (isScalarFPTypeInSSEReg(VA.getValVT()))
2156 ValToCopy = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f80, ValToCopy);
2157 RetOps.push_back(ValToCopy);
2158 // Don't emit a copytoreg.
2162 // 64-bit vector (MMX) values are returned in XMM0 / XMM1 except for v1i64
2163 // which is returned in RAX / RDX.
2164 if (Subtarget->is64Bit()) {
2165 if (ValVT == MVT::x86mmx) {
2166 if (VA.getLocReg() == X86::XMM0 || VA.getLocReg() == X86::XMM1) {
2167 ValToCopy = DAG.getBitcast(MVT::i64, ValToCopy);
2168 ValToCopy = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i64,
2170 // If we don't have SSE2 available, convert to v4f32 so the generated
2171 // register is legal.
2172 if (!Subtarget->hasSSE2())
2173 ValToCopy = DAG.getBitcast(MVT::v4f32, ValToCopy);
2178 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), ValToCopy, Flag);
2179 Flag = Chain.getValue(1);
2180 RetOps.push_back(DAG.getRegister(VA.getLocReg(), VA.getLocVT()));
2183 // All x86 ABIs require that for returning structs by value we copy
2184 // the sret argument into %rax/%eax (depending on ABI) for the return.
2185 // We saved the argument into a virtual register in the entry block,
2186 // so now we copy the value out and into %rax/%eax.
2188 // Checking Function.hasStructRetAttr() here is insufficient because the IR
2189 // may not have an explicit sret argument. If FuncInfo.CanLowerReturn is
2190 // false, then an sret argument may be implicitly inserted in the SelDAG. In
2191 // either case FuncInfo->setSRetReturnReg() will have been called.
2192 if (unsigned SRetReg = FuncInfo->getSRetReturnReg()) {
2193 SDValue Val = DAG.getCopyFromReg(Chain, dl, SRetReg,
2194 getPointerTy(MF.getDataLayout()));
2197 = (Subtarget->is64Bit() && !Subtarget->isTarget64BitILP32()) ?
2198 X86::RAX : X86::EAX;
2199 Chain = DAG.getCopyToReg(Chain, dl, RetValReg, Val, Flag);
2200 Flag = Chain.getValue(1);
2202 // RAX/EAX now acts like a return value.
2204 DAG.getRegister(RetValReg, getPointerTy(DAG.getDataLayout())));
2207 RetOps[0] = Chain; // Update chain.
2209 // Add the flag if we have it.
2211 RetOps.push_back(Flag);
2213 return DAG.getNode(X86ISD::RET_FLAG, dl, MVT::Other, RetOps);
2216 bool X86TargetLowering::isUsedByReturnOnly(SDNode *N, SDValue &Chain) const {
2217 if (N->getNumValues() != 1)
2219 if (!N->hasNUsesOfValue(1, 0))
2222 SDValue TCChain = Chain;
2223 SDNode *Copy = *N->use_begin();
2224 if (Copy->getOpcode() == ISD::CopyToReg) {
2225 // If the copy has a glue operand, we conservatively assume it isn't safe to
2226 // perform a tail call.
2227 if (Copy->getOperand(Copy->getNumOperands()-1).getValueType() == MVT::Glue)
2229 TCChain = Copy->getOperand(0);
2230 } else if (Copy->getOpcode() != ISD::FP_EXTEND)
2233 bool HasRet = false;
2234 for (SDNode::use_iterator UI = Copy->use_begin(), UE = Copy->use_end();
2236 if (UI->getOpcode() != X86ISD::RET_FLAG)
2238 // If we are returning more than one value, we can definitely
2239 // not make a tail call see PR19530
2240 if (UI->getNumOperands() > 4)
2242 if (UI->getNumOperands() == 4 &&
2243 UI->getOperand(UI->getNumOperands()-1).getValueType() != MVT::Glue)
2256 X86TargetLowering::getTypeForExtArgOrReturn(LLVMContext &Context, EVT VT,
2257 ISD::NodeType ExtendKind) const {
2259 // TODO: Is this also valid on 32-bit?
2260 if (Subtarget->is64Bit() && VT == MVT::i1 && ExtendKind == ISD::ZERO_EXTEND)
2261 ReturnMVT = MVT::i8;
2263 ReturnMVT = MVT::i32;
2265 EVT MinVT = getRegisterType(Context, ReturnMVT);
2266 return VT.bitsLT(MinVT) ? MinVT : VT;
2269 /// Lower the result values of a call into the
2270 /// appropriate copies out of appropriate physical registers.
2273 X86TargetLowering::LowerCallResult(SDValue Chain, SDValue InFlag,
2274 CallingConv::ID CallConv, bool isVarArg,
2275 const SmallVectorImpl<ISD::InputArg> &Ins,
2276 SDLoc dl, SelectionDAG &DAG,
2277 SmallVectorImpl<SDValue> &InVals) const {
2279 // Assign locations to each value returned by this call.
2280 SmallVector<CCValAssign, 16> RVLocs;
2281 bool Is64Bit = Subtarget->is64Bit();
2282 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(), RVLocs,
2284 CCInfo.AnalyzeCallResult(Ins, RetCC_X86);
2286 // Copy all of the result registers out of their specified physreg.
2287 for (unsigned i = 0, e = RVLocs.size(); i != e; ++i) {
2288 CCValAssign &VA = RVLocs[i];
2289 EVT CopyVT = VA.getLocVT();
2291 // If this is x86-64, and we disabled SSE, we can't return FP values
2292 if ((CopyVT == MVT::f32 || CopyVT == MVT::f64) &&
2293 ((Is64Bit || Ins[i].Flags.isInReg()) && !Subtarget->hasSSE1())) {
2294 report_fatal_error("SSE register return with SSE disabled");
2297 // If we prefer to use the value in xmm registers, copy it out as f80 and
2298 // use a truncate to move it from fp stack reg to xmm reg.
2299 bool RoundAfterCopy = false;
2300 if ((VA.getLocReg() == X86::FP0 || VA.getLocReg() == X86::FP1) &&
2301 isScalarFPTypeInSSEReg(VA.getValVT())) {
2303 RoundAfterCopy = (CopyVT != VA.getLocVT());
2306 Chain = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(),
2307 CopyVT, InFlag).getValue(1);
2308 SDValue Val = Chain.getValue(0);
2311 Val = DAG.getNode(ISD::FP_ROUND, dl, VA.getValVT(), Val,
2312 // This truncation won't change the value.
2313 DAG.getIntPtrConstant(1, dl));
2315 if (VA.isExtInLoc() && VA.getValVT().getScalarType() == MVT::i1)
2316 Val = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), Val);
2318 InFlag = Chain.getValue(2);
2319 InVals.push_back(Val);
2325 //===----------------------------------------------------------------------===//
2326 // C & StdCall & Fast Calling Convention implementation
2327 //===----------------------------------------------------------------------===//
2328 // StdCall calling convention seems to be standard for many Windows' API
2329 // routines and around. It differs from C calling convention just a little:
2330 // callee should clean up the stack, not caller. Symbols should be also
2331 // decorated in some fancy way :) It doesn't support any vector arguments.
2332 // For info on fast calling convention see Fast Calling Convention (tail call)
2333 // implementation LowerX86_32FastCCCallTo.
2335 /// CallIsStructReturn - Determines whether a call uses struct return
2337 enum StructReturnType {
2342 static StructReturnType
2343 callIsStructReturn(const SmallVectorImpl<ISD::OutputArg> &Outs) {
2345 return NotStructReturn;
2347 const ISD::ArgFlagsTy &Flags = Outs[0].Flags;
2348 if (!Flags.isSRet())
2349 return NotStructReturn;
2350 if (Flags.isInReg())
2351 return RegStructReturn;
2352 return StackStructReturn;
2355 /// Determines whether a function uses struct return semantics.
2356 static StructReturnType
2357 argsAreStructReturn(const SmallVectorImpl<ISD::InputArg> &Ins) {
2359 return NotStructReturn;
2361 const ISD::ArgFlagsTy &Flags = Ins[0].Flags;
2362 if (!Flags.isSRet())
2363 return NotStructReturn;
2364 if (Flags.isInReg())
2365 return RegStructReturn;
2366 return StackStructReturn;
2369 /// Make a copy of an aggregate at address specified by "Src" to address
2370 /// "Dst" with size and alignment information specified by the specific
2371 /// parameter attribute. The copy will be passed as a byval function parameter.
2373 CreateCopyOfByValArgument(SDValue Src, SDValue Dst, SDValue Chain,
2374 ISD::ArgFlagsTy Flags, SelectionDAG &DAG,
2376 SDValue SizeNode = DAG.getConstant(Flags.getByValSize(), dl, MVT::i32);
2378 return DAG.getMemcpy(Chain, dl, Dst, Src, SizeNode, Flags.getByValAlign(),
2379 /*isVolatile*/false, /*AlwaysInline=*/true,
2380 /*isTailCall*/false,
2381 MachinePointerInfo(), MachinePointerInfo());
2384 /// Return true if the calling convention is one that
2385 /// supports tail call optimization.
2386 static bool IsTailCallConvention(CallingConv::ID CC) {
2387 return (CC == CallingConv::Fast || CC == CallingConv::GHC ||
2388 CC == CallingConv::HiPE);
2391 /// \brief Return true if the calling convention is a C calling convention.
2392 static bool IsCCallConvention(CallingConv::ID CC) {
2393 return (CC == CallingConv::C || CC == CallingConv::X86_64_Win64 ||
2394 CC == CallingConv::X86_64_SysV);
2397 bool X86TargetLowering::mayBeEmittedAsTailCall(CallInst *CI) const {
2399 CI->getParent()->getParent()->getFnAttribute("disable-tail-calls");
2400 if (!CI->isTailCall() || Attr.getValueAsString() == "true")
2404 CallingConv::ID CalleeCC = CS.getCallingConv();
2405 if (!IsTailCallConvention(CalleeCC) && !IsCCallConvention(CalleeCC))
2411 /// Return true if the function is being made into
2412 /// a tailcall target by changing its ABI.
2413 static bool FuncIsMadeTailCallSafe(CallingConv::ID CC,
2414 bool GuaranteedTailCallOpt) {
2415 return GuaranteedTailCallOpt && IsTailCallConvention(CC);
2419 X86TargetLowering::LowerMemArgument(SDValue Chain,
2420 CallingConv::ID CallConv,
2421 const SmallVectorImpl<ISD::InputArg> &Ins,
2422 SDLoc dl, SelectionDAG &DAG,
2423 const CCValAssign &VA,
2424 MachineFrameInfo *MFI,
2426 // Create the nodes corresponding to a load from this parameter slot.
2427 ISD::ArgFlagsTy Flags = Ins[i].Flags;
2428 bool AlwaysUseMutable = FuncIsMadeTailCallSafe(
2429 CallConv, DAG.getTarget().Options.GuaranteedTailCallOpt);
2430 bool isImmutable = !AlwaysUseMutable && !Flags.isByVal();
2433 // If value is passed by pointer we have address passed instead of the value
2435 bool ExtendedInMem = VA.isExtInLoc() &&
2436 VA.getValVT().getScalarType() == MVT::i1;
2438 if (VA.getLocInfo() == CCValAssign::Indirect || ExtendedInMem)
2439 ValVT = VA.getLocVT();
2441 ValVT = VA.getValVT();
2443 // FIXME: For now, all byval parameter objects are marked mutable. This can be
2444 // changed with more analysis.
2445 // In case of tail call optimization mark all arguments mutable. Since they
2446 // could be overwritten by lowering of arguments in case of a tail call.
2447 if (Flags.isByVal()) {
2448 unsigned Bytes = Flags.getByValSize();
2449 if (Bytes == 0) Bytes = 1; // Don't create zero-sized stack objects.
2450 int FI = MFI->CreateFixedObject(Bytes, VA.getLocMemOffset(), isImmutable);
2451 return DAG.getFrameIndex(FI, getPointerTy(DAG.getDataLayout()));
2453 int FI = MFI->CreateFixedObject(ValVT.getSizeInBits()/8,
2454 VA.getLocMemOffset(), isImmutable);
2455 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy(DAG.getDataLayout()));
2456 SDValue Val = DAG.getLoad(
2457 ValVT, dl, Chain, FIN,
2458 MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), FI), false,
2460 return ExtendedInMem ?
2461 DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), Val) : Val;
2465 // FIXME: Get this from tablegen.
2466 static ArrayRef<MCPhysReg> get64BitArgumentGPRs(CallingConv::ID CallConv,
2467 const X86Subtarget *Subtarget) {
2468 assert(Subtarget->is64Bit());
2470 if (Subtarget->isCallingConvWin64(CallConv)) {
2471 static const MCPhysReg GPR64ArgRegsWin64[] = {
2472 X86::RCX, X86::RDX, X86::R8, X86::R9
2474 return makeArrayRef(std::begin(GPR64ArgRegsWin64), std::end(GPR64ArgRegsWin64));
2477 static const MCPhysReg GPR64ArgRegs64Bit[] = {
2478 X86::RDI, X86::RSI, X86::RDX, X86::RCX, X86::R8, X86::R9
2480 return makeArrayRef(std::begin(GPR64ArgRegs64Bit), std::end(GPR64ArgRegs64Bit));
2483 // FIXME: Get this from tablegen.
2484 static ArrayRef<MCPhysReg> get64BitArgumentXMMs(MachineFunction &MF,
2485 CallingConv::ID CallConv,
2486 const X86Subtarget *Subtarget) {
2487 assert(Subtarget->is64Bit());
2488 if (Subtarget->isCallingConvWin64(CallConv)) {
2489 // The XMM registers which might contain var arg parameters are shadowed
2490 // in their paired GPR. So we only need to save the GPR to their home
2492 // TODO: __vectorcall will change this.
2496 const Function *Fn = MF.getFunction();
2497 bool NoImplicitFloatOps = Fn->hasFnAttribute(Attribute::NoImplicitFloat);
2498 bool isSoftFloat = Subtarget->useSoftFloat();
2499 assert(!(isSoftFloat && NoImplicitFloatOps) &&
2500 "SSE register cannot be used when SSE is disabled!");
2501 if (isSoftFloat || NoImplicitFloatOps || !Subtarget->hasSSE1())
2502 // Kernel mode asks for SSE to be disabled, so there are no XMM argument
2506 static const MCPhysReg XMMArgRegs64Bit[] = {
2507 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
2508 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
2510 return makeArrayRef(std::begin(XMMArgRegs64Bit), std::end(XMMArgRegs64Bit));
2514 X86TargetLowering::LowerFormalArguments(SDValue Chain,
2515 CallingConv::ID CallConv,
2517 const SmallVectorImpl<ISD::InputArg> &Ins,
2520 SmallVectorImpl<SDValue> &InVals)
2522 MachineFunction &MF = DAG.getMachineFunction();
2523 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
2524 const TargetFrameLowering &TFI = *Subtarget->getFrameLowering();
2526 const Function* Fn = MF.getFunction();
2527 if (Fn->hasExternalLinkage() &&
2528 Subtarget->isTargetCygMing() &&
2529 Fn->getName() == "main")
2530 FuncInfo->setForceFramePointer(true);
2532 MachineFrameInfo *MFI = MF.getFrameInfo();
2533 bool Is64Bit = Subtarget->is64Bit();
2534 bool IsWin64 = Subtarget->isCallingConvWin64(CallConv);
2536 assert(!(isVarArg && IsTailCallConvention(CallConv)) &&
2537 "Var args not supported with calling convention fastcc, ghc or hipe");
2539 // Assign locations to all of the incoming arguments.
2540 SmallVector<CCValAssign, 16> ArgLocs;
2541 CCState CCInfo(CallConv, isVarArg, MF, ArgLocs, *DAG.getContext());
2543 // Allocate shadow area for Win64
2545 CCInfo.AllocateStack(32, 8);
2547 CCInfo.AnalyzeFormalArguments(Ins, CC_X86);
2549 unsigned LastVal = ~0U;
2551 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2552 CCValAssign &VA = ArgLocs[i];
2553 // TODO: If an arg is passed in two places (e.g. reg and stack), skip later
2555 assert(VA.getValNo() != LastVal &&
2556 "Don't support value assigned to multiple locs yet");
2558 LastVal = VA.getValNo();
2560 if (VA.isRegLoc()) {
2561 EVT RegVT = VA.getLocVT();
2562 const TargetRegisterClass *RC;
2563 if (RegVT == MVT::i32)
2564 RC = &X86::GR32RegClass;
2565 else if (Is64Bit && RegVT == MVT::i64)
2566 RC = &X86::GR64RegClass;
2567 else if (RegVT == MVT::f32)
2568 RC = &X86::FR32RegClass;
2569 else if (RegVT == MVT::f64)
2570 RC = &X86::FR64RegClass;
2571 else if (RegVT.is512BitVector())
2572 RC = &X86::VR512RegClass;
2573 else if (RegVT.is256BitVector())
2574 RC = &X86::VR256RegClass;
2575 else if (RegVT.is128BitVector())
2576 RC = &X86::VR128RegClass;
2577 else if (RegVT == MVT::x86mmx)
2578 RC = &X86::VR64RegClass;
2579 else if (RegVT == MVT::i1)
2580 RC = &X86::VK1RegClass;
2581 else if (RegVT == MVT::v8i1)
2582 RC = &X86::VK8RegClass;
2583 else if (RegVT == MVT::v16i1)
2584 RC = &X86::VK16RegClass;
2585 else if (RegVT == MVT::v32i1)
2586 RC = &X86::VK32RegClass;
2587 else if (RegVT == MVT::v64i1)
2588 RC = &X86::VK64RegClass;
2590 llvm_unreachable("Unknown argument type!");
2592 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
2593 ArgValue = DAG.getCopyFromReg(Chain, dl, Reg, RegVT);
2595 // If this is an 8 or 16-bit value, it is really passed promoted to 32
2596 // bits. Insert an assert[sz]ext to capture this, then truncate to the
2598 if (VA.getLocInfo() == CCValAssign::SExt)
2599 ArgValue = DAG.getNode(ISD::AssertSext, dl, RegVT, ArgValue,
2600 DAG.getValueType(VA.getValVT()));
2601 else if (VA.getLocInfo() == CCValAssign::ZExt)
2602 ArgValue = DAG.getNode(ISD::AssertZext, dl, RegVT, ArgValue,
2603 DAG.getValueType(VA.getValVT()));
2604 else if (VA.getLocInfo() == CCValAssign::BCvt)
2605 ArgValue = DAG.getBitcast(VA.getValVT(), ArgValue);
2607 if (VA.isExtInLoc()) {
2608 // Handle MMX values passed in XMM regs.
2609 if (RegVT.isVector() && VA.getValVT().getScalarType() != MVT::i1)
2610 ArgValue = DAG.getNode(X86ISD::MOVDQ2Q, dl, VA.getValVT(), ArgValue);
2612 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
2615 assert(VA.isMemLoc());
2616 ArgValue = LowerMemArgument(Chain, CallConv, Ins, dl, DAG, VA, MFI, i);
2619 // If value is passed via pointer - do a load.
2620 if (VA.getLocInfo() == CCValAssign::Indirect)
2621 ArgValue = DAG.getLoad(VA.getValVT(), dl, Chain, ArgValue,
2622 MachinePointerInfo(), false, false, false, 0);
2624 InVals.push_back(ArgValue);
2627 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2628 // All x86 ABIs require that for returning structs by value we copy the
2629 // sret argument into %rax/%eax (depending on ABI) for the return. Save
2630 // the argument into a virtual register so that we can access it from the
2632 if (Ins[i].Flags.isSRet()) {
2633 unsigned Reg = FuncInfo->getSRetReturnReg();
2635 MVT PtrTy = getPointerTy(DAG.getDataLayout());
2636 Reg = MF.getRegInfo().createVirtualRegister(getRegClassFor(PtrTy));
2637 FuncInfo->setSRetReturnReg(Reg);
2639 SDValue Copy = DAG.getCopyToReg(DAG.getEntryNode(), dl, Reg, InVals[i]);
2640 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Copy, Chain);
2645 unsigned StackSize = CCInfo.getNextStackOffset();
2646 // Align stack specially for tail calls.
2647 if (FuncIsMadeTailCallSafe(CallConv,
2648 MF.getTarget().Options.GuaranteedTailCallOpt))
2649 StackSize = GetAlignedArgumentStackSize(StackSize, DAG);
2651 // If the function takes variable number of arguments, make a frame index for
2652 // the start of the first vararg value... for expansion of llvm.va_start. We
2653 // can skip this if there are no va_start calls.
2654 if (MFI->hasVAStart() &&
2655 (Is64Bit || (CallConv != CallingConv::X86_FastCall &&
2656 CallConv != CallingConv::X86_ThisCall))) {
2657 FuncInfo->setVarArgsFrameIndex(
2658 MFI->CreateFixedObject(1, StackSize, true));
2661 MachineModuleInfo &MMI = MF.getMMI();
2662 const Function *WinEHParent = nullptr;
2663 if (MMI.hasWinEHFuncInfo(Fn))
2664 WinEHParent = MMI.getWinEHParent(Fn);
2665 bool IsWinEHParent = WinEHParent && WinEHParent == Fn;
2667 // Figure out if XMM registers are in use.
2668 assert(!(Subtarget->useSoftFloat() &&
2669 Fn->hasFnAttribute(Attribute::NoImplicitFloat)) &&
2670 "SSE register cannot be used when SSE is disabled!");
2672 // 64-bit calling conventions support varargs and register parameters, so we
2673 // have to do extra work to spill them in the prologue.
2674 if (Is64Bit && isVarArg && MFI->hasVAStart()) {
2675 // Find the first unallocated argument registers.
2676 ArrayRef<MCPhysReg> ArgGPRs = get64BitArgumentGPRs(CallConv, Subtarget);
2677 ArrayRef<MCPhysReg> ArgXMMs = get64BitArgumentXMMs(MF, CallConv, Subtarget);
2678 unsigned NumIntRegs = CCInfo.getFirstUnallocated(ArgGPRs);
2679 unsigned NumXMMRegs = CCInfo.getFirstUnallocated(ArgXMMs);
2680 assert(!(NumXMMRegs && !Subtarget->hasSSE1()) &&
2681 "SSE register cannot be used when SSE is disabled!");
2683 // Gather all the live in physical registers.
2684 SmallVector<SDValue, 6> LiveGPRs;
2685 SmallVector<SDValue, 8> LiveXMMRegs;
2687 for (MCPhysReg Reg : ArgGPRs.slice(NumIntRegs)) {
2688 unsigned GPR = MF.addLiveIn(Reg, &X86::GR64RegClass);
2690 DAG.getCopyFromReg(Chain, dl, GPR, MVT::i64));
2692 if (!ArgXMMs.empty()) {
2693 unsigned AL = MF.addLiveIn(X86::AL, &X86::GR8RegClass);
2694 ALVal = DAG.getCopyFromReg(Chain, dl, AL, MVT::i8);
2695 for (MCPhysReg Reg : ArgXMMs.slice(NumXMMRegs)) {
2696 unsigned XMMReg = MF.addLiveIn(Reg, &X86::VR128RegClass);
2697 LiveXMMRegs.push_back(
2698 DAG.getCopyFromReg(Chain, dl, XMMReg, MVT::v4f32));
2703 // Get to the caller-allocated home save location. Add 8 to account
2704 // for the return address.
2705 int HomeOffset = TFI.getOffsetOfLocalArea() + 8;
2706 FuncInfo->setRegSaveFrameIndex(
2707 MFI->CreateFixedObject(1, NumIntRegs * 8 + HomeOffset, false));
2708 // Fixup to set vararg frame on shadow area (4 x i64).
2710 FuncInfo->setVarArgsFrameIndex(FuncInfo->getRegSaveFrameIndex());
2712 // For X86-64, if there are vararg parameters that are passed via
2713 // registers, then we must store them to their spots on the stack so
2714 // they may be loaded by deferencing the result of va_next.
2715 FuncInfo->setVarArgsGPOffset(NumIntRegs * 8);
2716 FuncInfo->setVarArgsFPOffset(ArgGPRs.size() * 8 + NumXMMRegs * 16);
2717 FuncInfo->setRegSaveFrameIndex(MFI->CreateStackObject(
2718 ArgGPRs.size() * 8 + ArgXMMs.size() * 16, 16, false));
2721 // Store the integer parameter registers.
2722 SmallVector<SDValue, 8> MemOps;
2723 SDValue RSFIN = DAG.getFrameIndex(FuncInfo->getRegSaveFrameIndex(),
2724 getPointerTy(DAG.getDataLayout()));
2725 unsigned Offset = FuncInfo->getVarArgsGPOffset();
2726 for (SDValue Val : LiveGPRs) {
2727 SDValue FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(DAG.getDataLayout()),
2728 RSFIN, DAG.getIntPtrConstant(Offset, dl));
2730 DAG.getStore(Val.getValue(1), dl, Val, FIN,
2731 MachinePointerInfo::getFixedStack(
2732 DAG.getMachineFunction(),
2733 FuncInfo->getRegSaveFrameIndex(), Offset),
2735 MemOps.push_back(Store);
2739 if (!ArgXMMs.empty() && NumXMMRegs != ArgXMMs.size()) {
2740 // Now store the XMM (fp + vector) parameter registers.
2741 SmallVector<SDValue, 12> SaveXMMOps;
2742 SaveXMMOps.push_back(Chain);
2743 SaveXMMOps.push_back(ALVal);
2744 SaveXMMOps.push_back(DAG.getIntPtrConstant(
2745 FuncInfo->getRegSaveFrameIndex(), dl));
2746 SaveXMMOps.push_back(DAG.getIntPtrConstant(
2747 FuncInfo->getVarArgsFPOffset(), dl));
2748 SaveXMMOps.insert(SaveXMMOps.end(), LiveXMMRegs.begin(),
2750 MemOps.push_back(DAG.getNode(X86ISD::VASTART_SAVE_XMM_REGS, dl,
2751 MVT::Other, SaveXMMOps));
2754 if (!MemOps.empty())
2755 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, MemOps);
2758 if (isVarArg && MFI->hasMustTailInVarArgFunc()) {
2759 // Find the largest legal vector type.
2760 MVT VecVT = MVT::Other;
2761 // FIXME: Only some x86_32 calling conventions support AVX512.
2762 if (Subtarget->hasAVX512() &&
2763 (Is64Bit || (CallConv == CallingConv::X86_VectorCall ||
2764 CallConv == CallingConv::Intel_OCL_BI)))
2765 VecVT = MVT::v16f32;
2766 else if (Subtarget->hasAVX())
2768 else if (Subtarget->hasSSE2())
2771 // We forward some GPRs and some vector types.
2772 SmallVector<MVT, 2> RegParmTypes;
2773 MVT IntVT = Is64Bit ? MVT::i64 : MVT::i32;
2774 RegParmTypes.push_back(IntVT);
2775 if (VecVT != MVT::Other)
2776 RegParmTypes.push_back(VecVT);
2778 // Compute the set of forwarded registers. The rest are scratch.
2779 SmallVectorImpl<ForwardedRegister> &Forwards =
2780 FuncInfo->getForwardedMustTailRegParms();
2781 CCInfo.analyzeMustTailForwardedRegisters(Forwards, RegParmTypes, CC_X86);
2783 // Conservatively forward AL on x86_64, since it might be used for varargs.
2784 if (Is64Bit && !CCInfo.isAllocated(X86::AL)) {
2785 unsigned ALVReg = MF.addLiveIn(X86::AL, &X86::GR8RegClass);
2786 Forwards.push_back(ForwardedRegister(ALVReg, X86::AL, MVT::i8));
2789 // Copy all forwards from physical to virtual registers.
2790 for (ForwardedRegister &F : Forwards) {
2791 // FIXME: Can we use a less constrained schedule?
2792 SDValue RegVal = DAG.getCopyFromReg(Chain, dl, F.VReg, F.VT);
2793 F.VReg = MF.getRegInfo().createVirtualRegister(getRegClassFor(F.VT));
2794 Chain = DAG.getCopyToReg(Chain, dl, F.VReg, RegVal);
2798 // Some CCs need callee pop.
2799 if (X86::isCalleePop(CallConv, Is64Bit, isVarArg,
2800 MF.getTarget().Options.GuaranteedTailCallOpt)) {
2801 FuncInfo->setBytesToPopOnReturn(StackSize); // Callee pops everything.
2803 FuncInfo->setBytesToPopOnReturn(0); // Callee pops nothing.
2804 // If this is an sret function, the return should pop the hidden pointer.
2805 if (!Is64Bit && !IsTailCallConvention(CallConv) &&
2806 !Subtarget->getTargetTriple().isOSMSVCRT() &&
2807 argsAreStructReturn(Ins) == StackStructReturn)
2808 FuncInfo->setBytesToPopOnReturn(4);
2812 // RegSaveFrameIndex is X86-64 only.
2813 FuncInfo->setRegSaveFrameIndex(0xAAAAAAA);
2814 if (CallConv == CallingConv::X86_FastCall ||
2815 CallConv == CallingConv::X86_ThisCall)
2816 // fastcc functions can't have varargs.
2817 FuncInfo->setVarArgsFrameIndex(0xAAAAAAA);
2820 FuncInfo->setArgumentStackSize(StackSize);
2822 if (IsWinEHParent) {
2824 int UnwindHelpFI = MFI->CreateStackObject(8, 8, /*isSS=*/false);
2825 SDValue StackSlot = DAG.getFrameIndex(UnwindHelpFI, MVT::i64);
2826 MMI.getWinEHFuncInfo(MF.getFunction()).UnwindHelpFrameIdx = UnwindHelpFI;
2827 SDValue Neg2 = DAG.getConstant(-2, dl, MVT::i64);
2828 Chain = DAG.getStore(Chain, dl, Neg2, StackSlot,
2829 MachinePointerInfo::getFixedStack(
2830 DAG.getMachineFunction(), UnwindHelpFI),
2831 /*isVolatile=*/true,
2832 /*isNonTemporal=*/false, /*Alignment=*/0);
2834 // Functions using Win32 EH are considered to have opaque SP adjustments
2835 // to force local variables to be addressed from the frame or base
2837 MFI->setHasOpaqueSPAdjustment(true);
2845 X86TargetLowering::LowerMemOpCallTo(SDValue Chain,
2846 SDValue StackPtr, SDValue Arg,
2847 SDLoc dl, SelectionDAG &DAG,
2848 const CCValAssign &VA,
2849 ISD::ArgFlagsTy Flags) const {
2850 unsigned LocMemOffset = VA.getLocMemOffset();
2851 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset, dl);
2852 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(DAG.getDataLayout()),
2854 if (Flags.isByVal())
2855 return CreateCopyOfByValArgument(Arg, PtrOff, Chain, Flags, DAG, dl);
2857 return DAG.getStore(
2858 Chain, dl, Arg, PtrOff,
2859 MachinePointerInfo::getStack(DAG.getMachineFunction(), LocMemOffset),
2863 /// Emit a load of return address if tail call
2864 /// optimization is performed and it is required.
2866 X86TargetLowering::EmitTailCallLoadRetAddr(SelectionDAG &DAG,
2867 SDValue &OutRetAddr, SDValue Chain,
2868 bool IsTailCall, bool Is64Bit,
2869 int FPDiff, SDLoc dl) const {
2870 // Adjust the Return address stack slot.
2871 EVT VT = getPointerTy(DAG.getDataLayout());
2872 OutRetAddr = getReturnAddressFrameIndex(DAG);
2874 // Load the "old" Return address.
2875 OutRetAddr = DAG.getLoad(VT, dl, Chain, OutRetAddr, MachinePointerInfo(),
2876 false, false, false, 0);
2877 return SDValue(OutRetAddr.getNode(), 1);
2880 /// Emit a store of the return address if tail call
2881 /// optimization is performed and it is required (FPDiff!=0).
2882 static SDValue EmitTailCallStoreRetAddr(SelectionDAG &DAG, MachineFunction &MF,
2883 SDValue Chain, SDValue RetAddrFrIdx,
2884 EVT PtrVT, unsigned SlotSize,
2885 int FPDiff, SDLoc dl) {
2886 // Store the return address to the appropriate stack slot.
2887 if (!FPDiff) return Chain;
2888 // Calculate the new stack slot for the return address.
2889 int NewReturnAddrFI =
2890 MF.getFrameInfo()->CreateFixedObject(SlotSize, (int64_t)FPDiff - SlotSize,
2892 SDValue NewRetAddrFrIdx = DAG.getFrameIndex(NewReturnAddrFI, PtrVT);
2893 Chain = DAG.getStore(Chain, dl, RetAddrFrIdx, NewRetAddrFrIdx,
2894 MachinePointerInfo::getFixedStack(
2895 DAG.getMachineFunction(), NewReturnAddrFI),
2900 /// Returns a vector_shuffle mask for an movs{s|d}, movd
2901 /// operation of specified width.
2902 static SDValue getMOVL(SelectionDAG &DAG, SDLoc dl, EVT VT, SDValue V1,
2904 unsigned NumElems = VT.getVectorNumElements();
2905 SmallVector<int, 8> Mask;
2906 Mask.push_back(NumElems);
2907 for (unsigned i = 1; i != NumElems; ++i)
2909 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
2913 X86TargetLowering::LowerCall(TargetLowering::CallLoweringInfo &CLI,
2914 SmallVectorImpl<SDValue> &InVals) const {
2915 SelectionDAG &DAG = CLI.DAG;
2917 SmallVectorImpl<ISD::OutputArg> &Outs = CLI.Outs;
2918 SmallVectorImpl<SDValue> &OutVals = CLI.OutVals;
2919 SmallVectorImpl<ISD::InputArg> &Ins = CLI.Ins;
2920 SDValue Chain = CLI.Chain;
2921 SDValue Callee = CLI.Callee;
2922 CallingConv::ID CallConv = CLI.CallConv;
2923 bool &isTailCall = CLI.IsTailCall;
2924 bool isVarArg = CLI.IsVarArg;
2926 MachineFunction &MF = DAG.getMachineFunction();
2927 bool Is64Bit = Subtarget->is64Bit();
2928 bool IsWin64 = Subtarget->isCallingConvWin64(CallConv);
2929 StructReturnType SR = callIsStructReturn(Outs);
2930 bool IsSibcall = false;
2931 X86MachineFunctionInfo *X86Info = MF.getInfo<X86MachineFunctionInfo>();
2932 auto Attr = MF.getFunction()->getFnAttribute("disable-tail-calls");
2934 if (Attr.getValueAsString() == "true")
2937 if (Subtarget->isPICStyleGOT() &&
2938 !MF.getTarget().Options.GuaranteedTailCallOpt) {
2939 // If we are using a GOT, disable tail calls to external symbols with
2940 // default visibility. Tail calling such a symbol requires using a GOT
2941 // relocation, which forces early binding of the symbol. This breaks code
2942 // that require lazy function symbol resolution. Using musttail or
2943 // GuaranteedTailCallOpt will override this.
2944 GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee);
2945 if (!G || (!G->getGlobal()->hasLocalLinkage() &&
2946 G->getGlobal()->hasDefaultVisibility()))
2950 bool IsMustTail = CLI.CS && CLI.CS->isMustTailCall();
2952 // Force this to be a tail call. The verifier rules are enough to ensure
2953 // that we can lower this successfully without moving the return address
2956 } else if (isTailCall) {
2957 // Check if it's really possible to do a tail call.
2958 isTailCall = IsEligibleForTailCallOptimization(Callee, CallConv,
2959 isVarArg, SR != NotStructReturn,
2960 MF.getFunction()->hasStructRetAttr(), CLI.RetTy,
2961 Outs, OutVals, Ins, DAG);
2963 // Sibcalls are automatically detected tailcalls which do not require
2965 if (!MF.getTarget().Options.GuaranteedTailCallOpt && isTailCall)
2972 assert(!(isVarArg && IsTailCallConvention(CallConv)) &&
2973 "Var args not supported with calling convention fastcc, ghc or hipe");
2975 // Analyze operands of the call, assigning locations to each operand.
2976 SmallVector<CCValAssign, 16> ArgLocs;
2977 CCState CCInfo(CallConv, isVarArg, MF, ArgLocs, *DAG.getContext());
2979 // Allocate shadow area for Win64
2981 CCInfo.AllocateStack(32, 8);
2983 CCInfo.AnalyzeCallOperands(Outs, CC_X86);
2985 // Get a count of how many bytes are to be pushed on the stack.
2986 unsigned NumBytes = CCInfo.getNextStackOffset();
2988 // This is a sibcall. The memory operands are available in caller's
2989 // own caller's stack.
2991 else if (MF.getTarget().Options.GuaranteedTailCallOpt &&
2992 IsTailCallConvention(CallConv))
2993 NumBytes = GetAlignedArgumentStackSize(NumBytes, DAG);
2996 if (isTailCall && !IsSibcall && !IsMustTail) {
2997 // Lower arguments at fp - stackoffset + fpdiff.
2998 unsigned NumBytesCallerPushed = X86Info->getBytesToPopOnReturn();
3000 FPDiff = NumBytesCallerPushed - NumBytes;
3002 // Set the delta of movement of the returnaddr stackslot.
3003 // But only set if delta is greater than previous delta.
3004 if (FPDiff < X86Info->getTCReturnAddrDelta())
3005 X86Info->setTCReturnAddrDelta(FPDiff);
3008 unsigned NumBytesToPush = NumBytes;
3009 unsigned NumBytesToPop = NumBytes;
3011 // If we have an inalloca argument, all stack space has already been allocated
3012 // for us and be right at the top of the stack. We don't support multiple
3013 // arguments passed in memory when using inalloca.
3014 if (!Outs.empty() && Outs.back().Flags.isInAlloca()) {
3016 if (!ArgLocs.back().isMemLoc())
3017 report_fatal_error("cannot use inalloca attribute on a register "
3019 if (ArgLocs.back().getLocMemOffset() != 0)
3020 report_fatal_error("any parameter with the inalloca attribute must be "
3021 "the only memory argument");
3025 Chain = DAG.getCALLSEQ_START(
3026 Chain, DAG.getIntPtrConstant(NumBytesToPush, dl, true), dl);
3028 SDValue RetAddrFrIdx;
3029 // Load return address for tail calls.
3030 if (isTailCall && FPDiff)
3031 Chain = EmitTailCallLoadRetAddr(DAG, RetAddrFrIdx, Chain, isTailCall,
3032 Is64Bit, FPDiff, dl);
3034 SmallVector<std::pair<unsigned, SDValue>, 8> RegsToPass;
3035 SmallVector<SDValue, 8> MemOpChains;
3038 // Walk the register/memloc assignments, inserting copies/loads. In the case
3039 // of tail call optimization arguments are handle later.
3040 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
3041 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
3042 // Skip inalloca arguments, they have already been written.
3043 ISD::ArgFlagsTy Flags = Outs[i].Flags;
3044 if (Flags.isInAlloca())
3047 CCValAssign &VA = ArgLocs[i];
3048 EVT RegVT = VA.getLocVT();
3049 SDValue Arg = OutVals[i];
3050 bool isByVal = Flags.isByVal();
3052 // Promote the value if needed.
3053 switch (VA.getLocInfo()) {
3054 default: llvm_unreachable("Unknown loc info!");
3055 case CCValAssign::Full: break;
3056 case CCValAssign::SExt:
3057 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, RegVT, Arg);
3059 case CCValAssign::ZExt:
3060 Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, RegVT, Arg);
3062 case CCValAssign::AExt:
3063 if (Arg.getValueType().isVector() &&
3064 Arg.getValueType().getScalarType() == MVT::i1)
3065 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, RegVT, Arg);
3066 else if (RegVT.is128BitVector()) {
3067 // Special case: passing MMX values in XMM registers.
3068 Arg = DAG.getBitcast(MVT::i64, Arg);
3069 Arg = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i64, Arg);
3070 Arg = getMOVL(DAG, dl, MVT::v2i64, DAG.getUNDEF(MVT::v2i64), Arg);
3072 Arg = DAG.getNode(ISD::ANY_EXTEND, dl, RegVT, Arg);
3074 case CCValAssign::BCvt:
3075 Arg = DAG.getBitcast(RegVT, Arg);
3077 case CCValAssign::Indirect: {
3078 // Store the argument.
3079 SDValue SpillSlot = DAG.CreateStackTemporary(VA.getValVT());
3080 int FI = cast<FrameIndexSDNode>(SpillSlot)->getIndex();
3081 Chain = DAG.getStore(
3082 Chain, dl, Arg, SpillSlot,
3083 MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), FI),
3090 if (VA.isRegLoc()) {
3091 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
3092 if (isVarArg && IsWin64) {
3093 // Win64 ABI requires argument XMM reg to be copied to the corresponding
3094 // shadow reg if callee is a varargs function.
3095 unsigned ShadowReg = 0;
3096 switch (VA.getLocReg()) {
3097 case X86::XMM0: ShadowReg = X86::RCX; break;
3098 case X86::XMM1: ShadowReg = X86::RDX; break;
3099 case X86::XMM2: ShadowReg = X86::R8; break;
3100 case X86::XMM3: ShadowReg = X86::R9; break;
3103 RegsToPass.push_back(std::make_pair(ShadowReg, Arg));
3105 } else if (!IsSibcall && (!isTailCall || isByVal)) {
3106 assert(VA.isMemLoc());
3107 if (!StackPtr.getNode())
3108 StackPtr = DAG.getCopyFromReg(Chain, dl, RegInfo->getStackRegister(),
3109 getPointerTy(DAG.getDataLayout()));
3110 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Arg,
3111 dl, DAG, VA, Flags));
3115 if (!MemOpChains.empty())
3116 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, MemOpChains);
3118 if (Subtarget->isPICStyleGOT()) {
3119 // ELF / PIC requires GOT in the EBX register before function calls via PLT
3122 RegsToPass.push_back(std::make_pair(
3123 unsigned(X86::EBX), DAG.getNode(X86ISD::GlobalBaseReg, SDLoc(),
3124 getPointerTy(DAG.getDataLayout()))));
3126 // If we are tail calling and generating PIC/GOT style code load the
3127 // address of the callee into ECX. The value in ecx is used as target of
3128 // the tail jump. This is done to circumvent the ebx/callee-saved problem
3129 // for tail calls on PIC/GOT architectures. Normally we would just put the
3130 // address of GOT into ebx and then call target@PLT. But for tail calls
3131 // ebx would be restored (since ebx is callee saved) before jumping to the
3134 // Note: The actual moving to ECX is done further down.
3135 GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee);
3136 if (G && !G->getGlobal()->hasLocalLinkage() &&
3137 G->getGlobal()->hasDefaultVisibility())
3138 Callee = LowerGlobalAddress(Callee, DAG);
3139 else if (isa<ExternalSymbolSDNode>(Callee))
3140 Callee = LowerExternalSymbol(Callee, DAG);
3144 if (Is64Bit && isVarArg && !IsWin64 && !IsMustTail) {
3145 // From AMD64 ABI document:
3146 // For calls that may call functions that use varargs or stdargs
3147 // (prototype-less calls or calls to functions containing ellipsis (...) in
3148 // the declaration) %al is used as hidden argument to specify the number
3149 // of SSE registers used. The contents of %al do not need to match exactly
3150 // the number of registers, but must be an ubound on the number of SSE
3151 // registers used and is in the range 0 - 8 inclusive.
3153 // Count the number of XMM registers allocated.
3154 static const MCPhysReg XMMArgRegs[] = {
3155 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
3156 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
3158 unsigned NumXMMRegs = CCInfo.getFirstUnallocated(XMMArgRegs);
3159 assert((Subtarget->hasSSE1() || !NumXMMRegs)
3160 && "SSE registers cannot be used when SSE is disabled");
3162 RegsToPass.push_back(std::make_pair(unsigned(X86::AL),
3163 DAG.getConstant(NumXMMRegs, dl,
3167 if (isVarArg && IsMustTail) {
3168 const auto &Forwards = X86Info->getForwardedMustTailRegParms();
3169 for (const auto &F : Forwards) {
3170 SDValue Val = DAG.getCopyFromReg(Chain, dl, F.VReg, F.VT);
3171 RegsToPass.push_back(std::make_pair(unsigned(F.PReg), Val));
3175 // For tail calls lower the arguments to the 'real' stack slots. Sibcalls
3176 // don't need this because the eligibility check rejects calls that require
3177 // shuffling arguments passed in memory.
3178 if (!IsSibcall && isTailCall) {
3179 // Force all the incoming stack arguments to be loaded from the stack
3180 // before any new outgoing arguments are stored to the stack, because the
3181 // outgoing stack slots may alias the incoming argument stack slots, and
3182 // the alias isn't otherwise explicit. This is slightly more conservative
3183 // than necessary, because it means that each store effectively depends
3184 // on every argument instead of just those arguments it would clobber.
3185 SDValue ArgChain = DAG.getStackArgumentTokenFactor(Chain);
3187 SmallVector<SDValue, 8> MemOpChains2;
3190 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
3191 CCValAssign &VA = ArgLocs[i];
3194 assert(VA.isMemLoc());
3195 SDValue Arg = OutVals[i];
3196 ISD::ArgFlagsTy Flags = Outs[i].Flags;
3197 // Skip inalloca arguments. They don't require any work.
3198 if (Flags.isInAlloca())
3200 // Create frame index.
3201 int32_t Offset = VA.getLocMemOffset()+FPDiff;
3202 uint32_t OpSize = (VA.getLocVT().getSizeInBits()+7)/8;
3203 FI = MF.getFrameInfo()->CreateFixedObject(OpSize, Offset, true);
3204 FIN = DAG.getFrameIndex(FI, getPointerTy(DAG.getDataLayout()));
3206 if (Flags.isByVal()) {
3207 // Copy relative to framepointer.
3208 SDValue Source = DAG.getIntPtrConstant(VA.getLocMemOffset(), dl);
3209 if (!StackPtr.getNode())
3210 StackPtr = DAG.getCopyFromReg(Chain, dl, RegInfo->getStackRegister(),
3211 getPointerTy(DAG.getDataLayout()));
3212 Source = DAG.getNode(ISD::ADD, dl, getPointerTy(DAG.getDataLayout()),
3215 MemOpChains2.push_back(CreateCopyOfByValArgument(Source, FIN,
3219 // Store relative to framepointer.
3220 MemOpChains2.push_back(DAG.getStore(
3221 ArgChain, dl, Arg, FIN,
3222 MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), FI),
3227 if (!MemOpChains2.empty())
3228 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, MemOpChains2);
3230 // Store the return address to the appropriate stack slot.
3231 Chain = EmitTailCallStoreRetAddr(DAG, MF, Chain, RetAddrFrIdx,
3232 getPointerTy(DAG.getDataLayout()),
3233 RegInfo->getSlotSize(), FPDiff, dl);
3236 // Build a sequence of copy-to-reg nodes chained together with token chain
3237 // and flag operands which copy the outgoing args into registers.
3239 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
3240 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
3241 RegsToPass[i].second, InFlag);
3242 InFlag = Chain.getValue(1);
3245 if (DAG.getTarget().getCodeModel() == CodeModel::Large) {
3246 assert(Is64Bit && "Large code model is only legal in 64-bit mode.");
3247 // In the 64-bit large code model, we have to make all calls
3248 // through a register, since the call instruction's 32-bit
3249 // pc-relative offset may not be large enough to hold the whole
3251 } else if (Callee->getOpcode() == ISD::GlobalAddress) {
3252 // If the callee is a GlobalAddress node (quite common, every direct call
3253 // is) turn it into a TargetGlobalAddress node so that legalize doesn't hack
3255 GlobalAddressSDNode* G = cast<GlobalAddressSDNode>(Callee);
3257 // We should use extra load for direct calls to dllimported functions in
3259 const GlobalValue *GV = G->getGlobal();
3260 if (!GV->hasDLLImportStorageClass()) {
3261 unsigned char OpFlags = 0;
3262 bool ExtraLoad = false;
3263 unsigned WrapperKind = ISD::DELETED_NODE;
3265 // On ELF targets, in both X86-64 and X86-32 mode, direct calls to
3266 // external symbols most go through the PLT in PIC mode. If the symbol
3267 // has hidden or protected visibility, or if it is static or local, then
3268 // we don't need to use the PLT - we can directly call it.
3269 if (Subtarget->isTargetELF() &&
3270 DAG.getTarget().getRelocationModel() == Reloc::PIC_ &&
3271 GV->hasDefaultVisibility() && !GV->hasLocalLinkage()) {
3272 OpFlags = X86II::MO_PLT;
3273 } else if (Subtarget->isPICStyleStubAny() &&
3274 !GV->isStrongDefinitionForLinker() &&
3275 (!Subtarget->getTargetTriple().isMacOSX() ||
3276 Subtarget->getTargetTriple().isMacOSXVersionLT(10, 5))) {
3277 // PC-relative references to external symbols should go through $stub,
3278 // unless we're building with the leopard linker or later, which
3279 // automatically synthesizes these stubs.
3280 OpFlags = X86II::MO_DARWIN_STUB;
3281 } else if (Subtarget->isPICStyleRIPRel() && isa<Function>(GV) &&
3282 cast<Function>(GV)->hasFnAttribute(Attribute::NonLazyBind)) {
3283 // If the function is marked as non-lazy, generate an indirect call
3284 // which loads from the GOT directly. This avoids runtime overhead
3285 // at the cost of eager binding (and one extra byte of encoding).
3286 OpFlags = X86II::MO_GOTPCREL;
3287 WrapperKind = X86ISD::WrapperRIP;
3291 Callee = DAG.getTargetGlobalAddress(
3292 GV, dl, getPointerTy(DAG.getDataLayout()), G->getOffset(), OpFlags);
3294 // Add a wrapper if needed.
3295 if (WrapperKind != ISD::DELETED_NODE)
3296 Callee = DAG.getNode(X86ISD::WrapperRIP, dl,
3297 getPointerTy(DAG.getDataLayout()), Callee);
3298 // Add extra indirection if needed.
3300 Callee = DAG.getLoad(
3301 getPointerTy(DAG.getDataLayout()), dl, DAG.getEntryNode(), Callee,
3302 MachinePointerInfo::getGOT(DAG.getMachineFunction()), false, false,
3305 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
3306 unsigned char OpFlags = 0;
3308 // On ELF targets, in either X86-64 or X86-32 mode, direct calls to
3309 // external symbols should go through the PLT.
3310 if (Subtarget->isTargetELF() &&
3311 DAG.getTarget().getRelocationModel() == Reloc::PIC_) {
3312 OpFlags = X86II::MO_PLT;
3313 } else if (Subtarget->isPICStyleStubAny() &&
3314 (!Subtarget->getTargetTriple().isMacOSX() ||
3315 Subtarget->getTargetTriple().isMacOSXVersionLT(10, 5))) {
3316 // PC-relative references to external symbols should go through $stub,
3317 // unless we're building with the leopard linker or later, which
3318 // automatically synthesizes these stubs.
3319 OpFlags = X86II::MO_DARWIN_STUB;
3322 Callee = DAG.getTargetExternalSymbol(
3323 S->getSymbol(), getPointerTy(DAG.getDataLayout()), OpFlags);
3324 } else if (Subtarget->isTarget64BitILP32() &&
3325 Callee->getValueType(0) == MVT::i32) {
3326 // Zero-extend the 32-bit Callee address into a 64-bit according to x32 ABI
3327 Callee = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i64, Callee);
3330 // Returns a chain & a flag for retval copy to use.
3331 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
3332 SmallVector<SDValue, 8> Ops;
3334 if (!IsSibcall && isTailCall) {
3335 Chain = DAG.getCALLSEQ_END(Chain,
3336 DAG.getIntPtrConstant(NumBytesToPop, dl, true),
3337 DAG.getIntPtrConstant(0, dl, true), InFlag, dl);
3338 InFlag = Chain.getValue(1);
3341 Ops.push_back(Chain);
3342 Ops.push_back(Callee);
3345 Ops.push_back(DAG.getConstant(FPDiff, dl, MVT::i32));
3347 // Add argument registers to the end of the list so that they are known live
3349 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
3350 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
3351 RegsToPass[i].second.getValueType()));
3353 // Add a register mask operand representing the call-preserved registers.
3354 const uint32_t *Mask = RegInfo->getCallPreservedMask(MF, CallConv);
3355 assert(Mask && "Missing call preserved mask for calling convention");
3357 // If this is an invoke in a 32-bit function using an MSVC personality, assume
3358 // the function clobbers all registers. If an exception is thrown, the runtime
3359 // will not restore CSRs.
3360 // FIXME: Model this more precisely so that we can register allocate across
3361 // the normal edge and spill and fill across the exceptional edge.
3362 if (!Is64Bit && CLI.CS && CLI.CS->isInvoke()) {
3363 const Function *CallerFn = MF.getFunction();
3364 EHPersonality Pers =
3365 CallerFn->hasPersonalityFn()
3366 ? classifyEHPersonality(CallerFn->getPersonalityFn())
3367 : EHPersonality::Unknown;
3368 if (isMSVCEHPersonality(Pers))
3369 Mask = RegInfo->getNoPreservedMask();
3372 Ops.push_back(DAG.getRegisterMask(Mask));
3374 if (InFlag.getNode())
3375 Ops.push_back(InFlag);
3379 //// If this is the first return lowered for this function, add the regs
3380 //// to the liveout set for the function.
3381 // This isn't right, although it's probably harmless on x86; liveouts
3382 // should be computed from returns not tail calls. Consider a void
3383 // function making a tail call to a function returning int.
3384 MF.getFrameInfo()->setHasTailCall();
3385 return DAG.getNode(X86ISD::TC_RETURN, dl, NodeTys, Ops);
3388 Chain = DAG.getNode(X86ISD::CALL, dl, NodeTys, Ops);
3389 InFlag = Chain.getValue(1);
3391 // Create the CALLSEQ_END node.
3392 unsigned NumBytesForCalleeToPop;
3393 if (X86::isCalleePop(CallConv, Is64Bit, isVarArg,
3394 DAG.getTarget().Options.GuaranteedTailCallOpt))
3395 NumBytesForCalleeToPop = NumBytes; // Callee pops everything
3396 else if (!Is64Bit && !IsTailCallConvention(CallConv) &&
3397 !Subtarget->getTargetTriple().isOSMSVCRT() &&
3398 SR == StackStructReturn)
3399 // If this is a call to a struct-return function, the callee
3400 // pops the hidden struct pointer, so we have to push it back.
3401 // This is common for Darwin/X86, Linux & Mingw32 targets.
3402 // For MSVC Win32 targets, the caller pops the hidden struct pointer.
3403 NumBytesForCalleeToPop = 4;
3405 NumBytesForCalleeToPop = 0; // Callee pops nothing.
3407 // Returns a flag for retval copy to use.
3409 Chain = DAG.getCALLSEQ_END(Chain,
3410 DAG.getIntPtrConstant(NumBytesToPop, dl, true),
3411 DAG.getIntPtrConstant(NumBytesForCalleeToPop, dl,
3414 InFlag = Chain.getValue(1);
3417 // Handle result values, copying them out of physregs into vregs that we
3419 return LowerCallResult(Chain, InFlag, CallConv, isVarArg,
3420 Ins, dl, DAG, InVals);
3423 //===----------------------------------------------------------------------===//
3424 // Fast Calling Convention (tail call) implementation
3425 //===----------------------------------------------------------------------===//
3427 // Like std call, callee cleans arguments, convention except that ECX is
3428 // reserved for storing the tail called function address. Only 2 registers are
3429 // free for argument passing (inreg). Tail call optimization is performed
3431 // * tailcallopt is enabled
3432 // * caller/callee are fastcc
3433 // On X86_64 architecture with GOT-style position independent code only local
3434 // (within module) calls are supported at the moment.
3435 // To keep the stack aligned according to platform abi the function
3436 // GetAlignedArgumentStackSize ensures that argument delta is always multiples
3437 // of stack alignment. (Dynamic linkers need this - darwin's dyld for example)
3438 // If a tail called function callee has more arguments than the caller the
3439 // caller needs to make sure that there is room to move the RETADDR to. This is
3440 // achieved by reserving an area the size of the argument delta right after the
3441 // original RETADDR, but before the saved framepointer or the spilled registers
3442 // e.g. caller(arg1, arg2) calls callee(arg1, arg2,arg3,arg4)
3454 /// Make the stack size align e.g 16n + 12 aligned for a 16-byte align
3457 X86TargetLowering::GetAlignedArgumentStackSize(unsigned StackSize,
3458 SelectionDAG& DAG) const {
3459 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
3460 const TargetFrameLowering &TFI = *Subtarget->getFrameLowering();
3461 unsigned StackAlignment = TFI.getStackAlignment();
3462 uint64_t AlignMask = StackAlignment - 1;
3463 int64_t Offset = StackSize;
3464 unsigned SlotSize = RegInfo->getSlotSize();
3465 if ( (Offset & AlignMask) <= (StackAlignment - SlotSize) ) {
3466 // Number smaller than 12 so just add the difference.
3467 Offset += ((StackAlignment - SlotSize) - (Offset & AlignMask));
3469 // Mask out lower bits, add stackalignment once plus the 12 bytes.
3470 Offset = ((~AlignMask) & Offset) + StackAlignment +
3471 (StackAlignment-SlotSize);
3476 /// Return true if the given stack call argument is already available in the
3477 /// same position (relatively) of the caller's incoming argument stack.
3479 bool MatchingStackOffset(SDValue Arg, unsigned Offset, ISD::ArgFlagsTy Flags,
3480 MachineFrameInfo *MFI, const MachineRegisterInfo *MRI,
3481 const X86InstrInfo *TII) {
3482 unsigned Bytes = Arg.getValueType().getSizeInBits() / 8;
3484 if (Arg.getOpcode() == ISD::CopyFromReg) {
3485 unsigned VR = cast<RegisterSDNode>(Arg.getOperand(1))->getReg();
3486 if (!TargetRegisterInfo::isVirtualRegister(VR))
3488 MachineInstr *Def = MRI->getVRegDef(VR);
3491 if (!Flags.isByVal()) {
3492 if (!TII->isLoadFromStackSlot(Def, FI))
3495 unsigned Opcode = Def->getOpcode();
3496 if ((Opcode == X86::LEA32r || Opcode == X86::LEA64r ||
3497 Opcode == X86::LEA64_32r) &&
3498 Def->getOperand(1).isFI()) {
3499 FI = Def->getOperand(1).getIndex();
3500 Bytes = Flags.getByValSize();
3504 } else if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Arg)) {
3505 if (Flags.isByVal())
3506 // ByVal argument is passed in as a pointer but it's now being
3507 // dereferenced. e.g.
3508 // define @foo(%struct.X* %A) {
3509 // tail call @bar(%struct.X* byval %A)
3512 SDValue Ptr = Ld->getBasePtr();
3513 FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr);
3516 FI = FINode->getIndex();
3517 } else if (Arg.getOpcode() == ISD::FrameIndex && Flags.isByVal()) {
3518 FrameIndexSDNode *FINode = cast<FrameIndexSDNode>(Arg);
3519 FI = FINode->getIndex();
3520 Bytes = Flags.getByValSize();
3524 assert(FI != INT_MAX);
3525 if (!MFI->isFixedObjectIndex(FI))
3527 return Offset == MFI->getObjectOffset(FI) && Bytes == MFI->getObjectSize(FI);
3530 /// Check whether the call is eligible for tail call optimization. Targets
3531 /// that want to do tail call optimization should implement this function.
3533 X86TargetLowering::IsEligibleForTailCallOptimization(SDValue Callee,
3534 CallingConv::ID CalleeCC,
3536 bool isCalleeStructRet,
3537 bool isCallerStructRet,
3539 const SmallVectorImpl<ISD::OutputArg> &Outs,
3540 const SmallVectorImpl<SDValue> &OutVals,
3541 const SmallVectorImpl<ISD::InputArg> &Ins,
3542 SelectionDAG &DAG) const {
3543 if (!IsTailCallConvention(CalleeCC) && !IsCCallConvention(CalleeCC))
3546 // If -tailcallopt is specified, make fastcc functions tail-callable.
3547 const MachineFunction &MF = DAG.getMachineFunction();
3548 const Function *CallerF = MF.getFunction();
3550 // If the function return type is x86_fp80 and the callee return type is not,
3551 // then the FP_EXTEND of the call result is not a nop. It's not safe to
3552 // perform a tailcall optimization here.
3553 if (CallerF->getReturnType()->isX86_FP80Ty() && !RetTy->isX86_FP80Ty())
3556 CallingConv::ID CallerCC = CallerF->getCallingConv();
3557 bool CCMatch = CallerCC == CalleeCC;
3558 bool IsCalleeWin64 = Subtarget->isCallingConvWin64(CalleeCC);
3559 bool IsCallerWin64 = Subtarget->isCallingConvWin64(CallerCC);
3561 // Win64 functions have extra shadow space for argument homing. Don't do the
3562 // sibcall if the caller and callee have mismatched expectations for this
3564 if (IsCalleeWin64 != IsCallerWin64)
3567 if (DAG.getTarget().Options.GuaranteedTailCallOpt) {
3568 if (IsTailCallConvention(CalleeCC) && CCMatch)
3573 // Look for obvious safe cases to perform tail call optimization that do not
3574 // require ABI changes. This is what gcc calls sibcall.
3576 // Can't do sibcall if stack needs to be dynamically re-aligned. PEI needs to
3577 // emit a special epilogue.
3578 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
3579 if (RegInfo->needsStackRealignment(MF))
3582 // Also avoid sibcall optimization if either caller or callee uses struct
3583 // return semantics.
3584 if (isCalleeStructRet || isCallerStructRet)
3587 // An stdcall/thiscall caller is expected to clean up its arguments; the
3588 // callee isn't going to do that.
3589 // FIXME: this is more restrictive than needed. We could produce a tailcall
3590 // when the stack adjustment matches. For example, with a thiscall that takes
3591 // only one argument.
3592 if (!CCMatch && (CallerCC == CallingConv::X86_StdCall ||
3593 CallerCC == CallingConv::X86_ThisCall))
3596 // Do not sibcall optimize vararg calls unless all arguments are passed via
3598 if (isVarArg && !Outs.empty()) {
3600 // Optimizing for varargs on Win64 is unlikely to be safe without
3601 // additional testing.
3602 if (IsCalleeWin64 || IsCallerWin64)
3605 SmallVector<CCValAssign, 16> ArgLocs;
3606 CCState CCInfo(CalleeCC, isVarArg, DAG.getMachineFunction(), ArgLocs,
3609 CCInfo.AnalyzeCallOperands(Outs, CC_X86);
3610 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i)
3611 if (!ArgLocs[i].isRegLoc())
3615 // If the call result is in ST0 / ST1, it needs to be popped off the x87
3616 // stack. Therefore, if it's not used by the call it is not safe to optimize
3617 // this into a sibcall.
3618 bool Unused = false;
3619 for (unsigned i = 0, e = Ins.size(); i != e; ++i) {
3626 SmallVector<CCValAssign, 16> RVLocs;
3627 CCState CCInfo(CalleeCC, false, DAG.getMachineFunction(), RVLocs,
3629 CCInfo.AnalyzeCallResult(Ins, RetCC_X86);
3630 for (unsigned i = 0, e = RVLocs.size(); i != e; ++i) {
3631 CCValAssign &VA = RVLocs[i];
3632 if (VA.getLocReg() == X86::FP0 || VA.getLocReg() == X86::FP1)
3637 // If the calling conventions do not match, then we'd better make sure the
3638 // results are returned in the same way as what the caller expects.
3640 SmallVector<CCValAssign, 16> RVLocs1;
3641 CCState CCInfo1(CalleeCC, false, DAG.getMachineFunction(), RVLocs1,
3643 CCInfo1.AnalyzeCallResult(Ins, RetCC_X86);
3645 SmallVector<CCValAssign, 16> RVLocs2;
3646 CCState CCInfo2(CallerCC, false, DAG.getMachineFunction(), RVLocs2,
3648 CCInfo2.AnalyzeCallResult(Ins, RetCC_X86);
3650 if (RVLocs1.size() != RVLocs2.size())
3652 for (unsigned i = 0, e = RVLocs1.size(); i != e; ++i) {
3653 if (RVLocs1[i].isRegLoc() != RVLocs2[i].isRegLoc())
3655 if (RVLocs1[i].getLocInfo() != RVLocs2[i].getLocInfo())
3657 if (RVLocs1[i].isRegLoc()) {
3658 if (RVLocs1[i].getLocReg() != RVLocs2[i].getLocReg())
3661 if (RVLocs1[i].getLocMemOffset() != RVLocs2[i].getLocMemOffset())
3667 // If the callee takes no arguments then go on to check the results of the
3669 if (!Outs.empty()) {
3670 // Check if stack adjustment is needed. For now, do not do this if any
3671 // argument is passed on the stack.
3672 SmallVector<CCValAssign, 16> ArgLocs;
3673 CCState CCInfo(CalleeCC, isVarArg, DAG.getMachineFunction(), ArgLocs,
3676 // Allocate shadow area for Win64
3678 CCInfo.AllocateStack(32, 8);
3680 CCInfo.AnalyzeCallOperands(Outs, CC_X86);
3681 if (CCInfo.getNextStackOffset()) {
3682 MachineFunction &MF = DAG.getMachineFunction();
3683 if (MF.getInfo<X86MachineFunctionInfo>()->getBytesToPopOnReturn())
3686 // Check if the arguments are already laid out in the right way as
3687 // the caller's fixed stack objects.
3688 MachineFrameInfo *MFI = MF.getFrameInfo();
3689 const MachineRegisterInfo *MRI = &MF.getRegInfo();
3690 const X86InstrInfo *TII = Subtarget->getInstrInfo();
3691 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
3692 CCValAssign &VA = ArgLocs[i];
3693 SDValue Arg = OutVals[i];
3694 ISD::ArgFlagsTy Flags = Outs[i].Flags;
3695 if (VA.getLocInfo() == CCValAssign::Indirect)
3697 if (!VA.isRegLoc()) {
3698 if (!MatchingStackOffset(Arg, VA.getLocMemOffset(), Flags,
3705 // If the tailcall address may be in a register, then make sure it's
3706 // possible to register allocate for it. In 32-bit, the call address can
3707 // only target EAX, EDX, or ECX since the tail call must be scheduled after
3708 // callee-saved registers are restored. These happen to be the same
3709 // registers used to pass 'inreg' arguments so watch out for those.
3710 if (!Subtarget->is64Bit() &&
3711 ((!isa<GlobalAddressSDNode>(Callee) &&
3712 !isa<ExternalSymbolSDNode>(Callee)) ||
3713 DAG.getTarget().getRelocationModel() == Reloc::PIC_)) {
3714 unsigned NumInRegs = 0;
3715 // In PIC we need an extra register to formulate the address computation
3717 unsigned MaxInRegs =
3718 (DAG.getTarget().getRelocationModel() == Reloc::PIC_) ? 2 : 3;
3720 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
3721 CCValAssign &VA = ArgLocs[i];
3724 unsigned Reg = VA.getLocReg();
3727 case X86::EAX: case X86::EDX: case X86::ECX:
3728 if (++NumInRegs == MaxInRegs)
3740 X86TargetLowering::createFastISel(FunctionLoweringInfo &funcInfo,
3741 const TargetLibraryInfo *libInfo) const {
3742 return X86::createFastISel(funcInfo, libInfo);
3745 //===----------------------------------------------------------------------===//
3746 // Other Lowering Hooks
3747 //===----------------------------------------------------------------------===//
3749 static bool MayFoldLoad(SDValue Op) {
3750 return Op.hasOneUse() && ISD::isNormalLoad(Op.getNode());
3753 static bool MayFoldIntoStore(SDValue Op) {
3754 return Op.hasOneUse() && ISD::isNormalStore(*Op.getNode()->use_begin());
3757 static bool isTargetShuffle(unsigned Opcode) {
3759 default: return false;
3760 case X86ISD::BLENDI:
3761 case X86ISD::PSHUFB:
3762 case X86ISD::PSHUFD:
3763 case X86ISD::PSHUFHW:
3764 case X86ISD::PSHUFLW:
3766 case X86ISD::PALIGNR:
3767 case X86ISD::MOVLHPS:
3768 case X86ISD::MOVLHPD:
3769 case X86ISD::MOVHLPS:
3770 case X86ISD::MOVLPS:
3771 case X86ISD::MOVLPD:
3772 case X86ISD::MOVSHDUP:
3773 case X86ISD::MOVSLDUP:
3774 case X86ISD::MOVDDUP:
3777 case X86ISD::UNPCKL:
3778 case X86ISD::UNPCKH:
3779 case X86ISD::VPERMILPI:
3780 case X86ISD::VPERM2X128:
3781 case X86ISD::VPERMI:
3782 case X86ISD::VPERMV:
3783 case X86ISD::VPERMV3:
3788 static SDValue getTargetShuffleNode(unsigned Opc, SDLoc dl, EVT VT,
3789 SDValue V1, unsigned TargetMask,
3790 SelectionDAG &DAG) {
3792 default: llvm_unreachable("Unknown x86 shuffle node");
3793 case X86ISD::PSHUFD:
3794 case X86ISD::PSHUFHW:
3795 case X86ISD::PSHUFLW:
3796 case X86ISD::VPERMILPI:
3797 case X86ISD::VPERMI:
3798 return DAG.getNode(Opc, dl, VT, V1,
3799 DAG.getConstant(TargetMask, dl, MVT::i8));
3803 static SDValue getTargetShuffleNode(unsigned Opc, SDLoc dl, EVT VT,
3804 SDValue V1, SDValue V2, SelectionDAG &DAG) {
3806 default: llvm_unreachable("Unknown x86 shuffle node");
3807 case X86ISD::MOVLHPS:
3808 case X86ISD::MOVLHPD:
3809 case X86ISD::MOVHLPS:
3810 case X86ISD::MOVLPS:
3811 case X86ISD::MOVLPD:
3814 case X86ISD::UNPCKL:
3815 case X86ISD::UNPCKH:
3816 return DAG.getNode(Opc, dl, VT, V1, V2);
3820 SDValue X86TargetLowering::getReturnAddressFrameIndex(SelectionDAG &DAG) const {
3821 MachineFunction &MF = DAG.getMachineFunction();
3822 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
3823 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
3824 int ReturnAddrIndex = FuncInfo->getRAIndex();
3826 if (ReturnAddrIndex == 0) {
3827 // Set up a frame object for the return address.
3828 unsigned SlotSize = RegInfo->getSlotSize();
3829 ReturnAddrIndex = MF.getFrameInfo()->CreateFixedObject(SlotSize,
3832 FuncInfo->setRAIndex(ReturnAddrIndex);
3835 return DAG.getFrameIndex(ReturnAddrIndex, getPointerTy(DAG.getDataLayout()));
3838 bool X86::isOffsetSuitableForCodeModel(int64_t Offset, CodeModel::Model M,
3839 bool hasSymbolicDisplacement) {
3840 // Offset should fit into 32 bit immediate field.
3841 if (!isInt<32>(Offset))
3844 // If we don't have a symbolic displacement - we don't have any extra
3846 if (!hasSymbolicDisplacement)
3849 // FIXME: Some tweaks might be needed for medium code model.
3850 if (M != CodeModel::Small && M != CodeModel::Kernel)
3853 // For small code model we assume that latest object is 16MB before end of 31
3854 // bits boundary. We may also accept pretty large negative constants knowing
3855 // that all objects are in the positive half of address space.
3856 if (M == CodeModel::Small && Offset < 16*1024*1024)
3859 // For kernel code model we know that all object resist in the negative half
3860 // of 32bits address space. We may not accept negative offsets, since they may
3861 // be just off and we may accept pretty large positive ones.
3862 if (M == CodeModel::Kernel && Offset >= 0)
3868 /// Determines whether the callee is required to pop its own arguments.
3869 /// Callee pop is necessary to support tail calls.
3870 bool X86::isCalleePop(CallingConv::ID CallingConv,
3871 bool is64Bit, bool IsVarArg, bool TailCallOpt) {
3872 switch (CallingConv) {
3875 case CallingConv::X86_StdCall:
3876 case CallingConv::X86_FastCall:
3877 case CallingConv::X86_ThisCall:
3879 case CallingConv::Fast:
3880 case CallingConv::GHC:
3881 case CallingConv::HiPE:
3888 /// \brief Return true if the condition is an unsigned comparison operation.
3889 static bool isX86CCUnsigned(unsigned X86CC) {
3891 default: llvm_unreachable("Invalid integer condition!");
3892 case X86::COND_E: return true;
3893 case X86::COND_G: return false;
3894 case X86::COND_GE: return false;
3895 case X86::COND_L: return false;
3896 case X86::COND_LE: return false;
3897 case X86::COND_NE: return true;
3898 case X86::COND_B: return true;
3899 case X86::COND_A: return true;
3900 case X86::COND_BE: return true;
3901 case X86::COND_AE: return true;
3903 llvm_unreachable("covered switch fell through?!");
3906 /// Do a one-to-one translation of a ISD::CondCode to the X86-specific
3907 /// condition code, returning the condition code and the LHS/RHS of the
3908 /// comparison to make.
3909 static unsigned TranslateX86CC(ISD::CondCode SetCCOpcode, SDLoc DL, bool isFP,
3910 SDValue &LHS, SDValue &RHS, SelectionDAG &DAG) {
3912 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS)) {
3913 if (SetCCOpcode == ISD::SETGT && RHSC->isAllOnesValue()) {
3914 // X > -1 -> X == 0, jump !sign.
3915 RHS = DAG.getConstant(0, DL, RHS.getValueType());
3916 return X86::COND_NS;
3918 if (SetCCOpcode == ISD::SETLT && RHSC->isNullValue()) {
3919 // X < 0 -> X == 0, jump on sign.
3922 if (SetCCOpcode == ISD::SETLT && RHSC->getZExtValue() == 1) {
3924 RHS = DAG.getConstant(0, DL, RHS.getValueType());
3925 return X86::COND_LE;
3929 switch (SetCCOpcode) {
3930 default: llvm_unreachable("Invalid integer condition!");
3931 case ISD::SETEQ: return X86::COND_E;
3932 case ISD::SETGT: return X86::COND_G;
3933 case ISD::SETGE: return X86::COND_GE;
3934 case ISD::SETLT: return X86::COND_L;
3935 case ISD::SETLE: return X86::COND_LE;
3936 case ISD::SETNE: return X86::COND_NE;
3937 case ISD::SETULT: return X86::COND_B;
3938 case ISD::SETUGT: return X86::COND_A;
3939 case ISD::SETULE: return X86::COND_BE;
3940 case ISD::SETUGE: return X86::COND_AE;
3944 // First determine if it is required or is profitable to flip the operands.
3946 // If LHS is a foldable load, but RHS is not, flip the condition.
3947 if (ISD::isNON_EXTLoad(LHS.getNode()) &&
3948 !ISD::isNON_EXTLoad(RHS.getNode())) {
3949 SetCCOpcode = getSetCCSwappedOperands(SetCCOpcode);
3950 std::swap(LHS, RHS);
3953 switch (SetCCOpcode) {
3959 std::swap(LHS, RHS);
3963 // On a floating point condition, the flags are set as follows:
3965 // 0 | 0 | 0 | X > Y
3966 // 0 | 0 | 1 | X < Y
3967 // 1 | 0 | 0 | X == Y
3968 // 1 | 1 | 1 | unordered
3969 switch (SetCCOpcode) {
3970 default: llvm_unreachable("Condcode should be pre-legalized away");
3972 case ISD::SETEQ: return X86::COND_E;
3973 case ISD::SETOLT: // flipped
3975 case ISD::SETGT: return X86::COND_A;
3976 case ISD::SETOLE: // flipped
3978 case ISD::SETGE: return X86::COND_AE;
3979 case ISD::SETUGT: // flipped
3981 case ISD::SETLT: return X86::COND_B;
3982 case ISD::SETUGE: // flipped
3984 case ISD::SETLE: return X86::COND_BE;
3986 case ISD::SETNE: return X86::COND_NE;
3987 case ISD::SETUO: return X86::COND_P;
3988 case ISD::SETO: return X86::COND_NP;
3990 case ISD::SETUNE: return X86::COND_INVALID;
3994 /// Is there a floating point cmov for the specific X86 condition code?
3995 /// Current x86 isa includes the following FP cmov instructions:
3996 /// fcmovb, fcomvbe, fcomve, fcmovu, fcmovae, fcmova, fcmovne, fcmovnu.
3997 static bool hasFPCMov(unsigned X86CC) {
4013 /// Returns true if the target can instruction select the
4014 /// specified FP immediate natively. If false, the legalizer will
4015 /// materialize the FP immediate as a load from a constant pool.
4016 bool X86TargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
4017 for (unsigned i = 0, e = LegalFPImmediates.size(); i != e; ++i) {
4018 if (Imm.bitwiseIsEqual(LegalFPImmediates[i]))
4024 bool X86TargetLowering::shouldReduceLoadWidth(SDNode *Load,
4025 ISD::LoadExtType ExtTy,
4027 // "ELF Handling for Thread-Local Storage" specifies that R_X86_64_GOTTPOFF
4028 // relocation target a movq or addq instruction: don't let the load shrink.
4029 SDValue BasePtr = cast<LoadSDNode>(Load)->getBasePtr();
4030 if (BasePtr.getOpcode() == X86ISD::WrapperRIP)
4031 if (const auto *GA = dyn_cast<GlobalAddressSDNode>(BasePtr.getOperand(0)))
4032 return GA->getTargetFlags() != X86II::MO_GOTTPOFF;
4036 /// \brief Returns true if it is beneficial to convert a load of a constant
4037 /// to just the constant itself.
4038 bool X86TargetLowering::shouldConvertConstantLoadToIntImm(const APInt &Imm,
4040 assert(Ty->isIntegerTy());
4042 unsigned BitSize = Ty->getPrimitiveSizeInBits();
4043 if (BitSize == 0 || BitSize > 64)
4048 bool X86TargetLowering::isExtractSubvectorCheap(EVT ResVT,
4049 unsigned Index) const {
4050 if (!isOperationLegalOrCustom(ISD::EXTRACT_SUBVECTOR, ResVT))
4053 return (Index == 0 || Index == ResVT.getVectorNumElements());
4056 bool X86TargetLowering::isCheapToSpeculateCttz() const {
4057 // Speculate cttz only if we can directly use TZCNT.
4058 return Subtarget->hasBMI();
4061 bool X86TargetLowering::isCheapToSpeculateCtlz() const {
4062 // Speculate ctlz only if we can directly use LZCNT.
4063 return Subtarget->hasLZCNT();
4066 /// Return true if every element in Mask, beginning
4067 /// from position Pos and ending in Pos+Size is undef.
4068 static bool isUndefInRange(ArrayRef<int> Mask, unsigned Pos, unsigned Size) {
4069 for (unsigned i = Pos, e = Pos + Size; i != e; ++i)
4075 /// Return true if Val is undef or if its value falls within the
4076 /// specified range (L, H].
4077 static bool isUndefOrInRange(int Val, int Low, int Hi) {
4078 return (Val < 0) || (Val >= Low && Val < Hi);
4081 /// Val is either less than zero (undef) or equal to the specified value.
4082 static bool isUndefOrEqual(int Val, int CmpVal) {
4083 return (Val < 0 || Val == CmpVal);
4086 /// Return true if every element in Mask, beginning
4087 /// from position Pos and ending in Pos+Size, falls within the specified
4088 /// sequential range (Low, Low+Size]. or is undef.
4089 static bool isSequentialOrUndefInRange(ArrayRef<int> Mask,
4090 unsigned Pos, unsigned Size, int Low) {
4091 for (unsigned i = Pos, e = Pos+Size; i != e; ++i, ++Low)
4092 if (!isUndefOrEqual(Mask[i], Low))
4097 /// Return true if the specified EXTRACT_SUBVECTOR operand specifies a vector
4098 /// extract that is suitable for instruction that extract 128 or 256 bit vectors
4099 static bool isVEXTRACTIndex(SDNode *N, unsigned vecWidth) {
4100 assert((vecWidth == 128 || vecWidth == 256) && "Unexpected vector width");
4101 if (!isa<ConstantSDNode>(N->getOperand(1).getNode()))
4104 // The index should be aligned on a vecWidth-bit boundary.
4106 cast<ConstantSDNode>(N->getOperand(1).getNode())->getZExtValue();
4108 MVT VT = N->getSimpleValueType(0);
4109 unsigned ElSize = VT.getVectorElementType().getSizeInBits();
4110 bool Result = (Index * ElSize) % vecWidth == 0;
4115 /// Return true if the specified INSERT_SUBVECTOR
4116 /// operand specifies a subvector insert that is suitable for input to
4117 /// insertion of 128 or 256-bit subvectors
4118 static bool isVINSERTIndex(SDNode *N, unsigned vecWidth) {
4119 assert((vecWidth == 128 || vecWidth == 256) && "Unexpected vector width");
4120 if (!isa<ConstantSDNode>(N->getOperand(2).getNode()))
4122 // The index should be aligned on a vecWidth-bit boundary.
4124 cast<ConstantSDNode>(N->getOperand(2).getNode())->getZExtValue();
4126 MVT VT = N->getSimpleValueType(0);
4127 unsigned ElSize = VT.getVectorElementType().getSizeInBits();
4128 bool Result = (Index * ElSize) % vecWidth == 0;
4133 bool X86::isVINSERT128Index(SDNode *N) {
4134 return isVINSERTIndex(N, 128);
4137 bool X86::isVINSERT256Index(SDNode *N) {
4138 return isVINSERTIndex(N, 256);
4141 bool X86::isVEXTRACT128Index(SDNode *N) {
4142 return isVEXTRACTIndex(N, 128);
4145 bool X86::isVEXTRACT256Index(SDNode *N) {
4146 return isVEXTRACTIndex(N, 256);
4149 static unsigned getExtractVEXTRACTImmediate(SDNode *N, unsigned vecWidth) {
4150 assert((vecWidth == 128 || vecWidth == 256) && "Unsupported vector width");
4151 if (!isa<ConstantSDNode>(N->getOperand(1).getNode()))
4152 llvm_unreachable("Illegal extract subvector for VEXTRACT");
4155 cast<ConstantSDNode>(N->getOperand(1).getNode())->getZExtValue();
4157 MVT VecVT = N->getOperand(0).getSimpleValueType();
4158 MVT ElVT = VecVT.getVectorElementType();
4160 unsigned NumElemsPerChunk = vecWidth / ElVT.getSizeInBits();
4161 return Index / NumElemsPerChunk;
4164 static unsigned getInsertVINSERTImmediate(SDNode *N, unsigned vecWidth) {
4165 assert((vecWidth == 128 || vecWidth == 256) && "Unsupported vector width");
4166 if (!isa<ConstantSDNode>(N->getOperand(2).getNode()))
4167 llvm_unreachable("Illegal insert subvector for VINSERT");
4170 cast<ConstantSDNode>(N->getOperand(2).getNode())->getZExtValue();
4172 MVT VecVT = N->getSimpleValueType(0);
4173 MVT ElVT = VecVT.getVectorElementType();
4175 unsigned NumElemsPerChunk = vecWidth / ElVT.getSizeInBits();
4176 return Index / NumElemsPerChunk;
4179 /// Return the appropriate immediate to extract the specified
4180 /// EXTRACT_SUBVECTOR index with VEXTRACTF128 and VINSERTI128 instructions.
4181 unsigned X86::getExtractVEXTRACT128Immediate(SDNode *N) {
4182 return getExtractVEXTRACTImmediate(N, 128);
4185 /// Return the appropriate immediate to extract the specified
4186 /// EXTRACT_SUBVECTOR index with VEXTRACTF64x4 and VINSERTI64x4 instructions.
4187 unsigned X86::getExtractVEXTRACT256Immediate(SDNode *N) {
4188 return getExtractVEXTRACTImmediate(N, 256);
4191 /// Return the appropriate immediate to insert at the specified
4192 /// INSERT_SUBVECTOR index with VINSERTF128 and VINSERTI128 instructions.
4193 unsigned X86::getInsertVINSERT128Immediate(SDNode *N) {
4194 return getInsertVINSERTImmediate(N, 128);
4197 /// Return the appropriate immediate to insert at the specified
4198 /// INSERT_SUBVECTOR index with VINSERTF46x4 and VINSERTI64x4 instructions.
4199 unsigned X86::getInsertVINSERT256Immediate(SDNode *N) {
4200 return getInsertVINSERTImmediate(N, 256);
4203 /// Returns true if Elt is a constant integer zero
4204 static bool isZero(SDValue V) {
4205 ConstantSDNode *C = dyn_cast<ConstantSDNode>(V);
4206 return C && C->isNullValue();
4209 /// Returns true if Elt is a constant zero or a floating point constant +0.0.
4210 bool X86::isZeroNode(SDValue Elt) {
4213 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(Elt))
4214 return CFP->getValueAPF().isPosZero();
4218 /// Returns a vector of specified type with all zero elements.
4219 static SDValue getZeroVector(EVT VT, const X86Subtarget *Subtarget,
4220 SelectionDAG &DAG, SDLoc dl) {
4221 assert(VT.isVector() && "Expected a vector type");
4223 // Always build SSE zero vectors as <4 x i32> bitcasted
4224 // to their dest type. This ensures they get CSE'd.
4226 if (VT.is128BitVector()) { // SSE
4227 if (Subtarget->hasSSE2()) { // SSE2
4228 SDValue Cst = DAG.getConstant(0, dl, MVT::i32);
4229 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
4231 SDValue Cst = DAG.getConstantFP(+0.0, dl, MVT::f32);
4232 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4f32, Cst, Cst, Cst, Cst);
4234 } else if (VT.is256BitVector()) { // AVX
4235 if (Subtarget->hasInt256()) { // AVX2
4236 SDValue Cst = DAG.getConstant(0, dl, MVT::i32);
4237 SDValue Ops[] = { Cst, Cst, Cst, Cst, Cst, Cst, Cst, Cst };
4238 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v8i32, Ops);
4240 // 256-bit logic and arithmetic instructions in AVX are all
4241 // floating-point, no support for integer ops. Emit fp zeroed vectors.
4242 SDValue Cst = DAG.getConstantFP(+0.0, dl, MVT::f32);
4243 SDValue Ops[] = { Cst, Cst, Cst, Cst, Cst, Cst, Cst, Cst };
4244 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v8f32, Ops);
4246 } else if (VT.is512BitVector()) { // AVX-512
4247 SDValue Cst = DAG.getConstant(0, dl, MVT::i32);
4248 SDValue Ops[] = { Cst, Cst, Cst, Cst, Cst, Cst, Cst, Cst,
4249 Cst, Cst, Cst, Cst, Cst, Cst, Cst, Cst };
4250 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v16i32, Ops);
4251 } else if (VT.getScalarType() == MVT::i1) {
4253 assert((Subtarget->hasBWI() || VT.getVectorNumElements() <= 16)
4254 && "Unexpected vector type");
4255 assert((Subtarget->hasVLX() || VT.getVectorNumElements() >= 8)
4256 && "Unexpected vector type");
4257 SDValue Cst = DAG.getConstant(0, dl, MVT::i1);
4258 SmallVector<SDValue, 64> Ops(VT.getVectorNumElements(), Cst);
4259 return DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Ops);
4261 llvm_unreachable("Unexpected vector type");
4263 return DAG.getBitcast(VT, Vec);
4266 static SDValue ExtractSubVector(SDValue Vec, unsigned IdxVal,
4267 SelectionDAG &DAG, SDLoc dl,
4268 unsigned vectorWidth) {
4269 assert((vectorWidth == 128 || vectorWidth == 256) &&
4270 "Unsupported vector width");
4271 EVT VT = Vec.getValueType();
4272 EVT ElVT = VT.getVectorElementType();
4273 unsigned Factor = VT.getSizeInBits()/vectorWidth;
4274 EVT ResultVT = EVT::getVectorVT(*DAG.getContext(), ElVT,
4275 VT.getVectorNumElements()/Factor);
4277 // Extract from UNDEF is UNDEF.
4278 if (Vec.getOpcode() == ISD::UNDEF)
4279 return DAG.getUNDEF(ResultVT);
4281 // Extract the relevant vectorWidth bits. Generate an EXTRACT_SUBVECTOR
4282 unsigned ElemsPerChunk = vectorWidth / ElVT.getSizeInBits();
4284 // This is the index of the first element of the vectorWidth-bit chunk
4286 unsigned NormalizedIdxVal = (((IdxVal * ElVT.getSizeInBits()) / vectorWidth)
4289 // If the input is a buildvector just emit a smaller one.
4290 if (Vec.getOpcode() == ISD::BUILD_VECTOR)
4291 return DAG.getNode(ISD::BUILD_VECTOR, dl, ResultVT,
4292 makeArrayRef(Vec->op_begin() + NormalizedIdxVal,
4295 SDValue VecIdx = DAG.getIntPtrConstant(NormalizedIdxVal, dl);
4296 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, ResultVT, Vec, VecIdx);
4299 /// Generate a DAG to grab 128-bits from a vector > 128 bits. This
4300 /// sets things up to match to an AVX VEXTRACTF128 / VEXTRACTI128
4301 /// or AVX-512 VEXTRACTF32x4 / VEXTRACTI32x4
4302 /// instructions or a simple subregister reference. Idx is an index in the
4303 /// 128 bits we want. It need not be aligned to a 128-bit boundary. That makes
4304 /// lowering EXTRACT_VECTOR_ELT operations easier.
4305 static SDValue Extract128BitVector(SDValue Vec, unsigned IdxVal,
4306 SelectionDAG &DAG, SDLoc dl) {
4307 assert((Vec.getValueType().is256BitVector() ||
4308 Vec.getValueType().is512BitVector()) && "Unexpected vector size!");
4309 return ExtractSubVector(Vec, IdxVal, DAG, dl, 128);
4312 /// Generate a DAG to grab 256-bits from a 512-bit vector.
4313 static SDValue Extract256BitVector(SDValue Vec, unsigned IdxVal,
4314 SelectionDAG &DAG, SDLoc dl) {
4315 assert(Vec.getValueType().is512BitVector() && "Unexpected vector size!");
4316 return ExtractSubVector(Vec, IdxVal, DAG, dl, 256);
4319 static SDValue InsertSubVector(SDValue Result, SDValue Vec,
4320 unsigned IdxVal, SelectionDAG &DAG,
4321 SDLoc dl, unsigned vectorWidth) {
4322 assert((vectorWidth == 128 || vectorWidth == 256) &&
4323 "Unsupported vector width");
4324 // Inserting UNDEF is Result
4325 if (Vec.getOpcode() == ISD::UNDEF)
4327 EVT VT = Vec.getValueType();
4328 EVT ElVT = VT.getVectorElementType();
4329 EVT ResultVT = Result.getValueType();
4331 // Insert the relevant vectorWidth bits.
4332 unsigned ElemsPerChunk = vectorWidth/ElVT.getSizeInBits();
4334 // This is the index of the first element of the vectorWidth-bit chunk
4336 unsigned NormalizedIdxVal = (((IdxVal * ElVT.getSizeInBits())/vectorWidth)
4339 SDValue VecIdx = DAG.getIntPtrConstant(NormalizedIdxVal, dl);
4340 return DAG.getNode(ISD::INSERT_SUBVECTOR, dl, ResultVT, Result, Vec, VecIdx);
4343 /// Generate a DAG to put 128-bits into a vector > 128 bits. This
4344 /// sets things up to match to an AVX VINSERTF128/VINSERTI128 or
4345 /// AVX-512 VINSERTF32x4/VINSERTI32x4 instructions or a
4346 /// simple superregister reference. Idx is an index in the 128 bits
4347 /// we want. It need not be aligned to a 128-bit boundary. That makes
4348 /// lowering INSERT_VECTOR_ELT operations easier.
4349 static SDValue Insert128BitVector(SDValue Result, SDValue Vec, unsigned IdxVal,
4350 SelectionDAG &DAG, SDLoc dl) {
4351 assert(Vec.getValueType().is128BitVector() && "Unexpected vector size!");
4353 // For insertion into the zero index (low half) of a 256-bit vector, it is
4354 // more efficient to generate a blend with immediate instead of an insert*128.
4355 // We are still creating an INSERT_SUBVECTOR below with an undef node to
4356 // extend the subvector to the size of the result vector. Make sure that
4357 // we are not recursing on that node by checking for undef here.
4358 if (IdxVal == 0 && Result.getValueType().is256BitVector() &&
4359 Result.getOpcode() != ISD::UNDEF) {
4360 EVT ResultVT = Result.getValueType();
4361 SDValue ZeroIndex = DAG.getIntPtrConstant(0, dl);
4362 SDValue Undef = DAG.getUNDEF(ResultVT);
4363 SDValue Vec256 = DAG.getNode(ISD::INSERT_SUBVECTOR, dl, ResultVT, Undef,
4366 // The blend instruction, and therefore its mask, depend on the data type.
4367 MVT ScalarType = ResultVT.getScalarType().getSimpleVT();
4368 if (ScalarType.isFloatingPoint()) {
4369 // Choose either vblendps (float) or vblendpd (double).
4370 unsigned ScalarSize = ScalarType.getSizeInBits();
4371 assert((ScalarSize == 64 || ScalarSize == 32) && "Unknown float type");
4372 unsigned MaskVal = (ScalarSize == 64) ? 0x03 : 0x0f;
4373 SDValue Mask = DAG.getConstant(MaskVal, dl, MVT::i8);
4374 return DAG.getNode(X86ISD::BLENDI, dl, ResultVT, Result, Vec256, Mask);
4377 const X86Subtarget &Subtarget =
4378 static_cast<const X86Subtarget &>(DAG.getSubtarget());
4380 // AVX2 is needed for 256-bit integer blend support.
4381 // Integers must be cast to 32-bit because there is only vpblendd;
4382 // vpblendw can't be used for this because it has a handicapped mask.
4384 // If we don't have AVX2, then cast to float. Using a wrong domain blend
4385 // is still more efficient than using the wrong domain vinsertf128 that
4386 // will be created by InsertSubVector().
4387 MVT CastVT = Subtarget.hasAVX2() ? MVT::v8i32 : MVT::v8f32;
4389 SDValue Mask = DAG.getConstant(0x0f, dl, MVT::i8);
4390 Vec256 = DAG.getBitcast(CastVT, Vec256);
4391 Vec256 = DAG.getNode(X86ISD::BLENDI, dl, CastVT, Result, Vec256, Mask);
4392 return DAG.getBitcast(ResultVT, Vec256);
4395 return InsertSubVector(Result, Vec, IdxVal, DAG, dl, 128);
4398 static SDValue Insert256BitVector(SDValue Result, SDValue Vec, unsigned IdxVal,
4399 SelectionDAG &DAG, SDLoc dl) {
4400 assert(Vec.getValueType().is256BitVector() && "Unexpected vector size!");
4401 return InsertSubVector(Result, Vec, IdxVal, DAG, dl, 256);
4404 /// Concat two 128-bit vectors into a 256 bit vector using VINSERTF128
4405 /// instructions. This is used because creating CONCAT_VECTOR nodes of
4406 /// BUILD_VECTORS returns a larger BUILD_VECTOR while we're trying to lower
4407 /// large BUILD_VECTORS.
4408 static SDValue Concat128BitVectors(SDValue V1, SDValue V2, EVT VT,
4409 unsigned NumElems, SelectionDAG &DAG,
4411 SDValue V = Insert128BitVector(DAG.getUNDEF(VT), V1, 0, DAG, dl);
4412 return Insert128BitVector(V, V2, NumElems/2, DAG, dl);
4415 static SDValue Concat256BitVectors(SDValue V1, SDValue V2, EVT VT,
4416 unsigned NumElems, SelectionDAG &DAG,
4418 SDValue V = Insert256BitVector(DAG.getUNDEF(VT), V1, 0, DAG, dl);
4419 return Insert256BitVector(V, V2, NumElems/2, DAG, dl);
4422 /// Returns a vector of specified type with all bits set.
4423 /// Always build ones vectors as <4 x i32> or <8 x i32>. For 256-bit types with
4424 /// no AVX2 supprt, use two <4 x i32> inserted in a <8 x i32> appropriately.
4425 /// Then bitcast to their original type, ensuring they get CSE'd.
4426 static SDValue getOnesVector(EVT VT, const X86Subtarget *Subtarget,
4427 SelectionDAG &DAG, SDLoc dl) {
4428 assert(VT.isVector() && "Expected a vector type");
4430 SDValue Cst = DAG.getConstant(~0U, dl, MVT::i32);
4432 if (VT.is512BitVector()) {
4433 SDValue Ops[] = { Cst, Cst, Cst, Cst, Cst, Cst, Cst, Cst,
4434 Cst, Cst, Cst, Cst, Cst, Cst, Cst, Cst };
4435 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v16i32, Ops);
4436 } else if (VT.is256BitVector()) {
4437 if (Subtarget->hasInt256()) { // AVX2
4438 SDValue Ops[] = { Cst, Cst, Cst, Cst, Cst, Cst, Cst, Cst };
4439 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v8i32, Ops);
4441 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
4442 Vec = Concat128BitVectors(Vec, Vec, MVT::v8i32, 8, DAG, dl);
4444 } else if (VT.is128BitVector()) {
4445 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
4447 llvm_unreachable("Unexpected vector type");
4449 return DAG.getBitcast(VT, Vec);
4452 /// Returns a vector_shuffle node for an unpackl operation.
4453 static SDValue getUnpackl(SelectionDAG &DAG, SDLoc dl, MVT VT, SDValue V1,
4455 unsigned NumElems = VT.getVectorNumElements();
4456 SmallVector<int, 8> Mask;
4457 for (unsigned i = 0, e = NumElems/2; i != e; ++i) {
4459 Mask.push_back(i + NumElems);
4461 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
4464 /// Returns a vector_shuffle node for an unpackh operation.
4465 static SDValue getUnpackh(SelectionDAG &DAG, SDLoc dl, MVT VT, SDValue V1,
4467 unsigned NumElems = VT.getVectorNumElements();
4468 SmallVector<int, 8> Mask;
4469 for (unsigned i = 0, Half = NumElems/2; i != Half; ++i) {
4470 Mask.push_back(i + Half);
4471 Mask.push_back(i + NumElems + Half);
4473 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
4476 /// Return a vector_shuffle of the specified vector of zero or undef vector.
4477 /// This produces a shuffle where the low element of V2 is swizzled into the
4478 /// zero/undef vector, landing at element Idx.
4479 /// This produces a shuffle mask like 4,1,2,3 (idx=0) or 0,1,2,4 (idx=3).
4480 static SDValue getShuffleVectorZeroOrUndef(SDValue V2, unsigned Idx,
4482 const X86Subtarget *Subtarget,
4483 SelectionDAG &DAG) {
4484 MVT VT = V2.getSimpleValueType();
4486 ? getZeroVector(VT, Subtarget, DAG, SDLoc(V2)) : DAG.getUNDEF(VT);
4487 unsigned NumElems = VT.getVectorNumElements();
4488 SmallVector<int, 16> MaskVec;
4489 for (unsigned i = 0; i != NumElems; ++i)
4490 // If this is the insertion idx, put the low elt of V2 here.
4491 MaskVec.push_back(i == Idx ? NumElems : i);
4492 return DAG.getVectorShuffle(VT, SDLoc(V2), V1, V2, &MaskVec[0]);
4495 /// Calculates the shuffle mask corresponding to the target-specific opcode.
4496 /// Returns true if the Mask could be calculated. Sets IsUnary to true if only
4497 /// uses one source. Note that this will set IsUnary for shuffles which use a
4498 /// single input multiple times, and in those cases it will
4499 /// adjust the mask to only have indices within that single input.
4500 /// FIXME: Add support for Decode*Mask functions that return SM_SentinelZero.
4501 static bool getTargetShuffleMask(SDNode *N, MVT VT,
4502 SmallVectorImpl<int> &Mask, bool &IsUnary) {
4503 unsigned NumElems = VT.getVectorNumElements();
4507 bool IsFakeUnary = false;
4508 switch(N->getOpcode()) {
4509 case X86ISD::BLENDI:
4510 ImmN = N->getOperand(N->getNumOperands()-1);
4511 DecodeBLENDMask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
4514 ImmN = N->getOperand(N->getNumOperands()-1);
4515 DecodeSHUFPMask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
4516 IsUnary = IsFakeUnary = N->getOperand(0) == N->getOperand(1);
4518 case X86ISD::UNPCKH:
4519 DecodeUNPCKHMask(VT, Mask);
4520 IsUnary = IsFakeUnary = N->getOperand(0) == N->getOperand(1);
4522 case X86ISD::UNPCKL:
4523 DecodeUNPCKLMask(VT, Mask);
4524 IsUnary = IsFakeUnary = N->getOperand(0) == N->getOperand(1);
4526 case X86ISD::MOVHLPS:
4527 DecodeMOVHLPSMask(NumElems, Mask);
4528 IsUnary = IsFakeUnary = N->getOperand(0) == N->getOperand(1);
4530 case X86ISD::MOVLHPS:
4531 DecodeMOVLHPSMask(NumElems, Mask);
4532 IsUnary = IsFakeUnary = N->getOperand(0) == N->getOperand(1);
4534 case X86ISD::PALIGNR:
4535 ImmN = N->getOperand(N->getNumOperands()-1);
4536 DecodePALIGNRMask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
4538 case X86ISD::PSHUFD:
4539 case X86ISD::VPERMILPI:
4540 ImmN = N->getOperand(N->getNumOperands()-1);
4541 DecodePSHUFMask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
4544 case X86ISD::PSHUFHW:
4545 ImmN = N->getOperand(N->getNumOperands()-1);
4546 DecodePSHUFHWMask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
4549 case X86ISD::PSHUFLW:
4550 ImmN = N->getOperand(N->getNumOperands()-1);
4551 DecodePSHUFLWMask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
4554 case X86ISD::PSHUFB: {
4556 SDValue MaskNode = N->getOperand(1);
4557 while (MaskNode->getOpcode() == ISD::BITCAST)
4558 MaskNode = MaskNode->getOperand(0);
4560 if (MaskNode->getOpcode() == ISD::BUILD_VECTOR) {
4561 // If we have a build-vector, then things are easy.
4562 EVT VT = MaskNode.getValueType();
4563 assert(VT.isVector() &&
4564 "Can't produce a non-vector with a build_vector!");
4565 if (!VT.isInteger())
4568 int NumBytesPerElement = VT.getVectorElementType().getSizeInBits() / 8;
4570 SmallVector<uint64_t, 32> RawMask;
4571 for (int i = 0, e = MaskNode->getNumOperands(); i < e; ++i) {
4572 SDValue Op = MaskNode->getOperand(i);
4573 if (Op->getOpcode() == ISD::UNDEF) {
4574 RawMask.push_back((uint64_t)SM_SentinelUndef);
4577 auto *CN = dyn_cast<ConstantSDNode>(Op.getNode());
4580 APInt MaskElement = CN->getAPIntValue();
4582 // We now have to decode the element which could be any integer size and
4583 // extract each byte of it.
4584 for (int j = 0; j < NumBytesPerElement; ++j) {
4585 // Note that this is x86 and so always little endian: the low byte is
4586 // the first byte of the mask.
4587 RawMask.push_back(MaskElement.getLoBits(8).getZExtValue());
4588 MaskElement = MaskElement.lshr(8);
4591 DecodePSHUFBMask(RawMask, Mask);
4595 auto *MaskLoad = dyn_cast<LoadSDNode>(MaskNode);
4599 SDValue Ptr = MaskLoad->getBasePtr();
4600 if (Ptr->getOpcode() == X86ISD::Wrapper ||
4601 Ptr->getOpcode() == X86ISD::WrapperRIP)
4602 Ptr = Ptr->getOperand(0);
4604 auto *MaskCP = dyn_cast<ConstantPoolSDNode>(Ptr);
4605 if (!MaskCP || MaskCP->isMachineConstantPoolEntry())
4608 if (auto *C = dyn_cast<Constant>(MaskCP->getConstVal())) {
4609 DecodePSHUFBMask(C, Mask);
4617 case X86ISD::VPERMI:
4618 ImmN = N->getOperand(N->getNumOperands()-1);
4619 DecodeVPERMMask(cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
4624 DecodeScalarMoveMask(VT, /* IsLoad */ false, Mask);
4626 case X86ISD::VPERM2X128:
4627 ImmN = N->getOperand(N->getNumOperands()-1);
4628 DecodeVPERM2X128Mask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
4629 if (Mask.empty()) return false;
4630 // Mask only contains negative index if an element is zero.
4631 if (std::any_of(Mask.begin(), Mask.end(),
4632 [](int M){ return M == SM_SentinelZero; }))
4635 case X86ISD::MOVSLDUP:
4636 DecodeMOVSLDUPMask(VT, Mask);
4639 case X86ISD::MOVSHDUP:
4640 DecodeMOVSHDUPMask(VT, Mask);
4643 case X86ISD::MOVDDUP:
4644 DecodeMOVDDUPMask(VT, Mask);
4647 case X86ISD::MOVLHPD:
4648 case X86ISD::MOVLPD:
4649 case X86ISD::MOVLPS:
4650 // Not yet implemented
4652 case X86ISD::VPERMV: {
4654 SDValue MaskNode = N->getOperand(0);
4655 while (MaskNode->getOpcode() == ISD::BITCAST)
4656 MaskNode = MaskNode->getOperand(0);
4658 unsigned MaskLoBits = Log2_64(VT.getVectorNumElements());
4659 SmallVector<uint64_t, 32> RawMask;
4660 if (MaskNode->getOpcode() == ISD::BUILD_VECTOR) {
4661 // If we have a build-vector, then things are easy.
4662 assert(MaskNode.getValueType().isInteger() &&
4663 MaskNode.getValueType().getVectorNumElements() ==
4664 VT.getVectorNumElements());
4666 for (unsigned i = 0; i < MaskNode->getNumOperands(); ++i) {
4667 SDValue Op = MaskNode->getOperand(i);
4668 if (Op->getOpcode() == ISD::UNDEF)
4669 RawMask.push_back((uint64_t)SM_SentinelUndef);
4670 else if (isa<ConstantSDNode>(Op)) {
4671 APInt MaskElement = cast<ConstantSDNode>(Op)->getAPIntValue();
4672 RawMask.push_back(MaskElement.getLoBits(MaskLoBits).getZExtValue());
4676 DecodeVPERMVMask(RawMask, Mask);
4679 if (MaskNode->getOpcode() == X86ISD::VBROADCAST) {
4680 unsigned NumEltsInMask = MaskNode->getNumOperands();
4681 MaskNode = MaskNode->getOperand(0);
4682 auto *CN = dyn_cast<ConstantSDNode>(MaskNode);
4684 APInt MaskEltValue = CN->getAPIntValue();
4685 for (unsigned i = 0; i < NumEltsInMask; ++i)
4686 RawMask.push_back(MaskEltValue.getLoBits(MaskLoBits).getZExtValue());
4687 DecodeVPERMVMask(RawMask, Mask);
4690 // It may be a scalar load
4693 auto *MaskLoad = dyn_cast<LoadSDNode>(MaskNode);
4697 SDValue Ptr = MaskLoad->getBasePtr();
4698 if (Ptr->getOpcode() == X86ISD::Wrapper ||
4699 Ptr->getOpcode() == X86ISD::WrapperRIP)
4700 Ptr = Ptr->getOperand(0);
4702 auto *MaskCP = dyn_cast<ConstantPoolSDNode>(Ptr);
4703 if (!MaskCP || MaskCP->isMachineConstantPoolEntry())
4706 auto *C = dyn_cast<Constant>(MaskCP->getConstVal());
4708 DecodeVPERMVMask(C, VT, Mask);
4715 case X86ISD::VPERMV3: {
4717 SDValue MaskNode = N->getOperand(1);
4718 while (MaskNode->getOpcode() == ISD::BITCAST)
4719 MaskNode = MaskNode->getOperand(1);
4721 if (MaskNode->getOpcode() == ISD::BUILD_VECTOR) {
4722 // If we have a build-vector, then things are easy.
4723 assert(MaskNode.getValueType().isInteger() &&
4724 MaskNode.getValueType().getVectorNumElements() ==
4725 VT.getVectorNumElements());
4727 SmallVector<uint64_t, 32> RawMask;
4728 unsigned MaskLoBits = Log2_64(VT.getVectorNumElements()*2);
4730 for (unsigned i = 0; i < MaskNode->getNumOperands(); ++i) {
4731 SDValue Op = MaskNode->getOperand(i);
4732 if (Op->getOpcode() == ISD::UNDEF)
4733 RawMask.push_back((uint64_t)SM_SentinelUndef);
4735 auto *CN = dyn_cast<ConstantSDNode>(Op.getNode());
4738 APInt MaskElement = CN->getAPIntValue();
4739 RawMask.push_back(MaskElement.getLoBits(MaskLoBits).getZExtValue());
4742 DecodeVPERMV3Mask(RawMask, Mask);
4746 auto *MaskLoad = dyn_cast<LoadSDNode>(MaskNode);
4750 SDValue Ptr = MaskLoad->getBasePtr();
4751 if (Ptr->getOpcode() == X86ISD::Wrapper ||
4752 Ptr->getOpcode() == X86ISD::WrapperRIP)
4753 Ptr = Ptr->getOperand(0);
4755 auto *MaskCP = dyn_cast<ConstantPoolSDNode>(Ptr);
4756 if (!MaskCP || MaskCP->isMachineConstantPoolEntry())
4759 auto *C = dyn_cast<Constant>(MaskCP->getConstVal());
4761 DecodeVPERMV3Mask(C, VT, Mask);
4768 default: llvm_unreachable("unknown target shuffle node");
4771 // If we have a fake unary shuffle, the shuffle mask is spread across two
4772 // inputs that are actually the same node. Re-map the mask to always point
4773 // into the first input.
4776 if (M >= (int)Mask.size())
4782 /// Returns the scalar element that will make up the ith
4783 /// element of the result of the vector shuffle.
4784 static SDValue getShuffleScalarElt(SDNode *N, unsigned Index, SelectionDAG &DAG,
4787 return SDValue(); // Limit search depth.
4789 SDValue V = SDValue(N, 0);
4790 EVT VT = V.getValueType();
4791 unsigned Opcode = V.getOpcode();
4793 // Recurse into ISD::VECTOR_SHUFFLE node to find scalars.
4794 if (const ShuffleVectorSDNode *SV = dyn_cast<ShuffleVectorSDNode>(N)) {
4795 int Elt = SV->getMaskElt(Index);
4798 return DAG.getUNDEF(VT.getVectorElementType());
4800 unsigned NumElems = VT.getVectorNumElements();
4801 SDValue NewV = (Elt < (int)NumElems) ? SV->getOperand(0)
4802 : SV->getOperand(1);
4803 return getShuffleScalarElt(NewV.getNode(), Elt % NumElems, DAG, Depth+1);
4806 // Recurse into target specific vector shuffles to find scalars.
4807 if (isTargetShuffle(Opcode)) {
4808 MVT ShufVT = V.getSimpleValueType();
4809 unsigned NumElems = ShufVT.getVectorNumElements();
4810 SmallVector<int, 16> ShuffleMask;
4813 if (!getTargetShuffleMask(N, ShufVT, ShuffleMask, IsUnary))
4816 int Elt = ShuffleMask[Index];
4818 return DAG.getUNDEF(ShufVT.getVectorElementType());
4820 SDValue NewV = (Elt < (int)NumElems) ? N->getOperand(0)
4822 return getShuffleScalarElt(NewV.getNode(), Elt % NumElems, DAG,
4826 // Actual nodes that may contain scalar elements
4827 if (Opcode == ISD::BITCAST) {
4828 V = V.getOperand(0);
4829 EVT SrcVT = V.getValueType();
4830 unsigned NumElems = VT.getVectorNumElements();
4832 if (!SrcVT.isVector() || SrcVT.getVectorNumElements() != NumElems)
4836 if (V.getOpcode() == ISD::SCALAR_TO_VECTOR)
4837 return (Index == 0) ? V.getOperand(0)
4838 : DAG.getUNDEF(VT.getVectorElementType());
4840 if (V.getOpcode() == ISD::BUILD_VECTOR)
4841 return V.getOperand(Index);
4846 /// Custom lower build_vector of v16i8.
4847 static SDValue LowerBuildVectorv16i8(SDValue Op, unsigned NonZeros,
4848 unsigned NumNonZero, unsigned NumZero,
4850 const X86Subtarget* Subtarget,
4851 const TargetLowering &TLI) {
4859 // SSE4.1 - use PINSRB to insert each byte directly.
4860 if (Subtarget->hasSSE41()) {
4861 for (unsigned i = 0; i < 16; ++i) {
4862 bool isNonZero = (NonZeros & (1 << i)) != 0;
4866 V = getZeroVector(MVT::v16i8, Subtarget, DAG, dl);
4868 V = DAG.getUNDEF(MVT::v16i8);
4871 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl,
4872 MVT::v16i8, V, Op.getOperand(i),
4873 DAG.getIntPtrConstant(i, dl));
4880 // Pre-SSE4.1 - merge byte pairs and insert with PINSRW.
4881 for (unsigned i = 0; i < 16; ++i) {
4882 bool ThisIsNonZero = (NonZeros & (1 << i)) != 0;
4883 if (ThisIsNonZero && First) {
4885 V = getZeroVector(MVT::v8i16, Subtarget, DAG, dl);
4887 V = DAG.getUNDEF(MVT::v8i16);
4892 SDValue ThisElt, LastElt;
4893 bool LastIsNonZero = (NonZeros & (1 << (i-1))) != 0;
4894 if (LastIsNonZero) {
4895 LastElt = DAG.getNode(ISD::ZERO_EXTEND, dl,
4896 MVT::i16, Op.getOperand(i-1));
4898 if (ThisIsNonZero) {
4899 ThisElt = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i16, Op.getOperand(i));
4900 ThisElt = DAG.getNode(ISD::SHL, dl, MVT::i16,
4901 ThisElt, DAG.getConstant(8, dl, MVT::i8));
4903 ThisElt = DAG.getNode(ISD::OR, dl, MVT::i16, ThisElt, LastElt);
4907 if (ThisElt.getNode())
4908 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, V, ThisElt,
4909 DAG.getIntPtrConstant(i/2, dl));
4913 return DAG.getBitcast(MVT::v16i8, V);
4916 /// Custom lower build_vector of v8i16.
4917 static SDValue LowerBuildVectorv8i16(SDValue Op, unsigned NonZeros,
4918 unsigned NumNonZero, unsigned NumZero,
4920 const X86Subtarget* Subtarget,
4921 const TargetLowering &TLI) {
4928 for (unsigned i = 0; i < 8; ++i) {
4929 bool isNonZero = (NonZeros & (1 << i)) != 0;
4933 V = getZeroVector(MVT::v8i16, Subtarget, DAG, dl);
4935 V = DAG.getUNDEF(MVT::v8i16);
4938 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl,
4939 MVT::v8i16, V, Op.getOperand(i),
4940 DAG.getIntPtrConstant(i, dl));
4947 /// Custom lower build_vector of v4i32 or v4f32.
4948 static SDValue LowerBuildVectorv4x32(SDValue Op, SelectionDAG &DAG,
4949 const X86Subtarget *Subtarget,
4950 const TargetLowering &TLI) {
4951 // Find all zeroable elements.
4952 std::bitset<4> Zeroable;
4953 for (int i=0; i < 4; ++i) {
4954 SDValue Elt = Op->getOperand(i);
4955 Zeroable[i] = (Elt.getOpcode() == ISD::UNDEF || X86::isZeroNode(Elt));
4957 assert(Zeroable.size() - Zeroable.count() > 1 &&
4958 "We expect at least two non-zero elements!");
4960 // We only know how to deal with build_vector nodes where elements are either
4961 // zeroable or extract_vector_elt with constant index.
4962 SDValue FirstNonZero;
4963 unsigned FirstNonZeroIdx;
4964 for (unsigned i=0; i < 4; ++i) {
4967 SDValue Elt = Op->getOperand(i);
4968 if (Elt.getOpcode() != ISD::EXTRACT_VECTOR_ELT ||
4969 !isa<ConstantSDNode>(Elt.getOperand(1)))
4971 // Make sure that this node is extracting from a 128-bit vector.
4972 MVT VT = Elt.getOperand(0).getSimpleValueType();
4973 if (!VT.is128BitVector())
4975 if (!FirstNonZero.getNode()) {
4977 FirstNonZeroIdx = i;
4981 assert(FirstNonZero.getNode() && "Unexpected build vector of all zeros!");
4982 SDValue V1 = FirstNonZero.getOperand(0);
4983 MVT VT = V1.getSimpleValueType();
4985 // See if this build_vector can be lowered as a blend with zero.
4987 unsigned EltMaskIdx, EltIdx;
4989 for (EltIdx = 0; EltIdx < 4; ++EltIdx) {
4990 if (Zeroable[EltIdx]) {
4991 // The zero vector will be on the right hand side.
4992 Mask[EltIdx] = EltIdx+4;
4996 Elt = Op->getOperand(EltIdx);
4997 // By construction, Elt is a EXTRACT_VECTOR_ELT with constant index.
4998 EltMaskIdx = cast<ConstantSDNode>(Elt.getOperand(1))->getZExtValue();
4999 if (Elt.getOperand(0) != V1 || EltMaskIdx != EltIdx)
5001 Mask[EltIdx] = EltIdx;
5005 // Let the shuffle legalizer deal with blend operations.
5006 SDValue VZero = getZeroVector(VT, Subtarget, DAG, SDLoc(Op));
5007 if (V1.getSimpleValueType() != VT)
5008 V1 = DAG.getNode(ISD::BITCAST, SDLoc(V1), VT, V1);
5009 return DAG.getVectorShuffle(VT, SDLoc(V1), V1, VZero, &Mask[0]);
5012 // See if we can lower this build_vector to a INSERTPS.
5013 if (!Subtarget->hasSSE41())
5016 SDValue V2 = Elt.getOperand(0);
5017 if (Elt == FirstNonZero && EltIdx == FirstNonZeroIdx)
5020 bool CanFold = true;
5021 for (unsigned i = EltIdx + 1; i < 4 && CanFold; ++i) {
5025 SDValue Current = Op->getOperand(i);
5026 SDValue SrcVector = Current->getOperand(0);
5029 CanFold = SrcVector == V1 &&
5030 cast<ConstantSDNode>(Current.getOperand(1))->getZExtValue() == i;
5036 assert(V1.getNode() && "Expected at least two non-zero elements!");
5037 if (V1.getSimpleValueType() != MVT::v4f32)
5038 V1 = DAG.getNode(ISD::BITCAST, SDLoc(V1), MVT::v4f32, V1);
5039 if (V2.getSimpleValueType() != MVT::v4f32)
5040 V2 = DAG.getNode(ISD::BITCAST, SDLoc(V2), MVT::v4f32, V2);
5042 // Ok, we can emit an INSERTPS instruction.
5043 unsigned ZMask = Zeroable.to_ulong();
5045 unsigned InsertPSMask = EltMaskIdx << 6 | EltIdx << 4 | ZMask;
5046 assert((InsertPSMask & ~0xFFu) == 0 && "Invalid mask!");
5048 SDValue Result = DAG.getNode(X86ISD::INSERTPS, DL, MVT::v4f32, V1, V2,
5049 DAG.getIntPtrConstant(InsertPSMask, DL));
5050 return DAG.getBitcast(VT, Result);
5053 /// Return a vector logical shift node.
5054 static SDValue getVShift(bool isLeft, EVT VT, SDValue SrcOp,
5055 unsigned NumBits, SelectionDAG &DAG,
5056 const TargetLowering &TLI, SDLoc dl) {
5057 assert(VT.is128BitVector() && "Unknown type for VShift");
5058 MVT ShVT = MVT::v2i64;
5059 unsigned Opc = isLeft ? X86ISD::VSHLDQ : X86ISD::VSRLDQ;
5060 SrcOp = DAG.getBitcast(ShVT, SrcOp);
5061 MVT ScalarShiftTy = TLI.getScalarShiftAmountTy(DAG.getDataLayout(), VT);
5062 assert(NumBits % 8 == 0 && "Only support byte sized shifts");
5063 SDValue ShiftVal = DAG.getConstant(NumBits/8, dl, ScalarShiftTy);
5064 return DAG.getBitcast(VT, DAG.getNode(Opc, dl, ShVT, SrcOp, ShiftVal));
5068 LowerAsSplatVectorLoad(SDValue SrcOp, MVT VT, SDLoc dl, SelectionDAG &DAG) {
5070 // Check if the scalar load can be widened into a vector load. And if
5071 // the address is "base + cst" see if the cst can be "absorbed" into
5072 // the shuffle mask.
5073 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(SrcOp)) {
5074 SDValue Ptr = LD->getBasePtr();
5075 if (!ISD::isNormalLoad(LD) || LD->isVolatile())
5077 EVT PVT = LD->getValueType(0);
5078 if (PVT != MVT::i32 && PVT != MVT::f32)
5083 if (FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr)) {
5084 FI = FINode->getIndex();
5086 } else if (DAG.isBaseWithConstantOffset(Ptr) &&
5087 isa<FrameIndexSDNode>(Ptr.getOperand(0))) {
5088 FI = cast<FrameIndexSDNode>(Ptr.getOperand(0))->getIndex();
5089 Offset = Ptr.getConstantOperandVal(1);
5090 Ptr = Ptr.getOperand(0);
5095 // FIXME: 256-bit vector instructions don't require a strict alignment,
5096 // improve this code to support it better.
5097 unsigned RequiredAlign = VT.getSizeInBits()/8;
5098 SDValue Chain = LD->getChain();
5099 // Make sure the stack object alignment is at least 16 or 32.
5100 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
5101 if (DAG.InferPtrAlignment(Ptr) < RequiredAlign) {
5102 if (MFI->isFixedObjectIndex(FI)) {
5103 // Can't change the alignment. FIXME: It's possible to compute
5104 // the exact stack offset and reference FI + adjust offset instead.
5105 // If someone *really* cares about this. That's the way to implement it.
5108 MFI->setObjectAlignment(FI, RequiredAlign);
5112 // (Offset % 16 or 32) must be multiple of 4. Then address is then
5113 // Ptr + (Offset & ~15).
5116 if ((Offset % RequiredAlign) & 3)
5118 int64_t StartOffset = Offset & ~int64_t(RequiredAlign - 1);
5121 Ptr = DAG.getNode(ISD::ADD, DL, Ptr.getValueType(), Ptr,
5122 DAG.getConstant(StartOffset, DL, Ptr.getValueType()));
5125 int EltNo = (Offset - StartOffset) >> 2;
5126 unsigned NumElems = VT.getVectorNumElements();
5128 EVT NVT = EVT::getVectorVT(*DAG.getContext(), PVT, NumElems);
5129 SDValue V1 = DAG.getLoad(NVT, dl, Chain, Ptr,
5130 LD->getPointerInfo().getWithOffset(StartOffset),
5131 false, false, false, 0);
5133 SmallVector<int, 8> Mask(NumElems, EltNo);
5135 return DAG.getVectorShuffle(NVT, dl, V1, DAG.getUNDEF(NVT), &Mask[0]);
5141 /// Given the initializing elements 'Elts' of a vector of type 'VT', see if the
5142 /// elements can be replaced by a single large load which has the same value as
5143 /// a build_vector or insert_subvector whose loaded operands are 'Elts'.
5145 /// Example: <load i32 *a, load i32 *a+4, undef, undef> -> zextload a
5147 /// FIXME: we'd also like to handle the case where the last elements are zero
5148 /// rather than undef via VZEXT_LOAD, but we do not detect that case today.
5149 /// There's even a handy isZeroNode for that purpose.
5150 static SDValue EltsFromConsecutiveLoads(EVT VT, ArrayRef<SDValue> Elts,
5151 SDLoc &DL, SelectionDAG &DAG,
5152 bool isAfterLegalize) {
5153 unsigned NumElems = Elts.size();
5155 LoadSDNode *LDBase = nullptr;
5156 unsigned LastLoadedElt = -1U;
5158 // For each element in the initializer, see if we've found a load or an undef.
5159 // If we don't find an initial load element, or later load elements are
5160 // non-consecutive, bail out.
5161 for (unsigned i = 0; i < NumElems; ++i) {
5162 SDValue Elt = Elts[i];
5163 // Look through a bitcast.
5164 if (Elt.getNode() && Elt.getOpcode() == ISD::BITCAST)
5165 Elt = Elt.getOperand(0);
5166 if (!Elt.getNode() ||
5167 (Elt.getOpcode() != ISD::UNDEF && !ISD::isNON_EXTLoad(Elt.getNode())))
5170 if (Elt.getNode()->getOpcode() == ISD::UNDEF)
5172 LDBase = cast<LoadSDNode>(Elt.getNode());
5176 if (Elt.getOpcode() == ISD::UNDEF)
5179 LoadSDNode *LD = cast<LoadSDNode>(Elt);
5180 EVT LdVT = Elt.getValueType();
5181 // Each loaded element must be the correct fractional portion of the
5182 // requested vector load.
5183 if (LdVT.getSizeInBits() != VT.getSizeInBits() / NumElems)
5185 if (!DAG.isConsecutiveLoad(LD, LDBase, LdVT.getSizeInBits() / 8, i))
5190 // If we have found an entire vector of loads and undefs, then return a large
5191 // load of the entire vector width starting at the base pointer. If we found
5192 // consecutive loads for the low half, generate a vzext_load node.
5193 if (LastLoadedElt == NumElems - 1) {
5194 assert(LDBase && "Did not find base load for merging consecutive loads");
5195 EVT EltVT = LDBase->getValueType(0);
5196 // Ensure that the input vector size for the merged loads matches the
5197 // cumulative size of the input elements.
5198 if (VT.getSizeInBits() != EltVT.getSizeInBits() * NumElems)
5201 if (isAfterLegalize &&
5202 !DAG.getTargetLoweringInfo().isOperationLegal(ISD::LOAD, VT))
5205 SDValue NewLd = SDValue();
5207 NewLd = DAG.getLoad(VT, DL, LDBase->getChain(), LDBase->getBasePtr(),
5208 LDBase->getPointerInfo(), LDBase->isVolatile(),
5209 LDBase->isNonTemporal(), LDBase->isInvariant(),
5210 LDBase->getAlignment());
5212 if (LDBase->hasAnyUseOfValue(1)) {
5213 SDValue NewChain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other,
5215 SDValue(NewLd.getNode(), 1));
5216 DAG.ReplaceAllUsesOfValueWith(SDValue(LDBase, 1), NewChain);
5217 DAG.UpdateNodeOperands(NewChain.getNode(), SDValue(LDBase, 1),
5218 SDValue(NewLd.getNode(), 1));
5224 //TODO: The code below fires only for for loading the low v2i32 / v2f32
5225 //of a v4i32 / v4f32. It's probably worth generalizing.
5226 EVT EltVT = VT.getVectorElementType();
5227 if (NumElems == 4 && LastLoadedElt == 1 && (EltVT.getSizeInBits() == 32) &&
5228 DAG.getTargetLoweringInfo().isTypeLegal(MVT::v2i64)) {
5229 SDVTList Tys = DAG.getVTList(MVT::v2i64, MVT::Other);
5230 SDValue Ops[] = { LDBase->getChain(), LDBase->getBasePtr() };
5232 DAG.getMemIntrinsicNode(X86ISD::VZEXT_LOAD, DL, Tys, Ops, MVT::i64,
5233 LDBase->getPointerInfo(),
5234 LDBase->getAlignment(),
5235 false/*isVolatile*/, true/*ReadMem*/,
5238 // Make sure the newly-created LOAD is in the same position as LDBase in
5239 // terms of dependency. We create a TokenFactor for LDBase and ResNode, and
5240 // update uses of LDBase's output chain to use the TokenFactor.
5241 if (LDBase->hasAnyUseOfValue(1)) {
5242 SDValue NewChain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other,
5243 SDValue(LDBase, 1), SDValue(ResNode.getNode(), 1));
5244 DAG.ReplaceAllUsesOfValueWith(SDValue(LDBase, 1), NewChain);
5245 DAG.UpdateNodeOperands(NewChain.getNode(), SDValue(LDBase, 1),
5246 SDValue(ResNode.getNode(), 1));
5249 return DAG.getBitcast(VT, ResNode);
5254 /// LowerVectorBroadcast - Attempt to use the vbroadcast instruction
5255 /// to generate a splat value for the following cases:
5256 /// 1. A splat BUILD_VECTOR which uses a single scalar load, or a constant.
5257 /// 2. A splat shuffle which uses a scalar_to_vector node which comes from
5258 /// a scalar load, or a constant.
5259 /// The VBROADCAST node is returned when a pattern is found,
5260 /// or SDValue() otherwise.
5261 static SDValue LowerVectorBroadcast(SDValue Op, const X86Subtarget* Subtarget,
5262 SelectionDAG &DAG) {
5263 // VBROADCAST requires AVX.
5264 // TODO: Splats could be generated for non-AVX CPUs using SSE
5265 // instructions, but there's less potential gain for only 128-bit vectors.
5266 if (!Subtarget->hasAVX())
5269 MVT VT = Op.getSimpleValueType();
5272 assert((VT.is128BitVector() || VT.is256BitVector() || VT.is512BitVector()) &&
5273 "Unsupported vector type for broadcast.");
5278 switch (Op.getOpcode()) {
5280 // Unknown pattern found.
5283 case ISD::BUILD_VECTOR: {
5284 auto *BVOp = cast<BuildVectorSDNode>(Op.getNode());
5285 BitVector UndefElements;
5286 SDValue Splat = BVOp->getSplatValue(&UndefElements);
5288 // We need a splat of a single value to use broadcast, and it doesn't
5289 // make any sense if the value is only in one element of the vector.
5290 if (!Splat || (VT.getVectorNumElements() - UndefElements.count()) <= 1)
5294 ConstSplatVal = (Ld.getOpcode() == ISD::Constant ||
5295 Ld.getOpcode() == ISD::ConstantFP);
5297 // Make sure that all of the users of a non-constant load are from the
5298 // BUILD_VECTOR node.
5299 if (!ConstSplatVal && !BVOp->isOnlyUserOf(Ld.getNode()))
5304 case ISD::VECTOR_SHUFFLE: {
5305 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
5307 // Shuffles must have a splat mask where the first element is
5309 if ((!SVOp->isSplat()) || SVOp->getMaskElt(0) != 0)
5312 SDValue Sc = Op.getOperand(0);
5313 if (Sc.getOpcode() != ISD::SCALAR_TO_VECTOR &&
5314 Sc.getOpcode() != ISD::BUILD_VECTOR) {
5316 if (!Subtarget->hasInt256())
5319 // Use the register form of the broadcast instruction available on AVX2.
5320 if (VT.getSizeInBits() >= 256)
5321 Sc = Extract128BitVector(Sc, 0, DAG, dl);
5322 return DAG.getNode(X86ISD::VBROADCAST, dl, VT, Sc);
5325 Ld = Sc.getOperand(0);
5326 ConstSplatVal = (Ld.getOpcode() == ISD::Constant ||
5327 Ld.getOpcode() == ISD::ConstantFP);
5329 // The scalar_to_vector node and the suspected
5330 // load node must have exactly one user.
5331 // Constants may have multiple users.
5333 // AVX-512 has register version of the broadcast
5334 bool hasRegVer = Subtarget->hasAVX512() && VT.is512BitVector() &&
5335 Ld.getValueType().getSizeInBits() >= 32;
5336 if (!ConstSplatVal && ((!Sc.hasOneUse() || !Ld.hasOneUse()) &&
5343 unsigned ScalarSize = Ld.getValueType().getSizeInBits();
5344 bool IsGE256 = (VT.getSizeInBits() >= 256);
5346 // When optimizing for size, generate up to 5 extra bytes for a broadcast
5347 // instruction to save 8 or more bytes of constant pool data.
5348 // TODO: If multiple splats are generated to load the same constant,
5349 // it may be detrimental to overall size. There needs to be a way to detect
5350 // that condition to know if this is truly a size win.
5351 bool OptForSize = DAG.getMachineFunction().getFunction()->optForSize();
5353 // Handle broadcasting a single constant scalar from the constant pool
5355 // On Sandybridge (no AVX2), it is still better to load a constant vector
5356 // from the constant pool and not to broadcast it from a scalar.
5357 // But override that restriction when optimizing for size.
5358 // TODO: Check if splatting is recommended for other AVX-capable CPUs.
5359 if (ConstSplatVal && (Subtarget->hasAVX2() || OptForSize)) {
5360 EVT CVT = Ld.getValueType();
5361 assert(!CVT.isVector() && "Must not broadcast a vector type");
5363 // Splat f32, i32, v4f64, v4i64 in all cases with AVX2.
5364 // For size optimization, also splat v2f64 and v2i64, and for size opt
5365 // with AVX2, also splat i8 and i16.
5366 // With pattern matching, the VBROADCAST node may become a VMOVDDUP.
5367 if (ScalarSize == 32 || (IsGE256 && ScalarSize == 64) ||
5368 (OptForSize && (ScalarSize == 64 || Subtarget->hasAVX2()))) {
5369 const Constant *C = nullptr;
5370 if (ConstantSDNode *CI = dyn_cast<ConstantSDNode>(Ld))
5371 C = CI->getConstantIntValue();
5372 else if (ConstantFPSDNode *CF = dyn_cast<ConstantFPSDNode>(Ld))
5373 C = CF->getConstantFPValue();
5375 assert(C && "Invalid constant type");
5377 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
5379 DAG.getConstantPool(C, TLI.getPointerTy(DAG.getDataLayout()));
5380 unsigned Alignment = cast<ConstantPoolSDNode>(CP)->getAlignment();
5382 CVT, dl, DAG.getEntryNode(), CP,
5383 MachinePointerInfo::getConstantPool(DAG.getMachineFunction()), false,
5384 false, false, Alignment);
5386 return DAG.getNode(X86ISD::VBROADCAST, dl, VT, Ld);
5390 bool IsLoad = ISD::isNormalLoad(Ld.getNode());
5392 // Handle AVX2 in-register broadcasts.
5393 if (!IsLoad && Subtarget->hasInt256() &&
5394 (ScalarSize == 32 || (IsGE256 && ScalarSize == 64)))
5395 return DAG.getNode(X86ISD::VBROADCAST, dl, VT, Ld);
5397 // The scalar source must be a normal load.
5401 if (ScalarSize == 32 || (IsGE256 && ScalarSize == 64) ||
5402 (Subtarget->hasVLX() && ScalarSize == 64))
5403 return DAG.getNode(X86ISD::VBROADCAST, dl, VT, Ld);
5405 // The integer check is needed for the 64-bit into 128-bit so it doesn't match
5406 // double since there is no vbroadcastsd xmm
5407 if (Subtarget->hasInt256() && Ld.getValueType().isInteger()) {
5408 if (ScalarSize == 8 || ScalarSize == 16 || ScalarSize == 64)
5409 return DAG.getNode(X86ISD::VBROADCAST, dl, VT, Ld);
5412 // Unsupported broadcast.
5416 /// \brief For an EXTRACT_VECTOR_ELT with a constant index return the real
5417 /// underlying vector and index.
5419 /// Modifies \p ExtractedFromVec to the real vector and returns the real
5421 static int getUnderlyingExtractedFromVec(SDValue &ExtractedFromVec,
5423 int Idx = cast<ConstantSDNode>(ExtIdx)->getZExtValue();
5424 if (!isa<ShuffleVectorSDNode>(ExtractedFromVec))
5427 // For 256-bit vectors, LowerEXTRACT_VECTOR_ELT_SSE4 may have already
5429 // (extract_vector_elt (v8f32 %vreg1), Constant<6>)
5431 // (extract_vector_elt (vector_shuffle<2,u,u,u>
5432 // (extract_subvector (v8f32 %vreg0), Constant<4>),
5435 // In this case the vector is the extract_subvector expression and the index
5436 // is 2, as specified by the shuffle.
5437 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(ExtractedFromVec);
5438 SDValue ShuffleVec = SVOp->getOperand(0);
5439 MVT ShuffleVecVT = ShuffleVec.getSimpleValueType();
5440 assert(ShuffleVecVT.getVectorElementType() ==
5441 ExtractedFromVec.getSimpleValueType().getVectorElementType());
5443 int ShuffleIdx = SVOp->getMaskElt(Idx);
5444 if (isUndefOrInRange(ShuffleIdx, 0, ShuffleVecVT.getVectorNumElements())) {
5445 ExtractedFromVec = ShuffleVec;
5451 static SDValue buildFromShuffleMostly(SDValue Op, SelectionDAG &DAG) {
5452 MVT VT = Op.getSimpleValueType();
5454 // Skip if insert_vec_elt is not supported.
5455 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
5456 if (!TLI.isOperationLegalOrCustom(ISD::INSERT_VECTOR_ELT, VT))
5460 unsigned NumElems = Op.getNumOperands();
5464 SmallVector<unsigned, 4> InsertIndices;
5465 SmallVector<int, 8> Mask(NumElems, -1);
5467 for (unsigned i = 0; i != NumElems; ++i) {
5468 unsigned Opc = Op.getOperand(i).getOpcode();
5470 if (Opc == ISD::UNDEF)
5473 if (Opc != ISD::EXTRACT_VECTOR_ELT) {
5474 // Quit if more than 1 elements need inserting.
5475 if (InsertIndices.size() > 1)
5478 InsertIndices.push_back(i);
5482 SDValue ExtractedFromVec = Op.getOperand(i).getOperand(0);
5483 SDValue ExtIdx = Op.getOperand(i).getOperand(1);
5484 // Quit if non-constant index.
5485 if (!isa<ConstantSDNode>(ExtIdx))
5487 int Idx = getUnderlyingExtractedFromVec(ExtractedFromVec, ExtIdx);
5489 // Quit if extracted from vector of different type.
5490 if (ExtractedFromVec.getValueType() != VT)
5493 if (!VecIn1.getNode())
5494 VecIn1 = ExtractedFromVec;
5495 else if (VecIn1 != ExtractedFromVec) {
5496 if (!VecIn2.getNode())
5497 VecIn2 = ExtractedFromVec;
5498 else if (VecIn2 != ExtractedFromVec)
5499 // Quit if more than 2 vectors to shuffle
5503 if (ExtractedFromVec == VecIn1)
5505 else if (ExtractedFromVec == VecIn2)
5506 Mask[i] = Idx + NumElems;
5509 if (!VecIn1.getNode())
5512 VecIn2 = VecIn2.getNode() ? VecIn2 : DAG.getUNDEF(VT);
5513 SDValue NV = DAG.getVectorShuffle(VT, DL, VecIn1, VecIn2, &Mask[0]);
5514 for (unsigned i = 0, e = InsertIndices.size(); i != e; ++i) {
5515 unsigned Idx = InsertIndices[i];
5516 NV = DAG.getNode(ISD::INSERT_VECTOR_ELT, DL, VT, NV, Op.getOperand(Idx),
5517 DAG.getIntPtrConstant(Idx, DL));
5523 static SDValue ConvertI1VectorToInteger(SDValue Op, SelectionDAG &DAG) {
5524 assert(ISD::isBuildVectorOfConstantSDNodes(Op.getNode()) &&
5525 Op.getScalarValueSizeInBits() == 1 &&
5526 "Can not convert non-constant vector");
5527 uint64_t Immediate = 0;
5528 for (unsigned idx = 0, e = Op.getNumOperands(); idx < e; ++idx) {
5529 SDValue In = Op.getOperand(idx);
5530 if (In.getOpcode() != ISD::UNDEF)
5531 Immediate |= cast<ConstantSDNode>(In)->getZExtValue() << idx;
5535 MVT::getIntegerVT(std::max((int)Op.getValueType().getSizeInBits(), 8));
5536 return DAG.getConstant(Immediate, dl, VT);
5538 // Lower BUILD_VECTOR operation for v8i1 and v16i1 types.
