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 // Attempt to lower to a bitmask if we can. VPAND is faster than VPBLENDVB.
6667 if (SDValue Masked = lowerVectorShuffleAsBitMask(DL, VT, V1, V2, Mask, DAG))
6670 // Scale the blend by the number of bytes per element.
6671 int Scale = VT.getScalarSizeInBits() / 8;
6673 // This form of blend is always done on bytes. Compute the byte vector
6675 MVT BlendVT = MVT::getVectorVT(MVT::i8, VT.getSizeInBits() / 8);
6677 // Compute the VSELECT mask. Note that VSELECT is really confusing in the
6678 // mix of LLVM's code generator and the x86 backend. We tell the code
6679 // generator that boolean values in the elements of an x86 vector register
6680 // are -1 for true and 0 for false. We then use the LLVM semantics of 'true'
6681 // mapping a select to operand #1, and 'false' mapping to operand #2. The
6682 // reality in x86 is that vector masks (pre-AVX-512) use only the high bit
6683 // of the element (the remaining are ignored) and 0 in that high bit would
6684 // mean operand #1 while 1 in the high bit would mean operand #2. So while
6685 // the LLVM model for boolean values in vector elements gets the relevant
6686 // bit set, it is set backwards and over constrained relative to x86's
6688 SmallVector<SDValue, 32> VSELECTMask;
6689 for (int i = 0, Size = Mask.size(); i < Size; ++i)
6690 for (int j = 0; j < Scale; ++j)
6691 VSELECTMask.push_back(
6692 Mask[i] < 0 ? DAG.getUNDEF(MVT::i8)
6693 : DAG.getConstant(Mask[i] < Size ? -1 : 0, DL,
6696 V1 = DAG.getBitcast(BlendVT, V1);
6697 V2 = DAG.getBitcast(BlendVT, V2);
6698 return DAG.getBitcast(VT, DAG.getNode(ISD::VSELECT, DL, BlendVT,
6699 DAG.getNode(ISD::BUILD_VECTOR, DL,
6700 BlendVT, VSELECTMask),
6705 llvm_unreachable("Not a supported integer vector type!");
6709 /// \brief Try to lower as a blend of elements from two inputs followed by
6710 /// a single-input permutation.
6712 /// This matches the pattern where we can blend elements from two inputs and
6713 /// then reduce the shuffle to a single-input permutation.
6714 static SDValue lowerVectorShuffleAsBlendAndPermute(SDLoc DL, MVT VT, SDValue V1,
6717 SelectionDAG &DAG) {
6718 // We build up the blend mask while checking whether a blend is a viable way
6719 // to reduce the shuffle.
6720 SmallVector<int, 32> BlendMask(Mask.size(), -1);
6721 SmallVector<int, 32> PermuteMask(Mask.size(), -1);
6723 for (int i = 0, Size = Mask.size(); i < Size; ++i) {
6727 assert(Mask[i] < Size * 2 && "Shuffle input is out of bounds.");
6729 if (BlendMask[Mask[i] % Size] == -1)
6730 BlendMask[Mask[i] % Size] = Mask[i];
6731 else if (BlendMask[Mask[i] % Size] != Mask[i])
6732 return SDValue(); // Can't blend in the needed input!
6734 PermuteMask[i] = Mask[i] % Size;
6737 SDValue V = DAG.getVectorShuffle(VT, DL, V1, V2, BlendMask);
6738 return DAG.getVectorShuffle(VT, DL, V, DAG.getUNDEF(VT), PermuteMask);
6741 /// \brief Generic routine to decompose a shuffle and blend into indepndent
6742 /// blends and permutes.
6744 /// This matches the extremely common pattern for handling combined
6745 /// shuffle+blend operations on newer X86 ISAs where we have very fast blend
6746 /// operations. It will try to pick the best arrangement of shuffles and
6748 static SDValue lowerVectorShuffleAsDecomposedShuffleBlend(SDLoc DL, MVT VT,
6752 SelectionDAG &DAG) {
6753 // Shuffle the input elements into the desired positions in V1 and V2 and
6754 // blend them together.
6755 SmallVector<int, 32> V1Mask(Mask.size(), -1);
6756 SmallVector<int, 32> V2Mask(Mask.size(), -1);
6757 SmallVector<int, 32> BlendMask(Mask.size(), -1);
6758 for (int i = 0, Size = Mask.size(); i < Size; ++i)
6759 if (Mask[i] >= 0 && Mask[i] < Size) {
6760 V1Mask[i] = Mask[i];
6762 } else if (Mask[i] >= Size) {
6763 V2Mask[i] = Mask[i] - Size;
6764 BlendMask[i] = i + Size;
6767 // Try to lower with the simpler initial blend strategy unless one of the
6768 // input shuffles would be a no-op. We prefer to shuffle inputs as the
6769 // shuffle may be able to fold with a load or other benefit. However, when
6770 // we'll have to do 2x as many shuffles in order to achieve this, blending
6771 // first is a better strategy.
6772 if (!isNoopShuffleMask(V1Mask) && !isNoopShuffleMask(V2Mask))
6773 if (SDValue BlendPerm =
6774 lowerVectorShuffleAsBlendAndPermute(DL, VT, V1, V2, Mask, DAG))
6777 V1 = DAG.getVectorShuffle(VT, DL, V1, DAG.getUNDEF(VT), V1Mask);
6778 V2 = DAG.getVectorShuffle(VT, DL, V2, DAG.getUNDEF(VT), V2Mask);
6779 return DAG.getVectorShuffle(VT, DL, V1, V2, BlendMask);
6782 /// \brief Try to lower a vector shuffle as a byte rotation.
6784 /// SSSE3 has a generic PALIGNR instruction in x86 that will do an arbitrary
6785 /// byte-rotation of the concatenation of two vectors; pre-SSSE3 can use
6786 /// a PSRLDQ/PSLLDQ/POR pattern to get a similar effect. This routine will
6787 /// try to generically lower a vector shuffle through such an pattern. It
6788 /// does not check for the profitability of lowering either as PALIGNR or
6789 /// PSRLDQ/PSLLDQ/POR, only whether the mask is valid to lower in that form.
6790 /// This matches shuffle vectors that look like:
6792 /// v8i16 [11, 12, 13, 14, 15, 0, 1, 2]
6794 /// Essentially it concatenates V1 and V2, shifts right by some number of
6795 /// elements, and takes the low elements as the result. Note that while this is
6796 /// specified as a *right shift* because x86 is little-endian, it is a *left
6797 /// rotate* of the vector lanes.
6798 static SDValue lowerVectorShuffleAsByteRotate(SDLoc DL, MVT VT, SDValue V1,
6801 const X86Subtarget *Subtarget,
6802 SelectionDAG &DAG) {
6803 assert(!isNoopShuffleMask(Mask) && "We shouldn't lower no-op shuffles!");
6805 int NumElts = Mask.size();
6806 int NumLanes = VT.getSizeInBits() / 128;
6807 int NumLaneElts = NumElts / NumLanes;
6809 // We need to detect various ways of spelling a rotation:
6810 // [11, 12, 13, 14, 15, 0, 1, 2]
6811 // [-1, 12, 13, 14, -1, -1, 1, -1]
6812 // [-1, -1, -1, -1, -1, -1, 1, 2]
6813 // [ 3, 4, 5, 6, 7, 8, 9, 10]
6814 // [-1, 4, 5, 6, -1, -1, 9, -1]
6815 // [-1, 4, 5, 6, -1, -1, -1, -1]
6818 for (int l = 0; l < NumElts; l += NumLaneElts) {
6819 for (int i = 0; i < NumLaneElts; ++i) {
6820 if (Mask[l + i] == -1)
6822 assert(Mask[l + i] >= 0 && "Only -1 is a valid negative mask element!");
6824 // Get the mod-Size index and lane correct it.
6825 int LaneIdx = (Mask[l + i] % NumElts) - l;
6826 // Make sure it was in this lane.
6827 if (LaneIdx < 0 || LaneIdx >= NumLaneElts)
6830 // Determine where a rotated vector would have started.
6831 int StartIdx = i - LaneIdx;
6833 // The identity rotation isn't interesting, stop.
6836 // If we found the tail of a vector the rotation must be the missing
6837 // front. If we found the head of a vector, it must be how much of the
6839 int CandidateRotation = StartIdx < 0 ? -StartIdx : NumLaneElts - StartIdx;
6842 Rotation = CandidateRotation;
6843 else if (Rotation != CandidateRotation)
6844 // The rotations don't match, so we can't match this mask.
6847 // Compute which value this mask is pointing at.
6848 SDValue MaskV = Mask[l + i] < NumElts ? V1 : V2;
6850 // Compute which of the two target values this index should be assigned
6851 // to. This reflects whether the high elements are remaining or the low
6852 // elements are remaining.
6853 SDValue &TargetV = StartIdx < 0 ? Hi : Lo;
6855 // Either set up this value if we've not encountered it before, or check
6856 // that it remains consistent.
6859 else if (TargetV != MaskV)
6860 // This may be a rotation, but it pulls from the inputs in some
6861 // unsupported interleaving.
6866 // Check that we successfully analyzed the mask, and normalize the results.
6867 assert(Rotation != 0 && "Failed to locate a viable rotation!");
6868 assert((Lo || Hi) && "Failed to find a rotated input vector!");
6874 // The actual rotate instruction rotates bytes, so we need to scale the
6875 // rotation based on how many bytes are in the vector lane.
6876 int Scale = 16 / NumLaneElts;
6878 // SSSE3 targets can use the palignr instruction.
6879 if (Subtarget->hasSSSE3()) {
6880 // Cast the inputs to i8 vector of correct length to match PALIGNR.
6881 MVT AlignVT = MVT::getVectorVT(MVT::i8, 16 * NumLanes);
6882 Lo = DAG.getBitcast(AlignVT, Lo);
6883 Hi = DAG.getBitcast(AlignVT, Hi);
6885 return DAG.getBitcast(
6886 VT, DAG.getNode(X86ISD::PALIGNR, DL, AlignVT, Hi, Lo,
6887 DAG.getConstant(Rotation * Scale, DL, MVT::i8)));
6890 assert(VT.getSizeInBits() == 128 &&
6891 "Rotate-based lowering only supports 128-bit lowering!");
6892 assert(Mask.size() <= 16 &&
6893 "Can shuffle at most 16 bytes in a 128-bit vector!");
6895 // Default SSE2 implementation
6896 int LoByteShift = 16 - Rotation * Scale;
6897 int HiByteShift = Rotation * Scale;
6899 // Cast the inputs to v2i64 to match PSLLDQ/PSRLDQ.
6900 Lo = DAG.getBitcast(MVT::v2i64, Lo);
6901 Hi = DAG.getBitcast(MVT::v2i64, Hi);
6903 SDValue LoShift = DAG.getNode(X86ISD::VSHLDQ, DL, MVT::v2i64, Lo,
6904 DAG.getConstant(LoByteShift, DL, MVT::i8));
6905 SDValue HiShift = DAG.getNode(X86ISD::VSRLDQ, DL, MVT::v2i64, Hi,
6906 DAG.getConstant(HiByteShift, DL, MVT::i8));
6907 return DAG.getBitcast(VT,
6908 DAG.getNode(ISD::OR, DL, MVT::v2i64, LoShift, HiShift));
6911 /// \brief Try to lower a vector shuffle as a bit shift (shifts in zeros).
6913 /// Attempts to match a shuffle mask against the PSLL(W/D/Q/DQ) and
6914 /// PSRL(W/D/Q/DQ) SSE2 and AVX2 logical bit-shift instructions. The function
6915 /// matches elements from one of the input vectors shuffled to the left or
6916 /// right with zeroable elements 'shifted in'. It handles both the strictly
6917 /// bit-wise element shifts and the byte shift across an entire 128-bit double
6920 /// PSHL : (little-endian) left bit shift.
6921 /// [ zz, 0, zz, 2 ]
6922 /// [ -1, 4, zz, -1 ]
6923 /// PSRL : (little-endian) right bit shift.
6925 /// [ -1, -1, 7, zz]
6926 /// PSLLDQ : (little-endian) left byte shift
6927 /// [ zz, 0, 1, 2, 3, 4, 5, 6]
6928 /// [ zz, zz, -1, -1, 2, 3, 4, -1]
6929 /// [ zz, zz, zz, zz, zz, zz, -1, 1]
6930 /// PSRLDQ : (little-endian) right byte shift
6931 /// [ 5, 6, 7, zz, zz, zz, zz, zz]
6932 /// [ -1, 5, 6, 7, zz, zz, zz, zz]
6933 /// [ 1, 2, -1, -1, -1, -1, zz, zz]
6934 static SDValue lowerVectorShuffleAsShift(SDLoc DL, MVT VT, SDValue V1,
6935 SDValue V2, ArrayRef<int> Mask,
6936 SelectionDAG &DAG) {
6937 SmallBitVector Zeroable = computeZeroableShuffleElements(Mask, V1, V2);
6939 int Size = Mask.size();
6940 assert(Size == (int)VT.getVectorNumElements() && "Unexpected mask size");
6942 auto CheckZeros = [&](int Shift, int Scale, bool Left) {
6943 for (int i = 0; i < Size; i += Scale)
6944 for (int j = 0; j < Shift; ++j)
6945 if (!Zeroable[i + j + (Left ? 0 : (Scale - Shift))])
6951 auto MatchShift = [&](int Shift, int Scale, bool Left, SDValue V) {
6952 for (int i = 0; i != Size; i += Scale) {
6953 unsigned Pos = Left ? i + Shift : i;
6954 unsigned Low = Left ? i : i + Shift;
6955 unsigned Len = Scale - Shift;
6956 if (!isSequentialOrUndefInRange(Mask, Pos, Len,
6957 Low + (V == V1 ? 0 : Size)))
6961 int ShiftEltBits = VT.getScalarSizeInBits() * Scale;
6962 bool ByteShift = ShiftEltBits > 64;
6963 unsigned OpCode = Left ? (ByteShift ? X86ISD::VSHLDQ : X86ISD::VSHLI)
6964 : (ByteShift ? X86ISD::VSRLDQ : X86ISD::VSRLI);
6965 int ShiftAmt = Shift * VT.getScalarSizeInBits() / (ByteShift ? 8 : 1);
6967 // Normalize the scale for byte shifts to still produce an i64 element
6969 Scale = ByteShift ? Scale / 2 : Scale;
6971 // We need to round trip through the appropriate type for the shift.
6972 MVT ShiftSVT = MVT::getIntegerVT(VT.getScalarSizeInBits() * Scale);
6973 MVT ShiftVT = MVT::getVectorVT(ShiftSVT, Size / Scale);
6974 assert(DAG.getTargetLoweringInfo().isTypeLegal(ShiftVT) &&
6975 "Illegal integer vector type");
6976 V = DAG.getBitcast(ShiftVT, V);
6978 V = DAG.getNode(OpCode, DL, ShiftVT, V,
6979 DAG.getConstant(ShiftAmt, DL, MVT::i8));
6980 return DAG.getBitcast(VT, V);
6983 // SSE/AVX supports logical shifts up to 64-bit integers - so we can just
6984 // keep doubling the size of the integer elements up to that. We can
6985 // then shift the elements of the integer vector by whole multiples of
6986 // their width within the elements of the larger integer vector. Test each
6987 // multiple to see if we can find a match with the moved element indices
6988 // and that the shifted in elements are all zeroable.
6989 for (int Scale = 2; Scale * VT.getScalarSizeInBits() <= 128; Scale *= 2)
6990 for (int Shift = 1; Shift != Scale; ++Shift)
6991 for (bool Left : {true, false})
6992 if (CheckZeros(Shift, Scale, Left))
6993 for (SDValue V : {V1, V2})
6994 if (SDValue Match = MatchShift(Shift, Scale, Left, V))
7001 /// \brief Try to lower a vector shuffle using SSE4a EXTRQ/INSERTQ.
7002 static SDValue lowerVectorShuffleWithSSE4A(SDLoc DL, MVT VT, SDValue V1,
7003 SDValue V2, ArrayRef<int> Mask,
7004 SelectionDAG &DAG) {
7005 SmallBitVector Zeroable = computeZeroableShuffleElements(Mask, V1, V2);
7006 assert(!Zeroable.all() && "Fully zeroable shuffle mask");
7008 int Size = Mask.size();
7009 int HalfSize = Size / 2;
7010 assert(Size == (int)VT.getVectorNumElements() && "Unexpected mask size");
7012 // Upper half must be undefined.
7013 if (!isUndefInRange(Mask, HalfSize, HalfSize))
7016 // EXTRQ: Extract Len elements from lower half of source, starting at Idx.
7017 // Remainder of lower half result is zero and upper half is all undef.
7018 auto LowerAsEXTRQ = [&]() {
7019 // Determine the extraction length from the part of the
7020 // lower half that isn't zeroable.
7022 for (; Len >= 0; --Len)
7023 if (!Zeroable[Len - 1])
7025 assert(Len > 0 && "Zeroable shuffle mask");
7027 // Attempt to match first Len sequential elements from the lower half.
7030 for (int i = 0; i != Len; ++i) {
7034 SDValue &V = (M < Size ? V1 : V2);
7037 // All mask elements must be in the lower half.
7041 if (Idx < 0 || (Src == V && Idx == (M - i))) {
7052 assert((Idx + Len) <= HalfSize && "Illegal extraction mask");
7053 int BitLen = (Len * VT.getScalarSizeInBits()) & 0x3f;
7054 int BitIdx = (Idx * VT.getScalarSizeInBits()) & 0x3f;
7055 return DAG.getNode(X86ISD::EXTRQI, DL, VT, Src,
7056 DAG.getConstant(BitLen, DL, MVT::i8),
7057 DAG.getConstant(BitIdx, DL, MVT::i8));
7060 if (SDValue ExtrQ = LowerAsEXTRQ())
7063 // INSERTQ: Extract lowest Len elements from lower half of second source and
7064 // insert over first source, starting at Idx.
7065 // { A[0], .., A[Idx-1], B[0], .., B[Len-1], A[Idx+Len], .., UNDEF, ... }
7066 auto LowerAsInsertQ = [&]() {
7067 for (int Idx = 0; Idx != HalfSize; ++Idx) {
7070 // Attempt to match first source from mask before insertion point.
7071 if (isUndefInRange(Mask, 0, Idx)) {
7073 } else if (isSequentialOrUndefInRange(Mask, 0, Idx, 0)) {
7075 } else if (isSequentialOrUndefInRange(Mask, 0, Idx, Size)) {
7081 // Extend the extraction length looking to match both the insertion of
7082 // the second source and the remaining elements of the first.
7083 for (int Hi = Idx + 1; Hi <= HalfSize; ++Hi) {
7088 if (isSequentialOrUndefInRange(Mask, Idx, Len, 0)) {
7090 } else if (isSequentialOrUndefInRange(Mask, Idx, Len, Size)) {
7096 // Match the remaining elements of the lower half.
7097 if (isUndefInRange(Mask, Hi, HalfSize - Hi)) {
7099 } else if ((!Base || (Base == V1)) &&
7100 isSequentialOrUndefInRange(Mask, Hi, HalfSize - Hi, Hi)) {
7102 } else if ((!Base || (Base == V2)) &&
7103 isSequentialOrUndefInRange(Mask, Hi, HalfSize - Hi,
7110 // We may not have a base (first source) - this can safely be undefined.
7112 Base = DAG.getUNDEF(VT);
7114 int BitLen = (Len * VT.getScalarSizeInBits()) & 0x3f;
7115 int BitIdx = (Idx * VT.getScalarSizeInBits()) & 0x3f;
7116 return DAG.getNode(X86ISD::INSERTQI, DL, VT, Base, Insert,
7117 DAG.getConstant(BitLen, DL, MVT::i8),
7118 DAG.getConstant(BitIdx, DL, MVT::i8));
7125 if (SDValue InsertQ = LowerAsInsertQ())
7131 /// \brief Lower a vector shuffle as a zero or any extension.
7133 /// Given a specific number of elements, element bit width, and extension
7134 /// stride, produce either a zero or any extension based on the available
7135 /// features of the subtarget.
7136 static SDValue lowerVectorShuffleAsSpecificZeroOrAnyExtend(
7137 SDLoc DL, MVT VT, int Scale, bool AnyExt, SDValue InputV,
7138 ArrayRef<int> Mask, const X86Subtarget *Subtarget, SelectionDAG &DAG) {
7139 assert(Scale > 1 && "Need a scale to extend.");
7140 int NumElements = VT.getVectorNumElements();
7141 int EltBits = VT.getScalarSizeInBits();
7142 assert((EltBits == 8 || EltBits == 16 || EltBits == 32) &&
7143 "Only 8, 16, and 32 bit elements can be extended.");
7144 assert(Scale * EltBits <= 64 && "Cannot zero extend past 64 bits.");
7146 // Found a valid zext mask! Try various lowering strategies based on the
7147 // input type and available ISA extensions.
7148 if (Subtarget->hasSSE41()) {
7149 MVT ExtVT = MVT::getVectorVT(MVT::getIntegerVT(EltBits * Scale),
7150 NumElements / Scale);
7151 return DAG.getBitcast(VT, DAG.getNode(X86ISD::VZEXT, DL, ExtVT, InputV));
7154 // For any extends we can cheat for larger element sizes and use shuffle
7155 // instructions that can fold with a load and/or copy.
7156 if (AnyExt && EltBits == 32) {
7157 int PSHUFDMask[4] = {0, -1, 1, -1};
7158 return DAG.getBitcast(
7159 VT, DAG.getNode(X86ISD::PSHUFD, DL, MVT::v4i32,
7160 DAG.getBitcast(MVT::v4i32, InputV),
7161 getV4X86ShuffleImm8ForMask(PSHUFDMask, DL, DAG)));
7163 if (AnyExt && EltBits == 16 && Scale > 2) {
7164 int PSHUFDMask[4] = {0, -1, 0, -1};
7165 InputV = DAG.getNode(X86ISD::PSHUFD, DL, MVT::v4i32,
7166 DAG.getBitcast(MVT::v4i32, InputV),
7167 getV4X86ShuffleImm8ForMask(PSHUFDMask, DL, DAG));
7168 int PSHUFHWMask[4] = {1, -1, -1, -1};
7169 return DAG.getBitcast(
7170 VT, DAG.getNode(X86ISD::PSHUFHW, DL, MVT::v8i16,
7171 DAG.getBitcast(MVT::v8i16, InputV),
7172 getV4X86ShuffleImm8ForMask(PSHUFHWMask, DL, DAG)));
7175 // The SSE4A EXTRQ instruction can efficiently extend the first 2 lanes
7177 if ((Scale * EltBits) == 64 && EltBits < 32 && Subtarget->hasSSE4A()) {
7178 assert(NumElements == (int)Mask.size() && "Unexpected shuffle mask size!");
7179 assert(VT.getSizeInBits() == 128 && "Unexpected vector width!");
7181 SDValue Lo = DAG.getNode(ISD::BITCAST, DL, MVT::v2i64,
7182 DAG.getNode(X86ISD::EXTRQI, DL, VT, InputV,
7183 DAG.getConstant(EltBits, DL, MVT::i8),
7184 DAG.getConstant(0, DL, MVT::i8)));
7185 if (isUndefInRange(Mask, NumElements/2, NumElements/2))
7186 return DAG.getNode(ISD::BITCAST, DL, VT, Lo);
7189 DAG.getNode(ISD::BITCAST, DL, MVT::v2i64,
7190 DAG.getNode(X86ISD::EXTRQI, DL, VT, InputV,
7191 DAG.getConstant(EltBits, DL, MVT::i8),
7192 DAG.getConstant(EltBits, DL, MVT::i8)));
7193 return DAG.getNode(ISD::BITCAST, DL, VT,
7194 DAG.getNode(X86ISD::UNPCKL, DL, MVT::v2i64, Lo, Hi));
7197 // If this would require more than 2 unpack instructions to expand, use
7198 // pshufb when available. We can only use more than 2 unpack instructions
7199 // when zero extending i8 elements which also makes it easier to use pshufb.
7200 if (Scale > 4 && EltBits == 8 && Subtarget->hasSSSE3()) {
7201 assert(NumElements == 16 && "Unexpected byte vector width!");
7202 SDValue PSHUFBMask[16];
7203 for (int i = 0; i < 16; ++i)
7205 DAG.getConstant((i % Scale == 0) ? i / Scale : 0x80, DL, MVT::i8);
7206 InputV = DAG.getBitcast(MVT::v16i8, InputV);
7207 return DAG.getBitcast(VT,
7208 DAG.getNode(X86ISD::PSHUFB, DL, MVT::v16i8, InputV,
7209 DAG.getNode(ISD::BUILD_VECTOR, DL,
7210 MVT::v16i8, PSHUFBMask)));
7213 // Otherwise emit a sequence of unpacks.
7215 MVT InputVT = MVT::getVectorVT(MVT::getIntegerVT(EltBits), NumElements);
7216 SDValue Ext = AnyExt ? DAG.getUNDEF(InputVT)
7217 : getZeroVector(InputVT, Subtarget, DAG, DL);
7218 InputV = DAG.getBitcast(InputVT, InputV);
7219 InputV = DAG.getNode(X86ISD::UNPCKL, DL, InputVT, InputV, Ext);
7223 } while (Scale > 1);
7224 return DAG.getBitcast(VT, InputV);
7227 /// \brief Try to lower a vector shuffle as a zero extension on any microarch.
7229 /// This routine will try to do everything in its power to cleverly lower
7230 /// a shuffle which happens to match the pattern of a zero extend. It doesn't
7231 /// check for the profitability of this lowering, it tries to aggressively
7232 /// match this pattern. It will use all of the micro-architectural details it
7233 /// can to emit an efficient lowering. It handles both blends with all-zero
7234 /// inputs to explicitly zero-extend and undef-lanes (sometimes undef due to
7235 /// masking out later).
7237 /// The reason we have dedicated lowering for zext-style shuffles is that they
7238 /// are both incredibly common and often quite performance sensitive.
7239 static SDValue lowerVectorShuffleAsZeroOrAnyExtend(
7240 SDLoc DL, MVT VT, SDValue V1, SDValue V2, ArrayRef<int> Mask,
7241 const X86Subtarget *Subtarget, SelectionDAG &DAG) {
7242 SmallBitVector Zeroable = computeZeroableShuffleElements(Mask, V1, V2);
7244 int Bits = VT.getSizeInBits();
7245 int NumElements = VT.getVectorNumElements();
7246 assert(VT.getScalarSizeInBits() <= 32 &&
7247 "Exceeds 32-bit integer zero extension limit");
7248 assert((int)Mask.size() == NumElements && "Unexpected shuffle mask size");
7250 // Define a helper function to check a particular ext-scale and lower to it if
7252 auto Lower = [&](int Scale) -> SDValue {
7255 for (int i = 0; i < NumElements; ++i) {
7257 continue; // Valid anywhere but doesn't tell us anything.
7258 if (i % Scale != 0) {
7259 // Each of the extended elements need to be zeroable.
7263 // We no longer are in the anyext case.
7268 // Each of the base elements needs to be consecutive indices into the
7269 // same input vector.
7270 SDValue V = Mask[i] < NumElements ? V1 : V2;
7273 else if (InputV != V)
7274 return SDValue(); // Flip-flopping inputs.
7276 if (Mask[i] % NumElements != i / Scale)
7277 return SDValue(); // Non-consecutive strided elements.
7280 // If we fail to find an input, we have a zero-shuffle which should always
7281 // have already been handled.
7282 // FIXME: Maybe handle this here in case during blending we end up with one?
7286 return lowerVectorShuffleAsSpecificZeroOrAnyExtend(
7287 DL, VT, Scale, AnyExt, InputV, Mask, Subtarget, DAG);
7290 // The widest scale possible for extending is to a 64-bit integer.
7291 assert(Bits % 64 == 0 &&
7292 "The number of bits in a vector must be divisible by 64 on x86!");
7293 int NumExtElements = Bits / 64;
7295 // Each iteration, try extending the elements half as much, but into twice as
7297 for (; NumExtElements < NumElements; NumExtElements *= 2) {
7298 assert(NumElements % NumExtElements == 0 &&
7299 "The input vector size must be divisible by the extended size.");
7300 if (SDValue V = Lower(NumElements / NumExtElements))
7304 // General extends failed, but 128-bit vectors may be able to use MOVQ.
7308 // Returns one of the source operands if the shuffle can be reduced to a
7309 // MOVQ, copying the lower 64-bits and zero-extending to the upper 64-bits.
7310 auto CanZExtLowHalf = [&]() {
7311 for (int i = NumElements / 2; i != NumElements; ++i)
7314 if (isSequentialOrUndefInRange(Mask, 0, NumElements / 2, 0))
7316 if (isSequentialOrUndefInRange(Mask, 0, NumElements / 2, NumElements))
7321 if (SDValue V = CanZExtLowHalf()) {
7322 V = DAG.getBitcast(MVT::v2i64, V);
7323 V = DAG.getNode(X86ISD::VZEXT_MOVL, DL, MVT::v2i64, V);
7324 return DAG.getBitcast(VT, V);
7327 // No viable ext lowering found.
7331 /// \brief Try to get a scalar value for a specific element of a vector.
7333 /// Looks through BUILD_VECTOR and SCALAR_TO_VECTOR nodes to find a scalar.
7334 static SDValue getScalarValueForVectorElement(SDValue V, int Idx,
7335 SelectionDAG &DAG) {
7336 MVT VT = V.getSimpleValueType();
7337 MVT EltVT = VT.getVectorElementType();
7338 while (V.getOpcode() == ISD::BITCAST)
7339 V = V.getOperand(0);
7340 // If the bitcasts shift the element size, we can't extract an equivalent
7342 MVT NewVT = V.getSimpleValueType();
7343 if (!NewVT.isVector() || NewVT.getScalarSizeInBits() != VT.getScalarSizeInBits())
7346 if (V.getOpcode() == ISD::BUILD_VECTOR ||
7347 (Idx == 0 && V.getOpcode() == ISD::SCALAR_TO_VECTOR)) {
7348 // Ensure the scalar operand is the same size as the destination.
7349 // FIXME: Add support for scalar truncation where possible.
7350 SDValue S = V.getOperand(Idx);
7351 if (EltVT.getSizeInBits() == S.getSimpleValueType().getSizeInBits())
7352 return DAG.getNode(ISD::BITCAST, SDLoc(V), EltVT, S);
7358 /// \brief Helper to test for a load that can be folded with x86 shuffles.
7360 /// This is particularly important because the set of instructions varies
7361 /// significantly based on whether the operand is a load or not.
7362 static bool isShuffleFoldableLoad(SDValue V) {
7363 while (V.getOpcode() == ISD::BITCAST)
7364 V = V.getOperand(0);
7366 return ISD::isNON_EXTLoad(V.getNode());
7369 /// \brief Try to lower insertion of a single element into a zero vector.
7371 /// This is a common pattern that we have especially efficient patterns to lower
7372 /// across all subtarget feature sets.
7373 static SDValue lowerVectorShuffleAsElementInsertion(
7374 SDLoc DL, MVT VT, SDValue V1, SDValue V2, ArrayRef<int> Mask,
7375 const X86Subtarget *Subtarget, SelectionDAG &DAG) {
7376 SmallBitVector Zeroable = computeZeroableShuffleElements(Mask, V1, V2);
7378 MVT EltVT = VT.getVectorElementType();
7380 int V2Index = std::find_if(Mask.begin(), Mask.end(),
7381 [&Mask](int M) { return M >= (int)Mask.size(); }) -
7383 bool IsV1Zeroable = true;
7384 for (int i = 0, Size = Mask.size(); i < Size; ++i)
7385 if (i != V2Index && !Zeroable[i]) {
7386 IsV1Zeroable = false;
7390 // Check for a single input from a SCALAR_TO_VECTOR node.
7391 // FIXME: All of this should be canonicalized into INSERT_VECTOR_ELT and
7392 // all the smarts here sunk into that routine. However, the current
7393 // lowering of BUILD_VECTOR makes that nearly impossible until the old
7394 // vector shuffle lowering is dead.
7395 SDValue V2S = getScalarValueForVectorElement(V2, Mask[V2Index] - Mask.size(),
7397 if (V2S && DAG.getTargetLoweringInfo().isTypeLegal(V2S.getValueType())) {
7398 // We need to zext the scalar if it is smaller than an i32.
7399 V2S = DAG.getBitcast(EltVT, V2S);
7400 if (EltVT == MVT::i8 || EltVT == MVT::i16) {
7401 // Using zext to expand a narrow element won't work for non-zero
7406 // Zero-extend directly to i32.
7408 V2S = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::i32, V2S);
7410 V2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, ExtVT, V2S);
7411 } else if (Mask[V2Index] != (int)Mask.size() || EltVT == MVT::i8 ||
7412 EltVT == MVT::i16) {
7413 // Either not inserting from the low element of the input or the input
7414 // element size is too small to use VZEXT_MOVL to clear the high bits.
7418 if (!IsV1Zeroable) {
7419 // If V1 can't be treated as a zero vector we have fewer options to lower
7420 // this. We can't support integer vectors or non-zero targets cheaply, and
7421 // the V1 elements can't be permuted in any way.
7422 assert(VT == ExtVT && "Cannot change extended type when non-zeroable!");
7423 if (!VT.isFloatingPoint() || V2Index != 0)
7425 SmallVector<int, 8> V1Mask(Mask.begin(), Mask.end());
7426 V1Mask[V2Index] = -1;
7427 if (!isNoopShuffleMask(V1Mask))
7429 // This is essentially a special case blend operation, but if we have
7430 // general purpose blend operations, they are always faster. Bail and let
7431 // the rest of the lowering handle these as blends.
7432 if (Subtarget->hasSSE41())
7435 // Otherwise, use MOVSD or MOVSS.
7436 assert((EltVT == MVT::f32 || EltVT == MVT::f64) &&
7437 "Only two types of floating point element types to handle!");
7438 return DAG.getNode(EltVT == MVT::f32 ? X86ISD::MOVSS : X86ISD::MOVSD, DL,
7442 // This lowering only works for the low element with floating point vectors.
7443 if (VT.isFloatingPoint() && V2Index != 0)
7446 V2 = DAG.getNode(X86ISD::VZEXT_MOVL, DL, ExtVT, V2);
7448 V2 = DAG.getBitcast(VT, V2);
7451 // If we have 4 or fewer lanes we can cheaply shuffle the element into
7452 // the desired position. Otherwise it is more efficient to do a vector
7453 // shift left. We know that we can do a vector shift left because all
7454 // the inputs are zero.
7455 if (VT.isFloatingPoint() || VT.getVectorNumElements() <= 4) {
7456 SmallVector<int, 4> V2Shuffle(Mask.size(), 1);
7457 V2Shuffle[V2Index] = 0;
7458 V2 = DAG.getVectorShuffle(VT, DL, V2, DAG.getUNDEF(VT), V2Shuffle);
7460 V2 = DAG.getBitcast(MVT::v2i64, V2);
7462 X86ISD::VSHLDQ, DL, MVT::v2i64, V2,
7463 DAG.getConstant(V2Index * EltVT.getSizeInBits() / 8, DL,
7464 DAG.getTargetLoweringInfo().getScalarShiftAmountTy(
7465 DAG.getDataLayout(), VT)));
7466 V2 = DAG.getBitcast(VT, V2);
7472 /// \brief Try to lower broadcast of a single element.
7474 /// For convenience, this code also bundles all of the subtarget feature set
7475 /// filtering. While a little annoying to re-dispatch on type here, there isn't
7476 /// a convenient way to factor it out.
7477 static SDValue lowerVectorShuffleAsBroadcast(SDLoc DL, MVT VT, SDValue V,
7479 const X86Subtarget *Subtarget,
7480 SelectionDAG &DAG) {
7481 if (!Subtarget->hasAVX())
7483 if (VT.isInteger() && !Subtarget->hasAVX2())
7486 // Check that the mask is a broadcast.
7487 int BroadcastIdx = -1;
7489 if (M >= 0 && BroadcastIdx == -1)
7491 else if (M >= 0 && M != BroadcastIdx)
7494 assert(BroadcastIdx < (int)Mask.size() && "We only expect to be called with "
7495 "a sorted mask where the broadcast "
7498 // Go up the chain of (vector) values to find a scalar load that we can
7499 // combine with the broadcast.
7501 switch (V.getOpcode()) {
7502 case ISD::CONCAT_VECTORS: {
7503 int OperandSize = Mask.size() / V.getNumOperands();
7504 V = V.getOperand(BroadcastIdx / OperandSize);
7505 BroadcastIdx %= OperandSize;
7509 case ISD::INSERT_SUBVECTOR: {
7510 SDValue VOuter = V.getOperand(0), VInner = V.getOperand(1);
7511 auto ConstantIdx = dyn_cast<ConstantSDNode>(V.getOperand(2));
7515 int BeginIdx = (int)ConstantIdx->getZExtValue();
7517 BeginIdx + (int)VInner.getValueType().getVectorNumElements();
7518 if (BroadcastIdx >= BeginIdx && BroadcastIdx < EndIdx) {
7519 BroadcastIdx -= BeginIdx;
7530 // Check if this is a broadcast of a scalar. We special case lowering
7531 // for scalars so that we can more effectively fold with loads.
7532 if (V.getOpcode() == ISD::BUILD_VECTOR ||
7533 (V.getOpcode() == ISD::SCALAR_TO_VECTOR && BroadcastIdx == 0)) {
7534 V = V.getOperand(BroadcastIdx);
7536 // If the scalar isn't a load, we can't broadcast from it in AVX1.
7537 // Only AVX2 has register broadcasts.
7538 if (!Subtarget->hasAVX2() && !isShuffleFoldableLoad(V))
7540 } else if (BroadcastIdx != 0 || !Subtarget->hasAVX2()) {
7541 // We can't broadcast from a vector register without AVX2, and we can only
7542 // broadcast from the zero-element of a vector register.
7546 return DAG.getNode(X86ISD::VBROADCAST, DL, VT, V);
7549 // Check for whether we can use INSERTPS to perform the shuffle. We only use
7550 // INSERTPS when the V1 elements are already in the correct locations
7551 // because otherwise we can just always use two SHUFPS instructions which
7552 // are much smaller to encode than a SHUFPS and an INSERTPS. We can also
7553 // perform INSERTPS if a single V1 element is out of place and all V2
7554 // elements are zeroable.
7555 static SDValue lowerVectorShuffleAsInsertPS(SDValue Op, SDValue V1, SDValue V2,
7557 SelectionDAG &DAG) {
7558 assert(Op.getSimpleValueType() == MVT::v4f32 && "Bad shuffle type!");
7559 assert(V1.getSimpleValueType() == MVT::v4f32 && "Bad operand type!");
7560 assert(V2.getSimpleValueType() == MVT::v4f32 && "Bad operand type!");
7561 assert(Mask.size() == 4 && "Unexpected mask size for v4 shuffle!");
7563 SmallBitVector Zeroable = computeZeroableShuffleElements(Mask, V1, V2);
7566 int V1DstIndex = -1;
7567 int V2DstIndex = -1;
7568 bool V1UsedInPlace = false;
7570 for (int i = 0; i < 4; ++i) {
7571 // Synthesize a zero mask from the zeroable elements (includes undefs).
7577 // Flag if we use any V1 inputs in place.
7579 V1UsedInPlace = true;
7583 // We can only insert a single non-zeroable element.
7584 if (V1DstIndex != -1 || V2DstIndex != -1)
7588 // V1 input out of place for insertion.
7591 // V2 input for insertion.
7596 // Don't bother if we have no (non-zeroable) element for insertion.
7597 if (V1DstIndex == -1 && V2DstIndex == -1)
7600 // Determine element insertion src/dst indices. The src index is from the
7601 // start of the inserted vector, not the start of the concatenated vector.
7602 unsigned V2SrcIndex = 0;
7603 if (V1DstIndex != -1) {
7604 // If we have a V1 input out of place, we use V1 as the V2 element insertion
7605 // and don't use the original V2 at all.
7606 V2SrcIndex = Mask[V1DstIndex];
7607 V2DstIndex = V1DstIndex;
7610 V2SrcIndex = Mask[V2DstIndex] - 4;
7613 // If no V1 inputs are used in place, then the result is created only from
7614 // the zero mask and the V2 insertion - so remove V1 dependency.
7616 V1 = DAG.getUNDEF(MVT::v4f32);
7618 unsigned InsertPSMask = V2SrcIndex << 6 | V2DstIndex << 4 | ZMask;
7619 assert((InsertPSMask & ~0xFFu) == 0 && "Invalid mask!");
7621 // Insert the V2 element into the desired position.
7623 return DAG.getNode(X86ISD::INSERTPS, DL, MVT::v4f32, V1, V2,
7624 DAG.getConstant(InsertPSMask, DL, MVT::i8));
7627 /// \brief Try to lower a shuffle as a permute of the inputs followed by an
7628 /// UNPCK instruction.
7630 /// This specifically targets cases where we end up with alternating between
7631 /// the two inputs, and so can permute them into something that feeds a single
7632 /// UNPCK instruction. Note that this routine only targets integer vectors
7633 /// because for floating point vectors we have a generalized SHUFPS lowering
7634 /// strategy that handles everything that doesn't *exactly* match an unpack,
7635 /// making this clever lowering unnecessary.
7636 static SDValue lowerVectorShuffleAsUnpack(SDLoc DL, MVT VT, SDValue V1,
7637 SDValue V2, ArrayRef<int> Mask,
7638 SelectionDAG &DAG) {
7639 assert(!VT.isFloatingPoint() &&
7640 "This routine only supports integer vectors.");
7641 assert(!isSingleInputShuffleMask(Mask) &&
7642 "This routine should only be used when blending two inputs.");
7643 assert(Mask.size() >= 2 && "Single element masks are invalid.");
7645 int Size = Mask.size();
7647 int NumLoInputs = std::count_if(Mask.begin(), Mask.end(), [Size](int M) {
7648 return M >= 0 && M % Size < Size / 2;
7650 int NumHiInputs = std::count_if(
7651 Mask.begin(), Mask.end(), [Size](int M) { return M % Size >= Size / 2; });
7653 bool UnpackLo = NumLoInputs >= NumHiInputs;
7655 auto TryUnpack = [&](MVT UnpackVT, int Scale) {
7656 SmallVector<int, 32> V1Mask(Mask.size(), -1);
7657 SmallVector<int, 32> V2Mask(Mask.size(), -1);
7659 for (int i = 0; i < Size; ++i) {
7663 // Each element of the unpack contains Scale elements from this mask.
7664 int UnpackIdx = i / Scale;
7666 // We only handle the case where V1 feeds the first slots of the unpack.
7667 // We rely on canonicalization to ensure this is the case.
7668 if ((UnpackIdx % 2 == 0) != (Mask[i] < Size))
7671 // Setup the mask for this input. The indexing is tricky as we have to
7672 // handle the unpack stride.
7673 SmallVectorImpl<int> &VMask = (UnpackIdx % 2 == 0) ? V1Mask : V2Mask;
7674 VMask[(UnpackIdx / 2) * Scale + i % Scale + (UnpackLo ? 0 : Size / 2)] =
7678 // If we will have to shuffle both inputs to use the unpack, check whether
7679 // we can just unpack first and shuffle the result. If so, skip this unpack.
7680 if ((NumLoInputs == 0 || NumHiInputs == 0) && !isNoopShuffleMask(V1Mask) &&
7681 !isNoopShuffleMask(V2Mask))
7684 // Shuffle the inputs into place.
7685 V1 = DAG.getVectorShuffle(VT, DL, V1, DAG.getUNDEF(VT), V1Mask);
7686 V2 = DAG.getVectorShuffle(VT, DL, V2, DAG.getUNDEF(VT), V2Mask);
7688 // Cast the inputs to the type we will use to unpack them.
7689 V1 = DAG.getBitcast(UnpackVT, V1);
7690 V2 = DAG.getBitcast(UnpackVT, V2);
7692 // Unpack the inputs and cast the result back to the desired type.
7693 return DAG.getBitcast(
7694 VT, DAG.getNode(UnpackLo ? X86ISD::UNPCKL : X86ISD::UNPCKH, DL,
7698 // We try each unpack from the largest to the smallest to try and find one
7699 // that fits this mask.
7700 int OrigNumElements = VT.getVectorNumElements();
7701 int OrigScalarSize = VT.getScalarSizeInBits();
7702 for (int ScalarSize = 64; ScalarSize >= OrigScalarSize; ScalarSize /= 2) {
7703 int Scale = ScalarSize / OrigScalarSize;
7704 int NumElements = OrigNumElements / Scale;
7705 MVT UnpackVT = MVT::getVectorVT(MVT::getIntegerVT(ScalarSize), NumElements);
7706 if (SDValue Unpack = TryUnpack(UnpackVT, Scale))
7710 // If none of the unpack-rooted lowerings worked (or were profitable) try an
7712 if (NumLoInputs == 0 || NumHiInputs == 0) {
7713 assert((NumLoInputs > 0 || NumHiInputs > 0) &&
7714 "We have to have *some* inputs!");
7715 int HalfOffset = NumLoInputs == 0 ? Size / 2 : 0;
7717 // FIXME: We could consider the total complexity of the permute of each
7718 // possible unpacking. Or at the least we should consider how many
7719 // half-crossings are created.
7720 // FIXME: We could consider commuting the unpacks.
7722 SmallVector<int, 32> PermMask;
7723 PermMask.assign(Size, -1);
7724 for (int i = 0; i < Size; ++i) {
7728 assert(Mask[i] % Size >= HalfOffset && "Found input from wrong half!");
7731 2 * ((Mask[i] % Size) - HalfOffset) + (Mask[i] < Size ? 0 : 1);
7733 return DAG.getVectorShuffle(
7734 VT, DL, DAG.getNode(NumLoInputs == 0 ? X86ISD::UNPCKH : X86ISD::UNPCKL,
7736 DAG.getUNDEF(VT), PermMask);
7742 /// \brief Handle lowering of 2-lane 64-bit floating point shuffles.
7744 /// This is the basis function for the 2-lane 64-bit shuffles as we have full
7745 /// support for floating point shuffles but not integer shuffles. These
7746 /// instructions will incur a domain crossing penalty on some chips though so
7747 /// it is better to avoid lowering through this for integer vectors where
7749 static SDValue lowerV2F64VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
7750 const X86Subtarget *Subtarget,
7751 SelectionDAG &DAG) {
7753 assert(Op.getSimpleValueType() == MVT::v2f64 && "Bad shuffle type!");
7754 assert(V1.getSimpleValueType() == MVT::v2f64 && "Bad operand type!");
7755 assert(V2.getSimpleValueType() == MVT::v2f64 && "Bad operand type!");
7756 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
7757 ArrayRef<int> Mask = SVOp->getMask();
7758 assert(Mask.size() == 2 && "Unexpected mask size for v2 shuffle!");
7760 if (isSingleInputShuffleMask(Mask)) {
7761 // Use low duplicate instructions for masks that match their pattern.
7762 if (Subtarget->hasSSE3())
7763 if (isShuffleEquivalent(V1, V2, Mask, {0, 0}))
7764 return DAG.getNode(X86ISD::MOVDDUP, DL, MVT::v2f64, V1);
7766 // Straight shuffle of a single input vector. Simulate this by using the
7767 // single input as both of the "inputs" to this instruction..
7768 unsigned SHUFPDMask = (Mask[0] == 1) | ((Mask[1] == 1) << 1);
7770 if (Subtarget->hasAVX()) {
7771 // If we have AVX, we can use VPERMILPS which will allow folding a load
7772 // into the shuffle.
7773 return DAG.getNode(X86ISD::VPERMILPI, DL, MVT::v2f64, V1,
7774 DAG.getConstant(SHUFPDMask, DL, MVT::i8));
7777 return DAG.getNode(X86ISD::SHUFP, DL, MVT::v2f64, V1, V1,
7778 DAG.getConstant(SHUFPDMask, DL, MVT::i8));
7780 assert(Mask[0] >= 0 && Mask[0] < 2 && "Non-canonicalized blend!");
7781 assert(Mask[1] >= 2 && "Non-canonicalized blend!");
7783 // If we have a single input, insert that into V1 if we can do so cheaply.
7784 if ((Mask[0] >= 2) + (Mask[1] >= 2) == 1) {
7785 if (SDValue Insertion = lowerVectorShuffleAsElementInsertion(
7786 DL, MVT::v2f64, V1, V2, Mask, Subtarget, DAG))
7788 // Try inverting the insertion since for v2 masks it is easy to do and we
7789 // can't reliably sort the mask one way or the other.
7790 int InverseMask[2] = {Mask[0] < 0 ? -1 : (Mask[0] ^ 2),
7791 Mask[1] < 0 ? -1 : (Mask[1] ^ 2)};
7792 if (SDValue Insertion = lowerVectorShuffleAsElementInsertion(
7793 DL, MVT::v2f64, V2, V1, InverseMask, Subtarget, DAG))
7797 // Try to use one of the special instruction patterns to handle two common
7798 // blend patterns if a zero-blend above didn't work.
7799 if (isShuffleEquivalent(V1, V2, Mask, {0, 3}) ||
7800 isShuffleEquivalent(V1, V2, Mask, {1, 3}))
7801 if (SDValue V1S = getScalarValueForVectorElement(V1, Mask[0], DAG))
7802 // We can either use a special instruction to load over the low double or
7803 // to move just the low double.
7805 isShuffleFoldableLoad(V1S) ? X86ISD::MOVLPD : X86ISD::MOVSD,
7807 DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, MVT::v2f64, V1S));
7809 if (Subtarget->hasSSE41())
7810 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v2f64, V1, V2, Mask,
7814 // Use dedicated unpack instructions for masks that match their pattern.
7815 if (isShuffleEquivalent(V1, V2, Mask, {0, 2}))
7816 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v2f64, V1, V2);
7817 if (isShuffleEquivalent(V1, V2, Mask, {1, 3}))
7818 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v2f64, V1, V2);
7820 unsigned SHUFPDMask = (Mask[0] == 1) | (((Mask[1] - 2) == 1) << 1);
7821 return DAG.getNode(X86ISD::SHUFP, DL, MVT::v2f64, V1, V2,
7822 DAG.getConstant(SHUFPDMask, DL, MVT::i8));
7825 /// \brief Handle lowering of 2-lane 64-bit integer shuffles.
7827 /// Tries to lower a 2-lane 64-bit shuffle using shuffle operations provided by
7828 /// the integer unit to minimize domain crossing penalties. However, for blends
7829 /// it falls back to the floating point shuffle operation with appropriate bit
7831 static SDValue lowerV2I64VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
7832 const X86Subtarget *Subtarget,
7833 SelectionDAG &DAG) {
7835 assert(Op.getSimpleValueType() == MVT::v2i64 && "Bad shuffle type!");
7836 assert(V1.getSimpleValueType() == MVT::v2i64 && "Bad operand type!");
7837 assert(V2.getSimpleValueType() == MVT::v2i64 && "Bad operand type!");
7838 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
7839 ArrayRef<int> Mask = SVOp->getMask();
7840 assert(Mask.size() == 2 && "Unexpected mask size for v2 shuffle!");
7842 if (isSingleInputShuffleMask(Mask)) {
7843 // Check for being able to broadcast a single element.
7844 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v2i64, V1,
7845 Mask, Subtarget, DAG))
7848 // Straight shuffle of a single input vector. For everything from SSE2
7849 // onward this has a single fast instruction with no scary immediates.
7850 // We have to map the mask as it is actually a v4i32 shuffle instruction.
7851 V1 = DAG.getBitcast(MVT::v4i32, V1);
7852 int WidenedMask[4] = {
7853 std::max(Mask[0], 0) * 2, std::max(Mask[0], 0) * 2 + 1,
7854 std::max(Mask[1], 0) * 2, std::max(Mask[1], 0) * 2 + 1};
7855 return DAG.getBitcast(
7857 DAG.getNode(X86ISD::PSHUFD, DL, MVT::v4i32, V1,
7858 getV4X86ShuffleImm8ForMask(WidenedMask, DL, DAG)));
7860 assert(Mask[0] != -1 && "No undef lanes in multi-input v2 shuffles!");
7861 assert(Mask[1] != -1 && "No undef lanes in multi-input v2 shuffles!");
7862 assert(Mask[0] < 2 && "We sort V1 to be the first input.");
7863 assert(Mask[1] >= 2 && "We sort V2 to be the second input.");
7865 // If we have a blend of two PACKUS operations an the blend aligns with the
7866 // low and half halves, we can just merge the PACKUS operations. This is
7867 // particularly important as it lets us merge shuffles that this routine itself
7869 auto GetPackNode = [](SDValue V) {
7870 while (V.getOpcode() == ISD::BITCAST)
7871 V = V.getOperand(0);
7873 return V.getOpcode() == X86ISD::PACKUS ? V : SDValue();
7875 if (SDValue V1Pack = GetPackNode(V1))
7876 if (SDValue V2Pack = GetPackNode(V2))
7877 return DAG.getBitcast(MVT::v2i64,
7878 DAG.getNode(X86ISD::PACKUS, DL, MVT::v16i8,
7879 Mask[0] == 0 ? V1Pack.getOperand(0)
7880 : V1Pack.getOperand(1),
7881 Mask[1] == 2 ? V2Pack.getOperand(0)
7882 : V2Pack.getOperand(1)));
7884 // Try to use shift instructions.
7886 lowerVectorShuffleAsShift(DL, MVT::v2i64, V1, V2, Mask, DAG))
7889 // When loading a scalar and then shuffling it into a vector we can often do
7890 // the insertion cheaply.
7891 if (SDValue Insertion = lowerVectorShuffleAsElementInsertion(
7892 DL, MVT::v2i64, V1, V2, Mask, Subtarget, DAG))
7894 // Try inverting the insertion since for v2 masks it is easy to do and we
7895 // can't reliably sort the mask one way or the other.
7896 int InverseMask[2] = {Mask[0] ^ 2, Mask[1] ^ 2};
7897 if (SDValue Insertion = lowerVectorShuffleAsElementInsertion(
7898 DL, MVT::v2i64, V2, V1, InverseMask, Subtarget, DAG))
7901 // We have different paths for blend lowering, but they all must use the
7902 // *exact* same predicate.
7903 bool IsBlendSupported = Subtarget->hasSSE41();
7904 if (IsBlendSupported)
7905 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v2i64, V1, V2, Mask,
7909 // Use dedicated unpack instructions for masks that match their pattern.
7910 if (isShuffleEquivalent(V1, V2, Mask, {0, 2}))
7911 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v2i64, V1, V2);
7912 if (isShuffleEquivalent(V1, V2, Mask, {1, 3}))
7913 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v2i64, V1, V2);
7915 // Try to use byte rotation instructions.
7916 // Its more profitable for pre-SSSE3 to use shuffles/unpacks.
7917 if (Subtarget->hasSSSE3())
7918 if (SDValue Rotate = lowerVectorShuffleAsByteRotate(
7919 DL, MVT::v2i64, V1, V2, Mask, Subtarget, DAG))
7922 // If we have direct support for blends, we should lower by decomposing into
7923 // a permute. That will be faster than the domain cross.
7924 if (IsBlendSupported)
7925 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v2i64, V1, V2,
7928 // We implement this with SHUFPD which is pretty lame because it will likely
7929 // incur 2 cycles of stall for integer vectors on Nehalem and older chips.
7930 // However, all the alternatives are still more cycles and newer chips don't
7931 // have this problem. It would be really nice if x86 had better shuffles here.
7932 V1 = DAG.getBitcast(MVT::v2f64, V1);
7933 V2 = DAG.getBitcast(MVT::v2f64, V2);
7934 return DAG.getBitcast(MVT::v2i64,
7935 DAG.getVectorShuffle(MVT::v2f64, DL, V1, V2, Mask));
7938 /// \brief Test whether this can be lowered with a single SHUFPS instruction.
7940 /// This is used to disable more specialized lowerings when the shufps lowering
7941 /// will happen to be efficient.
7942 static bool isSingleSHUFPSMask(ArrayRef<int> Mask) {
7943 // This routine only handles 128-bit shufps.
7944 assert(Mask.size() == 4 && "Unsupported mask size!");
7946 // To lower with a single SHUFPS we need to have the low half and high half
7947 // each requiring a single input.
7948 if (Mask[0] != -1 && Mask[1] != -1 && (Mask[0] < 4) != (Mask[1] < 4))
7950 if (Mask[2] != -1 && Mask[3] != -1 && (Mask[2] < 4) != (Mask[3] < 4))
7956 /// \brief Lower a vector shuffle using the SHUFPS instruction.
7958 /// This is a helper routine dedicated to lowering vector shuffles using SHUFPS.
7959 /// It makes no assumptions about whether this is the *best* lowering, it simply
7961 static SDValue lowerVectorShuffleWithSHUFPS(SDLoc DL, MVT VT,
7962 ArrayRef<int> Mask, SDValue V1,
7963 SDValue V2, SelectionDAG &DAG) {
7964 SDValue LowV = V1, HighV = V2;
7965 int NewMask[4] = {Mask[0], Mask[1], Mask[2], Mask[3]};
7968 std::count_if(Mask.begin(), Mask.end(), [](int M) { return M >= 4; });
7970 if (NumV2Elements == 1) {
7972 std::find_if(Mask.begin(), Mask.end(), [](int M) { return M >= 4; }) -
7975 // Compute the index adjacent to V2Index and in the same half by toggling
7977 int V2AdjIndex = V2Index ^ 1;
7979 if (Mask[V2AdjIndex] == -1) {
7980 // Handles all the cases where we have a single V2 element and an undef.
7981 // This will only ever happen in the high lanes because we commute the
7982 // vector otherwise.
7984 std::swap(LowV, HighV);
7985 NewMask[V2Index] -= 4;
7987 // Handle the case where the V2 element ends up adjacent to a V1 element.
7988 // To make this work, blend them together as the first step.
7989 int V1Index = V2AdjIndex;
7990 int BlendMask[4] = {Mask[V2Index] - 4, 0, Mask[V1Index], 0};
7991 V2 = DAG.getNode(X86ISD::SHUFP, DL, VT, V2, V1,
7992 getV4X86ShuffleImm8ForMask(BlendMask, DL, DAG));
7994 // Now proceed to reconstruct the final blend as we have the necessary
7995 // high or low half formed.
8002 NewMask[V1Index] = 2; // We put the V1 element in V2[2].
8003 NewMask[V2Index] = 0; // We shifted the V2 element into V2[0].
8005 } else if (NumV2Elements == 2) {
8006 if (Mask[0] < 4 && Mask[1] < 4) {
8007 // Handle the easy case where we have V1 in the low lanes and V2 in the
8011 } else if (Mask[2] < 4 && Mask[3] < 4) {
8012 // We also handle the reversed case because this utility may get called
8013 // when we detect a SHUFPS pattern but can't easily commute the shuffle to
8014 // arrange things in the right direction.
8020 // We have a mixture of V1 and V2 in both low and high lanes. Rather than
8021 // trying to place elements directly, just blend them and set up the final
8022 // shuffle to place them.
8024 // The first two blend mask elements are for V1, the second two are for
8026 int BlendMask[4] = {Mask[0] < 4 ? Mask[0] : Mask[1],
8027 Mask[2] < 4 ? Mask[2] : Mask[3],
8028 (Mask[0] >= 4 ? Mask[0] : Mask[1]) - 4,
8029 (Mask[2] >= 4 ? Mask[2] : Mask[3]) - 4};
8030 V1 = DAG.getNode(X86ISD::SHUFP, DL, VT, V1, V2,
8031 getV4X86ShuffleImm8ForMask(BlendMask, DL, DAG));
8033 // Now we do a normal shuffle of V1 by giving V1 as both operands to
8036 NewMask[0] = Mask[0] < 4 ? 0 : 2;
8037 NewMask[1] = Mask[0] < 4 ? 2 : 0;
8038 NewMask[2] = Mask[2] < 4 ? 1 : 3;
8039 NewMask[3] = Mask[2] < 4 ? 3 : 1;
8042 return DAG.getNode(X86ISD::SHUFP, DL, VT, LowV, HighV,
8043 getV4X86ShuffleImm8ForMask(NewMask, DL, DAG));
8046 /// \brief Lower 4-lane 32-bit floating point shuffles.
8048 /// Uses instructions exclusively from the floating point unit to minimize
8049 /// domain crossing penalties, as these are sufficient to implement all v4f32
8051 static SDValue lowerV4F32VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
8052 const X86Subtarget *Subtarget,
8053 SelectionDAG &DAG) {
8055 assert(Op.getSimpleValueType() == MVT::v4f32 && "Bad shuffle type!");
8056 assert(V1.getSimpleValueType() == MVT::v4f32 && "Bad operand type!");
8057 assert(V2.getSimpleValueType() == MVT::v4f32 && "Bad operand type!");
8058 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
8059 ArrayRef<int> Mask = SVOp->getMask();
8060 assert(Mask.size() == 4 && "Unexpected mask size for v4 shuffle!");
8063 std::count_if(Mask.begin(), Mask.end(), [](int M) { return M >= 4; });
8065 if (NumV2Elements == 0) {
8066 // Check for being able to broadcast a single element.
8067 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v4f32, V1,
8068 Mask, Subtarget, DAG))
8071 // Use even/odd duplicate instructions for masks that match their pattern.
8072 if (Subtarget->hasSSE3()) {
8073 if (isShuffleEquivalent(V1, V2, Mask, {0, 0, 2, 2}))
8074 return DAG.getNode(X86ISD::MOVSLDUP, DL, MVT::v4f32, V1);
8075 if (isShuffleEquivalent(V1, V2, Mask, {1, 1, 3, 3}))
8076 return DAG.getNode(X86ISD::MOVSHDUP, DL, MVT::v4f32, V1);
8079 if (Subtarget->hasAVX()) {
8080 // If we have AVX, we can use VPERMILPS which will allow folding a load
8081 // into the shuffle.
8082 return DAG.getNode(X86ISD::VPERMILPI, DL, MVT::v4f32, V1,
8083 getV4X86ShuffleImm8ForMask(Mask, DL, DAG));
8086 // Otherwise, use a straight shuffle of a single input vector. We pass the
8087 // input vector to both operands to simulate this with a SHUFPS.
8088 return DAG.getNode(X86ISD::SHUFP, DL, MVT::v4f32, V1, V1,
8089 getV4X86ShuffleImm8ForMask(Mask, DL, DAG));
8092 // There are special ways we can lower some single-element blends. However, we
8093 // have custom ways we can lower more complex single-element blends below that
8094 // we defer to if both this and BLENDPS fail to match, so restrict this to
8095 // when the V2 input is targeting element 0 of the mask -- that is the fast
8097 if (NumV2Elements == 1 && Mask[0] >= 4)
8098 if (SDValue V = lowerVectorShuffleAsElementInsertion(DL, MVT::v4f32, V1, V2,
8099 Mask, Subtarget, DAG))
8102 if (Subtarget->hasSSE41()) {
8103 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v4f32, V1, V2, Mask,
8107 // Use INSERTPS if we can complete the shuffle efficiently.
8108 if (SDValue V = lowerVectorShuffleAsInsertPS(Op, V1, V2, Mask, DAG))
8111 if (!isSingleSHUFPSMask(Mask))
8112 if (SDValue BlendPerm = lowerVectorShuffleAsBlendAndPermute(
8113 DL, MVT::v4f32, V1, V2, Mask, DAG))
8117 // Use dedicated unpack instructions for masks that match their pattern.
8118 if (isShuffleEquivalent(V1, V2, Mask, {0, 4, 1, 5}))
8119 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v4f32, V1, V2);
8120 if (isShuffleEquivalent(V1, V2, Mask, {2, 6, 3, 7}))
8121 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v4f32, V1, V2);
8122 if (isShuffleEquivalent(V1, V2, Mask, {4, 0, 5, 1}))
8123 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v4f32, V2, V1);
8124 if (isShuffleEquivalent(V1, V2, Mask, {6, 2, 7, 3}))
8125 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v4f32, V2, V1);
8127 // Otherwise fall back to a SHUFPS lowering strategy.
8128 return lowerVectorShuffleWithSHUFPS(DL, MVT::v4f32, Mask, V1, V2, DAG);
8131 /// \brief Lower 4-lane i32 vector shuffles.
8133 /// We try to handle these with integer-domain shuffles where we can, but for
8134 /// blends we use the floating point domain blend instructions.
8135 static SDValue lowerV4I32VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
8136 const X86Subtarget *Subtarget,
8137 SelectionDAG &DAG) {
8139 assert(Op.getSimpleValueType() == MVT::v4i32 && "Bad shuffle type!");
8140 assert(V1.getSimpleValueType() == MVT::v4i32 && "Bad operand type!");
8141 assert(V2.getSimpleValueType() == MVT::v4i32 && "Bad operand type!");
8142 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
8143 ArrayRef<int> Mask = SVOp->getMask();
8144 assert(Mask.size() == 4 && "Unexpected mask size for v4 shuffle!");
8146 // Whenever we can lower this as a zext, that instruction is strictly faster
8147 // than any alternative. It also allows us to fold memory operands into the
8148 // shuffle in many cases.
8149 if (SDValue ZExt = lowerVectorShuffleAsZeroOrAnyExtend(DL, MVT::v4i32, V1, V2,
8150 Mask, Subtarget, DAG))
8154 std::count_if(Mask.begin(), Mask.end(), [](int M) { return M >= 4; });
8156 if (NumV2Elements == 0) {
8157 // Check for being able to broadcast a single element.
8158 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v4i32, V1,
8159 Mask, Subtarget, DAG))
8162 // Straight shuffle of a single input vector. For everything from SSE2
8163 // onward this has a single fast instruction with no scary immediates.
8164 // We coerce the shuffle pattern to be compatible with UNPCK instructions
8165 // but we aren't actually going to use the UNPCK instruction because doing
8166 // so prevents folding a load into this instruction or making a copy.
8167 const int UnpackLoMask[] = {0, 0, 1, 1};
8168 const int UnpackHiMask[] = {2, 2, 3, 3};
8169 if (isShuffleEquivalent(V1, V2, Mask, {0, 0, 1, 1}))
8170 Mask = UnpackLoMask;
8171 else if (isShuffleEquivalent(V1, V2, Mask, {2, 2, 3, 3}))
8172 Mask = UnpackHiMask;
8174 return DAG.getNode(X86ISD::PSHUFD, DL, MVT::v4i32, V1,
8175 getV4X86ShuffleImm8ForMask(Mask, DL, DAG));
8178 // Try to use shift instructions.
8180 lowerVectorShuffleAsShift(DL, MVT::v4i32, V1, V2, Mask, DAG))
8183 // There are special ways we can lower some single-element blends.
8184 if (NumV2Elements == 1)
8185 if (SDValue V = lowerVectorShuffleAsElementInsertion(DL, MVT::v4i32, V1, V2,
8186 Mask, Subtarget, DAG))
8189 // We have different paths for blend lowering, but they all must use the
8190 // *exact* same predicate.
8191 bool IsBlendSupported = Subtarget->hasSSE41();
8192 if (IsBlendSupported)
8193 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v4i32, V1, V2, Mask,
8197 if (SDValue Masked =
8198 lowerVectorShuffleAsBitMask(DL, MVT::v4i32, V1, V2, Mask, DAG))
8201 // Use dedicated unpack instructions for masks that match their pattern.
8202 if (isShuffleEquivalent(V1, V2, Mask, {0, 4, 1, 5}))
8203 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v4i32, V1, V2);
8204 if (isShuffleEquivalent(V1, V2, Mask, {2, 6, 3, 7}))
8205 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v4i32, V1, V2);
8206 if (isShuffleEquivalent(V1, V2, Mask, {4, 0, 5, 1}))
8207 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v4i32, V2, V1);
8208 if (isShuffleEquivalent(V1, V2, Mask, {6, 2, 7, 3}))
8209 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v4i32, V2, V1);
8211 // Try to use byte rotation instructions.
8212 // Its more profitable for pre-SSSE3 to use shuffles/unpacks.
8213 if (Subtarget->hasSSSE3())
8214 if (SDValue Rotate = lowerVectorShuffleAsByteRotate(
8215 DL, MVT::v4i32, V1, V2, Mask, Subtarget, DAG))
8218 // If we have direct support for blends, we should lower by decomposing into
8219 // a permute. That will be faster than the domain cross.
8220 if (IsBlendSupported)
8221 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v4i32, V1, V2,
8224 // Try to lower by permuting the inputs into an unpack instruction.
8225 if (SDValue Unpack =
8226 lowerVectorShuffleAsUnpack(DL, MVT::v4i32, V1, V2, Mask, DAG))
8229 // We implement this with SHUFPS because it can blend from two vectors.
8230 // Because we're going to eventually use SHUFPS, we use SHUFPS even to build
8231 // up the inputs, bypassing domain shift penalties that we would encur if we
8232 // directly used PSHUFD on Nehalem and older. For newer chips, this isn't
8234 return DAG.getBitcast(
8236 DAG.getVectorShuffle(MVT::v4f32, DL, DAG.getBitcast(MVT::v4f32, V1),
8237 DAG.getBitcast(MVT::v4f32, V2), Mask));
8240 /// \brief Lowering of single-input v8i16 shuffles is the cornerstone of SSE2
8241 /// shuffle lowering, and the most complex part.
8243 /// The lowering strategy is to try to form pairs of input lanes which are
8244 /// targeted at the same half of the final vector, and then use a dword shuffle
8245 /// to place them onto the right half, and finally unpack the paired lanes into
8246 /// their final position.
8248 /// The exact breakdown of how to form these dword pairs and align them on the
8249 /// correct sides is really tricky. See the comments within the function for
8250 /// more of the details.
8252 /// This code also handles repeated 128-bit lanes of v8i16 shuffles, but each
8253 /// lane must shuffle the *exact* same way. In fact, you must pass a v8 Mask to
8254 /// this routine for it to work correctly. To shuffle a 256-bit or 512-bit i16
8255 /// vector, form the analogous 128-bit 8-element Mask.
8256 static SDValue lowerV8I16GeneralSingleInputVectorShuffle(
8257 SDLoc DL, MVT VT, SDValue V, MutableArrayRef<int> Mask,
8258 const X86Subtarget *Subtarget, SelectionDAG &DAG) {
8259 assert(VT.getScalarType() == MVT::i16 && "Bad input type!");
8260 MVT PSHUFDVT = MVT::getVectorVT(MVT::i32, VT.getVectorNumElements() / 2);
8262 assert(Mask.size() == 8 && "Shuffle mask length doen't match!");
8263 MutableArrayRef<int> LoMask = Mask.slice(0, 4);
8264 MutableArrayRef<int> HiMask = Mask.slice(4, 4);
8266 SmallVector<int, 4> LoInputs;
8267 std::copy_if(LoMask.begin(), LoMask.end(), std::back_inserter(LoInputs),
8268 [](int M) { return M >= 0; });
8269 std::sort(LoInputs.begin(), LoInputs.end());
8270 LoInputs.erase(std::unique(LoInputs.begin(), LoInputs.end()), LoInputs.end());
8271 SmallVector<int, 4> HiInputs;
8272 std::copy_if(HiMask.begin(), HiMask.end(), std::back_inserter(HiInputs),
8273 [](int M) { return M >= 0; });
8274 std::sort(HiInputs.begin(), HiInputs.end());
8275 HiInputs.erase(std::unique(HiInputs.begin(), HiInputs.end()), HiInputs.end());
8277 std::lower_bound(LoInputs.begin(), LoInputs.end(), 4) - LoInputs.begin();
8278 int NumHToL = LoInputs.size() - NumLToL;
8280 std::lower_bound(HiInputs.begin(), HiInputs.end(), 4) - HiInputs.begin();
8281 int NumHToH = HiInputs.size() - NumLToH;
8282 MutableArrayRef<int> LToLInputs(LoInputs.data(), NumLToL);
8283 MutableArrayRef<int> LToHInputs(HiInputs.data(), NumLToH);
8284 MutableArrayRef<int> HToLInputs(LoInputs.data() + NumLToL, NumHToL);
8285 MutableArrayRef<int> HToHInputs(HiInputs.data() + NumLToH, NumHToH);
8287 // Simplify the 1-into-3 and 3-into-1 cases with a single pshufd. For all
8288 // such inputs we can swap two of the dwords across the half mark and end up
8289 // with <=2 inputs to each half in each half. Once there, we can fall through
8290 // to the generic code below. For example:
8292 // Input: [a, b, c, d, e, f, g, h] -PSHUFD[0,2,1,3]-> [a, b, e, f, c, d, g, h]
8293 // Mask: [0, 1, 2, 7, 4, 5, 6, 3] -----------------> [0, 1, 4, 7, 2, 3, 6, 5]
8295 // However in some very rare cases we have a 1-into-3 or 3-into-1 on one half
8296 // and an existing 2-into-2 on the other half. In this case we may have to
8297 // pre-shuffle the 2-into-2 half to avoid turning it into a 3-into-1 or
8298 // 1-into-3 which could cause us to cycle endlessly fixing each side in turn.
8299 // Fortunately, we don't have to handle anything but a 2-into-2 pattern
8300 // because any other situation (including a 3-into-1 or 1-into-3 in the other
8301 // half than the one we target for fixing) will be fixed when we re-enter this
8302 // path. We will also combine away any sequence of PSHUFD instructions that
8303 // result into a single instruction. Here is an example of the tricky case:
8305 // Input: [a, b, c, d, e, f, g, h] -PSHUFD[0,2,1,3]-> [a, b, e, f, c, d, g, h]
8306 // Mask: [3, 7, 1, 0, 2, 7, 3, 5] -THIS-IS-BAD!!!!-> [5, 7, 1, 0, 4, 7, 5, 3]
8308 // This now has a 1-into-3 in the high half! Instead, we do two shuffles:
8310 // Input: [a, b, c, d, e, f, g, h] PSHUFHW[0,2,1,3]-> [a, b, c, d, e, g, f, h]
8311 // Mask: [3, 7, 1, 0, 2, 7, 3, 5] -----------------> [3, 7, 1, 0, 2, 7, 3, 6]
8313 // Input: [a, b, c, d, e, g, f, h] -PSHUFD[0,2,1,3]-> [a, b, e, g, c, d, f, h]
8314 // Mask: [3, 7, 1, 0, 2, 7, 3, 6] -----------------> [5, 7, 1, 0, 4, 7, 5, 6]
8316 // The result is fine to be handled by the generic logic.
8317 auto balanceSides = [&](ArrayRef<int> AToAInputs, ArrayRef<int> BToAInputs,
8318 ArrayRef<int> BToBInputs, ArrayRef<int> AToBInputs,
8319 int AOffset, int BOffset) {
8320 assert((AToAInputs.size() == 3 || AToAInputs.size() == 1) &&
8321 "Must call this with A having 3 or 1 inputs from the A half.");
8322 assert((BToAInputs.size() == 1 || BToAInputs.size() == 3) &&
8323 "Must call this with B having 1 or 3 inputs from the B half.");
8324 assert(AToAInputs.size() + BToAInputs.size() == 4 &&
8325 "Must call this with either 3:1 or 1:3 inputs (summing to 4).");
8327 // Compute the index of dword with only one word among the three inputs in
8328 // a half by taking the sum of the half with three inputs and subtracting
8329 // the sum of the actual three inputs. The difference is the remaining
8332 int &TripleDWord = AToAInputs.size() == 3 ? ADWord : BDWord;
8333 int &OneInputDWord = AToAInputs.size() == 3 ? BDWord : ADWord;
8334 int TripleInputOffset = AToAInputs.size() == 3 ? AOffset : BOffset;
8335 ArrayRef<int> TripleInputs = AToAInputs.size() == 3 ? AToAInputs : BToAInputs;
8336 int OneInput = AToAInputs.size() == 3 ? BToAInputs[0] : AToAInputs[0];
8337 int TripleInputSum = 0 + 1 + 2 + 3 + (4 * TripleInputOffset);
8338 int TripleNonInputIdx =
8339 TripleInputSum - std::accumulate(TripleInputs.begin(), TripleInputs.end(), 0);
8340 TripleDWord = TripleNonInputIdx / 2;
8342 // We use xor with one to compute the adjacent DWord to whichever one the
8344 OneInputDWord = (OneInput / 2) ^ 1;
8346 // Check for one tricky case: We're fixing a 3<-1 or a 1<-3 shuffle for AToA
8347 // and BToA inputs. If there is also such a problem with the BToB and AToB
8348 // inputs, we don't try to fix it necessarily -- we'll recurse and see it in
8349 // the next pass. However, if we have a 2<-2 in the BToB and AToB inputs, it
8350 // is essential that we don't *create* a 3<-1 as then we might oscillate.
8351 if (BToBInputs.size() == 2 && AToBInputs.size() == 2) {
8352 // Compute how many inputs will be flipped by swapping these DWords. We
8354 // to balance this to ensure we don't form a 3-1 shuffle in the other
8356 int NumFlippedAToBInputs =
8357 std::count(AToBInputs.begin(), AToBInputs.end(), 2 * ADWord) +
8358 std::count(AToBInputs.begin(), AToBInputs.end(), 2 * ADWord + 1);
8359 int NumFlippedBToBInputs =
8360 std::count(BToBInputs.begin(), BToBInputs.end(), 2 * BDWord) +
8361 std::count(BToBInputs.begin(), BToBInputs.end(), 2 * BDWord + 1);
8362 if ((NumFlippedAToBInputs == 1 &&
8363 (NumFlippedBToBInputs == 0 || NumFlippedBToBInputs == 2)) ||
8364 (NumFlippedBToBInputs == 1 &&
8365 (NumFlippedAToBInputs == 0 || NumFlippedAToBInputs == 2))) {
8366 // We choose whether to fix the A half or B half based on whether that
8367 // half has zero flipped inputs. At zero, we may not be able to fix it
8368 // with that half. We also bias towards fixing the B half because that
8369 // will more commonly be the high half, and we have to bias one way.
8370 auto FixFlippedInputs = [&V, &DL, &Mask, &DAG](int PinnedIdx, int DWord,
8371 ArrayRef<int> Inputs) {
8372 int FixIdx = PinnedIdx ^ 1; // The adjacent slot to the pinned slot.
8373 bool IsFixIdxInput = std::find(Inputs.begin(), Inputs.end(),
8374 PinnedIdx ^ 1) != Inputs.end();
8375 // Determine whether the free index is in the flipped dword or the
8376 // unflipped dword based on where the pinned index is. We use this bit
8377 // in an xor to conditionally select the adjacent dword.
8378 int FixFreeIdx = 2 * (DWord ^ (PinnedIdx / 2 == DWord));
8379 bool IsFixFreeIdxInput = std::find(Inputs.begin(), Inputs.end(),
8380 FixFreeIdx) != Inputs.end();
8381 if (IsFixIdxInput == IsFixFreeIdxInput)
8383 IsFixFreeIdxInput = std::find(Inputs.begin(), Inputs.end(),
8384 FixFreeIdx) != Inputs.end();
8385 assert(IsFixIdxInput != IsFixFreeIdxInput &&
8386 "We need to be changing the number of flipped inputs!");
8387 int PSHUFHalfMask[] = {0, 1, 2, 3};
8388 std::swap(PSHUFHalfMask[FixFreeIdx % 4], PSHUFHalfMask[FixIdx % 4]);
8389 V = DAG.getNode(FixIdx < 4 ? X86ISD::PSHUFLW : X86ISD::PSHUFHW, DL,
8391 getV4X86ShuffleImm8ForMask(PSHUFHalfMask, DL, DAG));
8394 if (M != -1 && M == FixIdx)
8396 else if (M != -1 && M == FixFreeIdx)
8399 if (NumFlippedBToBInputs != 0) {
8401 BToAInputs.size() == 3 ? TripleNonInputIdx : OneInput;
8402 FixFlippedInputs(BPinnedIdx, BDWord, BToBInputs);
8404 assert(NumFlippedAToBInputs != 0 && "Impossible given predicates!");
8406 AToAInputs.size() == 3 ? TripleNonInputIdx : OneInput;
8407 FixFlippedInputs(APinnedIdx, ADWord, AToBInputs);
8412 int PSHUFDMask[] = {0, 1, 2, 3};
8413 PSHUFDMask[ADWord] = BDWord;
8414 PSHUFDMask[BDWord] = ADWord;
8417 DAG.getNode(X86ISD::PSHUFD, DL, PSHUFDVT, DAG.getBitcast(PSHUFDVT, V),
8418 getV4X86ShuffleImm8ForMask(PSHUFDMask, DL, DAG)));
8420 // Adjust the mask to match the new locations of A and B.
8422 if (M != -1 && M/2 == ADWord)
8423 M = 2 * BDWord + M % 2;
8424 else if (M != -1 && M/2 == BDWord)
8425 M = 2 * ADWord + M % 2;
8427 // Recurse back into this routine to re-compute state now that this isn't
8428 // a 3 and 1 problem.
8429 return lowerV8I16GeneralSingleInputVectorShuffle(DL, VT, V, Mask, Subtarget,
8432 if ((NumLToL == 3 && NumHToL == 1) || (NumLToL == 1 && NumHToL == 3))
8433 return balanceSides(LToLInputs, HToLInputs, HToHInputs, LToHInputs, 0, 4);
8434 else if ((NumHToH == 3 && NumLToH == 1) || (NumHToH == 1 && NumLToH == 3))
8435 return balanceSides(HToHInputs, LToHInputs, LToLInputs, HToLInputs, 4, 0);
8437 // At this point there are at most two inputs to the low and high halves from
8438 // each half. That means the inputs can always be grouped into dwords and
8439 // those dwords can then be moved to the correct half with a dword shuffle.
8440 // We use at most one low and one high word shuffle to collect these paired
8441 // inputs into dwords, and finally a dword shuffle to place them.
8442 int PSHUFLMask[4] = {-1, -1, -1, -1};
8443 int PSHUFHMask[4] = {-1, -1, -1, -1};
8444 int PSHUFDMask[4] = {-1, -1, -1, -1};
8446 // First fix the masks for all the inputs that are staying in their
8447 // original halves. This will then dictate the targets of the cross-half
8449 auto fixInPlaceInputs =
8450 [&PSHUFDMask](ArrayRef<int> InPlaceInputs, ArrayRef<int> IncomingInputs,
8451 MutableArrayRef<int> SourceHalfMask,
8452 MutableArrayRef<int> HalfMask, int HalfOffset) {
8453 if (InPlaceInputs.empty())
8455 if (InPlaceInputs.size() == 1) {
8456 SourceHalfMask[InPlaceInputs[0] - HalfOffset] =
8457 InPlaceInputs[0] - HalfOffset;
8458 PSHUFDMask[InPlaceInputs[0] / 2] = InPlaceInputs[0] / 2;
8461 if (IncomingInputs.empty()) {
8462 // Just fix all of the in place inputs.
8463 for (int Input : InPlaceInputs) {
8464 SourceHalfMask[Input - HalfOffset] = Input - HalfOffset;
8465 PSHUFDMask[Input / 2] = Input / 2;
8470 assert(InPlaceInputs.size() == 2 && "Cannot handle 3 or 4 inputs!");
8471 SourceHalfMask[InPlaceInputs[0] - HalfOffset] =
8472 InPlaceInputs[0] - HalfOffset;
8473 // Put the second input next to the first so that they are packed into
8474 // a dword. We find the adjacent index by toggling the low bit.
8475 int AdjIndex = InPlaceInputs[0] ^ 1;
8476 SourceHalfMask[AdjIndex - HalfOffset] = InPlaceInputs[1] - HalfOffset;
8477 std::replace(HalfMask.begin(), HalfMask.end(), InPlaceInputs[1], AdjIndex);
8478 PSHUFDMask[AdjIndex / 2] = AdjIndex / 2;
8480 fixInPlaceInputs(LToLInputs, HToLInputs, PSHUFLMask, LoMask, 0);
8481 fixInPlaceInputs(HToHInputs, LToHInputs, PSHUFHMask, HiMask, 4);
8483 // Now gather the cross-half inputs and place them into a free dword of
8484 // their target half.
8485 // FIXME: This operation could almost certainly be simplified dramatically to
8486 // look more like the 3-1 fixing operation.
8487 auto moveInputsToRightHalf = [&PSHUFDMask](
8488 MutableArrayRef<int> IncomingInputs, ArrayRef<int> ExistingInputs,
8489 MutableArrayRef<int> SourceHalfMask, MutableArrayRef<int> HalfMask,
8490 MutableArrayRef<int> FinalSourceHalfMask, int SourceOffset,
8492 auto isWordClobbered = [](ArrayRef<int> SourceHalfMask, int Word) {
8493 return SourceHalfMask[Word] != -1 && SourceHalfMask[Word] != Word;
8495 auto isDWordClobbered = [&isWordClobbered](ArrayRef<int> SourceHalfMask,
8497 int LowWord = Word & ~1;
8498 int HighWord = Word | 1;
8499 return isWordClobbered(SourceHalfMask, LowWord) ||
8500 isWordClobbered(SourceHalfMask, HighWord);
8503 if (IncomingInputs.empty())
8506 if (ExistingInputs.empty()) {
8507 // Map any dwords with inputs from them into the right half.
8508 for (int Input : IncomingInputs) {
8509 // If the source half mask maps over the inputs, turn those into
8510 // swaps and use the swapped lane.
8511 if (isWordClobbered(SourceHalfMask, Input - SourceOffset)) {
8512 if (SourceHalfMask[SourceHalfMask[Input - SourceOffset]] == -1) {
8513 SourceHalfMask[SourceHalfMask[Input - SourceOffset]] =
8514 Input - SourceOffset;
8515 // We have to swap the uses in our half mask in one sweep.
8516 for (int &M : HalfMask)
8517 if (M == SourceHalfMask[Input - SourceOffset] + SourceOffset)
8519 else if (M == Input)
8520 M = SourceHalfMask[Input - SourceOffset] + SourceOffset;
8522 assert(SourceHalfMask[SourceHalfMask[Input - SourceOffset]] ==
8523 Input - SourceOffset &&
8524 "Previous placement doesn't match!");
8526 // Note that this correctly re-maps both when we do a swap and when
8527 // we observe the other side of the swap above. We rely on that to
8528 // avoid swapping the members of the input list directly.
8529 Input = SourceHalfMask[Input - SourceOffset] + SourceOffset;
8532 // Map the input's dword into the correct half.
8533 if (PSHUFDMask[(Input - SourceOffset + DestOffset) / 2] == -1)
8534 PSHUFDMask[(Input - SourceOffset + DestOffset) / 2] = Input / 2;
8536 assert(PSHUFDMask[(Input - SourceOffset + DestOffset) / 2] ==
8538 "Previous placement doesn't match!");
8541 // And just directly shift any other-half mask elements to be same-half
8542 // as we will have mirrored the dword containing the element into the
8543 // same position within that half.
8544 for (int &M : HalfMask)
8545 if (M >= SourceOffset && M < SourceOffset + 4) {
8546 M = M - SourceOffset + DestOffset;
8547 assert(M >= 0 && "This should never wrap below zero!");
8552 // Ensure we have the input in a viable dword of its current half. This
8553 // is particularly tricky because the original position may be clobbered
8554 // by inputs being moved and *staying* in that half.
8555 if (IncomingInputs.size() == 1) {
8556 if (isWordClobbered(SourceHalfMask, IncomingInputs[0] - SourceOffset)) {
8557 int InputFixed = std::find(std::begin(SourceHalfMask),
8558 std::end(SourceHalfMask), -1) -
8559 std::begin(SourceHalfMask) + SourceOffset;
8560 SourceHalfMask[InputFixed - SourceOffset] =
8561 IncomingInputs[0] - SourceOffset;
8562 std::replace(HalfMask.begin(), HalfMask.end(), IncomingInputs[0],
8564 IncomingInputs[0] = InputFixed;
8566 } else if (IncomingInputs.size() == 2) {
8567 if (IncomingInputs[0] / 2 != IncomingInputs[1] / 2 ||
8568 isDWordClobbered(SourceHalfMask, IncomingInputs[0] - SourceOffset)) {
8569 // We have two non-adjacent or clobbered inputs we need to extract from
8570 // the source half. To do this, we need to map them into some adjacent
8571 // dword slot in the source mask.
8572 int InputsFixed[2] = {IncomingInputs[0] - SourceOffset,
8573 IncomingInputs[1] - SourceOffset};
8575 // If there is a free slot in the source half mask adjacent to one of
8576 // the inputs, place the other input in it. We use (Index XOR 1) to
8577 // compute an adjacent index.
8578 if (!isWordClobbered(SourceHalfMask, InputsFixed[0]) &&
8579 SourceHalfMask[InputsFixed[0] ^ 1] == -1) {
8580 SourceHalfMask[InputsFixed[0]] = InputsFixed[0];
8581 SourceHalfMask[InputsFixed[0] ^ 1] = InputsFixed[1];
8582 InputsFixed[1] = InputsFixed[0] ^ 1;
8583 } else if (!isWordClobbered(SourceHalfMask, InputsFixed[1]) &&
8584 SourceHalfMask[InputsFixed[1] ^ 1] == -1) {
8585 SourceHalfMask[InputsFixed[1]] = InputsFixed[1];
8586 SourceHalfMask[InputsFixed[1] ^ 1] = InputsFixed[0];
8587 InputsFixed[0] = InputsFixed[1] ^ 1;
8588 } else if (SourceHalfMask[2 * ((InputsFixed[0] / 2) ^ 1)] == -1 &&
8589 SourceHalfMask[2 * ((InputsFixed[0] / 2) ^ 1) + 1] == -1) {
8590 // The two inputs are in the same DWord but it is clobbered and the
8591 // adjacent DWord isn't used at all. Move both inputs to the free
8593 SourceHalfMask[2 * ((InputsFixed[0] / 2) ^ 1)] = InputsFixed[0];
8594 SourceHalfMask[2 * ((InputsFixed[0] / 2) ^ 1) + 1] = InputsFixed[1];
8595 InputsFixed[0] = 2 * ((InputsFixed[0] / 2) ^ 1);
8596 InputsFixed[1] = 2 * ((InputsFixed[0] / 2) ^ 1) + 1;
8598 // The only way we hit this point is if there is no clobbering
8599 // (because there are no off-half inputs to this half) and there is no
8600 // free slot adjacent to one of the inputs. In this case, we have to
8601 // swap an input with a non-input.
8602 for (int i = 0; i < 4; ++i)
8603 assert((SourceHalfMask[i] == -1 || SourceHalfMask[i] == i) &&
8604 "We can't handle any clobbers here!");
8605 assert(InputsFixed[1] != (InputsFixed[0] ^ 1) &&
8606 "Cannot have adjacent inputs here!");
8608 SourceHalfMask[InputsFixed[0] ^ 1] = InputsFixed[1];
8609 SourceHalfMask[InputsFixed[1]] = InputsFixed[0] ^ 1;
8611 // We also have to update the final source mask in this case because
8612 // it may need to undo the above swap.
8613 for (int &M : FinalSourceHalfMask)
8614 if (M == (InputsFixed[0] ^ 1) + SourceOffset)
8615 M = InputsFixed[1] + SourceOffset;
8616 else if (M == InputsFixed[1] + SourceOffset)
8617 M = (InputsFixed[0] ^ 1) + SourceOffset;
8619 InputsFixed[1] = InputsFixed[0] ^ 1;
8622 // Point everything at the fixed inputs.
8623 for (int &M : HalfMask)
8624 if (M == IncomingInputs[0])
8625 M = InputsFixed[0] + SourceOffset;
8626 else if (M == IncomingInputs[1])
8627 M = InputsFixed[1] + SourceOffset;
8629 IncomingInputs[0] = InputsFixed[0] + SourceOffset;
8630 IncomingInputs[1] = InputsFixed[1] + SourceOffset;
8633 llvm_unreachable("Unhandled input size!");
8636 // Now hoist the DWord down to the right half.
8637 int FreeDWord = (PSHUFDMask[DestOffset / 2] == -1 ? 0 : 1) + DestOffset / 2;
8638 assert(PSHUFDMask[FreeDWord] == -1 && "DWord not free");
8639 PSHUFDMask[FreeDWord] = IncomingInputs[0] / 2;
8640 for (int &M : HalfMask)
8641 for (int Input : IncomingInputs)
8643 M = FreeDWord * 2 + Input % 2;
8645 moveInputsToRightHalf(HToLInputs, LToLInputs, PSHUFHMask, LoMask, HiMask,
8646 /*SourceOffset*/ 4, /*DestOffset*/ 0);
8647 moveInputsToRightHalf(LToHInputs, HToHInputs, PSHUFLMask, HiMask, LoMask,
8648 /*SourceOffset*/ 0, /*DestOffset*/ 4);
8650 // Now enact all the shuffles we've computed to move the inputs into their
8652 if (!isNoopShuffleMask(PSHUFLMask))
8653 V = DAG.getNode(X86ISD::PSHUFLW, DL, VT, V,
8654 getV4X86ShuffleImm8ForMask(PSHUFLMask, DL, DAG));
8655 if (!isNoopShuffleMask(PSHUFHMask))
8656 V = DAG.getNode(X86ISD::PSHUFHW, DL, VT, V,
8657 getV4X86ShuffleImm8ForMask(PSHUFHMask, DL, DAG));
8658 if (!isNoopShuffleMask(PSHUFDMask))
8661 DAG.getNode(X86ISD::PSHUFD, DL, PSHUFDVT, DAG.getBitcast(PSHUFDVT, V),
8662 getV4X86ShuffleImm8ForMask(PSHUFDMask, DL, DAG)));
8664 // At this point, each half should contain all its inputs, and we can then
8665 // just shuffle them into their final position.
8666 assert(std::count_if(LoMask.begin(), LoMask.end(),
8667 [](int M) { return M >= 4; }) == 0 &&
8668 "Failed to lift all the high half inputs to the low mask!");
8669 assert(std::count_if(HiMask.begin(), HiMask.end(),
8670 [](int M) { return M >= 0 && M < 4; }) == 0 &&
8671 "Failed to lift all the low half inputs to the high mask!");
8673 // Do a half shuffle for the low mask.
8674 if (!isNoopShuffleMask(LoMask))
8675 V = DAG.getNode(X86ISD::PSHUFLW, DL, VT, V,
8676 getV4X86ShuffleImm8ForMask(LoMask, DL, DAG));
8678 // Do a half shuffle with the high mask after shifting its values down.
8679 for (int &M : HiMask)
8682 if (!isNoopShuffleMask(HiMask))
8683 V = DAG.getNode(X86ISD::PSHUFHW, DL, VT, V,
8684 getV4X86ShuffleImm8ForMask(HiMask, DL, DAG));
8689 /// \brief Helper to form a PSHUFB-based shuffle+blend.
8690 static SDValue lowerVectorShuffleAsPSHUFB(SDLoc DL, MVT VT, SDValue V1,
8691 SDValue V2, ArrayRef<int> Mask,
8692 SelectionDAG &DAG, bool &V1InUse,
8694 SmallBitVector Zeroable = computeZeroableShuffleElements(Mask, V1, V2);
8700 int Size = Mask.size();
8701 int Scale = 16 / Size;
8702 for (int i = 0; i < 16; ++i) {
8703 if (Mask[i / Scale] == -1) {
8704 V1Mask[i] = V2Mask[i] = DAG.getUNDEF(MVT::i8);
8706 const int ZeroMask = 0x80;
8707 int V1Idx = Mask[i / Scale] < Size ? Mask[i / Scale] * Scale + i % Scale
8709 int V2Idx = Mask[i / Scale] < Size
8711 : (Mask[i / Scale] - Size) * Scale + i % Scale;
8712 if (Zeroable[i / Scale])
8713 V1Idx = V2Idx = ZeroMask;
8714 V1Mask[i] = DAG.getConstant(V1Idx, DL, MVT::i8);
8715 V2Mask[i] = DAG.getConstant(V2Idx, DL, MVT::i8);
8716 V1InUse |= (ZeroMask != V1Idx);
8717 V2InUse |= (ZeroMask != V2Idx);
8722 V1 = DAG.getNode(X86ISD::PSHUFB, DL, MVT::v16i8,
8723 DAG.getBitcast(MVT::v16i8, V1),
8724 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v16i8, V1Mask));
8726 V2 = DAG.getNode(X86ISD::PSHUFB, DL, MVT::v16i8,
8727 DAG.getBitcast(MVT::v16i8, V2),
8728 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v16i8, V2Mask));
8730 // If we need shuffled inputs from both, blend the two.
8732 if (V1InUse && V2InUse)
8733 V = DAG.getNode(ISD::OR, DL, MVT::v16i8, V1, V2);
8735 V = V1InUse ? V1 : V2;
8737 // Cast the result back to the correct type.
8738 return DAG.getBitcast(VT, V);
8741 /// \brief Generic lowering of 8-lane i16 shuffles.
8743 /// This handles both single-input shuffles and combined shuffle/blends with
8744 /// two inputs. The single input shuffles are immediately delegated to
8745 /// a dedicated lowering routine.
8747 /// The blends are lowered in one of three fundamental ways. If there are few
8748 /// enough inputs, it delegates to a basic UNPCK-based strategy. If the shuffle
8749 /// of the input is significantly cheaper when lowered as an interleaving of
8750 /// the two inputs, try to interleave them. Otherwise, blend the low and high
8751 /// halves of the inputs separately (making them have relatively few inputs)
8752 /// and then concatenate them.
8753 static SDValue lowerV8I16VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
8754 const X86Subtarget *Subtarget,
8755 SelectionDAG &DAG) {
8757 assert(Op.getSimpleValueType() == MVT::v8i16 && "Bad shuffle type!");
8758 assert(V1.getSimpleValueType() == MVT::v8i16 && "Bad operand type!");
8759 assert(V2.getSimpleValueType() == MVT::v8i16 && "Bad operand type!");
8760 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
8761 ArrayRef<int> OrigMask = SVOp->getMask();
8762 int MaskStorage[8] = {OrigMask[0], OrigMask[1], OrigMask[2], OrigMask[3],
8763 OrigMask[4], OrigMask[5], OrigMask[6], OrigMask[7]};
8764 MutableArrayRef<int> Mask(MaskStorage);
8766 assert(Mask.size() == 8 && "Unexpected mask size for v8 shuffle!");
8768 // Whenever we can lower this as a zext, that instruction is strictly faster
8769 // than any alternative.
8770 if (SDValue ZExt = lowerVectorShuffleAsZeroOrAnyExtend(
8771 DL, MVT::v8i16, V1, V2, OrigMask, Subtarget, DAG))
8774 auto isV1 = [](int M) { return M >= 0 && M < 8; };
8776 auto isV2 = [](int M) { return M >= 8; };
8778 int NumV2Inputs = std::count_if(Mask.begin(), Mask.end(), isV2);
8780 if (NumV2Inputs == 0) {
8781 // Check for being able to broadcast a single element.
8782 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v8i16, V1,
8783 Mask, Subtarget, DAG))
8786 // Try to use shift instructions.
8788 lowerVectorShuffleAsShift(DL, MVT::v8i16, V1, V1, Mask, DAG))
8791 // Use dedicated unpack instructions for masks that match their pattern.
8792 if (isShuffleEquivalent(V1, V1, Mask, {0, 0, 1, 1, 2, 2, 3, 3}))
8793 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v8i16, V1, V1);
8794 if (isShuffleEquivalent(V1, V1, Mask, {4, 4, 5, 5, 6, 6, 7, 7}))
8795 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v8i16, V1, V1);
8797 // Try to use byte rotation instructions.
8798 if (SDValue Rotate = lowerVectorShuffleAsByteRotate(DL, MVT::v8i16, V1, V1,
8799 Mask, Subtarget, DAG))
8802 return lowerV8I16GeneralSingleInputVectorShuffle(DL, MVT::v8i16, V1, Mask,
8806 assert(std::any_of(Mask.begin(), Mask.end(), isV1) &&
8807 "All single-input shuffles should be canonicalized to be V1-input "
8810 // Try to use shift instructions.
8812 lowerVectorShuffleAsShift(DL, MVT::v8i16, V1, V2, Mask, DAG))
8815 // See if we can use SSE4A Extraction / Insertion.
8816 if (Subtarget->hasSSE4A())
8817 if (SDValue V = lowerVectorShuffleWithSSE4A(DL, MVT::v8i16, V1, V2, Mask, DAG))
8820 // There are special ways we can lower some single-element blends.
8821 if (NumV2Inputs == 1)
8822 if (SDValue V = lowerVectorShuffleAsElementInsertion(DL, MVT::v8i16, V1, V2,
8823 Mask, Subtarget, DAG))
8826 // We have different paths for blend lowering, but they all must use the
8827 // *exact* same predicate.
8828 bool IsBlendSupported = Subtarget->hasSSE41();
8829 if (IsBlendSupported)
8830 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v8i16, V1, V2, Mask,
8834 if (SDValue Masked =
8835 lowerVectorShuffleAsBitMask(DL, MVT::v8i16, V1, V2, Mask, DAG))
8838 // Use dedicated unpack instructions for masks that match their pattern.
8839 if (isShuffleEquivalent(V1, V2, Mask, {0, 8, 1, 9, 2, 10, 3, 11}))
8840 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v8i16, V1, V2);
8841 if (isShuffleEquivalent(V1, V2, Mask, {4, 12, 5, 13, 6, 14, 7, 15}))
8842 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v8i16, V1, V2);
8844 // Try to use byte rotation instructions.
8845 if (SDValue Rotate = lowerVectorShuffleAsByteRotate(
8846 DL, MVT::v8i16, V1, V2, Mask, Subtarget, DAG))
8849 if (SDValue BitBlend =
8850 lowerVectorShuffleAsBitBlend(DL, MVT::v8i16, V1, V2, Mask, DAG))
8853 if (SDValue Unpack =
8854 lowerVectorShuffleAsUnpack(DL, MVT::v8i16, V1, V2, Mask, DAG))
8857 // If we can't directly blend but can use PSHUFB, that will be better as it
8858 // can both shuffle and set up the inefficient blend.
8859 if (!IsBlendSupported && Subtarget->hasSSSE3()) {
8860 bool V1InUse, V2InUse;
8861 return lowerVectorShuffleAsPSHUFB(DL, MVT::v8i16, V1, V2, Mask, DAG,
8865 // We can always bit-blend if we have to so the fallback strategy is to
8866 // decompose into single-input permutes and blends.
8867 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v8i16, V1, V2,
8871 /// \brief Check whether a compaction lowering can be done by dropping even
8872 /// elements and compute how many times even elements must be dropped.
8874 /// This handles shuffles which take every Nth element where N is a power of
8875 /// two. Example shuffle masks:
8877 /// N = 1: 0, 2, 4, 6, 8, 10, 12, 14, 0, 2, 4, 6, 8, 10, 12, 14
8878 /// N = 1: 0, 2, 4, 6, 8, 10, 12, 14, 16, 18, 20, 22, 24, 26, 28, 30
8879 /// N = 2: 0, 4, 8, 12, 0, 4, 8, 12, 0, 4, 8, 12, 0, 4, 8, 12
8880 /// N = 2: 0, 4, 8, 12, 16, 20, 24, 28, 0, 4, 8, 12, 16, 20, 24, 28
8881 /// N = 3: 0, 8, 0, 8, 0, 8, 0, 8, 0, 8, 0, 8, 0, 8, 0, 8
8882 /// N = 3: 0, 8, 16, 24, 0, 8, 16, 24, 0, 8, 16, 24, 0, 8, 16, 24
8884 /// Any of these lanes can of course be undef.
8886 /// This routine only supports N <= 3.
8887 /// FIXME: Evaluate whether either AVX or AVX-512 have any opportunities here
8890 /// \returns N above, or the number of times even elements must be dropped if
8891 /// there is such a number. Otherwise returns zero.
8892 static int canLowerByDroppingEvenElements(ArrayRef<int> Mask) {
8893 // Figure out whether we're looping over two inputs or just one.
8894 bool IsSingleInput = isSingleInputShuffleMask(Mask);
8896 // The modulus for the shuffle vector entries is based on whether this is
8897 // a single input or not.
8898 int ShuffleModulus = Mask.size() * (IsSingleInput ? 1 : 2);
8899 assert(isPowerOf2_32((uint32_t)ShuffleModulus) &&
8900 "We should only be called with masks with a power-of-2 size!");
8902 uint64_t ModMask = (uint64_t)ShuffleModulus - 1;
8904 // We track whether the input is viable for all power-of-2 strides 2^1, 2^2,
8905 // and 2^3 simultaneously. This is because we may have ambiguity with
8906 // partially undef inputs.
8907 bool ViableForN[3] = {true, true, true};
8909 for (int i = 0, e = Mask.size(); i < e; ++i) {
8910 // Ignore undef lanes, we'll optimistically collapse them to the pattern we
8915 bool IsAnyViable = false;
8916 for (unsigned j = 0; j != array_lengthof(ViableForN); ++j)
8917 if (ViableForN[j]) {
8920 // The shuffle mask must be equal to (i * 2^N) % M.
8921 if ((uint64_t)Mask[i] == (((uint64_t)i << N) & ModMask))
8924 ViableForN[j] = false;
8926 // Early exit if we exhaust the possible powers of two.
8931 for (unsigned j = 0; j != array_lengthof(ViableForN); ++j)
8935 // Return 0 as there is no viable power of two.
8939 /// \brief Generic lowering of v16i8 shuffles.
8941 /// This is a hybrid strategy to lower v16i8 vectors. It first attempts to
8942 /// detect any complexity reducing interleaving. If that doesn't help, it uses
8943 /// UNPCK to spread the i8 elements across two i16-element vectors, and uses
8944 /// the existing lowering for v8i16 blends on each half, finally PACK-ing them
8946 static SDValue lowerV16I8VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
8947 const X86Subtarget *Subtarget,
8948 SelectionDAG &DAG) {
8950 assert(Op.getSimpleValueType() == MVT::v16i8 && "Bad shuffle type!");
8951 assert(V1.getSimpleValueType() == MVT::v16i8 && "Bad operand type!");
8952 assert(V2.getSimpleValueType() == MVT::v16i8 && "Bad operand type!");
8953 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
8954 ArrayRef<int> Mask = SVOp->getMask();
8955 assert(Mask.size() == 16 && "Unexpected mask size for v16 shuffle!");
8957 // Try to use shift instructions.
8959 lowerVectorShuffleAsShift(DL, MVT::v16i8, V1, V2, Mask, DAG))
8962 // Try to use byte rotation instructions.
8963 if (SDValue Rotate = lowerVectorShuffleAsByteRotate(
8964 DL, MVT::v16i8, V1, V2, Mask, Subtarget, DAG))
8967 // Try to use a zext lowering.
8968 if (SDValue ZExt = lowerVectorShuffleAsZeroOrAnyExtend(
8969 DL, MVT::v16i8, V1, V2, Mask, Subtarget, DAG))
8972 // See if we can use SSE4A Extraction / Insertion.
8973 if (Subtarget->hasSSE4A())
8974 if (SDValue V = lowerVectorShuffleWithSSE4A(DL, MVT::v16i8, V1, V2, Mask, DAG))
8978 std::count_if(Mask.begin(), Mask.end(), [](int M) { return M >= 16; });
8980 // For single-input shuffles, there are some nicer lowering tricks we can use.
8981 if (NumV2Elements == 0) {
8982 // Check for being able to broadcast a single element.
8983 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v16i8, V1,
8984 Mask, Subtarget, DAG))
8987 // Check whether we can widen this to an i16 shuffle by duplicating bytes.
8988 // Notably, this handles splat and partial-splat shuffles more efficiently.
8989 // However, it only makes sense if the pre-duplication shuffle simplifies
8990 // things significantly. Currently, this means we need to be able to
8991 // express the pre-duplication shuffle as an i16 shuffle.
8993 // FIXME: We should check for other patterns which can be widened into an
8994 // i16 shuffle as well.
8995 auto canWidenViaDuplication = [](ArrayRef<int> Mask) {
8996 for (int i = 0; i < 16; i += 2)
8997 if (Mask[i] != -1 && Mask[i + 1] != -1 && Mask[i] != Mask[i + 1])
9002 auto tryToWidenViaDuplication = [&]() -> SDValue {
9003 if (!canWidenViaDuplication(Mask))
9005 SmallVector<int, 4> LoInputs;
9006 std::copy_if(Mask.begin(), Mask.end(), std::back_inserter(LoInputs),
9007 [](int M) { return M >= 0 && M < 8; });
9008 std::sort(LoInputs.begin(), LoInputs.end());
9009 LoInputs.erase(std::unique(LoInputs.begin(), LoInputs.end()),
9011 SmallVector<int, 4> HiInputs;
9012 std::copy_if(Mask.begin(), Mask.end(), std::back_inserter(HiInputs),
9013 [](int M) { return M >= 8; });
9014 std::sort(HiInputs.begin(), HiInputs.end());
9015 HiInputs.erase(std::unique(HiInputs.begin(), HiInputs.end()),
9018 bool TargetLo = LoInputs.size() >= HiInputs.size();
9019 ArrayRef<int> InPlaceInputs = TargetLo ? LoInputs : HiInputs;
9020 ArrayRef<int> MovingInputs = TargetLo ? HiInputs : LoInputs;
9022 int PreDupI16Shuffle[] = {-1, -1, -1, -1, -1, -1, -1, -1};
9023 SmallDenseMap<int, int, 8> LaneMap;
9024 for (int I : InPlaceInputs) {
9025 PreDupI16Shuffle[I/2] = I/2;
9028 int j = TargetLo ? 0 : 4, je = j + 4;
9029 for (int i = 0, ie = MovingInputs.size(); i < ie; ++i) {
9030 // Check if j is already a shuffle of this input. This happens when
9031 // there are two adjacent bytes after we move the low one.
9032 if (PreDupI16Shuffle[j] != MovingInputs[i] / 2) {
9033 // If we haven't yet mapped the input, search for a slot into which
9035 while (j < je && PreDupI16Shuffle[j] != -1)
9039 // We can't place the inputs into a single half with a simple i16 shuffle, so bail.
9042 // Map this input with the i16 shuffle.
9043 PreDupI16Shuffle[j] = MovingInputs[i] / 2;
9046 // Update the lane map based on the mapping we ended up with.
9047 LaneMap[MovingInputs[i]] = 2 * j + MovingInputs[i] % 2;
9049 V1 = DAG.getBitcast(
9051 DAG.getVectorShuffle(MVT::v8i16, DL, DAG.getBitcast(MVT::v8i16, V1),
9052 DAG.getUNDEF(MVT::v8i16), PreDupI16Shuffle));
9054 // Unpack the bytes to form the i16s that will be shuffled into place.
9055 V1 = DAG.getNode(TargetLo ? X86ISD::UNPCKL : X86ISD::UNPCKH, DL,
9056 MVT::v16i8, V1, V1);
9058 int PostDupI16Shuffle[8] = {-1, -1, -1, -1, -1, -1, -1, -1};
9059 for (int i = 0; i < 16; ++i)
9060 if (Mask[i] != -1) {
9061 int MappedMask = LaneMap[Mask[i]] - (TargetLo ? 0 : 8);
9062 assert(MappedMask < 8 && "Invalid v8 shuffle mask!");
9063 if (PostDupI16Shuffle[i / 2] == -1)
9064 PostDupI16Shuffle[i / 2] = MappedMask;
9066 assert(PostDupI16Shuffle[i / 2] == MappedMask &&
9067 "Conflicting entrties in the original shuffle!");
9069 return DAG.getBitcast(
9071 DAG.getVectorShuffle(MVT::v8i16, DL, DAG.getBitcast(MVT::v8i16, V1),
9072 DAG.getUNDEF(MVT::v8i16), PostDupI16Shuffle));
9074 if (SDValue V = tryToWidenViaDuplication())
9078 if (SDValue Masked =
9079 lowerVectorShuffleAsBitMask(DL, MVT::v16i8, V1, V2, Mask, DAG))
9082 // Use dedicated unpack instructions for masks that match their pattern.
9083 if (isShuffleEquivalent(V1, V2, Mask, {// Low half.
9084 0, 16, 1, 17, 2, 18, 3, 19,
9086 4, 20, 5, 21, 6, 22, 7, 23}))
9087 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v16i8, V1, V2);
9088 if (isShuffleEquivalent(V1, V2, Mask, {// Low half.
9089 8, 24, 9, 25, 10, 26, 11, 27,
9091 12, 28, 13, 29, 14, 30, 15, 31}))
9092 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v16i8, V1, V2);
9094 // Check for SSSE3 which lets us lower all v16i8 shuffles much more directly
9095 // with PSHUFB. It is important to do this before we attempt to generate any
9096 // blends but after all of the single-input lowerings. If the single input
9097 // lowerings can find an instruction sequence that is faster than a PSHUFB, we
9098 // want to preserve that and we can DAG combine any longer sequences into
9099 // a PSHUFB in the end. But once we start blending from multiple inputs,
9100 // the complexity of DAG combining bad patterns back into PSHUFB is too high,
9101 // and there are *very* few patterns that would actually be faster than the
9102 // PSHUFB approach because of its ability to zero lanes.
