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.
79 // X86 is weird. It always uses i8 for shift amounts and setcc results.
80 setBooleanContents(ZeroOrOneBooleanContent);
81 // X86-SSE is even stranger. It uses -1 or 0 for vector masks.
82 setBooleanVectorContents(ZeroOrNegativeOneBooleanContent);
84 // For 64-bit, since we have so many registers, use the ILP scheduler.
85 // For 32-bit, use the register pressure specific scheduling.
86 // For Atom, always use ILP scheduling.
87 if (Subtarget->isAtom())
88 setSchedulingPreference(Sched::ILP);
89 else if (Subtarget->is64Bit())
90 setSchedulingPreference(Sched::ILP);
92 setSchedulingPreference(Sched::RegPressure);
93 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
94 setStackPointerRegisterToSaveRestore(RegInfo->getStackRegister());
96 // Bypass expensive divides on Atom when compiling with O2.
97 if (TM.getOptLevel() >= CodeGenOpt::Default) {
98 if (Subtarget->hasSlowDivide32())
99 addBypassSlowDiv(32, 8);
100 if (Subtarget->hasSlowDivide64() && Subtarget->is64Bit())
101 addBypassSlowDiv(64, 16);
104 if (Subtarget->isTargetKnownWindowsMSVC()) {
105 // Setup Windows compiler runtime calls.
106 setLibcallName(RTLIB::SDIV_I64, "_alldiv");
107 setLibcallName(RTLIB::UDIV_I64, "_aulldiv");
108 setLibcallName(RTLIB::SREM_I64, "_allrem");
109 setLibcallName(RTLIB::UREM_I64, "_aullrem");
110 setLibcallName(RTLIB::MUL_I64, "_allmul");
111 setLibcallCallingConv(RTLIB::SDIV_I64, CallingConv::X86_StdCall);
112 setLibcallCallingConv(RTLIB::UDIV_I64, CallingConv::X86_StdCall);
113 setLibcallCallingConv(RTLIB::SREM_I64, CallingConv::X86_StdCall);
114 setLibcallCallingConv(RTLIB::UREM_I64, CallingConv::X86_StdCall);
115 setLibcallCallingConv(RTLIB::MUL_I64, CallingConv::X86_StdCall);
118 if (Subtarget->isTargetDarwin()) {
119 // Darwin should use _setjmp/_longjmp instead of setjmp/longjmp.
120 setUseUnderscoreSetJmp(false);
121 setUseUnderscoreLongJmp(false);
122 } else if (Subtarget->isTargetWindowsGNU()) {
123 // MS runtime is weird: it exports _setjmp, but longjmp!
124 setUseUnderscoreSetJmp(true);
125 setUseUnderscoreLongJmp(false);
127 setUseUnderscoreSetJmp(true);
128 setUseUnderscoreLongJmp(true);
131 // Set up the register classes.
132 addRegisterClass(MVT::i8, &X86::GR8RegClass);
133 addRegisterClass(MVT::i16, &X86::GR16RegClass);
134 addRegisterClass(MVT::i32, &X86::GR32RegClass);
135 if (Subtarget->is64Bit())
136 addRegisterClass(MVT::i64, &X86::GR64RegClass);
138 for (MVT VT : MVT::integer_valuetypes())
139 setLoadExtAction(ISD::SEXTLOAD, VT, MVT::i1, Promote);
141 // We don't accept any truncstore of integer registers.
142 setTruncStoreAction(MVT::i64, MVT::i32, Expand);
143 setTruncStoreAction(MVT::i64, MVT::i16, Expand);
144 setTruncStoreAction(MVT::i64, MVT::i8 , Expand);
145 setTruncStoreAction(MVT::i32, MVT::i16, Expand);
146 setTruncStoreAction(MVT::i32, MVT::i8 , Expand);
147 setTruncStoreAction(MVT::i16, MVT::i8, Expand);
149 setTruncStoreAction(MVT::f64, MVT::f32, Expand);
151 // SETOEQ and SETUNE require checking two conditions.
152 setCondCodeAction(ISD::SETOEQ, MVT::f32, Expand);
153 setCondCodeAction(ISD::SETOEQ, MVT::f64, Expand);
154 setCondCodeAction(ISD::SETOEQ, MVT::f80, Expand);
155 setCondCodeAction(ISD::SETUNE, MVT::f32, Expand);
156 setCondCodeAction(ISD::SETUNE, MVT::f64, Expand);
157 setCondCodeAction(ISD::SETUNE, MVT::f80, Expand);
159 // Promote all UINT_TO_FP to larger SINT_TO_FP's, as X86 doesn't have this
161 setOperationAction(ISD::UINT_TO_FP , MVT::i1 , Promote);
162 setOperationAction(ISD::UINT_TO_FP , MVT::i8 , Promote);
163 setOperationAction(ISD::UINT_TO_FP , MVT::i16 , Promote);
165 if (Subtarget->is64Bit()) {
166 if (!Subtarget->useSoftFloat() && Subtarget->hasAVX512())
167 // f32/f64 are legal, f80 is custom.
168 setOperationAction(ISD::UINT_TO_FP , MVT::i32 , Custom);
170 setOperationAction(ISD::UINT_TO_FP , MVT::i32 , Promote);
171 setOperationAction(ISD::UINT_TO_FP , MVT::i64 , Custom);
172 } else if (!Subtarget->useSoftFloat()) {
173 // We have an algorithm for SSE2->double, and we turn this into a
174 // 64-bit FILD followed by conditional FADD for other targets.
175 setOperationAction(ISD::UINT_TO_FP , MVT::i64 , Custom);
176 // We have an algorithm for SSE2, and we turn this into a 64-bit
177 // FILD or VCVTUSI2SS/SD for other targets.
178 setOperationAction(ISD::UINT_TO_FP , MVT::i32 , Custom);
181 // Promote i1/i8 SINT_TO_FP to larger SINT_TO_FP's, as X86 doesn't have
183 setOperationAction(ISD::SINT_TO_FP , MVT::i1 , Promote);
184 setOperationAction(ISD::SINT_TO_FP , MVT::i8 , Promote);
186 if (!Subtarget->useSoftFloat()) {
187 // SSE has no i16 to fp conversion, only i32
188 if (X86ScalarSSEf32) {
189 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Promote);
190 // f32 and f64 cases are Legal, f80 case is not
191 setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Custom);
193 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Custom);
194 setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Custom);
197 setOperationAction(ISD::SINT_TO_FP , MVT::i16 , Promote);
198 setOperationAction(ISD::SINT_TO_FP , MVT::i32 , Promote);
201 // In 32-bit mode these are custom lowered. In 64-bit mode F32 and F64
202 // are Legal, f80 is custom lowered.
203 setOperationAction(ISD::FP_TO_SINT , MVT::i64 , Custom);
204 setOperationAction(ISD::SINT_TO_FP , MVT::i64 , Custom);
206 // Promote i1/i8 FP_TO_SINT to larger FP_TO_SINTS's, as X86 doesn't have
208 setOperationAction(ISD::FP_TO_SINT , MVT::i1 , Promote);
209 setOperationAction(ISD::FP_TO_SINT , MVT::i8 , Promote);
211 if (X86ScalarSSEf32) {
212 setOperationAction(ISD::FP_TO_SINT , MVT::i16 , Promote);
213 // f32 and f64 cases are Legal, f80 case is not
214 setOperationAction(ISD::FP_TO_SINT , MVT::i32 , Custom);
216 setOperationAction(ISD::FP_TO_SINT , MVT::i16 , Custom);
217 setOperationAction(ISD::FP_TO_SINT , MVT::i32 , Custom);
220 // Handle FP_TO_UINT by promoting the destination to a larger signed
222 setOperationAction(ISD::FP_TO_UINT , MVT::i1 , Promote);
223 setOperationAction(ISD::FP_TO_UINT , MVT::i8 , Promote);
224 setOperationAction(ISD::FP_TO_UINT , MVT::i16 , Promote);
226 if (Subtarget->is64Bit()) {
227 if (!Subtarget->useSoftFloat() && Subtarget->hasAVX512()) {
228 // FP_TO_UINT-i32/i64 is legal for f32/f64, but custom for f80.
229 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Custom);
230 setOperationAction(ISD::FP_TO_UINT , MVT::i64 , Custom);
232 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Promote);
233 setOperationAction(ISD::FP_TO_UINT , MVT::i64 , Expand);
235 } else if (!Subtarget->useSoftFloat()) {
236 // Since AVX is a superset of SSE3, only check for SSE here.
237 if (Subtarget->hasSSE1() && !Subtarget->hasSSE3())
238 // Expand FP_TO_UINT into a select.
239 // FIXME: We would like to use a Custom expander here eventually to do
240 // the optimal thing for SSE vs. the default expansion in the legalizer.
241 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Expand);
243 // With AVX512 we can use vcvts[ds]2usi for f32/f64->i32, f80 is custom.
244 // With SSE3 we can use fisttpll to convert to a signed i64; without
245 // SSE, we're stuck with a fistpll.
246 setOperationAction(ISD::FP_TO_UINT , MVT::i32 , Custom);
248 setOperationAction(ISD::FP_TO_UINT , MVT::i64 , Custom);
251 // TODO: when we have SSE, these could be more efficient, by using movd/movq.
252 if (!X86ScalarSSEf64) {
253 setOperationAction(ISD::BITCAST , MVT::f32 , Expand);
254 setOperationAction(ISD::BITCAST , MVT::i32 , Expand);
255 if (Subtarget->is64Bit()) {
256 setOperationAction(ISD::BITCAST , MVT::f64 , Expand);
257 // Without SSE, i64->f64 goes through memory.
258 setOperationAction(ISD::BITCAST , MVT::i64 , Expand);
262 // Scalar integer divide and remainder are lowered to use operations that
263 // produce two results, to match the available instructions. This exposes
264 // the two-result form to trivial CSE, which is able to combine x/y and x%y
265 // into a single instruction.
267 // Scalar integer multiply-high is also lowered to use two-result
268 // operations, to match the available instructions. However, plain multiply
269 // (low) operations are left as Legal, as there are single-result
270 // instructions for this in x86. Using the two-result multiply instructions
271 // when both high and low results are needed must be arranged by dagcombine.
272 for (auto VT : { MVT::i8, MVT::i16, MVT::i32, MVT::i64 }) {
273 setOperationAction(ISD::MULHS, VT, Expand);
274 setOperationAction(ISD::MULHU, VT, Expand);
275 setOperationAction(ISD::SDIV, VT, Expand);
276 setOperationAction(ISD::UDIV, VT, Expand);
277 setOperationAction(ISD::SREM, VT, Expand);
278 setOperationAction(ISD::UREM, VT, Expand);
280 // Add/Sub overflow ops with MVT::Glues are lowered to EFLAGS dependences.
281 setOperationAction(ISD::ADDC, VT, Custom);
282 setOperationAction(ISD::ADDE, VT, Custom);
283 setOperationAction(ISD::SUBC, VT, Custom);
284 setOperationAction(ISD::SUBE, VT, Custom);
287 setOperationAction(ISD::BR_JT , MVT::Other, Expand);
288 setOperationAction(ISD::BRCOND , MVT::Other, Custom);
289 setOperationAction(ISD::BR_CC , MVT::f32, Expand);
290 setOperationAction(ISD::BR_CC , MVT::f64, Expand);
291 setOperationAction(ISD::BR_CC , MVT::f80, Expand);
292 setOperationAction(ISD::BR_CC , MVT::i8, Expand);
293 setOperationAction(ISD::BR_CC , MVT::i16, Expand);
294 setOperationAction(ISD::BR_CC , MVT::i32, Expand);
295 setOperationAction(ISD::BR_CC , MVT::i64, Expand);
296 setOperationAction(ISD::SELECT_CC , MVT::f32, Expand);
297 setOperationAction(ISD::SELECT_CC , MVT::f64, Expand);
298 setOperationAction(ISD::SELECT_CC , MVT::f80, Expand);
299 setOperationAction(ISD::SELECT_CC , MVT::i8, Expand);
300 setOperationAction(ISD::SELECT_CC , MVT::i16, Expand);
301 setOperationAction(ISD::SELECT_CC , MVT::i32, Expand);
302 setOperationAction(ISD::SELECT_CC , MVT::i64, Expand);
303 if (Subtarget->is64Bit())
304 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i32, Legal);
305 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i16 , Legal);
306 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i8 , Legal);
307 setOperationAction(ISD::SIGN_EXTEND_INREG, MVT::i1 , Expand);
308 setOperationAction(ISD::FP_ROUND_INREG , MVT::f32 , Expand);
310 if (Subtarget->is32Bit() && Subtarget->isTargetKnownWindowsMSVC()) {
311 // On 32 bit MSVC, `fmodf(f32)` is not defined - only `fmod(f64)`
312 // is. We should promote the value to 64-bits to solve this.
313 // This is what the CRT headers do - `fmodf` is an inline header
314 // function casting to f64 and calling `fmod`.
315 setOperationAction(ISD::FREM , MVT::f32 , Promote);
317 setOperationAction(ISD::FREM , MVT::f32 , Expand);
320 setOperationAction(ISD::FREM , MVT::f64 , Expand);
321 setOperationAction(ISD::FREM , MVT::f80 , Expand);
322 setOperationAction(ISD::FLT_ROUNDS_ , MVT::i32 , Custom);
324 // Promote the i8 variants and force them on up to i32 which has a shorter
326 setOperationAction(ISD::CTTZ , MVT::i8 , Promote);
327 AddPromotedToType (ISD::CTTZ , MVT::i8 , MVT::i32);
328 setOperationAction(ISD::CTTZ_ZERO_UNDEF , MVT::i8 , Promote);
329 AddPromotedToType (ISD::CTTZ_ZERO_UNDEF , MVT::i8 , MVT::i32);
330 if (Subtarget->hasBMI()) {
331 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::i16 , Expand);
332 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::i32 , Expand);
333 if (Subtarget->is64Bit())
334 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::i64, Expand);
336 setOperationAction(ISD::CTTZ , MVT::i16 , Custom);
337 setOperationAction(ISD::CTTZ , MVT::i32 , Custom);
338 if (Subtarget->is64Bit())
339 setOperationAction(ISD::CTTZ , MVT::i64 , Custom);
342 if (Subtarget->hasLZCNT()) {
343 // When promoting the i8 variants, force them to i32 for a shorter
345 setOperationAction(ISD::CTLZ , MVT::i8 , Promote);
346 AddPromotedToType (ISD::CTLZ , MVT::i8 , MVT::i32);
347 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i8 , Promote);
348 AddPromotedToType (ISD::CTLZ_ZERO_UNDEF, MVT::i8 , MVT::i32);
349 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i16 , Expand);
350 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i32 , Expand);
351 if (Subtarget->is64Bit())
352 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i64, Expand);
354 setOperationAction(ISD::CTLZ , MVT::i8 , Custom);
355 setOperationAction(ISD::CTLZ , MVT::i16 , Custom);
356 setOperationAction(ISD::CTLZ , MVT::i32 , Custom);
357 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i8 , Custom);
358 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i16 , Custom);
359 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i32 , Custom);
360 if (Subtarget->is64Bit()) {
361 setOperationAction(ISD::CTLZ , MVT::i64 , Custom);
362 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::i64, Custom);
366 // Special handling for half-precision floating point conversions.
367 // If we don't have F16C support, then lower half float conversions
368 // into library calls.
369 if (Subtarget->useSoftFloat() || !Subtarget->hasF16C()) {
370 setOperationAction(ISD::FP16_TO_FP, MVT::f32, Expand);
371 setOperationAction(ISD::FP_TO_FP16, MVT::f32, Expand);
374 // There's never any support for operations beyond MVT::f32.
375 setOperationAction(ISD::FP16_TO_FP, MVT::f64, Expand);
376 setOperationAction(ISD::FP16_TO_FP, MVT::f80, Expand);
377 setOperationAction(ISD::FP_TO_FP16, MVT::f64, Expand);
378 setOperationAction(ISD::FP_TO_FP16, MVT::f80, Expand);
380 setLoadExtAction(ISD::EXTLOAD, MVT::f32, MVT::f16, Expand);
381 setLoadExtAction(ISD::EXTLOAD, MVT::f64, MVT::f16, Expand);
382 setLoadExtAction(ISD::EXTLOAD, MVT::f80, MVT::f16, Expand);
383 setTruncStoreAction(MVT::f32, MVT::f16, Expand);
384 setTruncStoreAction(MVT::f64, MVT::f16, Expand);
385 setTruncStoreAction(MVT::f80, MVT::f16, Expand);
387 if (Subtarget->hasPOPCNT()) {
388 setOperationAction(ISD::CTPOP , MVT::i8 , Promote);
390 setOperationAction(ISD::CTPOP , MVT::i8 , Expand);
391 setOperationAction(ISD::CTPOP , MVT::i16 , Expand);
392 setOperationAction(ISD::CTPOP , MVT::i32 , Expand);
393 if (Subtarget->is64Bit())
394 setOperationAction(ISD::CTPOP , MVT::i64 , Expand);
397 setOperationAction(ISD::READCYCLECOUNTER , MVT::i64 , Custom);
399 if (!Subtarget->hasMOVBE())
400 setOperationAction(ISD::BSWAP , MVT::i16 , Expand);
402 // These should be promoted to a larger select which is supported.
403 setOperationAction(ISD::SELECT , MVT::i1 , Promote);
404 // X86 wants to expand cmov itself.
405 setOperationAction(ISD::SELECT , MVT::i8 , Custom);
406 setOperationAction(ISD::SELECT , MVT::i16 , Custom);
407 setOperationAction(ISD::SELECT , MVT::i32 , Custom);
408 setOperationAction(ISD::SELECT , MVT::f32 , Custom);
409 setOperationAction(ISD::SELECT , MVT::f64 , Custom);
410 setOperationAction(ISD::SELECT , MVT::f80 , Custom);
411 setOperationAction(ISD::SETCC , MVT::i8 , Custom);
412 setOperationAction(ISD::SETCC , MVT::i16 , Custom);
413 setOperationAction(ISD::SETCC , MVT::i32 , Custom);
414 setOperationAction(ISD::SETCC , MVT::f32 , Custom);
415 setOperationAction(ISD::SETCC , MVT::f64 , Custom);
416 setOperationAction(ISD::SETCC , MVT::f80 , Custom);
417 if (Subtarget->is64Bit()) {
418 setOperationAction(ISD::SELECT , MVT::i64 , Custom);
419 setOperationAction(ISD::SETCC , MVT::i64 , Custom);
421 setOperationAction(ISD::EH_RETURN , MVT::Other, Custom);
422 // NOTE: EH_SJLJ_SETJMP/_LONGJMP supported here is NOT intended to support
423 // SjLj exception handling but a light-weight setjmp/longjmp replacement to
424 // support continuation, user-level threading, and etc.. As a result, no
425 // other SjLj exception interfaces are implemented and please don't build
426 // your own exception handling based on them.
427 // LLVM/Clang supports zero-cost DWARF exception handling.
428 setOperationAction(ISD::EH_SJLJ_SETJMP, MVT::i32, Custom);
429 setOperationAction(ISD::EH_SJLJ_LONGJMP, MVT::Other, Custom);
432 setOperationAction(ISD::ConstantPool , MVT::i32 , Custom);
433 setOperationAction(ISD::JumpTable , MVT::i32 , Custom);
434 setOperationAction(ISD::GlobalAddress , MVT::i32 , Custom);
435 setOperationAction(ISD::GlobalTLSAddress, MVT::i32 , Custom);
436 if (Subtarget->is64Bit())
437 setOperationAction(ISD::GlobalTLSAddress, MVT::i64, Custom);
438 setOperationAction(ISD::ExternalSymbol , MVT::i32 , Custom);
439 setOperationAction(ISD::BlockAddress , MVT::i32 , Custom);
440 if (Subtarget->is64Bit()) {
441 setOperationAction(ISD::ConstantPool , MVT::i64 , Custom);
442 setOperationAction(ISD::JumpTable , MVT::i64 , Custom);
443 setOperationAction(ISD::GlobalAddress , MVT::i64 , Custom);
444 setOperationAction(ISD::ExternalSymbol, MVT::i64 , Custom);
445 setOperationAction(ISD::BlockAddress , MVT::i64 , Custom);
447 // 64-bit addm sub, shl, sra, srl (iff 32-bit x86)
448 setOperationAction(ISD::SHL_PARTS , MVT::i32 , Custom);
449 setOperationAction(ISD::SRA_PARTS , MVT::i32 , Custom);
450 setOperationAction(ISD::SRL_PARTS , MVT::i32 , Custom);
451 if (Subtarget->is64Bit()) {
452 setOperationAction(ISD::SHL_PARTS , MVT::i64 , Custom);
453 setOperationAction(ISD::SRA_PARTS , MVT::i64 , Custom);
454 setOperationAction(ISD::SRL_PARTS , MVT::i64 , Custom);
457 if (Subtarget->hasSSE1())
458 setOperationAction(ISD::PREFETCH , MVT::Other, Legal);
460 setOperationAction(ISD::ATOMIC_FENCE , MVT::Other, Custom);
462 // Expand certain atomics
463 for (auto VT : { MVT::i8, MVT::i16, MVT::i32, MVT::i64 }) {
464 setOperationAction(ISD::ATOMIC_CMP_SWAP_WITH_SUCCESS, VT, Custom);
465 setOperationAction(ISD::ATOMIC_LOAD_SUB, VT, Custom);
466 setOperationAction(ISD::ATOMIC_STORE, VT, Custom);
469 if (Subtarget->hasCmpxchg16b()) {
470 setOperationAction(ISD::ATOMIC_CMP_SWAP_WITH_SUCCESS, MVT::i128, Custom);
473 // FIXME - use subtarget debug flags
474 if (!Subtarget->isTargetDarwin() && !Subtarget->isTargetELF() &&
475 !Subtarget->isTargetCygMing() && !Subtarget->isTargetWin64()) {
476 setOperationAction(ISD::EH_LABEL, MVT::Other, Expand);
479 if (Subtarget->isTarget64BitLP64()) {
480 setExceptionPointerRegister(X86::RAX);
481 setExceptionSelectorRegister(X86::RDX);
483 setExceptionPointerRegister(X86::EAX);
484 setExceptionSelectorRegister(X86::EDX);
486 setOperationAction(ISD::FRAME_TO_ARGS_OFFSET, MVT::i32, Custom);
487 setOperationAction(ISD::FRAME_TO_ARGS_OFFSET, MVT::i64, Custom);
489 setOperationAction(ISD::INIT_TRAMPOLINE, MVT::Other, Custom);
490 setOperationAction(ISD::ADJUST_TRAMPOLINE, MVT::Other, Custom);
492 setOperationAction(ISD::TRAP, MVT::Other, Legal);
493 setOperationAction(ISD::DEBUGTRAP, MVT::Other, Legal);
495 // VASTART needs to be custom lowered to use the VarArgsFrameIndex
496 setOperationAction(ISD::VASTART , MVT::Other, Custom);
497 setOperationAction(ISD::VAEND , MVT::Other, Expand);
498 if (Subtarget->is64Bit()) {
499 setOperationAction(ISD::VAARG , MVT::Other, Custom);
500 setOperationAction(ISD::VACOPY , MVT::Other, Custom);
502 // TargetInfo::CharPtrBuiltinVaList
503 setOperationAction(ISD::VAARG , MVT::Other, Expand);
504 setOperationAction(ISD::VACOPY , MVT::Other, Expand);
507 setOperationAction(ISD::STACKSAVE, MVT::Other, Expand);
508 setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand);
510 setOperationAction(ISD::DYNAMIC_STACKALLOC, PtrVT, Custom);
512 // GC_TRANSITION_START and GC_TRANSITION_END need custom lowering.
513 setOperationAction(ISD::GC_TRANSITION_START, MVT::Other, Custom);
514 setOperationAction(ISD::GC_TRANSITION_END, MVT::Other, Custom);
516 if (!Subtarget->useSoftFloat() && X86ScalarSSEf64) {
517 // f32 and f64 use SSE.
518 // Set up the FP register classes.
519 addRegisterClass(MVT::f32, &X86::FR32RegClass);
520 addRegisterClass(MVT::f64, &X86::FR64RegClass);
522 // Use ANDPD to simulate FABS.
523 setOperationAction(ISD::FABS , MVT::f64, Custom);
524 setOperationAction(ISD::FABS , MVT::f32, Custom);
526 // Use XORP to simulate FNEG.
527 setOperationAction(ISD::FNEG , MVT::f64, Custom);
528 setOperationAction(ISD::FNEG , MVT::f32, Custom);
530 // Use ANDPD and ORPD to simulate FCOPYSIGN.
531 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Custom);
532 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
534 // Lower this to FGETSIGNx86 plus an AND.
535 setOperationAction(ISD::FGETSIGN, MVT::i64, Custom);
536 setOperationAction(ISD::FGETSIGN, MVT::i32, Custom);
538 // We don't support sin/cos/fmod
539 setOperationAction(ISD::FSIN , MVT::f64, Expand);
540 setOperationAction(ISD::FCOS , MVT::f64, Expand);
541 setOperationAction(ISD::FSINCOS, MVT::f64, Expand);
542 setOperationAction(ISD::FSIN , MVT::f32, Expand);
543 setOperationAction(ISD::FCOS , MVT::f32, Expand);
544 setOperationAction(ISD::FSINCOS, MVT::f32, Expand);
546 // Expand FP immediates into loads from the stack, except for the special
548 addLegalFPImmediate(APFloat(+0.0)); // xorpd
549 addLegalFPImmediate(APFloat(+0.0f)); // xorps
550 } else if (!Subtarget->useSoftFloat() && X86ScalarSSEf32) {
551 // Use SSE for f32, x87 for f64.
552 // Set up the FP register classes.
553 addRegisterClass(MVT::f32, &X86::FR32RegClass);
554 addRegisterClass(MVT::f64, &X86::RFP64RegClass);
556 // Use ANDPS to simulate FABS.
557 setOperationAction(ISD::FABS , MVT::f32, Custom);
559 // Use XORP to simulate FNEG.
560 setOperationAction(ISD::FNEG , MVT::f32, Custom);
562 setOperationAction(ISD::UNDEF, MVT::f64, Expand);
564 // Use ANDPS and ORPS to simulate FCOPYSIGN.
565 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
566 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Custom);
568 // We don't support sin/cos/fmod
569 setOperationAction(ISD::FSIN , MVT::f32, Expand);
570 setOperationAction(ISD::FCOS , MVT::f32, Expand);
571 setOperationAction(ISD::FSINCOS, MVT::f32, Expand);
573 // Special cases we handle for FP constants.
574 addLegalFPImmediate(APFloat(+0.0f)); // xorps
575 addLegalFPImmediate(APFloat(+0.0)); // FLD0
576 addLegalFPImmediate(APFloat(+1.0)); // FLD1
577 addLegalFPImmediate(APFloat(-0.0)); // FLD0/FCHS
578 addLegalFPImmediate(APFloat(-1.0)); // FLD1/FCHS
580 if (!TM.Options.UnsafeFPMath) {
581 setOperationAction(ISD::FSIN , MVT::f64, Expand);
582 setOperationAction(ISD::FCOS , MVT::f64, Expand);
583 setOperationAction(ISD::FSINCOS, MVT::f64, Expand);
585 } else if (!Subtarget->useSoftFloat()) {
586 // f32 and f64 in x87.
587 // Set up the FP register classes.
588 addRegisterClass(MVT::f64, &X86::RFP64RegClass);
589 addRegisterClass(MVT::f32, &X86::RFP32RegClass);
591 setOperationAction(ISD::UNDEF, MVT::f64, Expand);
592 setOperationAction(ISD::UNDEF, MVT::f32, Expand);
593 setOperationAction(ISD::FCOPYSIGN, MVT::f64, Expand);
594 setOperationAction(ISD::FCOPYSIGN, MVT::f32, Expand);
596 if (!TM.Options.UnsafeFPMath) {
597 setOperationAction(ISD::FSIN , MVT::f64, Expand);
598 setOperationAction(ISD::FSIN , MVT::f32, Expand);
599 setOperationAction(ISD::FCOS , MVT::f64, Expand);
600 setOperationAction(ISD::FCOS , MVT::f32, Expand);
601 setOperationAction(ISD::FSINCOS, MVT::f64, Expand);
602 setOperationAction(ISD::FSINCOS, MVT::f32, Expand);
604 addLegalFPImmediate(APFloat(+0.0)); // FLD0
605 addLegalFPImmediate(APFloat(+1.0)); // FLD1
606 addLegalFPImmediate(APFloat(-0.0)); // FLD0/FCHS
607 addLegalFPImmediate(APFloat(-1.0)); // FLD1/FCHS
608 addLegalFPImmediate(APFloat(+0.0f)); // FLD0
609 addLegalFPImmediate(APFloat(+1.0f)); // FLD1
610 addLegalFPImmediate(APFloat(-0.0f)); // FLD0/FCHS
611 addLegalFPImmediate(APFloat(-1.0f)); // FLD1/FCHS
614 // We don't support FMA.
615 setOperationAction(ISD::FMA, MVT::f64, Expand);
616 setOperationAction(ISD::FMA, MVT::f32, Expand);
618 // Long double always uses X87.
619 if (!Subtarget->useSoftFloat()) {
620 addRegisterClass(MVT::f80, &X86::RFP80RegClass);
621 setOperationAction(ISD::UNDEF, MVT::f80, Expand);
622 setOperationAction(ISD::FCOPYSIGN, MVT::f80, Expand);
624 APFloat TmpFlt = APFloat::getZero(APFloat::x87DoubleExtended);
625 addLegalFPImmediate(TmpFlt); // FLD0
627 addLegalFPImmediate(TmpFlt); // FLD0/FCHS
630 APFloat TmpFlt2(+1.0);
631 TmpFlt2.convert(APFloat::x87DoubleExtended, APFloat::rmNearestTiesToEven,
633 addLegalFPImmediate(TmpFlt2); // FLD1
634 TmpFlt2.changeSign();
635 addLegalFPImmediate(TmpFlt2); // FLD1/FCHS
638 if (!TM.Options.UnsafeFPMath) {
639 setOperationAction(ISD::FSIN , MVT::f80, Expand);
640 setOperationAction(ISD::FCOS , MVT::f80, Expand);
641 setOperationAction(ISD::FSINCOS, MVT::f80, Expand);
644 setOperationAction(ISD::FFLOOR, MVT::f80, Expand);
645 setOperationAction(ISD::FCEIL, MVT::f80, Expand);
646 setOperationAction(ISD::FTRUNC, MVT::f80, Expand);
647 setOperationAction(ISD::FRINT, MVT::f80, Expand);
648 setOperationAction(ISD::FNEARBYINT, MVT::f80, Expand);
649 setOperationAction(ISD::FMA, MVT::f80, Expand);
652 // Always use a library call for pow.
653 setOperationAction(ISD::FPOW , MVT::f32 , Expand);
654 setOperationAction(ISD::FPOW , MVT::f64 , Expand);
655 setOperationAction(ISD::FPOW , MVT::f80 , Expand);
657 setOperationAction(ISD::FLOG, MVT::f80, Expand);
658 setOperationAction(ISD::FLOG2, MVT::f80, Expand);
659 setOperationAction(ISD::FLOG10, MVT::f80, Expand);
660 setOperationAction(ISD::FEXP, MVT::f80, Expand);
661 setOperationAction(ISD::FEXP2, MVT::f80, Expand);
662 setOperationAction(ISD::FMINNUM, MVT::f80, Expand);
663 setOperationAction(ISD::FMAXNUM, MVT::f80, Expand);
665 // First set operation action for all vector types to either promote
666 // (for widening) or expand (for scalarization). Then we will selectively
667 // turn on ones that can be effectively codegen'd.
668 for (MVT VT : MVT::vector_valuetypes()) {
669 setOperationAction(ISD::ADD , VT, Expand);
670 setOperationAction(ISD::SUB , VT, Expand);
671 setOperationAction(ISD::FADD, VT, Expand);
672 setOperationAction(ISD::FNEG, VT, Expand);
673 setOperationAction(ISD::FSUB, VT, Expand);
674 setOperationAction(ISD::MUL , VT, Expand);
675 setOperationAction(ISD::FMUL, VT, Expand);
676 setOperationAction(ISD::SDIV, VT, Expand);
677 setOperationAction(ISD::UDIV, VT, Expand);
678 setOperationAction(ISD::FDIV, VT, Expand);
679 setOperationAction(ISD::SREM, VT, Expand);
680 setOperationAction(ISD::UREM, VT, Expand);
681 setOperationAction(ISD::LOAD, VT, Expand);
682 setOperationAction(ISD::VECTOR_SHUFFLE, VT, Expand);
683 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT,Expand);
684 setOperationAction(ISD::INSERT_VECTOR_ELT, VT, Expand);
685 setOperationAction(ISD::EXTRACT_SUBVECTOR, VT,Expand);
686 setOperationAction(ISD::INSERT_SUBVECTOR, VT,Expand);
687 setOperationAction(ISD::FABS, VT, Expand);
688 setOperationAction(ISD::FSIN, VT, Expand);
689 setOperationAction(ISD::FSINCOS, VT, Expand);
690 setOperationAction(ISD::FCOS, VT, Expand);
691 setOperationAction(ISD::FSINCOS, VT, Expand);
692 setOperationAction(ISD::FREM, VT, Expand);
693 setOperationAction(ISD::FMA, VT, Expand);
694 setOperationAction(ISD::FPOWI, VT, Expand);
695 setOperationAction(ISD::FSQRT, VT, Expand);
696 setOperationAction(ISD::FCOPYSIGN, VT, Expand);
697 setOperationAction(ISD::FFLOOR, VT, Expand);
698 setOperationAction(ISD::FCEIL, VT, Expand);
699 setOperationAction(ISD::FTRUNC, VT, Expand);
700 setOperationAction(ISD::FRINT, VT, Expand);
701 setOperationAction(ISD::FNEARBYINT, VT, Expand);
702 setOperationAction(ISD::SMUL_LOHI, VT, Expand);
703 setOperationAction(ISD::MULHS, VT, Expand);
704 setOperationAction(ISD::UMUL_LOHI, VT, Expand);
705 setOperationAction(ISD::MULHU, VT, Expand);
706 setOperationAction(ISD::SDIVREM, VT, Expand);
707 setOperationAction(ISD::UDIVREM, VT, Expand);
708 setOperationAction(ISD::FPOW, VT, Expand);
709 setOperationAction(ISD::CTPOP, VT, Expand);
710 setOperationAction(ISD::CTTZ, VT, Expand);
711 setOperationAction(ISD::CTTZ_ZERO_UNDEF, VT, Expand);
712 setOperationAction(ISD::CTLZ, VT, Expand);
713 setOperationAction(ISD::CTLZ_ZERO_UNDEF, VT, Expand);
714 setOperationAction(ISD::SHL, VT, Expand);
715 setOperationAction(ISD::SRA, VT, Expand);
716 setOperationAction(ISD::SRL, VT, Expand);
717 setOperationAction(ISD::ROTL, VT, Expand);
718 setOperationAction(ISD::ROTR, VT, Expand);
719 setOperationAction(ISD::BSWAP, VT, Expand);
720 setOperationAction(ISD::SETCC, VT, Expand);
721 setOperationAction(ISD::FLOG, VT, Expand);
722 setOperationAction(ISD::FLOG2, VT, Expand);
723 setOperationAction(ISD::FLOG10, VT, Expand);
724 setOperationAction(ISD::FEXP, VT, Expand);
725 setOperationAction(ISD::FEXP2, VT, Expand);
726 setOperationAction(ISD::FP_TO_UINT, VT, Expand);
727 setOperationAction(ISD::FP_TO_SINT, VT, Expand);
728 setOperationAction(ISD::UINT_TO_FP, VT, Expand);
729 setOperationAction(ISD::SINT_TO_FP, VT, Expand);
730 setOperationAction(ISD::SIGN_EXTEND_INREG, VT,Expand);
731 setOperationAction(ISD::TRUNCATE, VT, Expand);
732 setOperationAction(ISD::SIGN_EXTEND, VT, Expand);
733 setOperationAction(ISD::ZERO_EXTEND, VT, Expand);
734 setOperationAction(ISD::ANY_EXTEND, VT, Expand);
735 setOperationAction(ISD::VSELECT, VT, Expand);
736 setOperationAction(ISD::SELECT_CC, VT, Expand);
737 for (MVT InnerVT : MVT::vector_valuetypes()) {
738 setTruncStoreAction(InnerVT, VT, Expand);
740 setLoadExtAction(ISD::SEXTLOAD, InnerVT, VT, Expand);
741 setLoadExtAction(ISD::ZEXTLOAD, InnerVT, VT, Expand);
743 // N.b. ISD::EXTLOAD legality is basically ignored except for i1-like
744 // types, we have to deal with them whether we ask for Expansion or not.
745 // Setting Expand causes its own optimisation problems though, so leave
747 if (VT.getVectorElementType() == MVT::i1)
748 setLoadExtAction(ISD::EXTLOAD, InnerVT, VT, Expand);
750 // EXTLOAD for MVT::f16 vectors is not legal because f16 vectors are
751 // split/scalarized right now.
752 if (VT.getVectorElementType() == MVT::f16)
753 setLoadExtAction(ISD::EXTLOAD, InnerVT, VT, Expand);
757 // FIXME: In order to prevent SSE instructions being expanded to MMX ones
758 // with -msoft-float, disable use of MMX as well.
759 if (!Subtarget->useSoftFloat() && Subtarget->hasMMX()) {
760 addRegisterClass(MVT::x86mmx, &X86::VR64RegClass);
761 // No operations on x86mmx supported, everything uses intrinsics.
764 // MMX-sized vectors (other than x86mmx) are expected to be expanded
765 // into smaller operations.
766 for (MVT MMXTy : {MVT::v8i8, MVT::v4i16, MVT::v2i32, MVT::v1i64}) {
767 setOperationAction(ISD::MULHS, MMXTy, Expand);
768 setOperationAction(ISD::AND, MMXTy, Expand);
769 setOperationAction(ISD::OR, MMXTy, Expand);
770 setOperationAction(ISD::XOR, MMXTy, Expand);
771 setOperationAction(ISD::SCALAR_TO_VECTOR, MMXTy, Expand);
772 setOperationAction(ISD::SELECT, MMXTy, Expand);
773 setOperationAction(ISD::BITCAST, MMXTy, Expand);
775 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v1i64, Expand);
777 if (!Subtarget->useSoftFloat() && Subtarget->hasSSE1()) {
778 addRegisterClass(MVT::v4f32, &X86::VR128RegClass);
780 setOperationAction(ISD::FADD, MVT::v4f32, Legal);
781 setOperationAction(ISD::FSUB, MVT::v4f32, Legal);
782 setOperationAction(ISD::FMUL, MVT::v4f32, Legal);
783 setOperationAction(ISD::FDIV, MVT::v4f32, Legal);
784 setOperationAction(ISD::FSQRT, MVT::v4f32, Legal);
785 setOperationAction(ISD::FNEG, MVT::v4f32, Custom);
786 setOperationAction(ISD::FABS, MVT::v4f32, Custom);
787 setOperationAction(ISD::LOAD, MVT::v4f32, Legal);
788 setOperationAction(ISD::BUILD_VECTOR, MVT::v4f32, Custom);
789 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4f32, Custom);
790 setOperationAction(ISD::VSELECT, MVT::v4f32, Custom);
791 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f32, Custom);
792 setOperationAction(ISD::SELECT, MVT::v4f32, Custom);
793 setOperationAction(ISD::UINT_TO_FP, MVT::v4i32, Custom);
796 if (!Subtarget->useSoftFloat() && Subtarget->hasSSE2()) {
797 addRegisterClass(MVT::v2f64, &X86::VR128RegClass);
799 // FIXME: Unfortunately, -soft-float and -no-implicit-float mean XMM
800 // registers cannot be used even for integer operations.
801 addRegisterClass(MVT::v16i8, &X86::VR128RegClass);
802 addRegisterClass(MVT::v8i16, &X86::VR128RegClass);
803 addRegisterClass(MVT::v4i32, &X86::VR128RegClass);
804 addRegisterClass(MVT::v2i64, &X86::VR128RegClass);
806 setOperationAction(ISD::ADD, MVT::v16i8, Legal);
807 setOperationAction(ISD::ADD, MVT::v8i16, Legal);
808 setOperationAction(ISD::ADD, MVT::v4i32, Legal);
809 setOperationAction(ISD::ADD, MVT::v2i64, Legal);
810 setOperationAction(ISD::MUL, MVT::v16i8, Custom);
811 setOperationAction(ISD::MUL, MVT::v4i32, Custom);
812 setOperationAction(ISD::MUL, MVT::v2i64, Custom);
813 setOperationAction(ISD::UMUL_LOHI, MVT::v4i32, Custom);
814 setOperationAction(ISD::SMUL_LOHI, MVT::v4i32, Custom);
815 setOperationAction(ISD::MULHU, MVT::v8i16, Legal);
816 setOperationAction(ISD::MULHS, MVT::v8i16, Legal);
817 setOperationAction(ISD::SUB, MVT::v16i8, Legal);
818 setOperationAction(ISD::SUB, MVT::v8i16, Legal);
819 setOperationAction(ISD::SUB, MVT::v4i32, Legal);
820 setOperationAction(ISD::SUB, MVT::v2i64, Legal);
821 setOperationAction(ISD::MUL, MVT::v8i16, Legal);
822 setOperationAction(ISD::FADD, MVT::v2f64, Legal);
823 setOperationAction(ISD::FSUB, MVT::v2f64, Legal);
824 setOperationAction(ISD::FMUL, MVT::v2f64, Legal);
825 setOperationAction(ISD::FDIV, MVT::v2f64, Legal);
826 setOperationAction(ISD::FSQRT, MVT::v2f64, Legal);
827 setOperationAction(ISD::FNEG, MVT::v2f64, Custom);
828 setOperationAction(ISD::FABS, MVT::v2f64, Custom);
830 setOperationAction(ISD::SMAX, MVT::v8i16, Legal);
831 setOperationAction(ISD::UMAX, MVT::v16i8, Legal);
832 setOperationAction(ISD::SMIN, MVT::v8i16, Legal);
833 setOperationAction(ISD::UMIN, MVT::v16i8, Legal);
835 setOperationAction(ISD::SETCC, MVT::v2i64, Custom);
836 setOperationAction(ISD::SETCC, MVT::v16i8, Custom);
837 setOperationAction(ISD::SETCC, MVT::v8i16, Custom);
838 setOperationAction(ISD::SETCC, MVT::v4i32, Custom);
840 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v16i8, Custom);
841 setOperationAction(ISD::SCALAR_TO_VECTOR, MVT::v8i16, Custom);
842 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i16, Custom);
843 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i32, Custom);
844 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f32, Custom);
846 setOperationAction(ISD::CTPOP, MVT::v16i8, Custom);
847 setOperationAction(ISD::CTPOP, MVT::v8i16, Custom);
848 setOperationAction(ISD::CTPOP, MVT::v4i32, Custom);
849 setOperationAction(ISD::CTPOP, MVT::v2i64, Custom);
851 setOperationAction(ISD::CTTZ, MVT::v16i8, Custom);
852 setOperationAction(ISD::CTTZ, MVT::v8i16, Custom);
853 setOperationAction(ISD::CTTZ, MVT::v4i32, Custom);
854 // ISD::CTTZ v2i64 - scalarization is faster.
855 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::v16i8, Custom);
856 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::v8i16, Custom);
857 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::v4i32, Custom);
858 // ISD::CTTZ_ZERO_UNDEF v2i64 - scalarization is faster.
860 // Custom lower build_vector, vector_shuffle, and extract_vector_elt.
861 for (int i = MVT::v16i8; i != MVT::v2i64; ++i) {
862 MVT VT = (MVT::SimpleValueType)i;
863 // Do not attempt to custom lower non-power-of-2 vectors
864 if (!isPowerOf2_32(VT.getVectorNumElements()))
866 // Do not attempt to custom lower non-128-bit vectors
867 if (!VT.is128BitVector())
869 setOperationAction(ISD::BUILD_VECTOR, VT, Custom);
870 setOperationAction(ISD::VECTOR_SHUFFLE, VT, Custom);
871 setOperationAction(ISD::VSELECT, VT, Custom);
872 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Custom);
875 // We support custom legalizing of sext and anyext loads for specific
876 // memory vector types which we can load as a scalar (or sequence of
877 // scalars) and extend in-register to a legal 128-bit vector type. For sext
878 // loads these must work with a single scalar load.
879 for (MVT VT : MVT::integer_vector_valuetypes()) {
880 setLoadExtAction(ISD::SEXTLOAD, VT, MVT::v4i8, Custom);
881 setLoadExtAction(ISD::SEXTLOAD, VT, MVT::v4i16, Custom);
882 setLoadExtAction(ISD::SEXTLOAD, VT, MVT::v8i8, Custom);
883 setLoadExtAction(ISD::EXTLOAD, VT, MVT::v2i8, Custom);
884 setLoadExtAction(ISD::EXTLOAD, VT, MVT::v2i16, Custom);
885 setLoadExtAction(ISD::EXTLOAD, VT, MVT::v2i32, Custom);
886 setLoadExtAction(ISD::EXTLOAD, VT, MVT::v4i8, Custom);
887 setLoadExtAction(ISD::EXTLOAD, VT, MVT::v4i16, Custom);
888 setLoadExtAction(ISD::EXTLOAD, VT, MVT::v8i8, Custom);
891 setOperationAction(ISD::BUILD_VECTOR, MVT::v2f64, Custom);
892 setOperationAction(ISD::BUILD_VECTOR, MVT::v2i64, Custom);
893 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2f64, Custom);
894 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2i64, Custom);
895 setOperationAction(ISD::VSELECT, MVT::v2f64, Custom);
896 setOperationAction(ISD::VSELECT, MVT::v2i64, Custom);
897 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2f64, Custom);
898 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2f64, Custom);
900 if (Subtarget->is64Bit()) {
901 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2i64, Custom);
902 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2i64, Custom);
905 // Promote v16i8, v8i16, v4i32 load, select, and, or, xor to v2i64.
906 for (int i = MVT::v16i8; i != MVT::v2i64; ++i) {
907 MVT VT = (MVT::SimpleValueType)i;
909 // Do not attempt to promote non-128-bit vectors
910 if (!VT.is128BitVector())
913 setOperationAction(ISD::AND, VT, Promote);
914 AddPromotedToType (ISD::AND, VT, MVT::v2i64);
915 setOperationAction(ISD::OR, VT, Promote);
916 AddPromotedToType (ISD::OR, VT, MVT::v2i64);
917 setOperationAction(ISD::XOR, VT, Promote);
918 AddPromotedToType (ISD::XOR, VT, MVT::v2i64);
919 setOperationAction(ISD::LOAD, VT, Promote);
920 AddPromotedToType (ISD::LOAD, VT, MVT::v2i64);
921 setOperationAction(ISD::SELECT, VT, Promote);
922 AddPromotedToType (ISD::SELECT, VT, MVT::v2i64);
925 // Custom lower v2i64 and v2f64 selects.
926 setOperationAction(ISD::LOAD, MVT::v2f64, Legal);
927 setOperationAction(ISD::LOAD, MVT::v2i64, Legal);
928 setOperationAction(ISD::SELECT, MVT::v2f64, Custom);
929 setOperationAction(ISD::SELECT, MVT::v2i64, Custom);
931 setOperationAction(ISD::FP_TO_SINT, MVT::v4i32, Legal);
932 setOperationAction(ISD::SINT_TO_FP, MVT::v4i32, Legal);
934 setOperationAction(ISD::SINT_TO_FP, MVT::v2i32, Custom);
936 setOperationAction(ISD::UINT_TO_FP, MVT::v4i8, Custom);
937 setOperationAction(ISD::UINT_TO_FP, MVT::v4i16, Custom);
938 // As there is no 64-bit GPR available, we need build a special custom
939 // sequence to convert from v2i32 to v2f32.
940 if (!Subtarget->is64Bit())
941 setOperationAction(ISD::UINT_TO_FP, MVT::v2f32, Custom);
943 setOperationAction(ISD::FP_EXTEND, MVT::v2f32, Custom);
944 setOperationAction(ISD::FP_ROUND, MVT::v2f32, Custom);
946 for (MVT VT : MVT::fp_vector_valuetypes())
947 setLoadExtAction(ISD::EXTLOAD, VT, MVT::v2f32, Legal);
949 setOperationAction(ISD::BITCAST, MVT::v2i32, Custom);
950 setOperationAction(ISD::BITCAST, MVT::v4i16, Custom);
951 setOperationAction(ISD::BITCAST, MVT::v8i8, Custom);
954 if (!Subtarget->useSoftFloat() && Subtarget->hasSSE41()) {
955 for (MVT RoundedTy : {MVT::f32, MVT::f64, MVT::v4f32, MVT::v2f64}) {
956 setOperationAction(ISD::FFLOOR, RoundedTy, Legal);
957 setOperationAction(ISD::FCEIL, RoundedTy, Legal);
958 setOperationAction(ISD::FTRUNC, RoundedTy, Legal);
959 setOperationAction(ISD::FRINT, RoundedTy, Legal);
960 setOperationAction(ISD::FNEARBYINT, RoundedTy, Legal);
963 setOperationAction(ISD::SMAX, MVT::v16i8, Legal);
964 setOperationAction(ISD::SMAX, MVT::v4i32, Legal);
965 setOperationAction(ISD::UMAX, MVT::v8i16, Legal);
966 setOperationAction(ISD::UMAX, MVT::v4i32, Legal);
967 setOperationAction(ISD::SMIN, MVT::v16i8, Legal);
968 setOperationAction(ISD::SMIN, MVT::v4i32, Legal);
969 setOperationAction(ISD::UMIN, MVT::v8i16, Legal);
970 setOperationAction(ISD::UMIN, MVT::v4i32, Legal);
972 // FIXME: Do we need to handle scalar-to-vector here?
973 setOperationAction(ISD::MUL, MVT::v4i32, Legal);
975 // We directly match byte blends in the backend as they match the VSELECT
977 setOperationAction(ISD::VSELECT, MVT::v16i8, Legal);
979 // SSE41 brings specific instructions for doing vector sign extend even in
980 // cases where we don't have SRA.
981 for (MVT VT : MVT::integer_vector_valuetypes()) {
982 setLoadExtAction(ISD::SEXTLOAD, VT, MVT::v2i8, Custom);
983 setLoadExtAction(ISD::SEXTLOAD, VT, MVT::v2i16, Custom);
984 setLoadExtAction(ISD::SEXTLOAD, VT, MVT::v2i32, Custom);
987 // SSE41 also has vector sign/zero extending loads, PMOV[SZ]X
988 setLoadExtAction(ISD::SEXTLOAD, MVT::v8i16, MVT::v8i8, Legal);
989 setLoadExtAction(ISD::SEXTLOAD, MVT::v4i32, MVT::v4i8, Legal);
990 setLoadExtAction(ISD::SEXTLOAD, MVT::v2i64, MVT::v2i8, Legal);
991 setLoadExtAction(ISD::SEXTLOAD, MVT::v4i32, MVT::v4i16, Legal);
992 setLoadExtAction(ISD::SEXTLOAD, MVT::v2i64, MVT::v2i16, Legal);
993 setLoadExtAction(ISD::SEXTLOAD, MVT::v2i64, MVT::v2i32, Legal);
995 setLoadExtAction(ISD::ZEXTLOAD, MVT::v8i16, MVT::v8i8, Legal);
996 setLoadExtAction(ISD::ZEXTLOAD, MVT::v4i32, MVT::v4i8, Legal);
997 setLoadExtAction(ISD::ZEXTLOAD, MVT::v2i64, MVT::v2i8, Legal);
998 setLoadExtAction(ISD::ZEXTLOAD, MVT::v4i32, MVT::v4i16, Legal);
999 setLoadExtAction(ISD::ZEXTLOAD, MVT::v2i64, MVT::v2i16, Legal);
1000 setLoadExtAction(ISD::ZEXTLOAD, MVT::v2i64, MVT::v2i32, Legal);
1002 // i8 and i16 vectors are custom because the source register and source
1003 // source memory operand types are not the same width. f32 vectors are
1004 // custom since the immediate controlling the insert encodes additional
1006 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v16i8, Custom);
1007 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i16, Custom);
1008 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i32, Custom);
1009 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f32, Custom);
1011 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v16i8, Custom);
1012 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v8i16, Custom);
1013 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4i32, Custom);
1014 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f32, Custom);
1016 // FIXME: these should be Legal, but that's only for the case where
1017 // the index is constant. For now custom expand to deal with that.
1018 if (Subtarget->is64Bit()) {
1019 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2i64, Custom);
1020 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2i64, Custom);
1024 if (Subtarget->hasSSE2()) {
1025 setOperationAction(ISD::SIGN_EXTEND_VECTOR_INREG, MVT::v2i64, Custom);
1026 setOperationAction(ISD::SIGN_EXTEND_VECTOR_INREG, MVT::v4i32, Custom);
1027 setOperationAction(ISD::SIGN_EXTEND_VECTOR_INREG, MVT::v8i16, Custom);
1029 setOperationAction(ISD::SRL, MVT::v8i16, Custom);
1030 setOperationAction(ISD::SRL, MVT::v16i8, Custom);
1032 setOperationAction(ISD::SHL, MVT::v8i16, Custom);
1033 setOperationAction(ISD::SHL, MVT::v16i8, Custom);
1035 setOperationAction(ISD::SRA, MVT::v8i16, Custom);
1036 setOperationAction(ISD::SRA, MVT::v16i8, Custom);
1038 // In the customized shift lowering, the legal cases in AVX2 will be
1040 setOperationAction(ISD::SRL, MVT::v2i64, Custom);
1041 setOperationAction(ISD::SRL, MVT::v4i32, Custom);
1043 setOperationAction(ISD::SHL, MVT::v2i64, Custom);
1044 setOperationAction(ISD::SHL, MVT::v4i32, Custom);
1046 setOperationAction(ISD::SRA, MVT::v2i64, Custom);
1047 setOperationAction(ISD::SRA, MVT::v4i32, Custom);
1050 if (Subtarget->hasXOP()) {
1051 setOperationAction(ISD::ROTL, MVT::v16i8, Custom);
1052 setOperationAction(ISD::ROTL, MVT::v8i16, Custom);
1053 setOperationAction(ISD::ROTL, MVT::v4i32, Custom);
1054 setOperationAction(ISD::ROTL, MVT::v2i64, Custom);
1055 setOperationAction(ISD::ROTL, MVT::v32i8, Custom);
1056 setOperationAction(ISD::ROTL, MVT::v16i16, Custom);
1057 setOperationAction(ISD::ROTL, MVT::v8i32, Custom);
1058 setOperationAction(ISD::ROTL, MVT::v4i64, Custom);
1061 if (!Subtarget->useSoftFloat() && Subtarget->hasFp256()) {
1062 addRegisterClass(MVT::v32i8, &X86::VR256RegClass);
1063 addRegisterClass(MVT::v16i16, &X86::VR256RegClass);
1064 addRegisterClass(MVT::v8i32, &X86::VR256RegClass);
1065 addRegisterClass(MVT::v8f32, &X86::VR256RegClass);
1066 addRegisterClass(MVT::v4i64, &X86::VR256RegClass);
1067 addRegisterClass(MVT::v4f64, &X86::VR256RegClass);
1069 setOperationAction(ISD::LOAD, MVT::v8f32, Legal);
1070 setOperationAction(ISD::LOAD, MVT::v4f64, Legal);
1071 setOperationAction(ISD::LOAD, MVT::v4i64, Legal);
1073 setOperationAction(ISD::FADD, MVT::v8f32, Legal);
1074 setOperationAction(ISD::FSUB, MVT::v8f32, Legal);
1075 setOperationAction(ISD::FMUL, MVT::v8f32, Legal);
1076 setOperationAction(ISD::FDIV, MVT::v8f32, Legal);
1077 setOperationAction(ISD::FSQRT, MVT::v8f32, Legal);
1078 setOperationAction(ISD::FFLOOR, MVT::v8f32, Legal);
1079 setOperationAction(ISD::FCEIL, MVT::v8f32, Legal);
1080 setOperationAction(ISD::FTRUNC, MVT::v8f32, Legal);
1081 setOperationAction(ISD::FRINT, MVT::v8f32, Legal);
1082 setOperationAction(ISD::FNEARBYINT, MVT::v8f32, Legal);
1083 setOperationAction(ISD::FNEG, MVT::v8f32, Custom);
1084 setOperationAction(ISD::FABS, MVT::v8f32, Custom);
1086 setOperationAction(ISD::FADD, MVT::v4f64, Legal);
1087 setOperationAction(ISD::FSUB, MVT::v4f64, Legal);
1088 setOperationAction(ISD::FMUL, MVT::v4f64, Legal);
1089 setOperationAction(ISD::FDIV, MVT::v4f64, Legal);
1090 setOperationAction(ISD::FSQRT, MVT::v4f64, Legal);
1091 setOperationAction(ISD::FFLOOR, MVT::v4f64, Legal);
1092 setOperationAction(ISD::FCEIL, MVT::v4f64, Legal);
1093 setOperationAction(ISD::FTRUNC, MVT::v4f64, Legal);
1094 setOperationAction(ISD::FRINT, MVT::v4f64, Legal);
1095 setOperationAction(ISD::FNEARBYINT, MVT::v4f64, Legal);
1096 setOperationAction(ISD::FNEG, MVT::v4f64, Custom);
1097 setOperationAction(ISD::FABS, MVT::v4f64, Custom);
1099 // (fp_to_int:v8i16 (v8f32 ..)) requires the result type to be promoted
1100 // even though v8i16 is a legal type.