5540 X86TargetLowering::LowerBUILD_VECTORvXi1(SDValue Op, SelectionDAG &DAG) const {
5542 MVT VT = Op.getSimpleValueType();
5543 assert((VT.getVectorElementType() == MVT::i1) &&
5544 "Unexpected type in LowerBUILD_VECTORvXi1!");
5547 if (ISD::isBuildVectorAllZeros(Op.getNode())) {
5548 SDValue Cst = DAG.getTargetConstant(0, dl, MVT::i1);
5549 SmallVector<SDValue, 16> Ops(VT.getVectorNumElements(), Cst);
5550 return DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Ops);
5553 if (ISD::isBuildVectorAllOnes(Op.getNode())) {
5554 SDValue Cst = DAG.getTargetConstant(1, dl, MVT::i1);
5555 SmallVector<SDValue, 16> Ops(VT.getVectorNumElements(), Cst);
5556 return DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Ops);
5559 if (ISD::isBuildVectorOfConstantSDNodes(Op.getNode())) {
5560 SDValue Imm = ConvertI1VectorToInteger(Op, DAG);
5561 if (Imm.getValueSizeInBits() == VT.getSizeInBits())
5562 return DAG.getBitcast(VT, Imm);
5563 SDValue ExtVec = DAG.getBitcast(MVT::v8i1, Imm);
5564 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT, ExtVec,
5565 DAG.getIntPtrConstant(0, dl));
5568 // Vector has one or more non-const elements
5569 uint64_t Immediate = 0;
5570 SmallVector<unsigned, 16> NonConstIdx;
5571 bool IsSplat = true;
5572 bool HasConstElts = false;
5574 for (unsigned idx = 0, e = Op.getNumOperands(); idx < e; ++idx) {
5575 SDValue In = Op.getOperand(idx);
5576 if (In.getOpcode() == ISD::UNDEF)
5578 if (!isa<ConstantSDNode>(In))
5579 NonConstIdx.push_back(idx);
5581 Immediate |= cast<ConstantSDNode>(In)->getZExtValue() << idx;
5582 HasConstElts = true;
5586 else if (In != Op.getOperand(SplatIdx))
5590 // for splat use " (select i1 splat_elt, all-ones, all-zeroes)"
5592 return DAG.getNode(ISD::SELECT, dl, VT, Op.getOperand(SplatIdx),
5593 DAG.getConstant(1, dl, VT),
5594 DAG.getConstant(0, dl, VT));
5596 // insert elements one by one
5600 MVT ImmVT = MVT::getIntegerVT(std::max((int)VT.getSizeInBits(), 8));
5601 Imm = DAG.getConstant(Immediate, dl, ImmVT);
5603 else if (HasConstElts)
5604 Imm = DAG.getConstant(0, dl, VT);
5606 Imm = DAG.getUNDEF(VT);
5607 if (Imm.getValueSizeInBits() == VT.getSizeInBits())
5608 DstVec = DAG.getBitcast(VT, Imm);
5610 SDValue ExtVec = DAG.getBitcast(MVT::v8i1, Imm);
5611 DstVec = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT, ExtVec,
5612 DAG.getIntPtrConstant(0, dl));
5615 for (unsigned i = 0; i < NonConstIdx.size(); ++i) {
5616 unsigned InsertIdx = NonConstIdx[i];
5617 DstVec = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, DstVec,
5618 Op.getOperand(InsertIdx),
5619 DAG.getIntPtrConstant(InsertIdx, dl));
5624 /// \brief Return true if \p N implements a horizontal binop and return the
5625 /// operands for the horizontal binop into V0 and V1.
5627 /// This is a helper function of LowerToHorizontalOp().
5628 /// This function checks that the build_vector \p N in input implements a
5629 /// horizontal operation. Parameter \p Opcode defines the kind of horizontal
5630 /// operation to match.
5631 /// For example, if \p Opcode is equal to ISD::ADD, then this function
5632 /// checks if \p N implements a horizontal arithmetic add; if instead \p Opcode
5633 /// is equal to ISD::SUB, then this function checks if this is a horizontal
5636 /// This function only analyzes elements of \p N whose indices are
5637 /// in range [BaseIdx, LastIdx).
5638 static bool isHorizontalBinOp(const BuildVectorSDNode *N, unsigned Opcode,
5640 unsigned BaseIdx, unsigned LastIdx,
5641 SDValue &V0, SDValue &V1) {
5642 EVT VT = N->getValueType(0);
5644 assert(BaseIdx * 2 <= LastIdx && "Invalid Indices in input!");
5645 assert(VT.isVector() && VT.getVectorNumElements() >= LastIdx &&
5646 "Invalid Vector in input!");
5648 bool IsCommutable = (Opcode == ISD::ADD || Opcode == ISD::FADD);
5649 bool CanFold = true;
5650 unsigned ExpectedVExtractIdx = BaseIdx;
5651 unsigned NumElts = LastIdx - BaseIdx;
5652 V0 = DAG.getUNDEF(VT);
5653 V1 = DAG.getUNDEF(VT);
5655 // Check if N implements a horizontal binop.
5656 for (unsigned i = 0, e = NumElts; i != e && CanFold; ++i) {
5657 SDValue Op = N->getOperand(i + BaseIdx);
5660 if (Op->getOpcode() == ISD::UNDEF) {
5661 // Update the expected vector extract index.
5662 if (i * 2 == NumElts)
5663 ExpectedVExtractIdx = BaseIdx;
5664 ExpectedVExtractIdx += 2;
5668 CanFold = Op->getOpcode() == Opcode && Op->hasOneUse();
5673 SDValue Op0 = Op.getOperand(0);
5674 SDValue Op1 = Op.getOperand(1);
5676 // Try to match the following pattern:
5677 // (BINOP (extract_vector_elt A, I), (extract_vector_elt A, I+1))
5678 CanFold = (Op0.getOpcode() == ISD::EXTRACT_VECTOR_ELT &&
5679 Op1.getOpcode() == ISD::EXTRACT_VECTOR_ELT &&
5680 Op0.getOperand(0) == Op1.getOperand(0) &&
5681 isa<ConstantSDNode>(Op0.getOperand(1)) &&
5682 isa<ConstantSDNode>(Op1.getOperand(1)));
5686 unsigned I0 = cast<ConstantSDNode>(Op0.getOperand(1))->getZExtValue();
5687 unsigned I1 = cast<ConstantSDNode>(Op1.getOperand(1))->getZExtValue();
5689 if (i * 2 < NumElts) {
5690 if (V0.getOpcode() == ISD::UNDEF) {
5691 V0 = Op0.getOperand(0);
5692 if (V0.getValueType() != VT)
5696 if (V1.getOpcode() == ISD::UNDEF) {
5697 V1 = Op0.getOperand(0);
5698 if (V1.getValueType() != VT)
5701 if (i * 2 == NumElts)
5702 ExpectedVExtractIdx = BaseIdx;
5705 SDValue Expected = (i * 2 < NumElts) ? V0 : V1;
5706 if (I0 == ExpectedVExtractIdx)
5707 CanFold = I1 == I0 + 1 && Op0.getOperand(0) == Expected;
5708 else if (IsCommutable && I1 == ExpectedVExtractIdx) {
5709 // Try to match the following dag sequence:
5710 // (BINOP (extract_vector_elt A, I+1), (extract_vector_elt A, I))
5711 CanFold = I0 == I1 + 1 && Op1.getOperand(0) == Expected;
5715 ExpectedVExtractIdx += 2;
5721 /// \brief Emit a sequence of two 128-bit horizontal add/sub followed by
5722 /// a concat_vector.
5724 /// This is a helper function of LowerToHorizontalOp().
5725 /// This function expects two 256-bit vectors called V0 and V1.
5726 /// At first, each vector is split into two separate 128-bit vectors.
5727 /// Then, the resulting 128-bit vectors are used to implement two
5728 /// horizontal binary operations.
5730 /// The kind of horizontal binary operation is defined by \p X86Opcode.
5732 /// \p Mode specifies how the 128-bit parts of V0 and V1 are passed in input to
5733 /// the two new horizontal binop.
5734 /// When Mode is set, the first horizontal binop dag node would take as input
5735 /// the lower 128-bit of V0 and the upper 128-bit of V0. The second
5736 /// horizontal binop dag node would take as input the lower 128-bit of V1
5737 /// and the upper 128-bit of V1.
5739 /// HADD V0_LO, V0_HI
5740 /// HADD V1_LO, V1_HI
5742 /// Otherwise, the first horizontal binop dag node takes as input the lower
5743 /// 128-bit of V0 and the lower 128-bit of V1, and the second horizontal binop
5744 /// dag node takes the upper 128-bit of V0 and the upper 128-bit of V1.
5746 /// HADD V0_LO, V1_LO
5747 /// HADD V0_HI, V1_HI
5749 /// If \p isUndefLO is set, then the algorithm propagates UNDEF to the lower
5750 /// 128-bits of the result. If \p isUndefHI is set, then UNDEF is propagated to
5751 /// the upper 128-bits of the result.
5752 static SDValue ExpandHorizontalBinOp(const SDValue &V0, const SDValue &V1,
5753 SDLoc DL, SelectionDAG &DAG,
5754 unsigned X86Opcode, bool Mode,
5755 bool isUndefLO, bool isUndefHI) {
5756 EVT VT = V0.getValueType();
5757 assert(VT.is256BitVector() && VT == V1.getValueType() &&
5758 "Invalid nodes in input!");
5760 unsigned NumElts = VT.getVectorNumElements();
5761 SDValue V0_LO = Extract128BitVector(V0, 0, DAG, DL);
5762 SDValue V0_HI = Extract128BitVector(V0, NumElts/2, DAG, DL);
5763 SDValue V1_LO = Extract128BitVector(V1, 0, DAG, DL);
5764 SDValue V1_HI = Extract128BitVector(V1, NumElts/2, DAG, DL);
5765 EVT NewVT = V0_LO.getValueType();
5767 SDValue LO = DAG.getUNDEF(NewVT);
5768 SDValue HI = DAG.getUNDEF(NewVT);
5771 // Don't emit a horizontal binop if the result is expected to be UNDEF.
5772 if (!isUndefLO && V0->getOpcode() != ISD::UNDEF)
5773 LO = DAG.getNode(X86Opcode, DL, NewVT, V0_LO, V0_HI);
5774 if (!isUndefHI && V1->getOpcode() != ISD::UNDEF)
5775 HI = DAG.getNode(X86Opcode, DL, NewVT, V1_LO, V1_HI);
5777 // Don't emit a horizontal binop if the result is expected to be UNDEF.
5778 if (!isUndefLO && (V0_LO->getOpcode() != ISD::UNDEF ||
5779 V1_LO->getOpcode() != ISD::UNDEF))
5780 LO = DAG.getNode(X86Opcode, DL, NewVT, V0_LO, V1_LO);
5782 if (!isUndefHI && (V0_HI->getOpcode() != ISD::UNDEF ||
5783 V1_HI->getOpcode() != ISD::UNDEF))
5784 HI = DAG.getNode(X86Opcode, DL, NewVT, V0_HI, V1_HI);
5787 return DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, LO, HI);
5790 /// Try to fold a build_vector that performs an 'addsub' to an X86ISD::ADDSUB
5792 static SDValue LowerToAddSub(const BuildVectorSDNode *BV,
5793 const X86Subtarget *Subtarget, SelectionDAG &DAG) {
5794 EVT VT = BV->getValueType(0);
5795 if ((!Subtarget->hasSSE3() || (VT != MVT::v4f32 && VT != MVT::v2f64)) &&
5796 (!Subtarget->hasAVX() || (VT != MVT::v8f32 && VT != MVT::v4f64)))
5800 unsigned NumElts = VT.getVectorNumElements();
5801 SDValue InVec0 = DAG.getUNDEF(VT);
5802 SDValue InVec1 = DAG.getUNDEF(VT);
5804 assert((VT == MVT::v8f32 || VT == MVT::v4f64 || VT == MVT::v4f32 ||
5805 VT == MVT::v2f64) && "build_vector with an invalid type found!");
5807 // Odd-numbered elements in the input build vector are obtained from
5808 // adding two integer/float elements.
5809 // Even-numbered elements in the input build vector are obtained from
5810 // subtracting two integer/float elements.
5811 unsigned ExpectedOpcode = ISD::FSUB;
5812 unsigned NextExpectedOpcode = ISD::FADD;
5813 bool AddFound = false;
5814 bool SubFound = false;
5816 for (unsigned i = 0, e = NumElts; i != e; ++i) {
5817 SDValue Op = BV->getOperand(i);
5819 // Skip 'undef' values.
5820 unsigned Opcode = Op.getOpcode();
5821 if (Opcode == ISD::UNDEF) {
5822 std::swap(ExpectedOpcode, NextExpectedOpcode);
5826 // Early exit if we found an unexpected opcode.
5827 if (Opcode != ExpectedOpcode)
5830 SDValue Op0 = Op.getOperand(0);
5831 SDValue Op1 = Op.getOperand(1);
5833 // Try to match the following pattern:
5834 // (BINOP (extract_vector_elt A, i), (extract_vector_elt B, i))
5835 // Early exit if we cannot match that sequence.
5836 if (Op0.getOpcode() != ISD::EXTRACT_VECTOR_ELT ||
5837 Op1.getOpcode() != ISD::EXTRACT_VECTOR_ELT ||
5838 !isa<ConstantSDNode>(Op0.getOperand(1)) ||
5839 !isa<ConstantSDNode>(Op1.getOperand(1)) ||
5840 Op0.getOperand(1) != Op1.getOperand(1))
5843 unsigned I0 = cast<ConstantSDNode>(Op0.getOperand(1))->getZExtValue();
5847 // We found a valid add/sub node. Update the information accordingly.
5853 // Update InVec0 and InVec1.
5854 if (InVec0.getOpcode() == ISD::UNDEF) {
5855 InVec0 = Op0.getOperand(0);
5856 if (InVec0.getValueType() != VT)
5859 if (InVec1.getOpcode() == ISD::UNDEF) {
5860 InVec1 = Op1.getOperand(0);
5861 if (InVec1.getValueType() != VT)
5865 // Make sure that operands in input to each add/sub node always
5866 // come from a same pair of vectors.
5867 if (InVec0 != Op0.getOperand(0)) {
5868 if (ExpectedOpcode == ISD::FSUB)
5871 // FADD is commutable. Try to commute the operands
5872 // and then test again.
5873 std::swap(Op0, Op1);
5874 if (InVec0 != Op0.getOperand(0))
5878 if (InVec1 != Op1.getOperand(0))
5881 // Update the pair of expected opcodes.
5882 std::swap(ExpectedOpcode, NextExpectedOpcode);
5885 // Don't try to fold this build_vector into an ADDSUB if the inputs are undef.
5886 if (AddFound && SubFound && InVec0.getOpcode() != ISD::UNDEF &&
5887 InVec1.getOpcode() != ISD::UNDEF)
5888 return DAG.getNode(X86ISD::ADDSUB, DL, VT, InVec0, InVec1);
5893 /// Lower BUILD_VECTOR to a horizontal add/sub operation if possible.
5894 static SDValue LowerToHorizontalOp(const BuildVectorSDNode *BV,
5895 const X86Subtarget *Subtarget,
5896 SelectionDAG &DAG) {
5897 EVT VT = BV->getValueType(0);
5898 unsigned NumElts = VT.getVectorNumElements();
5899 unsigned NumUndefsLO = 0;
5900 unsigned NumUndefsHI = 0;
5901 unsigned Half = NumElts/2;
5903 // Count the number of UNDEF operands in the build_vector in input.
5904 for (unsigned i = 0, e = Half; i != e; ++i)
5905 if (BV->getOperand(i)->getOpcode() == ISD::UNDEF)
5908 for (unsigned i = Half, e = NumElts; i != e; ++i)
5909 if (BV->getOperand(i)->getOpcode() == ISD::UNDEF)
5912 // Early exit if this is either a build_vector of all UNDEFs or all the
5913 // operands but one are UNDEF.
5914 if (NumUndefsLO + NumUndefsHI + 1 >= NumElts)
5918 SDValue InVec0, InVec1;
5919 if ((VT == MVT::v4f32 || VT == MVT::v2f64) && Subtarget->hasSSE3()) {
5920 // Try to match an SSE3 float HADD/HSUB.
5921 if (isHorizontalBinOp(BV, ISD::FADD, DAG, 0, NumElts, InVec0, InVec1))
5922 return DAG.getNode(X86ISD::FHADD, DL, VT, InVec0, InVec1);
5924 if (isHorizontalBinOp(BV, ISD::FSUB, DAG, 0, NumElts, InVec0, InVec1))
5925 return DAG.getNode(X86ISD::FHSUB, DL, VT, InVec0, InVec1);
5926 } else if ((VT == MVT::v4i32 || VT == MVT::v8i16) && Subtarget->hasSSSE3()) {
5927 // Try to match an SSSE3 integer HADD/HSUB.
5928 if (isHorizontalBinOp(BV, ISD::ADD, DAG, 0, NumElts, InVec0, InVec1))
5929 return DAG.getNode(X86ISD::HADD, DL, VT, InVec0, InVec1);
5931 if (isHorizontalBinOp(BV, ISD::SUB, DAG, 0, NumElts, InVec0, InVec1))
5932 return DAG.getNode(X86ISD::HSUB, DL, VT, InVec0, InVec1);
5935 if (!Subtarget->hasAVX())
5938 if ((VT == MVT::v8f32 || VT == MVT::v4f64)) {
5939 // Try to match an AVX horizontal add/sub of packed single/double
5940 // precision floating point values from 256-bit vectors.
5941 SDValue InVec2, InVec3;
5942 if (isHorizontalBinOp(BV, ISD::FADD, DAG, 0, Half, InVec0, InVec1) &&
5943 isHorizontalBinOp(BV, ISD::FADD, DAG, Half, NumElts, InVec2, InVec3) &&
5944 ((InVec0.getOpcode() == ISD::UNDEF ||
5945 InVec2.getOpcode() == ISD::UNDEF) || InVec0 == InVec2) &&
5946 ((InVec1.getOpcode() == ISD::UNDEF ||
5947 InVec3.getOpcode() == ISD::UNDEF) || InVec1 == InVec3))
5948 return DAG.getNode(X86ISD::FHADD, DL, VT, InVec0, InVec1);
5950 if (isHorizontalBinOp(BV, ISD::FSUB, DAG, 0, Half, InVec0, InVec1) &&
5951 isHorizontalBinOp(BV, ISD::FSUB, DAG, Half, NumElts, InVec2, InVec3) &&
5952 ((InVec0.getOpcode() == ISD::UNDEF ||
5953 InVec2.getOpcode() == ISD::UNDEF) || InVec0 == InVec2) &&
5954 ((InVec1.getOpcode() == ISD::UNDEF ||
5955 InVec3.getOpcode() == ISD::UNDEF) || InVec1 == InVec3))
5956 return DAG.getNode(X86ISD::FHSUB, DL, VT, InVec0, InVec1);
5957 } else if (VT == MVT::v8i32 || VT == MVT::v16i16) {
5958 // Try to match an AVX2 horizontal add/sub of signed integers.
5959 SDValue InVec2, InVec3;
5961 bool CanFold = true;
5963 if (isHorizontalBinOp(BV, ISD::ADD, DAG, 0, Half, InVec0, InVec1) &&
5964 isHorizontalBinOp(BV, ISD::ADD, DAG, Half, NumElts, InVec2, InVec3) &&
5965 ((InVec0.getOpcode() == ISD::UNDEF ||
5966 InVec2.getOpcode() == ISD::UNDEF) || InVec0 == InVec2) &&
5967 ((InVec1.getOpcode() == ISD::UNDEF ||
5968 InVec3.getOpcode() == ISD::UNDEF) || InVec1 == InVec3))
5969 X86Opcode = X86ISD::HADD;
5970 else if (isHorizontalBinOp(BV, ISD::SUB, DAG, 0, Half, InVec0, InVec1) &&
5971 isHorizontalBinOp(BV, ISD::SUB, DAG, Half, NumElts, InVec2, InVec3) &&
5972 ((InVec0.getOpcode() == ISD::UNDEF ||
5973 InVec2.getOpcode() == ISD::UNDEF) || InVec0 == InVec2) &&
5974 ((InVec1.getOpcode() == ISD::UNDEF ||
5975 InVec3.getOpcode() == ISD::UNDEF) || InVec1 == InVec3))
5976 X86Opcode = X86ISD::HSUB;
5981 // Fold this build_vector into a single horizontal add/sub.
5982 // Do this only if the target has AVX2.
5983 if (Subtarget->hasAVX2())
5984 return DAG.getNode(X86Opcode, DL, VT, InVec0, InVec1);
5986 // Do not try to expand this build_vector into a pair of horizontal
5987 // add/sub if we can emit a pair of scalar add/sub.
5988 if (NumUndefsLO + 1 == Half || NumUndefsHI + 1 == Half)
5991 // Convert this build_vector into a pair of horizontal binop followed by
5993 bool isUndefLO = NumUndefsLO == Half;
5994 bool isUndefHI = NumUndefsHI == Half;
5995 return ExpandHorizontalBinOp(InVec0, InVec1, DL, DAG, X86Opcode, false,
5996 isUndefLO, isUndefHI);
6000 if ((VT == MVT::v8f32 || VT == MVT::v4f64 || VT == MVT::v8i32 ||
6001 VT == MVT::v16i16) && Subtarget->hasAVX()) {
6003 if (isHorizontalBinOp(BV, ISD::ADD, DAG, 0, NumElts, InVec0, InVec1))
6004 X86Opcode = X86ISD::HADD;
6005 else if (isHorizontalBinOp(BV, ISD::SUB, DAG, 0, NumElts, InVec0, InVec1))
6006 X86Opcode = X86ISD::HSUB;
6007 else if (isHorizontalBinOp(BV, ISD::FADD, DAG, 0, NumElts, InVec0, InVec1))
6008 X86Opcode = X86ISD::FHADD;
6009 else if (isHorizontalBinOp(BV, ISD::FSUB, DAG, 0, NumElts, InVec0, InVec1))
6010 X86Opcode = X86ISD::FHSUB;
6014 // Don't try to expand this build_vector into a pair of horizontal add/sub
6015 // if we can simply emit a pair of scalar add/sub.
6016 if (NumUndefsLO + 1 == Half || NumUndefsHI + 1 == Half)
6019 // Convert this build_vector into two horizontal add/sub followed by
6021 bool isUndefLO = NumUndefsLO == Half;
6022 bool isUndefHI = NumUndefsHI == Half;
6023 return ExpandHorizontalBinOp(InVec0, InVec1, DL, DAG, X86Opcode, true,
6024 isUndefLO, isUndefHI);
6031 X86TargetLowering::LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG) const {
6034 MVT VT = Op.getSimpleValueType();
6035 MVT ExtVT = VT.getVectorElementType();
6036 unsigned NumElems = Op.getNumOperands();
6038 // Generate vectors for predicate vectors.
6039 if (VT.getScalarType() == MVT::i1 && Subtarget->hasAVX512())
6040 return LowerBUILD_VECTORvXi1(Op, DAG);
6042 // Vectors containing all zeros can be matched by pxor and xorps later
6043 if (ISD::isBuildVectorAllZeros(Op.getNode())) {
6044 // Canonicalize this to <4 x i32> to 1) ensure the zero vectors are CSE'd
6045 // and 2) ensure that i64 scalars are eliminated on x86-32 hosts.
6046 if (VT == MVT::v4i32 || VT == MVT::v8i32 || VT == MVT::v16i32)
6049 return getZeroVector(VT, Subtarget, DAG, dl);
6052 // Vectors containing all ones can be matched by pcmpeqd on 128-bit width
6053 // vectors or broken into v4i32 operations on 256-bit vectors. AVX2 can use
6054 // vpcmpeqd on 256-bit vectors.
6055 if (Subtarget->hasSSE2() && ISD::isBuildVectorAllOnes(Op.getNode())) {
6056 if (VT == MVT::v4i32 || (VT == MVT::v8i32 && Subtarget->hasInt256()))
6059 if (!VT.is512BitVector())
6060 return getOnesVector(VT, Subtarget, DAG, dl);
6063 BuildVectorSDNode *BV = cast<BuildVectorSDNode>(Op.getNode());
6064 if (SDValue AddSub = LowerToAddSub(BV, Subtarget, DAG))
6066 if (SDValue HorizontalOp = LowerToHorizontalOp(BV, Subtarget, DAG))
6067 return HorizontalOp;
6068 if (SDValue Broadcast = LowerVectorBroadcast(Op, Subtarget, DAG))
6071 unsigned EVTBits = ExtVT.getSizeInBits();
6073 unsigned NumZero = 0;
6074 unsigned NumNonZero = 0;
6075 unsigned NonZeros = 0;
6076 bool IsAllConstants = true;
6077 SmallSet<SDValue, 8> Values;
6078 for (unsigned i = 0; i < NumElems; ++i) {
6079 SDValue Elt = Op.getOperand(i);
6080 if (Elt.getOpcode() == ISD::UNDEF)
6083 if (Elt.getOpcode() != ISD::Constant &&
6084 Elt.getOpcode() != ISD::ConstantFP)
6085 IsAllConstants = false;
6086 if (X86::isZeroNode(Elt))
6089 NonZeros |= (1 << i);
6094 // All undef vector. Return an UNDEF. All zero vectors were handled above.
6095 if (NumNonZero == 0)
6096 return DAG.getUNDEF(VT);
6098 // Special case for single non-zero, non-undef, element.
6099 if (NumNonZero == 1) {
6100 unsigned Idx = countTrailingZeros(NonZeros);
6101 SDValue Item = Op.getOperand(Idx);
6103 // If this is an insertion of an i64 value on x86-32, and if the top bits of
6104 // the value are obviously zero, truncate the value to i32 and do the
6105 // insertion that way. Only do this if the value is non-constant or if the
6106 // value is a constant being inserted into element 0. It is cheaper to do
6107 // a constant pool load than it is to do a movd + shuffle.
6108 if (ExtVT == MVT::i64 && !Subtarget->is64Bit() &&
6109 (!IsAllConstants || Idx == 0)) {
6110 if (DAG.MaskedValueIsZero(Item, APInt::getBitsSet(64, 32, 64))) {
6112 assert(VT == MVT::v2i64 && "Expected an SSE value type!");
6113 EVT VecVT = MVT::v4i32;
6115 // Truncate the value (which may itself be a constant) to i32, and
6116 // convert it to a vector with movd (S2V+shuffle to zero extend).
6117 Item = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, Item);
6118 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VecVT, Item);
6119 return DAG.getBitcast(VT, getShuffleVectorZeroOrUndef(
6120 Item, Idx * 2, true, Subtarget, DAG));
6124 // If we have a constant or non-constant insertion into the low element of
6125 // a vector, we can do this with SCALAR_TO_VECTOR + shuffle of zero into
6126 // the rest of the elements. This will be matched as movd/movq/movss/movsd
6127 // depending on what the source datatype is.
6130 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
6132 if (ExtVT == MVT::i32 || ExtVT == MVT::f32 || ExtVT == MVT::f64 ||
6133 (ExtVT == MVT::i64 && Subtarget->is64Bit())) {
6134 if (VT.is512BitVector()) {
6135 SDValue ZeroVec = getZeroVector(VT, Subtarget, DAG, dl);
6136 return DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, ZeroVec,
6137 Item, DAG.getIntPtrConstant(0, dl));
6139 assert((VT.is128BitVector() || VT.is256BitVector()) &&
6140 "Expected an SSE value type!");
6141 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
6142 // Turn it into a MOVL (i.e. movss, movsd, or movd) to a zero vector.
6143 return getShuffleVectorZeroOrUndef(Item, 0, true, Subtarget, DAG);
6146 // We can't directly insert an i8 or i16 into a vector, so zero extend
6148 if (ExtVT == MVT::i16 || ExtVT == MVT::i8) {
6149 Item = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, Item);
6150 if (VT.is256BitVector()) {
6151 if (Subtarget->hasAVX()) {
6152 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v8i32, Item);
6153 Item = getShuffleVectorZeroOrUndef(Item, 0, true, Subtarget, DAG);
6155 // Without AVX, we need to extend to a 128-bit vector and then
6156 // insert into the 256-bit vector.
6157 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32, Item);
6158 SDValue ZeroVec = getZeroVector(MVT::v8i32, Subtarget, DAG, dl);
6159 Item = Insert128BitVector(ZeroVec, Item, 0, DAG, dl);
6162 assert(VT.is128BitVector() && "Expected an SSE value type!");
6163 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32, Item);
6164 Item = getShuffleVectorZeroOrUndef(Item, 0, true, Subtarget, DAG);
6166 return DAG.getBitcast(VT, Item);
6170 // Is it a vector logical left shift?
6171 if (NumElems == 2 && Idx == 1 &&
6172 X86::isZeroNode(Op.getOperand(0)) &&
6173 !X86::isZeroNode(Op.getOperand(1))) {
6174 unsigned NumBits = VT.getSizeInBits();
6175 return getVShift(true, VT,
6176 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
6177 VT, Op.getOperand(1)),
6178 NumBits/2, DAG, *this, dl);
6181 if (IsAllConstants) // Otherwise, it's better to do a constpool load.
6184 // Otherwise, if this is a vector with i32 or f32 elements, and the element
6185 // is a non-constant being inserted into an element other than the low one,
6186 // we can't use a constant pool load. Instead, use SCALAR_TO_VECTOR (aka
6187 // movd/movss) to move this into the low element, then shuffle it into
6189 if (EVTBits == 32) {
6190 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
6191 return getShuffleVectorZeroOrUndef(Item, Idx, NumZero > 0, Subtarget, DAG);
6195 // Splat is obviously ok. Let legalizer expand it to a shuffle.
6196 if (Values.size() == 1) {
6197 if (EVTBits == 32) {
6198 // Instead of a shuffle like this:
6199 // shuffle (scalar_to_vector (load (ptr + 4))), undef, <0, 0, 0, 0>
6200 // Check if it's possible to issue this instead.
6201 // shuffle (vload ptr)), undef, <1, 1, 1, 1>
6202 unsigned Idx = countTrailingZeros(NonZeros);
6203 SDValue Item = Op.getOperand(Idx);
6204 if (Op.getNode()->isOnlyUserOf(Item.getNode()))
6205 return LowerAsSplatVectorLoad(Item, VT, dl, DAG);
6210 // A vector full of immediates; various special cases are already
6211 // handled, so this is best done with a single constant-pool load.
6215 // For AVX-length vectors, see if we can use a vector load to get all of the
6216 // elements, otherwise build the individual 128-bit pieces and use
6217 // shuffles to put them in place.
6218 if (VT.is256BitVector() || VT.is512BitVector()) {
6219 SmallVector<SDValue, 64> V(Op->op_begin(), Op->op_begin() + NumElems);
6221 // Check for a build vector of consecutive loads.
6222 if (SDValue LD = EltsFromConsecutiveLoads(VT, V, dl, DAG, false))
6225 EVT HVT = EVT::getVectorVT(*DAG.getContext(), ExtVT, NumElems/2);
6227 // Build both the lower and upper subvector.
6228 SDValue Lower = DAG.getNode(ISD::BUILD_VECTOR, dl, HVT,
6229 makeArrayRef(&V[0], NumElems/2));
6230 SDValue Upper = DAG.getNode(ISD::BUILD_VECTOR, dl, HVT,
6231 makeArrayRef(&V[NumElems / 2], NumElems/2));
6233 // Recreate the wider vector with the lower and upper part.
6234 if (VT.is256BitVector())
6235 return Concat128BitVectors(Lower, Upper, VT, NumElems, DAG, dl);
6236 return Concat256BitVectors(Lower, Upper, VT, NumElems, DAG, dl);
6239 // Let legalizer expand 2-wide build_vectors.
6240 if (EVTBits == 64) {
6241 if (NumNonZero == 1) {
6242 // One half is zero or undef.
6243 unsigned Idx = countTrailingZeros(NonZeros);
6244 SDValue V2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT,
6245 Op.getOperand(Idx));
6246 return getShuffleVectorZeroOrUndef(V2, Idx, true, Subtarget, DAG);
6251 // If element VT is < 32 bits, convert it to inserts into a zero vector.
6252 if (EVTBits == 8 && NumElems == 16)
6253 if (SDValue V = LowerBuildVectorv16i8(Op, NonZeros,NumNonZero,NumZero, DAG,
6257 if (EVTBits == 16 && NumElems == 8)
6258 if (SDValue V = LowerBuildVectorv8i16(Op, NonZeros,NumNonZero,NumZero, DAG,
6262 // If element VT is == 32 bits and has 4 elems, try to generate an INSERTPS
6263 if (EVTBits == 32 && NumElems == 4)
6264 if (SDValue V = LowerBuildVectorv4x32(Op, DAG, Subtarget, *this))
6267 // If element VT is == 32 bits, turn it into a number of shuffles.
6268 SmallVector<SDValue, 8> V(NumElems);
6269 if (NumElems == 4 && NumZero > 0) {
6270 for (unsigned i = 0; i < 4; ++i) {
6271 bool isZero = !(NonZeros & (1 << i));
6273 V[i] = getZeroVector(VT, Subtarget, DAG, dl);
6275 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(i));
6278 for (unsigned i = 0; i < 2; ++i) {
6279 switch ((NonZeros & (0x3 << i*2)) >> (i*2)) {
6282 V[i] = V[i*2]; // Must be a zero vector.
6285 V[i] = getMOVL(DAG, dl, VT, V[i*2+1], V[i*2]);
6288 V[i] = getMOVL(DAG, dl, VT, V[i*2], V[i*2+1]);
6291 V[i] = getUnpackl(DAG, dl, VT, V[i*2], V[i*2+1]);
6296 bool Reverse1 = (NonZeros & 0x3) == 2;
6297 bool Reverse2 = ((NonZeros & (0x3 << 2)) >> 2) == 2;
6301 static_cast<int>(Reverse2 ? NumElems+1 : NumElems),
6302 static_cast<int>(Reverse2 ? NumElems : NumElems+1)
6304 return DAG.getVectorShuffle(VT, dl, V[0], V[1], &MaskVec[0]);
6307 if (Values.size() > 1 && VT.is128BitVector()) {
6308 // Check for a build vector of consecutive loads.
6309 for (unsigned i = 0; i < NumElems; ++i)
6310 V[i] = Op.getOperand(i);
6312 // Check for elements which are consecutive loads.
6313 if (SDValue LD = EltsFromConsecutiveLoads(VT, V, dl, DAG, false))
6316 // Check for a build vector from mostly shuffle plus few inserting.
6317 if (SDValue Sh = buildFromShuffleMostly(Op, DAG))
6320 // For SSE 4.1, use insertps to put the high elements into the low element.
6321 if (Subtarget->hasSSE41()) {
6323 if (Op.getOperand(0).getOpcode() != ISD::UNDEF)
6324 Result = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(0));
6326 Result = DAG.getUNDEF(VT);
6328 for (unsigned i = 1; i < NumElems; ++i) {
6329 if (Op.getOperand(i).getOpcode() == ISD::UNDEF) continue;
6330 Result = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, Result,
6331 Op.getOperand(i), DAG.getIntPtrConstant(i, dl));
6336 // Otherwise, expand into a number of unpckl*, start by extending each of
6337 // our (non-undef) elements to the full vector width with the element in the
6338 // bottom slot of the vector (which generates no code for SSE).
6339 for (unsigned i = 0; i < NumElems; ++i) {
6340 if (Op.getOperand(i).getOpcode() != ISD::UNDEF)
6341 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(i));
6343 V[i] = DAG.getUNDEF(VT);
6346 // Next, we iteratively mix elements, e.g. for v4f32:
6347 // Step 1: unpcklps 0, 2 ==> X: <?, ?, 2, 0>
6348 // : unpcklps 1, 3 ==> Y: <?, ?, 3, 1>
6349 // Step 2: unpcklps X, Y ==> <3, 2, 1, 0>
6350 unsigned EltStride = NumElems >> 1;
6351 while (EltStride != 0) {
6352 for (unsigned i = 0; i < EltStride; ++i) {
6353 // If V[i+EltStride] is undef and this is the first round of mixing,
6354 // then it is safe to just drop this shuffle: V[i] is already in the
6355 // right place, the one element (since it's the first round) being
6356 // inserted as undef can be dropped. This isn't safe for successive
6357 // rounds because they will permute elements within both vectors.
6358 if (V[i+EltStride].getOpcode() == ISD::UNDEF &&
6359 EltStride == NumElems/2)
6362 V[i] = getUnpackl(DAG, dl, VT, V[i], V[i + EltStride]);
6371 // 256-bit AVX can use the vinsertf128 instruction
6372 // to create 256-bit vectors from two other 128-bit ones.
6373 static SDValue LowerAVXCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) {
6375 MVT ResVT = Op.getSimpleValueType();
6377 assert((ResVT.is256BitVector() ||
6378 ResVT.is512BitVector()) && "Value type must be 256-/512-bit wide");
6380 SDValue V1 = Op.getOperand(0);
6381 SDValue V2 = Op.getOperand(1);
6382 unsigned NumElems = ResVT.getVectorNumElements();
6383 if (ResVT.is256BitVector())
6384 return Concat128BitVectors(V1, V2, ResVT, NumElems, DAG, dl);
6386 if (Op.getNumOperands() == 4) {
6387 MVT HalfVT = MVT::getVectorVT(ResVT.getScalarType(),
6388 ResVT.getVectorNumElements()/2);
6389 SDValue V3 = Op.getOperand(2);
6390 SDValue V4 = Op.getOperand(3);
6391 return Concat256BitVectors(Concat128BitVectors(V1, V2, HalfVT, NumElems/2, DAG, dl),
6392 Concat128BitVectors(V3, V4, HalfVT, NumElems/2, DAG, dl), ResVT, NumElems, DAG, dl);
6394 return Concat256BitVectors(V1, V2, ResVT, NumElems, DAG, dl);
6397 static SDValue LowerCONCAT_VECTORSvXi1(SDValue Op,
6398 const X86Subtarget *Subtarget,
6399 SelectionDAG & DAG) {
6401 MVT ResVT = Op.getSimpleValueType();
6402 unsigned NumOfOperands = Op.getNumOperands();
6404 assert(isPowerOf2_32(NumOfOperands) &&
6405 "Unexpected number of operands in CONCAT_VECTORS");
6407 if (NumOfOperands > 2) {
6408 MVT HalfVT = MVT::getVectorVT(ResVT.getScalarType(),
6409 ResVT.getVectorNumElements()/2);
6410 SmallVector<SDValue, 2> Ops;
6411 for (unsigned i = 0; i < NumOfOperands/2; i++)
6412 Ops.push_back(Op.getOperand(i));
6413 SDValue Lo = DAG.getNode(ISD::CONCAT_VECTORS, dl, HalfVT, Ops);
6415 for (unsigned i = NumOfOperands/2; i < NumOfOperands; i++)
6416 Ops.push_back(Op.getOperand(i));
6417 SDValue Hi = DAG.getNode(ISD::CONCAT_VECTORS, dl, HalfVT, Ops);
6418 return DAG.getNode(ISD::CONCAT_VECTORS, dl, ResVT, Lo, Hi);
6421 SDValue V1 = Op.getOperand(0);
6422 SDValue V2 = Op.getOperand(1);
6423 bool IsZeroV1 = ISD::isBuildVectorAllZeros(V1.getNode());
6424 bool IsZeroV2 = ISD::isBuildVectorAllZeros(V2.getNode());
6426 if (IsZeroV1 && IsZeroV2)
6427 return getZeroVector(ResVT, Subtarget, DAG, dl);
6429 SDValue ZeroIdx = DAG.getIntPtrConstant(0, dl);
6430 SDValue Undef = DAG.getUNDEF(ResVT);
6431 unsigned NumElems = ResVT.getVectorNumElements();
6432 SDValue ShiftBits = DAG.getConstant(NumElems/2, dl, MVT::i8);
6434 V2 = DAG.getNode(ISD::INSERT_SUBVECTOR, dl, ResVT, Undef, V2, ZeroIdx);
6435 V2 = DAG.getNode(X86ISD::VSHLI, dl, ResVT, V2, ShiftBits);
6439 V1 = DAG.getNode(ISD::INSERT_SUBVECTOR, dl, ResVT, Undef, V1, ZeroIdx);
6440 // Zero the upper bits of V1
6441 V1 = DAG.getNode(X86ISD::VSHLI, dl, ResVT, V1, ShiftBits);
6442 V1 = DAG.getNode(X86ISD::VSRLI, dl, ResVT, V1, ShiftBits);
6445 return DAG.getNode(ISD::OR, dl, ResVT, V1, V2);
6448 static SDValue LowerCONCAT_VECTORS(SDValue Op,
6449 const X86Subtarget *Subtarget,
6450 SelectionDAG &DAG) {
6451 MVT VT = Op.getSimpleValueType();
6452 if (VT.getVectorElementType() == MVT::i1)
6453 return LowerCONCAT_VECTORSvXi1(Op, Subtarget, DAG);
6455 assert((VT.is256BitVector() && Op.getNumOperands() == 2) ||
6456 (VT.is512BitVector() && (Op.getNumOperands() == 2 ||
6457 Op.getNumOperands() == 4)));
6459 // AVX can use the vinsertf128 instruction to create 256-bit vectors
6460 // from two other 128-bit ones.
6462 // 512-bit vector may contain 2 256-bit vectors or 4 128-bit vectors
6463 return LowerAVXCONCAT_VECTORS(Op, DAG);
6467 //===----------------------------------------------------------------------===//
6468 // Vector shuffle lowering
6470 // This is an experimental code path for lowering vector shuffles on x86. It is
6471 // designed to handle arbitrary vector shuffles and blends, gracefully
6472 // degrading performance as necessary. It works hard to recognize idiomatic
6473 // shuffles and lower them to optimal instruction patterns without leaving
6474 // a framework that allows reasonably efficient handling of all vector shuffle
6476 //===----------------------------------------------------------------------===//
6478 /// \brief Tiny helper function to identify a no-op mask.
6480 /// This is a somewhat boring predicate function. It checks whether the mask
6481 /// array input, which is assumed to be a single-input shuffle mask of the kind
6482 /// used by the X86 shuffle instructions (not a fully general
6483 /// ShuffleVectorSDNode mask) requires any shuffles to occur. Both undef and an
6484 /// in-place shuffle are 'no-op's.
6485 static bool isNoopShuffleMask(ArrayRef<int> Mask) {
6486 for (int i = 0, Size = Mask.size(); i < Size; ++i)
6487 if (Mask[i] != -1 && Mask[i] != i)
6492 /// \brief Helper function to classify a mask as a single-input mask.
6494 /// This isn't a generic single-input test because in the vector shuffle
6495 /// lowering we canonicalize single inputs to be the first input operand. This
6496 /// means we can more quickly test for a single input by only checking whether
6497 /// an input from the second operand exists. We also assume that the size of
6498 /// mask corresponds to the size of the input vectors which isn't true in the
6499 /// fully general case.
6500 static bool isSingleInputShuffleMask(ArrayRef<int> Mask) {
6502 if (M >= (int)Mask.size())
6507 /// \brief Test whether there are elements crossing 128-bit lanes in this
6510 /// X86 divides up its shuffles into in-lane and cross-lane shuffle operations
6511 /// and we routinely test for these.
6512 static bool is128BitLaneCrossingShuffleMask(MVT VT, ArrayRef<int> Mask) {
6513 int LaneSize = 128 / VT.getScalarSizeInBits();
6514 int Size = Mask.size();
6515 for (int i = 0; i < Size; ++i)
6516 if (Mask[i] >= 0 && (Mask[i] % Size) / LaneSize != i / LaneSize)
6521 /// \brief Test whether a shuffle mask is equivalent within each 128-bit lane.
6523 /// This checks a shuffle mask to see if it is performing the same
6524 /// 128-bit lane-relative shuffle in each 128-bit lane. This trivially implies
6525 /// that it is also not lane-crossing. It may however involve a blend from the
6526 /// same lane of a second vector.
6528 /// The specific repeated shuffle mask is populated in \p RepeatedMask, as it is
6529 /// non-trivial to compute in the face of undef lanes. The representation is
6530 /// *not* suitable for use with existing 128-bit shuffles as it will contain
6531 /// entries from both V1 and V2 inputs to the wider mask.
6533 is128BitLaneRepeatedShuffleMask(MVT VT, ArrayRef<int> Mask,
6534 SmallVectorImpl<int> &RepeatedMask) {
6535 int LaneSize = 128 / VT.getScalarSizeInBits();
6536 RepeatedMask.resize(LaneSize, -1);
6537 int Size = Mask.size();
6538 for (int i = 0; i < Size; ++i) {
6541 if ((Mask[i] % Size) / LaneSize != i / LaneSize)
6542 // This entry crosses lanes, so there is no way to model this shuffle.
6545 // Ok, handle the in-lane shuffles by detecting if and when they repeat.
6546 if (RepeatedMask[i % LaneSize] == -1)
6547 // This is the first non-undef entry in this slot of a 128-bit lane.
6548 RepeatedMask[i % LaneSize] =
6549 Mask[i] < Size ? Mask[i] % LaneSize : Mask[i] % LaneSize + Size;
6550 else if (RepeatedMask[i % LaneSize] + (i / LaneSize) * LaneSize != Mask[i])
6551 // Found a mismatch with the repeated mask.
6557 /// \brief Checks whether a shuffle mask is equivalent to an explicit list of
6560 /// This is a fast way to test a shuffle mask against a fixed pattern:
6562 /// if (isShuffleEquivalent(Mask, 3, 2, {1, 0})) { ... }
6564 /// It returns true if the mask is exactly as wide as the argument list, and
6565 /// each element of the mask is either -1 (signifying undef) or the value given
6566 /// in the argument.
6567 static bool isShuffleEquivalent(SDValue V1, SDValue V2, ArrayRef<int> Mask,
6568 ArrayRef<int> ExpectedMask) {
6569 if (Mask.size() != ExpectedMask.size())
6572 int Size = Mask.size();
6574 // If the values are build vectors, we can look through them to find
6575 // equivalent inputs that make the shuffles equivalent.
6576 auto *BV1 = dyn_cast<BuildVectorSDNode>(V1);
6577 auto *BV2 = dyn_cast<BuildVectorSDNode>(V2);
6579 for (int i = 0; i < Size; ++i)
6580 if (Mask[i] != -1 && Mask[i] != ExpectedMask[i]) {
6581 auto *MaskBV = Mask[i] < Size ? BV1 : BV2;
6582 auto *ExpectedBV = ExpectedMask[i] < Size ? BV1 : BV2;
6583 if (!MaskBV || !ExpectedBV ||
6584 MaskBV->getOperand(Mask[i] % Size) !=
6585 ExpectedBV->getOperand(ExpectedMask[i] % Size))
6592 /// \brief Get a 4-lane 8-bit shuffle immediate for a mask.
6594 /// This helper function produces an 8-bit shuffle immediate corresponding to
6595 /// the ubiquitous shuffle encoding scheme used in x86 instructions for
6596 /// shuffling 4 lanes. It can be used with most of the PSHUF instructions for
6599 /// NB: We rely heavily on "undef" masks preserving the input lane.
6600 static SDValue getV4X86ShuffleImm8ForMask(ArrayRef<int> Mask, SDLoc DL,
6601 SelectionDAG &DAG) {
6602 assert(Mask.size() == 4 && "Only 4-lane shuffle masks");
6603 assert(Mask[0] >= -1 && Mask[0] < 4 && "Out of bound mask element!");
6604 assert(Mask[1] >= -1 && Mask[1] < 4 && "Out of bound mask element!");
6605 assert(Mask[2] >= -1 && Mask[2] < 4 && "Out of bound mask element!");
6606 assert(Mask[3] >= -1 && Mask[3] < 4 && "Out of bound mask element!");
6609 Imm |= (Mask[0] == -1 ? 0 : Mask[0]) << 0;
6610 Imm |= (Mask[1] == -1 ? 1 : Mask[1]) << 2;
6611 Imm |= (Mask[2] == -1 ? 2 : Mask[2]) << 4;
6612 Imm |= (Mask[3] == -1 ? 3 : Mask[3]) << 6;
6613 return DAG.getConstant(Imm, DL, MVT::i8);
6616 /// \brief Compute whether each element of a shuffle is zeroable.
6618 /// A "zeroable" vector shuffle element is one which can be lowered to zero.
6619 /// Either it is an undef element in the shuffle mask, the element of the input
6620 /// referenced is undef, or the element of the input referenced is known to be
6621 /// zero. Many x86 shuffles can zero lanes cheaply and we often want to handle
6622 /// as many lanes with this technique as possible to simplify the remaining
6624 static SmallBitVector computeZeroableShuffleElements(ArrayRef<int> Mask,
6625 SDValue V1, SDValue V2) {
6626 SmallBitVector Zeroable(Mask.size(), false);
6628 while (V1.getOpcode() == ISD::BITCAST)
6629 V1 = V1->getOperand(0);
6630 while (V2.getOpcode() == ISD::BITCAST)
6631 V2 = V2->getOperand(0);
6633 bool V1IsZero = ISD::isBuildVectorAllZeros(V1.getNode());
6634 bool V2IsZero = ISD::isBuildVectorAllZeros(V2.getNode());
6636 for (int i = 0, Size = Mask.size(); i < Size; ++i) {
6638 // Handle the easy cases.
6639 if (M < 0 || (M >= 0 && M < Size && V1IsZero) || (M >= Size && V2IsZero)) {
6644 // If this is an index into a build_vector node (which has the same number
6645 // of elements), dig out the input value and use it.
6646 SDValue V = M < Size ? V1 : V2;
6647 if (V.getOpcode() != ISD::BUILD_VECTOR || Size != (int)V.getNumOperands())
6650 SDValue Input = V.getOperand(M % Size);
6651 // The UNDEF opcode check really should be dead code here, but not quite
6652 // worth asserting on (it isn't invalid, just unexpected).
6653 if (Input.getOpcode() == ISD::UNDEF || X86::isZeroNode(Input))
6660 // X86 has dedicated unpack instructions that can handle specific blend
6661 // operations: UNPCKH and UNPCKL.
6662 static SDValue lowerVectorShuffleWithUNPCK(SDLoc DL, MVT VT, ArrayRef<int> Mask,
6663 SDValue V1, SDValue V2,
6664 SelectionDAG &DAG) {
6665 int NumElts = VT.getVectorNumElements();
6668 bool UnpcklSwapped = true;
6669 bool UnpckhSwapped = true;
6670 int NumEltsInLane = 128 / VT.getScalarSizeInBits();
6672 for (int i = 0; i < NumElts; ++i) {
6673 unsigned LaneStart = (i / NumEltsInLane) * NumEltsInLane;
6675 int LoPos = (i % NumEltsInLane) / 2 + LaneStart + NumElts * (i % 2);
6676 int HiPos = LoPos + NumEltsInLane / 2;
6677 int LoPosSwapped = (LoPos + NumElts) % (NumElts * 2);
6678 int HiPosSwapped = (HiPos + NumElts) % (NumElts * 2);
6682 if (Mask[i] != LoPos)
6684 if (Mask[i] != HiPos)
6686 if (Mask[i] != LoPosSwapped)
6687 UnpcklSwapped = false;
6688 if (Mask[i] != HiPosSwapped)
6689 UnpckhSwapped = false;
6690 if (!Unpckl && !Unpckh && !UnpcklSwapped && !UnpckhSwapped)
6694 return DAG.getNode(X86ISD::UNPCKL, DL, VT, V1, V2);
6696 return DAG.getNode(X86ISD::UNPCKH, DL, VT, V1, V2);
6698 return DAG.getNode(X86ISD::UNPCKL, DL, VT, V2, V1);
6700 return DAG.getNode(X86ISD::UNPCKH, DL, VT, V2, V1);
6702 llvm_unreachable("Unexpected result of UNPCK mask analysis");
6706 /// \brief Try to emit a bitmask instruction for a shuffle.
6708 /// This handles cases where we can model a blend exactly as a bitmask due to
6709 /// one of the inputs being zeroable.
6710 static SDValue lowerVectorShuffleAsBitMask(SDLoc DL, MVT VT, SDValue V1,
6711 SDValue V2, ArrayRef<int> Mask,
6712 SelectionDAG &DAG) {
6713 MVT EltVT = VT.getScalarType();
6714 int NumEltBits = EltVT.getSizeInBits();
6715 MVT IntEltVT = MVT::getIntegerVT(NumEltBits);
6716 SDValue Zero = DAG.getConstant(0, DL, IntEltVT);
6717 SDValue AllOnes = DAG.getConstant(APInt::getAllOnesValue(NumEltBits), DL,
6719 if (EltVT.isFloatingPoint()) {
6720 Zero = DAG.getBitcast(EltVT, Zero);
6721 AllOnes = DAG.getBitcast(EltVT, AllOnes);
6723 SmallVector<SDValue, 16> VMaskOps(Mask.size(), Zero);
6724 SmallBitVector Zeroable = computeZeroableShuffleElements(Mask, V1, V2);
6726 for (int i = 0, Size = Mask.size(); i < Size; ++i) {
6729 if (Mask[i] % Size != i)
6730 return SDValue(); // Not a blend.
6732 V = Mask[i] < Size ? V1 : V2;
6733 else if (V != (Mask[i] < Size ? V1 : V2))
6734 return SDValue(); // Can only let one input through the mask.
6736 VMaskOps[i] = AllOnes;
6739 return SDValue(); // No non-zeroable elements!
6741 SDValue VMask = DAG.getNode(ISD::BUILD_VECTOR, DL, VT, VMaskOps);
6742 V = DAG.getNode(VT.isFloatingPoint()
6743 ? (unsigned) X86ISD::FAND : (unsigned) ISD::AND,
6748 /// \brief Try to emit a blend instruction for a shuffle using bit math.
6750 /// This is used as a fallback approach when first class blend instructions are
6751 /// unavailable. Currently it is only suitable for integer vectors, but could
6752 /// be generalized for floating point vectors if desirable.
6753 static SDValue lowerVectorShuffleAsBitBlend(SDLoc DL, MVT VT, SDValue V1,
6754 SDValue V2, ArrayRef<int> Mask,
6755 SelectionDAG &DAG) {
6756 assert(VT.isInteger() && "Only supports integer vector types!");
6757 MVT EltVT = VT.getScalarType();
6758 int NumEltBits = EltVT.getSizeInBits();
6759 SDValue Zero = DAG.getConstant(0, DL, EltVT);
6760 SDValue AllOnes = DAG.getConstant(APInt::getAllOnesValue(NumEltBits), DL,
6762 SmallVector<SDValue, 16> MaskOps;
6763 for (int i = 0, Size = Mask.size(); i < Size; ++i) {
6764 if (Mask[i] != -1 && Mask[i] != i && Mask[i] != i + Size)
6765 return SDValue(); // Shuffled input!
6766 MaskOps.push_back(Mask[i] < Size ? AllOnes : Zero);
6769 SDValue V1Mask = DAG.getNode(ISD::BUILD_VECTOR, DL, VT, MaskOps);
6770 V1 = DAG.getNode(ISD::AND, DL, VT, V1, V1Mask);
6771 // We have to cast V2 around.
6772 MVT MaskVT = MVT::getVectorVT(MVT::i64, VT.getSizeInBits() / 64);
6773 V2 = DAG.getBitcast(VT, DAG.getNode(X86ISD::ANDNP, DL, MaskVT,
6774 DAG.getBitcast(MaskVT, V1Mask),
6775 DAG.getBitcast(MaskVT, V2)));
6776 return DAG.getNode(ISD::OR, DL, VT, V1, V2);
6779 /// \brief Try to emit a blend instruction for a shuffle.
6781 /// This doesn't do any checks for the availability of instructions for blending
6782 /// these values. It relies on the availability of the X86ISD::BLENDI pattern to
6783 /// be matched in the backend with the type given. What it does check for is
6784 /// that the shuffle mask is in fact a blend.
6785 static SDValue lowerVectorShuffleAsBlend(SDLoc DL, MVT VT, SDValue V1,
6786 SDValue V2, ArrayRef<int> Mask,
6787 const X86Subtarget *Subtarget,
6788 SelectionDAG &DAG) {
6789 unsigned BlendMask = 0;
6790 for (int i = 0, Size = Mask.size(); i < Size; ++i) {
6791 if (Mask[i] >= Size) {
6792 if (Mask[i] != i + Size)
6793 return SDValue(); // Shuffled V2 input!
6794 BlendMask |= 1u << i;
6797 if (Mask[i] >= 0 && Mask[i] != i)
6798 return SDValue(); // Shuffled V1 input!
6800 switch (VT.SimpleTy) {
6805 return DAG.getNode(X86ISD::BLENDI, DL, VT, V1, V2,
6806 DAG.getConstant(BlendMask, DL, MVT::i8));
6810 assert(Subtarget->hasAVX2() && "256-bit integer blends require AVX2!");
6814 // If we have AVX2 it is faster to use VPBLENDD when the shuffle fits into
6815 // that instruction.
6816 if (Subtarget->hasAVX2()) {
6817 // Scale the blend by the number of 32-bit dwords per element.
6818 int Scale = VT.getScalarSizeInBits() / 32;
6820 for (int i = 0, Size = Mask.size(); i < Size; ++i)
6821 if (Mask[i] >= Size)
6822 for (int j = 0; j < Scale; ++j)
6823 BlendMask |= 1u << (i * Scale + j);
6825 MVT BlendVT = VT.getSizeInBits() > 128 ? MVT::v8i32 : MVT::v4i32;
6826 V1 = DAG.getBitcast(BlendVT, V1);
6827 V2 = DAG.getBitcast(BlendVT, V2);
6828 return DAG.getBitcast(
6829 VT, DAG.getNode(X86ISD::BLENDI, DL, BlendVT, V1, V2,
6830 DAG.getConstant(BlendMask, DL, MVT::i8)));
6834 // For integer shuffles we need to expand the mask and cast the inputs to
6835 // v8i16s prior to blending.
6836 int Scale = 8 / VT.getVectorNumElements();
6838 for (int i = 0, Size = Mask.size(); i < Size; ++i)
6839 if (Mask[i] >= Size)
6840 for (int j = 0; j < Scale; ++j)
6841 BlendMask |= 1u << (i * Scale + j);
6843 V1 = DAG.getBitcast(MVT::v8i16, V1);
6844 V2 = DAG.getBitcast(MVT::v8i16, V2);
6845 return DAG.getBitcast(VT,
6846 DAG.getNode(X86ISD::BLENDI, DL, MVT::v8i16, V1, V2,
6847 DAG.getConstant(BlendMask, DL, MVT::i8)));
6851 assert(Subtarget->hasAVX2() && "256-bit integer blends require AVX2!");
6852 SmallVector<int, 8> RepeatedMask;
6853 if (is128BitLaneRepeatedShuffleMask(MVT::v16i16, Mask, RepeatedMask)) {
6854 // We can lower these with PBLENDW which is mirrored across 128-bit lanes.
6855 assert(RepeatedMask.size() == 8 && "Repeated mask size doesn't match!");
6857 for (int i = 0; i < 8; ++i)
6858 if (RepeatedMask[i] >= 16)
6859 BlendMask |= 1u << i;
6860 return DAG.getNode(X86ISD::BLENDI, DL, MVT::v16i16, V1, V2,
6861 DAG.getConstant(BlendMask, DL, MVT::i8));
6867 assert((VT.getSizeInBits() == 128 || Subtarget->hasAVX2()) &&
6868 "256-bit byte-blends require AVX2 support!");
6870 // Attempt to lower to a bitmask if we can. VPAND is faster than VPBLENDVB.
6871 if (SDValue Masked = lowerVectorShuffleAsBitMask(DL, VT, V1, V2, Mask, DAG))
6874 // Scale the blend by the number of bytes per element.
6875 int Scale = VT.getScalarSizeInBits() / 8;
6877 // This form of blend is always done on bytes. Compute the byte vector
6879 MVT BlendVT = MVT::getVectorVT(MVT::i8, VT.getSizeInBits() / 8);
6881 // Compute the VSELECT mask. Note that VSELECT is really confusing in the
6882 // mix of LLVM's code generator and the x86 backend. We tell the code
6883 // generator that boolean values in the elements of an x86 vector register
6884 // are -1 for true and 0 for false. We then use the LLVM semantics of 'true'
6885 // mapping a select to operand #1, and 'false' mapping to operand #2. The
6886 // reality in x86 is that vector masks (pre-AVX-512) use only the high bit
6887 // of the element (the remaining are ignored) and 0 in that high bit would
6888 // mean operand #1 while 1 in the high bit would mean operand #2. So while
6889 // the LLVM model for boolean values in vector elements gets the relevant
6890 // bit set, it is set backwards and over constrained relative to x86's
6892 SmallVector<SDValue, 32> VSELECTMask;
6893 for (int i = 0, Size = Mask.size(); i < Size; ++i)
6894 for (int j = 0; j < Scale; ++j)
6895 VSELECTMask.push_back(
6896 Mask[i] < 0 ? DAG.getUNDEF(MVT::i8)
6897 : DAG.getConstant(Mask[i] < Size ? -1 : 0, DL,
6900 V1 = DAG.getBitcast(BlendVT, V1);
6901 V2 = DAG.getBitcast(BlendVT, V2);
6902 return DAG.getBitcast(VT, DAG.getNode(ISD::VSELECT, DL, BlendVT,
6903 DAG.getNode(ISD::BUILD_VECTOR, DL,
6904 BlendVT, VSELECTMask),
6909 llvm_unreachable("Not a supported integer vector type!");
6913 /// \brief Try to lower as a blend of elements from two inputs followed by
6914 /// a single-input permutation.
6916 /// This matches the pattern where we can blend elements from two inputs and
6917 /// then reduce the shuffle to a single-input permutation.
6918 static SDValue lowerVectorShuffleAsBlendAndPermute(SDLoc DL, MVT VT, SDValue V1,
6921 SelectionDAG &DAG) {
6922 // We build up the blend mask while checking whether a blend is a viable way
6923 // to reduce the shuffle.
6924 SmallVector<int, 32> BlendMask(Mask.size(), -1);
6925 SmallVector<int, 32> PermuteMask(Mask.size(), -1);
6927 for (int i = 0, Size = Mask.size(); i < Size; ++i) {
6931 assert(Mask[i] < Size * 2 && "Shuffle input is out of bounds.");
6933 if (BlendMask[Mask[i] % Size] == -1)
6934 BlendMask[Mask[i] % Size] = Mask[i];
6935 else if (BlendMask[Mask[i] % Size] != Mask[i])
6936 return SDValue(); // Can't blend in the needed input!
6938 PermuteMask[i] = Mask[i] % Size;
6941 SDValue V = DAG.getVectorShuffle(VT, DL, V1, V2, BlendMask);
6942 return DAG.getVectorShuffle(VT, DL, V, DAG.getUNDEF(VT), PermuteMask);
6945 /// \brief Generic routine to decompose a shuffle and blend into indepndent
6946 /// blends and permutes.
6948 /// This matches the extremely common pattern for handling combined
6949 /// shuffle+blend operations on newer X86 ISAs where we have very fast blend
6950 /// operations. It will try to pick the best arrangement of shuffles and
6952 static SDValue lowerVectorShuffleAsDecomposedShuffleBlend(SDLoc DL, MVT VT,
6956 SelectionDAG &DAG) {
6957 // Shuffle the input elements into the desired positions in V1 and V2 and
6958 // blend them together.
6959 SmallVector<int, 32> V1Mask(Mask.size(), -1);
6960 SmallVector<int, 32> V2Mask(Mask.size(), -1);
6961 SmallVector<int, 32> BlendMask(Mask.size(), -1);
6962 for (int i = 0, Size = Mask.size(); i < Size; ++i)
6963 if (Mask[i] >= 0 && Mask[i] < Size) {
6964 V1Mask[i] = Mask[i];
6966 } else if (Mask[i] >= Size) {
6967 V2Mask[i] = Mask[i] - Size;
6968 BlendMask[i] = i + Size;
6971 // Try to lower with the simpler initial blend strategy unless one of the
6972 // input shuffles would be a no-op. We prefer to shuffle inputs as the
6973 // shuffle may be able to fold with a load or other benefit. However, when
6974 // we'll have to do 2x as many shuffles in order to achieve this, blending
6975 // first is a better strategy.
6976 if (!isNoopShuffleMask(V1Mask) && !isNoopShuffleMask(V2Mask))
6977 if (SDValue BlendPerm =
6978 lowerVectorShuffleAsBlendAndPermute(DL, VT, V1, V2, Mask, DAG))
6981 V1 = DAG.getVectorShuffle(VT, DL, V1, DAG.getUNDEF(VT), V1Mask);
6982 V2 = DAG.getVectorShuffle(VT, DL, V2, DAG.getUNDEF(VT), V2Mask);
6983 return DAG.getVectorShuffle(VT, DL, V1, V2, BlendMask);
6986 /// \brief Try to lower a vector shuffle as a byte rotation.
6988 /// SSSE3 has a generic PALIGNR instruction in x86 that will do an arbitrary
6989 /// byte-rotation of the concatenation of two vectors; pre-SSSE3 can use
6990 /// a PSRLDQ/PSLLDQ/POR pattern to get a similar effect. This routine will
6991 /// try to generically lower a vector shuffle through such an pattern. It
6992 /// does not check for the profitability of lowering either as PALIGNR or
6993 /// PSRLDQ/PSLLDQ/POR, only whether the mask is valid to lower in that form.
6994 /// This matches shuffle vectors that look like:
6996 /// v8i16 [11, 12, 13, 14, 15, 0, 1, 2]
6998 /// Essentially it concatenates V1 and V2, shifts right by some number of
6999 /// elements, and takes the low elements as the result. Note that while this is
7000 /// specified as a *right shift* because x86 is little-endian, it is a *left
7001 /// rotate* of the vector lanes.
7002 static SDValue lowerVectorShuffleAsByteRotate(SDLoc DL, MVT VT, SDValue V1,
7005 const X86Subtarget *Subtarget,
7006 SelectionDAG &DAG) {
7007 assert(!isNoopShuffleMask(Mask) && "We shouldn't lower no-op shuffles!");
7009 int NumElts = Mask.size();
7010 int NumLanes = VT.getSizeInBits() / 128;
7011 int NumLaneElts = NumElts / NumLanes;
7013 // We need to detect various ways of spelling a rotation:
7014 // [11, 12, 13, 14, 15, 0, 1, 2]
7015 // [-1, 12, 13, 14, -1, -1, 1, -1]
7016 // [-1, -1, -1, -1, -1, -1, 1, 2]
7017 // [ 3, 4, 5, 6, 7, 8, 9, 10]
7018 // [-1, 4, 5, 6, -1, -1, 9, -1]
7019 // [-1, 4, 5, 6, -1, -1, -1, -1]
7022 for (int l = 0; l < NumElts; l += NumLaneElts) {
7023 for (int i = 0; i < NumLaneElts; ++i) {
7024 if (Mask[l + i] == -1)
7026 assert(Mask[l + i] >= 0 && "Only -1 is a valid negative mask element!");
7028 // Get the mod-Size index and lane correct it.
7029 int LaneIdx = (Mask[l + i] % NumElts) - l;
7030 // Make sure it was in this lane.
7031 if (LaneIdx < 0 || LaneIdx >= NumLaneElts)
7034 // Determine where a rotated vector would have started.
7035 int StartIdx = i - LaneIdx;
7037 // The identity rotation isn't interesting, stop.
7040 // If we found the tail of a vector the rotation must be the missing
7041 // front. If we found the head of a vector, it must be how much of the
7043 int CandidateRotation = StartIdx < 0 ? -StartIdx : NumLaneElts - StartIdx;
7046 Rotation = CandidateRotation;
7047 else if (Rotation != CandidateRotation)
7048 // The rotations don't match, so we can't match this mask.
7051 // Compute which value this mask is pointing at.
7052 SDValue MaskV = Mask[l + i] < NumElts ? V1 : V2;
7054 // Compute which of the two target values this index should be assigned
7055 // to. This reflects whether the high elements are remaining or the low
7056 // elements are remaining.
7057 SDValue &TargetV = StartIdx < 0 ? Hi : Lo;
7059 // Either set up this value if we've not encountered it before, or check
7060 // that it remains consistent.
7063 else if (TargetV != MaskV)
7064 // This may be a rotation, but it pulls from the inputs in some
7065 // unsupported interleaving.
7070 // Check that we successfully analyzed the mask, and normalize the results.
7071 assert(Rotation != 0 && "Failed to locate a viable rotation!");
7072 assert((Lo || Hi) && "Failed to find a rotated input vector!");
7078 // The actual rotate instruction rotates bytes, so we need to scale the
7079 // rotation based on how many bytes are in the vector lane.
7080 int Scale = 16 / NumLaneElts;
7082 // SSSE3 targets can use the palignr instruction.
7083 if (Subtarget->hasSSSE3()) {
7084 // Cast the inputs to i8 vector of correct length to match PALIGNR.
7085 MVT AlignVT = MVT::getVectorVT(MVT::i8, 16 * NumLanes);
7086 Lo = DAG.getBitcast(AlignVT, Lo);
7087 Hi = DAG.getBitcast(AlignVT, Hi);
7089 return DAG.getBitcast(
7090 VT, DAG.getNode(X86ISD::PALIGNR, DL, AlignVT, Lo, Hi,
7091 DAG.getConstant(Rotation * Scale, DL, MVT::i8)));
7094 assert(VT.getSizeInBits() == 128 &&
7095 "Rotate-based lowering only supports 128-bit lowering!");
7096 assert(Mask.size() <= 16 &&
7097 "Can shuffle at most 16 bytes in a 128-bit vector!");
7099 // Default SSE2 implementation
7100 int LoByteShift = 16 - Rotation * Scale;
7101 int HiByteShift = Rotation * Scale;
7103 // Cast the inputs to v2i64 to match PSLLDQ/PSRLDQ.
7104 Lo = DAG.getBitcast(MVT::v2i64, Lo);
7105 Hi = DAG.getBitcast(MVT::v2i64, Hi);
7107 SDValue LoShift = DAG.getNode(X86ISD::VSHLDQ, DL, MVT::v2i64, Lo,
7108 DAG.getConstant(LoByteShift, DL, MVT::i8));
7109 SDValue HiShift = DAG.getNode(X86ISD::VSRLDQ, DL, MVT::v2i64, Hi,
7110 DAG.getConstant(HiByteShift, DL, MVT::i8));
7111 return DAG.getBitcast(VT,
7112 DAG.getNode(ISD::OR, DL, MVT::v2i64, LoShift, HiShift));
7115 /// \brief Try to lower a vector shuffle as a bit shift (shifts in zeros).
7117 /// Attempts to match a shuffle mask against the PSLL(W/D/Q/DQ) and
7118 /// PSRL(W/D/Q/DQ) SSE2 and AVX2 logical bit-shift instructions. The function
7119 /// matches elements from one of the input vectors shuffled to the left or
7120 /// right with zeroable elements 'shifted in'. It handles both the strictly
7121 /// bit-wise element shifts and the byte shift across an entire 128-bit double
7124 /// PSHL : (little-endian) left bit shift.
7125 /// [ zz, 0, zz, 2 ]
7126 /// [ -1, 4, zz, -1 ]
7127 /// PSRL : (little-endian) right bit shift.
7129 /// [ -1, -1, 7, zz]
7130 /// PSLLDQ : (little-endian) left byte shift
7131 /// [ zz, 0, 1, 2, 3, 4, 5, 6]
7132 /// [ zz, zz, -1, -1, 2, 3, 4, -1]
7133 /// [ zz, zz, zz, zz, zz, zz, -1, 1]
7134 /// PSRLDQ : (little-endian) right byte shift
7135 /// [ 5, 6, 7, zz, zz, zz, zz, zz]
7136 /// [ -1, 5, 6, 7, zz, zz, zz, zz]
7137 /// [ 1, 2, -1, -1, -1, -1, zz, zz]
7138 static SDValue lowerVectorShuffleAsShift(SDLoc DL, MVT VT, SDValue V1,
7139 SDValue V2, ArrayRef<int> Mask,
7140 SelectionDAG &DAG) {
7141 SmallBitVector Zeroable = computeZeroableShuffleElements(Mask, V1, V2);
7143 int Size = Mask.size();
7144 assert(Size == (int)VT.getVectorNumElements() && "Unexpected mask size");
7146 auto CheckZeros = [&](int Shift, int Scale, bool Left) {
7147 for (int i = 0; i < Size; i += Scale)
7148 for (int j = 0; j < Shift; ++j)
7149 if (!Zeroable[i + j + (Left ? 0 : (Scale - Shift))])
7155 auto MatchShift = [&](int Shift, int Scale, bool Left, SDValue V) {
7156 for (int i = 0; i != Size; i += Scale) {
7157 unsigned Pos = Left ? i + Shift : i;
7158 unsigned Low = Left ? i : i + Shift;
7159 unsigned Len = Scale - Shift;
7160 if (!isSequentialOrUndefInRange(Mask, Pos, Len,
7161 Low + (V == V1 ? 0 : Size)))
7165 int ShiftEltBits = VT.getScalarSizeInBits() * Scale;
7166 bool ByteShift = ShiftEltBits > 64;
7167 unsigned OpCode = Left ? (ByteShift ? X86ISD::VSHLDQ : X86ISD::VSHLI)
7168 : (ByteShift ? X86ISD::VSRLDQ : X86ISD::VSRLI);
7169 int ShiftAmt = Shift * VT.getScalarSizeInBits() / (ByteShift ? 8 : 1);
7171 // Normalize the scale for byte shifts to still produce an i64 element
7173 Scale = ByteShift ? Scale / 2 : Scale;
7175 // We need to round trip through the appropriate type for the shift.
7176 MVT ShiftSVT = MVT::getIntegerVT(VT.getScalarSizeInBits() * Scale);
7177 MVT ShiftVT = MVT::getVectorVT(ShiftSVT, Size / Scale);
7178 assert(DAG.getTargetLoweringInfo().isTypeLegal(ShiftVT) &&
7179 "Illegal integer vector type");
7180 V = DAG.getBitcast(ShiftVT, V);
7182 V = DAG.getNode(OpCode, DL, ShiftVT, V,
7183 DAG.getConstant(ShiftAmt, DL, MVT::i8));
7184 return DAG.getBitcast(VT, V);
7187 // SSE/AVX supports logical shifts up to 64-bit integers - so we can just
7188 // keep doubling the size of the integer elements up to that. We can
7189 // then shift the elements of the integer vector by whole multiples of
7190 // their width within the elements of the larger integer vector. Test each
7191 // multiple to see if we can find a match with the moved element indices
7192 // and that the shifted in elements are all zeroable.
7193 for (int Scale = 2; Scale * VT.getScalarSizeInBits() <= 128; Scale *= 2)
7194 for (int Shift = 1; Shift != Scale; ++Shift)
7195 for (bool Left : {true, false})
7196 if (CheckZeros(Shift, Scale, Left))
7197 for (SDValue V : {V1, V2})
7198 if (SDValue Match = MatchShift(Shift, Scale, Left, V))
7205 /// \brief Try to lower a vector shuffle using SSE4a EXTRQ/INSERTQ.
7206 static SDValue lowerVectorShuffleWithSSE4A(SDLoc DL, MVT VT, SDValue V1,
7207 SDValue V2, ArrayRef<int> Mask,
7208 SelectionDAG &DAG) {
7209 SmallBitVector Zeroable = computeZeroableShuffleElements(Mask, V1, V2);
7210 assert(!Zeroable.all() && "Fully zeroable shuffle mask");
7212 int Size = Mask.size();
7213 int HalfSize = Size / 2;
7214 assert(Size == (int)VT.getVectorNumElements() && "Unexpected mask size");
7216 // Upper half must be undefined.
7217 if (!isUndefInRange(Mask, HalfSize, HalfSize))
7220 // EXTRQ: Extract Len elements from lower half of source, starting at Idx.
7221 // Remainder of lower half result is zero and upper half is all undef.
7222 auto LowerAsEXTRQ = [&]() {
7223 // Determine the extraction length from the part of the
7224 // lower half that isn't zeroable.
7226 for (; Len >= 0; --Len)
7227 if (!Zeroable[Len - 1])
7229 assert(Len > 0 && "Zeroable shuffle mask");
7231 // Attempt to match first Len sequential elements from the lower half.
7234 for (int i = 0; i != Len; ++i) {
7238 SDValue &V = (M < Size ? V1 : V2);
7241 // All mask elements must be in the lower half.
7245 if (Idx < 0 || (Src == V && Idx == (M - i))) {
7256 assert((Idx + Len) <= HalfSize && "Illegal extraction mask");
7257 int BitLen = (Len * VT.getScalarSizeInBits()) & 0x3f;
7258 int BitIdx = (Idx * VT.getScalarSizeInBits()) & 0x3f;
7259 return DAG.getNode(X86ISD::EXTRQI, DL, VT, Src,
7260 DAG.getConstant(BitLen, DL, MVT::i8),
7261 DAG.getConstant(BitIdx, DL, MVT::i8));
7264 if (SDValue ExtrQ = LowerAsEXTRQ())
7267 // INSERTQ: Extract lowest Len elements from lower half of second source and
7268 // insert over first source, starting at Idx.
7269 // { A[0], .., A[Idx-1], B[0], .., B[Len-1], A[Idx+Len], .., UNDEF, ... }
7270 auto LowerAsInsertQ = [&]() {
7271 for (int Idx = 0; Idx != HalfSize; ++Idx) {
7274 // Attempt to match first source from mask before insertion point.
7275 if (isUndefInRange(Mask, 0, Idx)) {
7277 } else if (isSequentialOrUndefInRange(Mask, 0, Idx, 0)) {
7279 } else if (isSequentialOrUndefInRange(Mask, 0, Idx, Size)) {
7285 // Extend the extraction length looking to match both the insertion of
7286 // the second source and the remaining elements of the first.
7287 for (int Hi = Idx + 1; Hi <= HalfSize; ++Hi) {
7292 if (isSequentialOrUndefInRange(Mask, Idx, Len, 0)) {
7294 } else if (isSequentialOrUndefInRange(Mask, Idx, Len, Size)) {
7300 // Match the remaining elements of the lower half.
7301 if (isUndefInRange(Mask, Hi, HalfSize - Hi)) {
7303 } else if ((!Base || (Base == V1)) &&
7304 isSequentialOrUndefInRange(Mask, Hi, HalfSize - Hi, Hi)) {
7306 } else if ((!Base || (Base == V2)) &&
7307 isSequentialOrUndefInRange(Mask, Hi, HalfSize - Hi,
7314 // We may not have a base (first source) - this can safely be undefined.
7316 Base = DAG.getUNDEF(VT);
7318 int BitLen = (Len * VT.getScalarSizeInBits()) & 0x3f;
7319 int BitIdx = (Idx * VT.getScalarSizeInBits()) & 0x3f;
7320 return DAG.getNode(X86ISD::INSERTQI, DL, VT, Base, Insert,
7321 DAG.getConstant(BitLen, DL, MVT::i8),
7322 DAG.getConstant(BitIdx, DL, MVT::i8));
7329 if (SDValue InsertQ = LowerAsInsertQ())
7335 /// \brief Lower a vector shuffle as a zero or any extension.
7337 /// Given a specific number of elements, element bit width, and extension
7338 /// stride, produce either a zero or any extension based on the available
7339 /// features of the subtarget.
7340 static SDValue lowerVectorShuffleAsSpecificZeroOrAnyExtend(
7341 SDLoc DL, MVT VT, int Scale, bool AnyExt, SDValue InputV,
7342 ArrayRef<int> Mask, const X86Subtarget *Subtarget, SelectionDAG &DAG) {
7343 assert(Scale > 1 && "Need a scale to extend.");
7344 int NumElements = VT.getVectorNumElements();
7345 int EltBits = VT.getScalarSizeInBits();
7346 assert((EltBits == 8 || EltBits == 16 || EltBits == 32) &&
7347 "Only 8, 16, and 32 bit elements can be extended.");
7348 assert(Scale * EltBits <= 64 && "Cannot zero extend past 64 bits.");
7350 // Found a valid zext mask! Try various lowering strategies based on the
7351 // input type and available ISA extensions.
7352 if (Subtarget->hasSSE41()) {
7353 MVT ExtVT = MVT::getVectorVT(MVT::getIntegerVT(EltBits * Scale),
7354 NumElements / Scale);
7355 return DAG.getBitcast(VT, DAG.getNode(X86ISD::VZEXT, DL, ExtVT, InputV));
7358 // For any extends we can cheat for larger element sizes and use shuffle
7359 // instructions that can fold with a load and/or copy.
7360 if (AnyExt && EltBits == 32) {
7361 int PSHUFDMask[4] = {0, -1, 1, -1};
7362 return DAG.getBitcast(
7363 VT, DAG.getNode(X86ISD::PSHUFD, DL, MVT::v4i32,
7364 DAG.getBitcast(MVT::v4i32, InputV),
7365 getV4X86ShuffleImm8ForMask(PSHUFDMask, DL, DAG)));
7367 if (AnyExt && EltBits == 16 && Scale > 2) {
7368 int PSHUFDMask[4] = {0, -1, 0, -1};
7369 InputV = DAG.getNode(X86ISD::PSHUFD, DL, MVT::v4i32,
7370 DAG.getBitcast(MVT::v4i32, InputV),
7371 getV4X86ShuffleImm8ForMask(PSHUFDMask, DL, DAG));
7372 int PSHUFHWMask[4] = {1, -1, -1, -1};
7373 return DAG.getBitcast(
7374 VT, DAG.getNode(X86ISD::PSHUFHW, DL, MVT::v8i16,
7375 DAG.getBitcast(MVT::v8i16, InputV),
7376 getV4X86ShuffleImm8ForMask(PSHUFHWMask, DL, DAG)));
7379 // The SSE4A EXTRQ instruction can efficiently extend the first 2 lanes
7381 if ((Scale * EltBits) == 64 && EltBits < 32 && Subtarget->hasSSE4A()) {
7382 assert(NumElements == (int)Mask.size() && "Unexpected shuffle mask size!");
7383 assert(VT.getSizeInBits() == 128 && "Unexpected vector width!");
7385 SDValue Lo = DAG.getNode(ISD::BITCAST, DL, MVT::v2i64,
7386 DAG.getNode(X86ISD::EXTRQI, DL, VT, InputV,
7387 DAG.getConstant(EltBits, DL, MVT::i8),
7388 DAG.getConstant(0, DL, MVT::i8)));
7389 if (isUndefInRange(Mask, NumElements/2, NumElements/2))
7390 return DAG.getNode(ISD::BITCAST, DL, VT, Lo);
7393 DAG.getNode(ISD::BITCAST, DL, MVT::v2i64,
7394 DAG.getNode(X86ISD::EXTRQI, DL, VT, InputV,
7395 DAG.getConstant(EltBits, DL, MVT::i8),
7396 DAG.getConstant(EltBits, DL, MVT::i8)));
7397 return DAG.getNode(ISD::BITCAST, DL, VT,
7398 DAG.getNode(X86ISD::UNPCKL, DL, MVT::v2i64, Lo, Hi));
7401 // If this would require more than 2 unpack instructions to expand, use
7402 // pshufb when available. We can only use more than 2 unpack instructions
7403 // when zero extending i8 elements which also makes it easier to use pshufb.
7404 if (Scale > 4 && EltBits == 8 && Subtarget->hasSSSE3()) {
7405 assert(NumElements == 16 && "Unexpected byte vector width!");
7406 SDValue PSHUFBMask[16];
7407 for (int i = 0; i < 16; ++i)
7409 DAG.getConstant((i % Scale == 0) ? i / Scale : 0x80, DL, MVT::i8);
7410 InputV = DAG.getBitcast(MVT::v16i8, InputV);
7411 return DAG.getBitcast(VT,
7412 DAG.getNode(X86ISD::PSHUFB, DL, MVT::v16i8, InputV,
7413 DAG.getNode(ISD::BUILD_VECTOR, DL,
7414 MVT::v16i8, PSHUFBMask)));
7417 // Otherwise emit a sequence of unpacks.
7419 MVT InputVT = MVT::getVectorVT(MVT::getIntegerVT(EltBits), NumElements);
7420 SDValue Ext = AnyExt ? DAG.getUNDEF(InputVT)
7421 : getZeroVector(InputVT, Subtarget, DAG, DL);
7422 InputV = DAG.getBitcast(InputVT, InputV);
7423 InputV = DAG.getNode(X86ISD::UNPCKL, DL, InputVT, InputV, Ext);
7427 } while (Scale > 1);
7428 return DAG.getBitcast(VT, InputV);
7431 /// \brief Try to lower a vector shuffle as a zero extension on any microarch.
7433 /// This routine will try to do everything in its power to cleverly lower
7434 /// a shuffle which happens to match the pattern of a zero extend. It doesn't
7435 /// check for the profitability of this lowering, it tries to aggressively
7436 /// match this pattern. It will use all of the micro-architectural details it
7437 /// can to emit an efficient lowering. It handles both blends with all-zero
7438 /// inputs to explicitly zero-extend and undef-lanes (sometimes undef due to
7439 /// masking out later).
7441 /// The reason we have dedicated lowering for zext-style shuffles is that they
7442 /// are both incredibly common and often quite performance sensitive.
7443 static SDValue lowerVectorShuffleAsZeroOrAnyExtend(
7444 SDLoc DL, MVT VT, SDValue V1, SDValue V2, ArrayRef<int> Mask,
7445 const X86Subtarget *Subtarget, SelectionDAG &DAG) {
7446 SmallBitVector Zeroable = computeZeroableShuffleElements(Mask, V1, V2);
7448 int Bits = VT.getSizeInBits();
7449 int NumElements = VT.getVectorNumElements();
7450 assert(VT.getScalarSizeInBits() <= 32 &&
7451 "Exceeds 32-bit integer zero extension limit");
7452 assert((int)Mask.size() == NumElements && "Unexpected shuffle mask size");
7454 // Define a helper function to check a particular ext-scale and lower to it if
7456 auto Lower = [&](int Scale) -> SDValue {
7459 for (int i = 0; i < NumElements; ++i) {
7461 continue; // Valid anywhere but doesn't tell us anything.
7462 if (i % Scale != 0) {
7463 // Each of the extended elements need to be zeroable.
7467 // We no longer are in the anyext case.
7472 // Each of the base elements needs to be consecutive indices into the
7473 // same input vector.
7474 SDValue V = Mask[i] < NumElements ? V1 : V2;
7477 else if (InputV != V)
7478 return SDValue(); // Flip-flopping inputs.
7480 if (Mask[i] % NumElements != i / Scale)
7481 return SDValue(); // Non-consecutive strided elements.
7484 // If we fail to find an input, we have a zero-shuffle which should always
7485 // have already been handled.
7486 // FIXME: Maybe handle this here in case during blending we end up with one?
7490 return lowerVectorShuffleAsSpecificZeroOrAnyExtend(
7491 DL, VT, Scale, AnyExt, InputV, Mask, Subtarget, DAG);
7494 // The widest scale possible for extending is to a 64-bit integer.
7495 assert(Bits % 64 == 0 &&
7496 "The number of bits in a vector must be divisible by 64 on x86!");
7497 int NumExtElements = Bits / 64;
7499 // Each iteration, try extending the elements half as much, but into twice as
7501 for (; NumExtElements < NumElements; NumExtElements *= 2) {
7502 assert(NumElements % NumExtElements == 0 &&
7503 "The input vector size must be divisible by the extended size.");
7504 if (SDValue V = Lower(NumElements / NumExtElements))
7508 // General extends failed, but 128-bit vectors may be able to use MOVQ.
7512 // Returns one of the source operands if the shuffle can be reduced to a
7513 // MOVQ, copying the lower 64-bits and zero-extending to the upper 64-bits.
7514 auto CanZExtLowHalf = [&]() {
7515 for (int i = NumElements / 2; i != NumElements; ++i)
7518 if (isSequentialOrUndefInRange(Mask, 0, NumElements / 2, 0))
7520 if (isSequentialOrUndefInRange(Mask, 0, NumElements / 2, NumElements))
7525 if (SDValue V = CanZExtLowHalf()) {
7526 V = DAG.getBitcast(MVT::v2i64, V);
7527 V = DAG.getNode(X86ISD::VZEXT_MOVL, DL, MVT::v2i64, V);
7528 return DAG.getBitcast(VT, V);
7531 // No viable ext lowering found.
7535 /// \brief Try to get a scalar value for a specific element of a vector.
7537 /// Looks through BUILD_VECTOR and SCALAR_TO_VECTOR nodes to find a scalar.
7538 static SDValue getScalarValueForVectorElement(SDValue V, int Idx,
7539 SelectionDAG &DAG) {
7540 MVT VT = V.getSimpleValueType();
7541 MVT EltVT = VT.getVectorElementType();
7542 while (V.getOpcode() == ISD::BITCAST)
7543 V = V.getOperand(0);
7544 // If the bitcasts shift the element size, we can't extract an equivalent
7546 MVT NewVT = V.getSimpleValueType();
7547 if (!NewVT.isVector() || NewVT.getScalarSizeInBits() != VT.getScalarSizeInBits())
7550 if (V.getOpcode() == ISD::BUILD_VECTOR ||
7551 (Idx == 0 && V.getOpcode() == ISD::SCALAR_TO_VECTOR)) {
7552 // Ensure the scalar operand is the same size as the destination.
7553 // FIXME: Add support for scalar truncation where possible.
7554 SDValue S = V.getOperand(Idx);
7555 if (EltVT.getSizeInBits() == S.getSimpleValueType().getSizeInBits())
7556 return DAG.getNode(ISD::BITCAST, SDLoc(V), EltVT, S);
7562 /// \brief Helper to test for a load that can be folded with x86 shuffles.
7564 /// This is particularly important because the set of instructions varies
7565 /// significantly based on whether the operand is a load or not.
7566 static bool isShuffleFoldableLoad(SDValue V) {
7567 while (V.getOpcode() == ISD::BITCAST)
7568 V = V.getOperand(0);
7570 return ISD::isNON_EXTLoad(V.getNode());
7573 /// \brief Try to lower insertion of a single element into a zero vector.
7575 /// This is a common pattern that we have especially efficient patterns to lower
7576 /// across all subtarget feature sets.
7577 static SDValue lowerVectorShuffleAsElementInsertion(
7578 SDLoc DL, MVT VT, SDValue V1, SDValue V2, ArrayRef<int> Mask,
7579 const X86Subtarget *Subtarget, SelectionDAG &DAG) {
7580 SmallBitVector Zeroable = computeZeroableShuffleElements(Mask, V1, V2);
7582 MVT EltVT = VT.getVectorElementType();
7584 int V2Index = std::find_if(Mask.begin(), Mask.end(),
7585 [&Mask](int M) { return M >= (int)Mask.size(); }) -
7587 bool IsV1Zeroable = true;
7588 for (int i = 0, Size = Mask.size(); i < Size; ++i)
7589 if (i != V2Index && !Zeroable[i]) {
7590 IsV1Zeroable = false;
7594 // Check for a single input from a SCALAR_TO_VECTOR node.
7595 // FIXME: All of this should be canonicalized into INSERT_VECTOR_ELT and
7596 // all the smarts here sunk into that routine. However, the current
7597 // lowering of BUILD_VECTOR makes that nearly impossible until the old
7598 // vector shuffle lowering is dead.
7599 SDValue V2S = getScalarValueForVectorElement(V2, Mask[V2Index] - Mask.size(),
7601 if (V2S && DAG.getTargetLoweringInfo().isTypeLegal(V2S.getValueType())) {
7602 // We need to zext the scalar if it is smaller than an i32.
7603 V2S = DAG.getBitcast(EltVT, V2S);
7604 if (EltVT == MVT::i8 || EltVT == MVT::i16) {
7605 // Using zext to expand a narrow element won't work for non-zero
7610 // Zero-extend directly to i32.
7612 V2S = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::i32, V2S);
7614 V2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, ExtVT, V2S);
7615 } else if (Mask[V2Index] != (int)Mask.size() || EltVT == MVT::i8 ||
7616 EltVT == MVT::i16) {
7617 // Either not inserting from the low element of the input or the input
7618 // element size is too small to use VZEXT_MOVL to clear the high bits.
7622 if (!IsV1Zeroable) {
7623 // If V1 can't be treated as a zero vector we have fewer options to lower
7624 // this. We can't support integer vectors or non-zero targets cheaply, and
7625 // the V1 elements can't be permuted in any way.
7626 assert(VT == ExtVT && "Cannot change extended type when non-zeroable!");
7627 if (!VT.isFloatingPoint() || V2Index != 0)
7629 SmallVector<int, 8> V1Mask(Mask.begin(), Mask.end());
7630 V1Mask[V2Index] = -1;
7631 if (!isNoopShuffleMask(V1Mask))
7633 // This is essentially a special case blend operation, but if we have
7634 // general purpose blend operations, they are always faster. Bail and let
7635 // the rest of the lowering handle these as blends.
7636 if (Subtarget->hasSSE41())
7639 // Otherwise, use MOVSD or MOVSS.
7640 assert((EltVT == MVT::f32 || EltVT == MVT::f64) &&
7641 "Only two types of floating point element types to handle!");
7642 return DAG.getNode(EltVT == MVT::f32 ? X86ISD::MOVSS : X86ISD::MOVSD, DL,
7646 // This lowering only works for the low element with floating point vectors.
7647 if (VT.isFloatingPoint() && V2Index != 0)
7650 V2 = DAG.getNode(X86ISD::VZEXT_MOVL, DL, ExtVT, V2);
7652 V2 = DAG.getBitcast(VT, V2);
7655 // If we have 4 or fewer lanes we can cheaply shuffle the element into
7656 // the desired position. Otherwise it is more efficient to do a vector
7657 // shift left. We know that we can do a vector shift left because all
7658 // the inputs are zero.
7659 if (VT.isFloatingPoint() || VT.getVectorNumElements() <= 4) {
7660 SmallVector<int, 4> V2Shuffle(Mask.size(), 1);
7661 V2Shuffle[V2Index] = 0;
7662 V2 = DAG.getVectorShuffle(VT, DL, V2, DAG.getUNDEF(VT), V2Shuffle);
7664 V2 = DAG.getBitcast(MVT::v2i64, V2);
7666 X86ISD::VSHLDQ, DL, MVT::v2i64, V2,
7667 DAG.getConstant(V2Index * EltVT.getSizeInBits() / 8, DL,
7668 DAG.getTargetLoweringInfo().getScalarShiftAmountTy(
7669 DAG.getDataLayout(), VT)));
7670 V2 = DAG.getBitcast(VT, V2);
7676 /// \brief Try to lower broadcast of a single element.
7678 /// For convenience, this code also bundles all of the subtarget feature set
7679 /// filtering. While a little annoying to re-dispatch on type here, there isn't
7680 /// a convenient way to factor it out.
7681 static SDValue lowerVectorShuffleAsBroadcast(SDLoc DL, MVT VT, SDValue V,
7683 const X86Subtarget *Subtarget,
7684 SelectionDAG &DAG) {
7685 if (!Subtarget->hasAVX())
7687 if (VT.isInteger() && !Subtarget->hasAVX2())
7690 // Check that the mask is a broadcast.
7691 int BroadcastIdx = -1;
7693 if (M >= 0 && BroadcastIdx == -1)
7695 else if (M >= 0 && M != BroadcastIdx)
7698 assert(BroadcastIdx < (int)Mask.size() && "We only expect to be called with "
7699 "a sorted mask where the broadcast "
7702 // Go up the chain of (vector) values to find a scalar load that we can
7703 // combine with the broadcast.
7705 switch (V.getOpcode()) {
7706 case ISD::CONCAT_VECTORS: {
7707 int OperandSize = Mask.size() / V.getNumOperands();
7708 V = V.getOperand(BroadcastIdx / OperandSize);
7709 BroadcastIdx %= OperandSize;
7713 case ISD::INSERT_SUBVECTOR: {
7714 SDValue VOuter = V.getOperand(0), VInner = V.getOperand(1);
7715 auto ConstantIdx = dyn_cast<ConstantSDNode>(V.getOperand(2));
7719 int BeginIdx = (int)ConstantIdx->getZExtValue();
7721 BeginIdx + (int)VInner.getValueType().getVectorNumElements();
7722 if (BroadcastIdx >= BeginIdx && BroadcastIdx < EndIdx) {
7723 BroadcastIdx -= BeginIdx;
7734 // Check if this is a broadcast of a scalar. We special case lowering
7735 // for scalars so that we can more effectively fold with loads.
7736 // First, look through bitcast: if the original value has a larger element
7737 // type than the shuffle, the broadcast element is in essence truncated.
7738 // Make that explicit to ease folding.
7739 if (V.getOpcode() == ISD::BITCAST && VT.isInteger()) {
7740 EVT EltVT = VT.getVectorElementType();
7741 SDValue V0 = V.getOperand(0);
7742 EVT V0VT = V0.getValueType();
7744 if (V0VT.isInteger() && V0VT.getVectorElementType().bitsGT(EltVT) &&
7745 ((V0.getOpcode() == ISD::BUILD_VECTOR ||
7746 (V0.getOpcode() == ISD::SCALAR_TO_VECTOR && BroadcastIdx == 0)))) {
7747 V = DAG.getNode(ISD::TRUNCATE, DL, EltVT, V0.getOperand(BroadcastIdx));
7752 // Also check the simpler case, where we can directly reuse the scalar.
7753 if (V.getOpcode() == ISD::BUILD_VECTOR ||
7754 (V.getOpcode() == ISD::SCALAR_TO_VECTOR && BroadcastIdx == 0)) {
7755 V = V.getOperand(BroadcastIdx);
7757 // If the scalar isn't a load, we can't broadcast from it in AVX1.
7758 // Only AVX2 has register broadcasts.
7759 if (!Subtarget->hasAVX2() && !isShuffleFoldableLoad(V))
7761 } else if (BroadcastIdx != 0 || !Subtarget->hasAVX2()) {
7762 // We can't broadcast from a vector register without AVX2, and we can only
7763 // broadcast from the zero-element of a vector register.
7767 return DAG.getNode(X86ISD::VBROADCAST, DL, VT, V);
7770 // Check for whether we can use INSERTPS to perform the shuffle. We only use
7771 // INSERTPS when the V1 elements are already in the correct locations
7772 // because otherwise we can just always use two SHUFPS instructions which
7773 // are much smaller to encode than a SHUFPS and an INSERTPS. We can also
7774 // perform INSERTPS if a single V1 element is out of place and all V2
7775 // elements are zeroable.
7776 static SDValue lowerVectorShuffleAsInsertPS(SDValue Op, SDValue V1, SDValue V2,
7778 SelectionDAG &DAG) {
7779 assert(Op.getSimpleValueType() == MVT::v4f32 && "Bad shuffle type!");
7780 assert(V1.getSimpleValueType() == MVT::v4f32 && "Bad operand type!");
7781 assert(V2.getSimpleValueType() == MVT::v4f32 && "Bad operand type!");
7782 assert(Mask.size() == 4 && "Unexpected mask size for v4 shuffle!");
7784 SmallBitVector Zeroable = computeZeroableShuffleElements(Mask, V1, V2);
7787 int V1DstIndex = -1;
7788 int V2DstIndex = -1;
7789 bool V1UsedInPlace = false;
7791 for (int i = 0; i < 4; ++i) {
7792 // Synthesize a zero mask from the zeroable elements (includes undefs).
7798 // Flag if we use any V1 inputs in place.
7800 V1UsedInPlace = true;
7804 // We can only insert a single non-zeroable element.
7805 if (V1DstIndex != -1 || V2DstIndex != -1)
7809 // V1 input out of place for insertion.
7812 // V2 input for insertion.
7817 // Don't bother if we have no (non-zeroable) element for insertion.
7818 if (V1DstIndex == -1 && V2DstIndex == -1)
7821 // Determine element insertion src/dst indices. The src index is from the
7822 // start of the inserted vector, not the start of the concatenated vector.
7823 unsigned V2SrcIndex = 0;
7824 if (V1DstIndex != -1) {
7825 // If we have a V1 input out of place, we use V1 as the V2 element insertion
7826 // and don't use the original V2 at all.
7827 V2SrcIndex = Mask[V1DstIndex];
7828 V2DstIndex = V1DstIndex;
7831 V2SrcIndex = Mask[V2DstIndex] - 4;
7834 // If no V1 inputs are used in place, then the result is created only from
7835 // the zero mask and the V2 insertion - so remove V1 dependency.
7837 V1 = DAG.getUNDEF(MVT::v4f32);
7839 unsigned InsertPSMask = V2SrcIndex << 6 | V2DstIndex << 4 | ZMask;
7840 assert((InsertPSMask & ~0xFFu) == 0 && "Invalid mask!");
7842 // Insert the V2 element into the desired position.
7844 return DAG.getNode(X86ISD::INSERTPS, DL, MVT::v4f32, V1, V2,
7845 DAG.getConstant(InsertPSMask, DL, MVT::i8));
7848 /// \brief Try to lower a shuffle as a permute of the inputs followed by an
7849 /// UNPCK instruction.
7851 /// This specifically targets cases where we end up with alternating between
7852 /// the two inputs, and so can permute them into something that feeds a single
7853 /// UNPCK instruction. Note that this routine only targets integer vectors
7854 /// because for floating point vectors we have a generalized SHUFPS lowering
7855 /// strategy that handles everything that doesn't *exactly* match an unpack,
7856 /// making this clever lowering unnecessary.
7857 static SDValue lowerVectorShuffleAsPermuteAndUnpack(SDLoc DL, MVT VT,
7858 SDValue V1, SDValue V2,
7860 SelectionDAG &DAG) {
7861 assert(!VT.isFloatingPoint() &&
7862 "This routine only supports integer vectors.");
7863 assert(!isSingleInputShuffleMask(Mask) &&
7864 "This routine should only be used when blending two inputs.");
7865 assert(Mask.size() >= 2 && "Single element masks are invalid.");
7867 int Size = Mask.size();
7869 int NumLoInputs = std::count_if(Mask.begin(), Mask.end(), [Size](int M) {
7870 return M >= 0 && M % Size < Size / 2;
7872 int NumHiInputs = std::count_if(
7873 Mask.begin(), Mask.end(), [Size](int M) { return M % Size >= Size / 2; });
7875 bool UnpackLo = NumLoInputs >= NumHiInputs;
7877 auto TryUnpack = [&](MVT UnpackVT, int Scale) {
7878 SmallVector<int, 32> V1Mask(Mask.size(), -1);
7879 SmallVector<int, 32> V2Mask(Mask.size(), -1);
7881 for (int i = 0; i < Size; ++i) {
7885 // Each element of the unpack contains Scale elements from this mask.
7886 int UnpackIdx = i / Scale;
7888 // We only handle the case where V1 feeds the first slots of the unpack.
7889 // We rely on canonicalization to ensure this is the case.
7890 if ((UnpackIdx % 2 == 0) != (Mask[i] < Size))
7893 // Setup the mask for this input. The indexing is tricky as we have to
7894 // handle the unpack stride.
7895 SmallVectorImpl<int> &VMask = (UnpackIdx % 2 == 0) ? V1Mask : V2Mask;
7896 VMask[(UnpackIdx / 2) * Scale + i % Scale + (UnpackLo ? 0 : Size / 2)] =
7900 // If we will have to shuffle both inputs to use the unpack, check whether
7901 // we can just unpack first and shuffle the result. If so, skip this unpack.
7902 if ((NumLoInputs == 0 || NumHiInputs == 0) && !isNoopShuffleMask(V1Mask) &&
7903 !isNoopShuffleMask(V2Mask))
7906 // Shuffle the inputs into place.
7907 V1 = DAG.getVectorShuffle(VT, DL, V1, DAG.getUNDEF(VT), V1Mask);
7908 V2 = DAG.getVectorShuffle(VT, DL, V2, DAG.getUNDEF(VT), V2Mask);
7910 // Cast the inputs to the type we will use to unpack them.
7911 V1 = DAG.getBitcast(UnpackVT, V1);
7912 V2 = DAG.getBitcast(UnpackVT, V2);
7914 // Unpack the inputs and cast the result back to the desired type.
7915 return DAG.getBitcast(
7916 VT, DAG.getNode(UnpackLo ? X86ISD::UNPCKL : X86ISD::UNPCKH, DL,
7920 // We try each unpack from the largest to the smallest to try and find one
7921 // that fits this mask.
7922 int OrigNumElements = VT.getVectorNumElements();
7923 int OrigScalarSize = VT.getScalarSizeInBits();
7924 for (int ScalarSize = 64; ScalarSize >= OrigScalarSize; ScalarSize /= 2) {
7925 int Scale = ScalarSize / OrigScalarSize;
7926 int NumElements = OrigNumElements / Scale;
7927 MVT UnpackVT = MVT::getVectorVT(MVT::getIntegerVT(ScalarSize), NumElements);
7928 if (SDValue Unpack = TryUnpack(UnpackVT, Scale))
7932 // If none of the unpack-rooted lowerings worked (or were profitable) try an
7934 if (NumLoInputs == 0 || NumHiInputs == 0) {
7935 assert((NumLoInputs > 0 || NumHiInputs > 0) &&
7936 "We have to have *some* inputs!");
7937 int HalfOffset = NumLoInputs == 0 ? Size / 2 : 0;
7939 // FIXME: We could consider the total complexity of the permute of each
7940 // possible unpacking. Or at the least we should consider how many
7941 // half-crossings are created.
7942 // FIXME: We could consider commuting the unpacks.
7944 SmallVector<int, 32> PermMask;
7945 PermMask.assign(Size, -1);
7946 for (int i = 0; i < Size; ++i) {
7950 assert(Mask[i] % Size >= HalfOffset && "Found input from wrong half!");
7953 2 * ((Mask[i] % Size) - HalfOffset) + (Mask[i] < Size ? 0 : 1);
7955 return DAG.getVectorShuffle(
7956 VT, DL, DAG.getNode(NumLoInputs == 0 ? X86ISD::UNPCKH : X86ISD::UNPCKL,
7958 DAG.getUNDEF(VT), PermMask);
7964 /// \brief Handle lowering of 2-lane 64-bit floating point shuffles.
7966 /// This is the basis function for the 2-lane 64-bit shuffles as we have full
7967 /// support for floating point shuffles but not integer shuffles. These
7968 /// instructions will incur a domain crossing penalty on some chips though so
7969 /// it is better to avoid lowering through this for integer vectors where
7971 static SDValue lowerV2F64VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
7972 const X86Subtarget *Subtarget,
7973 SelectionDAG &DAG) {
7975 assert(Op.getSimpleValueType() == MVT::v2f64 && "Bad shuffle type!");
7976 assert(V1.getSimpleValueType() == MVT::v2f64 && "Bad operand type!");
7977 assert(V2.getSimpleValueType() == MVT::v2f64 && "Bad operand type!");
7978 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
7979 ArrayRef<int> Mask = SVOp->getMask();
7980 assert(Mask.size() == 2 && "Unexpected mask size for v2 shuffle!");
7982 if (isSingleInputShuffleMask(Mask)) {
7983 // Use low duplicate instructions for masks that match their pattern.
7984 if (Subtarget->hasSSE3())
7985 if (isShuffleEquivalent(V1, V2, Mask, {0, 0}))
7986 return DAG.getNode(X86ISD::MOVDDUP, DL, MVT::v2f64, V1);
7988 // Straight shuffle of a single input vector. Simulate this by using the
7989 // single input as both of the "inputs" to this instruction..
7990 unsigned SHUFPDMask = (Mask[0] == 1) | ((Mask[1] == 1) << 1);
7992 if (Subtarget->hasAVX()) {
7993 // If we have AVX, we can use VPERMILPS which will allow folding a load
7994 // into the shuffle.
7995 return DAG.getNode(X86ISD::VPERMILPI, DL, MVT::v2f64, V1,
7996 DAG.getConstant(SHUFPDMask, DL, MVT::i8));
7999 return DAG.getNode(X86ISD::SHUFP, DL, MVT::v2f64, V1, V1,
8000 DAG.getConstant(SHUFPDMask, DL, MVT::i8));
8002 assert(Mask[0] >= 0 && Mask[0] < 2 && "Non-canonicalized blend!");
8003 assert(Mask[1] >= 2 && "Non-canonicalized blend!");
8005 // If we have a single input, insert that into V1 if we can do so cheaply.
8006 if ((Mask[0] >= 2) + (Mask[1] >= 2) == 1) {
8007 if (SDValue Insertion = lowerVectorShuffleAsElementInsertion(
8008 DL, MVT::v2f64, V1, V2, Mask, Subtarget, DAG))
8010 // Try inverting the insertion since for v2 masks it is easy to do and we
8011 // can't reliably sort the mask one way or the other.
8012 int InverseMask[2] = {Mask[0] < 0 ? -1 : (Mask[0] ^ 2),
8013 Mask[1] < 0 ? -1 : (Mask[1] ^ 2)};
8014 if (SDValue Insertion = lowerVectorShuffleAsElementInsertion(
8015 DL, MVT::v2f64, V2, V1, InverseMask, Subtarget, DAG))
8019 // Try to use one of the special instruction patterns to handle two common
8020 // blend patterns if a zero-blend above didn't work.
8021 if (isShuffleEquivalent(V1, V2, Mask, {0, 3}) ||
8022 isShuffleEquivalent(V1, V2, Mask, {1, 3}))
8023 if (SDValue V1S = getScalarValueForVectorElement(V1, Mask[0], DAG))
8024 // We can either use a special instruction to load over the low double or
8025 // to move just the low double.
8027 isShuffleFoldableLoad(V1S) ? X86ISD::MOVLPD : X86ISD::MOVSD,
8029 DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, MVT::v2f64, V1S));
8031 if (Subtarget->hasSSE41())
8032 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v2f64, V1, V2, Mask,
8036 // Use dedicated unpack instructions for masks that match their pattern.
8037 if (isShuffleEquivalent(V1, V2, Mask, {0, 2}))
8038 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v2f64, V1, V2);
8039 if (isShuffleEquivalent(V1, V2, Mask, {1, 3}))
8040 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v2f64, V1, V2);
8042 unsigned SHUFPDMask = (Mask[0] == 1) | (((Mask[1] - 2) == 1) << 1);
8043 return DAG.getNode(X86ISD::SHUFP, DL, MVT::v2f64, V1, V2,
8044 DAG.getConstant(SHUFPDMask, DL, MVT::i8));
8047 /// \brief Handle lowering of 2-lane 64-bit integer shuffles.
8049 /// Tries to lower a 2-lane 64-bit shuffle using shuffle operations provided by
8050 /// the integer unit to minimize domain crossing penalties. However, for blends
8051 /// it falls back to the floating point shuffle operation with appropriate bit
8053 static SDValue lowerV2I64VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
8054 const X86Subtarget *Subtarget,
8055 SelectionDAG &DAG) {
8057 assert(Op.getSimpleValueType() == MVT::v2i64 && "Bad shuffle type!");
8058 assert(V1.getSimpleValueType() == MVT::v2i64 && "Bad operand type!");
8059 assert(V2.getSimpleValueType() == MVT::v2i64 && "Bad operand type!");
8060 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
8061 ArrayRef<int> Mask = SVOp->getMask();
8062 assert(Mask.size() == 2 && "Unexpected mask size for v2 shuffle!");
8064 if (isSingleInputShuffleMask(Mask)) {
8065 // Check for being able to broadcast a single element.
8066 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v2i64, V1,
8067 Mask, Subtarget, DAG))
8070 // Straight shuffle of a single input vector. For everything from SSE2
8071 // onward this has a single fast instruction with no scary immediates.
8072 // We have to map the mask as it is actually a v4i32 shuffle instruction.
8073 V1 = DAG.getBitcast(MVT::v4i32, V1);
8074 int WidenedMask[4] = {
8075 std::max(Mask[0], 0) * 2, std::max(Mask[0], 0) * 2 + 1,
8076 std::max(Mask[1], 0) * 2, std::max(Mask[1], 0) * 2 + 1};
8077 return DAG.getBitcast(
8079 DAG.getNode(X86ISD::PSHUFD, DL, MVT::v4i32, V1,
8080 getV4X86ShuffleImm8ForMask(WidenedMask, DL, DAG)));
8082 assert(Mask[0] != -1 && "No undef lanes in multi-input v2 shuffles!");
8083 assert(Mask[1] != -1 && "No undef lanes in multi-input v2 shuffles!");
8084 assert(Mask[0] < 2 && "We sort V1 to be the first input.");
8085 assert(Mask[1] >= 2 && "We sort V2 to be the second input.");
8087 // If we have a blend of two PACKUS operations an the blend aligns with the
8088 // low and half halves, we can just merge the PACKUS operations. This is
8089 // particularly important as it lets us merge shuffles that this routine itself
8091 auto GetPackNode = [](SDValue V) {
8092 while (V.getOpcode() == ISD::BITCAST)
8093 V = V.getOperand(0);
8095 return V.getOpcode() == X86ISD::PACKUS ? V : SDValue();
8097 if (SDValue V1Pack = GetPackNode(V1))
8098 if (SDValue V2Pack = GetPackNode(V2))
8099 return DAG.getBitcast(MVT::v2i64,
8100 DAG.getNode(X86ISD::PACKUS, DL, MVT::v16i8,
8101 Mask[0] == 0 ? V1Pack.getOperand(0)
8102 : V1Pack.getOperand(1),
8103 Mask[1] == 2 ? V2Pack.getOperand(0)
8104 : V2Pack.getOperand(1)));
8106 // Try to use shift instructions.
8108 lowerVectorShuffleAsShift(DL, MVT::v2i64, V1, V2, Mask, DAG))
8111 // When loading a scalar and then shuffling it into a vector we can often do
8112 // the insertion cheaply.
8113 if (SDValue Insertion = lowerVectorShuffleAsElementInsertion(
8114 DL, MVT::v2i64, V1, V2, Mask, Subtarget, DAG))
8116 // Try inverting the insertion since for v2 masks it is easy to do and we
8117 // can't reliably sort the mask one way or the other.
8118 int InverseMask[2] = {Mask[0] ^ 2, Mask[1] ^ 2};
8119 if (SDValue Insertion = lowerVectorShuffleAsElementInsertion(
8120 DL, MVT::v2i64, V2, V1, InverseMask, Subtarget, DAG))
8123 // We have different paths for blend lowering, but they all must use the
8124 // *exact* same predicate.
8125 bool IsBlendSupported = Subtarget->hasSSE41();
8126 if (IsBlendSupported)
8127 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v2i64, V1, V2, Mask,
8131 // Use dedicated unpack instructions for masks that match their pattern.
8132 if (isShuffleEquivalent(V1, V2, Mask, {0, 2}))
8133 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v2i64, V1, V2);
8134 if (isShuffleEquivalent(V1, V2, Mask, {1, 3}))
8135 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v2i64, V1, V2);
8137 // Try to use byte rotation instructions.
8138 // Its more profitable for pre-SSSE3 to use shuffles/unpacks.
8139 if (Subtarget->hasSSSE3())
8140 if (SDValue Rotate = lowerVectorShuffleAsByteRotate(
8141 DL, MVT::v2i64, V1, V2, Mask, Subtarget, DAG))
8144 // If we have direct support for blends, we should lower by decomposing into
8145 // a permute. That will be faster than the domain cross.
8146 if (IsBlendSupported)
8147 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v2i64, V1, V2,
8150 // We implement this with SHUFPD which is pretty lame because it will likely
8151 // incur 2 cycles of stall for integer vectors on Nehalem and older chips.
8152 // However, all the alternatives are still more cycles and newer chips don't
8153 // have this problem. It would be really nice if x86 had better shuffles here.
8154 V1 = DAG.getBitcast(MVT::v2f64, V1);
8155 V2 = DAG.getBitcast(MVT::v2f64, V2);
8156 return DAG.getBitcast(MVT::v2i64,
8157 DAG.getVectorShuffle(MVT::v2f64, DL, V1, V2, Mask));
8160 /// \brief Test whether this can be lowered with a single SHUFPS instruction.
8162 /// This is used to disable more specialized lowerings when the shufps lowering
8163 /// will happen to be efficient.
8164 static bool isSingleSHUFPSMask(ArrayRef<int> Mask) {
8165 // This routine only handles 128-bit shufps.
8166 assert(Mask.size() == 4 && "Unsupported mask size!");
8168 // To lower with a single SHUFPS we need to have the low half and high half
8169 // each requiring a single input.
8170 if (Mask[0] != -1 && Mask[1] != -1 && (Mask[0] < 4) != (Mask[1] < 4))
8172 if (Mask[2] != -1 && Mask[3] != -1 && (Mask[2] < 4) != (Mask[3] < 4))
8178 /// \brief Lower a vector shuffle using the SHUFPS instruction.
8180 /// This is a helper routine dedicated to lowering vector shuffles using SHUFPS.
8181 /// It makes no assumptions about whether this is the *best* lowering, it simply
8183 static SDValue lowerVectorShuffleWithSHUFPS(SDLoc DL, MVT VT,
8184 ArrayRef<int> Mask, SDValue V1,
8185 SDValue V2, SelectionDAG &DAG) {
8186 SDValue LowV = V1, HighV = V2;
8187 int NewMask[4] = {Mask[0], Mask[1], Mask[2], Mask[3]};
8190 std::count_if(Mask.begin(), Mask.end(), [](int M) { return M >= 4; });
8192 if (NumV2Elements == 1) {
8194 std::find_if(Mask.begin(), Mask.end(), [](int M) { return M >= 4; }) -
8197 // Compute the index adjacent to V2Index and in the same half by toggling
8199 int V2AdjIndex = V2Index ^ 1;
8201 if (Mask[V2AdjIndex] == -1) {
8202 // Handles all the cases where we have a single V2 element and an undef.
8203 // This will only ever happen in the high lanes because we commute the
8204 // vector otherwise.
8206 std::swap(LowV, HighV);
8207 NewMask[V2Index] -= 4;
8209 // Handle the case where the V2 element ends up adjacent to a V1 element.
8210 // To make this work, blend them together as the first step.
8211 int V1Index = V2AdjIndex;
8212 int BlendMask[4] = {Mask[V2Index] - 4, 0, Mask[V1Index], 0};
8213 V2 = DAG.getNode(X86ISD::SHUFP, DL, VT, V2, V1,
8214 getV4X86ShuffleImm8ForMask(BlendMask, DL, DAG));
8216 // Now proceed to reconstruct the final blend as we have the necessary
8217 // high or low half formed.
8224 NewMask[V1Index] = 2; // We put the V1 element in V2[2].
8225 NewMask[V2Index] = 0; // We shifted the V2 element into V2[0].
8227 } else if (NumV2Elements == 2) {
8228 if (Mask[0] < 4 && Mask[1] < 4) {
8229 // Handle the easy case where we have V1 in the low lanes and V2 in the
8233 } else if (Mask[2] < 4 && Mask[3] < 4) {
8234 // We also handle the reversed case because this utility may get called
8235 // when we detect a SHUFPS pattern but can't easily commute the shuffle to
8236 // arrange things in the right direction.
8242 // We have a mixture of V1 and V2 in both low and high lanes. Rather than
8243 // trying to place elements directly, just blend them and set up the final
8244 // shuffle to place them.
8246 // The first two blend mask elements are for V1, the second two are for
8248 int BlendMask[4] = {Mask[0] < 4 ? Mask[0] : Mask[1],
8249 Mask[2] < 4 ? Mask[2] : Mask[3],
8250 (Mask[0] >= 4 ? Mask[0] : Mask[1]) - 4,
8251 (Mask[2] >= 4 ? Mask[2] : Mask[3]) - 4};
8252 V1 = DAG.getNode(X86ISD::SHUFP, DL, VT, V1, V2,
8253 getV4X86ShuffleImm8ForMask(BlendMask, DL, DAG));
8255 // Now we do a normal shuffle of V1 by giving V1 as both operands to
8258 NewMask[0] = Mask[0] < 4 ? 0 : 2;
8259 NewMask[1] = Mask[0] < 4 ? 2 : 0;
8260 NewMask[2] = Mask[2] < 4 ? 1 : 3;
8261 NewMask[3] = Mask[2] < 4 ? 3 : 1;
8264 return DAG.getNode(X86ISD::SHUFP, DL, VT, LowV, HighV,
8265 getV4X86ShuffleImm8ForMask(NewMask, DL, DAG));
8268 /// \brief Lower 4-lane 32-bit floating point shuffles.
8270 /// Uses instructions exclusively from the floating point unit to minimize
8271 /// domain crossing penalties, as these are sufficient to implement all v4f32
8273 static SDValue lowerV4F32VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
8274 const X86Subtarget *Subtarget,
8275 SelectionDAG &DAG) {
8277 assert(Op.getSimpleValueType() == MVT::v4f32 && "Bad shuffle type!");
8278 assert(V1.getSimpleValueType() == MVT::v4f32 && "Bad operand type!");
8279 assert(V2.getSimpleValueType() == MVT::v4f32 && "Bad operand type!");
8280 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
8281 ArrayRef<int> Mask = SVOp->getMask();
8282 assert(Mask.size() == 4 && "Unexpected mask size for v4 shuffle!");
8285 std::count_if(Mask.begin(), Mask.end(), [](int M) { return M >= 4; });
8287 if (NumV2Elements == 0) {
8288 // Check for being able to broadcast a single element.
8289 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v4f32, V1,
8290 Mask, Subtarget, DAG))
8293 // Use even/odd duplicate instructions for masks that match their pattern.
8294 if (Subtarget->hasSSE3()) {
8295 if (isShuffleEquivalent(V1, V2, Mask, {0, 0, 2, 2}))
8296 return DAG.getNode(X86ISD::MOVSLDUP, DL, MVT::v4f32, V1);
8297 if (isShuffleEquivalent(V1, V2, Mask, {1, 1, 3, 3}))
8298 return DAG.getNode(X86ISD::MOVSHDUP, DL, MVT::v4f32, V1);
8301 if (Subtarget->hasAVX()) {
8302 // If we have AVX, we can use VPERMILPS which will allow folding a load
8303 // into the shuffle.
8304 return DAG.getNode(X86ISD::VPERMILPI, DL, MVT::v4f32, V1,
8305 getV4X86ShuffleImm8ForMask(Mask, DL, DAG));
8308 // Otherwise, use a straight shuffle of a single input vector. We pass the
8309 // input vector to both operands to simulate this with a SHUFPS.
8310 return DAG.getNode(X86ISD::SHUFP, DL, MVT::v4f32, V1, V1,
8311 getV4X86ShuffleImm8ForMask(Mask, DL, DAG));
8314 // There are special ways we can lower some single-element blends. However, we
8315 // have custom ways we can lower more complex single-element blends below that
8316 // we defer to if both this and BLENDPS fail to match, so restrict this to
8317 // when the V2 input is targeting element 0 of the mask -- that is the fast
8319 if (NumV2Elements == 1 && Mask[0] >= 4)
8320 if (SDValue V = lowerVectorShuffleAsElementInsertion(DL, MVT::v4f32, V1, V2,
8321 Mask, Subtarget, DAG))
8324 if (Subtarget->hasSSE41()) {
8325 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v4f32, V1, V2, Mask,
8329 // Use INSERTPS if we can complete the shuffle efficiently.
8330 if (SDValue V = lowerVectorShuffleAsInsertPS(Op, V1, V2, Mask, DAG))
8333 if (!isSingleSHUFPSMask(Mask))
8334 if (SDValue BlendPerm = lowerVectorShuffleAsBlendAndPermute(
8335 DL, MVT::v4f32, V1, V2, Mask, DAG))
8339 // Use dedicated unpack instructions for masks that match their pattern.
8340 if (isShuffleEquivalent(V1, V2, Mask, {0, 4, 1, 5}))
8341 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v4f32, V1, V2);
8342 if (isShuffleEquivalent(V1, V2, Mask, {2, 6, 3, 7}))
8343 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v4f32, V1, V2);
8344 if (isShuffleEquivalent(V1, V2, Mask, {4, 0, 5, 1}))
8345 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v4f32, V2, V1);
8346 if (isShuffleEquivalent(V1, V2, Mask, {6, 2, 7, 3}))
8347 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v4f32, V2, V1);
8349 // Otherwise fall back to a SHUFPS lowering strategy.
8350 return lowerVectorShuffleWithSHUFPS(DL, MVT::v4f32, Mask, V1, V2, DAG);
8353 /// \brief Lower 4-lane i32 vector shuffles.
8355 /// We try to handle these with integer-domain shuffles where we can, but for
8356 /// blends we use the floating point domain blend instructions.
8357 static SDValue lowerV4I32VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
8358 const X86Subtarget *Subtarget,
8359 SelectionDAG &DAG) {
8361 assert(Op.getSimpleValueType() == MVT::v4i32 && "Bad shuffle type!");
8362 assert(V1.getSimpleValueType() == MVT::v4i32 && "Bad operand type!");
8363 assert(V2.getSimpleValueType() == MVT::v4i32 && "Bad operand type!");
8364 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
8365 ArrayRef<int> Mask = SVOp->getMask();
8366 assert(Mask.size() == 4 && "Unexpected mask size for v4 shuffle!");
8368 // Whenever we can lower this as a zext, that instruction is strictly faster
8369 // than any alternative. It also allows us to fold memory operands into the
8370 // shuffle in many cases.
8371 if (SDValue ZExt = lowerVectorShuffleAsZeroOrAnyExtend(DL, MVT::v4i32, V1, V2,
8372 Mask, Subtarget, DAG))
8376 std::count_if(Mask.begin(), Mask.end(), [](int M) { return M >= 4; });
8378 if (NumV2Elements == 0) {
8379 // Check for being able to broadcast a single element.
8380 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v4i32, V1,
8381 Mask, Subtarget, DAG))
8384 // Straight shuffle of a single input vector. For everything from SSE2
8385 // onward this has a single fast instruction with no scary immediates.
8386 // We coerce the shuffle pattern to be compatible with UNPCK instructions
8387 // but we aren't actually going to use the UNPCK instruction because doing
8388 // so prevents folding a load into this instruction or making a copy.
8389 const int UnpackLoMask[] = {0, 0, 1, 1};
8390 const int UnpackHiMask[] = {2, 2, 3, 3};
8391 if (isShuffleEquivalent(V1, V2, Mask, {0, 0, 1, 1}))
8392 Mask = UnpackLoMask;
8393 else if (isShuffleEquivalent(V1, V2, Mask, {2, 2, 3, 3}))
8394 Mask = UnpackHiMask;
8396 return DAG.getNode(X86ISD::PSHUFD, DL, MVT::v4i32, V1,
8397 getV4X86ShuffleImm8ForMask(Mask, DL, DAG));
8400 // Try to use shift instructions.
8402 lowerVectorShuffleAsShift(DL, MVT::v4i32, V1, V2, Mask, DAG))
8405 // There are special ways we can lower some single-element blends.
8406 if (NumV2Elements == 1)
8407 if (SDValue V = lowerVectorShuffleAsElementInsertion(DL, MVT::v4i32, V1, V2,
8408 Mask, Subtarget, DAG))
8411 // We have different paths for blend lowering, but they all must use the
8412 // *exact* same predicate.
8413 bool IsBlendSupported = Subtarget->hasSSE41();
8414 if (IsBlendSupported)
8415 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v4i32, V1, V2, Mask,
8419 if (SDValue Masked =
8420 lowerVectorShuffleAsBitMask(DL, MVT::v4i32, V1, V2, Mask, DAG))
8423 // Use dedicated unpack instructions for masks that match their pattern.
8424 if (isShuffleEquivalent(V1, V2, Mask, {0, 4, 1, 5}))
8425 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v4i32, V1, V2);
8426 if (isShuffleEquivalent(V1, V2, Mask, {2, 6, 3, 7}))
8427 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v4i32, V1, V2);
8428 if (isShuffleEquivalent(V1, V2, Mask, {4, 0, 5, 1}))
8429 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v4i32, V2, V1);
8430 if (isShuffleEquivalent(V1, V2, Mask, {6, 2, 7, 3}))
8431 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v4i32, V2, V1);
8433 // Try to use byte rotation instructions.
8434 // Its more profitable for pre-SSSE3 to use shuffles/unpacks.
8435 if (Subtarget->hasSSSE3())
8436 if (SDValue Rotate = lowerVectorShuffleAsByteRotate(
8437 DL, MVT::v4i32, V1, V2, Mask, Subtarget, DAG))
8440 // If we have direct support for blends, we should lower by decomposing into
8441 // a permute. That will be faster than the domain cross.
8442 if (IsBlendSupported)
8443 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v4i32, V1, V2,
8446 // Try to lower by permuting the inputs into an unpack instruction.
8447 if (SDValue Unpack = lowerVectorShuffleAsPermuteAndUnpack(DL, MVT::v4i32, V1,
8451 // We implement this with SHUFPS because it can blend from two vectors.
8452 // Because we're going to eventually use SHUFPS, we use SHUFPS even to build
8453 // up the inputs, bypassing domain shift penalties that we would encur if we
8454 // directly used PSHUFD on Nehalem and older. For newer chips, this isn't
8456 return DAG.getBitcast(
8458 DAG.getVectorShuffle(MVT::v4f32, DL, DAG.getBitcast(MVT::v4f32, V1),
8459 DAG.getBitcast(MVT::v4f32, V2), Mask));
8462 /// \brief Lowering of single-input v8i16 shuffles is the cornerstone of SSE2
8463 /// shuffle lowering, and the most complex part.
8465 /// The lowering strategy is to try to form pairs of input lanes which are
8466 /// targeted at the same half of the final vector, and then use a dword shuffle
8467 /// to place them onto the right half, and finally unpack the paired lanes into
8468 /// their final position.
8470 /// The exact breakdown of how to form these dword pairs and align them on the
8471 /// correct sides is really tricky. See the comments within the function for
8472 /// more of the details.
8474 /// This code also handles repeated 128-bit lanes of v8i16 shuffles, but each
8475 /// lane must shuffle the *exact* same way. In fact, you must pass a v8 Mask to
8476 /// this routine for it to work correctly. To shuffle a 256-bit or 512-bit i16
8477 /// vector, form the analogous 128-bit 8-element Mask.
8478 static SDValue lowerV8I16GeneralSingleInputVectorShuffle(
8479 SDLoc DL, MVT VT, SDValue V, MutableArrayRef<int> Mask,
8480 const X86Subtarget *Subtarget, SelectionDAG &DAG) {
8481 assert(VT.getScalarType() == MVT::i16 && "Bad input type!");
8482 MVT PSHUFDVT = MVT::getVectorVT(MVT::i32, VT.getVectorNumElements() / 2);
8484 assert(Mask.size() == 8 && "Shuffle mask length doen't match!");
8485 MutableArrayRef<int> LoMask = Mask.slice(0, 4);
8486 MutableArrayRef<int> HiMask = Mask.slice(4, 4);
8488 SmallVector<int, 4> LoInputs;
8489 std::copy_if(LoMask.begin(), LoMask.end(), std::back_inserter(LoInputs),
8490 [](int M) { return M >= 0; });
8491 std::sort(LoInputs.begin(), LoInputs.end());
8492 LoInputs.erase(std::unique(LoInputs.begin(), LoInputs.end()), LoInputs.end());
8493 SmallVector<int, 4> HiInputs;
8494 std::copy_if(HiMask.begin(), HiMask.end(), std::back_inserter(HiInputs),
8495 [](int M) { return M >= 0; });
8496 std::sort(HiInputs.begin(), HiInputs.end());
8497 HiInputs.erase(std::unique(HiInputs.begin(), HiInputs.end()), HiInputs.end());
8499 std::lower_bound(LoInputs.begin(), LoInputs.end(), 4) - LoInputs.begin();
8500 int NumHToL = LoInputs.size() - NumLToL;
8502 std::lower_bound(HiInputs.begin(), HiInputs.end(), 4) - HiInputs.begin();
8503 int NumHToH = HiInputs.size() - NumLToH;
8504 MutableArrayRef<int> LToLInputs(LoInputs.data(), NumLToL);
8505 MutableArrayRef<int> LToHInputs(HiInputs.data(), NumLToH);
8506 MutableArrayRef<int> HToLInputs(LoInputs.data() + NumLToL, NumHToL);
8507 MutableArrayRef<int> HToHInputs(HiInputs.data() + NumLToH, NumHToH);
8509 // Simplify the 1-into-3 and 3-into-1 cases with a single pshufd. For all
8510 // such inputs we can swap two of the dwords across the half mark and end up
8511 // with <=2 inputs to each half in each half. Once there, we can fall through
8512 // to the generic code below. For example:
8514 // Input: [a, b, c, d, e, f, g, h] -PSHUFD[0,2,1,3]-> [a, b, e, f, c, d, g, h]
8515 // Mask: [0, 1, 2, 7, 4, 5, 6, 3] -----------------> [0, 1, 4, 7, 2, 3, 6, 5]
8517 // However in some very rare cases we have a 1-into-3 or 3-into-1 on one half
8518 // and an existing 2-into-2 on the other half. In this case we may have to
8519 // pre-shuffle the 2-into-2 half to avoid turning it into a 3-into-1 or
8520 // 1-into-3 which could cause us to cycle endlessly fixing each side in turn.
8521 // Fortunately, we don't have to handle anything but a 2-into-2 pattern
8522 // because any other situation (including a 3-into-1 or 1-into-3 in the other
8523 // half than the one we target for fixing) will be fixed when we re-enter this
8524 // path. We will also combine away any sequence of PSHUFD instructions that
8525 // result into a single instruction. Here is an example of the tricky case:
8527 // Input: [a, b, c, d, e, f, g, h] -PSHUFD[0,2,1,3]-> [a, b, e, f, c, d, g, h]
8528 // Mask: [3, 7, 1, 0, 2, 7, 3, 5] -THIS-IS-BAD!!!!-> [5, 7, 1, 0, 4, 7, 5, 3]
8530 // This now has a 1-into-3 in the high half! Instead, we do two shuffles:
8532 // Input: [a, b, c, d, e, f, g, h] PSHUFHW[0,2,1,3]-> [a, b, c, d, e, g, f, h]
8533 // Mask: [3, 7, 1, 0, 2, 7, 3, 5] -----------------> [3, 7, 1, 0, 2, 7, 3, 6]
8535 // Input: [a, b, c, d, e, g, f, h] -PSHUFD[0,2,1,3]-> [a, b, e, g, c, d, f, h]
8536 // Mask: [3, 7, 1, 0, 2, 7, 3, 6] -----------------> [5, 7, 1, 0, 4, 7, 5, 6]
8538 // The result is fine to be handled by the generic logic.
8539 auto balanceSides = [&](ArrayRef<int> AToAInputs, ArrayRef<int> BToAInputs,
8540 ArrayRef<int> BToBInputs, ArrayRef<int> AToBInputs,
8541 int AOffset, int BOffset) {
8542 assert((AToAInputs.size() == 3 || AToAInputs.size() == 1) &&
8543 "Must call this with A having 3 or 1 inputs from the A half.");
8544 assert((BToAInputs.size() == 1 || BToAInputs.size() == 3) &&
8545 "Must call this with B having 1 or 3 inputs from the B half.");
8546 assert(AToAInputs.size() + BToAInputs.size() == 4 &&
8547 "Must call this with either 3:1 or 1:3 inputs (summing to 4).");
8549 bool ThreeAInputs = AToAInputs.size() == 3;
8551 // Compute the index of dword with only one word among the three inputs in
8552 // a half by taking the sum of the half with three inputs and subtracting
8553 // the sum of the actual three inputs. The difference is the remaining
8556 int &TripleDWord = ThreeAInputs ? ADWord : BDWord;
8557 int &OneInputDWord = ThreeAInputs ? BDWord : ADWord;
8558 int TripleInputOffset = ThreeAInputs ? AOffset : BOffset;
8559 ArrayRef<int> TripleInputs = ThreeAInputs ? AToAInputs : BToAInputs;
8560 int OneInput = ThreeAInputs ? BToAInputs[0] : AToAInputs[0];
8561 int TripleInputSum = 0 + 1 + 2 + 3 + (4 * TripleInputOffset);
8562 int TripleNonInputIdx =
8563 TripleInputSum - std::accumulate(TripleInputs.begin(), TripleInputs.end(), 0);
8564 TripleDWord = TripleNonInputIdx / 2;
8566 // We use xor with one to compute the adjacent DWord to whichever one the
8568 OneInputDWord = (OneInput / 2) ^ 1;
8570 // Check for one tricky case: We're fixing a 3<-1 or a 1<-3 shuffle for AToA
8571 // and BToA inputs. If there is also such a problem with the BToB and AToB
8572 // inputs, we don't try to fix it necessarily -- we'll recurse and see it in
8573 // the next pass. However, if we have a 2<-2 in the BToB and AToB inputs, it
8574 // is essential that we don't *create* a 3<-1 as then we might oscillate.
8575 if (BToBInputs.size() == 2 && AToBInputs.size() == 2) {
8576 // Compute how many inputs will be flipped by swapping these DWords. We
8578 // to balance this to ensure we don't form a 3-1 shuffle in the other
8580 int NumFlippedAToBInputs =
8581 std::count(AToBInputs.begin(), AToBInputs.end(), 2 * ADWord) +
8582 std::count(AToBInputs.begin(), AToBInputs.end(), 2 * ADWord + 1);
8583 int NumFlippedBToBInputs =
8584 std::count(BToBInputs.begin(), BToBInputs.end(), 2 * BDWord) +
8585 std::count(BToBInputs.begin(), BToBInputs.end(), 2 * BDWord + 1);
8586 if ((NumFlippedAToBInputs == 1 &&
8587 (NumFlippedBToBInputs == 0 || NumFlippedBToBInputs == 2)) ||
8588 (NumFlippedBToBInputs == 1 &&
8589 (NumFlippedAToBInputs == 0 || NumFlippedAToBInputs == 2))) {
8590 // We choose whether to fix the A half or B half based on whether that
8591 // half has zero flipped inputs. At zero, we may not be able to fix it
8592 // with that half. We also bias towards fixing the B half because that
8593 // will more commonly be the high half, and we have to bias one way.
8594 auto FixFlippedInputs = [&V, &DL, &Mask, &DAG](int PinnedIdx, int DWord,
8595 ArrayRef<int> Inputs) {
8596 int FixIdx = PinnedIdx ^ 1; // The adjacent slot to the pinned slot.
8597 bool IsFixIdxInput = std::find(Inputs.begin(), Inputs.end(),
8598 PinnedIdx ^ 1) != Inputs.end();
8599 // Determine whether the free index is in the flipped dword or the
8600 // unflipped dword based on where the pinned index is. We use this bit
8601 // in an xor to conditionally select the adjacent dword.
8602 int FixFreeIdx = 2 * (DWord ^ (PinnedIdx / 2 == DWord));
8603 bool IsFixFreeIdxInput = std::find(Inputs.begin(), Inputs.end(),
8604 FixFreeIdx) != Inputs.end();
8605 if (IsFixIdxInput == IsFixFreeIdxInput)
8607 IsFixFreeIdxInput = std::find(Inputs.begin(), Inputs.end(),
8608 FixFreeIdx) != Inputs.end();
8609 assert(IsFixIdxInput != IsFixFreeIdxInput &&
8610 "We need to be changing the number of flipped inputs!");
8611 int PSHUFHalfMask[] = {0, 1, 2, 3};
8612 std::swap(PSHUFHalfMask[FixFreeIdx % 4], PSHUFHalfMask[FixIdx % 4]);
8613 V = DAG.getNode(FixIdx < 4 ? X86ISD::PSHUFLW : X86ISD::PSHUFHW, DL,
8615 getV4X86ShuffleImm8ForMask(PSHUFHalfMask, DL, DAG));
8618 if (M != -1 && M == FixIdx)
8620 else if (M != -1 && M == FixFreeIdx)
8623 if (NumFlippedBToBInputs != 0) {
8625 BToAInputs.size() == 3 ? TripleNonInputIdx : OneInput;
8626 FixFlippedInputs(BPinnedIdx, BDWord, BToBInputs);
8628 assert(NumFlippedAToBInputs != 0 && "Impossible given predicates!");
8629 int APinnedIdx = ThreeAInputs ? TripleNonInputIdx : OneInput;
8630 FixFlippedInputs(APinnedIdx, ADWord, AToBInputs);
8635 int PSHUFDMask[] = {0, 1, 2, 3};
8636 PSHUFDMask[ADWord] = BDWord;
8637 PSHUFDMask[BDWord] = ADWord;
8640 DAG.getNode(X86ISD::PSHUFD, DL, PSHUFDVT, DAG.getBitcast(PSHUFDVT, V),
8641 getV4X86ShuffleImm8ForMask(PSHUFDMask, DL, DAG)));
8643 // Adjust the mask to match the new locations of A and B.