9104 // FIXME: The only exceptions to the above are blends which are exact
9105 // interleavings with direct instructions supporting them. We currently don't
9106 // handle those well here.
9107 if (Subtarget->hasSSSE3()) {
9108 bool V1InUse = false;
9109 bool V2InUse = false;
9111 SDValue PSHUFB = lowerVectorShuffleAsPSHUFB(DL, MVT::v16i8, V1, V2, Mask,
9112 DAG, V1InUse, V2InUse);
9114 // If both V1 and V2 are in use and we can use a direct blend or an unpack,
9115 // do so. This avoids using them to handle blends-with-zero which is
9116 // important as a single pshufb is significantly faster for that.
9117 if (V1InUse && V2InUse) {
9118 if (Subtarget->hasSSE41())
9119 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v16i8, V1, V2,
9120 Mask, Subtarget, DAG))
9123 // We can use an unpack to do the blending rather than an or in some
9124 // cases. Even though the or may be (very minorly) more efficient, we
9125 // preference this lowering because there are common cases where part of
9126 // the complexity of the shuffles goes away when we do the final blend as
9128 // FIXME: It might be worth trying to detect if the unpack-feeding
9129 // shuffles will both be pshufb, in which case we shouldn't bother with
9131 if (SDValue Unpack =
9132 lowerVectorShuffleAsUnpack(DL, MVT::v16i8, V1, V2, Mask, DAG))
9139 // There are special ways we can lower some single-element blends.
9140 if (NumV2Elements == 1)
9141 if (SDValue V = lowerVectorShuffleAsElementInsertion(DL, MVT::v16i8, V1, V2,
9142 Mask, Subtarget, DAG))
9145 if (SDValue BitBlend =
9146 lowerVectorShuffleAsBitBlend(DL, MVT::v16i8, V1, V2, Mask, DAG))
9149 // Check whether a compaction lowering can be done. This handles shuffles
9150 // which take every Nth element for some even N. See the helper function for
9153 // We special case these as they can be particularly efficiently handled with
9154 // the PACKUSB instruction on x86 and they show up in common patterns of
9155 // rearranging bytes to truncate wide elements.
9156 if (int NumEvenDrops = canLowerByDroppingEvenElements(Mask)) {
9157 // NumEvenDrops is the power of two stride of the elements. Another way of
9158 // thinking about it is that we need to drop the even elements this many
9159 // times to get the original input.
9160 bool IsSingleInput = isSingleInputShuffleMask(Mask);
9162 // First we need to zero all the dropped bytes.
9163 assert(NumEvenDrops <= 3 &&
9164 "No support for dropping even elements more than 3 times.");
9165 // We use the mask type to pick which bytes are preserved based on how many
9166 // elements are dropped.
9167 MVT MaskVTs[] = { MVT::v8i16, MVT::v4i32, MVT::v2i64 };
9168 SDValue ByteClearMask = DAG.getBitcast(
9169 MVT::v16i8, DAG.getConstant(0xFF, DL, MaskVTs[NumEvenDrops - 1]));
9170 V1 = DAG.getNode(ISD::AND, DL, MVT::v16i8, V1, ByteClearMask);
9172 V2 = DAG.getNode(ISD::AND, DL, MVT::v16i8, V2, ByteClearMask);
9174 // Now pack things back together.
9175 V1 = DAG.getBitcast(MVT::v8i16, V1);
9176 V2 = IsSingleInput ? V1 : DAG.getBitcast(MVT::v8i16, V2);
9177 SDValue Result = DAG.getNode(X86ISD::PACKUS, DL, MVT::v16i8, V1, V2);
9178 for (int i = 1; i < NumEvenDrops; ++i) {
9179 Result = DAG.getBitcast(MVT::v8i16, Result);
9180 Result = DAG.getNode(X86ISD::PACKUS, DL, MVT::v16i8, Result, Result);
9186 // Handle multi-input cases by blending single-input shuffles.
9187 if (NumV2Elements > 0)
9188 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v16i8, V1, V2,
9191 // The fallback path for single-input shuffles widens this into two v8i16
9192 // vectors with unpacks, shuffles those, and then pulls them back together
9196 int LoBlendMask[8] = {-1, -1, -1, -1, -1, -1, -1, -1};
9197 int HiBlendMask[8] = {-1, -1, -1, -1, -1, -1, -1, -1};
9198 for (int i = 0; i < 16; ++i)
9200 (i < 8 ? LoBlendMask[i] : HiBlendMask[i % 8]) = Mask[i];
9202 SDValue Zero = getZeroVector(MVT::v8i16, Subtarget, DAG, DL);
9204 SDValue VLoHalf, VHiHalf;
9205 // Check if any of the odd lanes in the v16i8 are used. If not, we can mask
9206 // them out and avoid using UNPCK{L,H} to extract the elements of V as
9208 if (std::none_of(std::begin(LoBlendMask), std::end(LoBlendMask),
9209 [](int M) { return M >= 0 && M % 2 == 1; }) &&
9210 std::none_of(std::begin(HiBlendMask), std::end(HiBlendMask),
9211 [](int M) { return M >= 0 && M % 2 == 1; })) {
9212 // Use a mask to drop the high bytes.
9213 VLoHalf = DAG.getBitcast(MVT::v8i16, V);
9214 VLoHalf = DAG.getNode(ISD::AND, DL, MVT::v8i16, VLoHalf,
9215 DAG.getConstant(0x00FF, DL, MVT::v8i16));
9217 // This will be a single vector shuffle instead of a blend so nuke VHiHalf.
9218 VHiHalf = DAG.getUNDEF(MVT::v8i16);
9220 // Squash the masks to point directly into VLoHalf.
9221 for (int &M : LoBlendMask)
9224 for (int &M : HiBlendMask)
9228 // Otherwise just unpack the low half of V into VLoHalf and the high half into
9229 // VHiHalf so that we can blend them as i16s.
9230 VLoHalf = DAG.getBitcast(
9231 MVT::v8i16, DAG.getNode(X86ISD::UNPCKL, DL, MVT::v16i8, V, Zero));
9232 VHiHalf = DAG.getBitcast(
9233 MVT::v8i16, DAG.getNode(X86ISD::UNPCKH, DL, MVT::v16i8, V, Zero));
9236 SDValue LoV = DAG.getVectorShuffle(MVT::v8i16, DL, VLoHalf, VHiHalf, LoBlendMask);
9237 SDValue HiV = DAG.getVectorShuffle(MVT::v8i16, DL, VLoHalf, VHiHalf, HiBlendMask);
9239 return DAG.getNode(X86ISD::PACKUS, DL, MVT::v16i8, LoV, HiV);
9242 /// \brief Dispatching routine to lower various 128-bit x86 vector shuffles.
9244 /// This routine breaks down the specific type of 128-bit shuffle and
9245 /// dispatches to the lowering routines accordingly.
9246 static SDValue lower128BitVectorShuffle(SDValue Op, SDValue V1, SDValue V2,
9247 MVT VT, const X86Subtarget *Subtarget,
9248 SelectionDAG &DAG) {
9249 switch (VT.SimpleTy) {
9251 return lowerV2I64VectorShuffle(Op, V1, V2, Subtarget, DAG);
9253 return lowerV2F64VectorShuffle(Op, V1, V2, Subtarget, DAG);
9255 return lowerV4I32VectorShuffle(Op, V1, V2, Subtarget, DAG);
9257 return lowerV4F32VectorShuffle(Op, V1, V2, Subtarget, DAG);
9259 return lowerV8I16VectorShuffle(Op, V1, V2, Subtarget, DAG);
9261 return lowerV16I8VectorShuffle(Op, V1, V2, Subtarget, DAG);
9264 llvm_unreachable("Unimplemented!");
9268 /// \brief Helper function to test whether a shuffle mask could be
9269 /// simplified by widening the elements being shuffled.
9271 /// Appends the mask for wider elements in WidenedMask if valid. Otherwise
9272 /// leaves it in an unspecified state.
9274 /// NOTE: This must handle normal vector shuffle masks and *target* vector
9275 /// shuffle masks. The latter have the special property of a '-2' representing
9276 /// a zero-ed lane of a vector.
9277 static bool canWidenShuffleElements(ArrayRef<int> Mask,
9278 SmallVectorImpl<int> &WidenedMask) {
9279 for (int i = 0, Size = Mask.size(); i < Size; i += 2) {
9280 // If both elements are undef, its trivial.
9281 if (Mask[i] == SM_SentinelUndef && Mask[i + 1] == SM_SentinelUndef) {
9282 WidenedMask.push_back(SM_SentinelUndef);
9286 // Check for an undef mask and a mask value properly aligned to fit with
9287 // a pair of values. If we find such a case, use the non-undef mask's value.
9288 if (Mask[i] == SM_SentinelUndef && Mask[i + 1] >= 0 && Mask[i + 1] % 2 == 1) {
9289 WidenedMask.push_back(Mask[i + 1] / 2);
9292 if (Mask[i + 1] == SM_SentinelUndef && Mask[i] >= 0 && Mask[i] % 2 == 0) {
9293 WidenedMask.push_back(Mask[i] / 2);
9297 // When zeroing, we need to spread the zeroing across both lanes to widen.
9298 if (Mask[i] == SM_SentinelZero || Mask[i + 1] == SM_SentinelZero) {
9299 if ((Mask[i] == SM_SentinelZero || Mask[i] == SM_SentinelUndef) &&
9300 (Mask[i + 1] == SM_SentinelZero || Mask[i + 1] == SM_SentinelUndef)) {
9301 WidenedMask.push_back(SM_SentinelZero);
9307 // Finally check if the two mask values are adjacent and aligned with
9309 if (Mask[i] != SM_SentinelUndef && Mask[i] % 2 == 0 && Mask[i] + 1 == Mask[i + 1]) {
9310 WidenedMask.push_back(Mask[i] / 2);
9314 // Otherwise we can't safely widen the elements used in this shuffle.
9317 assert(WidenedMask.size() == Mask.size() / 2 &&
9318 "Incorrect size of mask after widening the elements!");
9323 /// \brief Generic routine to split vector shuffle into half-sized shuffles.
9325 /// This routine just extracts two subvectors, shuffles them independently, and
9326 /// then concatenates them back together. This should work effectively with all
9327 /// AVX vector shuffle types.
9328 static SDValue splitAndLowerVectorShuffle(SDLoc DL, MVT VT, SDValue V1,
9329 SDValue V2, ArrayRef<int> Mask,
9330 SelectionDAG &DAG) {
9331 assert(VT.getSizeInBits() >= 256 &&
9332 "Only for 256-bit or wider vector shuffles!");
9333 assert(V1.getSimpleValueType() == VT && "Bad operand type!");
9334 assert(V2.getSimpleValueType() == VT && "Bad operand type!");
9336 ArrayRef<int> LoMask = Mask.slice(0, Mask.size() / 2);
9337 ArrayRef<int> HiMask = Mask.slice(Mask.size() / 2);
9339 int NumElements = VT.getVectorNumElements();
9340 int SplitNumElements = NumElements / 2;
9341 MVT ScalarVT = VT.getScalarType();
9342 MVT SplitVT = MVT::getVectorVT(ScalarVT, NumElements / 2);
9344 // Rather than splitting build-vectors, just build two narrower build
9345 // vectors. This helps shuffling with splats and zeros.
9346 auto SplitVector = [&](SDValue V) {
9347 while (V.getOpcode() == ISD::BITCAST)
9348 V = V->getOperand(0);
9350 MVT OrigVT = V.getSimpleValueType();
9351 int OrigNumElements = OrigVT.getVectorNumElements();
9352 int OrigSplitNumElements = OrigNumElements / 2;
9353 MVT OrigScalarVT = OrigVT.getScalarType();
9354 MVT OrigSplitVT = MVT::getVectorVT(OrigScalarVT, OrigNumElements / 2);
9358 auto *BV = dyn_cast<BuildVectorSDNode>(V);
9360 LoV = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, OrigSplitVT, V,
9361 DAG.getIntPtrConstant(0, DL));
9362 HiV = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, OrigSplitVT, V,
9363 DAG.getIntPtrConstant(OrigSplitNumElements, DL));
9366 SmallVector<SDValue, 16> LoOps, HiOps;
9367 for (int i = 0; i < OrigSplitNumElements; ++i) {
9368 LoOps.push_back(BV->getOperand(i));
9369 HiOps.push_back(BV->getOperand(i + OrigSplitNumElements));
9371 LoV = DAG.getNode(ISD::BUILD_VECTOR, DL, OrigSplitVT, LoOps);
9372 HiV = DAG.getNode(ISD::BUILD_VECTOR, DL, OrigSplitVT, HiOps);
9374 return std::make_pair(DAG.getBitcast(SplitVT, LoV),
9375 DAG.getBitcast(SplitVT, HiV));
9378 SDValue LoV1, HiV1, LoV2, HiV2;
9379 std::tie(LoV1, HiV1) = SplitVector(V1);
9380 std::tie(LoV2, HiV2) = SplitVector(V2);
9382 // Now create two 4-way blends of these half-width vectors.
9383 auto HalfBlend = [&](ArrayRef<int> HalfMask) {
9384 bool UseLoV1 = false, UseHiV1 = false, UseLoV2 = false, UseHiV2 = false;
9385 SmallVector<int, 32> V1BlendMask, V2BlendMask, BlendMask;
9386 for (int i = 0; i < SplitNumElements; ++i) {
9387 int M = HalfMask[i];
9388 if (M >= NumElements) {
9389 if (M >= NumElements + SplitNumElements)
9393 V2BlendMask.push_back(M - NumElements);
9394 V1BlendMask.push_back(-1);
9395 BlendMask.push_back(SplitNumElements + i);
9396 } else if (M >= 0) {
9397 if (M >= SplitNumElements)
9401 V2BlendMask.push_back(-1);
9402 V1BlendMask.push_back(M);
9403 BlendMask.push_back(i);
9405 V2BlendMask.push_back(-1);
9406 V1BlendMask.push_back(-1);
9407 BlendMask.push_back(-1);
9411 // Because the lowering happens after all combining takes place, we need to
9412 // manually combine these blend masks as much as possible so that we create
9413 // a minimal number of high-level vector shuffle nodes.
9415 // First try just blending the halves of V1 or V2.
9416 if (!UseLoV1 && !UseHiV1 && !UseLoV2 && !UseHiV2)
9417 return DAG.getUNDEF(SplitVT);
9418 if (!UseLoV2 && !UseHiV2)
9419 return DAG.getVectorShuffle(SplitVT, DL, LoV1, HiV1, V1BlendMask);
9420 if (!UseLoV1 && !UseHiV1)
9421 return DAG.getVectorShuffle(SplitVT, DL, LoV2, HiV2, V2BlendMask);
9423 SDValue V1Blend, V2Blend;
9424 if (UseLoV1 && UseHiV1) {
9426 DAG.getVectorShuffle(SplitVT, DL, LoV1, HiV1, V1BlendMask);
9428 // We only use half of V1 so map the usage down into the final blend mask.
9429 V1Blend = UseLoV1 ? LoV1 : HiV1;
9430 for (int i = 0; i < SplitNumElements; ++i)
9431 if (BlendMask[i] >= 0 && BlendMask[i] < SplitNumElements)
9432 BlendMask[i] = V1BlendMask[i] - (UseLoV1 ? 0 : SplitNumElements);
9434 if (UseLoV2 && UseHiV2) {
9436 DAG.getVectorShuffle(SplitVT, DL, LoV2, HiV2, V2BlendMask);
9438 // We only use half of V2 so map the usage down into the final blend mask.
9439 V2Blend = UseLoV2 ? LoV2 : HiV2;
9440 for (int i = 0; i < SplitNumElements; ++i)
9441 if (BlendMask[i] >= SplitNumElements)
9442 BlendMask[i] = V2BlendMask[i] + (UseLoV2 ? SplitNumElements : 0);
9444 return DAG.getVectorShuffle(SplitVT, DL, V1Blend, V2Blend, BlendMask);
9446 SDValue Lo = HalfBlend(LoMask);
9447 SDValue Hi = HalfBlend(HiMask);
9448 return DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, Lo, Hi);
9451 /// \brief Either split a vector in halves or decompose the shuffles and the
9454 /// This is provided as a good fallback for many lowerings of non-single-input
9455 /// shuffles with more than one 128-bit lane. In those cases, we want to select
9456 /// between splitting the shuffle into 128-bit components and stitching those
9457 /// back together vs. extracting the single-input shuffles and blending those
9459 static SDValue lowerVectorShuffleAsSplitOrBlend(SDLoc DL, MVT VT, SDValue V1,
9460 SDValue V2, ArrayRef<int> Mask,
9461 SelectionDAG &DAG) {
9462 assert(!isSingleInputShuffleMask(Mask) && "This routine must not be used to "
9463 "lower single-input shuffles as it "
9464 "could then recurse on itself.");
9465 int Size = Mask.size();
9467 // If this can be modeled as a broadcast of two elements followed by a blend,
9468 // prefer that lowering. This is especially important because broadcasts can
9469 // often fold with memory operands.
9470 auto DoBothBroadcast = [&] {
9471 int V1BroadcastIdx = -1, V2BroadcastIdx = -1;
9474 if (V2BroadcastIdx == -1)
9475 V2BroadcastIdx = M - Size;
9476 else if (M - Size != V2BroadcastIdx)
9478 } else if (M >= 0) {
9479 if (V1BroadcastIdx == -1)
9481 else if (M != V1BroadcastIdx)
9486 if (DoBothBroadcast())
9487 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, VT, V1, V2, Mask,
9490 // If the inputs all stem from a single 128-bit lane of each input, then we
9491 // split them rather than blending because the split will decompose to
9492 // unusually few instructions.
9493 int LaneCount = VT.getSizeInBits() / 128;
9494 int LaneSize = Size / LaneCount;
9495 SmallBitVector LaneInputs[2];
9496 LaneInputs[0].resize(LaneCount, false);
9497 LaneInputs[1].resize(LaneCount, false);
9498 for (int i = 0; i < Size; ++i)
9500 LaneInputs[Mask[i] / Size][(Mask[i] % Size) / LaneSize] = true;
9501 if (LaneInputs[0].count() <= 1 && LaneInputs[1].count() <= 1)
9502 return splitAndLowerVectorShuffle(DL, VT, V1, V2, Mask, DAG);
9504 // Otherwise, just fall back to decomposed shuffles and a blend. This requires
9505 // that the decomposed single-input shuffles don't end up here.
9506 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, VT, V1, V2, Mask, DAG);
9509 /// \brief Lower a vector shuffle crossing multiple 128-bit lanes as
9510 /// a permutation and blend of those lanes.
9512 /// This essentially blends the out-of-lane inputs to each lane into the lane
9513 /// from a permuted copy of the vector. This lowering strategy results in four
9514 /// instructions in the worst case for a single-input cross lane shuffle which
9515 /// is lower than any other fully general cross-lane shuffle strategy I'm aware
9516 /// of. Special cases for each particular shuffle pattern should be handled
9517 /// prior to trying this lowering.
9518 static SDValue lowerVectorShuffleAsLanePermuteAndBlend(SDLoc DL, MVT VT,
9519 SDValue V1, SDValue V2,
9521 SelectionDAG &DAG) {
9522 // FIXME: This should probably be generalized for 512-bit vectors as well.
9523 assert(VT.getSizeInBits() == 256 && "Only for 256-bit vector shuffles!");
9524 int LaneSize = Mask.size() / 2;
9526 // If there are only inputs from one 128-bit lane, splitting will in fact be
9527 // less expensive. The flags track whether the given lane contains an element
9528 // that crosses to another lane.
9529 bool LaneCrossing[2] = {false, false};
9530 for (int i = 0, Size = Mask.size(); i < Size; ++i)
9531 if (Mask[i] >= 0 && (Mask[i] % Size) / LaneSize != i / LaneSize)
9532 LaneCrossing[(Mask[i] % Size) / LaneSize] = true;
9533 if (!LaneCrossing[0] || !LaneCrossing[1])
9534 return splitAndLowerVectorShuffle(DL, VT, V1, V2, Mask, DAG);
9536 if (isSingleInputShuffleMask(Mask)) {
9537 SmallVector<int, 32> FlippedBlendMask;
9538 for (int i = 0, Size = Mask.size(); i < Size; ++i)
9539 FlippedBlendMask.push_back(
9540 Mask[i] < 0 ? -1 : (((Mask[i] % Size) / LaneSize == i / LaneSize)
9542 : Mask[i] % LaneSize +
9543 (i / LaneSize) * LaneSize + Size));
9545 // Flip the vector, and blend the results which should now be in-lane. The
9546 // VPERM2X128 mask uses the low 2 bits for the low source and bits 4 and
9547 // 5 for the high source. The value 3 selects the high half of source 2 and
9548 // the value 2 selects the low half of source 2. We only use source 2 to
9549 // allow folding it into a memory operand.
9550 unsigned PERMMask = 3 | 2 << 4;
9551 SDValue Flipped = DAG.getNode(X86ISD::VPERM2X128, DL, VT, DAG.getUNDEF(VT),
9552 V1, DAG.getConstant(PERMMask, DL, MVT::i8));
9553 return DAG.getVectorShuffle(VT, DL, V1, Flipped, FlippedBlendMask);
9556 // This now reduces to two single-input shuffles of V1 and V2 which at worst
9557 // will be handled by the above logic and a blend of the results, much like
9558 // other patterns in AVX.
9559 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, VT, V1, V2, Mask, DAG);
9562 /// \brief Handle lowering 2-lane 128-bit shuffles.
9563 static SDValue lowerV2X128VectorShuffle(SDLoc DL, MVT VT, SDValue V1,
9564 SDValue V2, ArrayRef<int> Mask,
9565 const X86Subtarget *Subtarget,
9566 SelectionDAG &DAG) {
9567 // TODO: If minimizing size and one of the inputs is a zero vector and the
9568 // the zero vector has only one use, we could use a VPERM2X128 to save the
9569 // instruction bytes needed to explicitly generate the zero vector.
9571 // Blends are faster and handle all the non-lane-crossing cases.
9572 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, VT, V1, V2, Mask,
9576 bool IsV1Zero = ISD::isBuildVectorAllZeros(V1.getNode());
9577 bool IsV2Zero = ISD::isBuildVectorAllZeros(V2.getNode());
9579 // If either input operand is a zero vector, use VPERM2X128 because its mask
9580 // allows us to replace the zero input with an implicit zero.
9581 if (!IsV1Zero && !IsV2Zero) {
9582 // Check for patterns which can be matched with a single insert of a 128-bit
9584 bool OnlyUsesV1 = isShuffleEquivalent(V1, V2, Mask, {0, 1, 0, 1});
9585 if (OnlyUsesV1 || isShuffleEquivalent(V1, V2, Mask, {0, 1, 4, 5})) {
9586 MVT SubVT = MVT::getVectorVT(VT.getVectorElementType(),
9587 VT.getVectorNumElements() / 2);
9588 SDValue LoV = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, SubVT, V1,
9589 DAG.getIntPtrConstant(0, DL));
9590 SDValue HiV = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, SubVT,
9591 OnlyUsesV1 ? V1 : V2,
9592 DAG.getIntPtrConstant(0, DL));
9593 return DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, LoV, HiV);
9597 // Otherwise form a 128-bit permutation. After accounting for undefs,
9598 // convert the 64-bit shuffle mask selection values into 128-bit
9599 // selection bits by dividing the indexes by 2 and shifting into positions
9600 // defined by a vperm2*128 instruction's immediate control byte.
9602 // The immediate permute control byte looks like this:
9603 // [1:0] - select 128 bits from sources for low half of destination
9605 // [3] - zero low half of destination
9606 // [5:4] - select 128 bits from sources for high half of destination
9608 // [7] - zero high half of destination
9610 int MaskLO = Mask[0];
9611 if (MaskLO == SM_SentinelUndef)
9612 MaskLO = Mask[1] == SM_SentinelUndef ? 0 : Mask[1];
9614 int MaskHI = Mask[2];
9615 if (MaskHI == SM_SentinelUndef)
9616 MaskHI = Mask[3] == SM_SentinelUndef ? 0 : Mask[3];
9618 unsigned PermMask = MaskLO / 2 | (MaskHI / 2) << 4;
9620 // If either input is a zero vector, replace it with an undef input.
9621 // Shuffle mask values < 4 are selecting elements of V1.
9622 // Shuffle mask values >= 4 are selecting elements of V2.
9623 // Adjust each half of the permute mask by clearing the half that was
9624 // selecting the zero vector and setting the zero mask bit.
9626 V1 = DAG.getUNDEF(VT);
9628 PermMask = (PermMask & 0xf0) | 0x08;
9630 PermMask = (PermMask & 0x0f) | 0x80;
9633 V2 = DAG.getUNDEF(VT);
9635 PermMask = (PermMask & 0xf0) | 0x08;
9637 PermMask = (PermMask & 0x0f) | 0x80;
9640 return DAG.getNode(X86ISD::VPERM2X128, DL, VT, V1, V2,
9641 DAG.getConstant(PermMask, DL, MVT::i8));
9644 /// \brief Lower a vector shuffle by first fixing the 128-bit lanes and then
9645 /// shuffling each lane.
9647 /// This will only succeed when the result of fixing the 128-bit lanes results
9648 /// in a single-input non-lane-crossing shuffle with a repeating shuffle mask in
9649 /// each 128-bit lanes. This handles many cases where we can quickly blend away
9650 /// the lane crosses early and then use simpler shuffles within each lane.
9652 /// FIXME: It might be worthwhile at some point to support this without
9653 /// requiring the 128-bit lane-relative shuffles to be repeating, but currently
9654 /// in x86 only floating point has interesting non-repeating shuffles, and even
9655 /// those are still *marginally* more expensive.
9656 static SDValue lowerVectorShuffleByMerging128BitLanes(
9657 SDLoc DL, MVT VT, SDValue V1, SDValue V2, ArrayRef<int> Mask,
9658 const X86Subtarget *Subtarget, SelectionDAG &DAG) {
9659 assert(!isSingleInputShuffleMask(Mask) &&
9660 "This is only useful with multiple inputs.");
9662 int Size = Mask.size();
9663 int LaneSize = 128 / VT.getScalarSizeInBits();
9664 int NumLanes = Size / LaneSize;
9665 assert(NumLanes > 1 && "Only handles 256-bit and wider shuffles.");
9667 // See if we can build a hypothetical 128-bit lane-fixing shuffle mask. Also
9668 // check whether the in-128-bit lane shuffles share a repeating pattern.
9669 SmallVector<int, 4> Lanes;
9670 Lanes.resize(NumLanes, -1);
9671 SmallVector<int, 4> InLaneMask;
9672 InLaneMask.resize(LaneSize, -1);
9673 for (int i = 0; i < Size; ++i) {
9677 int j = i / LaneSize;
9680 // First entry we've seen for this lane.
9681 Lanes[j] = Mask[i] / LaneSize;
9682 } else if (Lanes[j] != Mask[i] / LaneSize) {
9683 // This doesn't match the lane selected previously!
9687 // Check that within each lane we have a consistent shuffle mask.
9688 int k = i % LaneSize;
9689 if (InLaneMask[k] < 0) {
9690 InLaneMask[k] = Mask[i] % LaneSize;
9691 } else if (InLaneMask[k] != Mask[i] % LaneSize) {
9692 // This doesn't fit a repeating in-lane mask.
9697 // First shuffle the lanes into place.
9698 MVT LaneVT = MVT::getVectorVT(VT.isFloatingPoint() ? MVT::f64 : MVT::i64,
9699 VT.getSizeInBits() / 64);
9700 SmallVector<int, 8> LaneMask;
9701 LaneMask.resize(NumLanes * 2, -1);
9702 for (int i = 0; i < NumLanes; ++i)
9703 if (Lanes[i] >= 0) {
9704 LaneMask[2 * i + 0] = 2*Lanes[i] + 0;
9705 LaneMask[2 * i + 1] = 2*Lanes[i] + 1;
9708 V1 = DAG.getBitcast(LaneVT, V1);
9709 V2 = DAG.getBitcast(LaneVT, V2);
9710 SDValue LaneShuffle = DAG.getVectorShuffle(LaneVT, DL, V1, V2, LaneMask);
9712 // Cast it back to the type we actually want.
9713 LaneShuffle = DAG.getBitcast(VT, LaneShuffle);
9715 // Now do a simple shuffle that isn't lane crossing.
9716 SmallVector<int, 8> NewMask;
9717 NewMask.resize(Size, -1);
9718 for (int i = 0; i < Size; ++i)
9720 NewMask[i] = (i / LaneSize) * LaneSize + Mask[i] % LaneSize;
9721 assert(!is128BitLaneCrossingShuffleMask(VT, NewMask) &&
9722 "Must not introduce lane crosses at this point!");
9724 return DAG.getVectorShuffle(VT, DL, LaneShuffle, DAG.getUNDEF(VT), NewMask);
9727 /// \brief Test whether the specified input (0 or 1) is in-place blended by the
9730 /// This returns true if the elements from a particular input are already in the
9731 /// slot required by the given mask and require no permutation.
9732 static bool isShuffleMaskInputInPlace(int Input, ArrayRef<int> Mask) {
9733 assert((Input == 0 || Input == 1) && "Only two inputs to shuffles.");
9734 int Size = Mask.size();
9735 for (int i = 0; i < Size; ++i)
9736 if (Mask[i] >= 0 && Mask[i] / Size == Input && Mask[i] % Size != i)
9742 static SDValue lowerVectorShuffleWithSHUFPD(SDLoc DL, MVT VT,
9743 ArrayRef<int> Mask, SDValue V1,
9744 SDValue V2, SelectionDAG &DAG) {
9746 // Mask for V8F64: 0/1, 8/9, 2/3, 10/11, 4/5, ..
9747 // Mask for V4F64; 0/1, 4/5, 2/3, 6/7..
9748 assert(VT.getScalarSizeInBits() == 64 && "Unexpected data type for VSHUFPD");
9749 int NumElts = VT.getVectorNumElements();
9750 bool ShufpdMask = true;
9751 bool CommutableMask = true;
9752 unsigned Immediate = 0;
9753 for (int i = 0; i < NumElts; ++i) {
9756 int Val = (i & 6) + NumElts * (i & 1);
9757 int CommutVal = (i & 0xe) + NumElts * ((i & 1)^1);
9758 if (Mask[i] < Val || Mask[i] > Val + 1)
9760 if (Mask[i] < CommutVal || Mask[i] > CommutVal + 1)
9761 CommutableMask = false;
9762 Immediate |= (Mask[i] % 2) << i;
9765 return DAG.getNode(X86ISD::SHUFP, DL, VT, V1, V2,
9766 DAG.getConstant(Immediate, DL, MVT::i8));
9768 return DAG.getNode(X86ISD::SHUFP, DL, VT, V2, V1,
9769 DAG.getConstant(Immediate, DL, MVT::i8));
9773 /// \brief Handle lowering of 4-lane 64-bit floating point shuffles.
9775 /// Also ends up handling lowering of 4-lane 64-bit integer shuffles when AVX2
9776 /// isn't available.
9777 static SDValue lowerV4F64VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
9778 const X86Subtarget *Subtarget,
9779 SelectionDAG &DAG) {
9781 assert(V1.getSimpleValueType() == MVT::v4f64 && "Bad operand type!");
9782 assert(V2.getSimpleValueType() == MVT::v4f64 && "Bad operand type!");
9783 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
9784 ArrayRef<int> Mask = SVOp->getMask();
9785 assert(Mask.size() == 4 && "Unexpected mask size for v4 shuffle!");
9787 SmallVector<int, 4> WidenedMask;
9788 if (canWidenShuffleElements(Mask, WidenedMask))
9789 return lowerV2X128VectorShuffle(DL, MVT::v4f64, V1, V2, Mask, Subtarget,
9792 if (isSingleInputShuffleMask(Mask)) {
9793 // Check for being able to broadcast a single element.
9794 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v4f64, V1,
9795 Mask, Subtarget, DAG))
9798 // Use low duplicate instructions for masks that match their pattern.
9799 if (isShuffleEquivalent(V1, V2, Mask, {0, 0, 2, 2}))
9800 return DAG.getNode(X86ISD::MOVDDUP, DL, MVT::v4f64, V1);
9802 if (!is128BitLaneCrossingShuffleMask(MVT::v4f64, Mask)) {
9803 // Non-half-crossing single input shuffles can be lowerid with an
9804 // interleaved permutation.
9805 unsigned VPERMILPMask = (Mask[0] == 1) | ((Mask[1] == 1) << 1) |
9806 ((Mask[2] == 3) << 2) | ((Mask[3] == 3) << 3);
9807 return DAG.getNode(X86ISD::VPERMILPI, DL, MVT::v4f64, V1,
9808 DAG.getConstant(VPERMILPMask, DL, MVT::i8));
9811 // With AVX2 we have direct support for this permutation.
9812 if (Subtarget->hasAVX2())
9813 return DAG.getNode(X86ISD::VPERMI, DL, MVT::v4f64, V1,
9814 getV4X86ShuffleImm8ForMask(Mask, DL, DAG));
9816 // Otherwise, fall back.
9817 return lowerVectorShuffleAsLanePermuteAndBlend(DL, MVT::v4f64, V1, V2, Mask,
9821 // X86 has dedicated unpack instructions that can handle specific blend
9822 // operations: UNPCKH and UNPCKL.
9823 if (isShuffleEquivalent(V1, V2, Mask, {0, 4, 2, 6}))
9824 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v4f64, V1, V2);
9825 if (isShuffleEquivalent(V1, V2, Mask, {1, 5, 3, 7}))
9826 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v4f64, V1, V2);
9827 if (isShuffleEquivalent(V1, V2, Mask, {4, 0, 6, 2}))
9828 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v4f64, V2, V1);
9829 if (isShuffleEquivalent(V1, V2, Mask, {5, 1, 7, 3}))
9830 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v4f64, V2, V1);
9832 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v4f64, V1, V2, Mask,
9836 // Check if the blend happens to exactly fit that of SHUFPD.
9838 lowerVectorShuffleWithSHUFPD(DL, MVT::v4f64, Mask, V1, V2, DAG))
9841 // Try to simplify this by merging 128-bit lanes to enable a lane-based
9842 // shuffle. However, if we have AVX2 and either inputs are already in place,
9843 // we will be able to shuffle even across lanes the other input in a single
9844 // instruction so skip this pattern.
9845 if (!(Subtarget->hasAVX2() && (isShuffleMaskInputInPlace(0, Mask) ||
9846 isShuffleMaskInputInPlace(1, Mask))))
9847 if (SDValue Result = lowerVectorShuffleByMerging128BitLanes(
9848 DL, MVT::v4f64, V1, V2, Mask, Subtarget, DAG))
9851 // If we have AVX2 then we always want to lower with a blend because an v4 we
9852 // can fully permute the elements.
9853 if (Subtarget->hasAVX2())
9854 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v4f64, V1, V2,
9857 // Otherwise fall back on generic lowering.
9858 return lowerVectorShuffleAsSplitOrBlend(DL, MVT::v4f64, V1, V2, Mask, DAG);
9861 /// \brief Handle lowering of 4-lane 64-bit integer shuffles.
9863 /// This routine is only called when we have AVX2 and thus a reasonable
9864 /// instruction set for v4i64 shuffling..
9865 static SDValue lowerV4I64VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
9866 const X86Subtarget *Subtarget,
9867 SelectionDAG &DAG) {
9869 assert(V1.getSimpleValueType() == MVT::v4i64 && "Bad operand type!");
9870 assert(V2.getSimpleValueType() == MVT::v4i64 && "Bad operand type!");
9871 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
9872 ArrayRef<int> Mask = SVOp->getMask();
9873 assert(Mask.size() == 4 && "Unexpected mask size for v4 shuffle!");
9874 assert(Subtarget->hasAVX2() && "We can only lower v4i64 with AVX2!");
9876 SmallVector<int, 4> WidenedMask;
9877 if (canWidenShuffleElements(Mask, WidenedMask))
9878 return lowerV2X128VectorShuffle(DL, MVT::v4i64, V1, V2, Mask, Subtarget,
9881 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v4i64, V1, V2, Mask,
9885 // Check for being able to broadcast a single element.
9886 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v4i64, V1,
9887 Mask, Subtarget, DAG))
9890 // When the shuffle is mirrored between the 128-bit lanes of the unit, we can
9891 // use lower latency instructions that will operate on both 128-bit lanes.
9892 SmallVector<int, 2> RepeatedMask;
9893 if (is128BitLaneRepeatedShuffleMask(MVT::v4i64, Mask, RepeatedMask)) {
9894 if (isSingleInputShuffleMask(Mask)) {
9895 int PSHUFDMask[] = {-1, -1, -1, -1};
9896 for (int i = 0; i < 2; ++i)
9897 if (RepeatedMask[i] >= 0) {
9898 PSHUFDMask[2 * i] = 2 * RepeatedMask[i];
9899 PSHUFDMask[2 * i + 1] = 2 * RepeatedMask[i] + 1;
9901 return DAG.getBitcast(
9903 DAG.getNode(X86ISD::PSHUFD, DL, MVT::v8i32,
9904 DAG.getBitcast(MVT::v8i32, V1),
9905 getV4X86ShuffleImm8ForMask(PSHUFDMask, DL, DAG)));
9909 // AVX2 provides a direct instruction for permuting a single input across
9911 if (isSingleInputShuffleMask(Mask))
9912 return DAG.getNode(X86ISD::VPERMI, DL, MVT::v4i64, V1,
9913 getV4X86ShuffleImm8ForMask(Mask, DL, DAG));
9915 // Try to use shift instructions.
9917 lowerVectorShuffleAsShift(DL, MVT::v4i64, V1, V2, Mask, DAG))
9920 // Use dedicated unpack instructions for masks that match their pattern.
9921 if (isShuffleEquivalent(V1, V2, Mask, {0, 4, 2, 6}))
9922 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v4i64, V1, V2);
9923 if (isShuffleEquivalent(V1, V2, Mask, {1, 5, 3, 7}))
9924 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v4i64, V1, V2);
9925 if (isShuffleEquivalent(V1, V2, Mask, {4, 0, 6, 2}))
9926 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v4i64, V2, V1);
9927 if (isShuffleEquivalent(V1, V2, Mask, {5, 1, 7, 3}))
9928 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v4i64, V2, V1);
9930 // Try to simplify this by merging 128-bit lanes to enable a lane-based
9931 // shuffle. However, if we have AVX2 and either inputs are already in place,
9932 // we will be able to shuffle even across lanes the other input in a single
9933 // instruction so skip this pattern.
9934 if (!(Subtarget->hasAVX2() && (isShuffleMaskInputInPlace(0, Mask) ||
9935 isShuffleMaskInputInPlace(1, Mask))))
9936 if (SDValue Result = lowerVectorShuffleByMerging128BitLanes(
9937 DL, MVT::v4i64, V1, V2, Mask, Subtarget, DAG))
9940 // Otherwise fall back on generic blend lowering.
9941 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v4i64, V1, V2,
9945 /// \brief Handle lowering of 8-lane 32-bit floating point shuffles.
9947 /// Also ends up handling lowering of 8-lane 32-bit integer shuffles when AVX2
9948 /// isn't available.
9949 static SDValue lowerV8F32VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
9950 const X86Subtarget *Subtarget,
9951 SelectionDAG &DAG) {
9953 assert(V1.getSimpleValueType() == MVT::v8f32 && "Bad operand type!");
9954 assert(V2.getSimpleValueType() == MVT::v8f32 && "Bad operand type!");
9955 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
9956 ArrayRef<int> Mask = SVOp->getMask();
9957 assert(Mask.size() == 8 && "Unexpected mask size for v8 shuffle!");
9959 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v8f32, V1, V2, Mask,
9963 // Check for being able to broadcast a single element.
9964 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v8f32, V1,
9965 Mask, Subtarget, DAG))
9968 // If the shuffle mask is repeated in each 128-bit lane, we have many more
9969 // options to efficiently lower the shuffle.
9970 SmallVector<int, 4> RepeatedMask;
9971 if (is128BitLaneRepeatedShuffleMask(MVT::v8f32, Mask, RepeatedMask)) {
9972 assert(RepeatedMask.size() == 4 &&
9973 "Repeated masks must be half the mask width!");
9975 // Use even/odd duplicate instructions for masks that match their pattern.
9976 if (isShuffleEquivalent(V1, V2, Mask, {0, 0, 2, 2, 4, 4, 6, 6}))
9977 return DAG.getNode(X86ISD::MOVSLDUP, DL, MVT::v8f32, V1);
9978 if (isShuffleEquivalent(V1, V2, Mask, {1, 1, 3, 3, 5, 5, 7, 7}))
9979 return DAG.getNode(X86ISD::MOVSHDUP, DL, MVT::v8f32, V1);
9981 if (isSingleInputShuffleMask(Mask))
9982 return DAG.getNode(X86ISD::VPERMILPI, DL, MVT::v8f32, V1,
9983 getV4X86ShuffleImm8ForMask(RepeatedMask, DL, DAG));
9985 // Use dedicated unpack instructions for masks that match their pattern.
9986 if (isShuffleEquivalent(V1, V2, Mask, {0, 8, 1, 9, 4, 12, 5, 13}))
9987 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v8f32, V1, V2);
9988 if (isShuffleEquivalent(V1, V2, Mask, {2, 10, 3, 11, 6, 14, 7, 15}))
9989 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v8f32, V1, V2);
9990 if (isShuffleEquivalent(V1, V2, Mask, {8, 0, 9, 1, 12, 4, 13, 5}))
9991 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v8f32, V2, V1);
9992 if (isShuffleEquivalent(V1, V2, Mask, {10, 2, 11, 3, 14, 6, 15, 7}))
9993 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v8f32, V2, V1);
9995 // Otherwise, fall back to a SHUFPS sequence. Here it is important that we
9996 // have already handled any direct blends. We also need to squash the
9997 // repeated mask into a simulated v4f32 mask.
9998 for (int i = 0; i < 4; ++i)
9999 if (RepeatedMask[i] >= 8)
10000 RepeatedMask[i] -= 4;
10001 return lowerVectorShuffleWithSHUFPS(DL, MVT::v8f32, RepeatedMask, V1, V2, DAG);
10004 // If we have a single input shuffle with different shuffle patterns in the
10005 // two 128-bit lanes use the variable mask to VPERMILPS.
10006 if (isSingleInputShuffleMask(Mask)) {
10007 SDValue VPermMask[8];
10008 for (int i = 0; i < 8; ++i)
10009 VPermMask[i] = Mask[i] < 0 ? DAG.getUNDEF(MVT::i32)
10010 : DAG.getConstant(Mask[i], DL, MVT::i32);
10011 if (!is128BitLaneCrossingShuffleMask(MVT::v8f32, Mask))
10012 return DAG.getNode(
10013 X86ISD::VPERMILPV, DL, MVT::v8f32, V1,
10014 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v8i32, VPermMask));
10016 if (Subtarget->hasAVX2())
10017 return DAG.getNode(
10018 X86ISD::VPERMV, DL, MVT::v8f32,
10019 DAG.getBitcast(MVT::v8f32, DAG.getNode(ISD::BUILD_VECTOR, DL,
10020 MVT::v8i32, VPermMask)),
10023 // Otherwise, fall back.
10024 return lowerVectorShuffleAsLanePermuteAndBlend(DL, MVT::v8f32, V1, V2, Mask,
10028 // Try to simplify this by merging 128-bit lanes to enable a lane-based
10030 if (SDValue Result = lowerVectorShuffleByMerging128BitLanes(
10031 DL, MVT::v8f32, V1, V2, Mask, Subtarget, DAG))
10034 // If we have AVX2 then we always want to lower with a blend because at v8 we
10035 // can fully permute the elements.
10036 if (Subtarget->hasAVX2())
10037 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v8f32, V1, V2,
10040 // Otherwise fall back on generic lowering.
10041 return lowerVectorShuffleAsSplitOrBlend(DL, MVT::v8f32, V1, V2, Mask, DAG);
10044 /// \brief Handle lowering of 8-lane 32-bit integer shuffles.
10046 /// This routine is only called when we have AVX2 and thus a reasonable
10047 /// instruction set for v8i32 shuffling..
10048 static SDValue lowerV8I32VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10049 const X86Subtarget *Subtarget,
10050 SelectionDAG &DAG) {
10052 assert(V1.getSimpleValueType() == MVT::v8i32 && "Bad operand type!");
10053 assert(V2.getSimpleValueType() == MVT::v8i32 && "Bad operand type!");
10054 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10055 ArrayRef<int> Mask = SVOp->getMask();
10056 assert(Mask.size() == 8 && "Unexpected mask size for v8 shuffle!");
10057 assert(Subtarget->hasAVX2() && "We can only lower v8i32 with AVX2!");
10059 // Whenever we can lower this as a zext, that instruction is strictly faster
10060 // than any alternative. It also allows us to fold memory operands into the
10061 // shuffle in many cases.
10062 if (SDValue ZExt = lowerVectorShuffleAsZeroOrAnyExtend(DL, MVT::v8i32, V1, V2,
10063 Mask, Subtarget, DAG))
10066 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v8i32, V1, V2, Mask,
10070 // Check for being able to broadcast a single element.
10071 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v8i32, V1,
10072 Mask, Subtarget, DAG))
10075 // If the shuffle mask is repeated in each 128-bit lane we can use more
10076 // efficient instructions that mirror the shuffles across the two 128-bit
10078 SmallVector<int, 4> RepeatedMask;
10079 if (is128BitLaneRepeatedShuffleMask(MVT::v8i32, Mask, RepeatedMask)) {
10080 assert(RepeatedMask.size() == 4 && "Unexpected repeated mask size!");
10081 if (isSingleInputShuffleMask(Mask))
10082 return DAG.getNode(X86ISD::PSHUFD, DL, MVT::v8i32, V1,
10083 getV4X86ShuffleImm8ForMask(RepeatedMask, DL, DAG));
10085 // Use dedicated unpack instructions for masks that match their pattern.
10086 if (isShuffleEquivalent(V1, V2, Mask, {0, 8, 1, 9, 4, 12, 5, 13}))
10087 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v8i32, V1, V2);
10088 if (isShuffleEquivalent(V1, V2, Mask, {2, 10, 3, 11, 6, 14, 7, 15}))
10089 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v8i32, V1, V2);
10090 if (isShuffleEquivalent(V1, V2, Mask, {8, 0, 9, 1, 12, 4, 13, 5}))
10091 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v8i32, V2, V1);
10092 if (isShuffleEquivalent(V1, V2, Mask, {10, 2, 11, 3, 14, 6, 15, 7}))
10093 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v8i32, V2, V1);
10096 // Try to use shift instructions.
10097 if (SDValue Shift =
10098 lowerVectorShuffleAsShift(DL, MVT::v8i32, V1, V2, Mask, DAG))
10101 if (SDValue Rotate = lowerVectorShuffleAsByteRotate(
10102 DL, MVT::v8i32, V1, V2, Mask, Subtarget, DAG))
10105 // If the shuffle patterns aren't repeated but it is a single input, directly
10106 // generate a cross-lane VPERMD instruction.
10107 if (isSingleInputShuffleMask(Mask)) {
10108 SDValue VPermMask[8];
10109 for (int i = 0; i < 8; ++i)
10110 VPermMask[i] = Mask[i] < 0 ? DAG.getUNDEF(MVT::i32)
10111 : DAG.getConstant(Mask[i], DL, MVT::i32);
10112 return DAG.getNode(
10113 X86ISD::VPERMV, DL, MVT::v8i32,
10114 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v8i32, VPermMask), V1);
10117 // Try to simplify this by merging 128-bit lanes to enable a lane-based
10119 if (SDValue Result = lowerVectorShuffleByMerging128BitLanes(
10120 DL, MVT::v8i32, V1, V2, Mask, Subtarget, DAG))
10123 // Otherwise fall back on generic blend lowering.
10124 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v8i32, V1, V2,
10128 /// \brief Handle lowering of 16-lane 16-bit integer shuffles.
10130 /// This routine is only called when we have AVX2 and thus a reasonable
10131 /// instruction set for v16i16 shuffling..
10132 static SDValue lowerV16I16VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10133 const X86Subtarget *Subtarget,
10134 SelectionDAG &DAG) {
10136 assert(V1.getSimpleValueType() == MVT::v16i16 && "Bad operand type!");
10137 assert(V2.getSimpleValueType() == MVT::v16i16 && "Bad operand type!");
10138 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10139 ArrayRef<int> Mask = SVOp->getMask();
10140 assert(Mask.size() == 16 && "Unexpected mask size for v16 shuffle!");
10141 assert(Subtarget->hasAVX2() && "We can only lower v16i16 with AVX2!");
10143 // Whenever we can lower this as a zext, that instruction is strictly faster
10144 // than any alternative. It also allows us to fold memory operands into the
10145 // shuffle in many cases.
10146 if (SDValue ZExt = lowerVectorShuffleAsZeroOrAnyExtend(DL, MVT::v16i16, V1, V2,
10147 Mask, Subtarget, DAG))
10150 // Check for being able to broadcast a single element.
10151 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v16i16, V1,
10152 Mask, Subtarget, DAG))
10155 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v16i16, V1, V2, Mask,
10159 // Use dedicated unpack instructions for masks that match their pattern.
10160 if (isShuffleEquivalent(V1, V2, Mask,
10161 {// First 128-bit lane:
10162 0, 16, 1, 17, 2, 18, 3, 19,
10163 // Second 128-bit lane:
10164 8, 24, 9, 25, 10, 26, 11, 27}))
10165 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v16i16, V1, V2);
10166 if (isShuffleEquivalent(V1, V2, Mask,
10167 {// First 128-bit lane:
10168 4, 20, 5, 21, 6, 22, 7, 23,
10169 // Second 128-bit lane:
10170 12, 28, 13, 29, 14, 30, 15, 31}))
10171 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v16i16, V1, V2);
10173 // Try to use shift instructions.
10174 if (SDValue Shift =
10175 lowerVectorShuffleAsShift(DL, MVT::v16i16, V1, V2, Mask, DAG))
10178 // Try to use byte rotation instructions.
10179 if (SDValue Rotate = lowerVectorShuffleAsByteRotate(
10180 DL, MVT::v16i16, V1, V2, Mask, Subtarget, DAG))
10183 if (isSingleInputShuffleMask(Mask)) {
10184 // There are no generalized cross-lane shuffle operations available on i16
10186 if (is128BitLaneCrossingShuffleMask(MVT::v16i16, Mask))
10187 return lowerVectorShuffleAsLanePermuteAndBlend(DL, MVT::v16i16, V1, V2,
10190 SmallVector<int, 8> RepeatedMask;
10191 if (is128BitLaneRepeatedShuffleMask(MVT::v16i16, Mask, RepeatedMask)) {
10192 // As this is a single-input shuffle, the repeated mask should be
10193 // a strictly valid v8i16 mask that we can pass through to the v8i16
10194 // lowering to handle even the v16 case.
10195 return lowerV8I16GeneralSingleInputVectorShuffle(
10196 DL, MVT::v16i16, V1, RepeatedMask, Subtarget, DAG);
10199 SDValue PSHUFBMask[32];
10200 for (int i = 0; i < 16; ++i) {
10201 if (Mask[i] == -1) {
10202 PSHUFBMask[2 * i] = PSHUFBMask[2 * i + 1] = DAG.getUNDEF(MVT::i8);
10206 int M = i < 8 ? Mask[i] : Mask[i] - 8;
10207 assert(M >= 0 && M < 8 && "Invalid single-input mask!");
10208 PSHUFBMask[2 * i] = DAG.getConstant(2 * M, DL, MVT::i8);
10209 PSHUFBMask[2 * i + 1] = DAG.getConstant(2 * M + 1, DL, MVT::i8);
10211 return DAG.getBitcast(MVT::v16i16,
10212 DAG.getNode(X86ISD::PSHUFB, DL, MVT::v32i8,
10213 DAG.getBitcast(MVT::v32i8, V1),
10214 DAG.getNode(ISD::BUILD_VECTOR, DL,
10215 MVT::v32i8, PSHUFBMask)));
10218 // Try to simplify this by merging 128-bit lanes to enable a lane-based
10220 if (SDValue Result = lowerVectorShuffleByMerging128BitLanes(
10221 DL, MVT::v16i16, V1, V2, Mask, Subtarget, DAG))
10224 // Otherwise fall back on generic lowering.
10225 return lowerVectorShuffleAsSplitOrBlend(DL, MVT::v16i16, V1, V2, Mask, DAG);
10228 /// \brief Handle lowering of 32-lane 8-bit integer shuffles.
10230 /// This routine is only called when we have AVX2 and thus a reasonable
10231 /// instruction set for v32i8 shuffling..
10232 static SDValue lowerV32I8VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10233 const X86Subtarget *Subtarget,
10234 SelectionDAG &DAG) {
10236 assert(V1.getSimpleValueType() == MVT::v32i8 && "Bad operand type!");
10237 assert(V2.getSimpleValueType() == MVT::v32i8 && "Bad operand type!");
10238 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10239 ArrayRef<int> Mask = SVOp->getMask();
10240 assert(Mask.size() == 32 && "Unexpected mask size for v32 shuffle!");
10241 assert(Subtarget->hasAVX2() && "We can only lower v32i8 with AVX2!");
10243 // Whenever we can lower this as a zext, that instruction is strictly faster
10244 // than any alternative. It also allows us to fold memory operands into the
10245 // shuffle in many cases.
10246 if (SDValue ZExt = lowerVectorShuffleAsZeroOrAnyExtend(DL, MVT::v32i8, V1, V2,
10247 Mask, Subtarget, DAG))
10250 // Check for being able to broadcast a single element.
10251 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v32i8, V1,
10252 Mask, Subtarget, DAG))
10255 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v32i8, V1, V2, Mask,
10259 // Use dedicated unpack instructions for masks that match their pattern.
10260 // Note that these are repeated 128-bit lane unpacks, not unpacks across all
10262 if (isShuffleEquivalent(
10264 {// First 128-bit lane:
10265 0, 32, 1, 33, 2, 34, 3, 35, 4, 36, 5, 37, 6, 38, 7, 39,
10266 // Second 128-bit lane:
10267 16, 48, 17, 49, 18, 50, 19, 51, 20, 52, 21, 53, 22, 54, 23, 55}))
10268 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v32i8, V1, V2);
10269 if (isShuffleEquivalent(
10271 {// First 128-bit lane:
10272 8, 40, 9, 41, 10, 42, 11, 43, 12, 44, 13, 45, 14, 46, 15, 47,
10273 // Second 128-bit lane:
10274 24, 56, 25, 57, 26, 58, 27, 59, 28, 60, 29, 61, 30, 62, 31, 63}))
10275 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v32i8, V1, V2);
10277 // Try to use shift instructions.
10278 if (SDValue Shift =
10279 lowerVectorShuffleAsShift(DL, MVT::v32i8, V1, V2, Mask, DAG))
10282 // Try to use byte rotation instructions.
10283 if (SDValue Rotate = lowerVectorShuffleAsByteRotate(
10284 DL, MVT::v32i8, V1, V2, Mask, Subtarget, DAG))
10287 if (isSingleInputShuffleMask(Mask)) {
10288 // There are no generalized cross-lane shuffle operations available on i8
10290 if (is128BitLaneCrossingShuffleMask(MVT::v32i8, Mask))
10291 return lowerVectorShuffleAsLanePermuteAndBlend(DL, MVT::v32i8, V1, V2,
10294 SDValue PSHUFBMask[32];
10295 for (int i = 0; i < 32; ++i)
10298 ? DAG.getUNDEF(MVT::i8)
10299 : DAG.getConstant(Mask[i] < 16 ? Mask[i] : Mask[i] - 16, DL,
10302 return DAG.getNode(
10303 X86ISD::PSHUFB, DL, MVT::v32i8, V1,
10304 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v32i8, PSHUFBMask));
10307 // Try to simplify this by merging 128-bit lanes to enable a lane-based
10309 if (SDValue Result = lowerVectorShuffleByMerging128BitLanes(
10310 DL, MVT::v32i8, V1, V2, Mask, Subtarget, DAG))
10313 // Otherwise fall back on generic lowering.
10314 return lowerVectorShuffleAsSplitOrBlend(DL, MVT::v32i8, V1, V2, Mask, DAG);
10317 /// \brief High-level routine to lower various 256-bit x86 vector shuffles.
10319 /// This routine either breaks down the specific type of a 256-bit x86 vector
10320 /// shuffle or splits it into two 128-bit shuffles and fuses the results back
10321 /// together based on the available instructions.
10322 static SDValue lower256BitVectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10323 MVT VT, const X86Subtarget *Subtarget,
10324 SelectionDAG &DAG) {
10326 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10327 ArrayRef<int> Mask = SVOp->getMask();
10329 // If we have a single input to the zero element, insert that into V1 if we
10330 // can do so cheaply.
10331 int NumElts = VT.getVectorNumElements();
10332 int NumV2Elements = std::count_if(Mask.begin(), Mask.end(), [NumElts](int M) {
10333 return M >= NumElts;
10336 if (NumV2Elements == 1 && Mask[0] >= NumElts)
10337 if (SDValue Insertion = lowerVectorShuffleAsElementInsertion(
10338 DL, VT, V1, V2, Mask, Subtarget, DAG))
10341 // There is a really nice hard cut-over between AVX1 and AVX2 that means we can
10342 // check for those subtargets here and avoid much of the subtarget querying in
10343 // the per-vector-type lowering routines. With AVX1 we have essentially *zero*
10344 // ability to manipulate a 256-bit vector with integer types. Since we'll use
10345 // floating point types there eventually, just immediately cast everything to
10346 // a float and operate entirely in that domain.
10347 if (VT.isInteger() && !Subtarget->hasAVX2()) {
10348 int ElementBits = VT.getScalarSizeInBits();
10349 if (ElementBits < 32)
10350 // No floating point type available, decompose into 128-bit vectors.
10351 return splitAndLowerVectorShuffle(DL, VT, V1, V2, Mask, DAG);
10353 MVT FpVT = MVT::getVectorVT(MVT::getFloatingPointVT(ElementBits),
10354 VT.getVectorNumElements());
10355 V1 = DAG.getBitcast(FpVT, V1);
10356 V2 = DAG.getBitcast(FpVT, V2);
10357 return DAG.getBitcast(VT, DAG.getVectorShuffle(FpVT, DL, V1, V2, Mask));
10360 switch (VT.SimpleTy) {
10362 return lowerV4F64VectorShuffle(Op, V1, V2, Subtarget, DAG);
10364 return lowerV4I64VectorShuffle(Op, V1, V2, Subtarget, DAG);
10366 return lowerV8F32VectorShuffle(Op, V1, V2, Subtarget, DAG);
10368 return lowerV8I32VectorShuffle(Op, V1, V2, Subtarget, DAG);
10370 return lowerV16I16VectorShuffle(Op, V1, V2, Subtarget, DAG);
10372 return lowerV32I8VectorShuffle(Op, V1, V2, Subtarget, DAG);
10375 llvm_unreachable("Not a valid 256-bit x86 vector type!");
10379 /// \brief Handle lowering of 8-lane 64-bit floating point shuffles.
10380 static SDValue lowerV8F64VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10381 const X86Subtarget *Subtarget,
10382 SelectionDAG &DAG) {
10384 assert(V1.getSimpleValueType() == MVT::v8f64 && "Bad operand type!");
10385 assert(V2.getSimpleValueType() == MVT::v8f64 && "Bad operand type!");
10386 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10387 ArrayRef<int> Mask = SVOp->getMask();
10388 assert(Mask.size() == 8 && "Unexpected mask size for v8 shuffle!");
10390 // X86 has dedicated unpack instructions that can handle specific blend
10391 // operations: UNPCKH and UNPCKL.
10392 if (isShuffleEquivalent(V1, V2, Mask, {0, 8, 2, 10, 4, 12, 6, 14}))
10393 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v8f64, V1, V2);
10394 if (isShuffleEquivalent(V1, V2, Mask, {1, 9, 3, 11, 5, 13, 7, 15}))
10395 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v8f64, V1, V2);
10397 // FIXME: Implement direct support for this type!
10398 return splitAndLowerVectorShuffle(DL, MVT::v8f64, V1, V2, Mask, DAG);
10401 /// \brief Handle lowering of 16-lane 32-bit floating point shuffles.
10402 static SDValue lowerV16F32VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10403 const X86Subtarget *Subtarget,
10404 SelectionDAG &DAG) {
10406 assert(V1.getSimpleValueType() == MVT::v16f32 && "Bad operand type!");
10407 assert(V2.getSimpleValueType() == MVT::v16f32 && "Bad operand type!");
10408 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10409 ArrayRef<int> Mask = SVOp->getMask();
10410 assert(Mask.size() == 16 && "Unexpected mask size for v16 shuffle!");
10412 // Use dedicated unpack instructions for masks that match their pattern.
10413 if (isShuffleEquivalent(V1, V2, Mask,
10414 {// First 128-bit lane.
10415 0, 16, 1, 17, 4, 20, 5, 21,
10416 // Second 128-bit lane.
10417 8, 24, 9, 25, 12, 28, 13, 29}))
10418 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v16f32, V1, V2);
10419 if (isShuffleEquivalent(V1, V2, Mask,
10420 {// First 128-bit lane.
10421 2, 18, 3, 19, 6, 22, 7, 23,
10422 // Second 128-bit lane.
10423 10, 26, 11, 27, 14, 30, 15, 31}))
10424 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v16f32, V1, V2);
10426 // FIXME: Implement direct support for this type!
10427 return splitAndLowerVectorShuffle(DL, MVT::v16f32, V1, V2, Mask, DAG);
10430 /// \brief Handle lowering of 8-lane 64-bit integer shuffles.
10431 static SDValue lowerV8I64VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10432 const X86Subtarget *Subtarget,
10433 SelectionDAG &DAG) {
10435 assert(V1.getSimpleValueType() == MVT::v8i64 && "Bad operand type!");
10436 assert(V2.getSimpleValueType() == MVT::v8i64 && "Bad operand type!");
10437 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10438 ArrayRef<int> Mask = SVOp->getMask();
10439 assert(Mask.size() == 8 && "Unexpected mask size for v8 shuffle!");
10441 // X86 has dedicated unpack instructions that can handle specific blend
10442 // operations: UNPCKH and UNPCKL.
10443 if (isShuffleEquivalent(V1, V2, Mask, {0, 8, 2, 10, 4, 12, 6, 14}))
10444 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v8i64, V1, V2);
10445 if (isShuffleEquivalent(V1, V2, Mask, {1, 9, 3, 11, 5, 13, 7, 15}))
10446 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v8i64, V1, V2);
10448 // FIXME: Implement direct support for this type!
10449 return splitAndLowerVectorShuffle(DL, MVT::v8i64, V1, V2, Mask, DAG);
10452 /// \brief Handle lowering of 16-lane 32-bit integer shuffles.
10453 static SDValue lowerV16I32VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10454 const X86Subtarget *Subtarget,
10455 SelectionDAG &DAG) {
10457 assert(V1.getSimpleValueType() == MVT::v16i32 && "Bad operand type!");
10458 assert(V2.getSimpleValueType() == MVT::v16i32 && "Bad operand type!");
10459 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10460 ArrayRef<int> Mask = SVOp->getMask();
10461 assert(Mask.size() == 16 && "Unexpected mask size for v16 shuffle!");
10463 // Use dedicated unpack instructions for masks that match their pattern.
10464 if (isShuffleEquivalent(V1, V2, Mask,
10465 {// First 128-bit lane.
10466 0, 16, 1, 17, 4, 20, 5, 21,
10467 // Second 128-bit lane.
10468 8, 24, 9, 25, 12, 28, 13, 29}))
10469 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v16i32, V1, V2);
10470 if (isShuffleEquivalent(V1, V2, Mask,
10471 {// First 128-bit lane.
10472 2, 18, 3, 19, 6, 22, 7, 23,
10473 // Second 128-bit lane.
10474 10, 26, 11, 27, 14, 30, 15, 31}))
10475 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v16i32, V1, V2);
10477 // FIXME: Implement direct support for this type!
10478 return splitAndLowerVectorShuffle(DL, MVT::v16i32, V1, V2, Mask, DAG);
10481 /// \brief Handle lowering of 32-lane 16-bit integer shuffles.
10482 static SDValue lowerV32I16VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10483 const X86Subtarget *Subtarget,
10484 SelectionDAG &DAG) {
10486 assert(V1.getSimpleValueType() == MVT::v32i16 && "Bad operand type!");
10487 assert(V2.getSimpleValueType() == MVT::v32i16 && "Bad operand type!");
10488 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10489 ArrayRef<int> Mask = SVOp->getMask();
10490 assert(Mask.size() == 32 && "Unexpected mask size for v32 shuffle!");
10491 assert(Subtarget->hasBWI() && "We can only lower v32i16 with AVX-512-BWI!");
10493 // FIXME: Implement direct support for this type!
10494 return splitAndLowerVectorShuffle(DL, MVT::v32i16, V1, V2, Mask, DAG);
10497 /// \brief Handle lowering of 64-lane 8-bit integer shuffles.
10498 static SDValue lowerV64I8VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10499 const X86Subtarget *Subtarget,
10500 SelectionDAG &DAG) {
10502 assert(V1.getSimpleValueType() == MVT::v64i8 && "Bad operand type!");
10503 assert(V2.getSimpleValueType() == MVT::v64i8 && "Bad operand type!");
10504 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10505 ArrayRef<int> Mask = SVOp->getMask();
10506 assert(Mask.size() == 64 && "Unexpected mask size for v64 shuffle!");
10507 assert(Subtarget->hasBWI() && "We can only lower v64i8 with AVX-512-BWI!");
10509 // FIXME: Implement direct support for this type!
10510 return splitAndLowerVectorShuffle(DL, MVT::v64i8, V1, V2, Mask, DAG);
10513 /// \brief High-level routine to lower various 512-bit x86 vector shuffles.
10515 /// This routine either breaks down the specific type of a 512-bit x86 vector
10516 /// shuffle or splits it into two 256-bit shuffles and fuses the results back
10517 /// together based on the available instructions.
10518 static SDValue lower512BitVectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10519 MVT VT, const X86Subtarget *Subtarget,
10520 SelectionDAG &DAG) {
10522 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10523 ArrayRef<int> Mask = SVOp->getMask();
10524 assert(Subtarget->hasAVX512() &&
10525 "Cannot lower 512-bit vectors w/ basic ISA!");
10527 // Check for being able to broadcast a single element.
10528 if (SDValue Broadcast =
10529 lowerVectorShuffleAsBroadcast(DL, VT, V1, Mask, Subtarget, DAG))
10532 // Dispatch to each element type for lowering. If we don't have supprot for
10533 // specific element type shuffles at 512 bits, immediately split them and
10534 // lower them. Each lowering routine of a given type is allowed to assume that
10535 // the requisite ISA extensions for that element type are available.
10536 switch (VT.SimpleTy) {
10538 return lowerV8F64VectorShuffle(Op, V1, V2, Subtarget, DAG);
10540 return lowerV16F32VectorShuffle(Op, V1, V2, Subtarget, DAG);
10542 return lowerV8I64VectorShuffle(Op, V1, V2, Subtarget, DAG);
10544 return lowerV16I32VectorShuffle(Op, V1, V2, Subtarget, DAG);
10546 if (Subtarget->hasBWI())
10547 return lowerV32I16VectorShuffle(Op, V1, V2, Subtarget, DAG);
10550 if (Subtarget->hasBWI())
10551 return lowerV64I8VectorShuffle(Op, V1, V2, Subtarget, DAG);
10555 llvm_unreachable("Not a valid 512-bit x86 vector type!");
10558 // Otherwise fall back on splitting.
10559 return splitAndLowerVectorShuffle(DL, VT, V1, V2, Mask, DAG);
10562 /// \brief Top-level lowering for x86 vector shuffles.
10564 /// This handles decomposition, canonicalization, and lowering of all x86
10565 /// vector shuffles. Most of the specific lowering strategies are encapsulated
10566 /// above in helper routines. The canonicalization attempts to widen shuffles
10567 /// to involve fewer lanes of wider elements, consolidate symmetric patterns
10568 /// s.t. only one of the two inputs needs to be tested, etc.
10569 static SDValue lowerVectorShuffle(SDValue Op, const X86Subtarget *Subtarget,
10570 SelectionDAG &DAG) {
10571 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10572 ArrayRef<int> Mask = SVOp->getMask();
10573 SDValue V1 = Op.getOperand(0);
10574 SDValue V2 = Op.getOperand(1);
10575 MVT VT = Op.getSimpleValueType();
10576 int NumElements = VT.getVectorNumElements();
10579 assert(VT.getSizeInBits() != 64 && "Can't lower MMX shuffles");
10581 bool V1IsUndef = V1.getOpcode() == ISD::UNDEF;
10582 bool V2IsUndef = V2.getOpcode() == ISD::UNDEF;
10583 if (V1IsUndef && V2IsUndef)
10584 return DAG.getUNDEF(VT);
10586 // When we create a shuffle node we put the UNDEF node to second operand,
10587 // but in some cases the first operand may be transformed to UNDEF.
10588 // In this case we should just commute the node.
10590 return DAG.getCommutedVectorShuffle(*SVOp);
10592 // Check for non-undef masks pointing at an undef vector and make the masks
10593 // undef as well. This makes it easier to match the shuffle based solely on
10597 if (M >= NumElements) {
10598 SmallVector<int, 8> NewMask(Mask.begin(), Mask.end());
10599 for (int &M : NewMask)
10600 if (M >= NumElements)
10602 return DAG.getVectorShuffle(VT, dl, V1, V2, NewMask);
10605 // We actually see shuffles that are entirely re-arrangements of a set of
10606 // zero inputs. This mostly happens while decomposing complex shuffles into
10607 // simple ones. Directly lower these as a buildvector of zeros.
10608 SmallBitVector Zeroable = computeZeroableShuffleElements(Mask, V1, V2);
10609 if (Zeroable.all())
10610 return getZeroVector(VT, Subtarget, DAG, dl);
10612 // Try to collapse shuffles into using a vector type with fewer elements but
10613 // wider element types. We cap this to not form integers or floating point
10614 // elements wider than 64 bits, but it might be interesting to form i128
10615 // integers to handle flipping the low and high halves of AVX 256-bit vectors.
10616 SmallVector<int, 16> WidenedMask;
10617 if (VT.getScalarSizeInBits() < 64 &&
10618 canWidenShuffleElements(Mask, WidenedMask)) {
10619 MVT NewEltVT = VT.isFloatingPoint()
10620 ? MVT::getFloatingPointVT(VT.getScalarSizeInBits() * 2)
10621 : MVT::getIntegerVT(VT.getScalarSizeInBits() * 2);
10622 MVT NewVT = MVT::getVectorVT(NewEltVT, VT.getVectorNumElements() / 2);
10623 // Make sure that the new vector type is legal. For example, v2f64 isn't
10625 if (DAG.getTargetLoweringInfo().isTypeLegal(NewVT)) {
10626 V1 = DAG.getBitcast(NewVT, V1);
10627 V2 = DAG.getBitcast(NewVT, V2);
10628 return DAG.getBitcast(
10629 VT, DAG.getVectorShuffle(NewVT, dl, V1, V2, WidenedMask));
10633 int NumV1Elements = 0, NumUndefElements = 0, NumV2Elements = 0;
10634 for (int M : SVOp->getMask())
10636 ++NumUndefElements;
10637 else if (M < NumElements)
10642 // Commute the shuffle as needed such that more elements come from V1 than
10643 // V2. This allows us to match the shuffle pattern strictly on how many
10644 // elements come from V1 without handling the symmetric cases.
10645 if (NumV2Elements > NumV1Elements)
10646 return DAG.getCommutedVectorShuffle(*SVOp);
10648 // When the number of V1 and V2 elements are the same, try to minimize the
10649 // number of uses of V2 in the low half of the vector. When that is tied,
10650 // ensure that the sum of indices for V1 is equal to or lower than the sum
10651 // indices for V2. When those are equal, try to ensure that the number of odd
10652 // indices for V1 is lower than the number of odd indices for V2.
10653 if (NumV1Elements == NumV2Elements) {
10654 int LowV1Elements = 0, LowV2Elements = 0;
10655 for (int M : SVOp->getMask().slice(0, NumElements / 2))
10656 if (M >= NumElements)
10660 if (LowV2Elements > LowV1Elements) {
10661 return DAG.getCommutedVectorShuffle(*SVOp);
10662 } else if (LowV2Elements == LowV1Elements) {
10663 int SumV1Indices = 0, SumV2Indices = 0;
10664 for (int i = 0, Size = SVOp->getMask().size(); i < Size; ++i)
10665 if (SVOp->getMask()[i] >= NumElements)
10667 else if (SVOp->getMask()[i] >= 0)
10669 if (SumV2Indices < SumV1Indices) {
10670 return DAG.getCommutedVectorShuffle(*SVOp);
10671 } else if (SumV2Indices == SumV1Indices) {
10672 int NumV1OddIndices = 0, NumV2OddIndices = 0;
10673 for (int i = 0, Size = SVOp->getMask().size(); i < Size; ++i)
10674 if (SVOp->getMask()[i] >= NumElements)
10675 NumV2OddIndices += i % 2;
10676 else if (SVOp->getMask()[i] >= 0)
10677 NumV1OddIndices += i % 2;
10678 if (NumV2OddIndices < NumV1OddIndices)
10679 return DAG.getCommutedVectorShuffle(*SVOp);
10684 // For each vector width, delegate to a specialized lowering routine.
10685 if (VT.getSizeInBits() == 128)
10686 return lower128BitVectorShuffle(Op, V1, V2, VT, Subtarget, DAG);
10688 if (VT.getSizeInBits() == 256)
10689 return lower256BitVectorShuffle(Op, V1, V2, VT, Subtarget, DAG);
10691 // Force AVX-512 vectors to be scalarized for now.
10692 // FIXME: Implement AVX-512 support!
10693 if (VT.getSizeInBits() == 512)
10694 return lower512BitVectorShuffle(Op, V1, V2, VT, Subtarget, DAG);
10696 llvm_unreachable("Unimplemented!");
10699 // This function assumes its argument is a BUILD_VECTOR of constants or
10700 // undef SDNodes. i.e: ISD::isBuildVectorOfConstantSDNodes(BuildVector) is
10702 static bool BUILD_VECTORtoBlendMask(BuildVectorSDNode *BuildVector,
10703 unsigned &MaskValue) {
10705 unsigned NumElems = BuildVector->getNumOperands();
10706 // There are 2 lanes if (NumElems > 8), and 1 lane otherwise.
10707 unsigned NumLanes = (NumElems - 1) / 8 + 1;
10708 unsigned NumElemsInLane = NumElems / NumLanes;
10710 // Blend for v16i16 should be symetric for the both lanes.
10711 for (unsigned i = 0; i < NumElemsInLane; ++i) {
10712 SDValue EltCond = BuildVector->getOperand(i);
10713 SDValue SndLaneEltCond =
10714 (NumLanes == 2) ? BuildVector->getOperand(i + NumElemsInLane) : EltCond;
10716 int Lane1Cond = -1, Lane2Cond = -1;
10717 if (isa<ConstantSDNode>(EltCond))
10718 Lane1Cond = !isZero(EltCond);
10719 if (isa<ConstantSDNode>(SndLaneEltCond))
10720 Lane2Cond = !isZero(SndLaneEltCond);
10722 if (Lane1Cond == Lane2Cond || Lane2Cond < 0)
10723 // Lane1Cond != 0, means we want the first argument.
10724 // Lane1Cond == 0, means we want the second argument.
10725 // The encoding of this argument is 0 for the first argument, 1
10726 // for the second. Therefore, invert the condition.
10727 MaskValue |= !Lane1Cond << i;
10728 else if (Lane1Cond < 0)
10729 MaskValue |= !Lane2Cond << i;
10736 /// \brief Try to lower a VSELECT instruction to a vector shuffle.
10737 static SDValue lowerVSELECTtoVectorShuffle(SDValue Op,
10738 const X86Subtarget *Subtarget,
10739 SelectionDAG &DAG) {
10740 SDValue Cond = Op.getOperand(0);
10741 SDValue LHS = Op.getOperand(1);
10742 SDValue RHS = Op.getOperand(2);
10744 MVT VT = Op.getSimpleValueType();
10746 if (!ISD::isBuildVectorOfConstantSDNodes(Cond.getNode()))
10748 auto *CondBV = cast<BuildVectorSDNode>(Cond);
10750 // Only non-legal VSELECTs reach this lowering, convert those into generic
10751 // shuffles and re-use the shuffle lowering path for blends.
10752 SmallVector<int, 32> Mask;
10753 for (int i = 0, Size = VT.getVectorNumElements(); i < Size; ++i) {
10754 SDValue CondElt = CondBV->getOperand(i);
10756 isa<ConstantSDNode>(CondElt) ? i + (isZero(CondElt) ? Size : 0) : -1);
10758 return DAG.getVectorShuffle(VT, dl, LHS, RHS, Mask);
10761 SDValue X86TargetLowering::LowerVSELECT(SDValue Op, SelectionDAG &DAG) const {
10762 // A vselect where all conditions and data are constants can be optimized into
10763 // a single vector load by SelectionDAGLegalize::ExpandBUILD_VECTOR().
10764 if (ISD::isBuildVectorOfConstantSDNodes(Op.getOperand(0).getNode()) &&
10765 ISD::isBuildVectorOfConstantSDNodes(Op.getOperand(1).getNode()) &&
10766 ISD::isBuildVectorOfConstantSDNodes(Op.getOperand(2).getNode()))
10769 // Try to lower this to a blend-style vector shuffle. This can handle all
10770 // constant condition cases.
10771 if (SDValue BlendOp = lowerVSELECTtoVectorShuffle(Op, Subtarget, DAG))
10774 // Variable blends are only legal from SSE4.1 onward.
10775 if (!Subtarget->hasSSE41())
10778 // Only some types will be legal on some subtargets. If we can emit a legal
10779 // VSELECT-matching blend, return Op, and but if we need to expand, return
10781 switch (Op.getSimpleValueType().SimpleTy) {
10783 // Most of the vector types have blends past SSE4.1.
10787 // The byte blends for AVX vectors were introduced only in AVX2.
10788 if (Subtarget->hasAVX2())
10795 // AVX-512 BWI and VLX features support VSELECT with i16 elements.
10796 if (Subtarget->hasBWI() && Subtarget->hasVLX())
10799 // FIXME: We should custom lower this by fixing the condition and using i8
10805 static SDValue LowerEXTRACT_VECTOR_ELT_SSE4(SDValue Op, SelectionDAG &DAG) {
10806 MVT VT = Op.getSimpleValueType();
10809 if (!Op.getOperand(0).getSimpleValueType().is128BitVector())
10812 if (VT.getSizeInBits() == 8) {
10813 SDValue Extract = DAG.getNode(X86ISD::PEXTRB, dl, MVT::i32,
10814 Op.getOperand(0), Op.getOperand(1));
10815 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Extract,
10816 DAG.getValueType(VT));
10817 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
10820 if (VT.getSizeInBits() == 16) {
10821 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
10822 // If Idx is 0, it's cheaper to do a move instead of a pextrw.
10824 return DAG.getNode(
10825 ISD::TRUNCATE, dl, MVT::i16,
10826 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
10827 DAG.getBitcast(MVT::v4i32, Op.getOperand(0)),
10828 Op.getOperand(1)));
10829 SDValue Extract = DAG.getNode(X86ISD::PEXTRW, dl, MVT::i32,
10830 Op.getOperand(0), Op.getOperand(1));
10831 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Extract,
10832 DAG.getValueType(VT));
10833 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
10836 if (VT == MVT::f32) {
10837 // EXTRACTPS outputs to a GPR32 register which will require a movd to copy
10838 // the result back to FR32 register. It's only worth matching if the
10839 // result has a single use which is a store or a bitcast to i32. And in
10840 // the case of a store, it's not worth it if the index is a constant 0,
10841 // because a MOVSSmr can be used instead, which is smaller and faster.
10842 if (!Op.hasOneUse())
10844 SDNode *User = *Op.getNode()->use_begin();
10845 if ((User->getOpcode() != ISD::STORE ||
10846 (isa<ConstantSDNode>(Op.getOperand(1)) &&
10847 cast<ConstantSDNode>(Op.getOperand(1))->isNullValue())) &&
10848 (User->getOpcode() != ISD::BITCAST ||
10849 User->getValueType(0) != MVT::i32))
10851 SDValue Extract = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
10852 DAG.getBitcast(MVT::v4i32, Op.getOperand(0)),
10854 return DAG.getBitcast(MVT::f32, Extract);
10857 if (VT == MVT::i32 || VT == MVT::i64) {
10858 // ExtractPS/pextrq works with constant index.
10859 if (isa<ConstantSDNode>(Op.getOperand(1)))
10865 /// Extract one bit from mask vector, like v16i1 or v8i1.
10866 /// AVX-512 feature.
10868 X86TargetLowering::ExtractBitFromMaskVector(SDValue Op, SelectionDAG &DAG) const {
10869 SDValue Vec = Op.getOperand(0);
10871 MVT VecVT = Vec.getSimpleValueType();
10872 SDValue Idx = Op.getOperand(1);
10873 MVT EltVT = Op.getSimpleValueType();
10875 assert((EltVT == MVT::i1) && "Unexpected operands in ExtractBitFromMaskVector");
10876 assert((VecVT.getVectorNumElements() <= 16 || Subtarget->hasBWI()) &&
10877 "Unexpected vector type in ExtractBitFromMaskVector");
10879 // variable index can't be handled in mask registers,
10880 // extend vector to VR512
10881 if (!isa<ConstantSDNode>(Idx)) {
10882 MVT ExtVT = (VecVT == MVT::v8i1 ? MVT::v8i64 : MVT::v16i32);
10883 SDValue Ext = DAG.getNode(ISD::ZERO_EXTEND, dl, ExtVT, Vec);
10884 SDValue Elt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl,
10885 ExtVT.getVectorElementType(), Ext, Idx);
10886 return DAG.getNode(ISD::TRUNCATE, dl, EltVT, Elt);
10889 unsigned IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
10890 const TargetRegisterClass* rc = getRegClassFor(VecVT);
10891 if (!Subtarget->hasDQI() && (VecVT.getVectorNumElements() <= 8))
10892 rc = getRegClassFor(MVT::v16i1);
10893 unsigned MaxSift = rc->getSize()*8 - 1;
10894 Vec = DAG.getNode(X86ISD::VSHLI, dl, VecVT, Vec,
10895 DAG.getConstant(MaxSift - IdxVal, dl, MVT::i8));
10896 Vec = DAG.getNode(X86ISD::VSRLI, dl, VecVT, Vec,
10897 DAG.getConstant(MaxSift, dl, MVT::i8));
10898 return DAG.getNode(X86ISD::VEXTRACT, dl, MVT::i1, Vec,
10899 DAG.getIntPtrConstant(0, dl));
10903 X86TargetLowering::LowerEXTRACT_VECTOR_ELT(SDValue Op,
10904 SelectionDAG &DAG) const {
10906 SDValue Vec = Op.getOperand(0);
10907 MVT VecVT = Vec.getSimpleValueType();
10908 SDValue Idx = Op.getOperand(1);
10910 if (Op.getSimpleValueType() == MVT::i1)
10911 return ExtractBitFromMaskVector(Op, DAG);
10913 if (!isa<ConstantSDNode>(Idx)) {
10914 if (VecVT.is512BitVector() ||
10915 (VecVT.is256BitVector() && Subtarget->hasInt256() &&
10916 VecVT.getVectorElementType().getSizeInBits() == 32)) {
10919 MVT::getIntegerVT(VecVT.getVectorElementType().getSizeInBits());
10920 MVT MaskVT = MVT::getVectorVT(MaskEltVT, VecVT.getSizeInBits() /
10921 MaskEltVT.getSizeInBits());
10923 Idx = DAG.getZExtOrTrunc(Idx, dl, MaskEltVT);
10924 auto PtrVT = getPointerTy(DAG.getDataLayout());
10925 SDValue Mask = DAG.getNode(X86ISD::VINSERT, dl, MaskVT,
10926 getZeroVector(MaskVT, Subtarget, DAG, dl), Idx,
10927 DAG.getConstant(0, dl, PtrVT));
10928 SDValue Perm = DAG.getNode(X86ISD::VPERMV, dl, VecVT, Mask, Vec);
10929 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, Op.getValueType(), Perm,
10930 DAG.getConstant(0, dl, PtrVT));
10935 // If this is a 256-bit vector result, first extract the 128-bit vector and
10936 // then extract the element from the 128-bit vector.
10937 if (VecVT.is256BitVector() || VecVT.is512BitVector()) {
10939 unsigned IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
10940 // Get the 128-bit vector.
10941 Vec = Extract128BitVector(Vec, IdxVal, DAG, dl);
10942 MVT EltVT = VecVT.getVectorElementType();
10944 unsigned ElemsPerChunk = 128 / EltVT.getSizeInBits();
10946 //if (IdxVal >= NumElems/2)
10947 // IdxVal -= NumElems/2;
10948 IdxVal -= (IdxVal/ElemsPerChunk)*ElemsPerChunk;
10949 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, Op.getValueType(), Vec,
10950 DAG.getConstant(IdxVal, dl, MVT::i32));
10953 assert(VecVT.is128BitVector() && "Unexpected vector length");
10955 if (Subtarget->hasSSE41())
10956 if (SDValue Res = LowerEXTRACT_VECTOR_ELT_SSE4(Op, DAG))
10959 MVT VT = Op.getSimpleValueType();
10960 // TODO: handle v16i8.
10961 if (VT.getSizeInBits() == 16) {
10962 SDValue Vec = Op.getOperand(0);
10963 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
10965 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i16,
10966 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
10967 DAG.getBitcast(MVT::v4i32, Vec),
10968 Op.getOperand(1)));
10969 // Transform it so it match pextrw which produces a 32-bit result.
10970 MVT EltVT = MVT::i32;
10971 SDValue Extract = DAG.getNode(X86ISD::PEXTRW, dl, EltVT,
10972 Op.getOperand(0), Op.getOperand(1));
10973 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, EltVT, Extract,
10974 DAG.getValueType(VT));
10975 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
10978 if (VT.getSizeInBits() == 32) {
10979 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
10983 // SHUFPS the element to the lowest double word, then movss.
10984 int Mask[4] = { static_cast<int>(Idx), -1, -1, -1 };
10985 MVT VVT = Op.getOperand(0).getSimpleValueType();
10986 SDValue Vec = DAG.getVectorShuffle(VVT, dl, Op.getOperand(0),
10987 DAG.getUNDEF(VVT), Mask);
10988 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, Vec,
10989 DAG.getIntPtrConstant(0, dl));
10992 if (VT.getSizeInBits() == 64) {
10993 // FIXME: .td only matches this for <2 x f64>, not <2 x i64> on 32b
10994 // FIXME: seems like this should be unnecessary if mov{h,l}pd were taught
10995 // to match extract_elt for f64.
10996 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
11000 // UNPCKHPD the element to the lowest double word, then movsd.
11001 // Note if the lower 64 bits of the result of the UNPCKHPD is then stored
11002 // to a f64mem, the whole operation is folded into a single MOVHPDmr.
11003 int Mask[2] = { 1, -1 };
11004 MVT VVT = Op.getOperand(0).getSimpleValueType();
11005 SDValue Vec = DAG.getVectorShuffle(VVT, dl, Op.getOperand(0),
11006 DAG.getUNDEF(VVT), Mask);
11007 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, Vec,
11008 DAG.getIntPtrConstant(0, dl));
11014 /// Insert one bit to mask vector, like v16i1 or v8i1.
11015 /// AVX-512 feature.
11017 X86TargetLowering::InsertBitToMaskVector(SDValue Op, SelectionDAG &DAG) const {
11019 SDValue Vec = Op.getOperand(0);
11020 SDValue Elt = Op.getOperand(1);
11021 SDValue Idx = Op.getOperand(2);
11022 MVT VecVT = Vec.getSimpleValueType();
11024 if (!isa<ConstantSDNode>(Idx)) {
11025 // Non constant index. Extend source and destination,
11026 // insert element and then truncate the result.
11027 MVT ExtVecVT = (VecVT == MVT::v8i1 ? MVT::v8i64 : MVT::v16i32);
11028 MVT ExtEltVT = (VecVT == MVT::v8i1 ? MVT::i64 : MVT::i32);
11029 SDValue ExtOp = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, ExtVecVT,
11030 DAG.getNode(ISD::ZERO_EXTEND, dl, ExtVecVT, Vec),
11031 DAG.getNode(ISD::ZERO_EXTEND, dl, ExtEltVT, Elt), Idx);
11032 return DAG.getNode(ISD::TRUNCATE, dl, VecVT, ExtOp);
11035 unsigned IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
11036 SDValue EltInVec = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VecVT, Elt);
11038 EltInVec = DAG.getNode(X86ISD::VSHLI, dl, VecVT, EltInVec,
11039 DAG.getConstant(IdxVal, dl, MVT::i8));
11040 if (Vec.getOpcode() == ISD::UNDEF)
11042 return DAG.getNode(ISD::OR, dl, VecVT, Vec, EltInVec);
11045 SDValue X86TargetLowering::LowerINSERT_VECTOR_ELT(SDValue Op,
11046 SelectionDAG &DAG) const {
11047 MVT VT = Op.getSimpleValueType();
11048 MVT EltVT = VT.getVectorElementType();
11050 if (EltVT == MVT::i1)
11051 return InsertBitToMaskVector(Op, DAG);
11054 SDValue N0 = Op.getOperand(0);
11055 SDValue N1 = Op.getOperand(1);
11056 SDValue N2 = Op.getOperand(2);
11057 if (!isa<ConstantSDNode>(N2))
11059 auto *N2C = cast<ConstantSDNode>(N2);
11060 unsigned IdxVal = N2C->getZExtValue();
11062 // If the vector is wider than 128 bits, extract the 128-bit subvector, insert
11063 // into that, and then insert the subvector back into the result.
11064 if (VT.is256BitVector() || VT.is512BitVector()) {
11065 // With a 256-bit vector, we can insert into the zero element efficiently
11066 // using a blend if we have AVX or AVX2 and the right data type.
11067 if (VT.is256BitVector() && IdxVal == 0) {
11068 // TODO: It is worthwhile to cast integer to floating point and back
11069 // and incur a domain crossing penalty if that's what we'll end up
11070 // doing anyway after extracting to a 128-bit vector.
11071 if ((Subtarget->hasAVX() && (EltVT == MVT::f64 || EltVT == MVT::f32)) ||
11072 (Subtarget->hasAVX2() && EltVT == MVT::i32)) {
11073 SDValue N1Vec = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, N1);
11074 N2 = DAG.getIntPtrConstant(1, dl);
11075 return DAG.getNode(X86ISD::BLENDI, dl, VT, N0, N1Vec, N2);
11079 // Get the desired 128-bit vector chunk.
11080 SDValue V = Extract128BitVector(N0, IdxVal, DAG, dl);
11082 // Insert the element into the desired chunk.
11083 unsigned NumEltsIn128 = 128 / EltVT.getSizeInBits();
11084 unsigned IdxIn128 = IdxVal - (IdxVal / NumEltsIn128) * NumEltsIn128;
11086 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, V.getValueType(), V, N1,
11087 DAG.getConstant(IdxIn128, dl, MVT::i32));
11089 // Insert the changed part back into the bigger vector
11090 return Insert128BitVector(N0, V, IdxVal, DAG, dl);
11092 assert(VT.is128BitVector() && "Only 128-bit vector types should be left!");
11094 if (Subtarget->hasSSE41()) {
11095 if (EltVT.getSizeInBits() == 8 || EltVT.getSizeInBits() == 16) {
11097 if (VT == MVT::v8i16) {
11098 Opc = X86ISD::PINSRW;
11100 assert(VT == MVT::v16i8);
11101 Opc = X86ISD::PINSRB;
11104 // Transform it so it match pinsr{b,w} which expects a GR32 as its second
11106 if (N1.getValueType() != MVT::i32)
11107 N1 = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, N1);
11108 if (N2.getValueType() != MVT::i32)
11109 N2 = DAG.getIntPtrConstant(IdxVal, dl);
11110 return DAG.getNode(Opc, dl, VT, N0, N1, N2);
11113 if (EltVT == MVT::f32) {
11114 // Bits [7:6] of the constant are the source select. This will always be
11115 // zero here. The DAG Combiner may combine an extract_elt index into
11116 // these bits. For example (insert (extract, 3), 2) could be matched by
11117 // putting the '3' into bits [7:6] of X86ISD::INSERTPS.
11118 // Bits [5:4] of the constant are the destination select. This is the
11119 // value of the incoming immediate.
11120 // Bits [3:0] of the constant are the zero mask. The DAG Combiner may
11121 // combine either bitwise AND or insert of float 0.0 to set these bits.
11123 const Function *F = DAG.getMachineFunction().getFunction();
11124 bool MinSize = F->hasFnAttribute(Attribute::MinSize);
11125 if (IdxVal == 0 && (!MinSize || !MayFoldLoad(N1))) {
11126 // If this is an insertion of 32-bits into the low 32-bits of
11127 // a vector, we prefer to generate a blend with immediate rather
11128 // than an insertps. Blends are simpler operations in hardware and so
11129 // will always have equal or better performance than insertps.
11130 // But if optimizing for size and there's a load folding opportunity,
11131 // generate insertps because blendps does not have a 32-bit memory
11133 N2 = DAG.getIntPtrConstant(1, dl);
11134 N1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4f32, N1);
11135 return DAG.getNode(X86ISD::BLENDI, dl, VT, N0, N1, N2);
11137 N2 = DAG.getIntPtrConstant(IdxVal << 4, dl);
11138 // Create this as a scalar to vector..
11139 N1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4f32, N1);
11140 return DAG.getNode(X86ISD::INSERTPS, dl, VT, N0, N1, N2);
11143 if (EltVT == MVT::i32 || EltVT == MVT::i64) {
11144 // PINSR* works with constant index.
11149 if (EltVT == MVT::i8)
11152 if (EltVT.getSizeInBits() == 16) {
11153 // Transform it so it match pinsrw which expects a 16-bit value in a GR32
11154 // as its second argument.
11155 if (N1.getValueType() != MVT::i32)
11156 N1 = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, N1);
11157 if (N2.getValueType() != MVT::i32)
11158 N2 = DAG.getIntPtrConstant(IdxVal, dl);
11159 return DAG.getNode(X86ISD::PINSRW, dl, VT, N0, N1, N2);
11164 static SDValue LowerSCALAR_TO_VECTOR(SDValue Op, SelectionDAG &DAG) {
11166 MVT OpVT = Op.getSimpleValueType();
11168 // If this is a 256-bit vector result, first insert into a 128-bit
11169 // vector and then insert into the 256-bit vector.
11170 if (!OpVT.is128BitVector()) {
11171 // Insert into a 128-bit vector.
11172 unsigned SizeFactor = OpVT.getSizeInBits()/128;
11173 MVT VT128 = MVT::getVectorVT(OpVT.getVectorElementType(),
11174 OpVT.getVectorNumElements() / SizeFactor);
11176 Op = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT128, Op.getOperand(0));
11178 // Insert the 128-bit vector.
11179 return Insert128BitVector(DAG.getUNDEF(OpVT), Op, 0, DAG, dl);
11182 if (OpVT == MVT::v1i64 &&
11183 Op.getOperand(0).getValueType() == MVT::i64)
11184 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v1i64, Op.getOperand(0));
11186 SDValue AnyExt = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, Op.getOperand(0));
11187 assert(OpVT.is128BitVector() && "Expected an SSE type!");
11188 return DAG.getBitcast(
11189 OpVT, DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32, AnyExt));
11192 // Lower a node with an EXTRACT_SUBVECTOR opcode. This may result in
11193 // a simple subregister reference or explicit instructions to grab
11194 // upper bits of a vector.
11195 static SDValue LowerEXTRACT_SUBVECTOR(SDValue Op, const X86Subtarget *Subtarget,
11196 SelectionDAG &DAG) {
11198 SDValue In = Op.getOperand(0);
11199 SDValue Idx = Op.getOperand(1);
11200 unsigned IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
11201 MVT ResVT = Op.getSimpleValueType();
11202 MVT InVT = In.getSimpleValueType();
11204 if (Subtarget->hasFp256()) {
11205 if (ResVT.is128BitVector() &&
11206 (InVT.is256BitVector() || InVT.is512BitVector()) &&
11207 isa<ConstantSDNode>(Idx)) {
11208 return Extract128BitVector(In, IdxVal, DAG, dl);
11210 if (ResVT.is256BitVector() && InVT.is512BitVector() &&
11211 isa<ConstantSDNode>(Idx)) {
11212 return Extract256BitVector(In, IdxVal, DAG, dl);
11218 // Lower a node with an INSERT_SUBVECTOR opcode. This may result in a
11219 // simple superregister reference or explicit instructions to insert
11220 // the upper bits of a vector.
11221 static SDValue LowerINSERT_SUBVECTOR(SDValue Op, const X86Subtarget *Subtarget,
11222 SelectionDAG &DAG) {
11223 if (!Subtarget->hasAVX())
11227 SDValue Vec = Op.getOperand(0);
11228 SDValue SubVec = Op.getOperand(1);
11229 SDValue Idx = Op.getOperand(2);
11231 if (!isa<ConstantSDNode>(Idx))
11234 unsigned IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
11235 MVT OpVT = Op.getSimpleValueType();
11236 MVT SubVecVT = SubVec.getSimpleValueType();
11238 // Fold two 16-byte subvector loads into one 32-byte load:
11239 // (insert_subvector (insert_subvector undef, (load addr), 0),
11240 // (load addr + 16), Elts/2)
11242 if ((IdxVal == OpVT.getVectorNumElements() / 2) &&
11243 Vec.getOpcode() == ISD::INSERT_SUBVECTOR &&
11244 OpVT.is256BitVector() && SubVecVT.is128BitVector() &&
11245 !Subtarget->isUnalignedMem32Slow()) {
11246 SDValue SubVec2 = Vec.getOperand(1);
11247 if (auto *Idx2 = dyn_cast<ConstantSDNode>(Vec.getOperand(2))) {
11248 if (Idx2->getZExtValue() == 0) {
11249 SDValue Ops[] = { SubVec2, SubVec };
11250 if (SDValue Ld = EltsFromConsecutiveLoads(OpVT, Ops, dl, DAG, false))
11256 if ((OpVT.is256BitVector() || OpVT.is512BitVector()) &&
11257 SubVecVT.is128BitVector())
11258 return Insert128BitVector(Vec, SubVec, IdxVal, DAG, dl);
11260 if (OpVT.is512BitVector() && SubVecVT.is256BitVector())
11261 return Insert256BitVector(Vec, SubVec, IdxVal, DAG, dl);
11263 if (OpVT.getVectorElementType() == MVT::i1) {
11264 if (IdxVal == 0 && Vec.getOpcode() == ISD::UNDEF) // the operation is legal
11266 SDValue ZeroIdx = DAG.getIntPtrConstant(0, dl);
11267 SDValue Undef = DAG.getUNDEF(OpVT);
11268 unsigned NumElems = OpVT.getVectorNumElements();
11269 SDValue ShiftBits = DAG.getConstant(NumElems/2, dl, MVT::i8);
11271 if (IdxVal == OpVT.getVectorNumElements() / 2) {
11272 // Zero upper bits of the Vec
11273 Vec = DAG.getNode(X86ISD::VSHLI, dl, OpVT, Vec, ShiftBits);
11274 Vec = DAG.getNode(X86ISD::VSRLI, dl, OpVT, Vec, ShiftBits);
11276 SDValue Vec2 = DAG.getNode(ISD::INSERT_SUBVECTOR, dl, OpVT, Undef,
11278 Vec2 = DAG.getNode(X86ISD::VSHLI, dl, OpVT, Vec2, ShiftBits);
11279 return DAG.getNode(ISD::OR, dl, OpVT, Vec, Vec2);
11282 SDValue Vec2 = DAG.getNode(ISD::INSERT_SUBVECTOR, dl, OpVT, Undef,
11284 // Zero upper bits of the Vec2
11285 Vec2 = DAG.getNode(X86ISD::VSHLI, dl, OpVT, Vec2, ShiftBits);
11286 Vec2 = DAG.getNode(X86ISD::VSRLI, dl, OpVT, Vec2, ShiftBits);
11287 // Zero lower bits of the Vec
11288 Vec = DAG.getNode(X86ISD::VSRLI, dl, OpVT, Vec, ShiftBits);
11289 Vec = DAG.getNode(X86ISD::VSHLI, dl, OpVT, Vec, ShiftBits);
11290 // Merge them together
11291 return DAG.getNode(ISD::OR, dl, OpVT, Vec, Vec2);
11297 // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
11298 // their target countpart wrapped in the X86ISD::Wrapper node. Suppose N is
11299 // one of the above mentioned nodes. It has to be wrapped because otherwise
11300 // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
11301 // be used to form addressing mode. These wrapped nodes will be selected
11304 X86TargetLowering::LowerConstantPool(SDValue Op, SelectionDAG &DAG) const {
11305 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
11307 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
11308 // global base reg.
11309 unsigned char OpFlag = 0;
11310 unsigned WrapperKind = X86ISD::Wrapper;
11311 CodeModel::Model M = DAG.getTarget().getCodeModel();
11313 if (Subtarget->isPICStyleRIPRel() &&
11314 (M == CodeModel::Small || M == CodeModel::Kernel))
11315 WrapperKind = X86ISD::WrapperRIP;
11316 else if (Subtarget->isPICStyleGOT())
11317 OpFlag = X86II::MO_GOTOFF;
11318 else if (Subtarget->isPICStyleStubPIC())
11319 OpFlag = X86II::MO_PIC_BASE_OFFSET;
11321 auto PtrVT = getPointerTy(DAG.getDataLayout());
11322 SDValue Result = DAG.getTargetConstantPool(
11323 CP->getConstVal(), PtrVT, CP->getAlignment(), CP->getOffset(), OpFlag);
11325 Result = DAG.getNode(WrapperKind, DL, PtrVT, Result);
11326 // With PIC, the address is actually $g + Offset.
11329 DAG.getNode(ISD::ADD, DL, PtrVT,
11330 DAG.getNode(X86ISD::GlobalBaseReg, SDLoc(), PtrVT), Result);
11336 SDValue X86TargetLowering::LowerJumpTable(SDValue Op, SelectionDAG &DAG) const {
11337 JumpTableSDNode *JT = cast<JumpTableSDNode>(Op);
11339 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
11340 // global base reg.
11341 unsigned char OpFlag = 0;
11342 unsigned WrapperKind = X86ISD::Wrapper;
11343 CodeModel::Model M = DAG.getTarget().getCodeModel();
11345 if (Subtarget->isPICStyleRIPRel() &&
11346 (M == CodeModel::Small || M == CodeModel::Kernel))
11347 WrapperKind = X86ISD::WrapperRIP;
11348 else if (Subtarget->isPICStyleGOT())
11349 OpFlag = X86II::MO_GOTOFF;
11350 else if (Subtarget->isPICStyleStubPIC())
11351 OpFlag = X86II::MO_PIC_BASE_OFFSET;
11353 auto PtrVT = getPointerTy(DAG.getDataLayout());
11354 SDValue Result = DAG.getTargetJumpTable(JT->getIndex(), PtrVT, OpFlag);
11356 Result = DAG.getNode(WrapperKind, DL, PtrVT, Result);
11358 // With PIC, the address is actually $g + Offset.
11361 DAG.getNode(ISD::ADD, DL, PtrVT,
11362 DAG.getNode(X86ISD::GlobalBaseReg, SDLoc(), PtrVT), Result);
11368 X86TargetLowering::LowerExternalSymbol(SDValue Op, SelectionDAG &DAG) const {
11369 const char *Sym = cast<ExternalSymbolSDNode>(Op)->getSymbol();
11371 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
11372 // global base reg.
11373 unsigned char OpFlag = 0;
11374 unsigned WrapperKind = X86ISD::Wrapper;
11375 CodeModel::Model M = DAG.getTarget().getCodeModel();
11377 if (Subtarget->isPICStyleRIPRel() &&
11378 (M == CodeModel::Small || M == CodeModel::Kernel)) {
11379 if (Subtarget->isTargetDarwin() || Subtarget->isTargetELF())
11380 OpFlag = X86II::MO_GOTPCREL;
11381 WrapperKind = X86ISD::WrapperRIP;
11382 } else if (Subtarget->isPICStyleGOT()) {
11383 OpFlag = X86II::MO_GOT;
11384 } else if (Subtarget->isPICStyleStubPIC()) {
11385 OpFlag = X86II::MO_DARWIN_NONLAZY_PIC_BASE;
11386 } else if (Subtarget->isPICStyleStubNoDynamic()) {
11387 OpFlag = X86II::MO_DARWIN_NONLAZY;
11390 auto PtrVT = getPointerTy(DAG.getDataLayout());
11391 SDValue Result = DAG.getTargetExternalSymbol(Sym, PtrVT, OpFlag);
11394 Result = DAG.getNode(WrapperKind, DL, PtrVT, Result);
11396 // With PIC, the address is actually $g + Offset.
11397 if (DAG.getTarget().getRelocationModel() == Reloc::PIC_ &&
11398 !Subtarget->is64Bit()) {
11400 DAG.getNode(ISD::ADD, DL, PtrVT,
11401 DAG.getNode(X86ISD::GlobalBaseReg, SDLoc(), PtrVT), Result);
11404 // For symbols that require a load from a stub to get the address, emit the
11406 if (isGlobalStubReference(OpFlag))
11407 Result = DAG.getLoad(PtrVT, DL, DAG.getEntryNode(), Result,
11408 MachinePointerInfo::getGOT(), false, false, false, 0);
11414 X86TargetLowering::LowerBlockAddress(SDValue Op, SelectionDAG &DAG) const {
11415 // Create the TargetBlockAddressAddress node.
11416 unsigned char OpFlags =
11417 Subtarget->ClassifyBlockAddressReference();
11418 CodeModel::Model M = DAG.getTarget().getCodeModel();
11419 const BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress();
11420 int64_t Offset = cast<BlockAddressSDNode>(Op)->getOffset();
11422 auto PtrVT = getPointerTy(DAG.getDataLayout());
11423 SDValue Result = DAG.getTargetBlockAddress(BA, PtrVT, Offset, OpFlags);
11425 if (Subtarget->isPICStyleRIPRel() &&
11426 (M == CodeModel::Small || M == CodeModel::Kernel))
11427 Result = DAG.getNode(X86ISD::WrapperRIP, dl, PtrVT, Result);
11429 Result = DAG.getNode(X86ISD::Wrapper, dl, PtrVT, Result);
11431 // With PIC, the address is actually $g + Offset.
11432 if (isGlobalRelativeToPICBase(OpFlags)) {
11433 Result = DAG.getNode(ISD::ADD, dl, PtrVT,
11434 DAG.getNode(X86ISD::GlobalBaseReg, dl, PtrVT), Result);
11441 X86TargetLowering::LowerGlobalAddress(const GlobalValue *GV, SDLoc dl,
11442 int64_t Offset, SelectionDAG &DAG) const {
11443 // Create the TargetGlobalAddress node, folding in the constant
11444 // offset if it is legal.
11445 unsigned char OpFlags =
11446 Subtarget->ClassifyGlobalReference(GV, DAG.getTarget());
11447 CodeModel::Model M = DAG.getTarget().getCodeModel();
11448 auto PtrVT = getPointerTy(DAG.getDataLayout());
11450 if (OpFlags == X86II::MO_NO_FLAG &&
11451 X86::isOffsetSuitableForCodeModel(Offset, M)) {
11452 // A direct static reference to a global.
11453 Result = DAG.getTargetGlobalAddress(GV, dl, PtrVT, Offset);
11456 Result = DAG.getTargetGlobalAddress(GV, dl, PtrVT, 0, OpFlags);
11459 if (Subtarget->isPICStyleRIPRel() &&
11460 (M == CodeModel::Small || M == CodeModel::Kernel))
11461 Result = DAG.getNode(X86ISD::WrapperRIP, dl, PtrVT, Result);
11463 Result = DAG.getNode(X86ISD::Wrapper, dl, PtrVT, Result);
11465 // With PIC, the address is actually $g + Offset.
11466 if (isGlobalRelativeToPICBase(OpFlags)) {
11467 Result = DAG.getNode(ISD::ADD, dl, PtrVT,
11468 DAG.getNode(X86ISD::GlobalBaseReg, dl, PtrVT), Result);
11471 // For globals that require a load from a stub to get the address, emit the
11473 if (isGlobalStubReference(OpFlags))
11474 Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Result,
11475 MachinePointerInfo::getGOT(), false, false, false, 0);
11477 // If there was a non-zero offset that we didn't fold, create an explicit
11478 // addition for it.
11480 Result = DAG.getNode(ISD::ADD, dl, PtrVT, Result,
11481 DAG.getConstant(Offset, dl, PtrVT));
11487 X86TargetLowering::LowerGlobalAddress(SDValue Op, SelectionDAG &DAG) const {
11488 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
11489 int64_t Offset = cast<GlobalAddressSDNode>(Op)->getOffset();
11490 return LowerGlobalAddress(GV, SDLoc(Op), Offset, DAG);
11494 GetTLSADDR(SelectionDAG &DAG, SDValue Chain, GlobalAddressSDNode *GA,
11495 SDValue *InFlag, const EVT PtrVT, unsigned ReturnReg,
11496 unsigned char OperandFlags, bool LocalDynamic = false) {
11497 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
11498 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
11500 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(), dl,
11501 GA->getValueType(0),
11505 X86ISD::NodeType CallType = LocalDynamic ? X86ISD::TLSBASEADDR
11509 SDValue Ops[] = { Chain, TGA, *InFlag };
11510 Chain = DAG.getNode(CallType, dl, NodeTys, Ops);
11512 SDValue Ops[] = { Chain, TGA };
11513 Chain = DAG.getNode(CallType, dl, NodeTys, Ops);
11516 // TLSADDR will be codegen'ed as call. Inform MFI that function has calls.
11517 MFI->setAdjustsStack(true);
11518 MFI->setHasCalls(true);
11520 SDValue Flag = Chain.getValue(1);
11521 return DAG.getCopyFromReg(Chain, dl, ReturnReg, PtrVT, Flag);
11524 // Lower ISD::GlobalTLSAddress using the "general dynamic" model, 32 bit
11526 LowerToTLSGeneralDynamicModel32(GlobalAddressSDNode *GA, SelectionDAG &DAG,
11529 SDLoc dl(GA); // ? function entry point might be better
11530 SDValue Chain = DAG.getCopyToReg(DAG.getEntryNode(), dl, X86::EBX,
11531 DAG.getNode(X86ISD::GlobalBaseReg,
11532 SDLoc(), PtrVT), InFlag);
11533 InFlag = Chain.getValue(1);
11535 return GetTLSADDR(DAG, Chain, GA, &InFlag, PtrVT, X86::EAX, X86II::MO_TLSGD);
11538 // Lower ISD::GlobalTLSAddress using the "general dynamic" model, 64 bit
11540 LowerToTLSGeneralDynamicModel64(GlobalAddressSDNode *GA, SelectionDAG &DAG,
11542 return GetTLSADDR(DAG, DAG.getEntryNode(), GA, nullptr, PtrVT,
11543 X86::RAX, X86II::MO_TLSGD);
11546 static SDValue LowerToTLSLocalDynamicModel(GlobalAddressSDNode *GA,
11552 // Get the start address of the TLS block for this module.
11553 X86MachineFunctionInfo* MFI = DAG.getMachineFunction()
11554 .getInfo<X86MachineFunctionInfo>();
11555 MFI->incNumLocalDynamicTLSAccesses();
11559 Base = GetTLSADDR(DAG, DAG.getEntryNode(), GA, nullptr, PtrVT, X86::RAX,
11560 X86II::MO_TLSLD, /*LocalDynamic=*/true);
11563 SDValue Chain = DAG.getCopyToReg(DAG.getEntryNode(), dl, X86::EBX,
11564 DAG.getNode(X86ISD::GlobalBaseReg, SDLoc(), PtrVT), InFlag);
11565 InFlag = Chain.getValue(1);
11566 Base = GetTLSADDR(DAG, Chain, GA, &InFlag, PtrVT, X86::EAX,
11567 X86II::MO_TLSLDM, /*LocalDynamic=*/true);
11570 // Note: the CleanupLocalDynamicTLSPass will remove redundant computations
11574 unsigned char OperandFlags = X86II::MO_DTPOFF;
11575 unsigned WrapperKind = X86ISD::Wrapper;
11576 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(), dl,
11577 GA->getValueType(0),
11578 GA->getOffset(), OperandFlags);
11579 SDValue Offset = DAG.getNode(WrapperKind, dl, PtrVT, TGA);
11581 // Add x@dtpoff with the base.
11582 return DAG.getNode(ISD::ADD, dl, PtrVT, Offset, Base);
11585 // Lower ISD::GlobalTLSAddress using the "initial exec" or "local exec" model.
