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->isTarget64BitLP64()) {
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);
1202 setOperationAction(ISD::SMAX, MVT::v32i8, Custom);
1203 setOperationAction(ISD::SMAX, MVT::v16i16, Custom);
1204 setOperationAction(ISD::SMAX, MVT::v8i32, Custom);
1205 setOperationAction(ISD::UMAX, MVT::v32i8, Custom);
1206 setOperationAction(ISD::UMAX, MVT::v16i16, Custom);
1207 setOperationAction(ISD::UMAX, MVT::v8i32, Custom);
1208 setOperationAction(ISD::SMIN, MVT::v32i8, Custom);
1209 setOperationAction(ISD::SMIN, MVT::v16i16, Custom);
1210 setOperationAction(ISD::SMIN, MVT::v8i32, Custom);
1211 setOperationAction(ISD::UMIN, MVT::v32i8, Custom);
1212 setOperationAction(ISD::UMIN, MVT::v16i16, Custom);
1213 setOperationAction(ISD::UMIN, MVT::v8i32, Custom);
1216 // In the customized shift lowering, the legal cases in AVX2 will be
1218 setOperationAction(ISD::SRL, MVT::v4i64, Custom);
1219 setOperationAction(ISD::SRL, MVT::v8i32, Custom);
1221 setOperationAction(ISD::SHL, MVT::v4i64, Custom);
1222 setOperationAction(ISD::SHL, MVT::v8i32, Custom);
1224 setOperationAction(ISD::SRA, MVT::v4i64, Custom);
1225 setOperationAction(ISD::SRA, MVT::v8i32, Custom);
1227 // Custom lower several nodes for 256-bit types.
1228 for (MVT VT : MVT::vector_valuetypes()) {
1229 if (VT.getScalarSizeInBits() >= 32) {
1230 setOperationAction(ISD::MLOAD, VT, Legal);
1231 setOperationAction(ISD::MSTORE, VT, Legal);
1233 // Extract subvector is special because the value type
1234 // (result) is 128-bit but the source is 256-bit wide.
1235 if (VT.is128BitVector()) {
1236 setOperationAction(ISD::EXTRACT_SUBVECTOR, VT, Custom);
1238 // Do not attempt to custom lower other non-256-bit vectors
1239 if (!VT.is256BitVector())
1242 setOperationAction(ISD::BUILD_VECTOR, VT, Custom);
1243 setOperationAction(ISD::VECTOR_SHUFFLE, VT, Custom);
1244 setOperationAction(ISD::VSELECT, VT, Custom);
1245 setOperationAction(ISD::INSERT_VECTOR_ELT, VT, Custom);
1246 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Custom);
1247 setOperationAction(ISD::SCALAR_TO_VECTOR, VT, Custom);
1248 setOperationAction(ISD::INSERT_SUBVECTOR, VT, Custom);
1249 setOperationAction(ISD::CONCAT_VECTORS, VT, Custom);
1252 if (Subtarget->hasInt256())
1253 setOperationAction(ISD::VSELECT, MVT::v32i8, Legal);
1256 // Promote v32i8, v16i16, v8i32 select, and, or, xor to v4i64.
1257 for (int i = MVT::v32i8; i != MVT::v4i64; ++i) {
1258 MVT VT = (MVT::SimpleValueType)i;
1260 // Do not attempt to promote non-256-bit vectors
1261 if (!VT.is256BitVector())
1264 setOperationAction(ISD::AND, VT, Promote);
1265 AddPromotedToType (ISD::AND, VT, MVT::v4i64);
1266 setOperationAction(ISD::OR, VT, Promote);
1267 AddPromotedToType (ISD::OR, VT, MVT::v4i64);
1268 setOperationAction(ISD::XOR, VT, Promote);
1269 AddPromotedToType (ISD::XOR, VT, MVT::v4i64);
1270 setOperationAction(ISD::LOAD, VT, Promote);
1271 AddPromotedToType (ISD::LOAD, VT, MVT::v4i64);
1272 setOperationAction(ISD::SELECT, VT, Promote);
1273 AddPromotedToType (ISD::SELECT, VT, MVT::v4i64);
1277 if (!Subtarget->useSoftFloat() && Subtarget->hasAVX512()) {
1278 addRegisterClass(MVT::v16i32, &X86::VR512RegClass);
1279 addRegisterClass(MVT::v16f32, &X86::VR512RegClass);
1280 addRegisterClass(MVT::v8i64, &X86::VR512RegClass);
1281 addRegisterClass(MVT::v8f64, &X86::VR512RegClass);
1283 addRegisterClass(MVT::i1, &X86::VK1RegClass);
1284 addRegisterClass(MVT::v8i1, &X86::VK8RegClass);
1285 addRegisterClass(MVT::v16i1, &X86::VK16RegClass);
1287 for (MVT VT : MVT::fp_vector_valuetypes())
1288 setLoadExtAction(ISD::EXTLOAD, VT, MVT::v8f32, Legal);
1290 setLoadExtAction(ISD::ZEXTLOAD, MVT::v16i32, MVT::v16i8, Legal);
1291 setLoadExtAction(ISD::SEXTLOAD, MVT::v16i32, MVT::v16i8, Legal);
1292 setLoadExtAction(ISD::ZEXTLOAD, MVT::v16i32, MVT::v16i16, Legal);
1293 setLoadExtAction(ISD::SEXTLOAD, MVT::v16i32, MVT::v16i16, Legal);
1294 setLoadExtAction(ISD::ZEXTLOAD, MVT::v32i16, MVT::v32i8, Legal);
1295 setLoadExtAction(ISD::SEXTLOAD, MVT::v32i16, MVT::v32i8, Legal);
1296 setLoadExtAction(ISD::ZEXTLOAD, MVT::v8i64, MVT::v8i8, Legal);
1297 setLoadExtAction(ISD::SEXTLOAD, MVT::v8i64, MVT::v8i8, Legal);
1298 setLoadExtAction(ISD::ZEXTLOAD, MVT::v8i64, MVT::v8i16, Legal);
1299 setLoadExtAction(ISD::SEXTLOAD, MVT::v8i64, MVT::v8i16, Legal);
1300 setLoadExtAction(ISD::ZEXTLOAD, MVT::v8i64, MVT::v8i32, Legal);
1301 setLoadExtAction(ISD::SEXTLOAD, MVT::v8i64, MVT::v8i32, Legal);
1303 setOperationAction(ISD::BR_CC, MVT::i1, Expand);
1304 setOperationAction(ISD::SETCC, MVT::i1, Custom);
1305 setOperationAction(ISD::XOR, MVT::i1, Legal);
1306 setOperationAction(ISD::OR, MVT::i1, Legal);
1307 setOperationAction(ISD::AND, MVT::i1, Legal);
1308 setOperationAction(ISD::SUB, MVT::i1, Custom);
1309 setOperationAction(ISD::ADD, MVT::i1, Custom);
1310 setOperationAction(ISD::MUL, MVT::i1, Custom);
1311 setOperationAction(ISD::LOAD, MVT::v16f32, Legal);
1312 setOperationAction(ISD::LOAD, MVT::v8f64, Legal);
1313 setOperationAction(ISD::LOAD, MVT::v8i64, Legal);
1314 setOperationAction(ISD::LOAD, MVT::v16i32, Legal);
1315 setOperationAction(ISD::LOAD, MVT::v16i1, Legal);
1317 setOperationAction(ISD::FADD, MVT::v16f32, Legal);
1318 setOperationAction(ISD::FSUB, MVT::v16f32, Legal);
1319 setOperationAction(ISD::FMUL, MVT::v16f32, Legal);
1320 setOperationAction(ISD::FDIV, MVT::v16f32, Legal);
1321 setOperationAction(ISD::FSQRT, MVT::v16f32, Legal);
1322 setOperationAction(ISD::FNEG, MVT::v16f32, Custom);
1324 setOperationAction(ISD::FADD, MVT::v8f64, Legal);
1325 setOperationAction(ISD::FSUB, MVT::v8f64, Legal);
1326 setOperationAction(ISD::FMUL, MVT::v8f64, Legal);
1327 setOperationAction(ISD::FDIV, MVT::v8f64, Legal);
1328 setOperationAction(ISD::FSQRT, MVT::v8f64, Legal);
1329 setOperationAction(ISD::FNEG, MVT::v8f64, Custom);
1330 setOperationAction(ISD::FMA, MVT::v8f64, Legal);
1331 setOperationAction(ISD::FMA, MVT::v16f32, Legal);
1333 setOperationAction(ISD::FP_TO_SINT, MVT::i32, Legal);
1334 setOperationAction(ISD::FP_TO_UINT, MVT::i32, Legal);
1335 setOperationAction(ISD::SINT_TO_FP, MVT::i32, Legal);
1336 setOperationAction(ISD::UINT_TO_FP, MVT::i32, Legal);
1337 if (Subtarget->is64Bit()) {
1338 setOperationAction(ISD::FP_TO_UINT, MVT::i64, Legal);
1339 setOperationAction(ISD::FP_TO_SINT, MVT::i64, Legal);
1340 setOperationAction(ISD::SINT_TO_FP, MVT::i64, Legal);
1341 setOperationAction(ISD::UINT_TO_FP, MVT::i64, Legal);
1343 setOperationAction(ISD::FP_TO_SINT, MVT::v16i32, Legal);
1344 setOperationAction(ISD::FP_TO_UINT, MVT::v16i32, Legal);
1345 setOperationAction(ISD::FP_TO_UINT, MVT::v8i32, Legal);
1346 setOperationAction(ISD::FP_TO_UINT, MVT::v4i32, Legal);
1347 setOperationAction(ISD::SINT_TO_FP, MVT::v16i32, Legal);
1348 setOperationAction(ISD::SINT_TO_FP, MVT::v8i1, Custom);
1349 setOperationAction(ISD::SINT_TO_FP, MVT::v16i1, Custom);
1350 setOperationAction(ISD::SINT_TO_FP, MVT::v16i8, Promote);
1351 setOperationAction(ISD::SINT_TO_FP, MVT::v16i16, Promote);
1352 setOperationAction(ISD::UINT_TO_FP, MVT::v16i32, Legal);
1353 setOperationAction(ISD::UINT_TO_FP, MVT::v8i32, Legal);
1354 setOperationAction(ISD::UINT_TO_FP, MVT::v4i32, Legal);
1355 setOperationAction(ISD::UINT_TO_FP, MVT::v16i8, Custom);
1356 setOperationAction(ISD::UINT_TO_FP, MVT::v16i16, Custom);
1357 setOperationAction(ISD::FP_ROUND, MVT::v8f32, Legal);
1358 setOperationAction(ISD::FP_EXTEND, MVT::v8f32, Legal);
1360 setTruncStoreAction(MVT::v8i64, MVT::v8i8, Legal);
1361 setTruncStoreAction(MVT::v8i64, MVT::v8i16, Legal);
1362 setTruncStoreAction(MVT::v8i64, MVT::v8i32, Legal);
1363 setTruncStoreAction(MVT::v16i32, MVT::v16i8, Legal);
1364 setTruncStoreAction(MVT::v16i32, MVT::v16i16, Legal);
1365 if (Subtarget->hasVLX()){
1366 setTruncStoreAction(MVT::v4i64, MVT::v4i8, Legal);
1367 setTruncStoreAction(MVT::v4i64, MVT::v4i16, Legal);
1368 setTruncStoreAction(MVT::v4i64, MVT::v4i32, Legal);
1369 setTruncStoreAction(MVT::v8i32, MVT::v8i8, Legal);
1370 setTruncStoreAction(MVT::v8i32, MVT::v8i16, Legal);
1372 setTruncStoreAction(MVT::v2i64, MVT::v2i8, Legal);
1373 setTruncStoreAction(MVT::v2i64, MVT::v2i16, Legal);
1374 setTruncStoreAction(MVT::v2i64, MVT::v2i32, Legal);
1375 setTruncStoreAction(MVT::v4i32, MVT::v4i8, Legal);
1376 setTruncStoreAction(MVT::v4i32, MVT::v4i16, Legal);
1378 setOperationAction(ISD::TRUNCATE, MVT::i1, Custom);
1379 setOperationAction(ISD::TRUNCATE, MVT::v16i8, Custom);
1380 setOperationAction(ISD::TRUNCATE, MVT::v8i32, Custom);
1381 if (Subtarget->hasDQI()) {
1382 setOperationAction(ISD::TRUNCATE, MVT::v2i1, Custom);
1383 setOperationAction(ISD::TRUNCATE, MVT::v4i1, Custom);
1385 setOperationAction(ISD::SINT_TO_FP, MVT::v8i64, Legal);
1386 setOperationAction(ISD::UINT_TO_FP, MVT::v8i64, Legal);
1387 setOperationAction(ISD::FP_TO_SINT, MVT::v8i64, Legal);
1388 setOperationAction(ISD::FP_TO_UINT, MVT::v8i64, Legal);
1389 if (Subtarget->hasVLX()) {
1390 setOperationAction(ISD::SINT_TO_FP, MVT::v4i64, Legal);
1391 setOperationAction(ISD::SINT_TO_FP, MVT::v2i64, Legal);
1392 setOperationAction(ISD::UINT_TO_FP, MVT::v4i64, Legal);
1393 setOperationAction(ISD::UINT_TO_FP, MVT::v2i64, Legal);
1394 setOperationAction(ISD::FP_TO_SINT, MVT::v4i64, Legal);
1395 setOperationAction(ISD::FP_TO_SINT, MVT::v2i64, Legal);
1396 setOperationAction(ISD::FP_TO_UINT, MVT::v4i64, Legal);
1397 setOperationAction(ISD::FP_TO_UINT, MVT::v2i64, Legal);
1400 if (Subtarget->hasVLX()) {
1401 setOperationAction(ISD::SINT_TO_FP, MVT::v8i32, Legal);
1402 setOperationAction(ISD::UINT_TO_FP, MVT::v8i32, Legal);
1403 setOperationAction(ISD::FP_TO_SINT, MVT::v8i32, Legal);
1404 setOperationAction(ISD::FP_TO_UINT, MVT::v8i32, Legal);
1405 setOperationAction(ISD::SINT_TO_FP, MVT::v4i32, Legal);
1406 setOperationAction(ISD::UINT_TO_FP, MVT::v4i32, Legal);
1407 setOperationAction(ISD::FP_TO_SINT, MVT::v4i32, Legal);
1408 setOperationAction(ISD::FP_TO_UINT, MVT::v4i32, Legal);
1410 setOperationAction(ISD::TRUNCATE, MVT::v8i1, Custom);
1411 setOperationAction(ISD::TRUNCATE, MVT::v16i1, Custom);
1412 setOperationAction(ISD::TRUNCATE, MVT::v16i16, Custom);
1413 setOperationAction(ISD::ZERO_EXTEND, MVT::v16i32, Custom);
1414 setOperationAction(ISD::ZERO_EXTEND, MVT::v8i64, Custom);
1415 setOperationAction(ISD::ANY_EXTEND, MVT::v16i32, Custom);
1416 setOperationAction(ISD::ANY_EXTEND, MVT::v8i64, Custom);
1417 setOperationAction(ISD::SIGN_EXTEND, MVT::v16i32, Custom);
1418 setOperationAction(ISD::SIGN_EXTEND, MVT::v8i64, Custom);
1419 setOperationAction(ISD::SIGN_EXTEND, MVT::v16i8, Custom);
1420 setOperationAction(ISD::SIGN_EXTEND, MVT::v8i16, Custom);
1421 setOperationAction(ISD::SIGN_EXTEND, MVT::v16i16, Custom);
1422 if (Subtarget->hasDQI()) {
1423 setOperationAction(ISD::SIGN_EXTEND, MVT::v4i32, Custom);
1424 setOperationAction(ISD::SIGN_EXTEND, MVT::v2i64, Custom);
1426 setOperationAction(ISD::FFLOOR, MVT::v16f32, Legal);
1427 setOperationAction(ISD::FFLOOR, MVT::v8f64, Legal);
1428 setOperationAction(ISD::FCEIL, MVT::v16f32, Legal);
1429 setOperationAction(ISD::FCEIL, MVT::v8f64, Legal);
1430 setOperationAction(ISD::FTRUNC, MVT::v16f32, Legal);
1431 setOperationAction(ISD::FTRUNC, MVT::v8f64, Legal);
1432 setOperationAction(ISD::FRINT, MVT::v16f32, Legal);
1433 setOperationAction(ISD::FRINT, MVT::v8f64, Legal);
1434 setOperationAction(ISD::FNEARBYINT, MVT::v16f32, Legal);
1435 setOperationAction(ISD::FNEARBYINT, MVT::v8f64, Legal);
1437 setOperationAction(ISD::CONCAT_VECTORS, MVT::v8f64, Custom);
1438 setOperationAction(ISD::CONCAT_VECTORS, MVT::v8i64, Custom);
1439 setOperationAction(ISD::CONCAT_VECTORS, MVT::v16f32, Custom);
1440 setOperationAction(ISD::CONCAT_VECTORS, MVT::v16i32, Custom);
1441 setOperationAction(ISD::CONCAT_VECTORS, MVT::v16i1, Legal);
1443 setOperationAction(ISD::SETCC, MVT::v16i1, Custom);
1444 setOperationAction(ISD::SETCC, MVT::v8i1, Custom);
1446 setOperationAction(ISD::MUL, MVT::v8i64, Custom);
1448 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v8i1, Custom);
1449 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v16i1, Custom);
1450 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v16i1, Custom);
1451 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i1, Custom);
1452 setOperationAction(ISD::BUILD_VECTOR, MVT::v8i1, Custom);
1453 setOperationAction(ISD::BUILD_VECTOR, MVT::v16i1, Custom);
1454 setOperationAction(ISD::SELECT, MVT::v8f64, Custom);
1455 setOperationAction(ISD::SELECT, MVT::v8i64, Custom);
1456 setOperationAction(ISD::SELECT, MVT::v16f32, Custom);
1457 setOperationAction(ISD::SELECT, MVT::v16i1, Custom);
1458 setOperationAction(ISD::SELECT, MVT::v8i1, Custom);
1460 setOperationAction(ISD::SMAX, MVT::v16i32, Legal);
1461 setOperationAction(ISD::SMAX, MVT::v8i64, Legal);
1462 setOperationAction(ISD::UMAX, MVT::v16i32, Legal);
1463 setOperationAction(ISD::UMAX, MVT::v8i64, Legal);
1464 setOperationAction(ISD::SMIN, MVT::v16i32, Legal);
1465 setOperationAction(ISD::SMIN, MVT::v8i64, Legal);
1466 setOperationAction(ISD::UMIN, MVT::v16i32, Legal);
1467 setOperationAction(ISD::UMIN, MVT::v8i64, Legal);
1469 setOperationAction(ISD::ADD, MVT::v8i64, Legal);
1470 setOperationAction(ISD::ADD, MVT::v16i32, Legal);
1472 setOperationAction(ISD::SUB, MVT::v8i64, Legal);
1473 setOperationAction(ISD::SUB, MVT::v16i32, Legal);
1475 setOperationAction(ISD::MUL, MVT::v16i32, Legal);
1477 setOperationAction(ISD::SRL, MVT::v8i64, Custom);
1478 setOperationAction(ISD::SRL, MVT::v16i32, Custom);
1480 setOperationAction(ISD::SHL, MVT::v8i64, Custom);
1481 setOperationAction(ISD::SHL, MVT::v16i32, Custom);
1483 setOperationAction(ISD::SRA, MVT::v8i64, Custom);
1484 setOperationAction(ISD::SRA, MVT::v16i32, Custom);
1486 setOperationAction(ISD::AND, MVT::v8i64, Legal);
1487 setOperationAction(ISD::OR, MVT::v8i64, Legal);
1488 setOperationAction(ISD::XOR, MVT::v8i64, Legal);
1489 setOperationAction(ISD::AND, MVT::v16i32, Legal);
1490 setOperationAction(ISD::OR, MVT::v16i32, Legal);
1491 setOperationAction(ISD::XOR, MVT::v16i32, Legal);
1493 if (Subtarget->hasCDI()) {
1494 setOperationAction(ISD::CTLZ, MVT::v8i64, Legal);
1495 setOperationAction(ISD::CTLZ, MVT::v16i32, Legal);
1497 if (Subtarget->hasDQI()) {
1498 setOperationAction(ISD::MUL, MVT::v2i64, Legal);
1499 setOperationAction(ISD::MUL, MVT::v4i64, Legal);
1500 setOperationAction(ISD::MUL, MVT::v8i64, Legal);
1502 // Custom lower several nodes.
1503 for (MVT VT : MVT::vector_valuetypes()) {
1504 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
1506 setOperationAction(ISD::AND, VT, Legal);
1507 setOperationAction(ISD::OR, VT, Legal);
1508 setOperationAction(ISD::XOR, VT, Legal);
1510 if (EltSize >= 32 && VT.getSizeInBits() <= 512) {
1511 setOperationAction(ISD::MGATHER, VT, Custom);
1512 setOperationAction(ISD::MSCATTER, VT, Custom);
1514 // Extract subvector is special because the value type
1515 // (result) is 256/128-bit but the source is 512-bit wide.
1516 if (VT.is128BitVector() || VT.is256BitVector()) {
1517 setOperationAction(ISD::EXTRACT_SUBVECTOR, VT, Custom);
1519 if (VT.getVectorElementType() == MVT::i1)
1520 setOperationAction(ISD::EXTRACT_SUBVECTOR, VT, Legal);
1522 // Do not attempt to custom lower other non-512-bit vectors
1523 if (!VT.is512BitVector())
1526 if (EltSize >= 32) {
1527 setOperationAction(ISD::VECTOR_SHUFFLE, VT, Custom);
1528 setOperationAction(ISD::INSERT_VECTOR_ELT, VT, Custom);
1529 setOperationAction(ISD::BUILD_VECTOR, VT, Custom);
1530 setOperationAction(ISD::VSELECT, VT, Legal);
1531 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Custom);
1532 setOperationAction(ISD::SCALAR_TO_VECTOR, VT, Custom);
1533 setOperationAction(ISD::INSERT_SUBVECTOR, VT, Custom);
1534 setOperationAction(ISD::MLOAD, VT, Legal);
1535 setOperationAction(ISD::MSTORE, VT, Legal);
1538 for (int i = MVT::v32i8; i != MVT::v8i64; ++i) {
1539 MVT VT = (MVT::SimpleValueType)i;
1541 // Do not attempt to promote non-512-bit vectors.
1542 if (!VT.is512BitVector())
1545 setOperationAction(ISD::SELECT, VT, Promote);
1546 AddPromotedToType (ISD::SELECT, VT, MVT::v8i64);
1550 if (!Subtarget->useSoftFloat() && Subtarget->hasBWI()) {
1551 addRegisterClass(MVT::v32i16, &X86::VR512RegClass);
1552 addRegisterClass(MVT::v64i8, &X86::VR512RegClass);
1554 addRegisterClass(MVT::v32i1, &X86::VK32RegClass);
1555 addRegisterClass(MVT::v64i1, &X86::VK64RegClass);
1557 setOperationAction(ISD::LOAD, MVT::v32i16, Legal);
1558 setOperationAction(ISD::LOAD, MVT::v64i8, Legal);
1559 setOperationAction(ISD::SETCC, MVT::v32i1, Custom);
1560 setOperationAction(ISD::SETCC, MVT::v64i1, Custom);
1561 setOperationAction(ISD::ADD, MVT::v32i16, Legal);
1562 setOperationAction(ISD::ADD, MVT::v64i8, Legal);
1563 setOperationAction(ISD::SUB, MVT::v32i16, Legal);
1564 setOperationAction(ISD::SUB, MVT::v64i8, Legal);
1565 setOperationAction(ISD::MUL, MVT::v32i16, Legal);
1566 setOperationAction(ISD::MULHS, MVT::v32i16, Legal);
1567 setOperationAction(ISD::MULHU, MVT::v32i16, Legal);
1568 setOperationAction(ISD::CONCAT_VECTORS, MVT::v32i1, Custom);
1569 setOperationAction(ISD::CONCAT_VECTORS, MVT::v64i1, Custom);
1570 setOperationAction(ISD::INSERT_SUBVECTOR, MVT::v32i1, Custom);
1571 setOperationAction(ISD::INSERT_SUBVECTOR, MVT::v64i1, Custom);
1572 setOperationAction(ISD::SELECT, MVT::v32i1, Custom);
1573 setOperationAction(ISD::SELECT, MVT::v64i1, Custom);
1574 setOperationAction(ISD::SIGN_EXTEND, MVT::v32i8, Custom);
1575 setOperationAction(ISD::ZERO_EXTEND, MVT::v32i8, Custom);
1576 setOperationAction(ISD::SIGN_EXTEND, MVT::v32i16, Custom);
1577 setOperationAction(ISD::ZERO_EXTEND, MVT::v32i16, Custom);
1578 setOperationAction(ISD::SIGN_EXTEND, MVT::v64i8, Custom);
1579 setOperationAction(ISD::ZERO_EXTEND, MVT::v64i8, Custom);
1580 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v32i1, Custom);
1581 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v64i1, Custom);
1582 setOperationAction(ISD::VSELECT, MVT::v32i16, Legal);
1583 setOperationAction(ISD::VSELECT, MVT::v64i8, Legal);
1584 setOperationAction(ISD::TRUNCATE, MVT::v32i1, Custom);
1585 setOperationAction(ISD::TRUNCATE, MVT::v64i1, Custom);
1586 setOperationAction(ISD::TRUNCATE, MVT::v32i8, Custom);
1588 setOperationAction(ISD::SMAX, MVT::v64i8, Legal);
1589 setOperationAction(ISD::SMAX, MVT::v32i16, Legal);
1590 setOperationAction(ISD::UMAX, MVT::v64i8, Legal);
1591 setOperationAction(ISD::UMAX, MVT::v32i16, Legal);
1592 setOperationAction(ISD::SMIN, MVT::v64i8, Legal);
1593 setOperationAction(ISD::SMIN, MVT::v32i16, Legal);
1594 setOperationAction(ISD::UMIN, MVT::v64i8, Legal);
1595 setOperationAction(ISD::UMIN, MVT::v32i16, Legal);
1597 setTruncStoreAction(MVT::v32i16, MVT::v32i8, Legal);
1598 setTruncStoreAction(MVT::v16i16, MVT::v16i8, Legal);
1599 if (Subtarget->hasVLX())
1600 setTruncStoreAction(MVT::v8i16, MVT::v8i8, Legal);
1602 for (int i = MVT::v32i8; i != MVT::v8i64; ++i) {
1603 const MVT VT = (MVT::SimpleValueType)i;
1605 const unsigned EltSize = VT.getVectorElementType().getSizeInBits();
1607 // Do not attempt to promote non-512-bit vectors.
1608 if (!VT.is512BitVector())
1612 setOperationAction(ISD::BUILD_VECTOR, VT, Custom);
1613 setOperationAction(ISD::VSELECT, VT, Legal);
1618 if (!Subtarget->useSoftFloat() && Subtarget->hasVLX()) {
1619 addRegisterClass(MVT::v4i1, &X86::VK4RegClass);
1620 addRegisterClass(MVT::v2i1, &X86::VK2RegClass);
1622 setOperationAction(ISD::SETCC, MVT::v4i1, Custom);
1623 setOperationAction(ISD::SETCC, MVT::v2i1, Custom);
1624 setOperationAction(ISD::CONCAT_VECTORS, MVT::v4i1, Custom);
1625 setOperationAction(ISD::CONCAT_VECTORS, MVT::v8i1, Custom);
1626 setOperationAction(ISD::INSERT_SUBVECTOR, MVT::v8i1, Custom);
1627 setOperationAction(ISD::INSERT_SUBVECTOR, MVT::v4i1, Custom);
1628 setOperationAction(ISD::SELECT, MVT::v4i1, Custom);
1629 setOperationAction(ISD::SELECT, MVT::v2i1, Custom);
1630 setOperationAction(ISD::BUILD_VECTOR, MVT::v4i1, Custom);
1631 setOperationAction(ISD::BUILD_VECTOR, MVT::v2i1, Custom);
1633 setOperationAction(ISD::AND, MVT::v8i32, Legal);
1634 setOperationAction(ISD::OR, MVT::v8i32, Legal);
1635 setOperationAction(ISD::XOR, MVT::v8i32, Legal);
1636 setOperationAction(ISD::AND, MVT::v4i32, Legal);
1637 setOperationAction(ISD::OR, MVT::v4i32, Legal);
1638 setOperationAction(ISD::XOR, MVT::v4i32, Legal);
1639 setOperationAction(ISD::SRA, MVT::v2i64, Custom);
1640 setOperationAction(ISD::SRA, MVT::v4i64, Custom);
1642 setOperationAction(ISD::SMAX, MVT::v2i64, Legal);
1643 setOperationAction(ISD::SMAX, MVT::v4i64, Legal);
1644 setOperationAction(ISD::UMAX, MVT::v2i64, Legal);
1645 setOperationAction(ISD::UMAX, MVT::v4i64, Legal);
1646 setOperationAction(ISD::SMIN, MVT::v2i64, Legal);
1647 setOperationAction(ISD::SMIN, MVT::v4i64, Legal);
1648 setOperationAction(ISD::UMIN, MVT::v2i64, Legal);
1649 setOperationAction(ISD::UMIN, MVT::v4i64, Legal);
1652 // We want to custom lower some of our intrinsics.
1653 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
1654 setOperationAction(ISD::INTRINSIC_W_CHAIN, MVT::Other, Custom);
1655 setOperationAction(ISD::INTRINSIC_VOID, MVT::Other, Custom);
1656 if (!Subtarget->is64Bit())
1657 setOperationAction(ISD::INTRINSIC_W_CHAIN, MVT::i64, Custom);
1659 // Only custom-lower 64-bit SADDO and friends on 64-bit because we don't
1660 // handle type legalization for these operations here.
1662 // FIXME: We really should do custom legalization for addition and
1663 // subtraction on x86-32 once PR3203 is fixed. We really can't do much better
1664 // than generic legalization for 64-bit multiplication-with-overflow, though.
1665 for (unsigned i = 0, e = 3+Subtarget->is64Bit(); i != e; ++i) {
1666 // Add/Sub/Mul with overflow operations are custom lowered.
1668 setOperationAction(ISD::SADDO, VT, Custom);
1669 setOperationAction(ISD::UADDO, VT, Custom);
1670 setOperationAction(ISD::SSUBO, VT, Custom);
1671 setOperationAction(ISD::USUBO, VT, Custom);
1672 setOperationAction(ISD::SMULO, VT, Custom);
1673 setOperationAction(ISD::UMULO, VT, Custom);
1677 if (!Subtarget->is64Bit()) {
1678 // These libcalls are not available in 32-bit.
1679 setLibcallName(RTLIB::SHL_I128, nullptr);
1680 setLibcallName(RTLIB::SRL_I128, nullptr);
1681 setLibcallName(RTLIB::SRA_I128, nullptr);
1684 // Combine sin / cos into one node or libcall if possible.
1685 if (Subtarget->hasSinCos()) {
1686 setLibcallName(RTLIB::SINCOS_F32, "sincosf");
1687 setLibcallName(RTLIB::SINCOS_F64, "sincos");
1688 if (Subtarget->isTargetDarwin()) {
1689 // For MacOSX, we don't want the normal expansion of a libcall to sincos.
1690 // We want to issue a libcall to __sincos_stret to avoid memory traffic.
1691 setOperationAction(ISD::FSINCOS, MVT::f64, Custom);
1692 setOperationAction(ISD::FSINCOS, MVT::f32, Custom);
1696 if (Subtarget->isTargetWin64()) {
1697 setOperationAction(ISD::SDIV, MVT::i128, Custom);
1698 setOperationAction(ISD::UDIV, MVT::i128, Custom);
1699 setOperationAction(ISD::SREM, MVT::i128, Custom);
1700 setOperationAction(ISD::UREM, MVT::i128, Custom);
1701 setOperationAction(ISD::SDIVREM, MVT::i128, Custom);
1702 setOperationAction(ISD::UDIVREM, MVT::i128, Custom);
1705 // We have target-specific dag combine patterns for the following nodes:
1706 setTargetDAGCombine(ISD::VECTOR_SHUFFLE);
1707 setTargetDAGCombine(ISD::EXTRACT_VECTOR_ELT);
1708 setTargetDAGCombine(ISD::BITCAST);
1709 setTargetDAGCombine(ISD::VSELECT);
1710 setTargetDAGCombine(ISD::SELECT);
1711 setTargetDAGCombine(ISD::SHL);
1712 setTargetDAGCombine(ISD::SRA);
1713 setTargetDAGCombine(ISD::SRL);
1714 setTargetDAGCombine(ISD::OR);
1715 setTargetDAGCombine(ISD::AND);
1716 setTargetDAGCombine(ISD::ADD);
1717 setTargetDAGCombine(ISD::FADD);
1718 setTargetDAGCombine(ISD::FSUB);
1719 setTargetDAGCombine(ISD::FMA);
1720 setTargetDAGCombine(ISD::SUB);
1721 setTargetDAGCombine(ISD::LOAD);
1722 setTargetDAGCombine(ISD::MLOAD);
1723 setTargetDAGCombine(ISD::STORE);
1724 setTargetDAGCombine(ISD::MSTORE);
1725 setTargetDAGCombine(ISD::ZERO_EXTEND);
1726 setTargetDAGCombine(ISD::ANY_EXTEND);
1727 setTargetDAGCombine(ISD::SIGN_EXTEND);
1728 setTargetDAGCombine(ISD::SIGN_EXTEND_INREG);
1729 setTargetDAGCombine(ISD::SINT_TO_FP);
1730 setTargetDAGCombine(ISD::UINT_TO_FP);
1731 setTargetDAGCombine(ISD::SETCC);
1732 setTargetDAGCombine(ISD::BUILD_VECTOR);
1733 setTargetDAGCombine(ISD::MUL);
1734 setTargetDAGCombine(ISD::XOR);
1736 computeRegisterProperties(Subtarget->getRegisterInfo());
1738 MaxStoresPerMemset = 16; // For @llvm.memset -> sequence of stores
1739 MaxStoresPerMemsetOptSize = 8;
1740 MaxStoresPerMemcpy = 8; // For @llvm.memcpy -> sequence of stores
1741 MaxStoresPerMemcpyOptSize = 4;
1742 MaxStoresPerMemmove = 8; // For @llvm.memmove -> sequence of stores
1743 MaxStoresPerMemmoveOptSize = 4;
1744 setPrefLoopAlignment(4); // 2^4 bytes.
1746 // Predictable cmov don't hurt on atom because it's in-order.
1747 PredictableSelectIsExpensive = !Subtarget->isAtom();
1748 EnableExtLdPromotion = true;
1749 setPrefFunctionAlignment(4); // 2^4 bytes.
1751 verifyIntrinsicTables();
1754 // This has so far only been implemented for 64-bit MachO.
1755 bool X86TargetLowering::useLoadStackGuardNode() const {
1756 return Subtarget->isTargetMachO() && Subtarget->is64Bit();
1759 TargetLoweringBase::LegalizeTypeAction
1760 X86TargetLowering::getPreferredVectorAction(EVT VT) const {
1761 if (ExperimentalVectorWideningLegalization &&
1762 VT.getVectorNumElements() != 1 &&
1763 VT.getVectorElementType().getSimpleVT() != MVT::i1)
1764 return TypeWidenVector;
1766 return TargetLoweringBase::getPreferredVectorAction(VT);
1769 EVT X86TargetLowering::getSetCCResultType(const DataLayout &DL, LLVMContext &,
1772 return Subtarget->hasAVX512() ? MVT::i1: MVT::i8;
1774 const unsigned NumElts = VT.getVectorNumElements();
1775 const EVT EltVT = VT.getVectorElementType();
1776 if (VT.is512BitVector()) {
1777 if (Subtarget->hasAVX512())
1778 if (EltVT == MVT::i32 || EltVT == MVT::i64 ||
1779 EltVT == MVT::f32 || EltVT == MVT::f64)
1781 case 8: return MVT::v8i1;
1782 case 16: return MVT::v16i1;
1784 if (Subtarget->hasBWI())
1785 if (EltVT == MVT::i8 || EltVT == MVT::i16)
1787 case 32: return MVT::v32i1;
1788 case 64: return MVT::v64i1;
1792 if (VT.is256BitVector() || VT.is128BitVector()) {
1793 if (Subtarget->hasVLX())
1794 if (EltVT == MVT::i32 || EltVT == MVT::i64 ||
1795 EltVT == MVT::f32 || EltVT == MVT::f64)
1797 case 2: return MVT::v2i1;
1798 case 4: return MVT::v4i1;
1799 case 8: return MVT::v8i1;
1801 if (Subtarget->hasBWI() && Subtarget->hasVLX())
1802 if (EltVT == MVT::i8 || EltVT == MVT::i16)
1804 case 8: return MVT::v8i1;
1805 case 16: return MVT::v16i1;
1806 case 32: return MVT::v32i1;
1810 return VT.changeVectorElementTypeToInteger();
1813 /// Helper for getByValTypeAlignment to determine
1814 /// the desired ByVal argument alignment.
1815 static void getMaxByValAlign(Type *Ty, unsigned &MaxAlign) {
1818 if (VectorType *VTy = dyn_cast<VectorType>(Ty)) {
1819 if (VTy->getBitWidth() == 128)
1821 } else if (ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
1822 unsigned EltAlign = 0;
1823 getMaxByValAlign(ATy->getElementType(), EltAlign);
1824 if (EltAlign > MaxAlign)
1825 MaxAlign = EltAlign;
1826 } else if (StructType *STy = dyn_cast<StructType>(Ty)) {
1827 for (auto *EltTy : STy->elements()) {
1828 unsigned EltAlign = 0;
1829 getMaxByValAlign(EltTy, EltAlign);
1830 if (EltAlign > MaxAlign)
1831 MaxAlign = EltAlign;
1838 /// Return the desired alignment for ByVal aggregate
1839 /// function arguments in the caller parameter area. For X86, aggregates
1840 /// that contain SSE vectors are placed at 16-byte boundaries while the rest
1841 /// are at 4-byte boundaries.
1842 unsigned X86TargetLowering::getByValTypeAlignment(Type *Ty,
1843 const DataLayout &DL) const {
1844 if (Subtarget->is64Bit()) {
1845 // Max of 8 and alignment of type.
1846 unsigned TyAlign = DL.getABITypeAlignment(Ty);
1853 if (Subtarget->hasSSE1())
1854 getMaxByValAlign(Ty, Align);
1858 /// Returns the target specific optimal type for load
1859 /// and store operations as a result of memset, memcpy, and memmove
1860 /// lowering. If DstAlign is zero that means it's safe to destination
1861 /// alignment can satisfy any constraint. Similarly if SrcAlign is zero it
1862 /// means there isn't a need to check it against alignment requirement,
1863 /// probably because the source does not need to be loaded. If 'IsMemset' is
1864 /// true, that means it's expanding a memset. If 'ZeroMemset' is true, that
1865 /// means it's a memset of zero. 'MemcpyStrSrc' indicates whether the memcpy
1866 /// source is constant so it does not need to be loaded.
1867 /// It returns EVT::Other if the type should be determined using generic
1868 /// target-independent logic.
1870 X86TargetLowering::getOptimalMemOpType(uint64_t Size,
1871 unsigned DstAlign, unsigned SrcAlign,
1872 bool IsMemset, bool ZeroMemset,
1874 MachineFunction &MF) const {
1875 const Function *F = MF.getFunction();
1876 if ((!IsMemset || ZeroMemset) &&
1877 !F->hasFnAttribute(Attribute::NoImplicitFloat)) {
1879 (!Subtarget->isUnalignedMemUnder32Slow() ||
1880 ((DstAlign == 0 || DstAlign >= 16) &&
1881 (SrcAlign == 0 || SrcAlign >= 16)))) {
1883 // FIXME: Check if unaligned 32-byte accesses are slow.
1884 if (Subtarget->hasInt256())
1886 if (Subtarget->hasFp256())
1889 if (Subtarget->hasSSE2())
1891 if (Subtarget->hasSSE1())
1893 } else if (!MemcpyStrSrc && Size >= 8 &&
1894 !Subtarget->is64Bit() &&
1895 Subtarget->hasSSE2()) {
1896 // Do not use f64 to lower memcpy if source is string constant. It's
1897 // better to use i32 to avoid the loads.
1901 // This is a compromise. If we reach here, unaligned accesses may be slow on
1902 // this target. However, creating smaller, aligned accesses could be even
1903 // slower and would certainly be a lot more code.
1904 if (Subtarget->is64Bit() && Size >= 8)
1909 bool X86TargetLowering::isSafeMemOpType(MVT VT) const {
1911 return X86ScalarSSEf32;
1912 else if (VT == MVT::f64)
1913 return X86ScalarSSEf64;
1918 X86TargetLowering::allowsMisalignedMemoryAccesses(EVT VT,
1923 if (VT.getSizeInBits() == 256)
1924 *Fast = !Subtarget->isUnalignedMem32Slow();
1926 *Fast = !Subtarget->isUnalignedMemUnder32Slow();
1931 /// Return the entry encoding for a jump table in the
1932 /// current function. The returned value is a member of the
1933 /// MachineJumpTableInfo::JTEntryKind enum.
1934 unsigned X86TargetLowering::getJumpTableEncoding() const {
1935 // In GOT pic mode, each entry in the jump table is emitted as a @GOTOFF
1937 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
1938 Subtarget->isPICStyleGOT())
1939 return MachineJumpTableInfo::EK_Custom32;
1941 // Otherwise, use the normal jump table encoding heuristics.
1942 return TargetLowering::getJumpTableEncoding();
1945 bool X86TargetLowering::useSoftFloat() const {
1946 return Subtarget->useSoftFloat();
1950 X86TargetLowering::LowerCustomJumpTableEntry(const MachineJumpTableInfo *MJTI,
1951 const MachineBasicBlock *MBB,
1952 unsigned uid,MCContext &Ctx) const{
1953 assert(MBB->getParent()->getTarget().getRelocationModel() == Reloc::PIC_ &&
1954 Subtarget->isPICStyleGOT());
1955 // In 32-bit ELF systems, our jump table entries are formed with @GOTOFF
1957 return MCSymbolRefExpr::create(MBB->getSymbol(),
1958 MCSymbolRefExpr::VK_GOTOFF, Ctx);
1961 /// Returns relocation base for the given PIC jumptable.
1962 SDValue X86TargetLowering::getPICJumpTableRelocBase(SDValue Table,
1963 SelectionDAG &DAG) const {
1964 if (!Subtarget->is64Bit())
1965 // This doesn't have SDLoc associated with it, but is not really the
1966 // same as a Register.
1967 return DAG.getNode(X86ISD::GlobalBaseReg, SDLoc(),
1968 getPointerTy(DAG.getDataLayout()));
1972 /// This returns the relocation base for the given PIC jumptable,
1973 /// the same as getPICJumpTableRelocBase, but as an MCExpr.
1974 const MCExpr *X86TargetLowering::
1975 getPICJumpTableRelocBaseExpr(const MachineFunction *MF, unsigned JTI,
1976 MCContext &Ctx) const {
1977 // X86-64 uses RIP relative addressing based on the jump table label.
1978 if (Subtarget->isPICStyleRIPRel())
1979 return TargetLowering::getPICJumpTableRelocBaseExpr(MF, JTI, Ctx);
1981 // Otherwise, the reference is relative to the PIC base.
1982 return MCSymbolRefExpr::create(MF->getPICBaseSymbol(), Ctx);
1985 std::pair<const TargetRegisterClass *, uint8_t>
1986 X86TargetLowering::findRepresentativeClass(const TargetRegisterInfo *TRI,
1988 const TargetRegisterClass *RRC = nullptr;
1990 switch (VT.SimpleTy) {
1992 return TargetLowering::findRepresentativeClass(TRI, VT);
1993 case MVT::i8: case MVT::i16: case MVT::i32: case MVT::i64:
1994 RRC = Subtarget->is64Bit() ? &X86::GR64RegClass : &X86::GR32RegClass;
1997 RRC = &X86::VR64RegClass;
1999 case MVT::f32: case MVT::f64:
2000 case MVT::v16i8: case MVT::v8i16: case MVT::v4i32: case MVT::v2i64:
2001 case MVT::v4f32: case MVT::v2f64:
2002 case MVT::v32i8: case MVT::v8i32: case MVT::v4i64: case MVT::v8f32:
2004 RRC = &X86::VR128RegClass;
2007 return std::make_pair(RRC, Cost);
2010 bool X86TargetLowering::getStackCookieLocation(unsigned &AddressSpace,
2011 unsigned &Offset) const {
2012 if (!Subtarget->isTargetLinux())
2015 if (Subtarget->is64Bit()) {
2016 // %fs:0x28, unless we're using a Kernel code model, in which case it's %gs:
2018 if (getTargetMachine().getCodeModel() == CodeModel::Kernel)
2030 bool X86TargetLowering::isNoopAddrSpaceCast(unsigned SrcAS,
2031 unsigned DestAS) const {
2032 assert(SrcAS != DestAS && "Expected different address spaces!");
2034 return SrcAS < 256 && DestAS < 256;
2037 //===----------------------------------------------------------------------===//
2038 // Return Value Calling Convention Implementation
2039 //===----------------------------------------------------------------------===//
2041 #include "X86GenCallingConv.inc"
2044 X86TargetLowering::CanLowerReturn(CallingConv::ID CallConv,
2045 MachineFunction &MF, bool isVarArg,
2046 const SmallVectorImpl<ISD::OutputArg> &Outs,
2047 LLVMContext &Context) const {
2048 SmallVector<CCValAssign, 16> RVLocs;
2049 CCState CCInfo(CallConv, isVarArg, MF, RVLocs, Context);
2050 return CCInfo.CheckReturn(Outs, RetCC_X86);
2053 const MCPhysReg *X86TargetLowering::getScratchRegisters(CallingConv::ID) const {
2054 static const MCPhysReg ScratchRegs[] = { X86::R11, 0 };
2059 X86TargetLowering::LowerReturn(SDValue Chain,
2060 CallingConv::ID CallConv, bool isVarArg,
2061 const SmallVectorImpl<ISD::OutputArg> &Outs,
2062 const SmallVectorImpl<SDValue> &OutVals,
2063 SDLoc dl, SelectionDAG &DAG) const {
2064 MachineFunction &MF = DAG.getMachineFunction();
2065 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
2067 SmallVector<CCValAssign, 16> RVLocs;
2068 CCState CCInfo(CallConv, isVarArg, MF, RVLocs, *DAG.getContext());
2069 CCInfo.AnalyzeReturn(Outs, RetCC_X86);
2072 SmallVector<SDValue, 6> RetOps;
2073 RetOps.push_back(Chain); // Operand #0 = Chain (updated below)
2074 // Operand #1 = Bytes To Pop
2075 RetOps.push_back(DAG.getTargetConstant(FuncInfo->getBytesToPopOnReturn(), dl,
2078 // Copy the result values into the output registers.
2079 for (unsigned i = 0; i != RVLocs.size(); ++i) {
2080 CCValAssign &VA = RVLocs[i];
2081 assert(VA.isRegLoc() && "Can only return in registers!");
2082 SDValue ValToCopy = OutVals[i];
2083 EVT ValVT = ValToCopy.getValueType();
2085 // Promote values to the appropriate types.
2086 if (VA.getLocInfo() == CCValAssign::SExt)
2087 ValToCopy = DAG.getNode(ISD::SIGN_EXTEND, dl, VA.getLocVT(), ValToCopy);
2088 else if (VA.getLocInfo() == CCValAssign::ZExt)
2089 ValToCopy = DAG.getNode(ISD::ZERO_EXTEND, dl, VA.getLocVT(), ValToCopy);
2090 else if (VA.getLocInfo() == CCValAssign::AExt) {
2091 if (ValVT.isVector() && ValVT.getScalarType() == MVT::i1)
2092 ValToCopy = DAG.getNode(ISD::SIGN_EXTEND, dl, VA.getLocVT(), ValToCopy);
2094 ValToCopy = DAG.getNode(ISD::ANY_EXTEND, dl, VA.getLocVT(), ValToCopy);
2096 else if (VA.getLocInfo() == CCValAssign::BCvt)
2097 ValToCopy = DAG.getBitcast(VA.getLocVT(), ValToCopy);
2099 assert(VA.getLocInfo() != CCValAssign::FPExt &&
2100 "Unexpected FP-extend for return value.");
2102 // If this is x86-64, and we disabled SSE, we can't return FP values,
2103 // or SSE or MMX vectors.
2104 if ((ValVT == MVT::f32 || ValVT == MVT::f64 ||
2105 VA.getLocReg() == X86::XMM0 || VA.getLocReg() == X86::XMM1) &&
2106 (Subtarget->is64Bit() && !Subtarget->hasSSE1())) {
2107 report_fatal_error("SSE register return with SSE disabled");
2109 // Likewise we can't return F64 values with SSE1 only. gcc does so, but
2110 // llvm-gcc has never done it right and no one has noticed, so this
2111 // should be OK for now.
2112 if (ValVT == MVT::f64 &&
2113 (Subtarget->is64Bit() && !Subtarget->hasSSE2()))
2114 report_fatal_error("SSE2 register return with SSE2 disabled");
2116 // Returns in ST0/ST1 are handled specially: these are pushed as operands to
2117 // the RET instruction and handled by the FP Stackifier.
2118 if (VA.getLocReg() == X86::FP0 ||
2119 VA.getLocReg() == X86::FP1) {
2120 // If this is a copy from an xmm register to ST(0), use an FPExtend to
2121 // change the value to the FP stack register class.
2122 if (isScalarFPTypeInSSEReg(VA.getValVT()))
2123 ValToCopy = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f80, ValToCopy);
2124 RetOps.push_back(ValToCopy);
2125 // Don't emit a copytoreg.
2129 // 64-bit vector (MMX) values are returned in XMM0 / XMM1 except for v1i64
2130 // which is returned in RAX / RDX.
2131 if (Subtarget->is64Bit()) {
2132 if (ValVT == MVT::x86mmx) {
2133 if (VA.getLocReg() == X86::XMM0 || VA.getLocReg() == X86::XMM1) {
2134 ValToCopy = DAG.getBitcast(MVT::i64, ValToCopy);
2135 ValToCopy = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i64,
2137 // If we don't have SSE2 available, convert to v4f32 so the generated
2138 // register is legal.
2139 if (!Subtarget->hasSSE2())
2140 ValToCopy = DAG.getBitcast(MVT::v4f32, ValToCopy);
2145 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), ValToCopy, Flag);
2146 Flag = Chain.getValue(1);
2147 RetOps.push_back(DAG.getRegister(VA.getLocReg(), VA.getLocVT()));
2150 // All x86 ABIs require that for returning structs by value we copy
2151 // the sret argument into %rax/%eax (depending on ABI) for the return.
2152 // We saved the argument into a virtual register in the entry block,
2153 // so now we copy the value out and into %rax/%eax.
2155 // Checking Function.hasStructRetAttr() here is insufficient because the IR
2156 // may not have an explicit sret argument. If FuncInfo.CanLowerReturn is
2157 // false, then an sret argument may be implicitly inserted in the SelDAG. In
2158 // either case FuncInfo->setSRetReturnReg() will have been called.
2159 if (unsigned SRetReg = FuncInfo->getSRetReturnReg()) {
2160 SDValue Val = DAG.getCopyFromReg(Chain, dl, SRetReg,
2161 getPointerTy(MF.getDataLayout()));
2164 = (Subtarget->is64Bit() && !Subtarget->isTarget64BitILP32()) ?
2165 X86::RAX : X86::EAX;
2166 Chain = DAG.getCopyToReg(Chain, dl, RetValReg, Val, Flag);
2167 Flag = Chain.getValue(1);
2169 // RAX/EAX now acts like a return value.
2171 DAG.getRegister(RetValReg, getPointerTy(DAG.getDataLayout())));
2174 RetOps[0] = Chain; // Update chain.
2176 // Add the flag if we have it.
2178 RetOps.push_back(Flag);
2180 return DAG.getNode(X86ISD::RET_FLAG, dl, MVT::Other, RetOps);
2183 bool X86TargetLowering::isUsedByReturnOnly(SDNode *N, SDValue &Chain) const {
2184 if (N->getNumValues() != 1)
2186 if (!N->hasNUsesOfValue(1, 0))
2189 SDValue TCChain = Chain;
2190 SDNode *Copy = *N->use_begin();
2191 if (Copy->getOpcode() == ISD::CopyToReg) {
2192 // If the copy has a glue operand, we conservatively assume it isn't safe to
2193 // perform a tail call.
2194 if (Copy->getOperand(Copy->getNumOperands()-1).getValueType() == MVT::Glue)
2196 TCChain = Copy->getOperand(0);
2197 } else if (Copy->getOpcode() != ISD::FP_EXTEND)
2200 bool HasRet = false;
2201 for (SDNode::use_iterator UI = Copy->use_begin(), UE = Copy->use_end();
2203 if (UI->getOpcode() != X86ISD::RET_FLAG)
2205 // If we are returning more than one value, we can definitely
2206 // not make a tail call see PR19530
2207 if (UI->getNumOperands() > 4)
2209 if (UI->getNumOperands() == 4 &&
2210 UI->getOperand(UI->getNumOperands()-1).getValueType() != MVT::Glue)
2223 X86TargetLowering::getTypeForExtArgOrReturn(LLVMContext &Context, EVT VT,
2224 ISD::NodeType ExtendKind) const {
2226 // TODO: Is this also valid on 32-bit?
2227 if (Subtarget->is64Bit() && VT == MVT::i1 && ExtendKind == ISD::ZERO_EXTEND)
2228 ReturnMVT = MVT::i8;
2230 ReturnMVT = MVT::i32;
2232 EVT MinVT = getRegisterType(Context, ReturnMVT);
2233 return VT.bitsLT(MinVT) ? MinVT : VT;
2236 /// Lower the result values of a call into the
2237 /// appropriate copies out of appropriate physical registers.
2240 X86TargetLowering::LowerCallResult(SDValue Chain, SDValue InFlag,
2241 CallingConv::ID CallConv, bool isVarArg,
2242 const SmallVectorImpl<ISD::InputArg> &Ins,
2243 SDLoc dl, SelectionDAG &DAG,
2244 SmallVectorImpl<SDValue> &InVals) const {
2246 // Assign locations to each value returned by this call.
2247 SmallVector<CCValAssign, 16> RVLocs;
2248 bool Is64Bit = Subtarget->is64Bit();
2249 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(), RVLocs,
2251 CCInfo.AnalyzeCallResult(Ins, RetCC_X86);
2253 // Copy all of the result registers out of their specified physreg.
2254 for (unsigned i = 0, e = RVLocs.size(); i != e; ++i) {
2255 CCValAssign &VA = RVLocs[i];
2256 EVT CopyVT = VA.getLocVT();
2258 // If this is x86-64, and we disabled SSE, we can't return FP values
2259 if ((CopyVT == MVT::f32 || CopyVT == MVT::f64) &&
2260 ((Is64Bit || Ins[i].Flags.isInReg()) && !Subtarget->hasSSE1())) {
2261 report_fatal_error("SSE register return with SSE disabled");
2264 // If we prefer to use the value in xmm registers, copy it out as f80 and
2265 // use a truncate to move it from fp stack reg to xmm reg.
2266 bool RoundAfterCopy = false;
2267 if ((VA.getLocReg() == X86::FP0 || VA.getLocReg() == X86::FP1) &&
2268 isScalarFPTypeInSSEReg(VA.getValVT())) {
2270 RoundAfterCopy = (CopyVT != VA.getLocVT());
2273 Chain = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(),
2274 CopyVT, InFlag).getValue(1);
2275 SDValue Val = Chain.getValue(0);
2278 Val = DAG.getNode(ISD::FP_ROUND, dl, VA.getValVT(), Val,
2279 // This truncation won't change the value.
2280 DAG.getIntPtrConstant(1, dl));
2282 if (VA.isExtInLoc() && VA.getValVT().getScalarType() == MVT::i1)
2283 Val = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), Val);
2285 InFlag = Chain.getValue(2);
2286 InVals.push_back(Val);
2292 //===----------------------------------------------------------------------===//
2293 // C & StdCall & Fast Calling Convention implementation
2294 //===----------------------------------------------------------------------===//
2295 // StdCall calling convention seems to be standard for many Windows' API
2296 // routines and around. It differs from C calling convention just a little:
2297 // callee should clean up the stack, not caller. Symbols should be also
2298 // decorated in some fancy way :) It doesn't support any vector arguments.
2299 // For info on fast calling convention see Fast Calling Convention (tail call)
2300 // implementation LowerX86_32FastCCCallTo.
2302 /// CallIsStructReturn - Determines whether a call uses struct return
2304 enum StructReturnType {
2309 static StructReturnType
2310 callIsStructReturn(const SmallVectorImpl<ISD::OutputArg> &Outs) {
2312 return NotStructReturn;
2314 const ISD::ArgFlagsTy &Flags = Outs[0].Flags;
2315 if (!Flags.isSRet())
2316 return NotStructReturn;
2317 if (Flags.isInReg())
2318 return RegStructReturn;
2319 return StackStructReturn;
2322 /// Determines whether a function uses struct return semantics.
2323 static StructReturnType
2324 argsAreStructReturn(const SmallVectorImpl<ISD::InputArg> &Ins) {
2326 return NotStructReturn;
2328 const ISD::ArgFlagsTy &Flags = Ins[0].Flags;
2329 if (!Flags.isSRet())
2330 return NotStructReturn;
2331 if (Flags.isInReg())
2332 return RegStructReturn;
2333 return StackStructReturn;
2336 /// Make a copy of an aggregate at address specified by "Src" to address
2337 /// "Dst" with size and alignment information specified by the specific
2338 /// parameter attribute. The copy will be passed as a byval function parameter.
2340 CreateCopyOfByValArgument(SDValue Src, SDValue Dst, SDValue Chain,
2341 ISD::ArgFlagsTy Flags, SelectionDAG &DAG,
2343 SDValue SizeNode = DAG.getConstant(Flags.getByValSize(), dl, MVT::i32);
2345 return DAG.getMemcpy(Chain, dl, Dst, Src, SizeNode, Flags.getByValAlign(),
2346 /*isVolatile*/false, /*AlwaysInline=*/true,
2347 /*isTailCall*/false,
2348 MachinePointerInfo(), MachinePointerInfo());
2351 /// Return true if the calling convention is one that
2352 /// supports tail call optimization.
2353 static bool IsTailCallConvention(CallingConv::ID CC) {
2354 return (CC == CallingConv::Fast || CC == CallingConv::GHC ||
2355 CC == CallingConv::HiPE);
2358 /// \brief Return true if the calling convention is a C calling convention.
2359 static bool IsCCallConvention(CallingConv::ID CC) {
2360 return (CC == CallingConv::C || CC == CallingConv::X86_64_Win64 ||
2361 CC == CallingConv::X86_64_SysV);
2364 bool X86TargetLowering::mayBeEmittedAsTailCall(CallInst *CI) const {
2366 CI->getParent()->getParent()->getFnAttribute("disable-tail-calls");
2367 if (!CI->isTailCall() || Attr.getValueAsString() == "true")
2371 CallingConv::ID CalleeCC = CS.getCallingConv();
2372 if (!IsTailCallConvention(CalleeCC) && !IsCCallConvention(CalleeCC))
2378 /// Return true if the function is being made into
2379 /// a tailcall target by changing its ABI.
2380 static bool FuncIsMadeTailCallSafe(CallingConv::ID CC,
2381 bool GuaranteedTailCallOpt) {
2382 return GuaranteedTailCallOpt && IsTailCallConvention(CC);
2386 X86TargetLowering::LowerMemArgument(SDValue Chain,
2387 CallingConv::ID CallConv,
2388 const SmallVectorImpl<ISD::InputArg> &Ins,
2389 SDLoc dl, SelectionDAG &DAG,
2390 const CCValAssign &VA,
2391 MachineFrameInfo *MFI,
2393 // Create the nodes corresponding to a load from this parameter slot.
2394 ISD::ArgFlagsTy Flags = Ins[i].Flags;
2395 bool AlwaysUseMutable = FuncIsMadeTailCallSafe(
2396 CallConv, DAG.getTarget().Options.GuaranteedTailCallOpt);
2397 bool isImmutable = !AlwaysUseMutable && !Flags.isByVal();
2400 // If value is passed by pointer we have address passed instead of the value
2402 bool ExtendedInMem = VA.isExtInLoc() &&
2403 VA.getValVT().getScalarType() == MVT::i1;
2405 if (VA.getLocInfo() == CCValAssign::Indirect || ExtendedInMem)
2406 ValVT = VA.getLocVT();
2408 ValVT = VA.getValVT();
2410 // FIXME: For now, all byval parameter objects are marked mutable. This can be
2411 // changed with more analysis.
2412 // In case of tail call optimization mark all arguments mutable. Since they
2413 // could be overwritten by lowering of arguments in case of a tail call.
2414 if (Flags.isByVal()) {
2415 unsigned Bytes = Flags.getByValSize();
2416 if (Bytes == 0) Bytes = 1; // Don't create zero-sized stack objects.
2417 int FI = MFI->CreateFixedObject(Bytes, VA.getLocMemOffset(), isImmutable);
2418 return DAG.getFrameIndex(FI, getPointerTy(DAG.getDataLayout()));
2420 int FI = MFI->CreateFixedObject(ValVT.getSizeInBits()/8,
2421 VA.getLocMemOffset(), isImmutable);
2422 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy(DAG.getDataLayout()));
2423 SDValue Val = DAG.getLoad(
2424 ValVT, dl, Chain, FIN,
2425 MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), FI), false,
2427 return ExtendedInMem ?
2428 DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), Val) : Val;
2432 // FIXME: Get this from tablegen.
2433 static ArrayRef<MCPhysReg> get64BitArgumentGPRs(CallingConv::ID CallConv,
2434 const X86Subtarget *Subtarget) {
2435 assert(Subtarget->is64Bit());
2437 if (Subtarget->isCallingConvWin64(CallConv)) {
2438 static const MCPhysReg GPR64ArgRegsWin64[] = {
2439 X86::RCX, X86::RDX, X86::R8, X86::R9
2441 return makeArrayRef(std::begin(GPR64ArgRegsWin64), std::end(GPR64ArgRegsWin64));
2444 static const MCPhysReg GPR64ArgRegs64Bit[] = {
2445 X86::RDI, X86::RSI, X86::RDX, X86::RCX, X86::R8, X86::R9
2447 return makeArrayRef(std::begin(GPR64ArgRegs64Bit), std::end(GPR64ArgRegs64Bit));
2450 // FIXME: Get this from tablegen.
2451 static ArrayRef<MCPhysReg> get64BitArgumentXMMs(MachineFunction &MF,
2452 CallingConv::ID CallConv,
2453 const X86Subtarget *Subtarget) {
2454 assert(Subtarget->is64Bit());
2455 if (Subtarget->isCallingConvWin64(CallConv)) {
2456 // The XMM registers which might contain var arg parameters are shadowed
2457 // in their paired GPR. So we only need to save the GPR to their home
2459 // TODO: __vectorcall will change this.
2463 const Function *Fn = MF.getFunction();
2464 bool NoImplicitFloatOps = Fn->hasFnAttribute(Attribute::NoImplicitFloat);
2465 bool isSoftFloat = Subtarget->useSoftFloat();
2466 assert(!(isSoftFloat && NoImplicitFloatOps) &&
2467 "SSE register cannot be used when SSE is disabled!");
2468 if (isSoftFloat || NoImplicitFloatOps || !Subtarget->hasSSE1())
2469 // Kernel mode asks for SSE to be disabled, so there are no XMM argument
2473 static const MCPhysReg XMMArgRegs64Bit[] = {
2474 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
2475 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
2477 return makeArrayRef(std::begin(XMMArgRegs64Bit), std::end(XMMArgRegs64Bit));
2481 X86TargetLowering::LowerFormalArguments(SDValue Chain,
2482 CallingConv::ID CallConv,
2484 const SmallVectorImpl<ISD::InputArg> &Ins,
2487 SmallVectorImpl<SDValue> &InVals)
2489 MachineFunction &MF = DAG.getMachineFunction();
2490 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
2491 const TargetFrameLowering &TFI = *Subtarget->getFrameLowering();
2493 const Function* Fn = MF.getFunction();
2494 if (Fn->hasExternalLinkage() &&
2495 Subtarget->isTargetCygMing() &&
2496 Fn->getName() == "main")
2497 FuncInfo->setForceFramePointer(true);
2499 MachineFrameInfo *MFI = MF.getFrameInfo();
2500 bool Is64Bit = Subtarget->is64Bit();
2501 bool IsWin64 = Subtarget->isCallingConvWin64(CallConv);
2503 assert(!(isVarArg && IsTailCallConvention(CallConv)) &&
2504 "Var args not supported with calling convention fastcc, ghc or hipe");
2506 // Assign locations to all of the incoming arguments.
2507 SmallVector<CCValAssign, 16> ArgLocs;
2508 CCState CCInfo(CallConv, isVarArg, MF, ArgLocs, *DAG.getContext());
2510 // Allocate shadow area for Win64
2512 CCInfo.AllocateStack(32, 8);
2514 CCInfo.AnalyzeFormalArguments(Ins, CC_X86);
2516 unsigned LastVal = ~0U;
2518 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2519 CCValAssign &VA = ArgLocs[i];
2520 // TODO: If an arg is passed in two places (e.g. reg and stack), skip later
2522 assert(VA.getValNo() != LastVal &&
2523 "Don't support value assigned to multiple locs yet");
2525 LastVal = VA.getValNo();
2527 if (VA.isRegLoc()) {
2528 EVT RegVT = VA.getLocVT();
2529 const TargetRegisterClass *RC;
2530 if (RegVT == MVT::i32)
2531 RC = &X86::GR32RegClass;
2532 else if (Is64Bit && RegVT == MVT::i64)
2533 RC = &X86::GR64RegClass;
2534 else if (RegVT == MVT::f32)
2535 RC = &X86::FR32RegClass;
2536 else if (RegVT == MVT::f64)
2537 RC = &X86::FR64RegClass;
2538 else if (RegVT.is512BitVector())
2539 RC = &X86::VR512RegClass;
2540 else if (RegVT.is256BitVector())
2541 RC = &X86::VR256RegClass;
2542 else if (RegVT.is128BitVector())
2543 RC = &X86::VR128RegClass;
2544 else if (RegVT == MVT::x86mmx)
2545 RC = &X86::VR64RegClass;
2546 else if (RegVT == MVT::i1)
2547 RC = &X86::VK1RegClass;
2548 else if (RegVT == MVT::v8i1)
2549 RC = &X86::VK8RegClass;
2550 else if (RegVT == MVT::v16i1)
2551 RC = &X86::VK16RegClass;
2552 else if (RegVT == MVT::v32i1)
2553 RC = &X86::VK32RegClass;
2554 else if (RegVT == MVT::v64i1)
2555 RC = &X86::VK64RegClass;
2557 llvm_unreachable("Unknown argument type!");
2559 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
2560 ArgValue = DAG.getCopyFromReg(Chain, dl, Reg, RegVT);
2562 // If this is an 8 or 16-bit value, it is really passed promoted to 32
2563 // bits. Insert an assert[sz]ext to capture this, then truncate to the
2565 if (VA.getLocInfo() == CCValAssign::SExt)
2566 ArgValue = DAG.getNode(ISD::AssertSext, dl, RegVT, ArgValue,
2567 DAG.getValueType(VA.getValVT()));
2568 else if (VA.getLocInfo() == CCValAssign::ZExt)
2569 ArgValue = DAG.getNode(ISD::AssertZext, dl, RegVT, ArgValue,
2570 DAG.getValueType(VA.getValVT()));
2571 else if (VA.getLocInfo() == CCValAssign::BCvt)
2572 ArgValue = DAG.getBitcast(VA.getValVT(), ArgValue);
2574 if (VA.isExtInLoc()) {
2575 // Handle MMX values passed in XMM regs.
2576 if (RegVT.isVector() && VA.getValVT().getScalarType() != MVT::i1)
2577 ArgValue = DAG.getNode(X86ISD::MOVDQ2Q, dl, VA.getValVT(), ArgValue);
2579 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
2582 assert(VA.isMemLoc());
2583 ArgValue = LowerMemArgument(Chain, CallConv, Ins, dl, DAG, VA, MFI, i);
2586 // If value is passed via pointer - do a load.
2587 if (VA.getLocInfo() == CCValAssign::Indirect)
2588 ArgValue = DAG.getLoad(VA.getValVT(), dl, Chain, ArgValue,
2589 MachinePointerInfo(), false, false, false, 0);
2591 InVals.push_back(ArgValue);
2594 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2595 // All x86 ABIs require that for returning structs by value we copy the
2596 // sret argument into %rax/%eax (depending on ABI) for the return. Save
2597 // the argument into a virtual register so that we can access it from the
2599 if (Ins[i].Flags.isSRet()) {
2600 unsigned Reg = FuncInfo->getSRetReturnReg();
2602 MVT PtrTy = getPointerTy(DAG.getDataLayout());
2603 Reg = MF.getRegInfo().createVirtualRegister(getRegClassFor(PtrTy));
2604 FuncInfo->setSRetReturnReg(Reg);
2606 SDValue Copy = DAG.getCopyToReg(DAG.getEntryNode(), dl, Reg, InVals[i]);
2607 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Copy, Chain);
2612 unsigned StackSize = CCInfo.getNextStackOffset();
2613 // Align stack specially for tail calls.
2614 if (FuncIsMadeTailCallSafe(CallConv,
2615 MF.getTarget().Options.GuaranteedTailCallOpt))
2616 StackSize = GetAlignedArgumentStackSize(StackSize, DAG);
2618 // If the function takes variable number of arguments, make a frame index for
2619 // the start of the first vararg value... for expansion of llvm.va_start. We
2620 // can skip this if there are no va_start calls.
2621 if (MFI->hasVAStart() &&
2622 (Is64Bit || (CallConv != CallingConv::X86_FastCall &&
2623 CallConv != CallingConv::X86_ThisCall))) {
2624 FuncInfo->setVarArgsFrameIndex(
2625 MFI->CreateFixedObject(1, StackSize, true));
2628 MachineModuleInfo &MMI = MF.getMMI();
2629 const Function *WinEHParent = nullptr;
2630 if (MMI.hasWinEHFuncInfo(Fn))
2631 WinEHParent = MMI.getWinEHParent(Fn);
2632 bool IsWinEHOutlined = WinEHParent && WinEHParent != Fn;
2633 bool IsWinEHParent = WinEHParent && WinEHParent == Fn;
2635 // Figure out if XMM registers are in use.
2636 assert(!(Subtarget->useSoftFloat() &&
2637 Fn->hasFnAttribute(Attribute::NoImplicitFloat)) &&
2638 "SSE register cannot be used when SSE is disabled!");
2640 // 64-bit calling conventions support varargs and register parameters, so we
2641 // have to do extra work to spill them in the prologue.
2642 if (Is64Bit && isVarArg && MFI->hasVAStart()) {
2643 // Find the first unallocated argument registers.
2644 ArrayRef<MCPhysReg> ArgGPRs = get64BitArgumentGPRs(CallConv, Subtarget);
2645 ArrayRef<MCPhysReg> ArgXMMs = get64BitArgumentXMMs(MF, CallConv, Subtarget);
2646 unsigned NumIntRegs = CCInfo.getFirstUnallocated(ArgGPRs);
2647 unsigned NumXMMRegs = CCInfo.getFirstUnallocated(ArgXMMs);
2648 assert(!(NumXMMRegs && !Subtarget->hasSSE1()) &&
2649 "SSE register cannot be used when SSE is disabled!");
2651 // Gather all the live in physical registers.
2652 SmallVector<SDValue, 6> LiveGPRs;
2653 SmallVector<SDValue, 8> LiveXMMRegs;
2655 for (MCPhysReg Reg : ArgGPRs.slice(NumIntRegs)) {
2656 unsigned GPR = MF.addLiveIn(Reg, &X86::GR64RegClass);
2658 DAG.getCopyFromReg(Chain, dl, GPR, MVT::i64));
2660 if (!ArgXMMs.empty()) {
2661 unsigned AL = MF.addLiveIn(X86::AL, &X86::GR8RegClass);
2662 ALVal = DAG.getCopyFromReg(Chain, dl, AL, MVT::i8);
2663 for (MCPhysReg Reg : ArgXMMs.slice(NumXMMRegs)) {
2664 unsigned XMMReg = MF.addLiveIn(Reg, &X86::VR128RegClass);
2665 LiveXMMRegs.push_back(
2666 DAG.getCopyFromReg(Chain, dl, XMMReg, MVT::v4f32));
2671 // Get to the caller-allocated home save location. Add 8 to account
2672 // for the return address.
2673 int HomeOffset = TFI.getOffsetOfLocalArea() + 8;
2674 FuncInfo->setRegSaveFrameIndex(
2675 MFI->CreateFixedObject(1, NumIntRegs * 8 + HomeOffset, false));
2676 // Fixup to set vararg frame on shadow area (4 x i64).
2678 FuncInfo->setVarArgsFrameIndex(FuncInfo->getRegSaveFrameIndex());
2680 // For X86-64, if there are vararg parameters that are passed via
2681 // registers, then we must store them to their spots on the stack so
2682 // they may be loaded by deferencing the result of va_next.
2683 FuncInfo->setVarArgsGPOffset(NumIntRegs * 8);
2684 FuncInfo->setVarArgsFPOffset(ArgGPRs.size() * 8 + NumXMMRegs * 16);
2685 FuncInfo->setRegSaveFrameIndex(MFI->CreateStackObject(
2686 ArgGPRs.size() * 8 + ArgXMMs.size() * 16, 16, false));
2689 // Store the integer parameter registers.
2690 SmallVector<SDValue, 8> MemOps;
2691 SDValue RSFIN = DAG.getFrameIndex(FuncInfo->getRegSaveFrameIndex(),
2692 getPointerTy(DAG.getDataLayout()));
2693 unsigned Offset = FuncInfo->getVarArgsGPOffset();
2694 for (SDValue Val : LiveGPRs) {
2695 SDValue FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(DAG.getDataLayout()),
2696 RSFIN, DAG.getIntPtrConstant(Offset, dl));
2698 DAG.getStore(Val.getValue(1), dl, Val, FIN,
2699 MachinePointerInfo::getFixedStack(
2700 DAG.getMachineFunction(),
2701 FuncInfo->getRegSaveFrameIndex(), Offset),
2703 MemOps.push_back(Store);
2707 if (!ArgXMMs.empty() && NumXMMRegs != ArgXMMs.size()) {
2708 // Now store the XMM (fp + vector) parameter registers.
2709 SmallVector<SDValue, 12> SaveXMMOps;
2710 SaveXMMOps.push_back(Chain);
2711 SaveXMMOps.push_back(ALVal);
2712 SaveXMMOps.push_back(DAG.getIntPtrConstant(
2713 FuncInfo->getRegSaveFrameIndex(), dl));
2714 SaveXMMOps.push_back(DAG.getIntPtrConstant(
2715 FuncInfo->getVarArgsFPOffset(), dl));
2716 SaveXMMOps.insert(SaveXMMOps.end(), LiveXMMRegs.begin(),
2718 MemOps.push_back(DAG.getNode(X86ISD::VASTART_SAVE_XMM_REGS, dl,
2719 MVT::Other, SaveXMMOps));
2722 if (!MemOps.empty())
2723 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, MemOps);
2724 } else if (IsWin64 && IsWinEHOutlined) {
2725 // Get to the caller-allocated home save location. Add 8 to account
2726 // for the return address.
2727 int HomeOffset = TFI.getOffsetOfLocalArea() + 8;
2728 FuncInfo->setRegSaveFrameIndex(MFI->CreateFixedObject(
2729 /*Size=*/1, /*SPOffset=*/HomeOffset + 8, /*Immutable=*/false));
2731 MMI.getWinEHFuncInfo(Fn)
2732 .CatchHandlerParentFrameObjIdx[const_cast<Function *>(Fn)] =
2733 FuncInfo->getRegSaveFrameIndex();
2735 // Store the second integer parameter (rdx) into rsp+16 relative to the
2736 // stack pointer at the entry of the function.
2737 SDValue RSFIN = DAG.getFrameIndex(FuncInfo->getRegSaveFrameIndex(),
2738 getPointerTy(DAG.getDataLayout()));
2739 unsigned GPR = MF.addLiveIn(X86::RDX, &X86::GR64RegClass);
2740 SDValue Val = DAG.getCopyFromReg(Chain, dl, GPR, MVT::i64);
2741 Chain = DAG.getStore(
2742 Val.getValue(1), dl, Val, RSFIN,
2743 MachinePointerInfo::getFixedStack(DAG.getMachineFunction(),
2744 FuncInfo->getRegSaveFrameIndex()),
2745 /*isVolatile=*/true, /*isNonTemporal=*/false, /*Alignment=*/0);
2748 if (isVarArg && MFI->hasMustTailInVarArgFunc()) {
2749 // Find the largest legal vector type.
2750 MVT VecVT = MVT::Other;
2751 // FIXME: Only some x86_32 calling conventions support AVX512.
2752 if (Subtarget->hasAVX512() &&
2753 (Is64Bit || (CallConv == CallingConv::X86_VectorCall ||
2754 CallConv == CallingConv::Intel_OCL_BI)))
2755 VecVT = MVT::v16f32;
2756 else if (Subtarget->hasAVX())
2758 else if (Subtarget->hasSSE2())
2761 // We forward some GPRs and some vector types.
2762 SmallVector<MVT, 2> RegParmTypes;
2763 MVT IntVT = Is64Bit ? MVT::i64 : MVT::i32;
2764 RegParmTypes.push_back(IntVT);
2765 if (VecVT != MVT::Other)
2766 RegParmTypes.push_back(VecVT);
2768 // Compute the set of forwarded registers. The rest are scratch.
2769 SmallVectorImpl<ForwardedRegister> &Forwards =
2770 FuncInfo->getForwardedMustTailRegParms();
2771 CCInfo.analyzeMustTailForwardedRegisters(Forwards, RegParmTypes, CC_X86);
2773 // Conservatively forward AL on x86_64, since it might be used for varargs.
2774 if (Is64Bit && !CCInfo.isAllocated(X86::AL)) {
2775 unsigned ALVReg = MF.addLiveIn(X86::AL, &X86::GR8RegClass);
2776 Forwards.push_back(ForwardedRegister(ALVReg, X86::AL, MVT::i8));
2779 // Copy all forwards from physical to virtual registers.
2780 for (ForwardedRegister &F : Forwards) {
2781 // FIXME: Can we use a less constrained schedule?
2782 SDValue RegVal = DAG.getCopyFromReg(Chain, dl, F.VReg, F.VT);
2783 F.VReg = MF.getRegInfo().createVirtualRegister(getRegClassFor(F.VT));
2784 Chain = DAG.getCopyToReg(Chain, dl, F.VReg, RegVal);
2788 // Some CCs need callee pop.
2789 if (X86::isCalleePop(CallConv, Is64Bit, isVarArg,
2790 MF.getTarget().Options.GuaranteedTailCallOpt)) {
2791 FuncInfo->setBytesToPopOnReturn(StackSize); // Callee pops everything.
2793 FuncInfo->setBytesToPopOnReturn(0); // Callee pops nothing.
2794 // If this is an sret function, the return should pop the hidden pointer.
2795 if (!Is64Bit && !IsTailCallConvention(CallConv) &&
2796 !Subtarget->getTargetTriple().isOSMSVCRT() &&
2797 argsAreStructReturn(Ins) == StackStructReturn)
2798 FuncInfo->setBytesToPopOnReturn(4);
2802 // RegSaveFrameIndex is X86-64 only.
2803 FuncInfo->setRegSaveFrameIndex(0xAAAAAAA);
2804 if (CallConv == CallingConv::X86_FastCall ||
2805 CallConv == CallingConv::X86_ThisCall)
2806 // fastcc functions can't have varargs.
2807 FuncInfo->setVarArgsFrameIndex(0xAAAAAAA);
2810 FuncInfo->setArgumentStackSize(StackSize);
2812 if (IsWinEHParent) {
2814 int UnwindHelpFI = MFI->CreateStackObject(8, 8, /*isSS=*/false);
2815 SDValue StackSlot = DAG.getFrameIndex(UnwindHelpFI, MVT::i64);
2816 MMI.getWinEHFuncInfo(MF.getFunction()).UnwindHelpFrameIdx = UnwindHelpFI;
2817 SDValue Neg2 = DAG.getConstant(-2, dl, MVT::i64);
2818 Chain = DAG.getStore(Chain, dl, Neg2, StackSlot,
2819 MachinePointerInfo::getFixedStack(
2820 DAG.getMachineFunction(), UnwindHelpFI),
2821 /*isVolatile=*/true,
2822 /*isNonTemporal=*/false, /*Alignment=*/0);
2824 // Functions using Win32 EH are considered to have opaque SP adjustments
2825 // to force local variables to be addressed from the frame or base
2827 MFI->setHasOpaqueSPAdjustment(true);
2835 X86TargetLowering::LowerMemOpCallTo(SDValue Chain,
2836 SDValue StackPtr, SDValue Arg,
2837 SDLoc dl, SelectionDAG &DAG,
2838 const CCValAssign &VA,
2839 ISD::ArgFlagsTy Flags) const {
2840 unsigned LocMemOffset = VA.getLocMemOffset();
2841 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset, dl);
2842 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(DAG.getDataLayout()),
2844 if (Flags.isByVal())
2845 return CreateCopyOfByValArgument(Arg, PtrOff, Chain, Flags, DAG, dl);
2847 return DAG.getStore(
2848 Chain, dl, Arg, PtrOff,
2849 MachinePointerInfo::getStack(DAG.getMachineFunction(), LocMemOffset),
2853 /// Emit a load of return address if tail call
2854 /// optimization is performed and it is required.
2856 X86TargetLowering::EmitTailCallLoadRetAddr(SelectionDAG &DAG,
2857 SDValue &OutRetAddr, SDValue Chain,
2858 bool IsTailCall, bool Is64Bit,
2859 int FPDiff, SDLoc dl) const {
2860 // Adjust the Return address stack slot.
2861 EVT VT = getPointerTy(DAG.getDataLayout());
2862 OutRetAddr = getReturnAddressFrameIndex(DAG);
2864 // Load the "old" Return address.
2865 OutRetAddr = DAG.getLoad(VT, dl, Chain, OutRetAddr, MachinePointerInfo(),
2866 false, false, false, 0);
2867 return SDValue(OutRetAddr.getNode(), 1);
2870 /// Emit a store of the return address if tail call
2871 /// optimization is performed and it is required (FPDiff!=0).
2872 static SDValue EmitTailCallStoreRetAddr(SelectionDAG &DAG, MachineFunction &MF,
2873 SDValue Chain, SDValue RetAddrFrIdx,
2874 EVT PtrVT, unsigned SlotSize,
2875 int FPDiff, SDLoc dl) {
2876 // Store the return address to the appropriate stack slot.
2877 if (!FPDiff) return Chain;
2878 // Calculate the new stack slot for the return address.
2879 int NewReturnAddrFI =
2880 MF.getFrameInfo()->CreateFixedObject(SlotSize, (int64_t)FPDiff - SlotSize,
2882 SDValue NewRetAddrFrIdx = DAG.getFrameIndex(NewReturnAddrFI, PtrVT);
2883 Chain = DAG.getStore(Chain, dl, RetAddrFrIdx, NewRetAddrFrIdx,
2884 MachinePointerInfo::getFixedStack(
2885 DAG.getMachineFunction(), NewReturnAddrFI),
2890 /// Returns a vector_shuffle mask for an movs{s|d}, movd
2891 /// operation of specified width.
2892 static SDValue getMOVL(SelectionDAG &DAG, SDLoc dl, EVT VT, SDValue V1,
2894 unsigned NumElems = VT.getVectorNumElements();
2895 SmallVector<int, 8> Mask;
2896 Mask.push_back(NumElems);
2897 for (unsigned i = 1; i != NumElems; ++i)
2899 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
2903 X86TargetLowering::LowerCall(TargetLowering::CallLoweringInfo &CLI,
2904 SmallVectorImpl<SDValue> &InVals) const {
2905 SelectionDAG &DAG = CLI.DAG;
2907 SmallVectorImpl<ISD::OutputArg> &Outs = CLI.Outs;
2908 SmallVectorImpl<SDValue> &OutVals = CLI.OutVals;
2909 SmallVectorImpl<ISD::InputArg> &Ins = CLI.Ins;
2910 SDValue Chain = CLI.Chain;
2911 SDValue Callee = CLI.Callee;
2912 CallingConv::ID CallConv = CLI.CallConv;
2913 bool &isTailCall = CLI.IsTailCall;
2914 bool isVarArg = CLI.IsVarArg;
2916 MachineFunction &MF = DAG.getMachineFunction();
2917 bool Is64Bit = Subtarget->is64Bit();
2918 bool IsWin64 = Subtarget->isCallingConvWin64(CallConv);
2919 StructReturnType SR = callIsStructReturn(Outs);
2920 bool IsSibcall = false;
2921 X86MachineFunctionInfo *X86Info = MF.getInfo<X86MachineFunctionInfo>();
2922 auto Attr = MF.getFunction()->getFnAttribute("disable-tail-calls");
2924 if (Attr.getValueAsString() == "true")
2927 if (Subtarget->isPICStyleGOT() &&
2928 !MF.getTarget().Options.GuaranteedTailCallOpt) {
2929 // If we are using a GOT, disable tail calls to external symbols with
2930 // default visibility. Tail calling such a symbol requires using a GOT
2931 // relocation, which forces early binding of the symbol. This breaks code
2932 // that require lazy function symbol resolution. Using musttail or
2933 // GuaranteedTailCallOpt will override this.
2934 GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee);
2935 if (!G || (!G->getGlobal()->hasLocalLinkage() &&
2936 G->getGlobal()->hasDefaultVisibility()))
2940 bool IsMustTail = CLI.CS && CLI.CS->isMustTailCall();
2942 // Force this to be a tail call. The verifier rules are enough to ensure
2943 // that we can lower this successfully without moving the return address
2946 } else if (isTailCall) {
2947 // Check if it's really possible to do a tail call.
2948 isTailCall = IsEligibleForTailCallOptimization(Callee, CallConv,
2949 isVarArg, SR != NotStructReturn,
2950 MF.getFunction()->hasStructRetAttr(), CLI.RetTy,
2951 Outs, OutVals, Ins, DAG);
2953 // Sibcalls are automatically detected tailcalls which do not require
2955 if (!MF.getTarget().Options.GuaranteedTailCallOpt && isTailCall)
2962 assert(!(isVarArg && IsTailCallConvention(CallConv)) &&
2963 "Var args not supported with calling convention fastcc, ghc or hipe");
2965 // Analyze operands of the call, assigning locations to each operand.
2966 SmallVector<CCValAssign, 16> ArgLocs;
2967 CCState CCInfo(CallConv, isVarArg, MF, ArgLocs, *DAG.getContext());
2969 // Allocate shadow area for Win64
2971 CCInfo.AllocateStack(32, 8);
2973 CCInfo.AnalyzeCallOperands(Outs, CC_X86);
2975 // Get a count of how many bytes are to be pushed on the stack.
2976 unsigned NumBytes = CCInfo.getNextStackOffset();
2978 // This is a sibcall. The memory operands are available in caller's
2979 // own caller's stack.
2981 else if (MF.getTarget().Options.GuaranteedTailCallOpt &&
2982 IsTailCallConvention(CallConv))
2983 NumBytes = GetAlignedArgumentStackSize(NumBytes, DAG);
2986 if (isTailCall && !IsSibcall && !IsMustTail) {
2987 // Lower arguments at fp - stackoffset + fpdiff.
2988 unsigned NumBytesCallerPushed = X86Info->getBytesToPopOnReturn();
2990 FPDiff = NumBytesCallerPushed - NumBytes;
2992 // Set the delta of movement of the returnaddr stackslot.
2993 // But only set if delta is greater than previous delta.
2994 if (FPDiff < X86Info->getTCReturnAddrDelta())
2995 X86Info->setTCReturnAddrDelta(FPDiff);
2998 unsigned NumBytesToPush = NumBytes;
2999 unsigned NumBytesToPop = NumBytes;
3001 // If we have an inalloca argument, all stack space has already been allocated
3002 // for us and be right at the top of the stack. We don't support multiple
3003 // arguments passed in memory when using inalloca.
3004 if (!Outs.empty() && Outs.back().Flags.isInAlloca()) {
3006 if (!ArgLocs.back().isMemLoc())
3007 report_fatal_error("cannot use inalloca attribute on a register "
3009 if (ArgLocs.back().getLocMemOffset() != 0)
3010 report_fatal_error("any parameter with the inalloca attribute must be "
3011 "the only memory argument");
3015 Chain = DAG.getCALLSEQ_START(
3016 Chain, DAG.getIntPtrConstant(NumBytesToPush, dl, true), dl);
3018 SDValue RetAddrFrIdx;
3019 // Load return address for tail calls.
3020 if (isTailCall && FPDiff)
3021 Chain = EmitTailCallLoadRetAddr(DAG, RetAddrFrIdx, Chain, isTailCall,
3022 Is64Bit, FPDiff, dl);
3024 SmallVector<std::pair<unsigned, SDValue>, 8> RegsToPass;
3025 SmallVector<SDValue, 8> MemOpChains;
3028 // Walk the register/memloc assignments, inserting copies/loads. In the case
3029 // of tail call optimization arguments are handle later.
3030 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
3031 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
3032 // Skip inalloca arguments, they have already been written.
3033 ISD::ArgFlagsTy Flags = Outs[i].Flags;
3034 if (Flags.isInAlloca())
3037 CCValAssign &VA = ArgLocs[i];
3038 EVT RegVT = VA.getLocVT();
3039 SDValue Arg = OutVals[i];
3040 bool isByVal = Flags.isByVal();
3042 // Promote the value if needed.
3043 switch (VA.getLocInfo()) {
3044 default: llvm_unreachable("Unknown loc info!");
3045 case CCValAssign::Full: break;
3046 case CCValAssign::SExt:
3047 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, RegVT, Arg);
3049 case CCValAssign::ZExt:
3050 Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, RegVT, Arg);
3052 case CCValAssign::AExt:
3053 if (Arg.getValueType().isVector() &&
3054 Arg.getValueType().getScalarType() == MVT::i1)
3055 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, RegVT, Arg);
3056 else if (RegVT.is128BitVector()) {
3057 // Special case: passing MMX values in XMM registers.
3058 Arg = DAG.getBitcast(MVT::i64, Arg);
3059 Arg = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i64, Arg);
3060 Arg = getMOVL(DAG, dl, MVT::v2i64, DAG.getUNDEF(MVT::v2i64), Arg);
3062 Arg = DAG.getNode(ISD::ANY_EXTEND, dl, RegVT, Arg);
3064 case CCValAssign::BCvt:
3065 Arg = DAG.getBitcast(RegVT, Arg);
3067 case CCValAssign::Indirect: {
3068 // Store the argument.
3069 SDValue SpillSlot = DAG.CreateStackTemporary(VA.getValVT());
3070 int FI = cast<FrameIndexSDNode>(SpillSlot)->getIndex();
3071 Chain = DAG.getStore(
3072 Chain, dl, Arg, SpillSlot,
3073 MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), FI),
3080 if (VA.isRegLoc()) {
3081 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
3082 if (isVarArg && IsWin64) {
3083 // Win64 ABI requires argument XMM reg to be copied to the corresponding
3084 // shadow reg if callee is a varargs function.
3085 unsigned ShadowReg = 0;
3086 switch (VA.getLocReg()) {
3087 case X86::XMM0: ShadowReg = X86::RCX; break;
3088 case X86::XMM1: ShadowReg = X86::RDX; break;
3089 case X86::XMM2: ShadowReg = X86::R8; break;
3090 case X86::XMM3: ShadowReg = X86::R9; break;
3093 RegsToPass.push_back(std::make_pair(ShadowReg, Arg));
3095 } else if (!IsSibcall && (!isTailCall || isByVal)) {
3096 assert(VA.isMemLoc());
3097 if (!StackPtr.getNode())
3098 StackPtr = DAG.getCopyFromReg(Chain, dl, RegInfo->getStackRegister(),
3099 getPointerTy(DAG.getDataLayout()));
3100 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Arg,
3101 dl, DAG, VA, Flags));
3105 if (!MemOpChains.empty())
3106 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, MemOpChains);
3108 if (Subtarget->isPICStyleGOT()) {
3109 // ELF / PIC requires GOT in the EBX register before function calls via PLT
3112 RegsToPass.push_back(std::make_pair(
3113 unsigned(X86::EBX), DAG.getNode(X86ISD::GlobalBaseReg, SDLoc(),
3114 getPointerTy(DAG.getDataLayout()))));
3116 // If we are tail calling and generating PIC/GOT style code load the
3117 // address of the callee into ECX. The value in ecx is used as target of
3118 // the tail jump. This is done to circumvent the ebx/callee-saved problem
3119 // for tail calls on PIC/GOT architectures. Normally we would just put the
3120 // address of GOT into ebx and then call target@PLT. But for tail calls
3121 // ebx would be restored (since ebx is callee saved) before jumping to the
3124 // Note: The actual moving to ECX is done further down.
3125 GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee);
3126 if (G && !G->getGlobal()->hasLocalLinkage() &&
3127 G->getGlobal()->hasDefaultVisibility())
3128 Callee = LowerGlobalAddress(Callee, DAG);
3129 else if (isa<ExternalSymbolSDNode>(Callee))
3130 Callee = LowerExternalSymbol(Callee, DAG);
3134 if (Is64Bit && isVarArg && !IsWin64 && !IsMustTail) {
3135 // From AMD64 ABI document:
3136 // For calls that may call functions that use varargs or stdargs
3137 // (prototype-less calls or calls to functions containing ellipsis (...) in
3138 // the declaration) %al is used as hidden argument to specify the number
3139 // of SSE registers used. The contents of %al do not need to match exactly
3140 // the number of registers, but must be an ubound on the number of SSE
3141 // registers used and is in the range 0 - 8 inclusive.
3143 // Count the number of XMM registers allocated.
3144 static const MCPhysReg XMMArgRegs[] = {
3145 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
3146 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
3148 unsigned NumXMMRegs = CCInfo.getFirstUnallocated(XMMArgRegs);
3149 assert((Subtarget->hasSSE1() || !NumXMMRegs)
3150 && "SSE registers cannot be used when SSE is disabled");
3152 RegsToPass.push_back(std::make_pair(unsigned(X86::AL),
3153 DAG.getConstant(NumXMMRegs, dl,
3157 if (isVarArg && IsMustTail) {
3158 const auto &Forwards = X86Info->getForwardedMustTailRegParms();
3159 for (const auto &F : Forwards) {
3160 SDValue Val = DAG.getCopyFromReg(Chain, dl, F.VReg, F.VT);
3161 RegsToPass.push_back(std::make_pair(unsigned(F.PReg), Val));
3165 // For tail calls lower the arguments to the 'real' stack slots. Sibcalls
3166 // don't need this because the eligibility check rejects calls that require
3167 // shuffling arguments passed in memory.
3168 if (!IsSibcall && isTailCall) {
3169 // Force all the incoming stack arguments to be loaded from the stack
3170 // before any new outgoing arguments are stored to the stack, because the
3171 // outgoing stack slots may alias the incoming argument stack slots, and
3172 // the alias isn't otherwise explicit. This is slightly more conservative
3173 // than necessary, because it means that each store effectively depends
3174 // on every argument instead of just those arguments it would clobber.
3175 SDValue ArgChain = DAG.getStackArgumentTokenFactor(Chain);
3177 SmallVector<SDValue, 8> MemOpChains2;
3180 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
3181 CCValAssign &VA = ArgLocs[i];
3184 assert(VA.isMemLoc());
3185 SDValue Arg = OutVals[i];
3186 ISD::ArgFlagsTy Flags = Outs[i].Flags;
3187 // Skip inalloca arguments. They don't require any work.
3188 if (Flags.isInAlloca())
3190 // Create frame index.
3191 int32_t Offset = VA.getLocMemOffset()+FPDiff;
3192 uint32_t OpSize = (VA.getLocVT().getSizeInBits()+7)/8;
3193 FI = MF.getFrameInfo()->CreateFixedObject(OpSize, Offset, true);
3194 FIN = DAG.getFrameIndex(FI, getPointerTy(DAG.getDataLayout()));
3196 if (Flags.isByVal()) {
3197 // Copy relative to framepointer.
3198 SDValue Source = DAG.getIntPtrConstant(VA.getLocMemOffset(), dl);
3199 if (!StackPtr.getNode())
3200 StackPtr = DAG.getCopyFromReg(Chain, dl, RegInfo->getStackRegister(),
3201 getPointerTy(DAG.getDataLayout()));
3202 Source = DAG.getNode(ISD::ADD, dl, getPointerTy(DAG.getDataLayout()),
3205 MemOpChains2.push_back(CreateCopyOfByValArgument(Source, FIN,
3209 // Store relative to framepointer.
3210 MemOpChains2.push_back(DAG.getStore(
3211 ArgChain, dl, Arg, FIN,
3212 MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), FI),
3217 if (!MemOpChains2.empty())
3218 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, MemOpChains2);
3220 // Store the return address to the appropriate stack slot.
3221 Chain = EmitTailCallStoreRetAddr(DAG, MF, Chain, RetAddrFrIdx,
3222 getPointerTy(DAG.getDataLayout()),
3223 RegInfo->getSlotSize(), FPDiff, dl);
3226 // Build a sequence of copy-to-reg nodes chained together with token chain
3227 // and flag operands which copy the outgoing args into registers.
3229 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
3230 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
3231 RegsToPass[i].second, InFlag);
3232 InFlag = Chain.getValue(1);
3235 if (DAG.getTarget().getCodeModel() == CodeModel::Large) {
3236 assert(Is64Bit && "Large code model is only legal in 64-bit mode.");
3237 // In the 64-bit large code model, we have to make all calls
3238 // through a register, since the call instruction's 32-bit
3239 // pc-relative offset may not be large enough to hold the whole
3241 } else if (Callee->getOpcode() == ISD::GlobalAddress) {
3242 // If the callee is a GlobalAddress node (quite common, every direct call
3243 // is) turn it into a TargetGlobalAddress node so that legalize doesn't hack
3245 GlobalAddressSDNode* G = cast<GlobalAddressSDNode>(Callee);
3247 // We should use extra load for direct calls to dllimported functions in
3249 const GlobalValue *GV = G->getGlobal();
3250 if (!GV->hasDLLImportStorageClass()) {
3251 unsigned char OpFlags = 0;
3252 bool ExtraLoad = false;
3253 unsigned WrapperKind = ISD::DELETED_NODE;
3255 // On ELF targets, in both X86-64 and X86-32 mode, direct calls to
3256 // external symbols most go through the PLT in PIC mode. If the symbol
3257 // has hidden or protected visibility, or if it is static or local, then
3258 // we don't need to use the PLT - we can directly call it.
3259 if (Subtarget->isTargetELF() &&
3260 DAG.getTarget().getRelocationModel() == Reloc::PIC_ &&
3261 GV->hasDefaultVisibility() && !GV->hasLocalLinkage()) {
3262 OpFlags = X86II::MO_PLT;
3263 } else if (Subtarget->isPICStyleStubAny() &&
3264 !GV->isStrongDefinitionForLinker() &&
3265 (!Subtarget->getTargetTriple().isMacOSX() ||
3266 Subtarget->getTargetTriple().isMacOSXVersionLT(10, 5))) {
3267 // PC-relative references to external symbols should go through $stub,
3268 // unless we're building with the leopard linker or later, which
3269 // automatically synthesizes these stubs.
3270 OpFlags = X86II::MO_DARWIN_STUB;
3271 } else if (Subtarget->isPICStyleRIPRel() && isa<Function>(GV) &&
3272 cast<Function>(GV)->hasFnAttribute(Attribute::NonLazyBind)) {
3273 // If the function is marked as non-lazy, generate an indirect call
3274 // which loads from the GOT directly. This avoids runtime overhead
3275 // at the cost of eager binding (and one extra byte of encoding).
3276 OpFlags = X86II::MO_GOTPCREL;
3277 WrapperKind = X86ISD::WrapperRIP;
3281 Callee = DAG.getTargetGlobalAddress(
3282 GV, dl, getPointerTy(DAG.getDataLayout()), G->getOffset(), OpFlags);
3284 // Add a wrapper if needed.
3285 if (WrapperKind != ISD::DELETED_NODE)
3286 Callee = DAG.getNode(X86ISD::WrapperRIP, dl,
3287 getPointerTy(DAG.getDataLayout()), Callee);
3288 // Add extra indirection if needed.
3290 Callee = DAG.getLoad(
3291 getPointerTy(DAG.getDataLayout()), dl, DAG.getEntryNode(), Callee,
3292 MachinePointerInfo::getGOT(DAG.getMachineFunction()), false, false,
3295 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
3296 unsigned char OpFlags = 0;
3298 // On ELF targets, in either X86-64 or X86-32 mode, direct calls to
3299 // external symbols should go through the PLT.
3300 if (Subtarget->isTargetELF() &&
3301 DAG.getTarget().getRelocationModel() == Reloc::PIC_) {
3302 OpFlags = X86II::MO_PLT;
3303 } else if (Subtarget->isPICStyleStubAny() &&
3304 (!Subtarget->getTargetTriple().isMacOSX() ||
3305 Subtarget->getTargetTriple().isMacOSXVersionLT(10, 5))) {
3306 // PC-relative references to external symbols should go through $stub,
3307 // unless we're building with the leopard linker or later, which
3308 // automatically synthesizes these stubs.
3309 OpFlags = X86II::MO_DARWIN_STUB;
3312 Callee = DAG.getTargetExternalSymbol(
3313 S->getSymbol(), getPointerTy(DAG.getDataLayout()), OpFlags);
3314 } else if (Subtarget->isTarget64BitILP32() &&
3315 Callee->getValueType(0) == MVT::i32) {
3316 // Zero-extend the 32-bit Callee address into a 64-bit according to x32 ABI
3317 Callee = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i64, Callee);
3320 // Returns a chain & a flag for retval copy to use.
3321 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
3322 SmallVector<SDValue, 8> Ops;
3324 if (!IsSibcall && isTailCall) {
3325 Chain = DAG.getCALLSEQ_END(Chain,
3326 DAG.getIntPtrConstant(NumBytesToPop, dl, true),
3327 DAG.getIntPtrConstant(0, dl, true), InFlag, dl);
3328 InFlag = Chain.getValue(1);
3331 Ops.push_back(Chain);
3332 Ops.push_back(Callee);
3335 Ops.push_back(DAG.getConstant(FPDiff, dl, MVT::i32));
3337 // Add argument registers to the end of the list so that they are known live
3339 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
3340 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
3341 RegsToPass[i].second.getValueType()));
3343 // Add a register mask operand representing the call-preserved registers.
3344 const uint32_t *Mask = RegInfo->getCallPreservedMask(MF, CallConv);
3345 assert(Mask && "Missing call preserved mask for calling convention");
3347 // If this is an invoke in a 32-bit function using an MSVC personality, assume
3348 // the function clobbers all registers. If an exception is thrown, the runtime
3349 // will not restore CSRs.
3350 // FIXME: Model this more precisely so that we can register allocate across
3351 // the normal edge and spill and fill across the exceptional edge.
3352 if (!Is64Bit && CLI.CS && CLI.CS->isInvoke()) {
3353 const Function *CallerFn = MF.getFunction();
3354 EHPersonality Pers =
3355 CallerFn->hasPersonalityFn()
3356 ? classifyEHPersonality(CallerFn->getPersonalityFn())
3357 : EHPersonality::Unknown;
3358 if (isMSVCEHPersonality(Pers))
3359 Mask = RegInfo->getNoPreservedMask();
3362 Ops.push_back(DAG.getRegisterMask(Mask));
3364 if (InFlag.getNode())
3365 Ops.push_back(InFlag);
3369 //// If this is the first return lowered for this function, add the regs
3370 //// to the liveout set for the function.
3371 // This isn't right, although it's probably harmless on x86; liveouts
3372 // should be computed from returns not tail calls. Consider a void
3373 // function making a tail call to a function returning int.
3374 MF.getFrameInfo()->setHasTailCall();
3375 return DAG.getNode(X86ISD::TC_RETURN, dl, NodeTys, Ops);
3378 Chain = DAG.getNode(X86ISD::CALL, dl, NodeTys, Ops);
3379 InFlag = Chain.getValue(1);
3381 // Create the CALLSEQ_END node.
3382 unsigned NumBytesForCalleeToPop;
3383 if (X86::isCalleePop(CallConv, Is64Bit, isVarArg,
3384 DAG.getTarget().Options.GuaranteedTailCallOpt))
3385 NumBytesForCalleeToPop = NumBytes; // Callee pops everything
3386 else if (!Is64Bit && !IsTailCallConvention(CallConv) &&
3387 !Subtarget->getTargetTriple().isOSMSVCRT() &&
3388 SR == StackStructReturn)
3389 // If this is a call to a struct-return function, the callee
3390 // pops the hidden struct pointer, so we have to push it back.
3391 // This is common for Darwin/X86, Linux & Mingw32 targets.
3392 // For MSVC Win32 targets, the caller pops the hidden struct pointer.
3393 NumBytesForCalleeToPop = 4;
3395 NumBytesForCalleeToPop = 0; // Callee pops nothing.
3397 // Returns a flag for retval copy to use.
3399 Chain = DAG.getCALLSEQ_END(Chain,
3400 DAG.getIntPtrConstant(NumBytesToPop, dl, true),
3401 DAG.getIntPtrConstant(NumBytesForCalleeToPop, dl,
3404 InFlag = Chain.getValue(1);
3407 // Handle result values, copying them out of physregs into vregs that we
3409 return LowerCallResult(Chain, InFlag, CallConv, isVarArg,
3410 Ins, dl, DAG, InVals);
3413 //===----------------------------------------------------------------------===//
3414 // Fast Calling Convention (tail call) implementation
3415 //===----------------------------------------------------------------------===//
3417 // Like std call, callee cleans arguments, convention except that ECX is
3418 // reserved for storing the tail called function address. Only 2 registers are
3419 // free for argument passing (inreg). Tail call optimization is performed
3421 // * tailcallopt is enabled
3422 // * caller/callee are fastcc
3423 // On X86_64 architecture with GOT-style position independent code only local
3424 // (within module) calls are supported at the moment.
3425 // To keep the stack aligned according to platform abi the function
3426 // GetAlignedArgumentStackSize ensures that argument delta is always multiples
3427 // of stack alignment. (Dynamic linkers need this - darwin's dyld for example)
3428 // If a tail called function callee has more arguments than the caller the
3429 // caller needs to make sure that there is room to move the RETADDR to. This is
3430 // achieved by reserving an area the size of the argument delta right after the
3431 // original RETADDR, but before the saved framepointer or the spilled registers
3432 // e.g. caller(arg1, arg2) calls callee(arg1, arg2,arg3,arg4)
3444 /// Make the stack size align e.g 16n + 12 aligned for a 16-byte align
3447 X86TargetLowering::GetAlignedArgumentStackSize(unsigned StackSize,
3448 SelectionDAG& DAG) const {
3449 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
3450 const TargetFrameLowering &TFI = *Subtarget->getFrameLowering();
3451 unsigned StackAlignment = TFI.getStackAlignment();
3452 uint64_t AlignMask = StackAlignment - 1;
3453 int64_t Offset = StackSize;
3454 unsigned SlotSize = RegInfo->getSlotSize();
3455 if ( (Offset & AlignMask) <= (StackAlignment - SlotSize) ) {
3456 // Number smaller than 12 so just add the difference.
3457 Offset += ((StackAlignment - SlotSize) - (Offset & AlignMask));
3459 // Mask out lower bits, add stackalignment once plus the 12 bytes.
3460 Offset = ((~AlignMask) & Offset) + StackAlignment +
3461 (StackAlignment-SlotSize);
3466 /// Return true if the given stack call argument is already available in the
3467 /// same position (relatively) of the caller's incoming argument stack.
3469 bool MatchingStackOffset(SDValue Arg, unsigned Offset, ISD::ArgFlagsTy Flags,
3470 MachineFrameInfo *MFI, const MachineRegisterInfo *MRI,
3471 const X86InstrInfo *TII) {
3472 unsigned Bytes = Arg.getValueType().getSizeInBits() / 8;
3474 if (Arg.getOpcode() == ISD::CopyFromReg) {
3475 unsigned VR = cast<RegisterSDNode>(Arg.getOperand(1))->getReg();
3476 if (!TargetRegisterInfo::isVirtualRegister(VR))
3478 MachineInstr *Def = MRI->getVRegDef(VR);
3481 if (!Flags.isByVal()) {
3482 if (!TII->isLoadFromStackSlot(Def, FI))
3485 unsigned Opcode = Def->getOpcode();
3486 if ((Opcode == X86::LEA32r || Opcode == X86::LEA64r ||
3487 Opcode == X86::LEA64_32r) &&
3488 Def->getOperand(1).isFI()) {
3489 FI = Def->getOperand(1).getIndex();
3490 Bytes = Flags.getByValSize();
3494 } else if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Arg)) {
3495 if (Flags.isByVal())
3496 // ByVal argument is passed in as a pointer but it's now being
3497 // dereferenced. e.g.
3498 // define @foo(%struct.X* %A) {
3499 // tail call @bar(%struct.X* byval %A)
3502 SDValue Ptr = Ld->getBasePtr();
3503 FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr);
3506 FI = FINode->getIndex();
3507 } else if (Arg.getOpcode() == ISD::FrameIndex && Flags.isByVal()) {
3508 FrameIndexSDNode *FINode = cast<FrameIndexSDNode>(Arg);
3509 FI = FINode->getIndex();
3510 Bytes = Flags.getByValSize();
3514 assert(FI != INT_MAX);
3515 if (!MFI->isFixedObjectIndex(FI))
3517 return Offset == MFI->getObjectOffset(FI) && Bytes == MFI->getObjectSize(FI);
3520 /// Check whether the call is eligible for tail call optimization. Targets
3521 /// that want to do tail call optimization should implement this function.
3523 X86TargetLowering::IsEligibleForTailCallOptimization(SDValue Callee,
3524 CallingConv::ID CalleeCC,
3526 bool isCalleeStructRet,
3527 bool isCallerStructRet,
3529 const SmallVectorImpl<ISD::OutputArg> &Outs,
3530 const SmallVectorImpl<SDValue> &OutVals,
3531 const SmallVectorImpl<ISD::InputArg> &Ins,
3532 SelectionDAG &DAG) const {
3533 if (!IsTailCallConvention(CalleeCC) && !IsCCallConvention(CalleeCC))
3536 // If -tailcallopt is specified, make fastcc functions tail-callable.
3537 const MachineFunction &MF = DAG.getMachineFunction();
3538 const Function *CallerF = MF.getFunction();
3540 // If the function return type is x86_fp80 and the callee return type is not,
3541 // then the FP_EXTEND of the call result is not a nop. It's not safe to
3542 // perform a tailcall optimization here.
3543 if (CallerF->getReturnType()->isX86_FP80Ty() && !RetTy->isX86_FP80Ty())
3546 CallingConv::ID CallerCC = CallerF->getCallingConv();
3547 bool CCMatch = CallerCC == CalleeCC;
3548 bool IsCalleeWin64 = Subtarget->isCallingConvWin64(CalleeCC);
3549 bool IsCallerWin64 = Subtarget->isCallingConvWin64(CallerCC);
3551 // Win64 functions have extra shadow space for argument homing. Don't do the
3552 // sibcall if the caller and callee have mismatched expectations for this
3554 if (IsCalleeWin64 != IsCallerWin64)
3557 if (DAG.getTarget().Options.GuaranteedTailCallOpt) {
3558 if (IsTailCallConvention(CalleeCC) && CCMatch)
3563 // Look for obvious safe cases to perform tail call optimization that do not
3564 // require ABI changes. This is what gcc calls sibcall.
3566 // Can't do sibcall if stack needs to be dynamically re-aligned. PEI needs to
3567 // emit a special epilogue.
3568 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
3569 if (RegInfo->needsStackRealignment(MF))
3572 // Also avoid sibcall optimization if either caller or callee uses struct
3573 // return semantics.
3574 if (isCalleeStructRet || isCallerStructRet)
3577 // An stdcall/thiscall caller is expected to clean up its arguments; the
3578 // callee isn't going to do that.
3579 // FIXME: this is more restrictive than needed. We could produce a tailcall
3580 // when the stack adjustment matches. For example, with a thiscall that takes
3581 // only one argument.
3582 if (!CCMatch && (CallerCC == CallingConv::X86_StdCall ||
3583 CallerCC == CallingConv::X86_ThisCall))
3586 // Do not sibcall optimize vararg calls unless all arguments are passed via
3588 if (isVarArg && !Outs.empty()) {
3590 // Optimizing for varargs on Win64 is unlikely to be safe without
3591 // additional testing.
3592 if (IsCalleeWin64 || IsCallerWin64)
3595 SmallVector<CCValAssign, 16> ArgLocs;
3596 CCState CCInfo(CalleeCC, isVarArg, DAG.getMachineFunction(), ArgLocs,
3599 CCInfo.AnalyzeCallOperands(Outs, CC_X86);
3600 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i)
3601 if (!ArgLocs[i].isRegLoc())
3605 // If the call result is in ST0 / ST1, it needs to be popped off the x87
3606 // stack. Therefore, if it's not used by the call it is not safe to optimize
3607 // this into a sibcall.
3608 bool Unused = false;
3609 for (unsigned i = 0, e = Ins.size(); i != e; ++i) {
3616 SmallVector<CCValAssign, 16> RVLocs;
3617 CCState CCInfo(CalleeCC, false, DAG.getMachineFunction(), RVLocs,
3619 CCInfo.AnalyzeCallResult(Ins, RetCC_X86);
3620 for (unsigned i = 0, e = RVLocs.size(); i != e; ++i) {
3621 CCValAssign &VA = RVLocs[i];
3622 if (VA.getLocReg() == X86::FP0 || VA.getLocReg() == X86::FP1)
3627 // If the calling conventions do not match, then we'd better make sure the
3628 // results are returned in the same way as what the caller expects.
3630 SmallVector<CCValAssign, 16> RVLocs1;
3631 CCState CCInfo1(CalleeCC, false, DAG.getMachineFunction(), RVLocs1,
3633 CCInfo1.AnalyzeCallResult(Ins, RetCC_X86);
3635 SmallVector<CCValAssign, 16> RVLocs2;
3636 CCState CCInfo2(CallerCC, false, DAG.getMachineFunction(), RVLocs2,
3638 CCInfo2.AnalyzeCallResult(Ins, RetCC_X86);
3640 if (RVLocs1.size() != RVLocs2.size())
3642 for (unsigned i = 0, e = RVLocs1.size(); i != e; ++i) {
3643 if (RVLocs1[i].isRegLoc() != RVLocs2[i].isRegLoc())
3645 if (RVLocs1[i].getLocInfo() != RVLocs2[i].getLocInfo())
3647 if (RVLocs1[i].isRegLoc()) {
3648 if (RVLocs1[i].getLocReg() != RVLocs2[i].getLocReg())
3651 if (RVLocs1[i].getLocMemOffset() != RVLocs2[i].getLocMemOffset())
3657 // If the callee takes no arguments then go on to check the results of the
3659 if (!Outs.empty()) {
3660 // Check if stack adjustment is needed. For now, do not do this if any
3661 // argument is passed on the stack.
3662 SmallVector<CCValAssign, 16> ArgLocs;
3663 CCState CCInfo(CalleeCC, isVarArg, DAG.getMachineFunction(), ArgLocs,
3666 // Allocate shadow area for Win64
3668 CCInfo.AllocateStack(32, 8);
3670 CCInfo.AnalyzeCallOperands(Outs, CC_X86);
3671 if (CCInfo.getNextStackOffset()) {
3672 MachineFunction &MF = DAG.getMachineFunction();
3673 if (MF.getInfo<X86MachineFunctionInfo>()->getBytesToPopOnReturn())
3676 // Check if the arguments are already laid out in the right way as
3677 // the caller's fixed stack objects.
3678 MachineFrameInfo *MFI = MF.getFrameInfo();
3679 const MachineRegisterInfo *MRI = &MF.getRegInfo();
3680 const X86InstrInfo *TII = Subtarget->getInstrInfo();
3681 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
3682 CCValAssign &VA = ArgLocs[i];
3683 SDValue Arg = OutVals[i];
3684 ISD::ArgFlagsTy Flags = Outs[i].Flags;
3685 if (VA.getLocInfo() == CCValAssign::Indirect)
3687 if (!VA.isRegLoc()) {
3688 if (!MatchingStackOffset(Arg, VA.getLocMemOffset(), Flags,
3695 // If the tailcall address may be in a register, then make sure it's
3696 // possible to register allocate for it. In 32-bit, the call address can
3697 // only target EAX, EDX, or ECX since the tail call must be scheduled after
3698 // callee-saved registers are restored. These happen to be the same
3699 // registers used to pass 'inreg' arguments so watch out for those.
3700 if (!Subtarget->is64Bit() &&
3701 ((!isa<GlobalAddressSDNode>(Callee) &&
3702 !isa<ExternalSymbolSDNode>(Callee)) ||
3703 DAG.getTarget().getRelocationModel() == Reloc::PIC_)) {
3704 unsigned NumInRegs = 0;
3705 // In PIC we need an extra register to formulate the address computation
3707 unsigned MaxInRegs =
3708 (DAG.getTarget().getRelocationModel() == Reloc::PIC_) ? 2 : 3;
3710 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
3711 CCValAssign &VA = ArgLocs[i];
3714 unsigned Reg = VA.getLocReg();
3717 case X86::EAX: case X86::EDX: case X86::ECX:
3718 if (++NumInRegs == MaxInRegs)
3730 X86TargetLowering::createFastISel(FunctionLoweringInfo &funcInfo,
3731 const TargetLibraryInfo *libInfo) const {
3732 return X86::createFastISel(funcInfo, libInfo);
3735 //===----------------------------------------------------------------------===//
3736 // Other Lowering Hooks
3737 //===----------------------------------------------------------------------===//
3739 static bool MayFoldLoad(SDValue Op) {
3740 return Op.hasOneUse() && ISD::isNormalLoad(Op.getNode());
3743 static bool MayFoldIntoStore(SDValue Op) {
3744 return Op.hasOneUse() && ISD::isNormalStore(*Op.getNode()->use_begin());
3747 static bool isTargetShuffle(unsigned Opcode) {
3749 default: return false;
3750 case X86ISD::BLENDI:
3751 case X86ISD::PSHUFB:
3752 case X86ISD::PSHUFD:
3753 case X86ISD::PSHUFHW:
3754 case X86ISD::PSHUFLW:
3756 case X86ISD::PALIGNR:
3757 case X86ISD::MOVLHPS:
3758 case X86ISD::MOVLHPD:
3759 case X86ISD::MOVHLPS:
3760 case X86ISD::MOVLPS:
3761 case X86ISD::MOVLPD:
3762 case X86ISD::MOVSHDUP:
3763 case X86ISD::MOVSLDUP:
3764 case X86ISD::MOVDDUP:
3767 case X86ISD::UNPCKL:
3768 case X86ISD::UNPCKH:
3769 case X86ISD::VPERMILPI:
3770 case X86ISD::VPERM2X128:
3771 case X86ISD::VPERMI:
3776 static SDValue getTargetShuffleNode(unsigned Opc, SDLoc dl, EVT VT,
3777 SDValue V1, unsigned TargetMask,
3778 SelectionDAG &DAG) {
3780 default: llvm_unreachable("Unknown x86 shuffle node");
3781 case X86ISD::PSHUFD:
3782 case X86ISD::PSHUFHW:
3783 case X86ISD::PSHUFLW:
3784 case X86ISD::VPERMILPI:
3785 case X86ISD::VPERMI:
3786 return DAG.getNode(Opc, dl, VT, V1,
3787 DAG.getConstant(TargetMask, dl, MVT::i8));
3791 static SDValue getTargetShuffleNode(unsigned Opc, SDLoc dl, EVT VT,
3792 SDValue V1, SDValue V2, SelectionDAG &DAG) {
3794 default: llvm_unreachable("Unknown x86 shuffle node");
3795 case X86ISD::MOVLHPS:
3796 case X86ISD::MOVLHPD:
3797 case X86ISD::MOVHLPS:
3798 case X86ISD::MOVLPS:
3799 case X86ISD::MOVLPD:
3802 case X86ISD::UNPCKL:
3803 case X86ISD::UNPCKH:
3804 return DAG.getNode(Opc, dl, VT, V1, V2);
3808 SDValue X86TargetLowering::getReturnAddressFrameIndex(SelectionDAG &DAG) const {
3809 MachineFunction &MF = DAG.getMachineFunction();
3810 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
3811 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
3812 int ReturnAddrIndex = FuncInfo->getRAIndex();
3814 if (ReturnAddrIndex == 0) {
3815 // Set up a frame object for the return address.
3816 unsigned SlotSize = RegInfo->getSlotSize();
3817 ReturnAddrIndex = MF.getFrameInfo()->CreateFixedObject(SlotSize,
3820 FuncInfo->setRAIndex(ReturnAddrIndex);
3823 return DAG.getFrameIndex(ReturnAddrIndex, getPointerTy(DAG.getDataLayout()));
3826 bool X86::isOffsetSuitableForCodeModel(int64_t Offset, CodeModel::Model M,
3827 bool hasSymbolicDisplacement) {
3828 // Offset should fit into 32 bit immediate field.
3829 if (!isInt<32>(Offset))
3832 // If we don't have a symbolic displacement - we don't have any extra
3834 if (!hasSymbolicDisplacement)
3837 // FIXME: Some tweaks might be needed for medium code model.
3838 if (M != CodeModel::Small && M != CodeModel::Kernel)
3841 // For small code model we assume that latest object is 16MB before end of 31
3842 // bits boundary. We may also accept pretty large negative constants knowing
3843 // that all objects are in the positive half of address space.
3844 if (M == CodeModel::Small && Offset < 16*1024*1024)
3847 // For kernel code model we know that all object resist in the negative half
3848 // of 32bits address space. We may not accept negative offsets, since they may
3849 // be just off and we may accept pretty large positive ones.
3850 if (M == CodeModel::Kernel && Offset >= 0)
3856 /// Determines whether the callee is required to pop its own arguments.
3857 /// Callee pop is necessary to support tail calls.
3858 bool X86::isCalleePop(CallingConv::ID CallingConv,
3859 bool is64Bit, bool IsVarArg, bool TailCallOpt) {
3860 switch (CallingConv) {
3863 case CallingConv::X86_StdCall:
3864 case CallingConv::X86_FastCall:
3865 case CallingConv::X86_ThisCall:
3867 case CallingConv::Fast:
3868 case CallingConv::GHC:
3869 case CallingConv::HiPE:
3876 /// \brief Return true if the condition is an unsigned comparison operation.
3877 static bool isX86CCUnsigned(unsigned X86CC) {
3879 default: llvm_unreachable("Invalid integer condition!");
3880 case X86::COND_E: return true;
3881 case X86::COND_G: return false;
3882 case X86::COND_GE: return false;
3883 case X86::COND_L: return false;
3884 case X86::COND_LE: return false;
3885 case X86::COND_NE: return true;
3886 case X86::COND_B: return true;
3887 case X86::COND_A: return true;
3888 case X86::COND_BE: return true;
3889 case X86::COND_AE: return true;
3891 llvm_unreachable("covered switch fell through?!");
3894 /// Do a one-to-one translation of a ISD::CondCode to the X86-specific
3895 /// condition code, returning the condition code and the LHS/RHS of the
3896 /// comparison to make.
3897 static unsigned TranslateX86CC(ISD::CondCode SetCCOpcode, SDLoc DL, bool isFP,
3898 SDValue &LHS, SDValue &RHS, SelectionDAG &DAG) {
3900 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS)) {
3901 if (SetCCOpcode == ISD::SETGT && RHSC->isAllOnesValue()) {
3902 // X > -1 -> X == 0, jump !sign.
3903 RHS = DAG.getConstant(0, DL, RHS.getValueType());
3904 return X86::COND_NS;
3906 if (SetCCOpcode == ISD::SETLT && RHSC->isNullValue()) {
3907 // X < 0 -> X == 0, jump on sign.
3910 if (SetCCOpcode == ISD::SETLT && RHSC->getZExtValue() == 1) {
3912 RHS = DAG.getConstant(0, DL, RHS.getValueType());
3913 return X86::COND_LE;
3917 switch (SetCCOpcode) {
3918 default: llvm_unreachable("Invalid integer condition!");
3919 case ISD::SETEQ: return X86::COND_E;
3920 case ISD::SETGT: return X86::COND_G;
3921 case ISD::SETGE: return X86::COND_GE;
3922 case ISD::SETLT: return X86::COND_L;
3923 case ISD::SETLE: return X86::COND_LE;
3924 case ISD::SETNE: return X86::COND_NE;
3925 case ISD::SETULT: return X86::COND_B;
3926 case ISD::SETUGT: return X86::COND_A;
3927 case ISD::SETULE: return X86::COND_BE;
3928 case ISD::SETUGE: return X86::COND_AE;
3932 // First determine if it is required or is profitable to flip the operands.
3934 // If LHS is a foldable load, but RHS is not, flip the condition.
3935 if (ISD::isNON_EXTLoad(LHS.getNode()) &&
3936 !ISD::isNON_EXTLoad(RHS.getNode())) {
3937 SetCCOpcode = getSetCCSwappedOperands(SetCCOpcode);
3938 std::swap(LHS, RHS);
3941 switch (SetCCOpcode) {
3947 std::swap(LHS, RHS);
3951 // On a floating point condition, the flags are set as follows:
3953 // 0 | 0 | 0 | X > Y
3954 // 0 | 0 | 1 | X < Y
3955 // 1 | 0 | 0 | X == Y
3956 // 1 | 1 | 1 | unordered
3957 switch (SetCCOpcode) {
3958 default: llvm_unreachable("Condcode should be pre-legalized away");
3960 case ISD::SETEQ: return X86::COND_E;
3961 case ISD::SETOLT: // flipped
3963 case ISD::SETGT: return X86::COND_A;
3964 case ISD::SETOLE: // flipped
3966 case ISD::SETGE: return X86::COND_AE;
3967 case ISD::SETUGT: // flipped
3969 case ISD::SETLT: return X86::COND_B;
3970 case ISD::SETUGE: // flipped
3972 case ISD::SETLE: return X86::COND_BE;
3974 case ISD::SETNE: return X86::COND_NE;
3975 case ISD::SETUO: return X86::COND_P;
3976 case ISD::SETO: return X86::COND_NP;
3978 case ISD::SETUNE: return X86::COND_INVALID;
3982 /// Is there a floating point cmov for the specific X86 condition code?
3983 /// Current x86 isa includes the following FP cmov instructions:
3984 /// fcmovb, fcomvbe, fcomve, fcmovu, fcmovae, fcmova, fcmovne, fcmovnu.
3985 static bool hasFPCMov(unsigned X86CC) {
4001 /// Returns true if the target can instruction select the
4002 /// specified FP immediate natively. If false, the legalizer will
4003 /// materialize the FP immediate as a load from a constant pool.
4004 bool X86TargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
4005 for (unsigned i = 0, e = LegalFPImmediates.size(); i != e; ++i) {
4006 if (Imm.bitwiseIsEqual(LegalFPImmediates[i]))
4012 bool X86TargetLowering::shouldReduceLoadWidth(SDNode *Load,
4013 ISD::LoadExtType ExtTy,
4015 // "ELF Handling for Thread-Local Storage" specifies that R_X86_64_GOTTPOFF
4016 // relocation target a movq or addq instruction: don't let the load shrink.
4017 SDValue BasePtr = cast<LoadSDNode>(Load)->getBasePtr();
4018 if (BasePtr.getOpcode() == X86ISD::WrapperRIP)
4019 if (const auto *GA = dyn_cast<GlobalAddressSDNode>(BasePtr.getOperand(0)))
4020 return GA->getTargetFlags() != X86II::MO_GOTTPOFF;
4024 /// \brief Returns true if it is beneficial to convert a load of a constant
4025 /// to just the constant itself.
4026 bool X86TargetLowering::shouldConvertConstantLoadToIntImm(const APInt &Imm,
4028 assert(Ty->isIntegerTy());
4030 unsigned BitSize = Ty->getPrimitiveSizeInBits();
4031 if (BitSize == 0 || BitSize > 64)
4036 bool X86TargetLowering::isExtractSubvectorCheap(EVT ResVT,
4037 unsigned Index) const {
4038 if (!isOperationLegalOrCustom(ISD::EXTRACT_SUBVECTOR, ResVT))
4041 return (Index == 0 || Index == ResVT.getVectorNumElements());
4044 bool X86TargetLowering::isCheapToSpeculateCttz() const {
4045 // Speculate cttz only if we can directly use TZCNT.
4046 return Subtarget->hasBMI();
4049 bool X86TargetLowering::isCheapToSpeculateCtlz() const {
4050 // Speculate ctlz only if we can directly use LZCNT.
4051 return Subtarget->hasLZCNT();
4054 /// Return true if every element in Mask, beginning
4055 /// from position Pos and ending in Pos+Size is undef.
4056 static bool isUndefInRange(ArrayRef<int> Mask, unsigned Pos, unsigned Size) {
4057 for (unsigned i = Pos, e = Pos + Size; i != e; ++i)
4063 /// Return true if Val is undef or if its value falls within the
4064 /// specified range (L, H].
4065 static bool isUndefOrInRange(int Val, int Low, int Hi) {
4066 return (Val < 0) || (Val >= Low && Val < Hi);
4069 /// Val is either less than zero (undef) or equal to the specified value.
4070 static bool isUndefOrEqual(int Val, int CmpVal) {
4071 return (Val < 0 || Val == CmpVal);
4074 /// Return true if every element in Mask, beginning
4075 /// from position Pos and ending in Pos+Size, falls within the specified
4076 /// sequential range (Low, Low+Size]. or is undef.
4077 static bool isSequentialOrUndefInRange(ArrayRef<int> Mask,
4078 unsigned Pos, unsigned Size, int Low) {
4079 for (unsigned i = Pos, e = Pos+Size; i != e; ++i, ++Low)
4080 if (!isUndefOrEqual(Mask[i], Low))
4085 /// Return true if the specified EXTRACT_SUBVECTOR operand specifies a vector
4086 /// extract that is suitable for instruction that extract 128 or 256 bit vectors
4087 static bool isVEXTRACTIndex(SDNode *N, unsigned vecWidth) {
4088 assert((vecWidth == 128 || vecWidth == 256) && "Unexpected vector width");
4089 if (!isa<ConstantSDNode>(N->getOperand(1).getNode()))
4092 // The index should be aligned on a vecWidth-bit boundary.
4094 cast<ConstantSDNode>(N->getOperand(1).getNode())->getZExtValue();
4096 MVT VT = N->getSimpleValueType(0);
4097 unsigned ElSize = VT.getVectorElementType().getSizeInBits();
4098 bool Result = (Index * ElSize) % vecWidth == 0;
4103 /// Return true if the specified INSERT_SUBVECTOR
4104 /// operand specifies a subvector insert that is suitable for input to
4105 /// insertion of 128 or 256-bit subvectors
4106 static bool isVINSERTIndex(SDNode *N, unsigned vecWidth) {
4107 assert((vecWidth == 128 || vecWidth == 256) && "Unexpected vector width");
4108 if (!isa<ConstantSDNode>(N->getOperand(2).getNode()))
4110 // The index should be aligned on a vecWidth-bit boundary.
4112 cast<ConstantSDNode>(N->getOperand(2).getNode())->getZExtValue();
4114 MVT VT = N->getSimpleValueType(0);
4115 unsigned ElSize = VT.getVectorElementType().getSizeInBits();
4116 bool Result = (Index * ElSize) % vecWidth == 0;
4121 bool X86::isVINSERT128Index(SDNode *N) {
4122 return isVINSERTIndex(N, 128);
4125 bool X86::isVINSERT256Index(SDNode *N) {
4126 return isVINSERTIndex(N, 256);
4129 bool X86::isVEXTRACT128Index(SDNode *N) {
4130 return isVEXTRACTIndex(N, 128);
4133 bool X86::isVEXTRACT256Index(SDNode *N) {
4134 return isVEXTRACTIndex(N, 256);
4137 static unsigned getExtractVEXTRACTImmediate(SDNode *N, unsigned vecWidth) {
4138 assert((vecWidth == 128 || vecWidth == 256) && "Unsupported vector width");
4139 if (!isa<ConstantSDNode>(N->getOperand(1).getNode()))
4140 llvm_unreachable("Illegal extract subvector for VEXTRACT");
4143 cast<ConstantSDNode>(N->getOperand(1).getNode())->getZExtValue();
4145 MVT VecVT = N->getOperand(0).getSimpleValueType();
4146 MVT ElVT = VecVT.getVectorElementType();
4148 unsigned NumElemsPerChunk = vecWidth / ElVT.getSizeInBits();
4149 return Index / NumElemsPerChunk;
4152 static unsigned getInsertVINSERTImmediate(SDNode *N, unsigned vecWidth) {
4153 assert((vecWidth == 128 || vecWidth == 256) && "Unsupported vector width");
4154 if (!isa<ConstantSDNode>(N->getOperand(2).getNode()))
4155 llvm_unreachable("Illegal insert subvector for VINSERT");
4158 cast<ConstantSDNode>(N->getOperand(2).getNode())->getZExtValue();
4160 MVT VecVT = N->getSimpleValueType(0);
4161 MVT ElVT = VecVT.getVectorElementType();
4163 unsigned NumElemsPerChunk = vecWidth / ElVT.getSizeInBits();
4164 return Index / NumElemsPerChunk;
4167 /// Return the appropriate immediate to extract the specified
4168 /// EXTRACT_SUBVECTOR index with VEXTRACTF128 and VINSERTI128 instructions.
4169 unsigned X86::getExtractVEXTRACT128Immediate(SDNode *N) {
4170 return getExtractVEXTRACTImmediate(N, 128);
4173 /// Return the appropriate immediate to extract the specified
4174 /// EXTRACT_SUBVECTOR index with VEXTRACTF64x4 and VINSERTI64x4 instructions.
4175 unsigned X86::getExtractVEXTRACT256Immediate(SDNode *N) {
4176 return getExtractVEXTRACTImmediate(N, 256);
4179 /// Return the appropriate immediate to insert at the specified
4180 /// INSERT_SUBVECTOR index with VINSERTF128 and VINSERTI128 instructions.
4181 unsigned X86::getInsertVINSERT128Immediate(SDNode *N) {
4182 return getInsertVINSERTImmediate(N, 128);
4185 /// Return the appropriate immediate to insert at the specified
4186 /// INSERT_SUBVECTOR index with VINSERTF46x4 and VINSERTI64x4 instructions.
4187 unsigned X86::getInsertVINSERT256Immediate(SDNode *N) {
4188 return getInsertVINSERTImmediate(N, 256);
4191 /// Returns true if Elt is a constant integer zero
4192 static bool isZero(SDValue V) {
4193 ConstantSDNode *C = dyn_cast<ConstantSDNode>(V);
4194 return C && C->isNullValue();
4197 /// Returns true if Elt is a constant zero or a floating point constant +0.0.
4198 bool X86::isZeroNode(SDValue Elt) {
4201 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(Elt))
4202 return CFP->getValueAPF().isPosZero();
4206 /// Returns a vector of specified type with all zero elements.
4207 static SDValue getZeroVector(EVT VT, const X86Subtarget *Subtarget,
4208 SelectionDAG &DAG, SDLoc dl) {
4209 assert(VT.isVector() && "Expected a vector type");
4211 // Always build SSE zero vectors as <4 x i32> bitcasted
4212 // to their dest type. This ensures they get CSE'd.
4214 if (VT.is128BitVector()) { // SSE
4215 if (Subtarget->hasSSE2()) { // SSE2
4216 SDValue Cst = DAG.getConstant(0, dl, MVT::i32);
4217 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
4219 SDValue Cst = DAG.getConstantFP(+0.0, dl, MVT::f32);
4220 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4f32, Cst, Cst, Cst, Cst);
4222 } else if (VT.is256BitVector()) { // AVX
4223 if (Subtarget->hasInt256()) { // AVX2
4224 SDValue Cst = DAG.getConstant(0, dl, MVT::i32);
4225 SDValue Ops[] = { Cst, Cst, Cst, Cst, Cst, Cst, Cst, Cst };
4226 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v8i32, Ops);
4228 // 256-bit logic and arithmetic instructions in AVX are all
4229 // floating-point, no support for integer ops. Emit fp zeroed vectors.
4230 SDValue Cst = DAG.getConstantFP(+0.0, dl, MVT::f32);
4231 SDValue Ops[] = { Cst, Cst, Cst, Cst, Cst, Cst, Cst, Cst };
4232 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v8f32, Ops);
4234 } else if (VT.is512BitVector()) { // AVX-512
4235 SDValue Cst = DAG.getConstant(0, dl, MVT::i32);
4236 SDValue Ops[] = { Cst, Cst, Cst, Cst, Cst, Cst, Cst, Cst,
4237 Cst, Cst, Cst, Cst, Cst, Cst, Cst, Cst };
4238 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v16i32, Ops);
4239 } else if (VT.getScalarType() == MVT::i1) {
4241 assert((Subtarget->hasBWI() || VT.getVectorNumElements() <= 16)
4242 && "Unexpected vector type");
4243 assert((Subtarget->hasVLX() || VT.getVectorNumElements() >= 8)
4244 && "Unexpected vector type");
4245 SDValue Cst = DAG.getConstant(0, dl, MVT::i1);
4246 SmallVector<SDValue, 64> Ops(VT.getVectorNumElements(), Cst);
4247 return DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Ops);
4249 llvm_unreachable("Unexpected vector type");
4251 return DAG.getBitcast(VT, Vec);
4254 static SDValue ExtractSubVector(SDValue Vec, unsigned IdxVal,
4255 SelectionDAG &DAG, SDLoc dl,
4256 unsigned vectorWidth) {
4257 assert((vectorWidth == 128 || vectorWidth == 256) &&
4258 "Unsupported vector width");
4259 EVT VT = Vec.getValueType();
4260 EVT ElVT = VT.getVectorElementType();
4261 unsigned Factor = VT.getSizeInBits()/vectorWidth;
4262 EVT ResultVT = EVT::getVectorVT(*DAG.getContext(), ElVT,
4263 VT.getVectorNumElements()/Factor);
4265 // Extract from UNDEF is UNDEF.
4266 if (Vec.getOpcode() == ISD::UNDEF)
4267 return DAG.getUNDEF(ResultVT);
4269 // Extract the relevant vectorWidth bits. Generate an EXTRACT_SUBVECTOR
4270 unsigned ElemsPerChunk = vectorWidth / ElVT.getSizeInBits();
4272 // This is the index of the first element of the vectorWidth-bit chunk
4274 unsigned NormalizedIdxVal = (((IdxVal * ElVT.getSizeInBits()) / vectorWidth)
4277 // If the input is a buildvector just emit a smaller one.
4278 if (Vec.getOpcode() == ISD::BUILD_VECTOR)
4279 return DAG.getNode(ISD::BUILD_VECTOR, dl, ResultVT,
4280 makeArrayRef(Vec->op_begin() + NormalizedIdxVal,
4283 SDValue VecIdx = DAG.getIntPtrConstant(NormalizedIdxVal, dl);
4284 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, ResultVT, Vec, VecIdx);
4287 /// Generate a DAG to grab 128-bits from a vector > 128 bits. This
4288 /// sets things up to match to an AVX VEXTRACTF128 / VEXTRACTI128
4289 /// or AVX-512 VEXTRACTF32x4 / VEXTRACTI32x4
4290 /// instructions or a simple subregister reference. Idx is an index in the
4291 /// 128 bits we want. It need not be aligned to a 128-bit boundary. That makes
4292 /// lowering EXTRACT_VECTOR_ELT operations easier.
4293 static SDValue Extract128BitVector(SDValue Vec, unsigned IdxVal,
4294 SelectionDAG &DAG, SDLoc dl) {
4295 assert((Vec.getValueType().is256BitVector() ||
4296 Vec.getValueType().is512BitVector()) && "Unexpected vector size!");
4297 return ExtractSubVector(Vec, IdxVal, DAG, dl, 128);
4300 /// Generate a DAG to grab 256-bits from a 512-bit vector.
4301 static SDValue Extract256BitVector(SDValue Vec, unsigned IdxVal,
4302 SelectionDAG &DAG, SDLoc dl) {
4303 assert(Vec.getValueType().is512BitVector() && "Unexpected vector size!");
4304 return ExtractSubVector(Vec, IdxVal, DAG, dl, 256);
4307 static SDValue InsertSubVector(SDValue Result, SDValue Vec,
4308 unsigned IdxVal, SelectionDAG &DAG,
4309 SDLoc dl, unsigned vectorWidth) {
4310 assert((vectorWidth == 128 || vectorWidth == 256) &&
4311 "Unsupported vector width");
4312 // Inserting UNDEF is Result
4313 if (Vec.getOpcode() == ISD::UNDEF)
4315 EVT VT = Vec.getValueType();
4316 EVT ElVT = VT.getVectorElementType();
4317 EVT ResultVT = Result.getValueType();
4319 // Insert the relevant vectorWidth bits.
4320 unsigned ElemsPerChunk = vectorWidth/ElVT.getSizeInBits();
4322 // This is the index of the first element of the vectorWidth-bit chunk
4324 unsigned NormalizedIdxVal = (((IdxVal * ElVT.getSizeInBits())/vectorWidth)
4327 SDValue VecIdx = DAG.getIntPtrConstant(NormalizedIdxVal, dl);
4328 return DAG.getNode(ISD::INSERT_SUBVECTOR, dl, ResultVT, Result, Vec, VecIdx);
4331 /// Generate a DAG to put 128-bits into a vector > 128 bits. This
4332 /// sets things up to match to an AVX VINSERTF128/VINSERTI128 or
4333 /// AVX-512 VINSERTF32x4/VINSERTI32x4 instructions or a
4334 /// simple superregister reference. Idx is an index in the 128 bits
4335 /// we want. It need not be aligned to a 128-bit boundary. That makes
4336 /// lowering INSERT_VECTOR_ELT operations easier.
4337 static SDValue Insert128BitVector(SDValue Result, SDValue Vec, unsigned IdxVal,
4338 SelectionDAG &DAG, SDLoc dl) {
4339 assert(Vec.getValueType().is128BitVector() && "Unexpected vector size!");
4341 // For insertion into the zero index (low half) of a 256-bit vector, it is
4342 // more efficient to generate a blend with immediate instead of an insert*128.
4343 // We are still creating an INSERT_SUBVECTOR below with an undef node to
4344 // extend the subvector to the size of the result vector. Make sure that
4345 // we are not recursing on that node by checking for undef here.
4346 if (IdxVal == 0 && Result.getValueType().is256BitVector() &&
4347 Result.getOpcode() != ISD::UNDEF) {
4348 EVT ResultVT = Result.getValueType();
4349 SDValue ZeroIndex = DAG.getIntPtrConstant(0, dl);
4350 SDValue Undef = DAG.getUNDEF(ResultVT);
4351 SDValue Vec256 = DAG.getNode(ISD::INSERT_SUBVECTOR, dl, ResultVT, Undef,
4354 // The blend instruction, and therefore its mask, depend on the data type.
4355 MVT ScalarType = ResultVT.getScalarType().getSimpleVT();
4356 if (ScalarType.isFloatingPoint()) {
4357 // Choose either vblendps (float) or vblendpd (double).
4358 unsigned ScalarSize = ScalarType.getSizeInBits();
4359 assert((ScalarSize == 64 || ScalarSize == 32) && "Unknown float type");
4360 unsigned MaskVal = (ScalarSize == 64) ? 0x03 : 0x0f;
4361 SDValue Mask = DAG.getConstant(MaskVal, dl, MVT::i8);
4362 return DAG.getNode(X86ISD::BLENDI, dl, ResultVT, Result, Vec256, Mask);
4365 const X86Subtarget &Subtarget =
4366 static_cast<const X86Subtarget &>(DAG.getSubtarget());
4368 // AVX2 is needed for 256-bit integer blend support.