8645 if (M != -1 && M/2 == ADWord)
8646 M = 2 * BDWord + M % 2;
8647 else if (M != -1 && M/2 == BDWord)
8648 M = 2 * ADWord + M % 2;
8650 // Recurse back into this routine to re-compute state now that this isn't
8651 // a 3 and 1 problem.
8652 return lowerV8I16GeneralSingleInputVectorShuffle(DL, VT, V, Mask, Subtarget,
8655 if ((NumLToL == 3 && NumHToL == 1) || (NumLToL == 1 && NumHToL == 3))
8656 return balanceSides(LToLInputs, HToLInputs, HToHInputs, LToHInputs, 0, 4);
8657 else if ((NumHToH == 3 && NumLToH == 1) || (NumHToH == 1 && NumLToH == 3))
8658 return balanceSides(HToHInputs, LToHInputs, LToLInputs, HToLInputs, 4, 0);
8660 // At this point there are at most two inputs to the low and high halves from
8661 // each half. That means the inputs can always be grouped into dwords and
8662 // those dwords can then be moved to the correct half with a dword shuffle.
8663 // We use at most one low and one high word shuffle to collect these paired
8664 // inputs into dwords, and finally a dword shuffle to place them.
8665 int PSHUFLMask[4] = {-1, -1, -1, -1};
8666 int PSHUFHMask[4] = {-1, -1, -1, -1};
8667 int PSHUFDMask[4] = {-1, -1, -1, -1};
8669 // First fix the masks for all the inputs that are staying in their
8670 // original halves. This will then dictate the targets of the cross-half
8672 auto fixInPlaceInputs =
8673 [&PSHUFDMask](ArrayRef<int> InPlaceInputs, ArrayRef<int> IncomingInputs,
8674 MutableArrayRef<int> SourceHalfMask,
8675 MutableArrayRef<int> HalfMask, int HalfOffset) {
8676 if (InPlaceInputs.empty())
8678 if (InPlaceInputs.size() == 1) {
8679 SourceHalfMask[InPlaceInputs[0] - HalfOffset] =
8680 InPlaceInputs[0] - HalfOffset;
8681 PSHUFDMask[InPlaceInputs[0] / 2] = InPlaceInputs[0] / 2;
8684 if (IncomingInputs.empty()) {
8685 // Just fix all of the in place inputs.
8686 for (int Input : InPlaceInputs) {
8687 SourceHalfMask[Input - HalfOffset] = Input - HalfOffset;
8688 PSHUFDMask[Input / 2] = Input / 2;
8693 assert(InPlaceInputs.size() == 2 && "Cannot handle 3 or 4 inputs!");
8694 SourceHalfMask[InPlaceInputs[0] - HalfOffset] =
8695 InPlaceInputs[0] - HalfOffset;
8696 // Put the second input next to the first so that they are packed into
8697 // a dword. We find the adjacent index by toggling the low bit.
8698 int AdjIndex = InPlaceInputs[0] ^ 1;
8699 SourceHalfMask[AdjIndex - HalfOffset] = InPlaceInputs[1] - HalfOffset;
8700 std::replace(HalfMask.begin(), HalfMask.end(), InPlaceInputs[1], AdjIndex);
8701 PSHUFDMask[AdjIndex / 2] = AdjIndex / 2;
8703 fixInPlaceInputs(LToLInputs, HToLInputs, PSHUFLMask, LoMask, 0);
8704 fixInPlaceInputs(HToHInputs, LToHInputs, PSHUFHMask, HiMask, 4);
8706 // Now gather the cross-half inputs and place them into a free dword of
8707 // their target half.
8708 // FIXME: This operation could almost certainly be simplified dramatically to
8709 // look more like the 3-1 fixing operation.
8710 auto moveInputsToRightHalf = [&PSHUFDMask](
8711 MutableArrayRef<int> IncomingInputs, ArrayRef<int> ExistingInputs,
8712 MutableArrayRef<int> SourceHalfMask, MutableArrayRef<int> HalfMask,
8713 MutableArrayRef<int> FinalSourceHalfMask, int SourceOffset,
8715 auto isWordClobbered = [](ArrayRef<int> SourceHalfMask, int Word) {
8716 return SourceHalfMask[Word] != -1 && SourceHalfMask[Word] != Word;
8718 auto isDWordClobbered = [&isWordClobbered](ArrayRef<int> SourceHalfMask,
8720 int LowWord = Word & ~1;
8721 int HighWord = Word | 1;
8722 return isWordClobbered(SourceHalfMask, LowWord) ||
8723 isWordClobbered(SourceHalfMask, HighWord);
8726 if (IncomingInputs.empty())
8729 if (ExistingInputs.empty()) {
8730 // Map any dwords with inputs from them into the right half.
8731 for (int Input : IncomingInputs) {
8732 // If the source half mask maps over the inputs, turn those into
8733 // swaps and use the swapped lane.
8734 if (isWordClobbered(SourceHalfMask, Input - SourceOffset)) {
8735 if (SourceHalfMask[SourceHalfMask[Input - SourceOffset]] == -1) {
8736 SourceHalfMask[SourceHalfMask[Input - SourceOffset]] =
8737 Input - SourceOffset;
8738 // We have to swap the uses in our half mask in one sweep.
8739 for (int &M : HalfMask)
8740 if (M == SourceHalfMask[Input - SourceOffset] + SourceOffset)
8742 else if (M == Input)
8743 M = SourceHalfMask[Input - SourceOffset] + SourceOffset;
8745 assert(SourceHalfMask[SourceHalfMask[Input - SourceOffset]] ==
8746 Input - SourceOffset &&
8747 "Previous placement doesn't match!");
8749 // Note that this correctly re-maps both when we do a swap and when
8750 // we observe the other side of the swap above. We rely on that to
8751 // avoid swapping the members of the input list directly.
8752 Input = SourceHalfMask[Input - SourceOffset] + SourceOffset;
8755 // Map the input's dword into the correct half.
8756 if (PSHUFDMask[(Input - SourceOffset + DestOffset) / 2] == -1)
8757 PSHUFDMask[(Input - SourceOffset + DestOffset) / 2] = Input / 2;
8759 assert(PSHUFDMask[(Input - SourceOffset + DestOffset) / 2] ==
8761 "Previous placement doesn't match!");
8764 // And just directly shift any other-half mask elements to be same-half
8765 // as we will have mirrored the dword containing the element into the
8766 // same position within that half.
8767 for (int &M : HalfMask)
8768 if (M >= SourceOffset && M < SourceOffset + 4) {
8769 M = M - SourceOffset + DestOffset;
8770 assert(M >= 0 && "This should never wrap below zero!");
8775 // Ensure we have the input in a viable dword of its current half. This
8776 // is particularly tricky because the original position may be clobbered
8777 // by inputs being moved and *staying* in that half.
8778 if (IncomingInputs.size() == 1) {
8779 if (isWordClobbered(SourceHalfMask, IncomingInputs[0] - SourceOffset)) {
8780 int InputFixed = std::find(std::begin(SourceHalfMask),
8781 std::end(SourceHalfMask), -1) -
8782 std::begin(SourceHalfMask) + SourceOffset;
8783 SourceHalfMask[InputFixed - SourceOffset] =
8784 IncomingInputs[0] - SourceOffset;
8785 std::replace(HalfMask.begin(), HalfMask.end(), IncomingInputs[0],
8787 IncomingInputs[0] = InputFixed;
8789 } else if (IncomingInputs.size() == 2) {
8790 if (IncomingInputs[0] / 2 != IncomingInputs[1] / 2 ||
8791 isDWordClobbered(SourceHalfMask, IncomingInputs[0] - SourceOffset)) {
8792 // We have two non-adjacent or clobbered inputs we need to extract from
8793 // the source half. To do this, we need to map them into some adjacent
8794 // dword slot in the source mask.
8795 int InputsFixed[2] = {IncomingInputs[0] - SourceOffset,
8796 IncomingInputs[1] - SourceOffset};
8798 // If there is a free slot in the source half mask adjacent to one of
8799 // the inputs, place the other input in it. We use (Index XOR 1) to
8800 // compute an adjacent index.
8801 if (!isWordClobbered(SourceHalfMask, InputsFixed[0]) &&
8802 SourceHalfMask[InputsFixed[0] ^ 1] == -1) {
8803 SourceHalfMask[InputsFixed[0]] = InputsFixed[0];
8804 SourceHalfMask[InputsFixed[0] ^ 1] = InputsFixed[1];
8805 InputsFixed[1] = InputsFixed[0] ^ 1;
8806 } else if (!isWordClobbered(SourceHalfMask, InputsFixed[1]) &&
8807 SourceHalfMask[InputsFixed[1] ^ 1] == -1) {
8808 SourceHalfMask[InputsFixed[1]] = InputsFixed[1];
8809 SourceHalfMask[InputsFixed[1] ^ 1] = InputsFixed[0];
8810 InputsFixed[0] = InputsFixed[1] ^ 1;
8811 } else if (SourceHalfMask[2 * ((InputsFixed[0] / 2) ^ 1)] == -1 &&
8812 SourceHalfMask[2 * ((InputsFixed[0] / 2) ^ 1) + 1] == -1) {
8813 // The two inputs are in the same DWord but it is clobbered and the
8814 // adjacent DWord isn't used at all. Move both inputs to the free
8816 SourceHalfMask[2 * ((InputsFixed[0] / 2) ^ 1)] = InputsFixed[0];
8817 SourceHalfMask[2 * ((InputsFixed[0] / 2) ^ 1) + 1] = InputsFixed[1];
8818 InputsFixed[0] = 2 * ((InputsFixed[0] / 2) ^ 1);
8819 InputsFixed[1] = 2 * ((InputsFixed[0] / 2) ^ 1) + 1;
8821 // The only way we hit this point is if there is no clobbering
8822 // (because there are no off-half inputs to this half) and there is no
8823 // free slot adjacent to one of the inputs. In this case, we have to
8824 // swap an input with a non-input.
8825 for (int i = 0; i < 4; ++i)
8826 assert((SourceHalfMask[i] == -1 || SourceHalfMask[i] == i) &&
8827 "We can't handle any clobbers here!");
8828 assert(InputsFixed[1] != (InputsFixed[0] ^ 1) &&
8829 "Cannot have adjacent inputs here!");
8831 SourceHalfMask[InputsFixed[0] ^ 1] = InputsFixed[1];
8832 SourceHalfMask[InputsFixed[1]] = InputsFixed[0] ^ 1;
8834 // We also have to update the final source mask in this case because
8835 // it may need to undo the above swap.
8836 for (int &M : FinalSourceHalfMask)
8837 if (M == (InputsFixed[0] ^ 1) + SourceOffset)
8838 M = InputsFixed[1] + SourceOffset;
8839 else if (M == InputsFixed[1] + SourceOffset)
8840 M = (InputsFixed[0] ^ 1) + SourceOffset;
8842 InputsFixed[1] = InputsFixed[0] ^ 1;
8845 // Point everything at the fixed inputs.
8846 for (int &M : HalfMask)
8847 if (M == IncomingInputs[0])
8848 M = InputsFixed[0] + SourceOffset;
8849 else if (M == IncomingInputs[1])
8850 M = InputsFixed[1] + SourceOffset;
8852 IncomingInputs[0] = InputsFixed[0] + SourceOffset;
8853 IncomingInputs[1] = InputsFixed[1] + SourceOffset;
8856 llvm_unreachable("Unhandled input size!");
8859 // Now hoist the DWord down to the right half.
8860 int FreeDWord = (PSHUFDMask[DestOffset / 2] == -1 ? 0 : 1) + DestOffset / 2;
8861 assert(PSHUFDMask[FreeDWord] == -1 && "DWord not free");
8862 PSHUFDMask[FreeDWord] = IncomingInputs[0] / 2;
8863 for (int &M : HalfMask)
8864 for (int Input : IncomingInputs)
8866 M = FreeDWord * 2 + Input % 2;
8868 moveInputsToRightHalf(HToLInputs, LToLInputs, PSHUFHMask, LoMask, HiMask,
8869 /*SourceOffset*/ 4, /*DestOffset*/ 0);
8870 moveInputsToRightHalf(LToHInputs, HToHInputs, PSHUFLMask, HiMask, LoMask,
8871 /*SourceOffset*/ 0, /*DestOffset*/ 4);
8873 // Now enact all the shuffles we've computed to move the inputs into their
8875 if (!isNoopShuffleMask(PSHUFLMask))
8876 V = DAG.getNode(X86ISD::PSHUFLW, DL, VT, V,
8877 getV4X86ShuffleImm8ForMask(PSHUFLMask, DL, DAG));
8878 if (!isNoopShuffleMask(PSHUFHMask))
8879 V = DAG.getNode(X86ISD::PSHUFHW, DL, VT, V,
8880 getV4X86ShuffleImm8ForMask(PSHUFHMask, DL, DAG));
8881 if (!isNoopShuffleMask(PSHUFDMask))
8884 DAG.getNode(X86ISD::PSHUFD, DL, PSHUFDVT, DAG.getBitcast(PSHUFDVT, V),
8885 getV4X86ShuffleImm8ForMask(PSHUFDMask, DL, DAG)));
8887 // At this point, each half should contain all its inputs, and we can then
8888 // just shuffle them into their final position.
8889 assert(std::count_if(LoMask.begin(), LoMask.end(),
8890 [](int M) { return M >= 4; }) == 0 &&
8891 "Failed to lift all the high half inputs to the low mask!");
8892 assert(std::count_if(HiMask.begin(), HiMask.end(),
8893 [](int M) { return M >= 0 && M < 4; }) == 0 &&
8894 "Failed to lift all the low half inputs to the high mask!");
8896 // Do a half shuffle for the low mask.
8897 if (!isNoopShuffleMask(LoMask))
8898 V = DAG.getNode(X86ISD::PSHUFLW, DL, VT, V,
8899 getV4X86ShuffleImm8ForMask(LoMask, DL, DAG));
8901 // Do a half shuffle with the high mask after shifting its values down.
8902 for (int &M : HiMask)
8905 if (!isNoopShuffleMask(HiMask))
8906 V = DAG.getNode(X86ISD::PSHUFHW, DL, VT, V,
8907 getV4X86ShuffleImm8ForMask(HiMask, DL, DAG));
8912 /// \brief Helper to form a PSHUFB-based shuffle+blend.
8913 static SDValue lowerVectorShuffleAsPSHUFB(SDLoc DL, MVT VT, SDValue V1,
8914 SDValue V2, ArrayRef<int> Mask,
8915 SelectionDAG &DAG, bool &V1InUse,
8917 SmallBitVector Zeroable = computeZeroableShuffleElements(Mask, V1, V2);
8923 int Size = Mask.size();
8924 int Scale = 16 / Size;
8925 for (int i = 0; i < 16; ++i) {
8926 if (Mask[i / Scale] == -1) {
8927 V1Mask[i] = V2Mask[i] = DAG.getUNDEF(MVT::i8);
8929 const int ZeroMask = 0x80;
8930 int V1Idx = Mask[i / Scale] < Size ? Mask[i / Scale] * Scale + i % Scale
8932 int V2Idx = Mask[i / Scale] < Size
8934 : (Mask[i / Scale] - Size) * Scale + i % Scale;
8935 if (Zeroable[i / Scale])
8936 V1Idx = V2Idx = ZeroMask;
8937 V1Mask[i] = DAG.getConstant(V1Idx, DL, MVT::i8);
8938 V2Mask[i] = DAG.getConstant(V2Idx, DL, MVT::i8);
8939 V1InUse |= (ZeroMask != V1Idx);
8940 V2InUse |= (ZeroMask != V2Idx);
8945 V1 = DAG.getNode(X86ISD::PSHUFB, DL, MVT::v16i8,
8946 DAG.getBitcast(MVT::v16i8, V1),
8947 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v16i8, V1Mask));
8949 V2 = DAG.getNode(X86ISD::PSHUFB, DL, MVT::v16i8,
8950 DAG.getBitcast(MVT::v16i8, V2),
8951 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v16i8, V2Mask));
8953 // If we need shuffled inputs from both, blend the two.
8955 if (V1InUse && V2InUse)
8956 V = DAG.getNode(ISD::OR, DL, MVT::v16i8, V1, V2);
8958 V = V1InUse ? V1 : V2;
8960 // Cast the result back to the correct type.
8961 return DAG.getBitcast(VT, V);
8964 /// \brief Generic lowering of 8-lane i16 shuffles.
8966 /// This handles both single-input shuffles and combined shuffle/blends with
8967 /// two inputs. The single input shuffles are immediately delegated to
8968 /// a dedicated lowering routine.
8970 /// The blends are lowered in one of three fundamental ways. If there are few
8971 /// enough inputs, it delegates to a basic UNPCK-based strategy. If the shuffle
8972 /// of the input is significantly cheaper when lowered as an interleaving of
8973 /// the two inputs, try to interleave them. Otherwise, blend the low and high
8974 /// halves of the inputs separately (making them have relatively few inputs)
8975 /// and then concatenate them.
8976 static SDValue lowerV8I16VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
8977 const X86Subtarget *Subtarget,
8978 SelectionDAG &DAG) {
8980 assert(Op.getSimpleValueType() == MVT::v8i16 && "Bad shuffle type!");
8981 assert(V1.getSimpleValueType() == MVT::v8i16 && "Bad operand type!");
8982 assert(V2.getSimpleValueType() == MVT::v8i16 && "Bad operand type!");
8983 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
8984 ArrayRef<int> OrigMask = SVOp->getMask();
8985 int MaskStorage[8] = {OrigMask[0], OrigMask[1], OrigMask[2], OrigMask[3],
8986 OrigMask[4], OrigMask[5], OrigMask[6], OrigMask[7]};
8987 MutableArrayRef<int> Mask(MaskStorage);
8989 assert(Mask.size() == 8 && "Unexpected mask size for v8 shuffle!");
8991 // Whenever we can lower this as a zext, that instruction is strictly faster
8992 // than any alternative.
8993 if (SDValue ZExt = lowerVectorShuffleAsZeroOrAnyExtend(
8994 DL, MVT::v8i16, V1, V2, OrigMask, Subtarget, DAG))
8997 auto isV1 = [](int M) { return M >= 0 && M < 8; };
8999 auto isV2 = [](int M) { return M >= 8; };
9001 int NumV2Inputs = std::count_if(Mask.begin(), Mask.end(), isV2);
9003 if (NumV2Inputs == 0) {
9004 // Check for being able to broadcast a single element.
9005 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v8i16, V1,
9006 Mask, Subtarget, DAG))
9009 // Try to use shift instructions.
9011 lowerVectorShuffleAsShift(DL, MVT::v8i16, V1, V1, Mask, DAG))
9014 // Use dedicated unpack instructions for masks that match their pattern.
9015 if (isShuffleEquivalent(V1, V1, Mask, {0, 0, 1, 1, 2, 2, 3, 3}))
9016 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v8i16, V1, V1);
9017 if (isShuffleEquivalent(V1, V1, Mask, {4, 4, 5, 5, 6, 6, 7, 7}))
9018 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v8i16, V1, V1);
9020 // Try to use byte rotation instructions.
9021 if (SDValue Rotate = lowerVectorShuffleAsByteRotate(DL, MVT::v8i16, V1, V1,
9022 Mask, Subtarget, DAG))
9025 return lowerV8I16GeneralSingleInputVectorShuffle(DL, MVT::v8i16, V1, Mask,
9029 assert(std::any_of(Mask.begin(), Mask.end(), isV1) &&
9030 "All single-input shuffles should be canonicalized to be V1-input "
9033 // Try to use shift instructions.
9035 lowerVectorShuffleAsShift(DL, MVT::v8i16, V1, V2, Mask, DAG))
9038 // See if we can use SSE4A Extraction / Insertion.
9039 if (Subtarget->hasSSE4A())
9040 if (SDValue V = lowerVectorShuffleWithSSE4A(DL, MVT::v8i16, V1, V2, Mask, DAG))
9043 // There are special ways we can lower some single-element blends.
9044 if (NumV2Inputs == 1)
9045 if (SDValue V = lowerVectorShuffleAsElementInsertion(DL, MVT::v8i16, V1, V2,
9046 Mask, Subtarget, DAG))
9049 // We have different paths for blend lowering, but they all must use the
9050 // *exact* same predicate.
9051 bool IsBlendSupported = Subtarget->hasSSE41();
9052 if (IsBlendSupported)
9053 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v8i16, V1, V2, Mask,
9057 if (SDValue Masked =
9058 lowerVectorShuffleAsBitMask(DL, MVT::v8i16, V1, V2, Mask, DAG))
9061 // Use dedicated unpack instructions for masks that match their pattern.
9062 if (isShuffleEquivalent(V1, V2, Mask, {0, 8, 1, 9, 2, 10, 3, 11}))
9063 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v8i16, V1, V2);
9064 if (isShuffleEquivalent(V1, V2, Mask, {4, 12, 5, 13, 6, 14, 7, 15}))
9065 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v8i16, V1, V2);
9067 // Try to use byte rotation instructions.
9068 if (SDValue Rotate = lowerVectorShuffleAsByteRotate(
9069 DL, MVT::v8i16, V1, V2, Mask, Subtarget, DAG))
9072 if (SDValue BitBlend =
9073 lowerVectorShuffleAsBitBlend(DL, MVT::v8i16, V1, V2, Mask, DAG))
9076 if (SDValue Unpack = lowerVectorShuffleAsPermuteAndUnpack(DL, MVT::v8i16, V1,
9080 // If we can't directly blend but can use PSHUFB, that will be better as it
9081 // can both shuffle and set up the inefficient blend.
9082 if (!IsBlendSupported && Subtarget->hasSSSE3()) {
9083 bool V1InUse, V2InUse;
9084 return lowerVectorShuffleAsPSHUFB(DL, MVT::v8i16, V1, V2, Mask, DAG,
9088 // We can always bit-blend if we have to so the fallback strategy is to
9089 // decompose into single-input permutes and blends.
9090 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v8i16, V1, V2,
9094 /// \brief Check whether a compaction lowering can be done by dropping even
9095 /// elements and compute how many times even elements must be dropped.
9097 /// This handles shuffles which take every Nth element where N is a power of
9098 /// two. Example shuffle masks:
9100 /// N = 1: 0, 2, 4, 6, 8, 10, 12, 14, 0, 2, 4, 6, 8, 10, 12, 14
9101 /// N = 1: 0, 2, 4, 6, 8, 10, 12, 14, 16, 18, 20, 22, 24, 26, 28, 30
9102 /// N = 2: 0, 4, 8, 12, 0, 4, 8, 12, 0, 4, 8, 12, 0, 4, 8, 12
9103 /// N = 2: 0, 4, 8, 12, 16, 20, 24, 28, 0, 4, 8, 12, 16, 20, 24, 28
9104 /// N = 3: 0, 8, 0, 8, 0, 8, 0, 8, 0, 8, 0, 8, 0, 8, 0, 8
9105 /// N = 3: 0, 8, 16, 24, 0, 8, 16, 24, 0, 8, 16, 24, 0, 8, 16, 24
9107 /// Any of these lanes can of course be undef.
9109 /// This routine only supports N <= 3.
9110 /// FIXME: Evaluate whether either AVX or AVX-512 have any opportunities here
9113 /// \returns N above, or the number of times even elements must be dropped if
9114 /// there is such a number. Otherwise returns zero.
9115 static int canLowerByDroppingEvenElements(ArrayRef<int> Mask) {
9116 // Figure out whether we're looping over two inputs or just one.
9117 bool IsSingleInput = isSingleInputShuffleMask(Mask);
9119 // The modulus for the shuffle vector entries is based on whether this is
9120 // a single input or not.
9121 int ShuffleModulus = Mask.size() * (IsSingleInput ? 1 : 2);
9122 assert(isPowerOf2_32((uint32_t)ShuffleModulus) &&
9123 "We should only be called with masks with a power-of-2 size!");
9125 uint64_t ModMask = (uint64_t)ShuffleModulus - 1;
9127 // We track whether the input is viable for all power-of-2 strides 2^1, 2^2,
9128 // and 2^3 simultaneously. This is because we may have ambiguity with
9129 // partially undef inputs.
9130 bool ViableForN[3] = {true, true, true};
9132 for (int i = 0, e = Mask.size(); i < e; ++i) {
9133 // Ignore undef lanes, we'll optimistically collapse them to the pattern we
9138 bool IsAnyViable = false;
9139 for (unsigned j = 0; j != array_lengthof(ViableForN); ++j)
9140 if (ViableForN[j]) {
9143 // The shuffle mask must be equal to (i * 2^N) % M.
9144 if ((uint64_t)Mask[i] == (((uint64_t)i << N) & ModMask))
9147 ViableForN[j] = false;
9149 // Early exit if we exhaust the possible powers of two.
9154 for (unsigned j = 0; j != array_lengthof(ViableForN); ++j)
9158 // Return 0 as there is no viable power of two.
9162 /// \brief Generic lowering of v16i8 shuffles.
9164 /// This is a hybrid strategy to lower v16i8 vectors. It first attempts to
9165 /// detect any complexity reducing interleaving. If that doesn't help, it uses
9166 /// UNPCK to spread the i8 elements across two i16-element vectors, and uses
9167 /// the existing lowering for v8i16 blends on each half, finally PACK-ing them
9169 static SDValue lowerV16I8VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
9170 const X86Subtarget *Subtarget,
9171 SelectionDAG &DAG) {
9173 assert(Op.getSimpleValueType() == MVT::v16i8 && "Bad shuffle type!");
9174 assert(V1.getSimpleValueType() == MVT::v16i8 && "Bad operand type!");
9175 assert(V2.getSimpleValueType() == MVT::v16i8 && "Bad operand type!");
9176 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
9177 ArrayRef<int> Mask = SVOp->getMask();
9178 assert(Mask.size() == 16 && "Unexpected mask size for v16 shuffle!");
9180 // Try to use shift instructions.
9182 lowerVectorShuffleAsShift(DL, MVT::v16i8, V1, V2, Mask, DAG))
9185 // Try to use byte rotation instructions.
9186 if (SDValue Rotate = lowerVectorShuffleAsByteRotate(
9187 DL, MVT::v16i8, V1, V2, Mask, Subtarget, DAG))
9190 // Try to use a zext lowering.
9191 if (SDValue ZExt = lowerVectorShuffleAsZeroOrAnyExtend(
9192 DL, MVT::v16i8, V1, V2, Mask, Subtarget, DAG))
9195 // See if we can use SSE4A Extraction / Insertion.
9196 if (Subtarget->hasSSE4A())
9197 if (SDValue V = lowerVectorShuffleWithSSE4A(DL, MVT::v16i8, V1, V2, Mask, DAG))
9201 std::count_if(Mask.begin(), Mask.end(), [](int M) { return M >= 16; });
9203 // For single-input shuffles, there are some nicer lowering tricks we can use.
9204 if (NumV2Elements == 0) {
9205 // Check for being able to broadcast a single element.
9206 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v16i8, V1,
9207 Mask, Subtarget, DAG))
9210 // Check whether we can widen this to an i16 shuffle by duplicating bytes.
9211 // Notably, this handles splat and partial-splat shuffles more efficiently.
9212 // However, it only makes sense if the pre-duplication shuffle simplifies
9213 // things significantly. Currently, this means we need to be able to
9214 // express the pre-duplication shuffle as an i16 shuffle.
9216 // FIXME: We should check for other patterns which can be widened into an
9217 // i16 shuffle as well.
9218 auto canWidenViaDuplication = [](ArrayRef<int> Mask) {
9219 for (int i = 0; i < 16; i += 2)
9220 if (Mask[i] != -1 && Mask[i + 1] != -1 && Mask[i] != Mask[i + 1])
9225 auto tryToWidenViaDuplication = [&]() -> SDValue {
9226 if (!canWidenViaDuplication(Mask))
9228 SmallVector<int, 4> LoInputs;
9229 std::copy_if(Mask.begin(), Mask.end(), std::back_inserter(LoInputs),
9230 [](int M) { return M >= 0 && M < 8; });
9231 std::sort(LoInputs.begin(), LoInputs.end());
9232 LoInputs.erase(std::unique(LoInputs.begin(), LoInputs.end()),
9234 SmallVector<int, 4> HiInputs;
9235 std::copy_if(Mask.begin(), Mask.end(), std::back_inserter(HiInputs),
9236 [](int M) { return M >= 8; });
9237 std::sort(HiInputs.begin(), HiInputs.end());
9238 HiInputs.erase(std::unique(HiInputs.begin(), HiInputs.end()),
9241 bool TargetLo = LoInputs.size() >= HiInputs.size();
9242 ArrayRef<int> InPlaceInputs = TargetLo ? LoInputs : HiInputs;
9243 ArrayRef<int> MovingInputs = TargetLo ? HiInputs : LoInputs;
9245 int PreDupI16Shuffle[] = {-1, -1, -1, -1, -1, -1, -1, -1};
9246 SmallDenseMap<int, int, 8> LaneMap;
9247 for (int I : InPlaceInputs) {
9248 PreDupI16Shuffle[I/2] = I/2;
9251 int j = TargetLo ? 0 : 4, je = j + 4;
9252 for (int i = 0, ie = MovingInputs.size(); i < ie; ++i) {
9253 // Check if j is already a shuffle of this input. This happens when
9254 // there are two adjacent bytes after we move the low one.
9255 if (PreDupI16Shuffle[j] != MovingInputs[i] / 2) {
9256 // If we haven't yet mapped the input, search for a slot into which
9258 while (j < je && PreDupI16Shuffle[j] != -1)
9262 // We can't place the inputs into a single half with a simple i16 shuffle, so bail.
9265 // Map this input with the i16 shuffle.
9266 PreDupI16Shuffle[j] = MovingInputs[i] / 2;
9269 // Update the lane map based on the mapping we ended up with.
9270 LaneMap[MovingInputs[i]] = 2 * j + MovingInputs[i] % 2;
9272 V1 = DAG.getBitcast(
9274 DAG.getVectorShuffle(MVT::v8i16, DL, DAG.getBitcast(MVT::v8i16, V1),
9275 DAG.getUNDEF(MVT::v8i16), PreDupI16Shuffle));
9277 // Unpack the bytes to form the i16s that will be shuffled into place.
9278 V1 = DAG.getNode(TargetLo ? X86ISD::UNPCKL : X86ISD::UNPCKH, DL,
9279 MVT::v16i8, V1, V1);
9281 int PostDupI16Shuffle[8] = {-1, -1, -1, -1, -1, -1, -1, -1};
9282 for (int i = 0; i < 16; ++i)
9283 if (Mask[i] != -1) {
9284 int MappedMask = LaneMap[Mask[i]] - (TargetLo ? 0 : 8);
9285 assert(MappedMask < 8 && "Invalid v8 shuffle mask!");
9286 if (PostDupI16Shuffle[i / 2] == -1)
9287 PostDupI16Shuffle[i / 2] = MappedMask;
9289 assert(PostDupI16Shuffle[i / 2] == MappedMask &&
9290 "Conflicting entrties in the original shuffle!");
9292 return DAG.getBitcast(
9294 DAG.getVectorShuffle(MVT::v8i16, DL, DAG.getBitcast(MVT::v8i16, V1),
9295 DAG.getUNDEF(MVT::v8i16), PostDupI16Shuffle));
9297 if (SDValue V = tryToWidenViaDuplication())
9301 if (SDValue Masked =
9302 lowerVectorShuffleAsBitMask(DL, MVT::v16i8, V1, V2, Mask, DAG))
9305 // Use dedicated unpack instructions for masks that match their pattern.
9306 if (isShuffleEquivalent(V1, V2, Mask, {// Low half.
9307 0, 16, 1, 17, 2, 18, 3, 19,
9309 4, 20, 5, 21, 6, 22, 7, 23}))
9310 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v16i8, V1, V2);
9311 if (isShuffleEquivalent(V1, V2, Mask, {// Low half.
9312 8, 24, 9, 25, 10, 26, 11, 27,
9314 12, 28, 13, 29, 14, 30, 15, 31}))
9315 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v16i8, V1, V2);
9317 // Check for SSSE3 which lets us lower all v16i8 shuffles much more directly
9318 // with PSHUFB. It is important to do this before we attempt to generate any
9319 // blends but after all of the single-input lowerings. If the single input
9320 // lowerings can find an instruction sequence that is faster than a PSHUFB, we
9321 // want to preserve that and we can DAG combine any longer sequences into
9322 // a PSHUFB in the end. But once we start blending from multiple inputs,
9323 // the complexity of DAG combining bad patterns back into PSHUFB is too high,
9324 // and there are *very* few patterns that would actually be faster than the
9325 // PSHUFB approach because of its ability to zero lanes.
9327 // FIXME: The only exceptions to the above are blends which are exact
9328 // interleavings with direct instructions supporting them. We currently don't
9329 // handle those well here.
9330 if (Subtarget->hasSSSE3()) {
9331 bool V1InUse = false;
9332 bool V2InUse = false;
9334 SDValue PSHUFB = lowerVectorShuffleAsPSHUFB(DL, MVT::v16i8, V1, V2, Mask,
9335 DAG, V1InUse, V2InUse);
9337 // If both V1 and V2 are in use and we can use a direct blend or an unpack,
9338 // do so. This avoids using them to handle blends-with-zero which is
9339 // important as a single pshufb is significantly faster for that.
9340 if (V1InUse && V2InUse) {
9341 if (Subtarget->hasSSE41())
9342 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v16i8, V1, V2,
9343 Mask, Subtarget, DAG))
9346 // We can use an unpack to do the blending rather than an or in some
9347 // cases. Even though the or may be (very minorly) more efficient, we
9348 // preference this lowering because there are common cases where part of
9349 // the complexity of the shuffles goes away when we do the final blend as
9351 // FIXME: It might be worth trying to detect if the unpack-feeding
9352 // shuffles will both be pshufb, in which case we shouldn't bother with
9354 if (SDValue Unpack = lowerVectorShuffleAsPermuteAndUnpack(
9355 DL, MVT::v16i8, V1, V2, Mask, DAG))
9362 // There are special ways we can lower some single-element blends.
9363 if (NumV2Elements == 1)
9364 if (SDValue V = lowerVectorShuffleAsElementInsertion(DL, MVT::v16i8, V1, V2,
9365 Mask, Subtarget, DAG))
9368 if (SDValue BitBlend =
9369 lowerVectorShuffleAsBitBlend(DL, MVT::v16i8, V1, V2, Mask, DAG))
9372 // Check whether a compaction lowering can be done. This handles shuffles
9373 // which take every Nth element for some even N. See the helper function for
9376 // We special case these as they can be particularly efficiently handled with
9377 // the PACKUSB instruction on x86 and they show up in common patterns of
9378 // rearranging bytes to truncate wide elements.
9379 if (int NumEvenDrops = canLowerByDroppingEvenElements(Mask)) {
9380 // NumEvenDrops is the power of two stride of the elements. Another way of
9381 // thinking about it is that we need to drop the even elements this many
9382 // times to get the original input.
9383 bool IsSingleInput = isSingleInputShuffleMask(Mask);
9385 // First we need to zero all the dropped bytes.
9386 assert(NumEvenDrops <= 3 &&
9387 "No support for dropping even elements more than 3 times.");
9388 // We use the mask type to pick which bytes are preserved based on how many
9389 // elements are dropped.
9390 MVT MaskVTs[] = { MVT::v8i16, MVT::v4i32, MVT::v2i64 };
9391 SDValue ByteClearMask = DAG.getBitcast(
9392 MVT::v16i8, DAG.getConstant(0xFF, DL, MaskVTs[NumEvenDrops - 1]));
9393 V1 = DAG.getNode(ISD::AND, DL, MVT::v16i8, V1, ByteClearMask);
9395 V2 = DAG.getNode(ISD::AND, DL, MVT::v16i8, V2, ByteClearMask);
9397 // Now pack things back together.
9398 V1 = DAG.getBitcast(MVT::v8i16, V1);
9399 V2 = IsSingleInput ? V1 : DAG.getBitcast(MVT::v8i16, V2);
9400 SDValue Result = DAG.getNode(X86ISD::PACKUS, DL, MVT::v16i8, V1, V2);
9401 for (int i = 1; i < NumEvenDrops; ++i) {
9402 Result = DAG.getBitcast(MVT::v8i16, Result);
9403 Result = DAG.getNode(X86ISD::PACKUS, DL, MVT::v16i8, Result, Result);
9409 // Handle multi-input cases by blending single-input shuffles.
9410 if (NumV2Elements > 0)
9411 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v16i8, V1, V2,
9414 // The fallback path for single-input shuffles widens this into two v8i16
9415 // vectors with unpacks, shuffles those, and then pulls them back together
9419 int LoBlendMask[8] = {-1, -1, -1, -1, -1, -1, -1, -1};
9420 int HiBlendMask[8] = {-1, -1, -1, -1, -1, -1, -1, -1};
9421 for (int i = 0; i < 16; ++i)
9423 (i < 8 ? LoBlendMask[i] : HiBlendMask[i % 8]) = Mask[i];
9425 SDValue Zero = getZeroVector(MVT::v8i16, Subtarget, DAG, DL);
9427 SDValue VLoHalf, VHiHalf;
9428 // Check if any of the odd lanes in the v16i8 are used. If not, we can mask
9429 // them out and avoid using UNPCK{L,H} to extract the elements of V as
9431 if (std::none_of(std::begin(LoBlendMask), std::end(LoBlendMask),
9432 [](int M) { return M >= 0 && M % 2 == 1; }) &&
9433 std::none_of(std::begin(HiBlendMask), std::end(HiBlendMask),
9434 [](int M) { return M >= 0 && M % 2 == 1; })) {
9435 // Use a mask to drop the high bytes.
9436 VLoHalf = DAG.getBitcast(MVT::v8i16, V);
9437 VLoHalf = DAG.getNode(ISD::AND, DL, MVT::v8i16, VLoHalf,
9438 DAG.getConstant(0x00FF, DL, MVT::v8i16));
9440 // This will be a single vector shuffle instead of a blend so nuke VHiHalf.
9441 VHiHalf = DAG.getUNDEF(MVT::v8i16);
9443 // Squash the masks to point directly into VLoHalf.
9444 for (int &M : LoBlendMask)
9447 for (int &M : HiBlendMask)
9451 // Otherwise just unpack the low half of V into VLoHalf and the high half into
9452 // VHiHalf so that we can blend them as i16s.
9453 VLoHalf = DAG.getBitcast(
9454 MVT::v8i16, DAG.getNode(X86ISD::UNPCKL, DL, MVT::v16i8, V, Zero));
9455 VHiHalf = DAG.getBitcast(
9456 MVT::v8i16, DAG.getNode(X86ISD::UNPCKH, DL, MVT::v16i8, V, Zero));
9459 SDValue LoV = DAG.getVectorShuffle(MVT::v8i16, DL, VLoHalf, VHiHalf, LoBlendMask);
9460 SDValue HiV = DAG.getVectorShuffle(MVT::v8i16, DL, VLoHalf, VHiHalf, HiBlendMask);
9462 return DAG.getNode(X86ISD::PACKUS, DL, MVT::v16i8, LoV, HiV);
9465 /// \brief Dispatching routine to lower various 128-bit x86 vector shuffles.
9467 /// This routine breaks down the specific type of 128-bit shuffle and
9468 /// dispatches to the lowering routines accordingly.
9469 static SDValue lower128BitVectorShuffle(SDValue Op, SDValue V1, SDValue V2,
9470 MVT VT, const X86Subtarget *Subtarget,
9471 SelectionDAG &DAG) {
9472 switch (VT.SimpleTy) {
9474 return lowerV2I64VectorShuffle(Op, V1, V2, Subtarget, DAG);
9476 return lowerV2F64VectorShuffle(Op, V1, V2, Subtarget, DAG);
9478 return lowerV4I32VectorShuffle(Op, V1, V2, Subtarget, DAG);
9480 return lowerV4F32VectorShuffle(Op, V1, V2, Subtarget, DAG);
9482 return lowerV8I16VectorShuffle(Op, V1, V2, Subtarget, DAG);
9484 return lowerV16I8VectorShuffle(Op, V1, V2, Subtarget, DAG);
9487 llvm_unreachable("Unimplemented!");
9491 /// \brief Helper function to test whether a shuffle mask could be
9492 /// simplified by widening the elements being shuffled.
9494 /// Appends the mask for wider elements in WidenedMask if valid. Otherwise
9495 /// leaves it in an unspecified state.
9497 /// NOTE: This must handle normal vector shuffle masks and *target* vector
9498 /// shuffle masks. The latter have the special property of a '-2' representing
9499 /// a zero-ed lane of a vector.
9500 static bool canWidenShuffleElements(ArrayRef<int> Mask,
9501 SmallVectorImpl<int> &WidenedMask) {
9502 for (int i = 0, Size = Mask.size(); i < Size; i += 2) {
9503 // If both elements are undef, its trivial.
9504 if (Mask[i] == SM_SentinelUndef && Mask[i + 1] == SM_SentinelUndef) {
9505 WidenedMask.push_back(SM_SentinelUndef);
9509 // Check for an undef mask and a mask value properly aligned to fit with
9510 // a pair of values. If we find such a case, use the non-undef mask's value.
9511 if (Mask[i] == SM_SentinelUndef && Mask[i + 1] >= 0 && Mask[i + 1] % 2 == 1) {
9512 WidenedMask.push_back(Mask[i + 1] / 2);
9515 if (Mask[i + 1] == SM_SentinelUndef && Mask[i] >= 0 && Mask[i] % 2 == 0) {
9516 WidenedMask.push_back(Mask[i] / 2);
9520 // When zeroing, we need to spread the zeroing across both lanes to widen.
9521 if (Mask[i] == SM_SentinelZero || Mask[i + 1] == SM_SentinelZero) {
9522 if ((Mask[i] == SM_SentinelZero || Mask[i] == SM_SentinelUndef) &&
9523 (Mask[i + 1] == SM_SentinelZero || Mask[i + 1] == SM_SentinelUndef)) {
9524 WidenedMask.push_back(SM_SentinelZero);
9530 // Finally check if the two mask values are adjacent and aligned with
9532 if (Mask[i] != SM_SentinelUndef && Mask[i] % 2 == 0 && Mask[i] + 1 == Mask[i + 1]) {
9533 WidenedMask.push_back(Mask[i] / 2);
9537 // Otherwise we can't safely widen the elements used in this shuffle.
9540 assert(WidenedMask.size() == Mask.size() / 2 &&
9541 "Incorrect size of mask after widening the elements!");
9546 /// \brief Generic routine to split vector shuffle into half-sized shuffles.
9548 /// This routine just extracts two subvectors, shuffles them independently, and
9549 /// then concatenates them back together. This should work effectively with all
9550 /// AVX vector shuffle types.
9551 static SDValue splitAndLowerVectorShuffle(SDLoc DL, MVT VT, SDValue V1,
9552 SDValue V2, ArrayRef<int> Mask,
9553 SelectionDAG &DAG) {
9554 assert(VT.getSizeInBits() >= 256 &&
9555 "Only for 256-bit or wider vector shuffles!");
9556 assert(V1.getSimpleValueType() == VT && "Bad operand type!");
9557 assert(V2.getSimpleValueType() == VT && "Bad operand type!");
9559 ArrayRef<int> LoMask = Mask.slice(0, Mask.size() / 2);
9560 ArrayRef<int> HiMask = Mask.slice(Mask.size() / 2);
9562 int NumElements = VT.getVectorNumElements();
9563 int SplitNumElements = NumElements / 2;
9564 MVT ScalarVT = VT.getScalarType();
9565 MVT SplitVT = MVT::getVectorVT(ScalarVT, NumElements / 2);
9567 // Rather than splitting build-vectors, just build two narrower build
9568 // vectors. This helps shuffling with splats and zeros.
9569 auto SplitVector = [&](SDValue V) {
9570 while (V.getOpcode() == ISD::BITCAST)
9571 V = V->getOperand(0);
9573 MVT OrigVT = V.getSimpleValueType();
9574 int OrigNumElements = OrigVT.getVectorNumElements();
9575 int OrigSplitNumElements = OrigNumElements / 2;
9576 MVT OrigScalarVT = OrigVT.getScalarType();
9577 MVT OrigSplitVT = MVT::getVectorVT(OrigScalarVT, OrigNumElements / 2);
9581 auto *BV = dyn_cast<BuildVectorSDNode>(V);
9583 LoV = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, OrigSplitVT, V,
9584 DAG.getIntPtrConstant(0, DL));
9585 HiV = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, OrigSplitVT, V,
9586 DAG.getIntPtrConstant(OrigSplitNumElements, DL));
9589 SmallVector<SDValue, 16> LoOps, HiOps;
9590 for (int i = 0; i < OrigSplitNumElements; ++i) {
9591 LoOps.push_back(BV->getOperand(i));
9592 HiOps.push_back(BV->getOperand(i + OrigSplitNumElements));
9594 LoV = DAG.getNode(ISD::BUILD_VECTOR, DL, OrigSplitVT, LoOps);
9595 HiV = DAG.getNode(ISD::BUILD_VECTOR, DL, OrigSplitVT, HiOps);
9597 return std::make_pair(DAG.getBitcast(SplitVT, LoV),
9598 DAG.getBitcast(SplitVT, HiV));
9601 SDValue LoV1, HiV1, LoV2, HiV2;
9602 std::tie(LoV1, HiV1) = SplitVector(V1);
9603 std::tie(LoV2, HiV2) = SplitVector(V2);
9605 // Now create two 4-way blends of these half-width vectors.
9606 auto HalfBlend = [&](ArrayRef<int> HalfMask) {
9607 bool UseLoV1 = false, UseHiV1 = false, UseLoV2 = false, UseHiV2 = false;
9608 SmallVector<int, 32> V1BlendMask, V2BlendMask, BlendMask;
9609 for (int i = 0; i < SplitNumElements; ++i) {
9610 int M = HalfMask[i];
9611 if (M >= NumElements) {
9612 if (M >= NumElements + SplitNumElements)
9616 V2BlendMask.push_back(M - NumElements);
9617 V1BlendMask.push_back(-1);
9618 BlendMask.push_back(SplitNumElements + i);
9619 } else if (M >= 0) {
9620 if (M >= SplitNumElements)
9624 V2BlendMask.push_back(-1);
9625 V1BlendMask.push_back(M);
9626 BlendMask.push_back(i);
9628 V2BlendMask.push_back(-1);
9629 V1BlendMask.push_back(-1);
9630 BlendMask.push_back(-1);
9634 // Because the lowering happens after all combining takes place, we need to
9635 // manually combine these blend masks as much as possible so that we create
9636 // a minimal number of high-level vector shuffle nodes.
9638 // First try just blending the halves of V1 or V2.
9639 if (!UseLoV1 && !UseHiV1 && !UseLoV2 && !UseHiV2)
9640 return DAG.getUNDEF(SplitVT);
9641 if (!UseLoV2 && !UseHiV2)
9642 return DAG.getVectorShuffle(SplitVT, DL, LoV1, HiV1, V1BlendMask);
9643 if (!UseLoV1 && !UseHiV1)
9644 return DAG.getVectorShuffle(SplitVT, DL, LoV2, HiV2, V2BlendMask);
9646 SDValue V1Blend, V2Blend;
9647 if (UseLoV1 && UseHiV1) {
9649 DAG.getVectorShuffle(SplitVT, DL, LoV1, HiV1, V1BlendMask);
9651 // We only use half of V1 so map the usage down into the final blend mask.
9652 V1Blend = UseLoV1 ? LoV1 : HiV1;
9653 for (int i = 0; i < SplitNumElements; ++i)
9654 if (BlendMask[i] >= 0 && BlendMask[i] < SplitNumElements)
9655 BlendMask[i] = V1BlendMask[i] - (UseLoV1 ? 0 : SplitNumElements);
9657 if (UseLoV2 && UseHiV2) {
9659 DAG.getVectorShuffle(SplitVT, DL, LoV2, HiV2, V2BlendMask);
9661 // We only use half of V2 so map the usage down into the final blend mask.
9662 V2Blend = UseLoV2 ? LoV2 : HiV2;
9663 for (int i = 0; i < SplitNumElements; ++i)
9664 if (BlendMask[i] >= SplitNumElements)
9665 BlendMask[i] = V2BlendMask[i] + (UseLoV2 ? SplitNumElements : 0);
9667 return DAG.getVectorShuffle(SplitVT, DL, V1Blend, V2Blend, BlendMask);
9669 SDValue Lo = HalfBlend(LoMask);
9670 SDValue Hi = HalfBlend(HiMask);
9671 return DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, Lo, Hi);
9674 /// \brief Either split a vector in halves or decompose the shuffles and the
9677 /// This is provided as a good fallback for many lowerings of non-single-input
9678 /// shuffles with more than one 128-bit lane. In those cases, we want to select
9679 /// between splitting the shuffle into 128-bit components and stitching those
9680 /// back together vs. extracting the single-input shuffles and blending those
9682 static SDValue lowerVectorShuffleAsSplitOrBlend(SDLoc DL, MVT VT, SDValue V1,
9683 SDValue V2, ArrayRef<int> Mask,
9684 SelectionDAG &DAG) {
9685 assert(!isSingleInputShuffleMask(Mask) && "This routine must not be used to "
9686 "lower single-input shuffles as it "
9687 "could then recurse on itself.");
9688 int Size = Mask.size();
9690 // If this can be modeled as a broadcast of two elements followed by a blend,
9691 // prefer that lowering. This is especially important because broadcasts can
9692 // often fold with memory operands.
9693 auto DoBothBroadcast = [&] {
9694 int V1BroadcastIdx = -1, V2BroadcastIdx = -1;
9697 if (V2BroadcastIdx == -1)
9698 V2BroadcastIdx = M - Size;
9699 else if (M - Size != V2BroadcastIdx)
9701 } else if (M >= 0) {
9702 if (V1BroadcastIdx == -1)
9704 else if (M != V1BroadcastIdx)
9709 if (DoBothBroadcast())
9710 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, VT, V1, V2, Mask,
9713 // If the inputs all stem from a single 128-bit lane of each input, then we
9714 // split them rather than blending because the split will decompose to
9715 // unusually few instructions.
9716 int LaneCount = VT.getSizeInBits() / 128;
9717 int LaneSize = Size / LaneCount;
9718 SmallBitVector LaneInputs[2];
9719 LaneInputs[0].resize(LaneCount, false);
9720 LaneInputs[1].resize(LaneCount, false);
9721 for (int i = 0; i < Size; ++i)
9723 LaneInputs[Mask[i] / Size][(Mask[i] % Size) / LaneSize] = true;
9724 if (LaneInputs[0].count() <= 1 && LaneInputs[1].count() <= 1)
9725 return splitAndLowerVectorShuffle(DL, VT, V1, V2, Mask, DAG);
9727 // Otherwise, just fall back to decomposed shuffles and a blend. This requires
9728 // that the decomposed single-input shuffles don't end up here.
9729 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, VT, V1, V2, Mask, DAG);
9732 /// \brief Lower a vector shuffle crossing multiple 128-bit lanes as
9733 /// a permutation and blend of those lanes.
9735 /// This essentially blends the out-of-lane inputs to each lane into the lane
9736 /// from a permuted copy of the vector. This lowering strategy results in four
9737 /// instructions in the worst case for a single-input cross lane shuffle which
9738 /// is lower than any other fully general cross-lane shuffle strategy I'm aware
9739 /// of. Special cases for each particular shuffle pattern should be handled
9740 /// prior to trying this lowering.
9741 static SDValue lowerVectorShuffleAsLanePermuteAndBlend(SDLoc DL, MVT VT,
9742 SDValue V1, SDValue V2,
9744 SelectionDAG &DAG) {
9745 // FIXME: This should probably be generalized for 512-bit vectors as well.
9746 assert(VT.getSizeInBits() == 256 && "Only for 256-bit vector shuffles!");
9747 int LaneSize = Mask.size() / 2;
9749 // If there are only inputs from one 128-bit lane, splitting will in fact be
9750 // less expensive. The flags track whether the given lane contains an element
9751 // that crosses to another lane.
9752 bool LaneCrossing[2] = {false, false};
9753 for (int i = 0, Size = Mask.size(); i < Size; ++i)
9754 if (Mask[i] >= 0 && (Mask[i] % Size) / LaneSize != i / LaneSize)
9755 LaneCrossing[(Mask[i] % Size) / LaneSize] = true;
9756 if (!LaneCrossing[0] || !LaneCrossing[1])
9757 return splitAndLowerVectorShuffle(DL, VT, V1, V2, Mask, DAG);
9759 if (isSingleInputShuffleMask(Mask)) {
9760 SmallVector<int, 32> FlippedBlendMask;
9761 for (int i = 0, Size = Mask.size(); i < Size; ++i)
9762 FlippedBlendMask.push_back(
9763 Mask[i] < 0 ? -1 : (((Mask[i] % Size) / LaneSize == i / LaneSize)
9765 : Mask[i] % LaneSize +
9766 (i / LaneSize) * LaneSize + Size));
9768 // Flip the vector, and blend the results which should now be in-lane. The
9769 // VPERM2X128 mask uses the low 2 bits for the low source and bits 4 and
9770 // 5 for the high source. The value 3 selects the high half of source 2 and
9771 // the value 2 selects the low half of source 2. We only use source 2 to
9772 // allow folding it into a memory operand.
9773 unsigned PERMMask = 3 | 2 << 4;
9774 SDValue Flipped = DAG.getNode(X86ISD::VPERM2X128, DL, VT, DAG.getUNDEF(VT),
9775 V1, DAG.getConstant(PERMMask, DL, MVT::i8));
9776 return DAG.getVectorShuffle(VT, DL, V1, Flipped, FlippedBlendMask);
9779 // This now reduces to two single-input shuffles of V1 and V2 which at worst
9780 // will be handled by the above logic and a blend of the results, much like
9781 // other patterns in AVX.
9782 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, VT, V1, V2, Mask, DAG);
9785 /// \brief Handle lowering 2-lane 128-bit shuffles.
9786 static SDValue lowerV2X128VectorShuffle(SDLoc DL, MVT VT, SDValue V1,
9787 SDValue V2, ArrayRef<int> Mask,
9788 const X86Subtarget *Subtarget,
9789 SelectionDAG &DAG) {
9790 // TODO: If minimizing size and one of the inputs is a zero vector and the
9791 // the zero vector has only one use, we could use a VPERM2X128 to save the
9792 // instruction bytes needed to explicitly generate the zero vector.
9794 // Blends are faster and handle all the non-lane-crossing cases.
9795 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, VT, V1, V2, Mask,
9799 bool IsV1Zero = ISD::isBuildVectorAllZeros(V1.getNode());
9800 bool IsV2Zero = ISD::isBuildVectorAllZeros(V2.getNode());
9802 // If either input operand is a zero vector, use VPERM2X128 because its mask
9803 // allows us to replace the zero input with an implicit zero.
9804 if (!IsV1Zero && !IsV2Zero) {
9805 // Check for patterns which can be matched with a single insert of a 128-bit
9807 bool OnlyUsesV1 = isShuffleEquivalent(V1, V2, Mask, {0, 1, 0, 1});
9808 if (OnlyUsesV1 || isShuffleEquivalent(V1, V2, Mask, {0, 1, 4, 5})) {
9809 MVT SubVT = MVT::getVectorVT(VT.getVectorElementType(),
9810 VT.getVectorNumElements() / 2);
9811 SDValue LoV = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, SubVT, V1,
9812 DAG.getIntPtrConstant(0, DL));
9813 SDValue HiV = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, SubVT,
9814 OnlyUsesV1 ? V1 : V2,
9815 DAG.getIntPtrConstant(0, DL));
9816 return DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, LoV, HiV);
9820 // Otherwise form a 128-bit permutation. After accounting for undefs,
9821 // convert the 64-bit shuffle mask selection values into 128-bit
9822 // selection bits by dividing the indexes by 2 and shifting into positions
9823 // defined by a vperm2*128 instruction's immediate control byte.
9825 // The immediate permute control byte looks like this:
9826 // [1:0] - select 128 bits from sources for low half of destination
9828 // [3] - zero low half of destination
9829 // [5:4] - select 128 bits from sources for high half of destination
9831 // [7] - zero high half of destination
9833 int MaskLO = Mask[0];
9834 if (MaskLO == SM_SentinelUndef)
9835 MaskLO = Mask[1] == SM_SentinelUndef ? 0 : Mask[1];
9837 int MaskHI = Mask[2];
9838 if (MaskHI == SM_SentinelUndef)
9839 MaskHI = Mask[3] == SM_SentinelUndef ? 0 : Mask[3];
9841 unsigned PermMask = MaskLO / 2 | (MaskHI / 2) << 4;
9843 // If either input is a zero vector, replace it with an undef input.
9844 // Shuffle mask values < 4 are selecting elements of V1.
9845 // Shuffle mask values >= 4 are selecting elements of V2.
9846 // Adjust each half of the permute mask by clearing the half that was
9847 // selecting the zero vector and setting the zero mask bit.
9849 V1 = DAG.getUNDEF(VT);
9851 PermMask = (PermMask & 0xf0) | 0x08;
9853 PermMask = (PermMask & 0x0f) | 0x80;
9856 V2 = DAG.getUNDEF(VT);
9858 PermMask = (PermMask & 0xf0) | 0x08;
9860 PermMask = (PermMask & 0x0f) | 0x80;
9863 return DAG.getNode(X86ISD::VPERM2X128, DL, VT, V1, V2,
9864 DAG.getConstant(PermMask, DL, MVT::i8));
9867 /// \brief Lower a vector shuffle by first fixing the 128-bit lanes and then
9868 /// shuffling each lane.
9870 /// This will only succeed when the result of fixing the 128-bit lanes results
9871 /// in a single-input non-lane-crossing shuffle with a repeating shuffle mask in
9872 /// each 128-bit lanes. This handles many cases where we can quickly blend away
9873 /// the lane crosses early and then use simpler shuffles within each lane.
9875 /// FIXME: It might be worthwhile at some point to support this without
9876 /// requiring the 128-bit lane-relative shuffles to be repeating, but currently
9877 /// in x86 only floating point has interesting non-repeating shuffles, and even
9878 /// those are still *marginally* more expensive.
9879 static SDValue lowerVectorShuffleByMerging128BitLanes(
9880 SDLoc DL, MVT VT, SDValue V1, SDValue V2, ArrayRef<int> Mask,
9881 const X86Subtarget *Subtarget, SelectionDAG &DAG) {
9882 assert(!isSingleInputShuffleMask(Mask) &&
9883 "This is only useful with multiple inputs.");
9885 int Size = Mask.size();
9886 int LaneSize = 128 / VT.getScalarSizeInBits();
9887 int NumLanes = Size / LaneSize;
9888 assert(NumLanes > 1 && "Only handles 256-bit and wider shuffles.");
9890 // See if we can build a hypothetical 128-bit lane-fixing shuffle mask. Also
9891 // check whether the in-128-bit lane shuffles share a repeating pattern.
9892 SmallVector<int, 4> Lanes;
9893 Lanes.resize(NumLanes, -1);
9894 SmallVector<int, 4> InLaneMask;
9895 InLaneMask.resize(LaneSize, -1);
9896 for (int i = 0; i < Size; ++i) {
9900 int j = i / LaneSize;
9903 // First entry we've seen for this lane.
9904 Lanes[j] = Mask[i] / LaneSize;
9905 } else if (Lanes[j] != Mask[i] / LaneSize) {
9906 // This doesn't match the lane selected previously!
9910 // Check that within each lane we have a consistent shuffle mask.
9911 int k = i % LaneSize;
9912 if (InLaneMask[k] < 0) {
9913 InLaneMask[k] = Mask[i] % LaneSize;
9914 } else if (InLaneMask[k] != Mask[i] % LaneSize) {
9915 // This doesn't fit a repeating in-lane mask.
9920 // First shuffle the lanes into place.
9921 MVT LaneVT = MVT::getVectorVT(VT.isFloatingPoint() ? MVT::f64 : MVT::i64,
9922 VT.getSizeInBits() / 64);
9923 SmallVector<int, 8> LaneMask;
9924 LaneMask.resize(NumLanes * 2, -1);
9925 for (int i = 0; i < NumLanes; ++i)
9926 if (Lanes[i] >= 0) {
9927 LaneMask[2 * i + 0] = 2*Lanes[i] + 0;
9928 LaneMask[2 * i + 1] = 2*Lanes[i] + 1;
9931 V1 = DAG.getBitcast(LaneVT, V1);
9932 V2 = DAG.getBitcast(LaneVT, V2);
9933 SDValue LaneShuffle = DAG.getVectorShuffle(LaneVT, DL, V1, V2, LaneMask);
9935 // Cast it back to the type we actually want.
9936 LaneShuffle = DAG.getBitcast(VT, LaneShuffle);
9938 // Now do a simple shuffle that isn't lane crossing.
9939 SmallVector<int, 8> NewMask;
9940 NewMask.resize(Size, -1);
9941 for (int i = 0; i < Size; ++i)
9943 NewMask[i] = (i / LaneSize) * LaneSize + Mask[i] % LaneSize;
9944 assert(!is128BitLaneCrossingShuffleMask(VT, NewMask) &&
9945 "Must not introduce lane crosses at this point!");
9947 return DAG.getVectorShuffle(VT, DL, LaneShuffle, DAG.getUNDEF(VT), NewMask);
9950 /// \brief Test whether the specified input (0 or 1) is in-place blended by the
9953 /// This returns true if the elements from a particular input are already in the
9954 /// slot required by the given mask and require no permutation.
9955 static bool isShuffleMaskInputInPlace(int Input, ArrayRef<int> Mask) {
9956 assert((Input == 0 || Input == 1) && "Only two inputs to shuffles.");
9957 int Size = Mask.size();
9958 for (int i = 0; i < Size; ++i)
9959 if (Mask[i] >= 0 && Mask[i] / Size == Input && Mask[i] % Size != i)
9965 static SDValue lowerVectorShuffleWithSHUFPD(SDLoc DL, MVT VT,
9966 ArrayRef<int> Mask, SDValue V1,
9967 SDValue V2, SelectionDAG &DAG) {
9969 // Mask for V8F64: 0/1, 8/9, 2/3, 10/11, 4/5, ..
9970 // Mask for V4F64; 0/1, 4/5, 2/3, 6/7..
9971 assert(VT.getScalarSizeInBits() == 64 && "Unexpected data type for VSHUFPD");
9972 int NumElts = VT.getVectorNumElements();
9973 bool ShufpdMask = true;
9974 bool CommutableMask = true;
9975 unsigned Immediate = 0;
9976 for (int i = 0; i < NumElts; ++i) {
9979 int Val = (i & 6) + NumElts * (i & 1);
9980 int CommutVal = (i & 0xe) + NumElts * ((i & 1)^1);
9981 if (Mask[i] < Val || Mask[i] > Val + 1)
9983 if (Mask[i] < CommutVal || Mask[i] > CommutVal + 1)
9984 CommutableMask = false;
9985 Immediate |= (Mask[i] % 2) << i;
9988 return DAG.getNode(X86ISD::SHUFP, DL, VT, V1, V2,
9989 DAG.getConstant(Immediate, DL, MVT::i8));
9991 return DAG.getNode(X86ISD::SHUFP, DL, VT, V2, V1,
9992 DAG.getConstant(Immediate, DL, MVT::i8));
9996 /// \brief Handle lowering of 4-lane 64-bit floating point shuffles.
9998 /// Also ends up handling lowering of 4-lane 64-bit integer shuffles when AVX2
9999 /// isn't available.
10000 static SDValue lowerV4F64VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10001 const X86Subtarget *Subtarget,
10002 SelectionDAG &DAG) {
10004 assert(V1.getSimpleValueType() == MVT::v4f64 && "Bad operand type!");
10005 assert(V2.getSimpleValueType() == MVT::v4f64 && "Bad operand type!");
10006 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10007 ArrayRef<int> Mask = SVOp->getMask();
10008 assert(Mask.size() == 4 && "Unexpected mask size for v4 shuffle!");
10010 SmallVector<int, 4> WidenedMask;
10011 if (canWidenShuffleElements(Mask, WidenedMask))
10012 return lowerV2X128VectorShuffle(DL, MVT::v4f64, V1, V2, Mask, Subtarget,
10015 if (isSingleInputShuffleMask(Mask)) {
10016 // Check for being able to broadcast a single element.
10017 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v4f64, V1,
10018 Mask, Subtarget, DAG))
10021 // Use low duplicate instructions for masks that match their pattern.
10022 if (isShuffleEquivalent(V1, V2, Mask, {0, 0, 2, 2}))
10023 return DAG.getNode(X86ISD::MOVDDUP, DL, MVT::v4f64, V1);
10025 if (!is128BitLaneCrossingShuffleMask(MVT::v4f64, Mask)) {
10026 // Non-half-crossing single input shuffles can be lowerid with an
10027 // interleaved permutation.
10028 unsigned VPERMILPMask = (Mask[0] == 1) | ((Mask[1] == 1) << 1) |
10029 ((Mask[2] == 3) << 2) | ((Mask[3] == 3) << 3);
10030 return DAG.getNode(X86ISD::VPERMILPI, DL, MVT::v4f64, V1,
10031 DAG.getConstant(VPERMILPMask, DL, MVT::i8));
10034 // With AVX2 we have direct support for this permutation.
10035 if (Subtarget->hasAVX2())
10036 return DAG.getNode(X86ISD::VPERMI, DL, MVT::v4f64, V1,
10037 getV4X86ShuffleImm8ForMask(Mask, DL, DAG));
10039 // Otherwise, fall back.
10040 return lowerVectorShuffleAsLanePermuteAndBlend(DL, MVT::v4f64, V1, V2, Mask,
10044 // X86 has dedicated unpack instructions that can handle specific blend
10045 // operations: UNPCKH and UNPCKL.
10046 if (isShuffleEquivalent(V1, V2, Mask, {0, 4, 2, 6}))
10047 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v4f64, V1, V2);
10048 if (isShuffleEquivalent(V1, V2, Mask, {1, 5, 3, 7}))
10049 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v4f64, V1, V2);
10050 if (isShuffleEquivalent(V1, V2, Mask, {4, 0, 6, 2}))
10051 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v4f64, V2, V1);
10052 if (isShuffleEquivalent(V1, V2, Mask, {5, 1, 7, 3}))
10053 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v4f64, V2, V1);
10055 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v4f64, V1, V2, Mask,
10059 // Check if the blend happens to exactly fit that of SHUFPD.
10061 lowerVectorShuffleWithSHUFPD(DL, MVT::v4f64, Mask, V1, V2, DAG))
10064 // Try to simplify this by merging 128-bit lanes to enable a lane-based
10065 // shuffle. However, if we have AVX2 and either inputs are already in place,
10066 // we will be able to shuffle even across lanes the other input in a single
10067 // instruction so skip this pattern.
10068 if (!(Subtarget->hasAVX2() && (isShuffleMaskInputInPlace(0, Mask) ||
10069 isShuffleMaskInputInPlace(1, Mask))))
10070 if (SDValue Result = lowerVectorShuffleByMerging128BitLanes(
10071 DL, MVT::v4f64, V1, V2, Mask, Subtarget, DAG))
10074 // If we have AVX2 then we always want to lower with a blend because an v4 we
10075 // can fully permute the elements.
10076 if (Subtarget->hasAVX2())
10077 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v4f64, V1, V2,
10080 // Otherwise fall back on generic lowering.
10081 return lowerVectorShuffleAsSplitOrBlend(DL, MVT::v4f64, V1, V2, Mask, DAG);
10084 /// \brief Handle lowering of 4-lane 64-bit integer shuffles.
10086 /// This routine is only called when we have AVX2 and thus a reasonable
10087 /// instruction set for v4i64 shuffling..
10088 static SDValue lowerV4I64VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10089 const X86Subtarget *Subtarget,
10090 SelectionDAG &DAG) {
10092 assert(V1.getSimpleValueType() == MVT::v4i64 && "Bad operand type!");
10093 assert(V2.getSimpleValueType() == MVT::v4i64 && "Bad operand type!");
10094 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10095 ArrayRef<int> Mask = SVOp->getMask();
10096 assert(Mask.size() == 4 && "Unexpected mask size for v4 shuffle!");
10097 assert(Subtarget->hasAVX2() && "We can only lower v4i64 with AVX2!");
10099 SmallVector<int, 4> WidenedMask;
10100 if (canWidenShuffleElements(Mask, WidenedMask))
10101 return lowerV2X128VectorShuffle(DL, MVT::v4i64, V1, V2, Mask, Subtarget,
10104 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v4i64, V1, V2, Mask,
10108 // Check for being able to broadcast a single element.
10109 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v4i64, V1,
10110 Mask, Subtarget, DAG))
10113 // When the shuffle is mirrored between the 128-bit lanes of the unit, we can
10114 // use lower latency instructions that will operate on both 128-bit lanes.
10115 SmallVector<int, 2> RepeatedMask;
10116 if (is128BitLaneRepeatedShuffleMask(MVT::v4i64, Mask, RepeatedMask)) {
10117 if (isSingleInputShuffleMask(Mask)) {
10118 int PSHUFDMask[] = {-1, -1, -1, -1};
10119 for (int i = 0; i < 2; ++i)
10120 if (RepeatedMask[i] >= 0) {
10121 PSHUFDMask[2 * i] = 2 * RepeatedMask[i];
10122 PSHUFDMask[2 * i + 1] = 2 * RepeatedMask[i] + 1;
10124 return DAG.getBitcast(
10126 DAG.getNode(X86ISD::PSHUFD, DL, MVT::v8i32,
10127 DAG.getBitcast(MVT::v8i32, V1),
10128 getV4X86ShuffleImm8ForMask(PSHUFDMask, DL, DAG)));
10132 // AVX2 provides a direct instruction for permuting a single input across
10134 if (isSingleInputShuffleMask(Mask))
10135 return DAG.getNode(X86ISD::VPERMI, DL, MVT::v4i64, V1,
10136 getV4X86ShuffleImm8ForMask(Mask, DL, DAG));
10138 // Try to use shift instructions.
10139 if (SDValue Shift =
10140 lowerVectorShuffleAsShift(DL, MVT::v4i64, V1, V2, Mask, DAG))
10143 // Use dedicated unpack instructions for masks that match their pattern.
10144 if (isShuffleEquivalent(V1, V2, Mask, {0, 4, 2, 6}))
10145 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v4i64, V1, V2);
10146 if (isShuffleEquivalent(V1, V2, Mask, {1, 5, 3, 7}))
10147 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v4i64, V1, V2);
10148 if (isShuffleEquivalent(V1, V2, Mask, {4, 0, 6, 2}))
10149 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v4i64, V2, V1);
10150 if (isShuffleEquivalent(V1, V2, Mask, {5, 1, 7, 3}))
10151 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v4i64, V2, V1);
10153 // Try to simplify this by merging 128-bit lanes to enable a lane-based
10154 // shuffle. However, if we have AVX2 and either inputs are already in place,
10155 // we will be able to shuffle even across lanes the other input in a single
10156 // instruction so skip this pattern.
10157 if (!(Subtarget->hasAVX2() && (isShuffleMaskInputInPlace(0, Mask) ||
10158 isShuffleMaskInputInPlace(1, Mask))))
10159 if (SDValue Result = lowerVectorShuffleByMerging128BitLanes(
10160 DL, MVT::v4i64, V1, V2, Mask, Subtarget, DAG))
10163 // Otherwise fall back on generic blend lowering.
10164 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v4i64, V1, V2,
10168 /// \brief Handle lowering of 8-lane 32-bit floating point shuffles.
10170 /// Also ends up handling lowering of 8-lane 32-bit integer shuffles when AVX2
10171 /// isn't available.
10172 static SDValue lowerV8F32VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10173 const X86Subtarget *Subtarget,
10174 SelectionDAG &DAG) {
10176 assert(V1.getSimpleValueType() == MVT::v8f32 && "Bad operand type!");
10177 assert(V2.getSimpleValueType() == MVT::v8f32 && "Bad operand type!");
10178 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10179 ArrayRef<int> Mask = SVOp->getMask();
10180 assert(Mask.size() == 8 && "Unexpected mask size for v8 shuffle!");
10182 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v8f32, V1, V2, Mask,
10186 // Check for being able to broadcast a single element.
10187 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v8f32, V1,
10188 Mask, Subtarget, DAG))
10191 // If the shuffle mask is repeated in each 128-bit lane, we have many more
10192 // options to efficiently lower the shuffle.
10193 SmallVector<int, 4> RepeatedMask;
10194 if (is128BitLaneRepeatedShuffleMask(MVT::v8f32, Mask, RepeatedMask)) {
10195 assert(RepeatedMask.size() == 4 &&
10196 "Repeated masks must be half the mask width!");
10198 // Use even/odd duplicate instructions for masks that match their pattern.
10199 if (isShuffleEquivalent(V1, V2, Mask, {0, 0, 2, 2, 4, 4, 6, 6}))
10200 return DAG.getNode(X86ISD::MOVSLDUP, DL, MVT::v8f32, V1);
10201 if (isShuffleEquivalent(V1, V2, Mask, {1, 1, 3, 3, 5, 5, 7, 7}))
10202 return DAG.getNode(X86ISD::MOVSHDUP, DL, MVT::v8f32, V1);
10204 if (isSingleInputShuffleMask(Mask))
10205 return DAG.getNode(X86ISD::VPERMILPI, DL, MVT::v8f32, V1,
10206 getV4X86ShuffleImm8ForMask(RepeatedMask, DL, DAG));
10208 // Use dedicated unpack instructions for masks that match their pattern.
10209 if (isShuffleEquivalent(V1, V2, Mask, {0, 8, 1, 9, 4, 12, 5, 13}))
10210 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v8f32, V1, V2);
10211 if (isShuffleEquivalent(V1, V2, Mask, {2, 10, 3, 11, 6, 14, 7, 15}))
10212 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v8f32, V1, V2);
10213 if (isShuffleEquivalent(V1, V2, Mask, {8, 0, 9, 1, 12, 4, 13, 5}))
10214 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v8f32, V2, V1);
10215 if (isShuffleEquivalent(V1, V2, Mask, {10, 2, 11, 3, 14, 6, 15, 7}))
10216 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v8f32, V2, V1);
10218 // Otherwise, fall back to a SHUFPS sequence. Here it is important that we
10219 // have already handled any direct blends. We also need to squash the
10220 // repeated mask into a simulated v4f32 mask.
10221 for (int i = 0; i < 4; ++i)
10222 if (RepeatedMask[i] >= 8)
10223 RepeatedMask[i] -= 4;
10224 return lowerVectorShuffleWithSHUFPS(DL, MVT::v8f32, RepeatedMask, V1, V2, DAG);
10227 // If we have a single input shuffle with different shuffle patterns in the
10228 // two 128-bit lanes use the variable mask to VPERMILPS.
10229 if (isSingleInputShuffleMask(Mask)) {
10230 SDValue VPermMask[8];
10231 for (int i = 0; i < 8; ++i)
10232 VPermMask[i] = Mask[i] < 0 ? DAG.getUNDEF(MVT::i32)
10233 : DAG.getConstant(Mask[i], DL, MVT::i32);
10234 if (!is128BitLaneCrossingShuffleMask(MVT::v8f32, Mask))
10235 return DAG.getNode(
10236 X86ISD::VPERMILPV, DL, MVT::v8f32, V1,
10237 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v8i32, VPermMask));
10239 if (Subtarget->hasAVX2())
10240 return DAG.getNode(
10241 X86ISD::VPERMV, DL, MVT::v8f32,
10242 DAG.getBitcast(MVT::v8f32, DAG.getNode(ISD::BUILD_VECTOR, DL,
10243 MVT::v8i32, VPermMask)),
10246 // Otherwise, fall back.
10247 return lowerVectorShuffleAsLanePermuteAndBlend(DL, MVT::v8f32, V1, V2, Mask,
10251 // Try to simplify this by merging 128-bit lanes to enable a lane-based
10253 if (SDValue Result = lowerVectorShuffleByMerging128BitLanes(
10254 DL, MVT::v8f32, V1, V2, Mask, Subtarget, DAG))
10257 // If we have AVX2 then we always want to lower with a blend because at v8 we
10258 // can fully permute the elements.
10259 if (Subtarget->hasAVX2())
10260 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v8f32, V1, V2,
10263 // Otherwise fall back on generic lowering.
10264 return lowerVectorShuffleAsSplitOrBlend(DL, MVT::v8f32, V1, V2, Mask, DAG);
10267 /// \brief Handle lowering of 8-lane 32-bit integer shuffles.
10269 /// This routine is only called when we have AVX2 and thus a reasonable
10270 /// instruction set for v8i32 shuffling..
10271 static SDValue lowerV8I32VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10272 const X86Subtarget *Subtarget,
10273 SelectionDAG &DAG) {
10275 assert(V1.getSimpleValueType() == MVT::v8i32 && "Bad operand type!");
10276 assert(V2.getSimpleValueType() == MVT::v8i32 && "Bad operand type!");
10277 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10278 ArrayRef<int> Mask = SVOp->getMask();
10279 assert(Mask.size() == 8 && "Unexpected mask size for v8 shuffle!");
10280 assert(Subtarget->hasAVX2() && "We can only lower v8i32 with AVX2!");
10282 // Whenever we can lower this as a zext, that instruction is strictly faster
10283 // than any alternative. It also allows us to fold memory operands into the
10284 // shuffle in many cases.
10285 if (SDValue ZExt = lowerVectorShuffleAsZeroOrAnyExtend(DL, MVT::v8i32, V1, V2,
10286 Mask, Subtarget, DAG))
10289 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v8i32, V1, V2, Mask,
10293 // Check for being able to broadcast a single element.
10294 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v8i32, V1,
10295 Mask, Subtarget, DAG))
10298 // If the shuffle mask is repeated in each 128-bit lane we can use more
10299 // efficient instructions that mirror the shuffles across the two 128-bit
10301 SmallVector<int, 4> RepeatedMask;
10302 if (is128BitLaneRepeatedShuffleMask(MVT::v8i32, Mask, RepeatedMask)) {
10303 assert(RepeatedMask.size() == 4 && "Unexpected repeated mask size!");
10304 if (isSingleInputShuffleMask(Mask))
10305 return DAG.getNode(X86ISD::PSHUFD, DL, MVT::v8i32, V1,
10306 getV4X86ShuffleImm8ForMask(RepeatedMask, DL, DAG));
10308 // Use dedicated unpack instructions for masks that match their pattern.
10309 if (isShuffleEquivalent(V1, V2, Mask, {0, 8, 1, 9, 4, 12, 5, 13}))
10310 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v8i32, V1, V2);
10311 if (isShuffleEquivalent(V1, V2, Mask, {2, 10, 3, 11, 6, 14, 7, 15}))
10312 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v8i32, V1, V2);
10313 if (isShuffleEquivalent(V1, V2, Mask, {8, 0, 9, 1, 12, 4, 13, 5}))
10314 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v8i32, V2, V1);
10315 if (isShuffleEquivalent(V1, V2, Mask, {10, 2, 11, 3, 14, 6, 15, 7}))
10316 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v8i32, V2, V1);
10319 // Try to use shift instructions.
10320 if (SDValue Shift =
10321 lowerVectorShuffleAsShift(DL, MVT::v8i32, V1, V2, Mask, DAG))
10324 if (SDValue Rotate = lowerVectorShuffleAsByteRotate(
10325 DL, MVT::v8i32, V1, V2, Mask, Subtarget, DAG))
10328 // If the shuffle patterns aren't repeated but it is a single input, directly
10329 // generate a cross-lane VPERMD instruction.
10330 if (isSingleInputShuffleMask(Mask)) {
10331 SDValue VPermMask[8];
10332 for (int i = 0; i < 8; ++i)
10333 VPermMask[i] = Mask[i] < 0 ? DAG.getUNDEF(MVT::i32)
10334 : DAG.getConstant(Mask[i], DL, MVT::i32);
10335 return DAG.getNode(
10336 X86ISD::VPERMV, DL, MVT::v8i32,
10337 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v8i32, VPermMask), V1);
10340 // Try to simplify this by merging 128-bit lanes to enable a lane-based
10342 if (SDValue Result = lowerVectorShuffleByMerging128BitLanes(
10343 DL, MVT::v8i32, V1, V2, Mask, Subtarget, DAG))
10346 // Otherwise fall back on generic blend lowering.
10347 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v8i32, V1, V2,
10351 /// \brief Handle lowering of 16-lane 16-bit integer shuffles.
10353 /// This routine is only called when we have AVX2 and thus a reasonable
10354 /// instruction set for v16i16 shuffling..
10355 static SDValue lowerV16I16VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10356 const X86Subtarget *Subtarget,
10357 SelectionDAG &DAG) {
10359 assert(V1.getSimpleValueType() == MVT::v16i16 && "Bad operand type!");
10360 assert(V2.getSimpleValueType() == MVT::v16i16 && "Bad operand type!");
10361 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10362 ArrayRef<int> Mask = SVOp->getMask();
10363 assert(Mask.size() == 16 && "Unexpected mask size for v16 shuffle!");
10364 assert(Subtarget->hasAVX2() && "We can only lower v16i16 with AVX2!");
10366 // Whenever we can lower this as a zext, that instruction is strictly faster
10367 // than any alternative. It also allows us to fold memory operands into the
10368 // shuffle in many cases.
10369 if (SDValue ZExt = lowerVectorShuffleAsZeroOrAnyExtend(DL, MVT::v16i16, V1, V2,
10370 Mask, Subtarget, DAG))
10373 // Check for being able to broadcast a single element.
10374 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v16i16, V1,
10375 Mask, Subtarget, DAG))
10378 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v16i16, V1, V2, Mask,
10382 // Use dedicated unpack instructions for masks that match their pattern.
10383 if (isShuffleEquivalent(V1, V2, Mask,
10384 {// First 128-bit lane:
10385 0, 16, 1, 17, 2, 18, 3, 19,
10386 // Second 128-bit lane:
10387 8, 24, 9, 25, 10, 26, 11, 27}))
10388 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v16i16, V1, V2);
10389 if (isShuffleEquivalent(V1, V2, Mask,
10390 {// First 128-bit lane:
10391 4, 20, 5, 21, 6, 22, 7, 23,
10392 // Second 128-bit lane:
10393 12, 28, 13, 29, 14, 30, 15, 31}))
10394 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v16i16, V1, V2);
10396 // Try to use shift instructions.
10397 if (SDValue Shift =
10398 lowerVectorShuffleAsShift(DL, MVT::v16i16, V1, V2, Mask, DAG))
10401 // Try to use byte rotation instructions.
10402 if (SDValue Rotate = lowerVectorShuffleAsByteRotate(
10403 DL, MVT::v16i16, V1, V2, Mask, Subtarget, DAG))
10406 if (isSingleInputShuffleMask(Mask)) {
10407 // There are no generalized cross-lane shuffle operations available on i16
10409 if (is128BitLaneCrossingShuffleMask(MVT::v16i16, Mask))
10410 return lowerVectorShuffleAsLanePermuteAndBlend(DL, MVT::v16i16, V1, V2,
10413 SmallVector<int, 8> RepeatedMask;
10414 if (is128BitLaneRepeatedShuffleMask(MVT::v16i16, Mask, RepeatedMask)) {
10415 // As this is a single-input shuffle, the repeated mask should be
10416 // a strictly valid v8i16 mask that we can pass through to the v8i16
10417 // lowering to handle even the v16 case.
10418 return lowerV8I16GeneralSingleInputVectorShuffle(
10419 DL, MVT::v16i16, V1, RepeatedMask, Subtarget, DAG);
10422 SDValue PSHUFBMask[32];
10423 for (int i = 0; i < 16; ++i) {
10424 if (Mask[i] == -1) {
10425 PSHUFBMask[2 * i] = PSHUFBMask[2 * i + 1] = DAG.getUNDEF(MVT::i8);
10429 int M = i < 8 ? Mask[i] : Mask[i] - 8;
10430 assert(M >= 0 && M < 8 && "Invalid single-input mask!");
10431 PSHUFBMask[2 * i] = DAG.getConstant(2 * M, DL, MVT::i8);
10432 PSHUFBMask[2 * i + 1] = DAG.getConstant(2 * M + 1, DL, MVT::i8);
10434 return DAG.getBitcast(MVT::v16i16,
10435 DAG.getNode(X86ISD::PSHUFB, DL, MVT::v32i8,
10436 DAG.getBitcast(MVT::v32i8, V1),
10437 DAG.getNode(ISD::BUILD_VECTOR, DL,
10438 MVT::v32i8, PSHUFBMask)));
10441 // Try to simplify this by merging 128-bit lanes to enable a lane-based
10443 if (SDValue Result = lowerVectorShuffleByMerging128BitLanes(
10444 DL, MVT::v16i16, V1, V2, Mask, Subtarget, DAG))
10447 // Otherwise fall back on generic lowering.
10448 return lowerVectorShuffleAsSplitOrBlend(DL, MVT::v16i16, V1, V2, Mask, DAG);
10451 /// \brief Handle lowering of 32-lane 8-bit integer shuffles.
10453 /// This routine is only called when we have AVX2 and thus a reasonable
10454 /// instruction set for v32i8 shuffling..
10455 static SDValue lowerV32I8VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10456 const X86Subtarget *Subtarget,
10457 SelectionDAG &DAG) {
10459 assert(V1.getSimpleValueType() == MVT::v32i8 && "Bad operand type!");
10460 assert(V2.getSimpleValueType() == MVT::v32i8 && "Bad operand type!");
10461 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10462 ArrayRef<int> Mask = SVOp->getMask();
10463 assert(Mask.size() == 32 && "Unexpected mask size for v32 shuffle!");
10464 assert(Subtarget->hasAVX2() && "We can only lower v32i8 with AVX2!");
10466 // Whenever we can lower this as a zext, that instruction is strictly faster
10467 // than any alternative. It also allows us to fold memory operands into the
10468 // shuffle in many cases.
10469 if (SDValue ZExt = lowerVectorShuffleAsZeroOrAnyExtend(DL, MVT::v32i8, V1, V2,
10470 Mask, Subtarget, DAG))
10473 // Check for being able to broadcast a single element.
10474 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v32i8, V1,
10475 Mask, Subtarget, DAG))
10478 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v32i8, V1, V2, Mask,
10482 // Use dedicated unpack instructions for masks that match their pattern.
10483 // Note that these are repeated 128-bit lane unpacks, not unpacks across all
10485 if (isShuffleEquivalent(
10487 {// First 128-bit lane:
10488 0, 32, 1, 33, 2, 34, 3, 35, 4, 36, 5, 37, 6, 38, 7, 39,
10489 // Second 128-bit lane:
10490 16, 48, 17, 49, 18, 50, 19, 51, 20, 52, 21, 53, 22, 54, 23, 55}))
10491 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v32i8, V1, V2);
10492 if (isShuffleEquivalent(
10494 {// First 128-bit lane:
10495 8, 40, 9, 41, 10, 42, 11, 43, 12, 44, 13, 45, 14, 46, 15, 47,
10496 // Second 128-bit lane:
10497 24, 56, 25, 57, 26, 58, 27, 59, 28, 60, 29, 61, 30, 62, 31, 63}))
10498 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v32i8, V1, V2);
10500 // Try to use shift instructions.
10501 if (SDValue Shift =
10502 lowerVectorShuffleAsShift(DL, MVT::v32i8, V1, V2, Mask, DAG))
10505 // Try to use byte rotation instructions.
10506 if (SDValue Rotate = lowerVectorShuffleAsByteRotate(
10507 DL, MVT::v32i8, V1, V2, Mask, Subtarget, DAG))
10510 if (isSingleInputShuffleMask(Mask)) {
10511 // There are no generalized cross-lane shuffle operations available on i8
10513 if (is128BitLaneCrossingShuffleMask(MVT::v32i8, Mask))
10514 return lowerVectorShuffleAsLanePermuteAndBlend(DL, MVT::v32i8, V1, V2,
10517 SDValue PSHUFBMask[32];
10518 for (int i = 0; i < 32; ++i)
10521 ? DAG.getUNDEF(MVT::i8)
10522 : DAG.getConstant(Mask[i] < 16 ? Mask[i] : Mask[i] - 16, DL,
10525 return DAG.getNode(
10526 X86ISD::PSHUFB, DL, MVT::v32i8, V1,
10527 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v32i8, PSHUFBMask));
10530 // Try to simplify this by merging 128-bit lanes to enable a lane-based
10532 if (SDValue Result = lowerVectorShuffleByMerging128BitLanes(
10533 DL, MVT::v32i8, V1, V2, Mask, Subtarget, DAG))
10536 // Otherwise fall back on generic lowering.
10537 return lowerVectorShuffleAsSplitOrBlend(DL, MVT::v32i8, V1, V2, Mask, DAG);
10540 /// \brief High-level routine to lower various 256-bit x86 vector shuffles.
10542 /// This routine either breaks down the specific type of a 256-bit x86 vector
10543 /// shuffle or splits it into two 128-bit shuffles and fuses the results back
10544 /// together based on the available instructions.
10545 static SDValue lower256BitVectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10546 MVT VT, const X86Subtarget *Subtarget,
10547 SelectionDAG &DAG) {
10549 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10550 ArrayRef<int> Mask = SVOp->getMask();
10552 // If we have a single input to the zero element, insert that into V1 if we
10553 // can do so cheaply.
10554 int NumElts = VT.getVectorNumElements();
10555 int NumV2Elements = std::count_if(Mask.begin(), Mask.end(), [NumElts](int M) {
10556 return M >= NumElts;
10559 if (NumV2Elements == 1 && Mask[0] >= NumElts)
10560 if (SDValue Insertion = lowerVectorShuffleAsElementInsertion(
10561 DL, VT, V1, V2, Mask, Subtarget, DAG))
10564 // There is a really nice hard cut-over between AVX1 and AVX2 that means we can
10565 // check for those subtargets here and avoid much of the subtarget querying in
10566 // the per-vector-type lowering routines. With AVX1 we have essentially *zero*
10567 // ability to manipulate a 256-bit vector with integer types. Since we'll use
10568 // floating point types there eventually, just immediately cast everything to
10569 // a float and operate entirely in that domain.
10570 if (VT.isInteger() && !Subtarget->hasAVX2()) {
10571 int ElementBits = VT.getScalarSizeInBits();
10572 if (ElementBits < 32)
10573 // No floating point type available, decompose into 128-bit vectors.
10574 return splitAndLowerVectorShuffle(DL, VT, V1, V2, Mask, DAG);
10576 MVT FpVT = MVT::getVectorVT(MVT::getFloatingPointVT(ElementBits),
10577 VT.getVectorNumElements());
10578 V1 = DAG.getBitcast(FpVT, V1);
10579 V2 = DAG.getBitcast(FpVT, V2);
10580 return DAG.getBitcast(VT, DAG.getVectorShuffle(FpVT, DL, V1, V2, Mask));
10583 switch (VT.SimpleTy) {
10585 return lowerV4F64VectorShuffle(Op, V1, V2, Subtarget, DAG);
10587 return lowerV4I64VectorShuffle(Op, V1, V2, Subtarget, DAG);
10589 return lowerV8F32VectorShuffle(Op, V1, V2, Subtarget, DAG);
10591 return lowerV8I32VectorShuffle(Op, V1, V2, Subtarget, DAG);
10593 return lowerV16I16VectorShuffle(Op, V1, V2, Subtarget, DAG);
10595 return lowerV32I8VectorShuffle(Op, V1, V2, Subtarget, DAG);
10598 llvm_unreachable("Not a valid 256-bit x86 vector type!");
10602 static SDValue lowerVectorShuffleWithPERMV(SDLoc DL, MVT VT,
10603 ArrayRef<int> Mask, SDValue V1,
10604 SDValue V2, SelectionDAG &DAG) {
10606 assert(VT.getScalarSizeInBits() >= 16 && "Unexpected data type for PERMV");
10608 MVT MaskEltVT = MVT::getIntegerVT(VT.getScalarSizeInBits());
10609 MVT MaskVecVT = MVT::getVectorVT(MaskEltVT, VT.getVectorNumElements());
10611 SmallVector<SDValue, 32> VPermMask;
10612 for (unsigned i = 0; i < VT.getVectorNumElements(); ++i)
10613 VPermMask.push_back(Mask[i] < 0 ? DAG.getUNDEF(MaskEltVT) :
10614 DAG.getConstant(Mask[i], DL, MaskEltVT));
10615 SDValue MaskNode = DAG.getNode(ISD::BUILD_VECTOR, DL, MaskVecVT,
10617 if (isSingleInputShuffleMask(Mask))
10618 return DAG.getNode(X86ISD::VPERMV, DL, VT, MaskNode, V1);
10620 return DAG.getNode(X86ISD::VPERMV3, DL, VT, V1, MaskNode, V2);
10623 /// \brief Handle lowering of 8-lane 64-bit floating point shuffles.
10624 static SDValue lowerV8F64VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10625 const X86Subtarget *Subtarget,
10626 SelectionDAG &DAG) {
10628 assert(V1.getSimpleValueType() == MVT::v8f64 && "Bad operand type!");
10629 assert(V2.getSimpleValueType() == MVT::v8f64 && "Bad operand type!");
10630 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10631 ArrayRef<int> Mask = SVOp->getMask();
10632 assert(Mask.size() == 8 && "Unexpected mask size for v8 shuffle!");
10634 if (SDValue Unpck =
10635 lowerVectorShuffleWithUNPCK(DL, MVT::v8f64, Mask, V1, V2, DAG))
10638 return lowerVectorShuffleWithPERMV(DL, MVT::v8f64, Mask, V1, V2, DAG);
10641 /// \brief Handle lowering of 16-lane 32-bit floating point shuffles.
10642 static SDValue lowerV16F32VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10643 const X86Subtarget *Subtarget,
10644 SelectionDAG &DAG) {
10646 assert(V1.getSimpleValueType() == MVT::v16f32 && "Bad operand type!");
10647 assert(V2.getSimpleValueType() == MVT::v16f32 && "Bad operand type!");
10648 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10649 ArrayRef<int> Mask = SVOp->getMask();
10650 assert(Mask.size() == 16 && "Unexpected mask size for v16 shuffle!");
10652 if (SDValue Unpck =
10653 lowerVectorShuffleWithUNPCK(DL, MVT::v16f32, Mask, V1, V2, DAG))
10656 return lowerVectorShuffleWithPERMV(DL, MVT::v16f32, Mask, V1, V2, DAG);
10659 /// \brief Handle lowering of 8-lane 64-bit integer shuffles.
10660 static SDValue lowerV8I64VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10661 const X86Subtarget *Subtarget,
10662 SelectionDAG &DAG) {
10664 assert(V1.getSimpleValueType() == MVT::v8i64 && "Bad operand type!");
10665 assert(V2.getSimpleValueType() == MVT::v8i64 && "Bad operand type!");
10666 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10667 ArrayRef<int> Mask = SVOp->getMask();
10668 assert(Mask.size() == 8 && "Unexpected mask size for v8 shuffle!");
10670 if (SDValue Unpck =
10671 lowerVectorShuffleWithUNPCK(DL, MVT::v8i64, Mask, V1, V2, DAG))
10674 return lowerVectorShuffleWithPERMV(DL, MVT::v8i64, Mask, V1, V2, DAG);
10677 /// \brief Handle lowering of 16-lane 32-bit integer shuffles.
10678 static SDValue lowerV16I32VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10679 const X86Subtarget *Subtarget,
10680 SelectionDAG &DAG) {
10682 assert(V1.getSimpleValueType() == MVT::v16i32 && "Bad operand type!");
10683 assert(V2.getSimpleValueType() == MVT::v16i32 && "Bad operand type!");
10684 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10685 ArrayRef<int> Mask = SVOp->getMask();
10686 assert(Mask.size() == 16 && "Unexpected mask size for v16 shuffle!");
10688 if (SDValue Unpck =
10689 lowerVectorShuffleWithUNPCK(DL, MVT::v16i32, Mask, V1, V2, DAG))
10692 return lowerVectorShuffleWithPERMV(DL, MVT::v16i32, Mask, V1, V2, DAG);
10695 /// \brief Handle lowering of 32-lane 16-bit integer shuffles.
10696 static SDValue lowerV32I16VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10697 const X86Subtarget *Subtarget,
10698 SelectionDAG &DAG) {
10700 assert(V1.getSimpleValueType() == MVT::v32i16 && "Bad operand type!");
10701 assert(V2.getSimpleValueType() == MVT::v32i16 && "Bad operand type!");
10702 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10703 ArrayRef<int> Mask = SVOp->getMask();
10704 assert(Mask.size() == 32 && "Unexpected mask size for v32 shuffle!");
10705 assert(Subtarget->hasBWI() && "We can only lower v32i16 with AVX-512-BWI!");
10707 return lowerVectorShuffleWithPERMV(DL, MVT::v32i16, Mask, V1, V2, DAG);
10710 /// \brief Handle lowering of 64-lane 8-bit integer shuffles.
10711 static SDValue lowerV64I8VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10712 const X86Subtarget *Subtarget,
10713 SelectionDAG &DAG) {
10715 assert(V1.getSimpleValueType() == MVT::v64i8 && "Bad operand type!");
10716 assert(V2.getSimpleValueType() == MVT::v64i8 && "Bad operand type!");
10717 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10718 ArrayRef<int> Mask = SVOp->getMask();
10719 assert(Mask.size() == 64 && "Unexpected mask size for v64 shuffle!");
10720 assert(Subtarget->hasBWI() && "We can only lower v64i8 with AVX-512-BWI!");
10722 // FIXME: Implement direct support for this type!
10723 return splitAndLowerVectorShuffle(DL, MVT::v64i8, V1, V2, Mask, DAG);
10726 /// \brief High-level routine to lower various 512-bit x86 vector shuffles.
10728 /// This routine either breaks down the specific type of a 512-bit x86 vector
10729 /// shuffle or splits it into two 256-bit shuffles and fuses the results back
10730 /// together based on the available instructions.
10731 static SDValue lower512BitVectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10732 MVT VT, const X86Subtarget *Subtarget,
10733 SelectionDAG &DAG) {
10735 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10736 ArrayRef<int> Mask = SVOp->getMask();
10737 assert(Subtarget->hasAVX512() &&
10738 "Cannot lower 512-bit vectors w/ basic ISA!");
10740 // Check for being able to broadcast a single element.
10741 if (SDValue Broadcast =
10742 lowerVectorShuffleAsBroadcast(DL, VT, V1, Mask, Subtarget, DAG))
10745 // Dispatch to each element type for lowering. If we don't have supprot for
10746 // specific element type shuffles at 512 bits, immediately split them and
10747 // lower them. Each lowering routine of a given type is allowed to assume that
10748 // the requisite ISA extensions for that element type are available.
10749 switch (VT.SimpleTy) {
10751 return lowerV8F64VectorShuffle(Op, V1, V2, Subtarget, DAG);
10753 return lowerV16F32VectorShuffle(Op, V1, V2, Subtarget, DAG);
10755 return lowerV8I64VectorShuffle(Op, V1, V2, Subtarget, DAG);
10757 return lowerV16I32VectorShuffle(Op, V1, V2, Subtarget, DAG);
10759 if (Subtarget->hasBWI())
10760 return lowerV32I16VectorShuffle(Op, V1, V2, Subtarget, DAG);
10763 if (Subtarget->hasBWI())
10764 return lowerV64I8VectorShuffle(Op, V1, V2, Subtarget, DAG);
10768 llvm_unreachable("Not a valid 512-bit x86 vector type!");
10771 // Otherwise fall back on splitting.
10772 return splitAndLowerVectorShuffle(DL, VT, V1, V2, Mask, DAG);
10775 // Lower vXi1 vector shuffles.
10776 // There is no a dedicated instruction on AVX-512 that shuffles the masks.
10777 // The only way to shuffle bits is to sign-extend the mask vector to SIMD
10778 // vector, shuffle and then truncate it back.
10779 static SDValue lower1BitVectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10780 MVT VT, const X86Subtarget *Subtarget,
10781 SelectionDAG &DAG) {
10783 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10784 ArrayRef<int> Mask = SVOp->getMask();
10785 assert(Subtarget->hasAVX512() &&
10786 "Cannot lower 512-bit vectors w/o basic ISA!");
10788 switch (VT.SimpleTy) {
10790 assert(false && "Expected a vector of i1 elements");
10793 ExtVT = MVT::v2i64;
10796 ExtVT = MVT::v4i32;
10799 ExtVT = MVT::v8i64; // Take 512-bit type, more shuffles on KNL
10802 ExtVT = MVT::v16i32;
10805 ExtVT = MVT::v32i16;
10808 ExtVT = MVT::v64i8;
10812 if (ISD::isBuildVectorAllZeros(V1.getNode()))
10813 V1 = getZeroVector(ExtVT, Subtarget, DAG, DL);
10814 else if (ISD::isBuildVectorAllOnes(V1.getNode()))
10815 V1 = getOnesVector(ExtVT, Subtarget, DAG, DL);
10817 V1 = DAG.getNode(ISD::SIGN_EXTEND, DL, ExtVT, V1);
10820 V2 = DAG.getUNDEF(ExtVT);
10821 else if (ISD::isBuildVectorAllZeros(V2.getNode()))
10822 V2 = getZeroVector(ExtVT, Subtarget, DAG, DL);
10823 else if (ISD::isBuildVectorAllOnes(V2.getNode()))
10824 V2 = getOnesVector(ExtVT, Subtarget, DAG, DL);
10826 V2 = DAG.getNode(ISD::SIGN_EXTEND, DL, ExtVT, V2);
10827 return DAG.getNode(ISD::TRUNCATE, DL, VT,
10828 DAG.getVectorShuffle(ExtVT, DL, V1, V2, Mask));
10830 /// \brief Top-level lowering for x86 vector shuffles.
10832 /// This handles decomposition, canonicalization, and lowering of all x86
10833 /// vector shuffles. Most of the specific lowering strategies are encapsulated
10834 /// above in helper routines. The canonicalization attempts to widen shuffles
10835 /// to involve fewer lanes of wider elements, consolidate symmetric patterns
10836 /// s.t. only one of the two inputs needs to be tested, etc.
10837 static SDValue lowerVectorShuffle(SDValue Op, const X86Subtarget *Subtarget,
10838 SelectionDAG &DAG) {
10839 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10840 ArrayRef<int> Mask = SVOp->getMask();
10841 SDValue V1 = Op.getOperand(0);
10842 SDValue V2 = Op.getOperand(1);
10843 MVT VT = Op.getSimpleValueType();
10844 int NumElements = VT.getVectorNumElements();
10846 bool Is1BitVector = (VT.getScalarType() == MVT::i1);
10848 assert((VT.getSizeInBits() != 64 || Is1BitVector) &&
10849 "Can't lower MMX shuffles");
10851 bool V1IsUndef = V1.getOpcode() == ISD::UNDEF;
10852 bool V2IsUndef = V2.getOpcode() == ISD::UNDEF;
10853 if (V1IsUndef && V2IsUndef)
10854 return DAG.getUNDEF(VT);
10856 // When we create a shuffle node we put the UNDEF node to second operand,
10857 // but in some cases the first operand may be transformed to UNDEF.
10858 // In this case we should just commute the node.
10860 return DAG.getCommutedVectorShuffle(*SVOp);
10862 // Check for non-undef masks pointing at an undef vector and make the masks
10863 // undef as well. This makes it easier to match the shuffle based solely on
10867 if (M >= NumElements) {
10868 SmallVector<int, 8> NewMask(Mask.begin(), Mask.end());
10869 for (int &M : NewMask)
10870 if (M >= NumElements)
10872 return DAG.getVectorShuffle(VT, dl, V1, V2, NewMask);
10875 // We actually see shuffles that are entirely re-arrangements of a set of
10876 // zero inputs. This mostly happens while decomposing complex shuffles into
10877 // simple ones. Directly lower these as a buildvector of zeros.
10878 SmallBitVector Zeroable = computeZeroableShuffleElements(Mask, V1, V2);
10879 if (Zeroable.all())
10880 return getZeroVector(VT, Subtarget, DAG, dl);
10882 // Try to collapse shuffles into using a vector type with fewer elements but
10883 // wider element types. We cap this to not form integers or floating point
10884 // elements wider than 64 bits, but it might be interesting to form i128
10885 // integers to handle flipping the low and high halves of AVX 256-bit vectors.
10886 SmallVector<int, 16> WidenedMask;
10887 if (VT.getScalarSizeInBits() < 64 && !Is1BitVector &&
10888 canWidenShuffleElements(Mask, WidenedMask)) {
10889 MVT NewEltVT = VT.isFloatingPoint()
10890 ? MVT::getFloatingPointVT(VT.getScalarSizeInBits() * 2)
10891 : MVT::getIntegerVT(VT.getScalarSizeInBits() * 2);
10892 MVT NewVT = MVT::getVectorVT(NewEltVT, VT.getVectorNumElements() / 2);
10893 // Make sure that the new vector type is legal. For example, v2f64 isn't
10895 if (DAG.getTargetLoweringInfo().isTypeLegal(NewVT)) {
10896 V1 = DAG.getBitcast(NewVT, V1);
10897 V2 = DAG.getBitcast(NewVT, V2);
10898 return DAG.getBitcast(
10899 VT, DAG.getVectorShuffle(NewVT, dl, V1, V2, WidenedMask));
10903 int NumV1Elements = 0, NumUndefElements = 0, NumV2Elements = 0;
10904 for (int M : SVOp->getMask())
10906 ++NumUndefElements;
10907 else if (M < NumElements)
10912 // Commute the shuffle as needed such that more elements come from V1 than
10913 // V2. This allows us to match the shuffle pattern strictly on how many
10914 // elements come from V1 without handling the symmetric cases.
10915 if (NumV2Elements > NumV1Elements)
10916 return DAG.getCommutedVectorShuffle(*SVOp);
10918 // When the number of V1 and V2 elements are the same, try to minimize the
10919 // number of uses of V2 in the low half of the vector. When that is tied,
10920 // ensure that the sum of indices for V1 is equal to or lower than the sum
10921 // indices for V2. When those are equal, try to ensure that the number of odd
10922 // indices for V1 is lower than the number of odd indices for V2.
10923 if (NumV1Elements == NumV2Elements) {
10924 int LowV1Elements = 0, LowV2Elements = 0;
10925 for (int M : SVOp->getMask().slice(0, NumElements / 2))
10926 if (M >= NumElements)
10930 if (LowV2Elements > LowV1Elements) {
10931 return DAG.getCommutedVectorShuffle(*SVOp);
10932 } else if (LowV2Elements == LowV1Elements) {
10933 int SumV1Indices = 0, SumV2Indices = 0;
10934 for (int i = 0, Size = SVOp->getMask().size(); i < Size; ++i)
10935 if (SVOp->getMask()[i] >= NumElements)
10937 else if (SVOp->getMask()[i] >= 0)
10939 if (SumV2Indices < SumV1Indices) {
10940 return DAG.getCommutedVectorShuffle(*SVOp);
10941 } else if (SumV2Indices == SumV1Indices) {
10942 int NumV1OddIndices = 0, NumV2OddIndices = 0;
10943 for (int i = 0, Size = SVOp->getMask().size(); i < Size; ++i)
10944 if (SVOp->getMask()[i] >= NumElements)
10945 NumV2OddIndices += i % 2;
10946 else if (SVOp->getMask()[i] >= 0)
10947 NumV1OddIndices += i % 2;
10948 if (NumV2OddIndices < NumV1OddIndices)
10949 return DAG.getCommutedVectorShuffle(*SVOp);
10954 // For each vector width, delegate to a specialized lowering routine.
10955 if (VT.getSizeInBits() == 128)
10956 return lower128BitVectorShuffle(Op, V1, V2, VT, Subtarget, DAG);
10958 if (VT.getSizeInBits() == 256)
10959 return lower256BitVectorShuffle(Op, V1, V2, VT, Subtarget, DAG);
10961 if (VT.getSizeInBits() == 512)
10962 return lower512BitVectorShuffle(Op, V1, V2, VT, Subtarget, DAG);
10965 return lower1BitVectorShuffle(Op, V1, V2, VT, Subtarget, DAG);
10966 llvm_unreachable("Unimplemented!");
10969 // This function assumes its argument is a BUILD_VECTOR of constants or
10970 // undef SDNodes. i.e: ISD::isBuildVectorOfConstantSDNodes(BuildVector) is
10972 static bool BUILD_VECTORtoBlendMask(BuildVectorSDNode *BuildVector,
10973 unsigned &MaskValue) {
10975 unsigned NumElems = BuildVector->getNumOperands();
10976 // There are 2 lanes if (NumElems > 8), and 1 lane otherwise.
10977 unsigned NumLanes = (NumElems - 1) / 8 + 1;
10978 unsigned NumElemsInLane = NumElems / NumLanes;
10980 // Blend for v16i16 should be symmetric for the both lanes.
10981 for (unsigned i = 0; i < NumElemsInLane; ++i) {
10982 SDValue EltCond = BuildVector->getOperand(i);
10983 SDValue SndLaneEltCond =
10984 (NumLanes == 2) ? BuildVector->getOperand(i + NumElemsInLane) : EltCond;
10986 int Lane1Cond = -1, Lane2Cond = -1;
10987 if (isa<ConstantSDNode>(EltCond))
10988 Lane1Cond = !isZero(EltCond);
10989 if (isa<ConstantSDNode>(SndLaneEltCond))
10990 Lane2Cond = !isZero(SndLaneEltCond);
10992 if (Lane1Cond == Lane2Cond || Lane2Cond < 0)
10993 // Lane1Cond != 0, means we want the first argument.
10994 // Lane1Cond == 0, means we want the second argument.
10995 // The encoding of this argument is 0 for the first argument, 1
10996 // for the second. Therefore, invert the condition.
10997 MaskValue |= !Lane1Cond << i;
10998 else if (Lane1Cond < 0)
10999 MaskValue |= !Lane2Cond << i;
11006 /// \brief Try to lower a VSELECT instruction to a vector shuffle.
11007 static SDValue lowerVSELECTtoVectorShuffle(SDValue Op,
11008 const X86Subtarget *Subtarget,
11009 SelectionDAG &DAG) {
11010 SDValue Cond = Op.getOperand(0);
11011 SDValue LHS = Op.getOperand(1);
11012 SDValue RHS = Op.getOperand(2);
11014 MVT VT = Op.getSimpleValueType();
11016 if (!ISD::isBuildVectorOfConstantSDNodes(Cond.getNode()))
11018 auto *CondBV = cast<BuildVectorSDNode>(Cond);
11020 // Only non-legal VSELECTs reach this lowering, convert those into generic
11021 // shuffles and re-use the shuffle lowering path for blends.
11022 SmallVector<int, 32> Mask;
11023 for (int i = 0, Size = VT.getVectorNumElements(); i < Size; ++i) {
11024 SDValue CondElt = CondBV->getOperand(i);
11026 isa<ConstantSDNode>(CondElt) ? i + (isZero(CondElt) ? Size : 0) : -1);
11028 return DAG.getVectorShuffle(VT, dl, LHS, RHS, Mask);
11031 SDValue X86TargetLowering::LowerVSELECT(SDValue Op, SelectionDAG &DAG) const {
11032 // A vselect where all conditions and data are constants can be optimized into
11033 // a single vector load by SelectionDAGLegalize::ExpandBUILD_VECTOR().
11034 if (ISD::isBuildVectorOfConstantSDNodes(Op.getOperand(0).getNode()) &&
11035 ISD::isBuildVectorOfConstantSDNodes(Op.getOperand(1).getNode()) &&
11036 ISD::isBuildVectorOfConstantSDNodes(Op.getOperand(2).getNode()))
11039 // Try to lower this to a blend-style vector shuffle. This can handle all
11040 // constant condition cases.
11041 if (SDValue BlendOp = lowerVSELECTtoVectorShuffle(Op, Subtarget, DAG))
11044 // Variable blends are only legal from SSE4.1 onward.
11045 if (!Subtarget->hasSSE41())
11048 // Only some types will be legal on some subtargets. If we can emit a legal
11049 // VSELECT-matching blend, return Op, and but if we need to expand, return
11051 switch (Op.getSimpleValueType().SimpleTy) {
11053 // Most of the vector types have blends past SSE4.1.
11057 // The byte blends for AVX vectors were introduced only in AVX2.
11058 if (Subtarget->hasAVX2())
11065 // AVX-512 BWI and VLX features support VSELECT with i16 elements.
11066 if (Subtarget->hasBWI() && Subtarget->hasVLX())
11069 // FIXME: We should custom lower this by fixing the condition and using i8
11075 static SDValue LowerEXTRACT_VECTOR_ELT_SSE4(SDValue Op, SelectionDAG &DAG) {
11076 MVT VT = Op.getSimpleValueType();
11079 if (!Op.getOperand(0).getSimpleValueType().is128BitVector())
11082 if (VT.getSizeInBits() == 8) {
11083 SDValue Extract = DAG.getNode(X86ISD::PEXTRB, dl, MVT::i32,
11084 Op.getOperand(0), Op.getOperand(1));
11085 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Extract,
11086 DAG.getValueType(VT));
11087 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
11090 if (VT.getSizeInBits() == 16) {
11091 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
11092 // If Idx is 0, it's cheaper to do a move instead of a pextrw.
11094 return DAG.getNode(
11095 ISD::TRUNCATE, dl, MVT::i16,
11096 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
11097 DAG.getBitcast(MVT::v4i32, Op.getOperand(0)),
11098 Op.getOperand(1)));
11099 SDValue Extract = DAG.getNode(X86ISD::PEXTRW, dl, MVT::i32,
11100 Op.getOperand(0), Op.getOperand(1));
11101 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Extract,
11102 DAG.getValueType(VT));
11103 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
11106 if (VT == MVT::f32) {
11107 // EXTRACTPS outputs to a GPR32 register which will require a movd to copy
11108 // the result back to FR32 register. It's only worth matching if the
11109 // result has a single use which is a store or a bitcast to i32. And in
11110 // the case of a store, it's not worth it if the index is a constant 0,
11111 // because a MOVSSmr can be used instead, which is smaller and faster.
11112 if (!Op.hasOneUse())
11114 SDNode *User = *Op.getNode()->use_begin();
11115 if ((User->getOpcode() != ISD::STORE ||
11116 (isa<ConstantSDNode>(Op.getOperand(1)) &&
11117 cast<ConstantSDNode>(Op.getOperand(1))->isNullValue())) &&
11118 (User->getOpcode() != ISD::BITCAST ||
11119 User->getValueType(0) != MVT::i32))
11121 SDValue Extract = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
11122 DAG.getBitcast(MVT::v4i32, Op.getOperand(0)),
11124 return DAG.getBitcast(MVT::f32, Extract);
11127 if (VT == MVT::i32 || VT == MVT::i64) {
11128 // ExtractPS/pextrq works with constant index.
11129 if (isa<ConstantSDNode>(Op.getOperand(1)))
11135 /// Extract one bit from mask vector, like v16i1 or v8i1.
11136 /// AVX-512 feature.
11138 X86TargetLowering::ExtractBitFromMaskVector(SDValue Op, SelectionDAG &DAG) const {
11139 SDValue Vec = Op.getOperand(0);
11141 MVT VecVT = Vec.getSimpleValueType();
11142 SDValue Idx = Op.getOperand(1);
11143 MVT EltVT = Op.getSimpleValueType();
11145 assert((EltVT == MVT::i1) && "Unexpected operands in ExtractBitFromMaskVector");
11146 assert((VecVT.getVectorNumElements() <= 16 || Subtarget->hasBWI()) &&
11147 "Unexpected vector type in ExtractBitFromMaskVector");
11149 // variable index can't be handled in mask registers,
11150 // extend vector to VR512
11151 if (!isa<ConstantSDNode>(Idx)) {
11152 MVT ExtVT = (VecVT == MVT::v8i1 ? MVT::v8i64 : MVT::v16i32);
11153 SDValue Ext = DAG.getNode(ISD::ZERO_EXTEND, dl, ExtVT, Vec);
11154 SDValue Elt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl,
11155 ExtVT.getVectorElementType(), Ext, Idx);
11156 return DAG.getNode(ISD::TRUNCATE, dl, EltVT, Elt);
11159 unsigned IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
11160 const TargetRegisterClass* rc = getRegClassFor(VecVT);
11161 if (!Subtarget->hasDQI() && (VecVT.getVectorNumElements() <= 8))
11162 rc = getRegClassFor(MVT::v16i1);
11163 unsigned MaxSift = rc->getSize()*8 - 1;
11164 Vec = DAG.getNode(X86ISD::VSHLI, dl, VecVT, Vec,
11165 DAG.getConstant(MaxSift - IdxVal, dl, MVT::i8));
11166 Vec = DAG.getNode(X86ISD::VSRLI, dl, VecVT, Vec,
11167 DAG.getConstant(MaxSift, dl, MVT::i8));
11168 return DAG.getNode(X86ISD::VEXTRACT, dl, MVT::i1, Vec,
11169 DAG.getIntPtrConstant(0, dl));
11173 X86TargetLowering::LowerEXTRACT_VECTOR_ELT(SDValue Op,
11174 SelectionDAG &DAG) const {
11176 SDValue Vec = Op.getOperand(0);
11177 MVT VecVT = Vec.getSimpleValueType();
11178 SDValue Idx = Op.getOperand(1);
11180 if (Op.getSimpleValueType() == MVT::i1)
11181 return ExtractBitFromMaskVector(Op, DAG);
11183 if (!isa<ConstantSDNode>(Idx)) {
11184 if (VecVT.is512BitVector() ||
11185 (VecVT.is256BitVector() && Subtarget->hasInt256() &&
11186 VecVT.getVectorElementType().getSizeInBits() == 32)) {
11189 MVT::getIntegerVT(VecVT.getVectorElementType().getSizeInBits());
11190 MVT MaskVT = MVT::getVectorVT(MaskEltVT, VecVT.getSizeInBits() /
11191 MaskEltVT.getSizeInBits());
11193 Idx = DAG.getZExtOrTrunc(Idx, dl, MaskEltVT);
11194 auto PtrVT = getPointerTy(DAG.getDataLayout());
11195 SDValue Mask = DAG.getNode(X86ISD::VINSERT, dl, MaskVT,
11196 getZeroVector(MaskVT, Subtarget, DAG, dl), Idx,
11197 DAG.getConstant(0, dl, PtrVT));
11198 SDValue Perm = DAG.getNode(X86ISD::VPERMV, dl, VecVT, Mask, Vec);
11199 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, Op.getValueType(), Perm,
11200 DAG.getConstant(0, dl, PtrVT));
11205 // If this is a 256-bit vector result, first extract the 128-bit vector and
11206 // then extract the element from the 128-bit vector.
11207 if (VecVT.is256BitVector() || VecVT.is512BitVector()) {
11209 unsigned IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
11210 // Get the 128-bit vector.
11211 Vec = Extract128BitVector(Vec, IdxVal, DAG, dl);
11212 MVT EltVT = VecVT.getVectorElementType();
11214 unsigned ElemsPerChunk = 128 / EltVT.getSizeInBits();
11216 //if (IdxVal >= NumElems/2)
11217 // IdxVal -= NumElems/2;
11218 IdxVal -= (IdxVal/ElemsPerChunk)*ElemsPerChunk;
11219 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, Op.getValueType(), Vec,
11220 DAG.getConstant(IdxVal, dl, MVT::i32));
11223 assert(VecVT.is128BitVector() && "Unexpected vector length");
11225 if (Subtarget->hasSSE41())
11226 if (SDValue Res = LowerEXTRACT_VECTOR_ELT_SSE4(Op, DAG))
11229 MVT VT = Op.getSimpleValueType();
11230 // TODO: handle v16i8.
11231 if (VT.getSizeInBits() == 16) {
11232 SDValue Vec = Op.getOperand(0);
11233 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
11235 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i16,
11236 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
11237 DAG.getBitcast(MVT::v4i32, Vec),
11238 Op.getOperand(1)));
11239 // Transform it so it match pextrw which produces a 32-bit result.
11240 MVT EltVT = MVT::i32;
11241 SDValue Extract = DAG.getNode(X86ISD::PEXTRW, dl, EltVT,
11242 Op.getOperand(0), Op.getOperand(1));
11243 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, EltVT, Extract,
11244 DAG.getValueType(VT));
11245 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
11248 if (VT.getSizeInBits() == 32) {
11249 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
11253 // SHUFPS the element to the lowest double word, then movss.
11254 int Mask[4] = { static_cast<int>(Idx), -1, -1, -1 };
11255 MVT VVT = Op.getOperand(0).getSimpleValueType();
11256 SDValue Vec = DAG.getVectorShuffle(VVT, dl, Op.getOperand(0),
11257 DAG.getUNDEF(VVT), Mask);
11258 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, Vec,
11259 DAG.getIntPtrConstant(0, dl));
11262 if (VT.getSizeInBits() == 64) {
11263 // FIXME: .td only matches this for <2 x f64>, not <2 x i64> on 32b
11264 // FIXME: seems like this should be unnecessary if mov{h,l}pd were taught
11265 // to match extract_elt for f64.
11266 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
11270 // UNPCKHPD the element to the lowest double word, then movsd.
11271 // Note if the lower 64 bits of the result of the UNPCKHPD is then stored
11272 // to a f64mem, the whole operation is folded into a single MOVHPDmr.
11273 int Mask[2] = { 1, -1 };
11274 MVT VVT = Op.getOperand(0).getSimpleValueType();
11275 SDValue Vec = DAG.getVectorShuffle(VVT, dl, Op.getOperand(0),
11276 DAG.getUNDEF(VVT), Mask);
11277 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, Vec,
11278 DAG.getIntPtrConstant(0, dl));
11284 /// Insert one bit to mask vector, like v16i1 or v8i1.
11285 /// AVX-512 feature.
11287 X86TargetLowering::InsertBitToMaskVector(SDValue Op, SelectionDAG &DAG) const {
11289 SDValue Vec = Op.getOperand(0);
11290 SDValue Elt = Op.getOperand(1);
11291 SDValue Idx = Op.getOperand(2);
11292 MVT VecVT = Vec.getSimpleValueType();
11294 if (!isa<ConstantSDNode>(Idx)) {
11295 // Non constant index. Extend source and destination,
11296 // insert element and then truncate the result.
11297 MVT ExtVecVT = (VecVT == MVT::v8i1 ? MVT::v8i64 : MVT::v16i32);
11298 MVT ExtEltVT = (VecVT == MVT::v8i1 ? MVT::i64 : MVT::i32);
11299 SDValue ExtOp = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, ExtVecVT,
11300 DAG.getNode(ISD::ZERO_EXTEND, dl, ExtVecVT, Vec),
11301 DAG.getNode(ISD::ZERO_EXTEND, dl, ExtEltVT, Elt), Idx);
11302 return DAG.getNode(ISD::TRUNCATE, dl, VecVT, ExtOp);
11305 unsigned IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
11306 SDValue EltInVec = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VecVT, Elt);
11308 EltInVec = DAG.getNode(X86ISD::VSHLI, dl, VecVT, EltInVec,
11309 DAG.getConstant(IdxVal, dl, MVT::i8));
11310 if (Vec.getOpcode() == ISD::UNDEF)
11312 return DAG.getNode(ISD::OR, dl, VecVT, Vec, EltInVec);
11315 SDValue X86TargetLowering::LowerINSERT_VECTOR_ELT(SDValue Op,
11316 SelectionDAG &DAG) const {
11317 MVT VT = Op.getSimpleValueType();
11318 MVT EltVT = VT.getVectorElementType();
11320 if (EltVT == MVT::i1)
11321 return InsertBitToMaskVector(Op, DAG);
11324 SDValue N0 = Op.getOperand(0);
11325 SDValue N1 = Op.getOperand(1);
11326 SDValue N2 = Op.getOperand(2);
11327 if (!isa<ConstantSDNode>(N2))
11329 auto *N2C = cast<ConstantSDNode>(N2);
11330 unsigned IdxVal = N2C->getZExtValue();
11332 // If the vector is wider than 128 bits, extract the 128-bit subvector, insert
11333 // into that, and then insert the subvector back into the result.
11334 if (VT.is256BitVector() || VT.is512BitVector()) {
11335 // With a 256-bit vector, we can insert into the zero element efficiently
11336 // using a blend if we have AVX or AVX2 and the right data type.
11337 if (VT.is256BitVector() && IdxVal == 0) {
11338 // TODO: It is worthwhile to cast integer to floating point and back
11339 // and incur a domain crossing penalty if that's what we'll end up
11340 // doing anyway after extracting to a 128-bit vector.
11341 if ((Subtarget->hasAVX() && (EltVT == MVT::f64 || EltVT == MVT::f32)) ||
11342 (Subtarget->hasAVX2() && EltVT == MVT::i32)) {
11343 SDValue N1Vec = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, N1);
11344 N2 = DAG.getIntPtrConstant(1, dl);
11345 return DAG.getNode(X86ISD::BLENDI, dl, VT, N0, N1Vec, N2);
11349 // Get the desired 128-bit vector chunk.
11350 SDValue V = Extract128BitVector(N0, IdxVal, DAG, dl);
11352 // Insert the element into the desired chunk.
11353 unsigned NumEltsIn128 = 128 / EltVT.getSizeInBits();
11354 unsigned IdxIn128 = IdxVal - (IdxVal / NumEltsIn128) * NumEltsIn128;
11356 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, V.getValueType(), V, N1,
11357 DAG.getConstant(IdxIn128, dl, MVT::i32));
11359 // Insert the changed part back into the bigger vector
11360 return Insert128BitVector(N0, V, IdxVal, DAG, dl);
11362 assert(VT.is128BitVector() && "Only 128-bit vector types should be left!");
11364 if (Subtarget->hasSSE41()) {
11365 if (EltVT.getSizeInBits() == 8 || EltVT.getSizeInBits() == 16) {
11367 if (VT == MVT::v8i16) {
11368 Opc = X86ISD::PINSRW;
11370 assert(VT == MVT::v16i8);
11371 Opc = X86ISD::PINSRB;
11374 // Transform it so it match pinsr{b,w} which expects a GR32 as its second
11376 if (N1.getValueType() != MVT::i32)
11377 N1 = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, N1);
11378 if (N2.getValueType() != MVT::i32)
11379 N2 = DAG.getIntPtrConstant(IdxVal, dl);
11380 return DAG.getNode(Opc, dl, VT, N0, N1, N2);
11383 if (EltVT == MVT::f32) {
11384 // Bits [7:6] of the constant are the source select. This will always be
11385 // zero here. The DAG Combiner may combine an extract_elt index into
11386 // these bits. For example (insert (extract, 3), 2) could be matched by
11387 // putting the '3' into bits [7:6] of X86ISD::INSERTPS.
11388 // Bits [5:4] of the constant are the destination select. This is the
11389 // value of the incoming immediate.
11390 // Bits [3:0] of the constant are the zero mask. The DAG Combiner may
11391 // combine either bitwise AND or insert of float 0.0 to set these bits.
11393 bool MinSize = DAG.getMachineFunction().getFunction()->optForMinSize();
11394 if (IdxVal == 0 && (!MinSize || !MayFoldLoad(N1))) {
11395 // If this is an insertion of 32-bits into the low 32-bits of
11396 // a vector, we prefer to generate a blend with immediate rather
11397 // than an insertps. Blends are simpler operations in hardware and so
11398 // will always have equal or better performance than insertps.
11399 // But if optimizing for size and there's a load folding opportunity,
11400 // generate insertps because blendps does not have a 32-bit memory
11402 N2 = DAG.getIntPtrConstant(1, dl);
11403 N1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4f32, N1);
11404 return DAG.getNode(X86ISD::BLENDI, dl, VT, N0, N1, N2);
11406 N2 = DAG.getIntPtrConstant(IdxVal << 4, dl);
11407 // Create this as a scalar to vector..
11408 N1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4f32, N1);
11409 return DAG.getNode(X86ISD::INSERTPS, dl, VT, N0, N1, N2);
11412 if (EltVT == MVT::i32 || EltVT == MVT::i64) {
11413 // PINSR* works with constant index.
11418 if (EltVT == MVT::i8)
11421 if (EltVT.getSizeInBits() == 16) {
11422 // Transform it so it match pinsrw which expects a 16-bit value in a GR32
11423 // as its second argument.
11424 if (N1.getValueType() != MVT::i32)
11425 N1 = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, N1);
11426 if (N2.getValueType() != MVT::i32)
11427 N2 = DAG.getIntPtrConstant(IdxVal, dl);
11428 return DAG.getNode(X86ISD::PINSRW, dl, VT, N0, N1, N2);
11433 static SDValue LowerSCALAR_TO_VECTOR(SDValue Op, SelectionDAG &DAG) {
11435 MVT OpVT = Op.getSimpleValueType();
11437 // If this is a 256-bit vector result, first insert into a 128-bit
11438 // vector and then insert into the 256-bit vector.
11439 if (!OpVT.is128BitVector()) {
11440 // Insert into a 128-bit vector.
11441 unsigned SizeFactor = OpVT.getSizeInBits()/128;
11442 MVT VT128 = MVT::getVectorVT(OpVT.getVectorElementType(),
11443 OpVT.getVectorNumElements() / SizeFactor);
11445 Op = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT128, Op.getOperand(0));
11447 // Insert the 128-bit vector.
11448 return Insert128BitVector(DAG.getUNDEF(OpVT), Op, 0, DAG, dl);
11451 if (OpVT == MVT::v1i64 &&
11452 Op.getOperand(0).getValueType() == MVT::i64)
11453 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v1i64, Op.getOperand(0));
11455 SDValue AnyExt = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, Op.getOperand(0));
11456 assert(OpVT.is128BitVector() && "Expected an SSE type!");
11457 return DAG.getBitcast(
11458 OpVT, DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32, AnyExt));
11461 // Lower a node with an EXTRACT_SUBVECTOR opcode. This may result in
11462 // a simple subregister reference or explicit instructions to grab
11463 // upper bits of a vector.
11464 static SDValue LowerEXTRACT_SUBVECTOR(SDValue Op, const X86Subtarget *Subtarget,
11465 SelectionDAG &DAG) {
11467 SDValue In = Op.getOperand(0);
11468 SDValue Idx = Op.getOperand(1);
11469 unsigned IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
11470 MVT ResVT = Op.getSimpleValueType();
11471 MVT InVT = In.getSimpleValueType();
11473 if (Subtarget->hasFp256()) {
11474 if (ResVT.is128BitVector() &&
11475 (InVT.is256BitVector() || InVT.is512BitVector()) &&
11476 isa<ConstantSDNode>(Idx)) {
11477 return Extract128BitVector(In, IdxVal, DAG, dl);
11479 if (ResVT.is256BitVector() && InVT.is512BitVector() &&
11480 isa<ConstantSDNode>(Idx)) {
11481 return Extract256BitVector(In, IdxVal, DAG, dl);
11487 // Lower a node with an INSERT_SUBVECTOR opcode. This may result in a
11488 // simple superregister reference or explicit instructions to insert
11489 // the upper bits of a vector.
11490 static SDValue LowerINSERT_SUBVECTOR(SDValue Op, const X86Subtarget *Subtarget,
11491 SelectionDAG &DAG) {
11492 if (!Subtarget->hasAVX())
11496 SDValue Vec = Op.getOperand(0);
11497 SDValue SubVec = Op.getOperand(1);
11498 SDValue Idx = Op.getOperand(2);
11500 if (!isa<ConstantSDNode>(Idx))
11503 unsigned IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
11504 MVT OpVT = Op.getSimpleValueType();
11505 MVT SubVecVT = SubVec.getSimpleValueType();
11507 // Fold two 16-byte subvector loads into one 32-byte load:
11508 // (insert_subvector (insert_subvector undef, (load addr), 0),
11509 // (load addr + 16), Elts/2)
11511 if ((IdxVal == OpVT.getVectorNumElements() / 2) &&
11512 Vec.getOpcode() == ISD::INSERT_SUBVECTOR &&
11513 OpVT.is256BitVector() && SubVecVT.is128BitVector()) {
11514 auto *Idx2 = dyn_cast<ConstantSDNode>(Vec.getOperand(2));
11515 if (Idx2 && Idx2->getZExtValue() == 0) {
11516 SDValue SubVec2 = Vec.getOperand(1);
11517 // If needed, look through a bitcast to get to the load.
11518 if (SubVec2.getNode() && SubVec2.getOpcode() == ISD::BITCAST)
11519 SubVec2 = SubVec2.getOperand(0);
11521 if (auto *FirstLd = dyn_cast<LoadSDNode>(SubVec2)) {
11523 unsigned Alignment = FirstLd->getAlignment();
11524 unsigned AS = FirstLd->getAddressSpace();
11525 const X86TargetLowering *TLI = Subtarget->getTargetLowering();
11526 if (TLI->allowsMemoryAccess(*DAG.getContext(), DAG.getDataLayout(),
11527 OpVT, AS, Alignment, &Fast) && Fast) {
11528 SDValue Ops[] = { SubVec2, SubVec };
11529 if (SDValue Ld = EltsFromConsecutiveLoads(OpVT, Ops, dl, DAG, false))
11536 if ((OpVT.is256BitVector() || OpVT.is512BitVector()) &&
11537 SubVecVT.is128BitVector())
11538 return Insert128BitVector(Vec, SubVec, IdxVal, DAG, dl);
11540 if (OpVT.is512BitVector() && SubVecVT.is256BitVector())
11541 return Insert256BitVector(Vec, SubVec, IdxVal, DAG, dl);
11543 if (OpVT.getVectorElementType() == MVT::i1) {
11544 if (IdxVal == 0 && Vec.getOpcode() == ISD::UNDEF) // the operation is legal
11546 SDValue ZeroIdx = DAG.getIntPtrConstant(0, dl);
11547 SDValue Undef = DAG.getUNDEF(OpVT);
11548 unsigned NumElems = OpVT.getVectorNumElements();
11549 SDValue ShiftBits = DAG.getConstant(NumElems/2, dl, MVT::i8);
11551 if (IdxVal == OpVT.getVectorNumElements() / 2) {
11552 // Zero upper bits of the Vec
11553 Vec = DAG.getNode(X86ISD::VSHLI, dl, OpVT, Vec, ShiftBits);
11554 Vec = DAG.getNode(X86ISD::VSRLI, dl, OpVT, Vec, ShiftBits);
11556 SDValue Vec2 = DAG.getNode(ISD::INSERT_SUBVECTOR, dl, OpVT, Undef,
11558 Vec2 = DAG.getNode(X86ISD::VSHLI, dl, OpVT, Vec2, ShiftBits);
11559 return DAG.getNode(ISD::OR, dl, OpVT, Vec, Vec2);
11562 SDValue Vec2 = DAG.getNode(ISD::INSERT_SUBVECTOR, dl, OpVT, Undef,
11564 // Zero upper bits of the Vec2
11565 Vec2 = DAG.getNode(X86ISD::VSHLI, dl, OpVT, Vec2, ShiftBits);
11566 Vec2 = DAG.getNode(X86ISD::VSRLI, dl, OpVT, Vec2, ShiftBits);
11567 // Zero lower bits of the Vec
11568 Vec = DAG.getNode(X86ISD::VSRLI, dl, OpVT, Vec, ShiftBits);
11569 Vec = DAG.getNode(X86ISD::VSHLI, dl, OpVT, Vec, ShiftBits);
11570 // Merge them together
11571 return DAG.getNode(ISD::OR, dl, OpVT, Vec, Vec2);
11577 // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
11578 // their target countpart wrapped in the X86ISD::Wrapper node. Suppose N is
11579 // one of the above mentioned nodes. It has to be wrapped because otherwise
11580 // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
11581 // be used to form addressing mode. These wrapped nodes will be selected
11584 X86TargetLowering::LowerConstantPool(SDValue Op, SelectionDAG &DAG) const {
11585 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
11587 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
11588 // global base reg.
11589 unsigned char OpFlag = 0;
11590 unsigned WrapperKind = X86ISD::Wrapper;
11591 CodeModel::Model M = DAG.getTarget().getCodeModel();
11593 if (Subtarget->isPICStyleRIPRel() &&
11594 (M == CodeModel::Small || M == CodeModel::Kernel))
11595 WrapperKind = X86ISD::WrapperRIP;
11596 else if (Subtarget->isPICStyleGOT())
11597 OpFlag = X86II::MO_GOTOFF;
11598 else if (Subtarget->isPICStyleStubPIC())
11599 OpFlag = X86II::MO_PIC_BASE_OFFSET;
11601 auto PtrVT = getPointerTy(DAG.getDataLayout());
11602 SDValue Result = DAG.getTargetConstantPool(
11603 CP->getConstVal(), PtrVT, CP->getAlignment(), CP->getOffset(), OpFlag);
11605 Result = DAG.getNode(WrapperKind, DL, PtrVT, Result);
11606 // With PIC, the address is actually $g + Offset.
11609 DAG.getNode(ISD::ADD, DL, PtrVT,
11610 DAG.getNode(X86ISD::GlobalBaseReg, SDLoc(), PtrVT), Result);
11616 SDValue X86TargetLowering::LowerJumpTable(SDValue Op, SelectionDAG &DAG) const {
11617 JumpTableSDNode *JT = cast<JumpTableSDNode>(Op);
11619 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
11620 // global base reg.
11621 unsigned char OpFlag = 0;
11622 unsigned WrapperKind = X86ISD::Wrapper;
11623 CodeModel::Model M = DAG.getTarget().getCodeModel();
11625 if (Subtarget->isPICStyleRIPRel() &&
11626 (M == CodeModel::Small || M == CodeModel::Kernel))
11627 WrapperKind = X86ISD::WrapperRIP;
11628 else if (Subtarget->isPICStyleGOT())
11629 OpFlag = X86II::MO_GOTOFF;
11630 else if (Subtarget->isPICStyleStubPIC())
11631 OpFlag = X86II::MO_PIC_BASE_OFFSET;
11633 auto PtrVT = getPointerTy(DAG.getDataLayout());
11634 SDValue Result = DAG.getTargetJumpTable(JT->getIndex(), PtrVT, OpFlag);
11636 Result = DAG.getNode(WrapperKind, DL, PtrVT, Result);
11638 // With PIC, the address is actually $g + Offset.
11641 DAG.getNode(ISD::ADD, DL, PtrVT,
11642 DAG.getNode(X86ISD::GlobalBaseReg, SDLoc(), PtrVT), Result);
11648 X86TargetLowering::LowerExternalSymbol(SDValue Op, SelectionDAG &DAG) const {
11649 const char *Sym = cast<ExternalSymbolSDNode>(Op)->getSymbol();
11651 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
11652 // global base reg.
11653 unsigned char OpFlag = 0;
11654 unsigned WrapperKind = X86ISD::Wrapper;
11655 CodeModel::Model M = DAG.getTarget().getCodeModel();
11657 if (Subtarget->isPICStyleRIPRel() &&
11658 (M == CodeModel::Small || M == CodeModel::Kernel)) {
11659 if (Subtarget->isTargetDarwin() || Subtarget->isTargetELF())
11660 OpFlag = X86II::MO_GOTPCREL;
11661 WrapperKind = X86ISD::WrapperRIP;
11662 } else if (Subtarget->isPICStyleGOT()) {
11663 OpFlag = X86II::MO_GOT;
11664 } else if (Subtarget->isPICStyleStubPIC()) {
11665 OpFlag = X86II::MO_DARWIN_NONLAZY_PIC_BASE;
11666 } else if (Subtarget->isPICStyleStubNoDynamic()) {
11667 OpFlag = X86II::MO_DARWIN_NONLAZY;
11670 auto PtrVT = getPointerTy(DAG.getDataLayout());
11671 SDValue Result = DAG.getTargetExternalSymbol(Sym, PtrVT, OpFlag);
11674 Result = DAG.getNode(WrapperKind, DL, PtrVT, Result);
11676 // With PIC, the address is actually $g + Offset.