11586 static SDValue LowerToTLSExecModel(GlobalAddressSDNode *GA, SelectionDAG &DAG,
11587 const EVT PtrVT, TLSModel::Model model,
11588 bool is64Bit, bool isPIC) {
11591 // Get the Thread Pointer, which is %gs:0 (32-bit) or %fs:0 (64-bit).
11592 Value *Ptr = Constant::getNullValue(Type::getInt8PtrTy(*DAG.getContext(),
11593 is64Bit ? 257 : 256));
11595 SDValue ThreadPointer =
11596 DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), DAG.getIntPtrConstant(0, dl),
11597 MachinePointerInfo(Ptr), false, false, false, 0);
11599 unsigned char OperandFlags = 0;
11600 // Most TLS accesses are not RIP relative, even on x86-64. One exception is
11602 unsigned WrapperKind = X86ISD::Wrapper;
11603 if (model == TLSModel::LocalExec) {
11604 OperandFlags = is64Bit ? X86II::MO_TPOFF : X86II::MO_NTPOFF;
11605 } else if (model == TLSModel::InitialExec) {
11607 OperandFlags = X86II::MO_GOTTPOFF;
11608 WrapperKind = X86ISD::WrapperRIP;
11610 OperandFlags = isPIC ? X86II::MO_GOTNTPOFF : X86II::MO_INDNTPOFF;
11613 llvm_unreachable("Unexpected model");
11616 // emit "addl x@ntpoff,%eax" (local exec)
11617 // or "addl x@indntpoff,%eax" (initial exec)
11618 // or "addl x@gotntpoff(%ebx) ,%eax" (initial exec, 32-bit pic)
11620 DAG.getTargetGlobalAddress(GA->getGlobal(), dl, GA->getValueType(0),
11621 GA->getOffset(), OperandFlags);
11622 SDValue Offset = DAG.getNode(WrapperKind, dl, PtrVT, TGA);
11624 if (model == TLSModel::InitialExec) {
11625 if (isPIC && !is64Bit) {
11626 Offset = DAG.getNode(ISD::ADD, dl, PtrVT,
11627 DAG.getNode(X86ISD::GlobalBaseReg, SDLoc(), PtrVT),
11631 Offset = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Offset,
11632 MachinePointerInfo::getGOT(), false, false, false, 0);
11635 // The address of the thread local variable is the add of the thread
11636 // pointer with the offset of the variable.
11637 return DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, Offset);
11641 X86TargetLowering::LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const {
11643 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
11644 const GlobalValue *GV = GA->getGlobal();
11645 auto PtrVT = getPointerTy(DAG.getDataLayout());
11647 if (Subtarget->isTargetELF()) {
11648 TLSModel::Model model = DAG.getTarget().getTLSModel(GV);
11650 case TLSModel::GeneralDynamic:
11651 if (Subtarget->is64Bit())
11652 return LowerToTLSGeneralDynamicModel64(GA, DAG, PtrVT);
11653 return LowerToTLSGeneralDynamicModel32(GA, DAG, PtrVT);
11654 case TLSModel::LocalDynamic:
11655 return LowerToTLSLocalDynamicModel(GA, DAG, PtrVT,
11656 Subtarget->is64Bit());
11657 case TLSModel::InitialExec:
11658 case TLSModel::LocalExec:
11659 return LowerToTLSExecModel(GA, DAG, PtrVT, model, Subtarget->is64Bit(),
11660 DAG.getTarget().getRelocationModel() ==
11663 llvm_unreachable("Unknown TLS model.");
11666 if (Subtarget->isTargetDarwin()) {
11667 // Darwin only has one model of TLS. Lower to that.
11668 unsigned char OpFlag = 0;
11669 unsigned WrapperKind = Subtarget->isPICStyleRIPRel() ?
11670 X86ISD::WrapperRIP : X86ISD::Wrapper;
11672 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
11673 // global base reg.
11674 bool PIC32 = (DAG.getTarget().getRelocationModel() == Reloc::PIC_) &&
11675 !Subtarget->is64Bit();
11677 OpFlag = X86II::MO_TLVP_PIC_BASE;
11679 OpFlag = X86II::MO_TLVP;
11681 SDValue Result = DAG.getTargetGlobalAddress(GA->getGlobal(), DL,
11682 GA->getValueType(0),
11683 GA->getOffset(), OpFlag);
11684 SDValue Offset = DAG.getNode(WrapperKind, DL, PtrVT, Result);
11686 // With PIC32, the address is actually $g + Offset.
11688 Offset = DAG.getNode(ISD::ADD, DL, PtrVT,
11689 DAG.getNode(X86ISD::GlobalBaseReg, SDLoc(), PtrVT),
11692 // Lowering the machine isd will make sure everything is in the right
11694 SDValue Chain = DAG.getEntryNode();
11695 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
11696 SDValue Args[] = { Chain, Offset };
11697 Chain = DAG.getNode(X86ISD::TLSCALL, DL, NodeTys, Args);
11699 // TLSCALL will be codegen'ed as call. Inform MFI that function has calls.
11700 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
11701 MFI->setAdjustsStack(true);
11703 // And our return value (tls address) is in the standard call return value
11705 unsigned Reg = Subtarget->is64Bit() ? X86::RAX : X86::EAX;
11706 return DAG.getCopyFromReg(Chain, DL, Reg, PtrVT, Chain.getValue(1));
11709 if (Subtarget->isTargetKnownWindowsMSVC() ||
11710 Subtarget->isTargetWindowsGNU()) {
11711 // Just use the implicit TLS architecture
11712 // Need to generate someting similar to:
11713 // mov rdx, qword [gs:abs 58H]; Load pointer to ThreadLocalStorage
11715 // mov ecx, dword [rel _tls_index]: Load index (from C runtime)
11716 // mov rcx, qword [rdx+rcx*8]
11717 // mov eax, .tls$:tlsvar
11718 // [rax+rcx] contains the address
11719 // Windows 64bit: gs:0x58
11720 // Windows 32bit: fs:__tls_array
11723 SDValue Chain = DAG.getEntryNode();
11725 // Get the Thread Pointer, which is %fs:__tls_array (32-bit) or
11726 // %gs:0x58 (64-bit). On MinGW, __tls_array is not available, so directly
11727 // use its literal value of 0x2C.
11728 Value *Ptr = Constant::getNullValue(Subtarget->is64Bit()
11729 ? Type::getInt8PtrTy(*DAG.getContext(),
11731 : Type::getInt32PtrTy(*DAG.getContext(),
11734 SDValue TlsArray = Subtarget->is64Bit()
11735 ? DAG.getIntPtrConstant(0x58, dl)
11736 : (Subtarget->isTargetWindowsGNU()
11737 ? DAG.getIntPtrConstant(0x2C, dl)
11738 : DAG.getExternalSymbol("_tls_array", PtrVT));
11740 SDValue ThreadPointer =
11741 DAG.getLoad(PtrVT, dl, Chain, TlsArray, MachinePointerInfo(Ptr), false,
11745 if (GV->getThreadLocalMode() == GlobalVariable::LocalExecTLSModel) {
11746 res = ThreadPointer;
11748 // Load the _tls_index variable
11749 SDValue IDX = DAG.getExternalSymbol("_tls_index", PtrVT);
11750 if (Subtarget->is64Bit())
11751 IDX = DAG.getExtLoad(ISD::ZEXTLOAD, dl, PtrVT, Chain, IDX,
11752 MachinePointerInfo(), MVT::i32, false, false,
11755 IDX = DAG.getLoad(PtrVT, dl, Chain, IDX, MachinePointerInfo(), false,
11758 auto &DL = DAG.getDataLayout();
11760 DAG.getConstant(Log2_64_Ceil(DL.getPointerSize()), dl, PtrVT);
11761 IDX = DAG.getNode(ISD::SHL, dl, PtrVT, IDX, Scale);
11763 res = DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, IDX);
11766 res = DAG.getLoad(PtrVT, dl, Chain, res, MachinePointerInfo(), false, false,
11769 // Get the offset of start of .tls section
11770 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(), dl,
11771 GA->getValueType(0),
11772 GA->getOffset(), X86II::MO_SECREL);
11773 SDValue Offset = DAG.getNode(X86ISD::Wrapper, dl, PtrVT, TGA);
11775 // The address of the thread local variable is the add of the thread
11776 // pointer with the offset of the variable.
11777 return DAG.getNode(ISD::ADD, dl, PtrVT, res, Offset);
11780 llvm_unreachable("TLS not implemented for this target.");
11783 /// LowerShiftParts - Lower SRA_PARTS and friends, which return two i32 values
11784 /// and take a 2 x i32 value to shift plus a shift amount.
11785 static SDValue LowerShiftParts(SDValue Op, SelectionDAG &DAG) {
11786 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
11787 MVT VT = Op.getSimpleValueType();
11788 unsigned VTBits = VT.getSizeInBits();
11790 bool isSRA = Op.getOpcode() == ISD::SRA_PARTS;
11791 SDValue ShOpLo = Op.getOperand(0);
11792 SDValue ShOpHi = Op.getOperand(1);
11793 SDValue ShAmt = Op.getOperand(2);
11794 // X86ISD::SHLD and X86ISD::SHRD have defined overflow behavior but the
11795 // generic ISD nodes haven't. Insert an AND to be safe, it's optimized away
11797 SDValue SafeShAmt = DAG.getNode(ISD::AND, dl, MVT::i8, ShAmt,
11798 DAG.getConstant(VTBits - 1, dl, MVT::i8));
11799 SDValue Tmp1 = isSRA ? DAG.getNode(ISD::SRA, dl, VT, ShOpHi,
11800 DAG.getConstant(VTBits - 1, dl, MVT::i8))
11801 : DAG.getConstant(0, dl, VT);
11803 SDValue Tmp2, Tmp3;
11804 if (Op.getOpcode() == ISD::SHL_PARTS) {
11805 Tmp2 = DAG.getNode(X86ISD::SHLD, dl, VT, ShOpHi, ShOpLo, ShAmt);
11806 Tmp3 = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, SafeShAmt);
11808 Tmp2 = DAG.getNode(X86ISD::SHRD, dl, VT, ShOpLo, ShOpHi, ShAmt);
11809 Tmp3 = DAG.getNode(isSRA ? ISD::SRA : ISD::SRL, dl, VT, ShOpHi, SafeShAmt);
11812 // If the shift amount is larger or equal than the width of a part we can't
11813 // rely on the results of shld/shrd. Insert a test and select the appropriate
11814 // values for large shift amounts.
11815 SDValue AndNode = DAG.getNode(ISD::AND, dl, MVT::i8, ShAmt,
11816 DAG.getConstant(VTBits, dl, MVT::i8));
11817 SDValue Cond = DAG.getNode(X86ISD::CMP, dl, MVT::i32,
11818 AndNode, DAG.getConstant(0, dl, MVT::i8));
11821 SDValue CC = DAG.getConstant(X86::COND_NE, dl, MVT::i8);
11822 SDValue Ops0[4] = { Tmp2, Tmp3, CC, Cond };
11823 SDValue Ops1[4] = { Tmp3, Tmp1, CC, Cond };
11825 if (Op.getOpcode() == ISD::SHL_PARTS) {
11826 Hi = DAG.getNode(X86ISD::CMOV, dl, VT, Ops0);
11827 Lo = DAG.getNode(X86ISD::CMOV, dl, VT, Ops1);
11829 Lo = DAG.getNode(X86ISD::CMOV, dl, VT, Ops0);
11830 Hi = DAG.getNode(X86ISD::CMOV, dl, VT, Ops1);
11833 SDValue Ops[2] = { Lo, Hi };
11834 return DAG.getMergeValues(Ops, dl);
11837 SDValue X86TargetLowering::LowerSINT_TO_FP(SDValue Op,
11838 SelectionDAG &DAG) const {
11839 SDValue Src = Op.getOperand(0);
11840 MVT SrcVT = Src.getSimpleValueType();
11841 MVT VT = Op.getSimpleValueType();
11844 if (SrcVT.isVector()) {
11845 if (SrcVT == MVT::v2i32 && VT == MVT::v2f64) {
11846 return DAG.getNode(X86ISD::CVTDQ2PD, dl, VT,
11847 DAG.getNode(ISD::CONCAT_VECTORS, dl, MVT::v4i32, Src,
11848 DAG.getUNDEF(SrcVT)));
11850 if (SrcVT.getVectorElementType() == MVT::i1) {
11851 MVT IntegerVT = MVT::getVectorVT(MVT::i32, SrcVT.getVectorNumElements());
11852 return DAG.getNode(ISD::SINT_TO_FP, dl, Op.getValueType(),
11853 DAG.getNode(ISD::SIGN_EXTEND, dl, IntegerVT, Src));
11858 assert(SrcVT <= MVT::i64 && SrcVT >= MVT::i16 &&
11859 "Unknown SINT_TO_FP to lower!");
11861 // These are really Legal; return the operand so the caller accepts it as
11863 if (SrcVT == MVT::i32 && isScalarFPTypeInSSEReg(Op.getValueType()))
11865 if (SrcVT == MVT::i64 && isScalarFPTypeInSSEReg(Op.getValueType()) &&
11866 Subtarget->is64Bit()) {
11870 unsigned Size = SrcVT.getSizeInBits()/8;
11871 MachineFunction &MF = DAG.getMachineFunction();
11872 auto PtrVT = getPointerTy(MF.getDataLayout());
11873 int SSFI = MF.getFrameInfo()->CreateStackObject(Size, Size, false);
11874 SDValue StackSlot = DAG.getFrameIndex(SSFI, PtrVT);
11875 SDValue Chain = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
11877 MachinePointerInfo::getFixedStack(SSFI),
11879 return BuildFILD(Op, SrcVT, Chain, StackSlot, DAG);
11882 SDValue X86TargetLowering::BuildFILD(SDValue Op, EVT SrcVT, SDValue Chain,
11884 SelectionDAG &DAG) const {
11888 bool useSSE = isScalarFPTypeInSSEReg(Op.getValueType());
11890 Tys = DAG.getVTList(MVT::f64, MVT::Other, MVT::Glue);
11892 Tys = DAG.getVTList(Op.getValueType(), MVT::Other);
11894 unsigned ByteSize = SrcVT.getSizeInBits()/8;
11896 FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(StackSlot);
11897 MachineMemOperand *MMO;
11899 int SSFI = FI->getIndex();
11901 DAG.getMachineFunction()
11902 .getMachineMemOperand(MachinePointerInfo::getFixedStack(SSFI),
11903 MachineMemOperand::MOLoad, ByteSize, ByteSize);
11905 MMO = cast<LoadSDNode>(StackSlot)->getMemOperand();
11906 StackSlot = StackSlot.getOperand(1);
11908 SDValue Ops[] = { Chain, StackSlot, DAG.getValueType(SrcVT) };
11909 SDValue Result = DAG.getMemIntrinsicNode(useSSE ? X86ISD::FILD_FLAG :
11911 Tys, Ops, SrcVT, MMO);
11914 Chain = Result.getValue(1);
11915 SDValue InFlag = Result.getValue(2);
11917 // FIXME: Currently the FST is flagged to the FILD_FLAG. This
11918 // shouldn't be necessary except that RFP cannot be live across
11919 // multiple blocks. When stackifier is fixed, they can be uncoupled.
11920 MachineFunction &MF = DAG.getMachineFunction();
11921 unsigned SSFISize = Op.getValueType().getSizeInBits()/8;
11922 int SSFI = MF.getFrameInfo()->CreateStackObject(SSFISize, SSFISize, false);
11923 auto PtrVT = getPointerTy(MF.getDataLayout());
11924 SDValue StackSlot = DAG.getFrameIndex(SSFI, PtrVT);
11925 Tys = DAG.getVTList(MVT::Other);
11927 Chain, Result, StackSlot, DAG.getValueType(Op.getValueType()), InFlag
11929 MachineMemOperand *MMO =
11930 DAG.getMachineFunction()
11931 .getMachineMemOperand(MachinePointerInfo::getFixedStack(SSFI),
11932 MachineMemOperand::MOStore, SSFISize, SSFISize);
11934 Chain = DAG.getMemIntrinsicNode(X86ISD::FST, DL, Tys,
11935 Ops, Op.getValueType(), MMO);
11936 Result = DAG.getLoad(Op.getValueType(), DL, Chain, StackSlot,
11937 MachinePointerInfo::getFixedStack(SSFI),
11938 false, false, false, 0);
11944 // LowerUINT_TO_FP_i64 - 64-bit unsigned integer to double expansion.
11945 SDValue X86TargetLowering::LowerUINT_TO_FP_i64(SDValue Op,
11946 SelectionDAG &DAG) const {
11947 // This algorithm is not obvious. Here it is what we're trying to output:
11950 punpckldq (c0), %xmm0 // c0: (uint4){ 0x43300000U, 0x45300000U, 0U, 0U }
11951 subpd (c1), %xmm0 // c1: (double2){ 0x1.0p52, 0x1.0p52 * 0x1.0p32 }
11953 haddpd %xmm0, %xmm0
11955 pshufd $0x4e, %xmm0, %xmm1
11961 LLVMContext *Context = DAG.getContext();
11963 // Build some magic constants.
11964 static const uint32_t CV0[] = { 0x43300000, 0x45300000, 0, 0 };
11965 Constant *C0 = ConstantDataVector::get(*Context, CV0);
11966 auto PtrVT = getPointerTy(DAG.getDataLayout());
11967 SDValue CPIdx0 = DAG.getConstantPool(C0, PtrVT, 16);
11969 SmallVector<Constant*,2> CV1;
11971 ConstantFP::get(*Context, APFloat(APFloat::IEEEdouble,
11972 APInt(64, 0x4330000000000000ULL))));
11974 ConstantFP::get(*Context, APFloat(APFloat::IEEEdouble,
11975 APInt(64, 0x4530000000000000ULL))));
11976 Constant *C1 = ConstantVector::get(CV1);
11977 SDValue CPIdx1 = DAG.getConstantPool(C1, PtrVT, 16);
11979 // Load the 64-bit value into an XMM register.
11980 SDValue XR1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i64,
11982 SDValue CLod0 = DAG.getLoad(MVT::v4i32, dl, DAG.getEntryNode(), CPIdx0,
11983 MachinePointerInfo::getConstantPool(),
11984 false, false, false, 16);
11986 getUnpackl(DAG, dl, MVT::v4i32, DAG.getBitcast(MVT::v4i32, XR1), CLod0);
11988 SDValue CLod1 = DAG.getLoad(MVT::v2f64, dl, CLod0.getValue(1), CPIdx1,
11989 MachinePointerInfo::getConstantPool(),
11990 false, false, false, 16);
11991 SDValue XR2F = DAG.getBitcast(MVT::v2f64, Unpck1);
11992 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::v2f64, XR2F, CLod1);
11995 if (Subtarget->hasSSE3()) {
11996 // FIXME: The 'haddpd' instruction may be slower than 'movhlps + addsd'.
11997 Result = DAG.getNode(X86ISD::FHADD, dl, MVT::v2f64, Sub, Sub);
11999 SDValue S2F = DAG.getBitcast(MVT::v4i32, Sub);
12000 SDValue Shuffle = getTargetShuffleNode(X86ISD::PSHUFD, dl, MVT::v4i32,
12002 Result = DAG.getNode(ISD::FADD, dl, MVT::v2f64,
12003 DAG.getBitcast(MVT::v2f64, Shuffle), Sub);
12006 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Result,
12007 DAG.getIntPtrConstant(0, dl));
12010 // LowerUINT_TO_FP_i32 - 32-bit unsigned integer to float expansion.
12011 SDValue X86TargetLowering::LowerUINT_TO_FP_i32(SDValue Op,
12012 SelectionDAG &DAG) const {
12014 // FP constant to bias correct the final result.
12015 SDValue Bias = DAG.getConstantFP(BitsToDouble(0x4330000000000000ULL), dl,
12018 // Load the 32-bit value into an XMM register.
12019 SDValue Load = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
12022 // Zero out the upper parts of the register.
12023 Load = getShuffleVectorZeroOrUndef(Load, 0, true, Subtarget, DAG);
12025 Load = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
12026 DAG.getBitcast(MVT::v2f64, Load),
12027 DAG.getIntPtrConstant(0, dl));
12029 // Or the load with the bias.
12030 SDValue Or = DAG.getNode(
12031 ISD::OR, dl, MVT::v2i64,
12032 DAG.getBitcast(MVT::v2i64,
12033 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2f64, Load)),
12034 DAG.getBitcast(MVT::v2i64,
12035 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2f64, Bias)));
12037 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
12038 DAG.getBitcast(MVT::v2f64, Or), DAG.getIntPtrConstant(0, dl));
12040 // Subtract the bias.
12041 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::f64, Or, Bias);
12043 // Handle final rounding.
12044 EVT DestVT = Op.getValueType();
12046 if (DestVT.bitsLT(MVT::f64))
12047 return DAG.getNode(ISD::FP_ROUND, dl, DestVT, Sub,
12048 DAG.getIntPtrConstant(0, dl));
12049 if (DestVT.bitsGT(MVT::f64))
12050 return DAG.getNode(ISD::FP_EXTEND, dl, DestVT, Sub);
12052 // Handle final rounding.
12056 static SDValue lowerUINT_TO_FP_vXi32(SDValue Op, SelectionDAG &DAG,
12057 const X86Subtarget &Subtarget) {
12058 // The algorithm is the following:
12059 // #ifdef __SSE4_1__
12060 // uint4 lo = _mm_blend_epi16( v, (uint4) 0x4b000000, 0xaa);
12061 // uint4 hi = _mm_blend_epi16( _mm_srli_epi32(v,16),
12062 // (uint4) 0x53000000, 0xaa);
12064 // uint4 lo = (v & (uint4) 0xffff) | (uint4) 0x4b000000;
12065 // uint4 hi = (v >> 16) | (uint4) 0x53000000;
12067 // float4 fhi = (float4) hi - (0x1.0p39f + 0x1.0p23f);
12068 // return (float4) lo + fhi;
12071 SDValue V = Op->getOperand(0);
12072 EVT VecIntVT = V.getValueType();
12073 bool Is128 = VecIntVT == MVT::v4i32;
12074 EVT VecFloatVT = Is128 ? MVT::v4f32 : MVT::v8f32;
12075 // If we convert to something else than the supported type, e.g., to v4f64,
12077 if (VecFloatVT != Op->getValueType(0))
12080 unsigned NumElts = VecIntVT.getVectorNumElements();
12081 assert((VecIntVT == MVT::v4i32 || VecIntVT == MVT::v8i32) &&
12082 "Unsupported custom type");
12083 assert(NumElts <= 8 && "The size of the constant array must be fixed");
12085 // In the #idef/#else code, we have in common:
12086 // - The vector of constants:
12092 // Create the splat vector for 0x4b000000.
12093 SDValue CstLow = DAG.getConstant(0x4b000000, DL, MVT::i32);
12094 SDValue CstLowArray[] = {CstLow, CstLow, CstLow, CstLow,
12095 CstLow, CstLow, CstLow, CstLow};
12096 SDValue VecCstLow = DAG.getNode(ISD::BUILD_VECTOR, DL, VecIntVT,
12097 makeArrayRef(&CstLowArray[0], NumElts));
12098 // Create the splat vector for 0x53000000.
12099 SDValue CstHigh = DAG.getConstant(0x53000000, DL, MVT::i32);
12100 SDValue CstHighArray[] = {CstHigh, CstHigh, CstHigh, CstHigh,
12101 CstHigh, CstHigh, CstHigh, CstHigh};
12102 SDValue VecCstHigh = DAG.getNode(ISD::BUILD_VECTOR, DL, VecIntVT,
12103 makeArrayRef(&CstHighArray[0], NumElts));
12105 // Create the right shift.
12106 SDValue CstShift = DAG.getConstant(16, DL, MVT::i32);
12107 SDValue CstShiftArray[] = {CstShift, CstShift, CstShift, CstShift,
12108 CstShift, CstShift, CstShift, CstShift};
12109 SDValue VecCstShift = DAG.getNode(ISD::BUILD_VECTOR, DL, VecIntVT,
12110 makeArrayRef(&CstShiftArray[0], NumElts));
12111 SDValue HighShift = DAG.getNode(ISD::SRL, DL, VecIntVT, V, VecCstShift);
12114 if (Subtarget.hasSSE41()) {
12115 EVT VecI16VT = Is128 ? MVT::v8i16 : MVT::v16i16;
12116 // uint4 lo = _mm_blend_epi16( v, (uint4) 0x4b000000, 0xaa);
12117 SDValue VecCstLowBitcast = DAG.getBitcast(VecI16VT, VecCstLow);
12118 SDValue VecBitcast = DAG.getBitcast(VecI16VT, V);
12119 // Low will be bitcasted right away, so do not bother bitcasting back to its
12121 Low = DAG.getNode(X86ISD::BLENDI, DL, VecI16VT, VecBitcast,
12122 VecCstLowBitcast, DAG.getConstant(0xaa, DL, MVT::i32));
12123 // uint4 hi = _mm_blend_epi16( _mm_srli_epi32(v,16),
12124 // (uint4) 0x53000000, 0xaa);
12125 SDValue VecCstHighBitcast = DAG.getBitcast(VecI16VT, VecCstHigh);
12126 SDValue VecShiftBitcast = DAG.getBitcast(VecI16VT, HighShift);
12127 // High will be bitcasted right away, so do not bother bitcasting back to
12128 // its original type.
12129 High = DAG.getNode(X86ISD::BLENDI, DL, VecI16VT, VecShiftBitcast,
12130 VecCstHighBitcast, DAG.getConstant(0xaa, DL, MVT::i32));
12132 SDValue CstMask = DAG.getConstant(0xffff, DL, MVT::i32);
12133 SDValue VecCstMask = DAG.getNode(ISD::BUILD_VECTOR, DL, VecIntVT, CstMask,
12134 CstMask, CstMask, CstMask);
12135 // uint4 lo = (v & (uint4) 0xffff) | (uint4) 0x4b000000;
12136 SDValue LowAnd = DAG.getNode(ISD::AND, DL, VecIntVT, V, VecCstMask);
12137 Low = DAG.getNode(ISD::OR, DL, VecIntVT, LowAnd, VecCstLow);
12139 // uint4 hi = (v >> 16) | (uint4) 0x53000000;
12140 High = DAG.getNode(ISD::OR, DL, VecIntVT, HighShift, VecCstHigh);
12143 // Create the vector constant for -(0x1.0p39f + 0x1.0p23f).
12144 SDValue CstFAdd = DAG.getConstantFP(
12145 APFloat(APFloat::IEEEsingle, APInt(32, 0xD3000080)), DL, MVT::f32);
12146 SDValue CstFAddArray[] = {CstFAdd, CstFAdd, CstFAdd, CstFAdd,
12147 CstFAdd, CstFAdd, CstFAdd, CstFAdd};
12148 SDValue VecCstFAdd = DAG.getNode(ISD::BUILD_VECTOR, DL, VecFloatVT,
12149 makeArrayRef(&CstFAddArray[0], NumElts));
12151 // float4 fhi = (float4) hi - (0x1.0p39f + 0x1.0p23f);
12152 SDValue HighBitcast = DAG.getBitcast(VecFloatVT, High);
12154 DAG.getNode(ISD::FADD, DL, VecFloatVT, HighBitcast, VecCstFAdd);
12155 // return (float4) lo + fhi;
12156 SDValue LowBitcast = DAG.getBitcast(VecFloatVT, Low);
12157 return DAG.getNode(ISD::FADD, DL, VecFloatVT, LowBitcast, FHigh);
12160 SDValue X86TargetLowering::lowerUINT_TO_FP_vec(SDValue Op,
12161 SelectionDAG &DAG) const {
12162 SDValue N0 = Op.getOperand(0);
12163 MVT SVT = N0.getSimpleValueType();
12166 switch (SVT.SimpleTy) {
12168 llvm_unreachable("Custom UINT_TO_FP is not supported!");
12173 MVT NVT = MVT::getVectorVT(MVT::i32, SVT.getVectorNumElements());
12174 return DAG.getNode(ISD::SINT_TO_FP, dl, Op.getValueType(),
12175 DAG.getNode(ISD::ZERO_EXTEND, dl, NVT, N0));
12179 return lowerUINT_TO_FP_vXi32(Op, DAG, *Subtarget);
12182 if (Subtarget->hasAVX512())
12183 return DAG.getNode(ISD::UINT_TO_FP, dl, Op.getValueType(),
12184 DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v16i32, N0));
12186 llvm_unreachable(nullptr);
12189 SDValue X86TargetLowering::LowerUINT_TO_FP(SDValue Op,
12190 SelectionDAG &DAG) const {
12191 SDValue N0 = Op.getOperand(0);
12193 auto PtrVT = getPointerTy(DAG.getDataLayout());
12195 if (Op.getValueType().isVector())
12196 return lowerUINT_TO_FP_vec(Op, DAG);
12198 // Since UINT_TO_FP is legal (it's marked custom), dag combiner won't
12199 // optimize it to a SINT_TO_FP when the sign bit is known zero. Perform
12200 // the optimization here.
12201 if (DAG.SignBitIsZero(N0))
12202 return DAG.getNode(ISD::SINT_TO_FP, dl, Op.getValueType(), N0);
12204 MVT SrcVT = N0.getSimpleValueType();
12205 MVT DstVT = Op.getSimpleValueType();
12206 if (SrcVT == MVT::i64 && DstVT == MVT::f64 && X86ScalarSSEf64)
12207 return LowerUINT_TO_FP_i64(Op, DAG);
12208 if (SrcVT == MVT::i32 && X86ScalarSSEf64)
12209 return LowerUINT_TO_FP_i32(Op, DAG);
12210 if (Subtarget->is64Bit() && SrcVT == MVT::i64 && DstVT == MVT::f32)
12213 // Make a 64-bit buffer, and use it to build an FILD.
12214 SDValue StackSlot = DAG.CreateStackTemporary(MVT::i64);
12215 if (SrcVT == MVT::i32) {
12216 SDValue WordOff = DAG.getConstant(4, dl, PtrVT);
12217 SDValue OffsetSlot = DAG.getNode(ISD::ADD, dl, PtrVT, StackSlot, WordOff);
12218 SDValue Store1 = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
12219 StackSlot, MachinePointerInfo(),
12221 SDValue Store2 = DAG.getStore(Store1, dl, DAG.getConstant(0, dl, MVT::i32),
12222 OffsetSlot, MachinePointerInfo(),
12224 SDValue Fild = BuildFILD(Op, MVT::i64, Store2, StackSlot, DAG);
12228 assert(SrcVT == MVT::i64 && "Unexpected type in UINT_TO_FP");
12229 SDValue Store = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
12230 StackSlot, MachinePointerInfo(),
12232 // For i64 source, we need to add the appropriate power of 2 if the input
12233 // was negative. This is the same as the optimization in
12234 // DAGTypeLegalizer::ExpandIntOp_UNIT_TO_FP, and for it to be safe here,
12235 // we must be careful to do the computation in x87 extended precision, not
12236 // in SSE. (The generic code can't know it's OK to do this, or how to.)
12237 int SSFI = cast<FrameIndexSDNode>(StackSlot)->getIndex();
12238 MachineMemOperand *MMO =
12239 DAG.getMachineFunction()
12240 .getMachineMemOperand(MachinePointerInfo::getFixedStack(SSFI),
12241 MachineMemOperand::MOLoad, 8, 8);
12243 SDVTList Tys = DAG.getVTList(MVT::f80, MVT::Other);
12244 SDValue Ops[] = { Store, StackSlot, DAG.getValueType(MVT::i64) };
12245 SDValue Fild = DAG.getMemIntrinsicNode(X86ISD::FILD, dl, Tys, Ops,
12248 APInt FF(32, 0x5F800000ULL);
12250 // Check whether the sign bit is set.
12251 SDValue SignSet = DAG.getSetCC(
12252 dl, getSetCCResultType(DAG.getDataLayout(), *DAG.getContext(), MVT::i64),
12253 Op.getOperand(0), DAG.getConstant(0, dl, MVT::i64), ISD::SETLT);
12255 // Build a 64 bit pair (0, FF) in the constant pool, with FF in the lo bits.
12256 SDValue FudgePtr = DAG.getConstantPool(
12257 ConstantInt::get(*DAG.getContext(), FF.zext(64)), PtrVT);
12259 // Get a pointer to FF if the sign bit was set, or to 0 otherwise.
12260 SDValue Zero = DAG.getIntPtrConstant(0, dl);
12261 SDValue Four = DAG.getIntPtrConstant(4, dl);
12262 SDValue Offset = DAG.getNode(ISD::SELECT, dl, Zero.getValueType(), SignSet,
12264 FudgePtr = DAG.getNode(ISD::ADD, dl, PtrVT, FudgePtr, Offset);
12266 // Load the value out, extending it from f32 to f80.
12267 // FIXME: Avoid the extend by constructing the right constant pool?
12268 SDValue Fudge = DAG.getExtLoad(ISD::EXTLOAD, dl, MVT::f80, DAG.getEntryNode(),
12269 FudgePtr, MachinePointerInfo::getConstantPool(),
12270 MVT::f32, false, false, false, 4);
12271 // Extend everything to 80 bits to force it to be done on x87.
12272 SDValue Add = DAG.getNode(ISD::FADD, dl, MVT::f80, Fild, Fudge);
12273 return DAG.getNode(ISD::FP_ROUND, dl, DstVT, Add,
12274 DAG.getIntPtrConstant(0, dl));
12277 std::pair<SDValue,SDValue>
12278 X86TargetLowering:: FP_TO_INTHelper(SDValue Op, SelectionDAG &DAG,
12279 bool IsSigned, bool IsReplace) const {
12282 EVT DstTy = Op.getValueType();
12283 auto PtrVT = getPointerTy(DAG.getDataLayout());
12285 if (!IsSigned && !isIntegerTypeFTOL(DstTy)) {
12286 assert(DstTy == MVT::i32 && "Unexpected FP_TO_UINT");
12290 assert(DstTy.getSimpleVT() <= MVT::i64 &&
12291 DstTy.getSimpleVT() >= MVT::i16 &&
12292 "Unknown FP_TO_INT to lower!");
12294 // These are really Legal.
12295 if (DstTy == MVT::i32 &&
12296 isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType()))
12297 return std::make_pair(SDValue(), SDValue());
12298 if (Subtarget->is64Bit() &&
12299 DstTy == MVT::i64 &&
12300 isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType()))
12301 return std::make_pair(SDValue(), SDValue());
12303 // We lower FP->int64 either into FISTP64 followed by a load from a temporary
12304 // stack slot, or into the FTOL runtime function.
12305 MachineFunction &MF = DAG.getMachineFunction();
12306 unsigned MemSize = DstTy.getSizeInBits()/8;
12307 int SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize, false);
12308 SDValue StackSlot = DAG.getFrameIndex(SSFI, PtrVT);
12311 if (!IsSigned && isIntegerTypeFTOL(DstTy))
12312 Opc = X86ISD::WIN_FTOL;
12314 switch (DstTy.getSimpleVT().SimpleTy) {
12315 default: llvm_unreachable("Invalid FP_TO_SINT to lower!");
12316 case MVT::i16: Opc = X86ISD::FP_TO_INT16_IN_MEM; break;
12317 case MVT::i32: Opc = X86ISD::FP_TO_INT32_IN_MEM; break;
12318 case MVT::i64: Opc = X86ISD::FP_TO_INT64_IN_MEM; break;
12321 SDValue Chain = DAG.getEntryNode();
12322 SDValue Value = Op.getOperand(0);
12323 EVT TheVT = Op.getOperand(0).getValueType();
12324 // FIXME This causes a redundant load/store if the SSE-class value is already
12325 // in memory, such as if it is on the callstack.
12326 if (isScalarFPTypeInSSEReg(TheVT)) {
12327 assert(DstTy == MVT::i64 && "Invalid FP_TO_SINT to lower!");
12328 Chain = DAG.getStore(Chain, DL, Value, StackSlot,
12329 MachinePointerInfo::getFixedStack(SSFI),
12331 SDVTList Tys = DAG.getVTList(Op.getOperand(0).getValueType(), MVT::Other);
12333 Chain, StackSlot, DAG.getValueType(TheVT)
12336 MachineMemOperand *MMO =
12337 MF.getMachineMemOperand(MachinePointerInfo::getFixedStack(SSFI),
12338 MachineMemOperand::MOLoad, MemSize, MemSize);
12339 Value = DAG.getMemIntrinsicNode(X86ISD::FLD, DL, Tys, Ops, DstTy, MMO);
12340 Chain = Value.getValue(1);
12341 SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize, false);
12342 StackSlot = DAG.getFrameIndex(SSFI, PtrVT);
12345 MachineMemOperand *MMO =
12346 MF.getMachineMemOperand(MachinePointerInfo::getFixedStack(SSFI),
12347 MachineMemOperand::MOStore, MemSize, MemSize);
12349 if (Opc != X86ISD::WIN_FTOL) {
12350 // Build the FP_TO_INT*_IN_MEM
12351 SDValue Ops[] = { Chain, Value, StackSlot };
12352 SDValue FIST = DAG.getMemIntrinsicNode(Opc, DL, DAG.getVTList(MVT::Other),
12354 return std::make_pair(FIST, StackSlot);
12356 SDValue ftol = DAG.getNode(X86ISD::WIN_FTOL, DL,
12357 DAG.getVTList(MVT::Other, MVT::Glue),
12359 SDValue eax = DAG.getCopyFromReg(ftol, DL, X86::EAX,
12360 MVT::i32, ftol.getValue(1));
12361 SDValue edx = DAG.getCopyFromReg(eax.getValue(1), DL, X86::EDX,
12362 MVT::i32, eax.getValue(2));
12363 SDValue Ops[] = { eax, edx };
12364 SDValue pair = IsReplace
12365 ? DAG.getNode(ISD::BUILD_PAIR, DL, MVT::i64, Ops)
12366 : DAG.getMergeValues(Ops, DL);
12367 return std::make_pair(pair, SDValue());
12371 static SDValue LowerAVXExtend(SDValue Op, SelectionDAG &DAG,
12372 const X86Subtarget *Subtarget) {
12373 MVT VT = Op->getSimpleValueType(0);
12374 SDValue In = Op->getOperand(0);
12375 MVT InVT = In.getSimpleValueType();
12378 if (VT.is512BitVector() || InVT.getScalarType() == MVT::i1)
12379 return DAG.getNode(ISD::ZERO_EXTEND, dl, VT, In);
12381 // Optimize vectors in AVX mode:
12384 // Use vpunpcklwd for 4 lower elements v8i16 -> v4i32.
12385 // Use vpunpckhwd for 4 upper elements v8i16 -> v4i32.
12386 // Concat upper and lower parts.
12389 // Use vpunpckldq for 4 lower elements v4i32 -> v2i64.
12390 // Use vpunpckhdq for 4 upper elements v4i32 -> v2i64.
12391 // Concat upper and lower parts.
12394 if (((VT != MVT::v16i16) || (InVT != MVT::v16i8)) &&
12395 ((VT != MVT::v8i32) || (InVT != MVT::v8i16)) &&
12396 ((VT != MVT::v4i64) || (InVT != MVT::v4i32)))
12399 if (Subtarget->hasInt256())
12400 return DAG.getNode(X86ISD::VZEXT, dl, VT, In);
12402 SDValue ZeroVec = getZeroVector(InVT, Subtarget, DAG, dl);
12403 SDValue Undef = DAG.getUNDEF(InVT);
12404 bool NeedZero = Op.getOpcode() == ISD::ZERO_EXTEND;
12405 SDValue OpLo = getUnpackl(DAG, dl, InVT, In, NeedZero ? ZeroVec : Undef);
12406 SDValue OpHi = getUnpackh(DAG, dl, InVT, In, NeedZero ? ZeroVec : Undef);
12408 MVT HVT = MVT::getVectorVT(VT.getVectorElementType(),
12409 VT.getVectorNumElements()/2);
12411 OpLo = DAG.getBitcast(HVT, OpLo);
12412 OpHi = DAG.getBitcast(HVT, OpHi);
12414 return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT, OpLo, OpHi);
12417 static SDValue LowerZERO_EXTEND_AVX512(SDValue Op,
12418 const X86Subtarget *Subtarget, SelectionDAG &DAG) {
12419 MVT VT = Op->getSimpleValueType(0);
12420 SDValue In = Op->getOperand(0);
12421 MVT InVT = In.getSimpleValueType();
12423 unsigned int NumElts = VT.getVectorNumElements();
12424 if (NumElts != 8 && NumElts != 16 && !Subtarget->hasBWI())
12427 if (VT.is512BitVector() && InVT.getVectorElementType() != MVT::i1)
12428 return DAG.getNode(X86ISD::VZEXT, DL, VT, In);
12430 assert(InVT.getVectorElementType() == MVT::i1);
12431 MVT ExtVT = NumElts == 8 ? MVT::v8i64 : MVT::v16i32;
12433 DAG.getConstant(APInt(ExtVT.getScalarSizeInBits(), 1), DL, ExtVT);
12435 DAG.getConstant(APInt::getNullValue(ExtVT.getScalarSizeInBits()), DL, ExtVT);
12437 SDValue V = DAG.getNode(ISD::VSELECT, DL, ExtVT, In, One, Zero);
12438 if (VT.is512BitVector())
12440 return DAG.getNode(X86ISD::VTRUNC, DL, VT, V);
12443 static SDValue LowerANY_EXTEND(SDValue Op, const X86Subtarget *Subtarget,
12444 SelectionDAG &DAG) {
12445 if (Subtarget->hasFp256())
12446 if (SDValue Res = LowerAVXExtend(Op, DAG, Subtarget))
12452 static SDValue LowerZERO_EXTEND(SDValue Op, const X86Subtarget *Subtarget,
12453 SelectionDAG &DAG) {
12455 MVT VT = Op.getSimpleValueType();
12456 SDValue In = Op.getOperand(0);
12457 MVT SVT = In.getSimpleValueType();
12459 if (VT.is512BitVector() || SVT.getVectorElementType() == MVT::i1)
12460 return LowerZERO_EXTEND_AVX512(Op, Subtarget, DAG);
12462 if (Subtarget->hasFp256())
12463 if (SDValue Res = LowerAVXExtend(Op, DAG, Subtarget))
12466 assert(!VT.is256BitVector() || !SVT.is128BitVector() ||
12467 VT.getVectorNumElements() != SVT.getVectorNumElements());
12471 SDValue X86TargetLowering::LowerTRUNCATE(SDValue Op, SelectionDAG &DAG) const {
12473 MVT VT = Op.getSimpleValueType();
12474 SDValue In = Op.getOperand(0);
12475 MVT InVT = In.getSimpleValueType();
12477 if (VT == MVT::i1) {
12478 assert((InVT.isInteger() && (InVT.getSizeInBits() <= 64)) &&
12479 "Invalid scalar TRUNCATE operation");
12480 if (InVT.getSizeInBits() >= 32)
12482 In = DAG.getNode(ISD::ANY_EXTEND, DL, MVT::i32, In);
12483 return DAG.getNode(ISD::TRUNCATE, DL, VT, In);
12485 assert(VT.getVectorNumElements() == InVT.getVectorNumElements() &&
12486 "Invalid TRUNCATE operation");
12488 // move vector to mask - truncate solution for SKX
12489 if (VT.getVectorElementType() == MVT::i1) {
12490 if (InVT.is512BitVector() && InVT.getScalarSizeInBits() <= 16 &&
12491 Subtarget->hasBWI())
12492 return Op; // legal, will go to VPMOVB2M, VPMOVW2M
12493 if ((InVT.is256BitVector() || InVT.is128BitVector())
12494 && InVT.getScalarSizeInBits() <= 16 &&
12495 Subtarget->hasBWI() && Subtarget->hasVLX())
12496 return Op; // legal, will go to VPMOVB2M, VPMOVW2M
12497 if (InVT.is512BitVector() && InVT.getScalarSizeInBits() >= 32 &&
12498 Subtarget->hasDQI())
12499 return Op; // legal, will go to VPMOVD2M, VPMOVQ2M
12500 if ((InVT.is256BitVector() || InVT.is128BitVector())
12501 && InVT.getScalarSizeInBits() >= 32 &&
12502 Subtarget->hasDQI() && Subtarget->hasVLX())
12503 return Op; // legal, will go to VPMOVB2M, VPMOVQ2M
12506 if (VT.getVectorElementType() == MVT::i1) {
12507 assert(VT.getVectorElementType() == MVT::i1 && "Unexpected vector type");
12508 unsigned NumElts = InVT.getVectorNumElements();
12509 assert ((NumElts == 8 || NumElts == 16) && "Unexpected vector type");
12510 if (InVT.getSizeInBits() < 512) {
12511 MVT ExtVT = (NumElts == 16)? MVT::v16i32 : MVT::v8i64;
12512 In = DAG.getNode(ISD::SIGN_EXTEND, DL, ExtVT, In);
12517 DAG.getConstant(APInt::getSignBit(InVT.getScalarSizeInBits()), DL, InVT);
12518 SDValue And = DAG.getNode(ISD::AND, DL, InVT, OneV, In);
12519 return DAG.getNode(X86ISD::TESTM, DL, VT, And, And);
12522 // vpmovqb/w/d, vpmovdb/w, vpmovwb
12523 if (((!InVT.is512BitVector() && Subtarget->hasVLX()) || InVT.is512BitVector()) &&
12524 (InVT.getVectorElementType() != MVT::i16 || Subtarget->hasBWI()))
12525 return DAG.getNode(X86ISD::VTRUNC, DL, VT, In);
12527 if ((VT == MVT::v4i32) && (InVT == MVT::v4i64)) {
12528 // On AVX2, v4i64 -> v4i32 becomes VPERMD.
12529 if (Subtarget->hasInt256()) {
12530 static const int ShufMask[] = {0, 2, 4, 6, -1, -1, -1, -1};
12531 In = DAG.getBitcast(MVT::v8i32, In);
12532 In = DAG.getVectorShuffle(MVT::v8i32, DL, In, DAG.getUNDEF(MVT::v8i32),
12534 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, VT, In,
12535 DAG.getIntPtrConstant(0, DL));
12538 SDValue OpLo = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v2i64, In,
12539 DAG.getIntPtrConstant(0, DL));
12540 SDValue OpHi = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v2i64, In,
12541 DAG.getIntPtrConstant(2, DL));
12542 OpLo = DAG.getBitcast(MVT::v4i32, OpLo);
12543 OpHi = DAG.getBitcast(MVT::v4i32, OpHi);
12544 static const int ShufMask[] = {0, 2, 4, 6};
12545 return DAG.getVectorShuffle(VT, DL, OpLo, OpHi, ShufMask);
12548 if ((VT == MVT::v8i16) && (InVT == MVT::v8i32)) {
12549 // On AVX2, v8i32 -> v8i16 becomed PSHUFB.
12550 if (Subtarget->hasInt256()) {
12551 In = DAG.getBitcast(MVT::v32i8, In);
12553 SmallVector<SDValue,32> pshufbMask;
12554 for (unsigned i = 0; i < 2; ++i) {
12555 pshufbMask.push_back(DAG.getConstant(0x0, DL, MVT::i8));
12556 pshufbMask.push_back(DAG.getConstant(0x1, DL, MVT::i8));
12557 pshufbMask.push_back(DAG.getConstant(0x4, DL, MVT::i8));
12558 pshufbMask.push_back(DAG.getConstant(0x5, DL, MVT::i8));
12559 pshufbMask.push_back(DAG.getConstant(0x8, DL, MVT::i8));
12560 pshufbMask.push_back(DAG.getConstant(0x9, DL, MVT::i8));
12561 pshufbMask.push_back(DAG.getConstant(0xc, DL, MVT::i8));
12562 pshufbMask.push_back(DAG.getConstant(0xd, DL, MVT::i8));
12563 for (unsigned j = 0; j < 8; ++j)
12564 pshufbMask.push_back(DAG.getConstant(0x80, DL, MVT::i8));
12566 SDValue BV = DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v32i8, pshufbMask);
12567 In = DAG.getNode(X86ISD::PSHUFB, DL, MVT::v32i8, In, BV);
12568 In = DAG.getBitcast(MVT::v4i64, In);
12570 static const int ShufMask[] = {0, 2, -1, -1};
12571 In = DAG.getVectorShuffle(MVT::v4i64, DL, In, DAG.getUNDEF(MVT::v4i64),
12573 In = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v2i64, In,
12574 DAG.getIntPtrConstant(0, DL));
12575 return DAG.getBitcast(VT, In);
12578 SDValue OpLo = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v4i32, In,
12579 DAG.getIntPtrConstant(0, DL));
12581 SDValue OpHi = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v4i32, In,
12582 DAG.getIntPtrConstant(4, DL));
12584 OpLo = DAG.getBitcast(MVT::v16i8, OpLo);
12585 OpHi = DAG.getBitcast(MVT::v16i8, OpHi);
12587 // The PSHUFB mask:
12588 static const int ShufMask1[] = {0, 1, 4, 5, 8, 9, 12, 13,
12589 -1, -1, -1, -1, -1, -1, -1, -1};
12591 SDValue Undef = DAG.getUNDEF(MVT::v16i8);
12592 OpLo = DAG.getVectorShuffle(MVT::v16i8, DL, OpLo, Undef, ShufMask1);
12593 OpHi = DAG.getVectorShuffle(MVT::v16i8, DL, OpHi, Undef, ShufMask1);
12595 OpLo = DAG.getBitcast(MVT::v4i32, OpLo);
12596 OpHi = DAG.getBitcast(MVT::v4i32, OpHi);
12598 // The MOVLHPS Mask:
12599 static const int ShufMask2[] = {0, 1, 4, 5};
12600 SDValue res = DAG.getVectorShuffle(MVT::v4i32, DL, OpLo, OpHi, ShufMask2);
12601 return DAG.getBitcast(MVT::v8i16, res);
12604 // Handle truncation of V256 to V128 using shuffles.
12605 if (!VT.is128BitVector() || !InVT.is256BitVector())
12608 assert(Subtarget->hasFp256() && "256-bit vector without AVX!");
12610 unsigned NumElems = VT.getVectorNumElements();
12611 MVT NVT = MVT::getVectorVT(VT.getVectorElementType(), NumElems * 2);
12613 SmallVector<int, 16> MaskVec(NumElems * 2, -1);
12614 // Prepare truncation shuffle mask
12615 for (unsigned i = 0; i != NumElems; ++i)
12616 MaskVec[i] = i * 2;
12617 SDValue V = DAG.getVectorShuffle(NVT, DL, DAG.getBitcast(NVT, In),
12618 DAG.getUNDEF(NVT), &MaskVec[0]);
12619 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, VT, V,
12620 DAG.getIntPtrConstant(0, DL));
12623 SDValue X86TargetLowering::LowerFP_TO_SINT(SDValue Op,
12624 SelectionDAG &DAG) const {
12625 assert(!Op.getSimpleValueType().isVector());
12627 std::pair<SDValue,SDValue> Vals = FP_TO_INTHelper(Op, DAG,
12628 /*IsSigned=*/ true, /*IsReplace=*/ false);
12629 SDValue FIST = Vals.first, StackSlot = Vals.second;
12630 // If FP_TO_INTHelper failed, the node is actually supposed to be Legal.
12631 if (!FIST.getNode()) return Op;
12633 if (StackSlot.getNode())
12634 // Load the result.
12635 return DAG.getLoad(Op.getValueType(), SDLoc(Op),
12636 FIST, StackSlot, MachinePointerInfo(),
12637 false, false, false, 0);
12639 // The node is the result.
12643 SDValue X86TargetLowering::LowerFP_TO_UINT(SDValue Op,
12644 SelectionDAG &DAG) const {
12645 std::pair<SDValue,SDValue> Vals = FP_TO_INTHelper(Op, DAG,
12646 /*IsSigned=*/ false, /*IsReplace=*/ false);
12647 SDValue FIST = Vals.first, StackSlot = Vals.second;
12648 assert(FIST.getNode() && "Unexpected failure");
12650 if (StackSlot.getNode())
12651 // Load the result.
12652 return DAG.getLoad(Op.getValueType(), SDLoc(Op),
12653 FIST, StackSlot, MachinePointerInfo(),
12654 false, false, false, 0);
12656 // The node is the result.
12660 static SDValue LowerFP_EXTEND(SDValue Op, SelectionDAG &DAG) {
12662 MVT VT = Op.getSimpleValueType();
12663 SDValue In = Op.getOperand(0);
12664 MVT SVT = In.getSimpleValueType();
12666 assert(SVT == MVT::v2f32 && "Only customize MVT::v2f32 type legalization!");
12668 return DAG.getNode(X86ISD::VFPEXT, DL, VT,
12669 DAG.getNode(ISD::CONCAT_VECTORS, DL, MVT::v4f32,
12670 In, DAG.getUNDEF(SVT)));
12673 /// The only differences between FABS and FNEG are the mask and the logic op.
12674 /// FNEG also has a folding opportunity for FNEG(FABS(x)).
12675 static SDValue LowerFABSorFNEG(SDValue Op, SelectionDAG &DAG) {
12676 assert((Op.getOpcode() == ISD::FABS || Op.getOpcode() == ISD::FNEG) &&
12677 "Wrong opcode for lowering FABS or FNEG.");
12679 bool IsFABS = (Op.getOpcode() == ISD::FABS);
12681 // If this is a FABS and it has an FNEG user, bail out to fold the combination
12682 // into an FNABS. We'll lower the FABS after that if it is still in use.
12684 for (SDNode *User : Op->uses())
12685 if (User->getOpcode() == ISD::FNEG)
12689 MVT VT = Op.getSimpleValueType();
12691 // FIXME: Use function attribute "OptimizeForSize" and/or CodeGenOpt::Level to
12692 // decide if we should generate a 16-byte constant mask when we only need 4 or
12693 // 8 bytes for the scalar case.
12699 if (VT.isVector()) {
12701 EltVT = VT.getVectorElementType();
12702 NumElts = VT.getVectorNumElements();
12704 // There are no scalar bitwise logical SSE/AVX instructions, so we
12705 // generate a 16-byte vector constant and logic op even for the scalar case.
12706 // Using a 16-byte mask allows folding the load of the mask with
12707 // the logic op, so it can save (~4 bytes) on code size.
12708 LogicVT = (VT == MVT::f64) ? MVT::v2f64 : MVT::v4f32;
12710 NumElts = (VT == MVT::f64) ? 2 : 4;
12713 unsigned EltBits = EltVT.getSizeInBits();
12714 LLVMContext *Context = DAG.getContext();
12715 // For FABS, mask is 0x7f...; for FNEG, mask is 0x80...
12717 IsFABS ? APInt::getSignedMaxValue(EltBits) : APInt::getSignBit(EltBits);
12718 Constant *C = ConstantInt::get(*Context, MaskElt);
12719 C = ConstantVector::getSplat(NumElts, C);
12720 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
12721 SDValue CPIdx = DAG.getConstantPool(C, TLI.getPointerTy(DAG.getDataLayout()));
12722 unsigned Alignment = cast<ConstantPoolSDNode>(CPIdx)->getAlignment();
12723 SDValue Mask = DAG.getLoad(LogicVT, dl, DAG.getEntryNode(), CPIdx,
12724 MachinePointerInfo::getConstantPool(),
12725 false, false, false, Alignment);
12727 SDValue Op0 = Op.getOperand(0);
12728 bool IsFNABS = !IsFABS && (Op0.getOpcode() == ISD::FABS);
12730 IsFABS ? X86ISD::FAND : IsFNABS ? X86ISD::FOR : X86ISD::FXOR;
12731 SDValue Operand = IsFNABS ? Op0.getOperand(0) : Op0;
12734 return DAG.getNode(LogicOp, dl, LogicVT, Operand, Mask);
12736 // For the scalar case extend to a 128-bit vector, perform the logic op,
12737 // and extract the scalar result back out.
12738 Operand = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, LogicVT, Operand);
12739 SDValue LogicNode = DAG.getNode(LogicOp, dl, LogicVT, Operand, Mask);
12740 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, LogicNode,
12741 DAG.getIntPtrConstant(0, dl));
12744 static SDValue LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) {
12745 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
12746 LLVMContext *Context = DAG.getContext();
12747 SDValue Op0 = Op.getOperand(0);
12748 SDValue Op1 = Op.getOperand(1);
12750 MVT VT = Op.getSimpleValueType();
12751 MVT SrcVT = Op1.getSimpleValueType();
12753 // If second operand is smaller, extend it first.
12754 if (SrcVT.bitsLT(VT)) {
12755 Op1 = DAG.getNode(ISD::FP_EXTEND, dl, VT, Op1);
12758 // And if it is bigger, shrink it first.
12759 if (SrcVT.bitsGT(VT)) {
12760 Op1 = DAG.getNode(ISD::FP_ROUND, dl, VT, Op1, DAG.getIntPtrConstant(1, dl));
12764 // At this point the operands and the result should have the same
12765 // type, and that won't be f80 since that is not custom lowered.
12767 const fltSemantics &Sem =
12768 VT == MVT::f64 ? APFloat::IEEEdouble : APFloat::IEEEsingle;
12769 const unsigned SizeInBits = VT.getSizeInBits();
12771 SmallVector<Constant *, 4> CV(
12772 VT == MVT::f64 ? 2 : 4,
12773 ConstantFP::get(*Context, APFloat(Sem, APInt(SizeInBits, 0))));
12775 // First, clear all bits but the sign bit from the second operand (sign).
12776 CV[0] = ConstantFP::get(*Context,
12777 APFloat(Sem, APInt::getHighBitsSet(SizeInBits, 1)));
12778 Constant *C = ConstantVector::get(CV);
12779 auto PtrVT = TLI.getPointerTy(DAG.getDataLayout());
12780 SDValue CPIdx = DAG.getConstantPool(C, PtrVT, 16);
12782 // Perform all logic operations as 16-byte vectors because there are no
12783 // scalar FP logic instructions in SSE. This allows load folding of the
12784 // constants into the logic instructions.
12785 MVT LogicVT = (VT == MVT::f64) ? MVT::v2f64 : MVT::v4f32;
12786 SDValue Mask1 = DAG.getLoad(LogicVT, dl, DAG.getEntryNode(), CPIdx,
12787 MachinePointerInfo::getConstantPool(),
12788 false, false, false, 16);
12789 Op1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, LogicVT, Op1);
12790 SDValue SignBit = DAG.getNode(X86ISD::FAND, dl, LogicVT, Op1, Mask1);
12792 // Next, clear the sign bit from the first operand (magnitude).
12793 // If it's a constant, we can clear it here.
12794 if (ConstantFPSDNode *Op0CN = dyn_cast<ConstantFPSDNode>(Op0)) {
12795 APFloat APF = Op0CN->getValueAPF();
12796 // If the magnitude is a positive zero, the sign bit alone is enough.
12797 if (APF.isPosZero())
12798 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, SrcVT, SignBit,
12799 DAG.getIntPtrConstant(0, dl));
12801 CV[0] = ConstantFP::get(*Context, APF);
12803 CV[0] = ConstantFP::get(
12805 APFloat(Sem, APInt::getLowBitsSet(SizeInBits, SizeInBits - 1)));
12807 C = ConstantVector::get(CV);
12808 CPIdx = DAG.getConstantPool(C, PtrVT, 16);
12809 SDValue Val = DAG.getLoad(LogicVT, dl, DAG.getEntryNode(), CPIdx,
12810 MachinePointerInfo::getConstantPool(),
12811 false, false, false, 16);
12812 // If the magnitude operand wasn't a constant, we need to AND out the sign.
12813 if (!isa<ConstantFPSDNode>(Op0)) {
12814 Op0 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, LogicVT, Op0);
12815 Val = DAG.getNode(X86ISD::FAND, dl, LogicVT, Op0, Val);
12817 // OR the magnitude value with the sign bit.
12818 Val = DAG.getNode(X86ISD::FOR, dl, LogicVT, Val, SignBit);
12819 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, SrcVT, Val,
12820 DAG.getIntPtrConstant(0, dl));
12823 static SDValue LowerFGETSIGN(SDValue Op, SelectionDAG &DAG) {
12824 SDValue N0 = Op.getOperand(0);
12826 MVT VT = Op.getSimpleValueType();
12828 // Lower ISD::FGETSIGN to (AND (X86ISD::FGETSIGNx86 ...) 1).
12829 SDValue xFGETSIGN = DAG.getNode(X86ISD::FGETSIGNx86, dl, VT, N0,
12830 DAG.getConstant(1, dl, VT));
12831 return DAG.getNode(ISD::AND, dl, VT, xFGETSIGN, DAG.getConstant(1, dl, VT));
12834 // Check whether an OR'd tree is PTEST-able.
12835 static SDValue LowerVectorAllZeroTest(SDValue Op, const X86Subtarget *Subtarget,
12836 SelectionDAG &DAG) {
12837 assert(Op.getOpcode() == ISD::OR && "Only check OR'd tree.");
12839 if (!Subtarget->hasSSE41())
12842 if (!Op->hasOneUse())
12845 SDNode *N = Op.getNode();
12848 SmallVector<SDValue, 8> Opnds;
12849 DenseMap<SDValue, unsigned> VecInMap;
12850 SmallVector<SDValue, 8> VecIns;
12851 EVT VT = MVT::Other;
12853 // Recognize a special case where a vector is casted into wide integer to
12855 Opnds.push_back(N->getOperand(0));
12856 Opnds.push_back(N->getOperand(1));
12858 for (unsigned Slot = 0, e = Opnds.size(); Slot < e; ++Slot) {
12859 SmallVectorImpl<SDValue>::const_iterator I = Opnds.begin() + Slot;
12860 // BFS traverse all OR'd operands.
12861 if (I->getOpcode() == ISD::OR) {
12862 Opnds.push_back(I->getOperand(0));
12863 Opnds.push_back(I->getOperand(1));
12864 // Re-evaluate the number of nodes to be traversed.
12865 e += 2; // 2 more nodes (LHS and RHS) are pushed.
12869 // Quit if a non-EXTRACT_VECTOR_ELT
12870 if (I->getOpcode() != ISD::EXTRACT_VECTOR_ELT)
12873 // Quit if without a constant index.
12874 SDValue Idx = I->getOperand(1);
12875 if (!isa<ConstantSDNode>(Idx))
12878 SDValue ExtractedFromVec = I->getOperand(0);
12879 DenseMap<SDValue, unsigned>::iterator M = VecInMap.find(ExtractedFromVec);
12880 if (M == VecInMap.end()) {
12881 VT = ExtractedFromVec.getValueType();
12882 // Quit if not 128/256-bit vector.
12883 if (!VT.is128BitVector() && !VT.is256BitVector())
12885 // Quit if not the same type.
12886 if (VecInMap.begin() != VecInMap.end() &&
12887 VT != VecInMap.begin()->first.getValueType())
12889 M = VecInMap.insert(std::make_pair(ExtractedFromVec, 0)).first;
12890 VecIns.push_back(ExtractedFromVec);
12892 M->second |= 1U << cast<ConstantSDNode>(Idx)->getZExtValue();
12895 assert((VT.is128BitVector() || VT.is256BitVector()) &&
12896 "Not extracted from 128-/256-bit vector.");
12898 unsigned FullMask = (1U << VT.getVectorNumElements()) - 1U;
12900 for (DenseMap<SDValue, unsigned>::const_iterator
12901 I = VecInMap.begin(), E = VecInMap.end(); I != E; ++I) {
12902 // Quit if not all elements are used.
12903 if (I->second != FullMask)
12907 EVT TestVT = VT.is128BitVector() ? MVT::v2i64 : MVT::v4i64;
12909 // Cast all vectors into TestVT for PTEST.
12910 for (unsigned i = 0, e = VecIns.size(); i < e; ++i)
12911 VecIns[i] = DAG.getBitcast(TestVT, VecIns[i]);
12913 // If more than one full vectors are evaluated, OR them first before PTEST.
12914 for (unsigned Slot = 0, e = VecIns.size(); e - Slot > 1; Slot += 2, e += 1) {
12915 // Each iteration will OR 2 nodes and append the result until there is only
12916 // 1 node left, i.e. the final OR'd value of all vectors.
12917 SDValue LHS = VecIns[Slot];
12918 SDValue RHS = VecIns[Slot + 1];
12919 VecIns.push_back(DAG.getNode(ISD::OR, DL, TestVT, LHS, RHS));
12922 return DAG.getNode(X86ISD::PTEST, DL, MVT::i32,
12923 VecIns.back(), VecIns.back());
12926 /// \brief return true if \c Op has a use that doesn't just read flags.
12927 static bool hasNonFlagsUse(SDValue Op) {
12928 for (SDNode::use_iterator UI = Op->use_begin(), UE = Op->use_end(); UI != UE;
12930 SDNode *User = *UI;
12931 unsigned UOpNo = UI.getOperandNo();
12932 if (User->getOpcode() == ISD::TRUNCATE && User->hasOneUse()) {
12933 // Look pass truncate.
12934 UOpNo = User->use_begin().getOperandNo();
12935 User = *User->use_begin();
12938 if (User->getOpcode() != ISD::BRCOND && User->getOpcode() != ISD::SETCC &&
12939 !(User->getOpcode() == ISD::SELECT && UOpNo == 0))
12945 /// Emit nodes that will be selected as "test Op0,Op0", or something
12947 SDValue X86TargetLowering::EmitTest(SDValue Op, unsigned X86CC, SDLoc dl,
12948 SelectionDAG &DAG) const {
12949 if (Op.getValueType() == MVT::i1) {
12950 SDValue ExtOp = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i8, Op);
12951 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, ExtOp,
12952 DAG.getConstant(0, dl, MVT::i8));
12954 // CF and OF aren't always set the way we want. Determine which
12955 // of these we need.
12956 bool NeedCF = false;
12957 bool NeedOF = false;
12960 case X86::COND_A: case X86::COND_AE:
12961 case X86::COND_B: case X86::COND_BE:
12964 case X86::COND_G: case X86::COND_GE:
12965 case X86::COND_L: case X86::COND_LE:
12966 case X86::COND_O: case X86::COND_NO: {
12967 // Check if we really need to set the
12968 // Overflow flag. If NoSignedWrap is present
12969 // that is not actually needed.
12970 switch (Op->getOpcode()) {
12975 const auto *BinNode = cast<BinaryWithFlagsSDNode>(Op.getNode());
12976 if (BinNode->Flags.hasNoSignedWrap())
12986 // See if we can use the EFLAGS value from the operand instead of
12987 // doing a separate TEST. TEST always sets OF and CF to 0, so unless
12988 // we prove that the arithmetic won't overflow, we can't use OF or CF.
12989 if (Op.getResNo() != 0 || NeedOF || NeedCF) {
12990 // Emit a CMP with 0, which is the TEST pattern.
12991 //if (Op.getValueType() == MVT::i1)
12992 // return DAG.getNode(X86ISD::CMP, dl, MVT::i1, Op,
12993 // DAG.getConstant(0, MVT::i1));
12994 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op,
12995 DAG.getConstant(0, dl, Op.getValueType()));
12997 unsigned Opcode = 0;
12998 unsigned NumOperands = 0;
13000 // Truncate operations may prevent the merge of the SETCC instruction
13001 // and the arithmetic instruction before it. Attempt to truncate the operands
13002 // of the arithmetic instruction and use a reduced bit-width instruction.
13003 bool NeedTruncation = false;
13004 SDValue ArithOp = Op;
13005 if (Op->getOpcode() == ISD::TRUNCATE && Op->hasOneUse()) {
13006 SDValue Arith = Op->getOperand(0);
13007 // Both the trunc and the arithmetic op need to have one user each.
13008 if (Arith->hasOneUse())
13009 switch (Arith.getOpcode()) {
13016 NeedTruncation = true;
13022 // NOTICE: In the code below we use ArithOp to hold the arithmetic operation
13023 // which may be the result of a CAST. We use the variable 'Op', which is the
13024 // non-casted variable when we check for possible users.
13025 switch (ArithOp.getOpcode()) {
13027 // Due to an isel shortcoming, be conservative if this add is likely to be
13028 // selected as part of a load-modify-store instruction. When the root node
13029 // in a match is a store, isel doesn't know how to remap non-chain non-flag
13030 // uses of other nodes in the match, such as the ADD in this case. This
13031 // leads to the ADD being left around and reselected, with the result being
13032 // two adds in the output. Alas, even if none our users are stores, that
13033 // doesn't prove we're O.K. Ergo, if we have any parents that aren't
13034 // CopyToReg or SETCC, eschew INC/DEC. A better fix seems to require
13035 // climbing the DAG back to the root, and it doesn't seem to be worth the
13037 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
13038 UE = Op.getNode()->use_end(); UI != UE; ++UI)
13039 if (UI->getOpcode() != ISD::CopyToReg &&
13040 UI->getOpcode() != ISD::SETCC &&
13041 UI->getOpcode() != ISD::STORE)
13044 if (ConstantSDNode *C =
13045 dyn_cast<ConstantSDNode>(ArithOp.getNode()->getOperand(1))) {
13046 // An add of one will be selected as an INC.
13047 if (C->getAPIntValue() == 1 && !Subtarget->slowIncDec()) {
13048 Opcode = X86ISD::INC;
13053 // An add of negative one (subtract of one) will be selected as a DEC.
13054 if (C->getAPIntValue().isAllOnesValue() && !Subtarget->slowIncDec()) {
13055 Opcode = X86ISD::DEC;
13061 // Otherwise use a regular EFLAGS-setting add.
13062 Opcode = X86ISD::ADD;
13067 // If we have a constant logical shift that's only used in a comparison
13068 // against zero turn it into an equivalent AND. This allows turning it into
13069 // a TEST instruction later.
13070 if ((X86CC == X86::COND_E || X86CC == X86::COND_NE) && Op->hasOneUse() &&
13071 isa<ConstantSDNode>(Op->getOperand(1)) && !hasNonFlagsUse(Op)) {
13072 EVT VT = Op.getValueType();
13073 unsigned BitWidth = VT.getSizeInBits();
13074 unsigned ShAmt = Op->getConstantOperandVal(1);
13075 if (ShAmt >= BitWidth) // Avoid undefined shifts.
13077 APInt Mask = ArithOp.getOpcode() == ISD::SRL
13078 ? APInt::getHighBitsSet(BitWidth, BitWidth - ShAmt)
13079 : APInt::getLowBitsSet(BitWidth, BitWidth - ShAmt);
13080 if (!Mask.isSignedIntN(32)) // Avoid large immediates.
13082 SDValue New = DAG.getNode(ISD::AND, dl, VT, Op->getOperand(0),
13083 DAG.getConstant(Mask, dl, VT));
13084 DAG.ReplaceAllUsesWith(Op, New);
13090 // If the primary and result isn't used, don't bother using X86ISD::AND,
13091 // because a TEST instruction will be better.
13092 if (!hasNonFlagsUse(Op))
13098 // Due to the ISEL shortcoming noted above, be conservative if this op is
13099 // likely to be selected as part of a load-modify-store instruction.
13100 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
13101 UE = Op.getNode()->use_end(); UI != UE; ++UI)
13102 if (UI->getOpcode() == ISD::STORE)
13105 // Otherwise use a regular EFLAGS-setting instruction.
13106 switch (ArithOp.getOpcode()) {
13107 default: llvm_unreachable("unexpected operator!");
13108 case ISD::SUB: Opcode = X86ISD::SUB; break;
13109 case ISD::XOR: Opcode = X86ISD::XOR; break;
13110 case ISD::AND: Opcode = X86ISD::AND; break;
13112 if (!NeedTruncation && (X86CC == X86::COND_E || X86CC == X86::COND_NE)) {
13113 SDValue EFLAGS = LowerVectorAllZeroTest(Op, Subtarget, DAG);
13114 if (EFLAGS.getNode())
13117 Opcode = X86ISD::OR;
13131 return SDValue(Op.getNode(), 1);
13137 // If we found that truncation is beneficial, perform the truncation and
13139 if (NeedTruncation) {
13140 EVT VT = Op.getValueType();
13141 SDValue WideVal = Op->getOperand(0);
13142 EVT WideVT = WideVal.getValueType();
13143 unsigned ConvertedOp = 0;
13144 // Use a target machine opcode to prevent further DAGCombine
13145 // optimizations that may separate the arithmetic operations
13146 // from the setcc node.
13147 switch (WideVal.getOpcode()) {
13149 case ISD::ADD: ConvertedOp = X86ISD::ADD; break;
13150 case ISD::SUB: ConvertedOp = X86ISD::SUB; break;
13151 case ISD::AND: ConvertedOp = X86ISD::AND; break;
13152 case ISD::OR: ConvertedOp = X86ISD::OR; break;
13153 case ISD::XOR: ConvertedOp = X86ISD::XOR; break;
13157 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
13158 if (TLI.isOperationLegal(WideVal.getOpcode(), WideVT)) {
13159 SDValue V0 = DAG.getNode(ISD::TRUNCATE, dl, VT, WideVal.getOperand(0));
13160 SDValue V1 = DAG.getNode(ISD::TRUNCATE, dl, VT, WideVal.getOperand(1));
13161 Op = DAG.getNode(ConvertedOp, dl, VT, V0, V1);
13167 // Emit a CMP with 0, which is the TEST pattern.
13168 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op,
13169 DAG.getConstant(0, dl, Op.getValueType()));
13171 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::i32);
13172 SmallVector<SDValue, 4> Ops(Op->op_begin(), Op->op_begin() + NumOperands);
13174 SDValue New = DAG.getNode(Opcode, dl, VTs, Ops);
13175 DAG.ReplaceAllUsesWith(Op, New);
13176 return SDValue(New.getNode(), 1);
13179 /// Emit nodes that will be selected as "cmp Op0,Op1", or something
13181 SDValue X86TargetLowering::EmitCmp(SDValue Op0, SDValue Op1, unsigned X86CC,
13182 SDLoc dl, SelectionDAG &DAG) const {
13183 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op1)) {
13184 if (C->getAPIntValue() == 0)
13185 return EmitTest(Op0, X86CC, dl, DAG);
13187 if (Op0.getValueType() == MVT::i1)
13188 llvm_unreachable("Unexpected comparison operation for MVT::i1 operands");
13191 if ((Op0.getValueType() == MVT::i8 || Op0.getValueType() == MVT::i16 ||
13192 Op0.getValueType() == MVT::i32 || Op0.getValueType() == MVT::i64)) {
13193 // Do the comparison at i32 if it's smaller, besides the Atom case.
13194 // This avoids subregister aliasing issues. Keep the smaller reference
13195 // if we're optimizing for size, however, as that'll allow better folding
13196 // of memory operations.
13197 if (Op0.getValueType() != MVT::i32 && Op0.getValueType() != MVT::i64 &&
13198 !DAG.getMachineFunction().getFunction()->hasFnAttribute(
13199 Attribute::MinSize) &&
13200 !Subtarget->isAtom()) {
13201 unsigned ExtendOp =
13202 isX86CCUnsigned(X86CC) ? ISD::ZERO_EXTEND : ISD::SIGN_EXTEND;
13203 Op0 = DAG.getNode(ExtendOp, dl, MVT::i32, Op0);
13204 Op1 = DAG.getNode(ExtendOp, dl, MVT::i32, Op1);
13206 // Use SUB instead of CMP to enable CSE between SUB and CMP.
13207 SDVTList VTs = DAG.getVTList(Op0.getValueType(), MVT::i32);
13208 SDValue Sub = DAG.getNode(X86ISD::SUB, dl, VTs,
13210 return SDValue(Sub.getNode(), 1);
13212 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op0, Op1);
13215 /// Convert a comparison if required by the subtarget.
13216 SDValue X86TargetLowering::ConvertCmpIfNecessary(SDValue Cmp,
13217 SelectionDAG &DAG) const {
13218 // If the subtarget does not support the FUCOMI instruction, floating-point
13219 // comparisons have to be converted.
13220 if (Subtarget->hasCMov() ||
13221 Cmp.getOpcode() != X86ISD::CMP ||
13222 !Cmp.getOperand(0).getValueType().isFloatingPoint() ||
13223 !Cmp.getOperand(1).getValueType().isFloatingPoint())
13226 // The instruction selector will select an FUCOM instruction instead of
13227 // FUCOMI, which writes the comparison result to FPSW instead of EFLAGS. Hence
13228 // build an SDNode sequence that transfers the result from FPSW into EFLAGS:
13229 // (X86sahf (trunc (srl (X86fp_stsw (trunc (X86cmp ...)), 8))))
13231 SDValue TruncFPSW = DAG.getNode(ISD::TRUNCATE, dl, MVT::i16, Cmp);
13232 SDValue FNStSW = DAG.getNode(X86ISD::FNSTSW16r, dl, MVT::i16, TruncFPSW);
13233 SDValue Srl = DAG.getNode(ISD::SRL, dl, MVT::i16, FNStSW,
13234 DAG.getConstant(8, dl, MVT::i8));
13235 SDValue TruncSrl = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Srl);
13236 return DAG.getNode(X86ISD::SAHF, dl, MVT::i32, TruncSrl);
13239 /// The minimum architected relative accuracy is 2^-12. We need one
13240 /// Newton-Raphson step to have a good float result (24 bits of precision).
13241 SDValue X86TargetLowering::getRsqrtEstimate(SDValue Op,
13242 DAGCombinerInfo &DCI,
13243 unsigned &RefinementSteps,
13244 bool &UseOneConstNR) const {
13245 EVT VT = Op.getValueType();
13246 const char *RecipOp;
13248 // SSE1 has rsqrtss and rsqrtps. AVX adds a 256-bit variant for rsqrtps.
13249 // TODO: Add support for AVX512 (v16f32).
13250 // It is likely not profitable to do this for f64 because a double-precision
13251 // rsqrt estimate with refinement on x86 prior to FMA requires at least 16
13252 // instructions: convert to single, rsqrtss, convert back to double, refine
13253 // (3 steps = at least 13 insts). If an 'rsqrtsd' variant was added to the ISA
13254 // along with FMA, this could be a throughput win.
13255 if (VT == MVT::f32 && Subtarget->hasSSE1())
13257 else if ((VT == MVT::v4f32 && Subtarget->hasSSE1()) ||
13258 (VT == MVT::v8f32 && Subtarget->hasAVX()))
13259 RecipOp = "vec-sqrtf";
13263 TargetRecip Recips = DCI.DAG.getTarget().Options.Reciprocals;
13264 if (!Recips.isEnabled(RecipOp))
13267 RefinementSteps = Recips.getRefinementSteps(RecipOp);
13268 UseOneConstNR = false;
13269 return DCI.DAG.getNode(X86ISD::FRSQRT, SDLoc(Op), VT, Op);
13272 /// The minimum architected relative accuracy is 2^-12. We need one
13273 /// Newton-Raphson step to have a good float result (24 bits of precision).
13274 SDValue X86TargetLowering::getRecipEstimate(SDValue Op,
13275 DAGCombinerInfo &DCI,
13276 unsigned &RefinementSteps) const {
13277 EVT VT = Op.getValueType();
13278 const char *RecipOp;
13280 // SSE1 has rcpss and rcpps. AVX adds a 256-bit variant for rcpps.
13281 // TODO: Add support for AVX512 (v16f32).
13282 // It is likely not profitable to do this for f64 because a double-precision
13283 // reciprocal estimate with refinement on x86 prior to FMA requires
13284 // 15 instructions: convert to single, rcpss, convert back to double, refine
13285 // (3 steps = 12 insts). If an 'rcpsd' variant was added to the ISA
13286 // along with FMA, this could be a throughput win.
13287 if (VT == MVT::f32 && Subtarget->hasSSE1())
13289 else if ((VT == MVT::v4f32 && Subtarget->hasSSE1()) ||
13290 (VT == MVT::v8f32 && Subtarget->hasAVX()))
13291 RecipOp = "vec-divf";
13295 TargetRecip Recips = DCI.DAG.getTarget().Options.Reciprocals;
13296 if (!Recips.isEnabled(RecipOp))
13299 RefinementSteps = Recips.getRefinementSteps(RecipOp);
13300 return DCI.DAG.getNode(X86ISD::FRCP, SDLoc(Op), VT, Op);
13303 /// If we have at least two divisions that use the same divisor, convert to
13304 /// multplication by a reciprocal. This may need to be adjusted for a given
13305 /// CPU if a division's cost is not at least twice the cost of a multiplication.
13306 /// This is because we still need one division to calculate the reciprocal and
13307 /// then we need two multiplies by that reciprocal as replacements for the
13308 /// original divisions.
13309 bool X86TargetLowering::combineRepeatedFPDivisors(unsigned NumUsers) const {
13310 return NumUsers > 1;
13313 static bool isAllOnes(SDValue V) {
13314 ConstantSDNode *C = dyn_cast<ConstantSDNode>(V);
13315 return C && C->isAllOnesValue();
13318 /// LowerToBT - Result of 'and' is compared against zero. Turn it into a BT node
13319 /// if it's possible.
13320 SDValue X86TargetLowering::LowerToBT(SDValue And, ISD::CondCode CC,
13321 SDLoc dl, SelectionDAG &DAG) const {
13322 SDValue Op0 = And.getOperand(0);
13323 SDValue Op1 = And.getOperand(1);
13324 if (Op0.getOpcode() == ISD::TRUNCATE)
13325 Op0 = Op0.getOperand(0);
13326 if (Op1.getOpcode() == ISD::TRUNCATE)
13327 Op1 = Op1.getOperand(0);
13330 if (Op1.getOpcode() == ISD::SHL)
13331 std::swap(Op0, Op1);
13332 if (Op0.getOpcode() == ISD::SHL) {
13333 if (ConstantSDNode *And00C = dyn_cast<ConstantSDNode>(Op0.getOperand(0)))
13334 if (And00C->getZExtValue() == 1) {
13335 // If we looked past a truncate, check that it's only truncating away
13337 unsigned BitWidth = Op0.getValueSizeInBits();
13338 unsigned AndBitWidth = And.getValueSizeInBits();
13339 if (BitWidth > AndBitWidth) {
13341 DAG.computeKnownBits(Op0, Zeros, Ones);
13342 if (Zeros.countLeadingOnes() < BitWidth - AndBitWidth)
13346 RHS = Op0.getOperand(1);
13348 } else if (Op1.getOpcode() == ISD::Constant) {
13349 ConstantSDNode *AndRHS = cast<ConstantSDNode>(Op1);
13350 uint64_t AndRHSVal = AndRHS->getZExtValue();
13351 SDValue AndLHS = Op0;
13353 if (AndRHSVal == 1 && AndLHS.getOpcode() == ISD::SRL) {
13354 LHS = AndLHS.getOperand(0);
13355 RHS = AndLHS.getOperand(1);
13358 // Use BT if the immediate can't be encoded in a TEST instruction.
13359 if (!isUInt<32>(AndRHSVal) && isPowerOf2_64(AndRHSVal)) {
13361 RHS = DAG.getConstant(Log2_64_Ceil(AndRHSVal), dl, LHS.getValueType());
13365 if (LHS.getNode()) {
13366 // If LHS is i8, promote it to i32 with any_extend. There is no i8 BT
13367 // instruction. Since the shift amount is in-range-or-undefined, we know
13368 // that doing a bittest on the i32 value is ok. We extend to i32 because
13369 // the encoding for the i16 version is larger than the i32 version.
13370 // Also promote i16 to i32 for performance / code size reason.
13371 if (LHS.getValueType() == MVT::i8 ||
13372 LHS.getValueType() == MVT::i16)
13373 LHS = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, LHS);
13375 // If the operand types disagree, extend the shift amount to match. Since
13376 // BT ignores high bits (like shifts) we can use anyextend.
13377 if (LHS.getValueType() != RHS.getValueType())
13378 RHS = DAG.getNode(ISD::ANY_EXTEND, dl, LHS.getValueType(), RHS);
13380 SDValue BT = DAG.getNode(X86ISD::BT, dl, MVT::i32, LHS, RHS);
13381 X86::CondCode Cond = CC == ISD::SETEQ ? X86::COND_AE : X86::COND_B;
13382 return DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
13383 DAG.getConstant(Cond, dl, MVT::i8), BT);
13389 /// \brief - Turns an ISD::CondCode into a value suitable for SSE floating point
13391 static int translateX86FSETCC(ISD::CondCode SetCCOpcode, SDValue &Op0,
13396 // SSE Condition code mapping:
13405 switch (SetCCOpcode) {
13406 default: llvm_unreachable("Unexpected SETCC condition");
13408 case ISD::SETEQ: SSECC = 0; break;
13410 case ISD::SETGT: Swap = true; // Fallthrough
13412 case ISD::SETOLT: SSECC = 1; break;
13414 case ISD::SETGE: Swap = true; // Fallthrough
13416 case ISD::SETOLE: SSECC = 2; break;
13417 case ISD::SETUO: SSECC = 3; break;
13419 case ISD::SETNE: SSECC = 4; break;
13420 case ISD::SETULE: Swap = true; // Fallthrough
13421 case ISD::SETUGE: SSECC = 5; break;
13422 case ISD::SETULT: Swap = true; // Fallthrough
13423 case ISD::SETUGT: SSECC = 6; break;
13424 case ISD::SETO: SSECC = 7; break;
13426 case ISD::SETONE: SSECC = 8; break;
13429 std::swap(Op0, Op1);
13434 // Lower256IntVSETCC - Break a VSETCC 256-bit integer VSETCC into two new 128
13435 // ones, and then concatenate the result back.
13436 static SDValue Lower256IntVSETCC(SDValue Op, SelectionDAG &DAG) {
13437 MVT VT = Op.getSimpleValueType();
13439 assert(VT.is256BitVector() && Op.getOpcode() == ISD::SETCC &&
13440 "Unsupported value type for operation");
13442 unsigned NumElems = VT.getVectorNumElements();
13444 SDValue CC = Op.getOperand(2);
13446 // Extract the LHS vectors
13447 SDValue LHS = Op.getOperand(0);
13448 SDValue LHS1 = Extract128BitVector(LHS, 0, DAG, dl);
13449 SDValue LHS2 = Extract128BitVector(LHS, NumElems/2, DAG, dl);
13451 // Extract the RHS vectors
13452 SDValue RHS = Op.getOperand(1);
13453 SDValue RHS1 = Extract128BitVector(RHS, 0, DAG, dl);
13454 SDValue RHS2 = Extract128BitVector(RHS, NumElems/2, DAG, dl);
13456 // Issue the operation on the smaller types and concatenate the result back
13457 MVT EltVT = VT.getVectorElementType();
13458 MVT NewVT = MVT::getVectorVT(EltVT, NumElems/2);
13459 return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT,
13460 DAG.getNode(Op.getOpcode(), dl, NewVT, LHS1, RHS1, CC),
13461 DAG.getNode(Op.getOpcode(), dl, NewVT, LHS2, RHS2, CC));
13464 static SDValue LowerBoolVSETCC_AVX512(SDValue Op, SelectionDAG &DAG) {
13465 SDValue Op0 = Op.getOperand(0);
13466 SDValue Op1 = Op.getOperand(1);
13467 SDValue CC = Op.getOperand(2);
13468 MVT VT = Op.getSimpleValueType();
13471 assert(Op0.getValueType().getVectorElementType() == MVT::i1 &&
13472 "Unexpected type for boolean compare operation");
13473 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
13474 SDValue NotOp0 = DAG.getNode(ISD::XOR, dl, VT, Op0,
13475 DAG.getConstant(-1, dl, VT));
13476 SDValue NotOp1 = DAG.getNode(ISD::XOR, dl, VT, Op1,
13477 DAG.getConstant(-1, dl, VT));
13478 switch (SetCCOpcode) {
13479 default: llvm_unreachable("Unexpected SETCC condition");
13481 // (x == y) -> ~(x ^ y)
13482 return DAG.getNode(ISD::XOR, dl, VT,
13483 DAG.getNode(ISD::XOR, dl, VT, Op0, Op1),
13484 DAG.getConstant(-1, dl, VT));
13486 // (x != y) -> (x ^ y)
13487 return DAG.getNode(ISD::XOR, dl, VT, Op0, Op1);
13490 // (x > y) -> (x & ~y)
13491 return DAG.getNode(ISD::AND, dl, VT, Op0, NotOp1);
13494 // (x < y) -> (~x & y)
13495 return DAG.getNode(ISD::AND, dl, VT, NotOp0, Op1);
13498 // (x <= y) -> (~x | y)
13499 return DAG.getNode(ISD::OR, dl, VT, NotOp0, Op1);
13502 // (x >=y) -> (x | ~y)
13503 return DAG.getNode(ISD::OR, dl, VT, Op0, NotOp1);
13507 static SDValue LowerIntVSETCC_AVX512(SDValue Op, SelectionDAG &DAG,
13508 const X86Subtarget *Subtarget) {
13509 SDValue Op0 = Op.getOperand(0);
13510 SDValue Op1 = Op.getOperand(1);
13511 SDValue CC = Op.getOperand(2);
13512 MVT VT = Op.getSimpleValueType();
13515 assert(Op0.getValueType().getVectorElementType().getSizeInBits() >= 8 &&
13516 Op.getValueType().getScalarType() == MVT::i1 &&
13517 "Cannot set masked compare for this operation");
13519 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
13521 bool Unsigned = false;
13524 switch (SetCCOpcode) {
13525 default: llvm_unreachable("Unexpected SETCC condition");
13526 case ISD::SETNE: SSECC = 4; break;
13527 case ISD::SETEQ: Opc = X86ISD::PCMPEQM; break;
13528 case ISD::SETUGT: SSECC = 6; Unsigned = true; break;
13529 case ISD::SETLT: Swap = true; //fall-through
13530 case ISD::SETGT: Opc = X86ISD::PCMPGTM; break;
13531 case ISD::SETULT: SSECC = 1; Unsigned = true; break;
13532 case ISD::SETUGE: SSECC = 5; Unsigned = true; break; //NLT
13533 case ISD::SETGE: Swap = true; SSECC = 2; break; // LE + swap
13534 case ISD::SETULE: Unsigned = true; //fall-through
13535 case ISD::SETLE: SSECC = 2; break;
13539 std::swap(Op0, Op1);
13541 return DAG.getNode(Opc, dl, VT, Op0, Op1);
13542 Opc = Unsigned ? X86ISD::CMPMU: X86ISD::CMPM;
13543 return DAG.getNode(Opc, dl, VT, Op0, Op1,
13544 DAG.getConstant(SSECC, dl, MVT::i8));
13547 /// \brief Try to turn a VSETULT into a VSETULE by modifying its second
13548 /// operand \p Op1. If non-trivial (for example because it's not constant)
13549 /// return an empty value.
13550 static SDValue ChangeVSETULTtoVSETULE(SDLoc dl, SDValue Op1, SelectionDAG &DAG)
13552 BuildVectorSDNode *BV = dyn_cast<BuildVectorSDNode>(Op1.getNode());
13556 MVT VT = Op1.getSimpleValueType();
13557 MVT EVT = VT.getVectorElementType();
13558 unsigned n = VT.getVectorNumElements();
13559 SmallVector<SDValue, 8> ULTOp1;
13561 for (unsigned i = 0; i < n; ++i) {
13562 ConstantSDNode *Elt = dyn_cast<ConstantSDNode>(BV->getOperand(i));
13563 if (!Elt || Elt->isOpaque() || Elt->getValueType(0) != EVT)
13566 // Avoid underflow.
13567 APInt Val = Elt->getAPIntValue();
13571 ULTOp1.push_back(DAG.getConstant(Val - 1, dl, EVT));
13574 return DAG.getNode(ISD::BUILD_VECTOR, dl, VT, ULTOp1);
13577 static SDValue LowerVSETCC(SDValue Op, const X86Subtarget *Subtarget,
13578 SelectionDAG &DAG) {
13579 SDValue Op0 = Op.getOperand(0);
13580 SDValue Op1 = Op.getOperand(1);
13581 SDValue CC = Op.getOperand(2);
13582 MVT VT = Op.getSimpleValueType();
13583 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
13584 bool isFP = Op.getOperand(1).getSimpleValueType().isFloatingPoint();
13589 MVT EltVT = Op0.getSimpleValueType().getVectorElementType();
13590 assert(EltVT == MVT::f32 || EltVT == MVT::f64);
13593 unsigned SSECC = translateX86FSETCC(SetCCOpcode, Op0, Op1);
13594 unsigned Opc = X86ISD::CMPP;
13595 if (Subtarget->hasAVX512() && VT.getVectorElementType() == MVT::i1) {
13596 assert(VT.getVectorNumElements() <= 16);
13597 Opc = X86ISD::CMPM;
13599 // In the two special cases we can't handle, emit two comparisons.
13602 unsigned CombineOpc;
13603 if (SetCCOpcode == ISD::SETUEQ) {
13604 CC0 = 3; CC1 = 0; CombineOpc = ISD::OR;
13606 assert(SetCCOpcode == ISD::SETONE);
13607 CC0 = 7; CC1 = 4; CombineOpc = ISD::AND;
13610 SDValue Cmp0 = DAG.getNode(Opc, dl, VT, Op0, Op1,
13611 DAG.getConstant(CC0, dl, MVT::i8));
13612 SDValue Cmp1 = DAG.getNode(Opc, dl, VT, Op0, Op1,
13613 DAG.getConstant(CC1, dl, MVT::i8));
13614 return DAG.getNode(CombineOpc, dl, VT, Cmp0, Cmp1);
13616 // Handle all other FP comparisons here.
13617 return DAG.getNode(Opc, dl, VT, Op0, Op1,
13618 DAG.getConstant(SSECC, dl, MVT::i8));
13621 // Break 256-bit integer vector compare into smaller ones.
13622 if (VT.is256BitVector() && !Subtarget->hasInt256())
13623 return Lower256IntVSETCC(Op, DAG);
13625 EVT OpVT = Op1.getValueType();
13626 if (OpVT.getVectorElementType() == MVT::i1)
13627 return LowerBoolVSETCC_AVX512(Op, DAG);
13629 bool MaskResult = (VT.getVectorElementType() == MVT::i1);
13630 if (Subtarget->hasAVX512()) {
13631 if (Op1.getValueType().is512BitVector() ||
13632 (Subtarget->hasBWI() && Subtarget->hasVLX()) ||
13633 (MaskResult && OpVT.getVectorElementType().getSizeInBits() >= 32))
13634 return LowerIntVSETCC_AVX512(Op, DAG, Subtarget);
13636 // In AVX-512 architecture setcc returns mask with i1 elements,
13637 // But there is no compare instruction for i8 and i16 elements in KNL.
13638 // We are not talking about 512-bit operands in this case, these
13639 // types are illegal.
13641 (OpVT.getVectorElementType().getSizeInBits() < 32 &&
13642 OpVT.getVectorElementType().getSizeInBits() >= 8))
13643 return DAG.getNode(ISD::TRUNCATE, dl, VT,
13644 DAG.getNode(ISD::SETCC, dl, OpVT, Op0, Op1, CC));
13647 // We are handling one of the integer comparisons here. Since SSE only has
13648 // GT and EQ comparisons for integer, swapping operands and multiple
13649 // operations may be required for some comparisons.
13651 bool Swap = false, Invert = false, FlipSigns = false, MinMax = false;
13652 bool Subus = false;
13654 switch (SetCCOpcode) {
13655 default: llvm_unreachable("Unexpected SETCC condition");
13656 case ISD::SETNE: Invert = true;
13657 case ISD::SETEQ: Opc = X86ISD::PCMPEQ; break;
13658 case ISD::SETLT: Swap = true;
13659 case ISD::SETGT: Opc = X86ISD::PCMPGT; break;
13660 case ISD::SETGE: Swap = true;
13661 case ISD::SETLE: Opc = X86ISD::PCMPGT;
13662 Invert = true; break;
13663 case ISD::SETULT: Swap = true;
13664 case ISD::SETUGT: Opc = X86ISD::PCMPGT;
13665 FlipSigns = true; break;
13666 case ISD::SETUGE: Swap = true;
13667 case ISD::SETULE: Opc = X86ISD::PCMPGT;
13668 FlipSigns = true; Invert = true; break;
13671 // Special case: Use min/max operations for SETULE/SETUGE
13672 MVT VET = VT.getVectorElementType();
13674 (Subtarget->hasSSE41() && (VET >= MVT::i8 && VET <= MVT::i32))
13675 || (Subtarget->hasSSE2() && (VET == MVT::i8));
13678 switch (SetCCOpcode) {
13680 case ISD::SETULE: Opc = ISD::UMIN; MinMax = true; break;
13681 case ISD::SETUGE: Opc = ISD::UMAX; MinMax = true; break;
13684 if (MinMax) { Swap = false; Invert = false; FlipSigns = false; }
13687 bool hasSubus = Subtarget->hasSSE2() && (VET == MVT::i8 || VET == MVT::i16);
13688 if (!MinMax && hasSubus) {
13689 // As another special case, use PSUBUS[BW] when it's profitable. E.g. for
13691 // t = psubus Op0, Op1
13692 // pcmpeq t, <0..0>
13693 switch (SetCCOpcode) {
13695 case ISD::SETULT: {
13696 // If the comparison is against a constant we can turn this into a
13697 // setule. With psubus, setule does not require a swap. This is
13698 // beneficial because the constant in the register is no longer
13699 // destructed as the destination so it can be hoisted out of a loop.
13700 // Only do this pre-AVX since vpcmp* is no longer destructive.
13701 if (Subtarget->hasAVX())
13703 SDValue ULEOp1 = ChangeVSETULTtoVSETULE(dl, Op1, DAG);
13704 if (ULEOp1.getNode()) {
13706 Subus = true; Invert = false; Swap = false;
13710 // Psubus is better than flip-sign because it requires no inversion.
13711 case ISD::SETUGE: Subus = true; Invert = false; Swap = true; break;
13712 case ISD::SETULE: Subus = true; Invert = false; Swap = false; break;
13716 Opc = X86ISD::SUBUS;
13722 std::swap(Op0, Op1);
13724 // Check that the operation in question is available (most are plain SSE2,
13725 // but PCMPGTQ and PCMPEQQ have different requirements).
13726 if (VT == MVT::v2i64) {
13727 if (Opc == X86ISD::PCMPGT && !Subtarget->hasSSE42()) {
13728 assert(Subtarget->hasSSE2() && "Don't know how to lower!");
13730 // First cast everything to the right type.
13731 Op0 = DAG.getBitcast(MVT::v4i32, Op0);
13732 Op1 = DAG.getBitcast(MVT::v4i32, Op1);
13734 // Since SSE has no unsigned integer comparisons, we need to flip the sign
13735 // bits of the inputs before performing those operations. The lower
13736 // compare is always unsigned.
13739 SB = DAG.getConstant(0x80000000U, dl, MVT::v4i32);
13741 SDValue Sign = DAG.getConstant(0x80000000U, dl, MVT::i32);
13742 SDValue Zero = DAG.getConstant(0x00000000U, dl, MVT::i32);
13743 SB = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32,
13744 Sign, Zero, Sign, Zero);
13746 Op0 = DAG.getNode(ISD::XOR, dl, MVT::v4i32, Op0, SB);
13747 Op1 = DAG.getNode(ISD::XOR, dl, MVT::v4i32, Op1, SB);
13749 // Emulate PCMPGTQ with (hi1 > hi2) | ((hi1 == hi2) & (lo1 > lo2))
13750 SDValue GT = DAG.getNode(X86ISD::PCMPGT, dl, MVT::v4i32, Op0, Op1);
13751 SDValue EQ = DAG.getNode(X86ISD::PCMPEQ, dl, MVT::v4i32, Op0, Op1);
13753 // Create masks for only the low parts/high parts of the 64 bit integers.
13754 static const int MaskHi[] = { 1, 1, 3, 3 };
13755 static const int MaskLo[] = { 0, 0, 2, 2 };
13756 SDValue EQHi = DAG.getVectorShuffle(MVT::v4i32, dl, EQ, EQ, MaskHi);
13757 SDValue GTLo = DAG.getVectorShuffle(MVT::v4i32, dl, GT, GT, MaskLo);
13758 SDValue GTHi = DAG.getVectorShuffle(MVT::v4i32, dl, GT, GT, MaskHi);
13760 SDValue Result = DAG.getNode(ISD::AND, dl, MVT::v4i32, EQHi, GTLo);
13761 Result = DAG.getNode(ISD::OR, dl, MVT::v4i32, Result, GTHi);
13764 Result = DAG.getNOT(dl, Result, MVT::v4i32);
13766 return DAG.getBitcast(VT, Result);
13769 if (Opc == X86ISD::PCMPEQ && !Subtarget->hasSSE41()) {
13770 // If pcmpeqq is missing but pcmpeqd is available synthesize pcmpeqq with
13771 // pcmpeqd + pshufd + pand.
13772 assert(Subtarget->hasSSE2() && !FlipSigns && "Don't know how to lower!");
13774 // First cast everything to the right type.
13775 Op0 = DAG.getBitcast(MVT::v4i32, Op0);
13776 Op1 = DAG.getBitcast(MVT::v4i32, Op1);
13779 SDValue Result = DAG.getNode(Opc, dl, MVT::v4i32, Op0, Op1);
13781 // Make sure the lower and upper halves are both all-ones.
13782 static const int Mask[] = { 1, 0, 3, 2 };
13783 SDValue Shuf = DAG.getVectorShuffle(MVT::v4i32, dl, Result, Result, Mask);
13784 Result = DAG.getNode(ISD::AND, dl, MVT::v4i32, Result, Shuf);
13787 Result = DAG.getNOT(dl, Result, MVT::v4i32);
13789 return DAG.getBitcast(VT, Result);
13793 // Since SSE has no unsigned integer comparisons, we need to flip the sign
13794 // bits of the inputs before performing those operations.
13796 EVT EltVT = VT.getVectorElementType();
13797 SDValue SB = DAG.getConstant(APInt::getSignBit(EltVT.getSizeInBits()), dl,
13799 Op0 = DAG.getNode(ISD::XOR, dl, VT, Op0, SB);
13800 Op1 = DAG.getNode(ISD::XOR, dl, VT, Op1, SB);
13803 SDValue Result = DAG.getNode(Opc, dl, VT, Op0, Op1);
13805 // If the logical-not of the result is required, perform that now.
13807 Result = DAG.getNOT(dl, Result, VT);
13810 Result = DAG.getNode(X86ISD::PCMPEQ, dl, VT, Op0, Result);
13813 Result = DAG.getNode(X86ISD::PCMPEQ, dl, VT, Result,
13814 getZeroVector(VT, Subtarget, DAG, dl));
13819 SDValue X86TargetLowering::LowerSETCC(SDValue Op, SelectionDAG &DAG) const {
13821 MVT VT = Op.getSimpleValueType();
13823 if (VT.isVector()) return LowerVSETCC(Op, Subtarget, DAG);
13825 assert(((!Subtarget->hasAVX512() && VT == MVT::i8) || (VT == MVT::i1))
13826 && "SetCC type must be 8-bit or 1-bit integer");
13827 SDValue Op0 = Op.getOperand(0);
13828 SDValue Op1 = Op.getOperand(1);
13830 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
13832 // Optimize to BT if possible.
13833 // Lower (X & (1 << N)) == 0 to BT(X, N).
13834 // Lower ((X >>u N) & 1) != 0 to BT(X, N).
13835 // Lower ((X >>s N) & 1) != 0 to BT(X, N).
13836 if (Op0.getOpcode() == ISD::AND && Op0.hasOneUse() &&
13837 Op1.getOpcode() == ISD::Constant &&
13838 cast<ConstantSDNode>(Op1)->isNullValue() &&
13839 (CC == ISD::SETEQ || CC == ISD::SETNE)) {
13840 SDValue NewSetCC = LowerToBT(Op0, CC, dl, DAG);
13841 if (NewSetCC.getNode()) {
13843 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i1, NewSetCC);
13848 // Look for X == 0, X == 1, X != 0, or X != 1. We can simplify some forms of
13850 if (Op1.getOpcode() == ISD::Constant &&
13851 (cast<ConstantSDNode>(Op1)->getZExtValue() == 1 ||
13852 cast<ConstantSDNode>(Op1)->isNullValue()) &&
13853 (CC == ISD::SETEQ || CC == ISD::SETNE)) {
13855 // If the input is a setcc, then reuse the input setcc or use a new one with
13856 // the inverted condition.
13857 if (Op0.getOpcode() == X86ISD::SETCC) {
13858 X86::CondCode CCode = (X86::CondCode)Op0.getConstantOperandVal(0);
13859 bool Invert = (CC == ISD::SETNE) ^
13860 cast<ConstantSDNode>(Op1)->isNullValue();
13864 CCode = X86::GetOppositeBranchCondition(CCode);
13865 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
13866 DAG.getConstant(CCode, dl, MVT::i8),
13867 Op0.getOperand(1));
13869 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i1, SetCC);
13873 if ((Op0.getValueType() == MVT::i1) && (Op1.getOpcode() == ISD::Constant) &&
13874 (cast<ConstantSDNode>(Op1)->getZExtValue() == 1) &&
13875 (CC == ISD::SETEQ || CC == ISD::SETNE)) {
13877 ISD::CondCode NewCC = ISD::getSetCCInverse(CC, true);
13878 return DAG.getSetCC(dl, VT, Op0, DAG.getConstant(0, dl, MVT::i1), NewCC);
13881 bool isFP = Op1.getSimpleValueType().isFloatingPoint();
13882 unsigned X86CC = TranslateX86CC(CC, dl, isFP, Op0, Op1, DAG);
13883 if (X86CC == X86::COND_INVALID)
13886 SDValue EFLAGS = EmitCmp(Op0, Op1, X86CC, dl, DAG);
13887 EFLAGS = ConvertCmpIfNecessary(EFLAGS, DAG);
13888 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
13889 DAG.getConstant(X86CC, dl, MVT::i8), EFLAGS);
13891 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i1, SetCC);
13895 // isX86LogicalCmp - Return true if opcode is a X86 logical comparison.
13896 static bool isX86LogicalCmp(SDValue Op) {
13897 unsigned Opc = Op.getNode()->getOpcode();
13898 if (Opc == X86ISD::CMP || Opc == X86ISD::COMI || Opc == X86ISD::UCOMI ||
13899 Opc == X86ISD::SAHF)
13901 if (Op.getResNo() == 1 &&
13902 (Opc == X86ISD::ADD ||
13903 Opc == X86ISD::SUB ||
13904 Opc == X86ISD::ADC ||
13905 Opc == X86ISD::SBB ||
13906 Opc == X86ISD::SMUL ||
13907 Opc == X86ISD::UMUL ||
13908 Opc == X86ISD::INC ||
13909 Opc == X86ISD::DEC ||
13910 Opc == X86ISD::OR ||
13911 Opc == X86ISD::XOR ||
13912 Opc == X86ISD::AND))
13915 if (Op.getResNo() == 2 && Opc == X86ISD::UMUL)
13921 static bool isTruncWithZeroHighBitsInput(SDValue V, SelectionDAG &DAG) {
13922 if (V.getOpcode() != ISD::TRUNCATE)
13925 SDValue VOp0 = V.getOperand(0);
13926 unsigned InBits = VOp0.getValueSizeInBits();
13927 unsigned Bits = V.getValueSizeInBits();
13928 return DAG.MaskedValueIsZero(VOp0, APInt::getHighBitsSet(InBits,InBits-Bits));
13931 SDValue X86TargetLowering::LowerSELECT(SDValue Op, SelectionDAG &DAG) const {
13932 bool addTest = true;
13933 SDValue Cond = Op.getOperand(0);
13934 SDValue Op1 = Op.getOperand(1);
13935 SDValue Op2 = Op.getOperand(2);
13937 EVT VT = Op1.getValueType();
13940 // Lower FP selects into a CMP/AND/ANDN/OR sequence when the necessary SSE ops
13941 // are available or VBLENDV if AVX is available.
13942 // Otherwise FP cmovs get lowered into a less efficient branch sequence later.
13943 if (Cond.getOpcode() == ISD::SETCC &&
13944 ((Subtarget->hasSSE2() && (VT == MVT::f32 || VT == MVT::f64)) ||
13945 (Subtarget->hasSSE1() && VT == MVT::f32)) &&
13946 VT == Cond.getOperand(0).getValueType() && Cond->hasOneUse()) {
13947 SDValue CondOp0 = Cond.getOperand(0), CondOp1 = Cond.getOperand(1);
13948 int SSECC = translateX86FSETCC(
13949 cast<CondCodeSDNode>(Cond.getOperand(2))->get(), CondOp0, CondOp1);
13952 if (Subtarget->hasAVX512()) {
13953 SDValue Cmp = DAG.getNode(X86ISD::FSETCC, DL, MVT::i1, CondOp0, CondOp1,
13954 DAG.getConstant(SSECC, DL, MVT::i8));
13955 return DAG.getNode(X86ISD::SELECT, DL, VT, Cmp, Op1, Op2);
13958 SDValue Cmp = DAG.getNode(X86ISD::FSETCC, DL, VT, CondOp0, CondOp1,
13959 DAG.getConstant(SSECC, DL, MVT::i8));
13961 // If we have AVX, we can use a variable vector select (VBLENDV) instead
13962 // of 3 logic instructions for size savings and potentially speed.
13963 // Unfortunately, there is no scalar form of VBLENDV.
13965 // If either operand is a constant, don't try this. We can expect to
13966 // optimize away at least one of the logic instructions later in that
13967 // case, so that sequence would be faster than a variable blend.
13969 // BLENDV was introduced with SSE 4.1, but the 2 register form implicitly
13970 // uses XMM0 as the selection register. That may need just as many
13971 // instructions as the AND/ANDN/OR sequence due to register moves, so
13974 if (Subtarget->hasAVX() &&
13975 !isa<ConstantFPSDNode>(Op1) && !isa<ConstantFPSDNode>(Op2)) {
13977 // Convert to vectors, do a VSELECT, and convert back to scalar.
13978 // All of the conversions should be optimized away.
13980 EVT VecVT = VT == MVT::f32 ? MVT::v4f32 : MVT::v2f64;
13981 SDValue VOp1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, VecVT, Op1);
13982 SDValue VOp2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, VecVT, Op2);
13983 SDValue VCmp = DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, VecVT, Cmp);
13985 EVT VCmpVT = VT == MVT::f32 ? MVT::v4i32 : MVT::v2i64;
13986 VCmp = DAG.getBitcast(VCmpVT, VCmp);
13988 SDValue VSel = DAG.getNode(ISD::VSELECT, DL, VecVT, VCmp, VOp1, VOp2);
13990 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, VT,
13991 VSel, DAG.getIntPtrConstant(0, DL));
13993 SDValue AndN = DAG.getNode(X86ISD::FANDN, DL, VT, Cmp, Op2);
13994 SDValue And = DAG.getNode(X86ISD::FAND, DL, VT, Cmp, Op1);
13995 return DAG.getNode(X86ISD::FOR, DL, VT, AndN, And);
13999 if (VT.isVector() && VT.getScalarType() == MVT::i1) {
14001 if (ISD::isBuildVectorOfConstantSDNodes(Op1.getNode()))
14002 Op1Scalar = ConvertI1VectorToInteger(Op1, DAG);
14003 else if (Op1.getOpcode() == ISD::BITCAST && Op1.getOperand(0))
14004 Op1Scalar = Op1.getOperand(0);
14006 if (ISD::isBuildVectorOfConstantSDNodes(Op2.getNode()))
14007 Op2Scalar = ConvertI1VectorToInteger(Op2, DAG);
14008 else if (Op2.getOpcode() == ISD::BITCAST && Op2.getOperand(0))
14009 Op2Scalar = Op2.getOperand(0);
14010 if (Op1Scalar.getNode() && Op2Scalar.getNode()) {
14011 SDValue newSelect = DAG.getNode(ISD::SELECT, DL,
14012 Op1Scalar.getValueType(),
14013 Cond, Op1Scalar, Op2Scalar);
14014 if (newSelect.getValueSizeInBits() == VT.getSizeInBits())
14015 return DAG.getBitcast(VT, newSelect);
14016 SDValue ExtVec = DAG.getBitcast(MVT::v8i1, newSelect);
14017 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, VT, ExtVec,
14018 DAG.getIntPtrConstant(0, DL));
14022 if (VT == MVT::v4i1 || VT == MVT::v2i1) {
14023 SDValue zeroConst = DAG.getIntPtrConstant(0, DL);
14024 Op1 = DAG.getNode(ISD::INSERT_SUBVECTOR, DL, MVT::v8i1,
14025 DAG.getUNDEF(MVT::v8i1), Op1, zeroConst);
14026 Op2 = DAG.getNode(ISD::INSERT_SUBVECTOR, DL, MVT::v8i1,
14027 DAG.getUNDEF(MVT::v8i1), Op2, zeroConst);
14028 SDValue newSelect = DAG.getNode(ISD::SELECT, DL, MVT::v8i1,
14030 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, VT, newSelect, zeroConst);
14033 if (Cond.getOpcode() == ISD::SETCC) {
14034 SDValue NewCond = LowerSETCC(Cond, DAG);
14035 if (NewCond.getNode())
14039 // (select (x == 0), -1, y) -> (sign_bit (x - 1)) | y
14040 // (select (x == 0), y, -1) -> ~(sign_bit (x - 1)) | y
14041 // (select (x != 0), y, -1) -> (sign_bit (x - 1)) | y
14042 // (select (x != 0), -1, y) -> ~(sign_bit (x - 1)) | y
14043 if (Cond.getOpcode() == X86ISD::SETCC &&
14044 Cond.getOperand(1).getOpcode() == X86ISD::CMP &&
14045 isZero(Cond.getOperand(1).getOperand(1))) {
14046 SDValue Cmp = Cond.getOperand(1);
14048 unsigned CondCode =cast<ConstantSDNode>(Cond.getOperand(0))->getZExtValue();
14050 if ((isAllOnes(Op1) || isAllOnes(Op2)) &&
14051 (CondCode == X86::COND_E || CondCode == X86::COND_NE)) {
14052 SDValue Y = isAllOnes(Op2) ? Op1 : Op2;
14054 SDValue CmpOp0 = Cmp.getOperand(0);
14055 // Apply further optimizations for special cases
14056 // (select (x != 0), -1, 0) -> neg & sbb
14057 // (select (x == 0), 0, -1) -> neg & sbb
14058 if (ConstantSDNode *YC = dyn_cast<ConstantSDNode>(Y))
14059 if (YC->isNullValue() &&
14060 (isAllOnes(Op1) == (CondCode == X86::COND_NE))) {
14061 SDVTList VTs = DAG.getVTList(CmpOp0.getValueType(), MVT::i32);
14062 SDValue Neg = DAG.getNode(X86ISD::SUB, DL, VTs,
14063 DAG.getConstant(0, DL,
14064 CmpOp0.getValueType()),
14066 SDValue Res = DAG.getNode(X86ISD::SETCC_CARRY, DL, Op.getValueType(),
14067 DAG.getConstant(X86::COND_B, DL, MVT::i8),
14068 SDValue(Neg.getNode(), 1));
14072 Cmp = DAG.getNode(X86ISD::CMP, DL, MVT::i32,
14073 CmpOp0, DAG.getConstant(1, DL, CmpOp0.getValueType()));
14074 Cmp = ConvertCmpIfNecessary(Cmp, DAG);
14076 SDValue Res = // Res = 0 or -1.
14077 DAG.getNode(X86ISD::SETCC_CARRY, DL, Op.getValueType(),
14078 DAG.getConstant(X86::COND_B, DL, MVT::i8), Cmp);
14080 if (isAllOnes(Op1) != (CondCode == X86::COND_E))
14081 Res = DAG.getNOT(DL, Res, Res.getValueType());
14083 ConstantSDNode *N2C = dyn_cast<ConstantSDNode>(Op2);
14084 if (!N2C || !N2C->isNullValue())
14085 Res = DAG.getNode(ISD::OR, DL, Res.getValueType(), Res, Y);
14090 // Look past (and (setcc_carry (cmp ...)), 1).