1101 setOperationAction(ISD::FP_TO_SINT, MVT::v8i16, Promote);
1102 setOperationAction(ISD::FP_TO_UINT, MVT::v8i16, Promote);
1103 setOperationAction(ISD::FP_TO_SINT, MVT::v8i32, Legal);
1105 setOperationAction(ISD::SINT_TO_FP, MVT::v8i16, Promote);
1106 setOperationAction(ISD::SINT_TO_FP, MVT::v8i32, Legal);
1107 setOperationAction(ISD::FP_ROUND, MVT::v4f32, Legal);
1109 setOperationAction(ISD::UINT_TO_FP, MVT::v8i8, Custom);
1110 setOperationAction(ISD::UINT_TO_FP, MVT::v8i16, Custom);
1112 for (MVT VT : MVT::fp_vector_valuetypes())
1113 setLoadExtAction(ISD::EXTLOAD, VT, MVT::v4f32, Legal);
1115 setOperationAction(ISD::SRL, MVT::v16i16, Custom);
1116 setOperationAction(ISD::SRL, MVT::v32i8, Custom);
1118 setOperationAction(ISD::SHL, MVT::v16i16, Custom);
1119 setOperationAction(ISD::SHL, MVT::v32i8, Custom);
1121 setOperationAction(ISD::SRA, MVT::v16i16, Custom);
1122 setOperationAction(ISD::SRA, MVT::v32i8, Custom);
1124 setOperationAction(ISD::SETCC, MVT::v32i8, Custom);
1125 setOperationAction(ISD::SETCC, MVT::v16i16, Custom);
1126 setOperationAction(ISD::SETCC, MVT::v8i32, Custom);
1127 setOperationAction(ISD::SETCC, MVT::v4i64, Custom);
1129 setOperationAction(ISD::SELECT, MVT::v4f64, Custom);
1130 setOperationAction(ISD::SELECT, MVT::v4i64, Custom);
1131 setOperationAction(ISD::SELECT, MVT::v8f32, Custom);
1133 setOperationAction(ISD::SIGN_EXTEND, MVT::v4i64, Custom);
1134 setOperationAction(ISD::SIGN_EXTEND, MVT::v8i32, Custom);
1135 setOperationAction(ISD::SIGN_EXTEND, MVT::v16i16, Custom);
1136 setOperationAction(ISD::ZERO_EXTEND, MVT::v4i64, Custom);
1137 setOperationAction(ISD::ZERO_EXTEND, MVT::v8i32, Custom);
1138 setOperationAction(ISD::ZERO_EXTEND, MVT::v16i16, Custom);
1139 setOperationAction(ISD::ANY_EXTEND, MVT::v4i64, Custom);
1140 setOperationAction(ISD::ANY_EXTEND, MVT::v8i32, Custom);
1141 setOperationAction(ISD::ANY_EXTEND, MVT::v16i16, Custom);
1142 setOperationAction(ISD::TRUNCATE, MVT::v16i8, Custom);
1143 setOperationAction(ISD::TRUNCATE, MVT::v8i16, Custom);
1144 setOperationAction(ISD::TRUNCATE, MVT::v4i32, Custom);
1146 setOperationAction(ISD::CTPOP, MVT::v32i8, Custom);
1147 setOperationAction(ISD::CTPOP, MVT::v16i16, Custom);
1148 setOperationAction(ISD::CTPOP, MVT::v8i32, Custom);
1149 setOperationAction(ISD::CTPOP, MVT::v4i64, Custom);
1151 setOperationAction(ISD::CTTZ, MVT::v32i8, Custom);
1152 setOperationAction(ISD::CTTZ, MVT::v16i16, Custom);
1153 setOperationAction(ISD::CTTZ, MVT::v8i32, Custom);
1154 setOperationAction(ISD::CTTZ, MVT::v4i64, Custom);
1155 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::v32i8, Custom);
1156 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::v16i16, Custom);
1157 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::v8i32, Custom);
1158 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::v4i64, Custom);
1160 if (Subtarget->hasFMA() || Subtarget->hasFMA4() || Subtarget->hasAVX512()) {
1161 setOperationAction(ISD::FMA, MVT::v8f32, Legal);
1162 setOperationAction(ISD::FMA, MVT::v4f64, Legal);
1163 setOperationAction(ISD::FMA, MVT::v4f32, Legal);
1164 setOperationAction(ISD::FMA, MVT::v2f64, Legal);
1165 setOperationAction(ISD::FMA, MVT::f32, Legal);
1166 setOperationAction(ISD::FMA, MVT::f64, Legal);
1169 if (Subtarget->hasInt256()) {
1170 setOperationAction(ISD::ADD, MVT::v4i64, Legal);
1171 setOperationAction(ISD::ADD, MVT::v8i32, Legal);
1172 setOperationAction(ISD::ADD, MVT::v16i16, Legal);
1173 setOperationAction(ISD::ADD, MVT::v32i8, Legal);
1175 setOperationAction(ISD::SUB, MVT::v4i64, Legal);
1176 setOperationAction(ISD::SUB, MVT::v8i32, Legal);
1177 setOperationAction(ISD::SUB, MVT::v16i16, Legal);
1178 setOperationAction(ISD::SUB, MVT::v32i8, Legal);
1180 setOperationAction(ISD::MUL, MVT::v4i64, Custom);
1181 setOperationAction(ISD::MUL, MVT::v8i32, Legal);
1182 setOperationAction(ISD::MUL, MVT::v16i16, Legal);
1183 setOperationAction(ISD::MUL, MVT::v32i8, Custom);
1185 setOperationAction(ISD::UMUL_LOHI, MVT::v8i32, Custom);
1186 setOperationAction(ISD::SMUL_LOHI, MVT::v8i32, Custom);
1187 setOperationAction(ISD::MULHU, MVT::v16i16, Legal);
1188 setOperationAction(ISD::MULHS, MVT::v16i16, Legal);
1190 setOperationAction(ISD::SMAX, MVT::v32i8, Legal);
1191 setOperationAction(ISD::SMAX, MVT::v16i16, Legal);
1192 setOperationAction(ISD::SMAX, MVT::v8i32, Legal);
1193 setOperationAction(ISD::UMAX, MVT::v32i8, Legal);
1194 setOperationAction(ISD::UMAX, MVT::v16i16, Legal);
1195 setOperationAction(ISD::UMAX, MVT::v8i32, Legal);
1196 setOperationAction(ISD::SMIN, MVT::v32i8, Legal);
1197 setOperationAction(ISD::SMIN, MVT::v16i16, Legal);
1198 setOperationAction(ISD::SMIN, MVT::v8i32, Legal);
1199 setOperationAction(ISD::UMIN, MVT::v32i8, Legal);
1200 setOperationAction(ISD::UMIN, MVT::v16i16, Legal);
1201 setOperationAction(ISD::UMIN, MVT::v8i32, Legal);
1203 // The custom lowering for UINT_TO_FP for v8i32 becomes interesting
1204 // when we have a 256bit-wide blend with immediate.
1205 setOperationAction(ISD::UINT_TO_FP, MVT::v8i32, Custom);
1207 // AVX2 also has wider vector sign/zero extending loads, VPMOV[SZ]X
1208 setLoadExtAction(ISD::SEXTLOAD, MVT::v16i16, MVT::v16i8, Legal);
1209 setLoadExtAction(ISD::SEXTLOAD, MVT::v8i32, MVT::v8i8, Legal);
1210 setLoadExtAction(ISD::SEXTLOAD, MVT::v4i64, MVT::v4i8, Legal);
1211 setLoadExtAction(ISD::SEXTLOAD, MVT::v8i32, MVT::v8i16, Legal);
1212 setLoadExtAction(ISD::SEXTLOAD, MVT::v4i64, MVT::v4i16, Legal);
1213 setLoadExtAction(ISD::SEXTLOAD, MVT::v4i64, MVT::v4i32, Legal);
1215 setLoadExtAction(ISD::ZEXTLOAD, MVT::v16i16, MVT::v16i8, Legal);
1216 setLoadExtAction(ISD::ZEXTLOAD, MVT::v8i32, MVT::v8i8, Legal);
1217 setLoadExtAction(ISD::ZEXTLOAD, MVT::v4i64, MVT::v4i8, Legal);
1218 setLoadExtAction(ISD::ZEXTLOAD, MVT::v8i32, MVT::v8i16, Legal);
1219 setLoadExtAction(ISD::ZEXTLOAD, MVT::v4i64, MVT::v4i16, Legal);
1220 setLoadExtAction(ISD::ZEXTLOAD, MVT::v4i64, MVT::v4i32, Legal);
1222 setOperationAction(ISD::ADD, MVT::v4i64, Custom);
1223 setOperationAction(ISD::ADD, MVT::v8i32, Custom);
1224 setOperationAction(ISD::ADD, MVT::v16i16, Custom);
1225 setOperationAction(ISD::ADD, MVT::v32i8, Custom);
1227 setOperationAction(ISD::SUB, MVT::v4i64, Custom);
1228 setOperationAction(ISD::SUB, MVT::v8i32, Custom);
1229 setOperationAction(ISD::SUB, MVT::v16i16, Custom);
1230 setOperationAction(ISD::SUB, MVT::v32i8, Custom);
1232 setOperationAction(ISD::MUL, MVT::v4i64, Custom);
1233 setOperationAction(ISD::MUL, MVT::v8i32, Custom);
1234 setOperationAction(ISD::MUL, MVT::v16i16, Custom);
1235 setOperationAction(ISD::MUL, MVT::v32i8, Custom);
1237 setOperationAction(ISD::SMAX, MVT::v32i8, Custom);
1238 setOperationAction(ISD::SMAX, MVT::v16i16, Custom);
1239 setOperationAction(ISD::SMAX, MVT::v8i32, Custom);
1240 setOperationAction(ISD::UMAX, MVT::v32i8, Custom);
1241 setOperationAction(ISD::UMAX, MVT::v16i16, Custom);
1242 setOperationAction(ISD::UMAX, MVT::v8i32, Custom);
1243 setOperationAction(ISD::SMIN, MVT::v32i8, Custom);
1244 setOperationAction(ISD::SMIN, MVT::v16i16, Custom);
1245 setOperationAction(ISD::SMIN, MVT::v8i32, Custom);
1246 setOperationAction(ISD::UMIN, MVT::v32i8, Custom);
1247 setOperationAction(ISD::UMIN, MVT::v16i16, Custom);
1248 setOperationAction(ISD::UMIN, MVT::v8i32, Custom);
1251 // In the customized shift lowering, the legal cases in AVX2 will be
1253 setOperationAction(ISD::SRL, MVT::v4i64, Custom);
1254 setOperationAction(ISD::SRL, MVT::v8i32, Custom);
1256 setOperationAction(ISD::SHL, MVT::v4i64, Custom);
1257 setOperationAction(ISD::SHL, MVT::v8i32, Custom);
1259 setOperationAction(ISD::SRA, MVT::v4i64, Custom);
1260 setOperationAction(ISD::SRA, MVT::v8i32, Custom);
1262 // Custom lower several nodes for 256-bit types.
1263 for (MVT VT : MVT::vector_valuetypes()) {
1264 if (VT.getScalarSizeInBits() >= 32) {
1265 setOperationAction(ISD::MLOAD, VT, Legal);
1266 setOperationAction(ISD::MSTORE, VT, Legal);
1268 // Extract subvector is special because the value type
1269 // (result) is 128-bit but the source is 256-bit wide.
1270 if (VT.is128BitVector()) {
1271 setOperationAction(ISD::EXTRACT_SUBVECTOR, VT, Custom);
1273 // Do not attempt to custom lower other non-256-bit vectors
1274 if (!VT.is256BitVector())
1277 setOperationAction(ISD::BUILD_VECTOR, VT, Custom);
1278 setOperationAction(ISD::VECTOR_SHUFFLE, VT, Custom);
1279 setOperationAction(ISD::VSELECT, VT, Custom);
1280 setOperationAction(ISD::INSERT_VECTOR_ELT, VT, Custom);
1281 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Custom);
1282 setOperationAction(ISD::SCALAR_TO_VECTOR, VT, Custom);
1283 setOperationAction(ISD::INSERT_SUBVECTOR, VT, Custom);
1284 setOperationAction(ISD::CONCAT_VECTORS, VT, Custom);
1287 if (Subtarget->hasInt256())
1288 setOperationAction(ISD::VSELECT, MVT::v32i8, Legal);
1290 // Promote v32i8, v16i16, v8i32 select, and, or, xor to v4i64.
1291 for (int i = MVT::v32i8; i != MVT::v4i64; ++i) {
1292 MVT VT = (MVT::SimpleValueType)i;
1294 // Do not attempt to promote non-256-bit vectors
1295 if (!VT.is256BitVector())
1298 setOperationAction(ISD::AND, VT, Promote);
1299 AddPromotedToType (ISD::AND, VT, MVT::v4i64);
1300 setOperationAction(ISD::OR, VT, Promote);
1301 AddPromotedToType (ISD::OR, VT, MVT::v4i64);
1302 setOperationAction(ISD::XOR, VT, Promote);
1303 AddPromotedToType (ISD::XOR, VT, MVT::v4i64);
1304 setOperationAction(ISD::LOAD, VT, Promote);
1305 AddPromotedToType (ISD::LOAD, VT, MVT::v4i64);
1306 setOperationAction(ISD::SELECT, VT, Promote);
1307 AddPromotedToType (ISD::SELECT, VT, MVT::v4i64);
1311 if (!Subtarget->useSoftFloat() && Subtarget->hasAVX512()) {
1312 addRegisterClass(MVT::v16i32, &X86::VR512RegClass);
1313 addRegisterClass(MVT::v16f32, &X86::VR512RegClass);
1314 addRegisterClass(MVT::v8i64, &X86::VR512RegClass);
1315 addRegisterClass(MVT::v8f64, &X86::VR512RegClass);
1317 addRegisterClass(MVT::i1, &X86::VK1RegClass);
1318 addRegisterClass(MVT::v8i1, &X86::VK8RegClass);
1319 addRegisterClass(MVT::v16i1, &X86::VK16RegClass);
1321 for (MVT VT : MVT::fp_vector_valuetypes())
1322 setLoadExtAction(ISD::EXTLOAD, VT, MVT::v8f32, Legal);
1324 setLoadExtAction(ISD::ZEXTLOAD, MVT::v16i32, MVT::v16i8, Legal);
1325 setLoadExtAction(ISD::SEXTLOAD, MVT::v16i32, MVT::v16i8, Legal);
1326 setLoadExtAction(ISD::ZEXTLOAD, MVT::v16i32, MVT::v16i16, Legal);
1327 setLoadExtAction(ISD::SEXTLOAD, MVT::v16i32, MVT::v16i16, Legal);
1328 setLoadExtAction(ISD::ZEXTLOAD, MVT::v32i16, MVT::v32i8, Legal);
1329 setLoadExtAction(ISD::SEXTLOAD, MVT::v32i16, MVT::v32i8, Legal);
1330 setLoadExtAction(ISD::ZEXTLOAD, MVT::v8i64, MVT::v8i8, Legal);
1331 setLoadExtAction(ISD::SEXTLOAD, MVT::v8i64, MVT::v8i8, Legal);
1332 setLoadExtAction(ISD::ZEXTLOAD, MVT::v8i64, MVT::v8i16, Legal);
1333 setLoadExtAction(ISD::SEXTLOAD, MVT::v8i64, MVT::v8i16, Legal);
1334 setLoadExtAction(ISD::ZEXTLOAD, MVT::v8i64, MVT::v8i32, Legal);
1335 setLoadExtAction(ISD::SEXTLOAD, MVT::v8i64, MVT::v8i32, Legal);
1337 setOperationAction(ISD::BR_CC, MVT::i1, Expand);
1338 setOperationAction(ISD::SETCC, MVT::i1, Custom);
1339 setOperationAction(ISD::SELECT_CC, MVT::i1, Expand);
1340 setOperationAction(ISD::XOR, MVT::i1, Legal);
1341 setOperationAction(ISD::OR, MVT::i1, Legal);
1342 setOperationAction(ISD::AND, MVT::i1, Legal);
1343 setOperationAction(ISD::SUB, MVT::i1, Custom);
1344 setOperationAction(ISD::ADD, MVT::i1, Custom);
1345 setOperationAction(ISD::MUL, MVT::i1, Custom);
1346 setOperationAction(ISD::LOAD, MVT::v16f32, Legal);
1347 setOperationAction(ISD::LOAD, MVT::v8f64, Legal);
1348 setOperationAction(ISD::LOAD, MVT::v8i64, Legal);
1349 setOperationAction(ISD::LOAD, MVT::v16i32, Legal);
1350 setOperationAction(ISD::LOAD, MVT::v16i1, Legal);
1352 setOperationAction(ISD::FADD, MVT::v16f32, Legal);
1353 setOperationAction(ISD::FSUB, MVT::v16f32, Legal);
1354 setOperationAction(ISD::FMUL, MVT::v16f32, Legal);
1355 setOperationAction(ISD::FDIV, MVT::v16f32, Legal);
1356 setOperationAction(ISD::FSQRT, MVT::v16f32, Legal);
1357 setOperationAction(ISD::FNEG, MVT::v16f32, Custom);
1359 setOperationAction(ISD::FADD, MVT::v8f64, Legal);
1360 setOperationAction(ISD::FSUB, MVT::v8f64, Legal);
1361 setOperationAction(ISD::FMUL, MVT::v8f64, Legal);
1362 setOperationAction(ISD::FDIV, MVT::v8f64, Legal);
1363 setOperationAction(ISD::FSQRT, MVT::v8f64, Legal);
1364 setOperationAction(ISD::FNEG, MVT::v8f64, Custom);
1365 setOperationAction(ISD::FMA, MVT::v8f64, Legal);
1366 setOperationAction(ISD::FMA, MVT::v16f32, Legal);
1368 setOperationAction(ISD::FP_TO_SINT, MVT::v16i32, Legal);
1369 setOperationAction(ISD::FP_TO_UINT, MVT::v16i32, Legal);
1370 setOperationAction(ISD::FP_TO_UINT, MVT::v8i32, Legal);
1371 setOperationAction(ISD::FP_TO_UINT, MVT::v4i32, Legal);
1372 setOperationAction(ISD::SINT_TO_FP, MVT::v16i32, Legal);
1373 setOperationAction(ISD::SINT_TO_FP, MVT::v8i1, Custom);
1374 setOperationAction(ISD::SINT_TO_FP, MVT::v16i1, Custom);
1375 setOperationAction(ISD::SINT_TO_FP, MVT::v16i8, Promote);
1376 setOperationAction(ISD::SINT_TO_FP, MVT::v16i16, Promote);
1377 setOperationAction(ISD::UINT_TO_FP, MVT::v16i32, Legal);
1378 setOperationAction(ISD::UINT_TO_FP, MVT::v8i32, Legal);
1379 setOperationAction(ISD::UINT_TO_FP, MVT::v4i32, Legal);
1380 setOperationAction(ISD::UINT_TO_FP, MVT::v16i8, Custom);
1381 setOperationAction(ISD::UINT_TO_FP, MVT::v16i16, Custom);
1382 setOperationAction(ISD::FP_ROUND, MVT::v8f32, Legal);
1383 setOperationAction(ISD::FP_EXTEND, MVT::v8f32, Legal);
1385 setTruncStoreAction(MVT::v8i64, MVT::v8i8, Legal);
1386 setTruncStoreAction(MVT::v8i64, MVT::v8i16, Legal);
1387 setTruncStoreAction(MVT::v8i64, MVT::v8i32, Legal);
1388 setTruncStoreAction(MVT::v16i32, MVT::v16i8, Legal);
1389 setTruncStoreAction(MVT::v16i32, MVT::v16i16, Legal);
1390 if (Subtarget->hasVLX()){
1391 setTruncStoreAction(MVT::v4i64, MVT::v4i8, Legal);
1392 setTruncStoreAction(MVT::v4i64, MVT::v4i16, Legal);
1393 setTruncStoreAction(MVT::v4i64, MVT::v4i32, Legal);
1394 setTruncStoreAction(MVT::v8i32, MVT::v8i8, Legal);
1395 setTruncStoreAction(MVT::v8i32, MVT::v8i16, Legal);
1397 setTruncStoreAction(MVT::v2i64, MVT::v2i8, Legal);
1398 setTruncStoreAction(MVT::v2i64, MVT::v2i16, Legal);
1399 setTruncStoreAction(MVT::v2i64, MVT::v2i32, Legal);
1400 setTruncStoreAction(MVT::v4i32, MVT::v4i8, Legal);
1401 setTruncStoreAction(MVT::v4i32, MVT::v4i16, Legal);
1403 setOperationAction(ISD::TRUNCATE, MVT::i1, Custom);
1404 setOperationAction(ISD::TRUNCATE, MVT::v16i8, Custom);
1405 setOperationAction(ISD::TRUNCATE, MVT::v8i32, Custom);
1406 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v8i1, Custom);
1407 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v16i1, Custom);
1408 if (Subtarget->hasDQI()) {
1409 setOperationAction(ISD::TRUNCATE, MVT::v2i1, Custom);
1410 setOperationAction(ISD::TRUNCATE, MVT::v4i1, Custom);
1412 setOperationAction(ISD::SINT_TO_FP, MVT::v8i64, Legal);
1413 setOperationAction(ISD::UINT_TO_FP, MVT::v8i64, Legal);
1414 setOperationAction(ISD::FP_TO_SINT, MVT::v8i64, Legal);
1415 setOperationAction(ISD::FP_TO_UINT, MVT::v8i64, Legal);
1416 if (Subtarget->hasVLX()) {
1417 setOperationAction(ISD::SINT_TO_FP, MVT::v4i64, Legal);
1418 setOperationAction(ISD::SINT_TO_FP, MVT::v2i64, Legal);
1419 setOperationAction(ISD::UINT_TO_FP, MVT::v4i64, Legal);
1420 setOperationAction(ISD::UINT_TO_FP, MVT::v2i64, Legal);
1421 setOperationAction(ISD::FP_TO_SINT, MVT::v4i64, Legal);
1422 setOperationAction(ISD::FP_TO_SINT, MVT::v2i64, Legal);
1423 setOperationAction(ISD::FP_TO_UINT, MVT::v4i64, Legal);
1424 setOperationAction(ISD::FP_TO_UINT, MVT::v2i64, Legal);
1427 if (Subtarget->hasVLX()) {
1428 setOperationAction(ISD::SINT_TO_FP, MVT::v8i32, Legal);
1429 setOperationAction(ISD::UINT_TO_FP, MVT::v8i32, Legal);
1430 setOperationAction(ISD::FP_TO_SINT, MVT::v8i32, Legal);
1431 setOperationAction(ISD::FP_TO_UINT, MVT::v8i32, Legal);
1432 setOperationAction(ISD::SINT_TO_FP, MVT::v4i32, Legal);
1433 setOperationAction(ISD::UINT_TO_FP, MVT::v4i32, Legal);
1434 setOperationAction(ISD::FP_TO_SINT, MVT::v4i32, Legal);
1435 setOperationAction(ISD::FP_TO_UINT, MVT::v4i32, Legal);
1437 setOperationAction(ISD::TRUNCATE, MVT::v8i1, Custom);
1438 setOperationAction(ISD::TRUNCATE, MVT::v16i1, Custom);
1439 setOperationAction(ISD::TRUNCATE, MVT::v16i16, Custom);
1440 setOperationAction(ISD::ZERO_EXTEND, MVT::v16i32, Custom);
1441 setOperationAction(ISD::ZERO_EXTEND, MVT::v8i64, Custom);
1442 setOperationAction(ISD::ANY_EXTEND, MVT::v16i32, Custom);
1443 setOperationAction(ISD::ANY_EXTEND, MVT::v8i64, Custom);
1444 setOperationAction(ISD::SIGN_EXTEND, MVT::v16i32, Custom);
1445 setOperationAction(ISD::SIGN_EXTEND, MVT::v8i64, Custom);
1446 setOperationAction(ISD::SIGN_EXTEND, MVT::v16i8, Custom);
1447 setOperationAction(ISD::SIGN_EXTEND, MVT::v8i16, Custom);
1448 setOperationAction(ISD::SIGN_EXTEND, MVT::v16i16, Custom);
1449 if (Subtarget->hasDQI()) {
1450 setOperationAction(ISD::SIGN_EXTEND, MVT::v4i32, Custom);
1451 setOperationAction(ISD::SIGN_EXTEND, MVT::v2i64, Custom);
1453 setOperationAction(ISD::FFLOOR, MVT::v16f32, Legal);
1454 setOperationAction(ISD::FFLOOR, MVT::v8f64, Legal);
1455 setOperationAction(ISD::FCEIL, MVT::v16f32, Legal);
1456 setOperationAction(ISD::FCEIL, MVT::v8f64, Legal);
1457 setOperationAction(ISD::FTRUNC, MVT::v16f32, Legal);
1458 setOperationAction(ISD::FTRUNC, MVT::v8f64, Legal);
1459 setOperationAction(ISD::FRINT, MVT::v16f32, Legal);
1460 setOperationAction(ISD::FRINT, MVT::v8f64, Legal);
1461 setOperationAction(ISD::FNEARBYINT, MVT::v16f32, Legal);
1462 setOperationAction(ISD::FNEARBYINT, MVT::v8f64, Legal);
1464 setOperationAction(ISD::CONCAT_VECTORS, MVT::v8f64, Custom);
1465 setOperationAction(ISD::CONCAT_VECTORS, MVT::v8i64, Custom);
1466 setOperationAction(ISD::CONCAT_VECTORS, MVT::v16f32, Custom);
1467 setOperationAction(ISD::CONCAT_VECTORS, MVT::v16i32, Custom);
1468 setOperationAction(ISD::CONCAT_VECTORS, MVT::v16i1, Legal);
1470 setOperationAction(ISD::SETCC, MVT::v16i1, Custom);
1471 setOperationAction(ISD::SETCC, MVT::v8i1, Custom);
1473 setOperationAction(ISD::MUL, MVT::v8i64, Custom);
1475 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v8i1, Custom);
1476 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v16i1, Custom);
1477 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v16i1, Custom);
1478 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i1, Custom);
1479 setOperationAction(ISD::BUILD_VECTOR, MVT::v8i1, Custom);
1480 setOperationAction(ISD::BUILD_VECTOR, MVT::v16i1, Custom);
1481 setOperationAction(ISD::SELECT, MVT::v8f64, Custom);
1482 setOperationAction(ISD::SELECT, MVT::v8i64, Custom);
1483 setOperationAction(ISD::SELECT, MVT::v16f32, Custom);
1484 setOperationAction(ISD::SELECT, MVT::v16i1, Custom);
1485 setOperationAction(ISD::SELECT, MVT::v8i1, Custom);
1487 setOperationAction(ISD::SMAX, MVT::v16i32, Legal);
1488 setOperationAction(ISD::SMAX, MVT::v8i64, Legal);
1489 setOperationAction(ISD::UMAX, MVT::v16i32, Legal);
1490 setOperationAction(ISD::UMAX, MVT::v8i64, Legal);
1491 setOperationAction(ISD::SMIN, MVT::v16i32, Legal);
1492 setOperationAction(ISD::SMIN, MVT::v8i64, Legal);
1493 setOperationAction(ISD::UMIN, MVT::v16i32, Legal);
1494 setOperationAction(ISD::UMIN, MVT::v8i64, Legal);
1496 setOperationAction(ISD::ADD, MVT::v8i64, Legal);
1497 setOperationAction(ISD::ADD, MVT::v16i32, Legal);
1499 setOperationAction(ISD::SUB, MVT::v8i64, Legal);
1500 setOperationAction(ISD::SUB, MVT::v16i32, Legal);
1502 setOperationAction(ISD::MUL, MVT::v16i32, Legal);
1504 setOperationAction(ISD::SRL, MVT::v8i64, Custom);
1505 setOperationAction(ISD::SRL, MVT::v16i32, Custom);
1507 setOperationAction(ISD::SHL, MVT::v8i64, Custom);
1508 setOperationAction(ISD::SHL, MVT::v16i32, Custom);
1510 setOperationAction(ISD::SRA, MVT::v8i64, Custom);
1511 setOperationAction(ISD::SRA, MVT::v16i32, Custom);
1513 setOperationAction(ISD::AND, MVT::v8i64, Legal);
1514 setOperationAction(ISD::OR, MVT::v8i64, Legal);
1515 setOperationAction(ISD::XOR, MVT::v8i64, Legal);
1516 setOperationAction(ISD::AND, MVT::v16i32, Legal);
1517 setOperationAction(ISD::OR, MVT::v16i32, Legal);
1518 setOperationAction(ISD::XOR, MVT::v16i32, Legal);
1520 if (Subtarget->hasCDI()) {
1521 setOperationAction(ISD::CTLZ, MVT::v8i64, Legal);
1522 setOperationAction(ISD::CTLZ, MVT::v16i32, Legal);
1523 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::v8i64, Legal);
1524 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::v16i32, Legal);
1526 setOperationAction(ISD::CTLZ, MVT::v8i16, Custom);
1527 setOperationAction(ISD::CTLZ, MVT::v16i8, Custom);
1528 setOperationAction(ISD::CTLZ, MVT::v16i16, Custom);
1529 setOperationAction(ISD::CTLZ, MVT::v32i8, Custom);
1530 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::v8i16, Custom);
1531 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::v16i8, Custom);
1532 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::v16i16, Custom);
1533 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::v32i8, Custom);
1535 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::v8i64, Custom);
1536 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::v16i32, Custom);
1538 if (Subtarget->hasVLX()) {
1539 setOperationAction(ISD::CTLZ, MVT::v4i64, Legal);
1540 setOperationAction(ISD::CTLZ, MVT::v8i32, Legal);
1541 setOperationAction(ISD::CTLZ, MVT::v2i64, Legal);
1542 setOperationAction(ISD::CTLZ, MVT::v4i32, Legal);
1543 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::v4i64, Legal);
1544 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::v8i32, Legal);
1545 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::v2i64, Legal);
1546 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::v4i32, Legal);
1548 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::v4i64, Custom);
1549 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::v8i32, Custom);
1550 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::v2i64, Custom);
1551 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::v4i32, Custom);
1553 setOperationAction(ISD::CTLZ, MVT::v4i64, Custom);
1554 setOperationAction(ISD::CTLZ, MVT::v8i32, Custom);
1555 setOperationAction(ISD::CTLZ, MVT::v2i64, Custom);
1556 setOperationAction(ISD::CTLZ, MVT::v4i32, Custom);
1557 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::v4i64, Custom);
1558 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::v8i32, Custom);
1559 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::v2i64, Custom);
1560 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::v4i32, Custom);
1562 } // Subtarget->hasCDI()
1564 if (Subtarget->hasDQI()) {
1565 setOperationAction(ISD::MUL, MVT::v2i64, Legal);
1566 setOperationAction(ISD::MUL, MVT::v4i64, Legal);
1567 setOperationAction(ISD::MUL, MVT::v8i64, Legal);
1569 // Custom lower several nodes.
1570 for (MVT VT : MVT::vector_valuetypes()) {
1571 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
1573 setOperationAction(ISD::AND, VT, Legal);
1574 setOperationAction(ISD::OR, VT, Legal);
1575 setOperationAction(ISD::XOR, VT, Legal);
1577 if (EltSize >= 32 && VT.getSizeInBits() <= 512) {
1578 setOperationAction(ISD::MGATHER, VT, Custom);
1579 setOperationAction(ISD::MSCATTER, VT, Custom);
1581 // Extract subvector is special because the value type
1582 // (result) is 256/128-bit but the source is 512-bit wide.
1583 if (VT.is128BitVector() || VT.is256BitVector()) {
1584 setOperationAction(ISD::EXTRACT_SUBVECTOR, VT, Custom);
1586 if (VT.getVectorElementType() == MVT::i1)
1587 setOperationAction(ISD::EXTRACT_SUBVECTOR, VT, Legal);
1589 // Do not attempt to custom lower other non-512-bit vectors
1590 if (!VT.is512BitVector())
1593 if (EltSize >= 32) {
1594 setOperationAction(ISD::VECTOR_SHUFFLE, VT, Custom);
1595 setOperationAction(ISD::INSERT_VECTOR_ELT, VT, Custom);
1596 setOperationAction(ISD::BUILD_VECTOR, VT, Custom);
1597 setOperationAction(ISD::VSELECT, VT, Legal);
1598 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Custom);
1599 setOperationAction(ISD::SCALAR_TO_VECTOR, VT, Custom);
1600 setOperationAction(ISD::INSERT_SUBVECTOR, VT, Custom);
1601 setOperationAction(ISD::MLOAD, VT, Legal);
1602 setOperationAction(ISD::MSTORE, VT, Legal);
1605 for (int i = MVT::v32i8; i != MVT::v8i64; ++i) {
1606 MVT VT = (MVT::SimpleValueType)i;
1608 // Do not attempt to promote non-512-bit vectors.
1609 if (!VT.is512BitVector())
1612 setOperationAction(ISD::SELECT, VT, Promote);
1613 AddPromotedToType (ISD::SELECT, VT, MVT::v8i64);
1617 if (!Subtarget->useSoftFloat() && Subtarget->hasBWI()) {
1618 addRegisterClass(MVT::v32i16, &X86::VR512RegClass);
1619 addRegisterClass(MVT::v64i8, &X86::VR512RegClass);
1621 addRegisterClass(MVT::v32i1, &X86::VK32RegClass);
1622 addRegisterClass(MVT::v64i1, &X86::VK64RegClass);
1624 setOperationAction(ISD::LOAD, MVT::v32i16, Legal);
1625 setOperationAction(ISD::LOAD, MVT::v64i8, Legal);
1626 setOperationAction(ISD::SETCC, MVT::v32i1, Custom);
1627 setOperationAction(ISD::SETCC, MVT::v64i1, Custom);
1628 setOperationAction(ISD::ADD, MVT::v32i16, Legal);
1629 setOperationAction(ISD::ADD, MVT::v64i8, Legal);
1630 setOperationAction(ISD::SUB, MVT::v32i16, Legal);
1631 setOperationAction(ISD::SUB, MVT::v64i8, Legal);
1632 setOperationAction(ISD::MUL, MVT::v32i16, Legal);
1633 setOperationAction(ISD::MULHS, MVT::v32i16, Legal);
1634 setOperationAction(ISD::MULHU, MVT::v32i16, Legal);
1635 setOperationAction(ISD::CONCAT_VECTORS, MVT::v32i1, Legal);
1636 setOperationAction(ISD::CONCAT_VECTORS, MVT::v64i1, Legal);
1637 setOperationAction(ISD::CONCAT_VECTORS, MVT::v32i16, Custom);
1638 setOperationAction(ISD::CONCAT_VECTORS, MVT::v64i8, Custom);
1639 setOperationAction(ISD::INSERT_SUBVECTOR, MVT::v32i1, Custom);
1640 setOperationAction(ISD::INSERT_SUBVECTOR, MVT::v64i1, Custom);
1641 setOperationAction(ISD::INSERT_SUBVECTOR, MVT::v32i16, Custom);
1642 setOperationAction(ISD::INSERT_SUBVECTOR, MVT::v64i8, Custom);
1643 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v32i16, Custom);
1644 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v64i8, Custom);
1645 setOperationAction(ISD::SELECT, MVT::v32i1, Custom);
1646 setOperationAction(ISD::SELECT, MVT::v64i1, Custom);
1647 setOperationAction(ISD::SIGN_EXTEND, MVT::v32i8, Custom);
1648 setOperationAction(ISD::ZERO_EXTEND, MVT::v32i8, Custom);
1649 setOperationAction(ISD::SIGN_EXTEND, MVT::v32i16, Custom);
1650 setOperationAction(ISD::ZERO_EXTEND, MVT::v32i16, Custom);
1651 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v32i16, Custom);
1652 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v64i8, Custom);
1653 setOperationAction(ISD::SIGN_EXTEND, MVT::v64i8, Custom);
1654 setOperationAction(ISD::ZERO_EXTEND, MVT::v64i8, Custom);
1655 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v32i1, Custom);
1656 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v64i1, Custom);
1657 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v32i16, Custom);
1658 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v64i8, Custom);
1659 setOperationAction(ISD::VSELECT, MVT::v32i16, Legal);
1660 setOperationAction(ISD::VSELECT, MVT::v64i8, Legal);
1661 setOperationAction(ISD::TRUNCATE, MVT::v32i1, Custom);
1662 setOperationAction(ISD::TRUNCATE, MVT::v64i1, Custom);
1663 setOperationAction(ISD::TRUNCATE, MVT::v32i8, Custom);
1664 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v32i1, Custom);
1665 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v64i1, Custom);
1667 setOperationAction(ISD::SMAX, MVT::v64i8, Legal);
1668 setOperationAction(ISD::SMAX, MVT::v32i16, Legal);
1669 setOperationAction(ISD::UMAX, MVT::v64i8, Legal);
1670 setOperationAction(ISD::UMAX, MVT::v32i16, Legal);
1671 setOperationAction(ISD::SMIN, MVT::v64i8, Legal);
1672 setOperationAction(ISD::SMIN, MVT::v32i16, Legal);
1673 setOperationAction(ISD::UMIN, MVT::v64i8, Legal);
1674 setOperationAction(ISD::UMIN, MVT::v32i16, Legal);
1676 setTruncStoreAction(MVT::v32i16, MVT::v32i8, Legal);
1677 setTruncStoreAction(MVT::v16i16, MVT::v16i8, Legal);
1678 if (Subtarget->hasVLX())
1679 setTruncStoreAction(MVT::v8i16, MVT::v8i8, Legal);
1681 if (Subtarget->hasCDI()) {
1682 setOperationAction(ISD::CTLZ, MVT::v32i16, Custom);
1683 setOperationAction(ISD::CTLZ, MVT::v64i8, Custom);
1684 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::v32i16, Custom);
1685 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::v64i8, Custom);
1688 for (int i = MVT::v32i8; i != MVT::v8i64; ++i) {
1689 const MVT VT = (MVT::SimpleValueType)i;
1691 const unsigned EltSize = VT.getVectorElementType().getSizeInBits();
1693 // Do not attempt to promote non-512-bit vectors.
1694 if (!VT.is512BitVector())
1698 setOperationAction(ISD::BUILD_VECTOR, VT, Custom);
1699 setOperationAction(ISD::VSELECT, VT, Legal);
1704 if (!Subtarget->useSoftFloat() && Subtarget->hasVLX()) {
1705 addRegisterClass(MVT::v4i1, &X86::VK4RegClass);
1706 addRegisterClass(MVT::v2i1, &X86::VK2RegClass);
1708 setOperationAction(ISD::SETCC, MVT::v4i1, Custom);
1709 setOperationAction(ISD::SETCC, MVT::v2i1, Custom);
1710 setOperationAction(ISD::CONCAT_VECTORS, MVT::v4i1, Custom);
1711 setOperationAction(ISD::CONCAT_VECTORS, MVT::v8i1, Custom);
1712 setOperationAction(ISD::INSERT_SUBVECTOR, MVT::v8i1, Custom);
1713 setOperationAction(ISD::INSERT_SUBVECTOR, MVT::v4i1, Custom);
1714 setOperationAction(ISD::SELECT, MVT::v4i1, Custom);
1715 setOperationAction(ISD::SELECT, MVT::v2i1, Custom);
1716 setOperationAction(ISD::BUILD_VECTOR, MVT::v4i1, Custom);
1717 setOperationAction(ISD::BUILD_VECTOR, MVT::v2i1, Custom);
1718 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2i1, Custom);
1719 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4i1, Custom);
1721 setOperationAction(ISD::AND, MVT::v8i32, Legal);
1722 setOperationAction(ISD::OR, MVT::v8i32, Legal);
1723 setOperationAction(ISD::XOR, MVT::v8i32, Legal);
1724 setOperationAction(ISD::AND, MVT::v4i32, Legal);
1725 setOperationAction(ISD::OR, MVT::v4i32, Legal);
1726 setOperationAction(ISD::XOR, MVT::v4i32, Legal);
1727 setOperationAction(ISD::SRA, MVT::v2i64, Custom);
1728 setOperationAction(ISD::SRA, MVT::v4i64, Custom);
1730 setOperationAction(ISD::SMAX, MVT::v2i64, Legal);
1731 setOperationAction(ISD::SMAX, MVT::v4i64, Legal);
1732 setOperationAction(ISD::UMAX, MVT::v2i64, Legal);
1733 setOperationAction(ISD::UMAX, MVT::v4i64, Legal);
1734 setOperationAction(ISD::SMIN, MVT::v2i64, Legal);
1735 setOperationAction(ISD::SMIN, MVT::v4i64, Legal);
1736 setOperationAction(ISD::UMIN, MVT::v2i64, Legal);
1737 setOperationAction(ISD::UMIN, MVT::v4i64, Legal);
1740 // We want to custom lower some of our intrinsics.
1741 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
1742 setOperationAction(ISD::INTRINSIC_W_CHAIN, MVT::Other, Custom);
1743 setOperationAction(ISD::INTRINSIC_VOID, MVT::Other, Custom);
1744 if (!Subtarget->is64Bit())
1745 setOperationAction(ISD::INTRINSIC_W_CHAIN, MVT::i64, Custom);
1747 // Only custom-lower 64-bit SADDO and friends on 64-bit because we don't
1748 // handle type legalization for these operations here.
1750 // FIXME: We really should do custom legalization for addition and
1751 // subtraction on x86-32 once PR3203 is fixed. We really can't do much better
1752 // than generic legalization for 64-bit multiplication-with-overflow, though.
1753 for (auto VT : { MVT::i8, MVT::i16, MVT::i32, MVT::i64 }) {
1754 if (VT == MVT::i64 && !Subtarget->is64Bit())
1756 // Add/Sub/Mul with overflow operations are custom lowered.
1757 setOperationAction(ISD::SADDO, VT, Custom);
1758 setOperationAction(ISD::UADDO, VT, Custom);
1759 setOperationAction(ISD::SSUBO, VT, Custom);
1760 setOperationAction(ISD::USUBO, VT, Custom);
1761 setOperationAction(ISD::SMULO, VT, Custom);
1762 setOperationAction(ISD::UMULO, VT, Custom);
1765 if (!Subtarget->is64Bit()) {
1766 // These libcalls are not available in 32-bit.
1767 setLibcallName(RTLIB::SHL_I128, nullptr);
1768 setLibcallName(RTLIB::SRL_I128, nullptr);
1769 setLibcallName(RTLIB::SRA_I128, nullptr);
1772 // Combine sin / cos into one node or libcall if possible.
1773 if (Subtarget->hasSinCos()) {
1774 setLibcallName(RTLIB::SINCOS_F32, "sincosf");
1775 setLibcallName(RTLIB::SINCOS_F64, "sincos");
1776 if (Subtarget->isTargetDarwin()) {
1777 // For MacOSX, we don't want the normal expansion of a libcall to sincos.
1778 // We want to issue a libcall to __sincos_stret to avoid memory traffic.
1779 setOperationAction(ISD::FSINCOS, MVT::f64, Custom);
1780 setOperationAction(ISD::FSINCOS, MVT::f32, Custom);
1784 if (Subtarget->isTargetWin64()) {
1785 setOperationAction(ISD::SDIV, MVT::i128, Custom);
1786 setOperationAction(ISD::UDIV, MVT::i128, Custom);
1787 setOperationAction(ISD::SREM, MVT::i128, Custom);
1788 setOperationAction(ISD::UREM, MVT::i128, Custom);
1789 setOperationAction(ISD::SDIVREM, MVT::i128, Custom);
1790 setOperationAction(ISD::UDIVREM, MVT::i128, Custom);
1793 // We have target-specific dag combine patterns for the following nodes:
1794 setTargetDAGCombine(ISD::VECTOR_SHUFFLE);
1795 setTargetDAGCombine(ISD::EXTRACT_VECTOR_ELT);
1796 setTargetDAGCombine(ISD::BITCAST);
1797 setTargetDAGCombine(ISD::VSELECT);
1798 setTargetDAGCombine(ISD::SELECT);
1799 setTargetDAGCombine(ISD::SHL);
1800 setTargetDAGCombine(ISD::SRA);
1801 setTargetDAGCombine(ISD::SRL);
1802 setTargetDAGCombine(ISD::OR);
1803 setTargetDAGCombine(ISD::AND);
1804 setTargetDAGCombine(ISD::ADD);
1805 setTargetDAGCombine(ISD::FADD);
1806 setTargetDAGCombine(ISD::FSUB);
1807 setTargetDAGCombine(ISD::FMA);
1808 setTargetDAGCombine(ISD::SUB);
1809 setTargetDAGCombine(ISD::LOAD);
1810 setTargetDAGCombine(ISD::MLOAD);
1811 setTargetDAGCombine(ISD::STORE);
1812 setTargetDAGCombine(ISD::MSTORE);
1813 setTargetDAGCombine(ISD::ZERO_EXTEND);
1814 setTargetDAGCombine(ISD::ANY_EXTEND);
1815 setTargetDAGCombine(ISD::SIGN_EXTEND);
1816 setTargetDAGCombine(ISD::SIGN_EXTEND_INREG);
1817 setTargetDAGCombine(ISD::SINT_TO_FP);
1818 setTargetDAGCombine(ISD::UINT_TO_FP);
1819 setTargetDAGCombine(ISD::SETCC);
1820 setTargetDAGCombine(ISD::BUILD_VECTOR);
1821 setTargetDAGCombine(ISD::MUL);
1822 setTargetDAGCombine(ISD::XOR);
1824 computeRegisterProperties(Subtarget->getRegisterInfo());
1826 MaxStoresPerMemset = 16; // For @llvm.memset -> sequence of stores
1827 MaxStoresPerMemsetOptSize = 8;
1828 MaxStoresPerMemcpy = 8; // For @llvm.memcpy -> sequence of stores
1829 MaxStoresPerMemcpyOptSize = 4;
1830 MaxStoresPerMemmove = 8; // For @llvm.memmove -> sequence of stores
1831 MaxStoresPerMemmoveOptSize = 4;
1832 setPrefLoopAlignment(4); // 2^4 bytes.
1834 // A predictable cmov does not hurt on an in-order CPU.
1835 // FIXME: Use a CPU attribute to trigger this, not a CPU model.
1836 PredictableSelectIsExpensive = !Subtarget->isAtom();
1837 EnableExtLdPromotion = true;
1838 setPrefFunctionAlignment(4); // 2^4 bytes.
1840 verifyIntrinsicTables();
1843 // This has so far only been implemented for 64-bit MachO.
1844 bool X86TargetLowering::useLoadStackGuardNode() const {
1845 return Subtarget->isTargetMachO() && Subtarget->is64Bit();
1848 TargetLoweringBase::LegalizeTypeAction
1849 X86TargetLowering::getPreferredVectorAction(EVT VT) const {
1850 if (ExperimentalVectorWideningLegalization &&
1851 VT.getVectorNumElements() != 1 &&
1852 VT.getVectorElementType().getSimpleVT() != MVT::i1)
1853 return TypeWidenVector;
1855 return TargetLoweringBase::getPreferredVectorAction(VT);
1858 EVT X86TargetLowering::getSetCCResultType(const DataLayout &DL, LLVMContext &,
1861 return Subtarget->hasAVX512() ? MVT::i1: MVT::i8;
1863 const unsigned NumElts = VT.getVectorNumElements();
1864 const EVT EltVT = VT.getVectorElementType();
1865 if (VT.is512BitVector()) {
1866 if (Subtarget->hasAVX512())
1867 if (EltVT == MVT::i32 || EltVT == MVT::i64 ||
1868 EltVT == MVT::f32 || EltVT == MVT::f64)
1870 case 8: return MVT::v8i1;
1871 case 16: return MVT::v16i1;
1873 if (Subtarget->hasBWI())
1874 if (EltVT == MVT::i8 || EltVT == MVT::i16)
1876 case 32: return MVT::v32i1;
1877 case 64: return MVT::v64i1;
1881 if (VT.is256BitVector() || VT.is128BitVector()) {
1882 if (Subtarget->hasVLX())
1883 if (EltVT == MVT::i32 || EltVT == MVT::i64 ||
1884 EltVT == MVT::f32 || EltVT == MVT::f64)
1886 case 2: return MVT::v2i1;
1887 case 4: return MVT::v4i1;
1888 case 8: return MVT::v8i1;
1890 if (Subtarget->hasBWI() && Subtarget->hasVLX())
1891 if (EltVT == MVT::i8 || EltVT == MVT::i16)
1893 case 8: return MVT::v8i1;
1894 case 16: return MVT::v16i1;
1895 case 32: return MVT::v32i1;
1899 return VT.changeVectorElementTypeToInteger();
1902 /// Helper for getByValTypeAlignment to determine
1903 /// the desired ByVal argument alignment.
1904 static void getMaxByValAlign(Type *Ty, unsigned &MaxAlign) {
1907 if (VectorType *VTy = dyn_cast<VectorType>(Ty)) {
1908 if (VTy->getBitWidth() == 128)
1910 } else if (ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
1911 unsigned EltAlign = 0;
1912 getMaxByValAlign(ATy->getElementType(), EltAlign);
1913 if (EltAlign > MaxAlign)
1914 MaxAlign = EltAlign;
1915 } else if (StructType *STy = dyn_cast<StructType>(Ty)) {
1916 for (auto *EltTy : STy->elements()) {
1917 unsigned EltAlign = 0;
1918 getMaxByValAlign(EltTy, EltAlign);
1919 if (EltAlign > MaxAlign)
1920 MaxAlign = EltAlign;
1927 /// Return the desired alignment for ByVal aggregate
1928 /// function arguments in the caller parameter area. For X86, aggregates
1929 /// that contain SSE vectors are placed at 16-byte boundaries while the rest
1930 /// are at 4-byte boundaries.
1931 unsigned X86TargetLowering::getByValTypeAlignment(Type *Ty,
1932 const DataLayout &DL) const {
1933 if (Subtarget->is64Bit()) {
1934 // Max of 8 and alignment of type.
1935 unsigned TyAlign = DL.getABITypeAlignment(Ty);
1942 if (Subtarget->hasSSE1())
1943 getMaxByValAlign(Ty, Align);
1947 /// Returns the target specific optimal type for load
1948 /// and store operations as a result of memset, memcpy, and memmove
1949 /// lowering. If DstAlign is zero that means it's safe to destination
1950 /// alignment can satisfy any constraint. Similarly if SrcAlign is zero it
1951 /// means there isn't a need to check it against alignment requirement,
1952 /// probably because the source does not need to be loaded. If 'IsMemset' is
1953 /// true, that means it's expanding a memset. If 'ZeroMemset' is true, that
1954 /// means it's a memset of zero. 'MemcpyStrSrc' indicates whether the memcpy
1955 /// source is constant so it does not need to be loaded.
1956 /// It returns EVT::Other if the type should be determined using generic
1957 /// target-independent logic.
1959 X86TargetLowering::getOptimalMemOpType(uint64_t Size,
1960 unsigned DstAlign, unsigned SrcAlign,
1961 bool IsMemset, bool ZeroMemset,
1963 MachineFunction &MF) const {
1964 const Function *F = MF.getFunction();
1965 if ((!IsMemset || ZeroMemset) &&
1966 !F->hasFnAttribute(Attribute::NoImplicitFloat)) {
1968 (!Subtarget->isUnalignedMem16Slow() ||
1969 ((DstAlign == 0 || DstAlign >= 16) &&
1970 (SrcAlign == 0 || SrcAlign >= 16)))) {
1972 // FIXME: Check if unaligned 32-byte accesses are slow.
1973 if (Subtarget->hasInt256())
1975 if (Subtarget->hasFp256())
1978 if (Subtarget->hasSSE2())
1980 if (Subtarget->hasSSE1())
1982 } else if (!MemcpyStrSrc && Size >= 8 &&
1983 !Subtarget->is64Bit() &&
1984 Subtarget->hasSSE2()) {
1985 // Do not use f64 to lower memcpy if source is string constant. It's
1986 // better to use i32 to avoid the loads.
1990 // This is a compromise. If we reach here, unaligned accesses may be slow on
1991 // this target. However, creating smaller, aligned accesses could be even
1992 // slower and would certainly be a lot more code.
1993 if (Subtarget->is64Bit() && Size >= 8)
1998 bool X86TargetLowering::isSafeMemOpType(MVT VT) const {
2000 return X86ScalarSSEf32;
2001 else if (VT == MVT::f64)
2002 return X86ScalarSSEf64;
2007 X86TargetLowering::allowsMisalignedMemoryAccesses(EVT VT,
2012 switch (VT.getSizeInBits()) {
2014 // 8-byte and under are always assumed to be fast.
2018 *Fast = !Subtarget->isUnalignedMem16Slow();
2021 *Fast = !Subtarget->isUnalignedMem32Slow();
2023 // TODO: What about AVX-512 (512-bit) accesses?
2026 // Misaligned accesses of any size are always allowed.
2030 /// Return the entry encoding for a jump table in the
2031 /// current function. The returned value is a member of the
2032 /// MachineJumpTableInfo::JTEntryKind enum.
2033 unsigned X86TargetLowering::getJumpTableEncoding() const {
2034 // In GOT pic mode, each entry in the jump table is emitted as a @GOTOFF
2036 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
2037 Subtarget->isPICStyleGOT())
2038 return MachineJumpTableInfo::EK_Custom32;
2040 // Otherwise, use the normal jump table encoding heuristics.
2041 return TargetLowering::getJumpTableEncoding();
2044 bool X86TargetLowering::useSoftFloat() const {
2045 return Subtarget->useSoftFloat();
2049 X86TargetLowering::LowerCustomJumpTableEntry(const MachineJumpTableInfo *MJTI,
2050 const MachineBasicBlock *MBB,
2051 unsigned uid,MCContext &Ctx) const{
2052 assert(MBB->getParent()->getTarget().getRelocationModel() == Reloc::PIC_ &&
2053 Subtarget->isPICStyleGOT());
2054 // In 32-bit ELF systems, our jump table entries are formed with @GOTOFF
2056 return MCSymbolRefExpr::create(MBB->getSymbol(),
2057 MCSymbolRefExpr::VK_GOTOFF, Ctx);
2060 /// Returns relocation base for the given PIC jumptable.
2061 SDValue X86TargetLowering::getPICJumpTableRelocBase(SDValue Table,
2062 SelectionDAG &DAG) const {
2063 if (!Subtarget->is64Bit())
2064 // This doesn't have SDLoc associated with it, but is not really the
2065 // same as a Register.
2066 return DAG.getNode(X86ISD::GlobalBaseReg, SDLoc(),
2067 getPointerTy(DAG.getDataLayout()));
2071 /// This returns the relocation base for the given PIC jumptable,
2072 /// the same as getPICJumpTableRelocBase, but as an MCExpr.
2073 const MCExpr *X86TargetLowering::
2074 getPICJumpTableRelocBaseExpr(const MachineFunction *MF, unsigned JTI,
2075 MCContext &Ctx) const {
2076 // X86-64 uses RIP relative addressing based on the jump table label.
2077 if (Subtarget->isPICStyleRIPRel())
2078 return TargetLowering::getPICJumpTableRelocBaseExpr(MF, JTI, Ctx);
2080 // Otherwise, the reference is relative to the PIC base.