4369 // Integers must be cast to 32-bit because there is only vpblendd;
4370 // vpblendw can't be used for this because it has a handicapped mask.
4372 // If we don't have AVX2, then cast to float. Using a wrong domain blend
4373 // is still more efficient than using the wrong domain vinsertf128 that
4374 // will be created by InsertSubVector().
4375 MVT CastVT = Subtarget.hasAVX2() ? MVT::v8i32 : MVT::v8f32;
4377 SDValue Mask = DAG.getConstant(0x0f, dl, MVT::i8);
4378 Vec256 = DAG.getBitcast(CastVT, Vec256);
4379 Vec256 = DAG.getNode(X86ISD::BLENDI, dl, CastVT, Result, Vec256, Mask);
4380 return DAG.getBitcast(ResultVT, Vec256);
4383 return InsertSubVector(Result, Vec, IdxVal, DAG, dl, 128);
4386 static SDValue Insert256BitVector(SDValue Result, SDValue Vec, unsigned IdxVal,
4387 SelectionDAG &DAG, SDLoc dl) {
4388 assert(Vec.getValueType().is256BitVector() && "Unexpected vector size!");
4389 return InsertSubVector(Result, Vec, IdxVal, DAG, dl, 256);
4392 /// Concat two 128-bit vectors into a 256 bit vector using VINSERTF128
4393 /// instructions. This is used because creating CONCAT_VECTOR nodes of
4394 /// BUILD_VECTORS returns a larger BUILD_VECTOR while we're trying to lower
4395 /// large BUILD_VECTORS.
4396 static SDValue Concat128BitVectors(SDValue V1, SDValue V2, EVT VT,
4397 unsigned NumElems, SelectionDAG &DAG,
4399 SDValue V = Insert128BitVector(DAG.getUNDEF(VT), V1, 0, DAG, dl);
4400 return Insert128BitVector(V, V2, NumElems/2, DAG, dl);
4403 static SDValue Concat256BitVectors(SDValue V1, SDValue V2, EVT VT,
4404 unsigned NumElems, SelectionDAG &DAG,
4406 SDValue V = Insert256BitVector(DAG.getUNDEF(VT), V1, 0, DAG, dl);
4407 return Insert256BitVector(V, V2, NumElems/2, DAG, dl);
4410 /// Returns a vector of specified type with all bits set.
4411 /// Always build ones vectors as <4 x i32> or <8 x i32>. For 256-bit types with
4412 /// no AVX2 supprt, use two <4 x i32> inserted in a <8 x i32> appropriately.
4413 /// Then bitcast to their original type, ensuring they get CSE'd.
4414 static SDValue getOnesVector(MVT VT, bool HasInt256, SelectionDAG &DAG,
4416 assert(VT.isVector() && "Expected a vector type");
4418 SDValue Cst = DAG.getConstant(~0U, dl, MVT::i32);
4420 if (VT.is256BitVector()) {
4421 if (HasInt256) { // AVX2
4422 SDValue Ops[] = { Cst, Cst, Cst, Cst, Cst, Cst, Cst, Cst };
4423 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v8i32, Ops);
4425 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
4426 Vec = Concat128BitVectors(Vec, Vec, MVT::v8i32, 8, DAG, dl);
4428 } else if (VT.is128BitVector()) {
4429 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
4431 llvm_unreachable("Unexpected vector type");
4433 return DAG.getBitcast(VT, Vec);
4436 /// Returns a vector_shuffle node for an unpackl operation.
4437 static SDValue getUnpackl(SelectionDAG &DAG, SDLoc dl, MVT VT, SDValue V1,
4439 unsigned NumElems = VT.getVectorNumElements();
4440 SmallVector<int, 8> Mask;
4441 for (unsigned i = 0, e = NumElems/2; i != e; ++i) {
4443 Mask.push_back(i + NumElems);
4445 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
4448 /// Returns a vector_shuffle node for an unpackh operation.
4449 static SDValue getUnpackh(SelectionDAG &DAG, SDLoc dl, MVT VT, SDValue V1,
4451 unsigned NumElems = VT.getVectorNumElements();
4452 SmallVector<int, 8> Mask;
4453 for (unsigned i = 0, Half = NumElems/2; i != Half; ++i) {
4454 Mask.push_back(i + Half);
4455 Mask.push_back(i + NumElems + Half);
4457 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
4460 /// Return a vector_shuffle of the specified vector of zero or undef vector.
4461 /// This produces a shuffle where the low element of V2 is swizzled into the
4462 /// zero/undef vector, landing at element Idx.
4463 /// This produces a shuffle mask like 4,1,2,3 (idx=0) or 0,1,2,4 (idx=3).
4464 static SDValue getShuffleVectorZeroOrUndef(SDValue V2, unsigned Idx,
4466 const X86Subtarget *Subtarget,
4467 SelectionDAG &DAG) {
4468 MVT VT = V2.getSimpleValueType();
4470 ? getZeroVector(VT, Subtarget, DAG, SDLoc(V2)) : DAG.getUNDEF(VT);
4471 unsigned NumElems = VT.getVectorNumElements();
4472 SmallVector<int, 16> MaskVec;
4473 for (unsigned i = 0; i != NumElems; ++i)
4474 // If this is the insertion idx, put the low elt of V2 here.
4475 MaskVec.push_back(i == Idx ? NumElems : i);
4476 return DAG.getVectorShuffle(VT, SDLoc(V2), V1, V2, &MaskVec[0]);
4479 /// Calculates the shuffle mask corresponding to the target-specific opcode.
4480 /// Returns true if the Mask could be calculated. Sets IsUnary to true if only
4481 /// uses one source. Note that this will set IsUnary for shuffles which use a
4482 /// single input multiple times, and in those cases it will
4483 /// adjust the mask to only have indices within that single input.
4484 /// FIXME: Add support for Decode*Mask functions that return SM_SentinelZero.
4485 static bool getTargetShuffleMask(SDNode *N, MVT VT,
4486 SmallVectorImpl<int> &Mask, bool &IsUnary) {
4487 unsigned NumElems = VT.getVectorNumElements();
4491 bool IsFakeUnary = false;
4492 switch(N->getOpcode()) {
4493 case X86ISD::BLENDI:
4494 ImmN = N->getOperand(N->getNumOperands()-1);
4495 DecodeBLENDMask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
4498 ImmN = N->getOperand(N->getNumOperands()-1);
4499 DecodeSHUFPMask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
4500 IsUnary = IsFakeUnary = N->getOperand(0) == N->getOperand(1);
4502 case X86ISD::UNPCKH:
4503 DecodeUNPCKHMask(VT, Mask);
4504 IsUnary = IsFakeUnary = N->getOperand(0) == N->getOperand(1);
4506 case X86ISD::UNPCKL:
4507 DecodeUNPCKLMask(VT, Mask);
4508 IsUnary = IsFakeUnary = N->getOperand(0) == N->getOperand(1);
4510 case X86ISD::MOVHLPS:
4511 DecodeMOVHLPSMask(NumElems, Mask);
4512 IsUnary = IsFakeUnary = N->getOperand(0) == N->getOperand(1);
4514 case X86ISD::MOVLHPS:
4515 DecodeMOVLHPSMask(NumElems, Mask);
4516 IsUnary = IsFakeUnary = N->getOperand(0) == N->getOperand(1);
4518 case X86ISD::PALIGNR:
4519 ImmN = N->getOperand(N->getNumOperands()-1);
4520 DecodePALIGNRMask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
4522 case X86ISD::PSHUFD:
4523 case X86ISD::VPERMILPI:
4524 ImmN = N->getOperand(N->getNumOperands()-1);
4525 DecodePSHUFMask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
4528 case X86ISD::PSHUFHW:
4529 ImmN = N->getOperand(N->getNumOperands()-1);
4530 DecodePSHUFHWMask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
4533 case X86ISD::PSHUFLW:
4534 ImmN = N->getOperand(N->getNumOperands()-1);
4535 DecodePSHUFLWMask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
4538 case X86ISD::PSHUFB: {
4540 SDValue MaskNode = N->getOperand(1);
4541 while (MaskNode->getOpcode() == ISD::BITCAST)
4542 MaskNode = MaskNode->getOperand(0);
4544 if (MaskNode->getOpcode() == ISD::BUILD_VECTOR) {
4545 // If we have a build-vector, then things are easy.
4546 EVT VT = MaskNode.getValueType();
4547 assert(VT.isVector() &&
4548 "Can't produce a non-vector with a build_vector!");
4549 if (!VT.isInteger())
4552 int NumBytesPerElement = VT.getVectorElementType().getSizeInBits() / 8;
4554 SmallVector<uint64_t, 32> RawMask;
4555 for (int i = 0, e = MaskNode->getNumOperands(); i < e; ++i) {
4556 SDValue Op = MaskNode->getOperand(i);
4557 if (Op->getOpcode() == ISD::UNDEF) {
4558 RawMask.push_back((uint64_t)SM_SentinelUndef);
4561 auto *CN = dyn_cast<ConstantSDNode>(Op.getNode());
4564 APInt MaskElement = CN->getAPIntValue();
4566 // We now have to decode the element which could be any integer size and
4567 // extract each byte of it.
4568 for (int j = 0; j < NumBytesPerElement; ++j) {
4569 // Note that this is x86 and so always little endian: the low byte is
4570 // the first byte of the mask.
4571 RawMask.push_back(MaskElement.getLoBits(8).getZExtValue());
4572 MaskElement = MaskElement.lshr(8);
4575 DecodePSHUFBMask(RawMask, Mask);
4579 auto *MaskLoad = dyn_cast<LoadSDNode>(MaskNode);
4583 SDValue Ptr = MaskLoad->getBasePtr();
4584 if (Ptr->getOpcode() == X86ISD::Wrapper ||
4585 Ptr->getOpcode() == X86ISD::WrapperRIP)
4586 Ptr = Ptr->getOperand(0);
4588 auto *MaskCP = dyn_cast<ConstantPoolSDNode>(Ptr);
4589 if (!MaskCP || MaskCP->isMachineConstantPoolEntry())
4592 if (auto *C = dyn_cast<Constant>(MaskCP->getConstVal())) {
4593 DecodePSHUFBMask(C, Mask);
4601 case X86ISD::VPERMI:
4602 ImmN = N->getOperand(N->getNumOperands()-1);
4603 DecodeVPERMMask(cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
4608 DecodeScalarMoveMask(VT, /* IsLoad */ false, Mask);
4610 case X86ISD::VPERM2X128:
4611 ImmN = N->getOperand(N->getNumOperands()-1);
4612 DecodeVPERM2X128Mask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
4613 if (Mask.empty()) return false;
4614 // Mask only contains negative index if an element is zero.
4615 if (std::any_of(Mask.begin(), Mask.end(),
4616 [](int M){ return M == SM_SentinelZero; }))
4619 case X86ISD::MOVSLDUP:
4620 DecodeMOVSLDUPMask(VT, Mask);
4623 case X86ISD::MOVSHDUP:
4624 DecodeMOVSHDUPMask(VT, Mask);
4627 case X86ISD::MOVDDUP:
4628 DecodeMOVDDUPMask(VT, Mask);
4631 case X86ISD::MOVLHPD:
4632 case X86ISD::MOVLPD:
4633 case X86ISD::MOVLPS:
4634 // Not yet implemented
4636 default: llvm_unreachable("unknown target shuffle node");
4639 // If we have a fake unary shuffle, the shuffle mask is spread across two
4640 // inputs that are actually the same node. Re-map the mask to always point
4641 // into the first input.
4644 if (M >= (int)Mask.size())
4650 /// Returns the scalar element that will make up the ith
4651 /// element of the result of the vector shuffle.
4652 static SDValue getShuffleScalarElt(SDNode *N, unsigned Index, SelectionDAG &DAG,
4655 return SDValue(); // Limit search depth.
4657 SDValue V = SDValue(N, 0);
4658 EVT VT = V.getValueType();
4659 unsigned Opcode = V.getOpcode();
4661 // Recurse into ISD::VECTOR_SHUFFLE node to find scalars.
4662 if (const ShuffleVectorSDNode *SV = dyn_cast<ShuffleVectorSDNode>(N)) {
4663 int Elt = SV->getMaskElt(Index);
4666 return DAG.getUNDEF(VT.getVectorElementType());
4668 unsigned NumElems = VT.getVectorNumElements();
4669 SDValue NewV = (Elt < (int)NumElems) ? SV->getOperand(0)
4670 : SV->getOperand(1);
4671 return getShuffleScalarElt(NewV.getNode(), Elt % NumElems, DAG, Depth+1);
4674 // Recurse into target specific vector shuffles to find scalars.
4675 if (isTargetShuffle(Opcode)) {
4676 MVT ShufVT = V.getSimpleValueType();
4677 unsigned NumElems = ShufVT.getVectorNumElements();
4678 SmallVector<int, 16> ShuffleMask;
4681 if (!getTargetShuffleMask(N, ShufVT, ShuffleMask, IsUnary))
4684 int Elt = ShuffleMask[Index];
4686 return DAG.getUNDEF(ShufVT.getVectorElementType());
4688 SDValue NewV = (Elt < (int)NumElems) ? N->getOperand(0)
4690 return getShuffleScalarElt(NewV.getNode(), Elt % NumElems, DAG,
4694 // Actual nodes that may contain scalar elements
4695 if (Opcode == ISD::BITCAST) {
4696 V = V.getOperand(0);
4697 EVT SrcVT = V.getValueType();
4698 unsigned NumElems = VT.getVectorNumElements();
4700 if (!SrcVT.isVector() || SrcVT.getVectorNumElements() != NumElems)
4704 if (V.getOpcode() == ISD::SCALAR_TO_VECTOR)
4705 return (Index == 0) ? V.getOperand(0)
4706 : DAG.getUNDEF(VT.getVectorElementType());
4708 if (V.getOpcode() == ISD::BUILD_VECTOR)
4709 return V.getOperand(Index);
4714 /// Custom lower build_vector of v16i8.
4715 static SDValue LowerBuildVectorv16i8(SDValue Op, unsigned NonZeros,
4716 unsigned NumNonZero, unsigned NumZero,
4718 const X86Subtarget* Subtarget,
4719 const TargetLowering &TLI) {
4727 // SSE4.1 - use PINSRB to insert each byte directly.
4728 if (Subtarget->hasSSE41()) {
4729 for (unsigned i = 0; i < 16; ++i) {
4730 bool isNonZero = (NonZeros & (1 << i)) != 0;
4734 V = getZeroVector(MVT::v16i8, Subtarget, DAG, dl);
4736 V = DAG.getUNDEF(MVT::v16i8);
4739 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl,
4740 MVT::v16i8, V, Op.getOperand(i),
4741 DAG.getIntPtrConstant(i, dl));
4748 // Pre-SSE4.1 - merge byte pairs and insert with PINSRW.
4749 for (unsigned i = 0; i < 16; ++i) {
4750 bool ThisIsNonZero = (NonZeros & (1 << i)) != 0;
4751 if (ThisIsNonZero && First) {
4753 V = getZeroVector(MVT::v8i16, Subtarget, DAG, dl);
4755 V = DAG.getUNDEF(MVT::v8i16);
4760 SDValue ThisElt, LastElt;
4761 bool LastIsNonZero = (NonZeros & (1 << (i-1))) != 0;
4762 if (LastIsNonZero) {
4763 LastElt = DAG.getNode(ISD::ZERO_EXTEND, dl,
4764 MVT::i16, Op.getOperand(i-1));
4766 if (ThisIsNonZero) {
4767 ThisElt = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i16, Op.getOperand(i));
4768 ThisElt = DAG.getNode(ISD::SHL, dl, MVT::i16,
4769 ThisElt, DAG.getConstant(8, dl, MVT::i8));
4771 ThisElt = DAG.getNode(ISD::OR, dl, MVT::i16, ThisElt, LastElt);
4775 if (ThisElt.getNode())
4776 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, V, ThisElt,
4777 DAG.getIntPtrConstant(i/2, dl));
4781 return DAG.getBitcast(MVT::v16i8, V);
4784 /// Custom lower build_vector of v8i16.
4785 static SDValue LowerBuildVectorv8i16(SDValue Op, unsigned NonZeros,
4786 unsigned NumNonZero, unsigned NumZero,
4788 const X86Subtarget* Subtarget,
4789 const TargetLowering &TLI) {
4796 for (unsigned i = 0; i < 8; ++i) {
4797 bool isNonZero = (NonZeros & (1 << i)) != 0;
4801 V = getZeroVector(MVT::v8i16, Subtarget, DAG, dl);
4803 V = DAG.getUNDEF(MVT::v8i16);
4806 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl,
4807 MVT::v8i16, V, Op.getOperand(i),
4808 DAG.getIntPtrConstant(i, dl));
4815 /// Custom lower build_vector of v4i32 or v4f32.
4816 static SDValue LowerBuildVectorv4x32(SDValue Op, SelectionDAG &DAG,
4817 const X86Subtarget *Subtarget,
4818 const TargetLowering &TLI) {
4819 // Find all zeroable elements.
4820 std::bitset<4> Zeroable;
4821 for (int i=0; i < 4; ++i) {
4822 SDValue Elt = Op->getOperand(i);
4823 Zeroable[i] = (Elt.getOpcode() == ISD::UNDEF || X86::isZeroNode(Elt));
4825 assert(Zeroable.size() - Zeroable.count() > 1 &&
4826 "We expect at least two non-zero elements!");
4828 // We only know how to deal with build_vector nodes where elements are either
4829 // zeroable or extract_vector_elt with constant index.
4830 SDValue FirstNonZero;
4831 unsigned FirstNonZeroIdx;
4832 for (unsigned i=0; i < 4; ++i) {
4835 SDValue Elt = Op->getOperand(i);
4836 if (Elt.getOpcode() != ISD::EXTRACT_VECTOR_ELT ||
4837 !isa<ConstantSDNode>(Elt.getOperand(1)))
4839 // Make sure that this node is extracting from a 128-bit vector.
4840 MVT VT = Elt.getOperand(0).getSimpleValueType();
4841 if (!VT.is128BitVector())
4843 if (!FirstNonZero.getNode()) {
4845 FirstNonZeroIdx = i;
4849 assert(FirstNonZero.getNode() && "Unexpected build vector of all zeros!");
4850 SDValue V1 = FirstNonZero.getOperand(0);
4851 MVT VT = V1.getSimpleValueType();
4853 // See if this build_vector can be lowered as a blend with zero.
4855 unsigned EltMaskIdx, EltIdx;
4857 for (EltIdx = 0; EltIdx < 4; ++EltIdx) {
4858 if (Zeroable[EltIdx]) {
4859 // The zero vector will be on the right hand side.
4860 Mask[EltIdx] = EltIdx+4;
4864 Elt = Op->getOperand(EltIdx);
4865 // By construction, Elt is a EXTRACT_VECTOR_ELT with constant index.
4866 EltMaskIdx = cast<ConstantSDNode>(Elt.getOperand(1))->getZExtValue();
4867 if (Elt.getOperand(0) != V1 || EltMaskIdx != EltIdx)
4869 Mask[EltIdx] = EltIdx;
4873 // Let the shuffle legalizer deal with blend operations.
4874 SDValue VZero = getZeroVector(VT, Subtarget, DAG, SDLoc(Op));
4875 if (V1.getSimpleValueType() != VT)
4876 V1 = DAG.getNode(ISD::BITCAST, SDLoc(V1), VT, V1);
4877 return DAG.getVectorShuffle(VT, SDLoc(V1), V1, VZero, &Mask[0]);
4880 // See if we can lower this build_vector to a INSERTPS.
4881 if (!Subtarget->hasSSE41())
4884 SDValue V2 = Elt.getOperand(0);
4885 if (Elt == FirstNonZero && EltIdx == FirstNonZeroIdx)
4888 bool CanFold = true;
4889 for (unsigned i = EltIdx + 1; i < 4 && CanFold; ++i) {
4893 SDValue Current = Op->getOperand(i);
4894 SDValue SrcVector = Current->getOperand(0);
4897 CanFold = SrcVector == V1 &&
4898 cast<ConstantSDNode>(Current.getOperand(1))->getZExtValue() == i;
4904 assert(V1.getNode() && "Expected at least two non-zero elements!");
4905 if (V1.getSimpleValueType() != MVT::v4f32)
4906 V1 = DAG.getNode(ISD::BITCAST, SDLoc(V1), MVT::v4f32, V1);
4907 if (V2.getSimpleValueType() != MVT::v4f32)
4908 V2 = DAG.getNode(ISD::BITCAST, SDLoc(V2), MVT::v4f32, V2);
4910 // Ok, we can emit an INSERTPS instruction.
4911 unsigned ZMask = Zeroable.to_ulong();
4913 unsigned InsertPSMask = EltMaskIdx << 6 | EltIdx << 4 | ZMask;
4914 assert((InsertPSMask & ~0xFFu) == 0 && "Invalid mask!");
4916 SDValue Result = DAG.getNode(X86ISD::INSERTPS, DL, MVT::v4f32, V1, V2,
4917 DAG.getIntPtrConstant(InsertPSMask, DL));
4918 return DAG.getBitcast(VT, Result);
4921 /// Return a vector logical shift node.
4922 static SDValue getVShift(bool isLeft, EVT VT, SDValue SrcOp,
4923 unsigned NumBits, SelectionDAG &DAG,
4924 const TargetLowering &TLI, SDLoc dl) {
4925 assert(VT.is128BitVector() && "Unknown type for VShift");
4926 MVT ShVT = MVT::v2i64;
4927 unsigned Opc = isLeft ? X86ISD::VSHLDQ : X86ISD::VSRLDQ;
4928 SrcOp = DAG.getBitcast(ShVT, SrcOp);
4929 MVT ScalarShiftTy = TLI.getScalarShiftAmountTy(DAG.getDataLayout(), VT);
4930 assert(NumBits % 8 == 0 && "Only support byte sized shifts");
4931 SDValue ShiftVal = DAG.getConstant(NumBits/8, dl, ScalarShiftTy);
4932 return DAG.getBitcast(VT, DAG.getNode(Opc, dl, ShVT, SrcOp, ShiftVal));
4936 LowerAsSplatVectorLoad(SDValue SrcOp, MVT VT, SDLoc dl, SelectionDAG &DAG) {
4938 // Check if the scalar load can be widened into a vector load. And if
4939 // the address is "base + cst" see if the cst can be "absorbed" into
4940 // the shuffle mask.
4941 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(SrcOp)) {
4942 SDValue Ptr = LD->getBasePtr();
4943 if (!ISD::isNormalLoad(LD) || LD->isVolatile())
4945 EVT PVT = LD->getValueType(0);
4946 if (PVT != MVT::i32 && PVT != MVT::f32)
4951 if (FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr)) {
4952 FI = FINode->getIndex();
4954 } else if (DAG.isBaseWithConstantOffset(Ptr) &&
4955 isa<FrameIndexSDNode>(Ptr.getOperand(0))) {
4956 FI = cast<FrameIndexSDNode>(Ptr.getOperand(0))->getIndex();
4957 Offset = Ptr.getConstantOperandVal(1);
4958 Ptr = Ptr.getOperand(0);
4963 // FIXME: 256-bit vector instructions don't require a strict alignment,
4964 // improve this code to support it better.
4965 unsigned RequiredAlign = VT.getSizeInBits()/8;
4966 SDValue Chain = LD->getChain();
4967 // Make sure the stack object alignment is at least 16 or 32.
4968 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
4969 if (DAG.InferPtrAlignment(Ptr) < RequiredAlign) {
4970 if (MFI->isFixedObjectIndex(FI)) {
4971 // Can't change the alignment. FIXME: It's possible to compute
4972 // the exact stack offset and reference FI + adjust offset instead.
4973 // If someone *really* cares about this. That's the way to implement it.
4976 MFI->setObjectAlignment(FI, RequiredAlign);
4980 // (Offset % 16 or 32) must be multiple of 4. Then address is then
4981 // Ptr + (Offset & ~15).
4984 if ((Offset % RequiredAlign) & 3)
4986 int64_t StartOffset = Offset & ~int64_t(RequiredAlign - 1);
4989 Ptr = DAG.getNode(ISD::ADD, DL, Ptr.getValueType(), Ptr,
4990 DAG.getConstant(StartOffset, DL, Ptr.getValueType()));
4993 int EltNo = (Offset - StartOffset) >> 2;
4994 unsigned NumElems = VT.getVectorNumElements();
4996 EVT NVT = EVT::getVectorVT(*DAG.getContext(), PVT, NumElems);
4997 SDValue V1 = DAG.getLoad(NVT, dl, Chain, Ptr,
4998 LD->getPointerInfo().getWithOffset(StartOffset),
4999 false, false, false, 0);
5001 SmallVector<int, 8> Mask(NumElems, EltNo);
5003 return DAG.getVectorShuffle(NVT, dl, V1, DAG.getUNDEF(NVT), &Mask[0]);
5009 /// Given the initializing elements 'Elts' of a vector of type 'VT', see if the
5010 /// elements can be replaced by a single large load which has the same value as
5011 /// a build_vector or insert_subvector whose loaded operands are 'Elts'.
5013 /// Example: <load i32 *a, load i32 *a+4, undef, undef> -> zextload a
5015 /// FIXME: we'd also like to handle the case where the last elements are zero
5016 /// rather than undef via VZEXT_LOAD, but we do not detect that case today.
5017 /// There's even a handy isZeroNode for that purpose.
5018 static SDValue EltsFromConsecutiveLoads(EVT VT, ArrayRef<SDValue> Elts,
5019 SDLoc &DL, SelectionDAG &DAG,
5020 bool isAfterLegalize) {
5021 unsigned NumElems = Elts.size();
5023 LoadSDNode *LDBase = nullptr;
5024 unsigned LastLoadedElt = -1U;
5026 // For each element in the initializer, see if we've found a load or an undef.
5027 // If we don't find an initial load element, or later load elements are
5028 // non-consecutive, bail out.
5029 for (unsigned i = 0; i < NumElems; ++i) {
5030 SDValue Elt = Elts[i];
5031 // Look through a bitcast.
5032 if (Elt.getNode() && Elt.getOpcode() == ISD::BITCAST)
5033 Elt = Elt.getOperand(0);
5034 if (!Elt.getNode() ||
5035 (Elt.getOpcode() != ISD::UNDEF && !ISD::isNON_EXTLoad(Elt.getNode())))
5038 if (Elt.getNode()->getOpcode() == ISD::UNDEF)
5040 LDBase = cast<LoadSDNode>(Elt.getNode());
5044 if (Elt.getOpcode() == ISD::UNDEF)
5047 LoadSDNode *LD = cast<LoadSDNode>(Elt);
5048 EVT LdVT = Elt.getValueType();
5049 // Each loaded element must be the correct fractional portion of the
5050 // requested vector load.
5051 if (LdVT.getSizeInBits() != VT.getSizeInBits() / NumElems)
5053 if (!DAG.isConsecutiveLoad(LD, LDBase, LdVT.getSizeInBits() / 8, i))
5058 // If we have found an entire vector of loads and undefs, then return a large
5059 // load of the entire vector width starting at the base pointer. If we found
5060 // consecutive loads for the low half, generate a vzext_load node.
5061 if (LastLoadedElt == NumElems - 1) {
5062 assert(LDBase && "Did not find base load for merging consecutive loads");
5063 EVT EltVT = LDBase->getValueType(0);
5064 // Ensure that the input vector size for the merged loads matches the
5065 // cumulative size of the input elements.
5066 if (VT.getSizeInBits() != EltVT.getSizeInBits() * NumElems)
5069 if (isAfterLegalize &&
5070 !DAG.getTargetLoweringInfo().isOperationLegal(ISD::LOAD, VT))
5073 SDValue NewLd = SDValue();
5075 NewLd = DAG.getLoad(VT, DL, LDBase->getChain(), LDBase->getBasePtr(),
5076 LDBase->getPointerInfo(), LDBase->isVolatile(),
5077 LDBase->isNonTemporal(), LDBase->isInvariant(),
5078 LDBase->getAlignment());
5080 if (LDBase->hasAnyUseOfValue(1)) {
5081 SDValue NewChain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other,
5083 SDValue(NewLd.getNode(), 1));
5084 DAG.ReplaceAllUsesOfValueWith(SDValue(LDBase, 1), NewChain);
5085 DAG.UpdateNodeOperands(NewChain.getNode(), SDValue(LDBase, 1),
5086 SDValue(NewLd.getNode(), 1));
5092 //TODO: The code below fires only for for loading the low v2i32 / v2f32
5093 //of a v4i32 / v4f32. It's probably worth generalizing.
5094 EVT EltVT = VT.getVectorElementType();
5095 if (NumElems == 4 && LastLoadedElt == 1 && (EltVT.getSizeInBits() == 32) &&
5096 DAG.getTargetLoweringInfo().isTypeLegal(MVT::v2i64)) {
5097 SDVTList Tys = DAG.getVTList(MVT::v2i64, MVT::Other);
5098 SDValue Ops[] = { LDBase->getChain(), LDBase->getBasePtr() };
5100 DAG.getMemIntrinsicNode(X86ISD::VZEXT_LOAD, DL, Tys, Ops, MVT::i64,
5101 LDBase->getPointerInfo(),
5102 LDBase->getAlignment(),
5103 false/*isVolatile*/, true/*ReadMem*/,
5106 // Make sure the newly-created LOAD is in the same position as LDBase in
5107 // terms of dependency. We create a TokenFactor for LDBase and ResNode, and
5108 // update uses of LDBase's output chain to use the TokenFactor.
5109 if (LDBase->hasAnyUseOfValue(1)) {
5110 SDValue NewChain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other,
5111 SDValue(LDBase, 1), SDValue(ResNode.getNode(), 1));
5112 DAG.ReplaceAllUsesOfValueWith(SDValue(LDBase, 1), NewChain);
5113 DAG.UpdateNodeOperands(NewChain.getNode(), SDValue(LDBase, 1),
5114 SDValue(ResNode.getNode(), 1));
5117 return DAG.getBitcast(VT, ResNode);
5122 /// LowerVectorBroadcast - Attempt to use the vbroadcast instruction
5123 /// to generate a splat value for the following cases:
5124 /// 1. A splat BUILD_VECTOR which uses a single scalar load, or a constant.
5125 /// 2. A splat shuffle which uses a scalar_to_vector node which comes from
5126 /// a scalar load, or a constant.
5127 /// The VBROADCAST node is returned when a pattern is found,
5128 /// or SDValue() otherwise.
5129 static SDValue LowerVectorBroadcast(SDValue Op, const X86Subtarget* Subtarget,
5130 SelectionDAG &DAG) {
5131 // VBROADCAST requires AVX.
5132 // TODO: Splats could be generated for non-AVX CPUs using SSE
5133 // instructions, but there's less potential gain for only 128-bit vectors.
5134 if (!Subtarget->hasAVX())
5137 MVT VT = Op.getSimpleValueType();
5140 assert((VT.is128BitVector() || VT.is256BitVector() || VT.is512BitVector()) &&
5141 "Unsupported vector type for broadcast.");
5146 switch (Op.getOpcode()) {
5148 // Unknown pattern found.
5151 case ISD::BUILD_VECTOR: {
5152 auto *BVOp = cast<BuildVectorSDNode>(Op.getNode());
5153 BitVector UndefElements;
5154 SDValue Splat = BVOp->getSplatValue(&UndefElements);
5156 // We need a splat of a single value to use broadcast, and it doesn't
5157 // make any sense if the value is only in one element of the vector.
5158 if (!Splat || (VT.getVectorNumElements() - UndefElements.count()) <= 1)
5162 ConstSplatVal = (Ld.getOpcode() == ISD::Constant ||
5163 Ld.getOpcode() == ISD::ConstantFP);
5165 // Make sure that all of the users of a non-constant load are from the
5166 // BUILD_VECTOR node.
5167 if (!ConstSplatVal && !BVOp->isOnlyUserOf(Ld.getNode()))
5172 case ISD::VECTOR_SHUFFLE: {
5173 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
5175 // Shuffles must have a splat mask where the first element is
5177 if ((!SVOp->isSplat()) || SVOp->getMaskElt(0) != 0)
5180 SDValue Sc = Op.getOperand(0);
5181 if (Sc.getOpcode() != ISD::SCALAR_TO_VECTOR &&
5182 Sc.getOpcode() != ISD::BUILD_VECTOR) {
5184 if (!Subtarget->hasInt256())
5187 // Use the register form of the broadcast instruction available on AVX2.
5188 if (VT.getSizeInBits() >= 256)
5189 Sc = Extract128BitVector(Sc, 0, DAG, dl);
5190 return DAG.getNode(X86ISD::VBROADCAST, dl, VT, Sc);
5193 Ld = Sc.getOperand(0);
5194 ConstSplatVal = (Ld.getOpcode() == ISD::Constant ||
5195 Ld.getOpcode() == ISD::ConstantFP);
5197 // The scalar_to_vector node and the suspected
5198 // load node must have exactly one user.
5199 // Constants may have multiple users.
5201 // AVX-512 has register version of the broadcast
5202 bool hasRegVer = Subtarget->hasAVX512() && VT.is512BitVector() &&
5203 Ld.getValueType().getSizeInBits() >= 32;
5204 if (!ConstSplatVal && ((!Sc.hasOneUse() || !Ld.hasOneUse()) &&
5211 unsigned ScalarSize = Ld.getValueType().getSizeInBits();
5212 bool IsGE256 = (VT.getSizeInBits() >= 256);
5214 // When optimizing for size, generate up to 5 extra bytes for a broadcast
5215 // instruction to save 8 or more bytes of constant pool data.
5216 // TODO: If multiple splats are generated to load the same constant,
5217 // it may be detrimental to overall size. There needs to be a way to detect
5218 // that condition to know if this is truly a size win.
5219 bool OptForSize = DAG.getMachineFunction().getFunction()->optForSize();
5221 // Handle broadcasting a single constant scalar from the constant pool
5223 // On Sandybridge (no AVX2), it is still better to load a constant vector
5224 // from the constant pool and not to broadcast it from a scalar.
5225 // But override that restriction when optimizing for size.
5226 // TODO: Check if splatting is recommended for other AVX-capable CPUs.
5227 if (ConstSplatVal && (Subtarget->hasAVX2() || OptForSize)) {
5228 EVT CVT = Ld.getValueType();
5229 assert(!CVT.isVector() && "Must not broadcast a vector type");
5231 // Splat f32, i32, v4f64, v4i64 in all cases with AVX2.
5232 // For size optimization, also splat v2f64 and v2i64, and for size opt
5233 // with AVX2, also splat i8 and i16.
5234 // With pattern matching, the VBROADCAST node may become a VMOVDDUP.
5235 if (ScalarSize == 32 || (IsGE256 && ScalarSize == 64) ||
5236 (OptForSize && (ScalarSize == 64 || Subtarget->hasAVX2()))) {
5237 const Constant *C = nullptr;
5238 if (ConstantSDNode *CI = dyn_cast<ConstantSDNode>(Ld))
5239 C = CI->getConstantIntValue();
5240 else if (ConstantFPSDNode *CF = dyn_cast<ConstantFPSDNode>(Ld))
5241 C = CF->getConstantFPValue();
5243 assert(C && "Invalid constant type");
5245 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
5247 DAG.getConstantPool(C, TLI.getPointerTy(DAG.getDataLayout()));
5248 unsigned Alignment = cast<ConstantPoolSDNode>(CP)->getAlignment();
5250 CVT, dl, DAG.getEntryNode(), CP,
5251 MachinePointerInfo::getConstantPool(DAG.getMachineFunction()), false,
5252 false, false, Alignment);
5254 return DAG.getNode(X86ISD::VBROADCAST, dl, VT, Ld);
5258 bool IsLoad = ISD::isNormalLoad(Ld.getNode());
5260 // Handle AVX2 in-register broadcasts.
5261 if (!IsLoad && Subtarget->hasInt256() &&
5262 (ScalarSize == 32 || (IsGE256 && ScalarSize == 64)))
5263 return DAG.getNode(X86ISD::VBROADCAST, dl, VT, Ld);
5265 // The scalar source must be a normal load.
5269 if (ScalarSize == 32 || (IsGE256 && ScalarSize == 64) ||
5270 (Subtarget->hasVLX() && ScalarSize == 64))
5271 return DAG.getNode(X86ISD::VBROADCAST, dl, VT, Ld);
5273 // The integer check is needed for the 64-bit into 128-bit so it doesn't match
5274 // double since there is no vbroadcastsd xmm
5275 if (Subtarget->hasInt256() && Ld.getValueType().isInteger()) {
5276 if (ScalarSize == 8 || ScalarSize == 16 || ScalarSize == 64)
5277 return DAG.getNode(X86ISD::VBROADCAST, dl, VT, Ld);
5280 // Unsupported broadcast.
5284 /// \brief For an EXTRACT_VECTOR_ELT with a constant index return the real
5285 /// underlying vector and index.
5287 /// Modifies \p ExtractedFromVec to the real vector and returns the real
5289 static int getUnderlyingExtractedFromVec(SDValue &ExtractedFromVec,
5291 int Idx = cast<ConstantSDNode>(ExtIdx)->getZExtValue();
5292 if (!isa<ShuffleVectorSDNode>(ExtractedFromVec))
5295 // For 256-bit vectors, LowerEXTRACT_VECTOR_ELT_SSE4 may have already
5297 // (extract_vector_elt (v8f32 %vreg1), Constant<6>)
5299 // (extract_vector_elt (vector_shuffle<2,u,u,u>
5300 // (extract_subvector (v8f32 %vreg0), Constant<4>),
5303 // In this case the vector is the extract_subvector expression and the index
5304 // is 2, as specified by the shuffle.
5305 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(ExtractedFromVec);
5306 SDValue ShuffleVec = SVOp->getOperand(0);
5307 MVT ShuffleVecVT = ShuffleVec.getSimpleValueType();
5308 assert(ShuffleVecVT.getVectorElementType() ==
5309 ExtractedFromVec.getSimpleValueType().getVectorElementType());
5311 int ShuffleIdx = SVOp->getMaskElt(Idx);
5312 if (isUndefOrInRange(ShuffleIdx, 0, ShuffleVecVT.getVectorNumElements())) {
5313 ExtractedFromVec = ShuffleVec;
5319 static SDValue buildFromShuffleMostly(SDValue Op, SelectionDAG &DAG) {
5320 MVT VT = Op.getSimpleValueType();
5322 // Skip if insert_vec_elt is not supported.
5323 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
5324 if (!TLI.isOperationLegalOrCustom(ISD::INSERT_VECTOR_ELT, VT))
5328 unsigned NumElems = Op.getNumOperands();
5332 SmallVector<unsigned, 4> InsertIndices;
5333 SmallVector<int, 8> Mask(NumElems, -1);
5335 for (unsigned i = 0; i != NumElems; ++i) {
5336 unsigned Opc = Op.getOperand(i).getOpcode();
5338 if (Opc == ISD::UNDEF)
5341 if (Opc != ISD::EXTRACT_VECTOR_ELT) {
5342 // Quit if more than 1 elements need inserting.
5343 if (InsertIndices.size() > 1)
5346 InsertIndices.push_back(i);
5350 SDValue ExtractedFromVec = Op.getOperand(i).getOperand(0);
5351 SDValue ExtIdx = Op.getOperand(i).getOperand(1);
5352 // Quit if non-constant index.
5353 if (!isa<ConstantSDNode>(ExtIdx))
5355 int Idx = getUnderlyingExtractedFromVec(ExtractedFromVec, ExtIdx);
5357 // Quit if extracted from vector of different type.
5358 if (ExtractedFromVec.getValueType() != VT)
5361 if (!VecIn1.getNode())
5362 VecIn1 = ExtractedFromVec;
5363 else if (VecIn1 != ExtractedFromVec) {
5364 if (!VecIn2.getNode())
5365 VecIn2 = ExtractedFromVec;
5366 else if (VecIn2 != ExtractedFromVec)
5367 // Quit if more than 2 vectors to shuffle
5371 if (ExtractedFromVec == VecIn1)
5373 else if (ExtractedFromVec == VecIn2)
5374 Mask[i] = Idx + NumElems;
5377 if (!VecIn1.getNode())
5380 VecIn2 = VecIn2.getNode() ? VecIn2 : DAG.getUNDEF(VT);
5381 SDValue NV = DAG.getVectorShuffle(VT, DL, VecIn1, VecIn2, &Mask[0]);
5382 for (unsigned i = 0, e = InsertIndices.size(); i != e; ++i) {
5383 unsigned Idx = InsertIndices[i];
5384 NV = DAG.getNode(ISD::INSERT_VECTOR_ELT, DL, VT, NV, Op.getOperand(Idx),
5385 DAG.getIntPtrConstant(Idx, DL));
5391 static SDValue ConvertI1VectorToInteger(SDValue Op, SelectionDAG &DAG) {
5392 assert(ISD::isBuildVectorOfConstantSDNodes(Op.getNode()) &&
5393 Op.getScalarValueSizeInBits() == 1 &&
5394 "Can not convert non-constant vector");
5395 uint64_t Immediate = 0;
5396 for (unsigned idx = 0, e = Op.getNumOperands(); idx < e; ++idx) {
5397 SDValue In = Op.getOperand(idx);
5398 if (In.getOpcode() != ISD::UNDEF)
5399 Immediate |= cast<ConstantSDNode>(In)->getZExtValue() << idx;
5403 MVT::getIntegerVT(std::max((int)Op.getValueType().getSizeInBits(), 8));
5404 return DAG.getConstant(Immediate, dl, VT);
5406 // Lower BUILD_VECTOR operation for v8i1 and v16i1 types.
5408 X86TargetLowering::LowerBUILD_VECTORvXi1(SDValue Op, SelectionDAG &DAG) const {
5410 MVT VT = Op.getSimpleValueType();
5411 assert((VT.getVectorElementType() == MVT::i1) &&
5412 "Unexpected type in LowerBUILD_VECTORvXi1!");
5415 if (ISD::isBuildVectorAllZeros(Op.getNode())) {
5416 SDValue Cst = DAG.getTargetConstant(0, dl, MVT::i1);
5417 SmallVector<SDValue, 16> Ops(VT.getVectorNumElements(), Cst);
5418 return DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Ops);
5421 if (ISD::isBuildVectorAllOnes(Op.getNode())) {
5422 SDValue Cst = DAG.getTargetConstant(1, dl, MVT::i1);
5423 SmallVector<SDValue, 16> Ops(VT.getVectorNumElements(), Cst);
5424 return DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Ops);
5427 if (ISD::isBuildVectorOfConstantSDNodes(Op.getNode())) {
5428 SDValue Imm = ConvertI1VectorToInteger(Op, DAG);
5429 if (Imm.getValueSizeInBits() == VT.getSizeInBits())
5430 return DAG.getBitcast(VT, Imm);
5431 SDValue ExtVec = DAG.getBitcast(MVT::v8i1, Imm);
5432 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT, ExtVec,
5433 DAG.getIntPtrConstant(0, dl));
5436 // Vector has one or more non-const elements
5437 uint64_t Immediate = 0;
5438 SmallVector<unsigned, 16> NonConstIdx;
5439 bool IsSplat = true;
5440 bool HasConstElts = false;
5442 for (unsigned idx = 0, e = Op.getNumOperands(); idx < e; ++idx) {
5443 SDValue In = Op.getOperand(idx);
5444 if (In.getOpcode() == ISD::UNDEF)
5446 if (!isa<ConstantSDNode>(In))
5447 NonConstIdx.push_back(idx);
5449 Immediate |= cast<ConstantSDNode>(In)->getZExtValue() << idx;
5450 HasConstElts = true;
5454 else if (In != Op.getOperand(SplatIdx))
5458 // for splat use " (select i1 splat_elt, all-ones, all-zeroes)"
5460 return DAG.getNode(ISD::SELECT, dl, VT, Op.getOperand(SplatIdx),
5461 DAG.getConstant(1, dl, VT),
5462 DAG.getConstant(0, dl, VT));
5464 // insert elements one by one
5468 MVT ImmVT = MVT::getIntegerVT(std::max((int)VT.getSizeInBits(), 8));
5469 Imm = DAG.getConstant(Immediate, dl, ImmVT);
5471 else if (HasConstElts)
5472 Imm = DAG.getConstant(0, dl, VT);
5474 Imm = DAG.getUNDEF(VT);
5475 if (Imm.getValueSizeInBits() == VT.getSizeInBits())
5476 DstVec = DAG.getBitcast(VT, Imm);
5478 SDValue ExtVec = DAG.getBitcast(MVT::v8i1, Imm);
5479 DstVec = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT, ExtVec,
5480 DAG.getIntPtrConstant(0, dl));
5483 for (unsigned i = 0; i < NonConstIdx.size(); ++i) {
5484 unsigned InsertIdx = NonConstIdx[i];
5485 DstVec = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, DstVec,
5486 Op.getOperand(InsertIdx),
5487 DAG.getIntPtrConstant(InsertIdx, dl));
5492 /// \brief Return true if \p N implements a horizontal binop and return the
5493 /// operands for the horizontal binop into V0 and V1.
5495 /// This is a helper function of LowerToHorizontalOp().
5496 /// This function checks that the build_vector \p N in input implements a
5497 /// horizontal operation. Parameter \p Opcode defines the kind of horizontal
5498 /// operation to match.
5499 /// For example, if \p Opcode is equal to ISD::ADD, then this function
5500 /// checks if \p N implements a horizontal arithmetic add; if instead \p Opcode
5501 /// is equal to ISD::SUB, then this function checks if this is a horizontal
5504 /// This function only analyzes elements of \p N whose indices are
5505 /// in range [BaseIdx, LastIdx).
5506 static bool isHorizontalBinOp(const BuildVectorSDNode *N, unsigned Opcode,
5508 unsigned BaseIdx, unsigned LastIdx,
5509 SDValue &V0, SDValue &V1) {
5510 EVT VT = N->getValueType(0);
5512 assert(BaseIdx * 2 <= LastIdx && "Invalid Indices in input!");
5513 assert(VT.isVector() && VT.getVectorNumElements() >= LastIdx &&
5514 "Invalid Vector in input!");
5516 bool IsCommutable = (Opcode == ISD::ADD || Opcode == ISD::FADD);
5517 bool CanFold = true;
5518 unsigned ExpectedVExtractIdx = BaseIdx;
5519 unsigned NumElts = LastIdx - BaseIdx;
5520 V0 = DAG.getUNDEF(VT);
5521 V1 = DAG.getUNDEF(VT);
5523 // Check if N implements a horizontal binop.
5524 for (unsigned i = 0, e = NumElts; i != e && CanFold; ++i) {
5525 SDValue Op = N->getOperand(i + BaseIdx);
5528 if (Op->getOpcode() == ISD::UNDEF) {
5529 // Update the expected vector extract index.
5530 if (i * 2 == NumElts)
5531 ExpectedVExtractIdx = BaseIdx;
5532 ExpectedVExtractIdx += 2;
5536 CanFold = Op->getOpcode() == Opcode && Op->hasOneUse();
5541 SDValue Op0 = Op.getOperand(0);
5542 SDValue Op1 = Op.getOperand(1);
5544 // Try to match the following pattern:
5545 // (BINOP (extract_vector_elt A, I), (extract_vector_elt A, I+1))
5546 CanFold = (Op0.getOpcode() == ISD::EXTRACT_VECTOR_ELT &&
5547 Op1.getOpcode() == ISD::EXTRACT_VECTOR_ELT &&
5548 Op0.getOperand(0) == Op1.getOperand(0) &&
5549 isa<ConstantSDNode>(Op0.getOperand(1)) &&
5550 isa<ConstantSDNode>(Op1.getOperand(1)));
5554 unsigned I0 = cast<ConstantSDNode>(Op0.getOperand(1))->getZExtValue();
5555 unsigned I1 = cast<ConstantSDNode>(Op1.getOperand(1))->getZExtValue();
5557 if (i * 2 < NumElts) {
5558 if (V0.getOpcode() == ISD::UNDEF) {
5559 V0 = Op0.getOperand(0);
5560 if (V0.getValueType() != VT)
5564 if (V1.getOpcode() == ISD::UNDEF) {
5565 V1 = Op0.getOperand(0);
5566 if (V1.getValueType() != VT)
5569 if (i * 2 == NumElts)
5570 ExpectedVExtractIdx = BaseIdx;
5573 SDValue Expected = (i * 2 < NumElts) ? V0 : V1;
5574 if (I0 == ExpectedVExtractIdx)
5575 CanFold = I1 == I0 + 1 && Op0.getOperand(0) == Expected;
5576 else if (IsCommutable && I1 == ExpectedVExtractIdx) {
5577 // Try to match the following dag sequence:
5578 // (BINOP (extract_vector_elt A, I+1), (extract_vector_elt A, I))
5579 CanFold = I0 == I1 + 1 && Op1.getOperand(0) == Expected;
5583 ExpectedVExtractIdx += 2;
5589 /// \brief Emit a sequence of two 128-bit horizontal add/sub followed by
5590 /// a concat_vector.
5592 /// This is a helper function of LowerToHorizontalOp().
5593 /// This function expects two 256-bit vectors called V0 and V1.
5594 /// At first, each vector is split into two separate 128-bit vectors.
5595 /// Then, the resulting 128-bit vectors are used to implement two
5596 /// horizontal binary operations.
5598 /// The kind of horizontal binary operation is defined by \p X86Opcode.
5600 /// \p Mode specifies how the 128-bit parts of V0 and V1 are passed in input to
5601 /// the two new horizontal binop.
5602 /// When Mode is set, the first horizontal binop dag node would take as input
5603 /// the lower 128-bit of V0 and the upper 128-bit of V0. The second
5604 /// horizontal binop dag node would take as input the lower 128-bit of V1
5605 /// and the upper 128-bit of V1.
5607 /// HADD V0_LO, V0_HI
5608 /// HADD V1_LO, V1_HI
5610 /// Otherwise, the first horizontal binop dag node takes as input the lower
5611 /// 128-bit of V0 and the lower 128-bit of V1, and the second horizontal binop
5612 /// dag node takes the upper 128-bit of V0 and the upper 128-bit of V1.
5614 /// HADD V0_LO, V1_LO
5615 /// HADD V0_HI, V1_HI
5617 /// If \p isUndefLO is set, then the algorithm propagates UNDEF to the lower
5618 /// 128-bits of the result. If \p isUndefHI is set, then UNDEF is propagated to
5619 /// the upper 128-bits of the result.
5620 static SDValue ExpandHorizontalBinOp(const SDValue &V0, const SDValue &V1,
5621 SDLoc DL, SelectionDAG &DAG,
5622 unsigned X86Opcode, bool Mode,
5623 bool isUndefLO, bool isUndefHI) {
5624 EVT VT = V0.getValueType();
5625 assert(VT.is256BitVector() && VT == V1.getValueType() &&
5626 "Invalid nodes in input!");
5628 unsigned NumElts = VT.getVectorNumElements();
5629 SDValue V0_LO = Extract128BitVector(V0, 0, DAG, DL);
5630 SDValue V0_HI = Extract128BitVector(V0, NumElts/2, DAG, DL);
5631 SDValue V1_LO = Extract128BitVector(V1, 0, DAG, DL);
5632 SDValue V1_HI = Extract128BitVector(V1, NumElts/2, DAG, DL);
5633 EVT NewVT = V0_LO.getValueType();
5635 SDValue LO = DAG.getUNDEF(NewVT);
5636 SDValue HI = DAG.getUNDEF(NewVT);
5639 // Don't emit a horizontal binop if the result is expected to be UNDEF.
5640 if (!isUndefLO && V0->getOpcode() != ISD::UNDEF)
5641 LO = DAG.getNode(X86Opcode, DL, NewVT, V0_LO, V0_HI);
5642 if (!isUndefHI && V1->getOpcode() != ISD::UNDEF)
5643 HI = DAG.getNode(X86Opcode, DL, NewVT, V1_LO, V1_HI);
5645 // Don't emit a horizontal binop if the result is expected to be UNDEF.
5646 if (!isUndefLO && (V0_LO->getOpcode() != ISD::UNDEF ||
5647 V1_LO->getOpcode() != ISD::UNDEF))
5648 LO = DAG.getNode(X86Opcode, DL, NewVT, V0_LO, V1_LO);
5650 if (!isUndefHI && (V0_HI->getOpcode() != ISD::UNDEF ||
5651 V1_HI->getOpcode() != ISD::UNDEF))
5652 HI = DAG.getNode(X86Opcode, DL, NewVT, V0_HI, V1_HI);
5655 return DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, LO, HI);
5658 /// Try to fold a build_vector that performs an 'addsub' to an X86ISD::ADDSUB
5660 static SDValue LowerToAddSub(const BuildVectorSDNode *BV,
5661 const X86Subtarget *Subtarget, SelectionDAG &DAG) {
5662 EVT VT = BV->getValueType(0);
5663 if ((!Subtarget->hasSSE3() || (VT != MVT::v4f32 && VT != MVT::v2f64)) &&
5664 (!Subtarget->hasAVX() || (VT != MVT::v8f32 && VT != MVT::v4f64)))
5668 unsigned NumElts = VT.getVectorNumElements();
5669 SDValue InVec0 = DAG.getUNDEF(VT);
5670 SDValue InVec1 = DAG.getUNDEF(VT);
5672 assert((VT == MVT::v8f32 || VT == MVT::v4f64 || VT == MVT::v4f32 ||
5673 VT == MVT::v2f64) && "build_vector with an invalid type found!");
5675 // Odd-numbered elements in the input build vector are obtained from
5676 // adding two integer/float elements.
5677 // Even-numbered elements in the input build vector are obtained from
5678 // subtracting two integer/float elements.
5679 unsigned ExpectedOpcode = ISD::FSUB;
5680 unsigned NextExpectedOpcode = ISD::FADD;
5681 bool AddFound = false;
5682 bool SubFound = false;
5684 for (unsigned i = 0, e = NumElts; i != e; ++i) {
5685 SDValue Op = BV->getOperand(i);
5687 // Skip 'undef' values.
5688 unsigned Opcode = Op.getOpcode();
5689 if (Opcode == ISD::UNDEF) {
5690 std::swap(ExpectedOpcode, NextExpectedOpcode);
5694 // Early exit if we found an unexpected opcode.
5695 if (Opcode != ExpectedOpcode)
5698 SDValue Op0 = Op.getOperand(0);
5699 SDValue Op1 = Op.getOperand(1);
5701 // Try to match the following pattern:
5702 // (BINOP (extract_vector_elt A, i), (extract_vector_elt B, i))
5703 // Early exit if we cannot match that sequence.
5704 if (Op0.getOpcode() != ISD::EXTRACT_VECTOR_ELT ||
5705 Op1.getOpcode() != ISD::EXTRACT_VECTOR_ELT ||
5706 !isa<ConstantSDNode>(Op0.getOperand(1)) ||
5707 !isa<ConstantSDNode>(Op1.getOperand(1)) ||
5708 Op0.getOperand(1) != Op1.getOperand(1))
5711 unsigned I0 = cast<ConstantSDNode>(Op0.getOperand(1))->getZExtValue();
5715 // We found a valid add/sub node. Update the information accordingly.
5721 // Update InVec0 and InVec1.
5722 if (InVec0.getOpcode() == ISD::UNDEF) {
5723 InVec0 = Op0.getOperand(0);
5724 if (InVec0.getValueType() != VT)
5727 if (InVec1.getOpcode() == ISD::UNDEF) {
5728 InVec1 = Op1.getOperand(0);
5729 if (InVec1.getValueType() != VT)
5733 // Make sure that operands in input to each add/sub node always
5734 // come from a same pair of vectors.
5735 if (InVec0 != Op0.getOperand(0)) {
5736 if (ExpectedOpcode == ISD::FSUB)
5739 // FADD is commutable. Try to commute the operands
5740 // and then test again.
5741 std::swap(Op0, Op1);
5742 if (InVec0 != Op0.getOperand(0))
5746 if (InVec1 != Op1.getOperand(0))
5749 // Update the pair of expected opcodes.
5750 std::swap(ExpectedOpcode, NextExpectedOpcode);
5753 // Don't try to fold this build_vector into an ADDSUB if the inputs are undef.
5754 if (AddFound && SubFound && InVec0.getOpcode() != ISD::UNDEF &&
5755 InVec1.getOpcode() != ISD::UNDEF)
5756 return DAG.getNode(X86ISD::ADDSUB, DL, VT, InVec0, InVec1);
5761 /// Lower BUILD_VECTOR to a horizontal add/sub operation if possible.
5762 static SDValue LowerToHorizontalOp(const BuildVectorSDNode *BV,
5763 const X86Subtarget *Subtarget,
5764 SelectionDAG &DAG) {
5765 EVT VT = BV->getValueType(0);
5766 unsigned NumElts = VT.getVectorNumElements();
5767 unsigned NumUndefsLO = 0;
5768 unsigned NumUndefsHI = 0;
5769 unsigned Half = NumElts/2;
5771 // Count the number of UNDEF operands in the build_vector in input.
5772 for (unsigned i = 0, e = Half; i != e; ++i)
5773 if (BV->getOperand(i)->getOpcode() == ISD::UNDEF)
5776 for (unsigned i = Half, e = NumElts; i != e; ++i)
5777 if (BV->getOperand(i)->getOpcode() == ISD::UNDEF)
5780 // Early exit if this is either a build_vector of all UNDEFs or all the
5781 // operands but one are UNDEF.
5782 if (NumUndefsLO + NumUndefsHI + 1 >= NumElts)
5786 SDValue InVec0, InVec1;
5787 if ((VT == MVT::v4f32 || VT == MVT::v2f64) && Subtarget->hasSSE3()) {
5788 // Try to match an SSE3 float HADD/HSUB.
5789 if (isHorizontalBinOp(BV, ISD::FADD, DAG, 0, NumElts, InVec0, InVec1))
5790 return DAG.getNode(X86ISD::FHADD, DL, VT, InVec0, InVec1);
5792 if (isHorizontalBinOp(BV, ISD::FSUB, DAG, 0, NumElts, InVec0, InVec1))
5793 return DAG.getNode(X86ISD::FHSUB, DL, VT, InVec0, InVec1);
5794 } else if ((VT == MVT::v4i32 || VT == MVT::v8i16) && Subtarget->hasSSSE3()) {
5795 // Try to match an SSSE3 integer HADD/HSUB.
5796 if (isHorizontalBinOp(BV, ISD::ADD, DAG, 0, NumElts, InVec0, InVec1))
5797 return DAG.getNode(X86ISD::HADD, DL, VT, InVec0, InVec1);
5799 if (isHorizontalBinOp(BV, ISD::SUB, DAG, 0, NumElts, InVec0, InVec1))
5800 return DAG.getNode(X86ISD::HSUB, DL, VT, InVec0, InVec1);
5803 if (!Subtarget->hasAVX())
5806 if ((VT == MVT::v8f32 || VT == MVT::v4f64)) {
5807 // Try to match an AVX horizontal add/sub of packed single/double
5808 // precision floating point values from 256-bit vectors.
5809 SDValue InVec2, InVec3;
5810 if (isHorizontalBinOp(BV, ISD::FADD, DAG, 0, Half, InVec0, InVec1) &&
5811 isHorizontalBinOp(BV, ISD::FADD, DAG, Half, NumElts, InVec2, InVec3) &&
5812 ((InVec0.getOpcode() == ISD::UNDEF ||
5813 InVec2.getOpcode() == ISD::UNDEF) || InVec0 == InVec2) &&
5814 ((InVec1.getOpcode() == ISD::UNDEF ||
5815 InVec3.getOpcode() == ISD::UNDEF) || InVec1 == InVec3))
5816 return DAG.getNode(X86ISD::FHADD, DL, VT, InVec0, InVec1);
5818 if (isHorizontalBinOp(BV, ISD::FSUB, DAG, 0, Half, InVec0, InVec1) &&
5819 isHorizontalBinOp(BV, ISD::FSUB, DAG, Half, NumElts, InVec2, InVec3) &&
5820 ((InVec0.getOpcode() == ISD::UNDEF ||
5821 InVec2.getOpcode() == ISD::UNDEF) || InVec0 == InVec2) &&
5822 ((InVec1.getOpcode() == ISD::UNDEF ||
5823 InVec3.getOpcode() == ISD::UNDEF) || InVec1 == InVec3))
5824 return DAG.getNode(X86ISD::FHSUB, DL, VT, InVec0, InVec1);
5825 } else if (VT == MVT::v8i32 || VT == MVT::v16i16) {
5826 // Try to match an AVX2 horizontal add/sub of signed integers.
5827 SDValue InVec2, InVec3;
5829 bool CanFold = true;
5831 if (isHorizontalBinOp(BV, ISD::ADD, DAG, 0, Half, InVec0, InVec1) &&
5832 isHorizontalBinOp(BV, ISD::ADD, DAG, Half, NumElts, InVec2, InVec3) &&
5833 ((InVec0.getOpcode() == ISD::UNDEF ||
5834 InVec2.getOpcode() == ISD::UNDEF) || InVec0 == InVec2) &&
5835 ((InVec1.getOpcode() == ISD::UNDEF ||
5836 InVec3.getOpcode() == ISD::UNDEF) || InVec1 == InVec3))
5837 X86Opcode = X86ISD::HADD;
5838 else if (isHorizontalBinOp(BV, ISD::SUB, DAG, 0, Half, InVec0, InVec1) &&
5839 isHorizontalBinOp(BV, ISD::SUB, DAG, Half, NumElts, InVec2, InVec3) &&
5840 ((InVec0.getOpcode() == ISD::UNDEF ||
5841 InVec2.getOpcode() == ISD::UNDEF) || InVec0 == InVec2) &&
5842 ((InVec1.getOpcode() == ISD::UNDEF ||
5843 InVec3.getOpcode() == ISD::UNDEF) || InVec1 == InVec3))
5844 X86Opcode = X86ISD::HSUB;
5849 // Fold this build_vector into a single horizontal add/sub.
5850 // Do this only if the target has AVX2.
5851 if (Subtarget->hasAVX2())
5852 return DAG.getNode(X86Opcode, DL, VT, InVec0, InVec1);
5854 // Do not try to expand this build_vector into a pair of horizontal
5855 // add/sub if we can emit a pair of scalar add/sub.
5856 if (NumUndefsLO + 1 == Half || NumUndefsHI + 1 == Half)
5859 // Convert this build_vector into a pair of horizontal binop followed by
5861 bool isUndefLO = NumUndefsLO == Half;
5862 bool isUndefHI = NumUndefsHI == Half;
5863 return ExpandHorizontalBinOp(InVec0, InVec1, DL, DAG, X86Opcode, false,
5864 isUndefLO, isUndefHI);
5868 if ((VT == MVT::v8f32 || VT == MVT::v4f64 || VT == MVT::v8i32 ||
5869 VT == MVT::v16i16) && Subtarget->hasAVX()) {
5871 if (isHorizontalBinOp(BV, ISD::ADD, DAG, 0, NumElts, InVec0, InVec1))
5872 X86Opcode = X86ISD::HADD;
5873 else if (isHorizontalBinOp(BV, ISD::SUB, DAG, 0, NumElts, InVec0, InVec1))
5874 X86Opcode = X86ISD::HSUB;
5875 else if (isHorizontalBinOp(BV, ISD::FADD, DAG, 0, NumElts, InVec0, InVec1))
5876 X86Opcode = X86ISD::FHADD;
5877 else if (isHorizontalBinOp(BV, ISD::FSUB, DAG, 0, NumElts, InVec0, InVec1))
5878 X86Opcode = X86ISD::FHSUB;
5882 // Don't try to expand this build_vector into a pair of horizontal add/sub
5883 // if we can simply emit a pair of scalar add/sub.
5884 if (NumUndefsLO + 1 == Half || NumUndefsHI + 1 == Half)
5887 // Convert this build_vector into two horizontal add/sub followed by
5889 bool isUndefLO = NumUndefsLO == Half;
5890 bool isUndefHI = NumUndefsHI == Half;
5891 return ExpandHorizontalBinOp(InVec0, InVec1, DL, DAG, X86Opcode, true,
5892 isUndefLO, isUndefHI);
5899 X86TargetLowering::LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG) const {
5902 MVT VT = Op.getSimpleValueType();
5903 MVT ExtVT = VT.getVectorElementType();
5904 unsigned NumElems = Op.getNumOperands();
5906 // Generate vectors for predicate vectors.
5907 if (VT.getScalarType() == MVT::i1 && Subtarget->hasAVX512())
5908 return LowerBUILD_VECTORvXi1(Op, DAG);
5910 // Vectors containing all zeros can be matched by pxor and xorps later
5911 if (ISD::isBuildVectorAllZeros(Op.getNode())) {
5912 // Canonicalize this to <4 x i32> to 1) ensure the zero vectors are CSE'd
5913 // and 2) ensure that i64 scalars are eliminated on x86-32 hosts.
5914 if (VT == MVT::v4i32 || VT == MVT::v8i32 || VT == MVT::v16i32)
5917 return getZeroVector(VT, Subtarget, DAG, dl);
5920 // Vectors containing all ones can be matched by pcmpeqd on 128-bit width
5921 // vectors or broken into v4i32 operations on 256-bit vectors. AVX2 can use
5922 // vpcmpeqd on 256-bit vectors.
5923 if (Subtarget->hasSSE2() && ISD::isBuildVectorAllOnes(Op.getNode())) {
5924 if (VT == MVT::v4i32 || (VT == MVT::v8i32 && Subtarget->hasInt256()))
5927 if (!VT.is512BitVector())
5928 return getOnesVector(VT, Subtarget->hasInt256(), DAG, dl);
5931 BuildVectorSDNode *BV = cast<BuildVectorSDNode>(Op.getNode());
5932 if (SDValue AddSub = LowerToAddSub(BV, Subtarget, DAG))
5934 if (SDValue HorizontalOp = LowerToHorizontalOp(BV, Subtarget, DAG))
5935 return HorizontalOp;
5936 if (SDValue Broadcast = LowerVectorBroadcast(Op, Subtarget, DAG))
5939 unsigned EVTBits = ExtVT.getSizeInBits();
5941 unsigned NumZero = 0;
5942 unsigned NumNonZero = 0;
5943 unsigned NonZeros = 0;
5944 bool IsAllConstants = true;
5945 SmallSet<SDValue, 8> Values;
5946 for (unsigned i = 0; i < NumElems; ++i) {
5947 SDValue Elt = Op.getOperand(i);
5948 if (Elt.getOpcode() == ISD::UNDEF)
5951 if (Elt.getOpcode() != ISD::Constant &&
5952 Elt.getOpcode() != ISD::ConstantFP)
5953 IsAllConstants = false;
5954 if (X86::isZeroNode(Elt))
5957 NonZeros |= (1 << i);
5962 // All undef vector. Return an UNDEF. All zero vectors were handled above.
5963 if (NumNonZero == 0)
5964 return DAG.getUNDEF(VT);
5966 // Special case for single non-zero, non-undef, element.
5967 if (NumNonZero == 1) {
5968 unsigned Idx = countTrailingZeros(NonZeros);
5969 SDValue Item = Op.getOperand(Idx);
5971 // If this is an insertion of an i64 value on x86-32, and if the top bits of
5972 // the value are obviously zero, truncate the value to i32 and do the
5973 // insertion that way. Only do this if the value is non-constant or if the
5974 // value is a constant being inserted into element 0. It is cheaper to do
5975 // a constant pool load than it is to do a movd + shuffle.
5976 if (ExtVT == MVT::i64 && !Subtarget->is64Bit() &&
5977 (!IsAllConstants || Idx == 0)) {
5978 if (DAG.MaskedValueIsZero(Item, APInt::getBitsSet(64, 32, 64))) {
5980 assert(VT == MVT::v2i64 && "Expected an SSE value type!");
5981 EVT VecVT = MVT::v4i32;
5983 // Truncate the value (which may itself be a constant) to i32, and
5984 // convert it to a vector with movd (S2V+shuffle to zero extend).
5985 Item = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, Item);
5986 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VecVT, Item);
5987 return DAG.getBitcast(VT, getShuffleVectorZeroOrUndef(
5988 Item, Idx * 2, true, Subtarget, DAG));
5992 // If we have a constant or non-constant insertion into the low element of
5993 // a vector, we can do this with SCALAR_TO_VECTOR + shuffle of zero into
5994 // the rest of the elements. This will be matched as movd/movq/movss/movsd
5995 // depending on what the source datatype is.
5998 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
6000 if (ExtVT == MVT::i32 || ExtVT == MVT::f32 || ExtVT == MVT::f64 ||
6001 (ExtVT == MVT::i64 && Subtarget->is64Bit())) {
6002 if (VT.is512BitVector()) {
6003 SDValue ZeroVec = getZeroVector(VT, Subtarget, DAG, dl);
6004 return DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, ZeroVec,
6005 Item, DAG.getIntPtrConstant(0, dl));
6007 assert((VT.is128BitVector() || VT.is256BitVector()) &&
6008 "Expected an SSE value type!");
6009 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
6010 // Turn it into a MOVL (i.e. movss, movsd, or movd) to a zero vector.
6011 return getShuffleVectorZeroOrUndef(Item, 0, true, Subtarget, DAG);
6014 // We can't directly insert an i8 or i16 into a vector, so zero extend
6016 if (ExtVT == MVT::i16 || ExtVT == MVT::i8) {
6017 Item = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, Item);
6018 if (VT.is256BitVector()) {
6019 if (Subtarget->hasAVX()) {
6020 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v8i32, Item);
6021 Item = getShuffleVectorZeroOrUndef(Item, 0, true, Subtarget, DAG);
6023 // Without AVX, we need to extend to a 128-bit vector and then
6024 // insert into the 256-bit vector.
6025 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32, Item);
6026 SDValue ZeroVec = getZeroVector(MVT::v8i32, Subtarget, DAG, dl);
6027 Item = Insert128BitVector(ZeroVec, Item, 0, DAG, dl);
6030 assert(VT.is128BitVector() && "Expected an SSE value type!");
6031 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32, Item);
6032 Item = getShuffleVectorZeroOrUndef(Item, 0, true, Subtarget, DAG);
6034 return DAG.getBitcast(VT, Item);
6038 // Is it a vector logical left shift?
6039 if (NumElems == 2 && Idx == 1 &&
6040 X86::isZeroNode(Op.getOperand(0)) &&
6041 !X86::isZeroNode(Op.getOperand(1))) {
6042 unsigned NumBits = VT.getSizeInBits();
6043 return getVShift(true, VT,
6044 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
6045 VT, Op.getOperand(1)),
6046 NumBits/2, DAG, *this, dl);
6049 if (IsAllConstants) // Otherwise, it's better to do a constpool load.
6052 // Otherwise, if this is a vector with i32 or f32 elements, and the element
6053 // is a non-constant being inserted into an element other than the low one,
6054 // we can't use a constant pool load. Instead, use SCALAR_TO_VECTOR (aka
6055 // movd/movss) to move this into the low element, then shuffle it into
6057 if (EVTBits == 32) {
6058 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
6059 return getShuffleVectorZeroOrUndef(Item, Idx, NumZero > 0, Subtarget, DAG);
6063 // Splat is obviously ok. Let legalizer expand it to a shuffle.
6064 if (Values.size() == 1) {
6065 if (EVTBits == 32) {
6066 // Instead of a shuffle like this:
6067 // shuffle (scalar_to_vector (load (ptr + 4))), undef, <0, 0, 0, 0>
6068 // Check if it's possible to issue this instead.
6069 // shuffle (vload ptr)), undef, <1, 1, 1, 1>
6070 unsigned Idx = countTrailingZeros(NonZeros);
6071 SDValue Item = Op.getOperand(Idx);
6072 if (Op.getNode()->isOnlyUserOf(Item.getNode()))
6073 return LowerAsSplatVectorLoad(Item, VT, dl, DAG);
6078 // A vector full of immediates; various special cases are already
6079 // handled, so this is best done with a single constant-pool load.
6083 // For AVX-length vectors, see if we can use a vector load to get all of the
6084 // elements, otherwise build the individual 128-bit pieces and use
6085 // shuffles to put them in place.
6086 if (VT.is256BitVector() || VT.is512BitVector()) {
6087 SmallVector<SDValue, 64> V(Op->op_begin(), Op->op_begin() + NumElems);
6089 // Check for a build vector of consecutive loads.
6090 if (SDValue LD = EltsFromConsecutiveLoads(VT, V, dl, DAG, false))
6093 EVT HVT = EVT::getVectorVT(*DAG.getContext(), ExtVT, NumElems/2);
6095 // Build both the lower and upper subvector.
6096 SDValue Lower = DAG.getNode(ISD::BUILD_VECTOR, dl, HVT,
6097 makeArrayRef(&V[0], NumElems/2));
6098 SDValue Upper = DAG.getNode(ISD::BUILD_VECTOR, dl, HVT,
6099 makeArrayRef(&V[NumElems / 2], NumElems/2));
6101 // Recreate the wider vector with the lower and upper part.
6102 if (VT.is256BitVector())
6103 return Concat128BitVectors(Lower, Upper, VT, NumElems, DAG, dl);
6104 return Concat256BitVectors(Lower, Upper, VT, NumElems, DAG, dl);
6107 // Let legalizer expand 2-wide build_vectors.
6108 if (EVTBits == 64) {
6109 if (NumNonZero == 1) {
6110 // One half is zero or undef.
6111 unsigned Idx = countTrailingZeros(NonZeros);
6112 SDValue V2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT,
6113 Op.getOperand(Idx));
6114 return getShuffleVectorZeroOrUndef(V2, Idx, true, Subtarget, DAG);
6119 // If element VT is < 32 bits, convert it to inserts into a zero vector.
6120 if (EVTBits == 8 && NumElems == 16)
6121 if (SDValue V = LowerBuildVectorv16i8(Op, NonZeros,NumNonZero,NumZero, DAG,
6125 if (EVTBits == 16 && NumElems == 8)
6126 if (SDValue V = LowerBuildVectorv8i16(Op, NonZeros,NumNonZero,NumZero, DAG,
6130 // If element VT is == 32 bits and has 4 elems, try to generate an INSERTPS
6131 if (EVTBits == 32 && NumElems == 4)
6132 if (SDValue V = LowerBuildVectorv4x32(Op, DAG, Subtarget, *this))
6135 // If element VT is == 32 bits, turn it into a number of shuffles.
6136 SmallVector<SDValue, 8> V(NumElems);
6137 if (NumElems == 4 && NumZero > 0) {
6138 for (unsigned i = 0; i < 4; ++i) {
6139 bool isZero = !(NonZeros & (1 << i));
6141 V[i] = getZeroVector(VT, Subtarget, DAG, dl);
6143 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(i));
6146 for (unsigned i = 0; i < 2; ++i) {
6147 switch ((NonZeros & (0x3 << i*2)) >> (i*2)) {
6150 V[i] = V[i*2]; // Must be a zero vector.
6153 V[i] = getMOVL(DAG, dl, VT, V[i*2+1], V[i*2]);
6156 V[i] = getMOVL(DAG, dl, VT, V[i*2], V[i*2+1]);
6159 V[i] = getUnpackl(DAG, dl, VT, V[i*2], V[i*2+1]);
6164 bool Reverse1 = (NonZeros & 0x3) == 2;
6165 bool Reverse2 = ((NonZeros & (0x3 << 2)) >> 2) == 2;
6169 static_cast<int>(Reverse2 ? NumElems+1 : NumElems),
6170 static_cast<int>(Reverse2 ? NumElems : NumElems+1)
6172 return DAG.getVectorShuffle(VT, dl, V[0], V[1], &MaskVec[0]);
6175 if (Values.size() > 1 && VT.is128BitVector()) {
6176 // Check for a build vector of consecutive loads.
6177 for (unsigned i = 0; i < NumElems; ++i)
6178 V[i] = Op.getOperand(i);
6180 // Check for elements which are consecutive loads.
6181 if (SDValue LD = EltsFromConsecutiveLoads(VT, V, dl, DAG, false))
6184 // Check for a build vector from mostly shuffle plus few inserting.
6185 if (SDValue Sh = buildFromShuffleMostly(Op, DAG))
6188 // For SSE 4.1, use insertps to put the high elements into the low element.
6189 if (Subtarget->hasSSE41()) {
6191 if (Op.getOperand(0).getOpcode() != ISD::UNDEF)
6192 Result = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(0));
6194 Result = DAG.getUNDEF(VT);
6196 for (unsigned i = 1; i < NumElems; ++i) {
6197 if (Op.getOperand(i).getOpcode() == ISD::UNDEF) continue;
6198 Result = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, Result,
6199 Op.getOperand(i), DAG.getIntPtrConstant(i, dl));
6204 // Otherwise, expand into a number of unpckl*, start by extending each of
6205 // our (non-undef) elements to the full vector width with the element in the
6206 // bottom slot of the vector (which generates no code for SSE).
6207 for (unsigned i = 0; i < NumElems; ++i) {
6208 if (Op.getOperand(i).getOpcode() != ISD::UNDEF)
6209 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(i));
6211 V[i] = DAG.getUNDEF(VT);
6214 // Next, we iteratively mix elements, e.g. for v4f32:
6215 // Step 1: unpcklps 0, 2 ==> X: <?, ?, 2, 0>
6216 // : unpcklps 1, 3 ==> Y: <?, ?, 3, 1>
6217 // Step 2: unpcklps X, Y ==> <3, 2, 1, 0>
6218 unsigned EltStride = NumElems >> 1;
6219 while (EltStride != 0) {
6220 for (unsigned i = 0; i < EltStride; ++i) {
6221 // If V[i+EltStride] is undef and this is the first round of mixing,
6222 // then it is safe to just drop this shuffle: V[i] is already in the
6223 // right place, the one element (since it's the first round) being
6224 // inserted as undef can be dropped. This isn't safe for successive
6225 // rounds because they will permute elements within both vectors.
6226 if (V[i+EltStride].getOpcode() == ISD::UNDEF &&
6227 EltStride == NumElems/2)
6230 V[i] = getUnpackl(DAG, dl, VT, V[i], V[i + EltStride]);
6239 // 256-bit AVX can use the vinsertf128 instruction
6240 // to create 256-bit vectors from two other 128-bit ones.
6241 static SDValue LowerAVXCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) {
6243 MVT ResVT = Op.getSimpleValueType();
6245 assert((ResVT.is256BitVector() ||
6246 ResVT.is512BitVector()) && "Value type must be 256-/512-bit wide");
6248 SDValue V1 = Op.getOperand(0);
6249 SDValue V2 = Op.getOperand(1);
6250 unsigned NumElems = ResVT.getVectorNumElements();
6251 if (ResVT.is256BitVector())
6252 return Concat128BitVectors(V1, V2, ResVT, NumElems, DAG, dl);
6254 if (Op.getNumOperands() == 4) {
6255 MVT HalfVT = MVT::getVectorVT(ResVT.getScalarType(),
6256 ResVT.getVectorNumElements()/2);
6257 SDValue V3 = Op.getOperand(2);
6258 SDValue V4 = Op.getOperand(3);
6259 return Concat256BitVectors(Concat128BitVectors(V1, V2, HalfVT, NumElems/2, DAG, dl),
6260 Concat128BitVectors(V3, V4, HalfVT, NumElems/2, DAG, dl), ResVT, NumElems, DAG, dl);
6262 return Concat256BitVectors(V1, V2, ResVT, NumElems, DAG, dl);
6265 static SDValue LowerCONCAT_VECTORSvXi1(SDValue Op,
6266 const X86Subtarget *Subtarget,
6267 SelectionDAG & DAG) {
6269 MVT ResVT = Op.getSimpleValueType();
6270 unsigned NumOfOperands = Op.getNumOperands();
6272 assert(isPowerOf2_32(NumOfOperands) &&
6273 "Unexpected number of operands in CONCAT_VECTORS");
6275 if (NumOfOperands > 2) {
6276 MVT HalfVT = MVT::getVectorVT(ResVT.getScalarType(),
6277 ResVT.getVectorNumElements()/2);
6278 SmallVector<SDValue, 2> Ops;
6279 for (unsigned i = 0; i < NumOfOperands/2; i++)
6280 Ops.push_back(Op.getOperand(i));
6281 SDValue Lo = DAG.getNode(ISD::CONCAT_VECTORS, dl, HalfVT, Ops);
6283 for (unsigned i = NumOfOperands/2; i < NumOfOperands; i++)
6284 Ops.push_back(Op.getOperand(i));
6285 SDValue Hi = DAG.getNode(ISD::CONCAT_VECTORS, dl, HalfVT, Ops);
6286 return DAG.getNode(ISD::CONCAT_VECTORS, dl, ResVT, Lo, Hi);
6289 SDValue V1 = Op.getOperand(0);
6290 SDValue V2 = Op.getOperand(1);
6291 bool IsZeroV1 = ISD::isBuildVectorAllZeros(V1.getNode());
6292 bool IsZeroV2 = ISD::isBuildVectorAllZeros(V2.getNode());
6294 if (IsZeroV1 && IsZeroV2)
6295 return getZeroVector(ResVT, Subtarget, DAG, dl);
6297 SDValue ZeroIdx = DAG.getIntPtrConstant(0, dl);
6298 SDValue Undef = DAG.getUNDEF(ResVT);
6299 unsigned NumElems = ResVT.getVectorNumElements();
6300 SDValue ShiftBits = DAG.getConstant(NumElems/2, dl, MVT::i8);
6302 V2 = DAG.getNode(ISD::INSERT_SUBVECTOR, dl, ResVT, Undef, V2, ZeroIdx);
6303 V2 = DAG.getNode(X86ISD::VSHLI, dl, ResVT, V2, ShiftBits);
6307 V1 = DAG.getNode(ISD::INSERT_SUBVECTOR, dl, ResVT, Undef, V1, ZeroIdx);
6308 // Zero the upper bits of V1
6309 V1 = DAG.getNode(X86ISD::VSHLI, dl, ResVT, V1, ShiftBits);
6310 V1 = DAG.getNode(X86ISD::VSRLI, dl, ResVT, V1, ShiftBits);
6313 return DAG.getNode(ISD::OR, dl, ResVT, V1, V2);
6316 static SDValue LowerCONCAT_VECTORS(SDValue Op,
6317 const X86Subtarget *Subtarget,
6318 SelectionDAG &DAG) {
6319 MVT VT = Op.getSimpleValueType();
6320 if (VT.getVectorElementType() == MVT::i1)
6321 return LowerCONCAT_VECTORSvXi1(Op, Subtarget, DAG);
6323 assert((VT.is256BitVector() && Op.getNumOperands() == 2) ||
6324 (VT.is512BitVector() && (Op.getNumOperands() == 2 ||
6325 Op.getNumOperands() == 4)));
6327 // AVX can use the vinsertf128 instruction to create 256-bit vectors
6328 // from two other 128-bit ones.
6330 // 512-bit vector may contain 2 256-bit vectors or 4 128-bit vectors
6331 return LowerAVXCONCAT_VECTORS(Op, DAG);
6335 //===----------------------------------------------------------------------===//
6336 // Vector shuffle lowering
6338 // This is an experimental code path for lowering vector shuffles on x86. It is
6339 // designed to handle arbitrary vector shuffles and blends, gracefully
6340 // degrading performance as necessary. It works hard to recognize idiomatic
6341 // shuffles and lower them to optimal instruction patterns without leaving
6342 // a framework that allows reasonably efficient handling of all vector shuffle
6344 //===----------------------------------------------------------------------===//
6346 /// \brief Tiny helper function to identify a no-op mask.
6348 /// This is a somewhat boring predicate function. It checks whether the mask
6349 /// array input, which is assumed to be a single-input shuffle mask of the kind
6350 /// used by the X86 shuffle instructions (not a fully general
6351 /// ShuffleVectorSDNode mask) requires any shuffles to occur. Both undef and an
6352 /// in-place shuffle are 'no-op's.
6353 static bool isNoopShuffleMask(ArrayRef<int> Mask) {
6354 for (int i = 0, Size = Mask.size(); i < Size; ++i)
6355 if (Mask[i] != -1 && Mask[i] != i)
6360 /// \brief Helper function to classify a mask as a single-input mask.
6362 /// This isn't a generic single-input test because in the vector shuffle
6363 /// lowering we canonicalize single inputs to be the first input operand. This
6364 /// means we can more quickly test for a single input by only checking whether
6365 /// an input from the second operand exists. We also assume that the size of
6366 /// mask corresponds to the size of the input vectors which isn't true in the
6367 /// fully general case.
6368 static bool isSingleInputShuffleMask(ArrayRef<int> Mask) {
6370 if (M >= (int)Mask.size())
6375 /// \brief Test whether there are elements crossing 128-bit lanes in this
6378 /// X86 divides up its shuffles into in-lane and cross-lane shuffle operations
6379 /// and we routinely test for these.
6380 static bool is128BitLaneCrossingShuffleMask(MVT VT, ArrayRef<int> Mask) {
6381 int LaneSize = 128 / VT.getScalarSizeInBits();
6382 int Size = Mask.size();
6383 for (int i = 0; i < Size; ++i)
6384 if (Mask[i] >= 0 && (Mask[i] % Size) / LaneSize != i / LaneSize)
6389 /// \brief Test whether a shuffle mask is equivalent within each 128-bit lane.
6391 /// This checks a shuffle mask to see if it is performing the same
6392 /// 128-bit lane-relative shuffle in each 128-bit lane. This trivially implies
6393 /// that it is also not lane-crossing. It may however involve a blend from the
6394 /// same lane of a second vector.
6396 /// The specific repeated shuffle mask is populated in \p RepeatedMask, as it is
6397 /// non-trivial to compute in the face of undef lanes. The representation is
6398 /// *not* suitable for use with existing 128-bit shuffles as it will contain
6399 /// entries from both V1 and V2 inputs to the wider mask.
6401 is128BitLaneRepeatedShuffleMask(MVT VT, ArrayRef<int> Mask,
6402 SmallVectorImpl<int> &RepeatedMask) {
6403 int LaneSize = 128 / VT.getScalarSizeInBits();
6404 RepeatedMask.resize(LaneSize, -1);
6405 int Size = Mask.size();
6406 for (int i = 0; i < Size; ++i) {
6409 if ((Mask[i] % Size) / LaneSize != i / LaneSize)
6410 // This entry crosses lanes, so there is no way to model this shuffle.
6413 // Ok, handle the in-lane shuffles by detecting if and when they repeat.
6414 if (RepeatedMask[i % LaneSize] == -1)
6415 // This is the first non-undef entry in this slot of a 128-bit lane.
6416 RepeatedMask[i % LaneSize] =
6417 Mask[i] < Size ? Mask[i] % LaneSize : Mask[i] % LaneSize + Size;
6418 else if (RepeatedMask[i % LaneSize] + (i / LaneSize) * LaneSize != Mask[i])
6419 // Found a mismatch with the repeated mask.
6425 /// \brief Checks whether a shuffle mask is equivalent to an explicit list of
6428 /// This is a fast way to test a shuffle mask against a fixed pattern:
6430 /// if (isShuffleEquivalent(Mask, 3, 2, {1, 0})) { ... }
6432 /// It returns true if the mask is exactly as wide as the argument list, and
6433 /// each element of the mask is either -1 (signifying undef) or the value given
6434 /// in the argument.
6435 static bool isShuffleEquivalent(SDValue V1, SDValue V2, ArrayRef<int> Mask,
6436 ArrayRef<int> ExpectedMask) {
6437 if (Mask.size() != ExpectedMask.size())
6440 int Size = Mask.size();
6442 // If the values are build vectors, we can look through them to find
6443 // equivalent inputs that make the shuffles equivalent.
6444 auto *BV1 = dyn_cast<BuildVectorSDNode>(V1);
6445 auto *BV2 = dyn_cast<BuildVectorSDNode>(V2);
6447 for (int i = 0; i < Size; ++i)
6448 if (Mask[i] != -1 && Mask[i] != ExpectedMask[i]) {
6449 auto *MaskBV = Mask[i] < Size ? BV1 : BV2;
6450 auto *ExpectedBV = ExpectedMask[i] < Size ? BV1 : BV2;
6451 if (!MaskBV || !ExpectedBV ||
6452 MaskBV->getOperand(Mask[i] % Size) !=
6453 ExpectedBV->getOperand(ExpectedMask[i] % Size))
6460 /// \brief Get a 4-lane 8-bit shuffle immediate for a mask.
6462 /// This helper function produces an 8-bit shuffle immediate corresponding to
6463 /// the ubiquitous shuffle encoding scheme used in x86 instructions for
6464 /// shuffling 4 lanes. It can be used with most of the PSHUF instructions for
6467 /// NB: We rely heavily on "undef" masks preserving the input lane.
6468 static SDValue getV4X86ShuffleImm8ForMask(ArrayRef<int> Mask, SDLoc DL,
6469 SelectionDAG &DAG) {
6470 assert(Mask.size() == 4 && "Only 4-lane shuffle masks");
6471 assert(Mask[0] >= -1 && Mask[0] < 4 && "Out of bound mask element!");
6472 assert(Mask[1] >= -1 && Mask[1] < 4 && "Out of bound mask element!");
6473 assert(Mask[2] >= -1 && Mask[2] < 4 && "Out of bound mask element!");
6474 assert(Mask[3] >= -1 && Mask[3] < 4 && "Out of bound mask element!");
6477 Imm |= (Mask[0] == -1 ? 0 : Mask[0]) << 0;
6478 Imm |= (Mask[1] == -1 ? 1 : Mask[1]) << 2;
6479 Imm |= (Mask[2] == -1 ? 2 : Mask[2]) << 4;
6480 Imm |= (Mask[3] == -1 ? 3 : Mask[3]) << 6;
6481 return DAG.getConstant(Imm, DL, MVT::i8);
6484 /// \brief Compute whether each element of a shuffle is zeroable.
6486 /// A "zeroable" vector shuffle element is one which can be lowered to zero.
6487 /// Either it is an undef element in the shuffle mask, the element of the input
6488 /// referenced is undef, or the element of the input referenced is known to be
6489 /// zero. Many x86 shuffles can zero lanes cheaply and we often want to handle
6490 /// as many lanes with this technique as possible to simplify the remaining
6492 static SmallBitVector computeZeroableShuffleElements(ArrayRef<int> Mask,
6493 SDValue V1, SDValue V2) {
6494 SmallBitVector Zeroable(Mask.size(), false);
6496 while (V1.getOpcode() == ISD::BITCAST)
6497 V1 = V1->getOperand(0);
6498 while (V2.getOpcode() == ISD::BITCAST)
6499 V2 = V2->getOperand(0);
6501 bool V1IsZero = ISD::isBuildVectorAllZeros(V1.getNode());
6502 bool V2IsZero = ISD::isBuildVectorAllZeros(V2.getNode());
6504 for (int i = 0, Size = Mask.size(); i < Size; ++i) {
6506 // Handle the easy cases.
6507 if (M < 0 || (M >= 0 && M < Size && V1IsZero) || (M >= Size && V2IsZero)) {
6512 // If this is an index into a build_vector node (which has the same number
6513 // of elements), dig out the input value and use it.
6514 SDValue V = M < Size ? V1 : V2;
6515 if (V.getOpcode() != ISD::BUILD_VECTOR || Size != (int)V.getNumOperands())
6518 SDValue Input = V.getOperand(M % Size);
6519 // The UNDEF opcode check really should be dead code here, but not quite
6520 // worth asserting on (it isn't invalid, just unexpected).
6521 if (Input.getOpcode() == ISD::UNDEF || X86::isZeroNode(Input))
6528 /// \brief Try to emit a bitmask instruction for a shuffle.
6530 /// This handles cases where we can model a blend exactly as a bitmask due to
6531 /// one of the inputs being zeroable.
6532 static SDValue lowerVectorShuffleAsBitMask(SDLoc DL, MVT VT, SDValue V1,
6533 SDValue V2, ArrayRef<int> Mask,
6534 SelectionDAG &DAG) {
6535 MVT EltVT = VT.getScalarType();
6536 int NumEltBits = EltVT.getSizeInBits();
6537 MVT IntEltVT = MVT::getIntegerVT(NumEltBits);
6538 SDValue Zero = DAG.getConstant(0, DL, IntEltVT);
6539 SDValue AllOnes = DAG.getConstant(APInt::getAllOnesValue(NumEltBits), DL,
6541 if (EltVT.isFloatingPoint()) {
6542 Zero = DAG.getBitcast(EltVT, Zero);
6543 AllOnes = DAG.getBitcast(EltVT, AllOnes);
6545 SmallVector<SDValue, 16> VMaskOps(Mask.size(), Zero);
6546 SmallBitVector Zeroable = computeZeroableShuffleElements(Mask, V1, V2);
6548 for (int i = 0, Size = Mask.size(); i < Size; ++i) {
6551 if (Mask[i] % Size != i)
6552 return SDValue(); // Not a blend.
6554 V = Mask[i] < Size ? V1 : V2;
6555 else if (V != (Mask[i] < Size ? V1 : V2))
6556 return SDValue(); // Can only let one input through the mask.
6558 VMaskOps[i] = AllOnes;
6561 return SDValue(); // No non-zeroable elements!
6563 SDValue VMask = DAG.getNode(ISD::BUILD_VECTOR, DL, VT, VMaskOps);
6564 V = DAG.getNode(VT.isFloatingPoint()
6565 ? (unsigned) X86ISD::FAND : (unsigned) ISD::AND,
6570 /// \brief Try to emit a blend instruction for a shuffle using bit math.
6572 /// This is used as a fallback approach when first class blend instructions are
6573 /// unavailable. Currently it is only suitable for integer vectors, but could
6574 /// be generalized for floating point vectors if desirable.
6575 static SDValue lowerVectorShuffleAsBitBlend(SDLoc DL, MVT VT, SDValue V1,
6576 SDValue V2, ArrayRef<int> Mask,
6577 SelectionDAG &DAG) {
6578 assert(VT.isInteger() && "Only supports integer vector types!");
6579 MVT EltVT = VT.getScalarType();
6580 int NumEltBits = EltVT.getSizeInBits();
6581 SDValue Zero = DAG.getConstant(0, DL, EltVT);
6582 SDValue AllOnes = DAG.getConstant(APInt::getAllOnesValue(NumEltBits), DL,
6584 SmallVector<SDValue, 16> MaskOps;
6585 for (int i = 0, Size = Mask.size(); i < Size; ++i) {
6586 if (Mask[i] != -1 && Mask[i] != i && Mask[i] != i + Size)
6587 return SDValue(); // Shuffled input!
6588 MaskOps.push_back(Mask[i] < Size ? AllOnes : Zero);
6591 SDValue V1Mask = DAG.getNode(ISD::BUILD_VECTOR, DL, VT, MaskOps);
6592 V1 = DAG.getNode(ISD::AND, DL, VT, V1, V1Mask);
6593 // We have to cast V2 around.
6594 MVT MaskVT = MVT::getVectorVT(MVT::i64, VT.getSizeInBits() / 64);
6595 V2 = DAG.getBitcast(VT, DAG.getNode(X86ISD::ANDNP, DL, MaskVT,
6596 DAG.getBitcast(MaskVT, V1Mask),
6597 DAG.getBitcast(MaskVT, V2)));
6598 return DAG.getNode(ISD::OR, DL, VT, V1, V2);
6601 /// \brief Try to emit a blend instruction for a shuffle.
6603 /// This doesn't do any checks for the availability of instructions for blending
6604 /// these values. It relies on the availability of the X86ISD::BLENDI pattern to
6605 /// be matched in the backend with the type given. What it does check for is
6606 /// that the shuffle mask is in fact a blend.
6607 static SDValue lowerVectorShuffleAsBlend(SDLoc DL, MVT VT, SDValue V1,
6608 SDValue V2, ArrayRef<int> Mask,
6609 const X86Subtarget *Subtarget,
6610 SelectionDAG &DAG) {
6611 unsigned BlendMask = 0;
6612 for (int i = 0, Size = Mask.size(); i < Size; ++i) {
6613 if (Mask[i] >= Size) {
6614 if (Mask[i] != i + Size)
6615 return SDValue(); // Shuffled V2 input!
6616 BlendMask |= 1u << i;
6619 if (Mask[i] >= 0 && Mask[i] != i)
6620 return SDValue(); // Shuffled V1 input!
6622 switch (VT.SimpleTy) {
6627 return DAG.getNode(X86ISD::BLENDI, DL, VT, V1, V2,
6628 DAG.getConstant(BlendMask, DL, MVT::i8));
6632 assert(Subtarget->hasAVX2() && "256-bit integer blends require AVX2!");
6636 // If we have AVX2 it is faster to use VPBLENDD when the shuffle fits into
6637 // that instruction.
6638 if (Subtarget->hasAVX2()) {
6639 // Scale the blend by the number of 32-bit dwords per element.
6640 int Scale = VT.getScalarSizeInBits() / 32;
6642 for (int i = 0, Size = Mask.size(); i < Size; ++i)
6643 if (Mask[i] >= Size)
6644 for (int j = 0; j < Scale; ++j)
6645 BlendMask |= 1u << (i * Scale + j);
6647 MVT BlendVT = VT.getSizeInBits() > 128 ? MVT::v8i32 : MVT::v4i32;
6648 V1 = DAG.getBitcast(BlendVT, V1);
6649 V2 = DAG.getBitcast(BlendVT, V2);
6650 return DAG.getBitcast(
6651 VT, DAG.getNode(X86ISD::BLENDI, DL, BlendVT, V1, V2,
6652 DAG.getConstant(BlendMask, DL, MVT::i8)));
6656 // For integer shuffles we need to expand the mask and cast the inputs to
6657 // v8i16s prior to blending.
6658 int Scale = 8 / VT.getVectorNumElements();
6660 for (int i = 0, Size = Mask.size(); i < Size; ++i)
6661 if (Mask[i] >= Size)
6662 for (int j = 0; j < Scale; ++j)
6663 BlendMask |= 1u << (i * Scale + j);
6665 V1 = DAG.getBitcast(MVT::v8i16, V1);
6666 V2 = DAG.getBitcast(MVT::v8i16, V2);
6667 return DAG.getBitcast(VT,
6668 DAG.getNode(X86ISD::BLENDI, DL, MVT::v8i16, V1, V2,
6669 DAG.getConstant(BlendMask, DL, MVT::i8)));
6673 assert(Subtarget->hasAVX2() && "256-bit integer blends require AVX2!");
6674 SmallVector<int, 8> RepeatedMask;
6675 if (is128BitLaneRepeatedShuffleMask(MVT::v16i16, Mask, RepeatedMask)) {
6676 // We can lower these with PBLENDW which is mirrored across 128-bit lanes.
6677 assert(RepeatedMask.size() == 8 && "Repeated mask size doesn't match!");
6679 for (int i = 0; i < 8; ++i)
6680 if (RepeatedMask[i] >= 16)
6681 BlendMask |= 1u << i;
6682 return DAG.getNode(X86ISD::BLENDI, DL, MVT::v16i16, V1, V2,
6683 DAG.getConstant(BlendMask, DL, MVT::i8));
6689 assert((VT.getSizeInBits() == 128 || Subtarget->hasAVX2()) &&
6690 "256-bit byte-blends require AVX2 support!");
6692 // Attempt to lower to a bitmask if we can. VPAND is faster than VPBLENDVB.
6693 if (SDValue Masked = lowerVectorShuffleAsBitMask(DL, VT, V1, V2, Mask, DAG))
6696 // Scale the blend by the number of bytes per element.
6697 int Scale = VT.getScalarSizeInBits() / 8;
6699 // This form of blend is always done on bytes. Compute the byte vector
6701 MVT BlendVT = MVT::getVectorVT(MVT::i8, VT.getSizeInBits() / 8);
6703 // Compute the VSELECT mask. Note that VSELECT is really confusing in the
6704 // mix of LLVM's code generator and the x86 backend. We tell the code
6705 // generator that boolean values in the elements of an x86 vector register
6706 // are -1 for true and 0 for false. We then use the LLVM semantics of 'true'
6707 // mapping a select to operand #1, and 'false' mapping to operand #2. The
6708 // reality in x86 is that vector masks (pre-AVX-512) use only the high bit
6709 // of the element (the remaining are ignored) and 0 in that high bit would
6710 // mean operand #1 while 1 in the high bit would mean operand #2. So while
6711 // the LLVM model for boolean values in vector elements gets the relevant
6712 // bit set, it is set backwards and over constrained relative to x86's
6714 SmallVector<SDValue, 32> VSELECTMask;
6715 for (int i = 0, Size = Mask.size(); i < Size; ++i)
6716 for (int j = 0; j < Scale; ++j)
6717 VSELECTMask.push_back(
6718 Mask[i] < 0 ? DAG.getUNDEF(MVT::i8)
6719 : DAG.getConstant(Mask[i] < Size ? -1 : 0, DL,
6722 V1 = DAG.getBitcast(BlendVT, V1);
6723 V2 = DAG.getBitcast(BlendVT, V2);
6724 return DAG.getBitcast(VT, DAG.getNode(ISD::VSELECT, DL, BlendVT,
6725 DAG.getNode(ISD::BUILD_VECTOR, DL,
6726 BlendVT, VSELECTMask),
6731 llvm_unreachable("Not a supported integer vector type!");
6735 /// \brief Try to lower as a blend of elements from two inputs followed by
6736 /// a single-input permutation.
6738 /// This matches the pattern where we can blend elements from two inputs and
6739 /// then reduce the shuffle to a single-input permutation.
6740 static SDValue lowerVectorShuffleAsBlendAndPermute(SDLoc DL, MVT VT, SDValue V1,
6743 SelectionDAG &DAG) {
6744 // We build up the blend mask while checking whether a blend is a viable way
6745 // to reduce the shuffle.
6746 SmallVector<int, 32> BlendMask(Mask.size(), -1);
6747 SmallVector<int, 32> PermuteMask(Mask.size(), -1);
6749 for (int i = 0, Size = Mask.size(); i < Size; ++i) {
6753 assert(Mask[i] < Size * 2 && "Shuffle input is out of bounds.");
6755 if (BlendMask[Mask[i] % Size] == -1)
6756 BlendMask[Mask[i] % Size] = Mask[i];
6757 else if (BlendMask[Mask[i] % Size] != Mask[i])
6758 return SDValue(); // Can't blend in the needed input!
6760 PermuteMask[i] = Mask[i] % Size;
6763 SDValue V = DAG.getVectorShuffle(VT, DL, V1, V2, BlendMask);
6764 return DAG.getVectorShuffle(VT, DL, V, DAG.getUNDEF(VT), PermuteMask);
6767 /// \brief Generic routine to decompose a shuffle and blend into indepndent
6768 /// blends and permutes.
6770 /// This matches the extremely common pattern for handling combined
6771 /// shuffle+blend operations on newer X86 ISAs where we have very fast blend
6772 /// operations. It will try to pick the best arrangement of shuffles and
6774 static SDValue lowerVectorShuffleAsDecomposedShuffleBlend(SDLoc DL, MVT VT,
6778 SelectionDAG &DAG) {
6779 // Shuffle the input elements into the desired positions in V1 and V2 and
6780 // blend them together.
6781 SmallVector<int, 32> V1Mask(Mask.size(), -1);
6782 SmallVector<int, 32> V2Mask(Mask.size(), -1);
6783 SmallVector<int, 32> BlendMask(Mask.size(), -1);
6784 for (int i = 0, Size = Mask.size(); i < Size; ++i)
6785 if (Mask[i] >= 0 && Mask[i] < Size) {
6786 V1Mask[i] = Mask[i];
6788 } else if (Mask[i] >= Size) {
6789 V2Mask[i] = Mask[i] - Size;
6790 BlendMask[i] = i + Size;
6793 // Try to lower with the simpler initial blend strategy unless one of the
6794 // input shuffles would be a no-op. We prefer to shuffle inputs as the
6795 // shuffle may be able to fold with a load or other benefit. However, when
6796 // we'll have to do 2x as many shuffles in order to achieve this, blending
6797 // first is a better strategy.
6798 if (!isNoopShuffleMask(V1Mask) && !isNoopShuffleMask(V2Mask))
6799 if (SDValue BlendPerm =
6800 lowerVectorShuffleAsBlendAndPermute(DL, VT, V1, V2, Mask, DAG))
6803 V1 = DAG.getVectorShuffle(VT, DL, V1, DAG.getUNDEF(VT), V1Mask);
6804 V2 = DAG.getVectorShuffle(VT, DL, V2, DAG.getUNDEF(VT), V2Mask);
6805 return DAG.getVectorShuffle(VT, DL, V1, V2, BlendMask);
6808 /// \brief Try to lower a vector shuffle as a byte rotation.
6810 /// SSSE3 has a generic PALIGNR instruction in x86 that will do an arbitrary
6811 /// byte-rotation of the concatenation of two vectors; pre-SSSE3 can use
6812 /// a PSRLDQ/PSLLDQ/POR pattern to get a similar effect. This routine will
6813 /// try to generically lower a vector shuffle through such an pattern. It
6814 /// does not check for the profitability of lowering either as PALIGNR or
6815 /// PSRLDQ/PSLLDQ/POR, only whether the mask is valid to lower in that form.
6816 /// This matches shuffle vectors that look like:
6818 /// v8i16 [11, 12, 13, 14, 15, 0, 1, 2]
6820 /// Essentially it concatenates V1 and V2, shifts right by some number of
6821 /// elements, and takes the low elements as the result. Note that while this is
6822 /// specified as a *right shift* because x86 is little-endian, it is a *left
6823 /// rotate* of the vector lanes.
6824 static SDValue lowerVectorShuffleAsByteRotate(SDLoc DL, MVT VT, SDValue V1,
6827 const X86Subtarget *Subtarget,
6828 SelectionDAG &DAG) {
6829 assert(!isNoopShuffleMask(Mask) && "We shouldn't lower no-op shuffles!");
6831 int NumElts = Mask.size();
6832 int NumLanes = VT.getSizeInBits() / 128;
6833 int NumLaneElts = NumElts / NumLanes;
6835 // We need to detect various ways of spelling a rotation:
6836 // [11, 12, 13, 14, 15, 0, 1, 2]
6837 // [-1, 12, 13, 14, -1, -1, 1, -1]
6838 // [-1, -1, -1, -1, -1, -1, 1, 2]
6839 // [ 3, 4, 5, 6, 7, 8, 9, 10]
6840 // [-1, 4, 5, 6, -1, -1, 9, -1]
6841 // [-1, 4, 5, 6, -1, -1, -1, -1]
6844 for (int l = 0; l < NumElts; l += NumLaneElts) {
6845 for (int i = 0; i < NumLaneElts; ++i) {
6846 if (Mask[l + i] == -1)
6848 assert(Mask[l + i] >= 0 && "Only -1 is a valid negative mask element!");
6850 // Get the mod-Size index and lane correct it.
6851 int LaneIdx = (Mask[l + i] % NumElts) - l;
6852 // Make sure it was in this lane.
6853 if (LaneIdx < 0 || LaneIdx >= NumLaneElts)
6856 // Determine where a rotated vector would have started.
6857 int StartIdx = i - LaneIdx;
6859 // The identity rotation isn't interesting, stop.
6862 // If we found the tail of a vector the rotation must be the missing
6863 // front. If we found the head of a vector, it must be how much of the
6865 int CandidateRotation = StartIdx < 0 ? -StartIdx : NumLaneElts - StartIdx;
6868 Rotation = CandidateRotation;
6869 else if (Rotation != CandidateRotation)
6870 // The rotations don't match, so we can't match this mask.
6873 // Compute which value this mask is pointing at.
6874 SDValue MaskV = Mask[l + i] < NumElts ? V1 : V2;
6876 // Compute which of the two target values this index should be assigned
6877 // to. This reflects whether the high elements are remaining or the low
6878 // elements are remaining.
6879 SDValue &TargetV = StartIdx < 0 ? Hi : Lo;
6881 // Either set up this value if we've not encountered it before, or check
6882 // that it remains consistent.
6885 else if (TargetV != MaskV)
6886 // This may be a rotation, but it pulls from the inputs in some
6887 // unsupported interleaving.
6892 // Check that we successfully analyzed the mask, and normalize the results.
6893 assert(Rotation != 0 && "Failed to locate a viable rotation!");
6894 assert((Lo || Hi) && "Failed to find a rotated input vector!");
6900 // The actual rotate instruction rotates bytes, so we need to scale the
6901 // rotation based on how many bytes are in the vector lane.
6902 int Scale = 16 / NumLaneElts;
6904 // SSSE3 targets can use the palignr instruction.
6905 if (Subtarget->hasSSSE3()) {
6906 // Cast the inputs to i8 vector of correct length to match PALIGNR.
6907 MVT AlignVT = MVT::getVectorVT(MVT::i8, 16 * NumLanes);
6908 Lo = DAG.getBitcast(AlignVT, Lo);
6909 Hi = DAG.getBitcast(AlignVT, Hi);
6911 return DAG.getBitcast(
6912 VT, DAG.getNode(X86ISD::PALIGNR, DL, AlignVT, Hi, Lo,
6913 DAG.getConstant(Rotation * Scale, DL, MVT::i8)));
6916 assert(VT.getSizeInBits() == 128 &&
6917 "Rotate-based lowering only supports 128-bit lowering!");
6918 assert(Mask.size() <= 16 &&
6919 "Can shuffle at most 16 bytes in a 128-bit vector!");
6921 // Default SSE2 implementation
6922 int LoByteShift = 16 - Rotation * Scale;
6923 int HiByteShift = Rotation * Scale;
6925 // Cast the inputs to v2i64 to match PSLLDQ/PSRLDQ.
6926 Lo = DAG.getBitcast(MVT::v2i64, Lo);
6927 Hi = DAG.getBitcast(MVT::v2i64, Hi);
6929 SDValue LoShift = DAG.getNode(X86ISD::VSHLDQ, DL, MVT::v2i64, Lo,
6930 DAG.getConstant(LoByteShift, DL, MVT::i8));
6931 SDValue HiShift = DAG.getNode(X86ISD::VSRLDQ, DL, MVT::v2i64, Hi,
6932 DAG.getConstant(HiByteShift, DL, MVT::i8));
6933 return DAG.getBitcast(VT,
6934 DAG.getNode(ISD::OR, DL, MVT::v2i64, LoShift, HiShift));
6937 /// \brief Try to lower a vector shuffle as a bit shift (shifts in zeros).
6939 /// Attempts to match a shuffle mask against the PSLL(W/D/Q/DQ) and
6940 /// PSRL(W/D/Q/DQ) SSE2 and AVX2 logical bit-shift instructions. The function
6941 /// matches elements from one of the input vectors shuffled to the left or
6942 /// right with zeroable elements 'shifted in'. It handles both the strictly
6943 /// bit-wise element shifts and the byte shift across an entire 128-bit double
6946 /// PSHL : (little-endian) left bit shift.
6947 /// [ zz, 0, zz, 2 ]
6948 /// [ -1, 4, zz, -1 ]
6949 /// PSRL : (little-endian) right bit shift.
6951 /// [ -1, -1, 7, zz]
6952 /// PSLLDQ : (little-endian) left byte shift
6953 /// [ zz, 0, 1, 2, 3, 4, 5, 6]
6954 /// [ zz, zz, -1, -1, 2, 3, 4, -1]
6955 /// [ zz, zz, zz, zz, zz, zz, -1, 1]
6956 /// PSRLDQ : (little-endian) right byte shift
6957 /// [ 5, 6, 7, zz, zz, zz, zz, zz]
6958 /// [ -1, 5, 6, 7, zz, zz, zz, zz]
6959 /// [ 1, 2, -1, -1, -1, -1, zz, zz]
6960 static SDValue lowerVectorShuffleAsShift(SDLoc DL, MVT VT, SDValue V1,
6961 SDValue V2, ArrayRef<int> Mask,
6962 SelectionDAG &DAG) {
6963 SmallBitVector Zeroable = computeZeroableShuffleElements(Mask, V1, V2);
6965 int Size = Mask.size();
6966 assert(Size == (int)VT.getVectorNumElements() && "Unexpected mask size");
6968 auto CheckZeros = [&](int Shift, int Scale, bool Left) {
6969 for (int i = 0; i < Size; i += Scale)
6970 for (int j = 0; j < Shift; ++j)
6971 if (!Zeroable[i + j + (Left ? 0 : (Scale - Shift))])
6977 auto MatchShift = [&](int Shift, int Scale, bool Left, SDValue V) {
6978 for (int i = 0; i != Size; i += Scale) {
6979 unsigned Pos = Left ? i + Shift : i;
6980 unsigned Low = Left ? i : i + Shift;
6981 unsigned Len = Scale - Shift;
6982 if (!isSequentialOrUndefInRange(Mask, Pos, Len,
6983 Low + (V == V1 ? 0 : Size)))
6987 int ShiftEltBits = VT.getScalarSizeInBits() * Scale;
6988 bool ByteShift = ShiftEltBits > 64;
6989 unsigned OpCode = Left ? (ByteShift ? X86ISD::VSHLDQ : X86ISD::VSHLI)
6990 : (ByteShift ? X86ISD::VSRLDQ : X86ISD::VSRLI);
6991 int ShiftAmt = Shift * VT.getScalarSizeInBits() / (ByteShift ? 8 : 1);
6993 // Normalize the scale for byte shifts to still produce an i64 element
6995 Scale = ByteShift ? Scale / 2 : Scale;
6997 // We need to round trip through the appropriate type for the shift.
6998 MVT ShiftSVT = MVT::getIntegerVT(VT.getScalarSizeInBits() * Scale);
6999 MVT ShiftVT = MVT::getVectorVT(ShiftSVT, Size / Scale);
7000 assert(DAG.getTargetLoweringInfo().isTypeLegal(ShiftVT) &&
7001 "Illegal integer vector type");
7002 V = DAG.getBitcast(ShiftVT, V);
7004 V = DAG.getNode(OpCode, DL, ShiftVT, V,
7005 DAG.getConstant(ShiftAmt, DL, MVT::i8));
7006 return DAG.getBitcast(VT, V);
7009 // SSE/AVX supports logical shifts up to 64-bit integers - so we can just
7010 // keep doubling the size of the integer elements up to that. We can
7011 // then shift the elements of the integer vector by whole multiples of
7012 // their width within the elements of the larger integer vector. Test each
7013 // multiple to see if we can find a match with the moved element indices
7014 // and that the shifted in elements are all zeroable.
7015 for (int Scale = 2; Scale * VT.getScalarSizeInBits() <= 128; Scale *= 2)
7016 for (int Shift = 1; Shift != Scale; ++Shift)
7017 for (bool Left : {true, false})
7018 if (CheckZeros(Shift, Scale, Left))
7019 for (SDValue V : {V1, V2})
7020 if (SDValue Match = MatchShift(Shift, Scale, Left, V))
7027 /// \brief Try to lower a vector shuffle using SSE4a EXTRQ/INSERTQ.
7028 static SDValue lowerVectorShuffleWithSSE4A(SDLoc DL, MVT VT, SDValue V1,
7029 SDValue V2, ArrayRef<int> Mask,
7030 SelectionDAG &DAG) {
7031 SmallBitVector Zeroable = computeZeroableShuffleElements(Mask, V1, V2);
7032 assert(!Zeroable.all() && "Fully zeroable shuffle mask");
7034 int Size = Mask.size();
7035 int HalfSize = Size / 2;
7036 assert(Size == (int)VT.getVectorNumElements() && "Unexpected mask size");
7038 // Upper half must be undefined.
7039 if (!isUndefInRange(Mask, HalfSize, HalfSize))
7042 // EXTRQ: Extract Len elements from lower half of source, starting at Idx.
7043 // Remainder of lower half result is zero and upper half is all undef.
7044 auto LowerAsEXTRQ = [&]() {
7045 // Determine the extraction length from the part of the
7046 // lower half that isn't zeroable.
7048 for (; Len >= 0; --Len)
7049 if (!Zeroable[Len - 1])
7051 assert(Len > 0 && "Zeroable shuffle mask");
7053 // Attempt to match first Len sequential elements from the lower half.
7056 for (int i = 0; i != Len; ++i) {
7060 SDValue &V = (M < Size ? V1 : V2);
7063 // All mask elements must be in the lower half.
7067 if (Idx < 0 || (Src == V && Idx == (M - i))) {
7078 assert((Idx + Len) <= HalfSize && "Illegal extraction mask");
7079 int BitLen = (Len * VT.getScalarSizeInBits()) & 0x3f;
7080 int BitIdx = (Idx * VT.getScalarSizeInBits()) & 0x3f;
7081 return DAG.getNode(X86ISD::EXTRQI, DL, VT, Src,
7082 DAG.getConstant(BitLen, DL, MVT::i8),
7083 DAG.getConstant(BitIdx, DL, MVT::i8));
7086 if (SDValue ExtrQ = LowerAsEXTRQ())
7089 // INSERTQ: Extract lowest Len elements from lower half of second source and
7090 // insert over first source, starting at Idx.
7091 // { A[0], .., A[Idx-1], B[0], .., B[Len-1], A[Idx+Len], .., UNDEF, ... }
7092 auto LowerAsInsertQ = [&]() {
7093 for (int Idx = 0; Idx != HalfSize; ++Idx) {
7096 // Attempt to match first source from mask before insertion point.
7097 if (isUndefInRange(Mask, 0, Idx)) {
7099 } else if (isSequentialOrUndefInRange(Mask, 0, Idx, 0)) {
7101 } else if (isSequentialOrUndefInRange(Mask, 0, Idx, Size)) {
7107 // Extend the extraction length looking to match both the insertion of
7108 // the second source and the remaining elements of the first.
7109 for (int Hi = Idx + 1; Hi <= HalfSize; ++Hi) {
7114 if (isSequentialOrUndefInRange(Mask, Idx, Len, 0)) {
7116 } else if (isSequentialOrUndefInRange(Mask, Idx, Len, Size)) {
7122 // Match the remaining elements of the lower half.
7123 if (isUndefInRange(Mask, Hi, HalfSize - Hi)) {
7125 } else if ((!Base || (Base == V1)) &&
7126 isSequentialOrUndefInRange(Mask, Hi, HalfSize - Hi, Hi)) {
7128 } else if ((!Base || (Base == V2)) &&
7129 isSequentialOrUndefInRange(Mask, Hi, HalfSize - Hi,
7136 // We may not have a base (first source) - this can safely be undefined.
7138 Base = DAG.getUNDEF(VT);
7140 int BitLen = (Len * VT.getScalarSizeInBits()) & 0x3f;
7141 int BitIdx = (Idx * VT.getScalarSizeInBits()) & 0x3f;
7142 return DAG.getNode(X86ISD::INSERTQI, DL, VT, Base, Insert,
7143 DAG.getConstant(BitLen, DL, MVT::i8),
7144 DAG.getConstant(BitIdx, DL, MVT::i8));
7151 if (SDValue InsertQ = LowerAsInsertQ())
7157 /// \brief Lower a vector shuffle as a zero or any extension.
7159 /// Given a specific number of elements, element bit width, and extension
7160 /// stride, produce either a zero or any extension based on the available
7161 /// features of the subtarget.
7162 static SDValue lowerVectorShuffleAsSpecificZeroOrAnyExtend(
7163 SDLoc DL, MVT VT, int Scale, bool AnyExt, SDValue InputV,
7164 ArrayRef<int> Mask, const X86Subtarget *Subtarget, SelectionDAG &DAG) {
7165 assert(Scale > 1 && "Need a scale to extend.");
7166 int NumElements = VT.getVectorNumElements();
7167 int EltBits = VT.getScalarSizeInBits();
7168 assert((EltBits == 8 || EltBits == 16 || EltBits == 32) &&
7169 "Only 8, 16, and 32 bit elements can be extended.");
7170 assert(Scale * EltBits <= 64 && "Cannot zero extend past 64 bits.");
7172 // Found a valid zext mask! Try various lowering strategies based on the
7173 // input type and available ISA extensions.
7174 if (Subtarget->hasSSE41()) {
7175 MVT ExtVT = MVT::getVectorVT(MVT::getIntegerVT(EltBits * Scale),
7176 NumElements / Scale);
7177 return DAG.getBitcast(VT, DAG.getNode(X86ISD::VZEXT, DL, ExtVT, InputV));
7180 // For any extends we can cheat for larger element sizes and use shuffle
7181 // instructions that can fold with a load and/or copy.
7182 if (AnyExt && EltBits == 32) {
7183 int PSHUFDMask[4] = {0, -1, 1, -1};
7184 return DAG.getBitcast(
7185 VT, DAG.getNode(X86ISD::PSHUFD, DL, MVT::v4i32,
7186 DAG.getBitcast(MVT::v4i32, InputV),
7187 getV4X86ShuffleImm8ForMask(PSHUFDMask, DL, DAG)));
7189 if (AnyExt && EltBits == 16 && Scale > 2) {
7190 int PSHUFDMask[4] = {0, -1, 0, -1};
7191 InputV = DAG.getNode(X86ISD::PSHUFD, DL, MVT::v4i32,
7192 DAG.getBitcast(MVT::v4i32, InputV),
7193 getV4X86ShuffleImm8ForMask(PSHUFDMask, DL, DAG));
7194 int PSHUFHWMask[4] = {1, -1, -1, -1};
7195 return DAG.getBitcast(
7196 VT, DAG.getNode(X86ISD::PSHUFHW, DL, MVT::v8i16,
7197 DAG.getBitcast(MVT::v8i16, InputV),
7198 getV4X86ShuffleImm8ForMask(PSHUFHWMask, DL, DAG)));
7201 // The SSE4A EXTRQ instruction can efficiently extend the first 2 lanes
7203 if ((Scale * EltBits) == 64 && EltBits < 32 && Subtarget->hasSSE4A()) {
7204 assert(NumElements == (int)Mask.size() && "Unexpected shuffle mask size!");
7205 assert(VT.getSizeInBits() == 128 && "Unexpected vector width!");
7207 SDValue Lo = DAG.getNode(ISD::BITCAST, DL, MVT::v2i64,
7208 DAG.getNode(X86ISD::EXTRQI, DL, VT, InputV,
7209 DAG.getConstant(EltBits, DL, MVT::i8),
7210 DAG.getConstant(0, DL, MVT::i8)));
7211 if (isUndefInRange(Mask, NumElements/2, NumElements/2))
7212 return DAG.getNode(ISD::BITCAST, DL, VT, Lo);
7215 DAG.getNode(ISD::BITCAST, DL, MVT::v2i64,
7216 DAG.getNode(X86ISD::EXTRQI, DL, VT, InputV,
7217 DAG.getConstant(EltBits, DL, MVT::i8),
7218 DAG.getConstant(EltBits, DL, MVT::i8)));
7219 return DAG.getNode(ISD::BITCAST, DL, VT,
7220 DAG.getNode(X86ISD::UNPCKL, DL, MVT::v2i64, Lo, Hi));
7223 // If this would require more than 2 unpack instructions to expand, use
7224 // pshufb when available. We can only use more than 2 unpack instructions
7225 // when zero extending i8 elements which also makes it easier to use pshufb.
7226 if (Scale > 4 && EltBits == 8 && Subtarget->hasSSSE3()) {
7227 assert(NumElements == 16 && "Unexpected byte vector width!");
7228 SDValue PSHUFBMask[16];
7229 for (int i = 0; i < 16; ++i)
7231 DAG.getConstant((i % Scale == 0) ? i / Scale : 0x80, DL, MVT::i8);
7232 InputV = DAG.getBitcast(MVT::v16i8, InputV);
7233 return DAG.getBitcast(VT,
7234 DAG.getNode(X86ISD::PSHUFB, DL, MVT::v16i8, InputV,
7235 DAG.getNode(ISD::BUILD_VECTOR, DL,
7236 MVT::v16i8, PSHUFBMask)));
7239 // Otherwise emit a sequence of unpacks.
7241 MVT InputVT = MVT::getVectorVT(MVT::getIntegerVT(EltBits), NumElements);
7242 SDValue Ext = AnyExt ? DAG.getUNDEF(InputVT)
7243 : getZeroVector(InputVT, Subtarget, DAG, DL);
7244 InputV = DAG.getBitcast(InputVT, InputV);
7245 InputV = DAG.getNode(X86ISD::UNPCKL, DL, InputVT, InputV, Ext);
7249 } while (Scale > 1);
7250 return DAG.getBitcast(VT, InputV);
7253 /// \brief Try to lower a vector shuffle as a zero extension on any microarch.
7255 /// This routine will try to do everything in its power to cleverly lower
7256 /// a shuffle which happens to match the pattern of a zero extend. It doesn't
7257 /// check for the profitability of this lowering, it tries to aggressively
7258 /// match this pattern. It will use all of the micro-architectural details it
7259 /// can to emit an efficient lowering. It handles both blends with all-zero
7260 /// inputs to explicitly zero-extend and undef-lanes (sometimes undef due to
7261 /// masking out later).
7263 /// The reason we have dedicated lowering for zext-style shuffles is that they
7264 /// are both incredibly common and often quite performance sensitive.
7265 static SDValue lowerVectorShuffleAsZeroOrAnyExtend(
7266 SDLoc DL, MVT VT, SDValue V1, SDValue V2, ArrayRef<int> Mask,
7267 const X86Subtarget *Subtarget, SelectionDAG &DAG) {
7268 SmallBitVector Zeroable = computeZeroableShuffleElements(Mask, V1, V2);
7270 int Bits = VT.getSizeInBits();
7271 int NumElements = VT.getVectorNumElements();
7272 assert(VT.getScalarSizeInBits() <= 32 &&
7273 "Exceeds 32-bit integer zero extension limit");
7274 assert((int)Mask.size() == NumElements && "Unexpected shuffle mask size");
7276 // Define a helper function to check a particular ext-scale and lower to it if
7278 auto Lower = [&](int Scale) -> SDValue {
7281 for (int i = 0; i < NumElements; ++i) {
7283 continue; // Valid anywhere but doesn't tell us anything.
7284 if (i % Scale != 0) {
7285 // Each of the extended elements need to be zeroable.
7289 // We no longer are in the anyext case.
7294 // Each of the base elements needs to be consecutive indices into the
7295 // same input vector.
7296 SDValue V = Mask[i] < NumElements ? V1 : V2;
7299 else if (InputV != V)
7300 return SDValue(); // Flip-flopping inputs.
7302 if (Mask[i] % NumElements != i / Scale)
7303 return SDValue(); // Non-consecutive strided elements.
7306 // If we fail to find an input, we have a zero-shuffle which should always
7307 // have already been handled.
7308 // FIXME: Maybe handle this here in case during blending we end up with one?
7312 return lowerVectorShuffleAsSpecificZeroOrAnyExtend(
7313 DL, VT, Scale, AnyExt, InputV, Mask, Subtarget, DAG);
7316 // The widest scale possible for extending is to a 64-bit integer.
7317 assert(Bits % 64 == 0 &&
7318 "The number of bits in a vector must be divisible by 64 on x86!");
7319 int NumExtElements = Bits / 64;
7321 // Each iteration, try extending the elements half as much, but into twice as
7323 for (; NumExtElements < NumElements; NumExtElements *= 2) {
7324 assert(NumElements % NumExtElements == 0 &&
7325 "The input vector size must be divisible by the extended size.");
7326 if (SDValue V = Lower(NumElements / NumExtElements))
7330 // General extends failed, but 128-bit vectors may be able to use MOVQ.
7334 // Returns one of the source operands if the shuffle can be reduced to a
7335 // MOVQ, copying the lower 64-bits and zero-extending to the upper 64-bits.
7336 auto CanZExtLowHalf = [&]() {
7337 for (int i = NumElements / 2; i != NumElements; ++i)
7340 if (isSequentialOrUndefInRange(Mask, 0, NumElements / 2, 0))
7342 if (isSequentialOrUndefInRange(Mask, 0, NumElements / 2, NumElements))
7347 if (SDValue V = CanZExtLowHalf()) {
7348 V = DAG.getBitcast(MVT::v2i64, V);
7349 V = DAG.getNode(X86ISD::VZEXT_MOVL, DL, MVT::v2i64, V);
7350 return DAG.getBitcast(VT, V);
7353 // No viable ext lowering found.
7357 /// \brief Try to get a scalar value for a specific element of a vector.
7359 /// Looks through BUILD_VECTOR and SCALAR_TO_VECTOR nodes to find a scalar.
7360 static SDValue getScalarValueForVectorElement(SDValue V, int Idx,
7361 SelectionDAG &DAG) {
7362 MVT VT = V.getSimpleValueType();
7363 MVT EltVT = VT.getVectorElementType();
7364 while (V.getOpcode() == ISD::BITCAST)
7365 V = V.getOperand(0);
7366 // If the bitcasts shift the element size, we can't extract an equivalent
7368 MVT NewVT = V.getSimpleValueType();
7369 if (!NewVT.isVector() || NewVT.getScalarSizeInBits() != VT.getScalarSizeInBits())
7372 if (V.getOpcode() == ISD::BUILD_VECTOR ||
7373 (Idx == 0 && V.getOpcode() == ISD::SCALAR_TO_VECTOR)) {
7374 // Ensure the scalar operand is the same size as the destination.
7375 // FIXME: Add support for scalar truncation where possible.
7376 SDValue S = V.getOperand(Idx);
7377 if (EltVT.getSizeInBits() == S.getSimpleValueType().getSizeInBits())
7378 return DAG.getNode(ISD::BITCAST, SDLoc(V), EltVT, S);
7384 /// \brief Helper to test for a load that can be folded with x86 shuffles.
7386 /// This is particularly important because the set of instructions varies
7387 /// significantly based on whether the operand is a load or not.
7388 static bool isShuffleFoldableLoad(SDValue V) {
7389 while (V.getOpcode() == ISD::BITCAST)
7390 V = V.getOperand(0);
7392 return ISD::isNON_EXTLoad(V.getNode());
7395 /// \brief Try to lower insertion of a single element into a zero vector.
7397 /// This is a common pattern that we have especially efficient patterns to lower
7398 /// across all subtarget feature sets.
7399 static SDValue lowerVectorShuffleAsElementInsertion(
7400 SDLoc DL, MVT VT, SDValue V1, SDValue V2, ArrayRef<int> Mask,
7401 const X86Subtarget *Subtarget, SelectionDAG &DAG) {
7402 SmallBitVector Zeroable = computeZeroableShuffleElements(Mask, V1, V2);
7404 MVT EltVT = VT.getVectorElementType();
7406 int V2Index = std::find_if(Mask.begin(), Mask.end(),
7407 [&Mask](int M) { return M >= (int)Mask.size(); }) -
7409 bool IsV1Zeroable = true;
7410 for (int i = 0, Size = Mask.size(); i < Size; ++i)
7411 if (i != V2Index && !Zeroable[i]) {
7412 IsV1Zeroable = false;
7416 // Check for a single input from a SCALAR_TO_VECTOR node.
7417 // FIXME: All of this should be canonicalized into INSERT_VECTOR_ELT and
7418 // all the smarts here sunk into that routine. However, the current
7419 // lowering of BUILD_VECTOR makes that nearly impossible until the old
7420 // vector shuffle lowering is dead.
7421 SDValue V2S = getScalarValueForVectorElement(V2, Mask[V2Index] - Mask.size(),
7423 if (V2S && DAG.getTargetLoweringInfo().isTypeLegal(V2S.getValueType())) {
7424 // We need to zext the scalar if it is smaller than an i32.
7425 V2S = DAG.getBitcast(EltVT, V2S);
7426 if (EltVT == MVT::i8 || EltVT == MVT::i16) {
7427 // Using zext to expand a narrow element won't work for non-zero
7432 // Zero-extend directly to i32.
7434 V2S = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::i32, V2S);
7436 V2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, ExtVT, V2S);
7437 } else if (Mask[V2Index] != (int)Mask.size() || EltVT == MVT::i8 ||
7438 EltVT == MVT::i16) {
7439 // Either not inserting from the low element of the input or the input
7440 // element size is too small to use VZEXT_MOVL to clear the high bits.
7444 if (!IsV1Zeroable) {
7445 // If V1 can't be treated as a zero vector we have fewer options to lower
7446 // this. We can't support integer vectors or non-zero targets cheaply, and
7447 // the V1 elements can't be permuted in any way.
7448 assert(VT == ExtVT && "Cannot change extended type when non-zeroable!");
7449 if (!VT.isFloatingPoint() || V2Index != 0)
7451 SmallVector<int, 8> V1Mask(Mask.begin(), Mask.end());
7452 V1Mask[V2Index] = -1;
7453 if (!isNoopShuffleMask(V1Mask))
7455 // This is essentially a special case blend operation, but if we have
7456 // general purpose blend operations, they are always faster. Bail and let
7457 // the rest of the lowering handle these as blends.
7458 if (Subtarget->hasSSE41())
7461 // Otherwise, use MOVSD or MOVSS.
7462 assert((EltVT == MVT::f32 || EltVT == MVT::f64) &&
7463 "Only two types of floating point element types to handle!");
7464 return DAG.getNode(EltVT == MVT::f32 ? X86ISD::MOVSS : X86ISD::MOVSD, DL,
7468 // This lowering only works for the low element with floating point vectors.
7469 if (VT.isFloatingPoint() && V2Index != 0)
7472 V2 = DAG.getNode(X86ISD::VZEXT_MOVL, DL, ExtVT, V2);
7474 V2 = DAG.getBitcast(VT, V2);
7477 // If we have 4 or fewer lanes we can cheaply shuffle the element into
7478 // the desired position. Otherwise it is more efficient to do a vector
7479 // shift left. We know that we can do a vector shift left because all
7480 // the inputs are zero.
7481 if (VT.isFloatingPoint() || VT.getVectorNumElements() <= 4) {
7482 SmallVector<int, 4> V2Shuffle(Mask.size(), 1);
7483 V2Shuffle[V2Index] = 0;
7484 V2 = DAG.getVectorShuffle(VT, DL, V2, DAG.getUNDEF(VT), V2Shuffle);
7486 V2 = DAG.getBitcast(MVT::v2i64, V2);
7488 X86ISD::VSHLDQ, DL, MVT::v2i64, V2,
7489 DAG.getConstant(V2Index * EltVT.getSizeInBits() / 8, DL,
7490 DAG.getTargetLoweringInfo().getScalarShiftAmountTy(
7491 DAG.getDataLayout(), VT)));
7492 V2 = DAG.getBitcast(VT, V2);
7498 /// \brief Try to lower broadcast of a single element.
7500 /// For convenience, this code also bundles all of the subtarget feature set
7501 /// filtering. While a little annoying to re-dispatch on type here, there isn't
7502 /// a convenient way to factor it out.
7503 static SDValue lowerVectorShuffleAsBroadcast(SDLoc DL, MVT VT, SDValue V,
7505 const X86Subtarget *Subtarget,
7506 SelectionDAG &DAG) {
7507 if (!Subtarget->hasAVX())
7509 if (VT.isInteger() && !Subtarget->hasAVX2())
7512 // Check that the mask is a broadcast.
7513 int BroadcastIdx = -1;
7515 if (M >= 0 && BroadcastIdx == -1)
7517 else if (M >= 0 && M != BroadcastIdx)
7520 assert(BroadcastIdx < (int)Mask.size() && "We only expect to be called with "
7521 "a sorted mask where the broadcast "
7524 // Go up the chain of (vector) values to find a scalar load that we can
7525 // combine with the broadcast.
7527 switch (V.getOpcode()) {
7528 case ISD::CONCAT_VECTORS: {
7529 int OperandSize = Mask.size() / V.getNumOperands();
7530 V = V.getOperand(BroadcastIdx / OperandSize);
7531 BroadcastIdx %= OperandSize;
7535 case ISD::INSERT_SUBVECTOR: {
7536 SDValue VOuter = V.getOperand(0), VInner = V.getOperand(1);
7537 auto ConstantIdx = dyn_cast<ConstantSDNode>(V.getOperand(2));
7541 int BeginIdx = (int)ConstantIdx->getZExtValue();
7543 BeginIdx + (int)VInner.getValueType().getVectorNumElements();
7544 if (BroadcastIdx >= BeginIdx && BroadcastIdx < EndIdx) {
7545 BroadcastIdx -= BeginIdx;
7556 // Check if this is a broadcast of a scalar. We special case lowering
7557 // for scalars so that we can more effectively fold with loads.
7558 // First, look through bitcast: if the original value has a larger element
7559 // type than the shuffle, the broadcast element is in essence truncated.
7560 // Make that explicit to ease folding.
7561 if (V.getOpcode() == ISD::BITCAST && VT.isInteger()) {
7562 EVT EltVT = VT.getVectorElementType();
7563 SDValue V0 = V.getOperand(0);
7564 EVT V0VT = V0.getValueType();
7566 if (V0VT.isInteger() && V0VT.getVectorElementType().bitsGT(EltVT) &&
7567 ((V0.getOpcode() == ISD::BUILD_VECTOR ||
7568 (V0.getOpcode() == ISD::SCALAR_TO_VECTOR && BroadcastIdx == 0)))) {
7569 V = DAG.getNode(ISD::TRUNCATE, DL, EltVT, V0.getOperand(BroadcastIdx));
7574 // Also check the simpler case, where we can directly reuse the scalar.
7575 if (V.getOpcode() == ISD::BUILD_VECTOR ||
7576 (V.getOpcode() == ISD::SCALAR_TO_VECTOR && BroadcastIdx == 0)) {
7577 V = V.getOperand(BroadcastIdx);
7579 // If the scalar isn't a load, we can't broadcast from it in AVX1.
7580 // Only AVX2 has register broadcasts.
7581 if (!Subtarget->hasAVX2() && !isShuffleFoldableLoad(V))
7583 } else if (BroadcastIdx != 0 || !Subtarget->hasAVX2()) {
7584 // We can't broadcast from a vector register without AVX2, and we can only
7585 // broadcast from the zero-element of a vector register.
7589 return DAG.getNode(X86ISD::VBROADCAST, DL, VT, V);
7592 // Check for whether we can use INSERTPS to perform the shuffle. We only use
7593 // INSERTPS when the V1 elements are already in the correct locations
7594 // because otherwise we can just always use two SHUFPS instructions which
7595 // are much smaller to encode than a SHUFPS and an INSERTPS. We can also
7596 // perform INSERTPS if a single V1 element is out of place and all V2
7597 // elements are zeroable.
7598 static SDValue lowerVectorShuffleAsInsertPS(SDValue Op, SDValue V1, SDValue V2,
7600 SelectionDAG &DAG) {
7601 assert(Op.getSimpleValueType() == MVT::v4f32 && "Bad shuffle type!");
7602 assert(V1.getSimpleValueType() == MVT::v4f32 && "Bad operand type!");
7603 assert(V2.getSimpleValueType() == MVT::v4f32 && "Bad operand type!");
7604 assert(Mask.size() == 4 && "Unexpected mask size for v4 shuffle!");
7606 SmallBitVector Zeroable = computeZeroableShuffleElements(Mask, V1, V2);
7609 int V1DstIndex = -1;
7610 int V2DstIndex = -1;
7611 bool V1UsedInPlace = false;
7613 for (int i = 0; i < 4; ++i) {
7614 // Synthesize a zero mask from the zeroable elements (includes undefs).
7620 // Flag if we use any V1 inputs in place.
7622 V1UsedInPlace = true;
7626 // We can only insert a single non-zeroable element.
7627 if (V1DstIndex != -1 || V2DstIndex != -1)
7631 // V1 input out of place for insertion.
7634 // V2 input for insertion.
7639 // Don't bother if we have no (non-zeroable) element for insertion.
7640 if (V1DstIndex == -1 && V2DstIndex == -1)
7643 // Determine element insertion src/dst indices. The src index is from the
7644 // start of the inserted vector, not the start of the concatenated vector.
7645 unsigned V2SrcIndex = 0;
7646 if (V1DstIndex != -1) {
7647 // If we have a V1 input out of place, we use V1 as the V2 element insertion
7648 // and don't use the original V2 at all.
7649 V2SrcIndex = Mask[V1DstIndex];
7650 V2DstIndex = V1DstIndex;
7653 V2SrcIndex = Mask[V2DstIndex] - 4;
7656 // If no V1 inputs are used in place, then the result is created only from
7657 // the zero mask and the V2 insertion - so remove V1 dependency.
7659 V1 = DAG.getUNDEF(MVT::v4f32);
7661 unsigned InsertPSMask = V2SrcIndex << 6 | V2DstIndex << 4 | ZMask;
7662 assert((InsertPSMask & ~0xFFu) == 0 && "Invalid mask!");
7664 // Insert the V2 element into the desired position.
7666 return DAG.getNode(X86ISD::INSERTPS, DL, MVT::v4f32, V1, V2,
7667 DAG.getConstant(InsertPSMask, DL, MVT::i8));
7670 /// \brief Try to lower a shuffle as a permute of the inputs followed by an
7671 /// UNPCK instruction.
7673 /// This specifically targets cases where we end up with alternating between
7674 /// the two inputs, and so can permute them into something that feeds a single
7675 /// UNPCK instruction. Note that this routine only targets integer vectors
7676 /// because for floating point vectors we have a generalized SHUFPS lowering
7677 /// strategy that handles everything that doesn't *exactly* match an unpack,
7678 /// making this clever lowering unnecessary.
7679 static SDValue lowerVectorShuffleAsUnpack(SDLoc DL, MVT VT, SDValue V1,
7680 SDValue V2, ArrayRef<int> Mask,
7681 SelectionDAG &DAG) {
7682 assert(!VT.isFloatingPoint() &&
7683 "This routine only supports integer vectors.");
7684 assert(!isSingleInputShuffleMask(Mask) &&
7685 "This routine should only be used when blending two inputs.");
7686 assert(Mask.size() >= 2 && "Single element masks are invalid.");
7688 int Size = Mask.size();
7690 int NumLoInputs = std::count_if(Mask.begin(), Mask.end(), [Size](int M) {
7691 return M >= 0 && M % Size < Size / 2;
7693 int NumHiInputs = std::count_if(
7694 Mask.begin(), Mask.end(), [Size](int M) { return M % Size >= Size / 2; });
7696 bool UnpackLo = NumLoInputs >= NumHiInputs;
7698 auto TryUnpack = [&](MVT UnpackVT, int Scale) {
7699 SmallVector<int, 32> V1Mask(Mask.size(), -1);
7700 SmallVector<int, 32> V2Mask(Mask.size(), -1);
7702 for (int i = 0; i < Size; ++i) {
7706 // Each element of the unpack contains Scale elements from this mask.
7707 int UnpackIdx = i / Scale;
7709 // We only handle the case where V1 feeds the first slots of the unpack.
7710 // We rely on canonicalization to ensure this is the case.
7711 if ((UnpackIdx % 2 == 0) != (Mask[i] < Size))
7714 // Setup the mask for this input. The indexing is tricky as we have to
7715 // handle the unpack stride.
7716 SmallVectorImpl<int> &VMask = (UnpackIdx % 2 == 0) ? V1Mask : V2Mask;
7717 VMask[(UnpackIdx / 2) * Scale + i % Scale + (UnpackLo ? 0 : Size / 2)] =
7721 // If we will have to shuffle both inputs to use the unpack, check whether
7722 // we can just unpack first and shuffle the result. If so, skip this unpack.
7723 if ((NumLoInputs == 0 || NumHiInputs == 0) && !isNoopShuffleMask(V1Mask) &&
7724 !isNoopShuffleMask(V2Mask))
7727 // Shuffle the inputs into place.
7728 V1 = DAG.getVectorShuffle(VT, DL, V1, DAG.getUNDEF(VT), V1Mask);
7729 V2 = DAG.getVectorShuffle(VT, DL, V2, DAG.getUNDEF(VT), V2Mask);
7731 // Cast the inputs to the type we will use to unpack them.
7732 V1 = DAG.getBitcast(UnpackVT, V1);
7733 V2 = DAG.getBitcast(UnpackVT, V2);
7735 // Unpack the inputs and cast the result back to the desired type.
7736 return DAG.getBitcast(
7737 VT, DAG.getNode(UnpackLo ? X86ISD::UNPCKL : X86ISD::UNPCKH, DL,
7741 // We try each unpack from the largest to the smallest to try and find one
7742 // that fits this mask.