14091 if (Cond.getOpcode() == ISD::AND &&
14092 Cond.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY) {
14093 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Cond.getOperand(1));
14094 if (C && C->getAPIntValue() == 1)
14095 Cond = Cond.getOperand(0);
14098 // If condition flag is set by a X86ISD::CMP, then use it as the condition
14099 // setting operand in place of the X86ISD::SETCC.
14100 unsigned CondOpcode = Cond.getOpcode();
14101 if (CondOpcode == X86ISD::SETCC ||
14102 CondOpcode == X86ISD::SETCC_CARRY) {
14103 CC = Cond.getOperand(0);
14105 SDValue Cmp = Cond.getOperand(1);
14106 unsigned Opc = Cmp.getOpcode();
14107 MVT VT = Op.getSimpleValueType();
14109 bool IllegalFPCMov = false;
14110 if (VT.isFloatingPoint() && !VT.isVector() &&
14111 !isScalarFPTypeInSSEReg(VT)) // FPStack?
14112 IllegalFPCMov = !hasFPCMov(cast<ConstantSDNode>(CC)->getSExtValue());
14114 if ((isX86LogicalCmp(Cmp) && !IllegalFPCMov) ||
14115 Opc == X86ISD::BT) { // FIXME
14119 } else if (CondOpcode == ISD::USUBO || CondOpcode == ISD::SSUBO ||
14120 CondOpcode == ISD::UADDO || CondOpcode == ISD::SADDO ||
14121 ((CondOpcode == ISD::UMULO || CondOpcode == ISD::SMULO) &&
14122 Cond.getOperand(0).getValueType() != MVT::i8)) {
14123 SDValue LHS = Cond.getOperand(0);
14124 SDValue RHS = Cond.getOperand(1);
14125 unsigned X86Opcode;
14128 switch (CondOpcode) {
14129 case ISD::UADDO: X86Opcode = X86ISD::ADD; X86Cond = X86::COND_B; break;
14130 case ISD::SADDO: X86Opcode = X86ISD::ADD; X86Cond = X86::COND_O; break;
14131 case ISD::USUBO: X86Opcode = X86ISD::SUB; X86Cond = X86::COND_B; break;
14132 case ISD::SSUBO: X86Opcode = X86ISD::SUB; X86Cond = X86::COND_O; break;
14133 case ISD::UMULO: X86Opcode = X86ISD::UMUL; X86Cond = X86::COND_O; break;
14134 case ISD::SMULO: X86Opcode = X86ISD::SMUL; X86Cond = X86::COND_O; break;
14135 default: llvm_unreachable("unexpected overflowing operator");
14137 if (CondOpcode == ISD::UMULO)
14138 VTs = DAG.getVTList(LHS.getValueType(), LHS.getValueType(),
14141 VTs = DAG.getVTList(LHS.getValueType(), MVT::i32);
14143 SDValue X86Op = DAG.getNode(X86Opcode, DL, VTs, LHS, RHS);
14145 if (CondOpcode == ISD::UMULO)
14146 Cond = X86Op.getValue(2);
14148 Cond = X86Op.getValue(1);
14150 CC = DAG.getConstant(X86Cond, DL, MVT::i8);
14155 // Look past the truncate if the high bits are known zero.
14156 if (isTruncWithZeroHighBitsInput(Cond, DAG))
14157 Cond = Cond.getOperand(0);
14159 // We know the result of AND is compared against zero. Try to match
14161 if (Cond.getOpcode() == ISD::AND && Cond.hasOneUse()) {
14162 SDValue NewSetCC = LowerToBT(Cond, ISD::SETNE, DL, DAG);
14163 if (NewSetCC.getNode()) {
14164 CC = NewSetCC.getOperand(0);
14165 Cond = NewSetCC.getOperand(1);
14172 CC = DAG.getConstant(X86::COND_NE, DL, MVT::i8);
14173 Cond = EmitTest(Cond, X86::COND_NE, DL, DAG);
14176 // a < b ? -1 : 0 -> RES = ~setcc_carry
14177 // a < b ? 0 : -1 -> RES = setcc_carry
14178 // a >= b ? -1 : 0 -> RES = setcc_carry
14179 // a >= b ? 0 : -1 -> RES = ~setcc_carry
14180 if (Cond.getOpcode() == X86ISD::SUB) {
14181 Cond = ConvertCmpIfNecessary(Cond, DAG);
14182 unsigned CondCode = cast<ConstantSDNode>(CC)->getZExtValue();
14184 if ((CondCode == X86::COND_AE || CondCode == X86::COND_B) &&
14185 (isAllOnes(Op1) || isAllOnes(Op2)) && (isZero(Op1) || isZero(Op2))) {
14186 SDValue Res = DAG.getNode(X86ISD::SETCC_CARRY, DL, Op.getValueType(),
14187 DAG.getConstant(X86::COND_B, DL, MVT::i8),
14189 if (isAllOnes(Op1) != (CondCode == X86::COND_B))
14190 return DAG.getNOT(DL, Res, Res.getValueType());
14195 // X86 doesn't have an i8 cmov. If both operands are the result of a truncate
14196 // widen the cmov and push the truncate through. This avoids introducing a new
14197 // branch during isel and doesn't add any extensions.
14198 if (Op.getValueType() == MVT::i8 &&
14199 Op1.getOpcode() == ISD::TRUNCATE && Op2.getOpcode() == ISD::TRUNCATE) {
14200 SDValue T1 = Op1.getOperand(0), T2 = Op2.getOperand(0);
14201 if (T1.getValueType() == T2.getValueType() &&
14202 // Blacklist CopyFromReg to avoid partial register stalls.
14203 T1.getOpcode() != ISD::CopyFromReg && T2.getOpcode()!=ISD::CopyFromReg){
14204 SDVTList VTs = DAG.getVTList(T1.getValueType(), MVT::Glue);
14205 SDValue Cmov = DAG.getNode(X86ISD::CMOV, DL, VTs, T2, T1, CC, Cond);
14206 return DAG.getNode(ISD::TRUNCATE, DL, Op.getValueType(), Cmov);
14210 // X86ISD::CMOV means set the result (which is operand 1) to the RHS if
14211 // condition is true.
14212 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::Glue);
14213 SDValue Ops[] = { Op2, Op1, CC, Cond };
14214 return DAG.getNode(X86ISD::CMOV, DL, VTs, Ops);
14217 static SDValue LowerSIGN_EXTEND_AVX512(SDValue Op,
14218 const X86Subtarget *Subtarget,
14219 SelectionDAG &DAG) {
14220 MVT VT = Op->getSimpleValueType(0);
14221 SDValue In = Op->getOperand(0);
14222 MVT InVT = In.getSimpleValueType();
14223 MVT VTElt = VT.getVectorElementType();
14224 MVT InVTElt = InVT.getVectorElementType();
14228 if ((InVTElt == MVT::i1) &&
14229 (((Subtarget->hasBWI() && Subtarget->hasVLX() &&
14230 VT.getSizeInBits() <= 256 && VTElt.getSizeInBits() <= 16)) ||
14232 ((Subtarget->hasBWI() && VT.is512BitVector() &&
14233 VTElt.getSizeInBits() <= 16)) ||
14235 ((Subtarget->hasDQI() && Subtarget->hasVLX() &&
14236 VT.getSizeInBits() <= 256 && VTElt.getSizeInBits() >= 32)) ||
14238 ((Subtarget->hasDQI() && VT.is512BitVector() &&
14239 VTElt.getSizeInBits() >= 32))))
14240 return DAG.getNode(X86ISD::VSEXT, dl, VT, In);
14242 unsigned int NumElts = VT.getVectorNumElements();
14244 if (NumElts != 8 && NumElts != 16 && !Subtarget->hasBWI())
14247 if (VT.is512BitVector() && InVT.getVectorElementType() != MVT::i1) {
14248 if (In.getOpcode() == X86ISD::VSEXT || In.getOpcode() == X86ISD::VZEXT)
14249 return DAG.getNode(In.getOpcode(), dl, VT, In.getOperand(0));
14250 return DAG.getNode(X86ISD::VSEXT, dl, VT, In);
14253 assert (InVT.getVectorElementType() == MVT::i1 && "Unexpected vector type");
14254 MVT ExtVT = NumElts == 8 ? MVT::v8i64 : MVT::v16i32;
14256 DAG.getConstant(APInt::getAllOnesValue(ExtVT.getScalarSizeInBits()), dl,
14259 DAG.getConstant(APInt::getNullValue(ExtVT.getScalarSizeInBits()), dl, ExtVT);
14261 SDValue V = DAG.getNode(ISD::VSELECT, dl, ExtVT, In, NegOne, Zero);
14262 if (VT.is512BitVector())
14264 return DAG.getNode(X86ISD::VTRUNC, dl, VT, V);
14267 static SDValue LowerSIGN_EXTEND_VECTOR_INREG(SDValue Op,
14268 const X86Subtarget *Subtarget,
14269 SelectionDAG &DAG) {
14270 SDValue In = Op->getOperand(0);
14271 MVT VT = Op->getSimpleValueType(0);
14272 MVT InVT = In.getSimpleValueType();
14273 assert(VT.getSizeInBits() == InVT.getSizeInBits());
14275 MVT InSVT = InVT.getScalarType();
14276 assert(VT.getScalarType().getScalarSizeInBits() > InSVT.getScalarSizeInBits());
14278 if (VT != MVT::v2i64 && VT != MVT::v4i32 && VT != MVT::v8i16)
14280 if (InSVT != MVT::i32 && InSVT != MVT::i16 && InSVT != MVT::i8)
14285 // SSE41 targets can use the pmovsx* instructions directly.
14286 if (Subtarget->hasSSE41())
14287 return DAG.getNode(X86ISD::VSEXT, dl, VT, In);
14289 // pre-SSE41 targets unpack lower lanes and then sign-extend using SRAI.
14293 // As SRAI is only available on i16/i32 types, we expand only up to i32
14294 // and handle i64 separately.
14295 while (CurrVT != VT && CurrVT.getScalarType() != MVT::i32) {
14296 Curr = DAG.getNode(X86ISD::UNPCKL, dl, CurrVT, DAG.getUNDEF(CurrVT), Curr);
14297 MVT CurrSVT = MVT::getIntegerVT(CurrVT.getScalarSizeInBits() * 2);
14298 CurrVT = MVT::getVectorVT(CurrSVT, CurrVT.getVectorNumElements() / 2);
14299 Curr = DAG.getBitcast(CurrVT, Curr);
14302 SDValue SignExt = Curr;
14303 if (CurrVT != InVT) {
14304 unsigned SignExtShift =
14305 CurrVT.getScalarSizeInBits() - InSVT.getScalarSizeInBits();
14306 SignExt = DAG.getNode(X86ISD::VSRAI, dl, CurrVT, Curr,
14307 DAG.getConstant(SignExtShift, dl, MVT::i8));
14313 if (VT == MVT::v2i64 && CurrVT == MVT::v4i32) {
14314 SDValue Sign = DAG.getNode(X86ISD::VSRAI, dl, CurrVT, Curr,
14315 DAG.getConstant(31, dl, MVT::i8));
14316 SDValue Ext = DAG.getVectorShuffle(CurrVT, dl, SignExt, Sign, {0, 4, 1, 5});
14317 return DAG.getBitcast(VT, Ext);
14323 static SDValue LowerSIGN_EXTEND(SDValue Op, const X86Subtarget *Subtarget,
14324 SelectionDAG &DAG) {
14325 MVT VT = Op->getSimpleValueType(0);
14326 SDValue In = Op->getOperand(0);
14327 MVT InVT = In.getSimpleValueType();
14330 if (VT.is512BitVector() || InVT.getVectorElementType() == MVT::i1)
14331 return LowerSIGN_EXTEND_AVX512(Op, Subtarget, DAG);
14333 if ((VT != MVT::v4i64 || InVT != MVT::v4i32) &&
14334 (VT != MVT::v8i32 || InVT != MVT::v8i16) &&
14335 (VT != MVT::v16i16 || InVT != MVT::v16i8))
14338 if (Subtarget->hasInt256())
14339 return DAG.getNode(X86ISD::VSEXT, dl, VT, In);
14341 // Optimize vectors in AVX mode
14342 // Sign extend v8i16 to v8i32 and
14345 // Divide input vector into two parts
14346 // for v4i32 the shuffle mask will be { 0, 1, -1, -1} {2, 3, -1, -1}
14347 // use vpmovsx instruction to extend v4i32 -> v2i64; v8i16 -> v4i32
14348 // concat the vectors to original VT
14350 unsigned NumElems = InVT.getVectorNumElements();
14351 SDValue Undef = DAG.getUNDEF(InVT);
14353 SmallVector<int,8> ShufMask1(NumElems, -1);
14354 for (unsigned i = 0; i != NumElems/2; ++i)
14357 SDValue OpLo = DAG.getVectorShuffle(InVT, dl, In, Undef, &ShufMask1[0]);
14359 SmallVector<int,8> ShufMask2(NumElems, -1);
14360 for (unsigned i = 0; i != NumElems/2; ++i)
14361 ShufMask2[i] = i + NumElems/2;
14363 SDValue OpHi = DAG.getVectorShuffle(InVT, dl, In, Undef, &ShufMask2[0]);
14365 MVT HalfVT = MVT::getVectorVT(VT.getScalarType(),
14366 VT.getVectorNumElements()/2);
14368 OpLo = DAG.getNode(X86ISD::VSEXT, dl, HalfVT, OpLo);
14369 OpHi = DAG.getNode(X86ISD::VSEXT, dl, HalfVT, OpHi);
14371 return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT, OpLo, OpHi);
14374 // Lower vector extended loads using a shuffle. If SSSE3 is not available we
14375 // may emit an illegal shuffle but the expansion is still better than scalar
14376 // code. We generate X86ISD::VSEXT for SEXTLOADs if it's available, otherwise
14377 // we'll emit a shuffle and a arithmetic shift.
14378 // FIXME: Is the expansion actually better than scalar code? It doesn't seem so.
14379 // TODO: It is possible to support ZExt by zeroing the undef values during
14380 // the shuffle phase or after the shuffle.
14381 static SDValue LowerExtendedLoad(SDValue Op, const X86Subtarget *Subtarget,
14382 SelectionDAG &DAG) {
14383 MVT RegVT = Op.getSimpleValueType();
14384 assert(RegVT.isVector() && "We only custom lower vector sext loads.");
14385 assert(RegVT.isInteger() &&
14386 "We only custom lower integer vector sext loads.");
14388 // Nothing useful we can do without SSE2 shuffles.
14389 assert(Subtarget->hasSSE2() && "We only custom lower sext loads with SSE2.");
14391 LoadSDNode *Ld = cast<LoadSDNode>(Op.getNode());
14393 EVT MemVT = Ld->getMemoryVT();
14394 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
14395 unsigned RegSz = RegVT.getSizeInBits();
14397 ISD::LoadExtType Ext = Ld->getExtensionType();
14399 assert((Ext == ISD::EXTLOAD || Ext == ISD::SEXTLOAD)
14400 && "Only anyext and sext are currently implemented.");
14401 assert(MemVT != RegVT && "Cannot extend to the same type");
14402 assert(MemVT.isVector() && "Must load a vector from memory");
14404 unsigned NumElems = RegVT.getVectorNumElements();
14405 unsigned MemSz = MemVT.getSizeInBits();
14406 assert(RegSz > MemSz && "Register size must be greater than the mem size");
14408 if (Ext == ISD::SEXTLOAD && RegSz == 256 && !Subtarget->hasInt256()) {
14409 // The only way in which we have a legal 256-bit vector result but not the
14410 // integer 256-bit operations needed to directly lower a sextload is if we
14411 // have AVX1 but not AVX2. In that case, we can always emit a sextload to
14412 // a 128-bit vector and a normal sign_extend to 256-bits that should get
14413 // correctly legalized. We do this late to allow the canonical form of
14414 // sextload to persist throughout the rest of the DAG combiner -- it wants
14415 // to fold together any extensions it can, and so will fuse a sign_extend
14416 // of an sextload into a sextload targeting a wider value.
14418 if (MemSz == 128) {
14419 // Just switch this to a normal load.
14420 assert(TLI.isTypeLegal(MemVT) && "If the memory type is a 128-bit type, "
14421 "it must be a legal 128-bit vector "
14423 Load = DAG.getLoad(MemVT, dl, Ld->getChain(), Ld->getBasePtr(),
14424 Ld->getPointerInfo(), Ld->isVolatile(), Ld->isNonTemporal(),
14425 Ld->isInvariant(), Ld->getAlignment());
14427 assert(MemSz < 128 &&
14428 "Can't extend a type wider than 128 bits to a 256 bit vector!");
14429 // Do an sext load to a 128-bit vector type. We want to use the same
14430 // number of elements, but elements half as wide. This will end up being
14431 // recursively lowered by this routine, but will succeed as we definitely
14432 // have all the necessary features if we're using AVX1.
14434 EVT::getIntegerVT(*DAG.getContext(), RegVT.getScalarSizeInBits() / 2);
14435 EVT HalfVecVT = EVT::getVectorVT(*DAG.getContext(), HalfEltVT, NumElems);
14437 DAG.getExtLoad(Ext, dl, HalfVecVT, Ld->getChain(), Ld->getBasePtr(),
14438 Ld->getPointerInfo(), MemVT, Ld->isVolatile(),
14439 Ld->isNonTemporal(), Ld->isInvariant(),
14440 Ld->getAlignment());
14443 // Replace chain users with the new chain.
14444 assert(Load->getNumValues() == 2 && "Loads must carry a chain!");
14445 DAG.ReplaceAllUsesOfValueWith(SDValue(Ld, 1), Load.getValue(1));
14447 // Finally, do a normal sign-extend to the desired register.
14448 return DAG.getSExtOrTrunc(Load, dl, RegVT);
14451 // All sizes must be a power of two.
14452 assert(isPowerOf2_32(RegSz * MemSz * NumElems) &&
14453 "Non-power-of-two elements are not custom lowered!");
14455 // Attempt to load the original value using scalar loads.
14456 // Find the largest scalar type that divides the total loaded size.
14457 MVT SclrLoadTy = MVT::i8;
14458 for (MVT Tp : MVT::integer_valuetypes()) {
14459 if (TLI.isTypeLegal(Tp) && ((MemSz % Tp.getSizeInBits()) == 0)) {
14464 // On 32bit systems, we can't save 64bit integers. Try bitcasting to F64.
14465 if (TLI.isTypeLegal(MVT::f64) && SclrLoadTy.getSizeInBits() < 64 &&
14467 SclrLoadTy = MVT::f64;
14469 // Calculate the number of scalar loads that we need to perform
14470 // in order to load our vector from memory.
14471 unsigned NumLoads = MemSz / SclrLoadTy.getSizeInBits();
14473 assert((Ext != ISD::SEXTLOAD || NumLoads == 1) &&
14474 "Can only lower sext loads with a single scalar load!");
14476 unsigned loadRegZize = RegSz;
14477 if (Ext == ISD::SEXTLOAD && RegSz >= 256)
14480 // Represent our vector as a sequence of elements which are the
14481 // largest scalar that we can load.
14482 EVT LoadUnitVecVT = EVT::getVectorVT(
14483 *DAG.getContext(), SclrLoadTy, loadRegZize / SclrLoadTy.getSizeInBits());
14485 // Represent the data using the same element type that is stored in
14486 // memory. In practice, we ''widen'' MemVT.
14488 EVT::getVectorVT(*DAG.getContext(), MemVT.getScalarType(),
14489 loadRegZize / MemVT.getScalarType().getSizeInBits());
14491 assert(WideVecVT.getSizeInBits() == LoadUnitVecVT.getSizeInBits() &&
14492 "Invalid vector type");
14494 // We can't shuffle using an illegal type.
14495 assert(TLI.isTypeLegal(WideVecVT) &&
14496 "We only lower types that form legal widened vector types");
14498 SmallVector<SDValue, 8> Chains;
14499 SDValue Ptr = Ld->getBasePtr();
14500 SDValue Increment = DAG.getConstant(SclrLoadTy.getSizeInBits() / 8, dl,
14501 TLI.getPointerTy(DAG.getDataLayout()));
14502 SDValue Res = DAG.getUNDEF(LoadUnitVecVT);
14504 for (unsigned i = 0; i < NumLoads; ++i) {
14505 // Perform a single load.
14506 SDValue ScalarLoad =
14507 DAG.getLoad(SclrLoadTy, dl, Ld->getChain(), Ptr, Ld->getPointerInfo(),
14508 Ld->isVolatile(), Ld->isNonTemporal(), Ld->isInvariant(),
14509 Ld->getAlignment());
14510 Chains.push_back(ScalarLoad.getValue(1));
14511 // Create the first element type using SCALAR_TO_VECTOR in order to avoid
14512 // another round of DAGCombining.
14514 Res = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, LoadUnitVecVT, ScalarLoad);
14516 Res = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, LoadUnitVecVT, Res,
14517 ScalarLoad, DAG.getIntPtrConstant(i, dl));
14519 Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr, Increment);
14522 SDValue TF = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Chains);
14524 // Bitcast the loaded value to a vector of the original element type, in
14525 // the size of the target vector type.
14526 SDValue SlicedVec = DAG.getBitcast(WideVecVT, Res);
14527 unsigned SizeRatio = RegSz / MemSz;
14529 if (Ext == ISD::SEXTLOAD) {
14530 // If we have SSE4.1, we can directly emit a VSEXT node.
14531 if (Subtarget->hasSSE41()) {
14532 SDValue Sext = DAG.getNode(X86ISD::VSEXT, dl, RegVT, SlicedVec);
14533 DAG.ReplaceAllUsesOfValueWith(SDValue(Ld, 1), TF);
14537 // Otherwise we'll shuffle the small elements in the high bits of the
14538 // larger type and perform an arithmetic shift. If the shift is not legal
14539 // it's better to scalarize.
14540 assert(TLI.isOperationLegalOrCustom(ISD::SRA, RegVT) &&
14541 "We can't implement a sext load without an arithmetic right shift!");
14543 // Redistribute the loaded elements into the different locations.
14544 SmallVector<int, 16> ShuffleVec(NumElems * SizeRatio, -1);
14545 for (unsigned i = 0; i != NumElems; ++i)
14546 ShuffleVec[i * SizeRatio + SizeRatio - 1] = i;
14548 SDValue Shuff = DAG.getVectorShuffle(
14549 WideVecVT, dl, SlicedVec, DAG.getUNDEF(WideVecVT), &ShuffleVec[0]);
14551 Shuff = DAG.getBitcast(RegVT, Shuff);
14553 // Build the arithmetic shift.
14554 unsigned Amt = RegVT.getVectorElementType().getSizeInBits() -
14555 MemVT.getVectorElementType().getSizeInBits();
14557 DAG.getNode(ISD::SRA, dl, RegVT, Shuff,
14558 DAG.getConstant(Amt, dl, RegVT));
14560 DAG.ReplaceAllUsesOfValueWith(SDValue(Ld, 1), TF);
14564 // Redistribute the loaded elements into the different locations.
14565 SmallVector<int, 16> ShuffleVec(NumElems * SizeRatio, -1);
14566 for (unsigned i = 0; i != NumElems; ++i)
14567 ShuffleVec[i * SizeRatio] = i;
14569 SDValue Shuff = DAG.getVectorShuffle(WideVecVT, dl, SlicedVec,
14570 DAG.getUNDEF(WideVecVT), &ShuffleVec[0]);
14572 // Bitcast to the requested type.
14573 Shuff = DAG.getBitcast(RegVT, Shuff);
14574 DAG.ReplaceAllUsesOfValueWith(SDValue(Ld, 1), TF);
14578 // isAndOrOfSingleUseSetCCs - Return true if node is an ISD::AND or
14579 // ISD::OR of two X86ISD::SETCC nodes each of which has no other use apart
14580 // from the AND / OR.
14581 static bool isAndOrOfSetCCs(SDValue Op, unsigned &Opc) {
14582 Opc = Op.getOpcode();
14583 if (Opc != ISD::OR && Opc != ISD::AND)
14585 return (Op.getOperand(0).getOpcode() == X86ISD::SETCC &&
14586 Op.getOperand(0).hasOneUse() &&
14587 Op.getOperand(1).getOpcode() == X86ISD::SETCC &&
14588 Op.getOperand(1).hasOneUse());
14591 // isXor1OfSetCC - Return true if node is an ISD::XOR of a X86ISD::SETCC and
14592 // 1 and that the SETCC node has a single use.
14593 static bool isXor1OfSetCC(SDValue Op) {
14594 if (Op.getOpcode() != ISD::XOR)
14596 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(Op.getOperand(1));
14597 if (N1C && N1C->getAPIntValue() == 1) {
14598 return Op.getOperand(0).getOpcode() == X86ISD::SETCC &&
14599 Op.getOperand(0).hasOneUse();
14604 SDValue X86TargetLowering::LowerBRCOND(SDValue Op, SelectionDAG &DAG) const {
14605 bool addTest = true;
14606 SDValue Chain = Op.getOperand(0);
14607 SDValue Cond = Op.getOperand(1);
14608 SDValue Dest = Op.getOperand(2);
14611 bool Inverted = false;
14613 if (Cond.getOpcode() == ISD::SETCC) {
14614 // Check for setcc([su]{add,sub,mul}o == 0).
14615 if (cast<CondCodeSDNode>(Cond.getOperand(2))->get() == ISD::SETEQ &&
14616 isa<ConstantSDNode>(Cond.getOperand(1)) &&
14617 cast<ConstantSDNode>(Cond.getOperand(1))->isNullValue() &&
14618 Cond.getOperand(0).getResNo() == 1 &&
14619 (Cond.getOperand(0).getOpcode() == ISD::SADDO ||
14620 Cond.getOperand(0).getOpcode() == ISD::UADDO ||
14621 Cond.getOperand(0).getOpcode() == ISD::SSUBO ||
14622 Cond.getOperand(0).getOpcode() == ISD::USUBO ||
14623 Cond.getOperand(0).getOpcode() == ISD::SMULO ||
14624 Cond.getOperand(0).getOpcode() == ISD::UMULO)) {
14626 Cond = Cond.getOperand(0);
14628 SDValue NewCond = LowerSETCC(Cond, DAG);
14629 if (NewCond.getNode())
14634 // FIXME: LowerXALUO doesn't handle these!!
14635 else if (Cond.getOpcode() == X86ISD::ADD ||
14636 Cond.getOpcode() == X86ISD::SUB ||
14637 Cond.getOpcode() == X86ISD::SMUL ||
14638 Cond.getOpcode() == X86ISD::UMUL)
14639 Cond = LowerXALUO(Cond, DAG);
14642 // Look pass (and (setcc_carry (cmp ...)), 1).
14643 if (Cond.getOpcode() == ISD::AND &&
14644 Cond.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY) {
14645 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Cond.getOperand(1));
14646 if (C && C->getAPIntValue() == 1)
14647 Cond = Cond.getOperand(0);
14650 // If condition flag is set by a X86ISD::CMP, then use it as the condition
14651 // setting operand in place of the X86ISD::SETCC.
14652 unsigned CondOpcode = Cond.getOpcode();
14653 if (CondOpcode == X86ISD::SETCC ||
14654 CondOpcode == X86ISD::SETCC_CARRY) {
14655 CC = Cond.getOperand(0);
14657 SDValue Cmp = Cond.getOperand(1);
14658 unsigned Opc = Cmp.getOpcode();
14659 // FIXME: WHY THE SPECIAL CASING OF LogicalCmp??
14660 if (isX86LogicalCmp(Cmp) || Opc == X86ISD::BT) {
14664 switch (cast<ConstantSDNode>(CC)->getZExtValue()) {
14668 // These can only come from an arithmetic instruction with overflow,
14669 // e.g. SADDO, UADDO.
14670 Cond = Cond.getNode()->getOperand(1);
14676 CondOpcode = Cond.getOpcode();
14677 if (CondOpcode == ISD::UADDO || CondOpcode == ISD::SADDO ||
14678 CondOpcode == ISD::USUBO || CondOpcode == ISD::SSUBO ||
14679 ((CondOpcode == ISD::UMULO || CondOpcode == ISD::SMULO) &&
14680 Cond.getOperand(0).getValueType() != MVT::i8)) {
14681 SDValue LHS = Cond.getOperand(0);
14682 SDValue RHS = Cond.getOperand(1);
14683 unsigned X86Opcode;
14686 // Keep this in sync with LowerXALUO, otherwise we might create redundant
14687 // instructions that can't be removed afterwards (i.e. X86ISD::ADD and
14689 switch (CondOpcode) {
14690 case ISD::UADDO: X86Opcode = X86ISD::ADD; X86Cond = X86::COND_B; break;
14692 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(RHS))
14694 X86Opcode = X86ISD::INC; X86Cond = X86::COND_O;
14697 X86Opcode = X86ISD::ADD; X86Cond = X86::COND_O; break;
14698 case ISD::USUBO: X86Opcode = X86ISD::SUB; X86Cond = X86::COND_B; break;
14700 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(RHS))
14702 X86Opcode = X86ISD::DEC; X86Cond = X86::COND_O;
14705 X86Opcode = X86ISD::SUB; X86Cond = X86::COND_O; break;
14706 case ISD::UMULO: X86Opcode = X86ISD::UMUL; X86Cond = X86::COND_O; break;
14707 case ISD::SMULO: X86Opcode = X86ISD::SMUL; X86Cond = X86::COND_O; break;
14708 default: llvm_unreachable("unexpected overflowing operator");
14711 X86Cond = X86::GetOppositeBranchCondition((X86::CondCode)X86Cond);
14712 if (CondOpcode == ISD::UMULO)
14713 VTs = DAG.getVTList(LHS.getValueType(), LHS.getValueType(),
14716 VTs = DAG.getVTList(LHS.getValueType(), MVT::i32);
14718 SDValue X86Op = DAG.getNode(X86Opcode, dl, VTs, LHS, RHS);
14720 if (CondOpcode == ISD::UMULO)
14721 Cond = X86Op.getValue(2);
14723 Cond = X86Op.getValue(1);
14725 CC = DAG.getConstant(X86Cond, dl, MVT::i8);
14729 if (Cond.hasOneUse() && isAndOrOfSetCCs(Cond, CondOpc)) {
14730 SDValue Cmp = Cond.getOperand(0).getOperand(1);
14731 if (CondOpc == ISD::OR) {
14732 // Also, recognize the pattern generated by an FCMP_UNE. We can emit
14733 // two branches instead of an explicit OR instruction with a
14735 if (Cmp == Cond.getOperand(1).getOperand(1) &&
14736 isX86LogicalCmp(Cmp)) {
14737 CC = Cond.getOperand(0).getOperand(0);
14738 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
14739 Chain, Dest, CC, Cmp);
14740 CC = Cond.getOperand(1).getOperand(0);
14744 } else { // ISD::AND
14745 // Also, recognize the pattern generated by an FCMP_OEQ. We can emit
14746 // two branches instead of an explicit AND instruction with a
14747 // separate test. However, we only do this if this block doesn't
14748 // have a fall-through edge, because this requires an explicit
14749 // jmp when the condition is false.
14750 if (Cmp == Cond.getOperand(1).getOperand(1) &&
14751 isX86LogicalCmp(Cmp) &&
14752 Op.getNode()->hasOneUse()) {
14753 X86::CondCode CCode =
14754 (X86::CondCode)Cond.getOperand(0).getConstantOperandVal(0);
14755 CCode = X86::GetOppositeBranchCondition(CCode);
14756 CC = DAG.getConstant(CCode, dl, MVT::i8);
14757 SDNode *User = *Op.getNode()->use_begin();
14758 // Look for an unconditional branch following this conditional branch.
14759 // We need this because we need to reverse the successors in order
14760 // to implement FCMP_OEQ.
14761 if (User->getOpcode() == ISD::BR) {
14762 SDValue FalseBB = User->getOperand(1);
14764 DAG.UpdateNodeOperands(User, User->getOperand(0), Dest);
14765 assert(NewBR == User);
14769 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
14770 Chain, Dest, CC, Cmp);
14771 X86::CondCode CCode =
14772 (X86::CondCode)Cond.getOperand(1).getConstantOperandVal(0);
14773 CCode = X86::GetOppositeBranchCondition(CCode);
14774 CC = DAG.getConstant(CCode, dl, MVT::i8);
14780 } else if (Cond.hasOneUse() && isXor1OfSetCC(Cond)) {
14781 // Recognize for xorb (setcc), 1 patterns. The xor inverts the condition.
14782 // It should be transformed during dag combiner except when the condition
14783 // is set by a arithmetics with overflow node.
14784 X86::CondCode CCode =
14785 (X86::CondCode)Cond.getOperand(0).getConstantOperandVal(0);
14786 CCode = X86::GetOppositeBranchCondition(CCode);
14787 CC = DAG.getConstant(CCode, dl, MVT::i8);
14788 Cond = Cond.getOperand(0).getOperand(1);
14790 } else if (Cond.getOpcode() == ISD::SETCC &&
14791 cast<CondCodeSDNode>(Cond.getOperand(2))->get() == ISD::SETOEQ) {
14792 // For FCMP_OEQ, we can emit
14793 // two branches instead of an explicit AND instruction with a
14794 // separate test. However, we only do this if this block doesn't
14795 // have a fall-through edge, because this requires an explicit
14796 // jmp when the condition is false.
14797 if (Op.getNode()->hasOneUse()) {
14798 SDNode *User = *Op.getNode()->use_begin();
14799 // Look for an unconditional branch following this conditional branch.
14800 // We need this because we need to reverse the successors in order
14801 // to implement FCMP_OEQ.
14802 if (User->getOpcode() == ISD::BR) {
14803 SDValue FalseBB = User->getOperand(1);
14805 DAG.UpdateNodeOperands(User, User->getOperand(0), Dest);
14806 assert(NewBR == User);
14810 SDValue Cmp = DAG.getNode(X86ISD::CMP, dl, MVT::i32,
14811 Cond.getOperand(0), Cond.getOperand(1));
14812 Cmp = ConvertCmpIfNecessary(Cmp, DAG);
14813 CC = DAG.getConstant(X86::COND_NE, dl, MVT::i8);
14814 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
14815 Chain, Dest, CC, Cmp);
14816 CC = DAG.getConstant(X86::COND_P, dl, MVT::i8);
14821 } else if (Cond.getOpcode() == ISD::SETCC &&
14822 cast<CondCodeSDNode>(Cond.getOperand(2))->get() == ISD::SETUNE) {
14823 // For FCMP_UNE, we can emit
14824 // two branches instead of an explicit AND instruction with a
14825 // separate test. However, we only do this if this block doesn't
14826 // have a fall-through edge, because this requires an explicit
14827 // jmp when the condition is false.
14828 if (Op.getNode()->hasOneUse()) {
14829 SDNode *User = *Op.getNode()->use_begin();
14830 // Look for an unconditional branch following this conditional branch.
14831 // We need this because we need to reverse the successors in order
14832 // to implement FCMP_UNE.
14833 if (User->getOpcode() == ISD::BR) {
14834 SDValue FalseBB = User->getOperand(1);
14836 DAG.UpdateNodeOperands(User, User->getOperand(0), Dest);
14837 assert(NewBR == User);
14840 SDValue Cmp = DAG.getNode(X86ISD::CMP, dl, MVT::i32,
14841 Cond.getOperand(0), Cond.getOperand(1));
14842 Cmp = ConvertCmpIfNecessary(Cmp, DAG);
14843 CC = DAG.getConstant(X86::COND_NE, dl, MVT::i8);
14844 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
14845 Chain, Dest, CC, Cmp);
14846 CC = DAG.getConstant(X86::COND_NP, dl, MVT::i8);
14856 // Look pass the truncate if the high bits are known zero.
14857 if (isTruncWithZeroHighBitsInput(Cond, DAG))
14858 Cond = Cond.getOperand(0);
14860 // We know the result of AND is compared against zero. Try to match
14862 if (Cond.getOpcode() == ISD::AND && Cond.hasOneUse()) {
14863 SDValue NewSetCC = LowerToBT(Cond, ISD::SETNE, dl, DAG);
14864 if (NewSetCC.getNode()) {
14865 CC = NewSetCC.getOperand(0);
14866 Cond = NewSetCC.getOperand(1);
14873 X86::CondCode X86Cond = Inverted ? X86::COND_E : X86::COND_NE;
14874 CC = DAG.getConstant(X86Cond, dl, MVT::i8);
14875 Cond = EmitTest(Cond, X86Cond, dl, DAG);
14877 Cond = ConvertCmpIfNecessary(Cond, DAG);
14878 return DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
14879 Chain, Dest, CC, Cond);
14882 // Lower dynamic stack allocation to _alloca call for Cygwin/Mingw targets.
14883 // Calls to _alloca are needed to probe the stack when allocating more than 4k
14884 // bytes in one go. Touching the stack at 4K increments is necessary to ensure
14885 // that the guard pages used by the OS virtual memory manager are allocated in
14886 // correct sequence.
14888 X86TargetLowering::LowerDYNAMIC_STACKALLOC(SDValue Op,
14889 SelectionDAG &DAG) const {
14890 MachineFunction &MF = DAG.getMachineFunction();
14891 bool SplitStack = MF.shouldSplitStack();
14892 bool Lower = (Subtarget->isOSWindows() && !Subtarget->isTargetMachO()) ||
14897 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
14898 SDNode* Node = Op.getNode();
14900 unsigned SPReg = TLI.getStackPointerRegisterToSaveRestore();
14901 assert(SPReg && "Target cannot require DYNAMIC_STACKALLOC expansion and"
14902 " not tell us which reg is the stack pointer!");
14903 EVT VT = Node->getValueType(0);
14904 SDValue Tmp1 = SDValue(Node, 0);
14905 SDValue Tmp2 = SDValue(Node, 1);
14906 SDValue Tmp3 = Node->getOperand(2);
14907 SDValue Chain = Tmp1.getOperand(0);
14909 // Chain the dynamic stack allocation so that it doesn't modify the stack
14910 // pointer when other instructions are using the stack.
14911 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(0, dl, true),
14914 SDValue Size = Tmp2.getOperand(1);
14915 SDValue SP = DAG.getCopyFromReg(Chain, dl, SPReg, VT);
14916 Chain = SP.getValue(1);
14917 unsigned Align = cast<ConstantSDNode>(Tmp3)->getZExtValue();
14918 const TargetFrameLowering &TFI = *Subtarget->getFrameLowering();
14919 unsigned StackAlign = TFI.getStackAlignment();
14920 Tmp1 = DAG.getNode(ISD::SUB, dl, VT, SP, Size); // Value
14921 if (Align > StackAlign)
14922 Tmp1 = DAG.getNode(ISD::AND, dl, VT, Tmp1,
14923 DAG.getConstant(-(uint64_t)Align, dl, VT));
14924 Chain = DAG.getCopyToReg(Chain, dl, SPReg, Tmp1); // Output chain
14926 Tmp2 = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(0, dl, true),
14927 DAG.getIntPtrConstant(0, dl, true), SDValue(),
14930 SDValue Ops[2] = { Tmp1, Tmp2 };
14931 return DAG.getMergeValues(Ops, dl);
14935 SDValue Chain = Op.getOperand(0);
14936 SDValue Size = Op.getOperand(1);
14937 unsigned Align = cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue();
14938 EVT VT = Op.getNode()->getValueType(0);
14940 bool Is64Bit = Subtarget->is64Bit();
14941 MVT SPTy = getPointerTy(DAG.getDataLayout());
14944 MachineRegisterInfo &MRI = MF.getRegInfo();
14947 // The 64 bit implementation of segmented stacks needs to clobber both r10
14948 // r11. This makes it impossible to use it along with nested parameters.
14949 const Function *F = MF.getFunction();
14951 for (Function::const_arg_iterator I = F->arg_begin(), E = F->arg_end();
14953 if (I->hasNestAttr())
14954 report_fatal_error("Cannot use segmented stacks with functions that "
14955 "have nested arguments.");
14958 const TargetRegisterClass *AddrRegClass = getRegClassFor(SPTy);
14959 unsigned Vreg = MRI.createVirtualRegister(AddrRegClass);
14960 Chain = DAG.getCopyToReg(Chain, dl, Vreg, Size);
14961 SDValue Value = DAG.getNode(X86ISD::SEG_ALLOCA, dl, SPTy, Chain,
14962 DAG.getRegister(Vreg, SPTy));
14963 SDValue Ops1[2] = { Value, Chain };
14964 return DAG.getMergeValues(Ops1, dl);
14967 const unsigned Reg = (Subtarget->isTarget64BitLP64() ? X86::RAX : X86::EAX);
14969 Chain = DAG.getCopyToReg(Chain, dl, Reg, Size, Flag);
14970 Flag = Chain.getValue(1);
14971 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
14973 Chain = DAG.getNode(X86ISD::WIN_ALLOCA, dl, NodeTys, Chain, Flag);
14975 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
14976 unsigned SPReg = RegInfo->getStackRegister();
14977 SDValue SP = DAG.getCopyFromReg(Chain, dl, SPReg, SPTy);
14978 Chain = SP.getValue(1);
14981 SP = DAG.getNode(ISD::AND, dl, VT, SP.getValue(0),
14982 DAG.getConstant(-(uint64_t)Align, dl, VT));
14983 Chain = DAG.getCopyToReg(Chain, dl, SPReg, SP);
14986 SDValue Ops1[2] = { SP, Chain };
14987 return DAG.getMergeValues(Ops1, dl);
14991 SDValue X86TargetLowering::LowerVASTART(SDValue Op, SelectionDAG &DAG) const {
14992 MachineFunction &MF = DAG.getMachineFunction();
14993 auto PtrVT = getPointerTy(MF.getDataLayout());
14994 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
14996 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
14999 if (!Subtarget->is64Bit() || Subtarget->isTargetWin64()) {
15000 // vastart just stores the address of the VarArgsFrameIndex slot into the
15001 // memory location argument.
15002 SDValue FR = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(), PtrVT);
15003 return DAG.getStore(Op.getOperand(0), DL, FR, Op.getOperand(1),
15004 MachinePointerInfo(SV), false, false, 0);
15008 // gp_offset (0 - 6 * 8)
15009 // fp_offset (48 - 48 + 8 * 16)
15010 // overflow_arg_area (point to parameters coming in memory).
15012 SmallVector<SDValue, 8> MemOps;
15013 SDValue FIN = Op.getOperand(1);
15015 SDValue Store = DAG.getStore(Op.getOperand(0), DL,
15016 DAG.getConstant(FuncInfo->getVarArgsGPOffset(),
15018 FIN, MachinePointerInfo(SV), false, false, 0);
15019 MemOps.push_back(Store);
15022 FIN = DAG.getNode(ISD::ADD, DL, PtrVT, FIN, DAG.getIntPtrConstant(4, DL));
15023 Store = DAG.getStore(Op.getOperand(0), DL,
15024 DAG.getConstant(FuncInfo->getVarArgsFPOffset(), DL,
15026 FIN, MachinePointerInfo(SV, 4), false, false, 0);
15027 MemOps.push_back(Store);
15029 // Store ptr to overflow_arg_area
15030 FIN = DAG.getNode(ISD::ADD, DL, PtrVT, FIN, DAG.getIntPtrConstant(4, DL));
15031 SDValue OVFIN = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(), PtrVT);
15032 Store = DAG.getStore(Op.getOperand(0), DL, OVFIN, FIN,
15033 MachinePointerInfo(SV, 8),
15035 MemOps.push_back(Store);
15037 // Store ptr to reg_save_area.
15038 FIN = DAG.getNode(ISD::ADD, DL, PtrVT, FIN, DAG.getIntPtrConstant(8, DL));
15039 SDValue RSFIN = DAG.getFrameIndex(FuncInfo->getRegSaveFrameIndex(), PtrVT);
15040 Store = DAG.getStore(Op.getOperand(0), DL, RSFIN, FIN,
15041 MachinePointerInfo(SV, 16), false, false, 0);
15042 MemOps.push_back(Store);
15043 return DAG.getNode(ISD::TokenFactor, DL, MVT::Other, MemOps);
15046 SDValue X86TargetLowering::LowerVAARG(SDValue Op, SelectionDAG &DAG) const {
15047 assert(Subtarget->is64Bit() &&
15048 "LowerVAARG only handles 64-bit va_arg!");
15049 assert((Subtarget->isTargetLinux() ||
15050 Subtarget->isTargetDarwin()) &&
15051 "Unhandled target in LowerVAARG");
15052 assert(Op.getNode()->getNumOperands() == 4);
15053 SDValue Chain = Op.getOperand(0);
15054 SDValue SrcPtr = Op.getOperand(1);
15055 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
15056 unsigned Align = Op.getConstantOperandVal(3);
15059 EVT ArgVT = Op.getNode()->getValueType(0);
15060 Type *ArgTy = ArgVT.getTypeForEVT(*DAG.getContext());
15061 uint32_t ArgSize = DAG.getDataLayout().getTypeAllocSize(ArgTy);
15064 // Decide which area this value should be read from.
15065 // TODO: Implement the AMD64 ABI in its entirety. This simple
15066 // selection mechanism works only for the basic types.
15067 if (ArgVT == MVT::f80) {
15068 llvm_unreachable("va_arg for f80 not yet implemented");
15069 } else if (ArgVT.isFloatingPoint() && ArgSize <= 16 /*bytes*/) {
15070 ArgMode = 2; // Argument passed in XMM register. Use fp_offset.
15071 } else if (ArgVT.isInteger() && ArgSize <= 32 /*bytes*/) {
15072 ArgMode = 1; // Argument passed in GPR64 register(s). Use gp_offset.
15074 llvm_unreachable("Unhandled argument type in LowerVAARG");
15077 if (ArgMode == 2) {
15078 // Sanity Check: Make sure using fp_offset makes sense.
15079 assert(!Subtarget->useSoftFloat() &&
15080 !(DAG.getMachineFunction().getFunction()->hasFnAttribute(
15081 Attribute::NoImplicitFloat)) &&
15082 Subtarget->hasSSE1());
15085 // Insert VAARG_64 node into the DAG
15086 // VAARG_64 returns two values: Variable Argument Address, Chain
15087 SDValue InstOps[] = {Chain, SrcPtr, DAG.getConstant(ArgSize, dl, MVT::i32),
15088 DAG.getConstant(ArgMode, dl, MVT::i8),
15089 DAG.getConstant(Align, dl, MVT::i32)};
15090 SDVTList VTs = DAG.getVTList(getPointerTy(DAG.getDataLayout()), MVT::Other);
15091 SDValue VAARG = DAG.getMemIntrinsicNode(X86ISD::VAARG_64, dl,
15092 VTs, InstOps, MVT::i64,
15093 MachinePointerInfo(SV),
15095 /*Volatile=*/false,
15097 /*WriteMem=*/true);
15098 Chain = VAARG.getValue(1);
15100 // Load the next argument and return it
15101 return DAG.getLoad(ArgVT, dl,
15104 MachinePointerInfo(),
15105 false, false, false, 0);
15108 static SDValue LowerVACOPY(SDValue Op, const X86Subtarget *Subtarget,
15109 SelectionDAG &DAG) {
15110 // X86-64 va_list is a struct { i32, i32, i8*, i8* }.
15111 assert(Subtarget->is64Bit() && "This code only handles 64-bit va_copy!");
15112 SDValue Chain = Op.getOperand(0);
15113 SDValue DstPtr = Op.getOperand(1);
15114 SDValue SrcPtr = Op.getOperand(2);
15115 const Value *DstSV = cast<SrcValueSDNode>(Op.getOperand(3))->getValue();
15116 const Value *SrcSV = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
15119 return DAG.getMemcpy(Chain, DL, DstPtr, SrcPtr,
15120 DAG.getIntPtrConstant(24, DL), 8, /*isVolatile*/false,
15122 MachinePointerInfo(DstSV), MachinePointerInfo(SrcSV));
15125 // getTargetVShiftByConstNode - Handle vector element shifts where the shift
15126 // amount is a constant. Takes immediate version of shift as input.
15127 static SDValue getTargetVShiftByConstNode(unsigned Opc, SDLoc dl, MVT VT,
15128 SDValue SrcOp, uint64_t ShiftAmt,
15129 SelectionDAG &DAG) {
15130 MVT ElementType = VT.getVectorElementType();
15132 // Fold this packed shift into its first operand if ShiftAmt is 0.
15136 // Check for ShiftAmt >= element width
15137 if (ShiftAmt >= ElementType.getSizeInBits()) {
15138 if (Opc == X86ISD::VSRAI)
15139 ShiftAmt = ElementType.getSizeInBits() - 1;
15141 return DAG.getConstant(0, dl, VT);
15144 assert((Opc == X86ISD::VSHLI || Opc == X86ISD::VSRLI || Opc == X86ISD::VSRAI)
15145 && "Unknown target vector shift-by-constant node");
15147 // Fold this packed vector shift into a build vector if SrcOp is a
15148 // vector of Constants or UNDEFs, and SrcOp valuetype is the same as VT.
15149 if (VT == SrcOp.getSimpleValueType() &&
15150 ISD::isBuildVectorOfConstantSDNodes(SrcOp.getNode())) {
15151 SmallVector<SDValue, 8> Elts;
15152 unsigned NumElts = SrcOp->getNumOperands();
15153 ConstantSDNode *ND;
15156 default: llvm_unreachable(nullptr);
15157 case X86ISD::VSHLI:
15158 for (unsigned i=0; i!=NumElts; ++i) {
15159 SDValue CurrentOp = SrcOp->getOperand(i);
15160 if (CurrentOp->getOpcode() == ISD::UNDEF) {
15161 Elts.push_back(CurrentOp);
15164 ND = cast<ConstantSDNode>(CurrentOp);
15165 const APInt &C = ND->getAPIntValue();
15166 Elts.push_back(DAG.getConstant(C.shl(ShiftAmt), dl, ElementType));
15169 case X86ISD::VSRLI:
15170 for (unsigned i=0; i!=NumElts; ++i) {
15171 SDValue CurrentOp = SrcOp->getOperand(i);
15172 if (CurrentOp->getOpcode() == ISD::UNDEF) {
15173 Elts.push_back(CurrentOp);
15176 ND = cast<ConstantSDNode>(CurrentOp);
15177 const APInt &C = ND->getAPIntValue();
15178 Elts.push_back(DAG.getConstant(C.lshr(ShiftAmt), dl, ElementType));
15181 case X86ISD::VSRAI:
15182 for (unsigned i=0; i!=NumElts; ++i) {
15183 SDValue CurrentOp = SrcOp->getOperand(i);
15184 if (CurrentOp->getOpcode() == ISD::UNDEF) {
15185 Elts.push_back(CurrentOp);
15188 ND = cast<ConstantSDNode>(CurrentOp);
15189 const APInt &C = ND->getAPIntValue();
15190 Elts.push_back(DAG.getConstant(C.ashr(ShiftAmt), dl, ElementType));
15195 return DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Elts);
15198 return DAG.getNode(Opc, dl, VT, SrcOp,
15199 DAG.getConstant(ShiftAmt, dl, MVT::i8));
15202 // getTargetVShiftNode - Handle vector element shifts where the shift amount
15203 // may or may not be a constant. Takes immediate version of shift as input.
15204 static SDValue getTargetVShiftNode(unsigned Opc, SDLoc dl, MVT VT,
15205 SDValue SrcOp, SDValue ShAmt,
15206 SelectionDAG &DAG) {
15207 MVT SVT = ShAmt.getSimpleValueType();
15208 assert((SVT == MVT::i32 || SVT == MVT::i64) && "Unexpected value type!");
15210 // Catch shift-by-constant.
15211 if (ConstantSDNode *CShAmt = dyn_cast<ConstantSDNode>(ShAmt))
15212 return getTargetVShiftByConstNode(Opc, dl, VT, SrcOp,
15213 CShAmt->getZExtValue(), DAG);
15215 // Change opcode to non-immediate version
15217 default: llvm_unreachable("Unknown target vector shift node");
15218 case X86ISD::VSHLI: Opc = X86ISD::VSHL; break;
15219 case X86ISD::VSRLI: Opc = X86ISD::VSRL; break;
15220 case X86ISD::VSRAI: Opc = X86ISD::VSRA; break;
15223 const X86Subtarget &Subtarget =
15224 static_cast<const X86Subtarget &>(DAG.getSubtarget());
15225 if (Subtarget.hasSSE41() && ShAmt.getOpcode() == ISD::ZERO_EXTEND &&
15226 ShAmt.getOperand(0).getSimpleValueType() == MVT::i16) {
15227 // Let the shuffle legalizer expand this shift amount node.
15228 SDValue Op0 = ShAmt.getOperand(0);
15229 Op0 = DAG.getNode(ISD::SCALAR_TO_VECTOR, SDLoc(Op0), MVT::v8i16, Op0);
15230 ShAmt = getShuffleVectorZeroOrUndef(Op0, 0, true, &Subtarget, DAG);
15232 // Need to build a vector containing shift amount.
15233 // SSE/AVX packed shifts only use the lower 64-bit of the shift count.
15234 SmallVector<SDValue, 4> ShOps;
15235 ShOps.push_back(ShAmt);
15236 if (SVT == MVT::i32) {
15237 ShOps.push_back(DAG.getConstant(0, dl, SVT));
15238 ShOps.push_back(DAG.getUNDEF(SVT));
15240 ShOps.push_back(DAG.getUNDEF(SVT));
15242 MVT BVT = SVT == MVT::i32 ? MVT::v4i32 : MVT::v2i64;
15243 ShAmt = DAG.getNode(ISD::BUILD_VECTOR, dl, BVT, ShOps);
15246 // The return type has to be a 128-bit type with the same element
15247 // type as the input type.
15248 MVT EltVT = VT.getVectorElementType();
15249 EVT ShVT = MVT::getVectorVT(EltVT, 128/EltVT.getSizeInBits());
15251 ShAmt = DAG.getBitcast(ShVT, ShAmt);
15252 return DAG.getNode(Opc, dl, VT, SrcOp, ShAmt);
15255 /// \brief Return (and \p Op, \p Mask) for compare instructions or
15256 /// (vselect \p Mask, \p Op, \p PreservedSrc) for others along with the
15257 /// necessary casting or extending for \p Mask when lowering masking intrinsics
15258 static SDValue getVectorMaskingNode(SDValue Op, SDValue Mask,
15259 SDValue PreservedSrc,
15260 const X86Subtarget *Subtarget,
15261 SelectionDAG &DAG) {
15262 EVT VT = Op.getValueType();
15263 EVT MaskVT = EVT::getVectorVT(*DAG.getContext(),
15264 MVT::i1, VT.getVectorNumElements());
15265 SDValue VMask = SDValue();
15266 unsigned OpcodeSelect = ISD::VSELECT;
15269 assert(MaskVT.isSimple() && "invalid mask type");
15271 if (isAllOnes(Mask))
15274 if (MaskVT.bitsGT(Mask.getValueType())) {
15275 EVT newMaskVT = EVT::getIntegerVT(*DAG.getContext(),
15276 MaskVT.getSizeInBits());
15277 VMask = DAG.getBitcast(MaskVT,
15278 DAG.getNode(ISD::ANY_EXTEND, dl, newMaskVT, Mask));
15280 EVT BitcastVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
15281 Mask.getValueType().getSizeInBits());
15282 // In case when MaskVT equals v2i1 or v4i1, low 2 or 4 elements
15283 // are extracted by EXTRACT_SUBVECTOR.
15284 VMask = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MaskVT,
15285 DAG.getBitcast(BitcastVT, Mask),
15286 DAG.getIntPtrConstant(0, dl));
15289 switch (Op.getOpcode()) {
15291 case X86ISD::PCMPEQM:
15292 case X86ISD::PCMPGTM:
15294 case X86ISD::CMPMU:
15295 return DAG.getNode(ISD::AND, dl, VT, Op, VMask);
15296 case X86ISD::VTRUNC:
15297 case X86ISD::VTRUNCS:
15298 case X86ISD::VTRUNCUS:
15299 // We can't use ISD::VSELECT here because it is not always "Legal"
15300 // for the destination type. For example vpmovqb require only AVX512
15301 // and vselect that can operate on byte element type require BWI
15302 OpcodeSelect = X86ISD::SELECT;
15305 if (PreservedSrc.getOpcode() == ISD::UNDEF)
15306 PreservedSrc = getZeroVector(VT, Subtarget, DAG, dl);
15307 return DAG.getNode(OpcodeSelect, dl, VT, VMask, Op, PreservedSrc);
15310 /// \brief Creates an SDNode for a predicated scalar operation.
15311 /// \returns (X86vselect \p Mask, \p Op, \p PreservedSrc).
15312 /// The mask is comming as MVT::i8 and it should be truncated
15313 /// to MVT::i1 while lowering masking intrinsics.
15314 /// The main difference between ScalarMaskingNode and VectorMaskingNode is using
15315 /// "X86select" instead of "vselect". We just can't create the "vselect" node for
15316 /// a scalar instruction.
15317 static SDValue getScalarMaskingNode(SDValue Op, SDValue Mask,
15318 SDValue PreservedSrc,
15319 const X86Subtarget *Subtarget,
15320 SelectionDAG &DAG) {
15321 if (isAllOnes(Mask))
15324 EVT VT = Op.getValueType();
15326 // The mask should be of type MVT::i1
15327 SDValue IMask = DAG.getNode(ISD::TRUNCATE, dl, MVT::i1, Mask);
15329 if (PreservedSrc.getOpcode() == ISD::UNDEF)
15330 PreservedSrc = getZeroVector(VT, Subtarget, DAG, dl);
15331 return DAG.getNode(X86ISD::SELECT, dl, VT, IMask, Op, PreservedSrc);
15334 static int getSEHRegistrationNodeSize(const Function *Fn) {
15335 if (!Fn->hasPersonalityFn())
15336 report_fatal_error(
15337 "querying registration node size for function without personality");
15338 // The RegNodeSize is 6 32-bit words for SEH and 4 for C++ EH. See
15339 // WinEHStatePass for the full struct definition.
15340 switch (classifyEHPersonality(Fn->getPersonalityFn())) {
15341 case EHPersonality::MSVC_X86SEH: return 24;
15342 case EHPersonality::MSVC_CXX: return 16;
15345 report_fatal_error("can only recover FP for MSVC EH personality functions");
15348 /// When the 32-bit MSVC runtime transfers control to us, either to an outlined
15349 /// function or when returning to a parent frame after catching an exception, we
15350 /// recover the parent frame pointer by doing arithmetic on the incoming EBP.
15351 /// Here's the math:
15352 /// RegNodeBase = EntryEBP - RegNodeSize
15353 /// ParentFP = RegNodeBase - RegNodeFrameOffset
15354 /// Subtracting RegNodeSize takes us to the offset of the registration node, and
15355 /// subtracting the offset (negative on x86) takes us back to the parent FP.
15356 static SDValue recoverFramePointer(SelectionDAG &DAG, const Function *Fn,
15357 SDValue EntryEBP) {
15358 MachineFunction &MF = DAG.getMachineFunction();
15361 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
15362 MVT PtrVT = TLI.getPointerTy(DAG.getDataLayout());
15364 // It's possible that the parent function no longer has a personality function
15365 // if the exceptional code was optimized away, in which case we just return
15366 // the incoming EBP.
15367 if (!Fn->hasPersonalityFn())
15370 int RegNodeSize = getSEHRegistrationNodeSize(Fn);
15372 // Get an MCSymbol that will ultimately resolve to the frame offset of the EH
15374 MCSymbol *OffsetSym =
15375 MF.getMMI().getContext().getOrCreateParentFrameOffsetSymbol(
15376 GlobalValue::getRealLinkageName(Fn->getName()));
15377 SDValue OffsetSymVal = DAG.getMCSymbol(OffsetSym, PtrVT);
15378 SDValue RegNodeFrameOffset =
15379 DAG.getNode(ISD::LOCAL_RECOVER, dl, PtrVT, OffsetSymVal);
15381 // RegNodeBase = EntryEBP - RegNodeSize
15382 // ParentFP = RegNodeBase - RegNodeFrameOffset
15383 SDValue RegNodeBase = DAG.getNode(ISD::SUB, dl, PtrVT, EntryEBP,
15384 DAG.getConstant(RegNodeSize, dl, PtrVT));
15385 return DAG.getNode(ISD::SUB, dl, PtrVT, RegNodeBase, RegNodeFrameOffset);
15388 static SDValue LowerINTRINSIC_WO_CHAIN(SDValue Op, const X86Subtarget *Subtarget,
15389 SelectionDAG &DAG) {
15391 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
15392 EVT VT = Op.getValueType();
15393 const IntrinsicData* IntrData = getIntrinsicWithoutChain(IntNo);
15395 switch(IntrData->Type) {
15396 case INTR_TYPE_1OP:
15397 return DAG.getNode(IntrData->Opc0, dl, Op.getValueType(), Op.getOperand(1));
15398 case INTR_TYPE_2OP:
15399 return DAG.getNode(IntrData->Opc0, dl, Op.getValueType(), Op.getOperand(1),
15401 case INTR_TYPE_3OP:
15402 return DAG.getNode(IntrData->Opc0, dl, Op.getValueType(), Op.getOperand(1),
15403 Op.getOperand(2), Op.getOperand(3));
15404 case INTR_TYPE_4OP:
15405 return DAG.getNode(IntrData->Opc0, dl, Op.getValueType(), Op.getOperand(1),
15406 Op.getOperand(2), Op.getOperand(3), Op.getOperand(4));
15407 case INTR_TYPE_1OP_MASK_RM: {
15408 SDValue Src = Op.getOperand(1);
15409 SDValue PassThru = Op.getOperand(2);
15410 SDValue Mask = Op.getOperand(3);
15411 SDValue RoundingMode;
15412 // We allways add rounding mode to the Node.
15413 // If the rounding mode is not specified, we add the
15414 // "current direction" mode.
15415 if (Op.getNumOperands() == 4)
15417 DAG.getConstant(X86::STATIC_ROUNDING::CUR_DIRECTION, dl, MVT::i32);
15419 RoundingMode = Op.getOperand(4);
15420 unsigned IntrWithRoundingModeOpcode = IntrData->Opc1;
15421 if (IntrWithRoundingModeOpcode != 0)
15422 if (cast<ConstantSDNode>(RoundingMode)->getZExtValue() !=
15423 X86::STATIC_ROUNDING::CUR_DIRECTION)
15424 return getVectorMaskingNode(DAG.getNode(IntrWithRoundingModeOpcode,
15425 dl, Op.getValueType(), Src, RoundingMode),
15426 Mask, PassThru, Subtarget, DAG);
15427 return getVectorMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT, Src,
15429 Mask, PassThru, Subtarget, DAG);
15431 case INTR_TYPE_1OP_MASK: {
15432 SDValue Src = Op.getOperand(1);
15433 SDValue PassThru = Op.getOperand(2);
15434 SDValue Mask = Op.getOperand(3);
15435 // We add rounding mode to the Node when
15436 // - RM Opcode is specified and
15437 // - RM is not "current direction".
15438 unsigned IntrWithRoundingModeOpcode = IntrData->Opc1;
15439 if (IntrWithRoundingModeOpcode != 0) {
15440 SDValue Rnd = Op.getOperand(4);
15441 unsigned Round = cast<ConstantSDNode>(Rnd)->getZExtValue();
15442 if (Round != X86::STATIC_ROUNDING::CUR_DIRECTION) {
15443 return getVectorMaskingNode(DAG.getNode(IntrWithRoundingModeOpcode,
15444 dl, Op.getValueType(),
15446 Mask, PassThru, Subtarget, DAG);
15449 return getVectorMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT, Src),
15450 Mask, PassThru, Subtarget, DAG);
15452 case INTR_TYPE_SCALAR_MASK_RM: {
15453 SDValue Src1 = Op.getOperand(1);
15454 SDValue Src2 = Op.getOperand(2);
15455 SDValue Src0 = Op.getOperand(3);
15456 SDValue Mask = Op.getOperand(4);
15457 // There are 2 kinds of intrinsics in this group:
15458 // (1) With supress-all-exceptions (sae) or rounding mode- 6 operands
15459 // (2) With rounding mode and sae - 7 operands.
15460 if (Op.getNumOperands() == 6) {
15461 SDValue Sae = Op.getOperand(5);
15462 unsigned Opc = IntrData->Opc1 ? IntrData->Opc1 : IntrData->Opc0;
15463 return getScalarMaskingNode(DAG.getNode(Opc, dl, VT, Src1, Src2,
15465 Mask, Src0, Subtarget, DAG);
15467 assert(Op.getNumOperands() == 7 && "Unexpected intrinsic form");
15468 SDValue RoundingMode = Op.getOperand(5);
15469 SDValue Sae = Op.getOperand(6);
15470 return getScalarMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT, Src1, Src2,
15471 RoundingMode, Sae),
15472 Mask, Src0, Subtarget, DAG);
15474 case INTR_TYPE_2OP_MASK: {
15475 SDValue Src1 = Op.getOperand(1);
15476 SDValue Src2 = Op.getOperand(2);
15477 SDValue PassThru = Op.getOperand(3);
15478 SDValue Mask = Op.getOperand(4);
15479 // We specify 2 possible opcodes for intrinsics with rounding modes.
15480 // First, we check if the intrinsic may have non-default rounding mode,
15481 // (IntrData->Opc1 != 0), then we check the rounding mode operand.
15482 unsigned IntrWithRoundingModeOpcode = IntrData->Opc1;
15483 if (IntrWithRoundingModeOpcode != 0) {
15484 SDValue Rnd = Op.getOperand(5);
15485 unsigned Round = cast<ConstantSDNode>(Rnd)->getZExtValue();
15486 if (Round != X86::STATIC_ROUNDING::CUR_DIRECTION) {
15487 return getVectorMaskingNode(DAG.getNode(IntrWithRoundingModeOpcode,
15488 dl, Op.getValueType(),
15490 Mask, PassThru, Subtarget, DAG);
15493 return getVectorMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT,
15495 Mask, PassThru, Subtarget, DAG);
15497 case INTR_TYPE_2OP_MASK_RM: {
15498 SDValue Src1 = Op.getOperand(1);
15499 SDValue Src2 = Op.getOperand(2);
15500 SDValue PassThru = Op.getOperand(3);
15501 SDValue Mask = Op.getOperand(4);
15502 // We specify 2 possible modes for intrinsics, with/without rounding modes.
15503 // First, we check if the intrinsic have rounding mode (6 operands),
15504 // if not, we set rounding mode to "current".
15506 if (Op.getNumOperands() == 6)
15507 Rnd = Op.getOperand(5);
15509 Rnd = DAG.getConstant(X86::STATIC_ROUNDING::CUR_DIRECTION, dl, MVT::i32);
15510 return getVectorMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT,
15512 Mask, PassThru, Subtarget, DAG);
15514 case INTR_TYPE_3OP_MASK_RM: {
15515 SDValue Src1 = Op.getOperand(1);
15516 SDValue Src2 = Op.getOperand(2);
15517 SDValue Imm = Op.getOperand(3);
15518 SDValue PassThru = Op.getOperand(4);
15519 SDValue Mask = Op.getOperand(5);
15520 // We specify 2 possible modes for intrinsics, with/without rounding modes.
15521 // First, we check if the intrinsic have rounding mode (7 operands),
15522 // if not, we set rounding mode to "current".
15524 if (Op.getNumOperands() == 7)
15525 Rnd = Op.getOperand(6);
15527 Rnd = DAG.getConstant(X86::STATIC_ROUNDING::CUR_DIRECTION, dl, MVT::i32);
15528 return getVectorMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT,
15529 Src1, Src2, Imm, Rnd),
15530 Mask, PassThru, Subtarget, DAG);
15532 case INTR_TYPE_3OP_MASK: {
15533 SDValue Src1 = Op.getOperand(1);
15534 SDValue Src2 = Op.getOperand(2);
15535 SDValue Src3 = Op.getOperand(3);
15536 SDValue PassThru = Op.getOperand(4);
15537 SDValue Mask = Op.getOperand(5);
15538 // We specify 2 possible opcodes for intrinsics with rounding modes.
15539 // First, we check if the intrinsic may have non-default rounding mode,
15540 // (IntrData->Opc1 != 0), then we check the rounding mode operand.
15541 unsigned IntrWithRoundingModeOpcode = IntrData->Opc1;
15542 if (IntrWithRoundingModeOpcode != 0) {
15543 SDValue Rnd = Op.getOperand(6);
15544 unsigned Round = cast<ConstantSDNode>(Rnd)->getZExtValue();
15545 if (Round != X86::STATIC_ROUNDING::CUR_DIRECTION) {
15546 return getVectorMaskingNode(DAG.getNode(IntrWithRoundingModeOpcode,
15547 dl, Op.getValueType(),
15548 Src1, Src2, Src3, Rnd),
15549 Mask, PassThru, Subtarget, DAG);
15552 return getVectorMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT,
15554 Mask, PassThru, Subtarget, DAG);
15556 case VPERM_3OP_MASKZ:
15557 case VPERM_3OP_MASK:
15560 case FMA_OP_MASK: {
15561 SDValue Src1 = Op.getOperand(1);
15562 SDValue Src2 = Op.getOperand(2);
15563 SDValue Src3 = Op.getOperand(3);
15564 SDValue Mask = Op.getOperand(4);
15565 EVT VT = Op.getValueType();
15566 SDValue PassThru = SDValue();
15568 // set PassThru element
15569 if (IntrData->Type == VPERM_3OP_MASKZ || IntrData->Type == FMA_OP_MASKZ)
15570 PassThru = getZeroVector(VT, Subtarget, DAG, dl);
15571 else if (IntrData->Type == FMA_OP_MASK3)
15576 // We specify 2 possible opcodes for intrinsics with rounding modes.
15577 // First, we check if the intrinsic may have non-default rounding mode,
15578 // (IntrData->Opc1 != 0), then we check the rounding mode operand.
15579 unsigned IntrWithRoundingModeOpcode = IntrData->Opc1;
15580 if (IntrWithRoundingModeOpcode != 0) {
15581 SDValue Rnd = Op.getOperand(5);
15582 if (cast<ConstantSDNode>(Rnd)->getZExtValue() !=
15583 X86::STATIC_ROUNDING::CUR_DIRECTION)
15584 return getVectorMaskingNode(DAG.getNode(IntrWithRoundingModeOpcode,
15585 dl, Op.getValueType(),
15586 Src1, Src2, Src3, Rnd),
15587 Mask, PassThru, Subtarget, DAG);
15589 return getVectorMaskingNode(DAG.getNode(IntrData->Opc0,
15590 dl, Op.getValueType(),
15592 Mask, PassThru, Subtarget, DAG);
15595 case CMP_MASK_CC: {
15596 // Comparison intrinsics with masks.
15597 // Example of transformation:
15598 // (i8 (int_x86_avx512_mask_pcmpeq_q_128
15599 // (v2i64 %a), (v2i64 %b), (i8 %mask))) ->
15601 // (v8i1 (insert_subvector undef,
15602 // (v2i1 (and (PCMPEQM %a, %b),
15603 // (extract_subvector
15604 // (v8i1 (bitcast %mask)), 0))), 0))))
15605 EVT VT = Op.getOperand(1).getValueType();
15606 EVT MaskVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
15607 VT.getVectorNumElements());
15608 SDValue Mask = Op.getOperand((IntrData->Type == CMP_MASK_CC) ? 4 : 3);
15609 EVT BitcastVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
15610 Mask.getValueType().getSizeInBits());
15612 if (IntrData->Type == CMP_MASK_CC) {
15613 SDValue CC = Op.getOperand(3);
15614 CC = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, CC);
15615 // We specify 2 possible opcodes for intrinsics with rounding modes.
15616 // First, we check if the intrinsic may have non-default rounding mode,
15617 // (IntrData->Opc1 != 0), then we check the rounding mode operand.
15618 if (IntrData->Opc1 != 0) {
15619 SDValue Rnd = Op.getOperand(5);
15620 if (cast<ConstantSDNode>(Rnd)->getZExtValue() !=
15621 X86::STATIC_ROUNDING::CUR_DIRECTION)
15622 Cmp = DAG.getNode(IntrData->Opc1, dl, MaskVT, Op.getOperand(1),
15623 Op.getOperand(2), CC, Rnd);
15625 //default rounding mode
15627 Cmp = DAG.getNode(IntrData->Opc0, dl, MaskVT, Op.getOperand(1),
15628 Op.getOperand(2), CC);
15631 assert(IntrData->Type == CMP_MASK && "Unexpected intrinsic type!");
15632 Cmp = DAG.getNode(IntrData->Opc0, dl, MaskVT, Op.getOperand(1),
15635 SDValue CmpMask = getVectorMaskingNode(Cmp, Mask,
15636 DAG.getTargetConstant(0, dl,
15639 SDValue Res = DAG.getNode(ISD::INSERT_SUBVECTOR, dl, BitcastVT,
15640 DAG.getUNDEF(BitcastVT), CmpMask,
15641 DAG.getIntPtrConstant(0, dl));
15642 return DAG.getBitcast(Op.getValueType(), Res);
15644 case COMI: { // Comparison intrinsics
15645 ISD::CondCode CC = (ISD::CondCode)IntrData->Opc1;
15646 SDValue LHS = Op.getOperand(1);
15647 SDValue RHS = Op.getOperand(2);
15648 unsigned X86CC = TranslateX86CC(CC, dl, true, LHS, RHS, DAG);
15649 assert(X86CC != X86::COND_INVALID && "Unexpected illegal condition!");
15650 SDValue Cond = DAG.getNode(IntrData->Opc0, dl, MVT::i32, LHS, RHS);
15651 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
15652 DAG.getConstant(X86CC, dl, MVT::i8), Cond);
15653 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
15656 return getTargetVShiftNode(IntrData->Opc0, dl, Op.getSimpleValueType(),
15657 Op.getOperand(1), Op.getOperand(2), DAG);
15659 return getVectorMaskingNode(getTargetVShiftNode(IntrData->Opc0, dl,
15660 Op.getSimpleValueType(),
15662 Op.getOperand(2), DAG),
15663 Op.getOperand(4), Op.getOperand(3), Subtarget,
15665 case COMPRESS_EXPAND_IN_REG: {
15666 SDValue Mask = Op.getOperand(3);
15667 SDValue DataToCompress = Op.getOperand(1);
15668 SDValue PassThru = Op.getOperand(2);
15669 if (isAllOnes(Mask)) // return data as is
15670 return Op.getOperand(1);
15672 return getVectorMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT,
15674 Mask, PassThru, Subtarget, DAG);
15677 SDValue Mask = Op.getOperand(3);
15678 EVT VT = Op.getValueType();
15679 EVT MaskVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
15680 VT.getVectorNumElements());
15681 EVT BitcastVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
15682 Mask.getValueType().getSizeInBits());
15684 SDValue VMask = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MaskVT,
15685 DAG.getBitcast(BitcastVT, Mask),
15686 DAG.getIntPtrConstant(0, dl));
15687 return DAG.getNode(IntrData->Opc0, dl, VT, VMask, Op.getOperand(1),
15696 default: return SDValue(); // Don't custom lower most intrinsics.
15698 case Intrinsic::x86_avx2_permd:
15699 case Intrinsic::x86_avx2_permps:
15700 // Operands intentionally swapped. Mask is last operand to intrinsic,
15701 // but second operand for node/instruction.
15702 return DAG.getNode(X86ISD::VPERMV, dl, Op.getValueType(),
15703 Op.getOperand(2), Op.getOperand(1));
15705 // ptest and testp intrinsics. The intrinsic these come from are designed to
15706 // return an integer value, not just an instruction so lower it to the ptest
15707 // or testp pattern and a setcc for the result.
15708 case Intrinsic::x86_sse41_ptestz:
15709 case Intrinsic::x86_sse41_ptestc:
15710 case Intrinsic::x86_sse41_ptestnzc:
15711 case Intrinsic::x86_avx_ptestz_256:
15712 case Intrinsic::x86_avx_ptestc_256:
15713 case Intrinsic::x86_avx_ptestnzc_256:
15714 case Intrinsic::x86_avx_vtestz_ps:
15715 case Intrinsic::x86_avx_vtestc_ps:
15716 case Intrinsic::x86_avx_vtestnzc_ps:
15717 case Intrinsic::x86_avx_vtestz_pd:
15718 case Intrinsic::x86_avx_vtestc_pd:
15719 case Intrinsic::x86_avx_vtestnzc_pd:
15720 case Intrinsic::x86_avx_vtestz_ps_256:
15721 case Intrinsic::x86_avx_vtestc_ps_256:
15722 case Intrinsic::x86_avx_vtestnzc_ps_256:
15723 case Intrinsic::x86_avx_vtestz_pd_256:
15724 case Intrinsic::x86_avx_vtestc_pd_256:
15725 case Intrinsic::x86_avx_vtestnzc_pd_256: {
15726 bool IsTestPacked = false;
15729 default: llvm_unreachable("Bad fallthrough in Intrinsic lowering.");
15730 case Intrinsic::x86_avx_vtestz_ps:
15731 case Intrinsic::x86_avx_vtestz_pd:
15732 case Intrinsic::x86_avx_vtestz_ps_256:
15733 case Intrinsic::x86_avx_vtestz_pd_256:
15734 IsTestPacked = true; // Fallthrough
15735 case Intrinsic::x86_sse41_ptestz:
15736 case Intrinsic::x86_avx_ptestz_256:
15738 X86CC = X86::COND_E;
15740 case Intrinsic::x86_avx_vtestc_ps:
15741 case Intrinsic::x86_avx_vtestc_pd:
15742 case Intrinsic::x86_avx_vtestc_ps_256:
15743 case Intrinsic::x86_avx_vtestc_pd_256:
15744 IsTestPacked = true; // Fallthrough
15745 case Intrinsic::x86_sse41_ptestc:
15746 case Intrinsic::x86_avx_ptestc_256:
15748 X86CC = X86::COND_B;
15750 case Intrinsic::x86_avx_vtestnzc_ps:
15751 case Intrinsic::x86_avx_vtestnzc_pd:
15752 case Intrinsic::x86_avx_vtestnzc_ps_256:
15753 case Intrinsic::x86_avx_vtestnzc_pd_256:
15754 IsTestPacked = true; // Fallthrough
15755 case Intrinsic::x86_sse41_ptestnzc:
15756 case Intrinsic::x86_avx_ptestnzc_256:
15758 X86CC = X86::COND_A;
15762 SDValue LHS = Op.getOperand(1);
15763 SDValue RHS = Op.getOperand(2);
15764 unsigned TestOpc = IsTestPacked ? X86ISD::TESTP : X86ISD::PTEST;
15765 SDValue Test = DAG.getNode(TestOpc, dl, MVT::i32, LHS, RHS);
15766 SDValue CC = DAG.getConstant(X86CC, dl, MVT::i8);
15767 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8, CC, Test);
15768 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
15770 case Intrinsic::x86_avx512_kortestz_w:
15771 case Intrinsic::x86_avx512_kortestc_w: {
15772 unsigned X86CC = (IntNo == Intrinsic::x86_avx512_kortestz_w)? X86::COND_E: X86::COND_B;
15773 SDValue LHS = DAG.getBitcast(MVT::v16i1, Op.getOperand(1));
15774 SDValue RHS = DAG.getBitcast(MVT::v16i1, Op.getOperand(2));
15775 SDValue CC = DAG.getConstant(X86CC, dl, MVT::i8);
15776 SDValue Test = DAG.getNode(X86ISD::KORTEST, dl, MVT::i32, LHS, RHS);
15777 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i1, CC, Test);
15778 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
15781 case Intrinsic::x86_sse42_pcmpistria128:
15782 case Intrinsic::x86_sse42_pcmpestria128:
15783 case Intrinsic::x86_sse42_pcmpistric128:
15784 case Intrinsic::x86_sse42_pcmpestric128:
15785 case Intrinsic::x86_sse42_pcmpistrio128:
15786 case Intrinsic::x86_sse42_pcmpestrio128:
15787 case Intrinsic::x86_sse42_pcmpistris128:
15788 case Intrinsic::x86_sse42_pcmpestris128:
15789 case Intrinsic::x86_sse42_pcmpistriz128:
15790 case Intrinsic::x86_sse42_pcmpestriz128: {
15794 default: llvm_unreachable("Impossible intrinsic"); // Can't reach here.
15795 case Intrinsic::x86_sse42_pcmpistria128:
15796 Opcode = X86ISD::PCMPISTRI;
15797 X86CC = X86::COND_A;
15799 case Intrinsic::x86_sse42_pcmpestria128:
15800 Opcode = X86ISD::PCMPESTRI;
15801 X86CC = X86::COND_A;
15803 case Intrinsic::x86_sse42_pcmpistric128:
15804 Opcode = X86ISD::PCMPISTRI;
15805 X86CC = X86::COND_B;
15807 case Intrinsic::x86_sse42_pcmpestric128:
15808 Opcode = X86ISD::PCMPESTRI;
15809 X86CC = X86::COND_B;
15811 case Intrinsic::x86_sse42_pcmpistrio128:
15812 Opcode = X86ISD::PCMPISTRI;
15813 X86CC = X86::COND_O;
15815 case Intrinsic::x86_sse42_pcmpestrio128:
15816 Opcode = X86ISD::PCMPESTRI;
15817 X86CC = X86::COND_O;
15819 case Intrinsic::x86_sse42_pcmpistris128:
15820 Opcode = X86ISD::PCMPISTRI;
15821 X86CC = X86::COND_S;
15823 case Intrinsic::x86_sse42_pcmpestris128:
15824 Opcode = X86ISD::PCMPESTRI;
15825 X86CC = X86::COND_S;
15827 case Intrinsic::x86_sse42_pcmpistriz128:
15828 Opcode = X86ISD::PCMPISTRI;
15829 X86CC = X86::COND_E;
15831 case Intrinsic::x86_sse42_pcmpestriz128:
15832 Opcode = X86ISD::PCMPESTRI;
15833 X86CC = X86::COND_E;
15836 SmallVector<SDValue, 5> NewOps(Op->op_begin()+1, Op->op_end());
15837 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::i32);
15838 SDValue PCMP = DAG.getNode(Opcode, dl, VTs, NewOps);
15839 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
15840 DAG.getConstant(X86CC, dl, MVT::i8),
15841 SDValue(PCMP.getNode(), 1));
15842 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
15845 case Intrinsic::x86_sse42_pcmpistri128:
15846 case Intrinsic::x86_sse42_pcmpestri128: {
15848 if (IntNo == Intrinsic::x86_sse42_pcmpistri128)
15849 Opcode = X86ISD::PCMPISTRI;
15851 Opcode = X86ISD::PCMPESTRI;
15853 SmallVector<SDValue, 5> NewOps(Op->op_begin()+1, Op->op_end());
15854 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::i32);
15855 return DAG.getNode(Opcode, dl, VTs, NewOps);
15858 case Intrinsic::x86_seh_lsda: {
15859 // Compute the symbol for the LSDA. We know it'll get emitted later.
15860 MachineFunction &MF = DAG.getMachineFunction();
15861 SDValue Op1 = Op.getOperand(1);
15862 auto *Fn = cast<Function>(cast<GlobalAddressSDNode>(Op1)->getGlobal());
15863 MCSymbol *LSDASym = MF.getMMI().getContext().getOrCreateLSDASymbol(
15864 GlobalValue::getRealLinkageName(Fn->getName()));
15866 // Generate a simple absolute symbol reference. This intrinsic is only
15867 // supported on 32-bit Windows, which isn't PIC.
15868 SDValue Result = DAG.getMCSymbol(LSDASym, VT);
15869 return DAG.getNode(X86ISD::Wrapper, dl, VT, Result);
15872 case Intrinsic::x86_seh_recoverfp: {
15873 SDValue FnOp = Op.getOperand(1);
15874 SDValue IncomingFPOp = Op.getOperand(2);
15875 GlobalAddressSDNode *GSD = dyn_cast<GlobalAddressSDNode>(FnOp);
15876 auto *Fn = dyn_cast_or_null<Function>(GSD ? GSD->getGlobal() : nullptr);
15878 report_fatal_error(
15879 "llvm.x86.seh.recoverfp must take a function as the first argument");
15880 return recoverFramePointer(DAG, Fn, IncomingFPOp);
15883 case Intrinsic::localaddress: {
15884 // Returns one of the stack, base, or frame pointer registers, depending on
15885 // which is used to reference local variables.
15886 MachineFunction &MF = DAG.getMachineFunction();
15887 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
15889 if (RegInfo->hasBasePointer(MF))
15890 Reg = RegInfo->getBaseRegister();
15891 else // This function handles the SP or FP case.
15892 Reg = RegInfo->getPtrSizedFrameRegister(MF);
15893 return DAG.getCopyFromReg(DAG.getEntryNode(), dl, Reg, VT);
15898 static SDValue getGatherNode(unsigned Opc, SDValue Op, SelectionDAG &DAG,
15899 SDValue Src, SDValue Mask, SDValue Base,
15900 SDValue Index, SDValue ScaleOp, SDValue Chain,
15901 const X86Subtarget * Subtarget) {
15903 ConstantSDNode *C = dyn_cast<ConstantSDNode>(ScaleOp);
15905 llvm_unreachable("Invalid scale type");
15906 unsigned ScaleVal = C->getZExtValue();
15907 if (ScaleVal > 2 && ScaleVal != 4 && ScaleVal != 8)
15908 llvm_unreachable("Valid scale values are 1, 2, 4, 8");
15910 SDValue Scale = DAG.getTargetConstant(C->getZExtValue(), dl, MVT::i8);
15911 EVT MaskVT = MVT::getVectorVT(MVT::i1,
15912 Index.getSimpleValueType().getVectorNumElements());
15914 ConstantSDNode *MaskC = dyn_cast<ConstantSDNode>(Mask);
15916 MaskInReg = DAG.getTargetConstant(MaskC->getSExtValue(), dl, MaskVT);
15918 EVT BitcastVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
15919 Mask.getValueType().getSizeInBits());
15921 // In case when MaskVT equals v2i1 or v4i1, low 2 or 4 elements
15922 // are extracted by EXTRACT_SUBVECTOR.
15923 MaskInReg = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MaskVT,
15924 DAG.getBitcast(BitcastVT, Mask),
15925 DAG.getIntPtrConstant(0, dl));
15927 SDVTList VTs = DAG.getVTList(Op.getValueType(), MaskVT, MVT::Other);
15928 SDValue Disp = DAG.getTargetConstant(0, dl, MVT::i32);
15929 SDValue Segment = DAG.getRegister(0, MVT::i32);
15930 if (Src.getOpcode() == ISD::UNDEF)
15931 Src = getZeroVector(Op.getValueType(), Subtarget, DAG, dl);
15932 SDValue Ops[] = {Src, MaskInReg, Base, Scale, Index, Disp, Segment, Chain};
15933 SDNode *Res = DAG.getMachineNode(Opc, dl, VTs, Ops);
15934 SDValue RetOps[] = { SDValue(Res, 0), SDValue(Res, 2) };
15935 return DAG.getMergeValues(RetOps, dl);
15938 static SDValue getScatterNode(unsigned Opc, SDValue Op, SelectionDAG &DAG,
15939 SDValue Src, SDValue Mask, SDValue Base,
15940 SDValue Index, SDValue ScaleOp, SDValue Chain) {
15942 ConstantSDNode *C = dyn_cast<ConstantSDNode>(ScaleOp);
15944 llvm_unreachable("Invalid scale type");
15945 unsigned ScaleVal = C->getZExtValue();
15946 if (ScaleVal > 2 && ScaleVal != 4 && ScaleVal != 8)
15947 llvm_unreachable("Valid scale values are 1, 2, 4, 8");
15949 SDValue Scale = DAG.getTargetConstant(C->getZExtValue(), dl, MVT::i8);
15950 SDValue Disp = DAG.getTargetConstant(0, dl, MVT::i32);
15951 SDValue Segment = DAG.getRegister(0, MVT::i32);
15952 EVT MaskVT = MVT::getVectorVT(MVT::i1,
15953 Index.getSimpleValueType().getVectorNumElements());
15955 ConstantSDNode *MaskC = dyn_cast<ConstantSDNode>(Mask);
15957 MaskInReg = DAG.getTargetConstant(MaskC->getSExtValue(), dl, MaskVT);
15959 EVT BitcastVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
15960 Mask.getValueType().getSizeInBits());
15962 // In case when MaskVT equals v2i1 or v4i1, low 2 or 4 elements
15963 // are extracted by EXTRACT_SUBVECTOR.
15964 MaskInReg = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MaskVT,
15965 DAG.getBitcast(BitcastVT, Mask),
15966 DAG.getIntPtrConstant(0, dl));
15968 SDVTList VTs = DAG.getVTList(MaskVT, MVT::Other);
15969 SDValue Ops[] = {Base, Scale, Index, Disp, Segment, MaskInReg, Src, Chain};
15970 SDNode *Res = DAG.getMachineNode(Opc, dl, VTs, Ops);
15971 return SDValue(Res, 1);
15974 static SDValue getPrefetchNode(unsigned Opc, SDValue Op, SelectionDAG &DAG,
15975 SDValue Mask, SDValue Base, SDValue Index,
15976 SDValue ScaleOp, SDValue Chain) {
15978 ConstantSDNode *C = dyn_cast<ConstantSDNode>(ScaleOp);
15979 assert(C && "Invalid scale type");
15980 SDValue Scale = DAG.getTargetConstant(C->getZExtValue(), dl, MVT::i8);
15981 SDValue Disp = DAG.getTargetConstant(0, dl, MVT::i32);
15982 SDValue Segment = DAG.getRegister(0, MVT::i32);
15984 MVT::getVectorVT(MVT::i1, Index.getSimpleValueType().getVectorNumElements());
15986 ConstantSDNode *MaskC = dyn_cast<ConstantSDNode>(Mask);
15988 MaskInReg = DAG.getTargetConstant(MaskC->getSExtValue(), dl, MaskVT);
15990 MaskInReg = DAG.getBitcast(MaskVT, Mask);
15991 //SDVTList VTs = DAG.getVTList(MVT::Other);
15992 SDValue Ops[] = {MaskInReg, Base, Scale, Index, Disp, Segment, Chain};
15993 SDNode *Res = DAG.getMachineNode(Opc, dl, MVT::Other, Ops);
15994 return SDValue(Res, 0);
15997 // getReadPerformanceCounter - Handles the lowering of builtin intrinsics that
15998 // read performance monitor counters (x86_rdpmc).
15999 static void getReadPerformanceCounter(SDNode *N, SDLoc DL,
16000 SelectionDAG &DAG, const X86Subtarget *Subtarget,
16001 SmallVectorImpl<SDValue> &Results) {
16002 assert(N->getNumOperands() == 3 && "Unexpected number of operands!");
16003 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Glue);
16006 // The ECX register is used to select the index of the performance counter
16008 SDValue Chain = DAG.getCopyToReg(N->getOperand(0), DL, X86::ECX,
16010 SDValue rd = DAG.getNode(X86ISD::RDPMC_DAG, DL, Tys, Chain);
16012 // Reads the content of a 64-bit performance counter and returns it in the
16013 // registers EDX:EAX.
16014 if (Subtarget->is64Bit()) {
16015 LO = DAG.getCopyFromReg(rd, DL, X86::RAX, MVT::i64, rd.getValue(1));
16016 HI = DAG.getCopyFromReg(LO.getValue(1), DL, X86::RDX, MVT::i64,
16019 LO = DAG.getCopyFromReg(rd, DL, X86::EAX, MVT::i32, rd.getValue(1));
16020 HI = DAG.getCopyFromReg(LO.getValue(1), DL, X86::EDX, MVT::i32,
16023 Chain = HI.getValue(1);
16025 if (Subtarget->is64Bit()) {
16026 // The EAX register is loaded with the low-order 32 bits. The EDX register
16027 // is loaded with the supported high-order bits of the counter.
16028 SDValue Tmp = DAG.getNode(ISD::SHL, DL, MVT::i64, HI,
16029 DAG.getConstant(32, DL, MVT::i8));
16030 Results.push_back(DAG.getNode(ISD::OR, DL, MVT::i64, LO, Tmp));
16031 Results.push_back(Chain);
16035 // Use a buildpair to merge the two 32-bit values into a 64-bit one.
16036 SDValue Ops[] = { LO, HI };
16037 SDValue Pair = DAG.getNode(ISD::BUILD_PAIR, DL, MVT::i64, Ops);
16038 Results.push_back(Pair);
16039 Results.push_back(Chain);
16042 // getReadTimeStampCounter - Handles the lowering of builtin intrinsics that
16043 // read the time stamp counter (x86_rdtsc and x86_rdtscp). This function is
16044 // also used to custom lower READCYCLECOUNTER nodes.
16045 static void getReadTimeStampCounter(SDNode *N, SDLoc DL, unsigned Opcode,
16046 SelectionDAG &DAG, const X86Subtarget *Subtarget,
16047 SmallVectorImpl<SDValue> &Results) {
16048 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Glue);
16049 SDValue rd = DAG.getNode(Opcode, DL, Tys, N->getOperand(0));
16052 // The processor's time-stamp counter (a 64-bit MSR) is stored into the
16053 // EDX:EAX registers. EDX is loaded with the high-order 32 bits of the MSR
16054 // and the EAX register is loaded with the low-order 32 bits.
16055 if (Subtarget->is64Bit()) {
16056 LO = DAG.getCopyFromReg(rd, DL, X86::RAX, MVT::i64, rd.getValue(1));
16057 HI = DAG.getCopyFromReg(LO.getValue(1), DL, X86::RDX, MVT::i64,
16060 LO = DAG.getCopyFromReg(rd, DL, X86::EAX, MVT::i32, rd.getValue(1));
16061 HI = DAG.getCopyFromReg(LO.getValue(1), DL, X86::EDX, MVT::i32,
16064 SDValue Chain = HI.getValue(1);
16066 if (Opcode == X86ISD::RDTSCP_DAG) {
16067 assert(N->getNumOperands() == 3 && "Unexpected number of operands!");
16069 // Instruction RDTSCP loads the IA32:TSC_AUX_MSR (address C000_0103H) into
16070 // the ECX register. Add 'ecx' explicitly to the chain.
16071 SDValue ecx = DAG.getCopyFromReg(Chain, DL, X86::ECX, MVT::i32,
16073 // Explicitly store the content of ECX at the location passed in input
16074 // to the 'rdtscp' intrinsic.
16075 Chain = DAG.getStore(ecx.getValue(1), DL, ecx, N->getOperand(2),
16076 MachinePointerInfo(), false, false, 0);
16079 if (Subtarget->is64Bit()) {
16080 // The EDX register is loaded with the high-order 32 bits of the MSR, and
16081 // the EAX register is loaded with the low-order 32 bits.
16082 SDValue Tmp = DAG.getNode(ISD::SHL, DL, MVT::i64, HI,
16083 DAG.getConstant(32, DL, MVT::i8));
16084 Results.push_back(DAG.getNode(ISD::OR, DL, MVT::i64, LO, Tmp));
16085 Results.push_back(Chain);
16089 // Use a buildpair to merge the two 32-bit values into a 64-bit one.
16090 SDValue Ops[] = { LO, HI };
16091 SDValue Pair = DAG.getNode(ISD::BUILD_PAIR, DL, MVT::i64, Ops);
16092 Results.push_back(Pair);
16093 Results.push_back(Chain);
16096 static SDValue LowerREADCYCLECOUNTER(SDValue Op, const X86Subtarget *Subtarget,
16097 SelectionDAG &DAG) {
16098 SmallVector<SDValue, 2> Results;
16100 getReadTimeStampCounter(Op.getNode(), DL, X86ISD::RDTSC_DAG, DAG, Subtarget,
16102 return DAG.getMergeValues(Results, DL);
16105 static SDValue LowerSEHRESTOREFRAME(SDValue Op, const X86Subtarget *Subtarget,
16106 SelectionDAG &DAG) {
16107 MachineFunction &MF = DAG.getMachineFunction();
16108 const Function *Fn = MF.getFunction();
16110 SDValue Chain = Op.getOperand(0);
16112 assert(Subtarget->getFrameLowering()->hasFP(MF) &&
16113 "using llvm.x86.seh.restoreframe requires a frame pointer");
16115 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
16116 MVT VT = TLI.getPointerTy(DAG.getDataLayout());
16118 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
16119 unsigned FrameReg =
16120 RegInfo->getPtrSizedFrameRegister(DAG.getMachineFunction());
16121 unsigned SPReg = RegInfo->getStackRegister();
16122 unsigned SlotSize = RegInfo->getSlotSize();
16124 // Get incoming EBP.
16125 SDValue IncomingEBP =
16126 DAG.getCopyFromReg(Chain, dl, FrameReg, VT);
16128 // SP is saved in the first field of every registration node, so load
16129 // [EBP-RegNodeSize] into SP.
16130 int RegNodeSize = getSEHRegistrationNodeSize(Fn);
16131 SDValue SPAddr = DAG.getNode(ISD::ADD, dl, VT, IncomingEBP,
16132 DAG.getConstant(-RegNodeSize, dl, VT));
16134 DAG.getLoad(VT, dl, Chain, SPAddr, MachinePointerInfo(), false, false,
16135 false, VT.getScalarSizeInBits() / 8);
16136 Chain = DAG.getCopyToReg(Chain, dl, SPReg, NewSP);
16138 if (!RegInfo->needsStackRealignment(MF)) {
16139 // Adjust EBP to point back to the original frame position.
16140 SDValue NewFP = recoverFramePointer(DAG, Fn, IncomingEBP);
16141 Chain = DAG.getCopyToReg(Chain, dl, FrameReg, NewFP);
16143 assert(RegInfo->hasBasePointer(MF) &&
16144 "functions with Win32 EH must use frame or base pointer register");
16146 // Reload the base pointer (ESI) with the adjusted incoming EBP.
16147 SDValue NewBP = recoverFramePointer(DAG, Fn, IncomingEBP);
16148 Chain = DAG.getCopyToReg(Chain, dl, RegInfo->getBaseRegister(), NewBP);
16150 // Reload the spilled EBP value, now that the stack and base pointers are
16152 X86MachineFunctionInfo *X86FI = MF.getInfo<X86MachineFunctionInfo>();
16153 X86FI->setHasSEHFramePtrSave(true);
16154 int FI = MF.getFrameInfo()->CreateSpillStackObject(SlotSize, SlotSize);
16155 X86FI->setSEHFramePtrSaveIndex(FI);
16156 SDValue NewFP = DAG.getLoad(VT, dl, Chain, DAG.getFrameIndex(FI, VT),
16157 MachinePointerInfo(), false, false, false,
16158 VT.getScalarSizeInBits() / 8);
16159 Chain = DAG.getCopyToReg(NewFP, dl, FrameReg, NewFP);
16165 /// \brief Lower intrinsics for TRUNCATE_TO_MEM case
16166 /// return truncate Store/MaskedStore Node
16167 static SDValue LowerINTRINSIC_TRUNCATE_TO_MEM(const SDValue & Op,
16171 SDValue Mask = Op.getOperand(4);
16172 SDValue DataToTruncate = Op.getOperand(3);
16173 SDValue Addr = Op.getOperand(2);
16174 SDValue Chain = Op.getOperand(0);
16176 EVT VT = DataToTruncate.getValueType();
16177 EVT SVT = EVT::getVectorVT(*DAG.getContext(),
16178 ElementType, VT.getVectorNumElements());
16180 if (isAllOnes(Mask)) // return just a truncate store
16181 return DAG.getTruncStore(Chain, dl, DataToTruncate, Addr,
16182 MachinePointerInfo(), SVT, false, false,
16183 SVT.getScalarSizeInBits()/8);
16185 EVT MaskVT = EVT::getVectorVT(*DAG.getContext(),
16186 MVT::i1, VT.getVectorNumElements());
16187 EVT BitcastVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
16188 Mask.getValueType().getSizeInBits());
16189 // In case when MaskVT equals v2i1 or v4i1, low 2 or 4 elements
16190 // are extracted by EXTRACT_SUBVECTOR.
16191 SDValue VMask = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MaskVT,
16192 DAG.getBitcast(BitcastVT, Mask),
16193 DAG.getIntPtrConstant(0, dl));
16195 MachineMemOperand *MMO = DAG.getMachineFunction().
16196 getMachineMemOperand(MachinePointerInfo(),
16197 MachineMemOperand::MOStore, SVT.getStoreSize(),
16198 SVT.getScalarSizeInBits()/8);
16200 return DAG.getMaskedStore(Chain, dl, DataToTruncate, Addr,
16201 VMask, SVT, MMO, true);
16204 static SDValue LowerINTRINSIC_W_CHAIN(SDValue Op, const X86Subtarget *Subtarget,
16205 SelectionDAG &DAG) {
16206 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
16208 const IntrinsicData* IntrData = getIntrinsicWithChain(IntNo);
16210 if (IntNo == llvm::Intrinsic::x86_seh_restoreframe)
16211 return LowerSEHRESTOREFRAME(Op, Subtarget, DAG);
16216 switch(IntrData->Type) {
16218 llvm_unreachable("Unknown Intrinsic Type");
16222 // Emit the node with the right value type.
16223 SDVTList VTs = DAG.getVTList(Op->getValueType(0), MVT::Glue, MVT::Other);
16224 SDValue Result = DAG.getNode(IntrData->Opc0, dl, VTs, Op.getOperand(0));
16226 // If the value returned by RDRAND/RDSEED was valid (CF=1), return 1.
16227 // Otherwise return the value from Rand, which is always 0, casted to i32.
16228 SDValue Ops[] = { DAG.getZExtOrTrunc(Result, dl, Op->getValueType(1)),
16229 DAG.getConstant(1, dl, Op->getValueType(1)),
16230 DAG.getConstant(X86::COND_B, dl, MVT::i32),
16231 SDValue(Result.getNode(), 1) };
16232 SDValue isValid = DAG.getNode(X86ISD::CMOV, dl,
16233 DAG.getVTList(Op->getValueType(1), MVT::Glue),
16236 // Return { result, isValid, chain }.
16237 return DAG.getNode(ISD::MERGE_VALUES, dl, Op->getVTList(), Result, isValid,
16238 SDValue(Result.getNode(), 2));
16241 //gather(v1, mask, index, base, scale);
16242 SDValue Chain = Op.getOperand(0);
16243 SDValue Src = Op.getOperand(2);
16244 SDValue Base = Op.getOperand(3);
16245 SDValue Index = Op.getOperand(4);
16246 SDValue Mask = Op.getOperand(5);
16247 SDValue Scale = Op.getOperand(6);
16248 return getGatherNode(IntrData->Opc0, Op, DAG, Src, Mask, Base, Index, Scale,
16252 //scatter(base, mask, index, v1, scale);
16253 SDValue Chain = Op.getOperand(0);
16254 SDValue Base = Op.getOperand(2);
16255 SDValue Mask = Op.getOperand(3);
16256 SDValue Index = Op.getOperand(4);
16257 SDValue Src = Op.getOperand(5);
16258 SDValue Scale = Op.getOperand(6);
16259 return getScatterNode(IntrData->Opc0, Op, DAG, Src, Mask, Base, Index,
16263 SDValue Hint = Op.getOperand(6);
16264 unsigned HintVal = cast<ConstantSDNode>(Hint)->getZExtValue();
16265 assert(HintVal < 2 && "Wrong prefetch hint in intrinsic: should be 0 or 1");
16266 unsigned Opcode = (HintVal ? IntrData->Opc1 : IntrData->Opc0);
16267 SDValue Chain = Op.getOperand(0);
16268 SDValue Mask = Op.getOperand(2);
16269 SDValue Index = Op.getOperand(3);
16270 SDValue Base = Op.getOperand(4);
16271 SDValue Scale = Op.getOperand(5);
16272 return getPrefetchNode(Opcode, Op, DAG, Mask, Base, Index, Scale, Chain);
16274 // Read Time Stamp Counter (RDTSC) and Processor ID (RDTSCP).
16276 SmallVector<SDValue, 2> Results;
16277 getReadTimeStampCounter(Op.getNode(), dl, IntrData->Opc0, DAG, Subtarget,
16279 return DAG.getMergeValues(Results, dl);
16281 // Read Performance Monitoring Counters.
16283 SmallVector<SDValue, 2> Results;
16284 getReadPerformanceCounter(Op.getNode(), dl, DAG, Subtarget, Results);
16285 return DAG.getMergeValues(Results, dl);
16287 // XTEST intrinsics.
16289 SDVTList VTs = DAG.getVTList(Op->getValueType(0), MVT::Other);
16290 SDValue InTrans = DAG.getNode(IntrData->Opc0, dl, VTs, Op.getOperand(0));
16291 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
16292 DAG.getConstant(X86::COND_NE, dl, MVT::i8),
16294 SDValue Ret = DAG.getNode(ISD::ZERO_EXTEND, dl, Op->getValueType(0), SetCC);
16295 return DAG.getNode(ISD::MERGE_VALUES, dl, Op->getVTList(),
16296 Ret, SDValue(InTrans.getNode(), 1));
16300 SmallVector<SDValue, 2> Results;
16301 SDVTList CFVTs = DAG.getVTList(Op->getValueType(0), MVT::Other);
16302 SDVTList VTs = DAG.getVTList(Op.getOperand(3)->getValueType(0), MVT::Other);
16303 SDValue GenCF = DAG.getNode(X86ISD::ADD, dl, CFVTs, Op.getOperand(2),
16304 DAG.getConstant(-1, dl, MVT::i8));
16305 SDValue Res = DAG.getNode(IntrData->Opc0, dl, VTs, Op.getOperand(3),
16306 Op.getOperand(4), GenCF.getValue(1));
16307 SDValue Store = DAG.getStore(Op.getOperand(0), dl, Res.getValue(0),
16308 Op.getOperand(5), MachinePointerInfo(),
16310 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
16311 DAG.getConstant(X86::COND_B, dl, MVT::i8),
16313 Results.push_back(SetCC);
16314 Results.push_back(Store);
16315 return DAG.getMergeValues(Results, dl);
16317 case COMPRESS_TO_MEM: {
16319 SDValue Mask = Op.getOperand(4);
16320 SDValue DataToCompress = Op.getOperand(3);
16321 SDValue Addr = Op.getOperand(2);
16322 SDValue Chain = Op.getOperand(0);
16324 EVT VT = DataToCompress.getValueType();
16325 if (isAllOnes(Mask)) // return just a store
16326 return DAG.getStore(Chain, dl, DataToCompress, Addr,
16327 MachinePointerInfo(), false, false,
16328 VT.getScalarSizeInBits()/8);
16330 SDValue Compressed =
16331 getVectorMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT, DataToCompress),
16332 Mask, DAG.getUNDEF(VT), Subtarget, DAG);
16333 return DAG.getStore(Chain, dl, Compressed, Addr,
16334 MachinePointerInfo(), false, false,
16335 VT.getScalarSizeInBits()/8);
16337 case TRUNCATE_TO_MEM_VI8:
16338 return LowerINTRINSIC_TRUNCATE_TO_MEM(Op, DAG, MVT::i8);
16339 case TRUNCATE_TO_MEM_VI16:
16340 return LowerINTRINSIC_TRUNCATE_TO_MEM(Op, DAG, MVT::i16);
16341 case TRUNCATE_TO_MEM_VI32:
16342 return LowerINTRINSIC_TRUNCATE_TO_MEM(Op, DAG, MVT::i32);
16343 case EXPAND_FROM_MEM: {
16345 SDValue Mask = Op.getOperand(4);
16346 SDValue PassThru = Op.getOperand(3);
16347 SDValue Addr = Op.getOperand(2);
16348 SDValue Chain = Op.getOperand(0);
16349 EVT VT = Op.getValueType();
16351 if (isAllOnes(Mask)) // return just a load
16352 return DAG.getLoad(VT, dl, Chain, Addr, MachinePointerInfo(), false, false,
16353 false, VT.getScalarSizeInBits()/8);
16355 SDValue DataToExpand = DAG.getLoad(VT, dl, Chain, Addr, MachinePointerInfo(),
16356 false, false, false,
16357 VT.getScalarSizeInBits()/8);
16359 SDValue Results[] = {
16360 getVectorMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT, DataToExpand),
16361 Mask, PassThru, Subtarget, DAG), Chain};
16362 return DAG.getMergeValues(Results, dl);
16367 SDValue X86TargetLowering::LowerRETURNADDR(SDValue Op,
16368 SelectionDAG &DAG) const {
16369 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
16370 MFI->setReturnAddressIsTaken(true);
16372 if (verifyReturnAddressArgumentIsConstant(Op, DAG))
16375 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
16377 EVT PtrVT = getPointerTy(DAG.getDataLayout());
16380 SDValue FrameAddr = LowerFRAMEADDR(Op, DAG);
16381 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
16382 SDValue Offset = DAG.getConstant(RegInfo->getSlotSize(), dl, PtrVT);
16383 return DAG.getLoad(PtrVT, dl, DAG.getEntryNode(),
16384 DAG.getNode(ISD::ADD, dl, PtrVT,
16385 FrameAddr, Offset),
16386 MachinePointerInfo(), false, false, false, 0);
16389 // Just load the return address.
16390 SDValue RetAddrFI = getReturnAddressFrameIndex(DAG);
16391 return DAG.getLoad(PtrVT, dl, DAG.getEntryNode(),
16392 RetAddrFI, MachinePointerInfo(), false, false, false, 0);
16395 SDValue X86TargetLowering::LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const {
16396 MachineFunction &MF = DAG.getMachineFunction();
16397 MachineFrameInfo *MFI = MF.getFrameInfo();
16398 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
16399 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
16400 EVT VT = Op.getValueType();
16402 MFI->setFrameAddressIsTaken(true);
16404 if (MF.getTarget().getMCAsmInfo()->usesWindowsCFI()) {
16405 // Depth > 0 makes no sense on targets which use Windows unwind codes. It
16406 // is not possible to crawl up the stack without looking at the unwind codes
16408 int FrameAddrIndex = FuncInfo->getFAIndex();
16409 if (!FrameAddrIndex) {
16410 // Set up a frame object for the return address.
