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);
118 // The _ftol2 runtime function has an unusual calling conv, which
119 // is modeled by a special pseudo-instruction.
120 setLibcallName(RTLIB::FPTOUINT_F64_I64, nullptr);
121 setLibcallName(RTLIB::FPTOUINT_F32_I64, nullptr);
122 setLibcallName(RTLIB::FPTOUINT_F64_I32, nullptr);
123 setLibcallName(RTLIB::FPTOUINT_F32_I32, nullptr);
126 if (Subtarget->isTargetDarwin()) {
127 // Darwin should use _setjmp/_longjmp instead of setjmp/longjmp.
128 setUseUnderscoreSetJmp(false);
129 setUseUnderscoreLongJmp(false);
130 } else if (Subtarget->isTargetWindowsGNU()) {
131 // MS runtime is weird: it exports _setjmp, but longjmp!
132 setUseUnderscoreSetJmp(true);
133 setUseUnderscoreLongJmp(false);
135 setUseUnderscoreSetJmp(true);
136 setUseUnderscoreLongJmp(true);
139 // Set up the register classes.
140 addRegisterClass(MVT::i8, &X86::GR8RegClass);
141 addRegisterClass(MVT::i16, &X86::GR16RegClass);
142 addRegisterClass(MVT::i32, &X86::GR32RegClass);
143 if (Subtarget->is64Bit())
144 addRegisterClass(MVT::i64, &X86::GR64RegClass);
146 for (MVT VT : MVT::integer_valuetypes())
147 setLoadExtAction(ISD::SEXTLOAD, VT, MVT::i1, Promote);
149 // We don't accept any truncstore of integer registers.
150 setTruncStoreAction(MVT::i64, MVT::i32, Expand);
151 setTruncStoreAction(MVT::i64, MVT::i16, Expand);
152 setTruncStoreAction(MVT::i64, MVT::i8 , Expand);
153 setTruncStoreAction(MVT::i32, MVT::i16, Expand);
154 setTruncStoreAction(MVT::i32, MVT::i8 , Expand);
155 setTruncStoreAction(MVT::i16, MVT::i8, Expand);
157 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
159 // SETOEQ and SETUNE require checking two conditions.
160 setCondCodeAction(ISD::SETOEQ, MVT::f32, Expand);
161 setCondCodeAction(ISD::SETOEQ, MVT::f64, Expand);
162 setCondCodeAction(ISD::SETOEQ, MVT::f80, Expand);
163 setCondCodeAction(ISD::SETUNE, MVT::f32, Expand);
164 setCondCodeAction(ISD::SETUNE, MVT::f64, Expand);
165 setCondCodeAction(ISD::SETUNE, MVT::f80, Expand);
167 // Promote all UINT_TO_FP to larger SINT_TO_FP's, as X86 doesn't have this
169 setOperationAction(ISD::UINT_TO_FP , MVT::i1 , Promote);
170 setOperationAction(ISD::UINT_TO_FP , MVT::i8 , Promote);
171 setOperationAction(ISD::UINT_TO_FP , MVT::i16 , Promote);
173 if (Subtarget->is64Bit()) {
174 setOperationAction(ISD::UINT_TO_FP , MVT::i32 , Promote);
175 setOperationAction(ISD::UINT_TO_FP , MVT::i64 , Custom);
176 } else if (!Subtarget->useSoftFloat()) {
177 // We have an algorithm for SSE2->double, and we turn this into a
178 // 64-bit FILD followed by conditional FADD for other targets.
179 setOperationAction(ISD::UINT_TO_FP , MVT::i64 , Custom);
180 // We have an algorithm for SSE2, and we turn this into a 64-bit
181 // FILD for other targets.
182 setOperationAction(ISD::UINT_TO_FP , MVT::i32 , Custom);
185 // Promote i1/i8 SINT_TO_FP to larger SINT_TO_FP's, as X86 doesn't have
187 setOperationAction(ISD::SINT_TO_FP , MVT::i1 , Promote);
188 setOperationAction(ISD::SINT_TO_FP , MVT::i8 , Promote);
190 if (!Subtarget->useSoftFloat()) {
191 // SSE has no i16 to fp conversion, only i32
192 if (X86ScalarSSEf32) {
193 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Promote);
194 // f32 and f64 cases are Legal, f80 case is not
195 setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Custom);
197 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Custom);
198 setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Custom);
201 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Promote);
202 setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Promote);
205 // In 32-bit mode these are custom lowered. In 64-bit mode F32 and F64
206 // are Legal, f80 is custom lowered.
207 setOperationAction(ISD::FP_TO_SINT , MVT::i64 , Custom);
208 setOperationAction(ISD::SINT_TO_FP , MVT::i64 , Custom);
210 // Promote i1/i8 FP_TO_SINT to larger FP_TO_SINTS's, as X86 doesn't have
212 setOperationAction(ISD::FP_TO_SINT , MVT::i1 , Promote);
213 setOperationAction(ISD::FP_TO_SINT , MVT::i8 , Promote);
215 if (X86ScalarSSEf32) {
216 setOperationAction(ISD::FP_TO_SINT , MVT::i16 , Promote);
217 // f32 and f64 cases are Legal, f80 case is not
218 setOperationAction(ISD::FP_TO_SINT , MVT::i32 , Custom);
220 setOperationAction(ISD::FP_TO_SINT , MVT::i16 , Custom);
221 setOperationAction(ISD::FP_TO_SINT , MVT::i32 , Custom);
224 // Handle FP_TO_UINT by promoting the destination to a larger signed
226 setOperationAction(ISD::FP_TO_UINT , MVT::i1 , Promote);
227 setOperationAction(ISD::FP_TO_UINT , MVT::i8 , Promote);
228 setOperationAction(ISD::FP_TO_UINT , MVT::i16 , Promote);
230 if (Subtarget->is64Bit()) {
231 setOperationAction(ISD::FP_TO_UINT , MVT::i64 , Expand);
232 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Promote);
233 } else if (!Subtarget->useSoftFloat()) {
234 // Since AVX is a superset of SSE3, only check for SSE here.
235 if (Subtarget->hasSSE1() && !Subtarget->hasSSE3())
236 // Expand FP_TO_UINT into a select.
237 // FIXME: We would like to use a Custom expander here eventually to do
238 // the optimal thing for SSE vs. the default expansion in the legalizer.
239 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Expand);
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);
246 if (isTargetFTOL()) {
247 // Use the _ftol2 runtime function, which has a pseudo-instruction
248 // to handle its weird calling convention.
249 setOperationAction(ISD::FP_TO_UINT , MVT::i64 , Custom);
252 // TODO: when we have SSE, these could be more efficient, by using movd/movq.
253 if (!X86ScalarSSEf64) {
254 setOperationAction(ISD::BITCAST , MVT::f32 , Expand);
255 setOperationAction(ISD::BITCAST , MVT::i32 , Expand);
256 if (Subtarget->is64Bit()) {
257 setOperationAction(ISD::BITCAST , MVT::f64 , Expand);
258 // Without SSE, i64->f64 goes through memory.
259 setOperationAction(ISD::BITCAST , MVT::i64 , Expand);
263 // Scalar integer divide and remainder are lowered to use operations that
264 // produce two results, to match the available instructions. This exposes
265 // the two-result form to trivial CSE, which is able to combine x/y and x%y
266 // into a single instruction.
268 // Scalar integer multiply-high is also lowered to use two-result
269 // operations, to match the available instructions. However, plain multiply
270 // (low) operations are left as Legal, as there are single-result
271 // instructions for this in x86. Using the two-result multiply instructions
272 // when both high and low results are needed must be arranged by dagcombine.
273 for (unsigned i = 0; i != array_lengthof(IntVTs); ++i) {
275 setOperationAction(ISD::MULHS, VT, Expand);
276 setOperationAction(ISD::MULHU, VT, Expand);
277 setOperationAction(ISD::SDIV, VT, Expand);
278 setOperationAction(ISD::UDIV, VT, Expand);
279 setOperationAction(ISD::SREM, VT, Expand);
280 setOperationAction(ISD::UREM, VT, Expand);
282 // Add/Sub overflow ops with MVT::Glues are lowered to EFLAGS dependences.
283 setOperationAction(ISD::ADDC, VT, Custom);
284 setOperationAction(ISD::ADDE, VT, Custom);
285 setOperationAction(ISD::SUBC, VT, Custom);
286 setOperationAction(ISD::SUBE, VT, Custom);
289 setOperationAction(ISD::BR_JT , MVT::Other, Expand);
290 setOperationAction(ISD::BRCOND , MVT::Other, Custom);
291 setOperationAction(ISD::BR_CC , MVT::f32, Expand);
292 setOperationAction(ISD::BR_CC , MVT::f64, Expand);
293 setOperationAction(ISD::BR_CC , MVT::f80, Expand);
294 setOperationAction(ISD::BR_CC , MVT::i8, Expand);
295 setOperationAction(ISD::BR_CC , MVT::i16, Expand);
296 setOperationAction(ISD::BR_CC , MVT::i32, Expand);
297 setOperationAction(ISD::BR_CC , MVT::i64, Expand);
298 setOperationAction(ISD::SELECT_CC , MVT::f32, Expand);
299 setOperationAction(ISD::SELECT_CC , MVT::f64, Expand);
300 setOperationAction(ISD::SELECT_CC , MVT::f80, Expand);
301 setOperationAction(ISD::SELECT_CC , MVT::i8, Expand);
302 setOperationAction(ISD::SELECT_CC , MVT::i16, Expand);
303 setOperationAction(ISD::SELECT_CC , MVT::i32, Expand);
304 setOperationAction(ISD::SELECT_CC , MVT::i64, Expand);
305 if (Subtarget->is64Bit())
306 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i32, Legal);
307 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16 , Legal);
308 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8 , Legal);
309 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1 , Expand);
310 setOperationAction(ISD::FP_ROUND_INREG , MVT::f32 , Expand);
311 setOperationAction(ISD::FREM , MVT::f32 , Expand);
312 setOperationAction(ISD::FREM , MVT::f64 , Expand);
313 setOperationAction(ISD::FREM , MVT::f80 , Expand);
314 setOperationAction(ISD::FLT_ROUNDS_ , MVT::i32 , Custom);
316 // Promote the i8 variants and force them on up to i32 which has a shorter
318 setOperationAction(ISD::CTTZ , MVT::i8 , Promote);
319 AddPromotedToType (ISD::CTTZ , MVT::i8 , MVT::i32);
320 setOperationAction(ISD::CTTZ_ZERO_UNDEF , MVT::i8 , Promote);
321 AddPromotedToType (ISD::CTTZ_ZERO_UNDEF , MVT::i8 , MVT::i32);
322 if (Subtarget->hasBMI()) {
323 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::i16 , Expand);
324 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::i32 , Expand);
325 if (Subtarget->is64Bit())
326 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::i64, Expand);
328 setOperationAction(ISD::CTTZ , MVT::i16 , Custom);
329 setOperationAction(ISD::CTTZ , MVT::i32 , Custom);
330 if (Subtarget->is64Bit())
331 setOperationAction(ISD::CTTZ , MVT::i64 , Custom);
334 if (Subtarget->hasLZCNT()) {
335 // When promoting the i8 variants, force them to i32 for a shorter
337 setOperationAction(ISD::CTLZ , MVT::i8 , Promote);
338 AddPromotedToType (ISD::CTLZ , MVT::i8 , MVT::i32);
339 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i8 , Promote);
340 AddPromotedToType (ISD::CTLZ_ZERO_UNDEF, MVT::i8 , MVT::i32);
341 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i16 , Expand);
342 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i32 , Expand);
343 if (Subtarget->is64Bit())
344 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i64, Expand);
346 setOperationAction(ISD::CTLZ , MVT::i8 , Custom);
347 setOperationAction(ISD::CTLZ , MVT::i16 , Custom);
348 setOperationAction(ISD::CTLZ , MVT::i32 , Custom);
349 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i8 , Custom);
350 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i16 , Custom);
351 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i32 , Custom);
352 if (Subtarget->is64Bit()) {
353 setOperationAction(ISD::CTLZ , MVT::i64 , Custom);
354 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i64, Custom);
358 // Special handling for half-precision floating point conversions.
359 // If we don't have F16C support, then lower half float conversions
360 // into library calls.
361 if (Subtarget->useSoftFloat() || !Subtarget->hasF16C()) {
362 setOperationAction(ISD::FP16_TO_FP, MVT::f32, Expand);
363 setOperationAction(ISD::FP_TO_FP16, MVT::f32, Expand);
366 // There's never any support for operations beyond MVT::f32.
367 setOperationAction(ISD::FP16_TO_FP, MVT::f64, Expand);
368 setOperationAction(ISD::FP16_TO_FP, MVT::f80, Expand);
369 setOperationAction(ISD::FP_TO_FP16, MVT::f64, Expand);
370 setOperationAction(ISD::FP_TO_FP16, MVT::f80, Expand);
372 setLoadExtAction(ISD::EXTLOAD, MVT::f32, MVT::f16, Expand);
373 setLoadExtAction(ISD::EXTLOAD, MVT::f64, MVT::f16, Expand);
374 setLoadExtAction(ISD::EXTLOAD, MVT::f80, MVT::f16, Expand);
375 setTruncStoreAction(MVT::f32, MVT::f16, Expand);
376 setTruncStoreAction(MVT::f64, MVT::f16, Expand);
377 setTruncStoreAction(MVT::f80, MVT::f16, Expand);
379 if (Subtarget->hasPOPCNT()) {
380 setOperationAction(ISD::CTPOP , MVT::i8 , Promote);
382 setOperationAction(ISD::CTPOP , MVT::i8 , Expand);
383 setOperationAction(ISD::CTPOP , MVT::i16 , Expand);
384 setOperationAction(ISD::CTPOP , MVT::i32 , Expand);
385 if (Subtarget->is64Bit())
386 setOperationAction(ISD::CTPOP , MVT::i64 , Expand);
389 setOperationAction(ISD::READCYCLECOUNTER , MVT::i64 , Custom);
391 if (!Subtarget->hasMOVBE())
392 setOperationAction(ISD::BSWAP , MVT::i16 , Expand);
394 // These should be promoted to a larger select which is supported.
395 setOperationAction(ISD::SELECT , MVT::i1 , Promote);
396 // X86 wants to expand cmov itself.
397 setOperationAction(ISD::SELECT , MVT::i8 , Custom);
398 setOperationAction(ISD::SELECT , MVT::i16 , Custom);
399 setOperationAction(ISD::SELECT , MVT::i32 , Custom);
400 setOperationAction(ISD::SELECT , MVT::f32 , Custom);
401 setOperationAction(ISD::SELECT , MVT::f64 , Custom);
402 setOperationAction(ISD::SELECT , MVT::f80 , Custom);
403 setOperationAction(ISD::SETCC , MVT::i8 , Custom);
404 setOperationAction(ISD::SETCC , MVT::i16 , Custom);
405 setOperationAction(ISD::SETCC , MVT::i32 , Custom);
406 setOperationAction(ISD::SETCC , MVT::f32 , Custom);
407 setOperationAction(ISD::SETCC , MVT::f64 , Custom);
408 setOperationAction(ISD::SETCC , MVT::f80 , Custom);
409 if (Subtarget->is64Bit()) {
410 setOperationAction(ISD::SELECT , MVT::i64 , Custom);
411 setOperationAction(ISD::SETCC , MVT::i64 , Custom);
413 setOperationAction(ISD::EH_RETURN , MVT::Other, Custom);
414 // NOTE: EH_SJLJ_SETJMP/_LONGJMP supported here is NOT intended to support
415 // SjLj exception handling but a light-weight setjmp/longjmp replacement to
416 // support continuation, user-level threading, and etc.. As a result, no
417 // other SjLj exception interfaces are implemented and please don't build
418 // your own exception handling based on them.
419 // LLVM/Clang supports zero-cost DWARF exception handling.
420 setOperationAction(ISD::EH_SJLJ_SETJMP, MVT::i32, Custom);
421 setOperationAction(ISD::EH_SJLJ_LONGJMP, MVT::Other, Custom);
424 setOperationAction(ISD::ConstantPool , MVT::i32 , Custom);
425 setOperationAction(ISD::JumpTable , MVT::i32 , Custom);
426 setOperationAction(ISD::GlobalAddress , MVT::i32 , Custom);
427 setOperationAction(ISD::GlobalTLSAddress, MVT::i32 , Custom);
428 if (Subtarget->is64Bit())
429 setOperationAction(ISD::GlobalTLSAddress, MVT::i64, Custom);
430 setOperationAction(ISD::ExternalSymbol , MVT::i32 , Custom);
431 setOperationAction(ISD::BlockAddress , MVT::i32 , Custom);
432 if (Subtarget->is64Bit()) {
433 setOperationAction(ISD::ConstantPool , MVT::i64 , Custom);
434 setOperationAction(ISD::JumpTable , MVT::i64 , Custom);
435 setOperationAction(ISD::GlobalAddress , MVT::i64 , Custom);
436 setOperationAction(ISD::ExternalSymbol, MVT::i64 , Custom);
437 setOperationAction(ISD::BlockAddress , MVT::i64 , Custom);
439 // 64-bit addm sub, shl, sra, srl (iff 32-bit x86)
440 setOperationAction(ISD::SHL_PARTS , MVT::i32 , Custom);
441 setOperationAction(ISD::SRA_PARTS , MVT::i32 , Custom);
442 setOperationAction(ISD::SRL_PARTS , MVT::i32 , Custom);
443 if (Subtarget->is64Bit()) {
444 setOperationAction(ISD::SHL_PARTS , MVT::i64 , Custom);
445 setOperationAction(ISD::SRA_PARTS , MVT::i64 , Custom);
446 setOperationAction(ISD::SRL_PARTS , MVT::i64 , Custom);
449 if (Subtarget->hasSSE1())
450 setOperationAction(ISD::PREFETCH , MVT::Other, Legal);
452 setOperationAction(ISD::ATOMIC_FENCE , MVT::Other, Custom);
454 // Expand certain atomics
455 for (unsigned i = 0; i != array_lengthof(IntVTs); ++i) {
457 setOperationAction(ISD::ATOMIC_CMP_SWAP_WITH_SUCCESS, VT, Custom);
458 setOperationAction(ISD::ATOMIC_LOAD_SUB, VT, Custom);
459 setOperationAction(ISD::ATOMIC_STORE, VT, Custom);
462 if (Subtarget->hasCmpxchg16b()) {
463 setOperationAction(ISD::ATOMIC_CMP_SWAP_WITH_SUCCESS, MVT::i128, Custom);
466 // FIXME - use subtarget debug flags
467 if (!Subtarget->isTargetDarwin() && !Subtarget->isTargetELF() &&
468 !Subtarget->isTargetCygMing() && !Subtarget->isTargetWin64()) {
469 setOperationAction(ISD::EH_LABEL, MVT::Other, Expand);
472 if (Subtarget->is64Bit()) {
473 setExceptionPointerRegister(X86::RAX);
474 setExceptionSelectorRegister(X86::RDX);
476 setExceptionPointerRegister(X86::EAX);
477 setExceptionSelectorRegister(X86::EDX);
479 setOperationAction(ISD::FRAME_TO_ARGS_OFFSET, MVT::i32, Custom);
480 setOperationAction(ISD::FRAME_TO_ARGS_OFFSET, MVT::i64, Custom);
482 setOperationAction(ISD::INIT_TRAMPOLINE, MVT::Other, Custom);
483 setOperationAction(ISD::ADJUST_TRAMPOLINE, MVT::Other, Custom);
485 setOperationAction(ISD::TRAP, MVT::Other, Legal);
486 setOperationAction(ISD::DEBUGTRAP, MVT::Other, Legal);
488 // VASTART needs to be custom lowered to use the VarArgsFrameIndex
489 setOperationAction(ISD::VASTART , MVT::Other, Custom);
490 setOperationAction(ISD::VAEND , MVT::Other, Expand);
491 if (Subtarget->is64Bit() && !Subtarget->isTargetWin64()) {
492 // TargetInfo::X86_64ABIBuiltinVaList
493 setOperationAction(ISD::VAARG , MVT::Other, Custom);
494 setOperationAction(ISD::VACOPY , MVT::Other, Custom);
496 // TargetInfo::CharPtrBuiltinVaList
497 setOperationAction(ISD::VAARG , MVT::Other, Expand);
498 setOperationAction(ISD::VACOPY , MVT::Other, Expand);
501 setOperationAction(ISD::STACKSAVE, MVT::Other, Expand);
502 setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand);
504 setOperationAction(ISD::DYNAMIC_STACKALLOC, PtrVT, Custom);
506 // GC_TRANSITION_START and GC_TRANSITION_END need custom lowering.
507 setOperationAction(ISD::GC_TRANSITION_START, MVT::Other, Custom);
508 setOperationAction(ISD::GC_TRANSITION_END, MVT::Other, Custom);
510 if (!Subtarget->useSoftFloat() && X86ScalarSSEf64) {
511 // f32 and f64 use SSE.
512 // Set up the FP register classes.
513 addRegisterClass(MVT::f32, &X86::FR32RegClass);
514 addRegisterClass(MVT::f64, &X86::FR64RegClass);
516 // Use ANDPD to simulate FABS.
517 setOperationAction(ISD::FABS , MVT::f64, Custom);
518 setOperationAction(ISD::FABS , MVT::f32, Custom);
520 // Use XORP to simulate FNEG.
521 setOperationAction(ISD::FNEG , MVT::f64, Custom);
522 setOperationAction(ISD::FNEG , MVT::f32, Custom);
524 // Use ANDPD and ORPD to simulate FCOPYSIGN.
525 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Custom);
526 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
528 // Lower this to FGETSIGNx86 plus an AND.
529 setOperationAction(ISD::FGETSIGN, MVT::i64, Custom);
530 setOperationAction(ISD::FGETSIGN, MVT::i32, Custom);
532 // We don't support sin/cos/fmod
533 setOperationAction(ISD::FSIN , MVT::f64, Expand);
534 setOperationAction(ISD::FCOS , MVT::f64, Expand);
535 setOperationAction(ISD::FSINCOS, MVT::f64, Expand);
536 setOperationAction(ISD::FSIN , MVT::f32, Expand);
537 setOperationAction(ISD::FCOS , MVT::f32, Expand);
538 setOperationAction(ISD::FSINCOS, MVT::f32, Expand);
540 // Expand FP immediates into loads from the stack, except for the special
542 addLegalFPImmediate(APFloat(+0.0)); // xorpd
543 addLegalFPImmediate(APFloat(+0.0f)); // xorps
544 } else if (!Subtarget->useSoftFloat() && X86ScalarSSEf32) {
545 // Use SSE for f32, x87 for f64.
546 // Set up the FP register classes.
547 addRegisterClass(MVT::f32, &X86::FR32RegClass);
548 addRegisterClass(MVT::f64, &X86::RFP64RegClass);
550 // Use ANDPS to simulate FABS.
551 setOperationAction(ISD::FABS , MVT::f32, Custom);
553 // Use XORP to simulate FNEG.
554 setOperationAction(ISD::FNEG , MVT::f32, Custom);
556 setOperationAction(ISD::UNDEF, MVT::f64, Expand);
558 // Use ANDPS and ORPS to simulate FCOPYSIGN.
559 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
560 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
562 // We don't support sin/cos/fmod
563 setOperationAction(ISD::FSIN , MVT::f32, Expand);
564 setOperationAction(ISD::FCOS , MVT::f32, Expand);
565 setOperationAction(ISD::FSINCOS, MVT::f32, Expand);
567 // Special cases we handle for FP constants.
568 addLegalFPImmediate(APFloat(+0.0f)); // xorps
569 addLegalFPImmediate(APFloat(+0.0)); // FLD0
570 addLegalFPImmediate(APFloat(+1.0)); // FLD1
571 addLegalFPImmediate(APFloat(-0.0)); // FLD0/FCHS
572 addLegalFPImmediate(APFloat(-1.0)); // FLD1/FCHS
574 if (!TM.Options.UnsafeFPMath) {
575 setOperationAction(ISD::FSIN , MVT::f64, Expand);
576 setOperationAction(ISD::FCOS , MVT::f64, Expand);
577 setOperationAction(ISD::FSINCOS, MVT::f64, Expand);
579 } else if (!Subtarget->useSoftFloat()) {
580 // f32 and f64 in x87.
581 // Set up the FP register classes.
582 addRegisterClass(MVT::f64, &X86::RFP64RegClass);
583 addRegisterClass(MVT::f32, &X86::RFP32RegClass);
585 setOperationAction(ISD::UNDEF, MVT::f64, Expand);
586 setOperationAction(ISD::UNDEF, MVT::f32, Expand);
587 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
588 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Expand);
590 if (!TM.Options.UnsafeFPMath) {
591 setOperationAction(ISD::FSIN , MVT::f64, Expand);
592 setOperationAction(ISD::FSIN , MVT::f32, Expand);
593 setOperationAction(ISD::FCOS , MVT::f64, Expand);
594 setOperationAction(ISD::FCOS , MVT::f32, Expand);
595 setOperationAction(ISD::FSINCOS, MVT::f64, Expand);
596 setOperationAction(ISD::FSINCOS, MVT::f32, Expand);
598 addLegalFPImmediate(APFloat(+0.0)); // FLD0
599 addLegalFPImmediate(APFloat(+1.0)); // FLD1
600 addLegalFPImmediate(APFloat(-0.0)); // FLD0/FCHS
601 addLegalFPImmediate(APFloat(-1.0)); // FLD1/FCHS
602 addLegalFPImmediate(APFloat(+0.0f)); // FLD0
603 addLegalFPImmediate(APFloat(+1.0f)); // FLD1
604 addLegalFPImmediate(APFloat(-0.0f)); // FLD0/FCHS
605 addLegalFPImmediate(APFloat(-1.0f)); // FLD1/FCHS
608 // We don't support FMA.
609 setOperationAction(ISD::FMA, MVT::f64, Expand);
610 setOperationAction(ISD::FMA, MVT::f32, Expand);
612 // Long double always uses X87.
613 if (!Subtarget->useSoftFloat()) {
614 addRegisterClass(MVT::f80, &X86::RFP80RegClass);
615 setOperationAction(ISD::UNDEF, MVT::f80, Expand);
616 setOperationAction(ISD::FCOPYSIGN, MVT::f80, Expand);
618 APFloat TmpFlt = APFloat::getZero(APFloat::x87DoubleExtended);
619 addLegalFPImmediate(TmpFlt); // FLD0
621 addLegalFPImmediate(TmpFlt); // FLD0/FCHS
624 APFloat TmpFlt2(+1.0);
625 TmpFlt2.convert(APFloat::x87DoubleExtended, APFloat::rmNearestTiesToEven,
627 addLegalFPImmediate(TmpFlt2); // FLD1
628 TmpFlt2.changeSign();
629 addLegalFPImmediate(TmpFlt2); // FLD1/FCHS
632 if (!TM.Options.UnsafeFPMath) {
633 setOperationAction(ISD::FSIN , MVT::f80, Expand);
634 setOperationAction(ISD::FCOS , MVT::f80, Expand);
635 setOperationAction(ISD::FSINCOS, MVT::f80, Expand);
638 setOperationAction(ISD::FFLOOR, MVT::f80, Expand);
639 setOperationAction(ISD::FCEIL, MVT::f80, Expand);
640 setOperationAction(ISD::FTRUNC, MVT::f80, Expand);
641 setOperationAction(ISD::FRINT, MVT::f80, Expand);
642 setOperationAction(ISD::FNEARBYINT, MVT::f80, Expand);
643 setOperationAction(ISD::FMA, MVT::f80, Expand);
646 // Always use a library call for pow.
647 setOperationAction(ISD::FPOW , MVT::f32 , Expand);
648 setOperationAction(ISD::FPOW , MVT::f64 , Expand);
649 setOperationAction(ISD::FPOW , MVT::f80 , Expand);
651 setOperationAction(ISD::FLOG, MVT::f80, Expand);
652 setOperationAction(ISD::FLOG2, MVT::f80, Expand);
653 setOperationAction(ISD::FLOG10, MVT::f80, Expand);
654 setOperationAction(ISD::FEXP, MVT::f80, Expand);
655 setOperationAction(ISD::FEXP2, MVT::f80, Expand);
656 setOperationAction(ISD::FMINNUM, MVT::f80, Expand);
657 setOperationAction(ISD::FMAXNUM, MVT::f80, Expand);
659 // First set operation action for all vector types to either promote
660 // (for widening) or expand (for scalarization). Then we will selectively
661 // turn on ones that can be effectively codegen'd.
662 for (MVT VT : MVT::vector_valuetypes()) {
663 setOperationAction(ISD::ADD , VT, Expand);
664 setOperationAction(ISD::SUB , VT, Expand);
665 setOperationAction(ISD::FADD, VT, Expand);
666 setOperationAction(ISD::FNEG, VT, Expand);
667 setOperationAction(ISD::FSUB, VT, Expand);
668 setOperationAction(ISD::MUL , VT, Expand);
669 setOperationAction(ISD::FMUL, VT, Expand);
670 setOperationAction(ISD::SDIV, VT, Expand);
671 setOperationAction(ISD::UDIV, VT, Expand);
672 setOperationAction(ISD::FDIV, VT, Expand);
673 setOperationAction(ISD::SREM, VT, Expand);
674 setOperationAction(ISD::UREM, VT, Expand);
675 setOperationAction(ISD::LOAD, VT, Expand);
676 setOperationAction(ISD::VECTOR_SHUFFLE, VT, Expand);
677 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT,Expand);
678 setOperationAction(ISD::INSERT_VECTOR_ELT, VT, Expand);
679 setOperationAction(ISD::EXTRACT_SUBVECTOR, VT,Expand);
680 setOperationAction(ISD::INSERT_SUBVECTOR, VT,Expand);
681 setOperationAction(ISD::FABS, VT, Expand);
682 setOperationAction(ISD::FSIN, VT, Expand);
683 setOperationAction(ISD::FSINCOS, VT, Expand);
684 setOperationAction(ISD::FCOS, VT, Expand);
685 setOperationAction(ISD::FSINCOS, VT, Expand);
686 setOperationAction(ISD::FREM, VT, Expand);
687 setOperationAction(ISD::FMA, VT, Expand);
688 setOperationAction(ISD::FPOWI, VT, Expand);
689 setOperationAction(ISD::FSQRT, VT, Expand);
690 setOperationAction(ISD::FCOPYSIGN, VT, Expand);
691 setOperationAction(ISD::FFLOOR, VT, Expand);
692 setOperationAction(ISD::FCEIL, VT, Expand);
693 setOperationAction(ISD::FTRUNC, VT, Expand);
694 setOperationAction(ISD::FRINT, VT, Expand);
695 setOperationAction(ISD::FNEARBYINT, VT, Expand);
696 setOperationAction(ISD::SMUL_LOHI, VT, Expand);
697 setOperationAction(ISD::MULHS, VT, Expand);
698 setOperationAction(ISD::UMUL_LOHI, VT, Expand);
699 setOperationAction(ISD::MULHU, VT, Expand);
700 setOperationAction(ISD::SDIVREM, VT, Expand);
701 setOperationAction(ISD::UDIVREM, VT, Expand);
702 setOperationAction(ISD::FPOW, VT, Expand);
703 setOperationAction(ISD::CTPOP, VT, Expand);
704 setOperationAction(ISD::CTTZ, VT, Expand);
705 setOperationAction(ISD::CTTZ_ZERO_UNDEF, VT, Expand);
706 setOperationAction(ISD::CTLZ, VT, Expand);
707 setOperationAction(ISD::CTLZ_ZERO_UNDEF, VT, Expand);
708 setOperationAction(ISD::SHL, VT, Expand);
709 setOperationAction(ISD::SRA, VT, Expand);
710 setOperationAction(ISD::SRL, VT, Expand);
711 setOperationAction(ISD::ROTL, VT, Expand);
712 setOperationAction(ISD::ROTR, VT, Expand);
713 setOperationAction(ISD::BSWAP, VT, Expand);
714 setOperationAction(ISD::SETCC, VT, Expand);
715 setOperationAction(ISD::FLOG, VT, Expand);
716 setOperationAction(ISD::FLOG2, VT, Expand);
717 setOperationAction(ISD::FLOG10, VT, Expand);
718 setOperationAction(ISD::FEXP, VT, Expand);
719 setOperationAction(ISD::FEXP2, VT, Expand);
720 setOperationAction(ISD::FP_TO_UINT, VT, Expand);
721 setOperationAction(ISD::FP_TO_SINT, VT, Expand);
722 setOperationAction(ISD::UINT_TO_FP, VT, Expand);
723 setOperationAction(ISD::SINT_TO_FP, VT, Expand);
724 setOperationAction(ISD::SIGN_EXTEND_INREG, VT,Expand);
725 setOperationAction(ISD::TRUNCATE, VT, Expand);
726 setOperationAction(ISD::SIGN_EXTEND, VT, Expand);
727 setOperationAction(ISD::ZERO_EXTEND, VT, Expand);
728 setOperationAction(ISD::ANY_EXTEND, VT, Expand);
729 setOperationAction(ISD::VSELECT, VT, Expand);
730 setOperationAction(ISD::SELECT_CC, VT, Expand);
731 for (MVT InnerVT : MVT::vector_valuetypes()) {
732 setTruncStoreAction(InnerVT, VT, Expand);
734 setLoadExtAction(ISD::SEXTLOAD, InnerVT, VT, Expand);
735 setLoadExtAction(ISD::ZEXTLOAD, InnerVT, VT, Expand);
737 // N.b. ISD::EXTLOAD legality is basically ignored except for i1-like
738 // types, we have to deal with them whether we ask for Expansion or not.
739 // Setting Expand causes its own optimisation problems though, so leave
741 if (VT.getVectorElementType() == MVT::i1)
742 setLoadExtAction(ISD::EXTLOAD, InnerVT, VT, Expand);
744 // EXTLOAD for MVT::f16 vectors is not legal because f16 vectors are
745 // split/scalarized right now.
746 if (VT.getVectorElementType() == MVT::f16)
747 setLoadExtAction(ISD::EXTLOAD, InnerVT, VT, Expand);
751 // FIXME: In order to prevent SSE instructions being expanded to MMX ones
752 // with -msoft-float, disable use of MMX as well.
753 if (!Subtarget->useSoftFloat() && Subtarget->hasMMX()) {
754 addRegisterClass(MVT::x86mmx, &X86::VR64RegClass);
755 // No operations on x86mmx supported, everything uses intrinsics.
758 // MMX-sized vectors (other than x86mmx) are expected to be expanded
759 // into smaller operations.
760 for (MVT MMXTy : {MVT::v8i8, MVT::v4i16, MVT::v2i32, MVT::v1i64}) {
761 setOperationAction(ISD::MULHS, MMXTy, Expand);
762 setOperationAction(ISD::AND, MMXTy, Expand);
763 setOperationAction(ISD::OR, MMXTy, Expand);
764 setOperationAction(ISD::XOR, MMXTy, Expand);
765 setOperationAction(ISD::SCALAR_TO_VECTOR, MMXTy, Expand);
766 setOperationAction(ISD::SELECT, MMXTy, Expand);
767 setOperationAction(ISD::BITCAST, MMXTy, Expand);
769 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v1i64, Expand);
771 if (!Subtarget->useSoftFloat() && Subtarget->hasSSE1()) {
772 addRegisterClass(MVT::v4f32, &X86::VR128RegClass);
774 setOperationAction(ISD::FADD, MVT::v4f32, Legal);
775 setOperationAction(ISD::FSUB, MVT::v4f32, Legal);
776 setOperationAction(ISD::FMUL, MVT::v4f32, Legal);
777 setOperationAction(ISD::FDIV, MVT::v4f32, Legal);
778 setOperationAction(ISD::FSQRT, MVT::v4f32, Legal);
779 setOperationAction(ISD::FNEG, MVT::v4f32, Custom);
780 setOperationAction(ISD::FABS, MVT::v4f32, Custom);
781 setOperationAction(ISD::LOAD, MVT::v4f32, Legal);
782 setOperationAction(ISD::BUILD_VECTOR, MVT::v4f32, Custom);
783 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4f32, Custom);
784 setOperationAction(ISD::VSELECT, MVT::v4f32, Custom);
785 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f32, Custom);
786 setOperationAction(ISD::SELECT, MVT::v4f32, Custom);
787 setOperationAction(ISD::UINT_TO_FP, MVT::v4i32, Custom);
790 if (!Subtarget->useSoftFloat() && Subtarget->hasSSE2()) {
791 addRegisterClass(MVT::v2f64, &X86::VR128RegClass);
793 // FIXME: Unfortunately, -soft-float and -no-implicit-float mean XMM
794 // registers cannot be used even for integer operations.
795 addRegisterClass(MVT::v16i8, &X86::VR128RegClass);
796 addRegisterClass(MVT::v8i16, &X86::VR128RegClass);
797 addRegisterClass(MVT::v4i32, &X86::VR128RegClass);
798 addRegisterClass(MVT::v2i64, &X86::VR128RegClass);
800 setOperationAction(ISD::ADD, MVT::v16i8, Legal);
801 setOperationAction(ISD::ADD, MVT::v8i16, Legal);
802 setOperationAction(ISD::ADD, MVT::v4i32, Legal);
803 setOperationAction(ISD::ADD, MVT::v2i64, Legal);
804 setOperationAction(ISD::MUL, MVT::v16i8, Custom);
805 setOperationAction(ISD::MUL, MVT::v4i32, Custom);
806 setOperationAction(ISD::MUL, MVT::v2i64, Custom);
807 setOperationAction(ISD::UMUL_LOHI, MVT::v4i32, Custom);
808 setOperationAction(ISD::SMUL_LOHI, MVT::v4i32, Custom);
809 setOperationAction(ISD::MULHU, MVT::v8i16, Legal);
810 setOperationAction(ISD::MULHS, MVT::v8i16, Legal);
811 setOperationAction(ISD::SUB, MVT::v16i8, Legal);
812 setOperationAction(ISD::SUB, MVT::v8i16, Legal);
813 setOperationAction(ISD::SUB, MVT::v4i32, Legal);
814 setOperationAction(ISD::SUB, MVT::v2i64, Legal);
815 setOperationAction(ISD::MUL, MVT::v8i16, Legal);
816 setOperationAction(ISD::FADD, MVT::v2f64, Legal);
817 setOperationAction(ISD::FSUB, MVT::v2f64, Legal);
818 setOperationAction(ISD::FMUL, MVT::v2f64, Legal);
819 setOperationAction(ISD::FDIV, MVT::v2f64, Legal);
820 setOperationAction(ISD::FSQRT, MVT::v2f64, Legal);
821 setOperationAction(ISD::FNEG, MVT::v2f64, Custom);
822 setOperationAction(ISD::FABS, MVT::v2f64, Custom);
824 setOperationAction(ISD::SMAX, MVT::v8i16, Legal);
825 setOperationAction(ISD::UMAX, MVT::v16i8, Legal);
826 setOperationAction(ISD::SMIN, MVT::v8i16, Legal);
827 setOperationAction(ISD::UMIN, MVT::v16i8, Legal);
829 setOperationAction(ISD::SETCC, MVT::v2i64, Custom);
830 setOperationAction(ISD::SETCC, MVT::v16i8, Custom);
831 setOperationAction(ISD::SETCC, MVT::v8i16, Custom);
832 setOperationAction(ISD::SETCC, MVT::v4i32, Custom);
834 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v16i8, Custom);
835 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v8i16, Custom);
836 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i16, Custom);
837 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i32, Custom);
838 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f32, Custom);
840 setOperationAction(ISD::CTPOP, MVT::v16i8, Custom);
841 setOperationAction(ISD::CTPOP, MVT::v8i16, Custom);
842 setOperationAction(ISD::CTPOP, MVT::v4i32, Custom);
843 setOperationAction(ISD::CTPOP, MVT::v2i64, Custom);
845 // Custom lower build_vector, vector_shuffle, and extract_vector_elt.
846 for (int i = MVT::v16i8; i != MVT::v2i64; ++i) {
847 MVT VT = (MVT::SimpleValueType)i;
848 // Do not attempt to custom lower non-power-of-2 vectors
849 if (!isPowerOf2_32(VT.getVectorNumElements()))
851 // Do not attempt to custom lower non-128-bit vectors
852 if (!VT.is128BitVector())
854 setOperationAction(ISD::BUILD_VECTOR, VT, Custom);
855 setOperationAction(ISD::VECTOR_SHUFFLE, VT, Custom);
856 setOperationAction(ISD::VSELECT, VT, Custom);
857 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Custom);
860 // We support custom legalizing of sext and anyext loads for specific
861 // memory vector types which we can load as a scalar (or sequence of
862 // scalars) and extend in-register to a legal 128-bit vector type. For sext
863 // loads these must work with a single scalar load.
864 for (MVT VT : MVT::integer_vector_valuetypes()) {
865 setLoadExtAction(ISD::SEXTLOAD, VT, MVT::v4i8, Custom);
866 setLoadExtAction(ISD::SEXTLOAD, VT, MVT::v4i16, Custom);
867 setLoadExtAction(ISD::SEXTLOAD, VT, MVT::v8i8, Custom);
868 setLoadExtAction(ISD::EXTLOAD, VT, MVT::v2i8, Custom);
869 setLoadExtAction(ISD::EXTLOAD, VT, MVT::v2i16, Custom);
870 setLoadExtAction(ISD::EXTLOAD, VT, MVT::v2i32, Custom);
871 setLoadExtAction(ISD::EXTLOAD, VT, MVT::v4i8, Custom);
872 setLoadExtAction(ISD::EXTLOAD, VT, MVT::v4i16, Custom);
873 setLoadExtAction(ISD::EXTLOAD, VT, MVT::v8i8, Custom);
876 setOperationAction(ISD::BUILD_VECTOR, MVT::v2f64, Custom);
877 setOperationAction(ISD::BUILD_VECTOR, MVT::v2i64, Custom);
878 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2f64, Custom);
879 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2i64, Custom);
880 setOperationAction(ISD::VSELECT, MVT::v2f64, Custom);
881 setOperationAction(ISD::VSELECT, MVT::v2i64, Custom);
882 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2f64, Custom);
883 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2f64, Custom);
885 if (Subtarget->is64Bit()) {
886 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2i64, Custom);
887 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2i64, Custom);
890 // Promote v16i8, v8i16, v4i32 load, select, and, or, xor to v2i64.
891 for (int i = MVT::v16i8; i != MVT::v2i64; ++i) {
892 MVT VT = (MVT::SimpleValueType)i;
894 // Do not attempt to promote non-128-bit vectors
895 if (!VT.is128BitVector())
898 setOperationAction(ISD::AND, VT, Promote);
899 AddPromotedToType (ISD::AND, VT, MVT::v2i64);
900 setOperationAction(ISD::OR, VT, Promote);
901 AddPromotedToType (ISD::OR, VT, MVT::v2i64);
902 setOperationAction(ISD::XOR, VT, Promote);
903 AddPromotedToType (ISD::XOR, VT, MVT::v2i64);
904 setOperationAction(ISD::LOAD, VT, Promote);
905 AddPromotedToType (ISD::LOAD, VT, MVT::v2i64);
906 setOperationAction(ISD::SELECT, VT, Promote);
907 AddPromotedToType (ISD::SELECT, VT, MVT::v2i64);
910 // Custom lower v2i64 and v2f64 selects.
911 setOperationAction(ISD::LOAD, MVT::v2f64, Legal);
912 setOperationAction(ISD::LOAD, MVT::v2i64, Legal);
913 setOperationAction(ISD::SELECT, MVT::v2f64, Custom);
914 setOperationAction(ISD::SELECT, MVT::v2i64, Custom);
916 setOperationAction(ISD::FP_TO_SINT, MVT::v4i32, Legal);
917 setOperationAction(ISD::SINT_TO_FP, MVT::v4i32, Legal);
919 setOperationAction(ISD::SINT_TO_FP, MVT::v2i32, Custom);
921 setOperationAction(ISD::UINT_TO_FP, MVT::v4i8, Custom);
922 setOperationAction(ISD::UINT_TO_FP, MVT::v4i16, Custom);
923 // As there is no 64-bit GPR available, we need build a special custom
924 // sequence to convert from v2i32 to v2f32.
925 if (!Subtarget->is64Bit())
926 setOperationAction(ISD::UINT_TO_FP, MVT::v2f32, Custom);
928 setOperationAction(ISD::FP_EXTEND, MVT::v2f32, Custom);
929 setOperationAction(ISD::FP_ROUND, MVT::v2f32, Custom);
931 for (MVT VT : MVT::fp_vector_valuetypes())
932 setLoadExtAction(ISD::EXTLOAD, VT, MVT::v2f32, Legal);
934 setOperationAction(ISD::BITCAST, MVT::v2i32, Custom);
935 setOperationAction(ISD::BITCAST, MVT::v4i16, Custom);
936 setOperationAction(ISD::BITCAST, MVT::v8i8, Custom);
939 if (!Subtarget->useSoftFloat() && Subtarget->hasSSE41()) {
940 for (MVT RoundedTy : {MVT::f32, MVT::f64, MVT::v4f32, MVT::v2f64}) {
941 setOperationAction(ISD::FFLOOR, RoundedTy, Legal);
942 setOperationAction(ISD::FCEIL, RoundedTy, Legal);
943 setOperationAction(ISD::FTRUNC, RoundedTy, Legal);
944 setOperationAction(ISD::FRINT, RoundedTy, Legal);
945 setOperationAction(ISD::FNEARBYINT, RoundedTy, Legal);
948 setOperationAction(ISD::SMAX, MVT::v16i8, Legal);
949 setOperationAction(ISD::SMAX, MVT::v4i32, Legal);
950 setOperationAction(ISD::UMAX, MVT::v8i16, Legal);
951 setOperationAction(ISD::UMAX, MVT::v4i32, Legal);
952 setOperationAction(ISD::SMIN, MVT::v16i8, Legal);
953 setOperationAction(ISD::SMIN, MVT::v4i32, Legal);
954 setOperationAction(ISD::UMIN, MVT::v8i16, Legal);
955 setOperationAction(ISD::UMIN, MVT::v4i32, Legal);
957 // FIXME: Do we need to handle scalar-to-vector here?
958 setOperationAction(ISD::MUL, MVT::v4i32, Legal);
960 // We directly match byte blends in the backend as they match the VSELECT
962 setOperationAction(ISD::VSELECT, MVT::v16i8, Legal);
964 // SSE41 brings specific instructions for doing vector sign extend even in
965 // cases where we don't have SRA.
966 for (MVT VT : MVT::integer_vector_valuetypes()) {
967 setLoadExtAction(ISD::SEXTLOAD, VT, MVT::v2i8, Custom);
968 setLoadExtAction(ISD::SEXTLOAD, VT, MVT::v2i16, Custom);
969 setLoadExtAction(ISD::SEXTLOAD, VT, MVT::v2i32, Custom);
972 // SSE41 also has vector sign/zero extending loads, PMOV[SZ]X
973 setLoadExtAction(ISD::SEXTLOAD, MVT::v8i16, MVT::v8i8, Legal);
974 setLoadExtAction(ISD::SEXTLOAD, MVT::v4i32, MVT::v4i8, Legal);
975 setLoadExtAction(ISD::SEXTLOAD, MVT::v2i64, MVT::v2i8, Legal);
976 setLoadExtAction(ISD::SEXTLOAD, MVT::v4i32, MVT::v4i16, Legal);
977 setLoadExtAction(ISD::SEXTLOAD, MVT::v2i64, MVT::v2i16, Legal);
978 setLoadExtAction(ISD::SEXTLOAD, MVT::v2i64, MVT::v2i32, Legal);
980 setLoadExtAction(ISD::ZEXTLOAD, MVT::v8i16, MVT::v8i8, Legal);
981 setLoadExtAction(ISD::ZEXTLOAD, MVT::v4i32, MVT::v4i8, Legal);
982 setLoadExtAction(ISD::ZEXTLOAD, MVT::v2i64, MVT::v2i8, Legal);
983 setLoadExtAction(ISD::ZEXTLOAD, MVT::v4i32, MVT::v4i16, Legal);
984 setLoadExtAction(ISD::ZEXTLOAD, MVT::v2i64, MVT::v2i16, Legal);
985 setLoadExtAction(ISD::ZEXTLOAD, MVT::v2i64, MVT::v2i32, Legal);
987 // i8 and i16 vectors are custom because the source register and source
988 // source memory operand types are not the same width. f32 vectors are
989 // custom since the immediate controlling the insert encodes additional
991 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v16i8, Custom);
992 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i16, Custom);
993 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i32, Custom);
994 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f32, Custom);
996 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v16i8, Custom);
997 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v8i16, Custom);
998 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4i32, Custom);
999 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f32, Custom);
1001 // FIXME: these should be Legal, but that's only for the case where
1002 // the index is constant. For now custom expand to deal with that.
1003 if (Subtarget->is64Bit()) {
1004 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2i64, Custom);
1005 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2i64, Custom);
1009 if (Subtarget->hasSSE2()) {
1010 setOperationAction(ISD::SIGN_EXTEND_VECTOR_INREG, MVT::v2i64, Custom);
1011 setOperationAction(ISD::SIGN_EXTEND_VECTOR_INREG, MVT::v4i32, Custom);
1012 setOperationAction(ISD::SIGN_EXTEND_VECTOR_INREG, MVT::v8i16, Custom);
1014 setOperationAction(ISD::SRL, MVT::v8i16, Custom);
1015 setOperationAction(ISD::SRL, MVT::v16i8, Custom);
1017 setOperationAction(ISD::SHL, MVT::v8i16, Custom);
1018 setOperationAction(ISD::SHL, MVT::v16i8, Custom);
1020 setOperationAction(ISD::SRA, MVT::v8i16, Custom);
1021 setOperationAction(ISD::SRA, MVT::v16i8, Custom);
1023 // In the customized shift lowering, the legal cases in AVX2 will be
1025 setOperationAction(ISD::SRL, MVT::v2i64, Custom);
1026 setOperationAction(ISD::SRL, MVT::v4i32, Custom);
1028 setOperationAction(ISD::SHL, MVT::v2i64, Custom);
1029 setOperationAction(ISD::SHL, MVT::v4i32, Custom);
1031 setOperationAction(ISD::SRA, MVT::v2i64, Custom);
1032 setOperationAction(ISD::SRA, MVT::v4i32, Custom);
1035 if (!Subtarget->useSoftFloat() && Subtarget->hasFp256()) {
1036 addRegisterClass(MVT::v32i8, &X86::VR256RegClass);
1037 addRegisterClass(MVT::v16i16, &X86::VR256RegClass);
1038 addRegisterClass(MVT::v8i32, &X86::VR256RegClass);
1039 addRegisterClass(MVT::v8f32, &X86::VR256RegClass);
1040 addRegisterClass(MVT::v4i64, &X86::VR256RegClass);
1041 addRegisterClass(MVT::v4f64, &X86::VR256RegClass);
1043 setOperationAction(ISD::LOAD, MVT::v8f32, Legal);
1044 setOperationAction(ISD::LOAD, MVT::v4f64, Legal);
1045 setOperationAction(ISD::LOAD, MVT::v4i64, Legal);
1047 setOperationAction(ISD::FADD, MVT::v8f32, Legal);
1048 setOperationAction(ISD::FSUB, MVT::v8f32, Legal);
1049 setOperationAction(ISD::FMUL, MVT::v8f32, Legal);
1050 setOperationAction(ISD::FDIV, MVT::v8f32, Legal);
1051 setOperationAction(ISD::FSQRT, MVT::v8f32, Legal);
1052 setOperationAction(ISD::FFLOOR, MVT::v8f32, Legal);
1053 setOperationAction(ISD::FCEIL, MVT::v8f32, Legal);
1054 setOperationAction(ISD::FTRUNC, MVT::v8f32, Legal);
1055 setOperationAction(ISD::FRINT, MVT::v8f32, Legal);
1056 setOperationAction(ISD::FNEARBYINT, MVT::v8f32, Legal);
1057 setOperationAction(ISD::FNEG, MVT::v8f32, Custom);
1058 setOperationAction(ISD::FABS, MVT::v8f32, Custom);
1060 setOperationAction(ISD::FADD, MVT::v4f64, Legal);
1061 setOperationAction(ISD::FSUB, MVT::v4f64, Legal);
1062 setOperationAction(ISD::FMUL, MVT::v4f64, Legal);
1063 setOperationAction(ISD::FDIV, MVT::v4f64, Legal);
1064 setOperationAction(ISD::FSQRT, MVT::v4f64, Legal);
1065 setOperationAction(ISD::FFLOOR, MVT::v4f64, Legal);
1066 setOperationAction(ISD::FCEIL, MVT::v4f64, Legal);
1067 setOperationAction(ISD::FTRUNC, MVT::v4f64, Legal);
1068 setOperationAction(ISD::FRINT, MVT::v4f64, Legal);
1069 setOperationAction(ISD::FNEARBYINT, MVT::v4f64, Legal);
1070 setOperationAction(ISD::FNEG, MVT::v4f64, Custom);
1071 setOperationAction(ISD::FABS, MVT::v4f64, Custom);
1073 // (fp_to_int:v8i16 (v8f32 ..)) requires the result type to be promoted
1074 // even though v8i16 is a legal type.
1075 setOperationAction(ISD::FP_TO_SINT, MVT::v8i16, Promote);
1076 setOperationAction(ISD::FP_TO_UINT, MVT::v8i16, Promote);
1077 setOperationAction(ISD::FP_TO_SINT, MVT::v8i32, Legal);
1079 setOperationAction(ISD::SINT_TO_FP, MVT::v8i16, Promote);
1080 setOperationAction(ISD::SINT_TO_FP, MVT::v8i32, Legal);
1081 setOperationAction(ISD::FP_ROUND, MVT::v4f32, Legal);
1083 setOperationAction(ISD::UINT_TO_FP, MVT::v8i8, Custom);
1084 setOperationAction(ISD::UINT_TO_FP, MVT::v8i16, Custom);
1086 for (MVT VT : MVT::fp_vector_valuetypes())
1087 setLoadExtAction(ISD::EXTLOAD, VT, MVT::v4f32, Legal);
1089 setOperationAction(ISD::SRL, MVT::v16i16, Custom);
1090 setOperationAction(ISD::SRL, MVT::v32i8, Custom);
1092 setOperationAction(ISD::SHL, MVT::v16i16, Custom);
1093 setOperationAction(ISD::SHL, MVT::v32i8, Custom);
1095 setOperationAction(ISD::SRA, MVT::v16i16, Custom);
1096 setOperationAction(ISD::SRA, MVT::v32i8, Custom);
1098 setOperationAction(ISD::SETCC, MVT::v32i8, Custom);
1099 setOperationAction(ISD::SETCC, MVT::v16i16, Custom);
1100 setOperationAction(ISD::SETCC, MVT::v8i32, Custom);
1101 setOperationAction(ISD::SETCC, MVT::v4i64, Custom);
1103 setOperationAction(ISD::SELECT, MVT::v4f64, Custom);
1104 setOperationAction(ISD::SELECT, MVT::v4i64, Custom);
1105 setOperationAction(ISD::SELECT, MVT::v8f32, Custom);
1107 setOperationAction(ISD::SIGN_EXTEND, MVT::v4i64, Custom);
1108 setOperationAction(ISD::SIGN_EXTEND, MVT::v8i32, Custom);
1109 setOperationAction(ISD::SIGN_EXTEND, MVT::v16i16, Custom);
1110 setOperationAction(ISD::ZERO_EXTEND, MVT::v4i64, Custom);
1111 setOperationAction(ISD::ZERO_EXTEND, MVT::v8i32, Custom);
1112 setOperationAction(ISD::ZERO_EXTEND, MVT::v16i16, Custom);
1113 setOperationAction(ISD::ANY_EXTEND, MVT::v4i64, Custom);
1114 setOperationAction(ISD::ANY_EXTEND, MVT::v8i32, Custom);
1115 setOperationAction(ISD::ANY_EXTEND, MVT::v16i16, Custom);
1116 setOperationAction(ISD::TRUNCATE, MVT::v16i8, Custom);
1117 setOperationAction(ISD::TRUNCATE, MVT::v8i16, Custom);
1118 setOperationAction(ISD::TRUNCATE, MVT::v4i32, Custom);
1120 setOperationAction(ISD::CTPOP, MVT::v32i8, Custom);
1121 setOperationAction(ISD::CTPOP, MVT::v16i16, Custom);
1122 setOperationAction(ISD::CTPOP, MVT::v8i32, Custom);
1123 setOperationAction(ISD::CTPOP, MVT::v4i64, Custom);
1125 if (Subtarget->hasFMA() || Subtarget->hasFMA4() || Subtarget->hasAVX512()) {
1126 setOperationAction(ISD::FMA, MVT::v8f32, Legal);
1127 setOperationAction(ISD::FMA, MVT::v4f64, Legal);
1128 setOperationAction(ISD::FMA, MVT::v4f32, Legal);
1129 setOperationAction(ISD::FMA, MVT::v2f64, Legal);
1130 setOperationAction(ISD::FMA, MVT::f32, Legal);
1131 setOperationAction(ISD::FMA, MVT::f64, Legal);
1134 if (Subtarget->hasInt256()) {
1135 setOperationAction(ISD::ADD, MVT::v4i64, Legal);
1136 setOperationAction(ISD::ADD, MVT::v8i32, Legal);
1137 setOperationAction(ISD::ADD, MVT::v16i16, Legal);
1138 setOperationAction(ISD::ADD, MVT::v32i8, Legal);
1140 setOperationAction(ISD::SUB, MVT::v4i64, Legal);
1141 setOperationAction(ISD::SUB, MVT::v8i32, Legal);
1142 setOperationAction(ISD::SUB, MVT::v16i16, Legal);
1143 setOperationAction(ISD::SUB, MVT::v32i8, Legal);
1145 setOperationAction(ISD::MUL, MVT::v4i64, Custom);
1146 setOperationAction(ISD::MUL, MVT::v8i32, Legal);
1147 setOperationAction(ISD::MUL, MVT::v16i16, Legal);
1148 setOperationAction(ISD::MUL, MVT::v32i8, Custom);
1150 setOperationAction(ISD::UMUL_LOHI, MVT::v8i32, Custom);
1151 setOperationAction(ISD::SMUL_LOHI, MVT::v8i32, Custom);
1152 setOperationAction(ISD::MULHU, MVT::v16i16, Legal);
1153 setOperationAction(ISD::MULHS, MVT::v16i16, Legal);
1155 setOperationAction(ISD::SMAX, MVT::v32i8, Legal);
1156 setOperationAction(ISD::SMAX, MVT::v16i16, Legal);
1157 setOperationAction(ISD::SMAX, MVT::v8i32, Legal);
1158 setOperationAction(ISD::UMAX, MVT::v32i8, Legal);
1159 setOperationAction(ISD::UMAX, MVT::v16i16, Legal);
1160 setOperationAction(ISD::UMAX, MVT::v8i32, Legal);
1161 setOperationAction(ISD::SMIN, MVT::v32i8, Legal);
1162 setOperationAction(ISD::SMIN, MVT::v16i16, Legal);
1163 setOperationAction(ISD::SMIN, MVT::v8i32, Legal);
1164 setOperationAction(ISD::UMIN, MVT::v32i8, Legal);
1165 setOperationAction(ISD::UMIN, MVT::v16i16, Legal);
1166 setOperationAction(ISD::UMIN, MVT::v8i32, Legal);
1168 // The custom lowering for UINT_TO_FP for v8i32 becomes interesting
1169 // when we have a 256bit-wide blend with immediate.
1170 setOperationAction(ISD::UINT_TO_FP, MVT::v8i32, Custom);
1172 // AVX2 also has wider vector sign/zero extending loads, VPMOV[SZ]X
1173 setLoadExtAction(ISD::SEXTLOAD, MVT::v16i16, MVT::v16i8, Legal);
1174 setLoadExtAction(ISD::SEXTLOAD, MVT::v8i32, MVT::v8i8, Legal);
1175 setLoadExtAction(ISD::SEXTLOAD, MVT::v4i64, MVT::v4i8, Legal);
1176 setLoadExtAction(ISD::SEXTLOAD, MVT::v8i32, MVT::v8i16, Legal);
1177 setLoadExtAction(ISD::SEXTLOAD, MVT::v4i64, MVT::v4i16, Legal);
1178 setLoadExtAction(ISD::SEXTLOAD, MVT::v4i64, MVT::v4i32, Legal);
1180 setLoadExtAction(ISD::ZEXTLOAD, MVT::v16i16, MVT::v16i8, Legal);
1181 setLoadExtAction(ISD::ZEXTLOAD, MVT::v8i32, MVT::v8i8, Legal);
1182 setLoadExtAction(ISD::ZEXTLOAD, MVT::v4i64, MVT::v4i8, Legal);
1183 setLoadExtAction(ISD::ZEXTLOAD, MVT::v8i32, MVT::v8i16, Legal);
1184 setLoadExtAction(ISD::ZEXTLOAD, MVT::v4i64, MVT::v4i16, Legal);
1185 setLoadExtAction(ISD::ZEXTLOAD, MVT::v4i64, MVT::v4i32, Legal);
1187 setOperationAction(ISD::ADD, MVT::v4i64, Custom);
1188 setOperationAction(ISD::ADD, MVT::v8i32, Custom);
1189 setOperationAction(ISD::ADD, MVT::v16i16, Custom);
1190 setOperationAction(ISD::ADD, MVT::v32i8, Custom);
1192 setOperationAction(ISD::SUB, MVT::v4i64, Custom);
1193 setOperationAction(ISD::SUB, MVT::v8i32, Custom);
1194 setOperationAction(ISD::SUB, MVT::v16i16, Custom);
1195 setOperationAction(ISD::SUB, MVT::v32i8, Custom);
1197 setOperationAction(ISD::MUL, MVT::v4i64, Custom);
1198 setOperationAction(ISD::MUL, MVT::v8i32, Custom);
1199 setOperationAction(ISD::MUL, MVT::v16i16, Custom);
1200 setOperationAction(ISD::MUL, MVT::v32i8, Custom);
1203 // In the customized shift lowering, the legal cases in AVX2 will be
1205 setOperationAction(ISD::SRL, MVT::v4i64, Custom);
1206 setOperationAction(ISD::SRL, MVT::v8i32, Custom);
1208 setOperationAction(ISD::SHL, MVT::v4i64, Custom);
1209 setOperationAction(ISD::SHL, MVT::v8i32, Custom);
1211 setOperationAction(ISD::SRA, MVT::v4i64, Custom);
1212 setOperationAction(ISD::SRA, MVT::v8i32, Custom);
1214 // Custom lower several nodes for 256-bit types.
1215 for (MVT VT : MVT::vector_valuetypes()) {
1216 if (VT.getScalarSizeInBits() >= 32) {
1217 setOperationAction(ISD::MLOAD, VT, Legal);
1218 setOperationAction(ISD::MSTORE, VT, Legal);
1220 // Extract subvector is special because the value type
1221 // (result) is 128-bit but the source is 256-bit wide.
1222 if (VT.is128BitVector()) {
1223 setOperationAction(ISD::EXTRACT_SUBVECTOR, VT, Custom);
1225 // Do not attempt to custom lower other non-256-bit vectors
1226 if (!VT.is256BitVector())
1229 setOperationAction(ISD::BUILD_VECTOR, VT, Custom);
1230 setOperationAction(ISD::VECTOR_SHUFFLE, VT, Custom);
1231 setOperationAction(ISD::VSELECT, VT, Custom);
1232 setOperationAction(ISD::INSERT_VECTOR_ELT, VT, Custom);
1233 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Custom);
1234 setOperationAction(ISD::SCALAR_TO_VECTOR, VT, Custom);
1235 setOperationAction(ISD::INSERT_SUBVECTOR, VT, Custom);
1236 setOperationAction(ISD::CONCAT_VECTORS, VT, Custom);
1239 if (Subtarget->hasInt256())
1240 setOperationAction(ISD::VSELECT, MVT::v32i8, Legal);
1243 // Promote v32i8, v16i16, v8i32 select, and, or, xor to v4i64.
1244 for (int i = MVT::v32i8; i != MVT::v4i64; ++i) {
1245 MVT VT = (MVT::SimpleValueType)i;
1247 // Do not attempt to promote non-256-bit vectors
1248 if (!VT.is256BitVector())
1251 setOperationAction(ISD::AND, VT, Promote);
1252 AddPromotedToType (ISD::AND, VT, MVT::v4i64);
1253 setOperationAction(ISD::OR, VT, Promote);
1254 AddPromotedToType (ISD::OR, VT, MVT::v4i64);
1255 setOperationAction(ISD::XOR, VT, Promote);
1256 AddPromotedToType (ISD::XOR, VT, MVT::v4i64);
1257 setOperationAction(ISD::LOAD, VT, Promote);
1258 AddPromotedToType (ISD::LOAD, VT, MVT::v4i64);
1259 setOperationAction(ISD::SELECT, VT, Promote);
1260 AddPromotedToType (ISD::SELECT, VT, MVT::v4i64);
1264 if (!Subtarget->useSoftFloat() && Subtarget->hasAVX512()) {
1265 addRegisterClass(MVT::v16i32, &X86::VR512RegClass);
1266 addRegisterClass(MVT::v16f32, &X86::VR512RegClass);
1267 addRegisterClass(MVT::v8i64, &X86::VR512RegClass);
1268 addRegisterClass(MVT::v8f64, &X86::VR512RegClass);
1270 addRegisterClass(MVT::i1, &X86::VK1RegClass);
1271 addRegisterClass(MVT::v8i1, &X86::VK8RegClass);
1272 addRegisterClass(MVT::v16i1, &X86::VK16RegClass);
1274 for (MVT VT : MVT::fp_vector_valuetypes())
1275 setLoadExtAction(ISD::EXTLOAD, VT, MVT::v8f32, Legal);
1277 setLoadExtAction(ISD::ZEXTLOAD, MVT::v16i32, MVT::v16i8, Legal);
1278 setLoadExtAction(ISD::SEXTLOAD, MVT::v16i32, MVT::v16i8, Legal);
1279 setLoadExtAction(ISD::ZEXTLOAD, MVT::v16i32, MVT::v16i16, Legal);
1280 setLoadExtAction(ISD::SEXTLOAD, MVT::v16i32, MVT::v16i16, Legal);
1281 setLoadExtAction(ISD::ZEXTLOAD, MVT::v32i16, MVT::v32i8, Legal);
1282 setLoadExtAction(ISD::SEXTLOAD, MVT::v32i16, MVT::v32i8, Legal);
1283 setLoadExtAction(ISD::ZEXTLOAD, MVT::v8i64, MVT::v8i8, Legal);
1284 setLoadExtAction(ISD::SEXTLOAD, MVT::v8i64, MVT::v8i8, Legal);
1285 setLoadExtAction(ISD::ZEXTLOAD, MVT::v8i64, MVT::v8i16, Legal);
1286 setLoadExtAction(ISD::SEXTLOAD, MVT::v8i64, MVT::v8i16, Legal);
1287 setLoadExtAction(ISD::ZEXTLOAD, MVT::v8i64, MVT::v8i32, Legal);
1288 setLoadExtAction(ISD::SEXTLOAD, MVT::v8i64, MVT::v8i32, Legal);
1290 setOperationAction(ISD::BR_CC, MVT::i1, Expand);
1291 setOperationAction(ISD::SETCC, MVT::i1, Custom);
1292 setOperationAction(ISD::XOR, MVT::i1, Legal);
1293 setOperationAction(ISD::OR, MVT::i1, Legal);
1294 setOperationAction(ISD::AND, MVT::i1, Legal);
1295 setOperationAction(ISD::SUB, MVT::i1, Custom);
1296 setOperationAction(ISD::ADD, MVT::i1, Custom);
1297 setOperationAction(ISD::MUL, MVT::i1, Custom);
1298 setOperationAction(ISD::LOAD, MVT::v16f32, Legal);
1299 setOperationAction(ISD::LOAD, MVT::v8f64, Legal);
1300 setOperationAction(ISD::LOAD, MVT::v8i64, Legal);
1301 setOperationAction(ISD::LOAD, MVT::v16i32, Legal);
1302 setOperationAction(ISD::LOAD, MVT::v16i1, Legal);
1304 setOperationAction(ISD::FADD, MVT::v16f32, Legal);
1305 setOperationAction(ISD::FSUB, MVT::v16f32, Legal);
1306 setOperationAction(ISD::FMUL, MVT::v16f32, Legal);
1307 setOperationAction(ISD::FDIV, MVT::v16f32, Legal);
1308 setOperationAction(ISD::FSQRT, MVT::v16f32, Legal);
1309 setOperationAction(ISD::FNEG, MVT::v16f32, Custom);
1311 setOperationAction(ISD::FADD, MVT::v8f64, Legal);
1312 setOperationAction(ISD::FSUB, MVT::v8f64, Legal);
1313 setOperationAction(ISD::FMUL, MVT::v8f64, Legal);
1314 setOperationAction(ISD::FDIV, MVT::v8f64, Legal);
1315 setOperationAction(ISD::FSQRT, MVT::v8f64, Legal);
1316 setOperationAction(ISD::FNEG, MVT::v8f64, Custom);
1317 setOperationAction(ISD::FMA, MVT::v8f64, Legal);
1318 setOperationAction(ISD::FMA, MVT::v16f32, Legal);
1320 setOperationAction(ISD::FP_TO_SINT, MVT::i32, Legal);
1321 setOperationAction(ISD::FP_TO_UINT, MVT::i32, Legal);
1322 setOperationAction(ISD::SINT_TO_FP, MVT::i32, Legal);
1323 setOperationAction(ISD::UINT_TO_FP, MVT::i32, Legal);
1324 if (Subtarget->is64Bit()) {
1325 setOperationAction(ISD::FP_TO_UINT, MVT::i64, Legal);
1326 setOperationAction(ISD::FP_TO_SINT, MVT::i64, Legal);
1327 setOperationAction(ISD::SINT_TO_FP, MVT::i64, Legal);
1328 setOperationAction(ISD::UINT_TO_FP, MVT::i64, Legal);
1330 setOperationAction(ISD::FP_TO_SINT, MVT::v16i32, Legal);
1331 setOperationAction(ISD::FP_TO_UINT, MVT::v16i32, Legal);
1332 setOperationAction(ISD::FP_TO_UINT, MVT::v8i32, Legal);
1333 setOperationAction(ISD::FP_TO_UINT, MVT::v4i32, Legal);
1334 setOperationAction(ISD::SINT_TO_FP, MVT::v16i32, Legal);
1335 setOperationAction(ISD::SINT_TO_FP, MVT::v8i1, Custom);
1336 setOperationAction(ISD::SINT_TO_FP, MVT::v16i1, Custom);
1337 setOperationAction(ISD::SINT_TO_FP, MVT::v16i8, Promote);
1338 setOperationAction(ISD::SINT_TO_FP, MVT::v16i16, Promote);
1339 setOperationAction(ISD::UINT_TO_FP, MVT::v16i32, Legal);
1340 setOperationAction(ISD::UINT_TO_FP, MVT::v8i32, Legal);
1341 setOperationAction(ISD::UINT_TO_FP, MVT::v4i32, Legal);
1342 setOperationAction(ISD::UINT_TO_FP, MVT::v16i8, Custom);
1343 setOperationAction(ISD::UINT_TO_FP, MVT::v16i16, Custom);
1344 setOperationAction(ISD::FP_ROUND, MVT::v8f32, Legal);
1345 setOperationAction(ISD::FP_EXTEND, MVT::v8f32, Legal);
1347 setTruncStoreAction(MVT::v8i64, MVT::v8i8, Legal);
1348 setTruncStoreAction(MVT::v8i64, MVT::v8i16, Legal);
1349 setTruncStoreAction(MVT::v8i64, MVT::v8i32, Legal);
1350 setTruncStoreAction(MVT::v16i32, MVT::v16i8, Legal);
1351 setTruncStoreAction(MVT::v16i32, MVT::v16i16, Legal);
1352 if (Subtarget->hasVLX()){
1353 setTruncStoreAction(MVT::v4i64, MVT::v4i8, Legal);
1354 setTruncStoreAction(MVT::v4i64, MVT::v4i16, Legal);
1355 setTruncStoreAction(MVT::v4i64, MVT::v4i32, Legal);
1356 setTruncStoreAction(MVT::v8i32, MVT::v8i8, Legal);
1357 setTruncStoreAction(MVT::v8i32, MVT::v8i16, Legal);
1359 setTruncStoreAction(MVT::v2i64, MVT::v2i8, Legal);
1360 setTruncStoreAction(MVT::v2i64, MVT::v2i16, Legal);
1361 setTruncStoreAction(MVT::v2i64, MVT::v2i32, Legal);
1362 setTruncStoreAction(MVT::v4i32, MVT::v4i8, Legal);
1363 setTruncStoreAction(MVT::v4i32, MVT::v4i16, Legal);
1365 setOperationAction(ISD::TRUNCATE, MVT::i1, Custom);
1366 setOperationAction(ISD::TRUNCATE, MVT::v16i8, Custom);
1367 setOperationAction(ISD::TRUNCATE, MVT::v8i32, Custom);
1368 if (Subtarget->hasDQI()) {
1369 setOperationAction(ISD::TRUNCATE, MVT::v2i1, Custom);
1370 setOperationAction(ISD::TRUNCATE, MVT::v4i1, Custom);
1372 setOperationAction(ISD::SINT_TO_FP, MVT::v8i64, Legal);
1373 setOperationAction(ISD::UINT_TO_FP, MVT::v8i64, Legal);
1374 setOperationAction(ISD::FP_TO_SINT, MVT::v8i64, Legal);
1375 setOperationAction(ISD::FP_TO_UINT, MVT::v8i64, Legal);
1376 if (Subtarget->hasVLX()) {
1377 setOperationAction(ISD::SINT_TO_FP, MVT::v4i64, Legal);
1378 setOperationAction(ISD::SINT_TO_FP, MVT::v2i64, Legal);
1379 setOperationAction(ISD::UINT_TO_FP, MVT::v4i64, Legal);
1380 setOperationAction(ISD::UINT_TO_FP, MVT::v2i64, Legal);
1381 setOperationAction(ISD::FP_TO_SINT, MVT::v4i64, Legal);
1382 setOperationAction(ISD::FP_TO_SINT, MVT::v2i64, Legal);
1383 setOperationAction(ISD::FP_TO_UINT, MVT::v4i64, Legal);
1384 setOperationAction(ISD::FP_TO_UINT, MVT::v2i64, Legal);
1387 if (Subtarget->hasVLX()) {
1388 setOperationAction(ISD::SINT_TO_FP, MVT::v8i32, Legal);
1389 setOperationAction(ISD::UINT_TO_FP, MVT::v8i32, Legal);
1390 setOperationAction(ISD::FP_TO_SINT, MVT::v8i32, Legal);
1391 setOperationAction(ISD::FP_TO_UINT, MVT::v8i32, Legal);
1392 setOperationAction(ISD::SINT_TO_FP, MVT::v4i32, Legal);
1393 setOperationAction(ISD::UINT_TO_FP, MVT::v4i32, Legal);
1394 setOperationAction(ISD::FP_TO_SINT, MVT::v4i32, Legal);
1395 setOperationAction(ISD::FP_TO_UINT, MVT::v4i32, Legal);
1397 setOperationAction(ISD::TRUNCATE, MVT::v8i1, Custom);
1398 setOperationAction(ISD::TRUNCATE, MVT::v16i1, Custom);
1399 setOperationAction(ISD::TRUNCATE, MVT::v16i16, Custom);
1400 setOperationAction(ISD::ZERO_EXTEND, MVT::v16i32, Custom);
1401 setOperationAction(ISD::ZERO_EXTEND, MVT::v8i64, Custom);
1402 setOperationAction(ISD::ANY_EXTEND, MVT::v16i32, Custom);
1403 setOperationAction(ISD::ANY_EXTEND, MVT::v8i64, Custom);
1404 setOperationAction(ISD::SIGN_EXTEND, MVT::v16i32, Custom);
1405 setOperationAction(ISD::SIGN_EXTEND, MVT::v8i64, Custom);
1406 setOperationAction(ISD::SIGN_EXTEND, MVT::v16i8, Custom);
1407 setOperationAction(ISD::SIGN_EXTEND, MVT::v8i16, Custom);
1408 setOperationAction(ISD::SIGN_EXTEND, MVT::v16i16, Custom);
1409 if (Subtarget->hasDQI()) {
1410 setOperationAction(ISD::SIGN_EXTEND, MVT::v4i32, Custom);
1411 setOperationAction(ISD::SIGN_EXTEND, MVT::v2i64, Custom);
1413 setOperationAction(ISD::FFLOOR, MVT::v16f32, Legal);
1414 setOperationAction(ISD::FFLOOR, MVT::v8f64, Legal);
1415 setOperationAction(ISD::FCEIL, MVT::v16f32, Legal);
1416 setOperationAction(ISD::FCEIL, MVT::v8f64, Legal);
1417 setOperationAction(ISD::FTRUNC, MVT::v16f32, Legal);
1418 setOperationAction(ISD::FTRUNC, MVT::v8f64, Legal);
1419 setOperationAction(ISD::FRINT, MVT::v16f32, Legal);
1420 setOperationAction(ISD::FRINT, MVT::v8f64, Legal);
1421 setOperationAction(ISD::FNEARBYINT, MVT::v16f32, Legal);
1422 setOperationAction(ISD::FNEARBYINT, MVT::v8f64, Legal);
1424 setOperationAction(ISD::CONCAT_VECTORS, MVT::v8f64, Custom);
1425 setOperationAction(ISD::CONCAT_VECTORS, MVT::v8i64, Custom);
1426 setOperationAction(ISD::CONCAT_VECTORS, MVT::v16f32, Custom);
1427 setOperationAction(ISD::CONCAT_VECTORS, MVT::v16i32, Custom);
1428 setOperationAction(ISD::CONCAT_VECTORS, MVT::v16i1, Legal);
1430 setOperationAction(ISD::SETCC, MVT::v16i1, Custom);
1431 setOperationAction(ISD::SETCC, MVT::v8i1, Custom);
1433 setOperationAction(ISD::MUL, MVT::v8i64, Custom);
1435 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v8i1, Custom);
1436 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v16i1, Custom);
1437 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v16i1, Custom);
1438 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i1, Custom);
1439 setOperationAction(ISD::BUILD_VECTOR, MVT::v8i1, Custom);
1440 setOperationAction(ISD::BUILD_VECTOR, MVT::v16i1, Custom);
1441 setOperationAction(ISD::SELECT, MVT::v8f64, Custom);
1442 setOperationAction(ISD::SELECT, MVT::v8i64, Custom);
1443 setOperationAction(ISD::SELECT, MVT::v16f32, Custom);
1444 setOperationAction(ISD::SELECT, MVT::v16i1, Custom);
1445 setOperationAction(ISD::SELECT, MVT::v8i1, Custom);
1447 setOperationAction(ISD::SMAX, MVT::v16i32, Legal);
1448 setOperationAction(ISD::SMAX, MVT::v8i64, Legal);
1449 setOperationAction(ISD::UMAX, MVT::v16i32, Legal);
1450 setOperationAction(ISD::UMAX, MVT::v8i64, Legal);
1451 setOperationAction(ISD::SMIN, MVT::v16i32, Legal);
1452 setOperationAction(ISD::SMIN, MVT::v8i64, Legal);
1453 setOperationAction(ISD::UMIN, MVT::v16i32, Legal);
1454 setOperationAction(ISD::UMIN, MVT::v8i64, Legal);
1456 setOperationAction(ISD::ADD, MVT::v8i64, Legal);
1457 setOperationAction(ISD::ADD, MVT::v16i32, Legal);
1459 setOperationAction(ISD::SUB, MVT::v8i64, Legal);
1460 setOperationAction(ISD::SUB, MVT::v16i32, Legal);
1462 setOperationAction(ISD::MUL, MVT::v16i32, Legal);
1464 setOperationAction(ISD::SRL, MVT::v8i64, Custom);
1465 setOperationAction(ISD::SRL, MVT::v16i32, Custom);
1467 setOperationAction(ISD::SHL, MVT::v8i64, Custom);
1468 setOperationAction(ISD::SHL, MVT::v16i32, Custom);
1470 setOperationAction(ISD::SRA, MVT::v8i64, Custom);
1471 setOperationAction(ISD::SRA, MVT::v16i32, Custom);
1473 setOperationAction(ISD::AND, MVT::v8i64, Legal);
1474 setOperationAction(ISD::OR, MVT::v8i64, Legal);
1475 setOperationAction(ISD::XOR, MVT::v8i64, Legal);
1476 setOperationAction(ISD::AND, MVT::v16i32, Legal);
1477 setOperationAction(ISD::OR, MVT::v16i32, Legal);
1478 setOperationAction(ISD::XOR, MVT::v16i32, Legal);
1480 if (Subtarget->hasCDI()) {
1481 setOperationAction(ISD::CTLZ, MVT::v8i64, Legal);
1482 setOperationAction(ISD::CTLZ, MVT::v16i32, Legal);
1484 if (Subtarget->hasDQI()) {
1485 setOperationAction(ISD::MUL, MVT::v2i64, Legal);
1486 setOperationAction(ISD::MUL, MVT::v4i64, Legal);
1487 setOperationAction(ISD::MUL, MVT::v8i64, Legal);
1489 // Custom lower several nodes.
1490 for (MVT VT : MVT::vector_valuetypes()) {
1491 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
1493 setOperationAction(ISD::AND, VT, Legal);
1494 setOperationAction(ISD::OR, VT, Legal);
1495 setOperationAction(ISD::XOR, VT, Legal);
1497 if (EltSize >= 32 && VT.getSizeInBits() <= 512) {
1498 setOperationAction(ISD::MGATHER, VT, Custom);
1499 setOperationAction(ISD::MSCATTER, VT, Custom);
1501 // Extract subvector is special because the value type
1502 // (result) is 256/128-bit but the source is 512-bit wide.
1503 if (VT.is128BitVector() || VT.is256BitVector()) {
1504 setOperationAction(ISD::EXTRACT_SUBVECTOR, VT, Custom);
1506 if (VT.getVectorElementType() == MVT::i1)
1507 setOperationAction(ISD::EXTRACT_SUBVECTOR, VT, Legal);
1509 // Do not attempt to custom lower other non-512-bit vectors
1510 if (!VT.is512BitVector())
1513 if (EltSize >= 32) {
1514 setOperationAction(ISD::VECTOR_SHUFFLE, VT, Custom);
1515 setOperationAction(ISD::INSERT_VECTOR_ELT, VT, Custom);
1516 setOperationAction(ISD::BUILD_VECTOR, VT, Custom);
1517 setOperationAction(ISD::VSELECT, VT, Legal);
1518 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Custom);
1519 setOperationAction(ISD::SCALAR_TO_VECTOR, VT, Custom);
1520 setOperationAction(ISD::INSERT_SUBVECTOR, VT, Custom);
1521 setOperationAction(ISD::MLOAD, VT, Legal);
1522 setOperationAction(ISD::MSTORE, VT, Legal);
1525 for (int i = MVT::v32i8; i != MVT::v8i64; ++i) {
1526 MVT VT = (MVT::SimpleValueType)i;
1528 // Do not attempt to promote non-512-bit vectors.
1529 if (!VT.is512BitVector())
1532 setOperationAction(ISD::SELECT, VT, Promote);
1533 AddPromotedToType (ISD::SELECT, VT, MVT::v8i64);
1537 if (!Subtarget->useSoftFloat() && Subtarget->hasBWI()) {
1538 addRegisterClass(MVT::v32i16, &X86::VR512RegClass);
1539 addRegisterClass(MVT::v64i8, &X86::VR512RegClass);
1541 addRegisterClass(MVT::v32i1, &X86::VK32RegClass);
1542 addRegisterClass(MVT::v64i1, &X86::VK64RegClass);
1544 setOperationAction(ISD::LOAD, MVT::v32i16, Legal);
1545 setOperationAction(ISD::LOAD, MVT::v64i8, Legal);
1546 setOperationAction(ISD::SETCC, MVT::v32i1, Custom);
1547 setOperationAction(ISD::SETCC, MVT::v64i1, Custom);
1548 setOperationAction(ISD::ADD, MVT::v32i16, Legal);
1549 setOperationAction(ISD::ADD, MVT::v64i8, Legal);
1550 setOperationAction(ISD::SUB, MVT::v32i16, Legal);
1551 setOperationAction(ISD::SUB, MVT::v64i8, Legal);
1552 setOperationAction(ISD::MUL, MVT::v32i16, Legal);
1553 setOperationAction(ISD::MULHS, MVT::v32i16, Legal);
1554 setOperationAction(ISD::MULHU, MVT::v32i16, Legal);
1555 setOperationAction(ISD::CONCAT_VECTORS, MVT::v32i1, Custom);
1556 setOperationAction(ISD::CONCAT_VECTORS, MVT::v64i1, Custom);
1557 setOperationAction(ISD::INSERT_SUBVECTOR, MVT::v32i1, Custom);
1558 setOperationAction(ISD::INSERT_SUBVECTOR, MVT::v64i1, Custom);
1559 setOperationAction(ISD::SELECT, MVT::v32i1, Custom);
1560 setOperationAction(ISD::SELECT, MVT::v64i1, Custom);
1561 setOperationAction(ISD::SIGN_EXTEND, MVT::v32i8, Custom);
1562 setOperationAction(ISD::ZERO_EXTEND, MVT::v32i8, Custom);
1563 setOperationAction(ISD::SIGN_EXTEND, MVT::v32i16, Custom);
1564 setOperationAction(ISD::ZERO_EXTEND, MVT::v32i16, Custom);
1565 setOperationAction(ISD::SIGN_EXTEND, MVT::v64i8, Custom);
1566 setOperationAction(ISD::ZERO_EXTEND, MVT::v64i8, Custom);
1567 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v32i1, Custom);
1568 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v64i1, Custom);
1569 setOperationAction(ISD::VSELECT, MVT::v32i16, Legal);
1570 setOperationAction(ISD::VSELECT, MVT::v64i8, Legal);
1571 setOperationAction(ISD::TRUNCATE, MVT::v32i1, Custom);
1572 setOperationAction(ISD::TRUNCATE, MVT::v64i1, Custom);
1573 setOperationAction(ISD::TRUNCATE, MVT::v32i8, Custom);
1575 setOperationAction(ISD::SMAX, MVT::v64i8, Legal);
1576 setOperationAction(ISD::SMAX, MVT::v32i16, Legal);
1577 setOperationAction(ISD::UMAX, MVT::v64i8, Legal);
1578 setOperationAction(ISD::UMAX, MVT::v32i16, Legal);
1579 setOperationAction(ISD::SMIN, MVT::v64i8, Legal);
1580 setOperationAction(ISD::SMIN, MVT::v32i16, Legal);
1581 setOperationAction(ISD::UMIN, MVT::v64i8, Legal);
1582 setOperationAction(ISD::UMIN, MVT::v32i16, Legal);
1584 setTruncStoreAction(MVT::v32i16, MVT::v32i8, Legal);
1585 setTruncStoreAction(MVT::v16i16, MVT::v16i8, Legal);
1586 if (Subtarget->hasVLX())
1587 setTruncStoreAction(MVT::v8i16, MVT::v8i8, Legal);
1589 for (int i = MVT::v32i8; i != MVT::v8i64; ++i) {
1590 const MVT VT = (MVT::SimpleValueType)i;
1592 const unsigned EltSize = VT.getVectorElementType().getSizeInBits();
1594 // Do not attempt to promote non-512-bit vectors.
1595 if (!VT.is512BitVector())
1599 setOperationAction(ISD::BUILD_VECTOR, VT, Custom);
1600 setOperationAction(ISD::VSELECT, VT, Legal);
1605 if (!Subtarget->useSoftFloat() && Subtarget->hasVLX()) {
1606 addRegisterClass(MVT::v4i1, &X86::VK4RegClass);
1607 addRegisterClass(MVT::v2i1, &X86::VK2RegClass);
1609 setOperationAction(ISD::SETCC, MVT::v4i1, Custom);
1610 setOperationAction(ISD::SETCC, MVT::v2i1, Custom);
1611 setOperationAction(ISD::CONCAT_VECTORS, MVT::v4i1, Custom);
1612 setOperationAction(ISD::CONCAT_VECTORS, MVT::v8i1, Custom);
1613 setOperationAction(ISD::INSERT_SUBVECTOR, MVT::v8i1, Custom);
1614 setOperationAction(ISD::INSERT_SUBVECTOR, MVT::v4i1, Custom);
1615 setOperationAction(ISD::SELECT, MVT::v4i1, Custom);
1616 setOperationAction(ISD::SELECT, MVT::v2i1, Custom);
1617 setOperationAction(ISD::BUILD_VECTOR, MVT::v4i1, Custom);
1618 setOperationAction(ISD::BUILD_VECTOR, MVT::v2i1, Custom);
1620 setOperationAction(ISD::AND, MVT::v8i32, Legal);
1621 setOperationAction(ISD::OR, MVT::v8i32, Legal);
1622 setOperationAction(ISD::XOR, MVT::v8i32, Legal);
1623 setOperationAction(ISD::AND, MVT::v4i32, Legal);
1624 setOperationAction(ISD::OR, MVT::v4i32, Legal);
1625 setOperationAction(ISD::XOR, MVT::v4i32, Legal);
1626 setOperationAction(ISD::SRA, MVT::v2i64, Custom);
1627 setOperationAction(ISD::SRA, MVT::v4i64, Custom);
1629 setOperationAction(ISD::SMAX, MVT::v2i64, Legal);
1630 setOperationAction(ISD::SMAX, MVT::v4i64, Legal);
1631 setOperationAction(ISD::UMAX, MVT::v2i64, Legal);
1632 setOperationAction(ISD::UMAX, MVT::v4i64, Legal);
1633 setOperationAction(ISD::SMIN, MVT::v2i64, Legal);
1634 setOperationAction(ISD::SMIN, MVT::v4i64, Legal);
1635 setOperationAction(ISD::UMIN, MVT::v2i64, Legal);
1636 setOperationAction(ISD::UMIN, MVT::v4i64, Legal);
1639 // We want to custom lower some of our intrinsics.
1640 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
1641 setOperationAction(ISD::INTRINSIC_W_CHAIN, MVT::Other, Custom);
1642 setOperationAction(ISD::INTRINSIC_VOID, MVT::Other, Custom);
1643 if (!Subtarget->is64Bit())
1644 setOperationAction(ISD::INTRINSIC_W_CHAIN, MVT::i64, Custom);
1646 // Only custom-lower 64-bit SADDO and friends on 64-bit because we don't
1647 // handle type legalization for these operations here.
1649 // FIXME: We really should do custom legalization for addition and
1650 // subtraction on x86-32 once PR3203 is fixed. We really can't do much better
1651 // than generic legalization for 64-bit multiplication-with-overflow, though.
1652 for (unsigned i = 0, e = 3+Subtarget->is64Bit(); i != e; ++i) {
1653 // Add/Sub/Mul with overflow operations are custom lowered.
1655 setOperationAction(ISD::SADDO, VT, Custom);
1656 setOperationAction(ISD::UADDO, VT, Custom);
1657 setOperationAction(ISD::SSUBO, VT, Custom);
1658 setOperationAction(ISD::USUBO, VT, Custom);
1659 setOperationAction(ISD::SMULO, VT, Custom);
1660 setOperationAction(ISD::UMULO, VT, Custom);
1664 if (!Subtarget->is64Bit()) {
1665 // These libcalls are not available in 32-bit.
1666 setLibcallName(RTLIB::SHL_I128, nullptr);
1667 setLibcallName(RTLIB::SRL_I128, nullptr);
1668 setLibcallName(RTLIB::SRA_I128, nullptr);
1671 // Combine sin / cos into one node or libcall if possible.
1672 if (Subtarget->hasSinCos()) {
1673 setLibcallName(RTLIB::SINCOS_F32, "sincosf");
1674 setLibcallName(RTLIB::SINCOS_F64, "sincos");
1675 if (Subtarget->isTargetDarwin()) {
1676 // For MacOSX, we don't want the normal expansion of a libcall to sincos.
1677 // We want to issue a libcall to __sincos_stret to avoid memory traffic.
1678 setOperationAction(ISD::FSINCOS, MVT::f64, Custom);
1679 setOperationAction(ISD::FSINCOS, MVT::f32, Custom);
1683 if (Subtarget->isTargetWin64()) {
1684 setOperationAction(ISD::SDIV, MVT::i128, Custom);
1685 setOperationAction(ISD::UDIV, MVT::i128, Custom);
1686 setOperationAction(ISD::SREM, MVT::i128, Custom);
1687 setOperationAction(ISD::UREM, MVT::i128, Custom);
1688 setOperationAction(ISD::SDIVREM, MVT::i128, Custom);
1689 setOperationAction(ISD::UDIVREM, MVT::i128, Custom);
1692 // We have target-specific dag combine patterns for the following nodes:
1693 setTargetDAGCombine(ISD::VECTOR_SHUFFLE);
1694 setTargetDAGCombine(ISD::EXTRACT_VECTOR_ELT);
1695 setTargetDAGCombine(ISD::BITCAST);
1696 setTargetDAGCombine(ISD::VSELECT);
1697 setTargetDAGCombine(ISD::SELECT);
1698 setTargetDAGCombine(ISD::SHL);
1699 setTargetDAGCombine(ISD::SRA);
1700 setTargetDAGCombine(ISD::SRL);
1701 setTargetDAGCombine(ISD::OR);
1702 setTargetDAGCombine(ISD::AND);
1703 setTargetDAGCombine(ISD::ADD);
1704 setTargetDAGCombine(ISD::FADD);
1705 setTargetDAGCombine(ISD::FSUB);
1706 setTargetDAGCombine(ISD::FMA);
1707 setTargetDAGCombine(ISD::SUB);
1708 setTargetDAGCombine(ISD::LOAD);
1709 setTargetDAGCombine(ISD::MLOAD);
1710 setTargetDAGCombine(ISD::STORE);
1711 setTargetDAGCombine(ISD::MSTORE);
1712 setTargetDAGCombine(ISD::ZERO_EXTEND);
1713 setTargetDAGCombine(ISD::ANY_EXTEND);
1714 setTargetDAGCombine(ISD::SIGN_EXTEND);
1715 setTargetDAGCombine(ISD::SIGN_EXTEND_INREG);
1716 setTargetDAGCombine(ISD::SINT_TO_FP);
1717 setTargetDAGCombine(ISD::UINT_TO_FP);
1718 setTargetDAGCombine(ISD::SETCC);
1719 setTargetDAGCombine(ISD::INTRINSIC_WO_CHAIN);
1720 setTargetDAGCombine(ISD::BUILD_VECTOR);
1721 setTargetDAGCombine(ISD::MUL);
1722 setTargetDAGCombine(ISD::XOR);
1724 computeRegisterProperties(Subtarget->getRegisterInfo());
1726 // On Darwin, -Os means optimize for size without hurting performance,
1727 // do not reduce the limit.
1728 MaxStoresPerMemset = 16; // For @llvm.memset -> sequence of stores
1729 MaxStoresPerMemsetOptSize = Subtarget->isTargetDarwin() ? 16 : 8;
1730 MaxStoresPerMemcpy = 8; // For @llvm.memcpy -> sequence of stores
1731 MaxStoresPerMemcpyOptSize = Subtarget->isTargetDarwin() ? 8 : 4;
1732 MaxStoresPerMemmove = 8; // For @llvm.memmove -> sequence of stores
1733 MaxStoresPerMemmoveOptSize = Subtarget->isTargetDarwin() ? 8 : 4;
1734 setPrefLoopAlignment(4); // 2^4 bytes.
1736 // Predictable cmov don't hurt on atom because it's in-order.
1737 PredictableSelectIsExpensive = !Subtarget->isAtom();
1738 EnableExtLdPromotion = true;
1739 setPrefFunctionAlignment(4); // 2^4 bytes.
1741 verifyIntrinsicTables();
1744 // This has so far only been implemented for 64-bit MachO.
1745 bool X86TargetLowering::useLoadStackGuardNode() const {
1746 return Subtarget->isTargetMachO() && Subtarget->is64Bit();
1749 TargetLoweringBase::LegalizeTypeAction
1750 X86TargetLowering::getPreferredVectorAction(EVT VT) const {
1751 if (ExperimentalVectorWideningLegalization &&
1752 VT.getVectorNumElements() != 1 &&
1753 VT.getVectorElementType().getSimpleVT() != MVT::i1)
1754 return TypeWidenVector;
1756 return TargetLoweringBase::getPreferredVectorAction(VT);
1759 EVT X86TargetLowering::getSetCCResultType(const DataLayout &DL, LLVMContext &,
1762 return Subtarget->hasAVX512() ? MVT::i1: MVT::i8;
1764 const unsigned NumElts = VT.getVectorNumElements();
1765 const EVT EltVT = VT.getVectorElementType();
1766 if (VT.is512BitVector()) {
1767 if (Subtarget->hasAVX512())
1768 if (EltVT == MVT::i32 || EltVT == MVT::i64 ||
1769 EltVT == MVT::f32 || EltVT == MVT::f64)
1771 case 8: return MVT::v8i1;
1772 case 16: return MVT::v16i1;
1774 if (Subtarget->hasBWI())
1775 if (EltVT == MVT::i8 || EltVT == MVT::i16)
1777 case 32: return MVT::v32i1;
1778 case 64: return MVT::v64i1;
1782 if (VT.is256BitVector() || VT.is128BitVector()) {
1783 if (Subtarget->hasVLX())
1784 if (EltVT == MVT::i32 || EltVT == MVT::i64 ||
1785 EltVT == MVT::f32 || EltVT == MVT::f64)
1787 case 2: return MVT::v2i1;
1788 case 4: return MVT::v4i1;
1789 case 8: return MVT::v8i1;
1791 if (Subtarget->hasBWI() && Subtarget->hasVLX())
1792 if (EltVT == MVT::i8 || EltVT == MVT::i16)
1794 case 8: return MVT::v8i1;
1795 case 16: return MVT::v16i1;
1796 case 32: return MVT::v32i1;
1800 return VT.changeVectorElementTypeToInteger();
1803 /// Helper for getByValTypeAlignment to determine
1804 /// the desired ByVal argument alignment.
1805 static void getMaxByValAlign(Type *Ty, unsigned &MaxAlign) {
1808 if (VectorType *VTy = dyn_cast<VectorType>(Ty)) {
1809 if (VTy->getBitWidth() == 128)
1811 } else if (ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
1812 unsigned EltAlign = 0;
1813 getMaxByValAlign(ATy->getElementType(), EltAlign);
1814 if (EltAlign > MaxAlign)
1815 MaxAlign = EltAlign;
1816 } else if (StructType *STy = dyn_cast<StructType>(Ty)) {
1817 for (auto *EltTy : STy->elements()) {
1818 unsigned EltAlign = 0;
1819 getMaxByValAlign(EltTy, EltAlign);
1820 if (EltAlign > MaxAlign)
1821 MaxAlign = EltAlign;
1828 /// Return the desired alignment for ByVal aggregate
1829 /// function arguments in the caller parameter area. For X86, aggregates
1830 /// that contain SSE vectors are placed at 16-byte boundaries while the rest
1831 /// are at 4-byte boundaries.
1832 unsigned X86TargetLowering::getByValTypeAlignment(Type *Ty,
1833 const DataLayout &DL) const {
1834 if (Subtarget->is64Bit()) {
1835 // Max of 8 and alignment of type.
1836 unsigned TyAlign = DL.getABITypeAlignment(Ty);
1843 if (Subtarget->hasSSE1())
1844 getMaxByValAlign(Ty, Align);
1848 /// Returns the target specific optimal type for load
1849 /// and store operations as a result of memset, memcpy, and memmove
1850 /// lowering. If DstAlign is zero that means it's safe to destination
1851 /// alignment can satisfy any constraint. Similarly if SrcAlign is zero it
1852 /// means there isn't a need to check it against alignment requirement,
1853 /// probably because the source does not need to be loaded. If 'IsMemset' is
1854 /// true, that means it's expanding a memset. If 'ZeroMemset' is true, that
1855 /// means it's a memset of zero. 'MemcpyStrSrc' indicates whether the memcpy
1856 /// source is constant so it does not need to be loaded.
1857 /// It returns EVT::Other if the type should be determined using generic
1858 /// target-independent logic.
1860 X86TargetLowering::getOptimalMemOpType(uint64_t Size,
1861 unsigned DstAlign, unsigned SrcAlign,
1862 bool IsMemset, bool ZeroMemset,
1864 MachineFunction &MF) const {
1865 const Function *F = MF.getFunction();
1866 if ((!IsMemset || ZeroMemset) &&
1867 !F->hasFnAttribute(Attribute::NoImplicitFloat)) {
1869 (Subtarget->isUnalignedMemAccessFast() ||
1870 ((DstAlign == 0 || DstAlign >= 16) &&
1871 (SrcAlign == 0 || SrcAlign >= 16)))) {
1873 if (Subtarget->hasInt256())
1875 if (Subtarget->hasFp256())
1878 if (Subtarget->hasSSE2())
1880 if (Subtarget->hasSSE1())
1882 } else if (!MemcpyStrSrc && Size >= 8 &&
1883 !Subtarget->is64Bit() &&
1884 Subtarget->hasSSE2()) {
1885 // Do not use f64 to lower memcpy if source is string constant. It's
1886 // better to use i32 to avoid the loads.
1890 if (Subtarget->is64Bit() && Size >= 8)
1895 bool X86TargetLowering::isSafeMemOpType(MVT VT) const {
1897 return X86ScalarSSEf32;
1898 else if (VT == MVT::f64)
1899 return X86ScalarSSEf64;
1904 X86TargetLowering::allowsMisalignedMemoryAccesses(EVT VT,
1909 *Fast = Subtarget->isUnalignedMemAccessFast();
1913 /// Return the entry encoding for a jump table in the
1914 /// current function. The returned value is a member of the
1915 /// MachineJumpTableInfo::JTEntryKind enum.
1916 unsigned X86TargetLowering::getJumpTableEncoding() const {
1917 // In GOT pic mode, each entry in the jump table is emitted as a @GOTOFF
1919 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
1920 Subtarget->isPICStyleGOT())
1921 return MachineJumpTableInfo::EK_Custom32;
1923 // Otherwise, use the normal jump table encoding heuristics.
1924 return TargetLowering::getJumpTableEncoding();
1927 bool X86TargetLowering::useSoftFloat() const {
1928 return Subtarget->useSoftFloat();
1932 X86TargetLowering::LowerCustomJumpTableEntry(const MachineJumpTableInfo *MJTI,
1933 const MachineBasicBlock *MBB,
1934 unsigned uid,MCContext &Ctx) const{
1935 assert(MBB->getParent()->getTarget().getRelocationModel() == Reloc::PIC_ &&
1936 Subtarget->isPICStyleGOT());
1937 // In 32-bit ELF systems, our jump table entries are formed with @GOTOFF
1939 return MCSymbolRefExpr::create(MBB->getSymbol(),
1940 MCSymbolRefExpr::VK_GOTOFF, Ctx);
1943 /// Returns relocation base for the given PIC jumptable.
1944 SDValue X86TargetLowering::getPICJumpTableRelocBase(SDValue Table,
1945 SelectionDAG &DAG) const {
1946 if (!Subtarget->is64Bit())
1947 // This doesn't have SDLoc associated with it, but is not really the
1948 // same as a Register.
1949 return DAG.getNode(X86ISD::GlobalBaseReg, SDLoc(),
1950 getPointerTy(DAG.getDataLayout()));
1954 /// This returns the relocation base for the given PIC jumptable,
1955 /// the same as getPICJumpTableRelocBase, but as an MCExpr.
1956 const MCExpr *X86TargetLowering::
1957 getPICJumpTableRelocBaseExpr(const MachineFunction *MF, unsigned JTI,
1958 MCContext &Ctx) const {
1959 // X86-64 uses RIP relative addressing based on the jump table label.
1960 if (Subtarget->isPICStyleRIPRel())
1961 return TargetLowering::getPICJumpTableRelocBaseExpr(MF, JTI, Ctx);
1963 // Otherwise, the reference is relative to the PIC base.
1964 return MCSymbolRefExpr::create(MF->getPICBaseSymbol(), Ctx);
1967 std::pair<const TargetRegisterClass *, uint8_t>
1968 X86TargetLowering::findRepresentativeClass(const TargetRegisterInfo *TRI,
1970 const TargetRegisterClass *RRC = nullptr;
1972 switch (VT.SimpleTy) {
1974 return TargetLowering::findRepresentativeClass(TRI, VT);
1975 case MVT::i8: case MVT::i16: case MVT::i32: case MVT::i64:
1976 RRC = Subtarget->is64Bit() ? &X86::GR64RegClass : &X86::GR32RegClass;
1979 RRC = &X86::VR64RegClass;
1981 case MVT::f32: case MVT::f64:
1982 case MVT::v16i8: case MVT::v8i16: case MVT::v4i32: case MVT::v2i64:
1983 case MVT::v4f32: case MVT::v2f64:
1984 case MVT::v32i8: case MVT::v8i32: case MVT::v4i64: case MVT::v8f32:
1986 RRC = &X86::VR128RegClass;
1989 return std::make_pair(RRC, Cost);
1992 bool X86TargetLowering::getStackCookieLocation(unsigned &AddressSpace,
1993 unsigned &Offset) const {
1994 if (!Subtarget->isTargetLinux())
1997 if (Subtarget->is64Bit()) {
1998 // %fs:0x28, unless we're using a Kernel code model, in which case it's %gs:
2000 if (getTargetMachine().getCodeModel() == CodeModel::Kernel)
2012 bool X86TargetLowering::isNoopAddrSpaceCast(unsigned SrcAS,
2013 unsigned DestAS) const {
2014 assert(SrcAS != DestAS && "Expected different address spaces!");
2016 return SrcAS < 256 && DestAS < 256;
2019 //===----------------------------------------------------------------------===//
2020 // Return Value Calling Convention Implementation
2021 //===----------------------------------------------------------------------===//
2023 #include "X86GenCallingConv.inc"
2026 X86TargetLowering::CanLowerReturn(CallingConv::ID CallConv,
2027 MachineFunction &MF, bool isVarArg,
2028 const SmallVectorImpl<ISD::OutputArg> &Outs,
2029 LLVMContext &Context) const {
2030 SmallVector<CCValAssign, 16> RVLocs;
2031 CCState CCInfo(CallConv, isVarArg, MF, RVLocs, Context);
2032 return CCInfo.CheckReturn(Outs, RetCC_X86);
2035 const MCPhysReg *X86TargetLowering::getScratchRegisters(CallingConv::ID) const {
2036 static const MCPhysReg ScratchRegs[] = { X86::R11, 0 };
2041 X86TargetLowering::LowerReturn(SDValue Chain,
2042 CallingConv::ID CallConv, bool isVarArg,
2043 const SmallVectorImpl<ISD::OutputArg> &Outs,
2044 const SmallVectorImpl<SDValue> &OutVals,
2045 SDLoc dl, SelectionDAG &DAG) const {
2046 MachineFunction &MF = DAG.getMachineFunction();
2047 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
2049 SmallVector<CCValAssign, 16> RVLocs;
2050 CCState CCInfo(CallConv, isVarArg, MF, RVLocs, *DAG.getContext());
2051 CCInfo.AnalyzeReturn(Outs, RetCC_X86);
2054 SmallVector<SDValue, 6> RetOps;
2055 RetOps.push_back(Chain); // Operand #0 = Chain (updated below)
2056 // Operand #1 = Bytes To Pop
2057 RetOps.push_back(DAG.getTargetConstant(FuncInfo->getBytesToPopOnReturn(), dl,
2060 // Copy the result values into the output registers.
2061 for (unsigned i = 0; i != RVLocs.size(); ++i) {
2062 CCValAssign &VA = RVLocs[i];
2063 assert(VA.isRegLoc() && "Can only return in registers!");
2064 SDValue ValToCopy = OutVals[i];
2065 EVT ValVT = ValToCopy.getValueType();
2067 // Promote values to the appropriate types.
2068 if (VA.getLocInfo() == CCValAssign::SExt)
2069 ValToCopy = DAG.getNode(ISD::SIGN_EXTEND, dl, VA.getLocVT(), ValToCopy);
2070 else if (VA.getLocInfo() == CCValAssign::ZExt)
2071 ValToCopy = DAG.getNode(ISD::ZERO_EXTEND, dl, VA.getLocVT(), ValToCopy);
2072 else if (VA.getLocInfo() == CCValAssign::AExt) {
2073 if (ValVT.isVector() && ValVT.getScalarType() == MVT::i1)
2074 ValToCopy = DAG.getNode(ISD::SIGN_EXTEND, dl, VA.getLocVT(), ValToCopy);
2076 ValToCopy = DAG.getNode(ISD::ANY_EXTEND, dl, VA.getLocVT(), ValToCopy);
2078 else if (VA.getLocInfo() == CCValAssign::BCvt)
2079 ValToCopy = DAG.getBitcast(VA.getLocVT(), ValToCopy);
2081 assert(VA.getLocInfo() != CCValAssign::FPExt &&
2082 "Unexpected FP-extend for return value.");
2084 // If this is x86-64, and we disabled SSE, we can't return FP values,
2085 // or SSE or MMX vectors.
2086 if ((ValVT == MVT::f32 || ValVT == MVT::f64 ||
2087 VA.getLocReg() == X86::XMM0 || VA.getLocReg() == X86::XMM1) &&
2088 (Subtarget->is64Bit() && !Subtarget->hasSSE1())) {
2089 report_fatal_error("SSE register return with SSE disabled");
2091 // Likewise we can't return F64 values with SSE1 only. gcc does so, but
2092 // llvm-gcc has never done it right and no one has noticed, so this
2093 // should be OK for now.
2094 if (ValVT == MVT::f64 &&
2095 (Subtarget->is64Bit() && !Subtarget->hasSSE2()))
2096 report_fatal_error("SSE2 register return with SSE2 disabled");
2098 // Returns in ST0/ST1 are handled specially: these are pushed as operands to
2099 // the RET instruction and handled by the FP Stackifier.
2100 if (VA.getLocReg() == X86::FP0 ||
2101 VA.getLocReg() == X86::FP1) {
2102 // If this is a copy from an xmm register to ST(0), use an FPExtend to
2103 // change the value to the FP stack register class.
2104 if (isScalarFPTypeInSSEReg(VA.getValVT()))
2105 ValToCopy = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f80, ValToCopy);
2106 RetOps.push_back(ValToCopy);
2107 // Don't emit a copytoreg.
2111 // 64-bit vector (MMX) values are returned in XMM0 / XMM1 except for v1i64
2112 // which is returned in RAX / RDX.
2113 if (Subtarget->is64Bit()) {
2114 if (ValVT == MVT::x86mmx) {
2115 if (VA.getLocReg() == X86::XMM0 || VA.getLocReg() == X86::XMM1) {
2116 ValToCopy = DAG.getBitcast(MVT::i64, ValToCopy);
2117 ValToCopy = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i64,
2119 // If we don't have SSE2 available, convert to v4f32 so the generated
2120 // register is legal.
2121 if (!Subtarget->hasSSE2())
2122 ValToCopy = DAG.getBitcast(MVT::v4f32, ValToCopy);
2127 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), ValToCopy, Flag);
2128 Flag = Chain.getValue(1);
2129 RetOps.push_back(DAG.getRegister(VA.getLocReg(), VA.getLocVT()));
2132 // All x86 ABIs require that for returning structs by value we copy
2133 // the sret argument into %rax/%eax (depending on ABI) for the return.
2134 // We saved the argument into a virtual register in the entry block,
2135 // so now we copy the value out and into %rax/%eax.
2137 // Checking Function.hasStructRetAttr() here is insufficient because the IR
2138 // may not have an explicit sret argument. If FuncInfo.CanLowerReturn is
2139 // false, then an sret argument may be implicitly inserted in the SelDAG. In
2140 // either case FuncInfo->setSRetReturnReg() will have been called.
2141 if (unsigned SRetReg = FuncInfo->getSRetReturnReg()) {
2142 SDValue Val = DAG.getCopyFromReg(Chain, dl, SRetReg,
2143 getPointerTy(MF.getDataLayout()));
2146 = (Subtarget->is64Bit() && !Subtarget->isTarget64BitILP32()) ?
2147 X86::RAX : X86::EAX;
2148 Chain = DAG.getCopyToReg(Chain, dl, RetValReg, Val, Flag);
2149 Flag = Chain.getValue(1);
2151 // RAX/EAX now acts like a return value.
2153 DAG.getRegister(RetValReg, getPointerTy(DAG.getDataLayout())));
2156 RetOps[0] = Chain; // Update chain.
2158 // Add the flag if we have it.
2160 RetOps.push_back(Flag);
2162 return DAG.getNode(X86ISD::RET_FLAG, dl, MVT::Other, RetOps);
2165 bool X86TargetLowering::isUsedByReturnOnly(SDNode *N, SDValue &Chain) const {
2166 if (N->getNumValues() != 1)
2168 if (!N->hasNUsesOfValue(1, 0))
2171 SDValue TCChain = Chain;
2172 SDNode *Copy = *N->use_begin();
2173 if (Copy->getOpcode() == ISD::CopyToReg) {
2174 // If the copy has a glue operand, we conservatively assume it isn't safe to
2175 // perform a tail call.
2176 if (Copy->getOperand(Copy->getNumOperands()-1).getValueType() == MVT::Glue)
2178 TCChain = Copy->getOperand(0);
2179 } else if (Copy->getOpcode() != ISD::FP_EXTEND)
2182 bool HasRet = false;
2183 for (SDNode::use_iterator UI = Copy->use_begin(), UE = Copy->use_end();
2185 if (UI->getOpcode() != X86ISD::RET_FLAG)
2187 // If we are returning more than one value, we can definitely
2188 // not make a tail call see PR19530
2189 if (UI->getNumOperands() > 4)
2191 if (UI->getNumOperands() == 4 &&
2192 UI->getOperand(UI->getNumOperands()-1).getValueType() != MVT::Glue)
2205 X86TargetLowering::getTypeForExtArgOrReturn(LLVMContext &Context, EVT VT,
2206 ISD::NodeType ExtendKind) const {
2208 // TODO: Is this also valid on 32-bit?
2209 if (Subtarget->is64Bit() && VT == MVT::i1 && ExtendKind == ISD::ZERO_EXTEND)
2210 ReturnMVT = MVT::i8;
2212 ReturnMVT = MVT::i32;
2214 EVT MinVT = getRegisterType(Context, ReturnMVT);
2215 return VT.bitsLT(MinVT) ? MinVT : VT;
2218 /// Lower the result values of a call into the
2219 /// appropriate copies out of appropriate physical registers.
2222 X86TargetLowering::LowerCallResult(SDValue Chain, SDValue InFlag,
2223 CallingConv::ID CallConv, bool isVarArg,
2224 const SmallVectorImpl<ISD::InputArg> &Ins,
2225 SDLoc dl, SelectionDAG &DAG,
2226 SmallVectorImpl<SDValue> &InVals) const {
2228 // Assign locations to each value returned by this call.
2229 SmallVector<CCValAssign, 16> RVLocs;
2230 bool Is64Bit = Subtarget->is64Bit();
2231 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(), RVLocs,
2233 CCInfo.AnalyzeCallResult(Ins, RetCC_X86);
2235 // Copy all of the result registers out of their specified physreg.
2236 for (unsigned i = 0, e = RVLocs.size(); i != e; ++i) {
2237 CCValAssign &VA = RVLocs[i];
2238 EVT CopyVT = VA.getLocVT();
2240 // If this is x86-64, and we disabled SSE, we can't return FP values
2241 if ((CopyVT == MVT::f32 || CopyVT == MVT::f64) &&
2242 ((Is64Bit || Ins[i].Flags.isInReg()) && !Subtarget->hasSSE1())) {
2243 report_fatal_error("SSE register return with SSE disabled");
2246 // If we prefer to use the value in xmm registers, copy it out as f80 and
2247 // use a truncate to move it from fp stack reg to xmm reg.
2248 bool RoundAfterCopy = false;
2249 if ((VA.getLocReg() == X86::FP0 || VA.getLocReg() == X86::FP1) &&
2250 isScalarFPTypeInSSEReg(VA.getValVT())) {
2252 RoundAfterCopy = (CopyVT != VA.getLocVT());
2255 Chain = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(),
2256 CopyVT, InFlag).getValue(1);
2257 SDValue Val = Chain.getValue(0);
2260 Val = DAG.getNode(ISD::FP_ROUND, dl, VA.getValVT(), Val,
2261 // This truncation won't change the value.
2262 DAG.getIntPtrConstant(1, dl));
2264 if (VA.isExtInLoc() && VA.getValVT().getScalarType() == MVT::i1)
2265 Val = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), Val);
2267 InFlag = Chain.getValue(2);
2268 InVals.push_back(Val);
2274 //===----------------------------------------------------------------------===//
2275 // C & StdCall & Fast Calling Convention implementation
2276 //===----------------------------------------------------------------------===//
2277 // StdCall calling convention seems to be standard for many Windows' API
2278 // routines and around. It differs from C calling convention just a little:
2279 // callee should clean up the stack, not caller. Symbols should be also
2280 // decorated in some fancy way :) It doesn't support any vector arguments.
2281 // For info on fast calling convention see Fast Calling Convention (tail call)
2282 // implementation LowerX86_32FastCCCallTo.
2284 /// CallIsStructReturn - Determines whether a call uses struct return
2286 enum StructReturnType {
2291 static StructReturnType
2292 callIsStructReturn(const SmallVectorImpl<ISD::OutputArg> &Outs) {
2294 return NotStructReturn;
2296 const ISD::ArgFlagsTy &Flags = Outs[0].Flags;
2297 if (!Flags.isSRet())
2298 return NotStructReturn;
2299 if (Flags.isInReg())
2300 return RegStructReturn;
2301 return StackStructReturn;
2304 /// Determines whether a function uses struct return semantics.
2305 static StructReturnType
2306 argsAreStructReturn(const SmallVectorImpl<ISD::InputArg> &Ins) {
2308 return NotStructReturn;
2310 const ISD::ArgFlagsTy &Flags = Ins[0].Flags;
2311 if (!Flags.isSRet())
2312 return NotStructReturn;
2313 if (Flags.isInReg())
2314 return RegStructReturn;
2315 return StackStructReturn;
2318 /// Make a copy of an aggregate at address specified by "Src" to address
2319 /// "Dst" with size and alignment information specified by the specific
2320 /// parameter attribute. The copy will be passed as a byval function parameter.
2322 CreateCopyOfByValArgument(SDValue Src, SDValue Dst, SDValue Chain,
2323 ISD::ArgFlagsTy Flags, SelectionDAG &DAG,
2325 SDValue SizeNode = DAG.getConstant(Flags.getByValSize(), dl, MVT::i32);
2327 return DAG.getMemcpy(Chain, dl, Dst, Src, SizeNode, Flags.getByValAlign(),
2328 /*isVolatile*/false, /*AlwaysInline=*/true,
2329 /*isTailCall*/false,
2330 MachinePointerInfo(), MachinePointerInfo());
2333 /// Return true if the calling convention is one that
2334 /// supports tail call optimization.
2335 static bool IsTailCallConvention(CallingConv::ID CC) {
2336 return (CC == CallingConv::Fast || CC == CallingConv::GHC ||
2337 CC == CallingConv::HiPE);
2340 /// \brief Return true if the calling convention is a C calling convention.
2341 static bool IsCCallConvention(CallingConv::ID CC) {
2342 return (CC == CallingConv::C || CC == CallingConv::X86_64_Win64 ||
2343 CC == CallingConv::X86_64_SysV);
2346 bool X86TargetLowering::mayBeEmittedAsTailCall(CallInst *CI) const {
2348 CI->getParent()->getParent()->getFnAttribute("disable-tail-calls");
2349 if (!CI->isTailCall() || Attr.getValueAsString() == "true")
2353 CallingConv::ID CalleeCC = CS.getCallingConv();
2354 if (!IsTailCallConvention(CalleeCC) && !IsCCallConvention(CalleeCC))
2360 /// Return true if the function is being made into
2361 /// a tailcall target by changing its ABI.
2362 static bool FuncIsMadeTailCallSafe(CallingConv::ID CC,
2363 bool GuaranteedTailCallOpt) {
2364 return GuaranteedTailCallOpt && IsTailCallConvention(CC);
2368 X86TargetLowering::LowerMemArgument(SDValue Chain,
2369 CallingConv::ID CallConv,
2370 const SmallVectorImpl<ISD::InputArg> &Ins,
2371 SDLoc dl, SelectionDAG &DAG,
2372 const CCValAssign &VA,
2373 MachineFrameInfo *MFI,
2375 // Create the nodes corresponding to a load from this parameter slot.
2376 ISD::ArgFlagsTy Flags = Ins[i].Flags;
2377 bool AlwaysUseMutable = FuncIsMadeTailCallSafe(
2378 CallConv, DAG.getTarget().Options.GuaranteedTailCallOpt);
2379 bool isImmutable = !AlwaysUseMutable && !Flags.isByVal();
2382 // If value is passed by pointer we have address passed instead of the value
2384 bool ExtendedInMem = VA.isExtInLoc() &&
2385 VA.getValVT().getScalarType() == MVT::i1;
2387 if (VA.getLocInfo() == CCValAssign::Indirect || ExtendedInMem)
2388 ValVT = VA.getLocVT();
2390 ValVT = VA.getValVT();
2392 // FIXME: For now, all byval parameter objects are marked mutable. This can be
2393 // changed with more analysis.
2394 // In case of tail call optimization mark all arguments mutable. Since they
2395 // could be overwritten by lowering of arguments in case of a tail call.
2396 if (Flags.isByVal()) {
2397 unsigned Bytes = Flags.getByValSize();
2398 if (Bytes == 0) Bytes = 1; // Don't create zero-sized stack objects.
2399 int FI = MFI->CreateFixedObject(Bytes, VA.getLocMemOffset(), isImmutable);
2400 return DAG.getFrameIndex(FI, getPointerTy(DAG.getDataLayout()));
2402 int FI = MFI->CreateFixedObject(ValVT.getSizeInBits()/8,
2403 VA.getLocMemOffset(), isImmutable);
2404 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy(DAG.getDataLayout()));
2405 SDValue Val = DAG.getLoad(ValVT, dl, Chain, FIN,
2406 MachinePointerInfo::getFixedStack(FI),
2407 false, false, false, 0);
2408 return ExtendedInMem ?
2409 DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), Val) : Val;
2413 // FIXME: Get this from tablegen.
2414 static ArrayRef<MCPhysReg> get64BitArgumentGPRs(CallingConv::ID CallConv,
2415 const X86Subtarget *Subtarget) {
2416 assert(Subtarget->is64Bit());
2418 if (Subtarget->isCallingConvWin64(CallConv)) {
2419 static const MCPhysReg GPR64ArgRegsWin64[] = {
2420 X86::RCX, X86::RDX, X86::R8, X86::R9
2422 return makeArrayRef(std::begin(GPR64ArgRegsWin64), std::end(GPR64ArgRegsWin64));
2425 static const MCPhysReg GPR64ArgRegs64Bit[] = {
2426 X86::RDI, X86::RSI, X86::RDX, X86::RCX, X86::R8, X86::R9
2428 return makeArrayRef(std::begin(GPR64ArgRegs64Bit), std::end(GPR64ArgRegs64Bit));
2431 // FIXME: Get this from tablegen.
2432 static ArrayRef<MCPhysReg> get64BitArgumentXMMs(MachineFunction &MF,
2433 CallingConv::ID CallConv,
2434 const X86Subtarget *Subtarget) {
2435 assert(Subtarget->is64Bit());
2436 if (Subtarget->isCallingConvWin64(CallConv)) {
2437 // The XMM registers which might contain var arg parameters are shadowed
2438 // in their paired GPR. So we only need to save the GPR to their home
2440 // TODO: __vectorcall will change this.
2444 const Function *Fn = MF.getFunction();
2445 bool NoImplicitFloatOps = Fn->hasFnAttribute(Attribute::NoImplicitFloat);
2446 bool isSoftFloat = Subtarget->useSoftFloat();
2447 assert(!(isSoftFloat && NoImplicitFloatOps) &&
2448 "SSE register cannot be used when SSE is disabled!");
2449 if (isSoftFloat || NoImplicitFloatOps || !Subtarget->hasSSE1())
2450 // Kernel mode asks for SSE to be disabled, so there are no XMM argument
2454 static const MCPhysReg XMMArgRegs64Bit[] = {
2455 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
2456 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
2458 return makeArrayRef(std::begin(XMMArgRegs64Bit), std::end(XMMArgRegs64Bit));
2462 X86TargetLowering::LowerFormalArguments(SDValue Chain,
2463 CallingConv::ID CallConv,
2465 const SmallVectorImpl<ISD::InputArg> &Ins,
2468 SmallVectorImpl<SDValue> &InVals)
2470 MachineFunction &MF = DAG.getMachineFunction();
2471 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
2472 const TargetFrameLowering &TFI = *Subtarget->getFrameLowering();
2474 const Function* Fn = MF.getFunction();
2475 if (Fn->hasExternalLinkage() &&
2476 Subtarget->isTargetCygMing() &&
2477 Fn->getName() == "main")
2478 FuncInfo->setForceFramePointer(true);
2480 MachineFrameInfo *MFI = MF.getFrameInfo();
2481 bool Is64Bit = Subtarget->is64Bit();
2482 bool IsWin64 = Subtarget->isCallingConvWin64(CallConv);
2484 assert(!(isVarArg && IsTailCallConvention(CallConv)) &&
2485 "Var args not supported with calling convention fastcc, ghc or hipe");
2487 // Assign locations to all of the incoming arguments.
2488 SmallVector<CCValAssign, 16> ArgLocs;
2489 CCState CCInfo(CallConv, isVarArg, MF, ArgLocs, *DAG.getContext());
2491 // Allocate shadow area for Win64
2493 CCInfo.AllocateStack(32, 8);
2495 CCInfo.AnalyzeFormalArguments(Ins, CC_X86);
2497 unsigned LastVal = ~0U;
2499 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2500 CCValAssign &VA = ArgLocs[i];
2501 // TODO: If an arg is passed in two places (e.g. reg and stack), skip later
2503 assert(VA.getValNo() != LastVal &&
2504 "Don't support value assigned to multiple locs yet");
2506 LastVal = VA.getValNo();
2508 if (VA.isRegLoc()) {
2509 EVT RegVT = VA.getLocVT();
2510 const TargetRegisterClass *RC;
2511 if (RegVT == MVT::i32)
2512 RC = &X86::GR32RegClass;
2513 else if (Is64Bit && RegVT == MVT::i64)
2514 RC = &X86::GR64RegClass;
2515 else if (RegVT == MVT::f32)
2516 RC = &X86::FR32RegClass;
2517 else if (RegVT == MVT::f64)
2518 RC = &X86::FR64RegClass;
2519 else if (RegVT.is512BitVector())
2520 RC = &X86::VR512RegClass;
2521 else if (RegVT.is256BitVector())
2522 RC = &X86::VR256RegClass;
2523 else if (RegVT.is128BitVector())
2524 RC = &X86::VR128RegClass;
2525 else if (RegVT == MVT::x86mmx)
2526 RC = &X86::VR64RegClass;
2527 else if (RegVT == MVT::i1)
2528 RC = &X86::VK1RegClass;
2529 else if (RegVT == MVT::v8i1)
2530 RC = &X86::VK8RegClass;
2531 else if (RegVT == MVT::v16i1)
2532 RC = &X86::VK16RegClass;
2533 else if (RegVT == MVT::v32i1)
2534 RC = &X86::VK32RegClass;
2535 else if (RegVT == MVT::v64i1)
2536 RC = &X86::VK64RegClass;
2538 llvm_unreachable("Unknown argument type!");
2540 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
2541 ArgValue = DAG.getCopyFromReg(Chain, dl, Reg, RegVT);
2543 // If this is an 8 or 16-bit value, it is really passed promoted to 32
2544 // bits. Insert an assert[sz]ext to capture this, then truncate to the
2546 if (VA.getLocInfo() == CCValAssign::SExt)
2547 ArgValue = DAG.getNode(ISD::AssertSext, dl, RegVT, ArgValue,
2548 DAG.getValueType(VA.getValVT()));
2549 else if (VA.getLocInfo() == CCValAssign::ZExt)
2550 ArgValue = DAG.getNode(ISD::AssertZext, dl, RegVT, ArgValue,
2551 DAG.getValueType(VA.getValVT()));
2552 else if (VA.getLocInfo() == CCValAssign::BCvt)
2553 ArgValue = DAG.getBitcast(VA.getValVT(), ArgValue);
2555 if (VA.isExtInLoc()) {
2556 // Handle MMX values passed in XMM regs.
2557 if (RegVT.isVector() && VA.getValVT().getScalarType() != MVT::i1)
2558 ArgValue = DAG.getNode(X86ISD::MOVDQ2Q, dl, VA.getValVT(), ArgValue);
2560 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
2563 assert(VA.isMemLoc());
2564 ArgValue = LowerMemArgument(Chain, CallConv, Ins, dl, DAG, VA, MFI, i);
2567 // If value is passed via pointer - do a load.
2568 if (VA.getLocInfo() == CCValAssign::Indirect)
2569 ArgValue = DAG.getLoad(VA.getValVT(), dl, Chain, ArgValue,
2570 MachinePointerInfo(), false, false, false, 0);
2572 InVals.push_back(ArgValue);
2575 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2576 // All x86 ABIs require that for returning structs by value we copy the
2577 // sret argument into %rax/%eax (depending on ABI) for the return. Save
2578 // the argument into a virtual register so that we can access it from the
2580 if (Ins[i].Flags.isSRet()) {
2581 unsigned Reg = FuncInfo->getSRetReturnReg();
2583 MVT PtrTy = getPointerTy(DAG.getDataLayout());
2584 Reg = MF.getRegInfo().createVirtualRegister(getRegClassFor(PtrTy));
2585 FuncInfo->setSRetReturnReg(Reg);
2587 SDValue Copy = DAG.getCopyToReg(DAG.getEntryNode(), dl, Reg, InVals[i]);
2588 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Copy, Chain);
2593 unsigned StackSize = CCInfo.getNextStackOffset();
2594 // Align stack specially for tail calls.
2595 if (FuncIsMadeTailCallSafe(CallConv,
2596 MF.getTarget().Options.GuaranteedTailCallOpt))
2597 StackSize = GetAlignedArgumentStackSize(StackSize, DAG);
2599 // If the function takes variable number of arguments, make a frame index for
2600 // the start of the first vararg value... for expansion of llvm.va_start. We
2601 // can skip this if there are no va_start calls.
2602 if (MFI->hasVAStart() &&
2603 (Is64Bit || (CallConv != CallingConv::X86_FastCall &&
2604 CallConv != CallingConv::X86_ThisCall))) {
2605 FuncInfo->setVarArgsFrameIndex(
2606 MFI->CreateFixedObject(1, StackSize, true));
2609 MachineModuleInfo &MMI = MF.getMMI();
2610 const Function *WinEHParent = nullptr;
2611 if (MMI.hasWinEHFuncInfo(Fn))
2612 WinEHParent = MMI.getWinEHParent(Fn);
2613 bool IsWinEHOutlined = WinEHParent && WinEHParent != Fn;
2614 bool IsWinEHParent = WinEHParent && WinEHParent == Fn;
2616 // Figure out if XMM registers are in use.
2617 assert(!(Subtarget->useSoftFloat() &&
2618 Fn->hasFnAttribute(Attribute::NoImplicitFloat)) &&
2619 "SSE register cannot be used when SSE is disabled!");
2621 // 64-bit calling conventions support varargs and register parameters, so we
2622 // have to do extra work to spill them in the prologue.
2623 if (Is64Bit && isVarArg && MFI->hasVAStart()) {
2624 // Find the first unallocated argument registers.
2625 ArrayRef<MCPhysReg> ArgGPRs = get64BitArgumentGPRs(CallConv, Subtarget);
2626 ArrayRef<MCPhysReg> ArgXMMs = get64BitArgumentXMMs(MF, CallConv, Subtarget);
2627 unsigned NumIntRegs = CCInfo.getFirstUnallocated(ArgGPRs);
2628 unsigned NumXMMRegs = CCInfo.getFirstUnallocated(ArgXMMs);
2629 assert(!(NumXMMRegs && !Subtarget->hasSSE1()) &&
2630 "SSE register cannot be used when SSE is disabled!");
2632 // Gather all the live in physical registers.
2633 SmallVector<SDValue, 6> LiveGPRs;
2634 SmallVector<SDValue, 8> LiveXMMRegs;
2636 for (MCPhysReg Reg : ArgGPRs.slice(NumIntRegs)) {
2637 unsigned GPR = MF.addLiveIn(Reg, &X86::GR64RegClass);
2639 DAG.getCopyFromReg(Chain, dl, GPR, MVT::i64));
2641 if (!ArgXMMs.empty()) {
2642 unsigned AL = MF.addLiveIn(X86::AL, &X86::GR8RegClass);
2643 ALVal = DAG.getCopyFromReg(Chain, dl, AL, MVT::i8);
2644 for (MCPhysReg Reg : ArgXMMs.slice(NumXMMRegs)) {
2645 unsigned XMMReg = MF.addLiveIn(Reg, &X86::VR128RegClass);
2646 LiveXMMRegs.push_back(
2647 DAG.getCopyFromReg(Chain, dl, XMMReg, MVT::v4f32));
2652 // Get to the caller-allocated home save location. Add 8 to account
2653 // for the return address.
2654 int HomeOffset = TFI.getOffsetOfLocalArea() + 8;
2655 FuncInfo->setRegSaveFrameIndex(
2656 MFI->CreateFixedObject(1, NumIntRegs * 8 + HomeOffset, false));
2657 // Fixup to set vararg frame on shadow area (4 x i64).
2659 FuncInfo->setVarArgsFrameIndex(FuncInfo->getRegSaveFrameIndex());
2661 // For X86-64, if there are vararg parameters that are passed via
2662 // registers, then we must store them to their spots on the stack so
2663 // they may be loaded by deferencing the result of va_next.
2664 FuncInfo->setVarArgsGPOffset(NumIntRegs * 8);
2665 FuncInfo->setVarArgsFPOffset(ArgGPRs.size() * 8 + NumXMMRegs * 16);
2666 FuncInfo->setRegSaveFrameIndex(MFI->CreateStackObject(
2667 ArgGPRs.size() * 8 + ArgXMMs.size() * 16, 16, false));
2670 // Store the integer parameter registers.
2671 SmallVector<SDValue, 8> MemOps;
2672 SDValue RSFIN = DAG.getFrameIndex(FuncInfo->getRegSaveFrameIndex(),
2673 getPointerTy(DAG.getDataLayout()));
2674 unsigned Offset = FuncInfo->getVarArgsGPOffset();
2675 for (SDValue Val : LiveGPRs) {
2676 SDValue FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(DAG.getDataLayout()),
2677 RSFIN, DAG.getIntPtrConstant(Offset, dl));
2679 DAG.getStore(Val.getValue(1), dl, Val, FIN,
2680 MachinePointerInfo::getFixedStack(
2681 FuncInfo->getRegSaveFrameIndex(), Offset),
2683 MemOps.push_back(Store);
2687 if (!ArgXMMs.empty() && NumXMMRegs != ArgXMMs.size()) {
2688 // Now store the XMM (fp + vector) parameter registers.
2689 SmallVector<SDValue, 12> SaveXMMOps;
2690 SaveXMMOps.push_back(Chain);
2691 SaveXMMOps.push_back(ALVal);
2692 SaveXMMOps.push_back(DAG.getIntPtrConstant(
2693 FuncInfo->getRegSaveFrameIndex(), dl));
2694 SaveXMMOps.push_back(DAG.getIntPtrConstant(
2695 FuncInfo->getVarArgsFPOffset(), dl));
2696 SaveXMMOps.insert(SaveXMMOps.end(), LiveXMMRegs.begin(),
2698 MemOps.push_back(DAG.getNode(X86ISD::VASTART_SAVE_XMM_REGS, dl,
2699 MVT::Other, SaveXMMOps));
2702 if (!MemOps.empty())
2703 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, MemOps);
2704 } else if (IsWin64 && IsWinEHOutlined) {
2705 // Get to the caller-allocated home save location. Add 8 to account
2706 // for the return address.
2707 int HomeOffset = TFI.getOffsetOfLocalArea() + 8;
2708 FuncInfo->setRegSaveFrameIndex(MFI->CreateFixedObject(
2709 /*Size=*/1, /*SPOffset=*/HomeOffset + 8, /*Immutable=*/false));
2711 MMI.getWinEHFuncInfo(Fn)
2712 .CatchHandlerParentFrameObjIdx[const_cast<Function *>(Fn)] =
2713 FuncInfo->getRegSaveFrameIndex();
2715 // Store the second integer parameter (rdx) into rsp+16 relative to the
2716 // stack pointer at the entry of the function.
2717 SDValue RSFIN = DAG.getFrameIndex(FuncInfo->getRegSaveFrameIndex(),
2718 getPointerTy(DAG.getDataLayout()));
2719 unsigned GPR = MF.addLiveIn(X86::RDX, &X86::GR64RegClass);
2720 SDValue Val = DAG.getCopyFromReg(Chain, dl, GPR, MVT::i64);
2721 Chain = DAG.getStore(
2722 Val.getValue(1), dl, Val, RSFIN,
2723 MachinePointerInfo::getFixedStack(FuncInfo->getRegSaveFrameIndex()),
2724 /*isVolatile=*/true, /*isNonTemporal=*/false, /*Alignment=*/0);
2727 if (isVarArg && MFI->hasMustTailInVarArgFunc()) {
2728 // Find the largest legal vector type.
2729 MVT VecVT = MVT::Other;
2730 // FIXME: Only some x86_32 calling conventions support AVX512.
2731 if (Subtarget->hasAVX512() &&
2732 (Is64Bit || (CallConv == CallingConv::X86_VectorCall ||
2733 CallConv == CallingConv::Intel_OCL_BI)))
2734 VecVT = MVT::v16f32;
2735 else if (Subtarget->hasAVX())
2737 else if (Subtarget->hasSSE2())
2740 // We forward some GPRs and some vector types.
2741 SmallVector<MVT, 2> RegParmTypes;
2742 MVT IntVT = Is64Bit ? MVT::i64 : MVT::i32;
2743 RegParmTypes.push_back(IntVT);
2744 if (VecVT != MVT::Other)
2745 RegParmTypes.push_back(VecVT);
2747 // Compute the set of forwarded registers. The rest are scratch.
2748 SmallVectorImpl<ForwardedRegister> &Forwards =
2749 FuncInfo->getForwardedMustTailRegParms();
2750 CCInfo.analyzeMustTailForwardedRegisters(Forwards, RegParmTypes, CC_X86);
2752 // Conservatively forward AL on x86_64, since it might be used for varargs.
2753 if (Is64Bit && !CCInfo.isAllocated(X86::AL)) {
2754 unsigned ALVReg = MF.addLiveIn(X86::AL, &X86::GR8RegClass);
2755 Forwards.push_back(ForwardedRegister(ALVReg, X86::AL, MVT::i8));
2758 // Copy all forwards from physical to virtual registers.
2759 for (ForwardedRegister &F : Forwards) {
2760 // FIXME: Can we use a less constrained schedule?
2761 SDValue RegVal = DAG.getCopyFromReg(Chain, dl, F.VReg, F.VT);
2762 F.VReg = MF.getRegInfo().createVirtualRegister(getRegClassFor(F.VT));
2763 Chain = DAG.getCopyToReg(Chain, dl, F.VReg, RegVal);
2767 // Some CCs need callee pop.
2768 if (X86::isCalleePop(CallConv, Is64Bit, isVarArg,
2769 MF.getTarget().Options.GuaranteedTailCallOpt)) {
2770 FuncInfo->setBytesToPopOnReturn(StackSize); // Callee pops everything.
2772 FuncInfo->setBytesToPopOnReturn(0); // Callee pops nothing.
2773 // If this is an sret function, the return should pop the hidden pointer.
2774 if (!Is64Bit && !IsTailCallConvention(CallConv) &&
2775 !Subtarget->getTargetTriple().isOSMSVCRT() &&
2776 argsAreStructReturn(Ins) == StackStructReturn)
2777 FuncInfo->setBytesToPopOnReturn(4);
2781 // RegSaveFrameIndex is X86-64 only.
2782 FuncInfo->setRegSaveFrameIndex(0xAAAAAAA);
2783 if (CallConv == CallingConv::X86_FastCall ||
2784 CallConv == CallingConv::X86_ThisCall)
2785 // fastcc functions can't have varargs.
2786 FuncInfo->setVarArgsFrameIndex(0xAAAAAAA);
2789 FuncInfo->setArgumentStackSize(StackSize);
2791 if (IsWinEHParent) {
2793 int UnwindHelpFI = MFI->CreateStackObject(8, 8, /*isSS=*/false);
2794 SDValue StackSlot = DAG.getFrameIndex(UnwindHelpFI, MVT::i64);
2795 MMI.getWinEHFuncInfo(MF.getFunction()).UnwindHelpFrameIdx = UnwindHelpFI;
2796 SDValue Neg2 = DAG.getConstant(-2, dl, MVT::i64);
2797 Chain = DAG.getStore(Chain, dl, Neg2, StackSlot,
2798 MachinePointerInfo::getFixedStack(UnwindHelpFI),
2799 /*isVolatile=*/true,
2800 /*isNonTemporal=*/false, /*Alignment=*/0);
2802 // Functions using Win32 EH are considered to have opaque SP adjustments
2803 // to force local variables to be addressed from the frame or base
2805 MFI->setHasOpaqueSPAdjustment(true);
2813 X86TargetLowering::LowerMemOpCallTo(SDValue Chain,
2814 SDValue StackPtr, SDValue Arg,
2815 SDLoc dl, SelectionDAG &DAG,
2816 const CCValAssign &VA,
2817 ISD::ArgFlagsTy Flags) const {
2818 unsigned LocMemOffset = VA.getLocMemOffset();
2819 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset, dl);
2820 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(DAG.getDataLayout()),
2822 if (Flags.isByVal())
2823 return CreateCopyOfByValArgument(Arg, PtrOff, Chain, Flags, DAG, dl);
2825 return DAG.getStore(Chain, dl, Arg, PtrOff,
2826 MachinePointerInfo::getStack(LocMemOffset),
2830 /// Emit a load of return address if tail call
2831 /// optimization is performed and it is required.
2833 X86TargetLowering::EmitTailCallLoadRetAddr(SelectionDAG &DAG,
2834 SDValue &OutRetAddr, SDValue Chain,
2835 bool IsTailCall, bool Is64Bit,
2836 int FPDiff, SDLoc dl) const {
2837 // Adjust the Return address stack slot.
2838 EVT VT = getPointerTy(DAG.getDataLayout());
2839 OutRetAddr = getReturnAddressFrameIndex(DAG);
2841 // Load the "old" Return address.
2842 OutRetAddr = DAG.getLoad(VT, dl, Chain, OutRetAddr, MachinePointerInfo(),
2843 false, false, false, 0);
2844 return SDValue(OutRetAddr.getNode(), 1);
2847 /// Emit a store of the return address if tail call
2848 /// optimization is performed and it is required (FPDiff!=0).
2849 static SDValue EmitTailCallStoreRetAddr(SelectionDAG &DAG, MachineFunction &MF,
2850 SDValue Chain, SDValue RetAddrFrIdx,
2851 EVT PtrVT, unsigned SlotSize,
2852 int FPDiff, SDLoc dl) {
2853 // Store the return address to the appropriate stack slot.
2854 if (!FPDiff) return Chain;
2855 // Calculate the new stack slot for the return address.
2856 int NewReturnAddrFI =
2857 MF.getFrameInfo()->CreateFixedObject(SlotSize, (int64_t)FPDiff - SlotSize,
2859 SDValue NewRetAddrFrIdx = DAG.getFrameIndex(NewReturnAddrFI, PtrVT);
2860 Chain = DAG.getStore(Chain, dl, RetAddrFrIdx, NewRetAddrFrIdx,
2861 MachinePointerInfo::getFixedStack(NewReturnAddrFI),
2866 /// Returns a vector_shuffle mask for an movs{s|d}, movd
2867 /// operation of specified width.
2868 static SDValue getMOVL(SelectionDAG &DAG, SDLoc dl, EVT VT, SDValue V1,
2870 unsigned NumElems = VT.getVectorNumElements();
2871 SmallVector<int, 8> Mask;
2872 Mask.push_back(NumElems);
2873 for (unsigned i = 1; i != NumElems; ++i)
2875 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
2879 X86TargetLowering::LowerCall(TargetLowering::CallLoweringInfo &CLI,
2880 SmallVectorImpl<SDValue> &InVals) const {
2881 SelectionDAG &DAG = CLI.DAG;
2883 SmallVectorImpl<ISD::OutputArg> &Outs = CLI.Outs;
2884 SmallVectorImpl<SDValue> &OutVals = CLI.OutVals;
2885 SmallVectorImpl<ISD::InputArg> &Ins = CLI.Ins;
2886 SDValue Chain = CLI.Chain;
2887 SDValue Callee = CLI.Callee;
2888 CallingConv::ID CallConv = CLI.CallConv;
2889 bool &isTailCall = CLI.IsTailCall;
2890 bool isVarArg = CLI.IsVarArg;
2892 MachineFunction &MF = DAG.getMachineFunction();
2893 bool Is64Bit = Subtarget->is64Bit();
2894 bool IsWin64 = Subtarget->isCallingConvWin64(CallConv);
2895 StructReturnType SR = callIsStructReturn(Outs);
2896 bool IsSibcall = false;
2897 X86MachineFunctionInfo *X86Info = MF.getInfo<X86MachineFunctionInfo>();
2898 auto Attr = MF.getFunction()->getFnAttribute("disable-tail-calls");
2900 if (Attr.getValueAsString() == "true")
2903 if (Subtarget->isPICStyleGOT() &&
2904 !MF.getTarget().Options.GuaranteedTailCallOpt) {
2905 // If we are using a GOT, disable tail calls to external symbols with
2906 // default visibility. Tail calling such a symbol requires using a GOT
2907 // relocation, which forces early binding of the symbol. This breaks code
2908 // that require lazy function symbol resolution. Using musttail or
2909 // GuaranteedTailCallOpt will override this.
2910 GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee);
2911 if (!G || (!G->getGlobal()->hasLocalLinkage() &&
2912 G->getGlobal()->hasDefaultVisibility()))
2916 bool IsMustTail = CLI.CS && CLI.CS->isMustTailCall();
2918 // Force this to be a tail call. The verifier rules are enough to ensure
2919 // that we can lower this successfully without moving the return address
2922 } else if (isTailCall) {
2923 // Check if it's really possible to do a tail call.
2924 isTailCall = IsEligibleForTailCallOptimization(Callee, CallConv,
2925 isVarArg, SR != NotStructReturn,
2926 MF.getFunction()->hasStructRetAttr(), CLI.RetTy,
2927 Outs, OutVals, Ins, DAG);
2929 // Sibcalls are automatically detected tailcalls which do not require
2931 if (!MF.getTarget().Options.GuaranteedTailCallOpt && isTailCall)
2938 assert(!(isVarArg && IsTailCallConvention(CallConv)) &&
2939 "Var args not supported with calling convention fastcc, ghc or hipe");
2941 // Analyze operands of the call, assigning locations to each operand.
2942 SmallVector<CCValAssign, 16> ArgLocs;
2943 CCState CCInfo(CallConv, isVarArg, MF, ArgLocs, *DAG.getContext());
2945 // Allocate shadow area for Win64
2947 CCInfo.AllocateStack(32, 8);
2949 CCInfo.AnalyzeCallOperands(Outs, CC_X86);
2951 // Get a count of how many bytes are to be pushed on the stack.
2952 unsigned NumBytes = CCInfo.getNextStackOffset();
2954 // This is a sibcall. The memory operands are available in caller's
2955 // own caller's stack.
2957 else if (MF.getTarget().Options.GuaranteedTailCallOpt &&
2958 IsTailCallConvention(CallConv))
2959 NumBytes = GetAlignedArgumentStackSize(NumBytes, DAG);
2962 if (isTailCall && !IsSibcall && !IsMustTail) {
2963 // Lower arguments at fp - stackoffset + fpdiff.
2964 unsigned NumBytesCallerPushed = X86Info->getBytesToPopOnReturn();
2966 FPDiff = NumBytesCallerPushed - NumBytes;
2968 // Set the delta of movement of the returnaddr stackslot.
2969 // But only set if delta is greater than previous delta.
2970 if (FPDiff < X86Info->getTCReturnAddrDelta())
2971 X86Info->setTCReturnAddrDelta(FPDiff);
2974 unsigned NumBytesToPush = NumBytes;
2975 unsigned NumBytesToPop = NumBytes;
2977 // If we have an inalloca argument, all stack space has already been allocated
2978 // for us and be right at the top of the stack. We don't support multiple
2979 // arguments passed in memory when using inalloca.
2980 if (!Outs.empty() && Outs.back().Flags.isInAlloca()) {
2982 if (!ArgLocs.back().isMemLoc())
2983 report_fatal_error("cannot use inalloca attribute on a register "
2985 if (ArgLocs.back().getLocMemOffset() != 0)
2986 report_fatal_error("any parameter with the inalloca attribute must be "
2987 "the only memory argument");
2991 Chain = DAG.getCALLSEQ_START(
2992 Chain, DAG.getIntPtrConstant(NumBytesToPush, dl, true), dl);
2994 SDValue RetAddrFrIdx;
2995 // Load return address for tail calls.
2996 if (isTailCall && FPDiff)
2997 Chain = EmitTailCallLoadRetAddr(DAG, RetAddrFrIdx, Chain, isTailCall,
2998 Is64Bit, FPDiff, dl);
3000 SmallVector<std::pair<unsigned, SDValue>, 8> RegsToPass;
3001 SmallVector<SDValue, 8> MemOpChains;
3004 // Walk the register/memloc assignments, inserting copies/loads. In the case
3005 // of tail call optimization arguments are handle later.
3006 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
3007 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
3008 // Skip inalloca arguments, they have already been written.
3009 ISD::ArgFlagsTy Flags = Outs[i].Flags;
3010 if (Flags.isInAlloca())
3013 CCValAssign &VA = ArgLocs[i];
3014 EVT RegVT = VA.getLocVT();
3015 SDValue Arg = OutVals[i];
3016 bool isByVal = Flags.isByVal();
3018 // Promote the value if needed.