11677 if (DAG.getTarget().getRelocationModel() == Reloc::PIC_ &&
11678 !Subtarget->is64Bit()) {
11680 DAG.getNode(ISD::ADD, DL, PtrVT,
11681 DAG.getNode(X86ISD::GlobalBaseReg, SDLoc(), PtrVT), Result);
11684 // For symbols that require a load from a stub to get the address, emit the
11686 if (isGlobalStubReference(OpFlag))
11687 Result = DAG.getLoad(PtrVT, DL, DAG.getEntryNode(), Result,
11688 MachinePointerInfo::getGOT(DAG.getMachineFunction()),
11689 false, false, false, 0);
11695 X86TargetLowering::LowerBlockAddress(SDValue Op, SelectionDAG &DAG) const {
11696 // Create the TargetBlockAddressAddress node.
11697 unsigned char OpFlags =
11698 Subtarget->ClassifyBlockAddressReference();
11699 CodeModel::Model M = DAG.getTarget().getCodeModel();
11700 const BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress();
11701 int64_t Offset = cast<BlockAddressSDNode>(Op)->getOffset();
11703 auto PtrVT = getPointerTy(DAG.getDataLayout());
11704 SDValue Result = DAG.getTargetBlockAddress(BA, PtrVT, Offset, OpFlags);
11706 if (Subtarget->isPICStyleRIPRel() &&
11707 (M == CodeModel::Small || M == CodeModel::Kernel))
11708 Result = DAG.getNode(X86ISD::WrapperRIP, dl, PtrVT, Result);
11710 Result = DAG.getNode(X86ISD::Wrapper, dl, PtrVT, Result);
11712 // With PIC, the address is actually $g + Offset.
11713 if (isGlobalRelativeToPICBase(OpFlags)) {
11714 Result = DAG.getNode(ISD::ADD, dl, PtrVT,
11715 DAG.getNode(X86ISD::GlobalBaseReg, dl, PtrVT), Result);
11722 X86TargetLowering::LowerGlobalAddress(const GlobalValue *GV, SDLoc dl,
11723 int64_t Offset, SelectionDAG &DAG) const {
11724 // Create the TargetGlobalAddress node, folding in the constant
11725 // offset if it is legal.
11726 unsigned char OpFlags =
11727 Subtarget->ClassifyGlobalReference(GV, DAG.getTarget());
11728 CodeModel::Model M = DAG.getTarget().getCodeModel();
11729 auto PtrVT = getPointerTy(DAG.getDataLayout());
11731 if (OpFlags == X86II::MO_NO_FLAG &&
11732 X86::isOffsetSuitableForCodeModel(Offset, M)) {
11733 // A direct static reference to a global.
11734 Result = DAG.getTargetGlobalAddress(GV, dl, PtrVT, Offset);
11737 Result = DAG.getTargetGlobalAddress(GV, dl, PtrVT, 0, OpFlags);
11740 if (Subtarget->isPICStyleRIPRel() &&
11741 (M == CodeModel::Small || M == CodeModel::Kernel))
11742 Result = DAG.getNode(X86ISD::WrapperRIP, dl, PtrVT, Result);
11744 Result = DAG.getNode(X86ISD::Wrapper, dl, PtrVT, Result);
11746 // With PIC, the address is actually $g + Offset.
11747 if (isGlobalRelativeToPICBase(OpFlags)) {
11748 Result = DAG.getNode(ISD::ADD, dl, PtrVT,
11749 DAG.getNode(X86ISD::GlobalBaseReg, dl, PtrVT), Result);
11752 // For globals that require a load from a stub to get the address, emit the
11754 if (isGlobalStubReference(OpFlags))
11755 Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Result,
11756 MachinePointerInfo::getGOT(DAG.getMachineFunction()),
11757 false, false, false, 0);
11759 // If there was a non-zero offset that we didn't fold, create an explicit
11760 // addition for it.
11762 Result = DAG.getNode(ISD::ADD, dl, PtrVT, Result,
11763 DAG.getConstant(Offset, dl, PtrVT));
11769 X86TargetLowering::LowerGlobalAddress(SDValue Op, SelectionDAG &DAG) const {
11770 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
11771 int64_t Offset = cast<GlobalAddressSDNode>(Op)->getOffset();
11772 return LowerGlobalAddress(GV, SDLoc(Op), Offset, DAG);
11776 GetTLSADDR(SelectionDAG &DAG, SDValue Chain, GlobalAddressSDNode *GA,
11777 SDValue *InFlag, const EVT PtrVT, unsigned ReturnReg,
11778 unsigned char OperandFlags, bool LocalDynamic = false) {
11779 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
11780 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
11782 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(), dl,
11783 GA->getValueType(0),
11787 X86ISD::NodeType CallType = LocalDynamic ? X86ISD::TLSBASEADDR
11791 SDValue Ops[] = { Chain, TGA, *InFlag };
11792 Chain = DAG.getNode(CallType, dl, NodeTys, Ops);
11794 SDValue Ops[] = { Chain, TGA };
11795 Chain = DAG.getNode(CallType, dl, NodeTys, Ops);
11798 // TLSADDR will be codegen'ed as call. Inform MFI that function has calls.
11799 MFI->setAdjustsStack(true);
11800 MFI->setHasCalls(true);
11802 SDValue Flag = Chain.getValue(1);
11803 return DAG.getCopyFromReg(Chain, dl, ReturnReg, PtrVT, Flag);
11806 // Lower ISD::GlobalTLSAddress using the "general dynamic" model, 32 bit
11808 LowerToTLSGeneralDynamicModel32(GlobalAddressSDNode *GA, SelectionDAG &DAG,
11811 SDLoc dl(GA); // ? function entry point might be better
11812 SDValue Chain = DAG.getCopyToReg(DAG.getEntryNode(), dl, X86::EBX,
11813 DAG.getNode(X86ISD::GlobalBaseReg,
11814 SDLoc(), PtrVT), InFlag);
11815 InFlag = Chain.getValue(1);
11817 return GetTLSADDR(DAG, Chain, GA, &InFlag, PtrVT, X86::EAX, X86II::MO_TLSGD);
11820 // Lower ISD::GlobalTLSAddress using the "general dynamic" model, 64 bit
11822 LowerToTLSGeneralDynamicModel64(GlobalAddressSDNode *GA, SelectionDAG &DAG,
11824 return GetTLSADDR(DAG, DAG.getEntryNode(), GA, nullptr, PtrVT,
11825 X86::RAX, X86II::MO_TLSGD);
11828 static SDValue LowerToTLSLocalDynamicModel(GlobalAddressSDNode *GA,
11834 // Get the start address of the TLS block for this module.
11835 X86MachineFunctionInfo* MFI = DAG.getMachineFunction()
11836 .getInfo<X86MachineFunctionInfo>();
11837 MFI->incNumLocalDynamicTLSAccesses();
11841 Base = GetTLSADDR(DAG, DAG.getEntryNode(), GA, nullptr, PtrVT, X86::RAX,
11842 X86II::MO_TLSLD, /*LocalDynamic=*/true);
11845 SDValue Chain = DAG.getCopyToReg(DAG.getEntryNode(), dl, X86::EBX,
11846 DAG.getNode(X86ISD::GlobalBaseReg, SDLoc(), PtrVT), InFlag);
11847 InFlag = Chain.getValue(1);
11848 Base = GetTLSADDR(DAG, Chain, GA, &InFlag, PtrVT, X86::EAX,
11849 X86II::MO_TLSLDM, /*LocalDynamic=*/true);
11852 // Note: the CleanupLocalDynamicTLSPass will remove redundant computations
11856 unsigned char OperandFlags = X86II::MO_DTPOFF;
11857 unsigned WrapperKind = X86ISD::Wrapper;
11858 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(), dl,
11859 GA->getValueType(0),
11860 GA->getOffset(), OperandFlags);
11861 SDValue Offset = DAG.getNode(WrapperKind, dl, PtrVT, TGA);
11863 // Add x@dtpoff with the base.
11864 return DAG.getNode(ISD::ADD, dl, PtrVT, Offset, Base);
11867 // Lower ISD::GlobalTLSAddress using the "initial exec" or "local exec" model.
11868 static SDValue LowerToTLSExecModel(GlobalAddressSDNode *GA, SelectionDAG &DAG,
11869 const EVT PtrVT, TLSModel::Model model,
11870 bool is64Bit, bool isPIC) {
11873 // Get the Thread Pointer, which is %gs:0 (32-bit) or %fs:0 (64-bit).
11874 Value *Ptr = Constant::getNullValue(Type::getInt8PtrTy(*DAG.getContext(),
11875 is64Bit ? 257 : 256));
11877 SDValue ThreadPointer =
11878 DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), DAG.getIntPtrConstant(0, dl),
11879 MachinePointerInfo(Ptr), false, false, false, 0);
11881 unsigned char OperandFlags = 0;
11882 // Most TLS accesses are not RIP relative, even on x86-64. One exception is
11884 unsigned WrapperKind = X86ISD::Wrapper;
11885 if (model == TLSModel::LocalExec) {
11886 OperandFlags = is64Bit ? X86II::MO_TPOFF : X86II::MO_NTPOFF;
11887 } else if (model == TLSModel::InitialExec) {
11889 OperandFlags = X86II::MO_GOTTPOFF;
11890 WrapperKind = X86ISD::WrapperRIP;
11892 OperandFlags = isPIC ? X86II::MO_GOTNTPOFF : X86II::MO_INDNTPOFF;
11895 llvm_unreachable("Unexpected model");
11898 // emit "addl x@ntpoff,%eax" (local exec)
11899 // or "addl x@indntpoff,%eax" (initial exec)
11900 // or "addl x@gotntpoff(%ebx) ,%eax" (initial exec, 32-bit pic)
11902 DAG.getTargetGlobalAddress(GA->getGlobal(), dl, GA->getValueType(0),
11903 GA->getOffset(), OperandFlags);
11904 SDValue Offset = DAG.getNode(WrapperKind, dl, PtrVT, TGA);
11906 if (model == TLSModel::InitialExec) {
11907 if (isPIC && !is64Bit) {
11908 Offset = DAG.getNode(ISD::ADD, dl, PtrVT,
11909 DAG.getNode(X86ISD::GlobalBaseReg, SDLoc(), PtrVT),
11913 Offset = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Offset,
11914 MachinePointerInfo::getGOT(DAG.getMachineFunction()),
11915 false, false, false, 0);
11918 // The address of the thread local variable is the add of the thread
11919 // pointer with the offset of the variable.
11920 return DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, Offset);
11924 X86TargetLowering::LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const {
11926 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
11927 const GlobalValue *GV = GA->getGlobal();
11928 auto PtrVT = getPointerTy(DAG.getDataLayout());
11930 if (Subtarget->isTargetELF()) {
11931 if (DAG.getTarget().Options.EmulatedTLS)
11932 return LowerToTLSEmulatedModel(GA, DAG);
11933 TLSModel::Model model = DAG.getTarget().getTLSModel(GV);
11935 case TLSModel::GeneralDynamic:
11936 if (Subtarget->is64Bit())
11937 return LowerToTLSGeneralDynamicModel64(GA, DAG, PtrVT);
11938 return LowerToTLSGeneralDynamicModel32(GA, DAG, PtrVT);
11939 case TLSModel::LocalDynamic:
11940 return LowerToTLSLocalDynamicModel(GA, DAG, PtrVT,
11941 Subtarget->is64Bit());
11942 case TLSModel::InitialExec:
11943 case TLSModel::LocalExec:
11944 return LowerToTLSExecModel(GA, DAG, PtrVT, model, Subtarget->is64Bit(),
11945 DAG.getTarget().getRelocationModel() ==
11948 llvm_unreachable("Unknown TLS model.");
11951 if (Subtarget->isTargetDarwin()) {
11952 // Darwin only has one model of TLS. Lower to that.
11953 unsigned char OpFlag = 0;
11954 unsigned WrapperKind = Subtarget->isPICStyleRIPRel() ?
11955 X86ISD::WrapperRIP : X86ISD::Wrapper;
11957 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
11958 // global base reg.
11959 bool PIC32 = (DAG.getTarget().getRelocationModel() == Reloc::PIC_) &&
11960 !Subtarget->is64Bit();
11962 OpFlag = X86II::MO_TLVP_PIC_BASE;
11964 OpFlag = X86II::MO_TLVP;
11966 SDValue Result = DAG.getTargetGlobalAddress(GA->getGlobal(), DL,
11967 GA->getValueType(0),
11968 GA->getOffset(), OpFlag);
11969 SDValue Offset = DAG.getNode(WrapperKind, DL, PtrVT, Result);
11971 // With PIC32, the address is actually $g + Offset.
11973 Offset = DAG.getNode(ISD::ADD, DL, PtrVT,
11974 DAG.getNode(X86ISD::GlobalBaseReg, SDLoc(), PtrVT),
11977 // Lowering the machine isd will make sure everything is in the right
11979 SDValue Chain = DAG.getEntryNode();
11980 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
11981 SDValue Args[] = { Chain, Offset };
11982 Chain = DAG.getNode(X86ISD::TLSCALL, DL, NodeTys, Args);
11984 // TLSCALL will be codegen'ed as call. Inform MFI that function has calls.
11985 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
11986 MFI->setAdjustsStack(true);
11988 // And our return value (tls address) is in the standard call return value
11990 unsigned Reg = Subtarget->is64Bit() ? X86::RAX : X86::EAX;
11991 return DAG.getCopyFromReg(Chain, DL, Reg, PtrVT, Chain.getValue(1));
11994 if (Subtarget->isTargetKnownWindowsMSVC() ||
11995 Subtarget->isTargetWindowsGNU()) {
11996 // Just use the implicit TLS architecture
11997 // Need to generate someting similar to:
11998 // mov rdx, qword [gs:abs 58H]; Load pointer to ThreadLocalStorage
12000 // mov ecx, dword [rel _tls_index]: Load index (from C runtime)
12001 // mov rcx, qword [rdx+rcx*8]
12002 // mov eax, .tls$:tlsvar
12003 // [rax+rcx] contains the address
12004 // Windows 64bit: gs:0x58
12005 // Windows 32bit: fs:__tls_array
12008 SDValue Chain = DAG.getEntryNode();
12010 // Get the Thread Pointer, which is %fs:__tls_array (32-bit) or
12011 // %gs:0x58 (64-bit). On MinGW, __tls_array is not available, so directly
12012 // use its literal value of 0x2C.
12013 Value *Ptr = Constant::getNullValue(Subtarget->is64Bit()
12014 ? Type::getInt8PtrTy(*DAG.getContext(),
12016 : Type::getInt32PtrTy(*DAG.getContext(),
12019 SDValue TlsArray = Subtarget->is64Bit()
12020 ? DAG.getIntPtrConstant(0x58, dl)
12021 : (Subtarget->isTargetWindowsGNU()
12022 ? DAG.getIntPtrConstant(0x2C, dl)
12023 : DAG.getExternalSymbol("_tls_array", PtrVT));
12025 SDValue ThreadPointer =
12026 DAG.getLoad(PtrVT, dl, Chain, TlsArray, MachinePointerInfo(Ptr), false,
12030 if (GV->getThreadLocalMode() == GlobalVariable::LocalExecTLSModel) {
12031 res = ThreadPointer;
12033 // Load the _tls_index variable
12034 SDValue IDX = DAG.getExternalSymbol("_tls_index", PtrVT);
12035 if (Subtarget->is64Bit())
12036 IDX = DAG.getExtLoad(ISD::ZEXTLOAD, dl, PtrVT, Chain, IDX,
12037 MachinePointerInfo(), MVT::i32, false, false,
12040 IDX = DAG.getLoad(PtrVT, dl, Chain, IDX, MachinePointerInfo(), false,
12043 auto &DL = DAG.getDataLayout();
12045 DAG.getConstant(Log2_64_Ceil(DL.getPointerSize()), dl, PtrVT);
12046 IDX = DAG.getNode(ISD::SHL, dl, PtrVT, IDX, Scale);
12048 res = DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, IDX);
12051 res = DAG.getLoad(PtrVT, dl, Chain, res, MachinePointerInfo(), false, false,
12054 // Get the offset of start of .tls section
12055 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(), dl,
12056 GA->getValueType(0),
12057 GA->getOffset(), X86II::MO_SECREL);
12058 SDValue Offset = DAG.getNode(X86ISD::Wrapper, dl, PtrVT, TGA);
12060 // The address of the thread local variable is the add of the thread
12061 // pointer with the offset of the variable.
12062 return DAG.getNode(ISD::ADD, dl, PtrVT, res, Offset);
12065 llvm_unreachable("TLS not implemented for this target.");
12068 /// LowerShiftParts - Lower SRA_PARTS and friends, which return two i32 values
12069 /// and take a 2 x i32 value to shift plus a shift amount.
12070 static SDValue LowerShiftParts(SDValue Op, SelectionDAG &DAG) {
12071 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
12072 MVT VT = Op.getSimpleValueType();
12073 unsigned VTBits = VT.getSizeInBits();
12075 bool isSRA = Op.getOpcode() == ISD::SRA_PARTS;
12076 SDValue ShOpLo = Op.getOperand(0);
12077 SDValue ShOpHi = Op.getOperand(1);
12078 SDValue ShAmt = Op.getOperand(2);
12079 // X86ISD::SHLD and X86ISD::SHRD have defined overflow behavior but the
12080 // generic ISD nodes haven't. Insert an AND to be safe, it's optimized away
12082 SDValue SafeShAmt = DAG.getNode(ISD::AND, dl, MVT::i8, ShAmt,
12083 DAG.getConstant(VTBits - 1, dl, MVT::i8));
12084 SDValue Tmp1 = isSRA ? DAG.getNode(ISD::SRA, dl, VT, ShOpHi,
12085 DAG.getConstant(VTBits - 1, dl, MVT::i8))
12086 : DAG.getConstant(0, dl, VT);
12088 SDValue Tmp2, Tmp3;
12089 if (Op.getOpcode() == ISD::SHL_PARTS) {
12090 Tmp2 = DAG.getNode(X86ISD::SHLD, dl, VT, ShOpHi, ShOpLo, ShAmt);
12091 Tmp3 = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, SafeShAmt);
12093 Tmp2 = DAG.getNode(X86ISD::SHRD, dl, VT, ShOpLo, ShOpHi, ShAmt);
12094 Tmp3 = DAG.getNode(isSRA ? ISD::SRA : ISD::SRL, dl, VT, ShOpHi, SafeShAmt);
12097 // If the shift amount is larger or equal than the width of a part we can't
12098 // rely on the results of shld/shrd. Insert a test and select the appropriate
12099 // values for large shift amounts.
12100 SDValue AndNode = DAG.getNode(ISD::AND, dl, MVT::i8, ShAmt,
12101 DAG.getConstant(VTBits, dl, MVT::i8));
12102 SDValue Cond = DAG.getNode(X86ISD::CMP, dl, MVT::i32,
12103 AndNode, DAG.getConstant(0, dl, MVT::i8));
12106 SDValue CC = DAG.getConstant(X86::COND_NE, dl, MVT::i8);
12107 SDValue Ops0[4] = { Tmp2, Tmp3, CC, Cond };
12108 SDValue Ops1[4] = { Tmp3, Tmp1, CC, Cond };
12110 if (Op.getOpcode() == ISD::SHL_PARTS) {
12111 Hi = DAG.getNode(X86ISD::CMOV, dl, VT, Ops0);
12112 Lo = DAG.getNode(X86ISD::CMOV, dl, VT, Ops1);
12114 Lo = DAG.getNode(X86ISD::CMOV, dl, VT, Ops0);
12115 Hi = DAG.getNode(X86ISD::CMOV, dl, VT, Ops1);
12118 SDValue Ops[2] = { Lo, Hi };
12119 return DAG.getMergeValues(Ops, dl);
12122 SDValue X86TargetLowering::LowerSINT_TO_FP(SDValue Op,
12123 SelectionDAG &DAG) const {
12124 SDValue Src = Op.getOperand(0);
12125 MVT SrcVT = Src.getSimpleValueType();
12126 MVT VT = Op.getSimpleValueType();
12129 if (SrcVT.isVector()) {
12130 if (SrcVT == MVT::v2i32 && VT == MVT::v2f64) {
12131 return DAG.getNode(X86ISD::CVTDQ2PD, dl, VT,
12132 DAG.getNode(ISD::CONCAT_VECTORS, dl, MVT::v4i32, Src,
12133 DAG.getUNDEF(SrcVT)));
12135 if (SrcVT.getVectorElementType() == MVT::i1) {
12136 MVT IntegerVT = MVT::getVectorVT(MVT::i32, SrcVT.getVectorNumElements());
12137 return DAG.getNode(ISD::SINT_TO_FP, dl, Op.getValueType(),
12138 DAG.getNode(ISD::SIGN_EXTEND, dl, IntegerVT, Src));
12143 assert(SrcVT <= MVT::i64 && SrcVT >= MVT::i16 &&
12144 "Unknown SINT_TO_FP to lower!");
12146 // These are really Legal; return the operand so the caller accepts it as
12148 if (SrcVT == MVT::i32 && isScalarFPTypeInSSEReg(Op.getValueType()))
12150 if (SrcVT == MVT::i64 && isScalarFPTypeInSSEReg(Op.getValueType()) &&
12151 Subtarget->is64Bit()) {
12155 unsigned Size = SrcVT.getSizeInBits()/8;
12156 MachineFunction &MF = DAG.getMachineFunction();
12157 auto PtrVT = getPointerTy(MF.getDataLayout());
12158 int SSFI = MF.getFrameInfo()->CreateStackObject(Size, Size, false);
12159 SDValue StackSlot = DAG.getFrameIndex(SSFI, PtrVT);
12160 SDValue Chain = DAG.getStore(
12161 DAG.getEntryNode(), dl, Op.getOperand(0), StackSlot,
12162 MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), SSFI), false,
12164 return BuildFILD(Op, SrcVT, Chain, StackSlot, DAG);
12167 SDValue X86TargetLowering::BuildFILD(SDValue Op, EVT SrcVT, SDValue Chain,
12169 SelectionDAG &DAG) const {
12173 bool useSSE = isScalarFPTypeInSSEReg(Op.getValueType());
12175 Tys = DAG.getVTList(MVT::f64, MVT::Other, MVT::Glue);
12177 Tys = DAG.getVTList(Op.getValueType(), MVT::Other);
12179 unsigned ByteSize = SrcVT.getSizeInBits()/8;
12181 FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(StackSlot);
12182 MachineMemOperand *MMO;
12184 int SSFI = FI->getIndex();
12185 MMO = DAG.getMachineFunction().getMachineMemOperand(
12186 MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), SSFI),
12187 MachineMemOperand::MOLoad, ByteSize, ByteSize);
12189 MMO = cast<LoadSDNode>(StackSlot)->getMemOperand();
12190 StackSlot = StackSlot.getOperand(1);
12192 SDValue Ops[] = { Chain, StackSlot, DAG.getValueType(SrcVT) };
12193 SDValue Result = DAG.getMemIntrinsicNode(useSSE ? X86ISD::FILD_FLAG :
12195 Tys, Ops, SrcVT, MMO);
12198 Chain = Result.getValue(1);
12199 SDValue InFlag = Result.getValue(2);
12201 // FIXME: Currently the FST is flagged to the FILD_FLAG. This
12202 // shouldn't be necessary except that RFP cannot be live across
12203 // multiple blocks. When stackifier is fixed, they can be uncoupled.
12204 MachineFunction &MF = DAG.getMachineFunction();
12205 unsigned SSFISize = Op.getValueType().getSizeInBits()/8;
12206 int SSFI = MF.getFrameInfo()->CreateStackObject(SSFISize, SSFISize, false);
12207 auto PtrVT = getPointerTy(MF.getDataLayout());
12208 SDValue StackSlot = DAG.getFrameIndex(SSFI, PtrVT);
12209 Tys = DAG.getVTList(MVT::Other);
12211 Chain, Result, StackSlot, DAG.getValueType(Op.getValueType()), InFlag
12213 MachineMemOperand *MMO = DAG.getMachineFunction().getMachineMemOperand(
12214 MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), SSFI),
12215 MachineMemOperand::MOStore, SSFISize, SSFISize);
12217 Chain = DAG.getMemIntrinsicNode(X86ISD::FST, DL, Tys,
12218 Ops, Op.getValueType(), MMO);
12219 Result = DAG.getLoad(
12220 Op.getValueType(), DL, Chain, StackSlot,
12221 MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), SSFI),
12222 false, false, false, 0);
12228 // LowerUINT_TO_FP_i64 - 64-bit unsigned integer to double expansion.
12229 SDValue X86TargetLowering::LowerUINT_TO_FP_i64(SDValue Op,
12230 SelectionDAG &DAG) const {
12231 // This algorithm is not obvious. Here it is what we're trying to output:
12234 punpckldq (c0), %xmm0 // c0: (uint4){ 0x43300000U, 0x45300000U, 0U, 0U }
12235 subpd (c1), %xmm0 // c1: (double2){ 0x1.0p52, 0x1.0p52 * 0x1.0p32 }
12237 haddpd %xmm0, %xmm0
12239 pshufd $0x4e, %xmm0, %xmm1
12245 LLVMContext *Context = DAG.getContext();
12247 // Build some magic constants.
12248 static const uint32_t CV0[] = { 0x43300000, 0x45300000, 0, 0 };
12249 Constant *C0 = ConstantDataVector::get(*Context, CV0);
12250 auto PtrVT = getPointerTy(DAG.getDataLayout());
12251 SDValue CPIdx0 = DAG.getConstantPool(C0, PtrVT, 16);
12253 SmallVector<Constant*,2> CV1;
12255 ConstantFP::get(*Context, APFloat(APFloat::IEEEdouble,
12256 APInt(64, 0x4330000000000000ULL))));
12258 ConstantFP::get(*Context, APFloat(APFloat::IEEEdouble,
12259 APInt(64, 0x4530000000000000ULL))));
12260 Constant *C1 = ConstantVector::get(CV1);
12261 SDValue CPIdx1 = DAG.getConstantPool(C1, PtrVT, 16);
12263 // Load the 64-bit value into an XMM register.
12264 SDValue XR1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i64,
12267 DAG.getLoad(MVT::v4i32, dl, DAG.getEntryNode(), CPIdx0,
12268 MachinePointerInfo::getConstantPool(DAG.getMachineFunction()),
12269 false, false, false, 16);
12271 getUnpackl(DAG, dl, MVT::v4i32, DAG.getBitcast(MVT::v4i32, XR1), CLod0);
12274 DAG.getLoad(MVT::v2f64, dl, CLod0.getValue(1), CPIdx1,
12275 MachinePointerInfo::getConstantPool(DAG.getMachineFunction()),
12276 false, false, false, 16);
12277 SDValue XR2F = DAG.getBitcast(MVT::v2f64, Unpck1);
12278 // TODO: Are there any fast-math-flags to propagate here?
12279 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::v2f64, XR2F, CLod1);
12282 if (Subtarget->hasSSE3()) {
12283 // FIXME: The 'haddpd' instruction may be slower than 'movhlps + addsd'.
12284 Result = DAG.getNode(X86ISD::FHADD, dl, MVT::v2f64, Sub, Sub);
12286 SDValue S2F = DAG.getBitcast(MVT::v4i32, Sub);
12287 SDValue Shuffle = getTargetShuffleNode(X86ISD::PSHUFD, dl, MVT::v4i32,
12289 Result = DAG.getNode(ISD::FADD, dl, MVT::v2f64,
12290 DAG.getBitcast(MVT::v2f64, Shuffle), Sub);
12293 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Result,
12294 DAG.getIntPtrConstant(0, dl));
12297 // LowerUINT_TO_FP_i32 - 32-bit unsigned integer to float expansion.
12298 SDValue X86TargetLowering::LowerUINT_TO_FP_i32(SDValue Op,
12299 SelectionDAG &DAG) const {
12301 // FP constant to bias correct the final result.
12302 SDValue Bias = DAG.getConstantFP(BitsToDouble(0x4330000000000000ULL), dl,
12305 // Load the 32-bit value into an XMM register.
12306 SDValue Load = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
12309 // Zero out the upper parts of the register.
12310 Load = getShuffleVectorZeroOrUndef(Load, 0, true, Subtarget, DAG);
12312 Load = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
12313 DAG.getBitcast(MVT::v2f64, Load),
12314 DAG.getIntPtrConstant(0, dl));
12316 // Or the load with the bias.
12317 SDValue Or = DAG.getNode(
12318 ISD::OR, dl, MVT::v2i64,
12319 DAG.getBitcast(MVT::v2i64,
12320 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2f64, Load)),
12321 DAG.getBitcast(MVT::v2i64,
12322 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2f64, Bias)));
12324 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
12325 DAG.getBitcast(MVT::v2f64, Or), DAG.getIntPtrConstant(0, dl));
12327 // Subtract the bias.
12328 // TODO: Are there any fast-math-flags to propagate here?
12329 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::f64, Or, Bias);
12331 // Handle final rounding.
12332 EVT DestVT = Op.getValueType();
12334 if (DestVT.bitsLT(MVT::f64))
12335 return DAG.getNode(ISD::FP_ROUND, dl, DestVT, Sub,
12336 DAG.getIntPtrConstant(0, dl));
12337 if (DestVT.bitsGT(MVT::f64))
12338 return DAG.getNode(ISD::FP_EXTEND, dl, DestVT, Sub);
12340 // Handle final rounding.
12344 static SDValue lowerUINT_TO_FP_vXi32(SDValue Op, SelectionDAG &DAG,
12345 const X86Subtarget &Subtarget) {
12346 // The algorithm is the following:
12347 // #ifdef __SSE4_1__
12348 // uint4 lo = _mm_blend_epi16( v, (uint4) 0x4b000000, 0xaa);
12349 // uint4 hi = _mm_blend_epi16( _mm_srli_epi32(v,16),
12350 // (uint4) 0x53000000, 0xaa);
12352 // uint4 lo = (v & (uint4) 0xffff) | (uint4) 0x4b000000;
12353 // uint4 hi = (v >> 16) | (uint4) 0x53000000;
12355 // float4 fhi = (float4) hi - (0x1.0p39f + 0x1.0p23f);
12356 // return (float4) lo + fhi;
12359 SDValue V = Op->getOperand(0);
12360 EVT VecIntVT = V.getValueType();
12361 bool Is128 = VecIntVT == MVT::v4i32;
12362 EVT VecFloatVT = Is128 ? MVT::v4f32 : MVT::v8f32;
12363 // If we convert to something else than the supported type, e.g., to v4f64,
12365 if (VecFloatVT != Op->getValueType(0))
12368 unsigned NumElts = VecIntVT.getVectorNumElements();
12369 assert((VecIntVT == MVT::v4i32 || VecIntVT == MVT::v8i32) &&
12370 "Unsupported custom type");
12371 assert(NumElts <= 8 && "The size of the constant array must be fixed");
12373 // In the #idef/#else code, we have in common:
12374 // - The vector of constants:
12380 // Create the splat vector for 0x4b000000.
12381 SDValue CstLow = DAG.getConstant(0x4b000000, DL, MVT::i32);
12382 SDValue CstLowArray[] = {CstLow, CstLow, CstLow, CstLow,
12383 CstLow, CstLow, CstLow, CstLow};
12384 SDValue VecCstLow = DAG.getNode(ISD::BUILD_VECTOR, DL, VecIntVT,
12385 makeArrayRef(&CstLowArray[0], NumElts));
12386 // Create the splat vector for 0x53000000.
12387 SDValue CstHigh = DAG.getConstant(0x53000000, DL, MVT::i32);
12388 SDValue CstHighArray[] = {CstHigh, CstHigh, CstHigh, CstHigh,
12389 CstHigh, CstHigh, CstHigh, CstHigh};
12390 SDValue VecCstHigh = DAG.getNode(ISD::BUILD_VECTOR, DL, VecIntVT,
12391 makeArrayRef(&CstHighArray[0], NumElts));
12393 // Create the right shift.
12394 SDValue CstShift = DAG.getConstant(16, DL, MVT::i32);
12395 SDValue CstShiftArray[] = {CstShift, CstShift, CstShift, CstShift,
12396 CstShift, CstShift, CstShift, CstShift};
12397 SDValue VecCstShift = DAG.getNode(ISD::BUILD_VECTOR, DL, VecIntVT,
12398 makeArrayRef(&CstShiftArray[0], NumElts));
12399 SDValue HighShift = DAG.getNode(ISD::SRL, DL, VecIntVT, V, VecCstShift);
12402 if (Subtarget.hasSSE41()) {
12403 EVT VecI16VT = Is128 ? MVT::v8i16 : MVT::v16i16;
12404 // uint4 lo = _mm_blend_epi16( v, (uint4) 0x4b000000, 0xaa);
12405 SDValue VecCstLowBitcast = DAG.getBitcast(VecI16VT, VecCstLow);
12406 SDValue VecBitcast = DAG.getBitcast(VecI16VT, V);
12407 // Low will be bitcasted right away, so do not bother bitcasting back to its
12409 Low = DAG.getNode(X86ISD::BLENDI, DL, VecI16VT, VecBitcast,
12410 VecCstLowBitcast, DAG.getConstant(0xaa, DL, MVT::i32));
12411 // uint4 hi = _mm_blend_epi16( _mm_srli_epi32(v,16),
12412 // (uint4) 0x53000000, 0xaa);
12413 SDValue VecCstHighBitcast = DAG.getBitcast(VecI16VT, VecCstHigh);
12414 SDValue VecShiftBitcast = DAG.getBitcast(VecI16VT, HighShift);
12415 // High will be bitcasted right away, so do not bother bitcasting back to
12416 // its original type.
12417 High = DAG.getNode(X86ISD::BLENDI, DL, VecI16VT, VecShiftBitcast,
12418 VecCstHighBitcast, DAG.getConstant(0xaa, DL, MVT::i32));
12420 SDValue CstMask = DAG.getConstant(0xffff, DL, MVT::i32);
12421 SDValue VecCstMask = DAG.getNode(ISD::BUILD_VECTOR, DL, VecIntVT, CstMask,
12422 CstMask, CstMask, CstMask);
12423 // uint4 lo = (v & (uint4) 0xffff) | (uint4) 0x4b000000;
12424 SDValue LowAnd = DAG.getNode(ISD::AND, DL, VecIntVT, V, VecCstMask);
12425 Low = DAG.getNode(ISD::OR, DL, VecIntVT, LowAnd, VecCstLow);
12427 // uint4 hi = (v >> 16) | (uint4) 0x53000000;
12428 High = DAG.getNode(ISD::OR, DL, VecIntVT, HighShift, VecCstHigh);
12431 // Create the vector constant for -(0x1.0p39f + 0x1.0p23f).
12432 SDValue CstFAdd = DAG.getConstantFP(
12433 APFloat(APFloat::IEEEsingle, APInt(32, 0xD3000080)), DL, MVT::f32);
12434 SDValue CstFAddArray[] = {CstFAdd, CstFAdd, CstFAdd, CstFAdd,
12435 CstFAdd, CstFAdd, CstFAdd, CstFAdd};
12436 SDValue VecCstFAdd = DAG.getNode(ISD::BUILD_VECTOR, DL, VecFloatVT,
12437 makeArrayRef(&CstFAddArray[0], NumElts));
12439 // float4 fhi = (float4) hi - (0x1.0p39f + 0x1.0p23f);
12440 SDValue HighBitcast = DAG.getBitcast(VecFloatVT, High);
12441 // TODO: Are there any fast-math-flags to propagate here?
12443 DAG.getNode(ISD::FADD, DL, VecFloatVT, HighBitcast, VecCstFAdd);
12444 // return (float4) lo + fhi;
12445 SDValue LowBitcast = DAG.getBitcast(VecFloatVT, Low);
12446 return DAG.getNode(ISD::FADD, DL, VecFloatVT, LowBitcast, FHigh);
12449 SDValue X86TargetLowering::lowerUINT_TO_FP_vec(SDValue Op,
12450 SelectionDAG &DAG) const {
12451 SDValue N0 = Op.getOperand(0);
12452 MVT SVT = N0.getSimpleValueType();
12455 switch (SVT.SimpleTy) {
12457 llvm_unreachable("Custom UINT_TO_FP is not supported!");
12462 MVT NVT = MVT::getVectorVT(MVT::i32, SVT.getVectorNumElements());
12463 return DAG.getNode(ISD::SINT_TO_FP, dl, Op.getValueType(),
12464 DAG.getNode(ISD::ZERO_EXTEND, dl, NVT, N0));
12468 return lowerUINT_TO_FP_vXi32(Op, DAG, *Subtarget);
12471 if (Subtarget->hasAVX512())
12472 return DAG.getNode(ISD::UINT_TO_FP, dl, Op.getValueType(),
12473 DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v16i32, N0));
12475 llvm_unreachable(nullptr);
12478 SDValue X86TargetLowering::LowerUINT_TO_FP(SDValue Op,
12479 SelectionDAG &DAG) const {
12480 SDValue N0 = Op.getOperand(0);
12482 auto PtrVT = getPointerTy(DAG.getDataLayout());
12484 if (Op.getValueType().isVector())
12485 return lowerUINT_TO_FP_vec(Op, DAG);
12487 // Since UINT_TO_FP is legal (it's marked custom), dag combiner won't
12488 // optimize it to a SINT_TO_FP when the sign bit is known zero. Perform
12489 // the optimization here.
12490 if (DAG.SignBitIsZero(N0))
12491 return DAG.getNode(ISD::SINT_TO_FP, dl, Op.getValueType(), N0);
12493 MVT SrcVT = N0.getSimpleValueType();
12494 MVT DstVT = Op.getSimpleValueType();
12495 if (SrcVT == MVT::i64 && DstVT == MVT::f64 && X86ScalarSSEf64)
12496 return LowerUINT_TO_FP_i64(Op, DAG);
12497 if (SrcVT == MVT::i32 && X86ScalarSSEf64)
12498 return LowerUINT_TO_FP_i32(Op, DAG);
12499 if (Subtarget->is64Bit() && SrcVT == MVT::i64 && DstVT == MVT::f32)
12502 // Make a 64-bit buffer, and use it to build an FILD.
12503 SDValue StackSlot = DAG.CreateStackTemporary(MVT::i64);
12504 if (SrcVT == MVT::i32) {
12505 SDValue WordOff = DAG.getConstant(4, dl, PtrVT);
12506 SDValue OffsetSlot = DAG.getNode(ISD::ADD, dl, PtrVT, StackSlot, WordOff);
12507 SDValue Store1 = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
12508 StackSlot, MachinePointerInfo(),
12510 SDValue Store2 = DAG.getStore(Store1, dl, DAG.getConstant(0, dl, MVT::i32),
12511 OffsetSlot, MachinePointerInfo(),
12513 SDValue Fild = BuildFILD(Op, MVT::i64, Store2, StackSlot, DAG);
12517 assert(SrcVT == MVT::i64 && "Unexpected type in UINT_TO_FP");
12518 SDValue Store = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
12519 StackSlot, MachinePointerInfo(),
12521 // For i64 source, we need to add the appropriate power of 2 if the input
12522 // was negative. This is the same as the optimization in
12523 // DAGTypeLegalizer::ExpandIntOp_UNIT_TO_FP, and for it to be safe here,
12524 // we must be careful to do the computation in x87 extended precision, not
12525 // in SSE. (The generic code can't know it's OK to do this, or how to.)
12526 int SSFI = cast<FrameIndexSDNode>(StackSlot)->getIndex();
12527 MachineMemOperand *MMO = DAG.getMachineFunction().getMachineMemOperand(
12528 MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), SSFI),
12529 MachineMemOperand::MOLoad, 8, 8);
12531 SDVTList Tys = DAG.getVTList(MVT::f80, MVT::Other);
12532 SDValue Ops[] = { Store, StackSlot, DAG.getValueType(MVT::i64) };
12533 SDValue Fild = DAG.getMemIntrinsicNode(X86ISD::FILD, dl, Tys, Ops,
12536 APInt FF(32, 0x5F800000ULL);
12538 // Check whether the sign bit is set.
12539 SDValue SignSet = DAG.getSetCC(
12540 dl, getSetCCResultType(DAG.getDataLayout(), *DAG.getContext(), MVT::i64),
12541 Op.getOperand(0), DAG.getConstant(0, dl, MVT::i64), ISD::SETLT);
12543 // Build a 64 bit pair (0, FF) in the constant pool, with FF in the lo bits.
12544 SDValue FudgePtr = DAG.getConstantPool(
12545 ConstantInt::get(*DAG.getContext(), FF.zext(64)), PtrVT);
12547 // Get a pointer to FF if the sign bit was set, or to 0 otherwise.
12548 SDValue Zero = DAG.getIntPtrConstant(0, dl);
12549 SDValue Four = DAG.getIntPtrConstant(4, dl);
12550 SDValue Offset = DAG.getNode(ISD::SELECT, dl, Zero.getValueType(), SignSet,
12552 FudgePtr = DAG.getNode(ISD::ADD, dl, PtrVT, FudgePtr, Offset);
12554 // Load the value out, extending it from f32 to f80.
12555 // FIXME: Avoid the extend by constructing the right constant pool?
12556 SDValue Fudge = DAG.getExtLoad(
12557 ISD::EXTLOAD, dl, MVT::f80, DAG.getEntryNode(), FudgePtr,
12558 MachinePointerInfo::getConstantPool(DAG.getMachineFunction()), MVT::f32,
12559 false, false, false, 4);
12560 // Extend everything to 80 bits to force it to be done on x87.
12561 // TODO: Are there any fast-math-flags to propagate here?
12562 SDValue Add = DAG.getNode(ISD::FADD, dl, MVT::f80, Fild, Fudge);
12563 return DAG.getNode(ISD::FP_ROUND, dl, DstVT, Add,
12564 DAG.getIntPtrConstant(0, dl));
12567 // If the given FP_TO_SINT (IsSigned) or FP_TO_UINT (!IsSigned) operation
12568 // is legal, or has an f16 source (which needs to be promoted to f32),
12569 // just return an <SDValue(), SDValue()> pair.
12570 // Otherwise it is assumed to be a conversion from one of f32, f64 or f80
12571 // to i16, i32 or i64, and we lower it to a legal sequence.
12572 // If lowered to the final integer result we return a <result, SDValue()> pair.
12573 // Otherwise we lower it to a sequence ending with a FIST, return a
12574 // <FIST, StackSlot> pair, and the caller is responsible for loading
12575 // the final integer result from StackSlot.
12576 std::pair<SDValue,SDValue>
12577 X86TargetLowering::FP_TO_INTHelper(SDValue Op, SelectionDAG &DAG,
12578 bool IsSigned, bool IsReplace) const {
12581 EVT DstTy = Op.getValueType();
12582 EVT TheVT = Op.getOperand(0).getValueType();
12583 auto PtrVT = getPointerTy(DAG.getDataLayout());
12585 if (TheVT == MVT::f16)
12586 // We need to promote the f16 to f32 before using the lowering
12587 // in this routine.
12588 return std::make_pair(SDValue(), SDValue());
12590 assert((TheVT == MVT::f32 ||
12591 TheVT == MVT::f64 ||
12592 TheVT == MVT::f80) &&
12593 "Unexpected FP operand type in FP_TO_INTHelper");
12595 // If using FIST to compute an unsigned i64, we'll need some fixup
12596 // to handle values above the maximum signed i64. A FIST is always
12597 // used for the 32-bit subtarget, but also for f80 on a 64-bit target.
12598 bool UnsignedFixup = !IsSigned &&
12599 DstTy == MVT::i64 &&
12600 (!Subtarget->is64Bit() ||
12601 !isScalarFPTypeInSSEReg(TheVT));
12603 if (!IsSigned && DstTy != MVT::i64 && !Subtarget->hasAVX512()) {
12604 // Replace the fp-to-uint32 operation with an fp-to-sint64 FIST.
12605 // The low 32 bits of the fist result will have the correct uint32 result.
12606 assert(DstTy == MVT::i32 && "Unexpected FP_TO_UINT");
12610 assert(DstTy.getSimpleVT() <= MVT::i64 &&
12611 DstTy.getSimpleVT() >= MVT::i16 &&
12612 "Unknown FP_TO_INT to lower!");
12614 // These are really Legal.
12615 if (DstTy == MVT::i32 &&
12616 isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType()))
12617 return std::make_pair(SDValue(), SDValue());
12618 if (Subtarget->is64Bit() &&
12619 DstTy == MVT::i64 &&
12620 isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType()))
12621 return std::make_pair(SDValue(), SDValue());
12623 // We lower FP->int64 into FISTP64 followed by a load from a temporary
12625 MachineFunction &MF = DAG.getMachineFunction();
12626 unsigned MemSize = DstTy.getSizeInBits()/8;
12627 int SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize, false);
12628 SDValue StackSlot = DAG.getFrameIndex(SSFI, PtrVT);
12631 switch (DstTy.getSimpleVT().SimpleTy) {
12632 default: llvm_unreachable("Invalid FP_TO_SINT to lower!");
12633 case MVT::i16: Opc = X86ISD::FP_TO_INT16_IN_MEM; break;
12634 case MVT::i32: Opc = X86ISD::FP_TO_INT32_IN_MEM; break;
12635 case MVT::i64: Opc = X86ISD::FP_TO_INT64_IN_MEM; break;
12638 SDValue Chain = DAG.getEntryNode();
12639 SDValue Value = Op.getOperand(0);
12640 SDValue Adjust; // 0x0 or 0x80000000, for result sign bit adjustment.
12642 if (UnsignedFixup) {
12644 // Conversion to unsigned i64 is implemented with a select,
12645 // depending on whether the source value fits in the range
12646 // of a signed i64. Let Thresh be the FP equivalent of
12647 // 0x8000000000000000ULL.
12649 // Adjust i32 = (Value < Thresh) ? 0 : 0x80000000;
12650 // FistSrc = (Value < Thresh) ? Value : (Value - Thresh);
12651 // Fist-to-mem64 FistSrc
12652 // Add 0 or 0x800...0ULL to the 64-bit result, which is equivalent
12653 // to XOR'ing the high 32 bits with Adjust.
12655 // Being a power of 2, Thresh is exactly representable in all FP formats.
12656 // For X87 we'd like to use the smallest FP type for this constant, but
12657 // for DAG type consistency we have to match the FP operand type.
12659 APFloat Thresh(APFloat::IEEEsingle, APInt(32, 0x5f000000));
12660 APFloat::opStatus Status = APFloat::opOK;
12661 bool LosesInfo = false;
12662 if (TheVT == MVT::f64)
12663 // The rounding mode is irrelevant as the conversion should be exact.
12664 Status = Thresh.convert(APFloat::IEEEdouble, APFloat::rmNearestTiesToEven,
12666 else if (TheVT == MVT::f80)
12667 Status = Thresh.convert(APFloat::x87DoubleExtended,
12668 APFloat::rmNearestTiesToEven, &LosesInfo);
12670 assert(Status == APFloat::opOK && !LosesInfo &&
12671 "FP conversion should have been exact");
12673 SDValue ThreshVal = DAG.getConstantFP(Thresh, DL, TheVT);
12675 SDValue Cmp = DAG.getSetCC(DL,
12676 getSetCCResultType(DAG.getDataLayout(),
12677 *DAG.getContext(), TheVT),
12678 Value, ThreshVal, ISD::SETLT);
12679 Adjust = DAG.getSelect(DL, MVT::i32, Cmp,
12680 DAG.getConstant(0, DL, MVT::i32),
12681 DAG.getConstant(0x80000000, DL, MVT::i32));
12682 SDValue Sub = DAG.getNode(ISD::FSUB, DL, TheVT, Value, ThreshVal);
12683 Cmp = DAG.getSetCC(DL, getSetCCResultType(DAG.getDataLayout(),
12684 *DAG.getContext(), TheVT),
12685 Value, ThreshVal, ISD::SETLT);
12686 Value = DAG.getSelect(DL, TheVT, Cmp, Value, Sub);
12689 // FIXME This causes a redundant load/store if the SSE-class value is already
12690 // in memory, such as if it is on the callstack.
12691 if (isScalarFPTypeInSSEReg(TheVT)) {
12692 assert(DstTy == MVT::i64 && "Invalid FP_TO_SINT to lower!");
12693 Chain = DAG.getStore(Chain, DL, Value, StackSlot,
12694 MachinePointerInfo::getFixedStack(MF, SSFI), false,
12696 SDVTList Tys = DAG.getVTList(Op.getOperand(0).getValueType(), MVT::Other);
12698 Chain, StackSlot, DAG.getValueType(TheVT)
12701 MachineMemOperand *MMO =
12702 MF.getMachineMemOperand(MachinePointerInfo::getFixedStack(MF, SSFI),
12703 MachineMemOperand::MOLoad, MemSize, MemSize);
12704 Value = DAG.getMemIntrinsicNode(X86ISD::FLD, DL, Tys, Ops, DstTy, MMO);
12705 Chain = Value.getValue(1);
12706 SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize, false);
12707 StackSlot = DAG.getFrameIndex(SSFI, PtrVT);
12710 MachineMemOperand *MMO =
12711 MF.getMachineMemOperand(MachinePointerInfo::getFixedStack(MF, SSFI),
12712 MachineMemOperand::MOStore, MemSize, MemSize);
12714 if (UnsignedFixup) {
12716 // Insert the FIST, load its result as two i32's,
12717 // and XOR the high i32 with Adjust.
12719 SDValue FistOps[] = { Chain, Value, StackSlot };
12720 SDValue FIST = DAG.getMemIntrinsicNode(Opc, DL, DAG.getVTList(MVT::Other),
12721 FistOps, DstTy, MMO);
12723 SDValue Low32 = DAG.getLoad(MVT::i32, DL, FIST, StackSlot,
12724 MachinePointerInfo(),
12725 false, false, false, 0);
12726 SDValue HighAddr = DAG.getNode(ISD::ADD, DL, PtrVT, StackSlot,
12727 DAG.getConstant(4, DL, PtrVT));
12729 SDValue High32 = DAG.getLoad(MVT::i32, DL, FIST, HighAddr,
12730 MachinePointerInfo(),
12731 false, false, false, 0);
12732 High32 = DAG.getNode(ISD::XOR, DL, MVT::i32, High32, Adjust);
12734 if (Subtarget->is64Bit()) {
12735 // Join High32 and Low32 into a 64-bit result.
12736 // (High32 << 32) | Low32
12737 Low32 = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::i64, Low32);
12738 High32 = DAG.getNode(ISD::ANY_EXTEND, DL, MVT::i64, High32);
12739 High32 = DAG.getNode(ISD::SHL, DL, MVT::i64, High32,
12740 DAG.getConstant(32, DL, MVT::i8));
12741 SDValue Result = DAG.getNode(ISD::OR, DL, MVT::i64, High32, Low32);
12742 return std::make_pair(Result, SDValue());
12745 SDValue ResultOps[] = { Low32, High32 };
12747 SDValue pair = IsReplace
12748 ? DAG.getNode(ISD::BUILD_PAIR, DL, MVT::i64, ResultOps)
12749 : DAG.getMergeValues(ResultOps, DL);
12750 return std::make_pair(pair, SDValue());
12752 // Build the FP_TO_INT*_IN_MEM
12753 SDValue Ops[] = { Chain, Value, StackSlot };
12754 SDValue FIST = DAG.getMemIntrinsicNode(Opc, DL, DAG.getVTList(MVT::Other),
12756 return std::make_pair(FIST, StackSlot);
12760 static SDValue LowerAVXExtend(SDValue Op, SelectionDAG &DAG,
12761 const X86Subtarget *Subtarget) {
12762 MVT VT = Op->getSimpleValueType(0);
12763 SDValue In = Op->getOperand(0);
12764 MVT InVT = In.getSimpleValueType();
12767 if (VT.is512BitVector() || InVT.getScalarType() == MVT::i1)
12768 return DAG.getNode(ISD::ZERO_EXTEND, dl, VT, In);
12770 // Optimize vectors in AVX mode:
12773 // Use vpunpcklwd for 4 lower elements v8i16 -> v4i32.
12774 // Use vpunpckhwd for 4 upper elements v8i16 -> v4i32.
12775 // Concat upper and lower parts.
12778 // Use vpunpckldq for 4 lower elements v4i32 -> v2i64.
12779 // Use vpunpckhdq for 4 upper elements v4i32 -> v2i64.
12780 // Concat upper and lower parts.
12783 if (((VT != MVT::v16i16) || (InVT != MVT::v16i8)) &&
12784 ((VT != MVT::v8i32) || (InVT != MVT::v8i16)) &&
12785 ((VT != MVT::v4i64) || (InVT != MVT::v4i32)))
12788 if (Subtarget->hasInt256())
12789 return DAG.getNode(X86ISD::VZEXT, dl, VT, In);
12791 SDValue ZeroVec = getZeroVector(InVT, Subtarget, DAG, dl);
12792 SDValue Undef = DAG.getUNDEF(InVT);
12793 bool NeedZero = Op.getOpcode() == ISD::ZERO_EXTEND;
12794 SDValue OpLo = getUnpackl(DAG, dl, InVT, In, NeedZero ? ZeroVec : Undef);
12795 SDValue OpHi = getUnpackh(DAG, dl, InVT, In, NeedZero ? ZeroVec : Undef);
12797 MVT HVT = MVT::getVectorVT(VT.getVectorElementType(),
12798 VT.getVectorNumElements()/2);
12800 OpLo = DAG.getBitcast(HVT, OpLo);
12801 OpHi = DAG.getBitcast(HVT, OpHi);
12803 return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT, OpLo, OpHi);
12806 static SDValue LowerZERO_EXTEND_AVX512(SDValue Op,
12807 const X86Subtarget *Subtarget, SelectionDAG &DAG) {
12808 MVT VT = Op->getSimpleValueType(0);
12809 SDValue In = Op->getOperand(0);
12810 MVT InVT = In.getSimpleValueType();
12812 unsigned int NumElts = VT.getVectorNumElements();
12813 if (NumElts != 8 && NumElts != 16 && !Subtarget->hasBWI())
12816 if (VT.is512BitVector() && InVT.getVectorElementType() != MVT::i1)
12817 return DAG.getNode(X86ISD::VZEXT, DL, VT, In);
12819 assert(InVT.getVectorElementType() == MVT::i1);
12820 MVT ExtVT = NumElts == 8 ? MVT::v8i64 : MVT::v16i32;
12822 DAG.getConstant(APInt(ExtVT.getScalarSizeInBits(), 1), DL, ExtVT);
12824 DAG.getConstant(APInt::getNullValue(ExtVT.getScalarSizeInBits()), DL, ExtVT);
12826 SDValue V = DAG.getNode(ISD::VSELECT, DL, ExtVT, In, One, Zero);
12827 if (VT.is512BitVector())
12829 return DAG.getNode(X86ISD::VTRUNC, DL, VT, V);
12832 static SDValue LowerANY_EXTEND(SDValue Op, const X86Subtarget *Subtarget,
12833 SelectionDAG &DAG) {
12834 if (Subtarget->hasFp256())
12835 if (SDValue Res = LowerAVXExtend(Op, DAG, Subtarget))
12841 static SDValue LowerZERO_EXTEND(SDValue Op, const X86Subtarget *Subtarget,
12842 SelectionDAG &DAG) {
12844 MVT VT = Op.getSimpleValueType();
12845 SDValue In = Op.getOperand(0);
12846 MVT SVT = In.getSimpleValueType();
12848 if (VT.is512BitVector() || SVT.getVectorElementType() == MVT::i1)
12849 return LowerZERO_EXTEND_AVX512(Op, Subtarget, DAG);
12851 if (Subtarget->hasFp256())
12852 if (SDValue Res = LowerAVXExtend(Op, DAG, Subtarget))
12855 assert(!VT.is256BitVector() || !SVT.is128BitVector() ||
12856 VT.getVectorNumElements() != SVT.getVectorNumElements());
12860 SDValue X86TargetLowering::LowerTRUNCATE(SDValue Op, SelectionDAG &DAG) const {
12862 MVT VT = Op.getSimpleValueType();
12863 SDValue In = Op.getOperand(0);
12864 MVT InVT = In.getSimpleValueType();
12866 if (VT == MVT::i1) {
12867 assert((InVT.isInteger() && (InVT.getSizeInBits() <= 64)) &&
12868 "Invalid scalar TRUNCATE operation");
12869 if (InVT.getSizeInBits() >= 32)
12871 In = DAG.getNode(ISD::ANY_EXTEND, DL, MVT::i32, In);
12872 return DAG.getNode(ISD::TRUNCATE, DL, VT, In);
12874 assert(VT.getVectorNumElements() == InVT.getVectorNumElements() &&
12875 "Invalid TRUNCATE operation");
12877 // move vector to mask - truncate solution for SKX
12878 if (VT.getVectorElementType() == MVT::i1) {
12879 if (InVT.is512BitVector() && InVT.getScalarSizeInBits() <= 16 &&
12880 Subtarget->hasBWI())
12881 return Op; // legal, will go to VPMOVB2M, VPMOVW2M
12882 if ((InVT.is256BitVector() || InVT.is128BitVector())
12883 && InVT.getScalarSizeInBits() <= 16 &&
12884 Subtarget->hasBWI() && Subtarget->hasVLX())
12885 return Op; // legal, will go to VPMOVB2M, VPMOVW2M
12886 if (InVT.is512BitVector() && InVT.getScalarSizeInBits() >= 32 &&
12887 Subtarget->hasDQI())
12888 return Op; // legal, will go to VPMOVD2M, VPMOVQ2M
12889 if ((InVT.is256BitVector() || InVT.is128BitVector())
12890 && InVT.getScalarSizeInBits() >= 32 &&
12891 Subtarget->hasDQI() && Subtarget->hasVLX())
12892 return Op; // legal, will go to VPMOVB2M, VPMOVQ2M
12895 if (VT.getVectorElementType() == MVT::i1) {
12896 assert(VT.getVectorElementType() == MVT::i1 && "Unexpected vector type");
12897 unsigned NumElts = InVT.getVectorNumElements();
12898 assert ((NumElts == 8 || NumElts == 16) && "Unexpected vector type");
12899 if (InVT.getSizeInBits() < 512) {
12900 MVT ExtVT = (NumElts == 16)? MVT::v16i32 : MVT::v8i64;
12901 In = DAG.getNode(ISD::SIGN_EXTEND, DL, ExtVT, In);
12906 DAG.getConstant(APInt::getSignBit(InVT.getScalarSizeInBits()), DL, InVT);
12907 SDValue And = DAG.getNode(ISD::AND, DL, InVT, OneV, In);
12908 return DAG.getNode(X86ISD::TESTM, DL, VT, And, And);
12911 // vpmovqb/w/d, vpmovdb/w, vpmovwb
12912 if (((!InVT.is512BitVector() && Subtarget->hasVLX()) || InVT.is512BitVector()) &&
12913 (InVT.getVectorElementType() != MVT::i16 || Subtarget->hasBWI()))
12914 return DAG.getNode(X86ISD::VTRUNC, DL, VT, In);
12916 if ((VT == MVT::v4i32) && (InVT == MVT::v4i64)) {
12917 // On AVX2, v4i64 -> v4i32 becomes VPERMD.
12918 if (Subtarget->hasInt256()) {
12919 static const int ShufMask[] = {0, 2, 4, 6, -1, -1, -1, -1};
12920 In = DAG.getBitcast(MVT::v8i32, In);
12921 In = DAG.getVectorShuffle(MVT::v8i32, DL, In, DAG.getUNDEF(MVT::v8i32),
12923 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, VT, In,
12924 DAG.getIntPtrConstant(0, DL));
12927 SDValue OpLo = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v2i64, In,
12928 DAG.getIntPtrConstant(0, DL));
12929 SDValue OpHi = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v2i64, In,
12930 DAG.getIntPtrConstant(2, DL));
12931 OpLo = DAG.getBitcast(MVT::v4i32, OpLo);
12932 OpHi = DAG.getBitcast(MVT::v4i32, OpHi);
12933 static const int ShufMask[] = {0, 2, 4, 6};
12934 return DAG.getVectorShuffle(VT, DL, OpLo, OpHi, ShufMask);
12937 if ((VT == MVT::v8i16) && (InVT == MVT::v8i32)) {
12938 // On AVX2, v8i32 -> v8i16 becomed PSHUFB.
12939 if (Subtarget->hasInt256()) {
12940 In = DAG.getBitcast(MVT::v32i8, In);
12942 SmallVector<SDValue,32> pshufbMask;
12943 for (unsigned i = 0; i < 2; ++i) {
12944 pshufbMask.push_back(DAG.getConstant(0x0, DL, MVT::i8));
12945 pshufbMask.push_back(DAG.getConstant(0x1, DL, MVT::i8));
12946 pshufbMask.push_back(DAG.getConstant(0x4, DL, MVT::i8));
12947 pshufbMask.push_back(DAG.getConstant(0x5, DL, MVT::i8));
12948 pshufbMask.push_back(DAG.getConstant(0x8, DL, MVT::i8));
12949 pshufbMask.push_back(DAG.getConstant(0x9, DL, MVT::i8));
12950 pshufbMask.push_back(DAG.getConstant(0xc, DL, MVT::i8));
12951 pshufbMask.push_back(DAG.getConstant(0xd, DL, MVT::i8));
12952 for (unsigned j = 0; j < 8; ++j)
12953 pshufbMask.push_back(DAG.getConstant(0x80, DL, MVT::i8));
12955 SDValue BV = DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v32i8, pshufbMask);
12956 In = DAG.getNode(X86ISD::PSHUFB, DL, MVT::v32i8, In, BV);
12957 In = DAG.getBitcast(MVT::v4i64, In);
12959 static const int ShufMask[] = {0, 2, -1, -1};
12960 In = DAG.getVectorShuffle(MVT::v4i64, DL, In, DAG.getUNDEF(MVT::v4i64),
12962 In = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v2i64, In,
12963 DAG.getIntPtrConstant(0, DL));
12964 return DAG.getBitcast(VT, In);
12967 SDValue OpLo = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v4i32, In,
12968 DAG.getIntPtrConstant(0, DL));
12970 SDValue OpHi = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v4i32, In,
12971 DAG.getIntPtrConstant(4, DL));
12973 OpLo = DAG.getBitcast(MVT::v16i8, OpLo);
12974 OpHi = DAG.getBitcast(MVT::v16i8, OpHi);
12976 // The PSHUFB mask:
12977 static const int ShufMask1[] = {0, 1, 4, 5, 8, 9, 12, 13,
12978 -1, -1, -1, -1, -1, -1, -1, -1};
12980 SDValue Undef = DAG.getUNDEF(MVT::v16i8);
12981 OpLo = DAG.getVectorShuffle(MVT::v16i8, DL, OpLo, Undef, ShufMask1);
12982 OpHi = DAG.getVectorShuffle(MVT::v16i8, DL, OpHi, Undef, ShufMask1);
12984 OpLo = DAG.getBitcast(MVT::v4i32, OpLo);
12985 OpHi = DAG.getBitcast(MVT::v4i32, OpHi);
12987 // The MOVLHPS Mask:
12988 static const int ShufMask2[] = {0, 1, 4, 5};
12989 SDValue res = DAG.getVectorShuffle(MVT::v4i32, DL, OpLo, OpHi, ShufMask2);
12990 return DAG.getBitcast(MVT::v8i16, res);
12993 // Handle truncation of V256 to V128 using shuffles.
12994 if (!VT.is128BitVector() || !InVT.is256BitVector())
12997 assert(Subtarget->hasFp256() && "256-bit vector without AVX!");
12999 unsigned NumElems = VT.getVectorNumElements();
13000 MVT NVT = MVT::getVectorVT(VT.getVectorElementType(), NumElems * 2);
13002 SmallVector<int, 16> MaskVec(NumElems * 2, -1);
13003 // Prepare truncation shuffle mask
13004 for (unsigned i = 0; i != NumElems; ++i)
13005 MaskVec[i] = i * 2;
13006 SDValue V = DAG.getVectorShuffle(NVT, DL, DAG.getBitcast(NVT, In),
13007 DAG.getUNDEF(NVT), &MaskVec[0]);
13008 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, VT, V,
13009 DAG.getIntPtrConstant(0, DL));
13012 SDValue X86TargetLowering::LowerFP_TO_SINT(SDValue Op,
13013 SelectionDAG &DAG) const {
13014 assert(!Op.getSimpleValueType().isVector());
13016 std::pair<SDValue,SDValue> Vals = FP_TO_INTHelper(Op, DAG,
13017 /*IsSigned=*/ true, /*IsReplace=*/ false);
13018 SDValue FIST = Vals.first, StackSlot = Vals.second;
13019 // If FP_TO_INTHelper failed, the node is actually supposed to be Legal.
13020 if (!FIST.getNode())
13023 if (StackSlot.getNode())
13024 // Load the result.
13025 return DAG.getLoad(Op.getValueType(), SDLoc(Op),
13026 FIST, StackSlot, MachinePointerInfo(),
13027 false, false, false, 0);
13029 // The node is the result.
13033 SDValue X86TargetLowering::LowerFP_TO_UINT(SDValue Op,
13034 SelectionDAG &DAG) const {
13035 std::pair<SDValue,SDValue> Vals = FP_TO_INTHelper(Op, DAG,
13036 /*IsSigned=*/ false, /*IsReplace=*/ false);
13037 SDValue FIST = Vals.first, StackSlot = Vals.second;
13038 // If FP_TO_INTHelper failed, the node is actually supposed to be Legal.
13039 if (!FIST.getNode())
13042 if (StackSlot.getNode())
13043 // Load the result.
13044 return DAG.getLoad(Op.getValueType(), SDLoc(Op),
13045 FIST, StackSlot, MachinePointerInfo(),
13046 false, false, false, 0);
13048 // The node is the result.
13052 static SDValue LowerFP_EXTEND(SDValue Op, SelectionDAG &DAG) {
13054 MVT VT = Op.getSimpleValueType();
13055 SDValue In = Op.getOperand(0);
13056 MVT SVT = In.getSimpleValueType();
13058 assert(SVT == MVT::v2f32 && "Only customize MVT::v2f32 type legalization!");
13060 return DAG.getNode(X86ISD::VFPEXT, DL, VT,
13061 DAG.getNode(ISD::CONCAT_VECTORS, DL, MVT::v4f32,
13062 In, DAG.getUNDEF(SVT)));
13065 /// The only differences between FABS and FNEG are the mask and the logic op.
13066 /// FNEG also has a folding opportunity for FNEG(FABS(x)).
13067 static SDValue LowerFABSorFNEG(SDValue Op, SelectionDAG &DAG) {
13068 assert((Op.getOpcode() == ISD::FABS || Op.getOpcode() == ISD::FNEG) &&
13069 "Wrong opcode for lowering FABS or FNEG.");
13071 bool IsFABS = (Op.getOpcode() == ISD::FABS);
13073 // If this is a FABS and it has an FNEG user, bail out to fold the combination
13074 // into an FNABS. We'll lower the FABS after that if it is still in use.
13076 for (SDNode *User : Op->uses())
13077 if (User->getOpcode() == ISD::FNEG)
13081 MVT VT = Op.getSimpleValueType();
13083 // FIXME: Use function attribute "OptimizeForSize" and/or CodeGenOpt::Level to
13084 // decide if we should generate a 16-byte constant mask when we only need 4 or
13085 // 8 bytes for the scalar case.
13091 if (VT.isVector()) {
13093 EltVT = VT.getVectorElementType();
13094 NumElts = VT.getVectorNumElements();
13096 // There are no scalar bitwise logical SSE/AVX instructions, so we
13097 // generate a 16-byte vector constant and logic op even for the scalar case.
13098 // Using a 16-byte mask allows folding the load of the mask with
13099 // the logic op, so it can save (~4 bytes) on code size.
13100 LogicVT = (VT == MVT::f64) ? MVT::v2f64 : MVT::v4f32;
13102 NumElts = (VT == MVT::f64) ? 2 : 4;
13105 unsigned EltBits = EltVT.getSizeInBits();
13106 LLVMContext *Context = DAG.getContext();
13107 // For FABS, mask is 0x7f...; for FNEG, mask is 0x80...
13109 IsFABS ? APInt::getSignedMaxValue(EltBits) : APInt::getSignBit(EltBits);
13110 Constant *C = ConstantInt::get(*Context, MaskElt);
13111 C = ConstantVector::getSplat(NumElts, C);
13112 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
13113 SDValue CPIdx = DAG.getConstantPool(C, TLI.getPointerTy(DAG.getDataLayout()));
13114 unsigned Alignment = cast<ConstantPoolSDNode>(CPIdx)->getAlignment();
13116 DAG.getLoad(LogicVT, dl, DAG.getEntryNode(), CPIdx,
13117 MachinePointerInfo::getConstantPool(DAG.getMachineFunction()),
13118 false, false, false, Alignment);
13120 SDValue Op0 = Op.getOperand(0);
13121 bool IsFNABS = !IsFABS && (Op0.getOpcode() == ISD::FABS);
13123 IsFABS ? X86ISD::FAND : IsFNABS ? X86ISD::FOR : X86ISD::FXOR;
13124 SDValue Operand = IsFNABS ? Op0.getOperand(0) : Op0;
13127 return DAG.getNode(LogicOp, dl, LogicVT, Operand, Mask);
13129 // For the scalar case extend to a 128-bit vector, perform the logic op,
13130 // and extract the scalar result back out.
13131 Operand = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, LogicVT, Operand);
13132 SDValue LogicNode = DAG.getNode(LogicOp, dl, LogicVT, Operand, Mask);
13133 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, LogicNode,
13134 DAG.getIntPtrConstant(0, dl));
13137 static SDValue LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) {
13138 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
13139 LLVMContext *Context = DAG.getContext();
13140 SDValue Op0 = Op.getOperand(0);
13141 SDValue Op1 = Op.getOperand(1);
13143 MVT VT = Op.getSimpleValueType();
13144 MVT SrcVT = Op1.getSimpleValueType();
13146 // If second operand is smaller, extend it first.
13147 if (SrcVT.bitsLT(VT)) {
13148 Op1 = DAG.getNode(ISD::FP_EXTEND, dl, VT, Op1);
13151 // And if it is bigger, shrink it first.
13152 if (SrcVT.bitsGT(VT)) {
13153 Op1 = DAG.getNode(ISD::FP_ROUND, dl, VT, Op1, DAG.getIntPtrConstant(1, dl));
13157 // At this point the operands and the result should have the same
13158 // type, and that won't be f80 since that is not custom lowered.
13160 const fltSemantics &Sem =
13161 VT == MVT::f64 ? APFloat::IEEEdouble : APFloat::IEEEsingle;
13162 const unsigned SizeInBits = VT.getSizeInBits();
13164 SmallVector<Constant *, 4> CV(
13165 VT == MVT::f64 ? 2 : 4,
13166 ConstantFP::get(*Context, APFloat(Sem, APInt(SizeInBits, 0))));
13168 // First, clear all bits but the sign bit from the second operand (sign).
13169 CV[0] = ConstantFP::get(*Context,
13170 APFloat(Sem, APInt::getHighBitsSet(SizeInBits, 1)));
13171 Constant *C = ConstantVector::get(CV);
13172 auto PtrVT = TLI.getPointerTy(DAG.getDataLayout());
13173 SDValue CPIdx = DAG.getConstantPool(C, PtrVT, 16);
13175 // Perform all logic operations as 16-byte vectors because there are no
13176 // scalar FP logic instructions in SSE. This allows load folding of the
13177 // constants into the logic instructions.
13178 MVT LogicVT = (VT == MVT::f64) ? MVT::v2f64 : MVT::v4f32;
13180 DAG.getLoad(LogicVT, dl, DAG.getEntryNode(), CPIdx,
13181 MachinePointerInfo::getConstantPool(DAG.getMachineFunction()),
13182 false, false, false, 16);
13183 Op1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, LogicVT, Op1);
13184 SDValue SignBit = DAG.getNode(X86ISD::FAND, dl, LogicVT, Op1, Mask1);
13186 // Next, clear the sign bit from the first operand (magnitude).
13187 // If it's a constant, we can clear it here.
13188 if (ConstantFPSDNode *Op0CN = dyn_cast<ConstantFPSDNode>(Op0)) {
13189 APFloat APF = Op0CN->getValueAPF();
13190 // If the magnitude is a positive zero, the sign bit alone is enough.
13191 if (APF.isPosZero())
13192 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, SrcVT, SignBit,
13193 DAG.getIntPtrConstant(0, dl));
13195 CV[0] = ConstantFP::get(*Context, APF);
13197 CV[0] = ConstantFP::get(
13199 APFloat(Sem, APInt::getLowBitsSet(SizeInBits, SizeInBits - 1)));
13201 C = ConstantVector::get(CV);
13202 CPIdx = DAG.getConstantPool(C, PtrVT, 16);
13204 DAG.getLoad(LogicVT, dl, DAG.getEntryNode(), CPIdx,
13205 MachinePointerInfo::getConstantPool(DAG.getMachineFunction()),
13206 false, false, false, 16);
13207 // If the magnitude operand wasn't a constant, we need to AND out the sign.
13208 if (!isa<ConstantFPSDNode>(Op0)) {
13209 Op0 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, LogicVT, Op0);
13210 Val = DAG.getNode(X86ISD::FAND, dl, LogicVT, Op0, Val);
13212 // OR the magnitude value with the sign bit.
13213 Val = DAG.getNode(X86ISD::FOR, dl, LogicVT, Val, SignBit);
13214 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, SrcVT, Val,
13215 DAG.getIntPtrConstant(0, dl));
13218 static SDValue LowerFGETSIGN(SDValue Op, SelectionDAG &DAG) {
13219 SDValue N0 = Op.getOperand(0);
13221 MVT VT = Op.getSimpleValueType();
13223 // Lower ISD::FGETSIGN to (AND (X86ISD::FGETSIGNx86 ...) 1).
13224 SDValue xFGETSIGN = DAG.getNode(X86ISD::FGETSIGNx86, dl, VT, N0,
13225 DAG.getConstant(1, dl, VT));
13226 return DAG.getNode(ISD::AND, dl, VT, xFGETSIGN, DAG.getConstant(1, dl, VT));
13229 // Check whether an OR'd tree is PTEST-able.
13230 static SDValue LowerVectorAllZeroTest(SDValue Op, const X86Subtarget *Subtarget,
13231 SelectionDAG &DAG) {
13232 assert(Op.getOpcode() == ISD::OR && "Only check OR'd tree.");
13234 if (!Subtarget->hasSSE41())
13237 if (!Op->hasOneUse())
13240 SDNode *N = Op.getNode();
13243 SmallVector<SDValue, 8> Opnds;
13244 DenseMap<SDValue, unsigned> VecInMap;
13245 SmallVector<SDValue, 8> VecIns;
13246 EVT VT = MVT::Other;
13248 // Recognize a special case where a vector is casted into wide integer to
13250 Opnds.push_back(N->getOperand(0));
13251 Opnds.push_back(N->getOperand(1));
13253 for (unsigned Slot = 0, e = Opnds.size(); Slot < e; ++Slot) {
13254 SmallVectorImpl<SDValue>::const_iterator I = Opnds.begin() + Slot;
13255 // BFS traverse all OR'd operands.
13256 if (I->getOpcode() == ISD::OR) {
13257 Opnds.push_back(I->getOperand(0));
13258 Opnds.push_back(I->getOperand(1));
13259 // Re-evaluate the number of nodes to be traversed.
13260 e += 2; // 2 more nodes (LHS and RHS) are pushed.
13264 // Quit if a non-EXTRACT_VECTOR_ELT
13265 if (I->getOpcode() != ISD::EXTRACT_VECTOR_ELT)
13268 // Quit if without a constant index.
13269 SDValue Idx = I->getOperand(1);
13270 if (!isa<ConstantSDNode>(Idx))
13273 SDValue ExtractedFromVec = I->getOperand(0);
13274 DenseMap<SDValue, unsigned>::iterator M = VecInMap.find(ExtractedFromVec);
13275 if (M == VecInMap.end()) {
13276 VT = ExtractedFromVec.getValueType();
13277 // Quit if not 128/256-bit vector.
13278 if (!VT.is128BitVector() && !VT.is256BitVector())
13280 // Quit if not the same type.
13281 if (VecInMap.begin() != VecInMap.end() &&
13282 VT != VecInMap.begin()->first.getValueType())
13284 M = VecInMap.insert(std::make_pair(ExtractedFromVec, 0)).first;
13285 VecIns.push_back(ExtractedFromVec);
13287 M->second |= 1U << cast<ConstantSDNode>(Idx)->getZExtValue();
13290 assert((VT.is128BitVector() || VT.is256BitVector()) &&
13291 "Not extracted from 128-/256-bit vector.");
13293 unsigned FullMask = (1U << VT.getVectorNumElements()) - 1U;
13295 for (DenseMap<SDValue, unsigned>::const_iterator
13296 I = VecInMap.begin(), E = VecInMap.end(); I != E; ++I) {
13297 // Quit if not all elements are used.
13298 if (I->second != FullMask)
13302 EVT TestVT = VT.is128BitVector() ? MVT::v2i64 : MVT::v4i64;
13304 // Cast all vectors into TestVT for PTEST.
13305 for (unsigned i = 0, e = VecIns.size(); i < e; ++i)
13306 VecIns[i] = DAG.getBitcast(TestVT, VecIns[i]);
13308 // If more than one full vectors are evaluated, OR them first before PTEST.
13309 for (unsigned Slot = 0, e = VecIns.size(); e - Slot > 1; Slot += 2, e += 1) {
13310 // Each iteration will OR 2 nodes and append the result until there is only
13311 // 1 node left, i.e. the final OR'd value of all vectors.
13312 SDValue LHS = VecIns[Slot];
13313 SDValue RHS = VecIns[Slot + 1];
13314 VecIns.push_back(DAG.getNode(ISD::OR, DL, TestVT, LHS, RHS));
13317 return DAG.getNode(X86ISD::PTEST, DL, MVT::i32,
13318 VecIns.back(), VecIns.back());
13321 /// \brief return true if \c Op has a use that doesn't just read flags.
13322 static bool hasNonFlagsUse(SDValue Op) {
13323 for (SDNode::use_iterator UI = Op->use_begin(), UE = Op->use_end(); UI != UE;
13325 SDNode *User = *UI;
13326 unsigned UOpNo = UI.getOperandNo();
13327 if (User->getOpcode() == ISD::TRUNCATE && User->hasOneUse()) {
13328 // Look pass truncate.
13329 UOpNo = User->use_begin().getOperandNo();
13330 User = *User->use_begin();
13333 if (User->getOpcode() != ISD::BRCOND && User->getOpcode() != ISD::SETCC &&
13334 !(User->getOpcode() == ISD::SELECT && UOpNo == 0))
13340 /// Emit nodes that will be selected as "test Op0,Op0", or something
13342 SDValue X86TargetLowering::EmitTest(SDValue Op, unsigned X86CC, SDLoc dl,
13343 SelectionDAG &DAG) const {
13344 if (Op.getValueType() == MVT::i1) {
13345 SDValue ExtOp = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i8, Op);
13346 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, ExtOp,
13347 DAG.getConstant(0, dl, MVT::i8));
13349 // CF and OF aren't always set the way we want. Determine which
13350 // of these we need.
13351 bool NeedCF = false;
13352 bool NeedOF = false;
13355 case X86::COND_A: case X86::COND_AE:
13356 case X86::COND_B: case X86::COND_BE:
13359 case X86::COND_G: case X86::COND_GE:
13360 case X86::COND_L: case X86::COND_LE:
13361 case X86::COND_O: case X86::COND_NO: {
13362 // Check if we really need to set the
13363 // Overflow flag. If NoSignedWrap is present
13364 // that is not actually needed.
13365 switch (Op->getOpcode()) {
13370 const auto *BinNode = cast<BinaryWithFlagsSDNode>(Op.getNode());
13371 if (BinNode->Flags.hasNoSignedWrap())
13381 // See if we can use the EFLAGS value from the operand instead of
13382 // doing a separate TEST. TEST always sets OF and CF to 0, so unless
13383 // we prove that the arithmetic won't overflow, we can't use OF or CF.
13384 if (Op.getResNo() != 0 || NeedOF || NeedCF) {
13385 // Emit a CMP with 0, which is the TEST pattern.
13386 //if (Op.getValueType() == MVT::i1)
13387 // return DAG.getNode(X86ISD::CMP, dl, MVT::i1, Op,
13388 // DAG.getConstant(0, MVT::i1));
13389 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op,
13390 DAG.getConstant(0, dl, Op.getValueType()));
13392 unsigned Opcode = 0;
13393 unsigned NumOperands = 0;
13395 // Truncate operations may prevent the merge of the SETCC instruction
13396 // and the arithmetic instruction before it. Attempt to truncate the operands
13397 // of the arithmetic instruction and use a reduced bit-width instruction.
13398 bool NeedTruncation = false;
13399 SDValue ArithOp = Op;
13400 if (Op->getOpcode() == ISD::TRUNCATE && Op->hasOneUse()) {
13401 SDValue Arith = Op->getOperand(0);
13402 // Both the trunc and the arithmetic op need to have one user each.
13403 if (Arith->hasOneUse())
13404 switch (Arith.getOpcode()) {
13411 NeedTruncation = true;
13417 // NOTICE: In the code below we use ArithOp to hold the arithmetic operation
13418 // which may be the result of a CAST. We use the variable 'Op', which is the
13419 // non-casted variable when we check for possible users.
13420 switch (ArithOp.getOpcode()) {
13422 // Due to an isel shortcoming, be conservative if this add is likely to be
13423 // selected as part of a load-modify-store instruction. When the root node
13424 // in a match is a store, isel doesn't know how to remap non-chain non-flag
13425 // uses of other nodes in the match, such as the ADD in this case. This
13426 // leads to the ADD being left around and reselected, with the result being
13427 // two adds in the output. Alas, even if none our users are stores, that
13428 // doesn't prove we're O.K. Ergo, if we have any parents that aren't
13429 // CopyToReg or SETCC, eschew INC/DEC. A better fix seems to require
13430 // climbing the DAG back to the root, and it doesn't seem to be worth the
13432 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
13433 UE = Op.getNode()->use_end(); UI != UE; ++UI)
13434 if (UI->getOpcode() != ISD::CopyToReg &&
13435 UI->getOpcode() != ISD::SETCC &&
13436 UI->getOpcode() != ISD::STORE)
13439 if (ConstantSDNode *C =
13440 dyn_cast<ConstantSDNode>(ArithOp.getNode()->getOperand(1))) {
13441 // An add of one will be selected as an INC.
13442 if (C->getAPIntValue() == 1 && !Subtarget->slowIncDec()) {
13443 Opcode = X86ISD::INC;
13448 // An add of negative one (subtract of one) will be selected as a DEC.
13449 if (C->getAPIntValue().isAllOnesValue() && !Subtarget->slowIncDec()) {
13450 Opcode = X86ISD::DEC;
13456 // Otherwise use a regular EFLAGS-setting add.
13457 Opcode = X86ISD::ADD;
13462 // If we have a constant logical shift that's only used in a comparison
13463 // against zero turn it into an equivalent AND. This allows turning it into
13464 // a TEST instruction later.
13465 if ((X86CC == X86::COND_E || X86CC == X86::COND_NE) && Op->hasOneUse() &&
13466 isa<ConstantSDNode>(Op->getOperand(1)) && !hasNonFlagsUse(Op)) {
13467 EVT VT = Op.getValueType();
13468 unsigned BitWidth = VT.getSizeInBits();
13469 unsigned ShAmt = Op->getConstantOperandVal(1);
13470 if (ShAmt >= BitWidth) // Avoid undefined shifts.
13472 APInt Mask = ArithOp.getOpcode() == ISD::SRL
13473 ? APInt::getHighBitsSet(BitWidth, BitWidth - ShAmt)
13474 : APInt::getLowBitsSet(BitWidth, BitWidth - ShAmt);
13475 if (!Mask.isSignedIntN(32)) // Avoid large immediates.
13477 SDValue New = DAG.getNode(ISD::AND, dl, VT, Op->getOperand(0),
13478 DAG.getConstant(Mask, dl, VT));
13479 DAG.ReplaceAllUsesWith(Op, New);
13485 // If the primary and result isn't used, don't bother using X86ISD::AND,
13486 // because a TEST instruction will be better.
13487 if (!hasNonFlagsUse(Op))
13493 // Due to the ISEL shortcoming noted above, be conservative if this op is
13494 // likely to be selected as part of a load-modify-store instruction.
13495 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
13496 UE = Op.getNode()->use_end(); UI != UE; ++UI)
13497 if (UI->getOpcode() == ISD::STORE)
13500 // Otherwise use a regular EFLAGS-setting instruction.
13501 switch (ArithOp.getOpcode()) {
13502 default: llvm_unreachable("unexpected operator!");
13503 case ISD::SUB: Opcode = X86ISD::SUB; break;
13504 case ISD::XOR: Opcode = X86ISD::XOR; break;
13505 case ISD::AND: Opcode = X86ISD::AND; break;
13507 if (!NeedTruncation && (X86CC == X86::COND_E || X86CC == X86::COND_NE)) {
13508 SDValue EFLAGS = LowerVectorAllZeroTest(Op, Subtarget, DAG);
13509 if (EFLAGS.getNode())
13512 Opcode = X86ISD::OR;
13526 return SDValue(Op.getNode(), 1);
13532 // If we found that truncation is beneficial, perform the truncation and
13534 if (NeedTruncation) {
13535 EVT VT = Op.getValueType();
13536 SDValue WideVal = Op->getOperand(0);
13537 EVT WideVT = WideVal.getValueType();
13538 unsigned ConvertedOp = 0;
13539 // Use a target machine opcode to prevent further DAGCombine
13540 // optimizations that may separate the arithmetic operations
13541 // from the setcc node.
13542 switch (WideVal.getOpcode()) {
13544 case ISD::ADD: ConvertedOp = X86ISD::ADD; break;
13545 case ISD::SUB: ConvertedOp = X86ISD::SUB; break;
13546 case ISD::AND: ConvertedOp = X86ISD::AND; break;
13547 case ISD::OR: ConvertedOp = X86ISD::OR; break;
13548 case ISD::XOR: ConvertedOp = X86ISD::XOR; break;
13552 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
13553 if (TLI.isOperationLegal(WideVal.getOpcode(), WideVT)) {
13554 SDValue V0 = DAG.getNode(ISD::TRUNCATE, dl, VT, WideVal.getOperand(0));
13555 SDValue V1 = DAG.getNode(ISD::TRUNCATE, dl, VT, WideVal.getOperand(1));
13556 Op = DAG.getNode(ConvertedOp, dl, VT, V0, V1);
13562 // Emit a CMP with 0, which is the TEST pattern.
13563 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op,
13564 DAG.getConstant(0, dl, Op.getValueType()));
13566 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::i32);
13567 SmallVector<SDValue, 4> Ops(Op->op_begin(), Op->op_begin() + NumOperands);
13569 SDValue New = DAG.getNode(Opcode, dl, VTs, Ops);
13570 DAG.ReplaceAllUsesWith(Op, New);
13571 return SDValue(New.getNode(), 1);
13574 /// Emit nodes that will be selected as "cmp Op0,Op1", or something
13576 SDValue X86TargetLowering::EmitCmp(SDValue Op0, SDValue Op1, unsigned X86CC,
13577 SDLoc dl, SelectionDAG &DAG) const {
13578 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op1)) {
13579 if (C->getAPIntValue() == 0)
13580 return EmitTest(Op0, X86CC, dl, DAG);
13582 if (Op0.getValueType() == MVT::i1)
13583 llvm_unreachable("Unexpected comparison operation for MVT::i1 operands");
13586 if ((Op0.getValueType() == MVT::i8 || Op0.getValueType() == MVT::i16 ||
13587 Op0.getValueType() == MVT::i32 || Op0.getValueType() == MVT::i64)) {
13588 // Do the comparison at i32 if it's smaller, besides the Atom case.
13589 // This avoids subregister aliasing issues. Keep the smaller reference
13590 // if we're optimizing for size, however, as that'll allow better folding
13591 // of memory operations.
13592 if (Op0.getValueType() != MVT::i32 && Op0.getValueType() != MVT::i64 &&
13593 !DAG.getMachineFunction().getFunction()->optForMinSize() &&
13594 !Subtarget->isAtom()) {
13595 unsigned ExtendOp =
13596 isX86CCUnsigned(X86CC) ? ISD::ZERO_EXTEND : ISD::SIGN_EXTEND;
13597 Op0 = DAG.getNode(ExtendOp, dl, MVT::i32, Op0);
13598 Op1 = DAG.getNode(ExtendOp, dl, MVT::i32, Op1);
13600 // Use SUB instead of CMP to enable CSE between SUB and CMP.
13601 SDVTList VTs = DAG.getVTList(Op0.getValueType(), MVT::i32);
13602 SDValue Sub = DAG.getNode(X86ISD::SUB, dl, VTs,
13604 return SDValue(Sub.getNode(), 1);
13606 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op0, Op1);
13609 /// Convert a comparison if required by the subtarget.
13610 SDValue X86TargetLowering::ConvertCmpIfNecessary(SDValue Cmp,
13611 SelectionDAG &DAG) const {
13612 // If the subtarget does not support the FUCOMI instruction, floating-point
13613 // comparisons have to be converted.
13614 if (Subtarget->hasCMov() ||
13615 Cmp.getOpcode() != X86ISD::CMP ||
13616 !Cmp.getOperand(0).getValueType().isFloatingPoint() ||
13617 !Cmp.getOperand(1).getValueType().isFloatingPoint())
13620 // The instruction selector will select an FUCOM instruction instead of
13621 // FUCOMI, which writes the comparison result to FPSW instead of EFLAGS. Hence
13622 // build an SDNode sequence that transfers the result from FPSW into EFLAGS:
13623 // (X86sahf (trunc (srl (X86fp_stsw (trunc (X86cmp ...)), 8))))
13625 SDValue TruncFPSW = DAG.getNode(ISD::TRUNCATE, dl, MVT::i16, Cmp);
13626 SDValue FNStSW = DAG.getNode(X86ISD::FNSTSW16r, dl, MVT::i16, TruncFPSW);
13627 SDValue Srl = DAG.getNode(ISD::SRL, dl, MVT::i16, FNStSW,
13628 DAG.getConstant(8, dl, MVT::i8));
13629 SDValue TruncSrl = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Srl);
13630 return DAG.getNode(X86ISD::SAHF, dl, MVT::i32, TruncSrl);
13633 /// The minimum architected relative accuracy is 2^-12. We need one
13634 /// Newton-Raphson step to have a good float result (24 bits of precision).
13635 SDValue X86TargetLowering::getRsqrtEstimate(SDValue Op,
13636 DAGCombinerInfo &DCI,
13637 unsigned &RefinementSteps,
13638 bool &UseOneConstNR) const {
13639 EVT VT = Op.getValueType();
13640 const char *RecipOp;
13642 // SSE1 has rsqrtss and rsqrtps. AVX adds a 256-bit variant for rsqrtps.
13643 // TODO: Add support for AVX512 (v16f32).
13644 // It is likely not profitable to do this for f64 because a double-precision
13645 // rsqrt estimate with refinement on x86 prior to FMA requires at least 16
13646 // instructions: convert to single, rsqrtss, convert back to double, refine
13647 // (3 steps = at least 13 insts). If an 'rsqrtsd' variant was added to the ISA
13648 // along with FMA, this could be a throughput win.
13649 if (VT == MVT::f32 && Subtarget->hasSSE1())
13651 else if ((VT == MVT::v4f32 && Subtarget->hasSSE1()) ||
13652 (VT == MVT::v8f32 && Subtarget->hasAVX()))
13653 RecipOp = "vec-sqrtf";
13657 TargetRecip Recips = DCI.DAG.getTarget().Options.Reciprocals;
13658 if (!Recips.isEnabled(RecipOp))
13661 RefinementSteps = Recips.getRefinementSteps(RecipOp);
13662 UseOneConstNR = false;
13663 return DCI.DAG.getNode(X86ISD::FRSQRT, SDLoc(Op), VT, Op);
13666 /// The minimum architected relative accuracy is 2^-12. We need one
13667 /// Newton-Raphson step to have a good float result (24 bits of precision).
13668 SDValue X86TargetLowering::getRecipEstimate(SDValue Op,
13669 DAGCombinerInfo &DCI,
13670 unsigned &RefinementSteps) const {
13671 EVT VT = Op.getValueType();
13672 const char *RecipOp;
13674 // SSE1 has rcpss and rcpps. AVX adds a 256-bit variant for rcpps.
13675 // TODO: Add support for AVX512 (v16f32).
13676 // It is likely not profitable to do this for f64 because a double-precision
13677 // reciprocal estimate with refinement on x86 prior to FMA requires
13678 // 15 instructions: convert to single, rcpss, convert back to double, refine
13679 // (3 steps = 12 insts). If an 'rcpsd' variant was added to the ISA
13680 // along with FMA, this could be a throughput win.
13681 if (VT == MVT::f32 && Subtarget->hasSSE1())
13683 else if ((VT == MVT::v4f32 && Subtarget->hasSSE1()) ||
13684 (VT == MVT::v8f32 && Subtarget->hasAVX()))
13685 RecipOp = "vec-divf";
13689 TargetRecip Recips = DCI.DAG.getTarget().Options.Reciprocals;
13690 if (!Recips.isEnabled(RecipOp))
13693 RefinementSteps = Recips.getRefinementSteps(RecipOp);
13694 return DCI.DAG.getNode(X86ISD::FRCP, SDLoc(Op), VT, Op);
13697 /// If we have at least two divisions that use the same divisor, convert to
13698 /// multplication by a reciprocal. This may need to be adjusted for a given
13699 /// CPU if a division's cost is not at least twice the cost of a multiplication.
13700 /// This is because we still need one division to calculate the reciprocal and
13701 /// then we need two multiplies by that reciprocal as replacements for the
13702 /// original divisions.
13703 unsigned X86TargetLowering::combineRepeatedFPDivisors() const {
13707 static bool isAllOnes(SDValue V) {
13708 ConstantSDNode *C = dyn_cast<ConstantSDNode>(V);
13709 return C && C->isAllOnesValue();
13712 /// LowerToBT - Result of 'and' is compared against zero. Turn it into a BT node
13713 /// if it's possible.
13714 SDValue X86TargetLowering::LowerToBT(SDValue And, ISD::CondCode CC,
13715 SDLoc dl, SelectionDAG &DAG) const {
13716 SDValue Op0 = And.getOperand(0);
13717 SDValue Op1 = And.getOperand(1);
13718 if (Op0.getOpcode() == ISD::TRUNCATE)
13719 Op0 = Op0.getOperand(0);
13720 if (Op1.getOpcode() == ISD::TRUNCATE)
13721 Op1 = Op1.getOperand(0);
13724 if (Op1.getOpcode() == ISD::SHL)
13725 std::swap(Op0, Op1);
13726 if (Op0.getOpcode() == ISD::SHL) {
13727 if (ConstantSDNode *And00C = dyn_cast<ConstantSDNode>(Op0.getOperand(0)))
13728 if (And00C->getZExtValue() == 1) {
13729 // If we looked past a truncate, check that it's only truncating away
13731 unsigned BitWidth = Op0.getValueSizeInBits();
13732 unsigned AndBitWidth = And.getValueSizeInBits();
13733 if (BitWidth > AndBitWidth) {
13735 DAG.computeKnownBits(Op0, Zeros, Ones);
13736 if (Zeros.countLeadingOnes() < BitWidth - AndBitWidth)
13740 RHS = Op0.getOperand(1);
13742 } else if (Op1.getOpcode() == ISD::Constant) {
13743 ConstantSDNode *AndRHS = cast<ConstantSDNode>(Op1);
13744 uint64_t AndRHSVal = AndRHS->getZExtValue();
13745 SDValue AndLHS = Op0;
13747 if (AndRHSVal == 1 && AndLHS.getOpcode() == ISD::SRL) {
13748 LHS = AndLHS.getOperand(0);
13749 RHS = AndLHS.getOperand(1);
13752 // Use BT if the immediate can't be encoded in a TEST instruction.
13753 if (!isUInt<32>(AndRHSVal) && isPowerOf2_64(AndRHSVal)) {
13755 RHS = DAG.getConstant(Log2_64_Ceil(AndRHSVal), dl, LHS.getValueType());
13759 if (LHS.getNode()) {
13760 // If LHS is i8, promote it to i32 with any_extend. There is no i8 BT
13761 // instruction. Since the shift amount is in-range-or-undefined, we know
13762 // that doing a bittest on the i32 value is ok. We extend to i32 because
13763 // the encoding for the i16 version is larger than the i32 version.
13764 // Also promote i16 to i32 for performance / code size reason.
13765 if (LHS.getValueType() == MVT::i8 ||
13766 LHS.getValueType() == MVT::i16)
13767 LHS = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, LHS);
13769 // If the operand types disagree, extend the shift amount to match. Since
13770 // BT ignores high bits (like shifts) we can use anyextend.
13771 if (LHS.getValueType() != RHS.getValueType())
13772 RHS = DAG.getNode(ISD::ANY_EXTEND, dl, LHS.getValueType(), RHS);
13774 SDValue BT = DAG.getNode(X86ISD::BT, dl, MVT::i32, LHS, RHS);
13775 X86::CondCode Cond = CC == ISD::SETEQ ? X86::COND_AE : X86::COND_B;
13776 return DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
13777 DAG.getConstant(Cond, dl, MVT::i8), BT);
13783 /// \brief - Turns an ISD::CondCode into a value suitable for SSE floating point
13785 static int translateX86FSETCC(ISD::CondCode SetCCOpcode, SDValue &Op0,
13790 // SSE Condition code mapping:
13799 switch (SetCCOpcode) {
13800 default: llvm_unreachable("Unexpected SETCC condition");
13802 case ISD::SETEQ: SSECC = 0; break;
13804 case ISD::SETGT: Swap = true; // Fallthrough
13806 case ISD::SETOLT: SSECC = 1; break;
13808 case ISD::SETGE: Swap = true; // Fallthrough
13810 case ISD::SETOLE: SSECC = 2; break;
13811 case ISD::SETUO: SSECC = 3; break;
13813 case ISD::SETNE: SSECC = 4; break;
13814 case ISD::SETULE: Swap = true; // Fallthrough
13815 case ISD::SETUGE: SSECC = 5; break;
13816 case ISD::SETULT: Swap = true; // Fallthrough
13817 case ISD::SETUGT: SSECC = 6; break;
13818 case ISD::SETO: SSECC = 7; break;
13820 case ISD::SETONE: SSECC = 8; break;
13823 std::swap(Op0, Op1);
13828 // Lower256IntVSETCC - Break a VSETCC 256-bit integer VSETCC into two new 128
13829 // ones, and then concatenate the result back.
13830 static SDValue Lower256IntVSETCC(SDValue Op, SelectionDAG &DAG) {
13831 MVT VT = Op.getSimpleValueType();
13833 assert(VT.is256BitVector() && Op.getOpcode() == ISD::SETCC &&
13834 "Unsupported value type for operation");
13836 unsigned NumElems = VT.getVectorNumElements();
13838 SDValue CC = Op.getOperand(2);
13840 // Extract the LHS vectors
13841 SDValue LHS = Op.getOperand(0);
13842 SDValue LHS1 = Extract128BitVector(LHS, 0, DAG, dl);
13843 SDValue LHS2 = Extract128BitVector(LHS, NumElems/2, DAG, dl);
13845 // Extract the RHS vectors
13846 SDValue RHS = Op.getOperand(1);
13847 SDValue RHS1 = Extract128BitVector(RHS, 0, DAG, dl);
13848 SDValue RHS2 = Extract128BitVector(RHS, NumElems/2, DAG, dl);
13850 // Issue the operation on the smaller types and concatenate the result back
13851 MVT EltVT = VT.getVectorElementType();
13852 MVT NewVT = MVT::getVectorVT(EltVT, NumElems/2);
13853 return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT,
13854 DAG.getNode(Op.getOpcode(), dl, NewVT, LHS1, RHS1, CC),
13855 DAG.getNode(Op.getOpcode(), dl, NewVT, LHS2, RHS2, CC));
13858 static SDValue LowerBoolVSETCC_AVX512(SDValue Op, SelectionDAG &DAG) {
13859 SDValue Op0 = Op.getOperand(0);
13860 SDValue Op1 = Op.getOperand(1);
13861 SDValue CC = Op.getOperand(2);
13862 MVT VT = Op.getSimpleValueType();
13865 assert(Op0.getValueType().getVectorElementType() == MVT::i1 &&
13866 "Unexpected type for boolean compare operation");
13867 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
13868 SDValue NotOp0 = DAG.getNode(ISD::XOR, dl, VT, Op0,
13869 DAG.getConstant(-1, dl, VT));
13870 SDValue NotOp1 = DAG.getNode(ISD::XOR, dl, VT, Op1,
13871 DAG.getConstant(-1, dl, VT));
13872 switch (SetCCOpcode) {
13873 default: llvm_unreachable("Unexpected SETCC condition");
13875 // (x == y) -> ~(x ^ y)
13876 return DAG.getNode(ISD::XOR, dl, VT,
13877 DAG.getNode(ISD::XOR, dl, VT, Op0, Op1),
13878 DAG.getConstant(-1, dl, VT));
13880 // (x != y) -> (x ^ y)
13881 return DAG.getNode(ISD::XOR, dl, VT, Op0, Op1);
13884 // (x > y) -> (x & ~y)
13885 return DAG.getNode(ISD::AND, dl, VT, Op0, NotOp1);
13888 // (x < y) -> (~x & y)
13889 return DAG.getNode(ISD::AND, dl, VT, NotOp0, Op1);
13892 // (x <= y) -> (~x | y)
13893 return DAG.getNode(ISD::OR, dl, VT, NotOp0, Op1);
13896 // (x >=y) -> (x | ~y)
13897 return DAG.getNode(ISD::OR, dl, VT, Op0, NotOp1);
13901 static SDValue LowerIntVSETCC_AVX512(SDValue Op, SelectionDAG &DAG,
13902 const X86Subtarget *Subtarget) {
13903 SDValue Op0 = Op.getOperand(0);
13904 SDValue Op1 = Op.getOperand(1);
13905 SDValue CC = Op.getOperand(2);
13906 MVT VT = Op.getSimpleValueType();
13909 assert(Op0.getValueType().getVectorElementType().getSizeInBits() >= 8 &&
13910 Op.getValueType().getScalarType() == MVT::i1 &&
13911 "Cannot set masked compare for this operation");
13913 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
13915 bool Unsigned = false;
13918 switch (SetCCOpcode) {
13919 default: llvm_unreachable("Unexpected SETCC condition");
13920 case ISD::SETNE: SSECC = 4; break;
13921 case ISD::SETEQ: Opc = X86ISD::PCMPEQM; break;
13922 case ISD::SETUGT: SSECC = 6; Unsigned = true; break;
13923 case ISD::SETLT: Swap = true; //fall-through
13924 case ISD::SETGT: Opc = X86ISD::PCMPGTM; break;
13925 case ISD::SETULT: SSECC = 1; Unsigned = true; break;
13926 case ISD::SETUGE: SSECC = 5; Unsigned = true; break; //NLT
13927 case ISD::SETGE: Swap = true; SSECC = 2; break; // LE + swap
13928 case ISD::SETULE: Unsigned = true; //fall-through
13929 case ISD::SETLE: SSECC = 2; break;
13933 std::swap(Op0, Op1);
13935 return DAG.getNode(Opc, dl, VT, Op0, Op1);
13936 Opc = Unsigned ? X86ISD::CMPMU: X86ISD::CMPM;
13937 return DAG.getNode(Opc, dl, VT, Op0, Op1,
13938 DAG.getConstant(SSECC, dl, MVT::i8));
13941 /// \brief Try to turn a VSETULT into a VSETULE by modifying its second
13942 /// operand \p Op1. If non-trivial (for example because it's not constant)
13943 /// return an empty value.
13944 static SDValue ChangeVSETULTtoVSETULE(SDLoc dl, SDValue Op1, SelectionDAG &DAG)
13946 BuildVectorSDNode *BV = dyn_cast<BuildVectorSDNode>(Op1.getNode());
13950 MVT VT = Op1.getSimpleValueType();
13951 MVT EVT = VT.getVectorElementType();
13952 unsigned n = VT.getVectorNumElements();
13953 SmallVector<SDValue, 8> ULTOp1;
13955 for (unsigned i = 0; i < n; ++i) {
13956 ConstantSDNode *Elt = dyn_cast<ConstantSDNode>(BV->getOperand(i));
13957 if (!Elt || Elt->isOpaque() || Elt->getValueType(0) != EVT)
13960 // Avoid underflow.
13961 APInt Val = Elt->getAPIntValue();
13965 ULTOp1.push_back(DAG.getConstant(Val - 1, dl, EVT));
13968 return DAG.getNode(ISD::BUILD_VECTOR, dl, VT, ULTOp1);
13971 static SDValue LowerVSETCC(SDValue Op, const X86Subtarget *Subtarget,
13972 SelectionDAG &DAG) {
13973 SDValue Op0 = Op.getOperand(0);
13974 SDValue Op1 = Op.getOperand(1);
13975 SDValue CC = Op.getOperand(2);
13976 MVT VT = Op.getSimpleValueType();
13977 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
13978 bool isFP = Op.getOperand(1).getSimpleValueType().isFloatingPoint();
13983 MVT EltVT = Op0.getSimpleValueType().getVectorElementType();
13984 assert(EltVT == MVT::f32 || EltVT == MVT::f64);
13987 unsigned SSECC = translateX86FSETCC(SetCCOpcode, Op0, Op1);
13988 unsigned Opc = X86ISD::CMPP;
13989 if (Subtarget->hasAVX512() && VT.getVectorElementType() == MVT::i1) {
13990 assert(VT.getVectorNumElements() <= 16);
13991 Opc = X86ISD::CMPM;
13993 // In the two special cases we can't handle, emit two comparisons.
13996 unsigned CombineOpc;
13997 if (SetCCOpcode == ISD::SETUEQ) {
13998 CC0 = 3; CC1 = 0; CombineOpc = ISD::OR;
14000 assert(SetCCOpcode == ISD::SETONE);
14001 CC0 = 7; CC1 = 4; CombineOpc = ISD::AND;
14004 SDValue Cmp0 = DAG.getNode(Opc, dl, VT, Op0, Op1,
14005 DAG.getConstant(CC0, dl, MVT::i8));
14006 SDValue Cmp1 = DAG.getNode(Opc, dl, VT, Op0, Op1,
14007 DAG.getConstant(CC1, dl, MVT::i8));
14008 return DAG.getNode(CombineOpc, dl, VT, Cmp0, Cmp1);
14010 // Handle all other FP comparisons here.
14011 return DAG.getNode(Opc, dl, VT, Op0, Op1,
14012 DAG.getConstant(SSECC, dl, MVT::i8));
14015 // Break 256-bit integer vector compare into smaller ones.
14016 if (VT.is256BitVector() && !Subtarget->hasInt256())
14017 return Lower256IntVSETCC(Op, DAG);
14019 EVT OpVT = Op1.getValueType();
14020 if (OpVT.getVectorElementType() == MVT::i1)
14021 return LowerBoolVSETCC_AVX512(Op, DAG);
14023 bool MaskResult = (VT.getVectorElementType() == MVT::i1);
14024 if (Subtarget->hasAVX512()) {
14025 if (Op1.getValueType().is512BitVector() ||
14026 (Subtarget->hasBWI() && Subtarget->hasVLX()) ||
14027 (MaskResult && OpVT.getVectorElementType().getSizeInBits() >= 32))
14028 return LowerIntVSETCC_AVX512(Op, DAG, Subtarget);
14030 // In AVX-512 architecture setcc returns mask with i1 elements,
14031 // But there is no compare instruction for i8 and i16 elements in KNL.
14032 // We are not talking about 512-bit operands in this case, these
14033 // types are illegal.
14035 (OpVT.getVectorElementType().getSizeInBits() < 32 &&
14036 OpVT.getVectorElementType().getSizeInBits() >= 8))
14037 return DAG.getNode(ISD::TRUNCATE, dl, VT,
14038 DAG.getNode(ISD::SETCC, dl, OpVT, Op0, Op1, CC));
14041 // We are handling one of the integer comparisons here. Since SSE only has
14042 // GT and EQ comparisons for integer, swapping operands and multiple
14043 // operations may be required for some comparisons.
14045 bool Swap = false, Invert = false, FlipSigns = false, MinMax = false;
14046 bool Subus = false;
14048 switch (SetCCOpcode) {
14049 default: llvm_unreachable("Unexpected SETCC condition");
14050 case ISD::SETNE: Invert = true;
14051 case ISD::SETEQ: Opc = X86ISD::PCMPEQ; break;
14052 case ISD::SETLT: Swap = true;
14053 case ISD::SETGT: Opc = X86ISD::PCMPGT; break;
14054 case ISD::SETGE: Swap = true;
14055 case ISD::SETLE: Opc = X86ISD::PCMPGT;
14056 Invert = true; break;
14057 case ISD::SETULT: Swap = true;
14058 case ISD::SETUGT: Opc = X86ISD::PCMPGT;
14059 FlipSigns = true; break;
14060 case ISD::SETUGE: Swap = true;
14061 case ISD::SETULE: Opc = X86ISD::PCMPGT;
14062 FlipSigns = true; Invert = true; break;
14065 // Special case: Use min/max operations for SETULE/SETUGE
14066 MVT VET = VT.getVectorElementType();
14068 (Subtarget->hasSSE41() && (VET >= MVT::i8 && VET <= MVT::i32))
14069 || (Subtarget->hasSSE2() && (VET == MVT::i8));
14072 switch (SetCCOpcode) {
14074 case ISD::SETULE: Opc = ISD::UMIN; MinMax = true; break;
14075 case ISD::SETUGE: Opc = ISD::UMAX; MinMax = true; break;
14078 if (MinMax) { Swap = false; Invert = false; FlipSigns = false; }
14081 bool hasSubus = Subtarget->hasSSE2() && (VET == MVT::i8 || VET == MVT::i16);
14082 if (!MinMax && hasSubus) {
14083 // As another special case, use PSUBUS[BW] when it's profitable. E.g. for
14085 // t = psubus Op0, Op1
14086 // pcmpeq t, <0..0>
14087 switch (SetCCOpcode) {
14089 case ISD::SETULT: {
14090 // If the comparison is against a constant we can turn this into a
14091 // setule. With psubus, setule does not require a swap. This is
14092 // beneficial because the constant in the register is no longer
14093 // destructed as the destination so it can be hoisted out of a loop.
14094 // Only do this pre-AVX since vpcmp* is no longer destructive.
14095 if (Subtarget->hasAVX())
14097 SDValue ULEOp1 = ChangeVSETULTtoVSETULE(dl, Op1, DAG);
14098 if (ULEOp1.getNode()) {
14100 Subus = true; Invert = false; Swap = false;
14104 // Psubus is better than flip-sign because it requires no inversion.
14105 case ISD::SETUGE: Subus = true; Invert = false; Swap = true; break;
14106 case ISD::SETULE: Subus = true; Invert = false; Swap = false; break;
14110 Opc = X86ISD::SUBUS;
14116 std::swap(Op0, Op1);
14118 // Check that the operation in question is available (most are plain SSE2,
14119 // but PCMPGTQ and PCMPEQQ have different requirements).
14120 if (VT == MVT::v2i64) {
14121 if (Opc == X86ISD::PCMPGT && !Subtarget->hasSSE42()) {
14122 assert(Subtarget->hasSSE2() && "Don't know how to lower!");
14124 // First cast everything to the right type.
14125 Op0 = DAG.getBitcast(MVT::v4i32, Op0);
14126 Op1 = DAG.getBitcast(MVT::v4i32, Op1);
14128 // Since SSE has no unsigned integer comparisons, we need to flip the sign
14129 // bits of the inputs before performing those operations. The lower
14130 // compare is always unsigned.
14133 SB = DAG.getConstant(0x80000000U, dl, MVT::v4i32);
14135 SDValue Sign = DAG.getConstant(0x80000000U, dl, MVT::i32);
14136 SDValue Zero = DAG.getConstant(0x00000000U, dl, MVT::i32);
14137 SB = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32,
14138 Sign, Zero, Sign, Zero);
14140 Op0 = DAG.getNode(ISD::XOR, dl, MVT::v4i32, Op0, SB);
14141 Op1 = DAG.getNode(ISD::XOR, dl, MVT::v4i32, Op1, SB);
14143 // Emulate PCMPGTQ with (hi1 > hi2) | ((hi1 == hi2) & (lo1 > lo2))
14144 SDValue GT = DAG.getNode(X86ISD::PCMPGT, dl, MVT::v4i32, Op0, Op1);
14145 SDValue EQ = DAG.getNode(X86ISD::PCMPEQ, dl, MVT::v4i32, Op0, Op1);
14147 // Create masks for only the low parts/high parts of the 64 bit integers.
14148 static const int MaskHi[] = { 1, 1, 3, 3 };
14149 static const int MaskLo[] = { 0, 0, 2, 2 };
14150 SDValue EQHi = DAG.getVectorShuffle(MVT::v4i32, dl, EQ, EQ, MaskHi);
14151 SDValue GTLo = DAG.getVectorShuffle(MVT::v4i32, dl, GT, GT, MaskLo);
14152 SDValue GTHi = DAG.getVectorShuffle(MVT::v4i32, dl, GT, GT, MaskHi);
14154 SDValue Result = DAG.getNode(ISD::AND, dl, MVT::v4i32, EQHi, GTLo);
14155 Result = DAG.getNode(ISD::OR, dl, MVT::v4i32, Result, GTHi);
14158 Result = DAG.getNOT(dl, Result, MVT::v4i32);
14160 return DAG.getBitcast(VT, Result);
14163 if (Opc == X86ISD::PCMPEQ && !Subtarget->hasSSE41()) {
14164 // If pcmpeqq is missing but pcmpeqd is available synthesize pcmpeqq with
14165 // pcmpeqd + pshufd + pand.
14166 assert(Subtarget->hasSSE2() && !FlipSigns && "Don't know how to lower!");
14168 // First cast everything to the right type.
14169 Op0 = DAG.getBitcast(MVT::v4i32, Op0);
14170 Op1 = DAG.getBitcast(MVT::v4i32, Op1);
14173 SDValue Result = DAG.getNode(Opc, dl, MVT::v4i32, Op0, Op1);
14175 // Make sure the lower and upper halves are both all-ones.
14176 static const int Mask[] = { 1, 0, 3, 2 };
14177 SDValue Shuf = DAG.getVectorShuffle(MVT::v4i32, dl, Result, Result, Mask);
14178 Result = DAG.getNode(ISD::AND, dl, MVT::v4i32, Result, Shuf);
14181 Result = DAG.getNOT(dl, Result, MVT::v4i32);
14183 return DAG.getBitcast(VT, Result);
14187 // Since SSE has no unsigned integer comparisons, we need to flip the sign
14188 // bits of the inputs before performing those operations.
14190 EVT EltVT = VT.getVectorElementType();
14191 SDValue SB = DAG.getConstant(APInt::getSignBit(EltVT.getSizeInBits()), dl,
14193 Op0 = DAG.getNode(ISD::XOR, dl, VT, Op0, SB);
14194 Op1 = DAG.getNode(ISD::XOR, dl, VT, Op1, SB);
14197 SDValue Result = DAG.getNode(Opc, dl, VT, Op0, Op1);
14199 // If the logical-not of the result is required, perform that now.
14201 Result = DAG.getNOT(dl, Result, VT);
14204 Result = DAG.getNode(X86ISD::PCMPEQ, dl, VT, Op0, Result);
14207 Result = DAG.getNode(X86ISD::PCMPEQ, dl, VT, Result,
14208 getZeroVector(VT, Subtarget, DAG, dl));
14213 SDValue X86TargetLowering::LowerSETCC(SDValue Op, SelectionDAG &DAG) const {
14215 MVT VT = Op.getSimpleValueType();
14217 if (VT.isVector()) return LowerVSETCC(Op, Subtarget, DAG);
14219 assert(((!Subtarget->hasAVX512() && VT == MVT::i8) || (VT == MVT::i1))
14220 && "SetCC type must be 8-bit or 1-bit integer");
14221 SDValue Op0 = Op.getOperand(0);
14222 SDValue Op1 = Op.getOperand(1);
14224 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
14226 // Optimize to BT if possible.
14227 // Lower (X & (1 << N)) == 0 to BT(X, N).
14228 // Lower ((X >>u N) & 1) != 0 to BT(X, N).
14229 // Lower ((X >>s N) & 1) != 0 to BT(X, N).
14230 if (Op0.getOpcode() == ISD::AND && Op0.hasOneUse() &&
14231 Op1.getOpcode() == ISD::Constant &&
14232 cast<ConstantSDNode>(Op1)->isNullValue() &&
14233 (CC == ISD::SETEQ || CC == ISD::SETNE)) {
14234 SDValue NewSetCC = LowerToBT(Op0, CC, dl, DAG);
14235 if (NewSetCC.getNode()) {
14237 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i1, NewSetCC);
14242 // Look for X == 0, X == 1, X != 0, or X != 1. We can simplify some forms of
14244 if (Op1.getOpcode() == ISD::Constant &&
14245 (cast<ConstantSDNode>(Op1)->getZExtValue() == 1 ||
14246 cast<ConstantSDNode>(Op1)->isNullValue()) &&
14247 (CC == ISD::SETEQ || CC == ISD::SETNE)) {
14249 // If the input is a setcc, then reuse the input setcc or use a new one with
14250 // the inverted condition.
14251 if (Op0.getOpcode() == X86ISD::SETCC) {
14252 X86::CondCode CCode = (X86::CondCode)Op0.getConstantOperandVal(0);
14253 bool Invert = (CC == ISD::SETNE) ^
14254 cast<ConstantSDNode>(Op1)->isNullValue();
14258 CCode = X86::GetOppositeBranchCondition(CCode);
14259 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
14260 DAG.getConstant(CCode, dl, MVT::i8),
14261 Op0.getOperand(1));
14263 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i1, SetCC);
14267 if ((Op0.getValueType() == MVT::i1) && (Op1.getOpcode() == ISD::Constant) &&
14268 (cast<ConstantSDNode>(Op1)->getZExtValue() == 1) &&
14269 (CC == ISD::SETEQ || CC == ISD::SETNE)) {
14271 ISD::CondCode NewCC = ISD::getSetCCInverse(CC, true);
14272 return DAG.getSetCC(dl, VT, Op0, DAG.getConstant(0, dl, MVT::i1), NewCC);
14275 bool isFP = Op1.getSimpleValueType().isFloatingPoint();
14276 unsigned X86CC = TranslateX86CC(CC, dl, isFP, Op0, Op1, DAG);
14277 if (X86CC == X86::COND_INVALID)
14280 SDValue EFLAGS = EmitCmp(Op0, Op1, X86CC, dl, DAG);
14281 EFLAGS = ConvertCmpIfNecessary(EFLAGS, DAG);
14282 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
14283 DAG.getConstant(X86CC, dl, MVT::i8), EFLAGS);
14285 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i1, SetCC);
14289 // isX86LogicalCmp - Return true if opcode is a X86 logical comparison.
14290 static bool isX86LogicalCmp(SDValue Op) {
14291 unsigned Opc = Op.getNode()->getOpcode();
14292 if (Opc == X86ISD::CMP || Opc == X86ISD::COMI || Opc == X86ISD::UCOMI ||
14293 Opc == X86ISD::SAHF)
14295 if (Op.getResNo() == 1 &&
14296 (Opc == X86ISD::ADD ||
14297 Opc == X86ISD::SUB ||
14298 Opc == X86ISD::ADC ||
14299 Opc == X86ISD::SBB ||
14300 Opc == X86ISD::SMUL ||
14301 Opc == X86ISD::UMUL ||
14302 Opc == X86ISD::INC ||
14303 Opc == X86ISD::DEC ||
14304 Opc == X86ISD::OR ||
14305 Opc == X86ISD::XOR ||
14306 Opc == X86ISD::AND))
14309 if (Op.getResNo() == 2 && Opc == X86ISD::UMUL)
14315 static bool isTruncWithZeroHighBitsInput(SDValue V, SelectionDAG &DAG) {
14316 if (V.getOpcode() != ISD::TRUNCATE)
14319 SDValue VOp0 = V.getOperand(0);
14320 unsigned InBits = VOp0.getValueSizeInBits();
14321 unsigned Bits = V.getValueSizeInBits();
14322 return DAG.MaskedValueIsZero(VOp0, APInt::getHighBitsSet(InBits,InBits-Bits));
14325 SDValue X86TargetLowering::LowerSELECT(SDValue Op, SelectionDAG &DAG) const {
14326 bool addTest = true;
14327 SDValue Cond = Op.getOperand(0);
14328 SDValue Op1 = Op.getOperand(1);
14329 SDValue Op2 = Op.getOperand(2);
14331 EVT VT = Op1.getValueType();
14334 // Lower FP selects into a CMP/AND/ANDN/OR sequence when the necessary SSE ops
14335 // are available or VBLENDV if AVX is available.
14336 // Otherwise FP cmovs get lowered into a less efficient branch sequence later.
14337 if (Cond.getOpcode() == ISD::SETCC &&
14338 ((Subtarget->hasSSE2() && (VT == MVT::f32 || VT == MVT::f64)) ||
14339 (Subtarget->hasSSE1() && VT == MVT::f32)) &&
14340 VT == Cond.getOperand(0).getValueType() && Cond->hasOneUse()) {
14341 SDValue CondOp0 = Cond.getOperand(0), CondOp1 = Cond.getOperand(1);
14342 int SSECC = translateX86FSETCC(
14343 cast<CondCodeSDNode>(Cond.getOperand(2))->get(), CondOp0, CondOp1);
14346 if (Subtarget->hasAVX512()) {
14347 SDValue Cmp = DAG.getNode(X86ISD::FSETCC, DL, MVT::i1, CondOp0, CondOp1,
14348 DAG.getConstant(SSECC, DL, MVT::i8));
14349 return DAG.getNode(X86ISD::SELECT, DL, VT, Cmp, Op1, Op2);
14352 SDValue Cmp = DAG.getNode(X86ISD::FSETCC, DL, VT, CondOp0, CondOp1,
14353 DAG.getConstant(SSECC, DL, MVT::i8));
14355 // If we have AVX, we can use a variable vector select (VBLENDV) instead
14356 // of 3 logic instructions for size savings and potentially speed.
14357 // Unfortunately, there is no scalar form of VBLENDV.
14359 // If either operand is a constant, don't try this. We can expect to
14360 // optimize away at least one of the logic instructions later in that
14361 // case, so that sequence would be faster than a variable blend.
14363 // BLENDV was introduced with SSE 4.1, but the 2 register form implicitly
14364 // uses XMM0 as the selection register. That may need just as many
14365 // instructions as the AND/ANDN/OR sequence due to register moves, so
14368 if (Subtarget->hasAVX() &&
14369 !isa<ConstantFPSDNode>(Op1) && !isa<ConstantFPSDNode>(Op2)) {
14371 // Convert to vectors, do a VSELECT, and convert back to scalar.
14372 // All of the conversions should be optimized away.
14374 EVT VecVT = VT == MVT::f32 ? MVT::v4f32 : MVT::v2f64;
14375 SDValue VOp1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, VecVT, Op1);
14376 SDValue VOp2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, VecVT, Op2);
14377 SDValue VCmp = DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, VecVT, Cmp);
14379 EVT VCmpVT = VT == MVT::f32 ? MVT::v4i32 : MVT::v2i64;
14380 VCmp = DAG.getBitcast(VCmpVT, VCmp);
14382 SDValue VSel = DAG.getNode(ISD::VSELECT, DL, VecVT, VCmp, VOp1, VOp2);
14384 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, VT,
14385 VSel, DAG.getIntPtrConstant(0, DL));
14387 SDValue AndN = DAG.getNode(X86ISD::FANDN, DL, VT, Cmp, Op2);
14388 SDValue And = DAG.getNode(X86ISD::FAND, DL, VT, Cmp, Op1);
14389 return DAG.getNode(X86ISD::FOR, DL, VT, AndN, And);
14393 if (VT.isVector() && VT.getScalarType() == MVT::i1) {
14395 if (ISD::isBuildVectorOfConstantSDNodes(Op1.getNode()))
14396 Op1Scalar = ConvertI1VectorToInteger(Op1, DAG);
14397 else if (Op1.getOpcode() == ISD::BITCAST && Op1.getOperand(0))
14398 Op1Scalar = Op1.getOperand(0);
14400 if (ISD::isBuildVectorOfConstantSDNodes(Op2.getNode()))
14401 Op2Scalar = ConvertI1VectorToInteger(Op2, DAG);
14402 else if (Op2.getOpcode() == ISD::BITCAST && Op2.getOperand(0))
14403 Op2Scalar = Op2.getOperand(0);
14404 if (Op1Scalar.getNode() && Op2Scalar.getNode()) {
14405 SDValue newSelect = DAG.getNode(ISD::SELECT, DL,
14406 Op1Scalar.getValueType(),
14407 Cond, Op1Scalar, Op2Scalar);
14408 if (newSelect.getValueSizeInBits() == VT.getSizeInBits())
14409 return DAG.getBitcast(VT, newSelect);
14410 SDValue ExtVec = DAG.getBitcast(MVT::v8i1, newSelect);
14411 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, VT, ExtVec,
14412 DAG.getIntPtrConstant(0, DL));
14416 if (VT == MVT::v4i1 || VT == MVT::v2i1) {
14417 SDValue zeroConst = DAG.getIntPtrConstant(0, DL);
14418 Op1 = DAG.getNode(ISD::INSERT_SUBVECTOR, DL, MVT::v8i1,
14419 DAG.getUNDEF(MVT::v8i1), Op1, zeroConst);
14420 Op2 = DAG.getNode(ISD::INSERT_SUBVECTOR, DL, MVT::v8i1,
14421 DAG.getUNDEF(MVT::v8i1), Op2, zeroConst);
14422 SDValue newSelect = DAG.getNode(ISD::SELECT, DL, MVT::v8i1,
14424 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, VT, newSelect, zeroConst);
14427 if (Cond.getOpcode() == ISD::SETCC) {
14428 SDValue NewCond = LowerSETCC(Cond, DAG);
14429 if (NewCond.getNode())
14433 // (select (x == 0), -1, y) -> (sign_bit (x - 1)) | y
14434 // (select (x == 0), y, -1) -> ~(sign_bit (x - 1)) | y
14435 // (select (x != 0), y, -1) -> (sign_bit (x - 1)) | y
14436 // (select (x != 0), -1, y) -> ~(sign_bit (x - 1)) | y
14437 if (Cond.getOpcode() == X86ISD::SETCC &&
14438 Cond.getOperand(1).getOpcode() == X86ISD::CMP &&
14439 isZero(Cond.getOperand(1).getOperand(1))) {
14440 SDValue Cmp = Cond.getOperand(1);
14442 unsigned CondCode =cast<ConstantSDNode>(Cond.getOperand(0))->getZExtValue();
14444 if ((isAllOnes(Op1) || isAllOnes(Op2)) &&
14445 (CondCode == X86::COND_E || CondCode == X86::COND_NE)) {
14446 SDValue Y = isAllOnes(Op2) ? Op1 : Op2;
14448 SDValue CmpOp0 = Cmp.getOperand(0);
14449 // Apply further optimizations for special cases
14450 // (select (x != 0), -1, 0) -> neg & sbb
14451 // (select (x == 0), 0, -1) -> neg & sbb
14452 if (ConstantSDNode *YC = dyn_cast<ConstantSDNode>(Y))
14453 if (YC->isNullValue() &&
14454 (isAllOnes(Op1) == (CondCode == X86::COND_NE))) {
14455 SDVTList VTs = DAG.getVTList(CmpOp0.getValueType(), MVT::i32);
14456 SDValue Neg = DAG.getNode(X86ISD::SUB, DL, VTs,
14457 DAG.getConstant(0, DL,
14458 CmpOp0.getValueType()),
14460 SDValue Res = DAG.getNode(X86ISD::SETCC_CARRY, DL, Op.getValueType(),
14461 DAG.getConstant(X86::COND_B, DL, MVT::i8),
14462 SDValue(Neg.getNode(), 1));
14466 Cmp = DAG.getNode(X86ISD::CMP, DL, MVT::i32,
14467 CmpOp0, DAG.getConstant(1, DL, CmpOp0.getValueType()));
14468 Cmp = ConvertCmpIfNecessary(Cmp, DAG);
14470 SDValue Res = // Res = 0 or -1.
14471 DAG.getNode(X86ISD::SETCC_CARRY, DL, Op.getValueType(),
14472 DAG.getConstant(X86::COND_B, DL, MVT::i8), Cmp);
14474 if (isAllOnes(Op1) != (CondCode == X86::COND_E))
14475 Res = DAG.getNOT(DL, Res, Res.getValueType());
14477 ConstantSDNode *N2C = dyn_cast<ConstantSDNode>(Op2);
14478 if (!N2C || !N2C->isNullValue())
14479 Res = DAG.getNode(ISD::OR, DL, Res.getValueType(), Res, Y);
14484 // Look past (and (setcc_carry (cmp ...)), 1).
14485 if (Cond.getOpcode() == ISD::AND &&
14486 Cond.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY) {
14487 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Cond.getOperand(1));
14488 if (C && C->getAPIntValue() == 1)
14489 Cond = Cond.getOperand(0);
14492 // If condition flag is set by a X86ISD::CMP, then use it as the condition
14493 // setting operand in place of the X86ISD::SETCC.
14494 unsigned CondOpcode = Cond.getOpcode();
14495 if (CondOpcode == X86ISD::SETCC ||
14496 CondOpcode == X86ISD::SETCC_CARRY) {
14497 CC = Cond.getOperand(0);
14499 SDValue Cmp = Cond.getOperand(1);
14500 unsigned Opc = Cmp.getOpcode();
14501 MVT VT = Op.getSimpleValueType();
14503 bool IllegalFPCMov = false;
14504 if (VT.isFloatingPoint() && !VT.isVector() &&
14505 !isScalarFPTypeInSSEReg(VT)) // FPStack?
14506 IllegalFPCMov = !hasFPCMov(cast<ConstantSDNode>(CC)->getSExtValue());
14508 if ((isX86LogicalCmp(Cmp) && !IllegalFPCMov) ||
14509 Opc == X86ISD::BT) { // FIXME
14513 } else if (CondOpcode == ISD::USUBO || CondOpcode == ISD::SSUBO ||
14514 CondOpcode == ISD::UADDO || CondOpcode == ISD::SADDO ||
14515 ((CondOpcode == ISD::UMULO || CondOpcode == ISD::SMULO) &&
14516 Cond.getOperand(0).getValueType() != MVT::i8)) {
14517 SDValue LHS = Cond.getOperand(0);
14518 SDValue RHS = Cond.getOperand(1);
14519 unsigned X86Opcode;
14522 switch (CondOpcode) {
14523 case ISD::UADDO: X86Opcode = X86ISD::ADD; X86Cond = X86::COND_B; break;
14524 case ISD::SADDO: X86Opcode = X86ISD::ADD; X86Cond = X86::COND_O; break;
14525 case ISD::USUBO: X86Opcode = X86ISD::SUB; X86Cond = X86::COND_B; break;
14526 case ISD::SSUBO: X86Opcode = X86ISD::SUB; X86Cond = X86::COND_O; break;
14527 case ISD::UMULO: X86Opcode = X86ISD::UMUL; X86Cond = X86::COND_O; break;
14528 case ISD::SMULO: X86Opcode = X86ISD::SMUL; X86Cond = X86::COND_O; break;
14529 default: llvm_unreachable("unexpected overflowing operator");
14531 if (CondOpcode == ISD::UMULO)
14532 VTs = DAG.getVTList(LHS.getValueType(), LHS.getValueType(),
14535 VTs = DAG.getVTList(LHS.getValueType(), MVT::i32);
14537 SDValue X86Op = DAG.getNode(X86Opcode, DL, VTs, LHS, RHS);
14539 if (CondOpcode == ISD::UMULO)
14540 Cond = X86Op.getValue(2);
14542 Cond = X86Op.getValue(1);
14544 CC = DAG.getConstant(X86Cond, DL, MVT::i8);
14549 // Look past the truncate if the high bits are known zero.
14550 if (isTruncWithZeroHighBitsInput(Cond, DAG))
14551 Cond = Cond.getOperand(0);
14553 // We know the result of AND is compared against zero. Try to match
14555 if (Cond.getOpcode() == ISD::AND && Cond.hasOneUse()) {
14556 SDValue NewSetCC = LowerToBT(Cond, ISD::SETNE, DL, DAG);
14557 if (NewSetCC.getNode()) {
14558 CC = NewSetCC.getOperand(0);
14559 Cond = NewSetCC.getOperand(1);
14566 CC = DAG.getConstant(X86::COND_NE, DL, MVT::i8);
14567 Cond = EmitTest(Cond, X86::COND_NE, DL, DAG);
14570 // a < b ? -1 : 0 -> RES = ~setcc_carry
14571 // a < b ? 0 : -1 -> RES = setcc_carry
14572 // a >= b ? -1 : 0 -> RES = setcc_carry
14573 // a >= b ? 0 : -1 -> RES = ~setcc_carry
14574 if (Cond.getOpcode() == X86ISD::SUB) {
14575 Cond = ConvertCmpIfNecessary(Cond, DAG);
14576 unsigned CondCode = cast<ConstantSDNode>(CC)->getZExtValue();
14578 if ((CondCode == X86::COND_AE || CondCode == X86::COND_B) &&
14579 (isAllOnes(Op1) || isAllOnes(Op2)) && (isZero(Op1) || isZero(Op2))) {
14580 SDValue Res = DAG.getNode(X86ISD::SETCC_CARRY, DL, Op.getValueType(),
14581 DAG.getConstant(X86::COND_B, DL, MVT::i8),
14583 if (isAllOnes(Op1) != (CondCode == X86::COND_B))
14584 return DAG.getNOT(DL, Res, Res.getValueType());
14589 // X86 doesn't have an i8 cmov. If both operands are the result of a truncate
14590 // widen the cmov and push the truncate through. This avoids introducing a new
14591 // branch during isel and doesn't add any extensions.
14592 if (Op.getValueType() == MVT::i8 &&
14593 Op1.getOpcode() == ISD::TRUNCATE && Op2.getOpcode() == ISD::TRUNCATE) {
14594 SDValue T1 = Op1.getOperand(0), T2 = Op2.getOperand(0);
14595 if (T1.getValueType() == T2.getValueType() &&
14596 // Blacklist CopyFromReg to avoid partial register stalls.
14597 T1.getOpcode() != ISD::CopyFromReg && T2.getOpcode()!=ISD::CopyFromReg){
14598 SDVTList VTs = DAG.getVTList(T1.getValueType(), MVT::Glue);
14599 SDValue Cmov = DAG.getNode(X86ISD::CMOV, DL, VTs, T2, T1, CC, Cond);
14600 return DAG.getNode(ISD::TRUNCATE, DL, Op.getValueType(), Cmov);
14604 // X86ISD::CMOV means set the result (which is operand 1) to the RHS if
14605 // condition is true.
14606 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::Glue);
14607 SDValue Ops[] = { Op2, Op1, CC, Cond };
14608 return DAG.getNode(X86ISD::CMOV, DL, VTs, Ops);
14611 static SDValue LowerSIGN_EXTEND_AVX512(SDValue Op,
14612 const X86Subtarget *Subtarget,
14613 SelectionDAG &DAG) {
14614 MVT VT = Op->getSimpleValueType(0);
14615 SDValue In = Op->getOperand(0);
14616 MVT InVT = In.getSimpleValueType();
14617 MVT VTElt = VT.getVectorElementType();
14618 MVT InVTElt = InVT.getVectorElementType();
14622 if ((InVTElt == MVT::i1) &&
14623 (((Subtarget->hasBWI() && Subtarget->hasVLX() &&
14624 VT.getSizeInBits() <= 256 && VTElt.getSizeInBits() <= 16)) ||
14626 ((Subtarget->hasBWI() && VT.is512BitVector() &&
14627 VTElt.getSizeInBits() <= 16)) ||
14629 ((Subtarget->hasDQI() && Subtarget->hasVLX() &&
14630 VT.getSizeInBits() <= 256 && VTElt.getSizeInBits() >= 32)) ||
14632 ((Subtarget->hasDQI() && VT.is512BitVector() &&
14633 VTElt.getSizeInBits() >= 32))))
14634 return DAG.getNode(X86ISD::VSEXT, dl, VT, In);
14636 unsigned int NumElts = VT.getVectorNumElements();
14638 if (NumElts != 8 && NumElts != 16 && !Subtarget->hasBWI())
14641 if (VT.is512BitVector() && InVT.getVectorElementType() != MVT::i1) {
14642 if (In.getOpcode() == X86ISD::VSEXT || In.getOpcode() == X86ISD::VZEXT)
14643 return DAG.getNode(In.getOpcode(), dl, VT, In.getOperand(0));
14644 return DAG.getNode(X86ISD::VSEXT, dl, VT, In);
14647 assert (InVT.getVectorElementType() == MVT::i1 && "Unexpected vector type");
14648 MVT ExtVT = NumElts == 8 ? MVT::v8i64 : MVT::v16i32;
14650 DAG.getConstant(APInt::getAllOnesValue(ExtVT.getScalarSizeInBits()), dl,
14653 DAG.getConstant(APInt::getNullValue(ExtVT.getScalarSizeInBits()), dl, ExtVT);
14655 SDValue V = DAG.getNode(ISD::VSELECT, dl, ExtVT, In, NegOne, Zero);
14656 if (VT.is512BitVector())
14658 return DAG.getNode(X86ISD::VTRUNC, dl, VT, V);
14661 static SDValue LowerSIGN_EXTEND_VECTOR_INREG(SDValue Op,
14662 const X86Subtarget *Subtarget,
14663 SelectionDAG &DAG) {
14664 SDValue In = Op->getOperand(0);
14665 MVT VT = Op->getSimpleValueType(0);
14666 MVT InVT = In.getSimpleValueType();
14667 assert(VT.getSizeInBits() == InVT.getSizeInBits());
14669 MVT InSVT = InVT.getScalarType();
14670 assert(VT.getScalarType().getScalarSizeInBits() > InSVT.getScalarSizeInBits());
14672 if (VT != MVT::v2i64 && VT != MVT::v4i32 && VT != MVT::v8i16)
14674 if (InSVT != MVT::i32 && InSVT != MVT::i16 && InSVT != MVT::i8)
14679 // SSE41 targets can use the pmovsx* instructions directly.
14680 if (Subtarget->hasSSE41())
14681 return DAG.getNode(X86ISD::VSEXT, dl, VT, In);
14683 // pre-SSE41 targets unpack lower lanes and then sign-extend using SRAI.
14687 // As SRAI is only available on i16/i32 types, we expand only up to i32
14688 // and handle i64 separately.