16411 unsigned SlotSize = RegInfo->getSlotSize();
16412 FrameAddrIndex = MF.getFrameInfo()->CreateFixedObject(
16413 SlotSize, /*Offset=*/0, /*IsImmutable=*/false);
16414 FuncInfo->setFAIndex(FrameAddrIndex);
16416 return DAG.getFrameIndex(FrameAddrIndex, VT);
16419 unsigned FrameReg =
16420 RegInfo->getPtrSizedFrameRegister(DAG.getMachineFunction());
16421 SDLoc dl(Op); // FIXME probably not meaningful
16422 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
16423 assert(((FrameReg == X86::RBP && VT == MVT::i64) ||
16424 (FrameReg == X86::EBP && VT == MVT::i32)) &&
16425 "Invalid Frame Register!");
16426 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, VT);
16428 FrameAddr = DAG.getLoad(VT, dl, DAG.getEntryNode(), FrameAddr,
16429 MachinePointerInfo(),
16430 false, false, false, 0);
16434 // FIXME? Maybe this could be a TableGen attribute on some registers and
16435 // this table could be generated automatically from RegInfo.
16436 unsigned X86TargetLowering::getRegisterByName(const char* RegName, EVT VT,
16437 SelectionDAG &DAG) const {
16438 const TargetFrameLowering &TFI = *Subtarget->getFrameLowering();
16439 const MachineFunction &MF = DAG.getMachineFunction();
16441 unsigned Reg = StringSwitch<unsigned>(RegName)
16442 .Case("esp", X86::ESP)
16443 .Case("rsp", X86::RSP)
16444 .Case("ebp", X86::EBP)
16445 .Case("rbp", X86::RBP)
16448 if (Reg == X86::EBP || Reg == X86::RBP) {
16449 if (!TFI.hasFP(MF))
16450 report_fatal_error("register " + StringRef(RegName) +
16451 " is allocatable: function has no frame pointer");
16454 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
16455 unsigned FrameReg =
16456 RegInfo->getPtrSizedFrameRegister(DAG.getMachineFunction());
16457 assert((FrameReg == X86::EBP || FrameReg == X86::RBP) &&
16458 "Invalid Frame Register!");
16466 report_fatal_error("Invalid register name global variable");
16469 SDValue X86TargetLowering::LowerFRAME_TO_ARGS_OFFSET(SDValue Op,
16470 SelectionDAG &DAG) const {
16471 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
16472 return DAG.getIntPtrConstant(2 * RegInfo->getSlotSize(), SDLoc(Op));
16475 SDValue X86TargetLowering::LowerEH_RETURN(SDValue Op, SelectionDAG &DAG) const {
16476 SDValue Chain = Op.getOperand(0);
16477 SDValue Offset = Op.getOperand(1);
16478 SDValue Handler = Op.getOperand(2);
16481 EVT PtrVT = getPointerTy(DAG.getDataLayout());
16482 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
16483 unsigned FrameReg = RegInfo->getFrameRegister(DAG.getMachineFunction());
16484 assert(((FrameReg == X86::RBP && PtrVT == MVT::i64) ||
16485 (FrameReg == X86::EBP && PtrVT == MVT::i32)) &&
16486 "Invalid Frame Register!");
16487 SDValue Frame = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, PtrVT);
16488 unsigned StoreAddrReg = (PtrVT == MVT::i64) ? X86::RCX : X86::ECX;
16490 SDValue StoreAddr = DAG.getNode(ISD::ADD, dl, PtrVT, Frame,
16491 DAG.getIntPtrConstant(RegInfo->getSlotSize(),
16493 StoreAddr = DAG.getNode(ISD::ADD, dl, PtrVT, StoreAddr, Offset);
16494 Chain = DAG.getStore(Chain, dl, Handler, StoreAddr, MachinePointerInfo(),
16496 Chain = DAG.getCopyToReg(Chain, dl, StoreAddrReg, StoreAddr);
16498 return DAG.getNode(X86ISD::EH_RETURN, dl, MVT::Other, Chain,
16499 DAG.getRegister(StoreAddrReg, PtrVT));
16502 SDValue X86TargetLowering::lowerEH_SJLJ_SETJMP(SDValue Op,
16503 SelectionDAG &DAG) const {
16505 return DAG.getNode(X86ISD::EH_SJLJ_SETJMP, DL,
16506 DAG.getVTList(MVT::i32, MVT::Other),
16507 Op.getOperand(0), Op.getOperand(1));
16510 SDValue X86TargetLowering::lowerEH_SJLJ_LONGJMP(SDValue Op,
16511 SelectionDAG &DAG) const {
16513 return DAG.getNode(X86ISD::EH_SJLJ_LONGJMP, DL, MVT::Other,
16514 Op.getOperand(0), Op.getOperand(1));
16517 static SDValue LowerADJUST_TRAMPOLINE(SDValue Op, SelectionDAG &DAG) {
16518 return Op.getOperand(0);
16521 SDValue X86TargetLowering::LowerINIT_TRAMPOLINE(SDValue Op,
16522 SelectionDAG &DAG) const {
16523 SDValue Root = Op.getOperand(0);
16524 SDValue Trmp = Op.getOperand(1); // trampoline
16525 SDValue FPtr = Op.getOperand(2); // nested function
16526 SDValue Nest = Op.getOperand(3); // 'nest' parameter value
16529 const Value *TrmpAddr = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
16530 const TargetRegisterInfo *TRI = Subtarget->getRegisterInfo();
16532 if (Subtarget->is64Bit()) {
16533 SDValue OutChains[6];
16535 // Large code-model.
16536 const unsigned char JMP64r = 0xFF; // 64-bit jmp through register opcode.
16537 const unsigned char MOV64ri = 0xB8; // X86::MOV64ri opcode.
16539 const unsigned char N86R10 = TRI->getEncodingValue(X86::R10) & 0x7;
16540 const unsigned char N86R11 = TRI->getEncodingValue(X86::R11) & 0x7;
16542 const unsigned char REX_WB = 0x40 | 0x08 | 0x01; // REX prefix
16544 // Load the pointer to the nested function into R11.
16545 unsigned OpCode = ((MOV64ri | N86R11) << 8) | REX_WB; // movabsq r11
16546 SDValue Addr = Trmp;
16547 OutChains[0] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, dl, MVT::i16),
16548 Addr, MachinePointerInfo(TrmpAddr),
16551 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
16552 DAG.getConstant(2, dl, MVT::i64));
16553 OutChains[1] = DAG.getStore(Root, dl, FPtr, Addr,
16554 MachinePointerInfo(TrmpAddr, 2),
16557 // Load the 'nest' parameter value into R10.
16558 // R10 is specified in X86CallingConv.td
16559 OpCode = ((MOV64ri | N86R10) << 8) | REX_WB; // movabsq r10
16560 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
16561 DAG.getConstant(10, dl, MVT::i64));
16562 OutChains[2] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, dl, MVT::i16),
16563 Addr, MachinePointerInfo(TrmpAddr, 10),
16566 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
16567 DAG.getConstant(12, dl, MVT::i64));
16568 OutChains[3] = DAG.getStore(Root, dl, Nest, Addr,
16569 MachinePointerInfo(TrmpAddr, 12),
16572 // Jump to the nested function.
16573 OpCode = (JMP64r << 8) | REX_WB; // jmpq *...
16574 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
16575 DAG.getConstant(20, dl, MVT::i64));
16576 OutChains[4] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, dl, MVT::i16),
16577 Addr, MachinePointerInfo(TrmpAddr, 20),
16580 unsigned char ModRM = N86R11 | (4 << 3) | (3 << 6); // ...r11
16581 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
16582 DAG.getConstant(22, dl, MVT::i64));
16583 OutChains[5] = DAG.getStore(Root, dl, DAG.getConstant(ModRM, dl, MVT::i8),
16584 Addr, MachinePointerInfo(TrmpAddr, 22),
16587 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other, OutChains);
16589 const Function *Func =
16590 cast<Function>(cast<SrcValueSDNode>(Op.getOperand(5))->getValue());
16591 CallingConv::ID CC = Func->getCallingConv();
16596 llvm_unreachable("Unsupported calling convention");
16597 case CallingConv::C:
16598 case CallingConv::X86_StdCall: {
16599 // Pass 'nest' parameter in ECX.
16600 // Must be kept in sync with X86CallingConv.td
16601 NestReg = X86::ECX;
16603 // Check that ECX wasn't needed by an 'inreg' parameter.
16604 FunctionType *FTy = Func->getFunctionType();
16605 const AttributeSet &Attrs = Func->getAttributes();
16607 if (!Attrs.isEmpty() && !Func->isVarArg()) {
16608 unsigned InRegCount = 0;
16611 for (FunctionType::param_iterator I = FTy->param_begin(),
16612 E = FTy->param_end(); I != E; ++I, ++Idx)
16613 if (Attrs.hasAttribute(Idx, Attribute::InReg)) {
16614 auto &DL = DAG.getDataLayout();
16615 // FIXME: should only count parameters that are lowered to integers.
16616 InRegCount += (DL.getTypeSizeInBits(*I) + 31) / 32;
16619 if (InRegCount > 2) {
16620 report_fatal_error("Nest register in use - reduce number of inreg"
16626 case CallingConv::X86_FastCall:
16627 case CallingConv::X86_ThisCall:
16628 case CallingConv::Fast:
16629 // Pass 'nest' parameter in EAX.
16630 // Must be kept in sync with X86CallingConv.td
16631 NestReg = X86::EAX;
16635 SDValue OutChains[4];
16636 SDValue Addr, Disp;
16638 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
16639 DAG.getConstant(10, dl, MVT::i32));
16640 Disp = DAG.getNode(ISD::SUB, dl, MVT::i32, FPtr, Addr);
16642 // This is storing the opcode for MOV32ri.
16643 const unsigned char MOV32ri = 0xB8; // X86::MOV32ri's opcode byte.
16644 const unsigned char N86Reg = TRI->getEncodingValue(NestReg) & 0x7;
16645 OutChains[0] = DAG.getStore(Root, dl,
16646 DAG.getConstant(MOV32ri|N86Reg, dl, MVT::i8),
16647 Trmp, MachinePointerInfo(TrmpAddr),
16650 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
16651 DAG.getConstant(1, dl, MVT::i32));
16652 OutChains[1] = DAG.getStore(Root, dl, Nest, Addr,
16653 MachinePointerInfo(TrmpAddr, 1),
16656 const unsigned char JMP = 0xE9; // jmp <32bit dst> opcode.
16657 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
16658 DAG.getConstant(5, dl, MVT::i32));
16659 OutChains[2] = DAG.getStore(Root, dl, DAG.getConstant(JMP, dl, MVT::i8),
16660 Addr, MachinePointerInfo(TrmpAddr, 5),
16663 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
16664 DAG.getConstant(6, dl, MVT::i32));
16665 OutChains[3] = DAG.getStore(Root, dl, Disp, Addr,
16666 MachinePointerInfo(TrmpAddr, 6),
16669 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other, OutChains);
16673 SDValue X86TargetLowering::LowerFLT_ROUNDS_(SDValue Op,
16674 SelectionDAG &DAG) const {
16676 The rounding mode is in bits 11:10 of FPSR, and has the following
16678 00 Round to nearest
16683 FLT_ROUNDS, on the other hand, expects the following:
16690 To perform the conversion, we do:
16691 (((((FPSR & 0x800) >> 11) | ((FPSR & 0x400) >> 9)) + 1) & 3)
16694 MachineFunction &MF = DAG.getMachineFunction();
16695 const TargetFrameLowering &TFI = *Subtarget->getFrameLowering();
16696 unsigned StackAlignment = TFI.getStackAlignment();
16697 MVT VT = Op.getSimpleValueType();
16700 // Save FP Control Word to stack slot
16701 int SSFI = MF.getFrameInfo()->CreateStackObject(2, StackAlignment, false);
16702 SDValue StackSlot =
16703 DAG.getFrameIndex(SSFI, getPointerTy(DAG.getDataLayout()));
16705 MachineMemOperand *MMO =
16706 MF.getMachineMemOperand(MachinePointerInfo::getFixedStack(SSFI),
16707 MachineMemOperand::MOStore, 2, 2);
16709 SDValue Ops[] = { DAG.getEntryNode(), StackSlot };
16710 SDValue Chain = DAG.getMemIntrinsicNode(X86ISD::FNSTCW16m, DL,
16711 DAG.getVTList(MVT::Other),
16712 Ops, MVT::i16, MMO);
16714 // Load FP Control Word from stack slot
16715 SDValue CWD = DAG.getLoad(MVT::i16, DL, Chain, StackSlot,
16716 MachinePointerInfo(), false, false, false, 0);
16718 // Transform as necessary
16720 DAG.getNode(ISD::SRL, DL, MVT::i16,
16721 DAG.getNode(ISD::AND, DL, MVT::i16,
16722 CWD, DAG.getConstant(0x800, DL, MVT::i16)),
16723 DAG.getConstant(11, DL, MVT::i8));
16725 DAG.getNode(ISD::SRL, DL, MVT::i16,
16726 DAG.getNode(ISD::AND, DL, MVT::i16,
16727 CWD, DAG.getConstant(0x400, DL, MVT::i16)),
16728 DAG.getConstant(9, DL, MVT::i8));
16731 DAG.getNode(ISD::AND, DL, MVT::i16,
16732 DAG.getNode(ISD::ADD, DL, MVT::i16,
16733 DAG.getNode(ISD::OR, DL, MVT::i16, CWD1, CWD2),
16734 DAG.getConstant(1, DL, MVT::i16)),
16735 DAG.getConstant(3, DL, MVT::i16));
16737 return DAG.getNode((VT.getSizeInBits() < 16 ?
16738 ISD::TRUNCATE : ISD::ZERO_EXTEND), DL, VT, RetVal);
16741 static SDValue LowerCTLZ(SDValue Op, SelectionDAG &DAG) {
16742 MVT VT = Op.getSimpleValueType();
16744 unsigned NumBits = VT.getSizeInBits();
16747 Op = Op.getOperand(0);
16748 if (VT == MVT::i8) {
16749 // Zero extend to i32 since there is not an i8 bsr.
16751 Op = DAG.getNode(ISD::ZERO_EXTEND, dl, OpVT, Op);
16754 // Issue a bsr (scan bits in reverse) which also sets EFLAGS.
16755 SDVTList VTs = DAG.getVTList(OpVT, MVT::i32);
16756 Op = DAG.getNode(X86ISD::BSR, dl, VTs, Op);
16758 // If src is zero (i.e. bsr sets ZF), returns NumBits.
16761 DAG.getConstant(NumBits + NumBits - 1, dl, OpVT),
16762 DAG.getConstant(X86::COND_E, dl, MVT::i8),
16765 Op = DAG.getNode(X86ISD::CMOV, dl, OpVT, Ops);
16767 // Finally xor with NumBits-1.
16768 Op = DAG.getNode(ISD::XOR, dl, OpVT, Op,
16769 DAG.getConstant(NumBits - 1, dl, OpVT));
16772 Op = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op);
16776 static SDValue LowerCTLZ_ZERO_UNDEF(SDValue Op, SelectionDAG &DAG) {
16777 MVT VT = Op.getSimpleValueType();
16779 unsigned NumBits = VT.getSizeInBits();
16782 Op = Op.getOperand(0);
16783 if (VT == MVT::i8) {
16784 // Zero extend to i32 since there is not an i8 bsr.
16786 Op = DAG.getNode(ISD::ZERO_EXTEND, dl, OpVT, Op);
16789 // Issue a bsr (scan bits in reverse).
16790 SDVTList VTs = DAG.getVTList(OpVT, MVT::i32);
16791 Op = DAG.getNode(X86ISD::BSR, dl, VTs, Op);
16793 // And xor with NumBits-1.
16794 Op = DAG.getNode(ISD::XOR, dl, OpVT, Op,
16795 DAG.getConstant(NumBits - 1, dl, OpVT));
16798 Op = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op);
16802 static SDValue LowerCTTZ(SDValue Op, SelectionDAG &DAG) {
16803 MVT VT = Op.getSimpleValueType();
16804 unsigned NumBits = VT.getSizeInBits();
16806 Op = Op.getOperand(0);
16808 // Issue a bsf (scan bits forward) which also sets EFLAGS.
16809 SDVTList VTs = DAG.getVTList(VT, MVT::i32);
16810 Op = DAG.getNode(X86ISD::BSF, dl, VTs, Op);
16812 // If src is zero (i.e. bsf sets ZF), returns NumBits.
16815 DAG.getConstant(NumBits, dl, VT),
16816 DAG.getConstant(X86::COND_E, dl, MVT::i8),
16819 return DAG.getNode(X86ISD::CMOV, dl, VT, Ops);
16822 // Lower256IntArith - Break a 256-bit integer operation into two new 128-bit
16823 // ones, and then concatenate the result back.
16824 static SDValue Lower256IntArith(SDValue Op, SelectionDAG &DAG) {
16825 MVT VT = Op.getSimpleValueType();
16827 assert(VT.is256BitVector() && VT.isInteger() &&
16828 "Unsupported value type for operation");
16830 unsigned NumElems = VT.getVectorNumElements();
16833 // Extract the LHS vectors
16834 SDValue LHS = Op.getOperand(0);
16835 SDValue LHS1 = Extract128BitVector(LHS, 0, DAG, dl);
16836 SDValue LHS2 = Extract128BitVector(LHS, NumElems/2, DAG, dl);
16838 // Extract the RHS vectors
16839 SDValue RHS = Op.getOperand(1);
16840 SDValue RHS1 = Extract128BitVector(RHS, 0, DAG, dl);
16841 SDValue RHS2 = Extract128BitVector(RHS, NumElems/2, DAG, dl);
16843 MVT EltVT = VT.getVectorElementType();
16844 MVT NewVT = MVT::getVectorVT(EltVT, NumElems/2);
16846 return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT,
16847 DAG.getNode(Op.getOpcode(), dl, NewVT, LHS1, RHS1),
16848 DAG.getNode(Op.getOpcode(), dl, NewVT, LHS2, RHS2));
16851 static SDValue LowerADD(SDValue Op, SelectionDAG &DAG) {
16852 if (Op.getValueType() == MVT::i1)
16853 return DAG.getNode(ISD::XOR, SDLoc(Op), Op.getValueType(),
16854 Op.getOperand(0), Op.getOperand(1));
16855 assert(Op.getSimpleValueType().is256BitVector() &&
16856 Op.getSimpleValueType().isInteger() &&
16857 "Only handle AVX 256-bit vector integer operation");
16858 return Lower256IntArith(Op, DAG);
16861 static SDValue LowerSUB(SDValue Op, SelectionDAG &DAG) {
16862 if (Op.getValueType() == MVT::i1)
16863 return DAG.getNode(ISD::XOR, SDLoc(Op), Op.getValueType(),
16864 Op.getOperand(0), Op.getOperand(1));
16865 assert(Op.getSimpleValueType().is256BitVector() &&
16866 Op.getSimpleValueType().isInteger() &&
16867 "Only handle AVX 256-bit vector integer operation");
16868 return Lower256IntArith(Op, DAG);
16871 static SDValue LowerMUL(SDValue Op, const X86Subtarget *Subtarget,
16872 SelectionDAG &DAG) {
16874 MVT VT = Op.getSimpleValueType();
16877 return DAG.getNode(ISD::AND, dl, VT, Op.getOperand(0), Op.getOperand(1));
16879 // Decompose 256-bit ops into smaller 128-bit ops.
16880 if (VT.is256BitVector() && !Subtarget->hasInt256())
16881 return Lower256IntArith(Op, DAG);
16883 SDValue A = Op.getOperand(0);
16884 SDValue B = Op.getOperand(1);
16886 // Lower v16i8/v32i8 mul as promotion to v8i16/v16i16 vector
16887 // pairs, multiply and truncate.
16888 if (VT == MVT::v16i8 || VT == MVT::v32i8) {
16889 if (Subtarget->hasInt256()) {
16890 if (VT == MVT::v32i8) {
16891 MVT SubVT = MVT::getVectorVT(MVT::i8, VT.getVectorNumElements() / 2);
16892 SDValue Lo = DAG.getIntPtrConstant(0, dl);
16893 SDValue Hi = DAG.getIntPtrConstant(VT.getVectorNumElements() / 2, dl);
16894 SDValue ALo = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, SubVT, A, Lo);
16895 SDValue BLo = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, SubVT, B, Lo);
16896 SDValue AHi = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, SubVT, A, Hi);
16897 SDValue BHi = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, SubVT, B, Hi);
16898 return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT,
16899 DAG.getNode(ISD::MUL, dl, SubVT, ALo, BLo),
16900 DAG.getNode(ISD::MUL, dl, SubVT, AHi, BHi));
16903 MVT ExVT = MVT::getVectorVT(MVT::i16, VT.getVectorNumElements());
16904 return DAG.getNode(
16905 ISD::TRUNCATE, dl, VT,
16906 DAG.getNode(ISD::MUL, dl, ExVT,
16907 DAG.getNode(ISD::SIGN_EXTEND, dl, ExVT, A),
16908 DAG.getNode(ISD::SIGN_EXTEND, dl, ExVT, B)));
16911 assert(VT == MVT::v16i8 &&
16912 "Pre-AVX2 support only supports v16i8 multiplication");
16913 MVT ExVT = MVT::v8i16;
16915 // Extract the lo parts and sign extend to i16
16917 if (Subtarget->hasSSE41()) {
16918 ALo = DAG.getNode(X86ISD::VSEXT, dl, ExVT, A);
16919 BLo = DAG.getNode(X86ISD::VSEXT, dl, ExVT, B);
16921 const int ShufMask[] = {-1, 0, -1, 1, -1, 2, -1, 3,
16922 -1, 4, -1, 5, -1, 6, -1, 7};
16923 ALo = DAG.getVectorShuffle(VT, dl, A, A, ShufMask);
16924 BLo = DAG.getVectorShuffle(VT, dl, B, B, ShufMask);
16925 ALo = DAG.getBitcast(ExVT, ALo);
16926 BLo = DAG.getBitcast(ExVT, BLo);
16927 ALo = DAG.getNode(ISD::SRA, dl, ExVT, ALo, DAG.getConstant(8, dl, ExVT));
16928 BLo = DAG.getNode(ISD::SRA, dl, ExVT, BLo, DAG.getConstant(8, dl, ExVT));
16931 // Extract the hi parts and sign extend to i16
16933 if (Subtarget->hasSSE41()) {
16934 const int ShufMask[] = {8, 9, 10, 11, 12, 13, 14, 15,
16935 -1, -1, -1, -1, -1, -1, -1, -1};
16936 AHi = DAG.getVectorShuffle(VT, dl, A, A, ShufMask);
16937 BHi = DAG.getVectorShuffle(VT, dl, B, B, ShufMask);
16938 AHi = DAG.getNode(X86ISD::VSEXT, dl, ExVT, AHi);
16939 BHi = DAG.getNode(X86ISD::VSEXT, dl, ExVT, BHi);
16941 const int ShufMask[] = {-1, 8, -1, 9, -1, 10, -1, 11,
16942 -1, 12, -1, 13, -1, 14, -1, 15};
16943 AHi = DAG.getVectorShuffle(VT, dl, A, A, ShufMask);
16944 BHi = DAG.getVectorShuffle(VT, dl, B, B, ShufMask);
16945 AHi = DAG.getBitcast(ExVT, AHi);
16946 BHi = DAG.getBitcast(ExVT, BHi);
16947 AHi = DAG.getNode(ISD::SRA, dl, ExVT, AHi, DAG.getConstant(8, dl, ExVT));
16948 BHi = DAG.getNode(ISD::SRA, dl, ExVT, BHi, DAG.getConstant(8, dl, ExVT));
16951 // Multiply, mask the lower 8bits of the lo/hi results and pack
16952 SDValue RLo = DAG.getNode(ISD::MUL, dl, ExVT, ALo, BLo);
16953 SDValue RHi = DAG.getNode(ISD::MUL, dl, ExVT, AHi, BHi);
16954 RLo = DAG.getNode(ISD::AND, dl, ExVT, RLo, DAG.getConstant(255, dl, ExVT));
16955 RHi = DAG.getNode(ISD::AND, dl, ExVT, RHi, DAG.getConstant(255, dl, ExVT));
16956 return DAG.getNode(X86ISD::PACKUS, dl, VT, RLo, RHi);
16959 // Lower v4i32 mul as 2x shuffle, 2x pmuludq, 2x shuffle.
16960 if (VT == MVT::v4i32) {
16961 assert(Subtarget->hasSSE2() && !Subtarget->hasSSE41() &&
16962 "Should not custom lower when pmuldq is available!");
16964 // Extract the odd parts.
16965 static const int UnpackMask[] = { 1, -1, 3, -1 };
16966 SDValue Aodds = DAG.getVectorShuffle(VT, dl, A, A, UnpackMask);
16967 SDValue Bodds = DAG.getVectorShuffle(VT, dl, B, B, UnpackMask);
16969 // Multiply the even parts.
16970 SDValue Evens = DAG.getNode(X86ISD::PMULUDQ, dl, MVT::v2i64, A, B);
16971 // Now multiply odd parts.
16972 SDValue Odds = DAG.getNode(X86ISD::PMULUDQ, dl, MVT::v2i64, Aodds, Bodds);
16974 Evens = DAG.getBitcast(VT, Evens);
16975 Odds = DAG.getBitcast(VT, Odds);
16977 // Merge the two vectors back together with a shuffle. This expands into 2
16979 static const int ShufMask[] = { 0, 4, 2, 6 };
16980 return DAG.getVectorShuffle(VT, dl, Evens, Odds, ShufMask);
16983 assert((VT == MVT::v2i64 || VT == MVT::v4i64 || VT == MVT::v8i64) &&
16984 "Only know how to lower V2I64/V4I64/V8I64 multiply");
16986 // Ahi = psrlqi(a, 32);
16987 // Bhi = psrlqi(b, 32);
16989 // AloBlo = pmuludq(a, b);
16990 // AloBhi = pmuludq(a, Bhi);
16991 // AhiBlo = pmuludq(Ahi, b);
16993 // AloBhi = psllqi(AloBhi, 32);
16994 // AhiBlo = psllqi(AhiBlo, 32);
16995 // return AloBlo + AloBhi + AhiBlo;
16997 SDValue Ahi = getTargetVShiftByConstNode(X86ISD::VSRLI, dl, VT, A, 32, DAG);
16998 SDValue Bhi = getTargetVShiftByConstNode(X86ISD::VSRLI, dl, VT, B, 32, DAG);
17000 SDValue AhiBlo = Ahi;
17001 SDValue AloBhi = Bhi;
17002 // Bit cast to 32-bit vectors for MULUDQ
17003 EVT MulVT = (VT == MVT::v2i64) ? MVT::v4i32 :
17004 (VT == MVT::v4i64) ? MVT::v8i32 : MVT::v16i32;
17005 A = DAG.getBitcast(MulVT, A);
17006 B = DAG.getBitcast(MulVT, B);
17007 Ahi = DAG.getBitcast(MulVT, Ahi);
17008 Bhi = DAG.getBitcast(MulVT, Bhi);
17010 SDValue AloBlo = DAG.getNode(X86ISD::PMULUDQ, dl, VT, A, B);
17011 // After shifting right const values the result may be all-zero.
17012 if (!ISD::isBuildVectorAllZeros(Ahi.getNode())) {
17013 AhiBlo = DAG.getNode(X86ISD::PMULUDQ, dl, VT, Ahi, B);
17014 AhiBlo = getTargetVShiftByConstNode(X86ISD::VSHLI, dl, VT, AhiBlo, 32, DAG);
17016 if (!ISD::isBuildVectorAllZeros(Bhi.getNode())) {
17017 AloBhi = DAG.getNode(X86ISD::PMULUDQ, dl, VT, A, Bhi);
17018 AloBhi = getTargetVShiftByConstNode(X86ISD::VSHLI, dl, VT, AloBhi, 32, DAG);
17021 SDValue Res = DAG.getNode(ISD::ADD, dl, VT, AloBlo, AloBhi);
17022 return DAG.getNode(ISD::ADD, dl, VT, Res, AhiBlo);
17025 SDValue X86TargetLowering::LowerWin64_i128OP(SDValue Op, SelectionDAG &DAG) const {
17026 assert(Subtarget->isTargetWin64() && "Unexpected target");
17027 EVT VT = Op.getValueType();
17028 assert(VT.isInteger() && VT.getSizeInBits() == 128 &&
17029 "Unexpected return type for lowering");
17033 switch (Op->getOpcode()) {
17034 default: llvm_unreachable("Unexpected request for libcall!");
17035 case ISD::SDIV: isSigned = true; LC = RTLIB::SDIV_I128; break;
17036 case ISD::UDIV: isSigned = false; LC = RTLIB::UDIV_I128; break;
17037 case ISD::SREM: isSigned = true; LC = RTLIB::SREM_I128; break;
17038 case ISD::UREM: isSigned = false; LC = RTLIB::UREM_I128; break;
17039 case ISD::SDIVREM: isSigned = true; LC = RTLIB::SDIVREM_I128; break;
17040 case ISD::UDIVREM: isSigned = false; LC = RTLIB::UDIVREM_I128; break;
17044 SDValue InChain = DAG.getEntryNode();
17046 TargetLowering::ArgListTy Args;
17047 TargetLowering::ArgListEntry Entry;
17048 for (unsigned i = 0, e = Op->getNumOperands(); i != e; ++i) {
17049 EVT ArgVT = Op->getOperand(i).getValueType();
17050 assert(ArgVT.isInteger() && ArgVT.getSizeInBits() == 128 &&
17051 "Unexpected argument type for lowering");
17052 SDValue StackPtr = DAG.CreateStackTemporary(ArgVT, 16);
17053 Entry.Node = StackPtr;
17054 InChain = DAG.getStore(InChain, dl, Op->getOperand(i), StackPtr, MachinePointerInfo(),
17056 Type *ArgTy = ArgVT.getTypeForEVT(*DAG.getContext());
17057 Entry.Ty = PointerType::get(ArgTy,0);
17058 Entry.isSExt = false;
17059 Entry.isZExt = false;
17060 Args.push_back(Entry);
17063 SDValue Callee = DAG.getExternalSymbol(getLibcallName(LC),
17064 getPointerTy(DAG.getDataLayout()));
17066 TargetLowering::CallLoweringInfo CLI(DAG);
17067 CLI.setDebugLoc(dl).setChain(InChain)
17068 .setCallee(getLibcallCallingConv(LC),
17069 static_cast<EVT>(MVT::v2i64).getTypeForEVT(*DAG.getContext()),
17070 Callee, std::move(Args), 0)
17071 .setInRegister().setSExtResult(isSigned).setZExtResult(!isSigned);
17073 std::pair<SDValue, SDValue> CallInfo = LowerCallTo(CLI);
17074 return DAG.getBitcast(VT, CallInfo.first);
17077 static SDValue LowerMUL_LOHI(SDValue Op, const X86Subtarget *Subtarget,
17078 SelectionDAG &DAG) {
17079 SDValue Op0 = Op.getOperand(0), Op1 = Op.getOperand(1);
17080 EVT VT = Op0.getValueType();
17083 assert((VT == MVT::v4i32 && Subtarget->hasSSE2()) ||
17084 (VT == MVT::v8i32 && Subtarget->hasInt256()));
17086 // PMULxD operations multiply each even value (starting at 0) of LHS with
17087 // the related value of RHS and produce a widen result.
17088 // E.g., PMULUDQ <4 x i32> <a|b|c|d>, <4 x i32> <e|f|g|h>
17089 // => <2 x i64> <ae|cg>
17091 // In other word, to have all the results, we need to perform two PMULxD:
17092 // 1. one with the even values.
17093 // 2. one with the odd values.
17094 // To achieve #2, with need to place the odd values at an even position.
17096 // Place the odd value at an even position (basically, shift all values 1
17097 // step to the left):
17098 const int Mask[] = {1, -1, 3, -1, 5, -1, 7, -1};
17099 // <a|b|c|d> => <b|undef|d|undef>
17100 SDValue Odd0 = DAG.getVectorShuffle(VT, dl, Op0, Op0, Mask);
17101 // <e|f|g|h> => <f|undef|h|undef>
17102 SDValue Odd1 = DAG.getVectorShuffle(VT, dl, Op1, Op1, Mask);
17104 // Emit two multiplies, one for the lower 2 ints and one for the higher 2
17106 MVT MulVT = VT == MVT::v4i32 ? MVT::v2i64 : MVT::v4i64;
17107 bool IsSigned = Op->getOpcode() == ISD::SMUL_LOHI;
17109 (!IsSigned || !Subtarget->hasSSE41()) ? X86ISD::PMULUDQ : X86ISD::PMULDQ;
17110 // PMULUDQ <4 x i32> <a|b|c|d>, <4 x i32> <e|f|g|h>
17111 // => <2 x i64> <ae|cg>
17112 SDValue Mul1 = DAG.getBitcast(VT, DAG.getNode(Opcode, dl, MulVT, Op0, Op1));
17113 // PMULUDQ <4 x i32> <b|undef|d|undef>, <4 x i32> <f|undef|h|undef>
17114 // => <2 x i64> <bf|dh>
17115 SDValue Mul2 = DAG.getBitcast(VT, DAG.getNode(Opcode, dl, MulVT, Odd0, Odd1));
17117 // Shuffle it back into the right order.
17118 SDValue Highs, Lows;
17119 if (VT == MVT::v8i32) {
17120 const int HighMask[] = {1, 9, 3, 11, 5, 13, 7, 15};
17121 Highs = DAG.getVectorShuffle(VT, dl, Mul1, Mul2, HighMask);
17122 const int LowMask[] = {0, 8, 2, 10, 4, 12, 6, 14};
17123 Lows = DAG.getVectorShuffle(VT, dl, Mul1, Mul2, LowMask);
17125 const int HighMask[] = {1, 5, 3, 7};
17126 Highs = DAG.getVectorShuffle(VT, dl, Mul1, Mul2, HighMask);
17127 const int LowMask[] = {0, 4, 2, 6};
17128 Lows = DAG.getVectorShuffle(VT, dl, Mul1, Mul2, LowMask);
17131 // If we have a signed multiply but no PMULDQ fix up the high parts of a
17132 // unsigned multiply.
17133 if (IsSigned && !Subtarget->hasSSE41()) {
17134 SDValue ShAmt = DAG.getConstant(
17136 DAG.getTargetLoweringInfo().getShiftAmountTy(VT, DAG.getDataLayout()));
17137 SDValue T1 = DAG.getNode(ISD::AND, dl, VT,
17138 DAG.getNode(ISD::SRA, dl, VT, Op0, ShAmt), Op1);
17139 SDValue T2 = DAG.getNode(ISD::AND, dl, VT,
17140 DAG.getNode(ISD::SRA, dl, VT, Op1, ShAmt), Op0);
17142 SDValue Fixup = DAG.getNode(ISD::ADD, dl, VT, T1, T2);
17143 Highs = DAG.getNode(ISD::SUB, dl, VT, Highs, Fixup);
17146 // The first result of MUL_LOHI is actually the low value, followed by the
17148 SDValue Ops[] = {Lows, Highs};
17149 return DAG.getMergeValues(Ops, dl);
17152 // Return true if the required (according to Opcode) shift-imm form is natively
17153 // supported by the Subtarget
17154 static bool SupportedVectorShiftWithImm(MVT VT, const X86Subtarget *Subtarget,
17156 if (VT.getScalarSizeInBits() < 16)
17159 if (VT.is512BitVector() &&
17160 (VT.getScalarSizeInBits() > 16 || Subtarget->hasBWI()))
17163 bool LShift = VT.is128BitVector() ||
17164 (VT.is256BitVector() && Subtarget->hasInt256());
17166 bool AShift = LShift && (Subtarget->hasVLX() ||
17167 (VT != MVT::v2i64 && VT != MVT::v4i64));
17168 return (Opcode == ISD::SRA) ? AShift : LShift;
17171 // The shift amount is a variable, but it is the same for all vector lanes.
17172 // These instructions are defined together with shift-immediate.
17174 bool SupportedVectorShiftWithBaseAmnt(MVT VT, const X86Subtarget *Subtarget,
17176 return SupportedVectorShiftWithImm(VT, Subtarget, Opcode);
17179 // Return true if the required (according to Opcode) variable-shift form is
17180 // natively supported by the Subtarget
17181 static bool SupportedVectorVarShift(MVT VT, const X86Subtarget *Subtarget,
17184 if (!Subtarget->hasInt256() || VT.getScalarSizeInBits() < 16)
17187 // vXi16 supported only on AVX-512, BWI
17188 if (VT.getScalarSizeInBits() == 16 && !Subtarget->hasBWI())
17191 if (VT.is512BitVector() || Subtarget->hasVLX())
17194 bool LShift = VT.is128BitVector() || VT.is256BitVector();
17195 bool AShift = LShift && VT != MVT::v2i64 && VT != MVT::v4i64;
17196 return (Opcode == ISD::SRA) ? AShift : LShift;
17199 static SDValue LowerScalarImmediateShift(SDValue Op, SelectionDAG &DAG,
17200 const X86Subtarget *Subtarget) {
17201 MVT VT = Op.getSimpleValueType();
17203 SDValue R = Op.getOperand(0);
17204 SDValue Amt = Op.getOperand(1);
17206 unsigned X86Opc = (Op.getOpcode() == ISD::SHL) ? X86ISD::VSHLI :
17207 (Op.getOpcode() == ISD::SRL) ? X86ISD::VSRLI : X86ISD::VSRAI;
17209 auto ArithmeticShiftRight64 = [&](uint64_t ShiftAmt) {
17210 assert((VT == MVT::v2i64 || VT == MVT::v4i64) && "Unexpected SRA type");
17211 MVT ExVT = MVT::getVectorVT(MVT::i32, VT.getVectorNumElements() * 2);
17212 SDValue Ex = DAG.getBitcast(ExVT, R);
17214 if (ShiftAmt >= 32) {
17215 // Splat sign to upper i32 dst, and SRA upper i32 src to lower i32.
17217 getTargetVShiftByConstNode(X86ISD::VSRAI, dl, ExVT, Ex, 31, DAG);
17218 SDValue Lower = getTargetVShiftByConstNode(X86ISD::VSRAI, dl, ExVT, Ex,
17219 ShiftAmt - 32, DAG);
17220 if (VT == MVT::v2i64)
17221 Ex = DAG.getVectorShuffle(ExVT, dl, Upper, Lower, {5, 1, 7, 3});
17222 if (VT == MVT::v4i64)
17223 Ex = DAG.getVectorShuffle(ExVT, dl, Upper, Lower,
17224 {9, 1, 11, 3, 13, 5, 15, 7});
17226 // SRA upper i32, SHL whole i64 and select lower i32.
17227 SDValue Upper = getTargetVShiftByConstNode(X86ISD::VSRAI, dl, ExVT, Ex,
17230 getTargetVShiftByConstNode(X86ISD::VSRLI, dl, VT, R, ShiftAmt, DAG);
17231 Lower = DAG.getBitcast(ExVT, Lower);
17232 if (VT == MVT::v2i64)
17233 Ex = DAG.getVectorShuffle(ExVT, dl, Upper, Lower, {4, 1, 6, 3});
17234 if (VT == MVT::v4i64)
17235 Ex = DAG.getVectorShuffle(ExVT, dl, Upper, Lower,
17236 {8, 1, 10, 3, 12, 5, 14, 7});
17238 return DAG.getBitcast(VT, Ex);
17241 // Optimize shl/srl/sra with constant shift amount.
17242 if (auto *BVAmt = dyn_cast<BuildVectorSDNode>(Amt)) {
17243 if (auto *ShiftConst = BVAmt->getConstantSplatNode()) {
17244 uint64_t ShiftAmt = ShiftConst->getZExtValue();
17246 if (SupportedVectorShiftWithImm(VT, Subtarget, Op.getOpcode()))
17247 return getTargetVShiftByConstNode(X86Opc, dl, VT, R, ShiftAmt, DAG);
17249 // i64 SRA needs to be performed as partial shifts.
17250 if ((VT == MVT::v2i64 || (Subtarget->hasInt256() && VT == MVT::v4i64)) &&
17251 Op.getOpcode() == ISD::SRA)
17252 return ArithmeticShiftRight64(ShiftAmt);
17254 if (VT == MVT::v16i8 || (Subtarget->hasInt256() && VT == MVT::v32i8)) {
17255 unsigned NumElts = VT.getVectorNumElements();
17256 MVT ShiftVT = MVT::getVectorVT(MVT::i16, NumElts / 2);
17258 if (Op.getOpcode() == ISD::SHL) {
17259 // Simple i8 add case
17261 return DAG.getNode(ISD::ADD, dl, VT, R, R);
17263 // Make a large shift.
17264 SDValue SHL = getTargetVShiftByConstNode(X86ISD::VSHLI, dl, ShiftVT,
17266 SHL = DAG.getBitcast(VT, SHL);
17267 // Zero out the rightmost bits.
17268 SmallVector<SDValue, 32> V(
17269 NumElts, DAG.getConstant(uint8_t(-1U << ShiftAmt), dl, MVT::i8));
17270 return DAG.getNode(ISD::AND, dl, VT, SHL,
17271 DAG.getNode(ISD::BUILD_VECTOR, dl, VT, V));
17273 if (Op.getOpcode() == ISD::SRL) {
17274 // Make a large shift.
17275 SDValue SRL = getTargetVShiftByConstNode(X86ISD::VSRLI, dl, ShiftVT,
17277 SRL = DAG.getBitcast(VT, SRL);
17278 // Zero out the leftmost bits.
17279 SmallVector<SDValue, 32> V(
17280 NumElts, DAG.getConstant(uint8_t(-1U) >> ShiftAmt, dl, MVT::i8));
17281 return DAG.getNode(ISD::AND, dl, VT, SRL,
17282 DAG.getNode(ISD::BUILD_VECTOR, dl, VT, V));
17284 if (Op.getOpcode() == ISD::SRA) {
17285 if (ShiftAmt == 7) {
17286 // R s>> 7 === R s< 0
17287 SDValue Zeros = getZeroVector(VT, Subtarget, DAG, dl);
17288 return DAG.getNode(X86ISD::PCMPGT, dl, VT, Zeros, R);
17291 // R s>> a === ((R u>> a) ^ m) - m
17292 SDValue Res = DAG.getNode(ISD::SRL, dl, VT, R, Amt);
17293 SmallVector<SDValue, 32> V(NumElts,
17294 DAG.getConstant(128 >> ShiftAmt, dl,
17296 SDValue Mask = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, V);
17297 Res = DAG.getNode(ISD::XOR, dl, VT, Res, Mask);
17298 Res = DAG.getNode(ISD::SUB, dl, VT, Res, Mask);
17301 llvm_unreachable("Unknown shift opcode.");
17306 // Special case in 32-bit mode, where i64 is expanded into high and low parts.
17307 if (!Subtarget->is64Bit() &&
17308 (VT == MVT::v2i64 || (Subtarget->hasInt256() && VT == MVT::v4i64)) &&
17309 Amt.getOpcode() == ISD::BITCAST &&
17310 Amt.getOperand(0).getOpcode() == ISD::BUILD_VECTOR) {
17311 Amt = Amt.getOperand(0);
17312 unsigned Ratio = Amt.getSimpleValueType().getVectorNumElements() /
17313 VT.getVectorNumElements();
17314 unsigned RatioInLog2 = Log2_32_Ceil(Ratio);
17315 uint64_t ShiftAmt = 0;
17316 for (unsigned i = 0; i != Ratio; ++i) {
17317 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Amt.getOperand(i));
17321 ShiftAmt |= C->getZExtValue() << (i * (1 << (6 - RatioInLog2)));
17323 // Check remaining shift amounts.
17324 for (unsigned i = Ratio; i != Amt.getNumOperands(); i += Ratio) {
17325 uint64_t ShAmt = 0;
17326 for (unsigned j = 0; j != Ratio; ++j) {
17327 ConstantSDNode *C =
17328 dyn_cast<ConstantSDNode>(Amt.getOperand(i + j));
17332 ShAmt |= C->getZExtValue() << (j * (1 << (6 - RatioInLog2)));
17334 if (ShAmt != ShiftAmt)
17338 if (SupportedVectorShiftWithImm(VT, Subtarget, Op.getOpcode()))
17339 return getTargetVShiftByConstNode(X86Opc, dl, VT, R, ShiftAmt, DAG);
17341 if (Op.getOpcode() == ISD::SRA)
17342 return ArithmeticShiftRight64(ShiftAmt);
17348 static SDValue LowerScalarVariableShift(SDValue Op, SelectionDAG &DAG,
17349 const X86Subtarget* Subtarget) {
17350 MVT VT = Op.getSimpleValueType();
17352 SDValue R = Op.getOperand(0);
17353 SDValue Amt = Op.getOperand(1);
17355 unsigned X86OpcI = (Op.getOpcode() == ISD::SHL) ? X86ISD::VSHLI :
17356 (Op.getOpcode() == ISD::SRL) ? X86ISD::VSRLI : X86ISD::VSRAI;
17358 unsigned X86OpcV = (Op.getOpcode() == ISD::SHL) ? X86ISD::VSHL :
17359 (Op.getOpcode() == ISD::SRL) ? X86ISD::VSRL : X86ISD::VSRA;
17361 if (SupportedVectorShiftWithBaseAmnt(VT, Subtarget, Op.getOpcode())) {
17363 EVT EltVT = VT.getVectorElementType();
17365 if (BuildVectorSDNode *BV = dyn_cast<BuildVectorSDNode>(Amt)) {
17366 // Check if this build_vector node is doing a splat.
17367 // If so, then set BaseShAmt equal to the splat value.
17368 BaseShAmt = BV->getSplatValue();
17369 if (BaseShAmt && BaseShAmt.getOpcode() == ISD::UNDEF)
17370 BaseShAmt = SDValue();
17372 if (Amt.getOpcode() == ISD::EXTRACT_SUBVECTOR)
17373 Amt = Amt.getOperand(0);
17375 ShuffleVectorSDNode *SVN = dyn_cast<ShuffleVectorSDNode>(Amt);
17376 if (SVN && SVN->isSplat()) {
17377 unsigned SplatIdx = (unsigned)SVN->getSplatIndex();
17378 SDValue InVec = Amt.getOperand(0);
17379 if (InVec.getOpcode() == ISD::BUILD_VECTOR) {
17380 assert((SplatIdx < InVec.getValueType().getVectorNumElements()) &&
17381 "Unexpected shuffle index found!");
17382 BaseShAmt = InVec.getOperand(SplatIdx);
17383 } else if (InVec.getOpcode() == ISD::INSERT_VECTOR_ELT) {
17384 if (ConstantSDNode *C =
17385 dyn_cast<ConstantSDNode>(InVec.getOperand(2))) {
17386 if (C->getZExtValue() == SplatIdx)
17387 BaseShAmt = InVec.getOperand(1);
17392 // Avoid introducing an extract element from a shuffle.
17393 BaseShAmt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, EltVT, InVec,
17394 DAG.getIntPtrConstant(SplatIdx, dl));
17398 if (BaseShAmt.getNode()) {
17399 assert(EltVT.bitsLE(MVT::i64) && "Unexpected element type!");
17400 if (EltVT != MVT::i64 && EltVT.bitsGT(MVT::i32))
17401 BaseShAmt = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i64, BaseShAmt);
17402 else if (EltVT.bitsLT(MVT::i32))
17403 BaseShAmt = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, BaseShAmt);
17405 return getTargetVShiftNode(X86OpcI, dl, VT, R, BaseShAmt, DAG);
17409 // Special case in 32-bit mode, where i64 is expanded into high and low parts.
17410 if (!Subtarget->is64Bit() && VT == MVT::v2i64 &&
17411 Amt.getOpcode() == ISD::BITCAST &&
17412 Amt.getOperand(0).getOpcode() == ISD::BUILD_VECTOR) {
17413 Amt = Amt.getOperand(0);
17414 unsigned Ratio = Amt.getSimpleValueType().getVectorNumElements() /
17415 VT.getVectorNumElements();
17416 std::vector<SDValue> Vals(Ratio);
17417 for (unsigned i = 0; i != Ratio; ++i)
17418 Vals[i] = Amt.getOperand(i);
17419 for (unsigned i = Ratio; i != Amt.getNumOperands(); i += Ratio) {
17420 for (unsigned j = 0; j != Ratio; ++j)
17421 if (Vals[j] != Amt.getOperand(i + j))
17425 if (SupportedVectorShiftWithBaseAmnt(VT, Subtarget, Op.getOpcode()))
17426 return DAG.getNode(X86OpcV, dl, VT, R, Op.getOperand(1));
17431 static SDValue LowerShift(SDValue Op, const X86Subtarget* Subtarget,
17432 SelectionDAG &DAG) {
17433 MVT VT = Op.getSimpleValueType();
17435 SDValue R = Op.getOperand(0);
17436 SDValue Amt = Op.getOperand(1);
17438 assert(VT.isVector() && "Custom lowering only for vector shifts!");
17439 assert(Subtarget->hasSSE2() && "Only custom lower when we have SSE2!");
17441 if (SDValue V = LowerScalarImmediateShift(Op, DAG, Subtarget))
17444 if (SDValue V = LowerScalarVariableShift(Op, DAG, Subtarget))
17447 if (SupportedVectorVarShift(VT, Subtarget, Op.getOpcode()))
17450 // 2i64 vector logical shifts can efficiently avoid scalarization - do the
17451 // shifts per-lane and then shuffle the partial results back together.
17452 if (VT == MVT::v2i64 && Op.getOpcode() != ISD::SRA) {
17453 // Splat the shift amounts so the scalar shifts above will catch it.
17454 SDValue Amt0 = DAG.getVectorShuffle(VT, dl, Amt, Amt, {0, 0});
17455 SDValue Amt1 = DAG.getVectorShuffle(VT, dl, Amt, Amt, {1, 1});
17456 SDValue R0 = DAG.getNode(Op->getOpcode(), dl, VT, R, Amt0);
17457 SDValue R1 = DAG.getNode(Op->getOpcode(), dl, VT, R, Amt1);
17458 return DAG.getVectorShuffle(VT, dl, R0, R1, {0, 3});
17461 // If possible, lower this packed shift into a vector multiply instead of
17462 // expanding it into a sequence of scalar shifts.
17463 // Do this only if the vector shift count is a constant build_vector.
17464 if (Op.getOpcode() == ISD::SHL &&
17465 (VT == MVT::v8i16 || VT == MVT::v4i32 ||
17466 (Subtarget->hasInt256() && VT == MVT::v16i16)) &&
17467 ISD::isBuildVectorOfConstantSDNodes(Amt.getNode())) {
17468 SmallVector<SDValue, 8> Elts;
17469 EVT SVT = VT.getScalarType();
17470 unsigned SVTBits = SVT.getSizeInBits();
17471 const APInt &One = APInt(SVTBits, 1);
17472 unsigned NumElems = VT.getVectorNumElements();
17474 for (unsigned i=0; i !=NumElems; ++i) {
17475 SDValue Op = Amt->getOperand(i);
17476 if (Op->getOpcode() == ISD::UNDEF) {
17477 Elts.push_back(Op);
17481 ConstantSDNode *ND = cast<ConstantSDNode>(Op);
17482 const APInt &C = APInt(SVTBits, ND->getAPIntValue().getZExtValue());
17483 uint64_t ShAmt = C.getZExtValue();
17484 if (ShAmt >= SVTBits) {
17485 Elts.push_back(DAG.getUNDEF(SVT));
17488 Elts.push_back(DAG.getConstant(One.shl(ShAmt), dl, SVT));
17490 SDValue BV = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Elts);
17491 return DAG.getNode(ISD::MUL, dl, VT, R, BV);
17494 // Lower SHL with variable shift amount.
17495 if (VT == MVT::v4i32 && Op->getOpcode() == ISD::SHL) {
17496 Op = DAG.getNode(ISD::SHL, dl, VT, Amt, DAG.getConstant(23, dl, VT));
17498 Op = DAG.getNode(ISD::ADD, dl, VT, Op,
17499 DAG.getConstant(0x3f800000U, dl, VT));
17500 Op = DAG.getBitcast(MVT::v4f32, Op);
17501 Op = DAG.getNode(ISD::FP_TO_SINT, dl, VT, Op);
17502 return DAG.getNode(ISD::MUL, dl, VT, Op, R);
17505 // If possible, lower this shift as a sequence of two shifts by
17506 // constant plus a MOVSS/MOVSD instead of scalarizing it.
17508 // (v4i32 (srl A, (build_vector < X, Y, Y, Y>)))
17510 // Could be rewritten as:
17511 // (v4i32 (MOVSS (srl A, <Y,Y,Y,Y>), (srl A, <X,X,X,X>)))
17513 // The advantage is that the two shifts from the example would be
17514 // lowered as X86ISD::VSRLI nodes. This would be cheaper than scalarizing
17515 // the vector shift into four scalar shifts plus four pairs of vector
17517 if ((VT == MVT::v8i16 || VT == MVT::v4i32) &&
17518 ISD::isBuildVectorOfConstantSDNodes(Amt.getNode())) {
17519 unsigned TargetOpcode = X86ISD::MOVSS;
17520 bool CanBeSimplified;
17521 // The splat value for the first packed shift (the 'X' from the example).
17522 SDValue Amt1 = Amt->getOperand(0);
17523 // The splat value for the second packed shift (the 'Y' from the example).
17524 SDValue Amt2 = (VT == MVT::v4i32) ? Amt->getOperand(1) :
17525 Amt->getOperand(2);
17527 // See if it is possible to replace this node with a sequence of
17528 // two shifts followed by a MOVSS/MOVSD
17529 if (VT == MVT::v4i32) {
17530 // Check if it is legal to use a MOVSS.
17531 CanBeSimplified = Amt2 == Amt->getOperand(2) &&
17532 Amt2 == Amt->getOperand(3);
17533 if (!CanBeSimplified) {
17534 // Otherwise, check if we can still simplify this node using a MOVSD.
17535 CanBeSimplified = Amt1 == Amt->getOperand(1) &&
17536 Amt->getOperand(2) == Amt->getOperand(3);
17537 TargetOpcode = X86ISD::MOVSD;
17538 Amt2 = Amt->getOperand(2);
17541 // Do similar checks for the case where the machine value type
17543 CanBeSimplified = Amt1 == Amt->getOperand(1);
17544 for (unsigned i=3; i != 8 && CanBeSimplified; ++i)
17545 CanBeSimplified = Amt2 == Amt->getOperand(i);
17547 if (!CanBeSimplified) {
17548 TargetOpcode = X86ISD::MOVSD;
17549 CanBeSimplified = true;
17550 Amt2 = Amt->getOperand(4);
17551 for (unsigned i=0; i != 4 && CanBeSimplified; ++i)
17552 CanBeSimplified = Amt1 == Amt->getOperand(i);
17553 for (unsigned j=4; j != 8 && CanBeSimplified; ++j)
17554 CanBeSimplified = Amt2 == Amt->getOperand(j);
17558 if (CanBeSimplified && isa<ConstantSDNode>(Amt1) &&
17559 isa<ConstantSDNode>(Amt2)) {
17560 // Replace this node with two shifts followed by a MOVSS/MOVSD.
17561 EVT CastVT = MVT::v4i32;
17563 DAG.getConstant(cast<ConstantSDNode>(Amt1)->getAPIntValue(), dl, VT);
17564 SDValue Shift1 = DAG.getNode(Op->getOpcode(), dl, VT, R, Splat1);
17566 DAG.getConstant(cast<ConstantSDNode>(Amt2)->getAPIntValue(), dl, VT);
17567 SDValue Shift2 = DAG.getNode(Op->getOpcode(), dl, VT, R, Splat2);
17568 if (TargetOpcode == X86ISD::MOVSD)
17569 CastVT = MVT::v2i64;
17570 SDValue BitCast1 = DAG.getBitcast(CastVT, Shift1);
17571 SDValue BitCast2 = DAG.getBitcast(CastVT, Shift2);
17572 SDValue Result = getTargetShuffleNode(TargetOpcode, dl, CastVT, BitCast2,
17574 return DAG.getBitcast(VT, Result);
17578 // v4i32 Non Uniform Shifts.
17579 // If the shift amount is constant we can shift each lane using the SSE2
17580 // immediate shifts, else we need to zero-extend each lane to the lower i64
17581 // and shift using the SSE2 variable shifts.
17582 // The separate results can then be blended together.
17583 if (VT == MVT::v4i32) {
17584 unsigned Opc = Op.getOpcode();
17585 SDValue Amt0, Amt1, Amt2, Amt3;
17586 if (ISD::isBuildVectorOfConstantSDNodes(Amt.getNode())) {
17587 Amt0 = DAG.getVectorShuffle(VT, dl, Amt, DAG.getUNDEF(VT), {0, 0, 0, 0});
17588 Amt1 = DAG.getVectorShuffle(VT, dl, Amt, DAG.getUNDEF(VT), {1, 1, 1, 1});
17589 Amt2 = DAG.getVectorShuffle(VT, dl, Amt, DAG.getUNDEF(VT), {2, 2, 2, 2});
17590 Amt3 = DAG.getVectorShuffle(VT, dl, Amt, DAG.getUNDEF(VT), {3, 3, 3, 3});
17592 // ISD::SHL is handled above but we include it here for completeness.
17595 llvm_unreachable("Unknown target vector shift node");
17597 Opc = X86ISD::VSHL;
17600 Opc = X86ISD::VSRL;
17603 Opc = X86ISD::VSRA;
17606 // The SSE2 shifts use the lower i64 as the same shift amount for
17607 // all lanes and the upper i64 is ignored. These shuffle masks
17608 // optimally zero-extend each lanes on SSE2/SSE41/AVX targets.
17609 SDValue Z = getZeroVector(VT, Subtarget, DAG, dl);
17610 Amt0 = DAG.getVectorShuffle(VT, dl, Amt, Z, {0, 4, -1, -1});
17611 Amt1 = DAG.getVectorShuffle(VT, dl, Amt, Z, {1, 5, -1, -1});
17612 Amt2 = DAG.getVectorShuffle(VT, dl, Amt, Z, {2, 6, -1, -1});
17613 Amt3 = DAG.getVectorShuffle(VT, dl, Amt, Z, {3, 7, -1, -1});
17616 SDValue R0 = DAG.getNode(Opc, dl, VT, R, Amt0);
17617 SDValue R1 = DAG.getNode(Opc, dl, VT, R, Amt1);
17618 SDValue R2 = DAG.getNode(Opc, dl, VT, R, Amt2);
17619 SDValue R3 = DAG.getNode(Opc, dl, VT, R, Amt3);
17620 SDValue R02 = DAG.getVectorShuffle(VT, dl, R0, R2, {0, -1, 6, -1});
17621 SDValue R13 = DAG.getVectorShuffle(VT, dl, R1, R3, {-1, 1, -1, 7});
17622 return DAG.getVectorShuffle(VT, dl, R02, R13, {0, 5, 2, 7});
17625 if (VT == MVT::v16i8 || (VT == MVT::v32i8 && Subtarget->hasInt256())) {
17626 MVT ExtVT = MVT::getVectorVT(MVT::i16, VT.getVectorNumElements() / 2);
17627 unsigned ShiftOpcode = Op->getOpcode();
17629 auto SignBitSelect = [&](MVT SelVT, SDValue Sel, SDValue V0, SDValue V1) {
17630 // On SSE41 targets we make use of the fact that VSELECT lowers
17631 // to PBLENDVB which selects bytes based just on the sign bit.
17632 if (Subtarget->hasSSE41()) {
17633 V0 = DAG.getBitcast(VT, V0);
17634 V1 = DAG.getBitcast(VT, V1);
17635 Sel = DAG.getBitcast(VT, Sel);
17636 return DAG.getBitcast(SelVT,
17637 DAG.getNode(ISD::VSELECT, dl, VT, Sel, V0, V1));
17639 // On pre-SSE41 targets we test for the sign bit by comparing to
17640 // zero - a negative value will set all bits of the lanes to true
17641 // and VSELECT uses that in its OR(AND(V0,C),AND(V1,~C)) lowering.
17642 SDValue Z = getZeroVector(SelVT, Subtarget, DAG, dl);
17643 SDValue C = DAG.getNode(X86ISD::PCMPGT, dl, SelVT, Z, Sel);
17644 return DAG.getNode(ISD::VSELECT, dl, SelVT, C, V0, V1);
17647 // Turn 'a' into a mask suitable for VSELECT: a = a << 5;
17648 // We can safely do this using i16 shifts as we're only interested in
17649 // the 3 lower bits of each byte.
17650 Amt = DAG.getBitcast(ExtVT, Amt);
17651 Amt = DAG.getNode(ISD::SHL, dl, ExtVT, Amt, DAG.getConstant(5, dl, ExtVT));
17652 Amt = DAG.getBitcast(VT, Amt);
17654 if (Op->getOpcode() == ISD::SHL || Op->getOpcode() == ISD::SRL) {
17655 // r = VSELECT(r, shift(r, 4), a);
17657 DAG.getNode(ShiftOpcode, dl, VT, R, DAG.getConstant(4, dl, VT));
17658 R = SignBitSelect(VT, Amt, M, R);
17661 Amt = DAG.getNode(ISD::ADD, dl, VT, Amt, Amt);
17663 // r = VSELECT(r, shift(r, 2), a);
17664 M = DAG.getNode(ShiftOpcode, dl, VT, R, DAG.getConstant(2, dl, VT));
17665 R = SignBitSelect(VT, Amt, M, R);
17668 Amt = DAG.getNode(ISD::ADD, dl, VT, Amt, Amt);
17670 // return VSELECT(r, shift(r, 1), a);
17671 M = DAG.getNode(ShiftOpcode, dl, VT, R, DAG.getConstant(1, dl, VT));
17672 R = SignBitSelect(VT, Amt, M, R);
17676 if (Op->getOpcode() == ISD::SRA) {
17677 // For SRA we need to unpack each byte to the higher byte of a i16 vector
17678 // so we can correctly sign extend. We don't care what happens to the
17680 SDValue ALo = DAG.getNode(X86ISD::UNPCKL, dl, VT, DAG.getUNDEF(VT), Amt);
17681 SDValue AHi = DAG.getNode(X86ISD::UNPCKH, dl, VT, DAG.getUNDEF(VT), Amt);
17682 SDValue RLo = DAG.getNode(X86ISD::UNPCKL, dl, VT, DAG.getUNDEF(VT), R);
17683 SDValue RHi = DAG.getNode(X86ISD::UNPCKH, dl, VT, DAG.getUNDEF(VT), R);
17684 ALo = DAG.getBitcast(ExtVT, ALo);
17685 AHi = DAG.getBitcast(ExtVT, AHi);
17686 RLo = DAG.getBitcast(ExtVT, RLo);
17687 RHi = DAG.getBitcast(ExtVT, RHi);
17689 // r = VSELECT(r, shift(r, 4), a);
17690 SDValue MLo = DAG.getNode(ShiftOpcode, dl, ExtVT, RLo,
17691 DAG.getConstant(4, dl, ExtVT));
17692 SDValue MHi = DAG.getNode(ShiftOpcode, dl, ExtVT, RHi,
17693 DAG.getConstant(4, dl, ExtVT));
17694 RLo = SignBitSelect(ExtVT, ALo, MLo, RLo);
17695 RHi = SignBitSelect(ExtVT, AHi, MHi, RHi);
17698 ALo = DAG.getNode(ISD::ADD, dl, ExtVT, ALo, ALo);
17699 AHi = DAG.getNode(ISD::ADD, dl, ExtVT, AHi, AHi);
17701 // r = VSELECT(r, shift(r, 2), a);
17702 MLo = DAG.getNode(ShiftOpcode, dl, ExtVT, RLo,
17703 DAG.getConstant(2, dl, ExtVT));
17704 MHi = DAG.getNode(ShiftOpcode, dl, ExtVT, RHi,
17705 DAG.getConstant(2, dl, ExtVT));
17706 RLo = SignBitSelect(ExtVT, ALo, MLo, RLo);
17707 RHi = SignBitSelect(ExtVT, AHi, MHi, RHi);
17710 ALo = DAG.getNode(ISD::ADD, dl, ExtVT, ALo, ALo);
17711 AHi = DAG.getNode(ISD::ADD, dl, ExtVT, AHi, AHi);
17713 // r = VSELECT(r, shift(r, 1), a);
17714 MLo = DAG.getNode(ShiftOpcode, dl, ExtVT, RLo,
17715 DAG.getConstant(1, dl, ExtVT));
17716 MHi = DAG.getNode(ShiftOpcode, dl, ExtVT, RHi,
17717 DAG.getConstant(1, dl, ExtVT));
17718 RLo = SignBitSelect(ExtVT, ALo, MLo, RLo);
17719 RHi = SignBitSelect(ExtVT, AHi, MHi, RHi);
17721 // Logical shift the result back to the lower byte, leaving a zero upper
17723 // meaning that we can safely pack with PACKUSWB.
17725 DAG.getNode(ISD::SRL, dl, ExtVT, RLo, DAG.getConstant(8, dl, ExtVT));
17727 DAG.getNode(ISD::SRL, dl, ExtVT, RHi, DAG.getConstant(8, dl, ExtVT));
17728 return DAG.getNode(X86ISD::PACKUS, dl, VT, RLo, RHi);
17732 // It's worth extending once and using the v8i32 shifts for 16-bit types, but
17733 // the extra overheads to get from v16i8 to v8i32 make the existing SSE
17734 // solution better.
17735 if (Subtarget->hasInt256() && VT == MVT::v8i16) {
17736 MVT ExtVT = MVT::v8i32;
17738 Op.getOpcode() == ISD::SRA ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND;
17739 R = DAG.getNode(ExtOpc, dl, ExtVT, R);
17740 Amt = DAG.getNode(ISD::ANY_EXTEND, dl, ExtVT, Amt);
17741 return DAG.getNode(ISD::TRUNCATE, dl, VT,
17742 DAG.getNode(Op.getOpcode(), dl, ExtVT, R, Amt));
17745 if (Subtarget->hasInt256() && VT == MVT::v16i16) {
17746 MVT ExtVT = MVT::v8i32;
17747 SDValue Z = getZeroVector(VT, Subtarget, DAG, dl);
17748 SDValue ALo = DAG.getNode(X86ISD::UNPCKL, dl, VT, Amt, Z);
17749 SDValue AHi = DAG.getNode(X86ISD::UNPCKH, dl, VT, Amt, Z);
17750 SDValue RLo = DAG.getNode(X86ISD::UNPCKL, dl, VT, R, R);
17751 SDValue RHi = DAG.getNode(X86ISD::UNPCKH, dl, VT, R, R);
17752 ALo = DAG.getBitcast(ExtVT, ALo);
17753 AHi = DAG.getBitcast(ExtVT, AHi);
17754 RLo = DAG.getBitcast(ExtVT, RLo);
17755 RHi = DAG.getBitcast(ExtVT, RHi);
17756 SDValue Lo = DAG.getNode(Op.getOpcode(), dl, ExtVT, RLo, ALo);
17757 SDValue Hi = DAG.getNode(Op.getOpcode(), dl, ExtVT, RHi, AHi);
17758 Lo = DAG.getNode(ISD::SRL, dl, ExtVT, Lo, DAG.getConstant(16, dl, ExtVT));
17759 Hi = DAG.getNode(ISD::SRL, dl, ExtVT, Hi, DAG.getConstant(16, dl, ExtVT));
17760 return DAG.getNode(X86ISD::PACKUS, dl, VT, Lo, Hi);
17763 if (VT == MVT::v8i16) {
17764 unsigned ShiftOpcode = Op->getOpcode();
17766 auto SignBitSelect = [&](SDValue Sel, SDValue V0, SDValue V1) {
17767 // On SSE41 targets we make use of the fact that VSELECT lowers
17768 // to PBLENDVB which selects bytes based just on the sign bit.
17769 if (Subtarget->hasSSE41()) {
17770 MVT ExtVT = MVT::getVectorVT(MVT::i8, VT.getVectorNumElements() * 2);
17771 V0 = DAG.getBitcast(ExtVT, V0);
17772 V1 = DAG.getBitcast(ExtVT, V1);
17773 Sel = DAG.getBitcast(ExtVT, Sel);
17774 return DAG.getBitcast(
17775 VT, DAG.getNode(ISD::VSELECT, dl, ExtVT, Sel, V0, V1));
17777 // On pre-SSE41 targets we splat the sign bit - a negative value will
17778 // set all bits of the lanes to true and VSELECT uses that in
17779 // its OR(AND(V0,C),AND(V1,~C)) lowering.
17781 DAG.getNode(ISD::SRA, dl, VT, Sel, DAG.getConstant(15, dl, VT));
17782 return DAG.getNode(ISD::VSELECT, dl, VT, C, V0, V1);
17785 // Turn 'a' into a mask suitable for VSELECT: a = a << 12;
17786 if (Subtarget->hasSSE41()) {
17787 // On SSE41 targets we need to replicate the shift mask in both
17788 // bytes for PBLENDVB.
17791 DAG.getNode(ISD::SHL, dl, VT, Amt, DAG.getConstant(4, dl, VT)),
17792 DAG.getNode(ISD::SHL, dl, VT, Amt, DAG.getConstant(12, dl, VT)));
17794 Amt = DAG.getNode(ISD::SHL, dl, VT, Amt, DAG.getConstant(12, dl, VT));
17797 // r = VSELECT(r, shift(r, 8), a);
17798 SDValue M = DAG.getNode(ShiftOpcode, dl, VT, R, DAG.getConstant(8, dl, VT));
17799 R = SignBitSelect(Amt, M, R);
17802 Amt = DAG.getNode(ISD::ADD, dl, VT, Amt, Amt);
17804 // r = VSELECT(r, shift(r, 4), a);
17805 M = DAG.getNode(ShiftOpcode, dl, VT, R, DAG.getConstant(4, dl, VT));
17806 R = SignBitSelect(Amt, M, R);
17809 Amt = DAG.getNode(ISD::ADD, dl, VT, Amt, Amt);
17811 // r = VSELECT(r, shift(r, 2), a);
17812 M = DAG.getNode(ShiftOpcode, dl, VT, R, DAG.getConstant(2, dl, VT));
17813 R = SignBitSelect(Amt, M, R);
17816 Amt = DAG.getNode(ISD::ADD, dl, VT, Amt, Amt);
17818 // return VSELECT(r, shift(r, 1), a);
17819 M = DAG.getNode(ShiftOpcode, dl, VT, R, DAG.getConstant(1, dl, VT));
17820 R = SignBitSelect(Amt, M, R);
17824 // Decompose 256-bit shifts into smaller 128-bit shifts.
17825 if (VT.is256BitVector()) {
17826 unsigned NumElems = VT.getVectorNumElements();
17827 MVT EltVT = VT.getVectorElementType();
17828 EVT NewVT = MVT::getVectorVT(EltVT, NumElems/2);
17830 // Extract the two vectors
17831 SDValue V1 = Extract128BitVector(R, 0, DAG, dl);
17832 SDValue V2 = Extract128BitVector(R, NumElems/2, DAG, dl);
17834 // Recreate the shift amount vectors
17835 SDValue Amt1, Amt2;
17836 if (Amt.getOpcode() == ISD::BUILD_VECTOR) {
17837 // Constant shift amount
17838 SmallVector<SDValue, 8> Ops(Amt->op_begin(), Amt->op_begin() + NumElems);
17839 ArrayRef<SDValue> Amt1Csts = makeArrayRef(Ops).slice(0, NumElems / 2);
17840 ArrayRef<SDValue> Amt2Csts = makeArrayRef(Ops).slice(NumElems / 2);
17842 Amt1 = DAG.getNode(ISD::BUILD_VECTOR, dl, NewVT, Amt1Csts);
17843 Amt2 = DAG.getNode(ISD::BUILD_VECTOR, dl, NewVT, Amt2Csts);
17845 // Variable shift amount
17846 Amt1 = Extract128BitVector(Amt, 0, DAG, dl);
17847 Amt2 = Extract128BitVector(Amt, NumElems/2, DAG, dl);
17850 // Issue new vector shifts for the smaller types
17851 V1 = DAG.getNode(Op.getOpcode(), dl, NewVT, V1, Amt1);
17852 V2 = DAG.getNode(Op.getOpcode(), dl, NewVT, V2, Amt2);
17854 // Concatenate the result back
17855 return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT, V1, V2);
17861 static SDValue LowerXALUO(SDValue Op, SelectionDAG &DAG) {
17862 // Lower the "add/sub/mul with overflow" instruction into a regular ins plus
17863 // a "setcc" instruction that checks the overflow flag. The "brcond" lowering
17864 // looks for this combo and may remove the "setcc" instruction if the "setcc"
17865 // has only one use.
17866 SDNode *N = Op.getNode();
17867 SDValue LHS = N->getOperand(0);
17868 SDValue RHS = N->getOperand(1);
17869 unsigned BaseOp = 0;
17872 switch (Op.getOpcode()) {
17873 default: llvm_unreachable("Unknown ovf instruction!");
17875 // A subtract of one will be selected as a INC. Note that INC doesn't
17876 // set CF, so we can't do this for UADDO.
17877 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(RHS))
17879 BaseOp = X86ISD::INC;
17880 Cond = X86::COND_O;
17883 BaseOp = X86ISD::ADD;
17884 Cond = X86::COND_O;
17887 BaseOp = X86ISD::ADD;
17888 Cond = X86::COND_B;
17891 // A subtract of one will be selected as a DEC. Note that DEC doesn't
17892 // set CF, so we can't do this for USUBO.
17893 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(RHS))
17895 BaseOp = X86ISD::DEC;
17896 Cond = X86::COND_O;
17899 BaseOp = X86ISD::SUB;
17900 Cond = X86::COND_O;
17903 BaseOp = X86ISD::SUB;
17904 Cond = X86::COND_B;
17907 BaseOp = N->getValueType(0) == MVT::i8 ? X86ISD::SMUL8 : X86ISD::SMUL;
17908 Cond = X86::COND_O;
17910 case ISD::UMULO: { // i64, i8 = umulo lhs, rhs --> i64, i64, i32 umul lhs,rhs
17911 if (N->getValueType(0) == MVT::i8) {
17912 BaseOp = X86ISD::UMUL8;
17913 Cond = X86::COND_O;
17916 SDVTList VTs = DAG.getVTList(N->getValueType(0), N->getValueType(0),
17918 SDValue Sum = DAG.getNode(X86ISD::UMUL, DL, VTs, LHS, RHS);
17921 DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
17922 DAG.getConstant(X86::COND_O, DL, MVT::i32),
17923 SDValue(Sum.getNode(), 2));
17925 return DAG.getNode(ISD::MERGE_VALUES, DL, N->getVTList(), Sum, SetCC);
17929 // Also sets EFLAGS.
17930 SDVTList VTs = DAG.getVTList(N->getValueType(0), MVT::i32);
17931 SDValue Sum = DAG.getNode(BaseOp, DL, VTs, LHS, RHS);
17934 DAG.getNode(X86ISD::SETCC, DL, N->getValueType(1),
17935 DAG.getConstant(Cond, DL, MVT::i32),
17936 SDValue(Sum.getNode(), 1));
17938 return DAG.getNode(ISD::MERGE_VALUES, DL, N->getVTList(), Sum, SetCC);
17941 /// Returns true if the operand type is exactly twice the native width, and
17942 /// the corresponding cmpxchg8b or cmpxchg16b instruction is available.
17943 /// Used to know whether to use cmpxchg8/16b when expanding atomic operations
17944 /// (otherwise we leave them alone to become __sync_fetch_and_... calls).
17945 bool X86TargetLowering::needsCmpXchgNb(const Type *MemType) const {
17946 unsigned OpWidth = MemType->getPrimitiveSizeInBits();
17949 return !Subtarget->is64Bit(); // FIXME this should be Subtarget.hasCmpxchg8b
17950 else if (OpWidth == 128)
17951 return Subtarget->hasCmpxchg16b();
17956 bool X86TargetLowering::shouldExpandAtomicStoreInIR(StoreInst *SI) const {
17957 return needsCmpXchgNb(SI->getValueOperand()->getType());
17960 // Note: this turns large loads into lock cmpxchg8b/16b.
17961 // FIXME: On 32 bits x86, fild/movq might be faster than lock cmpxchg8b.
17962 bool X86TargetLowering::shouldExpandAtomicLoadInIR(LoadInst *LI) const {
17963 auto PTy = cast<PointerType>(LI->getPointerOperand()->getType());
17964 return needsCmpXchgNb(PTy->getElementType());
17967 TargetLoweringBase::AtomicRMWExpansionKind
17968 X86TargetLowering::shouldExpandAtomicRMWInIR(AtomicRMWInst *AI) const {
17969 unsigned NativeWidth = Subtarget->is64Bit() ? 64 : 32;
17970 const Type *MemType = AI->getType();
17972 // If the operand is too big, we must see if cmpxchg8/16b is available
17973 // and default to library calls otherwise.
17974 if (MemType->getPrimitiveSizeInBits() > NativeWidth) {
17975 return needsCmpXchgNb(MemType) ? AtomicRMWExpansionKind::CmpXChg
17976 : AtomicRMWExpansionKind::None;
17979 AtomicRMWInst::BinOp Op = AI->getOperation();
17982 llvm_unreachable("Unknown atomic operation");
17983 case AtomicRMWInst::Xchg:
17984 case AtomicRMWInst::Add:
17985 case AtomicRMWInst::Sub:
17986 // It's better to use xadd, xsub or xchg for these in all cases.
17987 return AtomicRMWExpansionKind::None;
17988 case AtomicRMWInst::Or:
17989 case AtomicRMWInst::And:
17990 case AtomicRMWInst::Xor:
17991 // If the atomicrmw's result isn't actually used, we can just add a "lock"
17992 // prefix to a normal instruction for these operations.
17993 return !AI->use_empty() ? AtomicRMWExpansionKind::CmpXChg
17994 : AtomicRMWExpansionKind::None;
17995 case AtomicRMWInst::Nand:
17996 case AtomicRMWInst::Max:
17997 case AtomicRMWInst::Min:
17998 case AtomicRMWInst::UMax:
17999 case AtomicRMWInst::UMin:
18000 // These always require a non-trivial set of data operations on x86. We must
18001 // use a cmpxchg loop.
18002 return AtomicRMWExpansionKind::CmpXChg;
18006 static bool hasMFENCE(const X86Subtarget& Subtarget) {
18007 // Use mfence if we have SSE2 or we're on x86-64 (even if we asked for
18008 // no-sse2). There isn't any reason to disable it if the target processor
18010 return Subtarget.hasSSE2() || Subtarget.is64Bit();
18014 X86TargetLowering::lowerIdempotentRMWIntoFencedLoad(AtomicRMWInst *AI) const {
18015 unsigned NativeWidth = Subtarget->is64Bit() ? 64 : 32;
18016 const Type *MemType = AI->getType();
18017 // Accesses larger than the native width are turned into cmpxchg/libcalls, so
18018 // there is no benefit in turning such RMWs into loads, and it is actually
18019 // harmful as it introduces a mfence.
18020 if (MemType->getPrimitiveSizeInBits() > NativeWidth)
18023 auto Builder = IRBuilder<>(AI);
18024 Module *M = Builder.GetInsertBlock()->getParent()->getParent();
18025 auto SynchScope = AI->getSynchScope();
18026 // We must restrict the ordering to avoid generating loads with Release or
18027 // ReleaseAcquire orderings.
18028 auto Order = AtomicCmpXchgInst::getStrongestFailureOrdering(AI->getOrdering());
18029 auto Ptr = AI->getPointerOperand();
18031 // Before the load we need a fence. Here is an example lifted from
18032 // http://www.hpl.hp.com/techreports/2012/HPL-2012-68.pdf showing why a fence
18035 // x.store(1, relaxed);
18036 // r1 = y.fetch_add(0, release);
18038 // y.fetch_add(42, acquire);
18039 // r2 = x.load(relaxed);
18040 // r1 = r2 = 0 is impossible, but becomes possible if the idempotent rmw is
18041 // lowered to just a load without a fence. A mfence flushes the store buffer,
18042 // making the optimization clearly correct.
18043 // FIXME: it is required if isAtLeastRelease(Order) but it is not clear
18044 // otherwise, we might be able to be more agressive on relaxed idempotent
18045 // rmw. In practice, they do not look useful, so we don't try to be
18046 // especially clever.
18047 if (SynchScope == SingleThread)
18048 // FIXME: we could just insert an X86ISD::MEMBARRIER here, except we are at
18049 // the IR level, so we must wrap it in an intrinsic.
18052 if (!hasMFENCE(*Subtarget))
18053 // FIXME: it might make sense to use a locked operation here but on a
18054 // different cache-line to prevent cache-line bouncing. In practice it
18055 // is probably a small win, and x86 processors without mfence are rare
18056 // enough that we do not bother.
18060 llvm::Intrinsic::getDeclaration(M, Intrinsic::x86_sse2_mfence);
18061 Builder.CreateCall(MFence, {});
18063 // Finally we can emit the atomic load.
18064 LoadInst *Loaded = Builder.CreateAlignedLoad(Ptr,
18065 AI->getType()->getPrimitiveSizeInBits());
18066 Loaded->setAtomic(Order, SynchScope);
18067 AI->replaceAllUsesWith(Loaded);
18068 AI->eraseFromParent();
18072 static SDValue LowerATOMIC_FENCE(SDValue Op, const X86Subtarget *Subtarget,
18073 SelectionDAG &DAG) {
18075 AtomicOrdering FenceOrdering = static_cast<AtomicOrdering>(
18076 cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue());
18077 SynchronizationScope FenceScope = static_cast<SynchronizationScope>(
18078 cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue());
18080 // The only fence that needs an instruction is a sequentially-consistent
18081 // cross-thread fence.
18082 if (FenceOrdering == SequentiallyConsistent && FenceScope == CrossThread) {
18083 if (hasMFENCE(*Subtarget))
18084 return DAG.getNode(X86ISD::MFENCE, dl, MVT::Other, Op.getOperand(0));
18086 SDValue Chain = Op.getOperand(0);
18087 SDValue Zero = DAG.getConstant(0, dl, MVT::i32);
18089 DAG.getRegister(X86::ESP, MVT::i32), // Base
18090 DAG.getTargetConstant(1, dl, MVT::i8), // Scale
18091 DAG.getRegister(0, MVT::i32), // Index
18092 DAG.getTargetConstant(0, dl, MVT::i32), // Disp
18093 DAG.getRegister(0, MVT::i32), // Segment.
18097 SDNode *Res = DAG.getMachineNode(X86::OR32mrLocked, dl, MVT::Other, Ops);
18098 return SDValue(Res, 0);
18101 // MEMBARRIER is a compiler barrier; it codegens to a no-op.
18102 return DAG.getNode(X86ISD::MEMBARRIER, dl, MVT::Other, Op.getOperand(0));
18105 static SDValue LowerCMP_SWAP(SDValue Op, const X86Subtarget *Subtarget,
18106 SelectionDAG &DAG) {
18107 MVT T = Op.getSimpleValueType();
18111 switch(T.SimpleTy) {
18112 default: llvm_unreachable("Invalid value type!");
18113 case MVT::i8: Reg = X86::AL; size = 1; break;
18114 case MVT::i16: Reg = X86::AX; size = 2; break;
18115 case MVT::i32: Reg = X86::EAX; size = 4; break;
18117 assert(Subtarget->is64Bit() && "Node not type legal!");
18118 Reg = X86::RAX; size = 8;
18121 SDValue cpIn = DAG.getCopyToReg(Op.getOperand(0), DL, Reg,
18122 Op.getOperand(2), SDValue());
18123 SDValue Ops[] = { cpIn.getValue(0),
18126 DAG.getTargetConstant(size, DL, MVT::i8),
18127 cpIn.getValue(1) };
18128 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Glue);
18129 MachineMemOperand *MMO = cast<AtomicSDNode>(Op)->getMemOperand();
18130 SDValue Result = DAG.getMemIntrinsicNode(X86ISD::LCMPXCHG_DAG, DL, Tys,
18134 DAG.getCopyFromReg(Result.getValue(0), DL, Reg, T, Result.getValue(1));
18135 SDValue EFLAGS = DAG.getCopyFromReg(cpOut.getValue(1), DL, X86::EFLAGS,
18136 MVT::i32, cpOut.getValue(2));
18137 SDValue Success = DAG.getNode(X86ISD::SETCC, DL, Op->getValueType(1),
18138 DAG.getConstant(X86::COND_E, DL, MVT::i8),
18141 DAG.ReplaceAllUsesOfValueWith(Op.getValue(0), cpOut);
18142 DAG.ReplaceAllUsesOfValueWith(Op.getValue(1), Success);
18143 DAG.ReplaceAllUsesOfValueWith(Op.getValue(2), EFLAGS.getValue(1));
18147 static SDValue LowerBITCAST(SDValue Op, const X86Subtarget *Subtarget,
18148 SelectionDAG &DAG) {
18149 MVT SrcVT = Op.getOperand(0).getSimpleValueType();
18150 MVT DstVT = Op.getSimpleValueType();
18152 if (SrcVT == MVT::v2i32 || SrcVT == MVT::v4i16 || SrcVT == MVT::v8i8) {
18153 assert(Subtarget->hasSSE2() && "Requires at least SSE2!");
18154 if (DstVT != MVT::f64)
18155 // This conversion needs to be expanded.
18158 SDValue InVec = Op->getOperand(0);
18160 unsigned NumElts = SrcVT.getVectorNumElements();
18161 EVT SVT = SrcVT.getVectorElementType();
18163 // Widen the vector in input in the case of MVT::v2i32.
18164 // Example: from MVT::v2i32 to MVT::v4i32.
18165 SmallVector<SDValue, 16> Elts;
18166 for (unsigned i = 0, e = NumElts; i != e; ++i)
18167 Elts.push_back(DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, SVT, InVec,
18168 DAG.getIntPtrConstant(i, dl)));
18170 // Explicitly mark the extra elements as Undef.
18171 Elts.append(NumElts, DAG.getUNDEF(SVT));
18173 EVT NewVT = EVT::getVectorVT(*DAG.getContext(), SVT, NumElts * 2);
18174 SDValue BV = DAG.getNode(ISD::BUILD_VECTOR, dl, NewVT, Elts);
18175 SDValue ToV2F64 = DAG.getBitcast(MVT::v2f64, BV);
18176 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, ToV2F64,
18177 DAG.getIntPtrConstant(0, dl));
18180 assert(Subtarget->is64Bit() && !Subtarget->hasSSE2() &&
18181 Subtarget->hasMMX() && "Unexpected custom BITCAST");
18182 assert((DstVT == MVT::i64 ||
18183 (DstVT.isVector() && DstVT.getSizeInBits()==64)) &&
18184 "Unexpected custom BITCAST");
18185 // i64 <=> MMX conversions are Legal.
18186 if (SrcVT==MVT::i64 && DstVT.isVector())
18188 if (DstVT==MVT::i64 && SrcVT.isVector())
18190 // MMX <=> MMX conversions are Legal.
18191 if (SrcVT.isVector() && DstVT.isVector())
18193 // All other conversions need to be expanded.
18197 /// Compute the horizontal sum of bytes in V for the elements of VT.
18199 /// Requires V to be a byte vector and VT to be an integer vector type with
18200 /// wider elements than V's type. The width of the elements of VT determines
18201 /// how many bytes of V are summed horizontally to produce each element of the
18203 static SDValue LowerHorizontalByteSum(SDValue V, MVT VT,
18204 const X86Subtarget *Subtarget,
18205 SelectionDAG &DAG) {
18207 MVT ByteVecVT = V.getSimpleValueType();
18208 MVT EltVT = VT.getVectorElementType();
18209 int NumElts = VT.getVectorNumElements();
18210 assert(ByteVecVT.getVectorElementType() == MVT::i8 &&
18211 "Expected value to have byte element type.");
18212 assert(EltVT != MVT::i8 &&
18213 "Horizontal byte sum only makes sense for wider elements!");
18214 unsigned VecSize = VT.getSizeInBits();
18215 assert(ByteVecVT.getSizeInBits() == VecSize && "Cannot change vector size!");
18217 // PSADBW instruction horizontally add all bytes and leave the result in i64
18218 // chunks, thus directly computes the pop count for v2i64 and v4i64.
18219 if (EltVT == MVT::i64) {
18220 SDValue Zeros = getZeroVector(ByteVecVT, Subtarget, DAG, DL);
18221 V = DAG.getNode(X86ISD::PSADBW, DL, ByteVecVT, V, Zeros);
18222 return DAG.getBitcast(VT, V);
18225 if (EltVT == MVT::i32) {
18226 // We unpack the low half and high half into i32s interleaved with zeros so
18227 // that we can use PSADBW to horizontally sum them. The most useful part of
18228 // this is that it lines up the results of two PSADBW instructions to be
18229 // two v2i64 vectors which concatenated are the 4 population counts. We can
18230 // then use PACKUSWB to shrink and concatenate them into a v4i32 again.
18231 SDValue Zeros = getZeroVector(VT, Subtarget, DAG, DL);
18232 SDValue Low = DAG.getNode(X86ISD::UNPCKL, DL, VT, V, Zeros);
18233 SDValue High = DAG.getNode(X86ISD::UNPCKH, DL, VT, V, Zeros);
18235 // Do the horizontal sums into two v2i64s.
18236 Zeros = getZeroVector(ByteVecVT, Subtarget, DAG, DL);
18237 Low = DAG.getNode(X86ISD::PSADBW, DL, ByteVecVT,
18238 DAG.getBitcast(ByteVecVT, Low), Zeros);
18239 High = DAG.getNode(X86ISD::PSADBW, DL, ByteVecVT,
18240 DAG.getBitcast(ByteVecVT, High), Zeros);
18242 // Merge them together.
18243 MVT ShortVecVT = MVT::getVectorVT(MVT::i16, VecSize / 16);
18244 V = DAG.getNode(X86ISD::PACKUS, DL, ByteVecVT,
18245 DAG.getBitcast(ShortVecVT, Low),
18246 DAG.getBitcast(ShortVecVT, High));
18248 return DAG.getBitcast(VT, V);
18251 // The only element type left is i16.
18252 assert(EltVT == MVT::i16 && "Unknown how to handle type");
18254 // To obtain pop count for each i16 element starting from the pop count for
18255 // i8 elements, shift the i16s left by 8, sum as i8s, and then shift as i16s
18256 // right by 8. It is important to shift as i16s as i8 vector shift isn't
18257 // directly supported.
18258 SmallVector<SDValue, 16> Shifters(NumElts, DAG.getConstant(8, DL, EltVT));
18259 SDValue Shifter = DAG.getNode(ISD::BUILD_VECTOR, DL, VT, Shifters);
18260 SDValue Shl = DAG.getNode(ISD::SHL, DL, VT, DAG.getBitcast(VT, V), Shifter);
18261 V = DAG.getNode(ISD::ADD, DL, ByteVecVT, DAG.getBitcast(ByteVecVT, Shl),
18262 DAG.getBitcast(ByteVecVT, V));
18263 return DAG.getNode(ISD::SRL, DL, VT, DAG.getBitcast(VT, V), Shifter);
18266 static SDValue LowerVectorCTPOPInRegLUT(SDValue Op, SDLoc DL,
18267 const X86Subtarget *Subtarget,
18268 SelectionDAG &DAG) {
18269 MVT VT = Op.getSimpleValueType();
18270 MVT EltVT = VT.getVectorElementType();
18271 unsigned VecSize = VT.getSizeInBits();
18273 // Implement a lookup table in register by using an algorithm based on:
18274 // http://wm.ite.pl/articles/sse-popcount.html
18276 // The general idea is that every lower byte nibble in the input vector is an
18277 // index into a in-register pre-computed pop count table. We then split up the
18278 // input vector in two new ones: (1) a vector with only the shifted-right
18279 // higher nibbles for each byte and (2) a vector with the lower nibbles (and
18280 // masked out higher ones) for each byte. PSHUB is used separately with both
18281 // to index the in-register table. Next, both are added and the result is a
18282 // i8 vector where each element contains the pop count for input byte.
18284 // To obtain the pop count for elements != i8, we follow up with the same
18285 // approach and use additional tricks as described below.
18287 const int LUT[16] = {/* 0 */ 0, /* 1 */ 1, /* 2 */ 1, /* 3 */ 2,
18288 /* 4 */ 1, /* 5 */ 2, /* 6 */ 2, /* 7 */ 3,
18289 /* 8 */ 1, /* 9 */ 2, /* a */ 2, /* b */ 3,
18290 /* c */ 2, /* d */ 3, /* e */ 3, /* f */ 4};
18292 int NumByteElts = VecSize / 8;
18293 MVT ByteVecVT = MVT::getVectorVT(MVT::i8, NumByteElts);
18294 SDValue In = DAG.getBitcast(ByteVecVT, Op);
18295 SmallVector<SDValue, 16> LUTVec;
18296 for (int i = 0; i < NumByteElts; ++i)
18297 LUTVec.push_back(DAG.getConstant(LUT[i % 16], DL, MVT::i8));
18298 SDValue InRegLUT = DAG.getNode(ISD::BUILD_VECTOR, DL, ByteVecVT, LUTVec);
18299 SmallVector<SDValue, 16> Mask0F(NumByteElts,
18300 DAG.getConstant(0x0F, DL, MVT::i8));
18301 SDValue M0F = DAG.getNode(ISD::BUILD_VECTOR, DL, ByteVecVT, Mask0F);
18304 SmallVector<SDValue, 16> Four(NumByteElts, DAG.getConstant(4, DL, MVT::i8));
18305 SDValue FourV = DAG.getNode(ISD::BUILD_VECTOR, DL, ByteVecVT, Four);
18306 SDValue HighNibbles = DAG.getNode(ISD::SRL, DL, ByteVecVT, In, FourV);
18309 SDValue LowNibbles = DAG.getNode(ISD::AND, DL, ByteVecVT, In, M0F);
18311 // The input vector is used as the shuffle mask that index elements into the
18312 // LUT. After counting low and high nibbles, add the vector to obtain the
18313 // final pop count per i8 element.
18314 SDValue HighPopCnt =
18315 DAG.getNode(X86ISD::PSHUFB, DL, ByteVecVT, InRegLUT, HighNibbles);
18316 SDValue LowPopCnt =
18317 DAG.getNode(X86ISD::PSHUFB, DL, ByteVecVT, InRegLUT, LowNibbles);
18318 SDValue PopCnt = DAG.getNode(ISD::ADD, DL, ByteVecVT, HighPopCnt, LowPopCnt);
18320 if (EltVT == MVT::i8)
18323 return LowerHorizontalByteSum(PopCnt, VT, Subtarget, DAG);
18326 static SDValue LowerVectorCTPOPBitmath(SDValue Op, SDLoc DL,
18327 const X86Subtarget *Subtarget,
18328 SelectionDAG &DAG) {
18329 MVT VT = Op.getSimpleValueType();
18330 assert(VT.is128BitVector() &&
18331 "Only 128-bit vector bitmath lowering supported.");
18333 int VecSize = VT.getSizeInBits();
18334 MVT EltVT = VT.getVectorElementType();
18335 int Len = EltVT.getSizeInBits();
18337 // This is the vectorized version of the "best" algorithm from
18338 // http://graphics.stanford.edu/~seander/bithacks.html#CountBitsSetParallel
18339 // with a minor tweak to use a series of adds + shifts instead of vector
18340 // multiplications. Implemented for all integer vector types. We only use
18341 // this when we don't have SSSE3 which allows a LUT-based lowering that is
18342 // much faster, even faster than using native popcnt instructions.
18344 auto GetShift = [&](unsigned OpCode, SDValue V, int Shifter) {
18345 MVT VT = V.getSimpleValueType();
18346 SmallVector<SDValue, 32> Shifters(
18347 VT.getVectorNumElements(),
18348 DAG.getConstant(Shifter, DL, VT.getVectorElementType()));
18349 return DAG.getNode(OpCode, DL, VT, V,
18350 DAG.getNode(ISD::BUILD_VECTOR, DL, VT, Shifters));
18352 auto GetMask = [&](SDValue V, APInt Mask) {
18353 MVT VT = V.getSimpleValueType();
18354 SmallVector<SDValue, 32> Masks(
18355 VT.getVectorNumElements(),
18356 DAG.getConstant(Mask, DL, VT.getVectorElementType()));
18357 return DAG.getNode(ISD::AND, DL, VT, V,
18358 DAG.getNode(ISD::BUILD_VECTOR, DL, VT, Masks));
18361 // We don't want to incur the implicit masks required to SRL vNi8 vectors on
18362 // x86, so set the SRL type to have elements at least i16 wide. This is
18363 // correct because all of our SRLs are followed immediately by a mask anyways
18364 // that handles any bits that sneak into the high bits of the byte elements.
18365 MVT SrlVT = Len > 8 ? VT : MVT::getVectorVT(MVT::i16, VecSize / 16);
18369 // v = v - ((v >> 1) & 0x55555555...)
18371 DAG.getBitcast(VT, GetShift(ISD::SRL, DAG.getBitcast(SrlVT, V), 1));
18372 SDValue And = GetMask(Srl, APInt::getSplat(Len, APInt(8, 0x55)));
18373 V = DAG.getNode(ISD::SUB, DL, VT, V, And);
18375 // v = (v & 0x33333333...) + ((v >> 2) & 0x33333333...)
18376 SDValue AndLHS = GetMask(V, APInt::getSplat(Len, APInt(8, 0x33)));
18377 Srl = DAG.getBitcast(VT, GetShift(ISD::SRL, DAG.getBitcast(SrlVT, V), 2));
18378 SDValue AndRHS = GetMask(Srl, APInt::getSplat(Len, APInt(8, 0x33)));
18379 V = DAG.getNode(ISD::ADD, DL, VT, AndLHS, AndRHS);
18381 // v = (v + (v >> 4)) & 0x0F0F0F0F...
18382 Srl = DAG.getBitcast(VT, GetShift(ISD::SRL, DAG.getBitcast(SrlVT, V), 4));
18383 SDValue Add = DAG.getNode(ISD::ADD, DL, VT, V, Srl);
18384 V = GetMask(Add, APInt::getSplat(Len, APInt(8, 0x0F)));
18386 // At this point, V contains the byte-wise population count, and we are
18387 // merely doing a horizontal sum if necessary to get the wider element
18389 if (EltVT == MVT::i8)
18392 return LowerHorizontalByteSum(
18393 DAG.getBitcast(MVT::getVectorVT(MVT::i8, VecSize / 8), V), VT, Subtarget,
18397 static SDValue LowerVectorCTPOP(SDValue Op, const X86Subtarget *Subtarget,
18398 SelectionDAG &DAG) {
18399 MVT VT = Op.getSimpleValueType();
18400 // FIXME: Need to add AVX-512 support here!
18401 assert((VT.is256BitVector() || VT.is128BitVector()) &&
18402 "Unknown CTPOP type to handle");
18403 SDLoc DL(Op.getNode());
18404 SDValue Op0 = Op.getOperand(0);
18406 if (!Subtarget->hasSSSE3()) {
18407 // We can't use the fast LUT approach, so fall back on vectorized bitmath.
18408 assert(VT.is128BitVector() && "Only 128-bit vectors supported in SSE!");
18409 return LowerVectorCTPOPBitmath(Op0, DL, Subtarget, DAG);
18412 if (VT.is256BitVector() && !Subtarget->hasInt256()) {
18413 unsigned NumElems = VT.getVectorNumElements();
18415 // Extract each 128-bit vector, compute pop count and concat the result.
18416 SDValue LHS = Extract128BitVector(Op0, 0, DAG, DL);
18417 SDValue RHS = Extract128BitVector(Op0, NumElems/2, DAG, DL);
18419 return DAG.getNode(ISD::CONCAT_VECTORS, DL, VT,
18420 LowerVectorCTPOPInRegLUT(LHS, DL, Subtarget, DAG),
18421 LowerVectorCTPOPInRegLUT(RHS, DL, Subtarget, DAG));
18424 return LowerVectorCTPOPInRegLUT(Op0, DL, Subtarget, DAG);
18427 static SDValue LowerCTPOP(SDValue Op, const X86Subtarget *Subtarget,
18428 SelectionDAG &DAG) {
18429 assert(Op.getValueType().isVector() &&
18430 "We only do custom lowering for vector population count.");
18431 return LowerVectorCTPOP(Op, Subtarget, DAG);
18434 static SDValue LowerLOAD_SUB(SDValue Op, SelectionDAG &DAG) {
18435 SDNode *Node = Op.getNode();
18437 EVT T = Node->getValueType(0);
18438 SDValue negOp = DAG.getNode(ISD::SUB, dl, T,
18439 DAG.getConstant(0, dl, T), Node->getOperand(2));
18440 return DAG.getAtomic(ISD::ATOMIC_LOAD_ADD, dl,
18441 cast<AtomicSDNode>(Node)->getMemoryVT(),
18442 Node->getOperand(0),
18443 Node->getOperand(1), negOp,
18444 cast<AtomicSDNode>(Node)->getMemOperand(),
18445 cast<AtomicSDNode>(Node)->getOrdering(),
18446 cast<AtomicSDNode>(Node)->getSynchScope());
18449 static SDValue LowerATOMIC_STORE(SDValue Op, SelectionDAG &DAG) {
18450 SDNode *Node = Op.getNode();
18452 EVT VT = cast<AtomicSDNode>(Node)->getMemoryVT();
18454 // Convert seq_cst store -> xchg
18455 // Convert wide store -> swap (-> cmpxchg8b/cmpxchg16b)
18456 // FIXME: On 32-bit, store -> fist or movq would be more efficient
18457 // (The only way to get a 16-byte store is cmpxchg16b)
18458 // FIXME: 16-byte ATOMIC_SWAP isn't actually hooked up at the moment.
18459 if (cast<AtomicSDNode>(Node)->getOrdering() == SequentiallyConsistent ||
18460 !DAG.getTargetLoweringInfo().isTypeLegal(VT)) {
18461 SDValue Swap = DAG.getAtomic(ISD::ATOMIC_SWAP, dl,
18462 cast<AtomicSDNode>(Node)->getMemoryVT(),
18463 Node->getOperand(0),
18464 Node->getOperand(1), Node->getOperand(2),
18465 cast<AtomicSDNode>(Node)->getMemOperand(),
18466 cast<AtomicSDNode>(Node)->getOrdering(),
18467 cast<AtomicSDNode>(Node)->getSynchScope());
18468 return Swap.getValue(1);
18470 // Other atomic stores have a simple pattern.
18474 static SDValue LowerADDC_ADDE_SUBC_SUBE(SDValue Op, SelectionDAG &DAG) {
18475 EVT VT = Op.getNode()->getSimpleValueType(0);
18477 // Let legalize expand this if it isn't a legal type yet.
18478 if (!DAG.getTargetLoweringInfo().isTypeLegal(VT))
18481 SDVTList VTs = DAG.getVTList(VT, MVT::i32);
18484 bool ExtraOp = false;
18485 switch (Op.getOpcode()) {
18486 default: llvm_unreachable("Invalid code");
18487 case ISD::ADDC: Opc = X86ISD::ADD; break;
18488 case ISD::ADDE: Opc = X86ISD::ADC; ExtraOp = true; break;
18489 case ISD::SUBC: Opc = X86ISD::SUB; break;
18490 case ISD::SUBE: Opc = X86ISD::SBB; ExtraOp = true; break;
18494 return DAG.getNode(Opc, SDLoc(Op), VTs, Op.getOperand(0),
18496 return DAG.getNode(Opc, SDLoc(Op), VTs, Op.getOperand(0),
18497 Op.getOperand(1), Op.getOperand(2));
18500 static SDValue LowerFSINCOS(SDValue Op, const X86Subtarget *Subtarget,
18501 SelectionDAG &DAG) {
18502 assert(Subtarget->isTargetDarwin() && Subtarget->is64Bit());
18504 // For MacOSX, we want to call an alternative entry point: __sincos_stret,
18505 // which returns the values as { float, float } (in XMM0) or
18506 // { double, double } (which is returned in XMM0, XMM1).
18508 SDValue Arg = Op.getOperand(0);
18509 EVT ArgVT = Arg.getValueType();
18510 Type *ArgTy = ArgVT.getTypeForEVT(*DAG.getContext());
18512 TargetLowering::ArgListTy Args;
18513 TargetLowering::ArgListEntry Entry;
18517 Entry.isSExt = false;
18518 Entry.isZExt = false;
18519 Args.push_back(Entry);
18521 bool isF64 = ArgVT == MVT::f64;
18522 // Only optimize x86_64 for now. i386 is a bit messy. For f32,
18523 // the small struct {f32, f32} is returned in (eax, edx). For f64,
18524 // the results are returned via SRet in memory.
18525 const char *LibcallName = isF64 ? "__sincos_stret" : "__sincosf_stret";
18526 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
18528 DAG.getExternalSymbol(LibcallName, TLI.getPointerTy(DAG.getDataLayout()));
18530 Type *RetTy = isF64
18531 ? (Type*)StructType::get(ArgTy, ArgTy, nullptr)
18532 : (Type*)VectorType::get(ArgTy, 4);
18534 TargetLowering::CallLoweringInfo CLI(DAG);
18535 CLI.setDebugLoc(dl).setChain(DAG.getEntryNode())
18536 .setCallee(CallingConv::C, RetTy, Callee, std::move(Args), 0);
18538 std::pair<SDValue, SDValue> CallResult = TLI.LowerCallTo(CLI);
18541 // Returned in xmm0 and xmm1.
18542 return CallResult.first;
18544 // Returned in bits 0:31 and 32:64 xmm0.
18545 SDValue SinVal = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, ArgVT,
18546 CallResult.first, DAG.getIntPtrConstant(0, dl));
18547 SDValue CosVal = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, ArgVT,
18548 CallResult.first, DAG.getIntPtrConstant(1, dl));
18549 SDVTList Tys = DAG.getVTList(ArgVT, ArgVT);
18550 return DAG.getNode(ISD::MERGE_VALUES, dl, Tys, SinVal, CosVal);
18553 static SDValue LowerMSCATTER(SDValue Op, const X86Subtarget *Subtarget,
18554 SelectionDAG &DAG) {
18555 assert(Subtarget->hasAVX512() &&
18556 "MGATHER/MSCATTER are supported on AVX-512 arch only");
18558 MaskedScatterSDNode *N = cast<MaskedScatterSDNode>(Op.getNode());
18559 EVT VT = N->getValue().getValueType();
18560 assert(VT.getScalarSizeInBits() >= 32 && "Unsupported scatter op");
18563 // X86 scatter kills mask register, so its type should be added to
18564 // the list of return values
18565 if (N->getNumValues() == 1) {
18566 SDValue Index = N->getIndex();
18567 if (!Subtarget->hasVLX() && !VT.is512BitVector() &&
18568 !Index.getValueType().is512BitVector())
18569 Index = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v8i64, Index);
18571 SDVTList VTs = DAG.getVTList(N->getMask().getValueType(), MVT::Other);
18572 SDValue Ops[] = { N->getOperand(0), N->getOperand(1), N->getOperand(2),
18573 N->getOperand(3), Index };
18575 SDValue NewScatter = DAG.getMaskedScatter(VTs, VT, dl, Ops, N->getMemOperand());
18576 DAG.ReplaceAllUsesWith(Op, SDValue(NewScatter.getNode(), 1));
18577 return SDValue(NewScatter.getNode(), 0);
18582 static SDValue LowerMGATHER(SDValue Op, const X86Subtarget *Subtarget,
18583 SelectionDAG &DAG) {
18584 assert(Subtarget->hasAVX512() &&
18585 "MGATHER/MSCATTER are supported on AVX-512 arch only");
18587 MaskedGatherSDNode *N = cast<MaskedGatherSDNode>(Op.getNode());
18588 EVT VT = Op.getValueType();
18589 assert(VT.getScalarSizeInBits() >= 32 && "Unsupported gather op");
18592 SDValue Index = N->getIndex();
18593 if (!Subtarget->hasVLX() && !VT.is512BitVector() &&
18594 !Index.getValueType().is512BitVector()) {
18595 Index = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v8i64, Index);
18596 SDValue Ops[] = { N->getOperand(0), N->getOperand(1), N->getOperand(2),
18597 N->getOperand(3), Index };
18598 DAG.UpdateNodeOperands(N, Ops);
18603 SDValue X86TargetLowering::LowerGC_TRANSITION_START(SDValue Op,
18604 SelectionDAG &DAG) const {
18605 // TODO: Eventually, the lowering of these nodes should be informed by or
18606 // deferred to the GC strategy for the function in which they appear. For
18607 // now, however, they must be lowered to something. Since they are logically
18608 // no-ops in the case of a null GC strategy (or a GC strategy which does not
18609 // require special handling for these nodes), lower them as literal NOOPs for
18611 SmallVector<SDValue, 2> Ops;
18613 Ops.push_back(Op.getOperand(0));
18614 if (Op->getGluedNode())
18615 Ops.push_back(Op->getOperand(Op->getNumOperands() - 1));
18618 SDVTList VTs = DAG.getVTList(MVT::Other, MVT::Glue);
18619 SDValue NOOP(DAG.getMachineNode(X86::NOOP, SDLoc(Op), VTs, Ops), 0);
18624 SDValue X86TargetLowering::LowerGC_TRANSITION_END(SDValue Op,
18625 SelectionDAG &DAG) const {
18626 // TODO: Eventually, the lowering of these nodes should be informed by or
18627 // deferred to the GC strategy for the function in which they appear. For
18628 // now, however, they must be lowered to something. Since they are logically
18629 // no-ops in the case of a null GC strategy (or a GC strategy which does not
18630 // require special handling for these nodes), lower them as literal NOOPs for
18632 SmallVector<SDValue, 2> Ops;
18634 Ops.push_back(Op.getOperand(0));
18635 if (Op->getGluedNode())
18636 Ops.push_back(Op->getOperand(Op->getNumOperands() - 1));
18639 SDVTList VTs = DAG.getVTList(MVT::Other, MVT::Glue);
18640 SDValue NOOP(DAG.getMachineNode(X86::NOOP, SDLoc(Op), VTs, Ops), 0);
18645 /// LowerOperation - Provide custom lowering hooks for some operations.