3019 switch (VA.getLocInfo()) {
3020 default: llvm_unreachable("Unknown loc info!");
3021 case CCValAssign::Full: break;
3022 case CCValAssign::SExt:
3023 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, RegVT, Arg);
3025 case CCValAssign::ZExt:
3026 Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, RegVT, Arg);
3028 case CCValAssign::AExt:
3029 if (Arg.getValueType().isVector() &&
3030 Arg.getValueType().getScalarType() == MVT::i1)
3031 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, RegVT, Arg);
3032 else if (RegVT.is128BitVector()) {
3033 // Special case: passing MMX values in XMM registers.
3034 Arg = DAG.getBitcast(MVT::i64, Arg);
3035 Arg = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i64, Arg);
3036 Arg = getMOVL(DAG, dl, MVT::v2i64, DAG.getUNDEF(MVT::v2i64), Arg);
3038 Arg = DAG.getNode(ISD::ANY_EXTEND, dl, RegVT, Arg);
3040 case CCValAssign::BCvt:
3041 Arg = DAG.getBitcast(RegVT, Arg);
3043 case CCValAssign::Indirect: {
3044 // Store the argument.
3045 SDValue SpillSlot = DAG.CreateStackTemporary(VA.getValVT());
3046 int FI = cast<FrameIndexSDNode>(SpillSlot)->getIndex();
3047 Chain = DAG.getStore(Chain, dl, Arg, SpillSlot,
3048 MachinePointerInfo::getFixedStack(FI),
3055 if (VA.isRegLoc()) {
3056 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
3057 if (isVarArg && IsWin64) {
3058 // Win64 ABI requires argument XMM reg to be copied to the corresponding
3059 // shadow reg if callee is a varargs function.
3060 unsigned ShadowReg = 0;
3061 switch (VA.getLocReg()) {
3062 case X86::XMM0: ShadowReg = X86::RCX; break;
3063 case X86::XMM1: ShadowReg = X86::RDX; break;
3064 case X86::XMM2: ShadowReg = X86::R8; break;
3065 case X86::XMM3: ShadowReg = X86::R9; break;
3068 RegsToPass.push_back(std::make_pair(ShadowReg, Arg));
3070 } else if (!IsSibcall && (!isTailCall || isByVal)) {
3071 assert(VA.isMemLoc());
3072 if (!StackPtr.getNode())
3073 StackPtr = DAG.getCopyFromReg(Chain, dl, RegInfo->getStackRegister(),
3074 getPointerTy(DAG.getDataLayout()));
3075 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Arg,
3076 dl, DAG, VA, Flags));
3080 if (!MemOpChains.empty())
3081 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, MemOpChains);
3083 if (Subtarget->isPICStyleGOT()) {
3084 // ELF / PIC requires GOT in the EBX register before function calls via PLT
3087 RegsToPass.push_back(std::make_pair(
3088 unsigned(X86::EBX), DAG.getNode(X86ISD::GlobalBaseReg, SDLoc(),
3089 getPointerTy(DAG.getDataLayout()))));
3091 // If we are tail calling and generating PIC/GOT style code load the
3092 // address of the callee into ECX. The value in ecx is used as target of
3093 // the tail jump. This is done to circumvent the ebx/callee-saved problem
3094 // for tail calls on PIC/GOT architectures. Normally we would just put the
3095 // address of GOT into ebx and then call target@PLT. But for tail calls
3096 // ebx would be restored (since ebx is callee saved) before jumping to the
3099 // Note: The actual moving to ECX is done further down.
3100 GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee);
3101 if (G && !G->getGlobal()->hasLocalLinkage() &&
3102 G->getGlobal()->hasDefaultVisibility())
3103 Callee = LowerGlobalAddress(Callee, DAG);
3104 else if (isa<ExternalSymbolSDNode>(Callee))
3105 Callee = LowerExternalSymbol(Callee, DAG);
3109 if (Is64Bit && isVarArg && !IsWin64 && !IsMustTail) {
3110 // From AMD64 ABI document:
3111 // For calls that may call functions that use varargs or stdargs
3112 // (prototype-less calls or calls to functions containing ellipsis (...) in
3113 // the declaration) %al is used as hidden argument to specify the number
3114 // of SSE registers used. The contents of %al do not need to match exactly
3115 // the number of registers, but must be an ubound on the number of SSE
3116 // registers used and is in the range 0 - 8 inclusive.
3118 // Count the number of XMM registers allocated.
3119 static const MCPhysReg XMMArgRegs[] = {
3120 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
3121 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
3123 unsigned NumXMMRegs = CCInfo.getFirstUnallocated(XMMArgRegs);
3124 assert((Subtarget->hasSSE1() || !NumXMMRegs)
3125 && "SSE registers cannot be used when SSE is disabled");
3127 RegsToPass.push_back(std::make_pair(unsigned(X86::AL),
3128 DAG.getConstant(NumXMMRegs, dl,
3132 if (isVarArg && IsMustTail) {
3133 const auto &Forwards = X86Info->getForwardedMustTailRegParms();
3134 for (const auto &F : Forwards) {
3135 SDValue Val = DAG.getCopyFromReg(Chain, dl, F.VReg, F.VT);
3136 RegsToPass.push_back(std::make_pair(unsigned(F.PReg), Val));
3140 // For tail calls lower the arguments to the 'real' stack slots. Sibcalls
3141 // don't need this because the eligibility check rejects calls that require
3142 // shuffling arguments passed in memory.
3143 if (!IsSibcall && isTailCall) {
3144 // Force all the incoming stack arguments to be loaded from the stack
3145 // before any new outgoing arguments are stored to the stack, because the
3146 // outgoing stack slots may alias the incoming argument stack slots, and
3147 // the alias isn't otherwise explicit. This is slightly more conservative
3148 // than necessary, because it means that each store effectively depends
3149 // on every argument instead of just those arguments it would clobber.
3150 SDValue ArgChain = DAG.getStackArgumentTokenFactor(Chain);
3152 SmallVector<SDValue, 8> MemOpChains2;
3155 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
3156 CCValAssign &VA = ArgLocs[i];
3159 assert(VA.isMemLoc());
3160 SDValue Arg = OutVals[i];
3161 ISD::ArgFlagsTy Flags = Outs[i].Flags;
3162 // Skip inalloca arguments. They don't require any work.
3163 if (Flags.isInAlloca())
3165 // Create frame index.
3166 int32_t Offset = VA.getLocMemOffset()+FPDiff;
3167 uint32_t OpSize = (VA.getLocVT().getSizeInBits()+7)/8;
3168 FI = MF.getFrameInfo()->CreateFixedObject(OpSize, Offset, true);
3169 FIN = DAG.getFrameIndex(FI, getPointerTy(DAG.getDataLayout()));
3171 if (Flags.isByVal()) {
3172 // Copy relative to framepointer.
3173 SDValue Source = DAG.getIntPtrConstant(VA.getLocMemOffset(), dl);
3174 if (!StackPtr.getNode())
3175 StackPtr = DAG.getCopyFromReg(Chain, dl, RegInfo->getStackRegister(),
3176 getPointerTy(DAG.getDataLayout()));
3177 Source = DAG.getNode(ISD::ADD, dl, getPointerTy(DAG.getDataLayout()),
3180 MemOpChains2.push_back(CreateCopyOfByValArgument(Source, FIN,
3184 // Store relative to framepointer.
3185 MemOpChains2.push_back(
3186 DAG.getStore(ArgChain, dl, Arg, FIN,
3187 MachinePointerInfo::getFixedStack(FI),
3192 if (!MemOpChains2.empty())
3193 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, MemOpChains2);
3195 // Store the return address to the appropriate stack slot.
3196 Chain = EmitTailCallStoreRetAddr(DAG, MF, Chain, RetAddrFrIdx,
3197 getPointerTy(DAG.getDataLayout()),
3198 RegInfo->getSlotSize(), FPDiff, dl);
3201 // Build a sequence of copy-to-reg nodes chained together with token chain
3202 // and flag operands which copy the outgoing args into registers.
3204 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
3205 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
3206 RegsToPass[i].second, InFlag);
3207 InFlag = Chain.getValue(1);
3210 if (DAG.getTarget().getCodeModel() == CodeModel::Large) {
3211 assert(Is64Bit && "Large code model is only legal in 64-bit mode.");
3212 // In the 64-bit large code model, we have to make all calls
3213 // through a register, since the call instruction's 32-bit
3214 // pc-relative offset may not be large enough to hold the whole
3216 } else if (Callee->getOpcode() == ISD::GlobalAddress) {
3217 // If the callee is a GlobalAddress node (quite common, every direct call
3218 // is) turn it into a TargetGlobalAddress node so that legalize doesn't hack
3220 GlobalAddressSDNode* G = cast<GlobalAddressSDNode>(Callee);
3222 // We should use extra load for direct calls to dllimported functions in
3224 const GlobalValue *GV = G->getGlobal();
3225 if (!GV->hasDLLImportStorageClass()) {
3226 unsigned char OpFlags = 0;
3227 bool ExtraLoad = false;
3228 unsigned WrapperKind = ISD::DELETED_NODE;
3230 // On ELF targets, in both X86-64 and X86-32 mode, direct calls to
3231 // external symbols most go through the PLT in PIC mode. If the symbol
3232 // has hidden or protected visibility, or if it is static or local, then
3233 // we don't need to use the PLT - we can directly call it.
3234 if (Subtarget->isTargetELF() &&
3235 DAG.getTarget().getRelocationModel() == Reloc::PIC_ &&
3236 GV->hasDefaultVisibility() && !GV->hasLocalLinkage()) {
3237 OpFlags = X86II::MO_PLT;
3238 } else if (Subtarget->isPICStyleStubAny() &&
3239 !GV->isStrongDefinitionForLinker() &&
3240 (!Subtarget->getTargetTriple().isMacOSX() ||
3241 Subtarget->getTargetTriple().isMacOSXVersionLT(10, 5))) {
3242 // PC-relative references to external symbols should go through $stub,
3243 // unless we're building with the leopard linker or later, which
3244 // automatically synthesizes these stubs.
3245 OpFlags = X86II::MO_DARWIN_STUB;
3246 } else if (Subtarget->isPICStyleRIPRel() && isa<Function>(GV) &&
3247 cast<Function>(GV)->hasFnAttribute(Attribute::NonLazyBind)) {
3248 // If the function is marked as non-lazy, generate an indirect call
3249 // which loads from the GOT directly. This avoids runtime overhead
3250 // at the cost of eager binding (and one extra byte of encoding).
3251 OpFlags = X86II::MO_GOTPCREL;
3252 WrapperKind = X86ISD::WrapperRIP;
3256 Callee = DAG.getTargetGlobalAddress(
3257 GV, dl, getPointerTy(DAG.getDataLayout()), G->getOffset(), OpFlags);
3259 // Add a wrapper if needed.
3260 if (WrapperKind != ISD::DELETED_NODE)
3261 Callee = DAG.getNode(X86ISD::WrapperRIP, dl,
3262 getPointerTy(DAG.getDataLayout()), Callee);
3263 // Add extra indirection if needed.
3265 Callee = DAG.getLoad(
3266 getPointerTy(DAG.getDataLayout()), dl, DAG.getEntryNode(), Callee,
3267 MachinePointerInfo::getGOT(), false, false, false, 0);
3269 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
3270 unsigned char OpFlags = 0;
3272 // On ELF targets, in either X86-64 or X86-32 mode, direct calls to
3273 // external symbols should go through the PLT.
3274 if (Subtarget->isTargetELF() &&
3275 DAG.getTarget().getRelocationModel() == Reloc::PIC_) {
3276 OpFlags = X86II::MO_PLT;
3277 } else if (Subtarget->isPICStyleStubAny() &&
3278 (!Subtarget->getTargetTriple().isMacOSX() ||
3279 Subtarget->getTargetTriple().isMacOSXVersionLT(10, 5))) {
3280 // PC-relative references to external symbols should go through $stub,
3281 // unless we're building with the leopard linker or later, which
3282 // automatically synthesizes these stubs.
3283 OpFlags = X86II::MO_DARWIN_STUB;
3286 Callee = DAG.getTargetExternalSymbol(
3287 S->getSymbol(), getPointerTy(DAG.getDataLayout()), OpFlags);
3288 } else if (Subtarget->isTarget64BitILP32() &&
3289 Callee->getValueType(0) == MVT::i32) {
3290 // Zero-extend the 32-bit Callee address into a 64-bit according to x32 ABI
3291 Callee = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i64, Callee);
3294 // Returns a chain & a flag for retval copy to use.
3295 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
3296 SmallVector<SDValue, 8> Ops;
3298 if (!IsSibcall && isTailCall) {
3299 Chain = DAG.getCALLSEQ_END(Chain,
3300 DAG.getIntPtrConstant(NumBytesToPop, dl, true),
3301 DAG.getIntPtrConstant(0, dl, true), InFlag, dl);
3302 InFlag = Chain.getValue(1);
3305 Ops.push_back(Chain);
3306 Ops.push_back(Callee);
3309 Ops.push_back(DAG.getConstant(FPDiff, dl, MVT::i32));
3311 // Add argument registers to the end of the list so that they are known live
3313 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
3314 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
3315 RegsToPass[i].second.getValueType()));
3317 // Add a register mask operand representing the call-preserved registers.
3318 const uint32_t *Mask = RegInfo->getCallPreservedMask(MF, CallConv);
3319 assert(Mask && "Missing call preserved mask for calling convention");
3321 // If this is an invoke in a 32-bit function using an MSVC personality, assume
3322 // the function clobbers all registers. If an exception is thrown, the runtime
3323 // will not restore CSRs.
3324 // FIXME: Model this more precisely so that we can register allocate across
3325 // the normal edge and spill and fill across the exceptional edge.
3326 if (!Is64Bit && CLI.CS && CLI.CS->isInvoke()) {
3327 const Function *CallerFn = MF.getFunction();
3328 EHPersonality Pers =
3329 CallerFn->hasPersonalityFn()
3330 ? classifyEHPersonality(CallerFn->getPersonalityFn())
3331 : EHPersonality::Unknown;
3332 if (isMSVCEHPersonality(Pers))
3333 Mask = RegInfo->getNoPreservedMask();
3336 Ops.push_back(DAG.getRegisterMask(Mask));
3338 if (InFlag.getNode())
3339 Ops.push_back(InFlag);
3343 //// If this is the first return lowered for this function, add the regs
3344 //// to the liveout set for the function.
3345 // This isn't right, although it's probably harmless on x86; liveouts
3346 // should be computed from returns not tail calls. Consider a void
3347 // function making a tail call to a function returning int.
3348 MF.getFrameInfo()->setHasTailCall();
3349 return DAG.getNode(X86ISD::TC_RETURN, dl, NodeTys, Ops);
3352 Chain = DAG.getNode(X86ISD::CALL, dl, NodeTys, Ops);
3353 InFlag = Chain.getValue(1);
3355 // Create the CALLSEQ_END node.
3356 unsigned NumBytesForCalleeToPop;
3357 if (X86::isCalleePop(CallConv, Is64Bit, isVarArg,
3358 DAG.getTarget().Options.GuaranteedTailCallOpt))
3359 NumBytesForCalleeToPop = NumBytes; // Callee pops everything
3360 else if (!Is64Bit && !IsTailCallConvention(CallConv) &&
3361 !Subtarget->getTargetTriple().isOSMSVCRT() &&
3362 SR == StackStructReturn)
3363 // If this is a call to a struct-return function, the callee
3364 // pops the hidden struct pointer, so we have to push it back.
3365 // This is common for Darwin/X86, Linux & Mingw32 targets.
3366 // For MSVC Win32 targets, the caller pops the hidden struct pointer.
3367 NumBytesForCalleeToPop = 4;
3369 NumBytesForCalleeToPop = 0; // Callee pops nothing.
3371 // Returns a flag for retval copy to use.
3373 Chain = DAG.getCALLSEQ_END(Chain,
3374 DAG.getIntPtrConstant(NumBytesToPop, dl, true),
3375 DAG.getIntPtrConstant(NumBytesForCalleeToPop, dl,
3378 InFlag = Chain.getValue(1);
3381 // Handle result values, copying them out of physregs into vregs that we
3383 return LowerCallResult(Chain, InFlag, CallConv, isVarArg,
3384 Ins, dl, DAG, InVals);
3387 //===----------------------------------------------------------------------===//
3388 // Fast Calling Convention (tail call) implementation
3389 //===----------------------------------------------------------------------===//
3391 // Like std call, callee cleans arguments, convention except that ECX is
3392 // reserved for storing the tail called function address. Only 2 registers are
3393 // free for argument passing (inreg). Tail call optimization is performed
3395 // * tailcallopt is enabled
3396 // * caller/callee are fastcc
3397 // On X86_64 architecture with GOT-style position independent code only local
3398 // (within module) calls are supported at the moment.
3399 // To keep the stack aligned according to platform abi the function
3400 // GetAlignedArgumentStackSize ensures that argument delta is always multiples
3401 // of stack alignment. (Dynamic linkers need this - darwin's dyld for example)
3402 // If a tail called function callee has more arguments than the caller the
3403 // caller needs to make sure that there is room to move the RETADDR to. This is
3404 // achieved by reserving an area the size of the argument delta right after the
3405 // original RETADDR, but before the saved framepointer or the spilled registers
3406 // e.g. caller(arg1, arg2) calls callee(arg1, arg2,arg3,arg4)
3418 /// Make the stack size align e.g 16n + 12 aligned for a 16-byte align
3421 X86TargetLowering::GetAlignedArgumentStackSize(unsigned StackSize,
3422 SelectionDAG& DAG) const {
3423 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
3424 const TargetFrameLowering &TFI = *Subtarget->getFrameLowering();
3425 unsigned StackAlignment = TFI.getStackAlignment();
3426 uint64_t AlignMask = StackAlignment - 1;
3427 int64_t Offset = StackSize;
3428 unsigned SlotSize = RegInfo->getSlotSize();
3429 if ( (Offset & AlignMask) <= (StackAlignment - SlotSize) ) {
3430 // Number smaller than 12 so just add the difference.
3431 Offset += ((StackAlignment - SlotSize) - (Offset & AlignMask));
3433 // Mask out lower bits, add stackalignment once plus the 12 bytes.
3434 Offset = ((~AlignMask) & Offset) + StackAlignment +
3435 (StackAlignment-SlotSize);
3440 /// Return true if the given stack call argument is already available in the
3441 /// same position (relatively) of the caller's incoming argument stack.
3443 bool MatchingStackOffset(SDValue Arg, unsigned Offset, ISD::ArgFlagsTy Flags,
3444 MachineFrameInfo *MFI, const MachineRegisterInfo *MRI,
3445 const X86InstrInfo *TII) {
3446 unsigned Bytes = Arg.getValueType().getSizeInBits() / 8;
3448 if (Arg.getOpcode() == ISD::CopyFromReg) {
3449 unsigned VR = cast<RegisterSDNode>(Arg.getOperand(1))->getReg();
3450 if (!TargetRegisterInfo::isVirtualRegister(VR))
3452 MachineInstr *Def = MRI->getVRegDef(VR);
3455 if (!Flags.isByVal()) {
3456 if (!TII->isLoadFromStackSlot(Def, FI))
3459 unsigned Opcode = Def->getOpcode();
3460 if ((Opcode == X86::LEA32r || Opcode == X86::LEA64r ||
3461 Opcode == X86::LEA64_32r) &&
3462 Def->getOperand(1).isFI()) {
3463 FI = Def->getOperand(1).getIndex();
3464 Bytes = Flags.getByValSize();
3468 } else if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Arg)) {
3469 if (Flags.isByVal())
3470 // ByVal argument is passed in as a pointer but it's now being
3471 // dereferenced. e.g.
3472 // define @foo(%struct.X* %A) {
3473 // tail call @bar(%struct.X* byval %A)
3476 SDValue Ptr = Ld->getBasePtr();
3477 FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr);
3480 FI = FINode->getIndex();
3481 } else if (Arg.getOpcode() == ISD::FrameIndex && Flags.isByVal()) {
3482 FrameIndexSDNode *FINode = cast<FrameIndexSDNode>(Arg);
3483 FI = FINode->getIndex();
3484 Bytes = Flags.getByValSize();
3488 assert(FI != INT_MAX);
3489 if (!MFI->isFixedObjectIndex(FI))
3491 return Offset == MFI->getObjectOffset(FI) && Bytes == MFI->getObjectSize(FI);
3494 /// Check whether the call is eligible for tail call optimization. Targets
3495 /// that want to do tail call optimization should implement this function.
3497 X86TargetLowering::IsEligibleForTailCallOptimization(SDValue Callee,
3498 CallingConv::ID CalleeCC,
3500 bool isCalleeStructRet,
3501 bool isCallerStructRet,
3503 const SmallVectorImpl<ISD::OutputArg> &Outs,
3504 const SmallVectorImpl<SDValue> &OutVals,
3505 const SmallVectorImpl<ISD::InputArg> &Ins,
3506 SelectionDAG &DAG) const {
3507 if (!IsTailCallConvention(CalleeCC) && !IsCCallConvention(CalleeCC))
3510 // If -tailcallopt is specified, make fastcc functions tail-callable.
3511 const MachineFunction &MF = DAG.getMachineFunction();
3512 const Function *CallerF = MF.getFunction();
3514 // If the function return type is x86_fp80 and the callee return type is not,
3515 // then the FP_EXTEND of the call result is not a nop. It's not safe to
3516 // perform a tailcall optimization here.
3517 if (CallerF->getReturnType()->isX86_FP80Ty() && !RetTy->isX86_FP80Ty())
3520 CallingConv::ID CallerCC = CallerF->getCallingConv();
3521 bool CCMatch = CallerCC == CalleeCC;
3522 bool IsCalleeWin64 = Subtarget->isCallingConvWin64(CalleeCC);
3523 bool IsCallerWin64 = Subtarget->isCallingConvWin64(CallerCC);
3525 // Win64 functions have extra shadow space for argument homing. Don't do the
3526 // sibcall if the caller and callee have mismatched expectations for this
3528 if (IsCalleeWin64 != IsCallerWin64)
3531 if (DAG.getTarget().Options.GuaranteedTailCallOpt) {
3532 if (IsTailCallConvention(CalleeCC) && CCMatch)
3537 // Look for obvious safe cases to perform tail call optimization that do not
3538 // require ABI changes. This is what gcc calls sibcall.
3540 // Can't do sibcall if stack needs to be dynamically re-aligned. PEI needs to
3541 // emit a special epilogue.
3542 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
3543 if (RegInfo->needsStackRealignment(MF))
3546 // Also avoid sibcall optimization if either caller or callee uses struct
3547 // return semantics.
3548 if (isCalleeStructRet || isCallerStructRet)
3551 // An stdcall/thiscall caller is expected to clean up its arguments; the
3552 // callee isn't going to do that.
3553 // FIXME: this is more restrictive than needed. We could produce a tailcall
3554 // when the stack adjustment matches. For example, with a thiscall that takes
3555 // only one argument.
3556 if (!CCMatch && (CallerCC == CallingConv::X86_StdCall ||
3557 CallerCC == CallingConv::X86_ThisCall))
3560 // Do not sibcall optimize vararg calls unless all arguments are passed via
3562 if (isVarArg && !Outs.empty()) {
3564 // Optimizing for varargs on Win64 is unlikely to be safe without
3565 // additional testing.
3566 if (IsCalleeWin64 || IsCallerWin64)
3569 SmallVector<CCValAssign, 16> ArgLocs;
3570 CCState CCInfo(CalleeCC, isVarArg, DAG.getMachineFunction(), ArgLocs,
3573 CCInfo.AnalyzeCallOperands(Outs, CC_X86);
3574 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i)
3575 if (!ArgLocs[i].isRegLoc())
3579 // If the call result is in ST0 / ST1, it needs to be popped off the x87
3580 // stack. Therefore, if it's not used by the call it is not safe to optimize
3581 // this into a sibcall.
3582 bool Unused = false;
3583 for (unsigned i = 0, e = Ins.size(); i != e; ++i) {
3590 SmallVector<CCValAssign, 16> RVLocs;
3591 CCState CCInfo(CalleeCC, false, DAG.getMachineFunction(), RVLocs,
3593 CCInfo.AnalyzeCallResult(Ins, RetCC_X86);
3594 for (unsigned i = 0, e = RVLocs.size(); i != e; ++i) {
3595 CCValAssign &VA = RVLocs[i];
3596 if (VA.getLocReg() == X86::FP0 || VA.getLocReg() == X86::FP1)
3601 // If the calling conventions do not match, then we'd better make sure the
3602 // results are returned in the same way as what the caller expects.
3604 SmallVector<CCValAssign, 16> RVLocs1;
3605 CCState CCInfo1(CalleeCC, false, DAG.getMachineFunction(), RVLocs1,
3607 CCInfo1.AnalyzeCallResult(Ins, RetCC_X86);
3609 SmallVector<CCValAssign, 16> RVLocs2;
3610 CCState CCInfo2(CallerCC, false, DAG.getMachineFunction(), RVLocs2,
3612 CCInfo2.AnalyzeCallResult(Ins, RetCC_X86);
3614 if (RVLocs1.size() != RVLocs2.size())
3616 for (unsigned i = 0, e = RVLocs1.size(); i != e; ++i) {
3617 if (RVLocs1[i].isRegLoc() != RVLocs2[i].isRegLoc())
3619 if (RVLocs1[i].getLocInfo() != RVLocs2[i].getLocInfo())
3621 if (RVLocs1[i].isRegLoc()) {
3622 if (RVLocs1[i].getLocReg() != RVLocs2[i].getLocReg())
3625 if (RVLocs1[i].getLocMemOffset() != RVLocs2[i].getLocMemOffset())
3631 // If the callee takes no arguments then go on to check the results of the
3633 if (!Outs.empty()) {
3634 // Check if stack adjustment is needed. For now, do not do this if any
3635 // argument is passed on the stack.
3636 SmallVector<CCValAssign, 16> ArgLocs;
3637 CCState CCInfo(CalleeCC, isVarArg, DAG.getMachineFunction(), ArgLocs,
3640 // Allocate shadow area for Win64
3642 CCInfo.AllocateStack(32, 8);
3644 CCInfo.AnalyzeCallOperands(Outs, CC_X86);
3645 if (CCInfo.getNextStackOffset()) {
3646 MachineFunction &MF = DAG.getMachineFunction();
3647 if (MF.getInfo<X86MachineFunctionInfo>()->getBytesToPopOnReturn())
3650 // Check if the arguments are already laid out in the right way as
3651 // the caller's fixed stack objects.
3652 MachineFrameInfo *MFI = MF.getFrameInfo();
3653 const MachineRegisterInfo *MRI = &MF.getRegInfo();
3654 const X86InstrInfo *TII = Subtarget->getInstrInfo();
3655 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
3656 CCValAssign &VA = ArgLocs[i];
3657 SDValue Arg = OutVals[i];
3658 ISD::ArgFlagsTy Flags = Outs[i].Flags;
3659 if (VA.getLocInfo() == CCValAssign::Indirect)
3661 if (!VA.isRegLoc()) {
3662 if (!MatchingStackOffset(Arg, VA.getLocMemOffset(), Flags,
3669 // If the tailcall address may be in a register, then make sure it's
3670 // possible to register allocate for it. In 32-bit, the call address can
3671 // only target EAX, EDX, or ECX since the tail call must be scheduled after
3672 // callee-saved registers are restored. These happen to be the same
3673 // registers used to pass 'inreg' arguments so watch out for those.
3674 if (!Subtarget->is64Bit() &&
3675 ((!isa<GlobalAddressSDNode>(Callee) &&
3676 !isa<ExternalSymbolSDNode>(Callee)) ||
3677 DAG.getTarget().getRelocationModel() == Reloc::PIC_)) {
3678 unsigned NumInRegs = 0;
3679 // In PIC we need an extra register to formulate the address computation
3681 unsigned MaxInRegs =
3682 (DAG.getTarget().getRelocationModel() == Reloc::PIC_) ? 2 : 3;
3684 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
3685 CCValAssign &VA = ArgLocs[i];
3688 unsigned Reg = VA.getLocReg();
3691 case X86::EAX: case X86::EDX: case X86::ECX:
3692 if (++NumInRegs == MaxInRegs)
3704 X86TargetLowering::createFastISel(FunctionLoweringInfo &funcInfo,
3705 const TargetLibraryInfo *libInfo) const {
3706 return X86::createFastISel(funcInfo, libInfo);
3709 //===----------------------------------------------------------------------===//
3710 // Other Lowering Hooks
3711 //===----------------------------------------------------------------------===//
3713 static bool MayFoldLoad(SDValue Op) {
3714 return Op.hasOneUse() && ISD::isNormalLoad(Op.getNode());
3717 static bool MayFoldIntoStore(SDValue Op) {
3718 return Op.hasOneUse() && ISD::isNormalStore(*Op.getNode()->use_begin());
3721 static bool isTargetShuffle(unsigned Opcode) {
3723 default: return false;
3724 case X86ISD::BLENDI:
3725 case X86ISD::PSHUFB:
3726 case X86ISD::PSHUFD:
3727 case X86ISD::PSHUFHW:
3728 case X86ISD::PSHUFLW:
3730 case X86ISD::PALIGNR:
3731 case X86ISD::MOVLHPS:
3732 case X86ISD::MOVLHPD:
3733 case X86ISD::MOVHLPS:
3734 case X86ISD::MOVLPS:
3735 case X86ISD::MOVLPD:
3736 case X86ISD::MOVSHDUP:
3737 case X86ISD::MOVSLDUP:
3738 case X86ISD::MOVDDUP:
3741 case X86ISD::UNPCKL:
3742 case X86ISD::UNPCKH:
3743 case X86ISD::VPERMILPI:
3744 case X86ISD::VPERM2X128:
3745 case X86ISD::VPERMI:
3750 static SDValue getTargetShuffleNode(unsigned Opc, SDLoc dl, EVT VT,
3751 SDValue V1, unsigned TargetMask,
3752 SelectionDAG &DAG) {
3754 default: llvm_unreachable("Unknown x86 shuffle node");
3755 case X86ISD::PSHUFD:
3756 case X86ISD::PSHUFHW:
3757 case X86ISD::PSHUFLW:
3758 case X86ISD::VPERMILPI:
3759 case X86ISD::VPERMI:
3760 return DAG.getNode(Opc, dl, VT, V1,
3761 DAG.getConstant(TargetMask, dl, MVT::i8));
3765 static SDValue getTargetShuffleNode(unsigned Opc, SDLoc dl, EVT VT,
3766 SDValue V1, SDValue V2, SelectionDAG &DAG) {
3768 default: llvm_unreachable("Unknown x86 shuffle node");
3769 case X86ISD::MOVLHPS:
3770 case X86ISD::MOVLHPD:
3771 case X86ISD::MOVHLPS:
3772 case X86ISD::MOVLPS:
3773 case X86ISD::MOVLPD:
3776 case X86ISD::UNPCKL:
3777 case X86ISD::UNPCKH:
3778 return DAG.getNode(Opc, dl, VT, V1, V2);
3782 SDValue X86TargetLowering::getReturnAddressFrameIndex(SelectionDAG &DAG) const {
3783 MachineFunction &MF = DAG.getMachineFunction();
3784 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
3785 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
3786 int ReturnAddrIndex = FuncInfo->getRAIndex();
3788 if (ReturnAddrIndex == 0) {
3789 // Set up a frame object for the return address.
3790 unsigned SlotSize = RegInfo->getSlotSize();
3791 ReturnAddrIndex = MF.getFrameInfo()->CreateFixedObject(SlotSize,
3794 FuncInfo->setRAIndex(ReturnAddrIndex);
3797 return DAG.getFrameIndex(ReturnAddrIndex, getPointerTy(DAG.getDataLayout()));
3800 bool X86::isOffsetSuitableForCodeModel(int64_t Offset, CodeModel::Model M,
3801 bool hasSymbolicDisplacement) {
3802 // Offset should fit into 32 bit immediate field.
3803 if (!isInt<32>(Offset))
3806 // If we don't have a symbolic displacement - we don't have any extra
3808 if (!hasSymbolicDisplacement)
3811 // FIXME: Some tweaks might be needed for medium code model.
3812 if (M != CodeModel::Small && M != CodeModel::Kernel)
3815 // For small code model we assume that latest object is 16MB before end of 31
3816 // bits boundary. We may also accept pretty large negative constants knowing
3817 // that all objects are in the positive half of address space.
3818 if (M == CodeModel::Small && Offset < 16*1024*1024)
3821 // For kernel code model we know that all object resist in the negative half
3822 // of 32bits address space. We may not accept negative offsets, since they may
3823 // be just off and we may accept pretty large positive ones.
3824 if (M == CodeModel::Kernel && Offset >= 0)
3830 /// Determines whether the callee is required to pop its own arguments.
3831 /// Callee pop is necessary to support tail calls.
3832 bool X86::isCalleePop(CallingConv::ID CallingConv,
3833 bool is64Bit, bool IsVarArg, bool TailCallOpt) {
3834 switch (CallingConv) {
3837 case CallingConv::X86_StdCall:
3838 case CallingConv::X86_FastCall:
3839 case CallingConv::X86_ThisCall:
3841 case CallingConv::Fast:
3842 case CallingConv::GHC:
3843 case CallingConv::HiPE:
3850 /// \brief Return true if the condition is an unsigned comparison operation.
3851 static bool isX86CCUnsigned(unsigned X86CC) {
3853 default: llvm_unreachable("Invalid integer condition!");
3854 case X86::COND_E: return true;
3855 case X86::COND_G: return false;
3856 case X86::COND_GE: return false;
3857 case X86::COND_L: return false;
3858 case X86::COND_LE: return false;
3859 case X86::COND_NE: return true;
3860 case X86::COND_B: return true;
3861 case X86::COND_A: return true;
3862 case X86::COND_BE: return true;
3863 case X86::COND_AE: return true;
3865 llvm_unreachable("covered switch fell through?!");
3868 /// Do a one-to-one translation of a ISD::CondCode to the X86-specific
3869 /// condition code, returning the condition code and the LHS/RHS of the
3870 /// comparison to make.
3871 static unsigned TranslateX86CC(ISD::CondCode SetCCOpcode, SDLoc DL, bool isFP,
3872 SDValue &LHS, SDValue &RHS, SelectionDAG &DAG) {
3874 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS)) {
3875 if (SetCCOpcode == ISD::SETGT && RHSC->isAllOnesValue()) {
3876 // X > -1 -> X == 0, jump !sign.
3877 RHS = DAG.getConstant(0, DL, RHS.getValueType());
3878 return X86::COND_NS;
3880 if (SetCCOpcode == ISD::SETLT && RHSC->isNullValue()) {
3881 // X < 0 -> X == 0, jump on sign.
3884 if (SetCCOpcode == ISD::SETLT && RHSC->getZExtValue() == 1) {
3886 RHS = DAG.getConstant(0, DL, RHS.getValueType());
3887 return X86::COND_LE;
3891 switch (SetCCOpcode) {
3892 default: llvm_unreachable("Invalid integer condition!");
3893 case ISD::SETEQ: return X86::COND_E;
3894 case ISD::SETGT: return X86::COND_G;
3895 case ISD::SETGE: return X86::COND_GE;
3896 case ISD::SETLT: return X86::COND_L;
3897 case ISD::SETLE: return X86::COND_LE;
3898 case ISD::SETNE: return X86::COND_NE;
3899 case ISD::SETULT: return X86::COND_B;
3900 case ISD::SETUGT: return X86::COND_A;
3901 case ISD::SETULE: return X86::COND_BE;
3902 case ISD::SETUGE: return X86::COND_AE;
3906 // First determine if it is required or is profitable to flip the operands.
3908 // If LHS is a foldable load, but RHS is not, flip the condition.
3909 if (ISD::isNON_EXTLoad(LHS.getNode()) &&
3910 !ISD::isNON_EXTLoad(RHS.getNode())) {
3911 SetCCOpcode = getSetCCSwappedOperands(SetCCOpcode);
3912 std::swap(LHS, RHS);
3915 switch (SetCCOpcode) {
3921 std::swap(LHS, RHS);
3925 // On a floating point condition, the flags are set as follows:
3927 // 0 | 0 | 0 | X > Y
3928 // 0 | 0 | 1 | X < Y
3929 // 1 | 0 | 0 | X == Y
3930 // 1 | 1 | 1 | unordered
3931 switch (SetCCOpcode) {
3932 default: llvm_unreachable("Condcode should be pre-legalized away");
3934 case ISD::SETEQ: return X86::COND_E;
3935 case ISD::SETOLT: // flipped
3937 case ISD::SETGT: return X86::COND_A;
3938 case ISD::SETOLE: // flipped
3940 case ISD::SETGE: return X86::COND_AE;
3941 case ISD::SETUGT: // flipped
3943 case ISD::SETLT: return X86::COND_B;
3944 case ISD::SETUGE: // flipped
3946 case ISD::SETLE: return X86::COND_BE;
3948 case ISD::SETNE: return X86::COND_NE;
3949 case ISD::SETUO: return X86::COND_P;
3950 case ISD::SETO: return X86::COND_NP;
3952 case ISD::SETUNE: return X86::COND_INVALID;
3956 /// Is there a floating point cmov for the specific X86 condition code?
3957 /// Current x86 isa includes the following FP cmov instructions:
3958 /// fcmovb, fcomvbe, fcomve, fcmovu, fcmovae, fcmova, fcmovne, fcmovnu.
3959 static bool hasFPCMov(unsigned X86CC) {
3975 /// Returns true if the target can instruction select the
3976 /// specified FP immediate natively. If false, the legalizer will
3977 /// materialize the FP immediate as a load from a constant pool.
3978 bool X86TargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
3979 for (unsigned i = 0, e = LegalFPImmediates.size(); i != e; ++i) {
3980 if (Imm.bitwiseIsEqual(LegalFPImmediates[i]))
3986 bool X86TargetLowering::shouldReduceLoadWidth(SDNode *Load,
3987 ISD::LoadExtType ExtTy,
3989 // "ELF Handling for Thread-Local Storage" specifies that R_X86_64_GOTTPOFF
3990 // relocation target a movq or addq instruction: don't let the load shrink.
3991 SDValue BasePtr = cast<LoadSDNode>(Load)->getBasePtr();
3992 if (BasePtr.getOpcode() == X86ISD::WrapperRIP)
3993 if (const auto *GA = dyn_cast<GlobalAddressSDNode>(BasePtr.getOperand(0)))
3994 return GA->getTargetFlags() != X86II::MO_GOTTPOFF;
3998 /// \brief Returns true if it is beneficial to convert a load of a constant
3999 /// to just the constant itself.
4000 bool X86TargetLowering::shouldConvertConstantLoadToIntImm(const APInt &Imm,
4002 assert(Ty->isIntegerTy());
4004 unsigned BitSize = Ty->getPrimitiveSizeInBits();
4005 if (BitSize == 0 || BitSize > 64)
4010 bool X86TargetLowering::isExtractSubvectorCheap(EVT ResVT,
4011 unsigned Index) const {
4012 if (!isOperationLegalOrCustom(ISD::EXTRACT_SUBVECTOR, ResVT))
4015 return (Index == 0 || Index == ResVT.getVectorNumElements());
4018 bool X86TargetLowering::isCheapToSpeculateCttz() const {
4019 // Speculate cttz only if we can directly use TZCNT.
4020 return Subtarget->hasBMI();
4023 bool X86TargetLowering::isCheapToSpeculateCtlz() const {
4024 // Speculate ctlz only if we can directly use LZCNT.
4025 return Subtarget->hasLZCNT();
4028 /// Return true if every element in Mask, beginning
4029 /// from position Pos and ending in Pos+Size is undef.
4030 static bool isUndefInRange(ArrayRef<int> Mask, unsigned Pos, unsigned Size) {
4031 for (unsigned i = Pos, e = Pos + Size; i != e; ++i)
4037 /// Return true if Val is undef or if its value falls within the
4038 /// specified range (L, H].
4039 static bool isUndefOrInRange(int Val, int Low, int Hi) {
4040 return (Val < 0) || (Val >= Low && Val < Hi);
4043 /// Val is either less than zero (undef) or equal to the specified value.
4044 static bool isUndefOrEqual(int Val, int CmpVal) {
4045 return (Val < 0 || Val == CmpVal);
4048 /// Return true if every element in Mask, beginning
4049 /// from position Pos and ending in Pos+Size, falls within the specified
4050 /// sequential range (Low, Low+Size]. or is undef.
4051 static bool isSequentialOrUndefInRange(ArrayRef<int> Mask,
4052 unsigned Pos, unsigned Size, int Low) {
4053 for (unsigned i = Pos, e = Pos+Size; i != e; ++i, ++Low)
4054 if (!isUndefOrEqual(Mask[i], Low))
4059 /// Return true if the specified EXTRACT_SUBVECTOR operand specifies a vector
4060 /// extract that is suitable for instruction that extract 128 or 256 bit vectors
4061 static bool isVEXTRACTIndex(SDNode *N, unsigned vecWidth) {
4062 assert((vecWidth == 128 || vecWidth == 256) && "Unexpected vector width");
4063 if (!isa<ConstantSDNode>(N->getOperand(1).getNode()))
4066 // The index should be aligned on a vecWidth-bit boundary.
4068 cast<ConstantSDNode>(N->getOperand(1).getNode())->getZExtValue();
4070 MVT VT = N->getSimpleValueType(0);
4071 unsigned ElSize = VT.getVectorElementType().getSizeInBits();
4072 bool Result = (Index * ElSize) % vecWidth == 0;
4077 /// Return true if the specified INSERT_SUBVECTOR
4078 /// operand specifies a subvector insert that is suitable for input to
4079 /// insertion of 128 or 256-bit subvectors
4080 static bool isVINSERTIndex(SDNode *N, unsigned vecWidth) {
4081 assert((vecWidth == 128 || vecWidth == 256) && "Unexpected vector width");
4082 if (!isa<ConstantSDNode>(N->getOperand(2).getNode()))
4084 // The index should be aligned on a vecWidth-bit boundary.
4086 cast<ConstantSDNode>(N->getOperand(2).getNode())->getZExtValue();
4088 MVT VT = N->getSimpleValueType(0);
4089 unsigned ElSize = VT.getVectorElementType().getSizeInBits();
4090 bool Result = (Index * ElSize) % vecWidth == 0;
4095 bool X86::isVINSERT128Index(SDNode *N) {
4096 return isVINSERTIndex(N, 128);
4099 bool X86::isVINSERT256Index(SDNode *N) {
4100 return isVINSERTIndex(N, 256);
4103 bool X86::isVEXTRACT128Index(SDNode *N) {
4104 return isVEXTRACTIndex(N, 128);
4107 bool X86::isVEXTRACT256Index(SDNode *N) {
4108 return isVEXTRACTIndex(N, 256);
4111 static unsigned getExtractVEXTRACTImmediate(SDNode *N, unsigned vecWidth) {
4112 assert((vecWidth == 128 || vecWidth == 256) && "Unsupported vector width");
4113 if (!isa<ConstantSDNode>(N->getOperand(1).getNode()))
4114 llvm_unreachable("Illegal extract subvector for VEXTRACT");
4117 cast<ConstantSDNode>(N->getOperand(1).getNode())->getZExtValue();
4119 MVT VecVT = N->getOperand(0).getSimpleValueType();
4120 MVT ElVT = VecVT.getVectorElementType();
4122 unsigned NumElemsPerChunk = vecWidth / ElVT.getSizeInBits();
4123 return Index / NumElemsPerChunk;
4126 static unsigned getInsertVINSERTImmediate(SDNode *N, unsigned vecWidth) {
4127 assert((vecWidth == 128 || vecWidth == 256) && "Unsupported vector width");
4128 if (!isa<ConstantSDNode>(N->getOperand(2).getNode()))
4129 llvm_unreachable("Illegal insert subvector for VINSERT");
4132 cast<ConstantSDNode>(N->getOperand(2).getNode())->getZExtValue();
4134 MVT VecVT = N->getSimpleValueType(0);
4135 MVT ElVT = VecVT.getVectorElementType();
4137 unsigned NumElemsPerChunk = vecWidth / ElVT.getSizeInBits();
4138 return Index / NumElemsPerChunk;
4141 /// Return the appropriate immediate to extract the specified
4142 /// EXTRACT_SUBVECTOR index with VEXTRACTF128 and VINSERTI128 instructions.
4143 unsigned X86::getExtractVEXTRACT128Immediate(SDNode *N) {
4144 return getExtractVEXTRACTImmediate(N, 128);
4147 /// Return the appropriate immediate to extract the specified
4148 /// EXTRACT_SUBVECTOR index with VEXTRACTF64x4 and VINSERTI64x4 instructions.
4149 unsigned X86::getExtractVEXTRACT256Immediate(SDNode *N) {
4150 return getExtractVEXTRACTImmediate(N, 256);
4153 /// Return the appropriate immediate to insert at the specified
4154 /// INSERT_SUBVECTOR index with VINSERTF128 and VINSERTI128 instructions.
4155 unsigned X86::getInsertVINSERT128Immediate(SDNode *N) {
4156 return getInsertVINSERTImmediate(N, 128);
4159 /// Return the appropriate immediate to insert at the specified
4160 /// INSERT_SUBVECTOR index with VINSERTF46x4 and VINSERTI64x4 instructions.
4161 unsigned X86::getInsertVINSERT256Immediate(SDNode *N) {
4162 return getInsertVINSERTImmediate(N, 256);
4165 /// Returns true if Elt is a constant integer zero
4166 static bool isZero(SDValue V) {
4167 ConstantSDNode *C = dyn_cast<ConstantSDNode>(V);
4168 return C && C->isNullValue();
4171 /// Returns true if Elt is a constant zero or a floating point constant +0.0.
4172 bool X86::isZeroNode(SDValue Elt) {
4175 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(Elt))
4176 return CFP->getValueAPF().isPosZero();
4180 /// Returns a vector of specified type with all zero elements.
4181 static SDValue getZeroVector(EVT VT, const X86Subtarget *Subtarget,
4182 SelectionDAG &DAG, SDLoc dl) {
4183 assert(VT.isVector() && "Expected a vector type");
4185 // Always build SSE zero vectors as <4 x i32> bitcasted
4186 // to their dest type. This ensures they get CSE'd.
4188 if (VT.is128BitVector()) { // SSE
4189 if (Subtarget->hasSSE2()) { // SSE2
4190 SDValue Cst = DAG.getConstant(0, dl, MVT::i32);
4191 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
4193 SDValue Cst = DAG.getConstantFP(+0.0, dl, MVT::f32);
4194 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4f32, Cst, Cst, Cst, Cst);
4196 } else if (VT.is256BitVector()) { // AVX
4197 if (Subtarget->hasInt256()) { // AVX2
4198 SDValue Cst = DAG.getConstant(0, dl, MVT::i32);
4199 SDValue Ops[] = { Cst, Cst, Cst, Cst, Cst, Cst, Cst, Cst };
4200 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v8i32, Ops);
4202 // 256-bit logic and arithmetic instructions in AVX are all
4203 // floating-point, no support for integer ops. Emit fp zeroed vectors.
4204 SDValue Cst = DAG.getConstantFP(+0.0, dl, MVT::f32);
4205 SDValue Ops[] = { Cst, Cst, Cst, Cst, Cst, Cst, Cst, Cst };
4206 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v8f32, Ops);
4208 } else if (VT.is512BitVector()) { // AVX-512
4209 SDValue Cst = DAG.getConstant(0, dl, MVT::i32);
4210 SDValue Ops[] = { Cst, Cst, Cst, Cst, Cst, Cst, Cst, Cst,
4211 Cst, Cst, Cst, Cst, Cst, Cst, Cst, Cst };
4212 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v16i32, Ops);
4213 } else if (VT.getScalarType() == MVT::i1) {
4215 assert((Subtarget->hasBWI() || VT.getVectorNumElements() <= 16)
4216 && "Unexpected vector type");
4217 assert((Subtarget->hasVLX() || VT.getVectorNumElements() >= 8)
4218 && "Unexpected vector type");
4219 SDValue Cst = DAG.getConstant(0, dl, MVT::i1);
4220 SmallVector<SDValue, 64> Ops(VT.getVectorNumElements(), Cst);
4221 return DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Ops);
4223 llvm_unreachable("Unexpected vector type");
4225 return DAG.getBitcast(VT, Vec);
4228 static SDValue ExtractSubVector(SDValue Vec, unsigned IdxVal,
4229 SelectionDAG &DAG, SDLoc dl,
4230 unsigned vectorWidth) {
4231 assert((vectorWidth == 128 || vectorWidth == 256) &&
4232 "Unsupported vector width");
4233 EVT VT = Vec.getValueType();
4234 EVT ElVT = VT.getVectorElementType();
4235 unsigned Factor = VT.getSizeInBits()/vectorWidth;
4236 EVT ResultVT = EVT::getVectorVT(*DAG.getContext(), ElVT,
4237 VT.getVectorNumElements()/Factor);
4239 // Extract from UNDEF is UNDEF.
4240 if (Vec.getOpcode() == ISD::UNDEF)
4241 return DAG.getUNDEF(ResultVT);
4243 // Extract the relevant vectorWidth bits. Generate an EXTRACT_SUBVECTOR
4244 unsigned ElemsPerChunk = vectorWidth / ElVT.getSizeInBits();
4246 // This is the index of the first element of the vectorWidth-bit chunk
4248 unsigned NormalizedIdxVal = (((IdxVal * ElVT.getSizeInBits()) / vectorWidth)
4251 // If the input is a buildvector just emit a smaller one.
4252 if (Vec.getOpcode() == ISD::BUILD_VECTOR)
4253 return DAG.getNode(ISD::BUILD_VECTOR, dl, ResultVT,
4254 makeArrayRef(Vec->op_begin() + NormalizedIdxVal,
4257 SDValue VecIdx = DAG.getIntPtrConstant(NormalizedIdxVal, dl);
4258 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, ResultVT, Vec, VecIdx);
4261 /// Generate a DAG to grab 128-bits from a vector > 128 bits. This
4262 /// sets things up to match to an AVX VEXTRACTF128 / VEXTRACTI128
4263 /// or AVX-512 VEXTRACTF32x4 / VEXTRACTI32x4
4264 /// instructions or a simple subregister reference. Idx is an index in the
4265 /// 128 bits we want. It need not be aligned to a 128-bit boundary. That makes
4266 /// lowering EXTRACT_VECTOR_ELT operations easier.
4267 static SDValue Extract128BitVector(SDValue Vec, unsigned IdxVal,
4268 SelectionDAG &DAG, SDLoc dl) {
4269 assert((Vec.getValueType().is256BitVector() ||
4270 Vec.getValueType().is512BitVector()) && "Unexpected vector size!");
4271 return ExtractSubVector(Vec, IdxVal, DAG, dl, 128);
4274 /// Generate a DAG to grab 256-bits from a 512-bit vector.
4275 static SDValue Extract256BitVector(SDValue Vec, unsigned IdxVal,
4276 SelectionDAG &DAG, SDLoc dl) {
4277 assert(Vec.getValueType().is512BitVector() && "Unexpected vector size!");
4278 return ExtractSubVector(Vec, IdxVal, DAG, dl, 256);
4281 static SDValue InsertSubVector(SDValue Result, SDValue Vec,
4282 unsigned IdxVal, SelectionDAG &DAG,
4283 SDLoc dl, unsigned vectorWidth) {
4284 assert((vectorWidth == 128 || vectorWidth == 256) &&
4285 "Unsupported vector width");
4286 // Inserting UNDEF is Result
4287 if (Vec.getOpcode() == ISD::UNDEF)
4289 EVT VT = Vec.getValueType();
4290 EVT ElVT = VT.getVectorElementType();
4291 EVT ResultVT = Result.getValueType();
4293 // Insert the relevant vectorWidth bits.
4294 unsigned ElemsPerChunk = vectorWidth/ElVT.getSizeInBits();
4296 // This is the index of the first element of the vectorWidth-bit chunk
4298 unsigned NormalizedIdxVal = (((IdxVal * ElVT.getSizeInBits())/vectorWidth)
4301 SDValue VecIdx = DAG.getIntPtrConstant(NormalizedIdxVal, dl);
4302 return DAG.getNode(ISD::INSERT_SUBVECTOR, dl, ResultVT, Result, Vec, VecIdx);
4305 /// Generate a DAG to put 128-bits into a vector > 128 bits. This
4306 /// sets things up to match to an AVX VINSERTF128/VINSERTI128 or
4307 /// AVX-512 VINSERTF32x4/VINSERTI32x4 instructions or a
4308 /// simple superregister reference. Idx is an index in the 128 bits
4309 /// we want. It need not be aligned to a 128-bit boundary. That makes
4310 /// lowering INSERT_VECTOR_ELT operations easier.
4311 static SDValue Insert128BitVector(SDValue Result, SDValue Vec, unsigned IdxVal,
4312 SelectionDAG &DAG, SDLoc dl) {
4313 assert(Vec.getValueType().is128BitVector() && "Unexpected vector size!");
4315 // For insertion into the zero index (low half) of a 256-bit vector, it is
4316 // more efficient to generate a blend with immediate instead of an insert*128.
4317 // We are still creating an INSERT_SUBVECTOR below with an undef node to
4318 // extend the subvector to the size of the result vector. Make sure that
4319 // we are not recursing on that node by checking for undef here.
4320 if (IdxVal == 0 && Result.getValueType().is256BitVector() &&
4321 Result.getOpcode() != ISD::UNDEF) {
4322 EVT ResultVT = Result.getValueType();
4323 SDValue ZeroIndex = DAG.getIntPtrConstant(0, dl);
4324 SDValue Undef = DAG.getUNDEF(ResultVT);
4325 SDValue Vec256 = DAG.getNode(ISD::INSERT_SUBVECTOR, dl, ResultVT, Undef,
4328 // The blend instruction, and therefore its mask, depend on the data type.
4329 MVT ScalarType = ResultVT.getScalarType().getSimpleVT();
4330 if (ScalarType.isFloatingPoint()) {
4331 // Choose either vblendps (float) or vblendpd (double).
4332 unsigned ScalarSize = ScalarType.getSizeInBits();
4333 assert((ScalarSize == 64 || ScalarSize == 32) && "Unknown float type");
4334 unsigned MaskVal = (ScalarSize == 64) ? 0x03 : 0x0f;
4335 SDValue Mask = DAG.getConstant(MaskVal, dl, MVT::i8);
4336 return DAG.getNode(X86ISD::BLENDI, dl, ResultVT, Result, Vec256, Mask);
4339 const X86Subtarget &Subtarget =
4340 static_cast<const X86Subtarget &>(DAG.getSubtarget());
4342 // AVX2 is needed for 256-bit integer blend support.
4343 // Integers must be cast to 32-bit because there is only vpblendd;
4344 // vpblendw can't be used for this because it has a handicapped mask.
4346 // If we don't have AVX2, then cast to float. Using a wrong domain blend
4347 // is still more efficient than using the wrong domain vinsertf128 that
4348 // will be created by InsertSubVector().
4349 MVT CastVT = Subtarget.hasAVX2() ? MVT::v8i32 : MVT::v8f32;
4351 SDValue Mask = DAG.getConstant(0x0f, dl, MVT::i8);
4352 Vec256 = DAG.getBitcast(CastVT, Vec256);
4353 Vec256 = DAG.getNode(X86ISD::BLENDI, dl, CastVT, Result, Vec256, Mask);
4354 return DAG.getBitcast(ResultVT, Vec256);
4357 return InsertSubVector(Result, Vec, IdxVal, DAG, dl, 128);
4360 static SDValue Insert256BitVector(SDValue Result, SDValue Vec, unsigned IdxVal,
4361 SelectionDAG &DAG, SDLoc dl) {
4362 assert(Vec.getValueType().is256BitVector() && "Unexpected vector size!");
4363 return InsertSubVector(Result, Vec, IdxVal, DAG, dl, 256);
4366 /// Concat two 128-bit vectors into a 256 bit vector using VINSERTF128
4367 /// instructions. This is used because creating CONCAT_VECTOR nodes of
4368 /// BUILD_VECTORS returns a larger BUILD_VECTOR while we're trying to lower
4369 /// large BUILD_VECTORS.
4370 static SDValue Concat128BitVectors(SDValue V1, SDValue V2, EVT VT,
4371 unsigned NumElems, SelectionDAG &DAG,
4373 SDValue V = Insert128BitVector(DAG.getUNDEF(VT), V1, 0, DAG, dl);
4374 return Insert128BitVector(V, V2, NumElems/2, DAG, dl);
4377 static SDValue Concat256BitVectors(SDValue V1, SDValue V2, EVT VT,
4378 unsigned NumElems, SelectionDAG &DAG,
4380 SDValue V = Insert256BitVector(DAG.getUNDEF(VT), V1, 0, DAG, dl);
4381 return Insert256BitVector(V, V2, NumElems/2, DAG, dl);
4384 /// Returns a vector of specified type with all bits set.
4385 /// Always build ones vectors as <4 x i32> or <8 x i32>. For 256-bit types with
4386 /// no AVX2 supprt, use two <4 x i32> inserted in a <8 x i32> appropriately.
4387 /// Then bitcast to their original type, ensuring they get CSE'd.
4388 static SDValue getOnesVector(MVT VT, bool HasInt256, SelectionDAG &DAG,
4390 assert(VT.isVector() && "Expected a vector type");
4392 SDValue Cst = DAG.getConstant(~0U, dl, MVT::i32);
4394 if (VT.is256BitVector()) {
4395 if (HasInt256) { // AVX2
4396 SDValue Ops[] = { Cst, Cst, Cst, Cst, Cst, Cst, Cst, Cst };
4397 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v8i32, Ops);
4399 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
4400 Vec = Concat128BitVectors(Vec, Vec, MVT::v8i32, 8, DAG, dl);
4402 } else if (VT.is128BitVector()) {
4403 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
4405 llvm_unreachable("Unexpected vector type");
4407 return DAG.getBitcast(VT, Vec);
4410 /// Returns a vector_shuffle node for an unpackl operation.
4411 static SDValue getUnpackl(SelectionDAG &DAG, SDLoc dl, MVT VT, SDValue V1,
4413 unsigned NumElems = VT.getVectorNumElements();
4414 SmallVector<int, 8> Mask;
4415 for (unsigned i = 0, e = NumElems/2; i != e; ++i) {
4417 Mask.push_back(i + NumElems);
4419 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
4422 /// Returns a vector_shuffle node for an unpackh operation.
4423 static SDValue getUnpackh(SelectionDAG &DAG, SDLoc dl, MVT VT, SDValue V1,
4425 unsigned NumElems = VT.getVectorNumElements();
4426 SmallVector<int, 8> Mask;
4427 for (unsigned i = 0, Half = NumElems/2; i != Half; ++i) {
4428 Mask.push_back(i + Half);
4429 Mask.push_back(i + NumElems + Half);
4431 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
4434 /// Return a vector_shuffle of the specified vector of zero or undef vector.
4435 /// This produces a shuffle where the low element of V2 is swizzled into the
4436 /// zero/undef vector, landing at element Idx.
4437 /// This produces a shuffle mask like 4,1,2,3 (idx=0) or 0,1,2,4 (idx=3).
4438 static SDValue getShuffleVectorZeroOrUndef(SDValue V2, unsigned Idx,
4440 const X86Subtarget *Subtarget,
4441 SelectionDAG &DAG) {
4442 MVT VT = V2.getSimpleValueType();
4444 ? getZeroVector(VT, Subtarget, DAG, SDLoc(V2)) : DAG.getUNDEF(VT);
4445 unsigned NumElems = VT.getVectorNumElements();
4446 SmallVector<int, 16> MaskVec;
4447 for (unsigned i = 0; i != NumElems; ++i)
4448 // If this is the insertion idx, put the low elt of V2 here.
4449 MaskVec.push_back(i == Idx ? NumElems : i);
4450 return DAG.getVectorShuffle(VT, SDLoc(V2), V1, V2, &MaskVec[0]);
4453 /// Calculates the shuffle mask corresponding to the target-specific opcode.
4454 /// Returns true if the Mask could be calculated. Sets IsUnary to true if only
4455 /// uses one source. Note that this will set IsUnary for shuffles which use a
4456 /// single input multiple times, and in those cases it will
4457 /// adjust the mask to only have indices within that single input.
4458 /// FIXME: Add support for Decode*Mask functions that return SM_SentinelZero.
4459 static bool getTargetShuffleMask(SDNode *N, MVT VT,
4460 SmallVectorImpl<int> &Mask, bool &IsUnary) {
4461 unsigned NumElems = VT.getVectorNumElements();
4465 bool IsFakeUnary = false;
4466 switch(N->getOpcode()) {
4467 case X86ISD::BLENDI:
4468 ImmN = N->getOperand(N->getNumOperands()-1);
4469 DecodeBLENDMask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
4472 ImmN = N->getOperand(N->getNumOperands()-1);
4473 DecodeSHUFPMask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
4474 IsUnary = IsFakeUnary = N->getOperand(0) == N->getOperand(1);
4476 case X86ISD::UNPCKH:
4477 DecodeUNPCKHMask(VT, Mask);
4478 IsUnary = IsFakeUnary = N->getOperand(0) == N->getOperand(1);
4480 case X86ISD::UNPCKL:
4481 DecodeUNPCKLMask(VT, Mask);
4482 IsUnary = IsFakeUnary = N->getOperand(0) == N->getOperand(1);
4484 case X86ISD::MOVHLPS:
4485 DecodeMOVHLPSMask(NumElems, Mask);
4486 IsUnary = IsFakeUnary = N->getOperand(0) == N->getOperand(1);
4488 case X86ISD::MOVLHPS:
4489 DecodeMOVLHPSMask(NumElems, Mask);
4490 IsUnary = IsFakeUnary = N->getOperand(0) == N->getOperand(1);
4492 case X86ISD::PALIGNR:
4493 ImmN = N->getOperand(N->getNumOperands()-1);
4494 DecodePALIGNRMask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
4496 case X86ISD::PSHUFD:
4497 case X86ISD::VPERMILPI:
4498 ImmN = N->getOperand(N->getNumOperands()-1);
4499 DecodePSHUFMask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
4502 case X86ISD::PSHUFHW:
4503 ImmN = N->getOperand(N->getNumOperands()-1);
4504 DecodePSHUFHWMask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
4507 case X86ISD::PSHUFLW:
4508 ImmN = N->getOperand(N->getNumOperands()-1);
4509 DecodePSHUFLWMask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
4512 case X86ISD::PSHUFB: {
4514 SDValue MaskNode = N->getOperand(1);
4515 while (MaskNode->getOpcode() == ISD::BITCAST)
4516 MaskNode = MaskNode->getOperand(0);
4518 if (MaskNode->getOpcode() == ISD::BUILD_VECTOR) {
4519 // If we have a build-vector, then things are easy.
4520 EVT VT = MaskNode.getValueType();
4521 assert(VT.isVector() &&
4522 "Can't produce a non-vector with a build_vector!");
4523 if (!VT.isInteger())
4526 int NumBytesPerElement = VT.getVectorElementType().getSizeInBits() / 8;
4528 SmallVector<uint64_t, 32> RawMask;
4529 for (int i = 0, e = MaskNode->getNumOperands(); i < e; ++i) {
4530 SDValue Op = MaskNode->getOperand(i);
4531 if (Op->getOpcode() == ISD::UNDEF) {
4532 RawMask.push_back((uint64_t)SM_SentinelUndef);
4535 auto *CN = dyn_cast<ConstantSDNode>(Op.getNode());
4538 APInt MaskElement = CN->getAPIntValue();
4540 // We now have to decode the element which could be any integer size and
4541 // extract each byte of it.
4542 for (int j = 0; j < NumBytesPerElement; ++j) {
4543 // Note that this is x86 and so always little endian: the low byte is
4544 // the first byte of the mask.
4545 RawMask.push_back(MaskElement.getLoBits(8).getZExtValue());
4546 MaskElement = MaskElement.lshr(8);
4549 DecodePSHUFBMask(RawMask, Mask);
4553 auto *MaskLoad = dyn_cast<LoadSDNode>(MaskNode);
4557 SDValue Ptr = MaskLoad->getBasePtr();
4558 if (Ptr->getOpcode() == X86ISD::Wrapper ||
4559 Ptr->getOpcode() == X86ISD::WrapperRIP)
4560 Ptr = Ptr->getOperand(0);
4562 auto *MaskCP = dyn_cast<ConstantPoolSDNode>(Ptr);
4563 if (!MaskCP || MaskCP->isMachineConstantPoolEntry())
4566 if (auto *C = dyn_cast<Constant>(MaskCP->getConstVal())) {
4567 DecodePSHUFBMask(C, Mask);
4575 case X86ISD::VPERMI:
4576 ImmN = N->getOperand(N->getNumOperands()-1);
4577 DecodeVPERMMask(cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
4582 DecodeScalarMoveMask(VT, /* IsLoad */ false, Mask);
4584 case X86ISD::VPERM2X128:
4585 ImmN = N->getOperand(N->getNumOperands()-1);
4586 DecodeVPERM2X128Mask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
4587 if (Mask.empty()) return false;
4588 // Mask only contains negative index if an element is zero.
4589 if (std::any_of(Mask.begin(), Mask.end(),
4590 [](int M){ return M == SM_SentinelZero; }))
4593 case X86ISD::MOVSLDUP:
4594 DecodeMOVSLDUPMask(VT, Mask);
4597 case X86ISD::MOVSHDUP:
4598 DecodeMOVSHDUPMask(VT, Mask);
4601 case X86ISD::MOVDDUP:
4602 DecodeMOVDDUPMask(VT, Mask);
4605 case X86ISD::MOVLHPD:
4606 case X86ISD::MOVLPD:
4607 case X86ISD::MOVLPS:
4608 // Not yet implemented
4610 default: llvm_unreachable("unknown target shuffle node");
4613 // If we have a fake unary shuffle, the shuffle mask is spread across two
4614 // inputs that are actually the same node. Re-map the mask to always point
4615 // into the first input.
4618 if (M >= (int)Mask.size())
4624 /// Returns the scalar element that will make up the ith
4625 /// element of the result of the vector shuffle.
4626 static SDValue getShuffleScalarElt(SDNode *N, unsigned Index, SelectionDAG &DAG,
4629 return SDValue(); // Limit search depth.
4631 SDValue V = SDValue(N, 0);
4632 EVT VT = V.getValueType();
4633 unsigned Opcode = V.getOpcode();
4635 // Recurse into ISD::VECTOR_SHUFFLE node to find scalars.
4636 if (const ShuffleVectorSDNode *SV = dyn_cast<ShuffleVectorSDNode>(N)) {
4637 int Elt = SV->getMaskElt(Index);
4640 return DAG.getUNDEF(VT.getVectorElementType());
4642 unsigned NumElems = VT.getVectorNumElements();
4643 SDValue NewV = (Elt < (int)NumElems) ? SV->getOperand(0)
4644 : SV->getOperand(1);
4645 return getShuffleScalarElt(NewV.getNode(), Elt % NumElems, DAG, Depth+1);
4648 // Recurse into target specific vector shuffles to find scalars.
4649 if (isTargetShuffle(Opcode)) {
4650 MVT ShufVT = V.getSimpleValueType();
4651 unsigned NumElems = ShufVT.getVectorNumElements();
4652 SmallVector<int, 16> ShuffleMask;
4655 if (!getTargetShuffleMask(N, ShufVT, ShuffleMask, IsUnary))
4658 int Elt = ShuffleMask[Index];
4660 return DAG.getUNDEF(ShufVT.getVectorElementType());
4662 SDValue NewV = (Elt < (int)NumElems) ? N->getOperand(0)
4664 return getShuffleScalarElt(NewV.getNode(), Elt % NumElems, DAG,
4668 // Actual nodes that may contain scalar elements
4669 if (Opcode == ISD::BITCAST) {
4670 V = V.getOperand(0);
4671 EVT SrcVT = V.getValueType();
4672 unsigned NumElems = VT.getVectorNumElements();
4674 if (!SrcVT.isVector() || SrcVT.getVectorNumElements() != NumElems)
4678 if (V.getOpcode() == ISD::SCALAR_TO_VECTOR)
4679 return (Index == 0) ? V.getOperand(0)
4680 : DAG.getUNDEF(VT.getVectorElementType());
4682 if (V.getOpcode() == ISD::BUILD_VECTOR)
4683 return V.getOperand(Index);
4688 /// Custom lower build_vector of v16i8.
4689 static SDValue LowerBuildVectorv16i8(SDValue Op, unsigned NonZeros,
4690 unsigned NumNonZero, unsigned NumZero,
4692 const X86Subtarget* Subtarget,
4693 const TargetLowering &TLI) {
4701 // SSE4.1 - use PINSRB to insert each byte directly.
4702 if (Subtarget->hasSSE41()) {
4703 for (unsigned i = 0; i < 16; ++i) {
4704 bool isNonZero = (NonZeros & (1 << i)) != 0;
4708 V = getZeroVector(MVT::v16i8, Subtarget, DAG, dl);
4710 V = DAG.getUNDEF(MVT::v16i8);
4713 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl,
4714 MVT::v16i8, V, Op.getOperand(i),
4715 DAG.getIntPtrConstant(i, dl));
4722 // Pre-SSE4.1 - merge byte pairs and insert with PINSRW.
4723 for (unsigned i = 0; i < 16; ++i) {
4724 bool ThisIsNonZero = (NonZeros & (1 << i)) != 0;
4725 if (ThisIsNonZero && First) {
4727 V = getZeroVector(MVT::v8i16, Subtarget, DAG, dl);
4729 V = DAG.getUNDEF(MVT::v8i16);
4734 SDValue ThisElt, LastElt;
4735 bool LastIsNonZero = (NonZeros & (1 << (i-1))) != 0;
4736 if (LastIsNonZero) {
4737 LastElt = DAG.getNode(ISD::ZERO_EXTEND, dl,
4738 MVT::i16, Op.getOperand(i-1));
4740 if (ThisIsNonZero) {
4741 ThisElt = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i16, Op.getOperand(i));
4742 ThisElt = DAG.getNode(ISD::SHL, dl, MVT::i16,
4743 ThisElt, DAG.getConstant(8, dl, MVT::i8));
4745 ThisElt = DAG.getNode(ISD::OR, dl, MVT::i16, ThisElt, LastElt);
4749 if (ThisElt.getNode())
4750 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, V, ThisElt,
4751 DAG.getIntPtrConstant(i/2, dl));
4755 return DAG.getBitcast(MVT::v16i8, V);
4758 /// Custom lower build_vector of v8i16.
4759 static SDValue LowerBuildVectorv8i16(SDValue Op, unsigned NonZeros,
4760 unsigned NumNonZero, unsigned NumZero,
4762 const X86Subtarget* Subtarget,
4763 const TargetLowering &TLI) {
4770 for (unsigned i = 0; i < 8; ++i) {
4771 bool isNonZero = (NonZeros & (1 << i)) != 0;
4775 V = getZeroVector(MVT::v8i16, Subtarget, DAG, dl);
4777 V = DAG.getUNDEF(MVT::v8i16);
4780 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl,
4781 MVT::v8i16, V, Op.getOperand(i),
4782 DAG.getIntPtrConstant(i, dl));
4789 /// Custom lower build_vector of v4i32 or v4f32.
4790 static SDValue LowerBuildVectorv4x32(SDValue Op, SelectionDAG &DAG,
4791 const X86Subtarget *Subtarget,
4792 const TargetLowering &TLI) {
4793 // Find all zeroable elements.
4794 std::bitset<4> Zeroable;
4795 for (int i=0; i < 4; ++i) {
4796 SDValue Elt = Op->getOperand(i);
4797 Zeroable[i] = (Elt.getOpcode() == ISD::UNDEF || X86::isZeroNode(Elt));
4799 assert(Zeroable.size() - Zeroable.count() > 1 &&
4800 "We expect at least two non-zero elements!");
4802 // We only know how to deal with build_vector nodes where elements are either
4803 // zeroable or extract_vector_elt with constant index.
4804 SDValue FirstNonZero;
4805 unsigned FirstNonZeroIdx;
4806 for (unsigned i=0; i < 4; ++i) {
4809 SDValue Elt = Op->getOperand(i);
4810 if (Elt.getOpcode() != ISD::EXTRACT_VECTOR_ELT ||
4811 !isa<ConstantSDNode>(Elt.getOperand(1)))
4813 // Make sure that this node is extracting from a 128-bit vector.
4814 MVT VT = Elt.getOperand(0).getSimpleValueType();
4815 if (!VT.is128BitVector())
4817 if (!FirstNonZero.getNode()) {
4819 FirstNonZeroIdx = i;
4823 assert(FirstNonZero.getNode() && "Unexpected build vector of all zeros!");
4824 SDValue V1 = FirstNonZero.getOperand(0);
4825 MVT VT = V1.getSimpleValueType();
4827 // See if this build_vector can be lowered as a blend with zero.
4829 unsigned EltMaskIdx, EltIdx;
4831 for (EltIdx = 0; EltIdx < 4; ++EltIdx) {
4832 if (Zeroable[EltIdx]) {
4833 // The zero vector will be on the right hand side.
4834 Mask[EltIdx] = EltIdx+4;
4838 Elt = Op->getOperand(EltIdx);
4839 // By construction, Elt is a EXTRACT_VECTOR_ELT with constant index.
4840 EltMaskIdx = cast<ConstantSDNode>(Elt.getOperand(1))->getZExtValue();
4841 if (Elt.getOperand(0) != V1 || EltMaskIdx != EltIdx)
4843 Mask[EltIdx] = EltIdx;
4847 // Let the shuffle legalizer deal with blend operations.
4848 SDValue VZero = getZeroVector(VT, Subtarget, DAG, SDLoc(Op));
4849 if (V1.getSimpleValueType() != VT)
4850 V1 = DAG.getNode(ISD::BITCAST, SDLoc(V1), VT, V1);
4851 return DAG.getVectorShuffle(VT, SDLoc(V1), V1, VZero, &Mask[0]);
4854 // See if we can lower this build_vector to a INSERTPS.
4855 if (!Subtarget->hasSSE41())
4858 SDValue V2 = Elt.getOperand(0);
4859 if (Elt == FirstNonZero && EltIdx == FirstNonZeroIdx)
4862 bool CanFold = true;
4863 for (unsigned i = EltIdx + 1; i < 4 && CanFold; ++i) {
4867 SDValue Current = Op->getOperand(i);
4868 SDValue SrcVector = Current->getOperand(0);
4871 CanFold = SrcVector == V1 &&
4872 cast<ConstantSDNode>(Current.getOperand(1))->getZExtValue() == i;
4878 assert(V1.getNode() && "Expected at least two non-zero elements!");
4879 if (V1.getSimpleValueType() != MVT::v4f32)
4880 V1 = DAG.getNode(ISD::BITCAST, SDLoc(V1), MVT::v4f32, V1);
4881 if (V2.getSimpleValueType() != MVT::v4f32)
4882 V2 = DAG.getNode(ISD::BITCAST, SDLoc(V2), MVT::v4f32, V2);
4884 // Ok, we can emit an INSERTPS instruction.
4885 unsigned ZMask = Zeroable.to_ulong();
4887 unsigned InsertPSMask = EltMaskIdx << 6 | EltIdx << 4 | ZMask;
4888 assert((InsertPSMask & ~0xFFu) == 0 && "Invalid mask!");
4890 SDValue Result = DAG.getNode(X86ISD::INSERTPS, DL, MVT::v4f32, V1, V2,
4891 DAG.getIntPtrConstant(InsertPSMask, DL));
4892 return DAG.getBitcast(VT, Result);
4895 /// Return a vector logical shift node.
4896 static SDValue getVShift(bool isLeft, EVT VT, SDValue SrcOp,
4897 unsigned NumBits, SelectionDAG &DAG,
4898 const TargetLowering &TLI, SDLoc dl) {
4899 assert(VT.is128BitVector() && "Unknown type for VShift");
4900 MVT ShVT = MVT::v2i64;
4901 unsigned Opc = isLeft ? X86ISD::VSHLDQ : X86ISD::VSRLDQ;
4902 SrcOp = DAG.getBitcast(ShVT, SrcOp);
4903 MVT ScalarShiftTy = TLI.getScalarShiftAmountTy(DAG.getDataLayout(), VT);
4904 assert(NumBits % 8 == 0 && "Only support byte sized shifts");
4905 SDValue ShiftVal = DAG.getConstant(NumBits/8, dl, ScalarShiftTy);
4906 return DAG.getBitcast(VT, DAG.getNode(Opc, dl, ShVT, SrcOp, ShiftVal));
4910 LowerAsSplatVectorLoad(SDValue SrcOp, MVT VT, SDLoc dl, SelectionDAG &DAG) {
4912 // Check if the scalar load can be widened into a vector load. And if
4913 // the address is "base + cst" see if the cst can be "absorbed" into
4914 // the shuffle mask.
4915 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(SrcOp)) {
4916 SDValue Ptr = LD->getBasePtr();
4917 if (!ISD::isNormalLoad(LD) || LD->isVolatile())
4919 EVT PVT = LD->getValueType(0);
4920 if (PVT != MVT::i32 && PVT != MVT::f32)
4925 if (FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr)) {
4926 FI = FINode->getIndex();
4928 } else if (DAG.isBaseWithConstantOffset(Ptr) &&
4929 isa<FrameIndexSDNode>(Ptr.getOperand(0))) {
4930 FI = cast<FrameIndexSDNode>(Ptr.getOperand(0))->getIndex();
4931 Offset = Ptr.getConstantOperandVal(1);
4932 Ptr = Ptr.getOperand(0);
4937 // FIXME: 256-bit vector instructions don't require a strict alignment,
4938 // improve this code to support it better.
4939 unsigned RequiredAlign = VT.getSizeInBits()/8;
4940 SDValue Chain = LD->getChain();
4941 // Make sure the stack object alignment is at least 16 or 32.
4942 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
4943 if (DAG.InferPtrAlignment(Ptr) < RequiredAlign) {
4944 if (MFI->isFixedObjectIndex(FI)) {
4945 // Can't change the alignment. FIXME: It's possible to compute
4946 // the exact stack offset and reference FI + adjust offset instead.
4947 // If someone *really* cares about this. That's the way to implement it.
4950 MFI->setObjectAlignment(FI, RequiredAlign);
4954 // (Offset % 16 or 32) must be multiple of 4. Then address is then
4955 // Ptr + (Offset & ~15).
4958 if ((Offset % RequiredAlign) & 3)
4960 int64_t StartOffset = Offset & ~(RequiredAlign-1);
4963 Ptr = DAG.getNode(ISD::ADD, DL, Ptr.getValueType(), Ptr,
4964 DAG.getConstant(StartOffset, DL, Ptr.getValueType()));
4967 int EltNo = (Offset - StartOffset) >> 2;
4968 unsigned NumElems = VT.getVectorNumElements();
4970 EVT NVT = EVT::getVectorVT(*DAG.getContext(), PVT, NumElems);
4971 SDValue V1 = DAG.getLoad(NVT, dl, Chain, Ptr,
4972 LD->getPointerInfo().getWithOffset(StartOffset),
4973 false, false, false, 0);
4975 SmallVector<int, 8> Mask(NumElems, EltNo);
4977 return DAG.getVectorShuffle(NVT, dl, V1, DAG.getUNDEF(NVT), &Mask[0]);
4983 /// Given the initializing elements 'Elts' of a vector of type 'VT', see if the
4984 /// elements can be replaced by a single large load which has the same value as
4985 /// a build_vector or insert_subvector whose loaded operands are 'Elts'.
4987 /// Example: <load i32 *a, load i32 *a+4, undef, undef> -> zextload a
4989 /// FIXME: we'd also like to handle the case where the last elements are zero
4990 /// rather than undef via VZEXT_LOAD, but we do not detect that case today.
4991 /// There's even a handy isZeroNode for that purpose.
4992 static SDValue EltsFromConsecutiveLoads(EVT VT, ArrayRef<SDValue> Elts,
4993 SDLoc &DL, SelectionDAG &DAG,
4994 bool isAfterLegalize) {
4995 unsigned NumElems = Elts.size();
4997 LoadSDNode *LDBase = nullptr;
4998 unsigned LastLoadedElt = -1U;
5000 // For each element in the initializer, see if we've found a load or an undef.
5001 // If we don't find an initial load element, or later load elements are
5002 // non-consecutive, bail out.
5003 for (unsigned i = 0; i < NumElems; ++i) {
5004 SDValue Elt = Elts[i];
5005 // Look through a bitcast.
5006 if (Elt.getNode() && Elt.getOpcode() == ISD::BITCAST)
5007 Elt = Elt.getOperand(0);
5008 if (!Elt.getNode() ||
5009 (Elt.getOpcode() != ISD::UNDEF && !ISD::isNON_EXTLoad(Elt.getNode())))
5012 if (Elt.getNode()->getOpcode() == ISD::UNDEF)
5014 LDBase = cast<LoadSDNode>(Elt.getNode());
5018 if (Elt.getOpcode() == ISD::UNDEF)
5021 LoadSDNode *LD = cast<LoadSDNode>(Elt);
5022 EVT LdVT = Elt.getValueType();
5023 // Each loaded element must be the correct fractional portion of the
5024 // requested vector load.
5025 if (LdVT.getSizeInBits() != VT.getSizeInBits() / NumElems)
5027 if (!DAG.isConsecutiveLoad(LD, LDBase, LdVT.getSizeInBits() / 8, i))
5032 // If we have found an entire vector of loads and undefs, then return a large
5033 // load of the entire vector width starting at the base pointer. If we found
5034 // consecutive loads for the low half, generate a vzext_load node.
5035 if (LastLoadedElt == NumElems - 1) {
5036 assert(LDBase && "Did not find base load for merging consecutive loads");
5037 EVT EltVT = LDBase->getValueType(0);
5038 // Ensure that the input vector size for the merged loads matches the
5039 // cumulative size of the input elements.
5040 if (VT.getSizeInBits() != EltVT.getSizeInBits() * NumElems)
5043 if (isAfterLegalize &&
5044 !DAG.getTargetLoweringInfo().isOperationLegal(ISD::LOAD, VT))
5047 SDValue NewLd = SDValue();
5049 NewLd = DAG.getLoad(VT, DL, LDBase->getChain(), LDBase->getBasePtr(),
5050 LDBase->getPointerInfo(), LDBase->isVolatile(),
5051 LDBase->isNonTemporal(), LDBase->isInvariant(),
5052 LDBase->getAlignment());
5054 if (LDBase->hasAnyUseOfValue(1)) {
5055 SDValue NewChain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other,
5057 SDValue(NewLd.getNode(), 1));
5058 DAG.ReplaceAllUsesOfValueWith(SDValue(LDBase, 1), NewChain);
5059 DAG.UpdateNodeOperands(NewChain.getNode(), SDValue(LDBase, 1),
5060 SDValue(NewLd.getNode(), 1));
5066 //TODO: The code below fires only for for loading the low v2i32 / v2f32
5067 //of a v4i32 / v4f32. It's probably worth generalizing.
5068 EVT EltVT = VT.getVectorElementType();
5069 if (NumElems == 4 && LastLoadedElt == 1 && (EltVT.getSizeInBits() == 32) &&
5070 DAG.getTargetLoweringInfo().isTypeLegal(MVT::v2i64)) {
5071 SDVTList Tys = DAG.getVTList(MVT::v2i64, MVT::Other);
5072 SDValue Ops[] = { LDBase->getChain(), LDBase->getBasePtr() };
5074 DAG.getMemIntrinsicNode(X86ISD::VZEXT_LOAD, DL, Tys, Ops, MVT::i64,
5075 LDBase->getPointerInfo(),
5076 LDBase->getAlignment(),
5077 false/*isVolatile*/, true/*ReadMem*/,
5080 // Make sure the newly-created LOAD is in the same position as LDBase in
5081 // terms of dependency. We create a TokenFactor for LDBase and ResNode, and
5082 // update uses of LDBase's output chain to use the TokenFactor.
5083 if (LDBase->hasAnyUseOfValue(1)) {
5084 SDValue NewChain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other,
5085 SDValue(LDBase, 1), SDValue(ResNode.getNode(), 1));
5086 DAG.ReplaceAllUsesOfValueWith(SDValue(LDBase, 1), NewChain);
5087 DAG.UpdateNodeOperands(NewChain.getNode(), SDValue(LDBase, 1),
5088 SDValue(ResNode.getNode(), 1));
5091 return DAG.getBitcast(VT, ResNode);
5096 /// LowerVectorBroadcast - Attempt to use the vbroadcast instruction
5097 /// to generate a splat value for the following cases:
5098 /// 1. A splat BUILD_VECTOR which uses a single scalar load, or a constant.
5099 /// 2. A splat shuffle which uses a scalar_to_vector node which comes from
5100 /// a scalar load, or a constant.
5101 /// The VBROADCAST node is returned when a pattern is found,
5102 /// or SDValue() otherwise.
5103 static SDValue LowerVectorBroadcast(SDValue Op, const X86Subtarget* Subtarget,
5104 SelectionDAG &DAG) {
5105 // VBROADCAST requires AVX.
5106 // TODO: Splats could be generated for non-AVX CPUs using SSE
5107 // instructions, but there's less potential gain for only 128-bit vectors.
5108 if (!Subtarget->hasAVX())
5111 MVT VT = Op.getSimpleValueType();
5114 assert((VT.is128BitVector() || VT.is256BitVector() || VT.is512BitVector()) &&
5115 "Unsupported vector type for broadcast.");
5120 switch (Op.getOpcode()) {
5122 // Unknown pattern found.
5125 case ISD::BUILD_VECTOR: {
5126 auto *BVOp = cast<BuildVectorSDNode>(Op.getNode());
5127 BitVector UndefElements;
5128 SDValue Splat = BVOp->getSplatValue(&UndefElements);
5130 // We need a splat of a single value to use broadcast, and it doesn't
5131 // make any sense if the value is only in one element of the vector.
5132 if (!Splat || (VT.getVectorNumElements() - UndefElements.count()) <= 1)
5136 ConstSplatVal = (Ld.getOpcode() == ISD::Constant ||
5137 Ld.getOpcode() == ISD::ConstantFP);
5139 // Make sure that all of the users of a non-constant load are from the
5140 // BUILD_VECTOR node.
5141 if (!ConstSplatVal && !BVOp->isOnlyUserOf(Ld.getNode()))
5146 case ISD::VECTOR_SHUFFLE: {
5147 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
5149 // Shuffles must have a splat mask where the first element is
5151 if ((!SVOp->isSplat()) || SVOp->getMaskElt(0) != 0)
5154 SDValue Sc = Op.getOperand(0);
5155 if (Sc.getOpcode() != ISD::SCALAR_TO_VECTOR &&
5156 Sc.getOpcode() != ISD::BUILD_VECTOR) {
5158 if (!Subtarget->hasInt256())
5161 // Use the register form of the broadcast instruction available on AVX2.
5162 if (VT.getSizeInBits() >= 256)
5163 Sc = Extract128BitVector(Sc, 0, DAG, dl);
5164 return DAG.getNode(X86ISD::VBROADCAST, dl, VT, Sc);
5167 Ld = Sc.getOperand(0);
5168 ConstSplatVal = (Ld.getOpcode() == ISD::Constant ||
5169 Ld.getOpcode() == ISD::ConstantFP);
5171 // The scalar_to_vector node and the suspected
5172 // load node must have exactly one user.
5173 // Constants may have multiple users.
5175 // AVX-512 has register version of the broadcast
5176 bool hasRegVer = Subtarget->hasAVX512() && VT.is512BitVector() &&
5177 Ld.getValueType().getSizeInBits() >= 32;
5178 if (!ConstSplatVal && ((!Sc.hasOneUse() || !Ld.hasOneUse()) &&
5185 unsigned ScalarSize = Ld.getValueType().getSizeInBits();
5186 bool IsGE256 = (VT.getSizeInBits() >= 256);
5188 // When optimizing for size, generate up to 5 extra bytes for a broadcast
5189 // instruction to save 8 or more bytes of constant pool data.
5190 // TODO: If multiple splats are generated to load the same constant,
5191 // it may be detrimental to overall size. There needs to be a way to detect
5192 // that condition to know if this is truly a size win.
5193 const Function *F = DAG.getMachineFunction().getFunction();
5194 bool OptForSize = F->hasFnAttribute(Attribute::OptimizeForSize);
5196 // Handle broadcasting a single constant scalar from the constant pool
5198 // On Sandybridge (no AVX2), it is still better to load a constant vector
5199 // from the constant pool and not to broadcast it from a scalar.
5200 // But override that restriction when optimizing for size.
5201 // TODO: Check if splatting is recommended for other AVX-capable CPUs.
5202 if (ConstSplatVal && (Subtarget->hasAVX2() || OptForSize)) {
5203 EVT CVT = Ld.getValueType();
5204 assert(!CVT.isVector() && "Must not broadcast a vector type");
5206 // Splat f32, i32, v4f64, v4i64 in all cases with AVX2.
5207 // For size optimization, also splat v2f64 and v2i64, and for size opt
5208 // with AVX2, also splat i8 and i16.
5209 // With pattern matching, the VBROADCAST node may become a VMOVDDUP.
5210 if (ScalarSize == 32 || (IsGE256 && ScalarSize == 64) ||
5211 (OptForSize && (ScalarSize == 64 || Subtarget->hasAVX2()))) {
5212 const Constant *C = nullptr;
5213 if (ConstantSDNode *CI = dyn_cast<ConstantSDNode>(Ld))
5214 C = CI->getConstantIntValue();
5215 else if (ConstantFPSDNode *CF = dyn_cast<ConstantFPSDNode>(Ld))
5216 C = CF->getConstantFPValue();
5218 assert(C && "Invalid constant type");
5220 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
5222 DAG.getConstantPool(C, TLI.getPointerTy(DAG.getDataLayout()));
5223 unsigned Alignment = cast<ConstantPoolSDNode>(CP)->getAlignment();
5224 Ld = DAG.getLoad(CVT, dl, DAG.getEntryNode(), CP,
5225 MachinePointerInfo::getConstantPool(),
5226 false, false, false, Alignment);
5228 return DAG.getNode(X86ISD::VBROADCAST, dl, VT, Ld);
5232 bool IsLoad = ISD::isNormalLoad(Ld.getNode());
5234 // Handle AVX2 in-register broadcasts.
5235 if (!IsLoad && Subtarget->hasInt256() &&
5236 (ScalarSize == 32 || (IsGE256 && ScalarSize == 64)))
5237 return DAG.getNode(X86ISD::VBROADCAST, dl, VT, Ld);
5239 // The scalar source must be a normal load.
5243 if (ScalarSize == 32 || (IsGE256 && ScalarSize == 64) ||
5244 (Subtarget->hasVLX() && ScalarSize == 64))
5245 return DAG.getNode(X86ISD::VBROADCAST, dl, VT, Ld);
5247 // The integer check is needed for the 64-bit into 128-bit so it doesn't match
5248 // double since there is no vbroadcastsd xmm
5249 if (Subtarget->hasInt256() && Ld.getValueType().isInteger()) {
5250 if (ScalarSize == 8 || ScalarSize == 16 || ScalarSize == 64)
5251 return DAG.getNode(X86ISD::VBROADCAST, dl, VT, Ld);
5254 // Unsupported broadcast.
5258 /// \brief For an EXTRACT_VECTOR_ELT with a constant index return the real
5259 /// underlying vector and index.
5261 /// Modifies \p ExtractedFromVec to the real vector and returns the real
5263 static int getUnderlyingExtractedFromVec(SDValue &ExtractedFromVec,
5265 int Idx = cast<ConstantSDNode>(ExtIdx)->getZExtValue();
5266 if (!isa<ShuffleVectorSDNode>(ExtractedFromVec))
5269 // For 256-bit vectors, LowerEXTRACT_VECTOR_ELT_SSE4 may have already
5271 // (extract_vector_elt (v8f32 %vreg1), Constant<6>)
5273 // (extract_vector_elt (vector_shuffle<2,u,u,u>
5274 // (extract_subvector (v8f32 %vreg0), Constant<4>),
5277 // In this case the vector is the extract_subvector expression and the index
5278 // is 2, as specified by the shuffle.
5279 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(ExtractedFromVec);
5280 SDValue ShuffleVec = SVOp->getOperand(0);
5281 MVT ShuffleVecVT = ShuffleVec.getSimpleValueType();
5282 assert(ShuffleVecVT.getVectorElementType() ==
5283 ExtractedFromVec.getSimpleValueType().getVectorElementType());
5285 int ShuffleIdx = SVOp->getMaskElt(Idx);
5286 if (isUndefOrInRange(ShuffleIdx, 0, ShuffleVecVT.getVectorNumElements())) {
5287 ExtractedFromVec = ShuffleVec;
5293 static SDValue buildFromShuffleMostly(SDValue Op, SelectionDAG &DAG) {
5294 MVT VT = Op.getSimpleValueType();
5296 // Skip if insert_vec_elt is not supported.
5297 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
5298 if (!TLI.isOperationLegalOrCustom(ISD::INSERT_VECTOR_ELT, VT))
5302 unsigned NumElems = Op.getNumOperands();
5306 SmallVector<unsigned, 4> InsertIndices;
5307 SmallVector<int, 8> Mask(NumElems, -1);
5309 for (unsigned i = 0; i != NumElems; ++i) {
5310 unsigned Opc = Op.getOperand(i).getOpcode();
5312 if (Opc == ISD::UNDEF)
5315 if (Opc != ISD::EXTRACT_VECTOR_ELT) {
5316 // Quit if more than 1 elements need inserting.
5317 if (InsertIndices.size() > 1)
5320 InsertIndices.push_back(i);
5324 SDValue ExtractedFromVec = Op.getOperand(i).getOperand(0);
5325 SDValue ExtIdx = Op.getOperand(i).getOperand(1);
5326 // Quit if non-constant index.
5327 if (!isa<ConstantSDNode>(ExtIdx))
5329 int Idx = getUnderlyingExtractedFromVec(ExtractedFromVec, ExtIdx);
5331 // Quit if extracted from vector of different type.
5332 if (ExtractedFromVec.getValueType() != VT)
5335 if (!VecIn1.getNode())
5336 VecIn1 = ExtractedFromVec;
5337 else if (VecIn1 != ExtractedFromVec) {
5338 if (!VecIn2.getNode())
5339 VecIn2 = ExtractedFromVec;
5340 else if (VecIn2 != ExtractedFromVec)
5341 // Quit if more than 2 vectors to shuffle
5345 if (ExtractedFromVec == VecIn1)
5347 else if (ExtractedFromVec == VecIn2)
5348 Mask[i] = Idx + NumElems;
5351 if (!VecIn1.getNode())
5354 VecIn2 = VecIn2.getNode() ? VecIn2 : DAG.getUNDEF(VT);
5355 SDValue NV = DAG.getVectorShuffle(VT, DL, VecIn1, VecIn2, &Mask[0]);
5356 for (unsigned i = 0, e = InsertIndices.size(); i != e; ++i) {
5357 unsigned Idx = InsertIndices[i];
5358 NV = DAG.getNode(ISD::INSERT_VECTOR_ELT, DL, VT, NV, Op.getOperand(Idx),
5359 DAG.getIntPtrConstant(Idx, DL));
5365 static SDValue ConvertI1VectorToInteger(SDValue Op, SelectionDAG &DAG) {
5366 assert(ISD::isBuildVectorOfConstantSDNodes(Op.getNode()) &&
5367 Op.getScalarValueSizeInBits() == 1 &&
5368 "Can not convert non-constant vector");
5369 uint64_t Immediate = 0;
5370 for (unsigned idx = 0, e = Op.getNumOperands(); idx < e; ++idx) {
5371 SDValue In = Op.getOperand(idx);
5372 if (In.getOpcode() != ISD::UNDEF)
5373 Immediate |= cast<ConstantSDNode>(In)->getZExtValue() << idx;
5377 MVT::getIntegerVT(std::max((int)Op.getValueType().getSizeInBits(), 8));
5378 return DAG.getConstant(Immediate, dl, VT);
5380 // Lower BUILD_VECTOR operation for v8i1 and v16i1 types.
5382 X86TargetLowering::LowerBUILD_VECTORvXi1(SDValue Op, SelectionDAG &DAG) const {
5384 MVT VT = Op.getSimpleValueType();
5385 assert((VT.getVectorElementType() == MVT::i1) &&
5386 "Unexpected type in LowerBUILD_VECTORvXi1!");
5389 if (ISD::isBuildVectorAllZeros(Op.getNode())) {
5390 SDValue Cst = DAG.getTargetConstant(0, dl, MVT::i1);
5391 SmallVector<SDValue, 16> Ops(VT.getVectorNumElements(), Cst);
5392 return DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Ops);
5395 if (ISD::isBuildVectorAllOnes(Op.getNode())) {
5396 SDValue Cst = DAG.getTargetConstant(1, dl, MVT::i1);
5397 SmallVector<SDValue, 16> Ops(VT.getVectorNumElements(), Cst);
5398 return DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Ops);
5401 if (ISD::isBuildVectorOfConstantSDNodes(Op.getNode())) {
5402 SDValue Imm = ConvertI1VectorToInteger(Op, DAG);
5403 if (Imm.getValueSizeInBits() == VT.getSizeInBits())
5404 return DAG.getBitcast(VT, Imm);
5405 SDValue ExtVec = DAG.getBitcast(MVT::v8i1, Imm);
5406 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT, ExtVec,
5407 DAG.getIntPtrConstant(0, dl));
5410 // Vector has one or more non-const elements
5411 uint64_t Immediate = 0;
5412 SmallVector<unsigned, 16> NonConstIdx;
5413 bool IsSplat = true;
5414 bool HasConstElts = false;
5416 for (unsigned idx = 0, e = Op.getNumOperands(); idx < e; ++idx) {
5417 SDValue In = Op.getOperand(idx);
5418 if (In.getOpcode() == ISD::UNDEF)
5420 if (!isa<ConstantSDNode>(In))
5421 NonConstIdx.push_back(idx);
5423 Immediate |= cast<ConstantSDNode>(In)->getZExtValue() << idx;
5424 HasConstElts = true;
5428 else if (In != Op.getOperand(SplatIdx))
5432 // for splat use " (select i1 splat_elt, all-ones, all-zeroes)"
5434 return DAG.getNode(ISD::SELECT, dl, VT, Op.getOperand(SplatIdx),
5435 DAG.getConstant(1, dl, VT),
5436 DAG.getConstant(0, dl, VT));
5438 // insert elements one by one
5442 MVT ImmVT = MVT::getIntegerVT(std::max((int)VT.getSizeInBits(), 8));
5443 Imm = DAG.getConstant(Immediate, dl, ImmVT);
5445 else if (HasConstElts)
5446 Imm = DAG.getConstant(0, dl, VT);
5448 Imm = DAG.getUNDEF(VT);
5449 if (Imm.getValueSizeInBits() == VT.getSizeInBits())
5450 DstVec = DAG.getBitcast(VT, Imm);
5452 SDValue ExtVec = DAG.getBitcast(MVT::v8i1, Imm);
5453 DstVec = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT, ExtVec,
5454 DAG.getIntPtrConstant(0, dl));
5457 for (unsigned i = 0; i < NonConstIdx.size(); ++i) {
5458 unsigned InsertIdx = NonConstIdx[i];
5459 DstVec = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, DstVec,
5460 Op.getOperand(InsertIdx),
5461 DAG.getIntPtrConstant(InsertIdx, dl));
5466 /// \brief Return true if \p N implements a horizontal binop and return the
5467 /// operands for the horizontal binop into V0 and V1.
5469 /// This is a helper function of LowerToHorizontalOp().
5470 /// This function checks that the build_vector \p N in input implements a
5471 /// horizontal operation. Parameter \p Opcode defines the kind of horizontal
5472 /// operation to match.
5473 /// For example, if \p Opcode is equal to ISD::ADD, then this function
5474 /// checks if \p N implements a horizontal arithmetic add; if instead \p Opcode
5475 /// is equal to ISD::SUB, then this function checks if this is a horizontal
5478 /// This function only analyzes elements of \p N whose indices are
5479 /// in range [BaseIdx, LastIdx).
5480 static bool isHorizontalBinOp(const BuildVectorSDNode *N, unsigned Opcode,
5482 unsigned BaseIdx, unsigned LastIdx,
5483 SDValue &V0, SDValue &V1) {
5484 EVT VT = N->getValueType(0);
5486 assert(BaseIdx * 2 <= LastIdx && "Invalid Indices in input!");
5487 assert(VT.isVector() && VT.getVectorNumElements() >= LastIdx &&
5488 "Invalid Vector in input!");
5490 bool IsCommutable = (Opcode == ISD::ADD || Opcode == ISD::FADD);
5491 bool CanFold = true;
5492 unsigned ExpectedVExtractIdx = BaseIdx;
5493 unsigned NumElts = LastIdx - BaseIdx;
5494 V0 = DAG.getUNDEF(VT);
5495 V1 = DAG.getUNDEF(VT);
5497 // Check if N implements a horizontal binop.
5498 for (unsigned i = 0, e = NumElts; i != e && CanFold; ++i) {
5499 SDValue Op = N->getOperand(i + BaseIdx);
5502 if (Op->getOpcode() == ISD::UNDEF) {
5503 // Update the expected vector extract index.
5504 if (i * 2 == NumElts)
5505 ExpectedVExtractIdx = BaseIdx;
5506 ExpectedVExtractIdx += 2;
5510 CanFold = Op->getOpcode() == Opcode && Op->hasOneUse();
5515 SDValue Op0 = Op.getOperand(0);
5516 SDValue Op1 = Op.getOperand(1);
5518 // Try to match the following pattern:
5519 // (BINOP (extract_vector_elt A, I), (extract_vector_elt A, I+1))
5520 CanFold = (Op0.getOpcode() == ISD::EXTRACT_VECTOR_ELT &&
5521 Op1.getOpcode() == ISD::EXTRACT_VECTOR_ELT &&
5522 Op0.getOperand(0) == Op1.getOperand(0) &&
5523 isa<ConstantSDNode>(Op0.getOperand(1)) &&
5524 isa<ConstantSDNode>(Op1.getOperand(1)));
5528 unsigned I0 = cast<ConstantSDNode>(Op0.getOperand(1))->getZExtValue();
5529 unsigned I1 = cast<ConstantSDNode>(Op1.getOperand(1))->getZExtValue();
5531 if (i * 2 < NumElts) {
5532 if (V0.getOpcode() == ISD::UNDEF) {
5533 V0 = Op0.getOperand(0);
5534 if (V0.getValueType() != VT)
5538 if (V1.getOpcode() == ISD::UNDEF) {
5539 V1 = Op0.getOperand(0);
5540 if (V1.getValueType() != VT)
5543 if (i * 2 == NumElts)
5544 ExpectedVExtractIdx = BaseIdx;
5547 SDValue Expected = (i * 2 < NumElts) ? V0 : V1;
5548 if (I0 == ExpectedVExtractIdx)
5549 CanFold = I1 == I0 + 1 && Op0.getOperand(0) == Expected;
5550 else if (IsCommutable && I1 == ExpectedVExtractIdx) {
5551 // Try to match the following dag sequence:
5552 // (BINOP (extract_vector_elt A, I+1), (extract_vector_elt A, I))
5553 CanFold = I0 == I1 + 1 && Op1.getOperand(0) == Expected;
5557 ExpectedVExtractIdx += 2;
5563 /// \brief Emit a sequence of two 128-bit horizontal add/sub followed by
5564 /// a concat_vector.
5566 /// This is a helper function of LowerToHorizontalOp().
5567 /// This function expects two 256-bit vectors called V0 and V1.
5568 /// At first, each vector is split into two separate 128-bit vectors.
5569 /// Then, the resulting 128-bit vectors are used to implement two
5570 /// horizontal binary operations.
5572 /// The kind of horizontal binary operation is defined by \p X86Opcode.
5574 /// \p Mode specifies how the 128-bit parts of V0 and V1 are passed in input to
5575 /// the two new horizontal binop.
5576 /// When Mode is set, the first horizontal binop dag node would take as input
5577 /// the lower 128-bit of V0 and the upper 128-bit of V0. The second
5578 /// horizontal binop dag node would take as input the lower 128-bit of V1
5579 /// and the upper 128-bit of V1.
5581 /// HADD V0_LO, V0_HI
5582 /// HADD V1_LO, V1_HI
5584 /// Otherwise, the first horizontal binop dag node takes as input the lower
5585 /// 128-bit of V0 and the lower 128-bit of V1, and the second horizontal binop
5586 /// dag node takes the upper 128-bit of V0 and the upper 128-bit of V1.
5588 /// HADD V0_LO, V1_LO
5589 /// HADD V0_HI, V1_HI
5591 /// If \p isUndefLO is set, then the algorithm propagates UNDEF to the lower
5592 /// 128-bits of the result. If \p isUndefHI is set, then UNDEF is propagated to
5593 /// the upper 128-bits of the result.
5594 static SDValue ExpandHorizontalBinOp(const SDValue &V0, const SDValue &V1,
5595 SDLoc DL, SelectionDAG &DAG,
5596 unsigned X86Opcode, bool Mode,
5597 bool isUndefLO, bool isUndefHI) {
5598 EVT VT = V0.getValueType();
5599 assert(VT.is256BitVector() && VT == V1.getValueType() &&
5600 "Invalid nodes in input!");
5602 unsigned NumElts = VT.getVectorNumElements();
5603 SDValue V0_LO = Extract128BitVector(V0, 0, DAG, DL);
5604 SDValue V0_HI = Extract128BitVector(V0, NumElts/2, DAG, DL);
5605 SDValue V1_LO = Extract128BitVector(V1, 0, DAG, DL);
5606 SDValue V1_HI = Extract128BitVector(V1, NumElts/2, DAG, DL);
5607 EVT NewVT = V0_LO.getValueType();
5609 SDValue LO = DAG.getUNDEF(NewVT);
5610 SDValue HI = DAG.getUNDEF(NewVT);
5613 // Don't emit a horizontal binop if the result is expected to be UNDEF.
5614 if (!isUndefLO && V0->getOpcode() != ISD::UNDEF)
5615 LO = DAG.getNode(X86Opcode, DL, NewVT, V0_LO, V0_HI);
5616 if (!isUndefHI && V1->getOpcode() != ISD::UNDEF)
5617 HI = DAG.getNode(X86Opcode, DL, NewVT, V1_LO, V1_HI);
5619 // Don't emit a horizontal binop if the result is expected to be UNDEF.
5620 if (!isUndefLO && (V0_LO->getOpcode() != ISD::UNDEF ||
5621 V1_LO->getOpcode() != ISD::UNDEF))
5622 LO = DAG.getNode(X86Opcode, DL, NewVT, V0_LO, V1_LO);
5624 if (!isUndefHI && (V0_HI->getOpcode() != ISD::UNDEF ||
5625 V1_HI->getOpcode() != ISD::UNDEF))
5626 HI = DAG.getNode(X86Opcode, DL, NewVT, V0_HI, V1_HI);
5629 return DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, LO, HI);
5632 /// Try to fold a build_vector that performs an 'addsub' to an X86ISD::ADDSUB
5634 static SDValue LowerToAddSub(const BuildVectorSDNode *BV,
5635 const X86Subtarget *Subtarget, SelectionDAG &DAG) {
5636 EVT VT = BV->getValueType(0);
5637 if ((!Subtarget->hasSSE3() || (VT != MVT::v4f32 && VT != MVT::v2f64)) &&
5638 (!Subtarget->hasAVX() || (VT != MVT::v8f32 && VT != MVT::v4f64)))
5642 unsigned NumElts = VT.getVectorNumElements();
5643 SDValue InVec0 = DAG.getUNDEF(VT);
5644 SDValue InVec1 = DAG.getUNDEF(VT);
5646 assert((VT == MVT::v8f32 || VT == MVT::v4f64 || VT == MVT::v4f32 ||
5647 VT == MVT::v2f64) && "build_vector with an invalid type found!");
5649 // Odd-numbered elements in the input build vector are obtained from
5650 // adding two integer/float elements.
5651 // Even-numbered elements in the input build vector are obtained from
5652 // subtracting two integer/float elements.
5653 unsigned ExpectedOpcode = ISD::FSUB;
5654 unsigned NextExpectedOpcode = ISD::FADD;
5655 bool AddFound = false;
5656 bool SubFound = false;
5658 for (unsigned i = 0, e = NumElts; i != e; ++i) {
5659 SDValue Op = BV->getOperand(i);
5661 // Skip 'undef' values.
5662 unsigned Opcode = Op.getOpcode();
5663 if (Opcode == ISD::UNDEF) {
5664 std::swap(ExpectedOpcode, NextExpectedOpcode);
5668 // Early exit if we found an unexpected opcode.
5669 if (Opcode != ExpectedOpcode)
5672 SDValue Op0 = Op.getOperand(0);
5673 SDValue Op1 = Op.getOperand(1);
5675 // Try to match the following pattern:
5676 // (BINOP (extract_vector_elt A, i), (extract_vector_elt B, i))
5677 // Early exit if we cannot match that sequence.
5678 if (Op0.getOpcode() != ISD::EXTRACT_VECTOR_ELT ||
5679 Op1.getOpcode() != ISD::EXTRACT_VECTOR_ELT ||
5680 !isa<ConstantSDNode>(Op0.getOperand(1)) ||
5681 !isa<ConstantSDNode>(Op1.getOperand(1)) ||
5682 Op0.getOperand(1) != Op1.getOperand(1))
5685 unsigned I0 = cast<ConstantSDNode>(Op0.getOperand(1))->getZExtValue();
5689 // We found a valid add/sub node. Update the information accordingly.
5695 // Update InVec0 and InVec1.
5696 if (InVec0.getOpcode() == ISD::UNDEF) {
5697 InVec0 = Op0.getOperand(0);
5698 if (InVec0.getValueType() != VT)
5701 if (InVec1.getOpcode() == ISD::UNDEF) {
5702 InVec1 = Op1.getOperand(0);
5703 if (InVec1.getValueType() != VT)
5707 // Make sure that operands in input to each add/sub node always
5708 // come from a same pair of vectors.
5709 if (InVec0 != Op0.getOperand(0)) {
5710 if (ExpectedOpcode == ISD::FSUB)
5713 // FADD is commutable. Try to commute the operands
5714 // and then test again.
5715 std::swap(Op0, Op1);
5716 if (InVec0 != Op0.getOperand(0))
5720 if (InVec1 != Op1.getOperand(0))
5723 // Update the pair of expected opcodes.
5724 std::swap(ExpectedOpcode, NextExpectedOpcode);
5727 // Don't try to fold this build_vector into an ADDSUB if the inputs are undef.
5728 if (AddFound && SubFound && InVec0.getOpcode() != ISD::UNDEF &&
5729 InVec1.getOpcode() != ISD::UNDEF)
5730 return DAG.getNode(X86ISD::ADDSUB, DL, VT, InVec0, InVec1);
5735 /// Lower BUILD_VECTOR to a horizontal add/sub operation if possible.
5736 static SDValue LowerToHorizontalOp(const BuildVectorSDNode *BV,
5737 const X86Subtarget *Subtarget,
5738 SelectionDAG &DAG) {
5739 EVT VT = BV->getValueType(0);
5740 unsigned NumElts = VT.getVectorNumElements();
5741 unsigned NumUndefsLO = 0;
5742 unsigned NumUndefsHI = 0;
5743 unsigned Half = NumElts/2;
5745 // Count the number of UNDEF operands in the build_vector in input.
5746 for (unsigned i = 0, e = Half; i != e; ++i)
5747 if (BV->getOperand(i)->getOpcode() == ISD::UNDEF)
5750 for (unsigned i = Half, e = NumElts; i != e; ++i)
5751 if (BV->getOperand(i)->getOpcode() == ISD::UNDEF)
5754 // Early exit if this is either a build_vector of all UNDEFs or all the
5755 // operands but one are UNDEF.
5756 if (NumUndefsLO + NumUndefsHI + 1 >= NumElts)
5760 SDValue InVec0, InVec1;
5761 if ((VT == MVT::v4f32 || VT == MVT::v2f64) && Subtarget->hasSSE3()) {
5762 // Try to match an SSE3 float HADD/HSUB.
5763 if (isHorizontalBinOp(BV, ISD::FADD, DAG, 0, NumElts, InVec0, InVec1))
5764 return DAG.getNode(X86ISD::FHADD, DL, VT, InVec0, InVec1);
5766 if (isHorizontalBinOp(BV, ISD::FSUB, DAG, 0, NumElts, InVec0, InVec1))
5767 return DAG.getNode(X86ISD::FHSUB, DL, VT, InVec0, InVec1);
5768 } else if ((VT == MVT::v4i32 || VT == MVT::v8i16) && Subtarget->hasSSSE3()) {
5769 // Try to match an SSSE3 integer HADD/HSUB.
5770 if (isHorizontalBinOp(BV, ISD::ADD, DAG, 0, NumElts, InVec0, InVec1))
5771 return DAG.getNode(X86ISD::HADD, DL, VT, InVec0, InVec1);
5773 if (isHorizontalBinOp(BV, ISD::SUB, DAG, 0, NumElts, InVec0, InVec1))
5774 return DAG.getNode(X86ISD::HSUB, DL, VT, InVec0, InVec1);
5777 if (!Subtarget->hasAVX())
5780 if ((VT == MVT::v8f32 || VT == MVT::v4f64)) {
5781 // Try to match an AVX horizontal add/sub of packed single/double
5782 // precision floating point values from 256-bit vectors.
5783 SDValue InVec2, InVec3;
5784 if (isHorizontalBinOp(BV, ISD::FADD, DAG, 0, Half, InVec0, InVec1) &&
5785 isHorizontalBinOp(BV, ISD::FADD, DAG, Half, NumElts, InVec2, InVec3) &&
5786 ((InVec0.getOpcode() == ISD::UNDEF ||
5787 InVec2.getOpcode() == ISD::UNDEF) || InVec0 == InVec2) &&
5788 ((InVec1.getOpcode() == ISD::UNDEF ||
5789 InVec3.getOpcode() == ISD::UNDEF) || InVec1 == InVec3))
5790 return DAG.getNode(X86ISD::FHADD, DL, VT, InVec0, InVec1);
5792 if (isHorizontalBinOp(BV, ISD::FSUB, DAG, 0, Half, InVec0, InVec1) &&
5793 isHorizontalBinOp(BV, ISD::FSUB, DAG, Half, NumElts, InVec2, InVec3) &&
5794 ((InVec0.getOpcode() == ISD::UNDEF ||
5795 InVec2.getOpcode() == ISD::UNDEF) || InVec0 == InVec2) &&
5796 ((InVec1.getOpcode() == ISD::UNDEF ||
5797 InVec3.getOpcode() == ISD::UNDEF) || InVec1 == InVec3))
5798 return DAG.getNode(X86ISD::FHSUB, DL, VT, InVec0, InVec1);
5799 } else if (VT == MVT::v8i32 || VT == MVT::v16i16) {
5800 // Try to match an AVX2 horizontal add/sub of signed integers.
5801 SDValue InVec2, InVec3;
5803 bool CanFold = true;
5805 if (isHorizontalBinOp(BV, ISD::ADD, DAG, 0, Half, InVec0, InVec1) &&
5806 isHorizontalBinOp(BV, ISD::ADD, DAG, Half, NumElts, InVec2, InVec3) &&
5807 ((InVec0.getOpcode() == ISD::UNDEF ||
5808 InVec2.getOpcode() == ISD::UNDEF) || InVec0 == InVec2) &&
5809 ((InVec1.getOpcode() == ISD::UNDEF ||
5810 InVec3.getOpcode() == ISD::UNDEF) || InVec1 == InVec3))
5811 X86Opcode = X86ISD::HADD;
5812 else if (isHorizontalBinOp(BV, ISD::SUB, DAG, 0, Half, InVec0, InVec1) &&
5813 isHorizontalBinOp(BV, ISD::SUB, DAG, Half, NumElts, InVec2, InVec3) &&
5814 ((InVec0.getOpcode() == ISD::UNDEF ||
5815 InVec2.getOpcode() == ISD::UNDEF) || InVec0 == InVec2) &&
5816 ((InVec1.getOpcode() == ISD::UNDEF ||
5817 InVec3.getOpcode() == ISD::UNDEF) || InVec1 == InVec3))
5818 X86Opcode = X86ISD::HSUB;
5823 // Fold this build_vector into a single horizontal add/sub.
5824 // Do this only if the target has AVX2.
5825 if (Subtarget->hasAVX2())
5826 return DAG.getNode(X86Opcode, DL, VT, InVec0, InVec1);
5828 // Do not try to expand this build_vector into a pair of horizontal
5829 // add/sub if we can emit a pair of scalar add/sub.
5830 if (NumUndefsLO + 1 == Half || NumUndefsHI + 1 == Half)
5833 // Convert this build_vector into a pair of horizontal binop followed by
5835 bool isUndefLO = NumUndefsLO == Half;
5836 bool isUndefHI = NumUndefsHI == Half;
5837 return ExpandHorizontalBinOp(InVec0, InVec1, DL, DAG, X86Opcode, false,
5838 isUndefLO, isUndefHI);
5842 if ((VT == MVT::v8f32 || VT == MVT::v4f64 || VT == MVT::v8i32 ||
5843 VT == MVT::v16i16) && Subtarget->hasAVX()) {
5845 if (isHorizontalBinOp(BV, ISD::ADD, DAG, 0, NumElts, InVec0, InVec1))
5846 X86Opcode = X86ISD::HADD;
5847 else if (isHorizontalBinOp(BV, ISD::SUB, DAG, 0, NumElts, InVec0, InVec1))
5848 X86Opcode = X86ISD::HSUB;
5849 else if (isHorizontalBinOp(BV, ISD::FADD, DAG, 0, NumElts, InVec0, InVec1))
5850 X86Opcode = X86ISD::FHADD;
5851 else if (isHorizontalBinOp(BV, ISD::FSUB, DAG, 0, NumElts, InVec0, InVec1))
5852 X86Opcode = X86ISD::FHSUB;
5856 // Don't try to expand this build_vector into a pair of horizontal add/sub
5857 // if we can simply emit a pair of scalar add/sub.
5858 if (NumUndefsLO + 1 == Half || NumUndefsHI + 1 == Half)
5861 // Convert this build_vector into two horizontal add/sub followed by
5863 bool isUndefLO = NumUndefsLO == Half;
5864 bool isUndefHI = NumUndefsHI == Half;
5865 return ExpandHorizontalBinOp(InVec0, InVec1, DL, DAG, X86Opcode, true,
5866 isUndefLO, isUndefHI);
5873 X86TargetLowering::LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG) const {
5876 MVT VT = Op.getSimpleValueType();
5877 MVT ExtVT = VT.getVectorElementType();
5878 unsigned NumElems = Op.getNumOperands();
5880 // Generate vectors for predicate vectors.
5881 if (VT.getScalarType() == MVT::i1 && Subtarget->hasAVX512())
5882 return LowerBUILD_VECTORvXi1(Op, DAG);
5884 // Vectors containing all zeros can be matched by pxor and xorps later
5885 if (ISD::isBuildVectorAllZeros(Op.getNode())) {
5886 // Canonicalize this to <4 x i32> to 1) ensure the zero vectors are CSE'd
5887 // and 2) ensure that i64 scalars are eliminated on x86-32 hosts.
5888 if (VT == MVT::v4i32 || VT == MVT::v8i32 || VT == MVT::v16i32)
5891 return getZeroVector(VT, Subtarget, DAG, dl);
5894 // Vectors containing all ones can be matched by pcmpeqd on 128-bit width
5895 // vectors or broken into v4i32 operations on 256-bit vectors. AVX2 can use
5896 // vpcmpeqd on 256-bit vectors.
5897 if (Subtarget->hasSSE2() && ISD::isBuildVectorAllOnes(Op.getNode())) {
5898 if (VT == MVT::v4i32 || (VT == MVT::v8i32 && Subtarget->hasInt256()))
5901 if (!VT.is512BitVector())
5902 return getOnesVector(VT, Subtarget->hasInt256(), DAG, dl);
5905 BuildVectorSDNode *BV = cast<BuildVectorSDNode>(Op.getNode());
5906 if (SDValue AddSub = LowerToAddSub(BV, Subtarget, DAG))
5908 if (SDValue HorizontalOp = LowerToHorizontalOp(BV, Subtarget, DAG))
5909 return HorizontalOp;
5910 if (SDValue Broadcast = LowerVectorBroadcast(Op, Subtarget, DAG))
5913 unsigned EVTBits = ExtVT.getSizeInBits();
5915 unsigned NumZero = 0;
5916 unsigned NumNonZero = 0;
5917 unsigned NonZeros = 0;
5918 bool IsAllConstants = true;
5919 SmallSet<SDValue, 8> Values;
5920 for (unsigned i = 0; i < NumElems; ++i) {
5921 SDValue Elt = Op.getOperand(i);
5922 if (Elt.getOpcode() == ISD::UNDEF)
5925 if (Elt.getOpcode() != ISD::Constant &&
5926 Elt.getOpcode() != ISD::ConstantFP)
5927 IsAllConstants = false;
5928 if (X86::isZeroNode(Elt))
5931 NonZeros |= (1 << i);
5936 // All undef vector. Return an UNDEF. All zero vectors were handled above.
5937 if (NumNonZero == 0)
5938 return DAG.getUNDEF(VT);
5940 // Special case for single non-zero, non-undef, element.
5941 if (NumNonZero == 1) {
5942 unsigned Idx = countTrailingZeros(NonZeros);
5943 SDValue Item = Op.getOperand(Idx);
5945 // If this is an insertion of an i64 value on x86-32, and if the top bits of
5946 // the value are obviously zero, truncate the value to i32 and do the
5947 // insertion that way. Only do this if the value is non-constant or if the
5948 // value is a constant being inserted into element 0. It is cheaper to do
5949 // a constant pool load than it is to do a movd + shuffle.
5950 if (ExtVT == MVT::i64 && !Subtarget->is64Bit() &&
5951 (!IsAllConstants || Idx == 0)) {
5952 if (DAG.MaskedValueIsZero(Item, APInt::getBitsSet(64, 32, 64))) {
5954 assert(VT == MVT::v2i64 && "Expected an SSE value type!");
5955 EVT VecVT = MVT::v4i32;
5957 // Truncate the value (which may itself be a constant) to i32, and
5958 // convert it to a vector with movd (S2V+shuffle to zero extend).
5959 Item = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, Item);
5960 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VecVT, Item);
5961 return DAG.getBitcast(VT, getShuffleVectorZeroOrUndef(
5962 Item, Idx * 2, true, Subtarget, DAG));
5966 // If we have a constant or non-constant insertion into the low element of
5967 // a vector, we can do this with SCALAR_TO_VECTOR + shuffle of zero into
5968 // the rest of the elements. This will be matched as movd/movq/movss/movsd
5969 // depending on what the source datatype is.
5972 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
5974 if (ExtVT == MVT::i32 || ExtVT == MVT::f32 || ExtVT == MVT::f64 ||
5975 (ExtVT == MVT::i64 && Subtarget->is64Bit())) {
5976 if (VT.is512BitVector()) {
5977 SDValue ZeroVec = getZeroVector(VT, Subtarget, DAG, dl);
5978 return DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, ZeroVec,
5979 Item, DAG.getIntPtrConstant(0, dl));
5981 assert((VT.is128BitVector() || VT.is256BitVector()) &&
5982 "Expected an SSE value type!");
5983 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
5984 // Turn it into a MOVL (i.e. movss, movsd, or movd) to a zero vector.
5985 return getShuffleVectorZeroOrUndef(Item, 0, true, Subtarget, DAG);
5988 // We can't directly insert an i8 or i16 into a vector, so zero extend
5990 if (ExtVT == MVT::i16 || ExtVT == MVT::i8) {
5991 Item = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, Item);
5992 if (VT.is256BitVector()) {
5993 if (Subtarget->hasAVX()) {
5994 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v8i32, Item);
5995 Item = getShuffleVectorZeroOrUndef(Item, 0, true, Subtarget, DAG);
5997 // Without AVX, we need to extend to a 128-bit vector and then
5998 // insert into the 256-bit vector.
5999 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32, Item);
6000 SDValue ZeroVec = getZeroVector(MVT::v8i32, Subtarget, DAG, dl);
6001 Item = Insert128BitVector(ZeroVec, Item, 0, DAG, dl);
6004 assert(VT.is128BitVector() && "Expected an SSE value type!");
6005 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32, Item);
6006 Item = getShuffleVectorZeroOrUndef(Item, 0, true, Subtarget, DAG);
6008 return DAG.getBitcast(VT, Item);
6012 // Is it a vector logical left shift?
6013 if (NumElems == 2 && Idx == 1 &&
6014 X86::isZeroNode(Op.getOperand(0)) &&
6015 !X86::isZeroNode(Op.getOperand(1))) {
6016 unsigned NumBits = VT.getSizeInBits();
6017 return getVShift(true, VT,
6018 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
6019 VT, Op.getOperand(1)),
6020 NumBits/2, DAG, *this, dl);
6023 if (IsAllConstants) // Otherwise, it's better to do a constpool load.
6026 // Otherwise, if this is a vector with i32 or f32 elements, and the element
6027 // is a non-constant being inserted into an element other than the low one,
6028 // we can't use a constant pool load. Instead, use SCALAR_TO_VECTOR (aka
6029 // movd/movss) to move this into the low element, then shuffle it into
6031 if (EVTBits == 32) {
6032 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
6033 return getShuffleVectorZeroOrUndef(Item, Idx, NumZero > 0, Subtarget, DAG);
6037 // Splat is obviously ok. Let legalizer expand it to a shuffle.
6038 if (Values.size() == 1) {
6039 if (EVTBits == 32) {
6040 // Instead of a shuffle like this:
6041 // shuffle (scalar_to_vector (load (ptr + 4))), undef, <0, 0, 0, 0>
6042 // Check if it's possible to issue this instead.
6043 // shuffle (vload ptr)), undef, <1, 1, 1, 1>
6044 unsigned Idx = countTrailingZeros(NonZeros);
6045 SDValue Item = Op.getOperand(Idx);
6046 if (Op.getNode()->isOnlyUserOf(Item.getNode()))
6047 return LowerAsSplatVectorLoad(Item, VT, dl, DAG);
6052 // A vector full of immediates; various special cases are already
6053 // handled, so this is best done with a single constant-pool load.
6057 // For AVX-length vectors, see if we can use a vector load to get all of the
6058 // elements, otherwise build the individual 128-bit pieces and use
6059 // shuffles to put them in place.
6060 if (VT.is256BitVector() || VT.is512BitVector()) {
6061 SmallVector<SDValue, 64> V(Op->op_begin(), Op->op_begin() + NumElems);
6063 // Check for a build vector of consecutive loads.
6064 if (SDValue LD = EltsFromConsecutiveLoads(VT, V, dl, DAG, false))
6067 EVT HVT = EVT::getVectorVT(*DAG.getContext(), ExtVT, NumElems/2);
6069 // Build both the lower and upper subvector.
6070 SDValue Lower = DAG.getNode(ISD::BUILD_VECTOR, dl, HVT,
6071 makeArrayRef(&V[0], NumElems/2));
6072 SDValue Upper = DAG.getNode(ISD::BUILD_VECTOR, dl, HVT,
6073 makeArrayRef(&V[NumElems / 2], NumElems/2));
6075 // Recreate the wider vector with the lower and upper part.
6076 if (VT.is256BitVector())
6077 return Concat128BitVectors(Lower, Upper, VT, NumElems, DAG, dl);
6078 return Concat256BitVectors(Lower, Upper, VT, NumElems, DAG, dl);
6081 // Let legalizer expand 2-wide build_vectors.
6082 if (EVTBits == 64) {
6083 if (NumNonZero == 1) {
6084 // One half is zero or undef.
6085 unsigned Idx = countTrailingZeros(NonZeros);
6086 SDValue V2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT,
6087 Op.getOperand(Idx));
6088 return getShuffleVectorZeroOrUndef(V2, Idx, true, Subtarget, DAG);
6093 // If element VT is < 32 bits, convert it to inserts into a zero vector.
6094 if (EVTBits == 8 && NumElems == 16)
6095 if (SDValue V = LowerBuildVectorv16i8(Op, NonZeros,NumNonZero,NumZero, DAG,
6099 if (EVTBits == 16 && NumElems == 8)
6100 if (SDValue V = LowerBuildVectorv8i16(Op, NonZeros,NumNonZero,NumZero, DAG,
6104 // If element VT is == 32 bits and has 4 elems, try to generate an INSERTPS
6105 if (EVTBits == 32 && NumElems == 4)
6106 if (SDValue V = LowerBuildVectorv4x32(Op, DAG, Subtarget, *this))
6109 // If element VT is == 32 bits, turn it into a number of shuffles.
6110 SmallVector<SDValue, 8> V(NumElems);
6111 if (NumElems == 4 && NumZero > 0) {
6112 for (unsigned i = 0; i < 4; ++i) {
6113 bool isZero = !(NonZeros & (1 << i));
6115 V[i] = getZeroVector(VT, Subtarget, DAG, dl);
6117 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(i));
6120 for (unsigned i = 0; i < 2; ++i) {
6121 switch ((NonZeros & (0x3 << i*2)) >> (i*2)) {
6124 V[i] = V[i*2]; // Must be a zero vector.
6127 V[i] = getMOVL(DAG, dl, VT, V[i*2+1], V[i*2]);
6130 V[i] = getMOVL(DAG, dl, VT, V[i*2], V[i*2+1]);
6133 V[i] = getUnpackl(DAG, dl, VT, V[i*2], V[i*2+1]);
6138 bool Reverse1 = (NonZeros & 0x3) == 2;
6139 bool Reverse2 = ((NonZeros & (0x3 << 2)) >> 2) == 2;
6143 static_cast<int>(Reverse2 ? NumElems+1 : NumElems),
6144 static_cast<int>(Reverse2 ? NumElems : NumElems+1)
6146 return DAG.getVectorShuffle(VT, dl, V[0], V[1], &MaskVec[0]);
6149 if (Values.size() > 1 && VT.is128BitVector()) {
6150 // Check for a build vector of consecutive loads.
6151 for (unsigned i = 0; i < NumElems; ++i)
6152 V[i] = Op.getOperand(i);
6154 // Check for elements which are consecutive loads.
6155 if (SDValue LD = EltsFromConsecutiveLoads(VT, V, dl, DAG, false))
6158 // Check for a build vector from mostly shuffle plus few inserting.
6159 if (SDValue Sh = buildFromShuffleMostly(Op, DAG))
6162 // For SSE 4.1, use insertps to put the high elements into the low element.
6163 if (Subtarget->hasSSE41()) {
6165 if (Op.getOperand(0).getOpcode() != ISD::UNDEF)
6166 Result = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(0));
6168 Result = DAG.getUNDEF(VT);
6170 for (unsigned i = 1; i < NumElems; ++i) {
6171 if (Op.getOperand(i).getOpcode() == ISD::UNDEF) continue;
6172 Result = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, Result,
6173 Op.getOperand(i), DAG.getIntPtrConstant(i, dl));
6178 // Otherwise, expand into a number of unpckl*, start by extending each of
6179 // our (non-undef) elements to the full vector width with the element in the
6180 // bottom slot of the vector (which generates no code for SSE).
6181 for (unsigned i = 0; i < NumElems; ++i) {
6182 if (Op.getOperand(i).getOpcode() != ISD::UNDEF)
6183 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(i));
6185 V[i] = DAG.getUNDEF(VT);
6188 // Next, we iteratively mix elements, e.g. for v4f32:
6189 // Step 1: unpcklps 0, 2 ==> X: <?, ?, 2, 0>
6190 // : unpcklps 1, 3 ==> Y: <?, ?, 3, 1>
6191 // Step 2: unpcklps X, Y ==> <3, 2, 1, 0>
6192 unsigned EltStride = NumElems >> 1;
6193 while (EltStride != 0) {
6194 for (unsigned i = 0; i < EltStride; ++i) {
6195 // If V[i+EltStride] is undef and this is the first round of mixing,
6196 // then it is safe to just drop this shuffle: V[i] is already in the
6197 // right place, the one element (since it's the first round) being
6198 // inserted as undef can be dropped. This isn't safe for successive
6199 // rounds because they will permute elements within both vectors.
6200 if (V[i+EltStride].getOpcode() == ISD::UNDEF &&
6201 EltStride == NumElems/2)
6204 V[i] = getUnpackl(DAG, dl, VT, V[i], V[i + EltStride]);
6213 // 256-bit AVX can use the vinsertf128 instruction
6214 // to create 256-bit vectors from two other 128-bit ones.
6215 static SDValue LowerAVXCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) {
6217 MVT ResVT = Op.getSimpleValueType();
6219 assert((ResVT.is256BitVector() ||
6220 ResVT.is512BitVector()) && "Value type must be 256-/512-bit wide");
6222 SDValue V1 = Op.getOperand(0);
6223 SDValue V2 = Op.getOperand(1);
6224 unsigned NumElems = ResVT.getVectorNumElements();
6225 if (ResVT.is256BitVector())
6226 return Concat128BitVectors(V1, V2, ResVT, NumElems, DAG, dl);
6228 if (Op.getNumOperands() == 4) {
6229 MVT HalfVT = MVT::getVectorVT(ResVT.getScalarType(),
6230 ResVT.getVectorNumElements()/2);
6231 SDValue V3 = Op.getOperand(2);
6232 SDValue V4 = Op.getOperand(3);
6233 return Concat256BitVectors(Concat128BitVectors(V1, V2, HalfVT, NumElems/2, DAG, dl),
6234 Concat128BitVectors(V3, V4, HalfVT, NumElems/2, DAG, dl), ResVT, NumElems, DAG, dl);
6236 return Concat256BitVectors(V1, V2, ResVT, NumElems, DAG, dl);
6239 static SDValue LowerCONCAT_VECTORSvXi1(SDValue Op,
6240 const X86Subtarget *Subtarget,
6241 SelectionDAG & DAG) {
6243 MVT ResVT = Op.getSimpleValueType();
6244 unsigned NumOfOperands = Op.getNumOperands();
6246 assert(isPowerOf2_32(NumOfOperands) &&
6247 "Unexpected number of operands in CONCAT_VECTORS");
6249 if (NumOfOperands > 2) {
6250 MVT HalfVT = MVT::getVectorVT(ResVT.getScalarType(),
6251 ResVT.getVectorNumElements()/2);
6252 SmallVector<SDValue, 2> Ops;
6253 for (unsigned i = 0; i < NumOfOperands/2; i++)
6254 Ops.push_back(Op.getOperand(i));
6255 SDValue Lo = DAG.getNode(ISD::CONCAT_VECTORS, dl, HalfVT, Ops);
6257 for (unsigned i = NumOfOperands/2; i < NumOfOperands; i++)
6258 Ops.push_back(Op.getOperand(i));
6259 SDValue Hi = DAG.getNode(ISD::CONCAT_VECTORS, dl, HalfVT, Ops);
6260 return DAG.getNode(ISD::CONCAT_VECTORS, dl, ResVT, Lo, Hi);
6263 SDValue V1 = Op.getOperand(0);
6264 SDValue V2 = Op.getOperand(1);
6265 bool IsZeroV1 = ISD::isBuildVectorAllZeros(V1.getNode());
6266 bool IsZeroV2 = ISD::isBuildVectorAllZeros(V2.getNode());
6268 if (IsZeroV1 && IsZeroV2)
6269 return getZeroVector(ResVT, Subtarget, DAG, dl);
6271 SDValue ZeroIdx = DAG.getIntPtrConstant(0, dl);
6272 SDValue Undef = DAG.getUNDEF(ResVT);
6273 unsigned NumElems = ResVT.getVectorNumElements();
6274 SDValue ShiftBits = DAG.getConstant(NumElems/2, dl, MVT::i8);
6276 V2 = DAG.getNode(ISD::INSERT_SUBVECTOR, dl, ResVT, Undef, V2, ZeroIdx);
6277 V2 = DAG.getNode(X86ISD::VSHLI, dl, ResVT, V2, ShiftBits);
6281 V1 = DAG.getNode(ISD::INSERT_SUBVECTOR, dl, ResVT, Undef, V1, ZeroIdx);
6282 // Zero the upper bits of V1
6283 V1 = DAG.getNode(X86ISD::VSHLI, dl, ResVT, V1, ShiftBits);
6284 V1 = DAG.getNode(X86ISD::VSRLI, dl, ResVT, V1, ShiftBits);
6287 return DAG.getNode(ISD::OR, dl, ResVT, V1, V2);
6290 static SDValue LowerCONCAT_VECTORS(SDValue Op,
6291 const X86Subtarget *Subtarget,
6292 SelectionDAG &DAG) {
6293 MVT VT = Op.getSimpleValueType();
6294 if (VT.getVectorElementType() == MVT::i1)
6295 return LowerCONCAT_VECTORSvXi1(Op, Subtarget, DAG);
6297 assert((VT.is256BitVector() && Op.getNumOperands() == 2) ||
6298 (VT.is512BitVector() && (Op.getNumOperands() == 2 ||
6299 Op.getNumOperands() == 4)));
6301 // AVX can use the vinsertf128 instruction to create 256-bit vectors
6302 // from two other 128-bit ones.
6304 // 512-bit vector may contain 2 256-bit vectors or 4 128-bit vectors
6305 return LowerAVXCONCAT_VECTORS(Op, DAG);
6309 //===----------------------------------------------------------------------===//
6310 // Vector shuffle lowering
6312 // This is an experimental code path for lowering vector shuffles on x86. It is
6313 // designed to handle arbitrary vector shuffles and blends, gracefully
6314 // degrading performance as necessary. It works hard to recognize idiomatic
6315 // shuffles and lower them to optimal instruction patterns without leaving
6316 // a framework that allows reasonably efficient handling of all vector shuffle
6318 //===----------------------------------------------------------------------===//
6320 /// \brief Tiny helper function to identify a no-op mask.
6322 /// This is a somewhat boring predicate function. It checks whether the mask
6323 /// array input, which is assumed to be a single-input shuffle mask of the kind
6324 /// used by the X86 shuffle instructions (not a fully general
6325 /// ShuffleVectorSDNode mask) requires any shuffles to occur. Both undef and an
6326 /// in-place shuffle are 'no-op's.
6327 static bool isNoopShuffleMask(ArrayRef<int> Mask) {
6328 for (int i = 0, Size = Mask.size(); i < Size; ++i)
6329 if (Mask[i] != -1 && Mask[i] != i)
6334 /// \brief Helper function to classify a mask as a single-input mask.
6336 /// This isn't a generic single-input test because in the vector shuffle
6337 /// lowering we canonicalize single inputs to be the first input operand. This
6338 /// means we can more quickly test for a single input by only checking whether
6339 /// an input from the second operand exists. We also assume that the size of
6340 /// mask corresponds to the size of the input vectors which isn't true in the
6341 /// fully general case.
6342 static bool isSingleInputShuffleMask(ArrayRef<int> Mask) {
6344 if (M >= (int)Mask.size())
6349 /// \brief Test whether there are elements crossing 128-bit lanes in this
6352 /// X86 divides up its shuffles into in-lane and cross-lane shuffle operations
6353 /// and we routinely test for these.
6354 static bool is128BitLaneCrossingShuffleMask(MVT VT, ArrayRef<int> Mask) {
6355 int LaneSize = 128 / VT.getScalarSizeInBits();
6356 int Size = Mask.size();
6357 for (int i = 0; i < Size; ++i)
6358 if (Mask[i] >= 0 && (Mask[i] % Size) / LaneSize != i / LaneSize)
6363 /// \brief Test whether a shuffle mask is equivalent within each 128-bit lane.
6365 /// This checks a shuffle mask to see if it is performing the same
6366 /// 128-bit lane-relative shuffle in each 128-bit lane. This trivially implies
6367 /// that it is also not lane-crossing. It may however involve a blend from the
6368 /// same lane of a second vector.
6370 /// The specific repeated shuffle mask is populated in \p RepeatedMask, as it is
6371 /// non-trivial to compute in the face of undef lanes. The representation is
6372 /// *not* suitable for use with existing 128-bit shuffles as it will contain
6373 /// entries from both V1 and V2 inputs to the wider mask.
6375 is128BitLaneRepeatedShuffleMask(MVT VT, ArrayRef<int> Mask,
6376 SmallVectorImpl<int> &RepeatedMask) {
6377 int LaneSize = 128 / VT.getScalarSizeInBits();
6378 RepeatedMask.resize(LaneSize, -1);
6379 int Size = Mask.size();
6380 for (int i = 0; i < Size; ++i) {
6383 if ((Mask[i] % Size) / LaneSize != i / LaneSize)
6384 // This entry crosses lanes, so there is no way to model this shuffle.
6387 // Ok, handle the in-lane shuffles by detecting if and when they repeat.
6388 if (RepeatedMask[i % LaneSize] == -1)
6389 // This is the first non-undef entry in this slot of a 128-bit lane.
6390 RepeatedMask[i % LaneSize] =
6391 Mask[i] < Size ? Mask[i] % LaneSize : Mask[i] % LaneSize + Size;
6392 else if (RepeatedMask[i % LaneSize] + (i / LaneSize) * LaneSize != Mask[i])
6393 // Found a mismatch with the repeated mask.
6399 /// \brief Checks whether a shuffle mask is equivalent to an explicit list of
6402 /// This is a fast way to test a shuffle mask against a fixed pattern:
6404 /// if (isShuffleEquivalent(Mask, 3, 2, {1, 0})) { ... }
6406 /// It returns true if the mask is exactly as wide as the argument list, and
6407 /// each element of the mask is either -1 (signifying undef) or the value given
6408 /// in the argument.
6409 static bool isShuffleEquivalent(SDValue V1, SDValue V2, ArrayRef<int> Mask,
6410 ArrayRef<int> ExpectedMask) {
6411 if (Mask.size() != ExpectedMask.size())
6414 int Size = Mask.size();
6416 // If the values are build vectors, we can look through them to find
6417 // equivalent inputs that make the shuffles equivalent.
6418 auto *BV1 = dyn_cast<BuildVectorSDNode>(V1);
6419 auto *BV2 = dyn_cast<BuildVectorSDNode>(V2);
6421 for (int i = 0; i < Size; ++i)
6422 if (Mask[i] != -1 && Mask[i] != ExpectedMask[i]) {
6423 auto *MaskBV = Mask[i] < Size ? BV1 : BV2;
6424 auto *ExpectedBV = ExpectedMask[i] < Size ? BV1 : BV2;
6425 if (!MaskBV || !ExpectedBV ||
6426 MaskBV->getOperand(Mask[i] % Size) !=
6427 ExpectedBV->getOperand(ExpectedMask[i] % Size))
6434 /// \brief Get a 4-lane 8-bit shuffle immediate for a mask.
6436 /// This helper function produces an 8-bit shuffle immediate corresponding to
6437 /// the ubiquitous shuffle encoding scheme used in x86 instructions for
6438 /// shuffling 4 lanes. It can be used with most of the PSHUF instructions for
6441 /// NB: We rely heavily on "undef" masks preserving the input lane.
6442 static SDValue getV4X86ShuffleImm8ForMask(ArrayRef<int> Mask, SDLoc DL,
6443 SelectionDAG &DAG) {
6444 assert(Mask.size() == 4 && "Only 4-lane shuffle masks");
6445 assert(Mask[0] >= -1 && Mask[0] < 4 && "Out of bound mask element!");
6446 assert(Mask[1] >= -1 && Mask[1] < 4 && "Out of bound mask element!");
6447 assert(Mask[2] >= -1 && Mask[2] < 4 && "Out of bound mask element!");
6448 assert(Mask[3] >= -1 && Mask[3] < 4 && "Out of bound mask element!");
6451 Imm |= (Mask[0] == -1 ? 0 : Mask[0]) << 0;
6452 Imm |= (Mask[1] == -1 ? 1 : Mask[1]) << 2;
6453 Imm |= (Mask[2] == -1 ? 2 : Mask[2]) << 4;
6454 Imm |= (Mask[3] == -1 ? 3 : Mask[3]) << 6;
6455 return DAG.getConstant(Imm, DL, MVT::i8);
6458 /// \brief Compute whether each element of a shuffle is zeroable.
6460 /// A "zeroable" vector shuffle element is one which can be lowered to zero.
6461 /// Either it is an undef element in the shuffle mask, the element of the input
6462 /// referenced is undef, or the element of the input referenced is known to be
6463 /// zero. Many x86 shuffles can zero lanes cheaply and we often want to handle
6464 /// as many lanes with this technique as possible to simplify the remaining
6466 static SmallBitVector computeZeroableShuffleElements(ArrayRef<int> Mask,
6467 SDValue V1, SDValue V2) {
6468 SmallBitVector Zeroable(Mask.size(), false);
6470 while (V1.getOpcode() == ISD::BITCAST)
6471 V1 = V1->getOperand(0);
6472 while (V2.getOpcode() == ISD::BITCAST)
6473 V2 = V2->getOperand(0);
6475 bool V1IsZero = ISD::isBuildVectorAllZeros(V1.getNode());
6476 bool V2IsZero = ISD::isBuildVectorAllZeros(V2.getNode());
6478 for (int i = 0, Size = Mask.size(); i < Size; ++i) {
6480 // Handle the easy cases.
6481 if (M < 0 || (M >= 0 && M < Size && V1IsZero) || (M >= Size && V2IsZero)) {
6486 // If this is an index into a build_vector node (which has the same number
6487 // of elements), dig out the input value and use it.
6488 SDValue V = M < Size ? V1 : V2;
6489 if (V.getOpcode() != ISD::BUILD_VECTOR || Size != (int)V.getNumOperands())
6492 SDValue Input = V.getOperand(M % Size);
6493 // The UNDEF opcode check really should be dead code here, but not quite
6494 // worth asserting on (it isn't invalid, just unexpected).
6495 if (Input.getOpcode() == ISD::UNDEF || X86::isZeroNode(Input))
6502 /// \brief Try to emit a bitmask instruction for a shuffle.
6504 /// This handles cases where we can model a blend exactly as a bitmask due to
6505 /// one of the inputs being zeroable.
6506 static SDValue lowerVectorShuffleAsBitMask(SDLoc DL, MVT VT, SDValue V1,
6507 SDValue V2, ArrayRef<int> Mask,
6508 SelectionDAG &DAG) {
6509 MVT EltVT = VT.getScalarType();
6510 int NumEltBits = EltVT.getSizeInBits();
6511 MVT IntEltVT = MVT::getIntegerVT(NumEltBits);
6512 SDValue Zero = DAG.getConstant(0, DL, IntEltVT);
6513 SDValue AllOnes = DAG.getConstant(APInt::getAllOnesValue(NumEltBits), DL,
6515 if (EltVT.isFloatingPoint()) {
6516 Zero = DAG.getBitcast(EltVT, Zero);
6517 AllOnes = DAG.getBitcast(EltVT, AllOnes);
6519 SmallVector<SDValue, 16> VMaskOps(Mask.size(), Zero);
6520 SmallBitVector Zeroable = computeZeroableShuffleElements(Mask, V1, V2);
6522 for (int i = 0, Size = Mask.size(); i < Size; ++i) {
6525 if (Mask[i] % Size != i)
6526 return SDValue(); // Not a blend.
6528 V = Mask[i] < Size ? V1 : V2;
6529 else if (V != (Mask[i] < Size ? V1 : V2))
6530 return SDValue(); // Can only let one input through the mask.
6532 VMaskOps[i] = AllOnes;
6535 return SDValue(); // No non-zeroable elements!
6537 SDValue VMask = DAG.getNode(ISD::BUILD_VECTOR, DL, VT, VMaskOps);
6538 V = DAG.getNode(VT.isFloatingPoint()
6539 ? (unsigned) X86ISD::FAND : (unsigned) ISD::AND,
6544 /// \brief Try to emit a blend instruction for a shuffle using bit math.
6546 /// This is used as a fallback approach when first class blend instructions are
6547 /// unavailable. Currently it is only suitable for integer vectors, but could
6548 /// be generalized for floating point vectors if desirable.
6549 static SDValue lowerVectorShuffleAsBitBlend(SDLoc DL, MVT VT, SDValue V1,
6550 SDValue V2, ArrayRef<int> Mask,
6551 SelectionDAG &DAG) {
6552 assert(VT.isInteger() && "Only supports integer vector types!");
6553 MVT EltVT = VT.getScalarType();
6554 int NumEltBits = EltVT.getSizeInBits();
6555 SDValue Zero = DAG.getConstant(0, DL, EltVT);
6556 SDValue AllOnes = DAG.getConstant(APInt::getAllOnesValue(NumEltBits), DL,
6558 SmallVector<SDValue, 16> MaskOps;
6559 for (int i = 0, Size = Mask.size(); i < Size; ++i) {
6560 if (Mask[i] != -1 && Mask[i] != i && Mask[i] != i + Size)
6561 return SDValue(); // Shuffled input!
6562 MaskOps.push_back(Mask[i] < Size ? AllOnes : Zero);
6565 SDValue V1Mask = DAG.getNode(ISD::BUILD_VECTOR, DL, VT, MaskOps);
6566 V1 = DAG.getNode(ISD::AND, DL, VT, V1, V1Mask);
6567 // We have to cast V2 around.
6568 MVT MaskVT = MVT::getVectorVT(MVT::i64, VT.getSizeInBits() / 64);
6569 V2 = DAG.getBitcast(VT, DAG.getNode(X86ISD::ANDNP, DL, MaskVT,
6570 DAG.getBitcast(MaskVT, V1Mask),
6571 DAG.getBitcast(MaskVT, V2)));
6572 return DAG.getNode(ISD::OR, DL, VT, V1, V2);
6575 /// \brief Try to emit a blend instruction for a shuffle.
6577 /// This doesn't do any checks for the availability of instructions for blending
6578 /// these values. It relies on the availability of the X86ISD::BLENDI pattern to
6579 /// be matched in the backend with the type given. What it does check for is
6580 /// that the shuffle mask is in fact a blend.
6581 static SDValue lowerVectorShuffleAsBlend(SDLoc DL, MVT VT, SDValue V1,
6582 SDValue V2, ArrayRef<int> Mask,
6583 const X86Subtarget *Subtarget,
6584 SelectionDAG &DAG) {
6585 unsigned BlendMask = 0;
6586 for (int i = 0, Size = Mask.size(); i < Size; ++i) {
6587 if (Mask[i] >= Size) {
6588 if (Mask[i] != i + Size)
6589 return SDValue(); // Shuffled V2 input!
6590 BlendMask |= 1u << i;
6593 if (Mask[i] >= 0 && Mask[i] != i)
6594 return SDValue(); // Shuffled V1 input!
6596 switch (VT.SimpleTy) {
6601 return DAG.getNode(X86ISD::BLENDI, DL, VT, V1, V2,
6602 DAG.getConstant(BlendMask, DL, MVT::i8));
6606 assert(Subtarget->hasAVX2() && "256-bit integer blends require AVX2!");
6610 // If we have AVX2 it is faster to use VPBLENDD when the shuffle fits into
6611 // that instruction.
6612 if (Subtarget->hasAVX2()) {
6613 // Scale the blend by the number of 32-bit dwords per element.