14689 while (CurrVT != VT && CurrVT.getScalarType() != MVT::i32) {
14690 Curr = DAG.getNode(X86ISD::UNPCKL, dl, CurrVT, DAG.getUNDEF(CurrVT), Curr);
14691 MVT CurrSVT = MVT::getIntegerVT(CurrVT.getScalarSizeInBits() * 2);
14692 CurrVT = MVT::getVectorVT(CurrSVT, CurrVT.getVectorNumElements() / 2);
14693 Curr = DAG.getBitcast(CurrVT, Curr);
14696 SDValue SignExt = Curr;
14697 if (CurrVT != InVT) {
14698 unsigned SignExtShift =
14699 CurrVT.getScalarSizeInBits() - InSVT.getScalarSizeInBits();
14700 SignExt = DAG.getNode(X86ISD::VSRAI, dl, CurrVT, Curr,
14701 DAG.getConstant(SignExtShift, dl, MVT::i8));
14707 if (VT == MVT::v2i64 && CurrVT == MVT::v4i32) {
14708 SDValue Sign = DAG.getNode(X86ISD::VSRAI, dl, CurrVT, Curr,
14709 DAG.getConstant(31, dl, MVT::i8));
14710 SDValue Ext = DAG.getVectorShuffle(CurrVT, dl, SignExt, Sign, {0, 4, 1, 5});
14711 return DAG.getBitcast(VT, Ext);
14717 static SDValue LowerSIGN_EXTEND(SDValue Op, const X86Subtarget *Subtarget,
14718 SelectionDAG &DAG) {
14719 MVT VT = Op->getSimpleValueType(0);
14720 SDValue In = Op->getOperand(0);
14721 MVT InVT = In.getSimpleValueType();
14724 if (VT.is512BitVector() || InVT.getVectorElementType() == MVT::i1)
14725 return LowerSIGN_EXTEND_AVX512(Op, Subtarget, DAG);
14727 if ((VT != MVT::v4i64 || InVT != MVT::v4i32) &&
14728 (VT != MVT::v8i32 || InVT != MVT::v8i16) &&
14729 (VT != MVT::v16i16 || InVT != MVT::v16i8))
14732 if (Subtarget->hasInt256())
14733 return DAG.getNode(X86ISD::VSEXT, dl, VT, In);
14735 // Optimize vectors in AVX mode
14736 // Sign extend v8i16 to v8i32 and
14739 // Divide input vector into two parts
14740 // for v4i32 the shuffle mask will be { 0, 1, -1, -1} {2, 3, -1, -1}
14741 // use vpmovsx instruction to extend v4i32 -> v2i64; v8i16 -> v4i32
14742 // concat the vectors to original VT
14744 unsigned NumElems = InVT.getVectorNumElements();
14745 SDValue Undef = DAG.getUNDEF(InVT);
14747 SmallVector<int,8> ShufMask1(NumElems, -1);
14748 for (unsigned i = 0; i != NumElems/2; ++i)
14751 SDValue OpLo = DAG.getVectorShuffle(InVT, dl, In, Undef, &ShufMask1[0]);
14753 SmallVector<int,8> ShufMask2(NumElems, -1);
14754 for (unsigned i = 0; i != NumElems/2; ++i)
14755 ShufMask2[i] = i + NumElems/2;
14757 SDValue OpHi = DAG.getVectorShuffle(InVT, dl, In, Undef, &ShufMask2[0]);
14759 MVT HalfVT = MVT::getVectorVT(VT.getScalarType(),
14760 VT.getVectorNumElements()/2);
14762 OpLo = DAG.getNode(X86ISD::VSEXT, dl, HalfVT, OpLo);
14763 OpHi = DAG.getNode(X86ISD::VSEXT, dl, HalfVT, OpHi);
14765 return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT, OpLo, OpHi);
14768 // Lower vector extended loads using a shuffle. If SSSE3 is not available we
14769 // may emit an illegal shuffle but the expansion is still better than scalar
14770 // code. We generate X86ISD::VSEXT for SEXTLOADs if it's available, otherwise
14771 // we'll emit a shuffle and a arithmetic shift.
14772 // FIXME: Is the expansion actually better than scalar code? It doesn't seem so.
14773 // TODO: It is possible to support ZExt by zeroing the undef values during
14774 // the shuffle phase or after the shuffle.
14775 static SDValue LowerExtendedLoad(SDValue Op, const X86Subtarget *Subtarget,
14776 SelectionDAG &DAG) {
14777 MVT RegVT = Op.getSimpleValueType();
14778 assert(RegVT.isVector() && "We only custom lower vector sext loads.");
14779 assert(RegVT.isInteger() &&
14780 "We only custom lower integer vector sext loads.");
14782 // Nothing useful we can do without SSE2 shuffles.
14783 assert(Subtarget->hasSSE2() && "We only custom lower sext loads with SSE2.");
14785 LoadSDNode *Ld = cast<LoadSDNode>(Op.getNode());
14787 EVT MemVT = Ld->getMemoryVT();
14788 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
14789 unsigned RegSz = RegVT.getSizeInBits();
14791 ISD::LoadExtType Ext = Ld->getExtensionType();
14793 assert((Ext == ISD::EXTLOAD || Ext == ISD::SEXTLOAD)
14794 && "Only anyext and sext are currently implemented.");
14795 assert(MemVT != RegVT && "Cannot extend to the same type");
14796 assert(MemVT.isVector() && "Must load a vector from memory");
14798 unsigned NumElems = RegVT.getVectorNumElements();
14799 unsigned MemSz = MemVT.getSizeInBits();
14800 assert(RegSz > MemSz && "Register size must be greater than the mem size");
14802 if (Ext == ISD::SEXTLOAD && RegSz == 256 && !Subtarget->hasInt256()) {
14803 // The only way in which we have a legal 256-bit vector result but not the
14804 // integer 256-bit operations needed to directly lower a sextload is if we
14805 // have AVX1 but not AVX2. In that case, we can always emit a sextload to
14806 // a 128-bit vector and a normal sign_extend to 256-bits that should get
14807 // correctly legalized. We do this late to allow the canonical form of
14808 // sextload to persist throughout the rest of the DAG combiner -- it wants
14809 // to fold together any extensions it can, and so will fuse a sign_extend
14810 // of an sextload into a sextload targeting a wider value.
14812 if (MemSz == 128) {
14813 // Just switch this to a normal load.
14814 assert(TLI.isTypeLegal(MemVT) && "If the memory type is a 128-bit type, "
14815 "it must be a legal 128-bit vector "
14817 Load = DAG.getLoad(MemVT, dl, Ld->getChain(), Ld->getBasePtr(),
14818 Ld->getPointerInfo(), Ld->isVolatile(), Ld->isNonTemporal(),
14819 Ld->isInvariant(), Ld->getAlignment());
14821 assert(MemSz < 128 &&
14822 "Can't extend a type wider than 128 bits to a 256 bit vector!");
14823 // Do an sext load to a 128-bit vector type. We want to use the same
14824 // number of elements, but elements half as wide. This will end up being
14825 // recursively lowered by this routine, but will succeed as we definitely
14826 // have all the necessary features if we're using AVX1.
14828 EVT::getIntegerVT(*DAG.getContext(), RegVT.getScalarSizeInBits() / 2);
14829 EVT HalfVecVT = EVT::getVectorVT(*DAG.getContext(), HalfEltVT, NumElems);
14831 DAG.getExtLoad(Ext, dl, HalfVecVT, Ld->getChain(), Ld->getBasePtr(),
14832 Ld->getPointerInfo(), MemVT, Ld->isVolatile(),
14833 Ld->isNonTemporal(), Ld->isInvariant(),
14834 Ld->getAlignment());
14837 // Replace chain users with the new chain.
14838 assert(Load->getNumValues() == 2 && "Loads must carry a chain!");
14839 DAG.ReplaceAllUsesOfValueWith(SDValue(Ld, 1), Load.getValue(1));
14841 // Finally, do a normal sign-extend to the desired register.
14842 return DAG.getSExtOrTrunc(Load, dl, RegVT);
14845 // All sizes must be a power of two.
14846 assert(isPowerOf2_32(RegSz * MemSz * NumElems) &&
14847 "Non-power-of-two elements are not custom lowered!");
14849 // Attempt to load the original value using scalar loads.
14850 // Find the largest scalar type that divides the total loaded size.
14851 MVT SclrLoadTy = MVT::i8;
14852 for (MVT Tp : MVT::integer_valuetypes()) {
14853 if (TLI.isTypeLegal(Tp) && ((MemSz % Tp.getSizeInBits()) == 0)) {
14858 // On 32bit systems, we can't save 64bit integers. Try bitcasting to F64.
14859 if (TLI.isTypeLegal(MVT::f64) && SclrLoadTy.getSizeInBits() < 64 &&
14861 SclrLoadTy = MVT::f64;
14863 // Calculate the number of scalar loads that we need to perform
14864 // in order to load our vector from memory.
14865 unsigned NumLoads = MemSz / SclrLoadTy.getSizeInBits();
14867 assert((Ext != ISD::SEXTLOAD || NumLoads == 1) &&
14868 "Can only lower sext loads with a single scalar load!");
14870 unsigned loadRegZize = RegSz;
14871 if (Ext == ISD::SEXTLOAD && RegSz >= 256)
14874 // Represent our vector as a sequence of elements which are the
14875 // largest scalar that we can load.
14876 EVT LoadUnitVecVT = EVT::getVectorVT(
14877 *DAG.getContext(), SclrLoadTy, loadRegZize / SclrLoadTy.getSizeInBits());
14879 // Represent the data using the same element type that is stored in
14880 // memory. In practice, we ''widen'' MemVT.
14882 EVT::getVectorVT(*DAG.getContext(), MemVT.getScalarType(),
14883 loadRegZize / MemVT.getScalarType().getSizeInBits());
14885 assert(WideVecVT.getSizeInBits() == LoadUnitVecVT.getSizeInBits() &&
14886 "Invalid vector type");
14888 // We can't shuffle using an illegal type.
14889 assert(TLI.isTypeLegal(WideVecVT) &&
14890 "We only lower types that form legal widened vector types");
14892 SmallVector<SDValue, 8> Chains;
14893 SDValue Ptr = Ld->getBasePtr();
14894 SDValue Increment = DAG.getConstant(SclrLoadTy.getSizeInBits() / 8, dl,
14895 TLI.getPointerTy(DAG.getDataLayout()));
14896 SDValue Res = DAG.getUNDEF(LoadUnitVecVT);
14898 for (unsigned i = 0; i < NumLoads; ++i) {
14899 // Perform a single load.
14900 SDValue ScalarLoad =
14901 DAG.getLoad(SclrLoadTy, dl, Ld->getChain(), Ptr, Ld->getPointerInfo(),
14902 Ld->isVolatile(), Ld->isNonTemporal(), Ld->isInvariant(),
14903 Ld->getAlignment());
14904 Chains.push_back(ScalarLoad.getValue(1));
14905 // Create the first element type using SCALAR_TO_VECTOR in order to avoid
14906 // another round of DAGCombining.
14908 Res = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, LoadUnitVecVT, ScalarLoad);
14910 Res = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, LoadUnitVecVT, Res,
14911 ScalarLoad, DAG.getIntPtrConstant(i, dl));
14913 Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr, Increment);
14916 SDValue TF = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Chains);
14918 // Bitcast the loaded value to a vector of the original element type, in
14919 // the size of the target vector type.
14920 SDValue SlicedVec = DAG.getBitcast(WideVecVT, Res);
14921 unsigned SizeRatio = RegSz / MemSz;
14923 if (Ext == ISD::SEXTLOAD) {
14924 // If we have SSE4.1, we can directly emit a VSEXT node.
14925 if (Subtarget->hasSSE41()) {
14926 SDValue Sext = DAG.getNode(X86ISD::VSEXT, dl, RegVT, SlicedVec);
14927 DAG.ReplaceAllUsesOfValueWith(SDValue(Ld, 1), TF);
14931 // Otherwise we'll shuffle the small elements in the high bits of the
14932 // larger type and perform an arithmetic shift. If the shift is not legal
14933 // it's better to scalarize.
14934 assert(TLI.isOperationLegalOrCustom(ISD::SRA, RegVT) &&
14935 "We can't implement a sext load without an arithmetic right shift!");
14937 // Redistribute the loaded elements into the different locations.
14938 SmallVector<int, 16> ShuffleVec(NumElems * SizeRatio, -1);
14939 for (unsigned i = 0; i != NumElems; ++i)
14940 ShuffleVec[i * SizeRatio + SizeRatio - 1] = i;
14942 SDValue Shuff = DAG.getVectorShuffle(
14943 WideVecVT, dl, SlicedVec, DAG.getUNDEF(WideVecVT), &ShuffleVec[0]);
14945 Shuff = DAG.getBitcast(RegVT, Shuff);
14947 // Build the arithmetic shift.
14948 unsigned Amt = RegVT.getVectorElementType().getSizeInBits() -
14949 MemVT.getVectorElementType().getSizeInBits();
14951 DAG.getNode(ISD::SRA, dl, RegVT, Shuff,
14952 DAG.getConstant(Amt, dl, RegVT));
14954 DAG.ReplaceAllUsesOfValueWith(SDValue(Ld, 1), TF);
14958 // Redistribute the loaded elements into the different locations.
14959 SmallVector<int, 16> ShuffleVec(NumElems * SizeRatio, -1);
14960 for (unsigned i = 0; i != NumElems; ++i)
14961 ShuffleVec[i * SizeRatio] = i;
14963 SDValue Shuff = DAG.getVectorShuffle(WideVecVT, dl, SlicedVec,
14964 DAG.getUNDEF(WideVecVT), &ShuffleVec[0]);
14966 // Bitcast to the requested type.
14967 Shuff = DAG.getBitcast(RegVT, Shuff);
14968 DAG.ReplaceAllUsesOfValueWith(SDValue(Ld, 1), TF);
14972 // isAndOrOfSingleUseSetCCs - Return true if node is an ISD::AND or
14973 // ISD::OR of two X86ISD::SETCC nodes each of which has no other use apart
14974 // from the AND / OR.
14975 static bool isAndOrOfSetCCs(SDValue Op, unsigned &Opc) {
14976 Opc = Op.getOpcode();
14977 if (Opc != ISD::OR && Opc != ISD::AND)
14979 return (Op.getOperand(0).getOpcode() == X86ISD::SETCC &&
14980 Op.getOperand(0).hasOneUse() &&
14981 Op.getOperand(1).getOpcode() == X86ISD::SETCC &&
14982 Op.getOperand(1).hasOneUse());
14985 // isXor1OfSetCC - Return true if node is an ISD::XOR of a X86ISD::SETCC and
14986 // 1 and that the SETCC node has a single use.
14987 static bool isXor1OfSetCC(SDValue Op) {
14988 if (Op.getOpcode() != ISD::XOR)
14990 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(Op.getOperand(1));
14991 if (N1C && N1C->getAPIntValue() == 1) {
14992 return Op.getOperand(0).getOpcode() == X86ISD::SETCC &&
14993 Op.getOperand(0).hasOneUse();
14998 SDValue X86TargetLowering::LowerBRCOND(SDValue Op, SelectionDAG &DAG) const {
14999 bool addTest = true;
15000 SDValue Chain = Op.getOperand(0);
15001 SDValue Cond = Op.getOperand(1);
15002 SDValue Dest = Op.getOperand(2);
15005 bool Inverted = false;
15007 if (Cond.getOpcode() == ISD::SETCC) {
15008 // Check for setcc([su]{add,sub,mul}o == 0).
15009 if (cast<CondCodeSDNode>(Cond.getOperand(2))->get() == ISD::SETEQ &&
15010 isa<ConstantSDNode>(Cond.getOperand(1)) &&
15011 cast<ConstantSDNode>(Cond.getOperand(1))->isNullValue() &&
15012 Cond.getOperand(0).getResNo() == 1 &&
15013 (Cond.getOperand(0).getOpcode() == ISD::SADDO ||
15014 Cond.getOperand(0).getOpcode() == ISD::UADDO ||
15015 Cond.getOperand(0).getOpcode() == ISD::SSUBO ||
15016 Cond.getOperand(0).getOpcode() == ISD::USUBO ||
15017 Cond.getOperand(0).getOpcode() == ISD::SMULO ||
15018 Cond.getOperand(0).getOpcode() == ISD::UMULO)) {
15020 Cond = Cond.getOperand(0);
15022 SDValue NewCond = LowerSETCC(Cond, DAG);
15023 if (NewCond.getNode())
15028 // FIXME: LowerXALUO doesn't handle these!!
15029 else if (Cond.getOpcode() == X86ISD::ADD ||
15030 Cond.getOpcode() == X86ISD::SUB ||
15031 Cond.getOpcode() == X86ISD::SMUL ||
15032 Cond.getOpcode() == X86ISD::UMUL)
15033 Cond = LowerXALUO(Cond, DAG);
15036 // Look pass (and (setcc_carry (cmp ...)), 1).
15037 if (Cond.getOpcode() == ISD::AND &&
15038 Cond.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY) {
15039 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Cond.getOperand(1));
15040 if (C && C->getAPIntValue() == 1)
15041 Cond = Cond.getOperand(0);
15044 // If condition flag is set by a X86ISD::CMP, then use it as the condition
15045 // setting operand in place of the X86ISD::SETCC.
15046 unsigned CondOpcode = Cond.getOpcode();
15047 if (CondOpcode == X86ISD::SETCC ||
15048 CondOpcode == X86ISD::SETCC_CARRY) {
15049 CC = Cond.getOperand(0);
15051 SDValue Cmp = Cond.getOperand(1);
15052 unsigned Opc = Cmp.getOpcode();
15053 // FIXME: WHY THE SPECIAL CASING OF LogicalCmp??
15054 if (isX86LogicalCmp(Cmp) || Opc == X86ISD::BT) {
15058 switch (cast<ConstantSDNode>(CC)->getZExtValue()) {
15062 // These can only come from an arithmetic instruction with overflow,
15063 // e.g. SADDO, UADDO.
15064 Cond = Cond.getNode()->getOperand(1);
15070 CondOpcode = Cond.getOpcode();
15071 if (CondOpcode == ISD::UADDO || CondOpcode == ISD::SADDO ||
15072 CondOpcode == ISD::USUBO || CondOpcode == ISD::SSUBO ||
15073 ((CondOpcode == ISD::UMULO || CondOpcode == ISD::SMULO) &&
15074 Cond.getOperand(0).getValueType() != MVT::i8)) {
15075 SDValue LHS = Cond.getOperand(0);
15076 SDValue RHS = Cond.getOperand(1);
15077 unsigned X86Opcode;
15080 // Keep this in sync with LowerXALUO, otherwise we might create redundant
15081 // instructions that can't be removed afterwards (i.e. X86ISD::ADD and
15083 switch (CondOpcode) {
15084 case ISD::UADDO: X86Opcode = X86ISD::ADD; X86Cond = X86::COND_B; break;
15086 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(RHS))
15088 X86Opcode = X86ISD::INC; X86Cond = X86::COND_O;
15091 X86Opcode = X86ISD::ADD; X86Cond = X86::COND_O; break;
15092 case ISD::USUBO: X86Opcode = X86ISD::SUB; X86Cond = X86::COND_B; break;
15094 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(RHS))
15096 X86Opcode = X86ISD::DEC; X86Cond = X86::COND_O;
15099 X86Opcode = X86ISD::SUB; X86Cond = X86::COND_O; break;
15100 case ISD::UMULO: X86Opcode = X86ISD::UMUL; X86Cond = X86::COND_O; break;
15101 case ISD::SMULO: X86Opcode = X86ISD::SMUL; X86Cond = X86::COND_O; break;
15102 default: llvm_unreachable("unexpected overflowing operator");
15105 X86Cond = X86::GetOppositeBranchCondition((X86::CondCode)X86Cond);
15106 if (CondOpcode == ISD::UMULO)
15107 VTs = DAG.getVTList(LHS.getValueType(), LHS.getValueType(),
15110 VTs = DAG.getVTList(LHS.getValueType(), MVT::i32);
15112 SDValue X86Op = DAG.getNode(X86Opcode, dl, VTs, LHS, RHS);
15114 if (CondOpcode == ISD::UMULO)
15115 Cond = X86Op.getValue(2);
15117 Cond = X86Op.getValue(1);
15119 CC = DAG.getConstant(X86Cond, dl, MVT::i8);
15123 if (Cond.hasOneUse() && isAndOrOfSetCCs(Cond, CondOpc)) {
15124 SDValue Cmp = Cond.getOperand(0).getOperand(1);
15125 if (CondOpc == ISD::OR) {
15126 // Also, recognize the pattern generated by an FCMP_UNE. We can emit
15127 // two branches instead of an explicit OR instruction with a
15129 if (Cmp == Cond.getOperand(1).getOperand(1) &&
15130 isX86LogicalCmp(Cmp)) {
15131 CC = Cond.getOperand(0).getOperand(0);
15132 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
15133 Chain, Dest, CC, Cmp);
15134 CC = Cond.getOperand(1).getOperand(0);
15138 } else { // ISD::AND
15139 // Also, recognize the pattern generated by an FCMP_OEQ. We can emit
15140 // two branches instead of an explicit AND instruction with a
15141 // separate test. However, we only do this if this block doesn't
15142 // have a fall-through edge, because this requires an explicit
15143 // jmp when the condition is false.
15144 if (Cmp == Cond.getOperand(1).getOperand(1) &&
15145 isX86LogicalCmp(Cmp) &&
15146 Op.getNode()->hasOneUse()) {
15147 X86::CondCode CCode =
15148 (X86::CondCode)Cond.getOperand(0).getConstantOperandVal(0);
15149 CCode = X86::GetOppositeBranchCondition(CCode);
15150 CC = DAG.getConstant(CCode, dl, MVT::i8);
15151 SDNode *User = *Op.getNode()->use_begin();
15152 // Look for an unconditional branch following this conditional branch.
15153 // We need this because we need to reverse the successors in order
15154 // to implement FCMP_OEQ.
15155 if (User->getOpcode() == ISD::BR) {
15156 SDValue FalseBB = User->getOperand(1);
15158 DAG.UpdateNodeOperands(User, User->getOperand(0), Dest);
15159 assert(NewBR == User);
15163 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
15164 Chain, Dest, CC, Cmp);
15165 X86::CondCode CCode =
15166 (X86::CondCode)Cond.getOperand(1).getConstantOperandVal(0);
15167 CCode = X86::GetOppositeBranchCondition(CCode);
15168 CC = DAG.getConstant(CCode, dl, MVT::i8);
15174 } else if (Cond.hasOneUse() && isXor1OfSetCC(Cond)) {
15175 // Recognize for xorb (setcc), 1 patterns. The xor inverts the condition.
15176 // It should be transformed during dag combiner except when the condition
15177 // is set by a arithmetics with overflow node.
15178 X86::CondCode CCode =
15179 (X86::CondCode)Cond.getOperand(0).getConstantOperandVal(0);
15180 CCode = X86::GetOppositeBranchCondition(CCode);
15181 CC = DAG.getConstant(CCode, dl, MVT::i8);
15182 Cond = Cond.getOperand(0).getOperand(1);
15184 } else if (Cond.getOpcode() == ISD::SETCC &&
15185 cast<CondCodeSDNode>(Cond.getOperand(2))->get() == ISD::SETOEQ) {
15186 // For FCMP_OEQ, we can emit
15187 // two branches instead of an explicit AND instruction with a
15188 // separate test. However, we only do this if this block doesn't
15189 // have a fall-through edge, because this requires an explicit
15190 // jmp when the condition is false.
15191 if (Op.getNode()->hasOneUse()) {
15192 SDNode *User = *Op.getNode()->use_begin();
15193 // Look for an unconditional branch following this conditional branch.
15194 // We need this because we need to reverse the successors in order
15195 // to implement FCMP_OEQ.
15196 if (User->getOpcode() == ISD::BR) {
15197 SDValue FalseBB = User->getOperand(1);
15199 DAG.UpdateNodeOperands(User, User->getOperand(0), Dest);
15200 assert(NewBR == User);
15204 SDValue Cmp = DAG.getNode(X86ISD::CMP, dl, MVT::i32,
15205 Cond.getOperand(0), Cond.getOperand(1));
15206 Cmp = ConvertCmpIfNecessary(Cmp, DAG);
15207 CC = DAG.getConstant(X86::COND_NE, dl, MVT::i8);
15208 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
15209 Chain, Dest, CC, Cmp);
15210 CC = DAG.getConstant(X86::COND_P, dl, MVT::i8);
15215 } else if (Cond.getOpcode() == ISD::SETCC &&
15216 cast<CondCodeSDNode>(Cond.getOperand(2))->get() == ISD::SETUNE) {
15217 // For FCMP_UNE, we can emit
15218 // two branches instead of an explicit AND instruction with a
15219 // separate test. However, we only do this if this block doesn't
15220 // have a fall-through edge, because this requires an explicit
15221 // jmp when the condition is false.
15222 if (Op.getNode()->hasOneUse()) {
15223 SDNode *User = *Op.getNode()->use_begin();
15224 // Look for an unconditional branch following this conditional branch.
15225 // We need this because we need to reverse the successors in order
15226 // to implement FCMP_UNE.
15227 if (User->getOpcode() == ISD::BR) {
15228 SDValue FalseBB = User->getOperand(1);
15230 DAG.UpdateNodeOperands(User, User->getOperand(0), Dest);
15231 assert(NewBR == User);
15234 SDValue Cmp = DAG.getNode(X86ISD::CMP, dl, MVT::i32,
15235 Cond.getOperand(0), Cond.getOperand(1));
15236 Cmp = ConvertCmpIfNecessary(Cmp, DAG);
15237 CC = DAG.getConstant(X86::COND_NE, dl, MVT::i8);
15238 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
15239 Chain, Dest, CC, Cmp);
15240 CC = DAG.getConstant(X86::COND_NP, dl, MVT::i8);
15250 // Look pass the truncate if the high bits are known zero.
15251 if (isTruncWithZeroHighBitsInput(Cond, DAG))
15252 Cond = Cond.getOperand(0);
15254 // We know the result of AND is compared against zero. Try to match
15256 if (Cond.getOpcode() == ISD::AND && Cond.hasOneUse()) {
15257 SDValue NewSetCC = LowerToBT(Cond, ISD::SETNE, dl, DAG);
15258 if (NewSetCC.getNode()) {
15259 CC = NewSetCC.getOperand(0);
15260 Cond = NewSetCC.getOperand(1);
15267 X86::CondCode X86Cond = Inverted ? X86::COND_E : X86::COND_NE;
15268 CC = DAG.getConstant(X86Cond, dl, MVT::i8);
15269 Cond = EmitTest(Cond, X86Cond, dl, DAG);
15271 Cond = ConvertCmpIfNecessary(Cond, DAG);
15272 return DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
15273 Chain, Dest, CC, Cond);
15276 // Lower dynamic stack allocation to _alloca call for Cygwin/Mingw targets.
15277 // Calls to _alloca are needed to probe the stack when allocating more than 4k
15278 // bytes in one go. Touching the stack at 4K increments is necessary to ensure
15279 // that the guard pages used by the OS virtual memory manager are allocated in
15280 // correct sequence.
15282 X86TargetLowering::LowerDYNAMIC_STACKALLOC(SDValue Op,
15283 SelectionDAG &DAG) const {
15284 MachineFunction &MF = DAG.getMachineFunction();
15285 bool SplitStack = MF.shouldSplitStack();
15286 bool Lower = (Subtarget->isOSWindows() && !Subtarget->isTargetMachO()) ||
15291 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
15292 SDNode* Node = Op.getNode();
15294 unsigned SPReg = TLI.getStackPointerRegisterToSaveRestore();
15295 assert(SPReg && "Target cannot require DYNAMIC_STACKALLOC expansion and"
15296 " not tell us which reg is the stack pointer!");
15297 EVT VT = Node->getValueType(0);
15298 SDValue Tmp1 = SDValue(Node, 0);
15299 SDValue Tmp2 = SDValue(Node, 1);
15300 SDValue Tmp3 = Node->getOperand(2);
15301 SDValue Chain = Tmp1.getOperand(0);
15303 // Chain the dynamic stack allocation so that it doesn't modify the stack
15304 // pointer when other instructions are using the stack.
15305 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(0, dl, true),
15308 SDValue Size = Tmp2.getOperand(1);
15309 SDValue SP = DAG.getCopyFromReg(Chain, dl, SPReg, VT);
15310 Chain = SP.getValue(1);
15311 unsigned Align = cast<ConstantSDNode>(Tmp3)->getZExtValue();
15312 const TargetFrameLowering &TFI = *Subtarget->getFrameLowering();
15313 unsigned StackAlign = TFI.getStackAlignment();
15314 Tmp1 = DAG.getNode(ISD::SUB, dl, VT, SP, Size); // Value
15315 if (Align > StackAlign)
15316 Tmp1 = DAG.getNode(ISD::AND, dl, VT, Tmp1,
15317 DAG.getConstant(-(uint64_t)Align, dl, VT));
15318 Chain = DAG.getCopyToReg(Chain, dl, SPReg, Tmp1); // Output chain
15320 Tmp2 = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(0, dl, true),
15321 DAG.getIntPtrConstant(0, dl, true), SDValue(),
15324 SDValue Ops[2] = { Tmp1, Tmp2 };
15325 return DAG.getMergeValues(Ops, dl);
15329 SDValue Chain = Op.getOperand(0);
15330 SDValue Size = Op.getOperand(1);
15331 unsigned Align = cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue();
15332 EVT VT = Op.getNode()->getValueType(0);
15334 bool Is64Bit = Subtarget->is64Bit();
15335 MVT SPTy = getPointerTy(DAG.getDataLayout());
15338 MachineRegisterInfo &MRI = MF.getRegInfo();
15341 // The 64 bit implementation of segmented stacks needs to clobber both r10
15342 // r11. This makes it impossible to use it along with nested parameters.
15343 const Function *F = MF.getFunction();
15345 for (Function::const_arg_iterator I = F->arg_begin(), E = F->arg_end();
15347 if (I->hasNestAttr())
15348 report_fatal_error("Cannot use segmented stacks with functions that "
15349 "have nested arguments.");
15352 const TargetRegisterClass *AddrRegClass = getRegClassFor(SPTy);
15353 unsigned Vreg = MRI.createVirtualRegister(AddrRegClass);
15354 Chain = DAG.getCopyToReg(Chain, dl, Vreg, Size);
15355 SDValue Value = DAG.getNode(X86ISD::SEG_ALLOCA, dl, SPTy, Chain,
15356 DAG.getRegister(Vreg, SPTy));
15357 SDValue Ops1[2] = { Value, Chain };
15358 return DAG.getMergeValues(Ops1, dl);
15361 const unsigned Reg = (Subtarget->isTarget64BitLP64() ? X86::RAX : X86::EAX);
15363 Chain = DAG.getCopyToReg(Chain, dl, Reg, Size, Flag);
15364 Flag = Chain.getValue(1);
15365 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
15367 Chain = DAG.getNode(X86ISD::WIN_ALLOCA, dl, NodeTys, Chain, Flag);
15369 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
15370 unsigned SPReg = RegInfo->getStackRegister();
15371 SDValue SP = DAG.getCopyFromReg(Chain, dl, SPReg, SPTy);
15372 Chain = SP.getValue(1);
15375 SP = DAG.getNode(ISD::AND, dl, VT, SP.getValue(0),
15376 DAG.getConstant(-(uint64_t)Align, dl, VT));
15377 Chain = DAG.getCopyToReg(Chain, dl, SPReg, SP);
15380 SDValue Ops1[2] = { SP, Chain };
15381 return DAG.getMergeValues(Ops1, dl);
15385 SDValue X86TargetLowering::LowerVASTART(SDValue Op, SelectionDAG &DAG) const {
15386 MachineFunction &MF = DAG.getMachineFunction();
15387 auto PtrVT = getPointerTy(MF.getDataLayout());
15388 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
15390 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
15393 if (!Subtarget->is64Bit() ||
15394 Subtarget->isCallingConvWin64(MF.getFunction()->getCallingConv())) {
15395 // vastart just stores the address of the VarArgsFrameIndex slot into the
15396 // memory location argument.
15397 SDValue FR = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(), PtrVT);
15398 return DAG.getStore(Op.getOperand(0), DL, FR, Op.getOperand(1),
15399 MachinePointerInfo(SV), false, false, 0);
15403 // gp_offset (0 - 6 * 8)
15404 // fp_offset (48 - 48 + 8 * 16)
15405 // overflow_arg_area (point to parameters coming in memory).
15407 SmallVector<SDValue, 8> MemOps;
15408 SDValue FIN = Op.getOperand(1);
15410 SDValue Store = DAG.getStore(Op.getOperand(0), DL,
15411 DAG.getConstant(FuncInfo->getVarArgsGPOffset(),
15413 FIN, MachinePointerInfo(SV), false, false, 0);
15414 MemOps.push_back(Store);
15417 FIN = DAG.getNode(ISD::ADD, DL, PtrVT, FIN, DAG.getIntPtrConstant(4, DL));
15418 Store = DAG.getStore(Op.getOperand(0), DL,
15419 DAG.getConstant(FuncInfo->getVarArgsFPOffset(), DL,
15421 FIN, MachinePointerInfo(SV, 4), false, false, 0);
15422 MemOps.push_back(Store);
15424 // Store ptr to overflow_arg_area
15425 FIN = DAG.getNode(ISD::ADD, DL, PtrVT, FIN, DAG.getIntPtrConstant(4, DL));
15426 SDValue OVFIN = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(), PtrVT);
15427 Store = DAG.getStore(Op.getOperand(0), DL, OVFIN, FIN,
15428 MachinePointerInfo(SV, 8),
15430 MemOps.push_back(Store);
15432 // Store ptr to reg_save_area.
15433 FIN = DAG.getNode(ISD::ADD, DL, PtrVT, FIN, DAG.getIntPtrConstant(
15434 Subtarget->isTarget64BitLP64() ? 8 : 4, DL));
15435 SDValue RSFIN = DAG.getFrameIndex(FuncInfo->getRegSaveFrameIndex(), PtrVT);
15436 Store = DAG.getStore(Op.getOperand(0), DL, RSFIN, FIN, MachinePointerInfo(
15437 SV, Subtarget->isTarget64BitLP64() ? 16 : 12), false, false, 0);
15438 MemOps.push_back(Store);
15439 return DAG.getNode(ISD::TokenFactor, DL, MVT::Other, MemOps);
15442 SDValue X86TargetLowering::LowerVAARG(SDValue Op, SelectionDAG &DAG) const {
15443 assert(Subtarget->is64Bit() &&
15444 "LowerVAARG only handles 64-bit va_arg!");
15445 assert(Op.getNode()->getNumOperands() == 4);
15447 MachineFunction &MF = DAG.getMachineFunction();
15448 if (Subtarget->isCallingConvWin64(MF.getFunction()->getCallingConv()))
15449 // The Win64 ABI uses char* instead of a structure.
15450 return DAG.expandVAArg(Op.getNode());
15452 SDValue Chain = Op.getOperand(0);
15453 SDValue SrcPtr = Op.getOperand(1);
15454 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
15455 unsigned Align = Op.getConstantOperandVal(3);
15458 EVT ArgVT = Op.getNode()->getValueType(0);
15459 Type *ArgTy = ArgVT.getTypeForEVT(*DAG.getContext());
15460 uint32_t ArgSize = DAG.getDataLayout().getTypeAllocSize(ArgTy);
15463 // Decide which area this value should be read from.
15464 // TODO: Implement the AMD64 ABI in its entirety. This simple
15465 // selection mechanism works only for the basic types.
15466 if (ArgVT == MVT::f80) {
15467 llvm_unreachable("va_arg for f80 not yet implemented");
15468 } else if (ArgVT.isFloatingPoint() && ArgSize <= 16 /*bytes*/) {
15469 ArgMode = 2; // Argument passed in XMM register. Use fp_offset.
15470 } else if (ArgVT.isInteger() && ArgSize <= 32 /*bytes*/) {
15471 ArgMode = 1; // Argument passed in GPR64 register(s). Use gp_offset.
15473 llvm_unreachable("Unhandled argument type in LowerVAARG");
15476 if (ArgMode == 2) {
15477 // Sanity Check: Make sure using fp_offset makes sense.
15478 assert(!Subtarget->useSoftFloat() &&
15479 !(MF.getFunction()->hasFnAttribute(Attribute::NoImplicitFloat)) &&
15480 Subtarget->hasSSE1());
15483 // Insert VAARG_64 node into the DAG
15484 // VAARG_64 returns two values: Variable Argument Address, Chain
15485 SDValue InstOps[] = {Chain, SrcPtr, DAG.getConstant(ArgSize, dl, MVT::i32),
15486 DAG.getConstant(ArgMode, dl, MVT::i8),
15487 DAG.getConstant(Align, dl, MVT::i32)};
15488 SDVTList VTs = DAG.getVTList(getPointerTy(DAG.getDataLayout()), MVT::Other);
15489 SDValue VAARG = DAG.getMemIntrinsicNode(X86ISD::VAARG_64, dl,
15490 VTs, InstOps, MVT::i64,
15491 MachinePointerInfo(SV),
15493 /*Volatile=*/false,
15495 /*WriteMem=*/true);
15496 Chain = VAARG.getValue(1);
15498 // Load the next argument and return it
15499 return DAG.getLoad(ArgVT, dl,
15502 MachinePointerInfo(),
15503 false, false, false, 0);
15506 static SDValue LowerVACOPY(SDValue Op, const X86Subtarget *Subtarget,
15507 SelectionDAG &DAG) {
15508 // X86-64 va_list is a struct { i32, i32, i8*, i8* }, except on Windows,
15509 // where a va_list is still an i8*.
15510 assert(Subtarget->is64Bit() && "This code only handles 64-bit va_copy!");
15511 if (Subtarget->isCallingConvWin64(
15512 DAG.getMachineFunction().getFunction()->getCallingConv()))
15513 // Probably a Win64 va_copy.
15514 return DAG.expandVACopy(Op.getNode());
15516 SDValue Chain = Op.getOperand(0);
15517 SDValue DstPtr = Op.getOperand(1);
15518 SDValue SrcPtr = Op.getOperand(2);
15519 const Value *DstSV = cast<SrcValueSDNode>(Op.getOperand(3))->getValue();
15520 const Value *SrcSV = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
15523 return DAG.getMemcpy(Chain, DL, DstPtr, SrcPtr,
15524 DAG.getIntPtrConstant(24, DL), 8, /*isVolatile*/false,
15526 MachinePointerInfo(DstSV), MachinePointerInfo(SrcSV));
15529 // getTargetVShiftByConstNode - Handle vector element shifts where the shift
15530 // amount is a constant. Takes immediate version of shift as input.
15531 static SDValue getTargetVShiftByConstNode(unsigned Opc, SDLoc dl, MVT VT,
15532 SDValue SrcOp, uint64_t ShiftAmt,
15533 SelectionDAG &DAG) {
15534 MVT ElementType = VT.getVectorElementType();
15536 // Fold this packed shift into its first operand if ShiftAmt is 0.
15540 // Check for ShiftAmt >= element width
15541 if (ShiftAmt >= ElementType.getSizeInBits()) {
15542 if (Opc == X86ISD::VSRAI)
15543 ShiftAmt = ElementType.getSizeInBits() - 1;
15545 return DAG.getConstant(0, dl, VT);
15548 assert((Opc == X86ISD::VSHLI || Opc == X86ISD::VSRLI || Opc == X86ISD::VSRAI)
15549 && "Unknown target vector shift-by-constant node");
15551 // Fold this packed vector shift into a build vector if SrcOp is a
15552 // vector of Constants or UNDEFs, and SrcOp valuetype is the same as VT.
15553 if (VT == SrcOp.getSimpleValueType() &&
15554 ISD::isBuildVectorOfConstantSDNodes(SrcOp.getNode())) {
15555 SmallVector<SDValue, 8> Elts;
15556 unsigned NumElts = SrcOp->getNumOperands();
15557 ConstantSDNode *ND;
15560 default: llvm_unreachable(nullptr);
15561 case X86ISD::VSHLI:
15562 for (unsigned i=0; i!=NumElts; ++i) {
15563 SDValue CurrentOp = SrcOp->getOperand(i);
15564 if (CurrentOp->getOpcode() == ISD::UNDEF) {
15565 Elts.push_back(CurrentOp);
15568 ND = cast<ConstantSDNode>(CurrentOp);
15569 const APInt &C = ND->getAPIntValue();
15570 Elts.push_back(DAG.getConstant(C.shl(ShiftAmt), dl, ElementType));
15573 case X86ISD::VSRLI:
15574 for (unsigned i=0; i!=NumElts; ++i) {
15575 SDValue CurrentOp = SrcOp->getOperand(i);
15576 if (CurrentOp->getOpcode() == ISD::UNDEF) {
15577 Elts.push_back(CurrentOp);
15580 ND = cast<ConstantSDNode>(CurrentOp);
15581 const APInt &C = ND->getAPIntValue();
15582 Elts.push_back(DAG.getConstant(C.lshr(ShiftAmt), dl, ElementType));
15585 case X86ISD::VSRAI:
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.ashr(ShiftAmt), dl, ElementType));
15599 return DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Elts);
15602 return DAG.getNode(Opc, dl, VT, SrcOp,
15603 DAG.getConstant(ShiftAmt, dl, MVT::i8));
15606 // getTargetVShiftNode - Handle vector element shifts where the shift amount
15607 // may or may not be a constant. Takes immediate version of shift as input.
15608 static SDValue getTargetVShiftNode(unsigned Opc, SDLoc dl, MVT VT,
15609 SDValue SrcOp, SDValue ShAmt,
15610 SelectionDAG &DAG) {
15611 MVT SVT = ShAmt.getSimpleValueType();
15612 assert((SVT == MVT::i32 || SVT == MVT::i64) && "Unexpected value type!");
15614 // Catch shift-by-constant.
15615 if (ConstantSDNode *CShAmt = dyn_cast<ConstantSDNode>(ShAmt))
15616 return getTargetVShiftByConstNode(Opc, dl, VT, SrcOp,
15617 CShAmt->getZExtValue(), DAG);
15619 // Change opcode to non-immediate version
15621 default: llvm_unreachable("Unknown target vector shift node");
15622 case X86ISD::VSHLI: Opc = X86ISD::VSHL; break;
15623 case X86ISD::VSRLI: Opc = X86ISD::VSRL; break;
15624 case X86ISD::VSRAI: Opc = X86ISD::VSRA; break;
15627 const X86Subtarget &Subtarget =
15628 static_cast<const X86Subtarget &>(DAG.getSubtarget());
15629 if (Subtarget.hasSSE41() && ShAmt.getOpcode() == ISD::ZERO_EXTEND &&
15630 ShAmt.getOperand(0).getSimpleValueType() == MVT::i16) {
15631 // Let the shuffle legalizer expand this shift amount node.
15632 SDValue Op0 = ShAmt.getOperand(0);
15633 Op0 = DAG.getNode(ISD::SCALAR_TO_VECTOR, SDLoc(Op0), MVT::v8i16, Op0);
15634 ShAmt = getShuffleVectorZeroOrUndef(Op0, 0, true, &Subtarget, DAG);
15636 // Need to build a vector containing shift amount.
15637 // SSE/AVX packed shifts only use the lower 64-bit of the shift count.
15638 SmallVector<SDValue, 4> ShOps;
15639 ShOps.push_back(ShAmt);
15640 if (SVT == MVT::i32) {
15641 ShOps.push_back(DAG.getConstant(0, dl, SVT));
15642 ShOps.push_back(DAG.getUNDEF(SVT));
15644 ShOps.push_back(DAG.getUNDEF(SVT));
15646 MVT BVT = SVT == MVT::i32 ? MVT::v4i32 : MVT::v2i64;
15647 ShAmt = DAG.getNode(ISD::BUILD_VECTOR, dl, BVT, ShOps);
15650 // The return type has to be a 128-bit type with the same element
15651 // type as the input type.
15652 MVT EltVT = VT.getVectorElementType();
15653 EVT ShVT = MVT::getVectorVT(EltVT, 128/EltVT.getSizeInBits());
15655 ShAmt = DAG.getBitcast(ShVT, ShAmt);
15656 return DAG.getNode(Opc, dl, VT, SrcOp, ShAmt);
15659 /// \brief Return (and \p Op, \p Mask) for compare instructions or
15660 /// (vselect \p Mask, \p Op, \p PreservedSrc) for others along with the
15661 /// necessary casting or extending for \p Mask when lowering masking intrinsics
15662 static SDValue getVectorMaskingNode(SDValue Op, SDValue Mask,
15663 SDValue PreservedSrc,
15664 const X86Subtarget *Subtarget,
15665 SelectionDAG &DAG) {
15666 EVT VT = Op.getValueType();
15667 EVT MaskVT = EVT::getVectorVT(*DAG.getContext(),
15668 MVT::i1, VT.getVectorNumElements());
15669 SDValue VMask = SDValue();
15670 unsigned OpcodeSelect = ISD::VSELECT;
15673 assert(MaskVT.isSimple() && "invalid mask type");
15675 if (isAllOnes(Mask))
15678 if (MaskVT.bitsGT(Mask.getValueType())) {
15679 EVT newMaskVT = EVT::getIntegerVT(*DAG.getContext(),
15680 MaskVT.getSizeInBits());
15681 VMask = DAG.getBitcast(MaskVT,
15682 DAG.getNode(ISD::ANY_EXTEND, dl, newMaskVT, Mask));
15684 EVT BitcastVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
15685 Mask.getValueType().getSizeInBits());
15686 // In case when MaskVT equals v2i1 or v4i1, low 2 or 4 elements
15687 // are extracted by EXTRACT_SUBVECTOR.
15688 VMask = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MaskVT,
15689 DAG.getBitcast(BitcastVT, Mask),
15690 DAG.getIntPtrConstant(0, dl));
15693 switch (Op.getOpcode()) {
15695 case X86ISD::PCMPEQM:
15696 case X86ISD::PCMPGTM:
15698 case X86ISD::CMPMU:
15699 return DAG.getNode(ISD::AND, dl, VT, Op, VMask);
15700 case X86ISD::VTRUNC:
15701 case X86ISD::VTRUNCS:
15702 case X86ISD::VTRUNCUS:
15703 // We can't use ISD::VSELECT here because it is not always "Legal"
15704 // for the destination type. For example vpmovqb require only AVX512
15705 // and vselect that can operate on byte element type require BWI
15706 OpcodeSelect = X86ISD::SELECT;
15709 if (PreservedSrc.getOpcode() == ISD::UNDEF)
15710 PreservedSrc = getZeroVector(VT, Subtarget, DAG, dl);
15711 return DAG.getNode(OpcodeSelect, dl, VT, VMask, Op, PreservedSrc);
15714 /// \brief Creates an SDNode for a predicated scalar operation.
15715 /// \returns (X86vselect \p Mask, \p Op, \p PreservedSrc).
15716 /// The mask is coming as MVT::i8 and it should be truncated
15717 /// to MVT::i1 while lowering masking intrinsics.
15718 /// The main difference between ScalarMaskingNode and VectorMaskingNode is using
15719 /// "X86select" instead of "vselect". We just can't create the "vselect" node
15720 /// for a scalar instruction.
15721 static SDValue getScalarMaskingNode(SDValue Op, SDValue Mask,
15722 SDValue PreservedSrc,
15723 const X86Subtarget *Subtarget,
15724 SelectionDAG &DAG) {
15725 if (isAllOnes(Mask))
15728 EVT VT = Op.getValueType();
15730 // The mask should be of type MVT::i1
15731 SDValue IMask = DAG.getNode(ISD::TRUNCATE, dl, MVT::i1, Mask);
15733 if (PreservedSrc.getOpcode() == ISD::UNDEF)
15734 PreservedSrc = getZeroVector(VT, Subtarget, DAG, dl);
15735 return DAG.getNode(X86ISD::SELECT, dl, VT, IMask, Op, PreservedSrc);
15738 static int getSEHRegistrationNodeSize(const Function *Fn) {
15739 if (!Fn->hasPersonalityFn())
15740 report_fatal_error(
15741 "querying registration node size for function without personality");
15742 // The RegNodeSize is 6 32-bit words for SEH and 4 for C++ EH. See
15743 // WinEHStatePass for the full struct definition.
15744 switch (classifyEHPersonality(Fn->getPersonalityFn())) {
15745 case EHPersonality::MSVC_X86SEH: return 24;
15746 case EHPersonality::MSVC_CXX: return 16;
15749 report_fatal_error("can only recover FP for MSVC EH personality functions");
15752 /// When the 32-bit MSVC runtime transfers control to us, either to an outlined
15753 /// function or when returning to a parent frame after catching an exception, we
15754 /// recover the parent frame pointer by doing arithmetic on the incoming EBP.
15755 /// Here's the math:
15756 /// RegNodeBase = EntryEBP - RegNodeSize
15757 /// ParentFP = RegNodeBase - RegNodeFrameOffset
15758 /// Subtracting RegNodeSize takes us to the offset of the registration node, and
15759 /// subtracting the offset (negative on x86) takes us back to the parent FP.
15760 static SDValue recoverFramePointer(SelectionDAG &DAG, const Function *Fn,
15761 SDValue EntryEBP) {
15762 MachineFunction &MF = DAG.getMachineFunction();
15765 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
15766 MVT PtrVT = TLI.getPointerTy(DAG.getDataLayout());
15768 // It's possible that the parent function no longer has a personality function
15769 // if the exceptional code was optimized away, in which case we just return
15770 // the incoming EBP.
15771 if (!Fn->hasPersonalityFn())
15774 int RegNodeSize = getSEHRegistrationNodeSize(Fn);
15776 // Get an MCSymbol that will ultimately resolve to the frame offset of the EH
15778 MCSymbol *OffsetSym =
15779 MF.getMMI().getContext().getOrCreateParentFrameOffsetSymbol(
15780 GlobalValue::getRealLinkageName(Fn->getName()));
15781 SDValue OffsetSymVal = DAG.getMCSymbol(OffsetSym, PtrVT);
15782 SDValue RegNodeFrameOffset =
15783 DAG.getNode(ISD::LOCAL_RECOVER, dl, PtrVT, OffsetSymVal);
15785 // RegNodeBase = EntryEBP - RegNodeSize
15786 // ParentFP = RegNodeBase - RegNodeFrameOffset
15787 SDValue RegNodeBase = DAG.getNode(ISD::SUB, dl, PtrVT, EntryEBP,
15788 DAG.getConstant(RegNodeSize, dl, PtrVT));
15789 return DAG.getNode(ISD::SUB, dl, PtrVT, RegNodeBase, RegNodeFrameOffset);
15792 static SDValue LowerINTRINSIC_WO_CHAIN(SDValue Op, const X86Subtarget *Subtarget,
15793 SelectionDAG &DAG) {
15795 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
15796 EVT VT = Op.getValueType();
15797 const IntrinsicData* IntrData = getIntrinsicWithoutChain(IntNo);
15799 switch(IntrData->Type) {
15800 case INTR_TYPE_1OP:
15801 return DAG.getNode(IntrData->Opc0, dl, Op.getValueType(), Op.getOperand(1));
15802 case INTR_TYPE_2OP:
15803 return DAG.getNode(IntrData->Opc0, dl, Op.getValueType(), Op.getOperand(1),
15805 case INTR_TYPE_2OP_IMM8:
15806 return DAG.getNode(IntrData->Opc0, dl, Op.getValueType(), Op.getOperand(1),
15807 DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op.getOperand(2)));
15808 case INTR_TYPE_3OP:
15809 return DAG.getNode(IntrData->Opc0, dl, Op.getValueType(), Op.getOperand(1),
15810 Op.getOperand(2), Op.getOperand(3));
15811 case INTR_TYPE_4OP:
15812 return DAG.getNode(IntrData->Opc0, dl, Op.getValueType(), Op.getOperand(1),
15813 Op.getOperand(2), Op.getOperand(3), Op.getOperand(4));
15814 case INTR_TYPE_1OP_MASK_RM: {
15815 SDValue Src = Op.getOperand(1);
15816 SDValue PassThru = Op.getOperand(2);
15817 SDValue Mask = Op.getOperand(3);
15818 SDValue RoundingMode;
15819 // We allways add rounding mode to the Node.
15820 // If the rounding mode is not specified, we add the
15821 // "current direction" mode.
15822 if (Op.getNumOperands() == 4)
15824 DAG.getConstant(X86::STATIC_ROUNDING::CUR_DIRECTION, dl, MVT::i32);
15826 RoundingMode = Op.getOperand(4);
15827 unsigned IntrWithRoundingModeOpcode = IntrData->Opc1;
15828 if (IntrWithRoundingModeOpcode != 0)
15829 if (cast<ConstantSDNode>(RoundingMode)->getZExtValue() !=
15830 X86::STATIC_ROUNDING::CUR_DIRECTION)
15831 return getVectorMaskingNode(DAG.getNode(IntrWithRoundingModeOpcode,
15832 dl, Op.getValueType(), Src, RoundingMode),
15833 Mask, PassThru, Subtarget, DAG);
15834 return getVectorMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT, Src,
15836 Mask, PassThru, Subtarget, DAG);
15838 case INTR_TYPE_1OP_MASK: {
15839 SDValue Src = Op.getOperand(1);
15840 SDValue PassThru = Op.getOperand(2);
15841 SDValue Mask = Op.getOperand(3);
15842 // We add rounding mode to the Node when
15843 // - RM Opcode is specified and
15844 // - RM is not "current direction".
15845 unsigned IntrWithRoundingModeOpcode = IntrData->Opc1;
15846 if (IntrWithRoundingModeOpcode != 0) {
15847 SDValue Rnd = Op.getOperand(4);
15848 unsigned Round = cast<ConstantSDNode>(Rnd)->getZExtValue();
15849 if (Round != X86::STATIC_ROUNDING::CUR_DIRECTION) {
15850 return getVectorMaskingNode(DAG.getNode(IntrWithRoundingModeOpcode,
15851 dl, Op.getValueType(),
15853 Mask, PassThru, Subtarget, DAG);
15856 return getVectorMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT, Src),
15857 Mask, PassThru, Subtarget, DAG);
15859 case INTR_TYPE_SCALAR_MASK_RM: {
15860 SDValue Src1 = Op.getOperand(1);
15861 SDValue Src2 = Op.getOperand(2);
15862 SDValue Src0 = Op.getOperand(3);
15863 SDValue Mask = Op.getOperand(4);
15864 // There are 2 kinds of intrinsics in this group:
15865 // (1) With suppress-all-exceptions (sae) or rounding mode- 6 operands
15866 // (2) With rounding mode and sae - 7 operands.
15867 if (Op.getNumOperands() == 6) {
15868 SDValue Sae = Op.getOperand(5);
15869 unsigned Opc = IntrData->Opc1 ? IntrData->Opc1 : IntrData->Opc0;
15870 return getScalarMaskingNode(DAG.getNode(Opc, dl, VT, Src1, Src2,
15872 Mask, Src0, Subtarget, DAG);
15874 assert(Op.getNumOperands() == 7 && "Unexpected intrinsic form");
15875 SDValue RoundingMode = Op.getOperand(5);
15876 SDValue Sae = Op.getOperand(6);
15877 return getScalarMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT, Src1, Src2,
15878 RoundingMode, Sae),
15879 Mask, Src0, Subtarget, DAG);
15881 case INTR_TYPE_2OP_MASK: {
15882 SDValue Src1 = Op.getOperand(1);
15883 SDValue Src2 = Op.getOperand(2);
15884 SDValue PassThru = Op.getOperand(3);
15885 SDValue Mask = Op.getOperand(4);
15886 // We specify 2 possible opcodes for intrinsics with rounding modes.
15887 // First, we check if the intrinsic may have non-default rounding mode,
15888 // (IntrData->Opc1 != 0), then we check the rounding mode operand.
15889 unsigned IntrWithRoundingModeOpcode = IntrData->Opc1;
15890 if (IntrWithRoundingModeOpcode != 0) {
15891 SDValue Rnd = Op.getOperand(5);
15892 unsigned Round = cast<ConstantSDNode>(Rnd)->getZExtValue();
15893 if (Round != X86::STATIC_ROUNDING::CUR_DIRECTION) {
15894 return getVectorMaskingNode(DAG.getNode(IntrWithRoundingModeOpcode,
15895 dl, Op.getValueType(),
15897 Mask, PassThru, Subtarget, DAG);
15900 // TODO: Intrinsics should have fast-math-flags to propagate.
15901 return getVectorMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT,Src1,Src2),
15902 Mask, PassThru, Subtarget, DAG);
15904 case INTR_TYPE_2OP_MASK_RM: {
15905 SDValue Src1 = Op.getOperand(1);
15906 SDValue Src2 = Op.getOperand(2);
15907 SDValue PassThru = Op.getOperand(3);
15908 SDValue Mask = Op.getOperand(4);
15909 // We specify 2 possible modes for intrinsics, with/without rounding modes.
15910 // First, we check if the intrinsic have rounding mode (6 operands),
15911 // if not, we set rounding mode to "current".
15913 if (Op.getNumOperands() == 6)
15914 Rnd = Op.getOperand(5);
15916 Rnd = DAG.getConstant(X86::STATIC_ROUNDING::CUR_DIRECTION, dl, MVT::i32);
15917 return getVectorMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT,
15919 Mask, PassThru, Subtarget, DAG);
15921 case INTR_TYPE_3OP_SCALAR_MASK_RM: {
15922 SDValue Src1 = Op.getOperand(1);
15923 SDValue Src2 = Op.getOperand(2);
15924 SDValue Src3 = Op.getOperand(3);
15925 SDValue PassThru = Op.getOperand(4);
15926 SDValue Mask = Op.getOperand(5);
15927 SDValue Sae = Op.getOperand(6);
15929 return getScalarMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT, Src1,
15931 Mask, PassThru, Subtarget, DAG);
15933 case INTR_TYPE_3OP_MASK_RM: {
15934 SDValue Src1 = Op.getOperand(1);
15935 SDValue Src2 = Op.getOperand(2);
15936 SDValue Imm = Op.getOperand(3);
15937 SDValue PassThru = Op.getOperand(4);
15938 SDValue Mask = Op.getOperand(5);
15939 // We specify 2 possible modes for intrinsics, with/without rounding modes.
15940 // First, we check if the intrinsic have rounding mode (7 operands),
15941 // if not, we set rounding mode to "current".
15943 if (Op.getNumOperands() == 7)
15944 Rnd = Op.getOperand(6);
15946 Rnd = DAG.getConstant(X86::STATIC_ROUNDING::CUR_DIRECTION, dl, MVT::i32);
15947 return getVectorMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT,
15948 Src1, Src2, Imm, Rnd),
15949 Mask, PassThru, Subtarget, DAG);
15951 case INTR_TYPE_3OP_IMM8_MASK:
15952 case INTR_TYPE_3OP_MASK: {
15953 SDValue Src1 = Op.getOperand(1);
15954 SDValue Src2 = Op.getOperand(2);
15955 SDValue Src3 = Op.getOperand(3);
15956 SDValue PassThru = Op.getOperand(4);
15957 SDValue Mask = Op.getOperand(5);
15959 if (IntrData->Type == INTR_TYPE_3OP_IMM8_MASK)
15960 Src3 = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Src3);
15961 // We specify 2 possible opcodes for intrinsics with rounding modes.
15962 // First, we check if the intrinsic may have non-default rounding mode,
15963 // (IntrData->Opc1 != 0), then we check the rounding mode operand.
15964 unsigned IntrWithRoundingModeOpcode = IntrData->Opc1;
15965 if (IntrWithRoundingModeOpcode != 0) {
15966 SDValue Rnd = Op.getOperand(6);
15967 unsigned Round = cast<ConstantSDNode>(Rnd)->getZExtValue();
15968 if (Round != X86::STATIC_ROUNDING::CUR_DIRECTION) {
15969 return getVectorMaskingNode(DAG.getNode(IntrWithRoundingModeOpcode,
15970 dl, Op.getValueType(),
15971 Src1, Src2, Src3, Rnd),
15972 Mask, PassThru, Subtarget, DAG);
15975 return getVectorMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT,
15977 Mask, PassThru, Subtarget, DAG);
15979 case VPERM_3OP_MASKZ:
15980 case VPERM_3OP_MASK:
15983 case FMA_OP_MASK: {
15984 SDValue Src1 = Op.getOperand(1);
15985 SDValue Src2 = Op.getOperand(2);
15986 SDValue Src3 = Op.getOperand(3);
15987 SDValue Mask = Op.getOperand(4);
15988 EVT VT = Op.getValueType();
15989 SDValue PassThru = SDValue();
15991 // set PassThru element
15992 if (IntrData->Type == VPERM_3OP_MASKZ || IntrData->Type == FMA_OP_MASKZ)
15993 PassThru = getZeroVector(VT, Subtarget, DAG, dl);
15994 else if (IntrData->Type == FMA_OP_MASK3)
15999 // We specify 2 possible opcodes for intrinsics with rounding modes.
16000 // First, we check if the intrinsic may have non-default rounding mode,
16001 // (IntrData->Opc1 != 0), then we check the rounding mode operand.
16002 unsigned IntrWithRoundingModeOpcode = IntrData->Opc1;
16003 if (IntrWithRoundingModeOpcode != 0) {
16004 SDValue Rnd = Op.getOperand(5);
16005 if (cast<ConstantSDNode>(Rnd)->getZExtValue() !=
16006 X86::STATIC_ROUNDING::CUR_DIRECTION)
16007 return getVectorMaskingNode(DAG.getNode(IntrWithRoundingModeOpcode,
16008 dl, Op.getValueType(),
16009 Src1, Src2, Src3, Rnd),
16010 Mask, PassThru, Subtarget, DAG);
16012 return getVectorMaskingNode(DAG.getNode(IntrData->Opc0,
16013 dl, Op.getValueType(),
16015 Mask, PassThru, Subtarget, DAG);
16018 case CMP_MASK_CC: {
16019 // Comparison intrinsics with masks.
16020 // Example of transformation:
16021 // (i8 (int_x86_avx512_mask_pcmpeq_q_128
16022 // (v2i64 %a), (v2i64 %b), (i8 %mask))) ->
16024 // (v8i1 (insert_subvector undef,
16025 // (v2i1 (and (PCMPEQM %a, %b),
16026 // (extract_subvector
16027 // (v8i1 (bitcast %mask)), 0))), 0))))
16028 EVT VT = Op.getOperand(1).getValueType();
16029 EVT MaskVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
16030 VT.getVectorNumElements());
16031 SDValue Mask = Op.getOperand((IntrData->Type == CMP_MASK_CC) ? 4 : 3);
16032 EVT BitcastVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
16033 Mask.getValueType().getSizeInBits());
16035 if (IntrData->Type == CMP_MASK_CC) {
16036 SDValue CC = Op.getOperand(3);
16037 CC = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, CC);
16038 // We specify 2 possible opcodes for intrinsics with rounding modes.
16039 // First, we check if the intrinsic may have non-default rounding mode,
16040 // (IntrData->Opc1 != 0), then we check the rounding mode operand.
16041 if (IntrData->Opc1 != 0) {
16042 SDValue Rnd = Op.getOperand(5);
16043 if (cast<ConstantSDNode>(Rnd)->getZExtValue() !=
16044 X86::STATIC_ROUNDING::CUR_DIRECTION)
16045 Cmp = DAG.getNode(IntrData->Opc1, dl, MaskVT, Op.getOperand(1),
16046 Op.getOperand(2), CC, Rnd);
16048 //default rounding mode
16050 Cmp = DAG.getNode(IntrData->Opc0, dl, MaskVT, Op.getOperand(1),
16051 Op.getOperand(2), CC);
16054 assert(IntrData->Type == CMP_MASK && "Unexpected intrinsic type!");
16055 Cmp = DAG.getNode(IntrData->Opc0, dl, MaskVT, Op.getOperand(1),
16058 SDValue CmpMask = getVectorMaskingNode(Cmp, Mask,
16059 DAG.getTargetConstant(0, dl,
16062 SDValue Res = DAG.getNode(ISD::INSERT_SUBVECTOR, dl, BitcastVT,
16063 DAG.getUNDEF(BitcastVT), CmpMask,
16064 DAG.getIntPtrConstant(0, dl));
16065 return DAG.getBitcast(Op.getValueType(), Res);
16067 case COMI: { // Comparison intrinsics
16068 ISD::CondCode CC = (ISD::CondCode)IntrData->Opc1;
16069 SDValue LHS = Op.getOperand(1);
16070 SDValue RHS = Op.getOperand(2);
16071 unsigned X86CC = TranslateX86CC(CC, dl, true, LHS, RHS, DAG);
16072 assert(X86CC != X86::COND_INVALID && "Unexpected illegal condition!");
16073 SDValue Cond = DAG.getNode(IntrData->Opc0, dl, MVT::i32, LHS, RHS);
16074 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
16075 DAG.getConstant(X86CC, dl, MVT::i8), Cond);
16076 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
16079 return getTargetVShiftNode(IntrData->Opc0, dl, Op.getSimpleValueType(),
16080 Op.getOperand(1), Op.getOperand(2), DAG);
16082 return getVectorMaskingNode(getTargetVShiftNode(IntrData->Opc0, dl,
16083 Op.getSimpleValueType(),
16085 Op.getOperand(2), DAG),
16086 Op.getOperand(4), Op.getOperand(3), Subtarget,
16088 case COMPRESS_EXPAND_IN_REG: {
16089 SDValue Mask = Op.getOperand(3);
16090 SDValue DataToCompress = Op.getOperand(1);
16091 SDValue PassThru = Op.getOperand(2);
16092 if (isAllOnes(Mask)) // return data as is
16093 return Op.getOperand(1);
16095 return getVectorMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT,
16097 Mask, PassThru, Subtarget, DAG);
16100 SDValue Mask = Op.getOperand(3);
16101 EVT VT = Op.getValueType();
16102 EVT MaskVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
16103 VT.getVectorNumElements());
16104 EVT BitcastVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
16105 Mask.getValueType().getSizeInBits());
16107 SDValue VMask = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MaskVT,
16108 DAG.getBitcast(BitcastVT, Mask),
16109 DAG.getIntPtrConstant(0, dl));
16110 return DAG.getNode(IntrData->Opc0, dl, VT, VMask, Op.getOperand(1),
16119 default: return SDValue(); // Don't custom lower most intrinsics.
16121 case Intrinsic::x86_avx2_permd:
16122 case Intrinsic::x86_avx2_permps:
16123 // Operands intentionally swapped. Mask is last operand to intrinsic,
16124 // but second operand for node/instruction.
16125 return DAG.getNode(X86ISD::VPERMV, dl, Op.getValueType(),
16126 Op.getOperand(2), Op.getOperand(1));
16128 // ptest and testp intrinsics. The intrinsic these come from are designed to
16129 // return an integer value, not just an instruction so lower it to the ptest
16130 // or testp pattern and a setcc for the result.
16131 case Intrinsic::x86_sse41_ptestz:
16132 case Intrinsic::x86_sse41_ptestc:
16133 case Intrinsic::x86_sse41_ptestnzc:
16134 case Intrinsic::x86_avx_ptestz_256:
16135 case Intrinsic::x86_avx_ptestc_256:
16136 case Intrinsic::x86_avx_ptestnzc_256:
16137 case Intrinsic::x86_avx_vtestz_ps:
16138 case Intrinsic::x86_avx_vtestc_ps:
16139 case Intrinsic::x86_avx_vtestnzc_ps:
16140 case Intrinsic::x86_avx_vtestz_pd:
16141 case Intrinsic::x86_avx_vtestc_pd:
16142 case Intrinsic::x86_avx_vtestnzc_pd:
16143 case Intrinsic::x86_avx_vtestz_ps_256:
16144 case Intrinsic::x86_avx_vtestc_ps_256:
16145 case Intrinsic::x86_avx_vtestnzc_ps_256:
16146 case Intrinsic::x86_avx_vtestz_pd_256:
16147 case Intrinsic::x86_avx_vtestc_pd_256:
16148 case Intrinsic::x86_avx_vtestnzc_pd_256: {
16149 bool IsTestPacked = false;
16152 default: llvm_unreachable("Bad fallthrough in Intrinsic lowering.");
16153 case Intrinsic::x86_avx_vtestz_ps:
16154 case Intrinsic::x86_avx_vtestz_pd:
16155 case Intrinsic::x86_avx_vtestz_ps_256:
16156 case Intrinsic::x86_avx_vtestz_pd_256:
16157 IsTestPacked = true; // Fallthrough
16158 case Intrinsic::x86_sse41_ptestz:
16159 case Intrinsic::x86_avx_ptestz_256:
16161 X86CC = X86::COND_E;
16163 case Intrinsic::x86_avx_vtestc_ps:
16164 case Intrinsic::x86_avx_vtestc_pd:
16165 case Intrinsic::x86_avx_vtestc_ps_256:
16166 case Intrinsic::x86_avx_vtestc_pd_256:
16167 IsTestPacked = true; // Fallthrough
16168 case Intrinsic::x86_sse41_ptestc:
16169 case Intrinsic::x86_avx_ptestc_256:
16171 X86CC = X86::COND_B;
16173 case Intrinsic::x86_avx_vtestnzc_ps:
16174 case Intrinsic::x86_avx_vtestnzc_pd:
16175 case Intrinsic::x86_avx_vtestnzc_ps_256:
16176 case Intrinsic::x86_avx_vtestnzc_pd_256:
16177 IsTestPacked = true; // Fallthrough
16178 case Intrinsic::x86_sse41_ptestnzc:
16179 case Intrinsic::x86_avx_ptestnzc_256:
16181 X86CC = X86::COND_A;
16185 SDValue LHS = Op.getOperand(1);
16186 SDValue RHS = Op.getOperand(2);
16187 unsigned TestOpc = IsTestPacked ? X86ISD::TESTP : X86ISD::PTEST;
16188 SDValue Test = DAG.getNode(TestOpc, dl, MVT::i32, LHS, RHS);
16189 SDValue CC = DAG.getConstant(X86CC, dl, MVT::i8);
16190 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8, CC, Test);
16191 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
16193 case Intrinsic::x86_avx512_kortestz_w:
16194 case Intrinsic::x86_avx512_kortestc_w: {
16195 unsigned X86CC = (IntNo == Intrinsic::x86_avx512_kortestz_w)? X86::COND_E: X86::COND_B;
16196 SDValue LHS = DAG.getBitcast(MVT::v16i1, Op.getOperand(1));
16197 SDValue RHS = DAG.getBitcast(MVT::v16i1, Op.getOperand(2));
16198 SDValue CC = DAG.getConstant(X86CC, dl, MVT::i8);
16199 SDValue Test = DAG.getNode(X86ISD::KORTEST, dl, MVT::i32, LHS, RHS);
16200 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i1, CC, Test);
16201 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
16204 case Intrinsic::x86_sse42_pcmpistria128:
16205 case Intrinsic::x86_sse42_pcmpestria128:
16206 case Intrinsic::x86_sse42_pcmpistric128:
16207 case Intrinsic::x86_sse42_pcmpestric128:
16208 case Intrinsic::x86_sse42_pcmpistrio128:
16209 case Intrinsic::x86_sse42_pcmpestrio128:
16210 case Intrinsic::x86_sse42_pcmpistris128:
16211 case Intrinsic::x86_sse42_pcmpestris128:
16212 case Intrinsic::x86_sse42_pcmpistriz128:
16213 case Intrinsic::x86_sse42_pcmpestriz128: {
16217 default: llvm_unreachable("Impossible intrinsic"); // Can't reach here.
16218 case Intrinsic::x86_sse42_pcmpistria128:
16219 Opcode = X86ISD::PCMPISTRI;
16220 X86CC = X86::COND_A;
16222 case Intrinsic::x86_sse42_pcmpestria128:
16223 Opcode = X86ISD::PCMPESTRI;
16224 X86CC = X86::COND_A;
16226 case Intrinsic::x86_sse42_pcmpistric128:
16227 Opcode = X86ISD::PCMPISTRI;
16228 X86CC = X86::COND_B;
16230 case Intrinsic::x86_sse42_pcmpestric128:
16231 Opcode = X86ISD::PCMPESTRI;
16232 X86CC = X86::COND_B;
16234 case Intrinsic::x86_sse42_pcmpistrio128:
16235 Opcode = X86ISD::PCMPISTRI;
16236 X86CC = X86::COND_O;
16238 case Intrinsic::x86_sse42_pcmpestrio128:
16239 Opcode = X86ISD::PCMPESTRI;
16240 X86CC = X86::COND_O;
16242 case Intrinsic::x86_sse42_pcmpistris128:
16243 Opcode = X86ISD::PCMPISTRI;
16244 X86CC = X86::COND_S;
16246 case Intrinsic::x86_sse42_pcmpestris128:
16247 Opcode = X86ISD::PCMPESTRI;
16248 X86CC = X86::COND_S;
16250 case Intrinsic::x86_sse42_pcmpistriz128:
16251 Opcode = X86ISD::PCMPISTRI;
16252 X86CC = X86::COND_E;
16254 case Intrinsic::x86_sse42_pcmpestriz128:
16255 Opcode = X86ISD::PCMPESTRI;
16256 X86CC = X86::COND_E;
16259 SmallVector<SDValue, 5> NewOps(Op->op_begin()+1, Op->op_end());
16260 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::i32);
16261 SDValue PCMP = DAG.getNode(Opcode, dl, VTs, NewOps);
16262 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
16263 DAG.getConstant(X86CC, dl, MVT::i8),
16264 SDValue(PCMP.getNode(), 1));
16265 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
16268 case Intrinsic::x86_sse42_pcmpistri128:
16269 case Intrinsic::x86_sse42_pcmpestri128: {
16271 if (IntNo == Intrinsic::x86_sse42_pcmpistri128)
16272 Opcode = X86ISD::PCMPISTRI;
16274 Opcode = X86ISD::PCMPESTRI;
16276 SmallVector<SDValue, 5> NewOps(Op->op_begin()+1, Op->op_end());
16277 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::i32);
16278 return DAG.getNode(Opcode, dl, VTs, NewOps);
16281 case Intrinsic::x86_seh_lsda: {
16282 // Compute the symbol for the LSDA. We know it'll get emitted later.
16283 MachineFunction &MF = DAG.getMachineFunction();
16284 SDValue Op1 = Op.getOperand(1);
16285 auto *Fn = cast<Function>(cast<GlobalAddressSDNode>(Op1)->getGlobal());
16286 MCSymbol *LSDASym = MF.getMMI().getContext().getOrCreateLSDASymbol(
16287 GlobalValue::getRealLinkageName(Fn->getName()));
16289 // Generate a simple absolute symbol reference. This intrinsic is only
16290 // supported on 32-bit Windows, which isn't PIC.
16291 SDValue Result = DAG.getMCSymbol(LSDASym, VT);
16292 return DAG.getNode(X86ISD::Wrapper, dl, VT, Result);
16295 case Intrinsic::x86_seh_recoverfp: {
16296 SDValue FnOp = Op.getOperand(1);
16297 SDValue IncomingFPOp = Op.getOperand(2);
16298 GlobalAddressSDNode *GSD = dyn_cast<GlobalAddressSDNode>(FnOp);
16299 auto *Fn = dyn_cast_or_null<Function>(GSD ? GSD->getGlobal() : nullptr);
16301 report_fatal_error(
16302 "llvm.x86.seh.recoverfp must take a function as the first argument");
16303 return recoverFramePointer(DAG, Fn, IncomingFPOp);
16306 case Intrinsic::localaddress: {
16307 // Returns one of the stack, base, or frame pointer registers, depending on
16308 // which is used to reference local variables.
16309 MachineFunction &MF = DAG.getMachineFunction();
16310 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
16312 if (RegInfo->hasBasePointer(MF))
16313 Reg = RegInfo->getBaseRegister();
16314 else // This function handles the SP or FP case.
16315 Reg = RegInfo->getPtrSizedFrameRegister(MF);
16316 return DAG.getCopyFromReg(DAG.getEntryNode(), dl, Reg, VT);
16321 static SDValue getGatherNode(unsigned Opc, SDValue Op, SelectionDAG &DAG,
16322 SDValue Src, SDValue Mask, SDValue Base,
16323 SDValue Index, SDValue ScaleOp, SDValue Chain,
16324 const X86Subtarget * Subtarget) {
16326 ConstantSDNode *C = dyn_cast<ConstantSDNode>(ScaleOp);
16328 llvm_unreachable("Invalid scale type");
16329 unsigned ScaleVal = C->getZExtValue();
16330 if (ScaleVal > 2 && ScaleVal != 4 && ScaleVal != 8)
16331 llvm_unreachable("Valid scale values are 1, 2, 4, 8");
16333 SDValue Scale = DAG.getTargetConstant(C->getZExtValue(), dl, MVT::i8);
16334 EVT MaskVT = MVT::getVectorVT(MVT::i1,
16335 Index.getSimpleValueType().getVectorNumElements());
16337 ConstantSDNode *MaskC = dyn_cast<ConstantSDNode>(Mask);
16339 MaskInReg = DAG.getTargetConstant(MaskC->getSExtValue(), dl, MaskVT);
16341 EVT BitcastVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
16342 Mask.getValueType().getSizeInBits());
16344 // In case when MaskVT equals v2i1 or v4i1, low 2 or 4 elements
16345 // are extracted by EXTRACT_SUBVECTOR.
16346 MaskInReg = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MaskVT,
16347 DAG.getBitcast(BitcastVT, Mask),
16348 DAG.getIntPtrConstant(0, dl));
16350 SDVTList VTs = DAG.getVTList(Op.getValueType(), MaskVT, MVT::Other);
16351 SDValue Disp = DAG.getTargetConstant(0, dl, MVT::i32);
16352 SDValue Segment = DAG.getRegister(0, MVT::i32);
16353 if (Src.getOpcode() == ISD::UNDEF)
16354 Src = getZeroVector(Op.getValueType(), Subtarget, DAG, dl);
16355 SDValue Ops[] = {Src, MaskInReg, Base, Scale, Index, Disp, Segment, Chain};
16356 SDNode *Res = DAG.getMachineNode(Opc, dl, VTs, Ops);
16357 SDValue RetOps[] = { SDValue(Res, 0), SDValue(Res, 2) };
16358 return DAG.getMergeValues(RetOps, dl);
16361 static SDValue getScatterNode(unsigned Opc, SDValue Op, SelectionDAG &DAG,
16362 SDValue Src, SDValue Mask, SDValue Base,
16363 SDValue Index, SDValue ScaleOp, SDValue Chain) {
16365 ConstantSDNode *C = dyn_cast<ConstantSDNode>(ScaleOp);
16367 llvm_unreachable("Invalid scale type");
16368 unsigned ScaleVal = C->getZExtValue();
16369 if (ScaleVal > 2 && ScaleVal != 4 && ScaleVal != 8)
16370 llvm_unreachable("Valid scale values are 1, 2, 4, 8");
16372 SDValue Scale = DAG.getTargetConstant(C->getZExtValue(), dl, MVT::i8);
16373 SDValue Disp = DAG.getTargetConstant(0, dl, MVT::i32);
16374 SDValue Segment = DAG.getRegister(0, MVT::i32);
16375 EVT MaskVT = MVT::getVectorVT(MVT::i1,
16376 Index.getSimpleValueType().getVectorNumElements());
16378 ConstantSDNode *MaskC = dyn_cast<ConstantSDNode>(Mask);
16380 MaskInReg = DAG.getTargetConstant(MaskC->getSExtValue(), dl, MaskVT);
16382 EVT BitcastVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
16383 Mask.getValueType().getSizeInBits());
16385 // In case when MaskVT equals v2i1 or v4i1, low 2 or 4 elements
16386 // are extracted by EXTRACT_SUBVECTOR.
16387 MaskInReg = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MaskVT,
16388 DAG.getBitcast(BitcastVT, Mask),
16389 DAG.getIntPtrConstant(0, dl));
16391 SDVTList VTs = DAG.getVTList(MaskVT, MVT::Other);
16392 SDValue Ops[] = {Base, Scale, Index, Disp, Segment, MaskInReg, Src, Chain};
16393 SDNode *Res = DAG.getMachineNode(Opc, dl, VTs, Ops);
16394 return SDValue(Res, 1);
16397 static SDValue getPrefetchNode(unsigned Opc, SDValue Op, SelectionDAG &DAG,
16398 SDValue Mask, SDValue Base, SDValue Index,
16399 SDValue ScaleOp, SDValue Chain) {
16401 ConstantSDNode *C = dyn_cast<ConstantSDNode>(ScaleOp);
16402 assert(C && "Invalid scale type");
16403 SDValue Scale = DAG.getTargetConstant(C->getZExtValue(), dl, MVT::i8);
16404 SDValue Disp = DAG.getTargetConstant(0, dl, MVT::i32);
16405 SDValue Segment = DAG.getRegister(0, MVT::i32);
16407 MVT::getVectorVT(MVT::i1, Index.getSimpleValueType().getVectorNumElements());
16409 ConstantSDNode *MaskC = dyn_cast<ConstantSDNode>(Mask);
16411 MaskInReg = DAG.getTargetConstant(MaskC->getSExtValue(), dl, MaskVT);
16413 MaskInReg = DAG.getBitcast(MaskVT, Mask);
16414 //SDVTList VTs = DAG.getVTList(MVT::Other);
16415 SDValue Ops[] = {MaskInReg, Base, Scale, Index, Disp, Segment, Chain};
16416 SDNode *Res = DAG.getMachineNode(Opc, dl, MVT::Other, Ops);
16417 return SDValue(Res, 0);
16420 // getReadPerformanceCounter - Handles the lowering of builtin intrinsics that
16421 // read performance monitor counters (x86_rdpmc).
16422 static void getReadPerformanceCounter(SDNode *N, SDLoc DL,
16423 SelectionDAG &DAG, const X86Subtarget *Subtarget,
16424 SmallVectorImpl<SDValue> &Results) {
16425 assert(N->getNumOperands() == 3 && "Unexpected number of operands!");
16426 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Glue);
16429 // The ECX register is used to select the index of the performance counter
16431 SDValue Chain = DAG.getCopyToReg(N->getOperand(0), DL, X86::ECX,
16433 SDValue rd = DAG.getNode(X86ISD::RDPMC_DAG, DL, Tys, Chain);
16435 // Reads the content of a 64-bit performance counter and returns it in the
16436 // registers EDX:EAX.
16437 if (Subtarget->is64Bit()) {
16438 LO = DAG.getCopyFromReg(rd, DL, X86::RAX, MVT::i64, rd.getValue(1));
16439 HI = DAG.getCopyFromReg(LO.getValue(1), DL, X86::RDX, MVT::i64,
16442 LO = DAG.getCopyFromReg(rd, DL, X86::EAX, MVT::i32, rd.getValue(1));
16443 HI = DAG.getCopyFromReg(LO.getValue(1), DL, X86::EDX, MVT::i32,
16446 Chain = HI.getValue(1);
16448 if (Subtarget->is64Bit()) {
16449 // The EAX register is loaded with the low-order 32 bits. The EDX register
16450 // is loaded with the supported high-order bits of the counter.
16451 SDValue Tmp = DAG.getNode(ISD::SHL, DL, MVT::i64, HI,
16452 DAG.getConstant(32, DL, MVT::i8));
16453 Results.push_back(DAG.getNode(ISD::OR, DL, MVT::i64, LO, Tmp));
16454 Results.push_back(Chain);
16458 // Use a buildpair to merge the two 32-bit values into a 64-bit one.
16459 SDValue Ops[] = { LO, HI };
16460 SDValue Pair = DAG.getNode(ISD::BUILD_PAIR, DL, MVT::i64, Ops);
16461 Results.push_back(Pair);
16462 Results.push_back(Chain);
16465 // getReadTimeStampCounter - Handles the lowering of builtin intrinsics that
16466 // read the time stamp counter (x86_rdtsc and x86_rdtscp). This function is
16467 // also used to custom lower READCYCLECOUNTER nodes.
16468 static void getReadTimeStampCounter(SDNode *N, SDLoc DL, unsigned Opcode,
16469 SelectionDAG &DAG, const X86Subtarget *Subtarget,
16470 SmallVectorImpl<SDValue> &Results) {
16471 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Glue);
16472 SDValue rd = DAG.getNode(Opcode, DL, Tys, N->getOperand(0));
16475 // The processor's time-stamp counter (a 64-bit MSR) is stored into the
16476 // EDX:EAX registers. EDX is loaded with the high-order 32 bits of the MSR
16477 // and the EAX register is loaded with the low-order 32 bits.
16478 if (Subtarget->is64Bit()) {
16479 LO = DAG.getCopyFromReg(rd, DL, X86::RAX, MVT::i64, rd.getValue(1));
16480 HI = DAG.getCopyFromReg(LO.getValue(1), DL, X86::RDX, MVT::i64,
16483 LO = DAG.getCopyFromReg(rd, DL, X86::EAX, MVT::i32, rd.getValue(1));
16484 HI = DAG.getCopyFromReg(LO.getValue(1), DL, X86::EDX, MVT::i32,
16487 SDValue Chain = HI.getValue(1);
16489 if (Opcode == X86ISD::RDTSCP_DAG) {
16490 assert(N->getNumOperands() == 3 && "Unexpected number of operands!");
16492 // Instruction RDTSCP loads the IA32:TSC_AUX_MSR (address C000_0103H) into
16493 // the ECX register. Add 'ecx' explicitly to the chain.
16494 SDValue ecx = DAG.getCopyFromReg(Chain, DL, X86::ECX, MVT::i32,
16496 // Explicitly store the content of ECX at the location passed in input
16497 // to the 'rdtscp' intrinsic.
16498 Chain = DAG.getStore(ecx.getValue(1), DL, ecx, N->getOperand(2),
16499 MachinePointerInfo(), false, false, 0);
16502 if (Subtarget->is64Bit()) {
16503 // The EDX register is loaded with the high-order 32 bits of the MSR, and
16504 // the EAX register is loaded with the low-order 32 bits.
16505 SDValue Tmp = DAG.getNode(ISD::SHL, DL, MVT::i64, HI,
16506 DAG.getConstant(32, DL, MVT::i8));
16507 Results.push_back(DAG.getNode(ISD::OR, DL, MVT::i64, LO, Tmp));
16508 Results.push_back(Chain);
16512 // Use a buildpair to merge the two 32-bit values into a 64-bit one.
16513 SDValue Ops[] = { LO, HI };
16514 SDValue Pair = DAG.getNode(ISD::BUILD_PAIR, DL, MVT::i64, Ops);
16515 Results.push_back(Pair);
16516 Results.push_back(Chain);
16519 static SDValue LowerREADCYCLECOUNTER(SDValue Op, const X86Subtarget *Subtarget,
16520 SelectionDAG &DAG) {
16521 SmallVector<SDValue, 2> Results;
16523 getReadTimeStampCounter(Op.getNode(), DL, X86ISD::RDTSC_DAG, DAG, Subtarget,
16525 return DAG.getMergeValues(Results, DL);
16528 static SDValue LowerSEHRESTOREFRAME(SDValue Op, const X86Subtarget *Subtarget,
16529 SelectionDAG &DAG) {
16530 MachineFunction &MF = DAG.getMachineFunction();
16531 const Function *Fn = MF.getFunction();
16533 SDValue Chain = Op.getOperand(0);
16535 assert(Subtarget->getFrameLowering()->hasFP(MF) &&
16536 "using llvm.x86.seh.restoreframe requires a frame pointer");
16538 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
16539 MVT VT = TLI.getPointerTy(DAG.getDataLayout());
16541 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
16542 unsigned FrameReg =
16543 RegInfo->getPtrSizedFrameRegister(DAG.getMachineFunction());
16544 unsigned SPReg = RegInfo->getStackRegister();
16545 unsigned SlotSize = RegInfo->getSlotSize();
16547 // Get incoming EBP.
16548 SDValue IncomingEBP =
16549 DAG.getCopyFromReg(Chain, dl, FrameReg, VT);
16551 // SP is saved in the first field of every registration node, so load
16552 // [EBP-RegNodeSize] into SP.
16553 int RegNodeSize = getSEHRegistrationNodeSize(Fn);
16554 SDValue SPAddr = DAG.getNode(ISD::ADD, dl, VT, IncomingEBP,
16555 DAG.getConstant(-RegNodeSize, dl, VT));
16557 DAG.getLoad(VT, dl, Chain, SPAddr, MachinePointerInfo(), false, false,
16558 false, VT.getScalarSizeInBits() / 8);
16559 Chain = DAG.getCopyToReg(Chain, dl, SPReg, NewSP);
16561 if (!RegInfo->needsStackRealignment(MF)) {
16562 // Adjust EBP to point back to the original frame position.
16563 SDValue NewFP = recoverFramePointer(DAG, Fn, IncomingEBP);
16564 Chain = DAG.getCopyToReg(Chain, dl, FrameReg, NewFP);
16566 assert(RegInfo->hasBasePointer(MF) &&
16567 "functions with Win32 EH must use frame or base pointer register");
16569 // Reload the base pointer (ESI) with the adjusted incoming EBP.
16570 SDValue NewBP = recoverFramePointer(DAG, Fn, IncomingEBP);
16571 Chain = DAG.getCopyToReg(Chain, dl, RegInfo->getBaseRegister(), NewBP);
16573 // Reload the spilled EBP value, now that the stack and base pointers are
16575 X86MachineFunctionInfo *X86FI = MF.getInfo<X86MachineFunctionInfo>();
16576 X86FI->setHasSEHFramePtrSave(true);
16577 int FI = MF.getFrameInfo()->CreateSpillStackObject(SlotSize, SlotSize);
16578 X86FI->setSEHFramePtrSaveIndex(FI);
16579 SDValue NewFP = DAG.getLoad(VT, dl, Chain, DAG.getFrameIndex(FI, VT),
16580 MachinePointerInfo(), false, false, false,
16581 VT.getScalarSizeInBits() / 8);
16582 Chain = DAG.getCopyToReg(NewFP, dl, FrameReg, NewFP);
16588 /// \brief Lower intrinsics for TRUNCATE_TO_MEM case
16589 /// return truncate Store/MaskedStore Node
16590 static SDValue LowerINTRINSIC_TRUNCATE_TO_MEM(const SDValue & Op,
16594 SDValue Mask = Op.getOperand(4);
16595 SDValue DataToTruncate = Op.getOperand(3);
16596 SDValue Addr = Op.getOperand(2);
16597 SDValue Chain = Op.getOperand(0);
16599 EVT VT = DataToTruncate.getValueType();
16600 EVT SVT = EVT::getVectorVT(*DAG.getContext(),
16601 ElementType, VT.getVectorNumElements());
16603 if (isAllOnes(Mask)) // return just a truncate store
16604 return DAG.getTruncStore(Chain, dl, DataToTruncate, Addr,
16605 MachinePointerInfo(), SVT, false, false,
16606 SVT.getScalarSizeInBits()/8);
16608 EVT MaskVT = EVT::getVectorVT(*DAG.getContext(),
16609 MVT::i1, VT.getVectorNumElements());
16610 EVT BitcastVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
16611 Mask.getValueType().getSizeInBits());
16612 // In case when MaskVT equals v2i1 or v4i1, low 2 or 4 elements
16613 // are extracted by EXTRACT_SUBVECTOR.
16614 SDValue VMask = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MaskVT,
16615 DAG.getBitcast(BitcastVT, Mask),
16616 DAG.getIntPtrConstant(0, dl));
16618 MachineMemOperand *MMO = DAG.getMachineFunction().
16619 getMachineMemOperand(MachinePointerInfo(),
16620 MachineMemOperand::MOStore, SVT.getStoreSize(),
16621 SVT.getScalarSizeInBits()/8);
16623 return DAG.getMaskedStore(Chain, dl, DataToTruncate, Addr,
16624 VMask, SVT, MMO, true);
16627 static SDValue LowerINTRINSIC_W_CHAIN(SDValue Op, const X86Subtarget *Subtarget,
16628 SelectionDAG &DAG) {
16629 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
16631 const IntrinsicData* IntrData = getIntrinsicWithChain(IntNo);
16633 if (IntNo == llvm::Intrinsic::x86_seh_restoreframe)
16634 return LowerSEHRESTOREFRAME(Op, Subtarget, DAG);
16639 switch(IntrData->Type) {
16641 llvm_unreachable("Unknown Intrinsic Type");
16645 // Emit the node with the right value type.
16646 SDVTList VTs = DAG.getVTList(Op->getValueType(0), MVT::Glue, MVT::Other);
16647 SDValue Result = DAG.getNode(IntrData->Opc0, dl, VTs, Op.getOperand(0));
16649 // If the value returned by RDRAND/RDSEED was valid (CF=1), return 1.
16650 // Otherwise return the value from Rand, which is always 0, casted to i32.
16651 SDValue Ops[] = { DAG.getZExtOrTrunc(Result, dl, Op->getValueType(1)),
16652 DAG.getConstant(1, dl, Op->getValueType(1)),
16653 DAG.getConstant(X86::COND_B, dl, MVT::i32),
16654 SDValue(Result.getNode(), 1) };
16655 SDValue isValid = DAG.getNode(X86ISD::CMOV, dl,
16656 DAG.getVTList(Op->getValueType(1), MVT::Glue),
16659 // Return { result, isValid, chain }.
16660 return DAG.getNode(ISD::MERGE_VALUES, dl, Op->getVTList(), Result, isValid,
16661 SDValue(Result.getNode(), 2));
16664 //gather(v1, mask, index, base, scale);
16665 SDValue Chain = Op.getOperand(0);
16666 SDValue Src = Op.getOperand(2);
16667 SDValue Base = Op.getOperand(3);
16668 SDValue Index = Op.getOperand(4);
16669 SDValue Mask = Op.getOperand(5);
16670 SDValue Scale = Op.getOperand(6);
16671 return getGatherNode(IntrData->Opc0, Op, DAG, Src, Mask, Base, Index, Scale,
16675 //scatter(base, mask, index, v1, scale);
16676 SDValue Chain = Op.getOperand(0);
16677 SDValue Base = Op.getOperand(2);
16678 SDValue Mask = Op.getOperand(3);
16679 SDValue Index = Op.getOperand(4);
16680 SDValue Src = Op.getOperand(5);
16681 SDValue Scale = Op.getOperand(6);
16682 return getScatterNode(IntrData->Opc0, Op, DAG, Src, Mask, Base, Index,
16686 SDValue Hint = Op.getOperand(6);
16687 unsigned HintVal = cast<ConstantSDNode>(Hint)->getZExtValue();
16688 assert(HintVal < 2 && "Wrong prefetch hint in intrinsic: should be 0 or 1");
16689 unsigned Opcode = (HintVal ? IntrData->Opc1 : IntrData->Opc0);
16690 SDValue Chain = Op.getOperand(0);
16691 SDValue Mask = Op.getOperand(2);
16692 SDValue Index = Op.getOperand(3);
16693 SDValue Base = Op.getOperand(4);
16694 SDValue Scale = Op.getOperand(5);
16695 return getPrefetchNode(Opcode, Op, DAG, Mask, Base, Index, Scale, Chain);
16697 // Read Time Stamp Counter (RDTSC) and Processor ID (RDTSCP).
16699 SmallVector<SDValue, 2> Results;
16700 getReadTimeStampCounter(Op.getNode(), dl, IntrData->Opc0, DAG, Subtarget,
16702 return DAG.getMergeValues(Results, dl);
16704 // Read Performance Monitoring Counters.
16706 SmallVector<SDValue, 2> Results;
16707 getReadPerformanceCounter(Op.getNode(), dl, DAG, Subtarget, Results);
16708 return DAG.getMergeValues(Results, dl);
16710 // XTEST intrinsics.
16712 SDVTList VTs = DAG.getVTList(Op->getValueType(0), MVT::Other);
16713 SDValue InTrans = DAG.getNode(IntrData->Opc0, dl, VTs, Op.getOperand(0));
16714 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
16715 DAG.getConstant(X86::COND_NE, dl, MVT::i8),
16717 SDValue Ret = DAG.getNode(ISD::ZERO_EXTEND, dl, Op->getValueType(0), SetCC);
16718 return DAG.getNode(ISD::MERGE_VALUES, dl, Op->getVTList(),
16719 Ret, SDValue(InTrans.getNode(), 1));
16723 SmallVector<SDValue, 2> Results;
16724 SDVTList CFVTs = DAG.getVTList(Op->getValueType(0), MVT::Other);
16725 SDVTList VTs = DAG.getVTList(Op.getOperand(3)->getValueType(0), MVT::Other);
16726 SDValue GenCF = DAG.getNode(X86ISD::ADD, dl, CFVTs, Op.getOperand(2),
16727 DAG.getConstant(-1, dl, MVT::i8));
16728 SDValue Res = DAG.getNode(IntrData->Opc0, dl, VTs, Op.getOperand(3),
16729 Op.getOperand(4), GenCF.getValue(1));
16730 SDValue Store = DAG.getStore(Op.getOperand(0), dl, Res.getValue(0),
16731 Op.getOperand(5), MachinePointerInfo(),
16733 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
16734 DAG.getConstant(X86::COND_B, dl, MVT::i8),
16736 Results.push_back(SetCC);
16737 Results.push_back(Store);
16738 return DAG.getMergeValues(Results, dl);
16740 case COMPRESS_TO_MEM: {
16742 SDValue Mask = Op.getOperand(4);
16743 SDValue DataToCompress = Op.getOperand(3);
16744 SDValue Addr = Op.getOperand(2);
16745 SDValue Chain = Op.getOperand(0);
16747 EVT VT = DataToCompress.getValueType();
16748 if (isAllOnes(Mask)) // return just a store
16749 return DAG.getStore(Chain, dl, DataToCompress, Addr,
16750 MachinePointerInfo(), false, false,
16751 VT.getScalarSizeInBits()/8);
16753 SDValue Compressed =
16754 getVectorMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT, DataToCompress),
16755 Mask, DAG.getUNDEF(VT), Subtarget, DAG);
16756 return DAG.getStore(Chain, dl, Compressed, Addr,
16757 MachinePointerInfo(), false, false,
16758 VT.getScalarSizeInBits()/8);
16760 case TRUNCATE_TO_MEM_VI8:
16761 return LowerINTRINSIC_TRUNCATE_TO_MEM(Op, DAG, MVT::i8);
16762 case TRUNCATE_TO_MEM_VI16:
16763 return LowerINTRINSIC_TRUNCATE_TO_MEM(Op, DAG, MVT::i16);
16764 case TRUNCATE_TO_MEM_VI32:
16765 return LowerINTRINSIC_TRUNCATE_TO_MEM(Op, DAG, MVT::i32);
16766 case EXPAND_FROM_MEM: {
16768 SDValue Mask = Op.getOperand(4);
16769 SDValue PassThru = Op.getOperand(3);
16770 SDValue Addr = Op.getOperand(2);
16771 SDValue Chain = Op.getOperand(0);
16772 EVT VT = Op.getValueType();
16774 if (isAllOnes(Mask)) // return just a load
16775 return DAG.getLoad(VT, dl, Chain, Addr, MachinePointerInfo(), false, false,
16776 false, VT.getScalarSizeInBits()/8);
16778 SDValue DataToExpand = DAG.getLoad(VT, dl, Chain, Addr, MachinePointerInfo(),
16779 false, false, false,
16780 VT.getScalarSizeInBits()/8);
16782 SDValue Results[] = {
16783 getVectorMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT, DataToExpand),
16784 Mask, PassThru, Subtarget, DAG), Chain};
16785 return DAG.getMergeValues(Results, dl);
16790 SDValue X86TargetLowering::LowerRETURNADDR(SDValue Op,
16791 SelectionDAG &DAG) const {
16792 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
16793 MFI->setReturnAddressIsTaken(true);
16795 if (verifyReturnAddressArgumentIsConstant(Op, DAG))
16798 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
16800 EVT PtrVT = getPointerTy(DAG.getDataLayout());
16803 SDValue FrameAddr = LowerFRAMEADDR(Op, DAG);
16804 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
16805 SDValue Offset = DAG.getConstant(RegInfo->getSlotSize(), dl, PtrVT);
16806 return DAG.getLoad(PtrVT, dl, DAG.getEntryNode(),
16807 DAG.getNode(ISD::ADD, dl, PtrVT,
16808 FrameAddr, Offset),
16809 MachinePointerInfo(), false, false, false, 0);
16812 // Just load the return address.
16813 SDValue RetAddrFI = getReturnAddressFrameIndex(DAG);
16814 return DAG.getLoad(PtrVT, dl, DAG.getEntryNode(),
16815 RetAddrFI, MachinePointerInfo(), false, false, false, 0);
16818 SDValue X86TargetLowering::LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const {
16819 MachineFunction &MF = DAG.getMachineFunction();
16820 MachineFrameInfo *MFI = MF.getFrameInfo();
16821 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
16822 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
16823 EVT VT = Op.getValueType();
16825 MFI->setFrameAddressIsTaken(true);
16827 if (MF.getTarget().getMCAsmInfo()->usesWindowsCFI()) {
16828 // Depth > 0 makes no sense on targets which use Windows unwind codes. It
16829 // is not possible to crawl up the stack without looking at the unwind codes
16831 int FrameAddrIndex = FuncInfo->getFAIndex();
16832 if (!FrameAddrIndex) {
16833 // Set up a frame object for the return address.
16834 unsigned SlotSize = RegInfo->getSlotSize();
16835 FrameAddrIndex = MF.getFrameInfo()->CreateFixedObject(
16836 SlotSize, /*Offset=*/0, /*IsImmutable=*/false);
16837 FuncInfo->setFAIndex(FrameAddrIndex);
16839 return DAG.getFrameIndex(FrameAddrIndex, VT);
16842 unsigned FrameReg =
16843 RegInfo->getPtrSizedFrameRegister(DAG.getMachineFunction());
16844 SDLoc dl(Op); // FIXME probably not meaningful
16845 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
16846 assert(((FrameReg == X86::RBP && VT == MVT::i64) ||
16847 (FrameReg == X86::EBP && VT == MVT::i32)) &&
16848 "Invalid Frame Register!");
16849 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, VT);
16851 FrameAddr = DAG.getLoad(VT, dl, DAG.getEntryNode(), FrameAddr,
16852 MachinePointerInfo(),
16853 false, false, false, 0);
16857 // FIXME? Maybe this could be a TableGen attribute on some registers and
16858 // this table could be generated automatically from RegInfo.
16859 unsigned X86TargetLowering::getRegisterByName(const char* RegName, EVT VT,
16860 SelectionDAG &DAG) const {
16861 const TargetFrameLowering &TFI = *Subtarget->getFrameLowering();
16862 const MachineFunction &MF = DAG.getMachineFunction();
16864 unsigned Reg = StringSwitch<unsigned>(RegName)
16865 .Case("esp", X86::ESP)
16866 .Case("rsp", X86::RSP)
16867 .Case("ebp", X86::EBP)
16868 .Case("rbp", X86::RBP)
16871 if (Reg == X86::EBP || Reg == X86::RBP) {
16872 if (!TFI.hasFP(MF))
16873 report_fatal_error("register " + StringRef(RegName) +
16874 " is allocatable: function has no frame pointer");
16877 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
16878 unsigned FrameReg =
16879 RegInfo->getPtrSizedFrameRegister(DAG.getMachineFunction());
16880 assert((FrameReg == X86::EBP || FrameReg == X86::RBP) &&
16881 "Invalid Frame Register!");
16889 report_fatal_error("Invalid register name global variable");
16892 SDValue X86TargetLowering::LowerFRAME_TO_ARGS_OFFSET(SDValue Op,
16893 SelectionDAG &DAG) const {
16894 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
16895 return DAG.getIntPtrConstant(2 * RegInfo->getSlotSize(), SDLoc(Op));
16898 SDValue X86TargetLowering::LowerEH_RETURN(SDValue Op, SelectionDAG &DAG) const {
16899 SDValue Chain = Op.getOperand(0);
16900 SDValue Offset = Op.getOperand(1);
16901 SDValue Handler = Op.getOperand(2);
16904 EVT PtrVT = getPointerTy(DAG.getDataLayout());
16905 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
16906 unsigned FrameReg = RegInfo->getFrameRegister(DAG.getMachineFunction());
16907 assert(((FrameReg == X86::RBP && PtrVT == MVT::i64) ||
16908 (FrameReg == X86::EBP && PtrVT == MVT::i32)) &&
16909 "Invalid Frame Register!");
16910 SDValue Frame = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, PtrVT);
16911 unsigned StoreAddrReg = (PtrVT == MVT::i64) ? X86::RCX : X86::ECX;
16913 SDValue StoreAddr = DAG.getNode(ISD::ADD, dl, PtrVT, Frame,
16914 DAG.getIntPtrConstant(RegInfo->getSlotSize(),
16916 StoreAddr = DAG.getNode(ISD::ADD, dl, PtrVT, StoreAddr, Offset);
16917 Chain = DAG.getStore(Chain, dl, Handler, StoreAddr, MachinePointerInfo(),
16919 Chain = DAG.getCopyToReg(Chain, dl, StoreAddrReg, StoreAddr);
16921 return DAG.getNode(X86ISD::EH_RETURN, dl, MVT::Other, Chain,
16922 DAG.getRegister(StoreAddrReg, PtrVT));
16925 SDValue X86TargetLowering::LowerCATCHRET(SDValue Op, SelectionDAG &DAG) const {
16926 SDValue Chain = Op.getOperand(0);
16927 SDValue Dest = Op.getOperand(1);
16930 MVT PtrVT = getPointerTy(DAG.getDataLayout());
16932 MachineFunction &MF = DAG.getMachineFunction();
16933 if (isAsynchronousEHPersonality(
16934 classifyEHPersonality(MF.getFunction()->getPersonalityFn()))) {
16935 // For SEH, codegen catchret as a branch for now.
16936 // FIXME: Insert something to restore the frame.
16937 return DAG.getNode(ISD::BR, DL, MVT::Other, Chain, Dest);
16940 unsigned ReturnReg = (PtrVT == MVT::i64 ? X86::RAX : X86::EAX);
16942 // Load the address of the destination block.
16943 // FIXME: Do this without creating a BlockAddress.
16944 MachineBasicBlock *DestMBB = cast<BasicBlockSDNode>(Dest)->getBasicBlock();
16946 BlockAddress::get(const_cast<Function *>(MF.getFunction()),
16947 const_cast<BasicBlock *>(DestMBB->getBasicBlock()));
16948 DestMBB->setHasAddressTaken();
16949 SDValue BlockPtr = DAG.getBlockAddress(BA, PtrVT);
16950 Chain = DAG.getCopyToReg(Chain, DL, ReturnReg, BlockPtr);
16951 return DAG.getNode(X86ISD::CATCHRET, DL, MVT::Other, Chain,
16952 DAG.getRegister(ReturnReg, PtrVT));
16955 SDValue X86TargetLowering::LowerCLEANUPRET(SDValue Op, SelectionDAG &DAG) const {
16956 return DAG.getNode(X86ISD::CLEANUPRET, SDLoc(Op), MVT::Other,
16960 SDValue X86TargetLowering::lowerEH_SJLJ_SETJMP(SDValue Op,
16961 SelectionDAG &DAG) const {
16963 return DAG.getNode(X86ISD::EH_SJLJ_SETJMP, DL,
16964 DAG.getVTList(MVT::i32, MVT::Other),
16965 Op.getOperand(0), Op.getOperand(1));
16968 SDValue X86TargetLowering::lowerEH_SJLJ_LONGJMP(SDValue Op,
16969 SelectionDAG &DAG) const {
16971 return DAG.getNode(X86ISD::EH_SJLJ_LONGJMP, DL, MVT::Other,
16972 Op.getOperand(0), Op.getOperand(1));
16975 static SDValue LowerADJUST_TRAMPOLINE(SDValue Op, SelectionDAG &DAG) {
16976 return Op.getOperand(0);
16979 SDValue X86TargetLowering::LowerINIT_TRAMPOLINE(SDValue Op,
16980 SelectionDAG &DAG) const {
16981 SDValue Root = Op.getOperand(0);
16982 SDValue Trmp = Op.getOperand(1); // trampoline
16983 SDValue FPtr = Op.getOperand(2); // nested function
16984 SDValue Nest = Op.getOperand(3); // 'nest' parameter value
16987 const Value *TrmpAddr = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
16988 const TargetRegisterInfo *TRI = Subtarget->getRegisterInfo();
16990 if (Subtarget->is64Bit()) {
16991 SDValue OutChains[6];
16993 // Large code-model.
16994 const unsigned char JMP64r = 0xFF; // 64-bit jmp through register opcode.
16995 const unsigned char MOV64ri = 0xB8; // X86::MOV64ri opcode.
16997 const unsigned char N86R10 = TRI->getEncodingValue(X86::R10) & 0x7;
16998 const unsigned char N86R11 = TRI->getEncodingValue(X86::R11) & 0x7;
17000 const unsigned char REX_WB = 0x40 | 0x08 | 0x01; // REX prefix
17002 // Load the pointer to the nested function into R11.
17003 unsigned OpCode = ((MOV64ri | N86R11) << 8) | REX_WB; // movabsq r11
17004 SDValue Addr = Trmp;
17005 OutChains[0] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, dl, MVT::i16),
17006 Addr, MachinePointerInfo(TrmpAddr),
17009 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
17010 DAG.getConstant(2, dl, MVT::i64));
17011 OutChains[1] = DAG.getStore(Root, dl, FPtr, Addr,
17012 MachinePointerInfo(TrmpAddr, 2),
17015 // Load the 'nest' parameter value into R10.
17016 // R10 is specified in X86CallingConv.td
17017 OpCode = ((MOV64ri | N86R10) << 8) | REX_WB; // movabsq r10
17018 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
17019 DAG.getConstant(10, dl, MVT::i64));
17020 OutChains[2] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, dl, MVT::i16),
17021 Addr, MachinePointerInfo(TrmpAddr, 10),
17024 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
17025 DAG.getConstant(12, dl, MVT::i64));
17026 OutChains[3] = DAG.getStore(Root, dl, Nest, Addr,
17027 MachinePointerInfo(TrmpAddr, 12),
17030 // Jump to the nested function.
17031 OpCode = (JMP64r << 8) | REX_WB; // jmpq *...
17032 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
17033 DAG.getConstant(20, dl, MVT::i64));
17034 OutChains[4] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, dl, MVT::i16),
17035 Addr, MachinePointerInfo(TrmpAddr, 20),
17038 unsigned char ModRM = N86R11 | (4 << 3) | (3 << 6); // ...r11
17039 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
17040 DAG.getConstant(22, dl, MVT::i64));
17041 OutChains[5] = DAG.getStore(Root, dl, DAG.getConstant(ModRM, dl, MVT::i8),
17042 Addr, MachinePointerInfo(TrmpAddr, 22),
17045 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other, OutChains);
17047 const Function *Func =
17048 cast<Function>(cast<SrcValueSDNode>(Op.getOperand(5))->getValue());
17049 CallingConv::ID CC = Func->getCallingConv();
17054 llvm_unreachable("Unsupported calling convention");
17055 case CallingConv::C:
17056 case CallingConv::X86_StdCall: {
17057 // Pass 'nest' parameter in ECX.
17058 // Must be kept in sync with X86CallingConv.td
17059 NestReg = X86::ECX;
17061 // Check that ECX wasn't needed by an 'inreg' parameter.
17062 FunctionType *FTy = Func->getFunctionType();
17063 const AttributeSet &Attrs = Func->getAttributes();
17065 if (!Attrs.isEmpty() && !Func->isVarArg()) {
17066 unsigned InRegCount = 0;
17069 for (FunctionType::param_iterator I = FTy->param_begin(),
17070 E = FTy->param_end(); I != E; ++I, ++Idx)
17071 if (Attrs.hasAttribute(Idx, Attribute::InReg)) {
17072 auto &DL = DAG.getDataLayout();
17073 // FIXME: should only count parameters that are lowered to integers.
17074 InRegCount += (DL.getTypeSizeInBits(*I) + 31) / 32;
17077 if (InRegCount > 2) {
17078 report_fatal_error("Nest register in use - reduce number of inreg"
17084 case CallingConv::X86_FastCall:
17085 case CallingConv::X86_ThisCall:
17086 case CallingConv::Fast:
17087 // Pass 'nest' parameter in EAX.
17088 // Must be kept in sync with X86CallingConv.td
17089 NestReg = X86::EAX;
17093 SDValue OutChains[4];
17094 SDValue Addr, Disp;
17096 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
17097 DAG.getConstant(10, dl, MVT::i32));
17098 Disp = DAG.getNode(ISD::SUB, dl, MVT::i32, FPtr, Addr);
17100 // This is storing the opcode for MOV32ri.
17101 const unsigned char MOV32ri = 0xB8; // X86::MOV32ri's opcode byte.
17102 const unsigned char N86Reg = TRI->getEncodingValue(NestReg) & 0x7;
17103 OutChains[0] = DAG.getStore(Root, dl,
17104 DAG.getConstant(MOV32ri|N86Reg, dl, MVT::i8),
17105 Trmp, MachinePointerInfo(TrmpAddr),
17108 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
17109 DAG.getConstant(1, dl, MVT::i32));
17110 OutChains[1] = DAG.getStore(Root, dl, Nest, Addr,
17111 MachinePointerInfo(TrmpAddr, 1),
17114 const unsigned char JMP = 0xE9; // jmp <32bit dst> opcode.
17115 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
17116 DAG.getConstant(5, dl, MVT::i32));
17117 OutChains[2] = DAG.getStore(Root, dl, DAG.getConstant(JMP, dl, MVT::i8),
17118 Addr, MachinePointerInfo(TrmpAddr, 5),
17121 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
17122 DAG.getConstant(6, dl, MVT::i32));
17123 OutChains[3] = DAG.getStore(Root, dl, Disp, Addr,
17124 MachinePointerInfo(TrmpAddr, 6),
17127 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other, OutChains);
17131 SDValue X86TargetLowering::LowerFLT_ROUNDS_(SDValue Op,
17132 SelectionDAG &DAG) const {
17134 The rounding mode is in bits 11:10 of FPSR, and has the following
17136 00 Round to nearest
17141 FLT_ROUNDS, on the other hand, expects the following:
17148 To perform the conversion, we do:
17149 (((((FPSR & 0x800) >> 11) | ((FPSR & 0x400) >> 9)) + 1) & 3)
17152 MachineFunction &MF = DAG.getMachineFunction();
17153 const TargetFrameLowering &TFI = *Subtarget->getFrameLowering();
17154 unsigned StackAlignment = TFI.getStackAlignment();
17155 MVT VT = Op.getSimpleValueType();
17158 // Save FP Control Word to stack slot
17159 int SSFI = MF.getFrameInfo()->CreateStackObject(2, StackAlignment, false);
17160 SDValue StackSlot =
17161 DAG.getFrameIndex(SSFI, getPointerTy(DAG.getDataLayout()));
17163 MachineMemOperand *MMO =
17164 MF.getMachineMemOperand(MachinePointerInfo::getFixedStack(MF, SSFI),
17165 MachineMemOperand::MOStore, 2, 2);
17167 SDValue Ops[] = { DAG.getEntryNode(), StackSlot };
17168 SDValue Chain = DAG.getMemIntrinsicNode(X86ISD::FNSTCW16m, DL,
17169 DAG.getVTList(MVT::Other),
17170 Ops, MVT::i16, MMO);
17172 // Load FP Control Word from stack slot
17173 SDValue CWD = DAG.getLoad(MVT::i16, DL, Chain, StackSlot,
17174 MachinePointerInfo(), false, false, false, 0);
17176 // Transform as necessary
17178 DAG.getNode(ISD::SRL, DL, MVT::i16,
17179 DAG.getNode(ISD::AND, DL, MVT::i16,
17180 CWD, DAG.getConstant(0x800, DL, MVT::i16)),
17181 DAG.getConstant(11, DL, MVT::i8));
17183 DAG.getNode(ISD::SRL, DL, MVT::i16,
17184 DAG.getNode(ISD::AND, DL, MVT::i16,
17185 CWD, DAG.getConstant(0x400, DL, MVT::i16)),
17186 DAG.getConstant(9, DL, MVT::i8));
17189 DAG.getNode(ISD::AND, DL, MVT::i16,
17190 DAG.getNode(ISD::ADD, DL, MVT::i16,
17191 DAG.getNode(ISD::OR, DL, MVT::i16, CWD1, CWD2),
17192 DAG.getConstant(1, DL, MVT::i16)),
17193 DAG.getConstant(3, DL, MVT::i16));
17195 return DAG.getNode((VT.getSizeInBits() < 16 ?
17196 ISD::TRUNCATE : ISD::ZERO_EXTEND), DL, VT, RetVal);
17199 static SDValue LowerCTLZ(SDValue Op, SelectionDAG &DAG) {
17200 MVT VT = Op.getSimpleValueType();
17202 unsigned NumBits = VT.getSizeInBits();
17205 Op = Op.getOperand(0);
17206 if (VT == MVT::i8) {
17207 // Zero extend to i32 since there is not an i8 bsr.
17209 Op = DAG.getNode(ISD::ZERO_EXTEND, dl, OpVT, Op);
17212 // Issue a bsr (scan bits in reverse) which also sets EFLAGS.
17213 SDVTList VTs = DAG.getVTList(OpVT, MVT::i32);
17214 Op = DAG.getNode(X86ISD::BSR, dl, VTs, Op);
17216 // If src is zero (i.e. bsr sets ZF), returns NumBits.
17219 DAG.getConstant(NumBits + NumBits - 1, dl, OpVT),
17220 DAG.getConstant(X86::COND_E, dl, MVT::i8),
17223 Op = DAG.getNode(X86ISD::CMOV, dl, OpVT, Ops);
17225 // Finally xor with NumBits-1.
17226 Op = DAG.getNode(ISD::XOR, dl, OpVT, Op,
17227 DAG.getConstant(NumBits - 1, dl, OpVT));
17230 Op = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op);
17234 static SDValue LowerCTLZ_ZERO_UNDEF(SDValue Op, SelectionDAG &DAG) {
17235 MVT VT = Op.getSimpleValueType();
17237 unsigned NumBits = VT.getSizeInBits();
17240 Op = Op.getOperand(0);
17241 if (VT == MVT::i8) {
17242 // Zero extend to i32 since there is not an i8 bsr.
17244 Op = DAG.getNode(ISD::ZERO_EXTEND, dl, OpVT, Op);
17247 // Issue a bsr (scan bits in reverse).
17248 SDVTList VTs = DAG.getVTList(OpVT, MVT::i32);
17249 Op = DAG.getNode(X86ISD::BSR, dl, VTs, Op);
17251 // And xor with NumBits-1.
17252 Op = DAG.getNode(ISD::XOR, dl, OpVT, Op,
17253 DAG.getConstant(NumBits - 1, dl, OpVT));
17256 Op = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op);
17260 static SDValue LowerCTTZ(SDValue Op, SelectionDAG &DAG) {
17261 MVT VT = Op.getSimpleValueType();
17262 unsigned NumBits = VT.getSizeInBits();
17264 Op = Op.getOperand(0);
17266 // Issue a bsf (scan bits forward) which also sets EFLAGS.
17267 SDVTList VTs = DAG.getVTList(VT, MVT::i32);
17268 Op = DAG.getNode(X86ISD::BSF, dl, VTs, Op);
17270 // If src is zero (i.e. bsf sets ZF), returns NumBits.
17273 DAG.getConstant(NumBits, dl, VT),
17274 DAG.getConstant(X86::COND_E, dl, MVT::i8),
17277 return DAG.getNode(X86ISD::CMOV, dl, VT, Ops);
17280 // Lower256IntArith - Break a 256-bit integer operation into two new 128-bit
17281 // ones, and then concatenate the result back.
17282 static SDValue Lower256IntArith(SDValue Op, SelectionDAG &DAG) {
17283 MVT VT = Op.getSimpleValueType();
17285 assert(VT.is256BitVector() && VT.isInteger() &&
17286 "Unsupported value type for operation");
17288 unsigned NumElems = VT.getVectorNumElements();
17291 // Extract the LHS vectors
17292 SDValue LHS = Op.getOperand(0);
17293 SDValue LHS1 = Extract128BitVector(LHS, 0, DAG, dl);
17294 SDValue LHS2 = Extract128BitVector(LHS, NumElems/2, DAG, dl);
17296 // Extract the RHS vectors
17297 SDValue RHS = Op.getOperand(1);
17298 SDValue RHS1 = Extract128BitVector(RHS, 0, DAG, dl);
17299 SDValue RHS2 = Extract128BitVector(RHS, NumElems/2, DAG, dl);
17301 MVT EltVT = VT.getVectorElementType();
17302 MVT NewVT = MVT::getVectorVT(EltVT, NumElems/2);
17304 return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT,
17305 DAG.getNode(Op.getOpcode(), dl, NewVT, LHS1, RHS1),
17306 DAG.getNode(Op.getOpcode(), dl, NewVT, LHS2, RHS2));
17309 static SDValue LowerADD(SDValue Op, SelectionDAG &DAG) {
17310 if (Op.getValueType() == MVT::i1)
17311 return DAG.getNode(ISD::XOR, SDLoc(Op), Op.getValueType(),
17312 Op.getOperand(0), Op.getOperand(1));
17313 assert(Op.getSimpleValueType().is256BitVector() &&
17314 Op.getSimpleValueType().isInteger() &&
17315 "Only handle AVX 256-bit vector integer operation");
17316 return Lower256IntArith(Op, DAG);
17319 static SDValue LowerSUB(SDValue Op, SelectionDAG &DAG) {
17320 if (Op.getValueType() == MVT::i1)
17321 return DAG.getNode(ISD::XOR, SDLoc(Op), Op.getValueType(),
17322 Op.getOperand(0), Op.getOperand(1));
17323 assert(Op.getSimpleValueType().is256BitVector() &&
17324 Op.getSimpleValueType().isInteger() &&
17325 "Only handle AVX 256-bit vector integer operation");
17326 return Lower256IntArith(Op, DAG);
17329 static SDValue LowerMINMAX(SDValue Op, SelectionDAG &DAG) {
17330 assert(Op.getSimpleValueType().is256BitVector() &&
17331 Op.getSimpleValueType().isInteger() &&
17332 "Only handle AVX 256-bit vector integer operation");
17333 return Lower256IntArith(Op, DAG);
17336 static SDValue LowerMUL(SDValue Op, const X86Subtarget *Subtarget,
17337 SelectionDAG &DAG) {
17339 MVT VT = Op.getSimpleValueType();
17342 return DAG.getNode(ISD::AND, dl, VT, Op.getOperand(0), Op.getOperand(1));
17344 // Decompose 256-bit ops into smaller 128-bit ops.
17345 if (VT.is256BitVector() && !Subtarget->hasInt256())
17346 return Lower256IntArith(Op, DAG);
17348 SDValue A = Op.getOperand(0);
17349 SDValue B = Op.getOperand(1);
17351 // Lower v16i8/v32i8 mul as promotion to v8i16/v16i16 vector
17352 // pairs, multiply and truncate.
17353 if (VT == MVT::v16i8 || VT == MVT::v32i8) {
17354 if (Subtarget->hasInt256()) {
17355 if (VT == MVT::v32i8) {
17356 MVT SubVT = MVT::getVectorVT(MVT::i8, VT.getVectorNumElements() / 2);
17357 SDValue Lo = DAG.getIntPtrConstant(0, dl);
17358 SDValue Hi = DAG.getIntPtrConstant(VT.getVectorNumElements() / 2, dl);
17359 SDValue ALo = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, SubVT, A, Lo);
17360 SDValue BLo = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, SubVT, B, Lo);
17361 SDValue AHi = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, SubVT, A, Hi);
17362 SDValue BHi = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, SubVT, B, Hi);
17363 return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT,
17364 DAG.getNode(ISD::MUL, dl, SubVT, ALo, BLo),
17365 DAG.getNode(ISD::MUL, dl, SubVT, AHi, BHi));
17368 MVT ExVT = MVT::getVectorVT(MVT::i16, VT.getVectorNumElements());
17369 return DAG.getNode(
17370 ISD::TRUNCATE, dl, VT,
17371 DAG.getNode(ISD::MUL, dl, ExVT,
17372 DAG.getNode(ISD::SIGN_EXTEND, dl, ExVT, A),
17373 DAG.getNode(ISD::SIGN_EXTEND, dl, ExVT, B)));
17376 assert(VT == MVT::v16i8 &&
17377 "Pre-AVX2 support only supports v16i8 multiplication");
17378 MVT ExVT = MVT::v8i16;
17380 // Extract the lo parts and sign extend to i16
17382 if (Subtarget->hasSSE41()) {
17383 ALo = DAG.getNode(X86ISD::VSEXT, dl, ExVT, A);
17384 BLo = DAG.getNode(X86ISD::VSEXT, dl, ExVT, B);
17386 const int ShufMask[] = {-1, 0, -1, 1, -1, 2, -1, 3,
17387 -1, 4, -1, 5, -1, 6, -1, 7};
17388 ALo = DAG.getVectorShuffle(VT, dl, A, A, ShufMask);
17389 BLo = DAG.getVectorShuffle(VT, dl, B, B, ShufMask);
17390 ALo = DAG.getBitcast(ExVT, ALo);
17391 BLo = DAG.getBitcast(ExVT, BLo);
17392 ALo = DAG.getNode(ISD::SRA, dl, ExVT, ALo, DAG.getConstant(8, dl, ExVT));
17393 BLo = DAG.getNode(ISD::SRA, dl, ExVT, BLo, DAG.getConstant(8, dl, ExVT));
17396 // Extract the hi parts and sign extend to i16
17398 if (Subtarget->hasSSE41()) {
17399 const int ShufMask[] = {8, 9, 10, 11, 12, 13, 14, 15,
17400 -1, -1, -1, -1, -1, -1, -1, -1};
17401 AHi = DAG.getVectorShuffle(VT, dl, A, A, ShufMask);
17402 BHi = DAG.getVectorShuffle(VT, dl, B, B, ShufMask);
17403 AHi = DAG.getNode(X86ISD::VSEXT, dl, ExVT, AHi);
17404 BHi = DAG.getNode(X86ISD::VSEXT, dl, ExVT, BHi);
17406 const int ShufMask[] = {-1, 8, -1, 9, -1, 10, -1, 11,
17407 -1, 12, -1, 13, -1, 14, -1, 15};
17408 AHi = DAG.getVectorShuffle(VT, dl, A, A, ShufMask);
17409 BHi = DAG.getVectorShuffle(VT, dl, B, B, ShufMask);
17410 AHi = DAG.getBitcast(ExVT, AHi);
17411 BHi = DAG.getBitcast(ExVT, BHi);
17412 AHi = DAG.getNode(ISD::SRA, dl, ExVT, AHi, DAG.getConstant(8, dl, ExVT));
17413 BHi = DAG.getNode(ISD::SRA, dl, ExVT, BHi, DAG.getConstant(8, dl, ExVT));
17416 // Multiply, mask the lower 8bits of the lo/hi results and pack
17417 SDValue RLo = DAG.getNode(ISD::MUL, dl, ExVT, ALo, BLo);
17418 SDValue RHi = DAG.getNode(ISD::MUL, dl, ExVT, AHi, BHi);
17419 RLo = DAG.getNode(ISD::AND, dl, ExVT, RLo, DAG.getConstant(255, dl, ExVT));
17420 RHi = DAG.getNode(ISD::AND, dl, ExVT, RHi, DAG.getConstant(255, dl, ExVT));
17421 return DAG.getNode(X86ISD::PACKUS, dl, VT, RLo, RHi);
17424 // Lower v4i32 mul as 2x shuffle, 2x pmuludq, 2x shuffle.
17425 if (VT == MVT::v4i32) {
17426 assert(Subtarget->hasSSE2() && !Subtarget->hasSSE41() &&
17427 "Should not custom lower when pmuldq is available!");
17429 // Extract the odd parts.
17430 static const int UnpackMask[] = { 1, -1, 3, -1 };
17431 SDValue Aodds = DAG.getVectorShuffle(VT, dl, A, A, UnpackMask);
17432 SDValue Bodds = DAG.getVectorShuffle(VT, dl, B, B, UnpackMask);
17434 // Multiply the even parts.
17435 SDValue Evens = DAG.getNode(X86ISD::PMULUDQ, dl, MVT::v2i64, A, B);
17436 // Now multiply odd parts.
17437 SDValue Odds = DAG.getNode(X86ISD::PMULUDQ, dl, MVT::v2i64, Aodds, Bodds);
17439 Evens = DAG.getBitcast(VT, Evens);
17440 Odds = DAG.getBitcast(VT, Odds);
17442 // Merge the two vectors back together with a shuffle. This expands into 2
17444 static const int ShufMask[] = { 0, 4, 2, 6 };
17445 return DAG.getVectorShuffle(VT, dl, Evens, Odds, ShufMask);
17448 assert((VT == MVT::v2i64 || VT == MVT::v4i64 || VT == MVT::v8i64) &&
17449 "Only know how to lower V2I64/V4I64/V8I64 multiply");
17451 // Ahi = psrlqi(a, 32);
17452 // Bhi = psrlqi(b, 32);
17454 // AloBlo = pmuludq(a, b);
17455 // AloBhi = pmuludq(a, Bhi);
17456 // AhiBlo = pmuludq(Ahi, b);
17458 // AloBhi = psllqi(AloBhi, 32);
17459 // AhiBlo = psllqi(AhiBlo, 32);
17460 // return AloBlo + AloBhi + AhiBlo;
17462 SDValue Ahi = getTargetVShiftByConstNode(X86ISD::VSRLI, dl, VT, A, 32, DAG);
17463 SDValue Bhi = getTargetVShiftByConstNode(X86ISD::VSRLI, dl, VT, B, 32, DAG);
17465 SDValue AhiBlo = Ahi;
17466 SDValue AloBhi = Bhi;
17467 // Bit cast to 32-bit vectors for MULUDQ
17468 EVT MulVT = (VT == MVT::v2i64) ? MVT::v4i32 :
17469 (VT == MVT::v4i64) ? MVT::v8i32 : MVT::v16i32;
17470 A = DAG.getBitcast(MulVT, A);
17471 B = DAG.getBitcast(MulVT, B);
17472 Ahi = DAG.getBitcast(MulVT, Ahi);
17473 Bhi = DAG.getBitcast(MulVT, Bhi);
17475 SDValue AloBlo = DAG.getNode(X86ISD::PMULUDQ, dl, VT, A, B);
17476 // After shifting right const values the result may be all-zero.
17477 if (!ISD::isBuildVectorAllZeros(Ahi.getNode())) {
17478 AhiBlo = DAG.getNode(X86ISD::PMULUDQ, dl, VT, Ahi, B);
17479 AhiBlo = getTargetVShiftByConstNode(X86ISD::VSHLI, dl, VT, AhiBlo, 32, DAG);
17481 if (!ISD::isBuildVectorAllZeros(Bhi.getNode())) {
17482 AloBhi = DAG.getNode(X86ISD::PMULUDQ, dl, VT, A, Bhi);
17483 AloBhi = getTargetVShiftByConstNode(X86ISD::VSHLI, dl, VT, AloBhi, 32, DAG);
17486 SDValue Res = DAG.getNode(ISD::ADD, dl, VT, AloBlo, AloBhi);
17487 return DAG.getNode(ISD::ADD, dl, VT, Res, AhiBlo);
17490 SDValue X86TargetLowering::LowerWin64_i128OP(SDValue Op, SelectionDAG &DAG) const {
17491 assert(Subtarget->isTargetWin64() && "Unexpected target");
17492 EVT VT = Op.getValueType();
17493 assert(VT.isInteger() && VT.getSizeInBits() == 128 &&
17494 "Unexpected return type for lowering");
17498 switch (Op->getOpcode()) {
17499 default: llvm_unreachable("Unexpected request for libcall!");
17500 case ISD::SDIV: isSigned = true; LC = RTLIB::SDIV_I128; break;
17501 case ISD::UDIV: isSigned = false; LC = RTLIB::UDIV_I128; break;
17502 case ISD::SREM: isSigned = true; LC = RTLIB::SREM_I128; break;
17503 case ISD::UREM: isSigned = false; LC = RTLIB::UREM_I128; break;
17504 case ISD::SDIVREM: isSigned = true; LC = RTLIB::SDIVREM_I128; break;
17505 case ISD::UDIVREM: isSigned = false; LC = RTLIB::UDIVREM_I128; break;
17509 SDValue InChain = DAG.getEntryNode();
17511 TargetLowering::ArgListTy Args;
17512 TargetLowering::ArgListEntry Entry;
17513 for (unsigned i = 0, e = Op->getNumOperands(); i != e; ++i) {
17514 EVT ArgVT = Op->getOperand(i).getValueType();
17515 assert(ArgVT.isInteger() && ArgVT.getSizeInBits() == 128 &&
17516 "Unexpected argument type for lowering");
17517 SDValue StackPtr = DAG.CreateStackTemporary(ArgVT, 16);
17518 Entry.Node = StackPtr;
17519 InChain = DAG.getStore(InChain, dl, Op->getOperand(i), StackPtr, MachinePointerInfo(),
17521 Type *ArgTy = ArgVT.getTypeForEVT(*DAG.getContext());
17522 Entry.Ty = PointerType::get(ArgTy,0);
17523 Entry.isSExt = false;
17524 Entry.isZExt = false;
17525 Args.push_back(Entry);
17528 SDValue Callee = DAG.getExternalSymbol(getLibcallName(LC),
17529 getPointerTy(DAG.getDataLayout()));
17531 TargetLowering::CallLoweringInfo CLI(DAG);
17532 CLI.setDebugLoc(dl).setChain(InChain)
17533 .setCallee(getLibcallCallingConv(LC),
17534 static_cast<EVT>(MVT::v2i64).getTypeForEVT(*DAG.getContext()),
17535 Callee, std::move(Args), 0)
17536 .setInRegister().setSExtResult(isSigned).setZExtResult(!isSigned);
17538 std::pair<SDValue, SDValue> CallInfo = LowerCallTo(CLI);
17539 return DAG.getBitcast(VT, CallInfo.first);
17542 static SDValue LowerMUL_LOHI(SDValue Op, const X86Subtarget *Subtarget,
17543 SelectionDAG &DAG) {
17544 SDValue Op0 = Op.getOperand(0), Op1 = Op.getOperand(1);
17545 EVT VT = Op0.getValueType();
17548 assert((VT == MVT::v4i32 && Subtarget->hasSSE2()) ||
17549 (VT == MVT::v8i32 && Subtarget->hasInt256()));
17551 // PMULxD operations multiply each even value (starting at 0) of LHS with
17552 // the related value of RHS and produce a widen result.
17553 // E.g., PMULUDQ <4 x i32> <a|b|c|d>, <4 x i32> <e|f|g|h>
17554 // => <2 x i64> <ae|cg>
17556 // In other word, to have all the results, we need to perform two PMULxD:
17557 // 1. one with the even values.
17558 // 2. one with the odd values.
17559 // To achieve #2, with need to place the odd values at an even position.
17561 // Place the odd value at an even position (basically, shift all values 1
17562 // step to the left):
17563 const int Mask[] = {1, -1, 3, -1, 5, -1, 7, -1};
17564 // <a|b|c|d> => <b|undef|d|undef>
17565 SDValue Odd0 = DAG.getVectorShuffle(VT, dl, Op0, Op0, Mask);
17566 // <e|f|g|h> => <f|undef|h|undef>
17567 SDValue Odd1 = DAG.getVectorShuffle(VT, dl, Op1, Op1, Mask);
17569 // Emit two multiplies, one for the lower 2 ints and one for the higher 2
17571 MVT MulVT = VT == MVT::v4i32 ? MVT::v2i64 : MVT::v4i64;
17572 bool IsSigned = Op->getOpcode() == ISD::SMUL_LOHI;
17574 (!IsSigned || !Subtarget->hasSSE41()) ? X86ISD::PMULUDQ : X86ISD::PMULDQ;
17575 // PMULUDQ <4 x i32> <a|b|c|d>, <4 x i32> <e|f|g|h>
17576 // => <2 x i64> <ae|cg>
17577 SDValue Mul1 = DAG.getBitcast(VT, DAG.getNode(Opcode, dl, MulVT, Op0, Op1));
17578 // PMULUDQ <4 x i32> <b|undef|d|undef>, <4 x i32> <f|undef|h|undef>
17579 // => <2 x i64> <bf|dh>
17580 SDValue Mul2 = DAG.getBitcast(VT, DAG.getNode(Opcode, dl, MulVT, Odd0, Odd1));
17582 // Shuffle it back into the right order.
17583 SDValue Highs, Lows;
17584 if (VT == MVT::v8i32) {
17585 const int HighMask[] = {1, 9, 3, 11, 5, 13, 7, 15};
17586 Highs = DAG.getVectorShuffle(VT, dl, Mul1, Mul2, HighMask);
17587 const int LowMask[] = {0, 8, 2, 10, 4, 12, 6, 14};
17588 Lows = DAG.getVectorShuffle(VT, dl, Mul1, Mul2, LowMask);
17590 const int HighMask[] = {1, 5, 3, 7};
17591 Highs = DAG.getVectorShuffle(VT, dl, Mul1, Mul2, HighMask);
17592 const int LowMask[] = {0, 4, 2, 6};
17593 Lows = DAG.getVectorShuffle(VT, dl, Mul1, Mul2, LowMask);
17596 // If we have a signed multiply but no PMULDQ fix up the high parts of a
17597 // unsigned multiply.
17598 if (IsSigned && !Subtarget->hasSSE41()) {
17599 SDValue ShAmt = DAG.getConstant(
17601 DAG.getTargetLoweringInfo().getShiftAmountTy(VT, DAG.getDataLayout()));
17602 SDValue T1 = DAG.getNode(ISD::AND, dl, VT,
17603 DAG.getNode(ISD::SRA, dl, VT, Op0, ShAmt), Op1);
17604 SDValue T2 = DAG.getNode(ISD::AND, dl, VT,
17605 DAG.getNode(ISD::SRA, dl, VT, Op1, ShAmt), Op0);
17607 SDValue Fixup = DAG.getNode(ISD::ADD, dl, VT, T1, T2);
17608 Highs = DAG.getNode(ISD::SUB, dl, VT, Highs, Fixup);
17611 // The first result of MUL_LOHI is actually the low value, followed by the
17613 SDValue Ops[] = {Lows, Highs};
17614 return DAG.getMergeValues(Ops, dl);
17617 // Return true if the required (according to Opcode) shift-imm form is natively
17618 // supported by the Subtarget
17619 static bool SupportedVectorShiftWithImm(MVT VT, const X86Subtarget *Subtarget,
17621 if (VT.getScalarSizeInBits() < 16)
17624 if (VT.is512BitVector() &&
17625 (VT.getScalarSizeInBits() > 16 || Subtarget->hasBWI()))
17628 bool LShift = VT.is128BitVector() ||
17629 (VT.is256BitVector() && Subtarget->hasInt256());
17631 bool AShift = LShift && (Subtarget->hasVLX() ||
17632 (VT != MVT::v2i64 && VT != MVT::v4i64));
17633 return (Opcode == ISD::SRA) ? AShift : LShift;
17636 // The shift amount is a variable, but it is the same for all vector lanes.
17637 // These instructions are defined together with shift-immediate.
17639 bool SupportedVectorShiftWithBaseAmnt(MVT VT, const X86Subtarget *Subtarget,
17641 return SupportedVectorShiftWithImm(VT, Subtarget, Opcode);
17644 // Return true if the required (according to Opcode) variable-shift form is
17645 // natively supported by the Subtarget
17646 static bool SupportedVectorVarShift(MVT VT, const X86Subtarget *Subtarget,
17649 if (!Subtarget->hasInt256() || VT.getScalarSizeInBits() < 16)
17652 // vXi16 supported only on AVX-512, BWI
17653 if (VT.getScalarSizeInBits() == 16 && !Subtarget->hasBWI())
17656 if (VT.is512BitVector() || Subtarget->hasVLX())
17659 bool LShift = VT.is128BitVector() || VT.is256BitVector();
17660 bool AShift = LShift && VT != MVT::v2i64 && VT != MVT::v4i64;
17661 return (Opcode == ISD::SRA) ? AShift : LShift;
17664 static SDValue LowerScalarImmediateShift(SDValue Op, SelectionDAG &DAG,
17665 const X86Subtarget *Subtarget) {
17666 MVT VT = Op.getSimpleValueType();
17668 SDValue R = Op.getOperand(0);
17669 SDValue Amt = Op.getOperand(1);
17671 unsigned X86Opc = (Op.getOpcode() == ISD::SHL) ? X86ISD::VSHLI :
17672 (Op.getOpcode() == ISD::SRL) ? X86ISD::VSRLI : X86ISD::VSRAI;
17674 auto ArithmeticShiftRight64 = [&](uint64_t ShiftAmt) {
17675 assert((VT == MVT::v2i64 || VT == MVT::v4i64) && "Unexpected SRA type");
17676 MVT ExVT = MVT::getVectorVT(MVT::i32, VT.getVectorNumElements() * 2);
17677 SDValue Ex = DAG.getBitcast(ExVT, R);
17679 if (ShiftAmt >= 32) {
17680 // Splat sign to upper i32 dst, and SRA upper i32 src to lower i32.