2081 return MCSymbolRefExpr::create(MF->getPICBaseSymbol(), Ctx);
2084 std::pair<const TargetRegisterClass *, uint8_t>
2085 X86TargetLowering::findRepresentativeClass(const TargetRegisterInfo *TRI,
2087 const TargetRegisterClass *RRC = nullptr;
2089 switch (VT.SimpleTy) {
2091 return TargetLowering::findRepresentativeClass(TRI, VT);
2092 case MVT::i8: case MVT::i16: case MVT::i32: case MVT::i64:
2093 RRC = Subtarget->is64Bit() ? &X86::GR64RegClass : &X86::GR32RegClass;
2096 RRC = &X86::VR64RegClass;
2098 case MVT::f32: case MVT::f64:
2099 case MVT::v16i8: case MVT::v8i16: case MVT::v4i32: case MVT::v2i64:
2100 case MVT::v4f32: case MVT::v2f64:
2101 case MVT::v32i8: case MVT::v8i32: case MVT::v4i64: case MVT::v8f32:
2103 RRC = &X86::VR128RegClass;
2106 return std::make_pair(RRC, Cost);
2109 bool X86TargetLowering::getStackCookieLocation(unsigned &AddressSpace,
2110 unsigned &Offset) const {
2111 if (!Subtarget->isTargetLinux())
2114 if (Subtarget->is64Bit()) {
2115 // %fs:0x28, unless we're using a Kernel code model, in which case it's %gs:
2117 if (getTargetMachine().getCodeModel() == CodeModel::Kernel)
2129 Value *X86TargetLowering::getSafeStackPointerLocation(IRBuilder<> &IRB) const {
2130 if (!Subtarget->isTargetAndroid())
2131 return TargetLowering::getSafeStackPointerLocation(IRB);
2133 // Android provides a fixed TLS slot for the SafeStack pointer. See the
2134 // definition of TLS_SLOT_SAFESTACK in
2135 // https://android.googlesource.com/platform/bionic/+/master/libc/private/bionic_tls.h
2136 unsigned AddressSpace, Offset;
2137 if (Subtarget->is64Bit()) {
2138 // %fs:0x48, unless we're using a Kernel code model, in which case it's %gs:
2140 if (getTargetMachine().getCodeModel() == CodeModel::Kernel)
2150 return ConstantExpr::getIntToPtr(
2151 ConstantInt::get(Type::getInt32Ty(IRB.getContext()), Offset),
2152 Type::getInt8PtrTy(IRB.getContext())->getPointerTo(AddressSpace));
2155 bool X86TargetLowering::isNoopAddrSpaceCast(unsigned SrcAS,
2156 unsigned DestAS) const {
2157 assert(SrcAS != DestAS && "Expected different address spaces!");
2159 return SrcAS < 256 && DestAS < 256;
2162 //===----------------------------------------------------------------------===//
2163 // Return Value Calling Convention Implementation
2164 //===----------------------------------------------------------------------===//
2166 #include "X86GenCallingConv.inc"
2168 bool X86TargetLowering::CanLowerReturn(
2169 CallingConv::ID CallConv, MachineFunction &MF, bool isVarArg,
2170 const SmallVectorImpl<ISD::OutputArg> &Outs, LLVMContext &Context) const {
2171 SmallVector<CCValAssign, 16> RVLocs;
2172 CCState CCInfo(CallConv, isVarArg, MF, RVLocs, Context);
2173 return CCInfo.CheckReturn(Outs, RetCC_X86);
2176 const MCPhysReg *X86TargetLowering::getScratchRegisters(CallingConv::ID) const {
2177 static const MCPhysReg ScratchRegs[] = { X86::R11, 0 };
2182 X86TargetLowering::LowerReturn(SDValue Chain,
2183 CallingConv::ID CallConv, bool isVarArg,
2184 const SmallVectorImpl<ISD::OutputArg> &Outs,
2185 const SmallVectorImpl<SDValue> &OutVals,
2186 SDLoc dl, SelectionDAG &DAG) const {
2187 MachineFunction &MF = DAG.getMachineFunction();
2188 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
2190 SmallVector<CCValAssign, 16> RVLocs;
2191 CCState CCInfo(CallConv, isVarArg, MF, RVLocs, *DAG.getContext());
2192 CCInfo.AnalyzeReturn(Outs, RetCC_X86);
2195 SmallVector<SDValue, 6> RetOps;
2196 RetOps.push_back(Chain); // Operand #0 = Chain (updated below)
2197 // Operand #1 = Bytes To Pop
2198 RetOps.push_back(DAG.getTargetConstant(FuncInfo->getBytesToPopOnReturn(), dl,
2201 // Copy the result values into the output registers.
2202 for (unsigned i = 0; i != RVLocs.size(); ++i) {
2203 CCValAssign &VA = RVLocs[i];
2204 assert(VA.isRegLoc() && "Can only return in registers!");
2205 SDValue ValToCopy = OutVals[i];
2206 EVT ValVT = ValToCopy.getValueType();
2208 // Promote values to the appropriate types.
2209 if (VA.getLocInfo() == CCValAssign::SExt)
2210 ValToCopy = DAG.getNode(ISD::SIGN_EXTEND, dl, VA.getLocVT(), ValToCopy);
2211 else if (VA.getLocInfo() == CCValAssign::ZExt)
2212 ValToCopy = DAG.getNode(ISD::ZERO_EXTEND, dl, VA.getLocVT(), ValToCopy);
2213 else if (VA.getLocInfo() == CCValAssign::AExt) {
2214 if (ValVT.isVector() && ValVT.getScalarType() == MVT::i1)
2215 ValToCopy = DAG.getNode(ISD::SIGN_EXTEND, dl, VA.getLocVT(), ValToCopy);
2217 ValToCopy = DAG.getNode(ISD::ANY_EXTEND, dl, VA.getLocVT(), ValToCopy);
2219 else if (VA.getLocInfo() == CCValAssign::BCvt)
2220 ValToCopy = DAG.getBitcast(VA.getLocVT(), ValToCopy);
2222 assert(VA.getLocInfo() != CCValAssign::FPExt &&
2223 "Unexpected FP-extend for return value.");
2225 // If this is x86-64, and we disabled SSE, we can't return FP values,
2226 // or SSE or MMX vectors.
2227 if ((ValVT == MVT::f32 || ValVT == MVT::f64 ||
2228 VA.getLocReg() == X86::XMM0 || VA.getLocReg() == X86::XMM1) &&
2229 (Subtarget->is64Bit() && !Subtarget->hasSSE1())) {
2230 report_fatal_error("SSE register return with SSE disabled");
2232 // Likewise we can't return F64 values with SSE1 only. gcc does so, but
2233 // llvm-gcc has never done it right and no one has noticed, so this
2234 // should be OK for now.
2235 if (ValVT == MVT::f64 &&
2236 (Subtarget->is64Bit() && !Subtarget->hasSSE2()))
2237 report_fatal_error("SSE2 register return with SSE2 disabled");
2239 // Returns in ST0/ST1 are handled specially: these are pushed as operands to
2240 // the RET instruction and handled by the FP Stackifier.
2241 if (VA.getLocReg() == X86::FP0 ||
2242 VA.getLocReg() == X86::FP1) {
2243 // If this is a copy from an xmm register to ST(0), use an FPExtend to
2244 // change the value to the FP stack register class.
2245 if (isScalarFPTypeInSSEReg(VA.getValVT()))
2246 ValToCopy = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f80, ValToCopy);
2247 RetOps.push_back(ValToCopy);
2248 // Don't emit a copytoreg.
2252 // 64-bit vector (MMX) values are returned in XMM0 / XMM1 except for v1i64
2253 // which is returned in RAX / RDX.
2254 if (Subtarget->is64Bit()) {
2255 if (ValVT == MVT::x86mmx) {
2256 if (VA.getLocReg() == X86::XMM0 || VA.getLocReg() == X86::XMM1) {
2257 ValToCopy = DAG.getBitcast(MVT::i64, ValToCopy);
2258 ValToCopy = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i64,
2260 // If we don't have SSE2 available, convert to v4f32 so the generated
2261 // register is legal.
2262 if (!Subtarget->hasSSE2())
2263 ValToCopy = DAG.getBitcast(MVT::v4f32, ValToCopy);
2268 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), ValToCopy, Flag);
2269 Flag = Chain.getValue(1);
2270 RetOps.push_back(DAG.getRegister(VA.getLocReg(), VA.getLocVT()));
2273 // All x86 ABIs require that for returning structs by value we copy
2274 // the sret argument into %rax/%eax (depending on ABI) for the return.
2275 // We saved the argument into a virtual register in the entry block,
2276 // so now we copy the value out and into %rax/%eax.
2278 // Checking Function.hasStructRetAttr() here is insufficient because the IR
2279 // may not have an explicit sret argument. If FuncInfo.CanLowerReturn is
2280 // false, then an sret argument may be implicitly inserted in the SelDAG. In
2281 // either case FuncInfo->setSRetReturnReg() will have been called.
2282 if (unsigned SRetReg = FuncInfo->getSRetReturnReg()) {
2283 SDValue Val = DAG.getCopyFromReg(Chain, dl, SRetReg,
2284 getPointerTy(MF.getDataLayout()));
2287 = (Subtarget->is64Bit() && !Subtarget->isTarget64BitILP32()) ?
2288 X86::RAX : X86::EAX;
2289 Chain = DAG.getCopyToReg(Chain, dl, RetValReg, Val, Flag);
2290 Flag = Chain.getValue(1);
2292 // RAX/EAX now acts like a return value.
2294 DAG.getRegister(RetValReg, getPointerTy(DAG.getDataLayout())));
2297 RetOps[0] = Chain; // Update chain.
2299 // Add the flag if we have it.
2301 RetOps.push_back(Flag);
2303 return DAG.getNode(X86ISD::RET_FLAG, dl, MVT::Other, RetOps);
2306 bool X86TargetLowering::isUsedByReturnOnly(SDNode *N, SDValue &Chain) const {
2307 if (N->getNumValues() != 1)
2309 if (!N->hasNUsesOfValue(1, 0))
2312 SDValue TCChain = Chain;
2313 SDNode *Copy = *N->use_begin();
2314 if (Copy->getOpcode() == ISD::CopyToReg) {
2315 // If the copy has a glue operand, we conservatively assume it isn't safe to
2316 // perform a tail call.
2317 if (Copy->getOperand(Copy->getNumOperands()-1).getValueType() == MVT::Glue)
2319 TCChain = Copy->getOperand(0);
2320 } else if (Copy->getOpcode() != ISD::FP_EXTEND)
2323 bool HasRet = false;
2324 for (SDNode::use_iterator UI = Copy->use_begin(), UE = Copy->use_end();
2326 if (UI->getOpcode() != X86ISD::RET_FLAG)
2328 // If we are returning more than one value, we can definitely
2329 // not make a tail call see PR19530
2330 if (UI->getNumOperands() > 4)
2332 if (UI->getNumOperands() == 4 &&
2333 UI->getOperand(UI->getNumOperands()-1).getValueType() != MVT::Glue)
2346 X86TargetLowering::getTypeForExtArgOrReturn(LLVMContext &Context, EVT VT,
2347 ISD::NodeType ExtendKind) const {
2349 // TODO: Is this also valid on 32-bit?
2350 if (Subtarget->is64Bit() && VT == MVT::i1 && ExtendKind == ISD::ZERO_EXTEND)
2351 ReturnMVT = MVT::i8;
2353 ReturnMVT = MVT::i32;
2355 EVT MinVT = getRegisterType(Context, ReturnMVT);
2356 return VT.bitsLT(MinVT) ? MinVT : VT;
2359 /// Lower the result values of a call into the
2360 /// appropriate copies out of appropriate physical registers.
2363 X86TargetLowering::LowerCallResult(SDValue Chain, SDValue InFlag,
2364 CallingConv::ID CallConv, bool isVarArg,
2365 const SmallVectorImpl<ISD::InputArg> &Ins,
2366 SDLoc dl, SelectionDAG &DAG,
2367 SmallVectorImpl<SDValue> &InVals) const {
2369 // Assign locations to each value returned by this call.
2370 SmallVector<CCValAssign, 16> RVLocs;
2371 bool Is64Bit = Subtarget->is64Bit();
2372 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(), RVLocs,
2374 CCInfo.AnalyzeCallResult(Ins, RetCC_X86);
2376 // Copy all of the result registers out of their specified physreg.
2377 for (unsigned i = 0, e = RVLocs.size(); i != e; ++i) {
2378 CCValAssign &VA = RVLocs[i];
2379 EVT CopyVT = VA.getLocVT();
2381 // If this is x86-64, and we disabled SSE, we can't return FP values
2382 if ((CopyVT == MVT::f32 || CopyVT == MVT::f64) &&
2383 ((Is64Bit || Ins[i].Flags.isInReg()) && !Subtarget->hasSSE1())) {
2384 report_fatal_error("SSE register return with SSE disabled");
2387 // If we prefer to use the value in xmm registers, copy it out as f80 and
2388 // use a truncate to move it from fp stack reg to xmm reg.
2389 bool RoundAfterCopy = false;
2390 if ((VA.getLocReg() == X86::FP0 || VA.getLocReg() == X86::FP1) &&
2391 isScalarFPTypeInSSEReg(VA.getValVT())) {
2393 RoundAfterCopy = (CopyVT != VA.getLocVT());
2396 Chain = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(),
2397 CopyVT, InFlag).getValue(1);
2398 SDValue Val = Chain.getValue(0);
2401 Val = DAG.getNode(ISD::FP_ROUND, dl, VA.getValVT(), Val,
2402 // This truncation won't change the value.
2403 DAG.getIntPtrConstant(1, dl));
2405 if (VA.isExtInLoc() && VA.getValVT().getScalarType() == MVT::i1)
2406 Val = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), Val);
2408 InFlag = Chain.getValue(2);
2409 InVals.push_back(Val);
2415 //===----------------------------------------------------------------------===//
2416 // C & StdCall & Fast Calling Convention implementation
2417 //===----------------------------------------------------------------------===//
2418 // StdCall calling convention seems to be standard for many Windows' API
2419 // routines and around. It differs from C calling convention just a little:
2420 // callee should clean up the stack, not caller. Symbols should be also
2421 // decorated in some fancy way :) It doesn't support any vector arguments.
2422 // For info on fast calling convention see Fast Calling Convention (tail call)
2423 // implementation LowerX86_32FastCCCallTo.
2425 /// CallIsStructReturn - Determines whether a call uses struct return
2427 enum StructReturnType {
2432 static StructReturnType
2433 callIsStructReturn(const SmallVectorImpl<ISD::OutputArg> &Outs) {
2435 return NotStructReturn;
2437 const ISD::ArgFlagsTy &Flags = Outs[0].Flags;
2438 if (!Flags.isSRet())
2439 return NotStructReturn;
2440 if (Flags.isInReg())
2441 return RegStructReturn;
2442 return StackStructReturn;
2445 /// Determines whether a function uses struct return semantics.
2446 static StructReturnType
2447 argsAreStructReturn(const SmallVectorImpl<ISD::InputArg> &Ins) {
2449 return NotStructReturn;
2451 const ISD::ArgFlagsTy &Flags = Ins[0].Flags;
2452 if (!Flags.isSRet())
2453 return NotStructReturn;
2454 if (Flags.isInReg())
2455 return RegStructReturn;
2456 return StackStructReturn;
2459 /// Make a copy of an aggregate at address specified by "Src" to address
2460 /// "Dst" with size and alignment information specified by the specific
2461 /// parameter attribute. The copy will be passed as a byval function parameter.
2463 CreateCopyOfByValArgument(SDValue Src, SDValue Dst, SDValue Chain,
2464 ISD::ArgFlagsTy Flags, SelectionDAG &DAG,
2466 SDValue SizeNode = DAG.getConstant(Flags.getByValSize(), dl, MVT::i32);
2468 return DAG.getMemcpy(Chain, dl, Dst, Src, SizeNode, Flags.getByValAlign(),
2469 /*isVolatile*/false, /*AlwaysInline=*/true,
2470 /*isTailCall*/false,
2471 MachinePointerInfo(), MachinePointerInfo());
2474 /// Return true if the calling convention is one that we can guarantee TCO for.
2475 static bool canGuaranteeTCO(CallingConv::ID CC) {
2476 return (CC == CallingConv::Fast || CC == CallingConv::GHC ||
2477 CC == CallingConv::HiPE || CC == CallingConv::HHVM);
2480 /// Return true if we might ever do TCO for calls with this calling convention.
2481 static bool mayTailCallThisCC(CallingConv::ID CC) {
2483 // C calling conventions:
2484 case CallingConv::C:
2485 case CallingConv::X86_64_Win64:
2486 case CallingConv::X86_64_SysV:
2487 // Callee pop conventions:
2488 case CallingConv::X86_ThisCall:
2489 case CallingConv::X86_StdCall:
2490 case CallingConv::X86_VectorCall:
2491 case CallingConv::X86_FastCall:
2494 return canGuaranteeTCO(CC);
2498 /// Return true if the function is being made into a tailcall target by
2499 /// changing its ABI.
2500 static bool shouldGuaranteeTCO(CallingConv::ID CC, bool GuaranteedTailCallOpt) {
2501 return GuaranteedTailCallOpt && canGuaranteeTCO(CC);
2504 bool X86TargetLowering::mayBeEmittedAsTailCall(CallInst *CI) const {
2506 CI->getParent()->getParent()->getFnAttribute("disable-tail-calls");
2507 if (!CI->isTailCall() || Attr.getValueAsString() == "true")
2511 CallingConv::ID CalleeCC = CS.getCallingConv();
2512 if (!mayTailCallThisCC(CalleeCC))
2519 X86TargetLowering::LowerMemArgument(SDValue Chain,
2520 CallingConv::ID CallConv,
2521 const SmallVectorImpl<ISD::InputArg> &Ins,
2522 SDLoc dl, SelectionDAG &DAG,
2523 const CCValAssign &VA,
2524 MachineFrameInfo *MFI,
2526 // Create the nodes corresponding to a load from this parameter slot.
2527 ISD::ArgFlagsTy Flags = Ins[i].Flags;
2528 bool AlwaysUseMutable = shouldGuaranteeTCO(
2529 CallConv, DAG.getTarget().Options.GuaranteedTailCallOpt);
2530 bool isImmutable = !AlwaysUseMutable && !Flags.isByVal();
2533 // If value is passed by pointer we have address passed instead of the value
2535 bool ExtendedInMem = VA.isExtInLoc() &&
2536 VA.getValVT().getScalarType() == MVT::i1;
2538 if (VA.getLocInfo() == CCValAssign::Indirect || ExtendedInMem)
2539 ValVT = VA.getLocVT();
2541 ValVT = VA.getValVT();
2543 // FIXME: For now, all byval parameter objects are marked mutable. This can be
2544 // changed with more analysis.
2545 // In case of tail call optimization mark all arguments mutable. Since they
2546 // could be overwritten by lowering of arguments in case of a tail call.
2547 if (Flags.isByVal()) {
2548 unsigned Bytes = Flags.getByValSize();
2549 if (Bytes == 0) Bytes = 1; // Don't create zero-sized stack objects.
2550 int FI = MFI->CreateFixedObject(Bytes, VA.getLocMemOffset(), isImmutable);
2551 return DAG.getFrameIndex(FI, getPointerTy(DAG.getDataLayout()));
2553 int FI = MFI->CreateFixedObject(ValVT.getSizeInBits()/8,
2554 VA.getLocMemOffset(), isImmutable);
2555 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy(DAG.getDataLayout()));
2556 SDValue Val = DAG.getLoad(
2557 ValVT, dl, Chain, FIN,
2558 MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), FI), false,
2560 return ExtendedInMem ?
2561 DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), Val) : Val;
2565 // FIXME: Get this from tablegen.
2566 static ArrayRef<MCPhysReg> get64BitArgumentGPRs(CallingConv::ID CallConv,
2567 const X86Subtarget *Subtarget) {
2568 assert(Subtarget->is64Bit());
2570 if (Subtarget->isCallingConvWin64(CallConv)) {
2571 static const MCPhysReg GPR64ArgRegsWin64[] = {
2572 X86::RCX, X86::RDX, X86::R8, X86::R9
2574 return makeArrayRef(std::begin(GPR64ArgRegsWin64), std::end(GPR64ArgRegsWin64));
2577 static const MCPhysReg GPR64ArgRegs64Bit[] = {
2578 X86::RDI, X86::RSI, X86::RDX, X86::RCX, X86::R8, X86::R9
2580 return makeArrayRef(std::begin(GPR64ArgRegs64Bit), std::end(GPR64ArgRegs64Bit));
2583 // FIXME: Get this from tablegen.
2584 static ArrayRef<MCPhysReg> get64BitArgumentXMMs(MachineFunction &MF,
2585 CallingConv::ID CallConv,
2586 const X86Subtarget *Subtarget) {
2587 assert(Subtarget->is64Bit());
2588 if (Subtarget->isCallingConvWin64(CallConv)) {
2589 // The XMM registers which might contain var arg parameters are shadowed
2590 // in their paired GPR. So we only need to save the GPR to their home
2592 // TODO: __vectorcall will change this.
2596 const Function *Fn = MF.getFunction();
2597 bool NoImplicitFloatOps = Fn->hasFnAttribute(Attribute::NoImplicitFloat);
2598 bool isSoftFloat = Subtarget->useSoftFloat();
2599 assert(!(isSoftFloat && NoImplicitFloatOps) &&
2600 "SSE register cannot be used when SSE is disabled!");
2601 if (isSoftFloat || NoImplicitFloatOps || !Subtarget->hasSSE1())
2602 // Kernel mode asks for SSE to be disabled, so there are no XMM argument
2606 static const MCPhysReg XMMArgRegs64Bit[] = {
2607 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
2608 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
2610 return makeArrayRef(std::begin(XMMArgRegs64Bit), std::end(XMMArgRegs64Bit));
2613 SDValue X86TargetLowering::LowerFormalArguments(
2614 SDValue Chain, CallingConv::ID CallConv, bool isVarArg,
2615 const SmallVectorImpl<ISD::InputArg> &Ins, SDLoc dl, SelectionDAG &DAG,
2616 SmallVectorImpl<SDValue> &InVals) const {
2617 MachineFunction &MF = DAG.getMachineFunction();
2618 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
2619 const TargetFrameLowering &TFI = *Subtarget->getFrameLowering();
2621 const Function* Fn = MF.getFunction();
2622 if (Fn->hasExternalLinkage() &&
2623 Subtarget->isTargetCygMing() &&
2624 Fn->getName() == "main")
2625 FuncInfo->setForceFramePointer(true);
2627 MachineFrameInfo *MFI = MF.getFrameInfo();
2628 bool Is64Bit = Subtarget->is64Bit();
2629 bool IsWin64 = Subtarget->isCallingConvWin64(CallConv);
2631 assert(!(isVarArg && canGuaranteeTCO(CallConv)) &&
2632 "Var args not supported with calling convention fastcc, ghc or hipe");
2634 // Assign locations to all of the incoming arguments.
2635 SmallVector<CCValAssign, 16> ArgLocs;
2636 CCState CCInfo(CallConv, isVarArg, MF, ArgLocs, *DAG.getContext());
2638 // Allocate shadow area for Win64
2640 CCInfo.AllocateStack(32, 8);
2642 CCInfo.AnalyzeFormalArguments(Ins, CC_X86);
2644 unsigned LastVal = ~0U;
2646 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2647 CCValAssign &VA = ArgLocs[i];
2648 // TODO: If an arg is passed in two places (e.g. reg and stack), skip later
2650 assert(VA.getValNo() != LastVal &&
2651 "Don't support value assigned to multiple locs yet");
2653 LastVal = VA.getValNo();
2655 if (VA.isRegLoc()) {
2656 EVT RegVT = VA.getLocVT();
2657 const TargetRegisterClass *RC;
2658 if (RegVT == MVT::i32)
2659 RC = &X86::GR32RegClass;
2660 else if (Is64Bit && RegVT == MVT::i64)
2661 RC = &X86::GR64RegClass;
2662 else if (RegVT == MVT::f32)
2663 RC = &X86::FR32RegClass;
2664 else if (RegVT == MVT::f64)
2665 RC = &X86::FR64RegClass;
2666 else if (RegVT.is512BitVector())
2667 RC = &X86::VR512RegClass;
2668 else if (RegVT.is256BitVector())
2669 RC = &X86::VR256RegClass;
2670 else if (RegVT.is128BitVector())
2671 RC = &X86::VR128RegClass;
2672 else if (RegVT == MVT::x86mmx)
2673 RC = &X86::VR64RegClass;
2674 else if (RegVT == MVT::i1)
2675 RC = &X86::VK1RegClass;
2676 else if (RegVT == MVT::v8i1)
2677 RC = &X86::VK8RegClass;
2678 else if (RegVT == MVT::v16i1)
2679 RC = &X86::VK16RegClass;
2680 else if (RegVT == MVT::v32i1)
2681 RC = &X86::VK32RegClass;
2682 else if (RegVT == MVT::v64i1)
2683 RC = &X86::VK64RegClass;
2685 llvm_unreachable("Unknown argument type!");
2687 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
2688 ArgValue = DAG.getCopyFromReg(Chain, dl, Reg, RegVT);
2690 // If this is an 8 or 16-bit value, it is really passed promoted to 32
2691 // bits. Insert an assert[sz]ext to capture this, then truncate to the
2693 if (VA.getLocInfo() == CCValAssign::SExt)
2694 ArgValue = DAG.getNode(ISD::AssertSext, dl, RegVT, ArgValue,
2695 DAG.getValueType(VA.getValVT()));
2696 else if (VA.getLocInfo() == CCValAssign::ZExt)
2697 ArgValue = DAG.getNode(ISD::AssertZext, dl, RegVT, ArgValue,
2698 DAG.getValueType(VA.getValVT()));
2699 else if (VA.getLocInfo() == CCValAssign::BCvt)
2700 ArgValue = DAG.getBitcast(VA.getValVT(), ArgValue);
2702 if (VA.isExtInLoc()) {
2703 // Handle MMX values passed in XMM regs.
2704 if (RegVT.isVector() && VA.getValVT().getScalarType() != MVT::i1)
2705 ArgValue = DAG.getNode(X86ISD::MOVDQ2Q, dl, VA.getValVT(), ArgValue);
2707 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
2710 assert(VA.isMemLoc());
2711 ArgValue = LowerMemArgument(Chain, CallConv, Ins, dl, DAG, VA, MFI, i);
2714 // If value is passed via pointer - do a load.
2715 if (VA.getLocInfo() == CCValAssign::Indirect)
2716 ArgValue = DAG.getLoad(VA.getValVT(), dl, Chain, ArgValue,
2717 MachinePointerInfo(), false, false, false, 0);
2719 InVals.push_back(ArgValue);
2722 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2723 // All x86 ABIs require that for returning structs by value we copy the
2724 // sret argument into %rax/%eax (depending on ABI) for the return. Save
2725 // the argument into a virtual register so that we can access it from the
2727 if (Ins[i].Flags.isSRet()) {
2728 unsigned Reg = FuncInfo->getSRetReturnReg();
2730 MVT PtrTy = getPointerTy(DAG.getDataLayout());
2731 Reg = MF.getRegInfo().createVirtualRegister(getRegClassFor(PtrTy));
2732 FuncInfo->setSRetReturnReg(Reg);
2734 SDValue Copy = DAG.getCopyToReg(DAG.getEntryNode(), dl, Reg, InVals[i]);
2735 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Copy, Chain);
2740 unsigned StackSize = CCInfo.getNextStackOffset();
2741 // Align stack specially for tail calls.
2742 if (shouldGuaranteeTCO(CallConv,
2743 MF.getTarget().Options.GuaranteedTailCallOpt))
2744 StackSize = GetAlignedArgumentStackSize(StackSize, DAG);
2746 // If the function takes variable number of arguments, make a frame index for
2747 // the start of the first vararg value... for expansion of llvm.va_start. We
2748 // can skip this if there are no va_start calls.
2749 if (MFI->hasVAStart() &&
2750 (Is64Bit || (CallConv != CallingConv::X86_FastCall &&
2751 CallConv != CallingConv::X86_ThisCall))) {
2752 FuncInfo->setVarArgsFrameIndex(
2753 MFI->CreateFixedObject(1, StackSize, true));
2756 MachineModuleInfo &MMI = MF.getMMI();
2758 // Figure out if XMM registers are in use.
2759 assert(!(Subtarget->useSoftFloat() &&
2760 Fn->hasFnAttribute(Attribute::NoImplicitFloat)) &&
2761 "SSE register cannot be used when SSE is disabled!");
2763 // 64-bit calling conventions support varargs and register parameters, so we
2764 // have to do extra work to spill them in the prologue.
2765 if (Is64Bit && isVarArg && MFI->hasVAStart()) {
2766 // Find the first unallocated argument registers.
2767 ArrayRef<MCPhysReg> ArgGPRs = get64BitArgumentGPRs(CallConv, Subtarget);
2768 ArrayRef<MCPhysReg> ArgXMMs = get64BitArgumentXMMs(MF, CallConv, Subtarget);
2769 unsigned NumIntRegs = CCInfo.getFirstUnallocated(ArgGPRs);
2770 unsigned NumXMMRegs = CCInfo.getFirstUnallocated(ArgXMMs);
2771 assert(!(NumXMMRegs && !Subtarget->hasSSE1()) &&
2772 "SSE register cannot be used when SSE is disabled!");
2774 // Gather all the live in physical registers.
2775 SmallVector<SDValue, 6> LiveGPRs;
2776 SmallVector<SDValue, 8> LiveXMMRegs;
2778 for (MCPhysReg Reg : ArgGPRs.slice(NumIntRegs)) {
2779 unsigned GPR = MF.addLiveIn(Reg, &X86::GR64RegClass);
2781 DAG.getCopyFromReg(Chain, dl, GPR, MVT::i64));
2783 if (!ArgXMMs.empty()) {
2784 unsigned AL = MF.addLiveIn(X86::AL, &X86::GR8RegClass);
2785 ALVal = DAG.getCopyFromReg(Chain, dl, AL, MVT::i8);
2786 for (MCPhysReg Reg : ArgXMMs.slice(NumXMMRegs)) {
2787 unsigned XMMReg = MF.addLiveIn(Reg, &X86::VR128RegClass);
2788 LiveXMMRegs.push_back(
2789 DAG.getCopyFromReg(Chain, dl, XMMReg, MVT::v4f32));
2794 // Get to the caller-allocated home save location. Add 8 to account
2795 // for the return address.
2796 int HomeOffset = TFI.getOffsetOfLocalArea() + 8;
2797 FuncInfo->setRegSaveFrameIndex(
2798 MFI->CreateFixedObject(1, NumIntRegs * 8 + HomeOffset, false));
2799 // Fixup to set vararg frame on shadow area (4 x i64).
2801 FuncInfo->setVarArgsFrameIndex(FuncInfo->getRegSaveFrameIndex());
2803 // For X86-64, if there are vararg parameters that are passed via
2804 // registers, then we must store them to their spots on the stack so
2805 // they may be loaded by deferencing the result of va_next.
2806 FuncInfo->setVarArgsGPOffset(NumIntRegs * 8);
2807 FuncInfo->setVarArgsFPOffset(ArgGPRs.size() * 8 + NumXMMRegs * 16);
2808 FuncInfo->setRegSaveFrameIndex(MFI->CreateStackObject(
2809 ArgGPRs.size() * 8 + ArgXMMs.size() * 16, 16, false));
2812 // Store the integer parameter registers.
2813 SmallVector<SDValue, 8> MemOps;
2814 SDValue RSFIN = DAG.getFrameIndex(FuncInfo->getRegSaveFrameIndex(),
2815 getPointerTy(DAG.getDataLayout()));
2816 unsigned Offset = FuncInfo->getVarArgsGPOffset();
2817 for (SDValue Val : LiveGPRs) {
2818 SDValue FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(DAG.getDataLayout()),
2819 RSFIN, DAG.getIntPtrConstant(Offset, dl));
2821 DAG.getStore(Val.getValue(1), dl, Val, FIN,
2822 MachinePointerInfo::getFixedStack(
2823 DAG.getMachineFunction(),
2824 FuncInfo->getRegSaveFrameIndex(), Offset),
2826 MemOps.push_back(Store);
2830 if (!ArgXMMs.empty() && NumXMMRegs != ArgXMMs.size()) {
2831 // Now store the XMM (fp + vector) parameter registers.
2832 SmallVector<SDValue, 12> SaveXMMOps;
2833 SaveXMMOps.push_back(Chain);
2834 SaveXMMOps.push_back(ALVal);
2835 SaveXMMOps.push_back(DAG.getIntPtrConstant(
2836 FuncInfo->getRegSaveFrameIndex(), dl));
2837 SaveXMMOps.push_back(DAG.getIntPtrConstant(
2838 FuncInfo->getVarArgsFPOffset(), dl));
2839 SaveXMMOps.insert(SaveXMMOps.end(), LiveXMMRegs.begin(),
2841 MemOps.push_back(DAG.getNode(X86ISD::VASTART_SAVE_XMM_REGS, dl,
2842 MVT::Other, SaveXMMOps));
2845 if (!MemOps.empty())
2846 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, MemOps);
2849 if (isVarArg && MFI->hasMustTailInVarArgFunc()) {
2850 // Find the largest legal vector type.
2851 MVT VecVT = MVT::Other;
2852 // FIXME: Only some x86_32 calling conventions support AVX512.
2853 if (Subtarget->hasAVX512() &&
2854 (Is64Bit || (CallConv == CallingConv::X86_VectorCall ||
2855 CallConv == CallingConv::Intel_OCL_BI)))
2856 VecVT = MVT::v16f32;
2857 else if (Subtarget->hasAVX())
2859 else if (Subtarget->hasSSE2())
2862 // We forward some GPRs and some vector types.
2863 SmallVector<MVT, 2> RegParmTypes;
2864 MVT IntVT = Is64Bit ? MVT::i64 : MVT::i32;
2865 RegParmTypes.push_back(IntVT);
2866 if (VecVT != MVT::Other)
2867 RegParmTypes.push_back(VecVT);
2869 // Compute the set of forwarded registers. The rest are scratch.
2870 SmallVectorImpl<ForwardedRegister> &Forwards =
2871 FuncInfo->getForwardedMustTailRegParms();
2872 CCInfo.analyzeMustTailForwardedRegisters(Forwards, RegParmTypes, CC_X86);
2874 // Conservatively forward AL on x86_64, since it might be used for varargs.
2875 if (Is64Bit && !CCInfo.isAllocated(X86::AL)) {
2876 unsigned ALVReg = MF.addLiveIn(X86::AL, &X86::GR8RegClass);
2877 Forwards.push_back(ForwardedRegister(ALVReg, X86::AL, MVT::i8));
2880 // Copy all forwards from physical to virtual registers.
2881 for (ForwardedRegister &F : Forwards) {
2882 // FIXME: Can we use a less constrained schedule?
2883 SDValue RegVal = DAG.getCopyFromReg(Chain, dl, F.VReg, F.VT);
2884 F.VReg = MF.getRegInfo().createVirtualRegister(getRegClassFor(F.VT));
2885 Chain = DAG.getCopyToReg(Chain, dl, F.VReg, RegVal);
2889 // Some CCs need callee pop.
2890 if (X86::isCalleePop(CallConv, Is64Bit, isVarArg,
2891 MF.getTarget().Options.GuaranteedTailCallOpt)) {
2892 FuncInfo->setBytesToPopOnReturn(StackSize); // Callee pops everything.
2894 FuncInfo->setBytesToPopOnReturn(0); // Callee pops nothing.
2895 // If this is an sret function, the return should pop the hidden pointer.
2896 if (!Is64Bit && !canGuaranteeTCO(CallConv) &&
2897 !Subtarget->getTargetTriple().isOSMSVCRT() &&
2898 argsAreStructReturn(Ins) == StackStructReturn)
2899 FuncInfo->setBytesToPopOnReturn(4);
2903 // RegSaveFrameIndex is X86-64 only.
2904 FuncInfo->setRegSaveFrameIndex(0xAAAAAAA);
2905 if (CallConv == CallingConv::X86_FastCall ||
2906 CallConv == CallingConv::X86_ThisCall)
2907 // fastcc functions can't have varargs.
2908 FuncInfo->setVarArgsFrameIndex(0xAAAAAAA);
2911 FuncInfo->setArgumentStackSize(StackSize);
2913 if (MMI.hasWinEHFuncInfo(Fn)) {
2915 int UnwindHelpFI = MFI->CreateStackObject(8, 8, /*isSS=*/false);
2916 SDValue StackSlot = DAG.getFrameIndex(UnwindHelpFI, MVT::i64);
2917 MMI.getWinEHFuncInfo(MF.getFunction()).UnwindHelpFrameIdx = UnwindHelpFI;
2918 SDValue Neg2 = DAG.getConstant(-2, dl, MVT::i64);
2919 Chain = DAG.getStore(Chain, dl, Neg2, StackSlot,
2920 MachinePointerInfo::getFixedStack(
2921 DAG.getMachineFunction(), UnwindHelpFI),
2922 /*isVolatile=*/true,
2923 /*isNonTemporal=*/false, /*Alignment=*/0);
2925 // Functions using Win32 EH are considered to have opaque SP adjustments
2926 // to force local variables to be addressed from the frame or base
2928 MFI->setHasOpaqueSPAdjustment(true);
2936 X86TargetLowering::LowerMemOpCallTo(SDValue Chain,
2937 SDValue StackPtr, SDValue Arg,
2938 SDLoc dl, SelectionDAG &DAG,
2939 const CCValAssign &VA,
2940 ISD::ArgFlagsTy Flags) const {
2941 unsigned LocMemOffset = VA.getLocMemOffset();
2942 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset, dl);
2943 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(DAG.getDataLayout()),
2945 if (Flags.isByVal())
2946 return CreateCopyOfByValArgument(Arg, PtrOff, Chain, Flags, DAG, dl);
2948 return DAG.getStore(
2949 Chain, dl, Arg, PtrOff,
2950 MachinePointerInfo::getStack(DAG.getMachineFunction(), LocMemOffset),
2954 /// Emit a load of return address if tail call
2955 /// optimization is performed and it is required.
2957 X86TargetLowering::EmitTailCallLoadRetAddr(SelectionDAG &DAG,
2958 SDValue &OutRetAddr, SDValue Chain,
2959 bool IsTailCall, bool Is64Bit,
2960 int FPDiff, SDLoc dl) const {
2961 // Adjust the Return address stack slot.
2962 EVT VT = getPointerTy(DAG.getDataLayout());
2963 OutRetAddr = getReturnAddressFrameIndex(DAG);
2965 // Load the "old" Return address.
2966 OutRetAddr = DAG.getLoad(VT, dl, Chain, OutRetAddr, MachinePointerInfo(),
2967 false, false, false, 0);
2968 return SDValue(OutRetAddr.getNode(), 1);
2971 /// Emit a store of the return address if tail call
2972 /// optimization is performed and it is required (FPDiff!=0).
2973 static SDValue EmitTailCallStoreRetAddr(SelectionDAG &DAG, MachineFunction &MF,
2974 SDValue Chain, SDValue RetAddrFrIdx,
2975 EVT PtrVT, unsigned SlotSize,
2976 int FPDiff, SDLoc dl) {
2977 // Store the return address to the appropriate stack slot.
2978 if (!FPDiff) return Chain;
2979 // Calculate the new stack slot for the return address.
2980 int NewReturnAddrFI =
2981 MF.getFrameInfo()->CreateFixedObject(SlotSize, (int64_t)FPDiff - SlotSize,
2983 SDValue NewRetAddrFrIdx = DAG.getFrameIndex(NewReturnAddrFI, PtrVT);
2984 Chain = DAG.getStore(Chain, dl, RetAddrFrIdx, NewRetAddrFrIdx,
2985 MachinePointerInfo::getFixedStack(
2986 DAG.getMachineFunction(), NewReturnAddrFI),
2991 /// Returns a vector_shuffle mask for an movs{s|d}, movd
2992 /// operation of specified width.
2993 static SDValue getMOVL(SelectionDAG &DAG, SDLoc dl, EVT VT, SDValue V1,
2995 unsigned NumElems = VT.getVectorNumElements();
2996 SmallVector<int, 8> Mask;
2997 Mask.push_back(NumElems);
2998 for (unsigned i = 1; i != NumElems; ++i)
3000 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
3004 X86TargetLowering::LowerCall(TargetLowering::CallLoweringInfo &CLI,
3005 SmallVectorImpl<SDValue> &InVals) const {
3006 SelectionDAG &DAG = CLI.DAG;
3008 SmallVectorImpl<ISD::OutputArg> &Outs = CLI.Outs;
3009 SmallVectorImpl<SDValue> &OutVals = CLI.OutVals;
3010 SmallVectorImpl<ISD::InputArg> &Ins = CLI.Ins;
3011 SDValue Chain = CLI.Chain;
3012 SDValue Callee = CLI.Callee;
3013 CallingConv::ID CallConv = CLI.CallConv;
3014 bool &isTailCall = CLI.IsTailCall;
3015 bool isVarArg = CLI.IsVarArg;
3017 MachineFunction &MF = DAG.getMachineFunction();
3018 bool Is64Bit = Subtarget->is64Bit();
3019 bool IsWin64 = Subtarget->isCallingConvWin64(CallConv);
3020 StructReturnType SR = callIsStructReturn(Outs);
3021 bool IsSibcall = false;
3022 X86MachineFunctionInfo *X86Info = MF.getInfo<X86MachineFunctionInfo>();
3023 auto Attr = MF.getFunction()->getFnAttribute("disable-tail-calls");
3025 if (Attr.getValueAsString() == "true")
3028 if (Subtarget->isPICStyleGOT() &&
3029 !MF.getTarget().Options.GuaranteedTailCallOpt) {
3030 // If we are using a GOT, disable tail calls to external symbols with
3031 // default visibility. Tail calling such a symbol requires using a GOT
3032 // relocation, which forces early binding of the symbol. This breaks code
3033 // that require lazy function symbol resolution. Using musttail or
3034 // GuaranteedTailCallOpt will override this.
3035 GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee);
3036 if (!G || (!G->getGlobal()->hasLocalLinkage() &&
3037 G->getGlobal()->hasDefaultVisibility()))
3041 bool IsMustTail = CLI.CS && CLI.CS->isMustTailCall();
3043 // Force this to be a tail call. The verifier rules are enough to ensure
3044 // that we can lower this successfully without moving the return address
3047 } else if (isTailCall) {
3048 // Check if it's really possible to do a tail call.
3049 isTailCall = IsEligibleForTailCallOptimization(Callee, CallConv,
3050 isVarArg, SR != NotStructReturn,
3051 MF.getFunction()->hasStructRetAttr(), CLI.RetTy,
3052 Outs, OutVals, Ins, DAG);
3054 // Sibcalls are automatically detected tailcalls which do not require
3056 if (!MF.getTarget().Options.GuaranteedTailCallOpt && isTailCall)
3063 assert(!(isVarArg && canGuaranteeTCO(CallConv)) &&
3064 "Var args not supported with calling convention fastcc, ghc or hipe");
3066 // Analyze operands of the call, assigning locations to each operand.
3067 SmallVector<CCValAssign, 16> ArgLocs;
3068 CCState CCInfo(CallConv, isVarArg, MF, ArgLocs, *DAG.getContext());
3070 // Allocate shadow area for Win64
3072 CCInfo.AllocateStack(32, 8);
3074 CCInfo.AnalyzeCallOperands(Outs, CC_X86);
3076 // Get a count of how many bytes are to be pushed on the stack.
3077 unsigned NumBytes = CCInfo.getAlignedCallFrameSize();
3079 // This is a sibcall. The memory operands are available in caller's
3080 // own caller's stack.
3082 else if (MF.getTarget().Options.GuaranteedTailCallOpt &&
3083 canGuaranteeTCO(CallConv))
3084 NumBytes = GetAlignedArgumentStackSize(NumBytes, DAG);
3087 if (isTailCall && !IsSibcall && !IsMustTail) {
3088 // Lower arguments at fp - stackoffset + fpdiff.
3089 unsigned NumBytesCallerPushed = X86Info->getBytesToPopOnReturn();
3091 FPDiff = NumBytesCallerPushed - NumBytes;
3093 // Set the delta of movement of the returnaddr stackslot.
3094 // But only set if delta is greater than previous delta.
3095 if (FPDiff < X86Info->getTCReturnAddrDelta())
3096 X86Info->setTCReturnAddrDelta(FPDiff);
3099 unsigned NumBytesToPush = NumBytes;
3100 unsigned NumBytesToPop = NumBytes;
3102 // If we have an inalloca argument, all stack space has already been allocated
3103 // for us and be right at the top of the stack. We don't support multiple
3104 // arguments passed in memory when using inalloca.
3105 if (!Outs.empty() && Outs.back().Flags.isInAlloca()) {
3107 if (!ArgLocs.back().isMemLoc())
3108 report_fatal_error("cannot use inalloca attribute on a register "
3110 if (ArgLocs.back().getLocMemOffset() != 0)
3111 report_fatal_error("any parameter with the inalloca attribute must be "
3112 "the only memory argument");
3116 Chain = DAG.getCALLSEQ_START(
3117 Chain, DAG.getIntPtrConstant(NumBytesToPush, dl, true), dl);
3119 SDValue RetAddrFrIdx;
3120 // Load return address for tail calls.
3121 if (isTailCall && FPDiff)
3122 Chain = EmitTailCallLoadRetAddr(DAG, RetAddrFrIdx, Chain, isTailCall,
3123 Is64Bit, FPDiff, dl);
3125 SmallVector<std::pair<unsigned, SDValue>, 8> RegsToPass;
3126 SmallVector<SDValue, 8> MemOpChains;
3129 // Walk the register/memloc assignments, inserting copies/loads. In the case
3130 // of tail call optimization arguments are handle later.
3131 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
3132 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
3133 // Skip inalloca arguments, they have already been written.
3134 ISD::ArgFlagsTy Flags = Outs[i].Flags;
3135 if (Flags.isInAlloca())
3138 CCValAssign &VA = ArgLocs[i];
3139 EVT RegVT = VA.getLocVT();
3140 SDValue Arg = OutVals[i];
3141 bool isByVal = Flags.isByVal();
3143 // Promote the value if needed.
3144 switch (VA.getLocInfo()) {
3145 default: llvm_unreachable("Unknown loc info!");
3146 case CCValAssign::Full: break;
3147 case CCValAssign::SExt:
3148 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, RegVT, Arg);
3150 case CCValAssign::ZExt:
3151 Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, RegVT, Arg);
3153 case CCValAssign::AExt:
3154 if (Arg.getValueType().isVector() &&
3155 Arg.getValueType().getScalarType() == MVT::i1)
3156 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, RegVT, Arg);
3157 else if (RegVT.is128BitVector()) {
3158 // Special case: passing MMX values in XMM registers.
3159 Arg = DAG.getBitcast(MVT::i64, Arg);
3160 Arg = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i64, Arg);
3161 Arg = getMOVL(DAG, dl, MVT::v2i64, DAG.getUNDEF(MVT::v2i64), Arg);
3163 Arg = DAG.getNode(ISD::ANY_EXTEND, dl, RegVT, Arg);
3165 case CCValAssign::BCvt:
3166 Arg = DAG.getBitcast(RegVT, Arg);
3168 case CCValAssign::Indirect: {
3169 // Store the argument.
3170 SDValue SpillSlot = DAG.CreateStackTemporary(VA.getValVT());
3171 int FI = cast<FrameIndexSDNode>(SpillSlot)->getIndex();
3172 Chain = DAG.getStore(
3173 Chain, dl, Arg, SpillSlot,
3174 MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), FI),
3181 if (VA.isRegLoc()) {
3182 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
3183 if (isVarArg && IsWin64) {
3184 // Win64 ABI requires argument XMM reg to be copied to the corresponding
3185 // shadow reg if callee is a varargs function.
3186 unsigned ShadowReg = 0;
3187 switch (VA.getLocReg()) {
3188 case X86::XMM0: ShadowReg = X86::RCX; break;
3189 case X86::XMM1: ShadowReg = X86::RDX; break;
3190 case X86::XMM2: ShadowReg = X86::R8; break;
3191 case X86::XMM3: ShadowReg = X86::R9; break;
3194 RegsToPass.push_back(std::make_pair(ShadowReg, Arg));
3196 } else if (!IsSibcall && (!isTailCall || isByVal)) {
3197 assert(VA.isMemLoc());
3198 if (!StackPtr.getNode())
3199 StackPtr = DAG.getCopyFromReg(Chain, dl, RegInfo->getStackRegister(),
3200 getPointerTy(DAG.getDataLayout()));
3201 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Arg,
3202 dl, DAG, VA, Flags));
3206 if (!MemOpChains.empty())
3207 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, MemOpChains);
3209 if (Subtarget->isPICStyleGOT()) {
3210 // ELF / PIC requires GOT in the EBX register before function calls via PLT
3213 RegsToPass.push_back(std::make_pair(
3214 unsigned(X86::EBX), DAG.getNode(X86ISD::GlobalBaseReg, SDLoc(),
3215 getPointerTy(DAG.getDataLayout()))));
3217 // If we are tail calling and generating PIC/GOT style code load the
3218 // address of the callee into ECX. The value in ecx is used as target of
3219 // the tail jump. This is done to circumvent the ebx/callee-saved problem
3220 // for tail calls on PIC/GOT architectures. Normally we would just put the
3221 // address of GOT into ebx and then call target@PLT. But for tail calls
3222 // ebx would be restored (since ebx is callee saved) before jumping to the
3225 // Note: The actual moving to ECX is done further down.
3226 GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee);
3227 if (G && !G->getGlobal()->hasLocalLinkage() &&
3228 G->getGlobal()->hasDefaultVisibility())
3229 Callee = LowerGlobalAddress(Callee, DAG);
3230 else if (isa<ExternalSymbolSDNode>(Callee))
3231 Callee = LowerExternalSymbol(Callee, DAG);
3235 if (Is64Bit && isVarArg && !IsWin64 && !IsMustTail) {
3236 // From AMD64 ABI document:
3237 // For calls that may call functions that use varargs or stdargs
3238 // (prototype-less calls or calls to functions containing ellipsis (...) in
3239 // the declaration) %al is used as hidden argument to specify the number
3240 // of SSE registers used. The contents of %al do not need to match exactly
3241 // the number of registers, but must be an ubound on the number of SSE
3242 // registers used and is in the range 0 - 8 inclusive.
3244 // Count the number of XMM registers allocated.
3245 static const MCPhysReg XMMArgRegs[] = {
3246 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
3247 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
3249 unsigned NumXMMRegs = CCInfo.getFirstUnallocated(XMMArgRegs);
3250 assert((Subtarget->hasSSE1() || !NumXMMRegs)
3251 && "SSE registers cannot be used when SSE is disabled");
3253 RegsToPass.push_back(std::make_pair(unsigned(X86::AL),
3254 DAG.getConstant(NumXMMRegs, dl,
3258 if (isVarArg && IsMustTail) {
3259 const auto &Forwards = X86Info->getForwardedMustTailRegParms();
3260 for (const auto &F : Forwards) {
3261 SDValue Val = DAG.getCopyFromReg(Chain, dl, F.VReg, F.VT);
3262 RegsToPass.push_back(std::make_pair(unsigned(F.PReg), Val));
3266 // For tail calls lower the arguments to the 'real' stack slots. Sibcalls
3267 // don't need this because the eligibility check rejects calls that require
3268 // shuffling arguments passed in memory.
3269 if (!IsSibcall && isTailCall) {
3270 // Force all the incoming stack arguments to be loaded from the stack
3271 // before any new outgoing arguments are stored to the stack, because the
3272 // outgoing stack slots may alias the incoming argument stack slots, and
3273 // the alias isn't otherwise explicit. This is slightly more conservative
3274 // than necessary, because it means that each store effectively depends
3275 // on every argument instead of just those arguments it would clobber.
3276 SDValue ArgChain = DAG.getStackArgumentTokenFactor(Chain);
3278 SmallVector<SDValue, 8> MemOpChains2;
3281 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
3282 CCValAssign &VA = ArgLocs[i];
3285 assert(VA.isMemLoc());
3286 SDValue Arg = OutVals[i];
3287 ISD::ArgFlagsTy Flags = Outs[i].Flags;
3288 // Skip inalloca arguments. They don't require any work.
3289 if (Flags.isInAlloca())
3291 // Create frame index.
3292 int32_t Offset = VA.getLocMemOffset()+FPDiff;
3293 uint32_t OpSize = (VA.getLocVT().getSizeInBits()+7)/8;
3294 FI = MF.getFrameInfo()->CreateFixedObject(OpSize, Offset, true);
3295 FIN = DAG.getFrameIndex(FI, getPointerTy(DAG.getDataLayout()));
3297 if (Flags.isByVal()) {
3298 // Copy relative to framepointer.
3299 SDValue Source = DAG.getIntPtrConstant(VA.getLocMemOffset(), dl);
3300 if (!StackPtr.getNode())
3301 StackPtr = DAG.getCopyFromReg(Chain, dl, RegInfo->getStackRegister(),
3302 getPointerTy(DAG.getDataLayout()));
3303 Source = DAG.getNode(ISD::ADD, dl, getPointerTy(DAG.getDataLayout()),
3306 MemOpChains2.push_back(CreateCopyOfByValArgument(Source, FIN,
3310 // Store relative to framepointer.
3311 MemOpChains2.push_back(DAG.getStore(
3312 ArgChain, dl, Arg, FIN,
3313 MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), FI),
3318 if (!MemOpChains2.empty())
3319 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, MemOpChains2);
3321 // Store the return address to the appropriate stack slot.
3322 Chain = EmitTailCallStoreRetAddr(DAG, MF, Chain, RetAddrFrIdx,
3323 getPointerTy(DAG.getDataLayout()),
3324 RegInfo->getSlotSize(), FPDiff, dl);
3327 // Build a sequence of copy-to-reg nodes chained together with token chain
3328 // and flag operands which copy the outgoing args into registers.
3330 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
3331 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
3332 RegsToPass[i].second, InFlag);
3333 InFlag = Chain.getValue(1);
3336 if (DAG.getTarget().getCodeModel() == CodeModel::Large) {
3337 assert(Is64Bit && "Large code model is only legal in 64-bit mode.");
3338 // In the 64-bit large code model, we have to make all calls
3339 // through a register, since the call instruction's 32-bit
3340 // pc-relative offset may not be large enough to hold the whole
3342 } else if (Callee->getOpcode() == ISD::GlobalAddress) {
3343 // If the callee is a GlobalAddress node (quite common, every direct call
3344 // is) turn it into a TargetGlobalAddress node so that legalize doesn't hack
3346 GlobalAddressSDNode* G = cast<GlobalAddressSDNode>(Callee);
3348 // We should use extra load for direct calls to dllimported functions in
3350 const GlobalValue *GV = G->getGlobal();
3351 if (!GV->hasDLLImportStorageClass()) {
3352 unsigned char OpFlags = 0;
3353 bool ExtraLoad = false;
3354 unsigned WrapperKind = ISD::DELETED_NODE;
3356 // On ELF targets, in both X86-64 and X86-32 mode, direct calls to
3357 // external symbols most go through the PLT in PIC mode. If the symbol
3358 // has hidden or protected visibility, or if it is static or local, then
3359 // we don't need to use the PLT - we can directly call it.
3360 if (Subtarget->isTargetELF() &&
3361 DAG.getTarget().getRelocationModel() == Reloc::PIC_ &&
3362 GV->hasDefaultVisibility() && !GV->hasLocalLinkage()) {
3363 OpFlags = X86II::MO_PLT;
3364 } else if (Subtarget->isPICStyleStubAny() &&
3365 !GV->isStrongDefinitionForLinker() &&
3366 (!Subtarget->getTargetTriple().isMacOSX() ||
3367 Subtarget->getTargetTriple().isMacOSXVersionLT(10, 5))) {
3368 // PC-relative references to external symbols should go through $stub,
3369 // unless we're building with the leopard linker or later, which
3370 // automatically synthesizes these stubs.
3371 OpFlags = X86II::MO_DARWIN_STUB;
3372 } else if (Subtarget->isPICStyleRIPRel() && isa<Function>(GV) &&
3373 cast<Function>(GV)->hasFnAttribute(Attribute::NonLazyBind)) {
3374 // If the function is marked as non-lazy, generate an indirect call
3375 // which loads from the GOT directly. This avoids runtime overhead
3376 // at the cost of eager binding (and one extra byte of encoding).
3377 OpFlags = X86II::MO_GOTPCREL;
3378 WrapperKind = X86ISD::WrapperRIP;
3382 Callee = DAG.getTargetGlobalAddress(
3383 GV, dl, getPointerTy(DAG.getDataLayout()), G->getOffset(), OpFlags);
3385 // Add a wrapper if needed.