18647 SDValue X86TargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
18648 switch (Op.getOpcode()) {
18649 default: llvm_unreachable("Should not custom lower this!");
18650 case ISD::ATOMIC_FENCE: return LowerATOMIC_FENCE(Op, Subtarget, DAG);
18651 case ISD::ATOMIC_CMP_SWAP_WITH_SUCCESS:
18652 return LowerCMP_SWAP(Op, Subtarget, DAG);
18653 case ISD::CTPOP: return LowerCTPOP(Op, Subtarget, DAG);
18654 case ISD::ATOMIC_LOAD_SUB: return LowerLOAD_SUB(Op,DAG);
18655 case ISD::ATOMIC_STORE: return LowerATOMIC_STORE(Op,DAG);
18656 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG);
18657 case ISD::CONCAT_VECTORS: return LowerCONCAT_VECTORS(Op, Subtarget, DAG);
18658 case ISD::VECTOR_SHUFFLE: return lowerVectorShuffle(Op, Subtarget, DAG);
18659 case ISD::VSELECT: return LowerVSELECT(Op, DAG);
18660 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
18661 case ISD::INSERT_VECTOR_ELT: return LowerINSERT_VECTOR_ELT(Op, DAG);
18662 case ISD::EXTRACT_SUBVECTOR: return LowerEXTRACT_SUBVECTOR(Op,Subtarget,DAG);
18663 case ISD::INSERT_SUBVECTOR: return LowerINSERT_SUBVECTOR(Op, Subtarget,DAG);
18664 case ISD::SCALAR_TO_VECTOR: return LowerSCALAR_TO_VECTOR(Op, DAG);
18665 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
18666 case ISD::GlobalAddress: return LowerGlobalAddress(Op, DAG);
18667 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
18668 case ISD::ExternalSymbol: return LowerExternalSymbol(Op, DAG);
18669 case ISD::BlockAddress: return LowerBlockAddress(Op, DAG);
18670 case ISD::SHL_PARTS:
18671 case ISD::SRA_PARTS:
18672 case ISD::SRL_PARTS: return LowerShiftParts(Op, DAG);
18673 case ISD::SINT_TO_FP: return LowerSINT_TO_FP(Op, DAG);
18674 case ISD::UINT_TO_FP: return LowerUINT_TO_FP(Op, DAG);
18675 case ISD::TRUNCATE: return LowerTRUNCATE(Op, DAG);
18676 case ISD::ZERO_EXTEND: return LowerZERO_EXTEND(Op, Subtarget, DAG);
18677 case ISD::SIGN_EXTEND: return LowerSIGN_EXTEND(Op, Subtarget, DAG);
18678 case ISD::ANY_EXTEND: return LowerANY_EXTEND(Op, Subtarget, DAG);
18679 case ISD::SIGN_EXTEND_VECTOR_INREG:
18680 return LowerSIGN_EXTEND_VECTOR_INREG(Op, Subtarget, DAG);
18681 case ISD::FP_TO_SINT: return LowerFP_TO_SINT(Op, DAG);
18682 case ISD::FP_TO_UINT: return LowerFP_TO_UINT(Op, DAG);
18683 case ISD::FP_EXTEND: return LowerFP_EXTEND(Op, DAG);
18684 case ISD::LOAD: return LowerExtendedLoad(Op, Subtarget, DAG);
18686 case ISD::FNEG: return LowerFABSorFNEG(Op, DAG);
18687 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
18688 case ISD::FGETSIGN: return LowerFGETSIGN(Op, DAG);
18689 case ISD::SETCC: return LowerSETCC(Op, DAG);
18690 case ISD::SELECT: return LowerSELECT(Op, DAG);
18691 case ISD::BRCOND: return LowerBRCOND(Op, DAG);
18692 case ISD::JumpTable: return LowerJumpTable(Op, DAG);
18693 case ISD::VASTART: return LowerVASTART(Op, DAG);
18694 case ISD::VAARG: return LowerVAARG(Op, DAG);
18695 case ISD::VACOPY: return LowerVACOPY(Op, Subtarget, DAG);
18696 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, Subtarget, DAG);
18697 case ISD::INTRINSIC_VOID:
18698 case ISD::INTRINSIC_W_CHAIN: return LowerINTRINSIC_W_CHAIN(Op, Subtarget, DAG);
18699 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
18700 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
18701 case ISD::FRAME_TO_ARGS_OFFSET:
18702 return LowerFRAME_TO_ARGS_OFFSET(Op, DAG);
18703 case ISD::DYNAMIC_STACKALLOC: return LowerDYNAMIC_STACKALLOC(Op, DAG);
18704 case ISD::EH_RETURN: return LowerEH_RETURN(Op, DAG);
18705 case ISD::EH_SJLJ_SETJMP: return lowerEH_SJLJ_SETJMP(Op, DAG);
18706 case ISD::EH_SJLJ_LONGJMP: return lowerEH_SJLJ_LONGJMP(Op, DAG);
18707 case ISD::INIT_TRAMPOLINE: return LowerINIT_TRAMPOLINE(Op, DAG);
18708 case ISD::ADJUST_TRAMPOLINE: return LowerADJUST_TRAMPOLINE(Op, DAG);
18709 case ISD::FLT_ROUNDS_: return LowerFLT_ROUNDS_(Op, DAG);
18710 case ISD::CTLZ: return LowerCTLZ(Op, DAG);
18711 case ISD::CTLZ_ZERO_UNDEF: return LowerCTLZ_ZERO_UNDEF(Op, DAG);
18712 case ISD::CTTZ: return LowerCTTZ(Op, DAG);
18713 case ISD::MUL: return LowerMUL(Op, Subtarget, DAG);
18714 case ISD::UMUL_LOHI:
18715 case ISD::SMUL_LOHI: return LowerMUL_LOHI(Op, Subtarget, DAG);
18718 case ISD::SHL: return LowerShift(Op, Subtarget, DAG);
18724 case ISD::UMULO: return LowerXALUO(Op, DAG);
18725 case ISD::READCYCLECOUNTER: return LowerREADCYCLECOUNTER(Op, Subtarget,DAG);
18726 case ISD::BITCAST: return LowerBITCAST(Op, Subtarget, DAG);
18730 case ISD::SUBE: return LowerADDC_ADDE_SUBC_SUBE(Op, DAG);
18731 case ISD::ADD: return LowerADD(Op, DAG);
18732 case ISD::SUB: return LowerSUB(Op, DAG);
18733 case ISD::FSINCOS: return LowerFSINCOS(Op, Subtarget, DAG);
18734 case ISD::MGATHER: return LowerMGATHER(Op, Subtarget, DAG);
18735 case ISD::MSCATTER: return LowerMSCATTER(Op, Subtarget, DAG);
18736 case ISD::GC_TRANSITION_START:
18737 return LowerGC_TRANSITION_START(Op, DAG);
18738 case ISD::GC_TRANSITION_END: return LowerGC_TRANSITION_END(Op, DAG);
18742 /// ReplaceNodeResults - Replace a node with an illegal result type
18743 /// with a new node built out of custom code.
18744 void X86TargetLowering::ReplaceNodeResults(SDNode *N,
18745 SmallVectorImpl<SDValue>&Results,
18746 SelectionDAG &DAG) const {
18748 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
18749 switch (N->getOpcode()) {
18751 llvm_unreachable("Do not know how to custom type legalize this operation!");
18752 // We might have generated v2f32 FMIN/FMAX operations. Widen them to v4f32.
18753 case X86ISD::FMINC:
18755 case X86ISD::FMAXC:
18756 case X86ISD::FMAX: {
18757 EVT VT = N->getValueType(0);
18758 if (VT != MVT::v2f32)
18759 llvm_unreachable("Unexpected type (!= v2f32) on FMIN/FMAX.");
18760 SDValue UNDEF = DAG.getUNDEF(VT);
18761 SDValue LHS = DAG.getNode(ISD::CONCAT_VECTORS, dl, MVT::v4f32,
18762 N->getOperand(0), UNDEF);
18763 SDValue RHS = DAG.getNode(ISD::CONCAT_VECTORS, dl, MVT::v4f32,
18764 N->getOperand(1), UNDEF);
18765 Results.push_back(DAG.getNode(N->getOpcode(), dl, MVT::v4f32, LHS, RHS));
18768 case ISD::SIGN_EXTEND_INREG:
18773 // We don't want to expand or promote these.
18780 case ISD::UDIVREM: {
18781 SDValue V = LowerWin64_i128OP(SDValue(N,0), DAG);
18782 Results.push_back(V);
18785 case ISD::FP_TO_SINT:
18786 // FP_TO_INT*_IN_MEM is not legal for f16 inputs. Do not convert
18787 // (FP_TO_SINT (load f16)) to FP_TO_INT*.
18788 if (N->getOperand(0).getValueType() == MVT::f16)
18791 case ISD::FP_TO_UINT: {
18792 bool IsSigned = N->getOpcode() == ISD::FP_TO_SINT;
18794 if (!IsSigned && !isIntegerTypeFTOL(SDValue(N, 0).getValueType()))
18797 std::pair<SDValue,SDValue> Vals =
18798 FP_TO_INTHelper(SDValue(N, 0), DAG, IsSigned, /*IsReplace=*/ true);
18799 SDValue FIST = Vals.first, StackSlot = Vals.second;
18800 if (FIST.getNode()) {
18801 EVT VT = N->getValueType(0);
18802 // Return a load from the stack slot.
18803 if (StackSlot.getNode())
18804 Results.push_back(DAG.getLoad(VT, dl, FIST, StackSlot,
18805 MachinePointerInfo(),
18806 false, false, false, 0));
18808 Results.push_back(FIST);
18812 case ISD::UINT_TO_FP: {
18813 assert(Subtarget->hasSSE2() && "Requires at least SSE2!");
18814 if (N->getOperand(0).getValueType() != MVT::v2i32 ||
18815 N->getValueType(0) != MVT::v2f32)
18817 SDValue ZExtIn = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v2i64,
18819 SDValue Bias = DAG.getConstantFP(BitsToDouble(0x4330000000000000ULL), dl,
18821 SDValue VBias = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v2f64, Bias, Bias);
18822 SDValue Or = DAG.getNode(ISD::OR, dl, MVT::v2i64, ZExtIn,
18823 DAG.getBitcast(MVT::v2i64, VBias));
18824 Or = DAG.getBitcast(MVT::v2f64, Or);
18825 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::v2f64, Or, VBias);
18826 Results.push_back(DAG.getNode(X86ISD::VFPROUND, dl, MVT::v4f32, Sub));
18829 case ISD::FP_ROUND: {
18830 if (!TLI.isTypeLegal(N->getOperand(0).getValueType()))
18832 SDValue V = DAG.getNode(X86ISD::VFPROUND, dl, MVT::v4f32, N->getOperand(0));
18833 Results.push_back(V);
18836 case ISD::FP_EXTEND: {
18837 // Right now, only MVT::v2f32 has OperationAction for FP_EXTEND.
18838 // No other ValueType for FP_EXTEND should reach this point.
18839 assert(N->getValueType(0) == MVT::v2f32 &&
18840 "Do not know how to legalize this Node");
18843 case ISD::INTRINSIC_W_CHAIN: {
18844 unsigned IntNo = cast<ConstantSDNode>(N->getOperand(1))->getZExtValue();
18846 default : llvm_unreachable("Do not know how to custom type "
18847 "legalize this intrinsic operation!");
18848 case Intrinsic::x86_rdtsc:
18849 return getReadTimeStampCounter(N, dl, X86ISD::RDTSC_DAG, DAG, Subtarget,
18851 case Intrinsic::x86_rdtscp:
18852 return getReadTimeStampCounter(N, dl, X86ISD::RDTSCP_DAG, DAG, Subtarget,
18854 case Intrinsic::x86_rdpmc:
18855 return getReadPerformanceCounter(N, dl, DAG, Subtarget, Results);
18858 case ISD::READCYCLECOUNTER: {
18859 return getReadTimeStampCounter(N, dl, X86ISD::RDTSC_DAG, DAG, Subtarget,
18862 case ISD::ATOMIC_CMP_SWAP_WITH_SUCCESS: {
18863 EVT T = N->getValueType(0);
18864 assert((T == MVT::i64 || T == MVT::i128) && "can only expand cmpxchg pair");
18865 bool Regs64bit = T == MVT::i128;
18866 EVT HalfT = Regs64bit ? MVT::i64 : MVT::i32;
18867 SDValue cpInL, cpInH;
18868 cpInL = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, HalfT, N->getOperand(2),
18869 DAG.getConstant(0, dl, HalfT));
18870 cpInH = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, HalfT, N->getOperand(2),
18871 DAG.getConstant(1, dl, HalfT));
18872 cpInL = DAG.getCopyToReg(N->getOperand(0), dl,
18873 Regs64bit ? X86::RAX : X86::EAX,
18875 cpInH = DAG.getCopyToReg(cpInL.getValue(0), dl,
18876 Regs64bit ? X86::RDX : X86::EDX,
18877 cpInH, cpInL.getValue(1));
18878 SDValue swapInL, swapInH;
18879 swapInL = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, HalfT, N->getOperand(3),
18880 DAG.getConstant(0, dl, HalfT));
18881 swapInH = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, HalfT, N->getOperand(3),
18882 DAG.getConstant(1, dl, HalfT));
18883 swapInL = DAG.getCopyToReg(cpInH.getValue(0), dl,
18884 Regs64bit ? X86::RBX : X86::EBX,
18885 swapInL, cpInH.getValue(1));
18886 swapInH = DAG.getCopyToReg(swapInL.getValue(0), dl,
18887 Regs64bit ? X86::RCX : X86::ECX,
18888 swapInH, swapInL.getValue(1));
18889 SDValue Ops[] = { swapInH.getValue(0),
18891 swapInH.getValue(1) };
18892 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Glue);
18893 MachineMemOperand *MMO = cast<AtomicSDNode>(N)->getMemOperand();
18894 unsigned Opcode = Regs64bit ? X86ISD::LCMPXCHG16_DAG :
18895 X86ISD::LCMPXCHG8_DAG;
18896 SDValue Result = DAG.getMemIntrinsicNode(Opcode, dl, Tys, Ops, T, MMO);
18897 SDValue cpOutL = DAG.getCopyFromReg(Result.getValue(0), dl,
18898 Regs64bit ? X86::RAX : X86::EAX,
18899 HalfT, Result.getValue(1));
18900 SDValue cpOutH = DAG.getCopyFromReg(cpOutL.getValue(1), dl,
18901 Regs64bit ? X86::RDX : X86::EDX,
18902 HalfT, cpOutL.getValue(2));
18903 SDValue OpsF[] = { cpOutL.getValue(0), cpOutH.getValue(0)};
18905 SDValue EFLAGS = DAG.getCopyFromReg(cpOutH.getValue(1), dl, X86::EFLAGS,
18906 MVT::i32, cpOutH.getValue(2));
18908 DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
18909 DAG.getConstant(X86::COND_E, dl, MVT::i8), EFLAGS);
18910 Success = DAG.getZExtOrTrunc(Success, dl, N->getValueType(1));
18912 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, T, OpsF));
18913 Results.push_back(Success);
18914 Results.push_back(EFLAGS.getValue(1));
18917 case ISD::ATOMIC_SWAP:
18918 case ISD::ATOMIC_LOAD_ADD:
18919 case ISD::ATOMIC_LOAD_SUB:
18920 case ISD::ATOMIC_LOAD_AND:
18921 case ISD::ATOMIC_LOAD_OR:
18922 case ISD::ATOMIC_LOAD_XOR:
18923 case ISD::ATOMIC_LOAD_NAND:
18924 case ISD::ATOMIC_LOAD_MIN:
18925 case ISD::ATOMIC_LOAD_MAX:
18926 case ISD::ATOMIC_LOAD_UMIN:
18927 case ISD::ATOMIC_LOAD_UMAX:
18928 case ISD::ATOMIC_LOAD: {
18929 // Delegate to generic TypeLegalization. Situations we can really handle
18930 // should have already been dealt with by AtomicExpandPass.cpp.
18933 case ISD::BITCAST: {
18934 assert(Subtarget->hasSSE2() && "Requires at least SSE2!");
18935 EVT DstVT = N->getValueType(0);
18936 EVT SrcVT = N->getOperand(0)->getValueType(0);
18938 if (SrcVT != MVT::f64 ||
18939 (DstVT != MVT::v2i32 && DstVT != MVT::v4i16 && DstVT != MVT::v8i8))
18942 unsigned NumElts = DstVT.getVectorNumElements();
18943 EVT SVT = DstVT.getVectorElementType();
18944 EVT WiderVT = EVT::getVectorVT(*DAG.getContext(), SVT, NumElts * 2);
18945 SDValue Expanded = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
18946 MVT::v2f64, N->getOperand(0));
18947 SDValue ToVecInt = DAG.getBitcast(WiderVT, Expanded);
18949 if (ExperimentalVectorWideningLegalization) {
18950 // If we are legalizing vectors by widening, we already have the desired
18951 // legal vector type, just return it.
18952 Results.push_back(ToVecInt);
18956 SmallVector<SDValue, 8> Elts;
18957 for (unsigned i = 0, e = NumElts; i != e; ++i)
18958 Elts.push_back(DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, SVT,
18959 ToVecInt, DAG.getIntPtrConstant(i, dl)));
18961 Results.push_back(DAG.getNode(ISD::BUILD_VECTOR, dl, DstVT, Elts));
18966 const char *X86TargetLowering::getTargetNodeName(unsigned Opcode) const {
18967 switch ((X86ISD::NodeType)Opcode) {
18968 case X86ISD::FIRST_NUMBER: break;
18969 case X86ISD::BSF: return "X86ISD::BSF";
18970 case X86ISD::BSR: return "X86ISD::BSR";
18971 case X86ISD::SHLD: return "X86ISD::SHLD";
18972 case X86ISD::SHRD: return "X86ISD::SHRD";
18973 case X86ISD::FAND: return "X86ISD::FAND";
18974 case X86ISD::FANDN: return "X86ISD::FANDN";
18975 case X86ISD::FOR: return "X86ISD::FOR";
18976 case X86ISD::FXOR: return "X86ISD::FXOR";
18977 case X86ISD::FILD: return "X86ISD::FILD";
18978 case X86ISD::FILD_FLAG: return "X86ISD::FILD_FLAG";
18979 case X86ISD::FP_TO_INT16_IN_MEM: return "X86ISD::FP_TO_INT16_IN_MEM";
18980 case X86ISD::FP_TO_INT32_IN_MEM: return "X86ISD::FP_TO_INT32_IN_MEM";
18981 case X86ISD::FP_TO_INT64_IN_MEM: return "X86ISD::FP_TO_INT64_IN_MEM";
18982 case X86ISD::FLD: return "X86ISD::FLD";
18983 case X86ISD::FST: return "X86ISD::FST";
18984 case X86ISD::CALL: return "X86ISD::CALL";
18985 case X86ISD::RDTSC_DAG: return "X86ISD::RDTSC_DAG";
18986 case X86ISD::RDTSCP_DAG: return "X86ISD::RDTSCP_DAG";
18987 case X86ISD::RDPMC_DAG: return "X86ISD::RDPMC_DAG";
18988 case X86ISD::BT: return "X86ISD::BT";
18989 case X86ISD::CMP: return "X86ISD::CMP";
18990 case X86ISD::COMI: return "X86ISD::COMI";
18991 case X86ISD::UCOMI: return "X86ISD::UCOMI";
18992 case X86ISD::CMPM: return "X86ISD::CMPM";
18993 case X86ISD::CMPMU: return "X86ISD::CMPMU";
18994 case X86ISD::CMPM_RND: return "X86ISD::CMPM_RND";
18995 case X86ISD::SETCC: return "X86ISD::SETCC";
18996 case X86ISD::SETCC_CARRY: return "X86ISD::SETCC_CARRY";
18997 case X86ISD::FSETCC: return "X86ISD::FSETCC";
18998 case X86ISD::FGETSIGNx86: return "X86ISD::FGETSIGNx86";
18999 case X86ISD::CMOV: return "X86ISD::CMOV";
19000 case X86ISD::BRCOND: return "X86ISD::BRCOND";
19001 case X86ISD::RET_FLAG: return "X86ISD::RET_FLAG";
19002 case X86ISD::REP_STOS: return "X86ISD::REP_STOS";
19003 case X86ISD::REP_MOVS: return "X86ISD::REP_MOVS";
19004 case X86ISD::GlobalBaseReg: return "X86ISD::GlobalBaseReg";
19005 case X86ISD::Wrapper: return "X86ISD::Wrapper";
19006 case X86ISD::WrapperRIP: return "X86ISD::WrapperRIP";
19007 case X86ISD::MOVDQ2Q: return "X86ISD::MOVDQ2Q";
19008 case X86ISD::MMX_MOVD2W: return "X86ISD::MMX_MOVD2W";
19009 case X86ISD::MMX_MOVW2D: return "X86ISD::MMX_MOVW2D";
19010 case X86ISD::PEXTRB: return "X86ISD::PEXTRB";
19011 case X86ISD::PEXTRW: return "X86ISD::PEXTRW";
19012 case X86ISD::INSERTPS: return "X86ISD::INSERTPS";
19013 case X86ISD::PINSRB: return "X86ISD::PINSRB";
19014 case X86ISD::PINSRW: return "X86ISD::PINSRW";
19015 case X86ISD::MMX_PINSRW: return "X86ISD::MMX_PINSRW";
19016 case X86ISD::PSHUFB: return "X86ISD::PSHUFB";
19017 case X86ISD::ANDNP: return "X86ISD::ANDNP";
19018 case X86ISD::PSIGN: return "X86ISD::PSIGN";
19019 case X86ISD::BLENDI: return "X86ISD::BLENDI";
19020 case X86ISD::SHRUNKBLEND: return "X86ISD::SHRUNKBLEND";
19021 case X86ISD::ADDUS: return "X86ISD::ADDUS";
19022 case X86ISD::SUBUS: return "X86ISD::SUBUS";
19023 case X86ISD::HADD: return "X86ISD::HADD";
19024 case X86ISD::HSUB: return "X86ISD::HSUB";
19025 case X86ISD::FHADD: return "X86ISD::FHADD";
19026 case X86ISD::FHSUB: return "X86ISD::FHSUB";
19027 case X86ISD::ABS: return "X86ISD::ABS";
19028 case X86ISD::FMAX: return "X86ISD::FMAX";
19029 case X86ISD::FMAX_RND: return "X86ISD::FMAX_RND";
19030 case X86ISD::FMIN: return "X86ISD::FMIN";
19031 case X86ISD::FMIN_RND: return "X86ISD::FMIN_RND";
19032 case X86ISD::FMAXC: return "X86ISD::FMAXC";
19033 case X86ISD::FMINC: return "X86ISD::FMINC";
19034 case X86ISD::FRSQRT: return "X86ISD::FRSQRT";
19035 case X86ISD::FRCP: return "X86ISD::FRCP";
19036 case X86ISD::EXTRQI: return "X86ISD::EXTRQI";
19037 case X86ISD::INSERTQI: return "X86ISD::INSERTQI";
19038 case X86ISD::TLSADDR: return "X86ISD::TLSADDR";
19039 case X86ISD::TLSBASEADDR: return "X86ISD::TLSBASEADDR";
19040 case X86ISD::TLSCALL: return "X86ISD::TLSCALL";
19041 case X86ISD::EH_SJLJ_SETJMP: return "X86ISD::EH_SJLJ_SETJMP";
19042 case X86ISD::EH_SJLJ_LONGJMP: return "X86ISD::EH_SJLJ_LONGJMP";
19043 case X86ISD::EH_RETURN: return "X86ISD::EH_RETURN";
19044 case X86ISD::TC_RETURN: return "X86ISD::TC_RETURN";
19045 case X86ISD::FNSTCW16m: return "X86ISD::FNSTCW16m";
19046 case X86ISD::FNSTSW16r: return "X86ISD::FNSTSW16r";
19047 case X86ISD::LCMPXCHG_DAG: return "X86ISD::LCMPXCHG_DAG";
19048 case X86ISD::LCMPXCHG8_DAG: return "X86ISD::LCMPXCHG8_DAG";
19049 case X86ISD::LCMPXCHG16_DAG: return "X86ISD::LCMPXCHG16_DAG";
19050 case X86ISD::VZEXT_MOVL: return "X86ISD::VZEXT_MOVL";
19051 case X86ISD::VZEXT_LOAD: return "X86ISD::VZEXT_LOAD";
19052 case X86ISD::VZEXT: return "X86ISD::VZEXT";
19053 case X86ISD::VSEXT: return "X86ISD::VSEXT";
19054 case X86ISD::VTRUNC: return "X86ISD::VTRUNC";
19055 case X86ISD::VTRUNCS: return "X86ISD::VTRUNCS";
19056 case X86ISD::VTRUNCUS: return "X86ISD::VTRUNCUS";
19057 case X86ISD::VINSERT: return "X86ISD::VINSERT";
19058 case X86ISD::VFPEXT: return "X86ISD::VFPEXT";
19059 case X86ISD::VFPROUND: return "X86ISD::VFPROUND";
19060 case X86ISD::CVTDQ2PD: return "X86ISD::CVTDQ2PD";
19061 case X86ISD::CVTUDQ2PD: return "X86ISD::CVTUDQ2PD";
19062 case X86ISD::VSHLDQ: return "X86ISD::VSHLDQ";
19063 case X86ISD::VSRLDQ: return "X86ISD::VSRLDQ";
19064 case X86ISD::VSHL: return "X86ISD::VSHL";
19065 case X86ISD::VSRL: return "X86ISD::VSRL";
19066 case X86ISD::VSRA: return "X86ISD::VSRA";
19067 case X86ISD::VSHLI: return "X86ISD::VSHLI";
19068 case X86ISD::VSRLI: return "X86ISD::VSRLI";
19069 case X86ISD::VSRAI: return "X86ISD::VSRAI";
19070 case X86ISD::CMPP: return "X86ISD::CMPP";
19071 case X86ISD::PCMPEQ: return "X86ISD::PCMPEQ";
19072 case X86ISD::PCMPGT: return "X86ISD::PCMPGT";
19073 case X86ISD::PCMPEQM: return "X86ISD::PCMPEQM";
19074 case X86ISD::PCMPGTM: return "X86ISD::PCMPGTM";
19075 case X86ISD::ADD: return "X86ISD::ADD";
19076 case X86ISD::SUB: return "X86ISD::SUB";
19077 case X86ISD::ADC: return "X86ISD::ADC";
19078 case X86ISD::SBB: return "X86ISD::SBB";
19079 case X86ISD::SMUL: return "X86ISD::SMUL";
19080 case X86ISD::UMUL: return "X86ISD::UMUL";
19081 case X86ISD::SMUL8: return "X86ISD::SMUL8";
19082 case X86ISD::UMUL8: return "X86ISD::UMUL8";
19083 case X86ISD::SDIVREM8_SEXT_HREG: return "X86ISD::SDIVREM8_SEXT_HREG";
19084 case X86ISD::UDIVREM8_ZEXT_HREG: return "X86ISD::UDIVREM8_ZEXT_HREG";
19085 case X86ISD::INC: return "X86ISD::INC";
19086 case X86ISD::DEC: return "X86ISD::DEC";
19087 case X86ISD::OR: return "X86ISD::OR";
19088 case X86ISD::XOR: return "X86ISD::XOR";
19089 case X86ISD::AND: return "X86ISD::AND";
19090 case X86ISD::BEXTR: return "X86ISD::BEXTR";
19091 case X86ISD::MUL_IMM: return "X86ISD::MUL_IMM";
19092 case X86ISD::PTEST: return "X86ISD::PTEST";
19093 case X86ISD::TESTP: return "X86ISD::TESTP";
19094 case X86ISD::TESTM: return "X86ISD::TESTM";
19095 case X86ISD::TESTNM: return "X86ISD::TESTNM";
19096 case X86ISD::KORTEST: return "X86ISD::KORTEST";
19097 case X86ISD::PACKSS: return "X86ISD::PACKSS";
19098 case X86ISD::PACKUS: return "X86ISD::PACKUS";
19099 case X86ISD::PALIGNR: return "X86ISD::PALIGNR";
19100 case X86ISD::VALIGN: return "X86ISD::VALIGN";
19101 case X86ISD::PSHUFD: return "X86ISD::PSHUFD";
19102 case X86ISD::PSHUFHW: return "X86ISD::PSHUFHW";
19103 case X86ISD::PSHUFLW: return "X86ISD::PSHUFLW";
19104 case X86ISD::SHUFP: return "X86ISD::SHUFP";
19105 case X86ISD::SHUF128: return "X86ISD::SHUF128";
19106 case X86ISD::MOVLHPS: return "X86ISD::MOVLHPS";
19107 case X86ISD::MOVLHPD: return "X86ISD::MOVLHPD";
19108 case X86ISD::MOVHLPS: return "X86ISD::MOVHLPS";
19109 case X86ISD::MOVLPS: return "X86ISD::MOVLPS";
19110 case X86ISD::MOVLPD: return "X86ISD::MOVLPD";
19111 case X86ISD::MOVDDUP: return "X86ISD::MOVDDUP";
19112 case X86ISD::MOVSHDUP: return "X86ISD::MOVSHDUP";
19113 case X86ISD::MOVSLDUP: return "X86ISD::MOVSLDUP";
19114 case X86ISD::MOVSD: return "X86ISD::MOVSD";
19115 case X86ISD::MOVSS: return "X86ISD::MOVSS";
19116 case X86ISD::UNPCKL: return "X86ISD::UNPCKL";
19117 case X86ISD::UNPCKH: return "X86ISD::UNPCKH";
19118 case X86ISD::VBROADCAST: return "X86ISD::VBROADCAST";
19119 case X86ISD::SUBV_BROADCAST: return "X86ISD::SUBV_BROADCAST";
19120 case X86ISD::VEXTRACT: return "X86ISD::VEXTRACT";
19121 case X86ISD::VPERMILPV: return "X86ISD::VPERMILPV";
19122 case X86ISD::VPERMILPI: return "X86ISD::VPERMILPI";
19123 case X86ISD::VPERM2X128: return "X86ISD::VPERM2X128";
19124 case X86ISD::VPERMV: return "X86ISD::VPERMV";
19125 case X86ISD::VPERMV3: return "X86ISD::VPERMV3";
19126 case X86ISD::VPERMIV3: return "X86ISD::VPERMIV3";
19127 case X86ISD::VPERMI: return "X86ISD::VPERMI";
19128 case X86ISD::VFIXUPIMM: return "X86ISD::VFIXUPIMM";
19129 case X86ISD::VRANGE: return "X86ISD::VRANGE";
19130 case X86ISD::PMULUDQ: return "X86ISD::PMULUDQ";
19131 case X86ISD::PMULDQ: return "X86ISD::PMULDQ";
19132 case X86ISD::PSADBW: return "X86ISD::PSADBW";
19133 case X86ISD::VASTART_SAVE_XMM_REGS: return "X86ISD::VASTART_SAVE_XMM_REGS";
19134 case X86ISD::VAARG_64: return "X86ISD::VAARG_64";
19135 case X86ISD::WIN_ALLOCA: return "X86ISD::WIN_ALLOCA";
19136 case X86ISD::MEMBARRIER: return "X86ISD::MEMBARRIER";
19137 case X86ISD::MFENCE: return "X86ISD::MFENCE";
19138 case X86ISD::SFENCE: return "X86ISD::SFENCE";
19139 case X86ISD::LFENCE: return "X86ISD::LFENCE";
19140 case X86ISD::SEG_ALLOCA: return "X86ISD::SEG_ALLOCA";
19141 case X86ISD::WIN_FTOL: return "X86ISD::WIN_FTOL";
19142 case X86ISD::SAHF: return "X86ISD::SAHF";
19143 case X86ISD::RDRAND: return "X86ISD::RDRAND";
19144 case X86ISD::RDSEED: return "X86ISD::RDSEED";
19145 case X86ISD::VPMADDUBSW: return "X86ISD::VPMADDUBSW";
19146 case X86ISD::VPMADDWD: return "X86ISD::VPMADDWD";
19147 case X86ISD::FMADD: return "X86ISD::FMADD";
19148 case X86ISD::FMSUB: return "X86ISD::FMSUB";
19149 case X86ISD::FNMADD: return "X86ISD::FNMADD";
19150 case X86ISD::FNMSUB: return "X86ISD::FNMSUB";
19151 case X86ISD::FMADDSUB: return "X86ISD::FMADDSUB";
19152 case X86ISD::FMSUBADD: return "X86ISD::FMSUBADD";
19153 case X86ISD::FMADD_RND: return "X86ISD::FMADD_RND";
19154 case X86ISD::FNMADD_RND: return "X86ISD::FNMADD_RND";
19155 case X86ISD::FMSUB_RND: return "X86ISD::FMSUB_RND";
19156 case X86ISD::FNMSUB_RND: return "X86ISD::FNMSUB_RND";
19157 case X86ISD::FMADDSUB_RND: return "X86ISD::FMADDSUB_RND";
19158 case X86ISD::FMSUBADD_RND: return "X86ISD::FMSUBADD_RND";
19159 case X86ISD::VRNDSCALE: return "X86ISD::VRNDSCALE";
19160 case X86ISD::VREDUCE: return "X86ISD::VREDUCE";
19161 case X86ISD::PCMPESTRI: return "X86ISD::PCMPESTRI";
19162 case X86ISD::PCMPISTRI: return "X86ISD::PCMPISTRI";
19163 case X86ISD::XTEST: return "X86ISD::XTEST";
19164 case X86ISD::COMPRESS: return "X86ISD::COMPRESS";
19165 case X86ISD::EXPAND: return "X86ISD::EXPAND";
19166 case X86ISD::SELECT: return "X86ISD::SELECT";
19167 case X86ISD::ADDSUB: return "X86ISD::ADDSUB";
19168 case X86ISD::RCP28: return "X86ISD::RCP28";
19169 case X86ISD::EXP2: return "X86ISD::EXP2";
19170 case X86ISD::RSQRT28: return "X86ISD::RSQRT28";
19171 case X86ISD::FADD_RND: return "X86ISD::FADD_RND";
19172 case X86ISD::FSUB_RND: return "X86ISD::FSUB_RND";
19173 case X86ISD::FMUL_RND: return "X86ISD::FMUL_RND";
19174 case X86ISD::FDIV_RND: return "X86ISD::FDIV_RND";
19175 case X86ISD::FSQRT_RND: return "X86ISD::FSQRT_RND";
19176 case X86ISD::FGETEXP_RND: return "X86ISD::FGETEXP_RND";
19177 case X86ISD::SCALEF: return "X86ISD::SCALEF";
19178 case X86ISD::ADDS: return "X86ISD::ADDS";
19179 case X86ISD::SUBS: return "X86ISD::SUBS";
19180 case X86ISD::AVG: return "X86ISD::AVG";
19181 case X86ISD::MULHRS: return "X86ISD::MULHRS";
19182 case X86ISD::SINT_TO_FP_RND: return "X86ISD::SINT_TO_FP_RND";
19183 case X86ISD::UINT_TO_FP_RND: return "X86ISD::UINT_TO_FP_RND";
19184 case X86ISD::FP_TO_SINT_RND: return "X86ISD::FP_TO_SINT_RND";
19185 case X86ISD::FP_TO_UINT_RND: return "X86ISD::FP_TO_UINT_RND";
19190 // isLegalAddressingMode - Return true if the addressing mode represented
19191 // by AM is legal for this target, for a load/store of the specified type.
19192 bool X86TargetLowering::isLegalAddressingMode(const DataLayout &DL,
19193 const AddrMode &AM, Type *Ty,
19194 unsigned AS) const {
19195 // X86 supports extremely general addressing modes.
19196 CodeModel::Model M = getTargetMachine().getCodeModel();
19197 Reloc::Model R = getTargetMachine().getRelocationModel();
19199 // X86 allows a sign-extended 32-bit immediate field as a displacement.
19200 if (!X86::isOffsetSuitableForCodeModel(AM.BaseOffs, M, AM.BaseGV != nullptr))
19205 Subtarget->ClassifyGlobalReference(AM.BaseGV, getTargetMachine());
19207 // If a reference to this global requires an extra load, we can't fold it.
19208 if (isGlobalStubReference(GVFlags))
19211 // If BaseGV requires a register for the PIC base, we cannot also have a
19212 // BaseReg specified.
19213 if (AM.HasBaseReg && isGlobalRelativeToPICBase(GVFlags))
19216 // If lower 4G is not available, then we must use rip-relative addressing.
19217 if ((M != CodeModel::Small || R != Reloc::Static) &&
19218 Subtarget->is64Bit() && (AM.BaseOffs || AM.Scale > 1))
19222 switch (AM.Scale) {
19228 // These scales always work.
19233 // These scales are formed with basereg+scalereg. Only accept if there is
19238 default: // Other stuff never works.
19245 bool X86TargetLowering::isVectorShiftByScalarCheap(Type *Ty) const {
19246 unsigned Bits = Ty->getScalarSizeInBits();
19248 // 8-bit shifts are always expensive, but versions with a scalar amount aren't
19249 // particularly cheaper than those without.
19253 // On AVX2 there are new vpsllv[dq] instructions (and other shifts), that make
19254 // variable shifts just as cheap as scalar ones.
19255 if (Subtarget->hasInt256() && (Bits == 32 || Bits == 64))
19258 // Otherwise, it's significantly cheaper to shift by a scalar amount than by a
19259 // fully general vector.
19263 bool X86TargetLowering::isTruncateFree(Type *Ty1, Type *Ty2) const {
19264 if (!Ty1->isIntegerTy() || !Ty2->isIntegerTy())
19266 unsigned NumBits1 = Ty1->getPrimitiveSizeInBits();
19267 unsigned NumBits2 = Ty2->getPrimitiveSizeInBits();
19268 return NumBits1 > NumBits2;
19271 bool X86TargetLowering::allowTruncateForTailCall(Type *Ty1, Type *Ty2) const {
19272 if (!Ty1->isIntegerTy() || !Ty2->isIntegerTy())
19275 if (!isTypeLegal(EVT::getEVT(Ty1)))
19278 assert(Ty1->getPrimitiveSizeInBits() <= 64 && "i128 is probably not a noop");
19280 // Assuming the caller doesn't have a zeroext or signext return parameter,
19281 // truncation all the way down to i1 is valid.
19285 bool X86TargetLowering::isLegalICmpImmediate(int64_t Imm) const {
19286 return isInt<32>(Imm);
19289 bool X86TargetLowering::isLegalAddImmediate(int64_t Imm) const {
19290 // Can also use sub to handle negated immediates.
19291 return isInt<32>(Imm);
19294 bool X86TargetLowering::isTruncateFree(EVT VT1, EVT VT2) const {
19295 if (!VT1.isInteger() || !VT2.isInteger())
19297 unsigned NumBits1 = VT1.getSizeInBits();
19298 unsigned NumBits2 = VT2.getSizeInBits();
19299 return NumBits1 > NumBits2;
19302 bool X86TargetLowering::isZExtFree(Type *Ty1, Type *Ty2) const {
19303 // x86-64 implicitly zero-extends 32-bit results in 64-bit registers.
19304 return Ty1->isIntegerTy(32) && Ty2->isIntegerTy(64) && Subtarget->is64Bit();
19307 bool X86TargetLowering::isZExtFree(EVT VT1, EVT VT2) const {
19308 // x86-64 implicitly zero-extends 32-bit results in 64-bit registers.
19309 return VT1 == MVT::i32 && VT2 == MVT::i64 && Subtarget->is64Bit();
19312 bool X86TargetLowering::isZExtFree(SDValue Val, EVT VT2) const {
19313 EVT VT1 = Val.getValueType();
19314 if (isZExtFree(VT1, VT2))
19317 if (Val.getOpcode() != ISD::LOAD)
19320 if (!VT1.isSimple() || !VT1.isInteger() ||
19321 !VT2.isSimple() || !VT2.isInteger())
19324 switch (VT1.getSimpleVT().SimpleTy) {
19329 // X86 has 8, 16, and 32-bit zero-extending loads.
19336 bool X86TargetLowering::isVectorLoadExtDesirable(SDValue) const { return true; }
19339 X86TargetLowering::isFMAFasterThanFMulAndFAdd(EVT VT) const {
19340 if (!(Subtarget->hasFMA() || Subtarget->hasFMA4() || Subtarget->hasAVX512()))
19343 VT = VT.getScalarType();
19345 if (!VT.isSimple())
19348 switch (VT.getSimpleVT().SimpleTy) {
19359 bool X86TargetLowering::isNarrowingProfitable(EVT VT1, EVT VT2) const {
19360 // i16 instructions are longer (0x66 prefix) and potentially slower.
19361 return !(VT1 == MVT::i32 && VT2 == MVT::i16);
19364 /// isShuffleMaskLegal - Targets can use this to indicate that they only
19365 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
19366 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
19367 /// are assumed to be legal.
19369 X86TargetLowering::isShuffleMaskLegal(const SmallVectorImpl<int> &M,
19371 if (!VT.isSimple())
19374 // Not for i1 vectors
19375 if (VT.getScalarType() == MVT::i1)
19378 // Very little shuffling can be done for 64-bit vectors right now.
19379 if (VT.getSizeInBits() == 64)
19382 // We only care that the types being shuffled are legal. The lowering can
19383 // handle any possible shuffle mask that results.
19384 return isTypeLegal(VT.getSimpleVT());
19388 X86TargetLowering::isVectorClearMaskLegal(const SmallVectorImpl<int> &Mask,
19390 // Just delegate to the generic legality, clear masks aren't special.
19391 return isShuffleMaskLegal(Mask, VT);
19394 //===----------------------------------------------------------------------===//
19395 // X86 Scheduler Hooks
19396 //===----------------------------------------------------------------------===//
19398 /// Utility function to emit xbegin specifying the start of an RTM region.
19399 static MachineBasicBlock *EmitXBegin(MachineInstr *MI, MachineBasicBlock *MBB,
19400 const TargetInstrInfo *TII) {
19401 DebugLoc DL = MI->getDebugLoc();
19403 const BasicBlock *BB = MBB->getBasicBlock();
19404 MachineFunction::iterator I = MBB;
19407 // For the v = xbegin(), we generate
19418 MachineBasicBlock *thisMBB = MBB;
19419 MachineFunction *MF = MBB->getParent();
19420 MachineBasicBlock *mainMBB = MF->CreateMachineBasicBlock(BB);
19421 MachineBasicBlock *sinkMBB = MF->CreateMachineBasicBlock(BB);
19422 MF->insert(I, mainMBB);
19423 MF->insert(I, sinkMBB);
19425 // Transfer the remainder of BB and its successor edges to sinkMBB.
19426 sinkMBB->splice(sinkMBB->begin(), MBB,
19427 std::next(MachineBasicBlock::iterator(MI)), MBB->end());
19428 sinkMBB->transferSuccessorsAndUpdatePHIs(MBB);
19432 // # fallthrough to mainMBB
19433 // # abortion to sinkMBB
19434 BuildMI(thisMBB, DL, TII->get(X86::XBEGIN_4)).addMBB(sinkMBB);
19435 thisMBB->addSuccessor(mainMBB);
19436 thisMBB->addSuccessor(sinkMBB);
19440 BuildMI(mainMBB, DL, TII->get(X86::MOV32ri), X86::EAX).addImm(-1);
19441 mainMBB->addSuccessor(sinkMBB);
19444 // EAX is live into the sinkMBB
19445 sinkMBB->addLiveIn(X86::EAX);
19446 BuildMI(*sinkMBB, sinkMBB->begin(), DL,
19447 TII->get(TargetOpcode::COPY), MI->getOperand(0).getReg())
19450 MI->eraseFromParent();
19454 // FIXME: When we get size specific XMM0 registers, i.e. XMM0_V16I8
19455 // or XMM0_V32I8 in AVX all of this code can be replaced with that
19456 // in the .td file.
19457 static MachineBasicBlock *EmitPCMPSTRM(MachineInstr *MI, MachineBasicBlock *BB,
19458 const TargetInstrInfo *TII) {
19460 switch (MI->getOpcode()) {
19461 default: llvm_unreachable("illegal opcode!");
19462 case X86::PCMPISTRM128REG: Opc = X86::PCMPISTRM128rr; break;
19463 case X86::VPCMPISTRM128REG: Opc = X86::VPCMPISTRM128rr; break;
19464 case X86::PCMPISTRM128MEM: Opc = X86::PCMPISTRM128rm; break;
19465 case X86::VPCMPISTRM128MEM: Opc = X86::VPCMPISTRM128rm; break;
19466 case X86::PCMPESTRM128REG: Opc = X86::PCMPESTRM128rr; break;
19467 case X86::VPCMPESTRM128REG: Opc = X86::VPCMPESTRM128rr; break;
19468 case X86::PCMPESTRM128MEM: Opc = X86::PCMPESTRM128rm; break;
19469 case X86::VPCMPESTRM128MEM: Opc = X86::VPCMPESTRM128rm; break;
19472 DebugLoc dl = MI->getDebugLoc();
19473 MachineInstrBuilder MIB = BuildMI(*BB, MI, dl, TII->get(Opc));
19475 unsigned NumArgs = MI->getNumOperands();
19476 for (unsigned i = 1; i < NumArgs; ++i) {
19477 MachineOperand &Op = MI->getOperand(i);
19478 if (!(Op.isReg() && Op.isImplicit()))
19479 MIB.addOperand(Op);
19481 if (MI->hasOneMemOperand())
19482 MIB->setMemRefs(MI->memoperands_begin(), MI->memoperands_end());
19484 BuildMI(*BB, MI, dl,
19485 TII->get(TargetOpcode::COPY), MI->getOperand(0).getReg())
19486 .addReg(X86::XMM0);
19488 MI->eraseFromParent();
19492 // FIXME: Custom handling because TableGen doesn't support multiple implicit
19493 // defs in an instruction pattern
19494 static MachineBasicBlock *EmitPCMPSTRI(MachineInstr *MI, MachineBasicBlock *BB,
19495 const TargetInstrInfo *TII) {
19497 switch (MI->getOpcode()) {
19498 default: llvm_unreachable("illegal opcode!");
19499 case X86::PCMPISTRIREG: Opc = X86::PCMPISTRIrr; break;
19500 case X86::VPCMPISTRIREG: Opc = X86::VPCMPISTRIrr; break;
19501 case X86::PCMPISTRIMEM: Opc = X86::PCMPISTRIrm; break;
19502 case X86::VPCMPISTRIMEM: Opc = X86::VPCMPISTRIrm; break;
19503 case X86::PCMPESTRIREG: Opc = X86::PCMPESTRIrr; break;
19504 case X86::VPCMPESTRIREG: Opc = X86::VPCMPESTRIrr; break;
19505 case X86::PCMPESTRIMEM: Opc = X86::PCMPESTRIrm; break;
19506 case X86::VPCMPESTRIMEM: Opc = X86::VPCMPESTRIrm; break;
19509 DebugLoc dl = MI->getDebugLoc();
19510 MachineInstrBuilder MIB = BuildMI(*BB, MI, dl, TII->get(Opc));
19512 unsigned NumArgs = MI->getNumOperands(); // remove the results
19513 for (unsigned i = 1; i < NumArgs; ++i) {
19514 MachineOperand &Op = MI->getOperand(i);
19515 if (!(Op.isReg() && Op.isImplicit()))
19516 MIB.addOperand(Op);
19518 if (MI->hasOneMemOperand())
19519 MIB->setMemRefs(MI->memoperands_begin(), MI->memoperands_end());
19521 BuildMI(*BB, MI, dl,
19522 TII->get(TargetOpcode::COPY), MI->getOperand(0).getReg())
19525 MI->eraseFromParent();
19529 static MachineBasicBlock *EmitMonitor(MachineInstr *MI, MachineBasicBlock *BB,
19530 const X86Subtarget *Subtarget) {
19531 DebugLoc dl = MI->getDebugLoc();
19532 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
19533 // Address into RAX/EAX, other two args into ECX, EDX.
19534 unsigned MemOpc = Subtarget->is64Bit() ? X86::LEA64r : X86::LEA32r;
19535 unsigned MemReg = Subtarget->is64Bit() ? X86::RAX : X86::EAX;
19536 MachineInstrBuilder MIB = BuildMI(*BB, MI, dl, TII->get(MemOpc), MemReg);
19537 for (int i = 0; i < X86::AddrNumOperands; ++i)
19538 MIB.addOperand(MI->getOperand(i));
19540 unsigned ValOps = X86::AddrNumOperands;
19541 BuildMI(*BB, MI, dl, TII->get(TargetOpcode::COPY), X86::ECX)
19542 .addReg(MI->getOperand(ValOps).getReg());
19543 BuildMI(*BB, MI, dl, TII->get(TargetOpcode::COPY), X86::EDX)
19544 .addReg(MI->getOperand(ValOps+1).getReg());
19546 // The instruction doesn't actually take any operands though.
19547 BuildMI(*BB, MI, dl, TII->get(X86::MONITORrrr));
19549 MI->eraseFromParent(); // The pseudo is gone now.
19553 MachineBasicBlock *
19554 X86TargetLowering::EmitVAARG64WithCustomInserter(MachineInstr *MI,
19555 MachineBasicBlock *MBB) const {
19556 // Emit va_arg instruction on X86-64.
19558 // Operands to this pseudo-instruction:
19559 // 0 ) Output : destination address (reg)
19560 // 1-5) Input : va_list address (addr, i64mem)
19561 // 6 ) ArgSize : Size (in bytes) of vararg type
19562 // 7 ) ArgMode : 0=overflow only, 1=use gp_offset, 2=use fp_offset
19563 // 8 ) Align : Alignment of type
19564 // 9 ) EFLAGS (implicit-def)
19566 assert(MI->getNumOperands() == 10 && "VAARG_64 should have 10 operands!");
19567 static_assert(X86::AddrNumOperands == 5,
19568 "VAARG_64 assumes 5 address operands");
19570 unsigned DestReg = MI->getOperand(0).getReg();
19571 MachineOperand &Base = MI->getOperand(1);
19572 MachineOperand &Scale = MI->getOperand(2);
19573 MachineOperand &Index = MI->getOperand(3);
19574 MachineOperand &Disp = MI->getOperand(4);
19575 MachineOperand &Segment = MI->getOperand(5);
19576 unsigned ArgSize = MI->getOperand(6).getImm();
19577 unsigned ArgMode = MI->getOperand(7).getImm();
19578 unsigned Align = MI->getOperand(8).getImm();
19580 // Memory Reference
19581 assert(MI->hasOneMemOperand() && "Expected VAARG_64 to have one memoperand");
19582 MachineInstr::mmo_iterator MMOBegin = MI->memoperands_begin();
19583 MachineInstr::mmo_iterator MMOEnd = MI->memoperands_end();
19585 // Machine Information
19586 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
19587 MachineRegisterInfo &MRI = MBB->getParent()->getRegInfo();
19588 const TargetRegisterClass *AddrRegClass = getRegClassFor(MVT::i64);
19589 const TargetRegisterClass *OffsetRegClass = getRegClassFor(MVT::i32);
19590 DebugLoc DL = MI->getDebugLoc();
19592 // struct va_list {
19595 // i64 overflow_area (address)
19596 // i64 reg_save_area (address)
19598 // sizeof(va_list) = 24
19599 // alignment(va_list) = 8
19601 unsigned TotalNumIntRegs = 6;
19602 unsigned TotalNumXMMRegs = 8;
19603 bool UseGPOffset = (ArgMode == 1);
19604 bool UseFPOffset = (ArgMode == 2);
19605 unsigned MaxOffset = TotalNumIntRegs * 8 +
19606 (UseFPOffset ? TotalNumXMMRegs * 16 : 0);
19608 /* Align ArgSize to a multiple of 8 */
19609 unsigned ArgSizeA8 = (ArgSize + 7) & ~7;
19610 bool NeedsAlign = (Align > 8);
19612 MachineBasicBlock *thisMBB = MBB;
19613 MachineBasicBlock *overflowMBB;
19614 MachineBasicBlock *offsetMBB;
19615 MachineBasicBlock *endMBB;
19617 unsigned OffsetDestReg = 0; // Argument address computed by offsetMBB
19618 unsigned OverflowDestReg = 0; // Argument address computed by overflowMBB
19619 unsigned OffsetReg = 0;
19621 if (!UseGPOffset && !UseFPOffset) {
19622 // If we only pull from the overflow region, we don't create a branch.
19623 // We don't need to alter control flow.
19624 OffsetDestReg = 0; // unused
19625 OverflowDestReg = DestReg;
19627 offsetMBB = nullptr;
19628 overflowMBB = thisMBB;
19631 // First emit code to check if gp_offset (or fp_offset) is below the bound.
19632 // If so, pull the argument from reg_save_area. (branch to offsetMBB)
19633 // If not, pull from overflow_area. (branch to overflowMBB)
19638 // offsetMBB overflowMBB
19643 // Registers for the PHI in endMBB
19644 OffsetDestReg = MRI.createVirtualRegister(AddrRegClass);
19645 OverflowDestReg = MRI.createVirtualRegister(AddrRegClass);
19647 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
19648 MachineFunction *MF = MBB->getParent();
19649 overflowMBB = MF->CreateMachineBasicBlock(LLVM_BB);
19650 offsetMBB = MF->CreateMachineBasicBlock(LLVM_BB);
19651 endMBB = MF->CreateMachineBasicBlock(LLVM_BB);
19653 MachineFunction::iterator MBBIter = MBB;
19656 // Insert the new basic blocks
19657 MF->insert(MBBIter, offsetMBB);
19658 MF->insert(MBBIter, overflowMBB);
19659 MF->insert(MBBIter, endMBB);
19661 // Transfer the remainder of MBB and its successor edges to endMBB.
19662 endMBB->splice(endMBB->begin(), thisMBB,
19663 std::next(MachineBasicBlock::iterator(MI)), thisMBB->end());
19664 endMBB->transferSuccessorsAndUpdatePHIs(thisMBB);
19666 // Make offsetMBB and overflowMBB successors of thisMBB
19667 thisMBB->addSuccessor(offsetMBB);
19668 thisMBB->addSuccessor(overflowMBB);
19670 // endMBB is a successor of both offsetMBB and overflowMBB
19671 offsetMBB->addSuccessor(endMBB);
19672 overflowMBB->addSuccessor(endMBB);
19674 // Load the offset value into a register
19675 OffsetReg = MRI.createVirtualRegister(OffsetRegClass);
19676 BuildMI(thisMBB, DL, TII->get(X86::MOV32rm), OffsetReg)
19680 .addDisp(Disp, UseFPOffset ? 4 : 0)
19681 .addOperand(Segment)
19682 .setMemRefs(MMOBegin, MMOEnd);
19684 // Check if there is enough room left to pull this argument.
19685 BuildMI(thisMBB, DL, TII->get(X86::CMP32ri))
19687 .addImm(MaxOffset + 8 - ArgSizeA8);
19689 // Branch to "overflowMBB" if offset >= max
19690 // Fall through to "offsetMBB" otherwise
19691 BuildMI(thisMBB, DL, TII->get(X86::GetCondBranchFromCond(X86::COND_AE)))
19692 .addMBB(overflowMBB);
19695 // In offsetMBB, emit code to use the reg_save_area.
19697 assert(OffsetReg != 0);
19699 // Read the reg_save_area address.
19700 unsigned RegSaveReg = MRI.createVirtualRegister(AddrRegClass);
19701 BuildMI(offsetMBB, DL, TII->get(X86::MOV64rm), RegSaveReg)
19706 .addOperand(Segment)
19707 .setMemRefs(MMOBegin, MMOEnd);
19709 // Zero-extend the offset
19710 unsigned OffsetReg64 = MRI.createVirtualRegister(AddrRegClass);
19711 BuildMI(offsetMBB, DL, TII->get(X86::SUBREG_TO_REG), OffsetReg64)
19714 .addImm(X86::sub_32bit);
19716 // Add the offset to the reg_save_area to get the final address.
19717 BuildMI(offsetMBB, DL, TII->get(X86::ADD64rr), OffsetDestReg)
19718 .addReg(OffsetReg64)
19719 .addReg(RegSaveReg);
19721 // Compute the offset for the next argument
19722 unsigned NextOffsetReg = MRI.createVirtualRegister(OffsetRegClass);
19723 BuildMI(offsetMBB, DL, TII->get(X86::ADD32ri), NextOffsetReg)
19725 .addImm(UseFPOffset ? 16 : 8);
19727 // Store it back into the va_list.
19728 BuildMI(offsetMBB, DL, TII->get(X86::MOV32mr))
19732 .addDisp(Disp, UseFPOffset ? 4 : 0)
19733 .addOperand(Segment)
19734 .addReg(NextOffsetReg)
19735 .setMemRefs(MMOBegin, MMOEnd);
19738 BuildMI(offsetMBB, DL, TII->get(X86::JMP_1))
19743 // Emit code to use overflow area
19746 // Load the overflow_area address into a register.
19747 unsigned OverflowAddrReg = MRI.createVirtualRegister(AddrRegClass);
19748 BuildMI(overflowMBB, DL, TII->get(X86::MOV64rm), OverflowAddrReg)
19753 .addOperand(Segment)
19754 .setMemRefs(MMOBegin, MMOEnd);
19756 // If we need to align it, do so. Otherwise, just copy the address
19757 // to OverflowDestReg.
19759 // Align the overflow address
19760 assert((Align & (Align-1)) == 0 && "Alignment must be a power of 2");
19761 unsigned TmpReg = MRI.createVirtualRegister(AddrRegClass);
19763 // aligned_addr = (addr + (align-1)) & ~(align-1)
19764 BuildMI(overflowMBB, DL, TII->get(X86::ADD64ri32), TmpReg)
19765 .addReg(OverflowAddrReg)
19768 BuildMI(overflowMBB, DL, TII->get(X86::AND64ri32), OverflowDestReg)
19770 .addImm(~(uint64_t)(Align-1));
19772 BuildMI(overflowMBB, DL, TII->get(TargetOpcode::COPY), OverflowDestReg)
19773 .addReg(OverflowAddrReg);
19776 // Compute the next overflow address after this argument.
19777 // (the overflow address should be kept 8-byte aligned)
19778 unsigned NextAddrReg = MRI.createVirtualRegister(AddrRegClass);
19779 BuildMI(overflowMBB, DL, TII->get(X86::ADD64ri32), NextAddrReg)
19780 .addReg(OverflowDestReg)
19781 .addImm(ArgSizeA8);
19783 // Store the new overflow address.
19784 BuildMI(overflowMBB, DL, TII->get(X86::MOV64mr))
19789 .addOperand(Segment)
19790 .addReg(NextAddrReg)
19791 .setMemRefs(MMOBegin, MMOEnd);
19793 // If we branched, emit the PHI to the front of endMBB.
19795 BuildMI(*endMBB, endMBB->begin(), DL,
19796 TII->get(X86::PHI), DestReg)
19797 .addReg(OffsetDestReg).addMBB(offsetMBB)
19798 .addReg(OverflowDestReg).addMBB(overflowMBB);
19801 // Erase the pseudo instruction
19802 MI->eraseFromParent();
19807 MachineBasicBlock *
19808 X86TargetLowering::EmitVAStartSaveXMMRegsWithCustomInserter(
19810 MachineBasicBlock *MBB) const {
19811 // Emit code to save XMM registers to the stack. The ABI says that the
19812 // number of registers to save is given in %al, so it's theoretically
19813 // possible to do an indirect jump trick to avoid saving all of them,
19814 // however this code takes a simpler approach and just executes all
19815 // of the stores if %al is non-zero. It's less code, and it's probably
19816 // easier on the hardware branch predictor, and stores aren't all that
19817 // expensive anyway.
19819 // Create the new basic blocks. One block contains all the XMM stores,
19820 // and one block is the final destination regardless of whether any
19821 // stores were performed.
19822 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
19823 MachineFunction *F = MBB->getParent();
19824 MachineFunction::iterator MBBIter = MBB;
19826 MachineBasicBlock *XMMSaveMBB = F->CreateMachineBasicBlock(LLVM_BB);
19827 MachineBasicBlock *EndMBB = F->CreateMachineBasicBlock(LLVM_BB);
19828 F->insert(MBBIter, XMMSaveMBB);
19829 F->insert(MBBIter, EndMBB);
19831 // Transfer the remainder of MBB and its successor edges to EndMBB.
19832 EndMBB->splice(EndMBB->begin(), MBB,
19833 std::next(MachineBasicBlock::iterator(MI)), MBB->end());
19834 EndMBB->transferSuccessorsAndUpdatePHIs(MBB);
19836 // The original block will now fall through to the XMM save block.
19837 MBB->addSuccessor(XMMSaveMBB);
19838 // The XMMSaveMBB will fall through to the end block.
19839 XMMSaveMBB->addSuccessor(EndMBB);
19841 // Now add the instructions.
19842 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
19843 DebugLoc DL = MI->getDebugLoc();
19845 unsigned CountReg = MI->getOperand(0).getReg();
19846 int64_t RegSaveFrameIndex = MI->getOperand(1).getImm();
19847 int64_t VarArgsFPOffset = MI->getOperand(2).getImm();
19849 if (!Subtarget->isTargetWin64()) {
19850 // If %al is 0, branch around the XMM save block.
19851 BuildMI(MBB, DL, TII->get(X86::TEST8rr)).addReg(CountReg).addReg(CountReg);
19852 BuildMI(MBB, DL, TII->get(X86::JE_1)).addMBB(EndMBB);
19853 MBB->addSuccessor(EndMBB);
19856 // Make sure the last operand is EFLAGS, which gets clobbered by the branch
19857 // that was just emitted, but clearly shouldn't be "saved".
19858 assert((MI->getNumOperands() <= 3 ||
19859 !MI->getOperand(MI->getNumOperands() - 1).isReg() ||
19860 MI->getOperand(MI->getNumOperands() - 1).getReg() == X86::EFLAGS)
19861 && "Expected last argument to be EFLAGS");
19862 unsigned MOVOpc = Subtarget->hasFp256() ? X86::VMOVAPSmr : X86::MOVAPSmr;
19863 // In the XMM save block, save all the XMM argument registers.
19864 for (int i = 3, e = MI->getNumOperands() - 1; i != e; ++i) {
19865 int64_t Offset = (i - 3) * 16 + VarArgsFPOffset;
19866 MachineMemOperand *MMO =
19867 F->getMachineMemOperand(
19868 MachinePointerInfo::getFixedStack(RegSaveFrameIndex, Offset),
19869 MachineMemOperand::MOStore,
19870 /*Size=*/16, /*Align=*/16);
19871 BuildMI(XMMSaveMBB, DL, TII->get(MOVOpc))
19872 .addFrameIndex(RegSaveFrameIndex)
19873 .addImm(/*Scale=*/1)
19874 .addReg(/*IndexReg=*/0)
19875 .addImm(/*Disp=*/Offset)
19876 .addReg(/*Segment=*/0)
19877 .addReg(MI->getOperand(i).getReg())
19878 .addMemOperand(MMO);
19881 MI->eraseFromParent(); // The pseudo instruction is gone now.
19886 // The EFLAGS operand of SelectItr might be missing a kill marker
19887 // because there were multiple uses of EFLAGS, and ISel didn't know
19888 // which to mark. Figure out whether SelectItr should have had a
19889 // kill marker, and set it if it should. Returns the correct kill
19891 static bool checkAndUpdateEFLAGSKill(MachineBasicBlock::iterator SelectItr,
19892 MachineBasicBlock* BB,
19893 const TargetRegisterInfo* TRI) {
19894 // Scan forward through BB for a use/def of EFLAGS.
19895 MachineBasicBlock::iterator miI(std::next(SelectItr));
19896 for (MachineBasicBlock::iterator miE = BB->end(); miI != miE; ++miI) {
19897 const MachineInstr& mi = *miI;
19898 if (mi.readsRegister(X86::EFLAGS))
19900 if (mi.definesRegister(X86::EFLAGS))
19901 break; // Should have kill-flag - update below.
19904 // If we hit the end of the block, check whether EFLAGS is live into a
19906 if (miI == BB->end()) {
19907 for (MachineBasicBlock::succ_iterator sItr = BB->succ_begin(),
19908 sEnd = BB->succ_end();
19909 sItr != sEnd; ++sItr) {
19910 MachineBasicBlock* succ = *sItr;
19911 if (succ->isLiveIn(X86::EFLAGS))
19916 // We found a def, or hit the end of the basic block and EFLAGS wasn't live
19917 // out. SelectMI should have a kill flag on EFLAGS.
19918 SelectItr->addRegisterKilled(X86::EFLAGS, TRI);
19922 MachineBasicBlock *
19923 X86TargetLowering::EmitLoweredSelect(MachineInstr *MI,
19924 MachineBasicBlock *BB) const {
19925 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
19926 DebugLoc DL = MI->getDebugLoc();
19928 // To "insert" a SELECT_CC instruction, we actually have to insert the
19929 // diamond control-flow pattern. The incoming instruction knows the
19930 // destination vreg to set, the condition code register to branch on, the
19931 // true/false values to select between, and a branch opcode to use.
19932 const BasicBlock *LLVM_BB = BB->getBasicBlock();
19933 MachineFunction::iterator It = BB;
19939 // cmpTY ccX, r1, r2
19941 // fallthrough --> copy0MBB
19942 MachineBasicBlock *thisMBB = BB;
19943 MachineFunction *F = BB->getParent();
19945 // We also lower double CMOVs:
19946 // (CMOV (CMOV F, T, cc1), T, cc2)
19947 // to two successives branches. For that, we look for another CMOV as the
19948 // following instruction.
19950 // Without this, we would add a PHI between the two jumps, which ends up
19951 // creating a few copies all around. For instance, for
19953 // (sitofp (zext (fcmp une)))
19955 // we would generate:
19957 // ucomiss %xmm1, %xmm0
19958 // movss <1.0f>, %xmm0
19959 // movaps %xmm0, %xmm1
19961 // xorps %xmm1, %xmm1
19964 // movaps %xmm1, %xmm0
19968 // because this custom-inserter would have generated:
19980 // A: X = ...; Y = ...
19982 // C: Z = PHI [X, A], [Y, B]
19984 // E: PHI [X, C], [Z, D]
19986 // If we lower both CMOVs in a single step, we can instead generate:
19998 // A: X = ...; Y = ...
20000 // E: PHI [X, A], [X, C], [Y, D]
20002 // Which, in our sitofp/fcmp example, gives us something like:
20004 // ucomiss %xmm1, %xmm0
20005 // movss <1.0f>, %xmm0
20008 // xorps %xmm0, %xmm0
20012 MachineInstr *NextCMOV = nullptr;
20013 MachineBasicBlock::iterator NextMIIt =
20014 std::next(MachineBasicBlock::iterator(MI));
20015 if (NextMIIt != BB->end() && NextMIIt->getOpcode() == MI->getOpcode() &&
20016 NextMIIt->getOperand(2).getReg() == MI->getOperand(2).getReg() &&
20017 NextMIIt->getOperand(1).getReg() == MI->getOperand(0).getReg())
20018 NextCMOV = &*NextMIIt;
20020 MachineBasicBlock *jcc1MBB = nullptr;
20022 // If we have a double CMOV, we lower it to two successive branches to
20023 // the same block. EFLAGS is used by both, so mark it as live in the second.
20025 jcc1MBB = F->CreateMachineBasicBlock(LLVM_BB);
20026 F->insert(It, jcc1MBB);
20027 jcc1MBB->addLiveIn(X86::EFLAGS);
20030 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
20031 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
20032 F->insert(It, copy0MBB);
20033 F->insert(It, sinkMBB);
20035 // If the EFLAGS register isn't dead in the terminator, then claim that it's
20036 // live into the sink and copy blocks.
20037 const TargetRegisterInfo *TRI = Subtarget->getRegisterInfo();
20039 MachineInstr *LastEFLAGSUser = NextCMOV ? NextCMOV : MI;
20040 if (!LastEFLAGSUser->killsRegister(X86::EFLAGS) &&
20041 !checkAndUpdateEFLAGSKill(LastEFLAGSUser, BB, TRI)) {
20042 copy0MBB->addLiveIn(X86::EFLAGS);
20043 sinkMBB->addLiveIn(X86::EFLAGS);
20046 // Transfer the remainder of BB and its successor edges to sinkMBB.
20047 sinkMBB->splice(sinkMBB->begin(), BB,
20048 std::next(MachineBasicBlock::iterator(MI)), BB->end());
20049 sinkMBB->transferSuccessorsAndUpdatePHIs(BB);
20051 // Add the true and fallthrough blocks as its successors.
20053 // The fallthrough block may be jcc1MBB, if we have a double CMOV.
20054 BB->addSuccessor(jcc1MBB);
20056 // In that case, jcc1MBB will itself fallthrough the copy0MBB, and
20057 // jump to the sinkMBB.
20058 jcc1MBB->addSuccessor(copy0MBB);
20059 jcc1MBB->addSuccessor(sinkMBB);
20061 BB->addSuccessor(copy0MBB);
20064 // The true block target of the first (or only) branch is always sinkMBB.
20065 BB->addSuccessor(sinkMBB);
20067 // Create the conditional branch instruction.
20069 X86::GetCondBranchFromCond((X86::CondCode)MI->getOperand(3).getImm());
20070 BuildMI(BB, DL, TII->get(Opc)).addMBB(sinkMBB);
20073 unsigned Opc2 = X86::GetCondBranchFromCond(
20074 (X86::CondCode)NextCMOV->getOperand(3).getImm());
20075 BuildMI(jcc1MBB, DL, TII->get(Opc2)).addMBB(sinkMBB);
20079 // %FalseValue = ...
20080 // # fallthrough to sinkMBB
20081 copy0MBB->addSuccessor(sinkMBB);
20084 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
20086 MachineInstrBuilder MIB =
20087 BuildMI(*sinkMBB, sinkMBB->begin(), DL, TII->get(X86::PHI),
20088 MI->getOperand(0).getReg())
20089 .addReg(MI->getOperand(1).getReg()).addMBB(copy0MBB)
20090 .addReg(MI->getOperand(2).getReg()).addMBB(thisMBB);
20092 // If we have a double CMOV, the second Jcc provides the same incoming
20093 // value as the first Jcc (the True operand of the SELECT_CC/CMOV nodes).
20095 MIB.addReg(MI->getOperand(2).getReg()).addMBB(jcc1MBB);
20096 // Copy the PHI result to the register defined by the second CMOV.
20097 BuildMI(*sinkMBB, std::next(MachineBasicBlock::iterator(MIB.getInstr())),
20098 DL, TII->get(TargetOpcode::COPY), NextCMOV->getOperand(0).getReg())
20099 .addReg(MI->getOperand(0).getReg());
20100 NextCMOV->eraseFromParent();
20103 MI->eraseFromParent(); // The pseudo instruction is gone now.
20107 MachineBasicBlock *
20108 X86TargetLowering::EmitLoweredSegAlloca(MachineInstr *MI,
20109 MachineBasicBlock *BB) const {
20110 MachineFunction *MF = BB->getParent();
20111 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
20112 DebugLoc DL = MI->getDebugLoc();
20113 const BasicBlock *LLVM_BB = BB->getBasicBlock();
20115 assert(MF->shouldSplitStack());
20117 const bool Is64Bit = Subtarget->is64Bit();
20118 const bool IsLP64 = Subtarget->isTarget64BitLP64();
20120 const unsigned TlsReg = Is64Bit ? X86::FS : X86::GS;
20121 const unsigned TlsOffset = IsLP64 ? 0x70 : Is64Bit ? 0x40 : 0x30;
20124 // ... [Till the alloca]
20125 // If stacklet is not large enough, jump to mallocMBB
20128 // Allocate by subtracting from RSP
20129 // Jump to continueMBB
20132 // Allocate by call to runtime
20136 // [rest of original BB]
20139 MachineBasicBlock *mallocMBB = MF->CreateMachineBasicBlock(LLVM_BB);
20140 MachineBasicBlock *bumpMBB = MF->CreateMachineBasicBlock(LLVM_BB);
20141 MachineBasicBlock *continueMBB = MF->CreateMachineBasicBlock(LLVM_BB);
20143 MachineRegisterInfo &MRI = MF->getRegInfo();
20144 const TargetRegisterClass *AddrRegClass =
20145 getRegClassFor(getPointerTy(MF->getDataLayout()));
20147 unsigned mallocPtrVReg = MRI.createVirtualRegister(AddrRegClass),
20148 bumpSPPtrVReg = MRI.createVirtualRegister(AddrRegClass),
20149 tmpSPVReg = MRI.createVirtualRegister(AddrRegClass),
20150 SPLimitVReg = MRI.createVirtualRegister(AddrRegClass),
20151 sizeVReg = MI->getOperand(1).getReg(),
20152 physSPReg = IsLP64 || Subtarget->isTargetNaCl64() ? X86::RSP : X86::ESP;
20154 MachineFunction::iterator MBBIter = BB;
20157 MF->insert(MBBIter, bumpMBB);
20158 MF->insert(MBBIter, mallocMBB);
20159 MF->insert(MBBIter, continueMBB);
20161 continueMBB->splice(continueMBB->begin(), BB,
20162 std::next(MachineBasicBlock::iterator(MI)), BB->end());
20163 continueMBB->transferSuccessorsAndUpdatePHIs(BB);
20165 // Add code to the main basic block to check if the stack limit has been hit,
20166 // and if so, jump to mallocMBB otherwise to bumpMBB.
20167 BuildMI(BB, DL, TII->get(TargetOpcode::COPY), tmpSPVReg).addReg(physSPReg);
20168 BuildMI(BB, DL, TII->get(IsLP64 ? X86::SUB64rr:X86::SUB32rr), SPLimitVReg)
20169 .addReg(tmpSPVReg).addReg(sizeVReg);
20170 BuildMI(BB, DL, TII->get(IsLP64 ? X86::CMP64mr:X86::CMP32mr))
20171 .addReg(0).addImm(1).addReg(0).addImm(TlsOffset).addReg(TlsReg)
20172 .addReg(SPLimitVReg);
20173 BuildMI(BB, DL, TII->get(X86::JG_1)).addMBB(mallocMBB);
20175 // bumpMBB simply decreases the stack pointer, since we know the current
20176 // stacklet has enough space.
20177 BuildMI(bumpMBB, DL, TII->get(TargetOpcode::COPY), physSPReg)
20178 .addReg(SPLimitVReg);
20179 BuildMI(bumpMBB, DL, TII->get(TargetOpcode::COPY), bumpSPPtrVReg)
20180 .addReg(SPLimitVReg);
20181 BuildMI(bumpMBB, DL, TII->get(X86::JMP_1)).addMBB(continueMBB);
20183 // Calls into a routine in libgcc to allocate more space from the heap.
20184 const uint32_t *RegMask =
20185 Subtarget->getRegisterInfo()->getCallPreservedMask(*MF, CallingConv::C);
20187 BuildMI(mallocMBB, DL, TII->get(X86::MOV64rr), X86::RDI)
20189 BuildMI(mallocMBB, DL, TII->get(X86::CALL64pcrel32))
20190 .addExternalSymbol("__morestack_allocate_stack_space")
20191 .addRegMask(RegMask)
20192 .addReg(X86::RDI, RegState::Implicit)
20193 .addReg(X86::RAX, RegState::ImplicitDefine);
20194 } else if (Is64Bit) {
20195 BuildMI(mallocMBB, DL, TII->get(X86::MOV32rr), X86::EDI)
20197 BuildMI(mallocMBB, DL, TII->get(X86::CALL64pcrel32))
20198 .addExternalSymbol("__morestack_allocate_stack_space")
20199 .addRegMask(RegMask)
20200 .addReg(X86::EDI, RegState::Implicit)
20201 .addReg(X86::EAX, RegState::ImplicitDefine);
20203 BuildMI(mallocMBB, DL, TII->get(X86::SUB32ri), physSPReg).addReg(physSPReg)
20205 BuildMI(mallocMBB, DL, TII->get(X86::PUSH32r)).addReg(sizeVReg);
20206 BuildMI(mallocMBB, DL, TII->get(X86::CALLpcrel32))
20207 .addExternalSymbol("__morestack_allocate_stack_space")
20208 .addRegMask(RegMask)
20209 .addReg(X86::EAX, RegState::ImplicitDefine);
20213 BuildMI(mallocMBB, DL, TII->get(X86::ADD32ri), physSPReg).addReg(physSPReg)
20216 BuildMI(mallocMBB, DL, TII->get(TargetOpcode::COPY), mallocPtrVReg)
20217 .addReg(IsLP64 ? X86::RAX : X86::EAX);
20218 BuildMI(mallocMBB, DL, TII->get(X86::JMP_1)).addMBB(continueMBB);
20220 // Set up the CFG correctly.
20221 BB->addSuccessor(bumpMBB);
20222 BB->addSuccessor(mallocMBB);
20223 mallocMBB->addSuccessor(continueMBB);
20224 bumpMBB->addSuccessor(continueMBB);
20226 // Take care of the PHI nodes.
20227 BuildMI(*continueMBB, continueMBB->begin(), DL, TII->get(X86::PHI),
20228 MI->getOperand(0).getReg())
20229 .addReg(mallocPtrVReg).addMBB(mallocMBB)
20230 .addReg(bumpSPPtrVReg).addMBB(bumpMBB);
20232 // Delete the original pseudo instruction.
20233 MI->eraseFromParent();
20236 return continueMBB;
20239 MachineBasicBlock *
20240 X86TargetLowering::EmitLoweredWinAlloca(MachineInstr *MI,
20241 MachineBasicBlock *BB) const {
20242 DebugLoc DL = MI->getDebugLoc();
20244 assert(!Subtarget->isTargetMachO());
20246 Subtarget->getFrameLowering()->emitStackProbeCall(*BB->getParent(), *BB, MI,
20249 MI->eraseFromParent(); // The pseudo instruction is gone now.
20253 MachineBasicBlock *
20254 X86TargetLowering::EmitLoweredTLSCall(MachineInstr *MI,
20255 MachineBasicBlock *BB) const {
20256 // This is pretty easy. We're taking the value that we received from
20257 // our load from the relocation, sticking it in either RDI (x86-64)
20258 // or EAX and doing an indirect call. The return value will then
20259 // be in the normal return register.
20260 MachineFunction *F = BB->getParent();
20261 const X86InstrInfo *TII = Subtarget->getInstrInfo();
20262 DebugLoc DL = MI->getDebugLoc();
20264 assert(Subtarget->isTargetDarwin() && "Darwin only instr emitted?");
20265 assert(MI->getOperand(3).isGlobal() && "This should be a global");
20267 // Get a register mask for the lowered call.
20268 // FIXME: The 32-bit calls have non-standard calling conventions. Use a
20269 // proper register mask.
20270 const uint32_t *RegMask =
20271 Subtarget->getRegisterInfo()->getCallPreservedMask(*F, CallingConv::C);
20272 if (Subtarget->is64Bit()) {
20273 MachineInstrBuilder MIB = BuildMI(*BB, MI, DL,
20274 TII->get(X86::MOV64rm), X86::RDI)
20276 .addImm(0).addReg(0)
20277 .addGlobalAddress(MI->getOperand(3).getGlobal(), 0,
20278 MI->getOperand(3).getTargetFlags())
20280 MIB = BuildMI(*BB, MI, DL, TII->get(X86::CALL64m));
20281 addDirectMem(MIB, X86::RDI);
20282 MIB.addReg(X86::RAX, RegState::ImplicitDefine).addRegMask(RegMask);
20283 } else if (F->getTarget().getRelocationModel() != Reloc::PIC_) {
20284 MachineInstrBuilder MIB = BuildMI(*BB, MI, DL,
20285 TII->get(X86::MOV32rm), X86::EAX)
20287 .addImm(0).addReg(0)
20288 .addGlobalAddress(MI->getOperand(3).getGlobal(), 0,
20289 MI->getOperand(3).getTargetFlags())
20291 MIB = BuildMI(*BB, MI, DL, TII->get(X86::CALL32m));
20292 addDirectMem(MIB, X86::EAX);
20293 MIB.addReg(X86::EAX, RegState::ImplicitDefine).addRegMask(RegMask);
20295 MachineInstrBuilder MIB = BuildMI(*BB, MI, DL,
20296 TII->get(X86::MOV32rm), X86::EAX)
20297 .addReg(TII->getGlobalBaseReg(F))
20298 .addImm(0).addReg(0)
20299 .addGlobalAddress(MI->getOperand(3).getGlobal(), 0,
20300 MI->getOperand(3).getTargetFlags())
20302 MIB = BuildMI(*BB, MI, DL, TII->get(X86::CALL32m));
20303 addDirectMem(MIB, X86::EAX);
20304 MIB.addReg(X86::EAX, RegState::ImplicitDefine).addRegMask(RegMask);
20307 MI->eraseFromParent(); // The pseudo instruction is gone now.
20311 MachineBasicBlock *
20312 X86TargetLowering::emitEHSjLjSetJmp(MachineInstr *MI,
20313 MachineBasicBlock *MBB) const {
20314 DebugLoc DL = MI->getDebugLoc();
20315 MachineFunction *MF = MBB->getParent();
20316 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
20317 MachineRegisterInfo &MRI = MF->getRegInfo();
20319 const BasicBlock *BB = MBB->getBasicBlock();
20320 MachineFunction::iterator I = MBB;
20323 // Memory Reference
20324 MachineInstr::mmo_iterator MMOBegin = MI->memoperands_begin();
20325 MachineInstr::mmo_iterator MMOEnd = MI->memoperands_end();
20328 unsigned MemOpndSlot = 0;
20330 unsigned CurOp = 0;
20332 DstReg = MI->getOperand(CurOp++).getReg();
20333 const TargetRegisterClass *RC = MRI.getRegClass(DstReg);
20334 assert(RC->hasType(MVT::i32) && "Invalid destination!");
20335 unsigned mainDstReg = MRI.createVirtualRegister(RC);
20336 unsigned restoreDstReg = MRI.createVirtualRegister(RC);
20338 MemOpndSlot = CurOp;
20340 MVT PVT = getPointerTy(MF->getDataLayout());
20341 assert((PVT == MVT::i64 || PVT == MVT::i32) &&
20342 "Invalid Pointer Size!");
20344 // For v = setjmp(buf), we generate
20347 // buf[LabelOffset] = restoreMBB
20348 // SjLjSetup restoreMBB
20354 // v = phi(main, restore)
20357 // if base pointer being used, load it from frame
20360 MachineBasicBlock *thisMBB = MBB;
20361 MachineBasicBlock *mainMBB = MF->CreateMachineBasicBlock(BB);
20362 MachineBasicBlock *sinkMBB = MF->CreateMachineBasicBlock(BB);
20363 MachineBasicBlock *restoreMBB = MF->CreateMachineBasicBlock(BB);
20364 MF->insert(I, mainMBB);
20365 MF->insert(I, sinkMBB);
20366 MF->push_back(restoreMBB);
20368 MachineInstrBuilder MIB;
20370 // Transfer the remainder of BB and its successor edges to sinkMBB.
20371 sinkMBB->splice(sinkMBB->begin(), MBB,
20372 std::next(MachineBasicBlock::iterator(MI)), MBB->end());
20373 sinkMBB->transferSuccessorsAndUpdatePHIs(MBB);
20376 unsigned PtrStoreOpc = 0;
20377 unsigned LabelReg = 0;
20378 const int64_t LabelOffset = 1 * PVT.getStoreSize();
20379 Reloc::Model RM = MF->getTarget().getRelocationModel();
20380 bool UseImmLabel = (MF->getTarget().getCodeModel() == CodeModel::Small) &&
20381 (RM == Reloc::Static || RM == Reloc::DynamicNoPIC);
20383 // Prepare IP either in reg or imm.
20384 if (!UseImmLabel) {
20385 PtrStoreOpc = (PVT == MVT::i64) ? X86::MOV64mr : X86::MOV32mr;
20386 const TargetRegisterClass *PtrRC = getRegClassFor(PVT);
20387 LabelReg = MRI.createVirtualRegister(PtrRC);
20388 if (Subtarget->is64Bit()) {
20389 MIB = BuildMI(*thisMBB, MI, DL, TII->get(X86::LEA64r), LabelReg)
20393 .addMBB(restoreMBB)
20396 const X86InstrInfo *XII = static_cast<const X86InstrInfo*>(TII);
20397 MIB = BuildMI(*thisMBB, MI, DL, TII->get(X86::LEA32r), LabelReg)
20398 .addReg(XII->getGlobalBaseReg(MF))
20401 .addMBB(restoreMBB, Subtarget->ClassifyBlockAddressReference())
20405 PtrStoreOpc = (PVT == MVT::i64) ? X86::MOV64mi32 : X86::MOV32mi;
20407 MIB = BuildMI(*thisMBB, MI, DL, TII->get(PtrStoreOpc));
20408 for (unsigned i = 0; i < X86::AddrNumOperands; ++i) {
20409 if (i == X86::AddrDisp)
20410 MIB.addDisp(MI->getOperand(MemOpndSlot + i), LabelOffset);
20412 MIB.addOperand(MI->getOperand(MemOpndSlot + i));
20415 MIB.addReg(LabelReg);
20417 MIB.addMBB(restoreMBB);
20418 MIB.setMemRefs(MMOBegin, MMOEnd);
20420 MIB = BuildMI(*thisMBB, MI, DL, TII->get(X86::EH_SjLj_Setup))
20421 .addMBB(restoreMBB);
20423 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
20424 MIB.addRegMask(RegInfo->getNoPreservedMask());
20425 thisMBB->addSuccessor(mainMBB);
20426 thisMBB->addSuccessor(restoreMBB);
20430 BuildMI(mainMBB, DL, TII->get(X86::MOV32r0), mainDstReg);
20431 mainMBB->addSuccessor(sinkMBB);
20434 BuildMI(*sinkMBB, sinkMBB->begin(), DL,
20435 TII->get(X86::PHI), DstReg)
20436 .addReg(mainDstReg).addMBB(mainMBB)
20437 .addReg(restoreDstReg).addMBB(restoreMBB);
20440 if (RegInfo->hasBasePointer(*MF)) {
20441 const bool Uses64BitFramePtr =
20442 Subtarget->isTarget64BitLP64() || Subtarget->isTargetNaCl64();
20443 X86MachineFunctionInfo *X86FI = MF->getInfo<X86MachineFunctionInfo>();
20444 X86FI->setRestoreBasePointer(MF);
20445 unsigned FramePtr = RegInfo->getFrameRegister(*MF);
20446 unsigned BasePtr = RegInfo->getBaseRegister();
20447 unsigned Opm = Uses64BitFramePtr ? X86::MOV64rm : X86::MOV32rm;
20448 addRegOffset(BuildMI(restoreMBB, DL, TII->get(Opm), BasePtr),
20449 FramePtr, true, X86FI->getRestoreBasePointerOffset())
20450 .setMIFlag(MachineInstr::FrameSetup);
20452 BuildMI(restoreMBB, DL, TII->get(X86::MOV32ri), restoreDstReg).addImm(1);
20453 BuildMI(restoreMBB, DL, TII->get(X86::JMP_1)).addMBB(sinkMBB);
20454 restoreMBB->addSuccessor(sinkMBB);
20456 MI->eraseFromParent();
20460 MachineBasicBlock *
20461 X86TargetLowering::emitEHSjLjLongJmp(MachineInstr *MI,
20462 MachineBasicBlock *MBB) const {
20463 DebugLoc DL = MI->getDebugLoc();
20464 MachineFunction *MF = MBB->getParent();
20465 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
20466 MachineRegisterInfo &MRI = MF->getRegInfo();
20468 // Memory Reference
20469 MachineInstr::mmo_iterator MMOBegin = MI->memoperands_begin();
20470 MachineInstr::mmo_iterator MMOEnd = MI->memoperands_end();
20472 MVT PVT = getPointerTy(MF->getDataLayout());
20473 assert((PVT == MVT::i64 || PVT == MVT::i32) &&
20474 "Invalid Pointer Size!");
20476 const TargetRegisterClass *RC =
20477 (PVT == MVT::i64) ? &X86::GR64RegClass : &X86::GR32RegClass;
20478 unsigned Tmp = MRI.createVirtualRegister(RC);
20479 // Since FP is only updated here but NOT referenced, it's treated as GPR.
20480 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
20481 unsigned FP = (PVT == MVT::i64) ? X86::RBP : X86::EBP;
20482 unsigned SP = RegInfo->getStackRegister();
20484 MachineInstrBuilder MIB;
20486 const int64_t LabelOffset = 1 * PVT.getStoreSize();
20487 const int64_t SPOffset = 2 * PVT.getStoreSize();
20489 unsigned PtrLoadOpc = (PVT == MVT::i64) ? X86::MOV64rm : X86::MOV32rm;
20490 unsigned IJmpOpc = (PVT == MVT::i64) ? X86::JMP64r : X86::JMP32r;
20493 MIB = BuildMI(*MBB, MI, DL, TII->get(PtrLoadOpc), FP);
20494 for (unsigned i = 0; i < X86::AddrNumOperands; ++i)
20495 MIB.addOperand(MI->getOperand(i));
20496 MIB.setMemRefs(MMOBegin, MMOEnd);
20498 MIB = BuildMI(*MBB, MI, DL, TII->get(PtrLoadOpc), Tmp);
20499 for (unsigned i = 0; i < X86::AddrNumOperands; ++i) {
20500 if (i == X86::AddrDisp)
20501 MIB.addDisp(MI->getOperand(i), LabelOffset);
20503 MIB.addOperand(MI->getOperand(i));
20505 MIB.setMemRefs(MMOBegin, MMOEnd);
20507 MIB = BuildMI(*MBB, MI, DL, TII->get(PtrLoadOpc), SP);
20508 for (unsigned i = 0; i < X86::AddrNumOperands; ++i) {
20509 if (i == X86::AddrDisp)
20510 MIB.addDisp(MI->getOperand(i), SPOffset);
20512 MIB.addOperand(MI->getOperand(i));
20514 MIB.setMemRefs(MMOBegin, MMOEnd);
20516 BuildMI(*MBB, MI, DL, TII->get(IJmpOpc)).addReg(Tmp);
20518 MI->eraseFromParent();
20522 // Replace 213-type (isel default) FMA3 instructions with 231-type for
20523 // accumulator loops. Writing back to the accumulator allows the coalescer
20524 // to remove extra copies in the loop.
20525 // FIXME: Do this on AVX512. We don't support 231 variants yet (PR23937).
20526 MachineBasicBlock *
20527 X86TargetLowering::emitFMA3Instr(MachineInstr *MI,
20528 MachineBasicBlock *MBB) const {
20529 MachineOperand &AddendOp = MI->getOperand(3);
20531 // Bail out early if the addend isn't a register - we can't switch these.
20532 if (!AddendOp.isReg())
20535 MachineFunction &MF = *MBB->getParent();
20536 MachineRegisterInfo &MRI = MF.getRegInfo();
20538 // Check whether the addend is defined by a PHI:
20539 assert(MRI.hasOneDef(AddendOp.getReg()) && "Multiple defs in SSA?");
20540 MachineInstr &AddendDef = *MRI.def_instr_begin(AddendOp.getReg());
20541 if (!AddendDef.isPHI())
20544 // Look for the following pattern:
20546 // %addend = phi [%entry, 0], [%loop, %result]
20548 // %result<tied1> = FMA213 %m2<tied0>, %m1, %addend
20552 // %addend = phi [%entry, 0], [%loop, %result]
20554 // %result<tied1> = FMA231 %addend<tied0>, %m1, %m2
20556 for (unsigned i = 1, e = AddendDef.getNumOperands(); i < e; i += 2) {
20557 assert(AddendDef.getOperand(i).isReg());
20558 MachineOperand PHISrcOp = AddendDef.getOperand(i);
20559 MachineInstr &PHISrcInst = *MRI.def_instr_begin(PHISrcOp.getReg());
20560 if (&PHISrcInst == MI) {
20561 // Found a matching instruction.
20562 unsigned NewFMAOpc = 0;
20563 switch (MI->getOpcode()) {
20564 case X86::VFMADDPDr213r: NewFMAOpc = X86::VFMADDPDr231r; break;
20565 case X86::VFMADDPSr213r: NewFMAOpc = X86::VFMADDPSr231r; break;
20566 case X86::VFMADDSDr213r: NewFMAOpc = X86::VFMADDSDr231r; break;
20567 case X86::VFMADDSSr213r: NewFMAOpc = X86::VFMADDSSr231r; break;
20568 case X86::VFMSUBPDr213r: NewFMAOpc = X86::VFMSUBPDr231r; break;
20569 case X86::VFMSUBPSr213r: NewFMAOpc = X86::VFMSUBPSr231r; break;
20570 case X86::VFMSUBSDr213r: NewFMAOpc = X86::VFMSUBSDr231r; break;
20571 case X86::VFMSUBSSr213r: NewFMAOpc = X86::VFMSUBSSr231r; break;
20572 case X86::VFNMADDPDr213r: NewFMAOpc = X86::VFNMADDPDr231r; break;
20573 case X86::VFNMADDPSr213r: NewFMAOpc = X86::VFNMADDPSr231r; break;
20574 case X86::VFNMADDSDr213r: NewFMAOpc = X86::VFNMADDSDr231r; break;
20575 case X86::VFNMADDSSr213r: NewFMAOpc = X86::VFNMADDSSr231r; break;
20576 case X86::VFNMSUBPDr213r: NewFMAOpc = X86::VFNMSUBPDr231r; break;
20577 case X86::VFNMSUBPSr213r: NewFMAOpc = X86::VFNMSUBPSr231r; break;
20578 case X86::VFNMSUBSDr213r: NewFMAOpc = X86::VFNMSUBSDr231r; break;
20579 case X86::VFNMSUBSSr213r: NewFMAOpc = X86::VFNMSUBSSr231r; break;
20580 case X86::VFMADDSUBPDr213r: NewFMAOpc = X86::VFMADDSUBPDr231r; break;
20581 case X86::VFMADDSUBPSr213r: NewFMAOpc = X86::VFMADDSUBPSr231r; break;
20582 case X86::VFMSUBADDPDr213r: NewFMAOpc = X86::VFMSUBADDPDr231r; break;
20583 case X86::VFMSUBADDPSr213r: NewFMAOpc = X86::VFMSUBADDPSr231r; break;
20585 case X86::VFMADDPDr213rY: NewFMAOpc = X86::VFMADDPDr231rY; break;
20586 case X86::VFMADDPSr213rY: NewFMAOpc = X86::VFMADDPSr231rY; break;
20587 case X86::VFMSUBPDr213rY: NewFMAOpc = X86::VFMSUBPDr231rY; break;
20588 case X86::VFMSUBPSr213rY: NewFMAOpc = X86::VFMSUBPSr231rY; break;
20589 case X86::VFNMADDPDr213rY: NewFMAOpc = X86::VFNMADDPDr231rY; break;
20590 case X86::VFNMADDPSr213rY: NewFMAOpc = X86::VFNMADDPSr231rY; break;
20591 case X86::VFNMSUBPDr213rY: NewFMAOpc = X86::VFNMSUBPDr231rY; break;
20592 case X86::VFNMSUBPSr213rY: NewFMAOpc = X86::VFNMSUBPSr231rY; break;
20593 case X86::VFMADDSUBPDr213rY: NewFMAOpc = X86::VFMADDSUBPDr231rY; break;
20594 case X86::VFMADDSUBPSr213rY: NewFMAOpc = X86::VFMADDSUBPSr231rY; break;
20595 case X86::VFMSUBADDPDr213rY: NewFMAOpc = X86::VFMSUBADDPDr231rY; break;
20596 case X86::VFMSUBADDPSr213rY: NewFMAOpc = X86::VFMSUBADDPSr231rY; break;
20597 default: llvm_unreachable("Unrecognized FMA variant.");
20600 const TargetInstrInfo &TII = *Subtarget->getInstrInfo();
20601 MachineInstrBuilder MIB =
20602 BuildMI(MF, MI->getDebugLoc(), TII.get(NewFMAOpc))
20603 .addOperand(MI->getOperand(0))
20604 .addOperand(MI->getOperand(3))
20605 .addOperand(MI->getOperand(2))
20606 .addOperand(MI->getOperand(1));
20607 MBB->insert(MachineBasicBlock::iterator(MI), MIB);
20608 MI->eraseFromParent();
20615 MachineBasicBlock *
20616 X86TargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
20617 MachineBasicBlock *BB) const {
20618 switch (MI->getOpcode()) {
20619 default: llvm_unreachable("Unexpected instr type to insert");
20620 case X86::TAILJMPd64:
20621 case X86::TAILJMPr64:
20622 case X86::TAILJMPm64:
20623 case X86::TAILJMPd64_REX:
20624 case X86::TAILJMPr64_REX:
20625 case X86::TAILJMPm64_REX:
20626 llvm_unreachable("TAILJMP64 would not be touched here.");
20627 case X86::TCRETURNdi64:
20628 case X86::TCRETURNri64:
20629 case X86::TCRETURNmi64:
20631 case X86::WIN_ALLOCA:
20632 return EmitLoweredWinAlloca(MI, BB);
20633 case X86::SEG_ALLOCA_32:
20634 case X86::SEG_ALLOCA_64:
20635 return EmitLoweredSegAlloca(MI, BB);
20636 case X86::TLSCall_32:
20637 case X86::TLSCall_64:
20638 return EmitLoweredTLSCall(MI, BB);
20639 case X86::CMOV_GR8:
20640 case X86::CMOV_FR32:
20641 case X86::CMOV_FR64:
20642 case X86::CMOV_V4F32:
20643 case X86::CMOV_V2F64:
20644 case X86::CMOV_V2I64:
20645 case X86::CMOV_V8F32:
20646 case X86::CMOV_V4F64:
20647 case X86::CMOV_V4I64:
20648 case X86::CMOV_V16F32:
20649 case X86::CMOV_V8F64:
20650 case X86::CMOV_V8I64:
20651 case X86::CMOV_GR16:
20652 case X86::CMOV_GR32:
20653 case X86::CMOV_RFP32:
20654 case X86::CMOV_RFP64:
20655 case X86::CMOV_RFP80:
20656 case X86::CMOV_V8I1:
20657 case X86::CMOV_V16I1:
20658 case X86::CMOV_V32I1:
20659 case X86::CMOV_V64I1:
20660 return EmitLoweredSelect(MI, BB);
20662 case X86::FP32_TO_INT16_IN_MEM:
20663 case X86::FP32_TO_INT32_IN_MEM:
20664 case X86::FP32_TO_INT64_IN_MEM:
20665 case X86::FP64_TO_INT16_IN_MEM:
20666 case X86::FP64_TO_INT32_IN_MEM:
20667 case X86::FP64_TO_INT64_IN_MEM:
20668 case X86::FP80_TO_INT16_IN_MEM:
20669 case X86::FP80_TO_INT32_IN_MEM:
20670 case X86::FP80_TO_INT64_IN_MEM: {
20671 MachineFunction *F = BB->getParent();
20672 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
20673 DebugLoc DL = MI->getDebugLoc();
20675 // Change the floating point control register to use "round towards zero"
20676 // mode when truncating to an integer value.
20677 int CWFrameIdx = F->getFrameInfo()->CreateStackObject(2, 2, false);
20678 addFrameReference(BuildMI(*BB, MI, DL,
20679 TII->get(X86::FNSTCW16m)), CWFrameIdx);
20681 // Load the old value of the high byte of the control word...
20683 F->getRegInfo().createVirtualRegister(&X86::GR16RegClass);
20684 addFrameReference(BuildMI(*BB, MI, DL, TII->get(X86::MOV16rm), OldCW),
20687 // Set the high part to be round to zero...
20688 addFrameReference(BuildMI(*BB, MI, DL, TII->get(X86::MOV16mi)), CWFrameIdx)
20691 // Reload the modified control word now...
20692 addFrameReference(BuildMI(*BB, MI, DL,
20693 TII->get(X86::FLDCW16m)), CWFrameIdx);
20695 // Restore the memory image of control word to original value
20696 addFrameReference(BuildMI(*BB, MI, DL, TII->get(X86::MOV16mr)), CWFrameIdx)
20699 // Get the X86 opcode to use.
20701 switch (MI->getOpcode()) {
20702 default: llvm_unreachable("illegal opcode!");
20703 case X86::FP32_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m32; break;
20704 case X86::FP32_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m32; break;
20705 case X86::FP32_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m32; break;
20706 case X86::FP64_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m64; break;
20707 case X86::FP64_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m64; break;
20708 case X86::FP64_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m64; break;
20709 case X86::FP80_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m80; break;
20710 case X86::FP80_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m80; break;
20711 case X86::FP80_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m80; break;
20715 MachineOperand &Op = MI->getOperand(0);
20717 AM.BaseType = X86AddressMode::RegBase;
20718 AM.Base.Reg = Op.getReg();
20720 AM.BaseType = X86AddressMode::FrameIndexBase;
20721 AM.Base.FrameIndex = Op.getIndex();
20723 Op = MI->getOperand(1);
20725 AM.Scale = Op.getImm();
20726 Op = MI->getOperand(2);
20728 AM.IndexReg = Op.getImm();
20729 Op = MI->getOperand(3);
20730 if (Op.isGlobal()) {
20731 AM.GV = Op.getGlobal();
20733 AM.Disp = Op.getImm();
20735 addFullAddress(BuildMI(*BB, MI, DL, TII->get(Opc)), AM)
20736 .addReg(MI->getOperand(X86::AddrNumOperands).getReg());
20738 // Reload the original control word now.
20739 addFrameReference(BuildMI(*BB, MI, DL,
20740 TII->get(X86::FLDCW16m)), CWFrameIdx);
20742 MI->eraseFromParent(); // The pseudo instruction is gone now.
20745 // String/text processing lowering.
20746 case X86::PCMPISTRM128REG:
20747 case X86::VPCMPISTRM128REG:
20748 case X86::PCMPISTRM128MEM:
20749 case X86::VPCMPISTRM128MEM:
20750 case X86::PCMPESTRM128REG:
20751 case X86::VPCMPESTRM128REG:
20752 case X86::PCMPESTRM128MEM:
20753 case X86::VPCMPESTRM128MEM:
20754 assert(Subtarget->hasSSE42() &&
20755 "Target must have SSE4.2 or AVX features enabled");
20756 return EmitPCMPSTRM(MI, BB, Subtarget->getInstrInfo());
20758 // String/text processing lowering.
20759 case X86::PCMPISTRIREG:
20760 case X86::VPCMPISTRIREG:
20761 case X86::PCMPISTRIMEM:
20762 case X86::VPCMPISTRIMEM:
20763 case X86::PCMPESTRIREG:
20764 case X86::VPCMPESTRIREG:
20765 case X86::PCMPESTRIMEM:
20766 case X86::VPCMPESTRIMEM:
20767 assert(Subtarget->hasSSE42() &&
20768 "Target must have SSE4.2 or AVX features enabled");
20769 return EmitPCMPSTRI(MI, BB, Subtarget->getInstrInfo());
20771 // Thread synchronization.
20773 return EmitMonitor(MI, BB, Subtarget);
20777 return EmitXBegin(MI, BB, Subtarget->getInstrInfo());
20779 case X86::VASTART_SAVE_XMM_REGS:
20780 return EmitVAStartSaveXMMRegsWithCustomInserter(MI, BB);
20782 case X86::VAARG_64:
20783 return EmitVAARG64WithCustomInserter(MI, BB);
20785 case X86::EH_SjLj_SetJmp32:
20786 case X86::EH_SjLj_SetJmp64:
20787 return emitEHSjLjSetJmp(MI, BB);
20789 case X86::EH_SjLj_LongJmp32:
20790 case X86::EH_SjLj_LongJmp64:
20791 return emitEHSjLjLongJmp(MI, BB);
20793 case TargetOpcode::STATEPOINT:
20794 // As an implementation detail, STATEPOINT shares the STACKMAP format at
20795 // this point in the process. We diverge later.
20796 return emitPatchPoint(MI, BB);
20798 case TargetOpcode::STACKMAP:
20799 case TargetOpcode::PATCHPOINT:
20800 return emitPatchPoint(MI, BB);
20802 case X86::VFMADDPDr213r:
20803 case X86::VFMADDPSr213r:
20804 case X86::VFMADDSDr213r:
20805 case X86::VFMADDSSr213r:
20806 case X86::VFMSUBPDr213r:
20807 case X86::VFMSUBPSr213r:
20808 case X86::VFMSUBSDr213r:
20809 case X86::VFMSUBSSr213r:
20810 case X86::VFNMADDPDr213r:
20811 case X86::VFNMADDPSr213r:
20812 case X86::VFNMADDSDr213r:
20813 case X86::VFNMADDSSr213r:
20814 case X86::VFNMSUBPDr213r:
20815 case X86::VFNMSUBPSr213r:
20816 case X86::VFNMSUBSDr213r:
20817 case X86::VFNMSUBSSr213r:
20818 case X86::VFMADDSUBPDr213r:
20819 case X86::VFMADDSUBPSr213r:
20820 case X86::VFMSUBADDPDr213r:
20821 case X86::VFMSUBADDPSr213r:
20822 case X86::VFMADDPDr213rY:
20823 case X86::VFMADDPSr213rY:
20824 case X86::VFMSUBPDr213rY:
20825 case X86::VFMSUBPSr213rY:
20826 case X86::VFNMADDPDr213rY:
20827 case X86::VFNMADDPSr213rY:
20828 case X86::VFNMSUBPDr213rY:
20829 case X86::VFNMSUBPSr213rY:
20830 case X86::VFMADDSUBPDr213rY:
20831 case X86::VFMADDSUBPSr213rY:
20832 case X86::VFMSUBADDPDr213rY:
20833 case X86::VFMSUBADDPSr213rY:
20834 return emitFMA3Instr(MI, BB);
20838 //===----------------------------------------------------------------------===//
20839 // X86 Optimization Hooks
20840 //===----------------------------------------------------------------------===//
20842 void X86TargetLowering::computeKnownBitsForTargetNode(const SDValue Op,
20845 const SelectionDAG &DAG,
20846 unsigned Depth) const {
20847 unsigned BitWidth = KnownZero.getBitWidth();
20848 unsigned Opc = Op.getOpcode();
20849 assert((Opc >= ISD::BUILTIN_OP_END ||
20850 Opc == ISD::INTRINSIC_WO_CHAIN ||
20851 Opc == ISD::INTRINSIC_W_CHAIN ||
20852 Opc == ISD::INTRINSIC_VOID) &&
20853 "Should use MaskedValueIsZero if you don't know whether Op"
20854 " is a target node!");
20856 KnownZero = KnownOne = APInt(BitWidth, 0); // Don't know anything.
20870 // These nodes' second result is a boolean.
20871 if (Op.getResNo() == 0)
20874 case X86ISD::SETCC:
20875 KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - 1);
20877 case ISD::INTRINSIC_WO_CHAIN: {
20878 unsigned IntId = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
20879 unsigned NumLoBits = 0;
20882 case Intrinsic::x86_sse_movmsk_ps:
20883 case Intrinsic::x86_avx_movmsk_ps_256:
20884 case Intrinsic::x86_sse2_movmsk_pd:
20885 case Intrinsic::x86_avx_movmsk_pd_256:
20886 case Intrinsic::x86_mmx_pmovmskb:
20887 case Intrinsic::x86_sse2_pmovmskb_128:
20888 case Intrinsic::x86_avx2_pmovmskb: {
20889 // High bits of movmskp{s|d}, pmovmskb are known zero.
20891 default: llvm_unreachable("Impossible intrinsic"); // Can't reach here.
20892 case Intrinsic::x86_sse_movmsk_ps: NumLoBits = 4; break;
20893 case Intrinsic::x86_avx_movmsk_ps_256: NumLoBits = 8; break;
20894 case Intrinsic::x86_sse2_movmsk_pd: NumLoBits = 2; break;
20895 case Intrinsic::x86_avx_movmsk_pd_256: NumLoBits = 4; break;
20896 case Intrinsic::x86_mmx_pmovmskb: NumLoBits = 8; break;
20897 case Intrinsic::x86_sse2_pmovmskb_128: NumLoBits = 16; break;
20898 case Intrinsic::x86_avx2_pmovmskb: NumLoBits = 32; break;
20900 KnownZero = APInt::getHighBitsSet(BitWidth, BitWidth - NumLoBits);
20909 unsigned X86TargetLowering::ComputeNumSignBitsForTargetNode(
20911 const SelectionDAG &,
20912 unsigned Depth) const {
20913 // SETCC_CARRY sets the dest to ~0 for true or 0 for false.
20914 if (Op.getOpcode() == X86ISD::SETCC_CARRY)
20915 return Op.getValueType().getScalarType().getSizeInBits();
20921 /// isGAPlusOffset - Returns true (and the GlobalValue and the offset) if the
20922 /// node is a GlobalAddress + offset.
20923 bool X86TargetLowering::isGAPlusOffset(SDNode *N,
20924 const GlobalValue* &GA,
20925 int64_t &Offset) const {
20926 if (N->getOpcode() == X86ISD::Wrapper) {
20927 if (isa<GlobalAddressSDNode>(N->getOperand(0))) {
20928 GA = cast<GlobalAddressSDNode>(N->getOperand(0))->getGlobal();
20929 Offset = cast<GlobalAddressSDNode>(N->getOperand(0))->getOffset();
20933 return TargetLowering::isGAPlusOffset(N, GA, Offset);
20936 /// isShuffleHigh128VectorInsertLow - Checks whether the shuffle node is the
20937 /// same as extracting the high 128-bit part of 256-bit vector and then
20938 /// inserting the result into the low part of a new 256-bit vector
20939 static bool isShuffleHigh128VectorInsertLow(ShuffleVectorSDNode *SVOp) {
20940 EVT VT = SVOp->getValueType(0);
20941 unsigned NumElems = VT.getVectorNumElements();
20943 // vector_shuffle <4, 5, 6, 7, u, u, u, u> or <2, 3, u, u>
20944 for (unsigned i = 0, j = NumElems/2; i != NumElems/2; ++i, ++j)
20945 if (!isUndefOrEqual(SVOp->getMaskElt(i), j) ||
20946 SVOp->getMaskElt(j) >= 0)
20952 /// isShuffleLow128VectorInsertHigh - Checks whether the shuffle node is the
20953 /// same as extracting the low 128-bit part of 256-bit vector and then
20954 /// inserting the result into the high part of a new 256-bit vector
20955 static bool isShuffleLow128VectorInsertHigh(ShuffleVectorSDNode *SVOp) {
20956 EVT VT = SVOp->getValueType(0);
20957 unsigned NumElems = VT.getVectorNumElements();
20959 // vector_shuffle <u, u, u, u, 0, 1, 2, 3> or <u, u, 0, 1>
20960 for (unsigned i = NumElems/2, j = 0; i != NumElems; ++i, ++j)
20961 if (!isUndefOrEqual(SVOp->getMaskElt(i), j) ||
20962 SVOp->getMaskElt(j) >= 0)
20968 /// PerformShuffleCombine256 - Performs shuffle combines for 256-bit vectors.
20969 static SDValue PerformShuffleCombine256(SDNode *N, SelectionDAG &DAG,
20970 TargetLowering::DAGCombinerInfo &DCI,
20971 const X86Subtarget* Subtarget) {
20973 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
20974 SDValue V1 = SVOp->getOperand(0);
20975 SDValue V2 = SVOp->getOperand(1);
20976 EVT VT = SVOp->getValueType(0);
20977 unsigned NumElems = VT.getVectorNumElements();
20979 if (V1.getOpcode() == ISD::CONCAT_VECTORS &&
20980 V2.getOpcode() == ISD::CONCAT_VECTORS) {
20984 // V UNDEF BUILD_VECTOR UNDEF
20986 // CONCAT_VECTOR CONCAT_VECTOR
20989 // RESULT: V + zero extended
20991 if (V2.getOperand(0).getOpcode() != ISD::BUILD_VECTOR ||
20992 V2.getOperand(1).getOpcode() != ISD::UNDEF ||
20993 V1.getOperand(1).getOpcode() != ISD::UNDEF)
20996 if (!ISD::isBuildVectorAllZeros(V2.getOperand(0).getNode()))
20999 // To match the shuffle mask, the first half of the mask should
21000 // be exactly the first vector, and all the rest a splat with the
21001 // first element of the second one.
21002 for (unsigned i = 0; i != NumElems/2; ++i)
21003 if (!isUndefOrEqual(SVOp->getMaskElt(i), i) ||
21004 !isUndefOrEqual(SVOp->getMaskElt(i+NumElems/2), NumElems))
21007 // If V1 is coming from a vector load then just fold to a VZEXT_LOAD.
21008 if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(V1.getOperand(0))) {
21009 if (Ld->hasNUsesOfValue(1, 0)) {
21010 SDVTList Tys = DAG.getVTList(MVT::v4i64, MVT::Other);
21011 SDValue Ops[] = { Ld->getChain(), Ld->getBasePtr() };
21013 DAG.getMemIntrinsicNode(X86ISD::VZEXT_LOAD, dl, Tys, Ops,
21015 Ld->getPointerInfo(),
21016 Ld->getAlignment(),
21017 false/*isVolatile*/, true/*ReadMem*/,
21018 false/*WriteMem*/);
21020 // Make sure the newly-created LOAD is in the same position as Ld in
21021 // terms of dependency. We create a TokenFactor for Ld and ResNode,
21022 // and update uses of Ld's output chain to use the TokenFactor.
21023 if (Ld->hasAnyUseOfValue(1)) {
21024 SDValue NewChain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
21025 SDValue(Ld, 1), SDValue(ResNode.getNode(), 1));
21026 DAG.ReplaceAllUsesOfValueWith(SDValue(Ld, 1), NewChain);
21027 DAG.UpdateNodeOperands(NewChain.getNode(), SDValue(Ld, 1),
21028 SDValue(ResNode.getNode(), 1));
21031 return DAG.getBitcast(VT, ResNode);
21035 // Emit a zeroed vector and insert the desired subvector on its
21037 SDValue Zeros = getZeroVector(VT, Subtarget, DAG, dl);
21038 SDValue InsV = Insert128BitVector(Zeros, V1.getOperand(0), 0, DAG, dl);
21039 return DCI.CombineTo(N, InsV);
21042 //===--------------------------------------------------------------------===//
21043 // Combine some shuffles into subvector extracts and inserts:
21046 // vector_shuffle <4, 5, 6, 7, u, u, u, u> or <2, 3, u, u>
21047 if (isShuffleHigh128VectorInsertLow(SVOp)) {
21048 SDValue V = Extract128BitVector(V1, NumElems/2, DAG, dl);
21049 SDValue InsV = Insert128BitVector(DAG.getUNDEF(VT), V, 0, DAG, dl);
21050 return DCI.CombineTo(N, InsV);
21053 // vector_shuffle <u, u, u, u, 0, 1, 2, 3> or <u, u, 0, 1>
21054 if (isShuffleLow128VectorInsertHigh(SVOp)) {
21055 SDValue V = Extract128BitVector(V1, 0, DAG, dl);
21056 SDValue InsV = Insert128BitVector(DAG.getUNDEF(VT), V, NumElems/2, DAG, dl);
21057 return DCI.CombineTo(N, InsV);
21063 /// \brief Combine an arbitrary chain of shuffles into a single instruction if
21066 /// This is the leaf of the recursive combinine below. When we have found some
21067 /// chain of single-use x86 shuffle instructions and accumulated the combined
21068 /// shuffle mask represented by them, this will try to pattern match that mask
21069 /// into either a single instruction if there is a special purpose instruction
21070 /// for this operation, or into a PSHUFB instruction which is a fully general
21071 /// instruction but should only be used to replace chains over a certain depth.
21072 static bool combineX86ShuffleChain(SDValue Op, SDValue Root, ArrayRef<int> Mask,
21073 int Depth, bool HasPSHUFB, SelectionDAG &DAG,
21074 TargetLowering::DAGCombinerInfo &DCI,
21075 const X86Subtarget *Subtarget) {
21076 assert(!Mask.empty() && "Cannot combine an empty shuffle mask!");
21078 // Find the operand that enters the chain. Note that multiple uses are OK
21079 // here, we're not going to remove the operand we find.