6614 int Scale = VT.getScalarSizeInBits() / 32;
6616 for (int i = 0, Size = Mask.size(); i < Size; ++i)
6617 if (Mask[i] >= Size)
6618 for (int j = 0; j < Scale; ++j)
6619 BlendMask |= 1u << (i * Scale + j);
6621 MVT BlendVT = VT.getSizeInBits() > 128 ? MVT::v8i32 : MVT::v4i32;
6622 V1 = DAG.getBitcast(BlendVT, V1);
6623 V2 = DAG.getBitcast(BlendVT, V2);
6624 return DAG.getBitcast(
6625 VT, DAG.getNode(X86ISD::BLENDI, DL, BlendVT, V1, V2,
6626 DAG.getConstant(BlendMask, DL, MVT::i8)));
6630 // For integer shuffles we need to expand the mask and cast the inputs to
6631 // v8i16s prior to blending.
6632 int Scale = 8 / VT.getVectorNumElements();
6634 for (int i = 0, Size = Mask.size(); i < Size; ++i)
6635 if (Mask[i] >= Size)
6636 for (int j = 0; j < Scale; ++j)
6637 BlendMask |= 1u << (i * Scale + j);
6639 V1 = DAG.getBitcast(MVT::v8i16, V1);
6640 V2 = DAG.getBitcast(MVT::v8i16, V2);
6641 return DAG.getBitcast(VT,
6642 DAG.getNode(X86ISD::BLENDI, DL, MVT::v8i16, V1, V2,
6643 DAG.getConstant(BlendMask, DL, MVT::i8)));
6647 assert(Subtarget->hasAVX2() && "256-bit integer blends require AVX2!");
6648 SmallVector<int, 8> RepeatedMask;
6649 if (is128BitLaneRepeatedShuffleMask(MVT::v16i16, Mask, RepeatedMask)) {
6650 // We can lower these with PBLENDW which is mirrored across 128-bit lanes.
6651 assert(RepeatedMask.size() == 8 && "Repeated mask size doesn't match!");
6653 for (int i = 0; i < 8; ++i)
6654 if (RepeatedMask[i] >= 16)
6655 BlendMask |= 1u << i;
6656 return DAG.getNode(X86ISD::BLENDI, DL, MVT::v16i16, V1, V2,
6657 DAG.getConstant(BlendMask, DL, MVT::i8));
6663 assert((VT.getSizeInBits() == 128 || Subtarget->hasAVX2()) &&
6664 "256-bit byte-blends require AVX2 support!");
6666 // Scale the blend by the number of bytes per element.
6667 int Scale = VT.getScalarSizeInBits() / 8;
6669 // This form of blend is always done on bytes. Compute the byte vector
6671 MVT BlendVT = MVT::getVectorVT(MVT::i8, VT.getSizeInBits() / 8);
6673 // Compute the VSELECT mask. Note that VSELECT is really confusing in the
6674 // mix of LLVM's code generator and the x86 backend. We tell the code
6675 // generator that boolean values in the elements of an x86 vector register
6676 // are -1 for true and 0 for false. We then use the LLVM semantics of 'true'
6677 // mapping a select to operand #1, and 'false' mapping to operand #2. The
6678 // reality in x86 is that vector masks (pre-AVX-512) use only the high bit
6679 // of the element (the remaining are ignored) and 0 in that high bit would
6680 // mean operand #1 while 1 in the high bit would mean operand #2. So while
6681 // the LLVM model for boolean values in vector elements gets the relevant
6682 // bit set, it is set backwards and over constrained relative to x86's
6684 SmallVector<SDValue, 32> VSELECTMask;
6685 for (int i = 0, Size = Mask.size(); i < Size; ++i)
6686 for (int j = 0; j < Scale; ++j)
6687 VSELECTMask.push_back(
6688 Mask[i] < 0 ? DAG.getUNDEF(MVT::i8)
6689 : DAG.getConstant(Mask[i] < Size ? -1 : 0, DL,
6692 V1 = DAG.getBitcast(BlendVT, V1);
6693 V2 = DAG.getBitcast(BlendVT, V2);
6694 return DAG.getBitcast(VT, DAG.getNode(ISD::VSELECT, DL, BlendVT,
6695 DAG.getNode(ISD::BUILD_VECTOR, DL,
6696 BlendVT, VSELECTMask),
6701 llvm_unreachable("Not a supported integer vector type!");
6705 /// \brief Try to lower as a blend of elements from two inputs followed by
6706 /// a single-input permutation.
6708 /// This matches the pattern where we can blend elements from two inputs and
6709 /// then reduce the shuffle to a single-input permutation.
6710 static SDValue lowerVectorShuffleAsBlendAndPermute(SDLoc DL, MVT VT, SDValue V1,
6713 SelectionDAG &DAG) {
6714 // We build up the blend mask while checking whether a blend is a viable way
6715 // to reduce the shuffle.
6716 SmallVector<int, 32> BlendMask(Mask.size(), -1);
6717 SmallVector<int, 32> PermuteMask(Mask.size(), -1);
6719 for (int i = 0, Size = Mask.size(); i < Size; ++i) {
6723 assert(Mask[i] < Size * 2 && "Shuffle input is out of bounds.");
6725 if (BlendMask[Mask[i] % Size] == -1)
6726 BlendMask[Mask[i] % Size] = Mask[i];
6727 else if (BlendMask[Mask[i] % Size] != Mask[i])
6728 return SDValue(); // Can't blend in the needed input!
6730 PermuteMask[i] = Mask[i] % Size;
6733 SDValue V = DAG.getVectorShuffle(VT, DL, V1, V2, BlendMask);
6734 return DAG.getVectorShuffle(VT, DL, V, DAG.getUNDEF(VT), PermuteMask);
6737 /// \brief Generic routine to decompose a shuffle and blend into indepndent
6738 /// blends and permutes.
6740 /// This matches the extremely common pattern for handling combined
6741 /// shuffle+blend operations on newer X86 ISAs where we have very fast blend
6742 /// operations. It will try to pick the best arrangement of shuffles and
6744 static SDValue lowerVectorShuffleAsDecomposedShuffleBlend(SDLoc DL, MVT VT,
6748 SelectionDAG &DAG) {
6749 // Shuffle the input elements into the desired positions in V1 and V2 and
6750 // blend them together.
6751 SmallVector<int, 32> V1Mask(Mask.size(), -1);
6752 SmallVector<int, 32> V2Mask(Mask.size(), -1);
6753 SmallVector<int, 32> BlendMask(Mask.size(), -1);
6754 for (int i = 0, Size = Mask.size(); i < Size; ++i)
6755 if (Mask[i] >= 0 && Mask[i] < Size) {
6756 V1Mask[i] = Mask[i];
6758 } else if (Mask[i] >= Size) {
6759 V2Mask[i] = Mask[i] - Size;
6760 BlendMask[i] = i + Size;
6763 // Try to lower with the simpler initial blend strategy unless one of the
6764 // input shuffles would be a no-op. We prefer to shuffle inputs as the
6765 // shuffle may be able to fold with a load or other benefit. However, when
6766 // we'll have to do 2x as many shuffles in order to achieve this, blending
6767 // first is a better strategy.
6768 if (!isNoopShuffleMask(V1Mask) && !isNoopShuffleMask(V2Mask))
6769 if (SDValue BlendPerm =
6770 lowerVectorShuffleAsBlendAndPermute(DL, VT, V1, V2, Mask, DAG))
6773 V1 = DAG.getVectorShuffle(VT, DL, V1, DAG.getUNDEF(VT), V1Mask);
6774 V2 = DAG.getVectorShuffle(VT, DL, V2, DAG.getUNDEF(VT), V2Mask);
6775 return DAG.getVectorShuffle(VT, DL, V1, V2, BlendMask);
6778 /// \brief Try to lower a vector shuffle as a byte rotation.
6780 /// SSSE3 has a generic PALIGNR instruction in x86 that will do an arbitrary
6781 /// byte-rotation of the concatenation of two vectors; pre-SSSE3 can use
6782 /// a PSRLDQ/PSLLDQ/POR pattern to get a similar effect. This routine will
6783 /// try to generically lower a vector shuffle through such an pattern. It
6784 /// does not check for the profitability of lowering either as PALIGNR or
6785 /// PSRLDQ/PSLLDQ/POR, only whether the mask is valid to lower in that form.
6786 /// This matches shuffle vectors that look like:
6788 /// v8i16 [11, 12, 13, 14, 15, 0, 1, 2]
6790 /// Essentially it concatenates V1 and V2, shifts right by some number of
6791 /// elements, and takes the low elements as the result. Note that while this is
6792 /// specified as a *right shift* because x86 is little-endian, it is a *left
6793 /// rotate* of the vector lanes.
6794 static SDValue lowerVectorShuffleAsByteRotate(SDLoc DL, MVT VT, SDValue V1,
6797 const X86Subtarget *Subtarget,
6798 SelectionDAG &DAG) {
6799 assert(!isNoopShuffleMask(Mask) && "We shouldn't lower no-op shuffles!");
6801 int NumElts = Mask.size();
6802 int NumLanes = VT.getSizeInBits() / 128;
6803 int NumLaneElts = NumElts / NumLanes;
6805 // We need to detect various ways of spelling a rotation:
6806 // [11, 12, 13, 14, 15, 0, 1, 2]
6807 // [-1, 12, 13, 14, -1, -1, 1, -1]
6808 // [-1, -1, -1, -1, -1, -1, 1, 2]
6809 // [ 3, 4, 5, 6, 7, 8, 9, 10]
6810 // [-1, 4, 5, 6, -1, -1, 9, -1]
6811 // [-1, 4, 5, 6, -1, -1, -1, -1]
6814 for (int l = 0; l < NumElts; l += NumLaneElts) {
6815 for (int i = 0; i < NumLaneElts; ++i) {
6816 if (Mask[l + i] == -1)
6818 assert(Mask[l + i] >= 0 && "Only -1 is a valid negative mask element!");
6820 // Get the mod-Size index and lane correct it.
6821 int LaneIdx = (Mask[l + i] % NumElts) - l;
6822 // Make sure it was in this lane.
6823 if (LaneIdx < 0 || LaneIdx >= NumLaneElts)
6826 // Determine where a rotated vector would have started.
6827 int StartIdx = i - LaneIdx;
6829 // The identity rotation isn't interesting, stop.
6832 // If we found the tail of a vector the rotation must be the missing
6833 // front. If we found the head of a vector, it must be how much of the
6835 int CandidateRotation = StartIdx < 0 ? -StartIdx : NumLaneElts - StartIdx;
6838 Rotation = CandidateRotation;
6839 else if (Rotation != CandidateRotation)
6840 // The rotations don't match, so we can't match this mask.
6843 // Compute which value this mask is pointing at.
6844 SDValue MaskV = Mask[l + i] < NumElts ? V1 : V2;
6846 // Compute which of the two target values this index should be assigned
6847 // to. This reflects whether the high elements are remaining or the low
6848 // elements are remaining.
6849 SDValue &TargetV = StartIdx < 0 ? Hi : Lo;
6851 // Either set up this value if we've not encountered it before, or check
6852 // that it remains consistent.
6855 else if (TargetV != MaskV)
6856 // This may be a rotation, but it pulls from the inputs in some
6857 // unsupported interleaving.
6862 // Check that we successfully analyzed the mask, and normalize the results.
6863 assert(Rotation != 0 && "Failed to locate a viable rotation!");
6864 assert((Lo || Hi) && "Failed to find a rotated input vector!");
6870 // The actual rotate instruction rotates bytes, so we need to scale the
6871 // rotation based on how many bytes are in the vector lane.
6872 int Scale = 16 / NumLaneElts;
6874 // SSSE3 targets can use the palignr instruction.
6875 if (Subtarget->hasSSSE3()) {
6876 // Cast the inputs to i8 vector of correct length to match PALIGNR.
6877 MVT AlignVT = MVT::getVectorVT(MVT::i8, 16 * NumLanes);
6878 Lo = DAG.getBitcast(AlignVT, Lo);
6879 Hi = DAG.getBitcast(AlignVT, Hi);
6881 return DAG.getBitcast(
6882 VT, DAG.getNode(X86ISD::PALIGNR, DL, AlignVT, Hi, Lo,
6883 DAG.getConstant(Rotation * Scale, DL, MVT::i8)));
6886 assert(VT.getSizeInBits() == 128 &&
6887 "Rotate-based lowering only supports 128-bit lowering!");
6888 assert(Mask.size() <= 16 &&
6889 "Can shuffle at most 16 bytes in a 128-bit vector!");
6891 // Default SSE2 implementation
6892 int LoByteShift = 16 - Rotation * Scale;
6893 int HiByteShift = Rotation * Scale;
6895 // Cast the inputs to v2i64 to match PSLLDQ/PSRLDQ.
6896 Lo = DAG.getBitcast(MVT::v2i64, Lo);
6897 Hi = DAG.getBitcast(MVT::v2i64, Hi);
6899 SDValue LoShift = DAG.getNode(X86ISD::VSHLDQ, DL, MVT::v2i64, Lo,
6900 DAG.getConstant(LoByteShift, DL, MVT::i8));
6901 SDValue HiShift = DAG.getNode(X86ISD::VSRLDQ, DL, MVT::v2i64, Hi,
6902 DAG.getConstant(HiByteShift, DL, MVT::i8));
6903 return DAG.getBitcast(VT,
6904 DAG.getNode(ISD::OR, DL, MVT::v2i64, LoShift, HiShift));
6907 /// \brief Try to lower a vector shuffle as a bit shift (shifts in zeros).
6909 /// Attempts to match a shuffle mask against the PSLL(W/D/Q/DQ) and
6910 /// PSRL(W/D/Q/DQ) SSE2 and AVX2 logical bit-shift instructions. The function
6911 /// matches elements from one of the input vectors shuffled to the left or
6912 /// right with zeroable elements 'shifted in'. It handles both the strictly
6913 /// bit-wise element shifts and the byte shift across an entire 128-bit double
6916 /// PSHL : (little-endian) left bit shift.
6917 /// [ zz, 0, zz, 2 ]
6918 /// [ -1, 4, zz, -1 ]
6919 /// PSRL : (little-endian) right bit shift.
6921 /// [ -1, -1, 7, zz]
6922 /// PSLLDQ : (little-endian) left byte shift
6923 /// [ zz, 0, 1, 2, 3, 4, 5, 6]
6924 /// [ zz, zz, -1, -1, 2, 3, 4, -1]
6925 /// [ zz, zz, zz, zz, zz, zz, -1, 1]
6926 /// PSRLDQ : (little-endian) right byte shift
6927 /// [ 5, 6, 7, zz, zz, zz, zz, zz]
6928 /// [ -1, 5, 6, 7, zz, zz, zz, zz]
6929 /// [ 1, 2, -1, -1, -1, -1, zz, zz]
6930 static SDValue lowerVectorShuffleAsShift(SDLoc DL, MVT VT, SDValue V1,
6931 SDValue V2, ArrayRef<int> Mask,
6932 SelectionDAG &DAG) {
6933 SmallBitVector Zeroable = computeZeroableShuffleElements(Mask, V1, V2);
6935 int Size = Mask.size();
6936 assert(Size == (int)VT.getVectorNumElements() && "Unexpected mask size");
6938 auto CheckZeros = [&](int Shift, int Scale, bool Left) {
6939 for (int i = 0; i < Size; i += Scale)
6940 for (int j = 0; j < Shift; ++j)
6941 if (!Zeroable[i + j + (Left ? 0 : (Scale - Shift))])
6947 auto MatchShift = [&](int Shift, int Scale, bool Left, SDValue V) {
6948 for (int i = 0; i != Size; i += Scale) {
6949 unsigned Pos = Left ? i + Shift : i;
6950 unsigned Low = Left ? i : i + Shift;
6951 unsigned Len = Scale - Shift;
6952 if (!isSequentialOrUndefInRange(Mask, Pos, Len,
6953 Low + (V == V1 ? 0 : Size)))
6957 int ShiftEltBits = VT.getScalarSizeInBits() * Scale;
6958 bool ByteShift = ShiftEltBits > 64;
6959 unsigned OpCode = Left ? (ByteShift ? X86ISD::VSHLDQ : X86ISD::VSHLI)
6960 : (ByteShift ? X86ISD::VSRLDQ : X86ISD::VSRLI);
6961 int ShiftAmt = Shift * VT.getScalarSizeInBits() / (ByteShift ? 8 : 1);
6963 // Normalize the scale for byte shifts to still produce an i64 element
6965 Scale = ByteShift ? Scale / 2 : Scale;
6967 // We need to round trip through the appropriate type for the shift.
6968 MVT ShiftSVT = MVT::getIntegerVT(VT.getScalarSizeInBits() * Scale);
6969 MVT ShiftVT = MVT::getVectorVT(ShiftSVT, Size / Scale);
6970 assert(DAG.getTargetLoweringInfo().isTypeLegal(ShiftVT) &&
6971 "Illegal integer vector type");
6972 V = DAG.getBitcast(ShiftVT, V);
6974 V = DAG.getNode(OpCode, DL, ShiftVT, V,
6975 DAG.getConstant(ShiftAmt, DL, MVT::i8));
6976 return DAG.getBitcast(VT, V);
6979 // SSE/AVX supports logical shifts up to 64-bit integers - so we can just
6980 // keep doubling the size of the integer elements up to that. We can
6981 // then shift the elements of the integer vector by whole multiples of
6982 // their width within the elements of the larger integer vector. Test each
6983 // multiple to see if we can find a match with the moved element indices
6984 // and that the shifted in elements are all zeroable.
6985 for (int Scale = 2; Scale * VT.getScalarSizeInBits() <= 128; Scale *= 2)
6986 for (int Shift = 1; Shift != Scale; ++Shift)
6987 for (bool Left : {true, false})
6988 if (CheckZeros(Shift, Scale, Left))
6989 for (SDValue V : {V1, V2})
6990 if (SDValue Match = MatchShift(Shift, Scale, Left, V))
6997 /// \brief Try to lower a vector shuffle using SSE4a EXTRQ/INSERTQ.
6998 static SDValue lowerVectorShuffleWithSSE4A(SDLoc DL, MVT VT, SDValue V1,
6999 SDValue V2, ArrayRef<int> Mask,
7000 SelectionDAG &DAG) {
7001 SmallBitVector Zeroable = computeZeroableShuffleElements(Mask, V1, V2);
7002 assert(!Zeroable.all() && "Fully zeroable shuffle mask");
7004 int Size = Mask.size();
7005 int HalfSize = Size / 2;
7006 assert(Size == (int)VT.getVectorNumElements() && "Unexpected mask size");
7008 // Upper half must be undefined.
7009 if (!isUndefInRange(Mask, HalfSize, HalfSize))
7012 // EXTRQ: Extract Len elements from lower half of source, starting at Idx.
7013 // Remainder of lower half result is zero and upper half is all undef.
7014 auto LowerAsEXTRQ = [&]() {
7015 // Determine the extraction length from the part of the
7016 // lower half that isn't zeroable.
7018 for (; Len >= 0; --Len)
7019 if (!Zeroable[Len - 1])
7021 assert(Len > 0 && "Zeroable shuffle mask");
7023 // Attempt to match first Len sequential elements from the lower half.
7026 for (int i = 0; i != Len; ++i) {
7030 SDValue &V = (M < Size ? V1 : V2);
7033 // All mask elements must be in the lower half.
7037 if (Idx < 0 || (Src == V && Idx == (M - i))) {
7048 assert((Idx + Len) <= HalfSize && "Illegal extraction mask");
7049 int BitLen = (Len * VT.getScalarSizeInBits()) & 0x3f;
7050 int BitIdx = (Idx * VT.getScalarSizeInBits()) & 0x3f;
7051 return DAG.getNode(X86ISD::EXTRQI, DL, VT, Src,
7052 DAG.getConstant(BitLen, DL, MVT::i8),
7053 DAG.getConstant(BitIdx, DL, MVT::i8));
7056 if (SDValue ExtrQ = LowerAsEXTRQ())
7059 // INSERTQ: Extract lowest Len elements from lower half of second source and
7060 // insert over first source, starting at Idx.
7061 // { A[0], .., A[Idx-1], B[0], .., B[Len-1], A[Idx+Len], .., UNDEF, ... }
7062 auto LowerAsInsertQ = [&]() {
7063 for (int Idx = 0; Idx != HalfSize; ++Idx) {
7066 // Attempt to match first source from mask before insertion point.
7067 if (isUndefInRange(Mask, 0, Idx)) {
7069 } else if (isSequentialOrUndefInRange(Mask, 0, Idx, 0)) {
7071 } else if (isSequentialOrUndefInRange(Mask, 0, Idx, Size)) {
7077 // Extend the extraction length looking to match both the insertion of
7078 // the second source and the remaining elements of the first.
7079 for (int Hi = Idx + 1; Hi <= HalfSize; ++Hi) {
7084 if (isSequentialOrUndefInRange(Mask, Idx, Len, 0)) {
7086 } else if (isSequentialOrUndefInRange(Mask, Idx, Len, Size)) {
7092 // Match the remaining elements of the lower half.
7093 if (isUndefInRange(Mask, Hi, HalfSize - Hi)) {
7095 } else if ((!Base || (Base == V1)) &&
7096 isSequentialOrUndefInRange(Mask, Hi, HalfSize - Hi, Hi)) {
7098 } else if ((!Base || (Base == V2)) &&
7099 isSequentialOrUndefInRange(Mask, Hi, HalfSize - Hi,
7106 // We may not have a base (first source) - this can safely be undefined.
7108 Base = DAG.getUNDEF(VT);
7110 int BitLen = (Len * VT.getScalarSizeInBits()) & 0x3f;
7111 int BitIdx = (Idx * VT.getScalarSizeInBits()) & 0x3f;
7112 return DAG.getNode(X86ISD::INSERTQI, DL, VT, Base, Insert,
7113 DAG.getConstant(BitLen, DL, MVT::i8),
7114 DAG.getConstant(BitIdx, DL, MVT::i8));
7121 if (SDValue InsertQ = LowerAsInsertQ())
7127 /// \brief Lower a vector shuffle as a zero or any extension.
7129 /// Given a specific number of elements, element bit width, and extension
7130 /// stride, produce either a zero or any extension based on the available
7131 /// features of the subtarget.
7132 static SDValue lowerVectorShuffleAsSpecificZeroOrAnyExtend(
7133 SDLoc DL, MVT VT, int Scale, bool AnyExt, SDValue InputV,
7134 ArrayRef<int> Mask, const X86Subtarget *Subtarget, SelectionDAG &DAG) {
7135 assert(Scale > 1 && "Need a scale to extend.");
7136 int NumElements = VT.getVectorNumElements();
7137 int EltBits = VT.getScalarSizeInBits();
7138 assert((EltBits == 8 || EltBits == 16 || EltBits == 32) &&
7139 "Only 8, 16, and 32 bit elements can be extended.");
7140 assert(Scale * EltBits <= 64 && "Cannot zero extend past 64 bits.");
7142 // Found a valid zext mask! Try various lowering strategies based on the
7143 // input type and available ISA extensions.
7144 if (Subtarget->hasSSE41()) {
7145 MVT ExtVT = MVT::getVectorVT(MVT::getIntegerVT(EltBits * Scale),
7146 NumElements / Scale);
7147 return DAG.getBitcast(VT, DAG.getNode(X86ISD::VZEXT, DL, ExtVT, InputV));
7150 // For any extends we can cheat for larger element sizes and use shuffle
7151 // instructions that can fold with a load and/or copy.
7152 if (AnyExt && EltBits == 32) {
7153 int PSHUFDMask[4] = {0, -1, 1, -1};
7154 return DAG.getBitcast(
7155 VT, DAG.getNode(X86ISD::PSHUFD, DL, MVT::v4i32,
7156 DAG.getBitcast(MVT::v4i32, InputV),
7157 getV4X86ShuffleImm8ForMask(PSHUFDMask, DL, DAG)));
7159 if (AnyExt && EltBits == 16 && Scale > 2) {
7160 int PSHUFDMask[4] = {0, -1, 0, -1};
7161 InputV = DAG.getNode(X86ISD::PSHUFD, DL, MVT::v4i32,
7162 DAG.getBitcast(MVT::v4i32, InputV),
7163 getV4X86ShuffleImm8ForMask(PSHUFDMask, DL, DAG));
7164 int PSHUFHWMask[4] = {1, -1, -1, -1};
7165 return DAG.getBitcast(
7166 VT, DAG.getNode(X86ISD::PSHUFHW, DL, MVT::v8i16,
7167 DAG.getBitcast(MVT::v8i16, InputV),
7168 getV4X86ShuffleImm8ForMask(PSHUFHWMask, DL, DAG)));
7171 // The SSE4A EXTRQ instruction can efficiently extend the first 2 lanes
7173 if ((Scale * EltBits) == 64 && EltBits < 32 && Subtarget->hasSSE4A()) {
7174 assert(NumElements == (int)Mask.size() && "Unexpected shuffle mask size!");
7175 assert(VT.getSizeInBits() == 128 && "Unexpected vector width!");
7177 SDValue Lo = DAG.getNode(ISD::BITCAST, DL, MVT::v2i64,
7178 DAG.getNode(X86ISD::EXTRQI, DL, VT, InputV,
7179 DAG.getConstant(EltBits, DL, MVT::i8),
7180 DAG.getConstant(0, DL, MVT::i8)));
7181 if (isUndefInRange(Mask, NumElements/2, NumElements/2))
7182 return DAG.getNode(ISD::BITCAST, DL, VT, Lo);
7185 DAG.getNode(ISD::BITCAST, DL, MVT::v2i64,
7186 DAG.getNode(X86ISD::EXTRQI, DL, VT, InputV,
7187 DAG.getConstant(EltBits, DL, MVT::i8),
7188 DAG.getConstant(EltBits, DL, MVT::i8)));
7189 return DAG.getNode(ISD::BITCAST, DL, VT,
7190 DAG.getNode(X86ISD::UNPCKL, DL, MVT::v2i64, Lo, Hi));
7193 // If this would require more than 2 unpack instructions to expand, use
7194 // pshufb when available. We can only use more than 2 unpack instructions
7195 // when zero extending i8 elements which also makes it easier to use pshufb.
7196 if (Scale > 4 && EltBits == 8 && Subtarget->hasSSSE3()) {
7197 assert(NumElements == 16 && "Unexpected byte vector width!");
7198 SDValue PSHUFBMask[16];
7199 for (int i = 0; i < 16; ++i)
7201 DAG.getConstant((i % Scale == 0) ? i / Scale : 0x80, DL, MVT::i8);
7202 InputV = DAG.getBitcast(MVT::v16i8, InputV);
7203 return DAG.getBitcast(VT,
7204 DAG.getNode(X86ISD::PSHUFB, DL, MVT::v16i8, InputV,
7205 DAG.getNode(ISD::BUILD_VECTOR, DL,
7206 MVT::v16i8, PSHUFBMask)));
7209 // Otherwise emit a sequence of unpacks.
7211 MVT InputVT = MVT::getVectorVT(MVT::getIntegerVT(EltBits), NumElements);
7212 SDValue Ext = AnyExt ? DAG.getUNDEF(InputVT)
7213 : getZeroVector(InputVT, Subtarget, DAG, DL);
7214 InputV = DAG.getBitcast(InputVT, InputV);
7215 InputV = DAG.getNode(X86ISD::UNPCKL, DL, InputVT, InputV, Ext);
7219 } while (Scale > 1);
7220 return DAG.getBitcast(VT, InputV);
7223 /// \brief Try to lower a vector shuffle as a zero extension on any microarch.
7225 /// This routine will try to do everything in its power to cleverly lower
7226 /// a shuffle which happens to match the pattern of a zero extend. It doesn't
7227 /// check for the profitability of this lowering, it tries to aggressively
7228 /// match this pattern. It will use all of the micro-architectural details it
7229 /// can to emit an efficient lowering. It handles both blends with all-zero
7230 /// inputs to explicitly zero-extend and undef-lanes (sometimes undef due to
7231 /// masking out later).
7233 /// The reason we have dedicated lowering for zext-style shuffles is that they
7234 /// are both incredibly common and often quite performance sensitive.
7235 static SDValue lowerVectorShuffleAsZeroOrAnyExtend(
7236 SDLoc DL, MVT VT, SDValue V1, SDValue V2, ArrayRef<int> Mask,
7237 const X86Subtarget *Subtarget, SelectionDAG &DAG) {
7238 SmallBitVector Zeroable = computeZeroableShuffleElements(Mask, V1, V2);
7240 int Bits = VT.getSizeInBits();
7241 int NumElements = VT.getVectorNumElements();
7242 assert(VT.getScalarSizeInBits() <= 32 &&
7243 "Exceeds 32-bit integer zero extension limit");
7244 assert((int)Mask.size() == NumElements && "Unexpected shuffle mask size");
7246 // Define a helper function to check a particular ext-scale and lower to it if
7248 auto Lower = [&](int Scale) -> SDValue {
7251 for (int i = 0; i < NumElements; ++i) {
7253 continue; // Valid anywhere but doesn't tell us anything.
7254 if (i % Scale != 0) {
7255 // Each of the extended elements need to be zeroable.
7259 // We no longer are in the anyext case.
7264 // Each of the base elements needs to be consecutive indices into the
7265 // same input vector.
7266 SDValue V = Mask[i] < NumElements ? V1 : V2;
7269 else if (InputV != V)
7270 return SDValue(); // Flip-flopping inputs.
7272 if (Mask[i] % NumElements != i / Scale)
7273 return SDValue(); // Non-consecutive strided elements.
7276 // If we fail to find an input, we have a zero-shuffle which should always
7277 // have already been handled.
7278 // FIXME: Maybe handle this here in case during blending we end up with one?
7282 return lowerVectorShuffleAsSpecificZeroOrAnyExtend(
7283 DL, VT, Scale, AnyExt, InputV, Mask, Subtarget, DAG);
7286 // The widest scale possible for extending is to a 64-bit integer.
7287 assert(Bits % 64 == 0 &&
7288 "The number of bits in a vector must be divisible by 64 on x86!");
7289 int NumExtElements = Bits / 64;
7291 // Each iteration, try extending the elements half as much, but into twice as
7293 for (; NumExtElements < NumElements; NumExtElements *= 2) {
7294 assert(NumElements % NumExtElements == 0 &&
7295 "The input vector size must be divisible by the extended size.");
7296 if (SDValue V = Lower(NumElements / NumExtElements))
7300 // General extends failed, but 128-bit vectors may be able to use MOVQ.
7304 // Returns one of the source operands if the shuffle can be reduced to a
7305 // MOVQ, copying the lower 64-bits and zero-extending to the upper 64-bits.
7306 auto CanZExtLowHalf = [&]() {
7307 for (int i = NumElements / 2; i != NumElements; ++i)
7310 if (isSequentialOrUndefInRange(Mask, 0, NumElements / 2, 0))
7312 if (isSequentialOrUndefInRange(Mask, 0, NumElements / 2, NumElements))
7317 if (SDValue V = CanZExtLowHalf()) {
7318 V = DAG.getBitcast(MVT::v2i64, V);
7319 V = DAG.getNode(X86ISD::VZEXT_MOVL, DL, MVT::v2i64, V);
7320 return DAG.getBitcast(VT, V);
7323 // No viable ext lowering found.
7327 /// \brief Try to get a scalar value for a specific element of a vector.
7329 /// Looks through BUILD_VECTOR and SCALAR_TO_VECTOR nodes to find a scalar.
7330 static SDValue getScalarValueForVectorElement(SDValue V, int Idx,
7331 SelectionDAG &DAG) {
7332 MVT VT = V.getSimpleValueType();
7333 MVT EltVT = VT.getVectorElementType();
7334 while (V.getOpcode() == ISD::BITCAST)
7335 V = V.getOperand(0);
7336 // If the bitcasts shift the element size, we can't extract an equivalent
7338 MVT NewVT = V.getSimpleValueType();
7339 if (!NewVT.isVector() || NewVT.getScalarSizeInBits() != VT.getScalarSizeInBits())
7342 if (V.getOpcode() == ISD::BUILD_VECTOR ||
7343 (Idx == 0 && V.getOpcode() == ISD::SCALAR_TO_VECTOR)) {
7344 // Ensure the scalar operand is the same size as the destination.
7345 // FIXME: Add support for scalar truncation where possible.
7346 SDValue S = V.getOperand(Idx);
7347 if (EltVT.getSizeInBits() == S.getSimpleValueType().getSizeInBits())
7348 return DAG.getNode(ISD::BITCAST, SDLoc(V), EltVT, S);
7354 /// \brief Helper to test for a load that can be folded with x86 shuffles.
7356 /// This is particularly important because the set of instructions varies
7357 /// significantly based on whether the operand is a load or not.
7358 static bool isShuffleFoldableLoad(SDValue V) {
7359 while (V.getOpcode() == ISD::BITCAST)
7360 V = V.getOperand(0);
7362 return ISD::isNON_EXTLoad(V.getNode());
7365 /// \brief Try to lower insertion of a single element into a zero vector.
7367 /// This is a common pattern that we have especially efficient patterns to lower
7368 /// across all subtarget feature sets.
7369 static SDValue lowerVectorShuffleAsElementInsertion(
7370 SDLoc DL, MVT VT, SDValue V1, SDValue V2, ArrayRef<int> Mask,
7371 const X86Subtarget *Subtarget, SelectionDAG &DAG) {
7372 SmallBitVector Zeroable = computeZeroableShuffleElements(Mask, V1, V2);
7374 MVT EltVT = VT.getVectorElementType();
7376 int V2Index = std::find_if(Mask.begin(), Mask.end(),
7377 [&Mask](int M) { return M >= (int)Mask.size(); }) -
7379 bool IsV1Zeroable = true;
7380 for (int i = 0, Size = Mask.size(); i < Size; ++i)
7381 if (i != V2Index && !Zeroable[i]) {
7382 IsV1Zeroable = false;
7386 // Check for a single input from a SCALAR_TO_VECTOR node.
7387 // FIXME: All of this should be canonicalized into INSERT_VECTOR_ELT and
7388 // all the smarts here sunk into that routine. However, the current
7389 // lowering of BUILD_VECTOR makes that nearly impossible until the old
7390 // vector shuffle lowering is dead.
7391 SDValue V2S = getScalarValueForVectorElement(V2, Mask[V2Index] - Mask.size(),
7393 if (V2S && DAG.getTargetLoweringInfo().isTypeLegal(V2S.getValueType())) {
7394 // We need to zext the scalar if it is smaller than an i32.
7395 V2S = DAG.getBitcast(EltVT, V2S);
7396 if (EltVT == MVT::i8 || EltVT == MVT::i16) {
7397 // Using zext to expand a narrow element won't work for non-zero
7402 // Zero-extend directly to i32.
7404 V2S = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::i32, V2S);
7406 V2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, ExtVT, V2S);
7407 } else if (Mask[V2Index] != (int)Mask.size() || EltVT == MVT::i8 ||
7408 EltVT == MVT::i16) {
7409 // Either not inserting from the low element of the input or the input
7410 // element size is too small to use VZEXT_MOVL to clear the high bits.
7414 if (!IsV1Zeroable) {
7415 // If V1 can't be treated as a zero vector we have fewer options to lower
7416 // this. We can't support integer vectors or non-zero targets cheaply, and
7417 // the V1 elements can't be permuted in any way.
7418 assert(VT == ExtVT && "Cannot change extended type when non-zeroable!");
7419 if (!VT.isFloatingPoint() || V2Index != 0)
7421 SmallVector<int, 8> V1Mask(Mask.begin(), Mask.end());
7422 V1Mask[V2Index] = -1;
7423 if (!isNoopShuffleMask(V1Mask))
7425 // This is essentially a special case blend operation, but if we have
7426 // general purpose blend operations, they are always faster. Bail and let
7427 // the rest of the lowering handle these as blends.
7428 if (Subtarget->hasSSE41())
7431 // Otherwise, use MOVSD or MOVSS.
7432 assert((EltVT == MVT::f32 || EltVT == MVT::f64) &&
7433 "Only two types of floating point element types to handle!");
7434 return DAG.getNode(EltVT == MVT::f32 ? X86ISD::MOVSS : X86ISD::MOVSD, DL,
7438 // This lowering only works for the low element with floating point vectors.
7439 if (VT.isFloatingPoint() && V2Index != 0)
7442 V2 = DAG.getNode(X86ISD::VZEXT_MOVL, DL, ExtVT, V2);
7444 V2 = DAG.getBitcast(VT, V2);
7447 // If we have 4 or fewer lanes we can cheaply shuffle the element into
7448 // the desired position. Otherwise it is more efficient to do a vector
7449 // shift left. We know that we can do a vector shift left because all
7450 // the inputs are zero.
7451 if (VT.isFloatingPoint() || VT.getVectorNumElements() <= 4) {
7452 SmallVector<int, 4> V2Shuffle(Mask.size(), 1);
7453 V2Shuffle[V2Index] = 0;
7454 V2 = DAG.getVectorShuffle(VT, DL, V2, DAG.getUNDEF(VT), V2Shuffle);
7456 V2 = DAG.getBitcast(MVT::v2i64, V2);
7458 X86ISD::VSHLDQ, DL, MVT::v2i64, V2,
7459 DAG.getConstant(V2Index * EltVT.getSizeInBits() / 8, DL,
7460 DAG.getTargetLoweringInfo().getScalarShiftAmountTy(
7461 DAG.getDataLayout(), VT)));
7462 V2 = DAG.getBitcast(VT, V2);
7468 /// \brief Try to lower broadcast of a single element.
7470 /// For convenience, this code also bundles all of the subtarget feature set
7471 /// filtering. While a little annoying to re-dispatch on type here, there isn't
7472 /// a convenient way to factor it out.
7473 static SDValue lowerVectorShuffleAsBroadcast(SDLoc DL, MVT VT, SDValue V,
7475 const X86Subtarget *Subtarget,
7476 SelectionDAG &DAG) {
7477 if (!Subtarget->hasAVX())
7479 if (VT.isInteger() && !Subtarget->hasAVX2())
7482 // Check that the mask is a broadcast.
7483 int BroadcastIdx = -1;
7485 if (M >= 0 && BroadcastIdx == -1)
7487 else if (M >= 0 && M != BroadcastIdx)
7490 assert(BroadcastIdx < (int)Mask.size() && "We only expect to be called with "
7491 "a sorted mask where the broadcast "
7494 // Go up the chain of (vector) values to find a scalar load that we can
7495 // combine with the broadcast.
7497 switch (V.getOpcode()) {
7498 case ISD::CONCAT_VECTORS: {
7499 int OperandSize = Mask.size() / V.getNumOperands();
7500 V = V.getOperand(BroadcastIdx / OperandSize);
7501 BroadcastIdx %= OperandSize;
7505 case ISD::INSERT_SUBVECTOR: {
7506 SDValue VOuter = V.getOperand(0), VInner = V.getOperand(1);
7507 auto ConstantIdx = dyn_cast<ConstantSDNode>(V.getOperand(2));
7511 int BeginIdx = (int)ConstantIdx->getZExtValue();
7513 BeginIdx + (int)VInner.getValueType().getVectorNumElements();
7514 if (BroadcastIdx >= BeginIdx && BroadcastIdx < EndIdx) {
7515 BroadcastIdx -= BeginIdx;
7526 // Check if this is a broadcast of a scalar. We special case lowering
7527 // for scalars so that we can more effectively fold with loads.
7528 if (V.getOpcode() == ISD::BUILD_VECTOR ||
7529 (V.getOpcode() == ISD::SCALAR_TO_VECTOR && BroadcastIdx == 0)) {
7530 V = V.getOperand(BroadcastIdx);
7532 // If the scalar isn't a load, we can't broadcast from it in AVX1.
7533 // Only AVX2 has register broadcasts.
7534 if (!Subtarget->hasAVX2() && !isShuffleFoldableLoad(V))
7536 } else if (BroadcastIdx != 0 || !Subtarget->hasAVX2()) {
7537 // We can't broadcast from a vector register without AVX2, and we can only
7538 // broadcast from the zero-element of a vector register.
7542 return DAG.getNode(X86ISD::VBROADCAST, DL, VT, V);
7545 // Check for whether we can use INSERTPS to perform the shuffle. We only use
7546 // INSERTPS when the V1 elements are already in the correct locations
7547 // because otherwise we can just always use two SHUFPS instructions which
7548 // are much smaller to encode than a SHUFPS and an INSERTPS. We can also
7549 // perform INSERTPS if a single V1 element is out of place and all V2
7550 // elements are zeroable.
7551 static SDValue lowerVectorShuffleAsInsertPS(SDValue Op, SDValue V1, SDValue V2,
7553 SelectionDAG &DAG) {
7554 assert(Op.getSimpleValueType() == MVT::v4f32 && "Bad shuffle type!");
7555 assert(V1.getSimpleValueType() == MVT::v4f32 && "Bad operand type!");
7556 assert(V2.getSimpleValueType() == MVT::v4f32 && "Bad operand type!");
7557 assert(Mask.size() == 4 && "Unexpected mask size for v4 shuffle!");
7559 SmallBitVector Zeroable = computeZeroableShuffleElements(Mask, V1, V2);
7562 int V1DstIndex = -1;
7563 int V2DstIndex = -1;
7564 bool V1UsedInPlace = false;
7566 for (int i = 0; i < 4; ++i) {
7567 // Synthesize a zero mask from the zeroable elements (includes undefs).
7573 // Flag if we use any V1 inputs in place.
7575 V1UsedInPlace = true;
7579 // We can only insert a single non-zeroable element.
7580 if (V1DstIndex != -1 || V2DstIndex != -1)
7584 // V1 input out of place for insertion.
7587 // V2 input for insertion.
7592 // Don't bother if we have no (non-zeroable) element for insertion.
7593 if (V1DstIndex == -1 && V2DstIndex == -1)
7596 // Determine element insertion src/dst indices. The src index is from the
7597 // start of the inserted vector, not the start of the concatenated vector.
7598 unsigned V2SrcIndex = 0;
7599 if (V1DstIndex != -1) {
7600 // If we have a V1 input out of place, we use V1 as the V2 element insertion
7601 // and don't use the original V2 at all.
7602 V2SrcIndex = Mask[V1DstIndex];
7603 V2DstIndex = V1DstIndex;
7606 V2SrcIndex = Mask[V2DstIndex] - 4;
7609 // If no V1 inputs are used in place, then the result is created only from
7610 // the zero mask and the V2 insertion - so remove V1 dependency.
7612 V1 = DAG.getUNDEF(MVT::v4f32);
7614 unsigned InsertPSMask = V2SrcIndex << 6 | V2DstIndex << 4 | ZMask;
7615 assert((InsertPSMask & ~0xFFu) == 0 && "Invalid mask!");
7617 // Insert the V2 element into the desired position.
7619 return DAG.getNode(X86ISD::INSERTPS, DL, MVT::v4f32, V1, V2,
7620 DAG.getConstant(InsertPSMask, DL, MVT::i8));
7623 /// \brief Try to lower a shuffle as a permute of the inputs followed by an
7624 /// UNPCK instruction.
7626 /// This specifically targets cases where we end up with alternating between
7627 /// the two inputs, and so can permute them into something that feeds a single
7628 /// UNPCK instruction. Note that this routine only targets integer vectors
7629 /// because for floating point vectors we have a generalized SHUFPS lowering
7630 /// strategy that handles everything that doesn't *exactly* match an unpack,
7631 /// making this clever lowering unnecessary.
7632 static SDValue lowerVectorShuffleAsUnpack(SDLoc DL, MVT VT, SDValue V1,
7633 SDValue V2, ArrayRef<int> Mask,
7634 SelectionDAG &DAG) {
7635 assert(!VT.isFloatingPoint() &&
7636 "This routine only supports integer vectors.");
7637 assert(!isSingleInputShuffleMask(Mask) &&
7638 "This routine should only be used when blending two inputs.");
7639 assert(Mask.size() >= 2 && "Single element masks are invalid.");
7641 int Size = Mask.size();
7643 int NumLoInputs = std::count_if(Mask.begin(), Mask.end(), [Size](int M) {
7644 return M >= 0 && M % Size < Size / 2;
7646 int NumHiInputs = std::count_if(
7647 Mask.begin(), Mask.end(), [Size](int M) { return M % Size >= Size / 2; });
7649 bool UnpackLo = NumLoInputs >= NumHiInputs;
7651 auto TryUnpack = [&](MVT UnpackVT, int Scale) {
7652 SmallVector<int, 32> V1Mask(Mask.size(), -1);
7653 SmallVector<int, 32> V2Mask(Mask.size(), -1);
7655 for (int i = 0; i < Size; ++i) {
7659 // Each element of the unpack contains Scale elements from this mask.
7660 int UnpackIdx = i / Scale;
7662 // We only handle the case where V1 feeds the first slots of the unpack.
7663 // We rely on canonicalization to ensure this is the case.
7664 if ((UnpackIdx % 2 == 0) != (Mask[i] < Size))
7667 // Setup the mask for this input. The indexing is tricky as we have to
7668 // handle the unpack stride.
7669 SmallVectorImpl<int> &VMask = (UnpackIdx % 2 == 0) ? V1Mask : V2Mask;
7670 VMask[(UnpackIdx / 2) * Scale + i % Scale + (UnpackLo ? 0 : Size / 2)] =
7674 // If we will have to shuffle both inputs to use the unpack, check whether
7675 // we can just unpack first and shuffle the result. If so, skip this unpack.
7676 if ((NumLoInputs == 0 || NumHiInputs == 0) && !isNoopShuffleMask(V1Mask) &&
7677 !isNoopShuffleMask(V2Mask))
7680 // Shuffle the inputs into place.
7681 V1 = DAG.getVectorShuffle(VT, DL, V1, DAG.getUNDEF(VT), V1Mask);
7682 V2 = DAG.getVectorShuffle(VT, DL, V2, DAG.getUNDEF(VT), V2Mask);
7684 // Cast the inputs to the type we will use to unpack them.
7685 V1 = DAG.getBitcast(UnpackVT, V1);
7686 V2 = DAG.getBitcast(UnpackVT, V2);
7688 // Unpack the inputs and cast the result back to the desired type.
7689 return DAG.getBitcast(
7690 VT, DAG.getNode(UnpackLo ? X86ISD::UNPCKL : X86ISD::UNPCKH, DL,
7694 // We try each unpack from the largest to the smallest to try and find one
7695 // that fits this mask.
7696 int OrigNumElements = VT.getVectorNumElements();
7697 int OrigScalarSize = VT.getScalarSizeInBits();
7698 for (int ScalarSize = 64; ScalarSize >= OrigScalarSize; ScalarSize /= 2) {
7699 int Scale = ScalarSize / OrigScalarSize;
7700 int NumElements = OrigNumElements / Scale;
7701 MVT UnpackVT = MVT::getVectorVT(MVT::getIntegerVT(ScalarSize), NumElements);
7702 if (SDValue Unpack = TryUnpack(UnpackVT, Scale))
7706 // If none of the unpack-rooted lowerings worked (or were profitable) try an
7708 if (NumLoInputs == 0 || NumHiInputs == 0) {
7709 assert((NumLoInputs > 0 || NumHiInputs > 0) &&
7710 "We have to have *some* inputs!");
7711 int HalfOffset = NumLoInputs == 0 ? Size / 2 : 0;
7713 // FIXME: We could consider the total complexity of the permute of each
7714 // possible unpacking. Or at the least we should consider how many
7715 // half-crossings are created.
7716 // FIXME: We could consider commuting the unpacks.
7718 SmallVector<int, 32> PermMask;
7719 PermMask.assign(Size, -1);
7720 for (int i = 0; i < Size; ++i) {
7724 assert(Mask[i] % Size >= HalfOffset && "Found input from wrong half!");
7727 2 * ((Mask[i] % Size) - HalfOffset) + (Mask[i] < Size ? 0 : 1);
7729 return DAG.getVectorShuffle(
7730 VT, DL, DAG.getNode(NumLoInputs == 0 ? X86ISD::UNPCKH : X86ISD::UNPCKL,
7732 DAG.getUNDEF(VT), PermMask);
7738 /// \brief Handle lowering of 2-lane 64-bit floating point shuffles.
7740 /// This is the basis function for the 2-lane 64-bit shuffles as we have full
7741 /// support for floating point shuffles but not integer shuffles. These
7742 /// instructions will incur a domain crossing penalty on some chips though so
7743 /// it is better to avoid lowering through this for integer vectors where
7745 static SDValue lowerV2F64VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
7746 const X86Subtarget *Subtarget,
7747 SelectionDAG &DAG) {
7749 assert(Op.getSimpleValueType() == MVT::v2f64 && "Bad shuffle type!");
7750 assert(V1.getSimpleValueType() == MVT::v2f64 && "Bad operand type!");
7751 assert(V2.getSimpleValueType() == MVT::v2f64 && "Bad operand type!");
7752 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
7753 ArrayRef<int> Mask = SVOp->getMask();
7754 assert(Mask.size() == 2 && "Unexpected mask size for v2 shuffle!");
7756 if (isSingleInputShuffleMask(Mask)) {
7757 // Use low duplicate instructions for masks that match their pattern.
7758 if (Subtarget->hasSSE3())
7759 if (isShuffleEquivalent(V1, V2, Mask, {0, 0}))
7760 return DAG.getNode(X86ISD::MOVDDUP, DL, MVT::v2f64, V1);
7762 // Straight shuffle of a single input vector. Simulate this by using the
7763 // single input as both of the "inputs" to this instruction..
7764 unsigned SHUFPDMask = (Mask[0] == 1) | ((Mask[1] == 1) << 1);
7766 if (Subtarget->hasAVX()) {
7767 // If we have AVX, we can use VPERMILPS which will allow folding a load
7768 // into the shuffle.
7769 return DAG.getNode(X86ISD::VPERMILPI, DL, MVT::v2f64, V1,
7770 DAG.getConstant(SHUFPDMask, DL, MVT::i8));
7773 return DAG.getNode(X86ISD::SHUFP, DL, MVT::v2f64, V1, V1,
7774 DAG.getConstant(SHUFPDMask, DL, MVT::i8));
7776 assert(Mask[0] >= 0 && Mask[0] < 2 && "Non-canonicalized blend!");
7777 assert(Mask[1] >= 2 && "Non-canonicalized blend!");
7779 // If we have a single input, insert that into V1 if we can do so cheaply.
7780 if ((Mask[0] >= 2) + (Mask[1] >= 2) == 1) {
7781 if (SDValue Insertion = lowerVectorShuffleAsElementInsertion(
7782 DL, MVT::v2f64, V1, V2, Mask, Subtarget, DAG))
7784 // Try inverting the insertion since for v2 masks it is easy to do and we
7785 // can't reliably sort the mask one way or the other.
7786 int InverseMask[2] = {Mask[0] < 0 ? -1 : (Mask[0] ^ 2),
7787 Mask[1] < 0 ? -1 : (Mask[1] ^ 2)};
7788 if (SDValue Insertion = lowerVectorShuffleAsElementInsertion(
7789 DL, MVT::v2f64, V2, V1, InverseMask, Subtarget, DAG))
7793 // Try to use one of the special instruction patterns to handle two common
7794 // blend patterns if a zero-blend above didn't work.
7795 if (isShuffleEquivalent(V1, V2, Mask, {0, 3}) ||
7796 isShuffleEquivalent(V1, V2, Mask, {1, 3}))
7797 if (SDValue V1S = getScalarValueForVectorElement(V1, Mask[0], DAG))
7798 // We can either use a special instruction to load over the low double or
7799 // to move just the low double.
7801 isShuffleFoldableLoad(V1S) ? X86ISD::MOVLPD : X86ISD::MOVSD,
7803 DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, MVT::v2f64, V1S));
7805 if (Subtarget->hasSSE41())
7806 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v2f64, V1, V2, Mask,
7810 // Use dedicated unpack instructions for masks that match their pattern.
7811 if (isShuffleEquivalent(V1, V2, Mask, {0, 2}))
7812 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v2f64, V1, V2);
7813 if (isShuffleEquivalent(V1, V2, Mask, {1, 3}))
7814 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v2f64, V1, V2);
7816 unsigned SHUFPDMask = (Mask[0] == 1) | (((Mask[1] - 2) == 1) << 1);
7817 return DAG.getNode(X86ISD::SHUFP, DL, MVT::v2f64, V1, V2,
7818 DAG.getConstant(SHUFPDMask, DL, MVT::i8));
7821 /// \brief Handle lowering of 2-lane 64-bit integer shuffles.
7823 /// Tries to lower a 2-lane 64-bit shuffle using shuffle operations provided by
7824 /// the integer unit to minimize domain crossing penalties. However, for blends
7825 /// it falls back to the floating point shuffle operation with appropriate bit
7827 static SDValue lowerV2I64VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
7828 const X86Subtarget *Subtarget,
7829 SelectionDAG &DAG) {
7831 assert(Op.getSimpleValueType() == MVT::v2i64 && "Bad shuffle type!");
7832 assert(V1.getSimpleValueType() == MVT::v2i64 && "Bad operand type!");
7833 assert(V2.getSimpleValueType() == MVT::v2i64 && "Bad operand type!");
7834 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
7835 ArrayRef<int> Mask = SVOp->getMask();
7836 assert(Mask.size() == 2 && "Unexpected mask size for v2 shuffle!");
7838 if (isSingleInputShuffleMask(Mask)) {
7839 // Check for being able to broadcast a single element.
7840 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v2i64, V1,
7841 Mask, Subtarget, DAG))
7844 // Straight shuffle of a single input vector. For everything from SSE2
7845 // onward this has a single fast instruction with no scary immediates.
7846 // We have to map the mask as it is actually a v4i32 shuffle instruction.
7847 V1 = DAG.getBitcast(MVT::v4i32, V1);
7848 int WidenedMask[4] = {
7849 std::max(Mask[0], 0) * 2, std::max(Mask[0], 0) * 2 + 1,
7850 std::max(Mask[1], 0) * 2, std::max(Mask[1], 0) * 2 + 1};
7851 return DAG.getBitcast(
7853 DAG.getNode(X86ISD::PSHUFD, DL, MVT::v4i32, V1,
7854 getV4X86ShuffleImm8ForMask(WidenedMask, DL, DAG)));
7856 assert(Mask[0] != -1 && "No undef lanes in multi-input v2 shuffles!");
7857 assert(Mask[1] != -1 && "No undef lanes in multi-input v2 shuffles!");
7858 assert(Mask[0] < 2 && "We sort V1 to be the first input.");
7859 assert(Mask[1] >= 2 && "We sort V2 to be the second input.");
7861 // If we have a blend of two PACKUS operations an the blend aligns with the
7862 // low and half halves, we can just merge the PACKUS operations. This is
7863 // particularly important as it lets us merge shuffles that this routine itself
7865 auto GetPackNode = [](SDValue V) {
7866 while (V.getOpcode() == ISD::BITCAST)
7867 V = V.getOperand(0);
7869 return V.getOpcode() == X86ISD::PACKUS ? V : SDValue();
7871 if (SDValue V1Pack = GetPackNode(V1))
7872 if (SDValue V2Pack = GetPackNode(V2))
7873 return DAG.getBitcast(MVT::v2i64,
7874 DAG.getNode(X86ISD::PACKUS, DL, MVT::v16i8,
7875 Mask[0] == 0 ? V1Pack.getOperand(0)
7876 : V1Pack.getOperand(1),
7877 Mask[1] == 2 ? V2Pack.getOperand(0)
7878 : V2Pack.getOperand(1)));
7880 // Try to use shift instructions.
7882 lowerVectorShuffleAsShift(DL, MVT::v2i64, V1, V2, Mask, DAG))
7885 // When loading a scalar and then shuffling it into a vector we can often do
7886 // the insertion cheaply.
7887 if (SDValue Insertion = lowerVectorShuffleAsElementInsertion(
7888 DL, MVT::v2i64, V1, V2, Mask, Subtarget, DAG))
7890 // Try inverting the insertion since for v2 masks it is easy to do and we
7891 // can't reliably sort the mask one way or the other.
7892 int InverseMask[2] = {Mask[0] ^ 2, Mask[1] ^ 2};
7893 if (SDValue Insertion = lowerVectorShuffleAsElementInsertion(
7894 DL, MVT::v2i64, V2, V1, InverseMask, Subtarget, DAG))
7897 // We have different paths for blend lowering, but they all must use the
7898 // *exact* same predicate.
7899 bool IsBlendSupported = Subtarget->hasSSE41();
7900 if (IsBlendSupported)
7901 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v2i64, V1, V2, Mask,
7905 // Use dedicated unpack instructions for masks that match their pattern.
7906 if (isShuffleEquivalent(V1, V2, Mask, {0, 2}))
7907 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v2i64, V1, V2);
7908 if (isShuffleEquivalent(V1, V2, Mask, {1, 3}))
7909 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v2i64, V1, V2);
7911 // Try to use byte rotation instructions.
7912 // Its more profitable for pre-SSSE3 to use shuffles/unpacks.
7913 if (Subtarget->hasSSSE3())
7914 if (SDValue Rotate = lowerVectorShuffleAsByteRotate(
7915 DL, MVT::v2i64, V1, V2, Mask, Subtarget, DAG))
7918 // If we have direct support for blends, we should lower by decomposing into
7919 // a permute. That will be faster than the domain cross.
7920 if (IsBlendSupported)
7921 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v2i64, V1, V2,
7924 // We implement this with SHUFPD which is pretty lame because it will likely
7925 // incur 2 cycles of stall for integer vectors on Nehalem and older chips.
7926 // However, all the alternatives are still more cycles and newer chips don't
7927 // have this problem. It would be really nice if x86 had better shuffles here.
7928 V1 = DAG.getBitcast(MVT::v2f64, V1);
7929 V2 = DAG.getBitcast(MVT::v2f64, V2);
7930 return DAG.getBitcast(MVT::v2i64,
7931 DAG.getVectorShuffle(MVT::v2f64, DL, V1, V2, Mask));
7934 /// \brief Test whether this can be lowered with a single SHUFPS instruction.
7936 /// This is used to disable more specialized lowerings when the shufps lowering
7937 /// will happen to be efficient.
7938 static bool isSingleSHUFPSMask(ArrayRef<int> Mask) {
7939 // This routine only handles 128-bit shufps.
7940 assert(Mask.size() == 4 && "Unsupported mask size!");
7942 // To lower with a single SHUFPS we need to have the low half and high half
7943 // each requiring a single input.
7944 if (Mask[0] != -1 && Mask[1] != -1 && (Mask[0] < 4) != (Mask[1] < 4))
7946 if (Mask[2] != -1 && Mask[3] != -1 && (Mask[2] < 4) != (Mask[3] < 4))
7952 /// \brief Lower a vector shuffle using the SHUFPS instruction.
7954 /// This is a helper routine dedicated to lowering vector shuffles using SHUFPS.
7955 /// It makes no assumptions about whether this is the *best* lowering, it simply
7957 static SDValue lowerVectorShuffleWithSHUFPS(SDLoc DL, MVT VT,
7958 ArrayRef<int> Mask, SDValue V1,
7959 SDValue V2, SelectionDAG &DAG) {
7960 SDValue LowV = V1, HighV = V2;
7961 int NewMask[4] = {Mask[0], Mask[1], Mask[2], Mask[3]};
7964 std::count_if(Mask.begin(), Mask.end(), [](int M) { return M >= 4; });
7966 if (NumV2Elements == 1) {
7968 std::find_if(Mask.begin(), Mask.end(), [](int M) { return M >= 4; }) -
7971 // Compute the index adjacent to V2Index and in the same half by toggling
7973 int V2AdjIndex = V2Index ^ 1;
7975 if (Mask[V2AdjIndex] == -1) {
7976 // Handles all the cases where we have a single V2 element and an undef.
7977 // This will only ever happen in the high lanes because we commute the
7978 // vector otherwise.
7980 std::swap(LowV, HighV);
7981 NewMask[V2Index] -= 4;
7983 // Handle the case where the V2 element ends up adjacent to a V1 element.
7984 // To make this work, blend them together as the first step.
7985 int V1Index = V2AdjIndex;
7986 int BlendMask[4] = {Mask[V2Index] - 4, 0, Mask[V1Index], 0};
7987 V2 = DAG.getNode(X86ISD::SHUFP, DL, VT, V2, V1,
7988 getV4X86ShuffleImm8ForMask(BlendMask, DL, DAG));
7990 // Now proceed to reconstruct the final blend as we have the necessary
7991 // high or low half formed.
7998 NewMask[V1Index] = 2; // We put the V1 element in V2[2].
7999 NewMask[V2Index] = 0; // We shifted the V2 element into V2[0].
8001 } else if (NumV2Elements == 2) {
8002 if (Mask[0] < 4 && Mask[1] < 4) {
8003 // Handle the easy case where we have V1 in the low lanes and V2 in the
8007 } else if (Mask[2] < 4 && Mask[3] < 4) {
8008 // We also handle the reversed case because this utility may get called
8009 // when we detect a SHUFPS pattern but can't easily commute the shuffle to
8010 // arrange things in the right direction.
8016 // We have a mixture of V1 and V2 in both low and high lanes. Rather than
8017 // trying to place elements directly, just blend them and set up the final
8018 // shuffle to place them.
8020 // The first two blend mask elements are for V1, the second two are for
8022 int BlendMask[4] = {Mask[0] < 4 ? Mask[0] : Mask[1],
8023 Mask[2] < 4 ? Mask[2] : Mask[3],
8024 (Mask[0] >= 4 ? Mask[0] : Mask[1]) - 4,
8025 (Mask[2] >= 4 ? Mask[2] : Mask[3]) - 4};
8026 V1 = DAG.getNode(X86ISD::SHUFP, DL, VT, V1, V2,
8027 getV4X86ShuffleImm8ForMask(BlendMask, DL, DAG));
8029 // Now we do a normal shuffle of V1 by giving V1 as both operands to
8032 NewMask[0] = Mask[0] < 4 ? 0 : 2;
8033 NewMask[1] = Mask[0] < 4 ? 2 : 0;
8034 NewMask[2] = Mask[2] < 4 ? 1 : 3;
8035 NewMask[3] = Mask[2] < 4 ? 3 : 1;
8038 return DAG.getNode(X86ISD::SHUFP, DL, VT, LowV, HighV,
8039 getV4X86ShuffleImm8ForMask(NewMask, DL, DAG));
8042 /// \brief Lower 4-lane 32-bit floating point shuffles.
8044 /// Uses instructions exclusively from the floating point unit to minimize
8045 /// domain crossing penalties, as these are sufficient to implement all v4f32
8047 static SDValue lowerV4F32VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
8048 const X86Subtarget *Subtarget,
8049 SelectionDAG &DAG) {
8051 assert(Op.getSimpleValueType() == MVT::v4f32 && "Bad shuffle type!");
8052 assert(V1.getSimpleValueType() == MVT::v4f32 && "Bad operand type!");
8053 assert(V2.getSimpleValueType() == MVT::v4f32 && "Bad operand type!");
8054 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
8055 ArrayRef<int> Mask = SVOp->getMask();
8056 assert(Mask.size() == 4 && "Unexpected mask size for v4 shuffle!");
8059 std::count_if(Mask.begin(), Mask.end(), [](int M) { return M >= 4; });
8061 if (NumV2Elements == 0) {
8062 // Check for being able to broadcast a single element.
8063 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v4f32, V1,
8064 Mask, Subtarget, DAG))
8067 // Use even/odd duplicate instructions for masks that match their pattern.
8068 if (Subtarget->hasSSE3()) {
8069 if (isShuffleEquivalent(V1, V2, Mask, {0, 0, 2, 2}))
8070 return DAG.getNode(X86ISD::MOVSLDUP, DL, MVT::v4f32, V1);
8071 if (isShuffleEquivalent(V1, V2, Mask, {1, 1, 3, 3}))
8072 return DAG.getNode(X86ISD::MOVSHDUP, DL, MVT::v4f32, V1);
8075 if (Subtarget->hasAVX()) {
8076 // If we have AVX, we can use VPERMILPS which will allow folding a load
8077 // into the shuffle.
8078 return DAG.getNode(X86ISD::VPERMILPI, DL, MVT::v4f32, V1,
8079 getV4X86ShuffleImm8ForMask(Mask, DL, DAG));
8082 // Otherwise, use a straight shuffle of a single input vector. We pass the
8083 // input vector to both operands to simulate this with a SHUFPS.
8084 return DAG.getNode(X86ISD::SHUFP, DL, MVT::v4f32, V1, V1,
8085 getV4X86ShuffleImm8ForMask(Mask, DL, DAG));
8088 // There are special ways we can lower some single-element blends. However, we
8089 // have custom ways we can lower more complex single-element blends below that
8090 // we defer to if both this and BLENDPS fail to match, so restrict this to
8091 // when the V2 input is targeting element 0 of the mask -- that is the fast
8093 if (NumV2Elements == 1 && Mask[0] >= 4)
8094 if (SDValue V = lowerVectorShuffleAsElementInsertion(DL, MVT::v4f32, V1, V2,
8095 Mask, Subtarget, DAG))
8098 if (Subtarget->hasSSE41()) {
8099 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v4f32, V1, V2, Mask,
8103 // Use INSERTPS if we can complete the shuffle efficiently.
8104 if (SDValue V = lowerVectorShuffleAsInsertPS(Op, V1, V2, Mask, DAG))
8107 if (!isSingleSHUFPSMask(Mask))
8108 if (SDValue BlendPerm = lowerVectorShuffleAsBlendAndPermute(
8109 DL, MVT::v4f32, V1, V2, Mask, DAG))
8113 // Use dedicated unpack instructions for masks that match their pattern.
8114 if (isShuffleEquivalent(V1, V2, Mask, {0, 4, 1, 5}))
8115 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v4f32, V1, V2);
8116 if (isShuffleEquivalent(V1, V2, Mask, {2, 6, 3, 7}))
8117 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v4f32, V1, V2);
8118 if (isShuffleEquivalent(V1, V2, Mask, {4, 0, 5, 1}))
8119 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v4f32, V2, V1);
8120 if (isShuffleEquivalent(V1, V2, Mask, {6, 2, 7, 3}))
8121 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v4f32, V2, V1);
8123 // Otherwise fall back to a SHUFPS lowering strategy.
8124 return lowerVectorShuffleWithSHUFPS(DL, MVT::v4f32, Mask, V1, V2, DAG);
8127 /// \brief Lower 4-lane i32 vector shuffles.
8129 /// We try to handle these with integer-domain shuffles where we can, but for
8130 /// blends we use the floating point domain blend instructions.
8131 static SDValue lowerV4I32VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
8132 const X86Subtarget *Subtarget,
8133 SelectionDAG &DAG) {
8135 assert(Op.getSimpleValueType() == MVT::v4i32 && "Bad shuffle type!");
8136 assert(V1.getSimpleValueType() == MVT::v4i32 && "Bad operand type!");
8137 assert(V2.getSimpleValueType() == MVT::v4i32 && "Bad operand type!");
8138 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
8139 ArrayRef<int> Mask = SVOp->getMask();
8140 assert(Mask.size() == 4 && "Unexpected mask size for v4 shuffle!");
8142 // Whenever we can lower this as a zext, that instruction is strictly faster
8143 // than any alternative. It also allows us to fold memory operands into the
8144 // shuffle in many cases.
8145 if (SDValue ZExt = lowerVectorShuffleAsZeroOrAnyExtend(DL, MVT::v4i32, V1, V2,
8146 Mask, Subtarget, DAG))
8150 std::count_if(Mask.begin(), Mask.end(), [](int M) { return M >= 4; });
8152 if (NumV2Elements == 0) {
8153 // Check for being able to broadcast a single element.
8154 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v4i32, V1,
8155 Mask, Subtarget, DAG))
8158 // Straight shuffle of a single input vector. For everything from SSE2
8159 // onward this has a single fast instruction with no scary immediates.
8160 // We coerce the shuffle pattern to be compatible with UNPCK instructions
8161 // but we aren't actually going to use the UNPCK instruction because doing
8162 // so prevents folding a load into this instruction or making a copy.
8163 const int UnpackLoMask[] = {0, 0, 1, 1};
8164 const int UnpackHiMask[] = {2, 2, 3, 3};
8165 if (isShuffleEquivalent(V1, V2, Mask, {0, 0, 1, 1}))
8166 Mask = UnpackLoMask;
8167 else if (isShuffleEquivalent(V1, V2, Mask, {2, 2, 3, 3}))
8168 Mask = UnpackHiMask;
8170 return DAG.getNode(X86ISD::PSHUFD, DL, MVT::v4i32, V1,
8171 getV4X86ShuffleImm8ForMask(Mask, DL, DAG));
8174 // Try to use shift instructions.
8176 lowerVectorShuffleAsShift(DL, MVT::v4i32, V1, V2, Mask, DAG))
8179 // There are special ways we can lower some single-element blends.
8180 if (NumV2Elements == 1)
8181 if (SDValue V = lowerVectorShuffleAsElementInsertion(DL, MVT::v4i32, V1, V2,
8182 Mask, Subtarget, DAG))
8185 // We have different paths for blend lowering, but they all must use the
8186 // *exact* same predicate.
8187 bool IsBlendSupported = Subtarget->hasSSE41();
8188 if (IsBlendSupported)
8189 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v4i32, V1, V2, Mask,
8193 if (SDValue Masked =
8194 lowerVectorShuffleAsBitMask(DL, MVT::v4i32, V1, V2, Mask, DAG))
8197 // Use dedicated unpack instructions for masks that match their pattern.
8198 if (isShuffleEquivalent(V1, V2, Mask, {0, 4, 1, 5}))
8199 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v4i32, V1, V2);
8200 if (isShuffleEquivalent(V1, V2, Mask, {2, 6, 3, 7}))
8201 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v4i32, V1, V2);
8202 if (isShuffleEquivalent(V1, V2, Mask, {4, 0, 5, 1}))
8203 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v4i32, V2, V1);
8204 if (isShuffleEquivalent(V1, V2, Mask, {6, 2, 7, 3}))
8205 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v4i32, V2, V1);
8207 // Try to use byte rotation instructions.
8208 // Its more profitable for pre-SSSE3 to use shuffles/unpacks.
8209 if (Subtarget->hasSSSE3())
8210 if (SDValue Rotate = lowerVectorShuffleAsByteRotate(
8211 DL, MVT::v4i32, V1, V2, Mask, Subtarget, DAG))
8214 // If we have direct support for blends, we should lower by decomposing into
8215 // a permute. That will be faster than the domain cross.
8216 if (IsBlendSupported)
8217 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v4i32, V1, V2,
8220 // Try to lower by permuting the inputs into an unpack instruction.
8221 if (SDValue Unpack =
8222 lowerVectorShuffleAsUnpack(DL, MVT::v4i32, V1, V2, Mask, DAG))
8225 // We implement this with SHUFPS because it can blend from two vectors.
8226 // Because we're going to eventually use SHUFPS, we use SHUFPS even to build
8227 // up the inputs, bypassing domain shift penalties that we would encur if we
8228 // directly used PSHUFD on Nehalem and older. For newer chips, this isn't
8230 return DAG.getBitcast(
8232 DAG.getVectorShuffle(MVT::v4f32, DL, DAG.getBitcast(MVT::v4f32, V1),
8233 DAG.getBitcast(MVT::v4f32, V2), Mask));
8236 /// \brief Lowering of single-input v8i16 shuffles is the cornerstone of SSE2
8237 /// shuffle lowering, and the most complex part.
8239 /// The lowering strategy is to try to form pairs of input lanes which are
8240 /// targeted at the same half of the final vector, and then use a dword shuffle
8241 /// to place them onto the right half, and finally unpack the paired lanes into
8242 /// their final position.
8244 /// The exact breakdown of how to form these dword pairs and align them on the
8245 /// correct sides is really tricky. See the comments within the function for
8246 /// more of the details.
8248 /// This code also handles repeated 128-bit lanes of v8i16 shuffles, but each
8249 /// lane must shuffle the *exact* same way. In fact, you must pass a v8 Mask to
8250 /// this routine for it to work correctly. To shuffle a 256-bit or 512-bit i16
8251 /// vector, form the analogous 128-bit 8-element Mask.
8252 static SDValue lowerV8I16GeneralSingleInputVectorShuffle(
8253 SDLoc DL, MVT VT, SDValue V, MutableArrayRef<int> Mask,
8254 const X86Subtarget *Subtarget, SelectionDAG &DAG) {
8255 assert(VT.getScalarType() == MVT::i16 && "Bad input type!");
8256 MVT PSHUFDVT = MVT::getVectorVT(MVT::i32, VT.getVectorNumElements() / 2);
8258 assert(Mask.size() == 8 && "Shuffle mask length doen't match!");
8259 MutableArrayRef<int> LoMask = Mask.slice(0, 4);
8260 MutableArrayRef<int> HiMask = Mask.slice(4, 4);
8262 SmallVector<int, 4> LoInputs;
8263 std::copy_if(LoMask.begin(), LoMask.end(), std::back_inserter(LoInputs),
8264 [](int M) { return M >= 0; });
8265 std::sort(LoInputs.begin(), LoInputs.end());
8266 LoInputs.erase(std::unique(LoInputs.begin(), LoInputs.end()), LoInputs.end());
8267 SmallVector<int, 4> HiInputs;
8268 std::copy_if(HiMask.begin(), HiMask.end(), std::back_inserter(HiInputs),
8269 [](int M) { return M >= 0; });
8270 std::sort(HiInputs.begin(), HiInputs.end());
8271 HiInputs.erase(std::unique(HiInputs.begin(), HiInputs.end()), HiInputs.end());
8273 std::lower_bound(LoInputs.begin(), LoInputs.end(), 4) - LoInputs.begin();
8274 int NumHToL = LoInputs.size() - NumLToL;
8276 std::lower_bound(HiInputs.begin(), HiInputs.end(), 4) - HiInputs.begin();
8277 int NumHToH = HiInputs.size() - NumLToH;
8278 MutableArrayRef<int> LToLInputs(LoInputs.data(), NumLToL);
8279 MutableArrayRef<int> LToHInputs(HiInputs.data(), NumLToH);
8280 MutableArrayRef<int> HToLInputs(LoInputs.data() + NumLToL, NumHToL);
8281 MutableArrayRef<int> HToHInputs(HiInputs.data() + NumLToH, NumHToH);
8283 // Simplify the 1-into-3 and 3-into-1 cases with a single pshufd. For all
8284 // such inputs we can swap two of the dwords across the half mark and end up
8285 // with <=2 inputs to each half in each half. Once there, we can fall through
8286 // to the generic code below. For example:
8288 // Input: [a, b, c, d, e, f, g, h] -PSHUFD[0,2,1,3]-> [a, b, e, f, c, d, g, h]
8289 // Mask: [0, 1, 2, 7, 4, 5, 6, 3] -----------------> [0, 1, 4, 7, 2, 3, 6, 5]
8291 // However in some very rare cases we have a 1-into-3 or 3-into-1 on one half
8292 // and an existing 2-into-2 on the other half. In this case we may have to
8293 // pre-shuffle the 2-into-2 half to avoid turning it into a 3-into-1 or
8294 // 1-into-3 which could cause us to cycle endlessly fixing each side in turn.
8295 // Fortunately, we don't have to handle anything but a 2-into-2 pattern
8296 // because any other situation (including a 3-into-1 or 1-into-3 in the other
8297 // half than the one we target for fixing) will be fixed when we re-enter this
8298 // path. We will also combine away any sequence of PSHUFD instructions that
8299 // result into a single instruction. Here is an example of the tricky case:
8301 // Input: [a, b, c, d, e, f, g, h] -PSHUFD[0,2,1,3]-> [a, b, e, f, c, d, g, h]
8302 // Mask: [3, 7, 1, 0, 2, 7, 3, 5] -THIS-IS-BAD!!!!-> [5, 7, 1, 0, 4, 7, 5, 3]
8304 // This now has a 1-into-3 in the high half! Instead, we do two shuffles:
8306 // Input: [a, b, c, d, e, f, g, h] PSHUFHW[0,2,1,3]-> [a, b, c, d, e, g, f, h]
8307 // Mask: [3, 7, 1, 0, 2, 7, 3, 5] -----------------> [3, 7, 1, 0, 2, 7, 3, 6]
8309 // Input: [a, b, c, d, e, g, f, h] -PSHUFD[0,2,1,3]-> [a, b, e, g, c, d, f, h]
8310 // Mask: [3, 7, 1, 0, 2, 7, 3, 6] -----------------> [5, 7, 1, 0, 4, 7, 5, 6]
8312 // The result is fine to be handled by the generic logic.
8313 auto balanceSides = [&](ArrayRef<int> AToAInputs, ArrayRef<int> BToAInputs,
8314 ArrayRef<int> BToBInputs, ArrayRef<int> AToBInputs,
8315 int AOffset, int BOffset) {
8316 assert((AToAInputs.size() == 3 || AToAInputs.size() == 1) &&
8317 "Must call this with A having 3 or 1 inputs from the A half.");
8318 assert((BToAInputs.size() == 1 || BToAInputs.size() == 3) &&
8319 "Must call this with B having 1 or 3 inputs from the B half.");
8320 assert(AToAInputs.size() + BToAInputs.size() == 4 &&
8321 "Must call this with either 3:1 or 1:3 inputs (summing to 4).");
8323 // Compute the index of dword with only one word among the three inputs in
8324 // a half by taking the sum of the half with three inputs and subtracting
8325 // the sum of the actual three inputs. The difference is the remaining
8328 int &TripleDWord = AToAInputs.size() == 3 ? ADWord : BDWord;
8329 int &OneInputDWord = AToAInputs.size() == 3 ? BDWord : ADWord;
8330 int TripleInputOffset = AToAInputs.size() == 3 ? AOffset : BOffset;
8331 ArrayRef<int> TripleInputs = AToAInputs.size() == 3 ? AToAInputs : BToAInputs;
8332 int OneInput = AToAInputs.size() == 3 ? BToAInputs[0] : AToAInputs[0];
8333 int TripleInputSum = 0 + 1 + 2 + 3 + (4 * TripleInputOffset);
8334 int TripleNonInputIdx =
8335 TripleInputSum - std::accumulate(TripleInputs.begin(), TripleInputs.end(), 0);
8336 TripleDWord = TripleNonInputIdx / 2;
8338 // We use xor with one to compute the adjacent DWord to whichever one the
8340 OneInputDWord = (OneInput / 2) ^ 1;
8342 // Check for one tricky case: We're fixing a 3<-1 or a 1<-3 shuffle for AToA
8343 // and BToA inputs. If there is also such a problem with the BToB and AToB
8344 // inputs, we don't try to fix it necessarily -- we'll recurse and see it in
8345 // the next pass. However, if we have a 2<-2 in the BToB and AToB inputs, it
8346 // is essential that we don't *create* a 3<-1 as then we might oscillate.
8347 if (BToBInputs.size() == 2 && AToBInputs.size() == 2) {
8348 // Compute how many inputs will be flipped by swapping these DWords. We
8350 // to balance this to ensure we don't form a 3-1 shuffle in the other
8352 int NumFlippedAToBInputs =
8353 std::count(AToBInputs.begin(), AToBInputs.end(), 2 * ADWord) +
8354 std::count(AToBInputs.begin(), AToBInputs.end(), 2 * ADWord + 1);
8355 int NumFlippedBToBInputs =
8356 std::count(BToBInputs.begin(), BToBInputs.end(), 2 * BDWord) +
8357 std::count(BToBInputs.begin(), BToBInputs.end(), 2 * BDWord + 1);
8358 if ((NumFlippedAToBInputs == 1 &&
8359 (NumFlippedBToBInputs == 0 || NumFlippedBToBInputs == 2)) ||
8360 (NumFlippedBToBInputs == 1 &&
8361 (NumFlippedAToBInputs == 0 || NumFlippedAToBInputs == 2))) {
8362 // We choose whether to fix the A half or B half based on whether that
8363 // half has zero flipped inputs. At zero, we may not be able to fix it
8364 // with that half. We also bias towards fixing the B half because that
8365 // will more commonly be the high half, and we have to bias one way.
8366 auto FixFlippedInputs = [&V, &DL, &Mask, &DAG](int PinnedIdx, int DWord,
8367 ArrayRef<int> Inputs) {
8368 int FixIdx = PinnedIdx ^ 1; // The adjacent slot to the pinned slot.
8369 bool IsFixIdxInput = std::find(Inputs.begin(), Inputs.end(),
8370 PinnedIdx ^ 1) != Inputs.end();
8371 // Determine whether the free index is in the flipped dword or the
8372 // unflipped dword based on where the pinned index is. We use this bit
8373 // in an xor to conditionally select the adjacent dword.
8374 int FixFreeIdx = 2 * (DWord ^ (PinnedIdx / 2 == DWord));
8375 bool IsFixFreeIdxInput = std::find(Inputs.begin(), Inputs.end(),
8376 FixFreeIdx) != Inputs.end();
8377 if (IsFixIdxInput == IsFixFreeIdxInput)
8379 IsFixFreeIdxInput = std::find(Inputs.begin(), Inputs.end(),
8380 FixFreeIdx) != Inputs.end();
8381 assert(IsFixIdxInput != IsFixFreeIdxInput &&
8382 "We need to be changing the number of flipped inputs!");
8383 int PSHUFHalfMask[] = {0, 1, 2, 3};
8384 std::swap(PSHUFHalfMask[FixFreeIdx % 4], PSHUFHalfMask[FixIdx % 4]);
8385 V = DAG.getNode(FixIdx < 4 ? X86ISD::PSHUFLW : X86ISD::PSHUFHW, DL,
8387 getV4X86ShuffleImm8ForMask(PSHUFHalfMask, DL, DAG));
8390 if (M != -1 && M == FixIdx)
8392 else if (M != -1 && M == FixFreeIdx)
8395 if (NumFlippedBToBInputs != 0) {
8397 BToAInputs.size() == 3 ? TripleNonInputIdx : OneInput;
8398 FixFlippedInputs(BPinnedIdx, BDWord, BToBInputs);
8400 assert(NumFlippedAToBInputs != 0 && "Impossible given predicates!");
8402 AToAInputs.size() == 3 ? TripleNonInputIdx : OneInput;
8403 FixFlippedInputs(APinnedIdx, ADWord, AToBInputs);
8408 int PSHUFDMask[] = {0, 1, 2, 3};
8409 PSHUFDMask[ADWord] = BDWord;
8410 PSHUFDMask[BDWord] = ADWord;
8413 DAG.getNode(X86ISD::PSHUFD, DL, PSHUFDVT, DAG.getBitcast(PSHUFDVT, V),
8414 getV4X86ShuffleImm8ForMask(PSHUFDMask, DL, DAG)));
8416 // Adjust the mask to match the new locations of A and B.
8418 if (M != -1 && M/2 == ADWord)
8419 M = 2 * BDWord + M % 2;
8420 else if (M != -1 && M/2 == BDWord)
8421 M = 2 * ADWord + M % 2;
8423 // Recurse back into this routine to re-compute state now that this isn't
8424 // a 3 and 1 problem.
8425 return lowerV8I16GeneralSingleInputVectorShuffle(DL, VT, V, Mask, Subtarget,
8428 if ((NumLToL == 3 && NumHToL == 1) || (NumLToL == 1 && NumHToL == 3))
8429 return balanceSides(LToLInputs, HToLInputs, HToHInputs, LToHInputs, 0, 4);
8430 else if ((NumHToH == 3 && NumLToH == 1) || (NumHToH == 1 && NumLToH == 3))
8431 return balanceSides(HToHInputs, LToHInputs, LToLInputs, HToLInputs, 4, 0);
8433 // At this point there are at most two inputs to the low and high halves from
8434 // each half. That means the inputs can always be grouped into dwords and
8435 // those dwords can then be moved to the correct half with a dword shuffle.
8436 // We use at most one low and one high word shuffle to collect these paired
8437 // inputs into dwords, and finally a dword shuffle to place them.
8438 int PSHUFLMask[4] = {-1, -1, -1, -1};
8439 int PSHUFHMask[4] = {-1, -1, -1, -1};
8440 int PSHUFDMask[4] = {-1, -1, -1, -1};
8442 // First fix the masks for all the inputs that are staying in their
8443 // original halves. This will then dictate the targets of the cross-half
8445 auto fixInPlaceInputs =
8446 [&PSHUFDMask](ArrayRef<int> InPlaceInputs, ArrayRef<int> IncomingInputs,
8447 MutableArrayRef<int> SourceHalfMask,
8448 MutableArrayRef<int> HalfMask, int HalfOffset) {
8449 if (InPlaceInputs.empty())
8451 if (InPlaceInputs.size() == 1) {
8452 SourceHalfMask[InPlaceInputs[0] - HalfOffset] =
8453 InPlaceInputs[0] - HalfOffset;
8454 PSHUFDMask[InPlaceInputs[0] / 2] = InPlaceInputs[0] / 2;
8457 if (IncomingInputs.empty()) {
8458 // Just fix all of the in place inputs.
8459 for (int Input : InPlaceInputs) {
8460 SourceHalfMask[Input - HalfOffset] = Input - HalfOffset;
8461 PSHUFDMask[Input / 2] = Input / 2;
8466 assert(InPlaceInputs.size() == 2 && "Cannot handle 3 or 4 inputs!");
8467 SourceHalfMask[InPlaceInputs[0] - HalfOffset] =
8468 InPlaceInputs[0] - HalfOffset;
8469 // Put the second input next to the first so that they are packed into
8470 // a dword. We find the adjacent index by toggling the low bit.
8471 int AdjIndex = InPlaceInputs[0] ^ 1;
8472 SourceHalfMask[AdjIndex - HalfOffset] = InPlaceInputs[1] - HalfOffset;
8473 std::replace(HalfMask.begin(), HalfMask.end(), InPlaceInputs[1], AdjIndex);
8474 PSHUFDMask[AdjIndex / 2] = AdjIndex / 2;
8476 fixInPlaceInputs(LToLInputs, HToLInputs, PSHUFLMask, LoMask, 0);
8477 fixInPlaceInputs(HToHInputs, LToHInputs, PSHUFHMask, HiMask, 4);
8479 // Now gather the cross-half inputs and place them into a free dword of
8480 // their target half.
8481 // FIXME: This operation could almost certainly be simplified dramatically to
8482 // look more like the 3-1 fixing operation.
8483 auto moveInputsToRightHalf = [&PSHUFDMask](
8484 MutableArrayRef<int> IncomingInputs, ArrayRef<int> ExistingInputs,
8485 MutableArrayRef<int> SourceHalfMask, MutableArrayRef<int> HalfMask,
8486 MutableArrayRef<int> FinalSourceHalfMask, int SourceOffset,
8488 auto isWordClobbered = [](ArrayRef<int> SourceHalfMask, int Word) {
8489 return SourceHalfMask[Word] != -1 && SourceHalfMask[Word] != Word;
8491 auto isDWordClobbered = [&isWordClobbered](ArrayRef<int> SourceHalfMask,
8493 int LowWord = Word & ~1;
8494 int HighWord = Word | 1;
8495 return isWordClobbered(SourceHalfMask, LowWord) ||
8496 isWordClobbered(SourceHalfMask, HighWord);
8499 if (IncomingInputs.empty())
8502 if (ExistingInputs.empty()) {
8503 // Map any dwords with inputs from them into the right half.
8504 for (int Input : IncomingInputs) {
8505 // If the source half mask maps over the inputs, turn those into
8506 // swaps and use the swapped lane.
8507 if (isWordClobbered(SourceHalfMask, Input - SourceOffset)) {
8508 if (SourceHalfMask[SourceHalfMask[Input - SourceOffset]] == -1) {
8509 SourceHalfMask[SourceHalfMask[Input - SourceOffset]] =
8510 Input - SourceOffset;
8511 // We have to swap the uses in our half mask in one sweep.
8512 for (int &M : HalfMask)
8513 if (M == SourceHalfMask[Input - SourceOffset] + SourceOffset)
8515 else if (M == Input)
8516 M = SourceHalfMask[Input - SourceOffset] + SourceOffset;
8518 assert(SourceHalfMask[SourceHalfMask[Input - SourceOffset]] ==
8519 Input - SourceOffset &&
8520 "Previous placement doesn't match!");
8522 // Note that this correctly re-maps both when we do a swap and when
8523 // we observe the other side of the swap above. We rely on that to
8524 // avoid swapping the members of the input list directly.
8525 Input = SourceHalfMask[Input - SourceOffset] + SourceOffset;
8528 // Map the input's dword into the correct half.
8529 if (PSHUFDMask[(Input - SourceOffset + DestOffset) / 2] == -1)
8530 PSHUFDMask[(Input - SourceOffset + DestOffset) / 2] = Input / 2;
8532 assert(PSHUFDMask[(Input - SourceOffset + DestOffset) / 2] ==
8534 "Previous placement doesn't match!");
8537 // And just directly shift any other-half mask elements to be same-half
8538 // as we will have mirrored the dword containing the element into the
8539 // same position within that half.
8540 for (int &M : HalfMask)
8541 if (M >= SourceOffset && M < SourceOffset + 4) {
8542 M = M - SourceOffset + DestOffset;
8543 assert(M >= 0 && "This should never wrap below zero!");
8548 // Ensure we have the input in a viable dword of its current half. This
8549 // is particularly tricky because the original position may be clobbered
8550 // by inputs being moved and *staying* in that half.
8551 if (IncomingInputs.size() == 1) {
8552 if (isWordClobbered(SourceHalfMask, IncomingInputs[0] - SourceOffset)) {
8553 int InputFixed = std::find(std::begin(SourceHalfMask),
8554 std::end(SourceHalfMask), -1) -
8555 std::begin(SourceHalfMask) + SourceOffset;
8556 SourceHalfMask[InputFixed - SourceOffset] =
8557 IncomingInputs[0] - SourceOffset;
8558 std::replace(HalfMask.begin(), HalfMask.end(), IncomingInputs[0],
8560 IncomingInputs[0] = InputFixed;
8562 } else if (IncomingInputs.size() == 2) {
8563 if (IncomingInputs[0] / 2 != IncomingInputs[1] / 2 ||
8564 isDWordClobbered(SourceHalfMask, IncomingInputs[0] - SourceOffset)) {
8565 // We have two non-adjacent or clobbered inputs we need to extract from
8566 // the source half. To do this, we need to map them into some adjacent
8567 // dword slot in the source mask.
8568 int InputsFixed[2] = {IncomingInputs[0] - SourceOffset,
8569 IncomingInputs[1] - SourceOffset};
8571 // If there is a free slot in the source half mask adjacent to one of
8572 // the inputs, place the other input in it. We use (Index XOR 1) to
8573 // compute an adjacent index.
8574 if (!isWordClobbered(SourceHalfMask, InputsFixed[0]) &&
8575 SourceHalfMask[InputsFixed[0] ^ 1] == -1) {
8576 SourceHalfMask[InputsFixed[0]] = InputsFixed[0];
8577 SourceHalfMask[InputsFixed[0] ^ 1] = InputsFixed[1];
8578 InputsFixed[1] = InputsFixed[0] ^ 1;
8579 } else if (!isWordClobbered(SourceHalfMask, InputsFixed[1]) &&
8580 SourceHalfMask[InputsFixed[1] ^ 1] == -1) {
8581 SourceHalfMask[InputsFixed[1]] = InputsFixed[1];
8582 SourceHalfMask[InputsFixed[1] ^ 1] = InputsFixed[0];
8583 InputsFixed[0] = InputsFixed[1] ^ 1;
8584 } else if (SourceHalfMask[2 * ((InputsFixed[0] / 2) ^ 1)] == -1 &&
8585 SourceHalfMask[2 * ((InputsFixed[0] / 2) ^ 1) + 1] == -1) {
8586 // The two inputs are in the same DWord but it is clobbered and the
8587 // adjacent DWord isn't used at all. Move both inputs to the free
8589 SourceHalfMask[2 * ((InputsFixed[0] / 2) ^ 1)] = InputsFixed[0];
8590 SourceHalfMask[2 * ((InputsFixed[0] / 2) ^ 1) + 1] = InputsFixed[1];
8591 InputsFixed[0] = 2 * ((InputsFixed[0] / 2) ^ 1);
8592 InputsFixed[1] = 2 * ((InputsFixed[0] / 2) ^ 1) + 1;
8594 // The only way we hit this point is if there is no clobbering
8595 // (because there are no off-half inputs to this half) and there is no
8596 // free slot adjacent to one of the inputs. In this case, we have to
8597 // swap an input with a non-input.
8598 for (int i = 0; i < 4; ++i)
8599 assert((SourceHalfMask[i] == -1 || SourceHalfMask[i] == i) &&
8600 "We can't handle any clobbers here!");
8601 assert(InputsFixed[1] != (InputsFixed[0] ^ 1) &&
8602 "Cannot have adjacent inputs here!");
8604 SourceHalfMask[InputsFixed[0] ^ 1] = InputsFixed[1];
8605 SourceHalfMask[InputsFixed[1]] = InputsFixed[0] ^ 1;
8607 // We also have to update the final source mask in this case because
8608 // it may need to undo the above swap.
8609 for (int &M : FinalSourceHalfMask)
8610 if (M == (InputsFixed[0] ^ 1) + SourceOffset)
8611 M = InputsFixed[1] + SourceOffset;
8612 else if (M == InputsFixed[1] + SourceOffset)
8613 M = (InputsFixed[0] ^ 1) + SourceOffset;
8615 InputsFixed[1] = InputsFixed[0] ^ 1;
8618 // Point everything at the fixed inputs.
8619 for (int &M : HalfMask)
8620 if (M == IncomingInputs[0])
8621 M = InputsFixed[0] + SourceOffset;
8622 else if (M == IncomingInputs[1])
8623 M = InputsFixed[1] + SourceOffset;
8625 IncomingInputs[0] = InputsFixed[0] + SourceOffset;
8626 IncomingInputs[1] = InputsFixed[1] + SourceOffset;
8629 llvm_unreachable("Unhandled input size!");
8632 // Now hoist the DWord down to the right half.
8633 int FreeDWord = (PSHUFDMask[DestOffset / 2] == -1 ? 0 : 1) + DestOffset / 2;
8634 assert(PSHUFDMask[FreeDWord] == -1 && "DWord not free");
8635 PSHUFDMask[FreeDWord] = IncomingInputs[0] / 2;
8636 for (int &M : HalfMask)
8637 for (int Input : IncomingInputs)
8639 M = FreeDWord * 2 + Input % 2;
8641 moveInputsToRightHalf(HToLInputs, LToLInputs, PSHUFHMask, LoMask, HiMask,
8642 /*SourceOffset*/ 4, /*DestOffset*/ 0);
8643 moveInputsToRightHalf(LToHInputs, HToHInputs, PSHUFLMask, HiMask, LoMask,
8644 /*SourceOffset*/ 0, /*DestOffset*/ 4);
8646 // Now enact all the shuffles we've computed to move the inputs into their
8648 if (!isNoopShuffleMask(PSHUFLMask))
8649 V = DAG.getNode(X86ISD::PSHUFLW, DL, VT, V,
8650 getV4X86ShuffleImm8ForMask(PSHUFLMask, DL, DAG));
8651 if (!isNoopShuffleMask(PSHUFHMask))
8652 V = DAG.getNode(X86ISD::PSHUFHW, DL, VT, V,
8653 getV4X86ShuffleImm8ForMask(PSHUFHMask, DL, DAG));
8654 if (!isNoopShuffleMask(PSHUFDMask))
8657 DAG.getNode(X86ISD::PSHUFD, DL, PSHUFDVT, DAG.getBitcast(PSHUFDVT, V),
8658 getV4X86ShuffleImm8ForMask(PSHUFDMask, DL, DAG)));
8660 // At this point, each half should contain all its inputs, and we can then
8661 // just shuffle them into their final position.
8662 assert(std::count_if(LoMask.begin(), LoMask.end(),
8663 [](int M) { return M >= 4; }) == 0 &&
8664 "Failed to lift all the high half inputs to the low mask!");
8665 assert(std::count_if(HiMask.begin(), HiMask.end(),
8666 [](int M) { return M >= 0 && M < 4; }) == 0 &&
8667 "Failed to lift all the low half inputs to the high mask!");
8669 // Do a half shuffle for the low mask.
8670 if (!isNoopShuffleMask(LoMask))
8671 V = DAG.getNode(X86ISD::PSHUFLW, DL, VT, V,
8672 getV4X86ShuffleImm8ForMask(LoMask, DL, DAG));
8674 // Do a half shuffle with the high mask after shifting its values down.
8675 for (int &M : HiMask)
8678 if (!isNoopShuffleMask(HiMask))
8679 V = DAG.getNode(X86ISD::PSHUFHW, DL, VT, V,
8680 getV4X86ShuffleImm8ForMask(HiMask, DL, DAG));
8685 /// \brief Helper to form a PSHUFB-based shuffle+blend.
8686 static SDValue lowerVectorShuffleAsPSHUFB(SDLoc DL, MVT VT, SDValue V1,
8687 SDValue V2, ArrayRef<int> Mask,
8688 SelectionDAG &DAG, bool &V1InUse,
8690 SmallBitVector Zeroable = computeZeroableShuffleElements(Mask, V1, V2);
8696 int Size = Mask.size();
8697 int Scale = 16 / Size;
8698 for (int i = 0; i < 16; ++i) {
8699 if (Mask[i / Scale] == -1) {
8700 V1Mask[i] = V2Mask[i] = DAG.getUNDEF(MVT::i8);
8702 const int ZeroMask = 0x80;
8703 int V1Idx = Mask[i / Scale] < Size ? Mask[i / Scale] * Scale + i % Scale
8705 int V2Idx = Mask[i / Scale] < Size
8707 : (Mask[i / Scale] - Size) * Scale + i % Scale;
8708 if (Zeroable[i / Scale])
8709 V1Idx = V2Idx = ZeroMask;
8710 V1Mask[i] = DAG.getConstant(V1Idx, DL, MVT::i8);
8711 V2Mask[i] = DAG.getConstant(V2Idx, DL, MVT::i8);
8712 V1InUse |= (ZeroMask != V1Idx);
8713 V2InUse |= (ZeroMask != V2Idx);
8718 V1 = DAG.getNode(X86ISD::PSHUFB, DL, MVT::v16i8,
8719 DAG.getBitcast(MVT::v16i8, V1),
8720 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v16i8, V1Mask));
8722 V2 = DAG.getNode(X86ISD::PSHUFB, DL, MVT::v16i8,
8723 DAG.getBitcast(MVT::v16i8, V2),
8724 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v16i8, V2Mask));
8726 // If we need shuffled inputs from both, blend the two.
8728 if (V1InUse && V2InUse)
8729 V = DAG.getNode(ISD::OR, DL, MVT::v16i8, V1, V2);
8731 V = V1InUse ? V1 : V2;
8733 // Cast the result back to the correct type.
8734 return DAG.getBitcast(VT, V);
8737 /// \brief Generic lowering of 8-lane i16 shuffles.
8739 /// This handles both single-input shuffles and combined shuffle/blends with
8740 /// two inputs. The single input shuffles are immediately delegated to
8741 /// a dedicated lowering routine.
8743 /// The blends are lowered in one of three fundamental ways. If there are few
8744 /// enough inputs, it delegates to a basic UNPCK-based strategy. If the shuffle
8745 /// of the input is significantly cheaper when lowered as an interleaving of
8746 /// the two inputs, try to interleave them. Otherwise, blend the low and high
8747 /// halves of the inputs separately (making them have relatively few inputs)
8748 /// and then concatenate them.
8749 static SDValue lowerV8I16VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
8750 const X86Subtarget *Subtarget,
8751 SelectionDAG &DAG) {
8753 assert(Op.getSimpleValueType() == MVT::v8i16 && "Bad shuffle type!");
8754 assert(V1.getSimpleValueType() == MVT::v8i16 && "Bad operand type!");
8755 assert(V2.getSimpleValueType() == MVT::v8i16 && "Bad operand type!");
8756 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
8757 ArrayRef<int> OrigMask = SVOp->getMask();
8758 int MaskStorage[8] = {OrigMask[0], OrigMask[1], OrigMask[2], OrigMask[3],
8759 OrigMask[4], OrigMask[5], OrigMask[6], OrigMask[7]};
8760 MutableArrayRef<int> Mask(MaskStorage);
8762 assert(Mask.size() == 8 && "Unexpected mask size for v8 shuffle!");
8764 // Whenever we can lower this as a zext, that instruction is strictly faster
8765 // than any alternative.
8766 if (SDValue ZExt = lowerVectorShuffleAsZeroOrAnyExtend(
8767 DL, MVT::v8i16, V1, V2, OrigMask, Subtarget, DAG))
8770 auto isV1 = [](int M) { return M >= 0 && M < 8; };
8772 auto isV2 = [](int M) { return M >= 8; };
8774 int NumV2Inputs = std::count_if(Mask.begin(), Mask.end(), isV2);
8776 if (NumV2Inputs == 0) {
8777 // Check for being able to broadcast a single element.
8778 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v8i16, V1,
8779 Mask, Subtarget, DAG))
8782 // Try to use shift instructions.
8784 lowerVectorShuffleAsShift(DL, MVT::v8i16, V1, V1, Mask, DAG))
8787 // Use dedicated unpack instructions for masks that match their pattern.
8788 if (isShuffleEquivalent(V1, V1, Mask, {0, 0, 1, 1, 2, 2, 3, 3}))
8789 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v8i16, V1, V1);
8790 if (isShuffleEquivalent(V1, V1, Mask, {4, 4, 5, 5, 6, 6, 7, 7}))
8791 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v8i16, V1, V1);
8793 // Try to use byte rotation instructions.
8794 if (SDValue Rotate = lowerVectorShuffleAsByteRotate(DL, MVT::v8i16, V1, V1,
8795 Mask, Subtarget, DAG))
8798 return lowerV8I16GeneralSingleInputVectorShuffle(DL, MVT::v8i16, V1, Mask,
8802 assert(std::any_of(Mask.begin(), Mask.end(), isV1) &&
8803 "All single-input shuffles should be canonicalized to be V1-input "
8806 // Try to use shift instructions.
8808 lowerVectorShuffleAsShift(DL, MVT::v8i16, V1, V2, Mask, DAG))
8811 // See if we can use SSE4A Extraction / Insertion.
8812 if (Subtarget->hasSSE4A())
8813 if (SDValue V = lowerVectorShuffleWithSSE4A(DL, MVT::v8i16, V1, V2, Mask, DAG))
8816 // There are special ways we can lower some single-element blends.
8817 if (NumV2Inputs == 1)
8818 if (SDValue V = lowerVectorShuffleAsElementInsertion(DL, MVT::v8i16, V1, V2,
8819 Mask, Subtarget, DAG))
8822 // We have different paths for blend lowering, but they all must use the
8823 // *exact* same predicate.
8824 bool IsBlendSupported = Subtarget->hasSSE41();
8825 if (IsBlendSupported)
8826 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v8i16, V1, V2, Mask,
8830 if (SDValue Masked =
8831 lowerVectorShuffleAsBitMask(DL, MVT::v8i16, V1, V2, Mask, DAG))
8834 // Use dedicated unpack instructions for masks that match their pattern.
8835 if (isShuffleEquivalent(V1, V2, Mask, {0, 8, 1, 9, 2, 10, 3, 11}))
8836 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v8i16, V1, V2);
8837 if (isShuffleEquivalent(V1, V2, Mask, {4, 12, 5, 13, 6, 14, 7, 15}))
8838 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v8i16, V1, V2);
8840 // Try to use byte rotation instructions.
8841 if (SDValue Rotate = lowerVectorShuffleAsByteRotate(
8842 DL, MVT::v8i16, V1, V2, Mask, Subtarget, DAG))
8845 if (SDValue BitBlend =
8846 lowerVectorShuffleAsBitBlend(DL, MVT::v8i16, V1, V2, Mask, DAG))
8849 if (SDValue Unpack =
8850 lowerVectorShuffleAsUnpack(DL, MVT::v8i16, V1, V2, Mask, DAG))
8853 // If we can't directly blend but can use PSHUFB, that will be better as it
8854 // can both shuffle and set up the inefficient blend.
8855 if (!IsBlendSupported && Subtarget->hasSSSE3()) {
8856 bool V1InUse, V2InUse;
8857 return lowerVectorShuffleAsPSHUFB(DL, MVT::v8i16, V1, V2, Mask, DAG,
8861 // We can always bit-blend if we have to so the fallback strategy is to
8862 // decompose into single-input permutes and blends.
8863 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v8i16, V1, V2,
8867 /// \brief Check whether a compaction lowering can be done by dropping even
8868 /// elements and compute how many times even elements must be dropped.
8870 /// This handles shuffles which take every Nth element where N is a power of
8871 /// two. Example shuffle masks:
8873 /// N = 1: 0, 2, 4, 6, 8, 10, 12, 14, 0, 2, 4, 6, 8, 10, 12, 14
8874 /// N = 1: 0, 2, 4, 6, 8, 10, 12, 14, 16, 18, 20, 22, 24, 26, 28, 30
8875 /// N = 2: 0, 4, 8, 12, 0, 4, 8, 12, 0, 4, 8, 12, 0, 4, 8, 12
8876 /// N = 2: 0, 4, 8, 12, 16, 20, 24, 28, 0, 4, 8, 12, 16, 20, 24, 28
8877 /// N = 3: 0, 8, 0, 8, 0, 8, 0, 8, 0, 8, 0, 8, 0, 8, 0, 8
8878 /// N = 3: 0, 8, 16, 24, 0, 8, 16, 24, 0, 8, 16, 24, 0, 8, 16, 24
8880 /// Any of these lanes can of course be undef.
8882 /// This routine only supports N <= 3.
8883 /// FIXME: Evaluate whether either AVX or AVX-512 have any opportunities here
8886 /// \returns N above, or the number of times even elements must be dropped if
8887 /// there is such a number. Otherwise returns zero.
8888 static int canLowerByDroppingEvenElements(ArrayRef<int> Mask) {
8889 // Figure out whether we're looping over two inputs or just one.
8890 bool IsSingleInput = isSingleInputShuffleMask(Mask);
8892 // The modulus for the shuffle vector entries is based on whether this is
8893 // a single input or not.
8894 int ShuffleModulus = Mask.size() * (IsSingleInput ? 1 : 2);
8895 assert(isPowerOf2_32((uint32_t)ShuffleModulus) &&
8896 "We should only be called with masks with a power-of-2 size!");
8898 uint64_t ModMask = (uint64_t)ShuffleModulus - 1;
8900 // We track whether the input is viable for all power-of-2 strides 2^1, 2^2,
8901 // and 2^3 simultaneously. This is because we may have ambiguity with
8902 // partially undef inputs.
8903 bool ViableForN[3] = {true, true, true};
8905 for (int i = 0, e = Mask.size(); i < e; ++i) {
8906 // Ignore undef lanes, we'll optimistically collapse them to the pattern we
8911 bool IsAnyViable = false;
8912 for (unsigned j = 0; j != array_lengthof(ViableForN); ++j)
8913 if (ViableForN[j]) {
8916 // The shuffle mask must be equal to (i * 2^N) % M.
8917 if ((uint64_t)Mask[i] == (((uint64_t)i << N) & ModMask))
8920 ViableForN[j] = false;
8922 // Early exit if we exhaust the possible powers of two.
8927 for (unsigned j = 0; j != array_lengthof(ViableForN); ++j)
8931 // Return 0 as there is no viable power of two.
8935 /// \brief Generic lowering of v16i8 shuffles.
8937 /// This is a hybrid strategy to lower v16i8 vectors. It first attempts to
8938 /// detect any complexity reducing interleaving. If that doesn't help, it uses
8939 /// UNPCK to spread the i8 elements across two i16-element vectors, and uses
8940 /// the existing lowering for v8i16 blends on each half, finally PACK-ing them
8942 static SDValue lowerV16I8VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
8943 const X86Subtarget *Subtarget,
8944 SelectionDAG &DAG) {
8946 assert(Op.getSimpleValueType() == MVT::v16i8 && "Bad shuffle type!");
8947 assert(V1.getSimpleValueType() == MVT::v16i8 && "Bad operand type!");
8948 assert(V2.getSimpleValueType() == MVT::v16i8 && "Bad operand type!");
8949 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
8950 ArrayRef<int> Mask = SVOp->getMask();
8951 assert(Mask.size() == 16 && "Unexpected mask size for v16 shuffle!");
8953 // Try to use shift instructions.
8955 lowerVectorShuffleAsShift(DL, MVT::v16i8, V1, V2, Mask, DAG))
8958 // Try to use byte rotation instructions.
8959 if (SDValue Rotate = lowerVectorShuffleAsByteRotate(
8960 DL, MVT::v16i8, V1, V2, Mask, Subtarget, DAG))
8963 // Try to use a zext lowering.
8964 if (SDValue ZExt = lowerVectorShuffleAsZeroOrAnyExtend(
8965 DL, MVT::v16i8, V1, V2, Mask, Subtarget, DAG))
8968 // See if we can use SSE4A Extraction / Insertion.
8969 if (Subtarget->hasSSE4A())
8970 if (SDValue V = lowerVectorShuffleWithSSE4A(DL, MVT::v16i8, V1, V2, Mask, DAG))
8974 std::count_if(Mask.begin(), Mask.end(), [](int M) { return M >= 16; });
8976 // For single-input shuffles, there are some nicer lowering tricks we can use.
8977 if (NumV2Elements == 0) {
8978 // Check for being able to broadcast a single element.
8979 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v16i8, V1,
8980 Mask, Subtarget, DAG))
8983 // Check whether we can widen this to an i16 shuffle by duplicating bytes.
8984 // Notably, this handles splat and partial-splat shuffles more efficiently.
8985 // However, it only makes sense if the pre-duplication shuffle simplifies
8986 // things significantly. Currently, this means we need to be able to
8987 // express the pre-duplication shuffle as an i16 shuffle.
8989 // FIXME: We should check for other patterns which can be widened into an
8990 // i16 shuffle as well.
8991 auto canWidenViaDuplication = [](ArrayRef<int> Mask) {
8992 for (int i = 0; i < 16; i += 2)
8993 if (Mask[i] != -1 && Mask[i + 1] != -1 && Mask[i] != Mask[i + 1])
8998 auto tryToWidenViaDuplication = [&]() -> SDValue {
8999 if (!canWidenViaDuplication(Mask))
9001 SmallVector<int, 4> LoInputs;
9002 std::copy_if(Mask.begin(), Mask.end(), std::back_inserter(LoInputs),
9003 [](int M) { return M >= 0 && M < 8; });
9004 std::sort(LoInputs.begin(), LoInputs.end());
9005 LoInputs.erase(std::unique(LoInputs.begin(), LoInputs.end()),
9007 SmallVector<int, 4> HiInputs;
9008 std::copy_if(Mask.begin(), Mask.end(), std::back_inserter(HiInputs),
9009 [](int M) { return M >= 8; });
9010 std::sort(HiInputs.begin(), HiInputs.end());
9011 HiInputs.erase(std::unique(HiInputs.begin(), HiInputs.end()),
9014 bool TargetLo = LoInputs.size() >= HiInputs.size();
9015 ArrayRef<int> InPlaceInputs = TargetLo ? LoInputs : HiInputs;
9016 ArrayRef<int> MovingInputs = TargetLo ? HiInputs : LoInputs;
9018 int PreDupI16Shuffle[] = {-1, -1, -1, -1, -1, -1, -1, -1};
9019 SmallDenseMap<int, int, 8> LaneMap;
9020 for (int I : InPlaceInputs) {
9021 PreDupI16Shuffle[I/2] = I/2;
9024 int j = TargetLo ? 0 : 4, je = j + 4;
9025 for (int i = 0, ie = MovingInputs.size(); i < ie; ++i) {
9026 // Check if j is already a shuffle of this input. This happens when
9027 // there are two adjacent bytes after we move the low one.
9028 if (PreDupI16Shuffle[j] != MovingInputs[i] / 2) {
9029 // If we haven't yet mapped the input, search for a slot into which
9031 while (j < je && PreDupI16Shuffle[j] != -1)
9035 // We can't place the inputs into a single half with a simple i16 shuffle, so bail.
9038 // Map this input with the i16 shuffle.
9039 PreDupI16Shuffle[j] = MovingInputs[i] / 2;
9042 // Update the lane map based on the mapping we ended up with.
9043 LaneMap[MovingInputs[i]] = 2 * j + MovingInputs[i] % 2;
9045 V1 = DAG.getBitcast(
9047 DAG.getVectorShuffle(MVT::v8i16, DL, DAG.getBitcast(MVT::v8i16, V1),
9048 DAG.getUNDEF(MVT::v8i16), PreDupI16Shuffle));
9050 // Unpack the bytes to form the i16s that will be shuffled into place.
9051 V1 = DAG.getNode(TargetLo ? X86ISD::UNPCKL : X86ISD::UNPCKH, DL,
9052 MVT::v16i8, V1, V1);
9054 int PostDupI16Shuffle[8] = {-1, -1, -1, -1, -1, -1, -1, -1};
9055 for (int i = 0; i < 16; ++i)
9056 if (Mask[i] != -1) {
9057 int MappedMask = LaneMap[Mask[i]] - (TargetLo ? 0 : 8);
9058 assert(MappedMask < 8 && "Invalid v8 shuffle mask!");
9059 if (PostDupI16Shuffle[i / 2] == -1)
9060 PostDupI16Shuffle[i / 2] = MappedMask;
9062 assert(PostDupI16Shuffle[i / 2] == MappedMask &&
9063 "Conflicting entrties in the original shuffle!");
9065 return DAG.getBitcast(
9067 DAG.getVectorShuffle(MVT::v8i16, DL, DAG.getBitcast(MVT::v8i16, V1),
9068 DAG.getUNDEF(MVT::v8i16), PostDupI16Shuffle));
9070 if (SDValue V = tryToWidenViaDuplication())
9074 // Use dedicated unpack instructions for masks that match their pattern.
9075 if (isShuffleEquivalent(V1, V2, Mask, {// Low half.
9076 0, 16, 1, 17, 2, 18, 3, 19,
9078 4, 20, 5, 21, 6, 22, 7, 23}))
9079 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v16i8, V1, V2);
9080 if (isShuffleEquivalent(V1, V2, Mask, {// Low half.
9081 8, 24, 9, 25, 10, 26, 11, 27,
9083 12, 28, 13, 29, 14, 30, 15, 31}))
9084 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v16i8, V1, V2);
9086 // Check for SSSE3 which lets us lower all v16i8 shuffles much more directly
9087 // with PSHUFB. It is important to do this before we attempt to generate any
9088 // blends but after all of the single-input lowerings. If the single input
9089 // lowerings can find an instruction sequence that is faster than a PSHUFB, we
9090 // want to preserve that and we can DAG combine any longer sequences into
9091 // a PSHUFB in the end. But once we start blending from multiple inputs,
9092 // the complexity of DAG combining bad patterns back into PSHUFB is too high,
9093 // and there are *very* few patterns that would actually be faster than the
9094 // PSHUFB approach because of its ability to zero lanes.
9096 // FIXME: The only exceptions to the above are blends which are exact
9097 // interleavings with direct instructions supporting them. We currently don't
9098 // handle those well here.
9099 if (Subtarget->hasSSSE3()) {
9100 bool V1InUse = false;
9101 bool V2InUse = false;
9103 SDValue PSHUFB = lowerVectorShuffleAsPSHUFB(DL, MVT::v16i8, V1, V2, Mask,
9104 DAG, V1InUse, V2InUse);
9106 // If both V1 and V2 are in use and we can use a direct blend or an unpack,
9107 // do so. This avoids using them to handle blends-with-zero which is
9108 // important as a single pshufb is significantly faster for that.
9109 if (V1InUse && V2InUse) {
9110 if (Subtarget->hasSSE41())
9111 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v16i8, V1, V2,
9112 Mask, Subtarget, DAG))
9115 // We can use an unpack to do the blending rather than an or in some
9116 // cases. Even though the or may be (very minorly) more efficient, we
9117 // preference this lowering because there are common cases where part of
9118 // the complexity of the shuffles goes away when we do the final blend as
9120 // FIXME: It might be worth trying to detect if the unpack-feeding
9121 // shuffles will both be pshufb, in which case we shouldn't bother with
9123 if (SDValue Unpack =
9124 lowerVectorShuffleAsUnpack(DL, MVT::v16i8, V1, V2, Mask, DAG))
9131 // There are special ways we can lower some single-element blends.
9132 if (NumV2Elements == 1)
9133 if (SDValue V = lowerVectorShuffleAsElementInsertion(DL, MVT::v16i8, V1, V2,
9134 Mask, Subtarget, DAG))
9137 if (SDValue BitBlend =
9138 lowerVectorShuffleAsBitBlend(DL, MVT::v16i8, V1, V2, Mask, DAG))
9141 // Check whether a compaction lowering can be done. This handles shuffles
9142 // which take every Nth element for some even N. See the helper function for
9145 // We special case these as they can be particularly efficiently handled with
9146 // the PACKUSB instruction on x86 and they show up in common patterns of
9147 // rearranging bytes to truncate wide elements.
9148 if (int NumEvenDrops = canLowerByDroppingEvenElements(Mask)) {
9149 // NumEvenDrops is the power of two stride of the elements. Another way of
9150 // thinking about it is that we need to drop the even elements this many
9151 // times to get the original input.
9152 bool IsSingleInput = isSingleInputShuffleMask(Mask);
9154 // First we need to zero all the dropped bytes.
9155 assert(NumEvenDrops <= 3 &&
9156 "No support for dropping even elements more than 3 times.");
9157 // We use the mask type to pick which bytes are preserved based on how many
9158 // elements are dropped.
9159 MVT MaskVTs[] = { MVT::v8i16, MVT::v4i32, MVT::v2i64 };
9160 SDValue ByteClearMask = DAG.getBitcast(
9161 MVT::v16i8, DAG.getConstant(0xFF, DL, MaskVTs[NumEvenDrops - 1]));
9162 V1 = DAG.getNode(ISD::AND, DL, MVT::v16i8, V1, ByteClearMask);
9164 V2 = DAG.getNode(ISD::AND, DL, MVT::v16i8, V2, ByteClearMask);
9166 // Now pack things back together.
9167 V1 = DAG.getBitcast(MVT::v8i16, V1);
9168 V2 = IsSingleInput ? V1 : DAG.getBitcast(MVT::v8i16, V2);
9169 SDValue Result = DAG.getNode(X86ISD::PACKUS, DL, MVT::v16i8, V1, V2);
9170 for (int i = 1; i < NumEvenDrops; ++i) {
9171 Result = DAG.getBitcast(MVT::v8i16, Result);
9172 Result = DAG.getNode(X86ISD::PACKUS, DL, MVT::v16i8, Result, Result);
9178 // Handle multi-input cases by blending single-input shuffles.
9179 if (NumV2Elements > 0)
9180 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v16i8, V1, V2,
9183 // The fallback path for single-input shuffles widens this into two v8i16
9184 // vectors with unpacks, shuffles those, and then pulls them back together
9188 int LoBlendMask[8] = {-1, -1, -1, -1, -1, -1, -1, -1};
9189 int HiBlendMask[8] = {-1, -1, -1, -1, -1, -1, -1, -1};
9190 for (int i = 0; i < 16; ++i)
9192 (i < 8 ? LoBlendMask[i] : HiBlendMask[i % 8]) = Mask[i];
9194 SDValue Zero = getZeroVector(MVT::v8i16, Subtarget, DAG, DL);
9196 SDValue VLoHalf, VHiHalf;
9197 // Check if any of the odd lanes in the v16i8 are used. If not, we can mask
9198 // them out and avoid using UNPCK{L,H} to extract the elements of V as
9200 if (std::none_of(std::begin(LoBlendMask), std::end(LoBlendMask),
9201 [](int M) { return M >= 0 && M % 2 == 1; }) &&
9202 std::none_of(std::begin(HiBlendMask), std::end(HiBlendMask),
9203 [](int M) { return M >= 0 && M % 2 == 1; })) {
9204 // Use a mask to drop the high bytes.
9205 VLoHalf = DAG.getBitcast(MVT::v8i16, V);
9206 VLoHalf = DAG.getNode(ISD::AND, DL, MVT::v8i16, VLoHalf,
9207 DAG.getConstant(0x00FF, DL, MVT::v8i16));
9209 // This will be a single vector shuffle instead of a blend so nuke VHiHalf.
9210 VHiHalf = DAG.getUNDEF(MVT::v8i16);
9212 // Squash the masks to point directly into VLoHalf.
9213 for (int &M : LoBlendMask)
9216 for (int &M : HiBlendMask)
9220 // Otherwise just unpack the low half of V into VLoHalf and the high half into
9221 // VHiHalf so that we can blend them as i16s.
9222 VLoHalf = DAG.getBitcast(
9223 MVT::v8i16, DAG.getNode(X86ISD::UNPCKL, DL, MVT::v16i8, V, Zero));
9224 VHiHalf = DAG.getBitcast(
9225 MVT::v8i16, DAG.getNode(X86ISD::UNPCKH, DL, MVT::v16i8, V, Zero));
9228 SDValue LoV = DAG.getVectorShuffle(MVT::v8i16, DL, VLoHalf, VHiHalf, LoBlendMask);
9229 SDValue HiV = DAG.getVectorShuffle(MVT::v8i16, DL, VLoHalf, VHiHalf, HiBlendMask);
9231 return DAG.getNode(X86ISD::PACKUS, DL, MVT::v16i8, LoV, HiV);
9234 /// \brief Dispatching routine to lower various 128-bit x86 vector shuffles.
9236 /// This routine breaks down the specific type of 128-bit shuffle and
9237 /// dispatches to the lowering routines accordingly.
9238 static SDValue lower128BitVectorShuffle(SDValue Op, SDValue V1, SDValue V2,
9239 MVT VT, const X86Subtarget *Subtarget,
9240 SelectionDAG &DAG) {
9241 switch (VT.SimpleTy) {
9243 return lowerV2I64VectorShuffle(Op, V1, V2, Subtarget, DAG);
9245 return lowerV2F64VectorShuffle(Op, V1, V2, Subtarget, DAG);
9247 return lowerV4I32VectorShuffle(Op, V1, V2, Subtarget, DAG);
9249 return lowerV4F32VectorShuffle(Op, V1, V2, Subtarget, DAG);
9251 return lowerV8I16VectorShuffle(Op, V1, V2, Subtarget, DAG);
9253 return lowerV16I8VectorShuffle(Op, V1, V2, Subtarget, DAG);
9256 llvm_unreachable("Unimplemented!");
9260 /// \brief Helper function to test whether a shuffle mask could be
9261 /// simplified by widening the elements being shuffled.
9263 /// Appends the mask for wider elements in WidenedMask if valid. Otherwise
9264 /// leaves it in an unspecified state.
9266 /// NOTE: This must handle normal vector shuffle masks and *target* vector
9267 /// shuffle masks. The latter have the special property of a '-2' representing
9268 /// a zero-ed lane of a vector.
9269 static bool canWidenShuffleElements(ArrayRef<int> Mask,
9270 SmallVectorImpl<int> &WidenedMask) {
9271 for (int i = 0, Size = Mask.size(); i < Size; i += 2) {
9272 // If both elements are undef, its trivial.
9273 if (Mask[i] == SM_SentinelUndef && Mask[i + 1] == SM_SentinelUndef) {
9274 WidenedMask.push_back(SM_SentinelUndef);
9278 // Check for an undef mask and a mask value properly aligned to fit with
9279 // a pair of values. If we find such a case, use the non-undef mask's value.
9280 if (Mask[i] == SM_SentinelUndef && Mask[i + 1] >= 0 && Mask[i + 1] % 2 == 1) {
9281 WidenedMask.push_back(Mask[i + 1] / 2);
9284 if (Mask[i + 1] == SM_SentinelUndef && Mask[i] >= 0 && Mask[i] % 2 == 0) {
9285 WidenedMask.push_back(Mask[i] / 2);
9289 // When zeroing, we need to spread the zeroing across both lanes to widen.
9290 if (Mask[i] == SM_SentinelZero || Mask[i + 1] == SM_SentinelZero) {
9291 if ((Mask[i] == SM_SentinelZero || Mask[i] == SM_SentinelUndef) &&
9292 (Mask[i + 1] == SM_SentinelZero || Mask[i + 1] == SM_SentinelUndef)) {
9293 WidenedMask.push_back(SM_SentinelZero);
9299 // Finally check if the two mask values are adjacent and aligned with
9301 if (Mask[i] != SM_SentinelUndef && Mask[i] % 2 == 0 && Mask[i] + 1 == Mask[i + 1]) {
9302 WidenedMask.push_back(Mask[i] / 2);
9306 // Otherwise we can't safely widen the elements used in this shuffle.
9309 assert(WidenedMask.size() == Mask.size() / 2 &&
9310 "Incorrect size of mask after widening the elements!");
9315 /// \brief Generic routine to split vector shuffle into half-sized shuffles.
9317 /// This routine just extracts two subvectors, shuffles them independently, and
9318 /// then concatenates them back together. This should work effectively with all
9319 /// AVX vector shuffle types.
9320 static SDValue splitAndLowerVectorShuffle(SDLoc DL, MVT VT, SDValue V1,
9321 SDValue V2, ArrayRef<int> Mask,
9322 SelectionDAG &DAG) {
9323 assert(VT.getSizeInBits() >= 256 &&
9324 "Only for 256-bit or wider vector shuffles!");
9325 assert(V1.getSimpleValueType() == VT && "Bad operand type!");
9326 assert(V2.getSimpleValueType() == VT && "Bad operand type!");
9328 ArrayRef<int> LoMask = Mask.slice(0, Mask.size() / 2);
9329 ArrayRef<int> HiMask = Mask.slice(Mask.size() / 2);
9331 int NumElements = VT.getVectorNumElements();
9332 int SplitNumElements = NumElements / 2;
9333 MVT ScalarVT = VT.getScalarType();
9334 MVT SplitVT = MVT::getVectorVT(ScalarVT, NumElements / 2);
9336 // Rather than splitting build-vectors, just build two narrower build
9337 // vectors. This helps shuffling with splats and zeros.
9338 auto SplitVector = [&](SDValue V) {
9339 while (V.getOpcode() == ISD::BITCAST)
9340 V = V->getOperand(0);
9342 MVT OrigVT = V.getSimpleValueType();
9343 int OrigNumElements = OrigVT.getVectorNumElements();
9344 int OrigSplitNumElements = OrigNumElements / 2;
9345 MVT OrigScalarVT = OrigVT.getScalarType();
9346 MVT OrigSplitVT = MVT::getVectorVT(OrigScalarVT, OrigNumElements / 2);
9350 auto *BV = dyn_cast<BuildVectorSDNode>(V);
9352 LoV = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, OrigSplitVT, V,
9353 DAG.getIntPtrConstant(0, DL));
9354 HiV = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, OrigSplitVT, V,
9355 DAG.getIntPtrConstant(OrigSplitNumElements, DL));
9358 SmallVector<SDValue, 16> LoOps, HiOps;
9359 for (int i = 0; i < OrigSplitNumElements; ++i) {
9360 LoOps.push_back(BV->getOperand(i));
9361 HiOps.push_back(BV->getOperand(i + OrigSplitNumElements));
9363 LoV = DAG.getNode(ISD::BUILD_VECTOR, DL, OrigSplitVT, LoOps);
9364 HiV = DAG.getNode(ISD::BUILD_VECTOR, DL, OrigSplitVT, HiOps);
9366 return std::make_pair(DAG.getBitcast(SplitVT, LoV),
9367 DAG.getBitcast(SplitVT, HiV));
9370 SDValue LoV1, HiV1, LoV2, HiV2;
9371 std::tie(LoV1, HiV1) = SplitVector(V1);
9372 std::tie(LoV2, HiV2) = SplitVector(V2);
9374 // Now create two 4-way blends of these half-width vectors.
9375 auto HalfBlend = [&](ArrayRef<int> HalfMask) {
9376 bool UseLoV1 = false, UseHiV1 = false, UseLoV2 = false, UseHiV2 = false;
9377 SmallVector<int, 32> V1BlendMask, V2BlendMask, BlendMask;
9378 for (int i = 0; i < SplitNumElements; ++i) {
9379 int M = HalfMask[i];
9380 if (M >= NumElements) {
9381 if (M >= NumElements + SplitNumElements)
9385 V2BlendMask.push_back(M - NumElements);
9386 V1BlendMask.push_back(-1);
9387 BlendMask.push_back(SplitNumElements + i);
9388 } else if (M >= 0) {
9389 if (M >= SplitNumElements)
9393 V2BlendMask.push_back(-1);
9394 V1BlendMask.push_back(M);
9395 BlendMask.push_back(i);
9397 V2BlendMask.push_back(-1);
9398 V1BlendMask.push_back(-1);
9399 BlendMask.push_back(-1);
9403 // Because the lowering happens after all combining takes place, we need to
9404 // manually combine these blend masks as much as possible so that we create
9405 // a minimal number of high-level vector shuffle nodes.
9407 // First try just blending the halves of V1 or V2.
9408 if (!UseLoV1 && !UseHiV1 && !UseLoV2 && !UseHiV2)
9409 return DAG.getUNDEF(SplitVT);
9410 if (!UseLoV2 && !UseHiV2)
9411 return DAG.getVectorShuffle(SplitVT, DL, LoV1, HiV1, V1BlendMask);
9412 if (!UseLoV1 && !UseHiV1)
9413 return DAG.getVectorShuffle(SplitVT, DL, LoV2, HiV2, V2BlendMask);
9415 SDValue V1Blend, V2Blend;
9416 if (UseLoV1 && UseHiV1) {
9418 DAG.getVectorShuffle(SplitVT, DL, LoV1, HiV1, V1BlendMask);
9420 // We only use half of V1 so map the usage down into the final blend mask.
9421 V1Blend = UseLoV1 ? LoV1 : HiV1;
9422 for (int i = 0; i < SplitNumElements; ++i)
9423 if (BlendMask[i] >= 0 && BlendMask[i] < SplitNumElements)
9424 BlendMask[i] = V1BlendMask[i] - (UseLoV1 ? 0 : SplitNumElements);
9426 if (UseLoV2 && UseHiV2) {
9428 DAG.getVectorShuffle(SplitVT, DL, LoV2, HiV2, V2BlendMask);
9430 // We only use half of V2 so map the usage down into the final blend mask.
9431 V2Blend = UseLoV2 ? LoV2 : HiV2;
9432 for (int i = 0; i < SplitNumElements; ++i)
9433 if (BlendMask[i] >= SplitNumElements)
9434 BlendMask[i] = V2BlendMask[i] + (UseLoV2 ? SplitNumElements : 0);
9436 return DAG.getVectorShuffle(SplitVT, DL, V1Blend, V2Blend, BlendMask);
9438 SDValue Lo = HalfBlend(LoMask);
9439 SDValue Hi = HalfBlend(HiMask);
9440 return DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, Lo, Hi);
9443 /// \brief Either split a vector in halves or decompose the shuffles and the
9446 /// This is provided as a good fallback for many lowerings of non-single-input
9447 /// shuffles with more than one 128-bit lane. In those cases, we want to select
9448 /// between splitting the shuffle into 128-bit components and stitching those
9449 /// back together vs. extracting the single-input shuffles and blending those
9451 static SDValue lowerVectorShuffleAsSplitOrBlend(SDLoc DL, MVT VT, SDValue V1,
9452 SDValue V2, ArrayRef<int> Mask,
9453 SelectionDAG &DAG) {
9454 assert(!isSingleInputShuffleMask(Mask) && "This routine must not be used to "
9455 "lower single-input shuffles as it "
9456 "could then recurse on itself.");
9457 int Size = Mask.size();
9459 // If this can be modeled as a broadcast of two elements followed by a blend,
9460 // prefer that lowering. This is especially important because broadcasts can
9461 // often fold with memory operands.
9462 auto DoBothBroadcast = [&] {
9463 int V1BroadcastIdx = -1, V2BroadcastIdx = -1;
9466 if (V2BroadcastIdx == -1)
9467 V2BroadcastIdx = M - Size;
9468 else if (M - Size != V2BroadcastIdx)
9470 } else if (M >= 0) {
9471 if (V1BroadcastIdx == -1)
9473 else if (M != V1BroadcastIdx)
9478 if (DoBothBroadcast())
9479 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, VT, V1, V2, Mask,
9482 // If the inputs all stem from a single 128-bit lane of each input, then we
9483 // split them rather than blending because the split will decompose to
9484 // unusually few instructions.
9485 int LaneCount = VT.getSizeInBits() / 128;
9486 int LaneSize = Size / LaneCount;
9487 SmallBitVector LaneInputs[2];
9488 LaneInputs[0].resize(LaneCount, false);
9489 LaneInputs[1].resize(LaneCount, false);
9490 for (int i = 0; i < Size; ++i)
9492 LaneInputs[Mask[i] / Size][(Mask[i] % Size) / LaneSize] = true;
9493 if (LaneInputs[0].count() <= 1 && LaneInputs[1].count() <= 1)
9494 return splitAndLowerVectorShuffle(DL, VT, V1, V2, Mask, DAG);
9496 // Otherwise, just fall back to decomposed shuffles and a blend. This requires
9497 // that the decomposed single-input shuffles don't end up here.
9498 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, VT, V1, V2, Mask, DAG);
9501 /// \brief Lower a vector shuffle crossing multiple 128-bit lanes as
9502 /// a permutation and blend of those lanes.
9504 /// This essentially blends the out-of-lane inputs to each lane into the lane
9505 /// from a permuted copy of the vector. This lowering strategy results in four
9506 /// instructions in the worst case for a single-input cross lane shuffle which
9507 /// is lower than any other fully general cross-lane shuffle strategy I'm aware
9508 /// of. Special cases for each particular shuffle pattern should be handled
9509 /// prior to trying this lowering.
9510 static SDValue lowerVectorShuffleAsLanePermuteAndBlend(SDLoc DL, MVT VT,
9511 SDValue V1, SDValue V2,
9513 SelectionDAG &DAG) {
9514 // FIXME: This should probably be generalized for 512-bit vectors as well.
9515 assert(VT.getSizeInBits() == 256 && "Only for 256-bit vector shuffles!");
9516 int LaneSize = Mask.size() / 2;
9518 // If there are only inputs from one 128-bit lane, splitting will in fact be
9519 // less expensive. The flags track whether the given lane contains an element
9520 // that crosses to another lane.
9521 bool LaneCrossing[2] = {false, false};
9522 for (int i = 0, Size = Mask.size(); i < Size; ++i)
9523 if (Mask[i] >= 0 && (Mask[i] % Size) / LaneSize != i / LaneSize)
9524 LaneCrossing[(Mask[i] % Size) / LaneSize] = true;
9525 if (!LaneCrossing[0] || !LaneCrossing[1])
9526 return splitAndLowerVectorShuffle(DL, VT, V1, V2, Mask, DAG);
9528 if (isSingleInputShuffleMask(Mask)) {
9529 SmallVector<int, 32> FlippedBlendMask;
9530 for (int i = 0, Size = Mask.size(); i < Size; ++i)
9531 FlippedBlendMask.push_back(
9532 Mask[i] < 0 ? -1 : (((Mask[i] % Size) / LaneSize == i / LaneSize)
9534 : Mask[i] % LaneSize +
9535 (i / LaneSize) * LaneSize + Size));
9537 // Flip the vector, and blend the results which should now be in-lane. The
9538 // VPERM2X128 mask uses the low 2 bits for the low source and bits 4 and
9539 // 5 for the high source. The value 3 selects the high half of source 2 and
9540 // the value 2 selects the low half of source 2. We only use source 2 to
9541 // allow folding it into a memory operand.
9542 unsigned PERMMask = 3 | 2 << 4;
9543 SDValue Flipped = DAG.getNode(X86ISD::VPERM2X128, DL, VT, DAG.getUNDEF(VT),
9544 V1, DAG.getConstant(PERMMask, DL, MVT::i8));
9545 return DAG.getVectorShuffle(VT, DL, V1, Flipped, FlippedBlendMask);
9548 // This now reduces to two single-input shuffles of V1 and V2 which at worst
9549 // will be handled by the above logic and a blend of the results, much like
9550 // other patterns in AVX.
9551 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, VT, V1, V2, Mask, DAG);
9554 /// \brief Handle lowering 2-lane 128-bit shuffles.
9555 static SDValue lowerV2X128VectorShuffle(SDLoc DL, MVT VT, SDValue V1,
9556 SDValue V2, ArrayRef<int> Mask,
9557 const X86Subtarget *Subtarget,
9558 SelectionDAG &DAG) {
9559 // TODO: If minimizing size and one of the inputs is a zero vector and the
9560 // the zero vector has only one use, we could use a VPERM2X128 to save the
9561 // instruction bytes needed to explicitly generate the zero vector.
9563 // Blends are faster and handle all the non-lane-crossing cases.
9564 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, VT, V1, V2, Mask,
9568 bool IsV1Zero = ISD::isBuildVectorAllZeros(V1.getNode());
9569 bool IsV2Zero = ISD::isBuildVectorAllZeros(V2.getNode());
9571 // If either input operand is a zero vector, use VPERM2X128 because its mask
9572 // allows us to replace the zero input with an implicit zero.
9573 if (!IsV1Zero && !IsV2Zero) {
9574 // Check for patterns which can be matched with a single insert of a 128-bit
9576 bool OnlyUsesV1 = isShuffleEquivalent(V1, V2, Mask, {0, 1, 0, 1});
9577 if (OnlyUsesV1 || isShuffleEquivalent(V1, V2, Mask, {0, 1, 4, 5})) {
9578 MVT SubVT = MVT::getVectorVT(VT.getVectorElementType(),
9579 VT.getVectorNumElements() / 2);
9580 SDValue LoV = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, SubVT, V1,
9581 DAG.getIntPtrConstant(0, DL));
9582 SDValue HiV = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, SubVT,
9583 OnlyUsesV1 ? V1 : V2,
9584 DAG.getIntPtrConstant(0, DL));
9585 return DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, LoV, HiV);
9589 // Otherwise form a 128-bit permutation. After accounting for undefs,
9590 // convert the 64-bit shuffle mask selection values into 128-bit
9591 // selection bits by dividing the indexes by 2 and shifting into positions
9592 // defined by a vperm2*128 instruction's immediate control byte.
9594 // The immediate permute control byte looks like this:
9595 // [1:0] - select 128 bits from sources for low half of destination
9597 // [3] - zero low half of destination
9598 // [5:4] - select 128 bits from sources for high half of destination
9600 // [7] - zero high half of destination
9602 int MaskLO = Mask[0];
9603 if (MaskLO == SM_SentinelUndef)
9604 MaskLO = Mask[1] == SM_SentinelUndef ? 0 : Mask[1];
9606 int MaskHI = Mask[2];
9607 if (MaskHI == SM_SentinelUndef)
9608 MaskHI = Mask[3] == SM_SentinelUndef ? 0 : Mask[3];
9610 unsigned PermMask = MaskLO / 2 | (MaskHI / 2) << 4;
9612 // If either input is a zero vector, replace it with an undef input.
9613 // Shuffle mask values < 4 are selecting elements of V1.
9614 // Shuffle mask values >= 4 are selecting elements of V2.
9615 // Adjust each half of the permute mask by clearing the half that was
9616 // selecting the zero vector and setting the zero mask bit.
9618 V1 = DAG.getUNDEF(VT);
9620 PermMask = (PermMask & 0xf0) | 0x08;
9622 PermMask = (PermMask & 0x0f) | 0x80;
9625 V2 = DAG.getUNDEF(VT);
9627 PermMask = (PermMask & 0xf0) | 0x08;
9629 PermMask = (PermMask & 0x0f) | 0x80;
9632 return DAG.getNode(X86ISD::VPERM2X128, DL, VT, V1, V2,
9633 DAG.getConstant(PermMask, DL, MVT::i8));
9636 /// \brief Lower a vector shuffle by first fixing the 128-bit lanes and then
9637 /// shuffling each lane.
9639 /// This will only succeed when the result of fixing the 128-bit lanes results
9640 /// in a single-input non-lane-crossing shuffle with a repeating shuffle mask in
9641 /// each 128-bit lanes. This handles many cases where we can quickly blend away
9642 /// the lane crosses early and then use simpler shuffles within each lane.
9644 /// FIXME: It might be worthwhile at some point to support this without
9645 /// requiring the 128-bit lane-relative shuffles to be repeating, but currently
9646 /// in x86 only floating point has interesting non-repeating shuffles, and even
9647 /// those are still *marginally* more expensive.
9648 static SDValue lowerVectorShuffleByMerging128BitLanes(
9649 SDLoc DL, MVT VT, SDValue V1, SDValue V2, ArrayRef<int> Mask,
9650 const X86Subtarget *Subtarget, SelectionDAG &DAG) {
9651 assert(!isSingleInputShuffleMask(Mask) &&
9652 "This is only useful with multiple inputs.");
9654 int Size = Mask.size();
9655 int LaneSize = 128 / VT.getScalarSizeInBits();
9656 int NumLanes = Size / LaneSize;
9657 assert(NumLanes > 1 && "Only handles 256-bit and wider shuffles.");
9659 // See if we can build a hypothetical 128-bit lane-fixing shuffle mask. Also
9660 // check whether the in-128-bit lane shuffles share a repeating pattern.
9661 SmallVector<int, 4> Lanes;
9662 Lanes.resize(NumLanes, -1);
9663 SmallVector<int, 4> InLaneMask;
9664 InLaneMask.resize(LaneSize, -1);
9665 for (int i = 0; i < Size; ++i) {
9669 int j = i / LaneSize;
9672 // First entry we've seen for this lane.
9673 Lanes[j] = Mask[i] / LaneSize;
9674 } else if (Lanes[j] != Mask[i] / LaneSize) {
9675 // This doesn't match the lane selected previously!
9679 // Check that within each lane we have a consistent shuffle mask.
9680 int k = i % LaneSize;
9681 if (InLaneMask[k] < 0) {
9682 InLaneMask[k] = Mask[i] % LaneSize;
9683 } else if (InLaneMask[k] != Mask[i] % LaneSize) {
9684 // This doesn't fit a repeating in-lane mask.
9689 // First shuffle the lanes into place.
9690 MVT LaneVT = MVT::getVectorVT(VT.isFloatingPoint() ? MVT::f64 : MVT::i64,
9691 VT.getSizeInBits() / 64);
9692 SmallVector<int, 8> LaneMask;
9693 LaneMask.resize(NumLanes * 2, -1);
9694 for (int i = 0; i < NumLanes; ++i)
9695 if (Lanes[i] >= 0) {
9696 LaneMask[2 * i + 0] = 2*Lanes[i] + 0;
9697 LaneMask[2 * i + 1] = 2*Lanes[i] + 1;
9700 V1 = DAG.getBitcast(LaneVT, V1);
9701 V2 = DAG.getBitcast(LaneVT, V2);
9702 SDValue LaneShuffle = DAG.getVectorShuffle(LaneVT, DL, V1, V2, LaneMask);
9704 // Cast it back to the type we actually want.
9705 LaneShuffle = DAG.getBitcast(VT, LaneShuffle);
9707 // Now do a simple shuffle that isn't lane crossing.
9708 SmallVector<int, 8> NewMask;
9709 NewMask.resize(Size, -1);
9710 for (int i = 0; i < Size; ++i)
9712 NewMask[i] = (i / LaneSize) * LaneSize + Mask[i] % LaneSize;
9713 assert(!is128BitLaneCrossingShuffleMask(VT, NewMask) &&
9714 "Must not introduce lane crosses at this point!");
9716 return DAG.getVectorShuffle(VT, DL, LaneShuffle, DAG.getUNDEF(VT), NewMask);
9719 /// \brief Test whether the specified input (0 or 1) is in-place blended by the
9722 /// This returns true if the elements from a particular input are already in the
9723 /// slot required by the given mask and require no permutation.
9724 static bool isShuffleMaskInputInPlace(int Input, ArrayRef<int> Mask) {
9725 assert((Input == 0 || Input == 1) && "Only two inputs to shuffles.");
9726 int Size = Mask.size();
9727 for (int i = 0; i < Size; ++i)
9728 if (Mask[i] >= 0 && Mask[i] / Size == Input && Mask[i] % Size != i)
9734 static SDValue lowerVectorShuffleWithSHUFPD(SDLoc DL, MVT VT,
9735 ArrayRef<int> Mask, SDValue V1,
9736 SDValue V2, SelectionDAG &DAG) {
9738 // Mask for V8F64: 0/1, 8/9, 2/3, 10/11, 4/5, ..
9739 // Mask for V4F64; 0/1, 4/5, 2/3, 6/7..
9740 assert(VT.getScalarSizeInBits() == 64 && "Unexpected data type for VSHUFPD");
9741 int NumElts = VT.getVectorNumElements();
9742 bool ShufpdMask = true;
9743 bool CommutableMask = true;
9744 unsigned Immediate = 0;
9745 for (int i = 0; i < NumElts; ++i) {
9748 int Val = (i & 6) + NumElts * (i & 1);
9749 int CommutVal = (i & 0xe) + NumElts * ((i & 1)^1);
9750 if (Mask[i] < Val || Mask[i] > Val + 1)
9752 if (Mask[i] < CommutVal || Mask[i] > CommutVal + 1)
9753 CommutableMask = false;
9754 Immediate |= (Mask[i] % 2) << i;
9757 return DAG.getNode(X86ISD::SHUFP, DL, VT, V1, V2,
9758 DAG.getConstant(Immediate, DL, MVT::i8));
9760 return DAG.getNode(X86ISD::SHUFP, DL, VT, V2, V1,
9761 DAG.getConstant(Immediate, DL, MVT::i8));
9765 /// \brief Handle lowering of 4-lane 64-bit floating point shuffles.
9767 /// Also ends up handling lowering of 4-lane 64-bit integer shuffles when AVX2
9768 /// isn't available.
9769 static SDValue lowerV4F64VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
9770 const X86Subtarget *Subtarget,
9771 SelectionDAG &DAG) {
9773 assert(V1.getSimpleValueType() == MVT::v4f64 && "Bad operand type!");
9774 assert(V2.getSimpleValueType() == MVT::v4f64 && "Bad operand type!");
9775 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
9776 ArrayRef<int> Mask = SVOp->getMask();
9777 assert(Mask.size() == 4 && "Unexpected mask size for v4 shuffle!");
9779 SmallVector<int, 4> WidenedMask;
9780 if (canWidenShuffleElements(Mask, WidenedMask))
9781 return lowerV2X128VectorShuffle(DL, MVT::v4f64, V1, V2, Mask, Subtarget,
9784 if (isSingleInputShuffleMask(Mask)) {
9785 // Check for being able to broadcast a single element.
9786 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v4f64, V1,
9787 Mask, Subtarget, DAG))
9790 // Use low duplicate instructions for masks that match their pattern.
9791 if (isShuffleEquivalent(V1, V2, Mask, {0, 0, 2, 2}))
9792 return DAG.getNode(X86ISD::MOVDDUP, DL, MVT::v4f64, V1);
9794 if (!is128BitLaneCrossingShuffleMask(MVT::v4f64, Mask)) {
9795 // Non-half-crossing single input shuffles can be lowerid with an
9796 // interleaved permutation.
9797 unsigned VPERMILPMask = (Mask[0] == 1) | ((Mask[1] == 1) << 1) |
9798 ((Mask[2] == 3) << 2) | ((Mask[3] == 3) << 3);
9799 return DAG.getNode(X86ISD::VPERMILPI, DL, MVT::v4f64, V1,
9800 DAG.getConstant(VPERMILPMask, DL, MVT::i8));
9803 // With AVX2 we have direct support for this permutation.
9804 if (Subtarget->hasAVX2())
9805 return DAG.getNode(X86ISD::VPERMI, DL, MVT::v4f64, V1,
9806 getV4X86ShuffleImm8ForMask(Mask, DL, DAG));
9808 // Otherwise, fall back.
9809 return lowerVectorShuffleAsLanePermuteAndBlend(DL, MVT::v4f64, V1, V2, Mask,
9813 // X86 has dedicated unpack instructions that can handle specific blend
9814 // operations: UNPCKH and UNPCKL.
9815 if (isShuffleEquivalent(V1, V2, Mask, {0, 4, 2, 6}))
9816 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v4f64, V1, V2);
9817 if (isShuffleEquivalent(V1, V2, Mask, {1, 5, 3, 7}))
9818 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v4f64, V1, V2);
9819 if (isShuffleEquivalent(V1, V2, Mask, {4, 0, 6, 2}))
9820 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v4f64, V2, V1);
9821 if (isShuffleEquivalent(V1, V2, Mask, {5, 1, 7, 3}))
9822 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v4f64, V2, V1);
9824 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v4f64, V1, V2, Mask,
9828 // Check if the blend happens to exactly fit that of SHUFPD.
9830 lowerVectorShuffleWithSHUFPD(DL, MVT::v4f64, Mask, V1, V2, DAG))
9833 // Try to simplify this by merging 128-bit lanes to enable a lane-based
9834 // shuffle. However, if we have AVX2 and either inputs are already in place,
9835 // we will be able to shuffle even across lanes the other input in a single
9836 // instruction so skip this pattern.
9837 if (!(Subtarget->hasAVX2() && (isShuffleMaskInputInPlace(0, Mask) ||
9838 isShuffleMaskInputInPlace(1, Mask))))
9839 if (SDValue Result = lowerVectorShuffleByMerging128BitLanes(
9840 DL, MVT::v4f64, V1, V2, Mask, Subtarget, DAG))
9843 // If we have AVX2 then we always want to lower with a blend because an v4 we
9844 // can fully permute the elements.
9845 if (Subtarget->hasAVX2())
9846 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v4f64, V1, V2,
9849 // Otherwise fall back on generic lowering.
9850 return lowerVectorShuffleAsSplitOrBlend(DL, MVT::v4f64, V1, V2, Mask, DAG);
9853 /// \brief Handle lowering of 4-lane 64-bit integer shuffles.
9855 /// This routine is only called when we have AVX2 and thus a reasonable
9856 /// instruction set for v4i64 shuffling..
9857 static SDValue lowerV4I64VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
9858 const X86Subtarget *Subtarget,
9859 SelectionDAG &DAG) {
9861 assert(V1.getSimpleValueType() == MVT::v4i64 && "Bad operand type!");
9862 assert(V2.getSimpleValueType() == MVT::v4i64 && "Bad operand type!");
9863 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
9864 ArrayRef<int> Mask = SVOp->getMask();
9865 assert(Mask.size() == 4 && "Unexpected mask size for v4 shuffle!");
9866 assert(Subtarget->hasAVX2() && "We can only lower v4i64 with AVX2!");
9868 SmallVector<int, 4> WidenedMask;
9869 if (canWidenShuffleElements(Mask, WidenedMask))
9870 return lowerV2X128VectorShuffle(DL, MVT::v4i64, V1, V2, Mask, Subtarget,
9873 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v4i64, V1, V2, Mask,
9877 // Check for being able to broadcast a single element.