3386 if (WrapperKind != ISD::DELETED_NODE)
3387 Callee = DAG.getNode(X86ISD::WrapperRIP, dl,
3388 getPointerTy(DAG.getDataLayout()), Callee);
3389 // Add extra indirection if needed.
3391 Callee = DAG.getLoad(
3392 getPointerTy(DAG.getDataLayout()), dl, DAG.getEntryNode(), Callee,
3393 MachinePointerInfo::getGOT(DAG.getMachineFunction()), false, false,
3396 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
3397 unsigned char OpFlags = 0;
3399 // On ELF targets, in either X86-64 or X86-32 mode, direct calls to
3400 // external symbols should go through the PLT.
3401 if (Subtarget->isTargetELF() &&
3402 DAG.getTarget().getRelocationModel() == Reloc::PIC_) {
3403 OpFlags = X86II::MO_PLT;
3404 } else if (Subtarget->isPICStyleStubAny() &&
3405 (!Subtarget->getTargetTriple().isMacOSX() ||
3406 Subtarget->getTargetTriple().isMacOSXVersionLT(10, 5))) {
3407 // PC-relative references to external symbols should go through $stub,
3408 // unless we're building with the leopard linker or later, which
3409 // automatically synthesizes these stubs.
3410 OpFlags = X86II::MO_DARWIN_STUB;
3413 Callee = DAG.getTargetExternalSymbol(
3414 S->getSymbol(), getPointerTy(DAG.getDataLayout()), OpFlags);
3415 } else if (Subtarget->isTarget64BitILP32() &&
3416 Callee->getValueType(0) == MVT::i32) {
3417 // Zero-extend the 32-bit Callee address into a 64-bit according to x32 ABI
3418 Callee = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i64, Callee);
3421 // Returns a chain & a flag for retval copy to use.
3422 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
3423 SmallVector<SDValue, 8> Ops;
3425 if (!IsSibcall && isTailCall) {
3426 Chain = DAG.getCALLSEQ_END(Chain,
3427 DAG.getIntPtrConstant(NumBytesToPop, dl, true),
3428 DAG.getIntPtrConstant(0, dl, true), InFlag, dl);
3429 InFlag = Chain.getValue(1);
3432 Ops.push_back(Chain);
3433 Ops.push_back(Callee);
3436 Ops.push_back(DAG.getConstant(FPDiff, dl, MVT::i32));
3438 // Add argument registers to the end of the list so that they are known live
3440 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
3441 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
3442 RegsToPass[i].second.getValueType()));
3444 // Add a register mask operand representing the call-preserved registers.
3445 const uint32_t *Mask = RegInfo->getCallPreservedMask(MF, CallConv);
3446 assert(Mask && "Missing call preserved mask for calling convention");
3448 // If this is an invoke in a 32-bit function using a funclet-based
3449 // personality, assume the function clobbers all registers. If an exception
3450 // is thrown, the runtime will not restore CSRs.
3451 // FIXME: Model this more precisely so that we can register allocate across
3452 // the normal edge and spill and fill across the exceptional edge.
3453 if (!Is64Bit && CLI.CS && CLI.CS->isInvoke()) {
3454 const Function *CallerFn = MF.getFunction();
3455 EHPersonality Pers =
3456 CallerFn->hasPersonalityFn()
3457 ? classifyEHPersonality(CallerFn->getPersonalityFn())
3458 : EHPersonality::Unknown;
3459 if (isFuncletEHPersonality(Pers))
3460 Mask = RegInfo->getNoPreservedMask();
3463 Ops.push_back(DAG.getRegisterMask(Mask));
3465 if (InFlag.getNode())
3466 Ops.push_back(InFlag);
3470 //// If this is the first return lowered for this function, add the regs
3471 //// to the liveout set for the function.
3472 // This isn't right, although it's probably harmless on x86; liveouts
3473 // should be computed from returns not tail calls. Consider a void
3474 // function making a tail call to a function returning int.
3475 MF.getFrameInfo()->setHasTailCall();
3476 return DAG.getNode(X86ISD::TC_RETURN, dl, NodeTys, Ops);
3479 Chain = DAG.getNode(X86ISD::CALL, dl, NodeTys, Ops);
3480 InFlag = Chain.getValue(1);
3482 // Create the CALLSEQ_END node.
3483 unsigned NumBytesForCalleeToPop;
3484 if (X86::isCalleePop(CallConv, Is64Bit, isVarArg,
3485 DAG.getTarget().Options.GuaranteedTailCallOpt))
3486 NumBytesForCalleeToPop = NumBytes; // Callee pops everything
3487 else if (!Is64Bit && !canGuaranteeTCO(CallConv) &&
3488 !Subtarget->getTargetTriple().isOSMSVCRT() &&
3489 SR == StackStructReturn)
3490 // If this is a call to a struct-return function, the callee
3491 // pops the hidden struct pointer, so we have to push it back.
3492 // This is common for Darwin/X86, Linux & Mingw32 targets.
3493 // For MSVC Win32 targets, the caller pops the hidden struct pointer.
3494 NumBytesForCalleeToPop = 4;
3496 NumBytesForCalleeToPop = 0; // Callee pops nothing.
3498 // Returns a flag for retval copy to use.
3500 Chain = DAG.getCALLSEQ_END(Chain,
3501 DAG.getIntPtrConstant(NumBytesToPop, dl, true),
3502 DAG.getIntPtrConstant(NumBytesForCalleeToPop, dl,
3505 InFlag = Chain.getValue(1);
3508 // Handle result values, copying them out of physregs into vregs that we
3510 return LowerCallResult(Chain, InFlag, CallConv, isVarArg,
3511 Ins, dl, DAG, InVals);
3514 //===----------------------------------------------------------------------===//
3515 // Fast Calling Convention (tail call) implementation
3516 //===----------------------------------------------------------------------===//
3518 // Like std call, callee cleans arguments, convention except that ECX is
3519 // reserved for storing the tail called function address. Only 2 registers are
3520 // free for argument passing (inreg). Tail call optimization is performed
3522 // * tailcallopt is enabled
3523 // * caller/callee are fastcc
3524 // On X86_64 architecture with GOT-style position independent code only local
3525 // (within module) calls are supported at the moment.
3526 // To keep the stack aligned according to platform abi the function
3527 // GetAlignedArgumentStackSize ensures that argument delta is always multiples
3528 // of stack alignment. (Dynamic linkers need this - darwin's dyld for example)
3529 // If a tail called function callee has more arguments than the caller the
3530 // caller needs to make sure that there is room to move the RETADDR to. This is
3531 // achieved by reserving an area the size of the argument delta right after the
3532 // original RETADDR, but before the saved framepointer or the spilled registers
3533 // e.g. caller(arg1, arg2) calls callee(arg1, arg2,arg3,arg4)
3545 /// Make the stack size align e.g 16n + 12 aligned for a 16-byte align
3548 X86TargetLowering::GetAlignedArgumentStackSize(unsigned StackSize,
3549 SelectionDAG& DAG) const {
3550 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
3551 const TargetFrameLowering &TFI = *Subtarget->getFrameLowering();
3552 unsigned StackAlignment = TFI.getStackAlignment();
3553 uint64_t AlignMask = StackAlignment - 1;
3554 int64_t Offset = StackSize;
3555 unsigned SlotSize = RegInfo->getSlotSize();
3556 if ( (Offset & AlignMask) <= (StackAlignment - SlotSize) ) {
3557 // Number smaller than 12 so just add the difference.
3558 Offset += ((StackAlignment - SlotSize) - (Offset & AlignMask));
3560 // Mask out lower bits, add stackalignment once plus the 12 bytes.
3561 Offset = ((~AlignMask) & Offset) + StackAlignment +
3562 (StackAlignment-SlotSize);
3567 /// Return true if the given stack call argument is already available in the
3568 /// same position (relatively) of the caller's incoming argument stack.
3570 bool MatchingStackOffset(SDValue Arg, unsigned Offset, ISD::ArgFlagsTy Flags,
3571 MachineFrameInfo *MFI, const MachineRegisterInfo *MRI,
3572 const X86InstrInfo *TII) {
3573 unsigned Bytes = Arg.getValueType().getSizeInBits() / 8;
3575 if (Arg.getOpcode() == ISD::CopyFromReg) {
3576 unsigned VR = cast<RegisterSDNode>(Arg.getOperand(1))->getReg();
3577 if (!TargetRegisterInfo::isVirtualRegister(VR))
3579 MachineInstr *Def = MRI->getVRegDef(VR);
3582 if (!Flags.isByVal()) {
3583 if (!TII->isLoadFromStackSlot(Def, FI))
3586 unsigned Opcode = Def->getOpcode();
3587 if ((Opcode == X86::LEA32r || Opcode == X86::LEA64r ||
3588 Opcode == X86::LEA64_32r) &&
3589 Def->getOperand(1).isFI()) {
3590 FI = Def->getOperand(1).getIndex();
3591 Bytes = Flags.getByValSize();
3595 } else if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Arg)) {
3596 if (Flags.isByVal())
3597 // ByVal argument is passed in as a pointer but it's now being
3598 // dereferenced. e.g.
3599 // define @foo(%struct.X* %A) {
3600 // tail call @bar(%struct.X* byval %A)
3603 SDValue Ptr = Ld->getBasePtr();
3604 FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr);
3607 FI = FINode->getIndex();
3608 } else if (Arg.getOpcode() == ISD::FrameIndex && Flags.isByVal()) {
3609 FrameIndexSDNode *FINode = cast<FrameIndexSDNode>(Arg);
3610 FI = FINode->getIndex();
3611 Bytes = Flags.getByValSize();
3615 assert(FI != INT_MAX);
3616 if (!MFI->isFixedObjectIndex(FI))
3618 return Offset == MFI->getObjectOffset(FI) && Bytes == MFI->getObjectSize(FI);
3621 /// Check whether the call is eligible for tail call optimization. Targets
3622 /// that want to do tail call optimization should implement this function.
3623 bool X86TargetLowering::IsEligibleForTailCallOptimization(
3624 SDValue Callee, CallingConv::ID CalleeCC, bool isVarArg,
3625 bool isCalleeStructRet, bool isCallerStructRet, Type *RetTy,
3626 const SmallVectorImpl<ISD::OutputArg> &Outs,
3627 const SmallVectorImpl<SDValue> &OutVals,
3628 const SmallVectorImpl<ISD::InputArg> &Ins, SelectionDAG &DAG) const {
3629 if (!mayTailCallThisCC(CalleeCC))
3632 // If -tailcallopt is specified, make fastcc functions tail-callable.
3633 MachineFunction &MF = DAG.getMachineFunction();
3634 const Function *CallerF = MF.getFunction();
3636 // If the function return type is x86_fp80 and the callee return type is not,
3637 // then the FP_EXTEND of the call result is not a nop. It's not safe to
3638 // perform a tailcall optimization here.
3639 if (CallerF->getReturnType()->isX86_FP80Ty() && !RetTy->isX86_FP80Ty())
3642 CallingConv::ID CallerCC = CallerF->getCallingConv();
3643 bool CCMatch = CallerCC == CalleeCC;
3644 bool IsCalleeWin64 = Subtarget->isCallingConvWin64(CalleeCC);
3645 bool IsCallerWin64 = Subtarget->isCallingConvWin64(CallerCC);
3647 // Win64 functions have extra shadow space for argument homing. Don't do the
3648 // sibcall if the caller and callee have mismatched expectations for this
3650 if (IsCalleeWin64 != IsCallerWin64)
3653 if (DAG.getTarget().Options.GuaranteedTailCallOpt) {
3654 if (canGuaranteeTCO(CalleeCC) && CCMatch)
3659 // Look for obvious safe cases to perform tail call optimization that do not
3660 // require ABI changes. This is what gcc calls sibcall.
3662 // Can't do sibcall if stack needs to be dynamically re-aligned. PEI needs to
3663 // emit a special epilogue.
3664 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
3665 if (RegInfo->needsStackRealignment(MF))
3668 // Also avoid sibcall optimization if either caller or callee uses struct
3669 // return semantics.
3670 if (isCalleeStructRet || isCallerStructRet)
3673 // Do not sibcall optimize vararg calls unless all arguments are passed via
3675 if (isVarArg && !Outs.empty()) {
3676 // Optimizing for varargs on Win64 is unlikely to be safe without
3677 // additional testing.
3678 if (IsCalleeWin64 || IsCallerWin64)
3681 SmallVector<CCValAssign, 16> ArgLocs;
3682 CCState CCInfo(CalleeCC, isVarArg, DAG.getMachineFunction(), ArgLocs,
3685 CCInfo.AnalyzeCallOperands(Outs, CC_X86);
3686 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i)
3687 if (!ArgLocs[i].isRegLoc())
3691 // If the call result is in ST0 / ST1, it needs to be popped off the x87
3692 // stack. Therefore, if it's not used by the call it is not safe to optimize
3693 // this into a sibcall.
3694 bool Unused = false;
3695 for (unsigned i = 0, e = Ins.size(); i != e; ++i) {
3702 SmallVector<CCValAssign, 16> RVLocs;
3703 CCState CCInfo(CalleeCC, false, DAG.getMachineFunction(), RVLocs,
3705 CCInfo.AnalyzeCallResult(Ins, RetCC_X86);
3706 for (unsigned i = 0, e = RVLocs.size(); i != e; ++i) {
3707 CCValAssign &VA = RVLocs[i];
3708 if (VA.getLocReg() == X86::FP0 || VA.getLocReg() == X86::FP1)
3713 // If the calling conventions do not match, then we'd better make sure the
3714 // results are returned in the same way as what the caller expects.
3716 SmallVector<CCValAssign, 16> RVLocs1;
3717 CCState CCInfo1(CalleeCC, false, DAG.getMachineFunction(), RVLocs1,
3719 CCInfo1.AnalyzeCallResult(Ins, RetCC_X86);
3721 SmallVector<CCValAssign, 16> RVLocs2;
3722 CCState CCInfo2(CallerCC, false, DAG.getMachineFunction(), RVLocs2,
3724 CCInfo2.AnalyzeCallResult(Ins, RetCC_X86);
3726 if (RVLocs1.size() != RVLocs2.size())
3728 for (unsigned i = 0, e = RVLocs1.size(); i != e; ++i) {
3729 if (RVLocs1[i].isRegLoc() != RVLocs2[i].isRegLoc())
3731 if (RVLocs1[i].getLocInfo() != RVLocs2[i].getLocInfo())
3733 if (RVLocs1[i].isRegLoc()) {
3734 if (RVLocs1[i].getLocReg() != RVLocs2[i].getLocReg())
3737 if (RVLocs1[i].getLocMemOffset() != RVLocs2[i].getLocMemOffset())
3743 unsigned StackArgsSize = 0;
3745 // If the callee takes no arguments then go on to check the results of the
3747 if (!Outs.empty()) {
3748 // Check if stack adjustment is needed. For now, do not do this if any
3749 // argument is passed on the stack.
3750 SmallVector<CCValAssign, 16> ArgLocs;
3751 CCState CCInfo(CalleeCC, isVarArg, DAG.getMachineFunction(), ArgLocs,
3754 // Allocate shadow area for Win64
3756 CCInfo.AllocateStack(32, 8);
3758 CCInfo.AnalyzeCallOperands(Outs, CC_X86);
3759 StackArgsSize = CCInfo.getNextStackOffset();
3761 if (CCInfo.getNextStackOffset()) {
3762 // Check if the arguments are already laid out in the right way as
3763 // the caller's fixed stack objects.
3764 MachineFrameInfo *MFI = MF.getFrameInfo();
3765 const MachineRegisterInfo *MRI = &MF.getRegInfo();
3766 const X86InstrInfo *TII = Subtarget->getInstrInfo();
3767 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
3768 CCValAssign &VA = ArgLocs[i];
3769 SDValue Arg = OutVals[i];
3770 ISD::ArgFlagsTy Flags = Outs[i].Flags;
3771 if (VA.getLocInfo() == CCValAssign::Indirect)
3773 if (!VA.isRegLoc()) {
3774 if (!MatchingStackOffset(Arg, VA.getLocMemOffset(), Flags,
3781 // If the tailcall address may be in a register, then make sure it's
3782 // possible to register allocate for it. In 32-bit, the call address can
3783 // only target EAX, EDX, or ECX since the tail call must be scheduled after
3784 // callee-saved registers are restored. These happen to be the same
3785 // registers used to pass 'inreg' arguments so watch out for those.
3786 if (!Subtarget->is64Bit() &&
3787 ((!isa<GlobalAddressSDNode>(Callee) &&
3788 !isa<ExternalSymbolSDNode>(Callee)) ||
3789 DAG.getTarget().getRelocationModel() == Reloc::PIC_)) {
3790 unsigned NumInRegs = 0;
3791 // In PIC we need an extra register to formulate the address computation
3793 unsigned MaxInRegs =
3794 (DAG.getTarget().getRelocationModel() == Reloc::PIC_) ? 2 : 3;
3796 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
3797 CCValAssign &VA = ArgLocs[i];
3800 unsigned Reg = VA.getLocReg();
3803 case X86::EAX: case X86::EDX: case X86::ECX:
3804 if (++NumInRegs == MaxInRegs)
3812 bool CalleeWillPop =
3813 X86::isCalleePop(CalleeCC, Subtarget->is64Bit(), isVarArg,
3814 MF.getTarget().Options.GuaranteedTailCallOpt);
3816 if (unsigned BytesToPop =
3817 MF.getInfo<X86MachineFunctionInfo>()->getBytesToPopOnReturn()) {
3818 // If we have bytes to pop, the callee must pop them.
3819 bool CalleePopMatches = CalleeWillPop && BytesToPop == StackArgsSize;
3820 if (!CalleePopMatches)
3822 } else if (CalleeWillPop && StackArgsSize > 0) {
3823 // If we don't have bytes to pop, make sure the callee doesn't pop any.
3831 X86TargetLowering::createFastISel(FunctionLoweringInfo &funcInfo,
3832 const TargetLibraryInfo *libInfo) const {
3833 return X86::createFastISel(funcInfo, libInfo);
3836 //===----------------------------------------------------------------------===//
3837 // Other Lowering Hooks
3838 //===----------------------------------------------------------------------===//
3840 static bool MayFoldLoad(SDValue Op) {
3841 return Op.hasOneUse() && ISD::isNormalLoad(Op.getNode());
3844 static bool MayFoldIntoStore(SDValue Op) {
3845 return Op.hasOneUse() && ISD::isNormalStore(*Op.getNode()->use_begin());
3848 static bool isTargetShuffle(unsigned Opcode) {
3850 default: return false;
3851 case X86ISD::BLENDI:
3852 case X86ISD::PSHUFB:
3853 case X86ISD::PSHUFD:
3854 case X86ISD::PSHUFHW:
3855 case X86ISD::PSHUFLW:
3857 case X86ISD::PALIGNR:
3858 case X86ISD::MOVLHPS:
3859 case X86ISD::MOVLHPD:
3860 case X86ISD::MOVHLPS:
3861 case X86ISD::MOVLPS:
3862 case X86ISD::MOVLPD:
3863 case X86ISD::MOVSHDUP:
3864 case X86ISD::MOVSLDUP:
3865 case X86ISD::MOVDDUP:
3868 case X86ISD::UNPCKL:
3869 case X86ISD::UNPCKH:
3870 case X86ISD::VPERMILPI:
3871 case X86ISD::VPERM2X128:
3872 case X86ISD::VPERMI:
3873 case X86ISD::VPERMV:
3874 case X86ISD::VPERMV3:
3879 static SDValue getTargetShuffleNode(unsigned Opc, SDLoc dl, EVT VT,
3880 SDValue V1, unsigned TargetMask,
3881 SelectionDAG &DAG) {
3883 default: llvm_unreachable("Unknown x86 shuffle node");
3884 case X86ISD::PSHUFD:
3885 case X86ISD::PSHUFHW:
3886 case X86ISD::PSHUFLW:
3887 case X86ISD::VPERMILPI:
3888 case X86ISD::VPERMI:
3889 return DAG.getNode(Opc, dl, VT, V1,
3890 DAG.getConstant(TargetMask, dl, MVT::i8));
3894 static SDValue getTargetShuffleNode(unsigned Opc, SDLoc dl, EVT VT,
3895 SDValue V1, SDValue V2, SelectionDAG &DAG) {
3897 default: llvm_unreachable("Unknown x86 shuffle node");
3898 case X86ISD::MOVLHPS:
3899 case X86ISD::MOVLHPD:
3900 case X86ISD::MOVHLPS:
3901 case X86ISD::MOVLPS:
3902 case X86ISD::MOVLPD:
3905 case X86ISD::UNPCKL:
3906 case X86ISD::UNPCKH:
3907 return DAG.getNode(Opc, dl, VT, V1, V2);
3911 SDValue X86TargetLowering::getReturnAddressFrameIndex(SelectionDAG &DAG) const {
3912 MachineFunction &MF = DAG.getMachineFunction();
3913 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
3914 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
3915 int ReturnAddrIndex = FuncInfo->getRAIndex();
3917 if (ReturnAddrIndex == 0) {
3918 // Set up a frame object for the return address.
3919 unsigned SlotSize = RegInfo->getSlotSize();
3920 ReturnAddrIndex = MF.getFrameInfo()->CreateFixedObject(SlotSize,
3923 FuncInfo->setRAIndex(ReturnAddrIndex);
3926 return DAG.getFrameIndex(ReturnAddrIndex, getPointerTy(DAG.getDataLayout()));
3929 bool X86::isOffsetSuitableForCodeModel(int64_t Offset, CodeModel::Model M,
3930 bool hasSymbolicDisplacement) {
3931 // Offset should fit into 32 bit immediate field.
3932 if (!isInt<32>(Offset))
3935 // If we don't have a symbolic displacement - we don't have any extra
3937 if (!hasSymbolicDisplacement)
3940 // FIXME: Some tweaks might be needed for medium code model.
3941 if (M != CodeModel::Small && M != CodeModel::Kernel)
3944 // For small code model we assume that latest object is 16MB before end of 31
3945 // bits boundary. We may also accept pretty large negative constants knowing
3946 // that all objects are in the positive half of address space.
3947 if (M == CodeModel::Small && Offset < 16*1024*1024)
3950 // For kernel code model we know that all object resist in the negative half
3951 // of 32bits address space. We may not accept negative offsets, since they may
3952 // be just off and we may accept pretty large positive ones.
3953 if (M == CodeModel::Kernel && Offset >= 0)
3959 /// Determines whether the callee is required to pop its own arguments.
3960 /// Callee pop is necessary to support tail calls.
3961 bool X86::isCalleePop(CallingConv::ID CallingConv,
3962 bool is64Bit, bool IsVarArg, bool GuaranteeTCO) {
3963 // If GuaranteeTCO is true, we force some calls to be callee pop so that we
3964 // can guarantee TCO.
3965 if (!IsVarArg && shouldGuaranteeTCO(CallingConv, GuaranteeTCO))
3968 switch (CallingConv) {
3971 case CallingConv::X86_StdCall:
3972 case CallingConv::X86_FastCall:
3973 case CallingConv::X86_ThisCall:
3974 case CallingConv::X86_VectorCall:
3979 /// \brief Return true if the condition is an unsigned comparison operation.
3980 static bool isX86CCUnsigned(unsigned X86CC) {
3982 default: llvm_unreachable("Invalid integer condition!");
3983 case X86::COND_E: return true;
3984 case X86::COND_G: return false;
3985 case X86::COND_GE: return false;
3986 case X86::COND_L: return false;
3987 case X86::COND_LE: return false;
3988 case X86::COND_NE: return true;
3989 case X86::COND_B: return true;
3990 case X86::COND_A: return true;
3991 case X86::COND_BE: return true;
3992 case X86::COND_AE: return true;
3994 llvm_unreachable("covered switch fell through?!");
3997 /// Do a one-to-one translation of a ISD::CondCode to the X86-specific
3998 /// condition code, returning the condition code and the LHS/RHS of the
3999 /// comparison to make.
4000 static unsigned TranslateX86CC(ISD::CondCode SetCCOpcode, SDLoc DL, bool isFP,
4001 SDValue &LHS, SDValue &RHS, SelectionDAG &DAG) {
4003 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS)) {
4004 if (SetCCOpcode == ISD::SETGT && RHSC->isAllOnesValue()) {
4005 // X > -1 -> X == 0, jump !sign.
4006 RHS = DAG.getConstant(0, DL, RHS.getValueType());
4007 return X86::COND_NS;
4009 if (SetCCOpcode == ISD::SETLT && RHSC->isNullValue()) {
4010 // X < 0 -> X == 0, jump on sign.
4013 if (SetCCOpcode == ISD::SETLT && RHSC->getZExtValue() == 1) {
4015 RHS = DAG.getConstant(0, DL, RHS.getValueType());
4016 return X86::COND_LE;
4020 switch (SetCCOpcode) {
4021 default: llvm_unreachable("Invalid integer condition!");
4022 case ISD::SETEQ: return X86::COND_E;
4023 case ISD::SETGT: return X86::COND_G;
4024 case ISD::SETGE: return X86::COND_GE;
4025 case ISD::SETLT: return X86::COND_L;
4026 case ISD::SETLE: return X86::COND_LE;
4027 case ISD::SETNE: return X86::COND_NE;
4028 case ISD::SETULT: return X86::COND_B;
4029 case ISD::SETUGT: return X86::COND_A;
4030 case ISD::SETULE: return X86::COND_BE;
4031 case ISD::SETUGE: return X86::COND_AE;
4035 // First determine if it is required or is profitable to flip the operands.
4037 // If LHS is a foldable load, but RHS is not, flip the condition.
4038 if (ISD::isNON_EXTLoad(LHS.getNode()) &&
4039 !ISD::isNON_EXTLoad(RHS.getNode())) {
4040 SetCCOpcode = getSetCCSwappedOperands(SetCCOpcode);
4041 std::swap(LHS, RHS);
4044 switch (SetCCOpcode) {
4050 std::swap(LHS, RHS);
4054 // On a floating point condition, the flags are set as follows:
4056 // 0 | 0 | 0 | X > Y
4057 // 0 | 0 | 1 | X < Y
4058 // 1 | 0 | 0 | X == Y
4059 // 1 | 1 | 1 | unordered
4060 switch (SetCCOpcode) {
4061 default: llvm_unreachable("Condcode should be pre-legalized away");
4063 case ISD::SETEQ: return X86::COND_E;
4064 case ISD::SETOLT: // flipped
4066 case ISD::SETGT: return X86::COND_A;
4067 case ISD::SETOLE: // flipped
4069 case ISD::SETGE: return X86::COND_AE;
4070 case ISD::SETUGT: // flipped
4072 case ISD::SETLT: return X86::COND_B;
4073 case ISD::SETUGE: // flipped
4075 case ISD::SETLE: return X86::COND_BE;
4077 case ISD::SETNE: return X86::COND_NE;
4078 case ISD::SETUO: return X86::COND_P;
4079 case ISD::SETO: return X86::COND_NP;
4081 case ISD::SETUNE: return X86::COND_INVALID;
4085 /// Is there a floating point cmov for the specific X86 condition code?
4086 /// Current x86 isa includes the following FP cmov instructions:
4087 /// fcmovb, fcomvbe, fcomve, fcmovu, fcmovae, fcmova, fcmovne, fcmovnu.
4088 static bool hasFPCMov(unsigned X86CC) {
4104 /// Returns true if the target can instruction select the
4105 /// specified FP immediate natively. If false, the legalizer will
4106 /// materialize the FP immediate as a load from a constant pool.
4107 bool X86TargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
4108 for (unsigned i = 0, e = LegalFPImmediates.size(); i != e; ++i) {
4109 if (Imm.bitwiseIsEqual(LegalFPImmediates[i]))
4115 bool X86TargetLowering::shouldReduceLoadWidth(SDNode *Load,
4116 ISD::LoadExtType ExtTy,
4118 // "ELF Handling for Thread-Local Storage" specifies that R_X86_64_GOTTPOFF
4119 // relocation target a movq or addq instruction: don't let the load shrink.
4120 SDValue BasePtr = cast<LoadSDNode>(Load)->getBasePtr();
4121 if (BasePtr.getOpcode() == X86ISD::WrapperRIP)
4122 if (const auto *GA = dyn_cast<GlobalAddressSDNode>(BasePtr.getOperand(0)))
4123 return GA->getTargetFlags() != X86II::MO_GOTTPOFF;
4127 /// \brief Returns true if it is beneficial to convert a load of a constant
4128 /// to just the constant itself.
4129 bool X86TargetLowering::shouldConvertConstantLoadToIntImm(const APInt &Imm,
4131 assert(Ty->isIntegerTy());
4133 unsigned BitSize = Ty->getPrimitiveSizeInBits();
4134 if (BitSize == 0 || BitSize > 64)
4139 bool X86TargetLowering::isExtractSubvectorCheap(EVT ResVT,
4140 unsigned Index) const {
4141 if (!isOperationLegalOrCustom(ISD::EXTRACT_SUBVECTOR, ResVT))
4144 return (Index == 0 || Index == ResVT.getVectorNumElements());
4147 bool X86TargetLowering::isCheapToSpeculateCttz() const {
4148 // Speculate cttz only if we can directly use TZCNT.
4149 return Subtarget->hasBMI();
4152 bool X86TargetLowering::isCheapToSpeculateCtlz() const {
4153 // Speculate ctlz only if we can directly use LZCNT.
4154 return Subtarget->hasLZCNT();
4157 /// Return true if every element in Mask, beginning
4158 /// from position Pos and ending in Pos+Size is undef.
4159 static bool isUndefInRange(ArrayRef<int> Mask, unsigned Pos, unsigned Size) {
4160 for (unsigned i = Pos, e = Pos + Size; i != e; ++i)
4166 /// Return true if Val is undef or if its value falls within the
4167 /// specified range (L, H].
4168 static bool isUndefOrInRange(int Val, int Low, int Hi) {
4169 return (Val < 0) || (Val >= Low && Val < Hi);
4172 /// Val is either less than zero (undef) or equal to the specified value.
4173 static bool isUndefOrEqual(int Val, int CmpVal) {
4174 return (Val < 0 || Val == CmpVal);
4177 /// Return true if every element in Mask, beginning
4178 /// from position Pos and ending in Pos+Size, falls within the specified
4179 /// sequential range (Low, Low+Size]. or is undef.
4180 static bool isSequentialOrUndefInRange(ArrayRef<int> Mask,
4181 unsigned Pos, unsigned Size, int Low) {
4182 for (unsigned i = Pos, e = Pos+Size; i != e; ++i, ++Low)
4183 if (!isUndefOrEqual(Mask[i], Low))
4188 /// Return true if the specified EXTRACT_SUBVECTOR operand specifies a vector
4189 /// extract that is suitable for instruction that extract 128 or 256 bit vectors
4190 static bool isVEXTRACTIndex(SDNode *N, unsigned vecWidth) {
4191 assert((vecWidth == 128 || vecWidth == 256) && "Unexpected vector width");
4192 if (!isa<ConstantSDNode>(N->getOperand(1).getNode()))
4195 // The index should be aligned on a vecWidth-bit boundary.
4197 cast<ConstantSDNode>(N->getOperand(1).getNode())->getZExtValue();
4199 MVT VT = N->getSimpleValueType(0);
4200 unsigned ElSize = VT.getVectorElementType().getSizeInBits();
4201 bool Result = (Index * ElSize) % vecWidth == 0;
4206 /// Return true if the specified INSERT_SUBVECTOR
4207 /// operand specifies a subvector insert that is suitable for input to
4208 /// insertion of 128 or 256-bit subvectors
4209 static bool isVINSERTIndex(SDNode *N, unsigned vecWidth) {
4210 assert((vecWidth == 128 || vecWidth == 256) && "Unexpected vector width");
4211 if (!isa<ConstantSDNode>(N->getOperand(2).getNode()))
4213 // The index should be aligned on a vecWidth-bit boundary.
4215 cast<ConstantSDNode>(N->getOperand(2).getNode())->getZExtValue();
4217 MVT VT = N->getSimpleValueType(0);
4218 unsigned ElSize = VT.getVectorElementType().getSizeInBits();
4219 bool Result = (Index * ElSize) % vecWidth == 0;
4224 bool X86::isVINSERT128Index(SDNode *N) {
4225 return isVINSERTIndex(N, 128);
4228 bool X86::isVINSERT256Index(SDNode *N) {
4229 return isVINSERTIndex(N, 256);
4232 bool X86::isVEXTRACT128Index(SDNode *N) {
4233 return isVEXTRACTIndex(N, 128);
4236 bool X86::isVEXTRACT256Index(SDNode *N) {
4237 return isVEXTRACTIndex(N, 256);
4240 static unsigned getExtractVEXTRACTImmediate(SDNode *N, unsigned vecWidth) {
4241 assert((vecWidth == 128 || vecWidth == 256) && "Unsupported vector width");
4242 if (!isa<ConstantSDNode>(N->getOperand(1).getNode()))
4243 llvm_unreachable("Illegal extract subvector for VEXTRACT");
4246 cast<ConstantSDNode>(N->getOperand(1).getNode())->getZExtValue();
4248 MVT VecVT = N->getOperand(0).getSimpleValueType();
4249 MVT ElVT = VecVT.getVectorElementType();
4251 unsigned NumElemsPerChunk = vecWidth / ElVT.getSizeInBits();
4252 return Index / NumElemsPerChunk;
4255 static unsigned getInsertVINSERTImmediate(SDNode *N, unsigned vecWidth) {
4256 assert((vecWidth == 128 || vecWidth == 256) && "Unsupported vector width");
4257 if (!isa<ConstantSDNode>(N->getOperand(2).getNode()))
4258 llvm_unreachable("Illegal insert subvector for VINSERT");
4261 cast<ConstantSDNode>(N->getOperand(2).getNode())->getZExtValue();
4263 MVT VecVT = N->getSimpleValueType(0);
4264 MVT ElVT = VecVT.getVectorElementType();
4266 unsigned NumElemsPerChunk = vecWidth / ElVT.getSizeInBits();
4267 return Index / NumElemsPerChunk;
4270 /// Return the appropriate immediate to extract the specified
4271 /// EXTRACT_SUBVECTOR index with VEXTRACTF128 and VINSERTI128 instructions.
4272 unsigned X86::getExtractVEXTRACT128Immediate(SDNode *N) {
4273 return getExtractVEXTRACTImmediate(N, 128);
4276 /// Return the appropriate immediate to extract the specified
4277 /// EXTRACT_SUBVECTOR index with VEXTRACTF64x4 and VINSERTI64x4 instructions.
4278 unsigned X86::getExtractVEXTRACT256Immediate(SDNode *N) {
4279 return getExtractVEXTRACTImmediate(N, 256);
4282 /// Return the appropriate immediate to insert at the specified
4283 /// INSERT_SUBVECTOR index with VINSERTF128 and VINSERTI128 instructions.
4284 unsigned X86::getInsertVINSERT128Immediate(SDNode *N) {
4285 return getInsertVINSERTImmediate(N, 128);
4288 /// Return the appropriate immediate to insert at the specified
4289 /// INSERT_SUBVECTOR index with VINSERTF46x4 and VINSERTI64x4 instructions.
4290 unsigned X86::getInsertVINSERT256Immediate(SDNode *N) {
4291 return getInsertVINSERTImmediate(N, 256);
4294 /// Returns true if V is a constant integer zero.
4295 static bool isZero(SDValue V) {
4296 ConstantSDNode *C = dyn_cast<ConstantSDNode>(V);
4297 return C && C->isNullValue();
4300 /// Returns true if Elt is a constant zero or a floating point constant +0.0.
4301 bool X86::isZeroNode(SDValue Elt) {
4304 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(Elt))
4305 return CFP->getValueAPF().isPosZero();
4309 // Build a vector of constants
4310 // Use an UNDEF node if MaskElt == -1.
4311 // Spilt 64-bit constants in the 32-bit mode.
4312 static SDValue getConstVector(ArrayRef<int> Values, EVT VT,
4314 SDLoc dl, bool IsMask = false) {
4316 SmallVector<SDValue, 32> Ops;
4319 EVT ConstVecVT = VT;
4320 unsigned NumElts = VT.getVectorNumElements();
4321 bool In64BitMode = DAG.getTargetLoweringInfo().isTypeLegal(MVT::i64);
4322 if (!In64BitMode && VT.getScalarType() == MVT::i64) {
4323 ConstVecVT = MVT::getVectorVT(MVT::i32, NumElts * 2);
4327 EVT EltVT = ConstVecVT.getScalarType();
4328 for (unsigned i = 0; i < NumElts; ++i) {
4329 bool IsUndef = Values[i] < 0 && IsMask;
4330 SDValue OpNode = IsUndef ? DAG.getUNDEF(EltVT) :
4331 DAG.getConstant(Values[i], dl, EltVT);
4332 Ops.push_back(OpNode);
4334 Ops.push_back(IsUndef ? DAG.getUNDEF(EltVT) :
4335 DAG.getConstant(0, dl, EltVT));
4337 SDValue ConstsNode = DAG.getNode(ISD::BUILD_VECTOR, dl, ConstVecVT, Ops);
4339 ConstsNode = DAG.getBitcast(VT, ConstsNode);
4343 /// Returns a vector of specified type with all zero elements.
4344 static SDValue getZeroVector(EVT VT, const X86Subtarget *Subtarget,
4345 SelectionDAG &DAG, SDLoc dl) {
4346 assert(VT.isVector() && "Expected a vector type");
4348 // Always build SSE zero vectors as <4 x i32> bitcasted
4349 // to their dest type. This ensures they get CSE'd.
4351 if (VT.is128BitVector()) { // SSE
4352 if (Subtarget->hasSSE2()) { // SSE2
4353 SDValue Cst = DAG.getConstant(0, dl, MVT::i32);
4354 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
4356 SDValue Cst = DAG.getConstantFP(+0.0, dl, MVT::f32);
4357 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4f32, Cst, Cst, Cst, Cst);
4359 } else if (VT.is256BitVector()) { // AVX
4360 if (Subtarget->hasInt256()) { // AVX2
4361 SDValue Cst = DAG.getConstant(0, dl, MVT::i32);
4362 SDValue Ops[] = { Cst, Cst, Cst, Cst, Cst, Cst, Cst, Cst };
4363 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v8i32, Ops);
4365 // 256-bit logic and arithmetic instructions in AVX are all
4366 // floating-point, no support for integer ops. Emit fp zeroed vectors.
4367 SDValue Cst = DAG.getConstantFP(+0.0, dl, MVT::f32);
4368 SDValue Ops[] = { Cst, Cst, Cst, Cst, Cst, Cst, Cst, Cst };
4369 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v8f32, Ops);
4371 } else if (VT.is512BitVector()) { // AVX-512
4372 SDValue Cst = DAG.getConstant(0, dl, MVT::i32);
4373 SDValue Ops[] = { Cst, Cst, Cst, Cst, Cst, Cst, Cst, Cst,
4374 Cst, Cst, Cst, Cst, Cst, Cst, Cst, Cst };
4375 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v16i32, Ops);
4376 } else if (VT.getScalarType() == MVT::i1) {
4378 assert((Subtarget->hasBWI() || VT.getVectorNumElements() <= 16)
4379 && "Unexpected vector type");
4380 assert((Subtarget->hasVLX() || VT.getVectorNumElements() >= 8)
4381 && "Unexpected vector type");
4382 SDValue Cst = DAG.getConstant(0, dl, MVT::i1);
4383 SmallVector<SDValue, 64> Ops(VT.getVectorNumElements(), Cst);
4384 return DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Ops);
4386 llvm_unreachable("Unexpected vector type");
4388 return DAG.getBitcast(VT, Vec);
4391 static SDValue ExtractSubVector(SDValue Vec, unsigned IdxVal,
4392 SelectionDAG &DAG, SDLoc dl,
4393 unsigned vectorWidth) {
4394 assert((vectorWidth == 128 || vectorWidth == 256) &&
4395 "Unsupported vector width");
4396 EVT VT = Vec.getValueType();
4397 EVT ElVT = VT.getVectorElementType();
4398 unsigned Factor = VT.getSizeInBits()/vectorWidth;
4399 EVT ResultVT = EVT::getVectorVT(*DAG.getContext(), ElVT,
4400 VT.getVectorNumElements()/Factor);
4402 // Extract from UNDEF is UNDEF.
4403 if (Vec.getOpcode() == ISD::UNDEF)
4404 return DAG.getUNDEF(ResultVT);
4406 // Extract the relevant vectorWidth bits. Generate an EXTRACT_SUBVECTOR
4407 unsigned ElemsPerChunk = vectorWidth / ElVT.getSizeInBits();
4409 // This is the index of the first element of the vectorWidth-bit chunk
4411 unsigned NormalizedIdxVal = (((IdxVal * ElVT.getSizeInBits()) / vectorWidth)
4414 // If the input is a buildvector just emit a smaller one.
4415 if (Vec.getOpcode() == ISD::BUILD_VECTOR)
4416 return DAG.getNode(ISD::BUILD_VECTOR, dl, ResultVT,
4417 makeArrayRef(Vec->op_begin() + NormalizedIdxVal,
4420 SDValue VecIdx = DAG.getIntPtrConstant(NormalizedIdxVal, dl);
4421 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, ResultVT, Vec, VecIdx);
4424 /// Generate a DAG to grab 128-bits from a vector > 128 bits. This
4425 /// sets things up to match to an AVX VEXTRACTF128 / VEXTRACTI128
4426 /// or AVX-512 VEXTRACTF32x4 / VEXTRACTI32x4
4427 /// instructions or a simple subregister reference. Idx is an index in the
4428 /// 128 bits we want. It need not be aligned to a 128-bit boundary. That makes
4429 /// lowering EXTRACT_VECTOR_ELT operations easier.
4430 static SDValue Extract128BitVector(SDValue Vec, unsigned IdxVal,
4431 SelectionDAG &DAG, SDLoc dl) {
4432 assert((Vec.getValueType().is256BitVector() ||
4433 Vec.getValueType().is512BitVector()) && "Unexpected vector size!");
4434 return ExtractSubVector(Vec, IdxVal, DAG, dl, 128);
4437 /// Generate a DAG to grab 256-bits from a 512-bit vector.
4438 static SDValue Extract256BitVector(SDValue Vec, unsigned IdxVal,
4439 SelectionDAG &DAG, SDLoc dl) {
4440 assert(Vec.getValueType().is512BitVector() && "Unexpected vector size!");
4441 return ExtractSubVector(Vec, IdxVal, DAG, dl, 256);
4444 static SDValue InsertSubVector(SDValue Result, SDValue Vec,
4445 unsigned IdxVal, SelectionDAG &DAG,
4446 SDLoc dl, unsigned vectorWidth) {
4447 assert((vectorWidth == 128 || vectorWidth == 256) &&
4448 "Unsupported vector width");
4449 // Inserting UNDEF is Result
4450 if (Vec.getOpcode() == ISD::UNDEF)
4452 EVT VT = Vec.getValueType();
4453 EVT ElVT = VT.getVectorElementType();
4454 EVT ResultVT = Result.getValueType();
4456 // Insert the relevant vectorWidth bits.
4457 unsigned ElemsPerChunk = vectorWidth/ElVT.getSizeInBits();
4459 // This is the index of the first element of the vectorWidth-bit chunk
4461 unsigned NormalizedIdxVal = (((IdxVal * ElVT.getSizeInBits())/vectorWidth)
4464 SDValue VecIdx = DAG.getIntPtrConstant(NormalizedIdxVal, dl);
4465 return DAG.getNode(ISD::INSERT_SUBVECTOR, dl, ResultVT, Result, Vec, VecIdx);
4468 /// Generate a DAG to put 128-bits into a vector > 128 bits. This
4469 /// sets things up to match to an AVX VINSERTF128/VINSERTI128 or
4470 /// AVX-512 VINSERTF32x4/VINSERTI32x4 instructions or a
4471 /// simple superregister reference. Idx is an index in the 128 bits
4472 /// we want. It need not be aligned to a 128-bit boundary. That makes
4473 /// lowering INSERT_VECTOR_ELT operations easier.
4474 static SDValue Insert128BitVector(SDValue Result, SDValue Vec, unsigned IdxVal,
4475 SelectionDAG &DAG, SDLoc dl) {
4476 assert(Vec.getValueType().is128BitVector() && "Unexpected vector size!");
4478 // For insertion into the zero index (low half) of a 256-bit vector, it is
4479 // more efficient to generate a blend with immediate instead of an insert*128.
4480 // We are still creating an INSERT_SUBVECTOR below with an undef node to
4481 // extend the subvector to the size of the result vector. Make sure that
4482 // we are not recursing on that node by checking for undef here.
4483 if (IdxVal == 0 && Result.getValueType().is256BitVector() &&
4484 Result.getOpcode() != ISD::UNDEF) {
4485 EVT ResultVT = Result.getValueType();
4486 SDValue ZeroIndex = DAG.getIntPtrConstant(0, dl);
4487 SDValue Undef = DAG.getUNDEF(ResultVT);
4488 SDValue Vec256 = DAG.getNode(ISD::INSERT_SUBVECTOR, dl, ResultVT, Undef,
4491 // The blend instruction, and therefore its mask, depend on the data type.
4492 MVT ScalarType = ResultVT.getScalarType().getSimpleVT();
4493 if (ScalarType.isFloatingPoint()) {
4494 // Choose either vblendps (float) or vblendpd (double).
4495 unsigned ScalarSize = ScalarType.getSizeInBits();
4496 assert((ScalarSize == 64 || ScalarSize == 32) && "Unknown float type");
4497 unsigned MaskVal = (ScalarSize == 64) ? 0x03 : 0x0f;
4498 SDValue Mask = DAG.getConstant(MaskVal, dl, MVT::i8);
4499 return DAG.getNode(X86ISD::BLENDI, dl, ResultVT, Result, Vec256, Mask);
4502 const X86Subtarget &Subtarget =
4503 static_cast<const X86Subtarget &>(DAG.getSubtarget());
4505 // AVX2 is needed for 256-bit integer blend support.
4506 // Integers must be cast to 32-bit because there is only vpblendd;
4507 // vpblendw can't be used for this because it has a handicapped mask.
4509 // If we don't have AVX2, then cast to float. Using a wrong domain blend
4510 // is still more efficient than using the wrong domain vinsertf128 that
4511 // will be created by InsertSubVector().
4512 MVT CastVT = Subtarget.hasAVX2() ? MVT::v8i32 : MVT::v8f32;
4514 SDValue Mask = DAG.getConstant(0x0f, dl, MVT::i8);
4515 Vec256 = DAG.getBitcast(CastVT, Vec256);
4516 Vec256 = DAG.getNode(X86ISD::BLENDI, dl, CastVT, Result, Vec256, Mask);
4517 return DAG.getBitcast(ResultVT, Vec256);
4520 return InsertSubVector(Result, Vec, IdxVal, DAG, dl, 128);
4523 static SDValue Insert256BitVector(SDValue Result, SDValue Vec, unsigned IdxVal,
4524 SelectionDAG &DAG, SDLoc dl) {
4525 assert(Vec.getValueType().is256BitVector() && "Unexpected vector size!");
4526 return InsertSubVector(Result, Vec, IdxVal, DAG, dl, 256);
4529 /// Concat two 128-bit vectors into a 256 bit vector using VINSERTF128
4530 /// instructions. This is used because creating CONCAT_VECTOR nodes of
4531 /// BUILD_VECTORS returns a larger BUILD_VECTOR while we're trying to lower
4532 /// large BUILD_VECTORS.
4533 static SDValue Concat128BitVectors(SDValue V1, SDValue V2, EVT VT,
4534 unsigned NumElems, SelectionDAG &DAG,
4536 SDValue V = Insert128BitVector(DAG.getUNDEF(VT), V1, 0, DAG, dl);
4537 return Insert128BitVector(V, V2, NumElems/2, DAG, dl);
4540 static SDValue Concat256BitVectors(SDValue V1, SDValue V2, EVT VT,
4541 unsigned NumElems, SelectionDAG &DAG,
4543 SDValue V = Insert256BitVector(DAG.getUNDEF(VT), V1, 0, DAG, dl);
4544 return Insert256BitVector(V, V2, NumElems/2, DAG, dl);
4547 /// Returns a vector of specified type with all bits set.
4548 /// Always build ones vectors as <4 x i32> or <8 x i32>. For 256-bit types with
4549 /// no AVX2 supprt, use two <4 x i32> inserted in a <8 x i32> appropriately.
4550 /// Then bitcast to their original type, ensuring they get CSE'd.
4551 static SDValue getOnesVector(EVT VT, const X86Subtarget *Subtarget,
4552 SelectionDAG &DAG, SDLoc dl) {
4553 assert(VT.isVector() && "Expected a vector type");
4555 SDValue Cst = DAG.getConstant(~0U, dl, MVT::i32);
4557 if (VT.is512BitVector()) {
4558 SDValue Ops[] = { Cst, Cst, Cst, Cst, Cst, Cst, Cst, Cst,
4559 Cst, Cst, Cst, Cst, Cst, Cst, Cst, Cst };
4560 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v16i32, Ops);
4561 } else if (VT.is256BitVector()) {
4562 if (Subtarget->hasInt256()) { // AVX2
4563 SDValue Ops[] = { Cst, Cst, Cst, Cst, Cst, Cst, Cst, Cst };
4564 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v8i32, Ops);
4566 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
4567 Vec = Concat128BitVectors(Vec, Vec, MVT::v8i32, 8, DAG, dl);
4569 } else if (VT.is128BitVector()) {
4570 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
4572 llvm_unreachable("Unexpected vector type");
4574 return DAG.getBitcast(VT, Vec);
4577 /// Returns a vector_shuffle node for an unpackl operation.
4578 static SDValue getUnpackl(SelectionDAG &DAG, SDLoc dl, MVT VT, SDValue V1,
4580 unsigned NumElems = VT.getVectorNumElements();
4581 SmallVector<int, 8> Mask;
4582 for (unsigned i = 0, e = NumElems/2; i != e; ++i) {
4584 Mask.push_back(i + NumElems);
4586 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
4589 /// Returns a vector_shuffle node for an unpackh operation.
4590 static SDValue getUnpackh(SelectionDAG &DAG, SDLoc dl, MVT VT, SDValue V1,
4592 unsigned NumElems = VT.getVectorNumElements();
4593 SmallVector<int, 8> Mask;
4594 for (unsigned i = 0, Half = NumElems/2; i != Half; ++i) {
4595 Mask.push_back(i + Half);
4596 Mask.push_back(i + NumElems + Half);
4598 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
4601 /// Return a vector_shuffle of the specified vector of zero or undef vector.
4602 /// This produces a shuffle where the low element of V2 is swizzled into the
4603 /// zero/undef vector, landing at element Idx.
4604 /// This produces a shuffle mask like 4,1,2,3 (idx=0) or 0,1,2,4 (idx=3).
4605 static SDValue getShuffleVectorZeroOrUndef(SDValue V2, unsigned Idx,
4607 const X86Subtarget *Subtarget,
4608 SelectionDAG &DAG) {
4609 MVT VT = V2.getSimpleValueType();
4611 ? getZeroVector(VT, Subtarget, DAG, SDLoc(V2)) : DAG.getUNDEF(VT);
4612 unsigned NumElems = VT.getVectorNumElements();
4613 SmallVector<int, 16> MaskVec;
4614 for (unsigned i = 0; i != NumElems; ++i)
4615 // If this is the insertion idx, put the low elt of V2 here.
4616 MaskVec.push_back(i == Idx ? NumElems : i);
4617 return DAG.getVectorShuffle(VT, SDLoc(V2), V1, V2, &MaskVec[0]);
4620 /// Calculates the shuffle mask corresponding to the target-specific opcode.
4621 /// Returns true if the Mask could be calculated. Sets IsUnary to true if only
4622 /// uses one source. Note that this will set IsUnary for shuffles which use a
4623 /// single input multiple times, and in those cases it will
4624 /// adjust the mask to only have indices within that single input.
4625 /// FIXME: Add support for Decode*Mask functions that return SM_SentinelZero.
4626 static bool getTargetShuffleMask(SDNode *N, MVT VT,
4627 SmallVectorImpl<int> &Mask, bool &IsUnary) {
4628 unsigned NumElems = VT.getVectorNumElements();
4632 bool IsFakeUnary = false;
4633 switch(N->getOpcode()) {
4634 case X86ISD::BLENDI:
4635 ImmN = N->getOperand(N->getNumOperands()-1);
4636 DecodeBLENDMask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
4639 ImmN = N->getOperand(N->getNumOperands()-1);
4640 DecodeSHUFPMask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
4641 IsUnary = IsFakeUnary = N->getOperand(0) == N->getOperand(1);
4643 case X86ISD::UNPCKH:
4644 DecodeUNPCKHMask(VT, Mask);
4645 IsUnary = IsFakeUnary = N->getOperand(0) == N->getOperand(1);
4647 case X86ISD::UNPCKL:
4648 DecodeUNPCKLMask(VT, Mask);
4649 IsUnary = IsFakeUnary = N->getOperand(0) == N->getOperand(1);
4651 case X86ISD::MOVHLPS:
4652 DecodeMOVHLPSMask(NumElems, Mask);
4653 IsUnary = IsFakeUnary = N->getOperand(0) == N->getOperand(1);
4655 case X86ISD::MOVLHPS:
4656 DecodeMOVLHPSMask(NumElems, Mask);
4657 IsUnary = IsFakeUnary = N->getOperand(0) == N->getOperand(1);
4659 case X86ISD::PALIGNR:
4660 ImmN = N->getOperand(N->getNumOperands()-1);
4661 DecodePALIGNRMask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
4663 case X86ISD::PSHUFD:
4664 case X86ISD::VPERMILPI:
4665 ImmN = N->getOperand(N->getNumOperands()-1);
4666 DecodePSHUFMask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
4669 case X86ISD::PSHUFHW:
4670 ImmN = N->getOperand(N->getNumOperands()-1);
4671 DecodePSHUFHWMask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
4674 case X86ISD::PSHUFLW:
4675 ImmN = N->getOperand(N->getNumOperands()-1);
4676 DecodePSHUFLWMask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
4679 case X86ISD::PSHUFB: {
4681 SDValue MaskNode = N->getOperand(1);
4682 while (MaskNode->getOpcode() == ISD::BITCAST)
4683 MaskNode = MaskNode->getOperand(0);
4685 if (MaskNode->getOpcode() == ISD::BUILD_VECTOR) {
4686 // If we have a build-vector, then things are easy.
4687 EVT VT = MaskNode.getValueType();
4688 assert(VT.isVector() &&
4689 "Can't produce a non-vector with a build_vector!");
4690 if (!VT.isInteger())
4693 int NumBytesPerElement = VT.getVectorElementType().getSizeInBits() / 8;
4695 SmallVector<uint64_t, 32> RawMask;
4696 for (int i = 0, e = MaskNode->getNumOperands(); i < e; ++i) {
4697 SDValue Op = MaskNode->getOperand(i);
4698 if (Op->getOpcode() == ISD::UNDEF) {
4699 RawMask.push_back((uint64_t)SM_SentinelUndef);
4702 auto *CN = dyn_cast<ConstantSDNode>(Op.getNode());
4705 APInt MaskElement = CN->getAPIntValue();
4707 // We now have to decode the element which could be any integer size and
4708 // extract each byte of it.
4709 for (int j = 0; j < NumBytesPerElement; ++j) {
4710 // Note that this is x86 and so always little endian: the low byte is
4711 // the first byte of the mask.
4712 RawMask.push_back(MaskElement.getLoBits(8).getZExtValue());
4713 MaskElement = MaskElement.lshr(8);
4716 DecodePSHUFBMask(RawMask, Mask);
4720 auto *MaskLoad = dyn_cast<LoadSDNode>(MaskNode);
4724 SDValue Ptr = MaskLoad->getBasePtr();
4725 if (Ptr->getOpcode() == X86ISD::Wrapper ||
4726 Ptr->getOpcode() == X86ISD::WrapperRIP)
4727 Ptr = Ptr->getOperand(0);
4729 auto *MaskCP = dyn_cast<ConstantPoolSDNode>(Ptr);
4730 if (!MaskCP || MaskCP->isMachineConstantPoolEntry())
4733 if (auto *C = dyn_cast<Constant>(MaskCP->getConstVal())) {
4734 DecodePSHUFBMask(C, Mask);
4742 case X86ISD::VPERMI:
4743 ImmN = N->getOperand(N->getNumOperands()-1);
4744 DecodeVPERMMask(cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
4749 DecodeScalarMoveMask(VT, /* IsLoad */ false, Mask);
4751 case X86ISD::VPERM2X128:
4752 ImmN = N->getOperand(N->getNumOperands()-1);
4753 DecodeVPERM2X128Mask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
4754 if (Mask.empty()) return false;
4755 // Mask only contains negative index if an element is zero.