7743 int OrigNumElements = VT.getVectorNumElements();
7744 int OrigScalarSize = VT.getScalarSizeInBits();
7745 for (int ScalarSize = 64; ScalarSize >= OrigScalarSize; ScalarSize /= 2) {
7746 int Scale = ScalarSize / OrigScalarSize;
7747 int NumElements = OrigNumElements / Scale;
7748 MVT UnpackVT = MVT::getVectorVT(MVT::getIntegerVT(ScalarSize), NumElements);
7749 if (SDValue Unpack = TryUnpack(UnpackVT, Scale))
7753 // If none of the unpack-rooted lowerings worked (or were profitable) try an
7755 if (NumLoInputs == 0 || NumHiInputs == 0) {
7756 assert((NumLoInputs > 0 || NumHiInputs > 0) &&
7757 "We have to have *some* inputs!");
7758 int HalfOffset = NumLoInputs == 0 ? Size / 2 : 0;
7760 // FIXME: We could consider the total complexity of the permute of each
7761 // possible unpacking. Or at the least we should consider how many
7762 // half-crossings are created.
7763 // FIXME: We could consider commuting the unpacks.
7765 SmallVector<int, 32> PermMask;
7766 PermMask.assign(Size, -1);
7767 for (int i = 0; i < Size; ++i) {
7771 assert(Mask[i] % Size >= HalfOffset && "Found input from wrong half!");
7774 2 * ((Mask[i] % Size) - HalfOffset) + (Mask[i] < Size ? 0 : 1);
7776 return DAG.getVectorShuffle(
7777 VT, DL, DAG.getNode(NumLoInputs == 0 ? X86ISD::UNPCKH : X86ISD::UNPCKL,
7779 DAG.getUNDEF(VT), PermMask);
7785 /// \brief Handle lowering of 2-lane 64-bit floating point shuffles.
7787 /// This is the basis function for the 2-lane 64-bit shuffles as we have full
7788 /// support for floating point shuffles but not integer shuffles. These
7789 /// instructions will incur a domain crossing penalty on some chips though so
7790 /// it is better to avoid lowering through this for integer vectors where
7792 static SDValue lowerV2F64VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
7793 const X86Subtarget *Subtarget,
7794 SelectionDAG &DAG) {
7796 assert(Op.getSimpleValueType() == MVT::v2f64 && "Bad shuffle type!");
7797 assert(V1.getSimpleValueType() == MVT::v2f64 && "Bad operand type!");
7798 assert(V2.getSimpleValueType() == MVT::v2f64 && "Bad operand type!");
7799 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
7800 ArrayRef<int> Mask = SVOp->getMask();
7801 assert(Mask.size() == 2 && "Unexpected mask size for v2 shuffle!");
7803 if (isSingleInputShuffleMask(Mask)) {
7804 // Use low duplicate instructions for masks that match their pattern.
7805 if (Subtarget->hasSSE3())
7806 if (isShuffleEquivalent(V1, V2, Mask, {0, 0}))
7807 return DAG.getNode(X86ISD::MOVDDUP, DL, MVT::v2f64, V1);
7809 // Straight shuffle of a single input vector. Simulate this by using the
7810 // single input as both of the "inputs" to this instruction..
7811 unsigned SHUFPDMask = (Mask[0] == 1) | ((Mask[1] == 1) << 1);
7813 if (Subtarget->hasAVX()) {
7814 // If we have AVX, we can use VPERMILPS which will allow folding a load
7815 // into the shuffle.
7816 return DAG.getNode(X86ISD::VPERMILPI, DL, MVT::v2f64, V1,
7817 DAG.getConstant(SHUFPDMask, DL, MVT::i8));
7820 return DAG.getNode(X86ISD::SHUFP, DL, MVT::v2f64, V1, V1,
7821 DAG.getConstant(SHUFPDMask, DL, MVT::i8));
7823 assert(Mask[0] >= 0 && Mask[0] < 2 && "Non-canonicalized blend!");
7824 assert(Mask[1] >= 2 && "Non-canonicalized blend!");
7826 // If we have a single input, insert that into V1 if we can do so cheaply.
7827 if ((Mask[0] >= 2) + (Mask[1] >= 2) == 1) {
7828 if (SDValue Insertion = lowerVectorShuffleAsElementInsertion(
7829 DL, MVT::v2f64, V1, V2, Mask, Subtarget, DAG))
7831 // Try inverting the insertion since for v2 masks it is easy to do and we
7832 // can't reliably sort the mask one way or the other.
7833 int InverseMask[2] = {Mask[0] < 0 ? -1 : (Mask[0] ^ 2),
7834 Mask[1] < 0 ? -1 : (Mask[1] ^ 2)};
7835 if (SDValue Insertion = lowerVectorShuffleAsElementInsertion(
7836 DL, MVT::v2f64, V2, V1, InverseMask, Subtarget, DAG))
7840 // Try to use one of the special instruction patterns to handle two common
7841 // blend patterns if a zero-blend above didn't work.
7842 if (isShuffleEquivalent(V1, V2, Mask, {0, 3}) ||
7843 isShuffleEquivalent(V1, V2, Mask, {1, 3}))
7844 if (SDValue V1S = getScalarValueForVectorElement(V1, Mask[0], DAG))
7845 // We can either use a special instruction to load over the low double or
7846 // to move just the low double.
7848 isShuffleFoldableLoad(V1S) ? X86ISD::MOVLPD : X86ISD::MOVSD,
7850 DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, MVT::v2f64, V1S));
7852 if (Subtarget->hasSSE41())
7853 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v2f64, V1, V2, Mask,
7857 // Use dedicated unpack instructions for masks that match their pattern.
7858 if (isShuffleEquivalent(V1, V2, Mask, {0, 2}))
7859 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v2f64, V1, V2);
7860 if (isShuffleEquivalent(V1, V2, Mask, {1, 3}))
7861 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v2f64, V1, V2);
7863 unsigned SHUFPDMask = (Mask[0] == 1) | (((Mask[1] - 2) == 1) << 1);
7864 return DAG.getNode(X86ISD::SHUFP, DL, MVT::v2f64, V1, V2,
7865 DAG.getConstant(SHUFPDMask, DL, MVT::i8));
7868 /// \brief Handle lowering of 2-lane 64-bit integer shuffles.
7870 /// Tries to lower a 2-lane 64-bit shuffle using shuffle operations provided by
7871 /// the integer unit to minimize domain crossing penalties. However, for blends
7872 /// it falls back to the floating point shuffle operation with appropriate bit
7874 static SDValue lowerV2I64VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
7875 const X86Subtarget *Subtarget,
7876 SelectionDAG &DAG) {
7878 assert(Op.getSimpleValueType() == MVT::v2i64 && "Bad shuffle type!");
7879 assert(V1.getSimpleValueType() == MVT::v2i64 && "Bad operand type!");
7880 assert(V2.getSimpleValueType() == MVT::v2i64 && "Bad operand type!");
7881 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
7882 ArrayRef<int> Mask = SVOp->getMask();
7883 assert(Mask.size() == 2 && "Unexpected mask size for v2 shuffle!");
7885 if (isSingleInputShuffleMask(Mask)) {
7886 // Check for being able to broadcast a single element.
7887 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v2i64, V1,
7888 Mask, Subtarget, DAG))
7891 // Straight shuffle of a single input vector. For everything from SSE2
7892 // onward this has a single fast instruction with no scary immediates.
7893 // We have to map the mask as it is actually a v4i32 shuffle instruction.
7894 V1 = DAG.getBitcast(MVT::v4i32, V1);
7895 int WidenedMask[4] = {
7896 std::max(Mask[0], 0) * 2, std::max(Mask[0], 0) * 2 + 1,
7897 std::max(Mask[1], 0) * 2, std::max(Mask[1], 0) * 2 + 1};
7898 return DAG.getBitcast(
7900 DAG.getNode(X86ISD::PSHUFD, DL, MVT::v4i32, V1,
7901 getV4X86ShuffleImm8ForMask(WidenedMask, DL, DAG)));
7903 assert(Mask[0] != -1 && "No undef lanes in multi-input v2 shuffles!");
7904 assert(Mask[1] != -1 && "No undef lanes in multi-input v2 shuffles!");
7905 assert(Mask[0] < 2 && "We sort V1 to be the first input.");
7906 assert(Mask[1] >= 2 && "We sort V2 to be the second input.");
7908 // If we have a blend of two PACKUS operations an the blend aligns with the
7909 // low and half halves, we can just merge the PACKUS operations. This is
7910 // particularly important as it lets us merge shuffles that this routine itself
7912 auto GetPackNode = [](SDValue V) {
7913 while (V.getOpcode() == ISD::BITCAST)
7914 V = V.getOperand(0);
7916 return V.getOpcode() == X86ISD::PACKUS ? V : SDValue();
7918 if (SDValue V1Pack = GetPackNode(V1))
7919 if (SDValue V2Pack = GetPackNode(V2))
7920 return DAG.getBitcast(MVT::v2i64,
7921 DAG.getNode(X86ISD::PACKUS, DL, MVT::v16i8,
7922 Mask[0] == 0 ? V1Pack.getOperand(0)
7923 : V1Pack.getOperand(1),
7924 Mask[1] == 2 ? V2Pack.getOperand(0)
7925 : V2Pack.getOperand(1)));
7927 // Try to use shift instructions.
7929 lowerVectorShuffleAsShift(DL, MVT::v2i64, V1, V2, Mask, DAG))
7932 // When loading a scalar and then shuffling it into a vector we can often do
7933 // the insertion cheaply.
7934 if (SDValue Insertion = lowerVectorShuffleAsElementInsertion(
7935 DL, MVT::v2i64, V1, V2, Mask, Subtarget, DAG))
7937 // Try inverting the insertion since for v2 masks it is easy to do and we
7938 // can't reliably sort the mask one way or the other.
7939 int InverseMask[2] = {Mask[0] ^ 2, Mask[1] ^ 2};
7940 if (SDValue Insertion = lowerVectorShuffleAsElementInsertion(
7941 DL, MVT::v2i64, V2, V1, InverseMask, Subtarget, DAG))
7944 // We have different paths for blend lowering, but they all must use the
7945 // *exact* same predicate.
7946 bool IsBlendSupported = Subtarget->hasSSE41();
7947 if (IsBlendSupported)
7948 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v2i64, V1, V2, Mask,
7952 // Use dedicated unpack instructions for masks that match their pattern.
7953 if (isShuffleEquivalent(V1, V2, Mask, {0, 2}))
7954 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v2i64, V1, V2);
7955 if (isShuffleEquivalent(V1, V2, Mask, {1, 3}))
7956 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v2i64, V1, V2);
7958 // Try to use byte rotation instructions.
7959 // Its more profitable for pre-SSSE3 to use shuffles/unpacks.
7960 if (Subtarget->hasSSSE3())
7961 if (SDValue Rotate = lowerVectorShuffleAsByteRotate(
7962 DL, MVT::v2i64, V1, V2, Mask, Subtarget, DAG))
7965 // If we have direct support for blends, we should lower by decomposing into
7966 // a permute. That will be faster than the domain cross.
7967 if (IsBlendSupported)
7968 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v2i64, V1, V2,
7971 // We implement this with SHUFPD which is pretty lame because it will likely
7972 // incur 2 cycles of stall for integer vectors on Nehalem and older chips.
7973 // However, all the alternatives are still more cycles and newer chips don't
7974 // have this problem. It would be really nice if x86 had better shuffles here.
7975 V1 = DAG.getBitcast(MVT::v2f64, V1);
7976 V2 = DAG.getBitcast(MVT::v2f64, V2);
7977 return DAG.getBitcast(MVT::v2i64,
7978 DAG.getVectorShuffle(MVT::v2f64, DL, V1, V2, Mask));
7981 /// \brief Test whether this can be lowered with a single SHUFPS instruction.
7983 /// This is used to disable more specialized lowerings when the shufps lowering
7984 /// will happen to be efficient.
7985 static bool isSingleSHUFPSMask(ArrayRef<int> Mask) {
7986 // This routine only handles 128-bit shufps.
7987 assert(Mask.size() == 4 && "Unsupported mask size!");
7989 // To lower with a single SHUFPS we need to have the low half and high half
7990 // each requiring a single input.
7991 if (Mask[0] != -1 && Mask[1] != -1 && (Mask[0] < 4) != (Mask[1] < 4))
7993 if (Mask[2] != -1 && Mask[3] != -1 && (Mask[2] < 4) != (Mask[3] < 4))
7999 /// \brief Lower a vector shuffle using the SHUFPS instruction.
8001 /// This is a helper routine dedicated to lowering vector shuffles using SHUFPS.
8002 /// It makes no assumptions about whether this is the *best* lowering, it simply
8004 static SDValue lowerVectorShuffleWithSHUFPS(SDLoc DL, MVT VT,
8005 ArrayRef<int> Mask, SDValue V1,
8006 SDValue V2, SelectionDAG &DAG) {
8007 SDValue LowV = V1, HighV = V2;
8008 int NewMask[4] = {Mask[0], Mask[1], Mask[2], Mask[3]};
8011 std::count_if(Mask.begin(), Mask.end(), [](int M) { return M >= 4; });
8013 if (NumV2Elements == 1) {
8015 std::find_if(Mask.begin(), Mask.end(), [](int M) { return M >= 4; }) -
8018 // Compute the index adjacent to V2Index and in the same half by toggling
8020 int V2AdjIndex = V2Index ^ 1;
8022 if (Mask[V2AdjIndex] == -1) {
8023 // Handles all the cases where we have a single V2 element and an undef.
8024 // This will only ever happen in the high lanes because we commute the
8025 // vector otherwise.
8027 std::swap(LowV, HighV);
8028 NewMask[V2Index] -= 4;
8030 // Handle the case where the V2 element ends up adjacent to a V1 element.
8031 // To make this work, blend them together as the first step.
8032 int V1Index = V2AdjIndex;
8033 int BlendMask[4] = {Mask[V2Index] - 4, 0, Mask[V1Index], 0};
8034 V2 = DAG.getNode(X86ISD::SHUFP, DL, VT, V2, V1,
8035 getV4X86ShuffleImm8ForMask(BlendMask, DL, DAG));
8037 // Now proceed to reconstruct the final blend as we have the necessary
8038 // high or low half formed.
8045 NewMask[V1Index] = 2; // We put the V1 element in V2[2].
8046 NewMask[V2Index] = 0; // We shifted the V2 element into V2[0].
8048 } else if (NumV2Elements == 2) {
8049 if (Mask[0] < 4 && Mask[1] < 4) {
8050 // Handle the easy case where we have V1 in the low lanes and V2 in the
8054 } else if (Mask[2] < 4 && Mask[3] < 4) {
8055 // We also handle the reversed case because this utility may get called
8056 // when we detect a SHUFPS pattern but can't easily commute the shuffle to
8057 // arrange things in the right direction.
8063 // We have a mixture of V1 and V2 in both low and high lanes. Rather than
8064 // trying to place elements directly, just blend them and set up the final
8065 // shuffle to place them.
8067 // The first two blend mask elements are for V1, the second two are for
8069 int BlendMask[4] = {Mask[0] < 4 ? Mask[0] : Mask[1],
8070 Mask[2] < 4 ? Mask[2] : Mask[3],
8071 (Mask[0] >= 4 ? Mask[0] : Mask[1]) - 4,
8072 (Mask[2] >= 4 ? Mask[2] : Mask[3]) - 4};
8073 V1 = DAG.getNode(X86ISD::SHUFP, DL, VT, V1, V2,
8074 getV4X86ShuffleImm8ForMask(BlendMask, DL, DAG));
8076 // Now we do a normal shuffle of V1 by giving V1 as both operands to
8079 NewMask[0] = Mask[0] < 4 ? 0 : 2;
8080 NewMask[1] = Mask[0] < 4 ? 2 : 0;
8081 NewMask[2] = Mask[2] < 4 ? 1 : 3;
8082 NewMask[3] = Mask[2] < 4 ? 3 : 1;
8085 return DAG.getNode(X86ISD::SHUFP, DL, VT, LowV, HighV,
8086 getV4X86ShuffleImm8ForMask(NewMask, DL, DAG));
8089 /// \brief Lower 4-lane 32-bit floating point shuffles.
8091 /// Uses instructions exclusively from the floating point unit to minimize
8092 /// domain crossing penalties, as these are sufficient to implement all v4f32
8094 static SDValue lowerV4F32VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
8095 const X86Subtarget *Subtarget,
8096 SelectionDAG &DAG) {
8098 assert(Op.getSimpleValueType() == MVT::v4f32 && "Bad shuffle type!");
8099 assert(V1.getSimpleValueType() == MVT::v4f32 && "Bad operand type!");
8100 assert(V2.getSimpleValueType() == MVT::v4f32 && "Bad operand type!");
8101 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
8102 ArrayRef<int> Mask = SVOp->getMask();
8103 assert(Mask.size() == 4 && "Unexpected mask size for v4 shuffle!");
8106 std::count_if(Mask.begin(), Mask.end(), [](int M) { return M >= 4; });
8108 if (NumV2Elements == 0) {
8109 // Check for being able to broadcast a single element.
8110 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v4f32, V1,
8111 Mask, Subtarget, DAG))
8114 // Use even/odd duplicate instructions for masks that match their pattern.
8115 if (Subtarget->hasSSE3()) {
8116 if (isShuffleEquivalent(V1, V2, Mask, {0, 0, 2, 2}))
8117 return DAG.getNode(X86ISD::MOVSLDUP, DL, MVT::v4f32, V1);
8118 if (isShuffleEquivalent(V1, V2, Mask, {1, 1, 3, 3}))
8119 return DAG.getNode(X86ISD::MOVSHDUP, DL, MVT::v4f32, V1);
8122 if (Subtarget->hasAVX()) {
8123 // If we have AVX, we can use VPERMILPS which will allow folding a load
8124 // into the shuffle.
8125 return DAG.getNode(X86ISD::VPERMILPI, DL, MVT::v4f32, V1,
8126 getV4X86ShuffleImm8ForMask(Mask, DL, DAG));
8129 // Otherwise, use a straight shuffle of a single input vector. We pass the
8130 // input vector to both operands to simulate this with a SHUFPS.
8131 return DAG.getNode(X86ISD::SHUFP, DL, MVT::v4f32, V1, V1,
8132 getV4X86ShuffleImm8ForMask(Mask, DL, DAG));
8135 // There are special ways we can lower some single-element blends. However, we
8136 // have custom ways we can lower more complex single-element blends below that
8137 // we defer to if both this and BLENDPS fail to match, so restrict this to
8138 // when the V2 input is targeting element 0 of the mask -- that is the fast
8140 if (NumV2Elements == 1 && Mask[0] >= 4)
8141 if (SDValue V = lowerVectorShuffleAsElementInsertion(DL, MVT::v4f32, V1, V2,
8142 Mask, Subtarget, DAG))
8145 if (Subtarget->hasSSE41()) {
8146 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v4f32, V1, V2, Mask,
8150 // Use INSERTPS if we can complete the shuffle efficiently.
8151 if (SDValue V = lowerVectorShuffleAsInsertPS(Op, V1, V2, Mask, DAG))
8154 if (!isSingleSHUFPSMask(Mask))
8155 if (SDValue BlendPerm = lowerVectorShuffleAsBlendAndPermute(
8156 DL, MVT::v4f32, V1, V2, Mask, DAG))
8160 // Use dedicated unpack instructions for masks that match their pattern.
8161 if (isShuffleEquivalent(V1, V2, Mask, {0, 4, 1, 5}))
8162 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v4f32, V1, V2);
8163 if (isShuffleEquivalent(V1, V2, Mask, {2, 6, 3, 7}))
8164 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v4f32, V1, V2);
8165 if (isShuffleEquivalent(V1, V2, Mask, {4, 0, 5, 1}))
8166 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v4f32, V2, V1);
8167 if (isShuffleEquivalent(V1, V2, Mask, {6, 2, 7, 3}))
8168 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v4f32, V2, V1);
8170 // Otherwise fall back to a SHUFPS lowering strategy.
8171 return lowerVectorShuffleWithSHUFPS(DL, MVT::v4f32, Mask, V1, V2, DAG);
8174 /// \brief Lower 4-lane i32 vector shuffles.
8176 /// We try to handle these with integer-domain shuffles where we can, but for
8177 /// blends we use the floating point domain blend instructions.
8178 static SDValue lowerV4I32VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
8179 const X86Subtarget *Subtarget,
8180 SelectionDAG &DAG) {
8182 assert(Op.getSimpleValueType() == MVT::v4i32 && "Bad shuffle type!");
8183 assert(V1.getSimpleValueType() == MVT::v4i32 && "Bad operand type!");
8184 assert(V2.getSimpleValueType() == MVT::v4i32 && "Bad operand type!");
8185 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
8186 ArrayRef<int> Mask = SVOp->getMask();
8187 assert(Mask.size() == 4 && "Unexpected mask size for v4 shuffle!");
8189 // Whenever we can lower this as a zext, that instruction is strictly faster
8190 // than any alternative. It also allows us to fold memory operands into the
8191 // shuffle in many cases.
8192 if (SDValue ZExt = lowerVectorShuffleAsZeroOrAnyExtend(DL, MVT::v4i32, V1, V2,
8193 Mask, Subtarget, DAG))
8197 std::count_if(Mask.begin(), Mask.end(), [](int M) { return M >= 4; });
8199 if (NumV2Elements == 0) {
8200 // Check for being able to broadcast a single element.
8201 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v4i32, V1,
8202 Mask, Subtarget, DAG))
8205 // Straight shuffle of a single input vector. For everything from SSE2
8206 // onward this has a single fast instruction with no scary immediates.
8207 // We coerce the shuffle pattern to be compatible with UNPCK instructions
8208 // but we aren't actually going to use the UNPCK instruction because doing
8209 // so prevents folding a load into this instruction or making a copy.
8210 const int UnpackLoMask[] = {0, 0, 1, 1};
8211 const int UnpackHiMask[] = {2, 2, 3, 3};
8212 if (isShuffleEquivalent(V1, V2, Mask, {0, 0, 1, 1}))
8213 Mask = UnpackLoMask;
8214 else if (isShuffleEquivalent(V1, V2, Mask, {2, 2, 3, 3}))
8215 Mask = UnpackHiMask;
8217 return DAG.getNode(X86ISD::PSHUFD, DL, MVT::v4i32, V1,
8218 getV4X86ShuffleImm8ForMask(Mask, DL, DAG));
8221 // Try to use shift instructions.
8223 lowerVectorShuffleAsShift(DL, MVT::v4i32, V1, V2, Mask, DAG))
8226 // There are special ways we can lower some single-element blends.
8227 if (NumV2Elements == 1)
8228 if (SDValue V = lowerVectorShuffleAsElementInsertion(DL, MVT::v4i32, V1, V2,
8229 Mask, Subtarget, DAG))
8232 // We have different paths for blend lowering, but they all must use the
8233 // *exact* same predicate.
8234 bool IsBlendSupported = Subtarget->hasSSE41();
8235 if (IsBlendSupported)
8236 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v4i32, V1, V2, Mask,
8240 if (SDValue Masked =
8241 lowerVectorShuffleAsBitMask(DL, MVT::v4i32, V1, V2, Mask, DAG))
8244 // Use dedicated unpack instructions for masks that match their pattern.
8245 if (isShuffleEquivalent(V1, V2, Mask, {0, 4, 1, 5}))
8246 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v4i32, V1, V2);
8247 if (isShuffleEquivalent(V1, V2, Mask, {2, 6, 3, 7}))
8248 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v4i32, V1, V2);
8249 if (isShuffleEquivalent(V1, V2, Mask, {4, 0, 5, 1}))
8250 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v4i32, V2, V1);
8251 if (isShuffleEquivalent(V1, V2, Mask, {6, 2, 7, 3}))
8252 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v4i32, V2, V1);
8254 // Try to use byte rotation instructions.
8255 // Its more profitable for pre-SSSE3 to use shuffles/unpacks.
8256 if (Subtarget->hasSSSE3())
8257 if (SDValue Rotate = lowerVectorShuffleAsByteRotate(
8258 DL, MVT::v4i32, V1, V2, Mask, Subtarget, DAG))
8261 // If we have direct support for blends, we should lower by decomposing into
8262 // a permute. That will be faster than the domain cross.
8263 if (IsBlendSupported)
8264 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v4i32, V1, V2,
8267 // Try to lower by permuting the inputs into an unpack instruction.
8268 if (SDValue Unpack =
8269 lowerVectorShuffleAsUnpack(DL, MVT::v4i32, V1, V2, Mask, DAG))
8272 // We implement this with SHUFPS because it can blend from two vectors.
8273 // Because we're going to eventually use SHUFPS, we use SHUFPS even to build
8274 // up the inputs, bypassing domain shift penalties that we would encur if we
8275 // directly used PSHUFD on Nehalem and older. For newer chips, this isn't
8277 return DAG.getBitcast(
8279 DAG.getVectorShuffle(MVT::v4f32, DL, DAG.getBitcast(MVT::v4f32, V1),
8280 DAG.getBitcast(MVT::v4f32, V2), Mask));
8283 /// \brief Lowering of single-input v8i16 shuffles is the cornerstone of SSE2
8284 /// shuffle lowering, and the most complex part.
8286 /// The lowering strategy is to try to form pairs of input lanes which are
8287 /// targeted at the same half of the final vector, and then use a dword shuffle
8288 /// to place them onto the right half, and finally unpack the paired lanes into
8289 /// their final position.
8291 /// The exact breakdown of how to form these dword pairs and align them on the
8292 /// correct sides is really tricky. See the comments within the function for
8293 /// more of the details.
8295 /// This code also handles repeated 128-bit lanes of v8i16 shuffles, but each
8296 /// lane must shuffle the *exact* same way. In fact, you must pass a v8 Mask to
8297 /// this routine for it to work correctly. To shuffle a 256-bit or 512-bit i16
8298 /// vector, form the analogous 128-bit 8-element Mask.
8299 static SDValue lowerV8I16GeneralSingleInputVectorShuffle(
8300 SDLoc DL, MVT VT, SDValue V, MutableArrayRef<int> Mask,
8301 const X86Subtarget *Subtarget, SelectionDAG &DAG) {
8302 assert(VT.getScalarType() == MVT::i16 && "Bad input type!");
8303 MVT PSHUFDVT = MVT::getVectorVT(MVT::i32, VT.getVectorNumElements() / 2);
8305 assert(Mask.size() == 8 && "Shuffle mask length doen't match!");
8306 MutableArrayRef<int> LoMask = Mask.slice(0, 4);
8307 MutableArrayRef<int> HiMask = Mask.slice(4, 4);
8309 SmallVector<int, 4> LoInputs;
8310 std::copy_if(LoMask.begin(), LoMask.end(), std::back_inserter(LoInputs),
8311 [](int M) { return M >= 0; });
8312 std::sort(LoInputs.begin(), LoInputs.end());
8313 LoInputs.erase(std::unique(LoInputs.begin(), LoInputs.end()), LoInputs.end());
8314 SmallVector<int, 4> HiInputs;
8315 std::copy_if(HiMask.begin(), HiMask.end(), std::back_inserter(HiInputs),
8316 [](int M) { return M >= 0; });
8317 std::sort(HiInputs.begin(), HiInputs.end());
8318 HiInputs.erase(std::unique(HiInputs.begin(), HiInputs.end()), HiInputs.end());
8320 std::lower_bound(LoInputs.begin(), LoInputs.end(), 4) - LoInputs.begin();
8321 int NumHToL = LoInputs.size() - NumLToL;
8323 std::lower_bound(HiInputs.begin(), HiInputs.end(), 4) - HiInputs.begin();
8324 int NumHToH = HiInputs.size() - NumLToH;
8325 MutableArrayRef<int> LToLInputs(LoInputs.data(), NumLToL);
8326 MutableArrayRef<int> LToHInputs(HiInputs.data(), NumLToH);
8327 MutableArrayRef<int> HToLInputs(LoInputs.data() + NumLToL, NumHToL);
8328 MutableArrayRef<int> HToHInputs(HiInputs.data() + NumLToH, NumHToH);
8330 // Simplify the 1-into-3 and 3-into-1 cases with a single pshufd. For all
8331 // such inputs we can swap two of the dwords across the half mark and end up
8332 // with <=2 inputs to each half in each half. Once there, we can fall through
8333 // to the generic code below. For example:
8335 // Input: [a, b, c, d, e, f, g, h] -PSHUFD[0,2,1,3]-> [a, b, e, f, c, d, g, h]
8336 // Mask: [0, 1, 2, 7, 4, 5, 6, 3] -----------------> [0, 1, 4, 7, 2, 3, 6, 5]
8338 // However in some very rare cases we have a 1-into-3 or 3-into-1 on one half
8339 // and an existing 2-into-2 on the other half. In this case we may have to
8340 // pre-shuffle the 2-into-2 half to avoid turning it into a 3-into-1 or
8341 // 1-into-3 which could cause us to cycle endlessly fixing each side in turn.
8342 // Fortunately, we don't have to handle anything but a 2-into-2 pattern
8343 // because any other situation (including a 3-into-1 or 1-into-3 in the other
8344 // half than the one we target for fixing) will be fixed when we re-enter this
8345 // path. We will also combine away any sequence of PSHUFD instructions that
8346 // result into a single instruction. Here is an example of the tricky case:
8348 // Input: [a, b, c, d, e, f, g, h] -PSHUFD[0,2,1,3]-> [a, b, e, f, c, d, g, h]
8349 // Mask: [3, 7, 1, 0, 2, 7, 3, 5] -THIS-IS-BAD!!!!-> [5, 7, 1, 0, 4, 7, 5, 3]
8351 // This now has a 1-into-3 in the high half! Instead, we do two shuffles:
8353 // Input: [a, b, c, d, e, f, g, h] PSHUFHW[0,2,1,3]-> [a, b, c, d, e, g, f, h]
8354 // Mask: [3, 7, 1, 0, 2, 7, 3, 5] -----------------> [3, 7, 1, 0, 2, 7, 3, 6]
8356 // Input: [a, b, c, d, e, g, f, h] -PSHUFD[0,2,1,3]-> [a, b, e, g, c, d, f, h]
8357 // Mask: [3, 7, 1, 0, 2, 7, 3, 6] -----------------> [5, 7, 1, 0, 4, 7, 5, 6]
8359 // The result is fine to be handled by the generic logic.
8360 auto balanceSides = [&](ArrayRef<int> AToAInputs, ArrayRef<int> BToAInputs,
8361 ArrayRef<int> BToBInputs, ArrayRef<int> AToBInputs,
8362 int AOffset, int BOffset) {
8363 assert((AToAInputs.size() == 3 || AToAInputs.size() == 1) &&
8364 "Must call this with A having 3 or 1 inputs from the A half.");
8365 assert((BToAInputs.size() == 1 || BToAInputs.size() == 3) &&
8366 "Must call this with B having 1 or 3 inputs from the B half.");
8367 assert(AToAInputs.size() + BToAInputs.size() == 4 &&
8368 "Must call this with either 3:1 or 1:3 inputs (summing to 4).");
8370 bool ThreeAInputs = AToAInputs.size() == 3;
8372 // Compute the index of dword with only one word among the three inputs in
8373 // a half by taking the sum of the half with three inputs and subtracting
8374 // the sum of the actual three inputs. The difference is the remaining
8377 int &TripleDWord = ThreeAInputs ? ADWord : BDWord;
8378 int &OneInputDWord = ThreeAInputs ? BDWord : ADWord;
8379 int TripleInputOffset = ThreeAInputs ? AOffset : BOffset;
8380 ArrayRef<int> TripleInputs = ThreeAInputs ? AToAInputs : BToAInputs;
8381 int OneInput = ThreeAInputs ? BToAInputs[0] : AToAInputs[0];
8382 int TripleInputSum = 0 + 1 + 2 + 3 + (4 * TripleInputOffset);
8383 int TripleNonInputIdx =
8384 TripleInputSum - std::accumulate(TripleInputs.begin(), TripleInputs.end(), 0);
8385 TripleDWord = TripleNonInputIdx / 2;
8387 // We use xor with one to compute the adjacent DWord to whichever one the
8389 OneInputDWord = (OneInput / 2) ^ 1;
8391 // Check for one tricky case: We're fixing a 3<-1 or a 1<-3 shuffle for AToA
8392 // and BToA inputs. If there is also such a problem with the BToB and AToB
8393 // inputs, we don't try to fix it necessarily -- we'll recurse and see it in
8394 // the next pass. However, if we have a 2<-2 in the BToB and AToB inputs, it
8395 // is essential that we don't *create* a 3<-1 as then we might oscillate.
8396 if (BToBInputs.size() == 2 && AToBInputs.size() == 2) {
8397 // Compute how many inputs will be flipped by swapping these DWords. We
8399 // to balance this to ensure we don't form a 3-1 shuffle in the other
8401 int NumFlippedAToBInputs =
8402 std::count(AToBInputs.begin(), AToBInputs.end(), 2 * ADWord) +
8403 std::count(AToBInputs.begin(), AToBInputs.end(), 2 * ADWord + 1);
8404 int NumFlippedBToBInputs =
8405 std::count(BToBInputs.begin(), BToBInputs.end(), 2 * BDWord) +
8406 std::count(BToBInputs.begin(), BToBInputs.end(), 2 * BDWord + 1);
8407 if ((NumFlippedAToBInputs == 1 &&
8408 (NumFlippedBToBInputs == 0 || NumFlippedBToBInputs == 2)) ||
8409 (NumFlippedBToBInputs == 1 &&
8410 (NumFlippedAToBInputs == 0 || NumFlippedAToBInputs == 2))) {
8411 // We choose whether to fix the A half or B half based on whether that
8412 // half has zero flipped inputs. At zero, we may not be able to fix it
8413 // with that half. We also bias towards fixing the B half because that
8414 // will more commonly be the high half, and we have to bias one way.
8415 auto FixFlippedInputs = [&V, &DL, &Mask, &DAG](int PinnedIdx, int DWord,
8416 ArrayRef<int> Inputs) {
8417 int FixIdx = PinnedIdx ^ 1; // The adjacent slot to the pinned slot.
8418 bool IsFixIdxInput = std::find(Inputs.begin(), Inputs.end(),
8419 PinnedIdx ^ 1) != Inputs.end();
8420 // Determine whether the free index is in the flipped dword or the
8421 // unflipped dword based on where the pinned index is. We use this bit
8422 // in an xor to conditionally select the adjacent dword.
8423 int FixFreeIdx = 2 * (DWord ^ (PinnedIdx / 2 == DWord));
8424 bool IsFixFreeIdxInput = std::find(Inputs.begin(), Inputs.end(),
8425 FixFreeIdx) != Inputs.end();
8426 if (IsFixIdxInput == IsFixFreeIdxInput)
8428 IsFixFreeIdxInput = std::find(Inputs.begin(), Inputs.end(),
8429 FixFreeIdx) != Inputs.end();
8430 assert(IsFixIdxInput != IsFixFreeIdxInput &&
8431 "We need to be changing the number of flipped inputs!");
8432 int PSHUFHalfMask[] = {0, 1, 2, 3};
8433 std::swap(PSHUFHalfMask[FixFreeIdx % 4], PSHUFHalfMask[FixIdx % 4]);
8434 V = DAG.getNode(FixIdx < 4 ? X86ISD::PSHUFLW : X86ISD::PSHUFHW, DL,
8436 getV4X86ShuffleImm8ForMask(PSHUFHalfMask, DL, DAG));
8439 if (M != -1 && M == FixIdx)
8441 else if (M != -1 && M == FixFreeIdx)
8444 if (NumFlippedBToBInputs != 0) {
8446 BToAInputs.size() == 3 ? TripleNonInputIdx : OneInput;
8447 FixFlippedInputs(BPinnedIdx, BDWord, BToBInputs);
8449 assert(NumFlippedAToBInputs != 0 && "Impossible given predicates!");
8450 int APinnedIdx = ThreeAInputs ? TripleNonInputIdx : OneInput;
8451 FixFlippedInputs(APinnedIdx, ADWord, AToBInputs);
8456 int PSHUFDMask[] = {0, 1, 2, 3};
8457 PSHUFDMask[ADWord] = BDWord;
8458 PSHUFDMask[BDWord] = ADWord;
8461 DAG.getNode(X86ISD::PSHUFD, DL, PSHUFDVT, DAG.getBitcast(PSHUFDVT, V),
8462 getV4X86ShuffleImm8ForMask(PSHUFDMask, DL, DAG)));
8464 // Adjust the mask to match the new locations of A and B.
8466 if (M != -1 && M/2 == ADWord)
8467 M = 2 * BDWord + M % 2;
8468 else if (M != -1 && M/2 == BDWord)
8469 M = 2 * ADWord + M % 2;
8471 // Recurse back into this routine to re-compute state now that this isn't
8472 // a 3 and 1 problem.
8473 return lowerV8I16GeneralSingleInputVectorShuffle(DL, VT, V, Mask, Subtarget,
8476 if ((NumLToL == 3 && NumHToL == 1) || (NumLToL == 1 && NumHToL == 3))
8477 return balanceSides(LToLInputs, HToLInputs, HToHInputs, LToHInputs, 0, 4);
8478 else if ((NumHToH == 3 && NumLToH == 1) || (NumHToH == 1 && NumLToH == 3))
8479 return balanceSides(HToHInputs, LToHInputs, LToLInputs, HToLInputs, 4, 0);
8481 // At this point there are at most two inputs to the low and high halves from
8482 // each half. That means the inputs can always be grouped into dwords and
8483 // those dwords can then be moved to the correct half with a dword shuffle.
8484 // We use at most one low and one high word shuffle to collect these paired
8485 // inputs into dwords, and finally a dword shuffle to place them.
8486 int PSHUFLMask[4] = {-1, -1, -1, -1};
8487 int PSHUFHMask[4] = {-1, -1, -1, -1};
8488 int PSHUFDMask[4] = {-1, -1, -1, -1};
8490 // First fix the masks for all the inputs that are staying in their
8491 // original halves. This will then dictate the targets of the cross-half
8493 auto fixInPlaceInputs =
8494 [&PSHUFDMask](ArrayRef<int> InPlaceInputs, ArrayRef<int> IncomingInputs,
8495 MutableArrayRef<int> SourceHalfMask,
8496 MutableArrayRef<int> HalfMask, int HalfOffset) {
8497 if (InPlaceInputs.empty())
8499 if (InPlaceInputs.size() == 1) {
8500 SourceHalfMask[InPlaceInputs[0] - HalfOffset] =
8501 InPlaceInputs[0] - HalfOffset;
8502 PSHUFDMask[InPlaceInputs[0] / 2] = InPlaceInputs[0] / 2;
8505 if (IncomingInputs.empty()) {
8506 // Just fix all of the in place inputs.
8507 for (int Input : InPlaceInputs) {
8508 SourceHalfMask[Input - HalfOffset] = Input - HalfOffset;
8509 PSHUFDMask[Input / 2] = Input / 2;
8514 assert(InPlaceInputs.size() == 2 && "Cannot handle 3 or 4 inputs!");
8515 SourceHalfMask[InPlaceInputs[0] - HalfOffset] =
8516 InPlaceInputs[0] - HalfOffset;
8517 // Put the second input next to the first so that they are packed into
8518 // a dword. We find the adjacent index by toggling the low bit.
8519 int AdjIndex = InPlaceInputs[0] ^ 1;
8520 SourceHalfMask[AdjIndex - HalfOffset] = InPlaceInputs[1] - HalfOffset;
8521 std::replace(HalfMask.begin(), HalfMask.end(), InPlaceInputs[1], AdjIndex);
8522 PSHUFDMask[AdjIndex / 2] = AdjIndex / 2;
8524 fixInPlaceInputs(LToLInputs, HToLInputs, PSHUFLMask, LoMask, 0);
8525 fixInPlaceInputs(HToHInputs, LToHInputs, PSHUFHMask, HiMask, 4);
8527 // Now gather the cross-half inputs and place them into a free dword of
8528 // their target half.
8529 // FIXME: This operation could almost certainly be simplified dramatically to
8530 // look more like the 3-1 fixing operation.
8531 auto moveInputsToRightHalf = [&PSHUFDMask](
8532 MutableArrayRef<int> IncomingInputs, ArrayRef<int> ExistingInputs,
8533 MutableArrayRef<int> SourceHalfMask, MutableArrayRef<int> HalfMask,
8534 MutableArrayRef<int> FinalSourceHalfMask, int SourceOffset,
8536 auto isWordClobbered = [](ArrayRef<int> SourceHalfMask, int Word) {
8537 return SourceHalfMask[Word] != -1 && SourceHalfMask[Word] != Word;
8539 auto isDWordClobbered = [&isWordClobbered](ArrayRef<int> SourceHalfMask,
8541 int LowWord = Word & ~1;
8542 int HighWord = Word | 1;
8543 return isWordClobbered(SourceHalfMask, LowWord) ||
8544 isWordClobbered(SourceHalfMask, HighWord);
8547 if (IncomingInputs.empty())
8550 if (ExistingInputs.empty()) {
8551 // Map any dwords with inputs from them into the right half.
8552 for (int Input : IncomingInputs) {
8553 // If the source half mask maps over the inputs, turn those into
8554 // swaps and use the swapped lane.
8555 if (isWordClobbered(SourceHalfMask, Input - SourceOffset)) {
8556 if (SourceHalfMask[SourceHalfMask[Input - SourceOffset]] == -1) {
8557 SourceHalfMask[SourceHalfMask[Input - SourceOffset]] =
8558 Input - SourceOffset;
8559 // We have to swap the uses in our half mask in one sweep.
8560 for (int &M : HalfMask)
8561 if (M == SourceHalfMask[Input - SourceOffset] + SourceOffset)
8563 else if (M == Input)
8564 M = SourceHalfMask[Input - SourceOffset] + SourceOffset;
8566 assert(SourceHalfMask[SourceHalfMask[Input - SourceOffset]] ==
8567 Input - SourceOffset &&
8568 "Previous placement doesn't match!");
8570 // Note that this correctly re-maps both when we do a swap and when
8571 // we observe the other side of the swap above. We rely on that to
8572 // avoid swapping the members of the input list directly.
8573 Input = SourceHalfMask[Input - SourceOffset] + SourceOffset;
8576 // Map the input's dword into the correct half.
8577 if (PSHUFDMask[(Input - SourceOffset + DestOffset) / 2] == -1)
8578 PSHUFDMask[(Input - SourceOffset + DestOffset) / 2] = Input / 2;
8580 assert(PSHUFDMask[(Input - SourceOffset + DestOffset) / 2] ==
8582 "Previous placement doesn't match!");
8585 // And just directly shift any other-half mask elements to be same-half
8586 // as we will have mirrored the dword containing the element into the
8587 // same position within that half.
8588 for (int &M : HalfMask)
8589 if (M >= SourceOffset && M < SourceOffset + 4) {
8590 M = M - SourceOffset + DestOffset;
8591 assert(M >= 0 && "This should never wrap below zero!");
8596 // Ensure we have the input in a viable dword of its current half. This
8597 // is particularly tricky because the original position may be clobbered
8598 // by inputs being moved and *staying* in that half.
8599 if (IncomingInputs.size() == 1) {
8600 if (isWordClobbered(SourceHalfMask, IncomingInputs[0] - SourceOffset)) {
8601 int InputFixed = std::find(std::begin(SourceHalfMask),
8602 std::end(SourceHalfMask), -1) -
8603 std::begin(SourceHalfMask) + SourceOffset;
8604 SourceHalfMask[InputFixed - SourceOffset] =
8605 IncomingInputs[0] - SourceOffset;
8606 std::replace(HalfMask.begin(), HalfMask.end(), IncomingInputs[0],
8608 IncomingInputs[0] = InputFixed;
8610 } else if (IncomingInputs.size() == 2) {
8611 if (IncomingInputs[0] / 2 != IncomingInputs[1] / 2 ||
8612 isDWordClobbered(SourceHalfMask, IncomingInputs[0] - SourceOffset)) {
8613 // We have two non-adjacent or clobbered inputs we need to extract from
8614 // the source half. To do this, we need to map them into some adjacent
8615 // dword slot in the source mask.
8616 int InputsFixed[2] = {IncomingInputs[0] - SourceOffset,
8617 IncomingInputs[1] - SourceOffset};
8619 // If there is a free slot in the source half mask adjacent to one of
8620 // the inputs, place the other input in it. We use (Index XOR 1) to
8621 // compute an adjacent index.
8622 if (!isWordClobbered(SourceHalfMask, InputsFixed[0]) &&
8623 SourceHalfMask[InputsFixed[0] ^ 1] == -1) {
8624 SourceHalfMask[InputsFixed[0]] = InputsFixed[0];
8625 SourceHalfMask[InputsFixed[0] ^ 1] = InputsFixed[1];
8626 InputsFixed[1] = InputsFixed[0] ^ 1;
8627 } else if (!isWordClobbered(SourceHalfMask, InputsFixed[1]) &&
8628 SourceHalfMask[InputsFixed[1] ^ 1] == -1) {
8629 SourceHalfMask[InputsFixed[1]] = InputsFixed[1];
8630 SourceHalfMask[InputsFixed[1] ^ 1] = InputsFixed[0];
8631 InputsFixed[0] = InputsFixed[1] ^ 1;
8632 } else if (SourceHalfMask[2 * ((InputsFixed[0] / 2) ^ 1)] == -1 &&
8633 SourceHalfMask[2 * ((InputsFixed[0] / 2) ^ 1) + 1] == -1) {
8634 // The two inputs are in the same DWord but it is clobbered and the
8635 // adjacent DWord isn't used at all. Move both inputs to the free
8637 SourceHalfMask[2 * ((InputsFixed[0] / 2) ^ 1)] = InputsFixed[0];
8638 SourceHalfMask[2 * ((InputsFixed[0] / 2) ^ 1) + 1] = InputsFixed[1];
8639 InputsFixed[0] = 2 * ((InputsFixed[0] / 2) ^ 1);
8640 InputsFixed[1] = 2 * ((InputsFixed[0] / 2) ^ 1) + 1;
8642 // The only way we hit this point is if there is no clobbering
8643 // (because there are no off-half inputs to this half) and there is no
8644 // free slot adjacent to one of the inputs. In this case, we have to
8645 // swap an input with a non-input.
8646 for (int i = 0; i < 4; ++i)
8647 assert((SourceHalfMask[i] == -1 || SourceHalfMask[i] == i) &&
8648 "We can't handle any clobbers here!");
8649 assert(InputsFixed[1] != (InputsFixed[0] ^ 1) &&
8650 "Cannot have adjacent inputs here!");
8652 SourceHalfMask[InputsFixed[0] ^ 1] = InputsFixed[1];
8653 SourceHalfMask[InputsFixed[1]] = InputsFixed[0] ^ 1;
8655 // We also have to update the final source mask in this case because
8656 // it may need to undo the above swap.
8657 for (int &M : FinalSourceHalfMask)
8658 if (M == (InputsFixed[0] ^ 1) + SourceOffset)
8659 M = InputsFixed[1] + SourceOffset;
8660 else if (M == InputsFixed[1] + SourceOffset)
8661 M = (InputsFixed[0] ^ 1) + SourceOffset;
8663 InputsFixed[1] = InputsFixed[0] ^ 1;
8666 // Point everything at the fixed inputs.
8667 for (int &M : HalfMask)
8668 if (M == IncomingInputs[0])
8669 M = InputsFixed[0] + SourceOffset;
8670 else if (M == IncomingInputs[1])
8671 M = InputsFixed[1] + SourceOffset;
8673 IncomingInputs[0] = InputsFixed[0] + SourceOffset;
8674 IncomingInputs[1] = InputsFixed[1] + SourceOffset;
8677 llvm_unreachable("Unhandled input size!");
8680 // Now hoist the DWord down to the right half.
8681 int FreeDWord = (PSHUFDMask[DestOffset / 2] == -1 ? 0 : 1) + DestOffset / 2;
8682 assert(PSHUFDMask[FreeDWord] == -1 && "DWord not free");
8683 PSHUFDMask[FreeDWord] = IncomingInputs[0] / 2;
8684 for (int &M : HalfMask)
8685 for (int Input : IncomingInputs)
8687 M = FreeDWord * 2 + Input % 2;
8689 moveInputsToRightHalf(HToLInputs, LToLInputs, PSHUFHMask, LoMask, HiMask,
8690 /*SourceOffset*/ 4, /*DestOffset*/ 0);
8691 moveInputsToRightHalf(LToHInputs, HToHInputs, PSHUFLMask, HiMask, LoMask,
8692 /*SourceOffset*/ 0, /*DestOffset*/ 4);
8694 // Now enact all the shuffles we've computed to move the inputs into their
8696 if (!isNoopShuffleMask(PSHUFLMask))
8697 V = DAG.getNode(X86ISD::PSHUFLW, DL, VT, V,
8698 getV4X86ShuffleImm8ForMask(PSHUFLMask, DL, DAG));
8699 if (!isNoopShuffleMask(PSHUFHMask))
8700 V = DAG.getNode(X86ISD::PSHUFHW, DL, VT, V,
8701 getV4X86ShuffleImm8ForMask(PSHUFHMask, DL, DAG));
8702 if (!isNoopShuffleMask(PSHUFDMask))
8705 DAG.getNode(X86ISD::PSHUFD, DL, PSHUFDVT, DAG.getBitcast(PSHUFDVT, V),
8706 getV4X86ShuffleImm8ForMask(PSHUFDMask, DL, DAG)));
8708 // At this point, each half should contain all its inputs, and we can then
8709 // just shuffle them into their final position.
8710 assert(std::count_if(LoMask.begin(), LoMask.end(),
8711 [](int M) { return M >= 4; }) == 0 &&
8712 "Failed to lift all the high half inputs to the low mask!");
8713 assert(std::count_if(HiMask.begin(), HiMask.end(),
8714 [](int M) { return M >= 0 && M < 4; }) == 0 &&
8715 "Failed to lift all the low half inputs to the high mask!");
8717 // Do a half shuffle for the low mask.
8718 if (!isNoopShuffleMask(LoMask))
8719 V = DAG.getNode(X86ISD::PSHUFLW, DL, VT, V,
8720 getV4X86ShuffleImm8ForMask(LoMask, DL, DAG));
8722 // Do a half shuffle with the high mask after shifting its values down.
8723 for (int &M : HiMask)
8726 if (!isNoopShuffleMask(HiMask))
8727 V = DAG.getNode(X86ISD::PSHUFHW, DL, VT, V,
8728 getV4X86ShuffleImm8ForMask(HiMask, DL, DAG));
8733 /// \brief Helper to form a PSHUFB-based shuffle+blend.
8734 static SDValue lowerVectorShuffleAsPSHUFB(SDLoc DL, MVT VT, SDValue V1,
8735 SDValue V2, ArrayRef<int> Mask,
8736 SelectionDAG &DAG, bool &V1InUse,
8738 SmallBitVector Zeroable = computeZeroableShuffleElements(Mask, V1, V2);
8744 int Size = Mask.size();
8745 int Scale = 16 / Size;
8746 for (int i = 0; i < 16; ++i) {
8747 if (Mask[i / Scale] == -1) {
8748 V1Mask[i] = V2Mask[i] = DAG.getUNDEF(MVT::i8);
8750 const int ZeroMask = 0x80;
8751 int V1Idx = Mask[i / Scale] < Size ? Mask[i / Scale] * Scale + i % Scale
8753 int V2Idx = Mask[i / Scale] < Size
8755 : (Mask[i / Scale] - Size) * Scale + i % Scale;
8756 if (Zeroable[i / Scale])
8757 V1Idx = V2Idx = ZeroMask;
8758 V1Mask[i] = DAG.getConstant(V1Idx, DL, MVT::i8);
8759 V2Mask[i] = DAG.getConstant(V2Idx, DL, MVT::i8);
8760 V1InUse |= (ZeroMask != V1Idx);
8761 V2InUse |= (ZeroMask != V2Idx);
8766 V1 = DAG.getNode(X86ISD::PSHUFB, DL, MVT::v16i8,
8767 DAG.getBitcast(MVT::v16i8, V1),
8768 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v16i8, V1Mask));
8770 V2 = DAG.getNode(X86ISD::PSHUFB, DL, MVT::v16i8,
8771 DAG.getBitcast(MVT::v16i8, V2),
8772 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v16i8, V2Mask));
8774 // If we need shuffled inputs from both, blend the two.
8776 if (V1InUse && V2InUse)
8777 V = DAG.getNode(ISD::OR, DL, MVT::v16i8, V1, V2);
8779 V = V1InUse ? V1 : V2;
8781 // Cast the result back to the correct type.
8782 return DAG.getBitcast(VT, V);
8785 /// \brief Generic lowering of 8-lane i16 shuffles.
8787 /// This handles both single-input shuffles and combined shuffle/blends with
8788 /// two inputs. The single input shuffles are immediately delegated to
8789 /// a dedicated lowering routine.
8791 /// The blends are lowered in one of three fundamental ways. If there are few
8792 /// enough inputs, it delegates to a basic UNPCK-based strategy. If the shuffle
8793 /// of the input is significantly cheaper when lowered as an interleaving of
8794 /// the two inputs, try to interleave them. Otherwise, blend the low and high
8795 /// halves of the inputs separately (making them have relatively few inputs)
8796 /// and then concatenate them.
8797 static SDValue lowerV8I16VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
8798 const X86Subtarget *Subtarget,
8799 SelectionDAG &DAG) {
8801 assert(Op.getSimpleValueType() == MVT::v8i16 && "Bad shuffle type!");
8802 assert(V1.getSimpleValueType() == MVT::v8i16 && "Bad operand type!");
8803 assert(V2.getSimpleValueType() == MVT::v8i16 && "Bad operand type!");
8804 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
8805 ArrayRef<int> OrigMask = SVOp->getMask();
8806 int MaskStorage[8] = {OrigMask[0], OrigMask[1], OrigMask[2], OrigMask[3],
8807 OrigMask[4], OrigMask[5], OrigMask[6], OrigMask[7]};
8808 MutableArrayRef<int> Mask(MaskStorage);
8810 assert(Mask.size() == 8 && "Unexpected mask size for v8 shuffle!");
8812 // Whenever we can lower this as a zext, that instruction is strictly faster
8813 // than any alternative.
8814 if (SDValue ZExt = lowerVectorShuffleAsZeroOrAnyExtend(
8815 DL, MVT::v8i16, V1, V2, OrigMask, Subtarget, DAG))
8818 auto isV1 = [](int M) { return M >= 0 && M < 8; };
8820 auto isV2 = [](int M) { return M >= 8; };
8822 int NumV2Inputs = std::count_if(Mask.begin(), Mask.end(), isV2);
8824 if (NumV2Inputs == 0) {
8825 // Check for being able to broadcast a single element.
8826 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v8i16, V1,
8827 Mask, Subtarget, DAG))
8830 // Try to use shift instructions.
8832 lowerVectorShuffleAsShift(DL, MVT::v8i16, V1, V1, Mask, DAG))
8835 // Use dedicated unpack instructions for masks that match their pattern.
8836 if (isShuffleEquivalent(V1, V1, Mask, {0, 0, 1, 1, 2, 2, 3, 3}))
8837 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v8i16, V1, V1);
8838 if (isShuffleEquivalent(V1, V1, Mask, {4, 4, 5, 5, 6, 6, 7, 7}))
8839 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v8i16, V1, V1);
8841 // Try to use byte rotation instructions.
8842 if (SDValue Rotate = lowerVectorShuffleAsByteRotate(DL, MVT::v8i16, V1, V1,
8843 Mask, Subtarget, DAG))
8846 return lowerV8I16GeneralSingleInputVectorShuffle(DL, MVT::v8i16, V1, Mask,
8850 assert(std::any_of(Mask.begin(), Mask.end(), isV1) &&
8851 "All single-input shuffles should be canonicalized to be V1-input "
8854 // Try to use shift instructions.
8856 lowerVectorShuffleAsShift(DL, MVT::v8i16, V1, V2, Mask, DAG))
8859 // See if we can use SSE4A Extraction / Insertion.
8860 if (Subtarget->hasSSE4A())
8861 if (SDValue V = lowerVectorShuffleWithSSE4A(DL, MVT::v8i16, V1, V2, Mask, DAG))
8864 // There are special ways we can lower some single-element blends.
8865 if (NumV2Inputs == 1)
8866 if (SDValue V = lowerVectorShuffleAsElementInsertion(DL, MVT::v8i16, V1, V2,
8867 Mask, Subtarget, DAG))
8870 // We have different paths for blend lowering, but they all must use the
8871 // *exact* same predicate.
8872 bool IsBlendSupported = Subtarget->hasSSE41();
8873 if (IsBlendSupported)
8874 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v8i16, V1, V2, Mask,
8878 if (SDValue Masked =
8879 lowerVectorShuffleAsBitMask(DL, MVT::v8i16, V1, V2, Mask, DAG))
8882 // Use dedicated unpack instructions for masks that match their pattern.
8883 if (isShuffleEquivalent(V1, V2, Mask, {0, 8, 1, 9, 2, 10, 3, 11}))
8884 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v8i16, V1, V2);
8885 if (isShuffleEquivalent(V1, V2, Mask, {4, 12, 5, 13, 6, 14, 7, 15}))
8886 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v8i16, V1, V2);
8888 // Try to use byte rotation instructions.
8889 if (SDValue Rotate = lowerVectorShuffleAsByteRotate(
8890 DL, MVT::v8i16, V1, V2, Mask, Subtarget, DAG))
8893 if (SDValue BitBlend =
8894 lowerVectorShuffleAsBitBlend(DL, MVT::v8i16, V1, V2, Mask, DAG))
8897 if (SDValue Unpack =
8898 lowerVectorShuffleAsUnpack(DL, MVT::v8i16, V1, V2, Mask, DAG))
8901 // If we can't directly blend but can use PSHUFB, that will be better as it
8902 // can both shuffle and set up the inefficient blend.
8903 if (!IsBlendSupported && Subtarget->hasSSSE3()) {
8904 bool V1InUse, V2InUse;
8905 return lowerVectorShuffleAsPSHUFB(DL, MVT::v8i16, V1, V2, Mask, DAG,
8909 // We can always bit-blend if we have to so the fallback strategy is to
8910 // decompose into single-input permutes and blends.
8911 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v8i16, V1, V2,
8915 /// \brief Check whether a compaction lowering can be done by dropping even
8916 /// elements and compute how many times even elements must be dropped.
8918 /// This handles shuffles which take every Nth element where N is a power of
8919 /// two. Example shuffle masks:
8921 /// N = 1: 0, 2, 4, 6, 8, 10, 12, 14, 0, 2, 4, 6, 8, 10, 12, 14
8922 /// N = 1: 0, 2, 4, 6, 8, 10, 12, 14, 16, 18, 20, 22, 24, 26, 28, 30
8923 /// N = 2: 0, 4, 8, 12, 0, 4, 8, 12, 0, 4, 8, 12, 0, 4, 8, 12
8924 /// N = 2: 0, 4, 8, 12, 16, 20, 24, 28, 0, 4, 8, 12, 16, 20, 24, 28
8925 /// N = 3: 0, 8, 0, 8, 0, 8, 0, 8, 0, 8, 0, 8, 0, 8, 0, 8
8926 /// N = 3: 0, 8, 16, 24, 0, 8, 16, 24, 0, 8, 16, 24, 0, 8, 16, 24
8928 /// Any of these lanes can of course be undef.
8930 /// This routine only supports N <= 3.
8931 /// FIXME: Evaluate whether either AVX or AVX-512 have any opportunities here
8934 /// \returns N above, or the number of times even elements must be dropped if
8935 /// there is such a number. Otherwise returns zero.
8936 static int canLowerByDroppingEvenElements(ArrayRef<int> Mask) {
8937 // Figure out whether we're looping over two inputs or just one.
8938 bool IsSingleInput = isSingleInputShuffleMask(Mask);
8940 // The modulus for the shuffle vector entries is based on whether this is
8941 // a single input or not.
8942 int ShuffleModulus = Mask.size() * (IsSingleInput ? 1 : 2);
8943 assert(isPowerOf2_32((uint32_t)ShuffleModulus) &&
8944 "We should only be called with masks with a power-of-2 size!");
8946 uint64_t ModMask = (uint64_t)ShuffleModulus - 1;
8948 // We track whether the input is viable for all power-of-2 strides 2^1, 2^2,
8949 // and 2^3 simultaneously. This is because we may have ambiguity with
8950 // partially undef inputs.
8951 bool ViableForN[3] = {true, true, true};
8953 for (int i = 0, e = Mask.size(); i < e; ++i) {
8954 // Ignore undef lanes, we'll optimistically collapse them to the pattern we
8959 bool IsAnyViable = false;
8960 for (unsigned j = 0; j != array_lengthof(ViableForN); ++j)
8961 if (ViableForN[j]) {
8964 // The shuffle mask must be equal to (i * 2^N) % M.
8965 if ((uint64_t)Mask[i] == (((uint64_t)i << N) & ModMask))
8968 ViableForN[j] = false;
8970 // Early exit if we exhaust the possible powers of two.
8975 for (unsigned j = 0; j != array_lengthof(ViableForN); ++j)
8979 // Return 0 as there is no viable power of two.
8983 /// \brief Generic lowering of v16i8 shuffles.
8985 /// This is a hybrid strategy to lower v16i8 vectors. It first attempts to
8986 /// detect any complexity reducing interleaving. If that doesn't help, it uses
8987 /// UNPCK to spread the i8 elements across two i16-element vectors, and uses
8988 /// the existing lowering for v8i16 blends on each half, finally PACK-ing them
8990 static SDValue lowerV16I8VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
8991 const X86Subtarget *Subtarget,
8992 SelectionDAG &DAG) {
8994 assert(Op.getSimpleValueType() == MVT::v16i8 && "Bad shuffle type!");
8995 assert(V1.getSimpleValueType() == MVT::v16i8 && "Bad operand type!");
8996 assert(V2.getSimpleValueType() == MVT::v16i8 && "Bad operand type!");
8997 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
8998 ArrayRef<int> Mask = SVOp->getMask();
8999 assert(Mask.size() == 16 && "Unexpected mask size for v16 shuffle!");
9001 // Try to use shift instructions.
9003 lowerVectorShuffleAsShift(DL, MVT::v16i8, V1, V2, Mask, DAG))
9006 // Try to use byte rotation instructions.
9007 if (SDValue Rotate = lowerVectorShuffleAsByteRotate(
9008 DL, MVT::v16i8, V1, V2, Mask, Subtarget, DAG))
9011 // Try to use a zext lowering.
9012 if (SDValue ZExt = lowerVectorShuffleAsZeroOrAnyExtend(
9013 DL, MVT::v16i8, V1, V2, Mask, Subtarget, DAG))
9016 // See if we can use SSE4A Extraction / Insertion.
9017 if (Subtarget->hasSSE4A())
9018 if (SDValue V = lowerVectorShuffleWithSSE4A(DL, MVT::v16i8, V1, V2, Mask, DAG))
9022 std::count_if(Mask.begin(), Mask.end(), [](int M) { return M >= 16; });
9024 // For single-input shuffles, there are some nicer lowering tricks we can use.
9025 if (NumV2Elements == 0) {
9026 // Check for being able to broadcast a single element.
9027 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v16i8, V1,
9028 Mask, Subtarget, DAG))
9031 // Check whether we can widen this to an i16 shuffle by duplicating bytes.
9032 // Notably, this handles splat and partial-splat shuffles more efficiently.
9033 // However, it only makes sense if the pre-duplication shuffle simplifies
9034 // things significantly. Currently, this means we need to be able to
9035 // express the pre-duplication shuffle as an i16 shuffle.
9037 // FIXME: We should check for other patterns which can be widened into an
9038 // i16 shuffle as well.
9039 auto canWidenViaDuplication = [](ArrayRef<int> Mask) {
9040 for (int i = 0; i < 16; i += 2)
9041 if (Mask[i] != -1 && Mask[i + 1] != -1 && Mask[i] != Mask[i + 1])
9046 auto tryToWidenViaDuplication = [&]() -> SDValue {
9047 if (!canWidenViaDuplication(Mask))
9049 SmallVector<int, 4> LoInputs;
9050 std::copy_if(Mask.begin(), Mask.end(), std::back_inserter(LoInputs),
9051 [](int M) { return M >= 0 && M < 8; });
9052 std::sort(LoInputs.begin(), LoInputs.end());
9053 LoInputs.erase(std::unique(LoInputs.begin(), LoInputs.end()),
9055 SmallVector<int, 4> HiInputs;
9056 std::copy_if(Mask.begin(), Mask.end(), std::back_inserter(HiInputs),
9057 [](int M) { return M >= 8; });
9058 std::sort(HiInputs.begin(), HiInputs.end());
9059 HiInputs.erase(std::unique(HiInputs.begin(), HiInputs.end()),
9062 bool TargetLo = LoInputs.size() >= HiInputs.size();
9063 ArrayRef<int> InPlaceInputs = TargetLo ? LoInputs : HiInputs;
9064 ArrayRef<int> MovingInputs = TargetLo ? HiInputs : LoInputs;
9066 int PreDupI16Shuffle[] = {-1, -1, -1, -1, -1, -1, -1, -1};
9067 SmallDenseMap<int, int, 8> LaneMap;
9068 for (int I : InPlaceInputs) {
9069 PreDupI16Shuffle[I/2] = I/2;
9072 int j = TargetLo ? 0 : 4, je = j + 4;
9073 for (int i = 0, ie = MovingInputs.size(); i < ie; ++i) {
9074 // Check if j is already a shuffle of this input. This happens when
9075 // there are two adjacent bytes after we move the low one.
9076 if (PreDupI16Shuffle[j] != MovingInputs[i] / 2) {
9077 // If we haven't yet mapped the input, search for a slot into which
9079 while (j < je && PreDupI16Shuffle[j] != -1)
9083 // We can't place the inputs into a single half with a simple i16 shuffle, so bail.
9086 // Map this input with the i16 shuffle.
9087 PreDupI16Shuffle[j] = MovingInputs[i] / 2;
9090 // Update the lane map based on the mapping we ended up with.
9091 LaneMap[MovingInputs[i]] = 2 * j + MovingInputs[i] % 2;
9093 V1 = DAG.getBitcast(
9095 DAG.getVectorShuffle(MVT::v8i16, DL, DAG.getBitcast(MVT::v8i16, V1),
9096 DAG.getUNDEF(MVT::v8i16), PreDupI16Shuffle));
9098 // Unpack the bytes to form the i16s that will be shuffled into place.
9099 V1 = DAG.getNode(TargetLo ? X86ISD::UNPCKL : X86ISD::UNPCKH, DL,
9100 MVT::v16i8, V1, V1);
9102 int PostDupI16Shuffle[8] = {-1, -1, -1, -1, -1, -1, -1, -1};
9103 for (int i = 0; i < 16; ++i)
9104 if (Mask[i] != -1) {
9105 int MappedMask = LaneMap[Mask[i]] - (TargetLo ? 0 : 8);
9106 assert(MappedMask < 8 && "Invalid v8 shuffle mask!");
9107 if (PostDupI16Shuffle[i / 2] == -1)
9108 PostDupI16Shuffle[i / 2] = MappedMask;
9110 assert(PostDupI16Shuffle[i / 2] == MappedMask &&
9111 "Conflicting entrties in the original shuffle!");
9113 return DAG.getBitcast(
9115 DAG.getVectorShuffle(MVT::v8i16, DL, DAG.getBitcast(MVT::v8i16, V1),
9116 DAG.getUNDEF(MVT::v8i16), PostDupI16Shuffle));
9118 if (SDValue V = tryToWidenViaDuplication())
9122 if (SDValue Masked =
9123 lowerVectorShuffleAsBitMask(DL, MVT::v16i8, V1, V2, Mask, DAG))
9126 // Use dedicated unpack instructions for masks that match their pattern.
9127 if (isShuffleEquivalent(V1, V2, Mask, {// Low half.
9128 0, 16, 1, 17, 2, 18, 3, 19,
9130 4, 20, 5, 21, 6, 22, 7, 23}))
9131 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v16i8, V1, V2);
9132 if (isShuffleEquivalent(V1, V2, Mask, {// Low half.
9133 8, 24, 9, 25, 10, 26, 11, 27,
9135 12, 28, 13, 29, 14, 30, 15, 31}))
9136 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v16i8, V1, V2);
9138 // Check for SSSE3 which lets us lower all v16i8 shuffles much more directly
9139 // with PSHUFB. It is important to do this before we attempt to generate any
9140 // blends but after all of the single-input lowerings. If the single input
9141 // lowerings can find an instruction sequence that is faster than a PSHUFB, we
9142 // want to preserve that and we can DAG combine any longer sequences into
9143 // a PSHUFB in the end. But once we start blending from multiple inputs,
9144 // the complexity of DAG combining bad patterns back into PSHUFB is too high,
9145 // and there are *very* few patterns that would actually be faster than the
9146 // PSHUFB approach because of its ability to zero lanes.
9148 // FIXME: The only exceptions to the above are blends which are exact
9149 // interleavings with direct instructions supporting them. We currently don't
9150 // handle those well here.
9151 if (Subtarget->hasSSSE3()) {
9152 bool V1InUse = false;
9153 bool V2InUse = false;
9155 SDValue PSHUFB = lowerVectorShuffleAsPSHUFB(DL, MVT::v16i8, V1, V2, Mask,
9156 DAG, V1InUse, V2InUse);
9158 // If both V1 and V2 are in use and we can use a direct blend or an unpack,
9159 // do so. This avoids using them to handle blends-with-zero which is
9160 // important as a single pshufb is significantly faster for that.
9161 if (V1InUse && V2InUse) {
9162 if (Subtarget->hasSSE41())
9163 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v16i8, V1, V2,
9164 Mask, Subtarget, DAG))
9167 // We can use an unpack to do the blending rather than an or in some
9168 // cases. Even though the or may be (very minorly) more efficient, we
9169 // preference this lowering because there are common cases where part of
9170 // the complexity of the shuffles goes away when we do the final blend as
9172 // FIXME: It might be worth trying to detect if the unpack-feeding
9173 // shuffles will both be pshufb, in which case we shouldn't bother with
9175 if (SDValue Unpack =
9176 lowerVectorShuffleAsUnpack(DL, MVT::v16i8, V1, V2, Mask, DAG))
9183 // There are special ways we can lower some single-element blends.
9184 if (NumV2Elements == 1)
9185 if (SDValue V = lowerVectorShuffleAsElementInsertion(DL, MVT::v16i8, V1, V2,
9186 Mask, Subtarget, DAG))
9189 if (SDValue BitBlend =
9190 lowerVectorShuffleAsBitBlend(DL, MVT::v16i8, V1, V2, Mask, DAG))
9193 // Check whether a compaction lowering can be done. This handles shuffles
9194 // which take every Nth element for some even N. See the helper function for
9197 // We special case these as they can be particularly efficiently handled with
9198 // the PACKUSB instruction on x86 and they show up in common patterns of
9199 // rearranging bytes to truncate wide elements.
9200 if (int NumEvenDrops = canLowerByDroppingEvenElements(Mask)) {
9201 // NumEvenDrops is the power of two stride of the elements. Another way of
9202 // thinking about it is that we need to drop the even elements this many
9203 // times to get the original input.
9204 bool IsSingleInput = isSingleInputShuffleMask(Mask);
9206 // First we need to zero all the dropped bytes.
9207 assert(NumEvenDrops <= 3 &&
9208 "No support for dropping even elements more than 3 times.");
9209 // We use the mask type to pick which bytes are preserved based on how many
9210 // elements are dropped.
9211 MVT MaskVTs[] = { MVT::v8i16, MVT::v4i32, MVT::v2i64 };
9212 SDValue ByteClearMask = DAG.getBitcast(
9213 MVT::v16i8, DAG.getConstant(0xFF, DL, MaskVTs[NumEvenDrops - 1]));
9214 V1 = DAG.getNode(ISD::AND, DL, MVT::v16i8, V1, ByteClearMask);
9216 V2 = DAG.getNode(ISD::AND, DL, MVT::v16i8, V2, ByteClearMask);
9218 // Now pack things back together.
9219 V1 = DAG.getBitcast(MVT::v8i16, V1);
9220 V2 = IsSingleInput ? V1 : DAG.getBitcast(MVT::v8i16, V2);
9221 SDValue Result = DAG.getNode(X86ISD::PACKUS, DL, MVT::v16i8, V1, V2);
9222 for (int i = 1; i < NumEvenDrops; ++i) {
9223 Result = DAG.getBitcast(MVT::v8i16, Result);
9224 Result = DAG.getNode(X86ISD::PACKUS, DL, MVT::v16i8, Result, Result);
9230 // Handle multi-input cases by blending single-input shuffles.
9231 if (NumV2Elements > 0)
9232 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v16i8, V1, V2,
9235 // The fallback path for single-input shuffles widens this into two v8i16
9236 // vectors with unpacks, shuffles those, and then pulls them back together
9240 int LoBlendMask[8] = {-1, -1, -1, -1, -1, -1, -1, -1};
9241 int HiBlendMask[8] = {-1, -1, -1, -1, -1, -1, -1, -1};
9242 for (int i = 0; i < 16; ++i)
9244 (i < 8 ? LoBlendMask[i] : HiBlendMask[i % 8]) = Mask[i];
9246 SDValue Zero = getZeroVector(MVT::v8i16, Subtarget, DAG, DL);
9248 SDValue VLoHalf, VHiHalf;
9249 // Check if any of the odd lanes in the v16i8 are used. If not, we can mask
9250 // them out and avoid using UNPCK{L,H} to extract the elements of V as
9252 if (std::none_of(std::begin(LoBlendMask), std::end(LoBlendMask),
9253 [](int M) { return M >= 0 && M % 2 == 1; }) &&
9254 std::none_of(std::begin(HiBlendMask), std::end(HiBlendMask),
9255 [](int M) { return M >= 0 && M % 2 == 1; })) {
9256 // Use a mask to drop the high bytes.
9257 VLoHalf = DAG.getBitcast(MVT::v8i16, V);
9258 VLoHalf = DAG.getNode(ISD::AND, DL, MVT::v8i16, VLoHalf,
9259 DAG.getConstant(0x00FF, DL, MVT::v8i16));
9261 // This will be a single vector shuffle instead of a blend so nuke VHiHalf.
9262 VHiHalf = DAG.getUNDEF(MVT::v8i16);
9264 // Squash the masks to point directly into VLoHalf.
9265 for (int &M : LoBlendMask)
9268 for (int &M : HiBlendMask)
9272 // Otherwise just unpack the low half of V into VLoHalf and the high half into
9273 // VHiHalf so that we can blend them as i16s.
9274 VLoHalf = DAG.getBitcast(
9275 MVT::v8i16, DAG.getNode(X86ISD::UNPCKL, DL, MVT::v16i8, V, Zero));
9276 VHiHalf = DAG.getBitcast(
9277 MVT::v8i16, DAG.getNode(X86ISD::UNPCKH, DL, MVT::v16i8, V, Zero));
9280 SDValue LoV = DAG.getVectorShuffle(MVT::v8i16, DL, VLoHalf, VHiHalf, LoBlendMask);
9281 SDValue HiV = DAG.getVectorShuffle(MVT::v8i16, DL, VLoHalf, VHiHalf, HiBlendMask);
9283 return DAG.getNode(X86ISD::PACKUS, DL, MVT::v16i8, LoV, HiV);
9286 /// \brief Dispatching routine to lower various 128-bit x86 vector shuffles.
9288 /// This routine breaks down the specific type of 128-bit shuffle and
9289 /// dispatches to the lowering routines accordingly.
9290 static SDValue lower128BitVectorShuffle(SDValue Op, SDValue V1, SDValue V2,
9291 MVT VT, const X86Subtarget *Subtarget,
9292 SelectionDAG &DAG) {
9293 switch (VT.SimpleTy) {
9295 return lowerV2I64VectorShuffle(Op, V1, V2, Subtarget, DAG);
9297 return lowerV2F64VectorShuffle(Op, V1, V2, Subtarget, DAG);
9299 return lowerV4I32VectorShuffle(Op, V1, V2, Subtarget, DAG);
9301 return lowerV4F32VectorShuffle(Op, V1, V2, Subtarget, DAG);
9303 return lowerV8I16VectorShuffle(Op, V1, V2, Subtarget, DAG);
9305 return lowerV16I8VectorShuffle(Op, V1, V2, Subtarget, DAG);
9308 llvm_unreachable("Unimplemented!");
9312 /// \brief Helper function to test whether a shuffle mask could be
9313 /// simplified by widening the elements being shuffled.
9315 /// Appends the mask for wider elements in WidenedMask if valid. Otherwise
9316 /// leaves it in an unspecified state.
9318 /// NOTE: This must handle normal vector shuffle masks and *target* vector
9319 /// shuffle masks. The latter have the special property of a '-2' representing
9320 /// a zero-ed lane of a vector.
9321 static bool canWidenShuffleElements(ArrayRef<int> Mask,
9322 SmallVectorImpl<int> &WidenedMask) {
9323 for (int i = 0, Size = Mask.size(); i < Size; i += 2) {
9324 // If both elements are undef, its trivial.
9325 if (Mask[i] == SM_SentinelUndef && Mask[i + 1] == SM_SentinelUndef) {
9326 WidenedMask.push_back(SM_SentinelUndef);
9330 // Check for an undef mask and a mask value properly aligned to fit with
9331 // a pair of values. If we find such a case, use the non-undef mask's value.
9332 if (Mask[i] == SM_SentinelUndef && Mask[i + 1] >= 0 && Mask[i + 1] % 2 == 1) {
9333 WidenedMask.push_back(Mask[i + 1] / 2);
9336 if (Mask[i + 1] == SM_SentinelUndef && Mask[i] >= 0 && Mask[i] % 2 == 0) {
9337 WidenedMask.push_back(Mask[i] / 2);
9341 // When zeroing, we need to spread the zeroing across both lanes to widen.
9342 if (Mask[i] == SM_SentinelZero || Mask[i + 1] == SM_SentinelZero) {
9343 if ((Mask[i] == SM_SentinelZero || Mask[i] == SM_SentinelUndef) &&
9344 (Mask[i + 1] == SM_SentinelZero || Mask[i + 1] == SM_SentinelUndef)) {
9345 WidenedMask.push_back(SM_SentinelZero);
9351 // Finally check if the two mask values are adjacent and aligned with
9353 if (Mask[i] != SM_SentinelUndef && Mask[i] % 2 == 0 && Mask[i] + 1 == Mask[i + 1]) {
9354 WidenedMask.push_back(Mask[i] / 2);
9358 // Otherwise we can't safely widen the elements used in this shuffle.
9361 assert(WidenedMask.size() == Mask.size() / 2 &&
9362 "Incorrect size of mask after widening the elements!");
9367 /// \brief Generic routine to split vector shuffle into half-sized shuffles.
9369 /// This routine just extracts two subvectors, shuffles them independently, and
9370 /// then concatenates them back together. This should work effectively with all
9371 /// AVX vector shuffle types.
9372 static SDValue splitAndLowerVectorShuffle(SDLoc DL, MVT VT, SDValue V1,
9373 SDValue V2, ArrayRef<int> Mask,
9374 SelectionDAG &DAG) {
9375 assert(VT.getSizeInBits() >= 256 &&
9376 "Only for 256-bit or wider vector shuffles!");
9377 assert(V1.getSimpleValueType() == VT && "Bad operand type!");
9378 assert(V2.getSimpleValueType() == VT && "Bad operand type!");
9380 ArrayRef<int> LoMask = Mask.slice(0, Mask.size() / 2);
9381 ArrayRef<int> HiMask = Mask.slice(Mask.size() / 2);
9383 int NumElements = VT.getVectorNumElements();
9384 int SplitNumElements = NumElements / 2;
9385 MVT ScalarVT = VT.getScalarType();
9386 MVT SplitVT = MVT::getVectorVT(ScalarVT, NumElements / 2);
9388 // Rather than splitting build-vectors, just build two narrower build
9389 // vectors. This helps shuffling with splats and zeros.
9390 auto SplitVector = [&](SDValue V) {
9391 while (V.getOpcode() == ISD::BITCAST)
9392 V = V->getOperand(0);
9394 MVT OrigVT = V.getSimpleValueType();
9395 int OrigNumElements = OrigVT.getVectorNumElements();
9396 int OrigSplitNumElements = OrigNumElements / 2;
9397 MVT OrigScalarVT = OrigVT.getScalarType();
9398 MVT OrigSplitVT = MVT::getVectorVT(OrigScalarVT, OrigNumElements / 2);
9402 auto *BV = dyn_cast<BuildVectorSDNode>(V);
9404 LoV = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, OrigSplitVT, V,
9405 DAG.getIntPtrConstant(0, DL));
9406 HiV = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, OrigSplitVT, V,
9407 DAG.getIntPtrConstant(OrigSplitNumElements, DL));
9410 SmallVector<SDValue, 16> LoOps, HiOps;
9411 for (int i = 0; i < OrigSplitNumElements; ++i) {
9412 LoOps.push_back(BV->getOperand(i));
9413 HiOps.push_back(BV->getOperand(i + OrigSplitNumElements));
9415 LoV = DAG.getNode(ISD::BUILD_VECTOR, DL, OrigSplitVT, LoOps);
9416 HiV = DAG.getNode(ISD::BUILD_VECTOR, DL, OrigSplitVT, HiOps);
9418 return std::make_pair(DAG.getBitcast(SplitVT, LoV),
9419 DAG.getBitcast(SplitVT, HiV));
9422 SDValue LoV1, HiV1, LoV2, HiV2;
9423 std::tie(LoV1, HiV1) = SplitVector(V1);
9424 std::tie(LoV2, HiV2) = SplitVector(V2);
9426 // Now create two 4-way blends of these half-width vectors.
9427 auto HalfBlend = [&](ArrayRef<int> HalfMask) {
9428 bool UseLoV1 = false, UseHiV1 = false, UseLoV2 = false, UseHiV2 = false;
9429 SmallVector<int, 32> V1BlendMask, V2BlendMask, BlendMask;
9430 for (int i = 0; i < SplitNumElements; ++i) {
9431 int M = HalfMask[i];
9432 if (M >= NumElements) {
9433 if (M >= NumElements + SplitNumElements)
9437 V2BlendMask.push_back(M - NumElements);
9438 V1BlendMask.push_back(-1);
9439 BlendMask.push_back(SplitNumElements + i);
9440 } else if (M >= 0) {
9441 if (M >= SplitNumElements)
9445 V2BlendMask.push_back(-1);
9446 V1BlendMask.push_back(M);
9447 BlendMask.push_back(i);
9449 V2BlendMask.push_back(-1);
9450 V1BlendMask.push_back(-1);
9451 BlendMask.push_back(-1);
9455 // Because the lowering happens after all combining takes place, we need to
9456 // manually combine these blend masks as much as possible so that we create
9457 // a minimal number of high-level vector shuffle nodes.
9459 // First try just blending the halves of V1 or V2.
9460 if (!UseLoV1 && !UseHiV1 && !UseLoV2 && !UseHiV2)
9461 return DAG.getUNDEF(SplitVT);
9462 if (!UseLoV2 && !UseHiV2)
9463 return DAG.getVectorShuffle(SplitVT, DL, LoV1, HiV1, V1BlendMask);
9464 if (!UseLoV1 && !UseHiV1)
9465 return DAG.getVectorShuffle(SplitVT, DL, LoV2, HiV2, V2BlendMask);
9467 SDValue V1Blend, V2Blend;
9468 if (UseLoV1 && UseHiV1) {
9470 DAG.getVectorShuffle(SplitVT, DL, LoV1, HiV1, V1BlendMask);
9472 // We only use half of V1 so map the usage down into the final blend mask.
9473 V1Blend = UseLoV1 ? LoV1 : HiV1;
9474 for (int i = 0; i < SplitNumElements; ++i)
9475 if (BlendMask[i] >= 0 && BlendMask[i] < SplitNumElements)
9476 BlendMask[i] = V1BlendMask[i] - (UseLoV1 ? 0 : SplitNumElements);
9478 if (UseLoV2 && UseHiV2) {
9480 DAG.getVectorShuffle(SplitVT, DL, LoV2, HiV2, V2BlendMask);
9482 // We only use half of V2 so map the usage down into the final blend mask.
9483 V2Blend = UseLoV2 ? LoV2 : HiV2;
9484 for (int i = 0; i < SplitNumElements; ++i)
9485 if (BlendMask[i] >= SplitNumElements)
9486 BlendMask[i] = V2BlendMask[i] + (UseLoV2 ? SplitNumElements : 0);
9488 return DAG.getVectorShuffle(SplitVT, DL, V1Blend, V2Blend, BlendMask);
9490 SDValue Lo = HalfBlend(LoMask);
9491 SDValue Hi = HalfBlend(HiMask);
9492 return DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, Lo, Hi);
9495 /// \brief Either split a vector in halves or decompose the shuffles and the
9498 /// This is provided as a good fallback for many lowerings of non-single-input
9499 /// shuffles with more than one 128-bit lane. In those cases, we want to select
9500 /// between splitting the shuffle into 128-bit components and stitching those
9501 /// back together vs. extracting the single-input shuffles and blending those
9503 static SDValue lowerVectorShuffleAsSplitOrBlend(SDLoc DL, MVT VT, SDValue V1,
9504 SDValue V2, ArrayRef<int> Mask,
9505 SelectionDAG &DAG) {
9506 assert(!isSingleInputShuffleMask(Mask) && "This routine must not be used to "
9507 "lower single-input shuffles as it "
9508 "could then recurse on itself.");
9509 int Size = Mask.size();
9511 // If this can be modeled as a broadcast of two elements followed by a blend,
9512 // prefer that lowering. This is especially important because broadcasts can
9513 // often fold with memory operands.
9514 auto DoBothBroadcast = [&] {
9515 int V1BroadcastIdx = -1, V2BroadcastIdx = -1;
9518 if (V2BroadcastIdx == -1)
9519 V2BroadcastIdx = M - Size;
9520 else if (M - Size != V2BroadcastIdx)
9522 } else if (M >= 0) {
9523 if (V1BroadcastIdx == -1)
9525 else if (M != V1BroadcastIdx)
9530 if (DoBothBroadcast())
9531 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, VT, V1, V2, Mask,
9534 // If the inputs all stem from a single 128-bit lane of each input, then we
9535 // split them rather than blending because the split will decompose to
9536 // unusually few instructions.
9537 int LaneCount = VT.getSizeInBits() / 128;
9538 int LaneSize = Size / LaneCount;
9539 SmallBitVector LaneInputs[2];
9540 LaneInputs[0].resize(LaneCount, false);
9541 LaneInputs[1].resize(LaneCount, false);
9542 for (int i = 0; i < Size; ++i)
9544 LaneInputs[Mask[i] / Size][(Mask[i] % Size) / LaneSize] = true;
9545 if (LaneInputs[0].count() <= 1 && LaneInputs[1].count() <= 1)
9546 return splitAndLowerVectorShuffle(DL, VT, V1, V2, Mask, DAG);
9548 // Otherwise, just fall back to decomposed shuffles and a blend. This requires
9549 // that the decomposed single-input shuffles don't end up here.
9550 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, VT, V1, V2, Mask, DAG);
9553 /// \brief Lower a vector shuffle crossing multiple 128-bit lanes as
9554 /// a permutation and blend of those lanes.
9556 /// This essentially blends the out-of-lane inputs to each lane into the lane
9557 /// from a permuted copy of the vector. This lowering strategy results in four
9558 /// instructions in the worst case for a single-input cross lane shuffle which
9559 /// is lower than any other fully general cross-lane shuffle strategy I'm aware
9560 /// of. Special cases for each particular shuffle pattern should be handled
9561 /// prior to trying this lowering.
9562 static SDValue lowerVectorShuffleAsLanePermuteAndBlend(SDLoc DL, MVT VT,
9563 SDValue V1, SDValue V2,
9565 SelectionDAG &DAG) {
9566 // FIXME: This should probably be generalized for 512-bit vectors as well.
9567 assert(VT.getSizeInBits() == 256 && "Only for 256-bit vector shuffles!");
9568 int LaneSize = Mask.size() / 2;
9570 // If there are only inputs from one 128-bit lane, splitting will in fact be
9571 // less expensive. The flags track whether the given lane contains an element
9572 // that crosses to another lane.
9573 bool LaneCrossing[2] = {false, false};
9574 for (int i = 0, Size = Mask.size(); i < Size; ++i)
9575 if (Mask[i] >= 0 && (Mask[i] % Size) / LaneSize != i / LaneSize)
9576 LaneCrossing[(Mask[i] % Size) / LaneSize] = true;
9577 if (!LaneCrossing[0] || !LaneCrossing[1])
9578 return splitAndLowerVectorShuffle(DL, VT, V1, V2, Mask, DAG);
9580 if (isSingleInputShuffleMask(Mask)) {
9581 SmallVector<int, 32> FlippedBlendMask;
9582 for (int i = 0, Size = Mask.size(); i < Size; ++i)
9583 FlippedBlendMask.push_back(
9584 Mask[i] < 0 ? -1 : (((Mask[i] % Size) / LaneSize == i / LaneSize)
9586 : Mask[i] % LaneSize +
9587 (i / LaneSize) * LaneSize + Size));
9589 // Flip the vector, and blend the results which should now be in-lane. The
9590 // VPERM2X128 mask uses the low 2 bits for the low source and bits 4 and
9591 // 5 for the high source. The value 3 selects the high half of source 2 and
9592 // the value 2 selects the low half of source 2. We only use source 2 to
9593 // allow folding it into a memory operand.
9594 unsigned PERMMask = 3 | 2 << 4;
9595 SDValue Flipped = DAG.getNode(X86ISD::VPERM2X128, DL, VT, DAG.getUNDEF(VT),
9596 V1, DAG.getConstant(PERMMask, DL, MVT::i8));
9597 return DAG.getVectorShuffle(VT, DL, V1, Flipped, FlippedBlendMask);
9600 // This now reduces to two single-input shuffles of V1 and V2 which at worst
9601 // will be handled by the above logic and a blend of the results, much like
9602 // other patterns in AVX.
9603 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, VT, V1, V2, Mask, DAG);
9606 /// \brief Handle lowering 2-lane 128-bit shuffles.
9607 static SDValue lowerV2X128VectorShuffle(SDLoc DL, MVT VT, SDValue V1,
9608 SDValue V2, ArrayRef<int> Mask,
9609 const X86Subtarget *Subtarget,
9610 SelectionDAG &DAG) {
9611 // TODO: If minimizing size and one of the inputs is a zero vector and the
9612 // the zero vector has only one use, we could use a VPERM2X128 to save the
9613 // instruction bytes needed to explicitly generate the zero vector.
9615 // Blends are faster and handle all the non-lane-crossing cases.
9616 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, VT, V1, V2, Mask,
9620 bool IsV1Zero = ISD::isBuildVectorAllZeros(V1.getNode());
9621 bool IsV2Zero = ISD::isBuildVectorAllZeros(V2.getNode());
9623 // If either input operand is a zero vector, use VPERM2X128 because its mask
9624 // allows us to replace the zero input with an implicit zero.
9625 if (!IsV1Zero && !IsV2Zero) {
9626 // Check for patterns which can be matched with a single insert of a 128-bit
9628 bool OnlyUsesV1 = isShuffleEquivalent(V1, V2, Mask, {0, 1, 0, 1});
9629 if (OnlyUsesV1 || isShuffleEquivalent(V1, V2, Mask, {0, 1, 4, 5})) {
9630 MVT SubVT = MVT::getVectorVT(VT.getVectorElementType(),
9631 VT.getVectorNumElements() / 2);
9632 SDValue LoV = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, SubVT, V1,
9633 DAG.getIntPtrConstant(0, DL));
9634 SDValue HiV = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, SubVT,
9635 OnlyUsesV1 ? V1 : V2,
9636 DAG.getIntPtrConstant(0, DL));
9637 return DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, LoV, HiV);
9641 // Otherwise form a 128-bit permutation. After accounting for undefs,
9642 // convert the 64-bit shuffle mask selection values into 128-bit
9643 // selection bits by dividing the indexes by 2 and shifting into positions
9644 // defined by a vperm2*128 instruction's immediate control byte.
9646 // The immediate permute control byte looks like this:
9647 // [1:0] - select 128 bits from sources for low half of destination
9649 // [3] - zero low half of destination
9650 // [5:4] - select 128 bits from sources for high half of destination
9652 // [7] - zero high half of destination
9654 int MaskLO = Mask[0];
9655 if (MaskLO == SM_SentinelUndef)
9656 MaskLO = Mask[1] == SM_SentinelUndef ? 0 : Mask[1];
9658 int MaskHI = Mask[2];
9659 if (MaskHI == SM_SentinelUndef)
9660 MaskHI = Mask[3] == SM_SentinelUndef ? 0 : Mask[3];
9662 unsigned PermMask = MaskLO / 2 | (MaskHI / 2) << 4;
9664 // If either input is a zero vector, replace it with an undef input.
9665 // Shuffle mask values < 4 are selecting elements of V1.
9666 // Shuffle mask values >= 4 are selecting elements of V2.
9667 // Adjust each half of the permute mask by clearing the half that was
9668 // selecting the zero vector and setting the zero mask bit.
9670 V1 = DAG.getUNDEF(VT);
9672 PermMask = (PermMask & 0xf0) | 0x08;
9674 PermMask = (PermMask & 0x0f) | 0x80;
9677 V2 = DAG.getUNDEF(VT);
9679 PermMask = (PermMask & 0xf0) | 0x08;
9681 PermMask = (PermMask & 0x0f) | 0x80;
9684 return DAG.getNode(X86ISD::VPERM2X128, DL, VT, V1, V2,
9685 DAG.getConstant(PermMask, DL, MVT::i8));
9688 /// \brief Lower a vector shuffle by first fixing the 128-bit lanes and then
9689 /// shuffling each lane.
9691 /// This will only succeed when the result of fixing the 128-bit lanes results
9692 /// in a single-input non-lane-crossing shuffle with a repeating shuffle mask in
9693 /// each 128-bit lanes. This handles many cases where we can quickly blend away
9694 /// the lane crosses early and then use simpler shuffles within each lane.
9696 /// FIXME: It might be worthwhile at some point to support this without
9697 /// requiring the 128-bit lane-relative shuffles to be repeating, but currently
9698 /// in x86 only floating point has interesting non-repeating shuffles, and even
9699 /// those are still *marginally* more expensive.
9700 static SDValue lowerVectorShuffleByMerging128BitLanes(
9701 SDLoc DL, MVT VT, SDValue V1, SDValue V2, ArrayRef<int> Mask,
9702 const X86Subtarget *Subtarget, SelectionDAG &DAG) {
9703 assert(!isSingleInputShuffleMask(Mask) &&
9704 "This is only useful with multiple inputs.");
9706 int Size = Mask.size();
9707 int LaneSize = 128 / VT.getScalarSizeInBits();
9708 int NumLanes = Size / LaneSize;
9709 assert(NumLanes > 1 && "Only handles 256-bit and wider shuffles.");
9711 // See if we can build a hypothetical 128-bit lane-fixing shuffle mask. Also
9712 // check whether the in-128-bit lane shuffles share a repeating pattern.
9713 SmallVector<int, 4> Lanes;
9714 Lanes.resize(NumLanes, -1);
9715 SmallVector<int, 4> InLaneMask;
9716 InLaneMask.resize(LaneSize, -1);
9717 for (int i = 0; i < Size; ++i) {
9721 int j = i / LaneSize;
9724 // First entry we've seen for this lane.
9725 Lanes[j] = Mask[i] / LaneSize;
9726 } else if (Lanes[j] != Mask[i] / LaneSize) {
9727 // This doesn't match the lane selected previously!
9731 // Check that within each lane we have a consistent shuffle mask.
9732 int k = i % LaneSize;
9733 if (InLaneMask[k] < 0) {
9734 InLaneMask[k] = Mask[i] % LaneSize;
9735 } else if (InLaneMask[k] != Mask[i] % LaneSize) {
9736 // This doesn't fit a repeating in-lane mask.
9741 // First shuffle the lanes into place.
9742 MVT LaneVT = MVT::getVectorVT(VT.isFloatingPoint() ? MVT::f64 : MVT::i64,
9743 VT.getSizeInBits() / 64);
9744 SmallVector<int, 8> LaneMask;
9745 LaneMask.resize(NumLanes * 2, -1);
9746 for (int i = 0; i < NumLanes; ++i)
9747 if (Lanes[i] >= 0) {
9748 LaneMask[2 * i + 0] = 2*Lanes[i] + 0;
9749 LaneMask[2 * i + 1] = 2*Lanes[i] + 1;
9752 V1 = DAG.getBitcast(LaneVT, V1);
9753 V2 = DAG.getBitcast(LaneVT, V2);
9754 SDValue LaneShuffle = DAG.getVectorShuffle(LaneVT, DL, V1, V2, LaneMask);
9756 // Cast it back to the type we actually want.
9757 LaneShuffle = DAG.getBitcast(VT, LaneShuffle);
9759 // Now do a simple shuffle that isn't lane crossing.
9760 SmallVector<int, 8> NewMask;
9761 NewMask.resize(Size, -1);
9762 for (int i = 0; i < Size; ++i)
9764 NewMask[i] = (i / LaneSize) * LaneSize + Mask[i] % LaneSize;
9765 assert(!is128BitLaneCrossingShuffleMask(VT, NewMask) &&
9766 "Must not introduce lane crosses at this point!");
9768 return DAG.getVectorShuffle(VT, DL, LaneShuffle, DAG.getUNDEF(VT), NewMask);
9771 /// \brief Test whether the specified input (0 or 1) is in-place blended by the
9774 /// This returns true if the elements from a particular input are already in the
9775 /// slot required by the given mask and require no permutation.
9776 static bool isShuffleMaskInputInPlace(int Input, ArrayRef<int> Mask) {
9777 assert((Input == 0 || Input == 1) && "Only two inputs to shuffles.");
9778 int Size = Mask.size();
9779 for (int i = 0; i < Size; ++i)
9780 if (Mask[i] >= 0 && Mask[i] / Size == Input && Mask[i] % Size != i)
9786 static SDValue lowerVectorShuffleWithSHUFPD(SDLoc DL, MVT VT,
9787 ArrayRef<int> Mask, SDValue V1,
9788 SDValue V2, SelectionDAG &DAG) {
9790 // Mask for V8F64: 0/1, 8/9, 2/3, 10/11, 4/5, ..
9791 // Mask for V4F64; 0/1, 4/5, 2/3, 6/7..
9792 assert(VT.getScalarSizeInBits() == 64 && "Unexpected data type for VSHUFPD");
9793 int NumElts = VT.getVectorNumElements();
9794 bool ShufpdMask = true;
9795 bool CommutableMask = true;
9796 unsigned Immediate = 0;
9797 for (int i = 0; i < NumElts; ++i) {
9800 int Val = (i & 6) + NumElts * (i & 1);
9801 int CommutVal = (i & 0xe) + NumElts * ((i & 1)^1);
9802 if (Mask[i] < Val || Mask[i] > Val + 1)
9804 if (Mask[i] < CommutVal || Mask[i] > CommutVal + 1)
9805 CommutableMask = false;
9806 Immediate |= (Mask[i] % 2) << i;
9809 return DAG.getNode(X86ISD::SHUFP, DL, VT, V1, V2,
9810 DAG.getConstant(Immediate, DL, MVT::i8));
9812 return DAG.getNode(X86ISD::SHUFP, DL, VT, V2, V1,
9813 DAG.getConstant(Immediate, DL, MVT::i8));
9817 /// \brief Handle lowering of 4-lane 64-bit floating point shuffles.
9819 /// Also ends up handling lowering of 4-lane 64-bit integer shuffles when AVX2
9820 /// isn't available.
9821 static SDValue lowerV4F64VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
9822 const X86Subtarget *Subtarget,
9823 SelectionDAG &DAG) {
9825 assert(V1.getSimpleValueType() == MVT::v4f64 && "Bad operand type!");
9826 assert(V2.getSimpleValueType() == MVT::v4f64 && "Bad operand type!");
9827 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
9828 ArrayRef<int> Mask = SVOp->getMask();
9829 assert(Mask.size() == 4 && "Unexpected mask size for v4 shuffle!");
9831 SmallVector<int, 4> WidenedMask;
9832 if (canWidenShuffleElements(Mask, WidenedMask))
9833 return lowerV2X128VectorShuffle(DL, MVT::v4f64, V1, V2, Mask, Subtarget,
9836 if (isSingleInputShuffleMask(Mask)) {
9837 // Check for being able to broadcast a single element.
9838 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v4f64, V1,
9839 Mask, Subtarget, DAG))
9842 // Use low duplicate instructions for masks that match their pattern.
9843 if (isShuffleEquivalent(V1, V2, Mask, {0, 0, 2, 2}))
9844 return DAG.getNode(X86ISD::MOVDDUP, DL, MVT::v4f64, V1);
9846 if (!is128BitLaneCrossingShuffleMask(MVT::v4f64, Mask)) {
9847 // Non-half-crossing single input shuffles can be lowerid with an
9848 // interleaved permutation.
9849 unsigned VPERMILPMask = (Mask[0] == 1) | ((Mask[1] == 1) << 1) |
9850 ((Mask[2] == 3) << 2) | ((Mask[3] == 3) << 3);
9851 return DAG.getNode(X86ISD::VPERMILPI, DL, MVT::v4f64, V1,
9852 DAG.getConstant(VPERMILPMask, DL, MVT::i8));
9855 // With AVX2 we have direct support for this permutation.
9856 if (Subtarget->hasAVX2())
9857 return DAG.getNode(X86ISD::VPERMI, DL, MVT::v4f64, V1,
9858 getV4X86ShuffleImm8ForMask(Mask, DL, DAG));
9860 // Otherwise, fall back.
9861 return lowerVectorShuffleAsLanePermuteAndBlend(DL, MVT::v4f64, V1, V2, Mask,
9865 // X86 has dedicated unpack instructions that can handle specific blend
9866 // operations: UNPCKH and UNPCKL.
9867 if (isShuffleEquivalent(V1, V2, Mask, {0, 4, 2, 6}))
9868 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v4f64, V1, V2);
9869 if (isShuffleEquivalent(V1, V2, Mask, {1, 5, 3, 7}))
9870 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v4f64, V1, V2);
9871 if (isShuffleEquivalent(V1, V2, Mask, {4, 0, 6, 2}))
9872 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v4f64, V2, V1);
9873 if (isShuffleEquivalent(V1, V2, Mask, {5, 1, 7, 3}))
9874 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v4f64, V2, V1);
9876 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v4f64, V1, V2, Mask,
9880 // Check if the blend happens to exactly fit that of SHUFPD.
9882 lowerVectorShuffleWithSHUFPD(DL, MVT::v4f64, Mask, V1, V2, DAG))
9885 // Try to simplify this by merging 128-bit lanes to enable a lane-based
9886 // shuffle. However, if we have AVX2 and either inputs are already in place,
9887 // we will be able to shuffle even across lanes the other input in a single
9888 // instruction so skip this pattern.
9889 if (!(Subtarget->hasAVX2() && (isShuffleMaskInputInPlace(0, Mask) ||
9890 isShuffleMaskInputInPlace(1, Mask))))
9891 if (SDValue Result = lowerVectorShuffleByMerging128BitLanes(
9892 DL, MVT::v4f64, V1, V2, Mask, Subtarget, DAG))
9895 // If we have AVX2 then we always want to lower with a blend because an v4 we
9896 // can fully permute the elements.
9897 if (Subtarget->hasAVX2())
9898 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v4f64, V1, V2,
9901 // Otherwise fall back on generic lowering.
9902 return lowerVectorShuffleAsSplitOrBlend(DL, MVT::v4f64, V1, V2, Mask, DAG);
9905 /// \brief Handle lowering of 4-lane 64-bit integer shuffles.
9907 /// This routine is only called when we have AVX2 and thus a reasonable
9908 /// instruction set for v4i64 shuffling..
9909 static SDValue lowerV4I64VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
9910 const X86Subtarget *Subtarget,
9911 SelectionDAG &DAG) {
9913 assert(V1.getSimpleValueType() == MVT::v4i64 && "Bad operand type!");
9914 assert(V2.getSimpleValueType() == MVT::v4i64 && "Bad operand type!");
9915 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
9916 ArrayRef<int> Mask = SVOp->getMask();
9917 assert(Mask.size() == 4 && "Unexpected mask size for v4 shuffle!");
9918 assert(Subtarget->hasAVX2() && "We can only lower v4i64 with AVX2!");
9920 SmallVector<int, 4> WidenedMask;
9921 if (canWidenShuffleElements(Mask, WidenedMask))
9922 return lowerV2X128VectorShuffle(DL, MVT::v4i64, V1, V2, Mask, Subtarget,
9925 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v4i64, V1, V2, Mask,
9929 // Check for being able to broadcast a single element.
9930 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v4i64, V1,
9931 Mask, Subtarget, DAG))
9934 // When the shuffle is mirrored between the 128-bit lanes of the unit, we can
9935 // use lower latency instructions that will operate on both 128-bit lanes.
9936 SmallVector<int, 2> RepeatedMask;
9937 if (is128BitLaneRepeatedShuffleMask(MVT::v4i64, Mask, RepeatedMask)) {
9938 if (isSingleInputShuffleMask(Mask)) {
9939 int PSHUFDMask[] = {-1, -1, -1, -1};
9940 for (int i = 0; i < 2; ++i)
9941 if (RepeatedMask[i] >= 0) {
9942 PSHUFDMask[2 * i] = 2 * RepeatedMask[i];
9943 PSHUFDMask[2 * i + 1] = 2 * RepeatedMask[i] + 1;
9945 return DAG.getBitcast(
9947 DAG.getNode(X86ISD::PSHUFD, DL, MVT::v8i32,
9948 DAG.getBitcast(MVT::v8i32, V1),
9949 getV4X86ShuffleImm8ForMask(PSHUFDMask, DL, DAG)));
9953 // AVX2 provides a direct instruction for permuting a single input across
9955 if (isSingleInputShuffleMask(Mask))
9956 return DAG.getNode(X86ISD::VPERMI, DL, MVT::v4i64, V1,
9957 getV4X86ShuffleImm8ForMask(Mask, DL, DAG));
9959 // Try to use shift instructions.
9961 lowerVectorShuffleAsShift(DL, MVT::v4i64, V1, V2, Mask, DAG))
9964 // Use dedicated unpack instructions for masks that match their pattern.
9965 if (isShuffleEquivalent(V1, V2, Mask, {0, 4, 2, 6}))
9966 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v4i64, V1, V2);
9967 if (isShuffleEquivalent(V1, V2, Mask, {1, 5, 3, 7}))
9968 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v4i64, V1, V2);
9969 if (isShuffleEquivalent(V1, V2, Mask, {4, 0, 6, 2}))
9970 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v4i64, V2, V1);
9971 if (isShuffleEquivalent(V1, V2, Mask, {5, 1, 7, 3}))
9972 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v4i64, V2, V1);
9974 // Try to simplify this by merging 128-bit lanes to enable a lane-based
9975 // shuffle. However, if we have AVX2 and either inputs are already in place,
9976 // we will be able to shuffle even across lanes the other input in a single
9977 // instruction so skip this pattern.
9978 if (!(Subtarget->hasAVX2() && (isShuffleMaskInputInPlace(0, Mask) ||
9979 isShuffleMaskInputInPlace(1, Mask))))
9980 if (SDValue Result = lowerVectorShuffleByMerging128BitLanes(
9981 DL, MVT::v4i64, V1, V2, Mask, Subtarget, DAG))
9984 // Otherwise fall back on generic blend lowering.
9985 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v4i64, V1, V2,
9989 /// \brief Handle lowering of 8-lane 32-bit floating point shuffles.
9991 /// Also ends up handling lowering of 8-lane 32-bit integer shuffles when AVX2
9992 /// isn't available.
9993 static SDValue lowerV8F32VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
9994 const X86Subtarget *Subtarget,
9995 SelectionDAG &DAG) {
9997 assert(V1.getSimpleValueType() == MVT::v8f32 && "Bad operand type!");
9998 assert(V2.getSimpleValueType() == MVT::v8f32 && "Bad operand type!");
9999 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10000 ArrayRef<int> Mask = SVOp->getMask();
10001 assert(Mask.size() == 8 && "Unexpected mask size for v8 shuffle!");
10003 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v8f32, V1, V2, Mask,
10007 // Check for being able to broadcast a single element.
10008 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v8f32, V1,
10009 Mask, Subtarget, DAG))
10012 // If the shuffle mask is repeated in each 128-bit lane, we have many more
10013 // options to efficiently lower the shuffle.
10014 SmallVector<int, 4> RepeatedMask;
10015 if (is128BitLaneRepeatedShuffleMask(MVT::v8f32, Mask, RepeatedMask)) {
10016 assert(RepeatedMask.size() == 4 &&
10017 "Repeated masks must be half the mask width!");
10019 // Use even/odd duplicate instructions for masks that match their pattern.
10020 if (isShuffleEquivalent(V1, V2, Mask, {0, 0, 2, 2, 4, 4, 6, 6}))
10021 return DAG.getNode(X86ISD::MOVSLDUP, DL, MVT::v8f32, V1);
10022 if (isShuffleEquivalent(V1, V2, Mask, {1, 1, 3, 3, 5, 5, 7, 7}))
10023 return DAG.getNode(X86ISD::MOVSHDUP, DL, MVT::v8f32, V1);
10025 if (isSingleInputShuffleMask(Mask))
10026 return DAG.getNode(X86ISD::VPERMILPI, DL, MVT::v8f32, V1,
10027 getV4X86ShuffleImm8ForMask(RepeatedMask, DL, DAG));
10029 // Use dedicated unpack instructions for masks that match their pattern.
10030 if (isShuffleEquivalent(V1, V2, Mask, {0, 8, 1, 9, 4, 12, 5, 13}))
10031 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v8f32, V1, V2);
10032 if (isShuffleEquivalent(V1, V2, Mask, {2, 10, 3, 11, 6, 14, 7, 15}))
10033 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v8f32, V1, V2);
10034 if (isShuffleEquivalent(V1, V2, Mask, {8, 0, 9, 1, 12, 4, 13, 5}))
10035 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v8f32, V2, V1);
10036 if (isShuffleEquivalent(V1, V2, Mask, {10, 2, 11, 3, 14, 6, 15, 7}))
10037 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v8f32, V2, V1);
10039 // Otherwise, fall back to a SHUFPS sequence. Here it is important that we
10040 // have already handled any direct blends. We also need to squash the
10041 // repeated mask into a simulated v4f32 mask.
10042 for (int i = 0; i < 4; ++i)
10043 if (RepeatedMask[i] >= 8)
10044 RepeatedMask[i] -= 4;
10045 return lowerVectorShuffleWithSHUFPS(DL, MVT::v8f32, RepeatedMask, V1, V2, DAG);
10048 // If we have a single input shuffle with different shuffle patterns in the
10049 // two 128-bit lanes use the variable mask to VPERMILPS.
10050 if (isSingleInputShuffleMask(Mask)) {
10051 SDValue VPermMask[8];
10052 for (int i = 0; i < 8; ++i)
10053 VPermMask[i] = Mask[i] < 0 ? DAG.getUNDEF(MVT::i32)
10054 : DAG.getConstant(Mask[i], DL, MVT::i32);
10055 if (!is128BitLaneCrossingShuffleMask(MVT::v8f32, Mask))
10056 return DAG.getNode(
10057 X86ISD::VPERMILPV, DL, MVT::v8f32, V1,
10058 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v8i32, VPermMask));
10060 if (Subtarget->hasAVX2())
10061 return DAG.getNode(
10062 X86ISD::VPERMV, DL, MVT::v8f32,
10063 DAG.getBitcast(MVT::v8f32, DAG.getNode(ISD::BUILD_VECTOR, DL,
10064 MVT::v8i32, VPermMask)),
10067 // Otherwise, fall back.
10068 return lowerVectorShuffleAsLanePermuteAndBlend(DL, MVT::v8f32, V1, V2, Mask,
10072 // Try to simplify this by merging 128-bit lanes to enable a lane-based
10074 if (SDValue Result = lowerVectorShuffleByMerging128BitLanes(
10075 DL, MVT::v8f32, V1, V2, Mask, Subtarget, DAG))
10078 // If we have AVX2 then we always want to lower with a blend because at v8 we
10079 // can fully permute the elements.
10080 if (Subtarget->hasAVX2())
10081 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v8f32, V1, V2,
10084 // Otherwise fall back on generic lowering.
10085 return lowerVectorShuffleAsSplitOrBlend(DL, MVT::v8f32, V1, V2, Mask, DAG);
10088 /// \brief Handle lowering of 8-lane 32-bit integer shuffles.
10090 /// This routine is only called when we have AVX2 and thus a reasonable
10091 /// instruction set for v8i32 shuffling..
10092 static SDValue lowerV8I32VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10093 const X86Subtarget *Subtarget,
10094 SelectionDAG &DAG) {
10096 assert(V1.getSimpleValueType() == MVT::v8i32 && "Bad operand type!");
10097 assert(V2.getSimpleValueType() == MVT::v8i32 && "Bad operand type!");
10098 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10099 ArrayRef<int> Mask = SVOp->getMask();
10100 assert(Mask.size() == 8 && "Unexpected mask size for v8 shuffle!");
10101 assert(Subtarget->hasAVX2() && "We can only lower v8i32 with AVX2!");
10103 // Whenever we can lower this as a zext, that instruction is strictly faster
10104 // than any alternative. It also allows us to fold memory operands into the
10105 // shuffle in many cases.
10106 if (SDValue ZExt = lowerVectorShuffleAsZeroOrAnyExtend(DL, MVT::v8i32, V1, V2,
10107 Mask, Subtarget, DAG))
10110 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v8i32, V1, V2, Mask,
10114 // Check for being able to broadcast a single element.
10115 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v8i32, V1,
10116 Mask, Subtarget, DAG))
10119 // If the shuffle mask is repeated in each 128-bit lane we can use more
10120 // efficient instructions that mirror the shuffles across the two 128-bit
10122 SmallVector<int, 4> RepeatedMask;
10123 if (is128BitLaneRepeatedShuffleMask(MVT::v8i32, Mask, RepeatedMask)) {
10124 assert(RepeatedMask.size() == 4 && "Unexpected repeated mask size!");
10125 if (isSingleInputShuffleMask(Mask))
10126 return DAG.getNode(X86ISD::PSHUFD, DL, MVT::v8i32, V1,
10127 getV4X86ShuffleImm8ForMask(RepeatedMask, DL, DAG));
10129 // Use dedicated unpack instructions for masks that match their pattern.
10130 if (isShuffleEquivalent(V1, V2, Mask, {0, 8, 1, 9, 4, 12, 5, 13}))
10131 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v8i32, V1, V2);
10132 if (isShuffleEquivalent(V1, V2, Mask, {2, 10, 3, 11, 6, 14, 7, 15}))
10133 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v8i32, V1, V2);
10134 if (isShuffleEquivalent(V1, V2, Mask, {8, 0, 9, 1, 12, 4, 13, 5}))
10135 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v8i32, V2, V1);
10136 if (isShuffleEquivalent(V1, V2, Mask, {10, 2, 11, 3, 14, 6, 15, 7}))
10137 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v8i32, V2, V1);
10140 // Try to use shift instructions.
10141 if (SDValue Shift =
10142 lowerVectorShuffleAsShift(DL, MVT::v8i32, V1, V2, Mask, DAG))
10145 if (SDValue Rotate = lowerVectorShuffleAsByteRotate(
10146 DL, MVT::v8i32, V1, V2, Mask, Subtarget, DAG))
10149 // If the shuffle patterns aren't repeated but it is a single input, directly
10150 // generate a cross-lane VPERMD instruction.
10151 if (isSingleInputShuffleMask(Mask)) {
10152 SDValue VPermMask[8];
10153 for (int i = 0; i < 8; ++i)
10154 VPermMask[i] = Mask[i] < 0 ? DAG.getUNDEF(MVT::i32)
10155 : DAG.getConstant(Mask[i], DL, MVT::i32);
10156 return DAG.getNode(
10157 X86ISD::VPERMV, DL, MVT::v8i32,
10158 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v8i32, VPermMask), V1);
10161 // Try to simplify this by merging 128-bit lanes to enable a lane-based
10163 if (SDValue Result = lowerVectorShuffleByMerging128BitLanes(
10164 DL, MVT::v8i32, V1, V2, Mask, Subtarget, DAG))
10167 // Otherwise fall back on generic blend lowering.
10168 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v8i32, V1, V2,
10172 /// \brief Handle lowering of 16-lane 16-bit integer shuffles.
10174 /// This routine is only called when we have AVX2 and thus a reasonable
10175 /// instruction set for v16i16 shuffling..
10176 static SDValue lowerV16I16VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10177 const X86Subtarget *Subtarget,
10178 SelectionDAG &DAG) {
10180 assert(V1.getSimpleValueType() == MVT::v16i16 && "Bad operand type!");
10181 assert(V2.getSimpleValueType() == MVT::v16i16 && "Bad operand type!");
10182 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10183 ArrayRef<int> Mask = SVOp->getMask();
10184 assert(Mask.size() == 16 && "Unexpected mask size for v16 shuffle!");
10185 assert(Subtarget->hasAVX2() && "We can only lower v16i16 with AVX2!");
10187 // Whenever we can lower this as a zext, that instruction is strictly faster
10188 // than any alternative. It also allows us to fold memory operands into the
10189 // shuffle in many cases.
10190 if (SDValue ZExt = lowerVectorShuffleAsZeroOrAnyExtend(DL, MVT::v16i16, V1, V2,
10191 Mask, Subtarget, DAG))
10194 // Check for being able to broadcast a single element.
10195 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v16i16, V1,
10196 Mask, Subtarget, DAG))
10199 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v16i16, V1, V2, Mask,
10203 // Use dedicated unpack instructions for masks that match their pattern.
10204 if (isShuffleEquivalent(V1, V2, Mask,
10205 {// First 128-bit lane:
10206 0, 16, 1, 17, 2, 18, 3, 19,
10207 // Second 128-bit lane:
10208 8, 24, 9, 25, 10, 26, 11, 27}))
10209 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v16i16, V1, V2);
10210 if (isShuffleEquivalent(V1, V2, Mask,
10211 {// First 128-bit lane:
10212 4, 20, 5, 21, 6, 22, 7, 23,
10213 // Second 128-bit lane:
10214 12, 28, 13, 29, 14, 30, 15, 31}))
10215 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v16i16, V1, V2);
10217 // Try to use shift instructions.
10218 if (SDValue Shift =
10219 lowerVectorShuffleAsShift(DL, MVT::v16i16, V1, V2, Mask, DAG))
10222 // Try to use byte rotation instructions.
10223 if (SDValue Rotate = lowerVectorShuffleAsByteRotate(
10224 DL, MVT::v16i16, V1, V2, Mask, Subtarget, DAG))
10227 if (isSingleInputShuffleMask(Mask)) {
10228 // There are no generalized cross-lane shuffle operations available on i16
10230 if (is128BitLaneCrossingShuffleMask(MVT::v16i16, Mask))
10231 return lowerVectorShuffleAsLanePermuteAndBlend(DL, MVT::v16i16, V1, V2,
10234 SmallVector<int, 8> RepeatedMask;
10235 if (is128BitLaneRepeatedShuffleMask(MVT::v16i16, Mask, RepeatedMask)) {
10236 // As this is a single-input shuffle, the repeated mask should be
10237 // a strictly valid v8i16 mask that we can pass through to the v8i16
10238 // lowering to handle even the v16 case.
10239 return lowerV8I16GeneralSingleInputVectorShuffle(
10240 DL, MVT::v16i16, V1, RepeatedMask, Subtarget, DAG);
10243 SDValue PSHUFBMask[32];
10244 for (int i = 0; i < 16; ++i) {
10245 if (Mask[i] == -1) {
10246 PSHUFBMask[2 * i] = PSHUFBMask[2 * i + 1] = DAG.getUNDEF(MVT::i8);
10250 int M = i < 8 ? Mask[i] : Mask[i] - 8;
10251 assert(M >= 0 && M < 8 && "Invalid single-input mask!");
10252 PSHUFBMask[2 * i] = DAG.getConstant(2 * M, DL, MVT::i8);
10253 PSHUFBMask[2 * i + 1] = DAG.getConstant(2 * M + 1, DL, MVT::i8);
10255 return DAG.getBitcast(MVT::v16i16,
10256 DAG.getNode(X86ISD::PSHUFB, DL, MVT::v32i8,
10257 DAG.getBitcast(MVT::v32i8, V1),
10258 DAG.getNode(ISD::BUILD_VECTOR, DL,
10259 MVT::v32i8, PSHUFBMask)));
10262 // Try to simplify this by merging 128-bit lanes to enable a lane-based
10264 if (SDValue Result = lowerVectorShuffleByMerging128BitLanes(
10265 DL, MVT::v16i16, V1, V2, Mask, Subtarget, DAG))
10268 // Otherwise fall back on generic lowering.
10269 return lowerVectorShuffleAsSplitOrBlend(DL, MVT::v16i16, V1, V2, Mask, DAG);
10272 /// \brief Handle lowering of 32-lane 8-bit integer shuffles.
10274 /// This routine is only called when we have AVX2 and thus a reasonable
10275 /// instruction set for v32i8 shuffling..
10276 static SDValue lowerV32I8VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10277 const X86Subtarget *Subtarget,
10278 SelectionDAG &DAG) {
10280 assert(V1.getSimpleValueType() == MVT::v32i8 && "Bad operand type!");
10281 assert(V2.getSimpleValueType() == MVT::v32i8 && "Bad operand type!");
10282 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10283 ArrayRef<int> Mask = SVOp->getMask();
10284 assert(Mask.size() == 32 && "Unexpected mask size for v32 shuffle!");
10285 assert(Subtarget->hasAVX2() && "We can only lower v32i8 with AVX2!");
10287 // Whenever we can lower this as a zext, that instruction is strictly faster
10288 // than any alternative. It also allows us to fold memory operands into the
10289 // shuffle in many cases.
10290 if (SDValue ZExt = lowerVectorShuffleAsZeroOrAnyExtend(DL, MVT::v32i8, V1, V2,
10291 Mask, Subtarget, DAG))
10294 // Check for being able to broadcast a single element.
10295 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v32i8, V1,
10296 Mask, Subtarget, DAG))
10299 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v32i8, V1, V2, Mask,
10303 // Use dedicated unpack instructions for masks that match their pattern.
10304 // Note that these are repeated 128-bit lane unpacks, not unpacks across all
10306 if (isShuffleEquivalent(
10308 {// First 128-bit lane:
10309 0, 32, 1, 33, 2, 34, 3, 35, 4, 36, 5, 37, 6, 38, 7, 39,
10310 // Second 128-bit lane:
10311 16, 48, 17, 49, 18, 50, 19, 51, 20, 52, 21, 53, 22, 54, 23, 55}))
10312 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v32i8, V1, V2);
10313 if (isShuffleEquivalent(
10315 {// First 128-bit lane:
10316 8, 40, 9, 41, 10, 42, 11, 43, 12, 44, 13, 45, 14, 46, 15, 47,
10317 // Second 128-bit lane:
10318 24, 56, 25, 57, 26, 58, 27, 59, 28, 60, 29, 61, 30, 62, 31, 63}))
10319 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v32i8, V1, V2);
10321 // Try to use shift instructions.
10322 if (SDValue Shift =
10323 lowerVectorShuffleAsShift(DL, MVT::v32i8, V1, V2, Mask, DAG))
10326 // Try to use byte rotation instructions.
10327 if (SDValue Rotate = lowerVectorShuffleAsByteRotate(
10328 DL, MVT::v32i8, V1, V2, Mask, Subtarget, DAG))
10331 if (isSingleInputShuffleMask(Mask)) {
10332 // There are no generalized cross-lane shuffle operations available on i8
10334 if (is128BitLaneCrossingShuffleMask(MVT::v32i8, Mask))
10335 return lowerVectorShuffleAsLanePermuteAndBlend(DL, MVT::v32i8, V1, V2,
10338 SDValue PSHUFBMask[32];
10339 for (int i = 0; i < 32; ++i)
10342 ? DAG.getUNDEF(MVT::i8)
10343 : DAG.getConstant(Mask[i] < 16 ? Mask[i] : Mask[i] - 16, DL,
10346 return DAG.getNode(
10347 X86ISD::PSHUFB, DL, MVT::v32i8, V1,
10348 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v32i8, PSHUFBMask));
10351 // Try to simplify this by merging 128-bit lanes to enable a lane-based
10353 if (SDValue Result = lowerVectorShuffleByMerging128BitLanes(
10354 DL, MVT::v32i8, V1, V2, Mask, Subtarget, DAG))
10357 // Otherwise fall back on generic lowering.
10358 return lowerVectorShuffleAsSplitOrBlend(DL, MVT::v32i8, V1, V2, Mask, DAG);
10361 /// \brief High-level routine to lower various 256-bit x86 vector shuffles.
10363 /// This routine either breaks down the specific type of a 256-bit x86 vector
10364 /// shuffle or splits it into two 128-bit shuffles and fuses the results back
10365 /// together based on the available instructions.
10366 static SDValue lower256BitVectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10367 MVT VT, const X86Subtarget *Subtarget,
10368 SelectionDAG &DAG) {
10370 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10371 ArrayRef<int> Mask = SVOp->getMask();
10373 // If we have a single input to the zero element, insert that into V1 if we
10374 // can do so cheaply.
10375 int NumElts = VT.getVectorNumElements();
10376 int NumV2Elements = std::count_if(Mask.begin(), Mask.end(), [NumElts](int M) {
10377 return M >= NumElts;
10380 if (NumV2Elements == 1 && Mask[0] >= NumElts)
10381 if (SDValue Insertion = lowerVectorShuffleAsElementInsertion(
10382 DL, VT, V1, V2, Mask, Subtarget, DAG))
10385 // There is a really nice hard cut-over between AVX1 and AVX2 that means we can
10386 // check for those subtargets here and avoid much of the subtarget querying in
10387 // the per-vector-type lowering routines. With AVX1 we have essentially *zero*
10388 // ability to manipulate a 256-bit vector with integer types. Since we'll use
10389 // floating point types there eventually, just immediately cast everything to
10390 // a float and operate entirely in that domain.
10391 if (VT.isInteger() && !Subtarget->hasAVX2()) {
10392 int ElementBits = VT.getScalarSizeInBits();
10393 if (ElementBits < 32)
10394 // No floating point type available, decompose into 128-bit vectors.
10395 return splitAndLowerVectorShuffle(DL, VT, V1, V2, Mask, DAG);
10397 MVT FpVT = MVT::getVectorVT(MVT::getFloatingPointVT(ElementBits),
10398 VT.getVectorNumElements());
10399 V1 = DAG.getBitcast(FpVT, V1);
10400 V2 = DAG.getBitcast(FpVT, V2);
10401 return DAG.getBitcast(VT, DAG.getVectorShuffle(FpVT, DL, V1, V2, Mask));
10404 switch (VT.SimpleTy) {
10406 return lowerV4F64VectorShuffle(Op, V1, V2, Subtarget, DAG);
10408 return lowerV4I64VectorShuffle(Op, V1, V2, Subtarget, DAG);
10410 return lowerV8F32VectorShuffle(Op, V1, V2, Subtarget, DAG);
10412 return lowerV8I32VectorShuffle(Op, V1, V2, Subtarget, DAG);
10414 return lowerV16I16VectorShuffle(Op, V1, V2, Subtarget, DAG);
10416 return lowerV32I8VectorShuffle(Op, V1, V2, Subtarget, DAG);
10419 llvm_unreachable("Not a valid 256-bit x86 vector type!");
10423 /// \brief Handle lowering of 8-lane 64-bit floating point shuffles.
10424 static SDValue lowerV8F64VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10425 const X86Subtarget *Subtarget,
10426 SelectionDAG &DAG) {
10428 assert(V1.getSimpleValueType() == MVT::v8f64 && "Bad operand type!");
10429 assert(V2.getSimpleValueType() == MVT::v8f64 && "Bad operand type!");
10430 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10431 ArrayRef<int> Mask = SVOp->getMask();
10432 assert(Mask.size() == 8 && "Unexpected mask size for v8 shuffle!");
10434 // X86 has dedicated unpack instructions that can handle specific blend
10435 // operations: UNPCKH and UNPCKL.
10436 if (isShuffleEquivalent(V1, V2, Mask, {0, 8, 2, 10, 4, 12, 6, 14}))
10437 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v8f64, V1, V2);
10438 if (isShuffleEquivalent(V1, V2, Mask, {1, 9, 3, 11, 5, 13, 7, 15}))
10439 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v8f64, V1, V2);
10441 // FIXME: Implement direct support for this type!
10442 return splitAndLowerVectorShuffle(DL, MVT::v8f64, V1, V2, Mask, DAG);
10445 /// \brief Handle lowering of 16-lane 32-bit floating point shuffles.
10446 static SDValue lowerV16F32VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10447 const X86Subtarget *Subtarget,
10448 SelectionDAG &DAG) {
10450 assert(V1.getSimpleValueType() == MVT::v16f32 && "Bad operand type!");
10451 assert(V2.getSimpleValueType() == MVT::v16f32 && "Bad operand type!");
10452 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10453 ArrayRef<int> Mask = SVOp->getMask();
10454 assert(Mask.size() == 16 && "Unexpected mask size for v16 shuffle!");
10456 // Use dedicated unpack instructions for masks that match their pattern.
10457 if (isShuffleEquivalent(V1, V2, Mask,
10458 {// First 128-bit lane.
10459 0, 16, 1, 17, 4, 20, 5, 21,
10460 // Second 128-bit lane.
10461 8, 24, 9, 25, 12, 28, 13, 29}))
10462 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v16f32, V1, V2);
10463 if (isShuffleEquivalent(V1, V2, Mask,
10464 {// First 128-bit lane.
10465 2, 18, 3, 19, 6, 22, 7, 23,
10466 // Second 128-bit lane.
10467 10, 26, 11, 27, 14, 30, 15, 31}))
10468 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v16f32, V1, V2);
10470 // FIXME: Implement direct support for this type!
10471 return splitAndLowerVectorShuffle(DL, MVT::v16f32, V1, V2, Mask, DAG);
10474 /// \brief Handle lowering of 8-lane 64-bit integer shuffles.
10475 static SDValue lowerV8I64VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10476 const X86Subtarget *Subtarget,
10477 SelectionDAG &DAG) {
10479 assert(V1.getSimpleValueType() == MVT::v8i64 && "Bad operand type!");
10480 assert(V2.getSimpleValueType() == MVT::v8i64 && "Bad operand type!");
10481 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10482 ArrayRef<int> Mask = SVOp->getMask();
10483 assert(Mask.size() == 8 && "Unexpected mask size for v8 shuffle!");
10485 // X86 has dedicated unpack instructions that can handle specific blend
10486 // operations: UNPCKH and UNPCKL.
10487 if (isShuffleEquivalent(V1, V2, Mask, {0, 8, 2, 10, 4, 12, 6, 14}))
10488 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v8i64, V1, V2);
10489 if (isShuffleEquivalent(V1, V2, Mask, {1, 9, 3, 11, 5, 13, 7, 15}))
10490 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v8i64, V1, V2);
10492 // FIXME: Implement direct support for this type!
10493 return splitAndLowerVectorShuffle(DL, MVT::v8i64, V1, V2, Mask, DAG);
10496 /// \brief Handle lowering of 16-lane 32-bit integer shuffles.
10497 static SDValue lowerV16I32VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10498 const X86Subtarget *Subtarget,
10499 SelectionDAG &DAG) {
10501 assert(V1.getSimpleValueType() == MVT::v16i32 && "Bad operand type!");
10502 assert(V2.getSimpleValueType() == MVT::v16i32 && "Bad operand type!");
10503 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10504 ArrayRef<int> Mask = SVOp->getMask();
10505 assert(Mask.size() == 16 && "Unexpected mask size for v16 shuffle!");
10507 // Use dedicated unpack instructions for masks that match their pattern.
10508 if (isShuffleEquivalent(V1, V2, Mask,
10509 {// First 128-bit lane.
10510 0, 16, 1, 17, 4, 20, 5, 21,
10511 // Second 128-bit lane.
10512 8, 24, 9, 25, 12, 28, 13, 29}))
10513 return DAG.getNode(X86ISD::UNPCKL, DL, MVT::v16i32, V1, V2);
10514 if (isShuffleEquivalent(V1, V2, Mask,
10515 {// First 128-bit lane.
10516 2, 18, 3, 19, 6, 22, 7, 23,
10517 // Second 128-bit lane.
10518 10, 26, 11, 27, 14, 30, 15, 31}))
10519 return DAG.getNode(X86ISD::UNPCKH, DL, MVT::v16i32, V1, V2);
10521 // FIXME: Implement direct support for this type!
10522 return splitAndLowerVectorShuffle(DL, MVT::v16i32, V1, V2, Mask, DAG);
10525 /// \brief Handle lowering of 32-lane 16-bit integer shuffles.
10526 static SDValue lowerV32I16VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10527 const X86Subtarget *Subtarget,
10528 SelectionDAG &DAG) {
10530 assert(V1.getSimpleValueType() == MVT::v32i16 && "Bad operand type!");
10531 assert(V2.getSimpleValueType() == MVT::v32i16 && "Bad operand type!");
10532 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10533 ArrayRef<int> Mask = SVOp->getMask();
10534 assert(Mask.size() == 32 && "Unexpected mask size for v32 shuffle!");
10535 assert(Subtarget->hasBWI() && "We can only lower v32i16 with AVX-512-BWI!");
10537 // FIXME: Implement direct support for this type!
10538 return splitAndLowerVectorShuffle(DL, MVT::v32i16, V1, V2, Mask, DAG);
10541 /// \brief Handle lowering of 64-lane 8-bit integer shuffles.
10542 static SDValue lowerV64I8VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10543 const X86Subtarget *Subtarget,
10544 SelectionDAG &DAG) {
10546 assert(V1.getSimpleValueType() == MVT::v64i8 && "Bad operand type!");
10547 assert(V2.getSimpleValueType() == MVT::v64i8 && "Bad operand type!");
10548 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10549 ArrayRef<int> Mask = SVOp->getMask();
10550 assert(Mask.size() == 64 && "Unexpected mask size for v64 shuffle!");
10551 assert(Subtarget->hasBWI() && "We can only lower v64i8 with AVX-512-BWI!");
10553 // FIXME: Implement direct support for this type!
10554 return splitAndLowerVectorShuffle(DL, MVT::v64i8, V1, V2, Mask, DAG);
10557 /// \brief High-level routine to lower various 512-bit x86 vector shuffles.
10559 /// This routine either breaks down the specific type of a 512-bit x86 vector
10560 /// shuffle or splits it into two 256-bit shuffles and fuses the results back
10561 /// together based on the available instructions.
10562 static SDValue lower512BitVectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10563 MVT VT, const X86Subtarget *Subtarget,
10564 SelectionDAG &DAG) {
10566 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10567 ArrayRef<int> Mask = SVOp->getMask();
10568 assert(Subtarget->hasAVX512() &&
10569 "Cannot lower 512-bit vectors w/ basic ISA!");
10571 // Check for being able to broadcast a single element.
10572 if (SDValue Broadcast =
10573 lowerVectorShuffleAsBroadcast(DL, VT, V1, Mask, Subtarget, DAG))
10576 // Dispatch to each element type for lowering. If we don't have supprot for
10577 // specific element type shuffles at 512 bits, immediately split them and
10578 // lower them. Each lowering routine of a given type is allowed to assume that
10579 // the requisite ISA extensions for that element type are available.
10580 switch (VT.SimpleTy) {
10582 return lowerV8F64VectorShuffle(Op, V1, V2, Subtarget, DAG);
10584 return lowerV16F32VectorShuffle(Op, V1, V2, Subtarget, DAG);
10586 return lowerV8I64VectorShuffle(Op, V1, V2, Subtarget, DAG);
10588 return lowerV16I32VectorShuffle(Op, V1, V2, Subtarget, DAG);
10590 if (Subtarget->hasBWI())
10591 return lowerV32I16VectorShuffle(Op, V1, V2, Subtarget, DAG);
10594 if (Subtarget->hasBWI())
10595 return lowerV64I8VectorShuffle(Op, V1, V2, Subtarget, DAG);
10599 llvm_unreachable("Not a valid 512-bit x86 vector type!");
10602 // Otherwise fall back on splitting.
10603 return splitAndLowerVectorShuffle(DL, VT, V1, V2, Mask, DAG);
10606 /// \brief Top-level lowering for x86 vector shuffles.
10608 /// This handles decomposition, canonicalization, and lowering of all x86
10609 /// vector shuffles. Most of the specific lowering strategies are encapsulated
10610 /// above in helper routines. The canonicalization attempts to widen shuffles
10611 /// to involve fewer lanes of wider elements, consolidate symmetric patterns
10612 /// s.t. only one of the two inputs needs to be tested, etc.
10613 static SDValue lowerVectorShuffle(SDValue Op, const X86Subtarget *Subtarget,
10614 SelectionDAG &DAG) {
10615 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10616 ArrayRef<int> Mask = SVOp->getMask();
10617 SDValue V1 = Op.getOperand(0);
10618 SDValue V2 = Op.getOperand(1);
10619 MVT VT = Op.getSimpleValueType();
10620 int NumElements = VT.getVectorNumElements();
10623 assert(VT.getSizeInBits() != 64 && "Can't lower MMX shuffles");
10625 bool V1IsUndef = V1.getOpcode() == ISD::UNDEF;
10626 bool V2IsUndef = V2.getOpcode() == ISD::UNDEF;
10627 if (V1IsUndef && V2IsUndef)
10628 return DAG.getUNDEF(VT);
10630 // When we create a shuffle node we put the UNDEF node to second operand,
10631 // but in some cases the first operand may be transformed to UNDEF.
10632 // In this case we should just commute the node.
10634 return DAG.getCommutedVectorShuffle(*SVOp);
10636 // Check for non-undef masks pointing at an undef vector and make the masks
10637 // undef as well. This makes it easier to match the shuffle based solely on
10641 if (M >= NumElements) {
10642 SmallVector<int, 8> NewMask(Mask.begin(), Mask.end());
10643 for (int &M : NewMask)
10644 if (M >= NumElements)
10646 return DAG.getVectorShuffle(VT, dl, V1, V2, NewMask);
10649 // We actually see shuffles that are entirely re-arrangements of a set of
10650 // zero inputs. This mostly happens while decomposing complex shuffles into
10651 // simple ones. Directly lower these as a buildvector of zeros.
10652 SmallBitVector Zeroable = computeZeroableShuffleElements(Mask, V1, V2);
10653 if (Zeroable.all())
10654 return getZeroVector(VT, Subtarget, DAG, dl);
10656 // Try to collapse shuffles into using a vector type with fewer elements but
10657 // wider element types. We cap this to not form integers or floating point
10658 // elements wider than 64 bits, but it might be interesting to form i128
10659 // integers to handle flipping the low and high halves of AVX 256-bit vectors.
10660 SmallVector<int, 16> WidenedMask;
10661 if (VT.getScalarSizeInBits() < 64 &&
10662 canWidenShuffleElements(Mask, WidenedMask)) {
10663 MVT NewEltVT = VT.isFloatingPoint()
10664 ? MVT::getFloatingPointVT(VT.getScalarSizeInBits() * 2)
10665 : MVT::getIntegerVT(VT.getScalarSizeInBits() * 2);
10666 MVT NewVT = MVT::getVectorVT(NewEltVT, VT.getVectorNumElements() / 2);
10667 // Make sure that the new vector type is legal. For example, v2f64 isn't
10669 if (DAG.getTargetLoweringInfo().isTypeLegal(NewVT)) {
10670 V1 = DAG.getBitcast(NewVT, V1);
10671 V2 = DAG.getBitcast(NewVT, V2);
10672 return DAG.getBitcast(
10673 VT, DAG.getVectorShuffle(NewVT, dl, V1, V2, WidenedMask));
10677 int NumV1Elements = 0, NumUndefElements = 0, NumV2Elements = 0;
10678 for (int M : SVOp->getMask())
10680 ++NumUndefElements;
10681 else if (M < NumElements)
10686 // Commute the shuffle as needed such that more elements come from V1 than
10687 // V2. This allows us to match the shuffle pattern strictly on how many
10688 // elements come from V1 without handling the symmetric cases.
10689 if (NumV2Elements > NumV1Elements)
10690 return DAG.getCommutedVectorShuffle(*SVOp);
10692 // When the number of V1 and V2 elements are the same, try to minimize the
10693 // number of uses of V2 in the low half of the vector. When that is tied,
10694 // ensure that the sum of indices for V1 is equal to or lower than the sum
10695 // indices for V2. When those are equal, try to ensure that the number of odd
10696 // indices for V1 is lower than the number of odd indices for V2.
10697 if (NumV1Elements == NumV2Elements) {
10698 int LowV1Elements = 0, LowV2Elements = 0;
10699 for (int M : SVOp->getMask().slice(0, NumElements / 2))
10700 if (M >= NumElements)
10704 if (LowV2Elements > LowV1Elements) {
10705 return DAG.getCommutedVectorShuffle(*SVOp);
10706 } else if (LowV2Elements == LowV1Elements) {
10707 int SumV1Indices = 0, SumV2Indices = 0;
10708 for (int i = 0, Size = SVOp->getMask().size(); i < Size; ++i)
10709 if (SVOp->getMask()[i] >= NumElements)
10711 else if (SVOp->getMask()[i] >= 0)
10713 if (SumV2Indices < SumV1Indices) {
10714 return DAG.getCommutedVectorShuffle(*SVOp);
10715 } else if (SumV2Indices == SumV1Indices) {
10716 int NumV1OddIndices = 0, NumV2OddIndices = 0;
10717 for (int i = 0, Size = SVOp->getMask().size(); i < Size; ++i)
10718 if (SVOp->getMask()[i] >= NumElements)
10719 NumV2OddIndices += i % 2;
10720 else if (SVOp->getMask()[i] >= 0)
10721 NumV1OddIndices += i % 2;
10722 if (NumV2OddIndices < NumV1OddIndices)
10723 return DAG.getCommutedVectorShuffle(*SVOp);
10728 // For each vector width, delegate to a specialized lowering routine.
10729 if (VT.getSizeInBits() == 128)
10730 return lower128BitVectorShuffle(Op, V1, V2, VT, Subtarget, DAG);
10732 if (VT.getSizeInBits() == 256)
10733 return lower256BitVectorShuffle(Op, V1, V2, VT, Subtarget, DAG);
10735 // Force AVX-512 vectors to be scalarized for now.
10736 // FIXME: Implement AVX-512 support!
10737 if (VT.getSizeInBits() == 512)
10738 return lower512BitVectorShuffle(Op, V1, V2, VT, Subtarget, DAG);
10740 llvm_unreachable("Unimplemented!");
10743 // This function assumes its argument is a BUILD_VECTOR of constants or
10744 // undef SDNodes. i.e: ISD::isBuildVectorOfConstantSDNodes(BuildVector) is
10746 static bool BUILD_VECTORtoBlendMask(BuildVectorSDNode *BuildVector,
10747 unsigned &MaskValue) {
10749 unsigned NumElems = BuildVector->getNumOperands();
10750 // There are 2 lanes if (NumElems > 8), and 1 lane otherwise.
10751 unsigned NumLanes = (NumElems - 1) / 8 + 1;
10752 unsigned NumElemsInLane = NumElems / NumLanes;
10754 // Blend for v16i16 should be symmetric for the both lanes.
10755 for (unsigned i = 0; i < NumElemsInLane; ++i) {
10756 SDValue EltCond = BuildVector->getOperand(i);
10757 SDValue SndLaneEltCond =
10758 (NumLanes == 2) ? BuildVector->getOperand(i + NumElemsInLane) : EltCond;
10760 int Lane1Cond = -1, Lane2Cond = -1;
10761 if (isa<ConstantSDNode>(EltCond))
10762 Lane1Cond = !isZero(EltCond);
10763 if (isa<ConstantSDNode>(SndLaneEltCond))
10764 Lane2Cond = !isZero(SndLaneEltCond);
10766 if (Lane1Cond == Lane2Cond || Lane2Cond < 0)
10767 // Lane1Cond != 0, means we want the first argument.
10768 // Lane1Cond == 0, means we want the second argument.
10769 // The encoding of this argument is 0 for the first argument, 1
10770 // for the second. Therefore, invert the condition.