9878 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v4i64, V1,
9879 Mask, Subtarget, DAG))
9882 // When the shuffle is mirrored between the 128-bit lanes of the unit, we can
9883 // use lower latency instructions that will operate on both 128-bit lanes.
9884 SmallVector<int, 2> RepeatedMask;
9885 if (is128BitLaneRepeatedShuffleMask(MVT::v4i64, Mask, RepeatedMask)) {
9886 if (isSingleInputShuffleMask(Mask)) {
9887 int PSHUFDMask[] = {-1, -1, -1, -1};
9888 for (int i = 0; i < 2; ++i)
9889 if (RepeatedMask[i] >= 0) {
9890 PSHUFDMask[2 * i] = 2 * RepeatedMask[i];
9891 PSHUFDMask[2 * i + 1] = 2 * RepeatedMask[i] + 1;
9893 return DAG.getBitcast(
9895 DAG.getNode(X86ISD::PSHUFD, DL, MVT::v8i32,
9896 DAG.getBitcast(MVT::v8i32, V1),
9897 getV4X86ShuffleImm8ForMask(PSHUFDMask, DL, DAG)));
9901 // AVX2 provides a direct instruction for permuting a single input across
9903 if (isSingleInputShuffleMask(Mask))
9904 return DAG.getNode(X86ISD::VPERMI, DL, MVT::v4i64, V1,
9905 getV4X86ShuffleImm8ForMask(Mask, DL, DAG));
9907 // Try to use shift instructions.
9909 lowerVectorShuffleAsShift(DL, MVT::v4i64, V1, V2, Mask, DAG))
9912 // Use dedicated unpack instructions for masks that match their pattern.
9913 if (isShuffleEquivalent(V1, V2, Mask, {0, 4, 2, 6}))
9914 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v4i64, V1, V2);
9915 if (isShuffleEquivalent(V1, V2, Mask, {1, 5, 3, 7}))
9916 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v4i64, V1, V2);
9917 if (isShuffleEquivalent(V1, V2, Mask, {4, 0, 6, 2}))
9918 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v4i64, V2, V1);
9919 if (isShuffleEquivalent(V1, V2, Mask, {5, 1, 7, 3}))
9920 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v4i64, V2, V1);
9922 // Try to simplify this by merging 128-bit lanes to enable a lane-based
9923 // shuffle. However, if we have AVX2 and either inputs are already in place,
9924 // we will be able to shuffle even across lanes the other input in a single
9925 // instruction so skip this pattern.
9926 if (!(Subtarget->hasAVX2() && (isShuffleMaskInputInPlace(0, Mask) ||
9927 isShuffleMaskInputInPlace(1, Mask))))
9928 if (SDValue Result = lowerVectorShuffleByMerging128BitLanes(
9929 DL, MVT::v4i64, V1, V2, Mask, Subtarget, DAG))
9932 // Otherwise fall back on generic blend lowering.
9933 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v4i64, V1, V2,
9937 /// \brief Handle lowering of 8-lane 32-bit floating point shuffles.
9939 /// Also ends up handling lowering of 8-lane 32-bit integer shuffles when AVX2
9940 /// isn't available.
9941 static SDValue lowerV8F32VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
9942 const X86Subtarget *Subtarget,
9943 SelectionDAG &DAG) {
9945 assert(V1.getSimpleValueType() == MVT::v8f32 && "Bad operand type!");
9946 assert(V2.getSimpleValueType() == MVT::v8f32 && "Bad operand type!");
9947 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
9948 ArrayRef<int> Mask = SVOp->getMask();
9949 assert(Mask.size() == 8 && "Unexpected mask size for v8 shuffle!");
9951 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v8f32, V1, V2, Mask,
9955 // Check for being able to broadcast a single element.
9956 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v8f32, V1,
9957 Mask, Subtarget, DAG))
9960 // If the shuffle mask is repeated in each 128-bit lane, we have many more
9961 // options to efficiently lower the shuffle.
9962 SmallVector<int, 4> RepeatedMask;
9963 if (is128BitLaneRepeatedShuffleMask(MVT::v8f32, Mask, RepeatedMask)) {
9964 assert(RepeatedMask.size() == 4 &&
9965 "Repeated masks must be half the mask width!");
9967 // Use even/odd duplicate instructions for masks that match their pattern.
9968 if (isShuffleEquivalent(V1, V2, Mask, {0, 0, 2, 2, 4, 4, 6, 6}))
9969 return DAG.getNode(X86ISD::MOVSLDUP, DL, MVT::v8f32, V1);
9970 if (isShuffleEquivalent(V1, V2, Mask, {1, 1, 3, 3, 5, 5, 7, 7}))
9971 return DAG.getNode(X86ISD::MOVSHDUP, DL, MVT::v8f32, V1);
9973 if (isSingleInputShuffleMask(Mask))
9974 return DAG.getNode(X86ISD::VPERMILPI, DL, MVT::v8f32, V1,
9975 getV4X86ShuffleImm8ForMask(RepeatedMask, DL, DAG));
9977 // Use dedicated unpack instructions for masks that match their pattern.
9978 if (isShuffleEquivalent(V1, V2, Mask, {0, 8, 1, 9, 4, 12, 5, 13}))
9979 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v8f32, V1, V2);
9980 if (isShuffleEquivalent(V1, V2, Mask, {2, 10, 3, 11, 6, 14, 7, 15}))
9981 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v8f32, V1, V2);
9982 if (isShuffleEquivalent(V1, V2, Mask, {8, 0, 9, 1, 12, 4, 13, 5}))
9983 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v8f32, V2, V1);
9984 if (isShuffleEquivalent(V1, V2, Mask, {10, 2, 11, 3, 14, 6, 15, 7}))
9985 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v8f32, V2, V1);
9987 // Otherwise, fall back to a SHUFPS sequence. Here it is important that we
9988 // have already handled any direct blends. We also need to squash the
9989 // repeated mask into a simulated v4f32 mask.
9990 for (int i = 0; i < 4; ++i)
9991 if (RepeatedMask[i] >= 8)
9992 RepeatedMask[i] -= 4;
9993 return lowerVectorShuffleWithSHUFPS(DL, MVT::v8f32, RepeatedMask, V1, V2, DAG);
9996 // If we have a single input shuffle with different shuffle patterns in the
9997 // two 128-bit lanes use the variable mask to VPERMILPS.
9998 if (isSingleInputShuffleMask(Mask)) {
9999 SDValue VPermMask[8];
10000 for (int i = 0; i < 8; ++i)
10001 VPermMask[i] = Mask[i] < 0 ? DAG.getUNDEF(MVT::i32)
10002 : DAG.getConstant(Mask[i], DL, MVT::i32);
10003 if (!is128BitLaneCrossingShuffleMask(MVT::v8f32, Mask))
10004 return DAG.getNode(
10005 X86ISD::VPERMILPV, DL, MVT::v8f32, V1,
10006 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v8i32, VPermMask));
10008 if (Subtarget->hasAVX2())
10009 return DAG.getNode(
10010 X86ISD::VPERMV, DL, MVT::v8f32,
10011 DAG.getBitcast(MVT::v8f32, DAG.getNode(ISD::BUILD_VECTOR, DL,
10012 MVT::v8i32, VPermMask)),
10015 // Otherwise, fall back.
10016 return lowerVectorShuffleAsLanePermuteAndBlend(DL, MVT::v8f32, V1, V2, Mask,
10020 // Try to simplify this by merging 128-bit lanes to enable a lane-based
10022 if (SDValue Result = lowerVectorShuffleByMerging128BitLanes(
10023 DL, MVT::v8f32, V1, V2, Mask, Subtarget, DAG))
10026 // If we have AVX2 then we always want to lower with a blend because at v8 we
10027 // can fully permute the elements.
10028 if (Subtarget->hasAVX2())
10029 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v8f32, V1, V2,
10032 // Otherwise fall back on generic lowering.
10033 return lowerVectorShuffleAsSplitOrBlend(DL, MVT::v8f32, V1, V2, Mask, DAG);
10036 /// \brief Handle lowering of 8-lane 32-bit integer shuffles.
10038 /// This routine is only called when we have AVX2 and thus a reasonable
10039 /// instruction set for v8i32 shuffling..
10040 static SDValue lowerV8I32VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10041 const X86Subtarget *Subtarget,
10042 SelectionDAG &DAG) {
10044 assert(V1.getSimpleValueType() == MVT::v8i32 && "Bad operand type!");
10045 assert(V2.getSimpleValueType() == MVT::v8i32 && "Bad operand type!");
10046 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10047 ArrayRef<int> Mask = SVOp->getMask();
10048 assert(Mask.size() == 8 && "Unexpected mask size for v8 shuffle!");
10049 assert(Subtarget->hasAVX2() && "We can only lower v8i32 with AVX2!");
10051 // Whenever we can lower this as a zext, that instruction is strictly faster
10052 // than any alternative. It also allows us to fold memory operands into the
10053 // shuffle in many cases.
10054 if (SDValue ZExt = lowerVectorShuffleAsZeroOrAnyExtend(DL, MVT::v8i32, V1, V2,
10055 Mask, Subtarget, DAG))
10058 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v8i32, V1, V2, Mask,
10062 // Check for being able to broadcast a single element.
10063 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v8i32, V1,
10064 Mask, Subtarget, DAG))
10067 // If the shuffle mask is repeated in each 128-bit lane we can use more
10068 // efficient instructions that mirror the shuffles across the two 128-bit
10070 SmallVector<int, 4> RepeatedMask;
10071 if (is128BitLaneRepeatedShuffleMask(MVT::v8i32, Mask, RepeatedMask)) {
10072 assert(RepeatedMask.size() == 4 && "Unexpected repeated mask size!");
10073 if (isSingleInputShuffleMask(Mask))
10074 return DAG.getNode(X86ISD::PSHUFD, DL, MVT::v8i32, V1,
10075 getV4X86ShuffleImm8ForMask(RepeatedMask, DL, DAG));
10077 // Use dedicated unpack instructions for masks that match their pattern.
10078 if (isShuffleEquivalent(V1, V2, Mask, {0, 8, 1, 9, 4, 12, 5, 13}))
10079 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v8i32, V1, V2);
10080 if (isShuffleEquivalent(V1, V2, Mask, {2, 10, 3, 11, 6, 14, 7, 15}))
10081 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v8i32, V1, V2);
10082 if (isShuffleEquivalent(V1, V2, Mask, {8, 0, 9, 1, 12, 4, 13, 5}))
10083 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v8i32, V2, V1);
10084 if (isShuffleEquivalent(V1, V2, Mask, {10, 2, 11, 3, 14, 6, 15, 7}))
10085 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v8i32, V2, V1);
10088 // Try to use shift instructions.
10089 if (SDValue Shift =
10090 lowerVectorShuffleAsShift(DL, MVT::v8i32, V1, V2, Mask, DAG))
10093 if (SDValue Rotate = lowerVectorShuffleAsByteRotate(
10094 DL, MVT::v8i32, V1, V2, Mask, Subtarget, DAG))
10097 // If the shuffle patterns aren't repeated but it is a single input, directly
10098 // generate a cross-lane VPERMD instruction.
10099 if (isSingleInputShuffleMask(Mask)) {
10100 SDValue VPermMask[8];
10101 for (int i = 0; i < 8; ++i)
10102 VPermMask[i] = Mask[i] < 0 ? DAG.getUNDEF(MVT::i32)
10103 : DAG.getConstant(Mask[i], DL, MVT::i32);
10104 return DAG.getNode(
10105 X86ISD::VPERMV, DL, MVT::v8i32,
10106 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v8i32, VPermMask), V1);
10109 // Try to simplify this by merging 128-bit lanes to enable a lane-based
10111 if (SDValue Result = lowerVectorShuffleByMerging128BitLanes(
10112 DL, MVT::v8i32, V1, V2, Mask, Subtarget, DAG))
10115 // Otherwise fall back on generic blend lowering.
10116 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v8i32, V1, V2,
10120 /// \brief Handle lowering of 16-lane 16-bit integer shuffles.
10122 /// This routine is only called when we have AVX2 and thus a reasonable
10123 /// instruction set for v16i16 shuffling..
10124 static SDValue lowerV16I16VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10125 const X86Subtarget *Subtarget,
10126 SelectionDAG &DAG) {
10128 assert(V1.getSimpleValueType() == MVT::v16i16 && "Bad operand type!");
10129 assert(V2.getSimpleValueType() == MVT::v16i16 && "Bad operand type!");
10130 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10131 ArrayRef<int> Mask = SVOp->getMask();
10132 assert(Mask.size() == 16 && "Unexpected mask size for v16 shuffle!");
10133 assert(Subtarget->hasAVX2() && "We can only lower v16i16 with AVX2!");
10135 // Whenever we can lower this as a zext, that instruction is strictly faster
10136 // than any alternative. It also allows us to fold memory operands into the
10137 // shuffle in many cases.
10138 if (SDValue ZExt = lowerVectorShuffleAsZeroOrAnyExtend(DL, MVT::v16i16, V1, V2,
10139 Mask, Subtarget, DAG))
10142 // Check for being able to broadcast a single element.
10143 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v16i16, V1,
10144 Mask, Subtarget, DAG))
10147 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v16i16, V1, V2, Mask,
10151 // Use dedicated unpack instructions for masks that match their pattern.
10152 if (isShuffleEquivalent(V1, V2, Mask,
10153 {// First 128-bit lane:
10154 0, 16, 1, 17, 2, 18, 3, 19,
10155 // Second 128-bit lane:
10156 8, 24, 9, 25, 10, 26, 11, 27}))
10157 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v16i16, V1, V2);
10158 if (isShuffleEquivalent(V1, V2, Mask,
10159 {// First 128-bit lane:
10160 4, 20, 5, 21, 6, 22, 7, 23,
10161 // Second 128-bit lane:
10162 12, 28, 13, 29, 14, 30, 15, 31}))
10163 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v16i16, V1, V2);
10165 // Try to use shift instructions.
10166 if (SDValue Shift =
10167 lowerVectorShuffleAsShift(DL, MVT::v16i16, V1, V2, Mask, DAG))
10170 // Try to use byte rotation instructions.
10171 if (SDValue Rotate = lowerVectorShuffleAsByteRotate(
10172 DL, MVT::v16i16, V1, V2, Mask, Subtarget, DAG))
10175 if (isSingleInputShuffleMask(Mask)) {
10176 // There are no generalized cross-lane shuffle operations available on i16
10178 if (is128BitLaneCrossingShuffleMask(MVT::v16i16, Mask))
10179 return lowerVectorShuffleAsLanePermuteAndBlend(DL, MVT::v16i16, V1, V2,
10182 SmallVector<int, 8> RepeatedMask;
10183 if (is128BitLaneRepeatedShuffleMask(MVT::v16i16, Mask, RepeatedMask)) {
10184 // As this is a single-input shuffle, the repeated mask should be
10185 // a strictly valid v8i16 mask that we can pass through to the v8i16
10186 // lowering to handle even the v16 case.
10187 return lowerV8I16GeneralSingleInputVectorShuffle(
10188 DL, MVT::v16i16, V1, RepeatedMask, Subtarget, DAG);
10191 SDValue PSHUFBMask[32];
10192 for (int i = 0; i < 16; ++i) {
10193 if (Mask[i] == -1) {
10194 PSHUFBMask[2 * i] = PSHUFBMask[2 * i + 1] = DAG.getUNDEF(MVT::i8);
10198 int M = i < 8 ? Mask[i] : Mask[i] - 8;
10199 assert(M >= 0 && M < 8 && "Invalid single-input mask!");
10200 PSHUFBMask[2 * i] = DAG.getConstant(2 * M, DL, MVT::i8);
10201 PSHUFBMask[2 * i + 1] = DAG.getConstant(2 * M + 1, DL, MVT::i8);
10203 return DAG.getBitcast(MVT::v16i16,
10204 DAG.getNode(X86ISD::PSHUFB, DL, MVT::v32i8,
10205 DAG.getBitcast(MVT::v32i8, V1),
10206 DAG.getNode(ISD::BUILD_VECTOR, DL,
10207 MVT::v32i8, PSHUFBMask)));
10210 // Try to simplify this by merging 128-bit lanes to enable a lane-based
10212 if (SDValue Result = lowerVectorShuffleByMerging128BitLanes(
10213 DL, MVT::v16i16, V1, V2, Mask, Subtarget, DAG))
10216 // Otherwise fall back on generic lowering.
10217 return lowerVectorShuffleAsSplitOrBlend(DL, MVT::v16i16, V1, V2, Mask, DAG);
10220 /// \brief Handle lowering of 32-lane 8-bit integer shuffles.
10222 /// This routine is only called when we have AVX2 and thus a reasonable
10223 /// instruction set for v32i8 shuffling..
10224 static SDValue lowerV32I8VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10225 const X86Subtarget *Subtarget,
10226 SelectionDAG &DAG) {
10228 assert(V1.getSimpleValueType() == MVT::v32i8 && "Bad operand type!");
10229 assert(V2.getSimpleValueType() == MVT::v32i8 && "Bad operand type!");
10230 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10231 ArrayRef<int> Mask = SVOp->getMask();
10232 assert(Mask.size() == 32 && "Unexpected mask size for v32 shuffle!");
10233 assert(Subtarget->hasAVX2() && "We can only lower v32i8 with AVX2!");
10235 // Whenever we can lower this as a zext, that instruction is strictly faster
10236 // than any alternative. It also allows us to fold memory operands into the
10237 // shuffle in many cases.
10238 if (SDValue ZExt = lowerVectorShuffleAsZeroOrAnyExtend(DL, MVT::v32i8, V1, V2,
10239 Mask, Subtarget, DAG))
10242 // Check for being able to broadcast a single element.
10243 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v32i8, V1,
10244 Mask, Subtarget, DAG))
10247 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v32i8, V1, V2, Mask,
10251 // Use dedicated unpack instructions for masks that match their pattern.
10252 // Note that these are repeated 128-bit lane unpacks, not unpacks across all
10254 if (isShuffleEquivalent(
10256 {// First 128-bit lane:
10257 0, 32, 1, 33, 2, 34, 3, 35, 4, 36, 5, 37, 6, 38, 7, 39,
10258 // Second 128-bit lane:
10259 16, 48, 17, 49, 18, 50, 19, 51, 20, 52, 21, 53, 22, 54, 23, 55}))
10260 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v32i8, V1, V2);
10261 if (isShuffleEquivalent(
10263 {// First 128-bit lane:
10264 8, 40, 9, 41, 10, 42, 11, 43, 12, 44, 13, 45, 14, 46, 15, 47,
10265 // Second 128-bit lane:
10266 24, 56, 25, 57, 26, 58, 27, 59, 28, 60, 29, 61, 30, 62, 31, 63}))
10267 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v32i8, V1, V2);
10269 // Try to use shift instructions.
10270 if (SDValue Shift =
10271 lowerVectorShuffleAsShift(DL, MVT::v32i8, V1, V2, Mask, DAG))
10274 // Try to use byte rotation instructions.
10275 if (SDValue Rotate = lowerVectorShuffleAsByteRotate(
10276 DL, MVT::v32i8, V1, V2, Mask, Subtarget, DAG))
10279 if (isSingleInputShuffleMask(Mask)) {
10280 // There are no generalized cross-lane shuffle operations available on i8
10282 if (is128BitLaneCrossingShuffleMask(MVT::v32i8, Mask))
10283 return lowerVectorShuffleAsLanePermuteAndBlend(DL, MVT::v32i8, V1, V2,
10286 SDValue PSHUFBMask[32];
10287 for (int i = 0; i < 32; ++i)
10290 ? DAG.getUNDEF(MVT::i8)
10291 : DAG.getConstant(Mask[i] < 16 ? Mask[i] : Mask[i] - 16, DL,
10294 return DAG.getNode(
10295 X86ISD::PSHUFB, DL, MVT::v32i8, V1,
10296 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v32i8, PSHUFBMask));
10299 // Try to simplify this by merging 128-bit lanes to enable a lane-based
10301 if (SDValue Result = lowerVectorShuffleByMerging128BitLanes(
10302 DL, MVT::v32i8, V1, V2, Mask, Subtarget, DAG))
10305 // Otherwise fall back on generic lowering.
10306 return lowerVectorShuffleAsSplitOrBlend(DL, MVT::v32i8, V1, V2, Mask, DAG);
10309 /// \brief High-level routine to lower various 256-bit x86 vector shuffles.
10311 /// This routine either breaks down the specific type of a 256-bit x86 vector
10312 /// shuffle or splits it into two 128-bit shuffles and fuses the results back
10313 /// together based on the available instructions.
10314 static SDValue lower256BitVectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10315 MVT VT, const X86Subtarget *Subtarget,
10316 SelectionDAG &DAG) {
10318 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10319 ArrayRef<int> Mask = SVOp->getMask();
10321 // If we have a single input to the zero element, insert that into V1 if we
10322 // can do so cheaply.
10323 int NumElts = VT.getVectorNumElements();
10324 int NumV2Elements = std::count_if(Mask.begin(), Mask.end(), [NumElts](int M) {
10325 return M >= NumElts;
10328 if (NumV2Elements == 1 && Mask[0] >= NumElts)
10329 if (SDValue Insertion = lowerVectorShuffleAsElementInsertion(
10330 DL, VT, V1, V2, Mask, Subtarget, DAG))
10333 // There is a really nice hard cut-over between AVX1 and AVX2 that means we can
10334 // check for those subtargets here and avoid much of the subtarget querying in
10335 // the per-vector-type lowering routines. With AVX1 we have essentially *zero*
10336 // ability to manipulate a 256-bit vector with integer types. Since we'll use
10337 // floating point types there eventually, just immediately cast everything to
10338 // a float and operate entirely in that domain.
10339 if (VT.isInteger() && !Subtarget->hasAVX2()) {
10340 int ElementBits = VT.getScalarSizeInBits();
10341 if (ElementBits < 32)
10342 // No floating point type available, decompose into 128-bit vectors.
10343 return splitAndLowerVectorShuffle(DL, VT, V1, V2, Mask, DAG);
10345 MVT FpVT = MVT::getVectorVT(MVT::getFloatingPointVT(ElementBits),
10346 VT.getVectorNumElements());
10347 V1 = DAG.getBitcast(FpVT, V1);
10348 V2 = DAG.getBitcast(FpVT, V2);
10349 return DAG.getBitcast(VT, DAG.getVectorShuffle(FpVT, DL, V1, V2, Mask));
10352 switch (VT.SimpleTy) {
10354 return lowerV4F64VectorShuffle(Op, V1, V2, Subtarget, DAG);
10356 return lowerV4I64VectorShuffle(Op, V1, V2, Subtarget, DAG);
10358 return lowerV8F32VectorShuffle(Op, V1, V2, Subtarget, DAG);
10360 return lowerV8I32VectorShuffle(Op, V1, V2, Subtarget, DAG);
10362 return lowerV16I16VectorShuffle(Op, V1, V2, Subtarget, DAG);
10364 return lowerV32I8VectorShuffle(Op, V1, V2, Subtarget, DAG);
10367 llvm_unreachable("Not a valid 256-bit x86 vector type!");
10371 /// \brief Handle lowering of 8-lane 64-bit floating point shuffles.
10372 static SDValue lowerV8F64VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10373 const X86Subtarget *Subtarget,
10374 SelectionDAG &DAG) {
10376 assert(V1.getSimpleValueType() == MVT::v8f64 && "Bad operand type!");
10377 assert(V2.getSimpleValueType() == MVT::v8f64 && "Bad operand type!");
10378 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10379 ArrayRef<int> Mask = SVOp->getMask();
10380 assert(Mask.size() == 8 && "Unexpected mask size for v8 shuffle!");
10382 // X86 has dedicated unpack instructions that can handle specific blend
10383 // operations: UNPCKH and UNPCKL.
10384 if (isShuffleEquivalent(V1, V2, Mask, {0, 8, 2, 10, 4, 12, 6, 14}))
10385 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v8f64, V1, V2);
10386 if (isShuffleEquivalent(V1, V2, Mask, {1, 9, 3, 11, 5, 13, 7, 15}))
10387 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v8f64, V1, V2);
10389 // FIXME: Implement direct support for this type!
10390 return splitAndLowerVectorShuffle(DL, MVT::v8f64, V1, V2, Mask, DAG);
10393 /// \brief Handle lowering of 16-lane 32-bit floating point shuffles.
10394 static SDValue lowerV16F32VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10395 const X86Subtarget *Subtarget,
10396 SelectionDAG &DAG) {
10398 assert(V1.getSimpleValueType() == MVT::v16f32 && "Bad operand type!");
10399 assert(V2.getSimpleValueType() == MVT::v16f32 && "Bad operand type!");
10400 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10401 ArrayRef<int> Mask = SVOp->getMask();
10402 assert(Mask.size() == 16 && "Unexpected mask size for v16 shuffle!");
10404 // Use dedicated unpack instructions for masks that match their pattern.
10405 if (isShuffleEquivalent(V1, V2, Mask,
10406 {// First 128-bit lane.
10407 0, 16, 1, 17, 4, 20, 5, 21,
10408 // Second 128-bit lane.
10409 8, 24, 9, 25, 12, 28, 13, 29}))
10410 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v16f32, V1, V2);
10411 if (isShuffleEquivalent(V1, V2, Mask,
10412 {// First 128-bit lane.
10413 2, 18, 3, 19, 6, 22, 7, 23,
10414 // Second 128-bit lane.
10415 10, 26, 11, 27, 14, 30, 15, 31}))
10416 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v16f32, V1, V2);
10418 // FIXME: Implement direct support for this type!
10419 return splitAndLowerVectorShuffle(DL, MVT::v16f32, V1, V2, Mask, DAG);
10422 /// \brief Handle lowering of 8-lane 64-bit integer shuffles.
10423 static SDValue lowerV8I64VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10424 const X86Subtarget *Subtarget,
10425 SelectionDAG &DAG) {
10427 assert(V1.getSimpleValueType() == MVT::v8i64 && "Bad operand type!");
10428 assert(V2.getSimpleValueType() == MVT::v8i64 && "Bad operand type!");
10429 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10430 ArrayRef<int> Mask = SVOp->getMask();
10431 assert(Mask.size() == 8 && "Unexpected mask size for v8 shuffle!");
10433 // X86 has dedicated unpack instructions that can handle specific blend
10434 // operations: UNPCKH and UNPCKL.
10435 if (isShuffleEquivalent(V1, V2, Mask, {0, 8, 2, 10, 4, 12, 6, 14}))
10436 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v8i64, V1, V2);
10437 if (isShuffleEquivalent(V1, V2, Mask, {1, 9, 3, 11, 5, 13, 7, 15}))
10438 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v8i64, V1, V2);
10440 // FIXME: Implement direct support for this type!
10441 return splitAndLowerVectorShuffle(DL, MVT::v8i64, V1, V2, Mask, DAG);
10444 /// \brief Handle lowering of 16-lane 32-bit integer shuffles.
10445 static SDValue lowerV16I32VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10446 const X86Subtarget *Subtarget,
10447 SelectionDAG &DAG) {
10449 assert(V1.getSimpleValueType() == MVT::v16i32 && "Bad operand type!");
10450 assert(V2.getSimpleValueType() == MVT::v16i32 && "Bad operand type!");
10451 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10452 ArrayRef<int> Mask = SVOp->getMask();
10453 assert(Mask.size() == 16 && "Unexpected mask size for v16 shuffle!");
10455 // Use dedicated unpack instructions for masks that match their pattern.
10456 if (isShuffleEquivalent(V1, V2, Mask,
10457 {// First 128-bit lane.
10458 0, 16, 1, 17, 4, 20, 5, 21,
10459 // Second 128-bit lane.
10460 8, 24, 9, 25, 12, 28, 13, 29}))
10461 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v16i32, V1, V2);
10462 if (isShuffleEquivalent(V1, V2, Mask,
10463 {// First 128-bit lane.
10464 2, 18, 3, 19, 6, 22, 7, 23,
10465 // Second 128-bit lane.
10466 10, 26, 11, 27, 14, 30, 15, 31}))
10467 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v16i32, V1, V2);
10469 // FIXME: Implement direct support for this type!
10470 return splitAndLowerVectorShuffle(DL, MVT::v16i32, V1, V2, Mask, DAG);
10473 /// \brief Handle lowering of 32-lane 16-bit integer shuffles.
10474 static SDValue lowerV32I16VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10475 const X86Subtarget *Subtarget,
10476 SelectionDAG &DAG) {
10478 assert(V1.getSimpleValueType() == MVT::v32i16 && "Bad operand type!");
10479 assert(V2.getSimpleValueType() == MVT::v32i16 && "Bad operand type!");
10480 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10481 ArrayRef<int> Mask = SVOp->getMask();
10482 assert(Mask.size() == 32 && "Unexpected mask size for v32 shuffle!");
10483 assert(Subtarget->hasBWI() && "We can only lower v32i16 with AVX-512-BWI!");
10485 // FIXME: Implement direct support for this type!
10486 return splitAndLowerVectorShuffle(DL, MVT::v32i16, V1, V2, Mask, DAG);
10489 /// \brief Handle lowering of 64-lane 8-bit integer shuffles.
10490 static SDValue lowerV64I8VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10491 const X86Subtarget *Subtarget,
10492 SelectionDAG &DAG) {
10494 assert(V1.getSimpleValueType() == MVT::v64i8 && "Bad operand type!");
10495 assert(V2.getSimpleValueType() == MVT::v64i8 && "Bad operand type!");
10496 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10497 ArrayRef<int> Mask = SVOp->getMask();
10498 assert(Mask.size() == 64 && "Unexpected mask size for v64 shuffle!");
10499 assert(Subtarget->hasBWI() && "We can only lower v64i8 with AVX-512-BWI!");
10501 // FIXME: Implement direct support for this type!
10502 return splitAndLowerVectorShuffle(DL, MVT::v64i8, V1, V2, Mask, DAG);
10505 /// \brief High-level routine to lower various 512-bit x86 vector shuffles.
10507 /// This routine either breaks down the specific type of a 512-bit x86 vector
10508 /// shuffle or splits it into two 256-bit shuffles and fuses the results back
10509 /// together based on the available instructions.
10510 static SDValue lower512BitVectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10511 MVT VT, const X86Subtarget *Subtarget,
10512 SelectionDAG &DAG) {
10514 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10515 ArrayRef<int> Mask = SVOp->getMask();
10516 assert(Subtarget->hasAVX512() &&
10517 "Cannot lower 512-bit vectors w/ basic ISA!");
10519 // Check for being able to broadcast a single element.
10520 if (SDValue Broadcast =
10521 lowerVectorShuffleAsBroadcast(DL, VT, V1, Mask, Subtarget, DAG))
10524 // Dispatch to each element type for lowering. If we don't have supprot for
10525 // specific element type shuffles at 512 bits, immediately split them and
10526 // lower them. Each lowering routine of a given type is allowed to assume that
10527 // the requisite ISA extensions for that element type are available.
10528 switch (VT.SimpleTy) {
10530 return lowerV8F64VectorShuffle(Op, V1, V2, Subtarget, DAG);
10532 return lowerV16F32VectorShuffle(Op, V1, V2, Subtarget, DAG);
10534 return lowerV8I64VectorShuffle(Op, V1, V2, Subtarget, DAG);
10536 return lowerV16I32VectorShuffle(Op, V1, V2, Subtarget, DAG);
10538 if (Subtarget->hasBWI())
10539 return lowerV32I16VectorShuffle(Op, V1, V2, Subtarget, DAG);
10542 if (Subtarget->hasBWI())
10543 return lowerV64I8VectorShuffle(Op, V1, V2, Subtarget, DAG);
10547 llvm_unreachable("Not a valid 512-bit x86 vector type!");
10550 // Otherwise fall back on splitting.
10551 return splitAndLowerVectorShuffle(DL, VT, V1, V2, Mask, DAG);
10554 /// \brief Top-level lowering for x86 vector shuffles.
10556 /// This handles decomposition, canonicalization, and lowering of all x86
10557 /// vector shuffles. Most of the specific lowering strategies are encapsulated
10558 /// above in helper routines. The canonicalization attempts to widen shuffles
10559 /// to involve fewer lanes of wider elements, consolidate symmetric patterns
10560 /// s.t. only one of the two inputs needs to be tested, etc.
10561 static SDValue lowerVectorShuffle(SDValue Op, const X86Subtarget *Subtarget,
10562 SelectionDAG &DAG) {
10563 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10564 ArrayRef<int> Mask = SVOp->getMask();
10565 SDValue V1 = Op.getOperand(0);
10566 SDValue V2 = Op.getOperand(1);
10567 MVT VT = Op.getSimpleValueType();
10568 int NumElements = VT.getVectorNumElements();
10571 assert(VT.getSizeInBits() != 64 && "Can't lower MMX shuffles");
10573 bool V1IsUndef = V1.getOpcode() == ISD::UNDEF;
10574 bool V2IsUndef = V2.getOpcode() == ISD::UNDEF;
10575 if (V1IsUndef && V2IsUndef)
10576 return DAG.getUNDEF(VT);
10578 // When we create a shuffle node we put the UNDEF node to second operand,
10579 // but in some cases the first operand may be transformed to UNDEF.
10580 // In this case we should just commute the node.
10582 return DAG.getCommutedVectorShuffle(*SVOp);
10584 // Check for non-undef masks pointing at an undef vector and make the masks
10585 // undef as well. This makes it easier to match the shuffle based solely on
10589 if (M >= NumElements) {
10590 SmallVector<int, 8> NewMask(Mask.begin(), Mask.end());
10591 for (int &M : NewMask)
10592 if (M >= NumElements)
10594 return DAG.getVectorShuffle(VT, dl, V1, V2, NewMask);
10597 // We actually see shuffles that are entirely re-arrangements of a set of
10598 // zero inputs. This mostly happens while decomposing complex shuffles into
10599 // simple ones. Directly lower these as a buildvector of zeros.
10600 SmallBitVector Zeroable = computeZeroableShuffleElements(Mask, V1, V2);
10601 if (Zeroable.all())
10602 return getZeroVector(VT, Subtarget, DAG, dl);
10604 // Try to collapse shuffles into using a vector type with fewer elements but
10605 // wider element types. We cap this to not form integers or floating point
10606 // elements wider than 64 bits, but it might be interesting to form i128
10607 // integers to handle flipping the low and high halves of AVX 256-bit vectors.
10608 SmallVector<int, 16> WidenedMask;
10609 if (VT.getScalarSizeInBits() < 64 &&
10610 canWidenShuffleElements(Mask, WidenedMask)) {
10611 MVT NewEltVT = VT.isFloatingPoint()
10612 ? MVT::getFloatingPointVT(VT.getScalarSizeInBits() * 2)
10613 : MVT::getIntegerVT(VT.getScalarSizeInBits() * 2);
10614 MVT NewVT = MVT::getVectorVT(NewEltVT, VT.getVectorNumElements() / 2);
10615 // Make sure that the new vector type is legal. For example, v2f64 isn't
10617 if (DAG.getTargetLoweringInfo().isTypeLegal(NewVT)) {
10618 V1 = DAG.getBitcast(NewVT, V1);
10619 V2 = DAG.getBitcast(NewVT, V2);
10620 return DAG.getBitcast(
10621 VT, DAG.getVectorShuffle(NewVT, dl, V1, V2, WidenedMask));
10625 int NumV1Elements = 0, NumUndefElements = 0, NumV2Elements = 0;
10626 for (int M : SVOp->getMask())
10628 ++NumUndefElements;
10629 else if (M < NumElements)
10634 // Commute the shuffle as needed such that more elements come from V1 than
10635 // V2. This allows us to match the shuffle pattern strictly on how many
10636 // elements come from V1 without handling the symmetric cases.
10637 if (NumV2Elements > NumV1Elements)
10638 return DAG.getCommutedVectorShuffle(*SVOp);
10640 // When the number of V1 and V2 elements are the same, try to minimize the
10641 // number of uses of V2 in the low half of the vector. When that is tied,
10642 // ensure that the sum of indices for V1 is equal to or lower than the sum
10643 // indices for V2. When those are equal, try to ensure that the number of odd
10644 // indices for V1 is lower than the number of odd indices for V2.
10645 if (NumV1Elements == NumV2Elements) {
10646 int LowV1Elements = 0, LowV2Elements = 0;
10647 for (int M : SVOp->getMask().slice(0, NumElements / 2))
10648 if (M >= NumElements)
10652 if (LowV2Elements > LowV1Elements) {
10653 return DAG.getCommutedVectorShuffle(*SVOp);
10654 } else if (LowV2Elements == LowV1Elements) {
10655 int SumV1Indices = 0, SumV2Indices = 0;
10656 for (int i = 0, Size = SVOp->getMask().size(); i < Size; ++i)
10657 if (SVOp->getMask()[i] >= NumElements)
10659 else if (SVOp->getMask()[i] >= 0)
10661 if (SumV2Indices < SumV1Indices) {
10662 return DAG.getCommutedVectorShuffle(*SVOp);
10663 } else if (SumV2Indices == SumV1Indices) {
10664 int NumV1OddIndices = 0, NumV2OddIndices = 0;
10665 for (int i = 0, Size = SVOp->getMask().size(); i < Size; ++i)
10666 if (SVOp->getMask()[i] >= NumElements)
10667 NumV2OddIndices += i % 2;
10668 else if (SVOp->getMask()[i] >= 0)
10669 NumV1OddIndices += i % 2;
10670 if (NumV2OddIndices < NumV1OddIndices)
10671 return DAG.getCommutedVectorShuffle(*SVOp);
10676 // For each vector width, delegate to a specialized lowering routine.
10677 if (VT.getSizeInBits() == 128)
10678 return lower128BitVectorShuffle(Op, V1, V2, VT, Subtarget, DAG);
10680 if (VT.getSizeInBits() == 256)
10681 return lower256BitVectorShuffle(Op, V1, V2, VT, Subtarget, DAG);
10683 // Force AVX-512 vectors to be scalarized for now.
10684 // FIXME: Implement AVX-512 support!
10685 if (VT.getSizeInBits() == 512)
10686 return lower512BitVectorShuffle(Op, V1, V2, VT, Subtarget, DAG);
10688 llvm_unreachable("Unimplemented!");
10691 // This function assumes its argument is a BUILD_VECTOR of constants or
10692 // undef SDNodes. i.e: ISD::isBuildVectorOfConstantSDNodes(BuildVector) is
10694 static bool BUILD_VECTORtoBlendMask(BuildVectorSDNode *BuildVector,
10695 unsigned &MaskValue) {
10697 unsigned NumElems = BuildVector->getNumOperands();
10698 // There are 2 lanes if (NumElems > 8), and 1 lane otherwise.
10699 unsigned NumLanes = (NumElems - 1) / 8 + 1;
10700 unsigned NumElemsInLane = NumElems / NumLanes;
10702 // Blend for v16i16 should be symetric for the both lanes.
10703 for (unsigned i = 0; i < NumElemsInLane; ++i) {
10704 SDValue EltCond = BuildVector->getOperand(i);
10705 SDValue SndLaneEltCond =
10706 (NumLanes == 2) ? BuildVector->getOperand(i + NumElemsInLane) : EltCond;
10708 int Lane1Cond = -1, Lane2Cond = -1;
10709 if (isa<ConstantSDNode>(EltCond))
10710 Lane1Cond = !isZero(EltCond);
10711 if (isa<ConstantSDNode>(SndLaneEltCond))
10712 Lane2Cond = !isZero(SndLaneEltCond);
10714 if (Lane1Cond == Lane2Cond || Lane2Cond < 0)
10715 // Lane1Cond != 0, means we want the first argument.
10716 // Lane1Cond == 0, means we want the second argument.
10717 // The encoding of this argument is 0 for the first argument, 1
10718 // for the second. Therefore, invert the condition.
10719 MaskValue |= !Lane1Cond << i;
10720 else if (Lane1Cond < 0)
10721 MaskValue |= !Lane2Cond << i;
10728 /// \brief Try to lower a VSELECT instruction to a vector shuffle.
10729 static SDValue lowerVSELECTtoVectorShuffle(SDValue Op,
10730 const X86Subtarget *Subtarget,
10731 SelectionDAG &DAG) {
10732 SDValue Cond = Op.getOperand(0);
10733 SDValue LHS = Op.getOperand(1);
10734 SDValue RHS = Op.getOperand(2);
10736 MVT VT = Op.getSimpleValueType();
10738 if (!ISD::isBuildVectorOfConstantSDNodes(Cond.getNode()))
10740 auto *CondBV = cast<BuildVectorSDNode>(Cond);
10742 // Only non-legal VSELECTs reach this lowering, convert those into generic
10743 // shuffles and re-use the shuffle lowering path for blends.
10744 SmallVector<int, 32> Mask;
10745 for (int i = 0, Size = VT.getVectorNumElements(); i < Size; ++i) {
10746 SDValue CondElt = CondBV->getOperand(i);
10748 isa<ConstantSDNode>(CondElt) ? i + (isZero(CondElt) ? Size : 0) : -1);
10750 return DAG.getVectorShuffle(VT, dl, LHS, RHS, Mask);
10753 SDValue X86TargetLowering::LowerVSELECT(SDValue Op, SelectionDAG &DAG) const {
10754 // A vselect where all conditions and data are constants can be optimized into
10755 // a single vector load by SelectionDAGLegalize::ExpandBUILD_VECTOR().
10756 if (ISD::isBuildVectorOfConstantSDNodes(Op.getOperand(0).getNode()) &&
10757 ISD::isBuildVectorOfConstantSDNodes(Op.getOperand(1).getNode()) &&
10758 ISD::isBuildVectorOfConstantSDNodes(Op.getOperand(2).getNode()))
10761 // Try to lower this to a blend-style vector shuffle. This can handle all
10762 // constant condition cases.
10763 if (SDValue BlendOp = lowerVSELECTtoVectorShuffle(Op, Subtarget, DAG))
10766 // Variable blends are only legal from SSE4.1 onward.
10767 if (!Subtarget->hasSSE41())
10770 // Only some types will be legal on some subtargets. If we can emit a legal
10771 // VSELECT-matching blend, return Op, and but if we need to expand, return
10773 switch (Op.getSimpleValueType().SimpleTy) {
10775 // Most of the vector types have blends past SSE4.1.
10779 // The byte blends for AVX vectors were introduced only in AVX2.
10780 if (Subtarget->hasAVX2())
10787 // AVX-512 BWI and VLX features support VSELECT with i16 elements.
10788 if (Subtarget->hasBWI() && Subtarget->hasVLX())
10791 // FIXME: We should custom lower this by fixing the condition and using i8
10797 static SDValue LowerEXTRACT_VECTOR_ELT_SSE4(SDValue Op, SelectionDAG &DAG) {
10798 MVT VT = Op.getSimpleValueType();
10801 if (!Op.getOperand(0).getSimpleValueType().is128BitVector())
10804 if (VT.getSizeInBits() == 8) {
10805 SDValue Extract = DAG.getNode(X86ISD::PEXTRB, dl, MVT::i32,
10806 Op.getOperand(0), Op.getOperand(1));
10807 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Extract,
10808 DAG.getValueType(VT));
10809 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
10812 if (VT.getSizeInBits() == 16) {
10813 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
10814 // If Idx is 0, it's cheaper to do a move instead of a pextrw.
10816 return DAG.getNode(
10817 ISD::TRUNCATE, dl, MVT::i16,
10818 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
10819 DAG.getBitcast(MVT::v4i32, Op.getOperand(0)),
10820 Op.getOperand(1)));
10821 SDValue Extract = DAG.getNode(X86ISD::PEXTRW, dl, MVT::i32,
10822 Op.getOperand(0), Op.getOperand(1));
10823 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Extract,
10824 DAG.getValueType(VT));
10825 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
10828 if (VT == MVT::f32) {
10829 // EXTRACTPS outputs to a GPR32 register which will require a movd to copy
10830 // the result back to FR32 register. It's only worth matching if the
10831 // result has a single use which is a store or a bitcast to i32. And in
10832 // the case of a store, it's not worth it if the index is a constant 0,
10833 // because a MOVSSmr can be used instead, which is smaller and faster.
10834 if (!Op.hasOneUse())
10836 SDNode *User = *Op.getNode()->use_begin();
10837 if ((User->getOpcode() != ISD::STORE ||
10838 (isa<ConstantSDNode>(Op.getOperand(1)) &&
10839 cast<ConstantSDNode>(Op.getOperand(1))->isNullValue())) &&
10840 (User->getOpcode() != ISD::BITCAST ||
10841 User->getValueType(0) != MVT::i32))
10843 SDValue Extract = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
10844 DAG.getBitcast(MVT::v4i32, Op.getOperand(0)),
10846 return DAG.getBitcast(MVT::f32, Extract);
10849 if (VT == MVT::i32 || VT == MVT::i64) {
10850 // ExtractPS/pextrq works with constant index.
10851 if (isa<ConstantSDNode>(Op.getOperand(1)))
10857 /// Extract one bit from mask vector, like v16i1 or v8i1.
10858 /// AVX-512 feature.
10860 X86TargetLowering::ExtractBitFromMaskVector(SDValue Op, SelectionDAG &DAG) const {
10861 SDValue Vec = Op.getOperand(0);
10863 MVT VecVT = Vec.getSimpleValueType();
10864 SDValue Idx = Op.getOperand(1);
10865 MVT EltVT = Op.getSimpleValueType();
10867 assert((EltVT == MVT::i1) && "Unexpected operands in ExtractBitFromMaskVector");
10868 assert((VecVT.getVectorNumElements() <= 16 || Subtarget->hasBWI()) &&
10869 "Unexpected vector type in ExtractBitFromMaskVector");
10871 // variable index can't be handled in mask registers,
10872 // extend vector to VR512
10873 if (!isa<ConstantSDNode>(Idx)) {
10874 MVT ExtVT = (VecVT == MVT::v8i1 ? MVT::v8i64 : MVT::v16i32);
10875 SDValue Ext = DAG.getNode(ISD::ZERO_EXTEND, dl, ExtVT, Vec);
10876 SDValue Elt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl,
10877 ExtVT.getVectorElementType(), Ext, Idx);
10878 return DAG.getNode(ISD::TRUNCATE, dl, EltVT, Elt);
10881 unsigned IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
10882 const TargetRegisterClass* rc = getRegClassFor(VecVT);
10883 if (!Subtarget->hasDQI() && (VecVT.getVectorNumElements() <= 8))
10884 rc = getRegClassFor(MVT::v16i1);
10885 unsigned MaxSift = rc->getSize()*8 - 1;
10886 Vec = DAG.getNode(X86ISD::VSHLI, dl, VecVT, Vec,
10887 DAG.getConstant(MaxSift - IdxVal, dl, MVT::i8));
10888 Vec = DAG.getNode(X86ISD::VSRLI, dl, VecVT, Vec,
10889 DAG.getConstant(MaxSift, dl, MVT::i8));
10890 return DAG.getNode(X86ISD::VEXTRACT, dl, MVT::i1, Vec,
10891 DAG.getIntPtrConstant(0, dl));
10895 X86TargetLowering::LowerEXTRACT_VECTOR_ELT(SDValue Op,
10896 SelectionDAG &DAG) const {
10898 SDValue Vec = Op.getOperand(0);
10899 MVT VecVT = Vec.getSimpleValueType();
10900 SDValue Idx = Op.getOperand(1);
10902 if (Op.getSimpleValueType() == MVT::i1)
10903 return ExtractBitFromMaskVector(Op, DAG);
10905 if (!isa<ConstantSDNode>(Idx)) {
10906 if (VecVT.is512BitVector() ||
10907 (VecVT.is256BitVector() && Subtarget->hasInt256() &&
10908 VecVT.getVectorElementType().getSizeInBits() == 32)) {
10911 MVT::getIntegerVT(VecVT.getVectorElementType().getSizeInBits());
10912 MVT MaskVT = MVT::getVectorVT(MaskEltVT, VecVT.getSizeInBits() /
10913 MaskEltVT.getSizeInBits());
10915 Idx = DAG.getZExtOrTrunc(Idx, dl, MaskEltVT);
10916 auto PtrVT = getPointerTy(DAG.getDataLayout());
10917 SDValue Mask = DAG.getNode(X86ISD::VINSERT, dl, MaskVT,
10918 getZeroVector(MaskVT, Subtarget, DAG, dl), Idx,
10919 DAG.getConstant(0, dl, PtrVT));
10920 SDValue Perm = DAG.getNode(X86ISD::VPERMV, dl, VecVT, Mask, Vec);
10921 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, Op.getValueType(), Perm,
10922 DAG.getConstant(0, dl, PtrVT));
10927 // If this is a 256-bit vector result, first extract the 128-bit vector and
10928 // then extract the element from the 128-bit vector.
10929 if (VecVT.is256BitVector() || VecVT.is512BitVector()) {
10931 unsigned IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
10932 // Get the 128-bit vector.
10933 Vec = Extract128BitVector(Vec, IdxVal, DAG, dl);
10934 MVT EltVT = VecVT.getVectorElementType();
10936 unsigned ElemsPerChunk = 128 / EltVT.getSizeInBits();
10938 //if (IdxVal >= NumElems/2)
10939 // IdxVal -= NumElems/2;
10940 IdxVal -= (IdxVal/ElemsPerChunk)*ElemsPerChunk;
10941 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, Op.getValueType(), Vec,
10942 DAG.getConstant(IdxVal, dl, MVT::i32));
10945 assert(VecVT.is128BitVector() && "Unexpected vector length");
10947 if (Subtarget->hasSSE41())
10948 if (SDValue Res = LowerEXTRACT_VECTOR_ELT_SSE4(Op, DAG))
10951 MVT VT = Op.getSimpleValueType();
10952 // TODO: handle v16i8.
10953 if (VT.getSizeInBits() == 16) {
10954 SDValue Vec = Op.getOperand(0);
10955 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
10957 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i16,
10958 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
10959 DAG.getBitcast(MVT::v4i32, Vec),
10960 Op.getOperand(1)));
10961 // Transform it so it match pextrw which produces a 32-bit result.
10962 MVT EltVT = MVT::i32;
10963 SDValue Extract = DAG.getNode(X86ISD::PEXTRW, dl, EltVT,
10964 Op.getOperand(0), Op.getOperand(1));
10965 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, EltVT, Extract,
10966 DAG.getValueType(VT));
10967 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
10970 if (VT.getSizeInBits() == 32) {
10971 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
10975 // SHUFPS the element to the lowest double word, then movss.
10976 int Mask[4] = { static_cast<int>(Idx), -1, -1, -1 };
10977 MVT VVT = Op.getOperand(0).getSimpleValueType();
10978 SDValue Vec = DAG.getVectorShuffle(VVT, dl, Op.getOperand(0),
10979 DAG.getUNDEF(VVT), Mask);
10980 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, Vec,
10981 DAG.getIntPtrConstant(0, dl));
10984 if (VT.getSizeInBits() == 64) {
10985 // FIXME: .td only matches this for <2 x f64>, not <2 x i64> on 32b
10986 // FIXME: seems like this should be unnecessary if mov{h,l}pd were taught
10987 // to match extract_elt for f64.
10988 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
10992 // UNPCKHPD the element to the lowest double word, then movsd.
10993 // Note if the lower 64 bits of the result of the UNPCKHPD is then stored
10994 // to a f64mem, the whole operation is folded into a single MOVHPDmr.
10995 int Mask[2] = { 1, -1 };
10996 MVT VVT = Op.getOperand(0).getSimpleValueType();
10997 SDValue Vec = DAG.getVectorShuffle(VVT, dl, Op.getOperand(0),
10998 DAG.getUNDEF(VVT), Mask);
10999 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, Vec,
11000 DAG.getIntPtrConstant(0, dl));
11006 /// Insert one bit to mask vector, like v16i1 or v8i1.
11007 /// AVX-512 feature.
11009 X86TargetLowering::InsertBitToMaskVector(SDValue Op, SelectionDAG &DAG) const {
11011 SDValue Vec = Op.getOperand(0);
11012 SDValue Elt = Op.getOperand(1);
11013 SDValue Idx = Op.getOperand(2);
11014 MVT VecVT = Vec.getSimpleValueType();
11016 if (!isa<ConstantSDNode>(Idx)) {
11017 // Non constant index. Extend source and destination,
11018 // insert element and then truncate the result.
11019 MVT ExtVecVT = (VecVT == MVT::v8i1 ? MVT::v8i64 : MVT::v16i32);
11020 MVT ExtEltVT = (VecVT == MVT::v8i1 ? MVT::i64 : MVT::i32);
11021 SDValue ExtOp = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, ExtVecVT,
11022 DAG.getNode(ISD::ZERO_EXTEND, dl, ExtVecVT, Vec),
11023 DAG.getNode(ISD::ZERO_EXTEND, dl, ExtEltVT, Elt), Idx);
11024 return DAG.getNode(ISD::TRUNCATE, dl, VecVT, ExtOp);
11027 unsigned IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
11028 SDValue EltInVec = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VecVT, Elt);
11030 EltInVec = DAG.getNode(X86ISD::VSHLI, dl, VecVT, EltInVec,
11031 DAG.getConstant(IdxVal, dl, MVT::i8));
11032 if (Vec.getOpcode() == ISD::UNDEF)
11034 return DAG.getNode(ISD::OR, dl, VecVT, Vec, EltInVec);
11037 SDValue X86TargetLowering::LowerINSERT_VECTOR_ELT(SDValue Op,
11038 SelectionDAG &DAG) const {
11039 MVT VT = Op.getSimpleValueType();
11040 MVT EltVT = VT.getVectorElementType();
11042 if (EltVT == MVT::i1)
11043 return InsertBitToMaskVector(Op, DAG);
11046 SDValue N0 = Op.getOperand(0);
11047 SDValue N1 = Op.getOperand(1);
11048 SDValue N2 = Op.getOperand(2);
11049 if (!isa<ConstantSDNode>(N2))
11051 auto *N2C = cast<ConstantSDNode>(N2);
11052 unsigned IdxVal = N2C->getZExtValue();
11054 // If the vector is wider than 128 bits, extract the 128-bit subvector, insert
11055 // into that, and then insert the subvector back into the result.
11056 if (VT.is256BitVector() || VT.is512BitVector()) {
11057 // With a 256-bit vector, we can insert into the zero element efficiently
11058 // using a blend if we have AVX or AVX2 and the right data type.
11059 if (VT.is256BitVector() && IdxVal == 0) {
11060 // TODO: It is worthwhile to cast integer to floating point and back
11061 // and incur a domain crossing penalty if that's what we'll end up
11062 // doing anyway after extracting to a 128-bit vector.
11063 if ((Subtarget->hasAVX() && (EltVT == MVT::f64 || EltVT == MVT::f32)) ||
11064 (Subtarget->hasAVX2() && EltVT == MVT::i32)) {
11065 SDValue N1Vec = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, N1);
11066 N2 = DAG.getIntPtrConstant(1, dl);
11067 return DAG.getNode(X86ISD::BLENDI, dl, VT, N0, N1Vec, N2);
11071 // Get the desired 128-bit vector chunk.
11072 SDValue V = Extract128BitVector(N0, IdxVal, DAG, dl);
11074 // Insert the element into the desired chunk.
11075 unsigned NumEltsIn128 = 128 / EltVT.getSizeInBits();
11076 unsigned IdxIn128 = IdxVal - (IdxVal / NumEltsIn128) * NumEltsIn128;
11078 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, V.getValueType(), V, N1,
11079 DAG.getConstant(IdxIn128, dl, MVT::i32));
11081 // Insert the changed part back into the bigger vector
11082 return Insert128BitVector(N0, V, IdxVal, DAG, dl);
11084 assert(VT.is128BitVector() && "Only 128-bit vector types should be left!");
11086 if (Subtarget->hasSSE41()) {
11087 if (EltVT.getSizeInBits() == 8 || EltVT.getSizeInBits() == 16) {
11089 if (VT == MVT::v8i16) {
11090 Opc = X86ISD::PINSRW;
11092 assert(VT == MVT::v16i8);
11093 Opc = X86ISD::PINSRB;
11096 // Transform it so it match pinsr{b,w} which expects a GR32 as its second
11098 if (N1.getValueType() != MVT::i32)
11099 N1 = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, N1);
11100 if (N2.getValueType() != MVT::i32)
11101 N2 = DAG.getIntPtrConstant(IdxVal, dl);
11102 return DAG.getNode(Opc, dl, VT, N0, N1, N2);
11105 if (EltVT == MVT::f32) {
11106 // Bits [7:6] of the constant are the source select. This will always be
11107 // zero here. The DAG Combiner may combine an extract_elt index into
11108 // these bits. For example (insert (extract, 3), 2) could be matched by
11109 // putting the '3' into bits [7:6] of X86ISD::INSERTPS.
11110 // Bits [5:4] of the constant are the destination select. This is the
11111 // value of the incoming immediate.
11112 // Bits [3:0] of the constant are the zero mask. The DAG Combiner may
11113 // combine either bitwise AND or insert of float 0.0 to set these bits.
11115 const Function *F = DAG.getMachineFunction().getFunction();
11116 bool MinSize = F->hasFnAttribute(Attribute::MinSize);
11117 if (IdxVal == 0 && (!MinSize || !MayFoldLoad(N1))) {
11118 // If this is an insertion of 32-bits into the low 32-bits of
11119 // a vector, we prefer to generate a blend with immediate rather
11120 // than an insertps. Blends are simpler operations in hardware and so
11121 // will always have equal or better performance than insertps.
11122 // But if optimizing for size and there's a load folding opportunity,
11123 // generate insertps because blendps does not have a 32-bit memory
11125 N2 = DAG.getIntPtrConstant(1, dl);
11126 N1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4f32, N1);
11127 return DAG.getNode(X86ISD::BLENDI, dl, VT, N0, N1, N2);
11129 N2 = DAG.getIntPtrConstant(IdxVal << 4, dl);
11130 // Create this as a scalar to vector..
11131 N1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4f32, N1);
11132 return DAG.getNode(X86ISD::INSERTPS, dl, VT, N0, N1, N2);
11135 if (EltVT == MVT::i32 || EltVT == MVT::i64) {
11136 // PINSR* works with constant index.
11141 if (EltVT == MVT::i8)
11144 if (EltVT.getSizeInBits() == 16) {
11145 // Transform it so it match pinsrw which expects a 16-bit value in a GR32
11146 // as its second argument.
11147 if (N1.getValueType() != MVT::i32)
11148 N1 = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, N1);
11149 if (N2.getValueType() != MVT::i32)
11150 N2 = DAG.getIntPtrConstant(IdxVal, dl);
11151 return DAG.getNode(X86ISD::PINSRW, dl, VT, N0, N1, N2);
11156 static SDValue LowerSCALAR_TO_VECTOR(SDValue Op, SelectionDAG &DAG) {
11158 MVT OpVT = Op.getSimpleValueType();
11160 // If this is a 256-bit vector result, first insert into a 128-bit
11161 // vector and then insert into the 256-bit vector.
11162 if (!OpVT.is128BitVector()) {
11163 // Insert into a 128-bit vector.
11164 unsigned SizeFactor = OpVT.getSizeInBits()/128;
11165 MVT VT128 = MVT::getVectorVT(OpVT.getVectorElementType(),
11166 OpVT.getVectorNumElements() / SizeFactor);
11168 Op = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT128, Op.getOperand(0));
11170 // Insert the 128-bit vector.
11171 return Insert128BitVector(DAG.getUNDEF(OpVT), Op, 0, DAG, dl);
11174 if (OpVT == MVT::v1i64 &&
11175 Op.getOperand(0).getValueType() == MVT::i64)
11176 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v1i64, Op.getOperand(0));
11178 SDValue AnyExt = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, Op.getOperand(0));
11179 assert(OpVT.is128BitVector() && "Expected an SSE type!");
11180 return DAG.getBitcast(
11181 OpVT, DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32, AnyExt));
11184 // Lower a node with an EXTRACT_SUBVECTOR opcode. This may result in
11185 // a simple subregister reference or explicit instructions to grab
11186 // upper bits of a vector.
11187 static SDValue LowerEXTRACT_SUBVECTOR(SDValue Op, const X86Subtarget *Subtarget,
11188 SelectionDAG &DAG) {
11190 SDValue In = Op.getOperand(0);
11191 SDValue Idx = Op.getOperand(1);
11192 unsigned IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
11193 MVT ResVT = Op.getSimpleValueType();
11194 MVT InVT = In.getSimpleValueType();
11196 if (Subtarget->hasFp256()) {
11197 if (ResVT.is128BitVector() &&
11198 (InVT.is256BitVector() || InVT.is512BitVector()) &&
11199 isa<ConstantSDNode>(Idx)) {
11200 return Extract128BitVector(In, IdxVal, DAG, dl);
11202 if (ResVT.is256BitVector() && InVT.is512BitVector() &&
11203 isa<ConstantSDNode>(Idx)) {
11204 return Extract256BitVector(In, IdxVal, DAG, dl);
11210 // Lower a node with an INSERT_SUBVECTOR opcode. This may result in a
11211 // simple superregister reference or explicit instructions to insert
11212 // the upper bits of a vector.
11213 static SDValue LowerINSERT_SUBVECTOR(SDValue Op, const X86Subtarget *Subtarget,
11214 SelectionDAG &DAG) {
11215 if (!Subtarget->hasAVX())
11219 SDValue Vec = Op.getOperand(0);
11220 SDValue SubVec = Op.getOperand(1);
11221 SDValue Idx = Op.getOperand(2);
11223 if (!isa<ConstantSDNode>(Idx))
11226 unsigned IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
11227 MVT OpVT = Op.getSimpleValueType();
11228 MVT SubVecVT = SubVec.getSimpleValueType();
11230 // Fold two 16-byte subvector loads into one 32-byte load:
11231 // (insert_subvector (insert_subvector undef, (load addr), 0),
11232 // (load addr + 16), Elts/2)
11234 if ((IdxVal == OpVT.getVectorNumElements() / 2) &&
11235 Vec.getOpcode() == ISD::INSERT_SUBVECTOR &&
11236 OpVT.is256BitVector() && SubVecVT.is128BitVector() &&
11237 !Subtarget->isUnalignedMem32Slow()) {
11238 SDValue SubVec2 = Vec.getOperand(1);
11239 if (auto *Idx2 = dyn_cast<ConstantSDNode>(Vec.getOperand(2))) {
11240 if (Idx2->getZExtValue() == 0) {
11241 SDValue Ops[] = { SubVec2, SubVec };
11242 if (SDValue Ld = EltsFromConsecutiveLoads(OpVT, Ops, dl, DAG, false))
11248 if ((OpVT.is256BitVector() || OpVT.is512BitVector()) &&
11249 SubVecVT.is128BitVector())
11250 return Insert128BitVector(Vec, SubVec, IdxVal, DAG, dl);
11252 if (OpVT.is512BitVector() && SubVecVT.is256BitVector())
11253 return Insert256BitVector(Vec, SubVec, IdxVal, DAG, dl);
11255 if (OpVT.getVectorElementType() == MVT::i1) {
11256 if (IdxVal == 0 && Vec.getOpcode() == ISD::UNDEF) // the operation is legal
11258 SDValue ZeroIdx = DAG.getIntPtrConstant(0, dl);
11259 SDValue Undef = DAG.getUNDEF(OpVT);
11260 unsigned NumElems = OpVT.getVectorNumElements();
11261 SDValue ShiftBits = DAG.getConstant(NumElems/2, dl, MVT::i8);
11263 if (IdxVal == OpVT.getVectorNumElements() / 2) {
11264 // Zero upper bits of the Vec
11265 Vec = DAG.getNode(X86ISD::VSHLI, dl, OpVT, Vec, ShiftBits);
11266 Vec = DAG.getNode(X86ISD::VSRLI, dl, OpVT, Vec, ShiftBits);
11268 SDValue Vec2 = DAG.getNode(ISD::INSERT_SUBVECTOR, dl, OpVT, Undef,
11270 Vec2 = DAG.getNode(X86ISD::VSHLI, dl, OpVT, Vec2, ShiftBits);
11271 return DAG.getNode(ISD::OR, dl, OpVT, Vec, Vec2);
11274 SDValue Vec2 = DAG.getNode(ISD::INSERT_SUBVECTOR, dl, OpVT, Undef,
11276 // Zero upper bits of the Vec2
11277 Vec2 = DAG.getNode(X86ISD::VSHLI, dl, OpVT, Vec2, ShiftBits);
11278 Vec2 = DAG.getNode(X86ISD::VSRLI, dl, OpVT, Vec2, ShiftBits);
11279 // Zero lower bits of the Vec
11280 Vec = DAG.getNode(X86ISD::VSRLI, dl, OpVT, Vec, ShiftBits);
11281 Vec = DAG.getNode(X86ISD::VSHLI, dl, OpVT, Vec, ShiftBits);
11282 // Merge them together
11283 return DAG.getNode(ISD::OR, dl, OpVT, Vec, Vec2);
11289 // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
11290 // their target countpart wrapped in the X86ISD::Wrapper node. Suppose N is
11291 // one of the above mentioned nodes. It has to be wrapped because otherwise
11292 // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
11293 // be used to form addressing mode. These wrapped nodes will be selected
11296 X86TargetLowering::LowerConstantPool(SDValue Op, SelectionDAG &DAG) const {
11297 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
11299 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
11300 // global base reg.
11301 unsigned char OpFlag = 0;
11302 unsigned WrapperKind = X86ISD::Wrapper;
11303 CodeModel::Model M = DAG.getTarget().getCodeModel();
11305 if (Subtarget->isPICStyleRIPRel() &&
11306 (M == CodeModel::Small || M == CodeModel::Kernel))
11307 WrapperKind = X86ISD::WrapperRIP;
11308 else if (Subtarget->isPICStyleGOT())
11309 OpFlag = X86II::MO_GOTOFF;
11310 else if (Subtarget->isPICStyleStubPIC())
11311 OpFlag = X86II::MO_PIC_BASE_OFFSET;
11313 auto PtrVT = getPointerTy(DAG.getDataLayout());
11314 SDValue Result = DAG.getTargetConstantPool(
11315 CP->getConstVal(), PtrVT, CP->getAlignment(), CP->getOffset(), OpFlag);
11317 Result = DAG.getNode(WrapperKind, DL, PtrVT, Result);
11318 // With PIC, the address is actually $g + Offset.
11321 DAG.getNode(ISD::ADD, DL, PtrVT,
11322 DAG.getNode(X86ISD::GlobalBaseReg, SDLoc(), PtrVT), Result);
11328 SDValue X86TargetLowering::LowerJumpTable(SDValue Op, SelectionDAG &DAG) const {
11329 JumpTableSDNode *JT = cast<JumpTableSDNode>(Op);
11331 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
11332 // global base reg.
11333 unsigned char OpFlag = 0;
11334 unsigned WrapperKind = X86ISD::Wrapper;
11335 CodeModel::Model M = DAG.getTarget().getCodeModel();
11337 if (Subtarget->isPICStyleRIPRel() &&
11338 (M == CodeModel::Small || M == CodeModel::Kernel))
11339 WrapperKind = X86ISD::WrapperRIP;
11340 else if (Subtarget->isPICStyleGOT())
11341 OpFlag = X86II::MO_GOTOFF;
11342 else if (Subtarget->isPICStyleStubPIC())
11343 OpFlag = X86II::MO_PIC_BASE_OFFSET;
11345 auto PtrVT = getPointerTy(DAG.getDataLayout());
11346 SDValue Result = DAG.getTargetJumpTable(JT->getIndex(), PtrVT, OpFlag);
11348 Result = DAG.getNode(WrapperKind, DL, PtrVT, Result);
11350 // With PIC, the address is actually $g + Offset.
11353 DAG.getNode(ISD::ADD, DL, PtrVT,
11354 DAG.getNode(X86ISD::GlobalBaseReg, SDLoc(), PtrVT), Result);
11360 X86TargetLowering::LowerExternalSymbol(SDValue Op, SelectionDAG &DAG) const {
11361 const char *Sym = cast<ExternalSymbolSDNode>(Op)->getSymbol();
11363 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
11364 // global base reg.
11365 unsigned char OpFlag = 0;
11366 unsigned WrapperKind = X86ISD::Wrapper;
11367 CodeModel::Model M = DAG.getTarget().getCodeModel();
11369 if (Subtarget->isPICStyleRIPRel() &&
11370 (M == CodeModel::Small || M == CodeModel::Kernel)) {
11371 if (Subtarget->isTargetDarwin() || Subtarget->isTargetELF())
11372 OpFlag = X86II::MO_GOTPCREL;
11373 WrapperKind = X86ISD::WrapperRIP;
11374 } else if (Subtarget->isPICStyleGOT()) {
11375 OpFlag = X86II::MO_GOT;
11376 } else if (Subtarget->isPICStyleStubPIC()) {
11377 OpFlag = X86II::MO_DARWIN_NONLAZY_PIC_BASE;
11378 } else if (Subtarget->isPICStyleStubNoDynamic()) {
11379 OpFlag = X86II::MO_DARWIN_NONLAZY;
11382 auto PtrVT = getPointerTy(DAG.getDataLayout());
11383 SDValue Result = DAG.getTargetExternalSymbol(Sym, PtrVT, OpFlag);
11386 Result = DAG.getNode(WrapperKind, DL, PtrVT, Result);
11388 // With PIC, the address is actually $g + Offset.
11389 if (DAG.getTarget().getRelocationModel() == Reloc::PIC_ &&
11390 !Subtarget->is64Bit()) {
11392 DAG.getNode(ISD::ADD, DL, PtrVT,
11393 DAG.getNode(X86ISD::GlobalBaseReg, SDLoc(), PtrVT), Result);
11396 // For symbols that require a load from a stub to get the address, emit the
11398 if (isGlobalStubReference(OpFlag))
11399 Result = DAG.getLoad(PtrVT, DL, DAG.getEntryNode(), Result,
11400 MachinePointerInfo::getGOT(), false, false, false, 0);
11406 X86TargetLowering::LowerBlockAddress(SDValue Op, SelectionDAG &DAG) const {
11407 // Create the TargetBlockAddressAddress node.
11408 unsigned char OpFlags =
11409 Subtarget->ClassifyBlockAddressReference();
11410 CodeModel::Model M = DAG.getTarget().getCodeModel();
11411 const BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress();
11412 int64_t Offset = cast<BlockAddressSDNode>(Op)->getOffset();
11414 auto PtrVT = getPointerTy(DAG.getDataLayout());
11415 SDValue Result = DAG.getTargetBlockAddress(BA, PtrVT, Offset, OpFlags);
11417 if (Subtarget->isPICStyleRIPRel() &&
11418 (M == CodeModel::Small || M == CodeModel::Kernel))
11419 Result = DAG.getNode(X86ISD::WrapperRIP, dl, PtrVT, Result);
11421 Result = DAG.getNode(X86ISD::Wrapper, dl, PtrVT, Result);
11423 // With PIC, the address is actually $g + Offset.
11424 if (isGlobalRelativeToPICBase(OpFlags)) {
11425 Result = DAG.getNode(ISD::ADD, dl, PtrVT,
11426 DAG.getNode(X86ISD::GlobalBaseReg, dl, PtrVT), Result);
11433 X86TargetLowering::LowerGlobalAddress(const GlobalValue *GV, SDLoc dl,
11434 int64_t Offset, SelectionDAG &DAG) const {
11435 // Create the TargetGlobalAddress node, folding in the constant
11436 // offset if it is legal.
11437 unsigned char OpFlags =
11438 Subtarget->ClassifyGlobalReference(GV, DAG.getTarget());
11439 CodeModel::Model M = DAG.getTarget().getCodeModel();
11440 auto PtrVT = getPointerTy(DAG.getDataLayout());
11442 if (OpFlags == X86II::MO_NO_FLAG &&
11443 X86::isOffsetSuitableForCodeModel(Offset, M)) {
11444 // A direct static reference to a global.
11445 Result = DAG.getTargetGlobalAddress(GV, dl, PtrVT, Offset);
11448 Result = DAG.getTargetGlobalAddress(GV, dl, PtrVT, 0, OpFlags);
11451 if (Subtarget->isPICStyleRIPRel() &&
11452 (M == CodeModel::Small || M == CodeModel::Kernel))
11453 Result = DAG.getNode(X86ISD::WrapperRIP, dl, PtrVT, Result);
11455 Result = DAG.getNode(X86ISD::Wrapper, dl, PtrVT, Result);
11457 // With PIC, the address is actually $g + Offset.
11458 if (isGlobalRelativeToPICBase(OpFlags)) {
11459 Result = DAG.getNode(ISD::ADD, dl, PtrVT,
11460 DAG.getNode(X86ISD::GlobalBaseReg, dl, PtrVT), Result);
11463 // For globals that require a load from a stub to get the address, emit the
11465 if (isGlobalStubReference(OpFlags))
11466 Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Result,
11467 MachinePointerInfo::getGOT(), false, false, false, 0);
11469 // If there was a non-zero offset that we didn't fold, create an explicit
11470 // addition for it.
11472 Result = DAG.getNode(ISD::ADD, dl, PtrVT, Result,
11473 DAG.getConstant(Offset, dl, PtrVT));
11479 X86TargetLowering::LowerGlobalAddress(SDValue Op, SelectionDAG &DAG) const {
11480 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
11481 int64_t Offset = cast<GlobalAddressSDNode>(Op)->getOffset();
11482 return LowerGlobalAddress(GV, SDLoc(Op), Offset, DAG);
11486 GetTLSADDR(SelectionDAG &DAG, SDValue Chain, GlobalAddressSDNode *GA,
11487 SDValue *InFlag, const EVT PtrVT, unsigned ReturnReg,
11488 unsigned char OperandFlags, bool LocalDynamic = false) {
11489 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
11490 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
11492 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(), dl,
11493 GA->getValueType(0),
11497 X86ISD::NodeType CallType = LocalDynamic ? X86ISD::TLSBASEADDR
11501 SDValue Ops[] = { Chain, TGA, *InFlag };
11502 Chain = DAG.getNode(CallType, dl, NodeTys, Ops);
11504 SDValue Ops[] = { Chain, TGA };
11505 Chain = DAG.getNode(CallType, dl, NodeTys, Ops);
11508 // TLSADDR will be codegen'ed as call. Inform MFI that function has calls.
11509 MFI->setAdjustsStack(true);
11510 MFI->setHasCalls(true);
11512 SDValue Flag = Chain.getValue(1);
11513 return DAG.getCopyFromReg(Chain, dl, ReturnReg, PtrVT, Flag);
11516 // Lower ISD::GlobalTLSAddress using the "general dynamic" model, 32 bit
11518 LowerToTLSGeneralDynamicModel32(GlobalAddressSDNode *GA, SelectionDAG &DAG,
11521 SDLoc dl(GA); // ? function entry point might be better
11522 SDValue Chain = DAG.getCopyToReg(DAG.getEntryNode(), dl, X86::EBX,
11523 DAG.getNode(X86ISD::GlobalBaseReg,
11524 SDLoc(), PtrVT), InFlag);
11525 InFlag = Chain.getValue(1);
11527 return GetTLSADDR(DAG, Chain, GA, &InFlag, PtrVT, X86::EAX, X86II::MO_TLSGD);
11530 // Lower ISD::GlobalTLSAddress using the "general dynamic" model, 64 bit
11532 LowerToTLSGeneralDynamicModel64(GlobalAddressSDNode *GA, SelectionDAG &DAG,
11534 return GetTLSADDR(DAG, DAG.getEntryNode(), GA, nullptr, PtrVT,
11535 X86::RAX, X86II::MO_TLSGD);
11538 static SDValue LowerToTLSLocalDynamicModel(GlobalAddressSDNode *GA,
11544 // Get the start address of the TLS block for this module.
11545 X86MachineFunctionInfo* MFI = DAG.getMachineFunction()
11546 .getInfo<X86MachineFunctionInfo>();
11547 MFI->incNumLocalDynamicTLSAccesses();
11551 Base = GetTLSADDR(DAG, DAG.getEntryNode(), GA, nullptr, PtrVT, X86::RAX,
11552 X86II::MO_TLSLD, /*LocalDynamic=*/true);
11555 SDValue Chain = DAG.getCopyToReg(DAG.getEntryNode(), dl, X86::EBX,
11556 DAG.getNode(X86ISD::GlobalBaseReg, SDLoc(), PtrVT), InFlag);
11557 InFlag = Chain.getValue(1);
11558 Base = GetTLSADDR(DAG, Chain, GA, &InFlag, PtrVT, X86::EAX,
11559 X86II::MO_TLSLDM, /*LocalDynamic=*/true);
11562 // Note: the CleanupLocalDynamicTLSPass will remove redundant computations
11566 unsigned char OperandFlags = X86II::MO_DTPOFF;
11567 unsigned WrapperKind = X86ISD::Wrapper;
11568 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(), dl,
11569 GA->getValueType(0),
11570 GA->getOffset(), OperandFlags);
11571 SDValue Offset = DAG.getNode(WrapperKind, dl, PtrVT, TGA);
11573 // Add x@dtpoff with the base.
11574 return DAG.getNode(ISD::ADD, dl, PtrVT, Offset, Base);
11577 // Lower ISD::GlobalTLSAddress using the "initial exec" or "local exec" model.
11578 static SDValue LowerToTLSExecModel(GlobalAddressSDNode *GA, SelectionDAG &DAG,
11579 const EVT PtrVT, TLSModel::Model model,
11580 bool is64Bit, bool isPIC) {
11583 // Get the Thread Pointer, which is %gs:0 (32-bit) or %fs:0 (64-bit).
11584 Value *Ptr = Constant::getNullValue(Type::getInt8PtrTy(*DAG.getContext(),
11585 is64Bit ? 257 : 256));
11587 SDValue ThreadPointer =
11588 DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), DAG.getIntPtrConstant(0, dl),
11589 MachinePointerInfo(Ptr), false, false, false, 0);
11591 unsigned char OperandFlags = 0;
11592 // Most TLS accesses are not RIP relative, even on x86-64. One exception is
11594 unsigned WrapperKind = X86ISD::Wrapper;
11595 if (model == TLSModel::LocalExec) {
11596 OperandFlags = is64Bit ? X86II::MO_TPOFF : X86II::MO_NTPOFF;
11597 } else if (model == TLSModel::InitialExec) {
11599 OperandFlags = X86II::MO_GOTTPOFF;
11600 WrapperKind = X86ISD::WrapperRIP;
11602 OperandFlags = isPIC ? X86II::MO_GOTNTPOFF : X86II::MO_INDNTPOFF;
11605 llvm_unreachable("Unexpected model");
11608 // emit "addl x@ntpoff,%eax" (local exec)
11609 // or "addl x@indntpoff,%eax" (initial exec)
11610 // or "addl x@gotntpoff(%ebx) ,%eax" (initial exec, 32-bit pic)
11612 DAG.getTargetGlobalAddress(GA->getGlobal(), dl, GA->getValueType(0),
11613 GA->getOffset(), OperandFlags);
11614 SDValue Offset = DAG.getNode(WrapperKind, dl, PtrVT, TGA);
11616 if (model == TLSModel::InitialExec) {
11617 if (isPIC && !is64Bit) {
11618 Offset = DAG.getNode(ISD::ADD, dl, PtrVT,
11619 DAG.getNode(X86ISD::GlobalBaseReg, SDLoc(), PtrVT),
11623 Offset = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Offset,
11624 MachinePointerInfo::getGOT(), false, false, false, 0);
11627 // The address of the thread local variable is the add of the thread
11628 // pointer with the offset of the variable.
11629 return DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, Offset);
11633 X86TargetLowering::LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const {
11635 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
11636 const GlobalValue *GV = GA->getGlobal();
11637 auto PtrVT = getPointerTy(DAG.getDataLayout());
11639 if (Subtarget->isTargetELF()) {
11640 TLSModel::Model model = DAG.getTarget().getTLSModel(GV);
11642 case TLSModel::GeneralDynamic:
11643 if (Subtarget->is64Bit())
11644 return LowerToTLSGeneralDynamicModel64(GA, DAG, PtrVT);
11645 return LowerToTLSGeneralDynamicModel32(GA, DAG, PtrVT);
11646 case TLSModel::LocalDynamic:
11647 return LowerToTLSLocalDynamicModel(GA, DAG, PtrVT,
11648 Subtarget->is64Bit());
11649 case TLSModel::InitialExec:
11650 case TLSModel::LocalExec:
11651 return LowerToTLSExecModel(GA, DAG, PtrVT, model, Subtarget->is64Bit(),
11652 DAG.getTarget().getRelocationModel() ==
11655 llvm_unreachable("Unknown TLS model.");
11658 if (Subtarget->isTargetDarwin()) {
11659 // Darwin only has one model of TLS. Lower to that.
11660 unsigned char OpFlag = 0;
11661 unsigned WrapperKind = Subtarget->isPICStyleRIPRel() ?
11662 X86ISD::WrapperRIP : X86ISD::Wrapper;
11664 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
11665 // global base reg.
11666 bool PIC32 = (DAG.getTarget().getRelocationModel() == Reloc::PIC_) &&
11667 !Subtarget->is64Bit();
11669 OpFlag = X86II::MO_TLVP_PIC_BASE;
11671 OpFlag = X86II::MO_TLVP;
11673 SDValue Result = DAG.getTargetGlobalAddress(GA->getGlobal(), DL,
11674 GA->getValueType(0),
11675 GA->getOffset(), OpFlag);
11676 SDValue Offset = DAG.getNode(WrapperKind, DL, PtrVT, Result);
11678 // With PIC32, the address is actually $g + Offset.
11680 Offset = DAG.getNode(ISD::ADD, DL, PtrVT,
11681 DAG.getNode(X86ISD::GlobalBaseReg, SDLoc(), PtrVT),
11684 // Lowering the machine isd will make sure everything is in the right
11686 SDValue Chain = DAG.getEntryNode();
11687 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
11688 SDValue Args[] = { Chain, Offset };
11689 Chain = DAG.getNode(X86ISD::TLSCALL, DL, NodeTys, Args);
11691 // TLSCALL will be codegen'ed as call. Inform MFI that function has calls.
11692 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
11693 MFI->setAdjustsStack(true);
11695 // And our return value (tls address) is in the standard call return value
11697 unsigned Reg = Subtarget->is64Bit() ? X86::RAX : X86::EAX;
11698 return DAG.getCopyFromReg(Chain, DL, Reg, PtrVT, Chain.getValue(1));
11701 if (Subtarget->isTargetKnownWindowsMSVC() ||
11702 Subtarget->isTargetWindowsGNU()) {
11703 // Just use the implicit TLS architecture
11704 // Need to generate someting similar to:
11705 // mov rdx, qword [gs:abs 58H]; Load pointer to ThreadLocalStorage
11707 // mov ecx, dword [rel _tls_index]: Load index (from C runtime)
11708 // mov rcx, qword [rdx+rcx*8]
11709 // mov eax, .tls$:tlsvar
11710 // [rax+rcx] contains the address
11711 // Windows 64bit: gs:0x58
11712 // Windows 32bit: fs:__tls_array
11715 SDValue Chain = DAG.getEntryNode();
11717 // Get the Thread Pointer, which is %fs:__tls_array (32-bit) or
11718 // %gs:0x58 (64-bit). On MinGW, __tls_array is not available, so directly
11719 // use its literal value of 0x2C.
11720 Value *Ptr = Constant::getNullValue(Subtarget->is64Bit()
11721 ? Type::getInt8PtrTy(*DAG.getContext(),
11723 : Type::getInt32PtrTy(*DAG.getContext(),
11726 SDValue TlsArray = Subtarget->is64Bit()
11727 ? DAG.getIntPtrConstant(0x58, dl)
11728 : (Subtarget->isTargetWindowsGNU()
11729 ? DAG.getIntPtrConstant(0x2C, dl)
11730 : DAG.getExternalSymbol("_tls_array", PtrVT));
11732 SDValue ThreadPointer =
11733 DAG.getLoad(PtrVT, dl, Chain, TlsArray, MachinePointerInfo(Ptr), false,
11737 if (GV->getThreadLocalMode() == GlobalVariable::LocalExecTLSModel) {
11738 res = ThreadPointer;
11740 // Load the _tls_index variable
11741 SDValue IDX = DAG.getExternalSymbol("_tls_index", PtrVT);
11742 if (Subtarget->is64Bit())
11743 IDX = DAG.getExtLoad(ISD::ZEXTLOAD, dl, PtrVT, Chain, IDX,
11744 MachinePointerInfo(), MVT::i32, false, false,
11747 IDX = DAG.getLoad(PtrVT, dl, Chain, IDX, MachinePointerInfo(), false,
11750 auto &DL = DAG.getDataLayout();
11752 DAG.getConstant(Log2_64_Ceil(DL.getPointerSize()), dl, PtrVT);
11753 IDX = DAG.getNode(ISD::SHL, dl, PtrVT, IDX, Scale);
11755 res = DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, IDX);
11758 res = DAG.getLoad(PtrVT, dl, Chain, res, MachinePointerInfo(), false, false,
11761 // Get the offset of start of .tls section
11762 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(), dl,
11763 GA->getValueType(0),
11764 GA->getOffset(), X86II::MO_SECREL);
11765 SDValue Offset = DAG.getNode(X86ISD::Wrapper, dl, PtrVT, TGA);
11767 // The address of the thread local variable is the add of the thread
11768 // pointer with the offset of the variable.
11769 return DAG.getNode(ISD::ADD, dl, PtrVT, res, Offset);
11772 llvm_unreachable("TLS not implemented for this target.");
11775 /// LowerShiftParts - Lower SRA_PARTS and friends, which return two i32 values
11776 /// and take a 2 x i32 value to shift plus a shift amount.
11777 static SDValue LowerShiftParts(SDValue Op, SelectionDAG &DAG) {
11778 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
11779 MVT VT = Op.getSimpleValueType();
11780 unsigned VTBits = VT.getSizeInBits();
11782 bool isSRA = Op.getOpcode() == ISD::SRA_PARTS;
11783 SDValue ShOpLo = Op.getOperand(0);
11784 SDValue ShOpHi = Op.getOperand(1);
11785 SDValue ShAmt = Op.getOperand(2);
11786 // X86ISD::SHLD and X86ISD::SHRD have defined overflow behavior but the
11787 // generic ISD nodes haven't. Insert an AND to be safe, it's optimized away
11789 SDValue SafeShAmt = DAG.getNode(ISD::AND, dl, MVT::i8, ShAmt,
11790 DAG.getConstant(VTBits - 1, dl, MVT::i8));
11791 SDValue Tmp1 = isSRA ? DAG.getNode(ISD::SRA, dl, VT, ShOpHi,
11792 DAG.getConstant(VTBits - 1, dl, MVT::i8))
11793 : DAG.getConstant(0, dl, VT);
11795 SDValue Tmp2, Tmp3;
11796 if (Op.getOpcode() == ISD::SHL_PARTS) {
11797 Tmp2 = DAG.getNode(X86ISD::SHLD, dl, VT, ShOpHi, ShOpLo, ShAmt);
11798 Tmp3 = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, SafeShAmt);
11800 Tmp2 = DAG.getNode(X86ISD::SHRD, dl, VT, ShOpLo, ShOpHi, ShAmt);
11801 Tmp3 = DAG.getNode(isSRA ? ISD::SRA : ISD::SRL, dl, VT, ShOpHi, SafeShAmt);
11804 // If the shift amount is larger or equal than the width of a part we can't
11805 // rely on the results of shld/shrd. Insert a test and select the appropriate
11806 // values for large shift amounts.
11807 SDValue AndNode = DAG.getNode(ISD::AND, dl, MVT::i8, ShAmt,
11808 DAG.getConstant(VTBits, dl, MVT::i8));
11809 SDValue Cond = DAG.getNode(X86ISD::CMP, dl, MVT::i32,
11810 AndNode, DAG.getConstant(0, dl, MVT::i8));
11813 SDValue CC = DAG.getConstant(X86::COND_NE, dl, MVT::i8);
11814 SDValue Ops0[4] = { Tmp2, Tmp3, CC, Cond };
11815 SDValue Ops1[4] = { Tmp3, Tmp1, CC, Cond };
11817 if (Op.getOpcode() == ISD::SHL_PARTS) {
11818 Hi = DAG.getNode(X86ISD::CMOV, dl, VT, Ops0);
11819 Lo = DAG.getNode(X86ISD::CMOV, dl, VT, Ops1);
11821 Lo = DAG.getNode(X86ISD::CMOV, dl, VT, Ops0);
11822 Hi = DAG.getNode(X86ISD::CMOV, dl, VT, Ops1);
11825 SDValue Ops[2] = { Lo, Hi };
11826 return DAG.getMergeValues(Ops, dl);
11829 SDValue X86TargetLowering::LowerSINT_TO_FP(SDValue Op,
11830 SelectionDAG &DAG) const {
11831 SDValue Src = Op.getOperand(0);
11832 MVT SrcVT = Src.getSimpleValueType();
11833 MVT VT = Op.getSimpleValueType();
11836 if (SrcVT.isVector()) {
11837 if (SrcVT == MVT::v2i32 && VT == MVT::v2f64) {
11838 return DAG.getNode(X86ISD::CVTDQ2PD, dl, VT,
11839 DAG.getNode(ISD::CONCAT_VECTORS, dl, MVT::v4i32, Src,
11840 DAG.getUNDEF(SrcVT)));
11842 if (SrcVT.getVectorElementType() == MVT::i1) {
11843 MVT IntegerVT = MVT::getVectorVT(MVT::i32, SrcVT.getVectorNumElements());
11844 return DAG.getNode(ISD::SINT_TO_FP, dl, Op.getValueType(),
11845 DAG.getNode(ISD::SIGN_EXTEND, dl, IntegerVT, Src));
11850 assert(SrcVT <= MVT::i64 && SrcVT >= MVT::i16 &&
11851 "Unknown SINT_TO_FP to lower!");
11853 // These are really Legal; return the operand so the caller accepts it as
11855 if (SrcVT == MVT::i32 && isScalarFPTypeInSSEReg(Op.getValueType()))
11857 if (SrcVT == MVT::i64 && isScalarFPTypeInSSEReg(Op.getValueType()) &&
11858 Subtarget->is64Bit()) {
11862 unsigned Size = SrcVT.getSizeInBits()/8;
11863 MachineFunction &MF = DAG.getMachineFunction();
11864 auto PtrVT = getPointerTy(MF.getDataLayout());
11865 int SSFI = MF.getFrameInfo()->CreateStackObject(Size, Size, false);
11866 SDValue StackSlot = DAG.getFrameIndex(SSFI, PtrVT);
11867 SDValue Chain = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
11869 MachinePointerInfo::getFixedStack(SSFI),
11871 return BuildFILD(Op, SrcVT, Chain, StackSlot, DAG);
11874 SDValue X86TargetLowering::BuildFILD(SDValue Op, EVT SrcVT, SDValue Chain,
11876 SelectionDAG &DAG) const {
11880 bool useSSE = isScalarFPTypeInSSEReg(Op.getValueType());
11882 Tys = DAG.getVTList(MVT::f64, MVT::Other, MVT::Glue);
11884 Tys = DAG.getVTList(Op.getValueType(), MVT::Other);
11886 unsigned ByteSize = SrcVT.getSizeInBits()/8;
11888 FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(StackSlot);
11889 MachineMemOperand *MMO;
11891 int SSFI = FI->getIndex();
11893 DAG.getMachineFunction()
11894 .getMachineMemOperand(MachinePointerInfo::getFixedStack(SSFI),
11895 MachineMemOperand::MOLoad, ByteSize, ByteSize);
11897 MMO = cast<LoadSDNode>(StackSlot)->getMemOperand();
11898 StackSlot = StackSlot.getOperand(1);
11900 SDValue Ops[] = { Chain, StackSlot, DAG.getValueType(SrcVT) };
11901 SDValue Result = DAG.getMemIntrinsicNode(useSSE ? X86ISD::FILD_FLAG :
11903 Tys, Ops, SrcVT, MMO);
11906 Chain = Result.getValue(1);
11907 SDValue InFlag = Result.getValue(2);
11909 // FIXME: Currently the FST is flagged to the FILD_FLAG. This
11910 // shouldn't be necessary except that RFP cannot be live across
11911 // multiple blocks. When stackifier is fixed, they can be uncoupled.
11912 MachineFunction &MF = DAG.getMachineFunction();
11913 unsigned SSFISize = Op.getValueType().getSizeInBits()/8;
11914 int SSFI = MF.getFrameInfo()->CreateStackObject(SSFISize, SSFISize, false);
11915 auto PtrVT = getPointerTy(MF.getDataLayout());
11916 SDValue StackSlot = DAG.getFrameIndex(SSFI, PtrVT);
11917 Tys = DAG.getVTList(MVT::Other);
11919 Chain, Result, StackSlot, DAG.getValueType(Op.getValueType()), InFlag
11921 MachineMemOperand *MMO =
11922 DAG.getMachineFunction()
11923 .getMachineMemOperand(MachinePointerInfo::getFixedStack(SSFI),
11924 MachineMemOperand::MOStore, SSFISize, SSFISize);
11926 Chain = DAG.getMemIntrinsicNode(X86ISD::FST, DL, Tys,
11927 Ops, Op.getValueType(), MMO);
11928 Result = DAG.getLoad(Op.getValueType(), DL, Chain, StackSlot,
11929 MachinePointerInfo::getFixedStack(SSFI),
11930 false, false, false, 0);
11936 // LowerUINT_TO_FP_i64 - 64-bit unsigned integer to double expansion.
11937 SDValue X86TargetLowering::LowerUINT_TO_FP_i64(SDValue Op,
11938 SelectionDAG &DAG) const {
11939 // This algorithm is not obvious. Here it is what we're trying to output:
11942 punpckldq (c0), %xmm0 // c0: (uint4){ 0x43300000U, 0x45300000U, 0U, 0U }
11943 subpd (c1), %xmm0 // c1: (double2){ 0x1.0p52, 0x1.0p52 * 0x1.0p32 }
11945 haddpd %xmm0, %xmm0
11947 pshufd $0x4e, %xmm0, %xmm1
11953 LLVMContext *Context = DAG.getContext();
11955 // Build some magic constants.
11956 static const uint32_t CV0[] = { 0x43300000, 0x45300000, 0, 0 };
11957 Constant *C0 = ConstantDataVector::get(*Context, CV0);
11958 auto PtrVT = getPointerTy(DAG.getDataLayout());
11959 SDValue CPIdx0 = DAG.getConstantPool(C0, PtrVT, 16);
11961 SmallVector<Constant*,2> CV1;
11963 ConstantFP::get(*Context, APFloat(APFloat::IEEEdouble,
11964 APInt(64, 0x4330000000000000ULL))));
11966 ConstantFP::get(*Context, APFloat(APFloat::IEEEdouble,
11967 APInt(64, 0x4530000000000000ULL))));
11968 Constant *C1 = ConstantVector::get(CV1);
11969 SDValue CPIdx1 = DAG.getConstantPool(C1, PtrVT, 16);
11971 // Load the 64-bit value into an XMM register.
11972 SDValue XR1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i64,
11974 SDValue CLod0 = DAG.getLoad(MVT::v4i32, dl, DAG.getEntryNode(), CPIdx0,
11975 MachinePointerInfo::getConstantPool(),
11976 false, false, false, 16);
11978 getUnpackl(DAG, dl, MVT::v4i32, DAG.getBitcast(MVT::v4i32, XR1), CLod0);
11980 SDValue CLod1 = DAG.getLoad(MVT::v2f64, dl, CLod0.getValue(1), CPIdx1,
11981 MachinePointerInfo::getConstantPool(),
11982 false, false, false, 16);
11983 SDValue XR2F = DAG.getBitcast(MVT::v2f64, Unpck1);
11984 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::v2f64, XR2F, CLod1);
11987 if (Subtarget->hasSSE3()) {
11988 // FIXME: The 'haddpd' instruction may be slower than 'movhlps + addsd'.
11989 Result = DAG.getNode(X86ISD::FHADD, dl, MVT::v2f64, Sub, Sub);
11991 SDValue S2F = DAG.getBitcast(MVT::v4i32, Sub);
11992 SDValue Shuffle = getTargetShuffleNode(X86ISD::PSHUFD, dl, MVT::v4i32,
11994 Result = DAG.getNode(ISD::FADD, dl, MVT::v2f64,
11995 DAG.getBitcast(MVT::v2f64, Shuffle), Sub);
11998 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Result,
11999 DAG.getIntPtrConstant(0, dl));
12002 // LowerUINT_TO_FP_i32 - 32-bit unsigned integer to float expansion.
12003 SDValue X86TargetLowering::LowerUINT_TO_FP_i32(SDValue Op,
12004 SelectionDAG &DAG) const {
12006 // FP constant to bias correct the final result.
12007 SDValue Bias = DAG.getConstantFP(BitsToDouble(0x4330000000000000ULL), dl,
12010 // Load the 32-bit value into an XMM register.
12011 SDValue Load = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
12014 // Zero out the upper parts of the register.
12015 Load = getShuffleVectorZeroOrUndef(Load, 0, true, Subtarget, DAG);
12017 Load = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
12018 DAG.getBitcast(MVT::v2f64, Load),
12019 DAG.getIntPtrConstant(0, dl));
12021 // Or the load with the bias.
12022 SDValue Or = DAG.getNode(
12023 ISD::OR, dl, MVT::v2i64,
12024 DAG.getBitcast(MVT::v2i64,
12025 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2f64, Load)),
12026 DAG.getBitcast(MVT::v2i64,
12027 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2f64, Bias)));
12029 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
12030 DAG.getBitcast(MVT::v2f64, Or), DAG.getIntPtrConstant(0, dl));
12032 // Subtract the bias.
12033 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::f64, Or, Bias);
12035 // Handle final rounding.
12036 EVT DestVT = Op.getValueType();
12038 if (DestVT.bitsLT(MVT::f64))
12039 return DAG.getNode(ISD::FP_ROUND, dl, DestVT, Sub,
12040 DAG.getIntPtrConstant(0, dl));
12041 if (DestVT.bitsGT(MVT::f64))
12042 return DAG.getNode(ISD::FP_EXTEND, dl, DestVT, Sub);
12044 // Handle final rounding.
12048 static SDValue lowerUINT_TO_FP_vXi32(SDValue Op, SelectionDAG &DAG,
12049 const X86Subtarget &Subtarget) {
12050 // The algorithm is the following:
12051 // #ifdef __SSE4_1__
12052 // uint4 lo = _mm_blend_epi16( v, (uint4) 0x4b000000, 0xaa);
12053 // uint4 hi = _mm_blend_epi16( _mm_srli_epi32(v,16),
12054 // (uint4) 0x53000000, 0xaa);
12056 // uint4 lo = (v & (uint4) 0xffff) | (uint4) 0x4b000000;
12057 // uint4 hi = (v >> 16) | (uint4) 0x53000000;
12059 // float4 fhi = (float4) hi - (0x1.0p39f + 0x1.0p23f);
12060 // return (float4) lo + fhi;
12063 SDValue V = Op->getOperand(0);
12064 EVT VecIntVT = V.getValueType();
12065 bool Is128 = VecIntVT == MVT::v4i32;
12066 EVT VecFloatVT = Is128 ? MVT::v4f32 : MVT::v8f32;
12067 // If we convert to something else than the supported type, e.g., to v4f64,
12069 if (VecFloatVT != Op->getValueType(0))
12072 unsigned NumElts = VecIntVT.getVectorNumElements();
12073 assert((VecIntVT == MVT::v4i32 || VecIntVT == MVT::v8i32) &&
12074 "Unsupported custom type");
12075 assert(NumElts <= 8 && "The size of the constant array must be fixed");
12077 // In the #idef/#else code, we have in common:
12078 // - The vector of constants:
12084 // Create the splat vector for 0x4b000000.
12085 SDValue CstLow = DAG.getConstant(0x4b000000, DL, MVT::i32);
12086 SDValue CstLowArray[] = {CstLow, CstLow, CstLow, CstLow,
12087 CstLow, CstLow, CstLow, CstLow};
12088 SDValue VecCstLow = DAG.getNode(ISD::BUILD_VECTOR, DL, VecIntVT,
12089 makeArrayRef(&CstLowArray[0], NumElts));
12090 // Create the splat vector for 0x53000000.
12091 SDValue CstHigh = DAG.getConstant(0x53000000, DL, MVT::i32);
12092 SDValue CstHighArray[] = {CstHigh, CstHigh, CstHigh, CstHigh,
12093 CstHigh, CstHigh, CstHigh, CstHigh};
12094 SDValue VecCstHigh = DAG.getNode(ISD::BUILD_VECTOR, DL, VecIntVT,
12095 makeArrayRef(&CstHighArray[0], NumElts));
12097 // Create the right shift.
12098 SDValue CstShift = DAG.getConstant(16, DL, MVT::i32);
12099 SDValue CstShiftArray[] = {CstShift, CstShift, CstShift, CstShift,
12100 CstShift, CstShift, CstShift, CstShift};
12101 SDValue VecCstShift = DAG.getNode(ISD::BUILD_VECTOR, DL, VecIntVT,
12102 makeArrayRef(&CstShiftArray[0], NumElts));
12103 SDValue HighShift = DAG.getNode(ISD::SRL, DL, VecIntVT, V, VecCstShift);
12106 if (Subtarget.hasSSE41()) {
12107 EVT VecI16VT = Is128 ? MVT::v8i16 : MVT::v16i16;
12108 // uint4 lo = _mm_blend_epi16( v, (uint4) 0x4b000000, 0xaa);
12109 SDValue VecCstLowBitcast = DAG.getBitcast(VecI16VT, VecCstLow);
12110 SDValue VecBitcast = DAG.getBitcast(VecI16VT, V);
12111 // Low will be bitcasted right away, so do not bother bitcasting back to its
12113 Low = DAG.getNode(X86ISD::BLENDI, DL, VecI16VT, VecBitcast,
12114 VecCstLowBitcast, DAG.getConstant(0xaa, DL, MVT::i32));
12115 // uint4 hi = _mm_blend_epi16( _mm_srli_epi32(v,16),
12116 // (uint4) 0x53000000, 0xaa);
12117 SDValue VecCstHighBitcast = DAG.getBitcast(VecI16VT, VecCstHigh);
12118 SDValue VecShiftBitcast = DAG.getBitcast(VecI16VT, HighShift);
12119 // High will be bitcasted right away, so do not bother bitcasting back to
12120 // its original type.
12121 High = DAG.getNode(X86ISD::BLENDI, DL, VecI16VT, VecShiftBitcast,
12122 VecCstHighBitcast, DAG.getConstant(0xaa, DL, MVT::i32));
12124 SDValue CstMask = DAG.getConstant(0xffff, DL, MVT::i32);
12125 SDValue VecCstMask = DAG.getNode(ISD::BUILD_VECTOR, DL, VecIntVT, CstMask,
12126 CstMask, CstMask, CstMask);
12127 // uint4 lo = (v & (uint4) 0xffff) | (uint4) 0x4b000000;
12128 SDValue LowAnd = DAG.getNode(ISD::AND, DL, VecIntVT, V, VecCstMask);
12129 Low = DAG.getNode(ISD::OR, DL, VecIntVT, LowAnd, VecCstLow);
12131 // uint4 hi = (v >> 16) | (uint4) 0x53000000;
12132 High = DAG.getNode(ISD::OR, DL, VecIntVT, HighShift, VecCstHigh);
12135 // Create the vector constant for -(0x1.0p39f + 0x1.0p23f).
12136 SDValue CstFAdd = DAG.getConstantFP(
12137 APFloat(APFloat::IEEEsingle, APInt(32, 0xD3000080)), DL, MVT::f32);
12138 SDValue CstFAddArray[] = {CstFAdd, CstFAdd, CstFAdd, CstFAdd,
12139 CstFAdd, CstFAdd, CstFAdd, CstFAdd};
12140 SDValue VecCstFAdd = DAG.getNode(ISD::BUILD_VECTOR, DL, VecFloatVT,
12141 makeArrayRef(&CstFAddArray[0], NumElts));
12143 // float4 fhi = (float4) hi - (0x1.0p39f + 0x1.0p23f);
12144 SDValue HighBitcast = DAG.getBitcast(VecFloatVT, High);
12146 DAG.getNode(ISD::FADD, DL, VecFloatVT, HighBitcast, VecCstFAdd);
12147 // return (float4) lo + fhi;
12148 SDValue LowBitcast = DAG.getBitcast(VecFloatVT, Low);
12149 return DAG.getNode(ISD::FADD, DL, VecFloatVT, LowBitcast, FHigh);
12152 SDValue X86TargetLowering::lowerUINT_TO_FP_vec(SDValue Op,
12153 SelectionDAG &DAG) const {
12154 SDValue N0 = Op.getOperand(0);
12155 MVT SVT = N0.getSimpleValueType();
12158 switch (SVT.SimpleTy) {
12160 llvm_unreachable("Custom UINT_TO_FP is not supported!");
12165 MVT NVT = MVT::getVectorVT(MVT::i32, SVT.getVectorNumElements());
12166 return DAG.getNode(ISD::SINT_TO_FP, dl, Op.getValueType(),
12167 DAG.getNode(ISD::ZERO_EXTEND, dl, NVT, N0));
12171 return lowerUINT_TO_FP_vXi32(Op, DAG, *Subtarget);
12174 if (Subtarget->hasAVX512())
12175 return DAG.getNode(ISD::UINT_TO_FP, dl, Op.getValueType(),
12176 DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v16i32, N0));
12178 llvm_unreachable(nullptr);
12181 SDValue X86TargetLowering::LowerUINT_TO_FP(SDValue Op,
12182 SelectionDAG &DAG) const {
12183 SDValue N0 = Op.getOperand(0);
12185 auto PtrVT = getPointerTy(DAG.getDataLayout());
12187 if (Op.getValueType().isVector())
12188 return lowerUINT_TO_FP_vec(Op, DAG);
12190 // Since UINT_TO_FP is legal (it's marked custom), dag combiner won't
12191 // optimize it to a SINT_TO_FP when the sign bit is known zero. Perform
12192 // the optimization here.
12193 if (DAG.SignBitIsZero(N0))
12194 return DAG.getNode(ISD::SINT_TO_FP, dl, Op.getValueType(), N0);
12196 MVT SrcVT = N0.getSimpleValueType();
12197 MVT DstVT = Op.getSimpleValueType();
12198 if (SrcVT == MVT::i64 && DstVT == MVT::f64 && X86ScalarSSEf64)
12199 return LowerUINT_TO_FP_i64(Op, DAG);
12200 if (SrcVT == MVT::i32 && X86ScalarSSEf64)
12201 return LowerUINT_TO_FP_i32(Op, DAG);
12202 if (Subtarget->is64Bit() && SrcVT == MVT::i64 && DstVT == MVT::f32)
12205 // Make a 64-bit buffer, and use it to build an FILD.
12206 SDValue StackSlot = DAG.CreateStackTemporary(MVT::i64);
12207 if (SrcVT == MVT::i32) {
12208 SDValue WordOff = DAG.getConstant(4, dl, PtrVT);
12209 SDValue OffsetSlot = DAG.getNode(ISD::ADD, dl, PtrVT, StackSlot, WordOff);
12210 SDValue Store1 = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
12211 StackSlot, MachinePointerInfo(),
12213 SDValue Store2 = DAG.getStore(Store1, dl, DAG.getConstant(0, dl, MVT::i32),
12214 OffsetSlot, MachinePointerInfo(),
12216 SDValue Fild = BuildFILD(Op, MVT::i64, Store2, StackSlot, DAG);
12220 assert(SrcVT == MVT::i64 && "Unexpected type in UINT_TO_FP");
12221 SDValue Store = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
12222 StackSlot, MachinePointerInfo(),
12224 // For i64 source, we need to add the appropriate power of 2 if the input
12225 // was negative. This is the same as the optimization in
12226 // DAGTypeLegalizer::ExpandIntOp_UNIT_TO_FP, and for it to be safe here,
12227 // we must be careful to do the computation in x87 extended precision, not
12228 // in SSE. (The generic code can't know it's OK to do this, or how to.)
12229 int SSFI = cast<FrameIndexSDNode>(StackSlot)->getIndex();
12230 MachineMemOperand *MMO =
12231 DAG.getMachineFunction()
12232 .getMachineMemOperand(MachinePointerInfo::getFixedStack(SSFI),
12233 MachineMemOperand::MOLoad, 8, 8);
12235 SDVTList Tys = DAG.getVTList(MVT::f80, MVT::Other);
12236 SDValue Ops[] = { Store, StackSlot, DAG.getValueType(MVT::i64) };
12237 SDValue Fild = DAG.getMemIntrinsicNode(X86ISD::FILD, dl, Tys, Ops,
12240 APInt FF(32, 0x5F800000ULL);
12242 // Check whether the sign bit is set.
12243 SDValue SignSet = DAG.getSetCC(
12244 dl, getSetCCResultType(DAG.getDataLayout(), *DAG.getContext(), MVT::i64),
12245 Op.getOperand(0), DAG.getConstant(0, dl, MVT::i64), ISD::SETLT);
12247 // Build a 64 bit pair (0, FF) in the constant pool, with FF in the lo bits.
12248 SDValue FudgePtr = DAG.getConstantPool(
12249 ConstantInt::get(*DAG.getContext(), FF.zext(64)), PtrVT);
12251 // Get a pointer to FF if the sign bit was set, or to 0 otherwise.
12252 SDValue Zero = DAG.getIntPtrConstant(0, dl);
12253 SDValue Four = DAG.getIntPtrConstant(4, dl);
12254 SDValue Offset = DAG.getNode(ISD::SELECT, dl, Zero.getValueType(), SignSet,
12256 FudgePtr = DAG.getNode(ISD::ADD, dl, PtrVT, FudgePtr, Offset);
12258 // Load the value out, extending it from f32 to f80.
12259 // FIXME: Avoid the extend by constructing the right constant pool?
12260 SDValue Fudge = DAG.getExtLoad(ISD::EXTLOAD, dl, MVT::f80, DAG.getEntryNode(),
12261 FudgePtr, MachinePointerInfo::getConstantPool(),
12262 MVT::f32, false, false, false, 4);
12263 // Extend everything to 80 bits to force it to be done on x87.
12264 SDValue Add = DAG.getNode(ISD::FADD, dl, MVT::f80, Fild, Fudge);
12265 return DAG.getNode(ISD::FP_ROUND, dl, DstVT, Add,
12266 DAG.getIntPtrConstant(0, dl));
12269 std::pair<SDValue,SDValue>
12270 X86TargetLowering:: FP_TO_INTHelper(SDValue Op, SelectionDAG &DAG,
12271 bool IsSigned, bool IsReplace) const {
12274 EVT DstTy = Op.getValueType();
12275 auto PtrVT = getPointerTy(DAG.getDataLayout());
12277 if (!IsSigned && !isIntegerTypeFTOL(DstTy)) {
12278 assert(DstTy == MVT::i32 && "Unexpected FP_TO_UINT");
12282 assert(DstTy.getSimpleVT() <= MVT::i64 &&
12283 DstTy.getSimpleVT() >= MVT::i16 &&
12284 "Unknown FP_TO_INT to lower!");
12286 // These are really Legal.
12287 if (DstTy == MVT::i32 &&
12288 isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType()))
12289 return std::make_pair(SDValue(), SDValue());
12290 if (Subtarget->is64Bit() &&
12291 DstTy == MVT::i64 &&
12292 isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType()))
12293 return std::make_pair(SDValue(), SDValue());
12295 // We lower FP->int64 either into FISTP64 followed by a load from a temporary
12296 // stack slot, or into the FTOL runtime function.
12297 MachineFunction &MF = DAG.getMachineFunction();
12298 unsigned MemSize = DstTy.getSizeInBits()/8;
12299 int SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize, false);
12300 SDValue StackSlot = DAG.getFrameIndex(SSFI, PtrVT);
12303 if (!IsSigned && isIntegerTypeFTOL(DstTy))
12304 Opc = X86ISD::WIN_FTOL;
12306 switch (DstTy.getSimpleVT().SimpleTy) {
12307 default: llvm_unreachable("Invalid FP_TO_SINT to lower!");
12308 case MVT::i16: Opc = X86ISD::FP_TO_INT16_IN_MEM; break;
12309 case MVT::i32: Opc = X86ISD::FP_TO_INT32_IN_MEM; break;
12310 case MVT::i64: Opc = X86ISD::FP_TO_INT64_IN_MEM; break;
12313 SDValue Chain = DAG.getEntryNode();
12314 SDValue Value = Op.getOperand(0);
12315 EVT TheVT = Op.getOperand(0).getValueType();
12316 // FIXME This causes a redundant load/store if the SSE-class value is already
12317 // in memory, such as if it is on the callstack.
12318 if (isScalarFPTypeInSSEReg(TheVT)) {
12319 assert(DstTy == MVT::i64 && "Invalid FP_TO_SINT to lower!");
12320 Chain = DAG.getStore(Chain, DL, Value, StackSlot,
12321 MachinePointerInfo::getFixedStack(SSFI),
12323 SDVTList Tys = DAG.getVTList(Op.getOperand(0).getValueType(), MVT::Other);
12325 Chain, StackSlot, DAG.getValueType(TheVT)
12328 MachineMemOperand *MMO =
12329 MF.getMachineMemOperand(MachinePointerInfo::getFixedStack(SSFI),
12330 MachineMemOperand::MOLoad, MemSize, MemSize);
12331 Value = DAG.getMemIntrinsicNode(X86ISD::FLD, DL, Tys, Ops, DstTy, MMO);
12332 Chain = Value.getValue(1);
12333 SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize, false);
12334 StackSlot = DAG.getFrameIndex(SSFI, PtrVT);
12337 MachineMemOperand *MMO =
12338 MF.getMachineMemOperand(MachinePointerInfo::getFixedStack(SSFI),
12339 MachineMemOperand::MOStore, MemSize, MemSize);
12341 if (Opc != X86ISD::WIN_FTOL) {
12342 // Build the FP_TO_INT*_IN_MEM
12343 SDValue Ops[] = { Chain, Value, StackSlot };
12344 SDValue FIST = DAG.getMemIntrinsicNode(Opc, DL, DAG.getVTList(MVT::Other),
12346 return std::make_pair(FIST, StackSlot);
12348 SDValue ftol = DAG.getNode(X86ISD::WIN_FTOL, DL,
12349 DAG.getVTList(MVT::Other, MVT::Glue),
12351 SDValue eax = DAG.getCopyFromReg(ftol, DL, X86::EAX,
12352 MVT::i32, ftol.getValue(1));
12353 SDValue edx = DAG.getCopyFromReg(eax.getValue(1), DL, X86::EDX,
12354 MVT::i32, eax.getValue(2));
12355 SDValue Ops[] = { eax, edx };
12356 SDValue pair = IsReplace
12357 ? DAG.getNode(ISD::BUILD_PAIR, DL, MVT::i64, Ops)
12358 : DAG.getMergeValues(Ops, DL);
12359 return std::make_pair(pair, SDValue());
12363 static SDValue LowerAVXExtend(SDValue Op, SelectionDAG &DAG,
12364 const X86Subtarget *Subtarget) {
12365 MVT VT = Op->getSimpleValueType(0);
12366 SDValue In = Op->getOperand(0);
12367 MVT InVT = In.getSimpleValueType();
12370 if (VT.is512BitVector() || InVT.getScalarType() == MVT::i1)
12371 return DAG.getNode(ISD::ZERO_EXTEND, dl, VT, In);
12373 // Optimize vectors in AVX mode:
12376 // Use vpunpcklwd for 4 lower elements v8i16 -> v4i32.
12377 // Use vpunpckhwd for 4 upper elements v8i16 -> v4i32.
12378 // Concat upper and lower parts.
12381 // Use vpunpckldq for 4 lower elements v4i32 -> v2i64.
12382 // Use vpunpckhdq for 4 upper elements v4i32 -> v2i64.
12383 // Concat upper and lower parts.
12386 if (((VT != MVT::v16i16) || (InVT != MVT::v16i8)) &&
12387 ((VT != MVT::v8i32) || (InVT != MVT::v8i16)) &&
12388 ((VT != MVT::v4i64) || (InVT != MVT::v4i32)))
12391 if (Subtarget->hasInt256())
12392 return DAG.getNode(X86ISD::VZEXT, dl, VT, In);
12394 SDValue ZeroVec = getZeroVector(InVT, Subtarget, DAG, dl);
12395 SDValue Undef = DAG.getUNDEF(InVT);
12396 bool NeedZero = Op.getOpcode() == ISD::ZERO_EXTEND;
12397 SDValue OpLo = getUnpackl(DAG, dl, InVT, In, NeedZero ? ZeroVec : Undef);
12398 SDValue OpHi = getUnpackh(DAG, dl, InVT, In, NeedZero ? ZeroVec : Undef);
12400 MVT HVT = MVT::getVectorVT(VT.getVectorElementType(),
12401 VT.getVectorNumElements()/2);
12403 OpLo = DAG.getBitcast(HVT, OpLo);
12404 OpHi = DAG.getBitcast(HVT, OpHi);
12406 return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT, OpLo, OpHi);
12409 static SDValue LowerZERO_EXTEND_AVX512(SDValue Op,
12410 const X86Subtarget *Subtarget, SelectionDAG &DAG) {
12411 MVT VT = Op->getSimpleValueType(0);
12412 SDValue In = Op->getOperand(0);
12413 MVT InVT = In.getSimpleValueType();
12415 unsigned int NumElts = VT.getVectorNumElements();
12416 if (NumElts != 8 && NumElts != 16 && !Subtarget->hasBWI())
12419 if (VT.is512BitVector() && InVT.getVectorElementType() != MVT::i1)
12420 return DAG.getNode(X86ISD::VZEXT, DL, VT, In);
12422 assert(InVT.getVectorElementType() == MVT::i1);
12423 MVT ExtVT = NumElts == 8 ? MVT::v8i64 : MVT::v16i32;
12425 DAG.getConstant(APInt(ExtVT.getScalarSizeInBits(), 1), DL, ExtVT);
12427 DAG.getConstant(APInt::getNullValue(ExtVT.getScalarSizeInBits()), DL, ExtVT);
12429 SDValue V = DAG.getNode(ISD::VSELECT, DL, ExtVT, In, One, Zero);
12430 if (VT.is512BitVector())
12432 return DAG.getNode(X86ISD::VTRUNC, DL, VT, V);
12435 static SDValue LowerANY_EXTEND(SDValue Op, const X86Subtarget *Subtarget,
12436 SelectionDAG &DAG) {
12437 if (Subtarget->hasFp256())
12438 if (SDValue Res = LowerAVXExtend(Op, DAG, Subtarget))
12444 static SDValue LowerZERO_EXTEND(SDValue Op, const X86Subtarget *Subtarget,
12445 SelectionDAG &DAG) {
12447 MVT VT = Op.getSimpleValueType();
12448 SDValue In = Op.getOperand(0);
12449 MVT SVT = In.getSimpleValueType();
12451 if (VT.is512BitVector() || SVT.getVectorElementType() == MVT::i1)
12452 return LowerZERO_EXTEND_AVX512(Op, Subtarget, DAG);
12454 if (Subtarget->hasFp256())
12455 if (SDValue Res = LowerAVXExtend(Op, DAG, Subtarget))
12458 assert(!VT.is256BitVector() || !SVT.is128BitVector() ||
12459 VT.getVectorNumElements() != SVT.getVectorNumElements());
12463 SDValue X86TargetLowering::LowerTRUNCATE(SDValue Op, SelectionDAG &DAG) const {
12465 MVT VT = Op.getSimpleValueType();
12466 SDValue In = Op.getOperand(0);
12467 MVT InVT = In.getSimpleValueType();
12469 if (VT == MVT::i1) {
12470 assert((InVT.isInteger() && (InVT.getSizeInBits() <= 64)) &&
12471 "Invalid scalar TRUNCATE operation");
12472 if (InVT.getSizeInBits() >= 32)
12474 In = DAG.getNode(ISD::ANY_EXTEND, DL, MVT::i32, In);
12475 return DAG.getNode(ISD::TRUNCATE, DL, VT, In);
12477 assert(VT.getVectorNumElements() == InVT.getVectorNumElements() &&
12478 "Invalid TRUNCATE operation");
12480 // move vector to mask - truncate solution for SKX
12481 if (VT.getVectorElementType() == MVT::i1) {
12482 if (InVT.is512BitVector() && InVT.getScalarSizeInBits() <= 16 &&
12483 Subtarget->hasBWI())
12484 return Op; // legal, will go to VPMOVB2M, VPMOVW2M
12485 if ((InVT.is256BitVector() || InVT.is128BitVector())
12486 && InVT.getScalarSizeInBits() <= 16 &&
12487 Subtarget->hasBWI() && Subtarget->hasVLX())
12488 return Op; // legal, will go to VPMOVB2M, VPMOVW2M
12489 if (InVT.is512BitVector() && InVT.getScalarSizeInBits() >= 32 &&
12490 Subtarget->hasDQI())
12491 return Op; // legal, will go to VPMOVD2M, VPMOVQ2M
12492 if ((InVT.is256BitVector() || InVT.is128BitVector())
12493 && InVT.getScalarSizeInBits() >= 32 &&
12494 Subtarget->hasDQI() && Subtarget->hasVLX())
12495 return Op; // legal, will go to VPMOVB2M, VPMOVQ2M
12498 if (VT.getVectorElementType() == MVT::i1) {
12499 assert(VT.getVectorElementType() == MVT::i1 && "Unexpected vector type");
12500 unsigned NumElts = InVT.getVectorNumElements();
12501 assert ((NumElts == 8 || NumElts == 16) && "Unexpected vector type");
12502 if (InVT.getSizeInBits() < 512) {
12503 MVT ExtVT = (NumElts == 16)? MVT::v16i32 : MVT::v8i64;
12504 In = DAG.getNode(ISD::SIGN_EXTEND, DL, ExtVT, In);
12509 DAG.getConstant(APInt::getSignBit(InVT.getScalarSizeInBits()), DL, InVT);
12510 SDValue And = DAG.getNode(ISD::AND, DL, InVT, OneV, In);
12511 return DAG.getNode(X86ISD::TESTM, DL, VT, And, And);
12514 // vpmovqb/w/d, vpmovdb/w, vpmovwb
12515 if (((!InVT.is512BitVector() && Subtarget->hasVLX()) || InVT.is512BitVector()) &&
12516 (InVT.getVectorElementType() != MVT::i16 || Subtarget->hasBWI()))
12517 return DAG.getNode(X86ISD::VTRUNC, DL, VT, In);
12519 if ((VT == MVT::v4i32) && (InVT == MVT::v4i64)) {
12520 // On AVX2, v4i64 -> v4i32 becomes VPERMD.
12521 if (Subtarget->hasInt256()) {
12522 static const int ShufMask[] = {0, 2, 4, 6, -1, -1, -1, -1};
12523 In = DAG.getBitcast(MVT::v8i32, In);
12524 In = DAG.getVectorShuffle(MVT::v8i32, DL, In, DAG.getUNDEF(MVT::v8i32),
12526 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, VT, In,
12527 DAG.getIntPtrConstant(0, DL));
12530 SDValue OpLo = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v2i64, In,
12531 DAG.getIntPtrConstant(0, DL));
12532 SDValue OpHi = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v2i64, In,
12533 DAG.getIntPtrConstant(2, DL));
12534 OpLo = DAG.getBitcast(MVT::v4i32, OpLo);
12535 OpHi = DAG.getBitcast(MVT::v4i32, OpHi);
12536 static const int ShufMask[] = {0, 2, 4, 6};
12537 return DAG.getVectorShuffle(VT, DL, OpLo, OpHi, ShufMask);
12540 if ((VT == MVT::v8i16) && (InVT == MVT::v8i32)) {
12541 // On AVX2, v8i32 -> v8i16 becomed PSHUFB.
12542 if (Subtarget->hasInt256()) {
12543 In = DAG.getBitcast(MVT::v32i8, In);
12545 SmallVector<SDValue,32> pshufbMask;
12546 for (unsigned i = 0; i < 2; ++i) {
12547 pshufbMask.push_back(DAG.getConstant(0x0, DL, MVT::i8));
12548 pshufbMask.push_back(DAG.getConstant(0x1, DL, MVT::i8));
12549 pshufbMask.push_back(DAG.getConstant(0x4, DL, MVT::i8));
12550 pshufbMask.push_back(DAG.getConstant(0x5, DL, MVT::i8));
12551 pshufbMask.push_back(DAG.getConstant(0x8, DL, MVT::i8));
12552 pshufbMask.push_back(DAG.getConstant(0x9, DL, MVT::i8));
12553 pshufbMask.push_back(DAG.getConstant(0xc, DL, MVT::i8));
12554 pshufbMask.push_back(DAG.getConstant(0xd, DL, MVT::i8));
12555 for (unsigned j = 0; j < 8; ++j)
12556 pshufbMask.push_back(DAG.getConstant(0x80, DL, MVT::i8));
12558 SDValue BV = DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v32i8, pshufbMask);
12559 In = DAG.getNode(X86ISD::PSHUFB, DL, MVT::v32i8, In, BV);
12560 In = DAG.getBitcast(MVT::v4i64, In);
12562 static const int ShufMask[] = {0, 2, -1, -1};
12563 In = DAG.getVectorShuffle(MVT::v4i64, DL, In, DAG.getUNDEF(MVT::v4i64),
12565 In = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v2i64, In,
12566 DAG.getIntPtrConstant(0, DL));
12567 return DAG.getBitcast(VT, In);
12570 SDValue OpLo = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v4i32, In,
12571 DAG.getIntPtrConstant(0, DL));
12573 SDValue OpHi = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v4i32, In,
12574 DAG.getIntPtrConstant(4, DL));
12576 OpLo = DAG.getBitcast(MVT::v16i8, OpLo);
12577 OpHi = DAG.getBitcast(MVT::v16i8, OpHi);
12579 // The PSHUFB mask:
12580 static const int ShufMask1[] = {0, 1, 4, 5, 8, 9, 12, 13,
12581 -1, -1, -1, -1, -1, -1, -1, -1};
12583 SDValue Undef = DAG.getUNDEF(MVT::v16i8);
12584 OpLo = DAG.getVectorShuffle(MVT::v16i8, DL, OpLo, Undef, ShufMask1);
12585 OpHi = DAG.getVectorShuffle(MVT::v16i8, DL, OpHi, Undef, ShufMask1);
12587 OpLo = DAG.getBitcast(MVT::v4i32, OpLo);
12588 OpHi = DAG.getBitcast(MVT::v4i32, OpHi);
12590 // The MOVLHPS Mask:
12591 static const int ShufMask2[] = {0, 1, 4, 5};
12592 SDValue res = DAG.getVectorShuffle(MVT::v4i32, DL, OpLo, OpHi, ShufMask2);
12593 return DAG.getBitcast(MVT::v8i16, res);
12596 // Handle truncation of V256 to V128 using shuffles.
12597 if (!VT.is128BitVector() || !InVT.is256BitVector())
12600 assert(Subtarget->hasFp256() && "256-bit vector without AVX!");
12602 unsigned NumElems = VT.getVectorNumElements();
12603 MVT NVT = MVT::getVectorVT(VT.getVectorElementType(), NumElems * 2);
12605 SmallVector<int, 16> MaskVec(NumElems * 2, -1);
12606 // Prepare truncation shuffle mask
12607 for (unsigned i = 0; i != NumElems; ++i)
12608 MaskVec[i] = i * 2;
12609 SDValue V = DAG.getVectorShuffle(NVT, DL, DAG.getBitcast(NVT, In),
12610 DAG.getUNDEF(NVT), &MaskVec[0]);
12611 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, VT, V,
12612 DAG.getIntPtrConstant(0, DL));
12615 SDValue X86TargetLowering::LowerFP_TO_SINT(SDValue Op,
12616 SelectionDAG &DAG) const {
12617 assert(!Op.getSimpleValueType().isVector());
12619 std::pair<SDValue,SDValue> Vals = FP_TO_INTHelper(Op, DAG,
12620 /*IsSigned=*/ true, /*IsReplace=*/ false);
12621 SDValue FIST = Vals.first, StackSlot = Vals.second;
12622 // If FP_TO_INTHelper failed, the node is actually supposed to be Legal.
12623 if (!FIST.getNode()) return Op;
12625 if (StackSlot.getNode())
12626 // Load the result.
12627 return DAG.getLoad(Op.getValueType(), SDLoc(Op),
12628 FIST, StackSlot, MachinePointerInfo(),
12629 false, false, false, 0);
12631 // The node is the result.
12635 SDValue X86TargetLowering::LowerFP_TO_UINT(SDValue Op,
12636 SelectionDAG &DAG) const {
12637 std::pair<SDValue,SDValue> Vals = FP_TO_INTHelper(Op, DAG,
12638 /*IsSigned=*/ false, /*IsReplace=*/ false);
12639 SDValue FIST = Vals.first, StackSlot = Vals.second;
12640 assert(FIST.getNode() && "Unexpected failure");
12642 if (StackSlot.getNode())
12643 // Load the result.
12644 return DAG.getLoad(Op.getValueType(), SDLoc(Op),
12645 FIST, StackSlot, MachinePointerInfo(),
12646 false, false, false, 0);
12648 // The node is the result.
12652 static SDValue LowerFP_EXTEND(SDValue Op, SelectionDAG &DAG) {
12654 MVT VT = Op.getSimpleValueType();
12655 SDValue In = Op.getOperand(0);
12656 MVT SVT = In.getSimpleValueType();
12658 assert(SVT == MVT::v2f32 && "Only customize MVT::v2f32 type legalization!");
12660 return DAG.getNode(X86ISD::VFPEXT, DL, VT,
12661 DAG.getNode(ISD::CONCAT_VECTORS, DL, MVT::v4f32,
12662 In, DAG.getUNDEF(SVT)));
12665 /// The only differences between FABS and FNEG are the mask and the logic op.
12666 /// FNEG also has a folding opportunity for FNEG(FABS(x)).
12667 static SDValue LowerFABSorFNEG(SDValue Op, SelectionDAG &DAG) {
12668 assert((Op.getOpcode() == ISD::FABS || Op.getOpcode() == ISD::FNEG) &&
12669 "Wrong opcode for lowering FABS or FNEG.");
12671 bool IsFABS = (Op.getOpcode() == ISD::FABS);
12673 // If this is a FABS and it has an FNEG user, bail out to fold the combination
12674 // into an FNABS. We'll lower the FABS after that if it is still in use.
12676 for (SDNode *User : Op->uses())
12677 if (User->getOpcode() == ISD::FNEG)
12681 MVT VT = Op.getSimpleValueType();
12683 // FIXME: Use function attribute "OptimizeForSize" and/or CodeGenOpt::Level to
12684 // decide if we should generate a 16-byte constant mask when we only need 4 or
12685 // 8 bytes for the scalar case.
12691 if (VT.isVector()) {
12693 EltVT = VT.getVectorElementType();
12694 NumElts = VT.getVectorNumElements();
12696 // There are no scalar bitwise logical SSE/AVX instructions, so we
12697 // generate a 16-byte vector constant and logic op even for the scalar case.
12698 // Using a 16-byte mask allows folding the load of the mask with
12699 // the logic op, so it can save (~4 bytes) on code size.
12700 LogicVT = (VT == MVT::f64) ? MVT::v2f64 : MVT::v4f32;
12702 NumElts = (VT == MVT::f64) ? 2 : 4;
12705 unsigned EltBits = EltVT.getSizeInBits();
12706 LLVMContext *Context = DAG.getContext();
12707 // For FABS, mask is 0x7f...; for FNEG, mask is 0x80...
12709 IsFABS ? APInt::getSignedMaxValue(EltBits) : APInt::getSignBit(EltBits);
12710 Constant *C = ConstantInt::get(*Context, MaskElt);
12711 C = ConstantVector::getSplat(NumElts, C);
12712 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
12713 SDValue CPIdx = DAG.getConstantPool(C, TLI.getPointerTy(DAG.getDataLayout()));
12714 unsigned Alignment = cast<ConstantPoolSDNode>(CPIdx)->getAlignment();
12715 SDValue Mask = DAG.getLoad(LogicVT, dl, DAG.getEntryNode(), CPIdx,
12716 MachinePointerInfo::getConstantPool(),
12717 false, false, false, Alignment);
12719 SDValue Op0 = Op.getOperand(0);
12720 bool IsFNABS = !IsFABS && (Op0.getOpcode() == ISD::FABS);
12722 IsFABS ? X86ISD::FAND : IsFNABS ? X86ISD::FOR : X86ISD::FXOR;
12723 SDValue Operand = IsFNABS ? Op0.getOperand(0) : Op0;
12726 return DAG.getNode(LogicOp, dl, LogicVT, Operand, Mask);
12728 // For the scalar case extend to a 128-bit vector, perform the logic op,
12729 // and extract the scalar result back out.
12730 Operand = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, LogicVT, Operand);
12731 SDValue LogicNode = DAG.getNode(LogicOp, dl, LogicVT, Operand, Mask);
12732 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, LogicNode,
12733 DAG.getIntPtrConstant(0, dl));
12736 static SDValue LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) {
12737 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
12738 LLVMContext *Context = DAG.getContext();
12739 SDValue Op0 = Op.getOperand(0);
12740 SDValue Op1 = Op.getOperand(1);
12742 MVT VT = Op.getSimpleValueType();
12743 MVT SrcVT = Op1.getSimpleValueType();
12745 // If second operand is smaller, extend it first.
12746 if (SrcVT.bitsLT(VT)) {
12747 Op1 = DAG.getNode(ISD::FP_EXTEND, dl, VT, Op1);
12750 // And if it is bigger, shrink it first.
12751 if (SrcVT.bitsGT(VT)) {
12752 Op1 = DAG.getNode(ISD::FP_ROUND, dl, VT, Op1, DAG.getIntPtrConstant(1, dl));
12756 // At this point the operands and the result should have the same
12757 // type, and that won't be f80 since that is not custom lowered.
12759 const fltSemantics &Sem =
12760 VT == MVT::f64 ? APFloat::IEEEdouble : APFloat::IEEEsingle;
12761 const unsigned SizeInBits = VT.getSizeInBits();
12763 SmallVector<Constant *, 4> CV(
12764 VT == MVT::f64 ? 2 : 4,
12765 ConstantFP::get(*Context, APFloat(Sem, APInt(SizeInBits, 0))));
12767 // First, clear all bits but the sign bit from the second operand (sign).
12768 CV[0] = ConstantFP::get(*Context,
12769 APFloat(Sem, APInt::getHighBitsSet(SizeInBits, 1)));
12770 Constant *C = ConstantVector::get(CV);
12771 auto PtrVT = TLI.getPointerTy(DAG.getDataLayout());
12772 SDValue CPIdx = DAG.getConstantPool(C, PtrVT, 16);
12774 // Perform all logic operations as 16-byte vectors because there are no
12775 // scalar FP logic instructions in SSE. This allows load folding of the
12776 // constants into the logic instructions.
12777 MVT LogicVT = (VT == MVT::f64) ? MVT::v2f64 : MVT::v4f32;
12778 SDValue Mask1 = DAG.getLoad(LogicVT, dl, DAG.getEntryNode(), CPIdx,
12779 MachinePointerInfo::getConstantPool(),
12780 false, false, false, 16);
12781 Op1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, LogicVT, Op1);
12782 SDValue SignBit = DAG.getNode(X86ISD::FAND, dl, LogicVT, Op1, Mask1);
12784 // Next, clear the sign bit from the first operand (magnitude).
12785 // If it's a constant, we can clear it here.
12786 if (ConstantFPSDNode *Op0CN = dyn_cast<ConstantFPSDNode>(Op0)) {
12787 APFloat APF = Op0CN->getValueAPF();
12788 // If the magnitude is a positive zero, the sign bit alone is enough.
12789 if (APF.isPosZero())
12790 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, SrcVT, SignBit,
12791 DAG.getIntPtrConstant(0, dl));
12793 CV[0] = ConstantFP::get(*Context, APF);
12795 CV[0] = ConstantFP::get(
12797 APFloat(Sem, APInt::getLowBitsSet(SizeInBits, SizeInBits - 1)));
12799 C = ConstantVector::get(CV);
12800 CPIdx = DAG.getConstantPool(C, PtrVT, 16);
12801 SDValue Val = DAG.getLoad(LogicVT, dl, DAG.getEntryNode(), CPIdx,
12802 MachinePointerInfo::getConstantPool(),
12803 false, false, false, 16);
12804 // If the magnitude operand wasn't a constant, we need to AND out the sign.
12805 if (!isa<ConstantFPSDNode>(Op0)) {
12806 Op0 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, LogicVT, Op0);
12807 Val = DAG.getNode(X86ISD::FAND, dl, LogicVT, Op0, Val);
12809 // OR the magnitude value with the sign bit.
12810 Val = DAG.getNode(X86ISD::FOR, dl, LogicVT, Val, SignBit);
12811 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, SrcVT, Val,
12812 DAG.getIntPtrConstant(0, dl));
12815 static SDValue LowerFGETSIGN(SDValue Op, SelectionDAG &DAG) {
12816 SDValue N0 = Op.getOperand(0);
12818 MVT VT = Op.getSimpleValueType();
12820 // Lower ISD::FGETSIGN to (AND (X86ISD::FGETSIGNx86 ...) 1).
12821 SDValue xFGETSIGN = DAG.getNode(X86ISD::FGETSIGNx86, dl, VT, N0,
12822 DAG.getConstant(1, dl, VT));
12823 return DAG.getNode(ISD::AND, dl, VT, xFGETSIGN, DAG.getConstant(1, dl, VT));
12826 // Check whether an OR'd tree is PTEST-able.
12827 static SDValue LowerVectorAllZeroTest(SDValue Op, const X86Subtarget *Subtarget,
12828 SelectionDAG &DAG) {
12829 assert(Op.getOpcode() == ISD::OR && "Only check OR'd tree.");
12831 if (!Subtarget->hasSSE41())
12834 if (!Op->hasOneUse())
12837 SDNode *N = Op.getNode();
12840 SmallVector<SDValue, 8> Opnds;
12841 DenseMap<SDValue, unsigned> VecInMap;
12842 SmallVector<SDValue, 8> VecIns;
12843 EVT VT = MVT::Other;
12845 // Recognize a special case where a vector is casted into wide integer to
12847 Opnds.push_back(N->getOperand(0));
12848 Opnds.push_back(N->getOperand(1));
12850 for (unsigned Slot = 0, e = Opnds.size(); Slot < e; ++Slot) {
12851 SmallVectorImpl<SDValue>::const_iterator I = Opnds.begin() + Slot;
12852 // BFS traverse all OR'd operands.
12853 if (I->getOpcode() == ISD::OR) {
12854 Opnds.push_back(I->getOperand(0));
12855 Opnds.push_back(I->getOperand(1));
12856 // Re-evaluate the number of nodes to be traversed.
12857 e += 2; // 2 more nodes (LHS and RHS) are pushed.
12861 // Quit if a non-EXTRACT_VECTOR_ELT
12862 if (I->getOpcode() != ISD::EXTRACT_VECTOR_ELT)
12865 // Quit if without a constant index.
12866 SDValue Idx = I->getOperand(1);
12867 if (!isa<ConstantSDNode>(Idx))
12870 SDValue ExtractedFromVec = I->getOperand(0);
12871 DenseMap<SDValue, unsigned>::iterator M = VecInMap.find(ExtractedFromVec);
12872 if (M == VecInMap.end()) {
12873 VT = ExtractedFromVec.getValueType();
12874 // Quit if not 128/256-bit vector.
12875 if (!VT.is128BitVector() && !VT.is256BitVector())
12877 // Quit if not the same type.
12878 if (VecInMap.begin() != VecInMap.end() &&
12879 VT != VecInMap.begin()->first.getValueType())
12881 M = VecInMap.insert(std::make_pair(ExtractedFromVec, 0)).first;
12882 VecIns.push_back(ExtractedFromVec);
12884 M->second |= 1U << cast<ConstantSDNode>(Idx)->getZExtValue();
12887 assert((VT.is128BitVector() || VT.is256BitVector()) &&
12888 "Not extracted from 128-/256-bit vector.");
12890 unsigned FullMask = (1U << VT.getVectorNumElements()) - 1U;
12892 for (DenseMap<SDValue, unsigned>::const_iterator
12893 I = VecInMap.begin(), E = VecInMap.end(); I != E; ++I) {
12894 // Quit if not all elements are used.
12895 if (I->second != FullMask)
12899 EVT TestVT = VT.is128BitVector() ? MVT::v2i64 : MVT::v4i64;
12901 // Cast all vectors into TestVT for PTEST.
12902 for (unsigned i = 0, e = VecIns.size(); i < e; ++i)
12903 VecIns[i] = DAG.getBitcast(TestVT, VecIns[i]);
12905 // If more than one full vectors are evaluated, OR them first before PTEST.
12906 for (unsigned Slot = 0, e = VecIns.size(); e - Slot > 1; Slot += 2, e += 1) {
12907 // Each iteration will OR 2 nodes and append the result until there is only
12908 // 1 node left, i.e. the final OR'd value of all vectors.
12909 SDValue LHS = VecIns[Slot];
12910 SDValue RHS = VecIns[Slot + 1];
12911 VecIns.push_back(DAG.getNode(ISD::OR, DL, TestVT, LHS, RHS));
12914 return DAG.getNode(X86ISD::PTEST, DL, MVT::i32,
12915 VecIns.back(), VecIns.back());
12918 /// \brief return true if \c Op has a use that doesn't just read flags.
12919 static bool hasNonFlagsUse(SDValue Op) {
12920 for (SDNode::use_iterator UI = Op->use_begin(), UE = Op->use_end(); UI != UE;
12922 SDNode *User = *UI;
12923 unsigned UOpNo = UI.getOperandNo();
12924 if (User->getOpcode() == ISD::TRUNCATE && User->hasOneUse()) {
12925 // Look pass truncate.
12926 UOpNo = User->use_begin().getOperandNo();
12927 User = *User->use_begin();
12930 if (User->getOpcode() != ISD::BRCOND && User->getOpcode() != ISD::SETCC &&
12931 !(User->getOpcode() == ISD::SELECT && UOpNo == 0))
12937 /// Emit nodes that will be selected as "test Op0,Op0", or something
12939 SDValue X86TargetLowering::EmitTest(SDValue Op, unsigned X86CC, SDLoc dl,
12940 SelectionDAG &DAG) const {
12941 if (Op.getValueType() == MVT::i1) {
12942 SDValue ExtOp = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i8, Op);
12943 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, ExtOp,
12944 DAG.getConstant(0, dl, MVT::i8));
12946 // CF and OF aren't always set the way we want. Determine which
12947 // of these we need.
12948 bool NeedCF = false;
12949 bool NeedOF = false;
12952 case X86::COND_A: case X86::COND_AE:
12953 case X86::COND_B: case X86::COND_BE:
12956 case X86::COND_G: case X86::COND_GE:
12957 case X86::COND_L: case X86::COND_LE:
12958 case X86::COND_O: case X86::COND_NO: {
12959 // Check if we really need to set the
12960 // Overflow flag. If NoSignedWrap is present
12961 // that is not actually needed.
12962 switch (Op->getOpcode()) {
12967 const auto *BinNode = cast<BinaryWithFlagsSDNode>(Op.getNode());
12968 if (BinNode->Flags.hasNoSignedWrap())
12978 // See if we can use the EFLAGS value from the operand instead of
12979 // doing a separate TEST. TEST always sets OF and CF to 0, so unless
12980 // we prove that the arithmetic won't overflow, we can't use OF or CF.
12981 if (Op.getResNo() != 0 || NeedOF || NeedCF) {
12982 // Emit a CMP with 0, which is the TEST pattern.
12983 //if (Op.getValueType() == MVT::i1)
12984 // return DAG.getNode(X86ISD::CMP, dl, MVT::i1, Op,
12985 // DAG.getConstant(0, MVT::i1));
12986 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op,
12987 DAG.getConstant(0, dl, Op.getValueType()));
12989 unsigned Opcode = 0;
12990 unsigned NumOperands = 0;
12992 // Truncate operations may prevent the merge of the SETCC instruction
12993 // and the arithmetic instruction before it. Attempt to truncate the operands
12994 // of the arithmetic instruction and use a reduced bit-width instruction.
12995 bool NeedTruncation = false;
12996 SDValue ArithOp = Op;
12997 if (Op->getOpcode() == ISD::TRUNCATE && Op->hasOneUse()) {
12998 SDValue Arith = Op->getOperand(0);
12999 // Both the trunc and the arithmetic op need to have one user each.
13000 if (Arith->hasOneUse())
13001 switch (Arith.getOpcode()) {
13008 NeedTruncation = true;
13014 // NOTICE: In the code below we use ArithOp to hold the arithmetic operation
13015 // which may be the result of a CAST. We use the variable 'Op', which is the
13016 // non-casted variable when we check for possible users.
13017 switch (ArithOp.getOpcode()) {
13019 // Due to an isel shortcoming, be conservative if this add is likely to be
13020 // selected as part of a load-modify-store instruction. When the root node
13021 // in a match is a store, isel doesn't know how to remap non-chain non-flag
13022 // uses of other nodes in the match, such as the ADD in this case. This
13023 // leads to the ADD being left around and reselected, with the result being
13024 // two adds in the output. Alas, even if none our users are stores, that
13025 // doesn't prove we're O.K. Ergo, if we have any parents that aren't
13026 // CopyToReg or SETCC, eschew INC/DEC. A better fix seems to require
13027 // climbing the DAG back to the root, and it doesn't seem to be worth the
13029 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
13030 UE = Op.getNode()->use_end(); UI != UE; ++UI)
13031 if (UI->getOpcode() != ISD::CopyToReg &&
13032 UI->getOpcode() != ISD::SETCC &&
13033 UI->getOpcode() != ISD::STORE)
13036 if (ConstantSDNode *C =
13037 dyn_cast<ConstantSDNode>(ArithOp.getNode()->getOperand(1))) {
13038 // An add of one will be selected as an INC.
13039 if (C->getAPIntValue() == 1 && !Subtarget->slowIncDec()) {
13040 Opcode = X86ISD::INC;
13045 // An add of negative one (subtract of one) will be selected as a DEC.
13046 if (C->getAPIntValue().isAllOnesValue() && !Subtarget->slowIncDec()) {
13047 Opcode = X86ISD::DEC;
13053 // Otherwise use a regular EFLAGS-setting add.
13054 Opcode = X86ISD::ADD;
13059 // If we have a constant logical shift that's only used in a comparison
13060 // against zero turn it into an equivalent AND. This allows turning it into
13061 // a TEST instruction later.
13062 if ((X86CC == X86::COND_E || X86CC == X86::COND_NE) && Op->hasOneUse() &&
13063 isa<ConstantSDNode>(Op->getOperand(1)) && !hasNonFlagsUse(Op)) {
13064 EVT VT = Op.getValueType();
13065 unsigned BitWidth = VT.getSizeInBits();
13066 unsigned ShAmt = Op->getConstantOperandVal(1);
13067 if (ShAmt >= BitWidth) // Avoid undefined shifts.
13069 APInt Mask = ArithOp.getOpcode() == ISD::SRL
13070 ? APInt::getHighBitsSet(BitWidth, BitWidth - ShAmt)
13071 : APInt::getLowBitsSet(BitWidth, BitWidth - ShAmt);
13072 if (!Mask.isSignedIntN(32)) // Avoid large immediates.
13074 SDValue New = DAG.getNode(ISD::AND, dl, VT, Op->getOperand(0),
13075 DAG.getConstant(Mask, dl, VT));
13076 DAG.ReplaceAllUsesWith(Op, New);
13082 // If the primary and result isn't used, don't bother using X86ISD::AND,
13083 // because a TEST instruction will be better.
13084 if (!hasNonFlagsUse(Op))
13090 // Due to the ISEL shortcoming noted above, be conservative if this op is
13091 // likely to be selected as part of a load-modify-store instruction.
13092 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
13093 UE = Op.getNode()->use_end(); UI != UE; ++UI)
13094 if (UI->getOpcode() == ISD::STORE)
13097 // Otherwise use a regular EFLAGS-setting instruction.
13098 switch (ArithOp.getOpcode()) {
13099 default: llvm_unreachable("unexpected operator!");
13100 case ISD::SUB: Opcode = X86ISD::SUB; break;
13101 case ISD::XOR: Opcode = X86ISD::XOR; break;
13102 case ISD::AND: Opcode = X86ISD::AND; break;
13104 if (!NeedTruncation && (X86CC == X86::COND_E || X86CC == X86::COND_NE)) {
13105 SDValue EFLAGS = LowerVectorAllZeroTest(Op, Subtarget, DAG);
13106 if (EFLAGS.getNode())
13109 Opcode = X86ISD::OR;
13123 return SDValue(Op.getNode(), 1);
13129 // If we found that truncation is beneficial, perform the truncation and
13131 if (NeedTruncation) {
13132 EVT VT = Op.getValueType();
13133 SDValue WideVal = Op->getOperand(0);
13134 EVT WideVT = WideVal.getValueType();
13135 unsigned ConvertedOp = 0;
13136 // Use a target machine opcode to prevent further DAGCombine
13137 // optimizations that may separate the arithmetic operations
13138 // from the setcc node.
13139 switch (WideVal.getOpcode()) {
13141 case ISD::ADD: ConvertedOp = X86ISD::ADD; break;
13142 case ISD::SUB: ConvertedOp = X86ISD::SUB; break;
13143 case ISD::AND: ConvertedOp = X86ISD::AND; break;
13144 case ISD::OR: ConvertedOp = X86ISD::OR; break;
13145 case ISD::XOR: ConvertedOp = X86ISD::XOR; break;
13149 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
13150 if (TLI.isOperationLegal(WideVal.getOpcode(), WideVT)) {
13151 SDValue V0 = DAG.getNode(ISD::TRUNCATE, dl, VT, WideVal.getOperand(0));
13152 SDValue V1 = DAG.getNode(ISD::TRUNCATE, dl, VT, WideVal.getOperand(1));
13153 Op = DAG.getNode(ConvertedOp, dl, VT, V0, V1);
13159 // Emit a CMP with 0, which is the TEST pattern.
13160 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op,
13161 DAG.getConstant(0, dl, Op.getValueType()));
13163 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::i32);
13164 SmallVector<SDValue, 4> Ops(Op->op_begin(), Op->op_begin() + NumOperands);
13166 SDValue New = DAG.getNode(Opcode, dl, VTs, Ops);
13167 DAG.ReplaceAllUsesWith(Op, New);
13168 return SDValue(New.getNode(), 1);
13171 /// Emit nodes that will be selected as "cmp Op0,Op1", or something
13173 SDValue X86TargetLowering::EmitCmp(SDValue Op0, SDValue Op1, unsigned X86CC,
13174 SDLoc dl, SelectionDAG &DAG) const {
13175 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op1)) {
13176 if (C->getAPIntValue() == 0)
13177 return EmitTest(Op0, X86CC, dl, DAG);
13179 if (Op0.getValueType() == MVT::i1)
13180 llvm_unreachable("Unexpected comparison operation for MVT::i1 operands");
13183 if ((Op0.getValueType() == MVT::i8 || Op0.getValueType() == MVT::i16 ||
13184 Op0.getValueType() == MVT::i32 || Op0.getValueType() == MVT::i64)) {
13185 // Do the comparison at i32 if it's smaller, besides the Atom case.