17682 getTargetVShiftByConstNode(X86ISD::VSRAI, dl, ExVT, Ex, 31, DAG);
17683 SDValue Lower = getTargetVShiftByConstNode(X86ISD::VSRAI, dl, ExVT, Ex,
17684 ShiftAmt - 32, DAG);
17685 if (VT == MVT::v2i64)
17686 Ex = DAG.getVectorShuffle(ExVT, dl, Upper, Lower, {5, 1, 7, 3});
17687 if (VT == MVT::v4i64)
17688 Ex = DAG.getVectorShuffle(ExVT, dl, Upper, Lower,
17689 {9, 1, 11, 3, 13, 5, 15, 7});
17691 // SRA upper i32, SHL whole i64 and select lower i32.
17692 SDValue Upper = getTargetVShiftByConstNode(X86ISD::VSRAI, dl, ExVT, Ex,
17695 getTargetVShiftByConstNode(X86ISD::VSRLI, dl, VT, R, ShiftAmt, DAG);
17696 Lower = DAG.getBitcast(ExVT, Lower);
17697 if (VT == MVT::v2i64)
17698 Ex = DAG.getVectorShuffle(ExVT, dl, Upper, Lower, {4, 1, 6, 3});
17699 if (VT == MVT::v4i64)
17700 Ex = DAG.getVectorShuffle(ExVT, dl, Upper, Lower,
17701 {8, 1, 10, 3, 12, 5, 14, 7});
17703 return DAG.getBitcast(VT, Ex);
17706 // Optimize shl/srl/sra with constant shift amount.
17707 if (auto *BVAmt = dyn_cast<BuildVectorSDNode>(Amt)) {
17708 if (auto *ShiftConst = BVAmt->getConstantSplatNode()) {
17709 uint64_t ShiftAmt = ShiftConst->getZExtValue();
17711 if (SupportedVectorShiftWithImm(VT, Subtarget, Op.getOpcode()))
17712 return getTargetVShiftByConstNode(X86Opc, dl, VT, R, ShiftAmt, DAG);
17714 // i64 SRA needs to be performed as partial shifts.
17715 if ((VT == MVT::v2i64 || (Subtarget->hasInt256() && VT == MVT::v4i64)) &&
17716 Op.getOpcode() == ISD::SRA)
17717 return ArithmeticShiftRight64(ShiftAmt);
17719 if (VT == MVT::v16i8 || (Subtarget->hasInt256() && VT == MVT::v32i8)) {
17720 unsigned NumElts = VT.getVectorNumElements();
17721 MVT ShiftVT = MVT::getVectorVT(MVT::i16, NumElts / 2);
17723 if (Op.getOpcode() == ISD::SHL) {
17724 // Simple i8 add case
17726 return DAG.getNode(ISD::ADD, dl, VT, R, R);
17728 // Make a large shift.
17729 SDValue SHL = getTargetVShiftByConstNode(X86ISD::VSHLI, dl, ShiftVT,
17731 SHL = DAG.getBitcast(VT, SHL);
17732 // Zero out the rightmost bits.
17733 SmallVector<SDValue, 32> V(
17734 NumElts, DAG.getConstant(uint8_t(-1U << ShiftAmt), dl, MVT::i8));
17735 return DAG.getNode(ISD::AND, dl, VT, SHL,
17736 DAG.getNode(ISD::BUILD_VECTOR, dl, VT, V));
17738 if (Op.getOpcode() == ISD::SRL) {
17739 // Make a large shift.
17740 SDValue SRL = getTargetVShiftByConstNode(X86ISD::VSRLI, dl, ShiftVT,
17742 SRL = DAG.getBitcast(VT, SRL);
17743 // Zero out the leftmost bits.
17744 SmallVector<SDValue, 32> V(
17745 NumElts, DAG.getConstant(uint8_t(-1U) >> ShiftAmt, dl, MVT::i8));
17746 return DAG.getNode(ISD::AND, dl, VT, SRL,
17747 DAG.getNode(ISD::BUILD_VECTOR, dl, VT, V));
17749 if (Op.getOpcode() == ISD::SRA) {
17750 if (ShiftAmt == 7) {
17751 // ashr(R, 7) === cmp_slt(R, 0)
17752 SDValue Zeros = getZeroVector(VT, Subtarget, DAG, dl);
17753 return DAG.getNode(X86ISD::PCMPGT, dl, VT, Zeros, R);
17756 // ashr(R, Amt) === sub(xor(lshr(R, Amt), Mask), Mask)
17757 SDValue Res = DAG.getNode(ISD::SRL, dl, VT, R, Amt);
17758 SmallVector<SDValue, 32> V(NumElts,
17759 DAG.getConstant(128 >> ShiftAmt, dl,
17761 SDValue Mask = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, V);
17762 Res = DAG.getNode(ISD::XOR, dl, VT, Res, Mask);
17763 Res = DAG.getNode(ISD::SUB, dl, VT, Res, Mask);
17766 llvm_unreachable("Unknown shift opcode.");
17771 // Special case in 32-bit mode, where i64 is expanded into high and low parts.
17772 if (!Subtarget->is64Bit() &&
17773 (VT == MVT::v2i64 || (Subtarget->hasInt256() && VT == MVT::v4i64))) {
17775 // Peek through any splat that was introduced for i64 shift vectorization.
17776 int SplatIndex = -1;
17777 if (ShuffleVectorSDNode *SVN = dyn_cast<ShuffleVectorSDNode>(Amt.getNode()))
17778 if (SVN->isSplat()) {
17779 SplatIndex = SVN->getSplatIndex();
17780 Amt = Amt.getOperand(0);
17781 assert(SplatIndex < (int)VT.getVectorNumElements() &&
17782 "Splat shuffle referencing second operand");
17785 if (Amt.getOpcode() != ISD::BITCAST ||
17786 Amt.getOperand(0).getOpcode() != ISD::BUILD_VECTOR)
17789 Amt = Amt.getOperand(0);
17790 unsigned Ratio = Amt.getSimpleValueType().getVectorNumElements() /
17791 VT.getVectorNumElements();
17792 unsigned RatioInLog2 = Log2_32_Ceil(Ratio);
17793 uint64_t ShiftAmt = 0;
17794 unsigned BaseOp = (SplatIndex < 0 ? 0 : SplatIndex * Ratio);
17795 for (unsigned i = 0; i != Ratio; ++i) {
17796 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Amt.getOperand(i + BaseOp));
17800 ShiftAmt |= C->getZExtValue() << (i * (1 << (6 - RatioInLog2)));
17803 // Check remaining shift amounts (if not a splat).
17804 if (SplatIndex < 0) {
17805 for (unsigned i = Ratio; i != Amt.getNumOperands(); i += Ratio) {
17806 uint64_t ShAmt = 0;
17807 for (unsigned j = 0; j != Ratio; ++j) {
17808 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Amt.getOperand(i + j));
17812 ShAmt |= C->getZExtValue() << (j * (1 << (6 - RatioInLog2)));
17814 if (ShAmt != ShiftAmt)
17819 if (SupportedVectorShiftWithImm(VT, Subtarget, Op.getOpcode()))
17820 return getTargetVShiftByConstNode(X86Opc, dl, VT, R, ShiftAmt, DAG);
17822 if (Op.getOpcode() == ISD::SRA)
17823 return ArithmeticShiftRight64(ShiftAmt);
17829 static SDValue LowerScalarVariableShift(SDValue Op, SelectionDAG &DAG,
17830 const X86Subtarget* Subtarget) {
17831 MVT VT = Op.getSimpleValueType();
17833 SDValue R = Op.getOperand(0);
17834 SDValue Amt = Op.getOperand(1);
17836 unsigned X86OpcI = (Op.getOpcode() == ISD::SHL) ? X86ISD::VSHLI :
17837 (Op.getOpcode() == ISD::SRL) ? X86ISD::VSRLI : X86ISD::VSRAI;
17839 unsigned X86OpcV = (Op.getOpcode() == ISD::SHL) ? X86ISD::VSHL :
17840 (Op.getOpcode() == ISD::SRL) ? X86ISD::VSRL : X86ISD::VSRA;
17842 if (SupportedVectorShiftWithBaseAmnt(VT, Subtarget, Op.getOpcode())) {
17844 EVT EltVT = VT.getVectorElementType();
17846 if (BuildVectorSDNode *BV = dyn_cast<BuildVectorSDNode>(Amt)) {
17847 // Check if this build_vector node is doing a splat.
17848 // If so, then set BaseShAmt equal to the splat value.
17849 BaseShAmt = BV->getSplatValue();
17850 if (BaseShAmt && BaseShAmt.getOpcode() == ISD::UNDEF)
17851 BaseShAmt = SDValue();
17853 if (Amt.getOpcode() == ISD::EXTRACT_SUBVECTOR)
17854 Amt = Amt.getOperand(0);
17856 ShuffleVectorSDNode *SVN = dyn_cast<ShuffleVectorSDNode>(Amt);
17857 if (SVN && SVN->isSplat()) {
17858 unsigned SplatIdx = (unsigned)SVN->getSplatIndex();
17859 SDValue InVec = Amt.getOperand(0);
17860 if (InVec.getOpcode() == ISD::BUILD_VECTOR) {
17861 assert((SplatIdx < InVec.getValueType().getVectorNumElements()) &&
17862 "Unexpected shuffle index found!");
17863 BaseShAmt = InVec.getOperand(SplatIdx);
17864 } else if (InVec.getOpcode() == ISD::INSERT_VECTOR_ELT) {
17865 if (ConstantSDNode *C =
17866 dyn_cast<ConstantSDNode>(InVec.getOperand(2))) {
17867 if (C->getZExtValue() == SplatIdx)
17868 BaseShAmt = InVec.getOperand(1);
17873 // Avoid introducing an extract element from a shuffle.
17874 BaseShAmt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, EltVT, InVec,
17875 DAG.getIntPtrConstant(SplatIdx, dl));
17879 if (BaseShAmt.getNode()) {
17880 assert(EltVT.bitsLE(MVT::i64) && "Unexpected element type!");
17881 if (EltVT != MVT::i64 && EltVT.bitsGT(MVT::i32))
17882 BaseShAmt = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i64, BaseShAmt);
17883 else if (EltVT.bitsLT(MVT::i32))
17884 BaseShAmt = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, BaseShAmt);
17886 return getTargetVShiftNode(X86OpcI, dl, VT, R, BaseShAmt, DAG);
17890 // Special case in 32-bit mode, where i64 is expanded into high and low parts.
17891 if (!Subtarget->is64Bit() && VT == MVT::v2i64 &&
17892 Amt.getOpcode() == ISD::BITCAST &&
17893 Amt.getOperand(0).getOpcode() == ISD::BUILD_VECTOR) {
17894 Amt = Amt.getOperand(0);
17895 unsigned Ratio = Amt.getSimpleValueType().getVectorNumElements() /
17896 VT.getVectorNumElements();
17897 std::vector<SDValue> Vals(Ratio);
17898 for (unsigned i = 0; i != Ratio; ++i)
17899 Vals[i] = Amt.getOperand(i);
17900 for (unsigned i = Ratio; i != Amt.getNumOperands(); i += Ratio) {
17901 for (unsigned j = 0; j != Ratio; ++j)
17902 if (Vals[j] != Amt.getOperand(i + j))
17906 if (SupportedVectorShiftWithBaseAmnt(VT, Subtarget, Op.getOpcode()))
17907 return DAG.getNode(X86OpcV, dl, VT, R, Op.getOperand(1));
17912 static SDValue LowerShift(SDValue Op, const X86Subtarget* Subtarget,
17913 SelectionDAG &DAG) {
17914 MVT VT = Op.getSimpleValueType();
17916 SDValue R = Op.getOperand(0);
17917 SDValue Amt = Op.getOperand(1);
17919 assert(VT.isVector() && "Custom lowering only for vector shifts!");
17920 assert(Subtarget->hasSSE2() && "Only custom lower when we have SSE2!");
17922 if (SDValue V = LowerScalarImmediateShift(Op, DAG, Subtarget))
17925 if (SDValue V = LowerScalarVariableShift(Op, DAG, Subtarget))
17928 if (SupportedVectorVarShift(VT, Subtarget, Op.getOpcode()))
17931 // 2i64 vector logical shifts can efficiently avoid scalarization - do the
17932 // shifts per-lane and then shuffle the partial results back together.
17933 if (VT == MVT::v2i64 && Op.getOpcode() != ISD::SRA) {
17934 // Splat the shift amounts so the scalar shifts above will catch it.
17935 SDValue Amt0 = DAG.getVectorShuffle(VT, dl, Amt, Amt, {0, 0});
17936 SDValue Amt1 = DAG.getVectorShuffle(VT, dl, Amt, Amt, {1, 1});
17937 SDValue R0 = DAG.getNode(Op->getOpcode(), dl, VT, R, Amt0);
17938 SDValue R1 = DAG.getNode(Op->getOpcode(), dl, VT, R, Amt1);
17939 return DAG.getVectorShuffle(VT, dl, R0, R1, {0, 3});
17942 // i64 vector arithmetic shift can be emulated with the transform:
17943 // M = lshr(SIGN_BIT, Amt)
17944 // ashr(R, Amt) === sub(xor(lshr(R, Amt), M), M)
17945 if ((VT == MVT::v2i64 || (VT == MVT::v4i64 && Subtarget->hasInt256())) &&
17946 Op.getOpcode() == ISD::SRA) {
17947 SDValue S = DAG.getConstant(APInt::getSignBit(64), dl, VT);
17948 SDValue M = DAG.getNode(ISD::SRL, dl, VT, S, Amt);
17949 R = DAG.getNode(ISD::SRL, dl, VT, R, Amt);
17950 R = DAG.getNode(ISD::XOR, dl, VT, R, M);
17951 R = DAG.getNode(ISD::SUB, dl, VT, R, M);
17955 // If possible, lower this packed shift into a vector multiply instead of
17956 // expanding it into a sequence of scalar shifts.
17957 // Do this only if the vector shift count is a constant build_vector.
17958 if (Op.getOpcode() == ISD::SHL &&
17959 (VT == MVT::v8i16 || VT == MVT::v4i32 ||
17960 (Subtarget->hasInt256() && VT == MVT::v16i16)) &&
17961 ISD::isBuildVectorOfConstantSDNodes(Amt.getNode())) {
17962 SmallVector<SDValue, 8> Elts;
17963 EVT SVT = VT.getScalarType();
17964 unsigned SVTBits = SVT.getSizeInBits();
17965 const APInt &One = APInt(SVTBits, 1);
17966 unsigned NumElems = VT.getVectorNumElements();
17968 for (unsigned i=0; i !=NumElems; ++i) {
17969 SDValue Op = Amt->getOperand(i);
17970 if (Op->getOpcode() == ISD::UNDEF) {
17971 Elts.push_back(Op);
17975 ConstantSDNode *ND = cast<ConstantSDNode>(Op);
17976 const APInt &C = APInt(SVTBits, ND->getAPIntValue().getZExtValue());
17977 uint64_t ShAmt = C.getZExtValue();
17978 if (ShAmt >= SVTBits) {
17979 Elts.push_back(DAG.getUNDEF(SVT));
17982 Elts.push_back(DAG.getConstant(One.shl(ShAmt), dl, SVT));
17984 SDValue BV = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Elts);
17985 return DAG.getNode(ISD::MUL, dl, VT, R, BV);
17988 // Lower SHL with variable shift amount.
17989 if (VT == MVT::v4i32 && Op->getOpcode() == ISD::SHL) {
17990 Op = DAG.getNode(ISD::SHL, dl, VT, Amt, DAG.getConstant(23, dl, VT));
17992 Op = DAG.getNode(ISD::ADD, dl, VT, Op,
17993 DAG.getConstant(0x3f800000U, dl, VT));
17994 Op = DAG.getBitcast(MVT::v4f32, Op);
17995 Op = DAG.getNode(ISD::FP_TO_SINT, dl, VT, Op);
17996 return DAG.getNode(ISD::MUL, dl, VT, Op, R);
17999 // If possible, lower this shift as a sequence of two shifts by
18000 // constant plus a MOVSS/MOVSD instead of scalarizing it.
18002 // (v4i32 (srl A, (build_vector < X, Y, Y, Y>)))
18004 // Could be rewritten as:
18005 // (v4i32 (MOVSS (srl A, <Y,Y,Y,Y>), (srl A, <X,X,X,X>)))
18007 // The advantage is that the two shifts from the example would be
18008 // lowered as X86ISD::VSRLI nodes. This would be cheaper than scalarizing
18009 // the vector shift into four scalar shifts plus four pairs of vector
18011 if ((VT == MVT::v8i16 || VT == MVT::v4i32) &&
18012 ISD::isBuildVectorOfConstantSDNodes(Amt.getNode())) {
18013 unsigned TargetOpcode = X86ISD::MOVSS;
18014 bool CanBeSimplified;
18015 // The splat value for the first packed shift (the 'X' from the example).
18016 SDValue Amt1 = Amt->getOperand(0);
18017 // The splat value for the second packed shift (the 'Y' from the example).
18018 SDValue Amt2 = (VT == MVT::v4i32) ? Amt->getOperand(1) :
18019 Amt->getOperand(2);
18021 // See if it is possible to replace this node with a sequence of
18022 // two shifts followed by a MOVSS/MOVSD
18023 if (VT == MVT::v4i32) {
18024 // Check if it is legal to use a MOVSS.
18025 CanBeSimplified = Amt2 == Amt->getOperand(2) &&
18026 Amt2 == Amt->getOperand(3);
18027 if (!CanBeSimplified) {
18028 // Otherwise, check if we can still simplify this node using a MOVSD.
18029 CanBeSimplified = Amt1 == Amt->getOperand(1) &&
18030 Amt->getOperand(2) == Amt->getOperand(3);
18031 TargetOpcode = X86ISD::MOVSD;
18032 Amt2 = Amt->getOperand(2);
18035 // Do similar checks for the case where the machine value type
18037 CanBeSimplified = Amt1 == Amt->getOperand(1);
18038 for (unsigned i=3; i != 8 && CanBeSimplified; ++i)
18039 CanBeSimplified = Amt2 == Amt->getOperand(i);
18041 if (!CanBeSimplified) {
18042 TargetOpcode = X86ISD::MOVSD;
18043 CanBeSimplified = true;
18044 Amt2 = Amt->getOperand(4);
18045 for (unsigned i=0; i != 4 && CanBeSimplified; ++i)
18046 CanBeSimplified = Amt1 == Amt->getOperand(i);
18047 for (unsigned j=4; j != 8 && CanBeSimplified; ++j)
18048 CanBeSimplified = Amt2 == Amt->getOperand(j);
18052 if (CanBeSimplified && isa<ConstantSDNode>(Amt1) &&
18053 isa<ConstantSDNode>(Amt2)) {
18054 // Replace this node with two shifts followed by a MOVSS/MOVSD.
18055 EVT CastVT = MVT::v4i32;
18057 DAG.getConstant(cast<ConstantSDNode>(Amt1)->getAPIntValue(), dl, VT);
18058 SDValue Shift1 = DAG.getNode(Op->getOpcode(), dl, VT, R, Splat1);
18060 DAG.getConstant(cast<ConstantSDNode>(Amt2)->getAPIntValue(), dl, VT);
18061 SDValue Shift2 = DAG.getNode(Op->getOpcode(), dl, VT, R, Splat2);
18062 if (TargetOpcode == X86ISD::MOVSD)
18063 CastVT = MVT::v2i64;
18064 SDValue BitCast1 = DAG.getBitcast(CastVT, Shift1);
18065 SDValue BitCast2 = DAG.getBitcast(CastVT, Shift2);
18066 SDValue Result = getTargetShuffleNode(TargetOpcode, dl, CastVT, BitCast2,
18068 return DAG.getBitcast(VT, Result);
18072 // v4i32 Non Uniform Shifts.
18073 // If the shift amount is constant we can shift each lane using the SSE2
18074 // immediate shifts, else we need to zero-extend each lane to the lower i64
18075 // and shift using the SSE2 variable shifts.
18076 // The separate results can then be blended together.
18077 if (VT == MVT::v4i32) {
18078 unsigned Opc = Op.getOpcode();
18079 SDValue Amt0, Amt1, Amt2, Amt3;
18080 if (ISD::isBuildVectorOfConstantSDNodes(Amt.getNode())) {
18081 Amt0 = DAG.getVectorShuffle(VT, dl, Amt, DAG.getUNDEF(VT), {0, 0, 0, 0});
18082 Amt1 = DAG.getVectorShuffle(VT, dl, Amt, DAG.getUNDEF(VT), {1, 1, 1, 1});
18083 Amt2 = DAG.getVectorShuffle(VT, dl, Amt, DAG.getUNDEF(VT), {2, 2, 2, 2});
18084 Amt3 = DAG.getVectorShuffle(VT, dl, Amt, DAG.getUNDEF(VT), {3, 3, 3, 3});
18086 // ISD::SHL is handled above but we include it here for completeness.
18089 llvm_unreachable("Unknown target vector shift node");
18091 Opc = X86ISD::VSHL;
18094 Opc = X86ISD::VSRL;
18097 Opc = X86ISD::VSRA;
18100 // The SSE2 shifts use the lower i64 as the same shift amount for
18101 // all lanes and the upper i64 is ignored. These shuffle masks
18102 // optimally zero-extend each lanes on SSE2/SSE41/AVX targets.
18103 SDValue Z = getZeroVector(VT, Subtarget, DAG, dl);
18104 Amt0 = DAG.getVectorShuffle(VT, dl, Amt, Z, {0, 4, -1, -1});
18105 Amt1 = DAG.getVectorShuffle(VT, dl, Amt, Z, {1, 5, -1, -1});
18106 Amt2 = DAG.getVectorShuffle(VT, dl, Amt, Z, {2, 6, -1, -1});
18107 Amt3 = DAG.getVectorShuffle(VT, dl, Amt, Z, {3, 7, -1, -1});
18110 SDValue R0 = DAG.getNode(Opc, dl, VT, R, Amt0);
18111 SDValue R1 = DAG.getNode(Opc, dl, VT, R, Amt1);
18112 SDValue R2 = DAG.getNode(Opc, dl, VT, R, Amt2);
18113 SDValue R3 = DAG.getNode(Opc, dl, VT, R, Amt3);
18114 SDValue R02 = DAG.getVectorShuffle(VT, dl, R0, R2, {0, -1, 6, -1});
18115 SDValue R13 = DAG.getVectorShuffle(VT, dl, R1, R3, {-1, 1, -1, 7});
18116 return DAG.getVectorShuffle(VT, dl, R02, R13, {0, 5, 2, 7});
18119 if (VT == MVT::v16i8 || (VT == MVT::v32i8 && Subtarget->hasInt256())) {
18120 MVT ExtVT = MVT::getVectorVT(MVT::i16, VT.getVectorNumElements() / 2);
18121 unsigned ShiftOpcode = Op->getOpcode();
18123 auto SignBitSelect = [&](MVT SelVT, SDValue Sel, SDValue V0, SDValue V1) {
18124 // On SSE41 targets we make use of the fact that VSELECT lowers
18125 // to PBLENDVB which selects bytes based just on the sign bit.
18126 if (Subtarget->hasSSE41()) {
18127 V0 = DAG.getBitcast(VT, V0);
18128 V1 = DAG.getBitcast(VT, V1);
18129 Sel = DAG.getBitcast(VT, Sel);
18130 return DAG.getBitcast(SelVT,
18131 DAG.getNode(ISD::VSELECT, dl, VT, Sel, V0, V1));
18133 // On pre-SSE41 targets we test for the sign bit by comparing to
18134 // zero - a negative value will set all bits of the lanes to true
18135 // and VSELECT uses that in its OR(AND(V0,C),AND(V1,~C)) lowering.
18136 SDValue Z = getZeroVector(SelVT, Subtarget, DAG, dl);
18137 SDValue C = DAG.getNode(X86ISD::PCMPGT, dl, SelVT, Z, Sel);
18138 return DAG.getNode(ISD::VSELECT, dl, SelVT, C, V0, V1);
18141 // Turn 'a' into a mask suitable for VSELECT: a = a << 5;
18142 // We can safely do this using i16 shifts as we're only interested in
18143 // the 3 lower bits of each byte.
18144 Amt = DAG.getBitcast(ExtVT, Amt);
18145 Amt = DAG.getNode(ISD::SHL, dl, ExtVT, Amt, DAG.getConstant(5, dl, ExtVT));
18146 Amt = DAG.getBitcast(VT, Amt);
18148 if (Op->getOpcode() == ISD::SHL || Op->getOpcode() == ISD::SRL) {
18149 // r = VSELECT(r, shift(r, 4), a);
18151 DAG.getNode(ShiftOpcode, dl, VT, R, DAG.getConstant(4, dl, VT));
18152 R = SignBitSelect(VT, Amt, M, R);
18155 Amt = DAG.getNode(ISD::ADD, dl, VT, Amt, Amt);
18157 // r = VSELECT(r, shift(r, 2), a);
18158 M = DAG.getNode(ShiftOpcode, dl, VT, R, DAG.getConstant(2, dl, VT));
18159 R = SignBitSelect(VT, Amt, M, R);
18162 Amt = DAG.getNode(ISD::ADD, dl, VT, Amt, Amt);
18164 // return VSELECT(r, shift(r, 1), a);
18165 M = DAG.getNode(ShiftOpcode, dl, VT, R, DAG.getConstant(1, dl, VT));
18166 R = SignBitSelect(VT, Amt, M, R);
18170 if (Op->getOpcode() == ISD::SRA) {
18171 // For SRA we need to unpack each byte to the higher byte of a i16 vector
18172 // so we can correctly sign extend. We don't care what happens to the
18174 SDValue ALo = DAG.getNode(X86ISD::UNPCKL, dl, VT, DAG.getUNDEF(VT), Amt);
18175 SDValue AHi = DAG.getNode(X86ISD::UNPCKH, dl, VT, DAG.getUNDEF(VT), Amt);
18176 SDValue RLo = DAG.getNode(X86ISD::UNPCKL, dl, VT, DAG.getUNDEF(VT), R);
18177 SDValue RHi = DAG.getNode(X86ISD::UNPCKH, dl, VT, DAG.getUNDEF(VT), R);
18178 ALo = DAG.getBitcast(ExtVT, ALo);
18179 AHi = DAG.getBitcast(ExtVT, AHi);
18180 RLo = DAG.getBitcast(ExtVT, RLo);
18181 RHi = DAG.getBitcast(ExtVT, RHi);
18183 // r = VSELECT(r, shift(r, 4), a);
18184 SDValue MLo = DAG.getNode(ShiftOpcode, dl, ExtVT, RLo,
18185 DAG.getConstant(4, dl, ExtVT));
18186 SDValue MHi = DAG.getNode(ShiftOpcode, dl, ExtVT, RHi,
18187 DAG.getConstant(4, dl, ExtVT));
18188 RLo = SignBitSelect(ExtVT, ALo, MLo, RLo);
18189 RHi = SignBitSelect(ExtVT, AHi, MHi, RHi);
18192 ALo = DAG.getNode(ISD::ADD, dl, ExtVT, ALo, ALo);
18193 AHi = DAG.getNode(ISD::ADD, dl, ExtVT, AHi, AHi);
18195 // r = VSELECT(r, shift(r, 2), a);
18196 MLo = DAG.getNode(ShiftOpcode, dl, ExtVT, RLo,
18197 DAG.getConstant(2, dl, ExtVT));
18198 MHi = DAG.getNode(ShiftOpcode, dl, ExtVT, RHi,
18199 DAG.getConstant(2, dl, ExtVT));
18200 RLo = SignBitSelect(ExtVT, ALo, MLo, RLo);
18201 RHi = SignBitSelect(ExtVT, AHi, MHi, RHi);
18204 ALo = DAG.getNode(ISD::ADD, dl, ExtVT, ALo, ALo);
18205 AHi = DAG.getNode(ISD::ADD, dl, ExtVT, AHi, AHi);
18207 // r = VSELECT(r, shift(r, 1), a);
18208 MLo = DAG.getNode(ShiftOpcode, dl, ExtVT, RLo,
18209 DAG.getConstant(1, dl, ExtVT));
18210 MHi = DAG.getNode(ShiftOpcode, dl, ExtVT, RHi,
18211 DAG.getConstant(1, dl, ExtVT));
18212 RLo = SignBitSelect(ExtVT, ALo, MLo, RLo);
18213 RHi = SignBitSelect(ExtVT, AHi, MHi, RHi);
18215 // Logical shift the result back to the lower byte, leaving a zero upper
18217 // meaning that we can safely pack with PACKUSWB.
18219 DAG.getNode(ISD::SRL, dl, ExtVT, RLo, DAG.getConstant(8, dl, ExtVT));
18221 DAG.getNode(ISD::SRL, dl, ExtVT, RHi, DAG.getConstant(8, dl, ExtVT));
18222 return DAG.getNode(X86ISD::PACKUS, dl, VT, RLo, RHi);
18226 // It's worth extending once and using the v8i32 shifts for 16-bit types, but
18227 // the extra overheads to get from v16i8 to v8i32 make the existing SSE
18228 // solution better.
18229 if (Subtarget->hasInt256() && VT == MVT::v8i16) {
18230 MVT ExtVT = MVT::v8i32;
18232 Op.getOpcode() == ISD::SRA ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND;
18233 R = DAG.getNode(ExtOpc, dl, ExtVT, R);
18234 Amt = DAG.getNode(ISD::ANY_EXTEND, dl, ExtVT, Amt);
18235 return DAG.getNode(ISD::TRUNCATE, dl, VT,
18236 DAG.getNode(Op.getOpcode(), dl, ExtVT, R, Amt));
18239 if (Subtarget->hasInt256() && VT == MVT::v16i16) {
18240 MVT ExtVT = MVT::v8i32;
18241 SDValue Z = getZeroVector(VT, Subtarget, DAG, dl);
18242 SDValue ALo = DAG.getNode(X86ISD::UNPCKL, dl, VT, Amt, Z);
18243 SDValue AHi = DAG.getNode(X86ISD::UNPCKH, dl, VT, Amt, Z);
18244 SDValue RLo = DAG.getNode(X86ISD::UNPCKL, dl, VT, R, R);
18245 SDValue RHi = DAG.getNode(X86ISD::UNPCKH, dl, VT, R, R);
18246 ALo = DAG.getBitcast(ExtVT, ALo);
18247 AHi = DAG.getBitcast(ExtVT, AHi);
18248 RLo = DAG.getBitcast(ExtVT, RLo);
18249 RHi = DAG.getBitcast(ExtVT, RHi);
18250 SDValue Lo = DAG.getNode(Op.getOpcode(), dl, ExtVT, RLo, ALo);
18251 SDValue Hi = DAG.getNode(Op.getOpcode(), dl, ExtVT, RHi, AHi);
18252 Lo = DAG.getNode(ISD::SRL, dl, ExtVT, Lo, DAG.getConstant(16, dl, ExtVT));
18253 Hi = DAG.getNode(ISD::SRL, dl, ExtVT, Hi, DAG.getConstant(16, dl, ExtVT));
18254 return DAG.getNode(X86ISD::PACKUS, dl, VT, Lo, Hi);
18257 if (VT == MVT::v8i16) {
18258 unsigned ShiftOpcode = Op->getOpcode();
18260 auto SignBitSelect = [&](SDValue Sel, SDValue V0, SDValue V1) {
18261 // On SSE41 targets we make use of the fact that VSELECT lowers
18262 // to PBLENDVB which selects bytes based just on the sign bit.
18263 if (Subtarget->hasSSE41()) {
18264 MVT ExtVT = MVT::getVectorVT(MVT::i8, VT.getVectorNumElements() * 2);
18265 V0 = DAG.getBitcast(ExtVT, V0);
18266 V1 = DAG.getBitcast(ExtVT, V1);
18267 Sel = DAG.getBitcast(ExtVT, Sel);
18268 return DAG.getBitcast(
18269 VT, DAG.getNode(ISD::VSELECT, dl, ExtVT, Sel, V0, V1));
18271 // On pre-SSE41 targets we splat the sign bit - a negative value will
18272 // set all bits of the lanes to true and VSELECT uses that in
18273 // its OR(AND(V0,C),AND(V1,~C)) lowering.
18275 DAG.getNode(ISD::SRA, dl, VT, Sel, DAG.getConstant(15, dl, VT));
18276 return DAG.getNode(ISD::VSELECT, dl, VT, C, V0, V1);
18279 // Turn 'a' into a mask suitable for VSELECT: a = a << 12;
18280 if (Subtarget->hasSSE41()) {
18281 // On SSE41 targets we need to replicate the shift mask in both
18282 // bytes for PBLENDVB.
18285 DAG.getNode(ISD::SHL, dl, VT, Amt, DAG.getConstant(4, dl, VT)),
18286 DAG.getNode(ISD::SHL, dl, VT, Amt, DAG.getConstant(12, dl, VT)));
18288 Amt = DAG.getNode(ISD::SHL, dl, VT, Amt, DAG.getConstant(12, dl, VT));
18291 // r = VSELECT(r, shift(r, 8), a);
18292 SDValue M = DAG.getNode(ShiftOpcode, dl, VT, R, DAG.getConstant(8, dl, VT));
18293 R = SignBitSelect(Amt, M, R);
18296 Amt = DAG.getNode(ISD::ADD, dl, VT, Amt, Amt);
18298 // r = VSELECT(r, shift(r, 4), a);
18299 M = DAG.getNode(ShiftOpcode, dl, VT, R, DAG.getConstant(4, dl, VT));
18300 R = SignBitSelect(Amt, M, R);
18303 Amt = DAG.getNode(ISD::ADD, dl, VT, Amt, Amt);
18305 // r = VSELECT(r, shift(r, 2), a);
18306 M = DAG.getNode(ShiftOpcode, dl, VT, R, DAG.getConstant(2, dl, VT));
18307 R = SignBitSelect(Amt, M, R);
18310 Amt = DAG.getNode(ISD::ADD, dl, VT, Amt, Amt);
18312 // return VSELECT(r, shift(r, 1), a);
18313 M = DAG.getNode(ShiftOpcode, dl, VT, R, DAG.getConstant(1, dl, VT));
18314 R = SignBitSelect(Amt, M, R);
18318 // Decompose 256-bit shifts into smaller 128-bit shifts.
18319 if (VT.is256BitVector()) {
18320 unsigned NumElems = VT.getVectorNumElements();
18321 MVT EltVT = VT.getVectorElementType();
18322 EVT NewVT = MVT::getVectorVT(EltVT, NumElems/2);
18324 // Extract the two vectors
18325 SDValue V1 = Extract128BitVector(R, 0, DAG, dl);
18326 SDValue V2 = Extract128BitVector(R, NumElems/2, DAG, dl);
18328 // Recreate the shift amount vectors
18329 SDValue Amt1, Amt2;
18330 if (Amt.getOpcode() == ISD::BUILD_VECTOR) {
18331 // Constant shift amount
18332 SmallVector<SDValue, 8> Ops(Amt->op_begin(), Amt->op_begin() + NumElems);
18333 ArrayRef<SDValue> Amt1Csts = makeArrayRef(Ops).slice(0, NumElems / 2);
18334 ArrayRef<SDValue> Amt2Csts = makeArrayRef(Ops).slice(NumElems / 2);
18336 Amt1 = DAG.getNode(ISD::BUILD_VECTOR, dl, NewVT, Amt1Csts);
18337 Amt2 = DAG.getNode(ISD::BUILD_VECTOR, dl, NewVT, Amt2Csts);
18339 // Variable shift amount
18340 Amt1 = Extract128BitVector(Amt, 0, DAG, dl);
18341 Amt2 = Extract128BitVector(Amt, NumElems/2, DAG, dl);
18344 // Issue new vector shifts for the smaller types
18345 V1 = DAG.getNode(Op.getOpcode(), dl, NewVT, V1, Amt1);
18346 V2 = DAG.getNode(Op.getOpcode(), dl, NewVT, V2, Amt2);
18348 // Concatenate the result back
18349 return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT, V1, V2);
18355 static SDValue LowerXALUO(SDValue Op, SelectionDAG &DAG) {
18356 // Lower the "add/sub/mul with overflow" instruction into a regular ins plus
18357 // a "setcc" instruction that checks the overflow flag. The "brcond" lowering
18358 // looks for this combo and may remove the "setcc" instruction if the "setcc"
18359 // has only one use.
18360 SDNode *N = Op.getNode();
18361 SDValue LHS = N->getOperand(0);
18362 SDValue RHS = N->getOperand(1);
18363 unsigned BaseOp = 0;
18366 switch (Op.getOpcode()) {
18367 default: llvm_unreachable("Unknown ovf instruction!");
18369 // A subtract of one will be selected as a INC. Note that INC doesn't
18370 // set CF, so we can't do this for UADDO.
18371 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(RHS))
18373 BaseOp = X86ISD::INC;
18374 Cond = X86::COND_O;
18377 BaseOp = X86ISD::ADD;
18378 Cond = X86::COND_O;
18381 BaseOp = X86ISD::ADD;
18382 Cond = X86::COND_B;
18385 // A subtract of one will be selected as a DEC. Note that DEC doesn't
18386 // set CF, so we can't do this for USUBO.
18387 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(RHS))
18389 BaseOp = X86ISD::DEC;
18390 Cond = X86::COND_O;
18393 BaseOp = X86ISD::SUB;
18394 Cond = X86::COND_O;
18397 BaseOp = X86ISD::SUB;
18398 Cond = X86::COND_B;
18401 BaseOp = N->getValueType(0) == MVT::i8 ? X86ISD::SMUL8 : X86ISD::SMUL;
18402 Cond = X86::COND_O;
18404 case ISD::UMULO: { // i64, i8 = umulo lhs, rhs --> i64, i64, i32 umul lhs,rhs
18405 if (N->getValueType(0) == MVT::i8) {
18406 BaseOp = X86ISD::UMUL8;
18407 Cond = X86::COND_O;
18410 SDVTList VTs = DAG.getVTList(N->getValueType(0), N->getValueType(0),
18412 SDValue Sum = DAG.getNode(X86ISD::UMUL, DL, VTs, LHS, RHS);
18415 DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
18416 DAG.getConstant(X86::COND_O, DL, MVT::i32),
18417 SDValue(Sum.getNode(), 2));
18419 return DAG.getNode(ISD::MERGE_VALUES, DL, N->getVTList(), Sum, SetCC);
18423 // Also sets EFLAGS.
18424 SDVTList VTs = DAG.getVTList(N->getValueType(0), MVT::i32);
18425 SDValue Sum = DAG.getNode(BaseOp, DL, VTs, LHS, RHS);
18428 DAG.getNode(X86ISD::SETCC, DL, N->getValueType(1),
18429 DAG.getConstant(Cond, DL, MVT::i32),
18430 SDValue(Sum.getNode(), 1));
18432 return DAG.getNode(ISD::MERGE_VALUES, DL, N->getVTList(), Sum, SetCC);
18435 /// Returns true if the operand type is exactly twice the native width, and
18436 /// the corresponding cmpxchg8b or cmpxchg16b instruction is available.
18437 /// Used to know whether to use cmpxchg8/16b when expanding atomic operations
18438 /// (otherwise we leave them alone to become __sync_fetch_and_... calls).
18439 bool X86TargetLowering::needsCmpXchgNb(Type *MemType) const {
18440 unsigned OpWidth = MemType->getPrimitiveSizeInBits();
18443 return !Subtarget->is64Bit(); // FIXME this should be Subtarget.hasCmpxchg8b
18444 else if (OpWidth == 128)
18445 return Subtarget->hasCmpxchg16b();
18450 bool X86TargetLowering::shouldExpandAtomicStoreInIR(StoreInst *SI) const {
18451 return needsCmpXchgNb(SI->getValueOperand()->getType());
18454 // Note: this turns large loads into lock cmpxchg8b/16b.
18455 // FIXME: On 32 bits x86, fild/movq might be faster than lock cmpxchg8b.
18456 TargetLowering::AtomicExpansionKind
18457 X86TargetLowering::shouldExpandAtomicLoadInIR(LoadInst *LI) const {
18458 auto PTy = cast<PointerType>(LI->getPointerOperand()->getType());
18459 return needsCmpXchgNb(PTy->getElementType()) ? AtomicExpansionKind::CmpXChg
18460 : AtomicExpansionKind::None;
18463 TargetLowering::AtomicExpansionKind
18464 X86TargetLowering::shouldExpandAtomicRMWInIR(AtomicRMWInst *AI) const {
18465 unsigned NativeWidth = Subtarget->is64Bit() ? 64 : 32;
18466 Type *MemType = AI->getType();
18468 // If the operand is too big, we must see if cmpxchg8/16b is available
18469 // and default to library calls otherwise.
18470 if (MemType->getPrimitiveSizeInBits() > NativeWidth) {
18471 return needsCmpXchgNb(MemType) ? AtomicExpansionKind::CmpXChg
18472 : AtomicExpansionKind::None;
18475 AtomicRMWInst::BinOp Op = AI->getOperation();
18478 llvm_unreachable("Unknown atomic operation");
18479 case AtomicRMWInst::Xchg:
18480 case AtomicRMWInst::Add:
18481 case AtomicRMWInst::Sub:
18482 // It's better to use xadd, xsub or xchg for these in all cases.
18483 return AtomicExpansionKind::None;
18484 case AtomicRMWInst::Or:
18485 case AtomicRMWInst::And:
18486 case AtomicRMWInst::Xor:
18487 // If the atomicrmw's result isn't actually used, we can just add a "lock"
18488 // prefix to a normal instruction for these operations.
18489 return !AI->use_empty() ? AtomicExpansionKind::CmpXChg
18490 : AtomicExpansionKind::None;
18491 case AtomicRMWInst::Nand:
18492 case AtomicRMWInst::Max:
18493 case AtomicRMWInst::Min:
18494 case AtomicRMWInst::UMax:
18495 case AtomicRMWInst::UMin:
18496 // These always require a non-trivial set of data operations on x86. We must
18497 // use a cmpxchg loop.
18498 return AtomicExpansionKind::CmpXChg;
18502 static bool hasMFENCE(const X86Subtarget& Subtarget) {
18503 // Use mfence if we have SSE2 or we're on x86-64 (even if we asked for
18504 // no-sse2). There isn't any reason to disable it if the target processor
18506 return Subtarget.hasSSE2() || Subtarget.is64Bit();
18510 X86TargetLowering::lowerIdempotentRMWIntoFencedLoad(AtomicRMWInst *AI) const {
18511 unsigned NativeWidth = Subtarget->is64Bit() ? 64 : 32;
18512 Type *MemType = AI->getType();
18513 // Accesses larger than the native width are turned into cmpxchg/libcalls, so
18514 // there is no benefit in turning such RMWs into loads, and it is actually
18515 // harmful as it introduces a mfence.
18516 if (MemType->getPrimitiveSizeInBits() > NativeWidth)
18519 auto Builder = IRBuilder<>(AI);
18520 Module *M = Builder.GetInsertBlock()->getParent()->getParent();
18521 auto SynchScope = AI->getSynchScope();
18522 // We must restrict the ordering to avoid generating loads with Release or
18523 // ReleaseAcquire orderings.
18524 auto Order = AtomicCmpXchgInst::getStrongestFailureOrdering(AI->getOrdering());
18525 auto Ptr = AI->getPointerOperand();
18527 // Before the load we need a fence. Here is an example lifted from
18528 // http://www.hpl.hp.com/techreports/2012/HPL-2012-68.pdf showing why a fence
18531 // x.store(1, relaxed);
18532 // r1 = y.fetch_add(0, release);
18534 // y.fetch_add(42, acquire);
18535 // r2 = x.load(relaxed);
18536 // r1 = r2 = 0 is impossible, but becomes possible if the idempotent rmw is
18537 // lowered to just a load without a fence. A mfence flushes the store buffer,
18538 // making the optimization clearly correct.
18539 // FIXME: it is required if isAtLeastRelease(Order) but it is not clear
18540 // otherwise, we might be able to be more aggressive on relaxed idempotent
18541 // rmw. In practice, they do not look useful, so we don't try to be
18542 // especially clever.
18543 if (SynchScope == SingleThread)
18544 // FIXME: we could just insert an X86ISD::MEMBARRIER here, except we are at
18545 // the IR level, so we must wrap it in an intrinsic.
18548 if (!hasMFENCE(*Subtarget))
18549 // FIXME: it might make sense to use a locked operation here but on a
18550 // different cache-line to prevent cache-line bouncing. In practice it
18551 // is probably a small win, and x86 processors without mfence are rare
18552 // enough that we do not bother.
18556 llvm::Intrinsic::getDeclaration(M, Intrinsic::x86_sse2_mfence);
18557 Builder.CreateCall(MFence, {});
18559 // Finally we can emit the atomic load.
18560 LoadInst *Loaded = Builder.CreateAlignedLoad(Ptr,
18561 AI->getType()->getPrimitiveSizeInBits());
18562 Loaded->setAtomic(Order, SynchScope);
18563 AI->replaceAllUsesWith(Loaded);
18564 AI->eraseFromParent();
18568 static SDValue LowerATOMIC_FENCE(SDValue Op, const X86Subtarget *Subtarget,
18569 SelectionDAG &DAG) {
18571 AtomicOrdering FenceOrdering = static_cast<AtomicOrdering>(
18572 cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue());
18573 SynchronizationScope FenceScope = static_cast<SynchronizationScope>(
18574 cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue());
18576 // The only fence that needs an instruction is a sequentially-consistent
18577 // cross-thread fence.
18578 if (FenceOrdering == SequentiallyConsistent && FenceScope == CrossThread) {
18579 if (hasMFENCE(*Subtarget))
18580 return DAG.getNode(X86ISD::MFENCE, dl, MVT::Other, Op.getOperand(0));
18582 SDValue Chain = Op.getOperand(0);
18583 SDValue Zero = DAG.getConstant(0, dl, MVT::i32);
18585 DAG.getRegister(X86::ESP, MVT::i32), // Base
18586 DAG.getTargetConstant(1, dl, MVT::i8), // Scale
18587 DAG.getRegister(0, MVT::i32), // Index
18588 DAG.getTargetConstant(0, dl, MVT::i32), // Disp
18589 DAG.getRegister(0, MVT::i32), // Segment.
18593 SDNode *Res = DAG.getMachineNode(X86::OR32mrLocked, dl, MVT::Other, Ops);
18594 return SDValue(Res, 0);
18597 // MEMBARRIER is a compiler barrier; it codegens to a no-op.
18598 return DAG.getNode(X86ISD::MEMBARRIER, dl, MVT::Other, Op.getOperand(0));
18601 static SDValue LowerCMP_SWAP(SDValue Op, const X86Subtarget *Subtarget,
18602 SelectionDAG &DAG) {
18603 MVT T = Op.getSimpleValueType();
18607 switch(T.SimpleTy) {
18608 default: llvm_unreachable("Invalid value type!");
18609 case MVT::i8: Reg = X86::AL; size = 1; break;
18610 case MVT::i16: Reg = X86::AX; size = 2; break;
18611 case MVT::i32: Reg = X86::EAX; size = 4; break;
18613 assert(Subtarget->is64Bit() && "Node not type legal!");
18614 Reg = X86::RAX; size = 8;
18617 SDValue cpIn = DAG.getCopyToReg(Op.getOperand(0), DL, Reg,
18618 Op.getOperand(2), SDValue());
18619 SDValue Ops[] = { cpIn.getValue(0),
18622 DAG.getTargetConstant(size, DL, MVT::i8),
18623 cpIn.getValue(1) };
18624 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Glue);
18625 MachineMemOperand *MMO = cast<AtomicSDNode>(Op)->getMemOperand();
18626 SDValue Result = DAG.getMemIntrinsicNode(X86ISD::LCMPXCHG_DAG, DL, Tys,
18630 DAG.getCopyFromReg(Result.getValue(0), DL, Reg, T, Result.getValue(1));
18631 SDValue EFLAGS = DAG.getCopyFromReg(cpOut.getValue(1), DL, X86::EFLAGS,
18632 MVT::i32, cpOut.getValue(2));
18633 SDValue Success = DAG.getNode(X86ISD::SETCC, DL, Op->getValueType(1),
18634 DAG.getConstant(X86::COND_E, DL, MVT::i8),
18637 DAG.ReplaceAllUsesOfValueWith(Op.getValue(0), cpOut);
18638 DAG.ReplaceAllUsesOfValueWith(Op.getValue(1), Success);
18639 DAG.ReplaceAllUsesOfValueWith(Op.getValue(2), EFLAGS.getValue(1));
18643 static SDValue LowerBITCAST(SDValue Op, const X86Subtarget *Subtarget,
18644 SelectionDAG &DAG) {
18645 MVT SrcVT = Op.getOperand(0).getSimpleValueType();
18646 MVT DstVT = Op.getSimpleValueType();
18648 if (SrcVT == MVT::v2i32 || SrcVT == MVT::v4i16 || SrcVT == MVT::v8i8) {
18649 assert(Subtarget->hasSSE2() && "Requires at least SSE2!");
18650 if (DstVT != MVT::f64)
18651 // This conversion needs to be expanded.
18654 SDValue InVec = Op->getOperand(0);
18656 unsigned NumElts = SrcVT.getVectorNumElements();
18657 EVT SVT = SrcVT.getVectorElementType();
18659 // Widen the vector in input in the case of MVT::v2i32.
18660 // Example: from MVT::v2i32 to MVT::v4i32.
18661 SmallVector<SDValue, 16> Elts;
18662 for (unsigned i = 0, e = NumElts; i != e; ++i)
18663 Elts.push_back(DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, SVT, InVec,
18664 DAG.getIntPtrConstant(i, dl)));
18666 // Explicitly mark the extra elements as Undef.
18667 Elts.append(NumElts, DAG.getUNDEF(SVT));
18669 EVT NewVT = EVT::getVectorVT(*DAG.getContext(), SVT, NumElts * 2);
18670 SDValue BV = DAG.getNode(ISD::BUILD_VECTOR, dl, NewVT, Elts);
18671 SDValue ToV2F64 = DAG.getBitcast(MVT::v2f64, BV);
18672 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, ToV2F64,
18673 DAG.getIntPtrConstant(0, dl));
18676 assert(Subtarget->is64Bit() && !Subtarget->hasSSE2() &&
18677 Subtarget->hasMMX() && "Unexpected custom BITCAST");
18678 assert((DstVT == MVT::i64 ||
18679 (DstVT.isVector() && DstVT.getSizeInBits()==64)) &&
18680 "Unexpected custom BITCAST");
18681 // i64 <=> MMX conversions are Legal.
18682 if (SrcVT==MVT::i64 && DstVT.isVector())
18684 if (DstVT==MVT::i64 && SrcVT.isVector())
18686 // MMX <=> MMX conversions are Legal.
18687 if (SrcVT.isVector() && DstVT.isVector())
18689 // All other conversions need to be expanded.
18693 /// Compute the horizontal sum of bytes in V for the elements of VT.
18695 /// Requires V to be a byte vector and VT to be an integer vector type with
18696 /// wider elements than V's type. The width of the elements of VT determines
18697 /// how many bytes of V are summed horizontally to produce each element of the
18699 static SDValue LowerHorizontalByteSum(SDValue V, MVT VT,
18700 const X86Subtarget *Subtarget,
18701 SelectionDAG &DAG) {
18703 MVT ByteVecVT = V.getSimpleValueType();
18704 MVT EltVT = VT.getVectorElementType();
18705 int NumElts = VT.getVectorNumElements();
18706 assert(ByteVecVT.getVectorElementType() == MVT::i8 &&
18707 "Expected value to have byte element type.");
18708 assert(EltVT != MVT::i8 &&
18709 "Horizontal byte sum only makes sense for wider elements!");
18710 unsigned VecSize = VT.getSizeInBits();
18711 assert(ByteVecVT.getSizeInBits() == VecSize && "Cannot change vector size!");
18713 // PSADBW instruction horizontally add all bytes and leave the result in i64
18714 // chunks, thus directly computes the pop count for v2i64 and v4i64.
18715 if (EltVT == MVT::i64) {
18716 SDValue Zeros = getZeroVector(ByteVecVT, Subtarget, DAG, DL);
18717 V = DAG.getNode(X86ISD::PSADBW, DL, ByteVecVT, V, Zeros);
18718 return DAG.getBitcast(VT, V);
18721 if (EltVT == MVT::i32) {
18722 // We unpack the low half and high half into i32s interleaved with zeros so
18723 // that we can use PSADBW to horizontally sum them. The most useful part of
18724 // this is that it lines up the results of two PSADBW instructions to be
18725 // two v2i64 vectors which concatenated are the 4 population counts. We can
18726 // then use PACKUSWB to shrink and concatenate them into a v4i32 again.
18727 SDValue Zeros = getZeroVector(VT, Subtarget, DAG, DL);
18728 SDValue Low = DAG.getNode(X86ISD::UNPCKL, DL, VT, V, Zeros);
18729 SDValue High = DAG.getNode(X86ISD::UNPCKH, DL, VT, V, Zeros);
18731 // Do the horizontal sums into two v2i64s.
18732 Zeros = getZeroVector(ByteVecVT, Subtarget, DAG, DL);
18733 Low = DAG.getNode(X86ISD::PSADBW, DL, ByteVecVT,
18734 DAG.getBitcast(ByteVecVT, Low), Zeros);
18735 High = DAG.getNode(X86ISD::PSADBW, DL, ByteVecVT,
18736 DAG.getBitcast(ByteVecVT, High), Zeros);
18738 // Merge them together.
18739 MVT ShortVecVT = MVT::getVectorVT(MVT::i16, VecSize / 16);
18740 V = DAG.getNode(X86ISD::PACKUS, DL, ByteVecVT,
18741 DAG.getBitcast(ShortVecVT, Low),
18742 DAG.getBitcast(ShortVecVT, High));
18744 return DAG.getBitcast(VT, V);
18747 // The only element type left is i16.
18748 assert(EltVT == MVT::i16 && "Unknown how to handle type");
18750 // To obtain pop count for each i16 element starting from the pop count for
18751 // i8 elements, shift the i16s left by 8, sum as i8s, and then shift as i16s
18752 // right by 8. It is important to shift as i16s as i8 vector shift isn't
18753 // directly supported.
18754 SmallVector<SDValue, 16> Shifters(NumElts, DAG.getConstant(8, DL, EltVT));
18755 SDValue Shifter = DAG.getNode(ISD::BUILD_VECTOR, DL, VT, Shifters);
18756 SDValue Shl = DAG.getNode(ISD::SHL, DL, VT, DAG.getBitcast(VT, V), Shifter);
18757 V = DAG.getNode(ISD::ADD, DL, ByteVecVT, DAG.getBitcast(ByteVecVT, Shl),
18758 DAG.getBitcast(ByteVecVT, V));
18759 return DAG.getNode(ISD::SRL, DL, VT, DAG.getBitcast(VT, V), Shifter);
18762 static SDValue LowerVectorCTPOPInRegLUT(SDValue Op, SDLoc DL,
18763 const X86Subtarget *Subtarget,
18764 SelectionDAG &DAG) {
18765 MVT VT = Op.getSimpleValueType();
18766 MVT EltVT = VT.getVectorElementType();
18767 unsigned VecSize = VT.getSizeInBits();
18769 // Implement a lookup table in register by using an algorithm based on:
18770 // http://wm.ite.pl/articles/sse-popcount.html
18772 // The general idea is that every lower byte nibble in the input vector is an
18773 // index into a in-register pre-computed pop count table. We then split up the
18774 // input vector in two new ones: (1) a vector with only the shifted-right
18775 // higher nibbles for each byte and (2) a vector with the lower nibbles (and
18776 // masked out higher ones) for each byte. PSHUB is used separately with both
18777 // to index the in-register table. Next, both are added and the result is a
18778 // i8 vector where each element contains the pop count for input byte.
18780 // To obtain the pop count for elements != i8, we follow up with the same
18781 // approach and use additional tricks as described below.
18783 const int LUT[16] = {/* 0 */ 0, /* 1 */ 1, /* 2 */ 1, /* 3 */ 2,
18784 /* 4 */ 1, /* 5 */ 2, /* 6 */ 2, /* 7 */ 3,
18785 /* 8 */ 1, /* 9 */ 2, /* a */ 2, /* b */ 3,
18786 /* c */ 2, /* d */ 3, /* e */ 3, /* f */ 4};
18788 int NumByteElts = VecSize / 8;
18789 MVT ByteVecVT = MVT::getVectorVT(MVT::i8, NumByteElts);
18790 SDValue In = DAG.getBitcast(ByteVecVT, Op);
18791 SmallVector<SDValue, 16> LUTVec;
18792 for (int i = 0; i < NumByteElts; ++i)
18793 LUTVec.push_back(DAG.getConstant(LUT[i % 16], DL, MVT::i8));
18794 SDValue InRegLUT = DAG.getNode(ISD::BUILD_VECTOR, DL, ByteVecVT, LUTVec);
18795 SmallVector<SDValue, 16> Mask0F(NumByteElts,
18796 DAG.getConstant(0x0F, DL, MVT::i8));
18797 SDValue M0F = DAG.getNode(ISD::BUILD_VECTOR, DL, ByteVecVT, Mask0F);
18800 SmallVector<SDValue, 16> Four(NumByteElts, DAG.getConstant(4, DL, MVT::i8));
18801 SDValue FourV = DAG.getNode(ISD::BUILD_VECTOR, DL, ByteVecVT, Four);
18802 SDValue HighNibbles = DAG.getNode(ISD::SRL, DL, ByteVecVT, In, FourV);
18805 SDValue LowNibbles = DAG.getNode(ISD::AND, DL, ByteVecVT, In, M0F);
18807 // The input vector is used as the shuffle mask that index elements into the
18808 // LUT. After counting low and high nibbles, add the vector to obtain the
18809 // final pop count per i8 element.
18810 SDValue HighPopCnt =
18811 DAG.getNode(X86ISD::PSHUFB, DL, ByteVecVT, InRegLUT, HighNibbles);
18812 SDValue LowPopCnt =
18813 DAG.getNode(X86ISD::PSHUFB, DL, ByteVecVT, InRegLUT, LowNibbles);
18814 SDValue PopCnt = DAG.getNode(ISD::ADD, DL, ByteVecVT, HighPopCnt, LowPopCnt);
18816 if (EltVT == MVT::i8)
18819 return LowerHorizontalByteSum(PopCnt, VT, Subtarget, DAG);
18822 static SDValue LowerVectorCTPOPBitmath(SDValue Op, SDLoc DL,
18823 const X86Subtarget *Subtarget,
18824 SelectionDAG &DAG) {
18825 MVT VT = Op.getSimpleValueType();
18826 assert(VT.is128BitVector() &&
18827 "Only 128-bit vector bitmath lowering supported.");
18829 int VecSize = VT.getSizeInBits();
18830 MVT EltVT = VT.getVectorElementType();
18831 int Len = EltVT.getSizeInBits();
18833 // This is the vectorized version of the "best" algorithm from
18834 // http://graphics.stanford.edu/~seander/bithacks.html#CountBitsSetParallel
18835 // with a minor tweak to use a series of adds + shifts instead of vector
18836 // multiplications. Implemented for all integer vector types. We only use
18837 // this when we don't have SSSE3 which allows a LUT-based lowering that is
18838 // much faster, even faster than using native popcnt instructions.
18840 auto GetShift = [&](unsigned OpCode, SDValue V, int Shifter) {
18841 MVT VT = V.getSimpleValueType();
18842 SmallVector<SDValue, 32> Shifters(
18843 VT.getVectorNumElements(),
18844 DAG.getConstant(Shifter, DL, VT.getVectorElementType()));
18845 return DAG.getNode(OpCode, DL, VT, V,
18846 DAG.getNode(ISD::BUILD_VECTOR, DL, VT, Shifters));
18848 auto GetMask = [&](SDValue V, APInt Mask) {
18849 MVT VT = V.getSimpleValueType();
18850 SmallVector<SDValue, 32> Masks(
18851 VT.getVectorNumElements(),
18852 DAG.getConstant(Mask, DL, VT.getVectorElementType()));
18853 return DAG.getNode(ISD::AND, DL, VT, V,
18854 DAG.getNode(ISD::BUILD_VECTOR, DL, VT, Masks));
18857 // We don't want to incur the implicit masks required to SRL vNi8 vectors on
18858 // x86, so set the SRL type to have elements at least i16 wide. This is
18859 // correct because all of our SRLs are followed immediately by a mask anyways
18860 // that handles any bits that sneak into the high bits of the byte elements.
18861 MVT SrlVT = Len > 8 ? VT : MVT::getVectorVT(MVT::i16, VecSize / 16);
18865 // v = v - ((v >> 1) & 0x55555555...)
18867 DAG.getBitcast(VT, GetShift(ISD::SRL, DAG.getBitcast(SrlVT, V), 1));
18868 SDValue And = GetMask(Srl, APInt::getSplat(Len, APInt(8, 0x55)));
18869 V = DAG.getNode(ISD::SUB, DL, VT, V, And);
18871 // v = (v & 0x33333333...) + ((v >> 2) & 0x33333333...)
18872 SDValue AndLHS = GetMask(V, APInt::getSplat(Len, APInt(8, 0x33)));
18873 Srl = DAG.getBitcast(VT, GetShift(ISD::SRL, DAG.getBitcast(SrlVT, V), 2));
18874 SDValue AndRHS = GetMask(Srl, APInt::getSplat(Len, APInt(8, 0x33)));
18875 V = DAG.getNode(ISD::ADD, DL, VT, AndLHS, AndRHS);
18877 // v = (v + (v >> 4)) & 0x0F0F0F0F...
18878 Srl = DAG.getBitcast(VT, GetShift(ISD::SRL, DAG.getBitcast(SrlVT, V), 4));
18879 SDValue Add = DAG.getNode(ISD::ADD, DL, VT, V, Srl);
18880 V = GetMask(Add, APInt::getSplat(Len, APInt(8, 0x0F)));
18882 // At this point, V contains the byte-wise population count, and we are
18883 // merely doing a horizontal sum if necessary to get the wider element
18885 if (EltVT == MVT::i8)
18888 return LowerHorizontalByteSum(
18889 DAG.getBitcast(MVT::getVectorVT(MVT::i8, VecSize / 8), V), VT, Subtarget,
18893 static SDValue LowerVectorCTPOP(SDValue Op, const X86Subtarget *Subtarget,
18894 SelectionDAG &DAG) {
18895 MVT VT = Op.getSimpleValueType();
18896 // FIXME: Need to add AVX-512 support here!
18897 assert((VT.is256BitVector() || VT.is128BitVector()) &&
18898 "Unknown CTPOP type to handle");
18899 SDLoc DL(Op.getNode());
18900 SDValue Op0 = Op.getOperand(0);
18902 if (!Subtarget->hasSSSE3()) {
18903 // We can't use the fast LUT approach, so fall back on vectorized bitmath.
18904 assert(VT.is128BitVector() && "Only 128-bit vectors supported in SSE!");
18905 return LowerVectorCTPOPBitmath(Op0, DL, Subtarget, DAG);
18908 if (VT.is256BitVector() && !Subtarget->hasInt256()) {
18909 unsigned NumElems = VT.getVectorNumElements();
18911 // Extract each 128-bit vector, compute pop count and concat the result.
18912 SDValue LHS = Extract128BitVector(Op0, 0, DAG, DL);
18913 SDValue RHS = Extract128BitVector(Op0, NumElems/2, DAG, DL);
18915 return DAG.getNode(ISD::CONCAT_VECTORS, DL, VT,
18916 LowerVectorCTPOPInRegLUT(LHS, DL, Subtarget, DAG),
18917 LowerVectorCTPOPInRegLUT(RHS, DL, Subtarget, DAG));
18920 return LowerVectorCTPOPInRegLUT(Op0, DL, Subtarget, DAG);
18923 static SDValue LowerCTPOP(SDValue Op, const X86Subtarget *Subtarget,
18924 SelectionDAG &DAG) {
18925 assert(Op.getValueType().isVector() &&
18926 "We only do custom lowering for vector population count.");
18927 return LowerVectorCTPOP(Op, Subtarget, DAG);
18930 static SDValue LowerLOAD_SUB(SDValue Op, SelectionDAG &DAG) {
18931 SDNode *Node = Op.getNode();
18933 EVT T = Node->getValueType(0);
18934 SDValue negOp = DAG.getNode(ISD::SUB, dl, T,
18935 DAG.getConstant(0, dl, T), Node->getOperand(2));
18936 return DAG.getAtomic(ISD::ATOMIC_LOAD_ADD, dl,
18937 cast<AtomicSDNode>(Node)->getMemoryVT(),
18938 Node->getOperand(0),
18939 Node->getOperand(1), negOp,
18940 cast<AtomicSDNode>(Node)->getMemOperand(),
18941 cast<AtomicSDNode>(Node)->getOrdering(),
18942 cast<AtomicSDNode>(Node)->getSynchScope());
18945 static SDValue LowerATOMIC_STORE(SDValue Op, SelectionDAG &DAG) {
18946 SDNode *Node = Op.getNode();
18948 EVT VT = cast<AtomicSDNode>(Node)->getMemoryVT();
18950 // Convert seq_cst store -> xchg
18951 // Convert wide store -> swap (-> cmpxchg8b/cmpxchg16b)
18952 // FIXME: On 32-bit, store -> fist or movq would be more efficient
18953 // (The only way to get a 16-byte store is cmpxchg16b)
18954 // FIXME: 16-byte ATOMIC_SWAP isn't actually hooked up at the moment.
18955 if (cast<AtomicSDNode>(Node)->getOrdering() == SequentiallyConsistent ||
18956 !DAG.getTargetLoweringInfo().isTypeLegal(VT)) {
18957 SDValue Swap = DAG.getAtomic(ISD::ATOMIC_SWAP, dl,
18958 cast<AtomicSDNode>(Node)->getMemoryVT(),
18959 Node->getOperand(0),
18960 Node->getOperand(1), Node->getOperand(2),
18961 cast<AtomicSDNode>(Node)->getMemOperand(),
18962 cast<AtomicSDNode>(Node)->getOrdering(),
18963 cast<AtomicSDNode>(Node)->getSynchScope());
18964 return Swap.getValue(1);
18966 // Other atomic stores have a simple pattern.
18970 static SDValue LowerADDC_ADDE_SUBC_SUBE(SDValue Op, SelectionDAG &DAG) {
18971 EVT VT = Op.getNode()->getSimpleValueType(0);
18973 // Let legalize expand this if it isn't a legal type yet.
18974 if (!DAG.getTargetLoweringInfo().isTypeLegal(VT))
18977 SDVTList VTs = DAG.getVTList(VT, MVT::i32);
18980 bool ExtraOp = false;
18981 switch (Op.getOpcode()) {
18982 default: llvm_unreachable("Invalid code");
18983 case ISD::ADDC: Opc = X86ISD::ADD; break;
18984 case ISD::ADDE: Opc = X86ISD::ADC; ExtraOp = true; break;
18985 case ISD::SUBC: Opc = X86ISD::SUB; break;
18986 case ISD::SUBE: Opc = X86ISD::SBB; ExtraOp = true; break;
18990 return DAG.getNode(Opc, SDLoc(Op), VTs, Op.getOperand(0),
18992 return DAG.getNode(Opc, SDLoc(Op), VTs, Op.getOperand(0),
18993 Op.getOperand(1), Op.getOperand(2));
18996 static SDValue LowerFSINCOS(SDValue Op, const X86Subtarget *Subtarget,
18997 SelectionDAG &DAG) {
18998 assert(Subtarget->isTargetDarwin() && Subtarget->is64Bit());
19000 // For MacOSX, we want to call an alternative entry point: __sincos_stret,
19001 // which returns the values as { float, float } (in XMM0) or
19002 // { double, double } (which is returned in XMM0, XMM1).
19004 SDValue Arg = Op.getOperand(0);
19005 EVT ArgVT = Arg.getValueType();
19006 Type *ArgTy = ArgVT.getTypeForEVT(*DAG.getContext());
19008 TargetLowering::ArgListTy Args;
19009 TargetLowering::ArgListEntry Entry;
19013 Entry.isSExt = false;
19014 Entry.isZExt = false;
19015 Args.push_back(Entry);
19017 bool isF64 = ArgVT == MVT::f64;
19018 // Only optimize x86_64 for now. i386 is a bit messy. For f32,
19019 // the small struct {f32, f32} is returned in (eax, edx). For f64,
19020 // the results are returned via SRet in memory.
19021 const char *LibcallName = isF64 ? "__sincos_stret" : "__sincosf_stret";
19022 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
19024 DAG.getExternalSymbol(LibcallName, TLI.getPointerTy(DAG.getDataLayout()));
19026 Type *RetTy = isF64
19027 ? (Type*)StructType::get(ArgTy, ArgTy, nullptr)
19028 : (Type*)VectorType::get(ArgTy, 4);
19030 TargetLowering::CallLoweringInfo CLI(DAG);
19031 CLI.setDebugLoc(dl).setChain(DAG.getEntryNode())
19032 .setCallee(CallingConv::C, RetTy, Callee, std::move(Args), 0);
19034 std::pair<SDValue, SDValue> CallResult = TLI.LowerCallTo(CLI);
19037 // Returned in xmm0 and xmm1.
19038 return CallResult.first;
19040 // Returned in bits 0:31 and 32:64 xmm0.
19041 SDValue SinVal = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, ArgVT,
19042 CallResult.first, DAG.getIntPtrConstant(0, dl));
19043 SDValue CosVal = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, ArgVT,
19044 CallResult.first, DAG.getIntPtrConstant(1, dl));
19045 SDVTList Tys = DAG.getVTList(ArgVT, ArgVT);
19046 return DAG.getNode(ISD::MERGE_VALUES, dl, Tys, SinVal, CosVal);
19049 static SDValue LowerMSCATTER(SDValue Op, const X86Subtarget *Subtarget,
19050 SelectionDAG &DAG) {
19051 assert(Subtarget->hasAVX512() &&
19052 "MGATHER/MSCATTER are supported on AVX-512 arch only");
19054 MaskedScatterSDNode *N = cast<MaskedScatterSDNode>(Op.getNode());
19055 EVT VT = N->getValue().getValueType();
19056 assert(VT.getScalarSizeInBits() >= 32 && "Unsupported scatter op");
19059 // X86 scatter kills mask register, so its type should be added to
19060 // the list of return values
19061 if (N->getNumValues() == 1) {
19062 SDValue Index = N->getIndex();
19063 if (!Subtarget->hasVLX() && !VT.is512BitVector() &&
19064 !Index.getValueType().is512BitVector())
19065 Index = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v8i64, Index);
19067 SDVTList VTs = DAG.getVTList(N->getMask().getValueType(), MVT::Other);
19068 SDValue Ops[] = { N->getOperand(0), N->getOperand(1), N->getOperand(2),
19069 N->getOperand(3), Index };
19071 SDValue NewScatter = DAG.getMaskedScatter(VTs, VT, dl, Ops, N->getMemOperand());
19072 DAG.ReplaceAllUsesWith(Op, SDValue(NewScatter.getNode(), 1));
19073 return SDValue(NewScatter.getNode(), 0);
19078 static SDValue LowerMGATHER(SDValue Op, const X86Subtarget *Subtarget,
19079 SelectionDAG &DAG) {
19080 assert(Subtarget->hasAVX512() &&
19081 "MGATHER/MSCATTER are supported on AVX-512 arch only");
19083 MaskedGatherSDNode *N = cast<MaskedGatherSDNode>(Op.getNode());
19084 EVT VT = Op.getValueType();
19085 assert(VT.getScalarSizeInBits() >= 32 && "Unsupported gather op");
19088 SDValue Index = N->getIndex();
19089 if (!Subtarget->hasVLX() && !VT.is512BitVector() &&
19090 !Index.getValueType().is512BitVector()) {
19091 Index = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v8i64, Index);
19092 SDValue Ops[] = { N->getOperand(0), N->getOperand(1), N->getOperand(2),
19093 N->getOperand(3), Index };
19094 DAG.UpdateNodeOperands(N, Ops);
19099 SDValue X86TargetLowering::LowerGC_TRANSITION_START(SDValue Op,
19100 SelectionDAG &DAG) const {
19101 // TODO: Eventually, the lowering of these nodes should be informed by or
19102 // deferred to the GC strategy for the function in which they appear. For
19103 // now, however, they must be lowered to something. Since they are logically
19104 // no-ops in the case of a null GC strategy (or a GC strategy which does not
19105 // require special handling for these nodes), lower them as literal NOOPs for
19107 SmallVector<SDValue, 2> Ops;
19109 Ops.push_back(Op.getOperand(0));
19110 if (Op->getGluedNode())
19111 Ops.push_back(Op->getOperand(Op->getNumOperands() - 1));
19114 SDVTList VTs = DAG.getVTList(MVT::Other, MVT::Glue);
19115 SDValue NOOP(DAG.getMachineNode(X86::NOOP, SDLoc(Op), VTs, Ops), 0);
19120 SDValue X86TargetLowering::LowerGC_TRANSITION_END(SDValue Op,
19121 SelectionDAG &DAG) const {
19122 // TODO: Eventually, the lowering of these nodes should be informed by or
19123 // deferred to the GC strategy for the function in which they appear. For
19124 // now, however, they must be lowered to something. Since they are logically
19125 // no-ops in the case of a null GC strategy (or a GC strategy which does not
19126 // require special handling for these nodes), lower them as literal NOOPs for
19128 SmallVector<SDValue, 2> Ops;
19130 Ops.push_back(Op.getOperand(0));
19131 if (Op->getGluedNode())
19132 Ops.push_back(Op->getOperand(Op->getNumOperands() - 1));
19135 SDVTList VTs = DAG.getVTList(MVT::Other, MVT::Glue);
19136 SDValue NOOP(DAG.getMachineNode(X86::NOOP, SDLoc(Op), VTs, Ops), 0);
19141 /// LowerOperation - Provide custom lowering hooks for some operations.
19143 SDValue X86TargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
19144 switch (Op.getOpcode()) {
19145 default: llvm_unreachable("Should not custom lower this!");
19146 case ISD::ATOMIC_FENCE: return LowerATOMIC_FENCE(Op, Subtarget, DAG);
19147 case ISD::ATOMIC_CMP_SWAP_WITH_SUCCESS:
19148 return LowerCMP_SWAP(Op, Subtarget, DAG);
19149 case ISD::CTPOP: return LowerCTPOP(Op, Subtarget, DAG);
19150 case ISD::ATOMIC_LOAD_SUB: return LowerLOAD_SUB(Op,DAG);
19151 case ISD::ATOMIC_STORE: return LowerATOMIC_STORE(Op,DAG);
19152 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG);
19153 case ISD::CONCAT_VECTORS: return LowerCONCAT_VECTORS(Op, Subtarget, DAG);
19154 case ISD::VECTOR_SHUFFLE: return lowerVectorShuffle(Op, Subtarget, DAG);
19155 case ISD::VSELECT: return LowerVSELECT(Op, DAG);
19156 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
19157 case ISD::INSERT_VECTOR_ELT: return LowerINSERT_VECTOR_ELT(Op, DAG);
19158 case ISD::EXTRACT_SUBVECTOR: return LowerEXTRACT_SUBVECTOR(Op,Subtarget,DAG);
19159 case ISD::INSERT_SUBVECTOR: return LowerINSERT_SUBVECTOR(Op, Subtarget,DAG);
19160 case ISD::SCALAR_TO_VECTOR: return LowerSCALAR_TO_VECTOR(Op, DAG);
19161 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
19162 case ISD::GlobalAddress: return LowerGlobalAddress(Op, DAG);
19163 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
19164 case ISD::ExternalSymbol: return LowerExternalSymbol(Op, DAG);
19165 case ISD::BlockAddress: return LowerBlockAddress(Op, DAG);
19166 case ISD::SHL_PARTS:
19167 case ISD::SRA_PARTS:
19168 case ISD::SRL_PARTS: return LowerShiftParts(Op, DAG);
19169 case ISD::SINT_TO_FP: return LowerSINT_TO_FP(Op, DAG);
19170 case ISD::UINT_TO_FP: return LowerUINT_TO_FP(Op, DAG);
19171 case ISD::TRUNCATE: return LowerTRUNCATE(Op, DAG);
19172 case ISD::ZERO_EXTEND: return LowerZERO_EXTEND(Op, Subtarget, DAG);
19173 case ISD::SIGN_EXTEND: return LowerSIGN_EXTEND(Op, Subtarget, DAG);
19174 case ISD::ANY_EXTEND: return LowerANY_EXTEND(Op, Subtarget, DAG);
19175 case ISD::SIGN_EXTEND_VECTOR_INREG:
19176 return LowerSIGN_EXTEND_VECTOR_INREG(Op, Subtarget, DAG);
19177 case ISD::FP_TO_SINT: return LowerFP_TO_SINT(Op, DAG);
19178 case ISD::FP_TO_UINT: return LowerFP_TO_UINT(Op, DAG);
19179 case ISD::FP_EXTEND: return LowerFP_EXTEND(Op, DAG);
19180 case ISD::LOAD: return LowerExtendedLoad(Op, Subtarget, DAG);
19182 case ISD::FNEG: return LowerFABSorFNEG(Op, DAG);
19183 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
19184 case ISD::FGETSIGN: return LowerFGETSIGN(Op, DAG);
19185 case ISD::SETCC: return LowerSETCC(Op, DAG);
19186 case ISD::SELECT: return LowerSELECT(Op, DAG);
19187 case ISD::BRCOND: return LowerBRCOND(Op, DAG);
19188 case ISD::JumpTable: return LowerJumpTable(Op, DAG);
19189 case ISD::VASTART: return LowerVASTART(Op, DAG);
19190 case ISD::VAARG: return LowerVAARG(Op, DAG);
19191 case ISD::VACOPY: return LowerVACOPY(Op, Subtarget, DAG);
19192 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, Subtarget, DAG);
19193 case ISD::INTRINSIC_VOID:
19194 case ISD::INTRINSIC_W_CHAIN: return LowerINTRINSIC_W_CHAIN(Op, Subtarget, DAG);
19195 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
19196 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
19197 case ISD::FRAME_TO_ARGS_OFFSET:
19198 return LowerFRAME_TO_ARGS_OFFSET(Op, DAG);
19199 case ISD::DYNAMIC_STACKALLOC: return LowerDYNAMIC_STACKALLOC(Op, DAG);
19200 case ISD::EH_RETURN: return LowerEH_RETURN(Op, DAG);
19201 case ISD::CLEANUPRET: return LowerCLEANUPRET(Op, DAG);
19202 case ISD::CATCHRET: return LowerCATCHRET(Op, DAG);
19203 case ISD::EH_SJLJ_SETJMP: return lowerEH_SJLJ_SETJMP(Op, DAG);
19204 case ISD::EH_SJLJ_LONGJMP: return lowerEH_SJLJ_LONGJMP(Op, DAG);
19205 case ISD::INIT_TRAMPOLINE: return LowerINIT_TRAMPOLINE(Op, DAG);
19206 case ISD::ADJUST_TRAMPOLINE: return LowerADJUST_TRAMPOLINE(Op, DAG);
19207 case ISD::FLT_ROUNDS_: return LowerFLT_ROUNDS_(Op, DAG);
19208 case ISD::CTLZ: return LowerCTLZ(Op, DAG);
19209 case ISD::CTLZ_ZERO_UNDEF: return LowerCTLZ_ZERO_UNDEF(Op, DAG);
19210 case ISD::CTTZ: return LowerCTTZ(Op, DAG);
19211 case ISD::MUL: return LowerMUL(Op, Subtarget, DAG);
19212 case ISD::UMUL_LOHI:
19213 case ISD::SMUL_LOHI: return LowerMUL_LOHI(Op, Subtarget, DAG);
19216 case ISD::SHL: return LowerShift(Op, Subtarget, DAG);
19222 case ISD::UMULO: return LowerXALUO(Op, DAG);
19223 case ISD::READCYCLECOUNTER: return LowerREADCYCLECOUNTER(Op, Subtarget,DAG);
19224 case ISD::BITCAST: return LowerBITCAST(Op, Subtarget, DAG);
19228 case ISD::SUBE: return LowerADDC_ADDE_SUBC_SUBE(Op, DAG);
19229 case ISD::ADD: return LowerADD(Op, DAG);
19230 case ISD::SUB: return LowerSUB(Op, DAG);
19234 case ISD::UMIN: return LowerMINMAX(Op, DAG);
19235 case ISD::FSINCOS: return LowerFSINCOS(Op, Subtarget, DAG);
19236 case ISD::MGATHER: return LowerMGATHER(Op, Subtarget, DAG);
19237 case ISD::MSCATTER: return LowerMSCATTER(Op, Subtarget, DAG);
19238 case ISD::GC_TRANSITION_START:
19239 return LowerGC_TRANSITION_START(Op, DAG);
19240 case ISD::GC_TRANSITION_END: return LowerGC_TRANSITION_END(Op, DAG);
19244 /// ReplaceNodeResults - Replace a node with an illegal result type
19245 /// with a new node built out of custom code.
19246 void X86TargetLowering::ReplaceNodeResults(SDNode *N,
19247 SmallVectorImpl<SDValue>&Results,
19248 SelectionDAG &DAG) const {
19250 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
19251 switch (N->getOpcode()) {
19253 llvm_unreachable("Do not know how to custom type legalize this operation!");
19254 // We might have generated v2f32 FMIN/FMAX operations. Widen them to v4f32.
19255 case X86ISD::FMINC:
19257 case X86ISD::FMAXC:
19258 case X86ISD::FMAX: {
19259 EVT VT = N->getValueType(0);
19260 if (VT != MVT::v2f32)
19261 llvm_unreachable("Unexpected type (!= v2f32) on FMIN/FMAX.");
19262 SDValue UNDEF = DAG.getUNDEF(VT);
19263 SDValue LHS = DAG.getNode(ISD::CONCAT_VECTORS, dl, MVT::v4f32,
19264 N->getOperand(0), UNDEF);
19265 SDValue RHS = DAG.getNode(ISD::CONCAT_VECTORS, dl, MVT::v4f32,
19266 N->getOperand(1), UNDEF);
19267 Results.push_back(DAG.getNode(N->getOpcode(), dl, MVT::v4f32, LHS, RHS));
19270 case ISD::SIGN_EXTEND_INREG:
19275 // We don't want to expand or promote these.
19282 case ISD::UDIVREM: {
19283 SDValue V = LowerWin64_i128OP(SDValue(N,0), DAG);
19284 Results.push_back(V);
19287 case ISD::FP_TO_SINT:
19288 case ISD::FP_TO_UINT: {
19289 bool IsSigned = N->getOpcode() == ISD::FP_TO_SINT;
19291 std::pair<SDValue,SDValue> Vals =
19292 FP_TO_INTHelper(SDValue(N, 0), DAG, IsSigned, /*IsReplace=*/ true);
19293 SDValue FIST = Vals.first, StackSlot = Vals.second;
19294 if (FIST.getNode()) {
19295 EVT VT = N->getValueType(0);
19296 // Return a load from the stack slot.
19297 if (StackSlot.getNode())
19298 Results.push_back(DAG.getLoad(VT, dl, FIST, StackSlot,
19299 MachinePointerInfo(),
19300 false, false, false, 0));
19302 Results.push_back(FIST);
19306 case ISD::UINT_TO_FP: {
19307 assert(Subtarget->hasSSE2() && "Requires at least SSE2!");
19308 if (N->getOperand(0).getValueType() != MVT::v2i32 ||
19309 N->getValueType(0) != MVT::v2f32)
19311 SDValue ZExtIn = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v2i64,
19313 SDValue Bias = DAG.getConstantFP(BitsToDouble(0x4330000000000000ULL), dl,
19315 SDValue VBias = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v2f64, Bias, Bias);
19316 SDValue Or = DAG.getNode(ISD::OR, dl, MVT::v2i64, ZExtIn,
19317 DAG.getBitcast(MVT::v2i64, VBias));
19318 Or = DAG.getBitcast(MVT::v2f64, Or);
19319 // TODO: Are there any fast-math-flags to propagate here?
19320 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::v2f64, Or, VBias);
19321 Results.push_back(DAG.getNode(X86ISD::VFPROUND, dl, MVT::v4f32, Sub));
19324 case ISD::FP_ROUND: {
19325 if (!TLI.isTypeLegal(N->getOperand(0).getValueType()))
19327 SDValue V = DAG.getNode(X86ISD::VFPROUND, dl, MVT::v4f32, N->getOperand(0));
19328 Results.push_back(V);
19331 case ISD::FP_EXTEND: {
19332 // Right now, only MVT::v2f32 has OperationAction for FP_EXTEND.
19333 // No other ValueType for FP_EXTEND should reach this point.
19334 assert(N->getValueType(0) == MVT::v2f32 &&
19335 "Do not know how to legalize this Node");
19338 case ISD::INTRINSIC_W_CHAIN: {
19339 unsigned IntNo = cast<ConstantSDNode>(N->getOperand(1))->getZExtValue();
19341 default : llvm_unreachable("Do not know how to custom type "
19342 "legalize this intrinsic operation!");
19343 case Intrinsic::x86_rdtsc:
19344 return getReadTimeStampCounter(N, dl, X86ISD::RDTSC_DAG, DAG, Subtarget,
19346 case Intrinsic::x86_rdtscp:
19347 return getReadTimeStampCounter(N, dl, X86ISD::RDTSCP_DAG, DAG, Subtarget,
19349 case Intrinsic::x86_rdpmc:
19350 return getReadPerformanceCounter(N, dl, DAG, Subtarget, Results);
19353 case ISD::READCYCLECOUNTER: {
19354 return getReadTimeStampCounter(N, dl, X86ISD::RDTSC_DAG, DAG, Subtarget,
19357 case ISD::ATOMIC_CMP_SWAP_WITH_SUCCESS: {
19358 EVT T = N->getValueType(0);
19359 assert((T == MVT::i64 || T == MVT::i128) && "can only expand cmpxchg pair");
19360 bool Regs64bit = T == MVT::i128;
19361 EVT HalfT = Regs64bit ? MVT::i64 : MVT::i32;
19362 SDValue cpInL, cpInH;
19363 cpInL = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, HalfT, N->getOperand(2),
19364 DAG.getConstant(0, dl, HalfT));
19365 cpInH = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, HalfT, N->getOperand(2),
19366 DAG.getConstant(1, dl, HalfT));
19367 cpInL = DAG.getCopyToReg(N->getOperand(0), dl,
19368 Regs64bit ? X86::RAX : X86::EAX,
19370 cpInH = DAG.getCopyToReg(cpInL.getValue(0), dl,
19371 Regs64bit ? X86::RDX : X86::EDX,
19372 cpInH, cpInL.getValue(1));
19373 SDValue swapInL, swapInH;
19374 swapInL = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, HalfT, N->getOperand(3),
19375 DAG.getConstant(0, dl, HalfT));
19376 swapInH = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, HalfT, N->getOperand(3),
19377 DAG.getConstant(1, dl, HalfT));
19378 swapInL = DAG.getCopyToReg(cpInH.getValue(0), dl,
19379 Regs64bit ? X86::RBX : X86::EBX,
19380 swapInL, cpInH.getValue(1));
19381 swapInH = DAG.getCopyToReg(swapInL.getValue(0), dl,
19382 Regs64bit ? X86::RCX : X86::ECX,
19383 swapInH, swapInL.getValue(1));
19384 SDValue Ops[] = { swapInH.getValue(0),
19386 swapInH.getValue(1) };
19387 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Glue);
19388 MachineMemOperand *MMO = cast<AtomicSDNode>(N)->getMemOperand();
19389 unsigned Opcode = Regs64bit ? X86ISD::LCMPXCHG16_DAG :
19390 X86ISD::LCMPXCHG8_DAG;
19391 SDValue Result = DAG.getMemIntrinsicNode(Opcode, dl, Tys, Ops, T, MMO);
19392 SDValue cpOutL = DAG.getCopyFromReg(Result.getValue(0), dl,
19393 Regs64bit ? X86::RAX : X86::EAX,
19394 HalfT, Result.getValue(1));
19395 SDValue cpOutH = DAG.getCopyFromReg(cpOutL.getValue(1), dl,
19396 Regs64bit ? X86::RDX : X86::EDX,
19397 HalfT, cpOutL.getValue(2));
19398 SDValue OpsF[] = { cpOutL.getValue(0), cpOutH.getValue(0)};
19400 SDValue EFLAGS = DAG.getCopyFromReg(cpOutH.getValue(1), dl, X86::EFLAGS,
19401 MVT::i32, cpOutH.getValue(2));
19403 DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
19404 DAG.getConstant(X86::COND_E, dl, MVT::i8), EFLAGS);
19405 Success = DAG.getZExtOrTrunc(Success, dl, N->getValueType(1));
19407 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, T, OpsF));
19408 Results.push_back(Success);
19409 Results.push_back(EFLAGS.getValue(1));
19412 case ISD::ATOMIC_SWAP:
19413 case ISD::ATOMIC_LOAD_ADD:
19414 case ISD::ATOMIC_LOAD_SUB:
19415 case ISD::ATOMIC_LOAD_AND:
19416 case ISD::ATOMIC_LOAD_OR:
19417 case ISD::ATOMIC_LOAD_XOR:
19418 case ISD::ATOMIC_LOAD_NAND:
19419 case ISD::ATOMIC_LOAD_MIN:
19420 case ISD::ATOMIC_LOAD_MAX:
19421 case ISD::ATOMIC_LOAD_UMIN:
19422 case ISD::ATOMIC_LOAD_UMAX:
19423 case ISD::ATOMIC_LOAD: {
19424 // Delegate to generic TypeLegalization. Situations we can really handle
19425 // should have already been dealt with by AtomicExpandPass.cpp.
19428 case ISD::BITCAST: {
19429 assert(Subtarget->hasSSE2() && "Requires at least SSE2!");
19430 EVT DstVT = N->getValueType(0);
19431 EVT SrcVT = N->getOperand(0)->getValueType(0);
19433 if (SrcVT != MVT::f64 ||
19434 (DstVT != MVT::v2i32 && DstVT != MVT::v4i16 && DstVT != MVT::v8i8))
19437 unsigned NumElts = DstVT.getVectorNumElements();
19438 EVT SVT = DstVT.getVectorElementType();
19439 EVT WiderVT = EVT::getVectorVT(*DAG.getContext(), SVT, NumElts * 2);
19440 SDValue Expanded = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
19441 MVT::v2f64, N->getOperand(0));
19442 SDValue ToVecInt = DAG.getBitcast(WiderVT, Expanded);
19444 if (ExperimentalVectorWideningLegalization) {
19445 // If we are legalizing vectors by widening, we already have the desired
19446 // legal vector type, just return it.
19447 Results.push_back(ToVecInt);
19451 SmallVector<SDValue, 8> Elts;
19452 for (unsigned i = 0, e = NumElts; i != e; ++i)
19453 Elts.push_back(DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, SVT,
19454 ToVecInt, DAG.getIntPtrConstant(i, dl)));
19456 Results.push_back(DAG.getNode(ISD::BUILD_VECTOR, dl, DstVT, Elts));
19461 const char *X86TargetLowering::getTargetNodeName(unsigned Opcode) const {
19462 switch ((X86ISD::NodeType)Opcode) {
19463 case X86ISD::FIRST_NUMBER: break;
19464 case X86ISD::BSF: return "X86ISD::BSF";
19465 case X86ISD::BSR: return "X86ISD::BSR";
19466 case X86ISD::SHLD: return "X86ISD::SHLD";
19467 case X86ISD::SHRD: return "X86ISD::SHRD";
19468 case X86ISD::FAND: return "X86ISD::FAND";
19469 case X86ISD::FANDN: return "X86ISD::FANDN";
19470 case X86ISD::FOR: return "X86ISD::FOR";
19471 case X86ISD::FXOR: return "X86ISD::FXOR";
19472 case X86ISD::FILD: return "X86ISD::FILD";
19473 case X86ISD::FILD_FLAG: return "X86ISD::FILD_FLAG";
19474 case X86ISD::FP_TO_INT16_IN_MEM: return "X86ISD::FP_TO_INT16_IN_MEM";
19475 case X86ISD::FP_TO_INT32_IN_MEM: return "X86ISD::FP_TO_INT32_IN_MEM";
19476 case X86ISD::FP_TO_INT64_IN_MEM: return "X86ISD::FP_TO_INT64_IN_MEM";
19477 case X86ISD::FLD: return "X86ISD::FLD";
19478 case X86ISD::FST: return "X86ISD::FST";
19479 case X86ISD::CALL: return "X86ISD::CALL";
19480 case X86ISD::RDTSC_DAG: return "X86ISD::RDTSC_DAG";
19481 case X86ISD::RDTSCP_DAG: return "X86ISD::RDTSCP_DAG";
19482 case X86ISD::RDPMC_DAG: return "X86ISD::RDPMC_DAG";
19483 case X86ISD::BT: return "X86ISD::BT";
19484 case X86ISD::CMP: return "X86ISD::CMP";
19485 case X86ISD::COMI: return "X86ISD::COMI";
19486 case X86ISD::UCOMI: return "X86ISD::UCOMI";
19487 case X86ISD::CMPM: return "X86ISD::CMPM";
19488 case X86ISD::CMPMU: return "X86ISD::CMPMU";
19489 case X86ISD::CMPM_RND: return "X86ISD::CMPM_RND";
19490 case X86ISD::SETCC: return "X86ISD::SETCC";
19491 case X86ISD::SETCC_CARRY: return "X86ISD::SETCC_CARRY";
19492 case X86ISD::FSETCC: return "X86ISD::FSETCC";
19493 case X86ISD::FGETSIGNx86: return "X86ISD::FGETSIGNx86";
19494 case X86ISD::CMOV: return "X86ISD::CMOV";
19495 case X86ISD::BRCOND: return "X86ISD::BRCOND";
19496 case X86ISD::RET_FLAG: return "X86ISD::RET_FLAG";
19497 case X86ISD::REP_STOS: return "X86ISD::REP_STOS";
19498 case X86ISD::REP_MOVS: return "X86ISD::REP_MOVS";
19499 case X86ISD::GlobalBaseReg: return "X86ISD::GlobalBaseReg";
19500 case X86ISD::Wrapper: return "X86ISD::Wrapper";
19501 case X86ISD::WrapperRIP: return "X86ISD::WrapperRIP";
19502 case X86ISD::MOVDQ2Q: return "X86ISD::MOVDQ2Q";
19503 case X86ISD::MMX_MOVD2W: return "X86ISD::MMX_MOVD2W";
19504 case X86ISD::MMX_MOVW2D: return "X86ISD::MMX_MOVW2D";
19505 case X86ISD::PEXTRB: return "X86ISD::PEXTRB";
19506 case X86ISD::PEXTRW: return "X86ISD::PEXTRW";
19507 case X86ISD::INSERTPS: return "X86ISD::INSERTPS";
19508 case X86ISD::PINSRB: return "X86ISD::PINSRB";
19509 case X86ISD::PINSRW: return "X86ISD::PINSRW";
19510 case X86ISD::MMX_PINSRW: return "X86ISD::MMX_PINSRW";
19511 case X86ISD::PSHUFB: return "X86ISD::PSHUFB";
19512 case X86ISD::ANDNP: return "X86ISD::ANDNP";
19513 case X86ISD::PSIGN: return "X86ISD::PSIGN";
19514 case X86ISD::BLENDI: return "X86ISD::BLENDI";
19515 case X86ISD::SHRUNKBLEND: return "X86ISD::SHRUNKBLEND";
19516 case X86ISD::ADDUS: return "X86ISD::ADDUS";
19517 case X86ISD::SUBUS: return "X86ISD::SUBUS";
19518 case X86ISD::HADD: return "X86ISD::HADD";
19519 case X86ISD::HSUB: return "X86ISD::HSUB";
19520 case X86ISD::FHADD: return "X86ISD::FHADD";
19521 case X86ISD::FHSUB: return "X86ISD::FHSUB";
19522 case X86ISD::ABS: return "X86ISD::ABS";
19523 case X86ISD::CONFLICT: return "X86ISD::CONFLICT";
19524 case X86ISD::FMAX: return "X86ISD::FMAX";
19525 case X86ISD::FMAX_RND: return "X86ISD::FMAX_RND";
19526 case X86ISD::FMIN: return "X86ISD::FMIN";
19527 case X86ISD::FMIN_RND: return "X86ISD::FMIN_RND";
19528 case X86ISD::FMAXC: return "X86ISD::FMAXC";
19529 case X86ISD::FMINC: return "X86ISD::FMINC";
19530 case X86ISD::FRSQRT: return "X86ISD::FRSQRT";
19531 case X86ISD::FRCP: return "X86ISD::FRCP";
19532 case X86ISD::EXTRQI: return "X86ISD::EXTRQI";
19533 case X86ISD::INSERTQI: return "X86ISD::INSERTQI";
19534 case X86ISD::TLSADDR: return "X86ISD::TLSADDR";
19535 case X86ISD::TLSBASEADDR: return "X86ISD::TLSBASEADDR";
19536 case X86ISD::TLSCALL: return "X86ISD::TLSCALL";
19537 case X86ISD::EH_SJLJ_SETJMP: return "X86ISD::EH_SJLJ_SETJMP";
19538 case X86ISD::EH_SJLJ_LONGJMP: return "X86ISD::EH_SJLJ_LONGJMP";
19539 case X86ISD::EH_RETURN: return "X86ISD::EH_RETURN";
19540 case X86ISD::CATCHRET: return "X86ISD::CATCHRET";
19541 case X86ISD::CLEANUPRET: return "X86ISD::CLEANUPRET";
19542 case X86ISD::TC_RETURN: return "X86ISD::TC_RETURN";
19543 case X86ISD::FNSTCW16m: return "X86ISD::FNSTCW16m";
19544 case X86ISD::FNSTSW16r: return "X86ISD::FNSTSW16r";
19545 case X86ISD::LCMPXCHG_DAG: return "X86ISD::LCMPXCHG_DAG";
19546 case X86ISD::LCMPXCHG8_DAG: return "X86ISD::LCMPXCHG8_DAG";
19547 case X86ISD::LCMPXCHG16_DAG: return "X86ISD::LCMPXCHG16_DAG";
19548 case X86ISD::VZEXT_MOVL: return "X86ISD::VZEXT_MOVL";
19549 case X86ISD::VZEXT_LOAD: return "X86ISD::VZEXT_LOAD";
19550 case X86ISD::VZEXT: return "X86ISD::VZEXT";
19551 case X86ISD::VSEXT: return "X86ISD::VSEXT";
19552 case X86ISD::VTRUNC: return "X86ISD::VTRUNC";
19553 case X86ISD::VTRUNCS: return "X86ISD::VTRUNCS";
19554 case X86ISD::VTRUNCUS: return "X86ISD::VTRUNCUS";
19555 case X86ISD::VINSERT: return "X86ISD::VINSERT";
19556 case X86ISD::VFPEXT: return "X86ISD::VFPEXT";
19557 case X86ISD::VFPROUND: return "X86ISD::VFPROUND";
19558 case X86ISD::CVTDQ2PD: return "X86ISD::CVTDQ2PD";
19559 case X86ISD::CVTUDQ2PD: return "X86ISD::CVTUDQ2PD";
19560 case X86ISD::VSHLDQ: return "X86ISD::VSHLDQ";
19561 case X86ISD::VSRLDQ: return "X86ISD::VSRLDQ";
19562 case X86ISD::VSHL: return "X86ISD::VSHL";
19563 case X86ISD::VSRL: return "X86ISD::VSRL";
19564 case X86ISD::VSRA: return "X86ISD::VSRA";
19565 case X86ISD::VSHLI: return "X86ISD::VSHLI";
19566 case X86ISD::VSRLI: return "X86ISD::VSRLI";
19567 case X86ISD::VSRAI: return "X86ISD::VSRAI";
19568 case X86ISD::CMPP: return "X86ISD::CMPP";
19569 case X86ISD::PCMPEQ: return "X86ISD::PCMPEQ";
19570 case X86ISD::PCMPGT: return "X86ISD::PCMPGT";
19571 case X86ISD::PCMPEQM: return "X86ISD::PCMPEQM";
19572 case X86ISD::PCMPGTM: return "X86ISD::PCMPGTM";
19573 case X86ISD::ADD: return "X86ISD::ADD";
19574 case X86ISD::SUB: return "X86ISD::SUB";
19575 case X86ISD::ADC: return "X86ISD::ADC";
19576 case X86ISD::SBB: return "X86ISD::SBB";
19577 case X86ISD::SMUL: return "X86ISD::SMUL";
19578 case X86ISD::UMUL: return "X86ISD::UMUL";
19579 case X86ISD::SMUL8: return "X86ISD::SMUL8";
19580 case X86ISD::UMUL8: return "X86ISD::UMUL8";
19581 case X86ISD::SDIVREM8_SEXT_HREG: return "X86ISD::SDIVREM8_SEXT_HREG";
19582 case X86ISD::UDIVREM8_ZEXT_HREG: return "X86ISD::UDIVREM8_ZEXT_HREG";
19583 case X86ISD::INC: return "X86ISD::INC";
19584 case X86ISD::DEC: return "X86ISD::DEC";
19585 case X86ISD::OR: return "X86ISD::OR";
19586 case X86ISD::XOR: return "X86ISD::XOR";
19587 case X86ISD::AND: return "X86ISD::AND";
19588 case X86ISD::BEXTR: return "X86ISD::BEXTR";
19589 case X86ISD::MUL_IMM: return "X86ISD::MUL_IMM";
19590 case X86ISD::PTEST: return "X86ISD::PTEST";
19591 case X86ISD::TESTP: return "X86ISD::TESTP";
19592 case X86ISD::TESTM: return "X86ISD::TESTM";
19593 case X86ISD::TESTNM: return "X86ISD::TESTNM";
19594 case X86ISD::KORTEST: return "X86ISD::KORTEST";
19595 case X86ISD::KTEST: return "X86ISD::KTEST";
19596 case X86ISD::PACKSS: return "X86ISD::PACKSS";
19597 case X86ISD::PACKUS: return "X86ISD::PACKUS";
19598 case X86ISD::PALIGNR: return "X86ISD::PALIGNR";
19599 case X86ISD::VALIGN: return "X86ISD::VALIGN";
19600 case X86ISD::PSHUFD: return "X86ISD::PSHUFD";
19601 case X86ISD::PSHUFHW: return "X86ISD::PSHUFHW";
19602 case X86ISD::PSHUFLW: return "X86ISD::PSHUFLW";
19603 case X86ISD::SHUFP: return "X86ISD::SHUFP";
19604 case X86ISD::SHUF128: return "X86ISD::SHUF128";
19605 case X86ISD::MOVLHPS: return "X86ISD::MOVLHPS";
19606 case X86ISD::MOVLHPD: return "X86ISD::MOVLHPD";
19607 case X86ISD::MOVHLPS: return "X86ISD::MOVHLPS";
19608 case X86ISD::MOVLPS: return "X86ISD::MOVLPS";
19609 case X86ISD::MOVLPD: return "X86ISD::MOVLPD";
19610 case X86ISD::MOVDDUP: return "X86ISD::MOVDDUP";
19611 case X86ISD::MOVSHDUP: return "X86ISD::MOVSHDUP";
19612 case X86ISD::MOVSLDUP: return "X86ISD::MOVSLDUP";
19613 case X86ISD::MOVSD: return "X86ISD::MOVSD";
19614 case X86ISD::MOVSS: return "X86ISD::MOVSS";
19615 case X86ISD::UNPCKL: return "X86ISD::UNPCKL";
19616 case X86ISD::UNPCKH: return "X86ISD::UNPCKH";
19617 case X86ISD::VBROADCAST: return "X86ISD::VBROADCAST";
19618 case X86ISD::SUBV_BROADCAST: return "X86ISD::SUBV_BROADCAST";
19619 case X86ISD::VEXTRACT: return "X86ISD::VEXTRACT";
19620 case X86ISD::VPERMILPV: return "X86ISD::VPERMILPV";
19621 case X86ISD::VPERMILPI: return "X86ISD::VPERMILPI";
19622 case X86ISD::VPERM2X128: return "X86ISD::VPERM2X128";
19623 case X86ISD::VPERMV: return "X86ISD::VPERMV";
19624 case X86ISD::VPERMV3: return "X86ISD::VPERMV3";
19625 case X86ISD::VPERMIV3: return "X86ISD::VPERMIV3";
19626 case X86ISD::VPERMI: return "X86ISD::VPERMI";
19627 case X86ISD::VFIXUPIMM: return "X86ISD::VFIXUPIMM";
19628 case X86ISD::VRANGE: return "X86ISD::VRANGE";
19629 case X86ISD::PMULUDQ: return "X86ISD::PMULUDQ";
19630 case X86ISD::PMULDQ: return "X86ISD::PMULDQ";
19631 case X86ISD::PSADBW: return "X86ISD::PSADBW";
19632 case X86ISD::DBPSADBW: return "X86ISD::DBPSADBW";
19633 case X86ISD::VASTART_SAVE_XMM_REGS: return "X86ISD::VASTART_SAVE_XMM_REGS";
19634 case X86ISD::VAARG_64: return "X86ISD::VAARG_64";
19635 case X86ISD::WIN_ALLOCA: return "X86ISD::WIN_ALLOCA";
19636 case X86ISD::MEMBARRIER: return "X86ISD::MEMBARRIER";
19637 case X86ISD::MFENCE: return "X86ISD::MFENCE";
19638 case X86ISD::SFENCE: return "X86ISD::SFENCE";
19639 case X86ISD::LFENCE: return "X86ISD::LFENCE";
19640 case X86ISD::SEG_ALLOCA: return "X86ISD::SEG_ALLOCA";
19641 case X86ISD::SAHF: return "X86ISD::SAHF";
19642 case X86ISD::RDRAND: return "X86ISD::RDRAND";
19643 case X86ISD::RDSEED: return "X86ISD::RDSEED";
19644 case X86ISD::VPMADDUBSW: return "X86ISD::VPMADDUBSW";
19645 case X86ISD::VPMADDWD: return "X86ISD::VPMADDWD";
19646 case X86ISD::FMADD: return "X86ISD::FMADD";
19647 case X86ISD::FMSUB: return "X86ISD::FMSUB";
19648 case X86ISD::FNMADD: return "X86ISD::FNMADD";
19649 case X86ISD::FNMSUB: return "X86ISD::FNMSUB";
19650 case X86ISD::FMADDSUB: return "X86ISD::FMADDSUB";
19651 case X86ISD::FMSUBADD: return "X86ISD::FMSUBADD";
19652 case X86ISD::FMADD_RND: return "X86ISD::FMADD_RND";
19653 case X86ISD::FNMADD_RND: return "X86ISD::FNMADD_RND";
19654 case X86ISD::FMSUB_RND: return "X86ISD::FMSUB_RND";
19655 case X86ISD::FNMSUB_RND: return "X86ISD::FNMSUB_RND";
19656 case X86ISD::FMADDSUB_RND: return "X86ISD::FMADDSUB_RND";
19657 case X86ISD::FMSUBADD_RND: return "X86ISD::FMSUBADD_RND";
19658 case X86ISD::VRNDSCALE: return "X86ISD::VRNDSCALE";
19659 case X86ISD::VREDUCE: return "X86ISD::VREDUCE";
19660 case X86ISD::VGETMANT: return "X86ISD::VGETMANT";
19661 case X86ISD::PCMPESTRI: return "X86ISD::PCMPESTRI";
19662 case X86ISD::PCMPISTRI: return "X86ISD::PCMPISTRI";
19663 case X86ISD::XTEST: return "X86ISD::XTEST";
19664 case X86ISD::COMPRESS: return "X86ISD::COMPRESS";
19665 case X86ISD::EXPAND: return "X86ISD::EXPAND";
19666 case X86ISD::SELECT: return "X86ISD::SELECT";
19667 case X86ISD::ADDSUB: return "X86ISD::ADDSUB";
19668 case X86ISD::RCP28: return "X86ISD::RCP28";
19669 case X86ISD::EXP2: return "X86ISD::EXP2";
19670 case X86ISD::RSQRT28: return "X86ISD::RSQRT28";
19671 case X86ISD::FADD_RND: return "X86ISD::FADD_RND";
19672 case X86ISD::FSUB_RND: return "X86ISD::FSUB_RND";
19673 case X86ISD::FMUL_RND: return "X86ISD::FMUL_RND";
19674 case X86ISD::FDIV_RND: return "X86ISD::FDIV_RND";
19675 case X86ISD::FSQRT_RND: return "X86ISD::FSQRT_RND";
19676 case X86ISD::FGETEXP_RND: return "X86ISD::FGETEXP_RND";
19677 case X86ISD::SCALEF: return "X86ISD::SCALEF";
19678 case X86ISD::ADDS: return "X86ISD::ADDS";
19679 case X86ISD::SUBS: return "X86ISD::SUBS";
19680 case X86ISD::AVG: return "X86ISD::AVG";
19681 case X86ISD::MULHRS: return "X86ISD::MULHRS";
19682 case X86ISD::SINT_TO_FP_RND: return "X86ISD::SINT_TO_FP_RND";
19683 case X86ISD::UINT_TO_FP_RND: return "X86ISD::UINT_TO_FP_RND";
19684 case X86ISD::FP_TO_SINT_RND: return "X86ISD::FP_TO_SINT_RND";
19685 case X86ISD::FP_TO_UINT_RND: return "X86ISD::FP_TO_UINT_RND";
19690 // isLegalAddressingMode - Return true if the addressing mode represented
19691 // by AM is legal for this target, for a load/store of the specified type.
19692 bool X86TargetLowering::isLegalAddressingMode(const DataLayout &DL,
19693 const AddrMode &AM, Type *Ty,
19694 unsigned AS) const {
19695 // X86 supports extremely general addressing modes.
19696 CodeModel::Model M = getTargetMachine().getCodeModel();
19697 Reloc::Model R = getTargetMachine().getRelocationModel();
19699 // X86 allows a sign-extended 32-bit immediate field as a displacement.
19700 if (!X86::isOffsetSuitableForCodeModel(AM.BaseOffs, M, AM.BaseGV != nullptr))
19705 Subtarget->ClassifyGlobalReference(AM.BaseGV, getTargetMachine());
19707 // If a reference to this global requires an extra load, we can't fold it.
19708 if (isGlobalStubReference(GVFlags))
19711 // If BaseGV requires a register for the PIC base, we cannot also have a
19712 // BaseReg specified.
19713 if (AM.HasBaseReg && isGlobalRelativeToPICBase(GVFlags))
19716 // If lower 4G is not available, then we must use rip-relative addressing.
19717 if ((M != CodeModel::Small || R != Reloc::Static) &&
19718 Subtarget->is64Bit() && (AM.BaseOffs || AM.Scale > 1))
19722 switch (AM.Scale) {
19728 // These scales always work.
19733 // These scales are formed with basereg+scalereg. Only accept if there is
19738 default: // Other stuff never works.
19745 bool X86TargetLowering::isVectorShiftByScalarCheap(Type *Ty) const {
19746 unsigned Bits = Ty->getScalarSizeInBits();
19748 // 8-bit shifts are always expensive, but versions with a scalar amount aren't
19749 // particularly cheaper than those without.
19753 // On AVX2 there are new vpsllv[dq] instructions (and other shifts), that make
19754 // variable shifts just as cheap as scalar ones.
19755 if (Subtarget->hasInt256() && (Bits == 32 || Bits == 64))
19758 // Otherwise, it's significantly cheaper to shift by a scalar amount than by a
19759 // fully general vector.
19763 bool X86TargetLowering::isTruncateFree(Type *Ty1, Type *Ty2) const {
19764 if (!Ty1->isIntegerTy() || !Ty2->isIntegerTy())
19766 unsigned NumBits1 = Ty1->getPrimitiveSizeInBits();
19767 unsigned NumBits2 = Ty2->getPrimitiveSizeInBits();
19768 return NumBits1 > NumBits2;
19771 bool X86TargetLowering::allowTruncateForTailCall(Type *Ty1, Type *Ty2) const {
19772 if (!Ty1->isIntegerTy() || !Ty2->isIntegerTy())
19775 if (!isTypeLegal(EVT::getEVT(Ty1)))
19778 assert(Ty1->getPrimitiveSizeInBits() <= 64 && "i128 is probably not a noop");
19780 // Assuming the caller doesn't have a zeroext or signext return parameter,
19781 // truncation all the way down to i1 is valid.
19785 bool X86TargetLowering::isLegalICmpImmediate(int64_t Imm) const {
19786 return isInt<32>(Imm);
19789 bool X86TargetLowering::isLegalAddImmediate(int64_t Imm) const {
19790 // Can also use sub to handle negated immediates.
19791 return isInt<32>(Imm);
19794 bool X86TargetLowering::isTruncateFree(EVT VT1, EVT VT2) const {
19795 if (!VT1.isInteger() || !VT2.isInteger())
19797 unsigned NumBits1 = VT1.getSizeInBits();
19798 unsigned NumBits2 = VT2.getSizeInBits();
19799 return NumBits1 > NumBits2;
19802 bool X86TargetLowering::isZExtFree(Type *Ty1, Type *Ty2) const {
19803 // x86-64 implicitly zero-extends 32-bit results in 64-bit registers.
19804 return Ty1->isIntegerTy(32) && Ty2->isIntegerTy(64) && Subtarget->is64Bit();
19807 bool X86TargetLowering::isZExtFree(EVT VT1, EVT VT2) const {
19808 // x86-64 implicitly zero-extends 32-bit results in 64-bit registers.
19809 return VT1 == MVT::i32 && VT2 == MVT::i64 && Subtarget->is64Bit();
19812 bool X86TargetLowering::isZExtFree(SDValue Val, EVT VT2) const {
19813 EVT VT1 = Val.getValueType();
19814 if (isZExtFree(VT1, VT2))
19817 if (Val.getOpcode() != ISD::LOAD)
19820 if (!VT1.isSimple() || !VT1.isInteger() ||
19821 !VT2.isSimple() || !VT2.isInteger())
19824 switch (VT1.getSimpleVT().SimpleTy) {
19829 // X86 has 8, 16, and 32-bit zero-extending loads.
19836 bool X86TargetLowering::isVectorLoadExtDesirable(SDValue) const { return true; }
19839 X86TargetLowering::isFMAFasterThanFMulAndFAdd(EVT VT) const {
19840 if (!(Subtarget->hasFMA() || Subtarget->hasFMA4() || Subtarget->hasAVX512()))
19843 VT = VT.getScalarType();
19845 if (!VT.isSimple())
19848 switch (VT.getSimpleVT().SimpleTy) {
19859 bool X86TargetLowering::isNarrowingProfitable(EVT VT1, EVT VT2) const {
19860 // i16 instructions are longer (0x66 prefix) and potentially slower.
19861 return !(VT1 == MVT::i32 && VT2 == MVT::i16);
19864 /// isShuffleMaskLegal - Targets can use this to indicate that they only
19865 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
19866 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
19867 /// are assumed to be legal.
19869 X86TargetLowering::isShuffleMaskLegal(const SmallVectorImpl<int> &M,
19871 if (!VT.isSimple())
19874 // Not for i1 vectors
19875 if (VT.getScalarType() == MVT::i1)
19878 // Very little shuffling can be done for 64-bit vectors right now.
19879 if (VT.getSizeInBits() == 64)
19882 // We only care that the types being shuffled are legal. The lowering can
19883 // handle any possible shuffle mask that results.
19884 return isTypeLegal(VT.getSimpleVT());
19888 X86TargetLowering::isVectorClearMaskLegal(const SmallVectorImpl<int> &Mask,
19890 // Just delegate to the generic legality, clear masks aren't special.
19891 return isShuffleMaskLegal(Mask, VT);
19894 //===----------------------------------------------------------------------===//
19895 // X86 Scheduler Hooks
19896 //===----------------------------------------------------------------------===//
19898 /// Utility function to emit xbegin specifying the start of an RTM region.
19899 static MachineBasicBlock *EmitXBegin(MachineInstr *MI, MachineBasicBlock *MBB,
19900 const TargetInstrInfo *TII) {
19901 DebugLoc DL = MI->getDebugLoc();
19903 const BasicBlock *BB = MBB->getBasicBlock();
19904 MachineFunction::iterator I = MBB;
19907 // For the v = xbegin(), we generate
19918 MachineBasicBlock *thisMBB = MBB;
19919 MachineFunction *MF = MBB->getParent();
19920 MachineBasicBlock *mainMBB = MF->CreateMachineBasicBlock(BB);
19921 MachineBasicBlock *sinkMBB = MF->CreateMachineBasicBlock(BB);
19922 MF->insert(I, mainMBB);
19923 MF->insert(I, sinkMBB);
19925 // Transfer the remainder of BB and its successor edges to sinkMBB.
19926 sinkMBB->splice(sinkMBB->begin(), MBB,
19927 std::next(MachineBasicBlock::iterator(MI)), MBB->end());
19928 sinkMBB->transferSuccessorsAndUpdatePHIs(MBB);
19932 // # fallthrough to mainMBB
19933 // # abortion to sinkMBB
19934 BuildMI(thisMBB, DL, TII->get(X86::XBEGIN_4)).addMBB(sinkMBB);
19935 thisMBB->addSuccessor(mainMBB);
19936 thisMBB->addSuccessor(sinkMBB);
19940 BuildMI(mainMBB, DL, TII->get(X86::MOV32ri), X86::EAX).addImm(-1);
19941 mainMBB->addSuccessor(sinkMBB);
19944 // EAX is live into the sinkMBB
19945 sinkMBB->addLiveIn(X86::EAX);
19946 BuildMI(*sinkMBB, sinkMBB->begin(), DL,
19947 TII->get(TargetOpcode::COPY), MI->getOperand(0).getReg())
19950 MI->eraseFromParent();
19954 // FIXME: When we get size specific XMM0 registers, i.e. XMM0_V16I8
19955 // or XMM0_V32I8 in AVX all of this code can be replaced with that
19956 // in the .td file.
19957 static MachineBasicBlock *EmitPCMPSTRM(MachineInstr *MI, MachineBasicBlock *BB,
19958 const TargetInstrInfo *TII) {
19960 switch (MI->getOpcode()) {
19961 default: llvm_unreachable("illegal opcode!");
19962 case X86::PCMPISTRM128REG: Opc = X86::PCMPISTRM128rr; break;
19963 case X86::VPCMPISTRM128REG: Opc = X86::VPCMPISTRM128rr; break;
19964 case X86::PCMPISTRM128MEM: Opc = X86::PCMPISTRM128rm; break;
19965 case X86::VPCMPISTRM128MEM: Opc = X86::VPCMPISTRM128rm; break;
19966 case X86::PCMPESTRM128REG: Opc = X86::PCMPESTRM128rr; break;
19967 case X86::VPCMPESTRM128REG: Opc = X86::VPCMPESTRM128rr; break;
19968 case X86::PCMPESTRM128MEM: Opc = X86::PCMPESTRM128rm; break;
19969 case X86::VPCMPESTRM128MEM: Opc = X86::VPCMPESTRM128rm; break;
19972 DebugLoc dl = MI->getDebugLoc();
19973 MachineInstrBuilder MIB = BuildMI(*BB, MI, dl, TII->get(Opc));
19975 unsigned NumArgs = MI->getNumOperands();
19976 for (unsigned i = 1; i < NumArgs; ++i) {
19977 MachineOperand &Op = MI->getOperand(i);
19978 if (!(Op.isReg() && Op.isImplicit()))
19979 MIB.addOperand(Op);
19981 if (MI->hasOneMemOperand())
19982 MIB->setMemRefs(MI->memoperands_begin(), MI->memoperands_end());
19984 BuildMI(*BB, MI, dl,
19985 TII->get(TargetOpcode::COPY), MI->getOperand(0).getReg())
19986 .addReg(X86::XMM0);
19988 MI->eraseFromParent();
19992 // FIXME: Custom handling because TableGen doesn't support multiple implicit
19993 // defs in an instruction pattern
19994 static MachineBasicBlock *EmitPCMPSTRI(MachineInstr *MI, MachineBasicBlock *BB,
19995 const TargetInstrInfo *TII) {
19997 switch (MI->getOpcode()) {
19998 default: llvm_unreachable("illegal opcode!");
19999 case X86::PCMPISTRIREG: Opc = X86::PCMPISTRIrr; break;
20000 case X86::VPCMPISTRIREG: Opc = X86::VPCMPISTRIrr; break;
20001 case X86::PCMPISTRIMEM: Opc = X86::PCMPISTRIrm; break;
20002 case X86::VPCMPISTRIMEM: Opc = X86::VPCMPISTRIrm; break;
20003 case X86::PCMPESTRIREG: Opc = X86::PCMPESTRIrr; break;
20004 case X86::VPCMPESTRIREG: Opc = X86::VPCMPESTRIrr; break;
20005 case X86::PCMPESTRIMEM: Opc = X86::PCMPESTRIrm; break;
20006 case X86::VPCMPESTRIMEM: Opc = X86::VPCMPESTRIrm; break;
20009 DebugLoc dl = MI->getDebugLoc();
20010 MachineInstrBuilder MIB = BuildMI(*BB, MI, dl, TII->get(Opc));
20012 unsigned NumArgs = MI->getNumOperands(); // remove the results
20013 for (unsigned i = 1; i < NumArgs; ++i) {
20014 MachineOperand &Op = MI->getOperand(i);
20015 if (!(Op.isReg() && Op.isImplicit()))
20016 MIB.addOperand(Op);
20018 if (MI->hasOneMemOperand())
20019 MIB->setMemRefs(MI->memoperands_begin(), MI->memoperands_end());
20021 BuildMI(*BB, MI, dl,
20022 TII->get(TargetOpcode::COPY), MI->getOperand(0).getReg())
20025 MI->eraseFromParent();
20029 static MachineBasicBlock *EmitMonitor(MachineInstr *MI, MachineBasicBlock *BB,
20030 const X86Subtarget *Subtarget) {
20031 DebugLoc dl = MI->getDebugLoc();
20032 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
20033 // Address into RAX/EAX, other two args into ECX, EDX.
20034 unsigned MemOpc = Subtarget->is64Bit() ? X86::LEA64r : X86::LEA32r;
20035 unsigned MemReg = Subtarget->is64Bit() ? X86::RAX : X86::EAX;
20036 MachineInstrBuilder MIB = BuildMI(*BB, MI, dl, TII->get(MemOpc), MemReg);
20037 for (int i = 0; i < X86::AddrNumOperands; ++i)
20038 MIB.addOperand(MI->getOperand(i));
20040 unsigned ValOps = X86::AddrNumOperands;
20041 BuildMI(*BB, MI, dl, TII->get(TargetOpcode::COPY), X86::ECX)
20042 .addReg(MI->getOperand(ValOps).getReg());
20043 BuildMI(*BB, MI, dl, TII->get(TargetOpcode::COPY), X86::EDX)
20044 .addReg(MI->getOperand(ValOps+1).getReg());
20046 // The instruction doesn't actually take any operands though.
20047 BuildMI(*BB, MI, dl, TII->get(X86::MONITORrrr));
20049 MI->eraseFromParent(); // The pseudo is gone now.
20053 MachineBasicBlock *
20054 X86TargetLowering::EmitVAARG64WithCustomInserter(MachineInstr *MI,
20055 MachineBasicBlock *MBB) const {
20056 // Emit va_arg instruction on X86-64.
20058 // Operands to this pseudo-instruction:
20059 // 0 ) Output : destination address (reg)
20060 // 1-5) Input : va_list address (addr, i64mem)
20061 // 6 ) ArgSize : Size (in bytes) of vararg type
20062 // 7 ) ArgMode : 0=overflow only, 1=use gp_offset, 2=use fp_offset
20063 // 8 ) Align : Alignment of type
20064 // 9 ) EFLAGS (implicit-def)
20066 assert(MI->getNumOperands() == 10 && "VAARG_64 should have 10 operands!");
20067 static_assert(X86::AddrNumOperands == 5,
20068 "VAARG_64 assumes 5 address operands");
20070 unsigned DestReg = MI->getOperand(0).getReg();
20071 MachineOperand &Base = MI->getOperand(1);
20072 MachineOperand &Scale = MI->getOperand(2);
20073 MachineOperand &Index = MI->getOperand(3);
20074 MachineOperand &Disp = MI->getOperand(4);
20075 MachineOperand &Segment = MI->getOperand(5);
20076 unsigned ArgSize = MI->getOperand(6).getImm();
20077 unsigned ArgMode = MI->getOperand(7).getImm();
20078 unsigned Align = MI->getOperand(8).getImm();
20080 // Memory Reference
20081 assert(MI->hasOneMemOperand() && "Expected VAARG_64 to have one memoperand");
20082 MachineInstr::mmo_iterator MMOBegin = MI->memoperands_begin();
20083 MachineInstr::mmo_iterator MMOEnd = MI->memoperands_end();
20085 // Machine Information
20086 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
20087 MachineRegisterInfo &MRI = MBB->getParent()->getRegInfo();
20088 const TargetRegisterClass *AddrRegClass = getRegClassFor(MVT::i64);
20089 const TargetRegisterClass *OffsetRegClass = getRegClassFor(MVT::i32);
20090 DebugLoc DL = MI->getDebugLoc();
20092 // struct va_list {
20095 // i64 overflow_area (address)
20096 // i64 reg_save_area (address)
20098 // sizeof(va_list) = 24
20099 // alignment(va_list) = 8
20101 unsigned TotalNumIntRegs = 6;
20102 unsigned TotalNumXMMRegs = 8;
20103 bool UseGPOffset = (ArgMode == 1);
20104 bool UseFPOffset = (ArgMode == 2);
20105 unsigned MaxOffset = TotalNumIntRegs * 8 +
20106 (UseFPOffset ? TotalNumXMMRegs * 16 : 0);
20108 /* Align ArgSize to a multiple of 8 */
20109 unsigned ArgSizeA8 = (ArgSize + 7) & ~7;
20110 bool NeedsAlign = (Align > 8);
20112 MachineBasicBlock *thisMBB = MBB;
20113 MachineBasicBlock *overflowMBB;
20114 MachineBasicBlock *offsetMBB;
20115 MachineBasicBlock *endMBB;
20117 unsigned OffsetDestReg = 0; // Argument address computed by offsetMBB
20118 unsigned OverflowDestReg = 0; // Argument address computed by overflowMBB
20119 unsigned OffsetReg = 0;
20121 if (!UseGPOffset && !UseFPOffset) {
20122 // If we only pull from the overflow region, we don't create a branch.
20123 // We don't need to alter control flow.
20124 OffsetDestReg = 0; // unused
20125 OverflowDestReg = DestReg;
20127 offsetMBB = nullptr;
20128 overflowMBB = thisMBB;
20131 // First emit code to check if gp_offset (or fp_offset) is below the bound.
20132 // If so, pull the argument from reg_save_area. (branch to offsetMBB)
20133 // If not, pull from overflow_area. (branch to overflowMBB)
20138 // offsetMBB overflowMBB
20143 // Registers for the PHI in endMBB
20144 OffsetDestReg = MRI.createVirtualRegister(AddrRegClass);
20145 OverflowDestReg = MRI.createVirtualRegister(AddrRegClass);
20147 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
20148 MachineFunction *MF = MBB->getParent();
20149 overflowMBB = MF->CreateMachineBasicBlock(LLVM_BB);
20150 offsetMBB = MF->CreateMachineBasicBlock(LLVM_BB);
20151 endMBB = MF->CreateMachineBasicBlock(LLVM_BB);
20153 MachineFunction::iterator MBBIter = MBB;
20156 // Insert the new basic blocks
20157 MF->insert(MBBIter, offsetMBB);
20158 MF->insert(MBBIter, overflowMBB);
20159 MF->insert(MBBIter, endMBB);
20161 // Transfer the remainder of MBB and its successor edges to endMBB.
20162 endMBB->splice(endMBB->begin(), thisMBB,
20163 std::next(MachineBasicBlock::iterator(MI)), thisMBB->end());
20164 endMBB->transferSuccessorsAndUpdatePHIs(thisMBB);
20166 // Make offsetMBB and overflowMBB successors of thisMBB
20167 thisMBB->addSuccessor(offsetMBB);
20168 thisMBB->addSuccessor(overflowMBB);
20170 // endMBB is a successor of both offsetMBB and overflowMBB
20171 offsetMBB->addSuccessor(endMBB);
20172 overflowMBB->addSuccessor(endMBB);
20174 // Load the offset value into a register
20175 OffsetReg = MRI.createVirtualRegister(OffsetRegClass);
20176 BuildMI(thisMBB, DL, TII->get(X86::MOV32rm), OffsetReg)
20180 .addDisp(Disp, UseFPOffset ? 4 : 0)
20181 .addOperand(Segment)
20182 .setMemRefs(MMOBegin, MMOEnd);
20184 // Check if there is enough room left to pull this argument.
20185 BuildMI(thisMBB, DL, TII->get(X86::CMP32ri))
20187 .addImm(MaxOffset + 8 - ArgSizeA8);
20189 // Branch to "overflowMBB" if offset >= max
20190 // Fall through to "offsetMBB" otherwise
20191 BuildMI(thisMBB, DL, TII->get(X86::GetCondBranchFromCond(X86::COND_AE)))
20192 .addMBB(overflowMBB);
20195 // In offsetMBB, emit code to use the reg_save_area.
20197 assert(OffsetReg != 0);
20199 // Read the reg_save_area address.
20200 unsigned RegSaveReg = MRI.createVirtualRegister(AddrRegClass);
20201 BuildMI(offsetMBB, DL, TII->get(X86::MOV64rm), RegSaveReg)
20206 .addOperand(Segment)
20207 .setMemRefs(MMOBegin, MMOEnd);
20209 // Zero-extend the offset
20210 unsigned OffsetReg64 = MRI.createVirtualRegister(AddrRegClass);
20211 BuildMI(offsetMBB, DL, TII->get(X86::SUBREG_TO_REG), OffsetReg64)
20214 .addImm(X86::sub_32bit);
20216 // Add the offset to the reg_save_area to get the final address.
20217 BuildMI(offsetMBB, DL, TII->get(X86::ADD64rr), OffsetDestReg)
20218 .addReg(OffsetReg64)
20219 .addReg(RegSaveReg);
20221 // Compute the offset for the next argument
20222 unsigned NextOffsetReg = MRI.createVirtualRegister(OffsetRegClass);
20223 BuildMI(offsetMBB, DL, TII->get(X86::ADD32ri), NextOffsetReg)
20225 .addImm(UseFPOffset ? 16 : 8);
20227 // Store it back into the va_list.
20228 BuildMI(offsetMBB, DL, TII->get(X86::MOV32mr))
20232 .addDisp(Disp, UseFPOffset ? 4 : 0)
20233 .addOperand(Segment)
20234 .addReg(NextOffsetReg)
20235 .setMemRefs(MMOBegin, MMOEnd);
20238 BuildMI(offsetMBB, DL, TII->get(X86::JMP_1))
20243 // Emit code to use overflow area
20246 // Load the overflow_area address into a register.
20247 unsigned OverflowAddrReg = MRI.createVirtualRegister(AddrRegClass);
20248 BuildMI(overflowMBB, DL, TII->get(X86::MOV64rm), OverflowAddrReg)
20253 .addOperand(Segment)
20254 .setMemRefs(MMOBegin, MMOEnd);
20256 // If we need to align it, do so. Otherwise, just copy the address
20257 // to OverflowDestReg.
20259 // Align the overflow address
20260 assert((Align & (Align-1)) == 0 && "Alignment must be a power of 2");
20261 unsigned TmpReg = MRI.createVirtualRegister(AddrRegClass);
20263 // aligned_addr = (addr + (align-1)) & ~(align-1)
20264 BuildMI(overflowMBB, DL, TII->get(X86::ADD64ri32), TmpReg)
20265 .addReg(OverflowAddrReg)
20268 BuildMI(overflowMBB, DL, TII->get(X86::AND64ri32), OverflowDestReg)
20270 .addImm(~(uint64_t)(Align-1));
20272 BuildMI(overflowMBB, DL, TII->get(TargetOpcode::COPY), OverflowDestReg)
20273 .addReg(OverflowAddrReg);
20276 // Compute the next overflow address after this argument.
20277 // (the overflow address should be kept 8-byte aligned)
20278 unsigned NextAddrReg = MRI.createVirtualRegister(AddrRegClass);
20279 BuildMI(overflowMBB, DL, TII->get(X86::ADD64ri32), NextAddrReg)
20280 .addReg(OverflowDestReg)
20281 .addImm(ArgSizeA8);
20283 // Store the new overflow address.
20284 BuildMI(overflowMBB, DL, TII->get(X86::MOV64mr))
20289 .addOperand(Segment)
20290 .addReg(NextAddrReg)
20291 .setMemRefs(MMOBegin, MMOEnd);
20293 // If we branched, emit the PHI to the front of endMBB.
20295 BuildMI(*endMBB, endMBB->begin(), DL,
20296 TII->get(X86::PHI), DestReg)
20297 .addReg(OffsetDestReg).addMBB(offsetMBB)
20298 .addReg(OverflowDestReg).addMBB(overflowMBB);
20301 // Erase the pseudo instruction
20302 MI->eraseFromParent();
20307 MachineBasicBlock *
20308 X86TargetLowering::EmitVAStartSaveXMMRegsWithCustomInserter(
20310 MachineBasicBlock *MBB) const {
20311 // Emit code to save XMM registers to the stack. The ABI says that the
20312 // number of registers to save is given in %al, so it's theoretically
20313 // possible to do an indirect jump trick to avoid saving all of them,
20314 // however this code takes a simpler approach and just executes all
20315 // of the stores if %al is non-zero. It's less code, and it's probably
20316 // easier on the hardware branch predictor, and stores aren't all that
20317 // expensive anyway.
20319 // Create the new basic blocks. One block contains all the XMM stores,
20320 // and one block is the final destination regardless of whether any
20321 // stores were performed.
20322 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
20323 MachineFunction *F = MBB->getParent();
20324 MachineFunction::iterator MBBIter = MBB;
20326 MachineBasicBlock *XMMSaveMBB = F->CreateMachineBasicBlock(LLVM_BB);
20327 MachineBasicBlock *EndMBB = F->CreateMachineBasicBlock(LLVM_BB);
20328 F->insert(MBBIter, XMMSaveMBB);
20329 F->insert(MBBIter, EndMBB);
20331 // Transfer the remainder of MBB and its successor edges to EndMBB.
20332 EndMBB->splice(EndMBB->begin(), MBB,
20333 std::next(MachineBasicBlock::iterator(MI)), MBB->end());
20334 EndMBB->transferSuccessorsAndUpdatePHIs(MBB);
20336 // The original block will now fall through to the XMM save block.
20337 MBB->addSuccessor(XMMSaveMBB);
20338 // The XMMSaveMBB will fall through to the end block.
20339 XMMSaveMBB->addSuccessor(EndMBB);
20341 // Now add the instructions.
20342 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
20343 DebugLoc DL = MI->getDebugLoc();
20345 unsigned CountReg = MI->getOperand(0).getReg();
20346 int64_t RegSaveFrameIndex = MI->getOperand(1).getImm();
20347 int64_t VarArgsFPOffset = MI->getOperand(2).getImm();
20349 if (!Subtarget->isCallingConvWin64(F->getFunction()->getCallingConv())) {
20350 // If %al is 0, branch around the XMM save block.
20351 BuildMI(MBB, DL, TII->get(X86::TEST8rr)).addReg(CountReg).addReg(CountReg);
20352 BuildMI(MBB, DL, TII->get(X86::JE_1)).addMBB(EndMBB);
20353 MBB->addSuccessor(EndMBB);
20356 // Make sure the last operand is EFLAGS, which gets clobbered by the branch
20357 // that was just emitted, but clearly shouldn't be "saved".
20358 assert((MI->getNumOperands() <= 3 ||
20359 !MI->getOperand(MI->getNumOperands() - 1).isReg() ||
20360 MI->getOperand(MI->getNumOperands() - 1).getReg() == X86::EFLAGS)
20361 && "Expected last argument to be EFLAGS");
20362 unsigned MOVOpc = Subtarget->hasFp256() ? X86::VMOVAPSmr : X86::MOVAPSmr;
20363 // In the XMM save block, save all the XMM argument registers.
20364 for (int i = 3, e = MI->getNumOperands() - 1; i != e; ++i) {
20365 int64_t Offset = (i - 3) * 16 + VarArgsFPOffset;
20366 MachineMemOperand *MMO = F->getMachineMemOperand(
20367 MachinePointerInfo::getFixedStack(*F, RegSaveFrameIndex, Offset),
20368 MachineMemOperand::MOStore,
20369 /*Size=*/16, /*Align=*/16);
20370 BuildMI(XMMSaveMBB, DL, TII->get(MOVOpc))
20371 .addFrameIndex(RegSaveFrameIndex)
20372 .addImm(/*Scale=*/1)
20373 .addReg(/*IndexReg=*/0)
20374 .addImm(/*Disp=*/Offset)
20375 .addReg(/*Segment=*/0)
20376 .addReg(MI->getOperand(i).getReg())
20377 .addMemOperand(MMO);
20380 MI->eraseFromParent(); // The pseudo instruction is gone now.
20385 // The EFLAGS operand of SelectItr might be missing a kill marker
20386 // because there were multiple uses of EFLAGS, and ISel didn't know
20387 // which to mark. Figure out whether SelectItr should have had a
20388 // kill marker, and set it if it should. Returns the correct kill
20390 static bool checkAndUpdateEFLAGSKill(MachineBasicBlock::iterator SelectItr,
20391 MachineBasicBlock* BB,
20392 const TargetRegisterInfo* TRI) {
20393 // Scan forward through BB for a use/def of EFLAGS.
20394 MachineBasicBlock::iterator miI(std::next(SelectItr));
20395 for (MachineBasicBlock::iterator miE = BB->end(); miI != miE; ++miI) {
20396 const MachineInstr& mi = *miI;
20397 if (mi.readsRegister(X86::EFLAGS))
20399 if (mi.definesRegister(X86::EFLAGS))
20400 break; // Should have kill-flag - update below.
20403 // If we hit the end of the block, check whether EFLAGS is live into a
20405 if (miI == BB->end()) {
20406 for (MachineBasicBlock::succ_iterator sItr = BB->succ_begin(),
20407 sEnd = BB->succ_end();
20408 sItr != sEnd; ++sItr) {
20409 MachineBasicBlock* succ = *sItr;
20410 if (succ->isLiveIn(X86::EFLAGS))
20415 // We found a def, or hit the end of the basic block and EFLAGS wasn't live
20416 // out. SelectMI should have a kill flag on EFLAGS.
20417 SelectItr->addRegisterKilled(X86::EFLAGS, TRI);
20421 // Return true if it is OK for this CMOV pseudo-opcode to be cascaded
20422 // together with other CMOV pseudo-opcodes into a single basic-block with
20423 // conditional jump around it.
20424 static bool isCMOVPseudo(MachineInstr *MI) {
20425 switch (MI->getOpcode()) {
20426 case X86::CMOV_FR32:
20427 case X86::CMOV_FR64:
20428 case X86::CMOV_GR8:
20429 case X86::CMOV_GR16:
20430 case X86::CMOV_GR32:
20431 case X86::CMOV_RFP32:
20432 case X86::CMOV_RFP64:
20433 case X86::CMOV_RFP80:
20434 case X86::CMOV_V2F64:
20435 case X86::CMOV_V2I64:
20436 case X86::CMOV_V4F32:
20437 case X86::CMOV_V4F64:
20438 case X86::CMOV_V4I64:
20439 case X86::CMOV_V16F32:
20440 case X86::CMOV_V8F32:
20441 case X86::CMOV_V8F64:
20442 case X86::CMOV_V8I64:
20443 case X86::CMOV_V8I1:
20444 case X86::CMOV_V16I1:
20445 case X86::CMOV_V32I1:
20446 case X86::CMOV_V64I1:
20454 MachineBasicBlock *
20455 X86TargetLowering::EmitLoweredSelect(MachineInstr *MI,
20456 MachineBasicBlock *BB) const {
20457 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
20458 DebugLoc DL = MI->getDebugLoc();
20460 // To "insert" a SELECT_CC instruction, we actually have to insert the
20461 // diamond control-flow pattern. The incoming instruction knows the
20462 // destination vreg to set, the condition code register to branch on, the
20463 // true/false values to select between, and a branch opcode to use.