10771 MaskValue |= !Lane1Cond << i;
10772 else if (Lane1Cond < 0)
10773 MaskValue |= !Lane2Cond << i;
10780 /// \brief Try to lower a VSELECT instruction to a vector shuffle.
10781 static SDValue lowerVSELECTtoVectorShuffle(SDValue Op,
10782 const X86Subtarget *Subtarget,
10783 SelectionDAG &DAG) {
10784 SDValue Cond = Op.getOperand(0);
10785 SDValue LHS = Op.getOperand(1);
10786 SDValue RHS = Op.getOperand(2);
10788 MVT VT = Op.getSimpleValueType();
10790 if (!ISD::isBuildVectorOfConstantSDNodes(Cond.getNode()))
10792 auto *CondBV = cast<BuildVectorSDNode>(Cond);
10794 // Only non-legal VSELECTs reach this lowering, convert those into generic
10795 // shuffles and re-use the shuffle lowering path for blends.
10796 SmallVector<int, 32> Mask;
10797 for (int i = 0, Size = VT.getVectorNumElements(); i < Size; ++i) {
10798 SDValue CondElt = CondBV->getOperand(i);
10800 isa<ConstantSDNode>(CondElt) ? i + (isZero(CondElt) ? Size : 0) : -1);
10802 return DAG.getVectorShuffle(VT, dl, LHS, RHS, Mask);
10805 SDValue X86TargetLowering::LowerVSELECT(SDValue Op, SelectionDAG &DAG) const {
10806 // A vselect where all conditions and data are constants can be optimized into
10807 // a single vector load by SelectionDAGLegalize::ExpandBUILD_VECTOR().
10808 if (ISD::isBuildVectorOfConstantSDNodes(Op.getOperand(0).getNode()) &&
10809 ISD::isBuildVectorOfConstantSDNodes(Op.getOperand(1).getNode()) &&
10810 ISD::isBuildVectorOfConstantSDNodes(Op.getOperand(2).getNode()))
10813 // Try to lower this to a blend-style vector shuffle. This can handle all
10814 // constant condition cases.
10815 if (SDValue BlendOp = lowerVSELECTtoVectorShuffle(Op, Subtarget, DAG))
10818 // Variable blends are only legal from SSE4.1 onward.
10819 if (!Subtarget->hasSSE41())
10822 // Only some types will be legal on some subtargets. If we can emit a legal
10823 // VSELECT-matching blend, return Op, and but if we need to expand, return
10825 switch (Op.getSimpleValueType().SimpleTy) {
10827 // Most of the vector types have blends past SSE4.1.
10831 // The byte blends for AVX vectors were introduced only in AVX2.
10832 if (Subtarget->hasAVX2())
10839 // AVX-512 BWI and VLX features support VSELECT with i16 elements.
10840 if (Subtarget->hasBWI() && Subtarget->hasVLX())
10843 // FIXME: We should custom lower this by fixing the condition and using i8
10849 static SDValue LowerEXTRACT_VECTOR_ELT_SSE4(SDValue Op, SelectionDAG &DAG) {
10850 MVT VT = Op.getSimpleValueType();
10853 if (!Op.getOperand(0).getSimpleValueType().is128BitVector())
10856 if (VT.getSizeInBits() == 8) {
10857 SDValue Extract = DAG.getNode(X86ISD::PEXTRB, dl, MVT::i32,
10858 Op.getOperand(0), Op.getOperand(1));
10859 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Extract,
10860 DAG.getValueType(VT));
10861 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
10864 if (VT.getSizeInBits() == 16) {
10865 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
10866 // If Idx is 0, it's cheaper to do a move instead of a pextrw.
10868 return DAG.getNode(
10869 ISD::TRUNCATE, dl, MVT::i16,
10870 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
10871 DAG.getBitcast(MVT::v4i32, Op.getOperand(0)),
10872 Op.getOperand(1)));
10873 SDValue Extract = DAG.getNode(X86ISD::PEXTRW, dl, MVT::i32,
10874 Op.getOperand(0), Op.getOperand(1));
10875 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Extract,
10876 DAG.getValueType(VT));
10877 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
10880 if (VT == MVT::f32) {
10881 // EXTRACTPS outputs to a GPR32 register which will require a movd to copy
10882 // the result back to FR32 register. It's only worth matching if the
10883 // result has a single use which is a store or a bitcast to i32. And in
10884 // the case of a store, it's not worth it if the index is a constant 0,
10885 // because a MOVSSmr can be used instead, which is smaller and faster.
10886 if (!Op.hasOneUse())
10888 SDNode *User = *Op.getNode()->use_begin();
10889 if ((User->getOpcode() != ISD::STORE ||
10890 (isa<ConstantSDNode>(Op.getOperand(1)) &&
10891 cast<ConstantSDNode>(Op.getOperand(1))->isNullValue())) &&
10892 (User->getOpcode() != ISD::BITCAST ||
10893 User->getValueType(0) != MVT::i32))
10895 SDValue Extract = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
10896 DAG.getBitcast(MVT::v4i32, Op.getOperand(0)),
10898 return DAG.getBitcast(MVT::f32, Extract);
10901 if (VT == MVT::i32 || VT == MVT::i64) {
10902 // ExtractPS/pextrq works with constant index.
10903 if (isa<ConstantSDNode>(Op.getOperand(1)))
10909 /// Extract one bit from mask vector, like v16i1 or v8i1.
10910 /// AVX-512 feature.
10912 X86TargetLowering::ExtractBitFromMaskVector(SDValue Op, SelectionDAG &DAG) const {
10913 SDValue Vec = Op.getOperand(0);
10915 MVT VecVT = Vec.getSimpleValueType();
10916 SDValue Idx = Op.getOperand(1);
10917 MVT EltVT = Op.getSimpleValueType();
10919 assert((EltVT == MVT::i1) && "Unexpected operands in ExtractBitFromMaskVector");
10920 assert((VecVT.getVectorNumElements() <= 16 || Subtarget->hasBWI()) &&
10921 "Unexpected vector type in ExtractBitFromMaskVector");
10923 // variable index can't be handled in mask registers,
10924 // extend vector to VR512
10925 if (!isa<ConstantSDNode>(Idx)) {
10926 MVT ExtVT = (VecVT == MVT::v8i1 ? MVT::v8i64 : MVT::v16i32);
10927 SDValue Ext = DAG.getNode(ISD::ZERO_EXTEND, dl, ExtVT, Vec);
10928 SDValue Elt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl,
10929 ExtVT.getVectorElementType(), Ext, Idx);
10930 return DAG.getNode(ISD::TRUNCATE, dl, EltVT, Elt);
10933 unsigned IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
10934 const TargetRegisterClass* rc = getRegClassFor(VecVT);
10935 if (!Subtarget->hasDQI() && (VecVT.getVectorNumElements() <= 8))
10936 rc = getRegClassFor(MVT::v16i1);
10937 unsigned MaxSift = rc->getSize()*8 - 1;
10938 Vec = DAG.getNode(X86ISD::VSHLI, dl, VecVT, Vec,
10939 DAG.getConstant(MaxSift - IdxVal, dl, MVT::i8));
10940 Vec = DAG.getNode(X86ISD::VSRLI, dl, VecVT, Vec,
10941 DAG.getConstant(MaxSift, dl, MVT::i8));
10942 return DAG.getNode(X86ISD::VEXTRACT, dl, MVT::i1, Vec,
10943 DAG.getIntPtrConstant(0, dl));
10947 X86TargetLowering::LowerEXTRACT_VECTOR_ELT(SDValue Op,
10948 SelectionDAG &DAG) const {
10950 SDValue Vec = Op.getOperand(0);
10951 MVT VecVT = Vec.getSimpleValueType();
10952 SDValue Idx = Op.getOperand(1);
10954 if (Op.getSimpleValueType() == MVT::i1)
10955 return ExtractBitFromMaskVector(Op, DAG);
10957 if (!isa<ConstantSDNode>(Idx)) {
10958 if (VecVT.is512BitVector() ||
10959 (VecVT.is256BitVector() && Subtarget->hasInt256() &&
10960 VecVT.getVectorElementType().getSizeInBits() == 32)) {
10963 MVT::getIntegerVT(VecVT.getVectorElementType().getSizeInBits());
10964 MVT MaskVT = MVT::getVectorVT(MaskEltVT, VecVT.getSizeInBits() /
10965 MaskEltVT.getSizeInBits());
10967 Idx = DAG.getZExtOrTrunc(Idx, dl, MaskEltVT);
10968 auto PtrVT = getPointerTy(DAG.getDataLayout());
10969 SDValue Mask = DAG.getNode(X86ISD::VINSERT, dl, MaskVT,
10970 getZeroVector(MaskVT, Subtarget, DAG, dl), Idx,
10971 DAG.getConstant(0, dl, PtrVT));
10972 SDValue Perm = DAG.getNode(X86ISD::VPERMV, dl, VecVT, Mask, Vec);
10973 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, Op.getValueType(), Perm,
10974 DAG.getConstant(0, dl, PtrVT));
10979 // If this is a 256-bit vector result, first extract the 128-bit vector and
10980 // then extract the element from the 128-bit vector.
10981 if (VecVT.is256BitVector() || VecVT.is512BitVector()) {
10983 unsigned IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
10984 // Get the 128-bit vector.
10985 Vec = Extract128BitVector(Vec, IdxVal, DAG, dl);
10986 MVT EltVT = VecVT.getVectorElementType();
10988 unsigned ElemsPerChunk = 128 / EltVT.getSizeInBits();
10990 //if (IdxVal >= NumElems/2)
10991 // IdxVal -= NumElems/2;
10992 IdxVal -= (IdxVal/ElemsPerChunk)*ElemsPerChunk;
10993 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, Op.getValueType(), Vec,
10994 DAG.getConstant(IdxVal, dl, MVT::i32));
10997 assert(VecVT.is128BitVector() && "Unexpected vector length");
10999 if (Subtarget->hasSSE41())
11000 if (SDValue Res = LowerEXTRACT_VECTOR_ELT_SSE4(Op, DAG))
11003 MVT VT = Op.getSimpleValueType();
11004 // TODO: handle v16i8.
11005 if (VT.getSizeInBits() == 16) {
11006 SDValue Vec = Op.getOperand(0);
11007 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
11009 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i16,
11010 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
11011 DAG.getBitcast(MVT::v4i32, Vec),
11012 Op.getOperand(1)));
11013 // Transform it so it match pextrw which produces a 32-bit result.
11014 MVT EltVT = MVT::i32;
11015 SDValue Extract = DAG.getNode(X86ISD::PEXTRW, dl, EltVT,
11016 Op.getOperand(0), Op.getOperand(1));
11017 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, EltVT, Extract,
11018 DAG.getValueType(VT));
11019 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
11022 if (VT.getSizeInBits() == 32) {
11023 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
11027 // SHUFPS the element to the lowest double word, then movss.
11028 int Mask[4] = { static_cast<int>(Idx), -1, -1, -1 };
11029 MVT VVT = Op.getOperand(0).getSimpleValueType();
11030 SDValue Vec = DAG.getVectorShuffle(VVT, dl, Op.getOperand(0),
11031 DAG.getUNDEF(VVT), Mask);
11032 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, Vec,
11033 DAG.getIntPtrConstant(0, dl));
11036 if (VT.getSizeInBits() == 64) {
11037 // FIXME: .td only matches this for <2 x f64>, not <2 x i64> on 32b
11038 // FIXME: seems like this should be unnecessary if mov{h,l}pd were taught
11039 // to match extract_elt for f64.
11040 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
11044 // UNPCKHPD the element to the lowest double word, then movsd.
11045 // Note if the lower 64 bits of the result of the UNPCKHPD is then stored
11046 // to a f64mem, the whole operation is folded into a single MOVHPDmr.
11047 int Mask[2] = { 1, -1 };
11048 MVT VVT = Op.getOperand(0).getSimpleValueType();
11049 SDValue Vec = DAG.getVectorShuffle(VVT, dl, Op.getOperand(0),
11050 DAG.getUNDEF(VVT), Mask);
11051 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, Vec,
11052 DAG.getIntPtrConstant(0, dl));
11058 /// Insert one bit to mask vector, like v16i1 or v8i1.
11059 /// AVX-512 feature.
11061 X86TargetLowering::InsertBitToMaskVector(SDValue Op, SelectionDAG &DAG) const {
11063 SDValue Vec = Op.getOperand(0);
11064 SDValue Elt = Op.getOperand(1);
11065 SDValue Idx = Op.getOperand(2);
11066 MVT VecVT = Vec.getSimpleValueType();
11068 if (!isa<ConstantSDNode>(Idx)) {
11069 // Non constant index. Extend source and destination,
11070 // insert element and then truncate the result.
11071 MVT ExtVecVT = (VecVT == MVT::v8i1 ? MVT::v8i64 : MVT::v16i32);
11072 MVT ExtEltVT = (VecVT == MVT::v8i1 ? MVT::i64 : MVT::i32);
11073 SDValue ExtOp = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, ExtVecVT,
11074 DAG.getNode(ISD::ZERO_EXTEND, dl, ExtVecVT, Vec),
11075 DAG.getNode(ISD::ZERO_EXTEND, dl, ExtEltVT, Elt), Idx);
11076 return DAG.getNode(ISD::TRUNCATE, dl, VecVT, ExtOp);
11079 unsigned IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
11080 SDValue EltInVec = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VecVT, Elt);
11082 EltInVec = DAG.getNode(X86ISD::VSHLI, dl, VecVT, EltInVec,
11083 DAG.getConstant(IdxVal, dl, MVT::i8));
11084 if (Vec.getOpcode() == ISD::UNDEF)
11086 return DAG.getNode(ISD::OR, dl, VecVT, Vec, EltInVec);
11089 SDValue X86TargetLowering::LowerINSERT_VECTOR_ELT(SDValue Op,
11090 SelectionDAG &DAG) const {
11091 MVT VT = Op.getSimpleValueType();
11092 MVT EltVT = VT.getVectorElementType();
11094 if (EltVT == MVT::i1)
11095 return InsertBitToMaskVector(Op, DAG);
11098 SDValue N0 = Op.getOperand(0);
11099 SDValue N1 = Op.getOperand(1);
11100 SDValue N2 = Op.getOperand(2);
11101 if (!isa<ConstantSDNode>(N2))
11103 auto *N2C = cast<ConstantSDNode>(N2);
11104 unsigned IdxVal = N2C->getZExtValue();
11106 // If the vector is wider than 128 bits, extract the 128-bit subvector, insert
11107 // into that, and then insert the subvector back into the result.
11108 if (VT.is256BitVector() || VT.is512BitVector()) {
11109 // With a 256-bit vector, we can insert into the zero element efficiently
11110 // using a blend if we have AVX or AVX2 and the right data type.
11111 if (VT.is256BitVector() && IdxVal == 0) {
11112 // TODO: It is worthwhile to cast integer to floating point and back
11113 // and incur a domain crossing penalty if that's what we'll end up
11114 // doing anyway after extracting to a 128-bit vector.
11115 if ((Subtarget->hasAVX() && (EltVT == MVT::f64 || EltVT == MVT::f32)) ||
11116 (Subtarget->hasAVX2() && EltVT == MVT::i32)) {
11117 SDValue N1Vec = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, N1);
11118 N2 = DAG.getIntPtrConstant(1, dl);
11119 return DAG.getNode(X86ISD::BLENDI, dl, VT, N0, N1Vec, N2);
11123 // Get the desired 128-bit vector chunk.
11124 SDValue V = Extract128BitVector(N0, IdxVal, DAG, dl);
11126 // Insert the element into the desired chunk.
11127 unsigned NumEltsIn128 = 128 / EltVT.getSizeInBits();
11128 unsigned IdxIn128 = IdxVal - (IdxVal / NumEltsIn128) * NumEltsIn128;
11130 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, V.getValueType(), V, N1,
11131 DAG.getConstant(IdxIn128, dl, MVT::i32));
11133 // Insert the changed part back into the bigger vector
11134 return Insert128BitVector(N0, V, IdxVal, DAG, dl);
11136 assert(VT.is128BitVector() && "Only 128-bit vector types should be left!");
11138 if (Subtarget->hasSSE41()) {
11139 if (EltVT.getSizeInBits() == 8 || EltVT.getSizeInBits() == 16) {
11141 if (VT == MVT::v8i16) {
11142 Opc = X86ISD::PINSRW;
11144 assert(VT == MVT::v16i8);
11145 Opc = X86ISD::PINSRB;
11148 // Transform it so it match pinsr{b,w} which expects a GR32 as its second
11150 if (N1.getValueType() != MVT::i32)
11151 N1 = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, N1);
11152 if (N2.getValueType() != MVT::i32)
11153 N2 = DAG.getIntPtrConstant(IdxVal, dl);
11154 return DAG.getNode(Opc, dl, VT, N0, N1, N2);
11157 if (EltVT == MVT::f32) {
11158 // Bits [7:6] of the constant are the source select. This will always be
11159 // zero here. The DAG Combiner may combine an extract_elt index into
11160 // these bits. For example (insert (extract, 3), 2) could be matched by
11161 // putting the '3' into bits [7:6] of X86ISD::INSERTPS.
11162 // Bits [5:4] of the constant are the destination select. This is the
11163 // value of the incoming immediate.
11164 // Bits [3:0] of the constant are the zero mask. The DAG Combiner may
11165 // combine either bitwise AND or insert of float 0.0 to set these bits.
11167 bool MinSize = DAG.getMachineFunction().getFunction()->optForMinSize();
11168 if (IdxVal == 0 && (!MinSize || !MayFoldLoad(N1))) {
11169 // If this is an insertion of 32-bits into the low 32-bits of
11170 // a vector, we prefer to generate a blend with immediate rather
11171 // than an insertps. Blends are simpler operations in hardware and so
11172 // will always have equal or better performance than insertps.
11173 // But if optimizing for size and there's a load folding opportunity,
11174 // generate insertps because blendps does not have a 32-bit memory
11176 N2 = DAG.getIntPtrConstant(1, dl);
11177 N1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4f32, N1);
11178 return DAG.getNode(X86ISD::BLENDI, dl, VT, N0, N1, N2);
11180 N2 = DAG.getIntPtrConstant(IdxVal << 4, dl);
11181 // Create this as a scalar to vector..
11182 N1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4f32, N1);
11183 return DAG.getNode(X86ISD::INSERTPS, dl, VT, N0, N1, N2);
11186 if (EltVT == MVT::i32 || EltVT == MVT::i64) {
11187 // PINSR* works with constant index.
11192 if (EltVT == MVT::i8)
11195 if (EltVT.getSizeInBits() == 16) {
11196 // Transform it so it match pinsrw which expects a 16-bit value in a GR32
11197 // as its second argument.
11198 if (N1.getValueType() != MVT::i32)
11199 N1 = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, N1);
11200 if (N2.getValueType() != MVT::i32)
11201 N2 = DAG.getIntPtrConstant(IdxVal, dl);
11202 return DAG.getNode(X86ISD::PINSRW, dl, VT, N0, N1, N2);
11207 static SDValue LowerSCALAR_TO_VECTOR(SDValue Op, SelectionDAG &DAG) {
11209 MVT OpVT = Op.getSimpleValueType();
11211 // If this is a 256-bit vector result, first insert into a 128-bit
11212 // vector and then insert into the 256-bit vector.
11213 if (!OpVT.is128BitVector()) {
11214 // Insert into a 128-bit vector.
11215 unsigned SizeFactor = OpVT.getSizeInBits()/128;
11216 MVT VT128 = MVT::getVectorVT(OpVT.getVectorElementType(),
11217 OpVT.getVectorNumElements() / SizeFactor);
11219 Op = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT128, Op.getOperand(0));
11221 // Insert the 128-bit vector.
11222 return Insert128BitVector(DAG.getUNDEF(OpVT), Op, 0, DAG, dl);
11225 if (OpVT == MVT::v1i64 &&
11226 Op.getOperand(0).getValueType() == MVT::i64)
11227 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v1i64, Op.getOperand(0));
11229 SDValue AnyExt = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, Op.getOperand(0));
11230 assert(OpVT.is128BitVector() && "Expected an SSE type!");
11231 return DAG.getBitcast(
11232 OpVT, DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32, AnyExt));
11235 // Lower a node with an EXTRACT_SUBVECTOR opcode. This may result in
11236 // a simple subregister reference or explicit instructions to grab
11237 // upper bits of a vector.
11238 static SDValue LowerEXTRACT_SUBVECTOR(SDValue Op, const X86Subtarget *Subtarget,
11239 SelectionDAG &DAG) {
11241 SDValue In = Op.getOperand(0);
11242 SDValue Idx = Op.getOperand(1);
11243 unsigned IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
11244 MVT ResVT = Op.getSimpleValueType();
11245 MVT InVT = In.getSimpleValueType();
11247 if (Subtarget->hasFp256()) {
11248 if (ResVT.is128BitVector() &&
11249 (InVT.is256BitVector() || InVT.is512BitVector()) &&
11250 isa<ConstantSDNode>(Idx)) {
11251 return Extract128BitVector(In, IdxVal, DAG, dl);
11253 if (ResVT.is256BitVector() && InVT.is512BitVector() &&
11254 isa<ConstantSDNode>(Idx)) {
11255 return Extract256BitVector(In, IdxVal, DAG, dl);
11261 // Lower a node with an INSERT_SUBVECTOR opcode. This may result in a
11262 // simple superregister reference or explicit instructions to insert
11263 // the upper bits of a vector.
11264 static SDValue LowerINSERT_SUBVECTOR(SDValue Op, const X86Subtarget *Subtarget,
11265 SelectionDAG &DAG) {
11266 if (!Subtarget->hasAVX())
11270 SDValue Vec = Op.getOperand(0);
11271 SDValue SubVec = Op.getOperand(1);
11272 SDValue Idx = Op.getOperand(2);
11274 if (!isa<ConstantSDNode>(Idx))
11277 unsigned IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
11278 MVT OpVT = Op.getSimpleValueType();
11279 MVT SubVecVT = SubVec.getSimpleValueType();
11281 // Fold two 16-byte subvector loads into one 32-byte load:
11282 // (insert_subvector (insert_subvector undef, (load addr), 0),
11283 // (load addr + 16), Elts/2)
11285 if ((IdxVal == OpVT.getVectorNumElements() / 2) &&
11286 Vec.getOpcode() == ISD::INSERT_SUBVECTOR &&
11287 OpVT.is256BitVector() && SubVecVT.is128BitVector()) {
11288 auto *Idx2 = dyn_cast<ConstantSDNode>(Vec.getOperand(2));
11289 if (Idx2 && Idx2->getZExtValue() == 0) {
11290 SDValue SubVec2 = Vec.getOperand(1);
11291 // If needed, look through a bitcast to get to the load.
11292 if (SubVec2.getNode() && SubVec2.getOpcode() == ISD::BITCAST)
11293 SubVec2 = SubVec2.getOperand(0);
11295 if (auto *FirstLd = dyn_cast<LoadSDNode>(SubVec2)) {
11297 unsigned Alignment = FirstLd->getAlignment();
11298 unsigned AS = FirstLd->getAddressSpace();
11299 const X86TargetLowering *TLI = Subtarget->getTargetLowering();
11300 if (TLI->allowsMemoryAccess(*DAG.getContext(), DAG.getDataLayout(),
11301 OpVT, AS, Alignment, &Fast) && Fast) {
11302 SDValue Ops[] = { SubVec2, SubVec };
11303 if (SDValue Ld = EltsFromConsecutiveLoads(OpVT, Ops, dl, DAG, false))
11310 if ((OpVT.is256BitVector() || OpVT.is512BitVector()) &&
11311 SubVecVT.is128BitVector())
11312 return Insert128BitVector(Vec, SubVec, IdxVal, DAG, dl);
11314 if (OpVT.is512BitVector() && SubVecVT.is256BitVector())
11315 return Insert256BitVector(Vec, SubVec, IdxVal, DAG, dl);
11317 if (OpVT.getVectorElementType() == MVT::i1) {
11318 if (IdxVal == 0 && Vec.getOpcode() == ISD::UNDEF) // the operation is legal
11320 SDValue ZeroIdx = DAG.getIntPtrConstant(0, dl);
11321 SDValue Undef = DAG.getUNDEF(OpVT);
11322 unsigned NumElems = OpVT.getVectorNumElements();
11323 SDValue ShiftBits = DAG.getConstant(NumElems/2, dl, MVT::i8);
11325 if (IdxVal == OpVT.getVectorNumElements() / 2) {
11326 // Zero upper bits of the Vec
11327 Vec = DAG.getNode(X86ISD::VSHLI, dl, OpVT, Vec, ShiftBits);
11328 Vec = DAG.getNode(X86ISD::VSRLI, dl, OpVT, Vec, ShiftBits);
11330 SDValue Vec2 = DAG.getNode(ISD::INSERT_SUBVECTOR, dl, OpVT, Undef,
11332 Vec2 = DAG.getNode(X86ISD::VSHLI, dl, OpVT, Vec2, ShiftBits);
11333 return DAG.getNode(ISD::OR, dl, OpVT, Vec, Vec2);
11336 SDValue Vec2 = DAG.getNode(ISD::INSERT_SUBVECTOR, dl, OpVT, Undef,
11338 // Zero upper bits of the Vec2
11339 Vec2 = DAG.getNode(X86ISD::VSHLI, dl, OpVT, Vec2, ShiftBits);
11340 Vec2 = DAG.getNode(X86ISD::VSRLI, dl, OpVT, Vec2, ShiftBits);
11341 // Zero lower bits of the Vec
11342 Vec = DAG.getNode(X86ISD::VSRLI, dl, OpVT, Vec, ShiftBits);
11343 Vec = DAG.getNode(X86ISD::VSHLI, dl, OpVT, Vec, ShiftBits);
11344 // Merge them together
11345 return DAG.getNode(ISD::OR, dl, OpVT, Vec, Vec2);
11351 // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
11352 // their target countpart wrapped in the X86ISD::Wrapper node. Suppose N is
11353 // one of the above mentioned nodes. It has to be wrapped because otherwise
11354 // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
11355 // be used to form addressing mode. These wrapped nodes will be selected
11358 X86TargetLowering::LowerConstantPool(SDValue Op, SelectionDAG &DAG) const {
11359 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
11361 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
11362 // global base reg.
11363 unsigned char OpFlag = 0;
11364 unsigned WrapperKind = X86ISD::Wrapper;
11365 CodeModel::Model M = DAG.getTarget().getCodeModel();
11367 if (Subtarget->isPICStyleRIPRel() &&
11368 (M == CodeModel::Small || M == CodeModel::Kernel))
11369 WrapperKind = X86ISD::WrapperRIP;
11370 else if (Subtarget->isPICStyleGOT())
11371 OpFlag = X86II::MO_GOTOFF;
11372 else if (Subtarget->isPICStyleStubPIC())
11373 OpFlag = X86II::MO_PIC_BASE_OFFSET;
11375 auto PtrVT = getPointerTy(DAG.getDataLayout());
11376 SDValue Result = DAG.getTargetConstantPool(
11377 CP->getConstVal(), PtrVT, CP->getAlignment(), CP->getOffset(), OpFlag);
11379 Result = DAG.getNode(WrapperKind, DL, PtrVT, Result);
11380 // With PIC, the address is actually $g + Offset.
11383 DAG.getNode(ISD::ADD, DL, PtrVT,
11384 DAG.getNode(X86ISD::GlobalBaseReg, SDLoc(), PtrVT), Result);
11390 SDValue X86TargetLowering::LowerJumpTable(SDValue Op, SelectionDAG &DAG) const {
11391 JumpTableSDNode *JT = cast<JumpTableSDNode>(Op);
11393 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
11394 // global base reg.
11395 unsigned char OpFlag = 0;
11396 unsigned WrapperKind = X86ISD::Wrapper;
11397 CodeModel::Model M = DAG.getTarget().getCodeModel();
11399 if (Subtarget->isPICStyleRIPRel() &&
11400 (M == CodeModel::Small || M == CodeModel::Kernel))
11401 WrapperKind = X86ISD::WrapperRIP;
11402 else if (Subtarget->isPICStyleGOT())
11403 OpFlag = X86II::MO_GOTOFF;
11404 else if (Subtarget->isPICStyleStubPIC())
11405 OpFlag = X86II::MO_PIC_BASE_OFFSET;
11407 auto PtrVT = getPointerTy(DAG.getDataLayout());
11408 SDValue Result = DAG.getTargetJumpTable(JT->getIndex(), PtrVT, OpFlag);
11410 Result = DAG.getNode(WrapperKind, DL, PtrVT, Result);
11412 // With PIC, the address is actually $g + Offset.
11415 DAG.getNode(ISD::ADD, DL, PtrVT,
11416 DAG.getNode(X86ISD::GlobalBaseReg, SDLoc(), PtrVT), Result);
11422 X86TargetLowering::LowerExternalSymbol(SDValue Op, SelectionDAG &DAG) const {
11423 const char *Sym = cast<ExternalSymbolSDNode>(Op)->getSymbol();
11425 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
11426 // global base reg.
11427 unsigned char OpFlag = 0;
11428 unsigned WrapperKind = X86ISD::Wrapper;
11429 CodeModel::Model M = DAG.getTarget().getCodeModel();
11431 if (Subtarget->isPICStyleRIPRel() &&
11432 (M == CodeModel::Small || M == CodeModel::Kernel)) {
11433 if (Subtarget->isTargetDarwin() || Subtarget->isTargetELF())
11434 OpFlag = X86II::MO_GOTPCREL;
11435 WrapperKind = X86ISD::WrapperRIP;
11436 } else if (Subtarget->isPICStyleGOT()) {
11437 OpFlag = X86II::MO_GOT;
11438 } else if (Subtarget->isPICStyleStubPIC()) {
11439 OpFlag = X86II::MO_DARWIN_NONLAZY_PIC_BASE;
11440 } else if (Subtarget->isPICStyleStubNoDynamic()) {
11441 OpFlag = X86II::MO_DARWIN_NONLAZY;
11444 auto PtrVT = getPointerTy(DAG.getDataLayout());
11445 SDValue Result = DAG.getTargetExternalSymbol(Sym, PtrVT, OpFlag);
11448 Result = DAG.getNode(WrapperKind, DL, PtrVT, Result);
11450 // With PIC, the address is actually $g + Offset.
11451 if (DAG.getTarget().getRelocationModel() == Reloc::PIC_ &&
11452 !Subtarget->is64Bit()) {
11454 DAG.getNode(ISD::ADD, DL, PtrVT,
11455 DAG.getNode(X86ISD::GlobalBaseReg, SDLoc(), PtrVT), Result);
11458 // For symbols that require a load from a stub to get the address, emit the
11460 if (isGlobalStubReference(OpFlag))
11461 Result = DAG.getLoad(PtrVT, DL, DAG.getEntryNode(), Result,
11462 MachinePointerInfo::getGOT(DAG.getMachineFunction()),
11463 false, false, false, 0);
11469 X86TargetLowering::LowerBlockAddress(SDValue Op, SelectionDAG &DAG) const {
11470 // Create the TargetBlockAddressAddress node.
11471 unsigned char OpFlags =
11472 Subtarget->ClassifyBlockAddressReference();
11473 CodeModel::Model M = DAG.getTarget().getCodeModel();
11474 const BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress();
11475 int64_t Offset = cast<BlockAddressSDNode>(Op)->getOffset();
11477 auto PtrVT = getPointerTy(DAG.getDataLayout());
11478 SDValue Result = DAG.getTargetBlockAddress(BA, PtrVT, Offset, OpFlags);
11480 if (Subtarget->isPICStyleRIPRel() &&
11481 (M == CodeModel::Small || M == CodeModel::Kernel))
11482 Result = DAG.getNode(X86ISD::WrapperRIP, dl, PtrVT, Result);
11484 Result = DAG.getNode(X86ISD::Wrapper, dl, PtrVT, Result);
11486 // With PIC, the address is actually $g + Offset.
11487 if (isGlobalRelativeToPICBase(OpFlags)) {
11488 Result = DAG.getNode(ISD::ADD, dl, PtrVT,
11489 DAG.getNode(X86ISD::GlobalBaseReg, dl, PtrVT), Result);
11496 X86TargetLowering::LowerGlobalAddress(const GlobalValue *GV, SDLoc dl,
11497 int64_t Offset, SelectionDAG &DAG) const {
11498 // Create the TargetGlobalAddress node, folding in the constant
11499 // offset if it is legal.
11500 unsigned char OpFlags =
11501 Subtarget->ClassifyGlobalReference(GV, DAG.getTarget());
11502 CodeModel::Model M = DAG.getTarget().getCodeModel();
11503 auto PtrVT = getPointerTy(DAG.getDataLayout());
11505 if (OpFlags == X86II::MO_NO_FLAG &&
11506 X86::isOffsetSuitableForCodeModel(Offset, M)) {
11507 // A direct static reference to a global.
11508 Result = DAG.getTargetGlobalAddress(GV, dl, PtrVT, Offset);
11511 Result = DAG.getTargetGlobalAddress(GV, dl, PtrVT, 0, OpFlags);
11514 if (Subtarget->isPICStyleRIPRel() &&
11515 (M == CodeModel::Small || M == CodeModel::Kernel))
11516 Result = DAG.getNode(X86ISD::WrapperRIP, dl, PtrVT, Result);
11518 Result = DAG.getNode(X86ISD::Wrapper, dl, PtrVT, Result);
11520 // With PIC, the address is actually $g + Offset.
11521 if (isGlobalRelativeToPICBase(OpFlags)) {
11522 Result = DAG.getNode(ISD::ADD, dl, PtrVT,
11523 DAG.getNode(X86ISD::GlobalBaseReg, dl, PtrVT), Result);
11526 // For globals that require a load from a stub to get the address, emit the
11528 if (isGlobalStubReference(OpFlags))
11529 Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Result,
11530 MachinePointerInfo::getGOT(DAG.getMachineFunction()),
11531 false, false, false, 0);
11533 // If there was a non-zero offset that we didn't fold, create an explicit
11534 // addition for it.
11536 Result = DAG.getNode(ISD::ADD, dl, PtrVT, Result,
11537 DAG.getConstant(Offset, dl, PtrVT));
11543 X86TargetLowering::LowerGlobalAddress(SDValue Op, SelectionDAG &DAG) const {
11544 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
11545 int64_t Offset = cast<GlobalAddressSDNode>(Op)->getOffset();
11546 return LowerGlobalAddress(GV, SDLoc(Op), Offset, DAG);
11550 GetTLSADDR(SelectionDAG &DAG, SDValue Chain, GlobalAddressSDNode *GA,
11551 SDValue *InFlag, const EVT PtrVT, unsigned ReturnReg,
11552 unsigned char OperandFlags, bool LocalDynamic = false) {
11553 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
11554 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
11556 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(), dl,
11557 GA->getValueType(0),
11561 X86ISD::NodeType CallType = LocalDynamic ? X86ISD::TLSBASEADDR
11565 SDValue Ops[] = { Chain, TGA, *InFlag };
11566 Chain = DAG.getNode(CallType, dl, NodeTys, Ops);
11568 SDValue Ops[] = { Chain, TGA };
11569 Chain = DAG.getNode(CallType, dl, NodeTys, Ops);
11572 // TLSADDR will be codegen'ed as call. Inform MFI that function has calls.
11573 MFI->setAdjustsStack(true);
11574 MFI->setHasCalls(true);
11576 SDValue Flag = Chain.getValue(1);
11577 return DAG.getCopyFromReg(Chain, dl, ReturnReg, PtrVT, Flag);
11580 // Lower ISD::GlobalTLSAddress using the "general dynamic" model, 32 bit
11582 LowerToTLSGeneralDynamicModel32(GlobalAddressSDNode *GA, SelectionDAG &DAG,
11585 SDLoc dl(GA); // ? function entry point might be better
11586 SDValue Chain = DAG.getCopyToReg(DAG.getEntryNode(), dl, X86::EBX,
11587 DAG.getNode(X86ISD::GlobalBaseReg,
11588 SDLoc(), PtrVT), InFlag);
11589 InFlag = Chain.getValue(1);
11591 return GetTLSADDR(DAG, Chain, GA, &InFlag, PtrVT, X86::EAX, X86II::MO_TLSGD);
11594 // Lower ISD::GlobalTLSAddress using the "general dynamic" model, 64 bit
11596 LowerToTLSGeneralDynamicModel64(GlobalAddressSDNode *GA, SelectionDAG &DAG,
11598 return GetTLSADDR(DAG, DAG.getEntryNode(), GA, nullptr, PtrVT,
11599 X86::RAX, X86II::MO_TLSGD);
11602 static SDValue LowerToTLSLocalDynamicModel(GlobalAddressSDNode *GA,
11608 // Get the start address of the TLS block for this module.
11609 X86MachineFunctionInfo* MFI = DAG.getMachineFunction()
11610 .getInfo<X86MachineFunctionInfo>();
11611 MFI->incNumLocalDynamicTLSAccesses();
11615 Base = GetTLSADDR(DAG, DAG.getEntryNode(), GA, nullptr, PtrVT, X86::RAX,
11616 X86II::MO_TLSLD, /*LocalDynamic=*/true);
11619 SDValue Chain = DAG.getCopyToReg(DAG.getEntryNode(), dl, X86::EBX,
11620 DAG.getNode(X86ISD::GlobalBaseReg, SDLoc(), PtrVT), InFlag);
11621 InFlag = Chain.getValue(1);
11622 Base = GetTLSADDR(DAG, Chain, GA, &InFlag, PtrVT, X86::EAX,
11623 X86II::MO_TLSLDM, /*LocalDynamic=*/true);
11626 // Note: the CleanupLocalDynamicTLSPass will remove redundant computations
11630 unsigned char OperandFlags = X86II::MO_DTPOFF;
11631 unsigned WrapperKind = X86ISD::Wrapper;
11632 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(), dl,
11633 GA->getValueType(0),
11634 GA->getOffset(), OperandFlags);
11635 SDValue Offset = DAG.getNode(WrapperKind, dl, PtrVT, TGA);
11637 // Add x@dtpoff with the base.
11638 return DAG.getNode(ISD::ADD, dl, PtrVT, Offset, Base);
11641 // Lower ISD::GlobalTLSAddress using the "initial exec" or "local exec" model.
11642 static SDValue LowerToTLSExecModel(GlobalAddressSDNode *GA, SelectionDAG &DAG,
11643 const EVT PtrVT, TLSModel::Model model,
11644 bool is64Bit, bool isPIC) {
11647 // Get the Thread Pointer, which is %gs:0 (32-bit) or %fs:0 (64-bit).
11648 Value *Ptr = Constant::getNullValue(Type::getInt8PtrTy(*DAG.getContext(),
11649 is64Bit ? 257 : 256));
11651 SDValue ThreadPointer =
11652 DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), DAG.getIntPtrConstant(0, dl),
11653 MachinePointerInfo(Ptr), false, false, false, 0);
11655 unsigned char OperandFlags = 0;
11656 // Most TLS accesses are not RIP relative, even on x86-64. One exception is
11658 unsigned WrapperKind = X86ISD::Wrapper;
11659 if (model == TLSModel::LocalExec) {
11660 OperandFlags = is64Bit ? X86II::MO_TPOFF : X86II::MO_NTPOFF;
11661 } else if (model == TLSModel::InitialExec) {
11663 OperandFlags = X86II::MO_GOTTPOFF;
11664 WrapperKind = X86ISD::WrapperRIP;
11666 OperandFlags = isPIC ? X86II::MO_GOTNTPOFF : X86II::MO_INDNTPOFF;
11669 llvm_unreachable("Unexpected model");
11672 // emit "addl x@ntpoff,%eax" (local exec)
11673 // or "addl x@indntpoff,%eax" (initial exec)
11674 // or "addl x@gotntpoff(%ebx) ,%eax" (initial exec, 32-bit pic)
11676 DAG.getTargetGlobalAddress(GA->getGlobal(), dl, GA->getValueType(0),
11677 GA->getOffset(), OperandFlags);
11678 SDValue Offset = DAG.getNode(WrapperKind, dl, PtrVT, TGA);
11680 if (model == TLSModel::InitialExec) {
11681 if (isPIC && !is64Bit) {
11682 Offset = DAG.getNode(ISD::ADD, dl, PtrVT,
11683 DAG.getNode(X86ISD::GlobalBaseReg, SDLoc(), PtrVT),
11687 Offset = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Offset,
11688 MachinePointerInfo::getGOT(DAG.getMachineFunction()),
11689 false, false, false, 0);
11692 // The address of the thread local variable is the add of the thread
11693 // pointer with the offset of the variable.
11694 return DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, Offset);
11698 X86TargetLowering::LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const {
11700 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
11701 const GlobalValue *GV = GA->getGlobal();
11702 auto PtrVT = getPointerTy(DAG.getDataLayout());
11704 if (Subtarget->isTargetELF()) {
11705 if (DAG.getTarget().Options.EmulatedTLS)
11706 return LowerToTLSEmulatedModel(GA, DAG);
11707 TLSModel::Model model = DAG.getTarget().getTLSModel(GV);
11709 case TLSModel::GeneralDynamic:
11710 if (Subtarget->is64Bit())
11711 return LowerToTLSGeneralDynamicModel64(GA, DAG, PtrVT);
11712 return LowerToTLSGeneralDynamicModel32(GA, DAG, PtrVT);
11713 case TLSModel::LocalDynamic:
11714 return LowerToTLSLocalDynamicModel(GA, DAG, PtrVT,
11715 Subtarget->is64Bit());
11716 case TLSModel::InitialExec:
11717 case TLSModel::LocalExec:
11718 return LowerToTLSExecModel(GA, DAG, PtrVT, model, Subtarget->is64Bit(),
11719 DAG.getTarget().getRelocationModel() ==
11722 llvm_unreachable("Unknown TLS model.");
11725 if (Subtarget->isTargetDarwin()) {
11726 // Darwin only has one model of TLS. Lower to that.
11727 unsigned char OpFlag = 0;
11728 unsigned WrapperKind = Subtarget->isPICStyleRIPRel() ?
11729 X86ISD::WrapperRIP : X86ISD::Wrapper;
11731 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
11732 // global base reg.
11733 bool PIC32 = (DAG.getTarget().getRelocationModel() == Reloc::PIC_) &&
11734 !Subtarget->is64Bit();
11736 OpFlag = X86II::MO_TLVP_PIC_BASE;
11738 OpFlag = X86II::MO_TLVP;
11740 SDValue Result = DAG.getTargetGlobalAddress(GA->getGlobal(), DL,
11741 GA->getValueType(0),
11742 GA->getOffset(), OpFlag);
11743 SDValue Offset = DAG.getNode(WrapperKind, DL, PtrVT, Result);
11745 // With PIC32, the address is actually $g + Offset.
11747 Offset = DAG.getNode(ISD::ADD, DL, PtrVT,
11748 DAG.getNode(X86ISD::GlobalBaseReg, SDLoc(), PtrVT),
11751 // Lowering the machine isd will make sure everything is in the right
11753 SDValue Chain = DAG.getEntryNode();
11754 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
11755 SDValue Args[] = { Chain, Offset };
11756 Chain = DAG.getNode(X86ISD::TLSCALL, DL, NodeTys, Args);
11758 // TLSCALL will be codegen'ed as call. Inform MFI that function has calls.
11759 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
11760 MFI->setAdjustsStack(true);
11762 // And our return value (tls address) is in the standard call return value
11764 unsigned Reg = Subtarget->is64Bit() ? X86::RAX : X86::EAX;
11765 return DAG.getCopyFromReg(Chain, DL, Reg, PtrVT, Chain.getValue(1));
11768 if (Subtarget->isTargetKnownWindowsMSVC() ||
11769 Subtarget->isTargetWindowsGNU()) {
11770 // Just use the implicit TLS architecture
11771 // Need to generate someting similar to:
11772 // mov rdx, qword [gs:abs 58H]; Load pointer to ThreadLocalStorage
11774 // mov ecx, dword [rel _tls_index]: Load index (from C runtime)
11775 // mov rcx, qword [rdx+rcx*8]
11776 // mov eax, .tls$:tlsvar
11777 // [rax+rcx] contains the address
11778 // Windows 64bit: gs:0x58
11779 // Windows 32bit: fs:__tls_array
11782 SDValue Chain = DAG.getEntryNode();
11784 // Get the Thread Pointer, which is %fs:__tls_array (32-bit) or
11785 // %gs:0x58 (64-bit). On MinGW, __tls_array is not available, so directly
11786 // use its literal value of 0x2C.
11787 Value *Ptr = Constant::getNullValue(Subtarget->is64Bit()
11788 ? Type::getInt8PtrTy(*DAG.getContext(),
11790 : Type::getInt32PtrTy(*DAG.getContext(),
11793 SDValue TlsArray = Subtarget->is64Bit()
11794 ? DAG.getIntPtrConstant(0x58, dl)
11795 : (Subtarget->isTargetWindowsGNU()
11796 ? DAG.getIntPtrConstant(0x2C, dl)
11797 : DAG.getExternalSymbol("_tls_array", PtrVT));
11799 SDValue ThreadPointer =
11800 DAG.getLoad(PtrVT, dl, Chain, TlsArray, MachinePointerInfo(Ptr), false,
11804 if (GV->getThreadLocalMode() == GlobalVariable::LocalExecTLSModel) {
11805 res = ThreadPointer;
11807 // Load the _tls_index variable
11808 SDValue IDX = DAG.getExternalSymbol("_tls_index", PtrVT);
11809 if (Subtarget->is64Bit())
11810 IDX = DAG.getExtLoad(ISD::ZEXTLOAD, dl, PtrVT, Chain, IDX,
11811 MachinePointerInfo(), MVT::i32, false, false,
11814 IDX = DAG.getLoad(PtrVT, dl, Chain, IDX, MachinePointerInfo(), false,
11817 auto &DL = DAG.getDataLayout();
11819 DAG.getConstant(Log2_64_Ceil(DL.getPointerSize()), dl, PtrVT);
11820 IDX = DAG.getNode(ISD::SHL, dl, PtrVT, IDX, Scale);
11822 res = DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, IDX);
11825 res = DAG.getLoad(PtrVT, dl, Chain, res, MachinePointerInfo(), false, false,
11828 // Get the offset of start of .tls section
11829 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(), dl,
11830 GA->getValueType(0),
11831 GA->getOffset(), X86II::MO_SECREL);
11832 SDValue Offset = DAG.getNode(X86ISD::Wrapper, dl, PtrVT, TGA);
11834 // The address of the thread local variable is the add of the thread
11835 // pointer with the offset of the variable.
11836 return DAG.getNode(ISD::ADD, dl, PtrVT, res, Offset);
11839 llvm_unreachable("TLS not implemented for this target.");
11842 /// LowerShiftParts - Lower SRA_PARTS and friends, which return two i32 values
11843 /// and take a 2 x i32 value to shift plus a shift amount.
11844 static SDValue LowerShiftParts(SDValue Op, SelectionDAG &DAG) {
11845 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
11846 MVT VT = Op.getSimpleValueType();
11847 unsigned VTBits = VT.getSizeInBits();
11849 bool isSRA = Op.getOpcode() == ISD::SRA_PARTS;
11850 SDValue ShOpLo = Op.getOperand(0);
11851 SDValue ShOpHi = Op.getOperand(1);
11852 SDValue ShAmt = Op.getOperand(2);
11853 // X86ISD::SHLD and X86ISD::SHRD have defined overflow behavior but the
11854 // generic ISD nodes haven't. Insert an AND to be safe, it's optimized away
11856 SDValue SafeShAmt = DAG.getNode(ISD::AND, dl, MVT::i8, ShAmt,
11857 DAG.getConstant(VTBits - 1, dl, MVT::i8));
11858 SDValue Tmp1 = isSRA ? DAG.getNode(ISD::SRA, dl, VT, ShOpHi,
11859 DAG.getConstant(VTBits - 1, dl, MVT::i8))
11860 : DAG.getConstant(0, dl, VT);
11862 SDValue Tmp2, Tmp3;
11863 if (Op.getOpcode() == ISD::SHL_PARTS) {
11864 Tmp2 = DAG.getNode(X86ISD::SHLD, dl, VT, ShOpHi, ShOpLo, ShAmt);
11865 Tmp3 = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, SafeShAmt);
11867 Tmp2 = DAG.getNode(X86ISD::SHRD, dl, VT, ShOpLo, ShOpHi, ShAmt);
11868 Tmp3 = DAG.getNode(isSRA ? ISD::SRA : ISD::SRL, dl, VT, ShOpHi, SafeShAmt);
11871 // If the shift amount is larger or equal than the width of a part we can't
11872 // rely on the results of shld/shrd. Insert a test and select the appropriate
11873 // values for large shift amounts.
11874 SDValue AndNode = DAG.getNode(ISD::AND, dl, MVT::i8, ShAmt,
11875 DAG.getConstant(VTBits, dl, MVT::i8));
11876 SDValue Cond = DAG.getNode(X86ISD::CMP, dl, MVT::i32,
11877 AndNode, DAG.getConstant(0, dl, MVT::i8));
11880 SDValue CC = DAG.getConstant(X86::COND_NE, dl, MVT::i8);
11881 SDValue Ops0[4] = { Tmp2, Tmp3, CC, Cond };
11882 SDValue Ops1[4] = { Tmp3, Tmp1, CC, Cond };
11884 if (Op.getOpcode() == ISD::SHL_PARTS) {
11885 Hi = DAG.getNode(X86ISD::CMOV, dl, VT, Ops0);
11886 Lo = DAG.getNode(X86ISD::CMOV, dl, VT, Ops1);
11888 Lo = DAG.getNode(X86ISD::CMOV, dl, VT, Ops0);
11889 Hi = DAG.getNode(X86ISD::CMOV, dl, VT, Ops1);
11892 SDValue Ops[2] = { Lo, Hi };
11893 return DAG.getMergeValues(Ops, dl);
11896 SDValue X86TargetLowering::LowerSINT_TO_FP(SDValue Op,
11897 SelectionDAG &DAG) const {
11898 SDValue Src = Op.getOperand(0);
11899 MVT SrcVT = Src.getSimpleValueType();
11900 MVT VT = Op.getSimpleValueType();
11903 if (SrcVT.isVector()) {
11904 if (SrcVT == MVT::v2i32 && VT == MVT::v2f64) {
11905 return DAG.getNode(X86ISD::CVTDQ2PD, dl, VT,
11906 DAG.getNode(ISD::CONCAT_VECTORS, dl, MVT::v4i32, Src,
11907 DAG.getUNDEF(SrcVT)));
11909 if (SrcVT.getVectorElementType() == MVT::i1) {
11910 MVT IntegerVT = MVT::getVectorVT(MVT::i32, SrcVT.getVectorNumElements());
11911 return DAG.getNode(ISD::SINT_TO_FP, dl, Op.getValueType(),
11912 DAG.getNode(ISD::SIGN_EXTEND, dl, IntegerVT, Src));
11917 assert(SrcVT <= MVT::i64 && SrcVT >= MVT::i16 &&
11918 "Unknown SINT_TO_FP to lower!");
11920 // These are really Legal; return the operand so the caller accepts it as
11922 if (SrcVT == MVT::i32 && isScalarFPTypeInSSEReg(Op.getValueType()))
11924 if (SrcVT == MVT::i64 && isScalarFPTypeInSSEReg(Op.getValueType()) &&
11925 Subtarget->is64Bit()) {
11929 unsigned Size = SrcVT.getSizeInBits()/8;
11930 MachineFunction &MF = DAG.getMachineFunction();
11931 auto PtrVT = getPointerTy(MF.getDataLayout());
11932 int SSFI = MF.getFrameInfo()->CreateStackObject(Size, Size, false);
11933 SDValue StackSlot = DAG.getFrameIndex(SSFI, PtrVT);
11934 SDValue Chain = DAG.getStore(
11935 DAG.getEntryNode(), dl, Op.getOperand(0), StackSlot,
11936 MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), SSFI), false,
11938 return BuildFILD(Op, SrcVT, Chain, StackSlot, DAG);
11941 SDValue X86TargetLowering::BuildFILD(SDValue Op, EVT SrcVT, SDValue Chain,
11943 SelectionDAG &DAG) const {
11947 bool useSSE = isScalarFPTypeInSSEReg(Op.getValueType());
11949 Tys = DAG.getVTList(MVT::f64, MVT::Other, MVT::Glue);
11951 Tys = DAG.getVTList(Op.getValueType(), MVT::Other);
11953 unsigned ByteSize = SrcVT.getSizeInBits()/8;
11955 FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(StackSlot);
11956 MachineMemOperand *MMO;
11958 int SSFI = FI->getIndex();
11959 MMO = DAG.getMachineFunction().getMachineMemOperand(
11960 MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), SSFI),
11961 MachineMemOperand::MOLoad, ByteSize, ByteSize);
11963 MMO = cast<LoadSDNode>(StackSlot)->getMemOperand();
11964 StackSlot = StackSlot.getOperand(1);
11966 SDValue Ops[] = { Chain, StackSlot, DAG.getValueType(SrcVT) };
11967 SDValue Result = DAG.getMemIntrinsicNode(useSSE ? X86ISD::FILD_FLAG :
11969 Tys, Ops, SrcVT, MMO);
11972 Chain = Result.getValue(1);
11973 SDValue InFlag = Result.getValue(2);
11975 // FIXME: Currently the FST is flagged to the FILD_FLAG. This
11976 // shouldn't be necessary except that RFP cannot be live across
11977 // multiple blocks. When stackifier is fixed, they can be uncoupled.
11978 MachineFunction &MF = DAG.getMachineFunction();
11979 unsigned SSFISize = Op.getValueType().getSizeInBits()/8;
11980 int SSFI = MF.getFrameInfo()->CreateStackObject(SSFISize, SSFISize, false);
11981 auto PtrVT = getPointerTy(MF.getDataLayout());
11982 SDValue StackSlot = DAG.getFrameIndex(SSFI, PtrVT);
11983 Tys = DAG.getVTList(MVT::Other);
11985 Chain, Result, StackSlot, DAG.getValueType(Op.getValueType()), InFlag
11987 MachineMemOperand *MMO = DAG.getMachineFunction().getMachineMemOperand(
11988 MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), SSFI),
11989 MachineMemOperand::MOStore, SSFISize, SSFISize);
11991 Chain = DAG.getMemIntrinsicNode(X86ISD::FST, DL, Tys,
11992 Ops, Op.getValueType(), MMO);
11993 Result = DAG.getLoad(
11994 Op.getValueType(), DL, Chain, StackSlot,
11995 MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), SSFI),
11996 false, false, false, 0);
12002 // LowerUINT_TO_FP_i64 - 64-bit unsigned integer to double expansion.
12003 SDValue X86TargetLowering::LowerUINT_TO_FP_i64(SDValue Op,
12004 SelectionDAG &DAG) const {
12005 // This algorithm is not obvious. Here it is what we're trying to output:
12008 punpckldq (c0), %xmm0 // c0: (uint4){ 0x43300000U, 0x45300000U, 0U, 0U }
12009 subpd (c1), %xmm0 // c1: (double2){ 0x1.0p52, 0x1.0p52 * 0x1.0p32 }
12011 haddpd %xmm0, %xmm0
12013 pshufd $0x4e, %xmm0, %xmm1
12019 LLVMContext *Context = DAG.getContext();
12021 // Build some magic constants.
12022 static const uint32_t CV0[] = { 0x43300000, 0x45300000, 0, 0 };
12023 Constant *C0 = ConstantDataVector::get(*Context, CV0);
12024 auto PtrVT = getPointerTy(DAG.getDataLayout());
12025 SDValue CPIdx0 = DAG.getConstantPool(C0, PtrVT, 16);
12027 SmallVector<Constant*,2> CV1;
12029 ConstantFP::get(*Context, APFloat(APFloat::IEEEdouble,
12030 APInt(64, 0x4330000000000000ULL))));
12032 ConstantFP::get(*Context, APFloat(APFloat::IEEEdouble,
12033 APInt(64, 0x4530000000000000ULL))));
12034 Constant *C1 = ConstantVector::get(CV1);
12035 SDValue CPIdx1 = DAG.getConstantPool(C1, PtrVT, 16);
12037 // Load the 64-bit value into an XMM register.
12038 SDValue XR1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i64,
12041 DAG.getLoad(MVT::v4i32, dl, DAG.getEntryNode(), CPIdx0,
12042 MachinePointerInfo::getConstantPool(DAG.getMachineFunction()),
12043 false, false, false, 16);
12045 getUnpackl(DAG, dl, MVT::v4i32, DAG.getBitcast(MVT::v4i32, XR1), CLod0);
12048 DAG.getLoad(MVT::v2f64, dl, CLod0.getValue(1), CPIdx1,
12049 MachinePointerInfo::getConstantPool(DAG.getMachineFunction()),
12050 false, false, false, 16);
12051 SDValue XR2F = DAG.getBitcast(MVT::v2f64, Unpck1);
12052 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::v2f64, XR2F, CLod1);
12055 if (Subtarget->hasSSE3()) {
12056 // FIXME: The 'haddpd' instruction may be slower than 'movhlps + addsd'.
12057 Result = DAG.getNode(X86ISD::FHADD, dl, MVT::v2f64, Sub, Sub);
12059 SDValue S2F = DAG.getBitcast(MVT::v4i32, Sub);
12060 SDValue Shuffle = getTargetShuffleNode(X86ISD::PSHUFD, dl, MVT::v4i32,
12062 Result = DAG.getNode(ISD::FADD, dl, MVT::v2f64,
12063 DAG.getBitcast(MVT::v2f64, Shuffle), Sub);
12066 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Result,
12067 DAG.getIntPtrConstant(0, dl));
12070 // LowerUINT_TO_FP_i32 - 32-bit unsigned integer to float expansion.
12071 SDValue X86TargetLowering::LowerUINT_TO_FP_i32(SDValue Op,
12072 SelectionDAG &DAG) const {
12074 // FP constant to bias correct the final result.
12075 SDValue Bias = DAG.getConstantFP(BitsToDouble(0x4330000000000000ULL), dl,
12078 // Load the 32-bit value into an XMM register.
12079 SDValue Load = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
12082 // Zero out the upper parts of the register.
12083 Load = getShuffleVectorZeroOrUndef(Load, 0, true, Subtarget, DAG);
12085 Load = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
12086 DAG.getBitcast(MVT::v2f64, Load),
12087 DAG.getIntPtrConstant(0, dl));
12089 // Or the load with the bias.
12090 SDValue Or = DAG.getNode(
12091 ISD::OR, dl, MVT::v2i64,
12092 DAG.getBitcast(MVT::v2i64,
12093 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2f64, Load)),
12094 DAG.getBitcast(MVT::v2i64,
12095 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2f64, Bias)));
12097 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
12098 DAG.getBitcast(MVT::v2f64, Or), DAG.getIntPtrConstant(0, dl));
12100 // Subtract the bias.
12101 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::f64, Or, Bias);
12103 // Handle final rounding.
12104 EVT DestVT = Op.getValueType();
12106 if (DestVT.bitsLT(MVT::f64))
12107 return DAG.getNode(ISD::FP_ROUND, dl, DestVT, Sub,
12108 DAG.getIntPtrConstant(0, dl));
12109 if (DestVT.bitsGT(MVT::f64))
12110 return DAG.getNode(ISD::FP_EXTEND, dl, DestVT, Sub);
12112 // Handle final rounding.
12116 static SDValue lowerUINT_TO_FP_vXi32(SDValue Op, SelectionDAG &DAG,
12117 const X86Subtarget &Subtarget) {
12118 // The algorithm is the following:
12119 // #ifdef __SSE4_1__
12120 // uint4 lo = _mm_blend_epi16( v, (uint4) 0x4b000000, 0xaa);
12121 // uint4 hi = _mm_blend_epi16( _mm_srli_epi32(v,16),
12122 // (uint4) 0x53000000, 0xaa);
12124 // uint4 lo = (v & (uint4) 0xffff) | (uint4) 0x4b000000;
12125 // uint4 hi = (v >> 16) | (uint4) 0x53000000;
12127 // float4 fhi = (float4) hi - (0x1.0p39f + 0x1.0p23f);
12128 // return (float4) lo + fhi;
12131 SDValue V = Op->getOperand(0);
12132 EVT VecIntVT = V.getValueType();
12133 bool Is128 = VecIntVT == MVT::v4i32;
12134 EVT VecFloatVT = Is128 ? MVT::v4f32 : MVT::v8f32;
12135 // If we convert to something else than the supported type, e.g., to v4f64,
12137 if (VecFloatVT != Op->getValueType(0))
12140 unsigned NumElts = VecIntVT.getVectorNumElements();
12141 assert((VecIntVT == MVT::v4i32 || VecIntVT == MVT::v8i32) &&
12142 "Unsupported custom type");
12143 assert(NumElts <= 8 && "The size of the constant array must be fixed");
12145 // In the #idef/#else code, we have in common:
12146 // - The vector of constants:
12152 // Create the splat vector for 0x4b000000.
12153 SDValue CstLow = DAG.getConstant(0x4b000000, DL, MVT::i32);
12154 SDValue CstLowArray[] = {CstLow, CstLow, CstLow, CstLow,
12155 CstLow, CstLow, CstLow, CstLow};
12156 SDValue VecCstLow = DAG.getNode(ISD::BUILD_VECTOR, DL, VecIntVT,
12157 makeArrayRef(&CstLowArray[0], NumElts));
12158 // Create the splat vector for 0x53000000.
12159 SDValue CstHigh = DAG.getConstant(0x53000000, DL, MVT::i32);
12160 SDValue CstHighArray[] = {CstHigh, CstHigh, CstHigh, CstHigh,
12161 CstHigh, CstHigh, CstHigh, CstHigh};
12162 SDValue VecCstHigh = DAG.getNode(ISD::BUILD_VECTOR, DL, VecIntVT,
12163 makeArrayRef(&CstHighArray[0], NumElts));
12165 // Create the right shift.
12166 SDValue CstShift = DAG.getConstant(16, DL, MVT::i32);
12167 SDValue CstShiftArray[] = {CstShift, CstShift, CstShift, CstShift,
12168 CstShift, CstShift, CstShift, CstShift};
12169 SDValue VecCstShift = DAG.getNode(ISD::BUILD_VECTOR, DL, VecIntVT,
12170 makeArrayRef(&CstShiftArray[0], NumElts));
12171 SDValue HighShift = DAG.getNode(ISD::SRL, DL, VecIntVT, V, VecCstShift);
12174 if (Subtarget.hasSSE41()) {
12175 EVT VecI16VT = Is128 ? MVT::v8i16 : MVT::v16i16;
12176 // uint4 lo = _mm_blend_epi16( v, (uint4) 0x4b000000, 0xaa);
12177 SDValue VecCstLowBitcast = DAG.getBitcast(VecI16VT, VecCstLow);
12178 SDValue VecBitcast = DAG.getBitcast(VecI16VT, V);
12179 // Low will be bitcasted right away, so do not bother bitcasting back to its
12181 Low = DAG.getNode(X86ISD::BLENDI, DL, VecI16VT, VecBitcast,
12182 VecCstLowBitcast, DAG.getConstant(0xaa, DL, MVT::i32));
12183 // uint4 hi = _mm_blend_epi16( _mm_srli_epi32(v,16),
12184 // (uint4) 0x53000000, 0xaa);
12185 SDValue VecCstHighBitcast = DAG.getBitcast(VecI16VT, VecCstHigh);
12186 SDValue VecShiftBitcast = DAG.getBitcast(VecI16VT, HighShift);
12187 // High will be bitcasted right away, so do not bother bitcasting back to
12188 // its original type.
12189 High = DAG.getNode(X86ISD::BLENDI, DL, VecI16VT, VecShiftBitcast,
12190 VecCstHighBitcast, DAG.getConstant(0xaa, DL, MVT::i32));
12192 SDValue CstMask = DAG.getConstant(0xffff, DL, MVT::i32);
12193 SDValue VecCstMask = DAG.getNode(ISD::BUILD_VECTOR, DL, VecIntVT, CstMask,
12194 CstMask, CstMask, CstMask);
12195 // uint4 lo = (v & (uint4) 0xffff) | (uint4) 0x4b000000;
12196 SDValue LowAnd = DAG.getNode(ISD::AND, DL, VecIntVT, V, VecCstMask);
12197 Low = DAG.getNode(ISD::OR, DL, VecIntVT, LowAnd, VecCstLow);
12199 // uint4 hi = (v >> 16) | (uint4) 0x53000000;
12200 High = DAG.getNode(ISD::OR, DL, VecIntVT, HighShift, VecCstHigh);
12203 // Create the vector constant for -(0x1.0p39f + 0x1.0p23f).
12204 SDValue CstFAdd = DAG.getConstantFP(
12205 APFloat(APFloat::IEEEsingle, APInt(32, 0xD3000080)), DL, MVT::f32);
12206 SDValue CstFAddArray[] = {CstFAdd, CstFAdd, CstFAdd, CstFAdd,
12207 CstFAdd, CstFAdd, CstFAdd, CstFAdd};
12208 SDValue VecCstFAdd = DAG.getNode(ISD::BUILD_VECTOR, DL, VecFloatVT,
12209 makeArrayRef(&CstFAddArray[0], NumElts));
12211 // float4 fhi = (float4) hi - (0x1.0p39f + 0x1.0p23f);
12212 SDValue HighBitcast = DAG.getBitcast(VecFloatVT, High);
12214 DAG.getNode(ISD::FADD, DL, VecFloatVT, HighBitcast, VecCstFAdd);
12215 // return (float4) lo + fhi;
12216 SDValue LowBitcast = DAG.getBitcast(VecFloatVT, Low);
12217 return DAG.getNode(ISD::FADD, DL, VecFloatVT, LowBitcast, FHigh);
12220 SDValue X86TargetLowering::lowerUINT_TO_FP_vec(SDValue Op,
12221 SelectionDAG &DAG) const {
12222 SDValue N0 = Op.getOperand(0);
12223 MVT SVT = N0.getSimpleValueType();
12226 switch (SVT.SimpleTy) {
12228 llvm_unreachable("Custom UINT_TO_FP is not supported!");
12233 MVT NVT = MVT::getVectorVT(MVT::i32, SVT.getVectorNumElements());
12234 return DAG.getNode(ISD::SINT_TO_FP, dl, Op.getValueType(),
12235 DAG.getNode(ISD::ZERO_EXTEND, dl, NVT, N0));
12239 return lowerUINT_TO_FP_vXi32(Op, DAG, *Subtarget);
12242 if (Subtarget->hasAVX512())
12243 return DAG.getNode(ISD::UINT_TO_FP, dl, Op.getValueType(),
12244 DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v16i32, N0));
12246 llvm_unreachable(nullptr);
12249 SDValue X86TargetLowering::LowerUINT_TO_FP(SDValue Op,
12250 SelectionDAG &DAG) const {
12251 SDValue N0 = Op.getOperand(0);
12253 auto PtrVT = getPointerTy(DAG.getDataLayout());
12255 if (Op.getValueType().isVector())
12256 return lowerUINT_TO_FP_vec(Op, DAG);
12258 // Since UINT_TO_FP is legal (it's marked custom), dag combiner won't
12259 // optimize it to a SINT_TO_FP when the sign bit is known zero. Perform
12260 // the optimization here.
12261 if (DAG.SignBitIsZero(N0))
12262 return DAG.getNode(ISD::SINT_TO_FP, dl, Op.getValueType(), N0);
12264 MVT SrcVT = N0.getSimpleValueType();
12265 MVT DstVT = Op.getSimpleValueType();
12266 if (SrcVT == MVT::i64 && DstVT == MVT::f64 && X86ScalarSSEf64)
12267 return LowerUINT_TO_FP_i64(Op, DAG);
12268 if (SrcVT == MVT::i32 && X86ScalarSSEf64)
12269 return LowerUINT_TO_FP_i32(Op, DAG);
12270 if (Subtarget->is64Bit() && SrcVT == MVT::i64 && DstVT == MVT::f32)
12273 // Make a 64-bit buffer, and use it to build an FILD.
12274 SDValue StackSlot = DAG.CreateStackTemporary(MVT::i64);
12275 if (SrcVT == MVT::i32) {
12276 SDValue WordOff = DAG.getConstant(4, dl, PtrVT);
12277 SDValue OffsetSlot = DAG.getNode(ISD::ADD, dl, PtrVT, StackSlot, WordOff);
12278 SDValue Store1 = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
12279 StackSlot, MachinePointerInfo(),
12281 SDValue Store2 = DAG.getStore(Store1, dl, DAG.getConstant(0, dl, MVT::i32),
12282 OffsetSlot, MachinePointerInfo(),
12284 SDValue Fild = BuildFILD(Op, MVT::i64, Store2, StackSlot, DAG);
12288 assert(SrcVT == MVT::i64 && "Unexpected type in UINT_TO_FP");
12289 SDValue Store = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
12290 StackSlot, MachinePointerInfo(),
12292 // For i64 source, we need to add the appropriate power of 2 if the input
12293 // was negative. This is the same as the optimization in
12294 // DAGTypeLegalizer::ExpandIntOp_UNIT_TO_FP, and for it to be safe here,
12295 // we must be careful to do the computation in x87 extended precision, not
12296 // in SSE. (The generic code can't know it's OK to do this, or how to.)
12297 int SSFI = cast<FrameIndexSDNode>(StackSlot)->getIndex();
12298 MachineMemOperand *MMO = DAG.getMachineFunction().getMachineMemOperand(
12299 MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), SSFI),
12300 MachineMemOperand::MOLoad, 8, 8);
12302 SDVTList Tys = DAG.getVTList(MVT::f80, MVT::Other);
12303 SDValue Ops[] = { Store, StackSlot, DAG.getValueType(MVT::i64) };
12304 SDValue Fild = DAG.getMemIntrinsicNode(X86ISD::FILD, dl, Tys, Ops,
12307 APInt FF(32, 0x5F800000ULL);
12309 // Check whether the sign bit is set.
12310 SDValue SignSet = DAG.getSetCC(
12311 dl, getSetCCResultType(DAG.getDataLayout(), *DAG.getContext(), MVT::i64),
12312 Op.getOperand(0), DAG.getConstant(0, dl, MVT::i64), ISD::SETLT);
12314 // Build a 64 bit pair (0, FF) in the constant pool, with FF in the lo bits.
12315 SDValue FudgePtr = DAG.getConstantPool(
12316 ConstantInt::get(*DAG.getContext(), FF.zext(64)), PtrVT);
12318 // Get a pointer to FF if the sign bit was set, or to 0 otherwise.
12319 SDValue Zero = DAG.getIntPtrConstant(0, dl);
12320 SDValue Four = DAG.getIntPtrConstant(4, dl);
12321 SDValue Offset = DAG.getNode(ISD::SELECT, dl, Zero.getValueType(), SignSet,
12323 FudgePtr = DAG.getNode(ISD::ADD, dl, PtrVT, FudgePtr, Offset);
12325 // Load the value out, extending it from f32 to f80.
12326 // FIXME: Avoid the extend by constructing the right constant pool?
12327 SDValue Fudge = DAG.getExtLoad(
12328 ISD::EXTLOAD, dl, MVT::f80, DAG.getEntryNode(), FudgePtr,
12329 MachinePointerInfo::getConstantPool(DAG.getMachineFunction()), MVT::f32,
12330 false, false, false, 4);
12331 // Extend everything to 80 bits to force it to be done on x87.
12332 SDValue Add = DAG.getNode(ISD::FADD, dl, MVT::f80, Fild, Fudge);
12333 return DAG.getNode(ISD::FP_ROUND, dl, DstVT, Add,
12334 DAG.getIntPtrConstant(0, dl));
12337 std::pair<SDValue,SDValue>
12338 X86TargetLowering:: FP_TO_INTHelper(SDValue Op, SelectionDAG &DAG,
12339 bool IsSigned, bool IsReplace) const {
12342 EVT DstTy = Op.getValueType();
12343 auto PtrVT = getPointerTy(DAG.getDataLayout());
12345 if (!IsSigned && !isIntegerTypeFTOL(DstTy)) {
12346 assert(DstTy == MVT::i32 && "Unexpected FP_TO_UINT");
12350 assert(DstTy.getSimpleVT() <= MVT::i64 &&
12351 DstTy.getSimpleVT() >= MVT::i16 &&
12352 "Unknown FP_TO_INT to lower!");
12354 // These are really Legal.
12355 if (DstTy == MVT::i32 &&
12356 isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType()))
12357 return std::make_pair(SDValue(), SDValue());
12358 if (Subtarget->is64Bit() &&
12359 DstTy == MVT::i64 &&
12360 isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType()))
12361 return std::make_pair(SDValue(), SDValue());
12363 // We lower FP->int64 either into FISTP64 followed by a load from a temporary
12364 // stack slot, or into the FTOL runtime function.
12365 MachineFunction &MF = DAG.getMachineFunction();
12366 unsigned MemSize = DstTy.getSizeInBits()/8;
12367 int SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize, false);
12368 SDValue StackSlot = DAG.getFrameIndex(SSFI, PtrVT);
12371 if (!IsSigned && isIntegerTypeFTOL(DstTy))
12372 Opc = X86ISD::WIN_FTOL;
12374 switch (DstTy.getSimpleVT().SimpleTy) {
12375 default: llvm_unreachable("Invalid FP_TO_SINT to lower!");
12376 case MVT::i16: Opc = X86ISD::FP_TO_INT16_IN_MEM; break;
12377 case MVT::i32: Opc = X86ISD::FP_TO_INT32_IN_MEM; break;
12378 case MVT::i64: Opc = X86ISD::FP_TO_INT64_IN_MEM; break;
12381 SDValue Chain = DAG.getEntryNode();
12382 SDValue Value = Op.getOperand(0);
12383 EVT TheVT = Op.getOperand(0).getValueType();
12384 // FIXME This causes a redundant load/store if the SSE-class value is already
12385 // in memory, such as if it is on the callstack.
12386 if (isScalarFPTypeInSSEReg(TheVT)) {
12387 assert(DstTy == MVT::i64 && "Invalid FP_TO_SINT to lower!");
12388 Chain = DAG.getStore(Chain, DL, Value, StackSlot,
12389 MachinePointerInfo::getFixedStack(MF, SSFI), false,
12391 SDVTList Tys = DAG.getVTList(Op.getOperand(0).getValueType(), MVT::Other);
12393 Chain, StackSlot, DAG.getValueType(TheVT)
12396 MachineMemOperand *MMO =
12397 MF.getMachineMemOperand(MachinePointerInfo::getFixedStack(MF, SSFI),
12398 MachineMemOperand::MOLoad, MemSize, MemSize);
12399 Value = DAG.getMemIntrinsicNode(X86ISD::FLD, DL, Tys, Ops, DstTy, MMO);
12400 Chain = Value.getValue(1);
12401 SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize, false);
12402 StackSlot = DAG.getFrameIndex(SSFI, PtrVT);
12405 MachineMemOperand *MMO =
12406 MF.getMachineMemOperand(MachinePointerInfo::getFixedStack(MF, SSFI),
12407 MachineMemOperand::MOStore, MemSize, MemSize);
12409 if (Opc != X86ISD::WIN_FTOL) {
12410 // Build the FP_TO_INT*_IN_MEM
12411 SDValue Ops[] = { Chain, Value, StackSlot };
12412 SDValue FIST = DAG.getMemIntrinsicNode(Opc, DL, DAG.getVTList(MVT::Other),
12414 return std::make_pair(FIST, StackSlot);
12416 SDValue ftol = DAG.getNode(X86ISD::WIN_FTOL, DL,
12417 DAG.getVTList(MVT::Other, MVT::Glue),
12419 SDValue eax = DAG.getCopyFromReg(ftol, DL, X86::EAX,
12420 MVT::i32, ftol.getValue(1));
12421 SDValue edx = DAG.getCopyFromReg(eax.getValue(1), DL, X86::EDX,
12422 MVT::i32, eax.getValue(2));
12423 SDValue Ops[] = { eax, edx };
12424 SDValue pair = IsReplace
12425 ? DAG.getNode(ISD::BUILD_PAIR, DL, MVT::i64, Ops)
12426 : DAG.getMergeValues(Ops, DL);
12427 return std::make_pair(pair, SDValue());
12431 static SDValue LowerAVXExtend(SDValue Op, SelectionDAG &DAG,
12432 const X86Subtarget *Subtarget) {
12433 MVT VT = Op->getSimpleValueType(0);
12434 SDValue In = Op->getOperand(0);
12435 MVT InVT = In.getSimpleValueType();
12438 if (VT.is512BitVector() || InVT.getScalarType() == MVT::i1)
12439 return DAG.getNode(ISD::ZERO_EXTEND, dl, VT, In);
12441 // Optimize vectors in AVX mode:
12444 // Use vpunpcklwd for 4 lower elements v8i16 -> v4i32.
12445 // Use vpunpckhwd for 4 upper elements v8i16 -> v4i32.
12446 // Concat upper and lower parts.
12449 // Use vpunpckldq for 4 lower elements v4i32 -> v2i64.
12450 // Use vpunpckhdq for 4 upper elements v4i32 -> v2i64.
12451 // Concat upper and lower parts.
12454 if (((VT != MVT::v16i16) || (InVT != MVT::v16i8)) &&
12455 ((VT != MVT::v8i32) || (InVT != MVT::v8i16)) &&
12456 ((VT != MVT::v4i64) || (InVT != MVT::v4i32)))
12459 if (Subtarget->hasInt256())
12460 return DAG.getNode(X86ISD::VZEXT, dl, VT, In);
12462 SDValue ZeroVec = getZeroVector(InVT, Subtarget, DAG, dl);
12463 SDValue Undef = DAG.getUNDEF(InVT);
12464 bool NeedZero = Op.getOpcode() == ISD::ZERO_EXTEND;
12465 SDValue OpLo = getUnpackl(DAG, dl, InVT, In, NeedZero ? ZeroVec : Undef);
12466 SDValue OpHi = getUnpackh(DAG, dl, InVT, In, NeedZero ? ZeroVec : Undef);
12468 MVT HVT = MVT::getVectorVT(VT.getVectorElementType(),
12469 VT.getVectorNumElements()/2);
12471 OpLo = DAG.getBitcast(HVT, OpLo);
12472 OpHi = DAG.getBitcast(HVT, OpHi);
12474 return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT, OpLo, OpHi);
12477 static SDValue LowerZERO_EXTEND_AVX512(SDValue Op,
12478 const X86Subtarget *Subtarget, SelectionDAG &DAG) {
12479 MVT VT = Op->getSimpleValueType(0);
12480 SDValue In = Op->getOperand(0);
12481 MVT InVT = In.getSimpleValueType();
12483 unsigned int NumElts = VT.getVectorNumElements();
12484 if (NumElts != 8 && NumElts != 16 && !Subtarget->hasBWI())
12487 if (VT.is512BitVector() && InVT.getVectorElementType() != MVT::i1)
12488 return DAG.getNode(X86ISD::VZEXT, DL, VT, In);
12490 assert(InVT.getVectorElementType() == MVT::i1);
12491 MVT ExtVT = NumElts == 8 ? MVT::v8i64 : MVT::v16i32;
12493 DAG.getConstant(APInt(ExtVT.getScalarSizeInBits(), 1), DL, ExtVT);
12495 DAG.getConstant(APInt::getNullValue(ExtVT.getScalarSizeInBits()), DL, ExtVT);
12497 SDValue V = DAG.getNode(ISD::VSELECT, DL, ExtVT, In, One, Zero);
12498 if (VT.is512BitVector())
12500 return DAG.getNode(X86ISD::VTRUNC, DL, VT, V);
12503 static SDValue LowerANY_EXTEND(SDValue Op, const X86Subtarget *Subtarget,
12504 SelectionDAG &DAG) {
12505 if (Subtarget->hasFp256())
12506 if (SDValue Res = LowerAVXExtend(Op, DAG, Subtarget))
12512 static SDValue LowerZERO_EXTEND(SDValue Op, const X86Subtarget *Subtarget,
12513 SelectionDAG &DAG) {
12515 MVT VT = Op.getSimpleValueType();
12516 SDValue In = Op.getOperand(0);
12517 MVT SVT = In.getSimpleValueType();
12519 if (VT.is512BitVector() || SVT.getVectorElementType() == MVT::i1)
12520 return LowerZERO_EXTEND_AVX512(Op, Subtarget, DAG);
12522 if (Subtarget->hasFp256())
12523 if (SDValue Res = LowerAVXExtend(Op, DAG, Subtarget))
12526 assert(!VT.is256BitVector() || !SVT.is128BitVector() ||
12527 VT.getVectorNumElements() != SVT.getVectorNumElements());
12531 SDValue X86TargetLowering::LowerTRUNCATE(SDValue Op, SelectionDAG &DAG) const {
12533 MVT VT = Op.getSimpleValueType();
12534 SDValue In = Op.getOperand(0);
12535 MVT InVT = In.getSimpleValueType();
12537 if (VT == MVT::i1) {
12538 assert((InVT.isInteger() && (InVT.getSizeInBits() <= 64)) &&
12539 "Invalid scalar TRUNCATE operation");
12540 if (InVT.getSizeInBits() >= 32)
12542 In = DAG.getNode(ISD::ANY_EXTEND, DL, MVT::i32, In);
12543 return DAG.getNode(ISD::TRUNCATE, DL, VT, In);
12545 assert(VT.getVectorNumElements() == InVT.getVectorNumElements() &&
12546 "Invalid TRUNCATE operation");
12548 // move vector to mask - truncate solution for SKX
12549 if (VT.getVectorElementType() == MVT::i1) {
12550 if (InVT.is512BitVector() && InVT.getScalarSizeInBits() <= 16 &&
12551 Subtarget->hasBWI())
12552 return Op; // legal, will go to VPMOVB2M, VPMOVW2M
12553 if ((InVT.is256BitVector() || InVT.is128BitVector())
12554 && InVT.getScalarSizeInBits() <= 16 &&
12555 Subtarget->hasBWI() && Subtarget->hasVLX())
12556 return Op; // legal, will go to VPMOVB2M, VPMOVW2M
12557 if (InVT.is512BitVector() && InVT.getScalarSizeInBits() >= 32 &&
12558 Subtarget->hasDQI())
12559 return Op; // legal, will go to VPMOVD2M, VPMOVQ2M
12560 if ((InVT.is256BitVector() || InVT.is128BitVector())
12561 && InVT.getScalarSizeInBits() >= 32 &&
12562 Subtarget->hasDQI() && Subtarget->hasVLX())
12563 return Op; // legal, will go to VPMOVB2M, VPMOVQ2M
12566 if (VT.getVectorElementType() == MVT::i1) {
12567 assert(VT.getVectorElementType() == MVT::i1 && "Unexpected vector type");
12568 unsigned NumElts = InVT.getVectorNumElements();
12569 assert ((NumElts == 8 || NumElts == 16) && "Unexpected vector type");
12570 if (InVT.getSizeInBits() < 512) {
12571 MVT ExtVT = (NumElts == 16)? MVT::v16i32 : MVT::v8i64;
12572 In = DAG.getNode(ISD::SIGN_EXTEND, DL, ExtVT, In);
12577 DAG.getConstant(APInt::getSignBit(InVT.getScalarSizeInBits()), DL, InVT);
12578 SDValue And = DAG.getNode(ISD::AND, DL, InVT, OneV, In);
12579 return DAG.getNode(X86ISD::TESTM, DL, VT, And, And);
12582 // vpmovqb/w/d, vpmovdb/w, vpmovwb
12583 if (((!InVT.is512BitVector() && Subtarget->hasVLX()) || InVT.is512BitVector()) &&
12584 (InVT.getVectorElementType() != MVT::i16 || Subtarget->hasBWI()))
12585 return DAG.getNode(X86ISD::VTRUNC, DL, VT, In);
12587 if ((VT == MVT::v4i32) && (InVT == MVT::v4i64)) {
12588 // On AVX2, v4i64 -> v4i32 becomes VPERMD.
12589 if (Subtarget->hasInt256()) {
12590 static const int ShufMask[] = {0, 2, 4, 6, -1, -1, -1, -1};
12591 In = DAG.getBitcast(MVT::v8i32, In);
12592 In = DAG.getVectorShuffle(MVT::v8i32, DL, In, DAG.getUNDEF(MVT::v8i32),
12594 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, VT, In,
12595 DAG.getIntPtrConstant(0, DL));
12598 SDValue OpLo = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v2i64, In,
12599 DAG.getIntPtrConstant(0, DL));
12600 SDValue OpHi = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v2i64, In,
12601 DAG.getIntPtrConstant(2, DL));
12602 OpLo = DAG.getBitcast(MVT::v4i32, OpLo);
12603 OpHi = DAG.getBitcast(MVT::v4i32, OpHi);
12604 static const int ShufMask[] = {0, 2, 4, 6};
12605 return DAG.getVectorShuffle(VT, DL, OpLo, OpHi, ShufMask);
12608 if ((VT == MVT::v8i16) && (InVT == MVT::v8i32)) {
12609 // On AVX2, v8i32 -> v8i16 becomed PSHUFB.
12610 if (Subtarget->hasInt256()) {
12611 In = DAG.getBitcast(MVT::v32i8, In);
12613 SmallVector<SDValue,32> pshufbMask;
12614 for (unsigned i = 0; i < 2; ++i) {
12615 pshufbMask.push_back(DAG.getConstant(0x0, DL, MVT::i8));
12616 pshufbMask.push_back(DAG.getConstant(0x1, DL, MVT::i8));
12617 pshufbMask.push_back(DAG.getConstant(0x4, DL, MVT::i8));
12618 pshufbMask.push_back(DAG.getConstant(0x5, DL, MVT::i8));
12619 pshufbMask.push_back(DAG.getConstant(0x8, DL, MVT::i8));
12620 pshufbMask.push_back(DAG.getConstant(0x9, DL, MVT::i8));
12621 pshufbMask.push_back(DAG.getConstant(0xc, DL, MVT::i8));
12622 pshufbMask.push_back(DAG.getConstant(0xd, DL, MVT::i8));
12623 for (unsigned j = 0; j < 8; ++j)
12624 pshufbMask.push_back(DAG.getConstant(0x80, DL, MVT::i8));
12626 SDValue BV = DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v32i8, pshufbMask);
12627 In = DAG.getNode(X86ISD::PSHUFB, DL, MVT::v32i8, In, BV);
12628 In = DAG.getBitcast(MVT::v4i64, In);
12630 static const int ShufMask[] = {0, 2, -1, -1};
12631 In = DAG.getVectorShuffle(MVT::v4i64, DL, In, DAG.getUNDEF(MVT::v4i64),
12633 In = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v2i64, In,
12634 DAG.getIntPtrConstant(0, DL));
12635 return DAG.getBitcast(VT, In);
12638 SDValue OpLo = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v4i32, In,
12639 DAG.getIntPtrConstant(0, DL));
12641 SDValue OpHi = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v4i32, In,
12642 DAG.getIntPtrConstant(4, DL));
12644 OpLo = DAG.getBitcast(MVT::v16i8, OpLo);
12645 OpHi = DAG.getBitcast(MVT::v16i8, OpHi);
12647 // The PSHUFB mask:
12648 static const int ShufMask1[] = {0, 1, 4, 5, 8, 9, 12, 13,
12649 -1, -1, -1, -1, -1, -1, -1, -1};
12651 SDValue Undef = DAG.getUNDEF(MVT::v16i8);
12652 OpLo = DAG.getVectorShuffle(MVT::v16i8, DL, OpLo, Undef, ShufMask1);
12653 OpHi = DAG.getVectorShuffle(MVT::v16i8, DL, OpHi, Undef, ShufMask1);
12655 OpLo = DAG.getBitcast(MVT::v4i32, OpLo);
12656 OpHi = DAG.getBitcast(MVT::v4i32, OpHi);
12658 // The MOVLHPS Mask:
12659 static const int ShufMask2[] = {0, 1, 4, 5};
12660 SDValue res = DAG.getVectorShuffle(MVT::v4i32, DL, OpLo, OpHi, ShufMask2);
12661 return DAG.getBitcast(MVT::v8i16, res);
12664 // Handle truncation of V256 to V128 using shuffles.
12665 if (!VT.is128BitVector() || !InVT.is256BitVector())
12668 assert(Subtarget->hasFp256() && "256-bit vector without AVX!");
12670 unsigned NumElems = VT.getVectorNumElements();
12671 MVT NVT = MVT::getVectorVT(VT.getVectorElementType(), NumElems * 2);
12673 SmallVector<int, 16> MaskVec(NumElems * 2, -1);
12674 // Prepare truncation shuffle mask
12675 for (unsigned i = 0; i != NumElems; ++i)
12676 MaskVec[i] = i * 2;
12677 SDValue V = DAG.getVectorShuffle(NVT, DL, DAG.getBitcast(NVT, In),
12678 DAG.getUNDEF(NVT), &MaskVec[0]);
12679 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, VT, V,
12680 DAG.getIntPtrConstant(0, DL));
12683 SDValue X86TargetLowering::LowerFP_TO_SINT(SDValue Op,
12684 SelectionDAG &DAG) const {
12685 assert(!Op.getSimpleValueType().isVector());
12687 std::pair<SDValue,SDValue> Vals = FP_TO_INTHelper(Op, DAG,
12688 /*IsSigned=*/ true, /*IsReplace=*/ false);
12689 SDValue FIST = Vals.first, StackSlot = Vals.second;
12690 // If FP_TO_INTHelper failed, the node is actually supposed to be Legal.
12691 if (!FIST.getNode()) return Op;
12693 if (StackSlot.getNode())
12694 // Load the result.
12695 return DAG.getLoad(Op.getValueType(), SDLoc(Op),
12696 FIST, StackSlot, MachinePointerInfo(),
12697 false, false, false, 0);
12699 // The node is the result.
12703 SDValue X86TargetLowering::LowerFP_TO_UINT(SDValue Op,
12704 SelectionDAG &DAG) const {
12705 std::pair<SDValue,SDValue> Vals = FP_TO_INTHelper(Op, DAG,
12706 /*IsSigned=*/ false, /*IsReplace=*/ false);
12707 SDValue FIST = Vals.first, StackSlot = Vals.second;
12708 assert(FIST.getNode() && "Unexpected failure");
12710 if (StackSlot.getNode())
12711 // Load the result.
12712 return DAG.getLoad(Op.getValueType(), SDLoc(Op),
12713 FIST, StackSlot, MachinePointerInfo(),
12714 false, false, false, 0);
12716 // The node is the result.
12720 static SDValue LowerFP_EXTEND(SDValue Op, SelectionDAG &DAG) {
12722 MVT VT = Op.getSimpleValueType();
12723 SDValue In = Op.getOperand(0);
12724 MVT SVT = In.getSimpleValueType();
12726 assert(SVT == MVT::v2f32 && "Only customize MVT::v2f32 type legalization!");
12728 return DAG.getNode(X86ISD::VFPEXT, DL, VT,
12729 DAG.getNode(ISD::CONCAT_VECTORS, DL, MVT::v4f32,
12730 In, DAG.getUNDEF(SVT)));
12733 /// The only differences between FABS and FNEG are the mask and the logic op.
12734 /// FNEG also has a folding opportunity for FNEG(FABS(x)).
12735 static SDValue LowerFABSorFNEG(SDValue Op, SelectionDAG &DAG) {
12736 assert((Op.getOpcode() == ISD::FABS || Op.getOpcode() == ISD::FNEG) &&
12737 "Wrong opcode for lowering FABS or FNEG.");
12739 bool IsFABS = (Op.getOpcode() == ISD::FABS);
12741 // If this is a FABS and it has an FNEG user, bail out to fold the combination
12742 // into an FNABS. We'll lower the FABS after that if it is still in use.
12744 for (SDNode *User : Op->uses())
12745 if (User->getOpcode() == ISD::FNEG)
12749 MVT VT = Op.getSimpleValueType();
12751 // FIXME: Use function attribute "OptimizeForSize" and/or CodeGenOpt::Level to
12752 // decide if we should generate a 16-byte constant mask when we only need 4 or
12753 // 8 bytes for the scalar case.
12759 if (VT.isVector()) {
12761 EltVT = VT.getVectorElementType();
12762 NumElts = VT.getVectorNumElements();
12764 // There are no scalar bitwise logical SSE/AVX instructions, so we
12765 // generate a 16-byte vector constant and logic op even for the scalar case.
12766 // Using a 16-byte mask allows folding the load of the mask with
12767 // the logic op, so it can save (~4 bytes) on code size.
12768 LogicVT = (VT == MVT::f64) ? MVT::v2f64 : MVT::v4f32;
12770 NumElts = (VT == MVT::f64) ? 2 : 4;
12773 unsigned EltBits = EltVT.getSizeInBits();
12774 LLVMContext *Context = DAG.getContext();
12775 // For FABS, mask is 0x7f...; for FNEG, mask is 0x80...
12777 IsFABS ? APInt::getSignedMaxValue(EltBits) : APInt::getSignBit(EltBits);
12778 Constant *C = ConstantInt::get(*Context, MaskElt);
12779 C = ConstantVector::getSplat(NumElts, C);
12780 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
12781 SDValue CPIdx = DAG.getConstantPool(C, TLI.getPointerTy(DAG.getDataLayout()));
12782 unsigned Alignment = cast<ConstantPoolSDNode>(CPIdx)->getAlignment();
12784 DAG.getLoad(LogicVT, dl, DAG.getEntryNode(), CPIdx,
12785 MachinePointerInfo::getConstantPool(DAG.getMachineFunction()),
12786 false, false, false, Alignment);
12788 SDValue Op0 = Op.getOperand(0);
12789 bool IsFNABS = !IsFABS && (Op0.getOpcode() == ISD::FABS);
12791 IsFABS ? X86ISD::FAND : IsFNABS ? X86ISD::FOR : X86ISD::FXOR;
12792 SDValue Operand = IsFNABS ? Op0.getOperand(0) : Op0;
12795 return DAG.getNode(LogicOp, dl, LogicVT, Operand, Mask);
12797 // For the scalar case extend to a 128-bit vector, perform the logic op,
12798 // and extract the scalar result back out.
12799 Operand = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, LogicVT, Operand);
12800 SDValue LogicNode = DAG.getNode(LogicOp, dl, LogicVT, Operand, Mask);
12801 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, LogicNode,
12802 DAG.getIntPtrConstant(0, dl));
12805 static SDValue LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) {
12806 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
12807 LLVMContext *Context = DAG.getContext();
12808 SDValue Op0 = Op.getOperand(0);
12809 SDValue Op1 = Op.getOperand(1);
12811 MVT VT = Op.getSimpleValueType();
12812 MVT SrcVT = Op1.getSimpleValueType();
12814 // If second operand is smaller, extend it first.
12815 if (SrcVT.bitsLT(VT)) {
12816 Op1 = DAG.getNode(ISD::FP_EXTEND, dl, VT, Op1);
12819 // And if it is bigger, shrink it first.
12820 if (SrcVT.bitsGT(VT)) {
12821 Op1 = DAG.getNode(ISD::FP_ROUND, dl, VT, Op1, DAG.getIntPtrConstant(1, dl));
12825 // At this point the operands and the result should have the same
12826 // type, and that won't be f80 since that is not custom lowered.
12828 const fltSemantics &Sem =
12829 VT == MVT::f64 ? APFloat::IEEEdouble : APFloat::IEEEsingle;
12830 const unsigned SizeInBits = VT.getSizeInBits();
12832 SmallVector<Constant *, 4> CV(
12833 VT == MVT::f64 ? 2 : 4,
12834 ConstantFP::get(*Context, APFloat(Sem, APInt(SizeInBits, 0))));
12836 // First, clear all bits but the sign bit from the second operand (sign).
12837 CV[0] = ConstantFP::get(*Context,
12838 APFloat(Sem, APInt::getHighBitsSet(SizeInBits, 1)));
12839 Constant *C = ConstantVector::get(CV);
12840 auto PtrVT = TLI.getPointerTy(DAG.getDataLayout());
12841 SDValue CPIdx = DAG.getConstantPool(C, PtrVT, 16);
12843 // Perform all logic operations as 16-byte vectors because there are no
12844 // scalar FP logic instructions in SSE. This allows load folding of the
12845 // constants into the logic instructions.
12846 MVT LogicVT = (VT == MVT::f64) ? MVT::v2f64 : MVT::v4f32;
12848 DAG.getLoad(LogicVT, dl, DAG.getEntryNode(), CPIdx,
12849 MachinePointerInfo::getConstantPool(DAG.getMachineFunction()),
12850 false, false, false, 16);
12851 Op1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, LogicVT, Op1);
12852 SDValue SignBit = DAG.getNode(X86ISD::FAND, dl, LogicVT, Op1, Mask1);
12854 // Next, clear the sign bit from the first operand (magnitude).
12855 // If it's a constant, we can clear it here.
12856 if (ConstantFPSDNode *Op0CN = dyn_cast<ConstantFPSDNode>(Op0)) {
12857 APFloat APF = Op0CN->getValueAPF();
12858 // If the magnitude is a positive zero, the sign bit alone is enough.
12859 if (APF.isPosZero())
12860 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, SrcVT, SignBit,
12861 DAG.getIntPtrConstant(0, dl));
12863 CV[0] = ConstantFP::get(*Context, APF);
12865 CV[0] = ConstantFP::get(
12867 APFloat(Sem, APInt::getLowBitsSet(SizeInBits, SizeInBits - 1)));
12869 C = ConstantVector::get(CV);
12870 CPIdx = DAG.getConstantPool(C, PtrVT, 16);
12872 DAG.getLoad(LogicVT, dl, DAG.getEntryNode(), CPIdx,
12873 MachinePointerInfo::getConstantPool(DAG.getMachineFunction()),
12874 false, false, false, 16);
12875 // If the magnitude operand wasn't a constant, we need to AND out the sign.
12876 if (!isa<ConstantFPSDNode>(Op0)) {
12877 Op0 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, LogicVT, Op0);
12878 Val = DAG.getNode(X86ISD::FAND, dl, LogicVT, Op0, Val);
12880 // OR the magnitude value with the sign bit.
12881 Val = DAG.getNode(X86ISD::FOR, dl, LogicVT, Val, SignBit);
12882 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, SrcVT, Val,
12883 DAG.getIntPtrConstant(0, dl));
12886 static SDValue LowerFGETSIGN(SDValue Op, SelectionDAG &DAG) {
12887 SDValue N0 = Op.getOperand(0);
12889 MVT VT = Op.getSimpleValueType();
12891 // Lower ISD::FGETSIGN to (AND (X86ISD::FGETSIGNx86 ...) 1).
12892 SDValue xFGETSIGN = DAG.getNode(X86ISD::FGETSIGNx86, dl, VT, N0,
12893 DAG.getConstant(1, dl, VT));
12894 return DAG.getNode(ISD::AND, dl, VT, xFGETSIGN, DAG.getConstant(1, dl, VT));
12897 // Check whether an OR'd tree is PTEST-able.
12898 static SDValue LowerVectorAllZeroTest(SDValue Op, const X86Subtarget *Subtarget,
12899 SelectionDAG &DAG) {
12900 assert(Op.getOpcode() == ISD::OR && "Only check OR'd tree.");
12902 if (!Subtarget->hasSSE41())
12905 if (!Op->hasOneUse())
12908 SDNode *N = Op.getNode();
12911 SmallVector<SDValue, 8> Opnds;
12912 DenseMap<SDValue, unsigned> VecInMap;
12913 SmallVector<SDValue, 8> VecIns;
12914 EVT VT = MVT::Other;
12916 // Recognize a special case where a vector is casted into wide integer to
12918 Opnds.push_back(N->getOperand(0));
12919 Opnds.push_back(N->getOperand(1));
12921 for (unsigned Slot = 0, e = Opnds.size(); Slot < e; ++Slot) {
12922 SmallVectorImpl<SDValue>::const_iterator I = Opnds.begin() + Slot;
12923 // BFS traverse all OR'd operands.
12924 if (I->getOpcode() == ISD::OR) {
12925 Opnds.push_back(I->getOperand(0));
12926 Opnds.push_back(I->getOperand(1));
12927 // Re-evaluate the number of nodes to be traversed.
12928 e += 2; // 2 more nodes (LHS and RHS) are pushed.
12932 // Quit if a non-EXTRACT_VECTOR_ELT
12933 if (I->getOpcode() != ISD::EXTRACT_VECTOR_ELT)
12936 // Quit if without a constant index.
12937 SDValue Idx = I->getOperand(1);
12938 if (!isa<ConstantSDNode>(Idx))
12941 SDValue ExtractedFromVec = I->getOperand(0);
12942 DenseMap<SDValue, unsigned>::iterator M = VecInMap.find(ExtractedFromVec);
12943 if (M == VecInMap.end()) {
12944 VT = ExtractedFromVec.getValueType();
12945 // Quit if not 128/256-bit vector.
12946 if (!VT.is128BitVector() && !VT.is256BitVector())
12948 // Quit if not the same type.
12949 if (VecInMap.begin() != VecInMap.end() &&
12950 VT != VecInMap.begin()->first.getValueType())
12952 M = VecInMap.insert(std::make_pair(ExtractedFromVec, 0)).first;
12953 VecIns.push_back(ExtractedFromVec);
12955 M->second |= 1U << cast<ConstantSDNode>(Idx)->getZExtValue();
12958 assert((VT.is128BitVector() || VT.is256BitVector()) &&
12959 "Not extracted from 128-/256-bit vector.");
12961 unsigned FullMask = (1U << VT.getVectorNumElements()) - 1U;
12963 for (DenseMap<SDValue, unsigned>::const_iterator
12964 I = VecInMap.begin(), E = VecInMap.end(); I != E; ++I) {
12965 // Quit if not all elements are used.
12966 if (I->second != FullMask)
12970 EVT TestVT = VT.is128BitVector() ? MVT::v2i64 : MVT::v4i64;
12972 // Cast all vectors into TestVT for PTEST.
12973 for (unsigned i = 0, e = VecIns.size(); i < e; ++i)
12974 VecIns[i] = DAG.getBitcast(TestVT, VecIns[i]);
12976 // If more than one full vectors are evaluated, OR them first before PTEST.
12977 for (unsigned Slot = 0, e = VecIns.size(); e - Slot > 1; Slot += 2, e += 1) {
12978 // Each iteration will OR 2 nodes and append the result until there is only
12979 // 1 node left, i.e. the final OR'd value of all vectors.
12980 SDValue LHS = VecIns[Slot];
12981 SDValue RHS = VecIns[Slot + 1];
12982 VecIns.push_back(DAG.getNode(ISD::OR, DL, TestVT, LHS, RHS));
12985 return DAG.getNode(X86ISD::PTEST, DL, MVT::i32,
12986 VecIns.back(), VecIns.back());
12989 /// \brief return true if \c Op has a use that doesn't just read flags.
12990 static bool hasNonFlagsUse(SDValue Op) {
12991 for (SDNode::use_iterator UI = Op->use_begin(), UE = Op->use_end(); UI != UE;
12993 SDNode *User = *UI;
12994 unsigned UOpNo = UI.getOperandNo();
12995 if (User->getOpcode() == ISD::TRUNCATE && User->hasOneUse()) {
12996 // Look pass truncate.
12997 UOpNo = User->use_begin().getOperandNo();
12998 User = *User->use_begin();
13001 if (User->getOpcode() != ISD::BRCOND && User->getOpcode() != ISD::SETCC &&
13002 !(User->getOpcode() == ISD::SELECT && UOpNo == 0))
13008 /// Emit nodes that will be selected as "test Op0,Op0", or something
13010 SDValue X86TargetLowering::EmitTest(SDValue Op, unsigned X86CC, SDLoc dl,
13011 SelectionDAG &DAG) const {
13012 if (Op.getValueType() == MVT::i1) {
13013 SDValue ExtOp = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i8, Op);
13014 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, ExtOp,
13015 DAG.getConstant(0, dl, MVT::i8));
13017 // CF and OF aren't always set the way we want. Determine which
13018 // of these we need.
13019 bool NeedCF = false;
13020 bool NeedOF = false;
13023 case X86::COND_A: case X86::COND_AE:
13024 case X86::COND_B: case X86::COND_BE:
13027 case X86::COND_G: case X86::COND_GE:
13028 case X86::COND_L: case X86::COND_LE:
13029 case X86::COND_O: case X86::COND_NO: {
13030 // Check if we really need to set the
13031 // Overflow flag. If NoSignedWrap is present
13032 // that is not actually needed.
13033 switch (Op->getOpcode()) {
13038 const auto *BinNode = cast<BinaryWithFlagsSDNode>(Op.getNode());
13039 if (BinNode->Flags.hasNoSignedWrap())
13049 // See if we can use the EFLAGS value from the operand instead of
13050 // doing a separate TEST. TEST always sets OF and CF to 0, so unless
13051 // we prove that the arithmetic won't overflow, we can't use OF or CF.
13052 if (Op.getResNo() != 0 || NeedOF || NeedCF) {
13053 // Emit a CMP with 0, which is the TEST pattern.
13054 //if (Op.getValueType() == MVT::i1)
13055 // return DAG.getNode(X86ISD::CMP, dl, MVT::i1, Op,
13056 // DAG.getConstant(0, MVT::i1));
13057 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op,
13058 DAG.getConstant(0, dl, Op.getValueType()));
13060 unsigned Opcode = 0;
13061 unsigned NumOperands = 0;
13063 // Truncate operations may prevent the merge of the SETCC instruction
13064 // and the arithmetic instruction before it. Attempt to truncate the operands
13065 // of the arithmetic instruction and use a reduced bit-width instruction.
13066 bool NeedTruncation = false;
13067 SDValue ArithOp = Op;
13068 if (Op->getOpcode() == ISD::TRUNCATE && Op->hasOneUse()) {
13069 SDValue Arith = Op->getOperand(0);
13070 // Both the trunc and the arithmetic op need to have one user each.
13071 if (Arith->hasOneUse())
13072 switch (Arith.getOpcode()) {
13079 NeedTruncation = true;
13085 // NOTICE: In the code below we use ArithOp to hold the arithmetic operation
13086 // which may be the result of a CAST. We use the variable 'Op', which is the
13087 // non-casted variable when we check for possible users.
13088 switch (ArithOp.getOpcode()) {
13090 // Due to an isel shortcoming, be conservative if this add is likely to be
13091 // selected as part of a load-modify-store instruction. When the root node
13092 // in a match is a store, isel doesn't know how to remap non-chain non-flag
13093 // uses of other nodes in the match, such as the ADD in this case. This
13094 // leads to the ADD being left around and reselected, with the result being
13095 // two adds in the output. Alas, even if none our users are stores, that
13096 // doesn't prove we're O.K. Ergo, if we have any parents that aren't
13097 // CopyToReg or SETCC, eschew INC/DEC. A better fix seems to require
13098 // climbing the DAG back to the root, and it doesn't seem to be worth the
13100 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
13101 UE = Op.getNode()->use_end(); UI != UE; ++UI)
13102 if (UI->getOpcode() != ISD::CopyToReg &&
13103 UI->getOpcode() != ISD::SETCC &&
13104 UI->getOpcode() != ISD::STORE)
13107 if (ConstantSDNode *C =
13108 dyn_cast<ConstantSDNode>(ArithOp.getNode()->getOperand(1))) {
13109 // An add of one will be selected as an INC.
13110 if (C->getAPIntValue() == 1 && !Subtarget->slowIncDec()) {
13111 Opcode = X86ISD::INC;
13116 // An add of negative one (subtract of one) will be selected as a DEC.
13117 if (C->getAPIntValue().isAllOnesValue() && !Subtarget->slowIncDec()) {
13118 Opcode = X86ISD::DEC;
13124 // Otherwise use a regular EFLAGS-setting add.
13125 Opcode = X86ISD::ADD;
13130 // If we have a constant logical shift that's only used in a comparison
13131 // against zero turn it into an equivalent AND. This allows turning it into
13132 // a TEST instruction later.
13133 if ((X86CC == X86::COND_E || X86CC == X86::COND_NE) && Op->hasOneUse() &&
13134 isa<ConstantSDNode>(Op->getOperand(1)) && !hasNonFlagsUse(Op)) {
13135 EVT VT = Op.getValueType();
13136 unsigned BitWidth = VT.getSizeInBits();
13137 unsigned ShAmt = Op->getConstantOperandVal(1);
13138 if (ShAmt >= BitWidth) // Avoid undefined shifts.
13140 APInt Mask = ArithOp.getOpcode() == ISD::SRL
13141 ? APInt::getHighBitsSet(BitWidth, BitWidth - ShAmt)
13142 : APInt::getLowBitsSet(BitWidth, BitWidth - ShAmt);
13143 if (!Mask.isSignedIntN(32)) // Avoid large immediates.
13145 SDValue New = DAG.getNode(ISD::AND, dl, VT, Op->getOperand(0),
13146 DAG.getConstant(Mask, dl, VT));
13147 DAG.ReplaceAllUsesWith(Op, New);
13153 // If the primary and result isn't used, don't bother using X86ISD::AND,
13154 // because a TEST instruction will be better.
13155 if (!hasNonFlagsUse(Op))
13161 // Due to the ISEL shortcoming noted above, be conservative if this op is
13162 // likely to be selected as part of a load-modify-store instruction.
13163 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
13164 UE = Op.getNode()->use_end(); UI != UE; ++UI)
13165 if (UI->getOpcode() == ISD::STORE)
13168 // Otherwise use a regular EFLAGS-setting instruction.
13169 switch (ArithOp.getOpcode()) {
13170 default: llvm_unreachable("unexpected operator!");
13171 case ISD::SUB: Opcode = X86ISD::SUB; break;
13172 case ISD::XOR: Opcode = X86ISD::XOR; break;
13173 case ISD::AND: Opcode = X86ISD::AND; break;
13175 if (!NeedTruncation && (X86CC == X86::COND_E || X86CC == X86::COND_NE)) {
13176 SDValue EFLAGS = LowerVectorAllZeroTest(Op, Subtarget, DAG);
13177 if (EFLAGS.getNode())
13180 Opcode = X86ISD::OR;
13194 return SDValue(Op.getNode(), 1);
13200 // If we found that truncation is beneficial, perform the truncation and
13202 if (NeedTruncation) {
13203 EVT VT = Op.getValueType();
13204 SDValue WideVal = Op->getOperand(0);
13205 EVT WideVT = WideVal.getValueType();
13206 unsigned ConvertedOp = 0;
13207 // Use a target machine opcode to prevent further DAGCombine
13208 // optimizations that may separate the arithmetic operations
13209 // from the setcc node.
13210 switch (WideVal.getOpcode()) {
13212 case ISD::ADD: ConvertedOp = X86ISD::ADD; break;
13213 case ISD::SUB: ConvertedOp = X86ISD::SUB; break;
13214 case ISD::AND: ConvertedOp = X86ISD::AND; break;
13215 case ISD::OR: ConvertedOp = X86ISD::OR; break;
13216 case ISD::XOR: ConvertedOp = X86ISD::XOR; break;
13220 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
13221 if (TLI.isOperationLegal(WideVal.getOpcode(), WideVT)) {
13222 SDValue V0 = DAG.getNode(ISD::TRUNCATE, dl, VT, WideVal.getOperand(0));
13223 SDValue V1 = DAG.getNode(ISD::TRUNCATE, dl, VT, WideVal.getOperand(1));
13224 Op = DAG.getNode(ConvertedOp, dl, VT, V0, V1);
13230 // Emit a CMP with 0, which is the TEST pattern.
13231 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op,
13232 DAG.getConstant(0, dl, Op.getValueType()));
13234 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::i32);
13235 SmallVector<SDValue, 4> Ops(Op->op_begin(), Op->op_begin() + NumOperands);
13237 SDValue New = DAG.getNode(Opcode, dl, VTs, Ops);
13238 DAG.ReplaceAllUsesWith(Op, New);
13239 return SDValue(New.getNode(), 1);
13242 /// Emit nodes that will be selected as "cmp Op0,Op1", or something
13244 SDValue X86TargetLowering::EmitCmp(SDValue Op0, SDValue Op1, unsigned X86CC,
13245 SDLoc dl, SelectionDAG &DAG) const {
13246 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op1)) {
13247 if (C->getAPIntValue() == 0)
13248 return EmitTest(Op0, X86CC, dl, DAG);
13250 if (Op0.getValueType() == MVT::i1)
13251 llvm_unreachable("Unexpected comparison operation for MVT::i1 operands");
13254 if ((Op0.getValueType() == MVT::i8 || Op0.getValueType() == MVT::i16 ||
13255 Op0.getValueType() == MVT::i32 || Op0.getValueType() == MVT::i64)) {
13256 // Do the comparison at i32 if it's smaller, besides the Atom case.
13257 // This avoids subregister aliasing issues. Keep the smaller reference
13258 // if we're optimizing for size, however, as that'll allow better folding
13259 // of memory operations.
13260 if (Op0.getValueType() != MVT::i32 && Op0.getValueType() != MVT::i64 &&
13261 !DAG.getMachineFunction().getFunction()->optForMinSize() &&
13262 !Subtarget->isAtom()) {
13263 unsigned ExtendOp =
13264 isX86CCUnsigned(X86CC) ? ISD::ZERO_EXTEND : ISD::SIGN_EXTEND;
13265 Op0 = DAG.getNode(ExtendOp, dl, MVT::i32, Op0);
13266 Op1 = DAG.getNode(ExtendOp, dl, MVT::i32, Op1);
13268 // Use SUB instead of CMP to enable CSE between SUB and CMP.
13269 SDVTList VTs = DAG.getVTList(Op0.getValueType(), MVT::i32);
13270 SDValue Sub = DAG.getNode(X86ISD::SUB, dl, VTs,
13272 return SDValue(Sub.getNode(), 1);
13274 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op0, Op1);
13277 /// Convert a comparison if required by the subtarget.
13278 SDValue X86TargetLowering::ConvertCmpIfNecessary(SDValue Cmp,
13279 SelectionDAG &DAG) const {
13280 // If the subtarget does not support the FUCOMI instruction, floating-point
13281 // comparisons have to be converted.
13282 if (Subtarget->hasCMov() ||
13283 Cmp.getOpcode() != X86ISD::CMP ||
13284 !Cmp.getOperand(0).getValueType().isFloatingPoint() ||
13285 !Cmp.getOperand(1).getValueType().isFloatingPoint())
13288 // The instruction selector will select an FUCOM instruction instead of
13289 // FUCOMI, which writes the comparison result to FPSW instead of EFLAGS. Hence
13290 // build an SDNode sequence that transfers the result from FPSW into EFLAGS:
13291 // (X86sahf (trunc (srl (X86fp_stsw (trunc (X86cmp ...)), 8))))
13293 SDValue TruncFPSW = DAG.getNode(ISD::TRUNCATE, dl, MVT::i16, Cmp);
13294 SDValue FNStSW = DAG.getNode(X86ISD::FNSTSW16r, dl, MVT::i16, TruncFPSW);
13295 SDValue Srl = DAG.getNode(ISD::SRL, dl, MVT::i16, FNStSW,
13296 DAG.getConstant(8, dl, MVT::i8));
13297 SDValue TruncSrl = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Srl);
13298 return DAG.getNode(X86ISD::SAHF, dl, MVT::i32, TruncSrl);
13301 /// The minimum architected relative accuracy is 2^-12. We need one
13302 /// Newton-Raphson step to have a good float result (24 bits of precision).
13303 SDValue X86TargetLowering::getRsqrtEstimate(SDValue Op,
13304 DAGCombinerInfo &DCI,
13305 unsigned &RefinementSteps,
13306 bool &UseOneConstNR) const {
13307 EVT VT = Op.getValueType();
13308 const char *RecipOp;
13310 // SSE1 has rsqrtss and rsqrtps. AVX adds a 256-bit variant for rsqrtps.
13311 // TODO: Add support for AVX512 (v16f32).
13312 // It is likely not profitable to do this for f64 because a double-precision
13313 // rsqrt estimate with refinement on x86 prior to FMA requires at least 16
13314 // instructions: convert to single, rsqrtss, convert back to double, refine
13315 // (3 steps = at least 13 insts). If an 'rsqrtsd' variant was added to the ISA
13316 // along with FMA, this could be a throughput win.
13317 if (VT == MVT::f32 && Subtarget->hasSSE1())
13319 else if ((VT == MVT::v4f32 && Subtarget->hasSSE1()) ||
13320 (VT == MVT::v8f32 && Subtarget->hasAVX()))
13321 RecipOp = "vec-sqrtf";
13325 TargetRecip Recips = DCI.DAG.getTarget().Options.Reciprocals;
13326 if (!Recips.isEnabled(RecipOp))
13329 RefinementSteps = Recips.getRefinementSteps(RecipOp);
13330 UseOneConstNR = false;
13331 return DCI.DAG.getNode(X86ISD::FRSQRT, SDLoc(Op), VT, Op);
13334 /// The minimum architected relative accuracy is 2^-12. We need one
13335 /// Newton-Raphson step to have a good float result (24 bits of precision).
13336 SDValue X86TargetLowering::getRecipEstimate(SDValue Op,
13337 DAGCombinerInfo &DCI,
13338 unsigned &RefinementSteps) const {
13339 EVT VT = Op.getValueType();
13340 const char *RecipOp;
13342 // SSE1 has rcpss and rcpps. AVX adds a 256-bit variant for rcpps.
13343 // TODO: Add support for AVX512 (v16f32).
13344 // It is likely not profitable to do this for f64 because a double-precision
13345 // reciprocal estimate with refinement on x86 prior to FMA requires
13346 // 15 instructions: convert to single, rcpss, convert back to double, refine
13347 // (3 steps = 12 insts). If an 'rcpsd' variant was added to the ISA
13348 // along with FMA, this could be a throughput win.
13349 if (VT == MVT::f32 && Subtarget->hasSSE1())
13351 else if ((VT == MVT::v4f32 && Subtarget->hasSSE1()) ||
13352 (VT == MVT::v8f32 && Subtarget->hasAVX()))
13353 RecipOp = "vec-divf";
13357 TargetRecip Recips = DCI.DAG.getTarget().Options.Reciprocals;
13358 if (!Recips.isEnabled(RecipOp))
13361 RefinementSteps = Recips.getRefinementSteps(RecipOp);
13362 return DCI.DAG.getNode(X86ISD::FRCP, SDLoc(Op), VT, Op);
13365 /// If we have at least two divisions that use the same divisor, convert to
13366 /// multplication by a reciprocal. This may need to be adjusted for a given
13367 /// CPU if a division's cost is not at least twice the cost of a multiplication.
13368 /// This is because we still need one division to calculate the reciprocal and
13369 /// then we need two multiplies by that reciprocal as replacements for the
13370 /// original divisions.
13371 unsigned X86TargetLowering::combineRepeatedFPDivisors() const {
13375 static bool isAllOnes(SDValue V) {
13376 ConstantSDNode *C = dyn_cast<ConstantSDNode>(V);
13377 return C && C->isAllOnesValue();
13380 /// LowerToBT - Result of 'and' is compared against zero. Turn it into a BT node
13381 /// if it's possible.
13382 SDValue X86TargetLowering::LowerToBT(SDValue And, ISD::CondCode CC,
13383 SDLoc dl, SelectionDAG &DAG) const {
13384 SDValue Op0 = And.getOperand(0);
13385 SDValue Op1 = And.getOperand(1);
13386 if (Op0.getOpcode() == ISD::TRUNCATE)
13387 Op0 = Op0.getOperand(0);
13388 if (Op1.getOpcode() == ISD::TRUNCATE)
13389 Op1 = Op1.getOperand(0);
13392 if (Op1.getOpcode() == ISD::SHL)
13393 std::swap(Op0, Op1);
13394 if (Op0.getOpcode() == ISD::SHL) {
13395 if (ConstantSDNode *And00C = dyn_cast<ConstantSDNode>(Op0.getOperand(0)))
13396 if (And00C->getZExtValue() == 1) {
13397 // If we looked past a truncate, check that it's only truncating away
13399 unsigned BitWidth = Op0.getValueSizeInBits();
13400 unsigned AndBitWidth = And.getValueSizeInBits();
13401 if (BitWidth > AndBitWidth) {
13403 DAG.computeKnownBits(Op0, Zeros, Ones);
13404 if (Zeros.countLeadingOnes() < BitWidth - AndBitWidth)
13408 RHS = Op0.getOperand(1);
13410 } else if (Op1.getOpcode() == ISD::Constant) {
13411 ConstantSDNode *AndRHS = cast<ConstantSDNode>(Op1);
13412 uint64_t AndRHSVal = AndRHS->getZExtValue();
13413 SDValue AndLHS = Op0;
13415 if (AndRHSVal == 1 && AndLHS.getOpcode() == ISD::SRL) {
13416 LHS = AndLHS.getOperand(0);
13417 RHS = AndLHS.getOperand(1);
13420 // Use BT if the immediate can't be encoded in a TEST instruction.
13421 if (!isUInt<32>(AndRHSVal) && isPowerOf2_64(AndRHSVal)) {
13423 RHS = DAG.getConstant(Log2_64_Ceil(AndRHSVal), dl, LHS.getValueType());
13427 if (LHS.getNode()) {
13428 // If LHS is i8, promote it to i32 with any_extend. There is no i8 BT
13429 // instruction. Since the shift amount is in-range-or-undefined, we know
13430 // that doing a bittest on the i32 value is ok. We extend to i32 because
13431 // the encoding for the i16 version is larger than the i32 version.
13432 // Also promote i16 to i32 for performance / code size reason.
13433 if (LHS.getValueType() == MVT::i8 ||
13434 LHS.getValueType() == MVT::i16)
13435 LHS = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, LHS);
13437 // If the operand types disagree, extend the shift amount to match. Since
13438 // BT ignores high bits (like shifts) we can use anyextend.
13439 if (LHS.getValueType() != RHS.getValueType())
13440 RHS = DAG.getNode(ISD::ANY_EXTEND, dl, LHS.getValueType(), RHS);
13442 SDValue BT = DAG.getNode(X86ISD::BT, dl, MVT::i32, LHS, RHS);
13443 X86::CondCode Cond = CC == ISD::SETEQ ? X86::COND_AE : X86::COND_B;
13444 return DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
13445 DAG.getConstant(Cond, dl, MVT::i8), BT);
13451 /// \brief - Turns an ISD::CondCode into a value suitable for SSE floating point
13453 static int translateX86FSETCC(ISD::CondCode SetCCOpcode, SDValue &Op0,
13458 // SSE Condition code mapping:
13467 switch (SetCCOpcode) {
13468 default: llvm_unreachable("Unexpected SETCC condition");
13470 case ISD::SETEQ: SSECC = 0; break;
13472 case ISD::SETGT: Swap = true; // Fallthrough
13474 case ISD::SETOLT: SSECC = 1; break;
13476 case ISD::SETGE: Swap = true; // Fallthrough
13478 case ISD::SETOLE: SSECC = 2; break;
13479 case ISD::SETUO: SSECC = 3; break;
13481 case ISD::SETNE: SSECC = 4; break;
13482 case ISD::SETULE: Swap = true; // Fallthrough
13483 case ISD::SETUGE: SSECC = 5; break;
13484 case ISD::SETULT: Swap = true; // Fallthrough
13485 case ISD::SETUGT: SSECC = 6; break;
13486 case ISD::SETO: SSECC = 7; break;
13488 case ISD::SETONE: SSECC = 8; break;
13491 std::swap(Op0, Op1);
13496 // Lower256IntVSETCC - Break a VSETCC 256-bit integer VSETCC into two new 128
13497 // ones, and then concatenate the result back.
13498 static SDValue Lower256IntVSETCC(SDValue Op, SelectionDAG &DAG) {
13499 MVT VT = Op.getSimpleValueType();
13501 assert(VT.is256BitVector() && Op.getOpcode() == ISD::SETCC &&
13502 "Unsupported value type for operation");
13504 unsigned NumElems = VT.getVectorNumElements();
13506 SDValue CC = Op.getOperand(2);
13508 // Extract the LHS vectors
13509 SDValue LHS = Op.getOperand(0);
13510 SDValue LHS1 = Extract128BitVector(LHS, 0, DAG, dl);
13511 SDValue LHS2 = Extract128BitVector(LHS, NumElems/2, DAG, dl);
13513 // Extract the RHS vectors
13514 SDValue RHS = Op.getOperand(1);
13515 SDValue RHS1 = Extract128BitVector(RHS, 0, DAG, dl);
13516 SDValue RHS2 = Extract128BitVector(RHS, NumElems/2, DAG, dl);
13518 // Issue the operation on the smaller types and concatenate the result back
13519 MVT EltVT = VT.getVectorElementType();
13520 MVT NewVT = MVT::getVectorVT(EltVT, NumElems/2);
13521 return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT,
13522 DAG.getNode(Op.getOpcode(), dl, NewVT, LHS1, RHS1, CC),
13523 DAG.getNode(Op.getOpcode(), dl, NewVT, LHS2, RHS2, CC));
13526 static SDValue LowerBoolVSETCC_AVX512(SDValue Op, SelectionDAG &DAG) {
13527 SDValue Op0 = Op.getOperand(0);
13528 SDValue Op1 = Op.getOperand(1);
13529 SDValue CC = Op.getOperand(2);
13530 MVT VT = Op.getSimpleValueType();
13533 assert(Op0.getValueType().getVectorElementType() == MVT::i1 &&
13534 "Unexpected type for boolean compare operation");
13535 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
13536 SDValue NotOp0 = DAG.getNode(ISD::XOR, dl, VT, Op0,
13537 DAG.getConstant(-1, dl, VT));
13538 SDValue NotOp1 = DAG.getNode(ISD::XOR, dl, VT, Op1,
13539 DAG.getConstant(-1, dl, VT));
13540 switch (SetCCOpcode) {
13541 default: llvm_unreachable("Unexpected SETCC condition");
13543 // (x == y) -> ~(x ^ y)
13544 return DAG.getNode(ISD::XOR, dl, VT,
13545 DAG.getNode(ISD::XOR, dl, VT, Op0, Op1),
13546 DAG.getConstant(-1, dl, VT));
13548 // (x != y) -> (x ^ y)
13549 return DAG.getNode(ISD::XOR, dl, VT, Op0, Op1);
13552 // (x > y) -> (x & ~y)
13553 return DAG.getNode(ISD::AND, dl, VT, Op0, NotOp1);
13556 // (x < y) -> (~x & y)
13557 return DAG.getNode(ISD::AND, dl, VT, NotOp0, Op1);
13560 // (x <= y) -> (~x | y)
13561 return DAG.getNode(ISD::OR, dl, VT, NotOp0, Op1);
13564 // (x >=y) -> (x | ~y)
13565 return DAG.getNode(ISD::OR, dl, VT, Op0, NotOp1);
13569 static SDValue LowerIntVSETCC_AVX512(SDValue Op, SelectionDAG &DAG,
13570 const X86Subtarget *Subtarget) {
13571 SDValue Op0 = Op.getOperand(0);
13572 SDValue Op1 = Op.getOperand(1);
13573 SDValue CC = Op.getOperand(2);
13574 MVT VT = Op.getSimpleValueType();
13577 assert(Op0.getValueType().getVectorElementType().getSizeInBits() >= 8 &&
13578 Op.getValueType().getScalarType() == MVT::i1 &&
13579 "Cannot set masked compare for this operation");
13581 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
13583 bool Unsigned = false;
13586 switch (SetCCOpcode) {
13587 default: llvm_unreachable("Unexpected SETCC condition");
13588 case ISD::SETNE: SSECC = 4; break;
13589 case ISD::SETEQ: Opc = X86ISD::PCMPEQM; break;
13590 case ISD::SETUGT: SSECC = 6; Unsigned = true; break;
13591 case ISD::SETLT: Swap = true; //fall-through
13592 case ISD::SETGT: Opc = X86ISD::PCMPGTM; break;
13593 case ISD::SETULT: SSECC = 1; Unsigned = true; break;
13594 case ISD::SETUGE: SSECC = 5; Unsigned = true; break; //NLT
13595 case ISD::SETGE: Swap = true; SSECC = 2; break; // LE + swap
13596 case ISD::SETULE: Unsigned = true; //fall-through
13597 case ISD::SETLE: SSECC = 2; break;
13601 std::swap(Op0, Op1);
13603 return DAG.getNode(Opc, dl, VT, Op0, Op1);
13604 Opc = Unsigned ? X86ISD::CMPMU: X86ISD::CMPM;
13605 return DAG.getNode(Opc, dl, VT, Op0, Op1,
13606 DAG.getConstant(SSECC, dl, MVT::i8));
13609 /// \brief Try to turn a VSETULT into a VSETULE by modifying its second
13610 /// operand \p Op1. If non-trivial (for example because it's not constant)
13611 /// return an empty value.
13612 static SDValue ChangeVSETULTtoVSETULE(SDLoc dl, SDValue Op1, SelectionDAG &DAG)
13614 BuildVectorSDNode *BV = dyn_cast<BuildVectorSDNode>(Op1.getNode());
13618 MVT VT = Op1.getSimpleValueType();
13619 MVT EVT = VT.getVectorElementType();
13620 unsigned n = VT.getVectorNumElements();
13621 SmallVector<SDValue, 8> ULTOp1;
13623 for (unsigned i = 0; i < n; ++i) {
13624 ConstantSDNode *Elt = dyn_cast<ConstantSDNode>(BV->getOperand(i));
13625 if (!Elt || Elt->isOpaque() || Elt->getValueType(0) != EVT)
13628 // Avoid underflow.
13629 APInt Val = Elt->getAPIntValue();
13633 ULTOp1.push_back(DAG.getConstant(Val - 1, dl, EVT));
13636 return DAG.getNode(ISD::BUILD_VECTOR, dl, VT, ULTOp1);
13639 static SDValue LowerVSETCC(SDValue Op, const X86Subtarget *Subtarget,
13640 SelectionDAG &DAG) {
13641 SDValue Op0 = Op.getOperand(0);
13642 SDValue Op1 = Op.getOperand(1);
13643 SDValue CC = Op.getOperand(2);
13644 MVT VT = Op.getSimpleValueType();
13645 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
13646 bool isFP = Op.getOperand(1).getSimpleValueType().isFloatingPoint();
13651 MVT EltVT = Op0.getSimpleValueType().getVectorElementType();
13652 assert(EltVT == MVT::f32 || EltVT == MVT::f64);
13655 unsigned SSECC = translateX86FSETCC(SetCCOpcode, Op0, Op1);
13656 unsigned Opc = X86ISD::CMPP;
13657 if (Subtarget->hasAVX512() && VT.getVectorElementType() == MVT::i1) {
13658 assert(VT.getVectorNumElements() <= 16);
13659 Opc = X86ISD::CMPM;
13661 // In the two special cases we can't handle, emit two comparisons.
13664 unsigned CombineOpc;
13665 if (SetCCOpcode == ISD::SETUEQ) {
13666 CC0 = 3; CC1 = 0; CombineOpc = ISD::OR;
13668 assert(SetCCOpcode == ISD::SETONE);
13669 CC0 = 7; CC1 = 4; CombineOpc = ISD::AND;
13672 SDValue Cmp0 = DAG.getNode(Opc, dl, VT, Op0, Op1,
13673 DAG.getConstant(CC0, dl, MVT::i8));
13674 SDValue Cmp1 = DAG.getNode(Opc, dl, VT, Op0, Op1,
13675 DAG.getConstant(CC1, dl, MVT::i8));
13676 return DAG.getNode(CombineOpc, dl, VT, Cmp0, Cmp1);
13678 // Handle all other FP comparisons here.
13679 return DAG.getNode(Opc, dl, VT, Op0, Op1,
13680 DAG.getConstant(SSECC, dl, MVT::i8));
13683 // Break 256-bit integer vector compare into smaller ones.
13684 if (VT.is256BitVector() && !Subtarget->hasInt256())
13685 return Lower256IntVSETCC(Op, DAG);
13687 EVT OpVT = Op1.getValueType();
13688 if (OpVT.getVectorElementType() == MVT::i1)
13689 return LowerBoolVSETCC_AVX512(Op, DAG);
13691 bool MaskResult = (VT.getVectorElementType() == MVT::i1);
13692 if (Subtarget->hasAVX512()) {
13693 if (Op1.getValueType().is512BitVector() ||
13694 (Subtarget->hasBWI() && Subtarget->hasVLX()) ||
13695 (MaskResult && OpVT.getVectorElementType().getSizeInBits() >= 32))
13696 return LowerIntVSETCC_AVX512(Op, DAG, Subtarget);
13698 // In AVX-512 architecture setcc returns mask with i1 elements,
13699 // But there is no compare instruction for i8 and i16 elements in KNL.
13700 // We are not talking about 512-bit operands in this case, these
13701 // types are illegal.
13703 (OpVT.getVectorElementType().getSizeInBits() < 32 &&
13704 OpVT.getVectorElementType().getSizeInBits() >= 8))
13705 return DAG.getNode(ISD::TRUNCATE, dl, VT,
13706 DAG.getNode(ISD::SETCC, dl, OpVT, Op0, Op1, CC));
13709 // We are handling one of the integer comparisons here. Since SSE only has
13710 // GT and EQ comparisons for integer, swapping operands and multiple
13711 // operations may be required for some comparisons.
13713 bool Swap = false, Invert = false, FlipSigns = false, MinMax = false;
13714 bool Subus = false;
13716 switch (SetCCOpcode) {
13717 default: llvm_unreachable("Unexpected SETCC condition");
13718 case ISD::SETNE: Invert = true;
13719 case ISD::SETEQ: Opc = X86ISD::PCMPEQ; break;
13720 case ISD::SETLT: Swap = true;
13721 case ISD::SETGT: Opc = X86ISD::PCMPGT; break;
13722 case ISD::SETGE: Swap = true;
13723 case ISD::SETLE: Opc = X86ISD::PCMPGT;
13724 Invert = true; break;
13725 case ISD::SETULT: Swap = true;
13726 case ISD::SETUGT: Opc = X86ISD::PCMPGT;
13727 FlipSigns = true; break;
13728 case ISD::SETUGE: Swap = true;
13729 case ISD::SETULE: Opc = X86ISD::PCMPGT;
13730 FlipSigns = true; Invert = true; break;
13733 // Special case: Use min/max operations for SETULE/SETUGE
13734 MVT VET = VT.getVectorElementType();
13736 (Subtarget->hasSSE41() && (VET >= MVT::i8 && VET <= MVT::i32))
13737 || (Subtarget->hasSSE2() && (VET == MVT::i8));
13740 switch (SetCCOpcode) {
13742 case ISD::SETULE: Opc = ISD::UMIN; MinMax = true; break;
13743 case ISD::SETUGE: Opc = ISD::UMAX; MinMax = true; break;
13746 if (MinMax) { Swap = false; Invert = false; FlipSigns = false; }
13749 bool hasSubus = Subtarget->hasSSE2() && (VET == MVT::i8 || VET == MVT::i16);
13750 if (!MinMax && hasSubus) {
13751 // As another special case, use PSUBUS[BW] when it's profitable. E.g. for
13753 // t = psubus Op0, Op1
13754 // pcmpeq t, <0..0>
13755 switch (SetCCOpcode) {
13757 case ISD::SETULT: {
13758 // If the comparison is against a constant we can turn this into a
13759 // setule. With psubus, setule does not require a swap. This is
13760 // beneficial because the constant in the register is no longer
13761 // destructed as the destination so it can be hoisted out of a loop.
13762 // Only do this pre-AVX since vpcmp* is no longer destructive.
13763 if (Subtarget->hasAVX())
13765 SDValue ULEOp1 = ChangeVSETULTtoVSETULE(dl, Op1, DAG);
13766 if (ULEOp1.getNode()) {
13768 Subus = true; Invert = false; Swap = false;
13772 // Psubus is better than flip-sign because it requires no inversion.
13773 case ISD::SETUGE: Subus = true; Invert = false; Swap = true; break;
13774 case ISD::SETULE: Subus = true; Invert = false; Swap = false; break;
13778 Opc = X86ISD::SUBUS;
13784 std::swap(Op0, Op1);
13786 // Check that the operation in question is available (most are plain SSE2,
13787 // but PCMPGTQ and PCMPEQQ have different requirements).
13788 if (VT == MVT::v2i64) {
13789 if (Opc == X86ISD::PCMPGT && !Subtarget->hasSSE42()) {
13790 assert(Subtarget->hasSSE2() && "Don't know how to lower!");
13792 // First cast everything to the right type.
13793 Op0 = DAG.getBitcast(MVT::v4i32, Op0);
13794 Op1 = DAG.getBitcast(MVT::v4i32, Op1);
13796 // Since SSE has no unsigned integer comparisons, we need to flip the sign
13797 // bits of the inputs before performing those operations. The lower
13798 // compare is always unsigned.
13801 SB = DAG.getConstant(0x80000000U, dl, MVT::v4i32);
13803 SDValue Sign = DAG.getConstant(0x80000000U, dl, MVT::i32);
13804 SDValue Zero = DAG.getConstant(0x00000000U, dl, MVT::i32);
13805 SB = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32,
13806 Sign, Zero, Sign, Zero);
13808 Op0 = DAG.getNode(ISD::XOR, dl, MVT::v4i32, Op0, SB);
13809 Op1 = DAG.getNode(ISD::XOR, dl, MVT::v4i32, Op1, SB);
13811 // Emulate PCMPGTQ with (hi1 > hi2) | ((hi1 == hi2) & (lo1 > lo2))
13812 SDValue GT = DAG.getNode(X86ISD::PCMPGT, dl, MVT::v4i32, Op0, Op1);
13813 SDValue EQ = DAG.getNode(X86ISD::PCMPEQ, dl, MVT::v4i32, Op0, Op1);
13815 // Create masks for only the low parts/high parts of the 64 bit integers.
13816 static const int MaskHi[] = { 1, 1, 3, 3 };
13817 static const int MaskLo[] = { 0, 0, 2, 2 };
13818 SDValue EQHi = DAG.getVectorShuffle(MVT::v4i32, dl, EQ, EQ, MaskHi);
13819 SDValue GTLo = DAG.getVectorShuffle(MVT::v4i32, dl, GT, GT, MaskLo);
13820 SDValue GTHi = DAG.getVectorShuffle(MVT::v4i32, dl, GT, GT, MaskHi);
13822 SDValue Result = DAG.getNode(ISD::AND, dl, MVT::v4i32, EQHi, GTLo);
13823 Result = DAG.getNode(ISD::OR, dl, MVT::v4i32, Result, GTHi);
13826 Result = DAG.getNOT(dl, Result, MVT::v4i32);
13828 return DAG.getBitcast(VT, Result);
13831 if (Opc == X86ISD::PCMPEQ && !Subtarget->hasSSE41()) {
13832 // If pcmpeqq is missing but pcmpeqd is available synthesize pcmpeqq with
13833 // pcmpeqd + pshufd + pand.
13834 assert(Subtarget->hasSSE2() && !FlipSigns && "Don't know how to lower!");
13836 // First cast everything to the right type.
13837 Op0 = DAG.getBitcast(MVT::v4i32, Op0);
13838 Op1 = DAG.getBitcast(MVT::v4i32, Op1);
13841 SDValue Result = DAG.getNode(Opc, dl, MVT::v4i32, Op0, Op1);
13843 // Make sure the lower and upper halves are both all-ones.
13844 static const int Mask[] = { 1, 0, 3, 2 };
13845 SDValue Shuf = DAG.getVectorShuffle(MVT::v4i32, dl, Result, Result, Mask);
13846 Result = DAG.getNode(ISD::AND, dl, MVT::v4i32, Result, Shuf);
13849 Result = DAG.getNOT(dl, Result, MVT::v4i32);
13851 return DAG.getBitcast(VT, Result);
13855 // Since SSE has no unsigned integer comparisons, we need to flip the sign
13856 // bits of the inputs before performing those operations.
13858 EVT EltVT = VT.getVectorElementType();
13859 SDValue SB = DAG.getConstant(APInt::getSignBit(EltVT.getSizeInBits()), dl,
13861 Op0 = DAG.getNode(ISD::XOR, dl, VT, Op0, SB);
13862 Op1 = DAG.getNode(ISD::XOR, dl, VT, Op1, SB);
13865 SDValue Result = DAG.getNode(Opc, dl, VT, Op0, Op1);
13867 // If the logical-not of the result is required, perform that now.
13869 Result = DAG.getNOT(dl, Result, VT);
13872 Result = DAG.getNode(X86ISD::PCMPEQ, dl, VT, Op0, Result);
13875 Result = DAG.getNode(X86ISD::PCMPEQ, dl, VT, Result,
13876 getZeroVector(VT, Subtarget, DAG, dl));
13881 SDValue X86TargetLowering::LowerSETCC(SDValue Op, SelectionDAG &DAG) const {
13883 MVT VT = Op.getSimpleValueType();
13885 if (VT.isVector()) return LowerVSETCC(Op, Subtarget, DAG);
13887 assert(((!Subtarget->hasAVX512() && VT == MVT::i8) || (VT == MVT::i1))
13888 && "SetCC type must be 8-bit or 1-bit integer");
13889 SDValue Op0 = Op.getOperand(0);
13890 SDValue Op1 = Op.getOperand(1);
13892 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
13894 // Optimize to BT if possible.
13895 // Lower (X & (1 << N)) == 0 to BT(X, N).
13896 // Lower ((X >>u N) & 1) != 0 to BT(X, N).
13897 // Lower ((X >>s N) & 1) != 0 to BT(X, N).
13898 if (Op0.getOpcode() == ISD::AND && Op0.hasOneUse() &&
13899 Op1.getOpcode() == ISD::Constant &&
13900 cast<ConstantSDNode>(Op1)->isNullValue() &&
13901 (CC == ISD::SETEQ || CC == ISD::SETNE)) {
13902 SDValue NewSetCC = LowerToBT(Op0, CC, dl, DAG);
13903 if (NewSetCC.getNode()) {
13905 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i1, NewSetCC);
13910 // Look for X == 0, X == 1, X != 0, or X != 1. We can simplify some forms of
13912 if (Op1.getOpcode() == ISD::Constant &&
13913 (cast<ConstantSDNode>(Op1)->getZExtValue() == 1 ||
13914 cast<ConstantSDNode>(Op1)->isNullValue()) &&
13915 (CC == ISD::SETEQ || CC == ISD::SETNE)) {
13917 // If the input is a setcc, then reuse the input setcc or use a new one with
13918 // the inverted condition.
13919 if (Op0.getOpcode() == X86ISD::SETCC) {
13920 X86::CondCode CCode = (X86::CondCode)Op0.getConstantOperandVal(0);
13921 bool Invert = (CC == ISD::SETNE) ^
13922 cast<ConstantSDNode>(Op1)->isNullValue();
13926 CCode = X86::GetOppositeBranchCondition(CCode);
13927 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
13928 DAG.getConstant(CCode, dl, MVT::i8),
13929 Op0.getOperand(1));
13931 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i1, SetCC);
13935 if ((Op0.getValueType() == MVT::i1) && (Op1.getOpcode() == ISD::Constant) &&
13936 (cast<ConstantSDNode>(Op1)->getZExtValue() == 1) &&
13937 (CC == ISD::SETEQ || CC == ISD::SETNE)) {
13939 ISD::CondCode NewCC = ISD::getSetCCInverse(CC, true);
13940 return DAG.getSetCC(dl, VT, Op0, DAG.getConstant(0, dl, MVT::i1), NewCC);
13943 bool isFP = Op1.getSimpleValueType().isFloatingPoint();
13944 unsigned X86CC = TranslateX86CC(CC, dl, isFP, Op0, Op1, DAG);
13945 if (X86CC == X86::COND_INVALID)
13948 SDValue EFLAGS = EmitCmp(Op0, Op1, X86CC, dl, DAG);
13949 EFLAGS = ConvertCmpIfNecessary(EFLAGS, DAG);
13950 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
13951 DAG.getConstant(X86CC, dl, MVT::i8), EFLAGS);
13953 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i1, SetCC);
13957 // isX86LogicalCmp - Return true if opcode is a X86 logical comparison.