13186 // This avoids subregister aliasing issues. Keep the smaller reference
13187 // if we're optimizing for size, however, as that'll allow better folding
13188 // of memory operations.
13189 if (Op0.getValueType() != MVT::i32 && Op0.getValueType() != MVT::i64 &&
13190 !DAG.getMachineFunction().getFunction()->hasFnAttribute(
13191 Attribute::MinSize) &&
13192 !Subtarget->isAtom()) {
13193 unsigned ExtendOp =
13194 isX86CCUnsigned(X86CC) ? ISD::ZERO_EXTEND : ISD::SIGN_EXTEND;
13195 Op0 = DAG.getNode(ExtendOp, dl, MVT::i32, Op0);
13196 Op1 = DAG.getNode(ExtendOp, dl, MVT::i32, Op1);
13198 // Use SUB instead of CMP to enable CSE between SUB and CMP.
13199 SDVTList VTs = DAG.getVTList(Op0.getValueType(), MVT::i32);
13200 SDValue Sub = DAG.getNode(X86ISD::SUB, dl, VTs,
13202 return SDValue(Sub.getNode(), 1);
13204 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op0, Op1);
13207 /// Convert a comparison if required by the subtarget.
13208 SDValue X86TargetLowering::ConvertCmpIfNecessary(SDValue Cmp,
13209 SelectionDAG &DAG) const {
13210 // If the subtarget does not support the FUCOMI instruction, floating-point
13211 // comparisons have to be converted.
13212 if (Subtarget->hasCMov() ||
13213 Cmp.getOpcode() != X86ISD::CMP ||
13214 !Cmp.getOperand(0).getValueType().isFloatingPoint() ||
13215 !Cmp.getOperand(1).getValueType().isFloatingPoint())
13218 // The instruction selector will select an FUCOM instruction instead of
13219 // FUCOMI, which writes the comparison result to FPSW instead of EFLAGS. Hence
13220 // build an SDNode sequence that transfers the result from FPSW into EFLAGS:
13221 // (X86sahf (trunc (srl (X86fp_stsw (trunc (X86cmp ...)), 8))))
13223 SDValue TruncFPSW = DAG.getNode(ISD::TRUNCATE, dl, MVT::i16, Cmp);
13224 SDValue FNStSW = DAG.getNode(X86ISD::FNSTSW16r, dl, MVT::i16, TruncFPSW);
13225 SDValue Srl = DAG.getNode(ISD::SRL, dl, MVT::i16, FNStSW,
13226 DAG.getConstant(8, dl, MVT::i8));
13227 SDValue TruncSrl = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Srl);
13228 return DAG.getNode(X86ISD::SAHF, dl, MVT::i32, TruncSrl);
13231 /// The minimum architected relative accuracy is 2^-12. We need one
13232 /// Newton-Raphson step to have a good float result (24 bits of precision).
13233 SDValue X86TargetLowering::getRsqrtEstimate(SDValue Op,
13234 DAGCombinerInfo &DCI,
13235 unsigned &RefinementSteps,
13236 bool &UseOneConstNR) const {
13237 EVT VT = Op.getValueType();
13238 const char *RecipOp;
13240 // SSE1 has rsqrtss and rsqrtps. AVX adds a 256-bit variant for rsqrtps.
13241 // TODO: Add support for AVX512 (v16f32).
13242 // It is likely not profitable to do this for f64 because a double-precision
13243 // rsqrt estimate with refinement on x86 prior to FMA requires at least 16
13244 // instructions: convert to single, rsqrtss, convert back to double, refine
13245 // (3 steps = at least 13 insts). If an 'rsqrtsd' variant was added to the ISA
13246 // along with FMA, this could be a throughput win.
13247 if (VT == MVT::f32 && Subtarget->hasSSE1())
13249 else if ((VT == MVT::v4f32 && Subtarget->hasSSE1()) ||
13250 (VT == MVT::v8f32 && Subtarget->hasAVX()))
13251 RecipOp = "vec-sqrtf";
13255 TargetRecip Recips = DCI.DAG.getTarget().Options.Reciprocals;
13256 if (!Recips.isEnabled(RecipOp))
13259 RefinementSteps = Recips.getRefinementSteps(RecipOp);
13260 UseOneConstNR = false;
13261 return DCI.DAG.getNode(X86ISD::FRSQRT, SDLoc(Op), VT, Op);
13264 /// The minimum architected relative accuracy is 2^-12. We need one
13265 /// Newton-Raphson step to have a good float result (24 bits of precision).
13266 SDValue X86TargetLowering::getRecipEstimate(SDValue Op,
13267 DAGCombinerInfo &DCI,
13268 unsigned &RefinementSteps) const {
13269 EVT VT = Op.getValueType();
13270 const char *RecipOp;
13272 // SSE1 has rcpss and rcpps. AVX adds a 256-bit variant for rcpps.
13273 // TODO: Add support for AVX512 (v16f32).
13274 // It is likely not profitable to do this for f64 because a double-precision
13275 // reciprocal estimate with refinement on x86 prior to FMA requires
13276 // 15 instructions: convert to single, rcpss, convert back to double, refine
13277 // (3 steps = 12 insts). If an 'rcpsd' variant was added to the ISA
13278 // along with FMA, this could be a throughput win.
13279 if (VT == MVT::f32 && Subtarget->hasSSE1())
13281 else if ((VT == MVT::v4f32 && Subtarget->hasSSE1()) ||
13282 (VT == MVT::v8f32 && Subtarget->hasAVX()))
13283 RecipOp = "vec-divf";
13287 TargetRecip Recips = DCI.DAG.getTarget().Options.Reciprocals;
13288 if (!Recips.isEnabled(RecipOp))
13291 RefinementSteps = Recips.getRefinementSteps(RecipOp);
13292 return DCI.DAG.getNode(X86ISD::FRCP, SDLoc(Op), VT, Op);
13295 /// If we have at least two divisions that use the same divisor, convert to
13296 /// multplication by a reciprocal. This may need to be adjusted for a given
13297 /// CPU if a division's cost is not at least twice the cost of a multiplication.
13298 /// This is because we still need one division to calculate the reciprocal and
13299 /// then we need two multiplies by that reciprocal as replacements for the
13300 /// original divisions.
13301 bool X86TargetLowering::combineRepeatedFPDivisors(unsigned NumUsers) const {
13302 return NumUsers > 1;
13305 static bool isAllOnes(SDValue V) {
13306 ConstantSDNode *C = dyn_cast<ConstantSDNode>(V);
13307 return C && C->isAllOnesValue();
13310 /// LowerToBT - Result of 'and' is compared against zero. Turn it into a BT node
13311 /// if it's possible.
13312 SDValue X86TargetLowering::LowerToBT(SDValue And, ISD::CondCode CC,
13313 SDLoc dl, SelectionDAG &DAG) const {
13314 SDValue Op0 = And.getOperand(0);
13315 SDValue Op1 = And.getOperand(1);
13316 if (Op0.getOpcode() == ISD::TRUNCATE)
13317 Op0 = Op0.getOperand(0);
13318 if (Op1.getOpcode() == ISD::TRUNCATE)
13319 Op1 = Op1.getOperand(0);
13322 if (Op1.getOpcode() == ISD::SHL)
13323 std::swap(Op0, Op1);
13324 if (Op0.getOpcode() == ISD::SHL) {
13325 if (ConstantSDNode *And00C = dyn_cast<ConstantSDNode>(Op0.getOperand(0)))
13326 if (And00C->getZExtValue() == 1) {
13327 // If we looked past a truncate, check that it's only truncating away
13329 unsigned BitWidth = Op0.getValueSizeInBits();
13330 unsigned AndBitWidth = And.getValueSizeInBits();
13331 if (BitWidth > AndBitWidth) {
13333 DAG.computeKnownBits(Op0, Zeros, Ones);
13334 if (Zeros.countLeadingOnes() < BitWidth - AndBitWidth)
13338 RHS = Op0.getOperand(1);
13340 } else if (Op1.getOpcode() == ISD::Constant) {
13341 ConstantSDNode *AndRHS = cast<ConstantSDNode>(Op1);
13342 uint64_t AndRHSVal = AndRHS->getZExtValue();
13343 SDValue AndLHS = Op0;
13345 if (AndRHSVal == 1 && AndLHS.getOpcode() == ISD::SRL) {
13346 LHS = AndLHS.getOperand(0);
13347 RHS = AndLHS.getOperand(1);
13350 // Use BT if the immediate can't be encoded in a TEST instruction.
13351 if (!isUInt<32>(AndRHSVal) && isPowerOf2_64(AndRHSVal)) {
13353 RHS = DAG.getConstant(Log2_64_Ceil(AndRHSVal), dl, LHS.getValueType());
13357 if (LHS.getNode()) {
13358 // If LHS is i8, promote it to i32 with any_extend. There is no i8 BT
13359 // instruction. Since the shift amount is in-range-or-undefined, we know
13360 // that doing a bittest on the i32 value is ok. We extend to i32 because
13361 // the encoding for the i16 version is larger than the i32 version.
13362 // Also promote i16 to i32 for performance / code size reason.
13363 if (LHS.getValueType() == MVT::i8 ||
13364 LHS.getValueType() == MVT::i16)
13365 LHS = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, LHS);
13367 // If the operand types disagree, extend the shift amount to match. Since
13368 // BT ignores high bits (like shifts) we can use anyextend.
13369 if (LHS.getValueType() != RHS.getValueType())
13370 RHS = DAG.getNode(ISD::ANY_EXTEND, dl, LHS.getValueType(), RHS);
13372 SDValue BT = DAG.getNode(X86ISD::BT, dl, MVT::i32, LHS, RHS);
13373 X86::CondCode Cond = CC == ISD::SETEQ ? X86::COND_AE : X86::COND_B;
13374 return DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
13375 DAG.getConstant(Cond, dl, MVT::i8), BT);
13381 /// \brief - Turns an ISD::CondCode into a value suitable for SSE floating point
13383 static int translateX86FSETCC(ISD::CondCode SetCCOpcode, SDValue &Op0,
13388 // SSE Condition code mapping:
13397 switch (SetCCOpcode) {
13398 default: llvm_unreachable("Unexpected SETCC condition");
13400 case ISD::SETEQ: SSECC = 0; break;
13402 case ISD::SETGT: Swap = true; // Fallthrough
13404 case ISD::SETOLT: SSECC = 1; break;
13406 case ISD::SETGE: Swap = true; // Fallthrough
13408 case ISD::SETOLE: SSECC = 2; break;
13409 case ISD::SETUO: SSECC = 3; break;
13411 case ISD::SETNE: SSECC = 4; break;
13412 case ISD::SETULE: Swap = true; // Fallthrough
13413 case ISD::SETUGE: SSECC = 5; break;
13414 case ISD::SETULT: Swap = true; // Fallthrough
13415 case ISD::SETUGT: SSECC = 6; break;
13416 case ISD::SETO: SSECC = 7; break;
13418 case ISD::SETONE: SSECC = 8; break;
13421 std::swap(Op0, Op1);
13426 // Lower256IntVSETCC - Break a VSETCC 256-bit integer VSETCC into two new 128
13427 // ones, and then concatenate the result back.
13428 static SDValue Lower256IntVSETCC(SDValue Op, SelectionDAG &DAG) {
13429 MVT VT = Op.getSimpleValueType();
13431 assert(VT.is256BitVector() && Op.getOpcode() == ISD::SETCC &&
13432 "Unsupported value type for operation");
13434 unsigned NumElems = VT.getVectorNumElements();
13436 SDValue CC = Op.getOperand(2);
13438 // Extract the LHS vectors
13439 SDValue LHS = Op.getOperand(0);
13440 SDValue LHS1 = Extract128BitVector(LHS, 0, DAG, dl);
13441 SDValue LHS2 = Extract128BitVector(LHS, NumElems/2, DAG, dl);
13443 // Extract the RHS vectors
13444 SDValue RHS = Op.getOperand(1);
13445 SDValue RHS1 = Extract128BitVector(RHS, 0, DAG, dl);
13446 SDValue RHS2 = Extract128BitVector(RHS, NumElems/2, DAG, dl);
13448 // Issue the operation on the smaller types and concatenate the result back
13449 MVT EltVT = VT.getVectorElementType();
13450 MVT NewVT = MVT::getVectorVT(EltVT, NumElems/2);
13451 return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT,
13452 DAG.getNode(Op.getOpcode(), dl, NewVT, LHS1, RHS1, CC),
13453 DAG.getNode(Op.getOpcode(), dl, NewVT, LHS2, RHS2, CC));
13456 static SDValue LowerBoolVSETCC_AVX512(SDValue Op, SelectionDAG &DAG) {
13457 SDValue Op0 = Op.getOperand(0);
13458 SDValue Op1 = Op.getOperand(1);
13459 SDValue CC = Op.getOperand(2);
13460 MVT VT = Op.getSimpleValueType();
13463 assert(Op0.getValueType().getVectorElementType() == MVT::i1 &&
13464 "Unexpected type for boolean compare operation");
13465 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
13466 SDValue NotOp0 = DAG.getNode(ISD::XOR, dl, VT, Op0,
13467 DAG.getConstant(-1, dl, VT));
13468 SDValue NotOp1 = DAG.getNode(ISD::XOR, dl, VT, Op1,
13469 DAG.getConstant(-1, dl, VT));
13470 switch (SetCCOpcode) {
13471 default: llvm_unreachable("Unexpected SETCC condition");
13473 // (x == y) -> ~(x ^ y)
13474 return DAG.getNode(ISD::XOR, dl, VT,
13475 DAG.getNode(ISD::XOR, dl, VT, Op0, Op1),
13476 DAG.getConstant(-1, dl, VT));
13478 // (x != y) -> (x ^ y)
13479 return DAG.getNode(ISD::XOR, dl, VT, Op0, Op1);
13482 // (x > y) -> (x & ~y)
13483 return DAG.getNode(ISD::AND, dl, VT, Op0, NotOp1);
13486 // (x < y) -> (~x & y)
13487 return DAG.getNode(ISD::AND, dl, VT, NotOp0, Op1);
13490 // (x <= y) -> (~x | y)
13491 return DAG.getNode(ISD::OR, dl, VT, NotOp0, Op1);
13494 // (x >=y) -> (x | ~y)
13495 return DAG.getNode(ISD::OR, dl, VT, Op0, NotOp1);
13499 static SDValue LowerIntVSETCC_AVX512(SDValue Op, SelectionDAG &DAG,
13500 const X86Subtarget *Subtarget) {
13501 SDValue Op0 = Op.getOperand(0);
13502 SDValue Op1 = Op.getOperand(1);
13503 SDValue CC = Op.getOperand(2);
13504 MVT VT = Op.getSimpleValueType();
13507 assert(Op0.getValueType().getVectorElementType().getSizeInBits() >= 8 &&
13508 Op.getValueType().getScalarType() == MVT::i1 &&
13509 "Cannot set masked compare for this operation");
13511 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
13513 bool Unsigned = false;
13516 switch (SetCCOpcode) {
13517 default: llvm_unreachable("Unexpected SETCC condition");
13518 case ISD::SETNE: SSECC = 4; break;
13519 case ISD::SETEQ: Opc = X86ISD::PCMPEQM; break;
13520 case ISD::SETUGT: SSECC = 6; Unsigned = true; break;
13521 case ISD::SETLT: Swap = true; //fall-through
13522 case ISD::SETGT: Opc = X86ISD::PCMPGTM; break;
13523 case ISD::SETULT: SSECC = 1; Unsigned = true; break;
13524 case ISD::SETUGE: SSECC = 5; Unsigned = true; break; //NLT
13525 case ISD::SETGE: Swap = true; SSECC = 2; break; // LE + swap
13526 case ISD::SETULE: Unsigned = true; //fall-through
13527 case ISD::SETLE: SSECC = 2; break;
13531 std::swap(Op0, Op1);
13533 return DAG.getNode(Opc, dl, VT, Op0, Op1);
13534 Opc = Unsigned ? X86ISD::CMPMU: X86ISD::CMPM;
13535 return DAG.getNode(Opc, dl, VT, Op0, Op1,
13536 DAG.getConstant(SSECC, dl, MVT::i8));
13539 /// \brief Try to turn a VSETULT into a VSETULE by modifying its second
13540 /// operand \p Op1. If non-trivial (for example because it's not constant)
13541 /// return an empty value.
13542 static SDValue ChangeVSETULTtoVSETULE(SDLoc dl, SDValue Op1, SelectionDAG &DAG)
13544 BuildVectorSDNode *BV = dyn_cast<BuildVectorSDNode>(Op1.getNode());
13548 MVT VT = Op1.getSimpleValueType();
13549 MVT EVT = VT.getVectorElementType();
13550 unsigned n = VT.getVectorNumElements();
13551 SmallVector<SDValue, 8> ULTOp1;
13553 for (unsigned i = 0; i < n; ++i) {
13554 ConstantSDNode *Elt = dyn_cast<ConstantSDNode>(BV->getOperand(i));
13555 if (!Elt || Elt->isOpaque() || Elt->getValueType(0) != EVT)
13558 // Avoid underflow.
13559 APInt Val = Elt->getAPIntValue();
13563 ULTOp1.push_back(DAG.getConstant(Val - 1, dl, EVT));
13566 return DAG.getNode(ISD::BUILD_VECTOR, dl, VT, ULTOp1);
13569 static SDValue LowerVSETCC(SDValue Op, const X86Subtarget *Subtarget,
13570 SelectionDAG &DAG) {
13571 SDValue Op0 = Op.getOperand(0);
13572 SDValue Op1 = Op.getOperand(1);
13573 SDValue CC = Op.getOperand(2);
13574 MVT VT = Op.getSimpleValueType();
13575 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
13576 bool isFP = Op.getOperand(1).getSimpleValueType().isFloatingPoint();
13581 MVT EltVT = Op0.getSimpleValueType().getVectorElementType();
13582 assert(EltVT == MVT::f32 || EltVT == MVT::f64);
13585 unsigned SSECC = translateX86FSETCC(SetCCOpcode, Op0, Op1);
13586 unsigned Opc = X86ISD::CMPP;
13587 if (Subtarget->hasAVX512() && VT.getVectorElementType() == MVT::i1) {
13588 assert(VT.getVectorNumElements() <= 16);
13589 Opc = X86ISD::CMPM;
13591 // In the two special cases we can't handle, emit two comparisons.
13594 unsigned CombineOpc;
13595 if (SetCCOpcode == ISD::SETUEQ) {
13596 CC0 = 3; CC1 = 0; CombineOpc = ISD::OR;
13598 assert(SetCCOpcode == ISD::SETONE);
13599 CC0 = 7; CC1 = 4; CombineOpc = ISD::AND;
13602 SDValue Cmp0 = DAG.getNode(Opc, dl, VT, Op0, Op1,
13603 DAG.getConstant(CC0, dl, MVT::i8));
13604 SDValue Cmp1 = DAG.getNode(Opc, dl, VT, Op0, Op1,
13605 DAG.getConstant(CC1, dl, MVT::i8));
13606 return DAG.getNode(CombineOpc, dl, VT, Cmp0, Cmp1);
13608 // Handle all other FP comparisons here.
13609 return DAG.getNode(Opc, dl, VT, Op0, Op1,
13610 DAG.getConstant(SSECC, dl, MVT::i8));
13613 // Break 256-bit integer vector compare into smaller ones.
13614 if (VT.is256BitVector() && !Subtarget->hasInt256())
13615 return Lower256IntVSETCC(Op, DAG);
13617 EVT OpVT = Op1.getValueType();
13618 if (OpVT.getVectorElementType() == MVT::i1)
13619 return LowerBoolVSETCC_AVX512(Op, DAG);
13621 bool MaskResult = (VT.getVectorElementType() == MVT::i1);
13622 if (Subtarget->hasAVX512()) {
13623 if (Op1.getValueType().is512BitVector() ||
13624 (Subtarget->hasBWI() && Subtarget->hasVLX()) ||
13625 (MaskResult && OpVT.getVectorElementType().getSizeInBits() >= 32))
13626 return LowerIntVSETCC_AVX512(Op, DAG, Subtarget);
13628 // In AVX-512 architecture setcc returns mask with i1 elements,
13629 // But there is no compare instruction for i8 and i16 elements in KNL.
13630 // We are not talking about 512-bit operands in this case, these
13631 // types are illegal.
13633 (OpVT.getVectorElementType().getSizeInBits() < 32 &&
13634 OpVT.getVectorElementType().getSizeInBits() >= 8))
13635 return DAG.getNode(ISD::TRUNCATE, dl, VT,
13636 DAG.getNode(ISD::SETCC, dl, OpVT, Op0, Op1, CC));
13639 // We are handling one of the integer comparisons here. Since SSE only has
13640 // GT and EQ comparisons for integer, swapping operands and multiple
13641 // operations may be required for some comparisons.
13643 bool Swap = false, Invert = false, FlipSigns = false, MinMax = false;
13644 bool Subus = false;
13646 switch (SetCCOpcode) {
13647 default: llvm_unreachable("Unexpected SETCC condition");
13648 case ISD::SETNE: Invert = true;
13649 case ISD::SETEQ: Opc = X86ISD::PCMPEQ; break;
13650 case ISD::SETLT: Swap = true;
13651 case ISD::SETGT: Opc = X86ISD::PCMPGT; break;
13652 case ISD::SETGE: Swap = true;
13653 case ISD::SETLE: Opc = X86ISD::PCMPGT;
13654 Invert = true; break;
13655 case ISD::SETULT: Swap = true;
13656 case ISD::SETUGT: Opc = X86ISD::PCMPGT;
13657 FlipSigns = true; break;
13658 case ISD::SETUGE: Swap = true;
13659 case ISD::SETULE: Opc = X86ISD::PCMPGT;
13660 FlipSigns = true; Invert = true; break;
13663 // Special case: Use min/max operations for SETULE/SETUGE
13664 MVT VET = VT.getVectorElementType();
13666 (Subtarget->hasSSE41() && (VET >= MVT::i8 && VET <= MVT::i32))
13667 || (Subtarget->hasSSE2() && (VET == MVT::i8));
13670 switch (SetCCOpcode) {
13672 case ISD::SETULE: Opc = ISD::UMIN; MinMax = true; break;
13673 case ISD::SETUGE: Opc = ISD::UMAX; MinMax = true; break;
13676 if (MinMax) { Swap = false; Invert = false; FlipSigns = false; }
13679 bool hasSubus = Subtarget->hasSSE2() && (VET == MVT::i8 || VET == MVT::i16);
13680 if (!MinMax && hasSubus) {
13681 // As another special case, use PSUBUS[BW] when it's profitable. E.g. for
13683 // t = psubus Op0, Op1
13684 // pcmpeq t, <0..0>
13685 switch (SetCCOpcode) {
13687 case ISD::SETULT: {
13688 // If the comparison is against a constant we can turn this into a
13689 // setule. With psubus, setule does not require a swap. This is
13690 // beneficial because the constant in the register is no longer
13691 // destructed as the destination so it can be hoisted out of a loop.
13692 // Only do this pre-AVX since vpcmp* is no longer destructive.
13693 if (Subtarget->hasAVX())
13695 SDValue ULEOp1 = ChangeVSETULTtoVSETULE(dl, Op1, DAG);
13696 if (ULEOp1.getNode()) {
13698 Subus = true; Invert = false; Swap = false;
13702 // Psubus is better than flip-sign because it requires no inversion.
13703 case ISD::SETUGE: Subus = true; Invert = false; Swap = true; break;
13704 case ISD::SETULE: Subus = true; Invert = false; Swap = false; break;
13708 Opc = X86ISD::SUBUS;
13714 std::swap(Op0, Op1);
13716 // Check that the operation in question is available (most are plain SSE2,
13717 // but PCMPGTQ and PCMPEQQ have different requirements).
13718 if (VT == MVT::v2i64) {
13719 if (Opc == X86ISD::PCMPGT && !Subtarget->hasSSE42()) {
13720 assert(Subtarget->hasSSE2() && "Don't know how to lower!");
13722 // First cast everything to the right type.
13723 Op0 = DAG.getBitcast(MVT::v4i32, Op0);
13724 Op1 = DAG.getBitcast(MVT::v4i32, Op1);
13726 // Since SSE has no unsigned integer comparisons, we need to flip the sign
13727 // bits of the inputs before performing those operations. The lower
13728 // compare is always unsigned.
13731 SB = DAG.getConstant(0x80000000U, dl, MVT::v4i32);
13733 SDValue Sign = DAG.getConstant(0x80000000U, dl, MVT::i32);
13734 SDValue Zero = DAG.getConstant(0x00000000U, dl, MVT::i32);
13735 SB = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32,
13736 Sign, Zero, Sign, Zero);
13738 Op0 = DAG.getNode(ISD::XOR, dl, MVT::v4i32, Op0, SB);
13739 Op1 = DAG.getNode(ISD::XOR, dl, MVT::v4i32, Op1, SB);
13741 // Emulate PCMPGTQ with (hi1 > hi2) | ((hi1 == hi2) & (lo1 > lo2))
13742 SDValue GT = DAG.getNode(X86ISD::PCMPGT, dl, MVT::v4i32, Op0, Op1);
13743 SDValue EQ = DAG.getNode(X86ISD::PCMPEQ, dl, MVT::v4i32, Op0, Op1);
13745 // Create masks for only the low parts/high parts of the 64 bit integers.
13746 static const int MaskHi[] = { 1, 1, 3, 3 };
13747 static const int MaskLo[] = { 0, 0, 2, 2 };
13748 SDValue EQHi = DAG.getVectorShuffle(MVT::v4i32, dl, EQ, EQ, MaskHi);
13749 SDValue GTLo = DAG.getVectorShuffle(MVT::v4i32, dl, GT, GT, MaskLo);
13750 SDValue GTHi = DAG.getVectorShuffle(MVT::v4i32, dl, GT, GT, MaskHi);
13752 SDValue Result = DAG.getNode(ISD::AND, dl, MVT::v4i32, EQHi, GTLo);
13753 Result = DAG.getNode(ISD::OR, dl, MVT::v4i32, Result, GTHi);
13756 Result = DAG.getNOT(dl, Result, MVT::v4i32);
13758 return DAG.getBitcast(VT, Result);
13761 if (Opc == X86ISD::PCMPEQ && !Subtarget->hasSSE41()) {
13762 // If pcmpeqq is missing but pcmpeqd is available synthesize pcmpeqq with
13763 // pcmpeqd + pshufd + pand.
13764 assert(Subtarget->hasSSE2() && !FlipSigns && "Don't know how to lower!");
13766 // First cast everything to the right type.
13767 Op0 = DAG.getBitcast(MVT::v4i32, Op0);
13768 Op1 = DAG.getBitcast(MVT::v4i32, Op1);
13771 SDValue Result = DAG.getNode(Opc, dl, MVT::v4i32, Op0, Op1);
13773 // Make sure the lower and upper halves are both all-ones.
13774 static const int Mask[] = { 1, 0, 3, 2 };
13775 SDValue Shuf = DAG.getVectorShuffle(MVT::v4i32, dl, Result, Result, Mask);
13776 Result = DAG.getNode(ISD::AND, dl, MVT::v4i32, Result, Shuf);
13779 Result = DAG.getNOT(dl, Result, MVT::v4i32);
13781 return DAG.getBitcast(VT, Result);
13785 // Since SSE has no unsigned integer comparisons, we need to flip the sign
13786 // bits of the inputs before performing those operations.
13788 EVT EltVT = VT.getVectorElementType();
13789 SDValue SB = DAG.getConstant(APInt::getSignBit(EltVT.getSizeInBits()), dl,
13791 Op0 = DAG.getNode(ISD::XOR, dl, VT, Op0, SB);
13792 Op1 = DAG.getNode(ISD::XOR, dl, VT, Op1, SB);
13795 SDValue Result = DAG.getNode(Opc, dl, VT, Op0, Op1);
13797 // If the logical-not of the result is required, perform that now.
13799 Result = DAG.getNOT(dl, Result, VT);
13802 Result = DAG.getNode(X86ISD::PCMPEQ, dl, VT, Op0, Result);
13805 Result = DAG.getNode(X86ISD::PCMPEQ, dl, VT, Result,
13806 getZeroVector(VT, Subtarget, DAG, dl));
13811 SDValue X86TargetLowering::LowerSETCC(SDValue Op, SelectionDAG &DAG) const {
13813 MVT VT = Op.getSimpleValueType();
13815 if (VT.isVector()) return LowerVSETCC(Op, Subtarget, DAG);
13817 assert(((!Subtarget->hasAVX512() && VT == MVT::i8) || (VT == MVT::i1))
13818 && "SetCC type must be 8-bit or 1-bit integer");
13819 SDValue Op0 = Op.getOperand(0);
13820 SDValue Op1 = Op.getOperand(1);
13822 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
13824 // Optimize to BT if possible.
13825 // Lower (X & (1 << N)) == 0 to BT(X, N).
13826 // Lower ((X >>u N) & 1) != 0 to BT(X, N).
13827 // Lower ((X >>s N) & 1) != 0 to BT(X, N).
13828 if (Op0.getOpcode() == ISD::AND && Op0.hasOneUse() &&
13829 Op1.getOpcode() == ISD::Constant &&
13830 cast<ConstantSDNode>(Op1)->isNullValue() &&
13831 (CC == ISD::SETEQ || CC == ISD::SETNE)) {
13832 SDValue NewSetCC = LowerToBT(Op0, CC, dl, DAG);
13833 if (NewSetCC.getNode()) {
13835 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i1, NewSetCC);
13840 // Look for X == 0, X == 1, X != 0, or X != 1. We can simplify some forms of
13842 if (Op1.getOpcode() == ISD::Constant &&
13843 (cast<ConstantSDNode>(Op1)->getZExtValue() == 1 ||
13844 cast<ConstantSDNode>(Op1)->isNullValue()) &&
13845 (CC == ISD::SETEQ || CC == ISD::SETNE)) {
13847 // If the input is a setcc, then reuse the input setcc or use a new one with
13848 // the inverted condition.
13849 if (Op0.getOpcode() == X86ISD::SETCC) {
13850 X86::CondCode CCode = (X86::CondCode)Op0.getConstantOperandVal(0);
13851 bool Invert = (CC == ISD::SETNE) ^
13852 cast<ConstantSDNode>(Op1)->isNullValue();
13856 CCode = X86::GetOppositeBranchCondition(CCode);
13857 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
13858 DAG.getConstant(CCode, dl, MVT::i8),
13859 Op0.getOperand(1));
13861 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i1, SetCC);
13865 if ((Op0.getValueType() == MVT::i1) && (Op1.getOpcode() == ISD::Constant) &&
13866 (cast<ConstantSDNode>(Op1)->getZExtValue() == 1) &&
13867 (CC == ISD::SETEQ || CC == ISD::SETNE)) {
13869 ISD::CondCode NewCC = ISD::getSetCCInverse(CC, true);
13870 return DAG.getSetCC(dl, VT, Op0, DAG.getConstant(0, dl, MVT::i1), NewCC);
13873 bool isFP = Op1.getSimpleValueType().isFloatingPoint();
13874 unsigned X86CC = TranslateX86CC(CC, dl, isFP, Op0, Op1, DAG);
13875 if (X86CC == X86::COND_INVALID)
13878 SDValue EFLAGS = EmitCmp(Op0, Op1, X86CC, dl, DAG);
13879 EFLAGS = ConvertCmpIfNecessary(EFLAGS, DAG);
13880 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
13881 DAG.getConstant(X86CC, dl, MVT::i8), EFLAGS);
13883 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i1, SetCC);
13887 // isX86LogicalCmp - Return true if opcode is a X86 logical comparison.
13888 static bool isX86LogicalCmp(SDValue Op) {
13889 unsigned Opc = Op.getNode()->getOpcode();
13890 if (Opc == X86ISD::CMP || Opc == X86ISD::COMI || Opc == X86ISD::UCOMI ||
13891 Opc == X86ISD::SAHF)
13893 if (Op.getResNo() == 1 &&
13894 (Opc == X86ISD::ADD ||
13895 Opc == X86ISD::SUB ||
13896 Opc == X86ISD::ADC ||
13897 Opc == X86ISD::SBB ||
13898 Opc == X86ISD::SMUL ||
13899 Opc == X86ISD::UMUL ||
13900 Opc == X86ISD::INC ||
13901 Opc == X86ISD::DEC ||
13902 Opc == X86ISD::OR ||
13903 Opc == X86ISD::XOR ||
13904 Opc == X86ISD::AND))
13907 if (Op.getResNo() == 2 && Opc == X86ISD::UMUL)
13913 static bool isTruncWithZeroHighBitsInput(SDValue V, SelectionDAG &DAG) {
13914 if (V.getOpcode() != ISD::TRUNCATE)
13917 SDValue VOp0 = V.getOperand(0);
13918 unsigned InBits = VOp0.getValueSizeInBits();
13919 unsigned Bits = V.getValueSizeInBits();
13920 return DAG.MaskedValueIsZero(VOp0, APInt::getHighBitsSet(InBits,InBits-Bits));
13923 SDValue X86TargetLowering::LowerSELECT(SDValue Op, SelectionDAG &DAG) const {
13924 bool addTest = true;
13925 SDValue Cond = Op.getOperand(0);
13926 SDValue Op1 = Op.getOperand(1);
13927 SDValue Op2 = Op.getOperand(2);
13929 EVT VT = Op1.getValueType();
13932 // Lower FP selects into a CMP/AND/ANDN/OR sequence when the necessary SSE ops
13933 // are available or VBLENDV if AVX is available.
13934 // Otherwise FP cmovs get lowered into a less efficient branch sequence later.
13935 if (Cond.getOpcode() == ISD::SETCC &&
13936 ((Subtarget->hasSSE2() && (VT == MVT::f32 || VT == MVT::f64)) ||
13937 (Subtarget->hasSSE1() && VT == MVT::f32)) &&
13938 VT == Cond.getOperand(0).getValueType() && Cond->hasOneUse()) {
13939 SDValue CondOp0 = Cond.getOperand(0), CondOp1 = Cond.getOperand(1);
13940 int SSECC = translateX86FSETCC(
13941 cast<CondCodeSDNode>(Cond.getOperand(2))->get(), CondOp0, CondOp1);
13944 if (Subtarget->hasAVX512()) {
13945 SDValue Cmp = DAG.getNode(X86ISD::FSETCC, DL, MVT::i1, CondOp0, CondOp1,
13946 DAG.getConstant(SSECC, DL, MVT::i8));
13947 return DAG.getNode(X86ISD::SELECT, DL, VT, Cmp, Op1, Op2);
13950 SDValue Cmp = DAG.getNode(X86ISD::FSETCC, DL, VT, CondOp0, CondOp1,
13951 DAG.getConstant(SSECC, DL, MVT::i8));
13953 // If we have AVX, we can use a variable vector select (VBLENDV) instead
13954 // of 3 logic instructions for size savings and potentially speed.
13955 // Unfortunately, there is no scalar form of VBLENDV.
13957 // If either operand is a constant, don't try this. We can expect to
13958 // optimize away at least one of the logic instructions later in that
13959 // case, so that sequence would be faster than a variable blend.
13961 // BLENDV was introduced with SSE 4.1, but the 2 register form implicitly
13962 // uses XMM0 as the selection register. That may need just as many
13963 // instructions as the AND/ANDN/OR sequence due to register moves, so
13966 if (Subtarget->hasAVX() &&
13967 !isa<ConstantFPSDNode>(Op1) && !isa<ConstantFPSDNode>(Op2)) {
13969 // Convert to vectors, do a VSELECT, and convert back to scalar.
13970 // All of the conversions should be optimized away.
13972 EVT VecVT = VT == MVT::f32 ? MVT::v4f32 : MVT::v2f64;
13973 SDValue VOp1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, VecVT, Op1);
13974 SDValue VOp2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, VecVT, Op2);
13975 SDValue VCmp = DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, VecVT, Cmp);
13977 EVT VCmpVT = VT == MVT::f32 ? MVT::v4i32 : MVT::v2i64;
13978 VCmp = DAG.getBitcast(VCmpVT, VCmp);
13980 SDValue VSel = DAG.getNode(ISD::VSELECT, DL, VecVT, VCmp, VOp1, VOp2);
13982 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, VT,
13983 VSel, DAG.getIntPtrConstant(0, DL));
13985 SDValue AndN = DAG.getNode(X86ISD::FANDN, DL, VT, Cmp, Op2);
13986 SDValue And = DAG.getNode(X86ISD::FAND, DL, VT, Cmp, Op1);
13987 return DAG.getNode(X86ISD::FOR, DL, VT, AndN, And);
13991 if (VT.isVector() && VT.getScalarType() == MVT::i1) {
13993 if (ISD::isBuildVectorOfConstantSDNodes(Op1.getNode()))
13994 Op1Scalar = ConvertI1VectorToInteger(Op1, DAG);
13995 else if (Op1.getOpcode() == ISD::BITCAST && Op1.getOperand(0))
13996 Op1Scalar = Op1.getOperand(0);
13998 if (ISD::isBuildVectorOfConstantSDNodes(Op2.getNode()))
13999 Op2Scalar = ConvertI1VectorToInteger(Op2, DAG);
14000 else if (Op2.getOpcode() == ISD::BITCAST && Op2.getOperand(0))
14001 Op2Scalar = Op2.getOperand(0);
14002 if (Op1Scalar.getNode() && Op2Scalar.getNode()) {
14003 SDValue newSelect = DAG.getNode(ISD::SELECT, DL,
14004 Op1Scalar.getValueType(),
14005 Cond, Op1Scalar, Op2Scalar);
14006 if (newSelect.getValueSizeInBits() == VT.getSizeInBits())
14007 return DAG.getBitcast(VT, newSelect);
14008 SDValue ExtVec = DAG.getBitcast(MVT::v8i1, newSelect);
14009 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, VT, ExtVec,
14010 DAG.getIntPtrConstant(0, DL));
14014 if (VT == MVT::v4i1 || VT == MVT::v2i1) {
14015 SDValue zeroConst = DAG.getIntPtrConstant(0, DL);
14016 Op1 = DAG.getNode(ISD::INSERT_SUBVECTOR, DL, MVT::v8i1,
14017 DAG.getUNDEF(MVT::v8i1), Op1, zeroConst);
14018 Op2 = DAG.getNode(ISD::INSERT_SUBVECTOR, DL, MVT::v8i1,
14019 DAG.getUNDEF(MVT::v8i1), Op2, zeroConst);
14020 SDValue newSelect = DAG.getNode(ISD::SELECT, DL, MVT::v8i1,
14022 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, VT, newSelect, zeroConst);
14025 if (Cond.getOpcode() == ISD::SETCC) {
14026 SDValue NewCond = LowerSETCC(Cond, DAG);
14027 if (NewCond.getNode())
14031 // (select (x == 0), -1, y) -> (sign_bit (x - 1)) | y
14032 // (select (x == 0), y, -1) -> ~(sign_bit (x - 1)) | y
14033 // (select (x != 0), y, -1) -> (sign_bit (x - 1)) | y
14034 // (select (x != 0), -1, y) -> ~(sign_bit (x - 1)) | y
14035 if (Cond.getOpcode() == X86ISD::SETCC &&
14036 Cond.getOperand(1).getOpcode() == X86ISD::CMP &&
14037 isZero(Cond.getOperand(1).getOperand(1))) {
14038 SDValue Cmp = Cond.getOperand(1);
14040 unsigned CondCode =cast<ConstantSDNode>(Cond.getOperand(0))->getZExtValue();
14042 if ((isAllOnes(Op1) || isAllOnes(Op2)) &&
14043 (CondCode == X86::COND_E || CondCode == X86::COND_NE)) {
14044 SDValue Y = isAllOnes(Op2) ? Op1 : Op2;
14046 SDValue CmpOp0 = Cmp.getOperand(0);
14047 // Apply further optimizations for special cases
14048 // (select (x != 0), -1, 0) -> neg & sbb
14049 // (select (x == 0), 0, -1) -> neg & sbb
14050 if (ConstantSDNode *YC = dyn_cast<ConstantSDNode>(Y))
14051 if (YC->isNullValue() &&
14052 (isAllOnes(Op1) == (CondCode == X86::COND_NE))) {
14053 SDVTList VTs = DAG.getVTList(CmpOp0.getValueType(), MVT::i32);
14054 SDValue Neg = DAG.getNode(X86ISD::SUB, DL, VTs,
14055 DAG.getConstant(0, DL,
14056 CmpOp0.getValueType()),
14058 SDValue Res = DAG.getNode(X86ISD::SETCC_CARRY, DL, Op.getValueType(),
14059 DAG.getConstant(X86::COND_B, DL, MVT::i8),
14060 SDValue(Neg.getNode(), 1));
14064 Cmp = DAG.getNode(X86ISD::CMP, DL, MVT::i32,
14065 CmpOp0, DAG.getConstant(1, DL, CmpOp0.getValueType()));
14066 Cmp = ConvertCmpIfNecessary(Cmp, DAG);
14068 SDValue Res = // Res = 0 or -1.
14069 DAG.getNode(X86ISD::SETCC_CARRY, DL, Op.getValueType(),
14070 DAG.getConstant(X86::COND_B, DL, MVT::i8), Cmp);
14072 if (isAllOnes(Op1) != (CondCode == X86::COND_E))
14073 Res = DAG.getNOT(DL, Res, Res.getValueType());
14075 ConstantSDNode *N2C = dyn_cast<ConstantSDNode>(Op2);
14076 if (!N2C || !N2C->isNullValue())
14077 Res = DAG.getNode(ISD::OR, DL, Res.getValueType(), Res, Y);
14082 // Look past (and (setcc_carry (cmp ...)), 1).
14083 if (Cond.getOpcode() == ISD::AND &&
14084 Cond.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY) {
14085 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Cond.getOperand(1));
14086 if (C && C->getAPIntValue() == 1)
14087 Cond = Cond.getOperand(0);
14090 // If condition flag is set by a X86ISD::CMP, then use it as the condition
14091 // setting operand in place of the X86ISD::SETCC.
14092 unsigned CondOpcode = Cond.getOpcode();
14093 if (CondOpcode == X86ISD::SETCC ||
14094 CondOpcode == X86ISD::SETCC_CARRY) {
14095 CC = Cond.getOperand(0);
14097 SDValue Cmp = Cond.getOperand(1);
14098 unsigned Opc = Cmp.getOpcode();
14099 MVT VT = Op.getSimpleValueType();
14101 bool IllegalFPCMov = false;
14102 if (VT.isFloatingPoint() && !VT.isVector() &&
14103 !isScalarFPTypeInSSEReg(VT)) // FPStack?
14104 IllegalFPCMov = !hasFPCMov(cast<ConstantSDNode>(CC)->getSExtValue());
14106 if ((isX86LogicalCmp(Cmp) && !IllegalFPCMov) ||
14107 Opc == X86ISD::BT) { // FIXME
14111 } else if (CondOpcode == ISD::USUBO || CondOpcode == ISD::SSUBO ||
14112 CondOpcode == ISD::UADDO || CondOpcode == ISD::SADDO ||
14113 ((CondOpcode == ISD::UMULO || CondOpcode == ISD::SMULO) &&
14114 Cond.getOperand(0).getValueType() != MVT::i8)) {
14115 SDValue LHS = Cond.getOperand(0);
14116 SDValue RHS = Cond.getOperand(1);
14117 unsigned X86Opcode;
14120 switch (CondOpcode) {
14121 case ISD::UADDO: X86Opcode = X86ISD::ADD; X86Cond = X86::COND_B; break;
14122 case ISD::SADDO: X86Opcode = X86ISD::ADD; X86Cond = X86::COND_O; break;
14123 case ISD::USUBO: X86Opcode = X86ISD::SUB; X86Cond = X86::COND_B; break;
14124 case ISD::SSUBO: X86Opcode = X86ISD::SUB; X86Cond = X86::COND_O; break;
14125 case ISD::UMULO: X86Opcode = X86ISD::UMUL; X86Cond = X86::COND_O; break;
14126 case ISD::SMULO: X86Opcode = X86ISD::SMUL; X86Cond = X86::COND_O; break;
14127 default: llvm_unreachable("unexpected overflowing operator");
14129 if (CondOpcode == ISD::UMULO)
14130 VTs = DAG.getVTList(LHS.getValueType(), LHS.getValueType(),
14133 VTs = DAG.getVTList(LHS.getValueType(), MVT::i32);
14135 SDValue X86Op = DAG.getNode(X86Opcode, DL, VTs, LHS, RHS);
14137 if (CondOpcode == ISD::UMULO)
14138 Cond = X86Op.getValue(2);
14140 Cond = X86Op.getValue(1);
14142 CC = DAG.getConstant(X86Cond, DL, MVT::i8);
14147 // Look past the truncate if the high bits are known zero.
14148 if (isTruncWithZeroHighBitsInput(Cond, DAG))
14149 Cond = Cond.getOperand(0);
14151 // We know the result of AND is compared against zero. Try to match
14153 if (Cond.getOpcode() == ISD::AND && Cond.hasOneUse()) {
14154 SDValue NewSetCC = LowerToBT(Cond, ISD::SETNE, DL, DAG);
14155 if (NewSetCC.getNode()) {
14156 CC = NewSetCC.getOperand(0);
14157 Cond = NewSetCC.getOperand(1);
14164 CC = DAG.getConstant(X86::COND_NE, DL, MVT::i8);
14165 Cond = EmitTest(Cond, X86::COND_NE, DL, DAG);
14168 // a < b ? -1 : 0 -> RES = ~setcc_carry
14169 // a < b ? 0 : -1 -> RES = setcc_carry
14170 // a >= b ? -1 : 0 -> RES = setcc_carry
14171 // a >= b ? 0 : -1 -> RES = ~setcc_carry
14172 if (Cond.getOpcode() == X86ISD::SUB) {
14173 Cond = ConvertCmpIfNecessary(Cond, DAG);
14174 unsigned CondCode = cast<ConstantSDNode>(CC)->getZExtValue();
14176 if ((CondCode == X86::COND_AE || CondCode == X86::COND_B) &&
14177 (isAllOnes(Op1) || isAllOnes(Op2)) && (isZero(Op1) || isZero(Op2))) {
14178 SDValue Res = DAG.getNode(X86ISD::SETCC_CARRY, DL, Op.getValueType(),
14179 DAG.getConstant(X86::COND_B, DL, MVT::i8),
14181 if (isAllOnes(Op1) != (CondCode == X86::COND_B))
14182 return DAG.getNOT(DL, Res, Res.getValueType());
14187 // X86 doesn't have an i8 cmov. If both operands are the result of a truncate
14188 // widen the cmov and push the truncate through. This avoids introducing a new
14189 // branch during isel and doesn't add any extensions.
14190 if (Op.getValueType() == MVT::i8 &&
14191 Op1.getOpcode() == ISD::TRUNCATE && Op2.getOpcode() == ISD::TRUNCATE) {
14192 SDValue T1 = Op1.getOperand(0), T2 = Op2.getOperand(0);
14193 if (T1.getValueType() == T2.getValueType() &&
14194 // Blacklist CopyFromReg to avoid partial register stalls.
14195 T1.getOpcode() != ISD::CopyFromReg && T2.getOpcode()!=ISD::CopyFromReg){
14196 SDVTList VTs = DAG.getVTList(T1.getValueType(), MVT::Glue);
14197 SDValue Cmov = DAG.getNode(X86ISD::CMOV, DL, VTs, T2, T1, CC, Cond);
14198 return DAG.getNode(ISD::TRUNCATE, DL, Op.getValueType(), Cmov);
14202 // X86ISD::CMOV means set the result (which is operand 1) to the RHS if
14203 // condition is true.
14204 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::Glue);
14205 SDValue Ops[] = { Op2, Op1, CC, Cond };
14206 return DAG.getNode(X86ISD::CMOV, DL, VTs, Ops);
14209 static SDValue LowerSIGN_EXTEND_AVX512(SDValue Op,
14210 const X86Subtarget *Subtarget,
14211 SelectionDAG &DAG) {
14212 MVT VT = Op->getSimpleValueType(0);
14213 SDValue In = Op->getOperand(0);
14214 MVT InVT = In.getSimpleValueType();
14215 MVT VTElt = VT.getVectorElementType();
14216 MVT InVTElt = InVT.getVectorElementType();
14220 if ((InVTElt == MVT::i1) &&
14221 (((Subtarget->hasBWI() && Subtarget->hasVLX() &&
14222 VT.getSizeInBits() <= 256 && VTElt.getSizeInBits() <= 16)) ||
14224 ((Subtarget->hasBWI() && VT.is512BitVector() &&
14225 VTElt.getSizeInBits() <= 16)) ||
14227 ((Subtarget->hasDQI() && Subtarget->hasVLX() &&
14228 VT.getSizeInBits() <= 256 && VTElt.getSizeInBits() >= 32)) ||
14230 ((Subtarget->hasDQI() && VT.is512BitVector() &&
14231 VTElt.getSizeInBits() >= 32))))
14232 return DAG.getNode(X86ISD::VSEXT, dl, VT, In);
14234 unsigned int NumElts = VT.getVectorNumElements();
14236 if (NumElts != 8 && NumElts != 16 && !Subtarget->hasBWI())
14239 if (VT.is512BitVector() && InVT.getVectorElementType() != MVT::i1) {
14240 if (In.getOpcode() == X86ISD::VSEXT || In.getOpcode() == X86ISD::VZEXT)
14241 return DAG.getNode(In.getOpcode(), dl, VT, In.getOperand(0));
14242 return DAG.getNode(X86ISD::VSEXT, dl, VT, In);
14245 assert (InVT.getVectorElementType() == MVT::i1 && "Unexpected vector type");
14246 MVT ExtVT = NumElts == 8 ? MVT::v8i64 : MVT::v16i32;
14248 DAG.getConstant(APInt::getAllOnesValue(ExtVT.getScalarSizeInBits()), dl,
14251 DAG.getConstant(APInt::getNullValue(ExtVT.getScalarSizeInBits()), dl, ExtVT);
14253 SDValue V = DAG.getNode(ISD::VSELECT, dl, ExtVT, In, NegOne, Zero);
14254 if (VT.is512BitVector())
14256 return DAG.getNode(X86ISD::VTRUNC, dl, VT, V);
14259 static SDValue LowerSIGN_EXTEND_VECTOR_INREG(SDValue Op,
14260 const X86Subtarget *Subtarget,
14261 SelectionDAG &DAG) {
14262 SDValue In = Op->getOperand(0);
14263 MVT VT = Op->getSimpleValueType(0);
14264 MVT InVT = In.getSimpleValueType();
14265 assert(VT.getSizeInBits() == InVT.getSizeInBits());
14267 MVT InSVT = InVT.getScalarType();
14268 assert(VT.getScalarType().getScalarSizeInBits() > InSVT.getScalarSizeInBits());
14270 if (VT != MVT::v2i64 && VT != MVT::v4i32 && VT != MVT::v8i16)
14272 if (InSVT != MVT::i32 && InSVT != MVT::i16 && InSVT != MVT::i8)
14277 // SSE41 targets can use the pmovsx* instructions directly.
14278 if (Subtarget->hasSSE41())
14279 return DAG.getNode(X86ISD::VSEXT, dl, VT, In);
14281 // pre-SSE41 targets unpack lower lanes and then sign-extend using SRAI.
14285 // As SRAI is only available on i16/i32 types, we expand only up to i32
14286 // and handle i64 separately.
14287 while (CurrVT != VT && CurrVT.getScalarType() != MVT::i32) {
14288 Curr = DAG.getNode(X86ISD::UNPCKL, dl, CurrVT, DAG.getUNDEF(CurrVT), Curr);
14289 MVT CurrSVT = MVT::getIntegerVT(CurrVT.getScalarSizeInBits() * 2);
14290 CurrVT = MVT::getVectorVT(CurrSVT, CurrVT.getVectorNumElements() / 2);
14291 Curr = DAG.getBitcast(CurrVT, Curr);
14294 SDValue SignExt = Curr;
14295 if (CurrVT != InVT) {
14296 unsigned SignExtShift =
14297 CurrVT.getScalarSizeInBits() - InSVT.getScalarSizeInBits();
14298 SignExt = DAG.getNode(X86ISD::VSRAI, dl, CurrVT, Curr,
14299 DAG.getConstant(SignExtShift, dl, MVT::i8));
14305 if (VT == MVT::v2i64 && CurrVT == MVT::v4i32) {
14306 SDValue Sign = DAG.getNode(X86ISD::VSRAI, dl, CurrVT, Curr,
14307 DAG.getConstant(31, dl, MVT::i8));
14308 SDValue Ext = DAG.getVectorShuffle(CurrVT, dl, SignExt, Sign, {0, 4, 1, 5});
14309 return DAG.getBitcast(VT, Ext);
14315 static SDValue LowerSIGN_EXTEND(SDValue Op, const X86Subtarget *Subtarget,
14316 SelectionDAG &DAG) {
14317 MVT VT = Op->getSimpleValueType(0);
14318 SDValue In = Op->getOperand(0);
14319 MVT InVT = In.getSimpleValueType();
14322 if (VT.is512BitVector() || InVT.getVectorElementType() == MVT::i1)
14323 return LowerSIGN_EXTEND_AVX512(Op, Subtarget, DAG);
14325 if ((VT != MVT::v4i64 || InVT != MVT::v4i32) &&
14326 (VT != MVT::v8i32 || InVT != MVT::v8i16) &&
14327 (VT != MVT::v16i16 || InVT != MVT::v16i8))
14330 if (Subtarget->hasInt256())
14331 return DAG.getNode(X86ISD::VSEXT, dl, VT, In);
14333 // Optimize vectors in AVX mode
14334 // Sign extend v8i16 to v8i32 and
14337 // Divide input vector into two parts
14338 // for v4i32 the shuffle mask will be { 0, 1, -1, -1} {2, 3, -1, -1}
14339 // use vpmovsx instruction to extend v4i32 -> v2i64; v8i16 -> v4i32
14340 // concat the vectors to original VT
14342 unsigned NumElems = InVT.getVectorNumElements();
14343 SDValue Undef = DAG.getUNDEF(InVT);
14345 SmallVector<int,8> ShufMask1(NumElems, -1);
14346 for (unsigned i = 0; i != NumElems/2; ++i)
14349 SDValue OpLo = DAG.getVectorShuffle(InVT, dl, In, Undef, &ShufMask1[0]);
14351 SmallVector<int,8> ShufMask2(NumElems, -1);
14352 for (unsigned i = 0; i != NumElems/2; ++i)
14353 ShufMask2[i] = i + NumElems/2;
14355 SDValue OpHi = DAG.getVectorShuffle(InVT, dl, In, Undef, &ShufMask2[0]);
14357 MVT HalfVT = MVT::getVectorVT(VT.getScalarType(),
14358 VT.getVectorNumElements()/2);
14360 OpLo = DAG.getNode(X86ISD::VSEXT, dl, HalfVT, OpLo);
14361 OpHi = DAG.getNode(X86ISD::VSEXT, dl, HalfVT, OpHi);
14363 return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT, OpLo, OpHi);
14366 // Lower vector extended loads using a shuffle. If SSSE3 is not available we
14367 // may emit an illegal shuffle but the expansion is still better than scalar
14368 // code. We generate X86ISD::VSEXT for SEXTLOADs if it's available, otherwise
14369 // we'll emit a shuffle and a arithmetic shift.
14370 // FIXME: Is the expansion actually better than scalar code? It doesn't seem so.
14371 // TODO: It is possible to support ZExt by zeroing the undef values during
14372 // the shuffle phase or after the shuffle.
14373 static SDValue LowerExtendedLoad(SDValue Op, const X86Subtarget *Subtarget,
14374 SelectionDAG &DAG) {
14375 MVT RegVT = Op.getSimpleValueType();
14376 assert(RegVT.isVector() && "We only custom lower vector sext loads.");
14377 assert(RegVT.isInteger() &&
14378 "We only custom lower integer vector sext loads.");
14380 // Nothing useful we can do without SSE2 shuffles.
14381 assert(Subtarget->hasSSE2() && "We only custom lower sext loads with SSE2.");
14383 LoadSDNode *Ld = cast<LoadSDNode>(Op.getNode());
14385 EVT MemVT = Ld->getMemoryVT();
14386 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
14387 unsigned RegSz = RegVT.getSizeInBits();
14389 ISD::LoadExtType Ext = Ld->getExtensionType();
14391 assert((Ext == ISD::EXTLOAD || Ext == ISD::SEXTLOAD)
14392 && "Only anyext and sext are currently implemented.");
14393 assert(MemVT != RegVT && "Cannot extend to the same type");
14394 assert(MemVT.isVector() && "Must load a vector from memory");
14396 unsigned NumElems = RegVT.getVectorNumElements();
14397 unsigned MemSz = MemVT.getSizeInBits();
14398 assert(RegSz > MemSz && "Register size must be greater than the mem size");
14400 if (Ext == ISD::SEXTLOAD && RegSz == 256 && !Subtarget->hasInt256()) {
14401 // The only way in which we have a legal 256-bit vector result but not the
14402 // integer 256-bit operations needed to directly lower a sextload is if we
14403 // have AVX1 but not AVX2. In that case, we can always emit a sextload to
14404 // a 128-bit vector and a normal sign_extend to 256-bits that should get
14405 // correctly legalized. We do this late to allow the canonical form of
14406 // sextload to persist throughout the rest of the DAG combiner -- it wants
14407 // to fold together any extensions it can, and so will fuse a sign_extend
14408 // of an sextload into a sextload targeting a wider value.
14410 if (MemSz == 128) {
14411 // Just switch this to a normal load.
14412 assert(TLI.isTypeLegal(MemVT) && "If the memory type is a 128-bit type, "
14413 "it must be a legal 128-bit vector "
14415 Load = DAG.getLoad(MemVT, dl, Ld->getChain(), Ld->getBasePtr(),
14416 Ld->getPointerInfo(), Ld->isVolatile(), Ld->isNonTemporal(),
14417 Ld->isInvariant(), Ld->getAlignment());
14419 assert(MemSz < 128 &&
14420 "Can't extend a type wider than 128 bits to a 256 bit vector!");
14421 // Do an sext load to a 128-bit vector type. We want to use the same
14422 // number of elements, but elements half as wide. This will end up being
14423 // recursively lowered by this routine, but will succeed as we definitely
14424 // have all the necessary features if we're using AVX1.
14426 EVT::getIntegerVT(*DAG.getContext(), RegVT.getScalarSizeInBits() / 2);
14427 EVT HalfVecVT = EVT::getVectorVT(*DAG.getContext(), HalfEltVT, NumElems);
14429 DAG.getExtLoad(Ext, dl, HalfVecVT, Ld->getChain(), Ld->getBasePtr(),
14430 Ld->getPointerInfo(), MemVT, Ld->isVolatile(),
14431 Ld->isNonTemporal(), Ld->isInvariant(),
14432 Ld->getAlignment());
14435 // Replace chain users with the new chain.
14436 assert(Load->getNumValues() == 2 && "Loads must carry a chain!");
14437 DAG.ReplaceAllUsesOfValueWith(SDValue(Ld, 1), Load.getValue(1));
14439 // Finally, do a normal sign-extend to the desired register.
14440 return DAG.getSExtOrTrunc(Load, dl, RegVT);
14443 // All sizes must be a power of two.
14444 assert(isPowerOf2_32(RegSz * MemSz * NumElems) &&
14445 "Non-power-of-two elements are not custom lowered!");
14447 // Attempt to load the original value using scalar loads.
14448 // Find the largest scalar type that divides the total loaded size.
14449 MVT SclrLoadTy = MVT::i8;
14450 for (MVT Tp : MVT::integer_valuetypes()) {
14451 if (TLI.isTypeLegal(Tp) && ((MemSz % Tp.getSizeInBits()) == 0)) {
14456 // On 32bit systems, we can't save 64bit integers. Try bitcasting to F64.
14457 if (TLI.isTypeLegal(MVT::f64) && SclrLoadTy.getSizeInBits() < 64 &&
14459 SclrLoadTy = MVT::f64;
14461 // Calculate the number of scalar loads that we need to perform
14462 // in order to load our vector from memory.
14463 unsigned NumLoads = MemSz / SclrLoadTy.getSizeInBits();
14465 assert((Ext != ISD::SEXTLOAD || NumLoads == 1) &&
14466 "Can only lower sext loads with a single scalar load!");
14468 unsigned loadRegZize = RegSz;
14469 if (Ext == ISD::SEXTLOAD && RegSz >= 256)
14472 // Represent our vector as a sequence of elements which are the
14473 // largest scalar that we can load.
14474 EVT LoadUnitVecVT = EVT::getVectorVT(
14475 *DAG.getContext(), SclrLoadTy, loadRegZize / SclrLoadTy.getSizeInBits());
14477 // Represent the data using the same element type that is stored in
14478 // memory. In practice, we ''widen'' MemVT.
14480 EVT::getVectorVT(*DAG.getContext(), MemVT.getScalarType(),
14481 loadRegZize / MemVT.getScalarType().getSizeInBits());
14483 assert(WideVecVT.getSizeInBits() == LoadUnitVecVT.getSizeInBits() &&
14484 "Invalid vector type");
14486 // We can't shuffle using an illegal type.
14487 assert(TLI.isTypeLegal(WideVecVT) &&
14488 "We only lower types that form legal widened vector types");
14490 SmallVector<SDValue, 8> Chains;
14491 SDValue Ptr = Ld->getBasePtr();
14492 SDValue Increment = DAG.getConstant(SclrLoadTy.getSizeInBits() / 8, dl,
14493 TLI.getPointerTy(DAG.getDataLayout()));
14494 SDValue Res = DAG.getUNDEF(LoadUnitVecVT);
14496 for (unsigned i = 0; i < NumLoads; ++i) {
14497 // Perform a single load.
14498 SDValue ScalarLoad =
14499 DAG.getLoad(SclrLoadTy, dl, Ld->getChain(), Ptr, Ld->getPointerInfo(),
14500 Ld->isVolatile(), Ld->isNonTemporal(), Ld->isInvariant(),
14501 Ld->getAlignment());
14502 Chains.push_back(ScalarLoad.getValue(1));
14503 // Create the first element type using SCALAR_TO_VECTOR in order to avoid
14504 // another round of DAGCombining.
14506 Res = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, LoadUnitVecVT, ScalarLoad);
14508 Res = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, LoadUnitVecVT, Res,
14509 ScalarLoad, DAG.getIntPtrConstant(i, dl));
14511 Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr, Increment);
14514 SDValue TF = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Chains);
14516 // Bitcast the loaded value to a vector of the original element type, in
14517 // the size of the target vector type.
14518 SDValue SlicedVec = DAG.getBitcast(WideVecVT, Res);
14519 unsigned SizeRatio = RegSz / MemSz;
14521 if (Ext == ISD::SEXTLOAD) {
14522 // If we have SSE4.1, we can directly emit a VSEXT node.
14523 if (Subtarget->hasSSE41()) {
14524 SDValue Sext = DAG.getNode(X86ISD::VSEXT, dl, RegVT, SlicedVec);
14525 DAG.ReplaceAllUsesOfValueWith(SDValue(Ld, 1), TF);
14529 // Otherwise we'll shuffle the small elements in the high bits of the
14530 // larger type and perform an arithmetic shift. If the shift is not legal
14531 // it's better to scalarize.
14532 assert(TLI.isOperationLegalOrCustom(ISD::SRA, RegVT) &&
14533 "We can't implement a sext load without an arithmetic right shift!");
14535 // Redistribute the loaded elements into the different locations.
14536 SmallVector<int, 16> ShuffleVec(NumElems * SizeRatio, -1);
14537 for (unsigned i = 0; i != NumElems; ++i)
14538 ShuffleVec[i * SizeRatio + SizeRatio - 1] = i;
14540 SDValue Shuff = DAG.getVectorShuffle(
14541 WideVecVT, dl, SlicedVec, DAG.getUNDEF(WideVecVT), &ShuffleVec[0]);
14543 Shuff = DAG.getBitcast(RegVT, Shuff);
14545 // Build the arithmetic shift.
14546 unsigned Amt = RegVT.getVectorElementType().getSizeInBits() -
14547 MemVT.getVectorElementType().getSizeInBits();
14549 DAG.getNode(ISD::SRA, dl, RegVT, Shuff,
14550 DAG.getConstant(Amt, dl, RegVT));
14552 DAG.ReplaceAllUsesOfValueWith(SDValue(Ld, 1), TF);
14556 // Redistribute the loaded elements into the different locations.
14557 SmallVector<int, 16> ShuffleVec(NumElems * SizeRatio, -1);
14558 for (unsigned i = 0; i != NumElems; ++i)
14559 ShuffleVec[i * SizeRatio] = i;
14561 SDValue Shuff = DAG.getVectorShuffle(WideVecVT, dl, SlicedVec,
14562 DAG.getUNDEF(WideVecVT), &ShuffleVec[0]);
14564 // Bitcast to the requested type.
14565 Shuff = DAG.getBitcast(RegVT, Shuff);
14566 DAG.ReplaceAllUsesOfValueWith(SDValue(Ld, 1), TF);
14570 // isAndOrOfSingleUseSetCCs - Return true if node is an ISD::AND or
14571 // ISD::OR of two X86ISD::SETCC nodes each of which has no other use apart
14572 // from the AND / OR.
14573 static bool isAndOrOfSetCCs(SDValue Op, unsigned &Opc) {
14574 Opc = Op.getOpcode();
14575 if (Opc != ISD::OR && Opc != ISD::AND)
14577 return (Op.getOperand(0).getOpcode() == X86ISD::SETCC &&
14578 Op.getOperand(0).hasOneUse() &&
14579 Op.getOperand(1).getOpcode() == X86ISD::SETCC &&
14580 Op.getOperand(1).hasOneUse());
14583 // isXor1OfSetCC - Return true if node is an ISD::XOR of a X86ISD::SETCC and
14584 // 1 and that the SETCC node has a single use.
14585 static bool isXor1OfSetCC(SDValue Op) {
14586 if (Op.getOpcode() != ISD::XOR)
14588 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(Op.getOperand(1));
14589 if (N1C && N1C->getAPIntValue() == 1) {
14590 return Op.getOperand(0).getOpcode() == X86ISD::SETCC &&
14591 Op.getOperand(0).hasOneUse();
14596 SDValue X86TargetLowering::LowerBRCOND(SDValue Op, SelectionDAG &DAG) const {
14597 bool addTest = true;
14598 SDValue Chain = Op.getOperand(0);
14599 SDValue Cond = Op.getOperand(1);
14600 SDValue Dest = Op.getOperand(2);
14603 bool Inverted = false;
14605 if (Cond.getOpcode() == ISD::SETCC) {
14606 // Check for setcc([su]{add,sub,mul}o == 0).
14607 if (cast<CondCodeSDNode>(Cond.getOperand(2))->get() == ISD::SETEQ &&
14608 isa<ConstantSDNode>(Cond.getOperand(1)) &&
14609 cast<ConstantSDNode>(Cond.getOperand(1))->isNullValue() &&
14610 Cond.getOperand(0).getResNo() == 1 &&
14611 (Cond.getOperand(0).getOpcode() == ISD::SADDO ||
14612 Cond.getOperand(0).getOpcode() == ISD::UADDO ||
14613 Cond.getOperand(0).getOpcode() == ISD::SSUBO ||
14614 Cond.getOperand(0).getOpcode() == ISD::USUBO ||
14615 Cond.getOperand(0).getOpcode() == ISD::SMULO ||
14616 Cond.getOperand(0).getOpcode() == ISD::UMULO)) {
14618 Cond = Cond.getOperand(0);
14620 SDValue NewCond = LowerSETCC(Cond, DAG);
14621 if (NewCond.getNode())
14626 // FIXME: LowerXALUO doesn't handle these!!
14627 else if (Cond.getOpcode() == X86ISD::ADD ||
14628 Cond.getOpcode() == X86ISD::SUB ||
14629 Cond.getOpcode() == X86ISD::SMUL ||
14630 Cond.getOpcode() == X86ISD::UMUL)
14631 Cond = LowerXALUO(Cond, DAG);
14634 // Look pass (and (setcc_carry (cmp ...)), 1).
14635 if (Cond.getOpcode() == ISD::AND &&
14636 Cond.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY) {
14637 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Cond.getOperand(1));
14638 if (C && C->getAPIntValue() == 1)
14639 Cond = Cond.getOperand(0);
14642 // If condition flag is set by a X86ISD::CMP, then use it as the condition
14643 // setting operand in place of the X86ISD::SETCC.
14644 unsigned CondOpcode = Cond.getOpcode();
14645 if (CondOpcode == X86ISD::SETCC ||
14646 CondOpcode == X86ISD::SETCC_CARRY) {
14647 CC = Cond.getOperand(0);
14649 SDValue Cmp = Cond.getOperand(1);
14650 unsigned Opc = Cmp.getOpcode();
14651 // FIXME: WHY THE SPECIAL CASING OF LogicalCmp??
14652 if (isX86LogicalCmp(Cmp) || Opc == X86ISD::BT) {
14656 switch (cast<ConstantSDNode>(CC)->getZExtValue()) {
14660 // These can only come from an arithmetic instruction with overflow,
14661 // e.g. SADDO, UADDO.
14662 Cond = Cond.getNode()->getOperand(1);
14668 CondOpcode = Cond.getOpcode();
14669 if (CondOpcode == ISD::UADDO || CondOpcode == ISD::SADDO ||
14670 CondOpcode == ISD::USUBO || CondOpcode == ISD::SSUBO ||
14671 ((CondOpcode == ISD::UMULO || CondOpcode == ISD::SMULO) &&
14672 Cond.getOperand(0).getValueType() != MVT::i8)) {
14673 SDValue LHS = Cond.getOperand(0);
14674 SDValue RHS = Cond.getOperand(1);
14675 unsigned X86Opcode;
14678 // Keep this in sync with LowerXALUO, otherwise we might create redundant
14679 // instructions that can't be removed afterwards (i.e. X86ISD::ADD and
14681 switch (CondOpcode) {
14682 case ISD::UADDO: X86Opcode = X86ISD::ADD; X86Cond = X86::COND_B; break;
14684 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(RHS))
14686 X86Opcode = X86ISD::INC; X86Cond = X86::COND_O;
14689 X86Opcode = X86ISD::ADD; X86Cond = X86::COND_O; break;
14690 case ISD::USUBO: X86Opcode = X86ISD::SUB; X86Cond = X86::COND_B; break;
14692 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(RHS))
14694 X86Opcode = X86ISD::DEC; X86Cond = X86::COND_O;
14697 X86Opcode = X86ISD::SUB; X86Cond = X86::COND_O; break;
14698 case ISD::UMULO: X86Opcode = X86ISD::UMUL; X86Cond = X86::COND_O; break;
14699 case ISD::SMULO: X86Opcode = X86ISD::SMUL; X86Cond = X86::COND_O; break;
14700 default: llvm_unreachable("unexpected overflowing operator");
14703 X86Cond = X86::GetOppositeBranchCondition((X86::CondCode)X86Cond);
14704 if (CondOpcode == ISD::UMULO)
14705 VTs = DAG.getVTList(LHS.getValueType(), LHS.getValueType(),
14708 VTs = DAG.getVTList(LHS.getValueType(), MVT::i32);
14710 SDValue X86Op = DAG.getNode(X86Opcode, dl, VTs, LHS, RHS);
14712 if (CondOpcode == ISD::UMULO)
14713 Cond = X86Op.getValue(2);
14715 Cond = X86Op.getValue(1);
14717 CC = DAG.getConstant(X86Cond, dl, MVT::i8);
14721 if (Cond.hasOneUse() && isAndOrOfSetCCs(Cond, CondOpc)) {
14722 SDValue Cmp = Cond.getOperand(0).getOperand(1);
14723 if (CondOpc == ISD::OR) {
14724 // Also, recognize the pattern generated by an FCMP_UNE. We can emit
14725 // two branches instead of an explicit OR instruction with a
14727 if (Cmp == Cond.getOperand(1).getOperand(1) &&
14728 isX86LogicalCmp(Cmp)) {
14729 CC = Cond.getOperand(0).getOperand(0);
14730 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
14731 Chain, Dest, CC, Cmp);
14732 CC = Cond.getOperand(1).getOperand(0);
14736 } else { // ISD::AND
14737 // Also, recognize the pattern generated by an FCMP_OEQ. We can emit
14738 // two branches instead of an explicit AND instruction with a
14739 // separate test. However, we only do this if this block doesn't
14740 // have a fall-through edge, because this requires an explicit
14741 // jmp when the condition is false.
14742 if (Cmp == Cond.getOperand(1).getOperand(1) &&
14743 isX86LogicalCmp(Cmp) &&
14744 Op.getNode()->hasOneUse()) {
14745 X86::CondCode CCode =
14746 (X86::CondCode)Cond.getOperand(0).getConstantOperandVal(0);
14747 CCode = X86::GetOppositeBranchCondition(CCode);
14748 CC = DAG.getConstant(CCode, dl, MVT::i8);
14749 SDNode *User = *Op.getNode()->use_begin();
14750 // Look for an unconditional branch following this conditional branch.
14751 // We need this because we need to reverse the successors in order
14752 // to implement FCMP_OEQ.
14753 if (User->getOpcode() == ISD::BR) {
14754 SDValue FalseBB = User->getOperand(1);
14756 DAG.UpdateNodeOperands(User, User->getOperand(0), Dest);
14757 assert(NewBR == User);
14761 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
14762 Chain, Dest, CC, Cmp);
14763 X86::CondCode CCode =
14764 (X86::CondCode)Cond.getOperand(1).getConstantOperandVal(0);
14765 CCode = X86::GetOppositeBranchCondition(CCode);
14766 CC = DAG.getConstant(CCode, dl, MVT::i8);
14772 } else if (Cond.hasOneUse() && isXor1OfSetCC(Cond)) {
14773 // Recognize for xorb (setcc), 1 patterns. The xor inverts the condition.
14774 // It should be transformed during dag combiner except when the condition
14775 // is set by a arithmetics with overflow node.
14776 X86::CondCode CCode =
14777 (X86::CondCode)Cond.getOperand(0).getConstantOperandVal(0);
14778 CCode = X86::GetOppositeBranchCondition(CCode);
14779 CC = DAG.getConstant(CCode, dl, MVT::i8);
14780 Cond = Cond.getOperand(0).getOperand(1);
14782 } else if (Cond.getOpcode() == ISD::SETCC &&
14783 cast<CondCodeSDNode>(Cond.getOperand(2))->get() == ISD::SETOEQ) {
14784 // For FCMP_OEQ, we can emit
14785 // two branches instead of an explicit AND instruction with a
14786 // separate test. However, we only do this if this block doesn't
14787 // have a fall-through edge, because this requires an explicit
14788 // jmp when the condition is false.
14789 if (Op.getNode()->hasOneUse()) {
14790 SDNode *User = *Op.getNode()->use_begin();
14791 // Look for an unconditional branch following this conditional branch.
14792 // We need this because we need to reverse the successors in order
14793 // to implement FCMP_OEQ.
14794 if (User->getOpcode() == ISD::BR) {
14795 SDValue FalseBB = User->getOperand(1);
14797 DAG.UpdateNodeOperands(User, User->getOperand(0), Dest);
14798 assert(NewBR == User);
14802 SDValue Cmp = DAG.getNode(X86ISD::CMP, dl, MVT::i32,
14803 Cond.getOperand(0), Cond.getOperand(1));
14804 Cmp = ConvertCmpIfNecessary(Cmp, DAG);
14805 CC = DAG.getConstant(X86::COND_NE, dl, MVT::i8);
14806 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
14807 Chain, Dest, CC, Cmp);
14808 CC = DAG.getConstant(X86::COND_P, dl, MVT::i8);
14813 } else if (Cond.getOpcode() == ISD::SETCC &&
14814 cast<CondCodeSDNode>(Cond.getOperand(2))->get() == ISD::SETUNE) {
14815 // For FCMP_UNE, we can emit
14816 // two branches instead of an explicit AND instruction with a
14817 // separate test. However, we only do this if this block doesn't
14818 // have a fall-through edge, because this requires an explicit
14819 // jmp when the condition is false.
14820 if (Op.getNode()->hasOneUse()) {
14821 SDNode *User = *Op.getNode()->use_begin();
14822 // Look for an unconditional branch following this conditional branch.
14823 // We need this because we need to reverse the successors in order
14824 // to implement FCMP_UNE.
14825 if (User->getOpcode() == ISD::BR) {
14826 SDValue FalseBB = User->getOperand(1);
14828 DAG.UpdateNodeOperands(User, User->getOperand(0), Dest);
14829 assert(NewBR == User);
14832 SDValue Cmp = DAG.getNode(X86ISD::CMP, dl, MVT::i32,
14833 Cond.getOperand(0), Cond.getOperand(1));
14834 Cmp = ConvertCmpIfNecessary(Cmp, DAG);
14835 CC = DAG.getConstant(X86::COND_NE, dl, MVT::i8);
14836 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
14837 Chain, Dest, CC, Cmp);
14838 CC = DAG.getConstant(X86::COND_NP, dl, MVT::i8);
14848 // Look pass the truncate if the high bits are known zero.
14849 if (isTruncWithZeroHighBitsInput(Cond, DAG))
14850 Cond = Cond.getOperand(0);
14852 // We know the result of AND is compared against zero. Try to match
14854 if (Cond.getOpcode() == ISD::AND && Cond.hasOneUse()) {
14855 SDValue NewSetCC = LowerToBT(Cond, ISD::SETNE, dl, DAG);
14856 if (NewSetCC.getNode()) {
14857 CC = NewSetCC.getOperand(0);
14858 Cond = NewSetCC.getOperand(1);
14865 X86::CondCode X86Cond = Inverted ? X86::COND_E : X86::COND_NE;
14866 CC = DAG.getConstant(X86Cond, dl, MVT::i8);
14867 Cond = EmitTest(Cond, X86Cond, dl, DAG);
14869 Cond = ConvertCmpIfNecessary(Cond, DAG);
14870 return DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
14871 Chain, Dest, CC, Cond);
14874 // Lower dynamic stack allocation to _alloca call for Cygwin/Mingw targets.
14875 // Calls to _alloca are needed to probe the stack when allocating more than 4k
14876 // bytes in one go. Touching the stack at 4K increments is necessary to ensure
14877 // that the guard pages used by the OS virtual memory manager are allocated in
14878 // correct sequence.
14880 X86TargetLowering::LowerDYNAMIC_STACKALLOC(SDValue Op,
14881 SelectionDAG &DAG) const {
14882 MachineFunction &MF = DAG.getMachineFunction();
14883 bool SplitStack = MF.shouldSplitStack();
14884 bool Lower = (Subtarget->isOSWindows() && !Subtarget->isTargetMachO()) ||
14889 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
14890 SDNode* Node = Op.getNode();
14892 unsigned SPReg = TLI.getStackPointerRegisterToSaveRestore();
14893 assert(SPReg && "Target cannot require DYNAMIC_STACKALLOC expansion and"
14894 " not tell us which reg is the stack pointer!");
14895 EVT VT = Node->getValueType(0);
14896 SDValue Tmp1 = SDValue(Node, 0);
14897 SDValue Tmp2 = SDValue(Node, 1);
14898 SDValue Tmp3 = Node->getOperand(2);
14899 SDValue Chain = Tmp1.getOperand(0);
14901 // Chain the dynamic stack allocation so that it doesn't modify the stack
14902 // pointer when other instructions are using the stack.
14903 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(0, dl, true),
14906 SDValue Size = Tmp2.getOperand(1);
14907 SDValue SP = DAG.getCopyFromReg(Chain, dl, SPReg, VT);
14908 Chain = SP.getValue(1);
14909 unsigned Align = cast<ConstantSDNode>(Tmp3)->getZExtValue();
14910 const TargetFrameLowering &TFI = *Subtarget->getFrameLowering();
14911 unsigned StackAlign = TFI.getStackAlignment();
14912 Tmp1 = DAG.getNode(ISD::SUB, dl, VT, SP, Size); // Value
14913 if (Align > StackAlign)
14914 Tmp1 = DAG.getNode(ISD::AND, dl, VT, Tmp1,
14915 DAG.getConstant(-(uint64_t)Align, dl, VT));
14916 Chain = DAG.getCopyToReg(Chain, dl, SPReg, Tmp1); // Output chain
14918 Tmp2 = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(0, dl, true),
14919 DAG.getIntPtrConstant(0, dl, true), SDValue(),
14922 SDValue Ops[2] = { Tmp1, Tmp2 };
14923 return DAG.getMergeValues(Ops, dl);
14927 SDValue Chain = Op.getOperand(0);
14928 SDValue Size = Op.getOperand(1);
14929 unsigned Align = cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue();
14930 EVT VT = Op.getNode()->getValueType(0);
14932 bool Is64Bit = Subtarget->is64Bit();
14933 MVT SPTy = getPointerTy(DAG.getDataLayout());
14936 MachineRegisterInfo &MRI = MF.getRegInfo();
14939 // The 64 bit implementation of segmented stacks needs to clobber both r10
14940 // r11. This makes it impossible to use it along with nested parameters.
14941 const Function *F = MF.getFunction();
14943 for (Function::const_arg_iterator I = F->arg_begin(), E = F->arg_end();
14945 if (I->hasNestAttr())
14946 report_fatal_error("Cannot use segmented stacks with functions that "
14947 "have nested arguments.");
14950 const TargetRegisterClass *AddrRegClass = getRegClassFor(SPTy);
14951 unsigned Vreg = MRI.createVirtualRegister(AddrRegClass);
14952 Chain = DAG.getCopyToReg(Chain, dl, Vreg, Size);
14953 SDValue Value = DAG.getNode(X86ISD::SEG_ALLOCA, dl, SPTy, Chain,
14954 DAG.getRegister(Vreg, SPTy));
14955 SDValue Ops1[2] = { Value, Chain };
14956 return DAG.getMergeValues(Ops1, dl);
14959 const unsigned Reg = (Subtarget->isTarget64BitLP64() ? X86::RAX : X86::EAX);
14961 Chain = DAG.getCopyToReg(Chain, dl, Reg, Size, Flag);
14962 Flag = Chain.getValue(1);
14963 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
14965 Chain = DAG.getNode(X86ISD::WIN_ALLOCA, dl, NodeTys, Chain, Flag);
14967 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
14968 unsigned SPReg = RegInfo->getStackRegister();
14969 SDValue SP = DAG.getCopyFromReg(Chain, dl, SPReg, SPTy);
14970 Chain = SP.getValue(1);
14973 SP = DAG.getNode(ISD::AND, dl, VT, SP.getValue(0),
14974 DAG.getConstant(-(uint64_t)Align, dl, VT));
14975 Chain = DAG.getCopyToReg(Chain, dl, SPReg, SP);
14978 SDValue Ops1[2] = { SP, Chain };
14979 return DAG.getMergeValues(Ops1, dl);
14983 SDValue X86TargetLowering::LowerVASTART(SDValue Op, SelectionDAG &DAG) const {
14984 MachineFunction &MF = DAG.getMachineFunction();
14985 auto PtrVT = getPointerTy(MF.getDataLayout());
14986 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
14988 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
14991 if (!Subtarget->is64Bit() || Subtarget->isTargetWin64()) {
14992 // vastart just stores the address of the VarArgsFrameIndex slot into the
14993 // memory location argument.
14994 SDValue FR = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(), PtrVT);
14995 return DAG.getStore(Op.getOperand(0), DL, FR, Op.getOperand(1),
14996 MachinePointerInfo(SV), false, false, 0);
15000 // gp_offset (0 - 6 * 8)
15001 // fp_offset (48 - 48 + 8 * 16)
15002 // overflow_arg_area (point to parameters coming in memory).
15004 SmallVector<SDValue, 8> MemOps;
15005 SDValue FIN = Op.getOperand(1);
15007 SDValue Store = DAG.getStore(Op.getOperand(0), DL,
15008 DAG.getConstant(FuncInfo->getVarArgsGPOffset(),
15010 FIN, MachinePointerInfo(SV), false, false, 0);
15011 MemOps.push_back(Store);
15014 FIN = DAG.getNode(ISD::ADD, DL, PtrVT, FIN, DAG.getIntPtrConstant(4, DL));
15015 Store = DAG.getStore(Op.getOperand(0), DL,
15016 DAG.getConstant(FuncInfo->getVarArgsFPOffset(), DL,
15018 FIN, MachinePointerInfo(SV, 4), false, false, 0);
15019 MemOps.push_back(Store);
15021 // Store ptr to overflow_arg_area
15022 FIN = DAG.getNode(ISD::ADD, DL, PtrVT, FIN, DAG.getIntPtrConstant(4, DL));
15023 SDValue OVFIN = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(), PtrVT);
15024 Store = DAG.getStore(Op.getOperand(0), DL, OVFIN, FIN,
15025 MachinePointerInfo(SV, 8),
15027 MemOps.push_back(Store);
15029 // Store ptr to reg_save_area.
15030 FIN = DAG.getNode(ISD::ADD, DL, PtrVT, FIN, DAG.getIntPtrConstant(8, DL));
15031 SDValue RSFIN = DAG.getFrameIndex(FuncInfo->getRegSaveFrameIndex(), PtrVT);
15032 Store = DAG.getStore(Op.getOperand(0), DL, RSFIN, FIN,
15033 MachinePointerInfo(SV, 16), false, false, 0);
15034 MemOps.push_back(Store);
15035 return DAG.getNode(ISD::TokenFactor, DL, MVT::Other, MemOps);
15038 SDValue X86TargetLowering::LowerVAARG(SDValue Op, SelectionDAG &DAG) const {
15039 assert(Subtarget->is64Bit() &&
15040 "LowerVAARG only handles 64-bit va_arg!");
15041 assert((Subtarget->isTargetLinux() ||
15042 Subtarget->isTargetDarwin()) &&
15043 "Unhandled target in LowerVAARG");
15044 assert(Op.getNode()->getNumOperands() == 4);
15045 SDValue Chain = Op.getOperand(0);
15046 SDValue SrcPtr = Op.getOperand(1);
15047 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
15048 unsigned Align = Op.getConstantOperandVal(3);
15051 EVT ArgVT = Op.getNode()->getValueType(0);
15052 Type *ArgTy = ArgVT.getTypeForEVT(*DAG.getContext());
15053 uint32_t ArgSize = DAG.getDataLayout().getTypeAllocSize(ArgTy);
15056 // Decide which area this value should be read from.
15057 // TODO: Implement the AMD64 ABI in its entirety. This simple
15058 // selection mechanism works only for the basic types.
15059 if (ArgVT == MVT::f80) {
15060 llvm_unreachable("va_arg for f80 not yet implemented");
15061 } else if (ArgVT.isFloatingPoint() && ArgSize <= 16 /*bytes*/) {
15062 ArgMode = 2; // Argument passed in XMM register. Use fp_offset.
15063 } else if (ArgVT.isInteger() && ArgSize <= 32 /*bytes*/) {
15064 ArgMode = 1; // Argument passed in GPR64 register(s). Use gp_offset.
15066 llvm_unreachable("Unhandled argument type in LowerVAARG");
15069 if (ArgMode == 2) {
15070 // Sanity Check: Make sure using fp_offset makes sense.
15071 assert(!Subtarget->useSoftFloat() &&
15072 !(DAG.getMachineFunction().getFunction()->hasFnAttribute(
15073 Attribute::NoImplicitFloat)) &&
15074 Subtarget->hasSSE1());
15077 // Insert VAARG_64 node into the DAG
15078 // VAARG_64 returns two values: Variable Argument Address, Chain
15079 SDValue InstOps[] = {Chain, SrcPtr, DAG.getConstant(ArgSize, dl, MVT::i32),
15080 DAG.getConstant(ArgMode, dl, MVT::i8),
15081 DAG.getConstant(Align, dl, MVT::i32)};
15082 SDVTList VTs = DAG.getVTList(getPointerTy(DAG.getDataLayout()), MVT::Other);
15083 SDValue VAARG = DAG.getMemIntrinsicNode(X86ISD::VAARG_64, dl,
15084 VTs, InstOps, MVT::i64,
15085 MachinePointerInfo(SV),
15087 /*Volatile=*/false,
15089 /*WriteMem=*/true);
15090 Chain = VAARG.getValue(1);
15092 // Load the next argument and return it
15093 return DAG.getLoad(ArgVT, dl,
15096 MachinePointerInfo(),
15097 false, false, false, 0);
15100 static SDValue LowerVACOPY(SDValue Op, const X86Subtarget *Subtarget,
15101 SelectionDAG &DAG) {
15102 // X86-64 va_list is a struct { i32, i32, i8*, i8* }.
15103 assert(Subtarget->is64Bit() && "This code only handles 64-bit va_copy!");
15104 SDValue Chain = Op.getOperand(0);
15105 SDValue DstPtr = Op.getOperand(1);
15106 SDValue SrcPtr = Op.getOperand(2);
15107 const Value *DstSV = cast<SrcValueSDNode>(Op.getOperand(3))->getValue();
15108 const Value *SrcSV = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
15111 return DAG.getMemcpy(Chain, DL, DstPtr, SrcPtr,
15112 DAG.getIntPtrConstant(24, DL), 8, /*isVolatile*/false,
15114 MachinePointerInfo(DstSV), MachinePointerInfo(SrcSV));
15117 // getTargetVShiftByConstNode - Handle vector element shifts where the shift
15118 // amount is a constant. Takes immediate version of shift as input.
15119 static SDValue getTargetVShiftByConstNode(unsigned Opc, SDLoc dl, MVT VT,
15120 SDValue SrcOp, uint64_t ShiftAmt,
15121 SelectionDAG &DAG) {
15122 MVT ElementType = VT.getVectorElementType();
15124 // Fold this packed shift into its first operand if ShiftAmt is 0.
15128 // Check for ShiftAmt >= element width
15129 if (ShiftAmt >= ElementType.getSizeInBits()) {
15130 if (Opc == X86ISD::VSRAI)
15131 ShiftAmt = ElementType.getSizeInBits() - 1;
15133 return DAG.getConstant(0, dl, VT);
15136 assert((Opc == X86ISD::VSHLI || Opc == X86ISD::VSRLI || Opc == X86ISD::VSRAI)
15137 && "Unknown target vector shift-by-constant node");
15139 // Fold this packed vector shift into a build vector if SrcOp is a
15140 // vector of Constants or UNDEFs, and SrcOp valuetype is the same as VT.
15141 if (VT == SrcOp.getSimpleValueType() &&
15142 ISD::isBuildVectorOfConstantSDNodes(SrcOp.getNode())) {
15143 SmallVector<SDValue, 8> Elts;
15144 unsigned NumElts = SrcOp->getNumOperands();
15145 ConstantSDNode *ND;
15148 default: llvm_unreachable(nullptr);
15149 case X86ISD::VSHLI:
15150 for (unsigned i=0; i!=NumElts; ++i) {
15151 SDValue CurrentOp = SrcOp->getOperand(i);
15152 if (CurrentOp->getOpcode() == ISD::UNDEF) {
15153 Elts.push_back(CurrentOp);
15156 ND = cast<ConstantSDNode>(CurrentOp);
15157 const APInt &C = ND->getAPIntValue();
15158 Elts.push_back(DAG.getConstant(C.shl(ShiftAmt), dl, ElementType));
15161 case X86ISD::VSRLI:
15162 for (unsigned i=0; i!=NumElts; ++i) {
15163 SDValue CurrentOp = SrcOp->getOperand(i);
15164 if (CurrentOp->getOpcode() == ISD::UNDEF) {
15165 Elts.push_back(CurrentOp);
15168 ND = cast<ConstantSDNode>(CurrentOp);
15169 const APInt &C = ND->getAPIntValue();
15170 Elts.push_back(DAG.getConstant(C.lshr(ShiftAmt), dl, ElementType));
15173 case X86ISD::VSRAI:
15174 for (unsigned i=0; i!=NumElts; ++i) {
15175 SDValue CurrentOp = SrcOp->getOperand(i);
15176 if (CurrentOp->getOpcode() == ISD::UNDEF) {
15177 Elts.push_back(CurrentOp);
15180 ND = cast<ConstantSDNode>(CurrentOp);
15181 const APInt &C = ND->getAPIntValue();
15182 Elts.push_back(DAG.getConstant(C.ashr(ShiftAmt), dl, ElementType));
15187 return DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Elts);
15190 return DAG.getNode(Opc, dl, VT, SrcOp,
15191 DAG.getConstant(ShiftAmt, dl, MVT::i8));
15194 // getTargetVShiftNode - Handle vector element shifts where the shift amount
15195 // may or may not be a constant. Takes immediate version of shift as input.
15196 static SDValue getTargetVShiftNode(unsigned Opc, SDLoc dl, MVT VT,
15197 SDValue SrcOp, SDValue ShAmt,
15198 SelectionDAG &DAG) {
15199 MVT SVT = ShAmt.getSimpleValueType();
15200 assert((SVT == MVT::i32 || SVT == MVT::i64) && "Unexpected value type!");
15202 // Catch shift-by-constant.
15203 if (ConstantSDNode *CShAmt = dyn_cast<ConstantSDNode>(ShAmt))
15204 return getTargetVShiftByConstNode(Opc, dl, VT, SrcOp,
15205 CShAmt->getZExtValue(), DAG);
15207 // Change opcode to non-immediate version
15209 default: llvm_unreachable("Unknown target vector shift node");
15210 case X86ISD::VSHLI: Opc = X86ISD::VSHL; break;
15211 case X86ISD::VSRLI: Opc = X86ISD::VSRL; break;
15212 case X86ISD::VSRAI: Opc = X86ISD::VSRA; break;
15215 const X86Subtarget &Subtarget =
15216 static_cast<const X86Subtarget &>(DAG.getSubtarget());
15217 if (Subtarget.hasSSE41() && ShAmt.getOpcode() == ISD::ZERO_EXTEND &&
15218 ShAmt.getOperand(0).getSimpleValueType() == MVT::i16) {
15219 // Let the shuffle legalizer expand this shift amount node.
15220 SDValue Op0 = ShAmt.getOperand(0);
15221 Op0 = DAG.getNode(ISD::SCALAR_TO_VECTOR, SDLoc(Op0), MVT::v8i16, Op0);
15222 ShAmt = getShuffleVectorZeroOrUndef(Op0, 0, true, &Subtarget, DAG);
15224 // Need to build a vector containing shift amount.
15225 // SSE/AVX packed shifts only use the lower 64-bit of the shift count.
15226 SmallVector<SDValue, 4> ShOps;
15227 ShOps.push_back(ShAmt);
15228 if (SVT == MVT::i32) {
15229 ShOps.push_back(DAG.getConstant(0, dl, SVT));
15230 ShOps.push_back(DAG.getUNDEF(SVT));
15232 ShOps.push_back(DAG.getUNDEF(SVT));
15234 MVT BVT = SVT == MVT::i32 ? MVT::v4i32 : MVT::v2i64;
15235 ShAmt = DAG.getNode(ISD::BUILD_VECTOR, dl, BVT, ShOps);
15238 // The return type has to be a 128-bit type with the same element
15239 // type as the input type.
15240 MVT EltVT = VT.getVectorElementType();
15241 EVT ShVT = MVT::getVectorVT(EltVT, 128/EltVT.getSizeInBits());
15243 ShAmt = DAG.getBitcast(ShVT, ShAmt);
15244 return DAG.getNode(Opc, dl, VT, SrcOp, ShAmt);
15247 /// \brief Return (and \p Op, \p Mask) for compare instructions or
15248 /// (vselect \p Mask, \p Op, \p PreservedSrc) for others along with the
15249 /// necessary casting or extending for \p Mask when lowering masking intrinsics
15250 static SDValue getVectorMaskingNode(SDValue Op, SDValue Mask,
15251 SDValue PreservedSrc,
15252 const X86Subtarget *Subtarget,
15253 SelectionDAG &DAG) {
15254 EVT VT = Op.getValueType();
15255 EVT MaskVT = EVT::getVectorVT(*DAG.getContext(),
15256 MVT::i1, VT.getVectorNumElements());
15257 SDValue VMask = SDValue();
15258 unsigned OpcodeSelect = ISD::VSELECT;
15261 assert(MaskVT.isSimple() && "invalid mask type");
15263 if (isAllOnes(Mask))
15266 if (MaskVT.bitsGT(Mask.getValueType())) {
15267 EVT newMaskVT = EVT::getIntegerVT(*DAG.getContext(),
15268 MaskVT.getSizeInBits());
15269 VMask = DAG.getBitcast(MaskVT,
15270 DAG.getNode(ISD::ANY_EXTEND, dl, newMaskVT, Mask));
15272 EVT BitcastVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
15273 Mask.getValueType().getSizeInBits());
15274 // In case when MaskVT equals v2i1 or v4i1, low 2 or 4 elements
15275 // are extracted by EXTRACT_SUBVECTOR.
15276 VMask = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MaskVT,
15277 DAG.getBitcast(BitcastVT, Mask),
15278 DAG.getIntPtrConstant(0, dl));
15281 switch (Op.getOpcode()) {
15283 case X86ISD::PCMPEQM:
15284 case X86ISD::PCMPGTM:
15286 case X86ISD::CMPMU:
15287 return DAG.getNode(ISD::AND, dl, VT, Op, VMask);
15288 case X86ISD::VTRUNC:
15289 case X86ISD::VTRUNCS:
15290 case X86ISD::VTRUNCUS:
15291 // We can't use ISD::VSELECT here because it is not always "Legal"
15292 // for the destination type. For example vpmovqb require only AVX512
15293 // and vselect that can operate on byte element type require BWI
15294 OpcodeSelect = X86ISD::SELECT;
15297 if (PreservedSrc.getOpcode() == ISD::UNDEF)
15298 PreservedSrc = getZeroVector(VT, Subtarget, DAG, dl);
15299 return DAG.getNode(OpcodeSelect, dl, VT, VMask, Op, PreservedSrc);
15302 /// \brief Creates an SDNode for a predicated scalar operation.
15303 /// \returns (X86vselect \p Mask, \p Op, \p PreservedSrc).
15304 /// The mask is comming as MVT::i8 and it should be truncated
15305 /// to MVT::i1 while lowering masking intrinsics.
15306 /// The main difference between ScalarMaskingNode and VectorMaskingNode is using
15307 /// "X86select" instead of "vselect". We just can't create the "vselect" node for
15308 /// a scalar instruction.
15309 static SDValue getScalarMaskingNode(SDValue Op, SDValue Mask,
15310 SDValue PreservedSrc,
15311 const X86Subtarget *Subtarget,
15312 SelectionDAG &DAG) {
15313 if (isAllOnes(Mask))
15316 EVT VT = Op.getValueType();
15318 // The mask should be of type MVT::i1
15319 SDValue IMask = DAG.getNode(ISD::TRUNCATE, dl, MVT::i1, Mask);
15321 if (PreservedSrc.getOpcode() == ISD::UNDEF)
15322 PreservedSrc = getZeroVector(VT, Subtarget, DAG, dl);
15323 return DAG.getNode(X86ISD::SELECT, dl, VT, IMask, Op, PreservedSrc);
15326 static int getSEHRegistrationNodeSize(const Function *Fn) {
15327 if (!Fn->hasPersonalityFn())
15328 report_fatal_error(
15329 "querying registration node size for function without personality");
15330 // The RegNodeSize is 6 32-bit words for SEH and 4 for C++ EH. See
15331 // WinEHStatePass for the full struct definition.
15332 switch (classifyEHPersonality(Fn->getPersonalityFn())) {
15333 case EHPersonality::MSVC_X86SEH: return 24;
15334 case EHPersonality::MSVC_CXX: return 16;
15337 report_fatal_error("can only recover FP for MSVC EH personality functions");
15340 /// When the 32-bit MSVC runtime transfers control to us, either to an outlined
15341 /// function or when returning to a parent frame after catching an exception, we
15342 /// recover the parent frame pointer by doing arithmetic on the incoming EBP.
15343 /// Here's the math:
15344 /// RegNodeBase = EntryEBP - RegNodeSize
15345 /// ParentFP = RegNodeBase - RegNodeFrameOffset
15346 /// Subtracting RegNodeSize takes us to the offset of the registration node, and
15347 /// subtracting the offset (negative on x86) takes us back to the parent FP.
15348 static SDValue recoverFramePointer(SelectionDAG &DAG, const Function *Fn,
15349 SDValue EntryEBP) {
15350 MachineFunction &MF = DAG.getMachineFunction();
15353 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
15354 MVT PtrVT = TLI.getPointerTy(DAG.getDataLayout());
15356 // It's possible that the parent function no longer has a personality function
15357 // if the exceptional code was optimized away, in which case we just return
15358 // the incoming EBP.
15359 if (!Fn->hasPersonalityFn())
15362 int RegNodeSize = getSEHRegistrationNodeSize(Fn);
15364 // Get an MCSymbol that will ultimately resolve to the frame offset of the EH
15366 MCSymbol *OffsetSym =
15367 MF.getMMI().getContext().getOrCreateParentFrameOffsetSymbol(
15368 GlobalValue::getRealLinkageName(Fn->getName()));
15369 SDValue OffsetSymVal = DAG.getMCSymbol(OffsetSym, PtrVT);
15370 SDValue RegNodeFrameOffset =
15371 DAG.getNode(ISD::LOCAL_RECOVER, dl, PtrVT, OffsetSymVal);
15373 // RegNodeBase = EntryEBP - RegNodeSize
15374 // ParentFP = RegNodeBase - RegNodeFrameOffset
15375 SDValue RegNodeBase = DAG.getNode(ISD::SUB, dl, PtrVT, EntryEBP,
15376 DAG.getConstant(RegNodeSize, dl, PtrVT));
15377 return DAG.getNode(ISD::SUB, dl, PtrVT, RegNodeBase, RegNodeFrameOffset);
15380 static SDValue LowerINTRINSIC_WO_CHAIN(SDValue Op, const X86Subtarget *Subtarget,
15381 SelectionDAG &DAG) {
15383 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
15384 EVT VT = Op.getValueType();
15385 const IntrinsicData* IntrData = getIntrinsicWithoutChain(IntNo);
15387 switch(IntrData->Type) {
15388 case INTR_TYPE_1OP:
15389 return DAG.getNode(IntrData->Opc0, dl, Op.getValueType(), Op.getOperand(1));
15390 case INTR_TYPE_2OP:
15391 return DAG.getNode(IntrData->Opc0, dl, Op.getValueType(), Op.getOperand(1),
15393 case INTR_TYPE_3OP:
15394 return DAG.getNode(IntrData->Opc0, dl, Op.getValueType(), Op.getOperand(1),
15395 Op.getOperand(2), Op.getOperand(3));
15396 case INTR_TYPE_4OP:
15397 return DAG.getNode(IntrData->Opc0, dl, Op.getValueType(), Op.getOperand(1),
15398 Op.getOperand(2), Op.getOperand(3), Op.getOperand(4));
15399 case INTR_TYPE_1OP_MASK_RM: {
15400 SDValue Src = Op.getOperand(1);
15401 SDValue PassThru = Op.getOperand(2);
15402 SDValue Mask = Op.getOperand(3);
15403 SDValue RoundingMode;
15404 // We allways add rounding mode to the Node.
15405 // If the rounding mode is not specified, we add the
15406 // "current direction" mode.
15407 if (Op.getNumOperands() == 4)
15409 DAG.getConstant(X86::STATIC_ROUNDING::CUR_DIRECTION, dl, MVT::i32);
15411 RoundingMode = Op.getOperand(4);
15412 unsigned IntrWithRoundingModeOpcode = IntrData->Opc1;
15413 if (IntrWithRoundingModeOpcode != 0)
15414 if (cast<ConstantSDNode>(RoundingMode)->getZExtValue() !=
15415 X86::STATIC_ROUNDING::CUR_DIRECTION)
15416 return getVectorMaskingNode(DAG.getNode(IntrWithRoundingModeOpcode,
15417 dl, Op.getValueType(), Src, RoundingMode),
15418 Mask, PassThru, Subtarget, DAG);
15419 return getVectorMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT, Src,
15421 Mask, PassThru, Subtarget, DAG);
15423 case INTR_TYPE_1OP_MASK: {
15424 SDValue Src = Op.getOperand(1);
15425 SDValue PassThru = Op.getOperand(2);
15426 SDValue Mask = Op.getOperand(3);
15427 // We add rounding mode to the Node when
15428 // - RM Opcode is specified and
15429 // - RM is not "current direction".
15430 unsigned IntrWithRoundingModeOpcode = IntrData->Opc1;
15431 if (IntrWithRoundingModeOpcode != 0) {
15432 SDValue Rnd = Op.getOperand(4);
15433 unsigned Round = cast<ConstantSDNode>(Rnd)->getZExtValue();
15434 if (Round != X86::STATIC_ROUNDING::CUR_DIRECTION) {
15435 return getVectorMaskingNode(DAG.getNode(IntrWithRoundingModeOpcode,
15436 dl, Op.getValueType(),
15438 Mask, PassThru, Subtarget, DAG);
15441 return getVectorMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT, Src),
15442 Mask, PassThru, Subtarget, DAG);
15444 case INTR_TYPE_SCALAR_MASK_RM: {
15445 SDValue Src1 = Op.getOperand(1);
15446 SDValue Src2 = Op.getOperand(2);
15447 SDValue Src0 = Op.getOperand(3);
15448 SDValue Mask = Op.getOperand(4);
15449 // There are 2 kinds of intrinsics in this group:
15450 // (1) With supress-all-exceptions (sae) or rounding mode- 6 operands
15451 // (2) With rounding mode and sae - 7 operands.
15452 if (Op.getNumOperands() == 6) {
15453 SDValue Sae = Op.getOperand(5);
15454 unsigned Opc = IntrData->Opc1 ? IntrData->Opc1 : IntrData->Opc0;
15455 return getScalarMaskingNode(DAG.getNode(Opc, dl, VT, Src1, Src2,
15457 Mask, Src0, Subtarget, DAG);
15459 assert(Op.getNumOperands() == 7 && "Unexpected intrinsic form");
15460 SDValue RoundingMode = Op.getOperand(5);
15461 SDValue Sae = Op.getOperand(6);
15462 return getScalarMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT, Src1, Src2,
15463 RoundingMode, Sae),
15464 Mask, Src0, Subtarget, DAG);
15466 case INTR_TYPE_2OP_MASK: {
15467 SDValue Src1 = Op.getOperand(1);
15468 SDValue Src2 = Op.getOperand(2);
15469 SDValue PassThru = Op.getOperand(3);
15470 SDValue Mask = Op.getOperand(4);
15471 // We specify 2 possible opcodes for intrinsics with rounding modes.
15472 // First, we check if the intrinsic may have non-default rounding mode,
15473 // (IntrData->Opc1 != 0), then we check the rounding mode operand.
15474 unsigned IntrWithRoundingModeOpcode = IntrData->Opc1;
15475 if (IntrWithRoundingModeOpcode != 0) {
15476 SDValue Rnd = Op.getOperand(5);
15477 unsigned Round = cast<ConstantSDNode>(Rnd)->getZExtValue();
15478 if (Round != X86::STATIC_ROUNDING::CUR_DIRECTION) {
15479 return getVectorMaskingNode(DAG.getNode(IntrWithRoundingModeOpcode,
15480 dl, Op.getValueType(),
15482 Mask, PassThru, Subtarget, DAG);
15485 return getVectorMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT,
15487 Mask, PassThru, Subtarget, DAG);
15489 case INTR_TYPE_2OP_MASK_RM: {
15490 SDValue Src1 = Op.getOperand(1);
15491 SDValue Src2 = Op.getOperand(2);
15492 SDValue PassThru = Op.getOperand(3);
15493 SDValue Mask = Op.getOperand(4);
15494 // We specify 2 possible modes for intrinsics, with/without rounding modes.
15495 // First, we check if the intrinsic have rounding mode (6 operands),
15496 // if not, we set rounding mode to "current".
15498 if (Op.getNumOperands() == 6)
15499 Rnd = Op.getOperand(5);
15501 Rnd = DAG.getConstant(X86::STATIC_ROUNDING::CUR_DIRECTION, dl, MVT::i32);
15502 return getVectorMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT,
15504 Mask, PassThru, Subtarget, DAG);
15506 case INTR_TYPE_3OP_MASK_RM: {
15507 SDValue Src1 = Op.getOperand(1);
15508 SDValue Src2 = Op.getOperand(2);
15509 SDValue Imm = Op.getOperand(3);
15510 SDValue PassThru = Op.getOperand(4);
15511 SDValue Mask = Op.getOperand(5);
15512 // We specify 2 possible modes for intrinsics, with/without rounding modes.
15513 // First, we check if the intrinsic have rounding mode (7 operands),
15514 // if not, we set rounding mode to "current".
15516 if (Op.getNumOperands() == 7)
15517 Rnd = Op.getOperand(6);
15519 Rnd = DAG.getConstant(X86::STATIC_ROUNDING::CUR_DIRECTION, dl, MVT::i32);
15520 return getVectorMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT,
15521 Src1, Src2, Imm, Rnd),
15522 Mask, PassThru, Subtarget, DAG);
15524 case INTR_TYPE_3OP_MASK: {
15525 SDValue Src1 = Op.getOperand(1);
15526 SDValue Src2 = Op.getOperand(2);
15527 SDValue Src3 = Op.getOperand(3);
15528 SDValue PassThru = Op.getOperand(4);
15529 SDValue Mask = Op.getOperand(5);
15530 // We specify 2 possible opcodes for intrinsics with rounding modes.
15531 // First, we check if the intrinsic may have non-default rounding mode,
15532 // (IntrData->Opc1 != 0), then we check the rounding mode operand.
15533 unsigned IntrWithRoundingModeOpcode = IntrData->Opc1;
15534 if (IntrWithRoundingModeOpcode != 0) {
15535 SDValue Rnd = Op.getOperand(6);
15536 unsigned Round = cast<ConstantSDNode>(Rnd)->getZExtValue();
15537 if (Round != X86::STATIC_ROUNDING::CUR_DIRECTION) {
15538 return getVectorMaskingNode(DAG.getNode(IntrWithRoundingModeOpcode,
15539 dl, Op.getValueType(),
15540 Src1, Src2, Src3, Rnd),
15541 Mask, PassThru, Subtarget, DAG);
15544 return getVectorMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT,
15546 Mask, PassThru, Subtarget, DAG);
15548 case VPERM_3OP_MASKZ:
15549 case VPERM_3OP_MASK:
15552 case FMA_OP_MASK: {
15553 SDValue Src1 = Op.getOperand(1);
15554 SDValue Src2 = Op.getOperand(2);
15555 SDValue Src3 = Op.getOperand(3);
15556 SDValue Mask = Op.getOperand(4);
15557 EVT VT = Op.getValueType();
15558 SDValue PassThru = SDValue();
15560 // set PassThru element
15561 if (IntrData->Type == VPERM_3OP_MASKZ || IntrData->Type == FMA_OP_MASKZ)
15562 PassThru = getZeroVector(VT, Subtarget, DAG, dl);
15563 else if (IntrData->Type == FMA_OP_MASK3)
15568 // We specify 2 possible opcodes for intrinsics with rounding modes.
15569 // First, we check if the intrinsic may have non-default rounding mode,
15570 // (IntrData->Opc1 != 0), then we check the rounding mode operand.
15571 unsigned IntrWithRoundingModeOpcode = IntrData->Opc1;
15572 if (IntrWithRoundingModeOpcode != 0) {
15573 SDValue Rnd = Op.getOperand(5);
15574 if (cast<ConstantSDNode>(Rnd)->getZExtValue() !=
15575 X86::STATIC_ROUNDING::CUR_DIRECTION)
15576 return getVectorMaskingNode(DAG.getNode(IntrWithRoundingModeOpcode,
15577 dl, Op.getValueType(),
15578 Src1, Src2, Src3, Rnd),
15579 Mask, PassThru, Subtarget, DAG);
15581 return getVectorMaskingNode(DAG.getNode(IntrData->Opc0,
15582 dl, Op.getValueType(),
15584 Mask, PassThru, Subtarget, DAG);
15587 case CMP_MASK_CC: {
15588 // Comparison intrinsics with masks.
15589 // Example of transformation:
15590 // (i8 (int_x86_avx512_mask_pcmpeq_q_128
15591 // (v2i64 %a), (v2i64 %b), (i8 %mask))) ->
15593 // (v8i1 (insert_subvector undef,
15594 // (v2i1 (and (PCMPEQM %a, %b),
15595 // (extract_subvector
15596 // (v8i1 (bitcast %mask)), 0))), 0))))
15597 EVT VT = Op.getOperand(1).getValueType();
15598 EVT MaskVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
15599 VT.getVectorNumElements());
15600 SDValue Mask = Op.getOperand((IntrData->Type == CMP_MASK_CC) ? 4 : 3);
15601 EVT BitcastVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
15602 Mask.getValueType().getSizeInBits());
15604 if (IntrData->Type == CMP_MASK_CC) {
15605 SDValue CC = Op.getOperand(3);
15606 CC = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, CC);
15607 // We specify 2 possible opcodes for intrinsics with rounding modes.
15608 // First, we check if the intrinsic may have non-default rounding mode,
15609 // (IntrData->Opc1 != 0), then we check the rounding mode operand.
15610 if (IntrData->Opc1 != 0) {
15611 SDValue Rnd = Op.getOperand(5);
15612 if (cast<ConstantSDNode>(Rnd)->getZExtValue() !=
15613 X86::STATIC_ROUNDING::CUR_DIRECTION)
15614 Cmp = DAG.getNode(IntrData->Opc1, dl, MaskVT, Op.getOperand(1),
15615 Op.getOperand(2), CC, Rnd);
15617 //default rounding mode
15619 Cmp = DAG.getNode(IntrData->Opc0, dl, MaskVT, Op.getOperand(1),
15620 Op.getOperand(2), CC);
15623 assert(IntrData->Type == CMP_MASK && "Unexpected intrinsic type!");
15624 Cmp = DAG.getNode(IntrData->Opc0, dl, MaskVT, Op.getOperand(1),
15627 SDValue CmpMask = getVectorMaskingNode(Cmp, Mask,
15628 DAG.getTargetConstant(0, dl,
15631 SDValue Res = DAG.getNode(ISD::INSERT_SUBVECTOR, dl, BitcastVT,
15632 DAG.getUNDEF(BitcastVT), CmpMask,
15633 DAG.getIntPtrConstant(0, dl));
15634 return DAG.getBitcast(Op.getValueType(), Res);
15636 case COMI: { // Comparison intrinsics
15637 ISD::CondCode CC = (ISD::CondCode)IntrData->Opc1;
15638 SDValue LHS = Op.getOperand(1);
15639 SDValue RHS = Op.getOperand(2);
15640 unsigned X86CC = TranslateX86CC(CC, dl, true, LHS, RHS, DAG);
15641 assert(X86CC != X86::COND_INVALID && "Unexpected illegal condition!");
15642 SDValue Cond = DAG.getNode(IntrData->Opc0, dl, MVT::i32, LHS, RHS);
15643 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
15644 DAG.getConstant(X86CC, dl, MVT::i8), Cond);
15645 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
15648 return getTargetVShiftNode(IntrData->Opc0, dl, Op.getSimpleValueType(),
15649 Op.getOperand(1), Op.getOperand(2), DAG);
15651 return getVectorMaskingNode(getTargetVShiftNode(IntrData->Opc0, dl,
15652 Op.getSimpleValueType(),
15654 Op.getOperand(2), DAG),
15655 Op.getOperand(4), Op.getOperand(3), Subtarget,
15657 case COMPRESS_EXPAND_IN_REG: {
15658 SDValue Mask = Op.getOperand(3);
15659 SDValue DataToCompress = Op.getOperand(1);
15660 SDValue PassThru = Op.getOperand(2);
15661 if (isAllOnes(Mask)) // return data as is
15662 return Op.getOperand(1);
15664 return getVectorMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT,
15666 Mask, PassThru, Subtarget, DAG);
15669 SDValue Mask = Op.getOperand(3);
15670 EVT VT = Op.getValueType();
15671 EVT MaskVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
15672 VT.getVectorNumElements());
15673 EVT BitcastVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
15674 Mask.getValueType().getSizeInBits());
15676 SDValue VMask = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MaskVT,
15677 DAG.getBitcast(BitcastVT, Mask),
15678 DAG.getIntPtrConstant(0, dl));
15679 return DAG.getNode(IntrData->Opc0, dl, VT, VMask, Op.getOperand(1),
15688 default: return SDValue(); // Don't custom lower most intrinsics.
15690 case Intrinsic::x86_avx2_permd:
15691 case Intrinsic::x86_avx2_permps:
15692 // Operands intentionally swapped. Mask is last operand to intrinsic,
15693 // but second operand for node/instruction.
15694 return DAG.getNode(X86ISD::VPERMV, dl, Op.getValueType(),
15695 Op.getOperand(2), Op.getOperand(1));
15697 // ptest and testp intrinsics. The intrinsic these come from are designed to
15698 // return an integer value, not just an instruction so lower it to the ptest
15699 // or testp pattern and a setcc for the result.
15700 case Intrinsic::x86_sse41_ptestz:
15701 case Intrinsic::x86_sse41_ptestc:
15702 case Intrinsic::x86_sse41_ptestnzc:
15703 case Intrinsic::x86_avx_ptestz_256:
15704 case Intrinsic::x86_avx_ptestc_256:
15705 case Intrinsic::x86_avx_ptestnzc_256:
15706 case Intrinsic::x86_avx_vtestz_ps:
15707 case Intrinsic::x86_avx_vtestc_ps:
15708 case Intrinsic::x86_avx_vtestnzc_ps:
15709 case Intrinsic::x86_avx_vtestz_pd:
15710 case Intrinsic::x86_avx_vtestc_pd:
15711 case Intrinsic::x86_avx_vtestnzc_pd:
15712 case Intrinsic::x86_avx_vtestz_ps_256:
15713 case Intrinsic::x86_avx_vtestc_ps_256:
15714 case Intrinsic::x86_avx_vtestnzc_ps_256:
15715 case Intrinsic::x86_avx_vtestz_pd_256:
15716 case Intrinsic::x86_avx_vtestc_pd_256:
15717 case Intrinsic::x86_avx_vtestnzc_pd_256: {
15718 bool IsTestPacked = false;
15721 default: llvm_unreachable("Bad fallthrough in Intrinsic lowering.");
15722 case Intrinsic::x86_avx_vtestz_ps:
15723 case Intrinsic::x86_avx_vtestz_pd:
15724 case Intrinsic::x86_avx_vtestz_ps_256:
15725 case Intrinsic::x86_avx_vtestz_pd_256:
15726 IsTestPacked = true; // Fallthrough
15727 case Intrinsic::x86_sse41_ptestz:
15728 case Intrinsic::x86_avx_ptestz_256:
15730 X86CC = X86::COND_E;
15732 case Intrinsic::x86_avx_vtestc_ps:
15733 case Intrinsic::x86_avx_vtestc_pd:
15734 case Intrinsic::x86_avx_vtestc_ps_256:
15735 case Intrinsic::x86_avx_vtestc_pd_256:
15736 IsTestPacked = true; // Fallthrough
15737 case Intrinsic::x86_sse41_ptestc:
15738 case Intrinsic::x86_avx_ptestc_256:
15740 X86CC = X86::COND_B;
15742 case Intrinsic::x86_avx_vtestnzc_ps:
15743 case Intrinsic::x86_avx_vtestnzc_pd:
15744 case Intrinsic::x86_avx_vtestnzc_ps_256:
15745 case Intrinsic::x86_avx_vtestnzc_pd_256:
15746 IsTestPacked = true; // Fallthrough
15747 case Intrinsic::x86_sse41_ptestnzc:
15748 case Intrinsic::x86_avx_ptestnzc_256:
15750 X86CC = X86::COND_A;
15754 SDValue LHS = Op.getOperand(1);
15755 SDValue RHS = Op.getOperand(2);
15756 unsigned TestOpc = IsTestPacked ? X86ISD::TESTP : X86ISD::PTEST;
15757 SDValue Test = DAG.getNode(TestOpc, dl, MVT::i32, LHS, RHS);
15758 SDValue CC = DAG.getConstant(X86CC, dl, MVT::i8);
15759 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8, CC, Test);
15760 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
15762 case Intrinsic::x86_avx512_kortestz_w:
15763 case Intrinsic::x86_avx512_kortestc_w: {
15764 unsigned X86CC = (IntNo == Intrinsic::x86_avx512_kortestz_w)? X86::COND_E: X86::COND_B;
15765 SDValue LHS = DAG.getBitcast(MVT::v16i1, Op.getOperand(1));
15766 SDValue RHS = DAG.getBitcast(MVT::v16i1, Op.getOperand(2));
15767 SDValue CC = DAG.getConstant(X86CC, dl, MVT::i8);
15768 SDValue Test = DAG.getNode(X86ISD::KORTEST, dl, MVT::i32, LHS, RHS);
15769 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i1, CC, Test);
15770 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
15773 case Intrinsic::x86_sse42_pcmpistria128:
15774 case Intrinsic::x86_sse42_pcmpestria128:
15775 case Intrinsic::x86_sse42_pcmpistric128:
15776 case Intrinsic::x86_sse42_pcmpestric128:
15777 case Intrinsic::x86_sse42_pcmpistrio128:
15778 case Intrinsic::x86_sse42_pcmpestrio128:
15779 case Intrinsic::x86_sse42_pcmpistris128:
15780 case Intrinsic::x86_sse42_pcmpestris128:
15781 case Intrinsic::x86_sse42_pcmpistriz128:
15782 case Intrinsic::x86_sse42_pcmpestriz128: {
15786 default: llvm_unreachable("Impossible intrinsic"); // Can't reach here.
15787 case Intrinsic::x86_sse42_pcmpistria128:
15788 Opcode = X86ISD::PCMPISTRI;
15789 X86CC = X86::COND_A;
15791 case Intrinsic::x86_sse42_pcmpestria128:
15792 Opcode = X86ISD::PCMPESTRI;
15793 X86CC = X86::COND_A;
15795 case Intrinsic::x86_sse42_pcmpistric128:
15796 Opcode = X86ISD::PCMPISTRI;
15797 X86CC = X86::COND_B;
15799 case Intrinsic::x86_sse42_pcmpestric128:
15800 Opcode = X86ISD::PCMPESTRI;
15801 X86CC = X86::COND_B;
15803 case Intrinsic::x86_sse42_pcmpistrio128:
15804 Opcode = X86ISD::PCMPISTRI;
15805 X86CC = X86::COND_O;
15807 case Intrinsic::x86_sse42_pcmpestrio128:
15808 Opcode = X86ISD::PCMPESTRI;
15809 X86CC = X86::COND_O;
15811 case Intrinsic::x86_sse42_pcmpistris128:
15812 Opcode = X86ISD::PCMPISTRI;
15813 X86CC = X86::COND_S;
15815 case Intrinsic::x86_sse42_pcmpestris128:
15816 Opcode = X86ISD::PCMPESTRI;
15817 X86CC = X86::COND_S;
15819 case Intrinsic::x86_sse42_pcmpistriz128:
15820 Opcode = X86ISD::PCMPISTRI;
15821 X86CC = X86::COND_E;
15823 case Intrinsic::x86_sse42_pcmpestriz128:
15824 Opcode = X86ISD::PCMPESTRI;
15825 X86CC = X86::COND_E;
15828 SmallVector<SDValue, 5> NewOps(Op->op_begin()+1, Op->op_end());
15829 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::i32);
15830 SDValue PCMP = DAG.getNode(Opcode, dl, VTs, NewOps);
15831 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
15832 DAG.getConstant(X86CC, dl, MVT::i8),
15833 SDValue(PCMP.getNode(), 1));
15834 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
15837 case Intrinsic::x86_sse42_pcmpistri128:
15838 case Intrinsic::x86_sse42_pcmpestri128: {
15840 if (IntNo == Intrinsic::x86_sse42_pcmpistri128)
15841 Opcode = X86ISD::PCMPISTRI;
15843 Opcode = X86ISD::PCMPESTRI;
15845 SmallVector<SDValue, 5> NewOps(Op->op_begin()+1, Op->op_end());
15846 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::i32);
15847 return DAG.getNode(Opcode, dl, VTs, NewOps);
15850 case Intrinsic::x86_seh_lsda: {
15851 // Compute the symbol for the LSDA. We know it'll get emitted later.
15852 MachineFunction &MF = DAG.getMachineFunction();
15853 SDValue Op1 = Op.getOperand(1);
15854 auto *Fn = cast<Function>(cast<GlobalAddressSDNode>(Op1)->getGlobal());
15855 MCSymbol *LSDASym = MF.getMMI().getContext().getOrCreateLSDASymbol(
15856 GlobalValue::getRealLinkageName(Fn->getName()));
15858 // Generate a simple absolute symbol reference. This intrinsic is only
15859 // supported on 32-bit Windows, which isn't PIC.
15860 SDValue Result = DAG.getMCSymbol(LSDASym, VT);
15861 return DAG.getNode(X86ISD::Wrapper, dl, VT, Result);
15864 case Intrinsic::x86_seh_recoverfp: {
15865 SDValue FnOp = Op.getOperand(1);
15866 SDValue IncomingFPOp = Op.getOperand(2);
15867 GlobalAddressSDNode *GSD = dyn_cast<GlobalAddressSDNode>(FnOp);
15868 auto *Fn = dyn_cast_or_null<Function>(GSD ? GSD->getGlobal() : nullptr);
15870 report_fatal_error(
15871 "llvm.x86.seh.recoverfp must take a function as the first argument");
15872 return recoverFramePointer(DAG, Fn, IncomingFPOp);
15875 case Intrinsic::localaddress: {
15876 // Returns one of the stack, base, or frame pointer registers, depending on
15877 // which is used to reference local variables.
15878 MachineFunction &MF = DAG.getMachineFunction();
15879 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
15881 if (RegInfo->hasBasePointer(MF))
15882 Reg = RegInfo->getBaseRegister();
15883 else // This function handles the SP or FP case.
15884 Reg = RegInfo->getPtrSizedFrameRegister(MF);
15885 return DAG.getCopyFromReg(DAG.getEntryNode(), dl, Reg, VT);
15890 static SDValue getGatherNode(unsigned Opc, SDValue Op, SelectionDAG &DAG,
15891 SDValue Src, SDValue Mask, SDValue Base,
15892 SDValue Index, SDValue ScaleOp, SDValue Chain,
15893 const X86Subtarget * Subtarget) {
15895 ConstantSDNode *C = dyn_cast<ConstantSDNode>(ScaleOp);
15897 llvm_unreachable("Invalid scale type");
15898 unsigned ScaleVal = C->getZExtValue();
15899 if (ScaleVal > 2 && ScaleVal != 4 && ScaleVal != 8)
15900 llvm_unreachable("Valid scale values are 1, 2, 4, 8");
15902 SDValue Scale = DAG.getTargetConstant(C->getZExtValue(), dl, MVT::i8);
15903 EVT MaskVT = MVT::getVectorVT(MVT::i1,
15904 Index.getSimpleValueType().getVectorNumElements());
15906 ConstantSDNode *MaskC = dyn_cast<ConstantSDNode>(Mask);
15908 MaskInReg = DAG.getTargetConstant(MaskC->getSExtValue(), dl, MaskVT);
15910 EVT BitcastVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
15911 Mask.getValueType().getSizeInBits());
15913 // In case when MaskVT equals v2i1 or v4i1, low 2 or 4 elements
15914 // are extracted by EXTRACT_SUBVECTOR.
15915 MaskInReg = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MaskVT,
15916 DAG.getBitcast(BitcastVT, Mask),
15917 DAG.getIntPtrConstant(0, dl));
15919 SDVTList VTs = DAG.getVTList(Op.getValueType(), MaskVT, MVT::Other);
15920 SDValue Disp = DAG.getTargetConstant(0, dl, MVT::i32);
15921 SDValue Segment = DAG.getRegister(0, MVT::i32);
15922 if (Src.getOpcode() == ISD::UNDEF)
15923 Src = getZeroVector(Op.getValueType(), Subtarget, DAG, dl);
15924 SDValue Ops[] = {Src, MaskInReg, Base, Scale, Index, Disp, Segment, Chain};
15925 SDNode *Res = DAG.getMachineNode(Opc, dl, VTs, Ops);
15926 SDValue RetOps[] = { SDValue(Res, 0), SDValue(Res, 2) };
15927 return DAG.getMergeValues(RetOps, dl);
15930 static SDValue getScatterNode(unsigned Opc, SDValue Op, SelectionDAG &DAG,
15931 SDValue Src, SDValue Mask, SDValue Base,
15932 SDValue Index, SDValue ScaleOp, SDValue Chain) {
15934 ConstantSDNode *C = dyn_cast<ConstantSDNode>(ScaleOp);
15936 llvm_unreachable("Invalid scale type");
15937 unsigned ScaleVal = C->getZExtValue();
15938 if (ScaleVal > 2 && ScaleVal != 4 && ScaleVal != 8)
15939 llvm_unreachable("Valid scale values are 1, 2, 4, 8");
15941 SDValue Scale = DAG.getTargetConstant(C->getZExtValue(), dl, MVT::i8);
15942 SDValue Disp = DAG.getTargetConstant(0, dl, MVT::i32);
15943 SDValue Segment = DAG.getRegister(0, MVT::i32);
15944 EVT MaskVT = MVT::getVectorVT(MVT::i1,
15945 Index.getSimpleValueType().getVectorNumElements());
15947 ConstantSDNode *MaskC = dyn_cast<ConstantSDNode>(Mask);
15949 MaskInReg = DAG.getTargetConstant(MaskC->getSExtValue(), dl, MaskVT);
15951 EVT BitcastVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
15952 Mask.getValueType().getSizeInBits());
15954 // In case when MaskVT equals v2i1 or v4i1, low 2 or 4 elements
15955 // are extracted by EXTRACT_SUBVECTOR.
15956 MaskInReg = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MaskVT,
15957 DAG.getBitcast(BitcastVT, Mask),
15958 DAG.getIntPtrConstant(0, dl));
15960 SDVTList VTs = DAG.getVTList(MaskVT, MVT::Other);
15961 SDValue Ops[] = {Base, Scale, Index, Disp, Segment, MaskInReg, Src, Chain};
15962 SDNode *Res = DAG.getMachineNode(Opc, dl, VTs, Ops);
15963 return SDValue(Res, 1);
15966 static SDValue getPrefetchNode(unsigned Opc, SDValue Op, SelectionDAG &DAG,
15967 SDValue Mask, SDValue Base, SDValue Index,
15968 SDValue ScaleOp, SDValue Chain) {
15970 ConstantSDNode *C = dyn_cast<ConstantSDNode>(ScaleOp);
15971 assert(C && "Invalid scale type");
15972 SDValue Scale = DAG.getTargetConstant(C->getZExtValue(), dl, MVT::i8);
15973 SDValue Disp = DAG.getTargetConstant(0, dl, MVT::i32);
15974 SDValue Segment = DAG.getRegister(0, MVT::i32);
15976 MVT::getVectorVT(MVT::i1, Index.getSimpleValueType().getVectorNumElements());
15978 ConstantSDNode *MaskC = dyn_cast<ConstantSDNode>(Mask);
15980 MaskInReg = DAG.getTargetConstant(MaskC->getSExtValue(), dl, MaskVT);
15982 MaskInReg = DAG.getBitcast(MaskVT, Mask);
15983 //SDVTList VTs = DAG.getVTList(MVT::Other);
15984 SDValue Ops[] = {MaskInReg, Base, Scale, Index, Disp, Segment, Chain};
15985 SDNode *Res = DAG.getMachineNode(Opc, dl, MVT::Other, Ops);
15986 return SDValue(Res, 0);
15989 // getReadPerformanceCounter - Handles the lowering of builtin intrinsics that
15990 // read performance monitor counters (x86_rdpmc).
15991 static void getReadPerformanceCounter(SDNode *N, SDLoc DL,
15992 SelectionDAG &DAG, const X86Subtarget *Subtarget,
15993 SmallVectorImpl<SDValue> &Results) {
15994 assert(N->getNumOperands() == 3 && "Unexpected number of operands!");
15995 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Glue);
15998 // The ECX register is used to select the index of the performance counter
16000 SDValue Chain = DAG.getCopyToReg(N->getOperand(0), DL, X86::ECX,
16002 SDValue rd = DAG.getNode(X86ISD::RDPMC_DAG, DL, Tys, Chain);
16004 // Reads the content of a 64-bit performance counter and returns it in the
16005 // registers EDX:EAX.
16006 if (Subtarget->is64Bit()) {
16007 LO = DAG.getCopyFromReg(rd, DL, X86::RAX, MVT::i64, rd.getValue(1));
16008 HI = DAG.getCopyFromReg(LO.getValue(1), DL, X86::RDX, MVT::i64,
16011 LO = DAG.getCopyFromReg(rd, DL, X86::EAX, MVT::i32, rd.getValue(1));
16012 HI = DAG.getCopyFromReg(LO.getValue(1), DL, X86::EDX, MVT::i32,
16015 Chain = HI.getValue(1);
16017 if (Subtarget->is64Bit()) {
16018 // The EAX register is loaded with the low-order 32 bits. The EDX register
16019 // is loaded with the supported high-order bits of the counter.
16020 SDValue Tmp = DAG.getNode(ISD::SHL, DL, MVT::i64, HI,
16021 DAG.getConstant(32, DL, MVT::i8));
16022 Results.push_back(DAG.getNode(ISD::OR, DL, MVT::i64, LO, Tmp));
16023 Results.push_back(Chain);
16027 // Use a buildpair to merge the two 32-bit values into a 64-bit one.
16028 SDValue Ops[] = { LO, HI };
16029 SDValue Pair = DAG.getNode(ISD::BUILD_PAIR, DL, MVT::i64, Ops);
16030 Results.push_back(Pair);
16031 Results.push_back(Chain);
16034 // getReadTimeStampCounter - Handles the lowering of builtin intrinsics that
16035 // read the time stamp counter (x86_rdtsc and x86_rdtscp). This function is
16036 // also used to custom lower READCYCLECOUNTER nodes.
16037 static void getReadTimeStampCounter(SDNode *N, SDLoc DL, unsigned Opcode,
16038 SelectionDAG &DAG, const X86Subtarget *Subtarget,
16039 SmallVectorImpl<SDValue> &Results) {
16040 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Glue);
16041 SDValue rd = DAG.getNode(Opcode, DL, Tys, N->getOperand(0));
16044 // The processor's time-stamp counter (a 64-bit MSR) is stored into the
16045 // EDX:EAX registers. EDX is loaded with the high-order 32 bits of the MSR
16046 // and the EAX register is loaded with the low-order 32 bits.
16047 if (Subtarget->is64Bit()) {
16048 LO = DAG.getCopyFromReg(rd, DL, X86::RAX, MVT::i64, rd.getValue(1));
16049 HI = DAG.getCopyFromReg(LO.getValue(1), DL, X86::RDX, MVT::i64,
16052 LO = DAG.getCopyFromReg(rd, DL, X86::EAX, MVT::i32, rd.getValue(1));
16053 HI = DAG.getCopyFromReg(LO.getValue(1), DL, X86::EDX, MVT::i32,
16056 SDValue Chain = HI.getValue(1);
16058 if (Opcode == X86ISD::RDTSCP_DAG) {
16059 assert(N->getNumOperands() == 3 && "Unexpected number of operands!");
16061 // Instruction RDTSCP loads the IA32:TSC_AUX_MSR (address C000_0103H) into
16062 // the ECX register. Add 'ecx' explicitly to the chain.
16063 SDValue ecx = DAG.getCopyFromReg(Chain, DL, X86::ECX, MVT::i32,
16065 // Explicitly store the content of ECX at the location passed in input
16066 // to the 'rdtscp' intrinsic.
16067 Chain = DAG.getStore(ecx.getValue(1), DL, ecx, N->getOperand(2),
16068 MachinePointerInfo(), false, false, 0);
16071 if (Subtarget->is64Bit()) {
16072 // The EDX register is loaded with the high-order 32 bits of the MSR, and
16073 // the EAX register is loaded with the low-order 32 bits.
16074 SDValue Tmp = DAG.getNode(ISD::SHL, DL, MVT::i64, HI,
16075 DAG.getConstant(32, DL, MVT::i8));
16076 Results.push_back(DAG.getNode(ISD::OR, DL, MVT::i64, LO, Tmp));
16077 Results.push_back(Chain);
16081 // Use a buildpair to merge the two 32-bit values into a 64-bit one.
16082 SDValue Ops[] = { LO, HI };
16083 SDValue Pair = DAG.getNode(ISD::BUILD_PAIR, DL, MVT::i64, Ops);
16084 Results.push_back(Pair);
16085 Results.push_back(Chain);
16088 static SDValue LowerREADCYCLECOUNTER(SDValue Op, const X86Subtarget *Subtarget,
16089 SelectionDAG &DAG) {
16090 SmallVector<SDValue, 2> Results;
16092 getReadTimeStampCounter(Op.getNode(), DL, X86ISD::RDTSC_DAG, DAG, Subtarget,
16094 return DAG.getMergeValues(Results, DL);
16097 static SDValue LowerSEHRESTOREFRAME(SDValue Op, const X86Subtarget *Subtarget,
16098 SelectionDAG &DAG) {
16099 MachineFunction &MF = DAG.getMachineFunction();
16100 const Function *Fn = MF.getFunction();
16102 SDValue Chain = Op.getOperand(0);
16104 assert(Subtarget->getFrameLowering()->hasFP(MF) &&
16105 "using llvm.x86.seh.restoreframe requires a frame pointer");
16107 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
16108 MVT VT = TLI.getPointerTy(DAG.getDataLayout());
16110 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
16111 unsigned FrameReg =
16112 RegInfo->getPtrSizedFrameRegister(DAG.getMachineFunction());
16113 unsigned SPReg = RegInfo->getStackRegister();
16114 unsigned SlotSize = RegInfo->getSlotSize();
16116 // Get incoming EBP.
16117 SDValue IncomingEBP =
16118 DAG.getCopyFromReg(Chain, dl, FrameReg, VT);
16120 // SP is saved in the first field of every registration node, so load
16121 // [EBP-RegNodeSize] into SP.
16122 int RegNodeSize = getSEHRegistrationNodeSize(Fn);
16123 SDValue SPAddr = DAG.getNode(ISD::ADD, dl, VT, IncomingEBP,
16124 DAG.getConstant(-RegNodeSize, dl, VT));
16126 DAG.getLoad(VT, dl, Chain, SPAddr, MachinePointerInfo(), false, false,
16127 false, VT.getScalarSizeInBits() / 8);
16128 Chain = DAG.getCopyToReg(Chain, dl, SPReg, NewSP);
16130 if (!RegInfo->needsStackRealignment(MF)) {
16131 // Adjust EBP to point back to the original frame position.
16132 SDValue NewFP = recoverFramePointer(DAG, Fn, IncomingEBP);
16133 Chain = DAG.getCopyToReg(Chain, dl, FrameReg, NewFP);
16135 assert(RegInfo->hasBasePointer(MF) &&
16136 "functions with Win32 EH must use frame or base pointer register");
16138 // Reload the base pointer (ESI) with the adjusted incoming EBP.
16139 SDValue NewBP = recoverFramePointer(DAG, Fn, IncomingEBP);
16140 Chain = DAG.getCopyToReg(Chain, dl, RegInfo->getBaseRegister(), NewBP);
16142 // Reload the spilled EBP value, now that the stack and base pointers are
16144 X86MachineFunctionInfo *X86FI = MF.getInfo<X86MachineFunctionInfo>();
16145 X86FI->setHasSEHFramePtrSave(true);
16146 int FI = MF.getFrameInfo()->CreateSpillStackObject(SlotSize, SlotSize);
16147 X86FI->setSEHFramePtrSaveIndex(FI);
16148 SDValue NewFP = DAG.getLoad(VT, dl, Chain, DAG.getFrameIndex(FI, VT),
16149 MachinePointerInfo(), false, false, false,
16150 VT.getScalarSizeInBits() / 8);
16151 Chain = DAG.getCopyToReg(NewFP, dl, FrameReg, NewFP);
16157 /// \brief Lower intrinsics for TRUNCATE_TO_MEM case
16158 /// return truncate Store/MaskedStore Node
16159 static SDValue LowerINTRINSIC_TRUNCATE_TO_MEM(const SDValue & Op,
16163 SDValue Mask = Op.getOperand(4);
16164 SDValue DataToTruncate = Op.getOperand(3);
16165 SDValue Addr = Op.getOperand(2);
16166 SDValue Chain = Op.getOperand(0);
16168 EVT VT = DataToTruncate.getValueType();
16169 EVT SVT = EVT::getVectorVT(*DAG.getContext(),
16170 ElementType, VT.getVectorNumElements());
16172 if (isAllOnes(Mask)) // return just a truncate store
16173 return DAG.getTruncStore(Chain, dl, DataToTruncate, Addr,
16174 MachinePointerInfo(), SVT, false, false,
16175 SVT.getScalarSizeInBits()/8);
16177 EVT MaskVT = EVT::getVectorVT(*DAG.getContext(),
16178 MVT::i1, VT.getVectorNumElements());
16179 EVT BitcastVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
16180 Mask.getValueType().getSizeInBits());
16181 // In case when MaskVT equals v2i1 or v4i1, low 2 or 4 elements
16182 // are extracted by EXTRACT_SUBVECTOR.
16183 SDValue VMask = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MaskVT,
16184 DAG.getBitcast(BitcastVT, Mask),
16185 DAG.getIntPtrConstant(0, dl));
16187 MachineMemOperand *MMO = DAG.getMachineFunction().
16188 getMachineMemOperand(MachinePointerInfo(),
16189 MachineMemOperand::MOStore, SVT.getStoreSize(),
16190 SVT.getScalarSizeInBits()/8);
16192 return DAG.getMaskedStore(Chain, dl, DataToTruncate, Addr,
16193 VMask, SVT, MMO, true);
16196 static SDValue LowerINTRINSIC_W_CHAIN(SDValue Op, const X86Subtarget *Subtarget,
16197 SelectionDAG &DAG) {
16198 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
16200 const IntrinsicData* IntrData = getIntrinsicWithChain(IntNo);
16202 if (IntNo == llvm::Intrinsic::x86_seh_restoreframe)
16203 return LowerSEHRESTOREFRAME(Op, Subtarget, DAG);
16208 switch(IntrData->Type) {
16210 llvm_unreachable("Unknown Intrinsic Type");
16214 // Emit the node with the right value type.
16215 SDVTList VTs = DAG.getVTList(Op->getValueType(0), MVT::Glue, MVT::Other);
16216 SDValue Result = DAG.getNode(IntrData->Opc0, dl, VTs, Op.getOperand(0));
16218 // If the value returned by RDRAND/RDSEED was valid (CF=1), return 1.
16219 // Otherwise return the value from Rand, which is always 0, casted to i32.
16220 SDValue Ops[] = { DAG.getZExtOrTrunc(Result, dl, Op->getValueType(1)),
16221 DAG.getConstant(1, dl, Op->getValueType(1)),
16222 DAG.getConstant(X86::COND_B, dl, MVT::i32),
16223 SDValue(Result.getNode(), 1) };
16224 SDValue isValid = DAG.getNode(X86ISD::CMOV, dl,
16225 DAG.getVTList(Op->getValueType(1), MVT::Glue),
16228 // Return { result, isValid, chain }.
16229 return DAG.getNode(ISD::MERGE_VALUES, dl, Op->getVTList(), Result, isValid,
16230 SDValue(Result.getNode(), 2));
16233 //gather(v1, mask, index, base, scale);
16234 SDValue Chain = Op.getOperand(0);
16235 SDValue Src = Op.getOperand(2);
16236 SDValue Base = Op.getOperand(3);
16237 SDValue Index = Op.getOperand(4);
16238 SDValue Mask = Op.getOperand(5);
16239 SDValue Scale = Op.getOperand(6);
16240 return getGatherNode(IntrData->Opc0, Op, DAG, Src, Mask, Base, Index, Scale,
16244 //scatter(base, mask, index, v1, scale);
16245 SDValue Chain = Op.getOperand(0);
16246 SDValue Base = Op.getOperand(2);
16247 SDValue Mask = Op.getOperand(3);
16248 SDValue Index = Op.getOperand(4);
16249 SDValue Src = Op.getOperand(5);
16250 SDValue Scale = Op.getOperand(6);
16251 return getScatterNode(IntrData->Opc0, Op, DAG, Src, Mask, Base, Index,
16255 SDValue Hint = Op.getOperand(6);
16256 unsigned HintVal = cast<ConstantSDNode>(Hint)->getZExtValue();
16257 assert(HintVal < 2 && "Wrong prefetch hint in intrinsic: should be 0 or 1");
16258 unsigned Opcode = (HintVal ? IntrData->Opc1 : IntrData->Opc0);
16259 SDValue Chain = Op.getOperand(0);
16260 SDValue Mask = Op.getOperand(2);
16261 SDValue Index = Op.getOperand(3);
16262 SDValue Base = Op.getOperand(4);
16263 SDValue Scale = Op.getOperand(5);
16264 return getPrefetchNode(Opcode, Op, DAG, Mask, Base, Index, Scale, Chain);
16266 // Read Time Stamp Counter (RDTSC) and Processor ID (RDTSCP).
16268 SmallVector<SDValue, 2> Results;
16269 getReadTimeStampCounter(Op.getNode(), dl, IntrData->Opc0, DAG, Subtarget,
16271 return DAG.getMergeValues(Results, dl);
16273 // Read Performance Monitoring Counters.
16275 SmallVector<SDValue, 2> Results;
16276 getReadPerformanceCounter(Op.getNode(), dl, DAG, Subtarget, Results);
16277 return DAG.getMergeValues(Results, dl);
16279 // XTEST intrinsics.
16281 SDVTList VTs = DAG.getVTList(Op->getValueType(0), MVT::Other);
16282 SDValue InTrans = DAG.getNode(IntrData->Opc0, dl, VTs, Op.getOperand(0));
16283 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
16284 DAG.getConstant(X86::COND_NE, dl, MVT::i8),
16286 SDValue Ret = DAG.getNode(ISD::ZERO_EXTEND, dl, Op->getValueType(0), SetCC);
16287 return DAG.getNode(ISD::MERGE_VALUES, dl, Op->getVTList(),
16288 Ret, SDValue(InTrans.getNode(), 1));
16292 SmallVector<SDValue, 2> Results;
16293 SDVTList CFVTs = DAG.getVTList(Op->getValueType(0), MVT::Other);
16294 SDVTList VTs = DAG.getVTList(Op.getOperand(3)->getValueType(0), MVT::Other);
16295 SDValue GenCF = DAG.getNode(X86ISD::ADD, dl, CFVTs, Op.getOperand(2),
16296 DAG.getConstant(-1, dl, MVT::i8));
16297 SDValue Res = DAG.getNode(IntrData->Opc0, dl, VTs, Op.getOperand(3),
16298 Op.getOperand(4), GenCF.getValue(1));
16299 SDValue Store = DAG.getStore(Op.getOperand(0), dl, Res.getValue(0),
16300 Op.getOperand(5), MachinePointerInfo(),
16302 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
16303 DAG.getConstant(X86::COND_B, dl, MVT::i8),
16305 Results.push_back(SetCC);
16306 Results.push_back(Store);
16307 return DAG.getMergeValues(Results, dl);
16309 case COMPRESS_TO_MEM: {
16311 SDValue Mask = Op.getOperand(4);
16312 SDValue DataToCompress = Op.getOperand(3);
16313 SDValue Addr = Op.getOperand(2);
16314 SDValue Chain = Op.getOperand(0);
16316 EVT VT = DataToCompress.getValueType();
16317 if (isAllOnes(Mask)) // return just a store
16318 return DAG.getStore(Chain, dl, DataToCompress, Addr,
16319 MachinePointerInfo(), false, false,
16320 VT.getScalarSizeInBits()/8);
16322 SDValue Compressed =
16323 getVectorMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT, DataToCompress),
16324 Mask, DAG.getUNDEF(VT), Subtarget, DAG);
16325 return DAG.getStore(Chain, dl, Compressed, Addr,
16326 MachinePointerInfo(), false, false,
16327 VT.getScalarSizeInBits()/8);
16329 case TRUNCATE_TO_MEM_VI8:
16330 return LowerINTRINSIC_TRUNCATE_TO_MEM(Op, DAG, MVT::i8);
16331 case TRUNCATE_TO_MEM_VI16:
16332 return LowerINTRINSIC_TRUNCATE_TO_MEM(Op, DAG, MVT::i16);
16333 case TRUNCATE_TO_MEM_VI32:
16334 return LowerINTRINSIC_TRUNCATE_TO_MEM(Op, DAG, MVT::i32);
16335 case EXPAND_FROM_MEM: {
16337 SDValue Mask = Op.getOperand(4);
16338 SDValue PassThru = Op.getOperand(3);
16339 SDValue Addr = Op.getOperand(2);
16340 SDValue Chain = Op.getOperand(0);
16341 EVT VT = Op.getValueType();
16343 if (isAllOnes(Mask)) // return just a load
16344 return DAG.getLoad(VT, dl, Chain, Addr, MachinePointerInfo(), false, false,
16345 false, VT.getScalarSizeInBits()/8);
16347 SDValue DataToExpand = DAG.getLoad(VT, dl, Chain, Addr, MachinePointerInfo(),
16348 false, false, false,
16349 VT.getScalarSizeInBits()/8);
16351 SDValue Results[] = {
16352 getVectorMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT, DataToExpand),
16353 Mask, PassThru, Subtarget, DAG), Chain};
16354 return DAG.getMergeValues(Results, dl);
16359 SDValue X86TargetLowering::LowerRETURNADDR(SDValue Op,
16360 SelectionDAG &DAG) const {
16361 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
16362 MFI->setReturnAddressIsTaken(true);
16364 if (verifyReturnAddressArgumentIsConstant(Op, DAG))
16367 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
16369 EVT PtrVT = getPointerTy(DAG.getDataLayout());
16372 SDValue FrameAddr = LowerFRAMEADDR(Op, DAG);
16373 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
16374 SDValue Offset = DAG.getConstant(RegInfo->getSlotSize(), dl, PtrVT);
16375 return DAG.getLoad(PtrVT, dl, DAG.getEntryNode(),
16376 DAG.getNode(ISD::ADD, dl, PtrVT,
16377 FrameAddr, Offset),
16378 MachinePointerInfo(), false, false, false, 0);
16381 // Just load the return address.