4756 if (std::any_of(Mask.begin(), Mask.end(),
4757 [](int M){ return M == SM_SentinelZero; }))
4760 case X86ISD::MOVSLDUP:
4761 DecodeMOVSLDUPMask(VT, Mask);
4764 case X86ISD::MOVSHDUP:
4765 DecodeMOVSHDUPMask(VT, Mask);
4768 case X86ISD::MOVDDUP:
4769 DecodeMOVDDUPMask(VT, Mask);
4772 case X86ISD::MOVLHPD:
4773 case X86ISD::MOVLPD:
4774 case X86ISD::MOVLPS:
4775 // Not yet implemented
4777 case X86ISD::VPERMV: {
4779 SDValue MaskNode = N->getOperand(0);
4780 while (MaskNode->getOpcode() == ISD::BITCAST)
4781 MaskNode = MaskNode->getOperand(0);
4783 unsigned MaskLoBits = Log2_64(VT.getVectorNumElements());
4784 SmallVector<uint64_t, 32> RawMask;
4785 if (MaskNode->getOpcode() == ISD::BUILD_VECTOR) {
4786 // If we have a build-vector, then things are easy.
4787 assert(MaskNode.getValueType().isInteger() &&
4788 MaskNode.getValueType().getVectorNumElements() ==
4789 VT.getVectorNumElements());
4791 for (unsigned i = 0; i < MaskNode->getNumOperands(); ++i) {
4792 SDValue Op = MaskNode->getOperand(i);
4793 if (Op->getOpcode() == ISD::UNDEF)
4794 RawMask.push_back((uint64_t)SM_SentinelUndef);
4795 else if (isa<ConstantSDNode>(Op)) {
4796 APInt MaskElement = cast<ConstantSDNode>(Op)->getAPIntValue();
4797 RawMask.push_back(MaskElement.getLoBits(MaskLoBits).getZExtValue());
4801 DecodeVPERMVMask(RawMask, Mask);
4804 if (MaskNode->getOpcode() == X86ISD::VBROADCAST) {
4805 unsigned NumEltsInMask = MaskNode->getNumOperands();
4806 MaskNode = MaskNode->getOperand(0);
4807 auto *CN = dyn_cast<ConstantSDNode>(MaskNode);
4809 APInt MaskEltValue = CN->getAPIntValue();
4810 for (unsigned i = 0; i < NumEltsInMask; ++i)
4811 RawMask.push_back(MaskEltValue.getLoBits(MaskLoBits).getZExtValue());
4812 DecodeVPERMVMask(RawMask, Mask);
4815 // It may be a scalar load
4818 auto *MaskLoad = dyn_cast<LoadSDNode>(MaskNode);
4822 SDValue Ptr = MaskLoad->getBasePtr();
4823 if (Ptr->getOpcode() == X86ISD::Wrapper ||
4824 Ptr->getOpcode() == X86ISD::WrapperRIP)
4825 Ptr = Ptr->getOperand(0);
4827 auto *MaskCP = dyn_cast<ConstantPoolSDNode>(Ptr);
4828 if (!MaskCP || MaskCP->isMachineConstantPoolEntry())
4831 auto *C = dyn_cast<Constant>(MaskCP->getConstVal());
4833 DecodeVPERMVMask(C, VT, Mask);
4840 case X86ISD::VPERMV3: {
4842 SDValue MaskNode = N->getOperand(1);
4843 while (MaskNode->getOpcode() == ISD::BITCAST)
4844 MaskNode = MaskNode->getOperand(1);
4846 if (MaskNode->getOpcode() == ISD::BUILD_VECTOR) {
4847 // If we have a build-vector, then things are easy.
4848 assert(MaskNode.getValueType().isInteger() &&
4849 MaskNode.getValueType().getVectorNumElements() ==
4850 VT.getVectorNumElements());
4852 SmallVector<uint64_t, 32> RawMask;
4853 unsigned MaskLoBits = Log2_64(VT.getVectorNumElements()*2);
4855 for (unsigned i = 0; i < MaskNode->getNumOperands(); ++i) {
4856 SDValue Op = MaskNode->getOperand(i);
4857 if (Op->getOpcode() == ISD::UNDEF)
4858 RawMask.push_back((uint64_t)SM_SentinelUndef);
4860 auto *CN = dyn_cast<ConstantSDNode>(Op.getNode());
4863 APInt MaskElement = CN->getAPIntValue();
4864 RawMask.push_back(MaskElement.getLoBits(MaskLoBits).getZExtValue());
4867 DecodeVPERMV3Mask(RawMask, Mask);
4871 auto *MaskLoad = dyn_cast<LoadSDNode>(MaskNode);
4875 SDValue Ptr = MaskLoad->getBasePtr();
4876 if (Ptr->getOpcode() == X86ISD::Wrapper ||
4877 Ptr->getOpcode() == X86ISD::WrapperRIP)
4878 Ptr = Ptr->getOperand(0);
4880 auto *MaskCP = dyn_cast<ConstantPoolSDNode>(Ptr);
4881 if (!MaskCP || MaskCP->isMachineConstantPoolEntry())
4884 auto *C = dyn_cast<Constant>(MaskCP->getConstVal());
4886 DecodeVPERMV3Mask(C, VT, Mask);
4893 default: llvm_unreachable("unknown target shuffle node");
4896 // If we have a fake unary shuffle, the shuffle mask is spread across two
4897 // inputs that are actually the same node. Re-map the mask to always point
4898 // into the first input.
4901 if (M >= (int)Mask.size())
4907 /// Returns the scalar element that will make up the ith
4908 /// element of the result of the vector shuffle.
4909 static SDValue getShuffleScalarElt(SDNode *N, unsigned Index, SelectionDAG &DAG,
4912 return SDValue(); // Limit search depth.
4914 SDValue V = SDValue(N, 0);
4915 EVT VT = V.getValueType();
4916 unsigned Opcode = V.getOpcode();
4918 // Recurse into ISD::VECTOR_SHUFFLE node to find scalars.
4919 if (const ShuffleVectorSDNode *SV = dyn_cast<ShuffleVectorSDNode>(N)) {
4920 int Elt = SV->getMaskElt(Index);
4923 return DAG.getUNDEF(VT.getVectorElementType());
4925 unsigned NumElems = VT.getVectorNumElements();
4926 SDValue NewV = (Elt < (int)NumElems) ? SV->getOperand(0)
4927 : SV->getOperand(1);
4928 return getShuffleScalarElt(NewV.getNode(), Elt % NumElems, DAG, Depth+1);
4931 // Recurse into target specific vector shuffles to find scalars.
4932 if (isTargetShuffle(Opcode)) {
4933 MVT ShufVT = V.getSimpleValueType();
4934 unsigned NumElems = ShufVT.getVectorNumElements();
4935 SmallVector<int, 16> ShuffleMask;
4938 if (!getTargetShuffleMask(N, ShufVT, ShuffleMask, IsUnary))
4941 int Elt = ShuffleMask[Index];
4943 return DAG.getUNDEF(ShufVT.getVectorElementType());
4945 SDValue NewV = (Elt < (int)NumElems) ? N->getOperand(0)
4947 return getShuffleScalarElt(NewV.getNode(), Elt % NumElems, DAG,
4951 // Actual nodes that may contain scalar elements
4952 if (Opcode == ISD::BITCAST) {
4953 V = V.getOperand(0);
4954 EVT SrcVT = V.getValueType();
4955 unsigned NumElems = VT.getVectorNumElements();
4957 if (!SrcVT.isVector() || SrcVT.getVectorNumElements() != NumElems)
4961 if (V.getOpcode() == ISD::SCALAR_TO_VECTOR)
4962 return (Index == 0) ? V.getOperand(0)
4963 : DAG.getUNDEF(VT.getVectorElementType());
4965 if (V.getOpcode() == ISD::BUILD_VECTOR)
4966 return V.getOperand(Index);
4971 /// Custom lower build_vector of v16i8.
4972 static SDValue LowerBuildVectorv16i8(SDValue Op, unsigned NonZeros,
4973 unsigned NumNonZero, unsigned NumZero,
4975 const X86Subtarget* Subtarget,
4976 const TargetLowering &TLI) {
4984 // SSE4.1 - use PINSRB to insert each byte directly.
4985 if (Subtarget->hasSSE41()) {
4986 for (unsigned i = 0; i < 16; ++i) {
4987 bool isNonZero = (NonZeros & (1 << i)) != 0;
4991 V = getZeroVector(MVT::v16i8, Subtarget, DAG, dl);
4993 V = DAG.getUNDEF(MVT::v16i8);
4996 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl,
4997 MVT::v16i8, V, Op.getOperand(i),
4998 DAG.getIntPtrConstant(i, dl));
5005 // Pre-SSE4.1 - merge byte pairs and insert with PINSRW.
5006 for (unsigned i = 0; i < 16; ++i) {
5007 bool ThisIsNonZero = (NonZeros & (1 << i)) != 0;
5008 if (ThisIsNonZero && First) {
5010 V = getZeroVector(MVT::v8i16, Subtarget, DAG, dl);
5012 V = DAG.getUNDEF(MVT::v8i16);
5017 SDValue ThisElt, LastElt;
5018 bool LastIsNonZero = (NonZeros & (1 << (i-1))) != 0;
5019 if (LastIsNonZero) {
5020 LastElt = DAG.getNode(ISD::ZERO_EXTEND, dl,
5021 MVT::i16, Op.getOperand(i-1));
5023 if (ThisIsNonZero) {
5024 ThisElt = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i16, Op.getOperand(i));
5025 ThisElt = DAG.getNode(ISD::SHL, dl, MVT::i16,
5026 ThisElt, DAG.getConstant(8, dl, MVT::i8));
5028 ThisElt = DAG.getNode(ISD::OR, dl, MVT::i16, ThisElt, LastElt);
5032 if (ThisElt.getNode())
5033 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, V, ThisElt,
5034 DAG.getIntPtrConstant(i/2, dl));
5038 return DAG.getBitcast(MVT::v16i8, V);
5041 /// Custom lower build_vector of v8i16.
5042 static SDValue LowerBuildVectorv8i16(SDValue Op, unsigned NonZeros,
5043 unsigned NumNonZero, unsigned NumZero,
5045 const X86Subtarget* Subtarget,
5046 const TargetLowering &TLI) {
5053 for (unsigned i = 0; i < 8; ++i) {
5054 bool isNonZero = (NonZeros & (1 << i)) != 0;
5058 V = getZeroVector(MVT::v8i16, Subtarget, DAG, dl);
5060 V = DAG.getUNDEF(MVT::v8i16);
5063 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl,
5064 MVT::v8i16, V, Op.getOperand(i),
5065 DAG.getIntPtrConstant(i, dl));
5072 /// Custom lower build_vector of v4i32 or v4f32.
5073 static SDValue LowerBuildVectorv4x32(SDValue Op, SelectionDAG &DAG,
5074 const X86Subtarget *Subtarget,
5075 const TargetLowering &TLI) {
5076 // Find all zeroable elements.
5077 std::bitset<4> Zeroable;
5078 for (int i=0; i < 4; ++i) {
5079 SDValue Elt = Op->getOperand(i);
5080 Zeroable[i] = (Elt.getOpcode() == ISD::UNDEF || X86::isZeroNode(Elt));
5082 assert(Zeroable.size() - Zeroable.count() > 1 &&
5083 "We expect at least two non-zero elements!");
5085 // We only know how to deal with build_vector nodes where elements are either
5086 // zeroable or extract_vector_elt with constant index.
5087 SDValue FirstNonZero;
5088 unsigned FirstNonZeroIdx;
5089 for (unsigned i=0; i < 4; ++i) {
5092 SDValue Elt = Op->getOperand(i);
5093 if (Elt.getOpcode() != ISD::EXTRACT_VECTOR_ELT ||
5094 !isa<ConstantSDNode>(Elt.getOperand(1)))
5096 // Make sure that this node is extracting from a 128-bit vector.
5097 MVT VT = Elt.getOperand(0).getSimpleValueType();
5098 if (!VT.is128BitVector())
5100 if (!FirstNonZero.getNode()) {
5102 FirstNonZeroIdx = i;
5106 assert(FirstNonZero.getNode() && "Unexpected build vector of all zeros!");
5107 SDValue V1 = FirstNonZero.getOperand(0);
5108 MVT VT = V1.getSimpleValueType();
5110 // See if this build_vector can be lowered as a blend with zero.
5112 unsigned EltMaskIdx, EltIdx;
5114 for (EltIdx = 0; EltIdx < 4; ++EltIdx) {
5115 if (Zeroable[EltIdx]) {
5116 // The zero vector will be on the right hand side.
5117 Mask[EltIdx] = EltIdx+4;
5121 Elt = Op->getOperand(EltIdx);
5122 // By construction, Elt is a EXTRACT_VECTOR_ELT with constant index.
5123 EltMaskIdx = cast<ConstantSDNode>(Elt.getOperand(1))->getZExtValue();
5124 if (Elt.getOperand(0) != V1 || EltMaskIdx != EltIdx)
5126 Mask[EltIdx] = EltIdx;
5130 // Let the shuffle legalizer deal with blend operations.
5131 SDValue VZero = getZeroVector(VT, Subtarget, DAG, SDLoc(Op));
5132 if (V1.getSimpleValueType() != VT)
5133 V1 = DAG.getNode(ISD::BITCAST, SDLoc(V1), VT, V1);
5134 return DAG.getVectorShuffle(VT, SDLoc(V1), V1, VZero, &Mask[0]);
5137 // See if we can lower this build_vector to a INSERTPS.
5138 if (!Subtarget->hasSSE41())
5141 SDValue V2 = Elt.getOperand(0);
5142 if (Elt == FirstNonZero && EltIdx == FirstNonZeroIdx)
5145 bool CanFold = true;
5146 for (unsigned i = EltIdx + 1; i < 4 && CanFold; ++i) {
5150 SDValue Current = Op->getOperand(i);
5151 SDValue SrcVector = Current->getOperand(0);
5154 CanFold = SrcVector == V1 &&
5155 cast<ConstantSDNode>(Current.getOperand(1))->getZExtValue() == i;
5161 assert(V1.getNode() && "Expected at least two non-zero elements!");
5162 if (V1.getSimpleValueType() != MVT::v4f32)
5163 V1 = DAG.getNode(ISD::BITCAST, SDLoc(V1), MVT::v4f32, V1);
5164 if (V2.getSimpleValueType() != MVT::v4f32)
5165 V2 = DAG.getNode(ISD::BITCAST, SDLoc(V2), MVT::v4f32, V2);
5167 // Ok, we can emit an INSERTPS instruction.
5168 unsigned ZMask = Zeroable.to_ulong();
5170 unsigned InsertPSMask = EltMaskIdx << 6 | EltIdx << 4 | ZMask;
5171 assert((InsertPSMask & ~0xFFu) == 0 && "Invalid mask!");
5173 SDValue Result = DAG.getNode(X86ISD::INSERTPS, DL, MVT::v4f32, V1, V2,
5174 DAG.getIntPtrConstant(InsertPSMask, DL));
5175 return DAG.getBitcast(VT, Result);
5178 /// Return a vector logical shift node.
5179 static SDValue getVShift(bool isLeft, EVT VT, SDValue SrcOp,
5180 unsigned NumBits, SelectionDAG &DAG,
5181 const TargetLowering &TLI, SDLoc dl) {
5182 assert(VT.is128BitVector() && "Unknown type for VShift");
5183 MVT ShVT = MVT::v2i64;
5184 unsigned Opc = isLeft ? X86ISD::VSHLDQ : X86ISD::VSRLDQ;
5185 SrcOp = DAG.getBitcast(ShVT, SrcOp);
5186 MVT ScalarShiftTy = TLI.getScalarShiftAmountTy(DAG.getDataLayout(), VT);
5187 assert(NumBits % 8 == 0 && "Only support byte sized shifts");
5188 SDValue ShiftVal = DAG.getConstant(NumBits/8, dl, ScalarShiftTy);
5189 return DAG.getBitcast(VT, DAG.getNode(Opc, dl, ShVT, SrcOp, ShiftVal));
5193 LowerAsSplatVectorLoad(SDValue SrcOp, MVT VT, SDLoc dl, SelectionDAG &DAG) {
5195 // Check if the scalar load can be widened into a vector load. And if
5196 // the address is "base + cst" see if the cst can be "absorbed" into
5197 // the shuffle mask.
5198 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(SrcOp)) {
5199 SDValue Ptr = LD->getBasePtr();
5200 if (!ISD::isNormalLoad(LD) || LD->isVolatile())
5202 EVT PVT = LD->getValueType(0);
5203 if (PVT != MVT::i32 && PVT != MVT::f32)
5208 if (FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr)) {
5209 FI = FINode->getIndex();
5211 } else if (DAG.isBaseWithConstantOffset(Ptr) &&
5212 isa<FrameIndexSDNode>(Ptr.getOperand(0))) {
5213 FI = cast<FrameIndexSDNode>(Ptr.getOperand(0))->getIndex();
5214 Offset = Ptr.getConstantOperandVal(1);
5215 Ptr = Ptr.getOperand(0);
5220 // FIXME: 256-bit vector instructions don't require a strict alignment,
5221 // improve this code to support it better.
5222 unsigned RequiredAlign = VT.getSizeInBits()/8;
5223 SDValue Chain = LD->getChain();
5224 // Make sure the stack object alignment is at least 16 or 32.
5225 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
5226 if (DAG.InferPtrAlignment(Ptr) < RequiredAlign) {
5227 if (MFI->isFixedObjectIndex(FI)) {
5228 // Can't change the alignment. FIXME: It's possible to compute
5229 // the exact stack offset and reference FI + adjust offset instead.
5230 // If someone *really* cares about this. That's the way to implement it.
5233 MFI->setObjectAlignment(FI, RequiredAlign);
5237 // (Offset % 16 or 32) must be multiple of 4. Then address is then
5238 // Ptr + (Offset & ~15).
5241 if ((Offset % RequiredAlign) & 3)
5243 int64_t StartOffset = Offset & ~int64_t(RequiredAlign - 1);
5246 Ptr = DAG.getNode(ISD::ADD, DL, Ptr.getValueType(), Ptr,
5247 DAG.getConstant(StartOffset, DL, Ptr.getValueType()));
5250 int EltNo = (Offset - StartOffset) >> 2;
5251 unsigned NumElems = VT.getVectorNumElements();
5253 EVT NVT = EVT::getVectorVT(*DAG.getContext(), PVT, NumElems);
5254 SDValue V1 = DAG.getLoad(NVT, dl, Chain, Ptr,
5255 LD->getPointerInfo().getWithOffset(StartOffset),
5256 false, false, false, 0);
5258 SmallVector<int, 8> Mask(NumElems, EltNo);
5260 return DAG.getVectorShuffle(NVT, dl, V1, DAG.getUNDEF(NVT), &Mask[0]);
5266 /// Given the initializing elements 'Elts' of a vector of type 'VT', see if the
5267 /// elements can be replaced by a single large load which has the same value as
5268 /// a build_vector or insert_subvector whose loaded operands are 'Elts'.
5270 /// Example: <load i32 *a, load i32 *a+4, undef, undef> -> zextload a
5272 /// FIXME: we'd also like to handle the case where the last elements are zero
5273 /// rather than undef via VZEXT_LOAD, but we do not detect that case today.
5274 /// There's even a handy isZeroNode for that purpose.
5275 static SDValue EltsFromConsecutiveLoads(EVT VT, ArrayRef<SDValue> Elts,
5276 SDLoc &DL, SelectionDAG &DAG,
5277 bool isAfterLegalize) {
5278 unsigned NumElems = Elts.size();
5280 LoadSDNode *LDBase = nullptr;
5281 unsigned LastLoadedElt = -1U;
5283 // For each element in the initializer, see if we've found a load or an undef.
5284 // If we don't find an initial load element, or later load elements are
5285 // non-consecutive, bail out.
5286 for (unsigned i = 0; i < NumElems; ++i) {
5287 SDValue Elt = Elts[i];
5288 // Look through a bitcast.
5289 if (Elt.getNode() && Elt.getOpcode() == ISD::BITCAST)
5290 Elt = Elt.getOperand(0);
5291 if (!Elt.getNode() ||
5292 (Elt.getOpcode() != ISD::UNDEF && !ISD::isNON_EXTLoad(Elt.getNode())))
5295 if (Elt.getNode()->getOpcode() == ISD::UNDEF)
5297 LDBase = cast<LoadSDNode>(Elt.getNode());
5301 if (Elt.getOpcode() == ISD::UNDEF)
5304 LoadSDNode *LD = cast<LoadSDNode>(Elt);
5305 EVT LdVT = Elt.getValueType();
5306 // Each loaded element must be the correct fractional portion of the
5307 // requested vector load.
5308 if (LdVT.getSizeInBits() != VT.getSizeInBits() / NumElems)
5310 if (!DAG.isConsecutiveLoad(LD, LDBase, LdVT.getSizeInBits() / 8, i))
5315 // If we have found an entire vector of loads and undefs, then return a large
5316 // load of the entire vector width starting at the base pointer. If we found
5317 // consecutive loads for the low half, generate a vzext_load node.
5318 if (LastLoadedElt == NumElems - 1) {
5319 assert(LDBase && "Did not find base load for merging consecutive loads");
5320 EVT EltVT = LDBase->getValueType(0);
5321 // Ensure that the input vector size for the merged loads matches the
5322 // cumulative size of the input elements.
5323 if (VT.getSizeInBits() != EltVT.getSizeInBits() * NumElems)
5326 if (isAfterLegalize &&
5327 !DAG.getTargetLoweringInfo().isOperationLegal(ISD::LOAD, VT))
5330 SDValue NewLd = SDValue();
5332 NewLd = DAG.getLoad(VT, DL, LDBase->getChain(), LDBase->getBasePtr(),
5333 LDBase->getPointerInfo(), LDBase->isVolatile(),
5334 LDBase->isNonTemporal(), LDBase->isInvariant(),
5335 LDBase->getAlignment());
5337 if (LDBase->hasAnyUseOfValue(1)) {
5338 SDValue NewChain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other,
5340 SDValue(NewLd.getNode(), 1));
5341 DAG.ReplaceAllUsesOfValueWith(SDValue(LDBase, 1), NewChain);
5342 DAG.UpdateNodeOperands(NewChain.getNode(), SDValue(LDBase, 1),
5343 SDValue(NewLd.getNode(), 1));
5349 //TODO: The code below fires only for for loading the low v2i32 / v2f32
5350 //of a v4i32 / v4f32. It's probably worth generalizing.
5351 EVT EltVT = VT.getVectorElementType();
5352 if (NumElems == 4 && LastLoadedElt == 1 && (EltVT.getSizeInBits() == 32) &&
5353 DAG.getTargetLoweringInfo().isTypeLegal(MVT::v2i64)) {
5354 SDVTList Tys = DAG.getVTList(MVT::v2i64, MVT::Other);
5355 SDValue Ops[] = { LDBase->getChain(), LDBase->getBasePtr() };
5357 DAG.getMemIntrinsicNode(X86ISD::VZEXT_LOAD, DL, Tys, Ops, MVT::i64,
5358 LDBase->getPointerInfo(),
5359 LDBase->getAlignment(),
5360 false/*isVolatile*/, true/*ReadMem*/,
5363 // Make sure the newly-created LOAD is in the same position as LDBase in
5364 // terms of dependency. We create a TokenFactor for LDBase and ResNode, and
5365 // update uses of LDBase's output chain to use the TokenFactor.
5366 if (LDBase->hasAnyUseOfValue(1)) {
5367 SDValue NewChain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other,
5368 SDValue(LDBase, 1), SDValue(ResNode.getNode(), 1));
5369 DAG.ReplaceAllUsesOfValueWith(SDValue(LDBase, 1), NewChain);
5370 DAG.UpdateNodeOperands(NewChain.getNode(), SDValue(LDBase, 1),
5371 SDValue(ResNode.getNode(), 1));
5374 return DAG.getBitcast(VT, ResNode);
5379 /// LowerVectorBroadcast - Attempt to use the vbroadcast instruction
5380 /// to generate a splat value for the following cases:
5381 /// 1. A splat BUILD_VECTOR which uses a single scalar load, or a constant.
5382 /// 2. A splat shuffle which uses a scalar_to_vector node which comes from
5383 /// a scalar load, or a constant.
5384 /// The VBROADCAST node is returned when a pattern is found,
5385 /// or SDValue() otherwise.
5386 static SDValue LowerVectorBroadcast(SDValue Op, const X86Subtarget* Subtarget,
5387 SelectionDAG &DAG) {
5388 // VBROADCAST requires AVX.
5389 // TODO: Splats could be generated for non-AVX CPUs using SSE
5390 // instructions, but there's less potential gain for only 128-bit vectors.
5391 if (!Subtarget->hasAVX())
5394 MVT VT = Op.getSimpleValueType();
5397 assert((VT.is128BitVector() || VT.is256BitVector() || VT.is512BitVector()) &&
5398 "Unsupported vector type for broadcast.");
5403 switch (Op.getOpcode()) {
5405 // Unknown pattern found.
5408 case ISD::BUILD_VECTOR: {
5409 auto *BVOp = cast<BuildVectorSDNode>(Op.getNode());
5410 BitVector UndefElements;
5411 SDValue Splat = BVOp->getSplatValue(&UndefElements);
5413 // We need a splat of a single value to use broadcast, and it doesn't
5414 // make any sense if the value is only in one element of the vector.
5415 if (!Splat || (VT.getVectorNumElements() - UndefElements.count()) <= 1)
5419 ConstSplatVal = (Ld.getOpcode() == ISD::Constant ||
5420 Ld.getOpcode() == ISD::ConstantFP);
5422 // Make sure that all of the users of a non-constant load are from the
5423 // BUILD_VECTOR node.
5424 if (!ConstSplatVal && !BVOp->isOnlyUserOf(Ld.getNode()))
5429 case ISD::VECTOR_SHUFFLE: {
5430 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
5432 // Shuffles must have a splat mask where the first element is
5434 if ((!SVOp->isSplat()) || SVOp->getMaskElt(0) != 0)
5437 SDValue Sc = Op.getOperand(0);
5438 if (Sc.getOpcode() != ISD::SCALAR_TO_VECTOR &&
5439 Sc.getOpcode() != ISD::BUILD_VECTOR) {
5441 if (!Subtarget->hasInt256())
5444 // Use the register form of the broadcast instruction available on AVX2.
5445 if (VT.getSizeInBits() >= 256)
5446 Sc = Extract128BitVector(Sc, 0, DAG, dl);
5447 return DAG.getNode(X86ISD::VBROADCAST, dl, VT, Sc);
5450 Ld = Sc.getOperand(0);
5451 ConstSplatVal = (Ld.getOpcode() == ISD::Constant ||
5452 Ld.getOpcode() == ISD::ConstantFP);
5454 // The scalar_to_vector node and the suspected
5455 // load node must have exactly one user.
5456 // Constants may have multiple users.
5458 // AVX-512 has register version of the broadcast
5459 bool hasRegVer = Subtarget->hasAVX512() && VT.is512BitVector() &&
5460 Ld.getValueType().getSizeInBits() >= 32;
5461 if (!ConstSplatVal && ((!Sc.hasOneUse() || !Ld.hasOneUse()) &&
5468 unsigned ScalarSize = Ld.getValueType().getSizeInBits();
5469 bool IsGE256 = (VT.getSizeInBits() >= 256);
5471 // When optimizing for size, generate up to 5 extra bytes for a broadcast
5472 // instruction to save 8 or more bytes of constant pool data.
5473 // TODO: If multiple splats are generated to load the same constant,
5474 // it may be detrimental to overall size. There needs to be a way to detect
5475 // that condition to know if this is truly a size win.
5476 bool OptForSize = DAG.getMachineFunction().getFunction()->optForSize();
5478 // Handle broadcasting a single constant scalar from the constant pool
5480 // On Sandybridge (no AVX2), it is still better to load a constant vector
5481 // from the constant pool and not to broadcast it from a scalar.
5482 // But override that restriction when optimizing for size.
5483 // TODO: Check if splatting is recommended for other AVX-capable CPUs.
5484 if (ConstSplatVal && (Subtarget->hasAVX2() || OptForSize)) {
5485 EVT CVT = Ld.getValueType();
5486 assert(!CVT.isVector() && "Must not broadcast a vector type");
5488 // Splat f32, i32, v4f64, v4i64 in all cases with AVX2.
5489 // For size optimization, also splat v2f64 and v2i64, and for size opt
5490 // with AVX2, also splat i8 and i16.
5491 // With pattern matching, the VBROADCAST node may become a VMOVDDUP.
5492 if (ScalarSize == 32 || (IsGE256 && ScalarSize == 64) ||
5493 (OptForSize && (ScalarSize == 64 || Subtarget->hasAVX2()))) {
5494 const Constant *C = nullptr;
5495 if (ConstantSDNode *CI = dyn_cast<ConstantSDNode>(Ld))
5496 C = CI->getConstantIntValue();
5497 else if (ConstantFPSDNode *CF = dyn_cast<ConstantFPSDNode>(Ld))
5498 C = CF->getConstantFPValue();
5500 assert(C && "Invalid constant type");
5502 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
5504 DAG.getConstantPool(C, TLI.getPointerTy(DAG.getDataLayout()));
5505 unsigned Alignment = cast<ConstantPoolSDNode>(CP)->getAlignment();
5507 CVT, dl, DAG.getEntryNode(), CP,
5508 MachinePointerInfo::getConstantPool(DAG.getMachineFunction()), false,
5509 false, false, Alignment);
5511 return DAG.getNode(X86ISD::VBROADCAST, dl, VT, Ld);
5515 bool IsLoad = ISD::isNormalLoad(Ld.getNode());
5517 // Handle AVX2 in-register broadcasts.
5518 if (!IsLoad && Subtarget->hasInt256() &&
5519 (ScalarSize == 32 || (IsGE256 && ScalarSize == 64)))
5520 return DAG.getNode(X86ISD::VBROADCAST, dl, VT, Ld);
5522 // The scalar source must be a normal load.
5526 if (ScalarSize == 32 || (IsGE256 && ScalarSize == 64) ||
5527 (Subtarget->hasVLX() && ScalarSize == 64))
5528 return DAG.getNode(X86ISD::VBROADCAST, dl, VT, Ld);
5530 // The integer check is needed for the 64-bit into 128-bit so it doesn't match
5531 // double since there is no vbroadcastsd xmm
5532 if (Subtarget->hasInt256() && Ld.getValueType().isInteger()) {
5533 if (ScalarSize == 8 || ScalarSize == 16 || ScalarSize == 64)
5534 return DAG.getNode(X86ISD::VBROADCAST, dl, VT, Ld);
5537 // Unsupported broadcast.
5541 /// \brief For an EXTRACT_VECTOR_ELT with a constant index return the real
5542 /// underlying vector and index.
5544 /// Modifies \p ExtractedFromVec to the real vector and returns the real
5546 static int getUnderlyingExtractedFromVec(SDValue &ExtractedFromVec,
5548 int Idx = cast<ConstantSDNode>(ExtIdx)->getZExtValue();
5549 if (!isa<ShuffleVectorSDNode>(ExtractedFromVec))
5552 // For 256-bit vectors, LowerEXTRACT_VECTOR_ELT_SSE4 may have already
5554 // (extract_vector_elt (v8f32 %vreg1), Constant<6>)
5556 // (extract_vector_elt (vector_shuffle<2,u,u,u>
5557 // (extract_subvector (v8f32 %vreg0), Constant<4>),
5560 // In this case the vector is the extract_subvector expression and the index
5561 // is 2, as specified by the shuffle.
5562 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(ExtractedFromVec);
5563 SDValue ShuffleVec = SVOp->getOperand(0);
5564 MVT ShuffleVecVT = ShuffleVec.getSimpleValueType();
5565 assert(ShuffleVecVT.getVectorElementType() ==
5566 ExtractedFromVec.getSimpleValueType().getVectorElementType());
5568 int ShuffleIdx = SVOp->getMaskElt(Idx);
5569 if (isUndefOrInRange(ShuffleIdx, 0, ShuffleVecVT.getVectorNumElements())) {
5570 ExtractedFromVec = ShuffleVec;
5576 static SDValue buildFromShuffleMostly(SDValue Op, SelectionDAG &DAG) {
5577 MVT VT = Op.getSimpleValueType();
5579 // Skip if insert_vec_elt is not supported.
5580 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
5581 if (!TLI.isOperationLegalOrCustom(ISD::INSERT_VECTOR_ELT, VT))
5585 unsigned NumElems = Op.getNumOperands();
5589 SmallVector<unsigned, 4> InsertIndices;
5590 SmallVector<int, 8> Mask(NumElems, -1);
5592 for (unsigned i = 0; i != NumElems; ++i) {
5593 unsigned Opc = Op.getOperand(i).getOpcode();
5595 if (Opc == ISD::UNDEF)
5598 if (Opc != ISD::EXTRACT_VECTOR_ELT) {
5599 // Quit if more than 1 elements need inserting.
5600 if (InsertIndices.size() > 1)
5603 InsertIndices.push_back(i);
5607 SDValue ExtractedFromVec = Op.getOperand(i).getOperand(0);
5608 SDValue ExtIdx = Op.getOperand(i).getOperand(1);
5609 // Quit if non-constant index.
5610 if (!isa<ConstantSDNode>(ExtIdx))
5612 int Idx = getUnderlyingExtractedFromVec(ExtractedFromVec, ExtIdx);
5614 // Quit if extracted from vector of different type.
5615 if (ExtractedFromVec.getValueType() != VT)
5618 if (!VecIn1.getNode())
5619 VecIn1 = ExtractedFromVec;
5620 else if (VecIn1 != ExtractedFromVec) {
5621 if (!VecIn2.getNode())
5622 VecIn2 = ExtractedFromVec;
5623 else if (VecIn2 != ExtractedFromVec)
5624 // Quit if more than 2 vectors to shuffle
5628 if (ExtractedFromVec == VecIn1)
5630 else if (ExtractedFromVec == VecIn2)
5631 Mask[i] = Idx + NumElems;
5634 if (!VecIn1.getNode())
5637 VecIn2 = VecIn2.getNode() ? VecIn2 : DAG.getUNDEF(VT);
5638 SDValue NV = DAG.getVectorShuffle(VT, DL, VecIn1, VecIn2, &Mask[0]);
5639 for (unsigned i = 0, e = InsertIndices.size(); i != e; ++i) {
5640 unsigned Idx = InsertIndices[i];
5641 NV = DAG.getNode(ISD::INSERT_VECTOR_ELT, DL, VT, NV, Op.getOperand(Idx),
5642 DAG.getIntPtrConstant(Idx, DL));
5648 static SDValue ConvertI1VectorToInteger(SDValue Op, SelectionDAG &DAG) {
5649 assert(ISD::isBuildVectorOfConstantSDNodes(Op.getNode()) &&
5650 Op.getScalarValueSizeInBits() == 1 &&
5651 "Can not convert non-constant vector");
5652 uint64_t Immediate = 0;
5653 for (unsigned idx = 0, e = Op.getNumOperands(); idx < e; ++idx) {
5654 SDValue In = Op.getOperand(idx);
5655 if (In.getOpcode() != ISD::UNDEF)
5656 Immediate |= cast<ConstantSDNode>(In)->getZExtValue() << idx;
5660 MVT::getIntegerVT(std::max((int)Op.getValueType().getSizeInBits(), 8));
5661 return DAG.getConstant(Immediate, dl, VT);
5663 // Lower BUILD_VECTOR operation for v8i1 and v16i1 types.
5665 X86TargetLowering::LowerBUILD_VECTORvXi1(SDValue Op, SelectionDAG &DAG) const {
5667 MVT VT = Op.getSimpleValueType();
5668 assert((VT.getVectorElementType() == MVT::i1) &&
5669 "Unexpected type in LowerBUILD_VECTORvXi1!");
5672 if (ISD::isBuildVectorAllZeros(Op.getNode())) {
5673 SDValue Cst = DAG.getTargetConstant(0, dl, MVT::i1);
5674 SmallVector<SDValue, 16> Ops(VT.getVectorNumElements(), Cst);
5675 return DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Ops);
5678 if (ISD::isBuildVectorAllOnes(Op.getNode())) {
5679 SDValue Cst = DAG.getTargetConstant(1, dl, MVT::i1);
5680 SmallVector<SDValue, 16> Ops(VT.getVectorNumElements(), Cst);
5681 return DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Ops);
5684 if (ISD::isBuildVectorOfConstantSDNodes(Op.getNode())) {
5685 SDValue Imm = ConvertI1VectorToInteger(Op, DAG);
5686 if (Imm.getValueSizeInBits() == VT.getSizeInBits())
5687 return DAG.getBitcast(VT, Imm);
5688 SDValue ExtVec = DAG.getBitcast(MVT::v8i1, Imm);
5689 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT, ExtVec,
5690 DAG.getIntPtrConstant(0, dl));
5693 // Vector has one or more non-const elements
5694 uint64_t Immediate = 0;
5695 SmallVector<unsigned, 16> NonConstIdx;
5696 bool IsSplat = true;
5697 bool HasConstElts = false;
5699 for (unsigned idx = 0, e = Op.getNumOperands(); idx < e; ++idx) {
5700 SDValue In = Op.getOperand(idx);
5701 if (In.getOpcode() == ISD::UNDEF)
5703 if (!isa<ConstantSDNode>(In))
5704 NonConstIdx.push_back(idx);
5706 Immediate |= cast<ConstantSDNode>(In)->getZExtValue() << idx;
5707 HasConstElts = true;
5711 else if (In != Op.getOperand(SplatIdx))
5715 // for splat use " (select i1 splat_elt, all-ones, all-zeroes)"
5717 return DAG.getNode(ISD::SELECT, dl, VT, Op.getOperand(SplatIdx),
5718 DAG.getConstant(1, dl, VT),
5719 DAG.getConstant(0, dl, VT));
5721 // insert elements one by one
5725 MVT ImmVT = MVT::getIntegerVT(std::max((int)VT.getSizeInBits(), 8));
5726 Imm = DAG.getConstant(Immediate, dl, ImmVT);
5728 else if (HasConstElts)
5729 Imm = DAG.getConstant(0, dl, VT);
5731 Imm = DAG.getUNDEF(VT);
5732 if (Imm.getValueSizeInBits() == VT.getSizeInBits())
5733 DstVec = DAG.getBitcast(VT, Imm);
5735 SDValue ExtVec = DAG.getBitcast(MVT::v8i1, Imm);
5736 DstVec = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT, ExtVec,
5737 DAG.getIntPtrConstant(0, dl));
5740 for (unsigned i = 0; i < NonConstIdx.size(); ++i) {
5741 unsigned InsertIdx = NonConstIdx[i];
5742 DstVec = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, DstVec,
5743 Op.getOperand(InsertIdx),
5744 DAG.getIntPtrConstant(InsertIdx, dl));
5749 /// \brief Return true if \p N implements a horizontal binop and return the
5750 /// operands for the horizontal binop into V0 and V1.
5752 /// This is a helper function of LowerToHorizontalOp().
5753 /// This function checks that the build_vector \p N in input implements a
5754 /// horizontal operation. Parameter \p Opcode defines the kind of horizontal
5755 /// operation to match.
5756 /// For example, if \p Opcode is equal to ISD::ADD, then this function
5757 /// checks if \p N implements a horizontal arithmetic add; if instead \p Opcode
5758 /// is equal to ISD::SUB, then this function checks if this is a horizontal
5761 /// This function only analyzes elements of \p N whose indices are
5762 /// in range [BaseIdx, LastIdx).
5763 static bool isHorizontalBinOp(const BuildVectorSDNode *N, unsigned Opcode,
5765 unsigned BaseIdx, unsigned LastIdx,
5766 SDValue &V0, SDValue &V1) {
5767 EVT VT = N->getValueType(0);
5769 assert(BaseIdx * 2 <= LastIdx && "Invalid Indices in input!");
5770 assert(VT.isVector() && VT.getVectorNumElements() >= LastIdx &&
5771 "Invalid Vector in input!");
5773 bool IsCommutable = (Opcode == ISD::ADD || Opcode == ISD::FADD);
5774 bool CanFold = true;
5775 unsigned ExpectedVExtractIdx = BaseIdx;
5776 unsigned NumElts = LastIdx - BaseIdx;
5777 V0 = DAG.getUNDEF(VT);
5778 V1 = DAG.getUNDEF(VT);
5780 // Check if N implements a horizontal binop.
5781 for (unsigned i = 0, e = NumElts; i != e && CanFold; ++i) {
5782 SDValue Op = N->getOperand(i + BaseIdx);
5785 if (Op->getOpcode() == ISD::UNDEF) {
5786 // Update the expected vector extract index.
5787 if (i * 2 == NumElts)
5788 ExpectedVExtractIdx = BaseIdx;
5789 ExpectedVExtractIdx += 2;
5793 CanFold = Op->getOpcode() == Opcode && Op->hasOneUse();
5798 SDValue Op0 = Op.getOperand(0);
5799 SDValue Op1 = Op.getOperand(1);
5801 // Try to match the following pattern:
5802 // (BINOP (extract_vector_elt A, I), (extract_vector_elt A, I+1))
5803 CanFold = (Op0.getOpcode() == ISD::EXTRACT_VECTOR_ELT &&
5804 Op1.getOpcode() == ISD::EXTRACT_VECTOR_ELT &&
5805 Op0.getOperand(0) == Op1.getOperand(0) &&
5806 isa<ConstantSDNode>(Op0.getOperand(1)) &&
5807 isa<ConstantSDNode>(Op1.getOperand(1)));
5811 unsigned I0 = cast<ConstantSDNode>(Op0.getOperand(1))->getZExtValue();
5812 unsigned I1 = cast<ConstantSDNode>(Op1.getOperand(1))->getZExtValue();
5814 if (i * 2 < NumElts) {
5815 if (V0.getOpcode() == ISD::UNDEF) {
5816 V0 = Op0.getOperand(0);
5817 if (V0.getValueType() != VT)
5821 if (V1.getOpcode() == ISD::UNDEF) {
5822 V1 = Op0.getOperand(0);
5823 if (V1.getValueType() != VT)
5826 if (i * 2 == NumElts)
5827 ExpectedVExtractIdx = BaseIdx;
5830 SDValue Expected = (i * 2 < NumElts) ? V0 : V1;
5831 if (I0 == ExpectedVExtractIdx)
5832 CanFold = I1 == I0 + 1 && Op0.getOperand(0) == Expected;
5833 else if (IsCommutable && I1 == ExpectedVExtractIdx) {
5834 // Try to match the following dag sequence:
5835 // (BINOP (extract_vector_elt A, I+1), (extract_vector_elt A, I))
5836 CanFold = I0 == I1 + 1 && Op1.getOperand(0) == Expected;
5840 ExpectedVExtractIdx += 2;
5846 /// \brief Emit a sequence of two 128-bit horizontal add/sub followed by
5847 /// a concat_vector.
5849 /// This is a helper function of LowerToHorizontalOp().
5850 /// This function expects two 256-bit vectors called V0 and V1.
5851 /// At first, each vector is split into two separate 128-bit vectors.
5852 /// Then, the resulting 128-bit vectors are used to implement two
5853 /// horizontal binary operations.
5855 /// The kind of horizontal binary operation is defined by \p X86Opcode.
5857 /// \p Mode specifies how the 128-bit parts of V0 and V1 are passed in input to
5858 /// the two new horizontal binop.
5859 /// When Mode is set, the first horizontal binop dag node would take as input
5860 /// the lower 128-bit of V0 and the upper 128-bit of V0. The second
5861 /// horizontal binop dag node would take as input the lower 128-bit of V1
5862 /// and the upper 128-bit of V1.
5864 /// HADD V0_LO, V0_HI
5865 /// HADD V1_LO, V1_HI
5867 /// Otherwise, the first horizontal binop dag node takes as input the lower
5868 /// 128-bit of V0 and the lower 128-bit of V1, and the second horizontal binop
5869 /// dag node takes the upper 128-bit of V0 and the upper 128-bit of V1.
5871 /// HADD V0_LO, V1_LO
5872 /// HADD V0_HI, V1_HI
5874 /// If \p isUndefLO is set, then the algorithm propagates UNDEF to the lower
5875 /// 128-bits of the result. If \p isUndefHI is set, then UNDEF is propagated to
5876 /// the upper 128-bits of the result.
5877 static SDValue ExpandHorizontalBinOp(const SDValue &V0, const SDValue &V1,
5878 SDLoc DL, SelectionDAG &DAG,
5879 unsigned X86Opcode, bool Mode,
5880 bool isUndefLO, bool isUndefHI) {
5881 EVT VT = V0.getValueType();
5882 assert(VT.is256BitVector() && VT == V1.getValueType() &&
5883 "Invalid nodes in input!");
5885 unsigned NumElts = VT.getVectorNumElements();
5886 SDValue V0_LO = Extract128BitVector(V0, 0, DAG, DL);
5887 SDValue V0_HI = Extract128BitVector(V0, NumElts/2, DAG, DL);
5888 SDValue V1_LO = Extract128BitVector(V1, 0, DAG, DL);
5889 SDValue V1_HI = Extract128BitVector(V1, NumElts/2, DAG, DL);
5890 EVT NewVT = V0_LO.getValueType();
5892 SDValue LO = DAG.getUNDEF(NewVT);
5893 SDValue HI = DAG.getUNDEF(NewVT);
5896 // Don't emit a horizontal binop if the result is expected to be UNDEF.
5897 if (!isUndefLO && V0->getOpcode() != ISD::UNDEF)
5898 LO = DAG.getNode(X86Opcode, DL, NewVT, V0_LO, V0_HI);
5899 if (!isUndefHI && V1->getOpcode() != ISD::UNDEF)
5900 HI = DAG.getNode(X86Opcode, DL, NewVT, V1_LO, V1_HI);
5902 // Don't emit a horizontal binop if the result is expected to be UNDEF.
5903 if (!isUndefLO && (V0_LO->getOpcode() != ISD::UNDEF ||
5904 V1_LO->getOpcode() != ISD::UNDEF))
5905 LO = DAG.getNode(X86Opcode, DL, NewVT, V0_LO, V1_LO);
5907 if (!isUndefHI && (V0_HI->getOpcode() != ISD::UNDEF ||
5908 V1_HI->getOpcode() != ISD::UNDEF))
5909 HI = DAG.getNode(X86Opcode, DL, NewVT, V0_HI, V1_HI);
5912 return DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, LO, HI);
5915 /// Try to fold a build_vector that performs an 'addsub' to an X86ISD::ADDSUB
5917 static SDValue LowerToAddSub(const BuildVectorSDNode *BV,
5918 const X86Subtarget *Subtarget, SelectionDAG &DAG) {
5919 EVT VT = BV->getValueType(0);
5920 if ((!Subtarget->hasSSE3() || (VT != MVT::v4f32 && VT != MVT::v2f64)) &&
5921 (!Subtarget->hasAVX() || (VT != MVT::v8f32 && VT != MVT::v4f64)))
5925 unsigned NumElts = VT.getVectorNumElements();
5926 SDValue InVec0 = DAG.getUNDEF(VT);
5927 SDValue InVec1 = DAG.getUNDEF(VT);
5929 assert((VT == MVT::v8f32 || VT == MVT::v4f64 || VT == MVT::v4f32 ||
5930 VT == MVT::v2f64) && "build_vector with an invalid type found!");
5932 // Odd-numbered elements in the input build vector are obtained from
5933 // adding two integer/float elements.
5934 // Even-numbered elements in the input build vector are obtained from
5935 // subtracting two integer/float elements.
5936 unsigned ExpectedOpcode = ISD::FSUB;
5937 unsigned NextExpectedOpcode = ISD::FADD;
5938 bool AddFound = false;
5939 bool SubFound = false;
5941 for (unsigned i = 0, e = NumElts; i != e; ++i) {
5942 SDValue Op = BV->getOperand(i);
5944 // Skip 'undef' values.
5945 unsigned Opcode = Op.getOpcode();
5946 if (Opcode == ISD::UNDEF) {
5947 std::swap(ExpectedOpcode, NextExpectedOpcode);
5951 // Early exit if we found an unexpected opcode.
5952 if (Opcode != ExpectedOpcode)
5955 SDValue Op0 = Op.getOperand(0);
5956 SDValue Op1 = Op.getOperand(1);
5958 // Try to match the following pattern:
5959 // (BINOP (extract_vector_elt A, i), (extract_vector_elt B, i))
5960 // Early exit if we cannot match that sequence.
5961 if (Op0.getOpcode() != ISD::EXTRACT_VECTOR_ELT ||
5962 Op1.getOpcode() != ISD::EXTRACT_VECTOR_ELT ||
5963 !isa<ConstantSDNode>(Op0.getOperand(1)) ||
5964 !isa<ConstantSDNode>(Op1.getOperand(1)) ||
5965 Op0.getOperand(1) != Op1.getOperand(1))
5968 unsigned I0 = cast<ConstantSDNode>(Op0.getOperand(1))->getZExtValue();
5972 // We found a valid add/sub node. Update the information accordingly.
5978 // Update InVec0 and InVec1.
5979 if (InVec0.getOpcode() == ISD::UNDEF) {
5980 InVec0 = Op0.getOperand(0);
5981 if (InVec0.getValueType() != VT)
5984 if (InVec1.getOpcode() == ISD::UNDEF) {
5985 InVec1 = Op1.getOperand(0);
5986 if (InVec1.getValueType() != VT)
5990 // Make sure that operands in input to each add/sub node always
5991 // come from a same pair of vectors.
5992 if (InVec0 != Op0.getOperand(0)) {
5993 if (ExpectedOpcode == ISD::FSUB)
5996 // FADD is commutable. Try to commute the operands
5997 // and then test again.
5998 std::swap(Op0, Op1);
5999 if (InVec0 != Op0.getOperand(0))
6003 if (InVec1 != Op1.getOperand(0))
6006 // Update the pair of expected opcodes.
6007 std::swap(ExpectedOpcode, NextExpectedOpcode);
6010 // Don't try to fold this build_vector into an ADDSUB if the inputs are undef.
6011 if (AddFound && SubFound && InVec0.getOpcode() != ISD::UNDEF &&
6012 InVec1.getOpcode() != ISD::UNDEF)
6013 return DAG.getNode(X86ISD::ADDSUB, DL, VT, InVec0, InVec1);
6018 /// Lower BUILD_VECTOR to a horizontal add/sub operation if possible.
6019 static SDValue LowerToHorizontalOp(const BuildVectorSDNode *BV,
6020 const X86Subtarget *Subtarget,
6021 SelectionDAG &DAG) {
6022 EVT VT = BV->getValueType(0);
6023 unsigned NumElts = VT.getVectorNumElements();
6024 unsigned NumUndefsLO = 0;
6025 unsigned NumUndefsHI = 0;
6026 unsigned Half = NumElts/2;
6028 // Count the number of UNDEF operands in the build_vector in input.
6029 for (unsigned i = 0, e = Half; i != e; ++i)
6030 if (BV->getOperand(i)->getOpcode() == ISD::UNDEF)
6033 for (unsigned i = Half, e = NumElts; i != e; ++i)
6034 if (BV->getOperand(i)->getOpcode() == ISD::UNDEF)
6037 // Early exit if this is either a build_vector of all UNDEFs or all the
6038 // operands but one are UNDEF.
6039 if (NumUndefsLO + NumUndefsHI + 1 >= NumElts)
6043 SDValue InVec0, InVec1;
6044 if ((VT == MVT::v4f32 || VT == MVT::v2f64) && Subtarget->hasSSE3()) {
6045 // Try to match an SSE3 float HADD/HSUB.
6046 if (isHorizontalBinOp(BV, ISD::FADD, DAG, 0, NumElts, InVec0, InVec1))
6047 return DAG.getNode(X86ISD::FHADD, DL, VT, InVec0, InVec1);
6049 if (isHorizontalBinOp(BV, ISD::FSUB, DAG, 0, NumElts, InVec0, InVec1))
6050 return DAG.getNode(X86ISD::FHSUB, DL, VT, InVec0, InVec1);
6051 } else if ((VT == MVT::v4i32 || VT == MVT::v8i16) && Subtarget->hasSSSE3()) {
6052 // Try to match an SSSE3 integer HADD/HSUB.
6053 if (isHorizontalBinOp(BV, ISD::ADD, DAG, 0, NumElts, InVec0, InVec1))
6054 return DAG.getNode(X86ISD::HADD, DL, VT, InVec0, InVec1);
6056 if (isHorizontalBinOp(BV, ISD::SUB, DAG, 0, NumElts, InVec0, InVec1))
6057 return DAG.getNode(X86ISD::HSUB, DL, VT, InVec0, InVec1);
6060 if (!Subtarget->hasAVX())
6063 if ((VT == MVT::v8f32 || VT == MVT::v4f64)) {
6064 // Try to match an AVX horizontal add/sub of packed single/double
6065 // precision floating point values from 256-bit vectors.
6066 SDValue InVec2, InVec3;
6067 if (isHorizontalBinOp(BV, ISD::FADD, DAG, 0, Half, InVec0, InVec1) &&
6068 isHorizontalBinOp(BV, ISD::FADD, DAG, Half, NumElts, InVec2, InVec3) &&
6069 ((InVec0.getOpcode() == ISD::UNDEF ||
6070 InVec2.getOpcode() == ISD::UNDEF) || InVec0 == InVec2) &&
6071 ((InVec1.getOpcode() == ISD::UNDEF ||
6072 InVec3.getOpcode() == ISD::UNDEF) || InVec1 == InVec3))
6073 return DAG.getNode(X86ISD::FHADD, DL, VT, InVec0, InVec1);
6075 if (isHorizontalBinOp(BV, ISD::FSUB, DAG, 0, Half, InVec0, InVec1) &&
6076 isHorizontalBinOp(BV, ISD::FSUB, DAG, Half, NumElts, InVec2, InVec3) &&
6077 ((InVec0.getOpcode() == ISD::UNDEF ||
6078 InVec2.getOpcode() == ISD::UNDEF) || InVec0 == InVec2) &&
6079 ((InVec1.getOpcode() == ISD::UNDEF ||
6080 InVec3.getOpcode() == ISD::UNDEF) || InVec1 == InVec3))
6081 return DAG.getNode(X86ISD::FHSUB, DL, VT, InVec0, InVec1);
6082 } else if (VT == MVT::v8i32 || VT == MVT::v16i16) {
6083 // Try to match an AVX2 horizontal add/sub of signed integers.
6084 SDValue InVec2, InVec3;
6086 bool CanFold = true;
6088 if (isHorizontalBinOp(BV, ISD::ADD, DAG, 0, Half, InVec0, InVec1) &&
6089 isHorizontalBinOp(BV, ISD::ADD, DAG, Half, NumElts, InVec2, InVec3) &&
6090 ((InVec0.getOpcode() == ISD::UNDEF ||
6091 InVec2.getOpcode() == ISD::UNDEF) || InVec0 == InVec2) &&
6092 ((InVec1.getOpcode() == ISD::UNDEF ||
6093 InVec3.getOpcode() == ISD::UNDEF) || InVec1 == InVec3))
6094 X86Opcode = X86ISD::HADD;
6095 else if (isHorizontalBinOp(BV, ISD::SUB, DAG, 0, Half, InVec0, InVec1) &&
6096 isHorizontalBinOp(BV, ISD::SUB, DAG, Half, NumElts, InVec2, InVec3) &&
6097 ((InVec0.getOpcode() == ISD::UNDEF ||
6098 InVec2.getOpcode() == ISD::UNDEF) || InVec0 == InVec2) &&
6099 ((InVec1.getOpcode() == ISD::UNDEF ||
6100 InVec3.getOpcode() == ISD::UNDEF) || InVec1 == InVec3))
6101 X86Opcode = X86ISD::HSUB;
6106 // Fold this build_vector into a single horizontal add/sub.
6107 // Do this only if the target has AVX2.
6108 if (Subtarget->hasAVX2())
6109 return DAG.getNode(X86Opcode, DL, VT, InVec0, InVec1);
6111 // Do not try to expand this build_vector into a pair of horizontal
6112 // add/sub if we can emit a pair of scalar add/sub.