13958 static bool isX86LogicalCmp(SDValue Op) {
13959 unsigned Opc = Op.getNode()->getOpcode();
13960 if (Opc == X86ISD::CMP || Opc == X86ISD::COMI || Opc == X86ISD::UCOMI ||
13961 Opc == X86ISD::SAHF)
13963 if (Op.getResNo() == 1 &&
13964 (Opc == X86ISD::ADD ||
13965 Opc == X86ISD::SUB ||
13966 Opc == X86ISD::ADC ||
13967 Opc == X86ISD::SBB ||
13968 Opc == X86ISD::SMUL ||
13969 Opc == X86ISD::UMUL ||
13970 Opc == X86ISD::INC ||
13971 Opc == X86ISD::DEC ||
13972 Opc == X86ISD::OR ||
13973 Opc == X86ISD::XOR ||
13974 Opc == X86ISD::AND))
13977 if (Op.getResNo() == 2 && Opc == X86ISD::UMUL)
13983 static bool isTruncWithZeroHighBitsInput(SDValue V, SelectionDAG &DAG) {
13984 if (V.getOpcode() != ISD::TRUNCATE)
13987 SDValue VOp0 = V.getOperand(0);
13988 unsigned InBits = VOp0.getValueSizeInBits();
13989 unsigned Bits = V.getValueSizeInBits();
13990 return DAG.MaskedValueIsZero(VOp0, APInt::getHighBitsSet(InBits,InBits-Bits));
13993 SDValue X86TargetLowering::LowerSELECT(SDValue Op, SelectionDAG &DAG) const {
13994 bool addTest = true;
13995 SDValue Cond = Op.getOperand(0);
13996 SDValue Op1 = Op.getOperand(1);
13997 SDValue Op2 = Op.getOperand(2);
13999 EVT VT = Op1.getValueType();
14002 // Lower FP selects into a CMP/AND/ANDN/OR sequence when the necessary SSE ops
14003 // are available or VBLENDV if AVX is available.
14004 // Otherwise FP cmovs get lowered into a less efficient branch sequence later.
14005 if (Cond.getOpcode() == ISD::SETCC &&
14006 ((Subtarget->hasSSE2() && (VT == MVT::f32 || VT == MVT::f64)) ||
14007 (Subtarget->hasSSE1() && VT == MVT::f32)) &&
14008 VT == Cond.getOperand(0).getValueType() && Cond->hasOneUse()) {
14009 SDValue CondOp0 = Cond.getOperand(0), CondOp1 = Cond.getOperand(1);
14010 int SSECC = translateX86FSETCC(
14011 cast<CondCodeSDNode>(Cond.getOperand(2))->get(), CondOp0, CondOp1);
14014 if (Subtarget->hasAVX512()) {
14015 SDValue Cmp = DAG.getNode(X86ISD::FSETCC, DL, MVT::i1, CondOp0, CondOp1,
14016 DAG.getConstant(SSECC, DL, MVT::i8));
14017 return DAG.getNode(X86ISD::SELECT, DL, VT, Cmp, Op1, Op2);
14020 SDValue Cmp = DAG.getNode(X86ISD::FSETCC, DL, VT, CondOp0, CondOp1,
14021 DAG.getConstant(SSECC, DL, MVT::i8));
14023 // If we have AVX, we can use a variable vector select (VBLENDV) instead
14024 // of 3 logic instructions for size savings and potentially speed.
14025 // Unfortunately, there is no scalar form of VBLENDV.
14027 // If either operand is a constant, don't try this. We can expect to
14028 // optimize away at least one of the logic instructions later in that
14029 // case, so that sequence would be faster than a variable blend.
14031 // BLENDV was introduced with SSE 4.1, but the 2 register form implicitly
14032 // uses XMM0 as the selection register. That may need just as many
14033 // instructions as the AND/ANDN/OR sequence due to register moves, so
14036 if (Subtarget->hasAVX() &&
14037 !isa<ConstantFPSDNode>(Op1) && !isa<ConstantFPSDNode>(Op2)) {
14039 // Convert to vectors, do a VSELECT, and convert back to scalar.
14040 // All of the conversions should be optimized away.
14042 EVT VecVT = VT == MVT::f32 ? MVT::v4f32 : MVT::v2f64;
14043 SDValue VOp1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, VecVT, Op1);
14044 SDValue VOp2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, VecVT, Op2);
14045 SDValue VCmp = DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, VecVT, Cmp);
14047 EVT VCmpVT = VT == MVT::f32 ? MVT::v4i32 : MVT::v2i64;
14048 VCmp = DAG.getBitcast(VCmpVT, VCmp);
14050 SDValue VSel = DAG.getNode(ISD::VSELECT, DL, VecVT, VCmp, VOp1, VOp2);
14052 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, VT,
14053 VSel, DAG.getIntPtrConstant(0, DL));
14055 SDValue AndN = DAG.getNode(X86ISD::FANDN, DL, VT, Cmp, Op2);
14056 SDValue And = DAG.getNode(X86ISD::FAND, DL, VT, Cmp, Op1);
14057 return DAG.getNode(X86ISD::FOR, DL, VT, AndN, And);
14061 if (VT.isVector() && VT.getScalarType() == MVT::i1) {
14063 if (ISD::isBuildVectorOfConstantSDNodes(Op1.getNode()))
14064 Op1Scalar = ConvertI1VectorToInteger(Op1, DAG);
14065 else if (Op1.getOpcode() == ISD::BITCAST && Op1.getOperand(0))
14066 Op1Scalar = Op1.getOperand(0);
14068 if (ISD::isBuildVectorOfConstantSDNodes(Op2.getNode()))
14069 Op2Scalar = ConvertI1VectorToInteger(Op2, DAG);
14070 else if (Op2.getOpcode() == ISD::BITCAST && Op2.getOperand(0))
14071 Op2Scalar = Op2.getOperand(0);
14072 if (Op1Scalar.getNode() && Op2Scalar.getNode()) {
14073 SDValue newSelect = DAG.getNode(ISD::SELECT, DL,
14074 Op1Scalar.getValueType(),
14075 Cond, Op1Scalar, Op2Scalar);
14076 if (newSelect.getValueSizeInBits() == VT.getSizeInBits())
14077 return DAG.getBitcast(VT, newSelect);
14078 SDValue ExtVec = DAG.getBitcast(MVT::v8i1, newSelect);
14079 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, VT, ExtVec,
14080 DAG.getIntPtrConstant(0, DL));
14084 if (VT == MVT::v4i1 || VT == MVT::v2i1) {
14085 SDValue zeroConst = DAG.getIntPtrConstant(0, DL);
14086 Op1 = DAG.getNode(ISD::INSERT_SUBVECTOR, DL, MVT::v8i1,
14087 DAG.getUNDEF(MVT::v8i1), Op1, zeroConst);
14088 Op2 = DAG.getNode(ISD::INSERT_SUBVECTOR, DL, MVT::v8i1,
14089 DAG.getUNDEF(MVT::v8i1), Op2, zeroConst);
14090 SDValue newSelect = DAG.getNode(ISD::SELECT, DL, MVT::v8i1,
14092 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, VT, newSelect, zeroConst);
14095 if (Cond.getOpcode() == ISD::SETCC) {
14096 SDValue NewCond = LowerSETCC(Cond, DAG);
14097 if (NewCond.getNode())
14101 // (select (x == 0), -1, y) -> (sign_bit (x - 1)) | y
14102 // (select (x == 0), y, -1) -> ~(sign_bit (x - 1)) | y
14103 // (select (x != 0), y, -1) -> (sign_bit (x - 1)) | y
14104 // (select (x != 0), -1, y) -> ~(sign_bit (x - 1)) | y
14105 if (Cond.getOpcode() == X86ISD::SETCC &&
14106 Cond.getOperand(1).getOpcode() == X86ISD::CMP &&
14107 isZero(Cond.getOperand(1).getOperand(1))) {
14108 SDValue Cmp = Cond.getOperand(1);
14110 unsigned CondCode =cast<ConstantSDNode>(Cond.getOperand(0))->getZExtValue();
14112 if ((isAllOnes(Op1) || isAllOnes(Op2)) &&
14113 (CondCode == X86::COND_E || CondCode == X86::COND_NE)) {
14114 SDValue Y = isAllOnes(Op2) ? Op1 : Op2;
14116 SDValue CmpOp0 = Cmp.getOperand(0);
14117 // Apply further optimizations for special cases
14118 // (select (x != 0), -1, 0) -> neg & sbb
14119 // (select (x == 0), 0, -1) -> neg & sbb
14120 if (ConstantSDNode *YC = dyn_cast<ConstantSDNode>(Y))
14121 if (YC->isNullValue() &&
14122 (isAllOnes(Op1) == (CondCode == X86::COND_NE))) {
14123 SDVTList VTs = DAG.getVTList(CmpOp0.getValueType(), MVT::i32);
14124 SDValue Neg = DAG.getNode(X86ISD::SUB, DL, VTs,
14125 DAG.getConstant(0, DL,
14126 CmpOp0.getValueType()),
14128 SDValue Res = DAG.getNode(X86ISD::SETCC_CARRY, DL, Op.getValueType(),
14129 DAG.getConstant(X86::COND_B, DL, MVT::i8),
14130 SDValue(Neg.getNode(), 1));
14134 Cmp = DAG.getNode(X86ISD::CMP, DL, MVT::i32,
14135 CmpOp0, DAG.getConstant(1, DL, CmpOp0.getValueType()));
14136 Cmp = ConvertCmpIfNecessary(Cmp, DAG);
14138 SDValue Res = // Res = 0 or -1.
14139 DAG.getNode(X86ISD::SETCC_CARRY, DL, Op.getValueType(),
14140 DAG.getConstant(X86::COND_B, DL, MVT::i8), Cmp);
14142 if (isAllOnes(Op1) != (CondCode == X86::COND_E))
14143 Res = DAG.getNOT(DL, Res, Res.getValueType());
14145 ConstantSDNode *N2C = dyn_cast<ConstantSDNode>(Op2);
14146 if (!N2C || !N2C->isNullValue())
14147 Res = DAG.getNode(ISD::OR, DL, Res.getValueType(), Res, Y);
14152 // Look past (and (setcc_carry (cmp ...)), 1).
14153 if (Cond.getOpcode() == ISD::AND &&
14154 Cond.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY) {
14155 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Cond.getOperand(1));
14156 if (C && C->getAPIntValue() == 1)
14157 Cond = Cond.getOperand(0);
14160 // If condition flag is set by a X86ISD::CMP, then use it as the condition
14161 // setting operand in place of the X86ISD::SETCC.
14162 unsigned CondOpcode = Cond.getOpcode();
14163 if (CondOpcode == X86ISD::SETCC ||
14164 CondOpcode == X86ISD::SETCC_CARRY) {
14165 CC = Cond.getOperand(0);
14167 SDValue Cmp = Cond.getOperand(1);
14168 unsigned Opc = Cmp.getOpcode();
14169 MVT VT = Op.getSimpleValueType();
14171 bool IllegalFPCMov = false;
14172 if (VT.isFloatingPoint() && !VT.isVector() &&
14173 !isScalarFPTypeInSSEReg(VT)) // FPStack?
14174 IllegalFPCMov = !hasFPCMov(cast<ConstantSDNode>(CC)->getSExtValue());
14176 if ((isX86LogicalCmp(Cmp) && !IllegalFPCMov) ||
14177 Opc == X86ISD::BT) { // FIXME
14181 } else if (CondOpcode == ISD::USUBO || CondOpcode == ISD::SSUBO ||
14182 CondOpcode == ISD::UADDO || CondOpcode == ISD::SADDO ||
14183 ((CondOpcode == ISD::UMULO || CondOpcode == ISD::SMULO) &&
14184 Cond.getOperand(0).getValueType() != MVT::i8)) {
14185 SDValue LHS = Cond.getOperand(0);
14186 SDValue RHS = Cond.getOperand(1);
14187 unsigned X86Opcode;
14190 switch (CondOpcode) {
14191 case ISD::UADDO: X86Opcode = X86ISD::ADD; X86Cond = X86::COND_B; break;
14192 case ISD::SADDO: X86Opcode = X86ISD::ADD; X86Cond = X86::COND_O; break;
14193 case ISD::USUBO: X86Opcode = X86ISD::SUB; X86Cond = X86::COND_B; break;
14194 case ISD::SSUBO: X86Opcode = X86ISD::SUB; X86Cond = X86::COND_O; break;
14195 case ISD::UMULO: X86Opcode = X86ISD::UMUL; X86Cond = X86::COND_O; break;
14196 case ISD::SMULO: X86Opcode = X86ISD::SMUL; X86Cond = X86::COND_O; break;
14197 default: llvm_unreachable("unexpected overflowing operator");
14199 if (CondOpcode == ISD::UMULO)
14200 VTs = DAG.getVTList(LHS.getValueType(), LHS.getValueType(),
14203 VTs = DAG.getVTList(LHS.getValueType(), MVT::i32);
14205 SDValue X86Op = DAG.getNode(X86Opcode, DL, VTs, LHS, RHS);
14207 if (CondOpcode == ISD::UMULO)
14208 Cond = X86Op.getValue(2);
14210 Cond = X86Op.getValue(1);
14212 CC = DAG.getConstant(X86Cond, DL, MVT::i8);
14217 // Look past the truncate if the high bits are known zero.
14218 if (isTruncWithZeroHighBitsInput(Cond, DAG))
14219 Cond = Cond.getOperand(0);
14221 // We know the result of AND is compared against zero. Try to match
14223 if (Cond.getOpcode() == ISD::AND && Cond.hasOneUse()) {
14224 SDValue NewSetCC = LowerToBT(Cond, ISD::SETNE, DL, DAG);
14225 if (NewSetCC.getNode()) {
14226 CC = NewSetCC.getOperand(0);
14227 Cond = NewSetCC.getOperand(1);
14234 CC = DAG.getConstant(X86::COND_NE, DL, MVT::i8);
14235 Cond = EmitTest(Cond, X86::COND_NE, DL, DAG);
14238 // a < b ? -1 : 0 -> RES = ~setcc_carry
14239 // a < b ? 0 : -1 -> RES = setcc_carry
14240 // a >= b ? -1 : 0 -> RES = setcc_carry
14241 // a >= b ? 0 : -1 -> RES = ~setcc_carry
14242 if (Cond.getOpcode() == X86ISD::SUB) {
14243 Cond = ConvertCmpIfNecessary(Cond, DAG);
14244 unsigned CondCode = cast<ConstantSDNode>(CC)->getZExtValue();
14246 if ((CondCode == X86::COND_AE || CondCode == X86::COND_B) &&
14247 (isAllOnes(Op1) || isAllOnes(Op2)) && (isZero(Op1) || isZero(Op2))) {
14248 SDValue Res = DAG.getNode(X86ISD::SETCC_CARRY, DL, Op.getValueType(),
14249 DAG.getConstant(X86::COND_B, DL, MVT::i8),
14251 if (isAllOnes(Op1) != (CondCode == X86::COND_B))
14252 return DAG.getNOT(DL, Res, Res.getValueType());
14257 // X86 doesn't have an i8 cmov. If both operands are the result of a truncate
14258 // widen the cmov and push the truncate through. This avoids introducing a new
14259 // branch during isel and doesn't add any extensions.
14260 if (Op.getValueType() == MVT::i8 &&
14261 Op1.getOpcode() == ISD::TRUNCATE && Op2.getOpcode() == ISD::TRUNCATE) {
14262 SDValue T1 = Op1.getOperand(0), T2 = Op2.getOperand(0);
14263 if (T1.getValueType() == T2.getValueType() &&
14264 // Blacklist CopyFromReg to avoid partial register stalls.
14265 T1.getOpcode() != ISD::CopyFromReg && T2.getOpcode()!=ISD::CopyFromReg){
14266 SDVTList VTs = DAG.getVTList(T1.getValueType(), MVT::Glue);
14267 SDValue Cmov = DAG.getNode(X86ISD::CMOV, DL, VTs, T2, T1, CC, Cond);
14268 return DAG.getNode(ISD::TRUNCATE, DL, Op.getValueType(), Cmov);
14272 // X86ISD::CMOV means set the result (which is operand 1) to the RHS if
14273 // condition is true.
14274 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::Glue);
14275 SDValue Ops[] = { Op2, Op1, CC, Cond };
14276 return DAG.getNode(X86ISD::CMOV, DL, VTs, Ops);
14279 static SDValue LowerSIGN_EXTEND_AVX512(SDValue Op,
14280 const X86Subtarget *Subtarget,
14281 SelectionDAG &DAG) {
14282 MVT VT = Op->getSimpleValueType(0);
14283 SDValue In = Op->getOperand(0);
14284 MVT InVT = In.getSimpleValueType();
14285 MVT VTElt = VT.getVectorElementType();
14286 MVT InVTElt = InVT.getVectorElementType();
14290 if ((InVTElt == MVT::i1) &&
14291 (((Subtarget->hasBWI() && Subtarget->hasVLX() &&
14292 VT.getSizeInBits() <= 256 && VTElt.getSizeInBits() <= 16)) ||
14294 ((Subtarget->hasBWI() && VT.is512BitVector() &&
14295 VTElt.getSizeInBits() <= 16)) ||
14297 ((Subtarget->hasDQI() && Subtarget->hasVLX() &&
14298 VT.getSizeInBits() <= 256 && VTElt.getSizeInBits() >= 32)) ||
14300 ((Subtarget->hasDQI() && VT.is512BitVector() &&
14301 VTElt.getSizeInBits() >= 32))))
14302 return DAG.getNode(X86ISD::VSEXT, dl, VT, In);
14304 unsigned int NumElts = VT.getVectorNumElements();
14306 if (NumElts != 8 && NumElts != 16 && !Subtarget->hasBWI())
14309 if (VT.is512BitVector() && InVT.getVectorElementType() != MVT::i1) {
14310 if (In.getOpcode() == X86ISD::VSEXT || In.getOpcode() == X86ISD::VZEXT)
14311 return DAG.getNode(In.getOpcode(), dl, VT, In.getOperand(0));
14312 return DAG.getNode(X86ISD::VSEXT, dl, VT, In);
14315 assert (InVT.getVectorElementType() == MVT::i1 && "Unexpected vector type");
14316 MVT ExtVT = NumElts == 8 ? MVT::v8i64 : MVT::v16i32;
14318 DAG.getConstant(APInt::getAllOnesValue(ExtVT.getScalarSizeInBits()), dl,
14321 DAG.getConstant(APInt::getNullValue(ExtVT.getScalarSizeInBits()), dl, ExtVT);
14323 SDValue V = DAG.getNode(ISD::VSELECT, dl, ExtVT, In, NegOne, Zero);
14324 if (VT.is512BitVector())
14326 return DAG.getNode(X86ISD::VTRUNC, dl, VT, V);
14329 static SDValue LowerSIGN_EXTEND_VECTOR_INREG(SDValue Op,
14330 const X86Subtarget *Subtarget,
14331 SelectionDAG &DAG) {
14332 SDValue In = Op->getOperand(0);
14333 MVT VT = Op->getSimpleValueType(0);
14334 MVT InVT = In.getSimpleValueType();
14335 assert(VT.getSizeInBits() == InVT.getSizeInBits());
14337 MVT InSVT = InVT.getScalarType();
14338 assert(VT.getScalarType().getScalarSizeInBits() > InSVT.getScalarSizeInBits());
14340 if (VT != MVT::v2i64 && VT != MVT::v4i32 && VT != MVT::v8i16)
14342 if (InSVT != MVT::i32 && InSVT != MVT::i16 && InSVT != MVT::i8)
14347 // SSE41 targets can use the pmovsx* instructions directly.
14348 if (Subtarget->hasSSE41())
14349 return DAG.getNode(X86ISD::VSEXT, dl, VT, In);
14351 // pre-SSE41 targets unpack lower lanes and then sign-extend using SRAI.
14355 // As SRAI is only available on i16/i32 types, we expand only up to i32
14356 // and handle i64 separately.
14357 while (CurrVT != VT && CurrVT.getScalarType() != MVT::i32) {
14358 Curr = DAG.getNode(X86ISD::UNPCKL, dl, CurrVT, DAG.getUNDEF(CurrVT), Curr);
14359 MVT CurrSVT = MVT::getIntegerVT(CurrVT.getScalarSizeInBits() * 2);
14360 CurrVT = MVT::getVectorVT(CurrSVT, CurrVT.getVectorNumElements() / 2);
14361 Curr = DAG.getBitcast(CurrVT, Curr);
14364 SDValue SignExt = Curr;
14365 if (CurrVT != InVT) {
14366 unsigned SignExtShift =
14367 CurrVT.getScalarSizeInBits() - InSVT.getScalarSizeInBits();
14368 SignExt = DAG.getNode(X86ISD::VSRAI, dl, CurrVT, Curr,
14369 DAG.getConstant(SignExtShift, dl, MVT::i8));
14375 if (VT == MVT::v2i64 && CurrVT == MVT::v4i32) {
14376 SDValue Sign = DAG.getNode(X86ISD::VSRAI, dl, CurrVT, Curr,
14377 DAG.getConstant(31, dl, MVT::i8));
14378 SDValue Ext = DAG.getVectorShuffle(CurrVT, dl, SignExt, Sign, {0, 4, 1, 5});
14379 return DAG.getBitcast(VT, Ext);
14385 static SDValue LowerSIGN_EXTEND(SDValue Op, const X86Subtarget *Subtarget,
14386 SelectionDAG &DAG) {
14387 MVT VT = Op->getSimpleValueType(0);
14388 SDValue In = Op->getOperand(0);
14389 MVT InVT = In.getSimpleValueType();
14392 if (VT.is512BitVector() || InVT.getVectorElementType() == MVT::i1)
14393 return LowerSIGN_EXTEND_AVX512(Op, Subtarget, DAG);
14395 if ((VT != MVT::v4i64 || InVT != MVT::v4i32) &&
14396 (VT != MVT::v8i32 || InVT != MVT::v8i16) &&
14397 (VT != MVT::v16i16 || InVT != MVT::v16i8))
14400 if (Subtarget->hasInt256())
14401 return DAG.getNode(X86ISD::VSEXT, dl, VT, In);
14403 // Optimize vectors in AVX mode
14404 // Sign extend v8i16 to v8i32 and
14407 // Divide input vector into two parts
14408 // for v4i32 the shuffle mask will be { 0, 1, -1, -1} {2, 3, -1, -1}
14409 // use vpmovsx instruction to extend v4i32 -> v2i64; v8i16 -> v4i32
14410 // concat the vectors to original VT
14412 unsigned NumElems = InVT.getVectorNumElements();
14413 SDValue Undef = DAG.getUNDEF(InVT);
14415 SmallVector<int,8> ShufMask1(NumElems, -1);
14416 for (unsigned i = 0; i != NumElems/2; ++i)
14419 SDValue OpLo = DAG.getVectorShuffle(InVT, dl, In, Undef, &ShufMask1[0]);
14421 SmallVector<int,8> ShufMask2(NumElems, -1);
14422 for (unsigned i = 0; i != NumElems/2; ++i)
14423 ShufMask2[i] = i + NumElems/2;
14425 SDValue OpHi = DAG.getVectorShuffle(InVT, dl, In, Undef, &ShufMask2[0]);
14427 MVT HalfVT = MVT::getVectorVT(VT.getScalarType(),
14428 VT.getVectorNumElements()/2);
14430 OpLo = DAG.getNode(X86ISD::VSEXT, dl, HalfVT, OpLo);
14431 OpHi = DAG.getNode(X86ISD::VSEXT, dl, HalfVT, OpHi);
14433 return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT, OpLo, OpHi);
14436 // Lower vector extended loads using a shuffle. If SSSE3 is not available we
14437 // may emit an illegal shuffle but the expansion is still better than scalar
14438 // code. We generate X86ISD::VSEXT for SEXTLOADs if it's available, otherwise
14439 // we'll emit a shuffle and a arithmetic shift.
14440 // FIXME: Is the expansion actually better than scalar code? It doesn't seem so.
14441 // TODO: It is possible to support ZExt by zeroing the undef values during
14442 // the shuffle phase or after the shuffle.
14443 static SDValue LowerExtendedLoad(SDValue Op, const X86Subtarget *Subtarget,
14444 SelectionDAG &DAG) {
14445 MVT RegVT = Op.getSimpleValueType();
14446 assert(RegVT.isVector() && "We only custom lower vector sext loads.");
14447 assert(RegVT.isInteger() &&
14448 "We only custom lower integer vector sext loads.");
14450 // Nothing useful we can do without SSE2 shuffles.
14451 assert(Subtarget->hasSSE2() && "We only custom lower sext loads with SSE2.");
14453 LoadSDNode *Ld = cast<LoadSDNode>(Op.getNode());
14455 EVT MemVT = Ld->getMemoryVT();
14456 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
14457 unsigned RegSz = RegVT.getSizeInBits();
14459 ISD::LoadExtType Ext = Ld->getExtensionType();
14461 assert((Ext == ISD::EXTLOAD || Ext == ISD::SEXTLOAD)
14462 && "Only anyext and sext are currently implemented.");
14463 assert(MemVT != RegVT && "Cannot extend to the same type");
14464 assert(MemVT.isVector() && "Must load a vector from memory");
14466 unsigned NumElems = RegVT.getVectorNumElements();
14467 unsigned MemSz = MemVT.getSizeInBits();
14468 assert(RegSz > MemSz && "Register size must be greater than the mem size");
14470 if (Ext == ISD::SEXTLOAD && RegSz == 256 && !Subtarget->hasInt256()) {
14471 // The only way in which we have a legal 256-bit vector result but not the
14472 // integer 256-bit operations needed to directly lower a sextload is if we
14473 // have AVX1 but not AVX2. In that case, we can always emit a sextload to
14474 // a 128-bit vector and a normal sign_extend to 256-bits that should get
14475 // correctly legalized. We do this late to allow the canonical form of
14476 // sextload to persist throughout the rest of the DAG combiner -- it wants
14477 // to fold together any extensions it can, and so will fuse a sign_extend
14478 // of an sextload into a sextload targeting a wider value.
14480 if (MemSz == 128) {
14481 // Just switch this to a normal load.
14482 assert(TLI.isTypeLegal(MemVT) && "If the memory type is a 128-bit type, "
14483 "it must be a legal 128-bit vector "
14485 Load = DAG.getLoad(MemVT, dl, Ld->getChain(), Ld->getBasePtr(),
14486 Ld->getPointerInfo(), Ld->isVolatile(), Ld->isNonTemporal(),
14487 Ld->isInvariant(), Ld->getAlignment());
14489 assert(MemSz < 128 &&
14490 "Can't extend a type wider than 128 bits to a 256 bit vector!");
14491 // Do an sext load to a 128-bit vector type. We want to use the same
14492 // number of elements, but elements half as wide. This will end up being
14493 // recursively lowered by this routine, but will succeed as we definitely
14494 // have all the necessary features if we're using AVX1.
14496 EVT::getIntegerVT(*DAG.getContext(), RegVT.getScalarSizeInBits() / 2);
14497 EVT HalfVecVT = EVT::getVectorVT(*DAG.getContext(), HalfEltVT, NumElems);
14499 DAG.getExtLoad(Ext, dl, HalfVecVT, Ld->getChain(), Ld->getBasePtr(),
14500 Ld->getPointerInfo(), MemVT, Ld->isVolatile(),
14501 Ld->isNonTemporal(), Ld->isInvariant(),
14502 Ld->getAlignment());
14505 // Replace chain users with the new chain.
14506 assert(Load->getNumValues() == 2 && "Loads must carry a chain!");
14507 DAG.ReplaceAllUsesOfValueWith(SDValue(Ld, 1), Load.getValue(1));
14509 // Finally, do a normal sign-extend to the desired register.
14510 return DAG.getSExtOrTrunc(Load, dl, RegVT);
14513 // All sizes must be a power of two.
14514 assert(isPowerOf2_32(RegSz * MemSz * NumElems) &&
14515 "Non-power-of-two elements are not custom lowered!");
14517 // Attempt to load the original value using scalar loads.
14518 // Find the largest scalar type that divides the total loaded size.
14519 MVT SclrLoadTy = MVT::i8;
14520 for (MVT Tp : MVT::integer_valuetypes()) {
14521 if (TLI.isTypeLegal(Tp) && ((MemSz % Tp.getSizeInBits()) == 0)) {
14526 // On 32bit systems, we can't save 64bit integers. Try bitcasting to F64.
14527 if (TLI.isTypeLegal(MVT::f64) && SclrLoadTy.getSizeInBits() < 64 &&
14529 SclrLoadTy = MVT::f64;
14531 // Calculate the number of scalar loads that we need to perform
14532 // in order to load our vector from memory.
14533 unsigned NumLoads = MemSz / SclrLoadTy.getSizeInBits();
14535 assert((Ext != ISD::SEXTLOAD || NumLoads == 1) &&
14536 "Can only lower sext loads with a single scalar load!");
14538 unsigned loadRegZize = RegSz;
14539 if (Ext == ISD::SEXTLOAD && RegSz >= 256)
14542 // Represent our vector as a sequence of elements which are the
14543 // largest scalar that we can load.
14544 EVT LoadUnitVecVT = EVT::getVectorVT(
14545 *DAG.getContext(), SclrLoadTy, loadRegZize / SclrLoadTy.getSizeInBits());
14547 // Represent the data using the same element type that is stored in
14548 // memory. In practice, we ''widen'' MemVT.
14550 EVT::getVectorVT(*DAG.getContext(), MemVT.getScalarType(),
14551 loadRegZize / MemVT.getScalarType().getSizeInBits());
14553 assert(WideVecVT.getSizeInBits() == LoadUnitVecVT.getSizeInBits() &&
14554 "Invalid vector type");
14556 // We can't shuffle using an illegal type.
14557 assert(TLI.isTypeLegal(WideVecVT) &&
14558 "We only lower types that form legal widened vector types");
14560 SmallVector<SDValue, 8> Chains;
14561 SDValue Ptr = Ld->getBasePtr();
14562 SDValue Increment = DAG.getConstant(SclrLoadTy.getSizeInBits() / 8, dl,
14563 TLI.getPointerTy(DAG.getDataLayout()));
14564 SDValue Res = DAG.getUNDEF(LoadUnitVecVT);
14566 for (unsigned i = 0; i < NumLoads; ++i) {
14567 // Perform a single load.
14568 SDValue ScalarLoad =
14569 DAG.getLoad(SclrLoadTy, dl, Ld->getChain(), Ptr, Ld->getPointerInfo(),
14570 Ld->isVolatile(), Ld->isNonTemporal(), Ld->isInvariant(),
14571 Ld->getAlignment());
14572 Chains.push_back(ScalarLoad.getValue(1));
14573 // Create the first element type using SCALAR_TO_VECTOR in order to avoid
14574 // another round of DAGCombining.
14576 Res = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, LoadUnitVecVT, ScalarLoad);
14578 Res = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, LoadUnitVecVT, Res,
14579 ScalarLoad, DAG.getIntPtrConstant(i, dl));
14581 Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr, Increment);
14584 SDValue TF = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Chains);
14586 // Bitcast the loaded value to a vector of the original element type, in
14587 // the size of the target vector type.
14588 SDValue SlicedVec = DAG.getBitcast(WideVecVT, Res);
14589 unsigned SizeRatio = RegSz / MemSz;
14591 if (Ext == ISD::SEXTLOAD) {
14592 // If we have SSE4.1, we can directly emit a VSEXT node.
14593 if (Subtarget->hasSSE41()) {
14594 SDValue Sext = DAG.getNode(X86ISD::VSEXT, dl, RegVT, SlicedVec);
14595 DAG.ReplaceAllUsesOfValueWith(SDValue(Ld, 1), TF);
14599 // Otherwise we'll shuffle the small elements in the high bits of the
14600 // larger type and perform an arithmetic shift. If the shift is not legal
14601 // it's better to scalarize.
14602 assert(TLI.isOperationLegalOrCustom(ISD::SRA, RegVT) &&
14603 "We can't implement a sext load without an arithmetic right shift!");
14605 // Redistribute the loaded elements into the different locations.
14606 SmallVector<int, 16> ShuffleVec(NumElems * SizeRatio, -1);
14607 for (unsigned i = 0; i != NumElems; ++i)
14608 ShuffleVec[i * SizeRatio + SizeRatio - 1] = i;
14610 SDValue Shuff = DAG.getVectorShuffle(
14611 WideVecVT, dl, SlicedVec, DAG.getUNDEF(WideVecVT), &ShuffleVec[0]);
14613 Shuff = DAG.getBitcast(RegVT, Shuff);
14615 // Build the arithmetic shift.
14616 unsigned Amt = RegVT.getVectorElementType().getSizeInBits() -
14617 MemVT.getVectorElementType().getSizeInBits();
14619 DAG.getNode(ISD::SRA, dl, RegVT, Shuff,
14620 DAG.getConstant(Amt, dl, RegVT));
14622 DAG.ReplaceAllUsesOfValueWith(SDValue(Ld, 1), TF);
14626 // Redistribute the loaded elements into the different locations.
14627 SmallVector<int, 16> ShuffleVec(NumElems * SizeRatio, -1);
14628 for (unsigned i = 0; i != NumElems; ++i)
14629 ShuffleVec[i * SizeRatio] = i;
14631 SDValue Shuff = DAG.getVectorShuffle(WideVecVT, dl, SlicedVec,
14632 DAG.getUNDEF(WideVecVT), &ShuffleVec[0]);
14634 // Bitcast to the requested type.
14635 Shuff = DAG.getBitcast(RegVT, Shuff);
14636 DAG.ReplaceAllUsesOfValueWith(SDValue(Ld, 1), TF);
14640 // isAndOrOfSingleUseSetCCs - Return true if node is an ISD::AND or
14641 // ISD::OR of two X86ISD::SETCC nodes each of which has no other use apart
14642 // from the AND / OR.
14643 static bool isAndOrOfSetCCs(SDValue Op, unsigned &Opc) {
14644 Opc = Op.getOpcode();
14645 if (Opc != ISD::OR && Opc != ISD::AND)
14647 return (Op.getOperand(0).getOpcode() == X86ISD::SETCC &&
14648 Op.getOperand(0).hasOneUse() &&
14649 Op.getOperand(1).getOpcode() == X86ISD::SETCC &&
14650 Op.getOperand(1).hasOneUse());
14653 // isXor1OfSetCC - Return true if node is an ISD::XOR of a X86ISD::SETCC and
14654 // 1 and that the SETCC node has a single use.
14655 static bool isXor1OfSetCC(SDValue Op) {
14656 if (Op.getOpcode() != ISD::XOR)
14658 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(Op.getOperand(1));
14659 if (N1C && N1C->getAPIntValue() == 1) {
14660 return Op.getOperand(0).getOpcode() == X86ISD::SETCC &&
14661 Op.getOperand(0).hasOneUse();
14666 SDValue X86TargetLowering::LowerBRCOND(SDValue Op, SelectionDAG &DAG) const {
14667 bool addTest = true;
14668 SDValue Chain = Op.getOperand(0);
14669 SDValue Cond = Op.getOperand(1);
14670 SDValue Dest = Op.getOperand(2);
14673 bool Inverted = false;
14675 if (Cond.getOpcode() == ISD::SETCC) {
14676 // Check for setcc([su]{add,sub,mul}o == 0).
14677 if (cast<CondCodeSDNode>(Cond.getOperand(2))->get() == ISD::SETEQ &&
14678 isa<ConstantSDNode>(Cond.getOperand(1)) &&
14679 cast<ConstantSDNode>(Cond.getOperand(1))->isNullValue() &&
14680 Cond.getOperand(0).getResNo() == 1 &&
14681 (Cond.getOperand(0).getOpcode() == ISD::SADDO ||
14682 Cond.getOperand(0).getOpcode() == ISD::UADDO ||
14683 Cond.getOperand(0).getOpcode() == ISD::SSUBO ||
14684 Cond.getOperand(0).getOpcode() == ISD::USUBO ||
14685 Cond.getOperand(0).getOpcode() == ISD::SMULO ||
14686 Cond.getOperand(0).getOpcode() == ISD::UMULO)) {
14688 Cond = Cond.getOperand(0);
14690 SDValue NewCond = LowerSETCC(Cond, DAG);
14691 if (NewCond.getNode())
14696 // FIXME: LowerXALUO doesn't handle these!!
14697 else if (Cond.getOpcode() == X86ISD::ADD ||
14698 Cond.getOpcode() == X86ISD::SUB ||
14699 Cond.getOpcode() == X86ISD::SMUL ||
14700 Cond.getOpcode() == X86ISD::UMUL)
14701 Cond = LowerXALUO(Cond, DAG);
14704 // Look pass (and (setcc_carry (cmp ...)), 1).
14705 if (Cond.getOpcode() == ISD::AND &&
14706 Cond.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY) {
14707 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Cond.getOperand(1));
14708 if (C && C->getAPIntValue() == 1)
14709 Cond = Cond.getOperand(0);
14712 // If condition flag is set by a X86ISD::CMP, then use it as the condition
14713 // setting operand in place of the X86ISD::SETCC.
14714 unsigned CondOpcode = Cond.getOpcode();
14715 if (CondOpcode == X86ISD::SETCC ||
14716 CondOpcode == X86ISD::SETCC_CARRY) {
14717 CC = Cond.getOperand(0);
14719 SDValue Cmp = Cond.getOperand(1);
14720 unsigned Opc = Cmp.getOpcode();
14721 // FIXME: WHY THE SPECIAL CASING OF LogicalCmp??
14722 if (isX86LogicalCmp(Cmp) || Opc == X86ISD::BT) {
14726 switch (cast<ConstantSDNode>(CC)->getZExtValue()) {
14730 // These can only come from an arithmetic instruction with overflow,
14731 // e.g. SADDO, UADDO.
14732 Cond = Cond.getNode()->getOperand(1);
14738 CondOpcode = Cond.getOpcode();
14739 if (CondOpcode == ISD::UADDO || CondOpcode == ISD::SADDO ||
14740 CondOpcode == ISD::USUBO || CondOpcode == ISD::SSUBO ||
14741 ((CondOpcode == ISD::UMULO || CondOpcode == ISD::SMULO) &&
14742 Cond.getOperand(0).getValueType() != MVT::i8)) {
14743 SDValue LHS = Cond.getOperand(0);
14744 SDValue RHS = Cond.getOperand(1);
14745 unsigned X86Opcode;
14748 // Keep this in sync with LowerXALUO, otherwise we might create redundant
14749 // instructions that can't be removed afterwards (i.e. X86ISD::ADD and
14751 switch (CondOpcode) {
14752 case ISD::UADDO: X86Opcode = X86ISD::ADD; X86Cond = X86::COND_B; break;
14754 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(RHS))
14756 X86Opcode = X86ISD::INC; X86Cond = X86::COND_O;
14759 X86Opcode = X86ISD::ADD; X86Cond = X86::COND_O; break;
14760 case ISD::USUBO: X86Opcode = X86ISD::SUB; X86Cond = X86::COND_B; break;
14762 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(RHS))
14764 X86Opcode = X86ISD::DEC; X86Cond = X86::COND_O;
14767 X86Opcode = X86ISD::SUB; X86Cond = X86::COND_O; break;
14768 case ISD::UMULO: X86Opcode = X86ISD::UMUL; X86Cond = X86::COND_O; break;
14769 case ISD::SMULO: X86Opcode = X86ISD::SMUL; X86Cond = X86::COND_O; break;
14770 default: llvm_unreachable("unexpected overflowing operator");
14773 X86Cond = X86::GetOppositeBranchCondition((X86::CondCode)X86Cond);
14774 if (CondOpcode == ISD::UMULO)
14775 VTs = DAG.getVTList(LHS.getValueType(), LHS.getValueType(),
14778 VTs = DAG.getVTList(LHS.getValueType(), MVT::i32);
14780 SDValue X86Op = DAG.getNode(X86Opcode, dl, VTs, LHS, RHS);
14782 if (CondOpcode == ISD::UMULO)
14783 Cond = X86Op.getValue(2);
14785 Cond = X86Op.getValue(1);
14787 CC = DAG.getConstant(X86Cond, dl, MVT::i8);
14791 if (Cond.hasOneUse() && isAndOrOfSetCCs(Cond, CondOpc)) {
14792 SDValue Cmp = Cond.getOperand(0).getOperand(1);
14793 if (CondOpc == ISD::OR) {
14794 // Also, recognize the pattern generated by an FCMP_UNE. We can emit
14795 // two branches instead of an explicit OR instruction with a
14797 if (Cmp == Cond.getOperand(1).getOperand(1) &&
14798 isX86LogicalCmp(Cmp)) {
14799 CC = Cond.getOperand(0).getOperand(0);
14800 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
14801 Chain, Dest, CC, Cmp);
14802 CC = Cond.getOperand(1).getOperand(0);
14806 } else { // ISD::AND
14807 // Also, recognize the pattern generated by an FCMP_OEQ. We can emit
14808 // two branches instead of an explicit AND instruction with a
14809 // separate test. However, we only do this if this block doesn't
14810 // have a fall-through edge, because this requires an explicit
14811 // jmp when the condition is false.
14812 if (Cmp == Cond.getOperand(1).getOperand(1) &&
14813 isX86LogicalCmp(Cmp) &&
14814 Op.getNode()->hasOneUse()) {
14815 X86::CondCode CCode =
14816 (X86::CondCode)Cond.getOperand(0).getConstantOperandVal(0);
14817 CCode = X86::GetOppositeBranchCondition(CCode);
14818 CC = DAG.getConstant(CCode, dl, MVT::i8);
14819 SDNode *User = *Op.getNode()->use_begin();
14820 // Look for an unconditional branch following this conditional branch.
14821 // We need this because we need to reverse the successors in order
14822 // to implement FCMP_OEQ.
14823 if (User->getOpcode() == ISD::BR) {
14824 SDValue FalseBB = User->getOperand(1);
14826 DAG.UpdateNodeOperands(User, User->getOperand(0), Dest);
14827 assert(NewBR == User);
14831 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
14832 Chain, Dest, CC, Cmp);
14833 X86::CondCode CCode =
14834 (X86::CondCode)Cond.getOperand(1).getConstantOperandVal(0);
14835 CCode = X86::GetOppositeBranchCondition(CCode);
14836 CC = DAG.getConstant(CCode, dl, MVT::i8);
14842 } else if (Cond.hasOneUse() && isXor1OfSetCC(Cond)) {
14843 // Recognize for xorb (setcc), 1 patterns. The xor inverts the condition.
14844 // It should be transformed during dag combiner except when the condition
14845 // is set by a arithmetics with overflow node.
14846 X86::CondCode CCode =
14847 (X86::CondCode)Cond.getOperand(0).getConstantOperandVal(0);
14848 CCode = X86::GetOppositeBranchCondition(CCode);
14849 CC = DAG.getConstant(CCode, dl, MVT::i8);
14850 Cond = Cond.getOperand(0).getOperand(1);
14852 } else if (Cond.getOpcode() == ISD::SETCC &&
14853 cast<CondCodeSDNode>(Cond.getOperand(2))->get() == ISD::SETOEQ) {
14854 // For FCMP_OEQ, we can emit
14855 // two branches instead of an explicit AND instruction with a
14856 // separate test. However, we only do this if this block doesn't
14857 // have a fall-through edge, because this requires an explicit
14858 // jmp when the condition is false.
14859 if (Op.getNode()->hasOneUse()) {
14860 SDNode *User = *Op.getNode()->use_begin();
14861 // Look for an unconditional branch following this conditional branch.
14862 // We need this because we need to reverse the successors in order
14863 // to implement FCMP_OEQ.
14864 if (User->getOpcode() == ISD::BR) {
14865 SDValue FalseBB = User->getOperand(1);
14867 DAG.UpdateNodeOperands(User, User->getOperand(0), Dest);
14868 assert(NewBR == User);
14872 SDValue Cmp = DAG.getNode(X86ISD::CMP, dl, MVT::i32,
14873 Cond.getOperand(0), Cond.getOperand(1));
14874 Cmp = ConvertCmpIfNecessary(Cmp, DAG);
14875 CC = DAG.getConstant(X86::COND_NE, dl, MVT::i8);
14876 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
14877 Chain, Dest, CC, Cmp);
14878 CC = DAG.getConstant(X86::COND_P, dl, MVT::i8);
14883 } else if (Cond.getOpcode() == ISD::SETCC &&
14884 cast<CondCodeSDNode>(Cond.getOperand(2))->get() == ISD::SETUNE) {
14885 // For FCMP_UNE, we can emit
14886 // two branches instead of an explicit AND instruction with a
14887 // separate test. However, we only do this if this block doesn't
14888 // have a fall-through edge, because this requires an explicit
14889 // jmp when the condition is false.
14890 if (Op.getNode()->hasOneUse()) {
14891 SDNode *User = *Op.getNode()->use_begin();
14892 // Look for an unconditional branch following this conditional branch.
14893 // We need this because we need to reverse the successors in order
14894 // to implement FCMP_UNE.
14895 if (User->getOpcode() == ISD::BR) {
14896 SDValue FalseBB = User->getOperand(1);
14898 DAG.UpdateNodeOperands(User, User->getOperand(0), Dest);
14899 assert(NewBR == User);
14902 SDValue Cmp = DAG.getNode(X86ISD::CMP, dl, MVT::i32,
14903 Cond.getOperand(0), Cond.getOperand(1));
14904 Cmp = ConvertCmpIfNecessary(Cmp, DAG);
14905 CC = DAG.getConstant(X86::COND_NE, dl, MVT::i8);
14906 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
14907 Chain, Dest, CC, Cmp);
14908 CC = DAG.getConstant(X86::COND_NP, dl, MVT::i8);
14918 // Look pass the truncate if the high bits are known zero.
14919 if (isTruncWithZeroHighBitsInput(Cond, DAG))
14920 Cond = Cond.getOperand(0);
14922 // We know the result of AND is compared against zero. Try to match
14924 if (Cond.getOpcode() == ISD::AND && Cond.hasOneUse()) {
14925 SDValue NewSetCC = LowerToBT(Cond, ISD::SETNE, dl, DAG);
14926 if (NewSetCC.getNode()) {
14927 CC = NewSetCC.getOperand(0);
14928 Cond = NewSetCC.getOperand(1);
14935 X86::CondCode X86Cond = Inverted ? X86::COND_E : X86::COND_NE;
14936 CC = DAG.getConstant(X86Cond, dl, MVT::i8);
14937 Cond = EmitTest(Cond, X86Cond, dl, DAG);
14939 Cond = ConvertCmpIfNecessary(Cond, DAG);
14940 return DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
14941 Chain, Dest, CC, Cond);
14944 // Lower dynamic stack allocation to _alloca call for Cygwin/Mingw targets.
14945 // Calls to _alloca are needed to probe the stack when allocating more than 4k
14946 // bytes in one go. Touching the stack at 4K increments is necessary to ensure
14947 // that the guard pages used by the OS virtual memory manager are allocated in
14948 // correct sequence.
14950 X86TargetLowering::LowerDYNAMIC_STACKALLOC(SDValue Op,
14951 SelectionDAG &DAG) const {
14952 MachineFunction &MF = DAG.getMachineFunction();
14953 bool SplitStack = MF.shouldSplitStack();
14954 bool Lower = (Subtarget->isOSWindows() && !Subtarget->isTargetMachO()) ||
14959 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
14960 SDNode* Node = Op.getNode();
14962 unsigned SPReg = TLI.getStackPointerRegisterToSaveRestore();
14963 assert(SPReg && "Target cannot require DYNAMIC_STACKALLOC expansion and"
14964 " not tell us which reg is the stack pointer!");
14965 EVT VT = Node->getValueType(0);
14966 SDValue Tmp1 = SDValue(Node, 0);
14967 SDValue Tmp2 = SDValue(Node, 1);
14968 SDValue Tmp3 = Node->getOperand(2);
14969 SDValue Chain = Tmp1.getOperand(0);
14971 // Chain the dynamic stack allocation so that it doesn't modify the stack
14972 // pointer when other instructions are using the stack.
14973 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(0, dl, true),
14976 SDValue Size = Tmp2.getOperand(1);
14977 SDValue SP = DAG.getCopyFromReg(Chain, dl, SPReg, VT);
14978 Chain = SP.getValue(1);
14979 unsigned Align = cast<ConstantSDNode>(Tmp3)->getZExtValue();
14980 const TargetFrameLowering &TFI = *Subtarget->getFrameLowering();
14981 unsigned StackAlign = TFI.getStackAlignment();
14982 Tmp1 = DAG.getNode(ISD::SUB, dl, VT, SP, Size); // Value
14983 if (Align > StackAlign)
14984 Tmp1 = DAG.getNode(ISD::AND, dl, VT, Tmp1,
14985 DAG.getConstant(-(uint64_t)Align, dl, VT));
14986 Chain = DAG.getCopyToReg(Chain, dl, SPReg, Tmp1); // Output chain
14988 Tmp2 = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(0, dl, true),
14989 DAG.getIntPtrConstant(0, dl, true), SDValue(),
14992 SDValue Ops[2] = { Tmp1, Tmp2 };
14993 return DAG.getMergeValues(Ops, dl);
14997 SDValue Chain = Op.getOperand(0);
14998 SDValue Size = Op.getOperand(1);
14999 unsigned Align = cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue();
15000 EVT VT = Op.getNode()->getValueType(0);
15002 bool Is64Bit = Subtarget->is64Bit();
15003 MVT SPTy = getPointerTy(DAG.getDataLayout());
15006 MachineRegisterInfo &MRI = MF.getRegInfo();
15009 // The 64 bit implementation of segmented stacks needs to clobber both r10
15010 // r11. This makes it impossible to use it along with nested parameters.
15011 const Function *F = MF.getFunction();
15013 for (Function::const_arg_iterator I = F->arg_begin(), E = F->arg_end();
15015 if (I->hasNestAttr())
15016 report_fatal_error("Cannot use segmented stacks with functions that "
15017 "have nested arguments.");
15020 const TargetRegisterClass *AddrRegClass = getRegClassFor(SPTy);
15021 unsigned Vreg = MRI.createVirtualRegister(AddrRegClass);
15022 Chain = DAG.getCopyToReg(Chain, dl, Vreg, Size);
15023 SDValue Value = DAG.getNode(X86ISD::SEG_ALLOCA, dl, SPTy, Chain,
15024 DAG.getRegister(Vreg, SPTy));
15025 SDValue Ops1[2] = { Value, Chain };
15026 return DAG.getMergeValues(Ops1, dl);
15029 const unsigned Reg = (Subtarget->isTarget64BitLP64() ? X86::RAX : X86::EAX);
15031 Chain = DAG.getCopyToReg(Chain, dl, Reg, Size, Flag);
15032 Flag = Chain.getValue(1);
15033 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
15035 Chain = DAG.getNode(X86ISD::WIN_ALLOCA, dl, NodeTys, Chain, Flag);
15037 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
15038 unsigned SPReg = RegInfo->getStackRegister();
15039 SDValue SP = DAG.getCopyFromReg(Chain, dl, SPReg, SPTy);
15040 Chain = SP.getValue(1);
15043 SP = DAG.getNode(ISD::AND, dl, VT, SP.getValue(0),
15044 DAG.getConstant(-(uint64_t)Align, dl, VT));
15045 Chain = DAG.getCopyToReg(Chain, dl, SPReg, SP);
15048 SDValue Ops1[2] = { SP, Chain };
15049 return DAG.getMergeValues(Ops1, dl);
15053 SDValue X86TargetLowering::LowerVASTART(SDValue Op, SelectionDAG &DAG) const {
15054 MachineFunction &MF = DAG.getMachineFunction();
15055 auto PtrVT = getPointerTy(MF.getDataLayout());
15056 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
15058 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
15061 if (!Subtarget->is64Bit() || Subtarget->isTargetWin64()) {
15062 // vastart just stores the address of the VarArgsFrameIndex slot into the
15063 // memory location argument.
15064 SDValue FR = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(), PtrVT);
15065 return DAG.getStore(Op.getOperand(0), DL, FR, Op.getOperand(1),
15066 MachinePointerInfo(SV), false, false, 0);
15070 // gp_offset (0 - 6 * 8)
15071 // fp_offset (48 - 48 + 8 * 16)
15072 // overflow_arg_area (point to parameters coming in memory).
15074 SmallVector<SDValue, 8> MemOps;
15075 SDValue FIN = Op.getOperand(1);
15077 SDValue Store = DAG.getStore(Op.getOperand(0), DL,
15078 DAG.getConstant(FuncInfo->getVarArgsGPOffset(),
15080 FIN, MachinePointerInfo(SV), false, false, 0);
15081 MemOps.push_back(Store);
15084 FIN = DAG.getNode(ISD::ADD, DL, PtrVT, FIN, DAG.getIntPtrConstant(4, DL));
15085 Store = DAG.getStore(Op.getOperand(0), DL,
15086 DAG.getConstant(FuncInfo->getVarArgsFPOffset(), DL,
15088 FIN, MachinePointerInfo(SV, 4), false, false, 0);
15089 MemOps.push_back(Store);
15091 // Store ptr to overflow_arg_area
15092 FIN = DAG.getNode(ISD::ADD, DL, PtrVT, FIN, DAG.getIntPtrConstant(4, DL));
15093 SDValue OVFIN = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(), PtrVT);
15094 Store = DAG.getStore(Op.getOperand(0), DL, OVFIN, FIN,
15095 MachinePointerInfo(SV, 8),
15097 MemOps.push_back(Store);
15099 // Store ptr to reg_save_area.
15100 FIN = DAG.getNode(ISD::ADD, DL, PtrVT, FIN, DAG.getIntPtrConstant(8, DL));
15101 SDValue RSFIN = DAG.getFrameIndex(FuncInfo->getRegSaveFrameIndex(), PtrVT);
15102 Store = DAG.getStore(Op.getOperand(0), DL, RSFIN, FIN,
15103 MachinePointerInfo(SV, 16), false, false, 0);
15104 MemOps.push_back(Store);
15105 return DAG.getNode(ISD::TokenFactor, DL, MVT::Other, MemOps);
15108 SDValue X86TargetLowering::LowerVAARG(SDValue Op, SelectionDAG &DAG) const {
15109 assert(Subtarget->is64Bit() &&
15110 "LowerVAARG only handles 64-bit va_arg!");
15111 assert((Subtarget->isTargetLinux() ||
15112 Subtarget->isTargetDarwin()) &&
15113 "Unhandled target in LowerVAARG");
15114 assert(Op.getNode()->getNumOperands() == 4);
15115 SDValue Chain = Op.getOperand(0);
15116 SDValue SrcPtr = Op.getOperand(1);
15117 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
15118 unsigned Align = Op.getConstantOperandVal(3);
15121 EVT ArgVT = Op.getNode()->getValueType(0);
15122 Type *ArgTy = ArgVT.getTypeForEVT(*DAG.getContext());
15123 uint32_t ArgSize = DAG.getDataLayout().getTypeAllocSize(ArgTy);
15126 // Decide which area this value should be read from.
15127 // TODO: Implement the AMD64 ABI in its entirety. This simple
15128 // selection mechanism works only for the basic types.
15129 if (ArgVT == MVT::f80) {
15130 llvm_unreachable("va_arg for f80 not yet implemented");
15131 } else if (ArgVT.isFloatingPoint() && ArgSize <= 16 /*bytes*/) {
15132 ArgMode = 2; // Argument passed in XMM register. Use fp_offset.
15133 } else if (ArgVT.isInteger() && ArgSize <= 32 /*bytes*/) {
15134 ArgMode = 1; // Argument passed in GPR64 register(s). Use gp_offset.
15136 llvm_unreachable("Unhandled argument type in LowerVAARG");
15139 if (ArgMode == 2) {
15140 // Sanity Check: Make sure using fp_offset makes sense.
15141 assert(!Subtarget->useSoftFloat() &&
15142 !(DAG.getMachineFunction().getFunction()->hasFnAttribute(
15143 Attribute::NoImplicitFloat)) &&
15144 Subtarget->hasSSE1());
15147 // Insert VAARG_64 node into the DAG
15148 // VAARG_64 returns two values: Variable Argument Address, Chain
15149 SDValue InstOps[] = {Chain, SrcPtr, DAG.getConstant(ArgSize, dl, MVT::i32),
15150 DAG.getConstant(ArgMode, dl, MVT::i8),
15151 DAG.getConstant(Align, dl, MVT::i32)};
15152 SDVTList VTs = DAG.getVTList(getPointerTy(DAG.getDataLayout()), MVT::Other);
15153 SDValue VAARG = DAG.getMemIntrinsicNode(X86ISD::VAARG_64, dl,
15154 VTs, InstOps, MVT::i64,
15155 MachinePointerInfo(SV),
15157 /*Volatile=*/false,
15159 /*WriteMem=*/true);
15160 Chain = VAARG.getValue(1);
15162 // Load the next argument and return it
15163 return DAG.getLoad(ArgVT, dl,
15166 MachinePointerInfo(),
15167 false, false, false, 0);
15170 static SDValue LowerVACOPY(SDValue Op, const X86Subtarget *Subtarget,
15171 SelectionDAG &DAG) {
15172 // X86-64 va_list is a struct { i32, i32, i8*, i8* }.
15173 assert(Subtarget->is64Bit() && "This code only handles 64-bit va_copy!");
15174 SDValue Chain = Op.getOperand(0);
15175 SDValue DstPtr = Op.getOperand(1);
15176 SDValue SrcPtr = Op.getOperand(2);
15177 const Value *DstSV = cast<SrcValueSDNode>(Op.getOperand(3))->getValue();
15178 const Value *SrcSV = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
15181 return DAG.getMemcpy(Chain, DL, DstPtr, SrcPtr,
15182 DAG.getIntPtrConstant(24, DL), 8, /*isVolatile*/false,
15184 MachinePointerInfo(DstSV), MachinePointerInfo(SrcSV));
15187 // getTargetVShiftByConstNode - Handle vector element shifts where the shift
15188 // amount is a constant. Takes immediate version of shift as input.
15189 static SDValue getTargetVShiftByConstNode(unsigned Opc, SDLoc dl, MVT VT,
15190 SDValue SrcOp, uint64_t ShiftAmt,
15191 SelectionDAG &DAG) {
15192 MVT ElementType = VT.getVectorElementType();
15194 // Fold this packed shift into its first operand if ShiftAmt is 0.
15198 // Check for ShiftAmt >= element width
15199 if (ShiftAmt >= ElementType.getSizeInBits()) {
15200 if (Opc == X86ISD::VSRAI)
15201 ShiftAmt = ElementType.getSizeInBits() - 1;
15203 return DAG.getConstant(0, dl, VT);
15206 assert((Opc == X86ISD::VSHLI || Opc == X86ISD::VSRLI || Opc == X86ISD::VSRAI)
15207 && "Unknown target vector shift-by-constant node");
15209 // Fold this packed vector shift into a build vector if SrcOp is a
15210 // vector of Constants or UNDEFs, and SrcOp valuetype is the same as VT.
15211 if (VT == SrcOp.getSimpleValueType() &&
15212 ISD::isBuildVectorOfConstantSDNodes(SrcOp.getNode())) {
15213 SmallVector<SDValue, 8> Elts;
15214 unsigned NumElts = SrcOp->getNumOperands();
15215 ConstantSDNode *ND;
15218 default: llvm_unreachable(nullptr);
15219 case X86ISD::VSHLI:
15220 for (unsigned i=0; i!=NumElts; ++i) {
15221 SDValue CurrentOp = SrcOp->getOperand(i);
15222 if (CurrentOp->getOpcode() == ISD::UNDEF) {
15223 Elts.push_back(CurrentOp);
15226 ND = cast<ConstantSDNode>(CurrentOp);
15227 const APInt &C = ND->getAPIntValue();
15228 Elts.push_back(DAG.getConstant(C.shl(ShiftAmt), dl, ElementType));
15231 case X86ISD::VSRLI:
15232 for (unsigned i=0; i!=NumElts; ++i) {
15233 SDValue CurrentOp = SrcOp->getOperand(i);
15234 if (CurrentOp->getOpcode() == ISD::UNDEF) {
15235 Elts.push_back(CurrentOp);
15238 ND = cast<ConstantSDNode>(CurrentOp);
15239 const APInt &C = ND->getAPIntValue();
15240 Elts.push_back(DAG.getConstant(C.lshr(ShiftAmt), dl, ElementType));
15243 case X86ISD::VSRAI:
15244 for (unsigned i=0; i!=NumElts; ++i) {
15245 SDValue CurrentOp = SrcOp->getOperand(i);
15246 if (CurrentOp->getOpcode() == ISD::UNDEF) {
15247 Elts.push_back(CurrentOp);
15250 ND = cast<ConstantSDNode>(CurrentOp);
15251 const APInt &C = ND->getAPIntValue();
15252 Elts.push_back(DAG.getConstant(C.ashr(ShiftAmt), dl, ElementType));
15257 return DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Elts);
15260 return DAG.getNode(Opc, dl, VT, SrcOp,
15261 DAG.getConstant(ShiftAmt, dl, MVT::i8));
15264 // getTargetVShiftNode - Handle vector element shifts where the shift amount
15265 // may or may not be a constant. Takes immediate version of shift as input.
15266 static SDValue getTargetVShiftNode(unsigned Opc, SDLoc dl, MVT VT,
15267 SDValue SrcOp, SDValue ShAmt,
15268 SelectionDAG &DAG) {
15269 MVT SVT = ShAmt.getSimpleValueType();
15270 assert((SVT == MVT::i32 || SVT == MVT::i64) && "Unexpected value type!");
15272 // Catch shift-by-constant.
15273 if (ConstantSDNode *CShAmt = dyn_cast<ConstantSDNode>(ShAmt))
15274 return getTargetVShiftByConstNode(Opc, dl, VT, SrcOp,
15275 CShAmt->getZExtValue(), DAG);
15277 // Change opcode to non-immediate version
15279 default: llvm_unreachable("Unknown target vector shift node");
15280 case X86ISD::VSHLI: Opc = X86ISD::VSHL; break;
15281 case X86ISD::VSRLI: Opc = X86ISD::VSRL; break;
15282 case X86ISD::VSRAI: Opc = X86ISD::VSRA; break;
15285 const X86Subtarget &Subtarget =
15286 static_cast<const X86Subtarget &>(DAG.getSubtarget());
15287 if (Subtarget.hasSSE41() && ShAmt.getOpcode() == ISD::ZERO_EXTEND &&
15288 ShAmt.getOperand(0).getSimpleValueType() == MVT::i16) {
15289 // Let the shuffle legalizer expand this shift amount node.
15290 SDValue Op0 = ShAmt.getOperand(0);
15291 Op0 = DAG.getNode(ISD::SCALAR_TO_VECTOR, SDLoc(Op0), MVT::v8i16, Op0);
15292 ShAmt = getShuffleVectorZeroOrUndef(Op0, 0, true, &Subtarget, DAG);
15294 // Need to build a vector containing shift amount.
15295 // SSE/AVX packed shifts only use the lower 64-bit of the shift count.
15296 SmallVector<SDValue, 4> ShOps;
15297 ShOps.push_back(ShAmt);
15298 if (SVT == MVT::i32) {
15299 ShOps.push_back(DAG.getConstant(0, dl, SVT));
15300 ShOps.push_back(DAG.getUNDEF(SVT));
15302 ShOps.push_back(DAG.getUNDEF(SVT));
15304 MVT BVT = SVT == MVT::i32 ? MVT::v4i32 : MVT::v2i64;
15305 ShAmt = DAG.getNode(ISD::BUILD_VECTOR, dl, BVT, ShOps);
15308 // The return type has to be a 128-bit type with the same element
15309 // type as the input type.
15310 MVT EltVT = VT.getVectorElementType();
15311 EVT ShVT = MVT::getVectorVT(EltVT, 128/EltVT.getSizeInBits());
15313 ShAmt = DAG.getBitcast(ShVT, ShAmt);
15314 return DAG.getNode(Opc, dl, VT, SrcOp, ShAmt);
15317 /// \brief Return (and \p Op, \p Mask) for compare instructions or
15318 /// (vselect \p Mask, \p Op, \p PreservedSrc) for others along with the
15319 /// necessary casting or extending for \p Mask when lowering masking intrinsics
15320 static SDValue getVectorMaskingNode(SDValue Op, SDValue Mask,
15321 SDValue PreservedSrc,
15322 const X86Subtarget *Subtarget,
15323 SelectionDAG &DAG) {
15324 EVT VT = Op.getValueType();
15325 EVT MaskVT = EVT::getVectorVT(*DAG.getContext(),
15326 MVT::i1, VT.getVectorNumElements());
15327 SDValue VMask = SDValue();
15328 unsigned OpcodeSelect = ISD::VSELECT;
15331 assert(MaskVT.isSimple() && "invalid mask type");
15333 if (isAllOnes(Mask))
15336 if (MaskVT.bitsGT(Mask.getValueType())) {
15337 EVT newMaskVT = EVT::getIntegerVT(*DAG.getContext(),
15338 MaskVT.getSizeInBits());
15339 VMask = DAG.getBitcast(MaskVT,
15340 DAG.getNode(ISD::ANY_EXTEND, dl, newMaskVT, Mask));
15342 EVT BitcastVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
15343 Mask.getValueType().getSizeInBits());
15344 // In case when MaskVT equals v2i1 or v4i1, low 2 or 4 elements
15345 // are extracted by EXTRACT_SUBVECTOR.
15346 VMask = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MaskVT,
15347 DAG.getBitcast(BitcastVT, Mask),
15348 DAG.getIntPtrConstant(0, dl));
15351 switch (Op.getOpcode()) {
15353 case X86ISD::PCMPEQM:
15354 case X86ISD::PCMPGTM:
15356 case X86ISD::CMPMU:
15357 return DAG.getNode(ISD::AND, dl, VT, Op, VMask);
15358 case X86ISD::VTRUNC:
15359 case X86ISD::VTRUNCS:
15360 case X86ISD::VTRUNCUS:
15361 // We can't use ISD::VSELECT here because it is not always "Legal"
15362 // for the destination type. For example vpmovqb require only AVX512
15363 // and vselect that can operate on byte element type require BWI
15364 OpcodeSelect = X86ISD::SELECT;
15367 if (PreservedSrc.getOpcode() == ISD::UNDEF)
15368 PreservedSrc = getZeroVector(VT, Subtarget, DAG, dl);
15369 return DAG.getNode(OpcodeSelect, dl, VT, VMask, Op, PreservedSrc);
15372 /// \brief Creates an SDNode for a predicated scalar operation.
15373 /// \returns (X86vselect \p Mask, \p Op, \p PreservedSrc).
15374 /// The mask is coming as MVT::i8 and it should be truncated
15375 /// to MVT::i1 while lowering masking intrinsics.
15376 /// The main difference between ScalarMaskingNode and VectorMaskingNode is using
15377 /// "X86select" instead of "vselect". We just can't create the "vselect" node
15378 /// for a scalar instruction.
15379 static SDValue getScalarMaskingNode(SDValue Op, SDValue Mask,
15380 SDValue PreservedSrc,
15381 const X86Subtarget *Subtarget,
15382 SelectionDAG &DAG) {
15383 if (isAllOnes(Mask))
15386 EVT VT = Op.getValueType();
15388 // The mask should be of type MVT::i1
15389 SDValue IMask = DAG.getNode(ISD::TRUNCATE, dl, MVT::i1, Mask);
15391 if (PreservedSrc.getOpcode() == ISD::UNDEF)
15392 PreservedSrc = getZeroVector(VT, Subtarget, DAG, dl);
15393 return DAG.getNode(X86ISD::SELECT, dl, VT, IMask, Op, PreservedSrc);
15396 static int getSEHRegistrationNodeSize(const Function *Fn) {
15397 if (!Fn->hasPersonalityFn())
15398 report_fatal_error(
15399 "querying registration node size for function without personality");
15400 // The RegNodeSize is 6 32-bit words for SEH and 4 for C++ EH. See
15401 // WinEHStatePass for the full struct definition.
15402 switch (classifyEHPersonality(Fn->getPersonalityFn())) {
15403 case EHPersonality::MSVC_X86SEH: return 24;
15404 case EHPersonality::MSVC_CXX: return 16;
15407 report_fatal_error("can only recover FP for MSVC EH personality functions");
15410 /// When the 32-bit MSVC runtime transfers control to us, either to an outlined
15411 /// function or when returning to a parent frame after catching an exception, we
15412 /// recover the parent frame pointer by doing arithmetic on the incoming EBP.
15413 /// Here's the math:
15414 /// RegNodeBase = EntryEBP - RegNodeSize
15415 /// ParentFP = RegNodeBase - RegNodeFrameOffset
15416 /// Subtracting RegNodeSize takes us to the offset of the registration node, and
15417 /// subtracting the offset (negative on x86) takes us back to the parent FP.
15418 static SDValue recoverFramePointer(SelectionDAG &DAG, const Function *Fn,
15419 SDValue EntryEBP) {
15420 MachineFunction &MF = DAG.getMachineFunction();
15423 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
15424 MVT PtrVT = TLI.getPointerTy(DAG.getDataLayout());
15426 // It's possible that the parent function no longer has a personality function
15427 // if the exceptional code was optimized away, in which case we just return
15428 // the incoming EBP.
15429 if (!Fn->hasPersonalityFn())
15432 int RegNodeSize = getSEHRegistrationNodeSize(Fn);
15434 // Get an MCSymbol that will ultimately resolve to the frame offset of the EH
15436 MCSymbol *OffsetSym =
15437 MF.getMMI().getContext().getOrCreateParentFrameOffsetSymbol(
15438 GlobalValue::getRealLinkageName(Fn->getName()));
15439 SDValue OffsetSymVal = DAG.getMCSymbol(OffsetSym, PtrVT);
15440 SDValue RegNodeFrameOffset =
15441 DAG.getNode(ISD::LOCAL_RECOVER, dl, PtrVT, OffsetSymVal);
15443 // RegNodeBase = EntryEBP - RegNodeSize
15444 // ParentFP = RegNodeBase - RegNodeFrameOffset
15445 SDValue RegNodeBase = DAG.getNode(ISD::SUB, dl, PtrVT, EntryEBP,
15446 DAG.getConstant(RegNodeSize, dl, PtrVT));
15447 return DAG.getNode(ISD::SUB, dl, PtrVT, RegNodeBase, RegNodeFrameOffset);
15450 static SDValue LowerINTRINSIC_WO_CHAIN(SDValue Op, const X86Subtarget *Subtarget,
15451 SelectionDAG &DAG) {
15453 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
15454 EVT VT = Op.getValueType();
15455 const IntrinsicData* IntrData = getIntrinsicWithoutChain(IntNo);
15457 switch(IntrData->Type) {
15458 case INTR_TYPE_1OP:
15459 return DAG.getNode(IntrData->Opc0, dl, Op.getValueType(), Op.getOperand(1));
15460 case INTR_TYPE_2OP:
15461 return DAG.getNode(IntrData->Opc0, dl, Op.getValueType(), Op.getOperand(1),
15463 case INTR_TYPE_3OP:
15464 return DAG.getNode(IntrData->Opc0, dl, Op.getValueType(), Op.getOperand(1),
15465 Op.getOperand(2), Op.getOperand(3));
15466 case INTR_TYPE_4OP:
15467 return DAG.getNode(IntrData->Opc0, dl, Op.getValueType(), Op.getOperand(1),
15468 Op.getOperand(2), Op.getOperand(3), Op.getOperand(4));
15469 case INTR_TYPE_1OP_MASK_RM: {
15470 SDValue Src = Op.getOperand(1);
15471 SDValue PassThru = Op.getOperand(2);
15472 SDValue Mask = Op.getOperand(3);
15473 SDValue RoundingMode;
15474 // We allways add rounding mode to the Node.
15475 // If the rounding mode is not specified, we add the
15476 // "current direction" mode.
15477 if (Op.getNumOperands() == 4)
15479 DAG.getConstant(X86::STATIC_ROUNDING::CUR_DIRECTION, dl, MVT::i32);
15481 RoundingMode = Op.getOperand(4);
15482 unsigned IntrWithRoundingModeOpcode = IntrData->Opc1;
15483 if (IntrWithRoundingModeOpcode != 0)
15484 if (cast<ConstantSDNode>(RoundingMode)->getZExtValue() !=
15485 X86::STATIC_ROUNDING::CUR_DIRECTION)
15486 return getVectorMaskingNode(DAG.getNode(IntrWithRoundingModeOpcode,
15487 dl, Op.getValueType(), Src, RoundingMode),
15488 Mask, PassThru, Subtarget, DAG);
15489 return getVectorMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT, Src,
15491 Mask, PassThru, Subtarget, DAG);
15493 case INTR_TYPE_1OP_MASK: {
15494 SDValue Src = Op.getOperand(1);
15495 SDValue PassThru = Op.getOperand(2);
15496 SDValue Mask = Op.getOperand(3);
15497 // We add rounding mode to the Node when
15498 // - RM Opcode is specified and
15499 // - RM is not "current direction".
15500 unsigned IntrWithRoundingModeOpcode = IntrData->Opc1;
15501 if (IntrWithRoundingModeOpcode != 0) {
15502 SDValue Rnd = Op.getOperand(4);
15503 unsigned Round = cast<ConstantSDNode>(Rnd)->getZExtValue();
15504 if (Round != X86::STATIC_ROUNDING::CUR_DIRECTION) {
15505 return getVectorMaskingNode(DAG.getNode(IntrWithRoundingModeOpcode,
15506 dl, Op.getValueType(),
15508 Mask, PassThru, Subtarget, DAG);
15511 return getVectorMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT, Src),
15512 Mask, PassThru, Subtarget, DAG);
15514 case INTR_TYPE_SCALAR_MASK_RM: {
15515 SDValue Src1 = Op.getOperand(1);
15516 SDValue Src2 = Op.getOperand(2);
15517 SDValue Src0 = Op.getOperand(3);
15518 SDValue Mask = Op.getOperand(4);
15519 // There are 2 kinds of intrinsics in this group:
15520 // (1) With suppress-all-exceptions (sae) or rounding mode- 6 operands
15521 // (2) With rounding mode and sae - 7 operands.
15522 if (Op.getNumOperands() == 6) {
15523 SDValue Sae = Op.getOperand(5);
15524 unsigned Opc = IntrData->Opc1 ? IntrData->Opc1 : IntrData->Opc0;
15525 return getScalarMaskingNode(DAG.getNode(Opc, dl, VT, Src1, Src2,
15527 Mask, Src0, Subtarget, DAG);
15529 assert(Op.getNumOperands() == 7 && "Unexpected intrinsic form");
15530 SDValue RoundingMode = Op.getOperand(5);
15531 SDValue Sae = Op.getOperand(6);
15532 return getScalarMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT, Src1, Src2,
15533 RoundingMode, Sae),
15534 Mask, Src0, Subtarget, DAG);
15536 case INTR_TYPE_2OP_MASK: {
15537 SDValue Src1 = Op.getOperand(1);
15538 SDValue Src2 = Op.getOperand(2);
15539 SDValue PassThru = Op.getOperand(3);
15540 SDValue Mask = Op.getOperand(4);
15541 // We specify 2 possible opcodes for intrinsics with rounding modes.
15542 // First, we check if the intrinsic may have non-default rounding mode,
15543 // (IntrData->Opc1 != 0), then we check the rounding mode operand.
15544 unsigned IntrWithRoundingModeOpcode = IntrData->Opc1;
15545 if (IntrWithRoundingModeOpcode != 0) {
15546 SDValue Rnd = Op.getOperand(5);
15547 unsigned Round = cast<ConstantSDNode>(Rnd)->getZExtValue();
15548 if (Round != X86::STATIC_ROUNDING::CUR_DIRECTION) {
15549 return getVectorMaskingNode(DAG.getNode(IntrWithRoundingModeOpcode,
15550 dl, Op.getValueType(),
15552 Mask, PassThru, Subtarget, DAG);
15555 return getVectorMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT,
15557 Mask, PassThru, Subtarget, DAG);
15559 case INTR_TYPE_2OP_MASK_RM: {
15560 SDValue Src1 = Op.getOperand(1);
15561 SDValue Src2 = Op.getOperand(2);
15562 SDValue PassThru = Op.getOperand(3);
15563 SDValue Mask = Op.getOperand(4);
15564 // We specify 2 possible modes for intrinsics, with/without rounding modes.
15565 // First, we check if the intrinsic have rounding mode (6 operands),
15566 // if not, we set rounding mode to "current".
15568 if (Op.getNumOperands() == 6)
15569 Rnd = Op.getOperand(5);
15571 Rnd = DAG.getConstant(X86::STATIC_ROUNDING::CUR_DIRECTION, dl, MVT::i32);
15572 return getVectorMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT,
15574 Mask, PassThru, Subtarget, DAG);
15576 case INTR_TYPE_3OP_MASK_RM: {
15577 SDValue Src1 = Op.getOperand(1);
15578 SDValue Src2 = Op.getOperand(2);
15579 SDValue Imm = Op.getOperand(3);
15580 SDValue PassThru = Op.getOperand(4);
15581 SDValue Mask = Op.getOperand(5);
15582 // We specify 2 possible modes for intrinsics, with/without rounding modes.
15583 // First, we check if the intrinsic have rounding mode (7 operands),
15584 // if not, we set rounding mode to "current".
15586 if (Op.getNumOperands() == 7)
15587 Rnd = Op.getOperand(6);
15589 Rnd = DAG.getConstant(X86::STATIC_ROUNDING::CUR_DIRECTION, dl, MVT::i32);
15590 return getVectorMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT,
15591 Src1, Src2, Imm, Rnd),
15592 Mask, PassThru, Subtarget, DAG);
15594 case INTR_TYPE_3OP_MASK: {
15595 SDValue Src1 = Op.getOperand(1);
15596 SDValue Src2 = Op.getOperand(2);
15597 SDValue Src3 = Op.getOperand(3);
15598 SDValue PassThru = Op.getOperand(4);
15599 SDValue Mask = Op.getOperand(5);
15600 // We specify 2 possible opcodes for intrinsics with rounding modes.
15601 // First, we check if the intrinsic may have non-default rounding mode,
15602 // (IntrData->Opc1 != 0), then we check the rounding mode operand.
15603 unsigned IntrWithRoundingModeOpcode = IntrData->Opc1;
15604 if (IntrWithRoundingModeOpcode != 0) {
15605 SDValue Rnd = Op.getOperand(6);
15606 unsigned Round = cast<ConstantSDNode>(Rnd)->getZExtValue();
15607 if (Round != X86::STATIC_ROUNDING::CUR_DIRECTION) {
15608 return getVectorMaskingNode(DAG.getNode(IntrWithRoundingModeOpcode,
15609 dl, Op.getValueType(),
15610 Src1, Src2, Src3, Rnd),
15611 Mask, PassThru, Subtarget, DAG);
15614 return getVectorMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT,
15616 Mask, PassThru, Subtarget, DAG);
15618 case VPERM_3OP_MASKZ:
15619 case VPERM_3OP_MASK:
15622 case FMA_OP_MASK: {
15623 SDValue Src1 = Op.getOperand(1);
15624 SDValue Src2 = Op.getOperand(2);
15625 SDValue Src3 = Op.getOperand(3);
15626 SDValue Mask = Op.getOperand(4);
15627 EVT VT = Op.getValueType();
15628 SDValue PassThru = SDValue();
15630 // set PassThru element
15631 if (IntrData->Type == VPERM_3OP_MASKZ || IntrData->Type == FMA_OP_MASKZ)
15632 PassThru = getZeroVector(VT, Subtarget, DAG, dl);
15633 else if (IntrData->Type == FMA_OP_MASK3)
15638 // We specify 2 possible opcodes for intrinsics with rounding modes.
15639 // First, we check if the intrinsic may have non-default rounding mode,
15640 // (IntrData->Opc1 != 0), then we check the rounding mode operand.
15641 unsigned IntrWithRoundingModeOpcode = IntrData->Opc1;
15642 if (IntrWithRoundingModeOpcode != 0) {
15643 SDValue Rnd = Op.getOperand(5);
15644 if (cast<ConstantSDNode>(Rnd)->getZExtValue() !=
15645 X86::STATIC_ROUNDING::CUR_DIRECTION)
15646 return getVectorMaskingNode(DAG.getNode(IntrWithRoundingModeOpcode,
15647 dl, Op.getValueType(),
15648 Src1, Src2, Src3, Rnd),
15649 Mask, PassThru, Subtarget, DAG);
15651 return getVectorMaskingNode(DAG.getNode(IntrData->Opc0,
15652 dl, Op.getValueType(),
15654 Mask, PassThru, Subtarget, DAG);
15657 case CMP_MASK_CC: {
15658 // Comparison intrinsics with masks.
15659 // Example of transformation:
15660 // (i8 (int_x86_avx512_mask_pcmpeq_q_128
15661 // (v2i64 %a), (v2i64 %b), (i8 %mask))) ->
15663 // (v8i1 (insert_subvector undef,
15664 // (v2i1 (and (PCMPEQM %a, %b),
15665 // (extract_subvector
15666 // (v8i1 (bitcast %mask)), 0))), 0))))
15667 EVT VT = Op.getOperand(1).getValueType();
15668 EVT MaskVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
15669 VT.getVectorNumElements());
15670 SDValue Mask = Op.getOperand((IntrData->Type == CMP_MASK_CC) ? 4 : 3);
15671 EVT BitcastVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
15672 Mask.getValueType().getSizeInBits());
15674 if (IntrData->Type == CMP_MASK_CC) {
15675 SDValue CC = Op.getOperand(3);
15676 CC = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, CC);
15677 // We specify 2 possible opcodes for intrinsics with rounding modes.
15678 // First, we check if the intrinsic may have non-default rounding mode,
15679 // (IntrData->Opc1 != 0), then we check the rounding mode operand.
15680 if (IntrData->Opc1 != 0) {
15681 SDValue Rnd = Op.getOperand(5);
15682 if (cast<ConstantSDNode>(Rnd)->getZExtValue() !=
15683 X86::STATIC_ROUNDING::CUR_DIRECTION)
15684 Cmp = DAG.getNode(IntrData->Opc1, dl, MaskVT, Op.getOperand(1),
15685 Op.getOperand(2), CC, Rnd);
15687 //default rounding mode
15689 Cmp = DAG.getNode(IntrData->Opc0, dl, MaskVT, Op.getOperand(1),
15690 Op.getOperand(2), CC);
15693 assert(IntrData->Type == CMP_MASK && "Unexpected intrinsic type!");
15694 Cmp = DAG.getNode(IntrData->Opc0, dl, MaskVT, Op.getOperand(1),
15697 SDValue CmpMask = getVectorMaskingNode(Cmp, Mask,
15698 DAG.getTargetConstant(0, dl,
15701 SDValue Res = DAG.getNode(ISD::INSERT_SUBVECTOR, dl, BitcastVT,
15702 DAG.getUNDEF(BitcastVT), CmpMask,
15703 DAG.getIntPtrConstant(0, dl));
15704 return DAG.getBitcast(Op.getValueType(), Res);
15706 case COMI: { // Comparison intrinsics
15707 ISD::CondCode CC = (ISD::CondCode)IntrData->Opc1;
15708 SDValue LHS = Op.getOperand(1);
15709 SDValue RHS = Op.getOperand(2);
15710 unsigned X86CC = TranslateX86CC(CC, dl, true, LHS, RHS, DAG);
15711 assert(X86CC != X86::COND_INVALID && "Unexpected illegal condition!");
15712 SDValue Cond = DAG.getNode(IntrData->Opc0, dl, MVT::i32, LHS, RHS);
15713 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
15714 DAG.getConstant(X86CC, dl, MVT::i8), Cond);
15715 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
15718 return getTargetVShiftNode(IntrData->Opc0, dl, Op.getSimpleValueType(),
15719 Op.getOperand(1), Op.getOperand(2), DAG);
15721 return getVectorMaskingNode(getTargetVShiftNode(IntrData->Opc0, dl,
15722 Op.getSimpleValueType(),
15724 Op.getOperand(2), DAG),
15725 Op.getOperand(4), Op.getOperand(3), Subtarget,
15727 case COMPRESS_EXPAND_IN_REG: {
15728 SDValue Mask = Op.getOperand(3);
15729 SDValue DataToCompress = Op.getOperand(1);
15730 SDValue PassThru = Op.getOperand(2);
15731 if (isAllOnes(Mask)) // return data as is
15732 return Op.getOperand(1);
15734 return getVectorMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT,
15736 Mask, PassThru, Subtarget, DAG);
15739 SDValue Mask = Op.getOperand(3);
15740 EVT VT = Op.getValueType();
15741 EVT MaskVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
15742 VT.getVectorNumElements());
15743 EVT BitcastVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
15744 Mask.getValueType().getSizeInBits());
15746 SDValue VMask = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MaskVT,
15747 DAG.getBitcast(BitcastVT, Mask),
15748 DAG.getIntPtrConstant(0, dl));
15749 return DAG.getNode(IntrData->Opc0, dl, VT, VMask, Op.getOperand(1),
15758 default: return SDValue(); // Don't custom lower most intrinsics.
15760 case Intrinsic::x86_avx2_permd:
15761 case Intrinsic::x86_avx2_permps:
15762 // Operands intentionally swapped. Mask is last operand to intrinsic,
15763 // but second operand for node/instruction.
15764 return DAG.getNode(X86ISD::VPERMV, dl, Op.getValueType(),
15765 Op.getOperand(2), Op.getOperand(1));
15767 // ptest and testp intrinsics. The intrinsic these come from are designed to
15768 // return an integer value, not just an instruction so lower it to the ptest
15769 // or testp pattern and a setcc for the result.
15770 case Intrinsic::x86_sse41_ptestz:
15771 case Intrinsic::x86_sse41_ptestc:
15772 case Intrinsic::x86_sse41_ptestnzc:
15773 case Intrinsic::x86_avx_ptestz_256:
15774 case Intrinsic::x86_avx_ptestc_256:
15775 case Intrinsic::x86_avx_ptestnzc_256:
15776 case Intrinsic::x86_avx_vtestz_ps:
15777 case Intrinsic::x86_avx_vtestc_ps:
15778 case Intrinsic::x86_avx_vtestnzc_ps:
15779 case Intrinsic::x86_avx_vtestz_pd:
15780 case Intrinsic::x86_avx_vtestc_pd:
15781 case Intrinsic::x86_avx_vtestnzc_pd:
15782 case Intrinsic::x86_avx_vtestz_ps_256:
15783 case Intrinsic::x86_avx_vtestc_ps_256:
15784 case Intrinsic::x86_avx_vtestnzc_ps_256:
15785 case Intrinsic::x86_avx_vtestz_pd_256:
15786 case Intrinsic::x86_avx_vtestc_pd_256:
15787 case Intrinsic::x86_avx_vtestnzc_pd_256: {
15788 bool IsTestPacked = false;
15791 default: llvm_unreachable("Bad fallthrough in Intrinsic lowering.");
15792 case Intrinsic::x86_avx_vtestz_ps:
15793 case Intrinsic::x86_avx_vtestz_pd:
15794 case Intrinsic::x86_avx_vtestz_ps_256:
15795 case Intrinsic::x86_avx_vtestz_pd_256:
15796 IsTestPacked = true; // Fallthrough
15797 case Intrinsic::x86_sse41_ptestz:
15798 case Intrinsic::x86_avx_ptestz_256:
15800 X86CC = X86::COND_E;
15802 case Intrinsic::x86_avx_vtestc_ps:
15803 case Intrinsic::x86_avx_vtestc_pd:
15804 case Intrinsic::x86_avx_vtestc_ps_256:
15805 case Intrinsic::x86_avx_vtestc_pd_256:
15806 IsTestPacked = true; // Fallthrough
15807 case Intrinsic::x86_sse41_ptestc:
15808 case Intrinsic::x86_avx_ptestc_256:
15810 X86CC = X86::COND_B;
15812 case Intrinsic::x86_avx_vtestnzc_ps:
15813 case Intrinsic::x86_avx_vtestnzc_pd:
15814 case Intrinsic::x86_avx_vtestnzc_ps_256:
15815 case Intrinsic::x86_avx_vtestnzc_pd_256:
15816 IsTestPacked = true; // Fallthrough
15817 case Intrinsic::x86_sse41_ptestnzc:
15818 case Intrinsic::x86_avx_ptestnzc_256:
15820 X86CC = X86::COND_A;
15824 SDValue LHS = Op.getOperand(1);
15825 SDValue RHS = Op.getOperand(2);
15826 unsigned TestOpc = IsTestPacked ? X86ISD::TESTP : X86ISD::PTEST;
15827 SDValue Test = DAG.getNode(TestOpc, dl, MVT::i32, LHS, RHS);
15828 SDValue CC = DAG.getConstant(X86CC, dl, MVT::i8);
15829 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8, CC, Test);
15830 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
15832 case Intrinsic::x86_avx512_kortestz_w:
15833 case Intrinsic::x86_avx512_kortestc_w: {
15834 unsigned X86CC = (IntNo == Intrinsic::x86_avx512_kortestz_w)? X86::COND_E: X86::COND_B;
15835 SDValue LHS = DAG.getBitcast(MVT::v16i1, Op.getOperand(1));
15836 SDValue RHS = DAG.getBitcast(MVT::v16i1, Op.getOperand(2));
15837 SDValue CC = DAG.getConstant(X86CC, dl, MVT::i8);
15838 SDValue Test = DAG.getNode(X86ISD::KORTEST, dl, MVT::i32, LHS, RHS);
15839 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i1, CC, Test);
15840 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
15843 case Intrinsic::x86_sse42_pcmpistria128:
15844 case Intrinsic::x86_sse42_pcmpestria128:
15845 case Intrinsic::x86_sse42_pcmpistric128:
15846 case Intrinsic::x86_sse42_pcmpestric128:
15847 case Intrinsic::x86_sse42_pcmpistrio128:
15848 case Intrinsic::x86_sse42_pcmpestrio128:
15849 case Intrinsic::x86_sse42_pcmpistris128:
15850 case Intrinsic::x86_sse42_pcmpestris128:
15851 case Intrinsic::x86_sse42_pcmpistriz128:
15852 case Intrinsic::x86_sse42_pcmpestriz128: {
15856 default: llvm_unreachable("Impossible intrinsic"); // Can't reach here.
15857 case Intrinsic::x86_sse42_pcmpistria128:
15858 Opcode = X86ISD::PCMPISTRI;
15859 X86CC = X86::COND_A;
15861 case Intrinsic::x86_sse42_pcmpestria128:
15862 Opcode = X86ISD::PCMPESTRI;
15863 X86CC = X86::COND_A;
15865 case Intrinsic::x86_sse42_pcmpistric128:
15866 Opcode = X86ISD::PCMPISTRI;
15867 X86CC = X86::COND_B;
15869 case Intrinsic::x86_sse42_pcmpestric128:
15870 Opcode = X86ISD::PCMPESTRI;
15871 X86CC = X86::COND_B;
15873 case Intrinsic::x86_sse42_pcmpistrio128:
15874 Opcode = X86ISD::PCMPISTRI;
15875 X86CC = X86::COND_O;
15877 case Intrinsic::x86_sse42_pcmpestrio128:
15878 Opcode = X86ISD::PCMPESTRI;
15879 X86CC = X86::COND_O;
15881 case Intrinsic::x86_sse42_pcmpistris128:
15882 Opcode = X86ISD::PCMPISTRI;
15883 X86CC = X86::COND_S;
15885 case Intrinsic::x86_sse42_pcmpestris128:
15886 Opcode = X86ISD::PCMPESTRI;
15887 X86CC = X86::COND_S;
15889 case Intrinsic::x86_sse42_pcmpistriz128:
15890 Opcode = X86ISD::PCMPISTRI;
15891 X86CC = X86::COND_E;
15893 case Intrinsic::x86_sse42_pcmpestriz128:
15894 Opcode = X86ISD::PCMPESTRI;
15895 X86CC = X86::COND_E;
15898 SmallVector<SDValue, 5> NewOps(Op->op_begin()+1, Op->op_end());
15899 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::i32);
15900 SDValue PCMP = DAG.getNode(Opcode, dl, VTs, NewOps);
15901 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
15902 DAG.getConstant(X86CC, dl, MVT::i8),
15903 SDValue(PCMP.getNode(), 1));
15904 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
15907 case Intrinsic::x86_sse42_pcmpistri128:
15908 case Intrinsic::x86_sse42_pcmpestri128: {
15910 if (IntNo == Intrinsic::x86_sse42_pcmpistri128)
15911 Opcode = X86ISD::PCMPISTRI;
15913 Opcode = X86ISD::PCMPESTRI;
15915 SmallVector<SDValue, 5> NewOps(Op->op_begin()+1, Op->op_end());
15916 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::i32);
15917 return DAG.getNode(Opcode, dl, VTs, NewOps);
15920 case Intrinsic::x86_seh_lsda: {
15921 // Compute the symbol for the LSDA. We know it'll get emitted later.
15922 MachineFunction &MF = DAG.getMachineFunction();
15923 SDValue Op1 = Op.getOperand(1);
15924 auto *Fn = cast<Function>(cast<GlobalAddressSDNode>(Op1)->getGlobal());
15925 MCSymbol *LSDASym = MF.getMMI().getContext().getOrCreateLSDASymbol(
15926 GlobalValue::getRealLinkageName(Fn->getName()));
15928 // Generate a simple absolute symbol reference. This intrinsic is only
15929 // supported on 32-bit Windows, which isn't PIC.
15930 SDValue Result = DAG.getMCSymbol(LSDASym, VT);
15931 return DAG.getNode(X86ISD::Wrapper, dl, VT, Result);
15934 case Intrinsic::x86_seh_recoverfp: {
15935 SDValue FnOp = Op.getOperand(1);
15936 SDValue IncomingFPOp = Op.getOperand(2);
15937 GlobalAddressSDNode *GSD = dyn_cast<GlobalAddressSDNode>(FnOp);
15938 auto *Fn = dyn_cast_or_null<Function>(GSD ? GSD->getGlobal() : nullptr);
15940 report_fatal_error(
15941 "llvm.x86.seh.recoverfp must take a function as the first argument");
15942 return recoverFramePointer(DAG, Fn, IncomingFPOp);
15945 case Intrinsic::localaddress: {
15946 // Returns one of the stack, base, or frame pointer registers, depending on
15947 // which is used to reference local variables.
15948 MachineFunction &MF = DAG.getMachineFunction();
15949 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
15951 if (RegInfo->hasBasePointer(MF))
15952 Reg = RegInfo->getBaseRegister();
15953 else // This function handles the SP or FP case.
15954 Reg = RegInfo->getPtrSizedFrameRegister(MF);
15955 return DAG.getCopyFromReg(DAG.getEntryNode(), dl, Reg, VT);
15960 static SDValue getGatherNode(unsigned Opc, SDValue Op, SelectionDAG &DAG,
15961 SDValue Src, SDValue Mask, SDValue Base,
15962 SDValue Index, SDValue ScaleOp, SDValue Chain,
15963 const X86Subtarget * Subtarget) {
15965 ConstantSDNode *C = dyn_cast<ConstantSDNode>(ScaleOp);
15967 llvm_unreachable("Invalid scale type");
15968 unsigned ScaleVal = C->getZExtValue();
15969 if (ScaleVal > 2 && ScaleVal != 4 && ScaleVal != 8)
15970 llvm_unreachable("Valid scale values are 1, 2, 4, 8");
15972 SDValue Scale = DAG.getTargetConstant(C->getZExtValue(), dl, MVT::i8);
15973 EVT MaskVT = MVT::getVectorVT(MVT::i1,
15974 Index.getSimpleValueType().getVectorNumElements());
15976 ConstantSDNode *MaskC = dyn_cast<ConstantSDNode>(Mask);
15978 MaskInReg = DAG.getTargetConstant(MaskC->getSExtValue(), dl, MaskVT);
15980 EVT BitcastVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
15981 Mask.getValueType().getSizeInBits());
15983 // In case when MaskVT equals v2i1 or v4i1, low 2 or 4 elements
15984 // are extracted by EXTRACT_SUBVECTOR.
15985 MaskInReg = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MaskVT,
15986 DAG.getBitcast(BitcastVT, Mask),
15987 DAG.getIntPtrConstant(0, dl));
15989 SDVTList VTs = DAG.getVTList(Op.getValueType(), MaskVT, MVT::Other);
15990 SDValue Disp = DAG.getTargetConstant(0, dl, MVT::i32);
15991 SDValue Segment = DAG.getRegister(0, MVT::i32);
15992 if (Src.getOpcode() == ISD::UNDEF)
15993 Src = getZeroVector(Op.getValueType(), Subtarget, DAG, dl);
15994 SDValue Ops[] = {Src, MaskInReg, Base, Scale, Index, Disp, Segment, Chain};
15995 SDNode *Res = DAG.getMachineNode(Opc, dl, VTs, Ops);
15996 SDValue RetOps[] = { SDValue(Res, 0), SDValue(Res, 2) };
15997 return DAG.getMergeValues(RetOps, dl);
16000 static SDValue getScatterNode(unsigned Opc, SDValue Op, SelectionDAG &DAG,
16001 SDValue Src, SDValue Mask, SDValue Base,
16002 SDValue Index, SDValue ScaleOp, SDValue Chain) {
16004 ConstantSDNode *C = dyn_cast<ConstantSDNode>(ScaleOp);
16006 llvm_unreachable("Invalid scale type");
16007 unsigned ScaleVal = C->getZExtValue();
16008 if (ScaleVal > 2 && ScaleVal != 4 && ScaleVal != 8)
16009 llvm_unreachable("Valid scale values are 1, 2, 4, 8");
16011 SDValue Scale = DAG.getTargetConstant(C->getZExtValue(), dl, MVT::i8);
16012 SDValue Disp = DAG.getTargetConstant(0, dl, MVT::i32);
16013 SDValue Segment = DAG.getRegister(0, MVT::i32);
16014 EVT MaskVT = MVT::getVectorVT(MVT::i1,
16015 Index.getSimpleValueType().getVectorNumElements());
16017 ConstantSDNode *MaskC = dyn_cast<ConstantSDNode>(Mask);
16019 MaskInReg = DAG.getTargetConstant(MaskC->getSExtValue(), dl, MaskVT);
16021 EVT BitcastVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
16022 Mask.getValueType().getSizeInBits());
16024 // In case when MaskVT equals v2i1 or v4i1, low 2 or 4 elements
16025 // are extracted by EXTRACT_SUBVECTOR.
16026 MaskInReg = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MaskVT,
16027 DAG.getBitcast(BitcastVT, Mask),
16028 DAG.getIntPtrConstant(0, dl));
16030 SDVTList VTs = DAG.getVTList(MaskVT, MVT::Other);
16031 SDValue Ops[] = {Base, Scale, Index, Disp, Segment, MaskInReg, Src, Chain};
16032 SDNode *Res = DAG.getMachineNode(Opc, dl, VTs, Ops);
16033 return SDValue(Res, 1);
16036 static SDValue getPrefetchNode(unsigned Opc, SDValue Op, SelectionDAG &DAG,
16037 SDValue Mask, SDValue Base, SDValue Index,
16038 SDValue ScaleOp, SDValue Chain) {
16040 ConstantSDNode *C = dyn_cast<ConstantSDNode>(ScaleOp);
16041 assert(C && "Invalid scale type");
16042 SDValue Scale = DAG.getTargetConstant(C->getZExtValue(), dl, MVT::i8);
16043 SDValue Disp = DAG.getTargetConstant(0, dl, MVT::i32);
16044 SDValue Segment = DAG.getRegister(0, MVT::i32);
16046 MVT::getVectorVT(MVT::i1, Index.getSimpleValueType().getVectorNumElements());
16048 ConstantSDNode *MaskC = dyn_cast<ConstantSDNode>(Mask);
16050 MaskInReg = DAG.getTargetConstant(MaskC->getSExtValue(), dl, MaskVT);
16052 MaskInReg = DAG.getBitcast(MaskVT, Mask);
16053 //SDVTList VTs = DAG.getVTList(MVT::Other);
16054 SDValue Ops[] = {MaskInReg, Base, Scale, Index, Disp, Segment, Chain};
16055 SDNode *Res = DAG.getMachineNode(Opc, dl, MVT::Other, Ops);
16056 return SDValue(Res, 0);
16059 // getReadPerformanceCounter - Handles the lowering of builtin intrinsics that
16060 // read performance monitor counters (x86_rdpmc).
16061 static void getReadPerformanceCounter(SDNode *N, SDLoc DL,
16062 SelectionDAG &DAG, const X86Subtarget *Subtarget,
16063 SmallVectorImpl<SDValue> &Results) {
16064 assert(N->getNumOperands() == 3 && "Unexpected number of operands!");
16065 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Glue);
16068 // The ECX register is used to select the index of the performance counter
16070 SDValue Chain = DAG.getCopyToReg(N->getOperand(0), DL, X86::ECX,
16072 SDValue rd = DAG.getNode(X86ISD::RDPMC_DAG, DL, Tys, Chain);
16074 // Reads the content of a 64-bit performance counter and returns it in the
16075 // registers EDX:EAX.
16076 if (Subtarget->is64Bit()) {
16077 LO = DAG.getCopyFromReg(rd, DL, X86::RAX, MVT::i64, rd.getValue(1));
16078 HI = DAG.getCopyFromReg(LO.getValue(1), DL, X86::RDX, MVT::i64,
16081 LO = DAG.getCopyFromReg(rd, DL, X86::EAX, MVT::i32, rd.getValue(1));
16082 HI = DAG.getCopyFromReg(LO.getValue(1), DL, X86::EDX, MVT::i32,
16085 Chain = HI.getValue(1);
16087 if (Subtarget->is64Bit()) {
16088 // The EAX register is loaded with the low-order 32 bits. The EDX register
16089 // is loaded with the supported high-order bits of the counter.
16090 SDValue Tmp = DAG.getNode(ISD::SHL, DL, MVT::i64, HI,
16091 DAG.getConstant(32, DL, MVT::i8));
16092 Results.push_back(DAG.getNode(ISD::OR, DL, MVT::i64, LO, Tmp));
16093 Results.push_back(Chain);
16097 // Use a buildpair to merge the two 32-bit values into a 64-bit one.
16098 SDValue Ops[] = { LO, HI };
16099 SDValue Pair = DAG.getNode(ISD::BUILD_PAIR, DL, MVT::i64, Ops);
16100 Results.push_back(Pair);
16101 Results.push_back(Chain);
16104 // getReadTimeStampCounter - Handles the lowering of builtin intrinsics that
16105 // read the time stamp counter (x86_rdtsc and x86_rdtscp). This function is
16106 // also used to custom lower READCYCLECOUNTER nodes.
16107 static void getReadTimeStampCounter(SDNode *N, SDLoc DL, unsigned Opcode,
16108 SelectionDAG &DAG, const X86Subtarget *Subtarget,
16109 SmallVectorImpl<SDValue> &Results) {
16110 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Glue);
16111 SDValue rd = DAG.getNode(Opcode, DL, Tys, N->getOperand(0));
16114 // The processor's time-stamp counter (a 64-bit MSR) is stored into the
16115 // EDX:EAX registers. EDX is loaded with the high-order 32 bits of the MSR
16116 // and the EAX register is loaded with the low-order 32 bits.
16117 if (Subtarget->is64Bit()) {
16118 LO = DAG.getCopyFromReg(rd, DL, X86::RAX, MVT::i64, rd.getValue(1));
16119 HI = DAG.getCopyFromReg(LO.getValue(1), DL, X86::RDX, MVT::i64,
16122 LO = DAG.getCopyFromReg(rd, DL, X86::EAX, MVT::i32, rd.getValue(1));
16123 HI = DAG.getCopyFromReg(LO.getValue(1), DL, X86::EDX, MVT::i32,
16126 SDValue Chain = HI.getValue(1);
16128 if (Opcode == X86ISD::RDTSCP_DAG) {
16129 assert(N->getNumOperands() == 3 && "Unexpected number of operands!");
16131 // Instruction RDTSCP loads the IA32:TSC_AUX_MSR (address C000_0103H) into
16132 // the ECX register. Add 'ecx' explicitly to the chain.
16133 SDValue ecx = DAG.getCopyFromReg(Chain, DL, X86::ECX, MVT::i32,
16135 // Explicitly store the content of ECX at the location passed in input
16136 // to the 'rdtscp' intrinsic.
16137 Chain = DAG.getStore(ecx.getValue(1), DL, ecx, N->getOperand(2),
16138 MachinePointerInfo(), false, false, 0);
16141 if (Subtarget->is64Bit()) {
16142 // The EDX register is loaded with the high-order 32 bits of the MSR, and
16143 // the EAX register is loaded with the low-order 32 bits.
16144 SDValue Tmp = DAG.getNode(ISD::SHL, DL, MVT::i64, HI,
16145 DAG.getConstant(32, DL, MVT::i8));
16146 Results.push_back(DAG.getNode(ISD::OR, DL, MVT::i64, LO, Tmp));
16147 Results.push_back(Chain);
16151 // Use a buildpair to merge the two 32-bit values into a 64-bit one.
16152 SDValue Ops[] = { LO, HI };
16153 SDValue Pair = DAG.getNode(ISD::BUILD_PAIR, DL, MVT::i64, Ops);
16154 Results.push_back(Pair);
16155 Results.push_back(Chain);
16158 static SDValue LowerREADCYCLECOUNTER(SDValue Op, const X86Subtarget *Subtarget,
16159 SelectionDAG &DAG) {
16160 SmallVector<SDValue, 2> Results;
16162 getReadTimeStampCounter(Op.getNode(), DL, X86ISD::RDTSC_DAG, DAG, Subtarget,
16164 return DAG.getMergeValues(Results, DL);
16167 static SDValue LowerSEHRESTOREFRAME(SDValue Op, const X86Subtarget *Subtarget,
16168 SelectionDAG &DAG) {
16169 MachineFunction &MF = DAG.getMachineFunction();
16170 const Function *Fn = MF.getFunction();
16172 SDValue Chain = Op.getOperand(0);
16174 assert(Subtarget->getFrameLowering()->hasFP(MF) &&
16175 "using llvm.x86.seh.restoreframe requires a frame pointer");
16177 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
16178 MVT VT = TLI.getPointerTy(DAG.getDataLayout());
16180 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
16181 unsigned FrameReg =
16182 RegInfo->getPtrSizedFrameRegister(DAG.getMachineFunction());
16183 unsigned SPReg = RegInfo->getStackRegister();
16184 unsigned SlotSize = RegInfo->getSlotSize();
16186 // Get incoming EBP.
16187 SDValue IncomingEBP =
16188 DAG.getCopyFromReg(Chain, dl, FrameReg, VT);
16190 // SP is saved in the first field of every registration node, so load
16191 // [EBP-RegNodeSize] into SP.
16192 int RegNodeSize = getSEHRegistrationNodeSize(Fn);
16193 SDValue SPAddr = DAG.getNode(ISD::ADD, dl, VT, IncomingEBP,
16194 DAG.getConstant(-RegNodeSize, dl, VT));
16196 DAG.getLoad(VT, dl, Chain, SPAddr, MachinePointerInfo(), false, false,
16197 false, VT.getScalarSizeInBits() / 8);
16198 Chain = DAG.getCopyToReg(Chain, dl, SPReg, NewSP);
16200 if (!RegInfo->needsStackRealignment(MF)) {
16201 // Adjust EBP to point back to the original frame position.
16202 SDValue NewFP = recoverFramePointer(DAG, Fn, IncomingEBP);
16203 Chain = DAG.getCopyToReg(Chain, dl, FrameReg, NewFP);
16205 assert(RegInfo->hasBasePointer(MF) &&
16206 "functions with Win32 EH must use frame or base pointer register");
16208 // Reload the base pointer (ESI) with the adjusted incoming EBP.
16209 SDValue NewBP = recoverFramePointer(DAG, Fn, IncomingEBP);
16210 Chain = DAG.getCopyToReg(Chain, dl, RegInfo->getBaseRegister(), NewBP);
16212 // Reload the spilled EBP value, now that the stack and base pointers are
16214 X86MachineFunctionInfo *X86FI = MF.getInfo<X86MachineFunctionInfo>();
16215 X86FI->setHasSEHFramePtrSave(true);
16216 int FI = MF.getFrameInfo()->CreateSpillStackObject(SlotSize, SlotSize);
16217 X86FI->setSEHFramePtrSaveIndex(FI);
16218 SDValue NewFP = DAG.getLoad(VT, dl, Chain, DAG.getFrameIndex(FI, VT),
16219 MachinePointerInfo(), false, false, false,
16220 VT.getScalarSizeInBits() / 8);
16221 Chain = DAG.getCopyToReg(NewFP, dl, FrameReg, NewFP);
16227 /// \brief Lower intrinsics for TRUNCATE_TO_MEM case
16228 /// return truncate Store/MaskedStore Node
16229 static SDValue LowerINTRINSIC_TRUNCATE_TO_MEM(const SDValue & Op,
16233 SDValue Mask = Op.getOperand(4);
16234 SDValue DataToTruncate = Op.getOperand(3);
16235 SDValue Addr = Op.getOperand(2);
16236 SDValue Chain = Op.getOperand(0);
16238 EVT VT = DataToTruncate.getValueType();
16239 EVT SVT = EVT::getVectorVT(*DAG.getContext(),
16240 ElementType, VT.getVectorNumElements());
16242 if (isAllOnes(Mask)) // return just a truncate store
16243 return DAG.getTruncStore(Chain, dl, DataToTruncate, Addr,
16244 MachinePointerInfo(), SVT, false, false,
16245 SVT.getScalarSizeInBits()/8);
16247 EVT MaskVT = EVT::getVectorVT(*DAG.getContext(),
16248 MVT::i1, VT.getVectorNumElements());
16249 EVT BitcastVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
16250 Mask.getValueType().getSizeInBits());
16251 // In case when MaskVT equals v2i1 or v4i1, low 2 or 4 elements
16252 // are extracted by EXTRACT_SUBVECTOR.
16253 SDValue VMask = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MaskVT,
16254 DAG.getBitcast(BitcastVT, Mask),
16255 DAG.getIntPtrConstant(0, dl));
16257 MachineMemOperand *MMO = DAG.getMachineFunction().
16258 getMachineMemOperand(MachinePointerInfo(),
16259 MachineMemOperand::MOStore, SVT.getStoreSize(),
16260 SVT.getScalarSizeInBits()/8);
16262 return DAG.getMaskedStore(Chain, dl, DataToTruncate, Addr,
16263 VMask, SVT, MMO, true);
16266 static SDValue LowerINTRINSIC_W_CHAIN(SDValue Op, const X86Subtarget *Subtarget,
16267 SelectionDAG &DAG) {
16268 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
16270 const IntrinsicData* IntrData = getIntrinsicWithChain(IntNo);
16272 if (IntNo == llvm::Intrinsic::x86_seh_restoreframe)
16273 return LowerSEHRESTOREFRAME(Op, Subtarget, DAG);
16278 switch(IntrData->Type) {
16280 llvm_unreachable("Unknown Intrinsic Type");
16284 // Emit the node with the right value type.
16285 SDVTList VTs = DAG.getVTList(Op->getValueType(0), MVT::Glue, MVT::Other);
16286 SDValue Result = DAG.getNode(IntrData->Opc0, dl, VTs, Op.getOperand(0));
16288 // If the value returned by RDRAND/RDSEED was valid (CF=1), return 1.
16289 // Otherwise return the value from Rand, which is always 0, casted to i32.
16290 SDValue Ops[] = { DAG.getZExtOrTrunc(Result, dl, Op->getValueType(1)),
16291 DAG.getConstant(1, dl, Op->getValueType(1)),
16292 DAG.getConstant(X86::COND_B, dl, MVT::i32),
16293 SDValue(Result.getNode(), 1) };
16294 SDValue isValid = DAG.getNode(X86ISD::CMOV, dl,
16295 DAG.getVTList(Op->getValueType(1), MVT::Glue),
16298 // Return { result, isValid, chain }.
16299 return DAG.getNode(ISD::MERGE_VALUES, dl, Op->getVTList(), Result, isValid,
16300 SDValue(Result.getNode(), 2));
16303 //gather(v1, mask, index, base, scale);
16304 SDValue Chain = Op.getOperand(0);
16305 SDValue Src = Op.getOperand(2);
16306 SDValue Base = Op.getOperand(3);
16307 SDValue Index = Op.getOperand(4);
16308 SDValue Mask = Op.getOperand(5);
16309 SDValue Scale = Op.getOperand(6);
16310 return getGatherNode(IntrData->Opc0, Op, DAG, Src, Mask, Base, Index, Scale,
16314 //scatter(base, mask, index, v1, scale);
16315 SDValue Chain = Op.getOperand(0);
16316 SDValue Base = Op.getOperand(2);
16317 SDValue Mask = Op.getOperand(3);
16318 SDValue Index = Op.getOperand(4);
16319 SDValue Src = Op.getOperand(5);
16320 SDValue Scale = Op.getOperand(6);
16321 return getScatterNode(IntrData->Opc0, Op, DAG, Src, Mask, Base, Index,
16325 SDValue Hint = Op.getOperand(6);
16326 unsigned HintVal = cast<ConstantSDNode>(Hint)->getZExtValue();
16327 assert(HintVal < 2 && "Wrong prefetch hint in intrinsic: should be 0 or 1");
16328 unsigned Opcode = (HintVal ? IntrData->Opc1 : IntrData->Opc0);
16329 SDValue Chain = Op.getOperand(0);
16330 SDValue Mask = Op.getOperand(2);
16331 SDValue Index = Op.getOperand(3);
16332 SDValue Base = Op.getOperand(4);
16333 SDValue Scale = Op.getOperand(5);
16334 return getPrefetchNode(Opcode, Op, DAG, Mask, Base, Index, Scale, Chain);
16336 // Read Time Stamp Counter (RDTSC) and Processor ID (RDTSCP).
16338 SmallVector<SDValue, 2> Results;
16339 getReadTimeStampCounter(Op.getNode(), dl, IntrData->Opc0, DAG, Subtarget,
16341 return DAG.getMergeValues(Results, dl);
16343 // Read Performance Monitoring Counters.
16345 SmallVector<SDValue, 2> Results;
16346 getReadPerformanceCounter(Op.getNode(), dl, DAG, Subtarget, Results);
16347 return DAG.getMergeValues(Results, dl);
16349 // XTEST intrinsics.
16351 SDVTList VTs = DAG.getVTList(Op->getValueType(0), MVT::Other);
16352 SDValue InTrans = DAG.getNode(IntrData->Opc0, dl, VTs, Op.getOperand(0));
16353 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
16354 DAG.getConstant(X86::COND_NE, dl, MVT::i8),
16356 SDValue Ret = DAG.getNode(ISD::ZERO_EXTEND, dl, Op->getValueType(0), SetCC);
16357 return DAG.getNode(ISD::MERGE_VALUES, dl, Op->getVTList(),
16358 Ret, SDValue(InTrans.getNode(), 1));
16362 SmallVector<SDValue, 2> Results;
16363 SDVTList CFVTs = DAG.getVTList(Op->getValueType(0), MVT::Other);
16364 SDVTList VTs = DAG.getVTList(Op.getOperand(3)->getValueType(0), MVT::Other);
16365 SDValue GenCF = DAG.getNode(X86ISD::ADD, dl, CFVTs, Op.getOperand(2),
16366 DAG.getConstant(-1, dl, MVT::i8));
16367 SDValue Res = DAG.getNode(IntrData->Opc0, dl, VTs, Op.getOperand(3),
16368 Op.getOperand(4), GenCF.getValue(1));
16369 SDValue Store = DAG.getStore(Op.getOperand(0), dl, Res.getValue(0),
16370 Op.getOperand(5), MachinePointerInfo(),
16372 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
16373 DAG.getConstant(X86::COND_B, dl, MVT::i8),
16375 Results.push_back(SetCC);
16376 Results.push_back(Store);
16377 return DAG.getMergeValues(Results, dl);
16379 case COMPRESS_TO_MEM: {
16381 SDValue Mask = Op.getOperand(4);
16382 SDValue DataToCompress = Op.getOperand(3);
16383 SDValue Addr = Op.getOperand(2);
16384 SDValue Chain = Op.getOperand(0);
16386 EVT VT = DataToCompress.getValueType();
16387 if (isAllOnes(Mask)) // return just a store
16388 return DAG.getStore(Chain, dl, DataToCompress, Addr,
16389 MachinePointerInfo(), false, false,
16390 VT.getScalarSizeInBits()/8);
16392 SDValue Compressed =
16393 getVectorMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT, DataToCompress),
16394 Mask, DAG.getUNDEF(VT), Subtarget, DAG);
16395 return DAG.getStore(Chain, dl, Compressed, Addr,
16396 MachinePointerInfo(), false, false,
16397 VT.getScalarSizeInBits()/8);
16399 case TRUNCATE_TO_MEM_VI8:
16400 return LowerINTRINSIC_TRUNCATE_TO_MEM(Op, DAG, MVT::i8);
16401 case TRUNCATE_TO_MEM_VI16:
16402 return LowerINTRINSIC_TRUNCATE_TO_MEM(Op, DAG, MVT::i16);
16403 case TRUNCATE_TO_MEM_VI32:
16404 return LowerINTRINSIC_TRUNCATE_TO_MEM(Op, DAG, MVT::i32);
16405 case EXPAND_FROM_MEM: {
16407 SDValue Mask = Op.getOperand(4);
16408 SDValue PassThru = Op.getOperand(3);
16409 SDValue Addr = Op.getOperand(2);
16410 SDValue Chain = Op.getOperand(0);
16411 EVT VT = Op.getValueType();
16413 if (isAllOnes(Mask)) // return just a load
16414 return DAG.getLoad(VT, dl, Chain, Addr, MachinePointerInfo(), false, false,
16415 false, VT.getScalarSizeInBits()/8);
16417 SDValue DataToExpand = DAG.getLoad(VT, dl, Chain, Addr, MachinePointerInfo(),
16418 false, false, false,
16419 VT.getScalarSizeInBits()/8);
16421 SDValue Results[] = {
16422 getVectorMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT, DataToExpand),
16423 Mask, PassThru, Subtarget, DAG), Chain};
16424 return DAG.getMergeValues(Results, dl);
16429 SDValue X86TargetLowering::LowerRETURNADDR(SDValue Op,
16430 SelectionDAG &DAG) const {
16431 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
16432 MFI->setReturnAddressIsTaken(true);
16434 if (verifyReturnAddressArgumentIsConstant(Op, DAG))
16437 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
16439 EVT PtrVT = getPointerTy(DAG.getDataLayout());
16442 SDValue FrameAddr = LowerFRAMEADDR(Op, DAG);
16443 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
16444 SDValue Offset = DAG.getConstant(RegInfo->getSlotSize(), dl, PtrVT);
16445 return DAG.getLoad(PtrVT, dl, DAG.getEntryNode(),
16446 DAG.getNode(ISD::ADD, dl, PtrVT,
16447 FrameAddr, Offset),
16448 MachinePointerInfo(), false, false, false, 0);
16451 // Just load the return address.
16452 SDValue RetAddrFI = getReturnAddressFrameIndex(DAG);
16453 return DAG.getLoad(PtrVT, dl, DAG.getEntryNode(),
16454 RetAddrFI, MachinePointerInfo(), false, false, false, 0);
16457 SDValue X86TargetLowering::LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const {
16458 MachineFunction &MF = DAG.getMachineFunction();
16459 MachineFrameInfo *MFI = MF.getFrameInfo();
16460 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
16461 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
16462 EVT VT = Op.getValueType();
16464 MFI->setFrameAddressIsTaken(true);
16466 if (MF.getTarget().getMCAsmInfo()->usesWindowsCFI()) {
16467 // Depth > 0 makes no sense on targets which use Windows unwind codes. It
16468 // is not possible to crawl up the stack without looking at the unwind codes
16470 int FrameAddrIndex = FuncInfo->getFAIndex();
16471 if (!FrameAddrIndex) {
16472 // Set up a frame object for the return address.
16473 unsigned SlotSize = RegInfo->getSlotSize();
16474 FrameAddrIndex = MF.getFrameInfo()->CreateFixedObject(
16475 SlotSize, /*Offset=*/0, /*IsImmutable=*/false);
16476 FuncInfo->setFAIndex(FrameAddrIndex);
16478 return DAG.getFrameIndex(FrameAddrIndex, VT);
16481 unsigned FrameReg =
16482 RegInfo->getPtrSizedFrameRegister(DAG.getMachineFunction());
16483 SDLoc dl(Op); // FIXME probably not meaningful
16484 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
16485 assert(((FrameReg == X86::RBP && VT == MVT::i64) ||
16486 (FrameReg == X86::EBP && VT == MVT::i32)) &&
16487 "Invalid Frame Register!");
16488 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, VT);
16490 FrameAddr = DAG.getLoad(VT, dl, DAG.getEntryNode(), FrameAddr,
16491 MachinePointerInfo(),
16492 false, false, false, 0);
16496 // FIXME? Maybe this could be a TableGen attribute on some registers and
16497 // this table could be generated automatically from RegInfo.
16498 unsigned X86TargetLowering::getRegisterByName(const char* RegName, EVT VT,
16499 SelectionDAG &DAG) const {
16500 const TargetFrameLowering &TFI = *Subtarget->getFrameLowering();
16501 const MachineFunction &MF = DAG.getMachineFunction();
16503 unsigned Reg = StringSwitch<unsigned>(RegName)
16504 .Case("esp", X86::ESP)
16505 .Case("rsp", X86::RSP)
16506 .Case("ebp", X86::EBP)
16507 .Case("rbp", X86::RBP)
16510 if (Reg == X86::EBP || Reg == X86::RBP) {
16511 if (!TFI.hasFP(MF))
16512 report_fatal_error("register " + StringRef(RegName) +
16513 " is allocatable: function has no frame pointer");
16516 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
16517 unsigned FrameReg =
16518 RegInfo->getPtrSizedFrameRegister(DAG.getMachineFunction());
16519 assert((FrameReg == X86::EBP || FrameReg == X86::RBP) &&
16520 "Invalid Frame Register!");
16528 report_fatal_error("Invalid register name global variable");
16531 SDValue X86TargetLowering::LowerFRAME_TO_ARGS_OFFSET(SDValue Op,
16532 SelectionDAG &DAG) const {
16533 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
16534 return DAG.getIntPtrConstant(2 * RegInfo->getSlotSize(), SDLoc(Op));
16537 SDValue X86TargetLowering::LowerEH_RETURN(SDValue Op, SelectionDAG &DAG) const {
16538 SDValue Chain = Op.getOperand(0);
16539 SDValue Offset = Op.getOperand(1);
16540 SDValue Handler = Op.getOperand(2);
16543 EVT PtrVT = getPointerTy(DAG.getDataLayout());
16544 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
16545 unsigned FrameReg = RegInfo->getFrameRegister(DAG.getMachineFunction());
16546 assert(((FrameReg == X86::RBP && PtrVT == MVT::i64) ||
16547 (FrameReg == X86::EBP && PtrVT == MVT::i32)) &&
16548 "Invalid Frame Register!");
16549 SDValue Frame = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, PtrVT);
16550 unsigned StoreAddrReg = (PtrVT == MVT::i64) ? X86::RCX : X86::ECX;
16552 SDValue StoreAddr = DAG.getNode(ISD::ADD, dl, PtrVT, Frame,
16553 DAG.getIntPtrConstant(RegInfo->getSlotSize(),
16555 StoreAddr = DAG.getNode(ISD::ADD, dl, PtrVT, StoreAddr, Offset);
16556 Chain = DAG.getStore(Chain, dl, Handler, StoreAddr, MachinePointerInfo(),
16558 Chain = DAG.getCopyToReg(Chain, dl, StoreAddrReg, StoreAddr);
16560 return DAG.getNode(X86ISD::EH_RETURN, dl, MVT::Other, Chain,
16561 DAG.getRegister(StoreAddrReg, PtrVT));
16564 SDValue X86TargetLowering::lowerEH_SJLJ_SETJMP(SDValue Op,
16565 SelectionDAG &DAG) const {
16567 return DAG.getNode(X86ISD::EH_SJLJ_SETJMP, DL,
16568 DAG.getVTList(MVT::i32, MVT::Other),
16569 Op.getOperand(0), Op.getOperand(1));
16572 SDValue X86TargetLowering::lowerEH_SJLJ_LONGJMP(SDValue Op,
16573 SelectionDAG &DAG) const {
16575 return DAG.getNode(X86ISD::EH_SJLJ_LONGJMP, DL, MVT::Other,
16576 Op.getOperand(0), Op.getOperand(1));
16579 static SDValue LowerADJUST_TRAMPOLINE(SDValue Op, SelectionDAG &DAG) {
16580 return Op.getOperand(0);
16583 SDValue X86TargetLowering::LowerINIT_TRAMPOLINE(SDValue Op,
16584 SelectionDAG &DAG) const {
16585 SDValue Root = Op.getOperand(0);
16586 SDValue Trmp = Op.getOperand(1); // trampoline
16587 SDValue FPtr = Op.getOperand(2); // nested function
16588 SDValue Nest = Op.getOperand(3); // 'nest' parameter value
16591 const Value *TrmpAddr = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
16592 const TargetRegisterInfo *TRI = Subtarget->getRegisterInfo();
16594 if (Subtarget->is64Bit()) {
16595 SDValue OutChains[6];
16597 // Large code-model.
16598 const unsigned char JMP64r = 0xFF; // 64-bit jmp through register opcode.
16599 const unsigned char MOV64ri = 0xB8; // X86::MOV64ri opcode.
16601 const unsigned char N86R10 = TRI->getEncodingValue(X86::R10) & 0x7;
16602 const unsigned char N86R11 = TRI->getEncodingValue(X86::R11) & 0x7;
16604 const unsigned char REX_WB = 0x40 | 0x08 | 0x01; // REX prefix
16606 // Load the pointer to the nested function into R11.
16607 unsigned OpCode = ((MOV64ri | N86R11) << 8) | REX_WB; // movabsq r11
16608 SDValue Addr = Trmp;
16609 OutChains[0] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, dl, MVT::i16),
16610 Addr, MachinePointerInfo(TrmpAddr),
16613 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
16614 DAG.getConstant(2, dl, MVT::i64));
16615 OutChains[1] = DAG.getStore(Root, dl, FPtr, Addr,
16616 MachinePointerInfo(TrmpAddr, 2),
16619 // Load the 'nest' parameter value into R10.
16620 // R10 is specified in X86CallingConv.td
16621 OpCode = ((MOV64ri | N86R10) << 8) | REX_WB; // movabsq r10
16622 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
16623 DAG.getConstant(10, dl, MVT::i64));
16624 OutChains[2] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, dl, MVT::i16),
16625 Addr, MachinePointerInfo(TrmpAddr, 10),
16628 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
16629 DAG.getConstant(12, dl, MVT::i64));
16630 OutChains[3] = DAG.getStore(Root, dl, Nest, Addr,
16631 MachinePointerInfo(TrmpAddr, 12),
16634 // Jump to the nested function.
16635 OpCode = (JMP64r << 8) | REX_WB; // jmpq *...
16636 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
16637 DAG.getConstant(20, dl, MVT::i64));
16638 OutChains[4] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, dl, MVT::i16),
16639 Addr, MachinePointerInfo(TrmpAddr, 20),
16642 unsigned char ModRM = N86R11 | (4 << 3) | (3 << 6); // ...r11
16643 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
16644 DAG.getConstant(22, dl, MVT::i64));
16645 OutChains[5] = DAG.getStore(Root, dl, DAG.getConstant(ModRM, dl, MVT::i8),
16646 Addr, MachinePointerInfo(TrmpAddr, 22),
16649 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other, OutChains);
16651 const Function *Func =
16652 cast<Function>(cast<SrcValueSDNode>(Op.getOperand(5))->getValue());
16653 CallingConv::ID CC = Func->getCallingConv();
16658 llvm_unreachable("Unsupported calling convention");
16659 case CallingConv::C:
16660 case CallingConv::X86_StdCall: {
16661 // Pass 'nest' parameter in ECX.
16662 // Must be kept in sync with X86CallingConv.td
16663 NestReg = X86::ECX;
16665 // Check that ECX wasn't needed by an 'inreg' parameter.
16666 FunctionType *FTy = Func->getFunctionType();
16667 const AttributeSet &Attrs = Func->getAttributes();
16669 if (!Attrs.isEmpty() && !Func->isVarArg()) {
16670 unsigned InRegCount = 0;
16673 for (FunctionType::param_iterator I = FTy->param_begin(),
16674 E = FTy->param_end(); I != E; ++I, ++Idx)
16675 if (Attrs.hasAttribute(Idx, Attribute::InReg)) {
16676 auto &DL = DAG.getDataLayout();
16677 // FIXME: should only count parameters that are lowered to integers.
16678 InRegCount += (DL.getTypeSizeInBits(*I) + 31) / 32;
16681 if (InRegCount > 2) {
16682 report_fatal_error("Nest register in use - reduce number of inreg"
16688 case CallingConv::X86_FastCall:
16689 case CallingConv::X86_ThisCall:
16690 case CallingConv::Fast:
16691 // Pass 'nest' parameter in EAX.
16692 // Must be kept in sync with X86CallingConv.td
16693 NestReg = X86::EAX;
16697 SDValue OutChains[4];
16698 SDValue Addr, Disp;
16700 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
16701 DAG.getConstant(10, dl, MVT::i32));
16702 Disp = DAG.getNode(ISD::SUB, dl, MVT::i32, FPtr, Addr);
16704 // This is storing the opcode for MOV32ri.
16705 const unsigned char MOV32ri = 0xB8; // X86::MOV32ri's opcode byte.
16706 const unsigned char N86Reg = TRI->getEncodingValue(NestReg) & 0x7;
16707 OutChains[0] = DAG.getStore(Root, dl,
16708 DAG.getConstant(MOV32ri|N86Reg, dl, MVT::i8),
16709 Trmp, MachinePointerInfo(TrmpAddr),
16712 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
16713 DAG.getConstant(1, dl, MVT::i32));
16714 OutChains[1] = DAG.getStore(Root, dl, Nest, Addr,
16715 MachinePointerInfo(TrmpAddr, 1),
16718 const unsigned char JMP = 0xE9; // jmp <32bit dst> opcode.
16719 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
16720 DAG.getConstant(5, dl, MVT::i32));
16721 OutChains[2] = DAG.getStore(Root, dl, DAG.getConstant(JMP, dl, MVT::i8),
16722 Addr, MachinePointerInfo(TrmpAddr, 5),
16725 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
16726 DAG.getConstant(6, dl, MVT::i32));
16727 OutChains[3] = DAG.getStore(Root, dl, Disp, Addr,
16728 MachinePointerInfo(TrmpAddr, 6),
16731 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other, OutChains);
16735 SDValue X86TargetLowering::LowerFLT_ROUNDS_(SDValue Op,
16736 SelectionDAG &DAG) const {
16738 The rounding mode is in bits 11:10 of FPSR, and has the following
16740 00 Round to nearest
16745 FLT_ROUNDS, on the other hand, expects the following:
16752 To perform the conversion, we do:
16753 (((((FPSR & 0x800) >> 11) | ((FPSR & 0x400) >> 9)) + 1) & 3)
16756 MachineFunction &MF = DAG.getMachineFunction();
16757 const TargetFrameLowering &TFI = *Subtarget->getFrameLowering();
16758 unsigned StackAlignment = TFI.getStackAlignment();
16759 MVT VT = Op.getSimpleValueType();
16762 // Save FP Control Word to stack slot
16763 int SSFI = MF.getFrameInfo()->CreateStackObject(2, StackAlignment, false);
16764 SDValue StackSlot =
16765 DAG.getFrameIndex(SSFI, getPointerTy(DAG.getDataLayout()));
16767 MachineMemOperand *MMO =
16768 MF.getMachineMemOperand(MachinePointerInfo::getFixedStack(MF, SSFI),
16769 MachineMemOperand::MOStore, 2, 2);
16771 SDValue Ops[] = { DAG.getEntryNode(), StackSlot };
16772 SDValue Chain = DAG.getMemIntrinsicNode(X86ISD::FNSTCW16m, DL,
16773 DAG.getVTList(MVT::Other),
16774 Ops, MVT::i16, MMO);
16776 // Load FP Control Word from stack slot
16777 SDValue CWD = DAG.getLoad(MVT::i16, DL, Chain, StackSlot,
16778 MachinePointerInfo(), false, false, false, 0);
16780 // Transform as necessary
16782 DAG.getNode(ISD::SRL, DL, MVT::i16,
16783 DAG.getNode(ISD::AND, DL, MVT::i16,
16784 CWD, DAG.getConstant(0x800, DL, MVT::i16)),
16785 DAG.getConstant(11, DL, MVT::i8));
16787 DAG.getNode(ISD::SRL, DL, MVT::i16,
16788 DAG.getNode(ISD::AND, DL, MVT::i16,
16789 CWD, DAG.getConstant(0x400, DL, MVT::i16)),
16790 DAG.getConstant(9, DL, MVT::i8));
16793 DAG.getNode(ISD::AND, DL, MVT::i16,
16794 DAG.getNode(ISD::ADD, DL, MVT::i16,
16795 DAG.getNode(ISD::OR, DL, MVT::i16, CWD1, CWD2),
16796 DAG.getConstant(1, DL, MVT::i16)),
16797 DAG.getConstant(3, DL, MVT::i16));
16799 return DAG.getNode((VT.getSizeInBits() < 16 ?
16800 ISD::TRUNCATE : ISD::ZERO_EXTEND), DL, VT, RetVal);
16803 static SDValue LowerCTLZ(SDValue Op, SelectionDAG &DAG) {
16804 MVT VT = Op.getSimpleValueType();
16806 unsigned NumBits = VT.getSizeInBits();
16809 Op = Op.getOperand(0);
16810 if (VT == MVT::i8) {
16811 // Zero extend to i32 since there is not an i8 bsr.
16813 Op = DAG.getNode(ISD::ZERO_EXTEND, dl, OpVT, Op);
16816 // Issue a bsr (scan bits in reverse) which also sets EFLAGS.
16817 SDVTList VTs = DAG.getVTList(OpVT, MVT::i32);
16818 Op = DAG.getNode(X86ISD::BSR, dl, VTs, Op);
16820 // If src is zero (i.e. bsr sets ZF), returns NumBits.
16823 DAG.getConstant(NumBits + NumBits - 1, dl, OpVT),
16824 DAG.getConstant(X86::COND_E, dl, MVT::i8),
16827 Op = DAG.getNode(X86ISD::CMOV, dl, OpVT, Ops);
16829 // Finally xor with NumBits-1.
16830 Op = DAG.getNode(ISD::XOR, dl, OpVT, Op,
16831 DAG.getConstant(NumBits - 1, dl, OpVT));
16834 Op = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op);
16838 static SDValue LowerCTLZ_ZERO_UNDEF(SDValue Op, SelectionDAG &DAG) {
16839 MVT VT = Op.getSimpleValueType();
16841 unsigned NumBits = VT.getSizeInBits();
16844 Op = Op.getOperand(0);
16845 if (VT == MVT::i8) {
16846 // Zero extend to i32 since there is not an i8 bsr.
16848 Op = DAG.getNode(ISD::ZERO_EXTEND, dl, OpVT, Op);
16851 // Issue a bsr (scan bits in reverse).
16852 SDVTList VTs = DAG.getVTList(OpVT, MVT::i32);
16853 Op = DAG.getNode(X86ISD::BSR, dl, VTs, Op);
16855 // And xor with NumBits-1.
16856 Op = DAG.getNode(ISD::XOR, dl, OpVT, Op,
16857 DAG.getConstant(NumBits - 1, dl, OpVT));
16860 Op = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op);
16864 static SDValue LowerCTTZ(SDValue Op, SelectionDAG &DAG) {
16865 MVT VT = Op.getSimpleValueType();
16866 unsigned NumBits = VT.getSizeInBits();
16868 Op = Op.getOperand(0);
16870 // Issue a bsf (scan bits forward) which also sets EFLAGS.
16871 SDVTList VTs = DAG.getVTList(VT, MVT::i32);
16872 Op = DAG.getNode(X86ISD::BSF, dl, VTs, Op);
16874 // If src is zero (i.e. bsf sets ZF), returns NumBits.
16877 DAG.getConstant(NumBits, dl, VT),
16878 DAG.getConstant(X86::COND_E, dl, MVT::i8),
16881 return DAG.getNode(X86ISD::CMOV, dl, VT, Ops);
16884 // Lower256IntArith - Break a 256-bit integer operation into two new 128-bit
16885 // ones, and then concatenate the result back.
16886 static SDValue Lower256IntArith(SDValue Op, SelectionDAG &DAG) {
16887 MVT VT = Op.getSimpleValueType();
16889 assert(VT.is256BitVector() && VT.isInteger() &&
16890 "Unsupported value type for operation");
16892 unsigned NumElems = VT.getVectorNumElements();
16895 // Extract the LHS vectors
16896 SDValue LHS = Op.getOperand(0);
16897 SDValue LHS1 = Extract128BitVector(LHS, 0, DAG, dl);
16898 SDValue LHS2 = Extract128BitVector(LHS, NumElems/2, DAG, dl);
16900 // Extract the RHS vectors
16901 SDValue RHS = Op.getOperand(1);
16902 SDValue RHS1 = Extract128BitVector(RHS, 0, DAG, dl);
16903 SDValue RHS2 = Extract128BitVector(RHS, NumElems/2, DAG, dl);
16905 MVT EltVT = VT.getVectorElementType();
16906 MVT NewVT = MVT::getVectorVT(EltVT, NumElems/2);
16908 return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT,
16909 DAG.getNode(Op.getOpcode(), dl, NewVT, LHS1, RHS1),
16910 DAG.getNode(Op.getOpcode(), dl, NewVT, LHS2, RHS2));
16913 static SDValue LowerADD(SDValue Op, SelectionDAG &DAG) {
16914 if (Op.getValueType() == MVT::i1)
16915 return DAG.getNode(ISD::XOR, SDLoc(Op), Op.getValueType(),
16916 Op.getOperand(0), Op.getOperand(1));
16917 assert(Op.getSimpleValueType().is256BitVector() &&
16918 Op.getSimpleValueType().isInteger() &&
16919 "Only handle AVX 256-bit vector integer operation");
16920 return Lower256IntArith(Op, DAG);
16923 static SDValue LowerSUB(SDValue Op, SelectionDAG &DAG) {
16924 if (Op.getValueType() == MVT::i1)
16925 return DAG.getNode(ISD::XOR, SDLoc(Op), Op.getValueType(),
16926 Op.getOperand(0), Op.getOperand(1));
16927 assert(Op.getSimpleValueType().is256BitVector() &&
16928 Op.getSimpleValueType().isInteger() &&
16929 "Only handle AVX 256-bit vector integer operation");
16930 return Lower256IntArith(Op, DAG);
16933 static SDValue LowerMINMAX(SDValue Op, SelectionDAG &DAG) {
16934 assert(Op.getSimpleValueType().is256BitVector() &&
16935 Op.getSimpleValueType().isInteger() &&
16936 "Only handle AVX 256-bit vector integer operation");
16937 return Lower256IntArith(Op, DAG);
16940 static SDValue LowerMUL(SDValue Op, const X86Subtarget *Subtarget,
16941 SelectionDAG &DAG) {
16943 MVT VT = Op.getSimpleValueType();
16946 return DAG.getNode(ISD::AND, dl, VT, Op.getOperand(0), Op.getOperand(1));
16948 // Decompose 256-bit ops into smaller 128-bit ops.
16949 if (VT.is256BitVector() && !Subtarget->hasInt256())
16950 return Lower256IntArith(Op, DAG);
16952 SDValue A = Op.getOperand(0);
16953 SDValue B = Op.getOperand(1);
16955 // Lower v16i8/v32i8 mul as promotion to v8i16/v16i16 vector
16956 // pairs, multiply and truncate.
16957 if (VT == MVT::v16i8 || VT == MVT::v32i8) {
16958 if (Subtarget->hasInt256()) {
16959 if (VT == MVT::v32i8) {
16960 MVT SubVT = MVT::getVectorVT(MVT::i8, VT.getVectorNumElements() / 2);
16961 SDValue Lo = DAG.getIntPtrConstant(0, dl);
16962 SDValue Hi = DAG.getIntPtrConstant(VT.getVectorNumElements() / 2, dl);
16963 SDValue ALo = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, SubVT, A, Lo);
16964 SDValue BLo = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, SubVT, B, Lo);
16965 SDValue AHi = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, SubVT, A, Hi);
16966 SDValue BHi = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, SubVT, B, Hi);
16967 return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT,
16968 DAG.getNode(ISD::MUL, dl, SubVT, ALo, BLo),
16969 DAG.getNode(ISD::MUL, dl, SubVT, AHi, BHi));
16972 MVT ExVT = MVT::getVectorVT(MVT::i16, VT.getVectorNumElements());
16973 return DAG.getNode(
16974 ISD::TRUNCATE, dl, VT,
16975 DAG.getNode(ISD::MUL, dl, ExVT,
16976 DAG.getNode(ISD::SIGN_EXTEND, dl, ExVT, A),
16977 DAG.getNode(ISD::SIGN_EXTEND, dl, ExVT, B)));
16980 assert(VT == MVT::v16i8 &&
16981 "Pre-AVX2 support only supports v16i8 multiplication");
16982 MVT ExVT = MVT::v8i16;
16984 // Extract the lo parts and sign extend to i16
16986 if (Subtarget->hasSSE41()) {
16987 ALo = DAG.getNode(X86ISD::VSEXT, dl, ExVT, A);
16988 BLo = DAG.getNode(X86ISD::VSEXT, dl, ExVT, B);
16990 const int ShufMask[] = {-1, 0, -1, 1, -1, 2, -1, 3,
16991 -1, 4, -1, 5, -1, 6, -1, 7};
16992 ALo = DAG.getVectorShuffle(VT, dl, A, A, ShufMask);
16993 BLo = DAG.getVectorShuffle(VT, dl, B, B, ShufMask);
16994 ALo = DAG.getBitcast(ExVT, ALo);
16995 BLo = DAG.getBitcast(ExVT, BLo);
16996 ALo = DAG.getNode(ISD::SRA, dl, ExVT, ALo, DAG.getConstant(8, dl, ExVT));
16997 BLo = DAG.getNode(ISD::SRA, dl, ExVT, BLo, DAG.getConstant(8, dl, ExVT));
17000 // Extract the hi parts and sign extend to i16
17002 if (Subtarget->hasSSE41()) {
17003 const int ShufMask[] = {8, 9, 10, 11, 12, 13, 14, 15,
17004 -1, -1, -1, -1, -1, -1, -1, -1};
17005 AHi = DAG.getVectorShuffle(VT, dl, A, A, ShufMask);
17006 BHi = DAG.getVectorShuffle(VT, dl, B, B, ShufMask);
17007 AHi = DAG.getNode(X86ISD::VSEXT, dl, ExVT, AHi);
17008 BHi = DAG.getNode(X86ISD::VSEXT, dl, ExVT, BHi);
17010 const int ShufMask[] = {-1, 8, -1, 9, -1, 10, -1, 11,
17011 -1, 12, -1, 13, -1, 14, -1, 15};
17012 AHi = DAG.getVectorShuffle(VT, dl, A, A, ShufMask);
17013 BHi = DAG.getVectorShuffle(VT, dl, B, B, ShufMask);
17014 AHi = DAG.getBitcast(ExVT, AHi);
17015 BHi = DAG.getBitcast(ExVT, BHi);
17016 AHi = DAG.getNode(ISD::SRA, dl, ExVT, AHi, DAG.getConstant(8, dl, ExVT));
17017 BHi = DAG.getNode(ISD::SRA, dl, ExVT, BHi, DAG.getConstant(8, dl, ExVT));
17020 // Multiply, mask the lower 8bits of the lo/hi results and pack
17021 SDValue RLo = DAG.getNode(ISD::MUL, dl, ExVT, ALo, BLo);
17022 SDValue RHi = DAG.getNode(ISD::MUL, dl, ExVT, AHi, BHi);
17023 RLo = DAG.getNode(ISD::AND, dl, ExVT, RLo, DAG.getConstant(255, dl, ExVT));
17024 RHi = DAG.getNode(ISD::AND, dl, ExVT, RHi, DAG.getConstant(255, dl, ExVT));
17025 return DAG.getNode(X86ISD::PACKUS, dl, VT, RLo, RHi);
17028 // Lower v4i32 mul as 2x shuffle, 2x pmuludq, 2x shuffle.
17029 if (VT == MVT::v4i32) {
17030 assert(Subtarget->hasSSE2() && !Subtarget->hasSSE41() &&
17031 "Should not custom lower when pmuldq is available!");
17033 // Extract the odd parts.