16382 SDValue RetAddrFI = getReturnAddressFrameIndex(DAG);
16383 return DAG.getLoad(PtrVT, dl, DAG.getEntryNode(),
16384 RetAddrFI, MachinePointerInfo(), false, false, false, 0);
16387 SDValue X86TargetLowering::LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const {
16388 MachineFunction &MF = DAG.getMachineFunction();
16389 MachineFrameInfo *MFI = MF.getFrameInfo();
16390 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
16391 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
16392 EVT VT = Op.getValueType();
16394 MFI->setFrameAddressIsTaken(true);
16396 if (MF.getTarget().getMCAsmInfo()->usesWindowsCFI()) {
16397 // Depth > 0 makes no sense on targets which use Windows unwind codes. It
16398 // is not possible to crawl up the stack without looking at the unwind codes
16400 int FrameAddrIndex = FuncInfo->getFAIndex();
16401 if (!FrameAddrIndex) {
16402 // Set up a frame object for the return address.
16403 unsigned SlotSize = RegInfo->getSlotSize();
16404 FrameAddrIndex = MF.getFrameInfo()->CreateFixedObject(
16405 SlotSize, /*Offset=*/0, /*IsImmutable=*/false);
16406 FuncInfo->setFAIndex(FrameAddrIndex);
16408 return DAG.getFrameIndex(FrameAddrIndex, VT);
16411 unsigned FrameReg =
16412 RegInfo->getPtrSizedFrameRegister(DAG.getMachineFunction());
16413 SDLoc dl(Op); // FIXME probably not meaningful
16414 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
16415 assert(((FrameReg == X86::RBP && VT == MVT::i64) ||
16416 (FrameReg == X86::EBP && VT == MVT::i32)) &&
16417 "Invalid Frame Register!");
16418 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, VT);
16420 FrameAddr = DAG.getLoad(VT, dl, DAG.getEntryNode(), FrameAddr,
16421 MachinePointerInfo(),
16422 false, false, false, 0);
16426 // FIXME? Maybe this could be a TableGen attribute on some registers and
16427 // this table could be generated automatically from RegInfo.
16428 unsigned X86TargetLowering::getRegisterByName(const char* RegName, EVT VT,
16429 SelectionDAG &DAG) const {
16430 const TargetFrameLowering &TFI = *Subtarget->getFrameLowering();
16431 const MachineFunction &MF = DAG.getMachineFunction();
16433 unsigned Reg = StringSwitch<unsigned>(RegName)
16434 .Case("esp", X86::ESP)
16435 .Case("rsp", X86::RSP)
16436 .Case("ebp", X86::EBP)
16437 .Case("rbp", X86::RBP)
16440 if (Reg == X86::EBP || Reg == X86::RBP) {
16441 if (!TFI.hasFP(MF))
16442 report_fatal_error("register " + StringRef(RegName) +
16443 " is allocatable: function has no frame pointer");
16446 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
16447 unsigned FrameReg =
16448 RegInfo->getPtrSizedFrameRegister(DAG.getMachineFunction());
16449 assert((FrameReg == X86::EBP || FrameReg == X86::RBP) &&
16450 "Invalid Frame Register!");
16458 report_fatal_error("Invalid register name global variable");
16461 SDValue X86TargetLowering::LowerFRAME_TO_ARGS_OFFSET(SDValue Op,
16462 SelectionDAG &DAG) const {
16463 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
16464 return DAG.getIntPtrConstant(2 * RegInfo->getSlotSize(), SDLoc(Op));
16467 SDValue X86TargetLowering::LowerEH_RETURN(SDValue Op, SelectionDAG &DAG) const {
16468 SDValue Chain = Op.getOperand(0);
16469 SDValue Offset = Op.getOperand(1);
16470 SDValue Handler = Op.getOperand(2);
16473 EVT PtrVT = getPointerTy(DAG.getDataLayout());
16474 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
16475 unsigned FrameReg = RegInfo->getFrameRegister(DAG.getMachineFunction());
16476 assert(((FrameReg == X86::RBP && PtrVT == MVT::i64) ||
16477 (FrameReg == X86::EBP && PtrVT == MVT::i32)) &&
16478 "Invalid Frame Register!");
16479 SDValue Frame = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, PtrVT);
16480 unsigned StoreAddrReg = (PtrVT == MVT::i64) ? X86::RCX : X86::ECX;
16482 SDValue StoreAddr = DAG.getNode(ISD::ADD, dl, PtrVT, Frame,
16483 DAG.getIntPtrConstant(RegInfo->getSlotSize(),
16485 StoreAddr = DAG.getNode(ISD::ADD, dl, PtrVT, StoreAddr, Offset);
16486 Chain = DAG.getStore(Chain, dl, Handler, StoreAddr, MachinePointerInfo(),
16488 Chain = DAG.getCopyToReg(Chain, dl, StoreAddrReg, StoreAddr);
16490 return DAG.getNode(X86ISD::EH_RETURN, dl, MVT::Other, Chain,
16491 DAG.getRegister(StoreAddrReg, PtrVT));
16494 SDValue X86TargetLowering::lowerEH_SJLJ_SETJMP(SDValue Op,
16495 SelectionDAG &DAG) const {
16497 return DAG.getNode(X86ISD::EH_SJLJ_SETJMP, DL,
16498 DAG.getVTList(MVT::i32, MVT::Other),
16499 Op.getOperand(0), Op.getOperand(1));
16502 SDValue X86TargetLowering::lowerEH_SJLJ_LONGJMP(SDValue Op,
16503 SelectionDAG &DAG) const {
16505 return DAG.getNode(X86ISD::EH_SJLJ_LONGJMP, DL, MVT::Other,
16506 Op.getOperand(0), Op.getOperand(1));
16509 static SDValue LowerADJUST_TRAMPOLINE(SDValue Op, SelectionDAG &DAG) {
16510 return Op.getOperand(0);
16513 SDValue X86TargetLowering::LowerINIT_TRAMPOLINE(SDValue Op,
16514 SelectionDAG &DAG) const {
16515 SDValue Root = Op.getOperand(0);
16516 SDValue Trmp = Op.getOperand(1); // trampoline
16517 SDValue FPtr = Op.getOperand(2); // nested function
16518 SDValue Nest = Op.getOperand(3); // 'nest' parameter value
16521 const Value *TrmpAddr = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
16522 const TargetRegisterInfo *TRI = Subtarget->getRegisterInfo();
16524 if (Subtarget->is64Bit()) {
16525 SDValue OutChains[6];
16527 // Large code-model.
16528 const unsigned char JMP64r = 0xFF; // 64-bit jmp through register opcode.
16529 const unsigned char MOV64ri = 0xB8; // X86::MOV64ri opcode.
16531 const unsigned char N86R10 = TRI->getEncodingValue(X86::R10) & 0x7;
16532 const unsigned char N86R11 = TRI->getEncodingValue(X86::R11) & 0x7;
16534 const unsigned char REX_WB = 0x40 | 0x08 | 0x01; // REX prefix
16536 // Load the pointer to the nested function into R11.
16537 unsigned OpCode = ((MOV64ri | N86R11) << 8) | REX_WB; // movabsq r11
16538 SDValue Addr = Trmp;
16539 OutChains[0] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, dl, MVT::i16),
16540 Addr, MachinePointerInfo(TrmpAddr),
16543 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
16544 DAG.getConstant(2, dl, MVT::i64));
16545 OutChains[1] = DAG.getStore(Root, dl, FPtr, Addr,
16546 MachinePointerInfo(TrmpAddr, 2),
16549 // Load the 'nest' parameter value into R10.
16550 // R10 is specified in X86CallingConv.td
16551 OpCode = ((MOV64ri | N86R10) << 8) | REX_WB; // movabsq r10
16552 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
16553 DAG.getConstant(10, dl, MVT::i64));
16554 OutChains[2] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, dl, MVT::i16),
16555 Addr, MachinePointerInfo(TrmpAddr, 10),
16558 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
16559 DAG.getConstant(12, dl, MVT::i64));
16560 OutChains[3] = DAG.getStore(Root, dl, Nest, Addr,
16561 MachinePointerInfo(TrmpAddr, 12),
16564 // Jump to the nested function.
16565 OpCode = (JMP64r << 8) | REX_WB; // jmpq *...
16566 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
16567 DAG.getConstant(20, dl, MVT::i64));
16568 OutChains[4] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, dl, MVT::i16),
16569 Addr, MachinePointerInfo(TrmpAddr, 20),
16572 unsigned char ModRM = N86R11 | (4 << 3) | (3 << 6); // ...r11
16573 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
16574 DAG.getConstant(22, dl, MVT::i64));
16575 OutChains[5] = DAG.getStore(Root, dl, DAG.getConstant(ModRM, dl, MVT::i8),
16576 Addr, MachinePointerInfo(TrmpAddr, 22),
16579 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other, OutChains);
16581 const Function *Func =
16582 cast<Function>(cast<SrcValueSDNode>(Op.getOperand(5))->getValue());
16583 CallingConv::ID CC = Func->getCallingConv();
16588 llvm_unreachable("Unsupported calling convention");
16589 case CallingConv::C:
16590 case CallingConv::X86_StdCall: {
16591 // Pass 'nest' parameter in ECX.
16592 // Must be kept in sync with X86CallingConv.td
16593 NestReg = X86::ECX;
16595 // Check that ECX wasn't needed by an 'inreg' parameter.
16596 FunctionType *FTy = Func->getFunctionType();
16597 const AttributeSet &Attrs = Func->getAttributes();
16599 if (!Attrs.isEmpty() && !Func->isVarArg()) {
16600 unsigned InRegCount = 0;
16603 for (FunctionType::param_iterator I = FTy->param_begin(),
16604 E = FTy->param_end(); I != E; ++I, ++Idx)
16605 if (Attrs.hasAttribute(Idx, Attribute::InReg)) {
16606 auto &DL = DAG.getDataLayout();
16607 // FIXME: should only count parameters that are lowered to integers.
16608 InRegCount += (DL.getTypeSizeInBits(*I) + 31) / 32;
16611 if (InRegCount > 2) {
16612 report_fatal_error("Nest register in use - reduce number of inreg"
16618 case CallingConv::X86_FastCall:
16619 case CallingConv::X86_ThisCall:
16620 case CallingConv::Fast:
16621 // Pass 'nest' parameter in EAX.
16622 // Must be kept in sync with X86CallingConv.td
16623 NestReg = X86::EAX;
16627 SDValue OutChains[4];
16628 SDValue Addr, Disp;
16630 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
16631 DAG.getConstant(10, dl, MVT::i32));
16632 Disp = DAG.getNode(ISD::SUB, dl, MVT::i32, FPtr, Addr);
16634 // This is storing the opcode for MOV32ri.
16635 const unsigned char MOV32ri = 0xB8; // X86::MOV32ri's opcode byte.
16636 const unsigned char N86Reg = TRI->getEncodingValue(NestReg) & 0x7;
16637 OutChains[0] = DAG.getStore(Root, dl,
16638 DAG.getConstant(MOV32ri|N86Reg, dl, MVT::i8),
16639 Trmp, MachinePointerInfo(TrmpAddr),
16642 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
16643 DAG.getConstant(1, dl, MVT::i32));
16644 OutChains[1] = DAG.getStore(Root, dl, Nest, Addr,
16645 MachinePointerInfo(TrmpAddr, 1),
16648 const unsigned char JMP = 0xE9; // jmp <32bit dst> opcode.
16649 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
16650 DAG.getConstant(5, dl, MVT::i32));
16651 OutChains[2] = DAG.getStore(Root, dl, DAG.getConstant(JMP, dl, MVT::i8),
16652 Addr, MachinePointerInfo(TrmpAddr, 5),
16655 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
16656 DAG.getConstant(6, dl, MVT::i32));
16657 OutChains[3] = DAG.getStore(Root, dl, Disp, Addr,
16658 MachinePointerInfo(TrmpAddr, 6),
16661 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other, OutChains);
16665 SDValue X86TargetLowering::LowerFLT_ROUNDS_(SDValue Op,
16666 SelectionDAG &DAG) const {
16668 The rounding mode is in bits 11:10 of FPSR, and has the following
16670 00 Round to nearest
16675 FLT_ROUNDS, on the other hand, expects the following:
16682 To perform the conversion, we do:
16683 (((((FPSR & 0x800) >> 11) | ((FPSR & 0x400) >> 9)) + 1) & 3)
16686 MachineFunction &MF = DAG.getMachineFunction();
16687 const TargetFrameLowering &TFI = *Subtarget->getFrameLowering();
16688 unsigned StackAlignment = TFI.getStackAlignment();
16689 MVT VT = Op.getSimpleValueType();
16692 // Save FP Control Word to stack slot
16693 int SSFI = MF.getFrameInfo()->CreateStackObject(2, StackAlignment, false);
16694 SDValue StackSlot =
16695 DAG.getFrameIndex(SSFI, getPointerTy(DAG.getDataLayout()));
16697 MachineMemOperand *MMO =
16698 MF.getMachineMemOperand(MachinePointerInfo::getFixedStack(SSFI),
16699 MachineMemOperand::MOStore, 2, 2);
16701 SDValue Ops[] = { DAG.getEntryNode(), StackSlot };
16702 SDValue Chain = DAG.getMemIntrinsicNode(X86ISD::FNSTCW16m, DL,
16703 DAG.getVTList(MVT::Other),
16704 Ops, MVT::i16, MMO);
16706 // Load FP Control Word from stack slot
16707 SDValue CWD = DAG.getLoad(MVT::i16, DL, Chain, StackSlot,
16708 MachinePointerInfo(), false, false, false, 0);
16710 // Transform as necessary
16712 DAG.getNode(ISD::SRL, DL, MVT::i16,
16713 DAG.getNode(ISD::AND, DL, MVT::i16,
16714 CWD, DAG.getConstant(0x800, DL, MVT::i16)),
16715 DAG.getConstant(11, DL, MVT::i8));
16717 DAG.getNode(ISD::SRL, DL, MVT::i16,
16718 DAG.getNode(ISD::AND, DL, MVT::i16,
16719 CWD, DAG.getConstant(0x400, DL, MVT::i16)),
16720 DAG.getConstant(9, DL, MVT::i8));
16723 DAG.getNode(ISD::AND, DL, MVT::i16,
16724 DAG.getNode(ISD::ADD, DL, MVT::i16,
16725 DAG.getNode(ISD::OR, DL, MVT::i16, CWD1, CWD2),
16726 DAG.getConstant(1, DL, MVT::i16)),
16727 DAG.getConstant(3, DL, MVT::i16));
16729 return DAG.getNode((VT.getSizeInBits() < 16 ?
16730 ISD::TRUNCATE : ISD::ZERO_EXTEND), DL, VT, RetVal);
16733 static SDValue LowerCTLZ(SDValue Op, SelectionDAG &DAG) {
16734 MVT VT = Op.getSimpleValueType();
16736 unsigned NumBits = VT.getSizeInBits();
16739 Op = Op.getOperand(0);
16740 if (VT == MVT::i8) {
16741 // Zero extend to i32 since there is not an i8 bsr.
16743 Op = DAG.getNode(ISD::ZERO_EXTEND, dl, OpVT, Op);
16746 // Issue a bsr (scan bits in reverse) which also sets EFLAGS.
16747 SDVTList VTs = DAG.getVTList(OpVT, MVT::i32);
16748 Op = DAG.getNode(X86ISD::BSR, dl, VTs, Op);
16750 // If src is zero (i.e. bsr sets ZF), returns NumBits.
16753 DAG.getConstant(NumBits + NumBits - 1, dl, OpVT),
16754 DAG.getConstant(X86::COND_E, dl, MVT::i8),
16757 Op = DAG.getNode(X86ISD::CMOV, dl, OpVT, Ops);
16759 // Finally xor with NumBits-1.
16760 Op = DAG.getNode(ISD::XOR, dl, OpVT, Op,
16761 DAG.getConstant(NumBits - 1, dl, OpVT));
16764 Op = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op);
16768 static SDValue LowerCTLZ_ZERO_UNDEF(SDValue Op, SelectionDAG &DAG) {
16769 MVT VT = Op.getSimpleValueType();
16771 unsigned NumBits = VT.getSizeInBits();
16774 Op = Op.getOperand(0);
16775 if (VT == MVT::i8) {
16776 // Zero extend to i32 since there is not an i8 bsr.
16778 Op = DAG.getNode(ISD::ZERO_EXTEND, dl, OpVT, Op);
16781 // Issue a bsr (scan bits in reverse).
16782 SDVTList VTs = DAG.getVTList(OpVT, MVT::i32);
16783 Op = DAG.getNode(X86ISD::BSR, dl, VTs, Op);
16785 // And xor with NumBits-1.
16786 Op = DAG.getNode(ISD::XOR, dl, OpVT, Op,
16787 DAG.getConstant(NumBits - 1, dl, OpVT));
16790 Op = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op);
16794 static SDValue LowerCTTZ(SDValue Op, SelectionDAG &DAG) {
16795 MVT VT = Op.getSimpleValueType();
16796 unsigned NumBits = VT.getSizeInBits();
16798 Op = Op.getOperand(0);
16800 // Issue a bsf (scan bits forward) which also sets EFLAGS.
16801 SDVTList VTs = DAG.getVTList(VT, MVT::i32);
16802 Op = DAG.getNode(X86ISD::BSF, dl, VTs, Op);
16804 // If src is zero (i.e. bsf sets ZF), returns NumBits.
16807 DAG.getConstant(NumBits, dl, VT),
16808 DAG.getConstant(X86::COND_E, dl, MVT::i8),
16811 return DAG.getNode(X86ISD::CMOV, dl, VT, Ops);
16814 // Lower256IntArith - Break a 256-bit integer operation into two new 128-bit
16815 // ones, and then concatenate the result back.
16816 static SDValue Lower256IntArith(SDValue Op, SelectionDAG &DAG) {
16817 MVT VT = Op.getSimpleValueType();
16819 assert(VT.is256BitVector() && VT.isInteger() &&
16820 "Unsupported value type for operation");
16822 unsigned NumElems = VT.getVectorNumElements();
16825 // Extract the LHS vectors
16826 SDValue LHS = Op.getOperand(0);
16827 SDValue LHS1 = Extract128BitVector(LHS, 0, DAG, dl);
16828 SDValue LHS2 = Extract128BitVector(LHS, NumElems/2, DAG, dl);
16830 // Extract the RHS vectors
16831 SDValue RHS = Op.getOperand(1);
16832 SDValue RHS1 = Extract128BitVector(RHS, 0, DAG, dl);
16833 SDValue RHS2 = Extract128BitVector(RHS, NumElems/2, DAG, dl);
16835 MVT EltVT = VT.getVectorElementType();
16836 MVT NewVT = MVT::getVectorVT(EltVT, NumElems/2);
16838 return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT,
16839 DAG.getNode(Op.getOpcode(), dl, NewVT, LHS1, RHS1),
16840 DAG.getNode(Op.getOpcode(), dl, NewVT, LHS2, RHS2));
16843 static SDValue LowerADD(SDValue Op, SelectionDAG &DAG) {
16844 if (Op.getValueType() == MVT::i1)
16845 return DAG.getNode(ISD::XOR, SDLoc(Op), Op.getValueType(),
16846 Op.getOperand(0), Op.getOperand(1));
16847 assert(Op.getSimpleValueType().is256BitVector() &&
16848 Op.getSimpleValueType().isInteger() &&
16849 "Only handle AVX 256-bit vector integer operation");
16850 return Lower256IntArith(Op, DAG);
16853 static SDValue LowerSUB(SDValue Op, SelectionDAG &DAG) {
16854 if (Op.getValueType() == MVT::i1)
16855 return DAG.getNode(ISD::XOR, SDLoc(Op), Op.getValueType(),
16856 Op.getOperand(0), Op.getOperand(1));
16857 assert(Op.getSimpleValueType().is256BitVector() &&
16858 Op.getSimpleValueType().isInteger() &&
16859 "Only handle AVX 256-bit vector integer operation");
16860 return Lower256IntArith(Op, DAG);
16863 static SDValue LowerMUL(SDValue Op, const X86Subtarget *Subtarget,
16864 SelectionDAG &DAG) {
16866 MVT VT = Op.getSimpleValueType();
16869 return DAG.getNode(ISD::AND, dl, VT, Op.getOperand(0), Op.getOperand(1));
16871 // Decompose 256-bit ops into smaller 128-bit ops.
16872 if (VT.is256BitVector() && !Subtarget->hasInt256())
16873 return Lower256IntArith(Op, DAG);
16875 SDValue A = Op.getOperand(0);
16876 SDValue B = Op.getOperand(1);
16878 // Lower v16i8/v32i8 mul as promotion to v8i16/v16i16 vector
16879 // pairs, multiply and truncate.
16880 if (VT == MVT::v16i8 || VT == MVT::v32i8) {
16881 if (Subtarget->hasInt256()) {
16882 if (VT == MVT::v32i8) {
16883 MVT SubVT = MVT::getVectorVT(MVT::i8, VT.getVectorNumElements() / 2);
16884 SDValue Lo = DAG.getIntPtrConstant(0, dl);
16885 SDValue Hi = DAG.getIntPtrConstant(VT.getVectorNumElements() / 2, dl);
16886 SDValue ALo = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, SubVT, A, Lo);
16887 SDValue BLo = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, SubVT, B, Lo);
16888 SDValue AHi = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, SubVT, A, Hi);
16889 SDValue BHi = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, SubVT, B, Hi);
16890 return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT,
16891 DAG.getNode(ISD::MUL, dl, SubVT, ALo, BLo),
16892 DAG.getNode(ISD::MUL, dl, SubVT, AHi, BHi));
16895 MVT ExVT = MVT::getVectorVT(MVT::i16, VT.getVectorNumElements());
16896 return DAG.getNode(
16897 ISD::TRUNCATE, dl, VT,
16898 DAG.getNode(ISD::MUL, dl, ExVT,
16899 DAG.getNode(ISD::SIGN_EXTEND, dl, ExVT, A),
16900 DAG.getNode(ISD::SIGN_EXTEND, dl, ExVT, B)));
16903 assert(VT == MVT::v16i8 &&
16904 "Pre-AVX2 support only supports v16i8 multiplication");
16905 MVT ExVT = MVT::v8i16;
16907 // Extract the lo parts and sign extend to i16
16909 if (Subtarget->hasSSE41()) {
16910 ALo = DAG.getNode(X86ISD::VSEXT, dl, ExVT, A);
16911 BLo = DAG.getNode(X86ISD::VSEXT, dl, ExVT, B);
16913 const int ShufMask[] = {-1, 0, -1, 1, -1, 2, -1, 3,
16914 -1, 4, -1, 5, -1, 6, -1, 7};
16915 ALo = DAG.getVectorShuffle(VT, dl, A, A, ShufMask);
16916 BLo = DAG.getVectorShuffle(VT, dl, B, B, ShufMask);
16917 ALo = DAG.getBitcast(ExVT, ALo);
16918 BLo = DAG.getBitcast(ExVT, BLo);
16919 ALo = DAG.getNode(ISD::SRA, dl, ExVT, ALo, DAG.getConstant(8, dl, ExVT));
16920 BLo = DAG.getNode(ISD::SRA, dl, ExVT, BLo, DAG.getConstant(8, dl, ExVT));
16923 // Extract the hi parts and sign extend to i16
16925 if (Subtarget->hasSSE41()) {
16926 const int ShufMask[] = {8, 9, 10, 11, 12, 13, 14, 15,
16927 -1, -1, -1, -1, -1, -1, -1, -1};
16928 AHi = DAG.getVectorShuffle(VT, dl, A, A, ShufMask);
16929 BHi = DAG.getVectorShuffle(VT, dl, B, B, ShufMask);
16930 AHi = DAG.getNode(X86ISD::VSEXT, dl, ExVT, AHi);
16931 BHi = DAG.getNode(X86ISD::VSEXT, dl, ExVT, BHi);
16933 const int ShufMask[] = {-1, 8, -1, 9, -1, 10, -1, 11,
16934 -1, 12, -1, 13, -1, 14, -1, 15};
16935 AHi = DAG.getVectorShuffle(VT, dl, A, A, ShufMask);
16936 BHi = DAG.getVectorShuffle(VT, dl, B, B, ShufMask);
16937 AHi = DAG.getBitcast(ExVT, AHi);
16938 BHi = DAG.getBitcast(ExVT, BHi);
16939 AHi = DAG.getNode(ISD::SRA, dl, ExVT, AHi, DAG.getConstant(8, dl, ExVT));
16940 BHi = DAG.getNode(ISD::SRA, dl, ExVT, BHi, DAG.getConstant(8, dl, ExVT));
16943 // Multiply, mask the lower 8bits of the lo/hi results and pack
16944 SDValue RLo = DAG.getNode(ISD::MUL, dl, ExVT, ALo, BLo);
16945 SDValue RHi = DAG.getNode(ISD::MUL, dl, ExVT, AHi, BHi);
16946 RLo = DAG.getNode(ISD::AND, dl, ExVT, RLo, DAG.getConstant(255, dl, ExVT));
16947 RHi = DAG.getNode(ISD::AND, dl, ExVT, RHi, DAG.getConstant(255, dl, ExVT));
16948 return DAG.getNode(X86ISD::PACKUS, dl, VT, RLo, RHi);
16951 // Lower v4i32 mul as 2x shuffle, 2x pmuludq, 2x shuffle.
16952 if (VT == MVT::v4i32) {
16953 assert(Subtarget->hasSSE2() && !Subtarget->hasSSE41() &&
16954 "Should not custom lower when pmuldq is available!");
16956 // Extract the odd parts.
16957 static const int UnpackMask[] = { 1, -1, 3, -1 };
16958 SDValue Aodds = DAG.getVectorShuffle(VT, dl, A, A, UnpackMask);
16959 SDValue Bodds = DAG.getVectorShuffle(VT, dl, B, B, UnpackMask);
16961 // Multiply the even parts.
16962 SDValue Evens = DAG.getNode(X86ISD::PMULUDQ, dl, MVT::v2i64, A, B);
16963 // Now multiply odd parts.
16964 SDValue Odds = DAG.getNode(X86ISD::PMULUDQ, dl, MVT::v2i64, Aodds, Bodds);
16966 Evens = DAG.getBitcast(VT, Evens);
16967 Odds = DAG.getBitcast(VT, Odds);
16969 // Merge the two vectors back together with a shuffle. This expands into 2
16971 static const int ShufMask[] = { 0, 4, 2, 6 };
16972 return DAG.getVectorShuffle(VT, dl, Evens, Odds, ShufMask);
16975 assert((VT == MVT::v2i64 || VT == MVT::v4i64 || VT == MVT::v8i64) &&
16976 "Only know how to lower V2I64/V4I64/V8I64 multiply");
16978 // Ahi = psrlqi(a, 32);
16979 // Bhi = psrlqi(b, 32);
16981 // AloBlo = pmuludq(a, b);
16982 // AloBhi = pmuludq(a, Bhi);
16983 // AhiBlo = pmuludq(Ahi, b);
16985 // AloBhi = psllqi(AloBhi, 32);
16986 // AhiBlo = psllqi(AhiBlo, 32);
16987 // return AloBlo + AloBhi + AhiBlo;
16989 SDValue Ahi = getTargetVShiftByConstNode(X86ISD::VSRLI, dl, VT, A, 32, DAG);
16990 SDValue Bhi = getTargetVShiftByConstNode(X86ISD::VSRLI, dl, VT, B, 32, DAG);
16992 SDValue AhiBlo = Ahi;
16993 SDValue AloBhi = Bhi;
16994 // Bit cast to 32-bit vectors for MULUDQ
16995 EVT MulVT = (VT == MVT::v2i64) ? MVT::v4i32 :
16996 (VT == MVT::v4i64) ? MVT::v8i32 : MVT::v16i32;
16997 A = DAG.getBitcast(MulVT, A);
16998 B = DAG.getBitcast(MulVT, B);
16999 Ahi = DAG.getBitcast(MulVT, Ahi);
17000 Bhi = DAG.getBitcast(MulVT, Bhi);
17002 SDValue AloBlo = DAG.getNode(X86ISD::PMULUDQ, dl, VT, A, B);
17003 // After shifting right const values the result may be all-zero.
17004 if (!ISD::isBuildVectorAllZeros(Ahi.getNode())) {
17005 AhiBlo = DAG.getNode(X86ISD::PMULUDQ, dl, VT, Ahi, B);
17006 AhiBlo = getTargetVShiftByConstNode(X86ISD::VSHLI, dl, VT, AhiBlo, 32, DAG);
17008 if (!ISD::isBuildVectorAllZeros(Bhi.getNode())) {
17009 AloBhi = DAG.getNode(X86ISD::PMULUDQ, dl, VT, A, Bhi);
17010 AloBhi = getTargetVShiftByConstNode(X86ISD::VSHLI, dl, VT, AloBhi, 32, DAG);
17013 SDValue Res = DAG.getNode(ISD::ADD, dl, VT, AloBlo, AloBhi);
17014 return DAG.getNode(ISD::ADD, dl, VT, Res, AhiBlo);
17017 SDValue X86TargetLowering::LowerWin64_i128OP(SDValue Op, SelectionDAG &DAG) const {
17018 assert(Subtarget->isTargetWin64() && "Unexpected target");
17019 EVT VT = Op.getValueType();
17020 assert(VT.isInteger() && VT.getSizeInBits() == 128 &&
17021 "Unexpected return type for lowering");
17025 switch (Op->getOpcode()) {
17026 default: llvm_unreachable("Unexpected request for libcall!");
17027 case ISD::SDIV: isSigned = true; LC = RTLIB::SDIV_I128; break;
17028 case ISD::UDIV: isSigned = false; LC = RTLIB::UDIV_I128; break;
17029 case ISD::SREM: isSigned = true; LC = RTLIB::SREM_I128; break;
17030 case ISD::UREM: isSigned = false; LC = RTLIB::UREM_I128; break;
17031 case ISD::SDIVREM: isSigned = true; LC = RTLIB::SDIVREM_I128; break;
17032 case ISD::UDIVREM: isSigned = false; LC = RTLIB::UDIVREM_I128; break;
17036 SDValue InChain = DAG.getEntryNode();
17038 TargetLowering::ArgListTy Args;
17039 TargetLowering::ArgListEntry Entry;
17040 for (unsigned i = 0, e = Op->getNumOperands(); i != e; ++i) {
17041 EVT ArgVT = Op->getOperand(i).getValueType();
17042 assert(ArgVT.isInteger() && ArgVT.getSizeInBits() == 128 &&
17043 "Unexpected argument type for lowering");
17044 SDValue StackPtr = DAG.CreateStackTemporary(ArgVT, 16);
17045 Entry.Node = StackPtr;
17046 InChain = DAG.getStore(InChain, dl, Op->getOperand(i), StackPtr, MachinePointerInfo(),
17048 Type *ArgTy = ArgVT.getTypeForEVT(*DAG.getContext());
17049 Entry.Ty = PointerType::get(ArgTy,0);
17050 Entry.isSExt = false;
17051 Entry.isZExt = false;
17052 Args.push_back(Entry);
17055 SDValue Callee = DAG.getExternalSymbol(getLibcallName(LC),
17056 getPointerTy(DAG.getDataLayout()));
17058 TargetLowering::CallLoweringInfo CLI(DAG);
17059 CLI.setDebugLoc(dl).setChain(InChain)
17060 .setCallee(getLibcallCallingConv(LC),
17061 static_cast<EVT>(MVT::v2i64).getTypeForEVT(*DAG.getContext()),
17062 Callee, std::move(Args), 0)
17063 .setInRegister().setSExtResult(isSigned).setZExtResult(!isSigned);
17065 std::pair<SDValue, SDValue> CallInfo = LowerCallTo(CLI);
17066 return DAG.getBitcast(VT, CallInfo.first);
17069 static SDValue LowerMUL_LOHI(SDValue Op, const X86Subtarget *Subtarget,
17070 SelectionDAG &DAG) {
17071 SDValue Op0 = Op.getOperand(0), Op1 = Op.getOperand(1);
17072 EVT VT = Op0.getValueType();
17075 assert((VT == MVT::v4i32 && Subtarget->hasSSE2()) ||
17076 (VT == MVT::v8i32 && Subtarget->hasInt256()));
17078 // PMULxD operations multiply each even value (starting at 0) of LHS with
17079 // the related value of RHS and produce a widen result.
17080 // E.g., PMULUDQ <4 x i32> <a|b|c|d>, <4 x i32> <e|f|g|h>
17081 // => <2 x i64> <ae|cg>
17083 // In other word, to have all the results, we need to perform two PMULxD:
17084 // 1. one with the even values.
17085 // 2. one with the odd values.
17086 // To achieve #2, with need to place the odd values at an even position.
17088 // Place the odd value at an even position (basically, shift all values 1
17089 // step to the left):
17090 const int Mask[] = {1, -1, 3, -1, 5, -1, 7, -1};
17091 // <a|b|c|d> => <b|undef|d|undef>
17092 SDValue Odd0 = DAG.getVectorShuffle(VT, dl, Op0, Op0, Mask);
17093 // <e|f|g|h> => <f|undef|h|undef>
17094 SDValue Odd1 = DAG.getVectorShuffle(VT, dl, Op1, Op1, Mask);
17096 // Emit two multiplies, one for the lower 2 ints and one for the higher 2
17098 MVT MulVT = VT == MVT::v4i32 ? MVT::v2i64 : MVT::v4i64;
17099 bool IsSigned = Op->getOpcode() == ISD::SMUL_LOHI;
17101 (!IsSigned || !Subtarget->hasSSE41()) ? X86ISD::PMULUDQ : X86ISD::PMULDQ;
17102 // PMULUDQ <4 x i32> <a|b|c|d>, <4 x i32> <e|f|g|h>
17103 // => <2 x i64> <ae|cg>
17104 SDValue Mul1 = DAG.getBitcast(VT, DAG.getNode(Opcode, dl, MulVT, Op0, Op1));
17105 // PMULUDQ <4 x i32> <b|undef|d|undef>, <4 x i32> <f|undef|h|undef>
17106 // => <2 x i64> <bf|dh>
17107 SDValue Mul2 = DAG.getBitcast(VT, DAG.getNode(Opcode, dl, MulVT, Odd0, Odd1));
17109 // Shuffle it back into the right order.
17110 SDValue Highs, Lows;
17111 if (VT == MVT::v8i32) {
17112 const int HighMask[] = {1, 9, 3, 11, 5, 13, 7, 15};
17113 Highs = DAG.getVectorShuffle(VT, dl, Mul1, Mul2, HighMask);
17114 const int LowMask[] = {0, 8, 2, 10, 4, 12, 6, 14};
17115 Lows = DAG.getVectorShuffle(VT, dl, Mul1, Mul2, LowMask);
17117 const int HighMask[] = {1, 5, 3, 7};
17118 Highs = DAG.getVectorShuffle(VT, dl, Mul1, Mul2, HighMask);
17119 const int LowMask[] = {0, 4, 2, 6};
17120 Lows = DAG.getVectorShuffle(VT, dl, Mul1, Mul2, LowMask);
17123 // If we have a signed multiply but no PMULDQ fix up the high parts of a
17124 // unsigned multiply.
17125 if (IsSigned && !Subtarget->hasSSE41()) {
17126 SDValue ShAmt = DAG.getConstant(
17128 DAG.getTargetLoweringInfo().getShiftAmountTy(VT, DAG.getDataLayout()));
17129 SDValue T1 = DAG.getNode(ISD::AND, dl, VT,
17130 DAG.getNode(ISD::SRA, dl, VT, Op0, ShAmt), Op1);
17131 SDValue T2 = DAG.getNode(ISD::AND, dl, VT,
17132 DAG.getNode(ISD::SRA, dl, VT, Op1, ShAmt), Op0);
17134 SDValue Fixup = DAG.getNode(ISD::ADD, dl, VT, T1, T2);
17135 Highs = DAG.getNode(ISD::SUB, dl, VT, Highs, Fixup);
17138 // The first result of MUL_LOHI is actually the low value, followed by the
17140 SDValue Ops[] = {Lows, Highs};
17141 return DAG.getMergeValues(Ops, dl);
17144 // Return true if the required (according to Opcode) shift-imm form is natively
17145 // supported by the Subtarget
17146 static bool SupportedVectorShiftWithImm(MVT VT, const X86Subtarget *Subtarget,
17148 if (VT.getScalarSizeInBits() < 16)
17151 if (VT.is512BitVector() &&
17152 (VT.getScalarSizeInBits() > 16 || Subtarget->hasBWI()))
17155 bool LShift = VT.is128BitVector() ||
17156 (VT.is256BitVector() && Subtarget->hasInt256());
17158 bool AShift = LShift && (Subtarget->hasVLX() ||
17159 (VT != MVT::v2i64 && VT != MVT::v4i64));
17160 return (Opcode == ISD::SRA) ? AShift : LShift;
17163 // The shift amount is a variable, but it is the same for all vector lanes.
17164 // These instructions are defined together with shift-immediate.
17166 bool SupportedVectorShiftWithBaseAmnt(MVT VT, const X86Subtarget *Subtarget,
17168 return SupportedVectorShiftWithImm(VT, Subtarget, Opcode);
17171 // Return true if the required (according to Opcode) variable-shift form is
17172 // natively supported by the Subtarget
17173 static bool SupportedVectorVarShift(MVT VT, const X86Subtarget *Subtarget,
17176 if (!Subtarget->hasInt256() || VT.getScalarSizeInBits() < 16)
17179 // vXi16 supported only on AVX-512, BWI
17180 if (VT.getScalarSizeInBits() == 16 && !Subtarget->hasBWI())
17183 if (VT.is512BitVector() || Subtarget->hasVLX())
17186 bool LShift = VT.is128BitVector() || VT.is256BitVector();
17187 bool AShift = LShift && VT != MVT::v2i64 && VT != MVT::v4i64;
17188 return (Opcode == ISD::SRA) ? AShift : LShift;
17191 static SDValue LowerScalarImmediateShift(SDValue Op, SelectionDAG &DAG,
17192 const X86Subtarget *Subtarget) {
17193 MVT VT = Op.getSimpleValueType();
17195 SDValue R = Op.getOperand(0);
17196 SDValue Amt = Op.getOperand(1);
17198 unsigned X86Opc = (Op.getOpcode() == ISD::SHL) ? X86ISD::VSHLI :
17199 (Op.getOpcode() == ISD::SRL) ? X86ISD::VSRLI : X86ISD::VSRAI;
17201 auto ArithmeticShiftRight64 = [&](uint64_t ShiftAmt) {
17202 assert((VT == MVT::v2i64 || VT == MVT::v4i64) && "Unexpected SRA type");
17203 MVT ExVT = MVT::getVectorVT(MVT::i32, VT.getVectorNumElements() * 2);
17204 SDValue Ex = DAG.getBitcast(ExVT, R);
17206 if (ShiftAmt >= 32) {
17207 // Splat sign to upper i32 dst, and SRA upper i32 src to lower i32.
17209 getTargetVShiftByConstNode(X86ISD::VSRAI, dl, ExVT, Ex, 31, DAG);
17210 SDValue Lower = getTargetVShiftByConstNode(X86ISD::VSRAI, dl, ExVT, Ex,
17211 ShiftAmt - 32, DAG);
17212 if (VT == MVT::v2i64)
17213 Ex = DAG.getVectorShuffle(ExVT, dl, Upper, Lower, {5, 1, 7, 3});
17214 if (VT == MVT::v4i64)
17215 Ex = DAG.getVectorShuffle(ExVT, dl, Upper, Lower,
17216 {9, 1, 11, 3, 13, 5, 15, 7});
17218 // SRA upper i32, SHL whole i64 and select lower i32.
17219 SDValue Upper = getTargetVShiftByConstNode(X86ISD::VSRAI, dl, ExVT, Ex,
17222 getTargetVShiftByConstNode(X86ISD::VSRLI, dl, VT, R, ShiftAmt, DAG);
17223 Lower = DAG.getBitcast(ExVT, Lower);
17224 if (VT == MVT::v2i64)
17225 Ex = DAG.getVectorShuffle(ExVT, dl, Upper, Lower, {4, 1, 6, 3});
17226 if (VT == MVT::v4i64)
17227 Ex = DAG.getVectorShuffle(ExVT, dl, Upper, Lower,
17228 {8, 1, 10, 3, 12, 5, 14, 7});
17230 return DAG.getBitcast(VT, Ex);
17233 // Optimize shl/srl/sra with constant shift amount.
17234 if (auto *BVAmt = dyn_cast<BuildVectorSDNode>(Amt)) {
17235 if (auto *ShiftConst = BVAmt->getConstantSplatNode()) {
17236 uint64_t ShiftAmt = ShiftConst->getZExtValue();
17238 if (SupportedVectorShiftWithImm(VT, Subtarget, Op.getOpcode()))
17239 return getTargetVShiftByConstNode(X86Opc, dl, VT, R, ShiftAmt, DAG);
17241 // i64 SRA needs to be performed as partial shifts.
17242 if ((VT == MVT::v2i64 || (Subtarget->hasInt256() && VT == MVT::v4i64)) &&
17243 Op.getOpcode() == ISD::SRA)
17244 return ArithmeticShiftRight64(ShiftAmt);
17246 if (VT == MVT::v16i8 || (Subtarget->hasInt256() && VT == MVT::v32i8)) {
17247 unsigned NumElts = VT.getVectorNumElements();
17248 MVT ShiftVT = MVT::getVectorVT(MVT::i16, NumElts / 2);
17250 if (Op.getOpcode() == ISD::SHL) {
17251 // Simple i8 add case
17253 return DAG.getNode(ISD::ADD, dl, VT, R, R);
17255 // Make a large shift.
17256 SDValue SHL = getTargetVShiftByConstNode(X86ISD::VSHLI, dl, ShiftVT,
17258 SHL = DAG.getBitcast(VT, SHL);
17259 // Zero out the rightmost bits.
17260 SmallVector<SDValue, 32> V(
17261 NumElts, DAG.getConstant(uint8_t(-1U << ShiftAmt), dl, MVT::i8));
17262 return DAG.getNode(ISD::AND, dl, VT, SHL,
17263 DAG.getNode(ISD::BUILD_VECTOR, dl, VT, V));
17265 if (Op.getOpcode() == ISD::SRL) {
17266 // Make a large shift.
17267 SDValue SRL = getTargetVShiftByConstNode(X86ISD::VSRLI, dl, ShiftVT,
17269 SRL = DAG.getBitcast(VT, SRL);
17270 // Zero out the leftmost bits.
17271 SmallVector<SDValue, 32> V(
17272 NumElts, DAG.getConstant(uint8_t(-1U) >> ShiftAmt, dl, MVT::i8));
17273 return DAG.getNode(ISD::AND, dl, VT, SRL,
17274 DAG.getNode(ISD::BUILD_VECTOR, dl, VT, V));
17276 if (Op.getOpcode() == ISD::SRA) {
17277 if (ShiftAmt == 7) {
17278 // R s>> 7 === R s< 0
17279 SDValue Zeros = getZeroVector(VT, Subtarget, DAG, dl);
17280 return DAG.getNode(X86ISD::PCMPGT, dl, VT, Zeros, R);
17283 // R s>> a === ((R u>> a) ^ m) - m
17284 SDValue Res = DAG.getNode(ISD::SRL, dl, VT, R, Amt);
17285 SmallVector<SDValue, 32> V(NumElts,
17286 DAG.getConstant(128 >> ShiftAmt, dl,
17288 SDValue Mask = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, V);
17289 Res = DAG.getNode(ISD::XOR, dl, VT, Res, Mask);
17290 Res = DAG.getNode(ISD::SUB, dl, VT, Res, Mask);
17293 llvm_unreachable("Unknown shift opcode.");
17298 // Special case in 32-bit mode, where i64 is expanded into high and low parts.
17299 if (!Subtarget->is64Bit() &&
17300 (VT == MVT::v2i64 || (Subtarget->hasInt256() && VT == MVT::v4i64)) &&
17301 Amt.getOpcode() == ISD::BITCAST &&
17302 Amt.getOperand(0).getOpcode() == ISD::BUILD_VECTOR) {
17303 Amt = Amt.getOperand(0);
17304 unsigned Ratio = Amt.getSimpleValueType().getVectorNumElements() /
17305 VT.getVectorNumElements();
17306 unsigned RatioInLog2 = Log2_32_Ceil(Ratio);
17307 uint64_t ShiftAmt = 0;
17308 for (unsigned i = 0; i != Ratio; ++i) {
17309 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Amt.getOperand(i));
17313 ShiftAmt |= C->getZExtValue() << (i * (1 << (6 - RatioInLog2)));
17315 // Check remaining shift amounts.
17316 for (unsigned i = Ratio; i != Amt.getNumOperands(); i += Ratio) {
17317 uint64_t ShAmt = 0;
17318 for (unsigned j = 0; j != Ratio; ++j) {
17319 ConstantSDNode *C =
17320 dyn_cast<ConstantSDNode>(Amt.getOperand(i + j));
17324 ShAmt |= C->getZExtValue() << (j * (1 << (6 - RatioInLog2)));
17326 if (ShAmt != ShiftAmt)
17330 if (SupportedVectorShiftWithImm(VT, Subtarget, Op.getOpcode()))
17331 return getTargetVShiftByConstNode(X86Opc, dl, VT, R, ShiftAmt, DAG);
17333 if (Op.getOpcode() == ISD::SRA)
17334 return ArithmeticShiftRight64(ShiftAmt);
17340 static SDValue LowerScalarVariableShift(SDValue Op, SelectionDAG &DAG,
17341 const X86Subtarget* Subtarget) {
17342 MVT VT = Op.getSimpleValueType();
17344 SDValue R = Op.getOperand(0);
17345 SDValue Amt = Op.getOperand(1);
17347 unsigned X86OpcI = (Op.getOpcode() == ISD::SHL) ? X86ISD::VSHLI :
17348 (Op.getOpcode() == ISD::SRL) ? X86ISD::VSRLI : X86ISD::VSRAI;
17350 unsigned X86OpcV = (Op.getOpcode() == ISD::SHL) ? X86ISD::VSHL :
17351 (Op.getOpcode() == ISD::SRL) ? X86ISD::VSRL : X86ISD::VSRA;
17353 if (SupportedVectorShiftWithBaseAmnt(VT, Subtarget, Op.getOpcode())) {
17355 EVT EltVT = VT.getVectorElementType();
17357 if (BuildVectorSDNode *BV = dyn_cast<BuildVectorSDNode>(Amt)) {
17358 // Check if this build_vector node is doing a splat.
17359 // If so, then set BaseShAmt equal to the splat value.
17360 BaseShAmt = BV->getSplatValue();
17361 if (BaseShAmt && BaseShAmt.getOpcode() == ISD::UNDEF)
17362 BaseShAmt = SDValue();
17364 if (Amt.getOpcode() == ISD::EXTRACT_SUBVECTOR)
17365 Amt = Amt.getOperand(0);
17367 ShuffleVectorSDNode *SVN = dyn_cast<ShuffleVectorSDNode>(Amt);
17368 if (SVN && SVN->isSplat()) {
17369 unsigned SplatIdx = (unsigned)SVN->getSplatIndex();
17370 SDValue InVec = Amt.getOperand(0);
17371 if (InVec.getOpcode() == ISD::BUILD_VECTOR) {
17372 assert((SplatIdx < InVec.getValueType().getVectorNumElements()) &&
17373 "Unexpected shuffle index found!");
17374 BaseShAmt = InVec.getOperand(SplatIdx);
17375 } else if (InVec.getOpcode() == ISD::INSERT_VECTOR_ELT) {
17376 if (ConstantSDNode *C =
17377 dyn_cast<ConstantSDNode>(InVec.getOperand(2))) {
17378 if (C->getZExtValue() == SplatIdx)
17379 BaseShAmt = InVec.getOperand(1);
17384 // Avoid introducing an extract element from a shuffle.
17385 BaseShAmt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, EltVT, InVec,
17386 DAG.getIntPtrConstant(SplatIdx, dl));
17390 if (BaseShAmt.getNode()) {
17391 assert(EltVT.bitsLE(MVT::i64) && "Unexpected element type!");
17392 if (EltVT != MVT::i64 && EltVT.bitsGT(MVT::i32))
17393 BaseShAmt = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i64, BaseShAmt);
17394 else if (EltVT.bitsLT(MVT::i32))
17395 BaseShAmt = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, BaseShAmt);
17397 return getTargetVShiftNode(X86OpcI, dl, VT, R, BaseShAmt, DAG);
17401 // Special case in 32-bit mode, where i64 is expanded into high and low parts.
17402 if (!Subtarget->is64Bit() && VT == MVT::v2i64 &&
17403 Amt.getOpcode() == ISD::BITCAST &&
17404 Amt.getOperand(0).getOpcode() == ISD::BUILD_VECTOR) {
17405 Amt = Amt.getOperand(0);
17406 unsigned Ratio = Amt.getSimpleValueType().getVectorNumElements() /
17407 VT.getVectorNumElements();
17408 std::vector<SDValue> Vals(Ratio);
17409 for (unsigned i = 0; i != Ratio; ++i)
17410 Vals[i] = Amt.getOperand(i);
17411 for (unsigned i = Ratio; i != Amt.getNumOperands(); i += Ratio) {
17412 for (unsigned j = 0; j != Ratio; ++j)
17413 if (Vals[j] != Amt.getOperand(i + j))
17417 if (SupportedVectorShiftWithBaseAmnt(VT, Subtarget, Op.getOpcode()))
17418 return DAG.getNode(X86OpcV, dl, VT, R, Op.getOperand(1));
17423 static SDValue LowerShift(SDValue Op, const X86Subtarget* Subtarget,
17424 SelectionDAG &DAG) {
17425 MVT VT = Op.getSimpleValueType();
17427 SDValue R = Op.getOperand(0);
17428 SDValue Amt = Op.getOperand(1);
17430 assert(VT.isVector() && "Custom lowering only for vector shifts!");
17431 assert(Subtarget->hasSSE2() && "Only custom lower when we have SSE2!");
17433 if (SDValue V = LowerScalarImmediateShift(Op, DAG, Subtarget))
17436 if (SDValue V = LowerScalarVariableShift(Op, DAG, Subtarget))
17439 if (SupportedVectorVarShift(VT, Subtarget, Op.getOpcode()))
17442 // 2i64 vector logical shifts can efficiently avoid scalarization - do the
17443 // shifts per-lane and then shuffle the partial results back together.
17444 if (VT == MVT::v2i64 && Op.getOpcode() != ISD::SRA) {
17445 // Splat the shift amounts so the scalar shifts above will catch it.
17446 SDValue Amt0 = DAG.getVectorShuffle(VT, dl, Amt, Amt, {0, 0});
17447 SDValue Amt1 = DAG.getVectorShuffle(VT, dl, Amt, Amt, {1, 1});
17448 SDValue R0 = DAG.getNode(Op->getOpcode(), dl, VT, R, Amt0);
17449 SDValue R1 = DAG.getNode(Op->getOpcode(), dl, VT, R, Amt1);
17450 return DAG.getVectorShuffle(VT, dl, R0, R1, {0, 3});
17453 // If possible, lower this packed shift into a vector multiply instead of
17454 // expanding it into a sequence of scalar shifts.
17455 // Do this only if the vector shift count is a constant build_vector.
17456 if (Op.getOpcode() == ISD::SHL &&
17457 (VT == MVT::v8i16 || VT == MVT::v4i32 ||
17458 (Subtarget->hasInt256() && VT == MVT::v16i16)) &&
17459 ISD::isBuildVectorOfConstantSDNodes(Amt.getNode())) {
17460 SmallVector<SDValue, 8> Elts;
17461 EVT SVT = VT.getScalarType();
17462 unsigned SVTBits = SVT.getSizeInBits();
17463 const APInt &One = APInt(SVTBits, 1);
17464 unsigned NumElems = VT.getVectorNumElements();
17466 for (unsigned i=0; i !=NumElems; ++i) {
17467 SDValue Op = Amt->getOperand(i);
17468 if (Op->getOpcode() == ISD::UNDEF) {
17469 Elts.push_back(Op);
17473 ConstantSDNode *ND = cast<ConstantSDNode>(Op);
17474 const APInt &C = APInt(SVTBits, ND->getAPIntValue().getZExtValue());
17475 uint64_t ShAmt = C.getZExtValue();
17476 if (ShAmt >= SVTBits) {
17477 Elts.push_back(DAG.getUNDEF(SVT));
17480 Elts.push_back(DAG.getConstant(One.shl(ShAmt), dl, SVT));
17482 SDValue BV = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Elts);
17483 return DAG.getNode(ISD::MUL, dl, VT, R, BV);
17486 // Lower SHL with variable shift amount.
17487 if (VT == MVT::v4i32 && Op->getOpcode() == ISD::SHL) {
17488 Op = DAG.getNode(ISD::SHL, dl, VT, Amt, DAG.getConstant(23, dl, VT));
17490 Op = DAG.getNode(ISD::ADD, dl, VT, Op,
17491 DAG.getConstant(0x3f800000U, dl, VT));
17492 Op = DAG.getBitcast(MVT::v4f32, Op);
17493 Op = DAG.getNode(ISD::FP_TO_SINT, dl, VT, Op);
17494 return DAG.getNode(ISD::MUL, dl, VT, Op, R);
17497 // If possible, lower this shift as a sequence of two shifts by
17498 // constant plus a MOVSS/MOVSD instead of scalarizing it.
17500 // (v4i32 (srl A, (build_vector < X, Y, Y, Y>)))
17502 // Could be rewritten as:
17503 // (v4i32 (MOVSS (srl A, <Y,Y,Y,Y>), (srl A, <X,X,X,X>)))
17505 // The advantage is that the two shifts from the example would be
17506 // lowered as X86ISD::VSRLI nodes. This would be cheaper than scalarizing
17507 // the vector shift into four scalar shifts plus four pairs of vector
17509 if ((VT == MVT::v8i16 || VT == MVT::v4i32) &&
17510 ISD::isBuildVectorOfConstantSDNodes(Amt.getNode())) {
17511 unsigned TargetOpcode = X86ISD::MOVSS;
17512 bool CanBeSimplified;
17513 // The splat value for the first packed shift (the 'X' from the example).
17514 SDValue Amt1 = Amt->getOperand(0);
17515 // The splat value for the second packed shift (the 'Y' from the example).
17516 SDValue Amt2 = (VT == MVT::v4i32) ? Amt->getOperand(1) :
17517 Amt->getOperand(2);
17519 // See if it is possible to replace this node with a sequence of
17520 // two shifts followed by a MOVSS/MOVSD
17521 if (VT == MVT::v4i32) {
17522 // Check if it is legal to use a MOVSS.
17523 CanBeSimplified = Amt2 == Amt->getOperand(2) &&
17524 Amt2 == Amt->getOperand(3);
17525 if (!CanBeSimplified) {
17526 // Otherwise, check if we can still simplify this node using a MOVSD.
17527 CanBeSimplified = Amt1 == Amt->getOperand(1) &&
17528 Amt->getOperand(2) == Amt->getOperand(3);
17529 TargetOpcode = X86ISD::MOVSD;
17530 Amt2 = Amt->getOperand(2);
17533 // Do similar checks for the case where the machine value type
17535 CanBeSimplified = Amt1 == Amt->getOperand(1);
17536 for (unsigned i=3; i != 8 && CanBeSimplified; ++i)
17537 CanBeSimplified = Amt2 == Amt->getOperand(i);
17539 if (!CanBeSimplified) {
17540 TargetOpcode = X86ISD::MOVSD;
17541 CanBeSimplified = true;
17542 Amt2 = Amt->getOperand(4);
17543 for (unsigned i=0; i != 4 && CanBeSimplified; ++i)
17544 CanBeSimplified = Amt1 == Amt->getOperand(i);
17545 for (unsigned j=4; j != 8 && CanBeSimplified; ++j)
17546 CanBeSimplified = Amt2 == Amt->getOperand(j);
17550 if (CanBeSimplified && isa<ConstantSDNode>(Amt1) &&
17551 isa<ConstantSDNode>(Amt2)) {
17552 // Replace this node with two shifts followed by a MOVSS/MOVSD.
17553 EVT CastVT = MVT::v4i32;
17555 DAG.getConstant(cast<ConstantSDNode>(Amt1)->getAPIntValue(), dl, VT);
17556 SDValue Shift1 = DAG.getNode(Op->getOpcode(), dl, VT, R, Splat1);
17558 DAG.getConstant(cast<ConstantSDNode>(Amt2)->getAPIntValue(), dl, VT);
17559 SDValue Shift2 = DAG.getNode(Op->getOpcode(), dl, VT, R, Splat2);
17560 if (TargetOpcode == X86ISD::MOVSD)
17561 CastVT = MVT::v2i64;
17562 SDValue BitCast1 = DAG.getBitcast(CastVT, Shift1);
17563 SDValue BitCast2 = DAG.getBitcast(CastVT, Shift2);
17564 SDValue Result = getTargetShuffleNode(TargetOpcode, dl, CastVT, BitCast2,
17566 return DAG.getBitcast(VT, Result);
17570 // v4i32 Non Uniform Shifts.
17571 // If the shift amount is constant we can shift each lane using the SSE2
17572 // immediate shifts, else we need to zero-extend each lane to the lower i64
17573 // and shift using the SSE2 variable shifts.
17574 // The separate results can then be blended together.
17575 if (VT == MVT::v4i32) {
17576 unsigned Opc = Op.getOpcode();
17577 SDValue Amt0, Amt1, Amt2, Amt3;
17578 if (ISD::isBuildVectorOfConstantSDNodes(Amt.getNode())) {
17579 Amt0 = DAG.getVectorShuffle(VT, dl, Amt, DAG.getUNDEF(VT), {0, 0, 0, 0});
17580 Amt1 = DAG.getVectorShuffle(VT, dl, Amt, DAG.getUNDEF(VT), {1, 1, 1, 1});
17581 Amt2 = DAG.getVectorShuffle(VT, dl, Amt, DAG.getUNDEF(VT), {2, 2, 2, 2});
17582 Amt3 = DAG.getVectorShuffle(VT, dl, Amt, DAG.getUNDEF(VT), {3, 3, 3, 3});
17584 // ISD::SHL is handled above but we include it here for completeness.
17587 llvm_unreachable("Unknown target vector shift node");
17589 Opc = X86ISD::VSHL;
17592 Opc = X86ISD::VSRL;
17595 Opc = X86ISD::VSRA;
17598 // The SSE2 shifts use the lower i64 as the same shift amount for
17599 // all lanes and the upper i64 is ignored. These shuffle masks
17600 // optimally zero-extend each lanes on SSE2/SSE41/AVX targets.
17601 SDValue Z = getZeroVector(VT, Subtarget, DAG, dl);
17602 Amt0 = DAG.getVectorShuffle(VT, dl, Amt, Z, {0, 4, -1, -1});
17603 Amt1 = DAG.getVectorShuffle(VT, dl, Amt, Z, {1, 5, -1, -1});
17604 Amt2 = DAG.getVectorShuffle(VT, dl, Amt, Z, {2, 6, -1, -1});
17605 Amt3 = DAG.getVectorShuffle(VT, dl, Amt, Z, {3, 7, -1, -1});
17608 SDValue R0 = DAG.getNode(Opc, dl, VT, R, Amt0);
17609 SDValue R1 = DAG.getNode(Opc, dl, VT, R, Amt1);
17610 SDValue R2 = DAG.getNode(Opc, dl, VT, R, Amt2);
17611 SDValue R3 = DAG.getNode(Opc, dl, VT, R, Amt3);
17612 SDValue R02 = DAG.getVectorShuffle(VT, dl, R0, R2, {0, -1, 6, -1});
17613 SDValue R13 = DAG.getVectorShuffle(VT, dl, R1, R3, {-1, 1, -1, 7});
17614 return DAG.getVectorShuffle(VT, dl, R02, R13, {0, 5, 2, 7});
17617 if (VT == MVT::v16i8 || (VT == MVT::v32i8 && Subtarget->hasInt256())) {
17618 MVT ExtVT = MVT::getVectorVT(MVT::i16, VT.getVectorNumElements() / 2);
17619 unsigned ShiftOpcode = Op->getOpcode();
17621 auto SignBitSelect = [&](MVT SelVT, SDValue Sel, SDValue V0, SDValue V1) {
17622 // On SSE41 targets we make use of the fact that VSELECT lowers
17623 // to PBLENDVB which selects bytes based just on the sign bit.
17624 if (Subtarget->hasSSE41()) {
17625 V0 = DAG.getBitcast(VT, V0);
17626 V1 = DAG.getBitcast(VT, V1);
17627 Sel = DAG.getBitcast(VT, Sel);
17628 return DAG.getBitcast(SelVT,
17629 DAG.getNode(ISD::VSELECT, dl, VT, Sel, V0, V1));
17631 // On pre-SSE41 targets we test for the sign bit by comparing to
17632 // zero - a negative value will set all bits of the lanes to true
17633 // and VSELECT uses that in its OR(AND(V0,C),AND(V1,~C)) lowering.
17634 SDValue Z = getZeroVector(SelVT, Subtarget, DAG, dl);
17635 SDValue C = DAG.getNode(X86ISD::PCMPGT, dl, SelVT, Z, Sel);
17636 return DAG.getNode(ISD::VSELECT, dl, SelVT, C, V0, V1);
17639 // Turn 'a' into a mask suitable for VSELECT: a = a << 5;
17640 // We can safely do this using i16 shifts as we're only interested in
17641 // the 3 lower bits of each byte.
17642 Amt = DAG.getBitcast(ExtVT, Amt);
17643 Amt = DAG.getNode(ISD::SHL, dl, ExtVT, Amt, DAG.getConstant(5, dl, ExtVT));
17644 Amt = DAG.getBitcast(VT, Amt);
17646 if (Op->getOpcode() == ISD::SHL || Op->getOpcode() == ISD::SRL) {
17647 // r = VSELECT(r, shift(r, 4), a);
17649 DAG.getNode(ShiftOpcode, dl, VT, R, DAG.getConstant(4, dl, VT));
17650 R = SignBitSelect(VT, Amt, M, R);
17653 Amt = DAG.getNode(ISD::ADD, dl, VT, Amt, Amt);
17655 // r = VSELECT(r, shift(r, 2), a);
17656 M = DAG.getNode(ShiftOpcode, dl, VT, R, DAG.getConstant(2, dl, VT));
17657 R = SignBitSelect(VT, Amt, M, R);
17660 Amt = DAG.getNode(ISD::ADD, dl, VT, Amt, Amt);
17662 // return VSELECT(r, shift(r, 1), a);
17663 M = DAG.getNode(ShiftOpcode, dl, VT, R, DAG.getConstant(1, dl, VT));
17664 R = SignBitSelect(VT, Amt, M, R);
17668 if (Op->getOpcode() == ISD::SRA) {
17669 // For SRA we need to unpack each byte to the higher byte of a i16 vector
17670 // so we can correctly sign extend. We don't care what happens to the
17672 SDValue ALo = DAG.getNode(X86ISD::UNPCKL, dl, VT, DAG.getUNDEF(VT), Amt);
17673 SDValue AHi = DAG.getNode(X86ISD::UNPCKH, dl, VT, DAG.getUNDEF(VT), Amt);
17674 SDValue RLo = DAG.getNode(X86ISD::UNPCKL, dl, VT, DAG.getUNDEF(VT), R);
17675 SDValue RHi = DAG.getNode(X86ISD::UNPCKH, dl, VT, DAG.getUNDEF(VT), R);
17676 ALo = DAG.getBitcast(ExtVT, ALo);
17677 AHi = DAG.getBitcast(ExtVT, AHi);
17678 RLo = DAG.getBitcast(ExtVT, RLo);
17679 RHi = DAG.getBitcast(ExtVT, RHi);
17681 // r = VSELECT(r, shift(r, 4), a);
17682 SDValue MLo = DAG.getNode(ShiftOpcode, dl, ExtVT, RLo,
17683 DAG.getConstant(4, dl, ExtVT));
17684 SDValue MHi = DAG.getNode(ShiftOpcode, dl, ExtVT, RHi,
17685 DAG.getConstant(4, dl, ExtVT));
17686 RLo = SignBitSelect(ExtVT, ALo, MLo, RLo);
17687 RHi = SignBitSelect(ExtVT, AHi, MHi, RHi);
17690 ALo = DAG.getNode(ISD::ADD, dl, ExtVT, ALo, ALo);
17691 AHi = DAG.getNode(ISD::ADD, dl, ExtVT, AHi, AHi);
17693 // r = VSELECT(r, shift(r, 2), a);
17694 MLo = DAG.getNode(ShiftOpcode, dl, ExtVT, RLo,
17695 DAG.getConstant(2, dl, ExtVT));
17696 MHi = DAG.getNode(ShiftOpcode, dl, ExtVT, RHi,
17697 DAG.getConstant(2, dl, ExtVT));
17698 RLo = SignBitSelect(ExtVT, ALo, MLo, RLo);
17699 RHi = SignBitSelect(ExtVT, AHi, MHi, RHi);
17702 ALo = DAG.getNode(ISD::ADD, dl, ExtVT, ALo, ALo);
17703 AHi = DAG.getNode(ISD::ADD, dl, ExtVT, AHi, AHi);
17705 // r = VSELECT(r, shift(r, 1), a);
17706 MLo = DAG.getNode(ShiftOpcode, dl, ExtVT, RLo,
17707 DAG.getConstant(1, dl, ExtVT));
17708 MHi = DAG.getNode(ShiftOpcode, dl, ExtVT, RHi,
17709 DAG.getConstant(1, dl, ExtVT));
17710 RLo = SignBitSelect(ExtVT, ALo, MLo, RLo);
17711 RHi = SignBitSelect(ExtVT, AHi, MHi, RHi);
17713 // Logical shift the result back to the lower byte, leaving a zero upper
17715 // meaning that we can safely pack with PACKUSWB.
17717 DAG.getNode(ISD::SRL, dl, ExtVT, RLo, DAG.getConstant(8, dl, ExtVT));
17719 DAG.getNode(ISD::SRL, dl, ExtVT, RHi, DAG.getConstant(8, dl, ExtVT));
17720 return DAG.getNode(X86ISD::PACKUS, dl, VT, RLo, RHi);
17724 // It's worth extending once and using the v8i32 shifts for 16-bit types, but
17725 // the extra overheads to get from v16i8 to v8i32 make the existing SSE
17726 // solution better.
17727 if (Subtarget->hasInt256() && VT == MVT::v8i16) {
17728 MVT ExtVT = MVT::v8i32;
17730 Op.getOpcode() == ISD::SRA ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND;
17731 R = DAG.getNode(ExtOpc, dl, ExtVT, R);
17732 Amt = DAG.getNode(ISD::ANY_EXTEND, dl, ExtVT, Amt);
17733 return DAG.getNode(ISD::TRUNCATE, dl, VT,
17734 DAG.getNode(Op.getOpcode(), dl, ExtVT, R, Amt));
17737 if (Subtarget->hasInt256() && VT == MVT::v16i16) {
17738 MVT ExtVT = MVT::v8i32;
17739 SDValue Z = getZeroVector(VT, Subtarget, DAG, dl);
17740 SDValue ALo = DAG.getNode(X86ISD::UNPCKL, dl, VT, Amt, Z);
17741 SDValue AHi = DAG.getNode(X86ISD::UNPCKH, dl, VT, Amt, Z);
17742 SDValue RLo = DAG.getNode(X86ISD::UNPCKL, dl, VT, R, R);
17743 SDValue RHi = DAG.getNode(X86ISD::UNPCKH, dl, VT, R, R);
17744 ALo = DAG.getBitcast(ExtVT, ALo);
17745 AHi = DAG.getBitcast(ExtVT, AHi);
17746 RLo = DAG.getBitcast(ExtVT, RLo);
17747 RHi = DAG.getBitcast(ExtVT, RHi);
17748 SDValue Lo = DAG.getNode(Op.getOpcode(), dl, ExtVT, RLo, ALo);
17749 SDValue Hi = DAG.getNode(Op.getOpcode(), dl, ExtVT, RHi, AHi);
17750 Lo = DAG.getNode(ISD::SRL, dl, ExtVT, Lo, DAG.getConstant(16, dl, ExtVT));
17751 Hi = DAG.getNode(ISD::SRL, dl, ExtVT, Hi, DAG.getConstant(16, dl, ExtVT));
17752 return DAG.getNode(X86ISD::PACKUS, dl, VT, Lo, Hi);
17755 if (VT == MVT::v8i16) {
17756 unsigned ShiftOpcode = Op->getOpcode();
17758 auto SignBitSelect = [&](SDValue Sel, SDValue V0, SDValue V1) {
17759 // On SSE41 targets we make use of the fact that VSELECT lowers
17760 // to PBLENDVB which selects bytes based just on the sign bit.
17761 if (Subtarget->hasSSE41()) {
17762 MVT ExtVT = MVT::getVectorVT(MVT::i8, VT.getVectorNumElements() * 2);
17763 V0 = DAG.getBitcast(ExtVT, V0);
17764 V1 = DAG.getBitcast(ExtVT, V1);
17765 Sel = DAG.getBitcast(ExtVT, Sel);
17766 return DAG.getBitcast(
17767 VT, DAG.getNode(ISD::VSELECT, dl, ExtVT, Sel, V0, V1));
17769 // On pre-SSE41 targets we splat the sign bit - a negative value will
17770 // set all bits of the lanes to true and VSELECT uses that in
17771 // its OR(AND(V0,C),AND(V1,~C)) lowering.
17773 DAG.getNode(ISD::SRA, dl, VT, Sel, DAG.getConstant(15, dl, VT));
17774 return DAG.getNode(ISD::VSELECT, dl, VT, C, V0, V1);
17777 // Turn 'a' into a mask suitable for VSELECT: a = a << 12;
17778 if (Subtarget->hasSSE41()) {
17779 // On SSE41 targets we need to replicate the shift mask in both
17780 // bytes for PBLENDVB.
17783 DAG.getNode(ISD::SHL, dl, VT, Amt, DAG.getConstant(4, dl, VT)),
17784 DAG.getNode(ISD::SHL, dl, VT, Amt, DAG.getConstant(12, dl, VT)));
17786 Amt = DAG.getNode(ISD::SHL, dl, VT, Amt, DAG.getConstant(12, dl, VT));
17789 // r = VSELECT(r, shift(r, 8), a);
17790 SDValue M = DAG.getNode(ShiftOpcode, dl, VT, R, DAG.getConstant(8, dl, VT));
17791 R = SignBitSelect(Amt, M, R);
17794 Amt = DAG.getNode(ISD::ADD, dl, VT, Amt, Amt);
17796 // r = VSELECT(r, shift(r, 4), a);
17797 M = DAG.getNode(ShiftOpcode, dl, VT, R, DAG.getConstant(4, dl, VT));
17798 R = SignBitSelect(Amt, M, R);
17801 Amt = DAG.getNode(ISD::ADD, dl, VT, Amt, Amt);
17803 // r = VSELECT(r, shift(r, 2), a);
17804 M = DAG.getNode(ShiftOpcode, dl, VT, R, DAG.getConstant(2, dl, VT));
17805 R = SignBitSelect(Amt, M, R);
17808 Amt = DAG.getNode(ISD::ADD, dl, VT, Amt, Amt);
17810 // return VSELECT(r, shift(r, 1), a);
17811 M = DAG.getNode(ShiftOpcode, dl, VT, R, DAG.getConstant(1, dl, VT));
17812 R = SignBitSelect(Amt, M, R);
17816 // Decompose 256-bit shifts into smaller 128-bit shifts.
17817 if (VT.is256BitVector()) {
17818 unsigned NumElems = VT.getVectorNumElements();
17819 MVT EltVT = VT.getVectorElementType();
17820 EVT NewVT = MVT::getVectorVT(EltVT, NumElems/2);
17822 // Extract the two vectors
17823 SDValue V1 = Extract128BitVector(R, 0, DAG, dl);
17824 SDValue V2 = Extract128BitVector(R, NumElems/2, DAG, dl);
17826 // Recreate the shift amount vectors
17827 SDValue Amt1, Amt2;
17828 if (Amt.getOpcode() == ISD::BUILD_VECTOR) {
17829 // Constant shift amount
17830 SmallVector<SDValue, 8> Ops(Amt->op_begin(), Amt->op_begin() + NumElems);
17831 ArrayRef<SDValue> Amt1Csts = makeArrayRef(Ops).slice(0, NumElems / 2);
17832 ArrayRef<SDValue> Amt2Csts = makeArrayRef(Ops).slice(NumElems / 2);
17834 Amt1 = DAG.getNode(ISD::BUILD_VECTOR, dl, NewVT, Amt1Csts);
17835 Amt2 = DAG.getNode(ISD::BUILD_VECTOR, dl, NewVT, Amt2Csts);
17837 // Variable shift amount
17838 Amt1 = Extract128BitVector(Amt, 0, DAG, dl);
17839 Amt2 = Extract128BitVector(Amt, NumElems/2, DAG, dl);
17842 // Issue new vector shifts for the smaller types
17843 V1 = DAG.getNode(Op.getOpcode(), dl, NewVT, V1, Amt1);
17844 V2 = DAG.getNode(Op.getOpcode(), dl, NewVT, V2, Amt2);
17846 // Concatenate the result back
17847 return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT, V1, V2);
17853 static SDValue LowerXALUO(SDValue Op, SelectionDAG &DAG) {
17854 // Lower the "add/sub/mul with overflow" instruction into a regular ins plus
17855 // a "setcc" instruction that checks the overflow flag. The "brcond" lowering
17856 // looks for this combo and may remove the "setcc" instruction if the "setcc"
17857 // has only one use.
17858 SDNode *N = Op.getNode();
17859 SDValue LHS = N->getOperand(0);
17860 SDValue RHS = N->getOperand(1);
17861 unsigned BaseOp = 0;
17864 switch (Op.getOpcode()) {
17865 default: llvm_unreachable("Unknown ovf instruction!");
17867 // A subtract of one will be selected as a INC. Note that INC doesn't
17868 // set CF, so we can't do this for UADDO.
17869 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(RHS))
17871 BaseOp = X86ISD::INC;
17872 Cond = X86::COND_O;
17875 BaseOp = X86ISD::ADD;
17876 Cond = X86::COND_O;
17879 BaseOp = X86ISD::ADD;
17880 Cond = X86::COND_B;
17883 // A subtract of one will be selected as a DEC. Note that DEC doesn't
17884 // set CF, so we can't do this for USUBO.
17885 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(RHS))
17887 BaseOp = X86ISD::DEC;
17888 Cond = X86::COND_O;
17891 BaseOp = X86ISD::SUB;
17892 Cond = X86::COND_O;
17895 BaseOp = X86ISD::SUB;
17896 Cond = X86::COND_B;
17899 BaseOp = N->getValueType(0) == MVT::i8 ? X86ISD::SMUL8 : X86ISD::SMUL;
17900 Cond = X86::COND_O;
17902 case ISD::UMULO: { // i64, i8 = umulo lhs, rhs --> i64, i64, i32 umul lhs,rhs
17903 if (N->getValueType(0) == MVT::i8) {
17904 BaseOp = X86ISD::UMUL8;
17905 Cond = X86::COND_O;
17908 SDVTList VTs = DAG.getVTList(N->getValueType(0), N->getValueType(0),
17910 SDValue Sum = DAG.getNode(X86ISD::UMUL, DL, VTs, LHS, RHS);
17913 DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
17914 DAG.getConstant(X86::COND_O, DL, MVT::i32),
17915 SDValue(Sum.getNode(), 2));
17917 return DAG.getNode(ISD::MERGE_VALUES, DL, N->getVTList(), Sum, SetCC);
17921 // Also sets EFLAGS.
17922 SDVTList VTs = DAG.getVTList(N->getValueType(0), MVT::i32);
17923 SDValue Sum = DAG.getNode(BaseOp, DL, VTs, LHS, RHS);
17926 DAG.getNode(X86ISD::SETCC, DL, N->getValueType(1),
17927 DAG.getConstant(Cond, DL, MVT::i32),
17928 SDValue(Sum.getNode(), 1));
17930 return DAG.getNode(ISD::MERGE_VALUES, DL, N->getVTList(), Sum, SetCC);
17933 /// Returns true if the operand type is exactly twice the native width, and
17934 /// the corresponding cmpxchg8b or cmpxchg16b instruction is available.
17935 /// Used to know whether to use cmpxchg8/16b when expanding atomic operations
17936 /// (otherwise we leave them alone to become __sync_fetch_and_... calls).
17937 bool X86TargetLowering::needsCmpXchgNb(const Type *MemType) const {
17938 unsigned OpWidth = MemType->getPrimitiveSizeInBits();
17941 return !Subtarget->is64Bit(); // FIXME this should be Subtarget.hasCmpxchg8b
17942 else if (OpWidth == 128)
17943 return Subtarget->hasCmpxchg16b();
17948 bool X86TargetLowering::shouldExpandAtomicStoreInIR(StoreInst *SI) const {
17949 return needsCmpXchgNb(SI->getValueOperand()->getType());
17952 // Note: this turns large loads into lock cmpxchg8b/16b.
17953 // FIXME: On 32 bits x86, fild/movq might be faster than lock cmpxchg8b.
17954 bool X86TargetLowering::shouldExpandAtomicLoadInIR(LoadInst *LI) const {
17955 auto PTy = cast<PointerType>(LI->getPointerOperand()->getType());
17956 return needsCmpXchgNb(PTy->getElementType());
17959 TargetLoweringBase::AtomicRMWExpansionKind
17960 X86TargetLowering::shouldExpandAtomicRMWInIR(AtomicRMWInst *AI) const {
17961 unsigned NativeWidth = Subtarget->is64Bit() ? 64 : 32;
17962 const Type *MemType = AI->getType();
17964 // If the operand is too big, we must see if cmpxchg8/16b is available
17965 // and default to library calls otherwise.
17966 if (MemType->getPrimitiveSizeInBits() > NativeWidth) {
17967 return needsCmpXchgNb(MemType) ? AtomicRMWExpansionKind::CmpXChg
17968 : AtomicRMWExpansionKind::None;
17971 AtomicRMWInst::BinOp Op = AI->getOperation();
17974 llvm_unreachable("Unknown atomic operation");
17975 case AtomicRMWInst::Xchg:
17976 case AtomicRMWInst::Add:
17977 case AtomicRMWInst::Sub:
17978 // It's better to use xadd, xsub or xchg for these in all cases.
17979 return AtomicRMWExpansionKind::None;
17980 case AtomicRMWInst::Or:
17981 case AtomicRMWInst::And:
17982 case AtomicRMWInst::Xor:
17983 // If the atomicrmw's result isn't actually used, we can just add a "lock"
17984 // prefix to a normal instruction for these operations.
17985 return !AI->use_empty() ? AtomicRMWExpansionKind::CmpXChg
17986 : AtomicRMWExpansionKind::None;
17987 case AtomicRMWInst::Nand:
17988 case AtomicRMWInst::Max:
17989 case AtomicRMWInst::Min:
17990 case AtomicRMWInst::UMax:
17991 case AtomicRMWInst::UMin:
17992 // These always require a non-trivial set of data operations on x86. We must
17993 // use a cmpxchg loop.
17994 return AtomicRMWExpansionKind::CmpXChg;
17998 static bool hasMFENCE(const X86Subtarget& Subtarget) {
17999 // Use mfence if we have SSE2 or we're on x86-64 (even if we asked for
18000 // no-sse2). There isn't any reason to disable it if the target processor
18002 return Subtarget.hasSSE2() || Subtarget.is64Bit();
18006 X86TargetLowering::lowerIdempotentRMWIntoFencedLoad(AtomicRMWInst *AI) const {
18007 unsigned NativeWidth = Subtarget->is64Bit() ? 64 : 32;
18008 const Type *MemType = AI->getType();
18009 // Accesses larger than the native width are turned into cmpxchg/libcalls, so
18010 // there is no benefit in turning such RMWs into loads, and it is actually
18011 // harmful as it introduces a mfence.
18012 if (MemType->getPrimitiveSizeInBits() > NativeWidth)
18015 auto Builder = IRBuilder<>(AI);
18016 Module *M = Builder.GetInsertBlock()->getParent()->getParent();
18017 auto SynchScope = AI->getSynchScope();
18018 // We must restrict the ordering to avoid generating loads with Release or
18019 // ReleaseAcquire orderings.
18020 auto Order = AtomicCmpXchgInst::getStrongestFailureOrdering(AI->getOrdering());
18021 auto Ptr = AI->getPointerOperand();
18023 // Before the load we need a fence. Here is an example lifted from
18024 // http://www.hpl.hp.com/techreports/2012/HPL-2012-68.pdf showing why a fence
18027 // x.store(1, relaxed);
18028 // r1 = y.fetch_add(0, release);
18030 // y.fetch_add(42, acquire);
18031 // r2 = x.load(relaxed);
18032 // r1 = r2 = 0 is impossible, but becomes possible if the idempotent rmw is
18033 // lowered to just a load without a fence. A mfence flushes the store buffer,
18034 // making the optimization clearly correct.
18035 // FIXME: it is required if isAtLeastRelease(Order) but it is not clear
18036 // otherwise, we might be able to be more agressive on relaxed idempotent
18037 // rmw. In practice, they do not look useful, so we don't try to be
18038 // especially clever.
18039 if (SynchScope == SingleThread)
18040 // FIXME: we could just insert an X86ISD::MEMBARRIER here, except we are at
18041 // the IR level, so we must wrap it in an intrinsic.
18044 if (!hasMFENCE(*Subtarget))
18045 // FIXME: it might make sense to use a locked operation here but on a
18046 // different cache-line to prevent cache-line bouncing. In practice it
18047 // is probably a small win, and x86 processors without mfence are rare
18048 // enough that we do not bother.
18052 llvm::Intrinsic::getDeclaration(M, Intrinsic::x86_sse2_mfence);
18053 Builder.CreateCall(MFence, {});
18055 // Finally we can emit the atomic load.
18056 LoadInst *Loaded = Builder.CreateAlignedLoad(Ptr,
18057 AI->getType()->getPrimitiveSizeInBits());
18058 Loaded->setAtomic(Order, SynchScope);
18059 AI->replaceAllUsesWith(Loaded);
18060 AI->eraseFromParent();
18064 static SDValue LowerATOMIC_FENCE(SDValue Op, const X86Subtarget *Subtarget,
18065 SelectionDAG &DAG) {
18067 AtomicOrdering FenceOrdering = static_cast<AtomicOrdering>(
18068 cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue());
18069 SynchronizationScope FenceScope = static_cast<SynchronizationScope>(
18070 cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue());
18072 // The only fence that needs an instruction is a sequentially-consistent
18073 // cross-thread fence.
18074 if (FenceOrdering == SequentiallyConsistent && FenceScope == CrossThread) {
18075 if (hasMFENCE(*Subtarget))
18076 return DAG.getNode(X86ISD::MFENCE, dl, MVT::Other, Op.getOperand(0));
18078 SDValue Chain = Op.getOperand(0);
18079 SDValue Zero = DAG.getConstant(0, dl, MVT::i32);
18081 DAG.getRegister(X86::ESP, MVT::i32), // Base
18082 DAG.getTargetConstant(1, dl, MVT::i8), // Scale
18083 DAG.getRegister(0, MVT::i32), // Index
18084 DAG.getTargetConstant(0, dl, MVT::i32), // Disp
18085 DAG.getRegister(0, MVT::i32), // Segment.
18089 SDNode *Res = DAG.getMachineNode(X86::OR32mrLocked, dl, MVT::Other, Ops);
18090 return SDValue(Res, 0);
18093 // MEMBARRIER is a compiler barrier; it codegens to a no-op.
18094 return DAG.getNode(X86ISD::MEMBARRIER, dl, MVT::Other, Op.getOperand(0));
18097 static SDValue LowerCMP_SWAP(SDValue Op, const X86Subtarget *Subtarget,
18098 SelectionDAG &DAG) {
18099 MVT T = Op.getSimpleValueType();
18103 switch(T.SimpleTy) {
18104 default: llvm_unreachable("Invalid value type!");
18105 case MVT::i8: Reg = X86::AL; size = 1; break;
18106 case MVT::i16: Reg = X86::AX; size = 2; break;
18107 case MVT::i32: Reg = X86::EAX; size = 4; break;
18109 assert(Subtarget->is64Bit() && "Node not type legal!");
18110 Reg = X86::RAX; size = 8;
18113 SDValue cpIn = DAG.getCopyToReg(Op.getOperand(0), DL, Reg,
18114 Op.getOperand(2), SDValue());
18115 SDValue Ops[] = { cpIn.getValue(0),
18118 DAG.getTargetConstant(size, DL, MVT::i8),
18119 cpIn.getValue(1) };
18120 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Glue);
18121 MachineMemOperand *MMO = cast<AtomicSDNode>(Op)->getMemOperand();
18122 SDValue Result = DAG.getMemIntrinsicNode(X86ISD::LCMPXCHG_DAG, DL, Tys,
18126 DAG.getCopyFromReg(Result.getValue(0), DL, Reg, T, Result.getValue(1));
18127 SDValue EFLAGS = DAG.getCopyFromReg(cpOut.getValue(1), DL, X86::EFLAGS,
18128 MVT::i32, cpOut.getValue(2));
18129 SDValue Success = DAG.getNode(X86ISD::SETCC, DL, Op->getValueType(1),
18130 DAG.getConstant(X86::COND_E, DL, MVT::i8),
18133 DAG.ReplaceAllUsesOfValueWith(Op.getValue(0), cpOut);
18134 DAG.ReplaceAllUsesOfValueWith(Op.getValue(1), Success);
18135 DAG.ReplaceAllUsesOfValueWith(Op.getValue(2), EFLAGS.getValue(1));
18139 static SDValue LowerBITCAST(SDValue Op, const X86Subtarget *Subtarget,
18140 SelectionDAG &DAG) {
18141 MVT SrcVT = Op.getOperand(0).getSimpleValueType();
18142 MVT DstVT = Op.getSimpleValueType();
18144 if (SrcVT == MVT::v2i32 || SrcVT == MVT::v4i16 || SrcVT == MVT::v8i8) {
18145 assert(Subtarget->hasSSE2() && "Requires at least SSE2!");
18146 if (DstVT != MVT::f64)
18147 // This conversion needs to be expanded.
18150 SDValue InVec = Op->getOperand(0);
18152 unsigned NumElts = SrcVT.getVectorNumElements();
18153 EVT SVT = SrcVT.getVectorElementType();
18155 // Widen the vector in input in the case of MVT::v2i32.
18156 // Example: from MVT::v2i32 to MVT::v4i32.
18157 SmallVector<SDValue, 16> Elts;
18158 for (unsigned i = 0, e = NumElts; i != e; ++i)
18159 Elts.push_back(DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, SVT, InVec,
18160 DAG.getIntPtrConstant(i, dl)));
18162 // Explicitly mark the extra elements as Undef.
18163 Elts.append(NumElts, DAG.getUNDEF(SVT));
18165 EVT NewVT = EVT::getVectorVT(*DAG.getContext(), SVT, NumElts * 2);
18166 SDValue BV = DAG.getNode(ISD::BUILD_VECTOR, dl, NewVT, Elts);
18167 SDValue ToV2F64 = DAG.getBitcast(MVT::v2f64, BV);
18168 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, ToV2F64,
18169 DAG.getIntPtrConstant(0, dl));
18172 assert(Subtarget->is64Bit() && !Subtarget->hasSSE2() &&
18173 Subtarget->hasMMX() && "Unexpected custom BITCAST");
18174 assert((DstVT == MVT::i64 ||
18175 (DstVT.isVector() && DstVT.getSizeInBits()==64)) &&
18176 "Unexpected custom BITCAST");
18177 // i64 <=> MMX conversions are Legal.
18178 if (SrcVT==MVT::i64 && DstVT.isVector())
18180 if (DstVT==MVT::i64 && SrcVT.isVector())
18182 // MMX <=> MMX conversions are Legal.
18183 if (SrcVT.isVector() && DstVT.isVector())
18185 // All other conversions need to be expanded.
18189 /// Compute the horizontal sum of bytes in V for the elements of VT.
18191 /// Requires V to be a byte vector and VT to be an integer vector type with
18192 /// wider elements than V's type. The width of the elements of VT determines
18193 /// how many bytes of V are summed horizontally to produce each element of the
18195 static SDValue LowerHorizontalByteSum(SDValue V, MVT VT,
18196 const X86Subtarget *Subtarget,
18197 SelectionDAG &DAG) {
18199 MVT ByteVecVT = V.getSimpleValueType();
18200 MVT EltVT = VT.getVectorElementType();
18201 int NumElts = VT.getVectorNumElements();
18202 assert(ByteVecVT.getVectorElementType() == MVT::i8 &&
18203 "Expected value to have byte element type.");
18204 assert(EltVT != MVT::i8 &&
18205 "Horizontal byte sum only makes sense for wider elements!");
18206 unsigned VecSize = VT.getSizeInBits();
18207 assert(ByteVecVT.getSizeInBits() == VecSize && "Cannot change vector size!");
18209 // PSADBW instruction horizontally add all bytes and leave the result in i64
18210 // chunks, thus directly computes the pop count for v2i64 and v4i64.
18211 if (EltVT == MVT::i64) {
18212 SDValue Zeros = getZeroVector(ByteVecVT, Subtarget, DAG, DL);
18213 V = DAG.getNode(X86ISD::PSADBW, DL, ByteVecVT, V, Zeros);
18214 return DAG.getBitcast(VT, V);
18217 if (EltVT == MVT::i32) {
18218 // We unpack the low half and high half into i32s interleaved with zeros so
18219 // that we can use PSADBW to horizontally sum them. The most useful part of
18220 // this is that it lines up the results of two PSADBW instructions to be
18221 // two v2i64 vectors which concatenated are the 4 population counts. We can
18222 // then use PACKUSWB to shrink and concatenate them into a v4i32 again.
18223 SDValue Zeros = getZeroVector(VT, Subtarget, DAG, DL);
18224 SDValue Low = DAG.getNode(X86ISD::UNPCKL, DL, VT, V, Zeros);
18225 SDValue High = DAG.getNode(X86ISD::UNPCKH, DL, VT, V, Zeros);
18227 // Do the horizontal sums into two v2i64s.
18228 Zeros = getZeroVector(ByteVecVT, Subtarget, DAG, DL);
18229 Low = DAG.getNode(X86ISD::PSADBW, DL, ByteVecVT,
18230 DAG.getBitcast(ByteVecVT, Low), Zeros);
18231 High = DAG.getNode(X86ISD::PSADBW, DL, ByteVecVT,
18232 DAG.getBitcast(ByteVecVT, High), Zeros);
18234 // Merge them together.
18235 MVT ShortVecVT = MVT::getVectorVT(MVT::i16, VecSize / 16);
18236 V = DAG.getNode(X86ISD::PACKUS, DL, ByteVecVT,
18237 DAG.getBitcast(ShortVecVT, Low),
18238 DAG.getBitcast(ShortVecVT, High));
18240 return DAG.getBitcast(VT, V);
18243 // The only element type left is i16.
18244 assert(EltVT == MVT::i16 && "Unknown how to handle type");
18246 // To obtain pop count for each i16 element starting from the pop count for
18247 // i8 elements, shift the i16s left by 8, sum as i8s, and then shift as i16s
18248 // right by 8. It is important to shift as i16s as i8 vector shift isn't
18249 // directly supported.
18250 SmallVector<SDValue, 16> Shifters(NumElts, DAG.getConstant(8, DL, EltVT));
18251 SDValue Shifter = DAG.getNode(ISD::BUILD_VECTOR, DL, VT, Shifters);
18252 SDValue Shl = DAG.getNode(ISD::SHL, DL, VT, DAG.getBitcast(VT, V), Shifter);
18253 V = DAG.getNode(ISD::ADD, DL, ByteVecVT, DAG.getBitcast(ByteVecVT, Shl),
18254 DAG.getBitcast(ByteVecVT, V));
18255 return DAG.getNode(ISD::SRL, DL, VT, DAG.getBitcast(VT, V), Shifter);
18258 static SDValue LowerVectorCTPOPInRegLUT(SDValue Op, SDLoc DL,
18259 const X86Subtarget *Subtarget,
18260 SelectionDAG &DAG) {
18261 MVT VT = Op.getSimpleValueType();
18262 MVT EltVT = VT.getVectorElementType();
18263 unsigned VecSize = VT.getSizeInBits();
18265 // Implement a lookup table in register by using an algorithm based on:
18266 // http://wm.ite.pl/articles/sse-popcount.html
18268 // The general idea is that every lower byte nibble in the input vector is an
18269 // index into a in-register pre-computed pop count table. We then split up the
18270 // input vector in two new ones: (1) a vector with only the shifted-right
18271 // higher nibbles for each byte and (2) a vector with the lower nibbles (and
18272 // masked out higher ones) for each byte. PSHUB is used separately with both
18273 // to index the in-register table. Next, both are added and the result is a
18274 // i8 vector where each element contains the pop count for input byte.
18276 // To obtain the pop count for elements != i8, we follow up with the same
18277 // approach and use additional tricks as described below.
18279 const int LUT[16] = {/* 0 */ 0, /* 1 */ 1, /* 2 */ 1, /* 3 */ 2,
18280 /* 4 */ 1, /* 5 */ 2, /* 6 */ 2, /* 7 */ 3,
18281 /* 8 */ 1, /* 9 */ 2, /* a */ 2, /* b */ 3,
18282 /* c */ 2, /* d */ 3, /* e */ 3, /* f */ 4};
18284 int NumByteElts = VecSize / 8;
18285 MVT ByteVecVT = MVT::getVectorVT(MVT::i8, NumByteElts);
18286 SDValue In = DAG.getBitcast(ByteVecVT, Op);
18287 SmallVector<SDValue, 16> LUTVec;
18288 for (int i = 0; i < NumByteElts; ++i)
18289 LUTVec.push_back(DAG.getConstant(LUT[i % 16], DL, MVT::i8));
18290 SDValue InRegLUT = DAG.getNode(ISD::BUILD_VECTOR, DL, ByteVecVT, LUTVec);
18291 SmallVector<SDValue, 16> Mask0F(NumByteElts,
18292 DAG.getConstant(0x0F, DL, MVT::i8));
18293 SDValue M0F = DAG.getNode(ISD::BUILD_VECTOR, DL, ByteVecVT, Mask0F);
18296 SmallVector<SDValue, 16> Four(NumByteElts, DAG.getConstant(4, DL, MVT::i8));
18297 SDValue FourV = DAG.getNode(ISD::BUILD_VECTOR, DL, ByteVecVT, Four);
18298 SDValue HighNibbles = DAG.getNode(ISD::SRL, DL, ByteVecVT, In, FourV);
18301 SDValue LowNibbles = DAG.getNode(ISD::AND, DL, ByteVecVT, In, M0F);
18303 // The input vector is used as the shuffle mask that index elements into the
18304 // LUT. After counting low and high nibbles, add the vector to obtain the
18305 // final pop count per i8 element.
18306 SDValue HighPopCnt =
18307 DAG.getNode(X86ISD::PSHUFB, DL, ByteVecVT, InRegLUT, HighNibbles);
18308 SDValue LowPopCnt =
18309 DAG.getNode(X86ISD::PSHUFB, DL, ByteVecVT, InRegLUT, LowNibbles);
18310 SDValue PopCnt = DAG.getNode(ISD::ADD, DL, ByteVecVT, HighPopCnt, LowPopCnt);
18312 if (EltVT == MVT::i8)
18315 return LowerHorizontalByteSum(PopCnt, VT, Subtarget, DAG);
18318 static SDValue LowerVectorCTPOPBitmath(SDValue Op, SDLoc DL,
18319 const X86Subtarget *Subtarget,
18320 SelectionDAG &DAG) {
18321 MVT VT = Op.getSimpleValueType();
18322 assert(VT.is128BitVector() &&
18323 "Only 128-bit vector bitmath lowering supported.");
18325 int VecSize = VT.getSizeInBits();
18326 MVT EltVT = VT.getVectorElementType();
18327 int Len = EltVT.getSizeInBits();
18329 // This is the vectorized version of the "best" algorithm from
18330 // http://graphics.stanford.edu/~seander/bithacks.html#CountBitsSetParallel
18331 // with a minor tweak to use a series of adds + shifts instead of vector
18332 // multiplications. Implemented for all integer vector types. We only use
18333 // this when we don't have SSSE3 which allows a LUT-based lowering that is
18334 // much faster, even faster than using native popcnt instructions.
18336 auto GetShift = [&](unsigned OpCode, SDValue V, int Shifter) {
18337 MVT VT = V.getSimpleValueType();
18338 SmallVector<SDValue, 32> Shifters(
18339 VT.getVectorNumElements(),
18340 DAG.getConstant(Shifter, DL, VT.getVectorElementType()));
18341 return DAG.getNode(OpCode, DL, VT, V,
18342 DAG.getNode(ISD::BUILD_VECTOR, DL, VT, Shifters));
18344 auto GetMask = [&](SDValue V, APInt Mask) {
18345 MVT VT = V.getSimpleValueType();
18346 SmallVector<SDValue, 32> Masks(
18347 VT.getVectorNumElements(),
18348 DAG.getConstant(Mask, DL, VT.getVectorElementType()));
18349 return DAG.getNode(ISD::AND, DL, VT, V,
18350 DAG.getNode(ISD::BUILD_VECTOR, DL, VT, Masks));
18353 // We don't want to incur the implicit masks required to SRL vNi8 vectors on
18354 // x86, so set the SRL type to have elements at least i16 wide. This is
18355 // correct because all of our SRLs are followed immediately by a mask anyways
18356 // that handles any bits that sneak into the high bits of the byte elements.
18357 MVT SrlVT = Len > 8 ? VT : MVT::getVectorVT(MVT::i16, VecSize / 16);
18361 // v = v - ((v >> 1) & 0x55555555...)
18363 DAG.getBitcast(VT, GetShift(ISD::SRL, DAG.getBitcast(SrlVT, V), 1));
18364 SDValue And = GetMask(Srl, APInt::getSplat(Len, APInt(8, 0x55)));
18365 V = DAG.getNode(ISD::SUB, DL, VT, V, And);
18367 // v = (v & 0x33333333...) + ((v >> 2) & 0x33333333...)
18368 SDValue AndLHS = GetMask(V, APInt::getSplat(Len, APInt(8, 0x33)));
18369 Srl = DAG.getBitcast(VT, GetShift(ISD::SRL, DAG.getBitcast(SrlVT, V), 2));
18370 SDValue AndRHS = GetMask(Srl, APInt::getSplat(Len, APInt(8, 0x33)));
18371 V = DAG.getNode(ISD::ADD, DL, VT, AndLHS, AndRHS);
18373 // v = (v + (v >> 4)) & 0x0F0F0F0F...
18374 Srl = DAG.getBitcast(VT, GetShift(ISD::SRL, DAG.getBitcast(SrlVT, V), 4));
18375 SDValue Add = DAG.getNode(ISD::ADD, DL, VT, V, Srl);
18376 V = GetMask(Add, APInt::getSplat(Len, APInt(8, 0x0F)));
18378 // At this point, V contains the byte-wise population count, and we are
18379 // merely doing a horizontal sum if necessary to get the wider element
18381 if (EltVT == MVT::i8)
18384 return LowerHorizontalByteSum(
18385 DAG.getBitcast(MVT::getVectorVT(MVT::i8, VecSize / 8), V), VT, Subtarget,
18389 static SDValue LowerVectorCTPOP(SDValue Op, const X86Subtarget *Subtarget,
18390 SelectionDAG &DAG) {
18391 MVT VT = Op.getSimpleValueType();
18392 // FIXME: Need to add AVX-512 support here!
18393 assert((VT.is256BitVector() || VT.is128BitVector()) &&
18394 "Unknown CTPOP type to handle");
18395 SDLoc DL(Op.getNode());
18396 SDValue Op0 = Op.getOperand(0);
18398 if (!Subtarget->hasSSSE3()) {
18399 // We can't use the fast LUT approach, so fall back on vectorized bitmath.
18400 assert(VT.is128BitVector() && "Only 128-bit vectors supported in SSE!");
18401 return LowerVectorCTPOPBitmath(Op0, DL, Subtarget, DAG);
18404 if (VT.is256BitVector() && !Subtarget->hasInt256()) {
18405 unsigned NumElems = VT.getVectorNumElements();
18407 // Extract each 128-bit vector, compute pop count and concat the result.
18408 SDValue LHS = Extract128BitVector(Op0, 0, DAG, DL);
18409 SDValue RHS = Extract128BitVector(Op0, NumElems/2, DAG, DL);
18411 return DAG.getNode(ISD::CONCAT_VECTORS, DL, VT,
18412 LowerVectorCTPOPInRegLUT(LHS, DL, Subtarget, DAG),
18413 LowerVectorCTPOPInRegLUT(RHS, DL, Subtarget, DAG));
18416 return LowerVectorCTPOPInRegLUT(Op0, DL, Subtarget, DAG);
18419 static SDValue LowerCTPOP(SDValue Op, const X86Subtarget *Subtarget,
18420 SelectionDAG &DAG) {
18421 assert(Op.getValueType().isVector() &&
18422 "We only do custom lowering for vector population count.");
18423 return LowerVectorCTPOP(Op, Subtarget, DAG);
18426 static SDValue LowerLOAD_SUB(SDValue Op, SelectionDAG &DAG) {
18427 SDNode *Node = Op.getNode();
18429 EVT T = Node->getValueType(0);
18430 SDValue negOp = DAG.getNode(ISD::SUB, dl, T,
18431 DAG.getConstant(0, dl, T), Node->getOperand(2));
18432 return DAG.getAtomic(ISD::ATOMIC_LOAD_ADD, dl,
18433 cast<AtomicSDNode>(Node)->getMemoryVT(),
18434 Node->getOperand(0),
18435 Node->getOperand(1), negOp,
18436 cast<AtomicSDNode>(Node)->getMemOperand(),
18437 cast<AtomicSDNode>(Node)->getOrdering(),
18438 cast<AtomicSDNode>(Node)->getSynchScope());
18441 static SDValue LowerATOMIC_STORE(SDValue Op, SelectionDAG &DAG) {
18442 SDNode *Node = Op.getNode();
18444 EVT VT = cast<AtomicSDNode>(Node)->getMemoryVT();
18446 // Convert seq_cst store -> xchg
18447 // Convert wide store -> swap (-> cmpxchg8b/cmpxchg16b)
18448 // FIXME: On 32-bit, store -> fist or movq would be more efficient
18449 // (The only way to get a 16-byte store is cmpxchg16b)
18450 // FIXME: 16-byte ATOMIC_SWAP isn't actually hooked up at the moment.
18451 if (cast<AtomicSDNode>(Node)->getOrdering() == SequentiallyConsistent ||
18452 !DAG.getTargetLoweringInfo().isTypeLegal(VT)) {
18453 SDValue Swap = DAG.getAtomic(ISD::ATOMIC_SWAP, dl,
18454 cast<AtomicSDNode>(Node)->getMemoryVT(),
18455 Node->getOperand(0),
18456 Node->getOperand(1), Node->getOperand(2),
18457 cast<AtomicSDNode>(Node)->getMemOperand(),
18458 cast<AtomicSDNode>(Node)->getOrdering(),
18459 cast<AtomicSDNode>(Node)->getSynchScope());
18460 return Swap.getValue(1);
18462 // Other atomic stores have a simple pattern.
18466 static SDValue LowerADDC_ADDE_SUBC_SUBE(SDValue Op, SelectionDAG &DAG) {
18467 EVT VT = Op.getNode()->getSimpleValueType(0);
18469 // Let legalize expand this if it isn't a legal type yet.
18470 if (!DAG.getTargetLoweringInfo().isTypeLegal(VT))
18473 SDVTList VTs = DAG.getVTList(VT, MVT::i32);
18476 bool ExtraOp = false;
18477 switch (Op.getOpcode()) {
18478 default: llvm_unreachable("Invalid code");
18479 case ISD::ADDC: Opc = X86ISD::ADD; break;
18480 case ISD::ADDE: Opc = X86ISD::ADC; ExtraOp = true; break;
18481 case ISD::SUBC: Opc = X86ISD::SUB; break;
18482 case ISD::SUBE: Opc = X86ISD::SBB; ExtraOp = true; break;
18486 return DAG.getNode(Opc, SDLoc(Op), VTs, Op.getOperand(0),
18488 return DAG.getNode(Opc, SDLoc(Op), VTs, Op.getOperand(0),
18489 Op.getOperand(1), Op.getOperand(2));
18492 static SDValue LowerFSINCOS(SDValue Op, const X86Subtarget *Subtarget,
18493 SelectionDAG &DAG) {
18494 assert(Subtarget->isTargetDarwin() && Subtarget->is64Bit());
18496 // For MacOSX, we want to call an alternative entry point: __sincos_stret,
18497 // which returns the values as { float, float } (in XMM0) or
18498 // { double, double } (which is returned in XMM0, XMM1).
18500 SDValue Arg = Op.getOperand(0);
18501 EVT ArgVT = Arg.getValueType();
18502 Type *ArgTy = ArgVT.getTypeForEVT(*DAG.getContext());
18504 TargetLowering::ArgListTy Args;
18505 TargetLowering::ArgListEntry Entry;
18509 Entry.isSExt = false;
18510 Entry.isZExt = false;
18511 Args.push_back(Entry);
18513 bool isF64 = ArgVT == MVT::f64;
18514 // Only optimize x86_64 for now. i386 is a bit messy. For f32,
18515 // the small struct {f32, f32} is returned in (eax, edx). For f64,
18516 // the results are returned via SRet in memory.
18517 const char *LibcallName = isF64 ? "__sincos_stret" : "__sincosf_stret";
18518 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
18520 DAG.getExternalSymbol(LibcallName, TLI.getPointerTy(DAG.getDataLayout()));
18522 Type *RetTy = isF64
18523 ? (Type*)StructType::get(ArgTy, ArgTy, nullptr)
18524 : (Type*)VectorType::get(ArgTy, 4);
18526 TargetLowering::CallLoweringInfo CLI(DAG);
18527 CLI.setDebugLoc(dl).setChain(DAG.getEntryNode())
18528 .setCallee(CallingConv::C, RetTy, Callee, std::move(Args), 0);
18530 std::pair<SDValue, SDValue> CallResult = TLI.LowerCallTo(CLI);
18533 // Returned in xmm0 and xmm1.
18534 return CallResult.first;
18536 // Returned in bits 0:31 and 32:64 xmm0.
18537 SDValue SinVal = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, ArgVT,
18538 CallResult.first, DAG.getIntPtrConstant(0, dl));
18539 SDValue CosVal = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, ArgVT,
18540 CallResult.first, DAG.getIntPtrConstant(1, dl));
18541 SDVTList Tys = DAG.getVTList(ArgVT, ArgVT);
18542 return DAG.getNode(ISD::MERGE_VALUES, dl, Tys, SinVal, CosVal);
18545 static SDValue LowerMSCATTER(SDValue Op, const X86Subtarget *Subtarget,
18546 SelectionDAG &DAG) {
18547 assert(Subtarget->hasAVX512() &&
18548 "MGATHER/MSCATTER are supported on AVX-512 arch only");
18550 MaskedScatterSDNode *N = cast<MaskedScatterSDNode>(Op.getNode());
18551 EVT VT = N->getValue().getValueType();
18552 assert(VT.getScalarSizeInBits() >= 32 && "Unsupported scatter op");
18555 // X86 scatter kills mask register, so its type should be added to
18556 // the list of return values
18557 if (N->getNumValues() == 1) {
18558 SDValue Index = N->getIndex();
18559 if (!Subtarget->hasVLX() && !VT.is512BitVector() &&
18560 !Index.getValueType().is512BitVector())
18561 Index = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v8i64, Index);
18563 SDVTList VTs = DAG.getVTList(N->getMask().getValueType(), MVT::Other);
18564 SDValue Ops[] = { N->getOperand(0), N->getOperand(1), N->getOperand(2),
18565 N->getOperand(3), Index };
18567 SDValue NewScatter = DAG.getMaskedScatter(VTs, VT, dl, Ops, N->getMemOperand());
18568 DAG.ReplaceAllUsesWith(Op, SDValue(NewScatter.getNode(), 1));
18569 return SDValue(NewScatter.getNode(), 0);
18574 static SDValue LowerMGATHER(SDValue Op, const X86Subtarget *Subtarget,
18575 SelectionDAG &DAG) {
18576 assert(Subtarget->hasAVX512() &&
18577 "MGATHER/MSCATTER are supported on AVX-512 arch only");
18579 MaskedGatherSDNode *N = cast<MaskedGatherSDNode>(Op.getNode());
18580 EVT VT = Op.getValueType();
18581 assert(VT.getScalarSizeInBits() >= 32 && "Unsupported gather op");
18584 SDValue Index = N->getIndex();
18585 if (!Subtarget->hasVLX() && !VT.is512BitVector() &&
18586 !Index.getValueType().is512BitVector()) {
18587 Index = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v8i64, Index);
18588 SDValue Ops[] = { N->getOperand(0), N->getOperand(1), N->getOperand(2),
18589 N->getOperand(3), Index };
18590 DAG.UpdateNodeOperands(N, Ops);
18595 SDValue X86TargetLowering::LowerGC_TRANSITION_START(SDValue Op,
18596 SelectionDAG &DAG) const {
18597 // TODO: Eventually, the lowering of these nodes should be informed by or
18598 // deferred to the GC strategy for the function in which they appear. For
18599 // now, however, they must be lowered to something. Since they are logically
18600 // no-ops in the case of a null GC strategy (or a GC strategy which does not
18601 // require special handling for these nodes), lower them as literal NOOPs for
18603 SmallVector<SDValue, 2> Ops;
18605 Ops.push_back(Op.getOperand(0));
18606 if (Op->getGluedNode())
18607 Ops.push_back(Op->getOperand(Op->getNumOperands() - 1));
18610 SDVTList VTs = DAG.getVTList(MVT::Other, MVT::Glue);
18611 SDValue NOOP(DAG.getMachineNode(X86::NOOP, SDLoc(Op), VTs, Ops), 0);
18616 SDValue X86TargetLowering::LowerGC_TRANSITION_END(SDValue Op,
18617 SelectionDAG &DAG) const {
18618 // TODO: Eventually, the lowering of these nodes should be informed by or
18619 // deferred to the GC strategy for the function in which they appear. For
18620 // now, however, they must be lowered to something. Since they are logically
18621 // no-ops in the case of a null GC strategy (or a GC strategy which does not
18622 // require special handling for these nodes), lower them as literal NOOPs for
18624 SmallVector<SDValue, 2> Ops;
18626 Ops.push_back(Op.getOperand(0));
18627 if (Op->getGluedNode())
18628 Ops.push_back(Op->getOperand(Op->getNumOperands() - 1));
18631 SDVTList VTs = DAG.getVTList(MVT::Other, MVT::Glue);
18632 SDValue NOOP(DAG.getMachineNode(X86::NOOP, SDLoc(Op), VTs, Ops), 0);
18637 /// LowerOperation - Provide custom lowering hooks for some operations.
18639 SDValue X86TargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
18640 switch (Op.getOpcode()) {
18641 default: llvm_unreachable("Should not custom lower this!");
18642 case ISD::ATOMIC_FENCE: return LowerATOMIC_FENCE(Op, Subtarget, DAG);
18643 case ISD::ATOMIC_CMP_SWAP_WITH_SUCCESS:
18644 return LowerCMP_SWAP(Op, Subtarget, DAG);
18645 case ISD::CTPOP: return LowerCTPOP(Op, Subtarget, DAG);
18646 case ISD::ATOMIC_LOAD_SUB: return LowerLOAD_SUB(Op,DAG);
18647 case ISD::ATOMIC_STORE: return LowerATOMIC_STORE(Op,DAG);
18648 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG);
18649 case ISD::CONCAT_VECTORS: return LowerCONCAT_VECTORS(Op, Subtarget, DAG);
18650 case ISD::VECTOR_SHUFFLE: return lowerVectorShuffle(Op, Subtarget, DAG);
18651 case ISD::VSELECT: return LowerVSELECT(Op, DAG);
18652 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
18653 case ISD::INSERT_VECTOR_ELT: return LowerINSERT_VECTOR_ELT(Op, DAG);
18654 case ISD::EXTRACT_SUBVECTOR: return LowerEXTRACT_SUBVECTOR(Op,Subtarget,DAG);
18655 case ISD::INSERT_SUBVECTOR: return LowerINSERT_SUBVECTOR(Op, Subtarget,DAG);
18656 case ISD::SCALAR_TO_VECTOR: return LowerSCALAR_TO_VECTOR(Op, DAG);
18657 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
18658 case ISD::GlobalAddress: return LowerGlobalAddress(Op, DAG);
18659 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
18660 case ISD::ExternalSymbol: return LowerExternalSymbol(Op, DAG);
18661 case ISD::BlockAddress: return LowerBlockAddress(Op, DAG);
18662 case ISD::SHL_PARTS:
18663 case ISD::SRA_PARTS:
18664 case ISD::SRL_PARTS: return LowerShiftParts(Op, DAG);
18665 case ISD::SINT_TO_FP: return LowerSINT_TO_FP(Op, DAG);
18666 case ISD::UINT_TO_FP: return LowerUINT_TO_FP(Op, DAG);
18667 case ISD::TRUNCATE: return LowerTRUNCATE(Op, DAG);
18668 case ISD::ZERO_EXTEND: return LowerZERO_EXTEND(Op, Subtarget, DAG);
18669 case ISD::SIGN_EXTEND: return LowerSIGN_EXTEND(Op, Subtarget, DAG);
18670 case ISD::ANY_EXTEND: return LowerANY_EXTEND(Op, Subtarget, DAG);
18671 case ISD::SIGN_EXTEND_VECTOR_INREG:
18672 return LowerSIGN_EXTEND_VECTOR_INREG(Op, Subtarget, DAG);
18673 case ISD::FP_TO_SINT: return LowerFP_TO_SINT(Op, DAG);
18674 case ISD::FP_TO_UINT: return LowerFP_TO_UINT(Op, DAG);
18675 case ISD::FP_EXTEND: return LowerFP_EXTEND(Op, DAG);
18676 case ISD::LOAD: return LowerExtendedLoad(Op, Subtarget, DAG);
18678 case ISD::FNEG: return LowerFABSorFNEG(Op, DAG);
18679 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
18680 case ISD::FGETSIGN: return LowerFGETSIGN(Op, DAG);
18681 case ISD::SETCC: return LowerSETCC(Op, DAG);
18682 case ISD::SELECT: return LowerSELECT(Op, DAG);
18683 case ISD::BRCOND: return LowerBRCOND(Op, DAG);
18684 case ISD::JumpTable: return LowerJumpTable(Op, DAG);
18685 case ISD::VASTART: return LowerVASTART(Op, DAG);
18686 case ISD::VAARG: return LowerVAARG(Op, DAG);
18687 case ISD::VACOPY: return LowerVACOPY(Op, Subtarget, DAG);
18688 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, Subtarget, DAG);
18689 case ISD::INTRINSIC_VOID:
18690 case ISD::INTRINSIC_W_CHAIN: return LowerINTRINSIC_W_CHAIN(Op, Subtarget, DAG);
18691 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
18692 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
18693 case ISD::FRAME_TO_ARGS_OFFSET:
18694 return LowerFRAME_TO_ARGS_OFFSET(Op, DAG);
18695 case ISD::DYNAMIC_STACKALLOC: return LowerDYNAMIC_STACKALLOC(Op, DAG);
18696 case ISD::EH_RETURN: return LowerEH_RETURN(Op, DAG);
18697 case ISD::EH_SJLJ_SETJMP: return lowerEH_SJLJ_SETJMP(Op, DAG);
18698 case ISD::EH_SJLJ_LONGJMP: return lowerEH_SJLJ_LONGJMP(Op, DAG);
18699 case ISD::INIT_TRAMPOLINE: return LowerINIT_TRAMPOLINE(Op, DAG);
18700 case ISD::ADJUST_TRAMPOLINE: return LowerADJUST_TRAMPOLINE(Op, DAG);
18701 case ISD::FLT_ROUNDS_: return LowerFLT_ROUNDS_(Op, DAG);
18702 case ISD::CTLZ: return LowerCTLZ(Op, DAG);
18703 case ISD::CTLZ_ZERO_UNDEF: return LowerCTLZ_ZERO_UNDEF(Op, DAG);
18704 case ISD::CTTZ: return LowerCTTZ(Op, DAG);
18705 case ISD::MUL: return LowerMUL(Op, Subtarget, DAG);
18706 case ISD::UMUL_LOHI:
18707 case ISD::SMUL_LOHI: return LowerMUL_LOHI(Op, Subtarget, DAG);
18710 case ISD::SHL: return LowerShift(Op, Subtarget, DAG);
18716 case ISD::UMULO: return LowerXALUO(Op, DAG);
18717 case ISD::READCYCLECOUNTER: return LowerREADCYCLECOUNTER(Op, Subtarget,DAG);
18718 case ISD::BITCAST: return LowerBITCAST(Op, Subtarget, DAG);
18722 case ISD::SUBE: return LowerADDC_ADDE_SUBC_SUBE(Op, DAG);
18723 case ISD::ADD: return LowerADD(Op, DAG);
18724 case ISD::SUB: return LowerSUB(Op, DAG);
18725 case ISD::FSINCOS: return LowerFSINCOS(Op, Subtarget, DAG);
18726 case ISD::MGATHER: return LowerMGATHER(Op, Subtarget, DAG);
18727 case ISD::MSCATTER: return LowerMSCATTER(Op, Subtarget, DAG);
18728 case ISD::GC_TRANSITION_START:
18729 return LowerGC_TRANSITION_START(Op, DAG);
18730 case ISD::GC_TRANSITION_END: return LowerGC_TRANSITION_END(Op, DAG);
18734 /// ReplaceNodeResults - Replace a node with an illegal result type
18735 /// with a new node built out of custom code.
18736 void X86TargetLowering::ReplaceNodeResults(SDNode *N,
18737 SmallVectorImpl<SDValue>&Results,
18738 SelectionDAG &DAG) const {
18740 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
18741 switch (N->getOpcode()) {
18743 llvm_unreachable("Do not know how to custom type legalize this operation!");
18744 // We might have generated v2f32 FMIN/FMAX operations. Widen them to v4f32.
18745 case X86ISD::FMINC:
18747 case X86ISD::FMAXC:
18748 case X86ISD::FMAX: {
18749 EVT VT = N->getValueType(0);
18750 if (VT != MVT::v2f32)
18751 llvm_unreachable("Unexpected type (!= v2f32) on FMIN/FMAX.");
18752 SDValue UNDEF = DAG.getUNDEF(VT);
18753 SDValue LHS = DAG.getNode(ISD::CONCAT_VECTORS, dl, MVT::v4f32,
18754 N->getOperand(0), UNDEF);
18755 SDValue RHS = DAG.getNode(ISD::CONCAT_VECTORS, dl, MVT::v4f32,
18756 N->getOperand(1), UNDEF);
18757 Results.push_back(DAG.getNode(N->getOpcode(), dl, MVT::v4f32, LHS, RHS));
18760 case ISD::SIGN_EXTEND_INREG:
18765 // We don't want to expand or promote these.
18772 case ISD::UDIVREM: {
18773 SDValue V = LowerWin64_i128OP(SDValue(N,0), DAG);
18774 Results.push_back(V);
18777 case ISD::FP_TO_SINT:
18778 // FP_TO_INT*_IN_MEM is not legal for f16 inputs. Do not convert
18779 // (FP_TO_SINT (load f16)) to FP_TO_INT*.
18780 if (N->getOperand(0).getValueType() == MVT::f16)
18783 case ISD::FP_TO_UINT: {
18784 bool IsSigned = N->getOpcode() == ISD::FP_TO_SINT;
18786 if (!IsSigned && !isIntegerTypeFTOL(SDValue(N, 0).getValueType()))
18789 std::pair<SDValue,SDValue> Vals =
18790 FP_TO_INTHelper(SDValue(N, 0), DAG, IsSigned, /*IsReplace=*/ true);
18791 SDValue FIST = Vals.first, StackSlot = Vals.second;
18792 if (FIST.getNode()) {
18793 EVT VT = N->getValueType(0);
18794 // Return a load from the stack slot.
18795 if (StackSlot.getNode())
18796 Results.push_back(DAG.getLoad(VT, dl, FIST, StackSlot,
18797 MachinePointerInfo(),
18798 false, false, false, 0));
18800 Results.push_back(FIST);
18804 case ISD::UINT_TO_FP: {
18805 assert(Subtarget->hasSSE2() && "Requires at least SSE2!");
18806 if (N->getOperand(0).getValueType() != MVT::v2i32 ||
18807 N->getValueType(0) != MVT::v2f32)
18809 SDValue ZExtIn = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v2i64,
18811 SDValue Bias = DAG.getConstantFP(BitsToDouble(0x4330000000000000ULL), dl,
18813 SDValue VBias = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v2f64, Bias, Bias);
18814 SDValue Or = DAG.getNode(ISD::OR, dl, MVT::v2i64, ZExtIn,
18815 DAG.getBitcast(MVT::v2i64, VBias));
18816 Or = DAG.getBitcast(MVT::v2f64, Or);
18817 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::v2f64, Or, VBias);
18818 Results.push_back(DAG.getNode(X86ISD::VFPROUND, dl, MVT::v4f32, Sub));
18821 case ISD::FP_ROUND: {
18822 if (!TLI.isTypeLegal(N->getOperand(0).getValueType()))
18824 SDValue V = DAG.getNode(X86ISD::VFPROUND, dl, MVT::v4f32, N->getOperand(0));
18825 Results.push_back(V);
18828 case ISD::FP_EXTEND: {
18829 // Right now, only MVT::v2f32 has OperationAction for FP_EXTEND.
18830 // No other ValueType for FP_EXTEND should reach this point.
18831 assert(N->getValueType(0) == MVT::v2f32 &&
18832 "Do not know how to legalize this Node");
18835 case ISD::INTRINSIC_W_CHAIN: {
18836 unsigned IntNo = cast<ConstantSDNode>(N->getOperand(1))->getZExtValue();
18838 default : llvm_unreachable("Do not know how to custom type "
18839 "legalize this intrinsic operation!");
18840 case Intrinsic::x86_rdtsc:
18841 return getReadTimeStampCounter(N, dl, X86ISD::RDTSC_DAG, DAG, Subtarget,
18843 case Intrinsic::x86_rdtscp:
18844 return getReadTimeStampCounter(N, dl, X86ISD::RDTSCP_DAG, DAG, Subtarget,
18846 case Intrinsic::x86_rdpmc:
18847 return getReadPerformanceCounter(N, dl, DAG, Subtarget, Results);
18850 case ISD::READCYCLECOUNTER: {
18851 return getReadTimeStampCounter(N, dl, X86ISD::RDTSC_DAG, DAG, Subtarget,
18854 case ISD::ATOMIC_CMP_SWAP_WITH_SUCCESS: {
18855 EVT T = N->getValueType(0);
18856 assert((T == MVT::i64 || T == MVT::i128) && "can only expand cmpxchg pair");
18857 bool Regs64bit = T == MVT::i128;
18858 EVT HalfT = Regs64bit ? MVT::i64 : MVT::i32;
18859 SDValue cpInL, cpInH;
18860 cpInL = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, HalfT, N->getOperand(2),
18861 DAG.getConstant(0, dl, HalfT));
18862 cpInH = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, HalfT, N->getOperand(2),
18863 DAG.getConstant(1, dl, HalfT));
18864 cpInL = DAG.getCopyToReg(N->getOperand(0), dl,
18865 Regs64bit ? X86::RAX : X86::EAX,
18867 cpInH = DAG.getCopyToReg(cpInL.getValue(0), dl,
18868 Regs64bit ? X86::RDX : X86::EDX,
18869 cpInH, cpInL.getValue(1));
18870 SDValue swapInL, swapInH;
18871 swapInL = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, HalfT, N->getOperand(3),
18872 DAG.getConstant(0, dl, HalfT));
18873 swapInH = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, HalfT, N->getOperand(3),
18874 DAG.getConstant(1, dl, HalfT));
18875 swapInL = DAG.getCopyToReg(cpInH.getValue(0), dl,
18876 Regs64bit ? X86::RBX : X86::EBX,
18877 swapInL, cpInH.getValue(1));
18878 swapInH = DAG.getCopyToReg(swapInL.getValue(0), dl,
18879 Regs64bit ? X86::RCX : X86::ECX,
18880 swapInH, swapInL.getValue(1));
18881 SDValue Ops[] = { swapInH.getValue(0),
18883 swapInH.getValue(1) };
18884 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Glue);
18885 MachineMemOperand *MMO = cast<AtomicSDNode>(N)->getMemOperand();
18886 unsigned Opcode = Regs64bit ? X86ISD::LCMPXCHG16_DAG :
18887 X86ISD::LCMPXCHG8_DAG;
18888 SDValue Result = DAG.getMemIntrinsicNode(Opcode, dl, Tys, Ops, T, MMO);
18889 SDValue cpOutL = DAG.getCopyFromReg(Result.getValue(0), dl,
18890 Regs64bit ? X86::RAX : X86::EAX,
18891 HalfT, Result.getValue(1));
18892 SDValue cpOutH = DAG.getCopyFromReg(cpOutL.getValue(1), dl,
18893 Regs64bit ? X86::RDX : X86::EDX,
18894 HalfT, cpOutL.getValue(2));
18895 SDValue OpsF[] = { cpOutL.getValue(0), cpOutH.getValue(0)};
18897 SDValue EFLAGS = DAG.getCopyFromReg(cpOutH.getValue(1), dl, X86::EFLAGS,
18898 MVT::i32, cpOutH.getValue(2));
18900 DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
18901 DAG.getConstant(X86::COND_E, dl, MVT::i8), EFLAGS);
18902 Success = DAG.getZExtOrTrunc(Success, dl, N->getValueType(1));
18904 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, T, OpsF));
18905 Results.push_back(Success);
18906 Results.push_back(EFLAGS.getValue(1));
18909 case ISD::ATOMIC_SWAP:
18910 case ISD::ATOMIC_LOAD_ADD:
18911 case ISD::ATOMIC_LOAD_SUB:
18912 case ISD::ATOMIC_LOAD_AND:
18913 case ISD::ATOMIC_LOAD_OR:
18914 case ISD::ATOMIC_LOAD_XOR:
18915 case ISD::ATOMIC_LOAD_NAND:
18916 case ISD::ATOMIC_LOAD_MIN:
18917 case ISD::ATOMIC_LOAD_MAX:
18918 case ISD::ATOMIC_LOAD_UMIN:
18919 case ISD::ATOMIC_LOAD_UMAX:
18920 case ISD::ATOMIC_LOAD: {
18921 // Delegate to generic TypeLegalization. Situations we can really handle
18922 // should have already been dealt with by AtomicExpandPass.cpp.
18925 case ISD::BITCAST: {
18926 assert(Subtarget->hasSSE2() && "Requires at least SSE2!");
18927 EVT DstVT = N->getValueType(0);
18928 EVT SrcVT = N->getOperand(0)->getValueType(0);
18930 if (SrcVT != MVT::f64 ||
18931 (DstVT != MVT::v2i32 && DstVT != MVT::v4i16 && DstVT != MVT::v8i8))
18934 unsigned NumElts = DstVT.getVectorNumElements();
18935 EVT SVT = DstVT.getVectorElementType();
18936 EVT WiderVT = EVT::getVectorVT(*DAG.getContext(), SVT, NumElts * 2);
18937 SDValue Expanded = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
18938 MVT::v2f64, N->getOperand(0));
18939 SDValue ToVecInt = DAG.getBitcast(WiderVT, Expanded);
18941 if (ExperimentalVectorWideningLegalization) {
18942 // If we are legalizing vectors by widening, we already have the desired
18943 // legal vector type, just return it.
18944 Results.push_back(ToVecInt);
18948 SmallVector<SDValue, 8> Elts;
18949 for (unsigned i = 0, e = NumElts; i != e; ++i)
18950 Elts.push_back(DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, SVT,
18951 ToVecInt, DAG.getIntPtrConstant(i, dl)));
18953 Results.push_back(DAG.getNode(ISD::BUILD_VECTOR, dl, DstVT, Elts));
18958 const char *X86TargetLowering::getTargetNodeName(unsigned Opcode) const {
18959 switch ((X86ISD::NodeType)Opcode) {
18960 case X86ISD::FIRST_NUMBER: break;
18961 case X86ISD::BSF: return "X86ISD::BSF";
18962 case X86ISD::BSR: return "X86ISD::BSR";
18963 case X86ISD::SHLD: return "X86ISD::SHLD";
18964 case X86ISD::SHRD: return "X86ISD::SHRD";
18965 case X86ISD::FAND: return "X86ISD::FAND";
18966 case X86ISD::FANDN: return "X86ISD::FANDN";
18967 case X86ISD::FOR: return "X86ISD::FOR";
18968 case X86ISD::FXOR: return "X86ISD::FXOR";
18969 case X86ISD::FILD: return "X86ISD::FILD";
18970 case X86ISD::FILD_FLAG: return "X86ISD::FILD_FLAG";
18971 case X86ISD::FP_TO_INT16_IN_MEM: return "X86ISD::FP_TO_INT16_IN_MEM";
18972 case X86ISD::FP_TO_INT32_IN_MEM: return "X86ISD::FP_TO_INT32_IN_MEM";
18973 case X86ISD::FP_TO_INT64_IN_MEM: return "X86ISD::FP_TO_INT64_IN_MEM";
18974 case X86ISD::FLD: return "X86ISD::FLD";
18975 case X86ISD::FST: return "X86ISD::FST";
18976 case X86ISD::CALL: return "X86ISD::CALL";
18977 case X86ISD::RDTSC_DAG: return "X86ISD::RDTSC_DAG";
18978 case X86ISD::RDTSCP_DAG: return "X86ISD::RDTSCP_DAG";
18979 case X86ISD::RDPMC_DAG: return "X86ISD::RDPMC_DAG";
18980 case X86ISD::BT: return "X86ISD::BT";
18981 case X86ISD::CMP: return "X86ISD::CMP";
18982 case X86ISD::COMI: return "X86ISD::COMI";
18983 case X86ISD::UCOMI: return "X86ISD::UCOMI";
18984 case X86ISD::CMPM: return "X86ISD::CMPM";
18985 case X86ISD::CMPMU: return "X86ISD::CMPMU";
18986 case X86ISD::CMPM_RND: return "X86ISD::CMPM_RND";
18987 case X86ISD::SETCC: return "X86ISD::SETCC";
18988 case X86ISD::SETCC_CARRY: return "X86ISD::SETCC_CARRY";
18989 case X86ISD::FSETCC: return "X86ISD::FSETCC";
18990 case X86ISD::FGETSIGNx86: return "X86ISD::FGETSIGNx86";
18991 case X86ISD::CMOV: return "X86ISD::CMOV";
18992 case X86ISD::BRCOND: return "X86ISD::BRCOND";
18993 case X86ISD::RET_FLAG: return "X86ISD::RET_FLAG";
18994 case X86ISD::REP_STOS: return "X86ISD::REP_STOS";
18995 case X86ISD::REP_MOVS: return "X86ISD::REP_MOVS";
18996 case X86ISD::GlobalBaseReg: return "X86ISD::GlobalBaseReg";
18997 case X86ISD::Wrapper: return "X86ISD::Wrapper";
18998 case X86ISD::WrapperRIP: return "X86ISD::WrapperRIP";
18999 case X86ISD::MOVDQ2Q: return "X86ISD::MOVDQ2Q";
19000 case X86ISD::MMX_MOVD2W: return "X86ISD::MMX_MOVD2W";
19001 case X86ISD::MMX_MOVW2D: return "X86ISD::MMX_MOVW2D";
19002 case X86ISD::PEXTRB: return "X86ISD::PEXTRB";
19003 case X86ISD::PEXTRW: return "X86ISD::PEXTRW";
19004 case X86ISD::INSERTPS: return "X86ISD::INSERTPS";
19005 case X86ISD::PINSRB: return "X86ISD::PINSRB";
19006 case X86ISD::PINSRW: return "X86ISD::PINSRW";
19007 case X86ISD::MMX_PINSRW: return "X86ISD::MMX_PINSRW";
19008 case X86ISD::PSHUFB: return "X86ISD::PSHUFB";
19009 case X86ISD::ANDNP: return "X86ISD::ANDNP";
19010 case X86ISD::PSIGN: return "X86ISD::PSIGN";
19011 case X86ISD::BLENDI: return "X86ISD::BLENDI";
19012 case X86ISD::SHRUNKBLEND: return "X86ISD::SHRUNKBLEND";
19013 case X86ISD::ADDUS: return "X86ISD::ADDUS";
19014 case X86ISD::SUBUS: return "X86ISD::SUBUS";
19015 case X86ISD::HADD: return "X86ISD::HADD";
19016 case X86ISD::HSUB: return "X86ISD::HSUB";
19017 case X86ISD::FHADD: return "X86ISD::FHADD";
19018 case X86ISD::FHSUB: return "X86ISD::FHSUB";
19019 case X86ISD::ABS: return "X86ISD::ABS";
19020 case X86ISD::FMAX: return "X86ISD::FMAX";
19021 case X86ISD::FMAX_RND: return "X86ISD::FMAX_RND";
19022 case X86ISD::FMIN: return "X86ISD::FMIN";
19023 case X86ISD::FMIN_RND: return "X86ISD::FMIN_RND";
19024 case X86ISD::FMAXC: return "X86ISD::FMAXC";
19025 case X86ISD::FMINC: return "X86ISD::FMINC";
19026 case X86ISD::FRSQRT: return "X86ISD::FRSQRT";
19027 case X86ISD::FRCP: return "X86ISD::FRCP";
19028 case X86ISD::EXTRQI: return "X86ISD::EXTRQI";
19029 case X86ISD::INSERTQI: return "X86ISD::INSERTQI";
19030 case X86ISD::TLSADDR: return "X86ISD::TLSADDR";
19031 case X86ISD::TLSBASEADDR: return "X86ISD::TLSBASEADDR";
19032 case X86ISD::TLSCALL: return "X86ISD::TLSCALL";
19033 case X86ISD::EH_SJLJ_SETJMP: return "X86ISD::EH_SJLJ_SETJMP";
19034 case X86ISD::EH_SJLJ_LONGJMP: return "X86ISD::EH_SJLJ_LONGJMP";
19035 case X86ISD::EH_RETURN: return "X86ISD::EH_RETURN";
19036 case X86ISD::TC_RETURN: return "X86ISD::TC_RETURN";
19037 case X86ISD::FNSTCW16m: return "X86ISD::FNSTCW16m";
19038 case X86ISD::FNSTSW16r: return "X86ISD::FNSTSW16r";
19039 case X86ISD::LCMPXCHG_DAG: return "X86ISD::LCMPXCHG_DAG";
19040 case X86ISD::LCMPXCHG8_DAG: return "X86ISD::LCMPXCHG8_DAG";
19041 case X86ISD::LCMPXCHG16_DAG: return "X86ISD::LCMPXCHG16_DAG";
19042 case X86ISD::VZEXT_MOVL: return "X86ISD::VZEXT_MOVL";
19043 case X86ISD::VZEXT_LOAD: return "X86ISD::VZEXT_LOAD";
19044 case X86ISD::VZEXT: return "X86ISD::VZEXT";
19045 case X86ISD::VSEXT: return "X86ISD::VSEXT";
19046 case X86ISD::VTRUNC: return "X86ISD::VTRUNC";
19047 case X86ISD::VTRUNCS: return "X86ISD::VTRUNCS";
19048 case X86ISD::VTRUNCUS: return "X86ISD::VTRUNCUS";
19049 case X86ISD::VINSERT: return "X86ISD::VINSERT";
19050 case X86ISD::VFPEXT: return "X86ISD::VFPEXT";
19051 case X86ISD::VFPROUND: return "X86ISD::VFPROUND";
19052 case X86ISD::CVTDQ2PD: return "X86ISD::CVTDQ2PD";
19053 case X86ISD::CVTUDQ2PD: return "X86ISD::CVTUDQ2PD";
19054 case X86ISD::VSHLDQ: return "X86ISD::VSHLDQ";
19055 case X86ISD::VSRLDQ: return "X86ISD::VSRLDQ";
19056 case X86ISD::VSHL: return "X86ISD::VSHL";
19057 case X86ISD::VSRL: return "X86ISD::VSRL";
19058 case X86ISD::VSRA: return "X86ISD::VSRA";
19059 case X86ISD::VSHLI: return "X86ISD::VSHLI";
19060 case X86ISD::VSRLI: return "X86ISD::VSRLI";
19061 case X86ISD::VSRAI: return "X86ISD::VSRAI";
19062 case X86ISD::CMPP: return "X86ISD::CMPP";
19063 case X86ISD::PCMPEQ: return "X86ISD::PCMPEQ";
19064 case X86ISD::PCMPGT: return "X86ISD::PCMPGT";
19065 case X86ISD::PCMPEQM: return "X86ISD::PCMPEQM";
19066 case X86ISD::PCMPGTM: return "X86ISD::PCMPGTM";
19067 case X86ISD::ADD: return "X86ISD::ADD";
19068 case X86ISD::SUB: return "X86ISD::SUB";
19069 case X86ISD::ADC: return "X86ISD::ADC";
19070 case X86ISD::SBB: return "X86ISD::SBB";
19071 case X86ISD::SMUL: return "X86ISD::SMUL";
19072 case X86ISD::UMUL: return "X86ISD::UMUL";
19073 case X86ISD::SMUL8: return "X86ISD::SMUL8";
19074 case X86ISD::UMUL8: return "X86ISD::UMUL8";
19075 case X86ISD::SDIVREM8_SEXT_HREG: return "X86ISD::SDIVREM8_SEXT_HREG";
19076 case X86ISD::UDIVREM8_ZEXT_HREG: return "X86ISD::UDIVREM8_ZEXT_HREG";
19077 case X86ISD::INC: return "X86ISD::INC";
19078 case X86ISD::DEC: return "X86ISD::DEC";
19079 case X86ISD::OR: return "X86ISD::OR";
19080 case X86ISD::XOR: return "X86ISD::XOR";
19081 case X86ISD::AND: return "X86ISD::AND";
19082 case X86ISD::BEXTR: return "X86ISD::BEXTR";
19083 case X86ISD::MUL_IMM: return "X86ISD::MUL_IMM";
19084 case X86ISD::PTEST: return "X86ISD::PTEST";
19085 case X86ISD::TESTP: return "X86ISD::TESTP";
19086 case X86ISD::TESTM: return "X86ISD::TESTM";
19087 case X86ISD::TESTNM: return "X86ISD::TESTNM";
19088 case X86ISD::KORTEST: return "X86ISD::KORTEST";
19089 case X86ISD::PACKSS: return "X86ISD::PACKSS";
19090 case X86ISD::PACKUS: return "X86ISD::PACKUS";
19091 case X86ISD::PALIGNR: return "X86ISD::PALIGNR";
19092 case X86ISD::VALIGN: return "X86ISD::VALIGN";
19093 case X86ISD::PSHUFD: return "X86ISD::PSHUFD";
19094 case X86ISD::PSHUFHW: return "X86ISD::PSHUFHW";
19095 case X86ISD::PSHUFLW: return "X86ISD::PSHUFLW";
19096 case X86ISD::SHUFP: return "X86ISD::SHUFP";
19097 case X86ISD::SHUF128: return "X86ISD::SHUF128";
19098 case X86ISD::MOVLHPS: return "X86ISD::MOVLHPS";
19099 case X86ISD::MOVLHPD: return "X86ISD::MOVLHPD";
19100 case X86ISD::MOVHLPS: return "X86ISD::MOVHLPS";
19101 case X86ISD::MOVLPS: return "X86ISD::MOVLPS";
19102 case X86ISD::MOVLPD: return "X86ISD::MOVLPD";
19103 case X86ISD::MOVDDUP: return "X86ISD::MOVDDUP";
19104 case X86ISD::MOVSHDUP: return "X86ISD::MOVSHDUP";
19105 case X86ISD::MOVSLDUP: return "X86ISD::MOVSLDUP";
19106 case X86ISD::MOVSD: return "X86ISD::MOVSD";
19107 case X86ISD::MOVSS: return "X86ISD::MOVSS";
19108 case X86ISD::UNPCKL: return "X86ISD::UNPCKL";
19109 case X86ISD::UNPCKH: return "X86ISD::UNPCKH";
19110 case X86ISD::VBROADCAST: return "X86ISD::VBROADCAST";
19111 case X86ISD::SUBV_BROADCAST: return "X86ISD::SUBV_BROADCAST";
19112 case X86ISD::VEXTRACT: return "X86ISD::VEXTRACT";
19113 case X86ISD::VPERMILPV: return "X86ISD::VPERMILPV";
19114 case X86ISD::VPERMILPI: return "X86ISD::VPERMILPI";
19115 case X86ISD::VPERM2X128: return "X86ISD::VPERM2X128";
19116 case X86ISD::VPERMV: return "X86ISD::VPERMV";
19117 case X86ISD::VPERMV3: return "X86ISD::VPERMV3";
19118 case X86ISD::VPERMIV3: return "X86ISD::VPERMIV3";
19119 case X86ISD::VPERMI: return "X86ISD::VPERMI";
19120 case X86ISD::VFIXUPIMM: return "X86ISD::VFIXUPIMM";
19121 case X86ISD::VRANGE: return "X86ISD::VRANGE";
19122 case X86ISD::PMULUDQ: return "X86ISD::PMULUDQ";
19123 case X86ISD::PMULDQ: return "X86ISD::PMULDQ";
19124 case X86ISD::PSADBW: return "X86ISD::PSADBW";
19125 case X86ISD::VASTART_SAVE_XMM_REGS: return "X86ISD::VASTART_SAVE_XMM_REGS";
19126 case X86ISD::VAARG_64: return "X86ISD::VAARG_64";
19127 case X86ISD::WIN_ALLOCA: return "X86ISD::WIN_ALLOCA";
19128 case X86ISD::MEMBARRIER: return "X86ISD::MEMBARRIER";
19129 case X86ISD::MFENCE: return "X86ISD::MFENCE";
19130 case X86ISD::SFENCE: return "X86ISD::SFENCE";
19131 case X86ISD::LFENCE: return "X86ISD::LFENCE";
19132 case X86ISD::SEG_ALLOCA: return "X86ISD::SEG_ALLOCA";
19133 case X86ISD::WIN_FTOL: return "X86ISD::WIN_FTOL";
19134 case X86ISD::SAHF: return "X86ISD::SAHF";
19135 case X86ISD::RDRAND: return "X86ISD::RDRAND";
19136 case X86ISD::RDSEED: return "X86ISD::RDSEED";
19137 case X86ISD::VPMADDUBSW: return "X86ISD::VPMADDUBSW";
19138 case X86ISD::VPMADDWD: return "X86ISD::VPMADDWD";
19139 case X86ISD::FMADD: return "X86ISD::FMADD";
19140 case X86ISD::FMSUB: return "X86ISD::FMSUB";
19141 case X86ISD::FNMADD: return "X86ISD::FNMADD";
19142 case X86ISD::FNMSUB: return "X86ISD::FNMSUB";
19143 case X86ISD::FMADDSUB: return "X86ISD::FMADDSUB";
19144 case X86ISD::FMSUBADD: return "X86ISD::FMSUBADD";
19145 case X86ISD::FMADD_RND: return "X86ISD::FMADD_RND";
19146 case X86ISD::FNMADD_RND: return "X86ISD::FNMADD_RND";
19147 case X86ISD::FMSUB_RND: return "X86ISD::FMSUB_RND";
19148 case X86ISD::FNMSUB_RND: return "X86ISD::FNMSUB_RND";
19149 case X86ISD::FMADDSUB_RND: return "X86ISD::FMADDSUB_RND";
19150 case X86ISD::FMSUBADD_RND: return "X86ISD::FMSUBADD_RND";
19151 case X86ISD::VRNDSCALE: return "X86ISD::VRNDSCALE";
19152 case X86ISD::VREDUCE: return "X86ISD::VREDUCE";
19153 case X86ISD::PCMPESTRI: return "X86ISD::PCMPESTRI";
19154 case X86ISD::PCMPISTRI: return "X86ISD::PCMPISTRI";
19155 case X86ISD::XTEST: return "X86ISD::XTEST";
19156 case X86ISD::COMPRESS: return "X86ISD::COMPRESS";
19157 case X86ISD::EXPAND: return "X86ISD::EXPAND";
19158 case X86ISD::SELECT: return "X86ISD::SELECT";
19159 case X86ISD::ADDSUB: return "X86ISD::ADDSUB";
19160 case X86ISD::RCP28: return "X86ISD::RCP28";
19161 case X86ISD::EXP2: return "X86ISD::EXP2";
19162 case X86ISD::RSQRT28: return "X86ISD::RSQRT28";
19163 case X86ISD::FADD_RND: return "X86ISD::FADD_RND";
19164 case X86ISD::FSUB_RND: return "X86ISD::FSUB_RND";
19165 case X86ISD::FMUL_RND: return "X86ISD::FMUL_RND";
19166 case X86ISD::FDIV_RND: return "X86ISD::FDIV_RND";
19167 case X86ISD::FSQRT_RND: return "X86ISD::FSQRT_RND";
19168 case X86ISD::FGETEXP_RND: return "X86ISD::FGETEXP_RND";
19169 case X86ISD::SCALEF: return "X86ISD::SCALEF";
19170 case X86ISD::ADDS: return "X86ISD::ADDS";
19171 case X86ISD::SUBS: return "X86ISD::SUBS";
19172 case X86ISD::AVG: return "X86ISD::AVG";
19173 case X86ISD::MULHRS: return "X86ISD::MULHRS";
19174 case X86ISD::SINT_TO_FP_RND: return "X86ISD::SINT_TO_FP_RND";
19175 case X86ISD::UINT_TO_FP_RND: return "X86ISD::UINT_TO_FP_RND";
19176 case X86ISD::FP_TO_SINT_RND: return "X86ISD::FP_TO_SINT_RND";
19177 case X86ISD::FP_TO_UINT_RND: return "X86ISD::FP_TO_UINT_RND";
19182 // isLegalAddressingMode - Return true if the addressing mode represented
19183 // by AM is legal for this target, for a load/store of the specified type.
19184 bool X86TargetLowering::isLegalAddressingMode(const DataLayout &DL,
19185 const AddrMode &AM, Type *Ty,
19186 unsigned AS) const {
19187 // X86 supports extremely general addressing modes.
19188 CodeModel::Model M = getTargetMachine().getCodeModel();
19189 Reloc::Model R = getTargetMachine().getRelocationModel();
19191 // X86 allows a sign-extended 32-bit immediate field as a displacement.
19192 if (!X86::isOffsetSuitableForCodeModel(AM.BaseOffs, M, AM.BaseGV != nullptr))
19197 Subtarget->ClassifyGlobalReference(AM.BaseGV, getTargetMachine());
19199 // If a reference to this global requires an extra load, we can't fold it.
19200 if (isGlobalStubReference(GVFlags))
19203 // If BaseGV requires a register for the PIC base, we cannot also have a
19204 // BaseReg specified.
19205 if (AM.HasBaseReg && isGlobalRelativeToPICBase(GVFlags))
19208 // If lower 4G is not available, then we must use rip-relative addressing.
19209 if ((M != CodeModel::Small || R != Reloc::Static) &&
19210 Subtarget->is64Bit() && (AM.BaseOffs || AM.Scale > 1))
19214 switch (AM.Scale) {
19220 // These scales always work.
19225 // These scales are formed with basereg+scalereg. Only accept if there is
19230 default: // Other stuff never works.
19237 bool X86TargetLowering::isVectorShiftByScalarCheap(Type *Ty) const {
19238 unsigned Bits = Ty->getScalarSizeInBits();
19240 // 8-bit shifts are always expensive, but versions with a scalar amount aren't
19241 // particularly cheaper than those without.
19245 // On AVX2 there are new vpsllv[dq] instructions (and other shifts), that make
19246 // variable shifts just as cheap as scalar ones.
19247 if (Subtarget->hasInt256() && (Bits == 32 || Bits == 64))
19250 // Otherwise, it's significantly cheaper to shift by a scalar amount than by a
19251 // fully general vector.
19255 bool X86TargetLowering::isTruncateFree(Type *Ty1, Type *Ty2) const {
19256 if (!Ty1->isIntegerTy() || !Ty2->isIntegerTy())
19258 unsigned NumBits1 = Ty1->getPrimitiveSizeInBits();
19259 unsigned NumBits2 = Ty2->getPrimitiveSizeInBits();
19260 return NumBits1 > NumBits2;
19263 bool X86TargetLowering::allowTruncateForTailCall(Type *Ty1, Type *Ty2) const {
19264 if (!Ty1->isIntegerTy() || !Ty2->isIntegerTy())
19267 if (!isTypeLegal(EVT::getEVT(Ty1)))
19270 assert(Ty1->getPrimitiveSizeInBits() <= 64 && "i128 is probably not a noop");
19272 // Assuming the caller doesn't have a zeroext or signext return parameter,
19273 // truncation all the way down to i1 is valid.
19277 bool X86TargetLowering::isLegalICmpImmediate(int64_t Imm) const {
19278 return isInt<32>(Imm);
19281 bool X86TargetLowering::isLegalAddImmediate(int64_t Imm) const {
19282 // Can also use sub to handle negated immediates.
19283 return isInt<32>(Imm);
19286 bool X86TargetLowering::isTruncateFree(EVT VT1, EVT VT2) const {
19287 if (!VT1.isInteger() || !VT2.isInteger())
19289 unsigned NumBits1 = VT1.getSizeInBits();
19290 unsigned NumBits2 = VT2.getSizeInBits();
19291 return NumBits1 > NumBits2;
19294 bool X86TargetLowering::isZExtFree(Type *Ty1, Type *Ty2) const {
19295 // x86-64 implicitly zero-extends 32-bit results in 64-bit registers.
19296 return Ty1->isIntegerTy(32) && Ty2->isIntegerTy(64) && Subtarget->is64Bit();
19299 bool X86TargetLowering::isZExtFree(EVT VT1, EVT VT2) const {
19300 // x86-64 implicitly zero-extends 32-bit results in 64-bit registers.
19301 return VT1 == MVT::i32 && VT2 == MVT::i64 && Subtarget->is64Bit();
19304 bool X86TargetLowering::isZExtFree(SDValue Val, EVT VT2) const {
19305 EVT VT1 = Val.getValueType();
19306 if (isZExtFree(VT1, VT2))
19309 if (Val.getOpcode() != ISD::LOAD)
19312 if (!VT1.isSimple() || !VT1.isInteger() ||
19313 !VT2.isSimple() || !VT2.isInteger())
19316 switch (VT1.getSimpleVT().SimpleTy) {
19321 // X86 has 8, 16, and 32-bit zero-extending loads.