21080 SDValue Input = Op.getOperand(0);
21081 while (Input.getOpcode() == ISD::BITCAST)
21082 Input = Input.getOperand(0);
21084 MVT VT = Input.getSimpleValueType();
21085 MVT RootVT = Root.getSimpleValueType();
21088 // Just remove no-op shuffle masks.
21089 if (Mask.size() == 1) {
21090 DCI.CombineTo(Root.getNode(), DAG.getBitcast(RootVT, Input),
21095 // Use the float domain if the operand type is a floating point type.
21096 bool FloatDomain = VT.isFloatingPoint();
21098 // For floating point shuffles, we don't have free copies in the shuffle
21099 // instructions or the ability to load as part of the instruction, so
21100 // canonicalize their shuffles to UNPCK or MOV variants.
21102 // Note that even with AVX we prefer the PSHUFD form of shuffle for integer
21103 // vectors because it can have a load folded into it that UNPCK cannot. This
21104 // doesn't preclude something switching to the shorter encoding post-RA.
21106 // FIXME: Should teach these routines about AVX vector widths.
21107 if (FloatDomain && VT.getSizeInBits() == 128) {
21108 if (Mask.equals({0, 0}) || Mask.equals({1, 1})) {
21109 bool Lo = Mask.equals({0, 0});
21112 // Check if we have SSE3 which will let us use MOVDDUP. That instruction
21113 // is no slower than UNPCKLPD but has the option to fold the input operand
21114 // into even an unaligned memory load.
21115 if (Lo && Subtarget->hasSSE3()) {
21116 Shuffle = X86ISD::MOVDDUP;
21117 ShuffleVT = MVT::v2f64;
21119 // We have MOVLHPS and MOVHLPS throughout SSE and they encode smaller
21120 // than the UNPCK variants.
21121 Shuffle = Lo ? X86ISD::MOVLHPS : X86ISD::MOVHLPS;
21122 ShuffleVT = MVT::v4f32;
21124 if (Depth == 1 && Root->getOpcode() == Shuffle)
21125 return false; // Nothing to do!
21126 Op = DAG.getBitcast(ShuffleVT, Input);
21127 DCI.AddToWorklist(Op.getNode());
21128 if (Shuffle == X86ISD::MOVDDUP)
21129 Op = DAG.getNode(Shuffle, DL, ShuffleVT, Op);
21131 Op = DAG.getNode(Shuffle, DL, ShuffleVT, Op, Op);
21132 DCI.AddToWorklist(Op.getNode());
21133 DCI.CombineTo(Root.getNode(), DAG.getBitcast(RootVT, Op),
21137 if (Subtarget->hasSSE3() &&
21138 (Mask.equals({0, 0, 2, 2}) || Mask.equals({1, 1, 3, 3}))) {
21139 bool Lo = Mask.equals({0, 0, 2, 2});
21140 unsigned Shuffle = Lo ? X86ISD::MOVSLDUP : X86ISD::MOVSHDUP;
21141 MVT ShuffleVT = MVT::v4f32;
21142 if (Depth == 1 && Root->getOpcode() == Shuffle)
21143 return false; // Nothing to do!
21144 Op = DAG.getBitcast(ShuffleVT, Input);
21145 DCI.AddToWorklist(Op.getNode());
21146 Op = DAG.getNode(Shuffle, DL, ShuffleVT, Op);
21147 DCI.AddToWorklist(Op.getNode());
21148 DCI.CombineTo(Root.getNode(), DAG.getBitcast(RootVT, Op),
21152 if (Mask.equals({0, 0, 1, 1}) || Mask.equals({2, 2, 3, 3})) {
21153 bool Lo = Mask.equals({0, 0, 1, 1});
21154 unsigned Shuffle = Lo ? X86ISD::UNPCKL : X86ISD::UNPCKH;
21155 MVT ShuffleVT = MVT::v4f32;
21156 if (Depth == 1 && Root->getOpcode() == Shuffle)
21157 return false; // Nothing to do!
21158 Op = DAG.getBitcast(ShuffleVT, Input);
21159 DCI.AddToWorklist(Op.getNode());
21160 Op = DAG.getNode(Shuffle, DL, ShuffleVT, Op, Op);
21161 DCI.AddToWorklist(Op.getNode());
21162 DCI.CombineTo(Root.getNode(), DAG.getBitcast(RootVT, Op),
21168 // We always canonicalize the 8 x i16 and 16 x i8 shuffles into their UNPCK
21169 // variants as none of these have single-instruction variants that are
21170 // superior to the UNPCK formulation.
21171 if (!FloatDomain && VT.getSizeInBits() == 128 &&
21172 (Mask.equals({0, 0, 1, 1, 2, 2, 3, 3}) ||
21173 Mask.equals({4, 4, 5, 5, 6, 6, 7, 7}) ||
21174 Mask.equals({0, 0, 1, 1, 2, 2, 3, 3, 4, 4, 5, 5, 6, 6, 7, 7}) ||
21176 {8, 8, 9, 9, 10, 10, 11, 11, 12, 12, 13, 13, 14, 14, 15, 15}))) {
21177 bool Lo = Mask[0] == 0;
21178 unsigned Shuffle = Lo ? X86ISD::UNPCKL : X86ISD::UNPCKH;
21179 if (Depth == 1 && Root->getOpcode() == Shuffle)
21180 return false; // Nothing to do!
21182 switch (Mask.size()) {
21184 ShuffleVT = MVT::v8i16;
21187 ShuffleVT = MVT::v16i8;
21190 llvm_unreachable("Impossible mask size!");
21192 Op = DAG.getBitcast(ShuffleVT, Input);
21193 DCI.AddToWorklist(Op.getNode());
21194 Op = DAG.getNode(Shuffle, DL, ShuffleVT, Op, Op);
21195 DCI.AddToWorklist(Op.getNode());
21196 DCI.CombineTo(Root.getNode(), DAG.getBitcast(RootVT, Op),
21201 // Don't try to re-form single instruction chains under any circumstances now
21202 // that we've done encoding canonicalization for them.
21206 // If we have 3 or more shuffle instructions or a chain involving PSHUFB, we
21207 // can replace them with a single PSHUFB instruction profitably. Intel's
21208 // manuals suggest only using PSHUFB if doing so replacing 5 instructions, but
21209 // in practice PSHUFB tends to be *very* fast so we're more aggressive.
21210 if ((Depth >= 3 || HasPSHUFB) && Subtarget->hasSSSE3()) {
21211 SmallVector<SDValue, 16> PSHUFBMask;
21212 int NumBytes = VT.getSizeInBits() / 8;
21213 int Ratio = NumBytes / Mask.size();
21214 for (int i = 0; i < NumBytes; ++i) {
21215 if (Mask[i / Ratio] == SM_SentinelUndef) {
21216 PSHUFBMask.push_back(DAG.getUNDEF(MVT::i8));
21219 int M = Mask[i / Ratio] != SM_SentinelZero
21220 ? Ratio * Mask[i / Ratio] + i % Ratio
21222 PSHUFBMask.push_back(DAG.getConstant(M, DL, MVT::i8));
21224 MVT ByteVT = MVT::getVectorVT(MVT::i8, NumBytes);
21225 Op = DAG.getBitcast(ByteVT, Input);
21226 DCI.AddToWorklist(Op.getNode());
21227 SDValue PSHUFBMaskOp =
21228 DAG.getNode(ISD::BUILD_VECTOR, DL, ByteVT, PSHUFBMask);
21229 DCI.AddToWorklist(PSHUFBMaskOp.getNode());
21230 Op = DAG.getNode(X86ISD::PSHUFB, DL, ByteVT, Op, PSHUFBMaskOp);
21231 DCI.AddToWorklist(Op.getNode());
21232 DCI.CombineTo(Root.getNode(), DAG.getBitcast(RootVT, Op),
21237 // Failed to find any combines.
21241 /// \brief Fully generic combining of x86 shuffle instructions.
21243 /// This should be the last combine run over the x86 shuffle instructions. Once
21244 /// they have been fully optimized, this will recursively consider all chains
21245 /// of single-use shuffle instructions, build a generic model of the cumulative
21246 /// shuffle operation, and check for simpler instructions which implement this
21247 /// operation. We use this primarily for two purposes:
21249 /// 1) Collapse generic shuffles to specialized single instructions when
21250 /// equivalent. In most cases, this is just an encoding size win, but
21251 /// sometimes we will collapse multiple generic shuffles into a single
21252 /// special-purpose shuffle.
21253 /// 2) Look for sequences of shuffle instructions with 3 or more total
21254 /// instructions, and replace them with the slightly more expensive SSSE3
21255 /// PSHUFB instruction if available. We do this as the last combining step
21256 /// to ensure we avoid using PSHUFB if we can implement the shuffle with
21257 /// a suitable short sequence of other instructions. The PHUFB will either
21258 /// use a register or have to read from memory and so is slightly (but only
21259 /// slightly) more expensive than the other shuffle instructions.
21261 /// Because this is inherently a quadratic operation (for each shuffle in
21262 /// a chain, we recurse up the chain), the depth is limited to 8 instructions.
21263 /// This should never be an issue in practice as the shuffle lowering doesn't
21264 /// produce sequences of more than 8 instructions.
21266 /// FIXME: We will currently miss some cases where the redundant shuffling
21267 /// would simplify under the threshold for PSHUFB formation because of
21268 /// combine-ordering. To fix this, we should do the redundant instruction
21269 /// combining in this recursive walk.
21270 static bool combineX86ShufflesRecursively(SDValue Op, SDValue Root,
21271 ArrayRef<int> RootMask,
21272 int Depth, bool HasPSHUFB,
21274 TargetLowering::DAGCombinerInfo &DCI,
21275 const X86Subtarget *Subtarget) {
21276 // Bound the depth of our recursive combine because this is ultimately
21277 // quadratic in nature.
21281 // Directly rip through bitcasts to find the underlying operand.
21282 while (Op.getOpcode() == ISD::BITCAST && Op.getOperand(0).hasOneUse())
21283 Op = Op.getOperand(0);
21285 MVT VT = Op.getSimpleValueType();
21286 if (!VT.isVector())
21287 return false; // Bail if we hit a non-vector.
21289 assert(Root.getSimpleValueType().isVector() &&
21290 "Shuffles operate on vector types!");
21291 assert(VT.getSizeInBits() == Root.getSimpleValueType().getSizeInBits() &&
21292 "Can only combine shuffles of the same vector register size.");
21294 if (!isTargetShuffle(Op.getOpcode()))
21296 SmallVector<int, 16> OpMask;
21298 bool HaveMask = getTargetShuffleMask(Op.getNode(), VT, OpMask, IsUnary);
21299 // We only can combine unary shuffles which we can decode the mask for.
21300 if (!HaveMask || !IsUnary)
21303 assert(VT.getVectorNumElements() == OpMask.size() &&
21304 "Different mask size from vector size!");
21305 assert(((RootMask.size() > OpMask.size() &&
21306 RootMask.size() % OpMask.size() == 0) ||
21307 (OpMask.size() > RootMask.size() &&
21308 OpMask.size() % RootMask.size() == 0) ||
21309 OpMask.size() == RootMask.size()) &&
21310 "The smaller number of elements must divide the larger.");
21311 int RootRatio = std::max<int>(1, OpMask.size() / RootMask.size());
21312 int OpRatio = std::max<int>(1, RootMask.size() / OpMask.size());
21313 assert(((RootRatio == 1 && OpRatio == 1) ||
21314 (RootRatio == 1) != (OpRatio == 1)) &&
21315 "Must not have a ratio for both incoming and op masks!");
21317 SmallVector<int, 16> Mask;
21318 Mask.reserve(std::max(OpMask.size(), RootMask.size()));
21320 // Merge this shuffle operation's mask into our accumulated mask. Note that
21321 // this shuffle's mask will be the first applied to the input, followed by the
21322 // root mask to get us all the way to the root value arrangement. The reason
21323 // for this order is that we are recursing up the operation chain.
21324 for (int i = 0, e = std::max(OpMask.size(), RootMask.size()); i < e; ++i) {
21325 int RootIdx = i / RootRatio;
21326 if (RootMask[RootIdx] < 0) {
21327 // This is a zero or undef lane, we're done.
21328 Mask.push_back(RootMask[RootIdx]);
21332 int RootMaskedIdx = RootMask[RootIdx] * RootRatio + i % RootRatio;
21333 int OpIdx = RootMaskedIdx / OpRatio;
21334 if (OpMask[OpIdx] < 0) {
21335 // The incoming lanes are zero or undef, it doesn't matter which ones we
21337 Mask.push_back(OpMask[OpIdx]);
21341 // Ok, we have non-zero lanes, map them through.
21342 Mask.push_back(OpMask[OpIdx] * OpRatio +
21343 RootMaskedIdx % OpRatio);
21346 // See if we can recurse into the operand to combine more things.
21347 switch (Op.getOpcode()) {
21348 case X86ISD::PSHUFB:
21350 case X86ISD::PSHUFD:
21351 case X86ISD::PSHUFHW:
21352 case X86ISD::PSHUFLW:
21353 if (Op.getOperand(0).hasOneUse() &&
21354 combineX86ShufflesRecursively(Op.getOperand(0), Root, Mask, Depth + 1,
21355 HasPSHUFB, DAG, DCI, Subtarget))
21359 case X86ISD::UNPCKL:
21360 case X86ISD::UNPCKH:
21361 assert(Op.getOperand(0) == Op.getOperand(1) && "We only combine unary shuffles!");
21362 // We can't check for single use, we have to check that this shuffle is the only user.
21363 if (Op->isOnlyUserOf(Op.getOperand(0).getNode()) &&
21364 combineX86ShufflesRecursively(Op.getOperand(0), Root, Mask, Depth + 1,
21365 HasPSHUFB, DAG, DCI, Subtarget))
21370 // Minor canonicalization of the accumulated shuffle mask to make it easier
21371 // to match below. All this does is detect masks with squential pairs of
21372 // elements, and shrink them to the half-width mask. It does this in a loop
21373 // so it will reduce the size of the mask to the minimal width mask which
21374 // performs an equivalent shuffle.
21375 SmallVector<int, 16> WidenedMask;
21376 while (Mask.size() > 1 && canWidenShuffleElements(Mask, WidenedMask)) {
21377 Mask = std::move(WidenedMask);
21378 WidenedMask.clear();
21381 return combineX86ShuffleChain(Op, Root, Mask, Depth, HasPSHUFB, DAG, DCI,
21385 /// \brief Get the PSHUF-style mask from PSHUF node.
21387 /// This is a very minor wrapper around getTargetShuffleMask to easy forming v4
21388 /// PSHUF-style masks that can be reused with such instructions.
21389 static SmallVector<int, 4> getPSHUFShuffleMask(SDValue N) {
21390 MVT VT = N.getSimpleValueType();
21391 SmallVector<int, 4> Mask;
21393 bool HaveMask = getTargetShuffleMask(N.getNode(), VT, Mask, IsUnary);
21397 // If we have more than 128-bits, only the low 128-bits of shuffle mask
21398 // matter. Check that the upper masks are repeats and remove them.
21399 if (VT.getSizeInBits() > 128) {
21400 int LaneElts = 128 / VT.getScalarSizeInBits();
21402 for (int i = 1, NumLanes = VT.getSizeInBits() / 128; i < NumLanes; ++i)
21403 for (int j = 0; j < LaneElts; ++j)
21404 assert(Mask[j] == Mask[i * LaneElts + j] - (LaneElts * i) &&
21405 "Mask doesn't repeat in high 128-bit lanes!");
21407 Mask.resize(LaneElts);
21410 switch (N.getOpcode()) {
21411 case X86ISD::PSHUFD:
21413 case X86ISD::PSHUFLW:
21416 case X86ISD::PSHUFHW:
21417 Mask.erase(Mask.begin(), Mask.begin() + 4);
21418 for (int &M : Mask)
21422 llvm_unreachable("No valid shuffle instruction found!");
21426 /// \brief Search for a combinable shuffle across a chain ending in pshufd.
21428 /// We walk up the chain and look for a combinable shuffle, skipping over
21429 /// shuffles that we could hoist this shuffle's transformation past without
21430 /// altering anything.
21432 combineRedundantDWordShuffle(SDValue N, MutableArrayRef<int> Mask,
21434 TargetLowering::DAGCombinerInfo &DCI) {
21435 assert(N.getOpcode() == X86ISD::PSHUFD &&
21436 "Called with something other than an x86 128-bit half shuffle!");
21439 // Walk up a single-use chain looking for a combinable shuffle. Keep a stack
21440 // of the shuffles in the chain so that we can form a fresh chain to replace
21442 SmallVector<SDValue, 8> Chain;
21443 SDValue V = N.getOperand(0);
21444 for (; V.hasOneUse(); V = V.getOperand(0)) {
21445 switch (V.getOpcode()) {
21447 return SDValue(); // Nothing combined!
21450 // Skip bitcasts as we always know the type for the target specific
21454 case X86ISD::PSHUFD:
21455 // Found another dword shuffle.
21458 case X86ISD::PSHUFLW:
21459 // Check that the low words (being shuffled) are the identity in the
21460 // dword shuffle, and the high words are self-contained.
21461 if (Mask[0] != 0 || Mask[1] != 1 ||
21462 !(Mask[2] >= 2 && Mask[2] < 4 && Mask[3] >= 2 && Mask[3] < 4))
21465 Chain.push_back(V);
21468 case X86ISD::PSHUFHW:
21469 // Check that the high words (being shuffled) are the identity in the
21470 // dword shuffle, and the low words are self-contained.
21471 if (Mask[2] != 2 || Mask[3] != 3 ||
21472 !(Mask[0] >= 0 && Mask[0] < 2 && Mask[1] >= 0 && Mask[1] < 2))
21475 Chain.push_back(V);
21478 case X86ISD::UNPCKL:
21479 case X86ISD::UNPCKH:
21480 // For either i8 -> i16 or i16 -> i32 unpacks, we can combine a dword
21481 // shuffle into a preceding word shuffle.
21482 if (V.getSimpleValueType().getScalarType() != MVT::i8 &&
21483 V.getSimpleValueType().getScalarType() != MVT::i16)
21486 // Search for a half-shuffle which we can combine with.
21487 unsigned CombineOp =
21488 V.getOpcode() == X86ISD::UNPCKL ? X86ISD::PSHUFLW : X86ISD::PSHUFHW;
21489 if (V.getOperand(0) != V.getOperand(1) ||
21490 !V->isOnlyUserOf(V.getOperand(0).getNode()))
21492 Chain.push_back(V);
21493 V = V.getOperand(0);
21495 switch (V.getOpcode()) {
21497 return SDValue(); // Nothing to combine.
21499 case X86ISD::PSHUFLW:
21500 case X86ISD::PSHUFHW:
21501 if (V.getOpcode() == CombineOp)
21504 Chain.push_back(V);
21508 V = V.getOperand(0);
21512 } while (V.hasOneUse());
21515 // Break out of the loop if we break out of the switch.
21519 if (!V.hasOneUse())
21520 // We fell out of the loop without finding a viable combining instruction.
21523 // Merge this node's mask and our incoming mask.
21524 SmallVector<int, 4> VMask = getPSHUFShuffleMask(V);
21525 for (int &M : Mask)
21527 V = DAG.getNode(V.getOpcode(), DL, V.getValueType(), V.getOperand(0),
21528 getV4X86ShuffleImm8ForMask(Mask, DL, DAG));
21530 // Rebuild the chain around this new shuffle.
21531 while (!Chain.empty()) {
21532 SDValue W = Chain.pop_back_val();
21534 if (V.getValueType() != W.getOperand(0).getValueType())
21535 V = DAG.getBitcast(W.getOperand(0).getValueType(), V);
21537 switch (W.getOpcode()) {
21539 llvm_unreachable("Only PSHUF and UNPCK instructions get here!");
21541 case X86ISD::UNPCKL:
21542 case X86ISD::UNPCKH:
21543 V = DAG.getNode(W.getOpcode(), DL, W.getValueType(), V, V);
21546 case X86ISD::PSHUFD:
21547 case X86ISD::PSHUFLW:
21548 case X86ISD::PSHUFHW:
21549 V = DAG.getNode(W.getOpcode(), DL, W.getValueType(), V, W.getOperand(1));
21553 if (V.getValueType() != N.getValueType())
21554 V = DAG.getBitcast(N.getValueType(), V);
21556 // Return the new chain to replace N.
21560 /// \brief Search for a combinable shuffle across a chain ending in pshuflw or pshufhw.
21562 /// We walk up the chain, skipping shuffles of the other half and looking
21563 /// through shuffles which switch halves trying to find a shuffle of the same
21564 /// pair of dwords.
21565 static bool combineRedundantHalfShuffle(SDValue N, MutableArrayRef<int> Mask,
21567 TargetLowering::DAGCombinerInfo &DCI) {
21569 (N.getOpcode() == X86ISD::PSHUFLW || N.getOpcode() == X86ISD::PSHUFHW) &&
21570 "Called with something other than an x86 128-bit half shuffle!");
21572 unsigned CombineOpcode = N.getOpcode();
21574 // Walk up a single-use chain looking for a combinable shuffle.
21575 SDValue V = N.getOperand(0);
21576 for (; V.hasOneUse(); V = V.getOperand(0)) {
21577 switch (V.getOpcode()) {
21579 return false; // Nothing combined!
21582 // Skip bitcasts as we always know the type for the target specific
21586 case X86ISD::PSHUFLW:
21587 case X86ISD::PSHUFHW:
21588 if (V.getOpcode() == CombineOpcode)
21591 // Other-half shuffles are no-ops.
21594 // Break out of the loop if we break out of the switch.
21598 if (!V.hasOneUse())
21599 // We fell out of the loop without finding a viable combining instruction.
21602 // Combine away the bottom node as its shuffle will be accumulated into
21603 // a preceding shuffle.
21604 DCI.CombineTo(N.getNode(), N.getOperand(0), /*AddTo*/ true);
21606 // Record the old value.
21609 // Merge this node's mask and our incoming mask (adjusted to account for all
21610 // the pshufd instructions encountered).
21611 SmallVector<int, 4> VMask = getPSHUFShuffleMask(V);
21612 for (int &M : Mask)
21614 V = DAG.getNode(V.getOpcode(), DL, MVT::v8i16, V.getOperand(0),
21615 getV4X86ShuffleImm8ForMask(Mask, DL, DAG));
21617 // Check that the shuffles didn't cancel each other out. If not, we need to
21618 // combine to the new one.
21620 // Replace the combinable shuffle with the combined one, updating all users
21621 // so that we re-evaluate the chain here.
21622 DCI.CombineTo(Old.getNode(), V, /*AddTo*/ true);
21627 /// \brief Try to combine x86 target specific shuffles.
21628 static SDValue PerformTargetShuffleCombine(SDValue N, SelectionDAG &DAG,
21629 TargetLowering::DAGCombinerInfo &DCI,
21630 const X86Subtarget *Subtarget) {
21632 MVT VT = N.getSimpleValueType();
21633 SmallVector<int, 4> Mask;
21635 switch (N.getOpcode()) {
21636 case X86ISD::PSHUFD:
21637 case X86ISD::PSHUFLW:
21638 case X86ISD::PSHUFHW:
21639 Mask = getPSHUFShuffleMask(N);
21640 assert(Mask.size() == 4);
21646 // Nuke no-op shuffles that show up after combining.
21647 if (isNoopShuffleMask(Mask))
21648 return DCI.CombineTo(N.getNode(), N.getOperand(0), /*AddTo*/ true);
21650 // Look for simplifications involving one or two shuffle instructions.
21651 SDValue V = N.getOperand(0);
21652 switch (N.getOpcode()) {
21655 case X86ISD::PSHUFLW:
21656 case X86ISD::PSHUFHW:
21657 assert(VT.getScalarType() == MVT::i16 && "Bad word shuffle type!");
21659 if (combineRedundantHalfShuffle(N, Mask, DAG, DCI))
21660 return SDValue(); // We combined away this shuffle, so we're done.
21662 // See if this reduces to a PSHUFD which is no more expensive and can
21663 // combine with more operations. Note that it has to at least flip the
21664 // dwords as otherwise it would have been removed as a no-op.
21665 if (makeArrayRef(Mask).equals({2, 3, 0, 1})) {
21666 int DMask[] = {0, 1, 2, 3};
21667 int DOffset = N.getOpcode() == X86ISD::PSHUFLW ? 0 : 2;
21668 DMask[DOffset + 0] = DOffset + 1;
21669 DMask[DOffset + 1] = DOffset + 0;
21670 MVT DVT = MVT::getVectorVT(MVT::i32, VT.getVectorNumElements() / 2);
21671 V = DAG.getBitcast(DVT, V);
21672 DCI.AddToWorklist(V.getNode());
21673 V = DAG.getNode(X86ISD::PSHUFD, DL, DVT, V,
21674 getV4X86ShuffleImm8ForMask(DMask, DL, DAG));
21675 DCI.AddToWorklist(V.getNode());
21676 return DAG.getBitcast(VT, V);
21679 // Look for shuffle patterns which can be implemented as a single unpack.
21680 // FIXME: This doesn't handle the location of the PSHUFD generically, and
21681 // only works when we have a PSHUFD followed by two half-shuffles.
21682 if (Mask[0] == Mask[1] && Mask[2] == Mask[3] &&
21683 (V.getOpcode() == X86ISD::PSHUFLW ||
21684 V.getOpcode() == X86ISD::PSHUFHW) &&
21685 V.getOpcode() != N.getOpcode() &&
21687 SDValue D = V.getOperand(0);
21688 while (D.getOpcode() == ISD::BITCAST && D.hasOneUse())
21689 D = D.getOperand(0);
21690 if (D.getOpcode() == X86ISD::PSHUFD && D.hasOneUse()) {
21691 SmallVector<int, 4> VMask = getPSHUFShuffleMask(V);
21692 SmallVector<int, 4> DMask = getPSHUFShuffleMask(D);
21693 int NOffset = N.getOpcode() == X86ISD::PSHUFLW ? 0 : 4;
21694 int VOffset = V.getOpcode() == X86ISD::PSHUFLW ? 0 : 4;
21696 for (int i = 0; i < 4; ++i) {
21697 WordMask[i + NOffset] = Mask[i] + NOffset;
21698 WordMask[i + VOffset] = VMask[i] + VOffset;
21700 // Map the word mask through the DWord mask.
21702 for (int i = 0; i < 8; ++i)
21703 MappedMask[i] = 2 * DMask[WordMask[i] / 2] + WordMask[i] % 2;
21704 if (makeArrayRef(MappedMask).equals({0, 0, 1, 1, 2, 2, 3, 3}) ||
21705 makeArrayRef(MappedMask).equals({4, 4, 5, 5, 6, 6, 7, 7})) {
21706 // We can replace all three shuffles with an unpack.
21707 V = DAG.getBitcast(VT, D.getOperand(0));
21708 DCI.AddToWorklist(V.getNode());
21709 return DAG.getNode(MappedMask[0] == 0 ? X86ISD::UNPCKL
21718 case X86ISD::PSHUFD:
21719 if (SDValue NewN = combineRedundantDWordShuffle(N, Mask, DAG, DCI))
21728 /// \brief Try to combine a shuffle into a target-specific add-sub node.
21730 /// We combine this directly on the abstract vector shuffle nodes so it is
21731 /// easier to generically match. We also insert dummy vector shuffle nodes for
21732 /// the operands which explicitly discard the lanes which are unused by this
21733 /// operation to try to flow through the rest of the combiner the fact that
21734 /// they're unused.
21735 static SDValue combineShuffleToAddSub(SDNode *N, SelectionDAG &DAG) {
21737 EVT VT = N->getValueType(0);
21739 // We only handle target-independent shuffles.
21740 // FIXME: It would be easy and harmless to use the target shuffle mask
21741 // extraction tool to support more.
21742 if (N->getOpcode() != ISD::VECTOR_SHUFFLE)
21745 auto *SVN = cast<ShuffleVectorSDNode>(N);
21746 ArrayRef<int> Mask = SVN->getMask();
21747 SDValue V1 = N->getOperand(0);
21748 SDValue V2 = N->getOperand(1);
21750 // We require the first shuffle operand to be the SUB node, and the second to
21751 // be the ADD node.
21752 // FIXME: We should support the commuted patterns.
21753 if (V1->getOpcode() != ISD::FSUB || V2->getOpcode() != ISD::FADD)
21756 // If there are other uses of these operations we can't fold them.
21757 if (!V1->hasOneUse() || !V2->hasOneUse())
21760 // Ensure that both operations have the same operands. Note that we can
21761 // commute the FADD operands.
21762 SDValue LHS = V1->getOperand(0), RHS = V1->getOperand(1);
21763 if ((V2->getOperand(0) != LHS || V2->getOperand(1) != RHS) &&
21764 (V2->getOperand(0) != RHS || V2->getOperand(1) != LHS))
21767 // We're looking for blends between FADD and FSUB nodes. We insist on these
21768 // nodes being lined up in a specific expected pattern.
21769 if (!(isShuffleEquivalent(V1, V2, Mask, {0, 3}) ||
21770 isShuffleEquivalent(V1, V2, Mask, {0, 5, 2, 7}) ||
21771 isShuffleEquivalent(V1, V2, Mask, {0, 9, 2, 11, 4, 13, 6, 15})))
21774 // Only specific types are legal at this point, assert so we notice if and
21775 // when these change.
21776 assert((VT == MVT::v4f32 || VT == MVT::v2f64 || VT == MVT::v8f32 ||
21777 VT == MVT::v4f64) &&
21778 "Unknown vector type encountered!");
21780 return DAG.getNode(X86ISD::ADDSUB, DL, VT, LHS, RHS);
21783 /// PerformShuffleCombine - Performs several different shuffle combines.
21784 static SDValue PerformShuffleCombine(SDNode *N, SelectionDAG &DAG,
21785 TargetLowering::DAGCombinerInfo &DCI,
21786 const X86Subtarget *Subtarget) {
21788 SDValue N0 = N->getOperand(0);
21789 SDValue N1 = N->getOperand(1);
21790 EVT VT = N->getValueType(0);
21792 // Don't create instructions with illegal types after legalize types has run.
21793 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
21794 if (!DCI.isBeforeLegalize() && !TLI.isTypeLegal(VT.getVectorElementType()))
21797 // If we have legalized the vector types, look for blends of FADD and FSUB
21798 // nodes that we can fuse into an ADDSUB node.
21799 if (TLI.isTypeLegal(VT) && Subtarget->hasSSE3())
21800 if (SDValue AddSub = combineShuffleToAddSub(N, DAG))
21803 // Combine 256-bit vector shuffles. This is only profitable when in AVX mode
21804 if (Subtarget->hasFp256() && VT.is256BitVector() &&
21805 N->getOpcode() == ISD::VECTOR_SHUFFLE)
21806 return PerformShuffleCombine256(N, DAG, DCI, Subtarget);
21808 // During Type Legalization, when promoting illegal vector types,
21809 // the backend might introduce new shuffle dag nodes and bitcasts.
21811 // This code performs the following transformation:
21812 // fold: (shuffle (bitcast (BINOP A, B)), Undef, <Mask>) ->
21813 // (shuffle (BINOP (bitcast A), (bitcast B)), Undef, <Mask>)
21815 // We do this only if both the bitcast and the BINOP dag nodes have
21816 // one use. Also, perform this transformation only if the new binary
21817 // operation is legal. This is to avoid introducing dag nodes that
21818 // potentially need to be further expanded (or custom lowered) into a
21819 // less optimal sequence of dag nodes.
21820 if (!DCI.isBeforeLegalize() && DCI.isBeforeLegalizeOps() &&
21821 N1.getOpcode() == ISD::UNDEF && N0.hasOneUse() &&
21822 N0.getOpcode() == ISD::BITCAST) {
21823 SDValue BC0 = N0.getOperand(0);
21824 EVT SVT = BC0.getValueType();
21825 unsigned Opcode = BC0.getOpcode();
21826 unsigned NumElts = VT.getVectorNumElements();
21828 if (BC0.hasOneUse() && SVT.isVector() &&
21829 SVT.getVectorNumElements() * 2 == NumElts &&
21830 TLI.isOperationLegal(Opcode, VT)) {
21831 bool CanFold = false;
21843 unsigned SVTNumElts = SVT.getVectorNumElements();
21844 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
21845 for (unsigned i = 0, e = SVTNumElts; i != e && CanFold; ++i)
21846 CanFold = SVOp->getMaskElt(i) == (int)(i * 2);
21847 for (unsigned i = SVTNumElts, e = NumElts; i != e && CanFold; ++i)
21848 CanFold = SVOp->getMaskElt(i) < 0;
21851 SDValue BC00 = DAG.getBitcast(VT, BC0.getOperand(0));
21852 SDValue BC01 = DAG.getBitcast(VT, BC0.getOperand(1));
21853 SDValue NewBinOp = DAG.getNode(BC0.getOpcode(), dl, VT, BC00, BC01);
21854 return DAG.getVectorShuffle(VT, dl, NewBinOp, N1, &SVOp->getMask()[0]);
21859 // Combine a vector_shuffle that is equal to build_vector load1, load2, load3,
21860 // load4, <0, 1, 2, 3> into a 128-bit load if the load addresses are
21861 // consecutive, non-overlapping, and in the right order.
21862 SmallVector<SDValue, 16> Elts;
21863 for (unsigned i = 0, e = VT.getVectorNumElements(); i != e; ++i)
21864 Elts.push_back(getShuffleScalarElt(N, i, DAG, 0));
21866 if (SDValue LD = EltsFromConsecutiveLoads(VT, Elts, dl, DAG, true))
21869 if (isTargetShuffle(N->getOpcode())) {
21871 PerformTargetShuffleCombine(SDValue(N, 0), DAG, DCI, Subtarget);
21872 if (Shuffle.getNode())
21875 // Try recursively combining arbitrary sequences of x86 shuffle
21876 // instructions into higher-order shuffles. We do this after combining
21877 // specific PSHUF instruction sequences into their minimal form so that we
21878 // can evaluate how many specialized shuffle instructions are involved in
21879 // a particular chain.
21880 SmallVector<int, 1> NonceMask; // Just a placeholder.
21881 NonceMask.push_back(0);
21882 if (combineX86ShufflesRecursively(SDValue(N, 0), SDValue(N, 0), NonceMask,
21883 /*Depth*/ 1, /*HasPSHUFB*/ false, DAG,
21885 return SDValue(); // This routine will use CombineTo to replace N.
21891 /// XFormVExtractWithShuffleIntoLoad - Check if a vector extract from a target
21892 /// specific shuffle of a load can be folded into a single element load.
21893 /// Similar handling for VECTOR_SHUFFLE is performed by DAGCombiner, but
21894 /// shuffles have been custom lowered so we need to handle those here.
21895 static SDValue XFormVExtractWithShuffleIntoLoad(SDNode *N, SelectionDAG &DAG,
21896 TargetLowering::DAGCombinerInfo &DCI) {
21897 if (DCI.isBeforeLegalizeOps())
21900 SDValue InVec = N->getOperand(0);
21901 SDValue EltNo = N->getOperand(1);
21903 if (!isa<ConstantSDNode>(EltNo))
21906 EVT OriginalVT = InVec.getValueType();
21908 if (InVec.getOpcode() == ISD::BITCAST) {
21909 // Don't duplicate a load with other uses.
21910 if (!InVec.hasOneUse())
21912 EVT BCVT = InVec.getOperand(0).getValueType();
21913 if (!BCVT.isVector() ||
21914 BCVT.getVectorNumElements() != OriginalVT.getVectorNumElements())
21916 InVec = InVec.getOperand(0);
21919 EVT CurrentVT = InVec.getValueType();
21921 if (!isTargetShuffle(InVec.getOpcode()))
21924 // Don't duplicate a load with other uses.
21925 if (!InVec.hasOneUse())
21928 SmallVector<int, 16> ShuffleMask;
21930 if (!getTargetShuffleMask(InVec.getNode(), CurrentVT.getSimpleVT(),
21931 ShuffleMask, UnaryShuffle))
21934 // Select the input vector, guarding against out of range extract vector.
21935 unsigned NumElems = CurrentVT.getVectorNumElements();
21936 int Elt = cast<ConstantSDNode>(EltNo)->getZExtValue();
21937 int Idx = (Elt > (int)NumElems) ? -1 : ShuffleMask[Elt];
21938 SDValue LdNode = (Idx < (int)NumElems) ? InVec.getOperand(0)
21939 : InVec.getOperand(1);
21941 // If inputs to shuffle are the same for both ops, then allow 2 uses
21942 unsigned AllowedUses = InVec.getNumOperands() > 1 &&
21943 InVec.getOperand(0) == InVec.getOperand(1) ? 2 : 1;
21945 if (LdNode.getOpcode() == ISD::BITCAST) {
21946 // Don't duplicate a load with other uses.
21947 if (!LdNode.getNode()->hasNUsesOfValue(AllowedUses, 0))
21950 AllowedUses = 1; // only allow 1 load use if we have a bitcast
21951 LdNode = LdNode.getOperand(0);
21954 if (!ISD::isNormalLoad(LdNode.getNode()))
21957 LoadSDNode *LN0 = cast<LoadSDNode>(LdNode);
21959 if (!LN0 ||!LN0->hasNUsesOfValue(AllowedUses, 0) || LN0->isVolatile())
21962 EVT EltVT = N->getValueType(0);
21963 // If there's a bitcast before the shuffle, check if the load type and
21964 // alignment is valid.
21965 unsigned Align = LN0->getAlignment();
21966 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
21967 unsigned NewAlign = DAG.getDataLayout().getABITypeAlignment(
21968 EltVT.getTypeForEVT(*DAG.getContext()));
21970 if (NewAlign > Align || !TLI.isOperationLegalOrCustom(ISD::LOAD, EltVT))
21973 // All checks match so transform back to vector_shuffle so that DAG combiner
21974 // can finish the job
21977 // Create shuffle node taking into account the case that its a unary shuffle
21978 SDValue Shuffle = (UnaryShuffle) ? DAG.getUNDEF(CurrentVT)
21979 : InVec.getOperand(1);
21980 Shuffle = DAG.getVectorShuffle(CurrentVT, dl,
21981 InVec.getOperand(0), Shuffle,
21983 Shuffle = DAG.getBitcast(OriginalVT, Shuffle);
21984 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, N->getValueType(0), Shuffle,
21988 /// \brief Detect bitcasts between i32 to x86mmx low word. Since MMX types are
21989 /// special and don't usually play with other vector types, it's better to
21990 /// handle them early to be sure we emit efficient code by avoiding
21991 /// store-load conversions.
21992 static SDValue PerformBITCASTCombine(SDNode *N, SelectionDAG &DAG) {
21993 if (N->getValueType(0) != MVT::x86mmx ||
21994 N->getOperand(0)->getOpcode() != ISD::BUILD_VECTOR ||
21995 N->getOperand(0)->getValueType(0) != MVT::v2i32)
21998 SDValue V = N->getOperand(0);
21999 ConstantSDNode *C = dyn_cast<ConstantSDNode>(V.getOperand(1));
22000 if (C && C->getZExtValue() == 0 && V.getOperand(0).getValueType() == MVT::i32)
22001 return DAG.getNode(X86ISD::MMX_MOVW2D, SDLoc(V.getOperand(0)),
22002 N->getValueType(0), V.getOperand(0));
22007 /// PerformEXTRACT_VECTOR_ELTCombine - Detect vector gather/scatter index
22008 /// generation and convert it from being a bunch of shuffles and extracts
22009 /// into a somewhat faster sequence. For i686, the best sequence is apparently
22010 /// storing the value and loading scalars back, while for x64 we should
22011 /// use 64-bit extracts and shifts.
22012 static SDValue PerformEXTRACT_VECTOR_ELTCombine(SDNode *N, SelectionDAG &DAG,
22013 TargetLowering::DAGCombinerInfo &DCI) {
22014 if (SDValue NewOp = XFormVExtractWithShuffleIntoLoad(N, DAG, DCI))
22017 SDValue InputVector = N->getOperand(0);
22018 SDLoc dl(InputVector);
22019 // Detect mmx to i32 conversion through a v2i32 elt extract.
22020 if (InputVector.getOpcode() == ISD::BITCAST && InputVector.hasOneUse() &&
22021 N->getValueType(0) == MVT::i32 &&
22022 InputVector.getValueType() == MVT::v2i32) {
22024 // The bitcast source is a direct mmx result.
22025 SDValue MMXSrc = InputVector.getNode()->getOperand(0);
22026 if (MMXSrc.getValueType() == MVT::x86mmx)
22027 return DAG.getNode(X86ISD::MMX_MOVD2W, SDLoc(InputVector),
22028 N->getValueType(0),
22029 InputVector.getNode()->getOperand(0));
22031 // The mmx is indirect: (i64 extract_elt (v1i64 bitcast (x86mmx ...))).
22032 SDValue MMXSrcOp = MMXSrc.getOperand(0);
22033 if (MMXSrc.getOpcode() == ISD::EXTRACT_VECTOR_ELT && MMXSrc.hasOneUse() &&
22034 MMXSrc.getValueType() == MVT::i64 && MMXSrcOp.hasOneUse() &&
22035 MMXSrcOp.getOpcode() == ISD::BITCAST &&
22036 MMXSrcOp.getValueType() == MVT::v1i64 &&
22037 MMXSrcOp.getOperand(0).getValueType() == MVT::x86mmx)
22038 return DAG.getNode(X86ISD::MMX_MOVD2W, SDLoc(InputVector),
22039 N->getValueType(0),
22040 MMXSrcOp.getOperand(0));
22043 EVT VT = N->getValueType(0);
22045 if (VT == MVT::i1 && dyn_cast<ConstantSDNode>(N->getOperand(1)) &&
22046 InputVector.getOpcode() == ISD::BITCAST &&
22047 dyn_cast<ConstantSDNode>(InputVector.getOperand(0))) {
22048 uint64_t ExtractedElt =
22049 cast<ConstantSDNode>(N->getOperand(1))->getZExtValue();
22050 uint64_t InputValue =
22051 cast<ConstantSDNode>(InputVector.getOperand(0))->getZExtValue();
22052 uint64_t Res = (InputValue >> ExtractedElt) & 1;
22053 return DAG.getConstant(Res, dl, MVT::i1);
22055 // Only operate on vectors of 4 elements, where the alternative shuffling
22056 // gets to be more expensive.
22057 if (InputVector.getValueType() != MVT::v4i32)
22060 // Check whether every use of InputVector is an EXTRACT_VECTOR_ELT with a
22061 // single use which is a sign-extend or zero-extend, and all elements are
22063 SmallVector<SDNode *, 4> Uses;
22064 unsigned ExtractedElements = 0;
22065 for (SDNode::use_iterator UI = InputVector.getNode()->use_begin(),
22066 UE = InputVector.getNode()->use_end(); UI != UE; ++UI) {
22067 if (UI.getUse().getResNo() != InputVector.getResNo())
22070 SDNode *Extract = *UI;
22071 if (Extract->getOpcode() != ISD::EXTRACT_VECTOR_ELT)
22074 if (Extract->getValueType(0) != MVT::i32)
22076 if (!Extract->hasOneUse())
22078 if (Extract->use_begin()->getOpcode() != ISD::SIGN_EXTEND &&
22079 Extract->use_begin()->getOpcode() != ISD::ZERO_EXTEND)
22081 if (!isa<ConstantSDNode>(Extract->getOperand(1)))
22084 // Record which element was extracted.
22085 ExtractedElements |=
22086 1 << cast<ConstantSDNode>(Extract->getOperand(1))->getZExtValue();
22088 Uses.push_back(Extract);
22091 // If not all the elements were used, this may not be worthwhile.
22092 if (ExtractedElements != 15)
22095 // Ok, we've now decided to do the transformation.
22096 // If 64-bit shifts are legal, use the extract-shift sequence,
22097 // otherwise bounce the vector off the cache.
22098 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
22101 if (TLI.isOperationLegal(ISD::SRA, MVT::i64)) {
22102 SDValue Cst = DAG.getBitcast(MVT::v2i64, InputVector);
22103 auto &DL = DAG.getDataLayout();
22104 EVT VecIdxTy = DAG.getTargetLoweringInfo().getVectorIdxTy(DL);
22105 SDValue BottomHalf = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i64, Cst,
22106 DAG.getConstant(0, dl, VecIdxTy));
22107 SDValue TopHalf = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i64, Cst,
22108 DAG.getConstant(1, dl, VecIdxTy));
22110 SDValue ShAmt = DAG.getConstant(
22111 32, dl, DAG.getTargetLoweringInfo().getShiftAmountTy(MVT::i64, DL));
22112 Vals[0] = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, BottomHalf);
22113 Vals[1] = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32,
22114 DAG.getNode(ISD::SRA, dl, MVT::i64, BottomHalf, ShAmt));
22115 Vals[2] = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, TopHalf);
22116 Vals[3] = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32,
22117 DAG.getNode(ISD::SRA, dl, MVT::i64, TopHalf, ShAmt));
22119 // Store the value to a temporary stack slot.
22120 SDValue StackPtr = DAG.CreateStackTemporary(InputVector.getValueType());
22121 SDValue Ch = DAG.getStore(DAG.getEntryNode(), dl, InputVector, StackPtr,
22122 MachinePointerInfo(), false, false, 0);
22124 EVT ElementType = InputVector.getValueType().getVectorElementType();
22125 unsigned EltSize = ElementType.getSizeInBits() / 8;
22127 // Replace each use (extract) with a load of the appropriate element.
22128 for (unsigned i = 0; i < 4; ++i) {
22129 uint64_t Offset = EltSize * i;
22130 auto PtrVT = TLI.getPointerTy(DAG.getDataLayout());
22131 SDValue OffsetVal = DAG.getConstant(Offset, dl, PtrVT);
22133 SDValue ScalarAddr =
22134 DAG.getNode(ISD::ADD, dl, PtrVT, StackPtr, OffsetVal);
22136 // Load the scalar.
22137 Vals[i] = DAG.getLoad(ElementType, dl, Ch,
22138 ScalarAddr, MachinePointerInfo(),
22139 false, false, false, 0);
22144 // Replace the extracts
22145 for (SmallVectorImpl<SDNode *>::iterator UI = Uses.begin(),
22146 UE = Uses.end(); UI != UE; ++UI) {
22147 SDNode *Extract = *UI;
22149 SDValue Idx = Extract->getOperand(1);
22150 uint64_t IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
22151 DAG.ReplaceAllUsesOfValueWith(SDValue(Extract, 0), Vals[IdxVal]);
22154 // The replacement was made in place; don't return anything.
22158 /// \brief Matches a VSELECT onto min/max or return 0 if the node doesn't match.
22159 static std::pair<unsigned, bool>
22160 matchIntegerMINMAX(SDValue Cond, EVT VT, SDValue LHS, SDValue RHS,
22161 SelectionDAG &DAG, const X86Subtarget *Subtarget) {
22162 if (!VT.isVector())
22163 return std::make_pair(0, false);
22165 bool NeedSplit = false;
22166 switch (VT.getSimpleVT().SimpleTy) {
22167 default: return std::make_pair(0, false);
22170 if (!Subtarget->hasVLX())
22171 return std::make_pair(0, false);
22175 if (!Subtarget->hasBWI())
22176 return std::make_pair(0, false);
22180 if (!Subtarget->hasAVX512())
22181 return std::make_pair(0, false);
22186 if (!Subtarget->hasAVX2())
22188 if (!Subtarget->hasAVX())
22189 return std::make_pair(0, false);
22194 if (!Subtarget->hasSSE2())
22195 return std::make_pair(0, false);
22198 // SSE2 has only a small subset of the operations.
22199 bool hasUnsigned = Subtarget->hasSSE41() ||
22200 (Subtarget->hasSSE2() && VT == MVT::v16i8);
22201 bool hasSigned = Subtarget->hasSSE41() ||
22202 (Subtarget->hasSSE2() && VT == MVT::v8i16);
22204 ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
22207 // Check for x CC y ? x : y.
22208 if (DAG.isEqualTo(LHS, Cond.getOperand(0)) &&
22209 DAG.isEqualTo(RHS, Cond.getOperand(1))) {
22214 Opc = hasUnsigned ? ISD::UMIN : 0; break;
22217 Opc = hasUnsigned ? ISD::UMAX : 0; break;
22220 Opc = hasSigned ? ISD::SMIN : 0; break;
22223 Opc = hasSigned ? ISD::SMAX : 0; break;
22225 // Check for x CC y ? y : x -- a min/max with reversed arms.
22226 } else if (DAG.isEqualTo(LHS, Cond.getOperand(1)) &&
22227 DAG.isEqualTo(RHS, Cond.getOperand(0))) {
22232 Opc = hasUnsigned ? ISD::UMAX : 0; break;
22235 Opc = hasUnsigned ? ISD::UMIN : 0; break;
22238 Opc = hasSigned ? ISD::SMAX : 0; break;
22241 Opc = hasSigned ? ISD::SMIN : 0; break;
22245 return std::make_pair(Opc, NeedSplit);
22249 transformVSELECTtoBlendVECTOR_SHUFFLE(SDNode *N, SelectionDAG &DAG,
22250 const X86Subtarget *Subtarget) {
22252 SDValue Cond = N->getOperand(0);
22253 SDValue LHS = N->getOperand(1);
22254 SDValue RHS = N->getOperand(2);
22256 if (Cond.getOpcode() == ISD::SIGN_EXTEND) {
22257 SDValue CondSrc = Cond->getOperand(0);
22258 if (CondSrc->getOpcode() == ISD::SIGN_EXTEND_INREG)
22259 Cond = CondSrc->getOperand(0);
22262 if (!ISD::isBuildVectorOfConstantSDNodes(Cond.getNode()))
22265 // A vselect where all conditions and data are constants can be optimized into
22266 // a single vector load by SelectionDAGLegalize::ExpandBUILD_VECTOR().
22267 if (ISD::isBuildVectorOfConstantSDNodes(LHS.getNode()) &&
22268 ISD::isBuildVectorOfConstantSDNodes(RHS.getNode()))
22271 unsigned MaskValue = 0;
22272 if (!BUILD_VECTORtoBlendMask(cast<BuildVectorSDNode>(Cond), MaskValue))
22275 MVT VT = N->getSimpleValueType(0);
22276 unsigned NumElems = VT.getVectorNumElements();
22277 SmallVector<int, 8> ShuffleMask(NumElems, -1);
22278 for (unsigned i = 0; i < NumElems; ++i) {
22279 // Be sure we emit undef where we can.
22280 if (Cond.getOperand(i)->getOpcode() == ISD::UNDEF)
22281 ShuffleMask[i] = -1;
22283 ShuffleMask[i] = i + NumElems * ((MaskValue >> i) & 1);
22286 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
22287 if (!TLI.isShuffleMaskLegal(ShuffleMask, VT))
22289 return DAG.getVectorShuffle(VT, dl, LHS, RHS, &ShuffleMask[0]);
22292 /// PerformSELECTCombine - Do target-specific dag combines on SELECT and VSELECT
22294 static SDValue PerformSELECTCombine(SDNode *N, SelectionDAG &DAG,
22295 TargetLowering::DAGCombinerInfo &DCI,
22296 const X86Subtarget *Subtarget) {
22298 SDValue Cond = N->getOperand(0);
22299 // Get the LHS/RHS of the select.
22300 SDValue LHS = N->getOperand(1);
22301 SDValue RHS = N->getOperand(2);
22302 EVT VT = LHS.getValueType();
22303 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
22305 // If we have SSE[12] support, try to form min/max nodes. SSE min/max
22306 // instructions match the semantics of the common C idiom x<y?x:y but not
22307 // x<=y?x:y, because of how they handle negative zero (which can be
22308 // ignored in unsafe-math mode).
22309 // We also try to create v2f32 min/max nodes, which we later widen to v4f32.
22310 if (Cond.getOpcode() == ISD::SETCC && VT.isFloatingPoint() &&
22311 VT != MVT::f80 && (TLI.isTypeLegal(VT) || VT == MVT::v2f32) &&
22312 (Subtarget->hasSSE2() ||
22313 (Subtarget->hasSSE1() && VT.getScalarType() == MVT::f32))) {
22314 ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
22316 unsigned Opcode = 0;
22317 // Check for x CC y ? x : y.
22318 if (DAG.isEqualTo(LHS, Cond.getOperand(0)) &&
22319 DAG.isEqualTo(RHS, Cond.getOperand(1))) {
22323 // Converting this to a min would handle NaNs incorrectly, and swapping
22324 // the operands would cause it to handle comparisons between positive
22325 // and negative zero incorrectly.
22326 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)) {
22327 if (!DAG.getTarget().Options.UnsafeFPMath &&
22328 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
22330 std::swap(LHS, RHS);
22332 Opcode = X86ISD::FMIN;
22335 // Converting this to a min would handle comparisons between positive
22336 // and negative zero incorrectly.
22337 if (!DAG.getTarget().Options.UnsafeFPMath &&
22338 !DAG.isKnownNeverZero(LHS) && !DAG.isKnownNeverZero(RHS))
22340 Opcode = X86ISD::FMIN;
22343 // Converting this to a min would handle both negative zeros and NaNs
22344 // incorrectly, but we can swap the operands to fix both.
22345 std::swap(LHS, RHS);
22349 Opcode = X86ISD::FMIN;
22353 // Converting this to a max would handle comparisons between positive
22354 // and negative zero incorrectly.
22355 if (!DAG.getTarget().Options.UnsafeFPMath &&
22356 !DAG.isKnownNeverZero(LHS) && !DAG.isKnownNeverZero(RHS))
22358 Opcode = X86ISD::FMAX;
22361 // Converting this to a max would handle NaNs incorrectly, and swapping
22362 // the operands would cause it to handle comparisons between positive
22363 // and negative zero incorrectly.
22364 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)) {
22365 if (!DAG.getTarget().Options.UnsafeFPMath &&
22366 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
22368 std::swap(LHS, RHS);
22370 Opcode = X86ISD::FMAX;
22373 // Converting this to a max would handle both negative zeros and NaNs
22374 // incorrectly, but we can swap the operands to fix both.
22375 std::swap(LHS, RHS);
22379 Opcode = X86ISD::FMAX;
22382 // Check for x CC y ? y : x -- a min/max with reversed arms.
22383 } else if (DAG.isEqualTo(LHS, Cond.getOperand(1)) &&
22384 DAG.isEqualTo(RHS, Cond.getOperand(0))) {
22388 // Converting this to a min would handle comparisons between positive
22389 // and negative zero incorrectly, and swapping the operands would
22390 // cause it to handle NaNs incorrectly.
22391 if (!DAG.getTarget().Options.UnsafeFPMath &&
22392 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS))) {
22393 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS))
22395 std::swap(LHS, RHS);
22397 Opcode = X86ISD::FMIN;
22400 // Converting this to a min would handle NaNs incorrectly.
22401 if (!DAG.getTarget().Options.UnsafeFPMath &&
22402 (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)))
22404 Opcode = X86ISD::FMIN;
22407 // Converting this to a min would handle both negative zeros and NaNs
22408 // incorrectly, but we can swap the operands to fix both.
22409 std::swap(LHS, RHS);
22413 Opcode = X86ISD::FMIN;
22417 // Converting this to a max would handle NaNs incorrectly.
22418 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS))
22420 Opcode = X86ISD::FMAX;
22423 // Converting this to a max would handle comparisons between positive
22424 // and negative zero incorrectly, and swapping the operands would
22425 // cause it to handle NaNs incorrectly.
22426 if (!DAG.getTarget().Options.UnsafeFPMath &&
22427 !DAG.isKnownNeverZero(LHS) && !DAG.isKnownNeverZero(RHS)) {
22428 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS))
22430 std::swap(LHS, RHS);
22432 Opcode = X86ISD::FMAX;
22435 // Converting this to a max would handle both negative zeros and NaNs
22436 // incorrectly, but we can swap the operands to fix both.
22437 std::swap(LHS, RHS);
22441 Opcode = X86ISD::FMAX;
22447 return DAG.getNode(Opcode, DL, N->getValueType(0), LHS, RHS);
22450 EVT CondVT = Cond.getValueType();
22451 if (Subtarget->hasAVX512() && VT.isVector() && CondVT.isVector() &&
22452 CondVT.getVectorElementType() == MVT::i1) {
22453 // v16i8 (select v16i1, v16i8, v16i8) does not have a proper
22454 // lowering on KNL. In this case we convert it to
22455 // v16i8 (select v16i8, v16i8, v16i8) and use AVX instruction.
22456 // The same situation for all 128 and 256-bit vectors of i8 and i16.
22457 // Since SKX these selects have a proper lowering.
22458 EVT OpVT = LHS.getValueType();
22459 if ((OpVT.is128BitVector() || OpVT.is256BitVector()) &&
22460 (OpVT.getVectorElementType() == MVT::i8 ||
22461 OpVT.getVectorElementType() == MVT::i16) &&
22462 !(Subtarget->hasBWI() && Subtarget->hasVLX())) {
22463 Cond = DAG.getNode(ISD::SIGN_EXTEND, DL, OpVT, Cond);
22464 DCI.AddToWorklist(Cond.getNode());
22465 return DAG.getNode(N->getOpcode(), DL, OpVT, Cond, LHS, RHS);
22468 // If this is a select between two integer constants, try to do some
22470 if (ConstantSDNode *TrueC = dyn_cast<ConstantSDNode>(LHS)) {
22471 if (ConstantSDNode *FalseC = dyn_cast<ConstantSDNode>(RHS))
22472 // Don't do this for crazy integer types.
22473 if (DAG.getTargetLoweringInfo().isTypeLegal(LHS.getValueType())) {
22474 // If this is efficiently invertible, canonicalize the LHSC/RHSC values
22475 // so that TrueC (the true value) is larger than FalseC.
22476 bool NeedsCondInvert = false;
22478 if (TrueC->getAPIntValue().ult(FalseC->getAPIntValue()) &&
22479 // Efficiently invertible.
22480 (Cond.getOpcode() == ISD::SETCC || // setcc -> invertible.
22481 (Cond.getOpcode() == ISD::XOR && // xor(X, C) -> invertible.
22482 isa<ConstantSDNode>(Cond.getOperand(1))))) {
22483 NeedsCondInvert = true;
22484 std::swap(TrueC, FalseC);
22487 // Optimize C ? 8 : 0 -> zext(C) << 3. Likewise for any pow2/0.
22488 if (FalseC->getAPIntValue() == 0 &&
22489 TrueC->getAPIntValue().isPowerOf2()) {
22490 if (NeedsCondInvert) // Invert the condition if needed.
22491 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
22492 DAG.getConstant(1, DL, Cond.getValueType()));
22494 // Zero extend the condition if needed.
22495 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, LHS.getValueType(), Cond);
22497 unsigned ShAmt = TrueC->getAPIntValue().logBase2();
22498 return DAG.getNode(ISD::SHL, DL, LHS.getValueType(), Cond,
22499 DAG.getConstant(ShAmt, DL, MVT::i8));
22502 // Optimize Cond ? cst+1 : cst -> zext(setcc(C)+cst.
22503 if (FalseC->getAPIntValue()+1 == TrueC->getAPIntValue()) {
22504 if (NeedsCondInvert) // Invert the condition if needed.
22505 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
22506 DAG.getConstant(1, DL, Cond.getValueType()));
22508 // Zero extend the condition if needed.
22509 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL,
22510 FalseC->getValueType(0), Cond);
22511 return DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
22512 SDValue(FalseC, 0));
22515 // Optimize cases that will turn into an LEA instruction. This requires
22516 // an i32 or i64 and an efficient multiplier (1, 2, 3, 4, 5, 8, 9).
22517 if (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i64) {
22518 uint64_t Diff = TrueC->getZExtValue()-FalseC->getZExtValue();
22519 if (N->getValueType(0) == MVT::i32) Diff = (unsigned)Diff;
22521 bool isFastMultiplier = false;
22523 switch ((unsigned char)Diff) {
22525 case 1: // result = add base, cond
22526 case 2: // result = lea base( , cond*2)
22527 case 3: // result = lea base(cond, cond*2)
22528 case 4: // result = lea base( , cond*4)
22529 case 5: // result = lea base(cond, cond*4)
22530 case 8: // result = lea base( , cond*8)
22531 case 9: // result = lea base(cond, cond*8)
22532 isFastMultiplier = true;
22537 if (isFastMultiplier) {
22538 APInt Diff = TrueC->getAPIntValue()-FalseC->getAPIntValue();
22539 if (NeedsCondInvert) // Invert the condition if needed.
22540 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
22541 DAG.getConstant(1, DL, Cond.getValueType()));
22543 // Zero extend the condition if needed.
22544 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, FalseC->getValueType(0),
22546 // Scale the condition by the difference.
22548 Cond = DAG.getNode(ISD::MUL, DL, Cond.getValueType(), Cond,
22549 DAG.getConstant(Diff, DL,
22550 Cond.getValueType()));
22552 // Add the base if non-zero.
22553 if (FalseC->getAPIntValue() != 0)
22554 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
22555 SDValue(FalseC, 0));
22562 // Canonicalize max and min:
22563 // (x > y) ? x : y -> (x >= y) ? x : y
22564 // (x < y) ? x : y -> (x <= y) ? x : y
22565 // This allows use of COND_S / COND_NS (see TranslateX86CC) which eliminates
22566 // the need for an extra compare
22567 // against zero. e.g.
22568 // (x - y) > 0 : (x - y) ? 0 -> (x - y) >= 0 : (x - y) ? 0
22570 // testl %edi, %edi
22572 // cmovgl %edi, %eax
22576 // cmovsl %eax, %edi
22577 if (N->getOpcode() == ISD::SELECT && Cond.getOpcode() == ISD::SETCC &&
22578 DAG.isEqualTo(LHS, Cond.getOperand(0)) &&
22579 DAG.isEqualTo(RHS, Cond.getOperand(1))) {
22580 ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
22585 ISD::CondCode NewCC = (CC == ISD::SETLT) ? ISD::SETLE : ISD::SETGE;
22586 Cond = DAG.getSetCC(SDLoc(Cond), Cond.getValueType(),
22587 Cond.getOperand(0), Cond.getOperand(1), NewCC);
22588 return DAG.getNode(ISD::SELECT, DL, VT, Cond, LHS, RHS);
22593 // Early exit check
22594 if (!TLI.isTypeLegal(VT))
22597 // Match VSELECTs into subs with unsigned saturation.
22598 if (N->getOpcode() == ISD::VSELECT && Cond.getOpcode() == ISD::SETCC &&
22599 // psubus is available in SSE2 and AVX2 for i8 and i16 vectors.
22600 ((Subtarget->hasSSE2() && (VT == MVT::v16i8 || VT == MVT::v8i16)) ||
22601 (Subtarget->hasAVX2() && (VT == MVT::v32i8 || VT == MVT::v16i16)))) {
22602 ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
22604 // Check if one of the arms of the VSELECT is a zero vector. If it's on the
22605 // left side invert the predicate to simplify logic below.
22607 if (ISD::isBuildVectorAllZeros(LHS.getNode())) {
22609 CC = ISD::getSetCCInverse(CC, true);
22610 } else if (ISD::isBuildVectorAllZeros(RHS.getNode())) {
22614 if (Other.getNode() && Other->getNumOperands() == 2 &&
22615 DAG.isEqualTo(Other->getOperand(0), Cond.getOperand(0))) {
22616 SDValue OpLHS = Other->getOperand(0), OpRHS = Other->getOperand(1);
22617 SDValue CondRHS = Cond->getOperand(1);
22619 // Look for a general sub with unsigned saturation first.
22620 // x >= y ? x-y : 0 --> subus x, y
22621 // x > y ? x-y : 0 --> subus x, y
22622 if ((CC == ISD::SETUGE || CC == ISD::SETUGT) &&
22623 Other->getOpcode() == ISD::SUB && DAG.isEqualTo(OpRHS, CondRHS))
22624 return DAG.getNode(X86ISD::SUBUS, DL, VT, OpLHS, OpRHS);
22626 if (auto *OpRHSBV = dyn_cast<BuildVectorSDNode>(OpRHS))
22627 if (auto *OpRHSConst = OpRHSBV->getConstantSplatNode()) {
22628 if (auto *CondRHSBV = dyn_cast<BuildVectorSDNode>(CondRHS))
22629 if (auto *CondRHSConst = CondRHSBV->getConstantSplatNode())
22630 // If the RHS is a constant we have to reverse the const
22631 // canonicalization.
22632 // x > C-1 ? x+-C : 0 --> subus x, C
22633 if (CC == ISD::SETUGT && Other->getOpcode() == ISD::ADD &&
22634 CondRHSConst->getAPIntValue() ==
22635 (-OpRHSConst->getAPIntValue() - 1))
22636 return DAG.getNode(
22637 X86ISD::SUBUS, DL, VT, OpLHS,
22638 DAG.getConstant(-OpRHSConst->getAPIntValue(), DL, VT));
22640 // Another special case: If C was a sign bit, the sub has been
22641 // canonicalized into a xor.
22642 // FIXME: Would it be better to use computeKnownBits to determine
22643 // whether it's safe to decanonicalize the xor?
22644 // x s< 0 ? x^C : 0 --> subus x, C
22645 if (CC == ISD::SETLT && Other->getOpcode() == ISD::XOR &&
22646 ISD::isBuildVectorAllZeros(CondRHS.getNode()) &&
22647 OpRHSConst->getAPIntValue().isSignBit())
22648 // Note that we have to rebuild the RHS constant here to ensure we
22649 // don't rely on particular values of undef lanes.
22650 return DAG.getNode(
22651 X86ISD::SUBUS, DL, VT, OpLHS,
22652 DAG.getConstant(OpRHSConst->getAPIntValue(), DL, VT));
22657 // Try to match a min/max vector operation.
22658 if (N->getOpcode() == ISD::VSELECT && Cond.getOpcode() == ISD::SETCC) {
22659 std::pair<unsigned, bool> ret = matchIntegerMINMAX(Cond, VT, LHS, RHS, DAG, Subtarget);
22660 unsigned Opc = ret.first;
22661 bool NeedSplit = ret.second;
22663 if (Opc && NeedSplit) {
22664 unsigned NumElems = VT.getVectorNumElements();
22665 // Extract the LHS vectors
22666 SDValue LHS1 = Extract128BitVector(LHS, 0, DAG, DL);
22667 SDValue LHS2 = Extract128BitVector(LHS, NumElems/2, DAG, DL);
22669 // Extract the RHS vectors
22670 SDValue RHS1 = Extract128BitVector(RHS, 0, DAG, DL);
22671 SDValue RHS2 = Extract128BitVector(RHS, NumElems/2, DAG, DL);
22673 // Create min/max for each subvector
22674 LHS = DAG.getNode(Opc, DL, LHS1.getValueType(), LHS1, RHS1);
22675 RHS = DAG.getNode(Opc, DL, LHS2.getValueType(), LHS2, RHS2);
22677 // Merge the result
22678 return DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, LHS, RHS);
22680 return DAG.getNode(Opc, DL, VT, LHS, RHS);
22683 // Simplify vector selection if condition value type matches vselect
22685 if (N->getOpcode() == ISD::VSELECT && CondVT == VT) {
22686 assert(Cond.getValueType().isVector() &&
22687 "vector select expects a vector selector!");
22689 bool TValIsAllOnes = ISD::isBuildVectorAllOnes(LHS.getNode());
22690 bool FValIsAllZeros = ISD::isBuildVectorAllZeros(RHS.getNode());
22692 // Try invert the condition if true value is not all 1s and false value
22694 if (!TValIsAllOnes && !FValIsAllZeros &&
22695 // Check if the selector will be produced by CMPP*/PCMP*
22696 Cond.getOpcode() == ISD::SETCC &&
22697 // Check if SETCC has already been promoted
22698 TLI.getSetCCResultType(DAG.getDataLayout(), *DAG.getContext(), VT) ==
22700 bool TValIsAllZeros = ISD::isBuildVectorAllZeros(LHS.getNode());
22701 bool FValIsAllOnes = ISD::isBuildVectorAllOnes(RHS.getNode());
22703 if (TValIsAllZeros || FValIsAllOnes) {
22704 SDValue CC = Cond.getOperand(2);
22705 ISD::CondCode NewCC =
22706 ISD::getSetCCInverse(cast<CondCodeSDNode>(CC)->get(),
22707 Cond.getOperand(0).getValueType().isInteger());
22708 Cond = DAG.getSetCC(DL, CondVT, Cond.getOperand(0), Cond.getOperand(1), NewCC);
22709 std::swap(LHS, RHS);
22710 TValIsAllOnes = FValIsAllOnes;
22711 FValIsAllZeros = TValIsAllZeros;
22715 if (TValIsAllOnes || FValIsAllZeros) {
22718 if (TValIsAllOnes && FValIsAllZeros)
22720 else if (TValIsAllOnes)
22722 DAG.getNode(ISD::OR, DL, CondVT, Cond, DAG.getBitcast(CondVT, RHS));
22723 else if (FValIsAllZeros)
22724 Ret = DAG.getNode(ISD::AND, DL, CondVT, Cond,
22725 DAG.getBitcast(CondVT, LHS));
22727 return DAG.getBitcast(VT, Ret);
22731 // We should generate an X86ISD::BLENDI from a vselect if its argument
22732 // is a sign_extend_inreg of an any_extend of a BUILD_VECTOR of
22733 // constants. This specific pattern gets generated when we split a
22734 // selector for a 512 bit vector in a machine without AVX512 (but with
22735 // 256-bit vectors), during legalization:
22737 // (vselect (sign_extend (any_extend (BUILD_VECTOR)) i1) LHS RHS)
22739 // Iff we find this pattern and the build_vectors are built from
22740 // constants, we translate the vselect into a shuffle_vector that we
22741 // know will be matched by LowerVECTOR_SHUFFLEtoBlend.
22742 if ((N->getOpcode() == ISD::VSELECT ||
22743 N->getOpcode() == X86ISD::SHRUNKBLEND) &&
22744 !DCI.isBeforeLegalize() && !VT.is512BitVector()) {
22745 SDValue Shuffle = transformVSELECTtoBlendVECTOR_SHUFFLE(N, DAG, Subtarget);
22746 if (Shuffle.getNode())
22750 // If this is a *dynamic* select (non-constant condition) and we can match
22751 // this node with one of the variable blend instructions, restructure the
22752 // condition so that the blends can use the high bit of each element and use
22753 // SimplifyDemandedBits to simplify the condition operand.
22754 if (N->getOpcode() == ISD::VSELECT && DCI.isBeforeLegalizeOps() &&
22755 !DCI.isBeforeLegalize() &&
22756 !ISD::isBuildVectorOfConstantSDNodes(Cond.getNode())) {
22757 unsigned BitWidth = Cond.getValueType().getScalarType().getSizeInBits();
22759 // Don't optimize vector selects that map to mask-registers.
22763 // We can only handle the cases where VSELECT is directly legal on the
22764 // subtarget. We custom lower VSELECT nodes with constant conditions and
22765 // this makes it hard to see whether a dynamic VSELECT will correctly
22766 // lower, so we both check the operation's status and explicitly handle the
22767 // cases where a *dynamic* blend will fail even though a constant-condition
22768 // blend could be custom lowered.
22769 // FIXME: We should find a better way to handle this class of problems.
22770 // Potentially, we should combine constant-condition vselect nodes
22771 // pre-legalization into shuffles and not mark as many types as custom
22773 if (!TLI.isOperationLegalOrCustom(ISD::VSELECT, VT))
22775 // FIXME: We don't support i16-element blends currently. We could and
22776 // should support them by making *all* the bits in the condition be set
22777 // rather than just the high bit and using an i8-element blend.
22778 if (VT.getScalarType() == MVT::i16)
22780 // Dynamic blending was only available from SSE4.1 onward.
22781 if (VT.getSizeInBits() == 128 && !Subtarget->hasSSE41())
22783 // Byte blends are only available in AVX2
22784 if (VT.getSizeInBits() == 256 && VT.getScalarType() == MVT::i8 &&
22785 !Subtarget->hasAVX2())
22788 assert(BitWidth >= 8 && BitWidth <= 64 && "Invalid mask size");
22789 APInt DemandedMask = APInt::getHighBitsSet(BitWidth, 1);
22791 APInt KnownZero, KnownOne;
22792 TargetLowering::TargetLoweringOpt TLO(DAG, DCI.isBeforeLegalize(),
22793 DCI.isBeforeLegalizeOps());
22794 if (TLO.ShrinkDemandedConstant(Cond, DemandedMask) ||
22795 TLI.SimplifyDemandedBits(Cond, DemandedMask, KnownZero, KnownOne,
22797 // If we changed the computation somewhere in the DAG, this change
22798 // will affect all users of Cond.
22799 // Make sure it is fine and update all the nodes so that we do not
22800 // use the generic VSELECT anymore. Otherwise, we may perform
22801 // wrong optimizations as we messed up with the actual expectation
22802 // for the vector boolean values.
22803 if (Cond != TLO.Old) {
22804 // Check all uses of that condition operand to check whether it will be
22805 // consumed by non-BLEND instructions, which may depend on all bits are
22807 for (SDNode::use_iterator I = Cond->use_begin(), E = Cond->use_end();
22809 if (I->getOpcode() != ISD::VSELECT)
22810 // TODO: Add other opcodes eventually lowered into BLEND.
22813 // Update all the users of the condition, before committing the change,
22814 // so that the VSELECT optimizations that expect the correct vector
22815 // boolean value will not be triggered.
22816 for (SDNode::use_iterator I = Cond->use_begin(), E = Cond->use_end();
22818 DAG.ReplaceAllUsesOfValueWith(
22820 DAG.getNode(X86ISD::SHRUNKBLEND, SDLoc(*I), I->getValueType(0),
22821 Cond, I->getOperand(1), I->getOperand(2)));
22822 DCI.CommitTargetLoweringOpt(TLO);
22825 // At this point, only Cond is changed. Change the condition
22826 // just for N to keep the opportunity to optimize all other
22827 // users their own way.
22828 DAG.ReplaceAllUsesOfValueWith(
22830 DAG.getNode(X86ISD::SHRUNKBLEND, SDLoc(N), N->getValueType(0),
22831 TLO.New, N->getOperand(1), N->getOperand(2)));
22839 // Check whether a boolean test is testing a boolean value generated by
22840 // X86ISD::SETCC. If so, return the operand of that SETCC and proper condition
22843 // Simplify the following patterns:
22844 // (Op (CMP (SETCC Cond EFLAGS) 1) EQ) or
22845 // (Op (CMP (SETCC Cond EFLAGS) 0) NEQ)
22846 // to (Op EFLAGS Cond)
22848 // (Op (CMP (SETCC Cond EFLAGS) 0) EQ) or
22849 // (Op (CMP (SETCC Cond EFLAGS) 1) NEQ)
22850 // to (Op EFLAGS !Cond)
22852 // where Op could be BRCOND or CMOV.
22854 static SDValue checkBoolTestSetCCCombine(SDValue Cmp, X86::CondCode &CC) {
22855 // Quit if not CMP and SUB with its value result used.
22856 if (Cmp.getOpcode() != X86ISD::CMP &&
22857 (Cmp.getOpcode() != X86ISD::SUB || Cmp.getNode()->hasAnyUseOfValue(0)))
22860 // Quit if not used as a boolean value.
22861 if (CC != X86::COND_E && CC != X86::COND_NE)
22864 // Check CMP operands. One of them should be 0 or 1 and the other should be
22865 // an SetCC or extended from it.
22866 SDValue Op1 = Cmp.getOperand(0);
22867 SDValue Op2 = Cmp.getOperand(1);
22870 const ConstantSDNode* C = nullptr;
22871 bool needOppositeCond = (CC == X86::COND_E);
22872 bool checkAgainstTrue = false; // Is it a comparison against 1?
22874 if ((C = dyn_cast<ConstantSDNode>(Op1)))
22876 else if ((C = dyn_cast<ConstantSDNode>(Op2)))
22878 else // Quit if all operands are not constants.
22881 if (C->getZExtValue() == 1) {
22882 needOppositeCond = !needOppositeCond;
22883 checkAgainstTrue = true;
22884 } else if (C->getZExtValue() != 0)
22885 // Quit if the constant is neither 0 or 1.
22888 bool truncatedToBoolWithAnd = false;
22889 // Skip (zext $x), (trunc $x), or (and $x, 1) node.
22890 while (SetCC.getOpcode() == ISD::ZERO_EXTEND ||
22891 SetCC.getOpcode() == ISD::TRUNCATE ||
22892 SetCC.getOpcode() == ISD::AND) {
22893 if (SetCC.getOpcode() == ISD::AND) {
22895 ConstantSDNode *CS;
22896 if ((CS = dyn_cast<ConstantSDNode>(SetCC.getOperand(0))) &&
22897 CS->getZExtValue() == 1)
22899 if ((CS = dyn_cast<ConstantSDNode>(SetCC.getOperand(1))) &&
22900 CS->getZExtValue() == 1)
22904 SetCC = SetCC.getOperand(OpIdx);
22905 truncatedToBoolWithAnd = true;
22907 SetCC = SetCC.getOperand(0);
22910 switch (SetCC.getOpcode()) {
22911 case X86ISD::SETCC_CARRY:
22912 // Since SETCC_CARRY gives output based on R = CF ? ~0 : 0, it's unsafe to
22913 // simplify it if the result of SETCC_CARRY is not canonicalized to 0 or 1,
22914 // i.e. it's a comparison against true but the result of SETCC_CARRY is not
22915 // truncated to i1 using 'and'.
22916 if (checkAgainstTrue && !truncatedToBoolWithAnd)
22918 assert(X86::CondCode(SetCC.getConstantOperandVal(0)) == X86::COND_B &&
22919 "Invalid use of SETCC_CARRY!");
22921 case X86ISD::SETCC:
22922 // Set the condition code or opposite one if necessary.
22923 CC = X86::CondCode(SetCC.getConstantOperandVal(0));
22924 if (needOppositeCond)
22925 CC = X86::GetOppositeBranchCondition(CC);
22926 return SetCC.getOperand(1);
22927 case X86ISD::CMOV: {
22928 // Check whether false/true value has canonical one, i.e. 0 or 1.
22929 ConstantSDNode *FVal = dyn_cast<ConstantSDNode>(SetCC.getOperand(0));
22930 ConstantSDNode *TVal = dyn_cast<ConstantSDNode>(SetCC.getOperand(1));
22931 // Quit if true value is not a constant.
22934 // Quit if false value is not a constant.
22936 SDValue Op = SetCC.getOperand(0);
22937 // Skip 'zext' or 'trunc' node.
22938 if (Op.getOpcode() == ISD::ZERO_EXTEND ||
22939 Op.getOpcode() == ISD::TRUNCATE)
22940 Op = Op.getOperand(0);
22941 // A special case for rdrand/rdseed, where 0 is set if false cond is
22943 if ((Op.getOpcode() != X86ISD::RDRAND &&
22944 Op.getOpcode() != X86ISD::RDSEED) || Op.getResNo() != 0)
22947 // Quit if false value is not the constant 0 or 1.
22948 bool FValIsFalse = true;
22949 if (FVal && FVal->getZExtValue() != 0) {
22950 if (FVal->getZExtValue() != 1)
22952 // If FVal is 1, opposite cond is needed.
22953 needOppositeCond = !needOppositeCond;
22954 FValIsFalse = false;
22956 // Quit if TVal is not the constant opposite of FVal.
22957 if (FValIsFalse && TVal->getZExtValue() != 1)
22959 if (!FValIsFalse && TVal->getZExtValue() != 0)
22961 CC = X86::CondCode(SetCC.getConstantOperandVal(2));
22962 if (needOppositeCond)
22963 CC = X86::GetOppositeBranchCondition(CC);
22964 return SetCC.getOperand(3);
22971 /// Check whether Cond is an AND/OR of SETCCs off of the same EFLAGS.
22973 /// (X86or (X86setcc) (X86setcc))
22974 /// (X86cmp (and (X86setcc) (X86setcc)), 0)
22975 static bool checkBoolTestAndOrSetCCCombine(SDValue Cond, X86::CondCode &CC0,
22976 X86::CondCode &CC1, SDValue &Flags,
22978 if (Cond->getOpcode() == X86ISD::CMP) {
22979 ConstantSDNode *CondOp1C = dyn_cast<ConstantSDNode>(Cond->getOperand(1));
22980 if (!CondOp1C || !CondOp1C->isNullValue())
22983 Cond = Cond->getOperand(0);
22988 SDValue SetCC0, SetCC1;
22989 switch (Cond->getOpcode()) {
22990 default: return false;
22997 SetCC0 = Cond->getOperand(0);
22998 SetCC1 = Cond->getOperand(1);
23002 // Make sure we have SETCC nodes, using the same flags value.
23003 if (SetCC0.getOpcode() != X86ISD::SETCC ||
23004 SetCC1.getOpcode() != X86ISD::SETCC ||
23005 SetCC0->getOperand(1) != SetCC1->getOperand(1))
23008 CC0 = (X86::CondCode)SetCC0->getConstantOperandVal(0);
23009 CC1 = (X86::CondCode)SetCC1->getConstantOperandVal(0);
23010 Flags = SetCC0->getOperand(1);
23014 /// Optimize X86ISD::CMOV [LHS, RHS, CONDCODE (e.g. X86::COND_NE), CONDVAL]
23015 static SDValue PerformCMOVCombine(SDNode *N, SelectionDAG &DAG,
23016 TargetLowering::DAGCombinerInfo &DCI,
23017 const X86Subtarget *Subtarget) {
23020 // If the flag operand isn't dead, don't touch this CMOV.
23021 if (N->getNumValues() == 2 && !SDValue(N, 1).use_empty())
23024 SDValue FalseOp = N->getOperand(0);
23025 SDValue TrueOp = N->getOperand(1);
23026 X86::CondCode CC = (X86::CondCode)N->getConstantOperandVal(2);
23027 SDValue Cond = N->getOperand(3);
23029 if (CC == X86::COND_E || CC == X86::COND_NE) {
23030 switch (Cond.getOpcode()) {
23034 // If operand of BSR / BSF are proven never zero, then ZF cannot be set.
23035 if (DAG.isKnownNeverZero(Cond.getOperand(0)))
23036 return (CC == X86::COND_E) ? FalseOp : TrueOp;
23042 Flags = checkBoolTestSetCCCombine(Cond, CC);
23043 if (Flags.getNode() &&
23044 // Extra check as FCMOV only supports a subset of X86 cond.
23045 (FalseOp.getValueType() != MVT::f80 || hasFPCMov(CC))) {
23046 SDValue Ops[] = { FalseOp, TrueOp,
23047 DAG.getConstant(CC, DL, MVT::i8), Flags };
23048 return DAG.getNode(X86ISD::CMOV, DL, N->getVTList(), Ops);
23051 // If this is a select between two integer constants, try to do some
23052 // optimizations. Note that the operands are ordered the opposite of SELECT
23054 if (ConstantSDNode *TrueC = dyn_cast<ConstantSDNode>(TrueOp)) {
23055 if (ConstantSDNode *FalseC = dyn_cast<ConstantSDNode>(FalseOp)) {
23056 // Canonicalize the TrueC/FalseC values so that TrueC (the true value) is
23057 // larger than FalseC (the false value).
23058 if (TrueC->getAPIntValue().ult(FalseC->getAPIntValue())) {
23059 CC = X86::GetOppositeBranchCondition(CC);
23060 std::swap(TrueC, FalseC);
23061 std::swap(TrueOp, FalseOp);
23064 // Optimize C ? 8 : 0 -> zext(setcc(C)) << 3. Likewise for any pow2/0.
23065 // This is efficient for any integer data type (including i8/i16) and
23067 if (FalseC->getAPIntValue() == 0 && TrueC->getAPIntValue().isPowerOf2()) {
23068 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
23069 DAG.getConstant(CC, DL, MVT::i8), Cond);
23071 // Zero extend the condition if needed.
23072 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, TrueC->getValueType(0), Cond);
23074 unsigned ShAmt = TrueC->getAPIntValue().logBase2();
23075 Cond = DAG.getNode(ISD::SHL, DL, Cond.getValueType(), Cond,
23076 DAG.getConstant(ShAmt, DL, MVT::i8));
23077 if (N->getNumValues() == 2) // Dead flag value?
23078 return DCI.CombineTo(N, Cond, SDValue());
23082 // Optimize Cond ? cst+1 : cst -> zext(setcc(C)+cst. This is efficient
23083 // for any integer data type, including i8/i16.
23084 if (FalseC->getAPIntValue()+1 == TrueC->getAPIntValue()) {
23085 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
23086 DAG.getConstant(CC, DL, MVT::i8), Cond);
23088 // Zero extend the condition if needed.
23089 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL,
23090 FalseC->getValueType(0), Cond);
23091 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
23092 SDValue(FalseC, 0));
23094 if (N->getNumValues() == 2) // Dead flag value?
23095 return DCI.CombineTo(N, Cond, SDValue());
23099 // Optimize cases that will turn into an LEA instruction. This requires
23100 // an i32 or i64 and an efficient multiplier (1, 2, 3, 4, 5, 8, 9).
23101 if (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i64) {
23102 uint64_t Diff = TrueC->getZExtValue()-FalseC->getZExtValue();
23103 if (N->getValueType(0) == MVT::i32) Diff = (unsigned)Diff;
23105 bool isFastMultiplier = false;
23107 switch ((unsigned char)Diff) {
23109 case 1: // result = add base, cond
23110 case 2: // result = lea base( , cond*2)
23111 case 3: // result = lea base(cond, cond*2)
23112 case 4: // result = lea base( , cond*4)
23113 case 5: // result = lea base(cond, cond*4)
23114 case 8: // result = lea base( , cond*8)
23115 case 9: // result = lea base(cond, cond*8)
23116 isFastMultiplier = true;
23121 if (isFastMultiplier) {
23122 APInt Diff = TrueC->getAPIntValue()-FalseC->getAPIntValue();
23123 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
23124 DAG.getConstant(CC, DL, MVT::i8), Cond);
23125 // Zero extend the condition if needed.
23126 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, FalseC->getValueType(0),
23128 // Scale the condition by the difference.
23130 Cond = DAG.getNode(ISD::MUL, DL, Cond.getValueType(), Cond,
23131 DAG.getConstant(Diff, DL, Cond.getValueType()));
23133 // Add the base if non-zero.
23134 if (FalseC->getAPIntValue() != 0)
23135 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
23136 SDValue(FalseC, 0));
23137 if (N->getNumValues() == 2) // Dead flag value?
23138 return DCI.CombineTo(N, Cond, SDValue());
23145 // Handle these cases:
23146 // (select (x != c), e, c) -> select (x != c), e, x),
23147 // (select (x == c), c, e) -> select (x == c), x, e)
23148 // where the c is an integer constant, and the "select" is the combination
23149 // of CMOV and CMP.
23151 // The rationale for this change is that the conditional-move from a constant
23152 // needs two instructions, however, conditional-move from a register needs
23153 // only one instruction.
23155 // CAVEAT: By replacing a constant with a symbolic value, it may obscure
23156 // some instruction-combining opportunities. This opt needs to be
23157 // postponed as late as possible.
23159 if (!DCI.isBeforeLegalize() && !DCI.isBeforeLegalizeOps()) {
23160 // the DCI.xxxx conditions are provided to postpone the optimization as
23161 // late as possible.
23163 ConstantSDNode *CmpAgainst = nullptr;
23164 if ((Cond.getOpcode() == X86ISD::CMP || Cond.getOpcode() == X86ISD::SUB) &&
23165 (CmpAgainst = dyn_cast<ConstantSDNode>(Cond.getOperand(1))) &&
23166 !isa<ConstantSDNode>(Cond.getOperand(0))) {
23168 if (CC == X86::COND_NE &&
23169 CmpAgainst == dyn_cast<ConstantSDNode>(FalseOp)) {
23170 CC = X86::GetOppositeBranchCondition(CC);
23171 std::swap(TrueOp, FalseOp);
23174 if (CC == X86::COND_E &&
23175 CmpAgainst == dyn_cast<ConstantSDNode>(TrueOp)) {
23176 SDValue Ops[] = { FalseOp, Cond.getOperand(0),
23177 DAG.getConstant(CC, DL, MVT::i8), Cond };
23178 return DAG.getNode(X86ISD::CMOV, DL, N->getVTList (), Ops);
23183 // Fold and/or of setcc's to double CMOV:
23184 // (CMOV F, T, ((cc1 | cc2) != 0)) -> (CMOV (CMOV F, T, cc1), T, cc2)
23185 // (CMOV F, T, ((cc1 & cc2) != 0)) -> (CMOV (CMOV T, F, !cc1), F, !cc2)
23187 // This combine lets us generate:
23188 // cmovcc1 (jcc1 if we don't have CMOV)
23194 // cmovne (jne if we don't have CMOV)
23195 // When we can't use the CMOV instruction, it might increase branch
23197 // When we can use CMOV, or when there is no mispredict, this improves
23198 // throughput and reduces register pressure.
23200 if (CC == X86::COND_NE) {
23202 X86::CondCode CC0, CC1;
23204 if (checkBoolTestAndOrSetCCCombine(Cond, CC0, CC1, Flags, isAndSetCC)) {
23206 std::swap(FalseOp, TrueOp);
23207 CC0 = X86::GetOppositeBranchCondition(CC0);
23208 CC1 = X86::GetOppositeBranchCondition(CC1);
23211 SDValue LOps[] = {FalseOp, TrueOp, DAG.getConstant(CC0, DL, MVT::i8),
23213 SDValue LCMOV = DAG.getNode(X86ISD::CMOV, DL, N->getVTList(), LOps);
23214 SDValue Ops[] = {LCMOV, TrueOp, DAG.getConstant(CC1, DL, MVT::i8), Flags};
23215 SDValue CMOV = DAG.getNode(X86ISD::CMOV, DL, N->getVTList(), Ops);
23216 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), SDValue(CMOV.getNode(), 1));
23224 static SDValue PerformINTRINSIC_WO_CHAINCombine(SDNode *N, SelectionDAG &DAG,
23225 const X86Subtarget *Subtarget) {
23226 unsigned IntNo = cast<ConstantSDNode>(N->getOperand(0))->getZExtValue();
23228 default: return SDValue();
23229 // SSE/AVX/AVX2 blend intrinsics.
23230 case Intrinsic::x86_avx2_pblendvb:
23231 // Don't try to simplify this intrinsic if we don't have AVX2.
23232 if (!Subtarget->hasAVX2())
23235 case Intrinsic::x86_avx_blendv_pd_256:
23236 case Intrinsic::x86_avx_blendv_ps_256:
23237 // Don't try to simplify this intrinsic if we don't have AVX.
23238 if (!Subtarget->hasAVX())
23241 case Intrinsic::x86_sse41_blendvps:
23242 case Intrinsic::x86_sse41_blendvpd:
23243 case Intrinsic::x86_sse41_pblendvb: {
23244 SDValue Op0 = N->getOperand(1);
23245 SDValue Op1 = N->getOperand(2);
23246 SDValue Mask = N->getOperand(3);
23248 // Don't try to simplify this intrinsic if we don't have SSE4.1.
23249 if (!Subtarget->hasSSE41())
23252 // fold (blend A, A, Mask) -> A
23255 // fold (blend A, B, allZeros) -> A
23256 if (ISD::isBuildVectorAllZeros(Mask.getNode()))
23258 // fold (blend A, B, allOnes) -> B
23259 if (ISD::isBuildVectorAllOnes(Mask.getNode()))
23262 // Simplify the case where the mask is a constant i32 value.
23263 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Mask)) {
23264 if (C->isNullValue())
23266 if (C->isAllOnesValue())
23273 // Packed SSE2/AVX2 arithmetic shift immediate intrinsics.
23274 case Intrinsic::x86_sse2_psrai_w:
23275 case Intrinsic::x86_sse2_psrai_d:
23276 case Intrinsic::x86_avx2_psrai_w:
23277 case Intrinsic::x86_avx2_psrai_d:
23278 case Intrinsic::x86_sse2_psra_w:
23279 case Intrinsic::x86_sse2_psra_d:
23280 case Intrinsic::x86_avx2_psra_w:
23281 case Intrinsic::x86_avx2_psra_d: {
23282 SDValue Op0 = N->getOperand(1);
23283 SDValue Op1 = N->getOperand(2);
23284 EVT VT = Op0.getValueType();
23285 assert(VT.isVector() && "Expected a vector type!");
23287 if (isa<BuildVectorSDNode>(Op1))
23288 Op1 = Op1.getOperand(0);
23290 if (!isa<ConstantSDNode>(Op1))
23293 EVT SVT = VT.getVectorElementType();
23294 unsigned SVTBits = SVT.getSizeInBits();
23296 ConstantSDNode *CND = cast<ConstantSDNode>(Op1);
23297 const APInt &C = APInt(SVTBits, CND->getAPIntValue().getZExtValue());
23298 uint64_t ShAmt = C.getZExtValue();
23300 // Don't try to convert this shift into a ISD::SRA if the shift
23301 // count is bigger than or equal to the element size.
23302 if (ShAmt >= SVTBits)
23305 // Trivial case: if the shift count is zero, then fold this
23306 // into the first operand.
23310 // Replace this packed shift intrinsic with a target independent
23313 SDValue Splat = DAG.getConstant(C, DL, VT);
23314 return DAG.getNode(ISD::SRA, DL, VT, Op0, Splat);
23319 /// PerformMulCombine - Optimize a single multiply with constant into two
23320 /// in order to implement it with two cheaper instructions, e.g.
23321 /// LEA + SHL, LEA + LEA.
23322 static SDValue PerformMulCombine(SDNode *N, SelectionDAG &DAG,
23323 TargetLowering::DAGCombinerInfo &DCI) {
23324 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
23327 EVT VT = N->getValueType(0);
23328 if (VT != MVT::i64 && VT != MVT::i32)
23331 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1));
23334 uint64_t MulAmt = C->getZExtValue();
23335 if (isPowerOf2_64(MulAmt) || MulAmt == 3 || MulAmt == 5 || MulAmt == 9)
23338 uint64_t MulAmt1 = 0;
23339 uint64_t MulAmt2 = 0;
23340 if ((MulAmt % 9) == 0) {
23342 MulAmt2 = MulAmt / 9;
23343 } else if ((MulAmt % 5) == 0) {
23345 MulAmt2 = MulAmt / 5;
23346 } else if ((MulAmt % 3) == 0) {
23348 MulAmt2 = MulAmt / 3;
23351 (isPowerOf2_64(MulAmt2) || MulAmt2 == 3 || MulAmt2 == 5 || MulAmt2 == 9)){
23354 if (isPowerOf2_64(MulAmt2) &&
23355 !(N->hasOneUse() && N->use_begin()->getOpcode() == ISD::ADD))
23356 // If second multiplifer is pow2, issue it first. We want the multiply by
23357 // 3, 5, or 9 to be folded into the addressing mode unless the lone use
23359 std::swap(MulAmt1, MulAmt2);
23362 if (isPowerOf2_64(MulAmt1))
23363 NewMul = DAG.getNode(ISD::SHL, DL, VT, N->getOperand(0),
23364 DAG.getConstant(Log2_64(MulAmt1), DL, MVT::i8));
23366 NewMul = DAG.getNode(X86ISD::MUL_IMM, DL, VT, N->getOperand(0),
23367 DAG.getConstant(MulAmt1, DL, VT));
23369 if (isPowerOf2_64(MulAmt2))
23370 NewMul = DAG.getNode(ISD::SHL, DL, VT, NewMul,
23371 DAG.getConstant(Log2_64(MulAmt2), DL, MVT::i8));
23373 NewMul = DAG.getNode(X86ISD::MUL_IMM, DL, VT, NewMul,
23374 DAG.getConstant(MulAmt2, DL, VT));
23376 // Do not add new nodes to DAG combiner worklist.
23377 DCI.CombineTo(N, NewMul, false);
23382 static SDValue PerformSHLCombine(SDNode *N, SelectionDAG &DAG) {
23383 SDValue N0 = N->getOperand(0);
23384 SDValue N1 = N->getOperand(1);
23385 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
23386 EVT VT = N0.getValueType();
23388 // fold (shl (and (setcc_c), c1), c2) -> (and setcc_c, (c1 << c2))
23389 // since the result of setcc_c is all zero's or all ones.
23390 if (VT.isInteger() && !VT.isVector() &&
23391 N1C && N0.getOpcode() == ISD::AND &&
23392 N0.getOperand(1).getOpcode() == ISD::Constant) {
23393 SDValue N00 = N0.getOperand(0);
23394 if (N00.getOpcode() == X86ISD::SETCC_CARRY ||
23395 ((N00.getOpcode() == ISD::ANY_EXTEND ||
23396 N00.getOpcode() == ISD::ZERO_EXTEND) &&
23397 N00.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY)) {
23398 APInt Mask = cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
23399 APInt ShAmt = N1C->getAPIntValue();
23400 Mask = Mask.shl(ShAmt);
23403 return DAG.getNode(ISD::AND, DL, VT,
23404 N00, DAG.getConstant(Mask, DL, VT));
23409 // Hardware support for vector shifts is sparse which makes us scalarize the
23410 // vector operations in many cases. Also, on sandybridge ADD is faster than
23412 // (shl V, 1) -> add V,V
23413 if (auto *N1BV = dyn_cast<BuildVectorSDNode>(N1))
23414 if (auto *N1SplatC = N1BV->getConstantSplatNode()) {
23415 assert(N0.getValueType().isVector() && "Invalid vector shift type");
23416 // We shift all of the values by one. In many cases we do not have
23417 // hardware support for this operation. This is better expressed as an ADD
23419 if (N1SplatC->getAPIntValue() == 1)
23420 return DAG.getNode(ISD::ADD, SDLoc(N), VT, N0, N0);
23426 /// \brief Returns a vector of 0s if the node in input is a vector logical
23427 /// shift by a constant amount which is known to be bigger than or equal
23428 /// to the vector element size in bits.
23429 static SDValue performShiftToAllZeros(SDNode *N, SelectionDAG &DAG,
23430 const X86Subtarget *Subtarget) {
23431 EVT VT = N->getValueType(0);
23433 if (VT != MVT::v2i64 && VT != MVT::v4i32 && VT != MVT::v8i16 &&
23434 (!Subtarget->hasInt256() ||
23435 (VT != MVT::v4i64 && VT != MVT::v8i32 && VT != MVT::v16i16)))
23438 SDValue Amt = N->getOperand(1);
23440 if (auto *AmtBV = dyn_cast<BuildVectorSDNode>(Amt))
23441 if (auto *AmtSplat = AmtBV->getConstantSplatNode()) {
23442 APInt ShiftAmt = AmtSplat->getAPIntValue();
23443 unsigned MaxAmount = VT.getVectorElementType().getSizeInBits();
23445 // SSE2/AVX2 logical shifts always return a vector of 0s
23446 // if the shift amount is bigger than or equal to
23447 // the element size. The constant shift amount will be
23448 // encoded as a 8-bit immediate.
23449 if (ShiftAmt.trunc(8).uge(MaxAmount))
23450 return getZeroVector(VT, Subtarget, DAG, DL);
23456 /// PerformShiftCombine - Combine shifts.
23457 static SDValue PerformShiftCombine(SDNode* N, SelectionDAG &DAG,
23458 TargetLowering::DAGCombinerInfo &DCI,
23459 const X86Subtarget *Subtarget) {
23460 if (N->getOpcode() == ISD::SHL)
23461 if (SDValue V = PerformSHLCombine(N, DAG))
23464 // Try to fold this logical shift into a zero vector.
23465 if (N->getOpcode() != ISD::SRA)
23466 if (SDValue V = performShiftToAllZeros(N, DAG, Subtarget))
23472 // CMPEQCombine - Recognize the distinctive (AND (setcc ...) (setcc ..))
23473 // where both setccs reference the same FP CMP, and rewrite for CMPEQSS
23474 // and friends. Likewise for OR -> CMPNEQSS.
23475 static SDValue CMPEQCombine(SDNode *N, SelectionDAG &DAG,
23476 TargetLowering::DAGCombinerInfo &DCI,
23477 const X86Subtarget *Subtarget) {
23480 // SSE1 supports CMP{eq|ne}SS, and SSE2 added CMP{eq|ne}SD, but
23481 // we're requiring SSE2 for both.
23482 if (Subtarget->hasSSE2() && isAndOrOfSetCCs(SDValue(N, 0U), opcode)) {
23483 SDValue N0 = N->getOperand(0);
23484 SDValue N1 = N->getOperand(1);
23485 SDValue CMP0 = N0->getOperand(1);
23486 SDValue CMP1 = N1->getOperand(1);
23489 // The SETCCs should both refer to the same CMP.
23490 if (CMP0.getOpcode() != X86ISD::CMP || CMP0 != CMP1)
23493 SDValue CMP00 = CMP0->getOperand(0);
23494 SDValue CMP01 = CMP0->getOperand(1);
23495 EVT VT = CMP00.getValueType();
23497 if (VT == MVT::f32 || VT == MVT::f64) {
23498 bool ExpectingFlags = false;
23499 // Check for any users that want flags:
23500 for (SDNode::use_iterator UI = N->use_begin(), UE = N->use_end();
23501 !ExpectingFlags && UI != UE; ++UI)
23502 switch (UI->getOpcode()) {
23507 ExpectingFlags = true;
23509 case ISD::CopyToReg:
23510 case ISD::SIGN_EXTEND:
23511 case ISD::ZERO_EXTEND:
23512 case ISD::ANY_EXTEND:
23516 if (!ExpectingFlags) {
23517 enum X86::CondCode cc0 = (enum X86::CondCode)N0.getConstantOperandVal(0);
23518 enum X86::CondCode cc1 = (enum X86::CondCode)N1.getConstantOperandVal(0);
23520 if (cc1 == X86::COND_E || cc1 == X86::COND_NE) {
23521 X86::CondCode tmp = cc0;
23526 if ((cc0 == X86::COND_E && cc1 == X86::COND_NP) ||
23527 (cc0 == X86::COND_NE && cc1 == X86::COND_P)) {
23528 // FIXME: need symbolic constants for these magic numbers.
23529 // See X86ATTInstPrinter.cpp:printSSECC().
23530 unsigned x86cc = (cc0 == X86::COND_E) ? 0 : 4;
23531 if (Subtarget->hasAVX512()) {
23532 SDValue FSetCC = DAG.getNode(X86ISD::FSETCC, DL, MVT::i1, CMP00,
23534 DAG.getConstant(x86cc, DL, MVT::i8));
23535 if (N->getValueType(0) != MVT::i1)
23536 return DAG.getNode(ISD::ZERO_EXTEND, DL, N->getValueType(0),
23540 SDValue OnesOrZeroesF = DAG.getNode(X86ISD::FSETCC, DL,
23541 CMP00.getValueType(), CMP00, CMP01,
23542 DAG.getConstant(x86cc, DL,
23545 bool is64BitFP = (CMP00.getValueType() == MVT::f64);
23546 MVT IntVT = is64BitFP ? MVT::i64 : MVT::i32;
23548 if (is64BitFP && !Subtarget->is64Bit()) {
23549 // On a 32-bit target, we cannot bitcast the 64-bit float to a
23550 // 64-bit integer, since that's not a legal type. Since
23551 // OnesOrZeroesF is all ones of all zeroes, we don't need all the
23552 // bits, but can do this little dance to extract the lowest 32 bits
23553 // and work with those going forward.
23554 SDValue Vector64 = DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, MVT::v2f64,
23556 SDValue Vector32 = DAG.getBitcast(MVT::v4f32, Vector64);
23557 OnesOrZeroesF = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::f32,
23558 Vector32, DAG.getIntPtrConstant(0, DL));
23562 SDValue OnesOrZeroesI = DAG.getBitcast(IntVT, OnesOrZeroesF);
23563 SDValue ANDed = DAG.getNode(ISD::AND, DL, IntVT, OnesOrZeroesI,
23564 DAG.getConstant(1, DL, IntVT));
23565 SDValue OneBitOfTruth = DAG.getNode(ISD::TRUNCATE, DL, MVT::i8,
23567 return OneBitOfTruth;
23575 /// CanFoldXORWithAllOnes - Test whether the XOR operand is a AllOnes vector
23576 /// so it can be folded inside ANDNP.
23577 static bool CanFoldXORWithAllOnes(const SDNode *N) {
23578 EVT VT = N->getValueType(0);
23580 // Match direct AllOnes for 128 and 256-bit vectors
23581 if (ISD::isBuildVectorAllOnes(N))
23584 // Look through a bit convert.
23585 if (N->getOpcode() == ISD::BITCAST)
23586 N = N->getOperand(0).getNode();
23588 // Sometimes the operand may come from a insert_subvector building a 256-bit
23590 if (VT.is256BitVector() &&
23591 N->getOpcode() == ISD::INSERT_SUBVECTOR) {
23592 SDValue V1 = N->getOperand(0);
23593 SDValue V2 = N->getOperand(1);
23595 if (V1.getOpcode() == ISD::INSERT_SUBVECTOR &&
23596 V1.getOperand(0).getOpcode() == ISD::UNDEF &&
23597 ISD::isBuildVectorAllOnes(V1.getOperand(1).getNode()) &&
23598 ISD::isBuildVectorAllOnes(V2.getNode()))
23605 // On AVX/AVX2 the type v8i1 is legalized to v8i16, which is an XMM sized
23606 // register. In most cases we actually compare or select YMM-sized registers
23607 // and mixing the two types creates horrible code. This method optimizes
23608 // some of the transition sequences.
23609 static SDValue WidenMaskArithmetic(SDNode *N, SelectionDAG &DAG,
23610 TargetLowering::DAGCombinerInfo &DCI,
23611 const X86Subtarget *Subtarget) {
23612 EVT VT = N->getValueType(0);
23613 if (!VT.is256BitVector())
23616 assert((N->getOpcode() == ISD::ANY_EXTEND ||
23617 N->getOpcode() == ISD::ZERO_EXTEND ||
23618 N->getOpcode() == ISD::SIGN_EXTEND) && "Invalid Node");
23620 SDValue Narrow = N->getOperand(0);
23621 EVT NarrowVT = Narrow->getValueType(0);
23622 if (!NarrowVT.is128BitVector())
23625 if (Narrow->getOpcode() != ISD::XOR &&
23626 Narrow->getOpcode() != ISD::AND &&
23627 Narrow->getOpcode() != ISD::OR)
23630 SDValue N0 = Narrow->getOperand(0);
23631 SDValue N1 = Narrow->getOperand(1);
23634 // The Left side has to be a trunc.
23635 if (N0.getOpcode() != ISD::TRUNCATE)
23638 // The type of the truncated inputs.
23639 EVT WideVT = N0->getOperand(0)->getValueType(0);
23643 // The right side has to be a 'trunc' or a constant vector.
23644 bool RHSTrunc = N1.getOpcode() == ISD::TRUNCATE;
23645 ConstantSDNode *RHSConstSplat = nullptr;
23646 if (auto *RHSBV = dyn_cast<BuildVectorSDNode>(N1))
23647 RHSConstSplat = RHSBV->getConstantSplatNode();
23648 if (!RHSTrunc && !RHSConstSplat)
23651 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
23653 if (!TLI.isOperationLegalOrPromote(Narrow->getOpcode(), WideVT))
23656 // Set N0 and N1 to hold the inputs to the new wide operation.
23657 N0 = N0->getOperand(0);
23658 if (RHSConstSplat) {
23659 N1 = DAG.getNode(ISD::ZERO_EXTEND, DL, WideVT.getScalarType(),
23660 SDValue(RHSConstSplat, 0));
23661 SmallVector<SDValue, 8> C(WideVT.getVectorNumElements(), N1);
23662 N1 = DAG.getNode(ISD::BUILD_VECTOR, DL, WideVT, C);
23663 } else if (RHSTrunc) {
23664 N1 = N1->getOperand(0);
23667 // Generate the wide operation.
23668 SDValue Op = DAG.getNode(Narrow->getOpcode(), DL, WideVT, N0, N1);
23669 unsigned Opcode = N->getOpcode();
23671 case ISD::ANY_EXTEND:
23673 case ISD::ZERO_EXTEND: {
23674 unsigned InBits = NarrowVT.getScalarType().getSizeInBits();
23675 APInt Mask = APInt::getAllOnesValue(InBits);
23676 Mask = Mask.zext(VT.getScalarType().getSizeInBits());
23677 return DAG.getNode(ISD::AND, DL, VT,
23678 Op, DAG.getConstant(Mask, DL, VT));
23680 case ISD::SIGN_EXTEND:
23681 return DAG.getNode(ISD::SIGN_EXTEND_INREG, DL, VT,
23682 Op, DAG.getValueType(NarrowVT));
23684 llvm_unreachable("Unexpected opcode");
23688 static SDValue VectorZextCombine(SDNode *N, SelectionDAG &DAG,
23689 TargetLowering::DAGCombinerInfo &DCI,
23690 const X86Subtarget *Subtarget) {
23691 SDValue N0 = N->getOperand(0);
23692 SDValue N1 = N->getOperand(1);
23695 // A vector zext_in_reg may be represented as a shuffle,
23696 // feeding into a bitcast (this represents anyext) feeding into
23697 // an and with a mask.
23698 // We'd like to try to combine that into a shuffle with zero
23699 // plus a bitcast, removing the and.
23700 if (N0.getOpcode() != ISD::BITCAST ||
23701 N0.getOperand(0).getOpcode() != ISD::VECTOR_SHUFFLE)
23704 // The other side of the AND should be a splat of 2^C, where C
23705 // is the number of bits in the source type.
23706 if (N1.getOpcode() == ISD::BITCAST)
23707 N1 = N1.getOperand(0);
23708 if (N1.getOpcode() != ISD::BUILD_VECTOR)
23710 BuildVectorSDNode *Vector = cast<BuildVectorSDNode>(N1);
23712 ShuffleVectorSDNode *Shuffle = cast<ShuffleVectorSDNode>(N0.getOperand(0));
23713 EVT SrcType = Shuffle->getValueType(0);
23715 // We expect a single-source shuffle
23716 if (Shuffle->getOperand(1)->getOpcode() != ISD::UNDEF)
23719 unsigned SrcSize = SrcType.getScalarSizeInBits();
23721 APInt SplatValue, SplatUndef;
23722 unsigned SplatBitSize;
23724 if (!Vector->isConstantSplat(SplatValue, SplatUndef,
23725 SplatBitSize, HasAnyUndefs))
23728 unsigned ResSize = N1.getValueType().getScalarSizeInBits();
23729 // Make sure the splat matches the mask we expect
23730 if (SplatBitSize > ResSize ||
23731 (SplatValue + 1).exactLogBase2() != (int)SrcSize)
23734 // Make sure the input and output size make sense
23735 if (SrcSize >= ResSize || ResSize % SrcSize)
23738 // We expect a shuffle of the form <0, u, u, u, 1, u, u, u...>
23739 // The number of u's between each two values depends on the ratio between
23740 // the source and dest type.
23741 unsigned ZextRatio = ResSize / SrcSize;
23742 bool IsZext = true;
23743 for (unsigned i = 0; i < SrcType.getVectorNumElements(); ++i) {
23744 if (i % ZextRatio) {
23745 if (Shuffle->getMaskElt(i) > 0) {
23751 if (Shuffle->getMaskElt(i) != (int)(i / ZextRatio)) {
23752 // Expected element number
23762 // Ok, perform the transformation - replace the shuffle with
23763 // a shuffle of the form <0, k, k, k, 1, k, k, k> with zero
23764 // (instead of undef) where the k elements come from the zero vector.
23765 SmallVector<int, 8> Mask;
23766 unsigned NumElems = SrcType.getVectorNumElements();
23767 for (unsigned i = 0; i < NumElems; ++i)
23769 Mask.push_back(NumElems);
23771 Mask.push_back(i / ZextRatio);
23773 SDValue NewShuffle = DAG.getVectorShuffle(Shuffle->getValueType(0), DL,
23774 Shuffle->getOperand(0), DAG.getConstant(0, DL, SrcType), Mask);
23775 return DAG.getBitcast(N0.getValueType(), NewShuffle);
23778 static SDValue PerformAndCombine(SDNode *N, SelectionDAG &DAG,
23779 TargetLowering::DAGCombinerInfo &DCI,
23780 const X86Subtarget *Subtarget) {
23781 if (DCI.isBeforeLegalizeOps())
23784 if (SDValue Zext = VectorZextCombine(N, DAG, DCI, Subtarget))
23787 if (SDValue R = CMPEQCombine(N, DAG, DCI, Subtarget))
23790 EVT VT = N->getValueType(0);
23791 SDValue N0 = N->getOperand(0);
23792 SDValue N1 = N->getOperand(1);
23795 // Create BEXTR instructions
23796 // BEXTR is ((X >> imm) & (2**size-1))
23797 if (VT == MVT::i32 || VT == MVT::i64) {
23798 // Check for BEXTR.