17034 static const int UnpackMask[] = { 1, -1, 3, -1 };
17035 SDValue Aodds = DAG.getVectorShuffle(VT, dl, A, A, UnpackMask);
17036 SDValue Bodds = DAG.getVectorShuffle(VT, dl, B, B, UnpackMask);
17038 // Multiply the even parts.
17039 SDValue Evens = DAG.getNode(X86ISD::PMULUDQ, dl, MVT::v2i64, A, B);
17040 // Now multiply odd parts.
17041 SDValue Odds = DAG.getNode(X86ISD::PMULUDQ, dl, MVT::v2i64, Aodds, Bodds);
17043 Evens = DAG.getBitcast(VT, Evens);
17044 Odds = DAG.getBitcast(VT, Odds);
17046 // Merge the two vectors back together with a shuffle. This expands into 2
17048 static const int ShufMask[] = { 0, 4, 2, 6 };
17049 return DAG.getVectorShuffle(VT, dl, Evens, Odds, ShufMask);
17052 assert((VT == MVT::v2i64 || VT == MVT::v4i64 || VT == MVT::v8i64) &&
17053 "Only know how to lower V2I64/V4I64/V8I64 multiply");
17055 // Ahi = psrlqi(a, 32);
17056 // Bhi = psrlqi(b, 32);
17058 // AloBlo = pmuludq(a, b);
17059 // AloBhi = pmuludq(a, Bhi);
17060 // AhiBlo = pmuludq(Ahi, b);
17062 // AloBhi = psllqi(AloBhi, 32);
17063 // AhiBlo = psllqi(AhiBlo, 32);
17064 // return AloBlo + AloBhi + AhiBlo;
17066 SDValue Ahi = getTargetVShiftByConstNode(X86ISD::VSRLI, dl, VT, A, 32, DAG);
17067 SDValue Bhi = getTargetVShiftByConstNode(X86ISD::VSRLI, dl, VT, B, 32, DAG);
17069 SDValue AhiBlo = Ahi;
17070 SDValue AloBhi = Bhi;
17071 // Bit cast to 32-bit vectors for MULUDQ
17072 EVT MulVT = (VT == MVT::v2i64) ? MVT::v4i32 :
17073 (VT == MVT::v4i64) ? MVT::v8i32 : MVT::v16i32;
17074 A = DAG.getBitcast(MulVT, A);
17075 B = DAG.getBitcast(MulVT, B);
17076 Ahi = DAG.getBitcast(MulVT, Ahi);
17077 Bhi = DAG.getBitcast(MulVT, Bhi);
17079 SDValue AloBlo = DAG.getNode(X86ISD::PMULUDQ, dl, VT, A, B);
17080 // After shifting right const values the result may be all-zero.
17081 if (!ISD::isBuildVectorAllZeros(Ahi.getNode())) {
17082 AhiBlo = DAG.getNode(X86ISD::PMULUDQ, dl, VT, Ahi, B);
17083 AhiBlo = getTargetVShiftByConstNode(X86ISD::VSHLI, dl, VT, AhiBlo, 32, DAG);
17085 if (!ISD::isBuildVectorAllZeros(Bhi.getNode())) {
17086 AloBhi = DAG.getNode(X86ISD::PMULUDQ, dl, VT, A, Bhi);
17087 AloBhi = getTargetVShiftByConstNode(X86ISD::VSHLI, dl, VT, AloBhi, 32, DAG);
17090 SDValue Res = DAG.getNode(ISD::ADD, dl, VT, AloBlo, AloBhi);
17091 return DAG.getNode(ISD::ADD, dl, VT, Res, AhiBlo);
17094 SDValue X86TargetLowering::LowerWin64_i128OP(SDValue Op, SelectionDAG &DAG) const {
17095 assert(Subtarget->isTargetWin64() && "Unexpected target");
17096 EVT VT = Op.getValueType();
17097 assert(VT.isInteger() && VT.getSizeInBits() == 128 &&
17098 "Unexpected return type for lowering");
17102 switch (Op->getOpcode()) {
17103 default: llvm_unreachable("Unexpected request for libcall!");
17104 case ISD::SDIV: isSigned = true; LC = RTLIB::SDIV_I128; break;
17105 case ISD::UDIV: isSigned = false; LC = RTLIB::UDIV_I128; break;
17106 case ISD::SREM: isSigned = true; LC = RTLIB::SREM_I128; break;
17107 case ISD::UREM: isSigned = false; LC = RTLIB::UREM_I128; break;
17108 case ISD::SDIVREM: isSigned = true; LC = RTLIB::SDIVREM_I128; break;
17109 case ISD::UDIVREM: isSigned = false; LC = RTLIB::UDIVREM_I128; break;
17113 SDValue InChain = DAG.getEntryNode();
17115 TargetLowering::ArgListTy Args;
17116 TargetLowering::ArgListEntry Entry;
17117 for (unsigned i = 0, e = Op->getNumOperands(); i != e; ++i) {
17118 EVT ArgVT = Op->getOperand(i).getValueType();
17119 assert(ArgVT.isInteger() && ArgVT.getSizeInBits() == 128 &&
17120 "Unexpected argument type for lowering");
17121 SDValue StackPtr = DAG.CreateStackTemporary(ArgVT, 16);
17122 Entry.Node = StackPtr;
17123 InChain = DAG.getStore(InChain, dl, Op->getOperand(i), StackPtr, MachinePointerInfo(),
17125 Type *ArgTy = ArgVT.getTypeForEVT(*DAG.getContext());
17126 Entry.Ty = PointerType::get(ArgTy,0);
17127 Entry.isSExt = false;
17128 Entry.isZExt = false;
17129 Args.push_back(Entry);
17132 SDValue Callee = DAG.getExternalSymbol(getLibcallName(LC),
17133 getPointerTy(DAG.getDataLayout()));
17135 TargetLowering::CallLoweringInfo CLI(DAG);
17136 CLI.setDebugLoc(dl).setChain(InChain)
17137 .setCallee(getLibcallCallingConv(LC),
17138 static_cast<EVT>(MVT::v2i64).getTypeForEVT(*DAG.getContext()),
17139 Callee, std::move(Args), 0)
17140 .setInRegister().setSExtResult(isSigned).setZExtResult(!isSigned);
17142 std::pair<SDValue, SDValue> CallInfo = LowerCallTo(CLI);
17143 return DAG.getBitcast(VT, CallInfo.first);
17146 static SDValue LowerMUL_LOHI(SDValue Op, const X86Subtarget *Subtarget,
17147 SelectionDAG &DAG) {
17148 SDValue Op0 = Op.getOperand(0), Op1 = Op.getOperand(1);
17149 EVT VT = Op0.getValueType();
17152 assert((VT == MVT::v4i32 && Subtarget->hasSSE2()) ||
17153 (VT == MVT::v8i32 && Subtarget->hasInt256()));
17155 // PMULxD operations multiply each even value (starting at 0) of LHS with
17156 // the related value of RHS and produce a widen result.
17157 // E.g., PMULUDQ <4 x i32> <a|b|c|d>, <4 x i32> <e|f|g|h>
17158 // => <2 x i64> <ae|cg>
17160 // In other word, to have all the results, we need to perform two PMULxD:
17161 // 1. one with the even values.
17162 // 2. one with the odd values.
17163 // To achieve #2, with need to place the odd values at an even position.
17165 // Place the odd value at an even position (basically, shift all values 1
17166 // step to the left):
17167 const int Mask[] = {1, -1, 3, -1, 5, -1, 7, -1};
17168 // <a|b|c|d> => <b|undef|d|undef>
17169 SDValue Odd0 = DAG.getVectorShuffle(VT, dl, Op0, Op0, Mask);
17170 // <e|f|g|h> => <f|undef|h|undef>
17171 SDValue Odd1 = DAG.getVectorShuffle(VT, dl, Op1, Op1, Mask);
17173 // Emit two multiplies, one for the lower 2 ints and one for the higher 2
17175 MVT MulVT = VT == MVT::v4i32 ? MVT::v2i64 : MVT::v4i64;
17176 bool IsSigned = Op->getOpcode() == ISD::SMUL_LOHI;
17178 (!IsSigned || !Subtarget->hasSSE41()) ? X86ISD::PMULUDQ : X86ISD::PMULDQ;
17179 // PMULUDQ <4 x i32> <a|b|c|d>, <4 x i32> <e|f|g|h>
17180 // => <2 x i64> <ae|cg>
17181 SDValue Mul1 = DAG.getBitcast(VT, DAG.getNode(Opcode, dl, MulVT, Op0, Op1));
17182 // PMULUDQ <4 x i32> <b|undef|d|undef>, <4 x i32> <f|undef|h|undef>
17183 // => <2 x i64> <bf|dh>
17184 SDValue Mul2 = DAG.getBitcast(VT, DAG.getNode(Opcode, dl, MulVT, Odd0, Odd1));
17186 // Shuffle it back into the right order.
17187 SDValue Highs, Lows;
17188 if (VT == MVT::v8i32) {
17189 const int HighMask[] = {1, 9, 3, 11, 5, 13, 7, 15};
17190 Highs = DAG.getVectorShuffle(VT, dl, Mul1, Mul2, HighMask);
17191 const int LowMask[] = {0, 8, 2, 10, 4, 12, 6, 14};
17192 Lows = DAG.getVectorShuffle(VT, dl, Mul1, Mul2, LowMask);
17194 const int HighMask[] = {1, 5, 3, 7};
17195 Highs = DAG.getVectorShuffle(VT, dl, Mul1, Mul2, HighMask);
17196 const int LowMask[] = {0, 4, 2, 6};
17197 Lows = DAG.getVectorShuffle(VT, dl, Mul1, Mul2, LowMask);
17200 // If we have a signed multiply but no PMULDQ fix up the high parts of a
17201 // unsigned multiply.
17202 if (IsSigned && !Subtarget->hasSSE41()) {
17203 SDValue ShAmt = DAG.getConstant(
17205 DAG.getTargetLoweringInfo().getShiftAmountTy(VT, DAG.getDataLayout()));
17206 SDValue T1 = DAG.getNode(ISD::AND, dl, VT,
17207 DAG.getNode(ISD::SRA, dl, VT, Op0, ShAmt), Op1);
17208 SDValue T2 = DAG.getNode(ISD::AND, dl, VT,
17209 DAG.getNode(ISD::SRA, dl, VT, Op1, ShAmt), Op0);
17211 SDValue Fixup = DAG.getNode(ISD::ADD, dl, VT, T1, T2);
17212 Highs = DAG.getNode(ISD::SUB, dl, VT, Highs, Fixup);
17215 // The first result of MUL_LOHI is actually the low value, followed by the
17217 SDValue Ops[] = {Lows, Highs};
17218 return DAG.getMergeValues(Ops, dl);
17221 // Return true if the required (according to Opcode) shift-imm form is natively
17222 // supported by the Subtarget
17223 static bool SupportedVectorShiftWithImm(MVT VT, const X86Subtarget *Subtarget,
17225 if (VT.getScalarSizeInBits() < 16)
17228 if (VT.is512BitVector() &&
17229 (VT.getScalarSizeInBits() > 16 || Subtarget->hasBWI()))
17232 bool LShift = VT.is128BitVector() ||
17233 (VT.is256BitVector() && Subtarget->hasInt256());
17235 bool AShift = LShift && (Subtarget->hasVLX() ||
17236 (VT != MVT::v2i64 && VT != MVT::v4i64));
17237 return (Opcode == ISD::SRA) ? AShift : LShift;
17240 // The shift amount is a variable, but it is the same for all vector lanes.
17241 // These instructions are defined together with shift-immediate.
17243 bool SupportedVectorShiftWithBaseAmnt(MVT VT, const X86Subtarget *Subtarget,
17245 return SupportedVectorShiftWithImm(VT, Subtarget, Opcode);
17248 // Return true if the required (according to Opcode) variable-shift form is
17249 // natively supported by the Subtarget
17250 static bool SupportedVectorVarShift(MVT VT, const X86Subtarget *Subtarget,
17253 if (!Subtarget->hasInt256() || VT.getScalarSizeInBits() < 16)
17256 // vXi16 supported only on AVX-512, BWI
17257 if (VT.getScalarSizeInBits() == 16 && !Subtarget->hasBWI())
17260 if (VT.is512BitVector() || Subtarget->hasVLX())
17263 bool LShift = VT.is128BitVector() || VT.is256BitVector();
17264 bool AShift = LShift && VT != MVT::v2i64 && VT != MVT::v4i64;
17265 return (Opcode == ISD::SRA) ? AShift : LShift;
17268 static SDValue LowerScalarImmediateShift(SDValue Op, SelectionDAG &DAG,
17269 const X86Subtarget *Subtarget) {
17270 MVT VT = Op.getSimpleValueType();
17272 SDValue R = Op.getOperand(0);
17273 SDValue Amt = Op.getOperand(1);
17275 unsigned X86Opc = (Op.getOpcode() == ISD::SHL) ? X86ISD::VSHLI :
17276 (Op.getOpcode() == ISD::SRL) ? X86ISD::VSRLI : X86ISD::VSRAI;
17278 auto ArithmeticShiftRight64 = [&](uint64_t ShiftAmt) {
17279 assert((VT == MVT::v2i64 || VT == MVT::v4i64) && "Unexpected SRA type");
17280 MVT ExVT = MVT::getVectorVT(MVT::i32, VT.getVectorNumElements() * 2);
17281 SDValue Ex = DAG.getBitcast(ExVT, R);
17283 if (ShiftAmt >= 32) {
17284 // Splat sign to upper i32 dst, and SRA upper i32 src to lower i32.
17286 getTargetVShiftByConstNode(X86ISD::VSRAI, dl, ExVT, Ex, 31, DAG);
17287 SDValue Lower = getTargetVShiftByConstNode(X86ISD::VSRAI, dl, ExVT, Ex,
17288 ShiftAmt - 32, DAG);
17289 if (VT == MVT::v2i64)
17290 Ex = DAG.getVectorShuffle(ExVT, dl, Upper, Lower, {5, 1, 7, 3});
17291 if (VT == MVT::v4i64)
17292 Ex = DAG.getVectorShuffle(ExVT, dl, Upper, Lower,
17293 {9, 1, 11, 3, 13, 5, 15, 7});
17295 // SRA upper i32, SHL whole i64 and select lower i32.
17296 SDValue Upper = getTargetVShiftByConstNode(X86ISD::VSRAI, dl, ExVT, Ex,
17299 getTargetVShiftByConstNode(X86ISD::VSRLI, dl, VT, R, ShiftAmt, DAG);
17300 Lower = DAG.getBitcast(ExVT, Lower);
17301 if (VT == MVT::v2i64)
17302 Ex = DAG.getVectorShuffle(ExVT, dl, Upper, Lower, {4, 1, 6, 3});
17303 if (VT == MVT::v4i64)
17304 Ex = DAG.getVectorShuffle(ExVT, dl, Upper, Lower,
17305 {8, 1, 10, 3, 12, 5, 14, 7});
17307 return DAG.getBitcast(VT, Ex);
17310 // Optimize shl/srl/sra with constant shift amount.
17311 if (auto *BVAmt = dyn_cast<BuildVectorSDNode>(Amt)) {
17312 if (auto *ShiftConst = BVAmt->getConstantSplatNode()) {
17313 uint64_t ShiftAmt = ShiftConst->getZExtValue();
17315 if (SupportedVectorShiftWithImm(VT, Subtarget, Op.getOpcode()))
17316 return getTargetVShiftByConstNode(X86Opc, dl, VT, R, ShiftAmt, DAG);
17318 // i64 SRA needs to be performed as partial shifts.
17319 if ((VT == MVT::v2i64 || (Subtarget->hasInt256() && VT == MVT::v4i64)) &&
17320 Op.getOpcode() == ISD::SRA)
17321 return ArithmeticShiftRight64(ShiftAmt);
17323 if (VT == MVT::v16i8 || (Subtarget->hasInt256() && VT == MVT::v32i8)) {
17324 unsigned NumElts = VT.getVectorNumElements();
17325 MVT ShiftVT = MVT::getVectorVT(MVT::i16, NumElts / 2);
17327 if (Op.getOpcode() == ISD::SHL) {
17328 // Simple i8 add case
17330 return DAG.getNode(ISD::ADD, dl, VT, R, R);
17332 // Make a large shift.
17333 SDValue SHL = getTargetVShiftByConstNode(X86ISD::VSHLI, dl, ShiftVT,
17335 SHL = DAG.getBitcast(VT, SHL);
17336 // Zero out the rightmost bits.
17337 SmallVector<SDValue, 32> V(
17338 NumElts, DAG.getConstant(uint8_t(-1U << ShiftAmt), dl, MVT::i8));
17339 return DAG.getNode(ISD::AND, dl, VT, SHL,
17340 DAG.getNode(ISD::BUILD_VECTOR, dl, VT, V));
17342 if (Op.getOpcode() == ISD::SRL) {
17343 // Make a large shift.
17344 SDValue SRL = getTargetVShiftByConstNode(X86ISD::VSRLI, dl, ShiftVT,
17346 SRL = DAG.getBitcast(VT, SRL);
17347 // Zero out the leftmost bits.
17348 SmallVector<SDValue, 32> V(
17349 NumElts, DAG.getConstant(uint8_t(-1U) >> ShiftAmt, dl, MVT::i8));
17350 return DAG.getNode(ISD::AND, dl, VT, SRL,
17351 DAG.getNode(ISD::BUILD_VECTOR, dl, VT, V));
17353 if (Op.getOpcode() == ISD::SRA) {
17354 if (ShiftAmt == 7) {
17355 // ashr(R, 7) === cmp_slt(R, 0)
17356 SDValue Zeros = getZeroVector(VT, Subtarget, DAG, dl);
17357 return DAG.getNode(X86ISD::PCMPGT, dl, VT, Zeros, R);
17360 // ashr(R, Amt) === sub(xor(lshr(R, Amt), Mask), Mask)
17361 SDValue Res = DAG.getNode(ISD::SRL, dl, VT, R, Amt);
17362 SmallVector<SDValue, 32> V(NumElts,
17363 DAG.getConstant(128 >> ShiftAmt, dl,
17365 SDValue Mask = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, V);
17366 Res = DAG.getNode(ISD::XOR, dl, VT, Res, Mask);
17367 Res = DAG.getNode(ISD::SUB, dl, VT, Res, Mask);
17370 llvm_unreachable("Unknown shift opcode.");
17375 // Special case in 32-bit mode, where i64 is expanded into high and low parts.
17376 if (!Subtarget->is64Bit() &&
17377 (VT == MVT::v2i64 || (Subtarget->hasInt256() && VT == MVT::v4i64))) {
17379 // Peek through any splat that was introduced for i64 shift vectorization.
17380 int SplatIndex = -1;
17381 if (ShuffleVectorSDNode *SVN = dyn_cast<ShuffleVectorSDNode>(Amt.getNode()))
17382 if (SVN->isSplat()) {
17383 SplatIndex = SVN->getSplatIndex();
17384 Amt = Amt.getOperand(0);
17385 assert(SplatIndex < (int)VT.getVectorNumElements() &&
17386 "Splat shuffle referencing second operand");
17389 if (Amt.getOpcode() != ISD::BITCAST ||
17390 Amt.getOperand(0).getOpcode() != ISD::BUILD_VECTOR)
17393 Amt = Amt.getOperand(0);
17394 unsigned Ratio = Amt.getSimpleValueType().getVectorNumElements() /
17395 VT.getVectorNumElements();
17396 unsigned RatioInLog2 = Log2_32_Ceil(Ratio);
17397 uint64_t ShiftAmt = 0;
17398 unsigned BaseOp = (SplatIndex < 0 ? 0 : SplatIndex * Ratio);
17399 for (unsigned i = 0; i != Ratio; ++i) {
17400 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Amt.getOperand(i + BaseOp));
17404 ShiftAmt |= C->getZExtValue() << (i * (1 << (6 - RatioInLog2)));
17407 // Check remaining shift amounts (if not a splat).
17408 if (SplatIndex < 0) {
17409 for (unsigned i = Ratio; i != Amt.getNumOperands(); i += Ratio) {
17410 uint64_t ShAmt = 0;
17411 for (unsigned j = 0; j != Ratio; ++j) {
17412 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Amt.getOperand(i + j));
17416 ShAmt |= C->getZExtValue() << (j * (1 << (6 - RatioInLog2)));
17418 if (ShAmt != ShiftAmt)
17423 if (SupportedVectorShiftWithImm(VT, Subtarget, Op.getOpcode()))
17424 return getTargetVShiftByConstNode(X86Opc, dl, VT, R, ShiftAmt, DAG);
17426 if (Op.getOpcode() == ISD::SRA)
17427 return ArithmeticShiftRight64(ShiftAmt);
17433 static SDValue LowerScalarVariableShift(SDValue Op, SelectionDAG &DAG,
17434 const X86Subtarget* Subtarget) {
17435 MVT VT = Op.getSimpleValueType();
17437 SDValue R = Op.getOperand(0);
17438 SDValue Amt = Op.getOperand(1);
17440 unsigned X86OpcI = (Op.getOpcode() == ISD::SHL) ? X86ISD::VSHLI :
17441 (Op.getOpcode() == ISD::SRL) ? X86ISD::VSRLI : X86ISD::VSRAI;
17443 unsigned X86OpcV = (Op.getOpcode() == ISD::SHL) ? X86ISD::VSHL :
17444 (Op.getOpcode() == ISD::SRL) ? X86ISD::VSRL : X86ISD::VSRA;
17446 if (SupportedVectorShiftWithBaseAmnt(VT, Subtarget, Op.getOpcode())) {
17448 EVT EltVT = VT.getVectorElementType();
17450 if (BuildVectorSDNode *BV = dyn_cast<BuildVectorSDNode>(Amt)) {
17451 // Check if this build_vector node is doing a splat.
17452 // If so, then set BaseShAmt equal to the splat value.
17453 BaseShAmt = BV->getSplatValue();
17454 if (BaseShAmt && BaseShAmt.getOpcode() == ISD::UNDEF)
17455 BaseShAmt = SDValue();
17457 if (Amt.getOpcode() == ISD::EXTRACT_SUBVECTOR)
17458 Amt = Amt.getOperand(0);
17460 ShuffleVectorSDNode *SVN = dyn_cast<ShuffleVectorSDNode>(Amt);
17461 if (SVN && SVN->isSplat()) {
17462 unsigned SplatIdx = (unsigned)SVN->getSplatIndex();
17463 SDValue InVec = Amt.getOperand(0);
17464 if (InVec.getOpcode() == ISD::BUILD_VECTOR) {
17465 assert((SplatIdx < InVec.getValueType().getVectorNumElements()) &&
17466 "Unexpected shuffle index found!");
17467 BaseShAmt = InVec.getOperand(SplatIdx);
17468 } else if (InVec.getOpcode() == ISD::INSERT_VECTOR_ELT) {
17469 if (ConstantSDNode *C =
17470 dyn_cast<ConstantSDNode>(InVec.getOperand(2))) {
17471 if (C->getZExtValue() == SplatIdx)
17472 BaseShAmt = InVec.getOperand(1);
17477 // Avoid introducing an extract element from a shuffle.
17478 BaseShAmt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, EltVT, InVec,
17479 DAG.getIntPtrConstant(SplatIdx, dl));
17483 if (BaseShAmt.getNode()) {
17484 assert(EltVT.bitsLE(MVT::i64) && "Unexpected element type!");
17485 if (EltVT != MVT::i64 && EltVT.bitsGT(MVT::i32))
17486 BaseShAmt = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i64, BaseShAmt);
17487 else if (EltVT.bitsLT(MVT::i32))
17488 BaseShAmt = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, BaseShAmt);
17490 return getTargetVShiftNode(X86OpcI, dl, VT, R, BaseShAmt, DAG);
17494 // Special case in 32-bit mode, where i64 is expanded into high and low parts.
17495 if (!Subtarget->is64Bit() && VT == MVT::v2i64 &&
17496 Amt.getOpcode() == ISD::BITCAST &&
17497 Amt.getOperand(0).getOpcode() == ISD::BUILD_VECTOR) {
17498 Amt = Amt.getOperand(0);
17499 unsigned Ratio = Amt.getSimpleValueType().getVectorNumElements() /
17500 VT.getVectorNumElements();
17501 std::vector<SDValue> Vals(Ratio);
17502 for (unsigned i = 0; i != Ratio; ++i)
17503 Vals[i] = Amt.getOperand(i);
17504 for (unsigned i = Ratio; i != Amt.getNumOperands(); i += Ratio) {
17505 for (unsigned j = 0; j != Ratio; ++j)
17506 if (Vals[j] != Amt.getOperand(i + j))
17510 if (SupportedVectorShiftWithBaseAmnt(VT, Subtarget, Op.getOpcode()))
17511 return DAG.getNode(X86OpcV, dl, VT, R, Op.getOperand(1));
17516 static SDValue LowerShift(SDValue Op, const X86Subtarget* Subtarget,
17517 SelectionDAG &DAG) {
17518 MVT VT = Op.getSimpleValueType();
17520 SDValue R = Op.getOperand(0);
17521 SDValue Amt = Op.getOperand(1);
17523 assert(VT.isVector() && "Custom lowering only for vector shifts!");
17524 assert(Subtarget->hasSSE2() && "Only custom lower when we have SSE2!");
17526 if (SDValue V = LowerScalarImmediateShift(Op, DAG, Subtarget))
17529 if (SDValue V = LowerScalarVariableShift(Op, DAG, Subtarget))
17532 if (SupportedVectorVarShift(VT, Subtarget, Op.getOpcode()))
17535 // 2i64 vector logical shifts can efficiently avoid scalarization - do the
17536 // shifts per-lane and then shuffle the partial results back together.
17537 if (VT == MVT::v2i64 && Op.getOpcode() != ISD::SRA) {
17538 // Splat the shift amounts so the scalar shifts above will catch it.
17539 SDValue Amt0 = DAG.getVectorShuffle(VT, dl, Amt, Amt, {0, 0});
17540 SDValue Amt1 = DAG.getVectorShuffle(VT, dl, Amt, Amt, {1, 1});
17541 SDValue R0 = DAG.getNode(Op->getOpcode(), dl, VT, R, Amt0);
17542 SDValue R1 = DAG.getNode(Op->getOpcode(), dl, VT, R, Amt1);
17543 return DAG.getVectorShuffle(VT, dl, R0, R1, {0, 3});
17546 // i64 vector arithmetic shift can be emulated with the transform:
17547 // M = lshr(SIGN_BIT, Amt)
17548 // ashr(R, Amt) === sub(xor(lshr(R, Amt), M), M)
17549 if ((VT == MVT::v2i64 || (VT == MVT::v4i64 && Subtarget->hasInt256())) &&
17550 Op.getOpcode() == ISD::SRA) {
17551 SDValue S = DAG.getConstant(APInt::getSignBit(64), dl, VT);
17552 SDValue M = DAG.getNode(ISD::SRL, dl, VT, S, Amt);
17553 R = DAG.getNode(ISD::SRL, dl, VT, R, Amt);
17554 R = DAG.getNode(ISD::XOR, dl, VT, R, M);
17555 R = DAG.getNode(ISD::SUB, dl, VT, R, M);
17559 // If possible, lower this packed shift into a vector multiply instead of
17560 // expanding it into a sequence of scalar shifts.
17561 // Do this only if the vector shift count is a constant build_vector.
17562 if (Op.getOpcode() == ISD::SHL &&
17563 (VT == MVT::v8i16 || VT == MVT::v4i32 ||
17564 (Subtarget->hasInt256() && VT == MVT::v16i16)) &&
17565 ISD::isBuildVectorOfConstantSDNodes(Amt.getNode())) {
17566 SmallVector<SDValue, 8> Elts;
17567 EVT SVT = VT.getScalarType();
17568 unsigned SVTBits = SVT.getSizeInBits();
17569 const APInt &One = APInt(SVTBits, 1);
17570 unsigned NumElems = VT.getVectorNumElements();
17572 for (unsigned i=0; i !=NumElems; ++i) {
17573 SDValue Op = Amt->getOperand(i);
17574 if (Op->getOpcode() == ISD::UNDEF) {
17575 Elts.push_back(Op);
17579 ConstantSDNode *ND = cast<ConstantSDNode>(Op);
17580 const APInt &C = APInt(SVTBits, ND->getAPIntValue().getZExtValue());
17581 uint64_t ShAmt = C.getZExtValue();
17582 if (ShAmt >= SVTBits) {
17583 Elts.push_back(DAG.getUNDEF(SVT));
17586 Elts.push_back(DAG.getConstant(One.shl(ShAmt), dl, SVT));
17588 SDValue BV = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Elts);
17589 return DAG.getNode(ISD::MUL, dl, VT, R, BV);
17592 // Lower SHL with variable shift amount.
17593 if (VT == MVT::v4i32 && Op->getOpcode() == ISD::SHL) {
17594 Op = DAG.getNode(ISD::SHL, dl, VT, Amt, DAG.getConstant(23, dl, VT));
17596 Op = DAG.getNode(ISD::ADD, dl, VT, Op,
17597 DAG.getConstant(0x3f800000U, dl, VT));
17598 Op = DAG.getBitcast(MVT::v4f32, Op);
17599 Op = DAG.getNode(ISD::FP_TO_SINT, dl, VT, Op);
17600 return DAG.getNode(ISD::MUL, dl, VT, Op, R);
17603 // If possible, lower this shift as a sequence of two shifts by
17604 // constant plus a MOVSS/MOVSD instead of scalarizing it.
17606 // (v4i32 (srl A, (build_vector < X, Y, Y, Y>)))
17608 // Could be rewritten as:
17609 // (v4i32 (MOVSS (srl A, <Y,Y,Y,Y>), (srl A, <X,X,X,X>)))
17611 // The advantage is that the two shifts from the example would be
17612 // lowered as X86ISD::VSRLI nodes. This would be cheaper than scalarizing
17613 // the vector shift into four scalar shifts plus four pairs of vector
17615 if ((VT == MVT::v8i16 || VT == MVT::v4i32) &&
17616 ISD::isBuildVectorOfConstantSDNodes(Amt.getNode())) {
17617 unsigned TargetOpcode = X86ISD::MOVSS;
17618 bool CanBeSimplified;
17619 // The splat value for the first packed shift (the 'X' from the example).
17620 SDValue Amt1 = Amt->getOperand(0);
17621 // The splat value for the second packed shift (the 'Y' from the example).
17622 SDValue Amt2 = (VT == MVT::v4i32) ? Amt->getOperand(1) :
17623 Amt->getOperand(2);
17625 // See if it is possible to replace this node with a sequence of
17626 // two shifts followed by a MOVSS/MOVSD
17627 if (VT == MVT::v4i32) {
17628 // Check if it is legal to use a MOVSS.
17629 CanBeSimplified = Amt2 == Amt->getOperand(2) &&
17630 Amt2 == Amt->getOperand(3);
17631 if (!CanBeSimplified) {
17632 // Otherwise, check if we can still simplify this node using a MOVSD.
17633 CanBeSimplified = Amt1 == Amt->getOperand(1) &&
17634 Amt->getOperand(2) == Amt->getOperand(3);
17635 TargetOpcode = X86ISD::MOVSD;
17636 Amt2 = Amt->getOperand(2);
17639 // Do similar checks for the case where the machine value type
17641 CanBeSimplified = Amt1 == Amt->getOperand(1);
17642 for (unsigned i=3; i != 8 && CanBeSimplified; ++i)
17643 CanBeSimplified = Amt2 == Amt->getOperand(i);
17645 if (!CanBeSimplified) {
17646 TargetOpcode = X86ISD::MOVSD;
17647 CanBeSimplified = true;
17648 Amt2 = Amt->getOperand(4);
17649 for (unsigned i=0; i != 4 && CanBeSimplified; ++i)
17650 CanBeSimplified = Amt1 == Amt->getOperand(i);
17651 for (unsigned j=4; j != 8 && CanBeSimplified; ++j)
17652 CanBeSimplified = Amt2 == Amt->getOperand(j);
17656 if (CanBeSimplified && isa<ConstantSDNode>(Amt1) &&
17657 isa<ConstantSDNode>(Amt2)) {
17658 // Replace this node with two shifts followed by a MOVSS/MOVSD.
17659 EVT CastVT = MVT::v4i32;
17661 DAG.getConstant(cast<ConstantSDNode>(Amt1)->getAPIntValue(), dl, VT);
17662 SDValue Shift1 = DAG.getNode(Op->getOpcode(), dl, VT, R, Splat1);
17664 DAG.getConstant(cast<ConstantSDNode>(Amt2)->getAPIntValue(), dl, VT);
17665 SDValue Shift2 = DAG.getNode(Op->getOpcode(), dl, VT, R, Splat2);
17666 if (TargetOpcode == X86ISD::MOVSD)
17667 CastVT = MVT::v2i64;
17668 SDValue BitCast1 = DAG.getBitcast(CastVT, Shift1);
17669 SDValue BitCast2 = DAG.getBitcast(CastVT, Shift2);
17670 SDValue Result = getTargetShuffleNode(TargetOpcode, dl, CastVT, BitCast2,
17672 return DAG.getBitcast(VT, Result);
17676 // v4i32 Non Uniform Shifts.
17677 // If the shift amount is constant we can shift each lane using the SSE2
17678 // immediate shifts, else we need to zero-extend each lane to the lower i64
17679 // and shift using the SSE2 variable shifts.
17680 // The separate results can then be blended together.
17681 if (VT == MVT::v4i32) {
17682 unsigned Opc = Op.getOpcode();
17683 SDValue Amt0, Amt1, Amt2, Amt3;
17684 if (ISD::isBuildVectorOfConstantSDNodes(Amt.getNode())) {
17685 Amt0 = DAG.getVectorShuffle(VT, dl, Amt, DAG.getUNDEF(VT), {0, 0, 0, 0});
17686 Amt1 = DAG.getVectorShuffle(VT, dl, Amt, DAG.getUNDEF(VT), {1, 1, 1, 1});
17687 Amt2 = DAG.getVectorShuffle(VT, dl, Amt, DAG.getUNDEF(VT), {2, 2, 2, 2});
17688 Amt3 = DAG.getVectorShuffle(VT, dl, Amt, DAG.getUNDEF(VT), {3, 3, 3, 3});
17690 // ISD::SHL is handled above but we include it here for completeness.
17693 llvm_unreachable("Unknown target vector shift node");
17695 Opc = X86ISD::VSHL;
17698 Opc = X86ISD::VSRL;
17701 Opc = X86ISD::VSRA;
17704 // The SSE2 shifts use the lower i64 as the same shift amount for
17705 // all lanes and the upper i64 is ignored. These shuffle masks
17706 // optimally zero-extend each lanes on SSE2/SSE41/AVX targets.
17707 SDValue Z = getZeroVector(VT, Subtarget, DAG, dl);
17708 Amt0 = DAG.getVectorShuffle(VT, dl, Amt, Z, {0, 4, -1, -1});
17709 Amt1 = DAG.getVectorShuffle(VT, dl, Amt, Z, {1, 5, -1, -1});
17710 Amt2 = DAG.getVectorShuffle(VT, dl, Amt, Z, {2, 6, -1, -1});
17711 Amt3 = DAG.getVectorShuffle(VT, dl, Amt, Z, {3, 7, -1, -1});
17714 SDValue R0 = DAG.getNode(Opc, dl, VT, R, Amt0);
17715 SDValue R1 = DAG.getNode(Opc, dl, VT, R, Amt1);
17716 SDValue R2 = DAG.getNode(Opc, dl, VT, R, Amt2);
17717 SDValue R3 = DAG.getNode(Opc, dl, VT, R, Amt3);
17718 SDValue R02 = DAG.getVectorShuffle(VT, dl, R0, R2, {0, -1, 6, -1});
17719 SDValue R13 = DAG.getVectorShuffle(VT, dl, R1, R3, {-1, 1, -1, 7});
17720 return DAG.getVectorShuffle(VT, dl, R02, R13, {0, 5, 2, 7});
17723 if (VT == MVT::v16i8 || (VT == MVT::v32i8 && Subtarget->hasInt256())) {
17724 MVT ExtVT = MVT::getVectorVT(MVT::i16, VT.getVectorNumElements() / 2);
17725 unsigned ShiftOpcode = Op->getOpcode();
17727 auto SignBitSelect = [&](MVT SelVT, SDValue Sel, SDValue V0, SDValue V1) {
17728 // On SSE41 targets we make use of the fact that VSELECT lowers
17729 // to PBLENDVB which selects bytes based just on the sign bit.
17730 if (Subtarget->hasSSE41()) {
17731 V0 = DAG.getBitcast(VT, V0);
17732 V1 = DAG.getBitcast(VT, V1);
17733 Sel = DAG.getBitcast(VT, Sel);
17734 return DAG.getBitcast(SelVT,
17735 DAG.getNode(ISD::VSELECT, dl, VT, Sel, V0, V1));
17737 // On pre-SSE41 targets we test for the sign bit by comparing to
17738 // zero - a negative value will set all bits of the lanes to true
17739 // and VSELECT uses that in its OR(AND(V0,C),AND(V1,~C)) lowering.
17740 SDValue Z = getZeroVector(SelVT, Subtarget, DAG, dl);
17741 SDValue C = DAG.getNode(X86ISD::PCMPGT, dl, SelVT, Z, Sel);
17742 return DAG.getNode(ISD::VSELECT, dl, SelVT, C, V0, V1);
17745 // Turn 'a' into a mask suitable for VSELECT: a = a << 5;
17746 // We can safely do this using i16 shifts as we're only interested in
17747 // the 3 lower bits of each byte.
17748 Amt = DAG.getBitcast(ExtVT, Amt);
17749 Amt = DAG.getNode(ISD::SHL, dl, ExtVT, Amt, DAG.getConstant(5, dl, ExtVT));
17750 Amt = DAG.getBitcast(VT, Amt);
17752 if (Op->getOpcode() == ISD::SHL || Op->getOpcode() == ISD::SRL) {
17753 // r = VSELECT(r, shift(r, 4), a);
17755 DAG.getNode(ShiftOpcode, dl, VT, R, DAG.getConstant(4, dl, VT));
17756 R = SignBitSelect(VT, Amt, M, R);
17759 Amt = DAG.getNode(ISD::ADD, dl, VT, Amt, Amt);
17761 // r = VSELECT(r, shift(r, 2), a);
17762 M = DAG.getNode(ShiftOpcode, dl, VT, R, DAG.getConstant(2, dl, VT));
17763 R = SignBitSelect(VT, Amt, M, R);
17766 Amt = DAG.getNode(ISD::ADD, dl, VT, Amt, Amt);
17768 // return VSELECT(r, shift(r, 1), a);
17769 M = DAG.getNode(ShiftOpcode, dl, VT, R, DAG.getConstant(1, dl, VT));
17770 R = SignBitSelect(VT, Amt, M, R);
17774 if (Op->getOpcode() == ISD::SRA) {
17775 // For SRA we need to unpack each byte to the higher byte of a i16 vector
17776 // so we can correctly sign extend. We don't care what happens to the
17778 SDValue ALo = DAG.getNode(X86ISD::UNPCKL, dl, VT, DAG.getUNDEF(VT), Amt);
17779 SDValue AHi = DAG.getNode(X86ISD::UNPCKH, dl, VT, DAG.getUNDEF(VT), Amt);
17780 SDValue RLo = DAG.getNode(X86ISD::UNPCKL, dl, VT, DAG.getUNDEF(VT), R);
17781 SDValue RHi = DAG.getNode(X86ISD::UNPCKH, dl, VT, DAG.getUNDEF(VT), R);
17782 ALo = DAG.getBitcast(ExtVT, ALo);
17783 AHi = DAG.getBitcast(ExtVT, AHi);
17784 RLo = DAG.getBitcast(ExtVT, RLo);
17785 RHi = DAG.getBitcast(ExtVT, RHi);
17787 // r = VSELECT(r, shift(r, 4), a);
17788 SDValue MLo = DAG.getNode(ShiftOpcode, dl, ExtVT, RLo,
17789 DAG.getConstant(4, dl, ExtVT));
17790 SDValue MHi = DAG.getNode(ShiftOpcode, dl, ExtVT, RHi,
17791 DAG.getConstant(4, dl, ExtVT));
17792 RLo = SignBitSelect(ExtVT, ALo, MLo, RLo);
17793 RHi = SignBitSelect(ExtVT, AHi, MHi, RHi);
17796 ALo = DAG.getNode(ISD::ADD, dl, ExtVT, ALo, ALo);
17797 AHi = DAG.getNode(ISD::ADD, dl, ExtVT, AHi, AHi);
17799 // r = VSELECT(r, shift(r, 2), a);
17800 MLo = DAG.getNode(ShiftOpcode, dl, ExtVT, RLo,
17801 DAG.getConstant(2, dl, ExtVT));
17802 MHi = DAG.getNode(ShiftOpcode, dl, ExtVT, RHi,
17803 DAG.getConstant(2, dl, ExtVT));
17804 RLo = SignBitSelect(ExtVT, ALo, MLo, RLo);
17805 RHi = SignBitSelect(ExtVT, AHi, MHi, RHi);
17808 ALo = DAG.getNode(ISD::ADD, dl, ExtVT, ALo, ALo);
17809 AHi = DAG.getNode(ISD::ADD, dl, ExtVT, AHi, AHi);
17811 // r = VSELECT(r, shift(r, 1), a);
17812 MLo = DAG.getNode(ShiftOpcode, dl, ExtVT, RLo,
17813 DAG.getConstant(1, dl, ExtVT));
17814 MHi = DAG.getNode(ShiftOpcode, dl, ExtVT, RHi,
17815 DAG.getConstant(1, dl, ExtVT));
17816 RLo = SignBitSelect(ExtVT, ALo, MLo, RLo);
17817 RHi = SignBitSelect(ExtVT, AHi, MHi, RHi);
17819 // Logical shift the result back to the lower byte, leaving a zero upper
17821 // meaning that we can safely pack with PACKUSWB.
17823 DAG.getNode(ISD::SRL, dl, ExtVT, RLo, DAG.getConstant(8, dl, ExtVT));
17825 DAG.getNode(ISD::SRL, dl, ExtVT, RHi, DAG.getConstant(8, dl, ExtVT));
17826 return DAG.getNode(X86ISD::PACKUS, dl, VT, RLo, RHi);
17830 // It's worth extending once and using the v8i32 shifts for 16-bit types, but
17831 // the extra overheads to get from v16i8 to v8i32 make the existing SSE
17832 // solution better.
17833 if (Subtarget->hasInt256() && VT == MVT::v8i16) {
17834 MVT ExtVT = MVT::v8i32;
17836 Op.getOpcode() == ISD::SRA ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND;
17837 R = DAG.getNode(ExtOpc, dl, ExtVT, R);
17838 Amt = DAG.getNode(ISD::ANY_EXTEND, dl, ExtVT, Amt);
17839 return DAG.getNode(ISD::TRUNCATE, dl, VT,
17840 DAG.getNode(Op.getOpcode(), dl, ExtVT, R, Amt));
17843 if (Subtarget->hasInt256() && VT == MVT::v16i16) {
17844 MVT ExtVT = MVT::v8i32;
17845 SDValue Z = getZeroVector(VT, Subtarget, DAG, dl);
17846 SDValue ALo = DAG.getNode(X86ISD::UNPCKL, dl, VT, Amt, Z);
17847 SDValue AHi = DAG.getNode(X86ISD::UNPCKH, dl, VT, Amt, Z);
17848 SDValue RLo = DAG.getNode(X86ISD::UNPCKL, dl, VT, R, R);
17849 SDValue RHi = DAG.getNode(X86ISD::UNPCKH, dl, VT, R, R);
17850 ALo = DAG.getBitcast(ExtVT, ALo);
17851 AHi = DAG.getBitcast(ExtVT, AHi);
17852 RLo = DAG.getBitcast(ExtVT, RLo);
17853 RHi = DAG.getBitcast(ExtVT, RHi);
17854 SDValue Lo = DAG.getNode(Op.getOpcode(), dl, ExtVT, RLo, ALo);
17855 SDValue Hi = DAG.getNode(Op.getOpcode(), dl, ExtVT, RHi, AHi);
17856 Lo = DAG.getNode(ISD::SRL, dl, ExtVT, Lo, DAG.getConstant(16, dl, ExtVT));
17857 Hi = DAG.getNode(ISD::SRL, dl, ExtVT, Hi, DAG.getConstant(16, dl, ExtVT));
17858 return DAG.getNode(X86ISD::PACKUS, dl, VT, Lo, Hi);
17861 if (VT == MVT::v8i16) {
17862 unsigned ShiftOpcode = Op->getOpcode();
17864 auto SignBitSelect = [&](SDValue Sel, SDValue V0, SDValue V1) {
17865 // On SSE41 targets we make use of the fact that VSELECT lowers
17866 // to PBLENDVB which selects bytes based just on the sign bit.
17867 if (Subtarget->hasSSE41()) {
17868 MVT ExtVT = MVT::getVectorVT(MVT::i8, VT.getVectorNumElements() * 2);
17869 V0 = DAG.getBitcast(ExtVT, V0);
17870 V1 = DAG.getBitcast(ExtVT, V1);
17871 Sel = DAG.getBitcast(ExtVT, Sel);
17872 return DAG.getBitcast(
17873 VT, DAG.getNode(ISD::VSELECT, dl, ExtVT, Sel, V0, V1));
17875 // On pre-SSE41 targets we splat the sign bit - a negative value will
17876 // set all bits of the lanes to true and VSELECT uses that in
17877 // its OR(AND(V0,C),AND(V1,~C)) lowering.
17879 DAG.getNode(ISD::SRA, dl, VT, Sel, DAG.getConstant(15, dl, VT));
17880 return DAG.getNode(ISD::VSELECT, dl, VT, C, V0, V1);
17883 // Turn 'a' into a mask suitable for VSELECT: a = a << 12;
17884 if (Subtarget->hasSSE41()) {
17885 // On SSE41 targets we need to replicate the shift mask in both
17886 // bytes for PBLENDVB.
17889 DAG.getNode(ISD::SHL, dl, VT, Amt, DAG.getConstant(4, dl, VT)),
17890 DAG.getNode(ISD::SHL, dl, VT, Amt, DAG.getConstant(12, dl, VT)));
17892 Amt = DAG.getNode(ISD::SHL, dl, VT, Amt, DAG.getConstant(12, dl, VT));
17895 // r = VSELECT(r, shift(r, 8), a);
17896 SDValue M = DAG.getNode(ShiftOpcode, dl, VT, R, DAG.getConstant(8, dl, VT));
17897 R = SignBitSelect(Amt, M, R);
17900 Amt = DAG.getNode(ISD::ADD, dl, VT, Amt, Amt);
17902 // r = VSELECT(r, shift(r, 4), a);
17903 M = DAG.getNode(ShiftOpcode, dl, VT, R, DAG.getConstant(4, dl, VT));
17904 R = SignBitSelect(Amt, M, R);
17907 Amt = DAG.getNode(ISD::ADD, dl, VT, Amt, Amt);
17909 // r = VSELECT(r, shift(r, 2), a);
17910 M = DAG.getNode(ShiftOpcode, dl, VT, R, DAG.getConstant(2, dl, VT));
17911 R = SignBitSelect(Amt, M, R);
17914 Amt = DAG.getNode(ISD::ADD, dl, VT, Amt, Amt);
17916 // return VSELECT(r, shift(r, 1), a);
17917 M = DAG.getNode(ShiftOpcode, dl, VT, R, DAG.getConstant(1, dl, VT));
17918 R = SignBitSelect(Amt, M, R);
17922 // Decompose 256-bit shifts into smaller 128-bit shifts.
17923 if (VT.is256BitVector()) {
17924 unsigned NumElems = VT.getVectorNumElements();
17925 MVT EltVT = VT.getVectorElementType();
17926 EVT NewVT = MVT::getVectorVT(EltVT, NumElems/2);
17928 // Extract the two vectors
17929 SDValue V1 = Extract128BitVector(R, 0, DAG, dl);
17930 SDValue V2 = Extract128BitVector(R, NumElems/2, DAG, dl);
17932 // Recreate the shift amount vectors
17933 SDValue Amt1, Amt2;
17934 if (Amt.getOpcode() == ISD::BUILD_VECTOR) {
17935 // Constant shift amount
17936 SmallVector<SDValue, 8> Ops(Amt->op_begin(), Amt->op_begin() + NumElems);
17937 ArrayRef<SDValue> Amt1Csts = makeArrayRef(Ops).slice(0, NumElems / 2);
17938 ArrayRef<SDValue> Amt2Csts = makeArrayRef(Ops).slice(NumElems / 2);
17940 Amt1 = DAG.getNode(ISD::BUILD_VECTOR, dl, NewVT, Amt1Csts);
17941 Amt2 = DAG.getNode(ISD::BUILD_VECTOR, dl, NewVT, Amt2Csts);
17943 // Variable shift amount
17944 Amt1 = Extract128BitVector(Amt, 0, DAG, dl);
17945 Amt2 = Extract128BitVector(Amt, NumElems/2, DAG, dl);
17948 // Issue new vector shifts for the smaller types
17949 V1 = DAG.getNode(Op.getOpcode(), dl, NewVT, V1, Amt1);
17950 V2 = DAG.getNode(Op.getOpcode(), dl, NewVT, V2, Amt2);
17952 // Concatenate the result back
17953 return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT, V1, V2);
17959 static SDValue LowerXALUO(SDValue Op, SelectionDAG &DAG) {
17960 // Lower the "add/sub/mul with overflow" instruction into a regular ins plus
17961 // a "setcc" instruction that checks the overflow flag. The "brcond" lowering
17962 // looks for this combo and may remove the "setcc" instruction if the "setcc"
17963 // has only one use.
17964 SDNode *N = Op.getNode();
17965 SDValue LHS = N->getOperand(0);
17966 SDValue RHS = N->getOperand(1);
17967 unsigned BaseOp = 0;
17970 switch (Op.getOpcode()) {
17971 default: llvm_unreachable("Unknown ovf instruction!");
17973 // A subtract of one will be selected as a INC. Note that INC doesn't
17974 // set CF, so we can't do this for UADDO.
17975 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(RHS))
17977 BaseOp = X86ISD::INC;
17978 Cond = X86::COND_O;
17981 BaseOp = X86ISD::ADD;
17982 Cond = X86::COND_O;
17985 BaseOp = X86ISD::ADD;
17986 Cond = X86::COND_B;
17989 // A subtract of one will be selected as a DEC. Note that DEC doesn't
17990 // set CF, so we can't do this for USUBO.
17991 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(RHS))
17993 BaseOp = X86ISD::DEC;
17994 Cond = X86::COND_O;
17997 BaseOp = X86ISD::SUB;
17998 Cond = X86::COND_O;
18001 BaseOp = X86ISD::SUB;
18002 Cond = X86::COND_B;
18005 BaseOp = N->getValueType(0) == MVT::i8 ? X86ISD::SMUL8 : X86ISD::SMUL;
18006 Cond = X86::COND_O;
18008 case ISD::UMULO: { // i64, i8 = umulo lhs, rhs --> i64, i64, i32 umul lhs,rhs
18009 if (N->getValueType(0) == MVT::i8) {
18010 BaseOp = X86ISD::UMUL8;
18011 Cond = X86::COND_O;
18014 SDVTList VTs = DAG.getVTList(N->getValueType(0), N->getValueType(0),
18016 SDValue Sum = DAG.getNode(X86ISD::UMUL, DL, VTs, LHS, RHS);
18019 DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
18020 DAG.getConstant(X86::COND_O, DL, MVT::i32),
18021 SDValue(Sum.getNode(), 2));
18023 return DAG.getNode(ISD::MERGE_VALUES, DL, N->getVTList(), Sum, SetCC);
18027 // Also sets EFLAGS.
18028 SDVTList VTs = DAG.getVTList(N->getValueType(0), MVT::i32);
18029 SDValue Sum = DAG.getNode(BaseOp, DL, VTs, LHS, RHS);
18032 DAG.getNode(X86ISD::SETCC, DL, N->getValueType(1),
18033 DAG.getConstant(Cond, DL, MVT::i32),
18034 SDValue(Sum.getNode(), 1));
18036 return DAG.getNode(ISD::MERGE_VALUES, DL, N->getVTList(), Sum, SetCC);
18039 /// Returns true if the operand type is exactly twice the native width, and
18040 /// the corresponding cmpxchg8b or cmpxchg16b instruction is available.
18041 /// Used to know whether to use cmpxchg8/16b when expanding atomic operations
18042 /// (otherwise we leave them alone to become __sync_fetch_and_... calls).
18043 bool X86TargetLowering::needsCmpXchgNb(Type *MemType) const {
18044 unsigned OpWidth = MemType->getPrimitiveSizeInBits();
18047 return !Subtarget->is64Bit(); // FIXME this should be Subtarget.hasCmpxchg8b
18048 else if (OpWidth == 128)
18049 return Subtarget->hasCmpxchg16b();
18054 bool X86TargetLowering::shouldExpandAtomicStoreInIR(StoreInst *SI) const {
18055 return needsCmpXchgNb(SI->getValueOperand()->getType());
18058 // Note: this turns large loads into lock cmpxchg8b/16b.
18059 // FIXME: On 32 bits x86, fild/movq might be faster than lock cmpxchg8b.
18060 bool X86TargetLowering::shouldExpandAtomicLoadInIR(LoadInst *LI) const {
18061 auto PTy = cast<PointerType>(LI->getPointerOperand()->getType());
18062 return needsCmpXchgNb(PTy->getElementType());
18065 TargetLoweringBase::AtomicRMWExpansionKind
18066 X86TargetLowering::shouldExpandAtomicRMWInIR(AtomicRMWInst *AI) const {
18067 unsigned NativeWidth = Subtarget->is64Bit() ? 64 : 32;
18068 Type *MemType = AI->getType();
18070 // If the operand is too big, we must see if cmpxchg8/16b is available
18071 // and default to library calls otherwise.
18072 if (MemType->getPrimitiveSizeInBits() > NativeWidth) {
18073 return needsCmpXchgNb(MemType) ? AtomicRMWExpansionKind::CmpXChg
18074 : AtomicRMWExpansionKind::None;
18077 AtomicRMWInst::BinOp Op = AI->getOperation();
18080 llvm_unreachable("Unknown atomic operation");
18081 case AtomicRMWInst::Xchg:
18082 case AtomicRMWInst::Add:
18083 case AtomicRMWInst::Sub:
18084 // It's better to use xadd, xsub or xchg for these in all cases.
18085 return AtomicRMWExpansionKind::None;
18086 case AtomicRMWInst::Or:
18087 case AtomicRMWInst::And:
18088 case AtomicRMWInst::Xor:
18089 // If the atomicrmw's result isn't actually used, we can just add a "lock"
18090 // prefix to a normal instruction for these operations.
18091 return !AI->use_empty() ? AtomicRMWExpansionKind::CmpXChg
18092 : AtomicRMWExpansionKind::None;
18093 case AtomicRMWInst::Nand:
18094 case AtomicRMWInst::Max:
18095 case AtomicRMWInst::Min:
18096 case AtomicRMWInst::UMax:
18097 case AtomicRMWInst::UMin:
18098 // These always require a non-trivial set of data operations on x86. We must
18099 // use a cmpxchg loop.
18100 return AtomicRMWExpansionKind::CmpXChg;
18104 static bool hasMFENCE(const X86Subtarget& Subtarget) {
18105 // Use mfence if we have SSE2 or we're on x86-64 (even if we asked for
18106 // no-sse2). There isn't any reason to disable it if the target processor
18108 return Subtarget.hasSSE2() || Subtarget.is64Bit();
18112 X86TargetLowering::lowerIdempotentRMWIntoFencedLoad(AtomicRMWInst *AI) const {
18113 unsigned NativeWidth = Subtarget->is64Bit() ? 64 : 32;
18114 Type *MemType = AI->getType();
18115 // Accesses larger than the native width are turned into cmpxchg/libcalls, so
18116 // there is no benefit in turning such RMWs into loads, and it is actually
18117 // harmful as it introduces a mfence.
18118 if (MemType->getPrimitiveSizeInBits() > NativeWidth)
18121 auto Builder = IRBuilder<>(AI);
18122 Module *M = Builder.GetInsertBlock()->getParent()->getParent();
18123 auto SynchScope = AI->getSynchScope();
18124 // We must restrict the ordering to avoid generating loads with Release or
18125 // ReleaseAcquire orderings.
18126 auto Order = AtomicCmpXchgInst::getStrongestFailureOrdering(AI->getOrdering());
18127 auto Ptr = AI->getPointerOperand();
18129 // Before the load we need a fence. Here is an example lifted from
18130 // http://www.hpl.hp.com/techreports/2012/HPL-2012-68.pdf showing why a fence
18133 // x.store(1, relaxed);
18134 // r1 = y.fetch_add(0, release);
18136 // y.fetch_add(42, acquire);
18137 // r2 = x.load(relaxed);
18138 // r1 = r2 = 0 is impossible, but becomes possible if the idempotent rmw is
18139 // lowered to just a load without a fence. A mfence flushes the store buffer,
18140 // making the optimization clearly correct.
18141 // FIXME: it is required if isAtLeastRelease(Order) but it is not clear
18142 // otherwise, we might be able to be more aggressive on relaxed idempotent
18143 // rmw. In practice, they do not look useful, so we don't try to be
18144 // especially clever.
18145 if (SynchScope == SingleThread)
18146 // FIXME: we could just insert an X86ISD::MEMBARRIER here, except we are at
18147 // the IR level, so we must wrap it in an intrinsic.
18150 if (!hasMFENCE(*Subtarget))
18151 // FIXME: it might make sense to use a locked operation here but on a
18152 // different cache-line to prevent cache-line bouncing. In practice it
18153 // is probably a small win, and x86 processors without mfence are rare
18154 // enough that we do not bother.
18158 llvm::Intrinsic::getDeclaration(M, Intrinsic::x86_sse2_mfence);
18159 Builder.CreateCall(MFence, {});
18161 // Finally we can emit the atomic load.
18162 LoadInst *Loaded = Builder.CreateAlignedLoad(Ptr,
18163 AI->getType()->getPrimitiveSizeInBits());
18164 Loaded->setAtomic(Order, SynchScope);
18165 AI->replaceAllUsesWith(Loaded);
18166 AI->eraseFromParent();
18170 static SDValue LowerATOMIC_FENCE(SDValue Op, const X86Subtarget *Subtarget,
18171 SelectionDAG &DAG) {
18173 AtomicOrdering FenceOrdering = static_cast<AtomicOrdering>(
18174 cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue());
18175 SynchronizationScope FenceScope = static_cast<SynchronizationScope>(
18176 cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue());
18178 // The only fence that needs an instruction is a sequentially-consistent
18179 // cross-thread fence.
18180 if (FenceOrdering == SequentiallyConsistent && FenceScope == CrossThread) {
18181 if (hasMFENCE(*Subtarget))
18182 return DAG.getNode(X86ISD::MFENCE, dl, MVT::Other, Op.getOperand(0));
18184 SDValue Chain = Op.getOperand(0);
18185 SDValue Zero = DAG.getConstant(0, dl, MVT::i32);
18187 DAG.getRegister(X86::ESP, MVT::i32), // Base
18188 DAG.getTargetConstant(1, dl, MVT::i8), // Scale
18189 DAG.getRegister(0, MVT::i32), // Index
18190 DAG.getTargetConstant(0, dl, MVT::i32), // Disp
18191 DAG.getRegister(0, MVT::i32), // Segment.
18195 SDNode *Res = DAG.getMachineNode(X86::OR32mrLocked, dl, MVT::Other, Ops);
18196 return SDValue(Res, 0);
18199 // MEMBARRIER is a compiler barrier; it codegens to a no-op.
18200 return DAG.getNode(X86ISD::MEMBARRIER, dl, MVT::Other, Op.getOperand(0));
18203 static SDValue LowerCMP_SWAP(SDValue Op, const X86Subtarget *Subtarget,
18204 SelectionDAG &DAG) {
18205 MVT T = Op.getSimpleValueType();
18209 switch(T.SimpleTy) {
18210 default: llvm_unreachable("Invalid value type!");
18211 case MVT::i8: Reg = X86::AL; size = 1; break;
18212 case MVT::i16: Reg = X86::AX; size = 2; break;
18213 case MVT::i32: Reg = X86::EAX; size = 4; break;
18215 assert(Subtarget->is64Bit() && "Node not type legal!");
18216 Reg = X86::RAX; size = 8;
18219 SDValue cpIn = DAG.getCopyToReg(Op.getOperand(0), DL, Reg,
18220 Op.getOperand(2), SDValue());
18221 SDValue Ops[] = { cpIn.getValue(0),
18224 DAG.getTargetConstant(size, DL, MVT::i8),
18225 cpIn.getValue(1) };
18226 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Glue);
18227 MachineMemOperand *MMO = cast<AtomicSDNode>(Op)->getMemOperand();
18228 SDValue Result = DAG.getMemIntrinsicNode(X86ISD::LCMPXCHG_DAG, DL, Tys,
18232 DAG.getCopyFromReg(Result.getValue(0), DL, Reg, T, Result.getValue(1));
18233 SDValue EFLAGS = DAG.getCopyFromReg(cpOut.getValue(1), DL, X86::EFLAGS,
18234 MVT::i32, cpOut.getValue(2));
18235 SDValue Success = DAG.getNode(X86ISD::SETCC, DL, Op->getValueType(1),
18236 DAG.getConstant(X86::COND_E, DL, MVT::i8),
18239 DAG.ReplaceAllUsesOfValueWith(Op.getValue(0), cpOut);
18240 DAG.ReplaceAllUsesOfValueWith(Op.getValue(1), Success);
18241 DAG.ReplaceAllUsesOfValueWith(Op.getValue(2), EFLAGS.getValue(1));
18245 static SDValue LowerBITCAST(SDValue Op, const X86Subtarget *Subtarget,
18246 SelectionDAG &DAG) {
18247 MVT SrcVT = Op.getOperand(0).getSimpleValueType();
18248 MVT DstVT = Op.getSimpleValueType();
18250 if (SrcVT == MVT::v2i32 || SrcVT == MVT::v4i16 || SrcVT == MVT::v8i8) {
18251 assert(Subtarget->hasSSE2() && "Requires at least SSE2!");
18252 if (DstVT != MVT::f64)
18253 // This conversion needs to be expanded.
18256 SDValue InVec = Op->getOperand(0);
18258 unsigned NumElts = SrcVT.getVectorNumElements();
18259 EVT SVT = SrcVT.getVectorElementType();
18261 // Widen the vector in input in the case of MVT::v2i32.
18262 // Example: from MVT::v2i32 to MVT::v4i32.
18263 SmallVector<SDValue, 16> Elts;
18264 for (unsigned i = 0, e = NumElts; i != e; ++i)
18265 Elts.push_back(DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, SVT, InVec,
18266 DAG.getIntPtrConstant(i, dl)));
18268 // Explicitly mark the extra elements as Undef.
18269 Elts.append(NumElts, DAG.getUNDEF(SVT));
18271 EVT NewVT = EVT::getVectorVT(*DAG.getContext(), SVT, NumElts * 2);
18272 SDValue BV = DAG.getNode(ISD::BUILD_VECTOR, dl, NewVT, Elts);
18273 SDValue ToV2F64 = DAG.getBitcast(MVT::v2f64, BV);
18274 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, ToV2F64,
18275 DAG.getIntPtrConstant(0, dl));
18278 assert(Subtarget->is64Bit() && !Subtarget->hasSSE2() &&
18279 Subtarget->hasMMX() && "Unexpected custom BITCAST");
18280 assert((DstVT == MVT::i64 ||
18281 (DstVT.isVector() && DstVT.getSizeInBits()==64)) &&
18282 "Unexpected custom BITCAST");
18283 // i64 <=> MMX conversions are Legal.
18284 if (SrcVT==MVT::i64 && DstVT.isVector())
18286 if (DstVT==MVT::i64 && SrcVT.isVector())
18288 // MMX <=> MMX conversions are Legal.
18289 if (SrcVT.isVector() && DstVT.isVector())
18291 // All other conversions need to be expanded.
18295 /// Compute the horizontal sum of bytes in V for the elements of VT.
18297 /// Requires V to be a byte vector and VT to be an integer vector type with
18298 /// wider elements than V's type. The width of the elements of VT determines
18299 /// how many bytes of V are summed horizontally to produce each element of the
18301 static SDValue LowerHorizontalByteSum(SDValue V, MVT VT,
18302 const X86Subtarget *Subtarget,
18303 SelectionDAG &DAG) {
18305 MVT ByteVecVT = V.getSimpleValueType();
18306 MVT EltVT = VT.getVectorElementType();
18307 int NumElts = VT.getVectorNumElements();
18308 assert(ByteVecVT.getVectorElementType() == MVT::i8 &&
18309 "Expected value to have byte element type.");
18310 assert(EltVT != MVT::i8 &&
18311 "Horizontal byte sum only makes sense for wider elements!");
18312 unsigned VecSize = VT.getSizeInBits();
18313 assert(ByteVecVT.getSizeInBits() == VecSize && "Cannot change vector size!");
18315 // PSADBW instruction horizontally add all bytes and leave the result in i64
18316 // chunks, thus directly computes the pop count for v2i64 and v4i64.
18317 if (EltVT == MVT::i64) {
18318 SDValue Zeros = getZeroVector(ByteVecVT, Subtarget, DAG, DL);
18319 V = DAG.getNode(X86ISD::PSADBW, DL, ByteVecVT, V, Zeros);
18320 return DAG.getBitcast(VT, V);
18323 if (EltVT == MVT::i32) {
18324 // We unpack the low half and high half into i32s interleaved with zeros so
18325 // that we can use PSADBW to horizontally sum them. The most useful part of
18326 // this is that it lines up the results of two PSADBW instructions to be
18327 // two v2i64 vectors which concatenated are the 4 population counts. We can
18328 // then use PACKUSWB to shrink and concatenate them into a v4i32 again.
18329 SDValue Zeros = getZeroVector(VT, Subtarget, DAG, DL);
18330 SDValue Low = DAG.getNode(X86ISD::UNPCKL, DL, VT, V, Zeros);
18331 SDValue High = DAG.getNode(X86ISD::UNPCKH, DL, VT, V, Zeros);
18333 // Do the horizontal sums into two v2i64s.
18334 Zeros = getZeroVector(ByteVecVT, Subtarget, DAG, DL);
18335 Low = DAG.getNode(X86ISD::PSADBW, DL, ByteVecVT,
18336 DAG.getBitcast(ByteVecVT, Low), Zeros);
18337 High = DAG.getNode(X86ISD::PSADBW, DL, ByteVecVT,
18338 DAG.getBitcast(ByteVecVT, High), Zeros);
18340 // Merge them together.
18341 MVT ShortVecVT = MVT::getVectorVT(MVT::i16, VecSize / 16);
18342 V = DAG.getNode(X86ISD::PACKUS, DL, ByteVecVT,
18343 DAG.getBitcast(ShortVecVT, Low),
18344 DAG.getBitcast(ShortVecVT, High));
18346 return DAG.getBitcast(VT, V);
18349 // The only element type left is i16.
18350 assert(EltVT == MVT::i16 && "Unknown how to handle type");
18352 // To obtain pop count for each i16 element starting from the pop count for
18353 // i8 elements, shift the i16s left by 8, sum as i8s, and then shift as i16s
18354 // right by 8. It is important to shift as i16s as i8 vector shift isn't
18355 // directly supported.
18356 SmallVector<SDValue, 16> Shifters(NumElts, DAG.getConstant(8, DL, EltVT));
18357 SDValue Shifter = DAG.getNode(ISD::BUILD_VECTOR, DL, VT, Shifters);
18358 SDValue Shl = DAG.getNode(ISD::SHL, DL, VT, DAG.getBitcast(VT, V), Shifter);
18359 V = DAG.getNode(ISD::ADD, DL, ByteVecVT, DAG.getBitcast(ByteVecVT, Shl),
18360 DAG.getBitcast(ByteVecVT, V));
18361 return DAG.getNode(ISD::SRL, DL, VT, DAG.getBitcast(VT, V), Shifter);
18364 static SDValue LowerVectorCTPOPInRegLUT(SDValue Op, SDLoc DL,
18365 const X86Subtarget *Subtarget,
18366 SelectionDAG &DAG) {
18367 MVT VT = Op.getSimpleValueType();
18368 MVT EltVT = VT.getVectorElementType();
18369 unsigned VecSize = VT.getSizeInBits();
18371 // Implement a lookup table in register by using an algorithm based on:
18372 // http://wm.ite.pl/articles/sse-popcount.html
18374 // The general idea is that every lower byte nibble in the input vector is an
18375 // index into a in-register pre-computed pop count table. We then split up the
18376 // input vector in two new ones: (1) a vector with only the shifted-right
18377 // higher nibbles for each byte and (2) a vector with the lower nibbles (and
18378 // masked out higher ones) for each byte. PSHUB is used separately with both
18379 // to index the in-register table. Next, both are added and the result is a
18380 // i8 vector where each element contains the pop count for input byte.
18382 // To obtain the pop count for elements != i8, we follow up with the same
18383 // approach and use additional tricks as described below.
18385 const int LUT[16] = {/* 0 */ 0, /* 1 */ 1, /* 2 */ 1, /* 3 */ 2,
18386 /* 4 */ 1, /* 5 */ 2, /* 6 */ 2, /* 7 */ 3,
18387 /* 8 */ 1, /* 9 */ 2, /* a */ 2, /* b */ 3,
18388 /* c */ 2, /* d */ 3, /* e */ 3, /* f */ 4};
18390 int NumByteElts = VecSize / 8;
18391 MVT ByteVecVT = MVT::getVectorVT(MVT::i8, NumByteElts);
18392 SDValue In = DAG.getBitcast(ByteVecVT, Op);
18393 SmallVector<SDValue, 16> LUTVec;
18394 for (int i = 0; i < NumByteElts; ++i)
18395 LUTVec.push_back(DAG.getConstant(LUT[i % 16], DL, MVT::i8));
18396 SDValue InRegLUT = DAG.getNode(ISD::BUILD_VECTOR, DL, ByteVecVT, LUTVec);
18397 SmallVector<SDValue, 16> Mask0F(NumByteElts,
18398 DAG.getConstant(0x0F, DL, MVT::i8));
18399 SDValue M0F = DAG.getNode(ISD::BUILD_VECTOR, DL, ByteVecVT, Mask0F);
18402 SmallVector<SDValue, 16> Four(NumByteElts, DAG.getConstant(4, DL, MVT::i8));
18403 SDValue FourV = DAG.getNode(ISD::BUILD_VECTOR, DL, ByteVecVT, Four);
18404 SDValue HighNibbles = DAG.getNode(ISD::SRL, DL, ByteVecVT, In, FourV);
18407 SDValue LowNibbles = DAG.getNode(ISD::AND, DL, ByteVecVT, In, M0F);
18409 // The input vector is used as the shuffle mask that index elements into the
18410 // LUT. After counting low and high nibbles, add the vector to obtain the
18411 // final pop count per i8 element.
18412 SDValue HighPopCnt =
18413 DAG.getNode(X86ISD::PSHUFB, DL, ByteVecVT, InRegLUT, HighNibbles);
18414 SDValue LowPopCnt =
18415 DAG.getNode(X86ISD::PSHUFB, DL, ByteVecVT, InRegLUT, LowNibbles);
18416 SDValue PopCnt = DAG.getNode(ISD::ADD, DL, ByteVecVT, HighPopCnt, LowPopCnt);
18418 if (EltVT == MVT::i8)
18421 return LowerHorizontalByteSum(PopCnt, VT, Subtarget, DAG);
18424 static SDValue LowerVectorCTPOPBitmath(SDValue Op, SDLoc DL,
18425 const X86Subtarget *Subtarget,
18426 SelectionDAG &DAG) {
18427 MVT VT = Op.getSimpleValueType();
18428 assert(VT.is128BitVector() &&
18429 "Only 128-bit vector bitmath lowering supported.");
18431 int VecSize = VT.getSizeInBits();
18432 MVT EltVT = VT.getVectorElementType();
18433 int Len = EltVT.getSizeInBits();
18435 // This is the vectorized version of the "best" algorithm from
18436 // http://graphics.stanford.edu/~seander/bithacks.html#CountBitsSetParallel
18437 // with a minor tweak to use a series of adds + shifts instead of vector
18438 // multiplications. Implemented for all integer vector types. We only use
18439 // this when we don't have SSSE3 which allows a LUT-based lowering that is
18440 // much faster, even faster than using native popcnt instructions.
18442 auto GetShift = [&](unsigned OpCode, SDValue V, int Shifter) {
18443 MVT VT = V.getSimpleValueType();
18444 SmallVector<SDValue, 32> Shifters(
18445 VT.getVectorNumElements(),
18446 DAG.getConstant(Shifter, DL, VT.getVectorElementType()));
18447 return DAG.getNode(OpCode, DL, VT, V,
18448 DAG.getNode(ISD::BUILD_VECTOR, DL, VT, Shifters));
18450 auto GetMask = [&](SDValue V, APInt Mask) {
18451 MVT VT = V.getSimpleValueType();
18452 SmallVector<SDValue, 32> Masks(
18453 VT.getVectorNumElements(),
18454 DAG.getConstant(Mask, DL, VT.getVectorElementType()));
18455 return DAG.getNode(ISD::AND, DL, VT, V,
18456 DAG.getNode(ISD::BUILD_VECTOR, DL, VT, Masks));
18459 // We don't want to incur the implicit masks required to SRL vNi8 vectors on
18460 // x86, so set the SRL type to have elements at least i16 wide. This is
18461 // correct because all of our SRLs are followed immediately by a mask anyways
18462 // that handles any bits that sneak into the high bits of the byte elements.
18463 MVT SrlVT = Len > 8 ? VT : MVT::getVectorVT(MVT::i16, VecSize / 16);
18467 // v = v - ((v >> 1) & 0x55555555...)
18469 DAG.getBitcast(VT, GetShift(ISD::SRL, DAG.getBitcast(SrlVT, V), 1));
18470 SDValue And = GetMask(Srl, APInt::getSplat(Len, APInt(8, 0x55)));
18471 V = DAG.getNode(ISD::SUB, DL, VT, V, And);
18473 // v = (v & 0x33333333...) + ((v >> 2) & 0x33333333...)
18474 SDValue AndLHS = GetMask(V, APInt::getSplat(Len, APInt(8, 0x33)));
18475 Srl = DAG.getBitcast(VT, GetShift(ISD::SRL, DAG.getBitcast(SrlVT, V), 2));
18476 SDValue AndRHS = GetMask(Srl, APInt::getSplat(Len, APInt(8, 0x33)));
18477 V = DAG.getNode(ISD::ADD, DL, VT, AndLHS, AndRHS);
18479 // v = (v + (v >> 4)) & 0x0F0F0F0F...
18480 Srl = DAG.getBitcast(VT, GetShift(ISD::SRL, DAG.getBitcast(SrlVT, V), 4));
18481 SDValue Add = DAG.getNode(ISD::ADD, DL, VT, V, Srl);
18482 V = GetMask(Add, APInt::getSplat(Len, APInt(8, 0x0F)));
18484 // At this point, V contains the byte-wise population count, and we are
18485 // merely doing a horizontal sum if necessary to get the wider element
18487 if (EltVT == MVT::i8)
18490 return LowerHorizontalByteSum(
18491 DAG.getBitcast(MVT::getVectorVT(MVT::i8, VecSize / 8), V), VT, Subtarget,
18495 static SDValue LowerVectorCTPOP(SDValue Op, const X86Subtarget *Subtarget,
18496 SelectionDAG &DAG) {
18497 MVT VT = Op.getSimpleValueType();
18498 // FIXME: Need to add AVX-512 support here!
18499 assert((VT.is256BitVector() || VT.is128BitVector()) &&
18500 "Unknown CTPOP type to handle");
18501 SDLoc DL(Op.getNode());
18502 SDValue Op0 = Op.getOperand(0);
18504 if (!Subtarget->hasSSSE3()) {
18505 // We can't use the fast LUT approach, so fall back on vectorized bitmath.
18506 assert(VT.is128BitVector() && "Only 128-bit vectors supported in SSE!");
18507 return LowerVectorCTPOPBitmath(Op0, DL, Subtarget, DAG);
18510 if (VT.is256BitVector() && !Subtarget->hasInt256()) {
18511 unsigned NumElems = VT.getVectorNumElements();
18513 // Extract each 128-bit vector, compute pop count and concat the result.
18514 SDValue LHS = Extract128BitVector(Op0, 0, DAG, DL);
18515 SDValue RHS = Extract128BitVector(Op0, NumElems/2, DAG, DL);
18517 return DAG.getNode(ISD::CONCAT_VECTORS, DL, VT,
18518 LowerVectorCTPOPInRegLUT(LHS, DL, Subtarget, DAG),
18519 LowerVectorCTPOPInRegLUT(RHS, DL, Subtarget, DAG));
18522 return LowerVectorCTPOPInRegLUT(Op0, DL, Subtarget, DAG);
18525 static SDValue LowerCTPOP(SDValue Op, const X86Subtarget *Subtarget,
18526 SelectionDAG &DAG) {
18527 assert(Op.getValueType().isVector() &&
18528 "We only do custom lowering for vector population count.");
18529 return LowerVectorCTPOP(Op, Subtarget, DAG);
18532 static SDValue LowerLOAD_SUB(SDValue Op, SelectionDAG &DAG) {
18533 SDNode *Node = Op.getNode();
18535 EVT T = Node->getValueType(0);
18536 SDValue negOp = DAG.getNode(ISD::SUB, dl, T,
18537 DAG.getConstant(0, dl, T), Node->getOperand(2));
18538 return DAG.getAtomic(ISD::ATOMIC_LOAD_ADD, dl,
18539 cast<AtomicSDNode>(Node)->getMemoryVT(),
18540 Node->getOperand(0),
18541 Node->getOperand(1), negOp,
18542 cast<AtomicSDNode>(Node)->getMemOperand(),
18543 cast<AtomicSDNode>(Node)->getOrdering(),
18544 cast<AtomicSDNode>(Node)->getSynchScope());
18547 static SDValue LowerATOMIC_STORE(SDValue Op, SelectionDAG &DAG) {
18548 SDNode *Node = Op.getNode();
18550 EVT VT = cast<AtomicSDNode>(Node)->getMemoryVT();
18552 // Convert seq_cst store -> xchg
18553 // Convert wide store -> swap (-> cmpxchg8b/cmpxchg16b)
18554 // FIXME: On 32-bit, store -> fist or movq would be more efficient
18555 // (The only way to get a 16-byte store is cmpxchg16b)
18556 // FIXME: 16-byte ATOMIC_SWAP isn't actually hooked up at the moment.
18557 if (cast<AtomicSDNode>(Node)->getOrdering() == SequentiallyConsistent ||
18558 !DAG.getTargetLoweringInfo().isTypeLegal(VT)) {
18559 SDValue Swap = DAG.getAtomic(ISD::ATOMIC_SWAP, dl,
18560 cast<AtomicSDNode>(Node)->getMemoryVT(),
18561 Node->getOperand(0),
18562 Node->getOperand(1), Node->getOperand(2),
18563 cast<AtomicSDNode>(Node)->getMemOperand(),
18564 cast<AtomicSDNode>(Node)->getOrdering(),
18565 cast<AtomicSDNode>(Node)->getSynchScope());
18566 return Swap.getValue(1);
18568 // Other atomic stores have a simple pattern.
18572 static SDValue LowerADDC_ADDE_SUBC_SUBE(SDValue Op, SelectionDAG &DAG) {
18573 EVT VT = Op.getNode()->getSimpleValueType(0);
18575 // Let legalize expand this if it isn't a legal type yet.
18576 if (!DAG.getTargetLoweringInfo().isTypeLegal(VT))
18579 SDVTList VTs = DAG.getVTList(VT, MVT::i32);
18582 bool ExtraOp = false;
18583 switch (Op.getOpcode()) {
18584 default: llvm_unreachable("Invalid code");
18585 case ISD::ADDC: Opc = X86ISD::ADD; break;
18586 case ISD::ADDE: Opc = X86ISD::ADC; ExtraOp = true; break;
18587 case ISD::SUBC: Opc = X86ISD::SUB; break;
18588 case ISD::SUBE: Opc = X86ISD::SBB; ExtraOp = true; break;
18592 return DAG.getNode(Opc, SDLoc(Op), VTs, Op.getOperand(0),
18594 return DAG.getNode(Opc, SDLoc(Op), VTs, Op.getOperand(0),
18595 Op.getOperand(1), Op.getOperand(2));
18598 static SDValue LowerFSINCOS(SDValue Op, const X86Subtarget *Subtarget,
18599 SelectionDAG &DAG) {
18600 assert(Subtarget->isTargetDarwin() && Subtarget->is64Bit());
18602 // For MacOSX, we want to call an alternative entry point: __sincos_stret,
18603 // which returns the values as { float, float } (in XMM0) or
18604 // { double, double } (which is returned in XMM0, XMM1).
18606 SDValue Arg = Op.getOperand(0);
18607 EVT ArgVT = Arg.getValueType();
18608 Type *ArgTy = ArgVT.getTypeForEVT(*DAG.getContext());
18610 TargetLowering::ArgListTy Args;
18611 TargetLowering::ArgListEntry Entry;
18615 Entry.isSExt = false;
18616 Entry.isZExt = false;
18617 Args.push_back(Entry);
18619 bool isF64 = ArgVT == MVT::f64;
18620 // Only optimize x86_64 for now. i386 is a bit messy. For f32,
18621 // the small struct {f32, f32} is returned in (eax, edx). For f64,
18622 // the results are returned via SRet in memory.
18623 const char *LibcallName = isF64 ? "__sincos_stret" : "__sincosf_stret";
18624 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
18626 DAG.getExternalSymbol(LibcallName, TLI.getPointerTy(DAG.getDataLayout()));
18628 Type *RetTy = isF64
18629 ? (Type*)StructType::get(ArgTy, ArgTy, nullptr)
18630 : (Type*)VectorType::get(ArgTy, 4);
18632 TargetLowering::CallLoweringInfo CLI(DAG);
18633 CLI.setDebugLoc(dl).setChain(DAG.getEntryNode())
18634 .setCallee(CallingConv::C, RetTy, Callee, std::move(Args), 0);
18636 std::pair<SDValue, SDValue> CallResult = TLI.LowerCallTo(CLI);
18639 // Returned in xmm0 and xmm1.
18640 return CallResult.first;
18642 // Returned in bits 0:31 and 32:64 xmm0.
18643 SDValue SinVal = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, ArgVT,
18644 CallResult.first, DAG.getIntPtrConstant(0, dl));
18645 SDValue CosVal = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, ArgVT,
18646 CallResult.first, DAG.getIntPtrConstant(1, dl));
18647 SDVTList Tys = DAG.getVTList(ArgVT, ArgVT);
18648 return DAG.getNode(ISD::MERGE_VALUES, dl, Tys, SinVal, CosVal);
18651 static SDValue LowerMSCATTER(SDValue Op, const X86Subtarget *Subtarget,
18652 SelectionDAG &DAG) {
18653 assert(Subtarget->hasAVX512() &&
18654 "MGATHER/MSCATTER are supported on AVX-512 arch only");
18656 MaskedScatterSDNode *N = cast<MaskedScatterSDNode>(Op.getNode());
18657 EVT VT = N->getValue().getValueType();
18658 assert(VT.getScalarSizeInBits() >= 32 && "Unsupported scatter op");
18661 // X86 scatter kills mask register, so its type should be added to
18662 // the list of return values
18663 if (N->getNumValues() == 1) {
18664 SDValue Index = N->getIndex();
18665 if (!Subtarget->hasVLX() && !VT.is512BitVector() &&
18666 !Index.getValueType().is512BitVector())
18667 Index = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v8i64, Index);
18669 SDVTList VTs = DAG.getVTList(N->getMask().getValueType(), MVT::Other);
18670 SDValue Ops[] = { N->getOperand(0), N->getOperand(1), N->getOperand(2),
18671 N->getOperand(3), Index };
18673 SDValue NewScatter = DAG.getMaskedScatter(VTs, VT, dl, Ops, N->getMemOperand());
18674 DAG.ReplaceAllUsesWith(Op, SDValue(NewScatter.getNode(), 1));
18675 return SDValue(NewScatter.getNode(), 0);
18680 static SDValue LowerMGATHER(SDValue Op, const X86Subtarget *Subtarget,
18681 SelectionDAG &DAG) {
18682 assert(Subtarget->hasAVX512() &&
18683 "MGATHER/MSCATTER are supported on AVX-512 arch only");
18685 MaskedGatherSDNode *N = cast<MaskedGatherSDNode>(Op.getNode());
18686 EVT VT = Op.getValueType();
18687 assert(VT.getScalarSizeInBits() >= 32 && "Unsupported gather op");
18690 SDValue Index = N->getIndex();
18691 if (!Subtarget->hasVLX() && !VT.is512BitVector() &&
18692 !Index.getValueType().is512BitVector()) {
18693 Index = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v8i64, Index);
18694 SDValue Ops[] = { N->getOperand(0), N->getOperand(1), N->getOperand(2),
18695 N->getOperand(3), Index };
18696 DAG.UpdateNodeOperands(N, Ops);
18701 SDValue X86TargetLowering::LowerGC_TRANSITION_START(SDValue Op,
18702 SelectionDAG &DAG) const {
18703 // TODO: Eventually, the lowering of these nodes should be informed by or
18704 // deferred to the GC strategy for the function in which they appear. For
18705 // now, however, they must be lowered to something. Since they are logically
18706 // no-ops in the case of a null GC strategy (or a GC strategy which does not
18707 // require special handling for these nodes), lower them as literal NOOPs for
18709 SmallVector<SDValue, 2> Ops;
18711 Ops.push_back(Op.getOperand(0));
18712 if (Op->getGluedNode())
18713 Ops.push_back(Op->getOperand(Op->getNumOperands() - 1));
18716 SDVTList VTs = DAG.getVTList(MVT::Other, MVT::Glue);
18717 SDValue NOOP(DAG.getMachineNode(X86::NOOP, SDLoc(Op), VTs, Ops), 0);
18722 SDValue X86TargetLowering::LowerGC_TRANSITION_END(SDValue Op,
18723 SelectionDAG &DAG) const {
18724 // TODO: Eventually, the lowering of these nodes should be informed by or
18725 // deferred to the GC strategy for the function in which they appear. For
18726 // now, however, they must be lowered to something. Since they are logically
18727 // no-ops in the case of a null GC strategy (or a GC strategy which does not
18728 // require special handling for these nodes), lower them as literal NOOPs for
18730 SmallVector<SDValue, 2> Ops;
18732 Ops.push_back(Op.getOperand(0));
18733 if (Op->getGluedNode())
18734 Ops.push_back(Op->getOperand(Op->getNumOperands() - 1));
18737 SDVTList VTs = DAG.getVTList(MVT::Other, MVT::Glue);
18738 SDValue NOOP(DAG.getMachineNode(X86::NOOP, SDLoc(Op), VTs, Ops), 0);
18743 /// LowerOperation - Provide custom lowering hooks for some operations.
18745 SDValue X86TargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
18746 switch (Op.getOpcode()) {
18747 default: llvm_unreachable("Should not custom lower this!");
18748 case ISD::ATOMIC_FENCE: return LowerATOMIC_FENCE(Op, Subtarget, DAG);
18749 case ISD::ATOMIC_CMP_SWAP_WITH_SUCCESS:
18750 return LowerCMP_SWAP(Op, Subtarget, DAG);
18751 case ISD::CTPOP: return LowerCTPOP(Op, Subtarget, DAG);
18752 case ISD::ATOMIC_LOAD_SUB: return LowerLOAD_SUB(Op,DAG);
18753 case ISD::ATOMIC_STORE: return LowerATOMIC_STORE(Op,DAG);
18754 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG);
18755 case ISD::CONCAT_VECTORS: return LowerCONCAT_VECTORS(Op, Subtarget, DAG);
18756 case ISD::VECTOR_SHUFFLE: return lowerVectorShuffle(Op, Subtarget, DAG);
18757 case ISD::VSELECT: return LowerVSELECT(Op, DAG);
18758 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
18759 case ISD::INSERT_VECTOR_ELT: return LowerINSERT_VECTOR_ELT(Op, DAG);
18760 case ISD::EXTRACT_SUBVECTOR: return LowerEXTRACT_SUBVECTOR(Op,Subtarget,DAG);
18761 case ISD::INSERT_SUBVECTOR: return LowerINSERT_SUBVECTOR(Op, Subtarget,DAG);
18762 case ISD::SCALAR_TO_VECTOR: return LowerSCALAR_TO_VECTOR(Op, DAG);
18763 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
18764 case ISD::GlobalAddress: return LowerGlobalAddress(Op, DAG);
18765 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
18766 case ISD::ExternalSymbol: return LowerExternalSymbol(Op, DAG);
18767 case ISD::BlockAddress: return LowerBlockAddress(Op, DAG);
18768 case ISD::SHL_PARTS:
18769 case ISD::SRA_PARTS:
18770 case ISD::SRL_PARTS: return LowerShiftParts(Op, DAG);
18771 case ISD::SINT_TO_FP: return LowerSINT_TO_FP(Op, DAG);
18772 case ISD::UINT_TO_FP: return LowerUINT_TO_FP(Op, DAG);
18773 case ISD::TRUNCATE: return LowerTRUNCATE(Op, DAG);
18774 case ISD::ZERO_EXTEND: return LowerZERO_EXTEND(Op, Subtarget, DAG);
18775 case ISD::SIGN_EXTEND: return LowerSIGN_EXTEND(Op, Subtarget, DAG);
18776 case ISD::ANY_EXTEND: return LowerANY_EXTEND(Op, Subtarget, DAG);
18777 case ISD::SIGN_EXTEND_VECTOR_INREG:
18778 return LowerSIGN_EXTEND_VECTOR_INREG(Op, Subtarget, DAG);
18779 case ISD::FP_TO_SINT: return LowerFP_TO_SINT(Op, DAG);
18780 case ISD::FP_TO_UINT: return LowerFP_TO_UINT(Op, DAG);
18781 case ISD::FP_EXTEND: return LowerFP_EXTEND(Op, DAG);
18782 case ISD::LOAD: return LowerExtendedLoad(Op, Subtarget, DAG);
18784 case ISD::FNEG: return LowerFABSorFNEG(Op, DAG);
18785 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
18786 case ISD::FGETSIGN: return LowerFGETSIGN(Op, DAG);
18787 case ISD::SETCC: return LowerSETCC(Op, DAG);
18788 case ISD::SELECT: return LowerSELECT(Op, DAG);
18789 case ISD::BRCOND: return LowerBRCOND(Op, DAG);
18790 case ISD::JumpTable: return LowerJumpTable(Op, DAG);
18791 case ISD::VASTART: return LowerVASTART(Op, DAG);
18792 case ISD::VAARG: return LowerVAARG(Op, DAG);
18793 case ISD::VACOPY: return LowerVACOPY(Op, Subtarget, DAG);
18794 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, Subtarget, DAG);
18795 case ISD::INTRINSIC_VOID:
18796 case ISD::INTRINSIC_W_CHAIN: return LowerINTRINSIC_W_CHAIN(Op, Subtarget, DAG);
18797 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
18798 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
18799 case ISD::FRAME_TO_ARGS_OFFSET:
18800 return LowerFRAME_TO_ARGS_OFFSET(Op, DAG);
18801 case ISD::DYNAMIC_STACKALLOC: return LowerDYNAMIC_STACKALLOC(Op, DAG);
18802 case ISD::EH_RETURN: return LowerEH_RETURN(Op, DAG);
18803 case ISD::EH_SJLJ_SETJMP: return lowerEH_SJLJ_SETJMP(Op, DAG);
18804 case ISD::EH_SJLJ_LONGJMP: return lowerEH_SJLJ_LONGJMP(Op, DAG);
18805 case ISD::INIT_TRAMPOLINE: return LowerINIT_TRAMPOLINE(Op, DAG);
18806 case ISD::ADJUST_TRAMPOLINE: return LowerADJUST_TRAMPOLINE(Op, DAG);
18807 case ISD::FLT_ROUNDS_: return LowerFLT_ROUNDS_(Op, DAG);
18808 case ISD::CTLZ: return LowerCTLZ(Op, DAG);
18809 case ISD::CTLZ_ZERO_UNDEF: return LowerCTLZ_ZERO_UNDEF(Op, DAG);
18810 case ISD::CTTZ: return LowerCTTZ(Op, DAG);
18811 case ISD::MUL: return LowerMUL(Op, Subtarget, DAG);
18812 case ISD::UMUL_LOHI:
18813 case ISD::SMUL_LOHI: return LowerMUL_LOHI(Op, Subtarget, DAG);
18816 case ISD::SHL: return LowerShift(Op, Subtarget, DAG);
18822 case ISD::UMULO: return LowerXALUO(Op, DAG);
18823 case ISD::READCYCLECOUNTER: return LowerREADCYCLECOUNTER(Op, Subtarget,DAG);
18824 case ISD::BITCAST: return LowerBITCAST(Op, Subtarget, DAG);
18828 case ISD::SUBE: return LowerADDC_ADDE_SUBC_SUBE(Op, DAG);
18829 case ISD::ADD: return LowerADD(Op, DAG);
18830 case ISD::SUB: return LowerSUB(Op, DAG);
18834 case ISD::UMIN: return LowerMINMAX(Op, DAG);
18835 case ISD::FSINCOS: return LowerFSINCOS(Op, Subtarget, DAG);
18836 case ISD::MGATHER: return LowerMGATHER(Op, Subtarget, DAG);
18837 case ISD::MSCATTER: return LowerMSCATTER(Op, Subtarget, DAG);
18838 case ISD::GC_TRANSITION_START:
18839 return LowerGC_TRANSITION_START(Op, DAG);
18840 case ISD::GC_TRANSITION_END: return LowerGC_TRANSITION_END(Op, DAG);
18844 /// ReplaceNodeResults - Replace a node with an illegal result type
18845 /// with a new node built out of custom code.
18846 void X86TargetLowering::ReplaceNodeResults(SDNode *N,
18847 SmallVectorImpl<SDValue>&Results,
18848 SelectionDAG &DAG) const {
18850 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
18851 switch (N->getOpcode()) {
18853 llvm_unreachable("Do not know how to custom type legalize this operation!");
18854 // We might have generated v2f32 FMIN/FMAX operations. Widen them to v4f32.
18855 case X86ISD::FMINC:
18857 case X86ISD::FMAXC:
18858 case X86ISD::FMAX: {
18859 EVT VT = N->getValueType(0);
18860 if (VT != MVT::v2f32)
18861 llvm_unreachable("Unexpected type (!= v2f32) on FMIN/FMAX.");
18862 SDValue UNDEF = DAG.getUNDEF(VT);
18863 SDValue LHS = DAG.getNode(ISD::CONCAT_VECTORS, dl, MVT::v4f32,
18864 N->getOperand(0), UNDEF);
18865 SDValue RHS = DAG.getNode(ISD::CONCAT_VECTORS, dl, MVT::v4f32,
18866 N->getOperand(1), UNDEF);
18867 Results.push_back(DAG.getNode(N->getOpcode(), dl, MVT::v4f32, LHS, RHS));
18870 case ISD::SIGN_EXTEND_INREG:
18875 // We don't want to expand or promote these.
18882 case ISD::UDIVREM: {
18883 SDValue V = LowerWin64_i128OP(SDValue(N,0), DAG);
18884 Results.push_back(V);
18887 case ISD::FP_TO_SINT:
18888 // FP_TO_INT*_IN_MEM is not legal for f16 inputs. Do not convert
18889 // (FP_TO_SINT (load f16)) to FP_TO_INT*.
18890 if (N->getOperand(0).getValueType() == MVT::f16)
18893 case ISD::FP_TO_UINT: {
18894 bool IsSigned = N->getOpcode() == ISD::FP_TO_SINT;
18896 if (!IsSigned && !isIntegerTypeFTOL(SDValue(N, 0).getValueType()))
18899 std::pair<SDValue,SDValue> Vals =
18900 FP_TO_INTHelper(SDValue(N, 0), DAG, IsSigned, /*IsReplace=*/ true);
18901 SDValue FIST = Vals.first, StackSlot = Vals.second;
18902 if (FIST.getNode()) {
18903 EVT VT = N->getValueType(0);
18904 // Return a load from the stack slot.
18905 if (StackSlot.getNode())
18906 Results.push_back(DAG.getLoad(VT, dl, FIST, StackSlot,
18907 MachinePointerInfo(),
18908 false, false, false, 0));
18910 Results.push_back(FIST);
18914 case ISD::UINT_TO_FP: {
18915 assert(Subtarget->hasSSE2() && "Requires at least SSE2!");
18916 if (N->getOperand(0).getValueType() != MVT::v2i32 ||
18917 N->getValueType(0) != MVT::v2f32)
18919 SDValue ZExtIn = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v2i64,
18921 SDValue Bias = DAG.getConstantFP(BitsToDouble(0x4330000000000000ULL), dl,
18923 SDValue VBias = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v2f64, Bias, Bias);
18924 SDValue Or = DAG.getNode(ISD::OR, dl, MVT::v2i64, ZExtIn,
18925 DAG.getBitcast(MVT::v2i64, VBias));
18926 Or = DAG.getBitcast(MVT::v2f64, Or);
18927 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::v2f64, Or, VBias);
18928 Results.push_back(DAG.getNode(X86ISD::VFPROUND, dl, MVT::v4f32, Sub));
18931 case ISD::FP_ROUND: {
18932 if (!TLI.isTypeLegal(N->getOperand(0).getValueType()))
18934 SDValue V = DAG.getNode(X86ISD::VFPROUND, dl, MVT::v4f32, N->getOperand(0));
18935 Results.push_back(V);
18938 case ISD::FP_EXTEND: {
18939 // Right now, only MVT::v2f32 has OperationAction for FP_EXTEND.
18940 // No other ValueType for FP_EXTEND should reach this point.
18941 assert(N->getValueType(0) == MVT::v2f32 &&
18942 "Do not know how to legalize this Node");
18945 case ISD::INTRINSIC_W_CHAIN: {
18946 unsigned IntNo = cast<ConstantSDNode>(N->getOperand(1))->getZExtValue();
18948 default : llvm_unreachable("Do not know how to custom type "
18949 "legalize this intrinsic operation!");
18950 case Intrinsic::x86_rdtsc:
18951 return getReadTimeStampCounter(N, dl, X86ISD::RDTSC_DAG, DAG, Subtarget,
18953 case Intrinsic::x86_rdtscp:
18954 return getReadTimeStampCounter(N, dl, X86ISD::RDTSCP_DAG, DAG, Subtarget,
18956 case Intrinsic::x86_rdpmc:
18957 return getReadPerformanceCounter(N, dl, DAG, Subtarget, Results);
18960 case ISD::READCYCLECOUNTER: {
18961 return getReadTimeStampCounter(N, dl, X86ISD::RDTSC_DAG, DAG, Subtarget,
18964 case ISD::ATOMIC_CMP_SWAP_WITH_SUCCESS: {
18965 EVT T = N->getValueType(0);
18966 assert((T == MVT::i64 || T == MVT::i128) && "can only expand cmpxchg pair");
18967 bool Regs64bit = T == MVT::i128;
18968 EVT HalfT = Regs64bit ? MVT::i64 : MVT::i32;
18969 SDValue cpInL, cpInH;
18970 cpInL = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, HalfT, N->getOperand(2),
18971 DAG.getConstant(0, dl, HalfT));
18972 cpInH = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, HalfT, N->getOperand(2),
18973 DAG.getConstant(1, dl, HalfT));
18974 cpInL = DAG.getCopyToReg(N->getOperand(0), dl,
18975 Regs64bit ? X86::RAX : X86::EAX,
18977 cpInH = DAG.getCopyToReg(cpInL.getValue(0), dl,
18978 Regs64bit ? X86::RDX : X86::EDX,
18979 cpInH, cpInL.getValue(1));
18980 SDValue swapInL, swapInH;
18981 swapInL = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, HalfT, N->getOperand(3),
18982 DAG.getConstant(0, dl, HalfT));
18983 swapInH = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, HalfT, N->getOperand(3),
18984 DAG.getConstant(1, dl, HalfT));
18985 swapInL = DAG.getCopyToReg(cpInH.getValue(0), dl,
18986 Regs64bit ? X86::RBX : X86::EBX,
18987 swapInL, cpInH.getValue(1));
18988 swapInH = DAG.getCopyToReg(swapInL.getValue(0), dl,
18989 Regs64bit ? X86::RCX : X86::ECX,
18990 swapInH, swapInL.getValue(1));
18991 SDValue Ops[] = { swapInH.getValue(0),
18993 swapInH.getValue(1) };
18994 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Glue);
18995 MachineMemOperand *MMO = cast<AtomicSDNode>(N)->getMemOperand();
18996 unsigned Opcode = Regs64bit ? X86ISD::LCMPXCHG16_DAG :
18997 X86ISD::LCMPXCHG8_DAG;
18998 SDValue Result = DAG.getMemIntrinsicNode(Opcode, dl, Tys, Ops, T, MMO);
18999 SDValue cpOutL = DAG.getCopyFromReg(Result.getValue(0), dl,
19000 Regs64bit ? X86::RAX : X86::EAX,
19001 HalfT, Result.getValue(1));
19002 SDValue cpOutH = DAG.getCopyFromReg(cpOutL.getValue(1), dl,
19003 Regs64bit ? X86::RDX : X86::EDX,
19004 HalfT, cpOutL.getValue(2));
19005 SDValue OpsF[] = { cpOutL.getValue(0), cpOutH.getValue(0)};
19007 SDValue EFLAGS = DAG.getCopyFromReg(cpOutH.getValue(1), dl, X86::EFLAGS,
19008 MVT::i32, cpOutH.getValue(2));
19010 DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
19011 DAG.getConstant(X86::COND_E, dl, MVT::i8), EFLAGS);
19012 Success = DAG.getZExtOrTrunc(Success, dl, N->getValueType(1));
19014 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, T, OpsF));
19015 Results.push_back(Success);
19016 Results.push_back(EFLAGS.getValue(1));
19019 case ISD::ATOMIC_SWAP:
19020 case ISD::ATOMIC_LOAD_ADD:
19021 case ISD::ATOMIC_LOAD_SUB:
19022 case ISD::ATOMIC_LOAD_AND:
19023 case ISD::ATOMIC_LOAD_OR:
19024 case ISD::ATOMIC_LOAD_XOR:
19025 case ISD::ATOMIC_LOAD_NAND:
19026 case ISD::ATOMIC_LOAD_MIN:
19027 case ISD::ATOMIC_LOAD_MAX:
19028 case ISD::ATOMIC_LOAD_UMIN:
19029 case ISD::ATOMIC_LOAD_UMAX:
19030 case ISD::ATOMIC_LOAD: {
19031 // Delegate to generic TypeLegalization. Situations we can really handle
19032 // should have already been dealt with by AtomicExpandPass.cpp.
19035 case ISD::BITCAST: {
19036 assert(Subtarget->hasSSE2() && "Requires at least SSE2!");
19037 EVT DstVT = N->getValueType(0);
19038 EVT SrcVT = N->getOperand(0)->getValueType(0);
19040 if (SrcVT != MVT::f64 ||
19041 (DstVT != MVT::v2i32 && DstVT != MVT::v4i16 && DstVT != MVT::v8i8))
19044 unsigned NumElts = DstVT.getVectorNumElements();
19045 EVT SVT = DstVT.getVectorElementType();
19046 EVT WiderVT = EVT::getVectorVT(*DAG.getContext(), SVT, NumElts * 2);
19047 SDValue Expanded = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
19048 MVT::v2f64, N->getOperand(0));
19049 SDValue ToVecInt = DAG.getBitcast(WiderVT, Expanded);
19051 if (ExperimentalVectorWideningLegalization) {
19052 // If we are legalizing vectors by widening, we already have the desired
19053 // legal vector type, just return it.
19054 Results.push_back(ToVecInt);
19058 SmallVector<SDValue, 8> Elts;
19059 for (unsigned i = 0, e = NumElts; i != e; ++i)
19060 Elts.push_back(DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, SVT,
19061 ToVecInt, DAG.getIntPtrConstant(i, dl)));
19063 Results.push_back(DAG.getNode(ISD::BUILD_VECTOR, dl, DstVT, Elts));
19068 const char *X86TargetLowering::getTargetNodeName(unsigned Opcode) const {
19069 switch ((X86ISD::NodeType)Opcode) {
19070 case X86ISD::FIRST_NUMBER: break;
19071 case X86ISD::BSF: return "X86ISD::BSF";
19072 case X86ISD::BSR: return "X86ISD::BSR";
19073 case X86ISD::SHLD: return "X86ISD::SHLD";
19074 case X86ISD::SHRD: return "X86ISD::SHRD";
19075 case X86ISD::FAND: return "X86ISD::FAND";
19076 case X86ISD::FANDN: return "X86ISD::FANDN";
19077 case X86ISD::FOR: return "X86ISD::FOR";
19078 case X86ISD::FXOR: return "X86ISD::FXOR";
19079 case X86ISD::FILD: return "X86ISD::FILD";
19080 case X86ISD::FILD_FLAG: return "X86ISD::FILD_FLAG";
19081 case X86ISD::FP_TO_INT16_IN_MEM: return "X86ISD::FP_TO_INT16_IN_MEM";
19082 case X86ISD::FP_TO_INT32_IN_MEM: return "X86ISD::FP_TO_INT32_IN_MEM";
19083 case X86ISD::FP_TO_INT64_IN_MEM: return "X86ISD::FP_TO_INT64_IN_MEM";
19084 case X86ISD::FLD: return "X86ISD::FLD";
19085 case X86ISD::FST: return "X86ISD::FST";
19086 case X86ISD::CALL: return "X86ISD::CALL";
19087 case X86ISD::RDTSC_DAG: return "X86ISD::RDTSC_DAG";
19088 case X86ISD::RDTSCP_DAG: return "X86ISD::RDTSCP_DAG";
19089 case X86ISD::RDPMC_DAG: return "X86ISD::RDPMC_DAG";
19090 case X86ISD::BT: return "X86ISD::BT";
19091 case X86ISD::CMP: return "X86ISD::CMP";
19092 case X86ISD::COMI: return "X86ISD::COMI";
19093 case X86ISD::UCOMI: return "X86ISD::UCOMI";
19094 case X86ISD::CMPM: return "X86ISD::CMPM";
19095 case X86ISD::CMPMU: return "X86ISD::CMPMU";
19096 case X86ISD::CMPM_RND: return "X86ISD::CMPM_RND";
19097 case X86ISD::SETCC: return "X86ISD::SETCC";
19098 case X86ISD::SETCC_CARRY: return "X86ISD::SETCC_CARRY";
19099 case X86ISD::FSETCC: return "X86ISD::FSETCC";
19100 case X86ISD::FGETSIGNx86: return "X86ISD::FGETSIGNx86";
19101 case X86ISD::CMOV: return "X86ISD::CMOV";
19102 case X86ISD::BRCOND: return "X86ISD::BRCOND";
19103 case X86ISD::RET_FLAG: return "X86ISD::RET_FLAG";
19104 case X86ISD::REP_STOS: return "X86ISD::REP_STOS";
19105 case X86ISD::REP_MOVS: return "X86ISD::REP_MOVS";
19106 case X86ISD::GlobalBaseReg: return "X86ISD::GlobalBaseReg";
19107 case X86ISD::Wrapper: return "X86ISD::Wrapper";
19108 case X86ISD::WrapperRIP: return "X86ISD::WrapperRIP";
19109 case X86ISD::MOVDQ2Q: return "X86ISD::MOVDQ2Q";
19110 case X86ISD::MMX_MOVD2W: return "X86ISD::MMX_MOVD2W";
19111 case X86ISD::MMX_MOVW2D: return "X86ISD::MMX_MOVW2D";
19112 case X86ISD::PEXTRB: return "X86ISD::PEXTRB";
19113 case X86ISD::PEXTRW: return "X86ISD::PEXTRW";
19114 case X86ISD::INSERTPS: return "X86ISD::INSERTPS";
19115 case X86ISD::PINSRB: return "X86ISD::PINSRB";
19116 case X86ISD::PINSRW: return "X86ISD::PINSRW";
19117 case X86ISD::MMX_PINSRW: return "X86ISD::MMX_PINSRW";
19118 case X86ISD::PSHUFB: return "X86ISD::PSHUFB";
19119 case X86ISD::ANDNP: return "X86ISD::ANDNP";
19120 case X86ISD::PSIGN: return "X86ISD::PSIGN";
19121 case X86ISD::BLENDI: return "X86ISD::BLENDI";
19122 case X86ISD::SHRUNKBLEND: return "X86ISD::SHRUNKBLEND";
19123 case X86ISD::ADDUS: return "X86ISD::ADDUS";
19124 case X86ISD::SUBUS: return "X86ISD::SUBUS";
19125 case X86ISD::HADD: return "X86ISD::HADD";
19126 case X86ISD::HSUB: return "X86ISD::HSUB";
19127 case X86ISD::FHADD: return "X86ISD::FHADD";
19128 case X86ISD::FHSUB: return "X86ISD::FHSUB";
19129 case X86ISD::ABS: return "X86ISD::ABS";
19130 case X86ISD::FMAX: return "X86ISD::FMAX";
19131 case X86ISD::FMAX_RND: return "X86ISD::FMAX_RND";
19132 case X86ISD::FMIN: return "X86ISD::FMIN";
19133 case X86ISD::FMIN_RND: return "X86ISD::FMIN_RND";
19134 case X86ISD::FMAXC: return "X86ISD::FMAXC";
19135 case X86ISD::FMINC: return "X86ISD::FMINC";
19136 case X86ISD::FRSQRT: return "X86ISD::FRSQRT";
19137 case X86ISD::FRCP: return "X86ISD::FRCP";
19138 case X86ISD::EXTRQI: return "X86ISD::EXTRQI";
19139 case X86ISD::INSERTQI: return "X86ISD::INSERTQI";
19140 case X86ISD::TLSADDR: return "X86ISD::TLSADDR";
19141 case X86ISD::TLSBASEADDR: return "X86ISD::TLSBASEADDR";
19142 case X86ISD::TLSCALL: return "X86ISD::TLSCALL";
19143 case X86ISD::EH_SJLJ_SETJMP: return "X86ISD::EH_SJLJ_SETJMP";
19144 case X86ISD::EH_SJLJ_LONGJMP: return "X86ISD::EH_SJLJ_LONGJMP";
19145 case X86ISD::EH_RETURN: return "X86ISD::EH_RETURN";
19146 case X86ISD::TC_RETURN: return "X86ISD::TC_RETURN";
19147 case X86ISD::FNSTCW16m: return "X86ISD::FNSTCW16m";
19148 case X86ISD::FNSTSW16r: return "X86ISD::FNSTSW16r";
19149 case X86ISD::LCMPXCHG_DAG: return "X86ISD::LCMPXCHG_DAG";
19150 case X86ISD::LCMPXCHG8_DAG: return "X86ISD::LCMPXCHG8_DAG";
19151 case X86ISD::LCMPXCHG16_DAG: return "X86ISD::LCMPXCHG16_DAG";
19152 case X86ISD::VZEXT_MOVL: return "X86ISD::VZEXT_MOVL";
19153 case X86ISD::VZEXT_LOAD: return "X86ISD::VZEXT_LOAD";
19154 case X86ISD::VZEXT: return "X86ISD::VZEXT";
19155 case X86ISD::VSEXT: return "X86ISD::VSEXT";
19156 case X86ISD::VTRUNC: return "X86ISD::VTRUNC";
19157 case X86ISD::VTRUNCS: return "X86ISD::VTRUNCS";
19158 case X86ISD::VTRUNCUS: return "X86ISD::VTRUNCUS";
19159 case X86ISD::VINSERT: return "X86ISD::VINSERT";
19160 case X86ISD::VFPEXT: return "X86ISD::VFPEXT";
19161 case X86ISD::VFPROUND: return "X86ISD::VFPROUND";
19162 case X86ISD::CVTDQ2PD: return "X86ISD::CVTDQ2PD";
19163 case X86ISD::CVTUDQ2PD: return "X86ISD::CVTUDQ2PD";
19164 case X86ISD::VSHLDQ: return "X86ISD::VSHLDQ";
19165 case X86ISD::VSRLDQ: return "X86ISD::VSRLDQ";
19166 case X86ISD::VSHL: return "X86ISD::VSHL";
19167 case X86ISD::VSRL: return "X86ISD::VSRL";
19168 case X86ISD::VSRA: return "X86ISD::VSRA";
19169 case X86ISD::VSHLI: return "X86ISD::VSHLI";
19170 case X86ISD::VSRLI: return "X86ISD::VSRLI";
19171 case X86ISD::VSRAI: return "X86ISD::VSRAI";
19172 case X86ISD::CMPP: return "X86ISD::CMPP";
19173 case X86ISD::PCMPEQ: return "X86ISD::PCMPEQ";
19174 case X86ISD::PCMPGT: return "X86ISD::PCMPGT";
19175 case X86ISD::PCMPEQM: return "X86ISD::PCMPEQM";
19176 case X86ISD::PCMPGTM: return "X86ISD::PCMPGTM";
19177 case X86ISD::ADD: return "X86ISD::ADD";
19178 case X86ISD::SUB: return "X86ISD::SUB";
19179 case X86ISD::ADC: return "X86ISD::ADC";
19180 case X86ISD::SBB: return "X86ISD::SBB";
19181 case X86ISD::SMUL: return "X86ISD::SMUL";
19182 case X86ISD::UMUL: return "X86ISD::UMUL";
19183 case X86ISD::SMUL8: return "X86ISD::SMUL8";
19184 case X86ISD::UMUL8: return "X86ISD::UMUL8";
19185 case X86ISD::SDIVREM8_SEXT_HREG: return "X86ISD::SDIVREM8_SEXT_HREG";
19186 case X86ISD::UDIVREM8_ZEXT_HREG: return "X86ISD::UDIVREM8_ZEXT_HREG";
19187 case X86ISD::INC: return "X86ISD::INC";
19188 case X86ISD::DEC: return "X86ISD::DEC";
19189 case X86ISD::OR: return "X86ISD::OR";
19190 case X86ISD::XOR: return "X86ISD::XOR";
19191 case X86ISD::AND: return "X86ISD::AND";
19192 case X86ISD::BEXTR: return "X86ISD::BEXTR";
19193 case X86ISD::MUL_IMM: return "X86ISD::MUL_IMM";
19194 case X86ISD::PTEST: return "X86ISD::PTEST";
19195 case X86ISD::TESTP: return "X86ISD::TESTP";
19196 case X86ISD::TESTM: return "X86ISD::TESTM";
19197 case X86ISD::TESTNM: return "X86ISD::TESTNM";
19198 case X86ISD::KORTEST: return "X86ISD::KORTEST";
19199 case X86ISD::PACKSS: return "X86ISD::PACKSS";
19200 case X86ISD::PACKUS: return "X86ISD::PACKUS";
19201 case X86ISD::PALIGNR: return "X86ISD::PALIGNR";
19202 case X86ISD::VALIGN: return "X86ISD::VALIGN";
19203 case X86ISD::PSHUFD: return "X86ISD::PSHUFD";
19204 case X86ISD::PSHUFHW: return "X86ISD::PSHUFHW";
19205 case X86ISD::PSHUFLW: return "X86ISD::PSHUFLW";
19206 case X86ISD::SHUFP: return "X86ISD::SHUFP";
19207 case X86ISD::SHUF128: return "X86ISD::SHUF128";
19208 case X86ISD::MOVLHPS: return "X86ISD::MOVLHPS";
19209 case X86ISD::MOVLHPD: return "X86ISD::MOVLHPD";
19210 case X86ISD::MOVHLPS: return "X86ISD::MOVHLPS";
19211 case X86ISD::MOVLPS: return "X86ISD::MOVLPS";
19212 case X86ISD::MOVLPD: return "X86ISD::MOVLPD";
19213 case X86ISD::MOVDDUP: return "X86ISD::MOVDDUP";
19214 case X86ISD::MOVSHDUP: return "X86ISD::MOVSHDUP";
19215 case X86ISD::MOVSLDUP: return "X86ISD::MOVSLDUP";
19216 case X86ISD::MOVSD: return "X86ISD::MOVSD";
19217 case X86ISD::MOVSS: return "X86ISD::MOVSS";
19218 case X86ISD::UNPCKL: return "X86ISD::UNPCKL";
19219 case X86ISD::UNPCKH: return "X86ISD::UNPCKH";
19220 case X86ISD::VBROADCAST: return "X86ISD::VBROADCAST";
19221 case X86ISD::SUBV_BROADCAST: return "X86ISD::SUBV_BROADCAST";
19222 case X86ISD::VEXTRACT: return "X86ISD::VEXTRACT";
19223 case X86ISD::VPERMILPV: return "X86ISD::VPERMILPV";
19224 case X86ISD::VPERMILPI: return "X86ISD::VPERMILPI";
19225 case X86ISD::VPERM2X128: return "X86ISD::VPERM2X128";
19226 case X86ISD::VPERMV: return "X86ISD::VPERMV";
19227 case X86ISD::VPERMV3: return "X86ISD::VPERMV3";
19228 case X86ISD::VPERMIV3: return "X86ISD::VPERMIV3";
19229 case X86ISD::VPERMI: return "X86ISD::VPERMI";
19230 case X86ISD::VFIXUPIMM: return "X86ISD::VFIXUPIMM";
19231 case X86ISD::VRANGE: return "X86ISD::VRANGE";
19232 case X86ISD::PMULUDQ: return "X86ISD::PMULUDQ";
19233 case X86ISD::PMULDQ: return "X86ISD::PMULDQ";
19234 case X86ISD::PSADBW: return "X86ISD::PSADBW";
19235 case X86ISD::VASTART_SAVE_XMM_REGS: return "X86ISD::VASTART_SAVE_XMM_REGS";
19236 case X86ISD::VAARG_64: return "X86ISD::VAARG_64";
19237 case X86ISD::WIN_ALLOCA: return "X86ISD::WIN_ALLOCA";
19238 case X86ISD::MEMBARRIER: return "X86ISD::MEMBARRIER";
19239 case X86ISD::MFENCE: return "X86ISD::MFENCE";
19240 case X86ISD::SFENCE: return "X86ISD::SFENCE";
19241 case X86ISD::LFENCE: return "X86ISD::LFENCE";
19242 case X86ISD::SEG_ALLOCA: return "X86ISD::SEG_ALLOCA";
19243 case X86ISD::WIN_FTOL: return "X86ISD::WIN_FTOL";
19244 case X86ISD::SAHF: return "X86ISD::SAHF";
19245 case X86ISD::RDRAND: return "X86ISD::RDRAND";
19246 case X86ISD::RDSEED: return "X86ISD::RDSEED";
19247 case X86ISD::VPMADDUBSW: return "X86ISD::VPMADDUBSW";
19248 case X86ISD::VPMADDWD: return "X86ISD::VPMADDWD";
19249 case X86ISD::FMADD: return "X86ISD::FMADD";
19250 case X86ISD::FMSUB: return "X86ISD::FMSUB";
19251 case X86ISD::FNMADD: return "X86ISD::FNMADD";
19252 case X86ISD::FNMSUB: return "X86ISD::FNMSUB";
19253 case X86ISD::FMADDSUB: return "X86ISD::FMADDSUB";
19254 case X86ISD::FMSUBADD: return "X86ISD::FMSUBADD";
19255 case X86ISD::FMADD_RND: return "X86ISD::FMADD_RND";
19256 case X86ISD::FNMADD_RND: return "X86ISD::FNMADD_RND";
19257 case X86ISD::FMSUB_RND: return "X86ISD::FMSUB_RND";
19258 case X86ISD::FNMSUB_RND: return "X86ISD::FNMSUB_RND";
19259 case X86ISD::FMADDSUB_RND: return "X86ISD::FMADDSUB_RND";
19260 case X86ISD::FMSUBADD_RND: return "X86ISD::FMSUBADD_RND";
19261 case X86ISD::VRNDSCALE: return "X86ISD::VRNDSCALE";
19262 case X86ISD::VREDUCE: return "X86ISD::VREDUCE";
19263 case X86ISD::PCMPESTRI: return "X86ISD::PCMPESTRI";
19264 case X86ISD::PCMPISTRI: return "X86ISD::PCMPISTRI";
19265 case X86ISD::XTEST: return "X86ISD::XTEST";
19266 case X86ISD::COMPRESS: return "X86ISD::COMPRESS";
19267 case X86ISD::EXPAND: return "X86ISD::EXPAND";
19268 case X86ISD::SELECT: return "X86ISD::SELECT";
19269 case X86ISD::ADDSUB: return "X86ISD::ADDSUB";
19270 case X86ISD::RCP28: return "X86ISD::RCP28";
19271 case X86ISD::EXP2: return "X86ISD::EXP2";
19272 case X86ISD::RSQRT28: return "X86ISD::RSQRT28";
19273 case X86ISD::FADD_RND: return "X86ISD::FADD_RND";
19274 case X86ISD::FSUB_RND: return "X86ISD::FSUB_RND";
19275 case X86ISD::FMUL_RND: return "X86ISD::FMUL_RND";
19276 case X86ISD::FDIV_RND: return "X86ISD::FDIV_RND";
19277 case X86ISD::FSQRT_RND: return "X86ISD::FSQRT_RND";
19278 case X86ISD::FGETEXP_RND: return "X86ISD::FGETEXP_RND";
19279 case X86ISD::SCALEF: return "X86ISD::SCALEF";
19280 case X86ISD::ADDS: return "X86ISD::ADDS";
19281 case X86ISD::SUBS: return "X86ISD::SUBS";
19282 case X86ISD::AVG: return "X86ISD::AVG";
19283 case X86ISD::MULHRS: return "X86ISD::MULHRS";
19284 case X86ISD::SINT_TO_FP_RND: return "X86ISD::SINT_TO_FP_RND";
19285 case X86ISD::UINT_TO_FP_RND: return "X86ISD::UINT_TO_FP_RND";
19286 case X86ISD::FP_TO_SINT_RND: return "X86ISD::FP_TO_SINT_RND";
19287 case X86ISD::FP_TO_UINT_RND: return "X86ISD::FP_TO_UINT_RND";
19292 // isLegalAddressingMode - Return true if the addressing mode represented
19293 // by AM is legal for this target, for a load/store of the specified type.
19294 bool X86TargetLowering::isLegalAddressingMode(const DataLayout &DL,
19295 const AddrMode &AM, Type *Ty,
19296 unsigned AS) const {
19297 // X86 supports extremely general addressing modes.
19298 CodeModel::Model M = getTargetMachine().getCodeModel();
19299 Reloc::Model R = getTargetMachine().getRelocationModel();
19301 // X86 allows a sign-extended 32-bit immediate field as a displacement.
19302 if (!X86::isOffsetSuitableForCodeModel(AM.BaseOffs, M, AM.BaseGV != nullptr))
19307 Subtarget->ClassifyGlobalReference(AM.BaseGV, getTargetMachine());
19309 // If a reference to this global requires an extra load, we can't fold it.
19310 if (isGlobalStubReference(GVFlags))
19313 // If BaseGV requires a register for the PIC base, we cannot also have a
19314 // BaseReg specified.
19315 if (AM.HasBaseReg && isGlobalRelativeToPICBase(GVFlags))
19318 // If lower 4G is not available, then we must use rip-relative addressing.
19319 if ((M != CodeModel::Small || R != Reloc::Static) &&
19320 Subtarget->is64Bit() && (AM.BaseOffs || AM.Scale > 1))
19324 switch (AM.Scale) {
19330 // These scales always work.
19335 // These scales are formed with basereg+scalereg. Only accept if there is
19340 default: // Other stuff never works.
19347 bool X86TargetLowering::isVectorShiftByScalarCheap(Type *Ty) const {
19348 unsigned Bits = Ty->getScalarSizeInBits();
19350 // 8-bit shifts are always expensive, but versions with a scalar amount aren't
19351 // particularly cheaper than those without.
19355 // On AVX2 there are new vpsllv[dq] instructions (and other shifts), that make
19356 // variable shifts just as cheap as scalar ones.
19357 if (Subtarget->hasInt256() && (Bits == 32 || Bits == 64))
19360 // Otherwise, it's significantly cheaper to shift by a scalar amount than by a
19361 // fully general vector.
19365 bool X86TargetLowering::isTruncateFree(Type *Ty1, Type *Ty2) const {
19366 if (!Ty1->isIntegerTy() || !Ty2->isIntegerTy())
19368 unsigned NumBits1 = Ty1->getPrimitiveSizeInBits();
19369 unsigned NumBits2 = Ty2->getPrimitiveSizeInBits();
19370 return NumBits1 > NumBits2;
19373 bool X86TargetLowering::allowTruncateForTailCall(Type *Ty1, Type *Ty2) const {
19374 if (!Ty1->isIntegerTy() || !Ty2->isIntegerTy())
19377 if (!isTypeLegal(EVT::getEVT(Ty1)))
19380 assert(Ty1->getPrimitiveSizeInBits() <= 64 && "i128 is probably not a noop");
19382 // Assuming the caller doesn't have a zeroext or signext return parameter,
19383 // truncation all the way down to i1 is valid.
19387 bool X86TargetLowering::isLegalICmpImmediate(int64_t Imm) const {
19388 return isInt<32>(Imm);
19391 bool X86TargetLowering::isLegalAddImmediate(int64_t Imm) const {
19392 // Can also use sub to handle negated immediates.
19393 return isInt<32>(Imm);
19396 bool X86TargetLowering::isTruncateFree(EVT VT1, EVT VT2) const {
19397 if (!VT1.isInteger() || !VT2.isInteger())
19399 unsigned NumBits1 = VT1.getSizeInBits();
19400 unsigned NumBits2 = VT2.getSizeInBits();
19401 return NumBits1 > NumBits2;
19404 bool X86TargetLowering::isZExtFree(Type *Ty1, Type *Ty2) const {
19405 // x86-64 implicitly zero-extends 32-bit results in 64-bit registers.
19406 return Ty1->isIntegerTy(32) && Ty2->isIntegerTy(64) && Subtarget->is64Bit();
19409 bool X86TargetLowering::isZExtFree(EVT VT1, EVT VT2) const {
19410 // x86-64 implicitly zero-extends 32-bit results in 64-bit registers.
19411 return VT1 == MVT::i32 && VT2 == MVT::i64 && Subtarget->is64Bit();
19414 bool X86TargetLowering::isZExtFree(SDValue Val, EVT VT2) const {
19415 EVT VT1 = Val.getValueType();
19416 if (isZExtFree(VT1, VT2))
19419 if (Val.getOpcode() != ISD::LOAD)
19422 if (!VT1.isSimple() || !VT1.isInteger() ||
19423 !VT2.isSimple() || !VT2.isInteger())
19426 switch (VT1.getSimpleVT().SimpleTy) {
19431 // X86 has 8, 16, and 32-bit zero-extending loads.
19438 bool X86TargetLowering::isVectorLoadExtDesirable(SDValue) const { return true; }
19441 X86TargetLowering::isFMAFasterThanFMulAndFAdd(EVT VT) const {
19442 if (!(Subtarget->hasFMA() || Subtarget->hasFMA4() || Subtarget->hasAVX512()))
19445 VT = VT.getScalarType();
19447 if (!VT.isSimple())
19450 switch (VT.getSimpleVT().SimpleTy) {
19461 bool X86TargetLowering::isNarrowingProfitable(EVT VT1, EVT VT2) const {
19462 // i16 instructions are longer (0x66 prefix) and potentially slower.
19463 return !(VT1 == MVT::i32 && VT2 == MVT::i16);
19466 /// isShuffleMaskLegal - Targets can use this to indicate that they only
19467 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
19468 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
19469 /// are assumed to be legal.
19471 X86TargetLowering::isShuffleMaskLegal(const SmallVectorImpl<int> &M,
19473 if (!VT.isSimple())
19476 // Not for i1 vectors
19477 if (VT.getScalarType() == MVT::i1)
19480 // Very little shuffling can be done for 64-bit vectors right now.
19481 if (VT.getSizeInBits() == 64)
19484 // We only care that the types being shuffled are legal. The lowering can
19485 // handle any possible shuffle mask that results.
19486 return isTypeLegal(VT.getSimpleVT());
19490 X86TargetLowering::isVectorClearMaskLegal(const SmallVectorImpl<int> &Mask,
19492 // Just delegate to the generic legality, clear masks aren't special.
19493 return isShuffleMaskLegal(Mask, VT);
19496 //===----------------------------------------------------------------------===//
19497 // X86 Scheduler Hooks
19498 //===----------------------------------------------------------------------===//
19500 /// Utility function to emit xbegin specifying the start of an RTM region.
19501 static MachineBasicBlock *EmitXBegin(MachineInstr *MI, MachineBasicBlock *MBB,
19502 const TargetInstrInfo *TII) {
19503 DebugLoc DL = MI->getDebugLoc();
19505 const BasicBlock *BB = MBB->getBasicBlock();
19506 MachineFunction::iterator I = MBB;
19509 // For the v = xbegin(), we generate
19520 MachineBasicBlock *thisMBB = MBB;
19521 MachineFunction *MF = MBB->getParent();
19522 MachineBasicBlock *mainMBB = MF->CreateMachineBasicBlock(BB);
19523 MachineBasicBlock *sinkMBB = MF->CreateMachineBasicBlock(BB);
19524 MF->insert(I, mainMBB);
19525 MF->insert(I, sinkMBB);
19527 // Transfer the remainder of BB and its successor edges to sinkMBB.
19528 sinkMBB->splice(sinkMBB->begin(), MBB,
19529 std::next(MachineBasicBlock::iterator(MI)), MBB->end());
19530 sinkMBB->transferSuccessorsAndUpdatePHIs(MBB);
19534 // # fallthrough to mainMBB
19535 // # abortion to sinkMBB
19536 BuildMI(thisMBB, DL, TII->get(X86::XBEGIN_4)).addMBB(sinkMBB);
19537 thisMBB->addSuccessor(mainMBB);
19538 thisMBB->addSuccessor(sinkMBB);
19542 BuildMI(mainMBB, DL, TII->get(X86::MOV32ri), X86::EAX).addImm(-1);
19543 mainMBB->addSuccessor(sinkMBB);
19546 // EAX is live into the sinkMBB
19547 sinkMBB->addLiveIn(X86::EAX);
19548 BuildMI(*sinkMBB, sinkMBB->begin(), DL,
19549 TII->get(TargetOpcode::COPY), MI->getOperand(0).getReg())
19552 MI->eraseFromParent();
19556 // FIXME: When we get size specific XMM0 registers, i.e. XMM0_V16I8
19557 // or XMM0_V32I8 in AVX all of this code can be replaced with that
19558 // in the .td file.
19559 static MachineBasicBlock *EmitPCMPSTRM(MachineInstr *MI, MachineBasicBlock *BB,
19560 const TargetInstrInfo *TII) {
19562 switch (MI->getOpcode()) {
19563 default: llvm_unreachable("illegal opcode!");
19564 case X86::PCMPISTRM128REG: Opc = X86::PCMPISTRM128rr; break;
19565 case X86::VPCMPISTRM128REG: Opc = X86::VPCMPISTRM128rr; break;
19566 case X86::PCMPISTRM128MEM: Opc = X86::PCMPISTRM128rm; break;
19567 case X86::VPCMPISTRM128MEM: Opc = X86::VPCMPISTRM128rm; break;
19568 case X86::PCMPESTRM128REG: Opc = X86::PCMPESTRM128rr; break;
19569 case X86::VPCMPESTRM128REG: Opc = X86::VPCMPESTRM128rr; break;
19570 case X86::PCMPESTRM128MEM: Opc = X86::PCMPESTRM128rm; break;
19571 case X86::VPCMPESTRM128MEM: Opc = X86::VPCMPESTRM128rm; break;
19574 DebugLoc dl = MI->getDebugLoc();
19575 MachineInstrBuilder MIB = BuildMI(*BB, MI, dl, TII->get(Opc));
19577 unsigned NumArgs = MI->getNumOperands();
19578 for (unsigned i = 1; i < NumArgs; ++i) {
19579 MachineOperand &Op = MI->getOperand(i);
19580 if (!(Op.isReg() && Op.isImplicit()))
19581 MIB.addOperand(Op);
19583 if (MI->hasOneMemOperand())
19584 MIB->setMemRefs(MI->memoperands_begin(), MI->memoperands_end());
19586 BuildMI(*BB, MI, dl,
19587 TII->get(TargetOpcode::COPY), MI->getOperand(0).getReg())
19588 .addReg(X86::XMM0);
19590 MI->eraseFromParent();
19594 // FIXME: Custom handling because TableGen doesn't support multiple implicit
19595 // defs in an instruction pattern
19596 static MachineBasicBlock *EmitPCMPSTRI(MachineInstr *MI, MachineBasicBlock *BB,
19597 const TargetInstrInfo *TII) {
19599 switch (MI->getOpcode()) {
19600 default: llvm_unreachable("illegal opcode!");
19601 case X86::PCMPISTRIREG: Opc = X86::PCMPISTRIrr; break;
19602 case X86::VPCMPISTRIREG: Opc = X86::VPCMPISTRIrr; break;
19603 case X86::PCMPISTRIMEM: Opc = X86::PCMPISTRIrm; break;
19604 case X86::VPCMPISTRIMEM: Opc = X86::VPCMPISTRIrm; break;
19605 case X86::PCMPESTRIREG: Opc = X86::PCMPESTRIrr; break;
19606 case X86::VPCMPESTRIREG: Opc = X86::VPCMPESTRIrr; break;
19607 case X86::PCMPESTRIMEM: Opc = X86::PCMPESTRIrm; break;
19608 case X86::VPCMPESTRIMEM: Opc = X86::VPCMPESTRIrm; break;
19611 DebugLoc dl = MI->getDebugLoc();
19612 MachineInstrBuilder MIB = BuildMI(*BB, MI, dl, TII->get(Opc));
19614 unsigned NumArgs = MI->getNumOperands(); // remove the results
19615 for (unsigned i = 1; i < NumArgs; ++i) {
19616 MachineOperand &Op = MI->getOperand(i);
19617 if (!(Op.isReg() && Op.isImplicit()))
19618 MIB.addOperand(Op);
19620 if (MI->hasOneMemOperand())
19621 MIB->setMemRefs(MI->memoperands_begin(), MI->memoperands_end());
19623 BuildMI(*BB, MI, dl,
19624 TII->get(TargetOpcode::COPY), MI->getOperand(0).getReg())
19627 MI->eraseFromParent();
19631 static MachineBasicBlock *EmitMonitor(MachineInstr *MI, MachineBasicBlock *BB,
19632 const X86Subtarget *Subtarget) {
19633 DebugLoc dl = MI->getDebugLoc();
19634 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
19635 // Address into RAX/EAX, other two args into ECX, EDX.
19636 unsigned MemOpc = Subtarget->is64Bit() ? X86::LEA64r : X86::LEA32r;
19637 unsigned MemReg = Subtarget->is64Bit() ? X86::RAX : X86::EAX;
19638 MachineInstrBuilder MIB = BuildMI(*BB, MI, dl, TII->get(MemOpc), MemReg);
19639 for (int i = 0; i < X86::AddrNumOperands; ++i)
19640 MIB.addOperand(MI->getOperand(i));
19642 unsigned ValOps = X86::AddrNumOperands;
19643 BuildMI(*BB, MI, dl, TII->get(TargetOpcode::COPY), X86::ECX)
19644 .addReg(MI->getOperand(ValOps).getReg());
19645 BuildMI(*BB, MI, dl, TII->get(TargetOpcode::COPY), X86::EDX)
19646 .addReg(MI->getOperand(ValOps+1).getReg());
19648 // The instruction doesn't actually take any operands though.
19649 BuildMI(*BB, MI, dl, TII->get(X86::MONITORrrr));
19651 MI->eraseFromParent(); // The pseudo is gone now.
19655 MachineBasicBlock *
19656 X86TargetLowering::EmitVAARG64WithCustomInserter(MachineInstr *MI,
19657 MachineBasicBlock *MBB) const {
19658 // Emit va_arg instruction on X86-64.
19660 // Operands to this pseudo-instruction:
19661 // 0 ) Output : destination address (reg)
19662 // 1-5) Input : va_list address (addr, i64mem)
19663 // 6 ) ArgSize : Size (in bytes) of vararg type
19664 // 7 ) ArgMode : 0=overflow only, 1=use gp_offset, 2=use fp_offset
19665 // 8 ) Align : Alignment of type
19666 // 9 ) EFLAGS (implicit-def)
19668 assert(MI->getNumOperands() == 10 && "VAARG_64 should have 10 operands!");
19669 static_assert(X86::AddrNumOperands == 5,
19670 "VAARG_64 assumes 5 address operands");
19672 unsigned DestReg = MI->getOperand(0).getReg();
19673 MachineOperand &Base = MI->getOperand(1);
19674 MachineOperand &Scale = MI->getOperand(2);
19675 MachineOperand &Index = MI->getOperand(3);
19676 MachineOperand &Disp = MI->getOperand(4);
19677 MachineOperand &Segment = MI->getOperand(5);
19678 unsigned ArgSize = MI->getOperand(6).getImm();
19679 unsigned ArgMode = MI->getOperand(7).getImm();
19680 unsigned Align = MI->getOperand(8).getImm();
19682 // Memory Reference
19683 assert(MI->hasOneMemOperand() && "Expected VAARG_64 to have one memoperand");
19684 MachineInstr::mmo_iterator MMOBegin = MI->memoperands_begin();
19685 MachineInstr::mmo_iterator MMOEnd = MI->memoperands_end();
19687 // Machine Information
19688 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
19689 MachineRegisterInfo &MRI = MBB->getParent()->getRegInfo();
19690 const TargetRegisterClass *AddrRegClass = getRegClassFor(MVT::i64);
19691 const TargetRegisterClass *OffsetRegClass = getRegClassFor(MVT::i32);
19692 DebugLoc DL = MI->getDebugLoc();
19694 // struct va_list {
19697 // i64 overflow_area (address)
19698 // i64 reg_save_area (address)
19700 // sizeof(va_list) = 24
19701 // alignment(va_list) = 8
19703 unsigned TotalNumIntRegs = 6;
19704 unsigned TotalNumXMMRegs = 8;
19705 bool UseGPOffset = (ArgMode == 1);
19706 bool UseFPOffset = (ArgMode == 2);
19707 unsigned MaxOffset = TotalNumIntRegs * 8 +
19708 (UseFPOffset ? TotalNumXMMRegs * 16 : 0);
19710 /* Align ArgSize to a multiple of 8 */
19711 unsigned ArgSizeA8 = (ArgSize + 7) & ~7;
19712 bool NeedsAlign = (Align > 8);
19714 MachineBasicBlock *thisMBB = MBB;
19715 MachineBasicBlock *overflowMBB;
19716 MachineBasicBlock *offsetMBB;
19717 MachineBasicBlock *endMBB;
19719 unsigned OffsetDestReg = 0; // Argument address computed by offsetMBB
19720 unsigned OverflowDestReg = 0; // Argument address computed by overflowMBB
19721 unsigned OffsetReg = 0;
19723 if (!UseGPOffset && !UseFPOffset) {
19724 // If we only pull from the overflow region, we don't create a branch.
19725 // We don't need to alter control flow.
19726 OffsetDestReg = 0; // unused
19727 OverflowDestReg = DestReg;
19729 offsetMBB = nullptr;
19730 overflowMBB = thisMBB;
19733 // First emit code to check if gp_offset (or fp_offset) is below the bound.
19734 // If so, pull the argument from reg_save_area. (branch to offsetMBB)
19735 // If not, pull from overflow_area. (branch to overflowMBB)
19740 // offsetMBB overflowMBB
19745 // Registers for the PHI in endMBB
19746 OffsetDestReg = MRI.createVirtualRegister(AddrRegClass);
19747 OverflowDestReg = MRI.createVirtualRegister(AddrRegClass);
19749 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
19750 MachineFunction *MF = MBB->getParent();
19751 overflowMBB = MF->CreateMachineBasicBlock(LLVM_BB);
19752 offsetMBB = MF->CreateMachineBasicBlock(LLVM_BB);
19753 endMBB = MF->CreateMachineBasicBlock(LLVM_BB);
19755 MachineFunction::iterator MBBIter = MBB;
19758 // Insert the new basic blocks
19759 MF->insert(MBBIter, offsetMBB);
19760 MF->insert(MBBIter, overflowMBB);
19761 MF->insert(MBBIter, endMBB);
19763 // Transfer the remainder of MBB and its successor edges to endMBB.
19764 endMBB->splice(endMBB->begin(), thisMBB,
19765 std::next(MachineBasicBlock::iterator(MI)), thisMBB->end());
19766 endMBB->transferSuccessorsAndUpdatePHIs(thisMBB);
19768 // Make offsetMBB and overflowMBB successors of thisMBB
19769 thisMBB->addSuccessor(offsetMBB);
19770 thisMBB->addSuccessor(overflowMBB);
19772 // endMBB is a successor of both offsetMBB and overflowMBB
19773 offsetMBB->addSuccessor(endMBB);
19774 overflowMBB->addSuccessor(endMBB);
19776 // Load the offset value into a register
19777 OffsetReg = MRI.createVirtualRegister(OffsetRegClass);
19778 BuildMI(thisMBB, DL, TII->get(X86::MOV32rm), OffsetReg)
19782 .addDisp(Disp, UseFPOffset ? 4 : 0)
19783 .addOperand(Segment)
19784 .setMemRefs(MMOBegin, MMOEnd);
19786 // Check if there is enough room left to pull this argument.
19787 BuildMI(thisMBB, DL, TII->get(X86::CMP32ri))
19789 .addImm(MaxOffset + 8 - ArgSizeA8);
19791 // Branch to "overflowMBB" if offset >= max
19792 // Fall through to "offsetMBB" otherwise
19793 BuildMI(thisMBB, DL, TII->get(X86::GetCondBranchFromCond(X86::COND_AE)))
19794 .addMBB(overflowMBB);
19797 // In offsetMBB, emit code to use the reg_save_area.
19799 assert(OffsetReg != 0);
19801 // Read the reg_save_area address.
19802 unsigned RegSaveReg = MRI.createVirtualRegister(AddrRegClass);
19803 BuildMI(offsetMBB, DL, TII->get(X86::MOV64rm), RegSaveReg)
19808 .addOperand(Segment)
19809 .setMemRefs(MMOBegin, MMOEnd);
19811 // Zero-extend the offset
19812 unsigned OffsetReg64 = MRI.createVirtualRegister(AddrRegClass);
19813 BuildMI(offsetMBB, DL, TII->get(X86::SUBREG_TO_REG), OffsetReg64)
19816 .addImm(X86::sub_32bit);
19818 // Add the offset to the reg_save_area to get the final address.
19819 BuildMI(offsetMBB, DL, TII->get(X86::ADD64rr), OffsetDestReg)
19820 .addReg(OffsetReg64)
19821 .addReg(RegSaveReg);
19823 // Compute the offset for the next argument
19824 unsigned NextOffsetReg = MRI.createVirtualRegister(OffsetRegClass);
19825 BuildMI(offsetMBB, DL, TII->get(X86::ADD32ri), NextOffsetReg)
19827 .addImm(UseFPOffset ? 16 : 8);
19829 // Store it back into the va_list.
19830 BuildMI(offsetMBB, DL, TII->get(X86::MOV32mr))
19834 .addDisp(Disp, UseFPOffset ? 4 : 0)
19835 .addOperand(Segment)
19836 .addReg(NextOffsetReg)
19837 .setMemRefs(MMOBegin, MMOEnd);
19840 BuildMI(offsetMBB, DL, TII->get(X86::JMP_1))
19845 // Emit code to use overflow area
19848 // Load the overflow_area address into a register.
19849 unsigned OverflowAddrReg = MRI.createVirtualRegister(AddrRegClass);
19850 BuildMI(overflowMBB, DL, TII->get(X86::MOV64rm), OverflowAddrReg)
19855 .addOperand(Segment)
19856 .setMemRefs(MMOBegin, MMOEnd);
19858 // If we need to align it, do so. Otherwise, just copy the address
19859 // to OverflowDestReg.