19328 bool X86TargetLowering::isVectorLoadExtDesirable(SDValue) const { return true; }
19331 X86TargetLowering::isFMAFasterThanFMulAndFAdd(EVT VT) const {
19332 if (!(Subtarget->hasFMA() || Subtarget->hasFMA4() || Subtarget->hasAVX512()))
19335 VT = VT.getScalarType();
19337 if (!VT.isSimple())
19340 switch (VT.getSimpleVT().SimpleTy) {
19351 bool X86TargetLowering::isNarrowingProfitable(EVT VT1, EVT VT2) const {
19352 // i16 instructions are longer (0x66 prefix) and potentially slower.
19353 return !(VT1 == MVT::i32 && VT2 == MVT::i16);
19356 /// isShuffleMaskLegal - Targets can use this to indicate that they only
19357 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
19358 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
19359 /// are assumed to be legal.
19361 X86TargetLowering::isShuffleMaskLegal(const SmallVectorImpl<int> &M,
19363 if (!VT.isSimple())
19366 // Not for i1 vectors
19367 if (VT.getScalarType() == MVT::i1)
19370 // Very little shuffling can be done for 64-bit vectors right now.
19371 if (VT.getSizeInBits() == 64)
19374 // We only care that the types being shuffled are legal. The lowering can
19375 // handle any possible shuffle mask that results.
19376 return isTypeLegal(VT.getSimpleVT());
19380 X86TargetLowering::isVectorClearMaskLegal(const SmallVectorImpl<int> &Mask,
19382 // Just delegate to the generic legality, clear masks aren't special.
19383 return isShuffleMaskLegal(Mask, VT);
19386 //===----------------------------------------------------------------------===//
19387 // X86 Scheduler Hooks
19388 //===----------------------------------------------------------------------===//
19390 /// Utility function to emit xbegin specifying the start of an RTM region.
19391 static MachineBasicBlock *EmitXBegin(MachineInstr *MI, MachineBasicBlock *MBB,
19392 const TargetInstrInfo *TII) {
19393 DebugLoc DL = MI->getDebugLoc();
19395 const BasicBlock *BB = MBB->getBasicBlock();
19396 MachineFunction::iterator I = MBB;
19399 // For the v = xbegin(), we generate
19410 MachineBasicBlock *thisMBB = MBB;
19411 MachineFunction *MF = MBB->getParent();
19412 MachineBasicBlock *mainMBB = MF->CreateMachineBasicBlock(BB);
19413 MachineBasicBlock *sinkMBB = MF->CreateMachineBasicBlock(BB);
19414 MF->insert(I, mainMBB);
19415 MF->insert(I, sinkMBB);
19417 // Transfer the remainder of BB and its successor edges to sinkMBB.
19418 sinkMBB->splice(sinkMBB->begin(), MBB,
19419 std::next(MachineBasicBlock::iterator(MI)), MBB->end());
19420 sinkMBB->transferSuccessorsAndUpdatePHIs(MBB);
19424 // # fallthrough to mainMBB
19425 // # abortion to sinkMBB
19426 BuildMI(thisMBB, DL, TII->get(X86::XBEGIN_4)).addMBB(sinkMBB);
19427 thisMBB->addSuccessor(mainMBB);
19428 thisMBB->addSuccessor(sinkMBB);
19432 BuildMI(mainMBB, DL, TII->get(X86::MOV32ri), X86::EAX).addImm(-1);
19433 mainMBB->addSuccessor(sinkMBB);
19436 // EAX is live into the sinkMBB
19437 sinkMBB->addLiveIn(X86::EAX);
19438 BuildMI(*sinkMBB, sinkMBB->begin(), DL,
19439 TII->get(TargetOpcode::COPY), MI->getOperand(0).getReg())
19442 MI->eraseFromParent();
19446 // FIXME: When we get size specific XMM0 registers, i.e. XMM0_V16I8
19447 // or XMM0_V32I8 in AVX all of this code can be replaced with that
19448 // in the .td file.
19449 static MachineBasicBlock *EmitPCMPSTRM(MachineInstr *MI, MachineBasicBlock *BB,
19450 const TargetInstrInfo *TII) {
19452 switch (MI->getOpcode()) {
19453 default: llvm_unreachable("illegal opcode!");
19454 case X86::PCMPISTRM128REG: Opc = X86::PCMPISTRM128rr; break;
19455 case X86::VPCMPISTRM128REG: Opc = X86::VPCMPISTRM128rr; break;
19456 case X86::PCMPISTRM128MEM: Opc = X86::PCMPISTRM128rm; break;
19457 case X86::VPCMPISTRM128MEM: Opc = X86::VPCMPISTRM128rm; break;
19458 case X86::PCMPESTRM128REG: Opc = X86::PCMPESTRM128rr; break;
19459 case X86::VPCMPESTRM128REG: Opc = X86::VPCMPESTRM128rr; break;
19460 case X86::PCMPESTRM128MEM: Opc = X86::PCMPESTRM128rm; break;
19461 case X86::VPCMPESTRM128MEM: Opc = X86::VPCMPESTRM128rm; break;
19464 DebugLoc dl = MI->getDebugLoc();
19465 MachineInstrBuilder MIB = BuildMI(*BB, MI, dl, TII->get(Opc));
19467 unsigned NumArgs = MI->getNumOperands();
19468 for (unsigned i = 1; i < NumArgs; ++i) {
19469 MachineOperand &Op = MI->getOperand(i);
19470 if (!(Op.isReg() && Op.isImplicit()))
19471 MIB.addOperand(Op);
19473 if (MI->hasOneMemOperand())
19474 MIB->setMemRefs(MI->memoperands_begin(), MI->memoperands_end());
19476 BuildMI(*BB, MI, dl,
19477 TII->get(TargetOpcode::COPY), MI->getOperand(0).getReg())
19478 .addReg(X86::XMM0);
19480 MI->eraseFromParent();
19484 // FIXME: Custom handling because TableGen doesn't support multiple implicit
19485 // defs in an instruction pattern
19486 static MachineBasicBlock *EmitPCMPSTRI(MachineInstr *MI, MachineBasicBlock *BB,
19487 const TargetInstrInfo *TII) {
19489 switch (MI->getOpcode()) {
19490 default: llvm_unreachable("illegal opcode!");
19491 case X86::PCMPISTRIREG: Opc = X86::PCMPISTRIrr; break;
19492 case X86::VPCMPISTRIREG: Opc = X86::VPCMPISTRIrr; break;
19493 case X86::PCMPISTRIMEM: Opc = X86::PCMPISTRIrm; break;
19494 case X86::VPCMPISTRIMEM: Opc = X86::VPCMPISTRIrm; break;
19495 case X86::PCMPESTRIREG: Opc = X86::PCMPESTRIrr; break;
19496 case X86::VPCMPESTRIREG: Opc = X86::VPCMPESTRIrr; break;
19497 case X86::PCMPESTRIMEM: Opc = X86::PCMPESTRIrm; break;
19498 case X86::VPCMPESTRIMEM: Opc = X86::VPCMPESTRIrm; break;
19501 DebugLoc dl = MI->getDebugLoc();
19502 MachineInstrBuilder MIB = BuildMI(*BB, MI, dl, TII->get(Opc));
19504 unsigned NumArgs = MI->getNumOperands(); // remove the results
19505 for (unsigned i = 1; i < NumArgs; ++i) {
19506 MachineOperand &Op = MI->getOperand(i);
19507 if (!(Op.isReg() && Op.isImplicit()))
19508 MIB.addOperand(Op);
19510 if (MI->hasOneMemOperand())
19511 MIB->setMemRefs(MI->memoperands_begin(), MI->memoperands_end());
19513 BuildMI(*BB, MI, dl,
19514 TII->get(TargetOpcode::COPY), MI->getOperand(0).getReg())
19517 MI->eraseFromParent();
19521 static MachineBasicBlock *EmitMonitor(MachineInstr *MI, MachineBasicBlock *BB,
19522 const X86Subtarget *Subtarget) {
19523 DebugLoc dl = MI->getDebugLoc();
19524 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
19525 // Address into RAX/EAX, other two args into ECX, EDX.
19526 unsigned MemOpc = Subtarget->is64Bit() ? X86::LEA64r : X86::LEA32r;
19527 unsigned MemReg = Subtarget->is64Bit() ? X86::RAX : X86::EAX;
19528 MachineInstrBuilder MIB = BuildMI(*BB, MI, dl, TII->get(MemOpc), MemReg);
19529 for (int i = 0; i < X86::AddrNumOperands; ++i)
19530 MIB.addOperand(MI->getOperand(i));
19532 unsigned ValOps = X86::AddrNumOperands;
19533 BuildMI(*BB, MI, dl, TII->get(TargetOpcode::COPY), X86::ECX)
19534 .addReg(MI->getOperand(ValOps).getReg());
19535 BuildMI(*BB, MI, dl, TII->get(TargetOpcode::COPY), X86::EDX)
19536 .addReg(MI->getOperand(ValOps+1).getReg());
19538 // The instruction doesn't actually take any operands though.
19539 BuildMI(*BB, MI, dl, TII->get(X86::MONITORrrr));
19541 MI->eraseFromParent(); // The pseudo is gone now.
19545 MachineBasicBlock *
19546 X86TargetLowering::EmitVAARG64WithCustomInserter(MachineInstr *MI,
19547 MachineBasicBlock *MBB) const {
19548 // Emit va_arg instruction on X86-64.
19550 // Operands to this pseudo-instruction:
19551 // 0 ) Output : destination address (reg)
19552 // 1-5) Input : va_list address (addr, i64mem)
19553 // 6 ) ArgSize : Size (in bytes) of vararg type
19554 // 7 ) ArgMode : 0=overflow only, 1=use gp_offset, 2=use fp_offset
19555 // 8 ) Align : Alignment of type
19556 // 9 ) EFLAGS (implicit-def)
19558 assert(MI->getNumOperands() == 10 && "VAARG_64 should have 10 operands!");
19559 static_assert(X86::AddrNumOperands == 5,
19560 "VAARG_64 assumes 5 address operands");
19562 unsigned DestReg = MI->getOperand(0).getReg();
19563 MachineOperand &Base = MI->getOperand(1);
19564 MachineOperand &Scale = MI->getOperand(2);
19565 MachineOperand &Index = MI->getOperand(3);
19566 MachineOperand &Disp = MI->getOperand(4);
19567 MachineOperand &Segment = MI->getOperand(5);
19568 unsigned ArgSize = MI->getOperand(6).getImm();
19569 unsigned ArgMode = MI->getOperand(7).getImm();
19570 unsigned Align = MI->getOperand(8).getImm();
19572 // Memory Reference
19573 assert(MI->hasOneMemOperand() && "Expected VAARG_64 to have one memoperand");
19574 MachineInstr::mmo_iterator MMOBegin = MI->memoperands_begin();
19575 MachineInstr::mmo_iterator MMOEnd = MI->memoperands_end();
19577 // Machine Information
19578 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
19579 MachineRegisterInfo &MRI = MBB->getParent()->getRegInfo();
19580 const TargetRegisterClass *AddrRegClass = getRegClassFor(MVT::i64);
19581 const TargetRegisterClass *OffsetRegClass = getRegClassFor(MVT::i32);
19582 DebugLoc DL = MI->getDebugLoc();
19584 // struct va_list {
19587 // i64 overflow_area (address)
19588 // i64 reg_save_area (address)
19590 // sizeof(va_list) = 24
19591 // alignment(va_list) = 8
19593 unsigned TotalNumIntRegs = 6;
19594 unsigned TotalNumXMMRegs = 8;
19595 bool UseGPOffset = (ArgMode == 1);
19596 bool UseFPOffset = (ArgMode == 2);
19597 unsigned MaxOffset = TotalNumIntRegs * 8 +
19598 (UseFPOffset ? TotalNumXMMRegs * 16 : 0);
19600 /* Align ArgSize to a multiple of 8 */
19601 unsigned ArgSizeA8 = (ArgSize + 7) & ~7;
19602 bool NeedsAlign = (Align > 8);
19604 MachineBasicBlock *thisMBB = MBB;
19605 MachineBasicBlock *overflowMBB;
19606 MachineBasicBlock *offsetMBB;
19607 MachineBasicBlock *endMBB;
19609 unsigned OffsetDestReg = 0; // Argument address computed by offsetMBB
19610 unsigned OverflowDestReg = 0; // Argument address computed by overflowMBB
19611 unsigned OffsetReg = 0;
19613 if (!UseGPOffset && !UseFPOffset) {
19614 // If we only pull from the overflow region, we don't create a branch.
19615 // We don't need to alter control flow.
19616 OffsetDestReg = 0; // unused
19617 OverflowDestReg = DestReg;
19619 offsetMBB = nullptr;
19620 overflowMBB = thisMBB;
19623 // First emit code to check if gp_offset (or fp_offset) is below the bound.
19624 // If so, pull the argument from reg_save_area. (branch to offsetMBB)
19625 // If not, pull from overflow_area. (branch to overflowMBB)
19630 // offsetMBB overflowMBB
19635 // Registers for the PHI in endMBB
19636 OffsetDestReg = MRI.createVirtualRegister(AddrRegClass);
19637 OverflowDestReg = MRI.createVirtualRegister(AddrRegClass);
19639 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
19640 MachineFunction *MF = MBB->getParent();
19641 overflowMBB = MF->CreateMachineBasicBlock(LLVM_BB);
19642 offsetMBB = MF->CreateMachineBasicBlock(LLVM_BB);
19643 endMBB = MF->CreateMachineBasicBlock(LLVM_BB);
19645 MachineFunction::iterator MBBIter = MBB;
19648 // Insert the new basic blocks
19649 MF->insert(MBBIter, offsetMBB);
19650 MF->insert(MBBIter, overflowMBB);
19651 MF->insert(MBBIter, endMBB);
19653 // Transfer the remainder of MBB and its successor edges to endMBB.
19654 endMBB->splice(endMBB->begin(), thisMBB,
19655 std::next(MachineBasicBlock::iterator(MI)), thisMBB->end());
19656 endMBB->transferSuccessorsAndUpdatePHIs(thisMBB);
19658 // Make offsetMBB and overflowMBB successors of thisMBB
19659 thisMBB->addSuccessor(offsetMBB);
19660 thisMBB->addSuccessor(overflowMBB);
19662 // endMBB is a successor of both offsetMBB and overflowMBB
19663 offsetMBB->addSuccessor(endMBB);
19664 overflowMBB->addSuccessor(endMBB);
19666 // Load the offset value into a register
19667 OffsetReg = MRI.createVirtualRegister(OffsetRegClass);
19668 BuildMI(thisMBB, DL, TII->get(X86::MOV32rm), OffsetReg)
19672 .addDisp(Disp, UseFPOffset ? 4 : 0)
19673 .addOperand(Segment)
19674 .setMemRefs(MMOBegin, MMOEnd);
19676 // Check if there is enough room left to pull this argument.
19677 BuildMI(thisMBB, DL, TII->get(X86::CMP32ri))
19679 .addImm(MaxOffset + 8 - ArgSizeA8);
19681 // Branch to "overflowMBB" if offset >= max
19682 // Fall through to "offsetMBB" otherwise
19683 BuildMI(thisMBB, DL, TII->get(X86::GetCondBranchFromCond(X86::COND_AE)))
19684 .addMBB(overflowMBB);
19687 // In offsetMBB, emit code to use the reg_save_area.
19689 assert(OffsetReg != 0);
19691 // Read the reg_save_area address.
19692 unsigned RegSaveReg = MRI.createVirtualRegister(AddrRegClass);
19693 BuildMI(offsetMBB, DL, TII->get(X86::MOV64rm), RegSaveReg)
19698 .addOperand(Segment)
19699 .setMemRefs(MMOBegin, MMOEnd);
19701 // Zero-extend the offset
19702 unsigned OffsetReg64 = MRI.createVirtualRegister(AddrRegClass);
19703 BuildMI(offsetMBB, DL, TII->get(X86::SUBREG_TO_REG), OffsetReg64)
19706 .addImm(X86::sub_32bit);
19708 // Add the offset to the reg_save_area to get the final address.
19709 BuildMI(offsetMBB, DL, TII->get(X86::ADD64rr), OffsetDestReg)
19710 .addReg(OffsetReg64)
19711 .addReg(RegSaveReg);
19713 // Compute the offset for the next argument
19714 unsigned NextOffsetReg = MRI.createVirtualRegister(OffsetRegClass);
19715 BuildMI(offsetMBB, DL, TII->get(X86::ADD32ri), NextOffsetReg)
19717 .addImm(UseFPOffset ? 16 : 8);
19719 // Store it back into the va_list.
19720 BuildMI(offsetMBB, DL, TII->get(X86::MOV32mr))
19724 .addDisp(Disp, UseFPOffset ? 4 : 0)
19725 .addOperand(Segment)
19726 .addReg(NextOffsetReg)
19727 .setMemRefs(MMOBegin, MMOEnd);
19730 BuildMI(offsetMBB, DL, TII->get(X86::JMP_1))
19735 // Emit code to use overflow area
19738 // Load the overflow_area address into a register.
19739 unsigned OverflowAddrReg = MRI.createVirtualRegister(AddrRegClass);
19740 BuildMI(overflowMBB, DL, TII->get(X86::MOV64rm), OverflowAddrReg)
19745 .addOperand(Segment)
19746 .setMemRefs(MMOBegin, MMOEnd);
19748 // If we need to align it, do so. Otherwise, just copy the address
19749 // to OverflowDestReg.
19751 // Align the overflow address
19752 assert((Align & (Align-1)) == 0 && "Alignment must be a power of 2");
19753 unsigned TmpReg = MRI.createVirtualRegister(AddrRegClass);
19755 // aligned_addr = (addr + (align-1)) & ~(align-1)
19756 BuildMI(overflowMBB, DL, TII->get(X86::ADD64ri32), TmpReg)
19757 .addReg(OverflowAddrReg)
19760 BuildMI(overflowMBB, DL, TII->get(X86::AND64ri32), OverflowDestReg)
19762 .addImm(~(uint64_t)(Align-1));
19764 BuildMI(overflowMBB, DL, TII->get(TargetOpcode::COPY), OverflowDestReg)
19765 .addReg(OverflowAddrReg);
19768 // Compute the next overflow address after this argument.
19769 // (the overflow address should be kept 8-byte aligned)
19770 unsigned NextAddrReg = MRI.createVirtualRegister(AddrRegClass);
19771 BuildMI(overflowMBB, DL, TII->get(X86::ADD64ri32), NextAddrReg)
19772 .addReg(OverflowDestReg)
19773 .addImm(ArgSizeA8);
19775 // Store the new overflow address.
19776 BuildMI(overflowMBB, DL, TII->get(X86::MOV64mr))
19781 .addOperand(Segment)
19782 .addReg(NextAddrReg)
19783 .setMemRefs(MMOBegin, MMOEnd);
19785 // If we branched, emit the PHI to the front of endMBB.
19787 BuildMI(*endMBB, endMBB->begin(), DL,
19788 TII->get(X86::PHI), DestReg)
19789 .addReg(OffsetDestReg).addMBB(offsetMBB)
19790 .addReg(OverflowDestReg).addMBB(overflowMBB);
19793 // Erase the pseudo instruction
19794 MI->eraseFromParent();
19799 MachineBasicBlock *
19800 X86TargetLowering::EmitVAStartSaveXMMRegsWithCustomInserter(
19802 MachineBasicBlock *MBB) const {
19803 // Emit code to save XMM registers to the stack. The ABI says that the
19804 // number of registers to save is given in %al, so it's theoretically
19805 // possible to do an indirect jump trick to avoid saving all of them,
19806 // however this code takes a simpler approach and just executes all
19807 // of the stores if %al is non-zero. It's less code, and it's probably
19808 // easier on the hardware branch predictor, and stores aren't all that
19809 // expensive anyway.
19811 // Create the new basic blocks. One block contains all the XMM stores,
19812 // and one block is the final destination regardless of whether any
19813 // stores were performed.
19814 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
19815 MachineFunction *F = MBB->getParent();
19816 MachineFunction::iterator MBBIter = MBB;
19818 MachineBasicBlock *XMMSaveMBB = F->CreateMachineBasicBlock(LLVM_BB);
19819 MachineBasicBlock *EndMBB = F->CreateMachineBasicBlock(LLVM_BB);
19820 F->insert(MBBIter, XMMSaveMBB);
19821 F->insert(MBBIter, EndMBB);
19823 // Transfer the remainder of MBB and its successor edges to EndMBB.
19824 EndMBB->splice(EndMBB->begin(), MBB,
19825 std::next(MachineBasicBlock::iterator(MI)), MBB->end());
19826 EndMBB->transferSuccessorsAndUpdatePHIs(MBB);
19828 // The original block will now fall through to the XMM save block.
19829 MBB->addSuccessor(XMMSaveMBB);
19830 // The XMMSaveMBB will fall through to the end block.
19831 XMMSaveMBB->addSuccessor(EndMBB);
19833 // Now add the instructions.
19834 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
19835 DebugLoc DL = MI->getDebugLoc();
19837 unsigned CountReg = MI->getOperand(0).getReg();
19838 int64_t RegSaveFrameIndex = MI->getOperand(1).getImm();
19839 int64_t VarArgsFPOffset = MI->getOperand(2).getImm();
19841 if (!Subtarget->isTargetWin64()) {
19842 // If %al is 0, branch around the XMM save block.
19843 BuildMI(MBB, DL, TII->get(X86::TEST8rr)).addReg(CountReg).addReg(CountReg);
19844 BuildMI(MBB, DL, TII->get(X86::JE_1)).addMBB(EndMBB);
19845 MBB->addSuccessor(EndMBB);
19848 // Make sure the last operand is EFLAGS, which gets clobbered by the branch
19849 // that was just emitted, but clearly shouldn't be "saved".
19850 assert((MI->getNumOperands() <= 3 ||
19851 !MI->getOperand(MI->getNumOperands() - 1).isReg() ||
19852 MI->getOperand(MI->getNumOperands() - 1).getReg() == X86::EFLAGS)
19853 && "Expected last argument to be EFLAGS");
19854 unsigned MOVOpc = Subtarget->hasFp256() ? X86::VMOVAPSmr : X86::MOVAPSmr;
19855 // In the XMM save block, save all the XMM argument registers.
19856 for (int i = 3, e = MI->getNumOperands() - 1; i != e; ++i) {
19857 int64_t Offset = (i - 3) * 16 + VarArgsFPOffset;
19858 MachineMemOperand *MMO =
19859 F->getMachineMemOperand(
19860 MachinePointerInfo::getFixedStack(RegSaveFrameIndex, Offset),
19861 MachineMemOperand::MOStore,
19862 /*Size=*/16, /*Align=*/16);
19863 BuildMI(XMMSaveMBB, DL, TII->get(MOVOpc))
19864 .addFrameIndex(RegSaveFrameIndex)
19865 .addImm(/*Scale=*/1)
19866 .addReg(/*IndexReg=*/0)
19867 .addImm(/*Disp=*/Offset)
19868 .addReg(/*Segment=*/0)
19869 .addReg(MI->getOperand(i).getReg())
19870 .addMemOperand(MMO);
19873 MI->eraseFromParent(); // The pseudo instruction is gone now.
19878 // The EFLAGS operand of SelectItr might be missing a kill marker
19879 // because there were multiple uses of EFLAGS, and ISel didn't know
19880 // which to mark. Figure out whether SelectItr should have had a
19881 // kill marker, and set it if it should. Returns the correct kill
19883 static bool checkAndUpdateEFLAGSKill(MachineBasicBlock::iterator SelectItr,
19884 MachineBasicBlock* BB,
19885 const TargetRegisterInfo* TRI) {
19886 // Scan forward through BB for a use/def of EFLAGS.
19887 MachineBasicBlock::iterator miI(std::next(SelectItr));
19888 for (MachineBasicBlock::iterator miE = BB->end(); miI != miE; ++miI) {
19889 const MachineInstr& mi = *miI;
19890 if (mi.readsRegister(X86::EFLAGS))
19892 if (mi.definesRegister(X86::EFLAGS))
19893 break; // Should have kill-flag - update below.
19896 // If we hit the end of the block, check whether EFLAGS is live into a
19898 if (miI == BB->end()) {
19899 for (MachineBasicBlock::succ_iterator sItr = BB->succ_begin(),
19900 sEnd = BB->succ_end();
19901 sItr != sEnd; ++sItr) {
19902 MachineBasicBlock* succ = *sItr;
19903 if (succ->isLiveIn(X86::EFLAGS))
19908 // We found a def, or hit the end of the basic block and EFLAGS wasn't live
19909 // out. SelectMI should have a kill flag on EFLAGS.
19910 SelectItr->addRegisterKilled(X86::EFLAGS, TRI);
19914 MachineBasicBlock *
19915 X86TargetLowering::EmitLoweredSelect(MachineInstr *MI,
19916 MachineBasicBlock *BB) const {
19917 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
19918 DebugLoc DL = MI->getDebugLoc();
19920 // To "insert" a SELECT_CC instruction, we actually have to insert the
19921 // diamond control-flow pattern. The incoming instruction knows the
19922 // destination vreg to set, the condition code register to branch on, the
19923 // true/false values to select between, and a branch opcode to use.
19924 const BasicBlock *LLVM_BB = BB->getBasicBlock();
19925 MachineFunction::iterator It = BB;
19931 // cmpTY ccX, r1, r2
19933 // fallthrough --> copy0MBB
19934 MachineBasicBlock *thisMBB = BB;
19935 MachineFunction *F = BB->getParent();
19937 // We also lower double CMOVs:
19938 // (CMOV (CMOV F, T, cc1), T, cc2)
19939 // to two successives branches. For that, we look for another CMOV as the
19940 // following instruction.
19942 // Without this, we would add a PHI between the two jumps, which ends up
19943 // creating a few copies all around. For instance, for
19945 // (sitofp (zext (fcmp une)))
19947 // we would generate:
19949 // ucomiss %xmm1, %xmm0
19950 // movss <1.0f>, %xmm0
19951 // movaps %xmm0, %xmm1
19953 // xorps %xmm1, %xmm1
19956 // movaps %xmm1, %xmm0
19960 // because this custom-inserter would have generated:
19972 // A: X = ...; Y = ...
19974 // C: Z = PHI [X, A], [Y, B]
19976 // E: PHI [X, C], [Z, D]
19978 // If we lower both CMOVs in a single step, we can instead generate:
19990 // A: X = ...; Y = ...
19992 // E: PHI [X, A], [X, C], [Y, D]
19994 // Which, in our sitofp/fcmp example, gives us something like:
19996 // ucomiss %xmm1, %xmm0
19997 // movss <1.0f>, %xmm0
20000 // xorps %xmm0, %xmm0
20004 MachineInstr *NextCMOV = nullptr;
20005 MachineBasicBlock::iterator NextMIIt =
20006 std::next(MachineBasicBlock::iterator(MI));
20007 if (NextMIIt != BB->end() && NextMIIt->getOpcode() == MI->getOpcode() &&
20008 NextMIIt->getOperand(2).getReg() == MI->getOperand(2).getReg() &&
20009 NextMIIt->getOperand(1).getReg() == MI->getOperand(0).getReg())
20010 NextCMOV = &*NextMIIt;
20012 MachineBasicBlock *jcc1MBB = nullptr;
20014 // If we have a double CMOV, we lower it to two successive branches to
20015 // the same block. EFLAGS is used by both, so mark it as live in the second.
20017 jcc1MBB = F->CreateMachineBasicBlock(LLVM_BB);
20018 F->insert(It, jcc1MBB);
20019 jcc1MBB->addLiveIn(X86::EFLAGS);
20022 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
20023 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
20024 F->insert(It, copy0MBB);
20025 F->insert(It, sinkMBB);
20027 // If the EFLAGS register isn't dead in the terminator, then claim that it's
20028 // live into the sink and copy blocks.
20029 const TargetRegisterInfo *TRI = Subtarget->getRegisterInfo();
20031 MachineInstr *LastEFLAGSUser = NextCMOV ? NextCMOV : MI;
20032 if (!LastEFLAGSUser->killsRegister(X86::EFLAGS) &&
20033 !checkAndUpdateEFLAGSKill(LastEFLAGSUser, BB, TRI)) {
20034 copy0MBB->addLiveIn(X86::EFLAGS);
20035 sinkMBB->addLiveIn(X86::EFLAGS);
20038 // Transfer the remainder of BB and its successor edges to sinkMBB.
20039 sinkMBB->splice(sinkMBB->begin(), BB,
20040 std::next(MachineBasicBlock::iterator(MI)), BB->end());
20041 sinkMBB->transferSuccessorsAndUpdatePHIs(BB);
20043 // Add the true and fallthrough blocks as its successors.
20045 // The fallthrough block may be jcc1MBB, if we have a double CMOV.
20046 BB->addSuccessor(jcc1MBB);
20048 // In that case, jcc1MBB will itself fallthrough the copy0MBB, and
20049 // jump to the sinkMBB.
20050 jcc1MBB->addSuccessor(copy0MBB);
20051 jcc1MBB->addSuccessor(sinkMBB);
20053 BB->addSuccessor(copy0MBB);
20056 // The true block target of the first (or only) branch is always sinkMBB.
20057 BB->addSuccessor(sinkMBB);
20059 // Create the conditional branch instruction.
20061 X86::GetCondBranchFromCond((X86::CondCode)MI->getOperand(3).getImm());
20062 BuildMI(BB, DL, TII->get(Opc)).addMBB(sinkMBB);
20065 unsigned Opc2 = X86::GetCondBranchFromCond(
20066 (X86::CondCode)NextCMOV->getOperand(3).getImm());
20067 BuildMI(jcc1MBB, DL, TII->get(Opc2)).addMBB(sinkMBB);
20071 // %FalseValue = ...
20072 // # fallthrough to sinkMBB
20073 copy0MBB->addSuccessor(sinkMBB);
20076 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
20078 MachineInstrBuilder MIB =
20079 BuildMI(*sinkMBB, sinkMBB->begin(), DL, TII->get(X86::PHI),
20080 MI->getOperand(0).getReg())
20081 .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB)
20082 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
20084 // If we have a double CMOV, the second Jcc provides the same incoming
20085 // value as the first Jcc (the True operand of the SELECT_CC/CMOV nodes).
20087 MIB.addReg(MI->getOperand(2).getReg()).addMBB(jcc1MBB);
20088 // Copy the PHI result to the register defined by the second CMOV.
20089 BuildMI(*sinkMBB, std::next(MachineBasicBlock::iterator(MIB.getInstr())),
20090 DL, TII->get(TargetOpcode::COPY), NextCMOV->getOperand(0).getReg())
20091 .addReg(MI->getOperand(0).getReg());
20092 NextCMOV->eraseFromParent();
20095 MI->eraseFromParent(); // The pseudo instruction is gone now.
20099 MachineBasicBlock *
20100 X86TargetLowering::EmitLoweredSegAlloca(MachineInstr *MI,
20101 MachineBasicBlock *BB) const {
20102 MachineFunction *MF = BB->getParent();
20103 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
20104 DebugLoc DL = MI->getDebugLoc();
20105 const BasicBlock *LLVM_BB = BB->getBasicBlock();
20107 assert(MF->shouldSplitStack());
20109 const bool Is64Bit = Subtarget->is64Bit();
20110 const bool IsLP64 = Subtarget->isTarget64BitLP64();
20112 const unsigned TlsReg = Is64Bit ? X86::FS : X86::GS;
20113 const unsigned TlsOffset = IsLP64 ? 0x70 : Is64Bit ? 0x40 : 0x30;
20116 // ... [Till the alloca]
20117 // If stacklet is not large enough, jump to mallocMBB
20120 // Allocate by subtracting from RSP
20121 // Jump to continueMBB
20124 // Allocate by call to runtime
20128 // [rest of original BB]
20131 MachineBasicBlock *mallocMBB = MF->CreateMachineBasicBlock(LLVM_BB);
20132 MachineBasicBlock *bumpMBB = MF->CreateMachineBasicBlock(LLVM_BB);
20133 MachineBasicBlock *continueMBB = MF->CreateMachineBasicBlock(LLVM_BB);
20135 MachineRegisterInfo &MRI = MF->getRegInfo();
20136 const TargetRegisterClass *AddrRegClass =
20137 getRegClassFor(getPointerTy(MF->getDataLayout()));
20139 unsigned mallocPtrVReg = MRI.createVirtualRegister(AddrRegClass),
20140 bumpSPPtrVReg = MRI.createVirtualRegister(AddrRegClass),
20141 tmpSPVReg = MRI.createVirtualRegister(AddrRegClass),
20142 SPLimitVReg = MRI.createVirtualRegister(AddrRegClass),
20143 sizeVReg = MI->getOperand(1).getReg(),
20144 physSPReg = IsLP64 || Subtarget->isTargetNaCl64() ? X86::RSP : X86::ESP;
20146 MachineFunction::iterator MBBIter = BB;
20149 MF->insert(MBBIter, bumpMBB);
20150 MF->insert(MBBIter, mallocMBB);
20151 MF->insert(MBBIter, continueMBB);
20153 continueMBB->splice(continueMBB->begin(), BB,
20154 std::next(MachineBasicBlock::iterator(MI)), BB->end());
20155 continueMBB->transferSuccessorsAndUpdatePHIs(BB);
20157 // Add code to the main basic block to check if the stack limit has been hit,
20158 // and if so, jump to mallocMBB otherwise to bumpMBB.
20159 BuildMI(BB, DL, TII->get(TargetOpcode::COPY), tmpSPVReg).addReg(physSPReg);
20160 BuildMI(BB, DL, TII->get(IsLP64 ? X86::SUB64rr:X86::SUB32rr), SPLimitVReg)
20161 .addReg(tmpSPVReg).addReg(sizeVReg);
20162 BuildMI(BB, DL, TII->get(IsLP64 ? X86::CMP64mr:X86::CMP32mr))
20163 .addReg(0).addImm(1).addReg(0).addImm(TlsOffset).addReg(TlsReg)
20164 .addReg(SPLimitVReg);
20165 BuildMI(BB, DL, TII->get(X86::JG_1)).addMBB(mallocMBB);
20167 // bumpMBB simply decreases the stack pointer, since we know the current
20168 // stacklet has enough space.
20169 BuildMI(bumpMBB, DL, TII->get(TargetOpcode::COPY), physSPReg)
20170 .addReg(SPLimitVReg);
20171 BuildMI(bumpMBB, DL, TII->get(TargetOpcode::COPY), bumpSPPtrVReg)
20172 .addReg(SPLimitVReg);
20173 BuildMI(bumpMBB, DL, TII->get(X86::JMP_1)).addMBB(continueMBB);
20175 // Calls into a routine in libgcc to allocate more space from the heap.
20176 const uint32_t *RegMask =
20177 Subtarget->getRegisterInfo()->getCallPreservedMask(*MF, CallingConv::C);
20179 BuildMI(mallocMBB, DL, TII->get(X86::MOV64rr), X86::RDI)
20181 BuildMI(mallocMBB, DL, TII->get(X86::CALL64pcrel32))
20182 .addExternalSymbol("__morestack_allocate_stack_space")
20183 .addRegMask(RegMask)
20184 .addReg(X86::RDI, RegState::Implicit)
20185 .addReg(X86::RAX, RegState::ImplicitDefine);
20186 } else if (Is64Bit) {
20187 BuildMI(mallocMBB, DL, TII->get(X86::MOV32rr), X86::EDI)
20189 BuildMI(mallocMBB, DL, TII->get(X86::CALL64pcrel32))
20190 .addExternalSymbol("__morestack_allocate_stack_space")
20191 .addRegMask(RegMask)
20192 .addReg(X86::EDI, RegState::Implicit)
20193 .addReg(X86::EAX, RegState::ImplicitDefine);
20195 BuildMI(mallocMBB, DL, TII->get(X86::SUB32ri), physSPReg).addReg(physSPReg)
20197 BuildMI(mallocMBB, DL, TII->get(X86::PUSH32r)).addReg(sizeVReg);
20198 BuildMI(mallocMBB, DL, TII->get(X86::CALLpcrel32))
20199 .addExternalSymbol("__morestack_allocate_stack_space")
20200 .addRegMask(RegMask)
20201 .addReg(X86::EAX, RegState::ImplicitDefine);
20205 BuildMI(mallocMBB, DL, TII->get(X86::ADD32ri), physSPReg).addReg(physSPReg)
20208 BuildMI(mallocMBB, DL, TII->get(TargetOpcode::COPY), mallocPtrVReg)
20209 .addReg(IsLP64 ? X86::RAX : X86::EAX);
20210 BuildMI(mallocMBB, DL, TII->get(X86::JMP_1)).addMBB(continueMBB);
20212 // Set up the CFG correctly.
20213 BB->addSuccessor(bumpMBB);
20214 BB->addSuccessor(mallocMBB);
20215 mallocMBB->addSuccessor(continueMBB);
20216 bumpMBB->addSuccessor(continueMBB);
20218 // Take care of the PHI nodes.
20219 BuildMI(*continueMBB, continueMBB->begin(), DL, TII->get(X86::PHI),
20220 MI->getOperand(0).getReg())
20221 .addReg(mallocPtrVReg).addMBB(mallocMBB)
20222 .addReg(bumpSPPtrVReg).addMBB(bumpMBB);
20224 // Delete the original pseudo instruction.
20225 MI->eraseFromParent();
20228 return continueMBB;
20231 MachineBasicBlock *
20232 X86TargetLowering::EmitLoweredWinAlloca(MachineInstr *MI,
20233 MachineBasicBlock *BB) const {
20234 DebugLoc DL = MI->getDebugLoc();
20236 assert(!Subtarget->isTargetMachO());
20238 Subtarget->getFrameLowering()->emitStackProbeCall(*BB->getParent(), *BB, MI,
20241 MI->eraseFromParent(); // The pseudo instruction is gone now.
20245 MachineBasicBlock *
20246 X86TargetLowering::EmitLoweredTLSCall(MachineInstr *MI,
20247 MachineBasicBlock *BB) const {
20248 // This is pretty easy. We're taking the value that we received from
20249 // our load from the relocation, sticking it in either RDI (x86-64)
20250 // or EAX and doing an indirect call. The return value will then
20251 // be in the normal return register.
20252 MachineFunction *F = BB->getParent();
20253 const X86InstrInfo *TII = Subtarget->getInstrInfo();
20254 DebugLoc DL = MI->getDebugLoc();
20256 assert(Subtarget->isTargetDarwin() && "Darwin only instr emitted?");
20257 assert(MI->getOperand(3).isGlobal() && "This should be a global");
20259 // Get a register mask for the lowered call.
20260 // FIXME: The 32-bit calls have non-standard calling conventions. Use a
20261 // proper register mask.
20262 const uint32_t *RegMask =
20263 Subtarget->getRegisterInfo()->getCallPreservedMask(*F, CallingConv::C);
20264 if (Subtarget->is64Bit()) {
20265 MachineInstrBuilder MIB = BuildMI(*BB, MI, DL,
20266 TII->get(X86::MOV64rm), X86::RDI)
20268 .addImm(0).addReg(0)
20269 .addGlobalAddress(MI->getOperand(3).getGlobal(), 0,
20270 MI->getOperand(3).getTargetFlags())
20272 MIB = BuildMI(*BB, MI, DL, TII->get(X86::CALL64m));
20273 addDirectMem(MIB, X86::RDI);
20274 MIB.addReg(X86::RAX, RegState::ImplicitDefine).addRegMask(RegMask);
20275 } else if (F->getTarget().getRelocationModel() != Reloc::PIC_) {
20276 MachineInstrBuilder MIB = BuildMI(*BB, MI, DL,
20277 TII->get(X86::MOV32rm), X86::EAX)
20279 .addImm(0).addReg(0)
20280 .addGlobalAddress(MI->getOperand(3).getGlobal(), 0,
20281 MI->getOperand(3).getTargetFlags())
20283 MIB = BuildMI(*BB, MI, DL, TII->get(X86::CALL32m));
20284 addDirectMem(MIB, X86::EAX);
20285 MIB.addReg(X86::EAX, RegState::ImplicitDefine).addRegMask(RegMask);
20287 MachineInstrBuilder MIB = BuildMI(*BB, MI, DL,
20288 TII->get(X86::MOV32rm), X86::EAX)
20289 .addReg(TII->getGlobalBaseReg(F))
20290 .addImm(0).addReg(0)
20291 .addGlobalAddress(MI->getOperand(3).getGlobal(), 0,
20292 MI->getOperand(3).getTargetFlags())
20294 MIB = BuildMI(*BB, MI, DL, TII->get(X86::CALL32m));
20295 addDirectMem(MIB, X86::EAX);
20296 MIB.addReg(X86::EAX, RegState::ImplicitDefine).addRegMask(RegMask);
20299 MI->eraseFromParent(); // The pseudo instruction is gone now.
20303 MachineBasicBlock *
20304 X86TargetLowering::emitEHSjLjSetJmp(MachineInstr *MI,
20305 MachineBasicBlock *MBB) const {
20306 DebugLoc DL = MI->getDebugLoc();
20307 MachineFunction *MF = MBB->getParent();
20308 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
20309 MachineRegisterInfo &MRI = MF->getRegInfo();
20311 const BasicBlock *BB = MBB->getBasicBlock();
20312 MachineFunction::iterator I = MBB;
20315 // Memory Reference
20316 MachineInstr::mmo_iterator MMOBegin = MI->memoperands_begin();
20317 MachineInstr::mmo_iterator MMOEnd = MI->memoperands_end();
20320 unsigned MemOpndSlot = 0;
20322 unsigned CurOp = 0;
20324 DstReg = MI->getOperand(CurOp++).getReg();
20325 const TargetRegisterClass *RC = MRI.getRegClass(DstReg);
20326 assert(RC->hasType(MVT::i32) && "Invalid destination!");
20327 unsigned mainDstReg = MRI.createVirtualRegister(RC);
20328 unsigned restoreDstReg = MRI.createVirtualRegister(RC);
20330 MemOpndSlot = CurOp;
20332 MVT PVT = getPointerTy(MF->getDataLayout());
20333 assert((PVT == MVT::i64 || PVT == MVT::i32) &&
20334 "Invalid Pointer Size!");
20336 // For v = setjmp(buf), we generate
20339 // buf[LabelOffset] = restoreMBB
20340 // SjLjSetup restoreMBB
20346 // v = phi(main, restore)
20349 // if base pointer being used, load it from frame
20352 MachineBasicBlock *thisMBB = MBB;
20353 MachineBasicBlock *mainMBB = MF->CreateMachineBasicBlock(BB);
20354 MachineBasicBlock *sinkMBB = MF->CreateMachineBasicBlock(BB);
20355 MachineBasicBlock *restoreMBB = MF->CreateMachineBasicBlock(BB);
20356 MF->insert(I, mainMBB);
20357 MF->insert(I, sinkMBB);
20358 MF->push_back(restoreMBB);
20360 MachineInstrBuilder MIB;
20362 // Transfer the remainder of BB and its successor edges to sinkMBB.
20363 sinkMBB->splice(sinkMBB->begin(), MBB,
20364 std::next(MachineBasicBlock::iterator(MI)), MBB->end());
20365 sinkMBB->transferSuccessorsAndUpdatePHIs(MBB);
20368 unsigned PtrStoreOpc = 0;
20369 unsigned LabelReg = 0;
20370 const int64_t LabelOffset = 1 * PVT.getStoreSize();
20371 Reloc::Model RM = MF->getTarget().getRelocationModel();
20372 bool UseImmLabel = (MF->getTarget().getCodeModel() == CodeModel::Small) &&
20373 (RM == Reloc::Static || RM == Reloc::DynamicNoPIC);
20375 // Prepare IP either in reg or imm.
20376 if (!UseImmLabel) {
20377 PtrStoreOpc = (PVT == MVT::i64) ? X86::MOV64mr : X86::MOV32mr;
20378 const TargetRegisterClass *PtrRC = getRegClassFor(PVT);
20379 LabelReg = MRI.createVirtualRegister(PtrRC);
20380 if (Subtarget->is64Bit()) {
20381 MIB = BuildMI(*thisMBB, MI, DL, TII->get(X86::LEA64r), LabelReg)
20385 .addMBB(restoreMBB)
20388 const X86InstrInfo *XII = static_cast<const X86InstrInfo*>(TII);
20389 MIB = BuildMI(*thisMBB, MI, DL, TII->get(X86::LEA32r), LabelReg)
20390 .addReg(XII->getGlobalBaseReg(MF))
20393 .addMBB(restoreMBB, Subtarget->ClassifyBlockAddressReference())
20397 PtrStoreOpc = (PVT == MVT::i64) ? X86::MOV64mi32 : X86::MOV32mi;
20399 MIB = BuildMI(*thisMBB, MI, DL, TII->get(PtrStoreOpc));
20400 for (unsigned i = 0; i < X86::AddrNumOperands; ++i) {
20401 if (i == X86::AddrDisp)
20402 MIB.addDisp(MI->getOperand(MemOpndSlot + i), LabelOffset);
20404 MIB.addOperand(MI->getOperand(MemOpndSlot + i));
20407 MIB.addReg(LabelReg);
20409 MIB.addMBB(restoreMBB);
20410 MIB.setMemRefs(MMOBegin, MMOEnd);
20412 MIB = BuildMI(*thisMBB, MI, DL, TII->get(X86::EH_SjLj_Setup))
20413 .addMBB(restoreMBB);
20415 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
20416 MIB.addRegMask(RegInfo->getNoPreservedMask());
20417 thisMBB->addSuccessor(mainMBB);
20418 thisMBB->addSuccessor(restoreMBB);
20422 BuildMI(mainMBB, DL, TII->get(X86::MOV32r0), mainDstReg);
20423 mainMBB->addSuccessor(sinkMBB);
20426 BuildMI(*sinkMBB, sinkMBB->begin(), DL,
20427 TII->get(X86::PHI), DstReg)
20428 .addReg(mainDstReg).addMBB(mainMBB)
20429 .addReg(restoreDstReg).addMBB(restoreMBB);
20432 if (RegInfo->hasBasePointer(*MF)) {
20433 const bool Uses64BitFramePtr =
20434 Subtarget->isTarget64BitLP64() || Subtarget->isTargetNaCl64();
20435 X86MachineFunctionInfo *X86FI = MF->getInfo<X86MachineFunctionInfo>();
20436 X86FI->setRestoreBasePointer(MF);
20437 unsigned FramePtr = RegInfo->getFrameRegister(*MF);
20438 unsigned BasePtr = RegInfo->getBaseRegister();
20439 unsigned Opm = Uses64BitFramePtr ? X86::MOV64rm : X86::MOV32rm;
20440 addRegOffset(BuildMI(restoreMBB, DL, TII->get(Opm), BasePtr),
20441 FramePtr, true, X86FI->getRestoreBasePointerOffset())
20442 .setMIFlag(MachineInstr::FrameSetup);
20444 BuildMI(restoreMBB, DL, TII->get(X86::MOV32ri), restoreDstReg).addImm(1);
20445 BuildMI(restoreMBB, DL, TII->get(X86::JMP_1)).addMBB(sinkMBB);
20446 restoreMBB->addSuccessor(sinkMBB);
20448 MI->eraseFromParent();
20452 MachineBasicBlock *
20453 X86TargetLowering::emitEHSjLjLongJmp(MachineInstr *MI,
20454 MachineBasicBlock *MBB) const {
20455 DebugLoc DL = MI->getDebugLoc();
20456 MachineFunction *MF = MBB->getParent();
20457 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
20458 MachineRegisterInfo &MRI = MF->getRegInfo();
20460 // Memory Reference
20461 MachineInstr::mmo_iterator MMOBegin = MI->memoperands_begin();
20462 MachineInstr::mmo_iterator MMOEnd = MI->memoperands_end();
20464 MVT PVT = getPointerTy(MF->getDataLayout());
20465 assert((PVT == MVT::i64 || PVT == MVT::i32) &&
20466 "Invalid Pointer Size!");
20468 const TargetRegisterClass *RC =
20469 (PVT == MVT::i64) ? &X86::GR64RegClass : &X86::GR32RegClass;
20470 unsigned Tmp = MRI.createVirtualRegister(RC);
20471 // Since FP is only updated here but NOT referenced, it's treated as GPR.
20472 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
20473 unsigned FP = (PVT == MVT::i64) ? X86::RBP : X86::EBP;
20474 unsigned SP = RegInfo->getStackRegister();
20476 MachineInstrBuilder MIB;
20478 const int64_t LabelOffset = 1 * PVT.getStoreSize();
20479 const int64_t SPOffset = 2 * PVT.getStoreSize();
20481 unsigned PtrLoadOpc = (PVT == MVT::i64) ? X86::MOV64rm : X86::MOV32rm;
20482 unsigned IJmpOpc = (PVT == MVT::i64) ? X86::JMP64r : X86::JMP32r;
20485 MIB = BuildMI(*MBB, MI, DL, TII->get(PtrLoadOpc), FP);
20486 for (unsigned i = 0; i < X86::AddrNumOperands; ++i)
20487 MIB.addOperand(MI->getOperand(i));
20488 MIB.setMemRefs(MMOBegin, MMOEnd);
20490 MIB = BuildMI(*MBB, MI, DL, TII->get(PtrLoadOpc), Tmp);
20491 for (unsigned i = 0; i < X86::AddrNumOperands; ++i) {
20492 if (i == X86::AddrDisp)
20493 MIB.addDisp(MI->getOperand(i), LabelOffset);
20495 MIB.addOperand(MI->getOperand(i));
20497 MIB.setMemRefs(MMOBegin, MMOEnd);
20499 MIB = BuildMI(*MBB, MI, DL, TII->get(PtrLoadOpc), SP);
20500 for (unsigned i = 0; i < X86::AddrNumOperands; ++i) {
20501 if (i == X86::AddrDisp)
20502 MIB.addDisp(MI->getOperand(i), SPOffset);
20504 MIB.addOperand(MI->getOperand(i));
20506 MIB.setMemRefs(MMOBegin, MMOEnd);
20508 BuildMI(*MBB, MI, DL, TII->get(IJmpOpc)).addReg(Tmp);
20510 MI->eraseFromParent();
20514 // Replace 213-type (isel default) FMA3 instructions with 231-type for
20515 // accumulator loops. Writing back to the accumulator allows the coalescer
20516 // to remove extra copies in the loop.
20517 // FIXME: Do this on AVX512. We don't support 231 variants yet (PR23937).
20518 MachineBasicBlock *
20519 X86TargetLowering::emitFMA3Instr(MachineInstr *MI,
20520 MachineBasicBlock *MBB) const {
20521 MachineOperand &AddendOp = MI->getOperand(3);
20523 // Bail out early if the addend isn't a register - we can't switch these.
20524 if (!AddendOp.isReg())
20527 MachineFunction &MF = *MBB->getParent();
20528 MachineRegisterInfo &MRI = MF.getRegInfo();
20530 // Check whether the addend is defined by a PHI:
20531 assert(MRI.hasOneDef(AddendOp.getReg()) && "Multiple defs in SSA?");
20532 MachineInstr &AddendDef = *MRI.def_instr_begin(AddendOp.getReg());
20533 if (!AddendDef.isPHI())
20536 // Look for the following pattern:
20538 // %addend = phi [%entry, 0], [%loop, %result]
20540 // %result<tied1> = FMA213 %m2<tied0>, %m1, %addend
20544 // %addend = phi [%entry, 0], [%loop, %result]
20546 // %result<tied1> = FMA231 %addend<tied0>, %m1, %m2
20548 for (unsigned i = 1, e = AddendDef.getNumOperands(); i < e; i += 2) {
20549 assert(AddendDef.getOperand(i).isReg());
20550 MachineOperand PHISrcOp = AddendDef.getOperand(i);
20551 MachineInstr &PHISrcInst = *MRI.def_instr_begin(PHISrcOp.getReg());
20552 if (&PHISrcInst == MI) {
20553 // Found a matching instruction.
20554 unsigned NewFMAOpc = 0;
20555 switch (MI->getOpcode()) {
20556 case X86::VFMADDPDr213r: NewFMAOpc = X86::VFMADDPDr231r; break;
20557 case X86::VFMADDPSr213r: NewFMAOpc = X86::VFMADDPSr231r; break;
20558 case X86::VFMADDSDr213r: NewFMAOpc = X86::VFMADDSDr231r; break;
20559 case X86::VFMADDSSr213r: NewFMAOpc = X86::VFMADDSSr231r; break;
20560 case X86::VFMSUBPDr213r: NewFMAOpc = X86::VFMSUBPDr231r; break;
20561 case X86::VFMSUBPSr213r: NewFMAOpc = X86::VFMSUBPSr231r; break;
20562 case X86::VFMSUBSDr213r: NewFMAOpc = X86::VFMSUBSDr231r; break;
20563 case X86::VFMSUBSSr213r: NewFMAOpc = X86::VFMSUBSSr231r; break;
20564 case X86::VFNMADDPDr213r: NewFMAOpc = X86::VFNMADDPDr231r; break;
20565 case X86::VFNMADDPSr213r: NewFMAOpc = X86::VFNMADDPSr231r; break;
20566 case X86::VFNMADDSDr213r: NewFMAOpc = X86::VFNMADDSDr231r; break;
20567 case X86::VFNMADDSSr213r: NewFMAOpc = X86::VFNMADDSSr231r; break;
20568 case X86::VFNMSUBPDr213r: NewFMAOpc = X86::VFNMSUBPDr231r; break;
20569 case X86::VFNMSUBPSr213r: NewFMAOpc = X86::VFNMSUBPSr231r; break;
20570 case X86::VFNMSUBSDr213r: NewFMAOpc = X86::VFNMSUBSDr231r; break;
20571 case X86::VFNMSUBSSr213r: NewFMAOpc = X86::VFNMSUBSSr231r; break;
20572 case X86::VFMADDSUBPDr213r: NewFMAOpc = X86::VFMADDSUBPDr231r; break;
20573 case X86::VFMADDSUBPSr213r: NewFMAOpc = X86::VFMADDSUBPSr231r; break;
20574 case X86::VFMSUBADDPDr213r: NewFMAOpc = X86::VFMSUBADDPDr231r; break;
20575 case X86::VFMSUBADDPSr213r: NewFMAOpc = X86::VFMSUBADDPSr231r; break;
20577 case X86::VFMADDPDr213rY: NewFMAOpc = X86::VFMADDPDr231rY; break;
20578 case X86::VFMADDPSr213rY: NewFMAOpc = X86::VFMADDPSr231rY; break;
20579 case X86::VFMSUBPDr213rY: NewFMAOpc = X86::VFMSUBPDr231rY; break;
20580 case X86::VFMSUBPSr213rY: NewFMAOpc = X86::VFMSUBPSr231rY; break;
20581 case X86::VFNMADDPDr213rY: NewFMAOpc = X86::VFNMADDPDr231rY; break;
20582 case X86::VFNMADDPSr213rY: NewFMAOpc = X86::VFNMADDPSr231rY; break;
20583 case X86::VFNMSUBPDr213rY: NewFMAOpc = X86::VFNMSUBPDr231rY; break;
20584 case X86::VFNMSUBPSr213rY: NewFMAOpc = X86::VFNMSUBPSr231rY; break;
20585 case X86::VFMADDSUBPDr213rY: NewFMAOpc = X86::VFMADDSUBPDr231rY; break;
20586 case X86::VFMADDSUBPSr213rY: NewFMAOpc = X86::VFMADDSUBPSr231rY; break;
20587 case X86::VFMSUBADDPDr213rY: NewFMAOpc = X86::VFMSUBADDPDr231rY; break;
20588 case X86::VFMSUBADDPSr213rY: NewFMAOpc = X86::VFMSUBADDPSr231rY; break;
20589 default: llvm_unreachable("Unrecognized FMA variant.");
20592 const TargetInstrInfo &TII = *Subtarget->getInstrInfo();
20593 MachineInstrBuilder MIB =
20594 BuildMI(MF, MI->getDebugLoc(), TII.get(NewFMAOpc))
20595 .addOperand(MI->getOperand(0))
20596 .addOperand(MI->getOperand(3))
20597 .addOperand(MI->getOperand(2))
20598 .addOperand(MI->getOperand(1));
20599 MBB->insert(MachineBasicBlock::iterator(MI), MIB);
20600 MI->eraseFromParent();
20607 MachineBasicBlock *
20608 X86TargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
20609 MachineBasicBlock *BB) const {
20610 switch (MI->getOpcode()) {
20611 default: llvm_unreachable("Unexpected instr type to insert");
20612 case X86::TAILJMPd64:
20613 case X86::TAILJMPr64:
20614 case X86::TAILJMPm64:
20615 case X86::TAILJMPd64_REX:
20616 case X86::TAILJMPr64_REX:
20617 case X86::TAILJMPm64_REX:
20618 llvm_unreachable("TAILJMP64 would not be touched here.");
20619 case X86::TCRETURNdi64:
20620 case X86::TCRETURNri64:
20621 case X86::TCRETURNmi64:
20623 case X86::WIN_ALLOCA:
20624 return EmitLoweredWinAlloca(MI, BB);
20625 case X86::SEG_ALLOCA_32:
20626 case X86::SEG_ALLOCA_64:
20627 return EmitLoweredSegAlloca(MI, BB);
20628 case X86::TLSCall_32:
20629 case X86::TLSCall_64:
20630 return EmitLoweredTLSCall(MI, BB);
20631 case X86::CMOV_GR8:
20632 case X86::CMOV_FR32:
20633 case X86::CMOV_FR64:
20634 case X86::CMOV_V4F32:
20635 case X86::CMOV_V2F64:
20636 case X86::CMOV_V2I64:
20637 case X86::CMOV_V8F32:
20638 case X86::CMOV_V4F64:
20639 case X86::CMOV_V4I64:
20640 case X86::CMOV_V16F32:
20641 case X86::CMOV_V8F64:
20642 case X86::CMOV_V8I64:
20643 case X86::CMOV_GR16:
20644 case X86::CMOV_GR32:
20645 case X86::CMOV_RFP32:
20646 case X86::CMOV_RFP64:
20647 case X86::CMOV_RFP80:
20648 case X86::CMOV_V8I1:
20649 case X86::CMOV_V16I1:
20650 case X86::CMOV_V32I1:
20651 case X86::CMOV_V64I1:
20652 return EmitLoweredSelect(MI, BB);
20654 case X86::FP32_TO_INT16_IN_MEM:
20655 case X86::FP32_TO_INT32_IN_MEM:
20656 case X86::FP32_TO_INT64_IN_MEM:
20657 case X86::FP64_TO_INT16_IN_MEM:
20658 case X86::FP64_TO_INT32_IN_MEM:
20659 case X86::FP64_TO_INT64_IN_MEM:
20660 case X86::FP80_TO_INT16_IN_MEM:
20661 case X86::FP80_TO_INT32_IN_MEM:
20662 case X86::FP80_TO_INT64_IN_MEM: {
20663 MachineFunction *F = BB->getParent();
20664 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
20665 DebugLoc DL = MI->getDebugLoc();
20667 // Change the floating point control register to use "round towards zero"
20668 // mode when truncating to an integer value.
20669 int CWFrameIdx = F->getFrameInfo()->CreateStackObject(2, 2, false);
20670 addFrameReference(BuildMI(*BB, MI, DL,
20671 TII->get(X86::FNSTCW16m)), CWFrameIdx);
20673 // Load the old value of the high byte of the control word...
20675 F->getRegInfo().createVirtualRegister(&X86::GR16RegClass);
20676 addFrameReference(BuildMI(*BB, MI, DL, TII->get(X86::MOV16rm), OldCW),
20679 // Set the high part to be round to zero...
20680 addFrameReference(BuildMI(*BB, MI, DL, TII->get(X86::MOV16mi)), CWFrameIdx)
20683 // Reload the modified control word now...
20684 addFrameReference(BuildMI(*BB, MI, DL,
20685 TII->get(X86::FLDCW16m)), CWFrameIdx);
20687 // Restore the memory image of control word to original value
20688 addFrameReference(BuildMI(*BB, MI, DL, TII->get(X86::MOV16mr)), CWFrameIdx)
20691 // Get the X86 opcode to use.
20693 switch (MI->getOpcode()) {
20694 default: llvm_unreachable("illegal opcode!");
20695 case X86::FP32_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m32; break;
20696 case X86::FP32_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m32; break;
20697 case X86::FP32_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m32; break;
20698 case X86::FP64_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m64; break;
20699 case X86::FP64_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m64; break;
20700 case X86::FP64_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m64; break;
20701 case X86::FP80_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m80; break;
20702 case X86::FP80_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m80; break;
20703 case X86::FP80_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m80; break;
20707 MachineOperand &Op = MI->getOperand(0);
20709 AM.BaseType = X86AddressMode::RegBase;
20710 AM.Base.Reg = Op.getReg();
20712 AM.BaseType = X86AddressMode::FrameIndexBase;
20713 AM.Base.FrameIndex = Op.getIndex();
20715 Op = MI->getOperand(1);
20717 AM.Scale = Op.getImm();
20718 Op = MI->getOperand(2);
20720 AM.IndexReg = Op.getImm();
20721 Op = MI->getOperand(3);
20722 if (Op.isGlobal()) {
20723 AM.GV = Op.getGlobal();
20725 AM.Disp = Op.getImm();
20727 addFullAddress(BuildMI(*BB, MI, DL, TII->get(Opc)), AM)
20728 .addReg(MI->getOperand(X86::AddrNumOperands).getReg());
20730 // Reload the original control word now.
20731 addFrameReference(BuildMI(*BB, MI, DL,
20732 TII->get(X86::FLDCW16m)), CWFrameIdx);
20734 MI->eraseFromParent(); // The pseudo instruction is gone now.
20737 // String/text processing lowering.
20738 case X86::PCMPISTRM128REG:
20739 case X86::VPCMPISTRM128REG:
20740 case X86::PCMPISTRM128MEM:
20741 case X86::VPCMPISTRM128MEM:
20742 case X86::PCMPESTRM128REG:
20743 case X86::VPCMPESTRM128REG:
20744 case X86::PCMPESTRM128MEM:
20745 case X86::VPCMPESTRM128MEM:
20746 assert(Subtarget->hasSSE42() &&
20747 "Target must have SSE4.2 or AVX features enabled");
20748 return EmitPCMPSTRM(MI, BB, Subtarget->getInstrInfo());
20750 // String/text processing lowering.
20751 case X86::PCMPISTRIREG:
20752 case X86::VPCMPISTRIREG:
20753 case X86::PCMPISTRIMEM:
20754 case X86::VPCMPISTRIMEM:
20755 case X86::PCMPESTRIREG:
20756 case X86::VPCMPESTRIREG:
20757 case X86::PCMPESTRIMEM:
20758 case X86::VPCMPESTRIMEM:
20759 assert(Subtarget->hasSSE42() &&
20760 "Target must have SSE4.2 or AVX features enabled");
20761 return EmitPCMPSTRI(MI, BB, Subtarget->getInstrInfo());
20763 // Thread synchronization.
20765 return EmitMonitor(MI, BB, Subtarget);
20769 return EmitXBegin(MI, BB, Subtarget->getInstrInfo());
20771 case X86::VASTART_SAVE_XMM_REGS:
20772 return EmitVAStartSaveXMMRegsWithCustomInserter(MI, BB);
20774 case X86::VAARG_64:
20775 return EmitVAARG64WithCustomInserter(MI, BB);
20777 case X86::EH_SjLj_SetJmp32:
20778 case X86::EH_SjLj_SetJmp64:
20779 return emitEHSjLjSetJmp(MI, BB);
20781 case X86::EH_SjLj_LongJmp32:
20782 case X86::EH_SjLj_LongJmp64:
20783 return emitEHSjLjLongJmp(MI, BB);
20785 case TargetOpcode::STATEPOINT:
20786 // As an implementation detail, STATEPOINT shares the STACKMAP format at
20787 // this point in the process. We diverge later.
20788 return emitPatchPoint(MI, BB);
20790 case TargetOpcode::STACKMAP:
20791 case TargetOpcode::PATCHPOINT:
20792 return emitPatchPoint(MI, BB);
20794 case X86::VFMADDPDr213r:
20795 case X86::VFMADDPSr213r:
20796 case X86::VFMADDSDr213r:
20797 case X86::VFMADDSSr213r:
20798 case X86::VFMSUBPDr213r:
20799 case X86::VFMSUBPSr213r:
20800 case X86::VFMSUBSDr213r:
20801 case X86::VFMSUBSSr213r:
20802 case X86::VFNMADDPDr213r:
20803 case X86::VFNMADDPSr213r:
20804 case X86::VFNMADDSDr213r:
20805 case X86::VFNMADDSSr213r:
20806 case X86::VFNMSUBPDr213r:
20807 case X86::VFNMSUBPSr213r:
20808 case X86::VFNMSUBSDr213r:
20809 case X86::VFNMSUBSSr213r:
20810 case X86::VFMADDSUBPDr213r:
20811 case X86::VFMADDSUBPSr213r:
20812 case X86::VFMSUBADDPDr213r:
20813 case X86::VFMSUBADDPSr213r:
20814 case X86::VFMADDPDr213rY:
20815 case X86::VFMADDPSr213rY:
20816 case X86::VFMSUBPDr213rY:
20817 case X86::VFMSUBPSr213rY:
20818 case X86::VFNMADDPDr213rY:
20819 case X86::VFNMADDPSr213rY:
20820 case X86::VFNMSUBPDr213rY:
20821 case X86::VFNMSUBPSr213rY:
20822 case X86::VFMADDSUBPDr213rY:
20823 case X86::VFMADDSUBPSr213rY:
20824 case X86::VFMSUBADDPDr213rY:
20825 case X86::VFMSUBADDPSr213rY:
20826 return emitFMA3Instr(MI, BB);
20830 //===----------------------------------------------------------------------===//
20831 // X86 Optimization Hooks
20832 //===----------------------------------------------------------------------===//
20834 void X86TargetLowering::computeKnownBitsForTargetNode(const SDValue Op,
20837 const SelectionDAG &DAG,
20838 unsigned Depth) const {
20839 unsigned BitWidth = KnownZero.getBitWidth();
20840 unsigned Opc = Op.getOpcode();
20841 assert((Opc >= ISD::BUILTIN_OP_END ||
20842 Opc == ISD::INTRINSIC_WO_CHAIN ||
20843 Opc == ISD::INTRINSIC_W_CHAIN ||
20844 Opc == ISD::INTRINSIC_VOID) &&
20845 "Should use MaskedValueIsZero if you don't know whether Op"
20846 " is a target node!");
20848 KnownZero = KnownOne = APInt(BitWidth, 0); // Don't know anything.
20862 // These nodes' second result is a boolean.
20863 if (Op.getResNo() == 0)
20866 case X86ISD::SETCC:
20867 KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - 1);
20869 case ISD::INTRINSIC_WO_CHAIN: {
20870 unsigned IntId = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
20871 unsigned NumLoBits = 0;
20874 case Intrinsic::x86_sse_movmsk_ps:
20875 case Intrinsic::x86_avx_movmsk_ps_256:
20876 case Intrinsic::x86_sse2_movmsk_pd:
20877 case Intrinsic::x86_avx_movmsk_pd_256:
20878 case Intrinsic::x86_mmx_pmovmskb:
20879 case Intrinsic::x86_sse2_pmovmskb_128:
20880 case Intrinsic::x86_avx2_pmovmskb: {
20881 // High bits of movmskp{s|d}, pmovmskb are known zero.
20883 default: llvm_unreachable("Impossible intrinsic"); // Can't reach here.
20884 case Intrinsic::x86_sse_movmsk_ps: NumLoBits = 4; break;
20885 case Intrinsic::x86_avx_movmsk_ps_256: NumLoBits = 8; break;
20886 case Intrinsic::x86_sse2_movmsk_pd: NumLoBits = 2; break;
20887 case Intrinsic::x86_avx_movmsk_pd_256: NumLoBits = 4; break;
20888 case Intrinsic::x86_mmx_pmovmskb: NumLoBits = 8; break;
20889 case Intrinsic::x86_sse2_pmovmskb_128: NumLoBits = 16; break;
20890 case Intrinsic::x86_avx2_pmovmskb: NumLoBits = 32; break;
20892 KnownZero = APInt::getHighBitsSet(BitWidth, BitWidth - NumLoBits);
20901 unsigned X86TargetLowering::ComputeNumSignBitsForTargetNode(
20903 const SelectionDAG &,
20904 unsigned Depth) const {
20905 // SETCC_CARRY sets the dest to ~0 for true or 0 for false.
20906 if (Op.getOpcode() == X86ISD::SETCC_CARRY)
20907 return Op.getValueType().getScalarType().getSizeInBits();
20913 /// isGAPlusOffset - Returns true (and the GlobalValue and the offset) if the
20914 /// node is a GlobalAddress + offset.
20915 bool X86TargetLowering::isGAPlusOffset(SDNode *N,
20916 const GlobalValue* &GA,
20917 int64_t &Offset) const {
20918 if (N->getOpcode() == X86ISD::Wrapper) {
20919 if (isa<GlobalAddressSDNode>(N->getOperand(0))) {
20920 GA = cast<GlobalAddressSDNode>(N->getOperand(0))->getGlobal();
20921 Offset = cast<GlobalAddressSDNode>(N->getOperand(0))->getOffset();
20925 return TargetLowering::isGAPlusOffset(N, GA, Offset);
20928 /// isShuffleHigh128VectorInsertLow - Checks whether the shuffle node is the
20929 /// same as extracting the high 128-bit part of 256-bit vector and then
20930 /// inserting the result into the low part of a new 256-bit vector
20931 static bool isShuffleHigh128VectorInsertLow(ShuffleVectorSDNode *SVOp) {
20932 EVT VT = SVOp->getValueType(0);
20933 unsigned NumElems = VT.getVectorNumElements();
20935 // vector_shuffle <4, 5, 6, 7, u, u, u, u> or <2, 3, u, u>
20936 for (unsigned i = 0, j = NumElems/2; i != NumElems/2; ++i, ++j)
20937 if (!isUndefOrEqual(SVOp->getMaskElt(i), j) ||
20938 SVOp->getMaskElt(j) >= 0)
20944 /// isShuffleLow128VectorInsertHigh - Checks whether the shuffle node is the
20945 /// same as extracting the low 128-bit part of 256-bit vector and then
20946 /// inserting the result into the high part of a new 256-bit vector
20947 static bool isShuffleLow128VectorInsertHigh(ShuffleVectorSDNode *SVOp) {
20948 EVT VT = SVOp->getValueType(0);
20949 unsigned NumElems = VT.getVectorNumElements();
20951 // vector_shuffle <u, u, u, u, 0, 1, 2, 3> or <u, u, 0, 1>
20952 for (unsigned i = NumElems/2, j = 0; i != NumElems; ++i, ++j)
20953 if (!isUndefOrEqual(SVOp->getMaskElt(i), j) ||
20954 SVOp->getMaskElt(j) >= 0)
20960 /// PerformShuffleCombine256 - Performs shuffle combines for 256-bit vectors.
20961 static SDValue PerformShuffleCombine256(SDNode *N, SelectionDAG &DAG,
20962 TargetLowering::DAGCombinerInfo &DCI,
20963 const X86Subtarget* Subtarget) {
20965 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
20966 SDValue V1 = SVOp->getOperand(0);
20967 SDValue V2 = SVOp->getOperand(1);
20968 EVT VT = SVOp->getValueType(0);
20969 unsigned NumElems = VT.getVectorNumElements();
20971 if (V1.getOpcode() == ISD::CONCAT_VECTORS &&
20972 V2.getOpcode() == ISD::CONCAT_VECTORS) {
20976 // V UNDEF BUILD_VECTOR UNDEF
20978 // CONCAT_VECTOR CONCAT_VECTOR
20981 // RESULT: V + zero extended
20983 if (V2.getOperand(0).getOpcode() != ISD::BUILD_VECTOR ||
20984 V2.getOperand(1).getOpcode() != ISD::UNDEF ||
20985 V1.getOperand(1).getOpcode() != ISD::UNDEF)
20988 if (!ISD::isBuildVectorAllZeros(V2.getOperand(0).getNode()))
20991 // To match the shuffle mask, the first half of the mask should
20992 // be exactly the first vector, and all the rest a splat with the
20993 // first element of the second one.
20994 for (unsigned i = 0; i != NumElems/2; ++i)
20995 if (!isUndefOrEqual(SVOp->getMaskElt(i), i) ||
20996 !isUndefOrEqual(SVOp->getMaskElt(i+NumElems/2), NumElems))
20999 // If V1 is coming from a vector load then just fold to a VZEXT_LOAD.
21000 if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(V1.getOperand(0))) {
21001 if (Ld->hasNUsesOfValue(1, 0)) {
21002 SDVTList Tys = DAG.getVTList(MVT::v4i64, MVT::Other);
21003 SDValue Ops[] = { Ld->getChain(), Ld->getBasePtr() };
21005 DAG.getMemIntrinsicNode(X86ISD::VZEXT_LOAD, dl, Tys, Ops,
21007 Ld->getPointerInfo(),
21008 Ld->getAlignment(),
21009 false/*isVolatile*/, true/*ReadMem*/,
21010 false/*WriteMem*/);
21012 // Make sure the newly-created LOAD is in the same position as Ld in
21013 // terms of dependency. We create a TokenFactor for Ld and ResNode,
21014 // and update uses of Ld's output chain to use the TokenFactor.
21015 if (Ld->hasAnyUseOfValue(1)) {
21016 SDValue NewChain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
21017 SDValue(Ld, 1), SDValue(ResNode.getNode(), 1));
21018 DAG.ReplaceAllUsesOfValueWith(SDValue(Ld, 1), NewChain);
21019 DAG.UpdateNodeOperands(NewChain.getNode(), SDValue(Ld, 1),
21020 SDValue(ResNode.getNode(), 1));
21023 return DAG.getBitcast(VT, ResNode);
21027 // Emit a zeroed vector and insert the desired subvector on its
21029 SDValue Zeros = getZeroVector(VT, Subtarget, DAG, dl);
21030 SDValue InsV = Insert128BitVector(Zeros, V1.getOperand(0), 0, DAG, dl);
21031 return DCI.CombineTo(N, InsV);
21034 //===--------------------------------------------------------------------===//
21035 // Combine some shuffles into subvector extracts and inserts:
21038 // vector_shuffle <4, 5, 6, 7, u, u, u, u> or <2, 3, u, u>
21039 if (isShuffleHigh128VectorInsertLow(SVOp)) {
21040 SDValue V = Extract128BitVector(V1, NumElems/2, DAG, dl);
21041 SDValue InsV = Insert128BitVector(DAG.getUNDEF(VT), V, 0, DAG, dl);
21042 return DCI.CombineTo(N, InsV);
21045 // vector_shuffle <u, u, u, u, 0, 1, 2, 3> or <u, u, 0, 1>
21046 if (isShuffleLow128VectorInsertHigh(SVOp)) {
21047 SDValue V = Extract128BitVector(V1, 0, DAG, dl);
21048 SDValue InsV = Insert128BitVector(DAG.getUNDEF(VT), V, NumElems/2, DAG, dl);
21049 return DCI.CombineTo(N, InsV);
21055 /// \brief Combine an arbitrary chain of shuffles into a single instruction if
21058 /// This is the leaf of the recursive combinine below. When we have found some
21059 /// chain of single-use x86 shuffle instructions and accumulated the combined
21060 /// shuffle mask represented by them, this will try to pattern match that mask
21061 /// into either a single instruction if there is a special purpose instruction
21062 /// for this operation, or into a PSHUFB instruction which is a fully general
21063 /// instruction but should only be used to replace chains over a certain depth.
21064 static bool combineX86ShuffleChain(SDValue Op, SDValue Root, ArrayRef<int> Mask,
21065 int Depth, bool HasPSHUFB, SelectionDAG &DAG,
21066 TargetLowering::DAGCombinerInfo &DCI,
21067 const X86Subtarget *Subtarget) {
21068 assert(!Mask.empty() && "Cannot combine an empty shuffle mask!");
21070 // Find the operand that enters the chain. Note that multiple uses are OK
21071 // here, we're not going to remove the operand we find.
21072 SDValue Input = Op.getOperand(0);
21073 while (Input.getOpcode() == ISD::BITCAST)
21074 Input = Input.getOperand(0);
21076 MVT VT = Input.getSimpleValueType();
21077 MVT RootVT = Root.getSimpleValueType();
21080 // Just remove no-op shuffle masks.
21081 if (Mask.size() == 1) {
21082 DCI.CombineTo(Root.getNode(), DAG.getBitcast(RootVT, Input),
21087 // Use the float domain if the operand type is a floating point type.
21088 bool FloatDomain = VT.isFloatingPoint();
21090 // For floating point shuffles, we don't have free copies in the shuffle
21091 // instructions or the ability to load as part of the instruction, so
21092 // canonicalize their shuffles to UNPCK or MOV variants.
21094 // Note that even with AVX we prefer the PSHUFD form of shuffle for integer
21095 // vectors because it can have a load folded into it that UNPCK cannot. This
21096 // doesn't preclude something switching to the shorter encoding post-RA.
21098 // FIXME: Should teach these routines about AVX vector widths.
21099 if (FloatDomain && VT.getSizeInBits() == 128) {
21100 if (Mask.equals({0, 0}) || Mask.equals({1, 1})) {
21101 bool Lo = Mask.equals({0, 0});
21104 // Check if we have SSE3 which will let us use MOVDDUP. That instruction
21105 // is no slower than UNPCKLPD but has the option to fold the input operand
21106 // into even an unaligned memory load.
21107 if (Lo && Subtarget->hasSSE3()) {
21108 Shuffle = X86ISD::MOVDDUP;
21109 ShuffleVT = MVT::v2f64;
21111 // We have MOVLHPS and MOVHLPS throughout SSE and they encode smaller
21112 // than the UNPCK variants.
21113 Shuffle = Lo ? X86ISD::MOVLHPS : X86ISD::MOVHLPS;
21114 ShuffleVT = MVT::v4f32;
21116 if (Depth == 1 && Root->getOpcode() == Shuffle)
21117 return false; // Nothing to do!
21118 Op = DAG.getBitcast(ShuffleVT, Input);
21119 DCI.AddToWorklist(Op.getNode());
21120 if (Shuffle == X86ISD::MOVDDUP)
21121 Op = DAG.getNode(Shuffle, DL, ShuffleVT, Op);
21123 Op = DAG.getNode(Shuffle, DL, ShuffleVT, Op, Op);
21124 DCI.AddToWorklist(Op.getNode());
21125 DCI.CombineTo(Root.getNode(), DAG.getBitcast(RootVT, Op),
21129 if (Subtarget->hasSSE3() &&
21130 (Mask.equals({0, 0, 2, 2}) || Mask.equals({1, 1, 3, 3}))) {
21131 bool Lo = Mask.equals({0, 0, 2, 2});
21132 unsigned Shuffle = Lo ? X86ISD::MOVSLDUP : X86ISD::MOVSHDUP;
21133 MVT ShuffleVT = MVT::v4f32;
21134 if (Depth == 1 && Root->getOpcode() == Shuffle)
21135 return false; // Nothing to do!
21136 Op = DAG.getBitcast(ShuffleVT, Input);
21137 DCI.AddToWorklist(Op.getNode());
21138 Op = DAG.getNode(Shuffle, DL, ShuffleVT, Op);
21139 DCI.AddToWorklist(Op.getNode());
21140 DCI.CombineTo(Root.getNode(), DAG.getBitcast(RootVT, Op),
21144 if (Mask.equals({0, 0, 1, 1}) || Mask.equals({2, 2, 3, 3})) {
21145 bool Lo = Mask.equals({0, 0, 1, 1});
21146 unsigned Shuffle = Lo ? X86ISD::UNPCKL : X86ISD::UNPCKH;
21147 MVT ShuffleVT = MVT::v4f32;
21148 if (Depth == 1 && Root->getOpcode() == Shuffle)
21149 return false; // Nothing to do!
21150 Op = DAG.getBitcast(ShuffleVT, Input);
21151 DCI.AddToWorklist(Op.getNode());
21152 Op = DAG.getNode(Shuffle, DL, ShuffleVT, Op, Op);
21153 DCI.AddToWorklist(Op.getNode());
21154 DCI.CombineTo(Root.getNode(), DAG.getBitcast(RootVT, Op),
21160 // We always canonicalize the 8 x i16 and 16 x i8 shuffles into their UNPCK
21161 // variants as none of these have single-instruction variants that are
21162 // superior to the UNPCK formulation.
21163 if (!FloatDomain && VT.getSizeInBits() == 128 &&
21164 (Mask.equals({0, 0, 1, 1, 2, 2, 3, 3}) ||
21165 Mask.equals({4, 4, 5, 5, 6, 6, 7, 7}) ||
21166 Mask.equals({0, 0, 1, 1, 2, 2, 3, 3, 4, 4, 5, 5, 6, 6, 7, 7}) ||
21168 {8, 8, 9, 9, 10, 10, 11, 11, 12, 12, 13, 13, 14, 14, 15, 15}))) {
21169 bool Lo = Mask[0] == 0;
21170 unsigned Shuffle = Lo ? X86ISD::UNPCKL : X86ISD::UNPCKH;
21171 if (Depth == 1 && Root->getOpcode() == Shuffle)
21172 return false; // Nothing to do!
21174 switch (Mask.size()) {
21176 ShuffleVT = MVT::v8i16;
21179 ShuffleVT = MVT::v16i8;
21182 llvm_unreachable("Impossible mask size!");
21184 Op = DAG.getBitcast(ShuffleVT, Input);
21185 DCI.AddToWorklist(Op.getNode());
21186 Op = DAG.getNode(Shuffle, DL, ShuffleVT, Op, Op);
21187 DCI.AddToWorklist(Op.getNode());
21188 DCI.CombineTo(Root.getNode(), DAG.getBitcast(RootVT, Op),
21193 // Don't try to re-form single instruction chains under any circumstances now
21194 // that we've done encoding canonicalization for them.
21198 // If we have 3 or more shuffle instructions or a chain involving PSHUFB, we
21199 // can replace them with a single PSHUFB instruction profitably. Intel's
21200 // manuals suggest only using PSHUFB if doing so replacing 5 instructions, but
21201 // in practice PSHUFB tends to be *very* fast so we're more aggressive.
21202 if ((Depth >= 3 || HasPSHUFB) && Subtarget->hasSSSE3()) {
21203 SmallVector<SDValue, 16> PSHUFBMask;
21204 int NumBytes = VT.getSizeInBits() / 8;
21205 int Ratio = NumBytes / Mask.size();
21206 for (int i = 0; i < NumBytes; ++i) {
21207 if (Mask[i / Ratio] == SM_SentinelUndef) {
21208 PSHUFBMask.push_back(DAG.getUNDEF(MVT::i8));
21211 int M = Mask[i / Ratio] != SM_SentinelZero
21212 ? Ratio * Mask[i / Ratio] + i % Ratio
21214 PSHUFBMask.push_back(DAG.getConstant(M, DL, MVT::i8));
21216 MVT ByteVT = MVT::getVectorVT(MVT::i8, NumBytes);
21217 Op = DAG.getBitcast(ByteVT, Input);
21218 DCI.AddToWorklist(Op.getNode());
21219 SDValue PSHUFBMaskOp =
21220 DAG.getNode(ISD::BUILD_VECTOR, DL, ByteVT, PSHUFBMask);
21221 DCI.AddToWorklist(PSHUFBMaskOp.getNode());
21222 Op = DAG.getNode(X86ISD::PSHUFB, DL, ByteVT, Op, PSHUFBMaskOp);
21223 DCI.AddToWorklist(Op.getNode());
21224 DCI.CombineTo(Root.getNode(), DAG.getBitcast(RootVT, Op),
21229 // Failed to find any combines.
21233 /// \brief Fully generic combining of x86 shuffle instructions.
21235 /// This should be the last combine run over the x86 shuffle instructions. Once
21236 /// they have been fully optimized, this will recursively consider all chains
21237 /// of single-use shuffle instructions, build a generic model of the cumulative
21238 /// shuffle operation, and check for simpler instructions which implement this
21239 /// operation. We use this primarily for two purposes:
21241 /// 1) Collapse generic shuffles to specialized single instructions when
21242 /// equivalent. In most cases, this is just an encoding size win, but
21243 /// sometimes we will collapse multiple generic shuffles into a single
21244 /// special-purpose shuffle.
21245 /// 2) Look for sequences of shuffle instructions with 3 or more total
21246 /// instructions, and replace them with the slightly more expensive SSSE3
21247 /// PSHUFB instruction if available. We do this as the last combining step
21248 /// to ensure we avoid using PSHUFB if we can implement the shuffle with
21249 /// a suitable short sequence of other instructions. The PHUFB will either
21250 /// use a register or have to read from memory and so is slightly (but only
21251 /// slightly) more expensive than the other shuffle instructions.
21253 /// Because this is inherently a quadratic operation (for each shuffle in
21254 /// a chain, we recurse up the chain), the depth is limited to 8 instructions.
21255 /// This should never be an issue in practice as the shuffle lowering doesn't
21256 /// produce sequences of more than 8 instructions.
21258 /// FIXME: We will currently miss some cases where the redundant shuffling
21259 /// would simplify under the threshold for PSHUFB formation because of
21260 /// combine-ordering. To fix this, we should do the redundant instruction
21261 /// combining in this recursive walk.
21262 static bool combineX86ShufflesRecursively(SDValue Op, SDValue Root,
21263 ArrayRef<int> RootMask,
21264 int Depth, bool HasPSHUFB,
21266 TargetLowering::DAGCombinerInfo &DCI,
21267 const X86Subtarget *Subtarget) {
21268 // Bound the depth of our recursive combine because this is ultimately
21269 // quadratic in nature.
21273 // Directly rip through bitcasts to find the underlying operand.
21274 while (Op.getOpcode() == ISD::BITCAST && Op.getOperand(0).hasOneUse())
21275 Op = Op.getOperand(0);
21277 MVT VT = Op.getSimpleValueType();
21278 if (!VT.isVector())
21279 return false; // Bail if we hit a non-vector.
21281 assert(Root.getSimpleValueType().isVector() &&
21282 "Shuffles operate on vector types!");
21283 assert(VT.getSizeInBits() == Root.getSimpleValueType().getSizeInBits() &&
21284 "Can only combine shuffles of the same vector register size.");
21286 if (!isTargetShuffle(Op.getOpcode()))
21288 SmallVector<int, 16> OpMask;
21290 bool HaveMask = getTargetShuffleMask(Op.getNode(), VT, OpMask, IsUnary);
21291 // We only can combine unary shuffles which we can decode the mask for.
21292 if (!HaveMask || !IsUnary)
21295 assert(VT.getVectorNumElements() == OpMask.size() &&
21296 "Different mask size from vector size!");
21297 assert(((RootMask.size() > OpMask.size() &&
21298 RootMask.size() % OpMask.size() == 0) ||
21299 (OpMask.size() > RootMask.size() &&
21300 OpMask.size() % RootMask.size() == 0) ||
21301 OpMask.size() == RootMask.size()) &&
21302 "The smaller number of elements must divide the larger.");
21303 int RootRatio = std::max<int>(1, OpMask.size() / RootMask.size());
21304 int OpRatio = std::max<int>(1, RootMask.size() / OpMask.size());
21305 assert(((RootRatio == 1 && OpRatio == 1) ||
21306 (RootRatio == 1) != (OpRatio == 1)) &&
21307 "Must not have a ratio for both incoming and op masks!");
21309 SmallVector<int, 16> Mask;
21310 Mask.reserve(std::max(OpMask.size(), RootMask.size()));
21312 // Merge this shuffle operation's mask into our accumulated mask. Note that
21313 // this shuffle's mask will be the first applied to the input, followed by the
21314 // root mask to get us all the way to the root value arrangement. The reason
21315 // for this order is that we are recursing up the operation chain.
21316 for (int i = 0, e = std::max(OpMask.size(), RootMask.size()); i < e; ++i) {
21317 int RootIdx = i / RootRatio;
21318 if (RootMask[RootIdx] < 0) {
21319 // This is a zero or undef lane, we're done.
21320 Mask.push_back(RootMask[RootIdx]);
21324 int RootMaskedIdx = RootMask[RootIdx] * RootRatio + i % RootRatio;
21325 int OpIdx = RootMaskedIdx / OpRatio;
21326 if (OpMask[OpIdx] < 0) {
21327 // The incoming lanes are zero or undef, it doesn't matter which ones we
21329 Mask.push_back(OpMask[OpIdx]);
21333 // Ok, we have non-zero lanes, map them through.
21334 Mask.push_back(OpMask[OpIdx] * OpRatio +
21335 RootMaskedIdx % OpRatio);
21338 // See if we can recurse into the operand to combine more things.
21339 switch (Op.getOpcode()) {
21340 case X86ISD::PSHUFB:
21342 case X86ISD::PSHUFD:
21343 case X86ISD::PSHUFHW:
21344 case X86ISD::PSHUFLW:
21345 if (Op.getOperand(0).hasOneUse() &&
21346 combineX86ShufflesRecursively(Op.getOperand(0), Root, Mask, Depth + 1,
21347 HasPSHUFB, DAG, DCI, Subtarget))
21351 case X86ISD::UNPCKL:
21352 case X86ISD::UNPCKH:
21353 assert(Op.getOperand(0) == Op.getOperand(1) && "We only combine unary shuffles!");
21354 // We can't check for single use, we have to check that this shuffle is the only user.
21355 if (Op->isOnlyUserOf(Op.getOperand(0).getNode()) &&
21356 combineX86ShufflesRecursively(Op.getOperand(0), Root, Mask, Depth + 1,
21357 HasPSHUFB, DAG, DCI, Subtarget))
21362 // Minor canonicalization of the accumulated shuffle mask to make it easier
21363 // to match below. All this does is detect masks with squential pairs of
21364 // elements, and shrink them to the half-width mask. It does this in a loop
21365 // so it will reduce the size of the mask to the minimal width mask which
21366 // performs an equivalent shuffle.
21367 SmallVector<int, 16> WidenedMask;
21368 while (Mask.size() > 1 && canWidenShuffleElements(Mask, WidenedMask)) {
21369 Mask = std::move(WidenedMask);
21370 WidenedMask.clear();
21373 return combineX86ShuffleChain(Op, Root, Mask, Depth, HasPSHUFB, DAG, DCI,
21377 /// \brief Get the PSHUF-style mask from PSHUF node.
21379 /// This is a very minor wrapper around getTargetShuffleMask to easy forming v4
21380 /// PSHUF-style masks that can be reused with such instructions.
21381 static SmallVector<int, 4> getPSHUFShuffleMask(SDValue N) {
21382 MVT VT = N.getSimpleValueType();
21383 SmallVector<int, 4> Mask;
21385 bool HaveMask = getTargetShuffleMask(N.getNode(), VT, Mask, IsUnary);
21389 // If we have more than 128-bits, only the low 128-bits of shuffle mask
21390 // matter. Check that the upper masks are repeats and remove them.
21391 if (VT.getSizeInBits() > 128) {
21392 int LaneElts = 128 / VT.getScalarSizeInBits();
21394 for (int i = 1, NumLanes = VT.getSizeInBits() / 128; i < NumLanes; ++i)
21395 for (int j = 0; j < LaneElts; ++j)
21396 assert(Mask[j] == Mask[i * LaneElts + j] - (LaneElts * i) &&
21397 "Mask doesn't repeat in high 128-bit lanes!");
21399 Mask.resize(LaneElts);
21402 switch (N.getOpcode()) {
21403 case X86ISD::PSHUFD:
21405 case X86ISD::PSHUFLW:
21408 case X86ISD::PSHUFHW:
21409 Mask.erase(Mask.begin(), Mask.begin() + 4);
21410 for (int &M : Mask)
21414 llvm_unreachable("No valid shuffle instruction found!");
21418 /// \brief Search for a combinable shuffle across a chain ending in pshufd.
21420 /// We walk up the chain and look for a combinable shuffle, skipping over
21421 /// shuffles that we could hoist this shuffle's transformation past without
21422 /// altering anything.
21424 combineRedundantDWordShuffle(SDValue N, MutableArrayRef<int> Mask,
21426 TargetLowering::DAGCombinerInfo &DCI) {
21427 assert(N.getOpcode() == X86ISD::PSHUFD &&
21428 "Called with something other than an x86 128-bit half shuffle!");
21431 // Walk up a single-use chain looking for a combinable shuffle. Keep a stack
21432 // of the shuffles in the chain so that we can form a fresh chain to replace
21434 SmallVector<SDValue, 8> Chain;
21435 SDValue V = N.getOperand(0);
21436 for (; V.hasOneUse(); V = V.getOperand(0)) {
21437 switch (V.getOpcode()) {
21439 return SDValue(); // Nothing combined!
21442 // Skip bitcasts as we always know the type for the target specific
21446 case X86ISD::PSHUFD:
21447 // Found another dword shuffle.
21450 case X86ISD::PSHUFLW:
21451 // Check that the low words (being shuffled) are the identity in the
21452 // dword shuffle, and the high words are self-contained.
21453 if (Mask[0] != 0 || Mask[1] != 1 ||
21454 !(Mask[2] >= 2 && Mask[2] < 4 && Mask[3] >= 2 && Mask[3] < 4))
21457 Chain.push_back(V);
21460 case X86ISD::PSHUFHW:
21461 // Check that the high words (being shuffled) are the identity in the
21462 // dword shuffle, and the low words are self-contained.
21463 if (Mask[2] != 2 || Mask[3] != 3 ||
21464 !(Mask[0] >= 0 && Mask[0] < 2 && Mask[1] >= 0 && Mask[1] < 2))
21467 Chain.push_back(V);
21470 case X86ISD::UNPCKL:
21471 case X86ISD::UNPCKH:
21472 // For either i8 -> i16 or i16 -> i32 unpacks, we can combine a dword
21473 // shuffle into a preceding word shuffle.
21474 if (V.getSimpleValueType().getScalarType() != MVT::i8 &&
21475 V.getSimpleValueType().getScalarType() != MVT::i16)
21478 // Search for a half-shuffle which we can combine with.
21479 unsigned CombineOp =
21480 V.getOpcode() == X86ISD::UNPCKL ? X86ISD::PSHUFLW : X86ISD::PSHUFHW;
21481 if (V.getOperand(0) != V.getOperand(1) ||
21482 !V->isOnlyUserOf(V.getOperand(0).getNode()))
21484 Chain.push_back(V);
21485 V = V.getOperand(0);
21487 switch (V.getOpcode()) {
21489 return SDValue(); // Nothing to combine.
21491 case X86ISD::PSHUFLW:
21492 case X86ISD::PSHUFHW:
21493 if (V.getOpcode() == CombineOp)
21496 Chain.push_back(V);
21500 V = V.getOperand(0);
21504 } while (V.hasOneUse());
21507 // Break out of the loop if we break out of the switch.
21511 if (!V.hasOneUse())
21512 // We fell out of the loop without finding a viable combining instruction.
21515 // Merge this node's mask and our incoming mask.
21516 SmallVector<int, 4> VMask = getPSHUFShuffleMask(V);
21517 for (int &M : Mask)
21519 V = DAG.getNode(V.getOpcode(), DL, V.getValueType(), V.getOperand(0),
21520 getV4X86ShuffleImm8ForMask(Mask, DL, DAG));
21522 // Rebuild the chain around this new shuffle.
21523 while (!Chain.empty()) {
21524 SDValue W = Chain.pop_back_val();
21526 if (V.getValueType() != W.getOperand(0).getValueType())
21527 V = DAG.getBitcast(W.getOperand(0).getValueType(), V);
21529 switch (W.getOpcode()) {
21531 llvm_unreachable("Only PSHUF and UNPCK instructions get here!");
21533 case X86ISD::UNPCKL:
21534 case X86ISD::UNPCKH:
21535 V = DAG.getNode(W.getOpcode(), DL, W.getValueType(), V, V);
21538 case X86ISD::PSHUFD:
21539 case X86ISD::PSHUFLW:
21540 case X86ISD::PSHUFHW:
21541 V = DAG.getNode(W.getOpcode(), DL, W.getValueType(), V, W.getOperand(1));
21545 if (V.getValueType() != N.getValueType())
21546 V = DAG.getBitcast(N.getValueType(), V);
21548 // Return the new chain to replace N.
21552 /// \brief Search for a combinable shuffle across a chain ending in pshuflw or pshufhw.
21554 /// We walk up the chain, skipping shuffles of the other half and looking
21555 /// through shuffles which switch halves trying to find a shuffle of the same
21556 /// pair of dwords.
21557 static bool combineRedundantHalfShuffle(SDValue N, MutableArrayRef<int> Mask,
21559 TargetLowering::DAGCombinerInfo &DCI) {
21561 (N.getOpcode() == X86ISD::PSHUFLW || N.getOpcode() == X86ISD::PSHUFHW) &&
21562 "Called with something other than an x86 128-bit half shuffle!");
21564 unsigned CombineOpcode = N.getOpcode();
21566 // Walk up a single-use chain looking for a combinable shuffle.
21567 SDValue V = N.getOperand(0);
21568 for (; V.hasOneUse(); V = V.getOperand(0)) {
21569 switch (V.getOpcode()) {
21571 return false; // Nothing combined!
21574 // Skip bitcasts as we always know the type for the target specific
21578 case X86ISD::PSHUFLW:
21579 case X86ISD::PSHUFHW:
21580 if (V.getOpcode() == CombineOpcode)
21583 // Other-half shuffles are no-ops.
21586 // Break out of the loop if we break out of the switch.
21590 if (!V.hasOneUse())
21591 // We fell out of the loop without finding a viable combining instruction.
21594 // Combine away the bottom node as its shuffle will be accumulated into
21595 // a preceding shuffle.
21596 DCI.CombineTo(N.getNode(), N.getOperand(0), /*AddTo*/ true);
21598 // Record the old value.
21601 // Merge this node's mask and our incoming mask (adjusted to account for all
21602 // the pshufd instructions encountered).
21603 SmallVector<int, 4> VMask = getPSHUFShuffleMask(V);
21604 for (int &M : Mask)
21606 V = DAG.getNode(V.getOpcode(), DL, MVT::v8i16, V.getOperand(0),
21607 getV4X86ShuffleImm8ForMask(Mask, DL, DAG));
21609 // Check that the shuffles didn't cancel each other out. If not, we need to
21610 // combine to the new one.
21612 // Replace the combinable shuffle with the combined one, updating all users
21613 // so that we re-evaluate the chain here.
21614 DCI.CombineTo(Old.getNode(), V, /*AddTo*/ true);
21619 /// \brief Try to combine x86 target specific shuffles.
21620 static SDValue PerformTargetShuffleCombine(SDValue N, SelectionDAG &DAG,
21621 TargetLowering::DAGCombinerInfo &DCI,
21622 const X86Subtarget *Subtarget) {
21624 MVT VT = N.getSimpleValueType();
21625 SmallVector<int, 4> Mask;
21627 switch (N.getOpcode()) {
21628 case X86ISD::PSHUFD:
21629 case X86ISD::PSHUFLW:
21630 case X86ISD::PSHUFHW:
21631 Mask = getPSHUFShuffleMask(N);
21632 assert(Mask.size() == 4);
21638 // Nuke no-op shuffles that show up after combining.
21639 if (isNoopShuffleMask(Mask))
21640 return DCI.CombineTo(N.getNode(), N.getOperand(0), /*AddTo*/ true);
21642 // Look for simplifications involving one or two shuffle instructions.
21643 SDValue V = N.getOperand(0);
21644 switch (N.getOpcode()) {
21647 case X86ISD::PSHUFLW:
21648 case X86ISD::PSHUFHW:
21649 assert(VT.getScalarType() == MVT::i16 && "Bad word shuffle type!");
21651 if (combineRedundantHalfShuffle(N, Mask, DAG, DCI))
21652 return SDValue(); // We combined away this shuffle, so we're done.
21654 // See if this reduces to a PSHUFD which is no more expensive and can
21655 // combine with more operations. Note that it has to at least flip the
21656 // dwords as otherwise it would have been removed as a no-op.
21657 if (makeArrayRef(Mask).equals({2, 3, 0, 1})) {
21658 int DMask[] = {0, 1, 2, 3};
21659 int DOffset = N.getOpcode() == X86ISD::PSHUFLW ? 0 : 2;
21660 DMask[DOffset + 0] = DOffset + 1;
21661 DMask[DOffset + 1] = DOffset + 0;
21662 MVT DVT = MVT::getVectorVT(MVT::i32, VT.getVectorNumElements() / 2);
21663 V = DAG.getBitcast(DVT, V);
21664 DCI.AddToWorklist(V.getNode());
21665 V = DAG.getNode(X86ISD::PSHUFD, DL, DVT, V,
21666 getV4X86ShuffleImm8ForMask(DMask, DL, DAG));
21667 DCI.AddToWorklist(V.getNode());
21668 return DAG.getBitcast(VT, V);
21671 // Look for shuffle patterns which can be implemented as a single unpack.
21672 // FIXME: This doesn't handle the location of the PSHUFD generically, and
21673 // only works when we have a PSHUFD followed by two half-shuffles.
21674 if (Mask[0] == Mask[1] && Mask[2] == Mask[3] &&
21675 (V.getOpcode() == X86ISD::PSHUFLW ||
21676 V.getOpcode() == X86ISD::PSHUFHW) &&
21677 V.getOpcode() != N.getOpcode() &&
21679 SDValue D = V.getOperand(0);
21680 while (D.getOpcode() == ISD::BITCAST && D.hasOneUse())
21681 D = D.getOperand(0);
21682 if (D.getOpcode() == X86ISD::PSHUFD && D.hasOneUse()) {
21683 SmallVector<int, 4> VMask = getPSHUFShuffleMask(V);
21684 SmallVector<int, 4> DMask = getPSHUFShuffleMask(D);
21685 int NOffset = N.getOpcode() == X86ISD::PSHUFLW ? 0 : 4;
21686 int VOffset = V.getOpcode() == X86ISD::PSHUFLW ? 0 : 4;
21688 for (int i = 0; i < 4; ++i) {
21689 WordMask[i + NOffset] = Mask[i] + NOffset;
21690 WordMask[i + VOffset] = VMask[i] + VOffset;
21692 // Map the word mask through the DWord mask.
21694 for (int i = 0; i < 8; ++i)
21695 MappedMask[i] = 2 * DMask[WordMask[i] / 2] + WordMask[i] % 2;
21696 if (makeArrayRef(MappedMask).equals({0, 0, 1, 1, 2, 2, 3, 3}) ||
21697 makeArrayRef(MappedMask).equals({4, 4, 5, 5, 6, 6, 7, 7})) {
21698 // We can replace all three shuffles with an unpack.
21699 V = DAG.getBitcast(VT, D.getOperand(0));
21700 DCI.AddToWorklist(V.getNode());
21701 return DAG.getNode(MappedMask[0] == 0 ? X86ISD::UNPCKL
21710 case X86ISD::PSHUFD:
21711 if (SDValue NewN = combineRedundantDWordShuffle(N, Mask, DAG, DCI))
21720 /// \brief Try to combine a shuffle into a target-specific add-sub node.
21722 /// We combine this directly on the abstract vector shuffle nodes so it is
21723 /// easier to generically match. We also insert dummy vector shuffle nodes for
21724 /// the operands which explicitly discard the lanes which are unused by this
21725 /// operation to try to flow through the rest of the combiner the fact that
21726 /// they're unused.
21727 static SDValue combineShuffleToAddSub(SDNode *N, SelectionDAG &DAG) {
21729 EVT VT = N->getValueType(0);
21731 // We only handle target-independent shuffles.
21732 // FIXME: It would be easy and harmless to use the target shuffle mask
21733 // extraction tool to support more.
21734 if (N->getOpcode() != ISD::VECTOR_SHUFFLE)
21737 auto *SVN = cast<ShuffleVectorSDNode>(N);
21738 ArrayRef<int> Mask = SVN->getMask();
21739 SDValue V1 = N->getOperand(0);
21740 SDValue V2 = N->getOperand(1);
21742 // We require the first shuffle operand to be the SUB node, and the second to
21743 // be the ADD node.
21744 // FIXME: We should support the commuted patterns.
21745 if (V1->getOpcode() != ISD::FSUB || V2->getOpcode() != ISD::FADD)
21748 // If there are other uses of these operations we can't fold them.
21749 if (!V1->hasOneUse() || !V2->hasOneUse())
21752 // Ensure that both operations have the same operands. Note that we can
21753 // commute the FADD operands.
21754 SDValue LHS = V1->getOperand(0), RHS = V1->getOperand(1);
21755 if ((V2->getOperand(0) != LHS || V2->getOperand(1) != RHS) &&
21756 (V2->getOperand(0) != RHS || V2->getOperand(1) != LHS))
21759 // We're looking for blends between FADD and FSUB nodes. We insist on these
21760 // nodes being lined up in a specific expected pattern.
21761 if (!(isShuffleEquivalent(V1, V2, Mask, {0, 3}) ||
21762 isShuffleEquivalent(V1, V2, Mask, {0, 5, 2, 7}) ||
21763 isShuffleEquivalent(V1, V2, Mask, {0, 9, 2, 11, 4, 13, 6, 15})))
21766 // Only specific types are legal at this point, assert so we notice if and
21767 // when these change.
21768 assert((VT == MVT::v4f32 || VT == MVT::v2f64 || VT == MVT::v8f32 ||
21769 VT == MVT::v4f64) &&
21770 "Unknown vector type encountered!");
21772 return DAG.getNode(X86ISD::ADDSUB, DL, VT, LHS, RHS);
21775 /// PerformShuffleCombine - Performs several different shuffle combines.
21776 static SDValue PerformShuffleCombine(SDNode *N, SelectionDAG &DAG,
21777 TargetLowering::DAGCombinerInfo &DCI,
21778 const X86Subtarget *Subtarget) {
21780 SDValue N0 = N->getOperand(0);
21781 SDValue N1 = N->getOperand(1);
21782 EVT VT = N->getValueType(0);
21784 // Don't create instructions with illegal types after legalize types has run.
21785 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
21786 if (!DCI.isBeforeLegalize() && !TLI.isTypeLegal(VT.getVectorElementType()))
21789 // If we have legalized the vector types, look for blends of FADD and FSUB
21790 // nodes that we can fuse into an ADDSUB node.
21791 if (TLI.isTypeLegal(VT) && Subtarget->hasSSE3())
21792 if (SDValue AddSub = combineShuffleToAddSub(N, DAG))
21795 // Combine 256-bit vector shuffles. This is only profitable when in AVX mode
21796 if (Subtarget->hasFp256() && VT.is256BitVector() &&
21797 N->getOpcode() == ISD::VECTOR_SHUFFLE)
21798 return PerformShuffleCombine256(N, DAG, DCI, Subtarget);
21800 // During Type Legalization, when promoting illegal vector types,
21801 // the backend might introduce new shuffle dag nodes and bitcasts.
21803 // This code performs the following transformation:
21804 // fold: (shuffle (bitcast (BINOP A, B)), Undef, <Mask>) ->
21805 // (shuffle (BINOP (bitcast A), (bitcast B)), Undef, <Mask>)
21807 // We do this only if both the bitcast and the BINOP dag nodes have
21808 // one use. Also, perform this transformation only if the new binary
21809 // operation is legal. This is to avoid introducing dag nodes that
21810 // potentially need to be further expanded (or custom lowered) into a
21811 // less optimal sequence of dag nodes.
21812 if (!DCI.isBeforeLegalize() && DCI.isBeforeLegalizeOps() &&
21813 N1.getOpcode() == ISD::UNDEF && N0.hasOneUse() &&
21814 N0.getOpcode() == ISD::BITCAST) {
21815 SDValue BC0 = N0.getOperand(0);
21816 EVT SVT = BC0.getValueType();
21817 unsigned Opcode = BC0.getOpcode();
21818 unsigned NumElts = VT.getVectorNumElements();
21820 if (BC0.hasOneUse() && SVT.isVector() &&
21821 SVT.getVectorNumElements() * 2 == NumElts &&
21822 TLI.isOperationLegal(Opcode, VT)) {
21823 bool CanFold = false;
21835 unsigned SVTNumElts = SVT.getVectorNumElements();
21836 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
21837 for (unsigned i = 0, e = SVTNumElts; i != e && CanFold; ++i)
21838 CanFold = SVOp->getMaskElt(i) == (int)(i * 2);
21839 for (unsigned i = SVTNumElts, e = NumElts; i != e && CanFold; ++i)
21840 CanFold = SVOp->getMaskElt(i) < 0;
21843 SDValue BC00 = DAG.getBitcast(VT, BC0.getOperand(0));
21844 SDValue BC01 = DAG.getBitcast(VT, BC0.getOperand(1));
21845 SDValue NewBinOp = DAG.getNode(BC0.getOpcode(), dl, VT, BC00, BC01);
21846 return DAG.getVectorShuffle(VT, dl, NewBinOp, N1, &SVOp->getMask()[0]);
21851 // Combine a vector_shuffle that is equal to build_vector load1, load2, load3,
21852 // load4, <0, 1, 2, 3> into a 128-bit load if the load addresses are
21853 // consecutive, non-overlapping, and in the right order.
21854 SmallVector<SDValue, 16> Elts;
21855 for (unsigned i = 0, e = VT.getVectorNumElements(); i != e; ++i)
21856 Elts.push_back(getShuffleScalarElt(N, i, DAG, 0));
21858 if (SDValue LD = EltsFromConsecutiveLoads(VT, Elts, dl, DAG, true))
21861 if (isTargetShuffle(N->getOpcode())) {
21863 PerformTargetShuffleCombine(SDValue(N, 0), DAG, DCI, Subtarget);
21864 if (Shuffle.getNode())
21867 // Try recursively combining arbitrary sequences of x86 shuffle
21868 // instructions into higher-order shuffles. We do this after combining
21869 // specific PSHUF instruction sequences into their minimal form so that we
21870 // can evaluate how many specialized shuffle instructions are involved in
21871 // a particular chain.
21872 SmallVector<int, 1> NonceMask; // Just a placeholder.
21873 NonceMask.push_back(0);
21874 if (combineX86ShufflesRecursively(SDValue(N, 0), SDValue(N, 0), NonceMask,
21875 /*Depth*/ 1, /*HasPSHUFB*/ false, DAG,
21877 return SDValue(); // This routine will use CombineTo to replace N.
21883 /// XFormVExtractWithShuffleIntoLoad - Check if a vector extract from a target
21884 /// specific shuffle of a load can be folded into a single element load.
21885 /// Similar handling for VECTOR_SHUFFLE is performed by DAGCombiner, but
21886 /// shuffles have been custom lowered so we need to handle those here.
21887 static SDValue XFormVExtractWithShuffleIntoLoad(SDNode *N, SelectionDAG &DAG,
21888 TargetLowering::DAGCombinerInfo &DCI) {
21889 if (DCI.isBeforeLegalizeOps())
21892 SDValue InVec = N->getOperand(0);
21893 SDValue EltNo = N->getOperand(1);
21895 if (!isa<ConstantSDNode>(EltNo))
21898 EVT OriginalVT = InVec.getValueType();
21900 if (InVec.getOpcode() == ISD::BITCAST) {
21901 // Don't duplicate a load with other uses.
21902 if (!InVec.hasOneUse())
21904 EVT BCVT = InVec.getOperand(0).getValueType();
21905 if (!BCVT.isVector() ||
21906 BCVT.getVectorNumElements() != OriginalVT.getVectorNumElements())
21908 InVec = InVec.getOperand(0);
21911 EVT CurrentVT = InVec.getValueType();
21913 if (!isTargetShuffle(InVec.getOpcode()))
21916 // Don't duplicate a load with other uses.
21917 if (!InVec.hasOneUse())
21920 SmallVector<int, 16> ShuffleMask;
21922 if (!getTargetShuffleMask(InVec.getNode(), CurrentVT.getSimpleVT(),
21923 ShuffleMask, UnaryShuffle))
21926 // Select the input vector, guarding against out of range extract vector.
21927 unsigned NumElems = CurrentVT.getVectorNumElements();
21928 int Elt = cast<ConstantSDNode>(EltNo)->getZExtValue();
21929 int Idx = (Elt > (int)NumElems) ? -1 : ShuffleMask[Elt];
21930 SDValue LdNode = (Idx < (int)NumElems) ? InVec.getOperand(0)
21931 : InVec.getOperand(1);
21933 // If inputs to shuffle are the same for both ops, then allow 2 uses
21934 unsigned AllowedUses = InVec.getNumOperands() > 1 &&
21935 InVec.getOperand(0) == InVec.getOperand(1) ? 2 : 1;
21937 if (LdNode.getOpcode() == ISD::BITCAST) {
21938 // Don't duplicate a load with other uses.
21939 if (!LdNode.getNode()->hasNUsesOfValue(AllowedUses, 0))
21942 AllowedUses = 1; // only allow 1 load use if we have a bitcast
21943 LdNode = LdNode.getOperand(0);
21946 if (!ISD::isNormalLoad(LdNode.getNode()))
21949 LoadSDNode *LN0 = cast<LoadSDNode>(LdNode);
21951 if (!LN0 ||!LN0->hasNUsesOfValue(AllowedUses, 0) || LN0->isVolatile())
21954 EVT EltVT = N->getValueType(0);
21955 // If there's a bitcast before the shuffle, check if the load type and
21956 // alignment is valid.
21957 unsigned Align = LN0->getAlignment();
21958 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
21959 unsigned NewAlign = DAG.getDataLayout().getABITypeAlignment(
21960 EltVT.getTypeForEVT(*DAG.getContext()));
21962 if (NewAlign > Align || !TLI.isOperationLegalOrCustom(ISD::LOAD, EltVT))
21965 // All checks match so transform back to vector_shuffle so that DAG combiner
21966 // can finish the job
21969 // Create shuffle node taking into account the case that its a unary shuffle
21970 SDValue Shuffle = (UnaryShuffle) ? DAG.getUNDEF(CurrentVT)
21971 : InVec.getOperand(1);
21972 Shuffle = DAG.getVectorShuffle(CurrentVT, dl,
21973 InVec.getOperand(0), Shuffle,
21975 Shuffle = DAG.getBitcast(OriginalVT, Shuffle);
21976 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, N->getValueType(0), Shuffle,
21980 /// \brief Detect bitcasts between i32 to x86mmx low word. Since MMX types are
21981 /// special and don't usually play with other vector types, it's better to
21982 /// handle them early to be sure we emit efficient code by avoiding
21983 /// store-load conversions.
21984 static SDValue PerformBITCASTCombine(SDNode *N, SelectionDAG &DAG) {
21985 if (N->getValueType(0) != MVT::x86mmx ||
21986 N->getOperand(0)->getOpcode() != ISD::BUILD_VECTOR ||
21987 N->getOperand(0)->getValueType(0) != MVT::v2i32)
21990 SDValue V = N->getOperand(0);
21991 ConstantSDNode *C = dyn_cast<ConstantSDNode>(V.getOperand(1));
21992 if (C && C->getZExtValue() == 0 && V.getOperand(0).getValueType() == MVT::i32)
21993 return DAG.getNode(X86ISD::MMX_MOVW2D, SDLoc(V.getOperand(0)),
21994 N->getValueType(0), V.getOperand(0));
21999 /// PerformEXTRACT_VECTOR_ELTCombine - Detect vector gather/scatter index
22000 /// generation and convert it from being a bunch of shuffles and extracts
22001 /// into a somewhat faster sequence. For i686, the best sequence is apparently
22002 /// storing the value and loading scalars back, while for x64 we should
22003 /// use 64-bit extracts and shifts.
22004 static SDValue PerformEXTRACT_VECTOR_ELTCombine(SDNode *N, SelectionDAG &DAG,
22005 TargetLowering::DAGCombinerInfo &DCI) {
22006 if (SDValue NewOp = XFormVExtractWithShuffleIntoLoad(N, DAG, DCI))
22009 SDValue InputVector = N->getOperand(0);
22010 SDLoc dl(InputVector);
22011 // Detect mmx to i32 conversion through a v2i32 elt extract.
22012 if (InputVector.getOpcode() == ISD::BITCAST && InputVector.hasOneUse() &&
22013 N->getValueType(0) == MVT::i32 &&
22014 InputVector.getValueType() == MVT::v2i32) {
22016 // The bitcast source is a direct mmx result.
22017 SDValue MMXSrc = InputVector.getNode()->getOperand(0);
22018 if (MMXSrc.getValueType() == MVT::x86mmx)
22019 return DAG.getNode(X86ISD::MMX_MOVD2W, SDLoc(InputVector),
22020 N->getValueType(0),
22021 InputVector.getNode()->getOperand(0));
22023 // The mmx is indirect: (i64 extract_elt (v1i64 bitcast (x86mmx ...))).
22024 SDValue MMXSrcOp = MMXSrc.getOperand(0);
22025 if (MMXSrc.getOpcode() == ISD::EXTRACT_VECTOR_ELT && MMXSrc.hasOneUse() &&
22026 MMXSrc.getValueType() == MVT::i64 && MMXSrcOp.hasOneUse() &&
22027 MMXSrcOp.getOpcode() == ISD::BITCAST &&
22028 MMXSrcOp.getValueType() == MVT::v1i64 &&
22029 MMXSrcOp.getOperand(0).getValueType() == MVT::x86mmx)
22030 return DAG.getNode(X86ISD::MMX_MOVD2W, SDLoc(InputVector),
22031 N->getValueType(0),
22032 MMXSrcOp.getOperand(0));
22035 EVT VT = N->getValueType(0);
22037 if (VT == MVT::i1 && dyn_cast<ConstantSDNode>(N->getOperand(1)) &&
22038 InputVector.getOpcode() == ISD::BITCAST &&
22039 dyn_cast<ConstantSDNode>(InputVector.getOperand(0))) {
22040 uint64_t ExtractedElt =
22041 cast<ConstantSDNode>(N->getOperand(1))->getZExtValue();
22042 uint64_t InputValue =
22043 cast<ConstantSDNode>(InputVector.getOperand(0))->getZExtValue();
22044 uint64_t Res = (InputValue >> ExtractedElt) & 1;
22045 return DAG.getConstant(Res, dl, MVT::i1);
22047 // Only operate on vectors of 4 elements, where the alternative shuffling
22048 // gets to be more expensive.
22049 if (InputVector.getValueType() != MVT::v4i32)
22052 // Check whether every use of InputVector is an EXTRACT_VECTOR_ELT with a
22053 // single use which is a sign-extend or zero-extend, and all elements are
22055 SmallVector<SDNode *, 4> Uses;
22056 unsigned ExtractedElements = 0;
22057 for (SDNode::use_iterator UI = InputVector.getNode()->use_begin(),
22058 UE = InputVector.getNode()->use_end(); UI != UE; ++UI) {
22059 if (UI.getUse().getResNo() != InputVector.getResNo())
22062 SDNode *Extract = *UI;
22063 if (Extract->getOpcode() != ISD::EXTRACT_VECTOR_ELT)
22066 if (Extract->getValueType(0) != MVT::i32)
22068 if (!Extract->hasOneUse())
22070 if (Extract->use_begin()->getOpcode() != ISD::SIGN_EXTEND &&
22071 Extract->use_begin()->getOpcode() != ISD::ZERO_EXTEND)
22073 if (!isa<ConstantSDNode>(Extract->getOperand(1)))
22076 // Record which element was extracted.
22077 ExtractedElements |=
22078 1 << cast<ConstantSDNode>(Extract->getOperand(1))->getZExtValue();
22080 Uses.push_back(Extract);
22083 // If not all the elements were used, this may not be worthwhile.
22084 if (ExtractedElements != 15)
22087 // Ok, we've now decided to do the transformation.
22088 // If 64-bit shifts are legal, use the extract-shift sequence,
22089 // otherwise bounce the vector off the cache.
22090 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
22093 if (TLI.isOperationLegal(ISD::SRA, MVT::i64)) {
22094 SDValue Cst = DAG.getBitcast(MVT::v2i64, InputVector);
22095 auto &DL = DAG.getDataLayout();
22096 EVT VecIdxTy = DAG.getTargetLoweringInfo().getVectorIdxTy(DL);
22097 SDValue BottomHalf = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i64, Cst,
22098 DAG.getConstant(0, dl, VecIdxTy));
22099 SDValue TopHalf = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i64, Cst,
22100 DAG.getConstant(1, dl, VecIdxTy));
22102 SDValue ShAmt = DAG.getConstant(
22103 32, dl, DAG.getTargetLoweringInfo().getShiftAmountTy(MVT::i64, DL));
22104 Vals[0] = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, BottomHalf);
22105 Vals[1] = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32,
22106 DAG.getNode(ISD::SRA, dl, MVT::i64, BottomHalf, ShAmt));
22107 Vals[2] = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, TopHalf);
22108 Vals[3] = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32,
22109 DAG.getNode(ISD::SRA, dl, MVT::i64, TopHalf, ShAmt));
22111 // Store the value to a temporary stack slot.
22112 SDValue StackPtr = DAG.CreateStackTemporary(InputVector.getValueType());
22113 SDValue Ch = DAG.getStore(DAG.getEntryNode(), dl, InputVector, StackPtr,
22114 MachinePointerInfo(), false, false, 0);
22116 EVT ElementType = InputVector.getValueType().getVectorElementType();
22117 unsigned EltSize = ElementType.getSizeInBits() / 8;
22119 // Replace each use (extract) with a load of the appropriate element.
22120 for (unsigned i = 0; i < 4; ++i) {
22121 uint64_t Offset = EltSize * i;
22122 auto PtrVT = TLI.getPointerTy(DAG.getDataLayout());
22123 SDValue OffsetVal = DAG.getConstant(Offset, dl, PtrVT);
22125 SDValue ScalarAddr =
22126 DAG.getNode(ISD::ADD, dl, PtrVT, StackPtr, OffsetVal);
22128 // Load the scalar.
22129 Vals[i] = DAG.getLoad(ElementType, dl, Ch,
22130 ScalarAddr, MachinePointerInfo(),
22131 false, false, false, 0);
22136 // Replace the extracts
22137 for (SmallVectorImpl<SDNode *>::iterator UI = Uses.begin(),
22138 UE = Uses.end(); UI != UE; ++UI) {
22139 SDNode *Extract = *UI;
22141 SDValue Idx = Extract->getOperand(1);
22142 uint64_t IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
22143 DAG.ReplaceAllUsesOfValueWith(SDValue(Extract, 0), Vals[IdxVal]);
22146 // The replacement was made in place; don't return anything.
22150 /// \brief Matches a VSELECT onto min/max or return 0 if the node doesn't match.
22151 static std::pair<unsigned, bool>
22152 matchIntegerMINMAX(SDValue Cond, EVT VT, SDValue LHS, SDValue RHS,
22153 SelectionDAG &DAG, const X86Subtarget *Subtarget) {
22154 if (!VT.isVector())
22155 return std::make_pair(0, false);
22157 bool NeedSplit = false;
22158 switch (VT.getSimpleVT().SimpleTy) {
22159 default: return std::make_pair(0, false);
22162 if (!Subtarget->hasVLX())
22163 return std::make_pair(0, false);
22167 if (!Subtarget->hasBWI())
22168 return std::make_pair(0, false);
22172 if (!Subtarget->hasAVX512())
22173 return std::make_pair(0, false);
22178 if (!Subtarget->hasAVX2())
22180 if (!Subtarget->hasAVX())
22181 return std::make_pair(0, false);
22186 if (!Subtarget->hasSSE2())
22187 return std::make_pair(0, false);
22190 // SSE2 has only a small subset of the operations.
22191 bool hasUnsigned = Subtarget->hasSSE41() ||
22192 (Subtarget->hasSSE2() && VT == MVT::v16i8);
22193 bool hasSigned = Subtarget->hasSSE41() ||
22194 (Subtarget->hasSSE2() && VT == MVT::v8i16);
22196 ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
22199 // Check for x CC y ? x : y.
22200 if (DAG.isEqualTo(LHS, Cond.getOperand(0)) &&
22201 DAG.isEqualTo(RHS, Cond.getOperand(1))) {
22206 Opc = hasUnsigned ? ISD::UMIN : 0; break;
22209 Opc = hasUnsigned ? ISD::UMAX : 0; break;
22212 Opc = hasSigned ? ISD::SMIN : 0; break;
22215 Opc = hasSigned ? ISD::SMAX : 0; break;
22217 // Check for x CC y ? y : x -- a min/max with reversed arms.
22218 } else if (DAG.isEqualTo(LHS, Cond.getOperand(1)) &&
22219 DAG.isEqualTo(RHS, Cond.getOperand(0))) {
22224 Opc = hasUnsigned ? ISD::UMAX : 0; break;
22227 Opc = hasUnsigned ? ISD::UMIN : 0; break;
22230 Opc = hasSigned ? ISD::SMAX : 0; break;
22233 Opc = hasSigned ? ISD::SMIN : 0; break;
22237 return std::make_pair(Opc, NeedSplit);
22241 transformVSELECTtoBlendVECTOR_SHUFFLE(SDNode *N, SelectionDAG &DAG,
22242 const X86Subtarget *Subtarget) {
22244 SDValue Cond = N->getOperand(0);
22245 SDValue LHS = N->getOperand(1);
22246 SDValue RHS = N->getOperand(2);
22248 if (Cond.getOpcode() == ISD::SIGN_EXTEND) {
22249 SDValue CondSrc = Cond->getOperand(0);
22250 if (CondSrc->getOpcode() == ISD::SIGN_EXTEND_INREG)
22251 Cond = CondSrc->getOperand(0);
22254 if (!ISD::isBuildVectorOfConstantSDNodes(Cond.getNode()))
22257 // A vselect where all conditions and data are constants can be optimized into
22258 // a single vector load by SelectionDAGLegalize::ExpandBUILD_VECTOR().
22259 if (ISD::isBuildVectorOfConstantSDNodes(LHS.getNode()) &&
22260 ISD::isBuildVectorOfConstantSDNodes(RHS.getNode()))
22263 unsigned MaskValue = 0;
22264 if (!BUILD_VECTORtoBlendMask(cast<BuildVectorSDNode>(Cond), MaskValue))
22267 MVT VT = N->getSimpleValueType(0);
22268 unsigned NumElems = VT.getVectorNumElements();
22269 SmallVector<int, 8> ShuffleMask(NumElems, -1);
22270 for (unsigned i = 0; i < NumElems; ++i) {
22271 // Be sure we emit undef where we can.
22272 if (Cond.getOperand(i)->getOpcode() == ISD::UNDEF)
22273 ShuffleMask[i] = -1;
22275 ShuffleMask[i] = i + NumElems * ((MaskValue >> i) & 1);
22278 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
22279 if (!TLI.isShuffleMaskLegal(ShuffleMask, VT))
22281 return DAG.getVectorShuffle(VT, dl, LHS, RHS, &ShuffleMask[0]);
22284 /// PerformSELECTCombine - Do target-specific dag combines on SELECT and VSELECT
22286 static SDValue PerformSELECTCombine(SDNode *N, SelectionDAG &DAG,
22287 TargetLowering::DAGCombinerInfo &DCI,
22288 const X86Subtarget *Subtarget) {
22290 SDValue Cond = N->getOperand(0);
22291 // Get the LHS/RHS of the select.
22292 SDValue LHS = N->getOperand(1);
22293 SDValue RHS = N->getOperand(2);
22294 EVT VT = LHS.getValueType();
22295 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
22297 // If we have SSE[12] support, try to form min/max nodes. SSE min/max
22298 // instructions match the semantics of the common C idiom x<y?x:y but not
22299 // x<=y?x:y, because of how they handle negative zero (which can be
22300 // ignored in unsafe-math mode).
22301 // We also try to create v2f32 min/max nodes, which we later widen to v4f32.
22302 if (Cond.getOpcode() == ISD::SETCC && VT.isFloatingPoint() &&
22303 VT != MVT::f80 && (TLI.isTypeLegal(VT) || VT == MVT::v2f32) &&
22304 (Subtarget->hasSSE2() ||
22305 (Subtarget->hasSSE1() && VT.getScalarType() == MVT::f32))) {
22306 ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
22308 unsigned Opcode = 0;
22309 // Check for x CC y ? x : y.
22310 if (DAG.isEqualTo(LHS, Cond.getOperand(0)) &&
22311 DAG.isEqualTo(RHS, Cond.getOperand(1))) {
22315 // Converting this to a min would handle NaNs incorrectly, and swapping
22316 // the operands would cause it to handle comparisons between positive
22317 // and negative zero incorrectly.
22318 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)) {
22319 if (!DAG.getTarget().Options.UnsafeFPMath &&
22320 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
22322 std::swap(LHS, RHS);
22324 Opcode = X86ISD::FMIN;
22327 // Converting this to a min would handle comparisons between positive
22328 // and negative zero incorrectly.
22329 if (!DAG.getTarget().Options.UnsafeFPMath &&
22330 !DAG.isKnownNeverZero(LHS) && !DAG.isKnownNeverZero(RHS))
22332 Opcode = X86ISD::FMIN;
22335 // Converting this to a min would handle both negative zeros and NaNs
22336 // incorrectly, but we can swap the operands to fix both.
22337 std::swap(LHS, RHS);
22341 Opcode = X86ISD::FMIN;
22345 // Converting this to a max would handle comparisons between positive
22346 // and negative zero incorrectly.
22347 if (!DAG.getTarget().Options.UnsafeFPMath &&
22348 !DAG.isKnownNeverZero(LHS) && !DAG.isKnownNeverZero(RHS))
22350 Opcode = X86ISD::FMAX;
22353 // Converting this to a max would handle NaNs incorrectly, and swapping
22354 // the operands would cause it to handle comparisons between positive
22355 // and negative zero incorrectly.
22356 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)) {
22357 if (!DAG.getTarget().Options.UnsafeFPMath &&
22358 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
22360 std::swap(LHS, RHS);
22362 Opcode = X86ISD::FMAX;
22365 // Converting this to a max would handle both negative zeros and NaNs
22366 // incorrectly, but we can swap the operands to fix both.
22367 std::swap(LHS, RHS);
22371 Opcode = X86ISD::FMAX;
22374 // Check for x CC y ? y : x -- a min/max with reversed arms.
22375 } else if (DAG.isEqualTo(LHS, Cond.getOperand(1)) &&
22376 DAG.isEqualTo(RHS, Cond.getOperand(0))) {
22380 // Converting this to a min would handle comparisons between positive
22381 // and negative zero incorrectly, and swapping the operands would
22382 // cause it to handle NaNs incorrectly.
22383 if (!DAG.getTarget().Options.UnsafeFPMath &&
22384 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS))) {
22385 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS))
22387 std::swap(LHS, RHS);
22389 Opcode = X86ISD::FMIN;
22392 // Converting this to a min would handle NaNs incorrectly.
22393 if (!DAG.getTarget().Options.UnsafeFPMath &&
22394 (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)))
22396 Opcode = X86ISD::FMIN;
22399 // Converting this to a min would handle both negative zeros and NaNs
22400 // incorrectly, but we can swap the operands to fix both.
22401 std::swap(LHS, RHS);
22405 Opcode = X86ISD::FMIN;
22409 // Converting this to a max would handle NaNs incorrectly.
22410 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS))
22412 Opcode = X86ISD::FMAX;
22415 // Converting this to a max would handle comparisons between positive
22416 // and negative zero incorrectly, and swapping the operands would
22417 // cause it to handle NaNs incorrectly.
22418 if (!DAG.getTarget().Options.UnsafeFPMath &&
22419 !DAG.isKnownNeverZero(LHS) && !DAG.isKnownNeverZero(RHS)) {
22420 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS))
22422 std::swap(LHS, RHS);
22424 Opcode = X86ISD::FMAX;
22427 // Converting this to a max would handle both negative zeros and NaNs
22428 // incorrectly, but we can swap the operands to fix both.
22429 std::swap(LHS, RHS);
22433 Opcode = X86ISD::FMAX;
22439 return DAG.getNode(Opcode, DL, N->getValueType(0), LHS, RHS);
22442 EVT CondVT = Cond.getValueType();
22443 if (Subtarget->hasAVX512() && VT.isVector() && CondVT.isVector() &&
22444 CondVT.getVectorElementType() == MVT::i1) {
22445 // v16i8 (select v16i1, v16i8, v16i8) does not have a proper
22446 // lowering on KNL. In this case we convert it to
22447 // v16i8 (select v16i8, v16i8, v16i8) and use AVX instruction.
22448 // The same situation for all 128 and 256-bit vectors of i8 and i16.
22449 // Since SKX these selects have a proper lowering.
22450 EVT OpVT = LHS.getValueType();
22451 if ((OpVT.is128BitVector() || OpVT.is256BitVector()) &&
22452 (OpVT.getVectorElementType() == MVT::i8 ||
22453 OpVT.getVectorElementType() == MVT::i16) &&
22454 !(Subtarget->hasBWI() && Subtarget->hasVLX())) {
22455 Cond = DAG.getNode(ISD::SIGN_EXTEND, DL, OpVT, Cond);
22456 DCI.AddToWorklist(Cond.getNode());
22457 return DAG.getNode(N->getOpcode(), DL, OpVT, Cond, LHS, RHS);
22460 // If this is a select between two integer constants, try to do some
22462 if (ConstantSDNode *TrueC = dyn_cast<ConstantSDNode>(LHS)) {
22463 if (ConstantSDNode *FalseC = dyn_cast<ConstantSDNode>(RHS))
22464 // Don't do this for crazy integer types.
22465 if (DAG.getTargetLoweringInfo().isTypeLegal(LHS.getValueType())) {
22466 // If this is efficiently invertible, canonicalize the LHSC/RHSC values
22467 // so that TrueC (the true value) is larger than FalseC.
22468 bool NeedsCondInvert = false;
22470 if (TrueC->getAPIntValue().ult(FalseC->getAPIntValue()) &&
22471 // Efficiently invertible.
22472 (Cond.getOpcode() == ISD::SETCC || // setcc -> invertible.
22473 (Cond.getOpcode() == ISD::XOR && // xor(X, C) -> invertible.
22474 isa<ConstantSDNode>(Cond.getOperand(1))))) {
22475 NeedsCondInvert = true;
22476 std::swap(TrueC, FalseC);
22479 // Optimize C ? 8 : 0 -> zext(C) << 3. Likewise for any pow2/0.
22480 if (FalseC->getAPIntValue() == 0 &&
22481 TrueC->getAPIntValue().isPowerOf2()) {
22482 if (NeedsCondInvert) // Invert the condition if needed.
22483 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
22484 DAG.getConstant(1, DL, Cond.getValueType()));
22486 // Zero extend the condition if needed.
22487 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, LHS.getValueType(), Cond);
22489 unsigned ShAmt = TrueC->getAPIntValue().logBase2();
22490 return DAG.getNode(ISD::SHL, DL, LHS.getValueType(), Cond,
22491 DAG.getConstant(ShAmt, DL, MVT::i8));
22494 // Optimize Cond ? cst+1 : cst -> zext(setcc(C)+cst.
22495 if (FalseC->getAPIntValue()+1 == TrueC->getAPIntValue()) {
22496 if (NeedsCondInvert) // Invert the condition if needed.
22497 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
22498 DAG.getConstant(1, DL, Cond.getValueType()));
22500 // Zero extend the condition if needed.
22501 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL,
22502 FalseC->getValueType(0), Cond);
22503 return DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
22504 SDValue(FalseC, 0));
22507 // Optimize cases that will turn into an LEA instruction. This requires
22508 // an i32 or i64 and an efficient multiplier (1, 2, 3, 4, 5, 8, 9).
22509 if (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i64) {
22510 uint64_t Diff = TrueC->getZExtValue()-FalseC->getZExtValue();
22511 if (N->getValueType(0) == MVT::i32) Diff = (unsigned)Diff;
22513 bool isFastMultiplier = false;
22515 switch ((unsigned char)Diff) {
22517 case 1: // result = add base, cond
22518 case 2: // result = lea base( , cond*2)
22519 case 3: // result = lea base(cond, cond*2)
22520 case 4: // result = lea base( , cond*4)
22521 case 5: // result = lea base(cond, cond*4)
22522 case 8: // result = lea base( , cond*8)
22523 case 9: // result = lea base(cond, cond*8)
22524 isFastMultiplier = true;
22529 if (isFastMultiplier) {
22530 APInt Diff = TrueC->getAPIntValue()-FalseC->getAPIntValue();
22531 if (NeedsCondInvert) // Invert the condition if needed.
22532 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
22533 DAG.getConstant(1, DL, Cond.getValueType()));
22535 // Zero extend the condition if needed.
22536 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, FalseC->getValueType(0),
22538 // Scale the condition by the difference.
22540 Cond = DAG.getNode(ISD::MUL, DL, Cond.getValueType(), Cond,
22541 DAG.getConstant(Diff, DL,
22542 Cond.getValueType()));
22544 // Add the base if non-zero.
22545 if (FalseC->getAPIntValue() != 0)
22546 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
22547 SDValue(FalseC, 0));
22554 // Canonicalize max and min:
22555 // (x > y) ? x : y -> (x >= y) ? x : y
22556 // (x < y) ? x : y -> (x <= y) ? x : y
22557 // This allows use of COND_S / COND_NS (see TranslateX86CC) which eliminates
22558 // the need for an extra compare
22559 // against zero. e.g.
22560 // (x - y) > 0 : (x - y) ? 0 -> (x - y) >= 0 : (x - y) ? 0
22562 // testl %edi, %edi
22564 // cmovgl %edi, %eax
22568 // cmovsl %eax, %edi
22569 if (N->getOpcode() == ISD::SELECT && Cond.getOpcode() == ISD::SETCC &&
22570 DAG.isEqualTo(LHS, Cond.getOperand(0)) &&
22571 DAG.isEqualTo(RHS, Cond.getOperand(1))) {
22572 ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
22577 ISD::CondCode NewCC = (CC == ISD::SETLT) ? ISD::SETLE : ISD::SETGE;
22578 Cond = DAG.getSetCC(SDLoc(Cond), Cond.getValueType(),
22579 Cond.getOperand(0), Cond.getOperand(1), NewCC);
22580 return DAG.getNode(ISD::SELECT, DL, VT, Cond, LHS, RHS);
22585 // Early exit check
22586 if (!TLI.isTypeLegal(VT))
22589 // Match VSELECTs into subs with unsigned saturation.
22590 if (N->getOpcode() == ISD::VSELECT && Cond.getOpcode() == ISD::SETCC &&
22591 // psubus is available in SSE2 and AVX2 for i8 and i16 vectors.
22592 ((Subtarget->hasSSE2() && (VT == MVT::v16i8 || VT == MVT::v8i16)) ||
22593 (Subtarget->hasAVX2() && (VT == MVT::v32i8 || VT == MVT::v16i16)))) {
22594 ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
22596 // Check if one of the arms of the VSELECT is a zero vector. If it's on the
22597 // left side invert the predicate to simplify logic below.
22599 if (ISD::isBuildVectorAllZeros(LHS.getNode())) {
22601 CC = ISD::getSetCCInverse(CC, true);
22602 } else if (ISD::isBuildVectorAllZeros(RHS.getNode())) {
22606 if (Other.getNode() && Other->getNumOperands() == 2 &&
22607 DAG.isEqualTo(Other->getOperand(0), Cond.getOperand(0))) {
22608 SDValue OpLHS = Other->getOperand(0), OpRHS = Other->getOperand(1);
22609 SDValue CondRHS = Cond->getOperand(1);
22611 // Look for a general sub with unsigned saturation first.
22612 // x >= y ? x-y : 0 --> subus x, y
22613 // x > y ? x-y : 0 --> subus x, y
22614 if ((CC == ISD::SETUGE || CC == ISD::SETUGT) &&
22615 Other->getOpcode() == ISD::SUB && DAG.isEqualTo(OpRHS, CondRHS))
22616 return DAG.getNode(X86ISD::SUBUS, DL, VT, OpLHS, OpRHS);
22618 if (auto *OpRHSBV = dyn_cast<BuildVectorSDNode>(OpRHS))
22619 if (auto *OpRHSConst = OpRHSBV->getConstantSplatNode()) {
22620 if (auto *CondRHSBV = dyn_cast<BuildVectorSDNode>(CondRHS))
22621 if (auto *CondRHSConst = CondRHSBV->getConstantSplatNode())
22622 // If the RHS is a constant we have to reverse the const
22623 // canonicalization.
22624 // x > C-1 ? x+-C : 0 --> subus x, C
22625 if (CC == ISD::SETUGT && Other->getOpcode() == ISD::ADD &&
22626 CondRHSConst->getAPIntValue() ==
22627 (-OpRHSConst->getAPIntValue() - 1))
22628 return DAG.getNode(
22629 X86ISD::SUBUS, DL, VT, OpLHS,
22630 DAG.getConstant(-OpRHSConst->getAPIntValue(), DL, VT));
22632 // Another special case: If C was a sign bit, the sub has been
22633 // canonicalized into a xor.
22634 // FIXME: Would it be better to use computeKnownBits to determine
22635 // whether it's safe to decanonicalize the xor?
22636 // x s< 0 ? x^C : 0 --> subus x, C
22637 if (CC == ISD::SETLT && Other->getOpcode() == ISD::XOR &&
22638 ISD::isBuildVectorAllZeros(CondRHS.getNode()) &&
22639 OpRHSConst->getAPIntValue().isSignBit())
22640 // Note that we have to rebuild the RHS constant here to ensure we
22641 // don't rely on particular values of undef lanes.
22642 return DAG.getNode(
22643 X86ISD::SUBUS, DL, VT, OpLHS,
22644 DAG.getConstant(OpRHSConst->getAPIntValue(), DL, VT));
22649 // Try to match a min/max vector operation.
22650 if (N->getOpcode() == ISD::VSELECT && Cond.getOpcode() == ISD::SETCC) {
22651 std::pair<unsigned, bool> ret = matchIntegerMINMAX(Cond, VT, LHS, RHS, DAG, Subtarget);
22652 unsigned Opc = ret.first;
22653 bool NeedSplit = ret.second;
22655 if (Opc && NeedSplit) {
22656 unsigned NumElems = VT.getVectorNumElements();
22657 // Extract the LHS vectors
22658 SDValue LHS1 = Extract128BitVector(LHS, 0, DAG, DL);
22659 SDValue LHS2 = Extract128BitVector(LHS, NumElems/2, DAG, DL);
22661 // Extract the RHS vectors
22662 SDValue RHS1 = Extract128BitVector(RHS, 0, DAG, DL);
22663 SDValue RHS2 = Extract128BitVector(RHS, NumElems/2, DAG, DL);
22665 // Create min/max for each subvector
22666 LHS = DAG.getNode(Opc, DL, LHS1.getValueType(), LHS1, RHS1);
22667 RHS = DAG.getNode(Opc, DL, LHS2.getValueType(), LHS2, RHS2);
22669 // Merge the result
22670 return DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, LHS, RHS);
22672 return DAG.getNode(Opc, DL, VT, LHS, RHS);
22675 // Simplify vector selection if condition value type matches vselect
22677 if (N->getOpcode() == ISD::VSELECT && CondVT == VT) {
22678 assert(Cond.getValueType().isVector() &&
22679 "vector select expects a vector selector!");
22681 bool TValIsAllOnes = ISD::isBuildVectorAllOnes(LHS.getNode());
22682 bool FValIsAllZeros = ISD::isBuildVectorAllZeros(RHS.getNode());
22684 // Try invert the condition if true value is not all 1s and false value
22686 if (!TValIsAllOnes && !FValIsAllZeros &&
22687 // Check if the selector will be produced by CMPP*/PCMP*
22688 Cond.getOpcode() == ISD::SETCC &&
22689 // Check if SETCC has already been promoted
22690 TLI.getSetCCResultType(DAG.getDataLayout(), *DAG.getContext(), VT) ==
22692 bool TValIsAllZeros = ISD::isBuildVectorAllZeros(LHS.getNode());
22693 bool FValIsAllOnes = ISD::isBuildVectorAllOnes(RHS.getNode());
22695 if (TValIsAllZeros || FValIsAllOnes) {
22696 SDValue CC = Cond.getOperand(2);
22697 ISD::CondCode NewCC =
22698 ISD::getSetCCInverse(cast<CondCodeSDNode>(CC)->get(),
22699 Cond.getOperand(0).getValueType().isInteger());
22700 Cond = DAG.getSetCC(DL, CondVT, Cond.getOperand(0), Cond.getOperand(1), NewCC);
22701 std::swap(LHS, RHS);
22702 TValIsAllOnes = FValIsAllOnes;
22703 FValIsAllZeros = TValIsAllZeros;
22707 if (TValIsAllOnes || FValIsAllZeros) {
22710 if (TValIsAllOnes && FValIsAllZeros)
22712 else if (TValIsAllOnes)
22714 DAG.getNode(ISD::OR, DL, CondVT, Cond, DAG.getBitcast(CondVT, RHS));
22715 else if (FValIsAllZeros)
22716 Ret = DAG.getNode(ISD::AND, DL, CondVT, Cond,
22717 DAG.getBitcast(CondVT, LHS));
22719 return DAG.getBitcast(VT, Ret);
22723 // We should generate an X86ISD::BLENDI from a vselect if its argument
22724 // is a sign_extend_inreg of an any_extend of a BUILD_VECTOR of
22725 // constants. This specific pattern gets generated when we split a
22726 // selector for a 512 bit vector in a machine without AVX512 (but with
22727 // 256-bit vectors), during legalization:
22729 // (vselect (sign_extend (any_extend (BUILD_VECTOR)) i1) LHS RHS)
22731 // Iff we find this pattern and the build_vectors are built from
22732 // constants, we translate the vselect into a shuffle_vector that we
22733 // know will be matched by LowerVECTOR_SHUFFLEtoBlend.
22734 if ((N->getOpcode() == ISD::VSELECT ||
22735 N->getOpcode() == X86ISD::SHRUNKBLEND) &&
22736 !DCI.isBeforeLegalize() && !VT.is512BitVector()) {
22737 SDValue Shuffle = transformVSELECTtoBlendVECTOR_SHUFFLE(N, DAG, Subtarget);
22738 if (Shuffle.getNode())
22742 // If this is a *dynamic* select (non-constant condition) and we can match
22743 // this node with one of the variable blend instructions, restructure the
22744 // condition so that the blends can use the high bit of each element and use
22745 // SimplifyDemandedBits to simplify the condition operand.
22746 if (N->getOpcode() == ISD::VSELECT && DCI.isBeforeLegalizeOps() &&
22747 !DCI.isBeforeLegalize() &&
22748 !ISD::isBuildVectorOfConstantSDNodes(Cond.getNode())) {
22749 unsigned BitWidth = Cond.getValueType().getScalarType().getSizeInBits();
22751 // Don't optimize vector selects that map to mask-registers.
22755 // We can only handle the cases where VSELECT is directly legal on the
22756 // subtarget. We custom lower VSELECT nodes with constant conditions and
22757 // this makes it hard to see whether a dynamic VSELECT will correctly
22758 // lower, so we both check the operation's status and explicitly handle the
22759 // cases where a *dynamic* blend will fail even though a constant-condition
22760 // blend could be custom lowered.
22761 // FIXME: We should find a better way to handle this class of problems.
22762 // Potentially, we should combine constant-condition vselect nodes
22763 // pre-legalization into shuffles and not mark as many types as custom
22765 if (!TLI.isOperationLegalOrCustom(ISD::VSELECT, VT))
22767 // FIXME: We don't support i16-element blends currently. We could and
22768 // should support them by making *all* the bits in the condition be set
22769 // rather than just the high bit and using an i8-element blend.
22770 if (VT.getScalarType() == MVT::i16)
22772 // Dynamic blending was only available from SSE4.1 onward.
22773 if (VT.getSizeInBits() == 128 && !Subtarget->hasSSE41())
22775 // Byte blends are only available in AVX2
22776 if (VT.getSizeInBits() == 256 && VT.getScalarType() == MVT::i8 &&
22777 !Subtarget->hasAVX2())
22780 assert(BitWidth >= 8 && BitWidth <= 64 && "Invalid mask size");
22781 APInt DemandedMask = APInt::getHighBitsSet(BitWidth, 1);
22783 APInt KnownZero, KnownOne;
22784 TargetLowering::TargetLoweringOpt TLO(DAG, DCI.isBeforeLegalize(),
22785 DCI.isBeforeLegalizeOps());
22786 if (TLO.ShrinkDemandedConstant(Cond, DemandedMask) ||
22787 TLI.SimplifyDemandedBits(Cond, DemandedMask, KnownZero, KnownOne,
22789 // If we changed the computation somewhere in the DAG, this change
22790 // will affect all users of Cond.
22791 // Make sure it is fine and update all the nodes so that we do not
22792 // use the generic VSELECT anymore. Otherwise, we may perform
22793 // wrong optimizations as we messed up with the actual expectation
22794 // for the vector boolean values.
22795 if (Cond != TLO.Old) {
22796 // Check all uses of that condition operand to check whether it will be
22797 // consumed by non-BLEND instructions, which may depend on all bits are
22799 for (SDNode::use_iterator I = Cond->use_begin(), E = Cond->use_end();
22801 if (I->getOpcode() != ISD::VSELECT)
22802 // TODO: Add other opcodes eventually lowered into BLEND.
22805 // Update all the users of the condition, before committing the change,
22806 // so that the VSELECT optimizations that expect the correct vector
22807 // boolean value will not be triggered.
22808 for (SDNode::use_iterator I = Cond->use_begin(), E = Cond->use_end();
22810 DAG.ReplaceAllUsesOfValueWith(
22812 DAG.getNode(X86ISD::SHRUNKBLEND, SDLoc(*I), I->getValueType(0),
22813 Cond, I->getOperand(1), I->getOperand(2)));
22814 DCI.CommitTargetLoweringOpt(TLO);
22817 // At this point, only Cond is changed. Change the condition
22818 // just for N to keep the opportunity to optimize all other
22819 // users their own way.
22820 DAG.ReplaceAllUsesOfValueWith(
22822 DAG.getNode(X86ISD::SHRUNKBLEND, SDLoc(N), N->getValueType(0),
22823 TLO.New, N->getOperand(1), N->getOperand(2)));
22831 // Check whether a boolean test is testing a boolean value generated by
22832 // X86ISD::SETCC. If so, return the operand of that SETCC and proper condition
22835 // Simplify the following patterns:
22836 // (Op (CMP (SETCC Cond EFLAGS) 1) EQ) or
22837 // (Op (CMP (SETCC Cond EFLAGS) 0) NEQ)
22838 // to (Op EFLAGS Cond)
22840 // (Op (CMP (SETCC Cond EFLAGS) 0) EQ) or
22841 // (Op (CMP (SETCC Cond EFLAGS) 1) NEQ)
22842 // to (Op EFLAGS !Cond)
22844 // where Op could be BRCOND or CMOV.
22846 static SDValue checkBoolTestSetCCCombine(SDValue Cmp, X86::CondCode &CC) {
22847 // Quit if not CMP and SUB with its value result used.