6113 if (NumUndefsLO + 1 == Half || NumUndefsHI + 1 == Half)
6116 // Convert this build_vector into a pair of horizontal binop followed by
6118 bool isUndefLO = NumUndefsLO == Half;
6119 bool isUndefHI = NumUndefsHI == Half;
6120 return ExpandHorizontalBinOp(InVec0, InVec1, DL, DAG, X86Opcode, false,
6121 isUndefLO, isUndefHI);
6125 if ((VT == MVT::v8f32 || VT == MVT::v4f64 || VT == MVT::v8i32 ||
6126 VT == MVT::v16i16) && Subtarget->hasAVX()) {
6128 if (isHorizontalBinOp(BV, ISD::ADD, DAG, 0, NumElts, InVec0, InVec1))
6129 X86Opcode = X86ISD::HADD;
6130 else if (isHorizontalBinOp(BV, ISD::SUB, DAG, 0, NumElts, InVec0, InVec1))
6131 X86Opcode = X86ISD::HSUB;
6132 else if (isHorizontalBinOp(BV, ISD::FADD, DAG, 0, NumElts, InVec0, InVec1))
6133 X86Opcode = X86ISD::FHADD;
6134 else if (isHorizontalBinOp(BV, ISD::FSUB, DAG, 0, NumElts, InVec0, InVec1))
6135 X86Opcode = X86ISD::FHSUB;
6139 // Don't try to expand this build_vector into a pair of horizontal add/sub
6140 // if we can simply emit a pair of scalar add/sub.
6141 if (NumUndefsLO + 1 == Half || NumUndefsHI + 1 == Half)
6144 // Convert this build_vector into two horizontal add/sub followed by
6146 bool isUndefLO = NumUndefsLO == Half;
6147 bool isUndefHI = NumUndefsHI == Half;
6148 return ExpandHorizontalBinOp(InVec0, InVec1, DL, DAG, X86Opcode, true,
6149 isUndefLO, isUndefHI);
6156 X86TargetLowering::LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG) const {
6159 MVT VT = Op.getSimpleValueType();
6160 MVT ExtVT = VT.getVectorElementType();
6161 unsigned NumElems = Op.getNumOperands();
6163 // Generate vectors for predicate vectors.
6164 if (VT.getScalarType() == MVT::i1 && Subtarget->hasAVX512())
6165 return LowerBUILD_VECTORvXi1(Op, DAG);
6167 // Vectors containing all zeros can be matched by pxor and xorps later
6168 if (ISD::isBuildVectorAllZeros(Op.getNode())) {
6169 // Canonicalize this to <4 x i32> to 1) ensure the zero vectors are CSE'd
6170 // and 2) ensure that i64 scalars are eliminated on x86-32 hosts.
6171 if (VT == MVT::v4i32 || VT == MVT::v8i32 || VT == MVT::v16i32)
6174 return getZeroVector(VT, Subtarget, DAG, dl);
6177 // Vectors containing all ones can be matched by pcmpeqd on 128-bit width
6178 // vectors or broken into v4i32 operations on 256-bit vectors. AVX2 can use
6179 // vpcmpeqd on 256-bit vectors.
6180 if (Subtarget->hasSSE2() && ISD::isBuildVectorAllOnes(Op.getNode())) {
6181 if (VT == MVT::v4i32 || (VT == MVT::v8i32 && Subtarget->hasInt256()))
6184 if (!VT.is512BitVector())
6185 return getOnesVector(VT, Subtarget, DAG, dl);
6188 BuildVectorSDNode *BV = cast<BuildVectorSDNode>(Op.getNode());
6189 if (SDValue AddSub = LowerToAddSub(BV, Subtarget, DAG))
6191 if (SDValue HorizontalOp = LowerToHorizontalOp(BV, Subtarget, DAG))
6192 return HorizontalOp;
6193 if (SDValue Broadcast = LowerVectorBroadcast(Op, Subtarget, DAG))
6196 unsigned EVTBits = ExtVT.getSizeInBits();
6198 unsigned NumZero = 0;
6199 unsigned NumNonZero = 0;
6200 unsigned NonZeros = 0;
6201 bool IsAllConstants = true;
6202 SmallSet<SDValue, 8> Values;
6203 for (unsigned i = 0; i < NumElems; ++i) {
6204 SDValue Elt = Op.getOperand(i);
6205 if (Elt.getOpcode() == ISD::UNDEF)
6208 if (Elt.getOpcode() != ISD::Constant &&
6209 Elt.getOpcode() != ISD::ConstantFP)
6210 IsAllConstants = false;
6211 if (X86::isZeroNode(Elt))
6214 NonZeros |= (1 << i);
6219 // All undef vector. Return an UNDEF. All zero vectors were handled above.
6220 if (NumNonZero == 0)
6221 return DAG.getUNDEF(VT);
6223 // Special case for single non-zero, non-undef, element.
6224 if (NumNonZero == 1) {
6225 unsigned Idx = countTrailingZeros(NonZeros);
6226 SDValue Item = Op.getOperand(Idx);
6228 // If this is an insertion of an i64 value on x86-32, and if the top bits of
6229 // the value are obviously zero, truncate the value to i32 and do the
6230 // insertion that way. Only do this if the value is non-constant or if the
6231 // value is a constant being inserted into element 0. It is cheaper to do
6232 // a constant pool load than it is to do a movd + shuffle.
6233 if (ExtVT == MVT::i64 && !Subtarget->is64Bit() &&
6234 (!IsAllConstants || Idx == 0)) {
6235 if (DAG.MaskedValueIsZero(Item, APInt::getBitsSet(64, 32, 64))) {
6237 assert(VT == MVT::v2i64 && "Expected an SSE value type!");
6238 EVT VecVT = MVT::v4i32;
6240 // Truncate the value (which may itself be a constant) to i32, and
6241 // convert it to a vector with movd (S2V+shuffle to zero extend).
6242 Item = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, Item);
6243 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VecVT, Item);
6244 return DAG.getBitcast(VT, getShuffleVectorZeroOrUndef(
6245 Item, Idx * 2, true, Subtarget, DAG));
6249 // If we have a constant or non-constant insertion into the low element of
6250 // a vector, we can do this with SCALAR_TO_VECTOR + shuffle of zero into
6251 // the rest of the elements. This will be matched as movd/movq/movss/movsd
6252 // depending on what the source datatype is.
6255 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
6257 if (ExtVT == MVT::i32 || ExtVT == MVT::f32 || ExtVT == MVT::f64 ||
6258 (ExtVT == MVT::i64 && Subtarget->is64Bit())) {
6259 if (VT.is512BitVector()) {
6260 SDValue ZeroVec = getZeroVector(VT, Subtarget, DAG, dl);
6261 return DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, ZeroVec,
6262 Item, DAG.getIntPtrConstant(0, dl));
6264 assert((VT.is128BitVector() || VT.is256BitVector()) &&
6265 "Expected an SSE value type!");
6266 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
6267 // Turn it into a MOVL (i.e. movss, movsd, or movd) to a zero vector.
6268 return getShuffleVectorZeroOrUndef(Item, 0, true, Subtarget, DAG);
6271 // We can't directly insert an i8 or i16 into a vector, so zero extend
6273 if (ExtVT == MVT::i16 || ExtVT == MVT::i8) {
6274 Item = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, Item);
6275 if (VT.is256BitVector()) {
6276 if (Subtarget->hasAVX()) {
6277 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v8i32, Item);
6278 Item = getShuffleVectorZeroOrUndef(Item, 0, true, Subtarget, DAG);
6280 // Without AVX, we need to extend to a 128-bit vector and then
6281 // insert into the 256-bit vector.
6282 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32, Item);
6283 SDValue ZeroVec = getZeroVector(MVT::v8i32, Subtarget, DAG, dl);
6284 Item = Insert128BitVector(ZeroVec, Item, 0, DAG, dl);
6287 assert(VT.is128BitVector() && "Expected an SSE value type!");
6288 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32, Item);
6289 Item = getShuffleVectorZeroOrUndef(Item, 0, true, Subtarget, DAG);
6291 return DAG.getBitcast(VT, Item);
6295 // Is it a vector logical left shift?
6296 if (NumElems == 2 && Idx == 1 &&
6297 X86::isZeroNode(Op.getOperand(0)) &&
6298 !X86::isZeroNode(Op.getOperand(1))) {
6299 unsigned NumBits = VT.getSizeInBits();
6300 return getVShift(true, VT,
6301 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
6302 VT, Op.getOperand(1)),
6303 NumBits/2, DAG, *this, dl);
6306 if (IsAllConstants) // Otherwise, it's better to do a constpool load.
6309 // Otherwise, if this is a vector with i32 or f32 elements, and the element
6310 // is a non-constant being inserted into an element other than the low one,
6311 // we can't use a constant pool load. Instead, use SCALAR_TO_VECTOR (aka
6312 // movd/movss) to move this into the low element, then shuffle it into
6314 if (EVTBits == 32) {
6315 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
6316 return getShuffleVectorZeroOrUndef(Item, Idx, NumZero > 0, Subtarget, DAG);
6320 // Splat is obviously ok. Let legalizer expand it to a shuffle.
6321 if (Values.size() == 1) {
6322 if (EVTBits == 32) {
6323 // Instead of a shuffle like this:
6324 // shuffle (scalar_to_vector (load (ptr + 4))), undef, <0, 0, 0, 0>
6325 // Check if it's possible to issue this instead.
6326 // shuffle (vload ptr)), undef, <1, 1, 1, 1>
6327 unsigned Idx = countTrailingZeros(NonZeros);
6328 SDValue Item = Op.getOperand(Idx);
6329 if (Op.getNode()->isOnlyUserOf(Item.getNode()))
6330 return LowerAsSplatVectorLoad(Item, VT, dl, DAG);
6335 // A vector full of immediates; various special cases are already
6336 // handled, so this is best done with a single constant-pool load.
6340 // For AVX-length vectors, see if we can use a vector load to get all of the
6341 // elements, otherwise build the individual 128-bit pieces and use
6342 // shuffles to put them in place.
6343 if (VT.is256BitVector() || VT.is512BitVector()) {
6344 SmallVector<SDValue, 64> V(Op->op_begin(), Op->op_begin() + NumElems);
6346 // Check for a build vector of consecutive loads.
6347 if (SDValue LD = EltsFromConsecutiveLoads(VT, V, dl, DAG, false))
6350 EVT HVT = EVT::getVectorVT(*DAG.getContext(), ExtVT, NumElems/2);
6352 // Build both the lower and upper subvector.
6353 SDValue Lower = DAG.getNode(ISD::BUILD_VECTOR, dl, HVT,
6354 makeArrayRef(&V[0], NumElems/2));
6355 SDValue Upper = DAG.getNode(ISD::BUILD_VECTOR, dl, HVT,
6356 makeArrayRef(&V[NumElems / 2], NumElems/2));
6358 // Recreate the wider vector with the lower and upper part.
6359 if (VT.is256BitVector())
6360 return Concat128BitVectors(Lower, Upper, VT, NumElems, DAG, dl);
6361 return Concat256BitVectors(Lower, Upper, VT, NumElems, DAG, dl);
6364 // Let legalizer expand 2-wide build_vectors.
6365 if (EVTBits == 64) {
6366 if (NumNonZero == 1) {
6367 // One half is zero or undef.
6368 unsigned Idx = countTrailingZeros(NonZeros);
6369 SDValue V2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT,
6370 Op.getOperand(Idx));
6371 return getShuffleVectorZeroOrUndef(V2, Idx, true, Subtarget, DAG);
6376 // If element VT is < 32 bits, convert it to inserts into a zero vector.
6377 if (EVTBits == 8 && NumElems == 16)
6378 if (SDValue V = LowerBuildVectorv16i8(Op, NonZeros,NumNonZero,NumZero, DAG,
6382 if (EVTBits == 16 && NumElems == 8)
6383 if (SDValue V = LowerBuildVectorv8i16(Op, NonZeros,NumNonZero,NumZero, DAG,
6387 // If element VT is == 32 bits and has 4 elems, try to generate an INSERTPS
6388 if (EVTBits == 32 && NumElems == 4)
6389 if (SDValue V = LowerBuildVectorv4x32(Op, DAG, Subtarget, *this))
6392 // If element VT is == 32 bits, turn it into a number of shuffles.
6393 SmallVector<SDValue, 8> V(NumElems);
6394 if (NumElems == 4 && NumZero > 0) {
6395 for (unsigned i = 0; i < 4; ++i) {
6396 bool isZero = !(NonZeros & (1 << i));
6398 V[i] = getZeroVector(VT, Subtarget, DAG, dl);
6400 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(i));
6403 for (unsigned i = 0; i < 2; ++i) {
6404 switch ((NonZeros & (0x3 << i*2)) >> (i*2)) {
6407 V[i] = V[i*2]; // Must be a zero vector.
6410 V[i] = getMOVL(DAG, dl, VT, V[i*2+1], V[i*2]);
6413 V[i] = getMOVL(DAG, dl, VT, V[i*2], V[i*2+1]);
6416 V[i] = getUnpackl(DAG, dl, VT, V[i*2], V[i*2+1]);
6421 bool Reverse1 = (NonZeros & 0x3) == 2;
6422 bool Reverse2 = ((NonZeros & (0x3 << 2)) >> 2) == 2;
6426 static_cast<int>(Reverse2 ? NumElems+1 : NumElems),
6427 static_cast<int>(Reverse2 ? NumElems : NumElems+1)
6429 return DAG.getVectorShuffle(VT, dl, V[0], V[1], &MaskVec[0]);
6432 if (Values.size() > 1 && VT.is128BitVector()) {
6433 // Check for a build vector of consecutive loads.
6434 for (unsigned i = 0; i < NumElems; ++i)
6435 V[i] = Op.getOperand(i);
6437 // Check for elements which are consecutive loads.
6438 if (SDValue LD = EltsFromConsecutiveLoads(VT, V, dl, DAG, false))
6441 // Check for a build vector from mostly shuffle plus few inserting.
6442 if (SDValue Sh = buildFromShuffleMostly(Op, DAG))
6445 // For SSE 4.1, use insertps to put the high elements into the low element.
6446 if (Subtarget->hasSSE41()) {
6448 if (Op.getOperand(0).getOpcode() != ISD::UNDEF)
6449 Result = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(0));
6451 Result = DAG.getUNDEF(VT);
6453 for (unsigned i = 1; i < NumElems; ++i) {
6454 if (Op.getOperand(i).getOpcode() == ISD::UNDEF) continue;
6455 Result = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, Result,
6456 Op.getOperand(i), DAG.getIntPtrConstant(i, dl));
6461 // Otherwise, expand into a number of unpckl*, start by extending each of
6462 // our (non-undef) elements to the full vector width with the element in the
6463 // bottom slot of the vector (which generates no code for SSE).
6464 for (unsigned i = 0; i < NumElems; ++i) {
6465 if (Op.getOperand(i).getOpcode() != ISD::UNDEF)
6466 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(i));
6468 V[i] = DAG.getUNDEF(VT);
6471 // Next, we iteratively mix elements, e.g. for v4f32:
6472 // Step 1: unpcklps 0, 2 ==> X: <?, ?, 2, 0>
6473 // : unpcklps 1, 3 ==> Y: <?, ?, 3, 1>
6474 // Step 2: unpcklps X, Y ==> <3, 2, 1, 0>
6475 unsigned EltStride = NumElems >> 1;
6476 while (EltStride != 0) {
6477 for (unsigned i = 0; i < EltStride; ++i) {
6478 // If V[i+EltStride] is undef and this is the first round of mixing,
6479 // then it is safe to just drop this shuffle: V[i] is already in the
6480 // right place, the one element (since it's the first round) being
6481 // inserted as undef can be dropped. This isn't safe for successive
6482 // rounds because they will permute elements within both vectors.
6483 if (V[i+EltStride].getOpcode() == ISD::UNDEF &&
6484 EltStride == NumElems/2)
6487 V[i] = getUnpackl(DAG, dl, VT, V[i], V[i + EltStride]);
6496 // 256-bit AVX can use the vinsertf128 instruction
6497 // to create 256-bit vectors from two other 128-bit ones.
6498 static SDValue LowerAVXCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) {
6500 MVT ResVT = Op.getSimpleValueType();
6502 assert((ResVT.is256BitVector() ||
6503 ResVT.is512BitVector()) && "Value type must be 256-/512-bit wide");
6505 SDValue V1 = Op.getOperand(0);
6506 SDValue V2 = Op.getOperand(1);
6507 unsigned NumElems = ResVT.getVectorNumElements();
6508 if (ResVT.is256BitVector())
6509 return Concat128BitVectors(V1, V2, ResVT, NumElems, DAG, dl);
6511 if (Op.getNumOperands() == 4) {
6512 MVT HalfVT = MVT::getVectorVT(ResVT.getScalarType(),
6513 ResVT.getVectorNumElements()/2);
6514 SDValue V3 = Op.getOperand(2);
6515 SDValue V4 = Op.getOperand(3);
6516 return Concat256BitVectors(Concat128BitVectors(V1, V2, HalfVT, NumElems/2, DAG, dl),
6517 Concat128BitVectors(V3, V4, HalfVT, NumElems/2, DAG, dl), ResVT, NumElems, DAG, dl);
6519 return Concat256BitVectors(V1, V2, ResVT, NumElems, DAG, dl);
6522 static SDValue LowerCONCAT_VECTORSvXi1(SDValue Op,
6523 const X86Subtarget *Subtarget,
6524 SelectionDAG & DAG) {
6526 MVT ResVT = Op.getSimpleValueType();
6527 unsigned NumOfOperands = Op.getNumOperands();
6529 assert(isPowerOf2_32(NumOfOperands) &&
6530 "Unexpected number of operands in CONCAT_VECTORS");
6532 if (NumOfOperands > 2) {
6533 MVT HalfVT = MVT::getVectorVT(ResVT.getScalarType(),
6534 ResVT.getVectorNumElements()/2);
6535 SmallVector<SDValue, 2> Ops;
6536 for (unsigned i = 0; i < NumOfOperands/2; i++)
6537 Ops.push_back(Op.getOperand(i));
6538 SDValue Lo = DAG.getNode(ISD::CONCAT_VECTORS, dl, HalfVT, Ops);
6540 for (unsigned i = NumOfOperands/2; i < NumOfOperands; i++)
6541 Ops.push_back(Op.getOperand(i));
6542 SDValue Hi = DAG.getNode(ISD::CONCAT_VECTORS, dl, HalfVT, Ops);
6543 return DAG.getNode(ISD::CONCAT_VECTORS, dl, ResVT, Lo, Hi);
6546 SDValue V1 = Op.getOperand(0);
6547 SDValue V2 = Op.getOperand(1);
6548 bool IsZeroV1 = ISD::isBuildVectorAllZeros(V1.getNode());
6549 bool IsZeroV2 = ISD::isBuildVectorAllZeros(V2.getNode());
6551 if (IsZeroV1 && IsZeroV2)
6552 return getZeroVector(ResVT, Subtarget, DAG, dl);
6554 SDValue ZeroIdx = DAG.getIntPtrConstant(0, dl);
6555 SDValue Undef = DAG.getUNDEF(ResVT);
6556 unsigned NumElems = ResVT.getVectorNumElements();
6557 SDValue ShiftBits = DAG.getConstant(NumElems/2, dl, MVT::i8);
6559 V2 = DAG.getNode(ISD::INSERT_SUBVECTOR, dl, ResVT, Undef, V2, ZeroIdx);
6560 V2 = DAG.getNode(X86ISD::VSHLI, dl, ResVT, V2, ShiftBits);
6564 V1 = DAG.getNode(ISD::INSERT_SUBVECTOR, dl, ResVT, Undef, V1, ZeroIdx);
6565 // Zero the upper bits of V1
6566 V1 = DAG.getNode(X86ISD::VSHLI, dl, ResVT, V1, ShiftBits);
6567 V1 = DAG.getNode(X86ISD::VSRLI, dl, ResVT, V1, ShiftBits);
6570 return DAG.getNode(ISD::OR, dl, ResVT, V1, V2);
6573 static SDValue LowerCONCAT_VECTORS(SDValue Op,
6574 const X86Subtarget *Subtarget,
6575 SelectionDAG &DAG) {
6576 MVT VT = Op.getSimpleValueType();
6577 if (VT.getVectorElementType() == MVT::i1)
6578 return LowerCONCAT_VECTORSvXi1(Op, Subtarget, DAG);
6580 assert((VT.is256BitVector() && Op.getNumOperands() == 2) ||
6581 (VT.is512BitVector() && (Op.getNumOperands() == 2 ||
6582 Op.getNumOperands() == 4)));
6584 // AVX can use the vinsertf128 instruction to create 256-bit vectors
6585 // from two other 128-bit ones.
6587 // 512-bit vector may contain 2 256-bit vectors or 4 128-bit vectors
6588 return LowerAVXCONCAT_VECTORS(Op, DAG);
6591 //===----------------------------------------------------------------------===//
6592 // Vector shuffle lowering
6594 // This is an experimental code path for lowering vector shuffles on x86. It is
6595 // designed to handle arbitrary vector shuffles and blends, gracefully
6596 // degrading performance as necessary. It works hard to recognize idiomatic
6597 // shuffles and lower them to optimal instruction patterns without leaving
6598 // a framework that allows reasonably efficient handling of all vector shuffle
6600 //===----------------------------------------------------------------------===//
6602 /// \brief Tiny helper function to identify a no-op mask.
6604 /// This is a somewhat boring predicate function. It checks whether the mask
6605 /// array input, which is assumed to be a single-input shuffle mask of the kind
6606 /// used by the X86 shuffle instructions (not a fully general
6607 /// ShuffleVectorSDNode mask) requires any shuffles to occur. Both undef and an
6608 /// in-place shuffle are 'no-op's.
6609 static bool isNoopShuffleMask(ArrayRef<int> Mask) {
6610 for (int i = 0, Size = Mask.size(); i < Size; ++i)
6611 if (Mask[i] != -1 && Mask[i] != i)
6616 /// \brief Helper function to classify a mask as a single-input mask.
6618 /// This isn't a generic single-input test because in the vector shuffle
6619 /// lowering we canonicalize single inputs to be the first input operand. This
6620 /// means we can more quickly test for a single input by only checking whether
6621 /// an input from the second operand exists. We also assume that the size of
6622 /// mask corresponds to the size of the input vectors which isn't true in the
6623 /// fully general case.
6624 static bool isSingleInputShuffleMask(ArrayRef<int> Mask) {
6626 if (M >= (int)Mask.size())
6631 /// \brief Test whether there are elements crossing 128-bit lanes in this
6634 /// X86 divides up its shuffles into in-lane and cross-lane shuffle operations
6635 /// and we routinely test for these.
6636 static bool is128BitLaneCrossingShuffleMask(MVT VT, ArrayRef<int> Mask) {
6637 int LaneSize = 128 / VT.getScalarSizeInBits();
6638 int Size = Mask.size();
6639 for (int i = 0; i < Size; ++i)
6640 if (Mask[i] >= 0 && (Mask[i] % Size) / LaneSize != i / LaneSize)
6645 /// \brief Test whether a shuffle mask is equivalent within each 128-bit lane.
6647 /// This checks a shuffle mask to see if it is performing the same
6648 /// 128-bit lane-relative shuffle in each 128-bit lane. This trivially implies
6649 /// that it is also not lane-crossing. It may however involve a blend from the
6650 /// same lane of a second vector.
6652 /// The specific repeated shuffle mask is populated in \p RepeatedMask, as it is
6653 /// non-trivial to compute in the face of undef lanes. The representation is
6654 /// *not* suitable for use with existing 128-bit shuffles as it will contain
6655 /// entries from both V1 and V2 inputs to the wider mask.
6657 is128BitLaneRepeatedShuffleMask(MVT VT, ArrayRef<int> Mask,
6658 SmallVectorImpl<int> &RepeatedMask) {
6659 int LaneSize = 128 / VT.getScalarSizeInBits();
6660 RepeatedMask.resize(LaneSize, -1);
6661 int Size = Mask.size();
6662 for (int i = 0; i < Size; ++i) {
6665 if ((Mask[i] % Size) / LaneSize != i / LaneSize)
6666 // This entry crosses lanes, so there is no way to model this shuffle.
6669 // Ok, handle the in-lane shuffles by detecting if and when they repeat.
6670 if (RepeatedMask[i % LaneSize] == -1)
6671 // This is the first non-undef entry in this slot of a 128-bit lane.
6672 RepeatedMask[i % LaneSize] =
6673 Mask[i] < Size ? Mask[i] % LaneSize : Mask[i] % LaneSize + Size;
6674 else if (RepeatedMask[i % LaneSize] + (i / LaneSize) * LaneSize != Mask[i])
6675 // Found a mismatch with the repeated mask.
6681 /// \brief Checks whether a shuffle mask is equivalent to an explicit list of
6684 /// This is a fast way to test a shuffle mask against a fixed pattern:
6686 /// if (isShuffleEquivalent(Mask, 3, 2, {1, 0})) { ... }
6688 /// It returns true if the mask is exactly as wide as the argument list, and
6689 /// each element of the mask is either -1 (signifying undef) or the value given
6690 /// in the argument.
6691 static bool isShuffleEquivalent(SDValue V1, SDValue V2, ArrayRef<int> Mask,
6692 ArrayRef<int> ExpectedMask) {
6693 if (Mask.size() != ExpectedMask.size())
6696 int Size = Mask.size();
6698 // If the values are build vectors, we can look through them to find
6699 // equivalent inputs that make the shuffles equivalent.
6700 auto *BV1 = dyn_cast<BuildVectorSDNode>(V1);
6701 auto *BV2 = dyn_cast<BuildVectorSDNode>(V2);
6703 for (int i = 0; i < Size; ++i)
6704 if (Mask[i] != -1 && Mask[i] != ExpectedMask[i]) {
6705 auto *MaskBV = Mask[i] < Size ? BV1 : BV2;
6706 auto *ExpectedBV = ExpectedMask[i] < Size ? BV1 : BV2;
6707 if (!MaskBV || !ExpectedBV ||
6708 MaskBV->getOperand(Mask[i] % Size) !=
6709 ExpectedBV->getOperand(ExpectedMask[i] % Size))
6716 /// \brief Get a 4-lane 8-bit shuffle immediate for a mask.
6718 /// This helper function produces an 8-bit shuffle immediate corresponding to
6719 /// the ubiquitous shuffle encoding scheme used in x86 instructions for
6720 /// shuffling 4 lanes. It can be used with most of the PSHUF instructions for
6723 /// NB: We rely heavily on "undef" masks preserving the input lane.
6724 static SDValue getV4X86ShuffleImm8ForMask(ArrayRef<int> Mask, SDLoc DL,
6725 SelectionDAG &DAG) {
6726 assert(Mask.size() == 4 && "Only 4-lane shuffle masks");
6727 assert(Mask[0] >= -1 && Mask[0] < 4 && "Out of bound mask element!");
6728 assert(Mask[1] >= -1 && Mask[1] < 4 && "Out of bound mask element!");
6729 assert(Mask[2] >= -1 && Mask[2] < 4 && "Out of bound mask element!");
6730 assert(Mask[3] >= -1 && Mask[3] < 4 && "Out of bound mask element!");
6733 Imm |= (Mask[0] == -1 ? 0 : Mask[0]) << 0;
6734 Imm |= (Mask[1] == -1 ? 1 : Mask[1]) << 2;
6735 Imm |= (Mask[2] == -1 ? 2 : Mask[2]) << 4;
6736 Imm |= (Mask[3] == -1 ? 3 : Mask[3]) << 6;
6737 return DAG.getConstant(Imm, DL, MVT::i8);
6740 /// \brief Compute whether each element of a shuffle is zeroable.
6742 /// A "zeroable" vector shuffle element is one which can be lowered to zero.
6743 /// Either it is an undef element in the shuffle mask, the element of the input
6744 /// referenced is undef, or the element of the input referenced is known to be
6745 /// zero. Many x86 shuffles can zero lanes cheaply and we often want to handle
6746 /// as many lanes with this technique as possible to simplify the remaining
6748 static SmallBitVector computeZeroableShuffleElements(ArrayRef<int> Mask,
6749 SDValue V1, SDValue V2) {
6750 SmallBitVector Zeroable(Mask.size(), false);
6752 while (V1.getOpcode() == ISD::BITCAST)
6753 V1 = V1->getOperand(0);
6754 while (V2.getOpcode() == ISD::BITCAST)
6755 V2 = V2->getOperand(0);
6757 bool V1IsZero = ISD::isBuildVectorAllZeros(V1.getNode());
6758 bool V2IsZero = ISD::isBuildVectorAllZeros(V2.getNode());
6760 for (int i = 0, Size = Mask.size(); i < Size; ++i) {
6762 // Handle the easy cases.
6763 if (M < 0 || (M >= 0 && M < Size && V1IsZero) || (M >= Size && V2IsZero)) {
6768 // If this is an index into a build_vector node (which has the same number
6769 // of elements), dig out the input value and use it.
6770 SDValue V = M < Size ? V1 : V2;
6771 if (V.getOpcode() != ISD::BUILD_VECTOR || Size != (int)V.getNumOperands())
6774 SDValue Input = V.getOperand(M % Size);
6775 // The UNDEF opcode check really should be dead code here, but not quite
6776 // worth asserting on (it isn't invalid, just unexpected).
6777 if (Input.getOpcode() == ISD::UNDEF || X86::isZeroNode(Input))
6784 // X86 has dedicated unpack instructions that can handle specific blend
6785 // operations: UNPCKH and UNPCKL.
6786 static SDValue lowerVectorShuffleWithUNPCK(SDLoc DL, MVT VT, ArrayRef<int> Mask,
6787 SDValue V1, SDValue V2,
6788 SelectionDAG &DAG) {
6789 int NumElts = VT.getVectorNumElements();
6790 int NumEltsInLane = 128 / VT.getScalarSizeInBits();
6791 SmallVector<int, 8> Unpckl;
6792 SmallVector<int, 8> Unpckh;
6794 for (int i = 0; i < NumElts; ++i) {
6795 unsigned LaneStart = (i / NumEltsInLane) * NumEltsInLane;
6796 int LoPos = (i % NumEltsInLane) / 2 + LaneStart + NumElts * (i % 2);
6797 int HiPos = LoPos + NumEltsInLane / 2;
6798 Unpckl.push_back(LoPos);
6799 Unpckh.push_back(HiPos);
6802 if (isShuffleEquivalent(V1, V2, Mask, Unpckl))
6803 return DAG.getNode(X86ISD::UNPCKL, DL, VT, V1, V2);
6804 if (isShuffleEquivalent(V1, V2, Mask, Unpckh))
6805 return DAG.getNode(X86ISD::UNPCKH, DL, VT, V1, V2);
6807 // Commute and try again.
6808 ShuffleVectorSDNode::commuteMask(Unpckl);
6809 if (isShuffleEquivalent(V1, V2, Mask, Unpckl))
6810 return DAG.getNode(X86ISD::UNPCKL, DL, VT, V2, V1);
6812 ShuffleVectorSDNode::commuteMask(Unpckh);
6813 if (isShuffleEquivalent(V1, V2, Mask, Unpckh))
6814 return DAG.getNode(X86ISD::UNPCKH, DL, VT, V2, V1);
6819 /// \brief Try to emit a bitmask instruction for a shuffle.
6821 /// This handles cases where we can model a blend exactly as a bitmask due to
6822 /// one of the inputs being zeroable.
6823 static SDValue lowerVectorShuffleAsBitMask(SDLoc DL, MVT VT, SDValue V1,
6824 SDValue V2, ArrayRef<int> Mask,
6825 SelectionDAG &DAG) {
6826 MVT EltVT = VT.getScalarType();
6827 int NumEltBits = EltVT.getSizeInBits();
6828 MVT IntEltVT = MVT::getIntegerVT(NumEltBits);
6829 SDValue Zero = DAG.getConstant(0, DL, IntEltVT);
6830 SDValue AllOnes = DAG.getConstant(APInt::getAllOnesValue(NumEltBits), DL,
6832 if (EltVT.isFloatingPoint()) {
6833 Zero = DAG.getBitcast(EltVT, Zero);
6834 AllOnes = DAG.getBitcast(EltVT, AllOnes);
6836 SmallVector<SDValue, 16> VMaskOps(Mask.size(), Zero);
6837 SmallBitVector Zeroable = computeZeroableShuffleElements(Mask, V1, V2);
6839 for (int i = 0, Size = Mask.size(); i < Size; ++i) {
6842 if (Mask[i] % Size != i)
6843 return SDValue(); // Not a blend.
6845 V = Mask[i] < Size ? V1 : V2;
6846 else if (V != (Mask[i] < Size ? V1 : V2))
6847 return SDValue(); // Can only let one input through the mask.
6849 VMaskOps[i] = AllOnes;
6852 return SDValue(); // No non-zeroable elements!
6854 SDValue VMask = DAG.getNode(ISD::BUILD_VECTOR, DL, VT, VMaskOps);
6855 V = DAG.getNode(VT.isFloatingPoint()
6856 ? (unsigned) X86ISD::FAND : (unsigned) ISD::AND,
6861 /// \brief Try to emit a blend instruction for a shuffle using bit math.
6863 /// This is used as a fallback approach when first class blend instructions are
6864 /// unavailable. Currently it is only suitable for integer vectors, but could
6865 /// be generalized for floating point vectors if desirable.
6866 static SDValue lowerVectorShuffleAsBitBlend(SDLoc DL, MVT VT, SDValue V1,
6867 SDValue V2, ArrayRef<int> Mask,
6868 SelectionDAG &DAG) {
6869 assert(VT.isInteger() && "Only supports integer vector types!");
6870 MVT EltVT = VT.getScalarType();
6871 int NumEltBits = EltVT.getSizeInBits();
6872 SDValue Zero = DAG.getConstant(0, DL, EltVT);
6873 SDValue AllOnes = DAG.getConstant(APInt::getAllOnesValue(NumEltBits), DL,
6875 SmallVector<SDValue, 16> MaskOps;
6876 for (int i = 0, Size = Mask.size(); i < Size; ++i) {
6877 if (Mask[i] != -1 && Mask[i] != i && Mask[i] != i + Size)
6878 return SDValue(); // Shuffled input!
6879 MaskOps.push_back(Mask[i] < Size ? AllOnes : Zero);
6882 SDValue V1Mask = DAG.getNode(ISD::BUILD_VECTOR, DL, VT, MaskOps);
6883 V1 = DAG.getNode(ISD::AND, DL, VT, V1, V1Mask);
6884 // We have to cast V2 around.
6885 MVT MaskVT = MVT::getVectorVT(MVT::i64, VT.getSizeInBits() / 64);
6886 V2 = DAG.getBitcast(VT, DAG.getNode(X86ISD::ANDNP, DL, MaskVT,
6887 DAG.getBitcast(MaskVT, V1Mask),
6888 DAG.getBitcast(MaskVT, V2)));
6889 return DAG.getNode(ISD::OR, DL, VT, V1, V2);
6892 /// \brief Try to emit a blend instruction for a shuffle.
6894 /// This doesn't do any checks for the availability of instructions for blending
6895 /// these values. It relies on the availability of the X86ISD::BLENDI pattern to
6896 /// be matched in the backend with the type given. What it does check for is
6897 /// that the shuffle mask is in fact a blend.
6898 static SDValue lowerVectorShuffleAsBlend(SDLoc DL, MVT VT, SDValue V1,
6899 SDValue V2, ArrayRef<int> Mask,
6900 const X86Subtarget *Subtarget,
6901 SelectionDAG &DAG) {
6902 unsigned BlendMask = 0;
6903 for (int i = 0, Size = Mask.size(); i < Size; ++i) {
6904 if (Mask[i] >= Size) {
6905 if (Mask[i] != i + Size)
6906 return SDValue(); // Shuffled V2 input!
6907 BlendMask |= 1u << i;
6910 if (Mask[i] >= 0 && Mask[i] != i)
6911 return SDValue(); // Shuffled V1 input!
6913 switch (VT.SimpleTy) {
6918 return DAG.getNode(X86ISD::BLENDI, DL, VT, V1, V2,
6919 DAG.getConstant(BlendMask, DL, MVT::i8));
6923 assert(Subtarget->hasAVX2() && "256-bit integer blends require AVX2!");
6927 // If we have AVX2 it is faster to use VPBLENDD when the shuffle fits into
6928 // that instruction.
6929 if (Subtarget->hasAVX2()) {
6930 // Scale the blend by the number of 32-bit dwords per element.
6931 int Scale = VT.getScalarSizeInBits() / 32;
6933 for (int i = 0, Size = Mask.size(); i < Size; ++i)
6934 if (Mask[i] >= Size)
6935 for (int j = 0; j < Scale; ++j)
6936 BlendMask |= 1u << (i * Scale + j);
6938 MVT BlendVT = VT.getSizeInBits() > 128 ? MVT::v8i32 : MVT::v4i32;
6939 V1 = DAG.getBitcast(BlendVT, V1);
6940 V2 = DAG.getBitcast(BlendVT, V2);
6941 return DAG.getBitcast(
6942 VT, DAG.getNode(X86ISD::BLENDI, DL, BlendVT, V1, V2,
6943 DAG.getConstant(BlendMask, DL, MVT::i8)));
6947 // For integer shuffles we need to expand the mask and cast the inputs to
6948 // v8i16s prior to blending.
6949 int Scale = 8 / VT.getVectorNumElements();
6951 for (int i = 0, Size = Mask.size(); i < Size; ++i)
6952 if (Mask[i] >= Size)
6953 for (int j = 0; j < Scale; ++j)
6954 BlendMask |= 1u << (i * Scale + j);
6956 V1 = DAG.getBitcast(MVT::v8i16, V1);
6957 V2 = DAG.getBitcast(MVT::v8i16, V2);
6958 return DAG.getBitcast(VT,
6959 DAG.getNode(X86ISD::BLENDI, DL, MVT::v8i16, V1, V2,
6960 DAG.getConstant(BlendMask, DL, MVT::i8)));
6964 assert(Subtarget->hasAVX2() && "256-bit integer blends require AVX2!");
6965 SmallVector<int, 8> RepeatedMask;
6966 if (is128BitLaneRepeatedShuffleMask(MVT::v16i16, Mask, RepeatedMask)) {
6967 // We can lower these with PBLENDW which is mirrored across 128-bit lanes.
6968 assert(RepeatedMask.size() == 8 && "Repeated mask size doesn't match!");
6970 for (int i = 0; i < 8; ++i)
6971 if (RepeatedMask[i] >= 16)
6972 BlendMask |= 1u << i;
6973 return DAG.getNode(X86ISD::BLENDI, DL, MVT::v16i16, V1, V2,
6974 DAG.getConstant(BlendMask, DL, MVT::i8));
6980 assert((VT.getSizeInBits() == 128 || Subtarget->hasAVX2()) &&
6981 "256-bit byte-blends require AVX2 support!");
6983 // Attempt to lower to a bitmask if we can. VPAND is faster than VPBLENDVB.
6984 if (SDValue Masked = lowerVectorShuffleAsBitMask(DL, VT, V1, V2, Mask, DAG))
6987 // Scale the blend by the number of bytes per element.
6988 int Scale = VT.getScalarSizeInBits() / 8;
6990 // This form of blend is always done on bytes. Compute the byte vector
6992 MVT BlendVT = MVT::getVectorVT(MVT::i8, VT.getSizeInBits() / 8);
6994 // Compute the VSELECT mask. Note that VSELECT is really confusing in the
6995 // mix of LLVM's code generator and the x86 backend. We tell the code
6996 // generator that boolean values in the elements of an x86 vector register
6997 // are -1 for true and 0 for false. We then use the LLVM semantics of 'true'
6998 // mapping a select to operand #1, and 'false' mapping to operand #2. The
6999 // reality in x86 is that vector masks (pre-AVX-512) use only the high bit
7000 // of the element (the remaining are ignored) and 0 in that high bit would
7001 // mean operand #1 while 1 in the high bit would mean operand #2. So while
7002 // the LLVM model for boolean values in vector elements gets the relevant
7003 // bit set, it is set backwards and over constrained relative to x86's
7005 SmallVector<SDValue, 32> VSELECTMask;
7006 for (int i = 0, Size = Mask.size(); i < Size; ++i)
7007 for (int j = 0; j < Scale; ++j)
7008 VSELECTMask.push_back(
7009 Mask[i] < 0 ? DAG.getUNDEF(MVT::i8)
7010 : DAG.getConstant(Mask[i] < Size ? -1 : 0, DL,
7013 V1 = DAG.getBitcast(BlendVT, V1);
7014 V2 = DAG.getBitcast(BlendVT, V2);
7015 return DAG.getBitcast(VT, DAG.getNode(ISD::VSELECT, DL, BlendVT,
7016 DAG.getNode(ISD::BUILD_VECTOR, DL,
7017 BlendVT, VSELECTMask),
7022 llvm_unreachable("Not a supported integer vector type!");
7026 /// \brief Try to lower as a blend of elements from two inputs followed by
7027 /// a single-input permutation.
7029 /// This matches the pattern where we can blend elements from two inputs and
7030 /// then reduce the shuffle to a single-input permutation.
7031 static SDValue lowerVectorShuffleAsBlendAndPermute(SDLoc DL, MVT VT, SDValue V1,
7034 SelectionDAG &DAG) {
7035 // We build up the blend mask while checking whether a blend is a viable way
7036 // to reduce the shuffle.
7037 SmallVector<int, 32> BlendMask(Mask.size(), -1);
7038 SmallVector<int, 32> PermuteMask(Mask.size(), -1);
7040 for (int i = 0, Size = Mask.size(); i < Size; ++i) {
7044 assert(Mask[i] < Size * 2 && "Shuffle input is out of bounds.");
7046 if (BlendMask[Mask[i] % Size] == -1)
7047 BlendMask[Mask[i] % Size] = Mask[i];
7048 else if (BlendMask[Mask[i] % Size] != Mask[i])
7049 return SDValue(); // Can't blend in the needed input!
7051 PermuteMask[i] = Mask[i] % Size;
7054 SDValue V = DAG.getVectorShuffle(VT, DL, V1, V2, BlendMask);
7055 return DAG.getVectorShuffle(VT, DL, V, DAG.getUNDEF(VT), PermuteMask);
7058 /// \brief Generic routine to decompose a shuffle and blend into indepndent
7059 /// blends and permutes.
7061 /// This matches the extremely common pattern for handling combined
7062 /// shuffle+blend operations on newer X86 ISAs where we have very fast blend
7063 /// operations. It will try to pick the best arrangement of shuffles and
7065 static SDValue lowerVectorShuffleAsDecomposedShuffleBlend(SDLoc DL, MVT VT,
7069 SelectionDAG &DAG) {
7070 // Shuffle the input elements into the desired positions in V1 and V2 and
7071 // blend them together.
7072 SmallVector<int, 32> V1Mask(Mask.size(), -1);
7073 SmallVector<int, 32> V2Mask(Mask.size(), -1);
7074 SmallVector<int, 32> BlendMask(Mask.size(), -1);
7075 for (int i = 0, Size = Mask.size(); i < Size; ++i)
7076 if (Mask[i] >= 0 && Mask[i] < Size) {
7077 V1Mask[i] = Mask[i];
7079 } else if (Mask[i] >= Size) {
7080 V2Mask[i] = Mask[i] - Size;
7081 BlendMask[i] = i + Size;
7084 // Try to lower with the simpler initial blend strategy unless one of the
7085 // input shuffles would be a no-op. We prefer to shuffle inputs as the
7086 // shuffle may be able to fold with a load or other benefit. However, when
7087 // we'll have to do 2x as many shuffles in order to achieve this, blending
7088 // first is a better strategy.
7089 if (!isNoopShuffleMask(V1Mask) && !isNoopShuffleMask(V2Mask))
7090 if (SDValue BlendPerm =
7091 lowerVectorShuffleAsBlendAndPermute(DL, VT, V1, V2, Mask, DAG))
7094 V1 = DAG.getVectorShuffle(VT, DL, V1, DAG.getUNDEF(VT), V1Mask);
7095 V2 = DAG.getVectorShuffle(VT, DL, V2, DAG.getUNDEF(VT), V2Mask);
7096 return DAG.getVectorShuffle(VT, DL, V1, V2, BlendMask);
7099 /// \brief Try to lower a vector shuffle as a byte rotation.
7101 /// SSSE3 has a generic PALIGNR instruction in x86 that will do an arbitrary
7102 /// byte-rotation of the concatenation of two vectors; pre-SSSE3 can use
7103 /// a PSRLDQ/PSLLDQ/POR pattern to get a similar effect. This routine will
7104 /// try to generically lower a vector shuffle through such an pattern. It
7105 /// does not check for the profitability of lowering either as PALIGNR or
7106 /// PSRLDQ/PSLLDQ/POR, only whether the mask is valid to lower in that form.
7107 /// This matches shuffle vectors that look like:
7109 /// v8i16 [11, 12, 13, 14, 15, 0, 1, 2]
7111 /// Essentially it concatenates V1 and V2, shifts right by some number of
7112 /// elements, and takes the low elements as the result. Note that while this is
7113 /// specified as a *right shift* because x86 is little-endian, it is a *left
7114 /// rotate* of the vector lanes.
7115 static SDValue lowerVectorShuffleAsByteRotate(SDLoc DL, MVT VT, SDValue V1,
7118 const X86Subtarget *Subtarget,
7119 SelectionDAG &DAG) {
7120 assert(!isNoopShuffleMask(Mask) && "We shouldn't lower no-op shuffles!");
7122 int NumElts = Mask.size();
7123 int NumLanes = VT.getSizeInBits() / 128;
7124 int NumLaneElts = NumElts / NumLanes;
7126 // We need to detect various ways of spelling a rotation:
7127 // [11, 12, 13, 14, 15, 0, 1, 2]
7128 // [-1, 12, 13, 14, -1, -1, 1, -1]
7129 // [-1, -1, -1, -1, -1, -1, 1, 2]
7130 // [ 3, 4, 5, 6, 7, 8, 9, 10]
7131 // [-1, 4, 5, 6, -1, -1, 9, -1]
7132 // [-1, 4, 5, 6, -1, -1, -1, -1]
7135 for (int l = 0; l < NumElts; l += NumLaneElts) {
7136 for (int i = 0; i < NumLaneElts; ++i) {
7137 if (Mask[l + i] == -1)
7139 assert(Mask[l + i] >= 0 && "Only -1 is a valid negative mask element!");
7141 // Get the mod-Size index and lane correct it.
7142 int LaneIdx = (Mask[l + i] % NumElts) - l;
7143 // Make sure it was in this lane.
7144 if (LaneIdx < 0 || LaneIdx >= NumLaneElts)
7147 // Determine where a rotated vector would have started.
7148 int StartIdx = i - LaneIdx;
7150 // The identity rotation isn't interesting, stop.
7153 // If we found the tail of a vector the rotation must be the missing
7154 // front. If we found the head of a vector, it must be how much of the
7156 int CandidateRotation = StartIdx < 0 ? -StartIdx : NumLaneElts - StartIdx;
7159 Rotation = CandidateRotation;
7160 else if (Rotation != CandidateRotation)
7161 // The rotations don't match, so we can't match this mask.
7164 // Compute which value this mask is pointing at.
7165 SDValue MaskV = Mask[l + i] < NumElts ? V1 : V2;
7167 // Compute which of the two target values this index should be assigned
7168 // to. This reflects whether the high elements are remaining or the low
7169 // elements are remaining.
7170 SDValue &TargetV = StartIdx < 0 ? Hi : Lo;
7172 // Either set up this value if we've not encountered it before, or check
7173 // that it remains consistent.
7176 else if (TargetV != MaskV)
7177 // This may be a rotation, but it pulls from the inputs in some
7178 // unsupported interleaving.
7183 // Check that we successfully analyzed the mask, and normalize the results.
7184 assert(Rotation != 0 && "Failed to locate a viable rotation!");
7185 assert((Lo || Hi) && "Failed to find a rotated input vector!");
7191 // The actual rotate instruction rotates bytes, so we need to scale the
7192 // rotation based on how many bytes are in the vector lane.
7193 int Scale = 16 / NumLaneElts;
7195 // SSSE3 targets can use the palignr instruction.
7196 if (Subtarget->hasSSSE3()) {
7197 // Cast the inputs to i8 vector of correct length to match PALIGNR.
7198 MVT AlignVT = MVT::getVectorVT(MVT::i8, 16 * NumLanes);
7199 Lo = DAG.getBitcast(AlignVT, Lo);
7200 Hi = DAG.getBitcast(AlignVT, Hi);
7202 return DAG.getBitcast(
7203 VT, DAG.getNode(X86ISD::PALIGNR, DL, AlignVT, Lo, Hi,
7204 DAG.getConstant(Rotation * Scale, DL, MVT::i8)));
7207 assert(VT.getSizeInBits() == 128 &&
7208 "Rotate-based lowering only supports 128-bit lowering!");
7209 assert(Mask.size() <= 16 &&
7210 "Can shuffle at most 16 bytes in a 128-bit vector!");
7212 // Default SSE2 implementation
7213 int LoByteShift = 16 - Rotation * Scale;
7214 int HiByteShift = Rotation * Scale;
7216 // Cast the inputs to v2i64 to match PSLLDQ/PSRLDQ.
7217 Lo = DAG.getBitcast(MVT::v2i64, Lo);
7218 Hi = DAG.getBitcast(MVT::v2i64, Hi);
7220 SDValue LoShift = DAG.getNode(X86ISD::VSHLDQ, DL, MVT::v2i64, Lo,
7221 DAG.getConstant(LoByteShift, DL, MVT::i8));
7222 SDValue HiShift = DAG.getNode(X86ISD::VSRLDQ, DL, MVT::v2i64, Hi,
7223 DAG.getConstant(HiByteShift, DL, MVT::i8));
7224 return DAG.getBitcast(VT,
7225 DAG.getNode(ISD::OR, DL, MVT::v2i64, LoShift, HiShift));
7228 /// \brief Try to lower a vector shuffle as a bit shift (shifts in zeros).
7230 /// Attempts to match a shuffle mask against the PSLL(W/D/Q/DQ) and
7231 /// PSRL(W/D/Q/DQ) SSE2 and AVX2 logical bit-shift instructions. The function
7232 /// matches elements from one of the input vectors shuffled to the left or
7233 /// right with zeroable elements 'shifted in'. It handles both the strictly
7234 /// bit-wise element shifts and the byte shift across an entire 128-bit double
7237 /// PSHL : (little-endian) left bit shift.
7238 /// [ zz, 0, zz, 2 ]
7239 /// [ -1, 4, zz, -1 ]
7240 /// PSRL : (little-endian) right bit shift.
7242 /// [ -1, -1, 7, zz]
7243 /// PSLLDQ : (little-endian) left byte shift
7244 /// [ zz, 0, 1, 2, 3, 4, 5, 6]
7245 /// [ zz, zz, -1, -1, 2, 3, 4, -1]
7246 /// [ zz, zz, zz, zz, zz, zz, -1, 1]
7247 /// PSRLDQ : (little-endian) right byte shift
7248 /// [ 5, 6, 7, zz, zz, zz, zz, zz]
7249 /// [ -1, 5, 6, 7, zz, zz, zz, zz]
7250 /// [ 1, 2, -1, -1, -1, -1, zz, zz]
7251 static SDValue lowerVectorShuffleAsShift(SDLoc DL, MVT VT, SDValue V1,
7252 SDValue V2, ArrayRef<int> Mask,
7253 SelectionDAG &DAG) {
7254 SmallBitVector Zeroable = computeZeroableShuffleElements(Mask, V1, V2);
7256 int Size = Mask.size();
7257 assert(Size == (int)VT.getVectorNumElements() && "Unexpected mask size");
7259 auto CheckZeros = [&](int Shift, int Scale, bool Left) {
7260 for (int i = 0; i < Size; i += Scale)
7261 for (int j = 0; j < Shift; ++j)
7262 if (!Zeroable[i + j + (Left ? 0 : (Scale - Shift))])
7268 auto MatchShift = [&](int Shift, int Scale, bool Left, SDValue V) {
7269 for (int i = 0; i != Size; i += Scale) {
7270 unsigned Pos = Left ? i + Shift : i;
7271 unsigned Low = Left ? i : i + Shift;
7272 unsigned Len = Scale - Shift;
7273 if (!isSequentialOrUndefInRange(Mask, Pos, Len,
7274 Low + (V == V1 ? 0 : Size)))
7278 int ShiftEltBits = VT.getScalarSizeInBits() * Scale;
7279 bool ByteShift = ShiftEltBits > 64;
7280 unsigned OpCode = Left ? (ByteShift ? X86ISD::VSHLDQ : X86ISD::VSHLI)
7281 : (ByteShift ? X86ISD::VSRLDQ : X86ISD::VSRLI);
7282 int ShiftAmt = Shift * VT.getScalarSizeInBits() / (ByteShift ? 8 : 1);
7284 // Normalize the scale for byte shifts to still produce an i64 element
7286 Scale = ByteShift ? Scale / 2 : Scale;
7288 // We need to round trip through the appropriate type for the shift.
7289 MVT ShiftSVT = MVT::getIntegerVT(VT.getScalarSizeInBits() * Scale);
7290 MVT ShiftVT = MVT::getVectorVT(ShiftSVT, Size / Scale);
7291 assert(DAG.getTargetLoweringInfo().isTypeLegal(ShiftVT) &&
7292 "Illegal integer vector type");
7293 V = DAG.getBitcast(ShiftVT, V);
7295 V = DAG.getNode(OpCode, DL, ShiftVT, V,
7296 DAG.getConstant(ShiftAmt, DL, MVT::i8));
7297 return DAG.getBitcast(VT, V);
7300 // SSE/AVX supports logical shifts up to 64-bit integers - so we can just
7301 // keep doubling the size of the integer elements up to that. We can
7302 // then shift the elements of the integer vector by whole multiples of
7303 // their width within the elements of the larger integer vector. Test each
7304 // multiple to see if we can find a match with the moved element indices
7305 // and that the shifted in elements are all zeroable.
7306 for (int Scale = 2; Scale * VT.getScalarSizeInBits() <= 128; Scale *= 2)
7307 for (int Shift = 1; Shift != Scale; ++Shift)
7308 for (bool Left : {true, false})
7309 if (CheckZeros(Shift, Scale, Left))
7310 for (SDValue V : {V1, V2})
7311 if (SDValue Match = MatchShift(Shift, Scale, Left, V))
7318 /// \brief Try to lower a vector shuffle using SSE4a EXTRQ/INSERTQ.
7319 static SDValue lowerVectorShuffleWithSSE4A(SDLoc DL, MVT VT, SDValue V1,
7320 SDValue V2, ArrayRef<int> Mask,
7321 SelectionDAG &DAG) {
7322 SmallBitVector Zeroable = computeZeroableShuffleElements(Mask, V1, V2);
7323 assert(!Zeroable.all() && "Fully zeroable shuffle mask");
7325 int Size = Mask.size();
7326 int HalfSize = Size / 2;
7327 assert(Size == (int)VT.getVectorNumElements() && "Unexpected mask size");
7329 // Upper half must be undefined.
7330 if (!isUndefInRange(Mask, HalfSize, HalfSize))
7333 // EXTRQ: Extract Len elements from lower half of source, starting at Idx.
7334 // Remainder of lower half result is zero and upper half is all undef.
7335 auto LowerAsEXTRQ = [&]() {
7336 // Determine the extraction length from the part of the
7337 // lower half that isn't zeroable.
7339 for (; Len > 0; --Len)
7340 if (!Zeroable[Len - 1])
7342 assert(Len > 0 && "Zeroable shuffle mask");
7344 // Attempt to match first Len sequential elements from the lower half.
7347 for (int i = 0; i != Len; ++i) {
7351 SDValue &V = (M < Size ? V1 : V2);
7354 // All mask elements must be in the lower half.
7358 if (Idx < 0 || (Src == V && Idx == (M - i))) {
7369 assert((Idx + Len) <= HalfSize && "Illegal extraction mask");
7370 int BitLen = (Len * VT.getScalarSizeInBits()) & 0x3f;
7371 int BitIdx = (Idx * VT.getScalarSizeInBits()) & 0x3f;
7372 return DAG.getNode(X86ISD::EXTRQI, DL, VT, Src,
7373 DAG.getConstant(BitLen, DL, MVT::i8),
7374 DAG.getConstant(BitIdx, DL, MVT::i8));
7377 if (SDValue ExtrQ = LowerAsEXTRQ())
7380 // INSERTQ: Extract lowest Len elements from lower half of second source and
7381 // insert over first source, starting at Idx.
7382 // { A[0], .., A[Idx-1], B[0], .., B[Len-1], A[Idx+Len], .., UNDEF, ... }
7383 auto LowerAsInsertQ = [&]() {
7384 for (int Idx = 0; Idx != HalfSize; ++Idx) {
7387 // Attempt to match first source from mask before insertion point.