19861 // Align the overflow address
19862 assert((Align & (Align-1)) == 0 && "Alignment must be a power of 2");
19863 unsigned TmpReg = MRI.createVirtualRegister(AddrRegClass);
19865 // aligned_addr = (addr + (align-1)) & ~(align-1)
19866 BuildMI(overflowMBB, DL, TII->get(X86::ADD64ri32), TmpReg)
19867 .addReg(OverflowAddrReg)
19870 BuildMI(overflowMBB, DL, TII->get(X86::AND64ri32), OverflowDestReg)
19872 .addImm(~(uint64_t)(Align-1));
19874 BuildMI(overflowMBB, DL, TII->get(TargetOpcode::COPY), OverflowDestReg)
19875 .addReg(OverflowAddrReg);
19878 // Compute the next overflow address after this argument.
19879 // (the overflow address should be kept 8-byte aligned)
19880 unsigned NextAddrReg = MRI.createVirtualRegister(AddrRegClass);
19881 BuildMI(overflowMBB, DL, TII->get(X86::ADD64ri32), NextAddrReg)
19882 .addReg(OverflowDestReg)
19883 .addImm(ArgSizeA8);
19885 // Store the new overflow address.
19886 BuildMI(overflowMBB, DL, TII->get(X86::MOV64mr))
19891 .addOperand(Segment)
19892 .addReg(NextAddrReg)
19893 .setMemRefs(MMOBegin, MMOEnd);
19895 // If we branched, emit the PHI to the front of endMBB.
19897 BuildMI(*endMBB, endMBB->begin(), DL,
19898 TII->get(X86::PHI), DestReg)
19899 .addReg(OffsetDestReg).addMBB(offsetMBB)
19900 .addReg(OverflowDestReg).addMBB(overflowMBB);
19903 // Erase the pseudo instruction
19904 MI->eraseFromParent();
19909 MachineBasicBlock *
19910 X86TargetLowering::EmitVAStartSaveXMMRegsWithCustomInserter(
19912 MachineBasicBlock *MBB) const {
19913 // Emit code to save XMM registers to the stack. The ABI says that the
19914 // number of registers to save is given in %al, so it's theoretically
19915 // possible to do an indirect jump trick to avoid saving all of them,
19916 // however this code takes a simpler approach and just executes all
19917 // of the stores if %al is non-zero. It's less code, and it's probably
19918 // easier on the hardware branch predictor, and stores aren't all that
19919 // expensive anyway.
19921 // Create the new basic blocks. One block contains all the XMM stores,
19922 // and one block is the final destination regardless of whether any
19923 // stores were performed.
19924 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
19925 MachineFunction *F = MBB->getParent();
19926 MachineFunction::iterator MBBIter = MBB;
19928 MachineBasicBlock *XMMSaveMBB = F->CreateMachineBasicBlock(LLVM_BB);
19929 MachineBasicBlock *EndMBB = F->CreateMachineBasicBlock(LLVM_BB);
19930 F->insert(MBBIter, XMMSaveMBB);
19931 F->insert(MBBIter, EndMBB);
19933 // Transfer the remainder of MBB and its successor edges to EndMBB.
19934 EndMBB->splice(EndMBB->begin(), MBB,
19935 std::next(MachineBasicBlock::iterator(MI)), MBB->end());
19936 EndMBB->transferSuccessorsAndUpdatePHIs(MBB);
19938 // The original block will now fall through to the XMM save block.
19939 MBB->addSuccessor(XMMSaveMBB);
19940 // The XMMSaveMBB will fall through to the end block.
19941 XMMSaveMBB->addSuccessor(EndMBB);
19943 // Now add the instructions.
19944 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
19945 DebugLoc DL = MI->getDebugLoc();
19947 unsigned CountReg = MI->getOperand(0).getReg();
19948 int64_t RegSaveFrameIndex = MI->getOperand(1).getImm();
19949 int64_t VarArgsFPOffset = MI->getOperand(2).getImm();
19951 if (!Subtarget->isTargetWin64()) {
19952 // If %al is 0, branch around the XMM save block.
19953 BuildMI(MBB, DL, TII->get(X86::TEST8rr)).addReg(CountReg).addReg(CountReg);
19954 BuildMI(MBB, DL, TII->get(X86::JE_1)).addMBB(EndMBB);
19955 MBB->addSuccessor(EndMBB);
19958 // Make sure the last operand is EFLAGS, which gets clobbered by the branch
19959 // that was just emitted, but clearly shouldn't be "saved".
19960 assert((MI->getNumOperands() <= 3 ||
19961 !MI->getOperand(MI->getNumOperands() - 1).isReg() ||
19962 MI->getOperand(MI->getNumOperands() - 1).getReg() == X86::EFLAGS)
19963 && "Expected last argument to be EFLAGS");
19964 unsigned MOVOpc = Subtarget->hasFp256() ? X86::VMOVAPSmr : X86::MOVAPSmr;
19965 // In the XMM save block, save all the XMM argument registers.
19966 for (int i = 3, e = MI->getNumOperands() - 1; i != e; ++i) {
19967 int64_t Offset = (i - 3) * 16 + VarArgsFPOffset;
19968 MachineMemOperand *MMO = F->getMachineMemOperand(
19969 MachinePointerInfo::getFixedStack(*F, RegSaveFrameIndex, Offset),
19970 MachineMemOperand::MOStore,
19971 /*Size=*/16, /*Align=*/16);
19972 BuildMI(XMMSaveMBB, DL, TII->get(MOVOpc))
19973 .addFrameIndex(RegSaveFrameIndex)
19974 .addImm(/*Scale=*/1)
19975 .addReg(/*IndexReg=*/0)
19976 .addImm(/*Disp=*/Offset)
19977 .addReg(/*Segment=*/0)
19978 .addReg(MI->getOperand(i).getReg())
19979 .addMemOperand(MMO);
19982 MI->eraseFromParent(); // The pseudo instruction is gone now.
19987 // The EFLAGS operand of SelectItr might be missing a kill marker
19988 // because there were multiple uses of EFLAGS, and ISel didn't know
19989 // which to mark. Figure out whether SelectItr should have had a
19990 // kill marker, and set it if it should. Returns the correct kill
19992 static bool checkAndUpdateEFLAGSKill(MachineBasicBlock::iterator SelectItr,
19993 MachineBasicBlock* BB,
19994 const TargetRegisterInfo* TRI) {
19995 // Scan forward through BB for a use/def of EFLAGS.
19996 MachineBasicBlock::iterator miI(std::next(SelectItr));
19997 for (MachineBasicBlock::iterator miE = BB->end(); miI != miE; ++miI) {
19998 const MachineInstr& mi = *miI;
19999 if (mi.readsRegister(X86::EFLAGS))
20001 if (mi.definesRegister(X86::EFLAGS))
20002 break; // Should have kill-flag - update below.
20005 // If we hit the end of the block, check whether EFLAGS is live into a
20007 if (miI == BB->end()) {
20008 for (MachineBasicBlock::succ_iterator sItr = BB->succ_begin(),
20009 sEnd = BB->succ_end();
20010 sItr != sEnd; ++sItr) {
20011 MachineBasicBlock* succ = *sItr;
20012 if (succ->isLiveIn(X86::EFLAGS))
20017 // We found a def, or hit the end of the basic block and EFLAGS wasn't live
20018 // out. SelectMI should have a kill flag on EFLAGS.
20019 SelectItr->addRegisterKilled(X86::EFLAGS, TRI);
20023 // Return true if it is OK for this CMOV pseudo-opcode to be cascaded
20024 // together with other CMOV pseudo-opcodes into a single basic-block with
20025 // conditional jump around it.
20026 static bool isCMOVPseudo(MachineInstr *MI) {
20027 switch (MI->getOpcode()) {
20028 case X86::CMOV_FR32:
20029 case X86::CMOV_FR64:
20030 case X86::CMOV_GR8:
20031 case X86::CMOV_GR16:
20032 case X86::CMOV_GR32:
20033 case X86::CMOV_RFP32:
20034 case X86::CMOV_RFP64:
20035 case X86::CMOV_RFP80:
20036 case X86::CMOV_V2F64:
20037 case X86::CMOV_V2I64:
20038 case X86::CMOV_V4F32:
20039 case X86::CMOV_V4F64:
20040 case X86::CMOV_V4I64:
20041 case X86::CMOV_V16F32:
20042 case X86::CMOV_V8F32:
20043 case X86::CMOV_V8F64:
20044 case X86::CMOV_V8I64:
20045 case X86::CMOV_V8I1:
20046 case X86::CMOV_V16I1:
20047 case X86::CMOV_V32I1:
20048 case X86::CMOV_V64I1:
20056 MachineBasicBlock *
20057 X86TargetLowering::EmitLoweredSelect(MachineInstr *MI,
20058 MachineBasicBlock *BB) const {
20059 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
20060 DebugLoc DL = MI->getDebugLoc();
20062 // To "insert" a SELECT_CC instruction, we actually have to insert the
20063 // diamond control-flow pattern. The incoming instruction knows the
20064 // destination vreg to set, the condition code register to branch on, the
20065 // true/false values to select between, and a branch opcode to use.
20066 const BasicBlock *LLVM_BB = BB->getBasicBlock();
20067 MachineFunction::iterator It = BB;
20073 // cmpTY ccX, r1, r2
20075 // fallthrough --> copy0MBB
20076 MachineBasicBlock *thisMBB = BB;
20077 MachineFunction *F = BB->getParent();
20079 // This code lowers all pseudo-CMOV instructions. Generally it lowers these
20080 // as described above, by inserting a BB, and then making a PHI at the join
20081 // point to select the true and false operands of the CMOV in the PHI.
20083 // The code also handles two different cases of multiple CMOV opcodes
20087 // In this case, there are multiple CMOVs in a row, all which are based on
20088 // the same condition setting (or the exact opposite condition setting).
20089 // In this case we can lower all the CMOVs using a single inserted BB, and
20090 // then make a number of PHIs at the join point to model the CMOVs. The only
20091 // trickiness here, is that in a case like:
20093 // t2 = CMOV cond1 t1, f1
20094 // t3 = CMOV cond1 t2, f2
20096 // when rewriting this into PHIs, we have to perform some renaming on the
20097 // temps since you cannot have a PHI operand refer to a PHI result earlier
20098 // in the same block. The "simple" but wrong lowering would be:
20100 // t2 = PHI t1(BB1), f1(BB2)
20101 // t3 = PHI t2(BB1), f2(BB2)
20103 // but clearly t2 is not defined in BB1, so that is incorrect. The proper
20104 // renaming is to note that on the path through BB1, t2 is really just a
20105 // copy of t1, and do that renaming, properly generating:
20107 // t2 = PHI t1(BB1), f1(BB2)
20108 // t3 = PHI t1(BB1), f2(BB2)
20110 // Case 2, we lower cascaded CMOVs such as
20112 // (CMOV (CMOV F, T, cc1), T, cc2)
20114 // to two successives branches. For that, we look for another CMOV as the
20115 // following instruction.
20117 // Without this, we would add a PHI between the two jumps, which ends up
20118 // creating a few copies all around. For instance, for
20120 // (sitofp (zext (fcmp une)))
20122 // we would generate:
20124 // ucomiss %xmm1, %xmm0
20125 // movss <1.0f>, %xmm0
20126 // movaps %xmm0, %xmm1
20128 // xorps %xmm1, %xmm1
20131 // movaps %xmm1, %xmm0
20135 // because this custom-inserter would have generated:
20147 // A: X = ...; Y = ...
20149 // C: Z = PHI [X, A], [Y, B]
20151 // E: PHI [X, C], [Z, D]
20153 // If we lower both CMOVs in a single step, we can instead generate:
20165 // A: X = ...; Y = ...
20167 // E: PHI [X, A], [X, C], [Y, D]
20169 // Which, in our sitofp/fcmp example, gives us something like:
20171 // ucomiss %xmm1, %xmm0
20172 // movss <1.0f>, %xmm0
20175 // xorps %xmm0, %xmm0
20179 MachineInstr *CascadedCMOV = nullptr;
20180 MachineInstr *LastCMOV = MI;
20181 X86::CondCode CC = X86::CondCode(MI->getOperand(3).getImm());
20182 X86::CondCode OppCC = X86::GetOppositeBranchCondition(CC);
20183 MachineBasicBlock::iterator NextMIIt =
20184 std::next(MachineBasicBlock::iterator(MI));
20186 // Check for case 1, where there are multiple CMOVs with the same condition
20187 // first. Of the two cases of multiple CMOV lowerings, case 1 reduces the
20188 // number of jumps the most.
20190 if (isCMOVPseudo(MI)) {
20191 // See if we have a string of CMOVS with the same condition.
20192 while (NextMIIt != BB->end() &&
20193 isCMOVPseudo(NextMIIt) &&
20194 (NextMIIt->getOperand(3).getImm() == CC ||
20195 NextMIIt->getOperand(3).getImm() == OppCC)) {
20196 LastCMOV = &*NextMIIt;
20201 // This checks for case 2, but only do this if we didn't already find
20202 // case 1, as indicated by LastCMOV == MI.
20203 if (LastCMOV == MI &&
20204 NextMIIt != BB->end() && NextMIIt->getOpcode() == MI->getOpcode() &&
20205 NextMIIt->getOperand(2).getReg() == MI->getOperand(2).getReg() &&
20206 NextMIIt->getOperand(1).getReg() == MI->getOperand(0).getReg()) {
20207 CascadedCMOV = &*NextMIIt;
20210 MachineBasicBlock *jcc1MBB = nullptr;
20212 // If we have a cascaded CMOV, we lower it to two successive branches to
20213 // the same block. EFLAGS is used by both, so mark it as live in the second.
20214 if (CascadedCMOV) {
20215 jcc1MBB = F->CreateMachineBasicBlock(LLVM_BB);
20216 F->insert(It, jcc1MBB);
20217 jcc1MBB->addLiveIn(X86::EFLAGS);
20220 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
20221 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
20222 F->insert(It, copy0MBB);
20223 F->insert(It, sinkMBB);
20225 // If the EFLAGS register isn't dead in the terminator, then claim that it's
20226 // live into the sink and copy blocks.
20227 const TargetRegisterInfo *TRI = Subtarget->getRegisterInfo();
20229 MachineInstr *LastEFLAGSUser = CascadedCMOV ? CascadedCMOV : LastCMOV;
20230 if (!LastEFLAGSUser->killsRegister(X86::EFLAGS) &&
20231 !checkAndUpdateEFLAGSKill(LastEFLAGSUser, BB, TRI)) {
20232 copy0MBB->addLiveIn(X86::EFLAGS);
20233 sinkMBB->addLiveIn(X86::EFLAGS);
20236 // Transfer the remainder of BB and its successor edges to sinkMBB.
20237 sinkMBB->splice(sinkMBB->begin(), BB,
20238 std::next(MachineBasicBlock::iterator(LastCMOV)), BB->end());
20239 sinkMBB->transferSuccessorsAndUpdatePHIs(BB);
20241 // Add the true and fallthrough blocks as its successors.
20242 if (CascadedCMOV) {
20243 // The fallthrough block may be jcc1MBB, if we have a cascaded CMOV.
20244 BB->addSuccessor(jcc1MBB);
20246 // In that case, jcc1MBB will itself fallthrough the copy0MBB, and
20247 // jump to the sinkMBB.
20248 jcc1MBB->addSuccessor(copy0MBB);
20249 jcc1MBB->addSuccessor(sinkMBB);
20251 BB->addSuccessor(copy0MBB);
20254 // The true block target of the first (or only) branch is always sinkMBB.
20255 BB->addSuccessor(sinkMBB);
20257 // Create the conditional branch instruction.
20258 unsigned Opc = X86::GetCondBranchFromCond(CC);
20259 BuildMI(BB, DL, TII->get(Opc)).addMBB(sinkMBB);
20261 if (CascadedCMOV) {
20262 unsigned Opc2 = X86::GetCondBranchFromCond(
20263 (X86::CondCode)CascadedCMOV->getOperand(3).getImm());
20264 BuildMI(jcc1MBB, DL, TII->get(Opc2)).addMBB(sinkMBB);
20268 // %FalseValue = ...
20269 // # fallthrough to sinkMBB
20270 copy0MBB->addSuccessor(sinkMBB);
20273 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
20275 MachineBasicBlock::iterator MIItBegin = MachineBasicBlock::iterator(MI);
20276 MachineBasicBlock::iterator MIItEnd =
20277 std::next(MachineBasicBlock::iterator(LastCMOV));
20278 MachineBasicBlock::iterator SinkInsertionPoint = sinkMBB->begin();
20279 DenseMap<unsigned, std::pair<unsigned, unsigned>> RegRewriteTable;
20280 MachineInstrBuilder MIB;
20282 // As we are creating the PHIs, we have to be careful if there is more than
20283 // one. Later CMOVs may reference the results of earlier CMOVs, but later
20284 // PHIs have to reference the individual true/false inputs from earlier PHIs.
20285 // That also means that PHI construction must work forward from earlier to
20286 // later, and that the code must maintain a mapping from earlier PHI's
20287 // destination registers, and the registers that went into the PHI.
20289 for (MachineBasicBlock::iterator MIIt = MIItBegin; MIIt != MIItEnd; ++MIIt) {
20290 unsigned DestReg = MIIt->getOperand(0).getReg();
20291 unsigned Op1Reg = MIIt->getOperand(1).getReg();
20292 unsigned Op2Reg = MIIt->getOperand(2).getReg();
20294 // If this CMOV we are generating is the opposite condition from
20295 // the jump we generated, then we have to swap the operands for the
20296 // PHI that is going to be generated.
20297 if (MIIt->getOperand(3).getImm() == OppCC)
20298 std::swap(Op1Reg, Op2Reg);
20300 if (RegRewriteTable.find(Op1Reg) != RegRewriteTable.end())
20301 Op1Reg = RegRewriteTable[Op1Reg].first;
20303 if (RegRewriteTable.find(Op2Reg) != RegRewriteTable.end())
20304 Op2Reg = RegRewriteTable[Op2Reg].second;
20306 MIB = BuildMI(*sinkMBB, SinkInsertionPoint, DL,
20307 TII->get(X86::PHI), DestReg)
20308 .addReg(Op1Reg).addMBB(copy0MBB)
20309 .addReg(Op2Reg).addMBB(thisMBB);
20311 // Add this PHI to the rewrite table.
20312 RegRewriteTable[DestReg] = std::make_pair(Op1Reg, Op2Reg);
20315 // If we have a cascaded CMOV, the second Jcc provides the same incoming
20316 // value as the first Jcc (the True operand of the SELECT_CC/CMOV nodes).
20317 if (CascadedCMOV) {
20318 MIB.addReg(MI->getOperand(2).getReg()).addMBB(jcc1MBB);
20319 // Copy the PHI result to the register defined by the second CMOV.
20320 BuildMI(*sinkMBB, std::next(MachineBasicBlock::iterator(MIB.getInstr())),
20321 DL, TII->get(TargetOpcode::COPY),
20322 CascadedCMOV->getOperand(0).getReg())
20323 .addReg(MI->getOperand(0).getReg());
20324 CascadedCMOV->eraseFromParent();
20327 // Now remove the CMOV(s).
20328 for (MachineBasicBlock::iterator MIIt = MIItBegin; MIIt != MIItEnd; )
20329 (MIIt++)->eraseFromParent();
20334 MachineBasicBlock *
20335 X86TargetLowering::EmitLoweredAtomicFP(MachineInstr *MI,
20336 MachineBasicBlock *BB) const {
20337 // Combine the following atomic floating-point modification pattern:
20338 // a.store(reg OP a.load(acquire), release)
20339 // Transform them into:
20340 // OPss (%gpr), %xmm
20341 // movss %xmm, (%gpr)
20342 // Or sd equivalent for 64-bit operations.
20344 switch (MI->getOpcode()) {
20345 default: llvm_unreachable("unexpected instr type for EmitLoweredAtomicFP");
20346 case X86::RELEASE_FADD32mr: MOp = X86::MOVSSmr; FOp = X86::ADDSSrm; break;
20347 case X86::RELEASE_FADD64mr: MOp = X86::MOVSDmr; FOp = X86::ADDSDrm; break;
20349 const X86InstrInfo *TII = Subtarget->getInstrInfo();
20350 DebugLoc DL = MI->getDebugLoc();
20351 MachineRegisterInfo &MRI = BB->getParent()->getRegInfo();
20352 unsigned MSrc = MI->getOperand(0).getReg();
20353 unsigned VSrc = MI->getOperand(5).getReg();
20354 MachineInstrBuilder MIM = BuildMI(*BB, MI, DL, TII->get(MOp))
20355 .addReg(/*Base=*/MSrc)
20356 .addImm(/*Scale=*/1)
20357 .addReg(/*Index=*/0)
20360 MachineInstr *MIO = BuildMI(*BB, (MachineInstr *)MIM, DL, TII->get(FOp),
20361 MRI.createVirtualRegister(MRI.getRegClass(VSrc)))
20363 .addReg(/*Base=*/MSrc)
20364 .addImm(/*Scale=*/1)
20365 .addReg(/*Index=*/0)
20366 .addImm(/*Disp=*/0)
20367 .addReg(/*Segment=*/0);
20368 MIM.addReg(MIO->getOperand(0).getReg(), RegState::Kill);
20369 MI->eraseFromParent(); // The pseudo instruction is gone now.
20373 MachineBasicBlock *
20374 X86TargetLowering::EmitLoweredSegAlloca(MachineInstr *MI,
20375 MachineBasicBlock *BB) const {
20376 MachineFunction *MF = BB->getParent();
20377 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
20378 DebugLoc DL = MI->getDebugLoc();
20379 const BasicBlock *LLVM_BB = BB->getBasicBlock();
20381 assert(MF->shouldSplitStack());
20383 const bool Is64Bit = Subtarget->is64Bit();
20384 const bool IsLP64 = Subtarget->isTarget64BitLP64();
20386 const unsigned TlsReg = Is64Bit ? X86::FS : X86::GS;
20387 const unsigned TlsOffset = IsLP64 ? 0x70 : Is64Bit ? 0x40 : 0x30;
20390 // ... [Till the alloca]
20391 // If stacklet is not large enough, jump to mallocMBB
20394 // Allocate by subtracting from RSP
20395 // Jump to continueMBB
20398 // Allocate by call to runtime
20402 // [rest of original BB]
20405 MachineBasicBlock *mallocMBB = MF->CreateMachineBasicBlock(LLVM_BB);
20406 MachineBasicBlock *bumpMBB = MF->CreateMachineBasicBlock(LLVM_BB);
20407 MachineBasicBlock *continueMBB = MF->CreateMachineBasicBlock(LLVM_BB);
20409 MachineRegisterInfo &MRI = MF->getRegInfo();
20410 const TargetRegisterClass *AddrRegClass =
20411 getRegClassFor(getPointerTy(MF->getDataLayout()));
20413 unsigned mallocPtrVReg = MRI.createVirtualRegister(AddrRegClass),
20414 bumpSPPtrVReg = MRI.createVirtualRegister(AddrRegClass),
20415 tmpSPVReg = MRI.createVirtualRegister(AddrRegClass),
20416 SPLimitVReg = MRI.createVirtualRegister(AddrRegClass),
20417 sizeVReg = MI->getOperand(1).getReg(),
20418 physSPReg = IsLP64 || Subtarget->isTargetNaCl64() ? X86::RSP : X86::ESP;
20420 MachineFunction::iterator MBBIter = BB;
20423 MF->insert(MBBIter, bumpMBB);
20424 MF->insert(MBBIter, mallocMBB);
20425 MF->insert(MBBIter, continueMBB);
20427 continueMBB->splice(continueMBB->begin(), BB,
20428 std::next(MachineBasicBlock::iterator(MI)), BB->end());
20429 continueMBB->transferSuccessorsAndUpdatePHIs(BB);
20431 // Add code to the main basic block to check if the stack limit has been hit,
20432 // and if so, jump to mallocMBB otherwise to bumpMBB.
20433 BuildMI(BB, DL, TII->get(TargetOpcode::COPY), tmpSPVReg).addReg(physSPReg);
20434 BuildMI(BB, DL, TII->get(IsLP64 ? X86::SUB64rr:X86::SUB32rr), SPLimitVReg)
20435 .addReg(tmpSPVReg).addReg(sizeVReg);
20436 BuildMI(BB, DL, TII->get(IsLP64 ? X86::CMP64mr:X86::CMP32mr))
20437 .addReg(0).addImm(1).addReg(0).addImm(TlsOffset).addReg(TlsReg)
20438 .addReg(SPLimitVReg);
20439 BuildMI(BB, DL, TII->get(X86::JG_1)).addMBB(mallocMBB);
20441 // bumpMBB simply decreases the stack pointer, since we know the current
20442 // stacklet has enough space.
20443 BuildMI(bumpMBB, DL, TII->get(TargetOpcode::COPY), physSPReg)
20444 .addReg(SPLimitVReg);
20445 BuildMI(bumpMBB, DL, TII->get(TargetOpcode::COPY), bumpSPPtrVReg)
20446 .addReg(SPLimitVReg);
20447 BuildMI(bumpMBB, DL, TII->get(X86::JMP_1)).addMBB(continueMBB);
20449 // Calls into a routine in libgcc to allocate more space from the heap.
20450 const uint32_t *RegMask =
20451 Subtarget->getRegisterInfo()->getCallPreservedMask(*MF, CallingConv::C);
20453 BuildMI(mallocMBB, DL, TII->get(X86::MOV64rr), X86::RDI)
20455 BuildMI(mallocMBB, DL, TII->get(X86::CALL64pcrel32))
20456 .addExternalSymbol("__morestack_allocate_stack_space")
20457 .addRegMask(RegMask)
20458 .addReg(X86::RDI, RegState::Implicit)
20459 .addReg(X86::RAX, RegState::ImplicitDefine);
20460 } else if (Is64Bit) {
20461 BuildMI(mallocMBB, DL, TII->get(X86::MOV32rr), X86::EDI)
20463 BuildMI(mallocMBB, DL, TII->get(X86::CALL64pcrel32))
20464 .addExternalSymbol("__morestack_allocate_stack_space")
20465 .addRegMask(RegMask)
20466 .addReg(X86::EDI, RegState::Implicit)
20467 .addReg(X86::EAX, RegState::ImplicitDefine);
20469 BuildMI(mallocMBB, DL, TII->get(X86::SUB32ri), physSPReg).addReg(physSPReg)
20471 BuildMI(mallocMBB, DL, TII->get(X86::PUSH32r)).addReg(sizeVReg);
20472 BuildMI(mallocMBB, DL, TII->get(X86::CALLpcrel32))
20473 .addExternalSymbol("__morestack_allocate_stack_space")
20474 .addRegMask(RegMask)
20475 .addReg(X86::EAX, RegState::ImplicitDefine);
20479 BuildMI(mallocMBB, DL, TII->get(X86::ADD32ri), physSPReg).addReg(physSPReg)
20482 BuildMI(mallocMBB, DL, TII->get(TargetOpcode::COPY), mallocPtrVReg)
20483 .addReg(IsLP64 ? X86::RAX : X86::EAX);
20484 BuildMI(mallocMBB, DL, TII->get(X86::JMP_1)).addMBB(continueMBB);
20486 // Set up the CFG correctly.
20487 BB->addSuccessor(bumpMBB);
20488 BB->addSuccessor(mallocMBB);
20489 mallocMBB->addSuccessor(continueMBB);
20490 bumpMBB->addSuccessor(continueMBB);
20492 // Take care of the PHI nodes.
20493 BuildMI(*continueMBB, continueMBB->begin(), DL, TII->get(X86::PHI),
20494 MI->getOperand(0).getReg())
20495 .addReg(mallocPtrVReg).addMBB(mallocMBB)
20496 .addReg(bumpSPPtrVReg).addMBB(bumpMBB);
20498 // Delete the original pseudo instruction.
20499 MI->eraseFromParent();
20502 return continueMBB;
20505 MachineBasicBlock *
20506 X86TargetLowering::EmitLoweredWinAlloca(MachineInstr *MI,
20507 MachineBasicBlock *BB) const {
20508 DebugLoc DL = MI->getDebugLoc();
20510 assert(!Subtarget->isTargetMachO());
20512 Subtarget->getFrameLowering()->emitStackProbeCall(*BB->getParent(), *BB, MI,
20515 MI->eraseFromParent(); // The pseudo instruction is gone now.
20519 MachineBasicBlock *
20520 X86TargetLowering::EmitLoweredTLSCall(MachineInstr *MI,
20521 MachineBasicBlock *BB) const {
20522 // This is pretty easy. We're taking the value that we received from
20523 // our load from the relocation, sticking it in either RDI (x86-64)
20524 // or EAX and doing an indirect call. The return value will then
20525 // be in the normal return register.
20526 MachineFunction *F = BB->getParent();
20527 const X86InstrInfo *TII = Subtarget->getInstrInfo();
20528 DebugLoc DL = MI->getDebugLoc();
20530 assert(Subtarget->isTargetDarwin() && "Darwin only instr emitted?");
20531 assert(MI->getOperand(3).isGlobal() && "This should be a global");
20533 // Get a register mask for the lowered call.
20534 // FIXME: The 32-bit calls have non-standard calling conventions. Use a
20535 // proper register mask.
20536 const uint32_t *RegMask =
20537 Subtarget->getRegisterInfo()->getCallPreservedMask(*F, CallingConv::C);
20538 if (Subtarget->is64Bit()) {
20539 MachineInstrBuilder MIB = BuildMI(*BB, MI, DL,
20540 TII->get(X86::MOV64rm), X86::RDI)
20542 .addImm(0).addReg(0)
20543 .addGlobalAddress(MI->getOperand(3).getGlobal(), 0,
20544 MI->getOperand(3).getTargetFlags())
20546 MIB = BuildMI(*BB, MI, DL, TII->get(X86::CALL64m));
20547 addDirectMem(MIB, X86::RDI);
20548 MIB.addReg(X86::RAX, RegState::ImplicitDefine).addRegMask(RegMask);
20549 } else if (F->getTarget().getRelocationModel() != Reloc::PIC_) {
20550 MachineInstrBuilder MIB = BuildMI(*BB, MI, DL,
20551 TII->get(X86::MOV32rm), X86::EAX)
20553 .addImm(0).addReg(0)
20554 .addGlobalAddress(MI->getOperand(3).getGlobal(), 0,
20555 MI->getOperand(3).getTargetFlags())
20557 MIB = BuildMI(*BB, MI, DL, TII->get(X86::CALL32m));
20558 addDirectMem(MIB, X86::EAX);
20559 MIB.addReg(X86::EAX, RegState::ImplicitDefine).addRegMask(RegMask);
20561 MachineInstrBuilder MIB = BuildMI(*BB, MI, DL,
20562 TII->get(X86::MOV32rm), X86::EAX)
20563 .addReg(TII->getGlobalBaseReg(F))
20564 .addImm(0).addReg(0)
20565 .addGlobalAddress(MI->getOperand(3).getGlobal(), 0,
20566 MI->getOperand(3).getTargetFlags())
20568 MIB = BuildMI(*BB, MI, DL, TII->get(X86::CALL32m));
20569 addDirectMem(MIB, X86::EAX);
20570 MIB.addReg(X86::EAX, RegState::ImplicitDefine).addRegMask(RegMask);
20573 MI->eraseFromParent(); // The pseudo instruction is gone now.
20577 MachineBasicBlock *
20578 X86TargetLowering::emitEHSjLjSetJmp(MachineInstr *MI,
20579 MachineBasicBlock *MBB) const {
20580 DebugLoc DL = MI->getDebugLoc();
20581 MachineFunction *MF = MBB->getParent();
20582 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
20583 MachineRegisterInfo &MRI = MF->getRegInfo();
20585 const BasicBlock *BB = MBB->getBasicBlock();
20586 MachineFunction::iterator I = MBB;
20589 // Memory Reference
20590 MachineInstr::mmo_iterator MMOBegin = MI->memoperands_begin();
20591 MachineInstr::mmo_iterator MMOEnd = MI->memoperands_end();
20594 unsigned MemOpndSlot = 0;
20596 unsigned CurOp = 0;
20598 DstReg = MI->getOperand(CurOp++).getReg();
20599 const TargetRegisterClass *RC = MRI.getRegClass(DstReg);
20600 assert(RC->hasType(MVT::i32) && "Invalid destination!");
20601 unsigned mainDstReg = MRI.createVirtualRegister(RC);
20602 unsigned restoreDstReg = MRI.createVirtualRegister(RC);
20604 MemOpndSlot = CurOp;
20606 MVT PVT = getPointerTy(MF->getDataLayout());
20607 assert((PVT == MVT::i64 || PVT == MVT::i32) &&
20608 "Invalid Pointer Size!");
20610 // For v = setjmp(buf), we generate
20613 // buf[LabelOffset] = restoreMBB
20614 // SjLjSetup restoreMBB
20620 // v = phi(main, restore)
20623 // if base pointer being used, load it from frame
20626 MachineBasicBlock *thisMBB = MBB;
20627 MachineBasicBlock *mainMBB = MF->CreateMachineBasicBlock(BB);
20628 MachineBasicBlock *sinkMBB = MF->CreateMachineBasicBlock(BB);
20629 MachineBasicBlock *restoreMBB = MF->CreateMachineBasicBlock(BB);
20630 MF->insert(I, mainMBB);
20631 MF->insert(I, sinkMBB);
20632 MF->push_back(restoreMBB);
20634 MachineInstrBuilder MIB;
20636 // Transfer the remainder of BB and its successor edges to sinkMBB.
20637 sinkMBB->splice(sinkMBB->begin(), MBB,
20638 std::next(MachineBasicBlock::iterator(MI)), MBB->end());
20639 sinkMBB->transferSuccessorsAndUpdatePHIs(MBB);
20642 unsigned PtrStoreOpc = 0;
20643 unsigned LabelReg = 0;
20644 const int64_t LabelOffset = 1 * PVT.getStoreSize();
20645 Reloc::Model RM = MF->getTarget().getRelocationModel();
20646 bool UseImmLabel = (MF->getTarget().getCodeModel() == CodeModel::Small) &&
20647 (RM == Reloc::Static || RM == Reloc::DynamicNoPIC);
20649 // Prepare IP either in reg or imm.
20650 if (!UseImmLabel) {
20651 PtrStoreOpc = (PVT == MVT::i64) ? X86::MOV64mr : X86::MOV32mr;
20652 const TargetRegisterClass *PtrRC = getRegClassFor(PVT);
20653 LabelReg = MRI.createVirtualRegister(PtrRC);
20654 if (Subtarget->is64Bit()) {
20655 MIB = BuildMI(*thisMBB, MI, DL, TII->get(X86::LEA64r), LabelReg)
20659 .addMBB(restoreMBB)
20662 const X86InstrInfo *XII = static_cast<const X86InstrInfo*>(TII);
20663 MIB = BuildMI(*thisMBB, MI, DL, TII->get(X86::LEA32r), LabelReg)
20664 .addReg(XII->getGlobalBaseReg(MF))
20667 .addMBB(restoreMBB, Subtarget->ClassifyBlockAddressReference())
20671 PtrStoreOpc = (PVT == MVT::i64) ? X86::MOV64mi32 : X86::MOV32mi;
20673 MIB = BuildMI(*thisMBB, MI, DL, TII->get(PtrStoreOpc));
20674 for (unsigned i = 0; i < X86::AddrNumOperands; ++i) {
20675 if (i == X86::AddrDisp)
20676 MIB.addDisp(MI->getOperand(MemOpndSlot + i), LabelOffset);
20678 MIB.addOperand(MI->getOperand(MemOpndSlot + i));
20681 MIB.addReg(LabelReg);
20683 MIB.addMBB(restoreMBB);
20684 MIB.setMemRefs(MMOBegin, MMOEnd);
20686 MIB = BuildMI(*thisMBB, MI, DL, TII->get(X86::EH_SjLj_Setup))
20687 .addMBB(restoreMBB);
20689 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
20690 MIB.addRegMask(RegInfo->getNoPreservedMask());
20691 thisMBB->addSuccessor(mainMBB);
20692 thisMBB->addSuccessor(restoreMBB);
20696 BuildMI(mainMBB, DL, TII->get(X86::MOV32r0), mainDstReg);
20697 mainMBB->addSuccessor(sinkMBB);
20700 BuildMI(*sinkMBB, sinkMBB->begin(), DL,
20701 TII->get(X86::PHI), DstReg)
20702 .addReg(mainDstReg).addMBB(mainMBB)
20703 .addReg(restoreDstReg).addMBB(restoreMBB);
20706 if (RegInfo->hasBasePointer(*MF)) {
20707 const bool Uses64BitFramePtr =
20708 Subtarget->isTarget64BitLP64() || Subtarget->isTargetNaCl64();
20709 X86MachineFunctionInfo *X86FI = MF->getInfo<X86MachineFunctionInfo>();
20710 X86FI->setRestoreBasePointer(MF);
20711 unsigned FramePtr = RegInfo->getFrameRegister(*MF);
20712 unsigned BasePtr = RegInfo->getBaseRegister();
20713 unsigned Opm = Uses64BitFramePtr ? X86::MOV64rm : X86::MOV32rm;
20714 addRegOffset(BuildMI(restoreMBB, DL, TII->get(Opm), BasePtr),
20715 FramePtr, true, X86FI->getRestoreBasePointerOffset())
20716 .setMIFlag(MachineInstr::FrameSetup);
20718 BuildMI(restoreMBB, DL, TII->get(X86::MOV32ri), restoreDstReg).addImm(1);
20719 BuildMI(restoreMBB, DL, TII->get(X86::JMP_1)).addMBB(sinkMBB);
20720 restoreMBB->addSuccessor(sinkMBB);
20722 MI->eraseFromParent();
20726 MachineBasicBlock *
20727 X86TargetLowering::emitEHSjLjLongJmp(MachineInstr *MI,
20728 MachineBasicBlock *MBB) const {
20729 DebugLoc DL = MI->getDebugLoc();
20730 MachineFunction *MF = MBB->getParent();
20731 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
20732 MachineRegisterInfo &MRI = MF->getRegInfo();
20734 // Memory Reference
20735 MachineInstr::mmo_iterator MMOBegin = MI->memoperands_begin();
20736 MachineInstr::mmo_iterator MMOEnd = MI->memoperands_end();
20738 MVT PVT = getPointerTy(MF->getDataLayout());
20739 assert((PVT == MVT::i64 || PVT == MVT::i32) &&
20740 "Invalid Pointer Size!");
20742 const TargetRegisterClass *RC =
20743 (PVT == MVT::i64) ? &X86::GR64RegClass : &X86::GR32RegClass;
20744 unsigned Tmp = MRI.createVirtualRegister(RC);
20745 // Since FP is only updated here but NOT referenced, it's treated as GPR.
20746 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
20747 unsigned FP = (PVT == MVT::i64) ? X86::RBP : X86::EBP;
20748 unsigned SP = RegInfo->getStackRegister();
20750 MachineInstrBuilder MIB;
20752 const int64_t LabelOffset = 1 * PVT.getStoreSize();
20753 const int64_t SPOffset = 2 * PVT.getStoreSize();
20755 unsigned PtrLoadOpc = (PVT == MVT::i64) ? X86::MOV64rm : X86::MOV32rm;
20756 unsigned IJmpOpc = (PVT == MVT::i64) ? X86::JMP64r : X86::JMP32r;
20759 MIB = BuildMI(*MBB, MI, DL, TII->get(PtrLoadOpc), FP);
20760 for (unsigned i = 0; i < X86::AddrNumOperands; ++i)
20761 MIB.addOperand(MI->getOperand(i));
20762 MIB.setMemRefs(MMOBegin, MMOEnd);
20764 MIB = BuildMI(*MBB, MI, DL, TII->get(PtrLoadOpc), Tmp);
20765 for (unsigned i = 0; i < X86::AddrNumOperands; ++i) {
20766 if (i == X86::AddrDisp)
20767 MIB.addDisp(MI->getOperand(i), LabelOffset);
20769 MIB.addOperand(MI->getOperand(i));
20771 MIB.setMemRefs(MMOBegin, MMOEnd);
20773 MIB = BuildMI(*MBB, MI, DL, TII->get(PtrLoadOpc), SP);
20774 for (unsigned i = 0; i < X86::AddrNumOperands; ++i) {
20775 if (i == X86::AddrDisp)
20776 MIB.addDisp(MI->getOperand(i), SPOffset);
20778 MIB.addOperand(MI->getOperand(i));
20780 MIB.setMemRefs(MMOBegin, MMOEnd);
20782 BuildMI(*MBB, MI, DL, TII->get(IJmpOpc)).addReg(Tmp);
20784 MI->eraseFromParent();
20788 // Replace 213-type (isel default) FMA3 instructions with 231-type for
20789 // accumulator loops. Writing back to the accumulator allows the coalescer
20790 // to remove extra copies in the loop.
20791 // FIXME: Do this on AVX512. We don't support 231 variants yet (PR23937).
20792 MachineBasicBlock *
20793 X86TargetLowering::emitFMA3Instr(MachineInstr *MI,
20794 MachineBasicBlock *MBB) const {
20795 MachineOperand &AddendOp = MI->getOperand(3);
20797 // Bail out early if the addend isn't a register - we can't switch these.
20798 if (!AddendOp.isReg())
20801 MachineFunction &MF = *MBB->getParent();
20802 MachineRegisterInfo &MRI = MF.getRegInfo();
20804 // Check whether the addend is defined by a PHI:
20805 assert(MRI.hasOneDef(AddendOp.getReg()) && "Multiple defs in SSA?");
20806 MachineInstr &AddendDef = *MRI.def_instr_begin(AddendOp.getReg());
20807 if (!AddendDef.isPHI())
20810 // Look for the following pattern:
20812 // %addend = phi [%entry, 0], [%loop, %result]
20814 // %result<tied1> = FMA213 %m2<tied0>, %m1, %addend
20818 // %addend = phi [%entry, 0], [%loop, %result]
20820 // %result<tied1> = FMA231 %addend<tied0>, %m1, %m2
20822 for (unsigned i = 1, e = AddendDef.getNumOperands(); i < e; i += 2) {
20823 assert(AddendDef.getOperand(i).isReg());
20824 MachineOperand PHISrcOp = AddendDef.getOperand(i);
20825 MachineInstr &PHISrcInst = *MRI.def_instr_begin(PHISrcOp.getReg());
20826 if (&PHISrcInst == MI) {
20827 // Found a matching instruction.
20828 unsigned NewFMAOpc = 0;
20829 switch (MI->getOpcode()) {
20830 case X86::VFMADDPDr213r: NewFMAOpc = X86::VFMADDPDr231r; break;
20831 case X86::VFMADDPSr213r: NewFMAOpc = X86::VFMADDPSr231r; break;
20832 case X86::VFMADDSDr213r: NewFMAOpc = X86::VFMADDSDr231r; break;
20833 case X86::VFMADDSSr213r: NewFMAOpc = X86::VFMADDSSr231r; break;
20834 case X86::VFMSUBPDr213r: NewFMAOpc = X86::VFMSUBPDr231r; break;
20835 case X86::VFMSUBPSr213r: NewFMAOpc = X86::VFMSUBPSr231r; break;
20836 case X86::VFMSUBSDr213r: NewFMAOpc = X86::VFMSUBSDr231r; break;
20837 case X86::VFMSUBSSr213r: NewFMAOpc = X86::VFMSUBSSr231r; break;
20838 case X86::VFNMADDPDr213r: NewFMAOpc = X86::VFNMADDPDr231r; break;
20839 case X86::VFNMADDPSr213r: NewFMAOpc = X86::VFNMADDPSr231r; break;
20840 case X86::VFNMADDSDr213r: NewFMAOpc = X86::VFNMADDSDr231r; break;
20841 case X86::VFNMADDSSr213r: NewFMAOpc = X86::VFNMADDSSr231r; break;
20842 case X86::VFNMSUBPDr213r: NewFMAOpc = X86::VFNMSUBPDr231r; break;
20843 case X86::VFNMSUBPSr213r: NewFMAOpc = X86::VFNMSUBPSr231r; break;
20844 case X86::VFNMSUBSDr213r: NewFMAOpc = X86::VFNMSUBSDr231r; break;
20845 case X86::VFNMSUBSSr213r: NewFMAOpc = X86::VFNMSUBSSr231r; break;
20846 case X86::VFMADDSUBPDr213r: NewFMAOpc = X86::VFMADDSUBPDr231r; break;
20847 case X86::VFMADDSUBPSr213r: NewFMAOpc = X86::VFMADDSUBPSr231r; break;
20848 case X86::VFMSUBADDPDr213r: NewFMAOpc = X86::VFMSUBADDPDr231r; break;
20849 case X86::VFMSUBADDPSr213r: NewFMAOpc = X86::VFMSUBADDPSr231r; break;
20851 case X86::VFMADDPDr213rY: NewFMAOpc = X86::VFMADDPDr231rY; break;
20852 case X86::VFMADDPSr213rY: NewFMAOpc = X86::VFMADDPSr231rY; break;
20853 case X86::VFMSUBPDr213rY: NewFMAOpc = X86::VFMSUBPDr231rY; break;
20854 case X86::VFMSUBPSr213rY: NewFMAOpc = X86::VFMSUBPSr231rY; break;
20855 case X86::VFNMADDPDr213rY: NewFMAOpc = X86::VFNMADDPDr231rY; break;
20856 case X86::VFNMADDPSr213rY: NewFMAOpc = X86::VFNMADDPSr231rY; break;
20857 case X86::VFNMSUBPDr213rY: NewFMAOpc = X86::VFNMSUBPDr231rY; break;
20858 case X86::VFNMSUBPSr213rY: NewFMAOpc = X86::VFNMSUBPSr231rY; break;
20859 case X86::VFMADDSUBPDr213rY: NewFMAOpc = X86::VFMADDSUBPDr231rY; break;
20860 case X86::VFMADDSUBPSr213rY: NewFMAOpc = X86::VFMADDSUBPSr231rY; break;
20861 case X86::VFMSUBADDPDr213rY: NewFMAOpc = X86::VFMSUBADDPDr231rY; break;
20862 case X86::VFMSUBADDPSr213rY: NewFMAOpc = X86::VFMSUBADDPSr231rY; break;
20863 default: llvm_unreachable("Unrecognized FMA variant.");
20866 const TargetInstrInfo &TII = *Subtarget->getInstrInfo();
20867 MachineInstrBuilder MIB =
20868 BuildMI(MF, MI->getDebugLoc(), TII.get(NewFMAOpc))
20869 .addOperand(MI->getOperand(0))
20870 .addOperand(MI->getOperand(3))
20871 .addOperand(MI->getOperand(2))
20872 .addOperand(MI->getOperand(1));
20873 MBB->insert(MachineBasicBlock::iterator(MI), MIB);
20874 MI->eraseFromParent();
20881 MachineBasicBlock *
20882 X86TargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
20883 MachineBasicBlock *BB) const {
20884 switch (MI->getOpcode()) {
20885 default: llvm_unreachable("Unexpected instr type to insert");
20886 case X86::TAILJMPd64:
20887 case X86::TAILJMPr64:
20888 case X86::TAILJMPm64:
20889 case X86::TAILJMPd64_REX:
20890 case X86::TAILJMPr64_REX:
20891 case X86::TAILJMPm64_REX:
20892 llvm_unreachable("TAILJMP64 would not be touched here.");
20893 case X86::TCRETURNdi64:
20894 case X86::TCRETURNri64:
20895 case X86::TCRETURNmi64:
20897 case X86::WIN_ALLOCA:
20898 return EmitLoweredWinAlloca(MI, BB);
20899 case X86::SEG_ALLOCA_32:
20900 case X86::SEG_ALLOCA_64:
20901 return EmitLoweredSegAlloca(MI, BB);
20902 case X86::TLSCall_32:
20903 case X86::TLSCall_64:
20904 return EmitLoweredTLSCall(MI, BB);
20905 case X86::CMOV_FR32:
20906 case X86::CMOV_FR64:
20907 case X86::CMOV_GR8:
20908 case X86::CMOV_GR16:
20909 case X86::CMOV_GR32:
20910 case X86::CMOV_RFP32:
20911 case X86::CMOV_RFP64:
20912 case X86::CMOV_RFP80:
20913 case X86::CMOV_V2F64:
20914 case X86::CMOV_V2I64:
20915 case X86::CMOV_V4F32:
20916 case X86::CMOV_V4F64:
20917 case X86::CMOV_V4I64:
20918 case X86::CMOV_V16F32:
20919 case X86::CMOV_V8F32:
20920 case X86::CMOV_V8F64:
20921 case X86::CMOV_V8I64:
20922 case X86::CMOV_V8I1:
20923 case X86::CMOV_V16I1:
20924 case X86::CMOV_V32I1:
20925 case X86::CMOV_V64I1:
20926 return EmitLoweredSelect(MI, BB);
20928 case X86::RELEASE_FADD32mr:
20929 case X86::RELEASE_FADD64mr:
20930 return EmitLoweredAtomicFP(MI, BB);
20932 case X86::FP32_TO_INT16_IN_MEM:
20933 case X86::FP32_TO_INT32_IN_MEM:
20934 case X86::FP32_TO_INT64_IN_MEM:
20935 case X86::FP64_TO_INT16_IN_MEM:
20936 case X86::FP64_TO_INT32_IN_MEM:
20937 case X86::FP64_TO_INT64_IN_MEM:
20938 case X86::FP80_TO_INT16_IN_MEM:
20939 case X86::FP80_TO_INT32_IN_MEM:
20940 case X86::FP80_TO_INT64_IN_MEM: {
20941 MachineFunction *F = BB->getParent();
20942 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
20943 DebugLoc DL = MI->getDebugLoc();
20945 // Change the floating point control register to use "round towards zero"
20946 // mode when truncating to an integer value.
20947 int CWFrameIdx = F->getFrameInfo()->CreateStackObject(2, 2, false);
20948 addFrameReference(BuildMI(*BB, MI, DL,
20949 TII->get(X86::FNSTCW16m)), CWFrameIdx);
20951 // Load the old value of the high byte of the control word...
20953 F->getRegInfo().createVirtualRegister(&X86::GR16RegClass);
20954 addFrameReference(BuildMI(*BB, MI, DL, TII->get(X86::MOV16rm), OldCW),
20957 // Set the high part to be round to zero...
20958 addFrameReference(BuildMI(*BB, MI, DL, TII->get(X86::MOV16mi)), CWFrameIdx)
20961 // Reload the modified control word now...
20962 addFrameReference(BuildMI(*BB, MI, DL,
20963 TII->get(X86::FLDCW16m)), CWFrameIdx);
20965 // Restore the memory image of control word to original value
20966 addFrameReference(BuildMI(*BB, MI, DL, TII->get(X86::MOV16mr)), CWFrameIdx)
20969 // Get the X86 opcode to use.
20971 switch (MI->getOpcode()) {
20972 default: llvm_unreachable("illegal opcode!");
20973 case X86::FP32_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m32; break;
20974 case X86::FP32_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m32; break;
20975 case X86::FP32_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m32; break;
20976 case X86::FP64_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m64; break;
20977 case X86::FP64_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m64; break;
20978 case X86::FP64_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m64; break;
20979 case X86::FP80_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m80; break;
20980 case X86::FP80_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m80; break;
20981 case X86::FP80_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m80; break;
20985 MachineOperand &Op = MI->getOperand(0);
20987 AM.BaseType = X86AddressMode::RegBase;
20988 AM.Base.Reg = Op.getReg();
20990 AM.BaseType = X86AddressMode::FrameIndexBase;
20991 AM.Base.FrameIndex = Op.getIndex();
20993 Op = MI->getOperand(1);
20995 AM.Scale = Op.getImm();
20996 Op = MI->getOperand(2);
20998 AM.IndexReg = Op.getImm();
20999 Op = MI->getOperand(3);
21000 if (Op.isGlobal()) {
21001 AM.GV = Op.getGlobal();
21003 AM.Disp = Op.getImm();
21005 addFullAddress(BuildMI(*BB, MI, DL, TII->get(Opc)), AM)
21006 .addReg(MI->getOperand(X86::AddrNumOperands).getReg());
21008 // Reload the original control word now.
21009 addFrameReference(BuildMI(*BB, MI, DL,
21010 TII->get(X86::FLDCW16m)), CWFrameIdx);
21012 MI->eraseFromParent(); // The pseudo instruction is gone now.
21015 // String/text processing lowering.
21016 case X86::PCMPISTRM128REG:
21017 case X86::VPCMPISTRM128REG:
21018 case X86::PCMPISTRM128MEM:
21019 case X86::VPCMPISTRM128MEM:
21020 case X86::PCMPESTRM128REG:
21021 case X86::VPCMPESTRM128REG:
21022 case X86::PCMPESTRM128MEM:
21023 case X86::VPCMPESTRM128MEM:
21024 assert(Subtarget->hasSSE42() &&
21025 "Target must have SSE4.2 or AVX features enabled");
21026 return EmitPCMPSTRM(MI, BB, Subtarget->getInstrInfo());
21028 // String/text processing lowering.
21029 case X86::PCMPISTRIREG:
21030 case X86::VPCMPISTRIREG:
21031 case X86::PCMPISTRIMEM:
21032 case X86::VPCMPISTRIMEM:
21033 case X86::PCMPESTRIREG:
21034 case X86::VPCMPESTRIREG:
21035 case X86::PCMPESTRIMEM:
21036 case X86::VPCMPESTRIMEM:
21037 assert(Subtarget->hasSSE42() &&
21038 "Target must have SSE4.2 or AVX features enabled");
21039 return EmitPCMPSTRI(MI, BB, Subtarget->getInstrInfo());
21041 // Thread synchronization.
21043 return EmitMonitor(MI, BB, Subtarget);
21047 return EmitXBegin(MI, BB, Subtarget->getInstrInfo());
21049 case X86::VASTART_SAVE_XMM_REGS:
21050 return EmitVAStartSaveXMMRegsWithCustomInserter(MI, BB);
21052 case X86::VAARG_64:
21053 return EmitVAARG64WithCustomInserter(MI, BB);
21055 case X86::EH_SjLj_SetJmp32:
21056 case X86::EH_SjLj_SetJmp64:
21057 return emitEHSjLjSetJmp(MI, BB);
21059 case X86::EH_SjLj_LongJmp32:
21060 case X86::EH_SjLj_LongJmp64:
21061 return emitEHSjLjLongJmp(MI, BB);
21063 case TargetOpcode::STATEPOINT:
21064 // As an implementation detail, STATEPOINT shares the STACKMAP format at
21065 // this point in the process. We diverge later.
21066 return emitPatchPoint(MI, BB);
21068 case TargetOpcode::STACKMAP:
21069 case TargetOpcode::PATCHPOINT:
21070 return emitPatchPoint(MI, BB);
21072 case X86::VFMADDPDr213r:
21073 case X86::VFMADDPSr213r:
21074 case X86::VFMADDSDr213r:
21075 case X86::VFMADDSSr213r:
21076 case X86::VFMSUBPDr213r:
21077 case X86::VFMSUBPSr213r:
21078 case X86::VFMSUBSDr213r:
21079 case X86::VFMSUBSSr213r:
21080 case X86::VFNMADDPDr213r:
21081 case X86::VFNMADDPSr213r:
21082 case X86::VFNMADDSDr213r:
21083 case X86::VFNMADDSSr213r:
21084 case X86::VFNMSUBPDr213r:
21085 case X86::VFNMSUBPSr213r:
21086 case X86::VFNMSUBSDr213r:
21087 case X86::VFNMSUBSSr213r:
21088 case X86::VFMADDSUBPDr213r:
21089 case X86::VFMADDSUBPSr213r:
21090 case X86::VFMSUBADDPDr213r:
21091 case X86::VFMSUBADDPSr213r:
21092 case X86::VFMADDPDr213rY:
21093 case X86::VFMADDPSr213rY:
21094 case X86::VFMSUBPDr213rY:
21095 case X86::VFMSUBPSr213rY:
21096 case X86::VFNMADDPDr213rY:
21097 case X86::VFNMADDPSr213rY:
21098 case X86::VFNMSUBPDr213rY:
21099 case X86::VFNMSUBPSr213rY:
21100 case X86::VFMADDSUBPDr213rY:
21101 case X86::VFMADDSUBPSr213rY:
21102 case X86::VFMSUBADDPDr213rY:
21103 case X86::VFMSUBADDPSr213rY:
21104 return emitFMA3Instr(MI, BB);
21108 //===----------------------------------------------------------------------===//
21109 // X86 Optimization Hooks
21110 //===----------------------------------------------------------------------===//
21112 void X86TargetLowering::computeKnownBitsForTargetNode(const SDValue Op,
21115 const SelectionDAG &DAG,
21116 unsigned Depth) const {
21117 unsigned BitWidth = KnownZero.getBitWidth();
21118 unsigned Opc = Op.getOpcode();
21119 assert((Opc >= ISD::BUILTIN_OP_END ||
21120 Opc == ISD::INTRINSIC_WO_CHAIN ||
21121 Opc == ISD::INTRINSIC_W_CHAIN ||
21122 Opc == ISD::INTRINSIC_VOID) &&
21123 "Should use MaskedValueIsZero if you don't know whether Op"
21124 " is a target node!");
21126 KnownZero = KnownOne = APInt(BitWidth, 0); // Don't know anything.
21140 // These nodes' second result is a boolean.
21141 if (Op.getResNo() == 0)
21144 case X86ISD::SETCC:
21145 KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - 1);
21147 case ISD::INTRINSIC_WO_CHAIN: {
21148 unsigned IntId = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
21149 unsigned NumLoBits = 0;
21152 case Intrinsic::x86_sse_movmsk_ps:
21153 case Intrinsic::x86_avx_movmsk_ps_256:
21154 case Intrinsic::x86_sse2_movmsk_pd:
21155 case Intrinsic::x86_avx_movmsk_pd_256:
21156 case Intrinsic::x86_mmx_pmovmskb:
21157 case Intrinsic::x86_sse2_pmovmskb_128:
21158 case Intrinsic::x86_avx2_pmovmskb: {
21159 // High bits of movmskp{s|d}, pmovmskb are known zero.
21161 default: llvm_unreachable("Impossible intrinsic"); // Can't reach here.
21162 case Intrinsic::x86_sse_movmsk_ps: NumLoBits = 4; break;
21163 case Intrinsic::x86_avx_movmsk_ps_256: NumLoBits = 8; break;
21164 case Intrinsic::x86_sse2_movmsk_pd: NumLoBits = 2; break;
21165 case Intrinsic::x86_avx_movmsk_pd_256: NumLoBits = 4; break;
21166 case Intrinsic::x86_mmx_pmovmskb: NumLoBits = 8; break;
21167 case Intrinsic::x86_sse2_pmovmskb_128: NumLoBits = 16; break;
21168 case Intrinsic::x86_avx2_pmovmskb: NumLoBits = 32; break;
21170 KnownZero = APInt::getHighBitsSet(BitWidth, BitWidth - NumLoBits);
21179 unsigned X86TargetLowering::ComputeNumSignBitsForTargetNode(
21181 const SelectionDAG &,
21182 unsigned Depth) const {
21183 // SETCC_CARRY sets the dest to ~0 for true or 0 for false.
21184 if (Op.getOpcode() == X86ISD::SETCC_CARRY)
21185 return Op.getValueType().getScalarType().getSizeInBits();
21191 /// isGAPlusOffset - Returns true (and the GlobalValue and the offset) if the
21192 /// node is a GlobalAddress + offset.
21193 bool X86TargetLowering::isGAPlusOffset(SDNode *N,
21194 const GlobalValue* &GA,
21195 int64_t &Offset) const {
21196 if (N->getOpcode() == X86ISD::Wrapper) {
21197 if (isa<GlobalAddressSDNode>(N->getOperand(0))) {
21198 GA = cast<GlobalAddressSDNode>(N->getOperand(0))->getGlobal();
21199 Offset = cast<GlobalAddressSDNode>(N->getOperand(0))->getOffset();
21203 return TargetLowering::isGAPlusOffset(N, GA, Offset);
21206 /// isShuffleHigh128VectorInsertLow - Checks whether the shuffle node is the
21207 /// same as extracting the high 128-bit part of 256-bit vector and then
21208 /// inserting the result into the low part of a new 256-bit vector
21209 static bool isShuffleHigh128VectorInsertLow(ShuffleVectorSDNode *SVOp) {
21210 EVT VT = SVOp->getValueType(0);
21211 unsigned NumElems = VT.getVectorNumElements();
21213 // vector_shuffle <4, 5, 6, 7, u, u, u, u> or <2, 3, u, u>
21214 for (unsigned i = 0, j = NumElems/2; i != NumElems/2; ++i, ++j)
21215 if (!isUndefOrEqual(SVOp->getMaskElt(i), j) ||
21216 SVOp->getMaskElt(j) >= 0)
21222 /// isShuffleLow128VectorInsertHigh - Checks whether the shuffle node is the
21223 /// same as extracting the low 128-bit part of 256-bit vector and then
21224 /// inserting the result into the high part of a new 256-bit vector
21225 static bool isShuffleLow128VectorInsertHigh(ShuffleVectorSDNode *SVOp) {
21226 EVT VT = SVOp->getValueType(0);
21227 unsigned NumElems = VT.getVectorNumElements();
21229 // vector_shuffle <u, u, u, u, 0, 1, 2, 3> or <u, u, 0, 1>
21230 for (unsigned i = NumElems/2, j = 0; i != NumElems; ++i, ++j)
21231 if (!isUndefOrEqual(SVOp->getMaskElt(i), j) ||
21232 SVOp->getMaskElt(j) >= 0)
21238 /// PerformShuffleCombine256 - Performs shuffle combines for 256-bit vectors.
21239 static SDValue PerformShuffleCombine256(SDNode *N, SelectionDAG &DAG,
21240 TargetLowering::DAGCombinerInfo &DCI,
21241 const X86Subtarget* Subtarget) {
21243 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
21244 SDValue V1 = SVOp->getOperand(0);
21245 SDValue V2 = SVOp->getOperand(1);
21246 EVT VT = SVOp->getValueType(0);
21247 unsigned NumElems = VT.getVectorNumElements();
21249 if (V1.getOpcode() == ISD::CONCAT_VECTORS &&
21250 V2.getOpcode() == ISD::CONCAT_VECTORS) {
21254 // V UNDEF BUILD_VECTOR UNDEF
21256 // CONCAT_VECTOR CONCAT_VECTOR
21259 // RESULT: V + zero extended
21261 if (V2.getOperand(0).getOpcode() != ISD::BUILD_VECTOR ||
21262 V2.getOperand(1).getOpcode() != ISD::UNDEF ||
21263 V1.getOperand(1).getOpcode() != ISD::UNDEF)
21266 if (!ISD::isBuildVectorAllZeros(V2.getOperand(0).getNode()))
21269 // To match the shuffle mask, the first half of the mask should
21270 // be exactly the first vector, and all the rest a splat with the
21271 // first element of the second one.
21272 for (unsigned i = 0; i != NumElems/2; ++i)
21273 if (!isUndefOrEqual(SVOp->getMaskElt(i), i) ||
21274 !isUndefOrEqual(SVOp->getMaskElt(i+NumElems/2), NumElems))
21277 // If V1 is coming from a vector load then just fold to a VZEXT_LOAD.
21278 if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(V1.getOperand(0))) {
21279 if (Ld->hasNUsesOfValue(1, 0)) {
21280 SDVTList Tys = DAG.getVTList(MVT::v4i64, MVT::Other);
21281 SDValue Ops[] = { Ld->getChain(), Ld->getBasePtr() };
21283 DAG.getMemIntrinsicNode(X86ISD::VZEXT_LOAD, dl, Tys, Ops,
21285 Ld->getPointerInfo(),
21286 Ld->getAlignment(),
21287 false/*isVolatile*/, true/*ReadMem*/,
21288 false/*WriteMem*/);
21290 // Make sure the newly-created LOAD is in the same position as Ld in
21291 // terms of dependency. We create a TokenFactor for Ld and ResNode,
21292 // and update uses of Ld's output chain to use the TokenFactor.
21293 if (Ld->hasAnyUseOfValue(1)) {
21294 SDValue NewChain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
21295 SDValue(Ld, 1), SDValue(ResNode.getNode(), 1));
21296 DAG.ReplaceAllUsesOfValueWith(SDValue(Ld, 1), NewChain);
21297 DAG.UpdateNodeOperands(NewChain.getNode(), SDValue(Ld, 1),
21298 SDValue(ResNode.getNode(), 1));
21301 return DAG.getBitcast(VT, ResNode);
21305 // Emit a zeroed vector and insert the desired subvector on its
21307 SDValue Zeros = getZeroVector(VT, Subtarget, DAG, dl);
21308 SDValue InsV = Insert128BitVector(Zeros, V1.getOperand(0), 0, DAG, dl);
21309 return DCI.CombineTo(N, InsV);
21312 //===--------------------------------------------------------------------===//
21313 // Combine some shuffles into subvector extracts and inserts:
21316 // vector_shuffle <4, 5, 6, 7, u, u, u, u> or <2, 3, u, u>
21317 if (isShuffleHigh128VectorInsertLow(SVOp)) {
21318 SDValue V = Extract128BitVector(V1, NumElems/2, DAG, dl);
21319 SDValue InsV = Insert128BitVector(DAG.getUNDEF(VT), V, 0, DAG, dl);
21320 return DCI.CombineTo(N, InsV);
21323 // vector_shuffle <u, u, u, u, 0, 1, 2, 3> or <u, u, 0, 1>
21324 if (isShuffleLow128VectorInsertHigh(SVOp)) {
21325 SDValue V = Extract128BitVector(V1, 0, DAG, dl);
21326 SDValue InsV = Insert128BitVector(DAG.getUNDEF(VT), V, NumElems/2, DAG, dl);
21327 return DCI.CombineTo(N, InsV);
21333 /// \brief Combine an arbitrary chain of shuffles into a single instruction if
21336 /// This is the leaf of the recursive combinine below. When we have found some
21337 /// chain of single-use x86 shuffle instructions and accumulated the combined
21338 /// shuffle mask represented by them, this will try to pattern match that mask
21339 /// into either a single instruction if there is a special purpose instruction
21340 /// for this operation, or into a PSHUFB instruction which is a fully general
21341 /// instruction but should only be used to replace chains over a certain depth.
21342 static bool combineX86ShuffleChain(SDValue Op, SDValue Root, ArrayRef<int> Mask,
21343 int Depth, bool HasPSHUFB, SelectionDAG &DAG,
21344 TargetLowering::DAGCombinerInfo &DCI,
21345 const X86Subtarget *Subtarget) {
21346 assert(!Mask.empty() && "Cannot combine an empty shuffle mask!");
21348 // Find the operand that enters the chain. Note that multiple uses are OK
21349 // here, we're not going to remove the operand we find.
21350 SDValue Input = Op.getOperand(0);
21351 while (Input.getOpcode() == ISD::BITCAST)
21352 Input = Input.getOperand(0);
21354 MVT VT = Input.getSimpleValueType();
21355 MVT RootVT = Root.getSimpleValueType();
21358 // Just remove no-op shuffle masks.
21359 if (Mask.size() == 1) {
21360 DCI.CombineTo(Root.getNode(), DAG.getBitcast(RootVT, Input),
21365 // Use the float domain if the operand type is a floating point type.
21366 bool FloatDomain = VT.isFloatingPoint();
21368 // For floating point shuffles, we don't have free copies in the shuffle
21369 // instructions or the ability to load as part of the instruction, so
21370 // canonicalize their shuffles to UNPCK or MOV variants.
21372 // Note that even with AVX we prefer the PSHUFD form of shuffle for integer
21373 // vectors because it can have a load folded into it that UNPCK cannot. This
21374 // doesn't preclude something switching to the shorter encoding post-RA.
21376 // FIXME: Should teach these routines about AVX vector widths.
21377 if (FloatDomain && VT.getSizeInBits() == 128) {
21378 if (Mask.equals({0, 0}) || Mask.equals({1, 1})) {
21379 bool Lo = Mask.equals({0, 0});
21382 // Check if we have SSE3 which will let us use MOVDDUP. That instruction
21383 // is no slower than UNPCKLPD but has the option to fold the input operand
21384 // into even an unaligned memory load.
21385 if (Lo && Subtarget->hasSSE3()) {
21386 Shuffle = X86ISD::MOVDDUP;
21387 ShuffleVT = MVT::v2f64;
21389 // We have MOVLHPS and MOVHLPS throughout SSE and they encode smaller
21390 // than the UNPCK variants.
21391 Shuffle = Lo ? X86ISD::MOVLHPS : X86ISD::MOVHLPS;
21392 ShuffleVT = MVT::v4f32;
21394 if (Depth == 1 && Root->getOpcode() == Shuffle)
21395 return false; // Nothing to do!
21396 Op = DAG.getBitcast(ShuffleVT, Input);
21397 DCI.AddToWorklist(Op.getNode());
21398 if (Shuffle == X86ISD::MOVDDUP)
21399 Op = DAG.getNode(Shuffle, DL, ShuffleVT, Op);
21401 Op = DAG.getNode(Shuffle, DL, ShuffleVT, Op, Op);
21402 DCI.AddToWorklist(Op.getNode());
21403 DCI.CombineTo(Root.getNode(), DAG.getBitcast(RootVT, Op),
21407 if (Subtarget->hasSSE3() &&
21408 (Mask.equals({0, 0, 2, 2}) || Mask.equals({1, 1, 3, 3}))) {
21409 bool Lo = Mask.equals({0, 0, 2, 2});
21410 unsigned Shuffle = Lo ? X86ISD::MOVSLDUP : X86ISD::MOVSHDUP;
21411 MVT ShuffleVT = MVT::v4f32;
21412 if (Depth == 1 && Root->getOpcode() == Shuffle)
21413 return false; // Nothing to do!
21414 Op = DAG.getBitcast(ShuffleVT, Input);
21415 DCI.AddToWorklist(Op.getNode());
21416 Op = DAG.getNode(Shuffle, DL, ShuffleVT, Op);
21417 DCI.AddToWorklist(Op.getNode());
21418 DCI.CombineTo(Root.getNode(), DAG.getBitcast(RootVT, Op),
21422 if (Mask.equals({0, 0, 1, 1}) || Mask.equals({2, 2, 3, 3})) {
21423 bool Lo = Mask.equals({0, 0, 1, 1});
21424 unsigned Shuffle = Lo ? X86ISD::UNPCKL : X86ISD::UNPCKH;
21425 MVT ShuffleVT = MVT::v4f32;
21426 if (Depth == 1 && Root->getOpcode() == Shuffle)
21427 return false; // Nothing to do!
21428 Op = DAG.getBitcast(ShuffleVT, Input);
21429 DCI.AddToWorklist(Op.getNode());
21430 Op = DAG.getNode(Shuffle, DL, ShuffleVT, Op, Op);
21431 DCI.AddToWorklist(Op.getNode());
21432 DCI.CombineTo(Root.getNode(), DAG.getBitcast(RootVT, Op),
21438 // We always canonicalize the 8 x i16 and 16 x i8 shuffles into their UNPCK
21439 // variants as none of these have single-instruction variants that are
21440 // superior to the UNPCK formulation.
21441 if (!FloatDomain && VT.getSizeInBits() == 128 &&
21442 (Mask.equals({0, 0, 1, 1, 2, 2, 3, 3}) ||
21443 Mask.equals({4, 4, 5, 5, 6, 6, 7, 7}) ||
21444 Mask.equals({0, 0, 1, 1, 2, 2, 3, 3, 4, 4, 5, 5, 6, 6, 7, 7}) ||
21446 {8, 8, 9, 9, 10, 10, 11, 11, 12, 12, 13, 13, 14, 14, 15, 15}))) {
21447 bool Lo = Mask[0] == 0;
21448 unsigned Shuffle = Lo ? X86ISD::UNPCKL : X86ISD::UNPCKH;
21449 if (Depth == 1 && Root->getOpcode() == Shuffle)
21450 return false; // Nothing to do!
21452 switch (Mask.size()) {
21454 ShuffleVT = MVT::v8i16;
21457 ShuffleVT = MVT::v16i8;
21460 llvm_unreachable("Impossible mask size!");
21462 Op = DAG.getBitcast(ShuffleVT, Input);
21463 DCI.AddToWorklist(Op.getNode());
21464 Op = DAG.getNode(Shuffle, DL, ShuffleVT, Op, Op);
21465 DCI.AddToWorklist(Op.getNode());
21466 DCI.CombineTo(Root.getNode(), DAG.getBitcast(RootVT, Op),
21471 // Don't try to re-form single instruction chains under any circumstances now
21472 // that we've done encoding canonicalization for them.
21476 // If we have 3 or more shuffle instructions or a chain involving PSHUFB, we
21477 // can replace them with a single PSHUFB instruction profitably. Intel's
21478 // manuals suggest only using PSHUFB if doing so replacing 5 instructions, but
21479 // in practice PSHUFB tends to be *very* fast so we're more aggressive.
21480 if ((Depth >= 3 || HasPSHUFB) && Subtarget->hasSSSE3()) {
21481 SmallVector<SDValue, 16> PSHUFBMask;
21482 int NumBytes = VT.getSizeInBits() / 8;
21483 int Ratio = NumBytes / Mask.size();
21484 for (int i = 0; i < NumBytes; ++i) {
21485 if (Mask[i / Ratio] == SM_SentinelUndef) {
21486 PSHUFBMask.push_back(DAG.getUNDEF(MVT::i8));
21489 int M = Mask[i / Ratio] != SM_SentinelZero
21490 ? Ratio * Mask[i / Ratio] + i % Ratio
21492 PSHUFBMask.push_back(DAG.getConstant(M, DL, MVT::i8));
21494 MVT ByteVT = MVT::getVectorVT(MVT::i8, NumBytes);
21495 Op = DAG.getBitcast(ByteVT, Input);
21496 DCI.AddToWorklist(Op.getNode());
21497 SDValue PSHUFBMaskOp =
21498 DAG.getNode(ISD::BUILD_VECTOR, DL, ByteVT, PSHUFBMask);
21499 DCI.AddToWorklist(PSHUFBMaskOp.getNode());
21500 Op = DAG.getNode(X86ISD::PSHUFB, DL, ByteVT, Op, PSHUFBMaskOp);
21501 DCI.AddToWorklist(Op.getNode());
21502 DCI.CombineTo(Root.getNode(), DAG.getBitcast(RootVT, Op),
21507 // Failed to find any combines.
21511 /// \brief Fully generic combining of x86 shuffle instructions.
21513 /// This should be the last combine run over the x86 shuffle instructions. Once
21514 /// they have been fully optimized, this will recursively consider all chains
21515 /// of single-use shuffle instructions, build a generic model of the cumulative
21516 /// shuffle operation, and check for simpler instructions which implement this
21517 /// operation. We use this primarily for two purposes:
21519 /// 1) Collapse generic shuffles to specialized single instructions when
21520 /// equivalent. In most cases, this is just an encoding size win, but
21521 /// sometimes we will collapse multiple generic shuffles into a single
21522 /// special-purpose shuffle.
21523 /// 2) Look for sequences of shuffle instructions with 3 or more total
21524 /// instructions, and replace them with the slightly more expensive SSSE3
21525 /// PSHUFB instruction if available. We do this as the last combining step
21526 /// to ensure we avoid using PSHUFB if we can implement the shuffle with
21527 /// a suitable short sequence of other instructions. The PHUFB will either
21528 /// use a register or have to read from memory and so is slightly (but only
21529 /// slightly) more expensive than the other shuffle instructions.
21531 /// Because this is inherently a quadratic operation (for each shuffle in
21532 /// a chain, we recurse up the chain), the depth is limited to 8 instructions.
21533 /// This should never be an issue in practice as the shuffle lowering doesn't
21534 /// produce sequences of more than 8 instructions.
21536 /// FIXME: We will currently miss some cases where the redundant shuffling
21537 /// would simplify under the threshold for PSHUFB formation because of
21538 /// combine-ordering. To fix this, we should do the redundant instruction
21539 /// combining in this recursive walk.
21540 static bool combineX86ShufflesRecursively(SDValue Op, SDValue Root,
21541 ArrayRef<int> RootMask,
21542 int Depth, bool HasPSHUFB,
21544 TargetLowering::DAGCombinerInfo &DCI,
21545 const X86Subtarget *Subtarget) {
21546 // Bound the depth of our recursive combine because this is ultimately
21547 // quadratic in nature.
21551 // Directly rip through bitcasts to find the underlying operand.
21552 while (Op.getOpcode() == ISD::BITCAST && Op.getOperand(0).hasOneUse())
21553 Op = Op.getOperand(0);
21555 MVT VT = Op.getSimpleValueType();
21556 if (!VT.isVector())
21557 return false; // Bail if we hit a non-vector.
21559 assert(Root.getSimpleValueType().isVector() &&
21560 "Shuffles operate on vector types!");
21561 assert(VT.getSizeInBits() == Root.getSimpleValueType().getSizeInBits() &&
21562 "Can only combine shuffles of the same vector register size.");
21564 if (!isTargetShuffle(Op.getOpcode()))
21566 SmallVector<int, 16> OpMask;
21568 bool HaveMask = getTargetShuffleMask(Op.getNode(), VT, OpMask, IsUnary);
21569 // We only can combine unary shuffles which we can decode the mask for.
21570 if (!HaveMask || !IsUnary)
21573 assert(VT.getVectorNumElements() == OpMask.size() &&
21574 "Different mask size from vector size!");
21575 assert(((RootMask.size() > OpMask.size() &&
21576 RootMask.size() % OpMask.size() == 0) ||
21577 (OpMask.size() > RootMask.size() &&
21578 OpMask.size() % RootMask.size() == 0) ||
21579 OpMask.size() == RootMask.size()) &&
21580 "The smaller number of elements must divide the larger.");
21581 int RootRatio = std::max<int>(1, OpMask.size() / RootMask.size());
21582 int OpRatio = std::max<int>(1, RootMask.size() / OpMask.size());
21583 assert(((RootRatio == 1 && OpRatio == 1) ||
21584 (RootRatio == 1) != (OpRatio == 1)) &&
21585 "Must not have a ratio for both incoming and op masks!");
21587 SmallVector<int, 16> Mask;
21588 Mask.reserve(std::max(OpMask.size(), RootMask.size()));
21590 // Merge this shuffle operation's mask into our accumulated mask. Note that
21591 // this shuffle's mask will be the first applied to the input, followed by the
21592 // root mask to get us all the way to the root value arrangement. The reason
21593 // for this order is that we are recursing up the operation chain.
21594 for (int i = 0, e = std::max(OpMask.size(), RootMask.size()); i < e; ++i) {
21595 int RootIdx = i / RootRatio;
21596 if (RootMask[RootIdx] < 0) {
21597 // This is a zero or undef lane, we're done.
21598 Mask.push_back(RootMask[RootIdx]);
21602 int RootMaskedIdx = RootMask[RootIdx] * RootRatio + i % RootRatio;
21603 int OpIdx = RootMaskedIdx / OpRatio;
21604 if (OpMask[OpIdx] < 0) {
21605 // The incoming lanes are zero or undef, it doesn't matter which ones we
21607 Mask.push_back(OpMask[OpIdx]);
21611 // Ok, we have non-zero lanes, map them through.
21612 Mask.push_back(OpMask[OpIdx] * OpRatio +
21613 RootMaskedIdx % OpRatio);
21616 // See if we can recurse into the operand to combine more things.
21617 switch (Op.getOpcode()) {
21618 case X86ISD::PSHUFB:
21620 case X86ISD::PSHUFD:
21621 case X86ISD::PSHUFHW:
21622 case X86ISD::PSHUFLW:
21623 if (Op.getOperand(0).hasOneUse() &&
21624 combineX86ShufflesRecursively(Op.getOperand(0), Root, Mask, Depth + 1,
21625 HasPSHUFB, DAG, DCI, Subtarget))
21629 case X86ISD::UNPCKL:
21630 case X86ISD::UNPCKH:
21631 assert(Op.getOperand(0) == Op.getOperand(1) && "We only combine unary shuffles!");
21632 // We can't check for single use, we have to check that this shuffle is the only user.
21633 if (Op->isOnlyUserOf(Op.getOperand(0).getNode()) &&
21634 combineX86ShufflesRecursively(Op.getOperand(0), Root, Mask, Depth + 1,
21635 HasPSHUFB, DAG, DCI, Subtarget))
21640 // Minor canonicalization of the accumulated shuffle mask to make it easier
21641 // to match below. All this does is detect masks with squential pairs of
21642 // elements, and shrink them to the half-width mask. It does this in a loop
21643 // so it will reduce the size of the mask to the minimal width mask which
21644 // performs an equivalent shuffle.
21645 SmallVector<int, 16> WidenedMask;
21646 while (Mask.size() > 1 && canWidenShuffleElements(Mask, WidenedMask)) {
21647 Mask = std::move(WidenedMask);
21648 WidenedMask.clear();
21651 return combineX86ShuffleChain(Op, Root, Mask, Depth, HasPSHUFB, DAG, DCI,
21655 /// \brief Get the PSHUF-style mask from PSHUF node.
21657 /// This is a very minor wrapper around getTargetShuffleMask to easy forming v4
21658 /// PSHUF-style masks that can be reused with such instructions.
21659 static SmallVector<int, 4> getPSHUFShuffleMask(SDValue N) {
21660 MVT VT = N.getSimpleValueType();
21661 SmallVector<int, 4> Mask;
21663 bool HaveMask = getTargetShuffleMask(N.getNode(), VT, Mask, IsUnary);
21667 // If we have more than 128-bits, only the low 128-bits of shuffle mask
21668 // matter. Check that the upper masks are repeats and remove them.
21669 if (VT.getSizeInBits() > 128) {
21670 int LaneElts = 128 / VT.getScalarSizeInBits();
21672 for (int i = 1, NumLanes = VT.getSizeInBits() / 128; i < NumLanes; ++i)
21673 for (int j = 0; j < LaneElts; ++j)
21674 assert(Mask[j] == Mask[i * LaneElts + j] - (LaneElts * i) &&
21675 "Mask doesn't repeat in high 128-bit lanes!");
21677 Mask.resize(LaneElts);
21680 switch (N.getOpcode()) {
21681 case X86ISD::PSHUFD:
21683 case X86ISD::PSHUFLW:
21686 case X86ISD::PSHUFHW:
21687 Mask.erase(Mask.begin(), Mask.begin() + 4);
21688 for (int &M : Mask)
21692 llvm_unreachable("No valid shuffle instruction found!");
21696 /// \brief Search for a combinable shuffle across a chain ending in pshufd.
21698 /// We walk up the chain and look for a combinable shuffle, skipping over
21699 /// shuffles that we could hoist this shuffle's transformation past without
21700 /// altering anything.
21702 combineRedundantDWordShuffle(SDValue N, MutableArrayRef<int> Mask,
21704 TargetLowering::DAGCombinerInfo &DCI) {
21705 assert(N.getOpcode() == X86ISD::PSHUFD &&
21706 "Called with something other than an x86 128-bit half shuffle!");
21709 // Walk up a single-use chain looking for a combinable shuffle. Keep a stack
21710 // of the shuffles in the chain so that we can form a fresh chain to replace
21712 SmallVector<SDValue, 8> Chain;
21713 SDValue V = N.getOperand(0);
21714 for (; V.hasOneUse(); V = V.getOperand(0)) {
21715 switch (V.getOpcode()) {
21717 return SDValue(); // Nothing combined!
21720 // Skip bitcasts as we always know the type for the target specific
21724 case X86ISD::PSHUFD:
21725 // Found another dword shuffle.
21728 case X86ISD::PSHUFLW:
21729 // Check that the low words (being shuffled) are the identity in the
21730 // dword shuffle, and the high words are self-contained.
21731 if (Mask[0] != 0 || Mask[1] != 1 ||
21732 !(Mask[2] >= 2 && Mask[2] < 4 && Mask[3] >= 2 && Mask[3] < 4))
21735 Chain.push_back(V);
21738 case X86ISD::PSHUFHW:
21739 // Check that the high words (being shuffled) are the identity in the
21740 // dword shuffle, and the low words are self-contained.
21741 if (Mask[2] != 2 || Mask[3] != 3 ||
21742 !(Mask[0] >= 0 && Mask[0] < 2 && Mask[1] >= 0 && Mask[1] < 2))
21745 Chain.push_back(V);
21748 case X86ISD::UNPCKL:
21749 case X86ISD::UNPCKH:
21750 // For either i8 -> i16 or i16 -> i32 unpacks, we can combine a dword
21751 // shuffle into a preceding word shuffle.
21752 if (V.getSimpleValueType().getScalarType() != MVT::i8 &&
21753 V.getSimpleValueType().getScalarType() != MVT::i16)
21756 // Search for a half-shuffle which we can combine with.
21757 unsigned CombineOp =
21758 V.getOpcode() == X86ISD::UNPCKL ? X86ISD::PSHUFLW : X86ISD::PSHUFHW;
21759 if (V.getOperand(0) != V.getOperand(1) ||
21760 !V->isOnlyUserOf(V.getOperand(0).getNode()))
21762 Chain.push_back(V);
21763 V = V.getOperand(0);
21765 switch (V.getOpcode()) {
21767 return SDValue(); // Nothing to combine.
21769 case X86ISD::PSHUFLW:
21770 case X86ISD::PSHUFHW:
21771 if (V.getOpcode() == CombineOp)
21774 Chain.push_back(V);
21778 V = V.getOperand(0);
21782 } while (V.hasOneUse());
21785 // Break out of the loop if we break out of the switch.
21789 if (!V.hasOneUse())
21790 // We fell out of the loop without finding a viable combining instruction.
21793 // Merge this node's mask and our incoming mask.
21794 SmallVector<int, 4> VMask = getPSHUFShuffleMask(V);
21795 for (int &M : Mask)
21797 V = DAG.getNode(V.getOpcode(), DL, V.getValueType(), V.getOperand(0),
21798 getV4X86ShuffleImm8ForMask(Mask, DL, DAG));
21800 // Rebuild the chain around this new shuffle.
21801 while (!Chain.empty()) {
21802 SDValue W = Chain.pop_back_val();
21804 if (V.getValueType() != W.getOperand(0).getValueType())
21805 V = DAG.getBitcast(W.getOperand(0).getValueType(), V);
21807 switch (W.getOpcode()) {
21809 llvm_unreachable("Only PSHUF and UNPCK instructions get here!");
21811 case X86ISD::UNPCKL:
21812 case X86ISD::UNPCKH:
21813 V = DAG.getNode(W.getOpcode(), DL, W.getValueType(), V, V);
21816 case X86ISD::PSHUFD:
21817 case X86ISD::PSHUFLW:
21818 case X86ISD::PSHUFHW:
21819 V = DAG.getNode(W.getOpcode(), DL, W.getValueType(), V, W.getOperand(1));
21823 if (V.getValueType() != N.getValueType())
21824 V = DAG.getBitcast(N.getValueType(), V);
21826 // Return the new chain to replace N.
21830 /// \brief Search for a combinable shuffle across a chain ending in pshuflw or pshufhw.
21832 /// We walk up the chain, skipping shuffles of the other half and looking
21833 /// through shuffles which switch halves trying to find a shuffle of the same
21834 /// pair of dwords.
21835 static bool combineRedundantHalfShuffle(SDValue N, MutableArrayRef<int> Mask,
21837 TargetLowering::DAGCombinerInfo &DCI) {
21839 (N.getOpcode() == X86ISD::PSHUFLW || N.getOpcode() == X86ISD::PSHUFHW) &&
21840 "Called with something other than an x86 128-bit half shuffle!");
21842 unsigned CombineOpcode = N.getOpcode();
21844 // Walk up a single-use chain looking for a combinable shuffle.
21845 SDValue V = N.getOperand(0);
21846 for (; V.hasOneUse(); V = V.getOperand(0)) {
21847 switch (V.getOpcode()) {
21849 return false; // Nothing combined!
21852 // Skip bitcasts as we always know the type for the target specific
21856 case X86ISD::PSHUFLW:
21857 case X86ISD::PSHUFHW:
21858 if (V.getOpcode() == CombineOpcode)
21861 // Other-half shuffles are no-ops.
21864 // Break out of the loop if we break out of the switch.
21868 if (!V.hasOneUse())
21869 // We fell out of the loop without finding a viable combining instruction.
21872 // Combine away the bottom node as its shuffle will be accumulated into
21873 // a preceding shuffle.
21874 DCI.CombineTo(N.getNode(), N.getOperand(0), /*AddTo*/ true);
21876 // Record the old value.
21879 // Merge this node's mask and our incoming mask (adjusted to account for all
21880 // the pshufd instructions encountered).
21881 SmallVector<int, 4> VMask = getPSHUFShuffleMask(V);
21882 for (int &M : Mask)
21884 V = DAG.getNode(V.getOpcode(), DL, MVT::v8i16, V.getOperand(0),
21885 getV4X86ShuffleImm8ForMask(Mask, DL, DAG));
21887 // Check that the shuffles didn't cancel each other out. If not, we need to
21888 // combine to the new one.
21890 // Replace the combinable shuffle with the combined one, updating all users
21891 // so that we re-evaluate the chain here.
21892 DCI.CombineTo(Old.getNode(), V, /*AddTo*/ true);
21897 /// \brief Try to combine x86 target specific shuffles.
21898 static SDValue PerformTargetShuffleCombine(SDValue N, SelectionDAG &DAG,
21899 TargetLowering::DAGCombinerInfo &DCI,
21900 const X86Subtarget *Subtarget) {
21902 MVT VT = N.getSimpleValueType();
21903 SmallVector<int, 4> Mask;
21905 switch (N.getOpcode()) {
21906 case X86ISD::PSHUFD:
21907 case X86ISD::PSHUFLW:
21908 case X86ISD::PSHUFHW:
21909 Mask = getPSHUFShuffleMask(N);
21910 assert(Mask.size() == 4);
21916 // Nuke no-op shuffles that show up after combining.
21917 if (isNoopShuffleMask(Mask))
21918 return DCI.CombineTo(N.getNode(), N.getOperand(0), /*AddTo*/ true);
21920 // Look for simplifications involving one or two shuffle instructions.
21921 SDValue V = N.getOperand(0);
21922 switch (N.getOpcode()) {
21925 case X86ISD::PSHUFLW:
21926 case X86ISD::PSHUFHW:
21927 assert(VT.getScalarType() == MVT::i16 && "Bad word shuffle type!");
21929 if (combineRedundantHalfShuffle(N, Mask, DAG, DCI))
21930 return SDValue(); // We combined away this shuffle, so we're done.
21932 // See if this reduces to a PSHUFD which is no more expensive and can
21933 // combine with more operations. Note that it has to at least flip the
21934 // dwords as otherwise it would have been removed as a no-op.
21935 if (makeArrayRef(Mask).equals({2, 3, 0, 1})) {
21936 int DMask[] = {0, 1, 2, 3};
21937 int DOffset = N.getOpcode() == X86ISD::PSHUFLW ? 0 : 2;
21938 DMask[DOffset + 0] = DOffset + 1;
21939 DMask[DOffset + 1] = DOffset + 0;
21940 MVT DVT = MVT::getVectorVT(MVT::i32, VT.getVectorNumElements() / 2);
21941 V = DAG.getBitcast(DVT, V);
21942 DCI.AddToWorklist(V.getNode());
21943 V = DAG.getNode(X86ISD::PSHUFD, DL, DVT, V,
21944 getV4X86ShuffleImm8ForMask(DMask, DL, DAG));
21945 DCI.AddToWorklist(V.getNode());
21946 return DAG.getBitcast(VT, V);
21949 // Look for shuffle patterns which can be implemented as a single unpack.
21950 // FIXME: This doesn't handle the location of the PSHUFD generically, and
21951 // only works when we have a PSHUFD followed by two half-shuffles.
21952 if (Mask[0] == Mask[1] && Mask[2] == Mask[3] &&
21953 (V.getOpcode() == X86ISD::PSHUFLW ||
21954 V.getOpcode() == X86ISD::PSHUFHW) &&
21955 V.getOpcode() != N.getOpcode() &&
21957 SDValue D = V.getOperand(0);
21958 while (D.getOpcode() == ISD::BITCAST && D.hasOneUse())
21959 D = D.getOperand(0);
21960 if (D.getOpcode() == X86ISD::PSHUFD && D.hasOneUse()) {
21961 SmallVector<int, 4> VMask = getPSHUFShuffleMask(V);
21962 SmallVector<int, 4> DMask = getPSHUFShuffleMask(D);
21963 int NOffset = N.getOpcode() == X86ISD::PSHUFLW ? 0 : 4;
21964 int VOffset = V.getOpcode() == X86ISD::PSHUFLW ? 0 : 4;
21966 for (int i = 0; i < 4; ++i) {
21967 WordMask[i + NOffset] = Mask[i] + NOffset;
21968 WordMask[i + VOffset] = VMask[i] + VOffset;
21970 // Map the word mask through the DWord mask.
21972 for (int i = 0; i < 8; ++i)
21973 MappedMask[i] = 2 * DMask[WordMask[i] / 2] + WordMask[i] % 2;
21974 if (makeArrayRef(MappedMask).equals({0, 0, 1, 1, 2, 2, 3, 3}) ||
21975 makeArrayRef(MappedMask).equals({4, 4, 5, 5, 6, 6, 7, 7})) {
21976 // We can replace all three shuffles with an unpack.
21977 V = DAG.getBitcast(VT, D.getOperand(0));
21978 DCI.AddToWorklist(V.getNode());
21979 return DAG.getNode(MappedMask[0] == 0 ? X86ISD::UNPCKL
21988 case X86ISD::PSHUFD:
21989 if (SDValue NewN = combineRedundantDWordShuffle(N, Mask, DAG, DCI))
21998 /// \brief Try to combine a shuffle into a target-specific add-sub node.
22000 /// We combine this directly on the abstract vector shuffle nodes so it is
22001 /// easier to generically match. We also insert dummy vector shuffle nodes for
22002 /// the operands which explicitly discard the lanes which are unused by this
22003 /// operation to try to flow through the rest of the combiner the fact that
22004 /// they're unused.
22005 static SDValue combineShuffleToAddSub(SDNode *N, SelectionDAG &DAG) {
22007 EVT VT = N->getValueType(0);
22009 // We only handle target-independent shuffles.
22010 // FIXME: It would be easy and harmless to use the target shuffle mask
22011 // extraction tool to support more.
22012 if (N->getOpcode() != ISD::VECTOR_SHUFFLE)
22015 auto *SVN = cast<ShuffleVectorSDNode>(N);
22016 ArrayRef<int> Mask = SVN->getMask();
22017 SDValue V1 = N->getOperand(0);
22018 SDValue V2 = N->getOperand(1);
22020 // We require the first shuffle operand to be the SUB node, and the second to
22021 // be the ADD node.
22022 // FIXME: We should support the commuted patterns.
22023 if (V1->getOpcode() != ISD::FSUB || V2->getOpcode() != ISD::FADD)
22026 // If there are other uses of these operations we can't fold them.
22027 if (!V1->hasOneUse() || !V2->hasOneUse())
22030 // Ensure that both operations have the same operands. Note that we can
22031 // commute the FADD operands.
22032 SDValue LHS = V1->getOperand(0), RHS = V1->getOperand(1);
22033 if ((V2->getOperand(0) != LHS || V2->getOperand(1) != RHS) &&
22034 (V2->getOperand(0) != RHS || V2->getOperand(1) != LHS))
22037 // We're looking for blends between FADD and FSUB nodes. We insist on these
22038 // nodes being lined up in a specific expected pattern.
22039 if (!(isShuffleEquivalent(V1, V2, Mask, {0, 3}) ||
22040 isShuffleEquivalent(V1, V2, Mask, {0, 5, 2, 7}) ||
22041 isShuffleEquivalent(V1, V2, Mask, {0, 9, 2, 11, 4, 13, 6, 15})))
22044 // Only specific types are legal at this point, assert so we notice if and
22045 // when these change.
22046 assert((VT == MVT::v4f32 || VT == MVT::v2f64 || VT == MVT::v8f32 ||
22047 VT == MVT::v4f64) &&
22048 "Unknown vector type encountered!");
22050 return DAG.getNode(X86ISD::ADDSUB, DL, VT, LHS, RHS);
22053 /// PerformShuffleCombine - Performs several different shuffle combines.
22054 static SDValue PerformShuffleCombine(SDNode *N, SelectionDAG &DAG,
22055 TargetLowering::DAGCombinerInfo &DCI,
22056 const X86Subtarget *Subtarget) {
22058 SDValue N0 = N->getOperand(0);
22059 SDValue N1 = N->getOperand(1);
22060 EVT VT = N->getValueType(0);
22062 // Don't create instructions with illegal types after legalize types has run.
22063 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
22064 if (!DCI.isBeforeLegalize() && !TLI.isTypeLegal(VT.getVectorElementType()))
22067 // If we have legalized the vector types, look for blends of FADD and FSUB
22068 // nodes that we can fuse into an ADDSUB node.
22069 if (TLI.isTypeLegal(VT) && Subtarget->hasSSE3())
22070 if (SDValue AddSub = combineShuffleToAddSub(N, DAG))
22073 // Combine 256-bit vector shuffles. This is only profitable when in AVX mode
22074 if (Subtarget->hasFp256() && VT.is256BitVector() &&
22075 N->getOpcode() == ISD::VECTOR_SHUFFLE)
22076 return PerformShuffleCombine256(N, DAG, DCI, Subtarget);
22078 // During Type Legalization, when promoting illegal vector types,
22079 // the backend might introduce new shuffle dag nodes and bitcasts.
22081 // This code performs the following transformation:
22082 // fold: (shuffle (bitcast (BINOP A, B)), Undef, <Mask>) ->
22083 // (shuffle (BINOP (bitcast A), (bitcast B)), Undef, <Mask>)
22085 // We do this only if both the bitcast and the BINOP dag nodes have
22086 // one use. Also, perform this transformation only if the new binary
22087 // operation is legal. This is to avoid introducing dag nodes that
22088 // potentially need to be further expanded (or custom lowered) into a
22089 // less optimal sequence of dag nodes.
22090 if (!DCI.isBeforeLegalize() && DCI.isBeforeLegalizeOps() &&
22091 N1.getOpcode() == ISD::UNDEF && N0.hasOneUse() &&
22092 N0.getOpcode() == ISD::BITCAST) {
22093 SDValue BC0 = N0.getOperand(0);
22094 EVT SVT = BC0.getValueType();
22095 unsigned Opcode = BC0.getOpcode();
22096 unsigned NumElts = VT.getVectorNumElements();
22098 if (BC0.hasOneUse() && SVT.isVector() &&
22099 SVT.getVectorNumElements() * 2 == NumElts &&
22100 TLI.isOperationLegal(Opcode, VT)) {
22101 bool CanFold = false;
22113 unsigned SVTNumElts = SVT.getVectorNumElements();
22114 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
22115 for (unsigned i = 0, e = SVTNumElts; i != e && CanFold; ++i)
22116 CanFold = SVOp->getMaskElt(i) == (int)(i * 2);
22117 for (unsigned i = SVTNumElts, e = NumElts; i != e && CanFold; ++i)
22118 CanFold = SVOp->getMaskElt(i) < 0;
22121 SDValue BC00 = DAG.getBitcast(VT, BC0.getOperand(0));
22122 SDValue BC01 = DAG.getBitcast(VT, BC0.getOperand(1));
22123 SDValue NewBinOp = DAG.getNode(BC0.getOpcode(), dl, VT, BC00, BC01);
22124 return DAG.getVectorShuffle(VT, dl, NewBinOp, N1, &SVOp->getMask()[0]);
22129 // Combine a vector_shuffle that is equal to build_vector load1, load2, load3,
22130 // load4, <0, 1, 2, 3> into a 128-bit load if the load addresses are
22131 // consecutive, non-overlapping, and in the right order.
22132 SmallVector<SDValue, 16> Elts;
22133 for (unsigned i = 0, e = VT.getVectorNumElements(); i != e; ++i)
22134 Elts.push_back(getShuffleScalarElt(N, i, DAG, 0));
22136 if (SDValue LD = EltsFromConsecutiveLoads(VT, Elts, dl, DAG, true))
22139 if (isTargetShuffle(N->getOpcode())) {
22141 PerformTargetShuffleCombine(SDValue(N, 0), DAG, DCI, Subtarget);
22142 if (Shuffle.getNode())
22145 // Try recursively combining arbitrary sequences of x86 shuffle
22146 // instructions into higher-order shuffles. We do this after combining
22147 // specific PSHUF instruction sequences into their minimal form so that we
22148 // can evaluate how many specialized shuffle instructions are involved in
22149 // a particular chain.
22150 SmallVector<int, 1> NonceMask; // Just a placeholder.
22151 NonceMask.push_back(0);
22152 if (combineX86ShufflesRecursively(SDValue(N, 0), SDValue(N, 0), NonceMask,
22153 /*Depth*/ 1, /*HasPSHUFB*/ false, DAG,
22155 return SDValue(); // This routine will use CombineTo to replace N.
22161 /// XFormVExtractWithShuffleIntoLoad - Check if a vector extract from a target
22162 /// specific shuffle of a load can be folded into a single element load.
22163 /// Similar handling for VECTOR_SHUFFLE is performed by DAGCombiner, but
22164 /// shuffles have been custom lowered so we need to handle those here.
22165 static SDValue XFormVExtractWithShuffleIntoLoad(SDNode *N, SelectionDAG &DAG,
22166 TargetLowering::DAGCombinerInfo &DCI) {
22167 if (DCI.isBeforeLegalizeOps())
22170 SDValue InVec = N->getOperand(0);
22171 SDValue EltNo = N->getOperand(1);
22173 if (!isa<ConstantSDNode>(EltNo))
22176 EVT OriginalVT = InVec.getValueType();
22178 if (InVec.getOpcode() == ISD::BITCAST) {
22179 // Don't duplicate a load with other uses.
22180 if (!InVec.hasOneUse())
22182 EVT BCVT = InVec.getOperand(0).getValueType();
22183 if (!BCVT.isVector() ||
22184 BCVT.getVectorNumElements() != OriginalVT.getVectorNumElements())
22186 InVec = InVec.getOperand(0);
22189 EVT CurrentVT = InVec.getValueType();
22191 if (!isTargetShuffle(InVec.getOpcode()))
22194 // Don't duplicate a load with other uses.
22195 if (!InVec.hasOneUse())
22198 SmallVector<int, 16> ShuffleMask;
22200 if (!getTargetShuffleMask(InVec.getNode(), CurrentVT.getSimpleVT(),
22201 ShuffleMask, UnaryShuffle))
22204 // Select the input vector, guarding against out of range extract vector.
22205 unsigned NumElems = CurrentVT.getVectorNumElements();
22206 int Elt = cast<ConstantSDNode>(EltNo)->getZExtValue();
22207 int Idx = (Elt > (int)NumElems) ? -1 : ShuffleMask[Elt];
22208 SDValue LdNode = (Idx < (int)NumElems) ? InVec.getOperand(0)
22209 : InVec.getOperand(1);
22211 // If inputs to shuffle are the same for both ops, then allow 2 uses
22212 unsigned AllowedUses = InVec.getNumOperands() > 1 &&
22213 InVec.getOperand(0) == InVec.getOperand(1) ? 2 : 1;
22215 if (LdNode.getOpcode() == ISD::BITCAST) {
22216 // Don't duplicate a load with other uses.
22217 if (!LdNode.getNode()->hasNUsesOfValue(AllowedUses, 0))
22220 AllowedUses = 1; // only allow 1 load use if we have a bitcast
22221 LdNode = LdNode.getOperand(0);
22224 if (!ISD::isNormalLoad(LdNode.getNode()))
22227 LoadSDNode *LN0 = cast<LoadSDNode>(LdNode);
22229 if (!LN0 ||!LN0->hasNUsesOfValue(AllowedUses, 0) || LN0->isVolatile())
22232 EVT EltVT = N->getValueType(0);
22233 // If there's a bitcast before the shuffle, check if the load type and
22234 // alignment is valid.
22235 unsigned Align = LN0->getAlignment();
22236 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
22237 unsigned NewAlign = DAG.getDataLayout().getABITypeAlignment(
22238 EltVT.getTypeForEVT(*DAG.getContext()));
22240 if (NewAlign > Align || !TLI.isOperationLegalOrCustom(ISD::LOAD, EltVT))
22243 // All checks match so transform back to vector_shuffle so that DAG combiner
22244 // can finish the job
22247 // Create shuffle node taking into account the case that its a unary shuffle
22248 SDValue Shuffle = (UnaryShuffle) ? DAG.getUNDEF(CurrentVT)
22249 : InVec.getOperand(1);
22250 Shuffle = DAG.getVectorShuffle(CurrentVT, dl,
22251 InVec.getOperand(0), Shuffle,
22253 Shuffle = DAG.getBitcast(OriginalVT, Shuffle);
22254 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, N->getValueType(0), Shuffle,
22258 /// \brief Detect bitcasts between i32 to x86mmx low word. Since MMX types are
22259 /// special and don't usually play with other vector types, it's better to
22260 /// handle them early to be sure we emit efficient code by avoiding
22261 /// store-load conversions.
22262 static SDValue PerformBITCASTCombine(SDNode *N, SelectionDAG &DAG) {
22263 if (N->getValueType(0) != MVT::x86mmx ||
22264 N->getOperand(0)->getOpcode() != ISD::BUILD_VECTOR ||
22265 N->getOperand(0)->getValueType(0) != MVT::v2i32)
22268 SDValue V = N->getOperand(0);
22269 ConstantSDNode *C = dyn_cast<ConstantSDNode>(V.getOperand(1));
22270 if (C && C->getZExtValue() == 0 && V.getOperand(0).getValueType() == MVT::i32)
22271 return DAG.getNode(X86ISD::MMX_MOVW2D, SDLoc(V.getOperand(0)),
22272 N->getValueType(0), V.getOperand(0));
22277 /// PerformEXTRACT_VECTOR_ELTCombine - Detect vector gather/scatter index
22278 /// generation and convert it from being a bunch of shuffles and extracts
22279 /// into a somewhat faster sequence. For i686, the best sequence is apparently
22280 /// storing the value and loading scalars back, while for x64 we should
22281 /// use 64-bit extracts and shifts.
22282 static SDValue PerformEXTRACT_VECTOR_ELTCombine(SDNode *N, SelectionDAG &DAG,
22283 TargetLowering::DAGCombinerInfo &DCI) {
22284 if (SDValue NewOp = XFormVExtractWithShuffleIntoLoad(N, DAG, DCI))
22287 SDValue InputVector = N->getOperand(0);
22288 SDLoc dl(InputVector);
22289 // Detect mmx to i32 conversion through a v2i32 elt extract.
22290 if (InputVector.getOpcode() == ISD::BITCAST && InputVector.hasOneUse() &&
22291 N->getValueType(0) == MVT::i32 &&
22292 InputVector.getValueType() == MVT::v2i32) {
22294 // The bitcast source is a direct mmx result.
22295 SDValue MMXSrc = InputVector.getNode()->getOperand(0);
22296 if (MMXSrc.getValueType() == MVT::x86mmx)
22297 return DAG.getNode(X86ISD::MMX_MOVD2W, SDLoc(InputVector),
22298 N->getValueType(0),
22299 InputVector.getNode()->getOperand(0));
22301 // The mmx is indirect: (i64 extract_elt (v1i64 bitcast (x86mmx ...))).
22302 SDValue MMXSrcOp = MMXSrc.getOperand(0);
22303 if (MMXSrc.getOpcode() == ISD::EXTRACT_VECTOR_ELT && MMXSrc.hasOneUse() &&
22304 MMXSrc.getValueType() == MVT::i64 && MMXSrcOp.hasOneUse() &&
22305 MMXSrcOp.getOpcode() == ISD::BITCAST &&
22306 MMXSrcOp.getValueType() == MVT::v1i64 &&
22307 MMXSrcOp.getOperand(0).getValueType() == MVT::x86mmx)
22308 return DAG.getNode(X86ISD::MMX_MOVD2W, SDLoc(InputVector),
22309 N->getValueType(0),
22310 MMXSrcOp.getOperand(0));
22313 EVT VT = N->getValueType(0);
22315 if (VT == MVT::i1 && dyn_cast<ConstantSDNode>(N->getOperand(1)) &&
22316 InputVector.getOpcode() == ISD::BITCAST &&
22317 dyn_cast<ConstantSDNode>(InputVector.getOperand(0))) {
22318 uint64_t ExtractedElt =
22319 cast<ConstantSDNode>(N->getOperand(1))->getZExtValue();
22320 uint64_t InputValue =
22321 cast<ConstantSDNode>(InputVector.getOperand(0))->getZExtValue();
22322 uint64_t Res = (InputValue >> ExtractedElt) & 1;
22323 return DAG.getConstant(Res, dl, MVT::i1);
22325 // Only operate on vectors of 4 elements, where the alternative shuffling
22326 // gets to be more expensive.
22327 if (InputVector.getValueType() != MVT::v4i32)
22330 // Check whether every use of InputVector is an EXTRACT_VECTOR_ELT with a
22331 // single use which is a sign-extend or zero-extend, and all elements are
22333 SmallVector<SDNode *, 4> Uses;
22334 unsigned ExtractedElements = 0;
22335 for (SDNode::use_iterator UI = InputVector.getNode()->use_begin(),
22336 UE = InputVector.getNode()->use_end(); UI != UE; ++UI) {
22337 if (UI.getUse().getResNo() != InputVector.getResNo())
22340 SDNode *Extract = *UI;
22341 if (Extract->getOpcode() != ISD::EXTRACT_VECTOR_ELT)
22344 if (Extract->getValueType(0) != MVT::i32)
22346 if (!Extract->hasOneUse())
22348 if (Extract->use_begin()->getOpcode() != ISD::SIGN_EXTEND &&
22349 Extract->use_begin()->getOpcode() != ISD::ZERO_EXTEND)
22351 if (!isa<ConstantSDNode>(Extract->getOperand(1)))
22354 // Record which element was extracted.
22355 ExtractedElements |=
22356 1 << cast<ConstantSDNode>(Extract->getOperand(1))->getZExtValue();
22358 Uses.push_back(Extract);
22361 // If not all the elements were used, this may not be worthwhile.
22362 if (ExtractedElements != 15)
22365 // Ok, we've now decided to do the transformation.
22366 // If 64-bit shifts are legal, use the extract-shift sequence,
22367 // otherwise bounce the vector off the cache.
22368 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
22371 if (TLI.isOperationLegal(ISD::SRA, MVT::i64)) {
22372 SDValue Cst = DAG.getBitcast(MVT::v2i64, InputVector);
22373 auto &DL = DAG.getDataLayout();
22374 EVT VecIdxTy = DAG.getTargetLoweringInfo().getVectorIdxTy(DL);
22375 SDValue BottomHalf = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i64, Cst,
22376 DAG.getConstant(0, dl, VecIdxTy));
22377 SDValue TopHalf = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i64, Cst,
22378 DAG.getConstant(1, dl, VecIdxTy));
22380 SDValue ShAmt = DAG.getConstant(
22381 32, dl, DAG.getTargetLoweringInfo().getShiftAmountTy(MVT::i64, DL));
22382 Vals[0] = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, BottomHalf);
22383 Vals[1] = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32,
22384 DAG.getNode(ISD::SRA, dl, MVT::i64, BottomHalf, ShAmt));
22385 Vals[2] = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, TopHalf);
22386 Vals[3] = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32,
22387 DAG.getNode(ISD::SRA, dl, MVT::i64, TopHalf, ShAmt));
22389 // Store the value to a temporary stack slot.
22390 SDValue StackPtr = DAG.CreateStackTemporary(InputVector.getValueType());
22391 SDValue Ch = DAG.getStore(DAG.getEntryNode(), dl, InputVector, StackPtr,
22392 MachinePointerInfo(), false, false, 0);
22394 EVT ElementType = InputVector.getValueType().getVectorElementType();
22395 unsigned EltSize = ElementType.getSizeInBits() / 8;
22397 // Replace each use (extract) with a load of the appropriate element.
22398 for (unsigned i = 0; i < 4; ++i) {
22399 uint64_t Offset = EltSize * i;
22400 auto PtrVT = TLI.getPointerTy(DAG.getDataLayout());
22401 SDValue OffsetVal = DAG.getConstant(Offset, dl, PtrVT);
22403 SDValue ScalarAddr =
22404 DAG.getNode(ISD::ADD, dl, PtrVT, StackPtr, OffsetVal);
22406 // Load the scalar.
22407 Vals[i] = DAG.getLoad(ElementType, dl, Ch,
22408 ScalarAddr, MachinePointerInfo(),
22409 false, false, false, 0);
22414 // Replace the extracts
22415 for (SmallVectorImpl<SDNode *>::iterator UI = Uses.begin(),
22416 UE = Uses.end(); UI != UE; ++UI) {
22417 SDNode *Extract = *UI;
22419 SDValue Idx = Extract->getOperand(1);
22420 uint64_t IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
22421 DAG.ReplaceAllUsesOfValueWith(SDValue(Extract, 0), Vals[IdxVal]);
22424 // The replacement was made in place; don't return anything.
22429 transformVSELECTtoBlendVECTOR_SHUFFLE(SDNode *N, SelectionDAG &DAG,
22430 const X86Subtarget *Subtarget) {
22432 SDValue Cond = N->getOperand(0);
22433 SDValue LHS = N->getOperand(1);
22434 SDValue RHS = N->getOperand(2);
22436 if (Cond.getOpcode() == ISD::SIGN_EXTEND) {
22437 SDValue CondSrc = Cond->getOperand(0);
22438 if (CondSrc->getOpcode() == ISD::SIGN_EXTEND_INREG)
22439 Cond = CondSrc->getOperand(0);
22442 if (!ISD::isBuildVectorOfConstantSDNodes(Cond.getNode()))
22445 // A vselect where all conditions and data are constants can be optimized into
22446 // a single vector load by SelectionDAGLegalize::ExpandBUILD_VECTOR().
22447 if (ISD::isBuildVectorOfConstantSDNodes(LHS.getNode()) &&
22448 ISD::isBuildVectorOfConstantSDNodes(RHS.getNode()))
22451 unsigned MaskValue = 0;
22452 if (!BUILD_VECTORtoBlendMask(cast<BuildVectorSDNode>(Cond), MaskValue))
22455 MVT VT = N->getSimpleValueType(0);
22456 unsigned NumElems = VT.getVectorNumElements();
22457 SmallVector<int, 8> ShuffleMask(NumElems, -1);
22458 for (unsigned i = 0; i < NumElems; ++i) {
22459 // Be sure we emit undef where we can.
22460 if (Cond.getOperand(i)->getOpcode() == ISD::UNDEF)
22461 ShuffleMask[i] = -1;
22463 ShuffleMask[i] = i + NumElems * ((MaskValue >> i) & 1);
22466 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
22467 if (!TLI.isShuffleMaskLegal(ShuffleMask, VT))
22469 return DAG.getVectorShuffle(VT, dl, LHS, RHS, &ShuffleMask[0]);
22472 /// PerformSELECTCombine - Do target-specific dag combines on SELECT and VSELECT
22474 static SDValue PerformSELECTCombine(SDNode *N, SelectionDAG &DAG,
22475 TargetLowering::DAGCombinerInfo &DCI,
22476 const X86Subtarget *Subtarget) {
22478 SDValue Cond = N->getOperand(0);
22479 // Get the LHS/RHS of the select.
22480 SDValue LHS = N->getOperand(1);
22481 SDValue RHS = N->getOperand(2);
22482 EVT VT = LHS.getValueType();
22483 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
22485 // If we have SSE[12] support, try to form min/max nodes. SSE min/max
22486 // instructions match the semantics of the common C idiom x<y?x:y but not
22487 // x<=y?x:y, because of how they handle negative zero (which can be
22488 // ignored in unsafe-math mode).
22489 // We also try to create v2f32 min/max nodes, which we later widen to v4f32.
22490 if (Cond.getOpcode() == ISD::SETCC && VT.isFloatingPoint() &&
22491 VT != MVT::f80 && (TLI.isTypeLegal(VT) || VT == MVT::v2f32) &&
22492 (Subtarget->hasSSE2() ||
22493 (Subtarget->hasSSE1() && VT.getScalarType() == MVT::f32))) {
22494 ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
22496 unsigned Opcode = 0;
22497 // Check for x CC y ? x : y.
22498 if (DAG.isEqualTo(LHS, Cond.getOperand(0)) &&
22499 DAG.isEqualTo(RHS, Cond.getOperand(1))) {
22503 // Converting this to a min would handle NaNs incorrectly, and swapping
22504 // the operands would cause it to handle comparisons between positive
22505 // and negative zero incorrectly.
22506 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)) {
22507 if (!DAG.getTarget().Options.UnsafeFPMath &&
22508 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
22510 std::swap(LHS, RHS);
22512 Opcode = X86ISD::FMIN;
22515 // Converting this to a min would handle comparisons between positive
22516 // and negative zero incorrectly.
22517 if (!DAG.getTarget().Options.UnsafeFPMath &&
22518 !DAG.isKnownNeverZero(LHS) && !DAG.isKnownNeverZero(RHS))
22520 Opcode = X86ISD::FMIN;
22523 // Converting this to a min would handle both negative zeros and NaNs
22524 // incorrectly, but we can swap the operands to fix both.
22525 std::swap(LHS, RHS);
22529 Opcode = X86ISD::FMIN;
22533 // Converting this to a max would handle comparisons between positive
22534 // and negative zero incorrectly.
22535 if (!DAG.getTarget().Options.UnsafeFPMath &&
22536 !DAG.isKnownNeverZero(LHS) && !DAG.isKnownNeverZero(RHS))
22538 Opcode = X86ISD::FMAX;
22541 // Converting this to a max would handle NaNs incorrectly, and swapping
22542 // the operands would cause it to handle comparisons between positive
22543 // and negative zero incorrectly.
22544 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)) {
22545 if (!DAG.getTarget().Options.UnsafeFPMath &&
22546 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
22548 std::swap(LHS, RHS);
22550 Opcode = X86ISD::FMAX;
22553 // Converting this to a max would handle both negative zeros and NaNs
22554 // incorrectly, but we can swap the operands to fix both.
22555 std::swap(LHS, RHS);
22559 Opcode = X86ISD::FMAX;
22562 // Check for x CC y ? y : x -- a min/max with reversed arms.
22563 } else if (DAG.isEqualTo(LHS, Cond.getOperand(1)) &&
22564 DAG.isEqualTo(RHS, Cond.getOperand(0))) {
22568 // Converting this to a min would handle comparisons between positive
22569 // and negative zero incorrectly, and swapping the operands would
22570 // cause it to handle NaNs incorrectly.
22571 if (!DAG.getTarget().Options.UnsafeFPMath &&
22572 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS))) {
22573 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS))
22575 std::swap(LHS, RHS);
22577 Opcode = X86ISD::FMIN;
22580 // Converting this to a min would handle NaNs incorrectly.
22581 if (!DAG.getTarget().Options.UnsafeFPMath &&
22582 (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)))
22584 Opcode = X86ISD::FMIN;
22587 // Converting this to a min would handle both negative zeros and NaNs
22588 // incorrectly, but we can swap the operands to fix both.
22589 std::swap(LHS, RHS);
22593 Opcode = X86ISD::FMIN;
22597 // Converting this to a max would handle NaNs incorrectly.
22598 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS))
22600 Opcode = X86ISD::FMAX;
22603 // Converting this to a max would handle comparisons between positive
22604 // and negative zero incorrectly, and swapping the operands would
22605 // cause it to handle NaNs incorrectly.
22606 if (!DAG.getTarget().Options.UnsafeFPMath &&
22607 !DAG.isKnownNeverZero(LHS) && !DAG.isKnownNeverZero(RHS)) {
22608 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS))
22610 std::swap(LHS, RHS);
22612 Opcode = X86ISD::FMAX;
22615 // Converting this to a max would handle both negative zeros and NaNs
22616 // incorrectly, but we can swap the operands to fix both.
22617 std::swap(LHS, RHS);
22621 Opcode = X86ISD::FMAX;
22627 return DAG.getNode(Opcode, DL, N->getValueType(0), LHS, RHS);
22630 EVT CondVT = Cond.getValueType();
22631 if (Subtarget->hasAVX512() && VT.isVector() && CondVT.isVector() &&
22632 CondVT.getVectorElementType() == MVT::i1) {
22633 // v16i8 (select v16i1, v16i8, v16i8) does not have a proper
22634 // lowering on KNL. In this case we convert it to
22635 // v16i8 (select v16i8, v16i8, v16i8) and use AVX instruction.
22636 // The same situation for all 128 and 256-bit vectors of i8 and i16.
22637 // Since SKX these selects have a proper lowering.
22638 EVT OpVT = LHS.getValueType();
22639 if ((OpVT.is128BitVector() || OpVT.is256BitVector()) &&
22640 (OpVT.getVectorElementType() == MVT::i8 ||
22641 OpVT.getVectorElementType() == MVT::i16) &&
22642 !(Subtarget->hasBWI() && Subtarget->hasVLX())) {
22643 Cond = DAG.getNode(ISD::SIGN_EXTEND, DL, OpVT, Cond);
22644 DCI.AddToWorklist(Cond.getNode());
22645 return DAG.getNode(N->getOpcode(), DL, OpVT, Cond, LHS, RHS);
22648 // If this is a select between two integer constants, try to do some
22650 if (ConstantSDNode *TrueC = dyn_cast<ConstantSDNode>(LHS)) {
22651 if (ConstantSDNode *FalseC = dyn_cast<ConstantSDNode>(RHS))
22652 // Don't do this for crazy integer types.
22653 if (DAG.getTargetLoweringInfo().isTypeLegal(LHS.getValueType())) {
22654 // If this is efficiently invertible, canonicalize the LHSC/RHSC values
22655 // so that TrueC (the true value) is larger than FalseC.
22656 bool NeedsCondInvert = false;
22658 if (TrueC->getAPIntValue().ult(FalseC->getAPIntValue()) &&
22659 // Efficiently invertible.
22660 (Cond.getOpcode() == ISD::SETCC || // setcc -> invertible.
22661 (Cond.getOpcode() == ISD::XOR && // xor(X, C) -> invertible.
22662 isa<ConstantSDNode>(Cond.getOperand(1))))) {
22663 NeedsCondInvert = true;
22664 std::swap(TrueC, FalseC);
22667 // Optimize C ? 8 : 0 -> zext(C) << 3. Likewise for any pow2/0.
22668 if (FalseC->getAPIntValue() == 0 &&
22669 TrueC->getAPIntValue().isPowerOf2()) {
22670 if (NeedsCondInvert) // Invert the condition if needed.
22671 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
22672 DAG.getConstant(1, DL, Cond.getValueType()));
22674 // Zero extend the condition if needed.
22675 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, LHS.getValueType(), Cond);
22677 unsigned ShAmt = TrueC->getAPIntValue().logBase2();
22678 return DAG.getNode(ISD::SHL, DL, LHS.getValueType(), Cond,
22679 DAG.getConstant(ShAmt, DL, MVT::i8));
22682 // Optimize Cond ? cst+1 : cst -> zext(setcc(C)+cst.
22683 if (FalseC->getAPIntValue()+1 == TrueC->getAPIntValue()) {
22684 if (NeedsCondInvert) // Invert the condition if needed.
22685 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
22686 DAG.getConstant(1, DL, Cond.getValueType()));
22688 // Zero extend the condition if needed.
22689 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL,
22690 FalseC->getValueType(0), Cond);
22691 return DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
22692 SDValue(FalseC, 0));
22695 // Optimize cases that will turn into an LEA instruction. This requires
22696 // an i32 or i64 and an efficient multiplier (1, 2, 3, 4, 5, 8, 9).
22697 if (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i64) {
22698 uint64_t Diff = TrueC->getZExtValue()-FalseC->getZExtValue();
22699 if (N->getValueType(0) == MVT::i32) Diff = (unsigned)Diff;
22701 bool isFastMultiplier = false;
22703 switch ((unsigned char)Diff) {
22705 case 1: // result = add base, cond
22706 case 2: // result = lea base( , cond*2)
22707 case 3: // result = lea base(cond, cond*2)
22708 case 4: // result = lea base( , cond*4)
22709 case 5: // result = lea base(cond, cond*4)
22710 case 8: // result = lea base( , cond*8)
22711 case 9: // result = lea base(cond, cond*8)
22712 isFastMultiplier = true;
22717 if (isFastMultiplier) {
22718 APInt Diff = TrueC->getAPIntValue()-FalseC->getAPIntValue();
22719 if (NeedsCondInvert) // Invert the condition if needed.
22720 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
22721 DAG.getConstant(1, DL, Cond.getValueType()));
22723 // Zero extend the condition if needed.
22724 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, FalseC->getValueType(0),
22726 // Scale the condition by the difference.
22728 Cond = DAG.getNode(ISD::MUL, DL, Cond.getValueType(), Cond,
22729 DAG.getConstant(Diff, DL,
22730 Cond.getValueType()));
22732 // Add the base if non-zero.
22733 if (FalseC->getAPIntValue() != 0)
22734 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
22735 SDValue(FalseC, 0));
22742 // Canonicalize max and min:
22743 // (x > y) ? x : y -> (x >= y) ? x : y
22744 // (x < y) ? x : y -> (x <= y) ? x : y
22745 // This allows use of COND_S / COND_NS (see TranslateX86CC) which eliminates
22746 // the need for an extra compare
22747 // against zero. e.g.
22748 // (x - y) > 0 : (x - y) ? 0 -> (x - y) >= 0 : (x - y) ? 0
22750 // testl %edi, %edi
22752 // cmovgl %edi, %eax
22756 // cmovsl %eax, %edi
22757 if (N->getOpcode() == ISD::SELECT && Cond.getOpcode() == ISD::SETCC &&
22758 DAG.isEqualTo(LHS, Cond.getOperand(0)) &&
22759 DAG.isEqualTo(RHS, Cond.getOperand(1))) {
22760 ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
22765 ISD::CondCode NewCC = (CC == ISD::SETLT) ? ISD::SETLE : ISD::SETGE;
22766 Cond = DAG.getSetCC(SDLoc(Cond), Cond.getValueType(),
22767 Cond.getOperand(0), Cond.getOperand(1), NewCC);
22768 return DAG.getNode(ISD::SELECT, DL, VT, Cond, LHS, RHS);
22773 // Early exit check
22774 if (!TLI.isTypeLegal(VT))
22777 // Match VSELECTs into subs with unsigned saturation.
22778 if (N->getOpcode() == ISD::VSELECT && Cond.getOpcode() == ISD::SETCC &&
22779 // psubus is available in SSE2 and AVX2 for i8 and i16 vectors.
22780 ((Subtarget->hasSSE2() && (VT == MVT::v16i8 || VT == MVT::v8i16)) ||
22781 (Subtarget->hasAVX2() && (VT == MVT::v32i8 || VT == MVT::v16i16)))) {
22782 ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
22784 // Check if one of the arms of the VSELECT is a zero vector. If it's on the
22785 // left side invert the predicate to simplify logic below.
22787 if (ISD::isBuildVectorAllZeros(LHS.getNode())) {
22789 CC = ISD::getSetCCInverse(CC, true);
22790 } else if (ISD::isBuildVectorAllZeros(RHS.getNode())) {
22794 if (Other.getNode() && Other->getNumOperands() == 2 &&
22795 DAG.isEqualTo(Other->getOperand(0), Cond.getOperand(0))) {
22796 SDValue OpLHS = Other->getOperand(0), OpRHS = Other->getOperand(1);
22797 SDValue CondRHS = Cond->getOperand(1);
22799 // Look for a general sub with unsigned saturation first.
22800 // x >= y ? x-y : 0 --> subus x, y
22801 // x > y ? x-y : 0 --> subus x, y
22802 if ((CC == ISD::SETUGE || CC == ISD::SETUGT) &&
22803 Other->getOpcode() == ISD::SUB && DAG.isEqualTo(OpRHS, CondRHS))
22804 return DAG.getNode(X86ISD::SUBUS, DL, VT, OpLHS, OpRHS);
22806 if (auto *OpRHSBV = dyn_cast<BuildVectorSDNode>(OpRHS))
22807 if (auto *OpRHSConst = OpRHSBV->getConstantSplatNode()) {
22808 if (auto *CondRHSBV = dyn_cast<BuildVectorSDNode>(CondRHS))
22809 if (auto *CondRHSConst = CondRHSBV->getConstantSplatNode())
22810 // If the RHS is a constant we have to reverse the const
22811 // canonicalization.
22812 // x > C-1 ? x+-C : 0 --> subus x, C
22813 if (CC == ISD::SETUGT && Other->getOpcode() == ISD::ADD &&
22814 CondRHSConst->getAPIntValue() ==
22815 (-OpRHSConst->getAPIntValue() - 1))
22816 return DAG.getNode(
22817 X86ISD::SUBUS, DL, VT, OpLHS,
22818 DAG.getConstant(-OpRHSConst->getAPIntValue(), DL, VT));
22820 // Another special case: If C was a sign bit, the sub has been
22821 // canonicalized into a xor.
22822 // FIXME: Would it be better to use computeKnownBits to determine
22823 // whether it's safe to decanonicalize the xor?
22824 // x s< 0 ? x^C : 0 --> subus x, C
22825 if (CC == ISD::SETLT && Other->getOpcode() == ISD::XOR &&
22826 ISD::isBuildVectorAllZeros(CondRHS.getNode()) &&
22827 OpRHSConst->getAPIntValue().isSignBit())
22828 // Note that we have to rebuild the RHS constant here to ensure we
22829 // don't rely on particular values of undef lanes.
22830 return DAG.getNode(
22831 X86ISD::SUBUS, DL, VT, OpLHS,
22832 DAG.getConstant(OpRHSConst->getAPIntValue(), DL, VT));
22837 // Simplify vector selection if condition value type matches vselect
22839 if (N->getOpcode() == ISD::VSELECT && CondVT == VT) {
22840 assert(Cond.getValueType().isVector() &&
22841 "vector select expects a vector selector!");
22843 bool TValIsAllOnes = ISD::isBuildVectorAllOnes(LHS.getNode());
22844 bool FValIsAllZeros = ISD::isBuildVectorAllZeros(RHS.getNode());
22846 // Try invert the condition if true value is not all 1s and false value
22848 if (!TValIsAllOnes && !FValIsAllZeros &&
22849 // Check if the selector will be produced by CMPP*/PCMP*
22850 Cond.getOpcode() == ISD::SETCC &&
22851 // Check if SETCC has already been promoted
22852 TLI.getSetCCResultType(DAG.getDataLayout(), *DAG.getContext(), VT) ==
22854 bool TValIsAllZeros = ISD::isBuildVectorAllZeros(LHS.getNode());
22855 bool FValIsAllOnes = ISD::isBuildVectorAllOnes(RHS.getNode());
22857 if (TValIsAllZeros || FValIsAllOnes) {
22858 SDValue CC = Cond.getOperand(2);
22859 ISD::CondCode NewCC =
22860 ISD::getSetCCInverse(cast<CondCodeSDNode>(CC)->get(),
22861 Cond.getOperand(0).getValueType().isInteger());
22862 Cond = DAG.getSetCC(DL, CondVT, Cond.getOperand(0), Cond.getOperand(1), NewCC);
22863 std::swap(LHS, RHS);
22864 TValIsAllOnes = FValIsAllOnes;
22865 FValIsAllZeros = TValIsAllZeros;
22869 if (TValIsAllOnes || FValIsAllZeros) {
22872 if (TValIsAllOnes && FValIsAllZeros)
22874 else if (TValIsAllOnes)
22876 DAG.getNode(ISD::OR, DL, CondVT, Cond, DAG.getBitcast(CondVT, RHS));
22877 else if (FValIsAllZeros)
22878 Ret = DAG.getNode(ISD::AND, DL, CondVT, Cond,
22879 DAG.getBitcast(CondVT, LHS));
22881 return DAG.getBitcast(VT, Ret);
22885 // We should generate an X86ISD::BLENDI from a vselect if its argument
22886 // is a sign_extend_inreg of an any_extend of a BUILD_VECTOR of
22887 // constants. This specific pattern gets generated when we split a
22888 // selector for a 512 bit vector in a machine without AVX512 (but with
22889 // 256-bit vectors), during legalization:
22891 // (vselect (sign_extend (any_extend (BUILD_VECTOR)) i1) LHS RHS)
22893 // Iff we find this pattern and the build_vectors are built from
22894 // constants, we translate the vselect into a shuffle_vector that we
22895 // know will be matched by LowerVECTOR_SHUFFLEtoBlend.
22896 if ((N->getOpcode() == ISD::VSELECT ||
22897 N->getOpcode() == X86ISD::SHRUNKBLEND) &&
22898 !DCI.isBeforeLegalize() && !VT.is512BitVector()) {
22899 SDValue Shuffle = transformVSELECTtoBlendVECTOR_SHUFFLE(N, DAG, Subtarget);
22900 if (Shuffle.getNode())
22904 // If this is a *dynamic* select (non-constant condition) and we can match
22905 // this node with one of the variable blend instructions, restructure the
22906 // condition so that the blends can use the high bit of each element and use
22907 // SimplifyDemandedBits to simplify the condition operand.
22908 if (N->getOpcode() == ISD::VSELECT && DCI.isBeforeLegalizeOps() &&
22909 !DCI.isBeforeLegalize() &&
22910 !ISD::isBuildVectorOfConstantSDNodes(Cond.getNode())) {
22911 unsigned BitWidth = Cond.getValueType().getScalarType().getSizeInBits();
22913 // Don't optimize vector selects that map to mask-registers.
22917 // We can only handle the cases where VSELECT is directly legal on the
22918 // subtarget. We custom lower VSELECT nodes with constant conditions and
22919 // this makes it hard to see whether a dynamic VSELECT will correctly
22920 // lower, so we both check the operation's status and explicitly handle the
22921 // cases where a *dynamic* blend will fail even though a constant-condition
22922 // blend could be custom lowered.
22923 // FIXME: We should find a better way to handle this class of problems.
22924 // Potentially, we should combine constant-condition vselect nodes
22925 // pre-legalization into shuffles and not mark as many types as custom
22927 if (!TLI.isOperationLegalOrCustom(ISD::VSELECT, VT))
22929 // FIXME: We don't support i16-element blends currently. We could and
22930 // should support them by making *all* the bits in the condition be set
22931 // rather than just the high bit and using an i8-element blend.
22932 if (VT.getScalarType() == MVT::i16)
22934 // Dynamic blending was only available from SSE4.1 onward.
22935 if (VT.getSizeInBits() == 128 && !Subtarget->hasSSE41())
22937 // Byte blends are only available in AVX2
22938 if (VT.getSizeInBits() == 256 && VT.getScalarType() == MVT::i8 &&
22939 !Subtarget->hasAVX2())
22942 assert(BitWidth >= 8 && BitWidth <= 64 && "Invalid mask size");
22943 APInt DemandedMask = APInt::getHighBitsSet(BitWidth, 1);
22945 APInt KnownZero, KnownOne;
22946 TargetLowering::TargetLoweringOpt TLO(DAG, DCI.isBeforeLegalize(),
22947 DCI.isBeforeLegalizeOps());
22948 if (TLO.ShrinkDemandedConstant(Cond, DemandedMask) ||
22949 TLI.SimplifyDemandedBits(Cond, DemandedMask, KnownZero, KnownOne,
22951 // If we changed the computation somewhere in the DAG, this change
22952 // will affect all users of Cond.
22953 // Make sure it is fine and update all the nodes so that we do not
22954 // use the generic VSELECT anymore. Otherwise, we may perform
22955 // wrong optimizations as we messed up with the actual expectation
22956 // for the vector boolean values.
22957 if (Cond != TLO.Old) {
22958 // Check all uses of that condition operand to check whether it will be
22959 // consumed by non-BLEND instructions, which may depend on all bits are
22961 for (SDNode::use_iterator I = Cond->use_begin(), E = Cond->use_end();
22963 if (I->getOpcode() != ISD::VSELECT)
22964 // TODO: Add other opcodes eventually lowered into BLEND.
22967 // Update all the users of the condition, before committing the change,
22968 // so that the VSELECT optimizations that expect the correct vector
22969 // boolean value will not be triggered.
22970 for (SDNode::use_iterator I = Cond->use_begin(), E = Cond->use_end();
22972 DAG.ReplaceAllUsesOfValueWith(
22974 DAG.getNode(X86ISD::SHRUNKBLEND, SDLoc(*I), I->getValueType(0),
22975 Cond, I->getOperand(1), I->getOperand(2)));
22976 DCI.CommitTargetLoweringOpt(TLO);
22979 // At this point, only Cond is changed. Change the condition
22980 // just for N to keep the opportunity to optimize all other
22981 // users their own way.
22982 DAG.ReplaceAllUsesOfValueWith(
22984 DAG.getNode(X86ISD::SHRUNKBLEND, SDLoc(N), N->getValueType(0),
22985 TLO.New, N->getOperand(1), N->getOperand(2)));
22993 // Check whether a boolean test is testing a boolean value generated by
22994 // X86ISD::SETCC. If so, return the operand of that SETCC and proper condition
22997 // Simplify the following patterns:
22998 // (Op (CMP (SETCC Cond EFLAGS) 1) EQ) or
22999 // (Op (CMP (SETCC Cond EFLAGS) 0) NEQ)
23000 // to (Op EFLAGS Cond)
23002 // (Op (CMP (SETCC Cond EFLAGS) 0) EQ) or
23003 // (Op (CMP (SETCC Cond EFLAGS) 1) NEQ)
23004 // to (Op EFLAGS !Cond)
23006 // where Op could be BRCOND or CMOV.
23008 static SDValue checkBoolTestSetCCCombine(SDValue Cmp, X86::CondCode &CC) {
23009 // Quit if not CMP and SUB with its value result used.
23010 if (Cmp.getOpcode() != X86ISD::CMP &&
23011 (Cmp.getOpcode() != X86ISD::SUB || Cmp.getNode()->hasAnyUseOfValue(0)))
23014 // Quit if not used as a boolean value.
23015 if (CC != X86::COND_E && CC != X86::COND_NE)
23018 // Check CMP operands. One of them should be 0 or 1 and the other should be
23019 // an SetCC or extended from it.
23020 SDValue Op1 = Cmp.getOperand(0);
23021 SDValue Op2 = Cmp.getOperand(1);
23024 const ConstantSDNode* C = nullptr;
23025 bool needOppositeCond = (CC == X86::COND_E);
23026 bool checkAgainstTrue = false; // Is it a comparison against 1?
23028 if ((C = dyn_cast<ConstantSDNode>(Op1)))
23030 else if ((C = dyn_cast<ConstantSDNode>(Op2)))
23032 else // Quit if all operands are not constants.
23035 if (C->getZExtValue() == 1) {
23036 needOppositeCond = !needOppositeCond;
23037 checkAgainstTrue = true;
23038 } else if (C->getZExtValue() != 0)
23039 // Quit if the constant is neither 0 or 1.
23042 bool truncatedToBoolWithAnd = false;
23043 // Skip (zext $x), (trunc $x), or (and $x, 1) node.
23044 while (SetCC.getOpcode() == ISD::ZERO_EXTEND ||
23045 SetCC.getOpcode() == ISD::TRUNCATE ||
23046 SetCC.getOpcode() == ISD::AND) {
23047 if (SetCC.getOpcode() == ISD::AND) {
23049 ConstantSDNode *CS;
23050 if ((CS = dyn_cast<ConstantSDNode>(SetCC.getOperand(0))) &&
23051 CS->getZExtValue() == 1)
23053 if ((CS = dyn_cast<ConstantSDNode>(SetCC.getOperand(1))) &&
23054 CS->getZExtValue() == 1)
23058 SetCC = SetCC.getOperand(OpIdx);
23059 truncatedToBoolWithAnd = true;
23061 SetCC = SetCC.getOperand(0);
23064 switch (SetCC.getOpcode()) {
23065 case X86ISD::SETCC_CARRY:
23066 // Since SETCC_CARRY gives output based on R = CF ? ~0 : 0, it's unsafe to
23067 // simplify it if the result of SETCC_CARRY is not canonicalized to 0 or 1,
23068 // i.e. it's a comparison against true but the result of SETCC_CARRY is not
23069 // truncated to i1 using 'and'.
23070 if (checkAgainstTrue && !truncatedToBoolWithAnd)
23072 assert(X86::CondCode(SetCC.getConstantOperandVal(0)) == X86::COND_B &&
23073 "Invalid use of SETCC_CARRY!");
23075 case X86ISD::SETCC:
23076 // Set the condition code or opposite one if necessary.
23077 CC = X86::CondCode(SetCC.getConstantOperandVal(0));
23078 if (needOppositeCond)
23079 CC = X86::GetOppositeBranchCondition(CC);
23080 return SetCC.getOperand(1);
23081 case X86ISD::CMOV: {
23082 // Check whether false/true value has canonical one, i.e. 0 or 1.
23083 ConstantSDNode *FVal = dyn_cast<ConstantSDNode>(SetCC.getOperand(0));
23084 ConstantSDNode *TVal = dyn_cast<ConstantSDNode>(SetCC.getOperand(1));
23085 // Quit if true value is not a constant.
23088 // Quit if false value is not a constant.
23090 SDValue Op = SetCC.getOperand(0);
23091 // Skip 'zext' or 'trunc' node.
23092 if (Op.getOpcode() == ISD::ZERO_EXTEND ||
23093 Op.getOpcode() == ISD::TRUNCATE)
23094 Op = Op.getOperand(0);
23095 // A special case for rdrand/rdseed, where 0 is set if false cond is
23097 if ((Op.getOpcode() != X86ISD::RDRAND &&
23098 Op.getOpcode() != X86ISD::RDSEED) || Op.getResNo() != 0)
23101 // Quit if false value is not the constant 0 or 1.
23102 bool FValIsFalse = true;
23103 if (FVal && FVal->getZExtValue() != 0) {
23104 if (FVal->getZExtValue() != 1)
23106 // If FVal is 1, opposite cond is needed.
23107 needOppositeCond = !needOppositeCond;
23108 FValIsFalse = false;
23110 // Quit if TVal is not the constant opposite of FVal.
23111 if (FValIsFalse && TVal->getZExtValue() != 1)
23113 if (!FValIsFalse && TVal->getZExtValue() != 0)
23115 CC = X86::CondCode(SetCC.getConstantOperandVal(2));
23116 if (needOppositeCond)
23117 CC = X86::GetOppositeBranchCondition(CC);
23118 return SetCC.getOperand(3);
23125 /// Check whether Cond is an AND/OR of SETCCs off of the same EFLAGS.
23127 /// (X86or (X86setcc) (X86setcc))
23128 /// (X86cmp (and (X86setcc) (X86setcc)), 0)
23129 static bool checkBoolTestAndOrSetCCCombine(SDValue Cond, X86::CondCode &CC0,
23130 X86::CondCode &CC1, SDValue &Flags,
23132 if (Cond->getOpcode() == X86ISD::CMP) {
23133 ConstantSDNode *CondOp1C = dyn_cast<ConstantSDNode>(Cond->getOperand(1));
23134 if (!CondOp1C || !CondOp1C->isNullValue())
23137 Cond = Cond->getOperand(0);
23142 SDValue SetCC0, SetCC1;
23143 switch (Cond->getOpcode()) {
23144 default: return false;
23151 SetCC0 = Cond->getOperand(0);
23152 SetCC1 = Cond->getOperand(1);
23156 // Make sure we have SETCC nodes, using the same flags value.
23157 if (SetCC0.getOpcode() != X86ISD::SETCC ||
23158 SetCC1.getOpcode() != X86ISD::SETCC ||
23159 SetCC0->getOperand(1) != SetCC1->getOperand(1))
23162 CC0 = (X86::CondCode)SetCC0->getConstantOperandVal(0);
23163 CC1 = (X86::CondCode)SetCC1->getConstantOperandVal(0);
23164 Flags = SetCC0->getOperand(1);
23168 /// Optimize X86ISD::CMOV [LHS, RHS, CONDCODE (e.g. X86::COND_NE), CONDVAL]
23169 static SDValue PerformCMOVCombine(SDNode *N, SelectionDAG &DAG,
23170 TargetLowering::DAGCombinerInfo &DCI,
23171 const X86Subtarget *Subtarget) {
23174 // If the flag operand isn't dead, don't touch this CMOV.