22848 if (Cmp.getOpcode() != X86ISD::CMP &&
22849 (Cmp.getOpcode() != X86ISD::SUB || Cmp.getNode()->hasAnyUseOfValue(0)))
22852 // Quit if not used as a boolean value.
22853 if (CC != X86::COND_E && CC != X86::COND_NE)
22856 // Check CMP operands. One of them should be 0 or 1 and the other should be
22857 // an SetCC or extended from it.
22858 SDValue Op1 = Cmp.getOperand(0);
22859 SDValue Op2 = Cmp.getOperand(1);
22862 const ConstantSDNode* C = nullptr;
22863 bool needOppositeCond = (CC == X86::COND_E);
22864 bool checkAgainstTrue = false; // Is it a comparison against 1?
22866 if ((C = dyn_cast<ConstantSDNode>(Op1)))
22868 else if ((C = dyn_cast<ConstantSDNode>(Op2)))
22870 else // Quit if all operands are not constants.
22873 if (C->getZExtValue() == 1) {
22874 needOppositeCond = !needOppositeCond;
22875 checkAgainstTrue = true;
22876 } else if (C->getZExtValue() != 0)
22877 // Quit if the constant is neither 0 or 1.
22880 bool truncatedToBoolWithAnd = false;
22881 // Skip (zext $x), (trunc $x), or (and $x, 1) node.
22882 while (SetCC.getOpcode() == ISD::ZERO_EXTEND ||
22883 SetCC.getOpcode() == ISD::TRUNCATE ||
22884 SetCC.getOpcode() == ISD::AND) {
22885 if (SetCC.getOpcode() == ISD::AND) {
22887 ConstantSDNode *CS;
22888 if ((CS = dyn_cast<ConstantSDNode>(SetCC.getOperand(0))) &&
22889 CS->getZExtValue() == 1)
22891 if ((CS = dyn_cast<ConstantSDNode>(SetCC.getOperand(1))) &&
22892 CS->getZExtValue() == 1)
22896 SetCC = SetCC.getOperand(OpIdx);
22897 truncatedToBoolWithAnd = true;
22899 SetCC = SetCC.getOperand(0);
22902 switch (SetCC.getOpcode()) {
22903 case X86ISD::SETCC_CARRY:
22904 // Since SETCC_CARRY gives output based on R = CF ? ~0 : 0, it's unsafe to
22905 // simplify it if the result of SETCC_CARRY is not canonicalized to 0 or 1,
22906 // i.e. it's a comparison against true but the result of SETCC_CARRY is not
22907 // truncated to i1 using 'and'.
22908 if (checkAgainstTrue && !truncatedToBoolWithAnd)
22910 assert(X86::CondCode(SetCC.getConstantOperandVal(0)) == X86::COND_B &&
22911 "Invalid use of SETCC_CARRY!");
22913 case X86ISD::SETCC:
22914 // Set the condition code or opposite one if necessary.
22915 CC = X86::CondCode(SetCC.getConstantOperandVal(0));
22916 if (needOppositeCond)
22917 CC = X86::GetOppositeBranchCondition(CC);
22918 return SetCC.getOperand(1);
22919 case X86ISD::CMOV: {
22920 // Check whether false/true value has canonical one, i.e. 0 or 1.
22921 ConstantSDNode *FVal = dyn_cast<ConstantSDNode>(SetCC.getOperand(0));
22922 ConstantSDNode *TVal = dyn_cast<ConstantSDNode>(SetCC.getOperand(1));
22923 // Quit if true value is not a constant.
22926 // Quit if false value is not a constant.
22928 SDValue Op = SetCC.getOperand(0);
22929 // Skip 'zext' or 'trunc' node.
22930 if (Op.getOpcode() == ISD::ZERO_EXTEND ||
22931 Op.getOpcode() == ISD::TRUNCATE)
22932 Op = Op.getOperand(0);
22933 // A special case for rdrand/rdseed, where 0 is set if false cond is
22935 if ((Op.getOpcode() != X86ISD::RDRAND &&
22936 Op.getOpcode() != X86ISD::RDSEED) || Op.getResNo() != 0)
22939 // Quit if false value is not the constant 0 or 1.
22940 bool FValIsFalse = true;
22941 if (FVal && FVal->getZExtValue() != 0) {
22942 if (FVal->getZExtValue() != 1)
22944 // If FVal is 1, opposite cond is needed.
22945 needOppositeCond = !needOppositeCond;
22946 FValIsFalse = false;
22948 // Quit if TVal is not the constant opposite of FVal.
22949 if (FValIsFalse && TVal->getZExtValue() != 1)
22951 if (!FValIsFalse && TVal->getZExtValue() != 0)
22953 CC = X86::CondCode(SetCC.getConstantOperandVal(2));
22954 if (needOppositeCond)
22955 CC = X86::GetOppositeBranchCondition(CC);
22956 return SetCC.getOperand(3);
22963 /// Check whether Cond is an AND/OR of SETCCs off of the same EFLAGS.
22965 /// (X86or (X86setcc) (X86setcc))
22966 /// (X86cmp (and (X86setcc) (X86setcc)), 0)
22967 static bool checkBoolTestAndOrSetCCCombine(SDValue Cond, X86::CondCode &CC0,
22968 X86::CondCode &CC1, SDValue &Flags,
22970 if (Cond->getOpcode() == X86ISD::CMP) {
22971 ConstantSDNode *CondOp1C = dyn_cast<ConstantSDNode>(Cond->getOperand(1));
22972 if (!CondOp1C || !CondOp1C->isNullValue())
22975 Cond = Cond->getOperand(0);
22980 SDValue SetCC0, SetCC1;
22981 switch (Cond->getOpcode()) {
22982 default: return false;
22989 SetCC0 = Cond->getOperand(0);
22990 SetCC1 = Cond->getOperand(1);
22994 // Make sure we have SETCC nodes, using the same flags value.
22995 if (SetCC0.getOpcode() != X86ISD::SETCC ||
22996 SetCC1.getOpcode() != X86ISD::SETCC ||
22997 SetCC0->getOperand(1) != SetCC1->getOperand(1))
23000 CC0 = (X86::CondCode)SetCC0->getConstantOperandVal(0);
23001 CC1 = (X86::CondCode)SetCC1->getConstantOperandVal(0);
23002 Flags = SetCC0->getOperand(1);
23006 /// Optimize X86ISD::CMOV [LHS, RHS, CONDCODE (e.g. X86::COND_NE), CONDVAL]
23007 static SDValue PerformCMOVCombine(SDNode *N, SelectionDAG &DAG,
23008 TargetLowering::DAGCombinerInfo &DCI,
23009 const X86Subtarget *Subtarget) {
23012 // If the flag operand isn't dead, don't touch this CMOV.
23013 if (N->getNumValues() == 2 && !SDValue(N, 1).use_empty())
23016 SDValue FalseOp = N->getOperand(0);
23017 SDValue TrueOp = N->getOperand(1);
23018 X86::CondCode CC = (X86::CondCode)N->getConstantOperandVal(2);
23019 SDValue Cond = N->getOperand(3);
23021 if (CC == X86::COND_E || CC == X86::COND_NE) {
23022 switch (Cond.getOpcode()) {
23026 // If operand of BSR / BSF are proven never zero, then ZF cannot be set.
23027 if (DAG.isKnownNeverZero(Cond.getOperand(0)))
23028 return (CC == X86::COND_E) ? FalseOp : TrueOp;
23034 Flags = checkBoolTestSetCCCombine(Cond, CC);
23035 if (Flags.getNode() &&
23036 // Extra check as FCMOV only supports a subset of X86 cond.
23037 (FalseOp.getValueType() != MVT::f80 || hasFPCMov(CC))) {
23038 SDValue Ops[] = { FalseOp, TrueOp,
23039 DAG.getConstant(CC, DL, MVT::i8), Flags };
23040 return DAG.getNode(X86ISD::CMOV, DL, N->getVTList(), Ops);
23043 // If this is a select between two integer constants, try to do some
23044 // optimizations. Note that the operands are ordered the opposite of SELECT
23046 if (ConstantSDNode *TrueC = dyn_cast<ConstantSDNode>(TrueOp)) {
23047 if (ConstantSDNode *FalseC = dyn_cast<ConstantSDNode>(FalseOp)) {
23048 // Canonicalize the TrueC/FalseC values so that TrueC (the true value) is
23049 // larger than FalseC (the false value).
23050 if (TrueC->getAPIntValue().ult(FalseC->getAPIntValue())) {
23051 CC = X86::GetOppositeBranchCondition(CC);
23052 std::swap(TrueC, FalseC);
23053 std::swap(TrueOp, FalseOp);
23056 // Optimize C ? 8 : 0 -> zext(setcc(C)) << 3. Likewise for any pow2/0.
23057 // This is efficient for any integer data type (including i8/i16) and
23059 if (FalseC->getAPIntValue() == 0 && TrueC->getAPIntValue().isPowerOf2()) {
23060 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
23061 DAG.getConstant(CC, DL, MVT::i8), Cond);
23063 // Zero extend the condition if needed.
23064 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, TrueC->getValueType(0), Cond);
23066 unsigned ShAmt = TrueC->getAPIntValue().logBase2();
23067 Cond = DAG.getNode(ISD::SHL, DL, Cond.getValueType(), Cond,
23068 DAG.getConstant(ShAmt, DL, MVT::i8));
23069 if (N->getNumValues() == 2) // Dead flag value?
23070 return DCI.CombineTo(N, Cond, SDValue());
23074 // Optimize Cond ? cst+1 : cst -> zext(setcc(C)+cst. This is efficient
23075 // for any integer data type, including i8/i16.
23076 if (FalseC->getAPIntValue()+1 == TrueC->getAPIntValue()) {
23077 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
23078 DAG.getConstant(CC, DL, MVT::i8), Cond);
23080 // Zero extend the condition if needed.
23081 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL,
23082 FalseC->getValueType(0), Cond);
23083 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
23084 SDValue(FalseC, 0));
23086 if (N->getNumValues() == 2) // Dead flag value?
23087 return DCI.CombineTo(N, Cond, SDValue());
23091 // Optimize cases that will turn into an LEA instruction. This requires
23092 // an i32 or i64 and an efficient multiplier (1, 2, 3, 4, 5, 8, 9).
23093 if (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i64) {
23094 uint64_t Diff = TrueC->getZExtValue()-FalseC->getZExtValue();
23095 if (N->getValueType(0) == MVT::i32) Diff = (unsigned)Diff;
23097 bool isFastMultiplier = false;
23099 switch ((unsigned char)Diff) {
23101 case 1: // result = add base, cond
23102 case 2: // result = lea base( , cond*2)
23103 case 3: // result = lea base(cond, cond*2)
23104 case 4: // result = lea base( , cond*4)
23105 case 5: // result = lea base(cond, cond*4)
23106 case 8: // result = lea base( , cond*8)
23107 case 9: // result = lea base(cond, cond*8)
23108 isFastMultiplier = true;
23113 if (isFastMultiplier) {
23114 APInt Diff = TrueC->getAPIntValue()-FalseC->getAPIntValue();
23115 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
23116 DAG.getConstant(CC, DL, MVT::i8), Cond);
23117 // Zero extend the condition if needed.
23118 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, FalseC->getValueType(0),
23120 // Scale the condition by the difference.
23122 Cond = DAG.getNode(ISD::MUL, DL, Cond.getValueType(), Cond,
23123 DAG.getConstant(Diff, DL, Cond.getValueType()));
23125 // Add the base if non-zero.
23126 if (FalseC->getAPIntValue() != 0)
23127 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
23128 SDValue(FalseC, 0));
23129 if (N->getNumValues() == 2) // Dead flag value?
23130 return DCI.CombineTo(N, Cond, SDValue());
23137 // Handle these cases:
23138 // (select (x != c), e, c) -> select (x != c), e, x),
23139 // (select (x == c), c, e) -> select (x == c), x, e)
23140 // where the c is an integer constant, and the "select" is the combination
23141 // of CMOV and CMP.
23143 // The rationale for this change is that the conditional-move from a constant
23144 // needs two instructions, however, conditional-move from a register needs
23145 // only one instruction.
23147 // CAVEAT: By replacing a constant with a symbolic value, it may obscure
23148 // some instruction-combining opportunities. This opt needs to be
23149 // postponed as late as possible.
23151 if (!DCI.isBeforeLegalize() && !DCI.isBeforeLegalizeOps()) {
23152 // the DCI.xxxx conditions are provided to postpone the optimization as
23153 // late as possible.
23155 ConstantSDNode *CmpAgainst = nullptr;
23156 if ((Cond.getOpcode() == X86ISD::CMP || Cond.getOpcode() == X86ISD::SUB) &&
23157 (CmpAgainst = dyn_cast<ConstantSDNode>(Cond.getOperand(1))) &&
23158 !isa<ConstantSDNode>(Cond.getOperand(0))) {
23160 if (CC == X86::COND_NE &&
23161 CmpAgainst == dyn_cast<ConstantSDNode>(FalseOp)) {
23162 CC = X86::GetOppositeBranchCondition(CC);
23163 std::swap(TrueOp, FalseOp);
23166 if (CC == X86::COND_E &&
23167 CmpAgainst == dyn_cast<ConstantSDNode>(TrueOp)) {
23168 SDValue Ops[] = { FalseOp, Cond.getOperand(0),
23169 DAG.getConstant(CC, DL, MVT::i8), Cond };
23170 return DAG.getNode(X86ISD::CMOV, DL, N->getVTList (), Ops);
23175 // Fold and/or of setcc's to double CMOV:
23176 // (CMOV F, T, ((cc1 | cc2) != 0)) -> (CMOV (CMOV F, T, cc1), T, cc2)
23177 // (CMOV F, T, ((cc1 & cc2) != 0)) -> (CMOV (CMOV T, F, !cc1), F, !cc2)
23179 // This combine lets us generate:
23180 // cmovcc1 (jcc1 if we don't have CMOV)
23186 // cmovne (jne if we don't have CMOV)
23187 // When we can't use the CMOV instruction, it might increase branch
23189 // When we can use CMOV, or when there is no mispredict, this improves
23190 // throughput and reduces register pressure.
23192 if (CC == X86::COND_NE) {
23194 X86::CondCode CC0, CC1;
23196 if (checkBoolTestAndOrSetCCCombine(Cond, CC0, CC1, Flags, isAndSetCC)) {
23198 std::swap(FalseOp, TrueOp);
23199 CC0 = X86::GetOppositeBranchCondition(CC0);
23200 CC1 = X86::GetOppositeBranchCondition(CC1);
23203 SDValue LOps[] = {FalseOp, TrueOp, DAG.getConstant(CC0, DL, MVT::i8),
23205 SDValue LCMOV = DAG.getNode(X86ISD::CMOV, DL, N->getVTList(), LOps);
23206 SDValue Ops[] = {LCMOV, TrueOp, DAG.getConstant(CC1, DL, MVT::i8), Flags};
23207 SDValue CMOV = DAG.getNode(X86ISD::CMOV, DL, N->getVTList(), Ops);
23208 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), SDValue(CMOV.getNode(), 1));
23216 static SDValue PerformINTRINSIC_WO_CHAINCombine(SDNode *N, SelectionDAG &DAG,
23217 const X86Subtarget *Subtarget) {
23218 unsigned IntNo = cast<ConstantSDNode>(N->getOperand(0))->getZExtValue();
23220 default: return SDValue();
23221 // SSE/AVX/AVX2 blend intrinsics.
23222 case Intrinsic::x86_avx2_pblendvb:
23223 // Don't try to simplify this intrinsic if we don't have AVX2.
23224 if (!Subtarget->hasAVX2())
23227 case Intrinsic::x86_avx_blendv_pd_256:
23228 case Intrinsic::x86_avx_blendv_ps_256:
23229 // Don't try to simplify this intrinsic if we don't have AVX.
23230 if (!Subtarget->hasAVX())
23233 case Intrinsic::x86_sse41_blendvps:
23234 case Intrinsic::x86_sse41_blendvpd:
23235 case Intrinsic::x86_sse41_pblendvb: {
23236 SDValue Op0 = N->getOperand(1);
23237 SDValue Op1 = N->getOperand(2);
23238 SDValue Mask = N->getOperand(3);
23240 // Don't try to simplify this intrinsic if we don't have SSE4.1.
23241 if (!Subtarget->hasSSE41())
23244 // fold (blend A, A, Mask) -> A
23247 // fold (blend A, B, allZeros) -> A
23248 if (ISD::isBuildVectorAllZeros(Mask.getNode()))
23250 // fold (blend A, B, allOnes) -> B
23251 if (ISD::isBuildVectorAllOnes(Mask.getNode()))
23254 // Simplify the case where the mask is a constant i32 value.
23255 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Mask)) {
23256 if (C->isNullValue())
23258 if (C->isAllOnesValue())
23265 // Packed SSE2/AVX2 arithmetic shift immediate intrinsics.
23266 case Intrinsic::x86_sse2_psrai_w:
23267 case Intrinsic::x86_sse2_psrai_d:
23268 case Intrinsic::x86_avx2_psrai_w:
23269 case Intrinsic::x86_avx2_psrai_d:
23270 case Intrinsic::x86_sse2_psra_w:
23271 case Intrinsic::x86_sse2_psra_d:
23272 case Intrinsic::x86_avx2_psra_w:
23273 case Intrinsic::x86_avx2_psra_d: {
23274 SDValue Op0 = N->getOperand(1);
23275 SDValue Op1 = N->getOperand(2);
23276 EVT VT = Op0.getValueType();
23277 assert(VT.isVector() && "Expected a vector type!");
23279 if (isa<BuildVectorSDNode>(Op1))
23280 Op1 = Op1.getOperand(0);
23282 if (!isa<ConstantSDNode>(Op1))
23285 EVT SVT = VT.getVectorElementType();
23286 unsigned SVTBits = SVT.getSizeInBits();
23288 ConstantSDNode *CND = cast<ConstantSDNode>(Op1);
23289 const APInt &C = APInt(SVTBits, CND->getAPIntValue().getZExtValue());
23290 uint64_t ShAmt = C.getZExtValue();
23292 // Don't try to convert this shift into a ISD::SRA if the shift
23293 // count is bigger than or equal to the element size.
23294 if (ShAmt >= SVTBits)
23297 // Trivial case: if the shift count is zero, then fold this
23298 // into the first operand.
23302 // Replace this packed shift intrinsic with a target independent
23305 SDValue Splat = DAG.getConstant(C, DL, VT);
23306 return DAG.getNode(ISD::SRA, DL, VT, Op0, Splat);
23311 /// PerformMulCombine - Optimize a single multiply with constant into two
23312 /// in order to implement it with two cheaper instructions, e.g.
23313 /// LEA + SHL, LEA + LEA.
23314 static SDValue PerformMulCombine(SDNode *N, SelectionDAG &DAG,
23315 TargetLowering::DAGCombinerInfo &DCI) {
23316 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
23319 EVT VT = N->getValueType(0);
23320 if (VT != MVT::i64 && VT != MVT::i32)
23323 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1));
23326 uint64_t MulAmt = C->getZExtValue();
23327 if (isPowerOf2_64(MulAmt) || MulAmt == 3 || MulAmt == 5 || MulAmt == 9)
23330 uint64_t MulAmt1 = 0;
23331 uint64_t MulAmt2 = 0;
23332 if ((MulAmt % 9) == 0) {
23334 MulAmt2 = MulAmt / 9;
23335 } else if ((MulAmt % 5) == 0) {
23337 MulAmt2 = MulAmt / 5;
23338 } else if ((MulAmt % 3) == 0) {
23340 MulAmt2 = MulAmt / 3;
23343 (isPowerOf2_64(MulAmt2) || MulAmt2 == 3 || MulAmt2 == 5 || MulAmt2 == 9)){
23346 if (isPowerOf2_64(MulAmt2) &&
23347 !(N->hasOneUse() && N->use_begin()->getOpcode() == ISD::ADD))
23348 // If second multiplifer is pow2, issue it first. We want the multiply by
23349 // 3, 5, or 9 to be folded into the addressing mode unless the lone use
23351 std::swap(MulAmt1, MulAmt2);
23354 if (isPowerOf2_64(MulAmt1))
23355 NewMul = DAG.getNode(ISD::SHL, DL, VT, N->getOperand(0),
23356 DAG.getConstant(Log2_64(MulAmt1), DL, MVT::i8));
23358 NewMul = DAG.getNode(X86ISD::MUL_IMM, DL, VT, N->getOperand(0),
23359 DAG.getConstant(MulAmt1, DL, VT));
23361 if (isPowerOf2_64(MulAmt2))
23362 NewMul = DAG.getNode(ISD::SHL, DL, VT, NewMul,
23363 DAG.getConstant(Log2_64(MulAmt2), DL, MVT::i8));
23365 NewMul = DAG.getNode(X86ISD::MUL_IMM, DL, VT, NewMul,
23366 DAG.getConstant(MulAmt2, DL, VT));
23368 // Do not add new nodes to DAG combiner worklist.
23369 DCI.CombineTo(N, NewMul, false);
23374 static SDValue PerformSHLCombine(SDNode *N, SelectionDAG &DAG) {
23375 SDValue N0 = N->getOperand(0);
23376 SDValue N1 = N->getOperand(1);
23377 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
23378 EVT VT = N0.getValueType();
23380 // fold (shl (and (setcc_c), c1), c2) -> (and setcc_c, (c1 << c2))
23381 // since the result of setcc_c is all zero's or all ones.
23382 if (VT.isInteger() && !VT.isVector() &&
23383 N1C && N0.getOpcode() == ISD::AND &&
23384 N0.getOperand(1).getOpcode() == ISD::Constant) {
23385 SDValue N00 = N0.getOperand(0);
23386 if (N00.getOpcode() == X86ISD::SETCC_CARRY ||
23387 ((N00.getOpcode() == ISD::ANY_EXTEND ||
23388 N00.getOpcode() == ISD::ZERO_EXTEND) &&
23389 N00.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY)) {
23390 APInt Mask = cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
23391 APInt ShAmt = N1C->getAPIntValue();
23392 Mask = Mask.shl(ShAmt);
23395 return DAG.getNode(ISD::AND, DL, VT,
23396 N00, DAG.getConstant(Mask, DL, VT));
23401 // Hardware support for vector shifts is sparse which makes us scalarize the
23402 // vector operations in many cases. Also, on sandybridge ADD is faster than
23404 // (shl V, 1) -> add V,V
23405 if (auto *N1BV = dyn_cast<BuildVectorSDNode>(N1))
23406 if (auto *N1SplatC = N1BV->getConstantSplatNode()) {
23407 assert(N0.getValueType().isVector() && "Invalid vector shift type");
23408 // We shift all of the values by one. In many cases we do not have
23409 // hardware support for this operation. This is better expressed as an ADD
23411 if (N1SplatC->getAPIntValue() == 1)
23412 return DAG.getNode(ISD::ADD, SDLoc(N), VT, N0, N0);
23418 /// \brief Returns a vector of 0s if the node in input is a vector logical
23419 /// shift by a constant amount which is known to be bigger than or equal
23420 /// to the vector element size in bits.
23421 static SDValue performShiftToAllZeros(SDNode *N, SelectionDAG &DAG,
23422 const X86Subtarget *Subtarget) {
23423 EVT VT = N->getValueType(0);
23425 if (VT != MVT::v2i64 && VT != MVT::v4i32 && VT != MVT::v8i16 &&
23426 (!Subtarget->hasInt256() ||
23427 (VT != MVT::v4i64 && VT != MVT::v8i32 && VT != MVT::v16i16)))
23430 SDValue Amt = N->getOperand(1);
23432 if (auto *AmtBV = dyn_cast<BuildVectorSDNode>(Amt))
23433 if (auto *AmtSplat = AmtBV->getConstantSplatNode()) {
23434 APInt ShiftAmt = AmtSplat->getAPIntValue();
23435 unsigned MaxAmount = VT.getVectorElementType().getSizeInBits();
23437 // SSE2/AVX2 logical shifts always return a vector of 0s
23438 // if the shift amount is bigger than or equal to
23439 // the element size. The constant shift amount will be
23440 // encoded as a 8-bit immediate.
23441 if (ShiftAmt.trunc(8).uge(MaxAmount))
23442 return getZeroVector(VT, Subtarget, DAG, DL);
23448 /// PerformShiftCombine - Combine shifts.
23449 static SDValue PerformShiftCombine(SDNode* N, SelectionDAG &DAG,
23450 TargetLowering::DAGCombinerInfo &DCI,
23451 const X86Subtarget *Subtarget) {
23452 if (N->getOpcode() == ISD::SHL)
23453 if (SDValue V = PerformSHLCombine(N, DAG))
23456 // Try to fold this logical shift into a zero vector.
23457 if (N->getOpcode() != ISD::SRA)
23458 if (SDValue V = performShiftToAllZeros(N, DAG, Subtarget))
23464 // CMPEQCombine - Recognize the distinctive (AND (setcc ...) (setcc ..))
23465 // where both setccs reference the same FP CMP, and rewrite for CMPEQSS
23466 // and friends. Likewise for OR -> CMPNEQSS.
23467 static SDValue CMPEQCombine(SDNode *N, SelectionDAG &DAG,
23468 TargetLowering::DAGCombinerInfo &DCI,
23469 const X86Subtarget *Subtarget) {
23472 // SSE1 supports CMP{eq|ne}SS, and SSE2 added CMP{eq|ne}SD, but
23473 // we're requiring SSE2 for both.
23474 if (Subtarget->hasSSE2() && isAndOrOfSetCCs(SDValue(N, 0U), opcode)) {
23475 SDValue N0 = N->getOperand(0);
23476 SDValue N1 = N->getOperand(1);
23477 SDValue CMP0 = N0->getOperand(1);
23478 SDValue CMP1 = N1->getOperand(1);
23481 // The SETCCs should both refer to the same CMP.
23482 if (CMP0.getOpcode() != X86ISD::CMP || CMP0 != CMP1)
23485 SDValue CMP00 = CMP0->getOperand(0);
23486 SDValue CMP01 = CMP0->getOperand(1);
23487 EVT VT = CMP00.getValueType();
23489 if (VT == MVT::f32 || VT == MVT::f64) {
23490 bool ExpectingFlags = false;
23491 // Check for any users that want flags:
23492 for (SDNode::use_iterator UI = N->use_begin(), UE = N->use_end();
23493 !ExpectingFlags && UI != UE; ++UI)
23494 switch (UI->getOpcode()) {
23499 ExpectingFlags = true;
23501 case ISD::CopyToReg:
23502 case ISD::SIGN_EXTEND:
23503 case ISD::ZERO_EXTEND:
23504 case ISD::ANY_EXTEND:
23508 if (!ExpectingFlags) {
23509 enum X86::CondCode cc0 = (enum X86::CondCode)N0.getConstantOperandVal(0);
23510 enum X86::CondCode cc1 = (enum X86::CondCode)N1.getConstantOperandVal(0);
23512 if (cc1 == X86::COND_E || cc1 == X86::COND_NE) {
23513 X86::CondCode tmp = cc0;
23518 if ((cc0 == X86::COND_E && cc1 == X86::COND_NP) ||
23519 (cc0 == X86::COND_NE && cc1 == X86::COND_P)) {
23520 // FIXME: need symbolic constants for these magic numbers.
23521 // See X86ATTInstPrinter.cpp:printSSECC().
23522 unsigned x86cc = (cc0 == X86::COND_E) ? 0 : 4;
23523 if (Subtarget->hasAVX512()) {
23524 SDValue FSetCC = DAG.getNode(X86ISD::FSETCC, DL, MVT::i1, CMP00,
23526 DAG.getConstant(x86cc, DL, MVT::i8));
23527 if (N->getValueType(0) != MVT::i1)
23528 return DAG.getNode(ISD::ZERO_EXTEND, DL, N->getValueType(0),
23532 SDValue OnesOrZeroesF = DAG.getNode(X86ISD::FSETCC, DL,
23533 CMP00.getValueType(), CMP00, CMP01,
23534 DAG.getConstant(x86cc, DL,
23537 bool is64BitFP = (CMP00.getValueType() == MVT::f64);
23538 MVT IntVT = is64BitFP ? MVT::i64 : MVT::i32;
23540 if (is64BitFP && !Subtarget->is64Bit()) {
23541 // On a 32-bit target, we cannot bitcast the 64-bit float to a
23542 // 64-bit integer, since that's not a legal type. Since
23543 // OnesOrZeroesF is all ones of all zeroes, we don't need all the
23544 // bits, but can do this little dance to extract the lowest 32 bits
23545 // and work with those going forward.
23546 SDValue Vector64 = DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, MVT::v2f64,
23548 SDValue Vector32 = DAG.getBitcast(MVT::v4f32, Vector64);
23549 OnesOrZeroesF = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::f32,
23550 Vector32, DAG.getIntPtrConstant(0, DL));
23554 SDValue OnesOrZeroesI = DAG.getBitcast(IntVT, OnesOrZeroesF);
23555 SDValue ANDed = DAG.getNode(ISD::AND, DL, IntVT, OnesOrZeroesI,
23556 DAG.getConstant(1, DL, IntVT));
23557 SDValue OneBitOfTruth = DAG.getNode(ISD::TRUNCATE, DL, MVT::i8,
23559 return OneBitOfTruth;
23567 /// CanFoldXORWithAllOnes - Test whether the XOR operand is a AllOnes vector
23568 /// so it can be folded inside ANDNP.
23569 static bool CanFoldXORWithAllOnes(const SDNode *N) {
23570 EVT VT = N->getValueType(0);
23572 // Match direct AllOnes for 128 and 256-bit vectors
23573 if (ISD::isBuildVectorAllOnes(N))
23576 // Look through a bit convert.
23577 if (N->getOpcode() == ISD::BITCAST)
23578 N = N->getOperand(0).getNode();
23580 // Sometimes the operand may come from a insert_subvector building a 256-bit
23582 if (VT.is256BitVector() &&
23583 N->getOpcode() == ISD::INSERT_SUBVECTOR) {
23584 SDValue V1 = N->getOperand(0);
23585 SDValue V2 = N->getOperand(1);
23587 if (V1.getOpcode() == ISD::INSERT_SUBVECTOR &&
23588 V1.getOperand(0).getOpcode() == ISD::UNDEF &&
23589 ISD::isBuildVectorAllOnes(V1.getOperand(1).getNode()) &&
23590 ISD::isBuildVectorAllOnes(V2.getNode()))
23597 // On AVX/AVX2 the type v8i1 is legalized to v8i16, which is an XMM sized
23598 // register. In most cases we actually compare or select YMM-sized registers
23599 // and mixing the two types creates horrible code. This method optimizes
23600 // some of the transition sequences.
23601 static SDValue WidenMaskArithmetic(SDNode *N, SelectionDAG &DAG,
23602 TargetLowering::DAGCombinerInfo &DCI,
23603 const X86Subtarget *Subtarget) {
23604 EVT VT = N->getValueType(0);
23605 if (!VT.is256BitVector())
23608 assert((N->getOpcode() == ISD::ANY_EXTEND ||
23609 N->getOpcode() == ISD::ZERO_EXTEND ||
23610 N->getOpcode() == ISD::SIGN_EXTEND) && "Invalid Node");
23612 SDValue Narrow = N->getOperand(0);
23613 EVT NarrowVT = Narrow->getValueType(0);
23614 if (!NarrowVT.is128BitVector())
23617 if (Narrow->getOpcode() != ISD::XOR &&
23618 Narrow->getOpcode() != ISD::AND &&
23619 Narrow->getOpcode() != ISD::OR)
23622 SDValue N0 = Narrow->getOperand(0);
23623 SDValue N1 = Narrow->getOperand(1);
23626 // The Left side has to be a trunc.
23627 if (N0.getOpcode() != ISD::TRUNCATE)
23630 // The type of the truncated inputs.
23631 EVT WideVT = N0->getOperand(0)->getValueType(0);
23635 // The right side has to be a 'trunc' or a constant vector.
23636 bool RHSTrunc = N1.getOpcode() == ISD::TRUNCATE;
23637 ConstantSDNode *RHSConstSplat = nullptr;
23638 if (auto *RHSBV = dyn_cast<BuildVectorSDNode>(N1))
23639 RHSConstSplat = RHSBV->getConstantSplatNode();
23640 if (!RHSTrunc && !RHSConstSplat)
23643 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
23645 if (!TLI.isOperationLegalOrPromote(Narrow->getOpcode(), WideVT))
23648 // Set N0 and N1 to hold the inputs to the new wide operation.
23649 N0 = N0->getOperand(0);
23650 if (RHSConstSplat) {
23651 N1 = DAG.getNode(ISD::ZERO_EXTEND, DL, WideVT.getScalarType(),
23652 SDValue(RHSConstSplat, 0));
23653 SmallVector<SDValue, 8> C(WideVT.getVectorNumElements(), N1);
23654 N1 = DAG.getNode(ISD::BUILD_VECTOR, DL, WideVT, C);
23655 } else if (RHSTrunc) {
23656 N1 = N1->getOperand(0);
23659 // Generate the wide operation.
23660 SDValue Op = DAG.getNode(Narrow->getOpcode(), DL, WideVT, N0, N1);
23661 unsigned Opcode = N->getOpcode();
23663 case ISD::ANY_EXTEND:
23665 case ISD::ZERO_EXTEND: {
23666 unsigned InBits = NarrowVT.getScalarType().getSizeInBits();
23667 APInt Mask = APInt::getAllOnesValue(InBits);
23668 Mask = Mask.zext(VT.getScalarType().getSizeInBits());
23669 return DAG.getNode(ISD::AND, DL, VT,
23670 Op, DAG.getConstant(Mask, DL, VT));
23672 case ISD::SIGN_EXTEND:
23673 return DAG.getNode(ISD::SIGN_EXTEND_INREG, DL, VT,
23674 Op, DAG.getValueType(NarrowVT));
23676 llvm_unreachable("Unexpected opcode");
23680 static SDValue VectorZextCombine(SDNode *N, SelectionDAG &DAG,
23681 TargetLowering::DAGCombinerInfo &DCI,
23682 const X86Subtarget *Subtarget) {
23683 SDValue N0 = N->getOperand(0);
23684 SDValue N1 = N->getOperand(1);
23687 // A vector zext_in_reg may be represented as a shuffle,
23688 // feeding into a bitcast (this represents anyext) feeding into
23689 // an and with a mask.
23690 // We'd like to try to combine that into a shuffle with zero
23691 // plus a bitcast, removing the and.
23692 if (N0.getOpcode() != ISD::BITCAST ||
23693 N0.getOperand(0).getOpcode() != ISD::VECTOR_SHUFFLE)
23696 // The other side of the AND should be a splat of 2^C, where C
23697 // is the number of bits in the source type.
23698 if (N1.getOpcode() == ISD::BITCAST)
23699 N1 = N1.getOperand(0);
23700 if (N1.getOpcode() != ISD::BUILD_VECTOR)
23702 BuildVectorSDNode *Vector = cast<BuildVectorSDNode>(N1);
23704 ShuffleVectorSDNode *Shuffle = cast<ShuffleVectorSDNode>(N0.getOperand(0));
23705 EVT SrcType = Shuffle->getValueType(0);
23707 // We expect a single-source shuffle
23708 if (Shuffle->getOperand(1)->getOpcode() != ISD::UNDEF)
23711 unsigned SrcSize = SrcType.getScalarSizeInBits();
23713 APInt SplatValue, SplatUndef;
23714 unsigned SplatBitSize;
23716 if (!Vector->isConstantSplat(SplatValue, SplatUndef,
23717 SplatBitSize, HasAnyUndefs))
23720 unsigned ResSize = N1.getValueType().getScalarSizeInBits();
23721 // Make sure the splat matches the mask we expect
23722 if (SplatBitSize > ResSize ||
23723 (SplatValue + 1).exactLogBase2() != (int)SrcSize)
23726 // Make sure the input and output size make sense
23727 if (SrcSize >= ResSize || ResSize % SrcSize)
23730 // We expect a shuffle of the form <0, u, u, u, 1, u, u, u...>
23731 // The number of u's between each two values depends on the ratio between
23732 // the source and dest type.
23733 unsigned ZextRatio = ResSize / SrcSize;
23734 bool IsZext = true;
23735 for (unsigned i = 0; i < SrcType.getVectorNumElements(); ++i) {
23736 if (i % ZextRatio) {
23737 if (Shuffle->getMaskElt(i) > 0) {
23743 if (Shuffle->getMaskElt(i) != (int)(i / ZextRatio)) {
23744 // Expected element number
23754 // Ok, perform the transformation - replace the shuffle with
23755 // a shuffle of the form <0, k, k, k, 1, k, k, k> with zero
23756 // (instead of undef) where the k elements come from the zero vector.
23757 SmallVector<int, 8> Mask;
23758 unsigned NumElems = SrcType.getVectorNumElements();
23759 for (unsigned i = 0; i < NumElems; ++i)
23761 Mask.push_back(NumElems);
23763 Mask.push_back(i / ZextRatio);
23765 SDValue NewShuffle = DAG.getVectorShuffle(Shuffle->getValueType(0), DL,
23766 Shuffle->getOperand(0), DAG.getConstant(0, DL, SrcType), Mask);
23767 return DAG.getBitcast(N0.getValueType(), NewShuffle);
23770 static SDValue PerformAndCombine(SDNode *N, SelectionDAG &DAG,
23771 TargetLowering::DAGCombinerInfo &DCI,
23772 const X86Subtarget *Subtarget) {
23773 if (DCI.isBeforeLegalizeOps())
23776 if (SDValue Zext = VectorZextCombine(N, DAG, DCI, Subtarget))
23779 if (SDValue R = CMPEQCombine(N, DAG, DCI, Subtarget))
23782 EVT VT = N->getValueType(0);
23783 SDValue N0 = N->getOperand(0);
23784 SDValue N1 = N->getOperand(1);
23787 // Create BEXTR instructions
23788 // BEXTR is ((X >> imm) & (2**size-1))
23789 if (VT == MVT::i32 || VT == MVT::i64) {
23790 // Check for BEXTR.
23791 if ((Subtarget->hasBMI() || Subtarget->hasTBM()) &&
23792 (N0.getOpcode() == ISD::SRA || N0.getOpcode() == ISD::SRL)) {
23793 ConstantSDNode *MaskNode = dyn_cast<ConstantSDNode>(N1);
23794 ConstantSDNode *ShiftNode = dyn_cast<ConstantSDNode>(N0.getOperand(1));
23795 if (MaskNode && ShiftNode) {
23796 uint64_t Mask = MaskNode->getZExtValue();
23797 uint64_t Shift = ShiftNode->getZExtValue();
23798 if (isMask_64(Mask)) {
23799 uint64_t MaskSize = countPopulation(Mask);
23800 if (Shift + MaskSize <= VT.getSizeInBits())
23801 return DAG.getNode(X86ISD::BEXTR, DL, VT, N0.getOperand(0),
23802 DAG.getConstant(Shift | (MaskSize << 8), DL,
23811 // Want to form ANDNP nodes:
23812 // 1) In the hopes of then easily combining them with OR and AND nodes
23813 // to form PBLEND/PSIGN.
23814 // 2) To match ANDN packed intrinsics
23815 if (VT != MVT::v2i64 && VT != MVT::v4i64)
23818 // Check LHS for vnot
23819 if (N0.getOpcode() == ISD::XOR &&
23820 //ISD::isBuildVectorAllOnes(N0.getOperand(1).getNode()))
23821 CanFoldXORWithAllOnes(N0.getOperand(1).getNode()))
23822 return DAG.getNode(X86ISD::ANDNP, DL, VT, N0.getOperand(0), N1);
23824 // Check RHS for vnot
23825 if (N1.getOpcode() == ISD::XOR &&
23826 //ISD::isBuildVectorAllOnes(N1.getOperand(1).getNode()))
23827 CanFoldXORWithAllOnes(N1.getOperand(1).getNode()))
23828 return DAG.getNode(X86ISD::ANDNP, DL, VT, N1.getOperand(0), N0);
23833 static SDValue PerformOrCombine(SDNode *N, SelectionDAG &DAG,
23834 TargetLowering::DAGCombinerInfo &DCI,
23835 const X86Subtarget *Subtarget) {
23836 if (DCI.isBeforeLegalizeOps())
23839 if (SDValue R = CMPEQCombine(N, DAG, DCI, Subtarget))
23842 SDValue N0 = N->getOperand(0);
23843 SDValue N1 = N->getOperand(1);
23844 EVT VT = N->getValueType(0);
23846 // look for psign/blend
23847 if (VT == MVT::v2i64 || VT == MVT::v4i64) {
23848 if (!Subtarget->hasSSSE3() ||
23849 (VT == MVT::v4i64 && !Subtarget->hasInt256()))
23852 // Canonicalize pandn to RHS
23853 if (N0.getOpcode() == X86ISD::ANDNP)
23855 // or (and (m, y), (pandn m, x))
23856 if (N0.getOpcode() == ISD::AND && N1.getOpcode() == X86ISD::ANDNP) {
23857 SDValue Mask = N1.getOperand(0);
23858 SDValue X = N1.getOperand(1);
23860 if (N0.getOperand(0) == Mask)
23861 Y = N0.getOperand(1);
23862 if (N0.getOperand(1) == Mask)
23863 Y = N0.getOperand(0);
23865 // Check to see if the mask appeared in both the AND and ANDNP and
23869 // Validate that X, Y, and Mask are BIT_CONVERTS, and see through them.
23870 // Look through mask bitcast.
23871 if (Mask.getOpcode() == ISD::BITCAST)
23872 Mask = Mask.getOperand(0);
23873 if (X.getOpcode() == ISD::BITCAST)
23874 X = X.getOperand(0);
23875 if (Y.getOpcode() == ISD::BITCAST)
23876 Y = Y.getOperand(0);
23878 EVT MaskVT = Mask.getValueType();
23880 // Validate that the Mask operand is a vector sra node.
23881 // FIXME: what to do for bytes, since there is a psignb/pblendvb, but
23882 // there is no psrai.b
23883 unsigned EltBits = MaskVT.getVectorElementType().getSizeInBits();
23884 unsigned SraAmt = ~0;
23885 if (Mask.getOpcode() == ISD::SRA) {
23886 if (auto *AmtBV = dyn_cast<BuildVectorSDNode>(Mask.getOperand(1)))
23887 if (auto *AmtConst = AmtBV->getConstantSplatNode())
23888 SraAmt = AmtConst->getZExtValue();
23889 } else if (Mask.getOpcode() == X86ISD::VSRAI) {
23890 SDValue SraC = Mask.getOperand(1);
23891 SraAmt = cast<ConstantSDNode>(SraC)->getZExtValue();
23893 if ((SraAmt + 1) != EltBits)
23898 // Now we know we at least have a plendvb with the mask val. See if
23899 // we can form a psignb/w/d.
23900 // psign = x.type == y.type == mask.type && y = sub(0, x);
23901 if (Y.getOpcode() == ISD::SUB && Y.getOperand(1) == X &&
23902 ISD::isBuildVectorAllZeros(Y.getOperand(0).getNode()) &&
23903 X.getValueType() == MaskVT && Y.getValueType() == MaskVT) {
23904 assert((EltBits == 8 || EltBits == 16 || EltBits == 32) &&
23905 "Unsupported VT for PSIGN");
23906 Mask = DAG.getNode(X86ISD::PSIGN, DL, MaskVT, X, Mask.getOperand(0));
23907 return DAG.getBitcast(VT, Mask);
23909 // PBLENDVB only available on SSE 4.1
23910 if (!Subtarget->hasSSE41())
23913 EVT BlendVT = (VT == MVT::v4i64) ? MVT::v32i8 : MVT::v16i8;
23915 X = DAG.getBitcast(BlendVT, X);
23916 Y = DAG.getBitcast(BlendVT, Y);
23917 Mask = DAG.getBitcast(BlendVT, Mask);
23918 Mask = DAG.getNode(ISD::VSELECT, DL, BlendVT, Mask, Y, X);
23919 return DAG.getBitcast(VT, Mask);
23923 if (VT != MVT::i16 && VT != MVT::i32 && VT != MVT::i64)
23926 // fold (or (x << c) | (y >> (64 - c))) ==> (shld64 x, y, c)
23927 MachineFunction &MF = DAG.getMachineFunction();
23929 MF.getFunction()->hasFnAttribute(Attribute::OptimizeForSize);
23931 // SHLD/SHRD instructions have lower register pressure, but on some
23932 // platforms they have higher latency than the equivalent
23933 // series of shifts/or that would otherwise be generated.
23934 // Don't fold (or (x << c) | (y >> (64 - c))) if SHLD/SHRD instructions
23935 // have higher latencies and we are not optimizing for size.
23936 if (!OptForSize && Subtarget->isSHLDSlow())
23939 if (N0.getOpcode() == ISD::SRL && N1.getOpcode() == ISD::SHL)
23941 if (N0.getOpcode() != ISD::SHL || N1.getOpcode() != ISD::SRL)
23943 if (!N0.hasOneUse() || !N1.hasOneUse())
23946 SDValue ShAmt0 = N0.getOperand(1);
23947 if (ShAmt0.getValueType() != MVT::i8)
23949 SDValue ShAmt1 = N1.getOperand(1);
23950 if (ShAmt1.getValueType() != MVT::i8)
23952 if (ShAmt0.getOpcode() == ISD::TRUNCATE)
23953 ShAmt0 = ShAmt0.getOperand(0);
23954 if (ShAmt1.getOpcode() == ISD::TRUNCATE)
23955 ShAmt1 = ShAmt1.getOperand(0);
23958 unsigned Opc = X86ISD::SHLD;
23959 SDValue Op0 = N0.getOperand(0);
23960 SDValue Op1 = N1.getOperand(0);
23961 if (ShAmt0.getOpcode() == ISD::SUB) {
23962 Opc = X86ISD::SHRD;
23963 std::swap(Op0, Op1);
23964 std::swap(ShAmt0, ShAmt1);
23967 unsigned Bits = VT.getSizeInBits();
23968 if (ShAmt1.getOpcode() == ISD::SUB) {
23969 SDValue Sum = ShAmt1.getOperand(0);
23970 if (ConstantSDNode *SumC = dyn_cast<ConstantSDNode>(Sum)) {
23971 SDValue ShAmt1Op1 = ShAmt1.getOperand(1);
23972 if (ShAmt1Op1.getNode()->getOpcode() == ISD::TRUNCATE)
23973 ShAmt1Op1 = ShAmt1Op1.getOperand(0);
23974 if (SumC->getSExtValue() == Bits && ShAmt1Op1 == ShAmt0)
23975 return DAG.getNode(Opc, DL, VT,
23977 DAG.getNode(ISD::TRUNCATE, DL,
23980 } else if (ConstantSDNode *ShAmt1C = dyn_cast<ConstantSDNode>(ShAmt1)) {
23981 ConstantSDNode *ShAmt0C = dyn_cast<ConstantSDNode>(ShAmt0);
23983 ShAmt0C->getSExtValue() + ShAmt1C->getSExtValue() == Bits)
23984 return DAG.getNode(Opc, DL, VT,
23985 N0.getOperand(0), N1.getOperand(0),
23986 DAG.getNode(ISD::TRUNCATE, DL,
23993 // Generate NEG and CMOV for integer abs.
23994 static SDValue performIntegerAbsCombine(SDNode *N, SelectionDAG &DAG) {
23995 EVT VT = N->getValueType(0);
23997 // Since X86 does not have CMOV for 8-bit integer, we don't convert
23998 // 8-bit integer abs to NEG and CMOV.
23999 if (VT.isInteger() && VT.getSizeInBits() == 8)
24002 SDValue N0 = N->getOperand(0);
24003 SDValue N1 = N->getOperand(1);
24006 // Check pattern of XOR(ADD(X,Y), Y) where Y is SRA(X, size(X)-1)
24007 // and change it to SUB and CMOV.
24008 if (VT.isInteger() && N->getOpcode() == ISD::XOR &&
24009 N0.getOpcode() == ISD::ADD &&
24010 N0.getOperand(1) == N1 &&
24011 N1.getOpcode() == ISD::SRA &&
24012 N1.getOperand(0) == N0.getOperand(0))
24013 if (ConstantSDNode *Y1C = dyn_cast<ConstantSDNode>(N1.getOperand(1)))
24014 if (Y1C->getAPIntValue() == VT.getSizeInBits()-1) {
24015 // Generate SUB & CMOV.
24016 SDValue Neg = DAG.getNode(X86ISD::SUB, DL, DAG.getVTList(VT, MVT::i32),
24017 DAG.getConstant(0, DL, VT), N0.getOperand(0));
24019 SDValue Ops[] = { N0.getOperand(0), Neg,
24020 DAG.getConstant(X86::COND_GE, DL, MVT::i8),
24021 SDValue(Neg.getNode(), 1) };
24022 return DAG.getNode(X86ISD::CMOV, DL, DAG.getVTList(VT, MVT::Glue), Ops);
24027 // PerformXorCombine - Attempts to turn XOR nodes into BLSMSK nodes
24028 static SDValue PerformXorCombine(SDNode *N, SelectionDAG &DAG,
24029 TargetLowering::DAGCombinerInfo &DCI,
24030 const X86Subtarget *Subtarget) {
24031 if (DCI.isBeforeLegalizeOps())
24034 if (Subtarget->hasCMov())
24035 if (SDValue RV = performIntegerAbsCombine(N, DAG))
24041 /// PerformLOADCombine - Do target-specific dag combines on LOAD nodes.
24042 static SDValue PerformLOADCombine(SDNode *N, SelectionDAG &DAG,
24043 TargetLowering::DAGCombinerInfo &DCI,
24044 const X86Subtarget *Subtarget) {
24045 LoadSDNode *Ld = cast<LoadSDNode>(N);
24046 EVT RegVT = Ld->getValueType(0);
24047 EVT MemVT = Ld->getMemoryVT();
24049 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
24051 // For chips with slow 32-byte unaligned loads, break the 32-byte operation
24052 // into two 16-byte operations.
24053 ISD::LoadExtType Ext = Ld->getExtensionType();
24054 unsigned Alignment = Ld->getAlignment();
24055 bool IsAligned = Alignment == 0 || Alignment >= MemVT.getSizeInBits()/8;
24056 if (RegVT.is256BitVector() && Subtarget->isUnalignedMem32Slow() &&
24057 !DCI.isBeforeLegalizeOps() && !IsAligned && Ext == ISD::NON_EXTLOAD) {
24058 unsigned NumElems = RegVT.getVectorNumElements();
24062 SDValue Ptr = Ld->getBasePtr();
24063 SDValue Increment =
24064 DAG.getConstant(16, dl, TLI.getPointerTy(DAG.getDataLayout()));
24066 EVT HalfVT = EVT::getVectorVT(*DAG.getContext(), MemVT.getScalarType(),
24068 SDValue Load1 = DAG.getLoad(HalfVT, dl, Ld->getChain(), Ptr,
24069 Ld->getPointerInfo(), Ld->isVolatile(),
24070 Ld->isNonTemporal(), Ld->isInvariant(),
24072 Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr, Increment);
24073 SDValue Load2 = DAG.getLoad(HalfVT, dl, Ld->getChain(), Ptr,
24074 Ld->getPointerInfo(), Ld->isVolatile(),
24075 Ld->isNonTemporal(), Ld->isInvariant(),
24076 std::min(16U, Alignment));
24077 SDValue TF = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
24079 Load2.getValue(1));
24081 SDValue NewVec = DAG.getUNDEF(RegVT);
24082 NewVec = Insert128BitVector(NewVec, Load1, 0, DAG, dl);
24083 NewVec = Insert128BitVector(NewVec, Load2, NumElems/2, DAG, dl);
24084 return DCI.CombineTo(N, NewVec, TF, true);
24090 /// PerformMLOADCombine - Resolve extending loads
24091 static SDValue PerformMLOADCombine(SDNode *N, SelectionDAG &DAG,
24092 TargetLowering::DAGCombinerInfo &DCI,
24093 const X86Subtarget *Subtarget) {
24094 MaskedLoadSDNode *Mld = cast<MaskedLoadSDNode>(N);
24095 if (Mld->getExtensionType() != ISD::SEXTLOAD)
24098 EVT VT = Mld->getValueType(0);
24099 unsigned NumElems = VT.getVectorNumElements();
24100 EVT LdVT = Mld->getMemoryVT();
24103 assert(LdVT != VT && "Cannot extend to the same type");
24104 unsigned ToSz = VT.getVectorElementType().getSizeInBits();
24105 unsigned FromSz = LdVT.getVectorElementType().getSizeInBits();
24106 // From, To sizes and ElemCount must be pow of two
24107 assert (isPowerOf2_32(NumElems * FromSz * ToSz) &&
24108 "Unexpected size for extending masked load");
24110 unsigned SizeRatio = ToSz / FromSz;
24111 assert(SizeRatio * NumElems * FromSz == VT.getSizeInBits());
24113 // Create a type on which we perform the shuffle
24114 EVT WideVecVT = EVT::getVectorVT(*DAG.getContext(),
24115 LdVT.getScalarType(), NumElems*SizeRatio);
24116 assert(WideVecVT.getSizeInBits() == VT.getSizeInBits());
24118 // Convert Src0 value
24119 SDValue WideSrc0 = DAG.getBitcast(WideVecVT, Mld->getSrc0());
24120 if (Mld->getSrc0().getOpcode() != ISD::UNDEF) {
24121 SmallVector<int, 16> ShuffleVec(NumElems * SizeRatio, -1);
24122 for (unsigned i = 0; i != NumElems; ++i)
24123 ShuffleVec[i] = i * SizeRatio;
24125 // Can't shuffle using an illegal type.
24126 assert (DAG.getTargetLoweringInfo().isTypeLegal(WideVecVT)
24127 && "WideVecVT should be legal");
24128 WideSrc0 = DAG.getVectorShuffle(WideVecVT, dl, WideSrc0,
24129 DAG.getUNDEF(WideVecVT), &ShuffleVec[0]);
24131 // Prepare the new mask
24133 SDValue Mask = Mld->getMask();
24134 if (Mask.getValueType() == VT) {
24135 // Mask and original value have the same type
24136 NewMask = DAG.getBitcast(WideVecVT, Mask);
24137 SmallVector<int, 16> ShuffleVec(NumElems * SizeRatio, -1);
24138 for (unsigned i = 0; i != NumElems; ++i)
24139 ShuffleVec[i] = i * SizeRatio;
24140 for (unsigned i = NumElems; i != NumElems*SizeRatio; ++i)
24141 ShuffleVec[i] = NumElems*SizeRatio;
24142 NewMask = DAG.getVectorShuffle(WideVecVT, dl, NewMask,
24143 DAG.getConstant(0, dl, WideVecVT),
24147 assert(Mask.getValueType().getVectorElementType() == MVT::i1);
24148 unsigned WidenNumElts = NumElems*SizeRatio;
24149 unsigned MaskNumElts = VT.getVectorNumElements();
24150 EVT NewMaskVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
24153 unsigned NumConcat = WidenNumElts / MaskNumElts;
24154 SmallVector<SDValue, 16> Ops(NumConcat);
24155 SDValue ZeroVal = DAG.getConstant(0, dl, Mask.getValueType());
24157 for (unsigned i = 1; i != NumConcat; ++i)
24160 NewMask = DAG.getNode(ISD::CONCAT_VECTORS, dl, NewMaskVT, Ops);
24163 SDValue WideLd = DAG.getMaskedLoad(WideVecVT, dl, Mld->getChain(),
24164 Mld->getBasePtr(), NewMask, WideSrc0,
24165 Mld->getMemoryVT(), Mld->getMemOperand(),
24167 SDValue NewVec = DAG.getNode(X86ISD::VSEXT, dl, VT, WideLd);
24168 return DCI.CombineTo(N, NewVec, WideLd.getValue(1), true);
24171 /// PerformMSTORECombine - Resolve truncating stores
24172 static SDValue PerformMSTORECombine(SDNode *N, SelectionDAG &DAG,
24173 const X86Subtarget *Subtarget) {
24174 MaskedStoreSDNode *Mst = cast<MaskedStoreSDNode>(N);
24175 if (!Mst->isTruncatingStore())
24178 EVT VT = Mst->getValue().getValueType();
24179 unsigned NumElems = VT.getVectorNumElements();
24180 EVT StVT = Mst->getMemoryVT();
24183 assert(StVT != VT && "Cannot truncate to the same type");
24184 unsigned FromSz = VT.getVectorElementType().getSizeInBits();
24185 unsigned ToSz = StVT.getVectorElementType().getSizeInBits();
24187 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
24189 // The truncating store is legal in some cases. For example
24190 // vpmovqb, vpmovqw, vpmovqd, vpmovdb, vpmovdw
24191 // are designated for truncate store.
24192 // In this case we don't need any further transformations.
24193 if (TLI.isTruncStoreLegal(VT, StVT))
24196 // From, To sizes and ElemCount must be pow of two
24197 assert (isPowerOf2_32(NumElems * FromSz * ToSz) &&
24198 "Unexpected size for truncating masked store");
24199 // We are going to use the original vector elt for storing.
24200 // Accumulated smaller vector elements must be a multiple of the store size.
24201 assert (((NumElems * FromSz) % ToSz) == 0 &&
24202 "Unexpected ratio for truncating masked store");
24204 unsigned SizeRatio = FromSz / ToSz;
24205 assert(SizeRatio * NumElems * ToSz == VT.getSizeInBits());
24207 // Create a type on which we perform the shuffle
24208 EVT WideVecVT = EVT::getVectorVT(*DAG.getContext(),
24209 StVT.getScalarType(), NumElems*SizeRatio);
24211 assert(WideVecVT.getSizeInBits() == VT.getSizeInBits());
24213 SDValue WideVec = DAG.getBitcast(WideVecVT, Mst->getValue());
24214 SmallVector<int, 16> ShuffleVec(NumElems * SizeRatio, -1);
24215 for (unsigned i = 0; i != NumElems; ++i)
24216 ShuffleVec[i] = i * SizeRatio;
24218 // Can't shuffle using an illegal type.
24219 assert (DAG.getTargetLoweringInfo().isTypeLegal(WideVecVT)
24220 && "WideVecVT should be legal");
24222 SDValue TruncatedVal = DAG.getVectorShuffle(WideVecVT, dl, WideVec,
24223 DAG.getUNDEF(WideVecVT),
24227 SDValue Mask = Mst->getMask();
24228 if (Mask.getValueType() == VT) {
24229 // Mask and original value have the same type
24230 NewMask = DAG.getBitcast(WideVecVT, Mask);
24231 for (unsigned i = 0; i != NumElems; ++i)
24232 ShuffleVec[i] = i * SizeRatio;
24233 for (unsigned i = NumElems; i != NumElems*SizeRatio; ++i)
24234 ShuffleVec[i] = NumElems*SizeRatio;
24235 NewMask = DAG.getVectorShuffle(WideVecVT, dl, NewMask,
24236 DAG.getConstant(0, dl, WideVecVT),
24240 assert(Mask.getValueType().getVectorElementType() == MVT::i1);
24241 unsigned WidenNumElts = NumElems*SizeRatio;
24242 unsigned MaskNumElts = VT.getVectorNumElements();
24243 EVT NewMaskVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
24246 unsigned NumConcat = WidenNumElts / MaskNumElts;
24247 SmallVector<SDValue, 16> Ops(NumConcat);
24248 SDValue ZeroVal = DAG.getConstant(0, dl, Mask.getValueType());
24250 for (unsigned i = 1; i != NumConcat; ++i)
24253 NewMask = DAG.getNode(ISD::CONCAT_VECTORS, dl, NewMaskVT, Ops);
24256 return DAG.getMaskedStore(Mst->getChain(), dl, TruncatedVal, Mst->getBasePtr(),
24257 NewMask, StVT, Mst->getMemOperand(), false);
24259 /// PerformSTORECombine - Do target-specific dag combines on STORE nodes.
24260 static SDValue PerformSTORECombine(SDNode *N, SelectionDAG &DAG,
24261 const X86Subtarget *Subtarget) {
24262 StoreSDNode *St = cast<StoreSDNode>(N);
24263 EVT VT = St->getValue().getValueType();
24264 EVT StVT = St->getMemoryVT();
24266 SDValue StoredVal = St->getOperand(1);
24267 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
24269 // If we are saving a concatenation of two XMM registers and 32-byte stores
24270 // are slow, such as on Sandy Bridge, perform two 16-byte stores.
24271 unsigned Alignment = St->getAlignment();
24272 bool IsAligned = Alignment == 0 || Alignment >= VT.getSizeInBits()/8;
24273 if (VT.is256BitVector() && Subtarget->isUnalignedMem32Slow() &&
24274 StVT == VT && !IsAligned) {
24275 unsigned NumElems = VT.getVectorNumElements();
24279 SDValue Value0 = Extract128BitVector(StoredVal, 0, DAG, dl);
24280 SDValue Value1 = Extract128BitVector(StoredVal, NumElems/2, DAG, dl);
24283 DAG.getConstant(16, dl, TLI.getPointerTy(DAG.getDataLayout()));
24284 SDValue Ptr0 = St->getBasePtr();
24285 SDValue Ptr1 = DAG.getNode(ISD::ADD, dl, Ptr0.getValueType(), Ptr0, Stride);
24287 SDValue Ch0 = DAG.getStore(St->getChain(), dl, Value0, Ptr0,
24288 St->getPointerInfo(), St->isVolatile(),
24289 St->isNonTemporal(), Alignment);
24290 SDValue Ch1 = DAG.getStore(St->getChain(), dl, Value1, Ptr1,
24291 St->getPointerInfo(), St->isVolatile(),
24292 St->isNonTemporal(),
24293 std::min(16U, Alignment));
24294 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Ch0, Ch1);
24297 // Optimize trunc store (of multiple scalars) to shuffle and store.
24298 // First, pack all of the elements in one place. Next, store to memory
24299 // in fewer chunks.
24300 if (St->isTruncatingStore() && VT.isVector()) {
24301 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
24302 unsigned NumElems = VT.getVectorNumElements();
24303 assert(StVT != VT && "Cannot truncate to the same type");
24304 unsigned FromSz = VT.getVectorElementType().getSizeInBits();
24305 unsigned ToSz = StVT.getVectorElementType().getSizeInBits();
24307 // The truncating store is legal in some cases. For example
24308 // vpmovqb, vpmovqw, vpmovqd, vpmovdb, vpmovdw
24309 // are designated for truncate store.
24310 // In this case we don't need any further transformations.
24311 if (TLI.isTruncStoreLegal(VT, StVT))
24314 // From, To sizes and ElemCount must be pow of two
24315 if (!isPowerOf2_32(NumElems * FromSz * ToSz)) return SDValue();
24316 // We are going to use the original vector elt for storing.
24317 // Accumulated smaller vector elements must be a multiple of the store size.
24318 if (0 != (NumElems * FromSz) % ToSz) return SDValue();
24320 unsigned SizeRatio = FromSz / ToSz;
24322 assert(SizeRatio * NumElems * ToSz == VT.getSizeInBits());
24324 // Create a type on which we perform the shuffle
24325 EVT WideVecVT = EVT::getVectorVT(*DAG.getContext(),
24326 StVT.getScalarType(), NumElems*SizeRatio);
24328 assert(WideVecVT.getSizeInBits() == VT.getSizeInBits());
24330 SDValue WideVec = DAG.getBitcast(WideVecVT, St->getValue());
24331 SmallVector<int, 8> ShuffleVec(NumElems * SizeRatio, -1);
24332 for (unsigned i = 0; i != NumElems; ++i)
24333 ShuffleVec[i] = i * SizeRatio;
24335 // Can't shuffle using an illegal type.
24336 if (!TLI.isTypeLegal(WideVecVT))
24339 SDValue Shuff = DAG.getVectorShuffle(WideVecVT, dl, WideVec,
24340 DAG.getUNDEF(WideVecVT),
24342 // At this point all of the data is stored at the bottom of the
24343 // register. We now need to save it to mem.
24345 // Find the largest store unit
24346 MVT StoreType = MVT::i8;
24347 for (MVT Tp : MVT::integer_valuetypes()) {
24348 if (TLI.isTypeLegal(Tp) && Tp.getSizeInBits() <= NumElems * ToSz)
24352 // On 32bit systems, we can't save 64bit integers. Try bitcasting to F64.
24353 if (TLI.isTypeLegal(MVT::f64) && StoreType.getSizeInBits() < 64 &&
24354 (64 <= NumElems * ToSz))
24355 StoreType = MVT::f64;
24357 // Bitcast the original vector into a vector of store-size units
24358 EVT StoreVecVT = EVT::getVectorVT(*DAG.getContext(),
24359 StoreType, VT.getSizeInBits()/StoreType.getSizeInBits());
24360 assert(StoreVecVT.getSizeInBits() == VT.getSizeInBits());
24361 SDValue ShuffWide = DAG.getBitcast(StoreVecVT, Shuff);
24362 SmallVector<SDValue, 8> Chains;
24363 SDValue Increment = DAG.getConstant(StoreType.getSizeInBits() / 8, dl,
24364 TLI.getPointerTy(DAG.getDataLayout()));
24365 SDValue Ptr = St->getBasePtr();
24367 // Perform one or more big stores into memory.
24368 for (unsigned i=0, e=(ToSz*NumElems)/StoreType.getSizeInBits(); i!=e; ++i) {
24369 SDValue SubVec = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl,
24370 StoreType, ShuffWide,
24371 DAG.getIntPtrConstant(i, dl));
24372 SDValue Ch = DAG.getStore(St->getChain(), dl, SubVec, Ptr,
24373 St->getPointerInfo(), St->isVolatile(),
24374 St->isNonTemporal(), St->getAlignment());
24375 Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr, Increment);
24376 Chains.push_back(Ch);
24379 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Chains);
24382 // Turn load->store of MMX types into GPR load/stores. This avoids clobbering
24383 // the FP state in cases where an emms may be missing.
24384 // A preferable solution to the general problem is to figure out the right
24385 // places to insert EMMS. This qualifies as a quick hack.
24387 // Similarly, turn load->store of i64 into double load/stores in 32-bit mode.
24388 if (VT.getSizeInBits() != 64)
24391 const Function *F = DAG.getMachineFunction().getFunction();
24392 bool NoImplicitFloatOps = F->hasFnAttribute(Attribute::NoImplicitFloat);
24394 !Subtarget->useSoftFloat() && !NoImplicitFloatOps && Subtarget->hasSSE2();
24395 if ((VT.isVector() ||
24396 (VT == MVT::i64 && F64IsLegal && !Subtarget->is64Bit())) &&
24397 isa<LoadSDNode>(St->getValue()) &&
24398 !cast<LoadSDNode>(St->getValue())->isVolatile() &&
24399 St->getChain().hasOneUse() && !St->isVolatile()) {
24400 SDNode* LdVal = St->getValue().getNode();
24401 LoadSDNode *Ld = nullptr;
24402 int TokenFactorIndex = -1;
24403 SmallVector<SDValue, 8> Ops;
24404 SDNode* ChainVal = St->getChain().getNode();
24405 // Must be a store of a load. We currently handle two cases: the load
24406 // is a direct child, and it's under an intervening TokenFactor. It is
24407 // possible to dig deeper under nested TokenFactors.
24408 if (ChainVal == LdVal)
24409 Ld = cast<LoadSDNode>(St->getChain());
24410 else if (St->getValue().hasOneUse() &&
24411 ChainVal->getOpcode() == ISD::TokenFactor) {
24412 for (unsigned i = 0, e = ChainVal->getNumOperands(); i != e; ++i) {
24413 if (ChainVal->getOperand(i).getNode() == LdVal) {
24414 TokenFactorIndex = i;
24415 Ld = cast<LoadSDNode>(St->getValue());
24417 Ops.push_back(ChainVal->getOperand(i));
24421 if (!Ld || !ISD::isNormalLoad(Ld))
24424 // If this is not the MMX case, i.e. we are just turning i64 load/store
24425 // into f64 load/store, avoid the transformation if there are multiple
24426 // uses of the loaded value.
24427 if (!VT.isVector() && !Ld->hasNUsesOfValue(1, 0))
24432 // If we are a 64-bit capable x86, lower to a single movq load/store pair.
24433 // Otherwise, if it's legal to use f64 SSE instructions, use f64 load/store
24435 if (Subtarget->is64Bit() || F64IsLegal) {
24436 EVT LdVT = Subtarget->is64Bit() ? MVT::i64 : MVT::f64;
24437 SDValue NewLd = DAG.getLoad(LdVT, LdDL, Ld->getChain(), Ld->getBasePtr(),
24438 Ld->getPointerInfo(), Ld->isVolatile(),
24439 Ld->isNonTemporal(), Ld->isInvariant(),
24440 Ld->getAlignment());
24441 SDValue NewChain = NewLd.getValue(1);
24442 if (TokenFactorIndex != -1) {
24443 Ops.push_back(NewChain);
24444 NewChain = DAG.getNode(ISD::TokenFactor, LdDL, MVT::Other, Ops);
24446 return DAG.getStore(NewChain, StDL, NewLd, St->getBasePtr(),
24447 St->getPointerInfo(),
24448 St->isVolatile(), St->isNonTemporal(),
24449 St->getAlignment());
24452 // Otherwise, lower to two pairs of 32-bit loads / stores.
24453 SDValue LoAddr = Ld->getBasePtr();
24454 SDValue HiAddr = DAG.getNode(ISD::ADD, LdDL, MVT::i32, LoAddr,
24455 DAG.getConstant(4, LdDL, MVT::i32));
24457 SDValue LoLd = DAG.getLoad(MVT::i32, LdDL, Ld->getChain(), LoAddr,
24458 Ld->getPointerInfo(),
24459 Ld->isVolatile(), Ld->isNonTemporal(),
24460 Ld->isInvariant(), Ld->getAlignment());
24461 SDValue HiLd = DAG.getLoad(MVT::i32, LdDL, Ld->getChain(), HiAddr,
24462 Ld->getPointerInfo().getWithOffset(4),
24463 Ld->isVolatile(), Ld->isNonTemporal(),
24465 MinAlign(Ld->getAlignment(), 4));
24467 SDValue NewChain = LoLd.getValue(1);
24468 if (TokenFactorIndex != -1) {
24469 Ops.push_back(LoLd);
24470 Ops.push_back(HiLd);
24471 NewChain = DAG.getNode(ISD::TokenFactor, LdDL, MVT::Other, Ops);
24474 LoAddr = St->getBasePtr();
24475 HiAddr = DAG.getNode(ISD::ADD, StDL, MVT::i32, LoAddr,
24476 DAG.getConstant(4, StDL, MVT::i32));
24478 SDValue LoSt = DAG.getStore(NewChain, StDL, LoLd, LoAddr,
24479 St->getPointerInfo(),
24480 St->isVolatile(), St->isNonTemporal(),
24481 St->getAlignment());
24482 SDValue HiSt = DAG.getStore(NewChain, StDL, HiLd, HiAddr,
24483 St->getPointerInfo().getWithOffset(4),
24485 St->isNonTemporal(),
24486 MinAlign(St->getAlignment(), 4));
24487 return DAG.getNode(ISD::TokenFactor, StDL, MVT::Other, LoSt, HiSt);
24490 // This is similar to the above case, but here we handle a scalar 64-bit
24491 // integer store that is extracted from a vector on a 32-bit target.
24492 // If we have SSE2, then we can treat it like a floating-point double
24493 // to get past legalization. The execution dependencies fixup pass will
24494 // choose the optimal machine instruction for the store if this really is
24495 // an integer or v2f32 rather than an f64.
24496 if (VT == MVT::i64 && F64IsLegal && !Subtarget->is64Bit() &&
24497 St->getOperand(1).getOpcode() == ISD::EXTRACT_VECTOR_ELT) {
24498 SDValue OldExtract = St->getOperand(1);
24499 SDValue ExtOp0 = OldExtract.getOperand(0);
24500 unsigned VecSize = ExtOp0.getValueSizeInBits();
24501 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), MVT::f64, VecSize / 64);
24502 SDValue BitCast = DAG.getBitcast(VecVT, ExtOp0);
24503 SDValue NewExtract = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
24504 BitCast, OldExtract.getOperand(1));
24505 return DAG.getStore(St->getChain(), dl, NewExtract, St->getBasePtr(),
24506 St->getPointerInfo(), St->isVolatile(),
24507 St->isNonTemporal(), St->getAlignment());
24513 /// Return 'true' if this vector operation is "horizontal"
24514 /// and return the operands for the horizontal operation in LHS and RHS. A
24515 /// horizontal operation performs the binary operation on successive elements
24516 /// of its first operand, then on successive elements of its second operand,
24517 /// returning the resulting values in a vector. For example, if
24518 /// A = < float a0, float a1, float a2, float a3 >
24520 /// B = < float b0, float b1, float b2, float b3 >
24521 /// then the result of doing a horizontal operation on A and B is
24522 /// A horizontal-op B = < a0 op a1, a2 op a3, b0 op b1, b2 op b3 >.
24523 /// In short, LHS and RHS are inspected to see if LHS op RHS is of the form
24524 /// A horizontal-op B, for some already available A and B, and if so then LHS is
24525 /// set to A, RHS to B, and the routine returns 'true'.
24526 /// Note that the binary operation should have the property that if one of the
24527 /// operands is UNDEF then the result is UNDEF.
24528 static bool isHorizontalBinOp(SDValue &LHS, SDValue &RHS, bool IsCommutative) {
24529 // Look for the following pattern: if
24530 // A = < float a0, float a1, float a2, float a3 >
24531 // B = < float b0, float b1, float b2, float b3 >
24533 // LHS = VECTOR_SHUFFLE A, B, <0, 2, 4, 6>
24534 // RHS = VECTOR_SHUFFLE A, B, <1, 3, 5, 7>
24535 // then LHS op RHS = < a0 op a1, a2 op a3, b0 op b1, b2 op b3 >
24536 // which is A horizontal-op B.
24538 // At least one of the operands should be a vector shuffle.
24539 if (LHS.getOpcode() != ISD::VECTOR_SHUFFLE &&
24540 RHS.getOpcode() != ISD::VECTOR_SHUFFLE)
24543 MVT VT = LHS.getSimpleValueType();
24545 assert((VT.is128BitVector() || VT.is256BitVector()) &&
24546 "Unsupported vector type for horizontal add/sub");
24548 // Handle 128 and 256-bit vector lengths. AVX defines horizontal add/sub to
24549 // operate independently on 128-bit lanes.
24550 unsigned NumElts = VT.getVectorNumElements();
24551 unsigned NumLanes = VT.getSizeInBits()/128;
24552 unsigned NumLaneElts = NumElts / NumLanes;
24553 assert((NumLaneElts % 2 == 0) &&
24554 "Vector type should have an even number of elements in each lane");
24555 unsigned HalfLaneElts = NumLaneElts/2;
24557 // View LHS in the form
24558 // LHS = VECTOR_SHUFFLE A, B, LMask
24559 // If LHS is not a shuffle then pretend it is the shuffle
24560 // LHS = VECTOR_SHUFFLE LHS, undef, <0, 1, ..., N-1>
24561 // NOTE: in what follows a default initialized SDValue represents an UNDEF of
24564 SmallVector<int, 16> LMask(NumElts);
24565 if (LHS.getOpcode() == ISD::VECTOR_SHUFFLE) {
24566 if (LHS.getOperand(0).getOpcode() != ISD::UNDEF)
24567 A = LHS.getOperand(0);
24568 if (LHS.getOperand(1).getOpcode() != ISD::UNDEF)
24569 B = LHS.getOperand(1);
24570 ArrayRef<int> Mask = cast<ShuffleVectorSDNode>(LHS.getNode())->getMask();
24571 std::copy(Mask.begin(), Mask.end(), LMask.begin());
24573 if (LHS.getOpcode() != ISD::UNDEF)
24575 for (unsigned i = 0; i != NumElts; ++i)
24579 // Likewise, view RHS in the form
24580 // RHS = VECTOR_SHUFFLE C, D, RMask
24582 SmallVector<int, 16> RMask(NumElts);
24583 if (RHS.getOpcode() == ISD::VECTOR_SHUFFLE) {
24584 if (RHS.getOperand(0).getOpcode() != ISD::UNDEF)
24585 C = RHS.getOperand(0);
24586 if (RHS.getOperand(1).getOpcode() != ISD::UNDEF)
24587 D = RHS.getOperand(1);
24588 ArrayRef<int> Mask = cast<ShuffleVectorSDNode>(RHS.getNode())->getMask();
24589 std::copy(Mask.begin(), Mask.end(), RMask.begin());
24591 if (RHS.getOpcode() != ISD::UNDEF)
24593 for (unsigned i = 0; i != NumElts; ++i)
24597 // Check that the shuffles are both shuffling the same vectors.
24598 if (!(A == C && B == D) && !(A == D && B == C))
24601 // If everything is UNDEF then bail out: it would be better to fold to UNDEF.
24602 if (!A.getNode() && !B.getNode())
24605 // If A and B occur in reverse order in RHS, then "swap" them (which means
24606 // rewriting the mask).
24608 ShuffleVectorSDNode::commuteMask(RMask);
24610 // At this point LHS and RHS are equivalent to
24611 // LHS = VECTOR_SHUFFLE A, B, LMask
24612 // RHS = VECTOR_SHUFFLE A, B, RMask
24613 // Check that the masks correspond to performing a horizontal operation.
24614 for (unsigned l = 0; l != NumElts; l += NumLaneElts) {
24615 for (unsigned i = 0; i != NumLaneElts; ++i) {
24616 int LIdx = LMask[i+l], RIdx = RMask[i+l];
24618 // Ignore any UNDEF components.
24619 if (LIdx < 0 || RIdx < 0 ||
24620 (!A.getNode() && (LIdx < (int)NumElts || RIdx < (int)NumElts)) ||
24621 (!B.getNode() && (LIdx >= (int)NumElts || RIdx >= (int)NumElts)))
24624 // Check that successive elements are being operated on. If not, this is
24625 // not a horizontal operation.
24626 unsigned Src = (i/HalfLaneElts); // each lane is split between srcs
24627 int Index = 2*(i%HalfLaneElts) + NumElts*Src + l;
24628 if (!(LIdx == Index && RIdx == Index + 1) &&
24629 !(IsCommutative && LIdx == Index + 1 && RIdx == Index))
24634 LHS = A.getNode() ? A : B; // If A is 'UNDEF', use B for it.
24635 RHS = B.getNode() ? B : A; // If B is 'UNDEF', use A for it.
24639 /// Do target-specific dag combines on floating point adds.
24640 static SDValue PerformFADDCombine(SDNode *N, SelectionDAG &DAG,
24641 const X86Subtarget *Subtarget) {
24642 EVT VT = N->getValueType(0);
24643 SDValue LHS = N->getOperand(0);
24644 SDValue RHS = N->getOperand(1);
24646 // Try to synthesize horizontal adds from adds of shuffles.
24647 if (((Subtarget->hasSSE3() && (VT == MVT::v4f32 || VT == MVT::v2f64)) ||
24648 (Subtarget->hasFp256() && (VT == MVT::v8f32 || VT == MVT::v4f64))) &&
24649 isHorizontalBinOp(LHS, RHS, true))
24650 return DAG.getNode(X86ISD::FHADD, SDLoc(N), VT, LHS, RHS);
24654 /// Do target-specific dag combines on floating point subs.
24655 static SDValue PerformFSUBCombine(SDNode *N, SelectionDAG &DAG,
24656 const X86Subtarget *Subtarget) {
24657 EVT VT = N->getValueType(0);
24658 SDValue LHS = N->getOperand(0);
24659 SDValue RHS = N->getOperand(1);
24661 // Try to synthesize horizontal subs from subs of shuffles.
24662 if (((Subtarget->hasSSE3() && (VT == MVT::v4f32 || VT == MVT::v2f64)) ||
24663 (Subtarget->hasFp256() && (VT == MVT::v8f32 || VT == MVT::v4f64))) &&
24664 isHorizontalBinOp(LHS, RHS, false))
24665 return DAG.getNode(X86ISD::FHSUB, SDLoc(N), VT, LHS, RHS);
24669 /// Do target-specific dag combines on X86ISD::FOR and X86ISD::FXOR nodes.
24670 static SDValue PerformFORCombine(SDNode *N, SelectionDAG &DAG) {
24671 assert(N->getOpcode() == X86ISD::FOR || N->getOpcode() == X86ISD::FXOR);
24673 // F[X]OR(0.0, x) -> x
24674 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
24675 if (C->getValueAPF().isPosZero())
24676 return N->getOperand(1);
24678 // F[X]OR(x, 0.0) -> x
24679 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
24680 if (C->getValueAPF().isPosZero())
24681 return N->getOperand(0);
24685 /// Do target-specific dag combines on X86ISD::FMIN and X86ISD::FMAX nodes.
24686 static SDValue PerformFMinFMaxCombine(SDNode *N, SelectionDAG &DAG) {
24687 assert(N->getOpcode() == X86ISD::FMIN || N->getOpcode() == X86ISD::FMAX);
24689 // Only perform optimizations if UnsafeMath is used.
24690 if (!DAG.getTarget().Options.UnsafeFPMath)
24693 // If we run in unsafe-math mode, then convert the FMAX and FMIN nodes
24694 // into FMINC and FMAXC, which are Commutative operations.
24695 unsigned NewOp = 0;
24696 switch (N->getOpcode()) {
24697 default: llvm_unreachable("unknown opcode");
24698 case X86ISD::FMIN: NewOp = X86ISD::FMINC; break;
24699 case X86ISD::FMAX: NewOp = X86ISD::FMAXC; break;
24702 return DAG.getNode(NewOp, SDLoc(N), N->getValueType(0),
24703 N->getOperand(0), N->getOperand(1));
24706 /// Do target-specific dag combines on X86ISD::FAND nodes.
24707 static SDValue PerformFANDCombine(SDNode *N, SelectionDAG &DAG) {
24708 // FAND(0.0, x) -> 0.0
24709 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
24710 if (C->getValueAPF().isPosZero())
24711 return N->getOperand(0);
24713 // FAND(x, 0.0) -> 0.0
24714 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
24715 if (C->getValueAPF().isPosZero())
24716 return N->getOperand(1);
24721 /// Do target-specific dag combines on X86ISD::FANDN nodes
24722 static SDValue PerformFANDNCombine(SDNode *N, SelectionDAG &DAG) {
24723 // FANDN(0.0, x) -> x
24724 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
24725 if (C->getValueAPF().isPosZero())
24726 return N->getOperand(1);
24728 // FANDN(x, 0.0) -> 0.0
24729 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
24730 if (C->getValueAPF().isPosZero())
24731 return N->getOperand(1);
24736 static SDValue PerformBTCombine(SDNode *N,
24738 TargetLowering::DAGCombinerInfo &DCI) {
24739 // BT ignores high bits in the bit index operand.
24740 SDValue Op1 = N->getOperand(1);
24741 if (Op1.hasOneUse()) {
24742 unsigned BitWidth = Op1.getValueSizeInBits();
24743 APInt DemandedMask = APInt::getLowBitsSet(BitWidth, Log2_32(BitWidth));
24744 APInt KnownZero, KnownOne;
24745 TargetLowering::TargetLoweringOpt TLO(DAG, !DCI.isBeforeLegalize(),
24746 !DCI.isBeforeLegalizeOps());
24747 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
24748 if (TLO.ShrinkDemandedConstant(Op1, DemandedMask) ||
24749 TLI.SimplifyDemandedBits(Op1, DemandedMask, KnownZero, KnownOne, TLO))
24750 DCI.CommitTargetLoweringOpt(TLO);
24755 static SDValue PerformVZEXT_MOVLCombine(SDNode *N, SelectionDAG &DAG) {
24756 SDValue Op = N->getOperand(0);
24757 if (Op.getOpcode() == ISD::BITCAST)
24758 Op = Op.getOperand(0);
24759 EVT VT = N->getValueType(0), OpVT = Op.getValueType();
24760 if (Op.getOpcode() == X86ISD::VZEXT_LOAD &&
24761 VT.getVectorElementType().getSizeInBits() ==
24762 OpVT.getVectorElementType().getSizeInBits()) {
24763 return DAG.getNode(ISD::BITCAST, SDLoc(N), VT, Op);
24768 static SDValue PerformSIGN_EXTEND_INREGCombine(SDNode *N, SelectionDAG &DAG,
24769 const X86Subtarget *Subtarget) {
24770 EVT VT = N->getValueType(0);
24771 if (!VT.isVector())
24774 SDValue N0 = N->getOperand(0);
24775 SDValue N1 = N->getOperand(1);
24776 EVT ExtraVT = cast<VTSDNode>(N1)->getVT();
24779 // The SIGN_EXTEND_INREG to v4i64 is expensive operation on the
24780 // both SSE and AVX2 since there is no sign-extended shift right
24781 // operation on a vector with 64-bit elements.
24782 //(sext_in_reg (v4i64 anyext (v4i32 x )), ExtraVT) ->
24783 // (v4i64 sext (v4i32 sext_in_reg (v4i32 x , ExtraVT)))
24784 if (VT == MVT::v4i64 && (N0.getOpcode() == ISD::ANY_EXTEND ||
24785 N0.getOpcode() == ISD::SIGN_EXTEND)) {
24786 SDValue N00 = N0.getOperand(0);
24788 // EXTLOAD has a better solution on AVX2,
24789 // it may be replaced with X86ISD::VSEXT node.
24790 if (N00.getOpcode() == ISD::LOAD && Subtarget->hasInt256())
24791 if (!ISD::isNormalLoad(N00.getNode()))
24794 if (N00.getValueType() == MVT::v4i32 && ExtraVT.getSizeInBits() < 128) {
24795 SDValue Tmp = DAG.getNode(ISD::SIGN_EXTEND_INREG, dl, MVT::v4i32,
24797 return DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i64, Tmp);
24803 static SDValue PerformSExtCombine(SDNode *N, SelectionDAG &DAG,
24804 TargetLowering::DAGCombinerInfo &DCI,
24805 const X86Subtarget *Subtarget) {
24806 SDValue N0 = N->getOperand(0);
24807 EVT VT = N->getValueType(0);
24808 EVT SVT = VT.getScalarType();
24809 EVT InVT = N0.getValueType();
24810 EVT InSVT = InVT.getScalarType();
24813 // (i8,i32 sext (sdivrem (i8 x, i8 y)) ->
24814 // (i8,i32 (sdivrem_sext_hreg (i8 x, i8 y)
24815 // This exposes the sext to the sdivrem lowering, so that it directly extends
24816 // from AH (which we otherwise need to do contortions to access).
24817 if (N0.getOpcode() == ISD::SDIVREM && N0.getResNo() == 1 &&
24818 InVT == MVT::i8 && VT == MVT::i32) {
24819 SDVTList NodeTys = DAG.getVTList(MVT::i8, VT);
24820 SDValue R = DAG.getNode(X86ISD::SDIVREM8_SEXT_HREG, DL, NodeTys,
24821 N0.getOperand(0), N0.getOperand(1));
24822 DAG.ReplaceAllUsesOfValueWith(N0.getValue(0), R.getValue(0));
24823 return R.getValue(1);
24826 if (!DCI.isBeforeLegalizeOps()) {
24827 if (InVT == MVT::i1) {
24828 SDValue Zero = DAG.getConstant(0, DL, VT);
24830 DAG.getConstant(APInt::getAllOnesValue(VT.getSizeInBits()), DL, VT);
24831 return DAG.getNode(ISD::SELECT, DL, VT, N0, AllOnes, Zero);
24836 if (VT.isVector() && Subtarget->hasSSE2()) {
24837 auto ExtendVecSize = [&DAG](SDLoc DL, SDValue N, unsigned Size) {
24838 EVT InVT = N.getValueType();
24839 EVT OutVT = EVT::getVectorVT(*DAG.getContext(), InVT.getScalarType(),
24840 Size / InVT.getScalarSizeInBits());
24841 SmallVector<SDValue, 8> Opnds(Size / InVT.getSizeInBits(),
24842 DAG.getUNDEF(InVT));
24844 return DAG.getNode(ISD::CONCAT_VECTORS, DL, OutVT, Opnds);
24847 // If target-size is less than 128-bits, extend to a type that would extend
24848 // to 128 bits, extend that and extract the original target vector.
24849 if (VT.getSizeInBits() < 128 && !(128 % VT.getSizeInBits()) &&
24850 (SVT == MVT::i64 || SVT == MVT::i32 || SVT == MVT::i16) &&
24851 (InSVT == MVT::i32 || InSVT == MVT::i16 || InSVT == MVT::i8)) {
24852 unsigned Scale = 128 / VT.getSizeInBits();
24854 EVT::getVectorVT(*DAG.getContext(), SVT, 128 / SVT.getSizeInBits());
24855 SDValue Ex = ExtendVecSize(DL, N0, Scale * InVT.getSizeInBits());
24856 SDValue SExt = DAG.getNode(ISD::SIGN_EXTEND, DL, ExVT, Ex);
24857 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, VT, SExt,
24858 DAG.getIntPtrConstant(0, DL));
24861 // If target-size is 128-bits, then convert to ISD::SIGN_EXTEND_VECTOR_INREG
24862 // which ensures lowering to X86ISD::VSEXT (pmovsx*).
24863 if (VT.getSizeInBits() == 128 &&
24864 (SVT == MVT::i64 || SVT == MVT::i32 || SVT == MVT::i16) &&
24865 (InSVT == MVT::i32 || InSVT == MVT::i16 || InSVT == MVT::i8)) {
24866 SDValue ExOp = ExtendVecSize(DL, N0, 128);
24867 return DAG.getSignExtendVectorInReg(ExOp, DL, VT);
24870 // On pre-AVX2 targets, split into 128-bit nodes of
24871 // ISD::SIGN_EXTEND_VECTOR_INREG.
24872 if (!Subtarget->hasInt256() && !(VT.getSizeInBits() % 128) &&
24873 (SVT == MVT::i64 || SVT == MVT::i32 || SVT == MVT::i16) &&
24874 (InSVT == MVT::i32 || InSVT == MVT::i16 || InSVT == MVT::i8)) {
24875 unsigned NumVecs = VT.getSizeInBits() / 128;
24876 unsigned NumSubElts = 128 / SVT.getSizeInBits();
24877 EVT SubVT = EVT::getVectorVT(*DAG.getContext(), SVT, NumSubElts);
24878 EVT InSubVT = EVT::getVectorVT(*DAG.getContext(), InSVT, NumSubElts);
24880 SmallVector<SDValue, 8> Opnds;
24881 for (unsigned i = 0, Offset = 0; i != NumVecs;
24882 ++i, Offset += NumSubElts) {
24883 SDValue SrcVec = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, InSubVT, N0,
24884 DAG.getIntPtrConstant(Offset, DL));
24885 SrcVec = ExtendVecSize(DL, SrcVec, 128);
24886 SrcVec = DAG.getSignExtendVectorInReg(SrcVec, DL, SubVT);
24887 Opnds.push_back(SrcVec);
24889 return DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, Opnds);
24893 if (!Subtarget->hasFp256())
24896 if (VT.isVector() && VT.getSizeInBits() == 256)
24897 if (SDValue R = WidenMaskArithmetic(N, DAG, DCI, Subtarget))
24903 static SDValue PerformFMACombine(SDNode *N, SelectionDAG &DAG,
24904 const X86Subtarget* Subtarget) {
24906 EVT VT = N->getValueType(0);
24908 // Let legalize expand this if it isn't a legal type yet.
24909 if (!DAG.getTargetLoweringInfo().isTypeLegal(VT))
24912 EVT ScalarVT = VT.getScalarType();
24913 if ((ScalarVT != MVT::f32 && ScalarVT != MVT::f64) ||
24914 (!Subtarget->hasFMA() && !Subtarget->hasFMA4() &&
24915 !Subtarget->hasAVX512()))
24918 SDValue A = N->getOperand(0);
24919 SDValue B = N->getOperand(1);
24920 SDValue C = N->getOperand(2);
24922 bool NegA = (A.getOpcode() == ISD::FNEG);
24923 bool NegB = (B.getOpcode() == ISD::FNEG);
24924 bool NegC = (C.getOpcode() == ISD::FNEG);
24926 // Negative multiplication when NegA xor NegB
24927 bool NegMul = (NegA != NegB);
24929 A = A.getOperand(0);
24931 B = B.getOperand(0);
24933 C = C.getOperand(0);
24937 Opcode = (!NegC) ? X86ISD::FMADD : X86ISD::FMSUB;
24939 Opcode = (!NegC) ? X86ISD::FNMADD : X86ISD::FNMSUB;
24941 return DAG.getNode(Opcode, dl, VT, A, B, C);
24944 static SDValue PerformZExtCombine(SDNode *N, SelectionDAG &DAG,
24945 TargetLowering::DAGCombinerInfo &DCI,
24946 const X86Subtarget *Subtarget) {
24947 // (i32 zext (and (i8 x86isd::setcc_carry), 1)) ->
24948 // (and (i32 x86isd::setcc_carry), 1)
24949 // This eliminates the zext. This transformation is necessary because
24950 // ISD::SETCC is always legalized to i8.
24952 SDValue N0 = N->getOperand(0);
24953 EVT VT = N->getValueType(0);
24955 if (N0.getOpcode() == ISD::AND &&
24957 N0.getOperand(0).hasOneUse()) {
24958 SDValue N00 = N0.getOperand(0);
24959 if (N00.getOpcode() == X86ISD::SETCC_CARRY) {
24960 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
24961 if (!C || C->getZExtValue() != 1)
24963 return DAG.getNode(ISD::AND, dl, VT,
24964 DAG.getNode(X86ISD::SETCC_CARRY, dl, VT,
24965 N00.getOperand(0), N00.getOperand(1)),
24966 DAG.getConstant(1, dl, VT));
24970 if (N0.getOpcode() == ISD::TRUNCATE &&
24972 N0.getOperand(0).hasOneUse()) {
24973 SDValue N00 = N0.getOperand(0);
24974 if (N00.getOpcode() == X86ISD::SETCC_CARRY) {
24975 return DAG.getNode(ISD::AND, dl, VT,
24976 DAG.getNode(X86ISD::SETCC_CARRY, dl, VT,
24977 N00.getOperand(0), N00.getOperand(1)),
24978 DAG.getConstant(1, dl, VT));
24982 if (VT.is256BitVector())
24983 if (SDValue R = WidenMaskArithmetic(N, DAG, DCI, Subtarget))
24986 // (i8,i32 zext (udivrem (i8 x, i8 y)) ->
24987 // (i8,i32 (udivrem_zext_hreg (i8 x, i8 y)
24988 // This exposes the zext to the udivrem lowering, so that it directly extends
24989 // from AH (which we otherwise need to do contortions to access).
24990 if (N0.getOpcode() == ISD::UDIVREM &&
24991 N0.getResNo() == 1 && N0.getValueType() == MVT::i8 &&
24992 (VT == MVT::i32 || VT == MVT::i64)) {
24993 SDVTList NodeTys = DAG.getVTList(MVT::i8, VT);
24994 SDValue R = DAG.getNode(X86ISD::UDIVREM8_ZEXT_HREG, dl, NodeTys,
24995 N0.getOperand(0), N0.getOperand(1));
24996 DAG.ReplaceAllUsesOfValueWith(N0.getValue(0), R.getValue(0));
24997 return R.getValue(1);
25003 // Optimize x == -y --> x+y == 0
25004 // x != -y --> x+y != 0
25005 static SDValue PerformISDSETCCCombine(SDNode *N, SelectionDAG &DAG,
25006 const X86Subtarget* Subtarget) {
25007 ISD::CondCode CC = cast<CondCodeSDNode>(N->getOperand(2))->get();
25008 SDValue LHS = N->getOperand(0);
25009 SDValue RHS = N->getOperand(1);
25010 EVT VT = N->getValueType(0);
25013 if ((CC == ISD::SETNE || CC == ISD::SETEQ) && LHS.getOpcode() == ISD::SUB)
25014 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(LHS.getOperand(0)))
25015 if (C->getAPIntValue() == 0 && LHS.hasOneUse()) {
25016 SDValue addV = DAG.getNode(ISD::ADD, DL, LHS.getValueType(), RHS,
25017 LHS.getOperand(1));
25018 return DAG.getSetCC(DL, N->getValueType(0), addV,
25019 DAG.getConstant(0, DL, addV.getValueType()), CC);
25021 if ((CC == ISD::SETNE || CC == ISD::SETEQ) && RHS.getOpcode() == ISD::SUB)
25022 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(RHS.getOperand(0)))
25023 if (C->getAPIntValue() == 0 && RHS.hasOneUse()) {
25024 SDValue addV = DAG.getNode(ISD::ADD, DL, RHS.getValueType(), LHS,
25025 RHS.getOperand(1));
25026 return DAG.getSetCC(DL, N->getValueType(0), addV,
25027 DAG.getConstant(0, DL, addV.getValueType()), CC);
25030 if (VT.getScalarType() == MVT::i1 &&
25031 (CC == ISD::SETNE || CC == ISD::SETEQ || ISD::isSignedIntSetCC(CC))) {
25033 (LHS.getOpcode() == ISD::SIGN_EXTEND) &&
25034 (LHS.getOperand(0).getValueType().getScalarType() == MVT::i1);
25035 bool IsVZero1 = ISD::isBuildVectorAllZeros(RHS.getNode());
25037 if (!IsSEXT0 || !IsVZero1) {
25038 // Swap the operands and update the condition code.
25039 std::swap(LHS, RHS);
25040 CC = ISD::getSetCCSwappedOperands(CC);
25042 IsSEXT0 = (LHS.getOpcode() == ISD::SIGN_EXTEND) &&
25043 (LHS.getOperand(0).getValueType().getScalarType() == MVT::i1);
25044 IsVZero1 = ISD::isBuildVectorAllZeros(RHS.getNode());
25047 if (IsSEXT0 && IsVZero1) {
25048 assert(VT == LHS.getOperand(0).getValueType() &&
25049 "Uexpected operand type");
25050 if (CC == ISD::SETGT)
25051 return DAG.getConstant(0, DL, VT);
25052 if (CC == ISD::SETLE)
25053 return DAG.getConstant(1, DL, VT);
25054 if (CC == ISD::SETEQ || CC == ISD::SETGE)
25055 return DAG.getNOT(DL, LHS.getOperand(0), VT);
25057 assert((CC == ISD::SETNE || CC == ISD::SETLT) &&
25058 "Unexpected condition code!");
25059 return LHS.getOperand(0);
25066 static SDValue NarrowVectorLoadToElement(LoadSDNode *Load, unsigned Index,
25067 SelectionDAG &DAG) {
25069 MVT VT = Load->getSimpleValueType(0);
25070 MVT EVT = VT.getVectorElementType();
25071 SDValue Addr = Load->getOperand(1);
25072 SDValue NewAddr = DAG.getNode(
25073 ISD::ADD, dl, Addr.getSimpleValueType(), Addr,
25074 DAG.getConstant(Index * EVT.getStoreSize(), dl,
25075 Addr.getSimpleValueType()));
25078 DAG.getLoad(EVT, dl, Load->getChain(), NewAddr,
25079 DAG.getMachineFunction().getMachineMemOperand(
25080 Load->getMemOperand(), 0, EVT.getStoreSize()));
25084 static SDValue PerformINSERTPSCombine(SDNode *N, SelectionDAG &DAG,
25085 const X86Subtarget *Subtarget) {
25087 MVT VT = N->getOperand(1)->getSimpleValueType(0);
25088 assert((VT == MVT::v4f32 || VT == MVT::v4i32) &&
25089 "X86insertps is only defined for v4x32");
25091 SDValue Ld = N->getOperand(1);
25092 if (MayFoldLoad(Ld)) {
25093 // Extract the countS bits from the immediate so we can get the proper
25094 // address when narrowing the vector load to a specific element.
25095 // When the second source op is a memory address, insertps doesn't use
25096 // countS and just gets an f32 from that address.
25097 unsigned DestIndex =
25098 cast<ConstantSDNode>(N->getOperand(2))->getZExtValue() >> 6;
25100 Ld = NarrowVectorLoadToElement(cast<LoadSDNode>(Ld), DestIndex, DAG);
25102 // Create this as a scalar to vector to match the instruction pattern.
25103 SDValue LoadScalarToVector = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Ld);
25104 // countS bits are ignored when loading from memory on insertps, which
25105 // means we don't need to explicitly set them to 0.
25106 return DAG.getNode(X86ISD::INSERTPS, dl, VT, N->getOperand(0),
25107 LoadScalarToVector, N->getOperand(2));
25112 static SDValue PerformBLENDICombine(SDNode *N, SelectionDAG &DAG) {
25113 SDValue V0 = N->getOperand(0);
25114 SDValue V1 = N->getOperand(1);
25116 EVT VT = N->getValueType(0);
25118 // Canonicalize a v2f64 blend with a mask of 2 by swapping the vector
25119 // operands and changing the mask to 1. This saves us a bunch of
25120 // pattern-matching possibilities related to scalar math ops in SSE/AVX.
25121 // x86InstrInfo knows how to commute this back after instruction selection
25122 // if it would help register allocation.
25124 // TODO: If optimizing for size or a processor that doesn't suffer from
25125 // partial register update stalls, this should be transformed into a MOVSD
25126 // instruction because a MOVSD is 1-2 bytes smaller than a BLENDPD.
25128 if (VT == MVT::v2f64)
25129 if (auto *Mask = dyn_cast<ConstantSDNode>(N->getOperand(2)))
25130 if (Mask->getZExtValue() == 2 && !isShuffleFoldableLoad(V0)) {
25131 SDValue NewMask = DAG.getConstant(1, DL, MVT::i8);
25132 return DAG.getNode(X86ISD::BLENDI, DL, VT, V1, V0, NewMask);
25138 // Helper function of PerformSETCCCombine. It is to materialize "setb reg"
25139 // as "sbb reg,reg", since it can be extended without zext and produces
25140 // an all-ones bit which is more useful than 0/1 in some cases.
25141 static SDValue MaterializeSETB(SDLoc DL, SDValue EFLAGS, SelectionDAG &DAG,
25144 return DAG.getNode(ISD::AND, DL, VT,
25145 DAG.getNode(X86ISD::SETCC_CARRY, DL, MVT::i8,
25146 DAG.getConstant(X86::COND_B, DL, MVT::i8),
25148 DAG.getConstant(1, DL, VT));
25149 assert (VT == MVT::i1 && "Unexpected type for SECCC node");
25150 return DAG.getNode(ISD::TRUNCATE, DL, MVT::i1,
25151 DAG.getNode(X86ISD::SETCC_CARRY, DL, MVT::i8,
25152 DAG.getConstant(X86::COND_B, DL, MVT::i8),
25156 // Optimize RES = X86ISD::SETCC CONDCODE, EFLAG_INPUT
25157 static SDValue PerformSETCCCombine(SDNode *N, SelectionDAG &DAG,
25158 TargetLowering::DAGCombinerInfo &DCI,
25159 const X86Subtarget *Subtarget) {
25161 X86::CondCode CC = X86::CondCode(N->getConstantOperandVal(0));
25162 SDValue EFLAGS = N->getOperand(1);
25164 if (CC == X86::COND_A) {
25165 // Try to convert COND_A into COND_B in an attempt to facilitate
25166 // materializing "setb reg".
25168 // Do not flip "e > c", where "c" is a constant, because Cmp instruction
25169 // cannot take an immediate as its first operand.
25171 if (EFLAGS.getOpcode() == X86ISD::SUB && EFLAGS.hasOneUse() &&
25172 EFLAGS.getValueType().isInteger() &&
25173 !isa<ConstantSDNode>(EFLAGS.getOperand(1))) {
25174 SDValue NewSub = DAG.getNode(X86ISD::SUB, SDLoc(EFLAGS),
25175 EFLAGS.getNode()->getVTList(),
25176 EFLAGS.getOperand(1), EFLAGS.getOperand(0));
25177 SDValue NewEFLAGS = SDValue(NewSub.getNode(), EFLAGS.getResNo());
25178 return MaterializeSETB(DL, NewEFLAGS, DAG, N->getSimpleValueType(0));
25182 // Materialize "setb reg" as "sbb reg,reg", since it can be extended without
25183 // a zext and produces an all-ones bit which is more useful than 0/1 in some
25185 if (CC == X86::COND_B)
25186 return MaterializeSETB(DL, EFLAGS, DAG, N->getSimpleValueType(0));
25188 if (SDValue Flags = checkBoolTestSetCCCombine(EFLAGS, CC)) {
25189 SDValue Cond = DAG.getConstant(CC, DL, MVT::i8);
25190 return DAG.getNode(X86ISD::SETCC, DL, N->getVTList(), Cond, Flags);
25196 // Optimize branch condition evaluation.
25198 static SDValue PerformBrCondCombine(SDNode *N, SelectionDAG &DAG,
25199 TargetLowering::DAGCombinerInfo &DCI,
25200 const X86Subtarget *Subtarget) {
25202 SDValue Chain = N->getOperand(0);
25203 SDValue Dest = N->getOperand(1);
25204 SDValue EFLAGS = N->getOperand(3);
25205 X86::CondCode CC = X86::CondCode(N->getConstantOperandVal(2));
25207 if (SDValue Flags = checkBoolTestSetCCCombine(EFLAGS, CC)) {
25208 SDValue Cond = DAG.getConstant(CC, DL, MVT::i8);
25209 return DAG.getNode(X86ISD::BRCOND, DL, N->getVTList(), Chain, Dest, Cond,
25216 static SDValue performVectorCompareAndMaskUnaryOpCombine(SDNode *N,
25217 SelectionDAG &DAG) {
25218 // Take advantage of vector comparisons producing 0 or -1 in each lane to
25219 // optimize away operation when it's from a constant.
25221 // The general transformation is:
25222 // UNARYOP(AND(VECTOR_CMP(x,y), constant)) -->
25223 // AND(VECTOR_CMP(x,y), constant2)
25224 // constant2 = UNARYOP(constant)
25226 // Early exit if this isn't a vector operation, the operand of the
25227 // unary operation isn't a bitwise AND, or if the sizes of the operations
25228 // aren't the same.
25229 EVT VT = N->getValueType(0);
25230 if (!VT.isVector() || N->getOperand(0)->getOpcode() != ISD::AND ||
25231 N->getOperand(0)->getOperand(0)->getOpcode() != ISD::SETCC ||
25232 VT.getSizeInBits() != N->getOperand(0)->getValueType(0).getSizeInBits())
25235 // Now check that the other operand of the AND is a constant. We could
25236 // make the transformation for non-constant splats as well, but it's unclear
25237 // that would be a benefit as it would not eliminate any operations, just
25238 // perform one more step in scalar code before moving to the vector unit.
25239 if (BuildVectorSDNode *BV =
25240 dyn_cast<BuildVectorSDNode>(N->getOperand(0)->getOperand(1))) {
25241 // Bail out if the vector isn't a constant.
25242 if (!BV->isConstant())
25245 // Everything checks out. Build up the new and improved node.
25247 EVT IntVT = BV->getValueType(0);
25248 // Create a new constant of the appropriate type for the transformed
25250 SDValue SourceConst = DAG.getNode(N->getOpcode(), DL, VT, SDValue(BV, 0));
25251 // The AND node needs bitcasts to/from an integer vector type around it.
25252 SDValue MaskConst = DAG.getBitcast(IntVT, SourceConst);
25253 SDValue NewAnd = DAG.getNode(ISD::AND, DL, IntVT,
25254 N->getOperand(0)->getOperand(0), MaskConst);
25255 SDValue Res = DAG.getBitcast(VT, NewAnd);
25262 static SDValue PerformUINT_TO_FPCombine(SDNode *N, SelectionDAG &DAG,
25263 const X86Subtarget *Subtarget) {
25264 SDValue Op0 = N->getOperand(0);
25265 EVT VT = N->getValueType(0);
25266 EVT InVT = Op0.getValueType();
25267 EVT InSVT = InVT.getScalarType();
25268 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
25270 // UINT_TO_FP(vXi8) -> SINT_TO_FP(ZEXT(vXi8 to vXi32))
25271 // UINT_TO_FP(vXi16) -> SINT_TO_FP(ZEXT(vXi16 to vXi32))
25272 if (InVT.isVector() && (InSVT == MVT::i8 || InSVT == MVT::i16)) {
25274 EVT DstVT = EVT::getVectorVT(*DAG.getContext(), MVT::i32,
25275 InVT.getVectorNumElements());
25276 SDValue P = DAG.getNode(ISD::ZERO_EXTEND, dl, DstVT, Op0);
25278 if (TLI.isOperationLegal(ISD::UINT_TO_FP, DstVT))
25279 return DAG.getNode(ISD::UINT_TO_FP, dl, VT, P);
25281 return DAG.getNode(ISD::SINT_TO_FP, dl, VT, P);
25287 static SDValue PerformSINT_TO_FPCombine(SDNode *N, SelectionDAG &DAG,
25288 const X86Subtarget *Subtarget) {
25289 // First try to optimize away the conversion entirely when it's
25290 // conditionally from a constant. Vectors only.
25291 if (SDValue Res = performVectorCompareAndMaskUnaryOpCombine(N, DAG))
25294 // Now move on to more general possibilities.
25295 SDValue Op0 = N->getOperand(0);
25296 EVT VT = N->getValueType(0);
25297 EVT InVT = Op0.getValueType();
25298 EVT InSVT = InVT.getScalarType();
25300 // SINT_TO_FP(vXi8) -> SINT_TO_FP(SEXT(vXi8 to vXi32))
25301 // SINT_TO_FP(vXi16) -> SINT_TO_FP(SEXT(vXi16 to vXi32))
25302 if (InVT.isVector() && (InSVT == MVT::i8 || InSVT == MVT::i16)) {
25304 EVT DstVT = EVT::getVectorVT(*DAG.getContext(), MVT::i32,
25305 InVT.getVectorNumElements());
25306 SDValue P = DAG.getNode(ISD::SIGN_EXTEND, dl, DstVT, Op0);
25307 return DAG.getNode(ISD::SINT_TO_FP, dl, VT, P);
25310 // Transform (SINT_TO_FP (i64 ...)) into an x87 operation if we have
25311 // a 32-bit target where SSE doesn't support i64->FP operations.
25312 if (Op0.getOpcode() == ISD::LOAD) {
25313 LoadSDNode *Ld = cast<LoadSDNode>(Op0.getNode());
25314 EVT LdVT = Ld->getValueType(0);
25316 // This transformation is not supported if the result type is f16
25317 if (VT == MVT::f16)
25320 if (!Ld->isVolatile() && !VT.isVector() &&
25321 ISD::isNON_EXTLoad(Op0.getNode()) && Op0.hasOneUse() &&
25322 !Subtarget->is64Bit() && LdVT == MVT::i64) {
25323 SDValue FILDChain = Subtarget->getTargetLowering()->BuildFILD(
25324 SDValue(N, 0), LdVT, Ld->getChain(), Op0, DAG);
25325 DAG.ReplaceAllUsesOfValueWith(Op0.getValue(1), FILDChain.getValue(1));
25332 // Optimize RES, EFLAGS = X86ISD::ADC LHS, RHS, EFLAGS
25333 static SDValue PerformADCCombine(SDNode *N, SelectionDAG &DAG,
25334 X86TargetLowering::DAGCombinerInfo &DCI) {
25335 // If the LHS and RHS of the ADC node are zero, then it can't overflow and
25336 // the result is either zero or one (depending on the input carry bit).
25337 // Strength reduce this down to a "set on carry" aka SETCC_CARRY&1.
25338 if (X86::isZeroNode(N->getOperand(0)) &&
25339 X86::isZeroNode(N->getOperand(1)) &&
25340 // We don't have a good way to replace an EFLAGS use, so only do this when
25342 SDValue(N, 1).use_empty()) {
25344 EVT VT = N->getValueType(0);
25345 SDValue CarryOut = DAG.getConstant(0, DL, N->getValueType(1));
25346 SDValue Res1 = DAG.getNode(ISD::AND, DL, VT,
25347 DAG.getNode(X86ISD::SETCC_CARRY, DL, VT,
25348 DAG.getConstant(X86::COND_B, DL,
25351 DAG.getConstant(1, DL, VT));
25352 return DCI.CombineTo(N, Res1, CarryOut);
25358 // fold (add Y, (sete X, 0)) -> adc 0, Y
25359 // (add Y, (setne X, 0)) -> sbb -1, Y
25360 // (sub (sete X, 0), Y) -> sbb 0, Y
25361 // (sub (setne X, 0), Y) -> adc -1, Y
25362 static SDValue OptimizeConditionalInDecrement(SDNode *N, SelectionDAG &DAG) {
25365 // Look through ZExts.
25366 SDValue Ext = N->getOperand(N->getOpcode() == ISD::SUB ? 1 : 0);
25367 if (Ext.getOpcode() != ISD::ZERO_EXTEND || !Ext.hasOneUse())
25370 SDValue SetCC = Ext.getOperand(0);
25371 if (SetCC.getOpcode() != X86ISD::SETCC || !SetCC.hasOneUse())
25374 X86::CondCode CC = (X86::CondCode)SetCC.getConstantOperandVal(0);
25375 if (CC != X86::COND_E && CC != X86::COND_NE)
25378 SDValue Cmp = SetCC.getOperand(1);
25379 if (Cmp.getOpcode() != X86ISD::CMP || !Cmp.hasOneUse() ||
25380 !X86::isZeroNode(Cmp.getOperand(1)) ||
25381 !Cmp.getOperand(0).getValueType().isInteger())
25384 SDValue CmpOp0 = Cmp.getOperand(0);
25385 SDValue NewCmp = DAG.getNode(X86ISD::CMP, DL, MVT::i32, CmpOp0,
25386 DAG.getConstant(1, DL, CmpOp0.getValueType()));
25388 SDValue OtherVal = N->getOperand(N->getOpcode() == ISD::SUB ? 0 : 1);
25389 if (CC == X86::COND_NE)
25390 return DAG.getNode(N->getOpcode() == ISD::SUB ? X86ISD::ADC : X86ISD::SBB,
25391 DL, OtherVal.getValueType(), OtherVal,
25392 DAG.getConstant(-1ULL, DL, OtherVal.getValueType()),
25394 return DAG.getNode(N->getOpcode() == ISD::SUB ? X86ISD::SBB : X86ISD::ADC,
25395 DL, OtherVal.getValueType(), OtherVal,
25396 DAG.getConstant(0, DL, OtherVal.getValueType()), NewCmp);
25399 /// PerformADDCombine - Do target-specific dag combines on integer adds.
25400 static SDValue PerformAddCombine(SDNode *N, SelectionDAG &DAG,
25401 const X86Subtarget *Subtarget) {
25402 EVT VT = N->getValueType(0);
25403 SDValue Op0 = N->getOperand(0);
25404 SDValue Op1 = N->getOperand(1);
25406 // Try to synthesize horizontal adds from adds of shuffles.
25407 if (((Subtarget->hasSSSE3() && (VT == MVT::v8i16 || VT == MVT::v4i32)) ||
25408 (Subtarget->hasInt256() && (VT == MVT::v16i16 || VT == MVT::v8i32))) &&
25409 isHorizontalBinOp(Op0, Op1, true))
25410 return DAG.getNode(X86ISD::HADD, SDLoc(N), VT, Op0, Op1);
25412 return OptimizeConditionalInDecrement(N, DAG);
25415 static SDValue PerformSubCombine(SDNode *N, SelectionDAG &DAG,
25416 const X86Subtarget *Subtarget) {
25417 SDValue Op0 = N->getOperand(0);
25418 SDValue Op1 = N->getOperand(1);
25420 // X86 can't encode an immediate LHS of a sub. See if we can push the
25421 // negation into a preceding instruction.
25422 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op0)) {
25423 // If the RHS of the sub is a XOR with one use and a constant, invert the
25424 // immediate. Then add one to the LHS of the sub so we can turn
25425 // X-Y -> X+~Y+1, saving one register.
25426 if (Op1->hasOneUse() && Op1.getOpcode() == ISD::XOR &&
25427 isa<ConstantSDNode>(Op1.getOperand(1))) {
25428 APInt XorC = cast<ConstantSDNode>(Op1.getOperand(1))->getAPIntValue();
25429 EVT VT = Op0.getValueType();
25430 SDValue NewXor = DAG.getNode(ISD::XOR, SDLoc(Op1), VT,
25432 DAG.getConstant(~XorC, SDLoc(Op1), VT));
25433 return DAG.getNode(ISD::ADD, SDLoc(N), VT, NewXor,
25434 DAG.getConstant(C->getAPIntValue() + 1, SDLoc(N), VT));
25438 // Try to synthesize horizontal adds from adds of shuffles.
25439 EVT VT = N->getValueType(0);
25440 if (((Subtarget->hasSSSE3() && (VT == MVT::v8i16 || VT == MVT::v4i32)) ||
25441 (Subtarget->hasInt256() && (VT == MVT::v16i16 || VT == MVT::v8i32))) &&
25442 isHorizontalBinOp(Op0, Op1, true))
25443 return DAG.getNode(X86ISD::HSUB, SDLoc(N), VT, Op0, Op1);
25445 return OptimizeConditionalInDecrement(N, DAG);
25448 /// performVZEXTCombine - Performs build vector combines
25449 static SDValue performVZEXTCombine(SDNode *N, SelectionDAG &DAG,
25450 TargetLowering::DAGCombinerInfo &DCI,
25451 const X86Subtarget *Subtarget) {
25453 MVT VT = N->getSimpleValueType(0);
25454 SDValue Op = N->getOperand(0);
25455 MVT OpVT = Op.getSimpleValueType();
25456 MVT OpEltVT = OpVT.getVectorElementType();
25457 unsigned InputBits = OpEltVT.getSizeInBits() * VT.getVectorNumElements();
25459 // (vzext (bitcast (vzext (x)) -> (vzext x)
25461 while (V.getOpcode() == ISD::BITCAST)
25462 V = V.getOperand(0);
25464 if (V != Op && V.getOpcode() == X86ISD::VZEXT) {
25465 MVT InnerVT = V.getSimpleValueType();
25466 MVT InnerEltVT = InnerVT.getVectorElementType();
25468 // If the element sizes match exactly, we can just do one larger vzext. This
25469 // is always an exact type match as vzext operates on integer types.
25470 if (OpEltVT == InnerEltVT) {
25471 assert(OpVT == InnerVT && "Types must match for vzext!");
25472 return DAG.getNode(X86ISD::VZEXT, DL, VT, V.getOperand(0));
25475 // The only other way we can combine them is if only a single element of the
25476 // inner vzext is used in the input to the outer vzext.
25477 if (InnerEltVT.getSizeInBits() < InputBits)
25480 // In this case, the inner vzext is completely dead because we're going to
25481 // only look at bits inside of the low element. Just do the outer vzext on
25482 // a bitcast of the input to the inner.
25483 return DAG.getNode(X86ISD::VZEXT, DL, VT, DAG.getBitcast(OpVT, V));
25486 // Check if we can bypass extracting and re-inserting an element of an input
25487 // vector. Essentialy:
25488 // (bitcast (sclr2vec (ext_vec_elt x))) -> (bitcast x)
25489 if (V.getOpcode() == ISD::SCALAR_TO_VECTOR &&
25490 V.getOperand(0).getOpcode() == ISD::EXTRACT_VECTOR_ELT &&
25491 V.getOperand(0).getSimpleValueType().getSizeInBits() == InputBits) {
25492 SDValue ExtractedV = V.getOperand(0);
25493 SDValue OrigV = ExtractedV.getOperand(0);
25494 if (auto *ExtractIdx = dyn_cast<ConstantSDNode>(ExtractedV.getOperand(1)))
25495 if (ExtractIdx->getZExtValue() == 0) {
25496 MVT OrigVT = OrigV.getSimpleValueType();
25497 // Extract a subvector if necessary...
25498 if (OrigVT.getSizeInBits() > OpVT.getSizeInBits()) {
25499 int Ratio = OrigVT.getSizeInBits() / OpVT.getSizeInBits();
25500 OrigVT = MVT::getVectorVT(OrigVT.getVectorElementType(),
25501 OrigVT.getVectorNumElements() / Ratio);
25502 OrigV = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, OrigVT, OrigV,
25503 DAG.getIntPtrConstant(0, DL));
25505 Op = DAG.getBitcast(OpVT, OrigV);
25506 return DAG.getNode(X86ISD::VZEXT, DL, VT, Op);
25513 SDValue X86TargetLowering::PerformDAGCombine(SDNode *N,
25514 DAGCombinerInfo &DCI) const {
25515 SelectionDAG &DAG = DCI.DAG;
25516 switch (N->getOpcode()) {
25518 case ISD::EXTRACT_VECTOR_ELT:
25519 return PerformEXTRACT_VECTOR_ELTCombine(N, DAG, DCI);
25522 case X86ISD::SHRUNKBLEND:
25523 return PerformSELECTCombine(N, DAG, DCI, Subtarget);
25524 case ISD::BITCAST: return PerformBITCASTCombine(N, DAG);
25525 case X86ISD::CMOV: return PerformCMOVCombine(N, DAG, DCI, Subtarget);
25526 case ISD::ADD: return PerformAddCombine(N, DAG, Subtarget);
25527 case ISD::SUB: return PerformSubCombine(N, DAG, Subtarget);
25528 case X86ISD::ADC: return PerformADCCombine(N, DAG, DCI);
25529 case ISD::MUL: return PerformMulCombine(N, DAG, DCI);
25532 case ISD::SRL: return PerformShiftCombine(N, DAG, DCI, Subtarget);
25533 case ISD::AND: return PerformAndCombine(N, DAG, DCI, Subtarget);
25534 case ISD::OR: return PerformOrCombine(N, DAG, DCI, Subtarget);
25535 case ISD::XOR: return PerformXorCombine(N, DAG, DCI, Subtarget);
25536 case ISD::LOAD: return PerformLOADCombine(N, DAG, DCI, Subtarget);
25537 case ISD::MLOAD: return PerformMLOADCombine(N, DAG, DCI, Subtarget);
25538 case ISD::STORE: return PerformSTORECombine(N, DAG, Subtarget);
25539 case ISD::MSTORE: return PerformMSTORECombine(N, DAG, Subtarget);
25540 case ISD::SINT_TO_FP: return PerformSINT_TO_FPCombine(N, DAG, Subtarget);
25541 case ISD::UINT_TO_FP: return PerformUINT_TO_FPCombine(N, DAG, Subtarget);
25542 case ISD::FADD: return PerformFADDCombine(N, DAG, Subtarget);
25543 case ISD::FSUB: return PerformFSUBCombine(N, DAG, Subtarget);
25545 case X86ISD::FOR: return PerformFORCombine(N, DAG);
25547 case X86ISD::FMAX: return PerformFMinFMaxCombine(N, DAG);
25548 case X86ISD::FAND: return PerformFANDCombine(N, DAG);
25549 case X86ISD::FANDN: return PerformFANDNCombine(N, DAG);
25550 case X86ISD::BT: return PerformBTCombine(N, DAG, DCI);
25551 case X86ISD::VZEXT_MOVL: return PerformVZEXT_MOVLCombine(N, DAG);
25552 case ISD::ANY_EXTEND:
25553 case ISD::ZERO_EXTEND: return PerformZExtCombine(N, DAG, DCI, Subtarget);
25554 case ISD::SIGN_EXTEND: return PerformSExtCombine(N, DAG, DCI, Subtarget);
25555 case ISD::SIGN_EXTEND_INREG:
25556 return PerformSIGN_EXTEND_INREGCombine(N, DAG, Subtarget);
25557 case ISD::SETCC: return PerformISDSETCCCombine(N, DAG, Subtarget);
25558 case X86ISD::SETCC: return PerformSETCCCombine(N, DAG, DCI, Subtarget);
25559 case X86ISD::BRCOND: return PerformBrCondCombine(N, DAG, DCI, Subtarget);
25560 case X86ISD::VZEXT: return performVZEXTCombine(N, DAG, DCI, Subtarget);
25561 case X86ISD::SHUFP: // Handle all target specific shuffles
25562 case X86ISD::PALIGNR:
25563 case X86ISD::UNPCKH:
25564 case X86ISD::UNPCKL:
25565 case X86ISD::MOVHLPS:
25566 case X86ISD::MOVLHPS:
25567 case X86ISD::PSHUFB:
25568 case X86ISD::PSHUFD:
25569 case X86ISD::PSHUFHW:
25570 case X86ISD::PSHUFLW:
25571 case X86ISD::MOVSS:
25572 case X86ISD::MOVSD:
25573 case X86ISD::VPERMILPI:
25574 case X86ISD::VPERM2X128:
25575 case ISD::VECTOR_SHUFFLE: return PerformShuffleCombine(N, DAG, DCI,Subtarget);
25576 case ISD::FMA: return PerformFMACombine(N, DAG, Subtarget);
25577 case ISD::INTRINSIC_WO_CHAIN:
25578 return PerformINTRINSIC_WO_CHAINCombine(N, DAG, Subtarget);
25579 case X86ISD::INSERTPS: {
25580 if (getTargetMachine().getOptLevel() > CodeGenOpt::None)
25581 return PerformINSERTPSCombine(N, DAG, Subtarget);
25584 case X86ISD::BLENDI: return PerformBLENDICombine(N, DAG);
25590 /// isTypeDesirableForOp - Return true if the target has native support for
25591 /// the specified value type and it is 'desirable' to use the type for the
25592 /// given node type. e.g. On x86 i16 is legal, but undesirable since i16
25593 /// instruction encodings are longer and some i16 instructions are slow.
25594 bool X86TargetLowering::isTypeDesirableForOp(unsigned Opc, EVT VT) const {
25595 if (!isTypeLegal(VT))
25597 if (VT != MVT::i16)
25604 case ISD::SIGN_EXTEND:
25605 case ISD::ZERO_EXTEND:
25606 case ISD::ANY_EXTEND:
25619 /// IsDesirableToPromoteOp - This method query the target whether it is
25620 /// beneficial for dag combiner to promote the specified node. If true, it
25621 /// should return the desired promotion type by reference.
25622 bool X86TargetLowering::IsDesirableToPromoteOp(SDValue Op, EVT &PVT) const {
25623 EVT VT = Op.getValueType();
25624 if (VT != MVT::i16)
25627 bool Promote = false;
25628 bool Commute = false;
25629 switch (Op.getOpcode()) {
25632 LoadSDNode *LD = cast<LoadSDNode>(Op);
25633 // If the non-extending load has a single use and it's not live out, then it
25634 // might be folded.
25635 if (LD->getExtensionType() == ISD::NON_EXTLOAD /*&&
25636 Op.hasOneUse()*/) {
25637 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
25638 UE = Op.getNode()->use_end(); UI != UE; ++UI) {
25639 // The only case where we'd want to promote LOAD (rather then it being
25640 // promoted as an operand is when it's only use is liveout.
25641 if (UI->getOpcode() != ISD::CopyToReg)
25648 case ISD::SIGN_EXTEND:
25649 case ISD::ZERO_EXTEND:
25650 case ISD::ANY_EXTEND:
25655 SDValue N0 = Op.getOperand(0);
25656 // Look out for (store (shl (load), x)).
25657 if (MayFoldLoad(N0) && MayFoldIntoStore(Op))
25670 SDValue N0 = Op.getOperand(0);
25671 SDValue N1 = Op.getOperand(1);
25672 if (!Commute && MayFoldLoad(N1))
25674 // Avoid disabling potential load folding opportunities.
25675 if (MayFoldLoad(N0) && (!isa<ConstantSDNode>(N1) || MayFoldIntoStore(Op)))
25677 if (MayFoldLoad(N1) && (!isa<ConstantSDNode>(N0) || MayFoldIntoStore(Op)))
25687 //===----------------------------------------------------------------------===//
25688 // X86 Inline Assembly Support
25689 //===----------------------------------------------------------------------===//
25691 // Helper to match a string separated by whitespace.
25692 static bool matchAsm(StringRef S, ArrayRef<const char *> Pieces) {
25693 S = S.substr(S.find_first_not_of(" \t")); // Skip leading whitespace.
25695 for (StringRef Piece : Pieces) {
25696 if (!S.startswith(Piece)) // Check if the piece matches.
25699 S = S.substr(Piece.size());
25700 StringRef::size_type Pos = S.find_first_not_of(" \t");
25701 if (Pos == 0) // We matched a prefix.
25710 static bool clobbersFlagRegisters(const SmallVector<StringRef, 4> &AsmPieces) {
25712 if (AsmPieces.size() == 3 || AsmPieces.size() == 4) {
25713 if (std::count(AsmPieces.begin(), AsmPieces.end(), "~{cc}") &&
25714 std::count(AsmPieces.begin(), AsmPieces.end(), "~{flags}") &&
25715 std::count(AsmPieces.begin(), AsmPieces.end(), "~{fpsr}")) {
25717 if (AsmPieces.size() == 3)
25719 else if (std::count(AsmPieces.begin(), AsmPieces.end(), "~{dirflag}"))
25726 bool X86TargetLowering::ExpandInlineAsm(CallInst *CI) const {
25727 InlineAsm *IA = cast<InlineAsm>(CI->getCalledValue());
25729 std::string AsmStr = IA->getAsmString();
25731 IntegerType *Ty = dyn_cast<IntegerType>(CI->getType());
25732 if (!Ty || Ty->getBitWidth() % 16 != 0)
25735 // TODO: should remove alternatives from the asmstring: "foo {a|b}" -> "foo a"
25736 SmallVector<StringRef, 4> AsmPieces;
25737 SplitString(AsmStr, AsmPieces, ";\n");
25739 switch (AsmPieces.size()) {
25740 default: return false;
25742 // FIXME: this should verify that we are targeting a 486 or better. If not,
25743 // we will turn this bswap into something that will be lowered to logical
25744 // ops instead of emitting the bswap asm. For now, we don't support 486 or
25745 // lower so don't worry about this.
25747 if (matchAsm(AsmPieces[0], {"bswap", "$0"}) ||
25748 matchAsm(AsmPieces[0], {"bswapl", "$0"}) ||
25749 matchAsm(AsmPieces[0], {"bswapq", "$0"}) ||
25750 matchAsm(AsmPieces[0], {"bswap", "${0:q}"}) ||
25751 matchAsm(AsmPieces[0], {"bswapl", "${0:q}"}) ||
25752 matchAsm(AsmPieces[0], {"bswapq", "${0:q}"})) {
25753 // No need to check constraints, nothing other than the equivalent of
25754 // "=r,0" would be valid here.
25755 return IntrinsicLowering::LowerToByteSwap(CI);
25758 // rorw $$8, ${0:w} --> llvm.bswap.i16
25759 if (CI->getType()->isIntegerTy(16) &&
25760 IA->getConstraintString().compare(0, 5, "=r,0,") == 0 &&
25761 (matchAsm(AsmPieces[0], {"rorw", "$$8,", "${0:w}"}) ||
25762 matchAsm(AsmPieces[0], {"rolw", "$$8,", "${0:w}"}))) {
25764 StringRef ConstraintsStr = IA->getConstraintString();
25765 SplitString(StringRef(ConstraintsStr).substr(5), AsmPieces, ",");
25766 array_pod_sort(AsmPieces.begin(), AsmPieces.end());
25767 if (clobbersFlagRegisters(AsmPieces))
25768 return IntrinsicLowering::LowerToByteSwap(CI);
25772 if (CI->getType()->isIntegerTy(32) &&
25773 IA->getConstraintString().compare(0, 5, "=r,0,") == 0 &&
25774 matchAsm(AsmPieces[0], {"rorw", "$$8,", "${0:w}"}) &&
25775 matchAsm(AsmPieces[1], {"rorl", "$$16,", "$0"}) &&
25776 matchAsm(AsmPieces[2], {"rorw", "$$8,", "${0:w}"})) {
25778 StringRef ConstraintsStr = IA->getConstraintString();
25779 SplitString(StringRef(ConstraintsStr).substr(5), AsmPieces, ",");
25780 array_pod_sort(AsmPieces.begin(), AsmPieces.end());
25781 if (clobbersFlagRegisters(AsmPieces))
25782 return IntrinsicLowering::LowerToByteSwap(CI);
25785 if (CI->getType()->isIntegerTy(64)) {
25786 InlineAsm::ConstraintInfoVector Constraints = IA->ParseConstraints();
25787 if (Constraints.size() >= 2 &&
25788 Constraints[0].Codes.size() == 1 && Constraints[0].Codes[0] == "A" &&
25789 Constraints[1].Codes.size() == 1 && Constraints[1].Codes[0] == "0") {
25790 // bswap %eax / bswap %edx / xchgl %eax, %edx -> llvm.bswap.i64
25791 if (matchAsm(AsmPieces[0], {"bswap", "%eax"}) &&
25792 matchAsm(AsmPieces[1], {"bswap", "%edx"}) &&
25793 matchAsm(AsmPieces[2], {"xchgl", "%eax,", "%edx"}))
25794 return IntrinsicLowering::LowerToByteSwap(CI);
25802 /// getConstraintType - Given a constraint letter, return the type of
25803 /// constraint it is for this target.
25804 X86TargetLowering::ConstraintType
25805 X86TargetLowering::getConstraintType(StringRef Constraint) const {
25806 if (Constraint.size() == 1) {
25807 switch (Constraint[0]) {
25818 return C_RegisterClass;
25842 return TargetLowering::getConstraintType(Constraint);
25845 /// Examine constraint type and operand type and determine a weight value.
25846 /// This object must already have been set up with the operand type
25847 /// and the current alternative constraint selected.
25848 TargetLowering::ConstraintWeight
25849 X86TargetLowering::getSingleConstraintMatchWeight(
25850 AsmOperandInfo &info, const char *constraint) const {
25851 ConstraintWeight weight = CW_Invalid;
25852 Value *CallOperandVal = info.CallOperandVal;
25853 // If we don't have a value, we can't do a match,
25854 // but allow it at the lowest weight.
25855 if (!CallOperandVal)
25857 Type *type = CallOperandVal->getType();
25858 // Look at the constraint type.
25859 switch (*constraint) {
25861 weight = TargetLowering::getSingleConstraintMatchWeight(info, constraint);
25872 if (CallOperandVal->getType()->isIntegerTy())
25873 weight = CW_SpecificReg;
25878 if (type->isFloatingPointTy())
25879 weight = CW_SpecificReg;
25882 if (type->isX86_MMXTy() && Subtarget->hasMMX())
25883 weight = CW_SpecificReg;
25887 if (((type->getPrimitiveSizeInBits() == 128) && Subtarget->hasSSE1()) ||
25888 ((type->getPrimitiveSizeInBits() == 256) && Subtarget->hasFp256()))
25889 weight = CW_Register;
25892 if (ConstantInt *C = dyn_cast<ConstantInt>(info.CallOperandVal)) {
25893 if (C->getZExtValue() <= 31)
25894 weight = CW_Constant;
25898 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
25899 if (C->getZExtValue() <= 63)
25900 weight = CW_Constant;
25904 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
25905 if ((C->getSExtValue() >= -0x80) && (C->getSExtValue() <= 0x7f))
25906 weight = CW_Constant;
25910 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
25911 if ((C->getZExtValue() == 0xff) || (C->getZExtValue() == 0xffff))
25912 weight = CW_Constant;
25916 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
25917 if (C->getZExtValue() <= 3)
25918 weight = CW_Constant;
25922 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
25923 if (C->getZExtValue() <= 0xff)
25924 weight = CW_Constant;
25929 if (isa<ConstantFP>(CallOperandVal)) {
25930 weight = CW_Constant;
25934 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
25935 if ((C->getSExtValue() >= -0x80000000LL) &&
25936 (C->getSExtValue() <= 0x7fffffffLL))
25937 weight = CW_Constant;
25941 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
25942 if (C->getZExtValue() <= 0xffffffff)
25943 weight = CW_Constant;
25950 /// LowerXConstraint - try to replace an X constraint, which matches anything,
25951 /// with another that has more specific requirements based on the type of the
25952 /// corresponding operand.
25953 const char *X86TargetLowering::
25954 LowerXConstraint(EVT ConstraintVT) const {
25955 // FP X constraints get lowered to SSE1/2 registers if available, otherwise
25956 // 'f' like normal targets.
25957 if (ConstraintVT.isFloatingPoint()) {
25958 if (Subtarget->hasSSE2())
25960 if (Subtarget->hasSSE1())
25964 return TargetLowering::LowerXConstraint(ConstraintVT);
25967 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
25968 /// vector. If it is invalid, don't add anything to Ops.
25969 void X86TargetLowering::LowerAsmOperandForConstraint(SDValue Op,
25970 std::string &Constraint,
25971 std::vector<SDValue>&Ops,
25972 SelectionDAG &DAG) const {
25975 // Only support length 1 constraints for now.
25976 if (Constraint.length() > 1) return;
25978 char ConstraintLetter = Constraint[0];
25979 switch (ConstraintLetter) {
25982 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
25983 if (C->getZExtValue() <= 31) {
25984 Result = DAG.getTargetConstant(C->getZExtValue(), SDLoc(Op),
25985 Op.getValueType());
25991 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
25992 if (C->getZExtValue() <= 63) {
25993 Result = DAG.getTargetConstant(C->getZExtValue(), SDLoc(Op),
25994 Op.getValueType());
26000 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
26001 if (isInt<8>(C->getSExtValue())) {
26002 Result = DAG.getTargetConstant(C->getZExtValue(), SDLoc(Op),
26003 Op.getValueType());
26009 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
26010 if (C->getZExtValue() == 0xff || C->getZExtValue() == 0xffff ||
26011 (Subtarget->is64Bit() && C->getZExtValue() == 0xffffffff)) {
26012 Result = DAG.getTargetConstant(C->getSExtValue(), SDLoc(Op),
26013 Op.getValueType());
26019 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
26020 if (C->getZExtValue() <= 3) {
26021 Result = DAG.getTargetConstant(C->getZExtValue(), SDLoc(Op),
26022 Op.getValueType());
26028 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
26029 if (C->getZExtValue() <= 255) {
26030 Result = DAG.getTargetConstant(C->getZExtValue(), SDLoc(Op),
26031 Op.getValueType());
26037 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
26038 if (C->getZExtValue() <= 127) {
26039 Result = DAG.getTargetConstant(C->getZExtValue(), SDLoc(Op),
26040 Op.getValueType());
26046 // 32-bit signed value
26047 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
26048 if (ConstantInt::isValueValidForType(Type::getInt32Ty(*DAG.getContext()),
26049 C->getSExtValue())) {
26050 // Widen to 64 bits here to get it sign extended.
26051 Result = DAG.getTargetConstant(C->getSExtValue(), SDLoc(Op), MVT::i64);
26054 // FIXME gcc accepts some relocatable values here too, but only in certain
26055 // memory models; it's complicated.
26060 // 32-bit unsigned value
26061 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
26062 if (ConstantInt::isValueValidForType(Type::getInt32Ty(*DAG.getContext()),
26063 C->getZExtValue())) {
26064 Result = DAG.getTargetConstant(C->getZExtValue(), SDLoc(Op),
26065 Op.getValueType());
26069 // FIXME gcc accepts some relocatable values here too, but only in certain
26070 // memory models; it's complicated.
26074 // Literal immediates are always ok.
26075 if (ConstantSDNode *CST = dyn_cast<ConstantSDNode>(Op)) {
26076 // Widen to 64 bits here to get it sign extended.
26077 Result = DAG.getTargetConstant(CST->getSExtValue(), SDLoc(Op), MVT::i64);
26081 // In any sort of PIC mode addresses need to be computed at runtime by
26082 // adding in a register or some sort of table lookup. These can't
26083 // be used as immediates.
26084 if (Subtarget->isPICStyleGOT() || Subtarget->isPICStyleStubPIC())
26087 // If we are in non-pic codegen mode, we allow the address of a global (with
26088 // an optional displacement) to be used with 'i'.
26089 GlobalAddressSDNode *GA = nullptr;
26090 int64_t Offset = 0;
26092 // Match either (GA), (GA+C), (GA+C1+C2), etc.
26094 if ((GA = dyn_cast<GlobalAddressSDNode>(Op))) {
26095 Offset += GA->getOffset();
26097 } else if (Op.getOpcode() == ISD::ADD) {
26098 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
26099 Offset += C->getZExtValue();
26100 Op = Op.getOperand(0);
26103 } else if (Op.getOpcode() == ISD::SUB) {
26104 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
26105 Offset += -C->getZExtValue();
26106 Op = Op.getOperand(0);
26111 // Otherwise, this isn't something we can handle, reject it.
26115 const GlobalValue *GV = GA->getGlobal();
26116 // If we require an extra load to get this address, as in PIC mode, we
26117 // can't accept it.
26118 if (isGlobalStubReference(
26119 Subtarget->ClassifyGlobalReference(GV, DAG.getTarget())))
26122 Result = DAG.getTargetGlobalAddress(GV, SDLoc(Op),
26123 GA->getValueType(0), Offset);
26128 if (Result.getNode()) {
26129 Ops.push_back(Result);
26132 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG);
26135 std::pair<unsigned, const TargetRegisterClass *>
26136 X86TargetLowering::getRegForInlineAsmConstraint(const TargetRegisterInfo *TRI,
26137 StringRef Constraint,
26139 // First, see if this is a constraint that directly corresponds to an LLVM
26141 if (Constraint.size() == 1) {
26142 // GCC Constraint Letters
26143 switch (Constraint[0]) {
26145 // TODO: Slight differences here in allocation order and leaving
26146 // RIP in the class. Do they matter any more here than they do
26147 // in the normal allocation?
26148 case 'q': // GENERAL_REGS in 64-bit mode, Q_REGS in 32-bit mode.
26149 if (Subtarget->is64Bit()) {
26150 if (VT == MVT::i32 || VT == MVT::f32)
26151 return std::make_pair(0U, &X86::GR32RegClass);
26152 if (VT == MVT::i16)
26153 return std::make_pair(0U, &X86::GR16RegClass);
26154 if (VT == MVT::i8 || VT == MVT::i1)
26155 return std::make_pair(0U, &X86::GR8RegClass);
26156 if (VT == MVT::i64 || VT == MVT::f64)
26157 return std::make_pair(0U, &X86::GR64RegClass);
26160 // 32-bit fallthrough
26161 case 'Q': // Q_REGS
26162 if (VT == MVT::i32 || VT == MVT::f32)
26163 return std::make_pair(0U, &X86::GR32_ABCDRegClass);
26164 if (VT == MVT::i16)
26165 return std::make_pair(0U, &X86::GR16_ABCDRegClass);
26166 if (VT == MVT::i8 || VT == MVT::i1)
26167 return std::make_pair(0U, &X86::GR8_ABCD_LRegClass);
26168 if (VT == MVT::i64)
26169 return std::make_pair(0U, &X86::GR64_ABCDRegClass);
26171 case 'r': // GENERAL_REGS
26172 case 'l': // INDEX_REGS
26173 if (VT == MVT::i8 || VT == MVT::i1)
26174 return std::make_pair(0U, &X86::GR8RegClass);
26175 if (VT == MVT::i16)
26176 return std::make_pair(0U, &X86::GR16RegClass);
26177 if (VT == MVT::i32 || VT == MVT::f32 || !Subtarget->is64Bit())
26178 return std::make_pair(0U, &X86::GR32RegClass);
26179 return std::make_pair(0U, &X86::GR64RegClass);
26180 case 'R': // LEGACY_REGS
26181 if (VT == MVT::i8 || VT == MVT::i1)
26182 return std::make_pair(0U, &X86::GR8_NOREXRegClass);
26183 if (VT == MVT::i16)
26184 return std::make_pair(0U, &X86::GR16_NOREXRegClass);
26185 if (VT == MVT::i32 || !Subtarget->is64Bit())
26186 return std::make_pair(0U, &X86::GR32_NOREXRegClass);
26187 return std::make_pair(0U, &X86::GR64_NOREXRegClass);
26188 case 'f': // FP Stack registers.
26189 // If SSE is enabled for this VT, use f80 to ensure the isel moves the
26190 // value to the correct fpstack register class.
26191 if (VT == MVT::f32 && !isScalarFPTypeInSSEReg(VT))
26192 return std::make_pair(0U, &X86::RFP32RegClass);
26193 if (VT == MVT::f64 && !isScalarFPTypeInSSEReg(VT))
26194 return std::make_pair(0U, &X86::RFP64RegClass);
26195 return std::make_pair(0U, &X86::RFP80RegClass);
26196 case 'y': // MMX_REGS if MMX allowed.
26197 if (!Subtarget->hasMMX()) break;
26198 return std::make_pair(0U, &X86::VR64RegClass);
26199 case 'Y': // SSE_REGS if SSE2 allowed
26200 if (!Subtarget->hasSSE2()) break;
26202 case 'x': // SSE_REGS if SSE1 allowed or AVX_REGS if AVX allowed
26203 if (!Subtarget->hasSSE1()) break;
26205 switch (VT.SimpleTy) {
26207 // Scalar SSE types.
26210 return std::make_pair(0U, &X86::FR32RegClass);
26213 return std::make_pair(0U, &X86::FR64RegClass);
26221 return std::make_pair(0U, &X86::VR128RegClass);
26229 return std::make_pair(0U, &X86::VR256RegClass);
26234 return std::make_pair(0U, &X86::VR512RegClass);
26240 // Use the default implementation in TargetLowering to convert the register
26241 // constraint into a member of a register class.
26242 std::pair<unsigned, const TargetRegisterClass*> Res;
26243 Res = TargetLowering::getRegForInlineAsmConstraint(TRI, Constraint, VT);
26245 // Not found as a standard register?
26247 // Map st(0) -> st(7) -> ST0
26248 if (Constraint.size() == 7 && Constraint[0] == '{' &&
26249 tolower(Constraint[1]) == 's' &&
26250 tolower(Constraint[2]) == 't' &&
26251 Constraint[3] == '(' &&
26252 (Constraint[4] >= '0' && Constraint[4] <= '7') &&
26253 Constraint[5] == ')' &&
26254 Constraint[6] == '}') {
26256 Res.first = X86::FP0+Constraint[4]-'0';
26257 Res.second = &X86::RFP80RegClass;
26261 // GCC allows "st(0)" to be called just plain "st".
26262 if (StringRef("{st}").equals_lower(Constraint)) {
26263 Res.first = X86::FP0;
26264 Res.second = &X86::RFP80RegClass;
26269 if (StringRef("{flags}").equals_lower(Constraint)) {
26270 Res.first = X86::EFLAGS;
26271 Res.second = &X86::CCRRegClass;
26275 // 'A' means EAX + EDX.
26276 if (Constraint == "A") {
26277 Res.first = X86::EAX;
26278 Res.second = &X86::GR32_ADRegClass;
26284 // Otherwise, check to see if this is a register class of the wrong value
26285 // type. For example, we want to map "{ax},i32" -> {eax}, we don't want it to
26286 // turn into {ax},{dx}.
26287 // MVT::Other is used to specify clobber names.
26288 if (Res.second->hasType(VT) || VT == MVT::Other)
26289 return Res; // Correct type already, nothing to do.
26291 // Get a matching integer of the correct size. i.e. "ax" with MVT::32 should
26292 // return "eax". This should even work for things like getting 64bit integer
26293 // registers when given an f64 type.
26294 const TargetRegisterClass *Class = Res.second;
26295 if (Class == &X86::GR8RegClass || Class == &X86::GR16RegClass ||
26296 Class == &X86::GR32RegClass || Class == &X86::GR64RegClass) {
26297 unsigned Size = VT.getSizeInBits();
26298 MVT::SimpleValueType SimpleTy = Size == 1 || Size == 8 ? MVT::i8
26299 : Size == 16 ? MVT::i16
26300 : Size == 32 ? MVT::i32
26301 : Size == 64 ? MVT::i64
26303 unsigned DestReg = getX86SubSuperRegisterOrZero(Res.first, SimpleTy);
26305 Res.first = DestReg;
26306 Res.second = SimpleTy == MVT::i8 ? &X86::GR8RegClass
26307 : SimpleTy == MVT::i16 ? &X86::GR16RegClass
26308 : SimpleTy == MVT::i32 ? &X86::GR32RegClass
26309 : &X86::GR64RegClass;
26310 assert(Res.second->contains(Res.first) && "Register in register class");
26312 // No register found/type mismatch.
26314 Res.second = nullptr;
26316 } else if (Class == &X86::FR32RegClass || Class == &X86::FR64RegClass ||
26317 Class == &X86::VR128RegClass || Class == &X86::VR256RegClass ||
26318 Class == &X86::FR32XRegClass || Class == &X86::FR64XRegClass ||
26319 Class == &X86::VR128XRegClass || Class == &X86::VR256XRegClass ||
26320 Class == &X86::VR512RegClass) {
26321 // Handle references to XMM physical registers that got mapped into the
26322 // wrong class. This can happen with constraints like {xmm0} where the
26323 // target independent register mapper will just pick the first match it can
26324 // find, ignoring the required type.
26326 if (VT == MVT::f32 || VT == MVT::i32)
26327 Res.second = &X86::FR32RegClass;
26328 else if (VT == MVT::f64 || VT == MVT::i64)
26329 Res.second = &X86::FR64RegClass;
26330 else if (X86::VR128RegClass.hasType(VT))
26331 Res.second = &X86::VR128RegClass;
26332 else if (X86::VR256RegClass.hasType(VT))
26333 Res.second = &X86::VR256RegClass;
26334 else if (X86::VR512RegClass.hasType(VT))
26335 Res.second = &X86::VR512RegClass;
26337 // Type mismatch and not a clobber: Return an error;
26339 Res.second = nullptr;
26346 int X86TargetLowering::getScalingFactorCost(const DataLayout &DL,
26347 const AddrMode &AM, Type *Ty,
26348 unsigned AS) const {
26349 // Scaling factors are not free at all.
26350 // An indexed folded instruction, i.e., inst (reg1, reg2, scale),
26351 // will take 2 allocations in the out of order engine instead of 1
26352 // for plain addressing mode, i.e. inst (reg1).
26354 // vaddps (%rsi,%drx), %ymm0, %ymm1
26355 // Requires two allocations (one for the load, one for the computation)
26357 // vaddps (%rsi), %ymm0, %ymm1
26358 // Requires just 1 allocation, i.e., freeing allocations for other operations
26359 // and having less micro operations to execute.
26361 // For some X86 architectures, this is even worse because for instance for
26362 // stores, the complex addressing mode forces the instruction to use the
26363 // "load" ports instead of the dedicated "store" port.
26364 // E.g., on Haswell:
26365 // vmovaps %ymm1, (%r8, %rdi) can use port 2 or 3.
26366 // vmovaps %ymm1, (%r8) can use port 2, 3, or 7.
26367 if (isLegalAddressingMode(DL, AM, Ty, AS))
26368 // Scale represents reg2 * scale, thus account for 1
26369 // as soon as we use a second register.
26370 return AM.Scale != 0;
26374 bool X86TargetLowering::isTargetFTOL() const {
26375 return Subtarget->isTargetKnownWindowsMSVC() && !Subtarget->is64Bit();