7388 if (isUndefInRange(Mask, 0, Idx)) {
7390 } else if (isSequentialOrUndefInRange(Mask, 0, Idx, 0)) {
7392 } else if (isSequentialOrUndefInRange(Mask, 0, Idx, Size)) {
7398 // Extend the extraction length looking to match both the insertion of
7399 // the second source and the remaining elements of the first.
7400 for (int Hi = Idx + 1; Hi <= HalfSize; ++Hi) {
7405 if (isSequentialOrUndefInRange(Mask, Idx, Len, 0)) {
7407 } else if (isSequentialOrUndefInRange(Mask, Idx, Len, Size)) {
7413 // Match the remaining elements of the lower half.
7414 if (isUndefInRange(Mask, Hi, HalfSize - Hi)) {
7416 } else if ((!Base || (Base == V1)) &&
7417 isSequentialOrUndefInRange(Mask, Hi, HalfSize - Hi, Hi)) {
7419 } else if ((!Base || (Base == V2)) &&
7420 isSequentialOrUndefInRange(Mask, Hi, HalfSize - Hi,
7427 // We may not have a base (first source) - this can safely be undefined.
7429 Base = DAG.getUNDEF(VT);
7431 int BitLen = (Len * VT.getScalarSizeInBits()) & 0x3f;
7432 int BitIdx = (Idx * VT.getScalarSizeInBits()) & 0x3f;
7433 return DAG.getNode(X86ISD::INSERTQI, DL, VT, Base, Insert,
7434 DAG.getConstant(BitLen, DL, MVT::i8),
7435 DAG.getConstant(BitIdx, DL, MVT::i8));
7442 if (SDValue InsertQ = LowerAsInsertQ())
7448 /// \brief Lower a vector shuffle as a zero or any extension.
7450 /// Given a specific number of elements, element bit width, and extension
7451 /// stride, produce either a zero or any extension based on the available
7452 /// features of the subtarget. The extended elements are consecutive and
7453 /// begin and can start from an offseted element index in the input; to
7454 /// avoid excess shuffling the offset must either being in the bottom lane
7455 /// or at the start of a higher lane. All extended elements must be from
7457 static SDValue lowerVectorShuffleAsSpecificZeroOrAnyExtend(
7458 SDLoc DL, MVT VT, int Scale, int Offset, bool AnyExt, SDValue InputV,
7459 ArrayRef<int> Mask, const X86Subtarget *Subtarget, SelectionDAG &DAG) {
7460 assert(Scale > 1 && "Need a scale to extend.");
7461 int EltBits = VT.getScalarSizeInBits();
7462 int NumElements = VT.getVectorNumElements();
7463 int NumEltsPerLane = 128 / EltBits;
7464 int OffsetLane = Offset / NumEltsPerLane;
7465 assert((EltBits == 8 || EltBits == 16 || EltBits == 32) &&
7466 "Only 8, 16, and 32 bit elements can be extended.");
7467 assert(Scale * EltBits <= 64 && "Cannot zero extend past 64 bits.");
7468 assert(0 <= Offset && "Extension offset must be positive.");
7469 assert((Offset < NumEltsPerLane || Offset % NumEltsPerLane == 0) &&
7470 "Extension offset must be in the first lane or start an upper lane.");
7472 // Check that an index is in same lane as the base offset.
7473 auto SafeOffset = [&](int Idx) {
7474 return OffsetLane == (Idx / NumEltsPerLane);
7477 // Shift along an input so that the offset base moves to the first element.
7478 auto ShuffleOffset = [&](SDValue V) {
7482 SmallVector<int, 8> ShMask((unsigned)NumElements, -1);
7483 for (int i = 0; i * Scale < NumElements; ++i) {
7484 int SrcIdx = i + Offset;
7485 ShMask[i] = SafeOffset(SrcIdx) ? SrcIdx : -1;
7487 return DAG.getVectorShuffle(VT, DL, V, DAG.getUNDEF(VT), ShMask);
7490 // Found a valid zext mask! Try various lowering strategies based on the
7491 // input type and available ISA extensions.
7492 if (Subtarget->hasSSE41()) {
7493 // Not worth offseting 128-bit vectors if scale == 2, a pattern using
7494 // PUNPCK will catch this in a later shuffle match.
7495 if (Offset && Scale == 2 && VT.getSizeInBits() == 128)
7497 MVT ExtVT = MVT::getVectorVT(MVT::getIntegerVT(EltBits * Scale),
7498 NumElements / Scale);
7499 InputV = DAG.getNode(X86ISD::VZEXT, DL, ExtVT, ShuffleOffset(InputV));
7500 return DAG.getBitcast(VT, InputV);
7503 assert(VT.getSizeInBits() == 128 && "Only 128-bit vectors can be extended.");
7505 // For any extends we can cheat for larger element sizes and use shuffle
7506 // instructions that can fold with a load and/or copy.
7507 if (AnyExt && EltBits == 32) {
7508 int PSHUFDMask[4] = {Offset, -1, SafeOffset(Offset + 1) ? Offset + 1 : -1,
7510 return DAG.getBitcast(
7511 VT, DAG.getNode(X86ISD::PSHUFD, DL, MVT::v4i32,
7512 DAG.getBitcast(MVT::v4i32, InputV),
7513 getV4X86ShuffleImm8ForMask(PSHUFDMask, DL, DAG)));
7515 if (AnyExt && EltBits == 16 && Scale > 2) {
7516 int PSHUFDMask[4] = {Offset / 2, -1,
7517 SafeOffset(Offset + 1) ? (Offset + 1) / 2 : -1, -1};
7518 InputV = DAG.getNode(X86ISD::PSHUFD, DL, MVT::v4i32,
7519 DAG.getBitcast(MVT::v4i32, InputV),
7520 getV4X86ShuffleImm8ForMask(PSHUFDMask, DL, DAG));
7521 int PSHUFWMask[4] = {1, -1, -1, -1};
7522 unsigned OddEvenOp = (Offset & 1 ? X86ISD::PSHUFLW : X86ISD::PSHUFHW);
7523 return DAG.getBitcast(
7524 VT, DAG.getNode(OddEvenOp, DL, MVT::v8i16,
7525 DAG.getBitcast(MVT::v8i16, InputV),
7526 getV4X86ShuffleImm8ForMask(PSHUFWMask, DL, DAG)));
7529 // The SSE4A EXTRQ instruction can efficiently extend the first 2 lanes
7531 if ((Scale * EltBits) == 64 && EltBits < 32 && Subtarget->hasSSE4A()) {
7532 assert(NumElements == (int)Mask.size() && "Unexpected shuffle mask size!");
7533 assert(VT.getSizeInBits() == 128 && "Unexpected vector width!");
7535 int LoIdx = Offset * EltBits;
7536 SDValue Lo = DAG.getNode(ISD::BITCAST, DL, MVT::v2i64,
7537 DAG.getNode(X86ISD::EXTRQI, DL, VT, InputV,
7538 DAG.getConstant(EltBits, DL, MVT::i8),
7539 DAG.getConstant(LoIdx, DL, MVT::i8)));
7541 if (isUndefInRange(Mask, NumElements / 2, NumElements / 2) ||
7542 !SafeOffset(Offset + 1))
7543 return DAG.getNode(ISD::BITCAST, DL, VT, Lo);
7545 int HiIdx = (Offset + 1) * EltBits;
7546 SDValue Hi = DAG.getNode(ISD::BITCAST, DL, MVT::v2i64,
7547 DAG.getNode(X86ISD::EXTRQI, DL, VT, InputV,
7548 DAG.getConstant(EltBits, DL, MVT::i8),
7549 DAG.getConstant(HiIdx, DL, MVT::i8)));
7550 return DAG.getNode(ISD::BITCAST, DL, VT,
7551 DAG.getNode(X86ISD::UNPCKL, DL, MVT::v2i64, Lo, Hi));
7554 // If this would require more than 2 unpack instructions to expand, use
7555 // pshufb when available. We can only use more than 2 unpack instructions
7556 // when zero extending i8 elements which also makes it easier to use pshufb.
7557 if (Scale > 4 && EltBits == 8 && Subtarget->hasSSSE3()) {
7558 assert(NumElements == 16 && "Unexpected byte vector width!");
7559 SDValue PSHUFBMask[16];
7560 for (int i = 0; i < 16; ++i) {
7561 int Idx = Offset + (i / Scale);
7562 PSHUFBMask[i] = DAG.getConstant(
7563 (i % Scale == 0 && SafeOffset(Idx)) ? Idx : 0x80, DL, MVT::i8);
7565 InputV = DAG.getBitcast(MVT::v16i8, InputV);
7566 return DAG.getBitcast(VT,
7567 DAG.getNode(X86ISD::PSHUFB, DL, MVT::v16i8, InputV,
7568 DAG.getNode(ISD::BUILD_VECTOR, DL,
7569 MVT::v16i8, PSHUFBMask)));
7572 // If we are extending from an offset, ensure we start on a boundary that
7573 // we can unpack from.
7574 int AlignToUnpack = Offset % (NumElements / Scale);
7575 if (AlignToUnpack) {
7576 SmallVector<int, 8> ShMask((unsigned)NumElements, -1);
7577 for (int i = AlignToUnpack; i < NumElements; ++i)
7578 ShMask[i - AlignToUnpack] = i;
7579 InputV = DAG.getVectorShuffle(VT, DL, InputV, DAG.getUNDEF(VT), ShMask);
7580 Offset -= AlignToUnpack;
7583 // Otherwise emit a sequence of unpacks.
7585 unsigned UnpackLoHi = X86ISD::UNPCKL;
7586 if (Offset >= (NumElements / 2)) {
7587 UnpackLoHi = X86ISD::UNPCKH;
7588 Offset -= (NumElements / 2);
7591 MVT InputVT = MVT::getVectorVT(MVT::getIntegerVT(EltBits), NumElements);
7592 SDValue Ext = AnyExt ? DAG.getUNDEF(InputVT)
7593 : getZeroVector(InputVT, Subtarget, DAG, DL);
7594 InputV = DAG.getBitcast(InputVT, InputV);
7595 InputV = DAG.getNode(UnpackLoHi, DL, InputVT, InputV, Ext);
7599 } while (Scale > 1);
7600 return DAG.getBitcast(VT, InputV);
7603 /// \brief Try to lower a vector shuffle as a zero extension on any microarch.
7605 /// This routine will try to do everything in its power to cleverly lower
7606 /// a shuffle which happens to match the pattern of a zero extend. It doesn't
7607 /// check for the profitability of this lowering, it tries to aggressively
7608 /// match this pattern. It will use all of the micro-architectural details it
7609 /// can to emit an efficient lowering. It handles both blends with all-zero
7610 /// inputs to explicitly zero-extend and undef-lanes (sometimes undef due to
7611 /// masking out later).
7613 /// The reason we have dedicated lowering for zext-style shuffles is that they
7614 /// are both incredibly common and often quite performance sensitive.
7615 static SDValue lowerVectorShuffleAsZeroOrAnyExtend(
7616 SDLoc DL, MVT VT, SDValue V1, SDValue V2, ArrayRef<int> Mask,
7617 const X86Subtarget *Subtarget, SelectionDAG &DAG) {
7618 SmallBitVector Zeroable = computeZeroableShuffleElements(Mask, V1, V2);
7620 int Bits = VT.getSizeInBits();
7621 int NumLanes = Bits / 128;
7622 int NumElements = VT.getVectorNumElements();
7623 int NumEltsPerLane = NumElements / NumLanes;
7624 assert(VT.getScalarSizeInBits() <= 32 &&
7625 "Exceeds 32-bit integer zero extension limit");
7626 assert((int)Mask.size() == NumElements && "Unexpected shuffle mask size");
7628 // Define a helper function to check a particular ext-scale and lower to it if
7630 auto Lower = [&](int Scale) -> SDValue {
7635 for (int i = 0; i < NumElements; ++i) {
7638 continue; // Valid anywhere but doesn't tell us anything.
7639 if (i % Scale != 0) {
7640 // Each of the extended elements need to be zeroable.
7644 // We no longer are in the anyext case.
7649 // Each of the base elements needs to be consecutive indices into the
7650 // same input vector.
7651 SDValue V = M < NumElements ? V1 : V2;
7652 M = M % NumElements;
7655 Offset = M - (i / Scale);
7656 } else if (InputV != V)
7657 return SDValue(); // Flip-flopping inputs.
7659 // Offset must start in the lowest 128-bit lane or at the start of an
7661 // FIXME: Is it ever worth allowing a negative base offset?
7662 if (!((0 <= Offset && Offset < NumEltsPerLane) ||
7663 (Offset % NumEltsPerLane) == 0))
7666 // If we are offsetting, all referenced entries must come from the same
7668 if (Offset && (Offset / NumEltsPerLane) != (M / NumEltsPerLane))
7671 if ((M % NumElements) != (Offset + (i / Scale)))
7672 return SDValue(); // Non-consecutive strided elements.
7676 // If we fail to find an input, we have a zero-shuffle which should always
7677 // have already been handled.
7678 // FIXME: Maybe handle this here in case during blending we end up with one?
7682 // If we are offsetting, don't extend if we only match a single input, we
7683 // can always do better by using a basic PSHUF or PUNPCK.
7684 if (Offset != 0 && Matches < 2)
7687 return lowerVectorShuffleAsSpecificZeroOrAnyExtend(
7688 DL, VT, Scale, Offset, AnyExt, InputV, Mask, Subtarget, DAG);
7691 // The widest scale possible for extending is to a 64-bit integer.
7692 assert(Bits % 64 == 0 &&
7693 "The number of bits in a vector must be divisible by 64 on x86!");
7694 int NumExtElements = Bits / 64;
7696 // Each iteration, try extending the elements half as much, but into twice as
7698 for (; NumExtElements < NumElements; NumExtElements *= 2) {
7699 assert(NumElements % NumExtElements == 0 &&
7700 "The input vector size must be divisible by the extended size.");
7701 if (SDValue V = Lower(NumElements / NumExtElements))
7705 // General extends failed, but 128-bit vectors may be able to use MOVQ.
7709 // Returns one of the source operands if the shuffle can be reduced to a
7710 // MOVQ, copying the lower 64-bits and zero-extending to the upper 64-bits.
7711 auto CanZExtLowHalf = [&]() {
7712 for (int i = NumElements / 2; i != NumElements; ++i)
7715 if (isSequentialOrUndefInRange(Mask, 0, NumElements / 2, 0))
7717 if (isSequentialOrUndefInRange(Mask, 0, NumElements / 2, NumElements))
7722 if (SDValue V = CanZExtLowHalf()) {
7723 V = DAG.getBitcast(MVT::v2i64, V);
7724 V = DAG.getNode(X86ISD::VZEXT_MOVL, DL, MVT::v2i64, V);
7725 return DAG.getBitcast(VT, V);
7728 // No viable ext lowering found.
7732 /// \brief Try to get a scalar value for a specific element of a vector.
7734 /// Looks through BUILD_VECTOR and SCALAR_TO_VECTOR nodes to find a scalar.
7735 static SDValue getScalarValueForVectorElement(SDValue V, int Idx,
7736 SelectionDAG &DAG) {
7737 MVT VT = V.getSimpleValueType();
7738 MVT EltVT = VT.getVectorElementType();
7739 while (V.getOpcode() == ISD::BITCAST)
7740 V = V.getOperand(0);
7741 // If the bitcasts shift the element size, we can't extract an equivalent
7743 MVT NewVT = V.getSimpleValueType();
7744 if (!NewVT.isVector() || NewVT.getScalarSizeInBits() != VT.getScalarSizeInBits())
7747 if (V.getOpcode() == ISD::BUILD_VECTOR ||
7748 (Idx == 0 && V.getOpcode() == ISD::SCALAR_TO_VECTOR)) {
7749 // Ensure the scalar operand is the same size as the destination.
7750 // FIXME: Add support for scalar truncation where possible.
7751 SDValue S = V.getOperand(Idx);
7752 if (EltVT.getSizeInBits() == S.getSimpleValueType().getSizeInBits())
7753 return DAG.getNode(ISD::BITCAST, SDLoc(V), EltVT, S);
7759 /// \brief Helper to test for a load that can be folded with x86 shuffles.
7761 /// This is particularly important because the set of instructions varies
7762 /// significantly based on whether the operand is a load or not.
7763 static bool isShuffleFoldableLoad(SDValue V) {
7764 while (V.getOpcode() == ISD::BITCAST)
7765 V = V.getOperand(0);
7767 return ISD::isNON_EXTLoad(V.getNode());
7770 /// \brief Try to lower insertion of a single element into a zero vector.
7772 /// This is a common pattern that we have especially efficient patterns to lower
7773 /// across all subtarget feature sets.
7774 static SDValue lowerVectorShuffleAsElementInsertion(
7775 SDLoc DL, MVT VT, SDValue V1, SDValue V2, ArrayRef<int> Mask,
7776 const X86Subtarget *Subtarget, SelectionDAG &DAG) {
7777 SmallBitVector Zeroable = computeZeroableShuffleElements(Mask, V1, V2);
7779 MVT EltVT = VT.getVectorElementType();
7781 int V2Index = std::find_if(Mask.begin(), Mask.end(),
7782 [&Mask](int M) { return M >= (int)Mask.size(); }) -
7784 bool IsV1Zeroable = true;
7785 for (int i = 0, Size = Mask.size(); i < Size; ++i)
7786 if (i != V2Index && !Zeroable[i]) {
7787 IsV1Zeroable = false;
7791 // Check for a single input from a SCALAR_TO_VECTOR node.
7792 // FIXME: All of this should be canonicalized into INSERT_VECTOR_ELT and
7793 // all the smarts here sunk into that routine. However, the current
7794 // lowering of BUILD_VECTOR makes that nearly impossible until the old
7795 // vector shuffle lowering is dead.
7796 SDValue V2S = getScalarValueForVectorElement(V2, Mask[V2Index] - Mask.size(),
7798 if (V2S && DAG.getTargetLoweringInfo().isTypeLegal(V2S.getValueType())) {
7799 // We need to zext the scalar if it is smaller than an i32.
7800 V2S = DAG.getBitcast(EltVT, V2S);
7801 if (EltVT == MVT::i8 || EltVT == MVT::i16) {
7802 // Using zext to expand a narrow element won't work for non-zero
7807 // Zero-extend directly to i32.
7809 V2S = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::i32, V2S);
7811 V2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, ExtVT, V2S);
7812 } else if (Mask[V2Index] != (int)Mask.size() || EltVT == MVT::i8 ||
7813 EltVT == MVT::i16) {
7814 // Either not inserting from the low element of the input or the input
7815 // element size is too small to use VZEXT_MOVL to clear the high bits.
7819 if (!IsV1Zeroable) {
7820 // If V1 can't be treated as a zero vector we have fewer options to lower
7821 // this. We can't support integer vectors or non-zero targets cheaply, and
7822 // the V1 elements can't be permuted in any way.
7823 assert(VT == ExtVT && "Cannot change extended type when non-zeroable!");
7824 if (!VT.isFloatingPoint() || V2Index != 0)
7826 SmallVector<int, 8> V1Mask(Mask.begin(), Mask.end());
7827 V1Mask[V2Index] = -1;
7828 if (!isNoopShuffleMask(V1Mask))
7830 // This is essentially a special case blend operation, but if we have
7831 // general purpose blend operations, they are always faster. Bail and let
7832 // the rest of the lowering handle these as blends.
7833 if (Subtarget->hasSSE41())
7836 // Otherwise, use MOVSD or MOVSS.
7837 assert((EltVT == MVT::f32 || EltVT == MVT::f64) &&
7838 "Only two types of floating point element types to handle!");
7839 return DAG.getNode(EltVT == MVT::f32 ? X86ISD::MOVSS : X86ISD::MOVSD, DL,
7843 // This lowering only works for the low element with floating point vectors.
7844 if (VT.isFloatingPoint() && V2Index != 0)
7847 V2 = DAG.getNode(X86ISD::VZEXT_MOVL, DL, ExtVT, V2);
7849 V2 = DAG.getBitcast(VT, V2);
7852 // If we have 4 or fewer lanes we can cheaply shuffle the element into
7853 // the desired position. Otherwise it is more efficient to do a vector
7854 // shift left. We know that we can do a vector shift left because all
7855 // the inputs are zero.
7856 if (VT.isFloatingPoint() || VT.getVectorNumElements() <= 4) {
7857 SmallVector<int, 4> V2Shuffle(Mask.size(), 1);
7858 V2Shuffle[V2Index] = 0;
7859 V2 = DAG.getVectorShuffle(VT, DL, V2, DAG.getUNDEF(VT), V2Shuffle);
7861 V2 = DAG.getBitcast(MVT::v2i64, V2);
7863 X86ISD::VSHLDQ, DL, MVT::v2i64, V2,
7864 DAG.getConstant(V2Index * EltVT.getSizeInBits() / 8, DL,
7865 DAG.getTargetLoweringInfo().getScalarShiftAmountTy(
7866 DAG.getDataLayout(), VT)));
7867 V2 = DAG.getBitcast(VT, V2);
7873 /// \brief Try to lower broadcast of a single element.
7875 /// For convenience, this code also bundles all of the subtarget feature set
7876 /// filtering. While a little annoying to re-dispatch on type here, there isn't
7877 /// a convenient way to factor it out.
7878 static SDValue lowerVectorShuffleAsBroadcast(SDLoc DL, MVT VT, SDValue V,
7880 const X86Subtarget *Subtarget,
7881 SelectionDAG &DAG) {
7882 if (!Subtarget->hasAVX())
7884 if (VT.isInteger() && !Subtarget->hasAVX2())
7887 // Check that the mask is a broadcast.
7888 int BroadcastIdx = -1;
7890 if (M >= 0 && BroadcastIdx == -1)
7892 else if (M >= 0 && M != BroadcastIdx)
7895 assert(BroadcastIdx < (int)Mask.size() && "We only expect to be called with "
7896 "a sorted mask where the broadcast "
7899 // Go up the chain of (vector) values to find a scalar load that we can
7900 // combine with the broadcast.
7902 switch (V.getOpcode()) {
7903 case ISD::CONCAT_VECTORS: {
7904 int OperandSize = Mask.size() / V.getNumOperands();
7905 V = V.getOperand(BroadcastIdx / OperandSize);
7906 BroadcastIdx %= OperandSize;
7910 case ISD::INSERT_SUBVECTOR: {
7911 SDValue VOuter = V.getOperand(0), VInner = V.getOperand(1);
7912 auto ConstantIdx = dyn_cast<ConstantSDNode>(V.getOperand(2));
7916 int BeginIdx = (int)ConstantIdx->getZExtValue();
7918 BeginIdx + (int)VInner.getValueType().getVectorNumElements();
7919 if (BroadcastIdx >= BeginIdx && BroadcastIdx < EndIdx) {
7920 BroadcastIdx -= BeginIdx;
7931 // Check if this is a broadcast of a scalar. We special case lowering
7932 // for scalars so that we can more effectively fold with loads.
7933 // First, look through bitcast: if the original value has a larger element
7934 // type than the shuffle, the broadcast element is in essence truncated.
7935 // Make that explicit to ease folding.
7936 if (V.getOpcode() == ISD::BITCAST && VT.isInteger()) {
7937 EVT EltVT = VT.getVectorElementType();
7938 SDValue V0 = V.getOperand(0);
7939 EVT V0VT = V0.getValueType();
7941 if (V0VT.isInteger() && V0VT.getVectorElementType().bitsGT(EltVT) &&
7942 ((V0.getOpcode() == ISD::BUILD_VECTOR ||
7943 (V0.getOpcode() == ISD::SCALAR_TO_VECTOR && BroadcastIdx == 0)))) {
7944 V = DAG.getNode(ISD::TRUNCATE, DL, EltVT, V0.getOperand(BroadcastIdx));
7949 // Also check the simpler case, where we can directly reuse the scalar.
7950 if (V.getOpcode() == ISD::BUILD_VECTOR ||
7951 (V.getOpcode() == ISD::SCALAR_TO_VECTOR && BroadcastIdx == 0)) {
7952 V = V.getOperand(BroadcastIdx);
7954 // If the scalar isn't a load, we can't broadcast from it in AVX1.
7955 // Only AVX2 has register broadcasts.
7956 if (!Subtarget->hasAVX2() && !isShuffleFoldableLoad(V))
7958 } else if (BroadcastIdx != 0 || !Subtarget->hasAVX2()) {
7959 // We can't broadcast from a vector register without AVX2, and we can only
7960 // broadcast from the zero-element of a vector register.
7964 return DAG.getNode(X86ISD::VBROADCAST, DL, VT, V);
7967 // Check for whether we can use INSERTPS to perform the shuffle. We only use
7968 // INSERTPS when the V1 elements are already in the correct locations
7969 // because otherwise we can just always use two SHUFPS instructions which
7970 // are much smaller to encode than a SHUFPS and an INSERTPS. We can also
7971 // perform INSERTPS if a single V1 element is out of place and all V2
7972 // elements are zeroable.
7973 static SDValue lowerVectorShuffleAsInsertPS(SDValue Op, SDValue V1, SDValue V2,
7975 SelectionDAG &DAG) {
7976 assert(Op.getSimpleValueType() == MVT::v4f32 && "Bad shuffle type!");
7977 assert(V1.getSimpleValueType() == MVT::v4f32 && "Bad operand type!");
7978 assert(V2.getSimpleValueType() == MVT::v4f32 && "Bad operand type!");
7979 assert(Mask.size() == 4 && "Unexpected mask size for v4 shuffle!");
7981 SmallBitVector Zeroable = computeZeroableShuffleElements(Mask, V1, V2);
7984 int V1DstIndex = -1;
7985 int V2DstIndex = -1;
7986 bool V1UsedInPlace = false;
7988 for (int i = 0; i < 4; ++i) {
7989 // Synthesize a zero mask from the zeroable elements (includes undefs).
7995 // Flag if we use any V1 inputs in place.
7997 V1UsedInPlace = true;
8001 // We can only insert a single non-zeroable element.
8002 if (V1DstIndex != -1 || V2DstIndex != -1)
8006 // V1 input out of place for insertion.
8009 // V2 input for insertion.
8014 // Don't bother if we have no (non-zeroable) element for insertion.
8015 if (V1DstIndex == -1 && V2DstIndex == -1)
8018 // Determine element insertion src/dst indices. The src index is from the
8019 // start of the inserted vector, not the start of the concatenated vector.
8020 unsigned V2SrcIndex = 0;
8021 if (V1DstIndex != -1) {
8022 // If we have a V1 input out of place, we use V1 as the V2 element insertion
8023 // and don't use the original V2 at all.
8024 V2SrcIndex = Mask[V1DstIndex];
8025 V2DstIndex = V1DstIndex;
8028 V2SrcIndex = Mask[V2DstIndex] - 4;
8031 // If no V1 inputs are used in place, then the result is created only from
8032 // the zero mask and the V2 insertion - so remove V1 dependency.
8034 V1 = DAG.getUNDEF(MVT::v4f32);
8036 unsigned InsertPSMask = V2SrcIndex << 6 | V2DstIndex << 4 | ZMask;
8037 assert((InsertPSMask & ~0xFFu) == 0 && "Invalid mask!");
8039 // Insert the V2 element into the desired position.
8041 return DAG.getNode(X86ISD::INSERTPS, DL, MVT::v4f32, V1, V2,
8042 DAG.getConstant(InsertPSMask, DL, MVT::i8));
8045 /// \brief Try to lower a shuffle as a permute of the inputs followed by an
8046 /// UNPCK instruction.
8048 /// This specifically targets cases where we end up with alternating between
8049 /// the two inputs, and so can permute them into something that feeds a single
8050 /// UNPCK instruction. Note that this routine only targets integer vectors
8051 /// because for floating point vectors we have a generalized SHUFPS lowering
8052 /// strategy that handles everything that doesn't *exactly* match an unpack,
8053 /// making this clever lowering unnecessary.
8054 static SDValue lowerVectorShuffleAsPermuteAndUnpack(SDLoc DL, MVT VT,
8055 SDValue V1, SDValue V2,
8057 SelectionDAG &DAG) {
8058 assert(!VT.isFloatingPoint() &&
8059 "This routine only supports integer vectors.");
8060 assert(!isSingleInputShuffleMask(Mask) &&
8061 "This routine should only be used when blending two inputs.");
8062 assert(Mask.size() >= 2 && "Single element masks are invalid.");
8064 int Size = Mask.size();
8066 int NumLoInputs = std::count_if(Mask.begin(), Mask.end(), [Size](int M) {
8067 return M >= 0 && M % Size < Size / 2;
8069 int NumHiInputs = std::count_if(
8070 Mask.begin(), Mask.end(), [Size](int M) { return M % Size >= Size / 2; });
8072 bool UnpackLo = NumLoInputs >= NumHiInputs;
8074 auto TryUnpack = [&](MVT UnpackVT, int Scale) {
8075 SmallVector<int, 32> V1Mask(Mask.size(), -1);
8076 SmallVector<int, 32> V2Mask(Mask.size(), -1);
8078 for (int i = 0; i < Size; ++i) {
8082 // Each element of the unpack contains Scale elements from this mask.
8083 int UnpackIdx = i / Scale;
8085 // We only handle the case where V1 feeds the first slots of the unpack.
8086 // We rely on canonicalization to ensure this is the case.
8087 if ((UnpackIdx % 2 == 0) != (Mask[i] < Size))
8090 // Setup the mask for this input. The indexing is tricky as we have to
8091 // handle the unpack stride.
8092 SmallVectorImpl<int> &VMask = (UnpackIdx % 2 == 0) ? V1Mask : V2Mask;
8093 VMask[(UnpackIdx / 2) * Scale + i % Scale + (UnpackLo ? 0 : Size / 2)] =
8097 // If we will have to shuffle both inputs to use the unpack, check whether
8098 // we can just unpack first and shuffle the result. If so, skip this unpack.
8099 if ((NumLoInputs == 0 || NumHiInputs == 0) && !isNoopShuffleMask(V1Mask) &&
8100 !isNoopShuffleMask(V2Mask))
8103 // Shuffle the inputs into place.
8104 V1 = DAG.getVectorShuffle(VT, DL, V1, DAG.getUNDEF(VT), V1Mask);
8105 V2 = DAG.getVectorShuffle(VT, DL, V2, DAG.getUNDEF(VT), V2Mask);
8107 // Cast the inputs to the type we will use to unpack them.
8108 V1 = DAG.getBitcast(UnpackVT, V1);
8109 V2 = DAG.getBitcast(UnpackVT, V2);
8111 // Unpack the inputs and cast the result back to the desired type.
8112 return DAG.getBitcast(
8113 VT, DAG.getNode(UnpackLo ? X86ISD::UNPCKL : X86ISD::UNPCKH, DL,
8117 // We try each unpack from the largest to the smallest to try and find one
8118 // that fits this mask.
8119 int OrigNumElements = VT.getVectorNumElements();
8120 int OrigScalarSize = VT.getScalarSizeInBits();
8121 for (int ScalarSize = 64; ScalarSize >= OrigScalarSize; ScalarSize /= 2) {
8122 int Scale = ScalarSize / OrigScalarSize;
8123 int NumElements = OrigNumElements / Scale;
8124 MVT UnpackVT = MVT::getVectorVT(MVT::getIntegerVT(ScalarSize), NumElements);
8125 if (SDValue Unpack = TryUnpack(UnpackVT, Scale))
8129 // If none of the unpack-rooted lowerings worked (or were profitable) try an
8131 if (NumLoInputs == 0 || NumHiInputs == 0) {
8132 assert((NumLoInputs > 0 || NumHiInputs > 0) &&
8133 "We have to have *some* inputs!");
8134 int HalfOffset = NumLoInputs == 0 ? Size / 2 : 0;
8136 // FIXME: We could consider the total complexity of the permute of each
8137 // possible unpacking. Or at the least we should consider how many
8138 // half-crossings are created.
8139 // FIXME: We could consider commuting the unpacks.
8141 SmallVector<int, 32> PermMask;
8142 PermMask.assign(Size, -1);
8143 for (int i = 0; i < Size; ++i) {
8147 assert(Mask[i] % Size >= HalfOffset && "Found input from wrong half!");
8150 2 * ((Mask[i] % Size) - HalfOffset) + (Mask[i] < Size ? 0 : 1);
8152 return DAG.getVectorShuffle(
8153 VT, DL, DAG.getNode(NumLoInputs == 0 ? X86ISD::UNPCKH : X86ISD::UNPCKL,
8155 DAG.getUNDEF(VT), PermMask);
8161 /// \brief Handle lowering of 2-lane 64-bit floating point shuffles.
8163 /// This is the basis function for the 2-lane 64-bit shuffles as we have full
8164 /// support for floating point shuffles but not integer shuffles. These
8165 /// instructions will incur a domain crossing penalty on some chips though so
8166 /// it is better to avoid lowering through this for integer vectors where
8168 static SDValue lowerV2F64VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
8169 const X86Subtarget *Subtarget,
8170 SelectionDAG &DAG) {
8172 assert(Op.getSimpleValueType() == MVT::v2f64 && "Bad shuffle type!");
8173 assert(V1.getSimpleValueType() == MVT::v2f64 && "Bad operand type!");
8174 assert(V2.getSimpleValueType() == MVT::v2f64 && "Bad operand type!");
8175 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
8176 ArrayRef<int> Mask = SVOp->getMask();
8177 assert(Mask.size() == 2 && "Unexpected mask size for v2 shuffle!");
8179 if (isSingleInputShuffleMask(Mask)) {
8180 // Use low duplicate instructions for masks that match their pattern.
8181 if (Subtarget->hasSSE3())
8182 if (isShuffleEquivalent(V1, V2, Mask, {0, 0}))
8183 return DAG.getNode(X86ISD::MOVDDUP, DL, MVT::v2f64, V1);
8185 // Straight shuffle of a single input vector. Simulate this by using the
8186 // single input as both of the "inputs" to this instruction..
8187 unsigned SHUFPDMask = (Mask[0] == 1) | ((Mask[1] == 1) << 1);
8189 if (Subtarget->hasAVX()) {
8190 // If we have AVX, we can use VPERMILPS which will allow folding a load
8191 // into the shuffle.
8192 return DAG.getNode(X86ISD::VPERMILPI, DL, MVT::v2f64, V1,
8193 DAG.getConstant(SHUFPDMask, DL, MVT::i8));
8196 return DAG.getNode(X86ISD::SHUFP, DL, MVT::v2f64, V1, V1,
8197 DAG.getConstant(SHUFPDMask, DL, MVT::i8));
8199 assert(Mask[0] >= 0 && Mask[0] < 2 && "Non-canonicalized blend!");
8200 assert(Mask[1] >= 2 && "Non-canonicalized blend!");
8202 // If we have a single input, insert that into V1 if we can do so cheaply.
8203 if ((Mask[0] >= 2) + (Mask[1] >= 2) == 1) {
8204 if (SDValue Insertion = lowerVectorShuffleAsElementInsertion(
8205 DL, MVT::v2f64, V1, V2, Mask, Subtarget, DAG))
8207 // Try inverting the insertion since for v2 masks it is easy to do and we
8208 // can't reliably sort the mask one way or the other.
8209 int InverseMask[2] = {Mask[0] < 0 ? -1 : (Mask[0] ^ 2),
8210 Mask[1] < 0 ? -1 : (Mask[1] ^ 2)};
8211 if (SDValue Insertion = lowerVectorShuffleAsElementInsertion(
8212 DL, MVT::v2f64, V2, V1, InverseMask, Subtarget, DAG))
8216 // Try to use one of the special instruction patterns to handle two common
8217 // blend patterns if a zero-blend above didn't work.
8218 if (isShuffleEquivalent(V1, V2, Mask, {0, 3}) ||
8219 isShuffleEquivalent(V1, V2, Mask, {1, 3}))
8220 if (SDValue V1S = getScalarValueForVectorElement(V1, Mask[0], DAG))
8221 // We can either use a special instruction to load over the low double or
8222 // to move just the low double.
8224 isShuffleFoldableLoad(V1S) ? X86ISD::MOVLPD : X86ISD::MOVSD,
8226 DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, MVT::v2f64, V1S));
8228 if (Subtarget->hasSSE41())
8229 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v2f64, V1, V2, Mask,
8233 // Use dedicated unpack instructions for masks that match their pattern.
8235 lowerVectorShuffleWithUNPCK(DL, MVT::v2f64, Mask, V1, V2, DAG))
8238 unsigned SHUFPDMask = (Mask[0] == 1) | (((Mask[1] - 2) == 1) << 1);
8239 return DAG.getNode(X86ISD::SHUFP, DL, MVT::v2f64, V1, V2,
8240 DAG.getConstant(SHUFPDMask, DL, MVT::i8));
8243 /// \brief Handle lowering of 2-lane 64-bit integer shuffles.
8245 /// Tries to lower a 2-lane 64-bit shuffle using shuffle operations provided by
8246 /// the integer unit to minimize domain crossing penalties. However, for blends
8247 /// it falls back to the floating point shuffle operation with appropriate bit
8249 static SDValue lowerV2I64VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
8250 const X86Subtarget *Subtarget,
8251 SelectionDAG &DAG) {
8253 assert(Op.getSimpleValueType() == MVT::v2i64 && "Bad shuffle type!");
8254 assert(V1.getSimpleValueType() == MVT::v2i64 && "Bad operand type!");
8255 assert(V2.getSimpleValueType() == MVT::v2i64 && "Bad operand type!");
8256 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
8257 ArrayRef<int> Mask = SVOp->getMask();
8258 assert(Mask.size() == 2 && "Unexpected mask size for v2 shuffle!");
8260 if (isSingleInputShuffleMask(Mask)) {
8261 // Check for being able to broadcast a single element.
8262 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v2i64, V1,
8263 Mask, Subtarget, DAG))
8266 // Straight shuffle of a single input vector. For everything from SSE2
8267 // onward this has a single fast instruction with no scary immediates.
8268 // We have to map the mask as it is actually a v4i32 shuffle instruction.
8269 V1 = DAG.getBitcast(MVT::v4i32, V1);
8270 int WidenedMask[4] = {
8271 std::max(Mask[0], 0) * 2, std::max(Mask[0], 0) * 2 + 1,
8272 std::max(Mask[1], 0) * 2, std::max(Mask[1], 0) * 2 + 1};
8273 return DAG.getBitcast(
8275 DAG.getNode(X86ISD::PSHUFD, DL, MVT::v4i32, V1,
8276 getV4X86ShuffleImm8ForMask(WidenedMask, DL, DAG)));
8278 assert(Mask[0] != -1 && "No undef lanes in multi-input v2 shuffles!");
8279 assert(Mask[1] != -1 && "No undef lanes in multi-input v2 shuffles!");
8280 assert(Mask[0] < 2 && "We sort V1 to be the first input.");
8281 assert(Mask[1] >= 2 && "We sort V2 to be the second input.");
8283 // If we have a blend of two PACKUS operations an the blend aligns with the
8284 // low and half halves, we can just merge the PACKUS operations. This is
8285 // particularly important as it lets us merge shuffles that this routine itself
8287 auto GetPackNode = [](SDValue V) {
8288 while (V.getOpcode() == ISD::BITCAST)
8289 V = V.getOperand(0);
8291 return V.getOpcode() == X86ISD::PACKUS ? V : SDValue();
8293 if (SDValue V1Pack = GetPackNode(V1))
8294 if (SDValue V2Pack = GetPackNode(V2))
8295 return DAG.getBitcast(MVT::v2i64,
8296 DAG.getNode(X86ISD::PACKUS, DL, MVT::v16i8,
8297 Mask[0] == 0 ? V1Pack.getOperand(0)
8298 : V1Pack.getOperand(1),
8299 Mask[1] == 2 ? V2Pack.getOperand(0)
8300 : V2Pack.getOperand(1)));
8302 // Try to use shift instructions.
8304 lowerVectorShuffleAsShift(DL, MVT::v2i64, V1, V2, Mask, DAG))
8307 // When loading a scalar and then shuffling it into a vector we can often do
8308 // the insertion cheaply.
8309 if (SDValue Insertion = lowerVectorShuffleAsElementInsertion(
8310 DL, MVT::v2i64, V1, V2, Mask, Subtarget, DAG))
8312 // Try inverting the insertion since for v2 masks it is easy to do and we
8313 // can't reliably sort the mask one way or the other.
8314 int InverseMask[2] = {Mask[0] ^ 2, Mask[1] ^ 2};
8315 if (SDValue Insertion = lowerVectorShuffleAsElementInsertion(
8316 DL, MVT::v2i64, V2, V1, InverseMask, Subtarget, DAG))
8319 // We have different paths for blend lowering, but they all must use the
8320 // *exact* same predicate.
8321 bool IsBlendSupported = Subtarget->hasSSE41();
8322 if (IsBlendSupported)
8323 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v2i64, V1, V2, Mask,
8327 // Use dedicated unpack instructions for masks that match their pattern.
8329 lowerVectorShuffleWithUNPCK(DL, MVT::v2i64, Mask, V1, V2, DAG))
8332 // Try to use byte rotation instructions.
8333 // Its more profitable for pre-SSSE3 to use shuffles/unpacks.
8334 if (Subtarget->hasSSSE3())
8335 if (SDValue Rotate = lowerVectorShuffleAsByteRotate(
8336 DL, MVT::v2i64, V1, V2, Mask, Subtarget, DAG))
8339 // If we have direct support for blends, we should lower by decomposing into
8340 // a permute. That will be faster than the domain cross.
8341 if (IsBlendSupported)
8342 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v2i64, V1, V2,
8345 // We implement this with SHUFPD which is pretty lame because it will likely
8346 // incur 2 cycles of stall for integer vectors on Nehalem and older chips.
8347 // However, all the alternatives are still more cycles and newer chips don't
8348 // have this problem. It would be really nice if x86 had better shuffles here.
8349 V1 = DAG.getBitcast(MVT::v2f64, V1);
8350 V2 = DAG.getBitcast(MVT::v2f64, V2);
8351 return DAG.getBitcast(MVT::v2i64,
8352 DAG.getVectorShuffle(MVT::v2f64, DL, V1, V2, Mask));
8355 /// \brief Test whether this can be lowered with a single SHUFPS instruction.
8357 /// This is used to disable more specialized lowerings when the shufps lowering
8358 /// will happen to be efficient.
8359 static bool isSingleSHUFPSMask(ArrayRef<int> Mask) {
8360 // This routine only handles 128-bit shufps.
8361 assert(Mask.size() == 4 && "Unsupported mask size!");
8363 // To lower with a single SHUFPS we need to have the low half and high half
8364 // each requiring a single input.
8365 if (Mask[0] != -1 && Mask[1] != -1 && (Mask[0] < 4) != (Mask[1] < 4))
8367 if (Mask[2] != -1 && Mask[3] != -1 && (Mask[2] < 4) != (Mask[3] < 4))
8373 /// \brief Lower a vector shuffle using the SHUFPS instruction.
8375 /// This is a helper routine dedicated to lowering vector shuffles using SHUFPS.
8376 /// It makes no assumptions about whether this is the *best* lowering, it simply
8378 static SDValue lowerVectorShuffleWithSHUFPS(SDLoc DL, MVT VT,
8379 ArrayRef<int> Mask, SDValue V1,
8380 SDValue V2, SelectionDAG &DAG) {
8381 SDValue LowV = V1, HighV = V2;
8382 int NewMask[4] = {Mask[0], Mask[1], Mask[2], Mask[3]};
8385 std::count_if(Mask.begin(), Mask.end(), [](int M) { return M >= 4; });
8387 if (NumV2Elements == 1) {
8389 std::find_if(Mask.begin(), Mask.end(), [](int M) { return M >= 4; }) -
8392 // Compute the index adjacent to V2Index and in the same half by toggling
8394 int V2AdjIndex = V2Index ^ 1;
8396 if (Mask[V2AdjIndex] == -1) {
8397 // Handles all the cases where we have a single V2 element and an undef.
8398 // This will only ever happen in the high lanes because we commute the
8399 // vector otherwise.
8401 std::swap(LowV, HighV);
8402 NewMask[V2Index] -= 4;
8404 // Handle the case where the V2 element ends up adjacent to a V1 element.
8405 // To make this work, blend them together as the first step.
8406 int V1Index = V2AdjIndex;
8407 int BlendMask[4] = {Mask[V2Index] - 4, 0, Mask[V1Index], 0};
8408 V2 = DAG.getNode(X86ISD::SHUFP, DL, VT, V2, V1,
8409 getV4X86ShuffleImm8ForMask(BlendMask, DL, DAG));
8411 // Now proceed to reconstruct the final blend as we have the necessary
8412 // high or low half formed.
8419 NewMask[V1Index] = 2; // We put the V1 element in V2[2].
8420 NewMask[V2Index] = 0; // We shifted the V2 element into V2[0].
8422 } else if (NumV2Elements == 2) {
8423 if (Mask[0] < 4 && Mask[1] < 4) {
8424 // Handle the easy case where we have V1 in the low lanes and V2 in the
8428 } else if (Mask[2] < 4 && Mask[3] < 4) {
8429 // We also handle the reversed case because this utility may get called
8430 // when we detect a SHUFPS pattern but can't easily commute the shuffle to
8431 // arrange things in the right direction.
8437 // We have a mixture of V1 and V2 in both low and high lanes. Rather than
8438 // trying to place elements directly, just blend them and set up the final
8439 // shuffle to place them.
8441 // The first two blend mask elements are for V1, the second two are for
8443 int BlendMask[4] = {Mask[0] < 4 ? Mask[0] : Mask[1],
8444 Mask[2] < 4 ? Mask[2] : Mask[3],
8445 (Mask[0] >= 4 ? Mask[0] : Mask[1]) - 4,
8446 (Mask[2] >= 4 ? Mask[2] : Mask[3]) - 4};
8447 V1 = DAG.getNode(X86ISD::SHUFP, DL, VT, V1, V2,
8448 getV4X86ShuffleImm8ForMask(BlendMask, DL, DAG));
8450 // Now we do a normal shuffle of V1 by giving V1 as both operands to
8453 NewMask[0] = Mask[0] < 4 ? 0 : 2;
8454 NewMask[1] = Mask[0] < 4 ? 2 : 0;
8455 NewMask[2] = Mask[2] < 4 ? 1 : 3;
8456 NewMask[3] = Mask[2] < 4 ? 3 : 1;
8459 return DAG.getNode(X86ISD::SHUFP, DL, VT, LowV, HighV,
8460 getV4X86ShuffleImm8ForMask(NewMask, DL, DAG));
8463 /// \brief Lower 4-lane 32-bit floating point shuffles.
8465 /// Uses instructions exclusively from the floating point unit to minimize
8466 /// domain crossing penalties, as these are sufficient to implement all v4f32
8468 static SDValue lowerV4F32VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
8469 const X86Subtarget *Subtarget,
8470 SelectionDAG &DAG) {
8472 assert(Op.getSimpleValueType() == MVT::v4f32 && "Bad shuffle type!");
8473 assert(V1.getSimpleValueType() == MVT::v4f32 && "Bad operand type!");
8474 assert(V2.getSimpleValueType() == MVT::v4f32 && "Bad operand type!");
8475 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
8476 ArrayRef<int> Mask = SVOp->getMask();
8477 assert(Mask.size() == 4 && "Unexpected mask size for v4 shuffle!");
8480 std::count_if(Mask.begin(), Mask.end(), [](int M) { return M >= 4; });
8482 if (NumV2Elements == 0) {
8483 // Check for being able to broadcast a single element.
8484 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v4f32, V1,
8485 Mask, Subtarget, DAG))
8488 // Use even/odd duplicate instructions for masks that match their pattern.
8489 if (Subtarget->hasSSE3()) {
8490 if (isShuffleEquivalent(V1, V2, Mask, {0, 0, 2, 2}))
8491 return DAG.getNode(X86ISD::MOVSLDUP, DL, MVT::v4f32, V1);
8492 if (isShuffleEquivalent(V1, V2, Mask, {1, 1, 3, 3}))
8493 return DAG.getNode(X86ISD::MOVSHDUP, DL, MVT::v4f32, V1);
8496 if (Subtarget->hasAVX()) {
8497 // If we have AVX, we can use VPERMILPS which will allow folding a load
8498 // into the shuffle.
8499 return DAG.getNode(X86ISD::VPERMILPI, DL, MVT::v4f32, V1,
8500 getV4X86ShuffleImm8ForMask(Mask, DL, DAG));
8503 // Otherwise, use a straight shuffle of a single input vector. We pass the
8504 // input vector to both operands to simulate this with a SHUFPS.
8505 return DAG.getNode(X86ISD::SHUFP, DL, MVT::v4f32, V1, V1,
8506 getV4X86ShuffleImm8ForMask(Mask, DL, DAG));
8509 // There are special ways we can lower some single-element blends. However, we
8510 // have custom ways we can lower more complex single-element blends below that
8511 // we defer to if both this and BLENDPS fail to match, so restrict this to
8512 // when the V2 input is targeting element 0 of the mask -- that is the fast
8514 if (NumV2Elements == 1 && Mask[0] >= 4)
8515 if (SDValue V = lowerVectorShuffleAsElementInsertion(DL, MVT::v4f32, V1, V2,
8516 Mask, Subtarget, DAG))
8519 if (Subtarget->hasSSE41()) {
8520 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v4f32, V1, V2, Mask,
8524 // Use INSERTPS if we can complete the shuffle efficiently.
8525 if (SDValue V = lowerVectorShuffleAsInsertPS(Op, V1, V2, Mask, DAG))
8528 if (!isSingleSHUFPSMask(Mask))
8529 if (SDValue BlendPerm = lowerVectorShuffleAsBlendAndPermute(
8530 DL, MVT::v4f32, V1, V2, Mask, DAG))
8534 // Use dedicated unpack instructions for masks that match their pattern.
8536 lowerVectorShuffleWithUNPCK(DL, MVT::v4f32, Mask, V1, V2, DAG))
8539 // Otherwise fall back to a SHUFPS lowering strategy.
8540 return lowerVectorShuffleWithSHUFPS(DL, MVT::v4f32, Mask, V1, V2, DAG);
8543 /// \brief Lower 4-lane i32 vector shuffles.
8545 /// We try to handle these with integer-domain shuffles where we can, but for
8546 /// blends we use the floating point domain blend instructions.
8547 static SDValue lowerV4I32VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
8548 const X86Subtarget *Subtarget,
8549 SelectionDAG &DAG) {
8551 assert(Op.getSimpleValueType() == MVT::v4i32 && "Bad shuffle type!");
8552 assert(V1.getSimpleValueType() == MVT::v4i32 && "Bad operand type!");
8553 assert(V2.getSimpleValueType() == MVT::v4i32 && "Bad operand type!");
8554 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
8555 ArrayRef<int> Mask = SVOp->getMask();
8556 assert(Mask.size() == 4 && "Unexpected mask size for v4 shuffle!");
8558 // Whenever we can lower this as a zext, that instruction is strictly faster
8559 // than any alternative. It also allows us to fold memory operands into the
8560 // shuffle in many cases.
8561 if (SDValue ZExt = lowerVectorShuffleAsZeroOrAnyExtend(DL, MVT::v4i32, V1, V2,
8562 Mask, Subtarget, DAG))
8566 std::count_if(Mask.begin(), Mask.end(), [](int M) { return M >= 4; });
8568 if (NumV2Elements == 0) {
8569 // Check for being able to broadcast a single element.
8570 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v4i32, V1,
8571 Mask, Subtarget, DAG))
8574 // Straight shuffle of a single input vector. For everything from SSE2
8575 // onward this has a single fast instruction with no scary immediates.
8576 // We coerce the shuffle pattern to be compatible with UNPCK instructions
8577 // but we aren't actually going to use the UNPCK instruction because doing
8578 // so prevents folding a load into this instruction or making a copy.
8579 const int UnpackLoMask[] = {0, 0, 1, 1};
8580 const int UnpackHiMask[] = {2, 2, 3, 3};
8581 if (isShuffleEquivalent(V1, V2, Mask, {0, 0, 1, 1}))
8582 Mask = UnpackLoMask;
8583 else if (isShuffleEquivalent(V1, V2, Mask, {2, 2, 3, 3}))
8584 Mask = UnpackHiMask;
8586 return DAG.getNode(X86ISD::PSHUFD, DL, MVT::v4i32, V1,
8587 getV4X86ShuffleImm8ForMask(Mask, DL, DAG));
8590 // Try to use shift instructions.
8592 lowerVectorShuffleAsShift(DL, MVT::v4i32, V1, V2, Mask, DAG))
8595 // There are special ways we can lower some single-element blends.
8596 if (NumV2Elements == 1)
8597 if (SDValue V = lowerVectorShuffleAsElementInsertion(DL, MVT::v4i32, V1, V2,
8598 Mask, Subtarget, DAG))
8601 // We have different paths for blend lowering, but they all must use the
8602 // *exact* same predicate.
8603 bool IsBlendSupported = Subtarget->hasSSE41();
8604 if (IsBlendSupported)
8605 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v4i32, V1, V2, Mask,
8609 if (SDValue Masked =
8610 lowerVectorShuffleAsBitMask(DL, MVT::v4i32, V1, V2, Mask, DAG))
8613 // Use dedicated unpack instructions for masks that match their pattern.
8615 lowerVectorShuffleWithUNPCK(DL, MVT::v4i32, Mask, V1, V2, DAG))
8618 // Try to use byte rotation instructions.
8619 // Its more profitable for pre-SSSE3 to use shuffles/unpacks.
8620 if (Subtarget->hasSSSE3())
8621 if (SDValue Rotate = lowerVectorShuffleAsByteRotate(
8622 DL, MVT::v4i32, V1, V2, Mask, Subtarget, DAG))
8625 // If we have direct support for blends, we should lower by decomposing into
8626 // a permute. That will be faster than the domain cross.
8627 if (IsBlendSupported)
8628 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v4i32, V1, V2,
8631 // Try to lower by permuting the inputs into an unpack instruction.
8632 if (SDValue Unpack = lowerVectorShuffleAsPermuteAndUnpack(DL, MVT::v4i32, V1,
8636 // We implement this with SHUFPS because it can blend from two vectors.
8637 // Because we're going to eventually use SHUFPS, we use SHUFPS even to build
8638 // up the inputs, bypassing domain shift penalties that we would encur if we
8639 // directly used PSHUFD on Nehalem and older. For newer chips, this isn't
8641 return DAG.getBitcast(
8643 DAG.getVectorShuffle(MVT::v4f32, DL, DAG.getBitcast(MVT::v4f32, V1),
8644 DAG.getBitcast(MVT::v4f32, V2), Mask));
8647 /// \brief Lowering of single-input v8i16 shuffles is the cornerstone of SSE2
8648 /// shuffle lowering, and the most complex part.
8650 /// The lowering strategy is to try to form pairs of input lanes which are
8651 /// targeted at the same half of the final vector, and then use a dword shuffle
8652 /// to place them onto the right half, and finally unpack the paired lanes into
8653 /// their final position.
8655 /// The exact breakdown of how to form these dword pairs and align them on the
8656 /// correct sides is really tricky. See the comments within the function for
8657 /// more of the details.
8659 /// This code also handles repeated 128-bit lanes of v8i16 shuffles, but each
8660 /// lane must shuffle the *exact* same way. In fact, you must pass a v8 Mask to
8661 /// this routine for it to work correctly. To shuffle a 256-bit or 512-bit i16
8662 /// vector, form the analogous 128-bit 8-element Mask.