23175 if (N->getNumValues() == 2 && !SDValue(N, 1).use_empty())
23178 SDValue FalseOp = N->getOperand(0);
23179 SDValue TrueOp = N->getOperand(1);
23180 X86::CondCode CC = (X86::CondCode)N->getConstantOperandVal(2);
23181 SDValue Cond = N->getOperand(3);
23183 if (CC == X86::COND_E || CC == X86::COND_NE) {
23184 switch (Cond.getOpcode()) {
23188 // If operand of BSR / BSF are proven never zero, then ZF cannot be set.
23189 if (DAG.isKnownNeverZero(Cond.getOperand(0)))
23190 return (CC == X86::COND_E) ? FalseOp : TrueOp;
23196 Flags = checkBoolTestSetCCCombine(Cond, CC);
23197 if (Flags.getNode() &&
23198 // Extra check as FCMOV only supports a subset of X86 cond.
23199 (FalseOp.getValueType() != MVT::f80 || hasFPCMov(CC))) {
23200 SDValue Ops[] = { FalseOp, TrueOp,
23201 DAG.getConstant(CC, DL, MVT::i8), Flags };
23202 return DAG.getNode(X86ISD::CMOV, DL, N->getVTList(), Ops);
23205 // If this is a select between two integer constants, try to do some
23206 // optimizations. Note that the operands are ordered the opposite of SELECT
23208 if (ConstantSDNode *TrueC = dyn_cast<ConstantSDNode>(TrueOp)) {
23209 if (ConstantSDNode *FalseC = dyn_cast<ConstantSDNode>(FalseOp)) {
23210 // Canonicalize the TrueC/FalseC values so that TrueC (the true value) is
23211 // larger than FalseC (the false value).
23212 if (TrueC->getAPIntValue().ult(FalseC->getAPIntValue())) {
23213 CC = X86::GetOppositeBranchCondition(CC);
23214 std::swap(TrueC, FalseC);
23215 std::swap(TrueOp, FalseOp);
23218 // Optimize C ? 8 : 0 -> zext(setcc(C)) << 3. Likewise for any pow2/0.
23219 // This is efficient for any integer data type (including i8/i16) and
23221 if (FalseC->getAPIntValue() == 0 && TrueC->getAPIntValue().isPowerOf2()) {
23222 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
23223 DAG.getConstant(CC, DL, MVT::i8), Cond);
23225 // Zero extend the condition if needed.
23226 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, TrueC->getValueType(0), Cond);
23228 unsigned ShAmt = TrueC->getAPIntValue().logBase2();
23229 Cond = DAG.getNode(ISD::SHL, DL, Cond.getValueType(), Cond,
23230 DAG.getConstant(ShAmt, DL, MVT::i8));
23231 if (N->getNumValues() == 2) // Dead flag value?
23232 return DCI.CombineTo(N, Cond, SDValue());
23236 // Optimize Cond ? cst+1 : cst -> zext(setcc(C)+cst. This is efficient
23237 // for any integer data type, including i8/i16.
23238 if (FalseC->getAPIntValue()+1 == TrueC->getAPIntValue()) {
23239 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
23240 DAG.getConstant(CC, DL, MVT::i8), Cond);
23242 // Zero extend the condition if needed.
23243 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL,
23244 FalseC->getValueType(0), Cond);
23245 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
23246 SDValue(FalseC, 0));
23248 if (N->getNumValues() == 2) // Dead flag value?
23249 return DCI.CombineTo(N, Cond, SDValue());
23253 // Optimize cases that will turn into an LEA instruction. This requires
23254 // an i32 or i64 and an efficient multiplier (1, 2, 3, 4, 5, 8, 9).
23255 if (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i64) {
23256 uint64_t Diff = TrueC->getZExtValue()-FalseC->getZExtValue();
23257 if (N->getValueType(0) == MVT::i32) Diff = (unsigned)Diff;
23259 bool isFastMultiplier = false;
23261 switch ((unsigned char)Diff) {
23263 case 1: // result = add base, cond
23264 case 2: // result = lea base( , cond*2)
23265 case 3: // result = lea base(cond, cond*2)
23266 case 4: // result = lea base( , cond*4)
23267 case 5: // result = lea base(cond, cond*4)
23268 case 8: // result = lea base( , cond*8)
23269 case 9: // result = lea base(cond, cond*8)
23270 isFastMultiplier = true;
23275 if (isFastMultiplier) {
23276 APInt Diff = TrueC->getAPIntValue()-FalseC->getAPIntValue();
23277 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
23278 DAG.getConstant(CC, DL, MVT::i8), Cond);
23279 // Zero extend the condition if needed.
23280 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, FalseC->getValueType(0),
23282 // Scale the condition by the difference.
23284 Cond = DAG.getNode(ISD::MUL, DL, Cond.getValueType(), Cond,
23285 DAG.getConstant(Diff, DL, Cond.getValueType()));
23287 // Add the base if non-zero.
23288 if (FalseC->getAPIntValue() != 0)
23289 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
23290 SDValue(FalseC, 0));
23291 if (N->getNumValues() == 2) // Dead flag value?
23292 return DCI.CombineTo(N, Cond, SDValue());
23299 // Handle these cases:
23300 // (select (x != c), e, c) -> select (x != c), e, x),
23301 // (select (x == c), c, e) -> select (x == c), x, e)
23302 // where the c is an integer constant, and the "select" is the combination
23303 // of CMOV and CMP.
23305 // The rationale for this change is that the conditional-move from a constant
23306 // needs two instructions, however, conditional-move from a register needs
23307 // only one instruction.
23309 // CAVEAT: By replacing a constant with a symbolic value, it may obscure
23310 // some instruction-combining opportunities. This opt needs to be
23311 // postponed as late as possible.
23313 if (!DCI.isBeforeLegalize() && !DCI.isBeforeLegalizeOps()) {
23314 // the DCI.xxxx conditions are provided to postpone the optimization as
23315 // late as possible.
23317 ConstantSDNode *CmpAgainst = nullptr;
23318 if ((Cond.getOpcode() == X86ISD::CMP || Cond.getOpcode() == X86ISD::SUB) &&
23319 (CmpAgainst = dyn_cast<ConstantSDNode>(Cond.getOperand(1))) &&
23320 !isa<ConstantSDNode>(Cond.getOperand(0))) {
23322 if (CC == X86::COND_NE &&
23323 CmpAgainst == dyn_cast<ConstantSDNode>(FalseOp)) {
23324 CC = X86::GetOppositeBranchCondition(CC);
23325 std::swap(TrueOp, FalseOp);
23328 if (CC == X86::COND_E &&
23329 CmpAgainst == dyn_cast<ConstantSDNode>(TrueOp)) {
23330 SDValue Ops[] = { FalseOp, Cond.getOperand(0),
23331 DAG.getConstant(CC, DL, MVT::i8), Cond };
23332 return DAG.getNode(X86ISD::CMOV, DL, N->getVTList (), Ops);
23337 // Fold and/or of setcc's to double CMOV:
23338 // (CMOV F, T, ((cc1 | cc2) != 0)) -> (CMOV (CMOV F, T, cc1), T, cc2)
23339 // (CMOV F, T, ((cc1 & cc2) != 0)) -> (CMOV (CMOV T, F, !cc1), F, !cc2)
23341 // This combine lets us generate:
23342 // cmovcc1 (jcc1 if we don't have CMOV)
23348 // cmovne (jne if we don't have CMOV)
23349 // When we can't use the CMOV instruction, it might increase branch
23351 // When we can use CMOV, or when there is no mispredict, this improves
23352 // throughput and reduces register pressure.
23354 if (CC == X86::COND_NE) {
23356 X86::CondCode CC0, CC1;
23358 if (checkBoolTestAndOrSetCCCombine(Cond, CC0, CC1, Flags, isAndSetCC)) {
23360 std::swap(FalseOp, TrueOp);
23361 CC0 = X86::GetOppositeBranchCondition(CC0);
23362 CC1 = X86::GetOppositeBranchCondition(CC1);
23365 SDValue LOps[] = {FalseOp, TrueOp, DAG.getConstant(CC0, DL, MVT::i8),
23367 SDValue LCMOV = DAG.getNode(X86ISD::CMOV, DL, N->getVTList(), LOps);
23368 SDValue Ops[] = {LCMOV, TrueOp, DAG.getConstant(CC1, DL, MVT::i8), Flags};
23369 SDValue CMOV = DAG.getNode(X86ISD::CMOV, DL, N->getVTList(), Ops);
23370 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), SDValue(CMOV.getNode(), 1));
23378 /// PerformMulCombine - Optimize a single multiply with constant into two
23379 /// in order to implement it with two cheaper instructions, e.g.
23380 /// LEA + SHL, LEA + LEA.
23381 static SDValue PerformMulCombine(SDNode *N, SelectionDAG &DAG,
23382 TargetLowering::DAGCombinerInfo &DCI) {
23383 // An imul is usually smaller than the alternative sequence.
23384 if (DAG.getMachineFunction().getFunction()->optForMinSize())
23387 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
23390 EVT VT = N->getValueType(0);
23391 if (VT != MVT::i64 && VT != MVT::i32)
23394 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1));
23397 uint64_t MulAmt = C->getZExtValue();
23398 if (isPowerOf2_64(MulAmt) || MulAmt == 3 || MulAmt == 5 || MulAmt == 9)
23401 uint64_t MulAmt1 = 0;
23402 uint64_t MulAmt2 = 0;
23403 if ((MulAmt % 9) == 0) {
23405 MulAmt2 = MulAmt / 9;
23406 } else if ((MulAmt % 5) == 0) {
23408 MulAmt2 = MulAmt / 5;
23409 } else if ((MulAmt % 3) == 0) {
23411 MulAmt2 = MulAmt / 3;
23414 (isPowerOf2_64(MulAmt2) || MulAmt2 == 3 || MulAmt2 == 5 || MulAmt2 == 9)){
23417 if (isPowerOf2_64(MulAmt2) &&
23418 !(N->hasOneUse() && N->use_begin()->getOpcode() == ISD::ADD))
23419 // If second multiplifer is pow2, issue it first. We want the multiply by
23420 // 3, 5, or 9 to be folded into the addressing mode unless the lone use
23422 std::swap(MulAmt1, MulAmt2);
23425 if (isPowerOf2_64(MulAmt1))
23426 NewMul = DAG.getNode(ISD::SHL, DL, VT, N->getOperand(0),
23427 DAG.getConstant(Log2_64(MulAmt1), DL, MVT::i8));
23429 NewMul = DAG.getNode(X86ISD::MUL_IMM, DL, VT, N->getOperand(0),
23430 DAG.getConstant(MulAmt1, DL, VT));
23432 if (isPowerOf2_64(MulAmt2))
23433 NewMul = DAG.getNode(ISD::SHL, DL, VT, NewMul,
23434 DAG.getConstant(Log2_64(MulAmt2), DL, MVT::i8));
23436 NewMul = DAG.getNode(X86ISD::MUL_IMM, DL, VT, NewMul,
23437 DAG.getConstant(MulAmt2, DL, VT));
23439 // Do not add new nodes to DAG combiner worklist.
23440 DCI.CombineTo(N, NewMul, false);
23445 static SDValue PerformSHLCombine(SDNode *N, SelectionDAG &DAG) {
23446 SDValue N0 = N->getOperand(0);
23447 SDValue N1 = N->getOperand(1);
23448 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
23449 EVT VT = N0.getValueType();
23451 // fold (shl (and (setcc_c), c1), c2) -> (and setcc_c, (c1 << c2))
23452 // since the result of setcc_c is all zero's or all ones.
23453 if (VT.isInteger() && !VT.isVector() &&
23454 N1C && N0.getOpcode() == ISD::AND &&
23455 N0.getOperand(1).getOpcode() == ISD::Constant) {
23456 SDValue N00 = N0.getOperand(0);
23457 APInt Mask = cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
23458 APInt ShAmt = N1C->getAPIntValue();
23459 Mask = Mask.shl(ShAmt);
23460 bool MaskOK = false;
23461 // We can handle cases concerning bit-widening nodes containing setcc_c if
23462 // we carefully interrogate the mask to make sure we are semantics
23464 // The transform is not safe if the result of C1 << C2 exceeds the bitwidth
23465 // of the underlying setcc_c operation if the setcc_c was zero extended.
23466 // Consider the following example:
23467 // zext(setcc_c) -> i32 0x0000FFFF
23468 // c1 -> i32 0x0000FFFF
23469 // c2 -> i32 0x00000001
23470 // (shl (and (setcc_c), c1), c2) -> i32 0x0001FFFE
23471 // (and setcc_c, (c1 << c2)) -> i32 0x0000FFFE
23472 if (N00.getOpcode() == X86ISD::SETCC_CARRY) {
23474 } else if (N00.getOpcode() == ISD::SIGN_EXTEND &&
23475 N00.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY) {
23477 } else if ((N00.getOpcode() == ISD::ZERO_EXTEND ||
23478 N00.getOpcode() == ISD::ANY_EXTEND) &&
23479 N00.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY) {
23480 MaskOK = Mask.isIntN(N00.getOperand(0).getValueSizeInBits());
23482 if (MaskOK && Mask != 0) {
23484 return DAG.getNode(ISD::AND, DL, VT, N00, DAG.getConstant(Mask, DL, VT));
23488 // Hardware support for vector shifts is sparse which makes us scalarize the
23489 // vector operations in many cases. Also, on sandybridge ADD is faster than
23491 // (shl V, 1) -> add V,V
23492 if (auto *N1BV = dyn_cast<BuildVectorSDNode>(N1))
23493 if (auto *N1SplatC = N1BV->getConstantSplatNode()) {
23494 assert(N0.getValueType().isVector() && "Invalid vector shift type");
23495 // We shift all of the values by one. In many cases we do not have
23496 // hardware support for this operation. This is better expressed as an ADD
23498 if (N1SplatC->getAPIntValue() == 1)
23499 return DAG.getNode(ISD::ADD, SDLoc(N), VT, N0, N0);
23505 /// \brief Returns a vector of 0s if the node in input is a vector logical
23506 /// shift by a constant amount which is known to be bigger than or equal
23507 /// to the vector element size in bits.
23508 static SDValue performShiftToAllZeros(SDNode *N, SelectionDAG &DAG,
23509 const X86Subtarget *Subtarget) {
23510 EVT VT = N->getValueType(0);
23512 if (VT != MVT::v2i64 && VT != MVT::v4i32 && VT != MVT::v8i16 &&
23513 (!Subtarget->hasInt256() ||
23514 (VT != MVT::v4i64 && VT != MVT::v8i32 && VT != MVT::v16i16)))
23517 SDValue Amt = N->getOperand(1);
23519 if (auto *AmtBV = dyn_cast<BuildVectorSDNode>(Amt))
23520 if (auto *AmtSplat = AmtBV->getConstantSplatNode()) {
23521 APInt ShiftAmt = AmtSplat->getAPIntValue();
23522 unsigned MaxAmount = VT.getVectorElementType().getSizeInBits();
23524 // SSE2/AVX2 logical shifts always return a vector of 0s
23525 // if the shift amount is bigger than or equal to
23526 // the element size. The constant shift amount will be
23527 // encoded as a 8-bit immediate.
23528 if (ShiftAmt.trunc(8).uge(MaxAmount))
23529 return getZeroVector(VT, Subtarget, DAG, DL);
23535 /// PerformShiftCombine - Combine shifts.
23536 static SDValue PerformShiftCombine(SDNode* N, SelectionDAG &DAG,
23537 TargetLowering::DAGCombinerInfo &DCI,
23538 const X86Subtarget *Subtarget) {
23539 if (N->getOpcode() == ISD::SHL)
23540 if (SDValue V = PerformSHLCombine(N, DAG))
23543 // Try to fold this logical shift into a zero vector.
23544 if (N->getOpcode() != ISD::SRA)
23545 if (SDValue V = performShiftToAllZeros(N, DAG, Subtarget))
23551 // CMPEQCombine - Recognize the distinctive (AND (setcc ...) (setcc ..))
23552 // where both setccs reference the same FP CMP, and rewrite for CMPEQSS
23553 // and friends. Likewise for OR -> CMPNEQSS.
23554 static SDValue CMPEQCombine(SDNode *N, SelectionDAG &DAG,
23555 TargetLowering::DAGCombinerInfo &DCI,
23556 const X86Subtarget *Subtarget) {
23559 // SSE1 supports CMP{eq|ne}SS, and SSE2 added CMP{eq|ne}SD, but
23560 // we're requiring SSE2 for both.
23561 if (Subtarget->hasSSE2() && isAndOrOfSetCCs(SDValue(N, 0U), opcode)) {
23562 SDValue N0 = N->getOperand(0);
23563 SDValue N1 = N->getOperand(1);
23564 SDValue CMP0 = N0->getOperand(1);
23565 SDValue CMP1 = N1->getOperand(1);
23568 // The SETCCs should both refer to the same CMP.
23569 if (CMP0.getOpcode() != X86ISD::CMP || CMP0 != CMP1)
23572 SDValue CMP00 = CMP0->getOperand(0);
23573 SDValue CMP01 = CMP0->getOperand(1);
23574 EVT VT = CMP00.getValueType();
23576 if (VT == MVT::f32 || VT == MVT::f64) {
23577 bool ExpectingFlags = false;
23578 // Check for any users that want flags:
23579 for (SDNode::use_iterator UI = N->use_begin(), UE = N->use_end();
23580 !ExpectingFlags && UI != UE; ++UI)
23581 switch (UI->getOpcode()) {
23586 ExpectingFlags = true;
23588 case ISD::CopyToReg:
23589 case ISD::SIGN_EXTEND:
23590 case ISD::ZERO_EXTEND:
23591 case ISD::ANY_EXTEND:
23595 if (!ExpectingFlags) {
23596 enum X86::CondCode cc0 = (enum X86::CondCode)N0.getConstantOperandVal(0);
23597 enum X86::CondCode cc1 = (enum X86::CondCode)N1.getConstantOperandVal(0);
23599 if (cc1 == X86::COND_E || cc1 == X86::COND_NE) {
23600 X86::CondCode tmp = cc0;
23605 if ((cc0 == X86::COND_E && cc1 == X86::COND_NP) ||
23606 (cc0 == X86::COND_NE && cc1 == X86::COND_P)) {
23607 // FIXME: need symbolic constants for these magic numbers.
23608 // See X86ATTInstPrinter.cpp:printSSECC().
23609 unsigned x86cc = (cc0 == X86::COND_E) ? 0 : 4;
23610 if (Subtarget->hasAVX512()) {
23611 SDValue FSetCC = DAG.getNode(X86ISD::FSETCC, DL, MVT::i1, CMP00,
23613 DAG.getConstant(x86cc, DL, MVT::i8));
23614 if (N->getValueType(0) != MVT::i1)
23615 return DAG.getNode(ISD::ZERO_EXTEND, DL, N->getValueType(0),
23619 SDValue OnesOrZeroesF = DAG.getNode(X86ISD::FSETCC, DL,
23620 CMP00.getValueType(), CMP00, CMP01,
23621 DAG.getConstant(x86cc, DL,
23624 bool is64BitFP = (CMP00.getValueType() == MVT::f64);
23625 MVT IntVT = is64BitFP ? MVT::i64 : MVT::i32;
23627 if (is64BitFP && !Subtarget->is64Bit()) {
23628 // On a 32-bit target, we cannot bitcast the 64-bit float to a
23629 // 64-bit integer, since that's not a legal type. Since
23630 // OnesOrZeroesF is all ones of all zeroes, we don't need all the
23631 // bits, but can do this little dance to extract the lowest 32 bits
23632 // and work with those going forward.
23633 SDValue Vector64 = DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, MVT::v2f64,
23635 SDValue Vector32 = DAG.getBitcast(MVT::v4f32, Vector64);
23636 OnesOrZeroesF = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::f32,
23637 Vector32, DAG.getIntPtrConstant(0, DL));
23641 SDValue OnesOrZeroesI = DAG.getBitcast(IntVT, OnesOrZeroesF);
23642 SDValue ANDed = DAG.getNode(ISD::AND, DL, IntVT, OnesOrZeroesI,
23643 DAG.getConstant(1, DL, IntVT));
23644 SDValue OneBitOfTruth = DAG.getNode(ISD::TRUNCATE, DL, MVT::i8,
23646 return OneBitOfTruth;
23654 /// CanFoldXORWithAllOnes - Test whether the XOR operand is a AllOnes vector
23655 /// so it can be folded inside ANDNP.
23656 static bool CanFoldXORWithAllOnes(const SDNode *N) {
23657 EVT VT = N->getValueType(0);
23659 // Match direct AllOnes for 128 and 256-bit vectors
23660 if (ISD::isBuildVectorAllOnes(N))
23663 // Look through a bit convert.
23664 if (N->getOpcode() == ISD::BITCAST)
23665 N = N->getOperand(0).getNode();
23667 // Sometimes the operand may come from a insert_subvector building a 256-bit
23669 if (VT.is256BitVector() &&
23670 N->getOpcode() == ISD::INSERT_SUBVECTOR) {
23671 SDValue V1 = N->getOperand(0);
23672 SDValue V2 = N->getOperand(1);
23674 if (V1.getOpcode() == ISD::INSERT_SUBVECTOR &&
23675 V1.getOperand(0).getOpcode() == ISD::UNDEF &&
23676 ISD::isBuildVectorAllOnes(V1.getOperand(1).getNode()) &&
23677 ISD::isBuildVectorAllOnes(V2.getNode()))
23684 // On AVX/AVX2 the type v8i1 is legalized to v8i16, which is an XMM sized
23685 // register. In most cases we actually compare or select YMM-sized registers
23686 // and mixing the two types creates horrible code. This method optimizes
23687 // some of the transition sequences.
23688 static SDValue WidenMaskArithmetic(SDNode *N, SelectionDAG &DAG,
23689 TargetLowering::DAGCombinerInfo &DCI,
23690 const X86Subtarget *Subtarget) {
23691 EVT VT = N->getValueType(0);
23692 if (!VT.is256BitVector())
23695 assert((N->getOpcode() == ISD::ANY_EXTEND ||
23696 N->getOpcode() == ISD::ZERO_EXTEND ||
23697 N->getOpcode() == ISD::SIGN_EXTEND) && "Invalid Node");
23699 SDValue Narrow = N->getOperand(0);
23700 EVT NarrowVT = Narrow->getValueType(0);
23701 if (!NarrowVT.is128BitVector())
23704 if (Narrow->getOpcode() != ISD::XOR &&
23705 Narrow->getOpcode() != ISD::AND &&
23706 Narrow->getOpcode() != ISD::OR)
23709 SDValue N0 = Narrow->getOperand(0);
23710 SDValue N1 = Narrow->getOperand(1);
23713 // The Left side has to be a trunc.
23714 if (N0.getOpcode() != ISD::TRUNCATE)
23717 // The type of the truncated inputs.
23718 EVT WideVT = N0->getOperand(0)->getValueType(0);
23722 // The right side has to be a 'trunc' or a constant vector.
23723 bool RHSTrunc = N1.getOpcode() == ISD::TRUNCATE;
23724 ConstantSDNode *RHSConstSplat = nullptr;
23725 if (auto *RHSBV = dyn_cast<BuildVectorSDNode>(N1))
23726 RHSConstSplat = RHSBV->getConstantSplatNode();
23727 if (!RHSTrunc && !RHSConstSplat)
23730 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
23732 if (!TLI.isOperationLegalOrPromote(Narrow->getOpcode(), WideVT))
23735 // Set N0 and N1 to hold the inputs to the new wide operation.
23736 N0 = N0->getOperand(0);
23737 if (RHSConstSplat) {
23738 N1 = DAG.getNode(ISD::ZERO_EXTEND, DL, WideVT.getScalarType(),
23739 SDValue(RHSConstSplat, 0));
23740 SmallVector<SDValue, 8> C(WideVT.getVectorNumElements(), N1);
23741 N1 = DAG.getNode(ISD::BUILD_VECTOR, DL, WideVT, C);
23742 } else if (RHSTrunc) {
23743 N1 = N1->getOperand(0);
23746 // Generate the wide operation.
23747 SDValue Op = DAG.getNode(Narrow->getOpcode(), DL, WideVT, N0, N1);
23748 unsigned Opcode = N->getOpcode();
23750 case ISD::ANY_EXTEND:
23752 case ISD::ZERO_EXTEND: {
23753 unsigned InBits = NarrowVT.getScalarType().getSizeInBits();
23754 APInt Mask = APInt::getAllOnesValue(InBits);
23755 Mask = Mask.zext(VT.getScalarType().getSizeInBits());
23756 return DAG.getNode(ISD::AND, DL, VT,
23757 Op, DAG.getConstant(Mask, DL, VT));
23759 case ISD::SIGN_EXTEND:
23760 return DAG.getNode(ISD::SIGN_EXTEND_INREG, DL, VT,
23761 Op, DAG.getValueType(NarrowVT));
23763 llvm_unreachable("Unexpected opcode");
23767 static SDValue VectorZextCombine(SDNode *N, SelectionDAG &DAG,
23768 TargetLowering::DAGCombinerInfo &DCI,
23769 const X86Subtarget *Subtarget) {
23770 SDValue N0 = N->getOperand(0);
23771 SDValue N1 = N->getOperand(1);
23774 // A vector zext_in_reg may be represented as a shuffle,
23775 // feeding into a bitcast (this represents anyext) feeding into
23776 // an and with a mask.
23777 // We'd like to try to combine that into a shuffle with zero
23778 // plus a bitcast, removing the and.
23779 if (N0.getOpcode() != ISD::BITCAST ||
23780 N0.getOperand(0).getOpcode() != ISD::VECTOR_SHUFFLE)
23783 // The other side of the AND should be a splat of 2^C, where C
23784 // is the number of bits in the source type.
23785 if (N1.getOpcode() == ISD::BITCAST)
23786 N1 = N1.getOperand(0);
23787 if (N1.getOpcode() != ISD::BUILD_VECTOR)
23789 BuildVectorSDNode *Vector = cast<BuildVectorSDNode>(N1);
23791 ShuffleVectorSDNode *Shuffle = cast<ShuffleVectorSDNode>(N0.getOperand(0));
23792 EVT SrcType = Shuffle->getValueType(0);
23794 // We expect a single-source shuffle
23795 if (Shuffle->getOperand(1)->getOpcode() != ISD::UNDEF)
23798 unsigned SrcSize = SrcType.getScalarSizeInBits();
23800 APInt SplatValue, SplatUndef;
23801 unsigned SplatBitSize;
23803 if (!Vector->isConstantSplat(SplatValue, SplatUndef,
23804 SplatBitSize, HasAnyUndefs))
23807 unsigned ResSize = N1.getValueType().getScalarSizeInBits();
23808 // Make sure the splat matches the mask we expect
23809 if (SplatBitSize > ResSize ||
23810 (SplatValue + 1).exactLogBase2() != (int)SrcSize)
23813 // Make sure the input and output size make sense
23814 if (SrcSize >= ResSize || ResSize % SrcSize)
23817 // We expect a shuffle of the form <0, u, u, u, 1, u, u, u...>
23818 // The number of u's between each two values depends on the ratio between
23819 // the source and dest type.
23820 unsigned ZextRatio = ResSize / SrcSize;
23821 bool IsZext = true;
23822 for (unsigned i = 0; i < SrcType.getVectorNumElements(); ++i) {
23823 if (i % ZextRatio) {
23824 if (Shuffle->getMaskElt(i) > 0) {
23830 if (Shuffle->getMaskElt(i) != (int)(i / ZextRatio)) {
23831 // Expected element number
23841 // Ok, perform the transformation - replace the shuffle with
23842 // a shuffle of the form <0, k, k, k, 1, k, k, k> with zero
23843 // (instead of undef) where the k elements come from the zero vector.
23844 SmallVector<int, 8> Mask;
23845 unsigned NumElems = SrcType.getVectorNumElements();
23846 for (unsigned i = 0; i < NumElems; ++i)
23848 Mask.push_back(NumElems);
23850 Mask.push_back(i / ZextRatio);
23852 SDValue NewShuffle = DAG.getVectorShuffle(Shuffle->getValueType(0), DL,
23853 Shuffle->getOperand(0), DAG.getConstant(0, DL, SrcType), Mask);
23854 return DAG.getBitcast(N0.getValueType(), NewShuffle);
23857 static SDValue PerformAndCombine(SDNode *N, SelectionDAG &DAG,
23858 TargetLowering::DAGCombinerInfo &DCI,
23859 const X86Subtarget *Subtarget) {
23860 if (DCI.isBeforeLegalizeOps())
23863 if (SDValue Zext = VectorZextCombine(N, DAG, DCI, Subtarget))
23866 if (SDValue R = CMPEQCombine(N, DAG, DCI, Subtarget))
23869 EVT VT = N->getValueType(0);
23870 SDValue N0 = N->getOperand(0);
23871 SDValue N1 = N->getOperand(1);
23874 // Create BEXTR instructions
23875 // BEXTR is ((X >> imm) & (2**size-1))
23876 if (VT == MVT::i32 || VT == MVT::i64) {
23877 // Check for BEXTR.
23878 if ((Subtarget->hasBMI() || Subtarget->hasTBM()) &&
23879 (N0.getOpcode() == ISD::SRA || N0.getOpcode() == ISD::SRL)) {
23880 ConstantSDNode *MaskNode = dyn_cast<ConstantSDNode>(N1);
23881 ConstantSDNode *ShiftNode = dyn_cast<ConstantSDNode>(N0.getOperand(1));
23882 if (MaskNode && ShiftNode) {
23883 uint64_t Mask = MaskNode->getZExtValue();
23884 uint64_t Shift = ShiftNode->getZExtValue();
23885 if (isMask_64(Mask)) {
23886 uint64_t MaskSize = countPopulation(Mask);
23887 if (Shift + MaskSize <= VT.getSizeInBits())
23888 return DAG.getNode(X86ISD::BEXTR, DL, VT, N0.getOperand(0),
23889 DAG.getConstant(Shift | (MaskSize << 8), DL,
23898 // Want to form ANDNP nodes:
23899 // 1) In the hopes of then easily combining them with OR and AND nodes
23900 // to form PBLEND/PSIGN.
23901 // 2) To match ANDN packed intrinsics
23902 if (VT != MVT::v2i64 && VT != MVT::v4i64)
23905 // Check LHS for vnot
23906 if (N0.getOpcode() == ISD::XOR &&
23907 //ISD::isBuildVectorAllOnes(N0.getOperand(1).getNode()))
23908 CanFoldXORWithAllOnes(N0.getOperand(1).getNode()))
23909 return DAG.getNode(X86ISD::ANDNP, DL, VT, N0.getOperand(0), N1);
23911 // Check RHS for vnot
23912 if (N1.getOpcode() == ISD::XOR &&
23913 //ISD::isBuildVectorAllOnes(N1.getOperand(1).getNode()))
23914 CanFoldXORWithAllOnes(N1.getOperand(1).getNode()))
23915 return DAG.getNode(X86ISD::ANDNP, DL, VT, N1.getOperand(0), N0);
23920 static SDValue PerformOrCombine(SDNode *N, SelectionDAG &DAG,
23921 TargetLowering::DAGCombinerInfo &DCI,
23922 const X86Subtarget *Subtarget) {
23923 if (DCI.isBeforeLegalizeOps())
23926 if (SDValue R = CMPEQCombine(N, DAG, DCI, Subtarget))
23929 SDValue N0 = N->getOperand(0);
23930 SDValue N1 = N->getOperand(1);
23931 EVT VT = N->getValueType(0);
23933 // look for psign/blend
23934 if (VT == MVT::v2i64 || VT == MVT::v4i64) {
23935 if (!Subtarget->hasSSSE3() ||
23936 (VT == MVT::v4i64 && !Subtarget->hasInt256()))
23939 // Canonicalize pandn to RHS
23940 if (N0.getOpcode() == X86ISD::ANDNP)
23942 // or (and (m, y), (pandn m, x))
23943 if (N0.getOpcode() == ISD::AND && N1.getOpcode() == X86ISD::ANDNP) {
23944 SDValue Mask = N1.getOperand(0);
23945 SDValue X = N1.getOperand(1);
23947 if (N0.getOperand(0) == Mask)
23948 Y = N0.getOperand(1);
23949 if (N0.getOperand(1) == Mask)
23950 Y = N0.getOperand(0);
23952 // Check to see if the mask appeared in both the AND and ANDNP and
23956 // Validate that X, Y, and Mask are BIT_CONVERTS, and see through them.
23957 // Look through mask bitcast.
23958 if (Mask.getOpcode() == ISD::BITCAST)
23959 Mask = Mask.getOperand(0);
23960 if (X.getOpcode() == ISD::BITCAST)
23961 X = X.getOperand(0);
23962 if (Y.getOpcode() == ISD::BITCAST)
23963 Y = Y.getOperand(0);
23965 EVT MaskVT = Mask.getValueType();
23967 // Validate that the Mask operand is a vector sra node.
23968 // FIXME: what to do for bytes, since there is a psignb/pblendvb, but
23969 // there is no psrai.b
23970 unsigned EltBits = MaskVT.getVectorElementType().getSizeInBits();
23971 unsigned SraAmt = ~0;
23972 if (Mask.getOpcode() == ISD::SRA) {
23973 if (auto *AmtBV = dyn_cast<BuildVectorSDNode>(Mask.getOperand(1)))
23974 if (auto *AmtConst = AmtBV->getConstantSplatNode())
23975 SraAmt = AmtConst->getZExtValue();
23976 } else if (Mask.getOpcode() == X86ISD::VSRAI) {
23977 SDValue SraC = Mask.getOperand(1);
23978 SraAmt = cast<ConstantSDNode>(SraC)->getZExtValue();
23980 if ((SraAmt + 1) != EltBits)
23985 // Now we know we at least have a plendvb with the mask val. See if
23986 // we can form a psignb/w/d.
23987 // psign = x.type == y.type == mask.type && y = sub(0, x);
23988 if (Y.getOpcode() == ISD::SUB && Y.getOperand(1) == X &&
23989 ISD::isBuildVectorAllZeros(Y.getOperand(0).getNode()) &&
23990 X.getValueType() == MaskVT && Y.getValueType() == MaskVT) {
23991 assert((EltBits == 8 || EltBits == 16 || EltBits == 32) &&
23992 "Unsupported VT for PSIGN");
23993 Mask = DAG.getNode(X86ISD::PSIGN, DL, MaskVT, X, Mask.getOperand(0));
23994 return DAG.getBitcast(VT, Mask);
23996 // PBLENDVB only available on SSE 4.1
23997 if (!Subtarget->hasSSE41())
24000 EVT BlendVT = (VT == MVT::v4i64) ? MVT::v32i8 : MVT::v16i8;
24002 X = DAG.getBitcast(BlendVT, X);
24003 Y = DAG.getBitcast(BlendVT, Y);
24004 Mask = DAG.getBitcast(BlendVT, Mask);
24005 Mask = DAG.getNode(ISD::VSELECT, DL, BlendVT, Mask, Y, X);
24006 return DAG.getBitcast(VT, Mask);
24010 if (VT != MVT::i16 && VT != MVT::i32 && VT != MVT::i64)
24013 // fold (or (x << c) | (y >> (64 - c))) ==> (shld64 x, y, c)
24014 bool OptForSize = DAG.getMachineFunction().getFunction()->optForSize();
24016 // SHLD/SHRD instructions have lower register pressure, but on some
24017 // platforms they have higher latency than the equivalent
24018 // series of shifts/or that would otherwise be generated.
24019 // Don't fold (or (x << c) | (y >> (64 - c))) if SHLD/SHRD instructions
24020 // have higher latencies and we are not optimizing for size.
24021 if (!OptForSize && Subtarget->isSHLDSlow())
24024 if (N0.getOpcode() == ISD::SRL && N1.getOpcode() == ISD::SHL)
24026 if (N0.getOpcode() != ISD::SHL || N1.getOpcode() != ISD::SRL)
24028 if (!N0.hasOneUse() || !N1.hasOneUse())
24031 SDValue ShAmt0 = N0.getOperand(1);
24032 if (ShAmt0.getValueType() != MVT::i8)
24034 SDValue ShAmt1 = N1.getOperand(1);
24035 if (ShAmt1.getValueType() != MVT::i8)
24037 if (ShAmt0.getOpcode() == ISD::TRUNCATE)
24038 ShAmt0 = ShAmt0.getOperand(0);
24039 if (ShAmt1.getOpcode() == ISD::TRUNCATE)
24040 ShAmt1 = ShAmt1.getOperand(0);
24043 unsigned Opc = X86ISD::SHLD;
24044 SDValue Op0 = N0.getOperand(0);
24045 SDValue Op1 = N1.getOperand(0);
24046 if (ShAmt0.getOpcode() == ISD::SUB) {
24047 Opc = X86ISD::SHRD;
24048 std::swap(Op0, Op1);
24049 std::swap(ShAmt0, ShAmt1);
24052 unsigned Bits = VT.getSizeInBits();
24053 if (ShAmt1.getOpcode() == ISD::SUB) {
24054 SDValue Sum = ShAmt1.getOperand(0);
24055 if (ConstantSDNode *SumC = dyn_cast<ConstantSDNode>(Sum)) {
24056 SDValue ShAmt1Op1 = ShAmt1.getOperand(1);
24057 if (ShAmt1Op1.getNode()->getOpcode() == ISD::TRUNCATE)
24058 ShAmt1Op1 = ShAmt1Op1.getOperand(0);
24059 if (SumC->getSExtValue() == Bits && ShAmt1Op1 == ShAmt0)
24060 return DAG.getNode(Opc, DL, VT,
24062 DAG.getNode(ISD::TRUNCATE, DL,
24065 } else if (ConstantSDNode *ShAmt1C = dyn_cast<ConstantSDNode>(ShAmt1)) {
24066 ConstantSDNode *ShAmt0C = dyn_cast<ConstantSDNode>(ShAmt0);
24068 ShAmt0C->getSExtValue() + ShAmt1C->getSExtValue() == Bits)
24069 return DAG.getNode(Opc, DL, VT,
24070 N0.getOperand(0), N1.getOperand(0),
24071 DAG.getNode(ISD::TRUNCATE, DL,
24078 // Generate NEG and CMOV for integer abs.
24079 static SDValue performIntegerAbsCombine(SDNode *N, SelectionDAG &DAG) {
24080 EVT VT = N->getValueType(0);
24082 // Since X86 does not have CMOV for 8-bit integer, we don't convert
24083 // 8-bit integer abs to NEG and CMOV.
24084 if (VT.isInteger() && VT.getSizeInBits() == 8)
24087 SDValue N0 = N->getOperand(0);
24088 SDValue N1 = N->getOperand(1);
24091 // Check pattern of XOR(ADD(X,Y), Y) where Y is SRA(X, size(X)-1)
24092 // and change it to SUB and CMOV.
24093 if (VT.isInteger() && N->getOpcode() == ISD::XOR &&
24094 N0.getOpcode() == ISD::ADD &&
24095 N0.getOperand(1) == N1 &&
24096 N1.getOpcode() == ISD::SRA &&
24097 N1.getOperand(0) == N0.getOperand(0))
24098 if (ConstantSDNode *Y1C = dyn_cast<ConstantSDNode>(N1.getOperand(1)))
24099 if (Y1C->getAPIntValue() == VT.getSizeInBits()-1) {
24100 // Generate SUB & CMOV.
24101 SDValue Neg = DAG.getNode(X86ISD::SUB, DL, DAG.getVTList(VT, MVT::i32),
24102 DAG.getConstant(0, DL, VT), N0.getOperand(0));
24104 SDValue Ops[] = { N0.getOperand(0), Neg,
24105 DAG.getConstant(X86::COND_GE, DL, MVT::i8),
24106 SDValue(Neg.getNode(), 1) };
24107 return DAG.getNode(X86ISD::CMOV, DL, DAG.getVTList(VT, MVT::Glue), Ops);
24112 // Try to turn tests against the signbit in the form of:
24113 // XOR(TRUNCATE(SRL(X, size(X)-1)), 1)
24116 static SDValue foldXorTruncShiftIntoCmp(SDNode *N, SelectionDAG &DAG) {
24117 // This is only worth doing if the output type is i8.
24118 if (N->getValueType(0) != MVT::i8)
24121 SDValue N0 = N->getOperand(0);
24122 SDValue N1 = N->getOperand(1);
24124 // We should be performing an xor against a truncated shift.
24125 if (N0.getOpcode() != ISD::TRUNCATE || !N0.hasOneUse())
24128 // Make sure we are performing an xor against one.
24129 if (!isa<ConstantSDNode>(N1) || !cast<ConstantSDNode>(N1)->isOne())
24132 // SetCC on x86 zero extends so only act on this if it's a logical shift.
24133 SDValue Shift = N0.getOperand(0);
24134 if (Shift.getOpcode() != ISD::SRL || !Shift.hasOneUse())
24137 // Make sure we are truncating from one of i16, i32 or i64.
24138 EVT ShiftTy = Shift.getValueType();
24139 if (ShiftTy != MVT::i16 && ShiftTy != MVT::i32 && ShiftTy != MVT::i64)
24142 // Make sure the shift amount extracts the sign bit.
24143 if (!isa<ConstantSDNode>(Shift.getOperand(1)) ||
24144 Shift.getConstantOperandVal(1) != ShiftTy.getSizeInBits() - 1)
24147 // Create a greater-than comparison against -1.
24148 // N.B. Using SETGE against 0 works but we want a canonical looking
24149 // comparison, using SETGT matches up with what TranslateX86CC.
24151 SDValue ShiftOp = Shift.getOperand(0);
24152 EVT ShiftOpTy = ShiftOp.getValueType();
24153 SDValue Cond = DAG.getSetCC(DL, MVT::i8, ShiftOp,
24154 DAG.getConstant(-1, DL, ShiftOpTy), ISD::SETGT);
24158 static SDValue PerformXorCombine(SDNode *N, SelectionDAG &DAG,
24159 TargetLowering::DAGCombinerInfo &DCI,
24160 const X86Subtarget *Subtarget) {
24161 if (DCI.isBeforeLegalizeOps())
24164 if (SDValue RV = foldXorTruncShiftIntoCmp(N, DAG))
24167 if (Subtarget->hasCMov())
24168 if (SDValue RV = performIntegerAbsCombine(N, DAG))
24174 /// PerformLOADCombine - Do target-specific dag combines on LOAD nodes.
24175 static SDValue PerformLOADCombine(SDNode *N, SelectionDAG &DAG,
24176 TargetLowering::DAGCombinerInfo &DCI,
24177 const X86Subtarget *Subtarget) {
24178 LoadSDNode *Ld = cast<LoadSDNode>(N);
24179 EVT RegVT = Ld->getValueType(0);
24180 EVT MemVT = Ld->getMemoryVT();
24182 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
24184 // For chips with slow 32-byte unaligned loads, break the 32-byte operation
24185 // into two 16-byte operations.
24186 ISD::LoadExtType Ext = Ld->getExtensionType();
24188 unsigned AddressSpace = Ld->getAddressSpace();
24189 unsigned Alignment = Ld->getAlignment();
24190 if (RegVT.is256BitVector() && !DCI.isBeforeLegalizeOps() &&
24191 Ext == ISD::NON_EXTLOAD &&
24192 TLI.allowsMemoryAccess(*DAG.getContext(), DAG.getDataLayout(), RegVT,
24193 AddressSpace, Alignment, &Fast) && !Fast) {
24194 unsigned NumElems = RegVT.getVectorNumElements();
24198 SDValue Ptr = Ld->getBasePtr();
24199 SDValue Increment =
24200 DAG.getConstant(16, dl, TLI.getPointerTy(DAG.getDataLayout()));
24202 EVT HalfVT = EVT::getVectorVT(*DAG.getContext(), MemVT.getScalarType(),
24204 SDValue Load1 = DAG.getLoad(HalfVT, dl, Ld->getChain(), Ptr,
24205 Ld->getPointerInfo(), Ld->isVolatile(),
24206 Ld->isNonTemporal(), Ld->isInvariant(),
24208 Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr, Increment);
24209 SDValue Load2 = DAG.getLoad(HalfVT, dl, Ld->getChain(), Ptr,
24210 Ld->getPointerInfo(), Ld->isVolatile(),
24211 Ld->isNonTemporal(), Ld->isInvariant(),
24212 std::min(16U, Alignment));
24213 SDValue TF = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
24215 Load2.getValue(1));
24217 SDValue NewVec = DAG.getUNDEF(RegVT);
24218 NewVec = Insert128BitVector(NewVec, Load1, 0, DAG, dl);
24219 NewVec = Insert128BitVector(NewVec, Load2, NumElems/2, DAG, dl);
24220 return DCI.CombineTo(N, NewVec, TF, true);
24226 /// PerformMLOADCombine - Resolve extending loads
24227 static SDValue PerformMLOADCombine(SDNode *N, SelectionDAG &DAG,
24228 TargetLowering::DAGCombinerInfo &DCI,
24229 const X86Subtarget *Subtarget) {
24230 MaskedLoadSDNode *Mld = cast<MaskedLoadSDNode>(N);
24231 if (Mld->getExtensionType() != ISD::SEXTLOAD)
24234 EVT VT = Mld->getValueType(0);
24235 unsigned NumElems = VT.getVectorNumElements();
24236 EVT LdVT = Mld->getMemoryVT();
24239 assert(LdVT != VT && "Cannot extend to the same type");
24240 unsigned ToSz = VT.getVectorElementType().getSizeInBits();
24241 unsigned FromSz = LdVT.getVectorElementType().getSizeInBits();
24242 // From, To sizes and ElemCount must be pow of two
24243 assert (isPowerOf2_32(NumElems * FromSz * ToSz) &&
24244 "Unexpected size for extending masked load");
24246 unsigned SizeRatio = ToSz / FromSz;
24247 assert(SizeRatio * NumElems * FromSz == VT.getSizeInBits());
24249 // Create a type on which we perform the shuffle
24250 EVT WideVecVT = EVT::getVectorVT(*DAG.getContext(),
24251 LdVT.getScalarType(), NumElems*SizeRatio);
24252 assert(WideVecVT.getSizeInBits() == VT.getSizeInBits());
24254 // Convert Src0 value
24255 SDValue WideSrc0 = DAG.getBitcast(WideVecVT, Mld->getSrc0());
24256 if (Mld->getSrc0().getOpcode() != ISD::UNDEF) {
24257 SmallVector<int, 16> ShuffleVec(NumElems * SizeRatio, -1);
24258 for (unsigned i = 0; i != NumElems; ++i)
24259 ShuffleVec[i] = i * SizeRatio;
24261 // Can't shuffle using an illegal type.
24262 assert (DAG.getTargetLoweringInfo().isTypeLegal(WideVecVT)
24263 && "WideVecVT should be legal");
24264 WideSrc0 = DAG.getVectorShuffle(WideVecVT, dl, WideSrc0,
24265 DAG.getUNDEF(WideVecVT), &ShuffleVec[0]);
24267 // Prepare the new mask
24269 SDValue Mask = Mld->getMask();
24270 if (Mask.getValueType() == VT) {
24271 // Mask and original value have the same type
24272 NewMask = DAG.getBitcast(WideVecVT, Mask);
24273 SmallVector<int, 16> ShuffleVec(NumElems * SizeRatio, -1);
24274 for (unsigned i = 0; i != NumElems; ++i)
24275 ShuffleVec[i] = i * SizeRatio;
24276 for (unsigned i = NumElems; i != NumElems*SizeRatio; ++i)
24277 ShuffleVec[i] = NumElems*SizeRatio;
24278 NewMask = DAG.getVectorShuffle(WideVecVT, dl, NewMask,
24279 DAG.getConstant(0, dl, WideVecVT),
24283 assert(Mask.getValueType().getVectorElementType() == MVT::i1);
24284 unsigned WidenNumElts = NumElems*SizeRatio;
24285 unsigned MaskNumElts = VT.getVectorNumElements();
24286 EVT NewMaskVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
24289 unsigned NumConcat = WidenNumElts / MaskNumElts;
24290 SmallVector<SDValue, 16> Ops(NumConcat);
24291 SDValue ZeroVal = DAG.getConstant(0, dl, Mask.getValueType());
24293 for (unsigned i = 1; i != NumConcat; ++i)
24296 NewMask = DAG.getNode(ISD::CONCAT_VECTORS, dl, NewMaskVT, Ops);
24299 SDValue WideLd = DAG.getMaskedLoad(WideVecVT, dl, Mld->getChain(),
24300 Mld->getBasePtr(), NewMask, WideSrc0,
24301 Mld->getMemoryVT(), Mld->getMemOperand(),
24303 SDValue NewVec = DAG.getNode(X86ISD::VSEXT, dl, VT, WideLd);
24304 return DCI.CombineTo(N, NewVec, WideLd.getValue(1), true);
24307 /// PerformMSTORECombine - Resolve truncating stores
24308 static SDValue PerformMSTORECombine(SDNode *N, SelectionDAG &DAG,
24309 const X86Subtarget *Subtarget) {
24310 MaskedStoreSDNode *Mst = cast<MaskedStoreSDNode>(N);
24311 if (!Mst->isTruncatingStore())
24314 EVT VT = Mst->getValue().getValueType();
24315 unsigned NumElems = VT.getVectorNumElements();
24316 EVT StVT = Mst->getMemoryVT();
24319 assert(StVT != VT && "Cannot truncate to the same type");
24320 unsigned FromSz = VT.getVectorElementType().getSizeInBits();
24321 unsigned ToSz = StVT.getVectorElementType().getSizeInBits();
24323 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
24325 // The truncating store is legal in some cases. For example
24326 // vpmovqb, vpmovqw, vpmovqd, vpmovdb, vpmovdw
24327 // are designated for truncate store.
24328 // In this case we don't need any further transformations.
24329 if (TLI.isTruncStoreLegal(VT, StVT))
24332 // From, To sizes and ElemCount must be pow of two
24333 assert (isPowerOf2_32(NumElems * FromSz * ToSz) &&
24334 "Unexpected size for truncating masked store");
24335 // We are going to use the original vector elt for storing.
24336 // Accumulated smaller vector elements must be a multiple of the store size.
24337 assert (((NumElems * FromSz) % ToSz) == 0 &&
24338 "Unexpected ratio for truncating masked store");
24340 unsigned SizeRatio = FromSz / ToSz;
24341 assert(SizeRatio * NumElems * ToSz == VT.getSizeInBits());
24343 // Create a type on which we perform the shuffle
24344 EVT WideVecVT = EVT::getVectorVT(*DAG.getContext(),
24345 StVT.getScalarType(), NumElems*SizeRatio);
24347 assert(WideVecVT.getSizeInBits() == VT.getSizeInBits());
24349 SDValue WideVec = DAG.getBitcast(WideVecVT, Mst->getValue());
24350 SmallVector<int, 16> ShuffleVec(NumElems * SizeRatio, -1);
24351 for (unsigned i = 0; i != NumElems; ++i)
24352 ShuffleVec[i] = i * SizeRatio;
24354 // Can't shuffle using an illegal type.
24355 assert (DAG.getTargetLoweringInfo().isTypeLegal(WideVecVT)
24356 && "WideVecVT should be legal");
24358 SDValue TruncatedVal = DAG.getVectorShuffle(WideVecVT, dl, WideVec,
24359 DAG.getUNDEF(WideVecVT),
24363 SDValue Mask = Mst->getMask();
24364 if (Mask.getValueType() == VT) {
24365 // Mask and original value have the same type
24366 NewMask = DAG.getBitcast(WideVecVT, Mask);
24367 for (unsigned i = 0; i != NumElems; ++i)
24368 ShuffleVec[i] = i * SizeRatio;
24369 for (unsigned i = NumElems; i != NumElems*SizeRatio; ++i)
24370 ShuffleVec[i] = NumElems*SizeRatio;
24371 NewMask = DAG.getVectorShuffle(WideVecVT, dl, NewMask,
24372 DAG.getConstant(0, dl, WideVecVT),
24376 assert(Mask.getValueType().getVectorElementType() == MVT::i1);
24377 unsigned WidenNumElts = NumElems*SizeRatio;
24378 unsigned MaskNumElts = VT.getVectorNumElements();
24379 EVT NewMaskVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
24382 unsigned NumConcat = WidenNumElts / MaskNumElts;
24383 SmallVector<SDValue, 16> Ops(NumConcat);
24384 SDValue ZeroVal = DAG.getConstant(0, dl, Mask.getValueType());
24386 for (unsigned i = 1; i != NumConcat; ++i)
24389 NewMask = DAG.getNode(ISD::CONCAT_VECTORS, dl, NewMaskVT, Ops);
24392 return DAG.getMaskedStore(Mst->getChain(), dl, TruncatedVal, Mst->getBasePtr(),
24393 NewMask, StVT, Mst->getMemOperand(), false);
24395 /// PerformSTORECombine - Do target-specific dag combines on STORE nodes.
24396 static SDValue PerformSTORECombine(SDNode *N, SelectionDAG &DAG,
24397 const X86Subtarget *Subtarget) {
24398 StoreSDNode *St = cast<StoreSDNode>(N);
24399 EVT VT = St->getValue().getValueType();
24400 EVT StVT = St->getMemoryVT();
24402 SDValue StoredVal = St->getOperand(1);
24403 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
24405 // If we are saving a concatenation of two XMM registers and 32-byte stores
24406 // are slow, such as on Sandy Bridge, perform two 16-byte stores.
24408 unsigned AddressSpace = St->getAddressSpace();
24409 unsigned Alignment = St->getAlignment();
24410 if (VT.is256BitVector() && StVT == VT &&
24411 TLI.allowsMemoryAccess(*DAG.getContext(), DAG.getDataLayout(), VT,
24412 AddressSpace, Alignment, &Fast) && !Fast) {
24413 unsigned NumElems = VT.getVectorNumElements();
24417 SDValue Value0 = Extract128BitVector(StoredVal, 0, DAG, dl);
24418 SDValue Value1 = Extract128BitVector(StoredVal, NumElems/2, DAG, dl);
24421 DAG.getConstant(16, dl, TLI.getPointerTy(DAG.getDataLayout()));
24422 SDValue Ptr0 = St->getBasePtr();
24423 SDValue Ptr1 = DAG.getNode(ISD::ADD, dl, Ptr0.getValueType(), Ptr0, Stride);
24425 SDValue Ch0 = DAG.getStore(St->getChain(), dl, Value0, Ptr0,
24426 St->getPointerInfo(), St->isVolatile(),
24427 St->isNonTemporal(), Alignment);
24428 SDValue Ch1 = DAG.getStore(St->getChain(), dl, Value1, Ptr1,
24429 St->getPointerInfo(), St->isVolatile(),
24430 St->isNonTemporal(),
24431 std::min(16U, Alignment));
24432 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Ch0, Ch1);
24435 // Optimize trunc store (of multiple scalars) to shuffle and store.
24436 // First, pack all of the elements in one place. Next, store to memory
24437 // in fewer chunks.
24438 if (St->isTruncatingStore() && VT.isVector()) {
24439 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
24440 unsigned NumElems = VT.getVectorNumElements();
24441 assert(StVT != VT && "Cannot truncate to the same type");
24442 unsigned FromSz = VT.getVectorElementType().getSizeInBits();
24443 unsigned ToSz = StVT.getVectorElementType().getSizeInBits();
24445 // The truncating store is legal in some cases. For example
24446 // vpmovqb, vpmovqw, vpmovqd, vpmovdb, vpmovdw
24447 // are designated for truncate store.
24448 // In this case we don't need any further transformations.
24449 if (TLI.isTruncStoreLegal(VT, StVT))
24452 // From, To sizes and ElemCount must be pow of two
24453 if (!isPowerOf2_32(NumElems * FromSz * ToSz)) return SDValue();
24454 // We are going to use the original vector elt for storing.
24455 // Accumulated smaller vector elements must be a multiple of the store size.
24456 if (0 != (NumElems * FromSz) % ToSz) return SDValue();
24458 unsigned SizeRatio = FromSz / ToSz;
24460 assert(SizeRatio * NumElems * ToSz == VT.getSizeInBits());
24462 // Create a type on which we perform the shuffle
24463 EVT WideVecVT = EVT::getVectorVT(*DAG.getContext(),
24464 StVT.getScalarType(), NumElems*SizeRatio);
24466 assert(WideVecVT.getSizeInBits() == VT.getSizeInBits());
24468 SDValue WideVec = DAG.getBitcast(WideVecVT, St->getValue());
24469 SmallVector<int, 8> ShuffleVec(NumElems * SizeRatio, -1);
24470 for (unsigned i = 0; i != NumElems; ++i)
24471 ShuffleVec[i] = i * SizeRatio;
24473 // Can't shuffle using an illegal type.
24474 if (!TLI.isTypeLegal(WideVecVT))
24477 SDValue Shuff = DAG.getVectorShuffle(WideVecVT, dl, WideVec,
24478 DAG.getUNDEF(WideVecVT),
24480 // At this point all of the data is stored at the bottom of the
24481 // register. We now need to save it to mem.
24483 // Find the largest store unit
24484 MVT StoreType = MVT::i8;
24485 for (MVT Tp : MVT::integer_valuetypes()) {
24486 if (TLI.isTypeLegal(Tp) && Tp.getSizeInBits() <= NumElems * ToSz)
24490 // On 32bit systems, we can't save 64bit integers. Try bitcasting to F64.
24491 if (TLI.isTypeLegal(MVT::f64) && StoreType.getSizeInBits() < 64 &&
24492 (64 <= NumElems * ToSz))
24493 StoreType = MVT::f64;
24495 // Bitcast the original vector into a vector of store-size units
24496 EVT StoreVecVT = EVT::getVectorVT(*DAG.getContext(),
24497 StoreType, VT.getSizeInBits()/StoreType.getSizeInBits());
24498 assert(StoreVecVT.getSizeInBits() == VT.getSizeInBits());
24499 SDValue ShuffWide = DAG.getBitcast(StoreVecVT, Shuff);
24500 SmallVector<SDValue, 8> Chains;
24501 SDValue Increment = DAG.getConstant(StoreType.getSizeInBits() / 8, dl,
24502 TLI.getPointerTy(DAG.getDataLayout()));
24503 SDValue Ptr = St->getBasePtr();
24505 // Perform one or more big stores into memory.
24506 for (unsigned i=0, e=(ToSz*NumElems)/StoreType.getSizeInBits(); i!=e; ++i) {
24507 SDValue SubVec = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl,
24508 StoreType, ShuffWide,
24509 DAG.getIntPtrConstant(i, dl));
24510 SDValue Ch = DAG.getStore(St->getChain(), dl, SubVec, Ptr,
24511 St->getPointerInfo(), St->isVolatile(),
24512 St->isNonTemporal(), St->getAlignment());
24513 Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr, Increment);
24514 Chains.push_back(Ch);
24517 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Chains);
24520 // Turn load->store of MMX types into GPR load/stores. This avoids clobbering
24521 // the FP state in cases where an emms may be missing.
24522 // A preferable solution to the general problem is to figure out the right
24523 // places to insert EMMS. This qualifies as a quick hack.
24525 // Similarly, turn load->store of i64 into double load/stores in 32-bit mode.
24526 if (VT.getSizeInBits() != 64)
24529 const Function *F = DAG.getMachineFunction().getFunction();
24530 bool NoImplicitFloatOps = F->hasFnAttribute(Attribute::NoImplicitFloat);
24532 !Subtarget->useSoftFloat() && !NoImplicitFloatOps && Subtarget->hasSSE2();
24533 if ((VT.isVector() ||
24534 (VT == MVT::i64 && F64IsLegal && !Subtarget->is64Bit())) &&
24535 isa<LoadSDNode>(St->getValue()) &&
24536 !cast<LoadSDNode>(St->getValue())->isVolatile() &&
24537 St->getChain().hasOneUse() && !St->isVolatile()) {
24538 SDNode* LdVal = St->getValue().getNode();
24539 LoadSDNode *Ld = nullptr;
24540 int TokenFactorIndex = -1;
24541 SmallVector<SDValue, 8> Ops;
24542 SDNode* ChainVal = St->getChain().getNode();
24543 // Must be a store of a load. We currently handle two cases: the load
24544 // is a direct child, and it's under an intervening TokenFactor. It is
24545 // possible to dig deeper under nested TokenFactors.
24546 if (ChainVal == LdVal)
24547 Ld = cast<LoadSDNode>(St->getChain());
24548 else if (St->getValue().hasOneUse() &&
24549 ChainVal->getOpcode() == ISD::TokenFactor) {
24550 for (unsigned i = 0, e = ChainVal->getNumOperands(); i != e; ++i) {
24551 if (ChainVal->getOperand(i).getNode() == LdVal) {
24552 TokenFactorIndex = i;
24553 Ld = cast<LoadSDNode>(St->getValue());
24555 Ops.push_back(ChainVal->getOperand(i));
24559 if (!Ld || !ISD::isNormalLoad(Ld))
24562 // If this is not the MMX case, i.e. we are just turning i64 load/store
24563 // into f64 load/store, avoid the transformation if there are multiple
24564 // uses of the loaded value.
24565 if (!VT.isVector() && !Ld->hasNUsesOfValue(1, 0))
24570 // If we are a 64-bit capable x86, lower to a single movq load/store pair.
24571 // Otherwise, if it's legal to use f64 SSE instructions, use f64 load/store
24573 if (Subtarget->is64Bit() || F64IsLegal) {
24574 EVT LdVT = Subtarget->is64Bit() ? MVT::i64 : MVT::f64;
24575 SDValue NewLd = DAG.getLoad(LdVT, LdDL, Ld->getChain(), Ld->getBasePtr(),
24576 Ld->getPointerInfo(), Ld->isVolatile(),
24577 Ld->isNonTemporal(), Ld->isInvariant(),
24578 Ld->getAlignment());
24579 SDValue NewChain = NewLd.getValue(1);
24580 if (TokenFactorIndex != -1) {
24581 Ops.push_back(NewChain);
24582 NewChain = DAG.getNode(ISD::TokenFactor, LdDL, MVT::Other, Ops);
24584 return DAG.getStore(NewChain, StDL, NewLd, St->getBasePtr(),
24585 St->getPointerInfo(),
24586 St->isVolatile(), St->isNonTemporal(),
24587 St->getAlignment());
24590 // Otherwise, lower to two pairs of 32-bit loads / stores.
24591 SDValue LoAddr = Ld->getBasePtr();
24592 SDValue HiAddr = DAG.getNode(ISD::ADD, LdDL, MVT::i32, LoAddr,
24593 DAG.getConstant(4, LdDL, MVT::i32));
24595 SDValue LoLd = DAG.getLoad(MVT::i32, LdDL, Ld->getChain(), LoAddr,
24596 Ld->getPointerInfo(),
24597 Ld->isVolatile(), Ld->isNonTemporal(),
24598 Ld->isInvariant(), Ld->getAlignment());
24599 SDValue HiLd = DAG.getLoad(MVT::i32, LdDL, Ld->getChain(), HiAddr,
24600 Ld->getPointerInfo().getWithOffset(4),
24601 Ld->isVolatile(), Ld->isNonTemporal(),
24603 MinAlign(Ld->getAlignment(), 4));
24605 SDValue NewChain = LoLd.getValue(1);
24606 if (TokenFactorIndex != -1) {
24607 Ops.push_back(LoLd);
24608 Ops.push_back(HiLd);
24609 NewChain = DAG.getNode(ISD::TokenFactor, LdDL, MVT::Other, Ops);
24612 LoAddr = St->getBasePtr();
24613 HiAddr = DAG.getNode(ISD::ADD, StDL, MVT::i32, LoAddr,
24614 DAG.getConstant(4, StDL, MVT::i32));
24616 SDValue LoSt = DAG.getStore(NewChain, StDL, LoLd, LoAddr,
24617 St->getPointerInfo(),
24618 St->isVolatile(), St->isNonTemporal(),
24619 St->getAlignment());
24620 SDValue HiSt = DAG.getStore(NewChain, StDL, HiLd, HiAddr,
24621 St->getPointerInfo().getWithOffset(4),
24623 St->isNonTemporal(),
24624 MinAlign(St->getAlignment(), 4));
24625 return DAG.getNode(ISD::TokenFactor, StDL, MVT::Other, LoSt, HiSt);
24628 // This is similar to the above case, but here we handle a scalar 64-bit
24629 // integer store that is extracted from a vector on a 32-bit target.
24630 // If we have SSE2, then we can treat it like a floating-point double
24631 // to get past legalization. The execution dependencies fixup pass will
24632 // choose the optimal machine instruction for the store if this really is
24633 // an integer or v2f32 rather than an f64.
24634 if (VT == MVT::i64 && F64IsLegal && !Subtarget->is64Bit() &&
24635 St->getOperand(1).getOpcode() == ISD::EXTRACT_VECTOR_ELT) {
24636 SDValue OldExtract = St->getOperand(1);
24637 SDValue ExtOp0 = OldExtract.getOperand(0);
24638 unsigned VecSize = ExtOp0.getValueSizeInBits();
24639 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), MVT::f64, VecSize / 64);
24640 SDValue BitCast = DAG.getBitcast(VecVT, ExtOp0);
24641 SDValue NewExtract = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
24642 BitCast, OldExtract.getOperand(1));
24643 return DAG.getStore(St->getChain(), dl, NewExtract, St->getBasePtr(),
24644 St->getPointerInfo(), St->isVolatile(),
24645 St->isNonTemporal(), St->getAlignment());
24651 /// Return 'true' if this vector operation is "horizontal"
24652 /// and return the operands for the horizontal operation in LHS and RHS. A
24653 /// horizontal operation performs the binary operation on successive elements
24654 /// of its first operand, then on successive elements of its second operand,
24655 /// returning the resulting values in a vector. For example, if
24656 /// A = < float a0, float a1, float a2, float a3 >
24658 /// B = < float b0, float b1, float b2, float b3 >
24659 /// then the result of doing a horizontal operation on A and B is
24660 /// A horizontal-op B = < a0 op a1, a2 op a3, b0 op b1, b2 op b3 >.
24661 /// In short, LHS and RHS are inspected to see if LHS op RHS is of the form
24662 /// A horizontal-op B, for some already available A and B, and if so then LHS is
24663 /// set to A, RHS to B, and the routine returns 'true'.
24664 /// Note that the binary operation should have the property that if one of the
24665 /// operands is UNDEF then the result is UNDEF.
24666 static bool isHorizontalBinOp(SDValue &LHS, SDValue &RHS, bool IsCommutative) {
24667 // Look for the following pattern: if
24668 // A = < float a0, float a1, float a2, float a3 >
24669 // B = < float b0, float b1, float b2, float b3 >
24671 // LHS = VECTOR_SHUFFLE A, B, <0, 2, 4, 6>
24672 // RHS = VECTOR_SHUFFLE A, B, <1, 3, 5, 7>
24673 // then LHS op RHS = < a0 op a1, a2 op a3, b0 op b1, b2 op b3 >
24674 // which is A horizontal-op B.
24676 // At least one of the operands should be a vector shuffle.
24677 if (LHS.getOpcode() != ISD::VECTOR_SHUFFLE &&
24678 RHS.getOpcode() != ISD::VECTOR_SHUFFLE)
24681 MVT VT = LHS.getSimpleValueType();
24683 assert((VT.is128BitVector() || VT.is256BitVector()) &&
24684 "Unsupported vector type for horizontal add/sub");
24686 // Handle 128 and 256-bit vector lengths. AVX defines horizontal add/sub to
24687 // operate independently on 128-bit lanes.
24688 unsigned NumElts = VT.getVectorNumElements();
24689 unsigned NumLanes = VT.getSizeInBits()/128;
24690 unsigned NumLaneElts = NumElts / NumLanes;
24691 assert((NumLaneElts % 2 == 0) &&
24692 "Vector type should have an even number of elements in each lane");
24693 unsigned HalfLaneElts = NumLaneElts/2;
24695 // View LHS in the form
24696 // LHS = VECTOR_SHUFFLE A, B, LMask
24697 // If LHS is not a shuffle then pretend it is the shuffle
24698 // LHS = VECTOR_SHUFFLE LHS, undef, <0, 1, ..., N-1>
24699 // NOTE: in what follows a default initialized SDValue represents an UNDEF of
24702 SmallVector<int, 16> LMask(NumElts);
24703 if (LHS.getOpcode() == ISD::VECTOR_SHUFFLE) {
24704 if (LHS.getOperand(0).getOpcode() != ISD::UNDEF)
24705 A = LHS.getOperand(0);
24706 if (LHS.getOperand(1).getOpcode() != ISD::UNDEF)
24707 B = LHS.getOperand(1);
24708 ArrayRef<int> Mask = cast<ShuffleVectorSDNode>(LHS.getNode())->getMask();
24709 std::copy(Mask.begin(), Mask.end(), LMask.begin());
24711 if (LHS.getOpcode() != ISD::UNDEF)
24713 for (unsigned i = 0; i != NumElts; ++i)
24717 // Likewise, view RHS in the form
24718 // RHS = VECTOR_SHUFFLE C, D, RMask
24720 SmallVector<int, 16> RMask(NumElts);
24721 if (RHS.getOpcode() == ISD::VECTOR_SHUFFLE) {
24722 if (RHS.getOperand(0).getOpcode() != ISD::UNDEF)
24723 C = RHS.getOperand(0);
24724 if (RHS.getOperand(1).getOpcode() != ISD::UNDEF)
24725 D = RHS.getOperand(1);
24726 ArrayRef<int> Mask = cast<ShuffleVectorSDNode>(RHS.getNode())->getMask();
24727 std::copy(Mask.begin(), Mask.end(), RMask.begin());
24729 if (RHS.getOpcode() != ISD::UNDEF)
24731 for (unsigned i = 0; i != NumElts; ++i)
24735 // Check that the shuffles are both shuffling the same vectors.
24736 if (!(A == C && B == D) && !(A == D && B == C))
24739 // If everything is UNDEF then bail out: it would be better to fold to UNDEF.
24740 if (!A.getNode() && !B.getNode())
24743 // If A and B occur in reverse order in RHS, then "swap" them (which means
24744 // rewriting the mask).
24746 ShuffleVectorSDNode::commuteMask(RMask);
24748 // At this point LHS and RHS are equivalent to
24749 // LHS = VECTOR_SHUFFLE A, B, LMask
24750 // RHS = VECTOR_SHUFFLE A, B, RMask
24751 // Check that the masks correspond to performing a horizontal operation.
24752 for (unsigned l = 0; l != NumElts; l += NumLaneElts) {
24753 for (unsigned i = 0; i != NumLaneElts; ++i) {
24754 int LIdx = LMask[i+l], RIdx = RMask[i+l];
24756 // Ignore any UNDEF components.
24757 if (LIdx < 0 || RIdx < 0 ||
24758 (!A.getNode() && (LIdx < (int)NumElts || RIdx < (int)NumElts)) ||
24759 (!B.getNode() && (LIdx >= (int)NumElts || RIdx >= (int)NumElts)))
24762 // Check that successive elements are being operated on. If not, this is
24763 // not a horizontal operation.
24764 unsigned Src = (i/HalfLaneElts); // each lane is split between srcs
24765 int Index = 2*(i%HalfLaneElts) + NumElts*Src + l;
24766 if (!(LIdx == Index && RIdx == Index + 1) &&
24767 !(IsCommutative && LIdx == Index + 1 && RIdx == Index))
24772 LHS = A.getNode() ? A : B; // If A is 'UNDEF', use B for it.
24773 RHS = B.getNode() ? B : A; // If B is 'UNDEF', use A for it.
24777 /// Do target-specific dag combines on floating point adds.
24778 static SDValue PerformFADDCombine(SDNode *N, SelectionDAG &DAG,
24779 const X86Subtarget *Subtarget) {
24780 EVT VT = N->getValueType(0);
24781 SDValue LHS = N->getOperand(0);
24782 SDValue RHS = N->getOperand(1);
24784 // Try to synthesize horizontal adds from adds of shuffles.
24785 if (((Subtarget->hasSSE3() && (VT == MVT::v4f32 || VT == MVT::v2f64)) ||
24786 (Subtarget->hasFp256() && (VT == MVT::v8f32 || VT == MVT::v4f64))) &&
24787 isHorizontalBinOp(LHS, RHS, true))
24788 return DAG.getNode(X86ISD::FHADD, SDLoc(N), VT, LHS, RHS);
24792 /// Do target-specific dag combines on floating point subs.
24793 static SDValue PerformFSUBCombine(SDNode *N, SelectionDAG &DAG,
24794 const X86Subtarget *Subtarget) {
24795 EVT VT = N->getValueType(0);
24796 SDValue LHS = N->getOperand(0);
24797 SDValue RHS = N->getOperand(1);
24799 // Try to synthesize horizontal subs from subs of shuffles.
24800 if (((Subtarget->hasSSE3() && (VT == MVT::v4f32 || VT == MVT::v2f64)) ||
24801 (Subtarget->hasFp256() && (VT == MVT::v8f32 || VT == MVT::v4f64))) &&
24802 isHorizontalBinOp(LHS, RHS, false))
24803 return DAG.getNode(X86ISD::FHSUB, SDLoc(N), VT, LHS, RHS);
24807 /// Do target-specific dag combines on X86ISD::FOR and X86ISD::FXOR nodes.
24808 static SDValue PerformFORCombine(SDNode *N, SelectionDAG &DAG) {
24809 assert(N->getOpcode() == X86ISD::FOR || N->getOpcode() == X86ISD::FXOR);
24811 // F[X]OR(0.0, x) -> x
24812 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
24813 if (C->getValueAPF().isPosZero())
24814 return N->getOperand(1);
24816 // F[X]OR(x, 0.0) -> x
24817 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
24818 if (C->getValueAPF().isPosZero())
24819 return N->getOperand(0);
24823 /// Do target-specific dag combines on X86ISD::FMIN and X86ISD::FMAX nodes.
24824 static SDValue PerformFMinFMaxCombine(SDNode *N, SelectionDAG &DAG) {
24825 assert(N->getOpcode() == X86ISD::FMIN || N->getOpcode() == X86ISD::FMAX);
24827 // Only perform optimizations if UnsafeMath is used.
24828 if (!DAG.getTarget().Options.UnsafeFPMath)
24831 // If we run in unsafe-math mode, then convert the FMAX and FMIN nodes
24832 // into FMINC and FMAXC, which are Commutative operations.
24833 unsigned NewOp = 0;
24834 switch (N->getOpcode()) {
24835 default: llvm_unreachable("unknown opcode");
24836 case X86ISD::FMIN: NewOp = X86ISD::FMINC; break;
24837 case X86ISD::FMAX: NewOp = X86ISD::FMAXC; break;
24840 return DAG.getNode(NewOp, SDLoc(N), N->getValueType(0),
24841 N->getOperand(0), N->getOperand(1));
24844 /// Do target-specific dag combines on X86ISD::FAND nodes.
24845 static SDValue PerformFANDCombine(SDNode *N, SelectionDAG &DAG) {
24846 // FAND(0.0, x) -> 0.0
24847 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
24848 if (C->getValueAPF().isPosZero())
24849 return N->getOperand(0);
24851 // FAND(x, 0.0) -> 0.0
24852 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
24853 if (C->getValueAPF().isPosZero())
24854 return N->getOperand(1);
24859 /// Do target-specific dag combines on X86ISD::FANDN nodes
24860 static SDValue PerformFANDNCombine(SDNode *N, SelectionDAG &DAG) {
24861 // FANDN(0.0, x) -> x
24862 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
24863 if (C->getValueAPF().isPosZero())
24864 return N->getOperand(1);
24866 // FANDN(x, 0.0) -> 0.0
24867 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
24868 if (C->getValueAPF().isPosZero())
24869 return N->getOperand(1);
24874 static SDValue PerformBTCombine(SDNode *N,
24876 TargetLowering::DAGCombinerInfo &DCI) {
24877 // BT ignores high bits in the bit index operand.
24878 SDValue Op1 = N->getOperand(1);
24879 if (Op1.hasOneUse()) {
24880 unsigned BitWidth = Op1.getValueSizeInBits();
24881 APInt DemandedMask = APInt::getLowBitsSet(BitWidth, Log2_32(BitWidth));
24882 APInt KnownZero, KnownOne;
24883 TargetLowering::TargetLoweringOpt TLO(DAG, !DCI.isBeforeLegalize(),
24884 !DCI.isBeforeLegalizeOps());
24885 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
24886 if (TLO.ShrinkDemandedConstant(Op1, DemandedMask) ||
24887 TLI.SimplifyDemandedBits(Op1, DemandedMask, KnownZero, KnownOne, TLO))
24888 DCI.CommitTargetLoweringOpt(TLO);
24893 static SDValue PerformVZEXT_MOVLCombine(SDNode *N, SelectionDAG &DAG) {
24894 SDValue Op = N->getOperand(0);
24895 if (Op.getOpcode() == ISD::BITCAST)
24896 Op = Op.getOperand(0);
24897 EVT VT = N->getValueType(0), OpVT = Op.getValueType();
24898 if (Op.getOpcode() == X86ISD::VZEXT_LOAD &&
24899 VT.getVectorElementType().getSizeInBits() ==
24900 OpVT.getVectorElementType().getSizeInBits()) {
24901 return DAG.getNode(ISD::BITCAST, SDLoc(N), VT, Op);
24906 static SDValue PerformSIGN_EXTEND_INREGCombine(SDNode *N, SelectionDAG &DAG,
24907 const X86Subtarget *Subtarget) {
24908 EVT VT = N->getValueType(0);
24909 if (!VT.isVector())
24912 SDValue N0 = N->getOperand(0);
24913 SDValue N1 = N->getOperand(1);
24914 EVT ExtraVT = cast<VTSDNode>(N1)->getVT();
24917 // The SIGN_EXTEND_INREG to v4i64 is expensive operation on the
24918 // both SSE and AVX2 since there is no sign-extended shift right
24919 // operation on a vector with 64-bit elements.
24920 //(sext_in_reg (v4i64 anyext (v4i32 x )), ExtraVT) ->
24921 // (v4i64 sext (v4i32 sext_in_reg (v4i32 x , ExtraVT)))
24922 if (VT == MVT::v4i64 && (N0.getOpcode() == ISD::ANY_EXTEND ||
24923 N0.getOpcode() == ISD::SIGN_EXTEND)) {
24924 SDValue N00 = N0.getOperand(0);
24926 // EXTLOAD has a better solution on AVX2,
24927 // it may be replaced with X86ISD::VSEXT node.
24928 if (N00.getOpcode() == ISD::LOAD && Subtarget->hasInt256())
24929 if (!ISD::isNormalLoad(N00.getNode()))
24932 if (N00.getValueType() == MVT::v4i32 && ExtraVT.getSizeInBits() < 128) {
24933 SDValue Tmp = DAG.getNode(ISD::SIGN_EXTEND_INREG, dl, MVT::v4i32,
24935 return DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i64, Tmp);
24941 static SDValue PerformSExtCombine(SDNode *N, SelectionDAG &DAG,
24942 TargetLowering::DAGCombinerInfo &DCI,
24943 const X86Subtarget *Subtarget) {
24944 SDValue N0 = N->getOperand(0);
24945 EVT VT = N->getValueType(0);
24946 EVT SVT = VT.getScalarType();
24947 EVT InVT = N0.getValueType();
24948 EVT InSVT = InVT.getScalarType();
24951 // (i8,i32 sext (sdivrem (i8 x, i8 y)) ->
24952 // (i8,i32 (sdivrem_sext_hreg (i8 x, i8 y)
24953 // This exposes the sext to the sdivrem lowering, so that it directly extends
24954 // from AH (which we otherwise need to do contortions to access).
24955 if (N0.getOpcode() == ISD::SDIVREM && N0.getResNo() == 1 &&
24956 InVT == MVT::i8 && VT == MVT::i32) {
24957 SDVTList NodeTys = DAG.getVTList(MVT::i8, VT);
24958 SDValue R = DAG.getNode(X86ISD::SDIVREM8_SEXT_HREG, DL, NodeTys,
24959 N0.getOperand(0), N0.getOperand(1));
24960 DAG.ReplaceAllUsesOfValueWith(N0.getValue(0), R.getValue(0));
24961 return R.getValue(1);
24964 if (!DCI.isBeforeLegalizeOps()) {
24965 if (InVT == MVT::i1) {
24966 SDValue Zero = DAG.getConstant(0, DL, VT);
24968 DAG.getConstant(APInt::getAllOnesValue(VT.getSizeInBits()), DL, VT);
24969 return DAG.getNode(ISD::SELECT, DL, VT, N0, AllOnes, Zero);
24974 if (VT.isVector() && Subtarget->hasSSE2()) {
24975 auto ExtendVecSize = [&DAG](SDLoc DL, SDValue N, unsigned Size) {
24976 EVT InVT = N.getValueType();
24977 EVT OutVT = EVT::getVectorVT(*DAG.getContext(), InVT.getScalarType(),
24978 Size / InVT.getScalarSizeInBits());
24979 SmallVector<SDValue, 8> Opnds(Size / InVT.getSizeInBits(),
24980 DAG.getUNDEF(InVT));
24982 return DAG.getNode(ISD::CONCAT_VECTORS, DL, OutVT, Opnds);
24985 // If target-size is less than 128-bits, extend to a type that would extend
24986 // to 128 bits, extend that and extract the original target vector.
24987 if (VT.getSizeInBits() < 128 && !(128 % VT.getSizeInBits()) &&
24988 (SVT == MVT::i64 || SVT == MVT::i32 || SVT == MVT::i16) &&
24989 (InSVT == MVT::i32 || InSVT == MVT::i16 || InSVT == MVT::i8)) {
24990 unsigned Scale = 128 / VT.getSizeInBits();
24992 EVT::getVectorVT(*DAG.getContext(), SVT, 128 / SVT.getSizeInBits());
24993 SDValue Ex = ExtendVecSize(DL, N0, Scale * InVT.getSizeInBits());
24994 SDValue SExt = DAG.getNode(ISD::SIGN_EXTEND, DL, ExVT, Ex);
24995 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, VT, SExt,
24996 DAG.getIntPtrConstant(0, DL));
24999 // If target-size is 128-bits, then convert to ISD::SIGN_EXTEND_VECTOR_INREG
25000 // which ensures lowering to X86ISD::VSEXT (pmovsx*).
25001 if (VT.getSizeInBits() == 128 &&
25002 (SVT == MVT::i64 || SVT == MVT::i32 || SVT == MVT::i16) &&
25003 (InSVT == MVT::i32 || InSVT == MVT::i16 || InSVT == MVT::i8)) {
25004 SDValue ExOp = ExtendVecSize(DL, N0, 128);
25005 return DAG.getSignExtendVectorInReg(ExOp, DL, VT);
25008 // On pre-AVX2 targets, split into 128-bit nodes of
25009 // ISD::SIGN_EXTEND_VECTOR_INREG.
25010 if (!Subtarget->hasInt256() && !(VT.getSizeInBits() % 128) &&
25011 (SVT == MVT::i64 || SVT == MVT::i32 || SVT == MVT::i16) &&
25012 (InSVT == MVT::i32 || InSVT == MVT::i16 || InSVT == MVT::i8)) {
25013 unsigned NumVecs = VT.getSizeInBits() / 128;
25014 unsigned NumSubElts = 128 / SVT.getSizeInBits();
25015 EVT SubVT = EVT::getVectorVT(*DAG.getContext(), SVT, NumSubElts);
25016 EVT InSubVT = EVT::getVectorVT(*DAG.getContext(), InSVT, NumSubElts);
25018 SmallVector<SDValue, 8> Opnds;
25019 for (unsigned i = 0, Offset = 0; i != NumVecs;
25020 ++i, Offset += NumSubElts) {
25021 SDValue SrcVec = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, InSubVT, N0,
25022 DAG.getIntPtrConstant(Offset, DL));
25023 SrcVec = ExtendVecSize(DL, SrcVec, 128);
25024 SrcVec = DAG.getSignExtendVectorInReg(SrcVec, DL, SubVT);
25025 Opnds.push_back(SrcVec);
25027 return DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, Opnds);
25031 if (!Subtarget->hasFp256())
25034 if (VT.isVector() && VT.getSizeInBits() == 256)
25035 if (SDValue R = WidenMaskArithmetic(N, DAG, DCI, Subtarget))
25041 static SDValue PerformFMACombine(SDNode *N, SelectionDAG &DAG,
25042 const X86Subtarget* Subtarget) {
25044 EVT VT = N->getValueType(0);
25046 // Let legalize expand this if it isn't a legal type yet.
25047 if (!DAG.getTargetLoweringInfo().isTypeLegal(VT))
25050 EVT ScalarVT = VT.getScalarType();
25051 if ((ScalarVT != MVT::f32 && ScalarVT != MVT::f64) ||
25052 (!Subtarget->hasFMA() && !Subtarget->hasFMA4() &&
25053 !Subtarget->hasAVX512()))
25056 SDValue A = N->getOperand(0);
25057 SDValue B = N->getOperand(1);
25058 SDValue C = N->getOperand(2);
25060 bool NegA = (A.getOpcode() == ISD::FNEG);
25061 bool NegB = (B.getOpcode() == ISD::FNEG);
25062 bool NegC = (C.getOpcode() == ISD::FNEG);
25064 // Negative multiplication when NegA xor NegB
25065 bool NegMul = (NegA != NegB);
25067 A = A.getOperand(0);
25069 B = B.getOperand(0);
25071 C = C.getOperand(0);
25075 Opcode = (!NegC) ? X86ISD::FMADD : X86ISD::FMSUB;
25077 Opcode = (!NegC) ? X86ISD::FNMADD : X86ISD::FNMSUB;
25079 return DAG.getNode(Opcode, dl, VT, A, B, C);
25082 static SDValue PerformZExtCombine(SDNode *N, SelectionDAG &DAG,
25083 TargetLowering::DAGCombinerInfo &DCI,
25084 const X86Subtarget *Subtarget) {
25085 // (i32 zext (and (i8 x86isd::setcc_carry), 1)) ->
25086 // (and (i32 x86isd::setcc_carry), 1)
25087 // This eliminates the zext. This transformation is necessary because
25088 // ISD::SETCC is always legalized to i8.
25090 SDValue N0 = N->getOperand(0);
25091 EVT VT = N->getValueType(0);
25093 if (N0.getOpcode() == ISD::AND &&
25095 N0.getOperand(0).hasOneUse()) {
25096 SDValue N00 = N0.getOperand(0);
25097 if (N00.getOpcode() == X86ISD::SETCC_CARRY) {
25098 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
25099 if (!C || C->getZExtValue() != 1)
25101 return DAG.getNode(ISD::AND, dl, VT,
25102 DAG.getNode(X86ISD::SETCC_CARRY, dl, VT,
25103 N00.getOperand(0), N00.getOperand(1)),
25104 DAG.getConstant(1, dl, VT));
25108 if (N0.getOpcode() == ISD::TRUNCATE &&
25110 N0.getOperand(0).hasOneUse()) {
25111 SDValue N00 = N0.getOperand(0);
25112 if (N00.getOpcode() == X86ISD::SETCC_CARRY) {
25113 return DAG.getNode(ISD::AND, dl, VT,
25114 DAG.getNode(X86ISD::SETCC_CARRY, dl, VT,
25115 N00.getOperand(0), N00.getOperand(1)),
25116 DAG.getConstant(1, dl, VT));
25120 if (VT.is256BitVector())
25121 if (SDValue R = WidenMaskArithmetic(N, DAG, DCI, Subtarget))
25124 // (i8,i32 zext (udivrem (i8 x, i8 y)) ->
25125 // (i8,i32 (udivrem_zext_hreg (i8 x, i8 y)
25126 // This exposes the zext to the udivrem lowering, so that it directly extends
25127 // from AH (which we otherwise need to do contortions to access).
25128 if (N0.getOpcode() == ISD::UDIVREM &&
25129 N0.getResNo() == 1 && N0.getValueType() == MVT::i8 &&
25130 (VT == MVT::i32 || VT == MVT::i64)) {
25131 SDVTList NodeTys = DAG.getVTList(MVT::i8, VT);
25132 SDValue R = DAG.getNode(X86ISD::UDIVREM8_ZEXT_HREG, dl, NodeTys,
25133 N0.getOperand(0), N0.getOperand(1));
25134 DAG.ReplaceAllUsesOfValueWith(N0.getValue(0), R.getValue(0));
25135 return R.getValue(1);
25141 // Optimize x == -y --> x+y == 0
25142 // x != -y --> x+y != 0
25143 static SDValue PerformISDSETCCCombine(SDNode *N, SelectionDAG &DAG,
25144 const X86Subtarget* Subtarget) {
25145 ISD::CondCode CC = cast<CondCodeSDNode>(N->getOperand(2))->get();
25146 SDValue LHS = N->getOperand(0);
25147 SDValue RHS = N->getOperand(1);
25148 EVT VT = N->getValueType(0);
25151 if ((CC == ISD::SETNE || CC == ISD::SETEQ) && LHS.getOpcode() == ISD::SUB)
25152 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(LHS.getOperand(0)))
25153 if (C->getAPIntValue() == 0 && LHS.hasOneUse()) {
25154 SDValue addV = DAG.getNode(ISD::ADD, DL, LHS.getValueType(), RHS,
25155 LHS.getOperand(1));
25156 return DAG.getSetCC(DL, N->getValueType(0), addV,
25157 DAG.getConstant(0, DL, addV.getValueType()), CC);
25159 if ((CC == ISD::SETNE || CC == ISD::SETEQ) && RHS.getOpcode() == ISD::SUB)
25160 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(RHS.getOperand(0)))
25161 if (C->getAPIntValue() == 0 && RHS.hasOneUse()) {
25162 SDValue addV = DAG.getNode(ISD::ADD, DL, RHS.getValueType(), LHS,
25163 RHS.getOperand(1));
25164 return DAG.getSetCC(DL, N->getValueType(0), addV,
25165 DAG.getConstant(0, DL, addV.getValueType()), CC);
25168 if (VT.getScalarType() == MVT::i1 &&
25169 (CC == ISD::SETNE || CC == ISD::SETEQ || ISD::isSignedIntSetCC(CC))) {
25171 (LHS.getOpcode() == ISD::SIGN_EXTEND) &&
25172 (LHS.getOperand(0).getValueType().getScalarType() == MVT::i1);
25173 bool IsVZero1 = ISD::isBuildVectorAllZeros(RHS.getNode());
25175 if (!IsSEXT0 || !IsVZero1) {
25176 // Swap the operands and update the condition code.
25177 std::swap(LHS, RHS);
25178 CC = ISD::getSetCCSwappedOperands(CC);
25180 IsSEXT0 = (LHS.getOpcode() == ISD::SIGN_EXTEND) &&
25181 (LHS.getOperand(0).getValueType().getScalarType() == MVT::i1);
25182 IsVZero1 = ISD::isBuildVectorAllZeros(RHS.getNode());
25185 if (IsSEXT0 && IsVZero1) {
25186 assert(VT == LHS.getOperand(0).getValueType() &&
25187 "Uexpected operand type");
25188 if (CC == ISD::SETGT)
25189 return DAG.getConstant(0, DL, VT);
25190 if (CC == ISD::SETLE)
25191 return DAG.getConstant(1, DL, VT);
25192 if (CC == ISD::SETEQ || CC == ISD::SETGE)
25193 return DAG.getNOT(DL, LHS.getOperand(0), VT);
25195 assert((CC == ISD::SETNE || CC == ISD::SETLT) &&
25196 "Unexpected condition code!");
25197 return LHS.getOperand(0);
25204 static SDValue NarrowVectorLoadToElement(LoadSDNode *Load, unsigned Index,
25205 SelectionDAG &DAG) {
25207 MVT VT = Load->getSimpleValueType(0);
25208 MVT EVT = VT.getVectorElementType();
25209 SDValue Addr = Load->getOperand(1);
25210 SDValue NewAddr = DAG.getNode(
25211 ISD::ADD, dl, Addr.getSimpleValueType(), Addr,
25212 DAG.getConstant(Index * EVT.getStoreSize(), dl,
25213 Addr.getSimpleValueType()));
25216 DAG.getLoad(EVT, dl, Load->getChain(), NewAddr,
25217 DAG.getMachineFunction().getMachineMemOperand(
25218 Load->getMemOperand(), 0, EVT.getStoreSize()));
25222 static SDValue PerformINSERTPSCombine(SDNode *N, SelectionDAG &DAG,
25223 const X86Subtarget *Subtarget) {
25225 MVT VT = N->getOperand(1)->getSimpleValueType(0);
25226 assert((VT == MVT::v4f32 || VT == MVT::v4i32) &&
25227 "X86insertps is only defined for v4x32");
25229 SDValue Ld = N->getOperand(1);
25230 if (MayFoldLoad(Ld)) {
25231 // Extract the countS bits from the immediate so we can get the proper
25232 // address when narrowing the vector load to a specific element.
25233 // When the second source op is a memory address, insertps doesn't use
25234 // countS and just gets an f32 from that address.
25235 unsigned DestIndex =
25236 cast<ConstantSDNode>(N->getOperand(2))->getZExtValue() >> 6;
25238 Ld = NarrowVectorLoadToElement(cast<LoadSDNode>(Ld), DestIndex, DAG);
25240 // Create this as a scalar to vector to match the instruction pattern.
25241 SDValue LoadScalarToVector = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Ld);
25242 // countS bits are ignored when loading from memory on insertps, which
25243 // means we don't need to explicitly set them to 0.
25244 return DAG.getNode(X86ISD::INSERTPS, dl, VT, N->getOperand(0),
25245 LoadScalarToVector, N->getOperand(2));
25250 static SDValue PerformBLENDICombine(SDNode *N, SelectionDAG &DAG) {
25251 SDValue V0 = N->getOperand(0);
25252 SDValue V1 = N->getOperand(1);
25254 EVT VT = N->getValueType(0);
25256 // Canonicalize a v2f64 blend with a mask of 2 by swapping the vector
25257 // operands and changing the mask to 1. This saves us a bunch of
25258 // pattern-matching possibilities related to scalar math ops in SSE/AVX.
25259 // x86InstrInfo knows how to commute this back after instruction selection
25260 // if it would help register allocation.
25262 // TODO: If optimizing for size or a processor that doesn't suffer from
25263 // partial register update stalls, this should be transformed into a MOVSD
25264 // instruction because a MOVSD is 1-2 bytes smaller than a BLENDPD.
25266 if (VT == MVT::v2f64)
25267 if (auto *Mask = dyn_cast<ConstantSDNode>(N->getOperand(2)))
25268 if (Mask->getZExtValue() == 2 && !isShuffleFoldableLoad(V0)) {
25269 SDValue NewMask = DAG.getConstant(1, DL, MVT::i8);
25270 return DAG.getNode(X86ISD::BLENDI, DL, VT, V1, V0, NewMask);
25276 // Helper function of PerformSETCCCombine. It is to materialize "setb reg"
25277 // as "sbb reg,reg", since it can be extended without zext and produces
25278 // an all-ones bit which is more useful than 0/1 in some cases.
25279 static SDValue MaterializeSETB(SDLoc DL, SDValue EFLAGS, SelectionDAG &DAG,
25282 return DAG.getNode(ISD::AND, DL, VT,
25283 DAG.getNode(X86ISD::SETCC_CARRY, DL, MVT::i8,
25284 DAG.getConstant(X86::COND_B, DL, MVT::i8),
25286 DAG.getConstant(1, DL, VT));
25287 assert (VT == MVT::i1 && "Unexpected type for SECCC node");
25288 return DAG.getNode(ISD::TRUNCATE, DL, MVT::i1,
25289 DAG.getNode(X86ISD::SETCC_CARRY, DL, MVT::i8,
25290 DAG.getConstant(X86::COND_B, DL, MVT::i8),
25294 // Optimize RES = X86ISD::SETCC CONDCODE, EFLAG_INPUT
25295 static SDValue PerformSETCCCombine(SDNode *N, SelectionDAG &DAG,
25296 TargetLowering::DAGCombinerInfo &DCI,
25297 const X86Subtarget *Subtarget) {
25299 X86::CondCode CC = X86::CondCode(N->getConstantOperandVal(0));
25300 SDValue EFLAGS = N->getOperand(1);
25302 if (CC == X86::COND_A) {
25303 // Try to convert COND_A into COND_B in an attempt to facilitate
25304 // materializing "setb reg".
25306 // Do not flip "e > c", where "c" is a constant, because Cmp instruction
25307 // cannot take an immediate as its first operand.
25309 if (EFLAGS.getOpcode() == X86ISD::SUB && EFLAGS.hasOneUse() &&
25310 EFLAGS.getValueType().isInteger() &&
25311 !isa<ConstantSDNode>(EFLAGS.getOperand(1))) {
25312 SDValue NewSub = DAG.getNode(X86ISD::SUB, SDLoc(EFLAGS),
25313 EFLAGS.getNode()->getVTList(),
25314 EFLAGS.getOperand(1), EFLAGS.getOperand(0));
25315 SDValue NewEFLAGS = SDValue(NewSub.getNode(), EFLAGS.getResNo());
25316 return MaterializeSETB(DL, NewEFLAGS, DAG, N->getSimpleValueType(0));
25320 // Materialize "setb reg" as "sbb reg,reg", since it can be extended without
25321 // a zext and produces an all-ones bit which is more useful than 0/1 in some
25323 if (CC == X86::COND_B)
25324 return MaterializeSETB(DL, EFLAGS, DAG, N->getSimpleValueType(0));
25326 if (SDValue Flags = checkBoolTestSetCCCombine(EFLAGS, CC)) {
25327 SDValue Cond = DAG.getConstant(CC, DL, MVT::i8);
25328 return DAG.getNode(X86ISD::SETCC, DL, N->getVTList(), Cond, Flags);
25334 // Optimize branch condition evaluation.
25336 static SDValue PerformBrCondCombine(SDNode *N, SelectionDAG &DAG,
25337 TargetLowering::DAGCombinerInfo &DCI,
25338 const X86Subtarget *Subtarget) {
25340 SDValue Chain = N->getOperand(0);
25341 SDValue Dest = N->getOperand(1);
25342 SDValue EFLAGS = N->getOperand(3);
25343 X86::CondCode CC = X86::CondCode(N->getConstantOperandVal(2));
25345 if (SDValue Flags = checkBoolTestSetCCCombine(EFLAGS, CC)) {
25346 SDValue Cond = DAG.getConstant(CC, DL, MVT::i8);
25347 return DAG.getNode(X86ISD::BRCOND, DL, N->getVTList(), Chain, Dest, Cond,
25354 static SDValue performVectorCompareAndMaskUnaryOpCombine(SDNode *N,
25355 SelectionDAG &DAG) {
25356 // Take advantage of vector comparisons producing 0 or -1 in each lane to
25357 // optimize away operation when it's from a constant.
25359 // The general transformation is:
25360 // UNARYOP(AND(VECTOR_CMP(x,y), constant)) -->
25361 // AND(VECTOR_CMP(x,y), constant2)
25362 // constant2 = UNARYOP(constant)
25364 // Early exit if this isn't a vector operation, the operand of the
25365 // unary operation isn't a bitwise AND, or if the sizes of the operations
25366 // aren't the same.
25367 EVT VT = N->getValueType(0);
25368 if (!VT.isVector() || N->getOperand(0)->getOpcode() != ISD::AND ||
25369 N->getOperand(0)->getOperand(0)->getOpcode() != ISD::SETCC ||
25370 VT.getSizeInBits() != N->getOperand(0)->getValueType(0).getSizeInBits())
25373 // Now check that the other operand of the AND is a constant. We could
25374 // make the transformation for non-constant splats as well, but it's unclear
25375 // that would be a benefit as it would not eliminate any operations, just
25376 // perform one more step in scalar code before moving to the vector unit.
25377 if (BuildVectorSDNode *BV =
25378 dyn_cast<BuildVectorSDNode>(N->getOperand(0)->getOperand(1))) {
25379 // Bail out if the vector isn't a constant.
25380 if (!BV->isConstant())
25383 // Everything checks out. Build up the new and improved node.
25385 EVT IntVT = BV->getValueType(0);
25386 // Create a new constant of the appropriate type for the transformed
25388 SDValue SourceConst = DAG.getNode(N->getOpcode(), DL, VT, SDValue(BV, 0));
25389 // The AND node needs bitcasts to/from an integer vector type around it.
25390 SDValue MaskConst = DAG.getBitcast(IntVT, SourceConst);
25391 SDValue NewAnd = DAG.getNode(ISD::AND, DL, IntVT,
25392 N->getOperand(0)->getOperand(0), MaskConst);
25393 SDValue Res = DAG.getBitcast(VT, NewAnd);
25400 static SDValue PerformUINT_TO_FPCombine(SDNode *N, SelectionDAG &DAG,
25401 const X86Subtarget *Subtarget) {
25402 SDValue Op0 = N->getOperand(0);
25403 EVT VT = N->getValueType(0);
25404 EVT InVT = Op0.getValueType();
25405 EVT InSVT = InVT.getScalarType();
25406 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
25408 // UINT_TO_FP(vXi8) -> SINT_TO_FP(ZEXT(vXi8 to vXi32))
25409 // UINT_TO_FP(vXi16) -> SINT_TO_FP(ZEXT(vXi16 to vXi32))
25410 if (InVT.isVector() && (InSVT == MVT::i8 || InSVT == MVT::i16)) {
25412 EVT DstVT = EVT::getVectorVT(*DAG.getContext(), MVT::i32,
25413 InVT.getVectorNumElements());
25414 SDValue P = DAG.getNode(ISD::ZERO_EXTEND, dl, DstVT, Op0);
25416 if (TLI.isOperationLegal(ISD::UINT_TO_FP, DstVT))
25417 return DAG.getNode(ISD::UINT_TO_FP, dl, VT, P);
25419 return DAG.getNode(ISD::SINT_TO_FP, dl, VT, P);
25425 static SDValue PerformSINT_TO_FPCombine(SDNode *N, SelectionDAG &DAG,
25426 const X86Subtarget *Subtarget) {
25427 // First try to optimize away the conversion entirely when it's
25428 // conditionally from a constant. Vectors only.
25429 if (SDValue Res = performVectorCompareAndMaskUnaryOpCombine(N, DAG))
25432 // Now move on to more general possibilities.
25433 SDValue Op0 = N->getOperand(0);
25434 EVT VT = N->getValueType(0);
25435 EVT InVT = Op0.getValueType();
25436 EVT InSVT = InVT.getScalarType();
25438 // SINT_TO_FP(vXi8) -> SINT_TO_FP(SEXT(vXi8 to vXi32))
25439 // SINT_TO_FP(vXi16) -> SINT_TO_FP(SEXT(vXi16 to vXi32))
25440 if (InVT.isVector() && (InSVT == MVT::i8 || InSVT == MVT::i16)) {
25442 EVT DstVT = EVT::getVectorVT(*DAG.getContext(), MVT::i32,
25443 InVT.getVectorNumElements());
25444 SDValue P = DAG.getNode(ISD::SIGN_EXTEND, dl, DstVT, Op0);
25445 return DAG.getNode(ISD::SINT_TO_FP, dl, VT, P);
25448 // Transform (SINT_TO_FP (i64 ...)) into an x87 operation if we have
25449 // a 32-bit target where SSE doesn't support i64->FP operations.
25450 if (Op0.getOpcode() == ISD::LOAD) {
25451 LoadSDNode *Ld = cast<LoadSDNode>(Op0.getNode());
25452 EVT LdVT = Ld->getValueType(0);
25454 // This transformation is not supported if the result type is f16
25455 if (VT == MVT::f16)
25458 if (!Ld->isVolatile() && !VT.isVector() &&
25459 ISD::isNON_EXTLoad(Op0.getNode()) && Op0.hasOneUse() &&
25460 !Subtarget->is64Bit() && LdVT == MVT::i64) {
25461 SDValue FILDChain = Subtarget->getTargetLowering()->BuildFILD(
25462 SDValue(N, 0), LdVT, Ld->getChain(), Op0, DAG);
25463 DAG.ReplaceAllUsesOfValueWith(Op0.getValue(1), FILDChain.getValue(1));
25470 // Optimize RES, EFLAGS = X86ISD::ADC LHS, RHS, EFLAGS
25471 static SDValue PerformADCCombine(SDNode *N, SelectionDAG &DAG,
25472 X86TargetLowering::DAGCombinerInfo &DCI) {
25473 // If the LHS and RHS of the ADC node are zero, then it can't overflow and
25474 // the result is either zero or one (depending on the input carry bit).
25475 // Strength reduce this down to a "set on carry" aka SETCC_CARRY&1.
25476 if (X86::isZeroNode(N->getOperand(0)) &&
25477 X86::isZeroNode(N->getOperand(1)) &&
25478 // We don't have a good way to replace an EFLAGS use, so only do this when
25480 SDValue(N, 1).use_empty()) {
25482 EVT VT = N->getValueType(0);
25483 SDValue CarryOut = DAG.getConstant(0, DL, N->getValueType(1));
25484 SDValue Res1 = DAG.getNode(ISD::AND, DL, VT,
25485 DAG.getNode(X86ISD::SETCC_CARRY, DL, VT,
25486 DAG.getConstant(X86::COND_B, DL,
25489 DAG.getConstant(1, DL, VT));
25490 return DCI.CombineTo(N, Res1, CarryOut);
25496 // fold (add Y, (sete X, 0)) -> adc 0, Y
25497 // (add Y, (setne X, 0)) -> sbb -1, Y
25498 // (sub (sete X, 0), Y) -> sbb 0, Y
25499 // (sub (setne X, 0), Y) -> adc -1, Y
25500 static SDValue OptimizeConditionalInDecrement(SDNode *N, SelectionDAG &DAG) {
25503 // Look through ZExts.
25504 SDValue Ext = N->getOperand(N->getOpcode() == ISD::SUB ? 1 : 0);
25505 if (Ext.getOpcode() != ISD::ZERO_EXTEND || !Ext.hasOneUse())
25508 SDValue SetCC = Ext.getOperand(0);
25509 if (SetCC.getOpcode() != X86ISD::SETCC || !SetCC.hasOneUse())
25512 X86::CondCode CC = (X86::CondCode)SetCC.getConstantOperandVal(0);
25513 if (CC != X86::COND_E && CC != X86::COND_NE)
25516 SDValue Cmp = SetCC.getOperand(1);
25517 if (Cmp.getOpcode() != X86ISD::CMP || !Cmp.hasOneUse() ||
25518 !X86::isZeroNode(Cmp.getOperand(1)) ||
25519 !Cmp.getOperand(0).getValueType().isInteger())
25522 SDValue CmpOp0 = Cmp.getOperand(0);
25523 SDValue NewCmp = DAG.getNode(X86ISD::CMP, DL, MVT::i32, CmpOp0,
25524 DAG.getConstant(1, DL, CmpOp0.getValueType()));
25526 SDValue OtherVal = N->getOperand(N->getOpcode() == ISD::SUB ? 0 : 1);
25527 if (CC == X86::COND_NE)
25528 return DAG.getNode(N->getOpcode() == ISD::SUB ? X86ISD::ADC : X86ISD::SBB,
25529 DL, OtherVal.getValueType(), OtherVal,
25530 DAG.getConstant(-1ULL, DL, OtherVal.getValueType()),
25532 return DAG.getNode(N->getOpcode() == ISD::SUB ? X86ISD::SBB : X86ISD::ADC,
25533 DL, OtherVal.getValueType(), OtherVal,
25534 DAG.getConstant(0, DL, OtherVal.getValueType()), NewCmp);
25537 /// PerformADDCombine - Do target-specific dag combines on integer adds.
25538 static SDValue PerformAddCombine(SDNode *N, SelectionDAG &DAG,
25539 const X86Subtarget *Subtarget) {
25540 EVT VT = N->getValueType(0);
25541 SDValue Op0 = N->getOperand(0);
25542 SDValue Op1 = N->getOperand(1);
25544 // Try to synthesize horizontal adds from adds of shuffles.
25545 if (((Subtarget->hasSSSE3() && (VT == MVT::v8i16 || VT == MVT::v4i32)) ||
25546 (Subtarget->hasInt256() && (VT == MVT::v16i16 || VT == MVT::v8i32))) &&
25547 isHorizontalBinOp(Op0, Op1, true))
25548 return DAG.getNode(X86ISD::HADD, SDLoc(N), VT, Op0, Op1);
25550 return OptimizeConditionalInDecrement(N, DAG);
25553 static SDValue PerformSubCombine(SDNode *N, SelectionDAG &DAG,
25554 const X86Subtarget *Subtarget) {
25555 SDValue Op0 = N->getOperand(0);
25556 SDValue Op1 = N->getOperand(1);
25558 // X86 can't encode an immediate LHS of a sub. See if we can push the
25559 // negation into a preceding instruction.
25560 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op0)) {
25561 // If the RHS of the sub is a XOR with one use and a constant, invert the
25562 // immediate. Then add one to the LHS of the sub so we can turn
25563 // X-Y -> X+~Y+1, saving one register.
25564 if (Op1->hasOneUse() && Op1.getOpcode() == ISD::XOR &&
25565 isa<ConstantSDNode>(Op1.getOperand(1))) {
25566 APInt XorC = cast<ConstantSDNode>(Op1.getOperand(1))->getAPIntValue();
25567 EVT VT = Op0.getValueType();
25568 SDValue NewXor = DAG.getNode(ISD::XOR, SDLoc(Op1), VT,
25570 DAG.getConstant(~XorC, SDLoc(Op1), VT));
25571 return DAG.getNode(ISD::ADD, SDLoc(N), VT, NewXor,
25572 DAG.getConstant(C->getAPIntValue() + 1, SDLoc(N), VT));
25576 // Try to synthesize horizontal adds from adds of shuffles.
25577 EVT VT = N->getValueType(0);
25578 if (((Subtarget->hasSSSE3() && (VT == MVT::v8i16 || VT == MVT::v4i32)) ||
25579 (Subtarget->hasInt256() && (VT == MVT::v16i16 || VT == MVT::v8i32))) &&
25580 isHorizontalBinOp(Op0, Op1, true))
25581 return DAG.getNode(X86ISD::HSUB, SDLoc(N), VT, Op0, Op1);
25583 return OptimizeConditionalInDecrement(N, DAG);
25586 /// performVZEXTCombine - Performs build vector combines
25587 static SDValue performVZEXTCombine(SDNode *N, SelectionDAG &DAG,
25588 TargetLowering::DAGCombinerInfo &DCI,
25589 const X86Subtarget *Subtarget) {
25591 MVT VT = N->getSimpleValueType(0);
25592 SDValue Op = N->getOperand(0);
25593 MVT OpVT = Op.getSimpleValueType();
25594 MVT OpEltVT = OpVT.getVectorElementType();
25595 unsigned InputBits = OpEltVT.getSizeInBits() * VT.getVectorNumElements();
25597 // (vzext (bitcast (vzext (x)) -> (vzext x)
25599 while (V.getOpcode() == ISD::BITCAST)
25600 V = V.getOperand(0);
25602 if (V != Op && V.getOpcode() == X86ISD::VZEXT) {
25603 MVT InnerVT = V.getSimpleValueType();
25604 MVT InnerEltVT = InnerVT.getVectorElementType();
25606 // If the element sizes match exactly, we can just do one larger vzext. This
25607 // is always an exact type match as vzext operates on integer types.
25608 if (OpEltVT == InnerEltVT) {
25609 assert(OpVT == InnerVT && "Types must match for vzext!");
25610 return DAG.getNode(X86ISD::VZEXT, DL, VT, V.getOperand(0));
25613 // The only other way we can combine them is if only a single element of the
25614 // inner vzext is used in the input to the outer vzext.
25615 if (InnerEltVT.getSizeInBits() < InputBits)
25618 // In this case, the inner vzext is completely dead because we're going to
25619 // only look at bits inside of the low element. Just do the outer vzext on
25620 // a bitcast of the input to the inner.
25621 return DAG.getNode(X86ISD::VZEXT, DL, VT, DAG.getBitcast(OpVT, V));
25624 // Check if we can bypass extracting and re-inserting an element of an input
25625 // vector. Essentially:
25626 // (bitcast (sclr2vec (ext_vec_elt x))) -> (bitcast x)
25627 if (V.getOpcode() == ISD::SCALAR_TO_VECTOR &&
25628 V.getOperand(0).getOpcode() == ISD::EXTRACT_VECTOR_ELT &&
25629 V.getOperand(0).getSimpleValueType().getSizeInBits() == InputBits) {
25630 SDValue ExtractedV = V.getOperand(0);
25631 SDValue OrigV = ExtractedV.getOperand(0);
25632 if (auto *ExtractIdx = dyn_cast<ConstantSDNode>(ExtractedV.getOperand(1)))
25633 if (ExtractIdx->getZExtValue() == 0) {
25634 MVT OrigVT = OrigV.getSimpleValueType();
25635 // Extract a subvector if necessary...
25636 if (OrigVT.getSizeInBits() > OpVT.getSizeInBits()) {
25637 int Ratio = OrigVT.getSizeInBits() / OpVT.getSizeInBits();
25638 OrigVT = MVT::getVectorVT(OrigVT.getVectorElementType(),
25639 OrigVT.getVectorNumElements() / Ratio);
25640 OrigV = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, OrigVT, OrigV,
25641 DAG.getIntPtrConstant(0, DL));
25643 Op = DAG.getBitcast(OpVT, OrigV);
25644 return DAG.getNode(X86ISD::VZEXT, DL, VT, Op);
25651 SDValue X86TargetLowering::PerformDAGCombine(SDNode *N,
25652 DAGCombinerInfo &DCI) const {
25653 SelectionDAG &DAG = DCI.DAG;
25654 switch (N->getOpcode()) {
25656 case ISD::EXTRACT_VECTOR_ELT:
25657 return PerformEXTRACT_VECTOR_ELTCombine(N, DAG, DCI);
25660 case X86ISD::SHRUNKBLEND:
25661 return PerformSELECTCombine(N, DAG, DCI, Subtarget);
25662 case ISD::BITCAST: return PerformBITCASTCombine(N, DAG);
25663 case X86ISD::CMOV: return PerformCMOVCombine(N, DAG, DCI, Subtarget);
25664 case ISD::ADD: return PerformAddCombine(N, DAG, Subtarget);
25665 case ISD::SUB: return PerformSubCombine(N, DAG, Subtarget);
25666 case X86ISD::ADC: return PerformADCCombine(N, DAG, DCI);
25667 case ISD::MUL: return PerformMulCombine(N, DAG, DCI);
25670 case ISD::SRL: return PerformShiftCombine(N, DAG, DCI, Subtarget);
25671 case ISD::AND: return PerformAndCombine(N, DAG, DCI, Subtarget);
25672 case ISD::OR: return PerformOrCombine(N, DAG, DCI, Subtarget);
25673 case ISD::XOR: return PerformXorCombine(N, DAG, DCI, Subtarget);
25674 case ISD::LOAD: return PerformLOADCombine(N, DAG, DCI, Subtarget);
25675 case ISD::MLOAD: return PerformMLOADCombine(N, DAG, DCI, Subtarget);
25676 case ISD::STORE: return PerformSTORECombine(N, DAG, Subtarget);
25677 case ISD::MSTORE: return PerformMSTORECombine(N, DAG, Subtarget);
25678 case ISD::SINT_TO_FP: return PerformSINT_TO_FPCombine(N, DAG, Subtarget);
25679 case ISD::UINT_TO_FP: return PerformUINT_TO_FPCombine(N, DAG, Subtarget);
25680 case ISD::FADD: return PerformFADDCombine(N, DAG, Subtarget);
25681 case ISD::FSUB: return PerformFSUBCombine(N, DAG, Subtarget);
25683 case X86ISD::FOR: return PerformFORCombine(N, DAG);
25685 case X86ISD::FMAX: return PerformFMinFMaxCombine(N, DAG);
25686 case X86ISD::FAND: return PerformFANDCombine(N, DAG);
25687 case X86ISD::FANDN: return PerformFANDNCombine(N, DAG);
25688 case X86ISD::BT: return PerformBTCombine(N, DAG, DCI);
25689 case X86ISD::VZEXT_MOVL: return PerformVZEXT_MOVLCombine(N, DAG);
25690 case ISD::ANY_EXTEND:
25691 case ISD::ZERO_EXTEND: return PerformZExtCombine(N, DAG, DCI, Subtarget);
25692 case ISD::SIGN_EXTEND: return PerformSExtCombine(N, DAG, DCI, Subtarget);
25693 case ISD::SIGN_EXTEND_INREG:
25694 return PerformSIGN_EXTEND_INREGCombine(N, DAG, Subtarget);
25695 case ISD::SETCC: return PerformISDSETCCCombine(N, DAG, Subtarget);
25696 case X86ISD::SETCC: return PerformSETCCCombine(N, DAG, DCI, Subtarget);
25697 case X86ISD::BRCOND: return PerformBrCondCombine(N, DAG, DCI, Subtarget);
25698 case X86ISD::VZEXT: return performVZEXTCombine(N, DAG, DCI, Subtarget);
25699 case X86ISD::SHUFP: // Handle all target specific shuffles
25700 case X86ISD::PALIGNR:
25701 case X86ISD::UNPCKH:
25702 case X86ISD::UNPCKL:
25703 case X86ISD::MOVHLPS:
25704 case X86ISD::MOVLHPS:
25705 case X86ISD::PSHUFB:
25706 case X86ISD::PSHUFD:
25707 case X86ISD::PSHUFHW:
25708 case X86ISD::PSHUFLW:
25709 case X86ISD::MOVSS:
25710 case X86ISD::MOVSD:
25711 case X86ISD::VPERMILPI:
25712 case X86ISD::VPERM2X128:
25713 case ISD::VECTOR_SHUFFLE: return PerformShuffleCombine(N, DAG, DCI,Subtarget);
25714 case ISD::FMA: return PerformFMACombine(N, DAG, Subtarget);
25715 case X86ISD::INSERTPS: {
25716 if (getTargetMachine().getOptLevel() > CodeGenOpt::None)
25717 return PerformINSERTPSCombine(N, DAG, Subtarget);
25720 case X86ISD::BLENDI: return PerformBLENDICombine(N, DAG);
25726 /// isTypeDesirableForOp - Return true if the target has native support for
25727 /// the specified value type and it is 'desirable' to use the type for the
25728 /// given node type. e.g. On x86 i16 is legal, but undesirable since i16
25729 /// instruction encodings are longer and some i16 instructions are slow.
25730 bool X86TargetLowering::isTypeDesirableForOp(unsigned Opc, EVT VT) const {
25731 if (!isTypeLegal(VT))
25733 if (VT != MVT::i16)
25740 case ISD::SIGN_EXTEND:
25741 case ISD::ZERO_EXTEND:
25742 case ISD::ANY_EXTEND:
25755 /// IsDesirableToPromoteOp - This method query the target whether it is
25756 /// beneficial for dag combiner to promote the specified node. If true, it
25757 /// should return the desired promotion type by reference.
25758 bool X86TargetLowering::IsDesirableToPromoteOp(SDValue Op, EVT &PVT) const {
25759 EVT VT = Op.getValueType();
25760 if (VT != MVT::i16)
25763 bool Promote = false;
25764 bool Commute = false;
25765 switch (Op.getOpcode()) {
25768 LoadSDNode *LD = cast<LoadSDNode>(Op);
25769 // If the non-extending load has a single use and it's not live out, then it
25770 // might be folded.
25771 if (LD->getExtensionType() == ISD::NON_EXTLOAD /*&&
25772 Op.hasOneUse()*/) {
25773 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
25774 UE = Op.getNode()->use_end(); UI != UE; ++UI) {
25775 // The only case where we'd want to promote LOAD (rather then it being
25776 // promoted as an operand is when it's only use is liveout.
25777 if (UI->getOpcode() != ISD::CopyToReg)
25784 case ISD::SIGN_EXTEND:
25785 case ISD::ZERO_EXTEND:
25786 case ISD::ANY_EXTEND:
25791 SDValue N0 = Op.getOperand(0);
25792 // Look out for (store (shl (load), x)).
25793 if (MayFoldLoad(N0) && MayFoldIntoStore(Op))
25806 SDValue N0 = Op.getOperand(0);
25807 SDValue N1 = Op.getOperand(1);
25808 if (!Commute && MayFoldLoad(N1))
25810 // Avoid disabling potential load folding opportunities.
25811 if (MayFoldLoad(N0) && (!isa<ConstantSDNode>(N1) || MayFoldIntoStore(Op)))
25813 if (MayFoldLoad(N1) && (!isa<ConstantSDNode>(N0) || MayFoldIntoStore(Op)))
25823 //===----------------------------------------------------------------------===//
25824 // X86 Inline Assembly Support
25825 //===----------------------------------------------------------------------===//
25827 // Helper to match a string separated by whitespace.
25828 static bool matchAsm(StringRef S, ArrayRef<const char *> Pieces) {
25829 S = S.substr(S.find_first_not_of(" \t")); // Skip leading whitespace.
25831 for (StringRef Piece : Pieces) {
25832 if (!S.startswith(Piece)) // Check if the piece matches.
25835 S = S.substr(Piece.size());
25836 StringRef::size_type Pos = S.find_first_not_of(" \t");
25837 if (Pos == 0) // We matched a prefix.
25846 static bool clobbersFlagRegisters(const SmallVector<StringRef, 4> &AsmPieces) {
25848 if (AsmPieces.size() == 3 || AsmPieces.size() == 4) {
25849 if (std::count(AsmPieces.begin(), AsmPieces.end(), "~{cc}") &&
25850 std::count(AsmPieces.begin(), AsmPieces.end(), "~{flags}") &&
25851 std::count(AsmPieces.begin(), AsmPieces.end(), "~{fpsr}")) {
25853 if (AsmPieces.size() == 3)
25855 else if (std::count(AsmPieces.begin(), AsmPieces.end(), "~{dirflag}"))
25862 bool X86TargetLowering::ExpandInlineAsm(CallInst *CI) const {
25863 InlineAsm *IA = cast<InlineAsm>(CI->getCalledValue());
25865 std::string AsmStr = IA->getAsmString();
25867 IntegerType *Ty = dyn_cast<IntegerType>(CI->getType());
25868 if (!Ty || Ty->getBitWidth() % 16 != 0)
25871 // TODO: should remove alternatives from the asmstring: "foo {a|b}" -> "foo a"
25872 SmallVector<StringRef, 4> AsmPieces;
25873 SplitString(AsmStr, AsmPieces, ";\n");
25875 switch (AsmPieces.size()) {
25876 default: return false;
25878 // FIXME: this should verify that we are targeting a 486 or better. If not,
25879 // we will turn this bswap into something that will be lowered to logical
25880 // ops instead of emitting the bswap asm. For now, we don't support 486 or
25881 // lower so don't worry about this.
25883 if (matchAsm(AsmPieces[0], {"bswap", "$0"}) ||
25884 matchAsm(AsmPieces[0], {"bswapl", "$0"}) ||
25885 matchAsm(AsmPieces[0], {"bswapq", "$0"}) ||
25886 matchAsm(AsmPieces[0], {"bswap", "${0:q}"}) ||
25887 matchAsm(AsmPieces[0], {"bswapl", "${0:q}"}) ||
25888 matchAsm(AsmPieces[0], {"bswapq", "${0:q}"})) {
25889 // No need to check constraints, nothing other than the equivalent of
25890 // "=r,0" would be valid here.
25891 return IntrinsicLowering::LowerToByteSwap(CI);
25894 // rorw $$8, ${0:w} --> llvm.bswap.i16
25895 if (CI->getType()->isIntegerTy(16) &&
25896 IA->getConstraintString().compare(0, 5, "=r,0,") == 0 &&
25897 (matchAsm(AsmPieces[0], {"rorw", "$$8,", "${0:w}"}) ||
25898 matchAsm(AsmPieces[0], {"rolw", "$$8,", "${0:w}"}))) {
25900 StringRef ConstraintsStr = IA->getConstraintString();
25901 SplitString(StringRef(ConstraintsStr).substr(5), AsmPieces, ",");
25902 array_pod_sort(AsmPieces.begin(), AsmPieces.end());
25903 if (clobbersFlagRegisters(AsmPieces))
25904 return IntrinsicLowering::LowerToByteSwap(CI);
25908 if (CI->getType()->isIntegerTy(32) &&
25909 IA->getConstraintString().compare(0, 5, "=r,0,") == 0 &&
25910 matchAsm(AsmPieces[0], {"rorw", "$$8,", "${0:w}"}) &&
25911 matchAsm(AsmPieces[1], {"rorl", "$$16,", "$0"}) &&
25912 matchAsm(AsmPieces[2], {"rorw", "$$8,", "${0:w}"})) {
25914 StringRef ConstraintsStr = IA->getConstraintString();
25915 SplitString(StringRef(ConstraintsStr).substr(5), AsmPieces, ",");
25916 array_pod_sort(AsmPieces.begin(), AsmPieces.end());
25917 if (clobbersFlagRegisters(AsmPieces))
25918 return IntrinsicLowering::LowerToByteSwap(CI);
25921 if (CI->getType()->isIntegerTy(64)) {
25922 InlineAsm::ConstraintInfoVector Constraints = IA->ParseConstraints();
25923 if (Constraints.size() >= 2 &&
25924 Constraints[0].Codes.size() == 1 && Constraints[0].Codes[0] == "A" &&
25925 Constraints[1].Codes.size() == 1 && Constraints[1].Codes[0] == "0") {
25926 // bswap %eax / bswap %edx / xchgl %eax, %edx -> llvm.bswap.i64
25927 if (matchAsm(AsmPieces[0], {"bswap", "%eax"}) &&
25928 matchAsm(AsmPieces[1], {"bswap", "%edx"}) &&
25929 matchAsm(AsmPieces[2], {"xchgl", "%eax,", "%edx"}))
25930 return IntrinsicLowering::LowerToByteSwap(CI);
25938 /// getConstraintType - Given a constraint letter, return the type of
25939 /// constraint it is for this target.
25940 X86TargetLowering::ConstraintType
25941 X86TargetLowering::getConstraintType(StringRef Constraint) const {
25942 if (Constraint.size() == 1) {
25943 switch (Constraint[0]) {
25954 return C_RegisterClass;
25978 return TargetLowering::getConstraintType(Constraint);
25981 /// Examine constraint type and operand type and determine a weight value.
25982 /// This object must already have been set up with the operand type
25983 /// and the current alternative constraint selected.
25984 TargetLowering::ConstraintWeight
25985 X86TargetLowering::getSingleConstraintMatchWeight(
25986 AsmOperandInfo &info, const char *constraint) const {
25987 ConstraintWeight weight = CW_Invalid;
25988 Value *CallOperandVal = info.CallOperandVal;
25989 // If we don't have a value, we can't do a match,
25990 // but allow it at the lowest weight.
25991 if (!CallOperandVal)
25993 Type *type = CallOperandVal->getType();
25994 // Look at the constraint type.
25995 switch (*constraint) {
25997 weight = TargetLowering::getSingleConstraintMatchWeight(info, constraint);
26008 if (CallOperandVal->getType()->isIntegerTy())
26009 weight = CW_SpecificReg;
26014 if (type->isFloatingPointTy())
26015 weight = CW_SpecificReg;
26018 if (type->isX86_MMXTy() && Subtarget->hasMMX())
26019 weight = CW_SpecificReg;
26023 if (((type->getPrimitiveSizeInBits() == 128) && Subtarget->hasSSE1()) ||
26024 ((type->getPrimitiveSizeInBits() == 256) && Subtarget->hasFp256()))
26025 weight = CW_Register;
26028 if (ConstantInt *C = dyn_cast<ConstantInt>(info.CallOperandVal)) {
26029 if (C->getZExtValue() <= 31)
26030 weight = CW_Constant;
26034 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
26035 if (C->getZExtValue() <= 63)
26036 weight = CW_Constant;
26040 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
26041 if ((C->getSExtValue() >= -0x80) && (C->getSExtValue() <= 0x7f))
26042 weight = CW_Constant;
26046 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
26047 if ((C->getZExtValue() == 0xff) || (C->getZExtValue() == 0xffff))
26048 weight = CW_Constant;
26052 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
26053 if (C->getZExtValue() <= 3)
26054 weight = CW_Constant;
26058 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
26059 if (C->getZExtValue() <= 0xff)
26060 weight = CW_Constant;
26065 if (isa<ConstantFP>(CallOperandVal)) {
26066 weight = CW_Constant;
26070 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
26071 if ((C->getSExtValue() >= -0x80000000LL) &&
26072 (C->getSExtValue() <= 0x7fffffffLL))
26073 weight = CW_Constant;
26077 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
26078 if (C->getZExtValue() <= 0xffffffff)
26079 weight = CW_Constant;
26086 /// LowerXConstraint - try to replace an X constraint, which matches anything,
26087 /// with another that has more specific requirements based on the type of the
26088 /// corresponding operand.
26089 const char *X86TargetLowering::
26090 LowerXConstraint(EVT ConstraintVT) const {
26091 // FP X constraints get lowered to SSE1/2 registers if available, otherwise
26092 // 'f' like normal targets.
26093 if (ConstraintVT.isFloatingPoint()) {
26094 if (Subtarget->hasSSE2())
26096 if (Subtarget->hasSSE1())
26100 return TargetLowering::LowerXConstraint(ConstraintVT);
26103 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
26104 /// vector. If it is invalid, don't add anything to Ops.
26105 void X86TargetLowering::LowerAsmOperandForConstraint(SDValue Op,
26106 std::string &Constraint,
26107 std::vector<SDValue>&Ops,
26108 SelectionDAG &DAG) const {
26111 // Only support length 1 constraints for now.
26112 if (Constraint.length() > 1) return;
26114 char ConstraintLetter = Constraint[0];
26115 switch (ConstraintLetter) {
26118 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
26119 if (C->getZExtValue() <= 31) {
26120 Result = DAG.getTargetConstant(C->getZExtValue(), SDLoc(Op),
26121 Op.getValueType());
26127 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
26128 if (C->getZExtValue() <= 63) {
26129 Result = DAG.getTargetConstant(C->getZExtValue(), SDLoc(Op),
26130 Op.getValueType());
26136 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
26137 if (isInt<8>(C->getSExtValue())) {
26138 Result = DAG.getTargetConstant(C->getZExtValue(), SDLoc(Op),
26139 Op.getValueType());
26145 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
26146 if (C->getZExtValue() == 0xff || C->getZExtValue() == 0xffff ||
26147 (Subtarget->is64Bit() && C->getZExtValue() == 0xffffffff)) {
26148 Result = DAG.getTargetConstant(C->getSExtValue(), SDLoc(Op),
26149 Op.getValueType());
26155 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
26156 if (C->getZExtValue() <= 3) {
26157 Result = DAG.getTargetConstant(C->getZExtValue(), SDLoc(Op),
26158 Op.getValueType());
26164 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
26165 if (C->getZExtValue() <= 255) {
26166 Result = DAG.getTargetConstant(C->getZExtValue(), SDLoc(Op),
26167 Op.getValueType());
26173 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
26174 if (C->getZExtValue() <= 127) {
26175 Result = DAG.getTargetConstant(C->getZExtValue(), SDLoc(Op),
26176 Op.getValueType());
26182 // 32-bit signed value
26183 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
26184 if (ConstantInt::isValueValidForType(Type::getInt32Ty(*DAG.getContext()),
26185 C->getSExtValue())) {
26186 // Widen to 64 bits here to get it sign extended.
26187 Result = DAG.getTargetConstant(C->getSExtValue(), SDLoc(Op), MVT::i64);
26190 // FIXME gcc accepts some relocatable values here too, but only in certain
26191 // memory models; it's complicated.
26196 // 32-bit unsigned value
26197 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
26198 if (ConstantInt::isValueValidForType(Type::getInt32Ty(*DAG.getContext()),
26199 C->getZExtValue())) {
26200 Result = DAG.getTargetConstant(C->getZExtValue(), SDLoc(Op),
26201 Op.getValueType());
26205 // FIXME gcc accepts some relocatable values here too, but only in certain
26206 // memory models; it's complicated.
26210 // Literal immediates are always ok.
26211 if (ConstantSDNode *CST = dyn_cast<ConstantSDNode>(Op)) {
26212 // Widen to 64 bits here to get it sign extended.
26213 Result = DAG.getTargetConstant(CST->getSExtValue(), SDLoc(Op), MVT::i64);
26217 // In any sort of PIC mode addresses need to be computed at runtime by
26218 // adding in a register or some sort of table lookup. These can't
26219 // be used as immediates.
26220 if (Subtarget->isPICStyleGOT() || Subtarget->isPICStyleStubPIC())
26223 // If we are in non-pic codegen mode, we allow the address of a global (with
26224 // an optional displacement) to be used with 'i'.
26225 GlobalAddressSDNode *GA = nullptr;
26226 int64_t Offset = 0;
26228 // Match either (GA), (GA+C), (GA+C1+C2), etc.
26230 if ((GA = dyn_cast<GlobalAddressSDNode>(Op))) {
26231 Offset += GA->getOffset();
26233 } else if (Op.getOpcode() == ISD::ADD) {
26234 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
26235 Offset += C->getZExtValue();
26236 Op = Op.getOperand(0);
26239 } else if (Op.getOpcode() == ISD::SUB) {
26240 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
26241 Offset += -C->getZExtValue();
26242 Op = Op.getOperand(0);
26247 // Otherwise, this isn't something we can handle, reject it.
26251 const GlobalValue *GV = GA->getGlobal();
26252 // If we require an extra load to get this address, as in PIC mode, we
26253 // can't accept it.
26254 if (isGlobalStubReference(
26255 Subtarget->ClassifyGlobalReference(GV, DAG.getTarget())))
26258 Result = DAG.getTargetGlobalAddress(GV, SDLoc(Op),
26259 GA->getValueType(0), Offset);
26264 if (Result.getNode()) {
26265 Ops.push_back(Result);
26268 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG);
26271 std::pair<unsigned, const TargetRegisterClass *>
26272 X86TargetLowering::getRegForInlineAsmConstraint(const TargetRegisterInfo *TRI,
26273 StringRef Constraint,
26275 // First, see if this is a constraint that directly corresponds to an LLVM
26277 if (Constraint.size() == 1) {
26278 // GCC Constraint Letters
26279 switch (Constraint[0]) {
26281 // TODO: Slight differences here in allocation order and leaving
26282 // RIP in the class. Do they matter any more here than they do
26283 // in the normal allocation?
26284 case 'q': // GENERAL_REGS in 64-bit mode, Q_REGS in 32-bit mode.
26285 if (Subtarget->is64Bit()) {
26286 if (VT == MVT::i32 || VT == MVT::f32)
26287 return std::make_pair(0U, &X86::GR32RegClass);
26288 if (VT == MVT::i16)
26289 return std::make_pair(0U, &X86::GR16RegClass);
26290 if (VT == MVT::i8 || VT == MVT::i1)
26291 return std::make_pair(0U, &X86::GR8RegClass);
26292 if (VT == MVT::i64 || VT == MVT::f64)
26293 return std::make_pair(0U, &X86::GR64RegClass);
26296 // 32-bit fallthrough
26297 case 'Q': // Q_REGS
26298 if (VT == MVT::i32 || VT == MVT::f32)
26299 return std::make_pair(0U, &X86::GR32_ABCDRegClass);
26300 if (VT == MVT::i16)
26301 return std::make_pair(0U, &X86::GR16_ABCDRegClass);
26302 if (VT == MVT::i8 || VT == MVT::i1)
26303 return std::make_pair(0U, &X86::GR8_ABCD_LRegClass);
26304 if (VT == MVT::i64)
26305 return std::make_pair(0U, &X86::GR64_ABCDRegClass);
26307 case 'r': // GENERAL_REGS
26308 case 'l': // INDEX_REGS
26309 if (VT == MVT::i8 || VT == MVT::i1)
26310 return std::make_pair(0U, &X86::GR8RegClass);
26311 if (VT == MVT::i16)
26312 return std::make_pair(0U, &X86::GR16RegClass);
26313 if (VT == MVT::i32 || VT == MVT::f32 || !Subtarget->is64Bit())
26314 return std::make_pair(0U, &X86::GR32RegClass);
26315 return std::make_pair(0U, &X86::GR64RegClass);
26316 case 'R': // LEGACY_REGS
26317 if (VT == MVT::i8 || VT == MVT::i1)
26318 return std::make_pair(0U, &X86::GR8_NOREXRegClass);
26319 if (VT == MVT::i16)
26320 return std::make_pair(0U, &X86::GR16_NOREXRegClass);
26321 if (VT == MVT::i32 || !Subtarget->is64Bit())
26322 return std::make_pair(0U, &X86::GR32_NOREXRegClass);
26323 return std::make_pair(0U, &X86::GR64_NOREXRegClass);
26324 case 'f': // FP Stack registers.
26325 // If SSE is enabled for this VT, use f80 to ensure the isel moves the
26326 // value to the correct fpstack register class.
26327 if (VT == MVT::f32 && !isScalarFPTypeInSSEReg(VT))
26328 return std::make_pair(0U, &X86::RFP32RegClass);
26329 if (VT == MVT::f64 && !isScalarFPTypeInSSEReg(VT))
26330 return std::make_pair(0U, &X86::RFP64RegClass);
26331 return std::make_pair(0U, &X86::RFP80RegClass);
26332 case 'y': // MMX_REGS if MMX allowed.
26333 if (!Subtarget->hasMMX()) break;
26334 return std::make_pair(0U, &X86::VR64RegClass);
26335 case 'Y': // SSE_REGS if SSE2 allowed
26336 if (!Subtarget->hasSSE2()) break;
26338 case 'x': // SSE_REGS if SSE1 allowed or AVX_REGS if AVX allowed
26339 if (!Subtarget->hasSSE1()) break;
26341 switch (VT.SimpleTy) {
26343 // Scalar SSE types.
26346 return std::make_pair(0U, &X86::FR32RegClass);
26349 return std::make_pair(0U, &X86::FR64RegClass);
26357 return std::make_pair(0U, &X86::VR128RegClass);
26365 return std::make_pair(0U, &X86::VR256RegClass);
26370 return std::make_pair(0U, &X86::VR512RegClass);
26376 // Use the default implementation in TargetLowering to convert the register
26377 // constraint into a member of a register class.
26378 std::pair<unsigned, const TargetRegisterClass*> Res;
26379 Res = TargetLowering::getRegForInlineAsmConstraint(TRI, Constraint, VT);
26381 // Not found as a standard register?
26383 // Map st(0) -> st(7) -> ST0
26384 if (Constraint.size() == 7 && Constraint[0] == '{' &&
26385 tolower(Constraint[1]) == 's' &&
26386 tolower(Constraint[2]) == 't' &&
26387 Constraint[3] == '(' &&
26388 (Constraint[4] >= '0' && Constraint[4] <= '7') &&
26389 Constraint[5] == ')' &&
26390 Constraint[6] == '}') {
26392 Res.first = X86::FP0+Constraint[4]-'0';
26393 Res.second = &X86::RFP80RegClass;
26397 // GCC allows "st(0)" to be called just plain "st".
26398 if (StringRef("{st}").equals_lower(Constraint)) {
26399 Res.first = X86::FP0;
26400 Res.second = &X86::RFP80RegClass;
26405 if (StringRef("{flags}").equals_lower(Constraint)) {
26406 Res.first = X86::EFLAGS;
26407 Res.second = &X86::CCRRegClass;
26411 // 'A' means EAX + EDX.
26412 if (Constraint == "A") {
26413 Res.first = X86::EAX;
26414 Res.second = &X86::GR32_ADRegClass;
26420 // Otherwise, check to see if this is a register class of the wrong value
26421 // type. For example, we want to map "{ax},i32" -> {eax}, we don't want it to
26422 // turn into {ax},{dx}.
26423 // MVT::Other is used to specify clobber names.
26424 if (Res.second->hasType(VT) || VT == MVT::Other)
26425 return Res; // Correct type already, nothing to do.
26427 // Get a matching integer of the correct size. i.e. "ax" with MVT::32 should
26428 // return "eax". This should even work for things like getting 64bit integer
26429 // registers when given an f64 type.
26430 const TargetRegisterClass *Class = Res.second;
26431 if (Class == &X86::GR8RegClass || Class == &X86::GR16RegClass ||
26432 Class == &X86::GR32RegClass || Class == &X86::GR64RegClass) {
26433 unsigned Size = VT.getSizeInBits();
26434 MVT::SimpleValueType SimpleTy = Size == 1 || Size == 8 ? MVT::i8
26435 : Size == 16 ? MVT::i16
26436 : Size == 32 ? MVT::i32
26437 : Size == 64 ? MVT::i64
26439 unsigned DestReg = getX86SubSuperRegisterOrZero(Res.first, SimpleTy);
26441 Res.first = DestReg;
26442 Res.second = SimpleTy == MVT::i8 ? &X86::GR8RegClass
26443 : SimpleTy == MVT::i16 ? &X86::GR16RegClass
26444 : SimpleTy == MVT::i32 ? &X86::GR32RegClass
26445 : &X86::GR64RegClass;
26446 assert(Res.second->contains(Res.first) && "Register in register class");
26448 // No register found/type mismatch.
26450 Res.second = nullptr;
26452 } else if (Class == &X86::FR32RegClass || Class == &X86::FR64RegClass ||
26453 Class == &X86::VR128RegClass || Class == &X86::VR256RegClass ||
26454 Class == &X86::FR32XRegClass || Class == &X86::FR64XRegClass ||
26455 Class == &X86::VR128XRegClass || Class == &X86::VR256XRegClass ||
26456 Class == &X86::VR512RegClass) {
26457 // Handle references to XMM physical registers that got mapped into the
26458 // wrong class. This can happen with constraints like {xmm0} where the
26459 // target independent register mapper will just pick the first match it can
26460 // find, ignoring the required type.
26462 if (VT == MVT::f32 || VT == MVT::i32)
26463 Res.second = &X86::FR32RegClass;
26464 else if (VT == MVT::f64 || VT == MVT::i64)
26465 Res.second = &X86::FR64RegClass;
26466 else if (X86::VR128RegClass.hasType(VT))
26467 Res.second = &X86::VR128RegClass;
26468 else if (X86::VR256RegClass.hasType(VT))
26469 Res.second = &X86::VR256RegClass;
26470 else if (X86::VR512RegClass.hasType(VT))
26471 Res.second = &X86::VR512RegClass;
26473 // Type mismatch and not a clobber: Return an error;
26475 Res.second = nullptr;
26482 int X86TargetLowering::getScalingFactorCost(const DataLayout &DL,
26483 const AddrMode &AM, Type *Ty,
26484 unsigned AS) const {
26485 // Scaling factors are not free at all.
26486 // An indexed folded instruction, i.e., inst (reg1, reg2, scale),
26487 // will take 2 allocations in the out of order engine instead of 1
26488 // for plain addressing mode, i.e. inst (reg1).
26490 // vaddps (%rsi,%drx), %ymm0, %ymm1
26491 // Requires two allocations (one for the load, one for the computation)
26493 // vaddps (%rsi), %ymm0, %ymm1
26494 // Requires just 1 allocation, i.e., freeing allocations for other operations
26495 // and having less micro operations to execute.
26497 // For some X86 architectures, this is even worse because for instance for
26498 // stores, the complex addressing mode forces the instruction to use the
26499 // "load" ports instead of the dedicated "store" port.
26500 // E.g., on Haswell:
26501 // vmovaps %ymm1, (%r8, %rdi) can use port 2 or 3.
26502 // vmovaps %ymm1, (%r8) can use port 2, 3, or 7.
26503 if (isLegalAddressingMode(DL, AM, Ty, AS))
26504 // Scale represents reg2 * scale, thus account for 1
26505 // as soon as we use a second register.
26506 return AM.Scale != 0;
26510 bool X86TargetLowering::isTargetFTOL() const {
26511 return Subtarget->isTargetKnownWindowsMSVC() && !Subtarget->is64Bit();
26514 bool X86TargetLowering::isIntDivCheap(EVT VT, bool OptSize) const {
26515 // Integer division on x86 is expensive. However, when aggressively optimizing
26516 // for code size, we prefer to use a div instruction, as it is usually smaller
26517 // than the alternative sequence.
26518 // The exception to this is vector division. Since x86 doesn't have vector
26519 // integer division, leaving the division as-is is a loss even in terms of
26520 // size, because it will have to be scalarized, while the alternative code
26521 // sequence can be performed in vector form.
26522 return OptSize && !VT.isVector();