23799 if ((Subtarget->hasBMI() || Subtarget->hasTBM()) &&
23800 (N0.getOpcode() == ISD::SRA || N0.getOpcode() == ISD::SRL)) {
23801 ConstantSDNode *MaskNode = dyn_cast<ConstantSDNode>(N1);
23802 ConstantSDNode *ShiftNode = dyn_cast<ConstantSDNode>(N0.getOperand(1));
23803 if (MaskNode && ShiftNode) {
23804 uint64_t Mask = MaskNode->getZExtValue();
23805 uint64_t Shift = ShiftNode->getZExtValue();
23806 if (isMask_64(Mask)) {
23807 uint64_t MaskSize = countPopulation(Mask);
23808 if (Shift + MaskSize <= VT.getSizeInBits())
23809 return DAG.getNode(X86ISD::BEXTR, DL, VT, N0.getOperand(0),
23810 DAG.getConstant(Shift | (MaskSize << 8), DL,
23819 // Want to form ANDNP nodes:
23820 // 1) In the hopes of then easily combining them with OR and AND nodes
23821 // to form PBLEND/PSIGN.
23822 // 2) To match ANDN packed intrinsics
23823 if (VT != MVT::v2i64 && VT != MVT::v4i64)
23826 // Check LHS for vnot
23827 if (N0.getOpcode() == ISD::XOR &&
23828 //ISD::isBuildVectorAllOnes(N0.getOperand(1).getNode()))
23829 CanFoldXORWithAllOnes(N0.getOperand(1).getNode()))
23830 return DAG.getNode(X86ISD::ANDNP, DL, VT, N0.getOperand(0), N1);
23832 // Check RHS for vnot
23833 if (N1.getOpcode() == ISD::XOR &&
23834 //ISD::isBuildVectorAllOnes(N1.getOperand(1).getNode()))
23835 CanFoldXORWithAllOnes(N1.getOperand(1).getNode()))
23836 return DAG.getNode(X86ISD::ANDNP, DL, VT, N1.getOperand(0), N0);
23841 static SDValue PerformOrCombine(SDNode *N, SelectionDAG &DAG,
23842 TargetLowering::DAGCombinerInfo &DCI,
23843 const X86Subtarget *Subtarget) {
23844 if (DCI.isBeforeLegalizeOps())
23847 if (SDValue R = CMPEQCombine(N, DAG, DCI, Subtarget))
23850 SDValue N0 = N->getOperand(0);
23851 SDValue N1 = N->getOperand(1);
23852 EVT VT = N->getValueType(0);
23854 // look for psign/blend
23855 if (VT == MVT::v2i64 || VT == MVT::v4i64) {
23856 if (!Subtarget->hasSSSE3() ||
23857 (VT == MVT::v4i64 && !Subtarget->hasInt256()))
23860 // Canonicalize pandn to RHS
23861 if (N0.getOpcode() == X86ISD::ANDNP)
23863 // or (and (m, y), (pandn m, x))
23864 if (N0.getOpcode() == ISD::AND && N1.getOpcode() == X86ISD::ANDNP) {
23865 SDValue Mask = N1.getOperand(0);
23866 SDValue X = N1.getOperand(1);
23868 if (N0.getOperand(0) == Mask)
23869 Y = N0.getOperand(1);
23870 if (N0.getOperand(1) == Mask)
23871 Y = N0.getOperand(0);
23873 // Check to see if the mask appeared in both the AND and ANDNP and
23877 // Validate that X, Y, and Mask are BIT_CONVERTS, and see through them.
23878 // Look through mask bitcast.
23879 if (Mask.getOpcode() == ISD::BITCAST)
23880 Mask = Mask.getOperand(0);
23881 if (X.getOpcode() == ISD::BITCAST)
23882 X = X.getOperand(0);
23883 if (Y.getOpcode() == ISD::BITCAST)
23884 Y = Y.getOperand(0);
23886 EVT MaskVT = Mask.getValueType();
23888 // Validate that the Mask operand is a vector sra node.
23889 // FIXME: what to do for bytes, since there is a psignb/pblendvb, but
23890 // there is no psrai.b
23891 unsigned EltBits = MaskVT.getVectorElementType().getSizeInBits();
23892 unsigned SraAmt = ~0;
23893 if (Mask.getOpcode() == ISD::SRA) {
23894 if (auto *AmtBV = dyn_cast<BuildVectorSDNode>(Mask.getOperand(1)))
23895 if (auto *AmtConst = AmtBV->getConstantSplatNode())
23896 SraAmt = AmtConst->getZExtValue();
23897 } else if (Mask.getOpcode() == X86ISD::VSRAI) {
23898 SDValue SraC = Mask.getOperand(1);
23899 SraAmt = cast<ConstantSDNode>(SraC)->getZExtValue();
23901 if ((SraAmt + 1) != EltBits)
23906 // Now we know we at least have a plendvb with the mask val. See if
23907 // we can form a psignb/w/d.
23908 // psign = x.type == y.type == mask.type && y = sub(0, x);
23909 if (Y.getOpcode() == ISD::SUB && Y.getOperand(1) == X &&
23910 ISD::isBuildVectorAllZeros(Y.getOperand(0).getNode()) &&
23911 X.getValueType() == MaskVT && Y.getValueType() == MaskVT) {
23912 assert((EltBits == 8 || EltBits == 16 || EltBits == 32) &&
23913 "Unsupported VT for PSIGN");
23914 Mask = DAG.getNode(X86ISD::PSIGN, DL, MaskVT, X, Mask.getOperand(0));
23915 return DAG.getBitcast(VT, Mask);
23917 // PBLENDVB only available on SSE 4.1
23918 if (!Subtarget->hasSSE41())
23921 EVT BlendVT = (VT == MVT::v4i64) ? MVT::v32i8 : MVT::v16i8;
23923 X = DAG.getBitcast(BlendVT, X);
23924 Y = DAG.getBitcast(BlendVT, Y);
23925 Mask = DAG.getBitcast(BlendVT, Mask);
23926 Mask = DAG.getNode(ISD::VSELECT, DL, BlendVT, Mask, Y, X);
23927 return DAG.getBitcast(VT, Mask);
23931 if (VT != MVT::i16 && VT != MVT::i32 && VT != MVT::i64)
23934 // fold (or (x << c) | (y >> (64 - c))) ==> (shld64 x, y, c)
23935 MachineFunction &MF = DAG.getMachineFunction();
23937 MF.getFunction()->hasFnAttribute(Attribute::OptimizeForSize);
23939 // SHLD/SHRD instructions have lower register pressure, but on some
23940 // platforms they have higher latency than the equivalent
23941 // series of shifts/or that would otherwise be generated.
23942 // Don't fold (or (x << c) | (y >> (64 - c))) if SHLD/SHRD instructions
23943 // have higher latencies and we are not optimizing for size.
23944 if (!OptForSize && Subtarget->isSHLDSlow())
23947 if (N0.getOpcode() == ISD::SRL && N1.getOpcode() == ISD::SHL)
23949 if (N0.getOpcode() != ISD::SHL || N1.getOpcode() != ISD::SRL)
23951 if (!N0.hasOneUse() || !N1.hasOneUse())
23954 SDValue ShAmt0 = N0.getOperand(1);
23955 if (ShAmt0.getValueType() != MVT::i8)
23957 SDValue ShAmt1 = N1.getOperand(1);
23958 if (ShAmt1.getValueType() != MVT::i8)
23960 if (ShAmt0.getOpcode() == ISD::TRUNCATE)
23961 ShAmt0 = ShAmt0.getOperand(0);
23962 if (ShAmt1.getOpcode() == ISD::TRUNCATE)
23963 ShAmt1 = ShAmt1.getOperand(0);
23966 unsigned Opc = X86ISD::SHLD;
23967 SDValue Op0 = N0.getOperand(0);
23968 SDValue Op1 = N1.getOperand(0);
23969 if (ShAmt0.getOpcode() == ISD::SUB) {
23970 Opc = X86ISD::SHRD;
23971 std::swap(Op0, Op1);
23972 std::swap(ShAmt0, ShAmt1);
23975 unsigned Bits = VT.getSizeInBits();
23976 if (ShAmt1.getOpcode() == ISD::SUB) {
23977 SDValue Sum = ShAmt1.getOperand(0);
23978 if (ConstantSDNode *SumC = dyn_cast<ConstantSDNode>(Sum)) {
23979 SDValue ShAmt1Op1 = ShAmt1.getOperand(1);
23980 if (ShAmt1Op1.getNode()->getOpcode() == ISD::TRUNCATE)
23981 ShAmt1Op1 = ShAmt1Op1.getOperand(0);
23982 if (SumC->getSExtValue() == Bits && ShAmt1Op1 == ShAmt0)
23983 return DAG.getNode(Opc, DL, VT,
23985 DAG.getNode(ISD::TRUNCATE, DL,
23988 } else if (ConstantSDNode *ShAmt1C = dyn_cast<ConstantSDNode>(ShAmt1)) {
23989 ConstantSDNode *ShAmt0C = dyn_cast<ConstantSDNode>(ShAmt0);
23991 ShAmt0C->getSExtValue() + ShAmt1C->getSExtValue() == Bits)
23992 return DAG.getNode(Opc, DL, VT,
23993 N0.getOperand(0), N1.getOperand(0),
23994 DAG.getNode(ISD::TRUNCATE, DL,
24001 // Generate NEG and CMOV for integer abs.
24002 static SDValue performIntegerAbsCombine(SDNode *N, SelectionDAG &DAG) {
24003 EVT VT = N->getValueType(0);
24005 // Since X86 does not have CMOV for 8-bit integer, we don't convert
24006 // 8-bit integer abs to NEG and CMOV.
24007 if (VT.isInteger() && VT.getSizeInBits() == 8)
24010 SDValue N0 = N->getOperand(0);
24011 SDValue N1 = N->getOperand(1);
24014 // Check pattern of XOR(ADD(X,Y), Y) where Y is SRA(X, size(X)-1)
24015 // and change it to SUB and CMOV.
24016 if (VT.isInteger() && N->getOpcode() == ISD::XOR &&
24017 N0.getOpcode() == ISD::ADD &&
24018 N0.getOperand(1) == N1 &&
24019 N1.getOpcode() == ISD::SRA &&
24020 N1.getOperand(0) == N0.getOperand(0))
24021 if (ConstantSDNode *Y1C = dyn_cast<ConstantSDNode>(N1.getOperand(1)))
24022 if (Y1C->getAPIntValue() == VT.getSizeInBits()-1) {
24023 // Generate SUB & CMOV.
24024 SDValue Neg = DAG.getNode(X86ISD::SUB, DL, DAG.getVTList(VT, MVT::i32),
24025 DAG.getConstant(0, DL, VT), N0.getOperand(0));
24027 SDValue Ops[] = { N0.getOperand(0), Neg,
24028 DAG.getConstant(X86::COND_GE, DL, MVT::i8),
24029 SDValue(Neg.getNode(), 1) };
24030 return DAG.getNode(X86ISD::CMOV, DL, DAG.getVTList(VT, MVT::Glue), Ops);
24035 // PerformXorCombine - Attempts to turn XOR nodes into BLSMSK nodes
24036 static SDValue PerformXorCombine(SDNode *N, SelectionDAG &DAG,
24037 TargetLowering::DAGCombinerInfo &DCI,
24038 const X86Subtarget *Subtarget) {
24039 if (DCI.isBeforeLegalizeOps())
24042 if (Subtarget->hasCMov())
24043 if (SDValue RV = performIntegerAbsCombine(N, DAG))
24049 /// PerformLOADCombine - Do target-specific dag combines on LOAD nodes.
24050 static SDValue PerformLOADCombine(SDNode *N, SelectionDAG &DAG,
24051 TargetLowering::DAGCombinerInfo &DCI,
24052 const X86Subtarget *Subtarget) {
24053 LoadSDNode *Ld = cast<LoadSDNode>(N);
24054 EVT RegVT = Ld->getValueType(0);
24055 EVT MemVT = Ld->getMemoryVT();
24057 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
24059 // For chips with slow 32-byte unaligned loads, break the 32-byte operation
24060 // into two 16-byte operations.
24061 ISD::LoadExtType Ext = Ld->getExtensionType();
24062 unsigned Alignment = Ld->getAlignment();
24063 bool IsAligned = Alignment == 0 || Alignment >= MemVT.getSizeInBits()/8;
24064 if (RegVT.is256BitVector() && Subtarget->isUnalignedMem32Slow() &&
24065 !DCI.isBeforeLegalizeOps() && !IsAligned && Ext == ISD::NON_EXTLOAD) {
24066 unsigned NumElems = RegVT.getVectorNumElements();
24070 SDValue Ptr = Ld->getBasePtr();
24071 SDValue Increment =
24072 DAG.getConstant(16, dl, TLI.getPointerTy(DAG.getDataLayout()));
24074 EVT HalfVT = EVT::getVectorVT(*DAG.getContext(), MemVT.getScalarType(),
24076 SDValue Load1 = DAG.getLoad(HalfVT, dl, Ld->getChain(), Ptr,
24077 Ld->getPointerInfo(), Ld->isVolatile(),
24078 Ld->isNonTemporal(), Ld->isInvariant(),
24080 Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr, Increment);
24081 SDValue Load2 = DAG.getLoad(HalfVT, dl, Ld->getChain(), Ptr,
24082 Ld->getPointerInfo(), Ld->isVolatile(),
24083 Ld->isNonTemporal(), Ld->isInvariant(),
24084 std::min(16U, Alignment));
24085 SDValue TF = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
24087 Load2.getValue(1));
24089 SDValue NewVec = DAG.getUNDEF(RegVT);
24090 NewVec = Insert128BitVector(NewVec, Load1, 0, DAG, dl);
24091 NewVec = Insert128BitVector(NewVec, Load2, NumElems/2, DAG, dl);
24092 return DCI.CombineTo(N, NewVec, TF, true);
24098 /// PerformMLOADCombine - Resolve extending loads
24099 static SDValue PerformMLOADCombine(SDNode *N, SelectionDAG &DAG,
24100 TargetLowering::DAGCombinerInfo &DCI,
24101 const X86Subtarget *Subtarget) {
24102 MaskedLoadSDNode *Mld = cast<MaskedLoadSDNode>(N);
24103 if (Mld->getExtensionType() != ISD::SEXTLOAD)
24106 EVT VT = Mld->getValueType(0);
24107 unsigned NumElems = VT.getVectorNumElements();
24108 EVT LdVT = Mld->getMemoryVT();
24111 assert(LdVT != VT && "Cannot extend to the same type");
24112 unsigned ToSz = VT.getVectorElementType().getSizeInBits();
24113 unsigned FromSz = LdVT.getVectorElementType().getSizeInBits();
24114 // From, To sizes and ElemCount must be pow of two
24115 assert (isPowerOf2_32(NumElems * FromSz * ToSz) &&
24116 "Unexpected size for extending masked load");
24118 unsigned SizeRatio = ToSz / FromSz;
24119 assert(SizeRatio * NumElems * FromSz == VT.getSizeInBits());
24121 // Create a type on which we perform the shuffle
24122 EVT WideVecVT = EVT::getVectorVT(*DAG.getContext(),
24123 LdVT.getScalarType(), NumElems*SizeRatio);
24124 assert(WideVecVT.getSizeInBits() == VT.getSizeInBits());
24126 // Convert Src0 value
24127 SDValue WideSrc0 = DAG.getBitcast(WideVecVT, Mld->getSrc0());
24128 if (Mld->getSrc0().getOpcode() != ISD::UNDEF) {
24129 SmallVector<int, 16> ShuffleVec(NumElems * SizeRatio, -1);
24130 for (unsigned i = 0; i != NumElems; ++i)
24131 ShuffleVec[i] = i * SizeRatio;
24133 // Can't shuffle using an illegal type.
24134 assert (DAG.getTargetLoweringInfo().isTypeLegal(WideVecVT)
24135 && "WideVecVT should be legal");
24136 WideSrc0 = DAG.getVectorShuffle(WideVecVT, dl, WideSrc0,
24137 DAG.getUNDEF(WideVecVT), &ShuffleVec[0]);
24139 // Prepare the new mask
24141 SDValue Mask = Mld->getMask();
24142 if (Mask.getValueType() == VT) {
24143 // Mask and original value have the same type
24144 NewMask = DAG.getBitcast(WideVecVT, Mask);
24145 SmallVector<int, 16> ShuffleVec(NumElems * SizeRatio, -1);
24146 for (unsigned i = 0; i != NumElems; ++i)
24147 ShuffleVec[i] = i * SizeRatio;
24148 for (unsigned i = NumElems; i != NumElems*SizeRatio; ++i)
24149 ShuffleVec[i] = NumElems*SizeRatio;
24150 NewMask = DAG.getVectorShuffle(WideVecVT, dl, NewMask,
24151 DAG.getConstant(0, dl, WideVecVT),
24155 assert(Mask.getValueType().getVectorElementType() == MVT::i1);
24156 unsigned WidenNumElts = NumElems*SizeRatio;
24157 unsigned MaskNumElts = VT.getVectorNumElements();
24158 EVT NewMaskVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
24161 unsigned NumConcat = WidenNumElts / MaskNumElts;
24162 SmallVector<SDValue, 16> Ops(NumConcat);
24163 SDValue ZeroVal = DAG.getConstant(0, dl, Mask.getValueType());
24165 for (unsigned i = 1; i != NumConcat; ++i)
24168 NewMask = DAG.getNode(ISD::CONCAT_VECTORS, dl, NewMaskVT, Ops);
24171 SDValue WideLd = DAG.getMaskedLoad(WideVecVT, dl, Mld->getChain(),
24172 Mld->getBasePtr(), NewMask, WideSrc0,
24173 Mld->getMemoryVT(), Mld->getMemOperand(),
24175 SDValue NewVec = DAG.getNode(X86ISD::VSEXT, dl, VT, WideLd);
24176 return DCI.CombineTo(N, NewVec, WideLd.getValue(1), true);
24179 /// PerformMSTORECombine - Resolve truncating stores
24180 static SDValue PerformMSTORECombine(SDNode *N, SelectionDAG &DAG,
24181 const X86Subtarget *Subtarget) {
24182 MaskedStoreSDNode *Mst = cast<MaskedStoreSDNode>(N);
24183 if (!Mst->isTruncatingStore())
24186 EVT VT = Mst->getValue().getValueType();
24187 unsigned NumElems = VT.getVectorNumElements();
24188 EVT StVT = Mst->getMemoryVT();
24191 assert(StVT != VT && "Cannot truncate to the same type");
24192 unsigned FromSz = VT.getVectorElementType().getSizeInBits();
24193 unsigned ToSz = StVT.getVectorElementType().getSizeInBits();
24195 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
24197 // The truncating store is legal in some cases. For example
24198 // vpmovqb, vpmovqw, vpmovqd, vpmovdb, vpmovdw
24199 // are designated for truncate store.
24200 // In this case we don't need any further transformations.
24201 if (TLI.isTruncStoreLegal(VT, StVT))
24204 // From, To sizes and ElemCount must be pow of two
24205 assert (isPowerOf2_32(NumElems * FromSz * ToSz) &&
24206 "Unexpected size for truncating masked store");
24207 // We are going to use the original vector elt for storing.
24208 // Accumulated smaller vector elements must be a multiple of the store size.
24209 assert (((NumElems * FromSz) % ToSz) == 0 &&
24210 "Unexpected ratio for truncating masked store");
24212 unsigned SizeRatio = FromSz / ToSz;
24213 assert(SizeRatio * NumElems * ToSz == VT.getSizeInBits());
24215 // Create a type on which we perform the shuffle
24216 EVT WideVecVT = EVT::getVectorVT(*DAG.getContext(),
24217 StVT.getScalarType(), NumElems*SizeRatio);
24219 assert(WideVecVT.getSizeInBits() == VT.getSizeInBits());
24221 SDValue WideVec = DAG.getBitcast(WideVecVT, Mst->getValue());
24222 SmallVector<int, 16> ShuffleVec(NumElems * SizeRatio, -1);
24223 for (unsigned i = 0; i != NumElems; ++i)
24224 ShuffleVec[i] = i * SizeRatio;
24226 // Can't shuffle using an illegal type.
24227 assert (DAG.getTargetLoweringInfo().isTypeLegal(WideVecVT)
24228 && "WideVecVT should be legal");
24230 SDValue TruncatedVal = DAG.getVectorShuffle(WideVecVT, dl, WideVec,
24231 DAG.getUNDEF(WideVecVT),
24235 SDValue Mask = Mst->getMask();
24236 if (Mask.getValueType() == VT) {
24237 // Mask and original value have the same type
24238 NewMask = DAG.getBitcast(WideVecVT, Mask);
24239 for (unsigned i = 0; i != NumElems; ++i)
24240 ShuffleVec[i] = i * SizeRatio;
24241 for (unsigned i = NumElems; i != NumElems*SizeRatio; ++i)
24242 ShuffleVec[i] = NumElems*SizeRatio;
24243 NewMask = DAG.getVectorShuffle(WideVecVT, dl, NewMask,
24244 DAG.getConstant(0, dl, WideVecVT),
24248 assert(Mask.getValueType().getVectorElementType() == MVT::i1);
24249 unsigned WidenNumElts = NumElems*SizeRatio;
24250 unsigned MaskNumElts = VT.getVectorNumElements();
24251 EVT NewMaskVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
24254 unsigned NumConcat = WidenNumElts / MaskNumElts;
24255 SmallVector<SDValue, 16> Ops(NumConcat);
24256 SDValue ZeroVal = DAG.getConstant(0, dl, Mask.getValueType());
24258 for (unsigned i = 1; i != NumConcat; ++i)
24261 NewMask = DAG.getNode(ISD::CONCAT_VECTORS, dl, NewMaskVT, Ops);
24264 return DAG.getMaskedStore(Mst->getChain(), dl, TruncatedVal, Mst->getBasePtr(),
24265 NewMask, StVT, Mst->getMemOperand(), false);
24267 /// PerformSTORECombine - Do target-specific dag combines on STORE nodes.
24268 static SDValue PerformSTORECombine(SDNode *N, SelectionDAG &DAG,
24269 const X86Subtarget *Subtarget) {
24270 StoreSDNode *St = cast<StoreSDNode>(N);
24271 EVT VT = St->getValue().getValueType();
24272 EVT StVT = St->getMemoryVT();
24274 SDValue StoredVal = St->getOperand(1);
24275 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
24277 // If we are saving a concatenation of two XMM registers and 32-byte stores
24278 // are slow, such as on Sandy Bridge, perform two 16-byte stores.
24279 unsigned Alignment = St->getAlignment();
24280 bool IsAligned = Alignment == 0 || Alignment >= VT.getSizeInBits()/8;
24281 if (VT.is256BitVector() && Subtarget->isUnalignedMem32Slow() &&
24282 StVT == VT && !IsAligned) {
24283 unsigned NumElems = VT.getVectorNumElements();
24287 SDValue Value0 = Extract128BitVector(StoredVal, 0, DAG, dl);
24288 SDValue Value1 = Extract128BitVector(StoredVal, NumElems/2, DAG, dl);
24291 DAG.getConstant(16, dl, TLI.getPointerTy(DAG.getDataLayout()));
24292 SDValue Ptr0 = St->getBasePtr();
24293 SDValue Ptr1 = DAG.getNode(ISD::ADD, dl, Ptr0.getValueType(), Ptr0, Stride);
24295 SDValue Ch0 = DAG.getStore(St->getChain(), dl, Value0, Ptr0,
24296 St->getPointerInfo(), St->isVolatile(),
24297 St->isNonTemporal(), Alignment);
24298 SDValue Ch1 = DAG.getStore(St->getChain(), dl, Value1, Ptr1,
24299 St->getPointerInfo(), St->isVolatile(),
24300 St->isNonTemporal(),
24301 std::min(16U, Alignment));
24302 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Ch0, Ch1);
24305 // Optimize trunc store (of multiple scalars) to shuffle and store.
24306 // First, pack all of the elements in one place. Next, store to memory
24307 // in fewer chunks.
24308 if (St->isTruncatingStore() && VT.isVector()) {
24309 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
24310 unsigned NumElems = VT.getVectorNumElements();
24311 assert(StVT != VT && "Cannot truncate to the same type");
24312 unsigned FromSz = VT.getVectorElementType().getSizeInBits();
24313 unsigned ToSz = StVT.getVectorElementType().getSizeInBits();
24315 // The truncating store is legal in some cases. For example
24316 // vpmovqb, vpmovqw, vpmovqd, vpmovdb, vpmovdw
24317 // are designated for truncate store.
24318 // In this case we don't need any further transformations.
24319 if (TLI.isTruncStoreLegal(VT, StVT))
24322 // From, To sizes and ElemCount must be pow of two
24323 if (!isPowerOf2_32(NumElems * FromSz * ToSz)) return SDValue();
24324 // We are going to use the original vector elt for storing.
24325 // Accumulated smaller vector elements must be a multiple of the store size.
24326 if (0 != (NumElems * FromSz) % ToSz) return SDValue();
24328 unsigned SizeRatio = FromSz / ToSz;
24330 assert(SizeRatio * NumElems * ToSz == VT.getSizeInBits());
24332 // Create a type on which we perform the shuffle
24333 EVT WideVecVT = EVT::getVectorVT(*DAG.getContext(),
24334 StVT.getScalarType(), NumElems*SizeRatio);
24336 assert(WideVecVT.getSizeInBits() == VT.getSizeInBits());
24338 SDValue WideVec = DAG.getBitcast(WideVecVT, St->getValue());
24339 SmallVector<int, 8> ShuffleVec(NumElems * SizeRatio, -1);
24340 for (unsigned i = 0; i != NumElems; ++i)
24341 ShuffleVec[i] = i * SizeRatio;
24343 // Can't shuffle using an illegal type.
24344 if (!TLI.isTypeLegal(WideVecVT))
24347 SDValue Shuff = DAG.getVectorShuffle(WideVecVT, dl, WideVec,
24348 DAG.getUNDEF(WideVecVT),
24350 // At this point all of the data is stored at the bottom of the
24351 // register. We now need to save it to mem.
24353 // Find the largest store unit
24354 MVT StoreType = MVT::i8;
24355 for (MVT Tp : MVT::integer_valuetypes()) {
24356 if (TLI.isTypeLegal(Tp) && Tp.getSizeInBits() <= NumElems * ToSz)
24360 // On 32bit systems, we can't save 64bit integers. Try bitcasting to F64.
24361 if (TLI.isTypeLegal(MVT::f64) && StoreType.getSizeInBits() < 64 &&
24362 (64 <= NumElems * ToSz))
24363 StoreType = MVT::f64;
24365 // Bitcast the original vector into a vector of store-size units
24366 EVT StoreVecVT = EVT::getVectorVT(*DAG.getContext(),
24367 StoreType, VT.getSizeInBits()/StoreType.getSizeInBits());
24368 assert(StoreVecVT.getSizeInBits() == VT.getSizeInBits());
24369 SDValue ShuffWide = DAG.getBitcast(StoreVecVT, Shuff);
24370 SmallVector<SDValue, 8> Chains;
24371 SDValue Increment = DAG.getConstant(StoreType.getSizeInBits() / 8, dl,
24372 TLI.getPointerTy(DAG.getDataLayout()));
24373 SDValue Ptr = St->getBasePtr();
24375 // Perform one or more big stores into memory.
24376 for (unsigned i=0, e=(ToSz*NumElems)/StoreType.getSizeInBits(); i!=e; ++i) {
24377 SDValue SubVec = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl,
24378 StoreType, ShuffWide,
24379 DAG.getIntPtrConstant(i, dl));
24380 SDValue Ch = DAG.getStore(St->getChain(), dl, SubVec, Ptr,
24381 St->getPointerInfo(), St->isVolatile(),
24382 St->isNonTemporal(), St->getAlignment());
24383 Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr, Increment);
24384 Chains.push_back(Ch);
24387 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Chains);
24390 // Turn load->store of MMX types into GPR load/stores. This avoids clobbering
24391 // the FP state in cases where an emms may be missing.
24392 // A preferable solution to the general problem is to figure out the right
24393 // places to insert EMMS. This qualifies as a quick hack.
24395 // Similarly, turn load->store of i64 into double load/stores in 32-bit mode.
24396 if (VT.getSizeInBits() != 64)
24399 const Function *F = DAG.getMachineFunction().getFunction();
24400 bool NoImplicitFloatOps = F->hasFnAttribute(Attribute::NoImplicitFloat);
24402 !Subtarget->useSoftFloat() && !NoImplicitFloatOps && Subtarget->hasSSE2();
24403 if ((VT.isVector() ||
24404 (VT == MVT::i64 && F64IsLegal && !Subtarget->is64Bit())) &&
24405 isa<LoadSDNode>(St->getValue()) &&
24406 !cast<LoadSDNode>(St->getValue())->isVolatile() &&
24407 St->getChain().hasOneUse() && !St->isVolatile()) {
24408 SDNode* LdVal = St->getValue().getNode();
24409 LoadSDNode *Ld = nullptr;
24410 int TokenFactorIndex = -1;
24411 SmallVector<SDValue, 8> Ops;
24412 SDNode* ChainVal = St->getChain().getNode();
24413 // Must be a store of a load. We currently handle two cases: the load
24414 // is a direct child, and it's under an intervening TokenFactor. It is
24415 // possible to dig deeper under nested TokenFactors.
24416 if (ChainVal == LdVal)
24417 Ld = cast<LoadSDNode>(St->getChain());
24418 else if (St->getValue().hasOneUse() &&
24419 ChainVal->getOpcode() == ISD::TokenFactor) {
24420 for (unsigned i = 0, e = ChainVal->getNumOperands(); i != e; ++i) {
24421 if (ChainVal->getOperand(i).getNode() == LdVal) {
24422 TokenFactorIndex = i;
24423 Ld = cast<LoadSDNode>(St->getValue());
24425 Ops.push_back(ChainVal->getOperand(i));
24429 if (!Ld || !ISD::isNormalLoad(Ld))
24432 // If this is not the MMX case, i.e. we are just turning i64 load/store
24433 // into f64 load/store, avoid the transformation if there are multiple
24434 // uses of the loaded value.
24435 if (!VT.isVector() && !Ld->hasNUsesOfValue(1, 0))
24440 // If we are a 64-bit capable x86, lower to a single movq load/store pair.
24441 // Otherwise, if it's legal to use f64 SSE instructions, use f64 load/store
24443 if (Subtarget->is64Bit() || F64IsLegal) {
24444 EVT LdVT = Subtarget->is64Bit() ? MVT::i64 : MVT::f64;
24445 SDValue NewLd = DAG.getLoad(LdVT, LdDL, Ld->getChain(), Ld->getBasePtr(),
24446 Ld->getPointerInfo(), Ld->isVolatile(),
24447 Ld->isNonTemporal(), Ld->isInvariant(),
24448 Ld->getAlignment());
24449 SDValue NewChain = NewLd.getValue(1);
24450 if (TokenFactorIndex != -1) {
24451 Ops.push_back(NewChain);
24452 NewChain = DAG.getNode(ISD::TokenFactor, LdDL, MVT::Other, Ops);
24454 return DAG.getStore(NewChain, StDL, NewLd, St->getBasePtr(),
24455 St->getPointerInfo(),
24456 St->isVolatile(), St->isNonTemporal(),
24457 St->getAlignment());
24460 // Otherwise, lower to two pairs of 32-bit loads / stores.
24461 SDValue LoAddr = Ld->getBasePtr();
24462 SDValue HiAddr = DAG.getNode(ISD::ADD, LdDL, MVT::i32, LoAddr,
24463 DAG.getConstant(4, LdDL, MVT::i32));
24465 SDValue LoLd = DAG.getLoad(MVT::i32, LdDL, Ld->getChain(), LoAddr,
24466 Ld->getPointerInfo(),
24467 Ld->isVolatile(), Ld->isNonTemporal(),
24468 Ld->isInvariant(), Ld->getAlignment());
24469 SDValue HiLd = DAG.getLoad(MVT::i32, LdDL, Ld->getChain(), HiAddr,
24470 Ld->getPointerInfo().getWithOffset(4),
24471 Ld->isVolatile(), Ld->isNonTemporal(),
24473 MinAlign(Ld->getAlignment(), 4));
24475 SDValue NewChain = LoLd.getValue(1);
24476 if (TokenFactorIndex != -1) {
24477 Ops.push_back(LoLd);
24478 Ops.push_back(HiLd);
24479 NewChain = DAG.getNode(ISD::TokenFactor, LdDL, MVT::Other, Ops);
24482 LoAddr = St->getBasePtr();
24483 HiAddr = DAG.getNode(ISD::ADD, StDL, MVT::i32, LoAddr,
24484 DAG.getConstant(4, StDL, MVT::i32));
24486 SDValue LoSt = DAG.getStore(NewChain, StDL, LoLd, LoAddr,
24487 St->getPointerInfo(),
24488 St->isVolatile(), St->isNonTemporal(),
24489 St->getAlignment());
24490 SDValue HiSt = DAG.getStore(NewChain, StDL, HiLd, HiAddr,
24491 St->getPointerInfo().getWithOffset(4),
24493 St->isNonTemporal(),
24494 MinAlign(St->getAlignment(), 4));
24495 return DAG.getNode(ISD::TokenFactor, StDL, MVT::Other, LoSt, HiSt);
24498 // This is similar to the above case, but here we handle a scalar 64-bit
24499 // integer store that is extracted from a vector on a 32-bit target.
24500 // If we have SSE2, then we can treat it like a floating-point double
24501 // to get past legalization. The execution dependencies fixup pass will
24502 // choose the optimal machine instruction for the store if this really is
24503 // an integer or v2f32 rather than an f64.
24504 if (VT == MVT::i64 && F64IsLegal && !Subtarget->is64Bit() &&
24505 St->getOperand(1).getOpcode() == ISD::EXTRACT_VECTOR_ELT) {
24506 SDValue OldExtract = St->getOperand(1);
24507 SDValue ExtOp0 = OldExtract.getOperand(0);
24508 unsigned VecSize = ExtOp0.getValueSizeInBits();
24509 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), MVT::f64, VecSize / 64);
24510 SDValue BitCast = DAG.getBitcast(VecVT, ExtOp0);
24511 SDValue NewExtract = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
24512 BitCast, OldExtract.getOperand(1));
24513 return DAG.getStore(St->getChain(), dl, NewExtract, St->getBasePtr(),
24514 St->getPointerInfo(), St->isVolatile(),
24515 St->isNonTemporal(), St->getAlignment());
24521 /// Return 'true' if this vector operation is "horizontal"
24522 /// and return the operands for the horizontal operation in LHS and RHS. A
24523 /// horizontal operation performs the binary operation on successive elements
24524 /// of its first operand, then on successive elements of its second operand,
24525 /// returning the resulting values in a vector. For example, if
24526 /// A = < float a0, float a1, float a2, float a3 >
24528 /// B = < float b0, float b1, float b2, float b3 >
24529 /// then the result of doing a horizontal operation on A and B is
24530 /// A horizontal-op B = < a0 op a1, a2 op a3, b0 op b1, b2 op b3 >.
24531 /// In short, LHS and RHS are inspected to see if LHS op RHS is of the form
24532 /// A horizontal-op B, for some already available A and B, and if so then LHS is
24533 /// set to A, RHS to B, and the routine returns 'true'.
24534 /// Note that the binary operation should have the property that if one of the
24535 /// operands is UNDEF then the result is UNDEF.
24536 static bool isHorizontalBinOp(SDValue &LHS, SDValue &RHS, bool IsCommutative) {
24537 // Look for the following pattern: if
24538 // A = < float a0, float a1, float a2, float a3 >
24539 // B = < float b0, float b1, float b2, float b3 >
24541 // LHS = VECTOR_SHUFFLE A, B, <0, 2, 4, 6>
24542 // RHS = VECTOR_SHUFFLE A, B, <1, 3, 5, 7>
24543 // then LHS op RHS = < a0 op a1, a2 op a3, b0 op b1, b2 op b3 >
24544 // which is A horizontal-op B.
24546 // At least one of the operands should be a vector shuffle.
24547 if (LHS.getOpcode() != ISD::VECTOR_SHUFFLE &&
24548 RHS.getOpcode() != ISD::VECTOR_SHUFFLE)
24551 MVT VT = LHS.getSimpleValueType();
24553 assert((VT.is128BitVector() || VT.is256BitVector()) &&
24554 "Unsupported vector type for horizontal add/sub");
24556 // Handle 128 and 256-bit vector lengths. AVX defines horizontal add/sub to
24557 // operate independently on 128-bit lanes.
24558 unsigned NumElts = VT.getVectorNumElements();
24559 unsigned NumLanes = VT.getSizeInBits()/128;
24560 unsigned NumLaneElts = NumElts / NumLanes;
24561 assert((NumLaneElts % 2 == 0) &&
24562 "Vector type should have an even number of elements in each lane");
24563 unsigned HalfLaneElts = NumLaneElts/2;
24565 // View LHS in the form
24566 // LHS = VECTOR_SHUFFLE A, B, LMask
24567 // If LHS is not a shuffle then pretend it is the shuffle
24568 // LHS = VECTOR_SHUFFLE LHS, undef, <0, 1, ..., N-1>
24569 // NOTE: in what follows a default initialized SDValue represents an UNDEF of
24572 SmallVector<int, 16> LMask(NumElts);
24573 if (LHS.getOpcode() == ISD::VECTOR_SHUFFLE) {
24574 if (LHS.getOperand(0).getOpcode() != ISD::UNDEF)
24575 A = LHS.getOperand(0);
24576 if (LHS.getOperand(1).getOpcode() != ISD::UNDEF)
24577 B = LHS.getOperand(1);
24578 ArrayRef<int> Mask = cast<ShuffleVectorSDNode>(LHS.getNode())->getMask();
24579 std::copy(Mask.begin(), Mask.end(), LMask.begin());
24581 if (LHS.getOpcode() != ISD::UNDEF)
24583 for (unsigned i = 0; i != NumElts; ++i)
24587 // Likewise, view RHS in the form
24588 // RHS = VECTOR_SHUFFLE C, D, RMask
24590 SmallVector<int, 16> RMask(NumElts);
24591 if (RHS.getOpcode() == ISD::VECTOR_SHUFFLE) {
24592 if (RHS.getOperand(0).getOpcode() != ISD::UNDEF)
24593 C = RHS.getOperand(0);
24594 if (RHS.getOperand(1).getOpcode() != ISD::UNDEF)
24595 D = RHS.getOperand(1);
24596 ArrayRef<int> Mask = cast<ShuffleVectorSDNode>(RHS.getNode())->getMask();
24597 std::copy(Mask.begin(), Mask.end(), RMask.begin());
24599 if (RHS.getOpcode() != ISD::UNDEF)
24601 for (unsigned i = 0; i != NumElts; ++i)
24605 // Check that the shuffles are both shuffling the same vectors.
24606 if (!(A == C && B == D) && !(A == D && B == C))
24609 // If everything is UNDEF then bail out: it would be better to fold to UNDEF.
24610 if (!A.getNode() && !B.getNode())
24613 // If A and B occur in reverse order in RHS, then "swap" them (which means
24614 // rewriting the mask).
24616 ShuffleVectorSDNode::commuteMask(RMask);
24618 // At this point LHS and RHS are equivalent to
24619 // LHS = VECTOR_SHUFFLE A, B, LMask
24620 // RHS = VECTOR_SHUFFLE A, B, RMask
24621 // Check that the masks correspond to performing a horizontal operation.
24622 for (unsigned l = 0; l != NumElts; l += NumLaneElts) {
24623 for (unsigned i = 0; i != NumLaneElts; ++i) {
24624 int LIdx = LMask[i+l], RIdx = RMask[i+l];
24626 // Ignore any UNDEF components.
24627 if (LIdx < 0 || RIdx < 0 ||
24628 (!A.getNode() && (LIdx < (int)NumElts || RIdx < (int)NumElts)) ||
24629 (!B.getNode() && (LIdx >= (int)NumElts || RIdx >= (int)NumElts)))
24632 // Check that successive elements are being operated on. If not, this is
24633 // not a horizontal operation.
24634 unsigned Src = (i/HalfLaneElts); // each lane is split between srcs
24635 int Index = 2*(i%HalfLaneElts) + NumElts*Src + l;
24636 if (!(LIdx == Index && RIdx == Index + 1) &&
24637 !(IsCommutative && LIdx == Index + 1 && RIdx == Index))
24642 LHS = A.getNode() ? A : B; // If A is 'UNDEF', use B for it.
24643 RHS = B.getNode() ? B : A; // If B is 'UNDEF', use A for it.
24647 /// Do target-specific dag combines on floating point adds.
24648 static SDValue PerformFADDCombine(SDNode *N, SelectionDAG &DAG,
24649 const X86Subtarget *Subtarget) {
24650 EVT VT = N->getValueType(0);
24651 SDValue LHS = N->getOperand(0);
24652 SDValue RHS = N->getOperand(1);
24654 // Try to synthesize horizontal adds from adds of shuffles.
24655 if (((Subtarget->hasSSE3() && (VT == MVT::v4f32 || VT == MVT::v2f64)) ||
24656 (Subtarget->hasFp256() && (VT == MVT::v8f32 || VT == MVT::v4f64))) &&
24657 isHorizontalBinOp(LHS, RHS, true))
24658 return DAG.getNode(X86ISD::FHADD, SDLoc(N), VT, LHS, RHS);
24662 /// Do target-specific dag combines on floating point subs.
24663 static SDValue PerformFSUBCombine(SDNode *N, SelectionDAG &DAG,
24664 const X86Subtarget *Subtarget) {
24665 EVT VT = N->getValueType(0);
24666 SDValue LHS = N->getOperand(0);
24667 SDValue RHS = N->getOperand(1);
24669 // Try to synthesize horizontal subs from subs of shuffles.
24670 if (((Subtarget->hasSSE3() && (VT == MVT::v4f32 || VT == MVT::v2f64)) ||
24671 (Subtarget->hasFp256() && (VT == MVT::v8f32 || VT == MVT::v4f64))) &&
24672 isHorizontalBinOp(LHS, RHS, false))
24673 return DAG.getNode(X86ISD::FHSUB, SDLoc(N), VT, LHS, RHS);
24677 /// Do target-specific dag combines on X86ISD::FOR and X86ISD::FXOR nodes.
24678 static SDValue PerformFORCombine(SDNode *N, SelectionDAG &DAG) {
24679 assert(N->getOpcode() == X86ISD::FOR || N->getOpcode() == X86ISD::FXOR);
24681 // F[X]OR(0.0, x) -> x
24682 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
24683 if (C->getValueAPF().isPosZero())
24684 return N->getOperand(1);
24686 // F[X]OR(x, 0.0) -> x
24687 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
24688 if (C->getValueAPF().isPosZero())
24689 return N->getOperand(0);
24693 /// Do target-specific dag combines on X86ISD::FMIN and X86ISD::FMAX nodes.
24694 static SDValue PerformFMinFMaxCombine(SDNode *N, SelectionDAG &DAG) {
24695 assert(N->getOpcode() == X86ISD::FMIN || N->getOpcode() == X86ISD::FMAX);
24697 // Only perform optimizations if UnsafeMath is used.
24698 if (!DAG.getTarget().Options.UnsafeFPMath)
24701 // If we run in unsafe-math mode, then convert the FMAX and FMIN nodes
24702 // into FMINC and FMAXC, which are Commutative operations.
24703 unsigned NewOp = 0;
24704 switch (N->getOpcode()) {
24705 default: llvm_unreachable("unknown opcode");
24706 case X86ISD::FMIN: NewOp = X86ISD::FMINC; break;
24707 case X86ISD::FMAX: NewOp = X86ISD::FMAXC; break;
24710 return DAG.getNode(NewOp, SDLoc(N), N->getValueType(0),
24711 N->getOperand(0), N->getOperand(1));
24714 /// Do target-specific dag combines on X86ISD::FAND nodes.
24715 static SDValue PerformFANDCombine(SDNode *N, SelectionDAG &DAG) {
24716 // FAND(0.0, x) -> 0.0
24717 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
24718 if (C->getValueAPF().isPosZero())
24719 return N->getOperand(0);
24721 // FAND(x, 0.0) -> 0.0
24722 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
24723 if (C->getValueAPF().isPosZero())
24724 return N->getOperand(1);
24729 /// Do target-specific dag combines on X86ISD::FANDN nodes
24730 static SDValue PerformFANDNCombine(SDNode *N, SelectionDAG &DAG) {
24731 // FANDN(0.0, x) -> x
24732 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
24733 if (C->getValueAPF().isPosZero())
24734 return N->getOperand(1);
24736 // FANDN(x, 0.0) -> 0.0
24737 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
24738 if (C->getValueAPF().isPosZero())
24739 return N->getOperand(1);
24744 static SDValue PerformBTCombine(SDNode *N,
24746 TargetLowering::DAGCombinerInfo &DCI) {
24747 // BT ignores high bits in the bit index operand.
24748 SDValue Op1 = N->getOperand(1);
24749 if (Op1.hasOneUse()) {
24750 unsigned BitWidth = Op1.getValueSizeInBits();
24751 APInt DemandedMask = APInt::getLowBitsSet(BitWidth, Log2_32(BitWidth));
24752 APInt KnownZero, KnownOne;
24753 TargetLowering::TargetLoweringOpt TLO(DAG, !DCI.isBeforeLegalize(),
24754 !DCI.isBeforeLegalizeOps());
24755 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
24756 if (TLO.ShrinkDemandedConstant(Op1, DemandedMask) ||
24757 TLI.SimplifyDemandedBits(Op1, DemandedMask, KnownZero, KnownOne, TLO))
24758 DCI.CommitTargetLoweringOpt(TLO);
24763 static SDValue PerformVZEXT_MOVLCombine(SDNode *N, SelectionDAG &DAG) {
24764 SDValue Op = N->getOperand(0);
24765 if (Op.getOpcode() == ISD::BITCAST)
24766 Op = Op.getOperand(0);
24767 EVT VT = N->getValueType(0), OpVT = Op.getValueType();
24768 if (Op.getOpcode() == X86ISD::VZEXT_LOAD &&
24769 VT.getVectorElementType().getSizeInBits() ==
24770 OpVT.getVectorElementType().getSizeInBits()) {
24771 return DAG.getNode(ISD::BITCAST, SDLoc(N), VT, Op);
24776 static SDValue PerformSIGN_EXTEND_INREGCombine(SDNode *N, SelectionDAG &DAG,
24777 const X86Subtarget *Subtarget) {
24778 EVT VT = N->getValueType(0);
24779 if (!VT.isVector())
24782 SDValue N0 = N->getOperand(0);
24783 SDValue N1 = N->getOperand(1);
24784 EVT ExtraVT = cast<VTSDNode>(N1)->getVT();
24787 // The SIGN_EXTEND_INREG to v4i64 is expensive operation on the
24788 // both SSE and AVX2 since there is no sign-extended shift right
24789 // operation on a vector with 64-bit elements.
24790 //(sext_in_reg (v4i64 anyext (v4i32 x )), ExtraVT) ->
24791 // (v4i64 sext (v4i32 sext_in_reg (v4i32 x , ExtraVT)))
24792 if (VT == MVT::v4i64 && (N0.getOpcode() == ISD::ANY_EXTEND ||
24793 N0.getOpcode() == ISD::SIGN_EXTEND)) {
24794 SDValue N00 = N0.getOperand(0);
24796 // EXTLOAD has a better solution on AVX2,
24797 // it may be replaced with X86ISD::VSEXT node.
24798 if (N00.getOpcode() == ISD::LOAD && Subtarget->hasInt256())
24799 if (!ISD::isNormalLoad(N00.getNode()))
24802 if (N00.getValueType() == MVT::v4i32 && ExtraVT.getSizeInBits() < 128) {
24803 SDValue Tmp = DAG.getNode(ISD::SIGN_EXTEND_INREG, dl, MVT::v4i32,
24805 return DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i64, Tmp);
24811 static SDValue PerformSExtCombine(SDNode *N, SelectionDAG &DAG,
24812 TargetLowering::DAGCombinerInfo &DCI,
24813 const X86Subtarget *Subtarget) {
24814 SDValue N0 = N->getOperand(0);
24815 EVT VT = N->getValueType(0);
24816 EVT SVT = VT.getScalarType();
24817 EVT InVT = N0.getValueType();
24818 EVT InSVT = InVT.getScalarType();
24821 // (i8,i32 sext (sdivrem (i8 x, i8 y)) ->
24822 // (i8,i32 (sdivrem_sext_hreg (i8 x, i8 y)
24823 // This exposes the sext to the sdivrem lowering, so that it directly extends
24824 // from AH (which we otherwise need to do contortions to access).
24825 if (N0.getOpcode() == ISD::SDIVREM && N0.getResNo() == 1 &&
24826 InVT == MVT::i8 && VT == MVT::i32) {
24827 SDVTList NodeTys = DAG.getVTList(MVT::i8, VT);
24828 SDValue R = DAG.getNode(X86ISD::SDIVREM8_SEXT_HREG, DL, NodeTys,
24829 N0.getOperand(0), N0.getOperand(1));
24830 DAG.ReplaceAllUsesOfValueWith(N0.getValue(0), R.getValue(0));
24831 return R.getValue(1);
24834 if (!DCI.isBeforeLegalizeOps()) {
24835 if (InVT == MVT::i1) {
24836 SDValue Zero = DAG.getConstant(0, DL, VT);
24838 DAG.getConstant(APInt::getAllOnesValue(VT.getSizeInBits()), DL, VT);
24839 return DAG.getNode(ISD::SELECT, DL, VT, N0, AllOnes, Zero);
24844 if (VT.isVector() && Subtarget->hasSSE2()) {
24845 auto ExtendVecSize = [&DAG](SDLoc DL, SDValue N, unsigned Size) {
24846 EVT InVT = N.getValueType();
24847 EVT OutVT = EVT::getVectorVT(*DAG.getContext(), InVT.getScalarType(),
24848 Size / InVT.getScalarSizeInBits());
24849 SmallVector<SDValue, 8> Opnds(Size / InVT.getSizeInBits(),
24850 DAG.getUNDEF(InVT));
24852 return DAG.getNode(ISD::CONCAT_VECTORS, DL, OutVT, Opnds);
24855 // If target-size is less than 128-bits, extend to a type that would extend
24856 // to 128 bits, extend that and extract the original target vector.
24857 if (VT.getSizeInBits() < 128 && !(128 % VT.getSizeInBits()) &&
24858 (SVT == MVT::i64 || SVT == MVT::i32 || SVT == MVT::i16) &&
24859 (InSVT == MVT::i32 || InSVT == MVT::i16 || InSVT == MVT::i8)) {
24860 unsigned Scale = 128 / VT.getSizeInBits();
24862 EVT::getVectorVT(*DAG.getContext(), SVT, 128 / SVT.getSizeInBits());
24863 SDValue Ex = ExtendVecSize(DL, N0, Scale * InVT.getSizeInBits());
24864 SDValue SExt = DAG.getNode(ISD::SIGN_EXTEND, DL, ExVT, Ex);
24865 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, VT, SExt,
24866 DAG.getIntPtrConstant(0, DL));
24869 // If target-size is 128-bits, then convert to ISD::SIGN_EXTEND_VECTOR_INREG
24870 // which ensures lowering to X86ISD::VSEXT (pmovsx*).
24871 if (VT.getSizeInBits() == 128 &&
24872 (SVT == MVT::i64 || SVT == MVT::i32 || SVT == MVT::i16) &&
24873 (InSVT == MVT::i32 || InSVT == MVT::i16 || InSVT == MVT::i8)) {
24874 SDValue ExOp = ExtendVecSize(DL, N0, 128);
24875 return DAG.getSignExtendVectorInReg(ExOp, DL, VT);
24878 // On pre-AVX2 targets, split into 128-bit nodes of
24879 // ISD::SIGN_EXTEND_VECTOR_INREG.
24880 if (!Subtarget->hasInt256() && !(VT.getSizeInBits() % 128) &&
24881 (SVT == MVT::i64 || SVT == MVT::i32 || SVT == MVT::i16) &&
24882 (InSVT == MVT::i32 || InSVT == MVT::i16 || InSVT == MVT::i8)) {
24883 unsigned NumVecs = VT.getSizeInBits() / 128;
24884 unsigned NumSubElts = 128 / SVT.getSizeInBits();
24885 EVT SubVT = EVT::getVectorVT(*DAG.getContext(), SVT, NumSubElts);
24886 EVT InSubVT = EVT::getVectorVT(*DAG.getContext(), InSVT, NumSubElts);
24888 SmallVector<SDValue, 8> Opnds;
24889 for (unsigned i = 0, Offset = 0; i != NumVecs;
24890 ++i, Offset += NumSubElts) {
24891 SDValue SrcVec = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, InSubVT, N0,
24892 DAG.getIntPtrConstant(Offset, DL));
24893 SrcVec = ExtendVecSize(DL, SrcVec, 128);
24894 SrcVec = DAG.getSignExtendVectorInReg(SrcVec, DL, SubVT);
24895 Opnds.push_back(SrcVec);
24897 return DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, Opnds);
24901 if (!Subtarget->hasFp256())
24904 if (VT.isVector() && VT.getSizeInBits() == 256)
24905 if (SDValue R = WidenMaskArithmetic(N, DAG, DCI, Subtarget))
24911 static SDValue PerformFMACombine(SDNode *N, SelectionDAG &DAG,
24912 const X86Subtarget* Subtarget) {
24914 EVT VT = N->getValueType(0);
24916 // Let legalize expand this if it isn't a legal type yet.
24917 if (!DAG.getTargetLoweringInfo().isTypeLegal(VT))
24920 EVT ScalarVT = VT.getScalarType();
24921 if ((ScalarVT != MVT::f32 && ScalarVT != MVT::f64) ||
24922 (!Subtarget->hasFMA() && !Subtarget->hasFMA4() &&
24923 !Subtarget->hasAVX512()))
24926 SDValue A = N->getOperand(0);
24927 SDValue B = N->getOperand(1);
24928 SDValue C = N->getOperand(2);
24930 bool NegA = (A.getOpcode() == ISD::FNEG);
24931 bool NegB = (B.getOpcode() == ISD::FNEG);
24932 bool NegC = (C.getOpcode() == ISD::FNEG);
24934 // Negative multiplication when NegA xor NegB
24935 bool NegMul = (NegA != NegB);
24937 A = A.getOperand(0);
24939 B = B.getOperand(0);
24941 C = C.getOperand(0);
24945 Opcode = (!NegC) ? X86ISD::FMADD : X86ISD::FMSUB;
24947 Opcode = (!NegC) ? X86ISD::FNMADD : X86ISD::FNMSUB;
24949 return DAG.getNode(Opcode, dl, VT, A, B, C);
24952 static SDValue PerformZExtCombine(SDNode *N, SelectionDAG &DAG,
24953 TargetLowering::DAGCombinerInfo &DCI,
24954 const X86Subtarget *Subtarget) {
24955 // (i32 zext (and (i8 x86isd::setcc_carry), 1)) ->
24956 // (and (i32 x86isd::setcc_carry), 1)
24957 // This eliminates the zext. This transformation is necessary because
24958 // ISD::SETCC is always legalized to i8.
24960 SDValue N0 = N->getOperand(0);
24961 EVT VT = N->getValueType(0);
24963 if (N0.getOpcode() == ISD::AND &&
24965 N0.getOperand(0).hasOneUse()) {
24966 SDValue N00 = N0.getOperand(0);
24967 if (N00.getOpcode() == X86ISD::SETCC_CARRY) {
24968 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
24969 if (!C || C->getZExtValue() != 1)
24971 return DAG.getNode(ISD::AND, dl, VT,
24972 DAG.getNode(X86ISD::SETCC_CARRY, dl, VT,
24973 N00.getOperand(0), N00.getOperand(1)),
24974 DAG.getConstant(1, dl, VT));
24978 if (N0.getOpcode() == ISD::TRUNCATE &&
24980 N0.getOperand(0).hasOneUse()) {
24981 SDValue N00 = N0.getOperand(0);
24982 if (N00.getOpcode() == X86ISD::SETCC_CARRY) {
24983 return DAG.getNode(ISD::AND, dl, VT,
24984 DAG.getNode(X86ISD::SETCC_CARRY, dl, VT,
24985 N00.getOperand(0), N00.getOperand(1)),
24986 DAG.getConstant(1, dl, VT));
24990 if (VT.is256BitVector())
24991 if (SDValue R = WidenMaskArithmetic(N, DAG, DCI, Subtarget))
24994 // (i8,i32 zext (udivrem (i8 x, i8 y)) ->
24995 // (i8,i32 (udivrem_zext_hreg (i8 x, i8 y)
24996 // This exposes the zext to the udivrem lowering, so that it directly extends
24997 // from AH (which we otherwise need to do contortions to access).
24998 if (N0.getOpcode() == ISD::UDIVREM &&
24999 N0.getResNo() == 1 && N0.getValueType() == MVT::i8 &&
25000 (VT == MVT::i32 || VT == MVT::i64)) {
25001 SDVTList NodeTys = DAG.getVTList(MVT::i8, VT);
25002 SDValue R = DAG.getNode(X86ISD::UDIVREM8_ZEXT_HREG, dl, NodeTys,
25003 N0.getOperand(0), N0.getOperand(1));
25004 DAG.ReplaceAllUsesOfValueWith(N0.getValue(0), R.getValue(0));
25005 return R.getValue(1);
25011 // Optimize x == -y --> x+y == 0
25012 // x != -y --> x+y != 0
25013 static SDValue PerformISDSETCCCombine(SDNode *N, SelectionDAG &DAG,
25014 const X86Subtarget* Subtarget) {
25015 ISD::CondCode CC = cast<CondCodeSDNode>(N->getOperand(2))->get();
25016 SDValue LHS = N->getOperand(0);
25017 SDValue RHS = N->getOperand(1);
25018 EVT VT = N->getValueType(0);
25021 if ((CC == ISD::SETNE || CC == ISD::SETEQ) && LHS.getOpcode() == ISD::SUB)
25022 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(LHS.getOperand(0)))
25023 if (C->getAPIntValue() == 0 && LHS.hasOneUse()) {
25024 SDValue addV = DAG.getNode(ISD::ADD, DL, LHS.getValueType(), RHS,
25025 LHS.getOperand(1));
25026 return DAG.getSetCC(DL, N->getValueType(0), addV,
25027 DAG.getConstant(0, DL, addV.getValueType()), CC);
25029 if ((CC == ISD::SETNE || CC == ISD::SETEQ) && RHS.getOpcode() == ISD::SUB)
25030 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(RHS.getOperand(0)))
25031 if (C->getAPIntValue() == 0 && RHS.hasOneUse()) {
25032 SDValue addV = DAG.getNode(ISD::ADD, DL, RHS.getValueType(), LHS,
25033 RHS.getOperand(1));
25034 return DAG.getSetCC(DL, N->getValueType(0), addV,
25035 DAG.getConstant(0, DL, addV.getValueType()), CC);
25038 if (VT.getScalarType() == MVT::i1 &&
25039 (CC == ISD::SETNE || CC == ISD::SETEQ || ISD::isSignedIntSetCC(CC))) {
25041 (LHS.getOpcode() == ISD::SIGN_EXTEND) &&
25042 (LHS.getOperand(0).getValueType().getScalarType() == MVT::i1);
25043 bool IsVZero1 = ISD::isBuildVectorAllZeros(RHS.getNode());
25045 if (!IsSEXT0 || !IsVZero1) {
25046 // Swap the operands and update the condition code.
25047 std::swap(LHS, RHS);
25048 CC = ISD::getSetCCSwappedOperands(CC);
25050 IsSEXT0 = (LHS.getOpcode() == ISD::SIGN_EXTEND) &&
25051 (LHS.getOperand(0).getValueType().getScalarType() == MVT::i1);
25052 IsVZero1 = ISD::isBuildVectorAllZeros(RHS.getNode());
25055 if (IsSEXT0 && IsVZero1) {
25056 assert(VT == LHS.getOperand(0).getValueType() &&
25057 "Uexpected operand type");
25058 if (CC == ISD::SETGT)
25059 return DAG.getConstant(0, DL, VT);
25060 if (CC == ISD::SETLE)
25061 return DAG.getConstant(1, DL, VT);
25062 if (CC == ISD::SETEQ || CC == ISD::SETGE)
25063 return DAG.getNOT(DL, LHS.getOperand(0), VT);
25065 assert((CC == ISD::SETNE || CC == ISD::SETLT) &&
25066 "Unexpected condition code!");
25067 return LHS.getOperand(0);
25074 static SDValue NarrowVectorLoadToElement(LoadSDNode *Load, unsigned Index,
25075 SelectionDAG &DAG) {
25077 MVT VT = Load->getSimpleValueType(0);
25078 MVT EVT = VT.getVectorElementType();
25079 SDValue Addr = Load->getOperand(1);
25080 SDValue NewAddr = DAG.getNode(
25081 ISD::ADD, dl, Addr.getSimpleValueType(), Addr,
25082 DAG.getConstant(Index * EVT.getStoreSize(), dl,
25083 Addr.getSimpleValueType()));
25086 DAG.getLoad(EVT, dl, Load->getChain(), NewAddr,
25087 DAG.getMachineFunction().getMachineMemOperand(
25088 Load->getMemOperand(), 0, EVT.getStoreSize()));
25092 static SDValue PerformINSERTPSCombine(SDNode *N, SelectionDAG &DAG,
25093 const X86Subtarget *Subtarget) {
25095 MVT VT = N->getOperand(1)->getSimpleValueType(0);
25096 assert((VT == MVT::v4f32 || VT == MVT::v4i32) &&
25097 "X86insertps is only defined for v4x32");
25099 SDValue Ld = N->getOperand(1);
25100 if (MayFoldLoad(Ld)) {
25101 // Extract the countS bits from the immediate so we can get the proper
25102 // address when narrowing the vector load to a specific element.
25103 // When the second source op is a memory address, insertps doesn't use
25104 // countS and just gets an f32 from that address.
25105 unsigned DestIndex =
25106 cast<ConstantSDNode>(N->getOperand(2))->getZExtValue() >> 6;
25108 Ld = NarrowVectorLoadToElement(cast<LoadSDNode>(Ld), DestIndex, DAG);
25110 // Create this as a scalar to vector to match the instruction pattern.
25111 SDValue LoadScalarToVector = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Ld);
25112 // countS bits are ignored when loading from memory on insertps, which
25113 // means we don't need to explicitly set them to 0.
25114 return DAG.getNode(X86ISD::INSERTPS, dl, VT, N->getOperand(0),
25115 LoadScalarToVector, N->getOperand(2));
25120 static SDValue PerformBLENDICombine(SDNode *N, SelectionDAG &DAG) {
25121 SDValue V0 = N->getOperand(0);
25122 SDValue V1 = N->getOperand(1);
25124 EVT VT = N->getValueType(0);
25126 // Canonicalize a v2f64 blend with a mask of 2 by swapping the vector
25127 // operands and changing the mask to 1. This saves us a bunch of
25128 // pattern-matching possibilities related to scalar math ops in SSE/AVX.
25129 // x86InstrInfo knows how to commute this back after instruction selection
25130 // if it would help register allocation.
25132 // TODO: If optimizing for size or a processor that doesn't suffer from
25133 // partial register update stalls, this should be transformed into a MOVSD
25134 // instruction because a MOVSD is 1-2 bytes smaller than a BLENDPD.
25136 if (VT == MVT::v2f64)
25137 if (auto *Mask = dyn_cast<ConstantSDNode>(N->getOperand(2)))
25138 if (Mask->getZExtValue() == 2 && !isShuffleFoldableLoad(V0)) {
25139 SDValue NewMask = DAG.getConstant(1, DL, MVT::i8);
25140 return DAG.getNode(X86ISD::BLENDI, DL, VT, V1, V0, NewMask);
25146 // Helper function of PerformSETCCCombine. It is to materialize "setb reg"
25147 // as "sbb reg,reg", since it can be extended without zext and produces
25148 // an all-ones bit which is more useful than 0/1 in some cases.
25149 static SDValue MaterializeSETB(SDLoc DL, SDValue EFLAGS, SelectionDAG &DAG,
25152 return DAG.getNode(ISD::AND, DL, VT,
25153 DAG.getNode(X86ISD::SETCC_CARRY, DL, MVT::i8,
25154 DAG.getConstant(X86::COND_B, DL, MVT::i8),
25156 DAG.getConstant(1, DL, VT));
25157 assert (VT == MVT::i1 && "Unexpected type for SECCC node");
25158 return DAG.getNode(ISD::TRUNCATE, DL, MVT::i1,
25159 DAG.getNode(X86ISD::SETCC_CARRY, DL, MVT::i8,
25160 DAG.getConstant(X86::COND_B, DL, MVT::i8),
25164 // Optimize RES = X86ISD::SETCC CONDCODE, EFLAG_INPUT
25165 static SDValue PerformSETCCCombine(SDNode *N, SelectionDAG &DAG,
25166 TargetLowering::DAGCombinerInfo &DCI,
25167 const X86Subtarget *Subtarget) {
25169 X86::CondCode CC = X86::CondCode(N->getConstantOperandVal(0));
25170 SDValue EFLAGS = N->getOperand(1);
25172 if (CC == X86::COND_A) {
25173 // Try to convert COND_A into COND_B in an attempt to facilitate
25174 // materializing "setb reg".
25176 // Do not flip "e > c", where "c" is a constant, because Cmp instruction
25177 // cannot take an immediate as its first operand.
25179 if (EFLAGS.getOpcode() == X86ISD::SUB && EFLAGS.hasOneUse() &&
25180 EFLAGS.getValueType().isInteger() &&
25181 !isa<ConstantSDNode>(EFLAGS.getOperand(1))) {
25182 SDValue NewSub = DAG.getNode(X86ISD::SUB, SDLoc(EFLAGS),
25183 EFLAGS.getNode()->getVTList(),
25184 EFLAGS.getOperand(1), EFLAGS.getOperand(0));
25185 SDValue NewEFLAGS = SDValue(NewSub.getNode(), EFLAGS.getResNo());
25186 return MaterializeSETB(DL, NewEFLAGS, DAG, N->getSimpleValueType(0));
25190 // Materialize "setb reg" as "sbb reg,reg", since it can be extended without
25191 // a zext and produces an all-ones bit which is more useful than 0/1 in some
25193 if (CC == X86::COND_B)
25194 return MaterializeSETB(DL, EFLAGS, DAG, N->getSimpleValueType(0));
25196 if (SDValue Flags = checkBoolTestSetCCCombine(EFLAGS, CC)) {
25197 SDValue Cond = DAG.getConstant(CC, DL, MVT::i8);
25198 return DAG.getNode(X86ISD::SETCC, DL, N->getVTList(), Cond, Flags);
25204 // Optimize branch condition evaluation.
25206 static SDValue PerformBrCondCombine(SDNode *N, SelectionDAG &DAG,
25207 TargetLowering::DAGCombinerInfo &DCI,
25208 const X86Subtarget *Subtarget) {
25210 SDValue Chain = N->getOperand(0);
25211 SDValue Dest = N->getOperand(1);
25212 SDValue EFLAGS = N->getOperand(3);
25213 X86::CondCode CC = X86::CondCode(N->getConstantOperandVal(2));
25215 if (SDValue Flags = checkBoolTestSetCCCombine(EFLAGS, CC)) {
25216 SDValue Cond = DAG.getConstant(CC, DL, MVT::i8);
25217 return DAG.getNode(X86ISD::BRCOND, DL, N->getVTList(), Chain, Dest, Cond,
25224 static SDValue performVectorCompareAndMaskUnaryOpCombine(SDNode *N,
25225 SelectionDAG &DAG) {
25226 // Take advantage of vector comparisons producing 0 or -1 in each lane to
25227 // optimize away operation when it's from a constant.
25229 // The general transformation is:
25230 // UNARYOP(AND(VECTOR_CMP(x,y), constant)) -->
25231 // AND(VECTOR_CMP(x,y), constant2)
25232 // constant2 = UNARYOP(constant)
25234 // Early exit if this isn't a vector operation, the operand of the
25235 // unary operation isn't a bitwise AND, or if the sizes of the operations
25236 // aren't the same.
25237 EVT VT = N->getValueType(0);
25238 if (!VT.isVector() || N->getOperand(0)->getOpcode() != ISD::AND ||
25239 N->getOperand(0)->getOperand(0)->getOpcode() != ISD::SETCC ||
25240 VT.getSizeInBits() != N->getOperand(0)->getValueType(0).getSizeInBits())
25243 // Now check that the other operand of the AND is a constant. We could
25244 // make the transformation for non-constant splats as well, but it's unclear
25245 // that would be a benefit as it would not eliminate any operations, just
25246 // perform one more step in scalar code before moving to the vector unit.
25247 if (BuildVectorSDNode *BV =
25248 dyn_cast<BuildVectorSDNode>(N->getOperand(0)->getOperand(1))) {
25249 // Bail out if the vector isn't a constant.
25250 if (!BV->isConstant())
25253 // Everything checks out. Build up the new and improved node.
25255 EVT IntVT = BV->getValueType(0);
25256 // Create a new constant of the appropriate type for the transformed
25258 SDValue SourceConst = DAG.getNode(N->getOpcode(), DL, VT, SDValue(BV, 0));
25259 // The AND node needs bitcasts to/from an integer vector type around it.
25260 SDValue MaskConst = DAG.getBitcast(IntVT, SourceConst);
25261 SDValue NewAnd = DAG.getNode(ISD::AND, DL, IntVT,
25262 N->getOperand(0)->getOperand(0), MaskConst);
25263 SDValue Res = DAG.getBitcast(VT, NewAnd);
25270 static SDValue PerformUINT_TO_FPCombine(SDNode *N, SelectionDAG &DAG,
25271 const X86Subtarget *Subtarget) {
25272 SDValue Op0 = N->getOperand(0);
25273 EVT VT = N->getValueType(0);
25274 EVT InVT = Op0.getValueType();
25275 EVT InSVT = InVT.getScalarType();
25276 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
25278 // UINT_TO_FP(vXi8) -> SINT_TO_FP(ZEXT(vXi8 to vXi32))
25279 // UINT_TO_FP(vXi16) -> SINT_TO_FP(ZEXT(vXi16 to vXi32))
25280 if (InVT.isVector() && (InSVT == MVT::i8 || InSVT == MVT::i16)) {
25282 EVT DstVT = EVT::getVectorVT(*DAG.getContext(), MVT::i32,
25283 InVT.getVectorNumElements());
25284 SDValue P = DAG.getNode(ISD::ZERO_EXTEND, dl, DstVT, Op0);
25286 if (TLI.isOperationLegal(ISD::UINT_TO_FP, DstVT))
25287 return DAG.getNode(ISD::UINT_TO_FP, dl, VT, P);
25289 return DAG.getNode(ISD::SINT_TO_FP, dl, VT, P);
25295 static SDValue PerformSINT_TO_FPCombine(SDNode *N, SelectionDAG &DAG,
25296 const X86Subtarget *Subtarget) {
25297 // First try to optimize away the conversion entirely when it's
25298 // conditionally from a constant. Vectors only.
25299 if (SDValue Res = performVectorCompareAndMaskUnaryOpCombine(N, DAG))
25302 // Now move on to more general possibilities.
25303 SDValue Op0 = N->getOperand(0);
25304 EVT VT = N->getValueType(0);
25305 EVT InVT = Op0.getValueType();
25306 EVT InSVT = InVT.getScalarType();
25308 // SINT_TO_FP(vXi8) -> SINT_TO_FP(SEXT(vXi8 to vXi32))
25309 // SINT_TO_FP(vXi16) -> SINT_TO_FP(SEXT(vXi16 to vXi32))
25310 if (InVT.isVector() && (InSVT == MVT::i8 || InSVT == MVT::i16)) {
25312 EVT DstVT = EVT::getVectorVT(*DAG.getContext(), MVT::i32,
25313 InVT.getVectorNumElements());
25314 SDValue P = DAG.getNode(ISD::SIGN_EXTEND, dl, DstVT, Op0);
25315 return DAG.getNode(ISD::SINT_TO_FP, dl, VT, P);
25318 // Transform (SINT_TO_FP (i64 ...)) into an x87 operation if we have
25319 // a 32-bit target where SSE doesn't support i64->FP operations.
25320 if (Op0.getOpcode() == ISD::LOAD) {
25321 LoadSDNode *Ld = cast<LoadSDNode>(Op0.getNode());
25322 EVT LdVT = Ld->getValueType(0);
25324 // This transformation is not supported if the result type is f16
25325 if (VT == MVT::f16)
25328 if (!Ld->isVolatile() && !VT.isVector() &&
25329 ISD::isNON_EXTLoad(Op0.getNode()) && Op0.hasOneUse() &&
25330 !Subtarget->is64Bit() && LdVT == MVT::i64) {
25331 SDValue FILDChain = Subtarget->getTargetLowering()->BuildFILD(
25332 SDValue(N, 0), LdVT, Ld->getChain(), Op0, DAG);
25333 DAG.ReplaceAllUsesOfValueWith(Op0.getValue(1), FILDChain.getValue(1));
25340 // Optimize RES, EFLAGS = X86ISD::ADC LHS, RHS, EFLAGS
25341 static SDValue PerformADCCombine(SDNode *N, SelectionDAG &DAG,
25342 X86TargetLowering::DAGCombinerInfo &DCI) {
25343 // If the LHS and RHS of the ADC node are zero, then it can't overflow and
25344 // the result is either zero or one (depending on the input carry bit).
25345 // Strength reduce this down to a "set on carry" aka SETCC_CARRY&1.
25346 if (X86::isZeroNode(N->getOperand(0)) &&
25347 X86::isZeroNode(N->getOperand(1)) &&
25348 // We don't have a good way to replace an EFLAGS use, so only do this when
25350 SDValue(N, 1).use_empty()) {
25352 EVT VT = N->getValueType(0);
25353 SDValue CarryOut = DAG.getConstant(0, DL, N->getValueType(1));
25354 SDValue Res1 = DAG.getNode(ISD::AND, DL, VT,
25355 DAG.getNode(X86ISD::SETCC_CARRY, DL, VT,
25356 DAG.getConstant(X86::COND_B, DL,
25359 DAG.getConstant(1, DL, VT));
25360 return DCI.CombineTo(N, Res1, CarryOut);
25366 // fold (add Y, (sete X, 0)) -> adc 0, Y
25367 // (add Y, (setne X, 0)) -> sbb -1, Y
25368 // (sub (sete X, 0), Y) -> sbb 0, Y
25369 // (sub (setne X, 0), Y) -> adc -1, Y
25370 static SDValue OptimizeConditionalInDecrement(SDNode *N, SelectionDAG &DAG) {
25373 // Look through ZExts.
25374 SDValue Ext = N->getOperand(N->getOpcode() == ISD::SUB ? 1 : 0);
25375 if (Ext.getOpcode() != ISD::ZERO_EXTEND || !Ext.hasOneUse())
25378 SDValue SetCC = Ext.getOperand(0);
25379 if (SetCC.getOpcode() != X86ISD::SETCC || !SetCC.hasOneUse())
25382 X86::CondCode CC = (X86::CondCode)SetCC.getConstantOperandVal(0);
25383 if (CC != X86::COND_E && CC != X86::COND_NE)
25386 SDValue Cmp = SetCC.getOperand(1);
25387 if (Cmp.getOpcode() != X86ISD::CMP || !Cmp.hasOneUse() ||
25388 !X86::isZeroNode(Cmp.getOperand(1)) ||
25389 !Cmp.getOperand(0).getValueType().isInteger())
25392 SDValue CmpOp0 = Cmp.getOperand(0);
25393 SDValue NewCmp = DAG.getNode(X86ISD::CMP, DL, MVT::i32, CmpOp0,
25394 DAG.getConstant(1, DL, CmpOp0.getValueType()));
25396 SDValue OtherVal = N->getOperand(N->getOpcode() == ISD::SUB ? 0 : 1);
25397 if (CC == X86::COND_NE)
25398 return DAG.getNode(N->getOpcode() == ISD::SUB ? X86ISD::ADC : X86ISD::SBB,
25399 DL, OtherVal.getValueType(), OtherVal,
25400 DAG.getConstant(-1ULL, DL, OtherVal.getValueType()),
25402 return DAG.getNode(N->getOpcode() == ISD::SUB ? X86ISD::SBB : X86ISD::ADC,
25403 DL, OtherVal.getValueType(), OtherVal,
25404 DAG.getConstant(0, DL, OtherVal.getValueType()), NewCmp);
25407 /// PerformADDCombine - Do target-specific dag combines on integer adds.
25408 static SDValue PerformAddCombine(SDNode *N, SelectionDAG &DAG,
25409 const X86Subtarget *Subtarget) {
25410 EVT VT = N->getValueType(0);
25411 SDValue Op0 = N->getOperand(0);
25412 SDValue Op1 = N->getOperand(1);
25414 // Try to synthesize horizontal adds from adds of shuffles.
25415 if (((Subtarget->hasSSSE3() && (VT == MVT::v8i16 || VT == MVT::v4i32)) ||
25416 (Subtarget->hasInt256() && (VT == MVT::v16i16 || VT == MVT::v8i32))) &&
25417 isHorizontalBinOp(Op0, Op1, true))
25418 return DAG.getNode(X86ISD::HADD, SDLoc(N), VT, Op0, Op1);
25420 return OptimizeConditionalInDecrement(N, DAG);
25423 static SDValue PerformSubCombine(SDNode *N, SelectionDAG &DAG,
25424 const X86Subtarget *Subtarget) {
25425 SDValue Op0 = N->getOperand(0);
25426 SDValue Op1 = N->getOperand(1);
25428 // X86 can't encode an immediate LHS of a sub. See if we can push the
25429 // negation into a preceding instruction.
25430 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op0)) {
25431 // If the RHS of the sub is a XOR with one use and a constant, invert the
25432 // immediate. Then add one to the LHS of the sub so we can turn
25433 // X-Y -> X+~Y+1, saving one register.
25434 if (Op1->hasOneUse() && Op1.getOpcode() == ISD::XOR &&
25435 isa<ConstantSDNode>(Op1.getOperand(1))) {
25436 APInt XorC = cast<ConstantSDNode>(Op1.getOperand(1))->getAPIntValue();
25437 EVT VT = Op0.getValueType();
25438 SDValue NewXor = DAG.getNode(ISD::XOR, SDLoc(Op1), VT,
25440 DAG.getConstant(~XorC, SDLoc(Op1), VT));
25441 return DAG.getNode(ISD::ADD, SDLoc(N), VT, NewXor,
25442 DAG.getConstant(C->getAPIntValue() + 1, SDLoc(N), VT));
25446 // Try to synthesize horizontal adds from adds of shuffles.
25447 EVT VT = N->getValueType(0);
25448 if (((Subtarget->hasSSSE3() && (VT == MVT::v8i16 || VT == MVT::v4i32)) ||
25449 (Subtarget->hasInt256() && (VT == MVT::v16i16 || VT == MVT::v8i32))) &&
25450 isHorizontalBinOp(Op0, Op1, true))
25451 return DAG.getNode(X86ISD::HSUB, SDLoc(N), VT, Op0, Op1);
25453 return OptimizeConditionalInDecrement(N, DAG);
25456 /// performVZEXTCombine - Performs build vector combines
25457 static SDValue performVZEXTCombine(SDNode *N, SelectionDAG &DAG,
25458 TargetLowering::DAGCombinerInfo &DCI,
25459 const X86Subtarget *Subtarget) {
25461 MVT VT = N->getSimpleValueType(0);
25462 SDValue Op = N->getOperand(0);
25463 MVT OpVT = Op.getSimpleValueType();
25464 MVT OpEltVT = OpVT.getVectorElementType();
25465 unsigned InputBits = OpEltVT.getSizeInBits() * VT.getVectorNumElements();
25467 // (vzext (bitcast (vzext (x)) -> (vzext x)
25469 while (V.getOpcode() == ISD::BITCAST)
25470 V = V.getOperand(0);
25472 if (V != Op && V.getOpcode() == X86ISD::VZEXT) {
25473 MVT InnerVT = V.getSimpleValueType();
25474 MVT InnerEltVT = InnerVT.getVectorElementType();
25476 // If the element sizes match exactly, we can just do one larger vzext. This
25477 // is always an exact type match as vzext operates on integer types.
25478 if (OpEltVT == InnerEltVT) {
25479 assert(OpVT == InnerVT && "Types must match for vzext!");
25480 return DAG.getNode(X86ISD::VZEXT, DL, VT, V.getOperand(0));
25483 // The only other way we can combine them is if only a single element of the
25484 // inner vzext is used in the input to the outer vzext.
25485 if (InnerEltVT.getSizeInBits() < InputBits)
25488 // In this case, the inner vzext is completely dead because we're going to
25489 // only look at bits inside of the low element. Just do the outer vzext on
25490 // a bitcast of the input to the inner.
25491 return DAG.getNode(X86ISD::VZEXT, DL, VT, DAG.getBitcast(OpVT, V));
25494 // Check if we can bypass extracting and re-inserting an element of an input
25495 // vector. Essentialy:
25496 // (bitcast (sclr2vec (ext_vec_elt x))) -> (bitcast x)
25497 if (V.getOpcode() == ISD::SCALAR_TO_VECTOR &&
25498 V.getOperand(0).getOpcode() == ISD::EXTRACT_VECTOR_ELT &&
25499 V.getOperand(0).getSimpleValueType().getSizeInBits() == InputBits) {
25500 SDValue ExtractedV = V.getOperand(0);
25501 SDValue OrigV = ExtractedV.getOperand(0);
25502 if (auto *ExtractIdx = dyn_cast<ConstantSDNode>(ExtractedV.getOperand(1)))
25503 if (ExtractIdx->getZExtValue() == 0) {
25504 MVT OrigVT = OrigV.getSimpleValueType();
25505 // Extract a subvector if necessary...
25506 if (OrigVT.getSizeInBits() > OpVT.getSizeInBits()) {
25507 int Ratio = OrigVT.getSizeInBits() / OpVT.getSizeInBits();
25508 OrigVT = MVT::getVectorVT(OrigVT.getVectorElementType(),
25509 OrigVT.getVectorNumElements() / Ratio);
25510 OrigV = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, OrigVT, OrigV,
25511 DAG.getIntPtrConstant(0, DL));
25513 Op = DAG.getBitcast(OpVT, OrigV);
25514 return DAG.getNode(X86ISD::VZEXT, DL, VT, Op);
25521 SDValue X86TargetLowering::PerformDAGCombine(SDNode *N,
25522 DAGCombinerInfo &DCI) const {
25523 SelectionDAG &DAG = DCI.DAG;
25524 switch (N->getOpcode()) {
25526 case ISD::EXTRACT_VECTOR_ELT:
25527 return PerformEXTRACT_VECTOR_ELTCombine(N, DAG, DCI);
25530 case X86ISD::SHRUNKBLEND:
25531 return PerformSELECTCombine(N, DAG, DCI, Subtarget);
25532 case ISD::BITCAST: return PerformBITCASTCombine(N, DAG);
25533 case X86ISD::CMOV: return PerformCMOVCombine(N, DAG, DCI, Subtarget);
25534 case ISD::ADD: return PerformAddCombine(N, DAG, Subtarget);
25535 case ISD::SUB: return PerformSubCombine(N, DAG, Subtarget);
25536 case X86ISD::ADC: return PerformADCCombine(N, DAG, DCI);
25537 case ISD::MUL: return PerformMulCombine(N, DAG, DCI);
25540 case ISD::SRL: return PerformShiftCombine(N, DAG, DCI, Subtarget);
25541 case ISD::AND: return PerformAndCombine(N, DAG, DCI, Subtarget);
25542 case ISD::OR: return PerformOrCombine(N, DAG, DCI, Subtarget);
25543 case ISD::XOR: return PerformXorCombine(N, DAG, DCI, Subtarget);
25544 case ISD::LOAD: return PerformLOADCombine(N, DAG, DCI, Subtarget);
25545 case ISD::MLOAD: return PerformMLOADCombine(N, DAG, DCI, Subtarget);
25546 case ISD::STORE: return PerformSTORECombine(N, DAG, Subtarget);
25547 case ISD::MSTORE: return PerformMSTORECombine(N, DAG, Subtarget);
25548 case ISD::SINT_TO_FP: return PerformSINT_TO_FPCombine(N, DAG, Subtarget);
25549 case ISD::UINT_TO_FP: return PerformUINT_TO_FPCombine(N, DAG, Subtarget);
25550 case ISD::FADD: return PerformFADDCombine(N, DAG, Subtarget);
25551 case ISD::FSUB: return PerformFSUBCombine(N, DAG, Subtarget);
25553 case X86ISD::FOR: return PerformFORCombine(N, DAG);
25555 case X86ISD::FMAX: return PerformFMinFMaxCombine(N, DAG);
25556 case X86ISD::FAND: return PerformFANDCombine(N, DAG);
25557 case X86ISD::FANDN: return PerformFANDNCombine(N, DAG);
25558 case X86ISD::BT: return PerformBTCombine(N, DAG, DCI);
25559 case X86ISD::VZEXT_MOVL: return PerformVZEXT_MOVLCombine(N, DAG);
25560 case ISD::ANY_EXTEND:
25561 case ISD::ZERO_EXTEND: return PerformZExtCombine(N, DAG, DCI, Subtarget);
25562 case ISD::SIGN_EXTEND: return PerformSExtCombine(N, DAG, DCI, Subtarget);
25563 case ISD::SIGN_EXTEND_INREG:
25564 return PerformSIGN_EXTEND_INREGCombine(N, DAG, Subtarget);
25565 case ISD::SETCC: return PerformISDSETCCCombine(N, DAG, Subtarget);
25566 case X86ISD::SETCC: return PerformSETCCCombine(N, DAG, DCI, Subtarget);
25567 case X86ISD::BRCOND: return PerformBrCondCombine(N, DAG, DCI, Subtarget);
25568 case X86ISD::VZEXT: return performVZEXTCombine(N, DAG, DCI, Subtarget);
25569 case X86ISD::SHUFP: // Handle all target specific shuffles
25570 case X86ISD::PALIGNR:
25571 case X86ISD::UNPCKH:
25572 case X86ISD::UNPCKL:
25573 case X86ISD::MOVHLPS:
25574 case X86ISD::MOVLHPS:
25575 case X86ISD::PSHUFB:
25576 case X86ISD::PSHUFD:
25577 case X86ISD::PSHUFHW:
25578 case X86ISD::PSHUFLW:
25579 case X86ISD::MOVSS:
25580 case X86ISD::MOVSD:
25581 case X86ISD::VPERMILPI:
25582 case X86ISD::VPERM2X128:
25583 case ISD::VECTOR_SHUFFLE: return PerformShuffleCombine(N, DAG, DCI,Subtarget);
25584 case ISD::FMA: return PerformFMACombine(N, DAG, Subtarget);
25585 case ISD::INTRINSIC_WO_CHAIN:
25586 return PerformINTRINSIC_WO_CHAINCombine(N, DAG, Subtarget);
25587 case X86ISD::INSERTPS: {
25588 if (getTargetMachine().getOptLevel() > CodeGenOpt::None)
25589 return PerformINSERTPSCombine(N, DAG, Subtarget);
25592 case X86ISD::BLENDI: return PerformBLENDICombine(N, DAG);
25598 /// isTypeDesirableForOp - Return true if the target has native support for
25599 /// the specified value type and it is 'desirable' to use the type for the
25600 /// given node type. e.g. On x86 i16 is legal, but undesirable since i16
25601 /// instruction encodings are longer and some i16 instructions are slow.
25602 bool X86TargetLowering::isTypeDesirableForOp(unsigned Opc, EVT VT) const {
25603 if (!isTypeLegal(VT))
25605 if (VT != MVT::i16)
25612 case ISD::SIGN_EXTEND:
25613 case ISD::ZERO_EXTEND:
25614 case ISD::ANY_EXTEND:
25627 /// IsDesirableToPromoteOp - This method query the target whether it is
25628 /// beneficial for dag combiner to promote the specified node. If true, it
25629 /// should return the desired promotion type by reference.
25630 bool X86TargetLowering::IsDesirableToPromoteOp(SDValue Op, EVT &PVT) const {
25631 EVT VT = Op.getValueType();
25632 if (VT != MVT::i16)
25635 bool Promote = false;
25636 bool Commute = false;
25637 switch (Op.getOpcode()) {
25640 LoadSDNode *LD = cast<LoadSDNode>(Op);
25641 // If the non-extending load has a single use and it's not live out, then it
25642 // might be folded.
25643 if (LD->getExtensionType() == ISD::NON_EXTLOAD /*&&
25644 Op.hasOneUse()*/) {
25645 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
25646 UE = Op.getNode()->use_end(); UI != UE; ++UI) {
25647 // The only case where we'd want to promote LOAD (rather then it being
25648 // promoted as an operand is when it's only use is liveout.
25649 if (UI->getOpcode() != ISD::CopyToReg)
25656 case ISD::SIGN_EXTEND:
25657 case ISD::ZERO_EXTEND:
25658 case ISD::ANY_EXTEND:
25663 SDValue N0 = Op.getOperand(0);
25664 // Look out for (store (shl (load), x)).
25665 if (MayFoldLoad(N0) && MayFoldIntoStore(Op))
25678 SDValue N0 = Op.getOperand(0);
25679 SDValue N1 = Op.getOperand(1);
25680 if (!Commute && MayFoldLoad(N1))
25682 // Avoid disabling potential load folding opportunities.
25683 if (MayFoldLoad(N0) && (!isa<ConstantSDNode>(N1) || MayFoldIntoStore(Op)))
25685 if (MayFoldLoad(N1) && (!isa<ConstantSDNode>(N0) || MayFoldIntoStore(Op)))
25695 //===----------------------------------------------------------------------===//
25696 // X86 Inline Assembly Support
25697 //===----------------------------------------------------------------------===//
25699 // Helper to match a string separated by whitespace.
25700 static bool matchAsm(StringRef S, ArrayRef<const char *> Pieces) {
25701 S = S.substr(S.find_first_not_of(" \t")); // Skip leading whitespace.
25703 for (StringRef Piece : Pieces) {
25704 if (!S.startswith(Piece)) // Check if the piece matches.
25707 S = S.substr(Piece.size());
25708 StringRef::size_type Pos = S.find_first_not_of(" \t");
25709 if (Pos == 0) // We matched a prefix.
25718 static bool clobbersFlagRegisters(const SmallVector<StringRef, 4> &AsmPieces) {
25720 if (AsmPieces.size() == 3 || AsmPieces.size() == 4) {
25721 if (std::count(AsmPieces.begin(), AsmPieces.end(), "~{cc}") &&
25722 std::count(AsmPieces.begin(), AsmPieces.end(), "~{flags}") &&
25723 std::count(AsmPieces.begin(), AsmPieces.end(), "~{fpsr}")) {
25725 if (AsmPieces.size() == 3)
25727 else if (std::count(AsmPieces.begin(), AsmPieces.end(), "~{dirflag}"))
25734 bool X86TargetLowering::ExpandInlineAsm(CallInst *CI) const {
25735 InlineAsm *IA = cast<InlineAsm>(CI->getCalledValue());
25737 std::string AsmStr = IA->getAsmString();
25739 IntegerType *Ty = dyn_cast<IntegerType>(CI->getType());
25740 if (!Ty || Ty->getBitWidth() % 16 != 0)
25743 // TODO: should remove alternatives from the asmstring: "foo {a|b}" -> "foo a"
25744 SmallVector<StringRef, 4> AsmPieces;
25745 SplitString(AsmStr, AsmPieces, ";\n");
25747 switch (AsmPieces.size()) {
25748 default: return false;
25750 // FIXME: this should verify that we are targeting a 486 or better. If not,
25751 // we will turn this bswap into something that will be lowered to logical
25752 // ops instead of emitting the bswap asm. For now, we don't support 486 or
25753 // lower so don't worry about this.
25755 if (matchAsm(AsmPieces[0], {"bswap", "$0"}) ||
25756 matchAsm(AsmPieces[0], {"bswapl", "$0"}) ||
25757 matchAsm(AsmPieces[0], {"bswapq", "$0"}) ||
25758 matchAsm(AsmPieces[0], {"bswap", "${0:q}"}) ||
25759 matchAsm(AsmPieces[0], {"bswapl", "${0:q}"}) ||
25760 matchAsm(AsmPieces[0], {"bswapq", "${0:q}"})) {
25761 // No need to check constraints, nothing other than the equivalent of
25762 // "=r,0" would be valid here.
25763 return IntrinsicLowering::LowerToByteSwap(CI);
25766 // rorw $$8, ${0:w} --> llvm.bswap.i16
25767 if (CI->getType()->isIntegerTy(16) &&
25768 IA->getConstraintString().compare(0, 5, "=r,0,") == 0 &&
25769 (matchAsm(AsmPieces[0], {"rorw", "$$8,", "${0:w}"}) ||
25770 matchAsm(AsmPieces[0], {"rolw", "$$8,", "${0:w}"}))) {
25772 StringRef ConstraintsStr = IA->getConstraintString();
25773 SplitString(StringRef(ConstraintsStr).substr(5), AsmPieces, ",");
25774 array_pod_sort(AsmPieces.begin(), AsmPieces.end());
25775 if (clobbersFlagRegisters(AsmPieces))
25776 return IntrinsicLowering::LowerToByteSwap(CI);
25780 if (CI->getType()->isIntegerTy(32) &&
25781 IA->getConstraintString().compare(0, 5, "=r,0,") == 0 &&
25782 matchAsm(AsmPieces[0], {"rorw", "$$8,", "${0:w}"}) &&
25783 matchAsm(AsmPieces[1], {"rorl", "$$16,", "$0"}) &&
25784 matchAsm(AsmPieces[2], {"rorw", "$$8,", "${0:w}"})) {
25786 StringRef ConstraintsStr = IA->getConstraintString();
25787 SplitString(StringRef(ConstraintsStr).substr(5), AsmPieces, ",");
25788 array_pod_sort(AsmPieces.begin(), AsmPieces.end());
25789 if (clobbersFlagRegisters(AsmPieces))
25790 return IntrinsicLowering::LowerToByteSwap(CI);
25793 if (CI->getType()->isIntegerTy(64)) {
25794 InlineAsm::ConstraintInfoVector Constraints = IA->ParseConstraints();
25795 if (Constraints.size() >= 2 &&
25796 Constraints[0].Codes.size() == 1 && Constraints[0].Codes[0] == "A" &&
25797 Constraints[1].Codes.size() == 1 && Constraints[1].Codes[0] == "0") {
25798 // bswap %eax / bswap %edx / xchgl %eax, %edx -> llvm.bswap.i64
25799 if (matchAsm(AsmPieces[0], {"bswap", "%eax"}) &&
25800 matchAsm(AsmPieces[1], {"bswap", "%edx"}) &&
25801 matchAsm(AsmPieces[2], {"xchgl", "%eax,", "%edx"}))
25802 return IntrinsicLowering::LowerToByteSwap(CI);
25810 /// getConstraintType - Given a constraint letter, return the type of
25811 /// constraint it is for this target.
25812 X86TargetLowering::ConstraintType
25813 X86TargetLowering::getConstraintType(StringRef Constraint) const {
25814 if (Constraint.size() == 1) {
25815 switch (Constraint[0]) {
25826 return C_RegisterClass;
25850 return TargetLowering::getConstraintType(Constraint);
25853 /// Examine constraint type and operand type and determine a weight value.
25854 /// This object must already have been set up with the operand type
25855 /// and the current alternative constraint selected.
25856 TargetLowering::ConstraintWeight
25857 X86TargetLowering::getSingleConstraintMatchWeight(
25858 AsmOperandInfo &info, const char *constraint) const {
25859 ConstraintWeight weight = CW_Invalid;
25860 Value *CallOperandVal = info.CallOperandVal;
25861 // If we don't have a value, we can't do a match,
25862 // but allow it at the lowest weight.
25863 if (!CallOperandVal)
25865 Type *type = CallOperandVal->getType();
25866 // Look at the constraint type.
25867 switch (*constraint) {
25869 weight = TargetLowering::getSingleConstraintMatchWeight(info, constraint);
25880 if (CallOperandVal->getType()->isIntegerTy())
25881 weight = CW_SpecificReg;
25886 if (type->isFloatingPointTy())
25887 weight = CW_SpecificReg;
25890 if (type->isX86_MMXTy() && Subtarget->hasMMX())
25891 weight = CW_SpecificReg;
25895 if (((type->getPrimitiveSizeInBits() == 128) && Subtarget->hasSSE1()) ||
25896 ((type->getPrimitiveSizeInBits() == 256) && Subtarget->hasFp256()))
25897 weight = CW_Register;
25900 if (ConstantInt *C = dyn_cast<ConstantInt>(info.CallOperandVal)) {
25901 if (C->getZExtValue() <= 31)
25902 weight = CW_Constant;
25906 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
25907 if (C->getZExtValue() <= 63)
25908 weight = CW_Constant;
25912 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
25913 if ((C->getSExtValue() >= -0x80) && (C->getSExtValue() <= 0x7f))
25914 weight = CW_Constant;
25918 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
25919 if ((C->getZExtValue() == 0xff) || (C->getZExtValue() == 0xffff))
25920 weight = CW_Constant;
25924 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
25925 if (C->getZExtValue() <= 3)
25926 weight = CW_Constant;
25930 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
25931 if (C->getZExtValue() <= 0xff)
25932 weight = CW_Constant;
25937 if (isa<ConstantFP>(CallOperandVal)) {
25938 weight = CW_Constant;
25942 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
25943 if ((C->getSExtValue() >= -0x80000000LL) &&
25944 (C->getSExtValue() <= 0x7fffffffLL))
25945 weight = CW_Constant;
25949 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
25950 if (C->getZExtValue() <= 0xffffffff)
25951 weight = CW_Constant;
25958 /// LowerXConstraint - try to replace an X constraint, which matches anything,
25959 /// with another that has more specific requirements based on the type of the
25960 /// corresponding operand.
25961 const char *X86TargetLowering::
25962 LowerXConstraint(EVT ConstraintVT) const {
25963 // FP X constraints get lowered to SSE1/2 registers if available, otherwise
25964 // 'f' like normal targets.
25965 if (ConstraintVT.isFloatingPoint()) {
25966 if (Subtarget->hasSSE2())
25968 if (Subtarget->hasSSE1())
25972 return TargetLowering::LowerXConstraint(ConstraintVT);
25975 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
25976 /// vector. If it is invalid, don't add anything to Ops.
25977 void X86TargetLowering::LowerAsmOperandForConstraint(SDValue Op,
25978 std::string &Constraint,
25979 std::vector<SDValue>&Ops,
25980 SelectionDAG &DAG) const {
25983 // Only support length 1 constraints for now.
25984 if (Constraint.length() > 1) return;
25986 char ConstraintLetter = Constraint[0];
25987 switch (ConstraintLetter) {
25990 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
25991 if (C->getZExtValue() <= 31) {
25992 Result = DAG.getTargetConstant(C->getZExtValue(), SDLoc(Op),
25993 Op.getValueType());
25999 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
26000 if (C->getZExtValue() <= 63) {
26001 Result = DAG.getTargetConstant(C->getZExtValue(), SDLoc(Op),
26002 Op.getValueType());
26008 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
26009 if (isInt<8>(C->getSExtValue())) {
26010 Result = DAG.getTargetConstant(C->getZExtValue(), SDLoc(Op),
26011 Op.getValueType());
26017 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
26018 if (C->getZExtValue() == 0xff || C->getZExtValue() == 0xffff ||
26019 (Subtarget->is64Bit() && C->getZExtValue() == 0xffffffff)) {
26020 Result = DAG.getTargetConstant(C->getSExtValue(), SDLoc(Op),
26021 Op.getValueType());
26027 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
26028 if (C->getZExtValue() <= 3) {
26029 Result = DAG.getTargetConstant(C->getZExtValue(), SDLoc(Op),
26030 Op.getValueType());
26036 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
26037 if (C->getZExtValue() <= 255) {
26038 Result = DAG.getTargetConstant(C->getZExtValue(), SDLoc(Op),
26039 Op.getValueType());
26045 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
26046 if (C->getZExtValue() <= 127) {
26047 Result = DAG.getTargetConstant(C->getZExtValue(), SDLoc(Op),
26048 Op.getValueType());
26054 // 32-bit signed value
26055 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
26056 if (ConstantInt::isValueValidForType(Type::getInt32Ty(*DAG.getContext()),
26057 C->getSExtValue())) {
26058 // Widen to 64 bits here to get it sign extended.
26059 Result = DAG.getTargetConstant(C->getSExtValue(), SDLoc(Op), MVT::i64);
26062 // FIXME gcc accepts some relocatable values here too, but only in certain
26063 // memory models; it's complicated.
26068 // 32-bit unsigned value
26069 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
26070 if (ConstantInt::isValueValidForType(Type::getInt32Ty(*DAG.getContext()),
26071 C->getZExtValue())) {
26072 Result = DAG.getTargetConstant(C->getZExtValue(), SDLoc(Op),
26073 Op.getValueType());
26077 // FIXME gcc accepts some relocatable values here too, but only in certain
26078 // memory models; it's complicated.
26082 // Literal immediates are always ok.
26083 if (ConstantSDNode *CST = dyn_cast<ConstantSDNode>(Op)) {
26084 // Widen to 64 bits here to get it sign extended.
26085 Result = DAG.getTargetConstant(CST->getSExtValue(), SDLoc(Op), MVT::i64);
26089 // In any sort of PIC mode addresses need to be computed at runtime by
26090 // adding in a register or some sort of table lookup. These can't
26091 // be used as immediates.
26092 if (Subtarget->isPICStyleGOT() || Subtarget->isPICStyleStubPIC())
26095 // If we are in non-pic codegen mode, we allow the address of a global (with
26096 // an optional displacement) to be used with 'i'.
26097 GlobalAddressSDNode *GA = nullptr;
26098 int64_t Offset = 0;
26100 // Match either (GA), (GA+C), (GA+C1+C2), etc.
26102 if ((GA = dyn_cast<GlobalAddressSDNode>(Op))) {
26103 Offset += GA->getOffset();
26105 } else if (Op.getOpcode() == ISD::ADD) {
26106 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
26107 Offset += C->getZExtValue();
26108 Op = Op.getOperand(0);
26111 } else if (Op.getOpcode() == ISD::SUB) {
26112 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
26113 Offset += -C->getZExtValue();
26114 Op = Op.getOperand(0);
26119 // Otherwise, this isn't something we can handle, reject it.
26123 const GlobalValue *GV = GA->getGlobal();
26124 // If we require an extra load to get this address, as in PIC mode, we
26125 // can't accept it.
26126 if (isGlobalStubReference(
26127 Subtarget->ClassifyGlobalReference(GV, DAG.getTarget())))
26130 Result = DAG.getTargetGlobalAddress(GV, SDLoc(Op),
26131 GA->getValueType(0), Offset);
26136 if (Result.getNode()) {
26137 Ops.push_back(Result);
26140 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG);
26143 std::pair<unsigned, const TargetRegisterClass *>
26144 X86TargetLowering::getRegForInlineAsmConstraint(const TargetRegisterInfo *TRI,
26145 StringRef Constraint,
26147 // First, see if this is a constraint that directly corresponds to an LLVM
26149 if (Constraint.size() == 1) {
26150 // GCC Constraint Letters
26151 switch (Constraint[0]) {
26153 // TODO: Slight differences here in allocation order and leaving
26154 // RIP in the class. Do they matter any more here than they do
26155 // in the normal allocation?
26156 case 'q': // GENERAL_REGS in 64-bit mode, Q_REGS in 32-bit mode.
26157 if (Subtarget->is64Bit()) {
26158 if (VT == MVT::i32 || VT == MVT::f32)
26159 return std::make_pair(0U, &X86::GR32RegClass);
26160 if (VT == MVT::i16)
26161 return std::make_pair(0U, &X86::GR16RegClass);
26162 if (VT == MVT::i8 || VT == MVT::i1)
26163 return std::make_pair(0U, &X86::GR8RegClass);
26164 if (VT == MVT::i64 || VT == MVT::f64)
26165 return std::make_pair(0U, &X86::GR64RegClass);
26168 // 32-bit fallthrough
26169 case 'Q': // Q_REGS
26170 if (VT == MVT::i32 || VT == MVT::f32)
26171 return std::make_pair(0U, &X86::GR32_ABCDRegClass);
26172 if (VT == MVT::i16)
26173 return std::make_pair(0U, &X86::GR16_ABCDRegClass);
26174 if (VT == MVT::i8 || VT == MVT::i1)
26175 return std::make_pair(0U, &X86::GR8_ABCD_LRegClass);
26176 if (VT == MVT::i64)
26177 return std::make_pair(0U, &X86::GR64_ABCDRegClass);
26179 case 'r': // GENERAL_REGS
26180 case 'l': // INDEX_REGS
26181 if (VT == MVT::i8 || VT == MVT::i1)
26182 return std::make_pair(0U, &X86::GR8RegClass);
26183 if (VT == MVT::i16)
26184 return std::make_pair(0U, &X86::GR16RegClass);
26185 if (VT == MVT::i32 || VT == MVT::f32 || !Subtarget->is64Bit())
26186 return std::make_pair(0U, &X86::GR32RegClass);
26187 return std::make_pair(0U, &X86::GR64RegClass);
26188 case 'R': // LEGACY_REGS
26189 if (VT == MVT::i8 || VT == MVT::i1)
26190 return std::make_pair(0U, &X86::GR8_NOREXRegClass);
26191 if (VT == MVT::i16)
26192 return std::make_pair(0U, &X86::GR16_NOREXRegClass);
26193 if (VT == MVT::i32 || !Subtarget->is64Bit())
26194 return std::make_pair(0U, &X86::GR32_NOREXRegClass);
26195 return std::make_pair(0U, &X86::GR64_NOREXRegClass);
26196 case 'f': // FP Stack registers.
26197 // If SSE is enabled for this VT, use f80 to ensure the isel moves the
26198 // value to the correct fpstack register class.
26199 if (VT == MVT::f32 && !isScalarFPTypeInSSEReg(VT))
26200 return std::make_pair(0U, &X86::RFP32RegClass);
26201 if (VT == MVT::f64 && !isScalarFPTypeInSSEReg(VT))
26202 return std::make_pair(0U, &X86::RFP64RegClass);
26203 return std::make_pair(0U, &X86::RFP80RegClass);
26204 case 'y': // MMX_REGS if MMX allowed.
26205 if (!Subtarget->hasMMX()) break;
26206 return std::make_pair(0U, &X86::VR64RegClass);
26207 case 'Y': // SSE_REGS if SSE2 allowed
26208 if (!Subtarget->hasSSE2()) break;
26210 case 'x': // SSE_REGS if SSE1 allowed or AVX_REGS if AVX allowed
26211 if (!Subtarget->hasSSE1()) break;
26213 switch (VT.SimpleTy) {
26215 // Scalar SSE types.
26218 return std::make_pair(0U, &X86::FR32RegClass);
26221 return std::make_pair(0U, &X86::FR64RegClass);
26229 return std::make_pair(0U, &X86::VR128RegClass);
26237 return std::make_pair(0U, &X86::VR256RegClass);
26242 return std::make_pair(0U, &X86::VR512RegClass);
26248 // Use the default implementation in TargetLowering to convert the register
26249 // constraint into a member of a register class.
26250 std::pair<unsigned, const TargetRegisterClass*> Res;
26251 Res = TargetLowering::getRegForInlineAsmConstraint(TRI, Constraint, VT);
26253 // Not found as a standard register?
26255 // Map st(0) -> st(7) -> ST0
26256 if (Constraint.size() == 7 && Constraint[0] == '{' &&
26257 tolower(Constraint[1]) == 's' &&
26258 tolower(Constraint[2]) == 't' &&
26259 Constraint[3] == '(' &&
26260 (Constraint[4] >= '0' && Constraint[4] <= '7') &&
26261 Constraint[5] == ')' &&
26262 Constraint[6] == '}') {
26264 Res.first = X86::FP0+Constraint[4]-'0';
26265 Res.second = &X86::RFP80RegClass;
26269 // GCC allows "st(0)" to be called just plain "st".
26270 if (StringRef("{st}").equals_lower(Constraint)) {
26271 Res.first = X86::FP0;
26272 Res.second = &X86::RFP80RegClass;
26277 if (StringRef("{flags}").equals_lower(Constraint)) {
26278 Res.first = X86::EFLAGS;
26279 Res.second = &X86::CCRRegClass;
26283 // 'A' means EAX + EDX.
26284 if (Constraint == "A") {
26285 Res.first = X86::EAX;
26286 Res.second = &X86::GR32_ADRegClass;
26292 // Otherwise, check to see if this is a register class of the wrong value
26293 // type. For example, we want to map "{ax},i32" -> {eax}, we don't want it to
26294 // turn into {ax},{dx}.
26295 // MVT::Other is used to specify clobber names.
26296 if (Res.second->hasType(VT) || VT == MVT::Other)
26297 return Res; // Correct type already, nothing to do.
26299 // Get a matching integer of the correct size. i.e. "ax" with MVT::32 should
26300 // return "eax". This should even work for things like getting 64bit integer
26301 // registers when given an f64 type.
26302 const TargetRegisterClass *Class = Res.second;
26303 if (Class == &X86::GR8RegClass || Class == &X86::GR16RegClass ||
26304 Class == &X86::GR32RegClass || Class == &X86::GR64RegClass) {
26305 unsigned Size = VT.getSizeInBits();
26306 MVT::SimpleValueType SimpleTy = Size == 1 || Size == 8 ? MVT::i8
26307 : Size == 16 ? MVT::i16
26308 : Size == 32 ? MVT::i32
26309 : Size == 64 ? MVT::i64
26311 unsigned DestReg = getX86SubSuperRegisterOrZero(Res.first, SimpleTy);
26313 Res.first = DestReg;
26314 Res.second = SimpleTy == MVT::i8 ? &X86::GR8RegClass
26315 : SimpleTy == MVT::i16 ? &X86::GR16RegClass
26316 : SimpleTy == MVT::i32 ? &X86::GR32RegClass
26317 : &X86::GR64RegClass;
26318 assert(Res.second->contains(Res.first) && "Register in register class");
26320 // No register found/type mismatch.
26322 Res.second = nullptr;
26324 } else if (Class == &X86::FR32RegClass || Class == &X86::FR64RegClass ||
26325 Class == &X86::VR128RegClass || Class == &X86::VR256RegClass ||
26326 Class == &X86::FR32XRegClass || Class == &X86::FR64XRegClass ||
26327 Class == &X86::VR128XRegClass || Class == &X86::VR256XRegClass ||
26328 Class == &X86::VR512RegClass) {
26329 // Handle references to XMM physical registers that got mapped into the
26330 // wrong class. This can happen with constraints like {xmm0} where the
26331 // target independent register mapper will just pick the first match it can
26332 // find, ignoring the required type.
26334 if (VT == MVT::f32 || VT == MVT::i32)
26335 Res.second = &X86::FR32RegClass;
26336 else if (VT == MVT::f64 || VT == MVT::i64)
26337 Res.second = &X86::FR64RegClass;
26338 else if (X86::VR128RegClass.hasType(VT))
26339 Res.second = &X86::VR128RegClass;
26340 else if (X86::VR256RegClass.hasType(VT))
26341 Res.second = &X86::VR256RegClass;
26342 else if (X86::VR512RegClass.hasType(VT))
26343 Res.second = &X86::VR512RegClass;
26345 // Type mismatch and not a clobber: Return an error;
26347 Res.second = nullptr;
26354 int X86TargetLowering::getScalingFactorCost(const DataLayout &DL,
26355 const AddrMode &AM, Type *Ty,
26356 unsigned AS) const {
26357 // Scaling factors are not free at all.
26358 // An indexed folded instruction, i.e., inst (reg1, reg2, scale),
26359 // will take 2 allocations in the out of order engine instead of 1
26360 // for plain addressing mode, i.e. inst (reg1).
26362 // vaddps (%rsi,%drx), %ymm0, %ymm1
26363 // Requires two allocations (one for the load, one for the computation)
26365 // vaddps (%rsi), %ymm0, %ymm1
26366 // Requires just 1 allocation, i.e., freeing allocations for other operations
26367 // and having less micro operations to execute.
26369 // For some X86 architectures, this is even worse because for instance for
26370 // stores, the complex addressing mode forces the instruction to use the
26371 // "load" ports instead of the dedicated "store" port.
26372 // E.g., on Haswell:
26373 // vmovaps %ymm1, (%r8, %rdi) can use port 2 or 3.
26374 // vmovaps %ymm1, (%r8) can use port 2, 3, or 7.
26375 if (isLegalAddressingMode(DL, AM, Ty, AS))
26376 // Scale represents reg2 * scale, thus account for 1
26377 // as soon as we use a second register.
26378 return AM.Scale != 0;
26382 bool X86TargetLowering::isTargetFTOL() const {
26383 return Subtarget->isTargetKnownWindowsMSVC() && !Subtarget->is64Bit();