20464 const BasicBlock *LLVM_BB = BB->getBasicBlock();
20465 MachineFunction::iterator It = BB;
20471 // cmpTY ccX, r1, r2
20473 // fallthrough --> copy0MBB
20474 MachineBasicBlock *thisMBB = BB;
20475 MachineFunction *F = BB->getParent();
20477 // This code lowers all pseudo-CMOV instructions. Generally it lowers these
20478 // as described above, by inserting a BB, and then making a PHI at the join
20479 // point to select the true and false operands of the CMOV in the PHI.
20481 // The code also handles two different cases of multiple CMOV opcodes
20485 // In this case, there are multiple CMOVs in a row, all which are based on
20486 // the same condition setting (or the exact opposite condition setting).
20487 // In this case we can lower all the CMOVs using a single inserted BB, and
20488 // then make a number of PHIs at the join point to model the CMOVs. The only
20489 // trickiness here, is that in a case like:
20491 // t2 = CMOV cond1 t1, f1
20492 // t3 = CMOV cond1 t2, f2
20494 // when rewriting this into PHIs, we have to perform some renaming on the
20495 // temps since you cannot have a PHI operand refer to a PHI result earlier
20496 // in the same block. The "simple" but wrong lowering would be:
20498 // t2 = PHI t1(BB1), f1(BB2)
20499 // t3 = PHI t2(BB1), f2(BB2)
20501 // but clearly t2 is not defined in BB1, so that is incorrect. The proper
20502 // renaming is to note that on the path through BB1, t2 is really just a
20503 // copy of t1, and do that renaming, properly generating:
20505 // t2 = PHI t1(BB1), f1(BB2)
20506 // t3 = PHI t1(BB1), f2(BB2)
20508 // Case 2, we lower cascaded CMOVs such as
20510 // (CMOV (CMOV F, T, cc1), T, cc2)
20512 // to two successives branches. For that, we look for another CMOV as the
20513 // following instruction.
20515 // Without this, we would add a PHI between the two jumps, which ends up
20516 // creating a few copies all around. For instance, for
20518 // (sitofp (zext (fcmp une)))
20520 // we would generate:
20522 // ucomiss %xmm1, %xmm0
20523 // movss <1.0f>, %xmm0
20524 // movaps %xmm0, %xmm1
20526 // xorps %xmm1, %xmm1
20529 // movaps %xmm1, %xmm0
20533 // because this custom-inserter would have generated:
20545 // A: X = ...; Y = ...
20547 // C: Z = PHI [X, A], [Y, B]
20549 // E: PHI [X, C], [Z, D]
20551 // If we lower both CMOVs in a single step, we can instead generate:
20563 // A: X = ...; Y = ...
20565 // E: PHI [X, A], [X, C], [Y, D]
20567 // Which, in our sitofp/fcmp example, gives us something like:
20569 // ucomiss %xmm1, %xmm0
20570 // movss <1.0f>, %xmm0
20573 // xorps %xmm0, %xmm0
20577 MachineInstr *CascadedCMOV = nullptr;
20578 MachineInstr *LastCMOV = MI;
20579 X86::CondCode CC = X86::CondCode(MI->getOperand(3).getImm());
20580 X86::CondCode OppCC = X86::GetOppositeBranchCondition(CC);
20581 MachineBasicBlock::iterator NextMIIt =
20582 std::next(MachineBasicBlock::iterator(MI));
20584 // Check for case 1, where there are multiple CMOVs with the same condition
20585 // first. Of the two cases of multiple CMOV lowerings, case 1 reduces the
20586 // number of jumps the most.
20588 if (isCMOVPseudo(MI)) {
20589 // See if we have a string of CMOVS with the same condition.
20590 while (NextMIIt != BB->end() &&
20591 isCMOVPseudo(NextMIIt) &&
20592 (NextMIIt->getOperand(3).getImm() == CC ||
20593 NextMIIt->getOperand(3).getImm() == OppCC)) {
20594 LastCMOV = &*NextMIIt;
20599 // This checks for case 2, but only do this if we didn't already find
20600 // case 1, as indicated by LastCMOV == MI.
20601 if (LastCMOV == MI &&
20602 NextMIIt != BB->end() && NextMIIt->getOpcode() == MI->getOpcode() &&
20603 NextMIIt->getOperand(2).getReg() == MI->getOperand(2).getReg() &&
20604 NextMIIt->getOperand(1).getReg() == MI->getOperand(0).getReg()) {
20605 CascadedCMOV = &*NextMIIt;
20608 MachineBasicBlock *jcc1MBB = nullptr;
20610 // If we have a cascaded CMOV, we lower it to two successive branches to
20611 // the same block. EFLAGS is used by both, so mark it as live in the second.
20612 if (CascadedCMOV) {
20613 jcc1MBB = F->CreateMachineBasicBlock(LLVM_BB);
20614 F->insert(It, jcc1MBB);
20615 jcc1MBB->addLiveIn(X86::EFLAGS);
20618 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
20619 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
20620 F->insert(It, copy0MBB);
20621 F->insert(It, sinkMBB);
20623 // If the EFLAGS register isn't dead in the terminator, then claim that it's
20624 // live into the sink and copy blocks.
20625 const TargetRegisterInfo *TRI = Subtarget->getRegisterInfo();
20627 MachineInstr *LastEFLAGSUser = CascadedCMOV ? CascadedCMOV : LastCMOV;
20628 if (!LastEFLAGSUser->killsRegister(X86::EFLAGS) &&
20629 !checkAndUpdateEFLAGSKill(LastEFLAGSUser, BB, TRI)) {
20630 copy0MBB->addLiveIn(X86::EFLAGS);
20631 sinkMBB->addLiveIn(X86::EFLAGS);
20634 // Transfer the remainder of BB and its successor edges to sinkMBB.
20635 sinkMBB->splice(sinkMBB->begin(), BB,
20636 std::next(MachineBasicBlock::iterator(LastCMOV)), BB->end());
20637 sinkMBB->transferSuccessorsAndUpdatePHIs(BB);
20639 // Add the true and fallthrough blocks as its successors.
20640 if (CascadedCMOV) {
20641 // The fallthrough block may be jcc1MBB, if we have a cascaded CMOV.
20642 BB->addSuccessor(jcc1MBB);
20644 // In that case, jcc1MBB will itself fallthrough the copy0MBB, and
20645 // jump to the sinkMBB.
20646 jcc1MBB->addSuccessor(copy0MBB);
20647 jcc1MBB->addSuccessor(sinkMBB);
20649 BB->addSuccessor(copy0MBB);
20652 // The true block target of the first (or only) branch is always sinkMBB.
20653 BB->addSuccessor(sinkMBB);
20655 // Create the conditional branch instruction.
20656 unsigned Opc = X86::GetCondBranchFromCond(CC);
20657 BuildMI(BB, DL, TII->get(Opc)).addMBB(sinkMBB);
20659 if (CascadedCMOV) {
20660 unsigned Opc2 = X86::GetCondBranchFromCond(
20661 (X86::CondCode)CascadedCMOV->getOperand(3).getImm());
20662 BuildMI(jcc1MBB, DL, TII->get(Opc2)).addMBB(sinkMBB);
20666 // %FalseValue = ...
20667 // # fallthrough to sinkMBB
20668 copy0MBB->addSuccessor(sinkMBB);
20671 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
20673 MachineBasicBlock::iterator MIItBegin = MachineBasicBlock::iterator(MI);
20674 MachineBasicBlock::iterator MIItEnd =
20675 std::next(MachineBasicBlock::iterator(LastCMOV));
20676 MachineBasicBlock::iterator SinkInsertionPoint = sinkMBB->begin();
20677 DenseMap<unsigned, std::pair<unsigned, unsigned>> RegRewriteTable;
20678 MachineInstrBuilder MIB;
20680 // As we are creating the PHIs, we have to be careful if there is more than
20681 // one. Later CMOVs may reference the results of earlier CMOVs, but later
20682 // PHIs have to reference the individual true/false inputs from earlier PHIs.
20683 // That also means that PHI construction must work forward from earlier to
20684 // later, and that the code must maintain a mapping from earlier PHI's
20685 // destination registers, and the registers that went into the PHI.
20687 for (MachineBasicBlock::iterator MIIt = MIItBegin; MIIt != MIItEnd; ++MIIt) {
20688 unsigned DestReg = MIIt->getOperand(0).getReg();
20689 unsigned Op1Reg = MIIt->getOperand(1).getReg();
20690 unsigned Op2Reg = MIIt->getOperand(2).getReg();
20692 // If this CMOV we are generating is the opposite condition from
20693 // the jump we generated, then we have to swap the operands for the
20694 // PHI that is going to be generated.
20695 if (MIIt->getOperand(3).getImm() == OppCC)
20696 std::swap(Op1Reg, Op2Reg);
20698 if (RegRewriteTable.find(Op1Reg) != RegRewriteTable.end())
20699 Op1Reg = RegRewriteTable[Op1Reg].first;
20701 if (RegRewriteTable.find(Op2Reg) != RegRewriteTable.end())
20702 Op2Reg = RegRewriteTable[Op2Reg].second;
20704 MIB = BuildMI(*sinkMBB, SinkInsertionPoint, DL,
20705 TII->get(X86::PHI), DestReg)
20706 .addReg(Op1Reg).addMBB(copy0MBB)
20707 .addReg(Op2Reg).addMBB(thisMBB);
20709 // Add this PHI to the rewrite table.
20710 RegRewriteTable[DestReg] = std::make_pair(Op1Reg, Op2Reg);
20713 // If we have a cascaded CMOV, the second Jcc provides the same incoming
20714 // value as the first Jcc (the True operand of the SELECT_CC/CMOV nodes).
20715 if (CascadedCMOV) {
20716 MIB.addReg(MI->getOperand(2).getReg()).addMBB(jcc1MBB);
20717 // Copy the PHI result to the register defined by the second CMOV.
20718 BuildMI(*sinkMBB, std::next(MachineBasicBlock::iterator(MIB.getInstr())),
20719 DL, TII->get(TargetOpcode::COPY),
20720 CascadedCMOV->getOperand(0).getReg())
20721 .addReg(MI->getOperand(0).getReg());
20722 CascadedCMOV->eraseFromParent();
20725 // Now remove the CMOV(s).
20726 for (MachineBasicBlock::iterator MIIt = MIItBegin; MIIt != MIItEnd; )
20727 (MIIt++)->eraseFromParent();
20732 MachineBasicBlock *
20733 X86TargetLowering::EmitLoweredAtomicFP(MachineInstr *MI,
20734 MachineBasicBlock *BB) const {
20735 // Combine the following atomic floating-point modification pattern:
20736 // a.store(reg OP a.load(acquire), release)
20737 // Transform them into:
20738 // OPss (%gpr), %xmm
20739 // movss %xmm, (%gpr)
20740 // Or sd equivalent for 64-bit operations.
20742 switch (MI->getOpcode()) {
20743 default: llvm_unreachable("unexpected instr type for EmitLoweredAtomicFP");
20744 case X86::RELEASE_FADD32mr: MOp = X86::MOVSSmr; FOp = X86::ADDSSrm; break;
20745 case X86::RELEASE_FADD64mr: MOp = X86::MOVSDmr; FOp = X86::ADDSDrm; break;
20747 const X86InstrInfo *TII = Subtarget->getInstrInfo();
20748 DebugLoc DL = MI->getDebugLoc();
20749 MachineRegisterInfo &MRI = BB->getParent()->getRegInfo();
20750 unsigned MSrc = MI->getOperand(0).getReg();
20751 unsigned VSrc = MI->getOperand(5).getReg();
20752 MachineInstrBuilder MIM = BuildMI(*BB, MI, DL, TII->get(MOp))
20753 .addReg(/*Base=*/MSrc)
20754 .addImm(/*Scale=*/1)
20755 .addReg(/*Index=*/0)
20758 MachineInstr *MIO = BuildMI(*BB, (MachineInstr *)MIM, DL, TII->get(FOp),
20759 MRI.createVirtualRegister(MRI.getRegClass(VSrc)))
20761 .addReg(/*Base=*/MSrc)
20762 .addImm(/*Scale=*/1)
20763 .addReg(/*Index=*/0)
20764 .addImm(/*Disp=*/0)
20765 .addReg(/*Segment=*/0);
20766 MIM.addReg(MIO->getOperand(0).getReg(), RegState::Kill);
20767 MI->eraseFromParent(); // The pseudo instruction is gone now.
20771 MachineBasicBlock *
20772 X86TargetLowering::EmitLoweredSegAlloca(MachineInstr *MI,
20773 MachineBasicBlock *BB) const {
20774 MachineFunction *MF = BB->getParent();
20775 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
20776 DebugLoc DL = MI->getDebugLoc();
20777 const BasicBlock *LLVM_BB = BB->getBasicBlock();
20779 assert(MF->shouldSplitStack());
20781 const bool Is64Bit = Subtarget->is64Bit();
20782 const bool IsLP64 = Subtarget->isTarget64BitLP64();
20784 const unsigned TlsReg = Is64Bit ? X86::FS : X86::GS;
20785 const unsigned TlsOffset = IsLP64 ? 0x70 : Is64Bit ? 0x40 : 0x30;
20788 // ... [Till the alloca]
20789 // If stacklet is not large enough, jump to mallocMBB
20792 // Allocate by subtracting from RSP
20793 // Jump to continueMBB
20796 // Allocate by call to runtime
20800 // [rest of original BB]
20803 MachineBasicBlock *mallocMBB = MF->CreateMachineBasicBlock(LLVM_BB);
20804 MachineBasicBlock *bumpMBB = MF->CreateMachineBasicBlock(LLVM_BB);
20805 MachineBasicBlock *continueMBB = MF->CreateMachineBasicBlock(LLVM_BB);
20807 MachineRegisterInfo &MRI = MF->getRegInfo();
20808 const TargetRegisterClass *AddrRegClass =
20809 getRegClassFor(getPointerTy(MF->getDataLayout()));
20811 unsigned mallocPtrVReg = MRI.createVirtualRegister(AddrRegClass),
20812 bumpSPPtrVReg = MRI.createVirtualRegister(AddrRegClass),
20813 tmpSPVReg = MRI.createVirtualRegister(AddrRegClass),
20814 SPLimitVReg = MRI.createVirtualRegister(AddrRegClass),
20815 sizeVReg = MI->getOperand(1).getReg(),
20816 physSPReg = IsLP64 || Subtarget->isTargetNaCl64() ? X86::RSP : X86::ESP;
20818 MachineFunction::iterator MBBIter = BB;
20821 MF->insert(MBBIter, bumpMBB);
20822 MF->insert(MBBIter, mallocMBB);
20823 MF->insert(MBBIter, continueMBB);
20825 continueMBB->splice(continueMBB->begin(), BB,
20826 std::next(MachineBasicBlock::iterator(MI)), BB->end());
20827 continueMBB->transferSuccessorsAndUpdatePHIs(BB);
20829 // Add code to the main basic block to check if the stack limit has been hit,
20830 // and if so, jump to mallocMBB otherwise to bumpMBB.
20831 BuildMI(BB, DL, TII->get(TargetOpcode::COPY), tmpSPVReg).addReg(physSPReg);
20832 BuildMI(BB, DL, TII->get(IsLP64 ? X86::SUB64rr:X86::SUB32rr), SPLimitVReg)
20833 .addReg(tmpSPVReg).addReg(sizeVReg);
20834 BuildMI(BB, DL, TII->get(IsLP64 ? X86::CMP64mr:X86::CMP32mr))
20835 .addReg(0).addImm(1).addReg(0).addImm(TlsOffset).addReg(TlsReg)
20836 .addReg(SPLimitVReg);
20837 BuildMI(BB, DL, TII->get(X86::JG_1)).addMBB(mallocMBB);
20839 // bumpMBB simply decreases the stack pointer, since we know the current
20840 // stacklet has enough space.
20841 BuildMI(bumpMBB, DL, TII->get(TargetOpcode::COPY), physSPReg)
20842 .addReg(SPLimitVReg);
20843 BuildMI(bumpMBB, DL, TII->get(TargetOpcode::COPY), bumpSPPtrVReg)
20844 .addReg(SPLimitVReg);
20845 BuildMI(bumpMBB, DL, TII->get(X86::JMP_1)).addMBB(continueMBB);
20847 // Calls into a routine in libgcc to allocate more space from the heap.
20848 const uint32_t *RegMask =
20849 Subtarget->getRegisterInfo()->getCallPreservedMask(*MF, CallingConv::C);
20851 BuildMI(mallocMBB, DL, TII->get(X86::MOV64rr), X86::RDI)
20853 BuildMI(mallocMBB, DL, TII->get(X86::CALL64pcrel32))
20854 .addExternalSymbol("__morestack_allocate_stack_space")
20855 .addRegMask(RegMask)
20856 .addReg(X86::RDI, RegState::Implicit)
20857 .addReg(X86::RAX, RegState::ImplicitDefine);
20858 } else if (Is64Bit) {
20859 BuildMI(mallocMBB, DL, TII->get(X86::MOV32rr), X86::EDI)
20861 BuildMI(mallocMBB, DL, TII->get(X86::CALL64pcrel32))
20862 .addExternalSymbol("__morestack_allocate_stack_space")
20863 .addRegMask(RegMask)
20864 .addReg(X86::EDI, RegState::Implicit)
20865 .addReg(X86::EAX, RegState::ImplicitDefine);
20867 BuildMI(mallocMBB, DL, TII->get(X86::SUB32ri), physSPReg).addReg(physSPReg)
20869 BuildMI(mallocMBB, DL, TII->get(X86::PUSH32r)).addReg(sizeVReg);
20870 BuildMI(mallocMBB, DL, TII->get(X86::CALLpcrel32))
20871 .addExternalSymbol("__morestack_allocate_stack_space")
20872 .addRegMask(RegMask)
20873 .addReg(X86::EAX, RegState::ImplicitDefine);
20877 BuildMI(mallocMBB, DL, TII->get(X86::ADD32ri), physSPReg).addReg(physSPReg)
20880 BuildMI(mallocMBB, DL, TII->get(TargetOpcode::COPY), mallocPtrVReg)
20881 .addReg(IsLP64 ? X86::RAX : X86::EAX);
20882 BuildMI(mallocMBB, DL, TII->get(X86::JMP_1)).addMBB(continueMBB);
20884 // Set up the CFG correctly.
20885 BB->addSuccessor(bumpMBB);
20886 BB->addSuccessor(mallocMBB);
20887 mallocMBB->addSuccessor(continueMBB);
20888 bumpMBB->addSuccessor(continueMBB);
20890 // Take care of the PHI nodes.
20891 BuildMI(*continueMBB, continueMBB->begin(), DL, TII->get(X86::PHI),
20892 MI->getOperand(0).getReg())
20893 .addReg(mallocPtrVReg).addMBB(mallocMBB)
20894 .addReg(bumpSPPtrVReg).addMBB(bumpMBB);
20896 // Delete the original pseudo instruction.
20897 MI->eraseFromParent();
20900 return continueMBB;
20903 MachineBasicBlock *
20904 X86TargetLowering::EmitLoweredWinAlloca(MachineInstr *MI,
20905 MachineBasicBlock *BB) const {
20906 DebugLoc DL = MI->getDebugLoc();
20908 assert(!Subtarget->isTargetMachO());
20910 Subtarget->getFrameLowering()->emitStackProbeCall(*BB->getParent(), *BB, MI,
20913 MI->eraseFromParent(); // The pseudo instruction is gone now.
20917 MachineBasicBlock *
20918 X86TargetLowering::EmitLoweredTLSCall(MachineInstr *MI,
20919 MachineBasicBlock *BB) const {
20920 // This is pretty easy. We're taking the value that we received from
20921 // our load from the relocation, sticking it in either RDI (x86-64)
20922 // or EAX and doing an indirect call. The return value will then
20923 // be in the normal return register.
20924 MachineFunction *F = BB->getParent();
20925 const X86InstrInfo *TII = Subtarget->getInstrInfo();
20926 DebugLoc DL = MI->getDebugLoc();
20928 assert(Subtarget->isTargetDarwin() && "Darwin only instr emitted?");
20929 assert(MI->getOperand(3).isGlobal() && "This should be a global");
20931 // Get a register mask for the lowered call.
20932 // FIXME: The 32-bit calls have non-standard calling conventions. Use a
20933 // proper register mask.
20934 const uint32_t *RegMask =
20935 Subtarget->getRegisterInfo()->getCallPreservedMask(*F, CallingConv::C);
20936 if (Subtarget->is64Bit()) {
20937 MachineInstrBuilder MIB = BuildMI(*BB, MI, DL,
20938 TII->get(X86::MOV64rm), X86::RDI)
20940 .addImm(0).addReg(0)
20941 .addGlobalAddress(MI->getOperand(3).getGlobal(), 0,
20942 MI->getOperand(3).getTargetFlags())
20944 MIB = BuildMI(*BB, MI, DL, TII->get(X86::CALL64m));
20945 addDirectMem(MIB, X86::RDI);
20946 MIB.addReg(X86::RAX, RegState::ImplicitDefine).addRegMask(RegMask);
20947 } else if (F->getTarget().getRelocationModel() != Reloc::PIC_) {
20948 MachineInstrBuilder MIB = BuildMI(*BB, MI, DL,
20949 TII->get(X86::MOV32rm), X86::EAX)
20951 .addImm(0).addReg(0)
20952 .addGlobalAddress(MI->getOperand(3).getGlobal(), 0,
20953 MI->getOperand(3).getTargetFlags())
20955 MIB = BuildMI(*BB, MI, DL, TII->get(X86::CALL32m));
20956 addDirectMem(MIB, X86::EAX);
20957 MIB.addReg(X86::EAX, RegState::ImplicitDefine).addRegMask(RegMask);
20959 MachineInstrBuilder MIB = BuildMI(*BB, MI, DL,
20960 TII->get(X86::MOV32rm), X86::EAX)
20961 .addReg(TII->getGlobalBaseReg(F))
20962 .addImm(0).addReg(0)
20963 .addGlobalAddress(MI->getOperand(3).getGlobal(), 0,
20964 MI->getOperand(3).getTargetFlags())
20966 MIB = BuildMI(*BB, MI, DL, TII->get(X86::CALL32m));
20967 addDirectMem(MIB, X86::EAX);
20968 MIB.addReg(X86::EAX, RegState::ImplicitDefine).addRegMask(RegMask);
20971 MI->eraseFromParent(); // The pseudo instruction is gone now.
20975 MachineBasicBlock *
20976 X86TargetLowering::emitEHSjLjSetJmp(MachineInstr *MI,
20977 MachineBasicBlock *MBB) const {
20978 DebugLoc DL = MI->getDebugLoc();
20979 MachineFunction *MF = MBB->getParent();
20980 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
20981 MachineRegisterInfo &MRI = MF->getRegInfo();
20983 const BasicBlock *BB = MBB->getBasicBlock();
20984 MachineFunction::iterator I = MBB;
20987 // Memory Reference
20988 MachineInstr::mmo_iterator MMOBegin = MI->memoperands_begin();
20989 MachineInstr::mmo_iterator MMOEnd = MI->memoperands_end();
20992 unsigned MemOpndSlot = 0;
20994 unsigned CurOp = 0;
20996 DstReg = MI->getOperand(CurOp++).getReg();
20997 const TargetRegisterClass *RC = MRI.getRegClass(DstReg);
20998 assert(RC->hasType(MVT::i32) && "Invalid destination!");
20999 unsigned mainDstReg = MRI.createVirtualRegister(RC);
21000 unsigned restoreDstReg = MRI.createVirtualRegister(RC);
21002 MemOpndSlot = CurOp;
21004 MVT PVT = getPointerTy(MF->getDataLayout());
21005 assert((PVT == MVT::i64 || PVT == MVT::i32) &&
21006 "Invalid Pointer Size!");
21008 // For v = setjmp(buf), we generate
21011 // buf[LabelOffset] = restoreMBB
21012 // SjLjSetup restoreMBB
21018 // v = phi(main, restore)
21021 // if base pointer being used, load it from frame
21024 MachineBasicBlock *thisMBB = MBB;
21025 MachineBasicBlock *mainMBB = MF->CreateMachineBasicBlock(BB);
21026 MachineBasicBlock *sinkMBB = MF->CreateMachineBasicBlock(BB);
21027 MachineBasicBlock *restoreMBB = MF->CreateMachineBasicBlock(BB);
21028 MF->insert(I, mainMBB);
21029 MF->insert(I, sinkMBB);
21030 MF->push_back(restoreMBB);
21032 MachineInstrBuilder MIB;
21034 // Transfer the remainder of BB and its successor edges to sinkMBB.
21035 sinkMBB->splice(sinkMBB->begin(), MBB,
21036 std::next(MachineBasicBlock::iterator(MI)), MBB->end());
21037 sinkMBB->transferSuccessorsAndUpdatePHIs(MBB);
21040 unsigned PtrStoreOpc = 0;
21041 unsigned LabelReg = 0;
21042 const int64_t LabelOffset = 1 * PVT.getStoreSize();
21043 Reloc::Model RM = MF->getTarget().getRelocationModel();
21044 bool UseImmLabel = (MF->getTarget().getCodeModel() == CodeModel::Small) &&
21045 (RM == Reloc::Static || RM == Reloc::DynamicNoPIC);
21047 // Prepare IP either in reg or imm.
21048 if (!UseImmLabel) {
21049 PtrStoreOpc = (PVT == MVT::i64) ? X86::MOV64mr : X86::MOV32mr;
21050 const TargetRegisterClass *PtrRC = getRegClassFor(PVT);
21051 LabelReg = MRI.createVirtualRegister(PtrRC);
21052 if (Subtarget->is64Bit()) {
21053 MIB = BuildMI(*thisMBB, MI, DL, TII->get(X86::LEA64r), LabelReg)
21057 .addMBB(restoreMBB)
21060 const X86InstrInfo *XII = static_cast<const X86InstrInfo*>(TII);
21061 MIB = BuildMI(*thisMBB, MI, DL, TII->get(X86::LEA32r), LabelReg)
21062 .addReg(XII->getGlobalBaseReg(MF))
21065 .addMBB(restoreMBB, Subtarget->ClassifyBlockAddressReference())
21069 PtrStoreOpc = (PVT == MVT::i64) ? X86::MOV64mi32 : X86::MOV32mi;
21071 MIB = BuildMI(*thisMBB, MI, DL, TII->get(PtrStoreOpc));
21072 for (unsigned i = 0; i < X86::AddrNumOperands; ++i) {
21073 if (i == X86::AddrDisp)
21074 MIB.addDisp(MI->getOperand(MemOpndSlot + i), LabelOffset);
21076 MIB.addOperand(MI->getOperand(MemOpndSlot + i));
21079 MIB.addReg(LabelReg);
21081 MIB.addMBB(restoreMBB);
21082 MIB.setMemRefs(MMOBegin, MMOEnd);
21084 MIB = BuildMI(*thisMBB, MI, DL, TII->get(X86::EH_SjLj_Setup))
21085 .addMBB(restoreMBB);
21087 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
21088 MIB.addRegMask(RegInfo->getNoPreservedMask());
21089 thisMBB->addSuccessor(mainMBB);
21090 thisMBB->addSuccessor(restoreMBB);
21094 BuildMI(mainMBB, DL, TII->get(X86::MOV32r0), mainDstReg);
21095 mainMBB->addSuccessor(sinkMBB);
21098 BuildMI(*sinkMBB, sinkMBB->begin(), DL,
21099 TII->get(X86::PHI), DstReg)
21100 .addReg(mainDstReg).addMBB(mainMBB)
21101 .addReg(restoreDstReg).addMBB(restoreMBB);
21104 if (RegInfo->hasBasePointer(*MF)) {
21105 const bool Uses64BitFramePtr =
21106 Subtarget->isTarget64BitLP64() || Subtarget->isTargetNaCl64();
21107 X86MachineFunctionInfo *X86FI = MF->getInfo<X86MachineFunctionInfo>();
21108 X86FI->setRestoreBasePointer(MF);
21109 unsigned FramePtr = RegInfo->getFrameRegister(*MF);
21110 unsigned BasePtr = RegInfo->getBaseRegister();
21111 unsigned Opm = Uses64BitFramePtr ? X86::MOV64rm : X86::MOV32rm;
21112 addRegOffset(BuildMI(restoreMBB, DL, TII->get(Opm), BasePtr),
21113 FramePtr, true, X86FI->getRestoreBasePointerOffset())
21114 .setMIFlag(MachineInstr::FrameSetup);
21116 BuildMI(restoreMBB, DL, TII->get(X86::MOV32ri), restoreDstReg).addImm(1);
21117 BuildMI(restoreMBB, DL, TII->get(X86::JMP_1)).addMBB(sinkMBB);
21118 restoreMBB->addSuccessor(sinkMBB);
21120 MI->eraseFromParent();
21124 MachineBasicBlock *
21125 X86TargetLowering::emitEHSjLjLongJmp(MachineInstr *MI,
21126 MachineBasicBlock *MBB) const {
21127 DebugLoc DL = MI->getDebugLoc();
21128 MachineFunction *MF = MBB->getParent();
21129 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
21130 MachineRegisterInfo &MRI = MF->getRegInfo();
21132 // Memory Reference
21133 MachineInstr::mmo_iterator MMOBegin = MI->memoperands_begin();
21134 MachineInstr::mmo_iterator MMOEnd = MI->memoperands_end();
21136 MVT PVT = getPointerTy(MF->getDataLayout());
21137 assert((PVT == MVT::i64 || PVT == MVT::i32) &&
21138 "Invalid Pointer Size!");
21140 const TargetRegisterClass *RC =
21141 (PVT == MVT::i64) ? &X86::GR64RegClass : &X86::GR32RegClass;
21142 unsigned Tmp = MRI.createVirtualRegister(RC);
21143 // Since FP is only updated here but NOT referenced, it's treated as GPR.
21144 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
21145 unsigned FP = (PVT == MVT::i64) ? X86::RBP : X86::EBP;
21146 unsigned SP = RegInfo->getStackRegister();
21148 MachineInstrBuilder MIB;
21150 const int64_t LabelOffset = 1 * PVT.getStoreSize();
21151 const int64_t SPOffset = 2 * PVT.getStoreSize();
21153 unsigned PtrLoadOpc = (PVT == MVT::i64) ? X86::MOV64rm : X86::MOV32rm;
21154 unsigned IJmpOpc = (PVT == MVT::i64) ? X86::JMP64r : X86::JMP32r;
21157 MIB = BuildMI(*MBB, MI, DL, TII->get(PtrLoadOpc), FP);
21158 for (unsigned i = 0; i < X86::AddrNumOperands; ++i)
21159 MIB.addOperand(MI->getOperand(i));
21160 MIB.setMemRefs(MMOBegin, MMOEnd);
21162 MIB = BuildMI(*MBB, MI, DL, TII->get(PtrLoadOpc), Tmp);
21163 for (unsigned i = 0; i < X86::AddrNumOperands; ++i) {
21164 if (i == X86::AddrDisp)
21165 MIB.addDisp(MI->getOperand(i), LabelOffset);
21167 MIB.addOperand(MI->getOperand(i));
21169 MIB.setMemRefs(MMOBegin, MMOEnd);
21171 MIB = BuildMI(*MBB, MI, DL, TII->get(PtrLoadOpc), SP);
21172 for (unsigned i = 0; i < X86::AddrNumOperands; ++i) {
21173 if (i == X86::AddrDisp)
21174 MIB.addDisp(MI->getOperand(i), SPOffset);
21176 MIB.addOperand(MI->getOperand(i));
21178 MIB.setMemRefs(MMOBegin, MMOEnd);
21180 BuildMI(*MBB, MI, DL, TII->get(IJmpOpc)).addReg(Tmp);
21182 MI->eraseFromParent();
21186 // Replace 213-type (isel default) FMA3 instructions with 231-type for
21187 // accumulator loops. Writing back to the accumulator allows the coalescer
21188 // to remove extra copies in the loop.
21189 // FIXME: Do this on AVX512. We don't support 231 variants yet (PR23937).
21190 MachineBasicBlock *
21191 X86TargetLowering::emitFMA3Instr(MachineInstr *MI,
21192 MachineBasicBlock *MBB) const {
21193 MachineOperand &AddendOp = MI->getOperand(3);
21195 // Bail out early if the addend isn't a register - we can't switch these.
21196 if (!AddendOp.isReg())
21199 MachineFunction &MF = *MBB->getParent();
21200 MachineRegisterInfo &MRI = MF.getRegInfo();
21202 // Check whether the addend is defined by a PHI:
21203 assert(MRI.hasOneDef(AddendOp.getReg()) && "Multiple defs in SSA?");
21204 MachineInstr &AddendDef = *MRI.def_instr_begin(AddendOp.getReg());
21205 if (!AddendDef.isPHI())
21208 // Look for the following pattern:
21210 // %addend = phi [%entry, 0], [%loop, %result]
21212 // %result<tied1> = FMA213 %m2<tied0>, %m1, %addend
21216 // %addend = phi [%entry, 0], [%loop, %result]
21218 // %result<tied1> = FMA231 %addend<tied0>, %m1, %m2
21220 for (unsigned i = 1, e = AddendDef.getNumOperands(); i < e; i += 2) {
21221 assert(AddendDef.getOperand(i).isReg());
21222 MachineOperand PHISrcOp = AddendDef.getOperand(i);
21223 MachineInstr &PHISrcInst = *MRI.def_instr_begin(PHISrcOp.getReg());
21224 if (&PHISrcInst == MI) {
21225 // Found a matching instruction.
21226 unsigned NewFMAOpc = 0;
21227 switch (MI->getOpcode()) {
21228 case X86::VFMADDPDr213r: NewFMAOpc = X86::VFMADDPDr231r; break;
21229 case X86::VFMADDPSr213r: NewFMAOpc = X86::VFMADDPSr231r; break;
21230 case X86::VFMADDSDr213r: NewFMAOpc = X86::VFMADDSDr231r; break;
21231 case X86::VFMADDSSr213r: NewFMAOpc = X86::VFMADDSSr231r; break;
21232 case X86::VFMSUBPDr213r: NewFMAOpc = X86::VFMSUBPDr231r; break;
21233 case X86::VFMSUBPSr213r: NewFMAOpc = X86::VFMSUBPSr231r; break;
21234 case X86::VFMSUBSDr213r: NewFMAOpc = X86::VFMSUBSDr231r; break;
21235 case X86::VFMSUBSSr213r: NewFMAOpc = X86::VFMSUBSSr231r; break;
21236 case X86::VFNMADDPDr213r: NewFMAOpc = X86::VFNMADDPDr231r; break;
21237 case X86::VFNMADDPSr213r: NewFMAOpc = X86::VFNMADDPSr231r; break;
21238 case X86::VFNMADDSDr213r: NewFMAOpc = X86::VFNMADDSDr231r; break;
21239 case X86::VFNMADDSSr213r: NewFMAOpc = X86::VFNMADDSSr231r; break;
21240 case X86::VFNMSUBPDr213r: NewFMAOpc = X86::VFNMSUBPDr231r; break;
21241 case X86::VFNMSUBPSr213r: NewFMAOpc = X86::VFNMSUBPSr231r; break;
21242 case X86::VFNMSUBSDr213r: NewFMAOpc = X86::VFNMSUBSDr231r; break;
21243 case X86::VFNMSUBSSr213r: NewFMAOpc = X86::VFNMSUBSSr231r; break;
21244 case X86::VFMADDSUBPDr213r: NewFMAOpc = X86::VFMADDSUBPDr231r; break;
21245 case X86::VFMADDSUBPSr213r: NewFMAOpc = X86::VFMADDSUBPSr231r; break;
21246 case X86::VFMSUBADDPDr213r: NewFMAOpc = X86::VFMSUBADDPDr231r; break;
21247 case X86::VFMSUBADDPSr213r: NewFMAOpc = X86::VFMSUBADDPSr231r; break;
21249 case X86::VFMADDPDr213rY: NewFMAOpc = X86::VFMADDPDr231rY; break;
21250 case X86::VFMADDPSr213rY: NewFMAOpc = X86::VFMADDPSr231rY; break;
21251 case X86::VFMSUBPDr213rY: NewFMAOpc = X86::VFMSUBPDr231rY; break;
21252 case X86::VFMSUBPSr213rY: NewFMAOpc = X86::VFMSUBPSr231rY; break;
21253 case X86::VFNMADDPDr213rY: NewFMAOpc = X86::VFNMADDPDr231rY; break;
21254 case X86::VFNMADDPSr213rY: NewFMAOpc = X86::VFNMADDPSr231rY; break;
21255 case X86::VFNMSUBPDr213rY: NewFMAOpc = X86::VFNMSUBPDr231rY; break;
21256 case X86::VFNMSUBPSr213rY: NewFMAOpc = X86::VFNMSUBPSr231rY; break;
21257 case X86::VFMADDSUBPDr213rY: NewFMAOpc = X86::VFMADDSUBPDr231rY; break;
21258 case X86::VFMADDSUBPSr213rY: NewFMAOpc = X86::VFMADDSUBPSr231rY; break;
21259 case X86::VFMSUBADDPDr213rY: NewFMAOpc = X86::VFMSUBADDPDr231rY; break;
21260 case X86::VFMSUBADDPSr213rY: NewFMAOpc = X86::VFMSUBADDPSr231rY; break;
21261 default: llvm_unreachable("Unrecognized FMA variant.");
21264 const TargetInstrInfo &TII = *Subtarget->getInstrInfo();
21265 MachineInstrBuilder MIB =
21266 BuildMI(MF, MI->getDebugLoc(), TII.get(NewFMAOpc))
21267 .addOperand(MI->getOperand(0))
21268 .addOperand(MI->getOperand(3))
21269 .addOperand(MI->getOperand(2))
21270 .addOperand(MI->getOperand(1));
21271 MBB->insert(MachineBasicBlock::iterator(MI), MIB);
21272 MI->eraseFromParent();
21279 MachineBasicBlock *
21280 X86TargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
21281 MachineBasicBlock *BB) const {
21282 switch (MI->getOpcode()) {
21283 default: llvm_unreachable("Unexpected instr type to insert");
21284 case X86::TAILJMPd64:
21285 case X86::TAILJMPr64:
21286 case X86::TAILJMPm64:
21287 case X86::TAILJMPd64_REX:
21288 case X86::TAILJMPr64_REX:
21289 case X86::TAILJMPm64_REX:
21290 llvm_unreachable("TAILJMP64 would not be touched here.");
21291 case X86::TCRETURNdi64:
21292 case X86::TCRETURNri64:
21293 case X86::TCRETURNmi64:
21295 case X86::WIN_ALLOCA:
21296 return EmitLoweredWinAlloca(MI, BB);
21297 case X86::SEG_ALLOCA_32:
21298 case X86::SEG_ALLOCA_64:
21299 return EmitLoweredSegAlloca(MI, BB);
21300 case X86::TLSCall_32:
21301 case X86::TLSCall_64:
21302 return EmitLoweredTLSCall(MI, BB);
21303 case X86::CMOV_FR32:
21304 case X86::CMOV_FR64:
21305 case X86::CMOV_GR8:
21306 case X86::CMOV_GR16:
21307 case X86::CMOV_GR32:
21308 case X86::CMOV_RFP32:
21309 case X86::CMOV_RFP64:
21310 case X86::CMOV_RFP80:
21311 case X86::CMOV_V2F64:
21312 case X86::CMOV_V2I64:
21313 case X86::CMOV_V4F32:
21314 case X86::CMOV_V4F64:
21315 case X86::CMOV_V4I64:
21316 case X86::CMOV_V16F32:
21317 case X86::CMOV_V8F32:
21318 case X86::CMOV_V8F64:
21319 case X86::CMOV_V8I64:
21320 case X86::CMOV_V8I1:
21321 case X86::CMOV_V16I1:
21322 case X86::CMOV_V32I1:
21323 case X86::CMOV_V64I1:
21324 return EmitLoweredSelect(MI, BB);
21326 case X86::RELEASE_FADD32mr:
21327 case X86::RELEASE_FADD64mr:
21328 return EmitLoweredAtomicFP(MI, BB);
21330 case X86::FP32_TO_INT16_IN_MEM:
21331 case X86::FP32_TO_INT32_IN_MEM:
21332 case X86::FP32_TO_INT64_IN_MEM:
21333 case X86::FP64_TO_INT16_IN_MEM:
21334 case X86::FP64_TO_INT32_IN_MEM:
21335 case X86::FP64_TO_INT64_IN_MEM:
21336 case X86::FP80_TO_INT16_IN_MEM:
21337 case X86::FP80_TO_INT32_IN_MEM:
21338 case X86::FP80_TO_INT64_IN_MEM: {
21339 MachineFunction *F = BB->getParent();
21340 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
21341 DebugLoc DL = MI->getDebugLoc();
21343 // Change the floating point control register to use "round towards zero"
21344 // mode when truncating to an integer value.
21345 int CWFrameIdx = F->getFrameInfo()->CreateStackObject(2, 2, false);
21346 addFrameReference(BuildMI(*BB, MI, DL,
21347 TII->get(X86::FNSTCW16m)), CWFrameIdx);
21349 // Load the old value of the high byte of the control word...
21351 F->getRegInfo().createVirtualRegister(&X86::GR16RegClass);
21352 addFrameReference(BuildMI(*BB, MI, DL, TII->get(X86::MOV16rm), OldCW),
21355 // Set the high part to be round to zero...
21356 addFrameReference(BuildMI(*BB, MI, DL, TII->get(X86::MOV16mi)), CWFrameIdx)
21359 // Reload the modified control word now...
21360 addFrameReference(BuildMI(*BB, MI, DL,
21361 TII->get(X86::FLDCW16m)), CWFrameIdx);
21363 // Restore the memory image of control word to original value
21364 addFrameReference(BuildMI(*BB, MI, DL, TII->get(X86::MOV16mr)), CWFrameIdx)
21367 // Get the X86 opcode to use.
21369 switch (MI->getOpcode()) {
21370 default: llvm_unreachable("illegal opcode!");
21371 case X86::FP32_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m32; break;
21372 case X86::FP32_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m32; break;
21373 case X86::FP32_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m32; break;
21374 case X86::FP64_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m64; break;
21375 case X86::FP64_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m64; break;
21376 case X86::FP64_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m64; break;
21377 case X86::FP80_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m80; break;
21378 case X86::FP80_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m80; break;
21379 case X86::FP80_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m80; break;
21383 MachineOperand &Op = MI->getOperand(0);
21385 AM.BaseType = X86AddressMode::RegBase;
21386 AM.Base.Reg = Op.getReg();
21388 AM.BaseType = X86AddressMode::FrameIndexBase;
21389 AM.Base.FrameIndex = Op.getIndex();
21391 Op = MI->getOperand(1);
21393 AM.Scale = Op.getImm();
21394 Op = MI->getOperand(2);
21396 AM.IndexReg = Op.getImm();
21397 Op = MI->getOperand(3);
21398 if (Op.isGlobal()) {
21399 AM.GV = Op.getGlobal();
21401 AM.Disp = Op.getImm();
21403 addFullAddress(BuildMI(*BB, MI, DL, TII->get(Opc)), AM)
21404 .addReg(MI->getOperand(X86::AddrNumOperands).getReg());
21406 // Reload the original control word now.
21407 addFrameReference(BuildMI(*BB, MI, DL,
21408 TII->get(X86::FLDCW16m)), CWFrameIdx);
21410 MI->eraseFromParent(); // The pseudo instruction is gone now.
21413 // String/text processing lowering.
21414 case X86::PCMPISTRM128REG:
21415 case X86::VPCMPISTRM128REG:
21416 case X86::PCMPISTRM128MEM:
21417 case X86::VPCMPISTRM128MEM:
21418 case X86::PCMPESTRM128REG:
21419 case X86::VPCMPESTRM128REG:
21420 case X86::PCMPESTRM128MEM:
21421 case X86::VPCMPESTRM128MEM:
21422 assert(Subtarget->hasSSE42() &&
21423 "Target must have SSE4.2 or AVX features enabled");
21424 return EmitPCMPSTRM(MI, BB, Subtarget->getInstrInfo());
21426 // String/text processing lowering.
21427 case X86::PCMPISTRIREG:
21428 case X86::VPCMPISTRIREG:
21429 case X86::PCMPISTRIMEM:
21430 case X86::VPCMPISTRIMEM:
21431 case X86::PCMPESTRIREG:
21432 case X86::VPCMPESTRIREG:
21433 case X86::PCMPESTRIMEM:
21434 case X86::VPCMPESTRIMEM:
21435 assert(Subtarget->hasSSE42() &&
21436 "Target must have SSE4.2 or AVX features enabled");
21437 return EmitPCMPSTRI(MI, BB, Subtarget->getInstrInfo());
21439 // Thread synchronization.
21441 return EmitMonitor(MI, BB, Subtarget);
21445 return EmitXBegin(MI, BB, Subtarget->getInstrInfo());
21447 case X86::VASTART_SAVE_XMM_REGS:
21448 return EmitVAStartSaveXMMRegsWithCustomInserter(MI, BB);
21450 case X86::VAARG_64:
21451 return EmitVAARG64WithCustomInserter(MI, BB);
21453 case X86::EH_SjLj_SetJmp32:
21454 case X86::EH_SjLj_SetJmp64:
21455 return emitEHSjLjSetJmp(MI, BB);
21457 case X86::EH_SjLj_LongJmp32:
21458 case X86::EH_SjLj_LongJmp64:
21459 return emitEHSjLjLongJmp(MI, BB);
21461 case TargetOpcode::STATEPOINT:
21462 // As an implementation detail, STATEPOINT shares the STACKMAP format at
21463 // this point in the process. We diverge later.
21464 return emitPatchPoint(MI, BB);
21466 case TargetOpcode::STACKMAP:
21467 case TargetOpcode::PATCHPOINT:
21468 return emitPatchPoint(MI, BB);
21470 case X86::VFMADDPDr213r:
21471 case X86::VFMADDPSr213r:
21472 case X86::VFMADDSDr213r:
21473 case X86::VFMADDSSr213r:
21474 case X86::VFMSUBPDr213r:
21475 case X86::VFMSUBPSr213r:
21476 case X86::VFMSUBSDr213r:
21477 case X86::VFMSUBSSr213r:
21478 case X86::VFNMADDPDr213r:
21479 case X86::VFNMADDPSr213r:
21480 case X86::VFNMADDSDr213r:
21481 case X86::VFNMADDSSr213r:
21482 case X86::VFNMSUBPDr213r:
21483 case X86::VFNMSUBPSr213r:
21484 case X86::VFNMSUBSDr213r:
21485 case X86::VFNMSUBSSr213r:
21486 case X86::VFMADDSUBPDr213r:
21487 case X86::VFMADDSUBPSr213r:
21488 case X86::VFMSUBADDPDr213r:
21489 case X86::VFMSUBADDPSr213r:
21490 case X86::VFMADDPDr213rY:
21491 case X86::VFMADDPSr213rY:
21492 case X86::VFMSUBPDr213rY:
21493 case X86::VFMSUBPSr213rY:
21494 case X86::VFNMADDPDr213rY:
21495 case X86::VFNMADDPSr213rY:
21496 case X86::VFNMSUBPDr213rY:
21497 case X86::VFNMSUBPSr213rY:
21498 case X86::VFMADDSUBPDr213rY:
21499 case X86::VFMADDSUBPSr213rY:
21500 case X86::VFMSUBADDPDr213rY:
21501 case X86::VFMSUBADDPSr213rY:
21502 return emitFMA3Instr(MI, BB);
21506 //===----------------------------------------------------------------------===//
21507 // X86 Optimization Hooks
21508 //===----------------------------------------------------------------------===//
21510 void X86TargetLowering::computeKnownBitsForTargetNode(const SDValue Op,
21513 const SelectionDAG &DAG,
21514 unsigned Depth) const {
21515 unsigned BitWidth = KnownZero.getBitWidth();
21516 unsigned Opc = Op.getOpcode();
21517 assert((Opc >= ISD::BUILTIN_OP_END ||
21518 Opc == ISD::INTRINSIC_WO_CHAIN ||
21519 Opc == ISD::INTRINSIC_W_CHAIN ||
21520 Opc == ISD::INTRINSIC_VOID) &&
21521 "Should use MaskedValueIsZero if you don't know whether Op"
21522 " is a target node!");
21524 KnownZero = KnownOne = APInt(BitWidth, 0); // Don't know anything.
21538 // These nodes' second result is a boolean.
21539 if (Op.getResNo() == 0)
21542 case X86ISD::SETCC:
21543 KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - 1);
21545 case ISD::INTRINSIC_WO_CHAIN: {
21546 unsigned IntId = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
21547 unsigned NumLoBits = 0;
21550 case Intrinsic::x86_sse_movmsk_ps:
21551 case Intrinsic::x86_avx_movmsk_ps_256:
21552 case Intrinsic::x86_sse2_movmsk_pd:
21553 case Intrinsic::x86_avx_movmsk_pd_256:
21554 case Intrinsic::x86_mmx_pmovmskb:
21555 case Intrinsic::x86_sse2_pmovmskb_128:
21556 case Intrinsic::x86_avx2_pmovmskb: {
21557 // High bits of movmskp{s|d}, pmovmskb are known zero.
21559 default: llvm_unreachable("Impossible intrinsic"); // Can't reach here.
21560 case Intrinsic::x86_sse_movmsk_ps: NumLoBits = 4; break;
21561 case Intrinsic::x86_avx_movmsk_ps_256: NumLoBits = 8; break;
21562 case Intrinsic::x86_sse2_movmsk_pd: NumLoBits = 2; break;
21563 case Intrinsic::x86_avx_movmsk_pd_256: NumLoBits = 4; break;
21564 case Intrinsic::x86_mmx_pmovmskb: NumLoBits = 8; break;
21565 case Intrinsic::x86_sse2_pmovmskb_128: NumLoBits = 16; break;
21566 case Intrinsic::x86_avx2_pmovmskb: NumLoBits = 32; break;
21568 KnownZero = APInt::getHighBitsSet(BitWidth, BitWidth - NumLoBits);
21577 unsigned X86TargetLowering::ComputeNumSignBitsForTargetNode(
21579 const SelectionDAG &,
21580 unsigned Depth) const {
21581 // SETCC_CARRY sets the dest to ~0 for true or 0 for false.
21582 if (Op.getOpcode() == X86ISD::SETCC_CARRY)
21583 return Op.getValueType().getScalarType().getSizeInBits();
21589 /// isGAPlusOffset - Returns true (and the GlobalValue and the offset) if the
21590 /// node is a GlobalAddress + offset.
21591 bool X86TargetLowering::isGAPlusOffset(SDNode *N,
21592 const GlobalValue* &GA,
21593 int64_t &Offset) const {
21594 if (N->getOpcode() == X86ISD::Wrapper) {
21595 if (isa<GlobalAddressSDNode>(N->getOperand(0))) {
21596 GA = cast<GlobalAddressSDNode>(N->getOperand(0))->getGlobal();
21597 Offset = cast<GlobalAddressSDNode>(N->getOperand(0))->getOffset();
21601 return TargetLowering::isGAPlusOffset(N, GA, Offset);
21604 /// isShuffleHigh128VectorInsertLow - Checks whether the shuffle node is the
21605 /// same as extracting the high 128-bit part of 256-bit vector and then
21606 /// inserting the result into the low part of a new 256-bit vector
21607 static bool isShuffleHigh128VectorInsertLow(ShuffleVectorSDNode *SVOp) {
21608 EVT VT = SVOp->getValueType(0);
21609 unsigned NumElems = VT.getVectorNumElements();
21611 // vector_shuffle <4, 5, 6, 7, u, u, u, u> or <2, 3, u, u>
21612 for (unsigned i = 0, j = NumElems/2; i != NumElems/2; ++i, ++j)
21613 if (!isUndefOrEqual(SVOp->getMaskElt(i), j) ||
21614 SVOp->getMaskElt(j) >= 0)
21620 /// isShuffleLow128VectorInsertHigh - Checks whether the shuffle node is the
21621 /// same as extracting the low 128-bit part of 256-bit vector and then
21622 /// inserting the result into the high part of a new 256-bit vector
21623 static bool isShuffleLow128VectorInsertHigh(ShuffleVectorSDNode *SVOp) {
21624 EVT VT = SVOp->getValueType(0);
21625 unsigned NumElems = VT.getVectorNumElements();
21627 // vector_shuffle <u, u, u, u, 0, 1, 2, 3> or <u, u, 0, 1>
21628 for (unsigned i = NumElems/2, j = 0; i != NumElems; ++i, ++j)
21629 if (!isUndefOrEqual(SVOp->getMaskElt(i), j) ||
21630 SVOp->getMaskElt(j) >= 0)
21636 /// PerformShuffleCombine256 - Performs shuffle combines for 256-bit vectors.
21637 static SDValue PerformShuffleCombine256(SDNode *N, SelectionDAG &DAG,
21638 TargetLowering::DAGCombinerInfo &DCI,
21639 const X86Subtarget* Subtarget) {
21641 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
21642 SDValue V1 = SVOp->getOperand(0);
21643 SDValue V2 = SVOp->getOperand(1);
21644 EVT VT = SVOp->getValueType(0);
21645 unsigned NumElems = VT.getVectorNumElements();
21647 if (V1.getOpcode() == ISD::CONCAT_VECTORS &&
21648 V2.getOpcode() == ISD::CONCAT_VECTORS) {
21652 // V UNDEF BUILD_VECTOR UNDEF
21654 // CONCAT_VECTOR CONCAT_VECTOR
21657 // RESULT: V + zero extended
21659 if (V2.getOperand(0).getOpcode() != ISD::BUILD_VECTOR ||
21660 V2.getOperand(1).getOpcode() != ISD::UNDEF ||
21661 V1.getOperand(1).getOpcode() != ISD::UNDEF)
21664 if (!ISD::isBuildVectorAllZeros(V2.getOperand(0).getNode()))
21667 // To match the shuffle mask, the first half of the mask should
21668 // be exactly the first vector, and all the rest a splat with the
21669 // first element of the second one.
21670 for (unsigned i = 0; i != NumElems/2; ++i)
21671 if (!isUndefOrEqual(SVOp->getMaskElt(i), i) ||
21672 !isUndefOrEqual(SVOp->getMaskElt(i+NumElems/2), NumElems))
21675 // If V1 is coming from a vector load then just fold to a VZEXT_LOAD.
21676 if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(V1.getOperand(0))) {
21677 if (Ld->hasNUsesOfValue(1, 0)) {
21678 SDVTList Tys = DAG.getVTList(MVT::v4i64, MVT::Other);
21679 SDValue Ops[] = { Ld->getChain(), Ld->getBasePtr() };
21681 DAG.getMemIntrinsicNode(X86ISD::VZEXT_LOAD, dl, Tys, Ops,
21683 Ld->getPointerInfo(),
21684 Ld->getAlignment(),
21685 false/*isVolatile*/, true/*ReadMem*/,
21686 false/*WriteMem*/);
21688 // Make sure the newly-created LOAD is in the same position as Ld in
21689 // terms of dependency. We create a TokenFactor for Ld and ResNode,
21690 // and update uses of Ld's output chain to use the TokenFactor.
21691 if (Ld->hasAnyUseOfValue(1)) {
21692 SDValue NewChain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
21693 SDValue(Ld, 1), SDValue(ResNode.getNode(), 1));
21694 DAG.ReplaceAllUsesOfValueWith(SDValue(Ld, 1), NewChain);
21695 DAG.UpdateNodeOperands(NewChain.getNode(), SDValue(Ld, 1),
21696 SDValue(ResNode.getNode(), 1));
21699 return DAG.getBitcast(VT, ResNode);
21703 // Emit a zeroed vector and insert the desired subvector on its
21705 SDValue Zeros = getZeroVector(VT, Subtarget, DAG, dl);
21706 SDValue InsV = Insert128BitVector(Zeros, V1.getOperand(0), 0, DAG, dl);
21707 return DCI.CombineTo(N, InsV);
21710 //===--------------------------------------------------------------------===//
21711 // Combine some shuffles into subvector extracts and inserts:
21714 // vector_shuffle <4, 5, 6, 7, u, u, u, u> or <2, 3, u, u>
21715 if (isShuffleHigh128VectorInsertLow(SVOp)) {
21716 SDValue V = Extract128BitVector(V1, NumElems/2, DAG, dl);
21717 SDValue InsV = Insert128BitVector(DAG.getUNDEF(VT), V, 0, DAG, dl);
21718 return DCI.CombineTo(N, InsV);
21721 // vector_shuffle <u, u, u, u, 0, 1, 2, 3> or <u, u, 0, 1>
21722 if (isShuffleLow128VectorInsertHigh(SVOp)) {
21723 SDValue V = Extract128BitVector(V1, 0, DAG, dl);
21724 SDValue InsV = Insert128BitVector(DAG.getUNDEF(VT), V, NumElems/2, DAG, dl);
21725 return DCI.CombineTo(N, InsV);
21731 /// \brief Combine an arbitrary chain of shuffles into a single instruction if
21734 /// This is the leaf of the recursive combinine below. When we have found some
21735 /// chain of single-use x86 shuffle instructions and accumulated the combined
21736 /// shuffle mask represented by them, this will try to pattern match that mask
21737 /// into either a single instruction if there is a special purpose instruction
21738 /// for this operation, or into a PSHUFB instruction which is a fully general
21739 /// instruction but should only be used to replace chains over a certain depth.
21740 static bool combineX86ShuffleChain(SDValue Op, SDValue Root, ArrayRef<int> Mask,
21741 int Depth, bool HasPSHUFB, SelectionDAG &DAG,
21742 TargetLowering::DAGCombinerInfo &DCI,
21743 const X86Subtarget *Subtarget) {
21744 assert(!Mask.empty() && "Cannot combine an empty shuffle mask!");
21746 // Find the operand that enters the chain. Note that multiple uses are OK
21747 // here, we're not going to remove the operand we find.
21748 SDValue Input = Op.getOperand(0);
21749 while (Input.getOpcode() == ISD::BITCAST)
21750 Input = Input.getOperand(0);
21752 MVT VT = Input.getSimpleValueType();
21753 MVT RootVT = Root.getSimpleValueType();
21756 // Just remove no-op shuffle masks.
21757 if (Mask.size() == 1) {
21758 DCI.CombineTo(Root.getNode(), DAG.getBitcast(RootVT, Input),
21763 // Use the float domain if the operand type is a floating point type.
21764 bool FloatDomain = VT.isFloatingPoint();
21766 // For floating point shuffles, we don't have free copies in the shuffle
21767 // instructions or the ability to load as part of the instruction, so
21768 // canonicalize their shuffles to UNPCK or MOV variants.
21770 // Note that even with AVX we prefer the PSHUFD form of shuffle for integer
21771 // vectors because it can have a load folded into it that UNPCK cannot. This
21772 // doesn't preclude something switching to the shorter encoding post-RA.
21774 // FIXME: Should teach these routines about AVX vector widths.
21775 if (FloatDomain && VT.getSizeInBits() == 128) {
21776 if (Mask.equals({0, 0}) || Mask.equals({1, 1})) {
21777 bool Lo = Mask.equals({0, 0});
21780 // Check if we have SSE3 which will let us use MOVDDUP. That instruction
21781 // is no slower than UNPCKLPD but has the option to fold the input operand
21782 // into even an unaligned memory load.
21783 if (Lo && Subtarget->hasSSE3()) {
21784 Shuffle = X86ISD::MOVDDUP;
21785 ShuffleVT = MVT::v2f64;
21787 // We have MOVLHPS and MOVHLPS throughout SSE and they encode smaller
21788 // than the UNPCK variants.
21789 Shuffle = Lo ? X86ISD::MOVLHPS : X86ISD::MOVHLPS;
21790 ShuffleVT = MVT::v4f32;
21792 if (Depth == 1 && Root->getOpcode() == Shuffle)
21793 return false; // Nothing to do!
21794 Op = DAG.getBitcast(ShuffleVT, Input);
21795 DCI.AddToWorklist(Op.getNode());
21796 if (Shuffle == X86ISD::MOVDDUP)
21797 Op = DAG.getNode(Shuffle, DL, ShuffleVT, Op);
21799 Op = DAG.getNode(Shuffle, DL, ShuffleVT, Op, Op);
21800 DCI.AddToWorklist(Op.getNode());
21801 DCI.CombineTo(Root.getNode(), DAG.getBitcast(RootVT, Op),
21805 if (Subtarget->hasSSE3() &&
21806 (Mask.equals({0, 0, 2, 2}) || Mask.equals({1, 1, 3, 3}))) {
21807 bool Lo = Mask.equals({0, 0, 2, 2});
21808 unsigned Shuffle = Lo ? X86ISD::MOVSLDUP : X86ISD::MOVSHDUP;
21809 MVT ShuffleVT = MVT::v4f32;
21810 if (Depth == 1 && Root->getOpcode() == Shuffle)
21811 return false; // Nothing to do!
21812 Op = DAG.getBitcast(ShuffleVT, Input);
21813 DCI.AddToWorklist(Op.getNode());
21814 Op = DAG.getNode(Shuffle, DL, ShuffleVT, Op);
21815 DCI.AddToWorklist(Op.getNode());
21816 DCI.CombineTo(Root.getNode(), DAG.getBitcast(RootVT, Op),
21820 if (Mask.equals({0, 0, 1, 1}) || Mask.equals({2, 2, 3, 3})) {
21821 bool Lo = Mask.equals({0, 0, 1, 1});
21822 unsigned Shuffle = Lo ? X86ISD::UNPCKL : X86ISD::UNPCKH;
21823 MVT ShuffleVT = MVT::v4f32;
21824 if (Depth == 1 && Root->getOpcode() == Shuffle)
21825 return false; // Nothing to do!
21826 Op = DAG.getBitcast(ShuffleVT, Input);
21827 DCI.AddToWorklist(Op.getNode());
21828 Op = DAG.getNode(Shuffle, DL, ShuffleVT, Op, Op);
21829 DCI.AddToWorklist(Op.getNode());
21830 DCI.CombineTo(Root.getNode(), DAG.getBitcast(RootVT, Op),
21836 // We always canonicalize the 8 x i16 and 16 x i8 shuffles into their UNPCK
21837 // variants as none of these have single-instruction variants that are
21838 // superior to the UNPCK formulation.
21839 if (!FloatDomain && VT.getSizeInBits() == 128 &&
21840 (Mask.equals({0, 0, 1, 1, 2, 2, 3, 3}) ||
21841 Mask.equals({4, 4, 5, 5, 6, 6, 7, 7}) ||
21842 Mask.equals({0, 0, 1, 1, 2, 2, 3, 3, 4, 4, 5, 5, 6, 6, 7, 7}) ||
21844 {8, 8, 9, 9, 10, 10, 11, 11, 12, 12, 13, 13, 14, 14, 15, 15}))) {
21845 bool Lo = Mask[0] == 0;
21846 unsigned Shuffle = Lo ? X86ISD::UNPCKL : X86ISD::UNPCKH;
21847 if (Depth == 1 && Root->getOpcode() == Shuffle)
21848 return false; // Nothing to do!
21850 switch (Mask.size()) {
21852 ShuffleVT = MVT::v8i16;
21855 ShuffleVT = MVT::v16i8;
21858 llvm_unreachable("Impossible mask size!");
21860 Op = DAG.getBitcast(ShuffleVT, Input);
21861 DCI.AddToWorklist(Op.getNode());
21862 Op = DAG.getNode(Shuffle, DL, ShuffleVT, Op, Op);
21863 DCI.AddToWorklist(Op.getNode());
21864 DCI.CombineTo(Root.getNode(), DAG.getBitcast(RootVT, Op),
21869 // Don't try to re-form single instruction chains under any circumstances now
21870 // that we've done encoding canonicalization for them.
21874 // If we have 3 or more shuffle instructions or a chain involving PSHUFB, we
21875 // can replace them with a single PSHUFB instruction profitably. Intel's
21876 // manuals suggest only using PSHUFB if doing so replacing 5 instructions, but
21877 // in practice PSHUFB tends to be *very* fast so we're more aggressive.
21878 if ((Depth >= 3 || HasPSHUFB) && Subtarget->hasSSSE3()) {
21879 SmallVector<SDValue, 16> PSHUFBMask;
21880 int NumBytes = VT.getSizeInBits() / 8;
21881 int Ratio = NumBytes / Mask.size();
21882 for (int i = 0; i < NumBytes; ++i) {
21883 if (Mask[i / Ratio] == SM_SentinelUndef) {
21884 PSHUFBMask.push_back(DAG.getUNDEF(MVT::i8));
21887 int M = Mask[i / Ratio] != SM_SentinelZero
21888 ? Ratio * Mask[i / Ratio] + i % Ratio
21890 PSHUFBMask.push_back(DAG.getConstant(M, DL, MVT::i8));
21892 MVT ByteVT = MVT::getVectorVT(MVT::i8, NumBytes);
21893 Op = DAG.getBitcast(ByteVT, Input);
21894 DCI.AddToWorklist(Op.getNode());
21895 SDValue PSHUFBMaskOp =
21896 DAG.getNode(ISD::BUILD_VECTOR, DL, ByteVT, PSHUFBMask);
21897 DCI.AddToWorklist(PSHUFBMaskOp.getNode());
21898 Op = DAG.getNode(X86ISD::PSHUFB, DL, ByteVT, Op, PSHUFBMaskOp);
21899 DCI.AddToWorklist(Op.getNode());
21900 DCI.CombineTo(Root.getNode(), DAG.getBitcast(RootVT, Op),
21905 // Failed to find any combines.
21909 /// \brief Fully generic combining of x86 shuffle instructions.
21911 /// This should be the last combine run over the x86 shuffle instructions. Once
21912 /// they have been fully optimized, this will recursively consider all chains
21913 /// of single-use shuffle instructions, build a generic model of the cumulative
21914 /// shuffle operation, and check for simpler instructions which implement this
21915 /// operation. We use this primarily for two purposes:
21917 /// 1) Collapse generic shuffles to specialized single instructions when
21918 /// equivalent. In most cases, this is just an encoding size win, but
21919 /// sometimes we will collapse multiple generic shuffles into a single
21920 /// special-purpose shuffle.
21921 /// 2) Look for sequences of shuffle instructions with 3 or more total
21922 /// instructions, and replace them with the slightly more expensive SSSE3
21923 /// PSHUFB instruction if available. We do this as the last combining step
21924 /// to ensure we avoid using PSHUFB if we can implement the shuffle with
21925 /// a suitable short sequence of other instructions. The PHUFB will either
21926 /// use a register or have to read from memory and so is slightly (but only
21927 /// slightly) more expensive than the other shuffle instructions.
21929 /// Because this is inherently a quadratic operation (for each shuffle in
21930 /// a chain, we recurse up the chain), the depth is limited to 8 instructions.
21931 /// This should never be an issue in practice as the shuffle lowering doesn't
21932 /// produce sequences of more than 8 instructions.
21934 /// FIXME: We will currently miss some cases where the redundant shuffling
21935 /// would simplify under the threshold for PSHUFB formation because of
21936 /// combine-ordering. To fix this, we should do the redundant instruction
21937 /// combining in this recursive walk.
21938 static bool combineX86ShufflesRecursively(SDValue Op, SDValue Root,
21939 ArrayRef<int> RootMask,
21940 int Depth, bool HasPSHUFB,
21942 TargetLowering::DAGCombinerInfo &DCI,
21943 const X86Subtarget *Subtarget) {
21944 // Bound the depth of our recursive combine because this is ultimately
21945 // quadratic in nature.
21949 // Directly rip through bitcasts to find the underlying operand.
21950 while (Op.getOpcode() == ISD::BITCAST && Op.getOperand(0).hasOneUse())
21951 Op = Op.getOperand(0);
21953 MVT VT = Op.getSimpleValueType();
21954 if (!VT.isVector())
21955 return false; // Bail if we hit a non-vector.
21957 assert(Root.getSimpleValueType().isVector() &&
21958 "Shuffles operate on vector types!");
21959 assert(VT.getSizeInBits() == Root.getSimpleValueType().getSizeInBits() &&
21960 "Can only combine shuffles of the same vector register size.");
21962 if (!isTargetShuffle(Op.getOpcode()))
21964 SmallVector<int, 16> OpMask;
21966 bool HaveMask = getTargetShuffleMask(Op.getNode(), VT, OpMask, IsUnary);
21967 // We only can combine unary shuffles which we can decode the mask for.
21968 if (!HaveMask || !IsUnary)
21971 assert(VT.getVectorNumElements() == OpMask.size() &&
21972 "Different mask size from vector size!");
21973 assert(((RootMask.size() > OpMask.size() &&
21974 RootMask.size() % OpMask.size() == 0) ||
21975 (OpMask.size() > RootMask.size() &&
21976 OpMask.size() % RootMask.size() == 0) ||
21977 OpMask.size() == RootMask.size()) &&
21978 "The smaller number of elements must divide the larger.");
21979 int RootRatio = std::max<int>(1, OpMask.size() / RootMask.size());
21980 int OpRatio = std::max<int>(1, RootMask.size() / OpMask.size());
21981 assert(((RootRatio == 1 && OpRatio == 1) ||
21982 (RootRatio == 1) != (OpRatio == 1)) &&
21983 "Must not have a ratio for both incoming and op masks!");
21985 SmallVector<int, 16> Mask;
21986 Mask.reserve(std::max(OpMask.size(), RootMask.size()));
21988 // Merge this shuffle operation's mask into our accumulated mask. Note that
21989 // this shuffle's mask will be the first applied to the input, followed by the
21990 // root mask to get us all the way to the root value arrangement. The reason
21991 // for this order is that we are recursing up the operation chain.
21992 for (int i = 0, e = std::max(OpMask.size(), RootMask.size()); i < e; ++i) {
21993 int RootIdx = i / RootRatio;
21994 if (RootMask[RootIdx] < 0) {
21995 // This is a zero or undef lane, we're done.
21996 Mask.push_back(RootMask[RootIdx]);
22000 int RootMaskedIdx = RootMask[RootIdx] * RootRatio + i % RootRatio;
22001 int OpIdx = RootMaskedIdx / OpRatio;
22002 if (OpMask[OpIdx] < 0) {
22003 // The incoming lanes are zero or undef, it doesn't matter which ones we
22005 Mask.push_back(OpMask[OpIdx]);
22009 // Ok, we have non-zero lanes, map them through.
22010 Mask.push_back(OpMask[OpIdx] * OpRatio +
22011 RootMaskedIdx % OpRatio);
22014 // See if we can recurse into the operand to combine more things.
22015 switch (Op.getOpcode()) {
22016 case X86ISD::PSHUFB:
22018 case X86ISD::PSHUFD:
22019 case X86ISD::PSHUFHW:
22020 case X86ISD::PSHUFLW:
22021 if (Op.getOperand(0).hasOneUse() &&
22022 combineX86ShufflesRecursively(Op.getOperand(0), Root, Mask, Depth + 1,
22023 HasPSHUFB, DAG, DCI, Subtarget))
22027 case X86ISD::UNPCKL:
22028 case X86ISD::UNPCKH:
22029 assert(Op.getOperand(0) == Op.getOperand(1) && "We only combine unary shuffles!");
22030 // We can't check for single use, we have to check that this shuffle is the only user.
22031 if (Op->isOnlyUserOf(Op.getOperand(0).getNode()) &&
22032 combineX86ShufflesRecursively(Op.getOperand(0), Root, Mask, Depth + 1,
22033 HasPSHUFB, DAG, DCI, Subtarget))
22038 // Minor canonicalization of the accumulated shuffle mask to make it easier
22039 // to match below. All this does is detect masks with squential pairs of
22040 // elements, and shrink them to the half-width mask. It does this in a loop
22041 // so it will reduce the size of the mask to the minimal width mask which
22042 // performs an equivalent shuffle.
22043 SmallVector<int, 16> WidenedMask;
22044 while (Mask.size() > 1 && canWidenShuffleElements(Mask, WidenedMask)) {
22045 Mask = std::move(WidenedMask);
22046 WidenedMask.clear();
22049 return combineX86ShuffleChain(Op, Root, Mask, Depth, HasPSHUFB, DAG, DCI,
22053 /// \brief Get the PSHUF-style mask from PSHUF node.
22055 /// This is a very minor wrapper around getTargetShuffleMask to easy forming v4
22056 /// PSHUF-style masks that can be reused with such instructions.
22057 static SmallVector<int, 4> getPSHUFShuffleMask(SDValue N) {
22058 MVT VT = N.getSimpleValueType();
22059 SmallVector<int, 4> Mask;
22061 bool HaveMask = getTargetShuffleMask(N.getNode(), VT, Mask, IsUnary);
22065 // If we have more than 128-bits, only the low 128-bits of shuffle mask
22066 // matter. Check that the upper masks are repeats and remove them.
22067 if (VT.getSizeInBits() > 128) {
22068 int LaneElts = 128 / VT.getScalarSizeInBits();
22070 for (int i = 1, NumLanes = VT.getSizeInBits() / 128; i < NumLanes; ++i)
22071 for (int j = 0; j < LaneElts; ++j)
22072 assert(Mask[j] == Mask[i * LaneElts + j] - (LaneElts * i) &&
22073 "Mask doesn't repeat in high 128-bit lanes!");
22075 Mask.resize(LaneElts);
22078 switch (N.getOpcode()) {
22079 case X86ISD::PSHUFD:
22081 case X86ISD::PSHUFLW:
22084 case X86ISD::PSHUFHW:
22085 Mask.erase(Mask.begin(), Mask.begin() + 4);
22086 for (int &M : Mask)
22090 llvm_unreachable("No valid shuffle instruction found!");
22094 /// \brief Search for a combinable shuffle across a chain ending in pshufd.
22096 /// We walk up the chain and look for a combinable shuffle, skipping over
22097 /// shuffles that we could hoist this shuffle's transformation past without
22098 /// altering anything.
22100 combineRedundantDWordShuffle(SDValue N, MutableArrayRef<int> Mask,
22102 TargetLowering::DAGCombinerInfo &DCI) {
22103 assert(N.getOpcode() == X86ISD::PSHUFD &&
22104 "Called with something other than an x86 128-bit half shuffle!");
22107 // Walk up a single-use chain looking for a combinable shuffle. Keep a stack
22108 // of the shuffles in the chain so that we can form a fresh chain to replace
22110 SmallVector<SDValue, 8> Chain;
22111 SDValue V = N.getOperand(0);
22112 for (; V.hasOneUse(); V = V.getOperand(0)) {
22113 switch (V.getOpcode()) {
22115 return SDValue(); // Nothing combined!
22118 // Skip bitcasts as we always know the type for the target specific
22122 case X86ISD::PSHUFD:
22123 // Found another dword shuffle.
22126 case X86ISD::PSHUFLW:
22127 // Check that the low words (being shuffled) are the identity in the
22128 // dword shuffle, and the high words are self-contained.
22129 if (Mask[0] != 0 || Mask[1] != 1 ||
22130 !(Mask[2] >= 2 && Mask[2] < 4 && Mask[3] >= 2 && Mask[3] < 4))
22133 Chain.push_back(V);
22136 case X86ISD::PSHUFHW:
22137 // Check that the high words (being shuffled) are the identity in the
22138 // dword shuffle, and the low words are self-contained.
22139 if (Mask[2] != 2 || Mask[3] != 3 ||
22140 !(Mask[0] >= 0 && Mask[0] < 2 && Mask[1] >= 0 && Mask[1] < 2))
22143 Chain.push_back(V);
22146 case X86ISD::UNPCKL:
22147 case X86ISD::UNPCKH:
22148 // For either i8 -> i16 or i16 -> i32 unpacks, we can combine a dword
22149 // shuffle into a preceding word shuffle.
22150 if (V.getSimpleValueType().getScalarType() != MVT::i8 &&
22151 V.getSimpleValueType().getScalarType() != MVT::i16)
22154 // Search for a half-shuffle which we can combine with.
22155 unsigned CombineOp =
22156 V.getOpcode() == X86ISD::UNPCKL ? X86ISD::PSHUFLW : X86ISD::PSHUFHW;
22157 if (V.getOperand(0) != V.getOperand(1) ||
22158 !V->isOnlyUserOf(V.getOperand(0).getNode()))
22160 Chain.push_back(V);
22161 V = V.getOperand(0);
22163 switch (V.getOpcode()) {
22165 return SDValue(); // Nothing to combine.
22167 case X86ISD::PSHUFLW:
22168 case X86ISD::PSHUFHW:
22169 if (V.getOpcode() == CombineOp)
22172 Chain.push_back(V);
22176 V = V.getOperand(0);
22180 } while (V.hasOneUse());
22183 // Break out of the loop if we break out of the switch.
22187 if (!V.hasOneUse())
22188 // We fell out of the loop without finding a viable combining instruction.
22191 // Merge this node's mask and our incoming mask.
22192 SmallVector<int, 4> VMask = getPSHUFShuffleMask(V);
22193 for (int &M : Mask)
22195 V = DAG.getNode(V.getOpcode(), DL, V.getValueType(), V.getOperand(0),
22196 getV4X86ShuffleImm8ForMask(Mask, DL, DAG));
22198 // Rebuild the chain around this new shuffle.
22199 while (!Chain.empty()) {
22200 SDValue W = Chain.pop_back_val();
22202 if (V.getValueType() != W.getOperand(0).getValueType())
22203 V = DAG.getBitcast(W.getOperand(0).getValueType(), V);
22205 switch (W.getOpcode()) {
22207 llvm_unreachable("Only PSHUF and UNPCK instructions get here!");
22209 case X86ISD::UNPCKL:
22210 case X86ISD::UNPCKH:
22211 V = DAG.getNode(W.getOpcode(), DL, W.getValueType(), V, V);
22214 case X86ISD::PSHUFD:
22215 case X86ISD::PSHUFLW:
22216 case X86ISD::PSHUFHW:
22217 V = DAG.getNode(W.getOpcode(), DL, W.getValueType(), V, W.getOperand(1));
22221 if (V.getValueType() != N.getValueType())
22222 V = DAG.getBitcast(N.getValueType(), V);
22224 // Return the new chain to replace N.
22228 /// \brief Search for a combinable shuffle across a chain ending in pshuflw or pshufhw.
22230 /// We walk up the chain, skipping shuffles of the other half and looking
22231 /// through shuffles which switch halves trying to find a shuffle of the same
22232 /// pair of dwords.
22233 static bool combineRedundantHalfShuffle(SDValue N, MutableArrayRef<int> Mask,
22235 TargetLowering::DAGCombinerInfo &DCI) {
22237 (N.getOpcode() == X86ISD::PSHUFLW || N.getOpcode() == X86ISD::PSHUFHW) &&
22238 "Called with something other than an x86 128-bit half shuffle!");
22240 unsigned CombineOpcode = N.getOpcode();
22242 // Walk up a single-use chain looking for a combinable shuffle.
22243 SDValue V = N.getOperand(0);
22244 for (; V.hasOneUse(); V = V.getOperand(0)) {
22245 switch (V.getOpcode()) {
22247 return false; // Nothing combined!
22250 // Skip bitcasts as we always know the type for the target specific
22254 case X86ISD::PSHUFLW:
22255 case X86ISD::PSHUFHW:
22256 if (V.getOpcode() == CombineOpcode)
22259 // Other-half shuffles are no-ops.
22262 // Break out of the loop if we break out of the switch.
22266 if (!V.hasOneUse())
22267 // We fell out of the loop without finding a viable combining instruction.
22270 // Combine away the bottom node as its shuffle will be accumulated into
22271 // a preceding shuffle.
22272 DCI.CombineTo(N.getNode(), N.getOperand(0), /*AddTo*/ true);
22274 // Record the old value.
22277 // Merge this node's mask and our incoming mask (adjusted to account for all
22278 // the pshufd instructions encountered).
22279 SmallVector<int, 4> VMask = getPSHUFShuffleMask(V);
22280 for (int &M : Mask)
22282 V = DAG.getNode(V.getOpcode(), DL, MVT::v8i16, V.getOperand(0),
22283 getV4X86ShuffleImm8ForMask(Mask, DL, DAG));
22285 // Check that the shuffles didn't cancel each other out. If not, we need to
22286 // combine to the new one.
22288 // Replace the combinable shuffle with the combined one, updating all users
22289 // so that we re-evaluate the chain here.
22290 DCI.CombineTo(Old.getNode(), V, /*AddTo*/ true);
22295 /// \brief Try to combine x86 target specific shuffles.
22296 static SDValue PerformTargetShuffleCombine(SDValue N, SelectionDAG &DAG,
22297 TargetLowering::DAGCombinerInfo &DCI,
22298 const X86Subtarget *Subtarget) {
22300 MVT VT = N.getSimpleValueType();
22301 SmallVector<int, 4> Mask;
22303 switch (N.getOpcode()) {
22304 case X86ISD::PSHUFD:
22305 case X86ISD::PSHUFLW:
22306 case X86ISD::PSHUFHW:
22307 Mask = getPSHUFShuffleMask(N);
22308 assert(Mask.size() == 4);
22314 // Nuke no-op shuffles that show up after combining.
22315 if (isNoopShuffleMask(Mask))
22316 return DCI.CombineTo(N.getNode(), N.getOperand(0), /*AddTo*/ true);
22318 // Look for simplifications involving one or two shuffle instructions.
22319 SDValue V = N.getOperand(0);
22320 switch (N.getOpcode()) {
22323 case X86ISD::PSHUFLW:
22324 case X86ISD::PSHUFHW:
22325 assert(VT.getScalarType() == MVT::i16 && "Bad word shuffle type!");
22327 if (combineRedundantHalfShuffle(N, Mask, DAG, DCI))
22328 return SDValue(); // We combined away this shuffle, so we're done.
22330 // See if this reduces to a PSHUFD which is no more expensive and can
22331 // combine with more operations. Note that it has to at least flip the
22332 // dwords as otherwise it would have been removed as a no-op.
22333 if (makeArrayRef(Mask).equals({2, 3, 0, 1})) {
22334 int DMask[] = {0, 1, 2, 3};
22335 int DOffset = N.getOpcode() == X86ISD::PSHUFLW ? 0 : 2;
22336 DMask[DOffset + 0] = DOffset + 1;
22337 DMask[DOffset + 1] = DOffset + 0;
22338 MVT DVT = MVT::getVectorVT(MVT::i32, VT.getVectorNumElements() / 2);
22339 V = DAG.getBitcast(DVT, V);
22340 DCI.AddToWorklist(V.getNode());
22341 V = DAG.getNode(X86ISD::PSHUFD, DL, DVT, V,
22342 getV4X86ShuffleImm8ForMask(DMask, DL, DAG));
22343 DCI.AddToWorklist(V.getNode());
22344 return DAG.getBitcast(VT, V);
22347 // Look for shuffle patterns which can be implemented as a single unpack.
22348 // FIXME: This doesn't handle the location of the PSHUFD generically, and
22349 // only works when we have a PSHUFD followed by two half-shuffles.
22350 if (Mask[0] == Mask[1] && Mask[2] == Mask[3] &&
22351 (V.getOpcode() == X86ISD::PSHUFLW ||
22352 V.getOpcode() == X86ISD::PSHUFHW) &&
22353 V.getOpcode() != N.getOpcode() &&
22355 SDValue D = V.getOperand(0);
22356 while (D.getOpcode() == ISD::BITCAST && D.hasOneUse())
22357 D = D.getOperand(0);
22358 if (D.getOpcode() == X86ISD::PSHUFD && D.hasOneUse()) {
22359 SmallVector<int, 4> VMask = getPSHUFShuffleMask(V);
22360 SmallVector<int, 4> DMask = getPSHUFShuffleMask(D);
22361 int NOffset = N.getOpcode() == X86ISD::PSHUFLW ? 0 : 4;
22362 int VOffset = V.getOpcode() == X86ISD::PSHUFLW ? 0 : 4;
22364 for (int i = 0; i < 4; ++i) {
22365 WordMask[i + NOffset] = Mask[i] + NOffset;
22366 WordMask[i + VOffset] = VMask[i] + VOffset;
22368 // Map the word mask through the DWord mask.
22370 for (int i = 0; i < 8; ++i)
22371 MappedMask[i] = 2 * DMask[WordMask[i] / 2] + WordMask[i] % 2;
22372 if (makeArrayRef(MappedMask).equals({0, 0, 1, 1, 2, 2, 3, 3}) ||
22373 makeArrayRef(MappedMask).equals({4, 4, 5, 5, 6, 6, 7, 7})) {
22374 // We can replace all three shuffles with an unpack.
22375 V = DAG.getBitcast(VT, D.getOperand(0));
22376 DCI.AddToWorklist(V.getNode());
22377 return DAG.getNode(MappedMask[0] == 0 ? X86ISD::UNPCKL
22386 case X86ISD::PSHUFD:
22387 if (SDValue NewN = combineRedundantDWordShuffle(N, Mask, DAG, DCI))
22396 /// \brief Try to combine a shuffle into a target-specific add-sub node.
22398 /// We combine this directly on the abstract vector shuffle nodes so it is
22399 /// easier to generically match. We also insert dummy vector shuffle nodes for
22400 /// the operands which explicitly discard the lanes which are unused by this
22401 /// operation to try to flow through the rest of the combiner the fact that
22402 /// they're unused.
22403 static SDValue combineShuffleToAddSub(SDNode *N, SelectionDAG &DAG) {
22405 EVT VT = N->getValueType(0);
22407 // We only handle target-independent shuffles.
22408 // FIXME: It would be easy and harmless to use the target shuffle mask
22409 // extraction tool to support more.
22410 if (N->getOpcode() != ISD::VECTOR_SHUFFLE)
22413 auto *SVN = cast<ShuffleVectorSDNode>(N);
22414 ArrayRef<int> Mask = SVN->getMask();
22415 SDValue V1 = N->getOperand(0);
22416 SDValue V2 = N->getOperand(1);
22418 // We require the first shuffle operand to be the SUB node, and the second to
22419 // be the ADD node.
22420 // FIXME: We should support the commuted patterns.
22421 if (V1->getOpcode() != ISD::FSUB || V2->getOpcode() != ISD::FADD)
22424 // If there are other uses of these operations we can't fold them.
22425 if (!V1->hasOneUse() || !V2->hasOneUse())
22428 // Ensure that both operations have the same operands. Note that we can
22429 // commute the FADD operands.
22430 SDValue LHS = V1->getOperand(0), RHS = V1->getOperand(1);
22431 if ((V2->getOperand(0) != LHS || V2->getOperand(1) != RHS) &&
22432 (V2->getOperand(0) != RHS || V2->getOperand(1) != LHS))
22435 // We're looking for blends between FADD and FSUB nodes. We insist on these
22436 // nodes being lined up in a specific expected pattern.
22437 if (!(isShuffleEquivalent(V1, V2, Mask, {0, 3}) ||
22438 isShuffleEquivalent(V1, V2, Mask, {0, 5, 2, 7}) ||
22439 isShuffleEquivalent(V1, V2, Mask, {0, 9, 2, 11, 4, 13, 6, 15})))
22442 // Only specific types are legal at this point, assert so we notice if and
22443 // when these change.
22444 assert((VT == MVT::v4f32 || VT == MVT::v2f64 || VT == MVT::v8f32 ||
22445 VT == MVT::v4f64) &&
22446 "Unknown vector type encountered!");
22448 return DAG.getNode(X86ISD::ADDSUB, DL, VT, LHS, RHS);
22451 /// PerformShuffleCombine - Performs several different shuffle combines.
22452 static SDValue PerformShuffleCombine(SDNode *N, SelectionDAG &DAG,
22453 TargetLowering::DAGCombinerInfo &DCI,
22454 const X86Subtarget *Subtarget) {
22456 SDValue N0 = N->getOperand(0);
22457 SDValue N1 = N->getOperand(1);
22458 EVT VT = N->getValueType(0);
22460 // Don't create instructions with illegal types after legalize types has run.
22461 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
22462 if (!DCI.isBeforeLegalize() && !TLI.isTypeLegal(VT.getVectorElementType()))
22465 // If we have legalized the vector types, look for blends of FADD and FSUB
22466 // nodes that we can fuse into an ADDSUB node.
22467 if (TLI.isTypeLegal(VT) && Subtarget->hasSSE3())
22468 if (SDValue AddSub = combineShuffleToAddSub(N, DAG))
22471 // Combine 256-bit vector shuffles. This is only profitable when in AVX mode
22472 if (Subtarget->hasFp256() && VT.is256BitVector() &&
22473 N->getOpcode() == ISD::VECTOR_SHUFFLE)
22474 return PerformShuffleCombine256(N, DAG, DCI, Subtarget);
22476 // During Type Legalization, when promoting illegal vector types,
22477 // the backend might introduce new shuffle dag nodes and bitcasts.
22479 // This code performs the following transformation:
22480 // fold: (shuffle (bitcast (BINOP A, B)), Undef, <Mask>) ->
22481 // (shuffle (BINOP (bitcast A), (bitcast B)), Undef, <Mask>)
22483 // We do this only if both the bitcast and the BINOP dag nodes have
22484 // one use. Also, perform this transformation only if the new binary
22485 // operation is legal. This is to avoid introducing dag nodes that
22486 // potentially need to be further expanded (or custom lowered) into a
22487 // less optimal sequence of dag nodes.
22488 if (!DCI.isBeforeLegalize() && DCI.isBeforeLegalizeOps() &&
22489 N1.getOpcode() == ISD::UNDEF && N0.hasOneUse() &&
22490 N0.getOpcode() == ISD::BITCAST) {
22491 SDValue BC0 = N0.getOperand(0);
22492 EVT SVT = BC0.getValueType();
22493 unsigned Opcode = BC0.getOpcode();
22494 unsigned NumElts = VT.getVectorNumElements();
22496 if (BC0.hasOneUse() && SVT.isVector() &&
22497 SVT.getVectorNumElements() * 2 == NumElts &&
22498 TLI.isOperationLegal(Opcode, VT)) {
22499 bool CanFold = false;
22511 unsigned SVTNumElts = SVT.getVectorNumElements();
22512 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
22513 for (unsigned i = 0, e = SVTNumElts; i != e && CanFold; ++i)
22514 CanFold = SVOp->getMaskElt(i) == (int)(i * 2);
22515 for (unsigned i = SVTNumElts, e = NumElts; i != e && CanFold; ++i)
22516 CanFold = SVOp->getMaskElt(i) < 0;
22519 SDValue BC00 = DAG.getBitcast(VT, BC0.getOperand(0));
22520 SDValue BC01 = DAG.getBitcast(VT, BC0.getOperand(1));
22521 SDValue NewBinOp = DAG.getNode(BC0.getOpcode(), dl, VT, BC00, BC01);
22522 return DAG.getVectorShuffle(VT, dl, NewBinOp, N1, &SVOp->getMask()[0]);
22527 // Combine a vector_shuffle that is equal to build_vector load1, load2, load3,
22528 // load4, <0, 1, 2, 3> into a 128-bit load if the load addresses are
22529 // consecutive, non-overlapping, and in the right order.
22530 SmallVector<SDValue, 16> Elts;
22531 for (unsigned i = 0, e = VT.getVectorNumElements(); i != e; ++i)
22532 Elts.push_back(getShuffleScalarElt(N, i, DAG, 0));
22534 if (SDValue LD = EltsFromConsecutiveLoads(VT, Elts, dl, DAG, true))
22537 if (isTargetShuffle(N->getOpcode())) {
22539 PerformTargetShuffleCombine(SDValue(N, 0), DAG, DCI, Subtarget);
22540 if (Shuffle.getNode())
22543 // Try recursively combining arbitrary sequences of x86 shuffle
22544 // instructions into higher-order shuffles. We do this after combining
22545 // specific PSHUF instruction sequences into their minimal form so that we
22546 // can evaluate how many specialized shuffle instructions are involved in
22547 // a particular chain.
22548 SmallVector<int, 1> NonceMask; // Just a placeholder.
22549 NonceMask.push_back(0);
22550 if (combineX86ShufflesRecursively(SDValue(N, 0), SDValue(N, 0), NonceMask,
22551 /*Depth*/ 1, /*HasPSHUFB*/ false, DAG,
22553 return SDValue(); // This routine will use CombineTo to replace N.
22559 /// XFormVExtractWithShuffleIntoLoad - Check if a vector extract from a target
22560 /// specific shuffle of a load can be folded into a single element load.
22561 /// Similar handling for VECTOR_SHUFFLE is performed by DAGCombiner, but
22562 /// shuffles have been custom lowered so we need to handle those here.
22563 static SDValue XFormVExtractWithShuffleIntoLoad(SDNode *N, SelectionDAG &DAG,
22564 TargetLowering::DAGCombinerInfo &DCI) {
22565 if (DCI.isBeforeLegalizeOps())
22568 SDValue InVec = N->getOperand(0);
22569 SDValue EltNo = N->getOperand(1);
22571 if (!isa<ConstantSDNode>(EltNo))
22574 EVT OriginalVT = InVec.getValueType();
22576 if (InVec.getOpcode() == ISD::BITCAST) {
22577 // Don't duplicate a load with other uses.
22578 if (!InVec.hasOneUse())
22580 EVT BCVT = InVec.getOperand(0).getValueType();
22581 if (!BCVT.isVector() ||
22582 BCVT.getVectorNumElements() != OriginalVT.getVectorNumElements())
22584 InVec = InVec.getOperand(0);
22587 EVT CurrentVT = InVec.getValueType();
22589 if (!isTargetShuffle(InVec.getOpcode()))
22592 // Don't duplicate a load with other uses.
22593 if (!InVec.hasOneUse())
22596 SmallVector<int, 16> ShuffleMask;
22598 if (!getTargetShuffleMask(InVec.getNode(), CurrentVT.getSimpleVT(),
22599 ShuffleMask, UnaryShuffle))
22602 // Select the input vector, guarding against out of range extract vector.
22603 unsigned NumElems = CurrentVT.getVectorNumElements();
22604 int Elt = cast<ConstantSDNode>(EltNo)->getZExtValue();
22605 int Idx = (Elt > (int)NumElems) ? -1 : ShuffleMask[Elt];
22606 SDValue LdNode = (Idx < (int)NumElems) ? InVec.getOperand(0)
22607 : InVec.getOperand(1);
22609 // If inputs to shuffle are the same for both ops, then allow 2 uses
22610 unsigned AllowedUses = InVec.getNumOperands() > 1 &&
22611 InVec.getOperand(0) == InVec.getOperand(1) ? 2 : 1;
22613 if (LdNode.getOpcode() == ISD::BITCAST) {
22614 // Don't duplicate a load with other uses.
22615 if (!LdNode.getNode()->hasNUsesOfValue(AllowedUses, 0))
22618 AllowedUses = 1; // only allow 1 load use if we have a bitcast
22619 LdNode = LdNode.getOperand(0);
22622 if (!ISD::isNormalLoad(LdNode.getNode()))
22625 LoadSDNode *LN0 = cast<LoadSDNode>(LdNode);
22627 if (!LN0 ||!LN0->hasNUsesOfValue(AllowedUses, 0) || LN0->isVolatile())
22630 EVT EltVT = N->getValueType(0);
22631 // If there's a bitcast before the shuffle, check if the load type and
22632 // alignment is valid.
22633 unsigned Align = LN0->getAlignment();
22634 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
22635 unsigned NewAlign = DAG.getDataLayout().getABITypeAlignment(
22636 EltVT.getTypeForEVT(*DAG.getContext()));
22638 if (NewAlign > Align || !TLI.isOperationLegalOrCustom(ISD::LOAD, EltVT))
22641 // All checks match so transform back to vector_shuffle so that DAG combiner
22642 // can finish the job
22645 // Create shuffle node taking into account the case that its a unary shuffle
22646 SDValue Shuffle = (UnaryShuffle) ? DAG.getUNDEF(CurrentVT)
22647 : InVec.getOperand(1);
22648 Shuffle = DAG.getVectorShuffle(CurrentVT, dl,
22649 InVec.getOperand(0), Shuffle,
22651 Shuffle = DAG.getBitcast(OriginalVT, Shuffle);
22652 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, N->getValueType(0), Shuffle,
22656 /// \brief Detect bitcasts between i32 to x86mmx low word. Since MMX types are
22657 /// special and don't usually play with other vector types, it's better to
22658 /// handle them early to be sure we emit efficient code by avoiding
22659 /// store-load conversions.
22660 static SDValue PerformBITCASTCombine(SDNode *N, SelectionDAG &DAG) {
22661 if (N->getValueType(0) != MVT::x86mmx ||
22662 N->getOperand(0)->getOpcode() != ISD::BUILD_VECTOR ||
22663 N->getOperand(0)->getValueType(0) != MVT::v2i32)
22666 SDValue V = N->getOperand(0);
22667 ConstantSDNode *C = dyn_cast<ConstantSDNode>(V.getOperand(1));
22668 if (C && C->getZExtValue() == 0 && V.getOperand(0).getValueType() == MVT::i32)
22669 return DAG.getNode(X86ISD::MMX_MOVW2D, SDLoc(V.getOperand(0)),
22670 N->getValueType(0), V.getOperand(0));
22675 /// PerformEXTRACT_VECTOR_ELTCombine - Detect vector gather/scatter index
22676 /// generation and convert it from being a bunch of shuffles and extracts
22677 /// into a somewhat faster sequence. For i686, the best sequence is apparently
22678 /// storing the value and loading scalars back, while for x64 we should
22679 /// use 64-bit extracts and shifts.
22680 static SDValue PerformEXTRACT_VECTOR_ELTCombine(SDNode *N, SelectionDAG &DAG,
22681 TargetLowering::DAGCombinerInfo &DCI) {
22682 if (SDValue NewOp = XFormVExtractWithShuffleIntoLoad(N, DAG, DCI))
22685 SDValue InputVector = N->getOperand(0);
22686 SDLoc dl(InputVector);
22687 // Detect mmx to i32 conversion through a v2i32 elt extract.
22688 if (InputVector.getOpcode() == ISD::BITCAST && InputVector.hasOneUse() &&
22689 N->getValueType(0) == MVT::i32 &&
22690 InputVector.getValueType() == MVT::v2i32) {
22692 // The bitcast source is a direct mmx result.
22693 SDValue MMXSrc = InputVector.getNode()->getOperand(0);
22694 if (MMXSrc.getValueType() == MVT::x86mmx)
22695 return DAG.getNode(X86ISD::MMX_MOVD2W, SDLoc(InputVector),
22696 N->getValueType(0),
22697 InputVector.getNode()->getOperand(0));
22699 // The mmx is indirect: (i64 extract_elt (v1i64 bitcast (x86mmx ...))).
22700 SDValue MMXSrcOp = MMXSrc.getOperand(0);
22701 if (MMXSrc.getOpcode() == ISD::EXTRACT_VECTOR_ELT && MMXSrc.hasOneUse() &&
22702 MMXSrc.getValueType() == MVT::i64 && MMXSrcOp.hasOneUse() &&
22703 MMXSrcOp.getOpcode() == ISD::BITCAST &&
22704 MMXSrcOp.getValueType() == MVT::v1i64 &&
22705 MMXSrcOp.getOperand(0).getValueType() == MVT::x86mmx)
22706 return DAG.getNode(X86ISD::MMX_MOVD2W, SDLoc(InputVector),
22707 N->getValueType(0),
22708 MMXSrcOp.getOperand(0));
22711 EVT VT = N->getValueType(0);
22713 if (VT == MVT::i1 && dyn_cast<ConstantSDNode>(N->getOperand(1)) &&
22714 InputVector.getOpcode() == ISD::BITCAST &&
22715 dyn_cast<ConstantSDNode>(InputVector.getOperand(0))) {
22716 uint64_t ExtractedElt =
22717 cast<ConstantSDNode>(N->getOperand(1))->getZExtValue();
22718 uint64_t InputValue =
22719 cast<ConstantSDNode>(InputVector.getOperand(0))->getZExtValue();
22720 uint64_t Res = (InputValue >> ExtractedElt) & 1;
22721 return DAG.getConstant(Res, dl, MVT::i1);
22723 // Only operate on vectors of 4 elements, where the alternative shuffling
22724 // gets to be more expensive.
22725 if (InputVector.getValueType() != MVT::v4i32)
22728 // Check whether every use of InputVector is an EXTRACT_VECTOR_ELT with a
22729 // single use which is a sign-extend or zero-extend, and all elements are
22731 SmallVector<SDNode *, 4> Uses;
22732 unsigned ExtractedElements = 0;
22733 for (SDNode::use_iterator UI = InputVector.getNode()->use_begin(),
22734 UE = InputVector.getNode()->use_end(); UI != UE; ++UI) {
22735 if (UI.getUse().getResNo() != InputVector.getResNo())
22738 SDNode *Extract = *UI;
22739 if (Extract->getOpcode() != ISD::EXTRACT_VECTOR_ELT)
22742 if (Extract->getValueType(0) != MVT::i32)
22744 if (!Extract->hasOneUse())
22746 if (Extract->use_begin()->getOpcode() != ISD::SIGN_EXTEND &&
22747 Extract->use_begin()->getOpcode() != ISD::ZERO_EXTEND)
22749 if (!isa<ConstantSDNode>(Extract->getOperand(1)))
22752 // Record which element was extracted.
22753 ExtractedElements |=
22754 1 << cast<ConstantSDNode>(Extract->getOperand(1))->getZExtValue();
22756 Uses.push_back(Extract);
22759 // If not all the elements were used, this may not be worthwhile.
22760 if (ExtractedElements != 15)
22763 // Ok, we've now decided to do the transformation.
22764 // If 64-bit shifts are legal, use the extract-shift sequence,
22765 // otherwise bounce the vector off the cache.
22766 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
22769 if (TLI.isOperationLegal(ISD::SRA, MVT::i64)) {
22770 SDValue Cst = DAG.getBitcast(MVT::v2i64, InputVector);
22771 auto &DL = DAG.getDataLayout();
22772 EVT VecIdxTy = DAG.getTargetLoweringInfo().getVectorIdxTy(DL);
22773 SDValue BottomHalf = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i64, Cst,
22774 DAG.getConstant(0, dl, VecIdxTy));
22775 SDValue TopHalf = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i64, Cst,
22776 DAG.getConstant(1, dl, VecIdxTy));
22778 SDValue ShAmt = DAG.getConstant(
22779 32, dl, DAG.getTargetLoweringInfo().getShiftAmountTy(MVT::i64, DL));
22780 Vals[0] = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, BottomHalf);
22781 Vals[1] = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32,
22782 DAG.getNode(ISD::SRA, dl, MVT::i64, BottomHalf, ShAmt));
22783 Vals[2] = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, TopHalf);
22784 Vals[3] = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32,
22785 DAG.getNode(ISD::SRA, dl, MVT::i64, TopHalf, ShAmt));
22787 // Store the value to a temporary stack slot.
22788 SDValue StackPtr = DAG.CreateStackTemporary(InputVector.getValueType());
22789 SDValue Ch = DAG.getStore(DAG.getEntryNode(), dl, InputVector, StackPtr,
22790 MachinePointerInfo(), false, false, 0);
22792 EVT ElementType = InputVector.getValueType().getVectorElementType();
22793 unsigned EltSize = ElementType.getSizeInBits() / 8;
22795 // Replace each use (extract) with a load of the appropriate element.
22796 for (unsigned i = 0; i < 4; ++i) {
22797 uint64_t Offset = EltSize * i;
22798 auto PtrVT = TLI.getPointerTy(DAG.getDataLayout());
22799 SDValue OffsetVal = DAG.getConstant(Offset, dl, PtrVT);
22801 SDValue ScalarAddr =
22802 DAG.getNode(ISD::ADD, dl, PtrVT, StackPtr, OffsetVal);
22804 // Load the scalar.
22805 Vals[i] = DAG.getLoad(ElementType, dl, Ch,
22806 ScalarAddr, MachinePointerInfo(),
22807 false, false, false, 0);
22812 // Replace the extracts
22813 for (SmallVectorImpl<SDNode *>::iterator UI = Uses.begin(),
22814 UE = Uses.end(); UI != UE; ++UI) {
22815 SDNode *Extract = *UI;
22817 SDValue Idx = Extract->getOperand(1);
22818 uint64_t IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
22819 DAG.ReplaceAllUsesOfValueWith(SDValue(Extract, 0), Vals[IdxVal]);
22822 // The replacement was made in place; don't return anything.
22827 transformVSELECTtoBlendVECTOR_SHUFFLE(SDNode *N, SelectionDAG &DAG,
22828 const X86Subtarget *Subtarget) {
22830 SDValue Cond = N->getOperand(0);
22831 SDValue LHS = N->getOperand(1);
22832 SDValue RHS = N->getOperand(2);
22834 if (Cond.getOpcode() == ISD::SIGN_EXTEND) {
22835 SDValue CondSrc = Cond->getOperand(0);
22836 if (CondSrc->getOpcode() == ISD::SIGN_EXTEND_INREG)
22837 Cond = CondSrc->getOperand(0);
22840 if (!ISD::isBuildVectorOfConstantSDNodes(Cond.getNode()))
22843 // A vselect where all conditions and data are constants can be optimized into
22844 // a single vector load by SelectionDAGLegalize::ExpandBUILD_VECTOR().
22845 if (ISD::isBuildVectorOfConstantSDNodes(LHS.getNode()) &&
22846 ISD::isBuildVectorOfConstantSDNodes(RHS.getNode()))
22849 unsigned MaskValue = 0;
22850 if (!BUILD_VECTORtoBlendMask(cast<BuildVectorSDNode>(Cond), MaskValue))
22853 MVT VT = N->getSimpleValueType(0);
22854 unsigned NumElems = VT.getVectorNumElements();
22855 SmallVector<int, 8> ShuffleMask(NumElems, -1);
22856 for (unsigned i = 0; i < NumElems; ++i) {
22857 // Be sure we emit undef where we can.
22858 if (Cond.getOperand(i)->getOpcode() == ISD::UNDEF)
22859 ShuffleMask[i] = -1;
22861 ShuffleMask[i] = i + NumElems * ((MaskValue >> i) & 1);
22864 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
22865 if (!TLI.isShuffleMaskLegal(ShuffleMask, VT))
22867 return DAG.getVectorShuffle(VT, dl, LHS, RHS, &ShuffleMask[0]);
22870 /// PerformSELECTCombine - Do target-specific dag combines on SELECT and VSELECT
22872 static SDValue PerformSELECTCombine(SDNode *N, SelectionDAG &DAG,
22873 TargetLowering::DAGCombinerInfo &DCI,
22874 const X86Subtarget *Subtarget) {
22876 SDValue Cond = N->getOperand(0);
22877 // Get the LHS/RHS of the select.
22878 SDValue LHS = N->getOperand(1);
22879 SDValue RHS = N->getOperand(2);
22880 EVT VT = LHS.getValueType();
22881 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
22883 // If we have SSE[12] support, try to form min/max nodes. SSE min/max
22884 // instructions match the semantics of the common C idiom x<y?x:y but not
22885 // x<=y?x:y, because of how they handle negative zero (which can be
22886 // ignored in unsafe-math mode).
22887 // We also try to create v2f32 min/max nodes, which we later widen to v4f32.
22888 if (Cond.getOpcode() == ISD::SETCC && VT.isFloatingPoint() &&
22889 VT != MVT::f80 && (TLI.isTypeLegal(VT) || VT == MVT::v2f32) &&
22890 (Subtarget->hasSSE2() ||
22891 (Subtarget->hasSSE1() && VT.getScalarType() == MVT::f32))) {
22892 ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
22894 unsigned Opcode = 0;
22895 // Check for x CC y ? x : y.
22896 if (DAG.isEqualTo(LHS, Cond.getOperand(0)) &&
22897 DAG.isEqualTo(RHS, Cond.getOperand(1))) {
22901 // Converting this to a min would handle NaNs incorrectly, and swapping
22902 // the operands would cause it to handle comparisons between positive
22903 // and negative zero incorrectly.
22904 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)) {
22905 if (!DAG.getTarget().Options.UnsafeFPMath &&
22906 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
22908 std::swap(LHS, RHS);
22910 Opcode = X86ISD::FMIN;
22913 // Converting this to a min would handle comparisons between positive
22914 // and negative zero incorrectly.
22915 if (!DAG.getTarget().Options.UnsafeFPMath &&
22916 !DAG.isKnownNeverZero(LHS) && !DAG.isKnownNeverZero(RHS))
22918 Opcode = X86ISD::FMIN;
22921 // Converting this to a min would handle both negative zeros and NaNs
22922 // incorrectly, but we can swap the operands to fix both.
22923 std::swap(LHS, RHS);
22927 Opcode = X86ISD::FMIN;
22931 // Converting this to a max would handle comparisons between positive
22932 // and negative zero incorrectly.
22933 if (!DAG.getTarget().Options.UnsafeFPMath &&
22934 !DAG.isKnownNeverZero(LHS) && !DAG.isKnownNeverZero(RHS))
22936 Opcode = X86ISD::FMAX;
22939 // Converting this to a max would handle NaNs incorrectly, and swapping
22940 // the operands would cause it to handle comparisons between positive
22941 // and negative zero incorrectly.
22942 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)) {
22943 if (!DAG.getTarget().Options.UnsafeFPMath &&
22944 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
22946 std::swap(LHS, RHS);
22948 Opcode = X86ISD::FMAX;
22951 // Converting this to a max would handle both negative zeros and NaNs
22952 // incorrectly, but we can swap the operands to fix both.
22953 std::swap(LHS, RHS);
22957 Opcode = X86ISD::FMAX;
22960 // Check for x CC y ? y : x -- a min/max with reversed arms.
22961 } else if (DAG.isEqualTo(LHS, Cond.getOperand(1)) &&
22962 DAG.isEqualTo(RHS, Cond.getOperand(0))) {
22966 // Converting this to a min would handle comparisons between positive
22967 // and negative zero incorrectly, and swapping the operands would
22968 // cause it to handle NaNs incorrectly.
22969 if (!DAG.getTarget().Options.UnsafeFPMath &&
22970 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS))) {
22971 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS))
22973 std::swap(LHS, RHS);
22975 Opcode = X86ISD::FMIN;
22978 // Converting this to a min would handle NaNs incorrectly.
22979 if (!DAG.getTarget().Options.UnsafeFPMath &&
22980 (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)))
22982 Opcode = X86ISD::FMIN;
22985 // Converting this to a min would handle both negative zeros and NaNs
22986 // incorrectly, but we can swap the operands to fix both.
22987 std::swap(LHS, RHS);
22991 Opcode = X86ISD::FMIN;
22995 // Converting this to a max would handle NaNs incorrectly.
22996 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS))
22998 Opcode = X86ISD::FMAX;
23001 // Converting this to a max would handle comparisons between positive
23002 // and negative zero incorrectly, and swapping the operands would
23003 // cause it to handle NaNs incorrectly.
23004 if (!DAG.getTarget().Options.UnsafeFPMath &&
23005 !DAG.isKnownNeverZero(LHS) && !DAG.isKnownNeverZero(RHS)) {
23006 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS))
23008 std::swap(LHS, RHS);
23010 Opcode = X86ISD::FMAX;
23013 // Converting this to a max would handle both negative zeros and NaNs
23014 // incorrectly, but we can swap the operands to fix both.
23015 std::swap(LHS, RHS);
23019 Opcode = X86ISD::FMAX;
23025 return DAG.getNode(Opcode, DL, N->getValueType(0), LHS, RHS);
23028 EVT CondVT = Cond.getValueType();
23029 if (Subtarget->hasAVX512() && VT.isVector() && CondVT.isVector() &&
23030 CondVT.getVectorElementType() == MVT::i1) {
23031 // v16i8 (select v16i1, v16i8, v16i8) does not have a proper
23032 // lowering on KNL. In this case we convert it to
23033 // v16i8 (select v16i8, v16i8, v16i8) and use AVX instruction.
23034 // The same situation for all 128 and 256-bit vectors of i8 and i16.
23035 // Since SKX these selects have a proper lowering.
23036 EVT OpVT = LHS.getValueType();
23037 if ((OpVT.is128BitVector() || OpVT.is256BitVector()) &&
23038 (OpVT.getVectorElementType() == MVT::i8 ||
23039 OpVT.getVectorElementType() == MVT::i16) &&
23040 !(Subtarget->hasBWI() && Subtarget->hasVLX())) {
23041 Cond = DAG.getNode(ISD::SIGN_EXTEND, DL, OpVT, Cond);
23042 DCI.AddToWorklist(Cond.getNode());
23043 return DAG.getNode(N->getOpcode(), DL, OpVT, Cond, LHS, RHS);
23046 // If this is a select between two integer constants, try to do some
23048 if (ConstantSDNode *TrueC = dyn_cast<ConstantSDNode>(LHS)) {
23049 if (ConstantSDNode *FalseC = dyn_cast<ConstantSDNode>(RHS))
23050 // Don't do this for crazy integer types.
23051 if (DAG.getTargetLoweringInfo().isTypeLegal(LHS.getValueType())) {
23052 // If this is efficiently invertible, canonicalize the LHSC/RHSC values
23053 // so that TrueC (the true value) is larger than FalseC.
23054 bool NeedsCondInvert = false;
23056 if (TrueC->getAPIntValue().ult(FalseC->getAPIntValue()) &&
23057 // Efficiently invertible.
23058 (Cond.getOpcode() == ISD::SETCC || // setcc -> invertible.
23059 (Cond.getOpcode() == ISD::XOR && // xor(X, C) -> invertible.
23060 isa<ConstantSDNode>(Cond.getOperand(1))))) {
23061 NeedsCondInvert = true;
23062 std::swap(TrueC, FalseC);
23065 // Optimize C ? 8 : 0 -> zext(C) << 3. Likewise for any pow2/0.
23066 if (FalseC->getAPIntValue() == 0 &&
23067 TrueC->getAPIntValue().isPowerOf2()) {
23068 if (NeedsCondInvert) // Invert the condition if needed.
23069 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
23070 DAG.getConstant(1, DL, Cond.getValueType()));
23072 // Zero extend the condition if needed.
23073 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, LHS.getValueType(), Cond);
23075 unsigned ShAmt = TrueC->getAPIntValue().logBase2();
23076 return DAG.getNode(ISD::SHL, DL, LHS.getValueType(), Cond,
23077 DAG.getConstant(ShAmt, DL, MVT::i8));
23080 // Optimize Cond ? cst+1 : cst -> zext(setcc(C)+cst.
23081 if (FalseC->getAPIntValue()+1 == TrueC->getAPIntValue()) {
23082 if (NeedsCondInvert) // Invert the condition if needed.
23083 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
23084 DAG.getConstant(1, DL, Cond.getValueType()));
23086 // Zero extend the condition if needed.
23087 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL,
23088 FalseC->getValueType(0), Cond);
23089 return DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
23090 SDValue(FalseC, 0));
23093 // Optimize cases that will turn into an LEA instruction. This requires
23094 // an i32 or i64 and an efficient multiplier (1, 2, 3, 4, 5, 8, 9).
23095 if (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i64) {
23096 uint64_t Diff = TrueC->getZExtValue()-FalseC->getZExtValue();
23097 if (N->getValueType(0) == MVT::i32) Diff = (unsigned)Diff;
23099 bool isFastMultiplier = false;
23101 switch ((unsigned char)Diff) {
23103 case 1: // result = add base, cond
23104 case 2: // result = lea base( , cond*2)
23105 case 3: // result = lea base(cond, cond*2)
23106 case 4: // result = lea base( , cond*4)
23107 case 5: // result = lea base(cond, cond*4)
23108 case 8: // result = lea base( , cond*8)
23109 case 9: // result = lea base(cond, cond*8)
23110 isFastMultiplier = true;
23115 if (isFastMultiplier) {
23116 APInt Diff = TrueC->getAPIntValue()-FalseC->getAPIntValue();
23117 if (NeedsCondInvert) // Invert the condition if needed.
23118 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
23119 DAG.getConstant(1, DL, Cond.getValueType()));
23121 // Zero extend the condition if needed.
23122 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, FalseC->getValueType(0),
23124 // Scale the condition by the difference.
23126 Cond = DAG.getNode(ISD::MUL, DL, Cond.getValueType(), Cond,
23127 DAG.getConstant(Diff, DL,
23128 Cond.getValueType()));
23130 // Add the base if non-zero.
23131 if (FalseC->getAPIntValue() != 0)
23132 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
23133 SDValue(FalseC, 0));
23140 // Canonicalize max and min:
23141 // (x > y) ? x : y -> (x >= y) ? x : y
23142 // (x < y) ? x : y -> (x <= y) ? x : y
23143 // This allows use of COND_S / COND_NS (see TranslateX86CC) which eliminates
23144 // the need for an extra compare
23145 // against zero. e.g.
23146 // (x - y) > 0 : (x - y) ? 0 -> (x - y) >= 0 : (x - y) ? 0
23148 // testl %edi, %edi
23150 // cmovgl %edi, %eax
23154 // cmovsl %eax, %edi
23155 if (N->getOpcode() == ISD::SELECT && Cond.getOpcode() == ISD::SETCC &&
23156 DAG.isEqualTo(LHS, Cond.getOperand(0)) &&
23157 DAG.isEqualTo(RHS, Cond.getOperand(1))) {
23158 ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
23163 ISD::CondCode NewCC = (CC == ISD::SETLT) ? ISD::SETLE : ISD::SETGE;
23164 Cond = DAG.getSetCC(SDLoc(Cond), Cond.getValueType(),
23165 Cond.getOperand(0), Cond.getOperand(1), NewCC);
23166 return DAG.getNode(ISD::SELECT, DL, VT, Cond, LHS, RHS);
23171 // Early exit check
23172 if (!TLI.isTypeLegal(VT))
23175 // Match VSELECTs into subs with unsigned saturation.
23176 if (N->getOpcode() == ISD::VSELECT && Cond.getOpcode() == ISD::SETCC &&
23177 // psubus is available in SSE2 and AVX2 for i8 and i16 vectors.
23178 ((Subtarget->hasSSE2() && (VT == MVT::v16i8 || VT == MVT::v8i16)) ||
23179 (Subtarget->hasAVX2() && (VT == MVT::v32i8 || VT == MVT::v16i16)))) {
23180 ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
23182 // Check if one of the arms of the VSELECT is a zero vector. If it's on the
23183 // left side invert the predicate to simplify logic below.
23185 if (ISD::isBuildVectorAllZeros(LHS.getNode())) {
23187 CC = ISD::getSetCCInverse(CC, true);
23188 } else if (ISD::isBuildVectorAllZeros(RHS.getNode())) {
23192 if (Other.getNode() && Other->getNumOperands() == 2 &&
23193 DAG.isEqualTo(Other->getOperand(0), Cond.getOperand(0))) {
23194 SDValue OpLHS = Other->getOperand(0), OpRHS = Other->getOperand(1);
23195 SDValue CondRHS = Cond->getOperand(1);
23197 // Look for a general sub with unsigned saturation first.
23198 // x >= y ? x-y : 0 --> subus x, y
23199 // x > y ? x-y : 0 --> subus x, y
23200 if ((CC == ISD::SETUGE || CC == ISD::SETUGT) &&
23201 Other->getOpcode() == ISD::SUB && DAG.isEqualTo(OpRHS, CondRHS))
23202 return DAG.getNode(X86ISD::SUBUS, DL, VT, OpLHS, OpRHS);
23204 if (auto *OpRHSBV = dyn_cast<BuildVectorSDNode>(OpRHS))
23205 if (auto *OpRHSConst = OpRHSBV->getConstantSplatNode()) {
23206 if (auto *CondRHSBV = dyn_cast<BuildVectorSDNode>(CondRHS))
23207 if (auto *CondRHSConst = CondRHSBV->getConstantSplatNode())
23208 // If the RHS is a constant we have to reverse the const
23209 // canonicalization.
23210 // x > C-1 ? x+-C : 0 --> subus x, C
23211 if (CC == ISD::SETUGT && Other->getOpcode() == ISD::ADD &&
23212 CondRHSConst->getAPIntValue() ==
23213 (-OpRHSConst->getAPIntValue() - 1))
23214 return DAG.getNode(
23215 X86ISD::SUBUS, DL, VT, OpLHS,
23216 DAG.getConstant(-OpRHSConst->getAPIntValue(), DL, VT));
23218 // Another special case: If C was a sign bit, the sub has been
23219 // canonicalized into a xor.
23220 // FIXME: Would it be better to use computeKnownBits to determine
23221 // whether it's safe to decanonicalize the xor?
23222 // x s< 0 ? x^C : 0 --> subus x, C
23223 if (CC == ISD::SETLT && Other->getOpcode() == ISD::XOR &&
23224 ISD::isBuildVectorAllZeros(CondRHS.getNode()) &&
23225 OpRHSConst->getAPIntValue().isSignBit())
23226 // Note that we have to rebuild the RHS constant here to ensure we
23227 // don't rely on particular values of undef lanes.
23228 return DAG.getNode(
23229 X86ISD::SUBUS, DL, VT, OpLHS,
23230 DAG.getConstant(OpRHSConst->getAPIntValue(), DL, VT));
23235 // Simplify vector selection if condition value type matches vselect
23237 if (N->getOpcode() == ISD::VSELECT && CondVT == VT) {
23238 assert(Cond.getValueType().isVector() &&
23239 "vector select expects a vector selector!");
23241 bool TValIsAllOnes = ISD::isBuildVectorAllOnes(LHS.getNode());
23242 bool FValIsAllZeros = ISD::isBuildVectorAllZeros(RHS.getNode());
23244 // Try invert the condition if true value is not all 1s and false value
23246 if (!TValIsAllOnes && !FValIsAllZeros &&
23247 // Check if the selector will be produced by CMPP*/PCMP*
23248 Cond.getOpcode() == ISD::SETCC &&
23249 // Check if SETCC has already been promoted
23250 TLI.getSetCCResultType(DAG.getDataLayout(), *DAG.getContext(), VT) ==
23252 bool TValIsAllZeros = ISD::isBuildVectorAllZeros(LHS.getNode());
23253 bool FValIsAllOnes = ISD::isBuildVectorAllOnes(RHS.getNode());
23255 if (TValIsAllZeros || FValIsAllOnes) {
23256 SDValue CC = Cond.getOperand(2);
23257 ISD::CondCode NewCC =
23258 ISD::getSetCCInverse(cast<CondCodeSDNode>(CC)->get(),
23259 Cond.getOperand(0).getValueType().isInteger());
23260 Cond = DAG.getSetCC(DL, CondVT, Cond.getOperand(0), Cond.getOperand(1), NewCC);
23261 std::swap(LHS, RHS);
23262 TValIsAllOnes = FValIsAllOnes;
23263 FValIsAllZeros = TValIsAllZeros;
23267 if (TValIsAllOnes || FValIsAllZeros) {
23270 if (TValIsAllOnes && FValIsAllZeros)
23272 else if (TValIsAllOnes)
23274 DAG.getNode(ISD::OR, DL, CondVT, Cond, DAG.getBitcast(CondVT, RHS));
23275 else if (FValIsAllZeros)
23276 Ret = DAG.getNode(ISD::AND, DL, CondVT, Cond,
23277 DAG.getBitcast(CondVT, LHS));
23279 return DAG.getBitcast(VT, Ret);
23283 // We should generate an X86ISD::BLENDI from a vselect if its argument
23284 // is a sign_extend_inreg of an any_extend of a BUILD_VECTOR of
23285 // constants. This specific pattern gets generated when we split a
23286 // selector for a 512 bit vector in a machine without AVX512 (but with
23287 // 256-bit vectors), during legalization:
23289 // (vselect (sign_extend (any_extend (BUILD_VECTOR)) i1) LHS RHS)
23291 // Iff we find this pattern and the build_vectors are built from
23292 // constants, we translate the vselect into a shuffle_vector that we
23293 // know will be matched by LowerVECTOR_SHUFFLEtoBlend.
23294 if ((N->getOpcode() == ISD::VSELECT ||
23295 N->getOpcode() == X86ISD::SHRUNKBLEND) &&
23296 !DCI.isBeforeLegalize() && !VT.is512BitVector()) {
23297 SDValue Shuffle = transformVSELECTtoBlendVECTOR_SHUFFLE(N, DAG, Subtarget);
23298 if (Shuffle.getNode())
23302 // If this is a *dynamic* select (non-constant condition) and we can match
23303 // this node with one of the variable blend instructions, restructure the
23304 // condition so that the blends can use the high bit of each element and use
23305 // SimplifyDemandedBits to simplify the condition operand.
23306 if (N->getOpcode() == ISD::VSELECT && DCI.isBeforeLegalizeOps() &&
23307 !DCI.isBeforeLegalize() &&
23308 !ISD::isBuildVectorOfConstantSDNodes(Cond.getNode())) {
23309 unsigned BitWidth = Cond.getValueType().getScalarType().getSizeInBits();
23311 // Don't optimize vector selects that map to mask-registers.
23315 // We can only handle the cases where VSELECT is directly legal on the
23316 // subtarget. We custom lower VSELECT nodes with constant conditions and
23317 // this makes it hard to see whether a dynamic VSELECT will correctly
23318 // lower, so we both check the operation's status and explicitly handle the
23319 // cases where a *dynamic* blend will fail even though a constant-condition
23320 // blend could be custom lowered.
23321 // FIXME: We should find a better way to handle this class of problems.
23322 // Potentially, we should combine constant-condition vselect nodes
23323 // pre-legalization into shuffles and not mark as many types as custom
23325 if (!TLI.isOperationLegalOrCustom(ISD::VSELECT, VT))
23327 // FIXME: We don't support i16-element blends currently. We could and
23328 // should support them by making *all* the bits in the condition be set
23329 // rather than just the high bit and using an i8-element blend.
23330 if (VT.getScalarType() == MVT::i16)
23332 // Dynamic blending was only available from SSE4.1 onward.
23333 if (VT.getSizeInBits() == 128 && !Subtarget->hasSSE41())
23335 // Byte blends are only available in AVX2
23336 if (VT.getSizeInBits() == 256 && VT.getScalarType() == MVT::i8 &&
23337 !Subtarget->hasAVX2())
23340 assert(BitWidth >= 8 && BitWidth <= 64 && "Invalid mask size");
23341 APInt DemandedMask = APInt::getHighBitsSet(BitWidth, 1);
23343 APInt KnownZero, KnownOne;
23344 TargetLowering::TargetLoweringOpt TLO(DAG, DCI.isBeforeLegalize(),
23345 DCI.isBeforeLegalizeOps());
23346 if (TLO.ShrinkDemandedConstant(Cond, DemandedMask) ||
23347 TLI.SimplifyDemandedBits(Cond, DemandedMask, KnownZero, KnownOne,
23349 // If we changed the computation somewhere in the DAG, this change
23350 // will affect all users of Cond.
23351 // Make sure it is fine and update all the nodes so that we do not
23352 // use the generic VSELECT anymore. Otherwise, we may perform
23353 // wrong optimizations as we messed up with the actual expectation
23354 // for the vector boolean values.
23355 if (Cond != TLO.Old) {
23356 // Check all uses of that condition operand to check whether it will be
23357 // consumed by non-BLEND instructions, which may depend on all bits are
23359 for (SDNode::use_iterator I = Cond->use_begin(), E = Cond->use_end();
23361 if (I->getOpcode() != ISD::VSELECT)
23362 // TODO: Add other opcodes eventually lowered into BLEND.
23365 // Update all the users of the condition, before committing the change,
23366 // so that the VSELECT optimizations that expect the correct vector
23367 // boolean value will not be triggered.
23368 for (SDNode::use_iterator I = Cond->use_begin(), E = Cond->use_end();
23370 DAG.ReplaceAllUsesOfValueWith(
23372 DAG.getNode(X86ISD::SHRUNKBLEND, SDLoc(*I), I->getValueType(0),
23373 Cond, I->getOperand(1), I->getOperand(2)));
23374 DCI.CommitTargetLoweringOpt(TLO);
23377 // At this point, only Cond is changed. Change the condition
23378 // just for N to keep the opportunity to optimize all other
23379 // users their own way.
23380 DAG.ReplaceAllUsesOfValueWith(
23382 DAG.getNode(X86ISD::SHRUNKBLEND, SDLoc(N), N->getValueType(0),
23383 TLO.New, N->getOperand(1), N->getOperand(2)));
23391 // Check whether a boolean test is testing a boolean value generated by
23392 // X86ISD::SETCC. If so, return the operand of that SETCC and proper condition
23395 // Simplify the following patterns:
23396 // (Op (CMP (SETCC Cond EFLAGS) 1) EQ) or
23397 // (Op (CMP (SETCC Cond EFLAGS) 0) NEQ)
23398 // to (Op EFLAGS Cond)
23400 // (Op (CMP (SETCC Cond EFLAGS) 0) EQ) or
23401 // (Op (CMP (SETCC Cond EFLAGS) 1) NEQ)
23402 // to (Op EFLAGS !Cond)
23404 // where Op could be BRCOND or CMOV.
23406 static SDValue checkBoolTestSetCCCombine(SDValue Cmp, X86::CondCode &CC) {
23407 // Quit if not CMP and SUB with its value result used.
23408 if (Cmp.getOpcode() != X86ISD::CMP &&
23409 (Cmp.getOpcode() != X86ISD::SUB || Cmp.getNode()->hasAnyUseOfValue(0)))
23412 // Quit if not used as a boolean value.
23413 if (CC != X86::COND_E && CC != X86::COND_NE)
23416 // Check CMP operands. One of them should be 0 or 1 and the other should be
23417 // an SetCC or extended from it.
23418 SDValue Op1 = Cmp.getOperand(0);
23419 SDValue Op2 = Cmp.getOperand(1);
23422 const ConstantSDNode* C = nullptr;
23423 bool needOppositeCond = (CC == X86::COND_E);
23424 bool checkAgainstTrue = false; // Is it a comparison against 1?
23426 if ((C = dyn_cast<ConstantSDNode>(Op1)))
23428 else if ((C = dyn_cast<ConstantSDNode>(Op2)))
23430 else // Quit if all operands are not constants.
23433 if (C->getZExtValue() == 1) {
23434 needOppositeCond = !needOppositeCond;
23435 checkAgainstTrue = true;
23436 } else if (C->getZExtValue() != 0)
23437 // Quit if the constant is neither 0 or 1.
23440 bool truncatedToBoolWithAnd = false;
23441 // Skip (zext $x), (trunc $x), or (and $x, 1) node.
23442 while (SetCC.getOpcode() == ISD::ZERO_EXTEND ||
23443 SetCC.getOpcode() == ISD::TRUNCATE ||
23444 SetCC.getOpcode() == ISD::AND) {
23445 if (SetCC.getOpcode() == ISD::AND) {
23447 ConstantSDNode *CS;
23448 if ((CS = dyn_cast<ConstantSDNode>(SetCC.getOperand(0))) &&
23449 CS->getZExtValue() == 1)
23451 if ((CS = dyn_cast<ConstantSDNode>(SetCC.getOperand(1))) &&
23452 CS->getZExtValue() == 1)
23456 SetCC = SetCC.getOperand(OpIdx);
23457 truncatedToBoolWithAnd = true;
23459 SetCC = SetCC.getOperand(0);
23462 switch (SetCC.getOpcode()) {
23463 case X86ISD::SETCC_CARRY:
23464 // Since SETCC_CARRY gives output based on R = CF ? ~0 : 0, it's unsafe to
23465 // simplify it if the result of SETCC_CARRY is not canonicalized to 0 or 1,
23466 // i.e. it's a comparison against true but the result of SETCC_CARRY is not
23467 // truncated to i1 using 'and'.
23468 if (checkAgainstTrue && !truncatedToBoolWithAnd)
23470 assert(X86::CondCode(SetCC.getConstantOperandVal(0)) == X86::COND_B &&
23471 "Invalid use of SETCC_CARRY!");
23473 case X86ISD::SETCC:
23474 // Set the condition code or opposite one if necessary.
23475 CC = X86::CondCode(SetCC.getConstantOperandVal(0));
23476 if (needOppositeCond)
23477 CC = X86::GetOppositeBranchCondition(CC);
23478 return SetCC.getOperand(1);
23479 case X86ISD::CMOV: {
23480 // Check whether false/true value has canonical one, i.e. 0 or 1.
23481 ConstantSDNode *FVal = dyn_cast<ConstantSDNode>(SetCC.getOperand(0));
23482 ConstantSDNode *TVal = dyn_cast<ConstantSDNode>(SetCC.getOperand(1));
23483 // Quit if true value is not a constant.
23486 // Quit if false value is not a constant.
23488 SDValue Op = SetCC.getOperand(0);
23489 // Skip 'zext' or 'trunc' node.
23490 if (Op.getOpcode() == ISD::ZERO_EXTEND ||
23491 Op.getOpcode() == ISD::TRUNCATE)
23492 Op = Op.getOperand(0);
23493 // A special case for rdrand/rdseed, where 0 is set if false cond is
23495 if ((Op.getOpcode() != X86ISD::RDRAND &&
23496 Op.getOpcode() != X86ISD::RDSEED) || Op.getResNo() != 0)
23499 // Quit if false value is not the constant 0 or 1.
23500 bool FValIsFalse = true;
23501 if (FVal && FVal->getZExtValue() != 0) {
23502 if (FVal->getZExtValue() != 1)
23504 // If FVal is 1, opposite cond is needed.
23505 needOppositeCond = !needOppositeCond;
23506 FValIsFalse = false;
23508 // Quit if TVal is not the constant opposite of FVal.
23509 if (FValIsFalse && TVal->getZExtValue() != 1)
23511 if (!FValIsFalse && TVal->getZExtValue() != 0)
23513 CC = X86::CondCode(SetCC.getConstantOperandVal(2));
23514 if (needOppositeCond)
23515 CC = X86::GetOppositeBranchCondition(CC);
23516 return SetCC.getOperand(3);
23523 /// Check whether Cond is an AND/OR of SETCCs off of the same EFLAGS.
23525 /// (X86or (X86setcc) (X86setcc))
23526 /// (X86cmp (and (X86setcc) (X86setcc)), 0)
23527 static bool checkBoolTestAndOrSetCCCombine(SDValue Cond, X86::CondCode &CC0,
23528 X86::CondCode &CC1, SDValue &Flags,
23530 if (Cond->getOpcode() == X86ISD::CMP) {
23531 ConstantSDNode *CondOp1C = dyn_cast<ConstantSDNode>(Cond->getOperand(1));
23532 if (!CondOp1C || !CondOp1C->isNullValue())
23535 Cond = Cond->getOperand(0);
23540 SDValue SetCC0, SetCC1;
23541 switch (Cond->getOpcode()) {
23542 default: return false;
23549 SetCC0 = Cond->getOperand(0);
23550 SetCC1 = Cond->getOperand(1);
23554 // Make sure we have SETCC nodes, using the same flags value.
23555 if (SetCC0.getOpcode() != X86ISD::SETCC ||
23556 SetCC1.getOpcode() != X86ISD::SETCC ||
23557 SetCC0->getOperand(1) != SetCC1->getOperand(1))
23560 CC0 = (X86::CondCode)SetCC0->getConstantOperandVal(0);
23561 CC1 = (X86::CondCode)SetCC1->getConstantOperandVal(0);
23562 Flags = SetCC0->getOperand(1);
23566 /// Optimize X86ISD::CMOV [LHS, RHS, CONDCODE (e.g. X86::COND_NE), CONDVAL]
23567 static SDValue PerformCMOVCombine(SDNode *N, SelectionDAG &DAG,
23568 TargetLowering::DAGCombinerInfo &DCI,
23569 const X86Subtarget *Subtarget) {
23572 // If the flag operand isn't dead, don't touch this CMOV.
23573 if (N->getNumValues() == 2 && !SDValue(N, 1).use_empty())
23576 SDValue FalseOp = N->getOperand(0);
23577 SDValue TrueOp = N->getOperand(1);
23578 X86::CondCode CC = (X86::CondCode)N->getConstantOperandVal(2);
23579 SDValue Cond = N->getOperand(3);
23581 if (CC == X86::COND_E || CC == X86::COND_NE) {
23582 switch (Cond.getOpcode()) {
23586 // If operand of BSR / BSF are proven never zero, then ZF cannot be set.
23587 if (DAG.isKnownNeverZero(Cond.getOperand(0)))
23588 return (CC == X86::COND_E) ? FalseOp : TrueOp;
23594 Flags = checkBoolTestSetCCCombine(Cond, CC);
23595 if (Flags.getNode() &&
23596 // Extra check as FCMOV only supports a subset of X86 cond.
23597 (FalseOp.getValueType() != MVT::f80 || hasFPCMov(CC))) {
23598 SDValue Ops[] = { FalseOp, TrueOp,
23599 DAG.getConstant(CC, DL, MVT::i8), Flags };
23600 return DAG.getNode(X86ISD::CMOV, DL, N->getVTList(), Ops);
23603 // If this is a select between two integer constants, try to do some
23604 // optimizations. Note that the operands are ordered the opposite of SELECT
23606 if (ConstantSDNode *TrueC = dyn_cast<ConstantSDNode>(TrueOp)) {
23607 if (ConstantSDNode *FalseC = dyn_cast<ConstantSDNode>(FalseOp)) {
23608 // Canonicalize the TrueC/FalseC values so that TrueC (the true value) is
23609 // larger than FalseC (the false value).
23610 if (TrueC->getAPIntValue().ult(FalseC->getAPIntValue())) {
23611 CC = X86::GetOppositeBranchCondition(CC);
23612 std::swap(TrueC, FalseC);
23613 std::swap(TrueOp, FalseOp);
23616 // Optimize C ? 8 : 0 -> zext(setcc(C)) << 3. Likewise for any pow2/0.
23617 // This is efficient for any integer data type (including i8/i16) and
23619 if (FalseC->getAPIntValue() == 0 && TrueC->getAPIntValue().isPowerOf2()) {
23620 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
23621 DAG.getConstant(CC, DL, MVT::i8), Cond);
23623 // Zero extend the condition if needed.
23624 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, TrueC->getValueType(0), Cond);
23626 unsigned ShAmt = TrueC->getAPIntValue().logBase2();
23627 Cond = DAG.getNode(ISD::SHL, DL, Cond.getValueType(), Cond,
23628 DAG.getConstant(ShAmt, DL, MVT::i8));
23629 if (N->getNumValues() == 2) // Dead flag value?
23630 return DCI.CombineTo(N, Cond, SDValue());
23634 // Optimize Cond ? cst+1 : cst -> zext(setcc(C)+cst. This is efficient
23635 // for any integer data type, including i8/i16.
23636 if (FalseC->getAPIntValue()+1 == TrueC->getAPIntValue()) {
23637 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
23638 DAG.getConstant(CC, DL, MVT::i8), Cond);
23640 // Zero extend the condition if needed.
23641 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL,
23642 FalseC->getValueType(0), Cond);
23643 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
23644 SDValue(FalseC, 0));
23646 if (N->getNumValues() == 2) // Dead flag value?
23647 return DCI.CombineTo(N, Cond, SDValue());
23651 // Optimize cases that will turn into an LEA instruction. This requires
23652 // an i32 or i64 and an efficient multiplier (1, 2, 3, 4, 5, 8, 9).
23653 if (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i64) {
23654 uint64_t Diff = TrueC->getZExtValue()-FalseC->getZExtValue();
23655 if (N->getValueType(0) == MVT::i32) Diff = (unsigned)Diff;
23657 bool isFastMultiplier = false;
23659 switch ((unsigned char)Diff) {
23661 case 1: // result = add base, cond
23662 case 2: // result = lea base( , cond*2)
23663 case 3: // result = lea base(cond, cond*2)
23664 case 4: // result = lea base( , cond*4)
23665 case 5: // result = lea base(cond, cond*4)
23666 case 8: // result = lea base( , cond*8)
23667 case 9: // result = lea base(cond, cond*8)
23668 isFastMultiplier = true;
23673 if (isFastMultiplier) {
23674 APInt Diff = TrueC->getAPIntValue()-FalseC->getAPIntValue();
23675 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
23676 DAG.getConstant(CC, DL, MVT::i8), Cond);
23677 // Zero extend the condition if needed.
23678 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, FalseC->getValueType(0),
23680 // Scale the condition by the difference.
23682 Cond = DAG.getNode(ISD::MUL, DL, Cond.getValueType(), Cond,
23683 DAG.getConstant(Diff, DL, Cond.getValueType()));
23685 // Add the base if non-zero.
23686 if (FalseC->getAPIntValue() != 0)
23687 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
23688 SDValue(FalseC, 0));
23689 if (N->getNumValues() == 2) // Dead flag value?
23690 return DCI.CombineTo(N, Cond, SDValue());
23697 // Handle these cases:
23698 // (select (x != c), e, c) -> select (x != c), e, x),
23699 // (select (x == c), c, e) -> select (x == c), x, e)
23700 // where the c is an integer constant, and the "select" is the combination
23701 // of CMOV and CMP.
23703 // The rationale for this change is that the conditional-move from a constant
23704 // needs two instructions, however, conditional-move from a register needs
23705 // only one instruction.
23707 // CAVEAT: By replacing a constant with a symbolic value, it may obscure
23708 // some instruction-combining opportunities. This opt needs to be
23709 // postponed as late as possible.
23711 if (!DCI.isBeforeLegalize() && !DCI.isBeforeLegalizeOps()) {
23712 // the DCI.xxxx conditions are provided to postpone the optimization as
23713 // late as possible.
23715 ConstantSDNode *CmpAgainst = nullptr;
23716 if ((Cond.getOpcode() == X86ISD::CMP || Cond.getOpcode() == X86ISD::SUB) &&
23717 (CmpAgainst = dyn_cast<ConstantSDNode>(Cond.getOperand(1))) &&
23718 !isa<ConstantSDNode>(Cond.getOperand(0))) {
23720 if (CC == X86::COND_NE &&
23721 CmpAgainst == dyn_cast<ConstantSDNode>(FalseOp)) {
23722 CC = X86::GetOppositeBranchCondition(CC);
23723 std::swap(TrueOp, FalseOp);
23726 if (CC == X86::COND_E &&
23727 CmpAgainst == dyn_cast<ConstantSDNode>(TrueOp)) {
23728 SDValue Ops[] = { FalseOp, Cond.getOperand(0),
23729 DAG.getConstant(CC, DL, MVT::i8), Cond };
23730 return DAG.getNode(X86ISD::CMOV, DL, N->getVTList (), Ops);
23735 // Fold and/or of setcc's to double CMOV:
23736 // (CMOV F, T, ((cc1 | cc2) != 0)) -> (CMOV (CMOV F, T, cc1), T, cc2)
23737 // (CMOV F, T, ((cc1 & cc2) != 0)) -> (CMOV (CMOV T, F, !cc1), F, !cc2)
23739 // This combine lets us generate:
23740 // cmovcc1 (jcc1 if we don't have CMOV)
23746 // cmovne (jne if we don't have CMOV)
23747 // When we can't use the CMOV instruction, it might increase branch
23749 // When we can use CMOV, or when there is no mispredict, this improves
23750 // throughput and reduces register pressure.
23752 if (CC == X86::COND_NE) {
23754 X86::CondCode CC0, CC1;
23756 if (checkBoolTestAndOrSetCCCombine(Cond, CC0, CC1, Flags, isAndSetCC)) {
23758 std::swap(FalseOp, TrueOp);
23759 CC0 = X86::GetOppositeBranchCondition(CC0);
23760 CC1 = X86::GetOppositeBranchCondition(CC1);
23763 SDValue LOps[] = {FalseOp, TrueOp, DAG.getConstant(CC0, DL, MVT::i8),
23765 SDValue LCMOV = DAG.getNode(X86ISD::CMOV, DL, N->getVTList(), LOps);
23766 SDValue Ops[] = {LCMOV, TrueOp, DAG.getConstant(CC1, DL, MVT::i8), Flags};
23767 SDValue CMOV = DAG.getNode(X86ISD::CMOV, DL, N->getVTList(), Ops);
23768 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), SDValue(CMOV.getNode(), 1));
23776 /// PerformMulCombine - Optimize a single multiply with constant into two
23777 /// in order to implement it with two cheaper instructions, e.g.
23778 /// LEA + SHL, LEA + LEA.
23779 static SDValue PerformMulCombine(SDNode *N, SelectionDAG &DAG,
23780 TargetLowering::DAGCombinerInfo &DCI) {
23781 // An imul is usually smaller than the alternative sequence.
23782 if (DAG.getMachineFunction().getFunction()->optForMinSize())
23785 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
23788 EVT VT = N->getValueType(0);
23789 if (VT != MVT::i64 && VT != MVT::i32)
23792 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1));
23795 uint64_t MulAmt = C->getZExtValue();
23796 if (isPowerOf2_64(MulAmt) || MulAmt == 3 || MulAmt == 5 || MulAmt == 9)
23799 uint64_t MulAmt1 = 0;
23800 uint64_t MulAmt2 = 0;
23801 if ((MulAmt % 9) == 0) {
23803 MulAmt2 = MulAmt / 9;
23804 } else if ((MulAmt % 5) == 0) {
23806 MulAmt2 = MulAmt / 5;
23807 } else if ((MulAmt % 3) == 0) {
23809 MulAmt2 = MulAmt / 3;
23812 (isPowerOf2_64(MulAmt2) || MulAmt2 == 3 || MulAmt2 == 5 || MulAmt2 == 9)){
23815 if (isPowerOf2_64(MulAmt2) &&
23816 !(N->hasOneUse() && N->use_begin()->getOpcode() == ISD::ADD))
23817 // If second multiplifer is pow2, issue it first. We want the multiply by
23818 // 3, 5, or 9 to be folded into the addressing mode unless the lone use
23820 std::swap(MulAmt1, MulAmt2);
23823 if (isPowerOf2_64(MulAmt1))
23824 NewMul = DAG.getNode(ISD::SHL, DL, VT, N->getOperand(0),
23825 DAG.getConstant(Log2_64(MulAmt1), DL, MVT::i8));
23827 NewMul = DAG.getNode(X86ISD::MUL_IMM, DL, VT, N->getOperand(0),
23828 DAG.getConstant(MulAmt1, DL, VT));
23830 if (isPowerOf2_64(MulAmt2))
23831 NewMul = DAG.getNode(ISD::SHL, DL, VT, NewMul,
23832 DAG.getConstant(Log2_64(MulAmt2), DL, MVT::i8));
23834 NewMul = DAG.getNode(X86ISD::MUL_IMM, DL, VT, NewMul,
23835 DAG.getConstant(MulAmt2, DL, VT));
23837 // Do not add new nodes to DAG combiner worklist.
23838 DCI.CombineTo(N, NewMul, false);
23843 static SDValue PerformSHLCombine(SDNode *N, SelectionDAG &DAG) {
23844 SDValue N0 = N->getOperand(0);
23845 SDValue N1 = N->getOperand(1);
23846 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
23847 EVT VT = N0.getValueType();
23849 // fold (shl (and (setcc_c), c1), c2) -> (and setcc_c, (c1 << c2))
23850 // since the result of setcc_c is all zero's or all ones.
23851 if (VT.isInteger() && !VT.isVector() &&
23852 N1C && N0.getOpcode() == ISD::AND &&
23853 N0.getOperand(1).getOpcode() == ISD::Constant) {
23854 SDValue N00 = N0.getOperand(0);
23855 APInt Mask = cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
23856 APInt ShAmt = N1C->getAPIntValue();
23857 Mask = Mask.shl(ShAmt);
23858 bool MaskOK = false;
23859 // We can handle cases concerning bit-widening nodes containing setcc_c if
23860 // we carefully interrogate the mask to make sure we are semantics
23862 // The transform is not safe if the result of C1 << C2 exceeds the bitwidth
23863 // of the underlying setcc_c operation if the setcc_c was zero extended.
23864 // Consider the following example:
23865 // zext(setcc_c) -> i32 0x0000FFFF
23866 // c1 -> i32 0x0000FFFF
23867 // c2 -> i32 0x00000001
23868 // (shl (and (setcc_c), c1), c2) -> i32 0x0001FFFE
23869 // (and setcc_c, (c1 << c2)) -> i32 0x0000FFFE
23870 if (N00.getOpcode() == X86ISD::SETCC_CARRY) {
23872 } else if (N00.getOpcode() == ISD::SIGN_EXTEND &&
23873 N00.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY) {
23875 } else if ((N00.getOpcode() == ISD::ZERO_EXTEND ||
23876 N00.getOpcode() == ISD::ANY_EXTEND) &&
23877 N00.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY) {
23878 MaskOK = Mask.isIntN(N00.getOperand(0).getValueSizeInBits());
23880 if (MaskOK && Mask != 0) {
23882 return DAG.getNode(ISD::AND, DL, VT, N00, DAG.getConstant(Mask, DL, VT));
23886 // Hardware support for vector shifts is sparse which makes us scalarize the
23887 // vector operations in many cases. Also, on sandybridge ADD is faster than
23889 // (shl V, 1) -> add V,V
23890 if (auto *N1BV = dyn_cast<BuildVectorSDNode>(N1))
23891 if (auto *N1SplatC = N1BV->getConstantSplatNode()) {
23892 assert(N0.getValueType().isVector() && "Invalid vector shift type");
23893 // We shift all of the values by one. In many cases we do not have
23894 // hardware support for this operation. This is better expressed as an ADD
23896 if (N1SplatC->getAPIntValue() == 1)
23897 return DAG.getNode(ISD::ADD, SDLoc(N), VT, N0, N0);
23903 /// \brief Returns a vector of 0s if the node in input is a vector logical
23904 /// shift by a constant amount which is known to be bigger than or equal
23905 /// to the vector element size in bits.
23906 static SDValue performShiftToAllZeros(SDNode *N, SelectionDAG &DAG,
23907 const X86Subtarget *Subtarget) {
23908 EVT VT = N->getValueType(0);
23910 if (VT != MVT::v2i64 && VT != MVT::v4i32 && VT != MVT::v8i16 &&
23911 (!Subtarget->hasInt256() ||
23912 (VT != MVT::v4i64 && VT != MVT::v8i32 && VT != MVT::v16i16)))
23915 SDValue Amt = N->getOperand(1);
23917 if (auto *AmtBV = dyn_cast<BuildVectorSDNode>(Amt))
23918 if (auto *AmtSplat = AmtBV->getConstantSplatNode()) {
23919 APInt ShiftAmt = AmtSplat->getAPIntValue();
23920 unsigned MaxAmount = VT.getVectorElementType().getSizeInBits();
23922 // SSE2/AVX2 logical shifts always return a vector of 0s
23923 // if the shift amount is bigger than or equal to
23924 // the element size. The constant shift amount will be
23925 // encoded as a 8-bit immediate.
23926 if (ShiftAmt.trunc(8).uge(MaxAmount))
23927 return getZeroVector(VT, Subtarget, DAG, DL);
23933 /// PerformShiftCombine - Combine shifts.
23934 static SDValue PerformShiftCombine(SDNode* N, SelectionDAG &DAG,
23935 TargetLowering::DAGCombinerInfo &DCI,
23936 const X86Subtarget *Subtarget) {
23937 if (N->getOpcode() == ISD::SHL)
23938 if (SDValue V = PerformSHLCombine(N, DAG))
23941 // Try to fold this logical shift into a zero vector.
23942 if (N->getOpcode() != ISD::SRA)
23943 if (SDValue V = performShiftToAllZeros(N, DAG, Subtarget))
23949 // CMPEQCombine - Recognize the distinctive (AND (setcc ...) (setcc ..))
23950 // where both setccs reference the same FP CMP, and rewrite for CMPEQSS
23951 // and friends. Likewise for OR -> CMPNEQSS.
23952 static SDValue CMPEQCombine(SDNode *N, SelectionDAG &DAG,
23953 TargetLowering::DAGCombinerInfo &DCI,
23954 const X86Subtarget *Subtarget) {
23957 // SSE1 supports CMP{eq|ne}SS, and SSE2 added CMP{eq|ne}SD, but
23958 // we're requiring SSE2 for both.
23959 if (Subtarget->hasSSE2() && isAndOrOfSetCCs(SDValue(N, 0U), opcode)) {
23960 SDValue N0 = N->getOperand(0);
23961 SDValue N1 = N->getOperand(1);
23962 SDValue CMP0 = N0->getOperand(1);
23963 SDValue CMP1 = N1->getOperand(1);
23966 // The SETCCs should both refer to the same CMP.
23967 if (CMP0.getOpcode() != X86ISD::CMP || CMP0 != CMP1)
23970 SDValue CMP00 = CMP0->getOperand(0);
23971 SDValue CMP01 = CMP0->getOperand(1);
23972 EVT VT = CMP00.getValueType();
23974 if (VT == MVT::f32 || VT == MVT::f64) {
23975 bool ExpectingFlags = false;
23976 // Check for any users that want flags:
23977 for (SDNode::use_iterator UI = N->use_begin(), UE = N->use_end();
23978 !ExpectingFlags && UI != UE; ++UI)
23979 switch (UI->getOpcode()) {
23984 ExpectingFlags = true;
23986 case ISD::CopyToReg:
23987 case ISD::SIGN_EXTEND:
23988 case ISD::ZERO_EXTEND:
23989 case ISD::ANY_EXTEND:
23993 if (!ExpectingFlags) {
23994 enum X86::CondCode cc0 = (enum X86::CondCode)N0.getConstantOperandVal(0);
23995 enum X86::CondCode cc1 = (enum X86::CondCode)N1.getConstantOperandVal(0);
23997 if (cc1 == X86::COND_E || cc1 == X86::COND_NE) {
23998 X86::CondCode tmp = cc0;
24003 if ((cc0 == X86::COND_E && cc1 == X86::COND_NP) ||
24004 (cc0 == X86::COND_NE && cc1 == X86::COND_P)) {
24005 // FIXME: need symbolic constants for these magic numbers.
24006 // See X86ATTInstPrinter.cpp:printSSECC().
24007 unsigned x86cc = (cc0 == X86::COND_E) ? 0 : 4;
24008 if (Subtarget->hasAVX512()) {
24009 SDValue FSetCC = DAG.getNode(X86ISD::FSETCC, DL, MVT::i1, CMP00,
24011 DAG.getConstant(x86cc, DL, MVT::i8));
24012 if (N->getValueType(0) != MVT::i1)
24013 return DAG.getNode(ISD::ZERO_EXTEND, DL, N->getValueType(0),
24017 SDValue OnesOrZeroesF = DAG.getNode(X86ISD::FSETCC, DL,
24018 CMP00.getValueType(), CMP00, CMP01,
24019 DAG.getConstant(x86cc, DL,
24022 bool is64BitFP = (CMP00.getValueType() == MVT::f64);
24023 MVT IntVT = is64BitFP ? MVT::i64 : MVT::i32;
24025 if (is64BitFP && !Subtarget->is64Bit()) {
24026 // On a 32-bit target, we cannot bitcast the 64-bit float to a
24027 // 64-bit integer, since that's not a legal type. Since
24028 // OnesOrZeroesF is all ones of all zeroes, we don't need all the
24029 // bits, but can do this little dance to extract the lowest 32 bits
24030 // and work with those going forward.
24031 SDValue Vector64 = DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, MVT::v2f64,
24033 SDValue Vector32 = DAG.getBitcast(MVT::v4f32, Vector64);
24034 OnesOrZeroesF = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::f32,
24035 Vector32, DAG.getIntPtrConstant(0, DL));
24039 SDValue OnesOrZeroesI = DAG.getBitcast(IntVT, OnesOrZeroesF);
24040 SDValue ANDed = DAG.getNode(ISD::AND, DL, IntVT, OnesOrZeroesI,
24041 DAG.getConstant(1, DL, IntVT));
24042 SDValue OneBitOfTruth = DAG.getNode(ISD::TRUNCATE, DL, MVT::i8,
24044 return OneBitOfTruth;
24052 /// CanFoldXORWithAllOnes - Test whether the XOR operand is a AllOnes vector
24053 /// so it can be folded inside ANDNP.
24054 static bool CanFoldXORWithAllOnes(const SDNode *N) {
24055 EVT VT = N->getValueType(0);
24057 // Match direct AllOnes for 128 and 256-bit vectors
24058 if (ISD::isBuildVectorAllOnes(N))
24061 // Look through a bit convert.
24062 if (N->getOpcode() == ISD::BITCAST)
24063 N = N->getOperand(0).getNode();
24065 // Sometimes the operand may come from a insert_subvector building a 256-bit
24067 if (VT.is256BitVector() &&
24068 N->getOpcode() == ISD::INSERT_SUBVECTOR) {
24069 SDValue V1 = N->getOperand(0);
24070 SDValue V2 = N->getOperand(1);
24072 if (V1.getOpcode() == ISD::INSERT_SUBVECTOR &&
24073 V1.getOperand(0).getOpcode() == ISD::UNDEF &&
24074 ISD::isBuildVectorAllOnes(V1.getOperand(1).getNode()) &&
24075 ISD::isBuildVectorAllOnes(V2.getNode()))
24082 // On AVX/AVX2 the type v8i1 is legalized to v8i16, which is an XMM sized
24083 // register. In most cases we actually compare or select YMM-sized registers
24084 // and mixing the two types creates horrible code. This method optimizes
24085 // some of the transition sequences.
24086 static SDValue WidenMaskArithmetic(SDNode *N, SelectionDAG &DAG,
24087 TargetLowering::DAGCombinerInfo &DCI,
24088 const X86Subtarget *Subtarget) {
24089 EVT VT = N->getValueType(0);
24090 if (!VT.is256BitVector())
24093 assert((N->getOpcode() == ISD::ANY_EXTEND ||
24094 N->getOpcode() == ISD::ZERO_EXTEND ||
24095 N->getOpcode() == ISD::SIGN_EXTEND) && "Invalid Node");
24097 SDValue Narrow = N->getOperand(0);
24098 EVT NarrowVT = Narrow->getValueType(0);
24099 if (!NarrowVT.is128BitVector())
24102 if (Narrow->getOpcode() != ISD::XOR &&
24103 Narrow->getOpcode() != ISD::AND &&
24104 Narrow->getOpcode() != ISD::OR)
24107 SDValue N0 = Narrow->getOperand(0);
24108 SDValue N1 = Narrow->getOperand(1);
24111 // The Left side has to be a trunc.
24112 if (N0.getOpcode() != ISD::TRUNCATE)
24115 // The type of the truncated inputs.
24116 EVT WideVT = N0->getOperand(0)->getValueType(0);
24120 // The right side has to be a 'trunc' or a constant vector.
24121 bool RHSTrunc = N1.getOpcode() == ISD::TRUNCATE;
24122 ConstantSDNode *RHSConstSplat = nullptr;
24123 if (auto *RHSBV = dyn_cast<BuildVectorSDNode>(N1))
24124 RHSConstSplat = RHSBV->getConstantSplatNode();
24125 if (!RHSTrunc && !RHSConstSplat)
24128 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
24130 if (!TLI.isOperationLegalOrPromote(Narrow->getOpcode(), WideVT))
24133 // Set N0 and N1 to hold the inputs to the new wide operation.
24134 N0 = N0->getOperand(0);
24135 if (RHSConstSplat) {
24136 N1 = DAG.getNode(ISD::ZERO_EXTEND, DL, WideVT.getScalarType(),
24137 SDValue(RHSConstSplat, 0));
24138 SmallVector<SDValue, 8> C(WideVT.getVectorNumElements(), N1);
24139 N1 = DAG.getNode(ISD::BUILD_VECTOR, DL, WideVT, C);
24140 } else if (RHSTrunc) {
24141 N1 = N1->getOperand(0);
24144 // Generate the wide operation.
24145 SDValue Op = DAG.getNode(Narrow->getOpcode(), DL, WideVT, N0, N1);
24146 unsigned Opcode = N->getOpcode();
24148 case ISD::ANY_EXTEND:
24150 case ISD::ZERO_EXTEND: {
24151 unsigned InBits = NarrowVT.getScalarType().getSizeInBits();
24152 APInt Mask = APInt::getAllOnesValue(InBits);
24153 Mask = Mask.zext(VT.getScalarType().getSizeInBits());
24154 return DAG.getNode(ISD::AND, DL, VT,
24155 Op, DAG.getConstant(Mask, DL, VT));
24157 case ISD::SIGN_EXTEND:
24158 return DAG.getNode(ISD::SIGN_EXTEND_INREG, DL, VT,
24159 Op, DAG.getValueType(NarrowVT));
24161 llvm_unreachable("Unexpected opcode");
24165 static SDValue VectorZextCombine(SDNode *N, SelectionDAG &DAG,
24166 TargetLowering::DAGCombinerInfo &DCI,
24167 const X86Subtarget *Subtarget) {
24168 SDValue N0 = N->getOperand(0);
24169 SDValue N1 = N->getOperand(1);
24172 // A vector zext_in_reg may be represented as a shuffle,
24173 // feeding into a bitcast (this represents anyext) feeding into
24174 // an and with a mask.
24175 // We'd like to try to combine that into a shuffle with zero
24176 // plus a bitcast, removing the and.
24177 if (N0.getOpcode() != ISD::BITCAST ||
24178 N0.getOperand(0).getOpcode() != ISD::VECTOR_SHUFFLE)
24181 // The other side of the AND should be a splat of 2^C, where C
24182 // is the number of bits in the source type.
24183 if (N1.getOpcode() == ISD::BITCAST)
24184 N1 = N1.getOperand(0);
24185 if (N1.getOpcode() != ISD::BUILD_VECTOR)
24187 BuildVectorSDNode *Vector = cast<BuildVectorSDNode>(N1);
24189 ShuffleVectorSDNode *Shuffle = cast<ShuffleVectorSDNode>(N0.getOperand(0));
24190 EVT SrcType = Shuffle->getValueType(0);
24192 // We expect a single-source shuffle
24193 if (Shuffle->getOperand(1)->getOpcode() != ISD::UNDEF)
24196 unsigned SrcSize = SrcType.getScalarSizeInBits();
24198 APInt SplatValue, SplatUndef;
24199 unsigned SplatBitSize;
24201 if (!Vector->isConstantSplat(SplatValue, SplatUndef,
24202 SplatBitSize, HasAnyUndefs))
24205 unsigned ResSize = N1.getValueType().getScalarSizeInBits();
24206 // Make sure the splat matches the mask we expect
24207 if (SplatBitSize > ResSize ||
24208 (SplatValue + 1).exactLogBase2() != (int)SrcSize)
24211 // Make sure the input and output size make sense
24212 if (SrcSize >= ResSize || ResSize % SrcSize)
24215 // We expect a shuffle of the form <0, u, u, u, 1, u, u, u...>
24216 // The number of u's between each two values depends on the ratio between
24217 // the source and dest type.
24218 unsigned ZextRatio = ResSize / SrcSize;
24219 bool IsZext = true;
24220 for (unsigned i = 0; i < SrcType.getVectorNumElements(); ++i) {
24221 if (i % ZextRatio) {
24222 if (Shuffle->getMaskElt(i) > 0) {
24228 if (Shuffle->getMaskElt(i) != (int)(i / ZextRatio)) {
24229 // Expected element number
24239 // Ok, perform the transformation - replace the shuffle with
24240 // a shuffle of the form <0, k, k, k, 1, k, k, k> with zero
24241 // (instead of undef) where the k elements come from the zero vector.
24242 SmallVector<int, 8> Mask;
24243 unsigned NumElems = SrcType.getVectorNumElements();
24244 for (unsigned i = 0; i < NumElems; ++i)
24246 Mask.push_back(NumElems);
24248 Mask.push_back(i / ZextRatio);
24250 SDValue NewShuffle = DAG.getVectorShuffle(Shuffle->getValueType(0), DL,
24251 Shuffle->getOperand(0), DAG.getConstant(0, DL, SrcType), Mask);
24252 return DAG.getBitcast(N0.getValueType(), NewShuffle);
24255 static SDValue PerformAndCombine(SDNode *N, SelectionDAG &DAG,
24256 TargetLowering::DAGCombinerInfo &DCI,
24257 const X86Subtarget *Subtarget) {
24258 if (DCI.isBeforeLegalizeOps())
24261 if (SDValue Zext = VectorZextCombine(N, DAG, DCI, Subtarget))
24264 if (SDValue R = CMPEQCombine(N, DAG, DCI, Subtarget))
24267 EVT VT = N->getValueType(0);
24268 SDValue N0 = N->getOperand(0);
24269 SDValue N1 = N->getOperand(1);
24272 // Create BEXTR instructions
24273 // BEXTR is ((X >> imm) & (2**size-1))
24274 if (VT == MVT::i32 || VT == MVT::i64) {
24275 // Check for BEXTR.
24276 if ((Subtarget->hasBMI() || Subtarget->hasTBM()) &&
24277 (N0.getOpcode() == ISD::SRA || N0.getOpcode() == ISD::SRL)) {
24278 ConstantSDNode *MaskNode = dyn_cast<ConstantSDNode>(N1);
24279 ConstantSDNode *ShiftNode = dyn_cast<ConstantSDNode>(N0.getOperand(1));
24280 if (MaskNode && ShiftNode) {
24281 uint64_t Mask = MaskNode->getZExtValue();
24282 uint64_t Shift = ShiftNode->getZExtValue();
24283 if (isMask_64(Mask)) {
24284 uint64_t MaskSize = countPopulation(Mask);
24285 if (Shift + MaskSize <= VT.getSizeInBits())
24286 return DAG.getNode(X86ISD::BEXTR, DL, VT, N0.getOperand(0),
24287 DAG.getConstant(Shift | (MaskSize << 8), DL,
24296 // Want to form ANDNP nodes:
24297 // 1) In the hopes of then easily combining them with OR and AND nodes
24298 // to form PBLEND/PSIGN.
24299 // 2) To match ANDN packed intrinsics
24300 if (VT != MVT::v2i64 && VT != MVT::v4i64)
24303 // Check LHS for vnot
24304 if (N0.getOpcode() == ISD::XOR &&
24305 //ISD::isBuildVectorAllOnes(N0.getOperand(1).getNode()))
24306 CanFoldXORWithAllOnes(N0.getOperand(1).getNode()))
24307 return DAG.getNode(X86ISD::ANDNP, DL, VT, N0.getOperand(0), N1);
24309 // Check RHS for vnot
24310 if (N1.getOpcode() == ISD::XOR &&
24311 //ISD::isBuildVectorAllOnes(N1.getOperand(1).getNode()))
24312 CanFoldXORWithAllOnes(N1.getOperand(1).getNode()))
24313 return DAG.getNode(X86ISD::ANDNP, DL, VT, N1.getOperand(0), N0);
24318 static SDValue PerformOrCombine(SDNode *N, SelectionDAG &DAG,
24319 TargetLowering::DAGCombinerInfo &DCI,
24320 const X86Subtarget *Subtarget) {
24321 if (DCI.isBeforeLegalizeOps())
24324 if (SDValue R = CMPEQCombine(N, DAG, DCI, Subtarget))
24327 SDValue N0 = N->getOperand(0);
24328 SDValue N1 = N->getOperand(1);
24329 EVT VT = N->getValueType(0);
24331 // look for psign/blend
24332 if (VT == MVT::v2i64 || VT == MVT::v4i64) {
24333 if (!Subtarget->hasSSSE3() ||
24334 (VT == MVT::v4i64 && !Subtarget->hasInt256()))
24337 // Canonicalize pandn to RHS
24338 if (N0.getOpcode() == X86ISD::ANDNP)
24340 // or (and (m, y), (pandn m, x))
24341 if (N0.getOpcode() == ISD::AND && N1.getOpcode() == X86ISD::ANDNP) {
24342 SDValue Mask = N1.getOperand(0);
24343 SDValue X = N1.getOperand(1);
24345 if (N0.getOperand(0) == Mask)
24346 Y = N0.getOperand(1);
24347 if (N0.getOperand(1) == Mask)
24348 Y = N0.getOperand(0);
24350 // Check to see if the mask appeared in both the AND and ANDNP and
24354 // Validate that X, Y, and Mask are BIT_CONVERTS, and see through them.
24355 // Look through mask bitcast.
24356 if (Mask.getOpcode() == ISD::BITCAST)
24357 Mask = Mask.getOperand(0);
24358 if (X.getOpcode() == ISD::BITCAST)
24359 X = X.getOperand(0);
24360 if (Y.getOpcode() == ISD::BITCAST)
24361 Y = Y.getOperand(0);
24363 EVT MaskVT = Mask.getValueType();
24365 // Validate that the Mask operand is a vector sra node.
24366 // FIXME: what to do for bytes, since there is a psignb/pblendvb, but
24367 // there is no psrai.b
24368 unsigned EltBits = MaskVT.getVectorElementType().getSizeInBits();
24369 unsigned SraAmt = ~0;
24370 if (Mask.getOpcode() == ISD::SRA) {
24371 if (auto *AmtBV = dyn_cast<BuildVectorSDNode>(Mask.getOperand(1)))
24372 if (auto *AmtConst = AmtBV->getConstantSplatNode())
24373 SraAmt = AmtConst->getZExtValue();
24374 } else if (Mask.getOpcode() == X86ISD::VSRAI) {
24375 SDValue SraC = Mask.getOperand(1);
24376 SraAmt = cast<ConstantSDNode>(SraC)->getZExtValue();
24378 if ((SraAmt + 1) != EltBits)
24383 // Now we know we at least have a plendvb with the mask val. See if
24384 // we can form a psignb/w/d.
24385 // psign = x.type == y.type == mask.type && y = sub(0, x);
24386 if (Y.getOpcode() == ISD::SUB && Y.getOperand(1) == X &&
24387 ISD::isBuildVectorAllZeros(Y.getOperand(0).getNode()) &&
24388 X.getValueType() == MaskVT && Y.getValueType() == MaskVT) {
24389 assert((EltBits == 8 || EltBits == 16 || EltBits == 32) &&
24390 "Unsupported VT for PSIGN");
24391 Mask = DAG.getNode(X86ISD::PSIGN, DL, MaskVT, X, Mask.getOperand(0));
24392 return DAG.getBitcast(VT, Mask);
24394 // PBLENDVB only available on SSE 4.1
24395 if (!Subtarget->hasSSE41())
24398 EVT BlendVT = (VT == MVT::v4i64) ? MVT::v32i8 : MVT::v16i8;
24400 X = DAG.getBitcast(BlendVT, X);
24401 Y = DAG.getBitcast(BlendVT, Y);
24402 Mask = DAG.getBitcast(BlendVT, Mask);
24403 Mask = DAG.getNode(ISD::VSELECT, DL, BlendVT, Mask, Y, X);
24404 return DAG.getBitcast(VT, Mask);
24408 if (VT != MVT::i16 && VT != MVT::i32 && VT != MVT::i64)
24411 // fold (or (x << c) | (y >> (64 - c))) ==> (shld64 x, y, c)
24412 bool OptForSize = DAG.getMachineFunction().getFunction()->optForSize();
24414 // SHLD/SHRD instructions have lower register pressure, but on some
24415 // platforms they have higher latency than the equivalent
24416 // series of shifts/or that would otherwise be generated.
24417 // Don't fold (or (x << c) | (y >> (64 - c))) if SHLD/SHRD instructions
24418 // have higher latencies and we are not optimizing for size.
24419 if (!OptForSize && Subtarget->isSHLDSlow())
24422 if (N0.getOpcode() == ISD::SRL && N1.getOpcode() == ISD::SHL)
24424 if (N0.getOpcode() != ISD::SHL || N1.getOpcode() != ISD::SRL)
24426 if (!N0.hasOneUse() || !N1.hasOneUse())
24429 SDValue ShAmt0 = N0.getOperand(1);
24430 if (ShAmt0.getValueType() != MVT::i8)
24432 SDValue ShAmt1 = N1.getOperand(1);
24433 if (ShAmt1.getValueType() != MVT::i8)
24435 if (ShAmt0.getOpcode() == ISD::TRUNCATE)
24436 ShAmt0 = ShAmt0.getOperand(0);
24437 if (ShAmt1.getOpcode() == ISD::TRUNCATE)
24438 ShAmt1 = ShAmt1.getOperand(0);
24441 unsigned Opc = X86ISD::SHLD;
24442 SDValue Op0 = N0.getOperand(0);
24443 SDValue Op1 = N1.getOperand(0);
24444 if (ShAmt0.getOpcode() == ISD::SUB) {
24445 Opc = X86ISD::SHRD;
24446 std::swap(Op0, Op1);
24447 std::swap(ShAmt0, ShAmt1);
24450 unsigned Bits = VT.getSizeInBits();
24451 if (ShAmt1.getOpcode() == ISD::SUB) {
24452 SDValue Sum = ShAmt1.getOperand(0);
24453 if (ConstantSDNode *SumC = dyn_cast<ConstantSDNode>(Sum)) {
24454 SDValue ShAmt1Op1 = ShAmt1.getOperand(1);
24455 if (ShAmt1Op1.getNode()->getOpcode() == ISD::TRUNCATE)
24456 ShAmt1Op1 = ShAmt1Op1.getOperand(0);
24457 if (SumC->getSExtValue() == Bits && ShAmt1Op1 == ShAmt0)
24458 return DAG.getNode(Opc, DL, VT,
24460 DAG.getNode(ISD::TRUNCATE, DL,
24463 } else if (ConstantSDNode *ShAmt1C = dyn_cast<ConstantSDNode>(ShAmt1)) {
24464 ConstantSDNode *ShAmt0C = dyn_cast<ConstantSDNode>(ShAmt0);
24466 ShAmt0C->getSExtValue() + ShAmt1C->getSExtValue() == Bits)
24467 return DAG.getNode(Opc, DL, VT,
24468 N0.getOperand(0), N1.getOperand(0),
24469 DAG.getNode(ISD::TRUNCATE, DL,
24476 // Generate NEG and CMOV for integer abs.
24477 static SDValue performIntegerAbsCombine(SDNode *N, SelectionDAG &DAG) {
24478 EVT VT = N->getValueType(0);
24480 // Since X86 does not have CMOV for 8-bit integer, we don't convert
24481 // 8-bit integer abs to NEG and CMOV.
24482 if (VT.isInteger() && VT.getSizeInBits() == 8)
24485 SDValue N0 = N->getOperand(0);
24486 SDValue N1 = N->getOperand(1);
24489 // Check pattern of XOR(ADD(X,Y), Y) where Y is SRA(X, size(X)-1)
24490 // and change it to SUB and CMOV.
24491 if (VT.isInteger() && N->getOpcode() == ISD::XOR &&
24492 N0.getOpcode() == ISD::ADD &&
24493 N0.getOperand(1) == N1 &&
24494 N1.getOpcode() == ISD::SRA &&
24495 N1.getOperand(0) == N0.getOperand(0))
24496 if (ConstantSDNode *Y1C = dyn_cast<ConstantSDNode>(N1.getOperand(1)))
24497 if (Y1C->getAPIntValue() == VT.getSizeInBits()-1) {
24498 // Generate SUB & CMOV.
24499 SDValue Neg = DAG.getNode(X86ISD::SUB, DL, DAG.getVTList(VT, MVT::i32),
24500 DAG.getConstant(0, DL, VT), N0.getOperand(0));
24502 SDValue Ops[] = { N0.getOperand(0), Neg,
24503 DAG.getConstant(X86::COND_GE, DL, MVT::i8),
24504 SDValue(Neg.getNode(), 1) };
24505 return DAG.getNode(X86ISD::CMOV, DL, DAG.getVTList(VT, MVT::Glue), Ops);
24510 // Try to turn tests against the signbit in the form of:
24511 // XOR(TRUNCATE(SRL(X, size(X)-1)), 1)
24514 static SDValue foldXorTruncShiftIntoCmp(SDNode *N, SelectionDAG &DAG) {
24515 // This is only worth doing if the output type is i8.
24516 if (N->getValueType(0) != MVT::i8)
24519 SDValue N0 = N->getOperand(0);
24520 SDValue N1 = N->getOperand(1);
24522 // We should be performing an xor against a truncated shift.
24523 if (N0.getOpcode() != ISD::TRUNCATE || !N0.hasOneUse())
24526 // Make sure we are performing an xor against one.
24527 if (!isa<ConstantSDNode>(N1) || !cast<ConstantSDNode>(N1)->isOne())
24530 // SetCC on x86 zero extends so only act on this if it's a logical shift.
24531 SDValue Shift = N0.getOperand(0);
24532 if (Shift.getOpcode() != ISD::SRL || !Shift.hasOneUse())
24535 // Make sure we are truncating from one of i16, i32 or i64.
24536 EVT ShiftTy = Shift.getValueType();
24537 if (ShiftTy != MVT::i16 && ShiftTy != MVT::i32 && ShiftTy != MVT::i64)
24540 // Make sure the shift amount extracts the sign bit.
24541 if (!isa<ConstantSDNode>(Shift.getOperand(1)) ||
24542 Shift.getConstantOperandVal(1) != ShiftTy.getSizeInBits() - 1)
24545 // Create a greater-than comparison against -1.
24546 // N.B. Using SETGE against 0 works but we want a canonical looking
24547 // comparison, using SETGT matches up with what TranslateX86CC.
24549 SDValue ShiftOp = Shift.getOperand(0);
24550 EVT ShiftOpTy = ShiftOp.getValueType();
24551 SDValue Cond = DAG.getSetCC(DL, MVT::i8, ShiftOp,
24552 DAG.getConstant(-1, DL, ShiftOpTy), ISD::SETGT);
24556 static SDValue PerformXorCombine(SDNode *N, SelectionDAG &DAG,
24557 TargetLowering::DAGCombinerInfo &DCI,
24558 const X86Subtarget *Subtarget) {
24559 if (DCI.isBeforeLegalizeOps())
24562 if (SDValue RV = foldXorTruncShiftIntoCmp(N, DAG))
24565 if (Subtarget->hasCMov())
24566 if (SDValue RV = performIntegerAbsCombine(N, DAG))
24572 /// PerformLOADCombine - Do target-specific dag combines on LOAD nodes.
24573 static SDValue PerformLOADCombine(SDNode *N, SelectionDAG &DAG,
24574 TargetLowering::DAGCombinerInfo &DCI,
24575 const X86Subtarget *Subtarget) {
24576 LoadSDNode *Ld = cast<LoadSDNode>(N);
24577 EVT RegVT = Ld->getValueType(0);
24578 EVT MemVT = Ld->getMemoryVT();
24580 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
24582 // For chips with slow 32-byte unaligned loads, break the 32-byte operation
24583 // into two 16-byte operations.
24584 ISD::LoadExtType Ext = Ld->getExtensionType();
24586 unsigned AddressSpace = Ld->getAddressSpace();
24587 unsigned Alignment = Ld->getAlignment();
24588 if (RegVT.is256BitVector() && !DCI.isBeforeLegalizeOps() &&
24589 Ext == ISD::NON_EXTLOAD &&
24590 TLI.allowsMemoryAccess(*DAG.getContext(), DAG.getDataLayout(), RegVT,
24591 AddressSpace, Alignment, &Fast) && !Fast) {
24592 unsigned NumElems = RegVT.getVectorNumElements();
24596 SDValue Ptr = Ld->getBasePtr();
24597 SDValue Increment =
24598 DAG.getConstant(16, dl, TLI.getPointerTy(DAG.getDataLayout()));
24600 EVT HalfVT = EVT::getVectorVT(*DAG.getContext(), MemVT.getScalarType(),
24602 SDValue Load1 = DAG.getLoad(HalfVT, dl, Ld->getChain(), Ptr,
24603 Ld->getPointerInfo(), Ld->isVolatile(),
24604 Ld->isNonTemporal(), Ld->isInvariant(),
24606 Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr, Increment);
24607 SDValue Load2 = DAG.getLoad(HalfVT, dl, Ld->getChain(), Ptr,
24608 Ld->getPointerInfo(), Ld->isVolatile(),
24609 Ld->isNonTemporal(), Ld->isInvariant(),
24610 std::min(16U, Alignment));
24611 SDValue TF = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
24613 Load2.getValue(1));
24615 SDValue NewVec = DAG.getUNDEF(RegVT);
24616 NewVec = Insert128BitVector(NewVec, Load1, 0, DAG, dl);
24617 NewVec = Insert128BitVector(NewVec, Load2, NumElems/2, DAG, dl);
24618 return DCI.CombineTo(N, NewVec, TF, true);
24624 /// PerformMLOADCombine - Resolve extending loads
24625 static SDValue PerformMLOADCombine(SDNode *N, SelectionDAG &DAG,
24626 TargetLowering::DAGCombinerInfo &DCI,
24627 const X86Subtarget *Subtarget) {
24628 MaskedLoadSDNode *Mld = cast<MaskedLoadSDNode>(N);
24629 if (Mld->getExtensionType() != ISD::SEXTLOAD)
24632 EVT VT = Mld->getValueType(0);
24633 unsigned NumElems = VT.getVectorNumElements();
24634 EVT LdVT = Mld->getMemoryVT();
24637 assert(LdVT != VT && "Cannot extend to the same type");
24638 unsigned ToSz = VT.getVectorElementType().getSizeInBits();
24639 unsigned FromSz = LdVT.getVectorElementType().getSizeInBits();
24640 // From, To sizes and ElemCount must be pow of two
24641 assert (isPowerOf2_32(NumElems * FromSz * ToSz) &&
24642 "Unexpected size for extending masked load");
24644 unsigned SizeRatio = ToSz / FromSz;
24645 assert(SizeRatio * NumElems * FromSz == VT.getSizeInBits());
24647 // Create a type on which we perform the shuffle
24648 EVT WideVecVT = EVT::getVectorVT(*DAG.getContext(),
24649 LdVT.getScalarType(), NumElems*SizeRatio);
24650 assert(WideVecVT.getSizeInBits() == VT.getSizeInBits());
24652 // Convert Src0 value
24653 SDValue WideSrc0 = DAG.getBitcast(WideVecVT, Mld->getSrc0());
24654 if (Mld->getSrc0().getOpcode() != ISD::UNDEF) {
24655 SmallVector<int, 16> ShuffleVec(NumElems * SizeRatio, -1);
24656 for (unsigned i = 0; i != NumElems; ++i)
24657 ShuffleVec[i] = i * SizeRatio;
24659 // Can't shuffle using an illegal type.
24660 assert (DAG.getTargetLoweringInfo().isTypeLegal(WideVecVT)
24661 && "WideVecVT should be legal");
24662 WideSrc0 = DAG.getVectorShuffle(WideVecVT, dl, WideSrc0,
24663 DAG.getUNDEF(WideVecVT), &ShuffleVec[0]);
24665 // Prepare the new mask
24667 SDValue Mask = Mld->getMask();
24668 if (Mask.getValueType() == VT) {
24669 // Mask and original value have the same type
24670 NewMask = DAG.getBitcast(WideVecVT, Mask);
24671 SmallVector<int, 16> ShuffleVec(NumElems * SizeRatio, -1);
24672 for (unsigned i = 0; i != NumElems; ++i)
24673 ShuffleVec[i] = i * SizeRatio;
24674 for (unsigned i = NumElems; i != NumElems*SizeRatio; ++i)
24675 ShuffleVec[i] = NumElems*SizeRatio;
24676 NewMask = DAG.getVectorShuffle(WideVecVT, dl, NewMask,
24677 DAG.getConstant(0, dl, WideVecVT),
24681 assert(Mask.getValueType().getVectorElementType() == MVT::i1);
24682 unsigned WidenNumElts = NumElems*SizeRatio;
24683 unsigned MaskNumElts = VT.getVectorNumElements();
24684 EVT NewMaskVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
24687 unsigned NumConcat = WidenNumElts / MaskNumElts;
24688 SmallVector<SDValue, 16> Ops(NumConcat);
24689 SDValue ZeroVal = DAG.getConstant(0, dl, Mask.getValueType());
24691 for (unsigned i = 1; i != NumConcat; ++i)
24694 NewMask = DAG.getNode(ISD::CONCAT_VECTORS, dl, NewMaskVT, Ops);
24697 SDValue WideLd = DAG.getMaskedLoad(WideVecVT, dl, Mld->getChain(),
24698 Mld->getBasePtr(), NewMask, WideSrc0,
24699 Mld->getMemoryVT(), Mld->getMemOperand(),
24701 SDValue NewVec = DAG.getNode(X86ISD::VSEXT, dl, VT, WideLd);
24702 return DCI.CombineTo(N, NewVec, WideLd.getValue(1), true);
24705 /// PerformMSTORECombine - Resolve truncating stores
24706 static SDValue PerformMSTORECombine(SDNode *N, SelectionDAG &DAG,
24707 const X86Subtarget *Subtarget) {
24708 MaskedStoreSDNode *Mst = cast<MaskedStoreSDNode>(N);
24709 if (!Mst->isTruncatingStore())
24712 EVT VT = Mst->getValue().getValueType();
24713 unsigned NumElems = VT.getVectorNumElements();
24714 EVT StVT = Mst->getMemoryVT();
24717 assert(StVT != VT && "Cannot truncate to the same type");
24718 unsigned FromSz = VT.getVectorElementType().getSizeInBits();
24719 unsigned ToSz = StVT.getVectorElementType().getSizeInBits();
24721 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
24723 // The truncating store is legal in some cases. For example
24724 // vpmovqb, vpmovqw, vpmovqd, vpmovdb, vpmovdw
24725 // are designated for truncate store.
24726 // In this case we don't need any further transformations.
24727 if (TLI.isTruncStoreLegal(VT, StVT))
24730 // From, To sizes and ElemCount must be pow of two
24731 assert (isPowerOf2_32(NumElems * FromSz * ToSz) &&
24732 "Unexpected size for truncating masked store");
24733 // We are going to use the original vector elt for storing.
24734 // Accumulated smaller vector elements must be a multiple of the store size.
24735 assert (((NumElems * FromSz) % ToSz) == 0 &&
24736 "Unexpected ratio for truncating masked store");
24738 unsigned SizeRatio = FromSz / ToSz;
24739 assert(SizeRatio * NumElems * ToSz == VT.getSizeInBits());
24741 // Create a type on which we perform the shuffle
24742 EVT WideVecVT = EVT::getVectorVT(*DAG.getContext(),
24743 StVT.getScalarType(), NumElems*SizeRatio);
24745 assert(WideVecVT.getSizeInBits() == VT.getSizeInBits());
24747 SDValue WideVec = DAG.getBitcast(WideVecVT, Mst->getValue());
24748 SmallVector<int, 16> ShuffleVec(NumElems * SizeRatio, -1);
24749 for (unsigned i = 0; i != NumElems; ++i)
24750 ShuffleVec[i] = i * SizeRatio;
24752 // Can't shuffle using an illegal type.
24753 assert (DAG.getTargetLoweringInfo().isTypeLegal(WideVecVT)
24754 && "WideVecVT should be legal");
24756 SDValue TruncatedVal = DAG.getVectorShuffle(WideVecVT, dl, WideVec,
24757 DAG.getUNDEF(WideVecVT),
24761 SDValue Mask = Mst->getMask();
24762 if (Mask.getValueType() == VT) {
24763 // Mask and original value have the same type
24764 NewMask = DAG.getBitcast(WideVecVT, Mask);
24765 for (unsigned i = 0; i != NumElems; ++i)
24766 ShuffleVec[i] = i * SizeRatio;
24767 for (unsigned i = NumElems; i != NumElems*SizeRatio; ++i)
24768 ShuffleVec[i] = NumElems*SizeRatio;
24769 NewMask = DAG.getVectorShuffle(WideVecVT, dl, NewMask,
24770 DAG.getConstant(0, dl, WideVecVT),
24774 assert(Mask.getValueType().getVectorElementType() == MVT::i1);
24775 unsigned WidenNumElts = NumElems*SizeRatio;
24776 unsigned MaskNumElts = VT.getVectorNumElements();
24777 EVT NewMaskVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
24780 unsigned NumConcat = WidenNumElts / MaskNumElts;
24781 SmallVector<SDValue, 16> Ops(NumConcat);
24782 SDValue ZeroVal = DAG.getConstant(0, dl, Mask.getValueType());
24784 for (unsigned i = 1; i != NumConcat; ++i)
24787 NewMask = DAG.getNode(ISD::CONCAT_VECTORS, dl, NewMaskVT, Ops);
24790 return DAG.getMaskedStore(Mst->getChain(), dl, TruncatedVal, Mst->getBasePtr(),
24791 NewMask, StVT, Mst->getMemOperand(), false);
24793 /// PerformSTORECombine - Do target-specific dag combines on STORE nodes.
24794 static SDValue PerformSTORECombine(SDNode *N, SelectionDAG &DAG,
24795 const X86Subtarget *Subtarget) {
24796 StoreSDNode *St = cast<StoreSDNode>(N);
24797 EVT VT = St->getValue().getValueType();
24798 EVT StVT = St->getMemoryVT();
24800 SDValue StoredVal = St->getOperand(1);
24801 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
24803 // If we are saving a concatenation of two XMM registers and 32-byte stores
24804 // are slow, such as on Sandy Bridge, perform two 16-byte stores.
24806 unsigned AddressSpace = St->getAddressSpace();
24807 unsigned Alignment = St->getAlignment();
24808 if (VT.is256BitVector() && StVT == VT &&
24809 TLI.allowsMemoryAccess(*DAG.getContext(), DAG.getDataLayout(), VT,
24810 AddressSpace, Alignment, &Fast) && !Fast) {
24811 unsigned NumElems = VT.getVectorNumElements();
24815 SDValue Value0 = Extract128BitVector(StoredVal, 0, DAG, dl);
24816 SDValue Value1 = Extract128BitVector(StoredVal, NumElems/2, DAG, dl);
24819 DAG.getConstant(16, dl, TLI.getPointerTy(DAG.getDataLayout()));
24820 SDValue Ptr0 = St->getBasePtr();
24821 SDValue Ptr1 = DAG.getNode(ISD::ADD, dl, Ptr0.getValueType(), Ptr0, Stride);
24823 SDValue Ch0 = DAG.getStore(St->getChain(), dl, Value0, Ptr0,
24824 St->getPointerInfo(), St->isVolatile(),
24825 St->isNonTemporal(), Alignment);
24826 SDValue Ch1 = DAG.getStore(St->getChain(), dl, Value1, Ptr1,
24827 St->getPointerInfo(), St->isVolatile(),
24828 St->isNonTemporal(),
24829 std::min(16U, Alignment));
24830 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Ch0, Ch1);
24833 // Optimize trunc store (of multiple scalars) to shuffle and store.
24834 // First, pack all of the elements in one place. Next, store to memory
24835 // in fewer chunks.
24836 if (St->isTruncatingStore() && VT.isVector()) {
24837 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
24838 unsigned NumElems = VT.getVectorNumElements();
24839 assert(StVT != VT && "Cannot truncate to the same type");
24840 unsigned FromSz = VT.getVectorElementType().getSizeInBits();
24841 unsigned ToSz = StVT.getVectorElementType().getSizeInBits();
24843 // The truncating store is legal in some cases. For example
24844 // vpmovqb, vpmovqw, vpmovqd, vpmovdb, vpmovdw
24845 // are designated for truncate store.
24846 // In this case we don't need any further transformations.
24847 if (TLI.isTruncStoreLegal(VT, StVT))
24850 // From, To sizes and ElemCount must be pow of two
24851 if (!isPowerOf2_32(NumElems * FromSz * ToSz)) return SDValue();
24852 // We are going to use the original vector elt for storing.
24853 // Accumulated smaller vector elements must be a multiple of the store size.
24854 if (0 != (NumElems * FromSz) % ToSz) return SDValue();
24856 unsigned SizeRatio = FromSz / ToSz;
24858 assert(SizeRatio * NumElems * ToSz == VT.getSizeInBits());
24860 // Create a type on which we perform the shuffle
24861 EVT WideVecVT = EVT::getVectorVT(*DAG.getContext(),
24862 StVT.getScalarType(), NumElems*SizeRatio);
24864 assert(WideVecVT.getSizeInBits() == VT.getSizeInBits());
24866 SDValue WideVec = DAG.getBitcast(WideVecVT, St->getValue());
24867 SmallVector<int, 8> ShuffleVec(NumElems * SizeRatio, -1);
24868 for (unsigned i = 0; i != NumElems; ++i)
24869 ShuffleVec[i] = i * SizeRatio;
24871 // Can't shuffle using an illegal type.
24872 if (!TLI.isTypeLegal(WideVecVT))
24875 SDValue Shuff = DAG.getVectorShuffle(WideVecVT, dl, WideVec,
24876 DAG.getUNDEF(WideVecVT),
24878 // At this point all of the data is stored at the bottom of the
24879 // register. We now need to save it to mem.
24881 // Find the largest store unit
24882 MVT StoreType = MVT::i8;
24883 for (MVT Tp : MVT::integer_valuetypes()) {
24884 if (TLI.isTypeLegal(Tp) && Tp.getSizeInBits() <= NumElems * ToSz)
24888 // On 32bit systems, we can't save 64bit integers. Try bitcasting to F64.
24889 if (TLI.isTypeLegal(MVT::f64) && StoreType.getSizeInBits() < 64 &&
24890 (64 <= NumElems * ToSz))
24891 StoreType = MVT::f64;
24893 // Bitcast the original vector into a vector of store-size units
24894 EVT StoreVecVT = EVT::getVectorVT(*DAG.getContext(),
24895 StoreType, VT.getSizeInBits()/StoreType.getSizeInBits());
24896 assert(StoreVecVT.getSizeInBits() == VT.getSizeInBits());
24897 SDValue ShuffWide = DAG.getBitcast(StoreVecVT, Shuff);
24898 SmallVector<SDValue, 8> Chains;
24899 SDValue Increment = DAG.getConstant(StoreType.getSizeInBits() / 8, dl,
24900 TLI.getPointerTy(DAG.getDataLayout()));
24901 SDValue Ptr = St->getBasePtr();
24903 // Perform one or more big stores into memory.
24904 for (unsigned i=0, e=(ToSz*NumElems)/StoreType.getSizeInBits(); i!=e; ++i) {
24905 SDValue SubVec = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl,
24906 StoreType, ShuffWide,
24907 DAG.getIntPtrConstant(i, dl));
24908 SDValue Ch = DAG.getStore(St->getChain(), dl, SubVec, Ptr,
24909 St->getPointerInfo(), St->isVolatile(),
24910 St->isNonTemporal(), St->getAlignment());
24911 Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr, Increment);
24912 Chains.push_back(Ch);
24915 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Chains);
24918 // Turn load->store of MMX types into GPR load/stores. This avoids clobbering
24919 // the FP state in cases where an emms may be missing.
24920 // A preferable solution to the general problem is to figure out the right
24921 // places to insert EMMS. This qualifies as a quick hack.
24923 // Similarly, turn load->store of i64 into double load/stores in 32-bit mode.
24924 if (VT.getSizeInBits() != 64)
24927 const Function *F = DAG.getMachineFunction().getFunction();
24928 bool NoImplicitFloatOps = F->hasFnAttribute(Attribute::NoImplicitFloat);
24930 !Subtarget->useSoftFloat() && !NoImplicitFloatOps && Subtarget->hasSSE2();
24931 if ((VT.isVector() ||
24932 (VT == MVT::i64 && F64IsLegal && !Subtarget->is64Bit())) &&
24933 isa<LoadSDNode>(St->getValue()) &&
24934 !cast<LoadSDNode>(St->getValue())->isVolatile() &&
24935 St->getChain().hasOneUse() && !St->isVolatile()) {
24936 SDNode* LdVal = St->getValue().getNode();
24937 LoadSDNode *Ld = nullptr;
24938 int TokenFactorIndex = -1;
24939 SmallVector<SDValue, 8> Ops;
24940 SDNode* ChainVal = St->getChain().getNode();
24941 // Must be a store of a load. We currently handle two cases: the load
24942 // is a direct child, and it's under an intervening TokenFactor. It is
24943 // possible to dig deeper under nested TokenFactors.
24944 if (ChainVal == LdVal)
24945 Ld = cast<LoadSDNode>(St->getChain());
24946 else if (St->getValue().hasOneUse() &&
24947 ChainVal->getOpcode() == ISD::TokenFactor) {
24948 for (unsigned i = 0, e = ChainVal->getNumOperands(); i != e; ++i) {
24949 if (ChainVal->getOperand(i).getNode() == LdVal) {
24950 TokenFactorIndex = i;
24951 Ld = cast<LoadSDNode>(St->getValue());
24953 Ops.push_back(ChainVal->getOperand(i));
24957 if (!Ld || !ISD::isNormalLoad(Ld))
24960 // If this is not the MMX case, i.e. we are just turning i64 load/store
24961 // into f64 load/store, avoid the transformation if there are multiple
24962 // uses of the loaded value.
24963 if (!VT.isVector() && !Ld->hasNUsesOfValue(1, 0))
24968 // If we are a 64-bit capable x86, lower to a single movq load/store pair.
24969 // Otherwise, if it's legal to use f64 SSE instructions, use f64 load/store
24971 if (Subtarget->is64Bit() || F64IsLegal) {
24972 EVT LdVT = Subtarget->is64Bit() ? MVT::i64 : MVT::f64;
24973 SDValue NewLd = DAG.getLoad(LdVT, LdDL, Ld->getChain(), Ld->getBasePtr(),
24974 Ld->getPointerInfo(), Ld->isVolatile(),
24975 Ld->isNonTemporal(), Ld->isInvariant(),
24976 Ld->getAlignment());
24977 SDValue NewChain = NewLd.getValue(1);
24978 if (TokenFactorIndex != -1) {
24979 Ops.push_back(NewChain);
24980 NewChain = DAG.getNode(ISD::TokenFactor, LdDL, MVT::Other, Ops);
24982 return DAG.getStore(NewChain, StDL, NewLd, St->getBasePtr(),
24983 St->getPointerInfo(),
24984 St->isVolatile(), St->isNonTemporal(),
24985 St->getAlignment());
24988 // Otherwise, lower to two pairs of 32-bit loads / stores.
24989 SDValue LoAddr = Ld->getBasePtr();
24990 SDValue HiAddr = DAG.getNode(ISD::ADD, LdDL, MVT::i32, LoAddr,
24991 DAG.getConstant(4, LdDL, MVT::i32));
24993 SDValue LoLd = DAG.getLoad(MVT::i32, LdDL, Ld->getChain(), LoAddr,
24994 Ld->getPointerInfo(),
24995 Ld->isVolatile(), Ld->isNonTemporal(),
24996 Ld->isInvariant(), Ld->getAlignment());
24997 SDValue HiLd = DAG.getLoad(MVT::i32, LdDL, Ld->getChain(), HiAddr,
24998 Ld->getPointerInfo().getWithOffset(4),
24999 Ld->isVolatile(), Ld->isNonTemporal(),
25001 MinAlign(Ld->getAlignment(), 4));
25003 SDValue NewChain = LoLd.getValue(1);
25004 if (TokenFactorIndex != -1) {
25005 Ops.push_back(LoLd);
25006 Ops.push_back(HiLd);
25007 NewChain = DAG.getNode(ISD::TokenFactor, LdDL, MVT::Other, Ops);
25010 LoAddr = St->getBasePtr();
25011 HiAddr = DAG.getNode(ISD::ADD, StDL, MVT::i32, LoAddr,
25012 DAG.getConstant(4, StDL, MVT::i32));
25014 SDValue LoSt = DAG.getStore(NewChain, StDL, LoLd, LoAddr,
25015 St->getPointerInfo(),
25016 St->isVolatile(), St->isNonTemporal(),
25017 St->getAlignment());
25018 SDValue HiSt = DAG.getStore(NewChain, StDL, HiLd, HiAddr,
25019 St->getPointerInfo().getWithOffset(4),
25021 St->isNonTemporal(),
25022 MinAlign(St->getAlignment(), 4));
25023 return DAG.getNode(ISD::TokenFactor, StDL, MVT::Other, LoSt, HiSt);
25026 // This is similar to the above case, but here we handle a scalar 64-bit
25027 // integer store that is extracted from a vector on a 32-bit target.
25028 // If we have SSE2, then we can treat it like a floating-point double
25029 // to get past legalization. The execution dependencies fixup pass will
25030 // choose the optimal machine instruction for the store if this really is
25031 // an integer or v2f32 rather than an f64.
25032 if (VT == MVT::i64 && F64IsLegal && !Subtarget->is64Bit() &&
25033 St->getOperand(1).getOpcode() == ISD::EXTRACT_VECTOR_ELT) {
25034 SDValue OldExtract = St->getOperand(1);
25035 SDValue ExtOp0 = OldExtract.getOperand(0);
25036 unsigned VecSize = ExtOp0.getValueSizeInBits();
25037 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), MVT::f64, VecSize / 64);
25038 SDValue BitCast = DAG.getBitcast(VecVT, ExtOp0);
25039 SDValue NewExtract = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
25040 BitCast, OldExtract.getOperand(1));
25041 return DAG.getStore(St->getChain(), dl, NewExtract, St->getBasePtr(),
25042 St->getPointerInfo(), St->isVolatile(),
25043 St->isNonTemporal(), St->getAlignment());
25049 /// Return 'true' if this vector operation is "horizontal"
25050 /// and return the operands for the horizontal operation in LHS and RHS. A
25051 /// horizontal operation performs the binary operation on successive elements
25052 /// of its first operand, then on successive elements of its second operand,
25053 /// returning the resulting values in a vector. For example, if
25054 /// A = < float a0, float a1, float a2, float a3 >
25056 /// B = < float b0, float b1, float b2, float b3 >
25057 /// then the result of doing a horizontal operation on A and B is
25058 /// A horizontal-op B = < a0 op a1, a2 op a3, b0 op b1, b2 op b3 >.
25059 /// In short, LHS and RHS are inspected to see if LHS op RHS is of the form
25060 /// A horizontal-op B, for some already available A and B, and if so then LHS is
25061 /// set to A, RHS to B, and the routine returns 'true'.
25062 /// Note that the binary operation should have the property that if one of the
25063 /// operands is UNDEF then the result is UNDEF.
25064 static bool isHorizontalBinOp(SDValue &LHS, SDValue &RHS, bool IsCommutative) {
25065 // Look for the following pattern: if
25066 // A = < float a0, float a1, float a2, float a3 >
25067 // B = < float b0, float b1, float b2, float b3 >
25069 // LHS = VECTOR_SHUFFLE A, B, <0, 2, 4, 6>
25070 // RHS = VECTOR_SHUFFLE A, B, <1, 3, 5, 7>
25071 // then LHS op RHS = < a0 op a1, a2 op a3, b0 op b1, b2 op b3 >
25072 // which is A horizontal-op B.
25074 // At least one of the operands should be a vector shuffle.
25075 if (LHS.getOpcode() != ISD::VECTOR_SHUFFLE &&
25076 RHS.getOpcode() != ISD::VECTOR_SHUFFLE)
25079 MVT VT = LHS.getSimpleValueType();
25081 assert((VT.is128BitVector() || VT.is256BitVector()) &&
25082 "Unsupported vector type for horizontal add/sub");
25084 // Handle 128 and 256-bit vector lengths. AVX defines horizontal add/sub to
25085 // operate independently on 128-bit lanes.
25086 unsigned NumElts = VT.getVectorNumElements();
25087 unsigned NumLanes = VT.getSizeInBits()/128;
25088 unsigned NumLaneElts = NumElts / NumLanes;
25089 assert((NumLaneElts % 2 == 0) &&
25090 "Vector type should have an even number of elements in each lane");
25091 unsigned HalfLaneElts = NumLaneElts/2;
25093 // View LHS in the form
25094 // LHS = VECTOR_SHUFFLE A, B, LMask
25095 // If LHS is not a shuffle then pretend it is the shuffle
25096 // LHS = VECTOR_SHUFFLE LHS, undef, <0, 1, ..., N-1>
25097 // NOTE: in what follows a default initialized SDValue represents an UNDEF of
25100 SmallVector<int, 16> LMask(NumElts);
25101 if (LHS.getOpcode() == ISD::VECTOR_SHUFFLE) {
25102 if (LHS.getOperand(0).getOpcode() != ISD::UNDEF)
25103 A = LHS.getOperand(0);
25104 if (LHS.getOperand(1).getOpcode() != ISD::UNDEF)
25105 B = LHS.getOperand(1);
25106 ArrayRef<int> Mask = cast<ShuffleVectorSDNode>(LHS.getNode())->getMask();
25107 std::copy(Mask.begin(), Mask.end(), LMask.begin());
25109 if (LHS.getOpcode() != ISD::UNDEF)
25111 for (unsigned i = 0; i != NumElts; ++i)
25115 // Likewise, view RHS in the form
25116 // RHS = VECTOR_SHUFFLE C, D, RMask
25118 SmallVector<int, 16> RMask(NumElts);
25119 if (RHS.getOpcode() == ISD::VECTOR_SHUFFLE) {
25120 if (RHS.getOperand(0).getOpcode() != ISD::UNDEF)
25121 C = RHS.getOperand(0);
25122 if (RHS.getOperand(1).getOpcode() != ISD::UNDEF)
25123 D = RHS.getOperand(1);
25124 ArrayRef<int> Mask = cast<ShuffleVectorSDNode>(RHS.getNode())->getMask();
25125 std::copy(Mask.begin(), Mask.end(), RMask.begin());
25127 if (RHS.getOpcode() != ISD::UNDEF)
25129 for (unsigned i = 0; i != NumElts; ++i)
25133 // Check that the shuffles are both shuffling the same vectors.
25134 if (!(A == C && B == D) && !(A == D && B == C))
25137 // If everything is UNDEF then bail out: it would be better to fold to UNDEF.
25138 if (!A.getNode() && !B.getNode())
25141 // If A and B occur in reverse order in RHS, then "swap" them (which means
25142 // rewriting the mask).
25144 ShuffleVectorSDNode::commuteMask(RMask);
25146 // At this point LHS and RHS are equivalent to
25147 // LHS = VECTOR_SHUFFLE A, B, LMask
25148 // RHS = VECTOR_SHUFFLE A, B, RMask
25149 // Check that the masks correspond to performing a horizontal operation.
25150 for (unsigned l = 0; l != NumElts; l += NumLaneElts) {
25151 for (unsigned i = 0; i != NumLaneElts; ++i) {
25152 int LIdx = LMask[i+l], RIdx = RMask[i+l];
25154 // Ignore any UNDEF components.
25155 if (LIdx < 0 || RIdx < 0 ||
25156 (!A.getNode() && (LIdx < (int)NumElts || RIdx < (int)NumElts)) ||
25157 (!B.getNode() && (LIdx >= (int)NumElts || RIdx >= (int)NumElts)))
25160 // Check that successive elements are being operated on. If not, this is
25161 // not a horizontal operation.
25162 unsigned Src = (i/HalfLaneElts); // each lane is split between srcs
25163 int Index = 2*(i%HalfLaneElts) + NumElts*Src + l;
25164 if (!(LIdx == Index && RIdx == Index + 1) &&
25165 !(IsCommutative && LIdx == Index + 1 && RIdx == Index))
25170 LHS = A.getNode() ? A : B; // If A is 'UNDEF', use B for it.
25171 RHS = B.getNode() ? B : A; // If B is 'UNDEF', use A for it.
25175 /// Do target-specific dag combines on floating point adds.
25176 static SDValue PerformFADDCombine(SDNode *N, SelectionDAG &DAG,
25177 const X86Subtarget *Subtarget) {
25178 EVT VT = N->getValueType(0);
25179 SDValue LHS = N->getOperand(0);
25180 SDValue RHS = N->getOperand(1);
25182 // Try to synthesize horizontal adds from adds of shuffles.
25183 if (((Subtarget->hasSSE3() && (VT == MVT::v4f32 || VT == MVT::v2f64)) ||
25184 (Subtarget->hasFp256() && (VT == MVT::v8f32 || VT == MVT::v4f64))) &&
25185 isHorizontalBinOp(LHS, RHS, true))
25186 return DAG.getNode(X86ISD::FHADD, SDLoc(N), VT, LHS, RHS);
25190 /// Do target-specific dag combines on floating point subs.
25191 static SDValue PerformFSUBCombine(SDNode *N, SelectionDAG &DAG,
25192 const X86Subtarget *Subtarget) {
25193 EVT VT = N->getValueType(0);
25194 SDValue LHS = N->getOperand(0);
25195 SDValue RHS = N->getOperand(1);
25197 // Try to synthesize horizontal subs from subs of shuffles.
25198 if (((Subtarget->hasSSE3() && (VT == MVT::v4f32 || VT == MVT::v2f64)) ||
25199 (Subtarget->hasFp256() && (VT == MVT::v8f32 || VT == MVT::v4f64))) &&
25200 isHorizontalBinOp(LHS, RHS, false))
25201 return DAG.getNode(X86ISD::FHSUB, SDLoc(N), VT, LHS, RHS);
25205 /// Do target-specific dag combines on X86ISD::FOR and X86ISD::FXOR nodes.
25206 static SDValue PerformFORCombine(SDNode *N, SelectionDAG &DAG,
25207 const X86Subtarget *Subtarget) {
25208 assert(N->getOpcode() == X86ISD::FOR || N->getOpcode() == X86ISD::FXOR);
25210 // F[X]OR(0.0, x) -> x
25211 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
25212 if (C->getValueAPF().isPosZero())
25213 return N->getOperand(1);
25215 // F[X]OR(x, 0.0) -> x
25216 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
25217 if (C->getValueAPF().isPosZero())
25218 return N->getOperand(0);
25220 EVT VT = N->getValueType(0);
25221 if (VT.is512BitVector() && !Subtarget->hasDQI()) {
25223 MVT IntScalar = MVT::getIntegerVT(VT.getScalarSizeInBits());
25224 MVT IntVT = MVT::getVectorVT(IntScalar, VT.getVectorNumElements());
25226 SDValue Op0 = DAG.getNode(ISD::BITCAST, dl, IntVT, N->getOperand(0));
25227 SDValue Op1 = DAG.getNode(ISD::BITCAST, dl, IntVT, N->getOperand(1));
25228 unsigned IntOpcode = (N->getOpcode() == X86ISD::FOR) ? ISD::OR : ISD::XOR;
25229 SDValue IntOp = DAG.getNode(IntOpcode, dl, IntVT, Op0, Op1);
25230 return DAG.getNode(ISD::BITCAST, dl, VT, IntOp);
25235 /// Do target-specific dag combines on X86ISD::FMIN and X86ISD::FMAX nodes.
25236 static SDValue PerformFMinFMaxCombine(SDNode *N, SelectionDAG &DAG) {
25237 assert(N->getOpcode() == X86ISD::FMIN || N->getOpcode() == X86ISD::FMAX);
25239 // Only perform optimizations if UnsafeMath is used.
25240 if (!DAG.getTarget().Options.UnsafeFPMath)
25243 // If we run in unsafe-math mode, then convert the FMAX and FMIN nodes
25244 // into FMINC and FMAXC, which are Commutative operations.
25245 unsigned NewOp = 0;
25246 switch (N->getOpcode()) {
25247 default: llvm_unreachable("unknown opcode");
25248 case X86ISD::FMIN: NewOp = X86ISD::FMINC; break;
25249 case X86ISD::FMAX: NewOp = X86ISD::FMAXC; break;
25252 return DAG.getNode(NewOp, SDLoc(N), N->getValueType(0),
25253 N->getOperand(0), N->getOperand(1));
25256 /// Do target-specific dag combines on X86ISD::FAND nodes.
25257 static SDValue PerformFANDCombine(SDNode *N, SelectionDAG &DAG) {
25258 // FAND(0.0, x) -> 0.0
25259 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
25260 if (C->getValueAPF().isPosZero())
25261 return N->getOperand(0);
25263 // FAND(x, 0.0) -> 0.0
25264 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
25265 if (C->getValueAPF().isPosZero())
25266 return N->getOperand(1);
25271 /// Do target-specific dag combines on X86ISD::FANDN nodes
25272 static SDValue PerformFANDNCombine(SDNode *N, SelectionDAG &DAG) {
25273 // FANDN(0.0, x) -> x
25274 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
25275 if (C->getValueAPF().isPosZero())
25276 return N->getOperand(1);
25278 // FANDN(x, 0.0) -> 0.0
25279 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
25280 if (C->getValueAPF().isPosZero())
25281 return N->getOperand(1);
25286 static SDValue PerformBTCombine(SDNode *N,
25288 TargetLowering::DAGCombinerInfo &DCI) {
25289 // BT ignores high bits in the bit index operand.
25290 SDValue Op1 = N->getOperand(1);
25291 if (Op1.hasOneUse()) {
25292 unsigned BitWidth = Op1.getValueSizeInBits();
25293 APInt DemandedMask = APInt::getLowBitsSet(BitWidth, Log2_32(BitWidth));
25294 APInt KnownZero, KnownOne;
25295 TargetLowering::TargetLoweringOpt TLO(DAG, !DCI.isBeforeLegalize(),
25296 !DCI.isBeforeLegalizeOps());
25297 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
25298 if (TLO.ShrinkDemandedConstant(Op1, DemandedMask) ||
25299 TLI.SimplifyDemandedBits(Op1, DemandedMask, KnownZero, KnownOne, TLO))
25300 DCI.CommitTargetLoweringOpt(TLO);
25305 static SDValue PerformVZEXT_MOVLCombine(SDNode *N, SelectionDAG &DAG) {
25306 SDValue Op = N->getOperand(0);
25307 if (Op.getOpcode() == ISD::BITCAST)
25308 Op = Op.getOperand(0);
25309 EVT VT = N->getValueType(0), OpVT = Op.getValueType();
25310 if (Op.getOpcode() == X86ISD::VZEXT_LOAD &&
25311 VT.getVectorElementType().getSizeInBits() ==
25312 OpVT.getVectorElementType().getSizeInBits()) {
25313 return DAG.getNode(ISD::BITCAST, SDLoc(N), VT, Op);
25318 static SDValue PerformSIGN_EXTEND_INREGCombine(SDNode *N, SelectionDAG &DAG,
25319 const X86Subtarget *Subtarget) {
25320 EVT VT = N->getValueType(0);
25321 if (!VT.isVector())
25324 SDValue N0 = N->getOperand(0);
25325 SDValue N1 = N->getOperand(1);
25326 EVT ExtraVT = cast<VTSDNode>(N1)->getVT();
25329 // The SIGN_EXTEND_INREG to v4i64 is expensive operation on the
25330 // both SSE and AVX2 since there is no sign-extended shift right
25331 // operation on a vector with 64-bit elements.
25332 //(sext_in_reg (v4i64 anyext (v4i32 x )), ExtraVT) ->
25333 // (v4i64 sext (v4i32 sext_in_reg (v4i32 x , ExtraVT)))
25334 if (VT == MVT::v4i64 && (N0.getOpcode() == ISD::ANY_EXTEND ||
25335 N0.getOpcode() == ISD::SIGN_EXTEND)) {
25336 SDValue N00 = N0.getOperand(0);
25338 // EXTLOAD has a better solution on AVX2,
25339 // it may be replaced with X86ISD::VSEXT node.
25340 if (N00.getOpcode() == ISD::LOAD && Subtarget->hasInt256())
25341 if (!ISD::isNormalLoad(N00.getNode()))
25344 if (N00.getValueType() == MVT::v4i32 && ExtraVT.getSizeInBits() < 128) {
25345 SDValue Tmp = DAG.getNode(ISD::SIGN_EXTEND_INREG, dl, MVT::v4i32,
25347 return DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i64, Tmp);
25353 static SDValue PerformSExtCombine(SDNode *N, SelectionDAG &DAG,
25354 TargetLowering::DAGCombinerInfo &DCI,
25355 const X86Subtarget *Subtarget) {
25356 SDValue N0 = N->getOperand(0);
25357 EVT VT = N->getValueType(0);
25358 EVT SVT = VT.getScalarType();
25359 EVT InVT = N0.getValueType();
25360 EVT InSVT = InVT.getScalarType();
25363 // (i8,i32 sext (sdivrem (i8 x, i8 y)) ->
25364 // (i8,i32 (sdivrem_sext_hreg (i8 x, i8 y)
25365 // This exposes the sext to the sdivrem lowering, so that it directly extends
25366 // from AH (which we otherwise need to do contortions to access).
25367 if (N0.getOpcode() == ISD::SDIVREM && N0.getResNo() == 1 &&
25368 InVT == MVT::i8 && VT == MVT::i32) {
25369 SDVTList NodeTys = DAG.getVTList(MVT::i8, VT);
25370 SDValue R = DAG.getNode(X86ISD::SDIVREM8_SEXT_HREG, DL, NodeTys,
25371 N0.getOperand(0), N0.getOperand(1));
25372 DAG.ReplaceAllUsesOfValueWith(N0.getValue(0), R.getValue(0));
25373 return R.getValue(1);
25376 if (!DCI.isBeforeLegalizeOps()) {
25377 if (InVT == MVT::i1) {
25378 SDValue Zero = DAG.getConstant(0, DL, VT);
25380 DAG.getConstant(APInt::getAllOnesValue(VT.getSizeInBits()), DL, VT);
25381 return DAG.getNode(ISD::SELECT, DL, VT, N0, AllOnes, Zero);
25386 if (VT.isVector() && Subtarget->hasSSE2()) {
25387 auto ExtendVecSize = [&DAG](SDLoc DL, SDValue N, unsigned Size) {
25388 EVT InVT = N.getValueType();
25389 EVT OutVT = EVT::getVectorVT(*DAG.getContext(), InVT.getScalarType(),
25390 Size / InVT.getScalarSizeInBits());
25391 SmallVector<SDValue, 8> Opnds(Size / InVT.getSizeInBits(),
25392 DAG.getUNDEF(InVT));
25394 return DAG.getNode(ISD::CONCAT_VECTORS, DL, OutVT, Opnds);
25397 // If target-size is less than 128-bits, extend to a type that would extend
25398 // to 128 bits, extend that and extract the original target vector.
25399 if (VT.getSizeInBits() < 128 && !(128 % VT.getSizeInBits()) &&
25400 (SVT == MVT::i64 || SVT == MVT::i32 || SVT == MVT::i16) &&
25401 (InSVT == MVT::i32 || InSVT == MVT::i16 || InSVT == MVT::i8)) {
25402 unsigned Scale = 128 / VT.getSizeInBits();
25404 EVT::getVectorVT(*DAG.getContext(), SVT, 128 / SVT.getSizeInBits());
25405 SDValue Ex = ExtendVecSize(DL, N0, Scale * InVT.getSizeInBits());
25406 SDValue SExt = DAG.getNode(ISD::SIGN_EXTEND, DL, ExVT, Ex);
25407 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, VT, SExt,
25408 DAG.getIntPtrConstant(0, DL));
25411 // If target-size is 128-bits, then convert to ISD::SIGN_EXTEND_VECTOR_INREG
25412 // which ensures lowering to X86ISD::VSEXT (pmovsx*).
25413 if (VT.getSizeInBits() == 128 &&
25414 (SVT == MVT::i64 || SVT == MVT::i32 || SVT == MVT::i16) &&
25415 (InSVT == MVT::i32 || InSVT == MVT::i16 || InSVT == MVT::i8)) {
25416 SDValue ExOp = ExtendVecSize(DL, N0, 128);
25417 return DAG.getSignExtendVectorInReg(ExOp, DL, VT);
25420 // On pre-AVX2 targets, split into 128-bit nodes of
25421 // ISD::SIGN_EXTEND_VECTOR_INREG.
25422 if (!Subtarget->hasInt256() && !(VT.getSizeInBits() % 128) &&
25423 (SVT == MVT::i64 || SVT == MVT::i32 || SVT == MVT::i16) &&
25424 (InSVT == MVT::i32 || InSVT == MVT::i16 || InSVT == MVT::i8)) {
25425 unsigned NumVecs = VT.getSizeInBits() / 128;
25426 unsigned NumSubElts = 128 / SVT.getSizeInBits();
25427 EVT SubVT = EVT::getVectorVT(*DAG.getContext(), SVT, NumSubElts);
25428 EVT InSubVT = EVT::getVectorVT(*DAG.getContext(), InSVT, NumSubElts);
25430 SmallVector<SDValue, 8> Opnds;
25431 for (unsigned i = 0, Offset = 0; i != NumVecs;
25432 ++i, Offset += NumSubElts) {
25433 SDValue SrcVec = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, InSubVT, N0,
25434 DAG.getIntPtrConstant(Offset, DL));
25435 SrcVec = ExtendVecSize(DL, SrcVec, 128);
25436 SrcVec = DAG.getSignExtendVectorInReg(SrcVec, DL, SubVT);
25437 Opnds.push_back(SrcVec);
25439 return DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, Opnds);
25443 if (!Subtarget->hasFp256())
25446 if (VT.isVector() && VT.getSizeInBits() == 256)
25447 if (SDValue R = WidenMaskArithmetic(N, DAG, DCI, Subtarget))
25453 static SDValue PerformFMACombine(SDNode *N, SelectionDAG &DAG,
25454 const X86Subtarget* Subtarget) {
25456 EVT VT = N->getValueType(0);
25458 // Let legalize expand this if it isn't a legal type yet.
25459 if (!DAG.getTargetLoweringInfo().isTypeLegal(VT))
25462 EVT ScalarVT = VT.getScalarType();
25463 if ((ScalarVT != MVT::f32 && ScalarVT != MVT::f64) ||
25464 (!Subtarget->hasFMA() && !Subtarget->hasFMA4() &&
25465 !Subtarget->hasAVX512()))
25468 SDValue A = N->getOperand(0);
25469 SDValue B = N->getOperand(1);
25470 SDValue C = N->getOperand(2);
25472 bool NegA = (A.getOpcode() == ISD::FNEG);
25473 bool NegB = (B.getOpcode() == ISD::FNEG);
25474 bool NegC = (C.getOpcode() == ISD::FNEG);
25476 // Negative multiplication when NegA xor NegB
25477 bool NegMul = (NegA != NegB);
25479 A = A.getOperand(0);
25481 B = B.getOperand(0);
25483 C = C.getOperand(0);
25487 Opcode = (!NegC) ? X86ISD::FMADD : X86ISD::FMSUB;
25489 Opcode = (!NegC) ? X86ISD::FNMADD : X86ISD::FNMSUB;
25491 return DAG.getNode(Opcode, dl, VT, A, B, C);
25494 static SDValue PerformZExtCombine(SDNode *N, SelectionDAG &DAG,
25495 TargetLowering::DAGCombinerInfo &DCI,
25496 const X86Subtarget *Subtarget) {
25497 // (i32 zext (and (i8 x86isd::setcc_carry), 1)) ->
25498 // (and (i32 x86isd::setcc_carry), 1)
25499 // This eliminates the zext. This transformation is necessary because
25500 // ISD::SETCC is always legalized to i8.
25502 SDValue N0 = N->getOperand(0);
25503 EVT VT = N->getValueType(0);
25505 if (N0.getOpcode() == ISD::AND &&
25507 N0.getOperand(0).hasOneUse()) {
25508 SDValue N00 = N0.getOperand(0);
25509 if (N00.getOpcode() == X86ISD::SETCC_CARRY) {
25510 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
25511 if (!C || C->getZExtValue() != 1)
25513 return DAG.getNode(ISD::AND, dl, VT,
25514 DAG.getNode(X86ISD::SETCC_CARRY, dl, VT,
25515 N00.getOperand(0), N00.getOperand(1)),
25516 DAG.getConstant(1, dl, VT));
25520 if (N0.getOpcode() == ISD::TRUNCATE &&
25522 N0.getOperand(0).hasOneUse()) {
25523 SDValue N00 = N0.getOperand(0);
25524 if (N00.getOpcode() == X86ISD::SETCC_CARRY) {
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 (VT.is256BitVector())
25533 if (SDValue R = WidenMaskArithmetic(N, DAG, DCI, Subtarget))
25536 // (i8,i32 zext (udivrem (i8 x, i8 y)) ->
25537 // (i8,i32 (udivrem_zext_hreg (i8 x, i8 y)
25538 // This exposes the zext to the udivrem lowering, so that it directly extends
25539 // from AH (which we otherwise need to do contortions to access).
25540 if (N0.getOpcode() == ISD::UDIVREM &&
25541 N0.getResNo() == 1 && N0.getValueType() == MVT::i8 &&
25542 (VT == MVT::i32 || VT == MVT::i64)) {
25543 SDVTList NodeTys = DAG.getVTList(MVT::i8, VT);
25544 SDValue R = DAG.getNode(X86ISD::UDIVREM8_ZEXT_HREG, dl, NodeTys,
25545 N0.getOperand(0), N0.getOperand(1));
25546 DAG.ReplaceAllUsesOfValueWith(N0.getValue(0), R.getValue(0));
25547 return R.getValue(1);
25553 // Optimize x == -y --> x+y == 0
25554 // x != -y --> x+y != 0
25555 static SDValue PerformISDSETCCCombine(SDNode *N, SelectionDAG &DAG,
25556 const X86Subtarget* Subtarget) {
25557 ISD::CondCode CC = cast<CondCodeSDNode>(N->getOperand(2))->get();
25558 SDValue LHS = N->getOperand(0);
25559 SDValue RHS = N->getOperand(1);
25560 EVT VT = N->getValueType(0);
25563 if ((CC == ISD::SETNE || CC == ISD::SETEQ) && LHS.getOpcode() == ISD::SUB)
25564 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(LHS.getOperand(0)))
25565 if (C->getAPIntValue() == 0 && LHS.hasOneUse()) {
25566 SDValue addV = DAG.getNode(ISD::ADD, DL, LHS.getValueType(), RHS,
25567 LHS.getOperand(1));
25568 return DAG.getSetCC(DL, N->getValueType(0), addV,
25569 DAG.getConstant(0, DL, addV.getValueType()), CC);
25571 if ((CC == ISD::SETNE || CC == ISD::SETEQ) && RHS.getOpcode() == ISD::SUB)
25572 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(RHS.getOperand(0)))
25573 if (C->getAPIntValue() == 0 && RHS.hasOneUse()) {
25574 SDValue addV = DAG.getNode(ISD::ADD, DL, RHS.getValueType(), LHS,
25575 RHS.getOperand(1));
25576 return DAG.getSetCC(DL, N->getValueType(0), addV,
25577 DAG.getConstant(0, DL, addV.getValueType()), CC);
25580 if (VT.getScalarType() == MVT::i1 &&
25581 (CC == ISD::SETNE || CC == ISD::SETEQ || ISD::isSignedIntSetCC(CC))) {
25583 (LHS.getOpcode() == ISD::SIGN_EXTEND) &&
25584 (LHS.getOperand(0).getValueType().getScalarType() == MVT::i1);
25585 bool IsVZero1 = ISD::isBuildVectorAllZeros(RHS.getNode());
25587 if (!IsSEXT0 || !IsVZero1) {
25588 // Swap the operands and update the condition code.
25589 std::swap(LHS, RHS);
25590 CC = ISD::getSetCCSwappedOperands(CC);
25592 IsSEXT0 = (LHS.getOpcode() == ISD::SIGN_EXTEND) &&
25593 (LHS.getOperand(0).getValueType().getScalarType() == MVT::i1);
25594 IsVZero1 = ISD::isBuildVectorAllZeros(RHS.getNode());
25597 if (IsSEXT0 && IsVZero1) {
25598 assert(VT == LHS.getOperand(0).getValueType() &&
25599 "Uexpected operand type");
25600 if (CC == ISD::SETGT)
25601 return DAG.getConstant(0, DL, VT);
25602 if (CC == ISD::SETLE)
25603 return DAG.getConstant(1, DL, VT);
25604 if (CC == ISD::SETEQ || CC == ISD::SETGE)
25605 return DAG.getNOT(DL, LHS.getOperand(0), VT);
25607 assert((CC == ISD::SETNE || CC == ISD::SETLT) &&
25608 "Unexpected condition code!");
25609 return LHS.getOperand(0);
25616 static SDValue NarrowVectorLoadToElement(LoadSDNode *Load, unsigned Index,
25617 SelectionDAG &DAG) {
25619 MVT VT = Load->getSimpleValueType(0);
25620 MVT EVT = VT.getVectorElementType();
25621 SDValue Addr = Load->getOperand(1);
25622 SDValue NewAddr = DAG.getNode(
25623 ISD::ADD, dl, Addr.getSimpleValueType(), Addr,
25624 DAG.getConstant(Index * EVT.getStoreSize(), dl,
25625 Addr.getSimpleValueType()));
25628 DAG.getLoad(EVT, dl, Load->getChain(), NewAddr,
25629 DAG.getMachineFunction().getMachineMemOperand(
25630 Load->getMemOperand(), 0, EVT.getStoreSize()));
25634 static SDValue PerformINSERTPSCombine(SDNode *N, SelectionDAG &DAG,
25635 const X86Subtarget *Subtarget) {
25637 MVT VT = N->getOperand(1)->getSimpleValueType(0);
25638 assert((VT == MVT::v4f32 || VT == MVT::v4i32) &&
25639 "X86insertps is only defined for v4x32");
25641 SDValue Ld = N->getOperand(1);
25642 if (MayFoldLoad(Ld)) {
25643 // Extract the countS bits from the immediate so we can get the proper
25644 // address when narrowing the vector load to a specific element.
25645 // When the second source op is a memory address, insertps doesn't use
25646 // countS and just gets an f32 from that address.
25647 unsigned DestIndex =
25648 cast<ConstantSDNode>(N->getOperand(2))->getZExtValue() >> 6;
25650 Ld = NarrowVectorLoadToElement(cast<LoadSDNode>(Ld), DestIndex, DAG);
25652 // Create this as a scalar to vector to match the instruction pattern.
25653 SDValue LoadScalarToVector = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Ld);
25654 // countS bits are ignored when loading from memory on insertps, which
25655 // means we don't need to explicitly set them to 0.
25656 return DAG.getNode(X86ISD::INSERTPS, dl, VT, N->getOperand(0),
25657 LoadScalarToVector, N->getOperand(2));
25662 static SDValue PerformBLENDICombine(SDNode *N, SelectionDAG &DAG) {
25663 SDValue V0 = N->getOperand(0);
25664 SDValue V1 = N->getOperand(1);
25666 EVT VT = N->getValueType(0);
25668 // Canonicalize a v2f64 blend with a mask of 2 by swapping the vector
25669 // operands and changing the mask to 1. This saves us a bunch of
25670 // pattern-matching possibilities related to scalar math ops in SSE/AVX.
25671 // x86InstrInfo knows how to commute this back after instruction selection
25672 // if it would help register allocation.
25674 // TODO: If optimizing for size or a processor that doesn't suffer from
25675 // partial register update stalls, this should be transformed into a MOVSD
25676 // instruction because a MOVSD is 1-2 bytes smaller than a BLENDPD.
25678 if (VT == MVT::v2f64)
25679 if (auto *Mask = dyn_cast<ConstantSDNode>(N->getOperand(2)))
25680 if (Mask->getZExtValue() == 2 && !isShuffleFoldableLoad(V0)) {
25681 SDValue NewMask = DAG.getConstant(1, DL, MVT::i8);
25682 return DAG.getNode(X86ISD::BLENDI, DL, VT, V1, V0, NewMask);
25688 // Helper function of PerformSETCCCombine. It is to materialize "setb reg"
25689 // as "sbb reg,reg", since it can be extended without zext and produces
25690 // an all-ones bit which is more useful than 0/1 in some cases.
25691 static SDValue MaterializeSETB(SDLoc DL, SDValue EFLAGS, SelectionDAG &DAG,
25694 return DAG.getNode(ISD::AND, DL, VT,
25695 DAG.getNode(X86ISD::SETCC_CARRY, DL, MVT::i8,
25696 DAG.getConstant(X86::COND_B, DL, MVT::i8),
25698 DAG.getConstant(1, DL, VT));
25699 assert (VT == MVT::i1 && "Unexpected type for SECCC node");
25700 return DAG.getNode(ISD::TRUNCATE, DL, MVT::i1,
25701 DAG.getNode(X86ISD::SETCC_CARRY, DL, MVT::i8,
25702 DAG.getConstant(X86::COND_B, DL, MVT::i8),
25706 // Optimize RES = X86ISD::SETCC CONDCODE, EFLAG_INPUT
25707 static SDValue PerformSETCCCombine(SDNode *N, SelectionDAG &DAG,
25708 TargetLowering::DAGCombinerInfo &DCI,
25709 const X86Subtarget *Subtarget) {
25711 X86::CondCode CC = X86::CondCode(N->getConstantOperandVal(0));
25712 SDValue EFLAGS = N->getOperand(1);
25714 if (CC == X86::COND_A) {
25715 // Try to convert COND_A into COND_B in an attempt to facilitate
25716 // materializing "setb reg".
25718 // Do not flip "e > c", where "c" is a constant, because Cmp instruction
25719 // cannot take an immediate as its first operand.
25721 if (EFLAGS.getOpcode() == X86ISD::SUB && EFLAGS.hasOneUse() &&
25722 EFLAGS.getValueType().isInteger() &&
25723 !isa<ConstantSDNode>(EFLAGS.getOperand(1))) {
25724 SDValue NewSub = DAG.getNode(X86ISD::SUB, SDLoc(EFLAGS),
25725 EFLAGS.getNode()->getVTList(),
25726 EFLAGS.getOperand(1), EFLAGS.getOperand(0));
25727 SDValue NewEFLAGS = SDValue(NewSub.getNode(), EFLAGS.getResNo());
25728 return MaterializeSETB(DL, NewEFLAGS, DAG, N->getSimpleValueType(0));
25732 // Materialize "setb reg" as "sbb reg,reg", since it can be extended without
25733 // a zext and produces an all-ones bit which is more useful than 0/1 in some
25735 if (CC == X86::COND_B)
25736 return MaterializeSETB(DL, EFLAGS, DAG, N->getSimpleValueType(0));
25738 if (SDValue Flags = checkBoolTestSetCCCombine(EFLAGS, CC)) {
25739 SDValue Cond = DAG.getConstant(CC, DL, MVT::i8);
25740 return DAG.getNode(X86ISD::SETCC, DL, N->getVTList(), Cond, Flags);
25746 // Optimize branch condition evaluation.
25748 static SDValue PerformBrCondCombine(SDNode *N, SelectionDAG &DAG,
25749 TargetLowering::DAGCombinerInfo &DCI,
25750 const X86Subtarget *Subtarget) {
25752 SDValue Chain = N->getOperand(0);
25753 SDValue Dest = N->getOperand(1);
25754 SDValue EFLAGS = N->getOperand(3);
25755 X86::CondCode CC = X86::CondCode(N->getConstantOperandVal(2));
25757 if (SDValue Flags = checkBoolTestSetCCCombine(EFLAGS, CC)) {
25758 SDValue Cond = DAG.getConstant(CC, DL, MVT::i8);
25759 return DAG.getNode(X86ISD::BRCOND, DL, N->getVTList(), Chain, Dest, Cond,
25766 static SDValue performVectorCompareAndMaskUnaryOpCombine(SDNode *N,
25767 SelectionDAG &DAG) {
25768 // Take advantage of vector comparisons producing 0 or -1 in each lane to
25769 // optimize away operation when it's from a constant.
25771 // The general transformation is:
25772 // UNARYOP(AND(VECTOR_CMP(x,y), constant)) -->
25773 // AND(VECTOR_CMP(x,y), constant2)
25774 // constant2 = UNARYOP(constant)
25776 // Early exit if this isn't a vector operation, the operand of the
25777 // unary operation isn't a bitwise AND, or if the sizes of the operations
25778 // aren't the same.
25779 EVT VT = N->getValueType(0);
25780 if (!VT.isVector() || N->getOperand(0)->getOpcode() != ISD::AND ||
25781 N->getOperand(0)->getOperand(0)->getOpcode() != ISD::SETCC ||
25782 VT.getSizeInBits() != N->getOperand(0)->getValueType(0).getSizeInBits())
25785 // Now check that the other operand of the AND is a constant. We could
25786 // make the transformation for non-constant splats as well, but it's unclear
25787 // that would be a benefit as it would not eliminate any operations, just
25788 // perform one more step in scalar code before moving to the vector unit.
25789 if (BuildVectorSDNode *BV =
25790 dyn_cast<BuildVectorSDNode>(N->getOperand(0)->getOperand(1))) {
25791 // Bail out if the vector isn't a constant.
25792 if (!BV->isConstant())
25795 // Everything checks out. Build up the new and improved node.
25797 EVT IntVT = BV->getValueType(0);
25798 // Create a new constant of the appropriate type for the transformed
25800 SDValue SourceConst = DAG.getNode(N->getOpcode(), DL, VT, SDValue(BV, 0));
25801 // The AND node needs bitcasts to/from an integer vector type around it.
25802 SDValue MaskConst = DAG.getBitcast(IntVT, SourceConst);
25803 SDValue NewAnd = DAG.getNode(ISD::AND, DL, IntVT,
25804 N->getOperand(0)->getOperand(0), MaskConst);
25805 SDValue Res = DAG.getBitcast(VT, NewAnd);
25812 static SDValue PerformUINT_TO_FPCombine(SDNode *N, SelectionDAG &DAG,
25813 const X86Subtarget *Subtarget) {
25814 SDValue Op0 = N->getOperand(0);
25815 EVT VT = N->getValueType(0);
25816 EVT InVT = Op0.getValueType();
25817 EVT InSVT = InVT.getScalarType();
25818 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
25820 // UINT_TO_FP(vXi8) -> SINT_TO_FP(ZEXT(vXi8 to vXi32))
25821 // UINT_TO_FP(vXi16) -> SINT_TO_FP(ZEXT(vXi16 to vXi32))
25822 if (InVT.isVector() && (InSVT == MVT::i8 || InSVT == MVT::i16)) {
25824 EVT DstVT = EVT::getVectorVT(*DAG.getContext(), MVT::i32,
25825 InVT.getVectorNumElements());
25826 SDValue P = DAG.getNode(ISD::ZERO_EXTEND, dl, DstVT, Op0);
25828 if (TLI.isOperationLegal(ISD::UINT_TO_FP, DstVT))
25829 return DAG.getNode(ISD::UINT_TO_FP, dl, VT, P);
25831 return DAG.getNode(ISD::SINT_TO_FP, dl, VT, P);
25837 static SDValue PerformSINT_TO_FPCombine(SDNode *N, SelectionDAG &DAG,
25838 const X86Subtarget *Subtarget) {
25839 // First try to optimize away the conversion entirely when it's
25840 // conditionally from a constant. Vectors only.
25841 if (SDValue Res = performVectorCompareAndMaskUnaryOpCombine(N, DAG))
25844 // Now move on to more general possibilities.
25845 SDValue Op0 = N->getOperand(0);
25846 EVT VT = N->getValueType(0);
25847 EVT InVT = Op0.getValueType();
25848 EVT InSVT = InVT.getScalarType();
25850 // SINT_TO_FP(vXi8) -> SINT_TO_FP(SEXT(vXi8 to vXi32))
25851 // SINT_TO_FP(vXi16) -> SINT_TO_FP(SEXT(vXi16 to vXi32))
25852 if (InVT.isVector() && (InSVT == MVT::i8 || InSVT == MVT::i16)) {
25854 EVT DstVT = EVT::getVectorVT(*DAG.getContext(), MVT::i32,
25855 InVT.getVectorNumElements());
25856 SDValue P = DAG.getNode(ISD::SIGN_EXTEND, dl, DstVT, Op0);
25857 return DAG.getNode(ISD::SINT_TO_FP, dl, VT, P);
25860 // Transform (SINT_TO_FP (i64 ...)) into an x87 operation if we have
25861 // a 32-bit target where SSE doesn't support i64->FP operations.
25862 if (Op0.getOpcode() == ISD::LOAD) {
25863 LoadSDNode *Ld = cast<LoadSDNode>(Op0.getNode());
25864 EVT LdVT = Ld->getValueType(0);
25866 // This transformation is not supported if the result type is f16
25867 if (VT == MVT::f16)
25870 if (!Ld->isVolatile() && !VT.isVector() &&
25871 ISD::isNON_EXTLoad(Op0.getNode()) && Op0.hasOneUse() &&
25872 !Subtarget->is64Bit() && LdVT == MVT::i64) {
25873 SDValue FILDChain = Subtarget->getTargetLowering()->BuildFILD(
25874 SDValue(N, 0), LdVT, Ld->getChain(), Op0, DAG);
25875 DAG.ReplaceAllUsesOfValueWith(Op0.getValue(1), FILDChain.getValue(1));
25882 // Optimize RES, EFLAGS = X86ISD::ADC LHS, RHS, EFLAGS
25883 static SDValue PerformADCCombine(SDNode *N, SelectionDAG &DAG,
25884 X86TargetLowering::DAGCombinerInfo &DCI) {
25885 // If the LHS and RHS of the ADC node are zero, then it can't overflow and
25886 // the result is either zero or one (depending on the input carry bit).
25887 // Strength reduce this down to a "set on carry" aka SETCC_CARRY&1.
25888 if (X86::isZeroNode(N->getOperand(0)) &&
25889 X86::isZeroNode(N->getOperand(1)) &&
25890 // We don't have a good way to replace an EFLAGS use, so only do this when
25892 SDValue(N, 1).use_empty()) {
25894 EVT VT = N->getValueType(0);
25895 SDValue CarryOut = DAG.getConstant(0, DL, N->getValueType(1));
25896 SDValue Res1 = DAG.getNode(ISD::AND, DL, VT,
25897 DAG.getNode(X86ISD::SETCC_CARRY, DL, VT,
25898 DAG.getConstant(X86::COND_B, DL,
25901 DAG.getConstant(1, DL, VT));
25902 return DCI.CombineTo(N, Res1, CarryOut);
25908 // fold (add Y, (sete X, 0)) -> adc 0, Y
25909 // (add Y, (setne X, 0)) -> sbb -1, Y
25910 // (sub (sete X, 0), Y) -> sbb 0, Y
25911 // (sub (setne X, 0), Y) -> adc -1, Y
25912 static SDValue OptimizeConditionalInDecrement(SDNode *N, SelectionDAG &DAG) {
25915 // Look through ZExts.
25916 SDValue Ext = N->getOperand(N->getOpcode() == ISD::SUB ? 1 : 0);
25917 if (Ext.getOpcode() != ISD::ZERO_EXTEND || !Ext.hasOneUse())
25920 SDValue SetCC = Ext.getOperand(0);
25921 if (SetCC.getOpcode() != X86ISD::SETCC || !SetCC.hasOneUse())
25924 X86::CondCode CC = (X86::CondCode)SetCC.getConstantOperandVal(0);
25925 if (CC != X86::COND_E && CC != X86::COND_NE)
25928 SDValue Cmp = SetCC.getOperand(1);
25929 if (Cmp.getOpcode() != X86ISD::CMP || !Cmp.hasOneUse() ||
25930 !X86::isZeroNode(Cmp.getOperand(1)) ||
25931 !Cmp.getOperand(0).getValueType().isInteger())
25934 SDValue CmpOp0 = Cmp.getOperand(0);
25935 SDValue NewCmp = DAG.getNode(X86ISD::CMP, DL, MVT::i32, CmpOp0,
25936 DAG.getConstant(1, DL, CmpOp0.getValueType()));
25938 SDValue OtherVal = N->getOperand(N->getOpcode() == ISD::SUB ? 0 : 1);
25939 if (CC == X86::COND_NE)
25940 return DAG.getNode(N->getOpcode() == ISD::SUB ? X86ISD::ADC : X86ISD::SBB,
25941 DL, OtherVal.getValueType(), OtherVal,
25942 DAG.getConstant(-1ULL, DL, OtherVal.getValueType()),
25944 return DAG.getNode(N->getOpcode() == ISD::SUB ? X86ISD::SBB : X86ISD::ADC,
25945 DL, OtherVal.getValueType(), OtherVal,
25946 DAG.getConstant(0, DL, OtherVal.getValueType()), NewCmp);
25949 /// PerformADDCombine - Do target-specific dag combines on integer adds.
25950 static SDValue PerformAddCombine(SDNode *N, SelectionDAG &DAG,
25951 const X86Subtarget *Subtarget) {
25952 EVT VT = N->getValueType(0);
25953 SDValue Op0 = N->getOperand(0);
25954 SDValue Op1 = N->getOperand(1);
25956 // Try to synthesize horizontal adds from adds of shuffles.
25957 if (((Subtarget->hasSSSE3() && (VT == MVT::v8i16 || VT == MVT::v4i32)) ||
25958 (Subtarget->hasInt256() && (VT == MVT::v16i16 || VT == MVT::v8i32))) &&
25959 isHorizontalBinOp(Op0, Op1, true))
25960 return DAG.getNode(X86ISD::HADD, SDLoc(N), VT, Op0, Op1);
25962 return OptimizeConditionalInDecrement(N, DAG);
25965 static SDValue PerformSubCombine(SDNode *N, SelectionDAG &DAG,
25966 const X86Subtarget *Subtarget) {
25967 SDValue Op0 = N->getOperand(0);
25968 SDValue Op1 = N->getOperand(1);
25970 // X86 can't encode an immediate LHS of a sub. See if we can push the
25971 // negation into a preceding instruction.
25972 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op0)) {
25973 // If the RHS of the sub is a XOR with one use and a constant, invert the
25974 // immediate. Then add one to the LHS of the sub so we can turn
25975 // X-Y -> X+~Y+1, saving one register.
25976 if (Op1->hasOneUse() && Op1.getOpcode() == ISD::XOR &&
25977 isa<ConstantSDNode>(Op1.getOperand(1))) {
25978 APInt XorC = cast<ConstantSDNode>(Op1.getOperand(1))->getAPIntValue();
25979 EVT VT = Op0.getValueType();
25980 SDValue NewXor = DAG.getNode(ISD::XOR, SDLoc(Op1), VT,
25982 DAG.getConstant(~XorC, SDLoc(Op1), VT));
25983 return DAG.getNode(ISD::ADD, SDLoc(N), VT, NewXor,
25984 DAG.getConstant(C->getAPIntValue() + 1, SDLoc(N), VT));
25988 // Try to synthesize horizontal adds from adds of shuffles.
25989 EVT VT = N->getValueType(0);
25990 if (((Subtarget->hasSSSE3() && (VT == MVT::v8i16 || VT == MVT::v4i32)) ||
25991 (Subtarget->hasInt256() && (VT == MVT::v16i16 || VT == MVT::v8i32))) &&
25992 isHorizontalBinOp(Op0, Op1, true))
25993 return DAG.getNode(X86ISD::HSUB, SDLoc(N), VT, Op0, Op1);
25995 return OptimizeConditionalInDecrement(N, DAG);
25998 /// performVZEXTCombine - Performs build vector combines
25999 static SDValue performVZEXTCombine(SDNode *N, SelectionDAG &DAG,
26000 TargetLowering::DAGCombinerInfo &DCI,
26001 const X86Subtarget *Subtarget) {
26003 MVT VT = N->getSimpleValueType(0);
26004 SDValue Op = N->getOperand(0);
26005 MVT OpVT = Op.getSimpleValueType();
26006 MVT OpEltVT = OpVT.getVectorElementType();
26007 unsigned InputBits = OpEltVT.getSizeInBits() * VT.getVectorNumElements();
26009 // (vzext (bitcast (vzext (x)) -> (vzext x)
26011 while (V.getOpcode() == ISD::BITCAST)
26012 V = V.getOperand(0);
26014 if (V != Op && V.getOpcode() == X86ISD::VZEXT) {
26015 MVT InnerVT = V.getSimpleValueType();
26016 MVT InnerEltVT = InnerVT.getVectorElementType();
26018 // If the element sizes match exactly, we can just do one larger vzext. This
26019 // is always an exact type match as vzext operates on integer types.
26020 if (OpEltVT == InnerEltVT) {
26021 assert(OpVT == InnerVT && "Types must match for vzext!");
26022 return DAG.getNode(X86ISD::VZEXT, DL, VT, V.getOperand(0));
26025 // The only other way we can combine them is if only a single element of the
26026 // inner vzext is used in the input to the outer vzext.
26027 if (InnerEltVT.getSizeInBits() < InputBits)
26030 // In this case, the inner vzext is completely dead because we're going to
26031 // only look at bits inside of the low element. Just do the outer vzext on
26032 // a bitcast of the input to the inner.
26033 return DAG.getNode(X86ISD::VZEXT, DL, VT, DAG.getBitcast(OpVT, V));
26036 // Check if we can bypass extracting and re-inserting an element of an input
26037 // vector. Essentially:
26038 // (bitcast (sclr2vec (ext_vec_elt x))) -> (bitcast x)
26039 if (V.getOpcode() == ISD::SCALAR_TO_VECTOR &&
26040 V.getOperand(0).getOpcode() == ISD::EXTRACT_VECTOR_ELT &&
26041 V.getOperand(0).getSimpleValueType().getSizeInBits() == InputBits) {
26042 SDValue ExtractedV = V.getOperand(0);
26043 SDValue OrigV = ExtractedV.getOperand(0);
26044 if (auto *ExtractIdx = dyn_cast<ConstantSDNode>(ExtractedV.getOperand(1)))
26045 if (ExtractIdx->getZExtValue() == 0) {
26046 MVT OrigVT = OrigV.getSimpleValueType();
26047 // Extract a subvector if necessary...
26048 if (OrigVT.getSizeInBits() > OpVT.getSizeInBits()) {
26049 int Ratio = OrigVT.getSizeInBits() / OpVT.getSizeInBits();
26050 OrigVT = MVT::getVectorVT(OrigVT.getVectorElementType(),
26051 OrigVT.getVectorNumElements() / Ratio);
26052 OrigV = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, OrigVT, OrigV,
26053 DAG.getIntPtrConstant(0, DL));
26055 Op = DAG.getBitcast(OpVT, OrigV);
26056 return DAG.getNode(X86ISD::VZEXT, DL, VT, Op);
26063 SDValue X86TargetLowering::PerformDAGCombine(SDNode *N,
26064 DAGCombinerInfo &DCI) const {
26065 SelectionDAG &DAG = DCI.DAG;
26066 switch (N->getOpcode()) {
26068 case ISD::EXTRACT_VECTOR_ELT:
26069 return PerformEXTRACT_VECTOR_ELTCombine(N, DAG, DCI);
26072 case X86ISD::SHRUNKBLEND:
26073 return PerformSELECTCombine(N, DAG, DCI, Subtarget);
26074 case ISD::BITCAST: return PerformBITCASTCombine(N, DAG);
26075 case X86ISD::CMOV: return PerformCMOVCombine(N, DAG, DCI, Subtarget);
26076 case ISD::ADD: return PerformAddCombine(N, DAG, Subtarget);
26077 case ISD::SUB: return PerformSubCombine(N, DAG, Subtarget);
26078 case X86ISD::ADC: return PerformADCCombine(N, DAG, DCI);
26079 case ISD::MUL: return PerformMulCombine(N, DAG, DCI);
26082 case ISD::SRL: return PerformShiftCombine(N, DAG, DCI, Subtarget);
26083 case ISD::AND: return PerformAndCombine(N, DAG, DCI, Subtarget);
26084 case ISD::OR: return PerformOrCombine(N, DAG, DCI, Subtarget);
26085 case ISD::XOR: return PerformXorCombine(N, DAG, DCI, Subtarget);
26086 case ISD::LOAD: return PerformLOADCombine(N, DAG, DCI, Subtarget);
26087 case ISD::MLOAD: return PerformMLOADCombine(N, DAG, DCI, Subtarget);
26088 case ISD::STORE: return PerformSTORECombine(N, DAG, Subtarget);
26089 case ISD::MSTORE: return PerformMSTORECombine(N, DAG, Subtarget);
26090 case ISD::SINT_TO_FP: return PerformSINT_TO_FPCombine(N, DAG, Subtarget);
26091 case ISD::UINT_TO_FP: return PerformUINT_TO_FPCombine(N, DAG, Subtarget);
26092 case ISD::FADD: return PerformFADDCombine(N, DAG, Subtarget);
26093 case ISD::FSUB: return PerformFSUBCombine(N, DAG, Subtarget);
26095 case X86ISD::FOR: return PerformFORCombine(N, DAG, Subtarget);
26097 case X86ISD::FMAX: return PerformFMinFMaxCombine(N, DAG);
26098 case X86ISD::FAND: return PerformFANDCombine(N, DAG);
26099 case X86ISD::FANDN: return PerformFANDNCombine(N, DAG);
26100 case X86ISD::BT: return PerformBTCombine(N, DAG, DCI);
26101 case X86ISD::VZEXT_MOVL: return PerformVZEXT_MOVLCombine(N, DAG);
26102 case ISD::ANY_EXTEND:
26103 case ISD::ZERO_EXTEND: return PerformZExtCombine(N, DAG, DCI, Subtarget);
26104 case ISD::SIGN_EXTEND: return PerformSExtCombine(N, DAG, DCI, Subtarget);
26105 case ISD::SIGN_EXTEND_INREG:
26106 return PerformSIGN_EXTEND_INREGCombine(N, DAG, Subtarget);
26107 case ISD::SETCC: return PerformISDSETCCCombine(N, DAG, Subtarget);
26108 case X86ISD::SETCC: return PerformSETCCCombine(N, DAG, DCI, Subtarget);
26109 case X86ISD::BRCOND: return PerformBrCondCombine(N, DAG, DCI, Subtarget);
26110 case X86ISD::VZEXT: return performVZEXTCombine(N, DAG, DCI, Subtarget);
26111 case X86ISD::SHUFP: // Handle all target specific shuffles
26112 case X86ISD::PALIGNR:
26113 case X86ISD::UNPCKH:
26114 case X86ISD::UNPCKL:
26115 case X86ISD::MOVHLPS:
26116 case X86ISD::MOVLHPS:
26117 case X86ISD::PSHUFB:
26118 case X86ISD::PSHUFD:
26119 case X86ISD::PSHUFHW:
26120 case X86ISD::PSHUFLW:
26121 case X86ISD::MOVSS:
26122 case X86ISD::MOVSD:
26123 case X86ISD::VPERMILPI:
26124 case X86ISD::VPERM2X128:
26125 case ISD::VECTOR_SHUFFLE: return PerformShuffleCombine(N, DAG, DCI,Subtarget);
26126 case ISD::FMA: return PerformFMACombine(N, DAG, Subtarget);
26127 case X86ISD::INSERTPS: {
26128 if (getTargetMachine().getOptLevel() > CodeGenOpt::None)
26129 return PerformINSERTPSCombine(N, DAG, Subtarget);
26132 case X86ISD::BLENDI: return PerformBLENDICombine(N, DAG);
26138 /// isTypeDesirableForOp - Return true if the target has native support for
26139 /// the specified value type and it is 'desirable' to use the type for the
26140 /// given node type. e.g. On x86 i16 is legal, but undesirable since i16
26141 /// instruction encodings are longer and some i16 instructions are slow.
26142 bool X86TargetLowering::isTypeDesirableForOp(unsigned Opc, EVT VT) const {
26143 if (!isTypeLegal(VT))
26145 if (VT != MVT::i16)
26152 case ISD::SIGN_EXTEND:
26153 case ISD::ZERO_EXTEND:
26154 case ISD::ANY_EXTEND:
26167 /// IsDesirableToPromoteOp - This method query the target whether it is
26168 /// beneficial for dag combiner to promote the specified node. If true, it
26169 /// should return the desired promotion type by reference.
26170 bool X86TargetLowering::IsDesirableToPromoteOp(SDValue Op, EVT &PVT) const {
26171 EVT VT = Op.getValueType();
26172 if (VT != MVT::i16)
26175 bool Promote = false;
26176 bool Commute = false;
26177 switch (Op.getOpcode()) {
26180 LoadSDNode *LD = cast<LoadSDNode>(Op);
26181 // If the non-extending load has a single use and it's not live out, then it
26182 // might be folded.
26183 if (LD->getExtensionType() == ISD::NON_EXTLOAD /*&&
26184 Op.hasOneUse()*/) {
26185 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
26186 UE = Op.getNode()->use_end(); UI != UE; ++UI) {
26187 // The only case where we'd want to promote LOAD (rather then it being
26188 // promoted as an operand is when it's only use is liveout.
26189 if (UI->getOpcode() != ISD::CopyToReg)
26196 case ISD::SIGN_EXTEND:
26197 case ISD::ZERO_EXTEND:
26198 case ISD::ANY_EXTEND:
26203 SDValue N0 = Op.getOperand(0);
26204 // Look out for (store (shl (load), x)).
26205 if (MayFoldLoad(N0) && MayFoldIntoStore(Op))
26218 SDValue N0 = Op.getOperand(0);
26219 SDValue N1 = Op.getOperand(1);
26220 if (!Commute && MayFoldLoad(N1))
26222 // Avoid disabling potential load folding opportunities.
26223 if (MayFoldLoad(N0) && (!isa<ConstantSDNode>(N1) || MayFoldIntoStore(Op)))
26225 if (MayFoldLoad(N1) && (!isa<ConstantSDNode>(N0) || MayFoldIntoStore(Op)))
26235 //===----------------------------------------------------------------------===//
26236 // X86 Inline Assembly Support
26237 //===----------------------------------------------------------------------===//
26239 // Helper to match a string separated by whitespace.
26240 static bool matchAsm(StringRef S, ArrayRef<const char *> Pieces) {
26241 S = S.substr(S.find_first_not_of(" \t")); // Skip leading whitespace.
26243 for (StringRef Piece : Pieces) {
26244 if (!S.startswith(Piece)) // Check if the piece matches.
26247 S = S.substr(Piece.size());
26248 StringRef::size_type Pos = S.find_first_not_of(" \t");
26249 if (Pos == 0) // We matched a prefix.
26258 static bool clobbersFlagRegisters(const SmallVector<StringRef, 4> &AsmPieces) {
26260 if (AsmPieces.size() == 3 || AsmPieces.size() == 4) {
26261 if (std::count(AsmPieces.begin(), AsmPieces.end(), "~{cc}") &&
26262 std::count(AsmPieces.begin(), AsmPieces.end(), "~{flags}") &&
26263 std::count(AsmPieces.begin(), AsmPieces.end(), "~{fpsr}")) {
26265 if (AsmPieces.size() == 3)
26267 else if (std::count(AsmPieces.begin(), AsmPieces.end(), "~{dirflag}"))
26274 bool X86TargetLowering::ExpandInlineAsm(CallInst *CI) const {
26275 InlineAsm *IA = cast<InlineAsm>(CI->getCalledValue());
26277 std::string AsmStr = IA->getAsmString();
26279 IntegerType *Ty = dyn_cast<IntegerType>(CI->getType());
26280 if (!Ty || Ty->getBitWidth() % 16 != 0)
26283 // TODO: should remove alternatives from the asmstring: "foo {a|b}" -> "foo a"
26284 SmallVector<StringRef, 4> AsmPieces;
26285 SplitString(AsmStr, AsmPieces, ";\n");
26287 switch (AsmPieces.size()) {
26288 default: return false;
26290 // FIXME: this should verify that we are targeting a 486 or better. If not,
26291 // we will turn this bswap into something that will be lowered to logical
26292 // ops instead of emitting the bswap asm. For now, we don't support 486 or
26293 // lower so don't worry about this.
26295 if (matchAsm(AsmPieces[0], {"bswap", "$0"}) ||
26296 matchAsm(AsmPieces[0], {"bswapl", "$0"}) ||
26297 matchAsm(AsmPieces[0], {"bswapq", "$0"}) ||
26298 matchAsm(AsmPieces[0], {"bswap", "${0:q}"}) ||
26299 matchAsm(AsmPieces[0], {"bswapl", "${0:q}"}) ||
26300 matchAsm(AsmPieces[0], {"bswapq", "${0:q}"})) {
26301 // No need to check constraints, nothing other than the equivalent of
26302 // "=r,0" would be valid here.
26303 return IntrinsicLowering::LowerToByteSwap(CI);
26306 // rorw $$8, ${0:w} --> llvm.bswap.i16
26307 if (CI->getType()->isIntegerTy(16) &&
26308 IA->getConstraintString().compare(0, 5, "=r,0,") == 0 &&
26309 (matchAsm(AsmPieces[0], {"rorw", "$$8,", "${0:w}"}) ||
26310 matchAsm(AsmPieces[0], {"rolw", "$$8,", "${0:w}"}))) {
26312 StringRef ConstraintsStr = IA->getConstraintString();
26313 SplitString(StringRef(ConstraintsStr).substr(5), AsmPieces, ",");
26314 array_pod_sort(AsmPieces.begin(), AsmPieces.end());
26315 if (clobbersFlagRegisters(AsmPieces))
26316 return IntrinsicLowering::LowerToByteSwap(CI);
26320 if (CI->getType()->isIntegerTy(32) &&
26321 IA->getConstraintString().compare(0, 5, "=r,0,") == 0 &&
26322 matchAsm(AsmPieces[0], {"rorw", "$$8,", "${0:w}"}) &&
26323 matchAsm(AsmPieces[1], {"rorl", "$$16,", "$0"}) &&
26324 matchAsm(AsmPieces[2], {"rorw", "$$8,", "${0:w}"})) {
26326 StringRef ConstraintsStr = IA->getConstraintString();
26327 SplitString(StringRef(ConstraintsStr).substr(5), AsmPieces, ",");
26328 array_pod_sort(AsmPieces.begin(), AsmPieces.end());
26329 if (clobbersFlagRegisters(AsmPieces))
26330 return IntrinsicLowering::LowerToByteSwap(CI);
26333 if (CI->getType()->isIntegerTy(64)) {
26334 InlineAsm::ConstraintInfoVector Constraints = IA->ParseConstraints();
26335 if (Constraints.size() >= 2 &&
26336 Constraints[0].Codes.size() == 1 && Constraints[0].Codes[0] == "A" &&
26337 Constraints[1].Codes.size() == 1 && Constraints[1].Codes[0] == "0") {
26338 // bswap %eax / bswap %edx / xchgl %eax, %edx -> llvm.bswap.i64
26339 if (matchAsm(AsmPieces[0], {"bswap", "%eax"}) &&
26340 matchAsm(AsmPieces[1], {"bswap", "%edx"}) &&
26341 matchAsm(AsmPieces[2], {"xchgl", "%eax,", "%edx"}))
26342 return IntrinsicLowering::LowerToByteSwap(CI);
26350 /// getConstraintType - Given a constraint letter, return the type of
26351 /// constraint it is for this target.
26352 X86TargetLowering::ConstraintType
26353 X86TargetLowering::getConstraintType(StringRef Constraint) const {
26354 if (Constraint.size() == 1) {
26355 switch (Constraint[0]) {
26366 return C_RegisterClass;
26390 return TargetLowering::getConstraintType(Constraint);
26393 /// Examine constraint type and operand type and determine a weight value.
26394 /// This object must already have been set up with the operand type
26395 /// and the current alternative constraint selected.
26396 TargetLowering::ConstraintWeight
26397 X86TargetLowering::getSingleConstraintMatchWeight(
26398 AsmOperandInfo &info, const char *constraint) const {
26399 ConstraintWeight weight = CW_Invalid;
26400 Value *CallOperandVal = info.CallOperandVal;
26401 // If we don't have a value, we can't do a match,
26402 // but allow it at the lowest weight.
26403 if (!CallOperandVal)
26405 Type *type = CallOperandVal->getType();
26406 // Look at the constraint type.
26407 switch (*constraint) {
26409 weight = TargetLowering::getSingleConstraintMatchWeight(info, constraint);
26420 if (CallOperandVal->getType()->isIntegerTy())
26421 weight = CW_SpecificReg;
26426 if (type->isFloatingPointTy())
26427 weight = CW_SpecificReg;
26430 if (type->isX86_MMXTy() && Subtarget->hasMMX())
26431 weight = CW_SpecificReg;
26435 if (((type->getPrimitiveSizeInBits() == 128) && Subtarget->hasSSE1()) ||
26436 ((type->getPrimitiveSizeInBits() == 256) && Subtarget->hasFp256()))
26437 weight = CW_Register;
26440 if (ConstantInt *C = dyn_cast<ConstantInt>(info.CallOperandVal)) {
26441 if (C->getZExtValue() <= 31)
26442 weight = CW_Constant;
26446 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
26447 if (C->getZExtValue() <= 63)
26448 weight = CW_Constant;
26452 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
26453 if ((C->getSExtValue() >= -0x80) && (C->getSExtValue() <= 0x7f))
26454 weight = CW_Constant;
26458 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
26459 if ((C->getZExtValue() == 0xff) || (C->getZExtValue() == 0xffff))
26460 weight = CW_Constant;
26464 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
26465 if (C->getZExtValue() <= 3)
26466 weight = CW_Constant;
26470 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
26471 if (C->getZExtValue() <= 0xff)
26472 weight = CW_Constant;
26477 if (isa<ConstantFP>(CallOperandVal)) {
26478 weight = CW_Constant;
26482 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
26483 if ((C->getSExtValue() >= -0x80000000LL) &&
26484 (C->getSExtValue() <= 0x7fffffffLL))
26485 weight = CW_Constant;
26489 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
26490 if (C->getZExtValue() <= 0xffffffff)
26491 weight = CW_Constant;
26498 /// LowerXConstraint - try to replace an X constraint, which matches anything,
26499 /// with another that has more specific requirements based on the type of the
26500 /// corresponding operand.
26501 const char *X86TargetLowering::
26502 LowerXConstraint(EVT ConstraintVT) const {
26503 // FP X constraints get lowered to SSE1/2 registers if available, otherwise
26504 // 'f' like normal targets.
26505 if (ConstraintVT.isFloatingPoint()) {
26506 if (Subtarget->hasSSE2())
26508 if (Subtarget->hasSSE1())
26512 return TargetLowering::LowerXConstraint(ConstraintVT);
26515 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
26516 /// vector. If it is invalid, don't add anything to Ops.
26517 void X86TargetLowering::LowerAsmOperandForConstraint(SDValue Op,
26518 std::string &Constraint,
26519 std::vector<SDValue>&Ops,
26520 SelectionDAG &DAG) const {
26523 // Only support length 1 constraints for now.
26524 if (Constraint.length() > 1) return;
26526 char ConstraintLetter = Constraint[0];
26527 switch (ConstraintLetter) {
26530 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
26531 if (C->getZExtValue() <= 31) {
26532 Result = DAG.getTargetConstant(C->getZExtValue(), SDLoc(Op),
26533 Op.getValueType());
26539 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
26540 if (C->getZExtValue() <= 63) {
26541 Result = DAG.getTargetConstant(C->getZExtValue(), SDLoc(Op),
26542 Op.getValueType());
26548 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
26549 if (isInt<8>(C->getSExtValue())) {
26550 Result = DAG.getTargetConstant(C->getZExtValue(), SDLoc(Op),
26551 Op.getValueType());
26557 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
26558 if (C->getZExtValue() == 0xff || C->getZExtValue() == 0xffff ||
26559 (Subtarget->is64Bit() && C->getZExtValue() == 0xffffffff)) {
26560 Result = DAG.getTargetConstant(C->getSExtValue(), SDLoc(Op),
26561 Op.getValueType());
26567 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
26568 if (C->getZExtValue() <= 3) {
26569 Result = DAG.getTargetConstant(C->getZExtValue(), SDLoc(Op),
26570 Op.getValueType());
26576 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
26577 if (C->getZExtValue() <= 255) {
26578 Result = DAG.getTargetConstant(C->getZExtValue(), SDLoc(Op),
26579 Op.getValueType());
26585 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
26586 if (C->getZExtValue() <= 127) {
26587 Result = DAG.getTargetConstant(C->getZExtValue(), SDLoc(Op),
26588 Op.getValueType());
26594 // 32-bit signed value
26595 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
26596 if (ConstantInt::isValueValidForType(Type::getInt32Ty(*DAG.getContext()),
26597 C->getSExtValue())) {
26598 // Widen to 64 bits here to get it sign extended.
26599 Result = DAG.getTargetConstant(C->getSExtValue(), SDLoc(Op), MVT::i64);
26602 // FIXME gcc accepts some relocatable values here too, but only in certain
26603 // memory models; it's complicated.
26608 // 32-bit unsigned value
26609 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
26610 if (ConstantInt::isValueValidForType(Type::getInt32Ty(*DAG.getContext()),
26611 C->getZExtValue())) {
26612 Result = DAG.getTargetConstant(C->getZExtValue(), SDLoc(Op),
26613 Op.getValueType());
26617 // FIXME gcc accepts some relocatable values here too, but only in certain
26618 // memory models; it's complicated.
26622 // Literal immediates are always ok.
26623 if (ConstantSDNode *CST = dyn_cast<ConstantSDNode>(Op)) {
26624 // Widen to 64 bits here to get it sign extended.
26625 Result = DAG.getTargetConstant(CST->getSExtValue(), SDLoc(Op), MVT::i64);
26629 // In any sort of PIC mode addresses need to be computed at runtime by
26630 // adding in a register or some sort of table lookup. These can't
26631 // be used as immediates.
26632 if (Subtarget->isPICStyleGOT() || Subtarget->isPICStyleStubPIC())
26635 // If we are in non-pic codegen mode, we allow the address of a global (with
26636 // an optional displacement) to be used with 'i'.
26637 GlobalAddressSDNode *GA = nullptr;
26638 int64_t Offset = 0;
26640 // Match either (GA), (GA+C), (GA+C1+C2), etc.
26642 if ((GA = dyn_cast<GlobalAddressSDNode>(Op))) {
26643 Offset += GA->getOffset();
26645 } else if (Op.getOpcode() == ISD::ADD) {
26646 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
26647 Offset += C->getZExtValue();
26648 Op = Op.getOperand(0);
26651 } else if (Op.getOpcode() == ISD::SUB) {
26652 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
26653 Offset += -C->getZExtValue();
26654 Op = Op.getOperand(0);
26659 // Otherwise, this isn't something we can handle, reject it.
26663 const GlobalValue *GV = GA->getGlobal();
26664 // If we require an extra load to get this address, as in PIC mode, we
26665 // can't accept it.
26666 if (isGlobalStubReference(
26667 Subtarget->ClassifyGlobalReference(GV, DAG.getTarget())))
26670 Result = DAG.getTargetGlobalAddress(GV, SDLoc(Op),
26671 GA->getValueType(0), Offset);
26676 if (Result.getNode()) {
26677 Ops.push_back(Result);
26680 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG);
26683 std::pair<unsigned, const TargetRegisterClass *>
26684 X86TargetLowering::getRegForInlineAsmConstraint(const TargetRegisterInfo *TRI,
26685 StringRef Constraint,
26687 // First, see if this is a constraint that directly corresponds to an LLVM
26689 if (Constraint.size() == 1) {
26690 // GCC Constraint Letters
26691 switch (Constraint[0]) {
26693 // TODO: Slight differences here in allocation order and leaving
26694 // RIP in the class. Do they matter any more here than they do
26695 // in the normal allocation?
26696 case 'q': // GENERAL_REGS in 64-bit mode, Q_REGS in 32-bit mode.
26697 if (Subtarget->is64Bit()) {
26698 if (VT == MVT::i32 || VT == MVT::f32)
26699 return std::make_pair(0U, &X86::GR32RegClass);
26700 if (VT == MVT::i16)
26701 return std::make_pair(0U, &X86::GR16RegClass);
26702 if (VT == MVT::i8 || VT == MVT::i1)
26703 return std::make_pair(0U, &X86::GR8RegClass);
26704 if (VT == MVT::i64 || VT == MVT::f64)
26705 return std::make_pair(0U, &X86::GR64RegClass);
26708 // 32-bit fallthrough
26709 case 'Q': // Q_REGS
26710 if (VT == MVT::i32 || VT == MVT::f32)
26711 return std::make_pair(0U, &X86::GR32_ABCDRegClass);
26712 if (VT == MVT::i16)
26713 return std::make_pair(0U, &X86::GR16_ABCDRegClass);
26714 if (VT == MVT::i8 || VT == MVT::i1)
26715 return std::make_pair(0U, &X86::GR8_ABCD_LRegClass);
26716 if (VT == MVT::i64)
26717 return std::make_pair(0U, &X86::GR64_ABCDRegClass);
26719 case 'r': // GENERAL_REGS
26720 case 'l': // INDEX_REGS
26721 if (VT == MVT::i8 || VT == MVT::i1)
26722 return std::make_pair(0U, &X86::GR8RegClass);
26723 if (VT == MVT::i16)
26724 return std::make_pair(0U, &X86::GR16RegClass);
26725 if (VT == MVT::i32 || VT == MVT::f32 || !Subtarget->is64Bit())
26726 return std::make_pair(0U, &X86::GR32RegClass);
26727 return std::make_pair(0U, &X86::GR64RegClass);
26728 case 'R': // LEGACY_REGS
26729 if (VT == MVT::i8 || VT == MVT::i1)
26730 return std::make_pair(0U, &X86::GR8_NOREXRegClass);
26731 if (VT == MVT::i16)
26732 return std::make_pair(0U, &X86::GR16_NOREXRegClass);
26733 if (VT == MVT::i32 || !Subtarget->is64Bit())
26734 return std::make_pair(0U, &X86::GR32_NOREXRegClass);
26735 return std::make_pair(0U, &X86::GR64_NOREXRegClass);
26736 case 'f': // FP Stack registers.
26737 // If SSE is enabled for this VT, use f80 to ensure the isel moves the
26738 // value to the correct fpstack register class.
26739 if (VT == MVT::f32 && !isScalarFPTypeInSSEReg(VT))
26740 return std::make_pair(0U, &X86::RFP32RegClass);
26741 if (VT == MVT::f64 && !isScalarFPTypeInSSEReg(VT))
26742 return std::make_pair(0U, &X86::RFP64RegClass);
26743 return std::make_pair(0U, &X86::RFP80RegClass);
26744 case 'y': // MMX_REGS if MMX allowed.
26745 if (!Subtarget->hasMMX()) break;
26746 return std::make_pair(0U, &X86::VR64RegClass);
26747 case 'Y': // SSE_REGS if SSE2 allowed
26748 if (!Subtarget->hasSSE2()) break;
26750 case 'x': // SSE_REGS if SSE1 allowed or AVX_REGS if AVX allowed
26751 if (!Subtarget->hasSSE1()) break;
26753 switch (VT.SimpleTy) {
26755 // Scalar SSE types.
26758 return std::make_pair(0U, &X86::FR32RegClass);
26761 return std::make_pair(0U, &X86::FR64RegClass);
26769 return std::make_pair(0U, &X86::VR128RegClass);
26777 return std::make_pair(0U, &X86::VR256RegClass);
26782 return std::make_pair(0U, &X86::VR512RegClass);
26788 // Use the default implementation in TargetLowering to convert the register
26789 // constraint into a member of a register class.
26790 std::pair<unsigned, const TargetRegisterClass*> Res;
26791 Res = TargetLowering::getRegForInlineAsmConstraint(TRI, Constraint, VT);
26793 // Not found as a standard register?
26795 // Map st(0) -> st(7) -> ST0
26796 if (Constraint.size() == 7 && Constraint[0] == '{' &&
26797 tolower(Constraint[1]) == 's' &&
26798 tolower(Constraint[2]) == 't' &&
26799 Constraint[3] == '(' &&
26800 (Constraint[4] >= '0' && Constraint[4] <= '7') &&
26801 Constraint[5] == ')' &&
26802 Constraint[6] == '}') {
26804 Res.first = X86::FP0+Constraint[4]-'0';
26805 Res.second = &X86::RFP80RegClass;
26809 // GCC allows "st(0)" to be called just plain "st".
26810 if (StringRef("{st}").equals_lower(Constraint)) {
26811 Res.first = X86::FP0;
26812 Res.second = &X86::RFP80RegClass;
26817 if (StringRef("{flags}").equals_lower(Constraint)) {
26818 Res.first = X86::EFLAGS;
26819 Res.second = &X86::CCRRegClass;
26823 // 'A' means EAX + EDX.
26824 if (Constraint == "A") {
26825 Res.first = X86::EAX;
26826 Res.second = &X86::GR32_ADRegClass;
26832 // Otherwise, check to see if this is a register class of the wrong value
26833 // type. For example, we want to map "{ax},i32" -> {eax}, we don't want it to
26834 // turn into {ax},{dx}.
26835 // MVT::Other is used to specify clobber names.
26836 if (Res.second->hasType(VT) || VT == MVT::Other)
26837 return Res; // Correct type already, nothing to do.
26839 // Get a matching integer of the correct size. i.e. "ax" with MVT::32 should
26840 // return "eax". This should even work for things like getting 64bit integer
26841 // registers when given an f64 type.
26842 const TargetRegisterClass *Class = Res.second;
26843 if (Class == &X86::GR8RegClass || Class == &X86::GR16RegClass ||
26844 Class == &X86::GR32RegClass || Class == &X86::GR64RegClass) {
26845 unsigned Size = VT.getSizeInBits();
26846 MVT::SimpleValueType SimpleTy = Size == 1 || Size == 8 ? MVT::i8
26847 : Size == 16 ? MVT::i16
26848 : Size == 32 ? MVT::i32
26849 : Size == 64 ? MVT::i64
26851 unsigned DestReg = getX86SubSuperRegisterOrZero(Res.first, SimpleTy);
26853 Res.first = DestReg;
26854 Res.second = SimpleTy == MVT::i8 ? &X86::GR8RegClass
26855 : SimpleTy == MVT::i16 ? &X86::GR16RegClass
26856 : SimpleTy == MVT::i32 ? &X86::GR32RegClass
26857 : &X86::GR64RegClass;
26858 assert(Res.second->contains(Res.first) && "Register in register class");
26860 // No register found/type mismatch.
26862 Res.second = nullptr;
26864 } else if (Class == &X86::FR32RegClass || Class == &X86::FR64RegClass ||
26865 Class == &X86::VR128RegClass || Class == &X86::VR256RegClass ||
26866 Class == &X86::FR32XRegClass || Class == &X86::FR64XRegClass ||
26867 Class == &X86::VR128XRegClass || Class == &X86::VR256XRegClass ||
26868 Class == &X86::VR512RegClass) {
26869 // Handle references to XMM physical registers that got mapped into the
26870 // wrong class. This can happen with constraints like {xmm0} where the
26871 // target independent register mapper will just pick the first match it can
26872 // find, ignoring the required type.
26874 if (VT == MVT::f32 || VT == MVT::i32)
26875 Res.second = &X86::FR32RegClass;
26876 else if (VT == MVT::f64 || VT == MVT::i64)
26877 Res.second = &X86::FR64RegClass;
26878 else if (X86::VR128RegClass.hasType(VT))
26879 Res.second = &X86::VR128RegClass;
26880 else if (X86::VR256RegClass.hasType(VT))
26881 Res.second = &X86::VR256RegClass;
26882 else if (X86::VR512RegClass.hasType(VT))
26883 Res.second = &X86::VR512RegClass;
26885 // Type mismatch and not a clobber: Return an error;
26887 Res.second = nullptr;
26894 int X86TargetLowering::getScalingFactorCost(const DataLayout &DL,
26895 const AddrMode &AM, Type *Ty,
26896 unsigned AS) const {
26897 // Scaling factors are not free at all.
26898 // An indexed folded instruction, i.e., inst (reg1, reg2, scale),
26899 // will take 2 allocations in the out of order engine instead of 1
26900 // for plain addressing mode, i.e. inst (reg1).
26902 // vaddps (%rsi,%drx), %ymm0, %ymm1
26903 // Requires two allocations (one for the load, one for the computation)
26905 // vaddps (%rsi), %ymm0, %ymm1
26906 // Requires just 1 allocation, i.e., freeing allocations for other operations
26907 // and having less micro operations to execute.
26909 // For some X86 architectures, this is even worse because for instance for
26910 // stores, the complex addressing mode forces the instruction to use the
26911 // "load" ports instead of the dedicated "store" port.
26912 // E.g., on Haswell:
26913 // vmovaps %ymm1, (%r8, %rdi) can use port 2 or 3.
26914 // vmovaps %ymm1, (%r8) can use port 2, 3, or 7.
26915 if (isLegalAddressingMode(DL, AM, Ty, AS))
26916 // Scale represents reg2 * scale, thus account for 1
26917 // as soon as we use a second register.
26918 return AM.Scale != 0;
26922 bool X86TargetLowering::isIntDivCheap(EVT VT, AttributeSet Attr) const {
26923 // Integer division on x86 is expensive. However, when aggressively optimizing
26924 // for code size, we prefer to use a div instruction, as it is usually smaller
26925 // than the alternative sequence.
26926 // The exception to this is vector division. Since x86 doesn't have vector
26927 // integer division, leaving the division as-is is a loss even in terms of
26928 // size, because it will have to be scalarized, while the alternative code
26929 // sequence can be performed in vector form.
26930 bool OptSize = Attr.hasAttribute(AttributeSet::FunctionIndex,
26931 Attribute::MinSize);
26932 return OptSize && !VT.isVector();