8663 static SDValue lowerV8I16GeneralSingleInputVectorShuffle(
8664 SDLoc DL, MVT VT, SDValue V, MutableArrayRef<int> Mask,
8665 const X86Subtarget *Subtarget, SelectionDAG &DAG) {
8666 assert(VT.getScalarType() == MVT::i16 && "Bad input type!");
8667 MVT PSHUFDVT = MVT::getVectorVT(MVT::i32, VT.getVectorNumElements() / 2);
8669 assert(Mask.size() == 8 && "Shuffle mask length doen't match!");
8670 MutableArrayRef<int> LoMask = Mask.slice(0, 4);
8671 MutableArrayRef<int> HiMask = Mask.slice(4, 4);
8673 SmallVector<int, 4> LoInputs;
8674 std::copy_if(LoMask.begin(), LoMask.end(), std::back_inserter(LoInputs),
8675 [](int M) { return M >= 0; });
8676 std::sort(LoInputs.begin(), LoInputs.end());
8677 LoInputs.erase(std::unique(LoInputs.begin(), LoInputs.end()), LoInputs.end());
8678 SmallVector<int, 4> HiInputs;
8679 std::copy_if(HiMask.begin(), HiMask.end(), std::back_inserter(HiInputs),
8680 [](int M) { return M >= 0; });
8681 std::sort(HiInputs.begin(), HiInputs.end());
8682 HiInputs.erase(std::unique(HiInputs.begin(), HiInputs.end()), HiInputs.end());
8684 std::lower_bound(LoInputs.begin(), LoInputs.end(), 4) - LoInputs.begin();
8685 int NumHToL = LoInputs.size() - NumLToL;
8687 std::lower_bound(HiInputs.begin(), HiInputs.end(), 4) - HiInputs.begin();
8688 int NumHToH = HiInputs.size() - NumLToH;
8689 MutableArrayRef<int> LToLInputs(LoInputs.data(), NumLToL);
8690 MutableArrayRef<int> LToHInputs(HiInputs.data(), NumLToH);
8691 MutableArrayRef<int> HToLInputs(LoInputs.data() + NumLToL, NumHToL);
8692 MutableArrayRef<int> HToHInputs(HiInputs.data() + NumLToH, NumHToH);
8694 // Simplify the 1-into-3 and 3-into-1 cases with a single pshufd. For all
8695 // such inputs we can swap two of the dwords across the half mark and end up
8696 // with <=2 inputs to each half in each half. Once there, we can fall through
8697 // to the generic code below. For example:
8699 // Input: [a, b, c, d, e, f, g, h] -PSHUFD[0,2,1,3]-> [a, b, e, f, c, d, g, h]
8700 // Mask: [0, 1, 2, 7, 4, 5, 6, 3] -----------------> [0, 1, 4, 7, 2, 3, 6, 5]
8702 // However in some very rare cases we have a 1-into-3 or 3-into-1 on one half
8703 // and an existing 2-into-2 on the other half. In this case we may have to
8704 // pre-shuffle the 2-into-2 half to avoid turning it into a 3-into-1 or
8705 // 1-into-3 which could cause us to cycle endlessly fixing each side in turn.
8706 // Fortunately, we don't have to handle anything but a 2-into-2 pattern
8707 // because any other situation (including a 3-into-1 or 1-into-3 in the other
8708 // half than the one we target for fixing) will be fixed when we re-enter this
8709 // path. We will also combine away any sequence of PSHUFD instructions that
8710 // result into a single instruction. Here is an example of the tricky case:
8712 // Input: [a, b, c, d, e, f, g, h] -PSHUFD[0,2,1,3]-> [a, b, e, f, c, d, g, h]
8713 // Mask: [3, 7, 1, 0, 2, 7, 3, 5] -THIS-IS-BAD!!!!-> [5, 7, 1, 0, 4, 7, 5, 3]
8715 // This now has a 1-into-3 in the high half! Instead, we do two shuffles:
8717 // Input: [a, b, c, d, e, f, g, h] PSHUFHW[0,2,1,3]-> [a, b, c, d, e, g, f, h]
8718 // Mask: [3, 7, 1, 0, 2, 7, 3, 5] -----------------> [3, 7, 1, 0, 2, 7, 3, 6]
8720 // Input: [a, b, c, d, e, g, f, h] -PSHUFD[0,2,1,3]-> [a, b, e, g, c, d, f, h]
8721 // Mask: [3, 7, 1, 0, 2, 7, 3, 6] -----------------> [5, 7, 1, 0, 4, 7, 5, 6]
8723 // The result is fine to be handled by the generic logic.
8724 auto balanceSides = [&](ArrayRef<int> AToAInputs, ArrayRef<int> BToAInputs,
8725 ArrayRef<int> BToBInputs, ArrayRef<int> AToBInputs,
8726 int AOffset, int BOffset) {
8727 assert((AToAInputs.size() == 3 || AToAInputs.size() == 1) &&
8728 "Must call this with A having 3 or 1 inputs from the A half.");
8729 assert((BToAInputs.size() == 1 || BToAInputs.size() == 3) &&
8730 "Must call this with B having 1 or 3 inputs from the B half.");
8731 assert(AToAInputs.size() + BToAInputs.size() == 4 &&
8732 "Must call this with either 3:1 or 1:3 inputs (summing to 4).");
8734 bool ThreeAInputs = AToAInputs.size() == 3;
8736 // Compute the index of dword with only one word among the three inputs in
8737 // a half by taking the sum of the half with three inputs and subtracting
8738 // the sum of the actual three inputs. The difference is the remaining
8741 int &TripleDWord = ThreeAInputs ? ADWord : BDWord;
8742 int &OneInputDWord = ThreeAInputs ? BDWord : ADWord;
8743 int TripleInputOffset = ThreeAInputs ? AOffset : BOffset;
8744 ArrayRef<int> TripleInputs = ThreeAInputs ? AToAInputs : BToAInputs;
8745 int OneInput = ThreeAInputs ? BToAInputs[0] : AToAInputs[0];
8746 int TripleInputSum = 0 + 1 + 2 + 3 + (4 * TripleInputOffset);
8747 int TripleNonInputIdx =
8748 TripleInputSum - std::accumulate(TripleInputs.begin(), TripleInputs.end(), 0);
8749 TripleDWord = TripleNonInputIdx / 2;
8751 // We use xor with one to compute the adjacent DWord to whichever one the
8753 OneInputDWord = (OneInput / 2) ^ 1;
8755 // Check for one tricky case: We're fixing a 3<-1 or a 1<-3 shuffle for AToA
8756 // and BToA inputs. If there is also such a problem with the BToB and AToB
8757 // inputs, we don't try to fix it necessarily -- we'll recurse and see it in
8758 // the next pass. However, if we have a 2<-2 in the BToB and AToB inputs, it
8759 // is essential that we don't *create* a 3<-1 as then we might oscillate.
8760 if (BToBInputs.size() == 2 && AToBInputs.size() == 2) {
8761 // Compute how many inputs will be flipped by swapping these DWords. We
8763 // to balance this to ensure we don't form a 3-1 shuffle in the other
8765 int NumFlippedAToBInputs =
8766 std::count(AToBInputs.begin(), AToBInputs.end(), 2 * ADWord) +
8767 std::count(AToBInputs.begin(), AToBInputs.end(), 2 * ADWord + 1);
8768 int NumFlippedBToBInputs =
8769 std::count(BToBInputs.begin(), BToBInputs.end(), 2 * BDWord) +
8770 std::count(BToBInputs.begin(), BToBInputs.end(), 2 * BDWord + 1);
8771 if ((NumFlippedAToBInputs == 1 &&
8772 (NumFlippedBToBInputs == 0 || NumFlippedBToBInputs == 2)) ||
8773 (NumFlippedBToBInputs == 1 &&
8774 (NumFlippedAToBInputs == 0 || NumFlippedAToBInputs == 2))) {
8775 // We choose whether to fix the A half or B half based on whether that
8776 // half has zero flipped inputs. At zero, we may not be able to fix it
8777 // with that half. We also bias towards fixing the B half because that
8778 // will more commonly be the high half, and we have to bias one way.
8779 auto FixFlippedInputs = [&V, &DL, &Mask, &DAG](int PinnedIdx, int DWord,
8780 ArrayRef<int> Inputs) {
8781 int FixIdx = PinnedIdx ^ 1; // The adjacent slot to the pinned slot.
8782 bool IsFixIdxInput = std::find(Inputs.begin(), Inputs.end(),
8783 PinnedIdx ^ 1) != Inputs.end();
8784 // Determine whether the free index is in the flipped dword or the
8785 // unflipped dword based on where the pinned index is. We use this bit
8786 // in an xor to conditionally select the adjacent dword.
8787 int FixFreeIdx = 2 * (DWord ^ (PinnedIdx / 2 == DWord));
8788 bool IsFixFreeIdxInput = std::find(Inputs.begin(), Inputs.end(),
8789 FixFreeIdx) != Inputs.end();
8790 if (IsFixIdxInput == IsFixFreeIdxInput)
8792 IsFixFreeIdxInput = std::find(Inputs.begin(), Inputs.end(),
8793 FixFreeIdx) != Inputs.end();
8794 assert(IsFixIdxInput != IsFixFreeIdxInput &&
8795 "We need to be changing the number of flipped inputs!");
8796 int PSHUFHalfMask[] = {0, 1, 2, 3};
8797 std::swap(PSHUFHalfMask[FixFreeIdx % 4], PSHUFHalfMask[FixIdx % 4]);
8798 V = DAG.getNode(FixIdx < 4 ? X86ISD::PSHUFLW : X86ISD::PSHUFHW, DL,
8800 getV4X86ShuffleImm8ForMask(PSHUFHalfMask, DL, DAG));
8803 if (M != -1 && M == FixIdx)
8805 else if (M != -1 && M == FixFreeIdx)
8808 if (NumFlippedBToBInputs != 0) {
8810 BToAInputs.size() == 3 ? TripleNonInputIdx : OneInput;
8811 FixFlippedInputs(BPinnedIdx, BDWord, BToBInputs);
8813 assert(NumFlippedAToBInputs != 0 && "Impossible given predicates!");
8814 int APinnedIdx = ThreeAInputs ? TripleNonInputIdx : OneInput;
8815 FixFlippedInputs(APinnedIdx, ADWord, AToBInputs);
8820 int PSHUFDMask[] = {0, 1, 2, 3};
8821 PSHUFDMask[ADWord] = BDWord;
8822 PSHUFDMask[BDWord] = ADWord;
8825 DAG.getNode(X86ISD::PSHUFD, DL, PSHUFDVT, DAG.getBitcast(PSHUFDVT, V),
8826 getV4X86ShuffleImm8ForMask(PSHUFDMask, DL, DAG)));
8828 // Adjust the mask to match the new locations of A and B.
8830 if (M != -1 && M/2 == ADWord)
8831 M = 2 * BDWord + M % 2;
8832 else if (M != -1 && M/2 == BDWord)
8833 M = 2 * ADWord + M % 2;
8835 // Recurse back into this routine to re-compute state now that this isn't
8836 // a 3 and 1 problem.
8837 return lowerV8I16GeneralSingleInputVectorShuffle(DL, VT, V, Mask, Subtarget,
8840 if ((NumLToL == 3 && NumHToL == 1) || (NumLToL == 1 && NumHToL == 3))
8841 return balanceSides(LToLInputs, HToLInputs, HToHInputs, LToHInputs, 0, 4);
8842 else if ((NumHToH == 3 && NumLToH == 1) || (NumHToH == 1 && NumLToH == 3))
8843 return balanceSides(HToHInputs, LToHInputs, LToLInputs, HToLInputs, 4, 0);
8845 // At this point there are at most two inputs to the low and high halves from
8846 // each half. That means the inputs can always be grouped into dwords and
8847 // those dwords can then be moved to the correct half with a dword shuffle.
8848 // We use at most one low and one high word shuffle to collect these paired
8849 // inputs into dwords, and finally a dword shuffle to place them.
8850 int PSHUFLMask[4] = {-1, -1, -1, -1};
8851 int PSHUFHMask[4] = {-1, -1, -1, -1};
8852 int PSHUFDMask[4] = {-1, -1, -1, -1};
8854 // First fix the masks for all the inputs that are staying in their
8855 // original halves. This will then dictate the targets of the cross-half
8857 auto fixInPlaceInputs =
8858 [&PSHUFDMask](ArrayRef<int> InPlaceInputs, ArrayRef<int> IncomingInputs,
8859 MutableArrayRef<int> SourceHalfMask,
8860 MutableArrayRef<int> HalfMask, int HalfOffset) {
8861 if (InPlaceInputs.empty())
8863 if (InPlaceInputs.size() == 1) {
8864 SourceHalfMask[InPlaceInputs[0] - HalfOffset] =
8865 InPlaceInputs[0] - HalfOffset;
8866 PSHUFDMask[InPlaceInputs[0] / 2] = InPlaceInputs[0] / 2;
8869 if (IncomingInputs.empty()) {
8870 // Just fix all of the in place inputs.
8871 for (int Input : InPlaceInputs) {
8872 SourceHalfMask[Input - HalfOffset] = Input - HalfOffset;
8873 PSHUFDMask[Input / 2] = Input / 2;
8878 assert(InPlaceInputs.size() == 2 && "Cannot handle 3 or 4 inputs!");
8879 SourceHalfMask[InPlaceInputs[0] - HalfOffset] =
8880 InPlaceInputs[0] - HalfOffset;
8881 // Put the second input next to the first so that they are packed into
8882 // a dword. We find the adjacent index by toggling the low bit.
8883 int AdjIndex = InPlaceInputs[0] ^ 1;
8884 SourceHalfMask[AdjIndex - HalfOffset] = InPlaceInputs[1] - HalfOffset;
8885 std::replace(HalfMask.begin(), HalfMask.end(), InPlaceInputs[1], AdjIndex);
8886 PSHUFDMask[AdjIndex / 2] = AdjIndex / 2;
8888 fixInPlaceInputs(LToLInputs, HToLInputs, PSHUFLMask, LoMask, 0);
8889 fixInPlaceInputs(HToHInputs, LToHInputs, PSHUFHMask, HiMask, 4);
8891 // Now gather the cross-half inputs and place them into a free dword of
8892 // their target half.
8893 // FIXME: This operation could almost certainly be simplified dramatically to
8894 // look more like the 3-1 fixing operation.
8895 auto moveInputsToRightHalf = [&PSHUFDMask](
8896 MutableArrayRef<int> IncomingInputs, ArrayRef<int> ExistingInputs,
8897 MutableArrayRef<int> SourceHalfMask, MutableArrayRef<int> HalfMask,
8898 MutableArrayRef<int> FinalSourceHalfMask, int SourceOffset,
8900 auto isWordClobbered = [](ArrayRef<int> SourceHalfMask, int Word) {
8901 return SourceHalfMask[Word] != -1 && SourceHalfMask[Word] != Word;
8903 auto isDWordClobbered = [&isWordClobbered](ArrayRef<int> SourceHalfMask,
8905 int LowWord = Word & ~1;
8906 int HighWord = Word | 1;
8907 return isWordClobbered(SourceHalfMask, LowWord) ||
8908 isWordClobbered(SourceHalfMask, HighWord);
8911 if (IncomingInputs.empty())
8914 if (ExistingInputs.empty()) {
8915 // Map any dwords with inputs from them into the right half.
8916 for (int Input : IncomingInputs) {
8917 // If the source half mask maps over the inputs, turn those into
8918 // swaps and use the swapped lane.
8919 if (isWordClobbered(SourceHalfMask, Input - SourceOffset)) {
8920 if (SourceHalfMask[SourceHalfMask[Input - SourceOffset]] == -1) {
8921 SourceHalfMask[SourceHalfMask[Input - SourceOffset]] =
8922 Input - SourceOffset;
8923 // We have to swap the uses in our half mask in one sweep.
8924 for (int &M : HalfMask)
8925 if (M == SourceHalfMask[Input - SourceOffset] + SourceOffset)
8927 else if (M == Input)
8928 M = SourceHalfMask[Input - SourceOffset] + SourceOffset;
8930 assert(SourceHalfMask[SourceHalfMask[Input - SourceOffset]] ==
8931 Input - SourceOffset &&
8932 "Previous placement doesn't match!");
8934 // Note that this correctly re-maps both when we do a swap and when
8935 // we observe the other side of the swap above. We rely on that to
8936 // avoid swapping the members of the input list directly.
8937 Input = SourceHalfMask[Input - SourceOffset] + SourceOffset;
8940 // Map the input's dword into the correct half.
8941 if (PSHUFDMask[(Input - SourceOffset + DestOffset) / 2] == -1)
8942 PSHUFDMask[(Input - SourceOffset + DestOffset) / 2] = Input / 2;
8944 assert(PSHUFDMask[(Input - SourceOffset + DestOffset) / 2] ==
8946 "Previous placement doesn't match!");
8949 // And just directly shift any other-half mask elements to be same-half
8950 // as we will have mirrored the dword containing the element into the
8951 // same position within that half.
8952 for (int &M : HalfMask)
8953 if (M >= SourceOffset && M < SourceOffset + 4) {
8954 M = M - SourceOffset + DestOffset;
8955 assert(M >= 0 && "This should never wrap below zero!");
8960 // Ensure we have the input in a viable dword of its current half. This
8961 // is particularly tricky because the original position may be clobbered
8962 // by inputs being moved and *staying* in that half.
8963 if (IncomingInputs.size() == 1) {
8964 if (isWordClobbered(SourceHalfMask, IncomingInputs[0] - SourceOffset)) {
8965 int InputFixed = std::find(std::begin(SourceHalfMask),
8966 std::end(SourceHalfMask), -1) -
8967 std::begin(SourceHalfMask) + SourceOffset;
8968 SourceHalfMask[InputFixed - SourceOffset] =
8969 IncomingInputs[0] - SourceOffset;
8970 std::replace(HalfMask.begin(), HalfMask.end(), IncomingInputs[0],
8972 IncomingInputs[0] = InputFixed;
8974 } else if (IncomingInputs.size() == 2) {
8975 if (IncomingInputs[0] / 2 != IncomingInputs[1] / 2 ||
8976 isDWordClobbered(SourceHalfMask, IncomingInputs[0] - SourceOffset)) {
8977 // We have two non-adjacent or clobbered inputs we need to extract from
8978 // the source half. To do this, we need to map them into some adjacent
8979 // dword slot in the source mask.
8980 int InputsFixed[2] = {IncomingInputs[0] - SourceOffset,
8981 IncomingInputs[1] - SourceOffset};
8983 // If there is a free slot in the source half mask adjacent to one of
8984 // the inputs, place the other input in it. We use (Index XOR 1) to
8985 // compute an adjacent index.
8986 if (!isWordClobbered(SourceHalfMask, InputsFixed[0]) &&
8987 SourceHalfMask[InputsFixed[0] ^ 1] == -1) {
8988 SourceHalfMask[InputsFixed[0]] = InputsFixed[0];
8989 SourceHalfMask[InputsFixed[0] ^ 1] = InputsFixed[1];
8990 InputsFixed[1] = InputsFixed[0] ^ 1;
8991 } else if (!isWordClobbered(SourceHalfMask, InputsFixed[1]) &&
8992 SourceHalfMask[InputsFixed[1] ^ 1] == -1) {
8993 SourceHalfMask[InputsFixed[1]] = InputsFixed[1];
8994 SourceHalfMask[InputsFixed[1] ^ 1] = InputsFixed[0];
8995 InputsFixed[0] = InputsFixed[1] ^ 1;
8996 } else if (SourceHalfMask[2 * ((InputsFixed[0] / 2) ^ 1)] == -1 &&
8997 SourceHalfMask[2 * ((InputsFixed[0] / 2) ^ 1) + 1] == -1) {
8998 // The two inputs are in the same DWord but it is clobbered and the
8999 // adjacent DWord isn't used at all. Move both inputs to the free
9001 SourceHalfMask[2 * ((InputsFixed[0] / 2) ^ 1)] = InputsFixed[0];
9002 SourceHalfMask[2 * ((InputsFixed[0] / 2) ^ 1) + 1] = InputsFixed[1];
9003 InputsFixed[0] = 2 * ((InputsFixed[0] / 2) ^ 1);
9004 InputsFixed[1] = 2 * ((InputsFixed[0] / 2) ^ 1) + 1;
9006 // The only way we hit this point is if there is no clobbering
9007 // (because there are no off-half inputs to this half) and there is no
9008 // free slot adjacent to one of the inputs. In this case, we have to
9009 // swap an input with a non-input.
9010 for (int i = 0; i < 4; ++i)
9011 assert((SourceHalfMask[i] == -1 || SourceHalfMask[i] == i) &&
9012 "We can't handle any clobbers here!");
9013 assert(InputsFixed[1] != (InputsFixed[0] ^ 1) &&
9014 "Cannot have adjacent inputs here!");
9016 SourceHalfMask[InputsFixed[0] ^ 1] = InputsFixed[1];
9017 SourceHalfMask[InputsFixed[1]] = InputsFixed[0] ^ 1;
9019 // We also have to update the final source mask in this case because
9020 // it may need to undo the above swap.
9021 for (int &M : FinalSourceHalfMask)
9022 if (M == (InputsFixed[0] ^ 1) + SourceOffset)
9023 M = InputsFixed[1] + SourceOffset;
9024 else if (M == InputsFixed[1] + SourceOffset)
9025 M = (InputsFixed[0] ^ 1) + SourceOffset;
9027 InputsFixed[1] = InputsFixed[0] ^ 1;
9030 // Point everything at the fixed inputs.
9031 for (int &M : HalfMask)
9032 if (M == IncomingInputs[0])
9033 M = InputsFixed[0] + SourceOffset;
9034 else if (M == IncomingInputs[1])
9035 M = InputsFixed[1] + SourceOffset;
9037 IncomingInputs[0] = InputsFixed[0] + SourceOffset;
9038 IncomingInputs[1] = InputsFixed[1] + SourceOffset;
9041 llvm_unreachable("Unhandled input size!");
9044 // Now hoist the DWord down to the right half.
9045 int FreeDWord = (PSHUFDMask[DestOffset / 2] == -1 ? 0 : 1) + DestOffset / 2;
9046 assert(PSHUFDMask[FreeDWord] == -1 && "DWord not free");
9047 PSHUFDMask[FreeDWord] = IncomingInputs[0] / 2;
9048 for (int &M : HalfMask)
9049 for (int Input : IncomingInputs)
9051 M = FreeDWord * 2 + Input % 2;
9053 moveInputsToRightHalf(HToLInputs, LToLInputs, PSHUFHMask, LoMask, HiMask,
9054 /*SourceOffset*/ 4, /*DestOffset*/ 0);
9055 moveInputsToRightHalf(LToHInputs, HToHInputs, PSHUFLMask, HiMask, LoMask,
9056 /*SourceOffset*/ 0, /*DestOffset*/ 4);
9058 // Now enact all the shuffles we've computed to move the inputs into their
9060 if (!isNoopShuffleMask(PSHUFLMask))
9061 V = DAG.getNode(X86ISD::PSHUFLW, DL, VT, V,
9062 getV4X86ShuffleImm8ForMask(PSHUFLMask, DL, DAG));
9063 if (!isNoopShuffleMask(PSHUFHMask))
9064 V = DAG.getNode(X86ISD::PSHUFHW, DL, VT, V,
9065 getV4X86ShuffleImm8ForMask(PSHUFHMask, DL, DAG));
9066 if (!isNoopShuffleMask(PSHUFDMask))
9069 DAG.getNode(X86ISD::PSHUFD, DL, PSHUFDVT, DAG.getBitcast(PSHUFDVT, V),
9070 getV4X86ShuffleImm8ForMask(PSHUFDMask, DL, DAG)));
9072 // At this point, each half should contain all its inputs, and we can then
9073 // just shuffle them into their final position.
9074 assert(std::count_if(LoMask.begin(), LoMask.end(),
9075 [](int M) { return M >= 4; }) == 0 &&
9076 "Failed to lift all the high half inputs to the low mask!");
9077 assert(std::count_if(HiMask.begin(), HiMask.end(),
9078 [](int M) { return M >= 0 && M < 4; }) == 0 &&
9079 "Failed to lift all the low half inputs to the high mask!");
9081 // Do a half shuffle for the low mask.
9082 if (!isNoopShuffleMask(LoMask))
9083 V = DAG.getNode(X86ISD::PSHUFLW, DL, VT, V,
9084 getV4X86ShuffleImm8ForMask(LoMask, DL, DAG));
9086 // Do a half shuffle with the high mask after shifting its values down.
9087 for (int &M : HiMask)
9090 if (!isNoopShuffleMask(HiMask))
9091 V = DAG.getNode(X86ISD::PSHUFHW, DL, VT, V,
9092 getV4X86ShuffleImm8ForMask(HiMask, DL, DAG));
9097 /// \brief Helper to form a PSHUFB-based shuffle+blend.
9098 static SDValue lowerVectorShuffleAsPSHUFB(SDLoc DL, MVT VT, SDValue V1,
9099 SDValue V2, ArrayRef<int> Mask,
9100 SelectionDAG &DAG, bool &V1InUse,
9102 SmallBitVector Zeroable = computeZeroableShuffleElements(Mask, V1, V2);
9108 int Size = Mask.size();
9109 int Scale = 16 / Size;
9110 for (int i = 0; i < 16; ++i) {
9111 if (Mask[i / Scale] == -1) {
9112 V1Mask[i] = V2Mask[i] = DAG.getUNDEF(MVT::i8);
9114 const int ZeroMask = 0x80;
9115 int V1Idx = Mask[i / Scale] < Size ? Mask[i / Scale] * Scale + i % Scale
9117 int V2Idx = Mask[i / Scale] < Size
9119 : (Mask[i / Scale] - Size) * Scale + i % Scale;
9120 if (Zeroable[i / Scale])
9121 V1Idx = V2Idx = ZeroMask;
9122 V1Mask[i] = DAG.getConstant(V1Idx, DL, MVT::i8);
9123 V2Mask[i] = DAG.getConstant(V2Idx, DL, MVT::i8);
9124 V1InUse |= (ZeroMask != V1Idx);
9125 V2InUse |= (ZeroMask != V2Idx);
9130 V1 = DAG.getNode(X86ISD::PSHUFB, DL, MVT::v16i8,
9131 DAG.getBitcast(MVT::v16i8, V1),
9132 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v16i8, V1Mask));
9134 V2 = DAG.getNode(X86ISD::PSHUFB, DL, MVT::v16i8,
9135 DAG.getBitcast(MVT::v16i8, V2),
9136 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v16i8, V2Mask));
9138 // If we need shuffled inputs from both, blend the two.
9140 if (V1InUse && V2InUse)
9141 V = DAG.getNode(ISD::OR, DL, MVT::v16i8, V1, V2);
9143 V = V1InUse ? V1 : V2;
9145 // Cast the result back to the correct type.
9146 return DAG.getBitcast(VT, V);
9149 /// \brief Generic lowering of 8-lane i16 shuffles.
9151 /// This handles both single-input shuffles and combined shuffle/blends with
9152 /// two inputs. The single input shuffles are immediately delegated to
9153 /// a dedicated lowering routine.
9155 /// The blends are lowered in one of three fundamental ways. If there are few
9156 /// enough inputs, it delegates to a basic UNPCK-based strategy. If the shuffle
9157 /// of the input is significantly cheaper when lowered as an interleaving of
9158 /// the two inputs, try to interleave them. Otherwise, blend the low and high
9159 /// halves of the inputs separately (making them have relatively few inputs)
9160 /// and then concatenate them.
9161 static SDValue lowerV8I16VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
9162 const X86Subtarget *Subtarget,
9163 SelectionDAG &DAG) {
9165 assert(Op.getSimpleValueType() == MVT::v8i16 && "Bad shuffle type!");
9166 assert(V1.getSimpleValueType() == MVT::v8i16 && "Bad operand type!");
9167 assert(V2.getSimpleValueType() == MVT::v8i16 && "Bad operand type!");
9168 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
9169 ArrayRef<int> OrigMask = SVOp->getMask();
9170 int MaskStorage[8] = {OrigMask[0], OrigMask[1], OrigMask[2], OrigMask[3],
9171 OrigMask[4], OrigMask[5], OrigMask[6], OrigMask[7]};
9172 MutableArrayRef<int> Mask(MaskStorage);
9174 assert(Mask.size() == 8 && "Unexpected mask size for v8 shuffle!");
9176 // Whenever we can lower this as a zext, that instruction is strictly faster
9177 // than any alternative.
9178 if (SDValue ZExt = lowerVectorShuffleAsZeroOrAnyExtend(
9179 DL, MVT::v8i16, V1, V2, OrigMask, Subtarget, DAG))
9182 auto isV1 = [](int M) { return M >= 0 && M < 8; };
9184 auto isV2 = [](int M) { return M >= 8; };
9186 int NumV2Inputs = std::count_if(Mask.begin(), Mask.end(), isV2);
9188 if (NumV2Inputs == 0) {
9189 // Check for being able to broadcast a single element.
9190 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v8i16, V1,
9191 Mask, Subtarget, DAG))
9194 // Try to use shift instructions.
9196 lowerVectorShuffleAsShift(DL, MVT::v8i16, V1, V1, Mask, DAG))
9199 // Use dedicated unpack instructions for masks that match their pattern.
9201 lowerVectorShuffleWithUNPCK(DL, MVT::v8i16, Mask, V1, V2, DAG))
9204 // Try to use byte rotation instructions.
9205 if (SDValue Rotate = lowerVectorShuffleAsByteRotate(DL, MVT::v8i16, V1, V1,
9206 Mask, Subtarget, DAG))
9209 return lowerV8I16GeneralSingleInputVectorShuffle(DL, MVT::v8i16, V1, Mask,
9213 assert(std::any_of(Mask.begin(), Mask.end(), isV1) &&
9214 "All single-input shuffles should be canonicalized to be V1-input "
9217 // Try to use shift instructions.
9219 lowerVectorShuffleAsShift(DL, MVT::v8i16, V1, V2, Mask, DAG))
9222 // See if we can use SSE4A Extraction / Insertion.
9223 if (Subtarget->hasSSE4A())
9224 if (SDValue V = lowerVectorShuffleWithSSE4A(DL, MVT::v8i16, V1, V2, Mask, DAG))
9227 // There are special ways we can lower some single-element blends.
9228 if (NumV2Inputs == 1)
9229 if (SDValue V = lowerVectorShuffleAsElementInsertion(DL, MVT::v8i16, V1, V2,
9230 Mask, Subtarget, DAG))
9233 // We have different paths for blend lowering, but they all must use the
9234 // *exact* same predicate.
9235 bool IsBlendSupported = Subtarget->hasSSE41();
9236 if (IsBlendSupported)
9237 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v8i16, V1, V2, Mask,
9241 if (SDValue Masked =
9242 lowerVectorShuffleAsBitMask(DL, MVT::v8i16, V1, V2, Mask, DAG))
9245 // Use dedicated unpack instructions for masks that match their pattern.
9247 lowerVectorShuffleWithUNPCK(DL, MVT::v8i16, Mask, V1, V2, DAG))
9250 // Try to use byte rotation instructions.
9251 if (SDValue Rotate = lowerVectorShuffleAsByteRotate(
9252 DL, MVT::v8i16, V1, V2, Mask, Subtarget, DAG))
9255 if (SDValue BitBlend =
9256 lowerVectorShuffleAsBitBlend(DL, MVT::v8i16, V1, V2, Mask, DAG))
9259 if (SDValue Unpack = lowerVectorShuffleAsPermuteAndUnpack(DL, MVT::v8i16, V1,
9263 // If we can't directly blend but can use PSHUFB, that will be better as it
9264 // can both shuffle and set up the inefficient blend.
9265 if (!IsBlendSupported && Subtarget->hasSSSE3()) {
9266 bool V1InUse, V2InUse;
9267 return lowerVectorShuffleAsPSHUFB(DL, MVT::v8i16, V1, V2, Mask, DAG,
9271 // We can always bit-blend if we have to so the fallback strategy is to
9272 // decompose into single-input permutes and blends.
9273 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v8i16, V1, V2,
9277 /// \brief Check whether a compaction lowering can be done by dropping even
9278 /// elements and compute how many times even elements must be dropped.
9280 /// This handles shuffles which take every Nth element where N is a power of
9281 /// two. Example shuffle masks:
9283 /// N = 1: 0, 2, 4, 6, 8, 10, 12, 14, 0, 2, 4, 6, 8, 10, 12, 14
9284 /// N = 1: 0, 2, 4, 6, 8, 10, 12, 14, 16, 18, 20, 22, 24, 26, 28, 30
9285 /// N = 2: 0, 4, 8, 12, 0, 4, 8, 12, 0, 4, 8, 12, 0, 4, 8, 12
9286 /// N = 2: 0, 4, 8, 12, 16, 20, 24, 28, 0, 4, 8, 12, 16, 20, 24, 28
9287 /// N = 3: 0, 8, 0, 8, 0, 8, 0, 8, 0, 8, 0, 8, 0, 8, 0, 8
9288 /// N = 3: 0, 8, 16, 24, 0, 8, 16, 24, 0, 8, 16, 24, 0, 8, 16, 24
9290 /// Any of these lanes can of course be undef.
9292 /// This routine only supports N <= 3.
9293 /// FIXME: Evaluate whether either AVX or AVX-512 have any opportunities here
9296 /// \returns N above, or the number of times even elements must be dropped if
9297 /// there is such a number. Otherwise returns zero.
9298 static int canLowerByDroppingEvenElements(ArrayRef<int> Mask) {
9299 // Figure out whether we're looping over two inputs or just one.
9300 bool IsSingleInput = isSingleInputShuffleMask(Mask);
9302 // The modulus for the shuffle vector entries is based on whether this is
9303 // a single input or not.
9304 int ShuffleModulus = Mask.size() * (IsSingleInput ? 1 : 2);
9305 assert(isPowerOf2_32((uint32_t)ShuffleModulus) &&
9306 "We should only be called with masks with a power-of-2 size!");
9308 uint64_t ModMask = (uint64_t)ShuffleModulus - 1;
9310 // We track whether the input is viable for all power-of-2 strides 2^1, 2^2,
9311 // and 2^3 simultaneously. This is because we may have ambiguity with
9312 // partially undef inputs.
9313 bool ViableForN[3] = {true, true, true};
9315 for (int i = 0, e = Mask.size(); i < e; ++i) {
9316 // Ignore undef lanes, we'll optimistically collapse them to the pattern we
9321 bool IsAnyViable = false;
9322 for (unsigned j = 0; j != array_lengthof(ViableForN); ++j)
9323 if (ViableForN[j]) {
9326 // The shuffle mask must be equal to (i * 2^N) % M.
9327 if ((uint64_t)Mask[i] == (((uint64_t)i << N) & ModMask))
9330 ViableForN[j] = false;
9332 // Early exit if we exhaust the possible powers of two.
9337 for (unsigned j = 0; j != array_lengthof(ViableForN); ++j)
9341 // Return 0 as there is no viable power of two.
9345 /// \brief Generic lowering of v16i8 shuffles.
9347 /// This is a hybrid strategy to lower v16i8 vectors. It first attempts to
9348 /// detect any complexity reducing interleaving. If that doesn't help, it uses
9349 /// UNPCK to spread the i8 elements across two i16-element vectors, and uses
9350 /// the existing lowering for v8i16 blends on each half, finally PACK-ing them
9352 static SDValue lowerV16I8VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
9353 const X86Subtarget *Subtarget,
9354 SelectionDAG &DAG) {
9356 assert(Op.getSimpleValueType() == MVT::v16i8 && "Bad shuffle type!");
9357 assert(V1.getSimpleValueType() == MVT::v16i8 && "Bad operand type!");
9358 assert(V2.getSimpleValueType() == MVT::v16i8 && "Bad operand type!");
9359 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
9360 ArrayRef<int> Mask = SVOp->getMask();
9361 assert(Mask.size() == 16 && "Unexpected mask size for v16 shuffle!");
9363 // Try to use shift instructions.
9365 lowerVectorShuffleAsShift(DL, MVT::v16i8, V1, V2, Mask, DAG))
9368 // Try to use byte rotation instructions.
9369 if (SDValue Rotate = lowerVectorShuffleAsByteRotate(
9370 DL, MVT::v16i8, V1, V2, Mask, Subtarget, DAG))
9373 // Try to use a zext lowering.
9374 if (SDValue ZExt = lowerVectorShuffleAsZeroOrAnyExtend(
9375 DL, MVT::v16i8, V1, V2, Mask, Subtarget, DAG))
9378 // See if we can use SSE4A Extraction / Insertion.
9379 if (Subtarget->hasSSE4A())
9380 if (SDValue V = lowerVectorShuffleWithSSE4A(DL, MVT::v16i8, V1, V2, Mask, DAG))
9384 std::count_if(Mask.begin(), Mask.end(), [](int M) { return M >= 16; });
9386 // For single-input shuffles, there are some nicer lowering tricks we can use.
9387 if (NumV2Elements == 0) {
9388 // Check for being able to broadcast a single element.
9389 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v16i8, V1,
9390 Mask, Subtarget, DAG))
9393 // Check whether we can widen this to an i16 shuffle by duplicating bytes.
9394 // Notably, this handles splat and partial-splat shuffles more efficiently.
9395 // However, it only makes sense if the pre-duplication shuffle simplifies
9396 // things significantly. Currently, this means we need to be able to
9397 // express the pre-duplication shuffle as an i16 shuffle.
9399 // FIXME: We should check for other patterns which can be widened into an
9400 // i16 shuffle as well.
9401 auto canWidenViaDuplication = [](ArrayRef<int> Mask) {
9402 for (int i = 0; i < 16; i += 2)
9403 if (Mask[i] != -1 && Mask[i + 1] != -1 && Mask[i] != Mask[i + 1])
9408 auto tryToWidenViaDuplication = [&]() -> SDValue {
9409 if (!canWidenViaDuplication(Mask))
9411 SmallVector<int, 4> LoInputs;
9412 std::copy_if(Mask.begin(), Mask.end(), std::back_inserter(LoInputs),
9413 [](int M) { return M >= 0 && M < 8; });
9414 std::sort(LoInputs.begin(), LoInputs.end());
9415 LoInputs.erase(std::unique(LoInputs.begin(), LoInputs.end()),
9417 SmallVector<int, 4> HiInputs;
9418 std::copy_if(Mask.begin(), Mask.end(), std::back_inserter(HiInputs),
9419 [](int M) { return M >= 8; });
9420 std::sort(HiInputs.begin(), HiInputs.end());
9421 HiInputs.erase(std::unique(HiInputs.begin(), HiInputs.end()),
9424 bool TargetLo = LoInputs.size() >= HiInputs.size();
9425 ArrayRef<int> InPlaceInputs = TargetLo ? LoInputs : HiInputs;
9426 ArrayRef<int> MovingInputs = TargetLo ? HiInputs : LoInputs;
9428 int PreDupI16Shuffle[] = {-1, -1, -1, -1, -1, -1, -1, -1};
9429 SmallDenseMap<int, int, 8> LaneMap;
9430 for (int I : InPlaceInputs) {
9431 PreDupI16Shuffle[I/2] = I/2;
9434 int j = TargetLo ? 0 : 4, je = j + 4;
9435 for (int i = 0, ie = MovingInputs.size(); i < ie; ++i) {
9436 // Check if j is already a shuffle of this input. This happens when
9437 // there are two adjacent bytes after we move the low one.
9438 if (PreDupI16Shuffle[j] != MovingInputs[i] / 2) {
9439 // If we haven't yet mapped the input, search for a slot into which
9441 while (j < je && PreDupI16Shuffle[j] != -1)
9445 // We can't place the inputs into a single half with a simple i16 shuffle, so bail.
9448 // Map this input with the i16 shuffle.
9449 PreDupI16Shuffle[j] = MovingInputs[i] / 2;
9452 // Update the lane map based on the mapping we ended up with.
9453 LaneMap[MovingInputs[i]] = 2 * j + MovingInputs[i] % 2;
9455 V1 = DAG.getBitcast(
9457 DAG.getVectorShuffle(MVT::v8i16, DL, DAG.getBitcast(MVT::v8i16, V1),
9458 DAG.getUNDEF(MVT::v8i16), PreDupI16Shuffle));
9460 // Unpack the bytes to form the i16s that will be shuffled into place.
9461 V1 = DAG.getNode(TargetLo ? X86ISD::UNPCKL : X86ISD::UNPCKH, DL,
9462 MVT::v16i8, V1, V1);
9464 int PostDupI16Shuffle[8] = {-1, -1, -1, -1, -1, -1, -1, -1};
9465 for (int i = 0; i < 16; ++i)
9466 if (Mask[i] != -1) {
9467 int MappedMask = LaneMap[Mask[i]] - (TargetLo ? 0 : 8);
9468 assert(MappedMask < 8 && "Invalid v8 shuffle mask!");
9469 if (PostDupI16Shuffle[i / 2] == -1)
9470 PostDupI16Shuffle[i / 2] = MappedMask;
9472 assert(PostDupI16Shuffle[i / 2] == MappedMask &&
9473 "Conflicting entrties in the original shuffle!");
9475 return DAG.getBitcast(
9477 DAG.getVectorShuffle(MVT::v8i16, DL, DAG.getBitcast(MVT::v8i16, V1),
9478 DAG.getUNDEF(MVT::v8i16), PostDupI16Shuffle));
9480 if (SDValue V = tryToWidenViaDuplication())
9484 if (SDValue Masked =
9485 lowerVectorShuffleAsBitMask(DL, MVT::v16i8, V1, V2, Mask, DAG))
9488 // Use dedicated unpack instructions for masks that match their pattern.
9490 lowerVectorShuffleWithUNPCK(DL, MVT::v16i8, Mask, V1, V2, DAG))
9493 // Check for SSSE3 which lets us lower all v16i8 shuffles much more directly
9494 // with PSHUFB. It is important to do this before we attempt to generate any
9495 // blends but after all of the single-input lowerings. If the single input
9496 // lowerings can find an instruction sequence that is faster than a PSHUFB, we
9497 // want to preserve that and we can DAG combine any longer sequences into
9498 // a PSHUFB in the end. But once we start blending from multiple inputs,
9499 // the complexity of DAG combining bad patterns back into PSHUFB is too high,
9500 // and there are *very* few patterns that would actually be faster than the
9501 // PSHUFB approach because of its ability to zero lanes.
9503 // FIXME: The only exceptions to the above are blends which are exact
9504 // interleavings with direct instructions supporting them. We currently don't
9505 // handle those well here.
9506 if (Subtarget->hasSSSE3()) {
9507 bool V1InUse = false;
9508 bool V2InUse = false;
9510 SDValue PSHUFB = lowerVectorShuffleAsPSHUFB(DL, MVT::v16i8, V1, V2, Mask,
9511 DAG, V1InUse, V2InUse);
9513 // If both V1 and V2 are in use and we can use a direct blend or an unpack,
9514 // do so. This avoids using them to handle blends-with-zero which is
9515 // important as a single pshufb is significantly faster for that.
9516 if (V1InUse && V2InUse) {
9517 if (Subtarget->hasSSE41())
9518 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v16i8, V1, V2,
9519 Mask, Subtarget, DAG))
9522 // We can use an unpack to do the blending rather than an or in some
9523 // cases. Even though the or may be (very minorly) more efficient, we
9524 // preference this lowering because there are common cases where part of
9525 // the complexity of the shuffles goes away when we do the final blend as
9527 // FIXME: It might be worth trying to detect if the unpack-feeding
9528 // shuffles will both be pshufb, in which case we shouldn't bother with
9530 if (SDValue Unpack = lowerVectorShuffleAsPermuteAndUnpack(
9531 DL, MVT::v16i8, V1, V2, Mask, DAG))
9538 // There are special ways we can lower some single-element blends.
9539 if (NumV2Elements == 1)
9540 if (SDValue V = lowerVectorShuffleAsElementInsertion(DL, MVT::v16i8, V1, V2,
9541 Mask, Subtarget, DAG))
9544 if (SDValue BitBlend =
9545 lowerVectorShuffleAsBitBlend(DL, MVT::v16i8, V1, V2, Mask, DAG))
9548 // Check whether a compaction lowering can be done. This handles shuffles
9549 // which take every Nth element for some even N. See the helper function for
9552 // We special case these as they can be particularly efficiently handled with
9553 // the PACKUSB instruction on x86 and they show up in common patterns of
9554 // rearranging bytes to truncate wide elements.
9555 if (int NumEvenDrops = canLowerByDroppingEvenElements(Mask)) {
9556 // NumEvenDrops is the power of two stride of the elements. Another way of
9557 // thinking about it is that we need to drop the even elements this many
9558 // times to get the original input.
9559 bool IsSingleInput = isSingleInputShuffleMask(Mask);
9561 // First we need to zero all the dropped bytes.
9562 assert(NumEvenDrops <= 3 &&
9563 "No support for dropping even elements more than 3 times.");
9564 // We use the mask type to pick which bytes are preserved based on how many
9565 // elements are dropped.
9566 MVT MaskVTs[] = { MVT::v8i16, MVT::v4i32, MVT::v2i64 };
9567 SDValue ByteClearMask = DAG.getBitcast(
9568 MVT::v16i8, DAG.getConstant(0xFF, DL, MaskVTs[NumEvenDrops - 1]));
9569 V1 = DAG.getNode(ISD::AND, DL, MVT::v16i8, V1, ByteClearMask);
9571 V2 = DAG.getNode(ISD::AND, DL, MVT::v16i8, V2, ByteClearMask);
9573 // Now pack things back together.
9574 V1 = DAG.getBitcast(MVT::v8i16, V1);
9575 V2 = IsSingleInput ? V1 : DAG.getBitcast(MVT::v8i16, V2);
9576 SDValue Result = DAG.getNode(X86ISD::PACKUS, DL, MVT::v16i8, V1, V2);
9577 for (int i = 1; i < NumEvenDrops; ++i) {
9578 Result = DAG.getBitcast(MVT::v8i16, Result);
9579 Result = DAG.getNode(X86ISD::PACKUS, DL, MVT::v16i8, Result, Result);
9585 // Handle multi-input cases by blending single-input shuffles.
9586 if (NumV2Elements > 0)
9587 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v16i8, V1, V2,
9590 // The fallback path for single-input shuffles widens this into two v8i16
9591 // vectors with unpacks, shuffles those, and then pulls them back together
9595 int LoBlendMask[8] = {-1, -1, -1, -1, -1, -1, -1, -1};
9596 int HiBlendMask[8] = {-1, -1, -1, -1, -1, -1, -1, -1};
9597 for (int i = 0; i < 16; ++i)
9599 (i < 8 ? LoBlendMask[i] : HiBlendMask[i % 8]) = Mask[i];
9601 SDValue Zero = getZeroVector(MVT::v8i16, Subtarget, DAG, DL);
9603 SDValue VLoHalf, VHiHalf;
9604 // Check if any of the odd lanes in the v16i8 are used. If not, we can mask
9605 // them out and avoid using UNPCK{L,H} to extract the elements of V as
9607 if (std::none_of(std::begin(LoBlendMask), std::end(LoBlendMask),
9608 [](int M) { return M >= 0 && M % 2 == 1; }) &&
9609 std::none_of(std::begin(HiBlendMask), std::end(HiBlendMask),
9610 [](int M) { return M >= 0 && M % 2 == 1; })) {
9611 // Use a mask to drop the high bytes.
9612 VLoHalf = DAG.getBitcast(MVT::v8i16, V);
9613 VLoHalf = DAG.getNode(ISD::AND, DL, MVT::v8i16, VLoHalf,
9614 DAG.getConstant(0x00FF, DL, MVT::v8i16));
9616 // This will be a single vector shuffle instead of a blend so nuke VHiHalf.
9617 VHiHalf = DAG.getUNDEF(MVT::v8i16);
9619 // Squash the masks to point directly into VLoHalf.
9620 for (int &M : LoBlendMask)
9623 for (int &M : HiBlendMask)
9627 // Otherwise just unpack the low half of V into VLoHalf and the high half into
9628 // VHiHalf so that we can blend them as i16s.
9629 VLoHalf = DAG.getBitcast(
9630 MVT::v8i16, DAG.getNode(X86ISD::UNPCKL, DL, MVT::v16i8, V, Zero));
9631 VHiHalf = DAG.getBitcast(
9632 MVT::v8i16, DAG.getNode(X86ISD::UNPCKH, DL, MVT::v16i8, V, Zero));
9635 SDValue LoV = DAG.getVectorShuffle(MVT::v8i16, DL, VLoHalf, VHiHalf, LoBlendMask);
9636 SDValue HiV = DAG.getVectorShuffle(MVT::v8i16, DL, VLoHalf, VHiHalf, HiBlendMask);
9638 return DAG.getNode(X86ISD::PACKUS, DL, MVT::v16i8, LoV, HiV);
9641 /// \brief Dispatching routine to lower various 128-bit x86 vector shuffles.
9643 /// This routine breaks down the specific type of 128-bit shuffle and
9644 /// dispatches to the lowering routines accordingly.
9645 static SDValue lower128BitVectorShuffle(SDValue Op, SDValue V1, SDValue V2,
9646 MVT VT, const X86Subtarget *Subtarget,
9647 SelectionDAG &DAG) {
9648 switch (VT.SimpleTy) {
9650 return lowerV2I64VectorShuffle(Op, V1, V2, Subtarget, DAG);
9652 return lowerV2F64VectorShuffle(Op, V1, V2, Subtarget, DAG);
9654 return lowerV4I32VectorShuffle(Op, V1, V2, Subtarget, DAG);
9656 return lowerV4F32VectorShuffle(Op, V1, V2, Subtarget, DAG);
9658 return lowerV8I16VectorShuffle(Op, V1, V2, Subtarget, DAG);
9660 return lowerV16I8VectorShuffle(Op, V1, V2, Subtarget, DAG);
9663 llvm_unreachable("Unimplemented!");
9667 /// \brief Helper function to test whether a shuffle mask could be
9668 /// simplified by widening the elements being shuffled.
9670 /// Appends the mask for wider elements in WidenedMask if valid. Otherwise
9671 /// leaves it in an unspecified state.
9673 /// NOTE: This must handle normal vector shuffle masks and *target* vector
9674 /// shuffle masks. The latter have the special property of a '-2' representing
9675 /// a zero-ed lane of a vector.
9676 static bool canWidenShuffleElements(ArrayRef<int> Mask,
9677 SmallVectorImpl<int> &WidenedMask) {
9678 for (int i = 0, Size = Mask.size(); i < Size; i += 2) {
9679 // If both elements are undef, its trivial.
9680 if (Mask[i] == SM_SentinelUndef && Mask[i + 1] == SM_SentinelUndef) {
9681 WidenedMask.push_back(SM_SentinelUndef);
9685 // Check for an undef mask and a mask value properly aligned to fit with
9686 // a pair of values. If we find such a case, use the non-undef mask's value.
9687 if (Mask[i] == SM_SentinelUndef && Mask[i + 1] >= 0 && Mask[i + 1] % 2 == 1) {
9688 WidenedMask.push_back(Mask[i + 1] / 2);
9691 if (Mask[i + 1] == SM_SentinelUndef && Mask[i] >= 0 && Mask[i] % 2 == 0) {
9692 WidenedMask.push_back(Mask[i] / 2);
9696 // When zeroing, we need to spread the zeroing across both lanes to widen.
9697 if (Mask[i] == SM_SentinelZero || Mask[i + 1] == SM_SentinelZero) {
9698 if ((Mask[i] == SM_SentinelZero || Mask[i] == SM_SentinelUndef) &&
9699 (Mask[i + 1] == SM_SentinelZero || Mask[i + 1] == SM_SentinelUndef)) {
9700 WidenedMask.push_back(SM_SentinelZero);
9706 // Finally check if the two mask values are adjacent and aligned with
9708 if (Mask[i] != SM_SentinelUndef && Mask[i] % 2 == 0 && Mask[i] + 1 == Mask[i + 1]) {
9709 WidenedMask.push_back(Mask[i] / 2);
9713 // Otherwise we can't safely widen the elements used in this shuffle.
9716 assert(WidenedMask.size() == Mask.size() / 2 &&
9717 "Incorrect size of mask after widening the elements!");
9722 /// \brief Generic routine to split vector shuffle into half-sized shuffles.
9724 /// This routine just extracts two subvectors, shuffles them independently, and
9725 /// then concatenates them back together. This should work effectively with all
9726 /// AVX vector shuffle types.
9727 static SDValue splitAndLowerVectorShuffle(SDLoc DL, MVT VT, SDValue V1,
9728 SDValue V2, ArrayRef<int> Mask,
9729 SelectionDAG &DAG) {
9730 assert(VT.getSizeInBits() >= 256 &&
9731 "Only for 256-bit or wider vector shuffles!");
9732 assert(V1.getSimpleValueType() == VT && "Bad operand type!");
9733 assert(V2.getSimpleValueType() == VT && "Bad operand type!");
9735 ArrayRef<int> LoMask = Mask.slice(0, Mask.size() / 2);
9736 ArrayRef<int> HiMask = Mask.slice(Mask.size() / 2);
9738 int NumElements = VT.getVectorNumElements();
9739 int SplitNumElements = NumElements / 2;
9740 MVT ScalarVT = VT.getScalarType();
9741 MVT SplitVT = MVT::getVectorVT(ScalarVT, NumElements / 2);
9743 // Rather than splitting build-vectors, just build two narrower build
9744 // vectors. This helps shuffling with splats and zeros.
9745 auto SplitVector = [&](SDValue V) {
9746 while (V.getOpcode() == ISD::BITCAST)
9747 V = V->getOperand(0);
9749 MVT OrigVT = V.getSimpleValueType();
9750 int OrigNumElements = OrigVT.getVectorNumElements();
9751 int OrigSplitNumElements = OrigNumElements / 2;
9752 MVT OrigScalarVT = OrigVT.getScalarType();
9753 MVT OrigSplitVT = MVT::getVectorVT(OrigScalarVT, OrigNumElements / 2);
9757 auto *BV = dyn_cast<BuildVectorSDNode>(V);
9759 LoV = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, OrigSplitVT, V,
9760 DAG.getIntPtrConstant(0, DL));
9761 HiV = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, OrigSplitVT, V,
9762 DAG.getIntPtrConstant(OrigSplitNumElements, DL));
9765 SmallVector<SDValue, 16> LoOps, HiOps;
9766 for (int i = 0; i < OrigSplitNumElements; ++i) {
9767 LoOps.push_back(BV->getOperand(i));
9768 HiOps.push_back(BV->getOperand(i + OrigSplitNumElements));
9770 LoV = DAG.getNode(ISD::BUILD_VECTOR, DL, OrigSplitVT, LoOps);
9771 HiV = DAG.getNode(ISD::BUILD_VECTOR, DL, OrigSplitVT, HiOps);
9773 return std::make_pair(DAG.getBitcast(SplitVT, LoV),
9774 DAG.getBitcast(SplitVT, HiV));
9777 SDValue LoV1, HiV1, LoV2, HiV2;
9778 std::tie(LoV1, HiV1) = SplitVector(V1);
9779 std::tie(LoV2, HiV2) = SplitVector(V2);
9781 // Now create two 4-way blends of these half-width vectors.
9782 auto HalfBlend = [&](ArrayRef<int> HalfMask) {
9783 bool UseLoV1 = false, UseHiV1 = false, UseLoV2 = false, UseHiV2 = false;
9784 SmallVector<int, 32> V1BlendMask, V2BlendMask, BlendMask;
9785 for (int i = 0; i < SplitNumElements; ++i) {
9786 int M = HalfMask[i];
9787 if (M >= NumElements) {
9788 if (M >= NumElements + SplitNumElements)
9792 V2BlendMask.push_back(M - NumElements);
9793 V1BlendMask.push_back(-1);
9794 BlendMask.push_back(SplitNumElements + i);
9795 } else if (M >= 0) {
9796 if (M >= SplitNumElements)
9800 V2BlendMask.push_back(-1);
9801 V1BlendMask.push_back(M);
9802 BlendMask.push_back(i);
9804 V2BlendMask.push_back(-1);
9805 V1BlendMask.push_back(-1);
9806 BlendMask.push_back(-1);
9810 // Because the lowering happens after all combining takes place, we need to
9811 // manually combine these blend masks as much as possible so that we create
9812 // a minimal number of high-level vector shuffle nodes.
9814 // First try just blending the halves of V1 or V2.
9815 if (!UseLoV1 && !UseHiV1 && !UseLoV2 && !UseHiV2)
9816 return DAG.getUNDEF(SplitVT);
9817 if (!UseLoV2 && !UseHiV2)
9818 return DAG.getVectorShuffle(SplitVT, DL, LoV1, HiV1, V1BlendMask);
9819 if (!UseLoV1 && !UseHiV1)
9820 return DAG.getVectorShuffle(SplitVT, DL, LoV2, HiV2, V2BlendMask);
9822 SDValue V1Blend, V2Blend;
9823 if (UseLoV1 && UseHiV1) {
9825 DAG.getVectorShuffle(SplitVT, DL, LoV1, HiV1, V1BlendMask);
9827 // We only use half of V1 so map the usage down into the final blend mask.
9828 V1Blend = UseLoV1 ? LoV1 : HiV1;
9829 for (int i = 0; i < SplitNumElements; ++i)
9830 if (BlendMask[i] >= 0 && BlendMask[i] < SplitNumElements)
9831 BlendMask[i] = V1BlendMask[i] - (UseLoV1 ? 0 : SplitNumElements);
9833 if (UseLoV2 && UseHiV2) {
9835 DAG.getVectorShuffle(SplitVT, DL, LoV2, HiV2, V2BlendMask);
9837 // We only use half of V2 so map the usage down into the final blend mask.
9838 V2Blend = UseLoV2 ? LoV2 : HiV2;
9839 for (int i = 0; i < SplitNumElements; ++i)
9840 if (BlendMask[i] >= SplitNumElements)
9841 BlendMask[i] = V2BlendMask[i] + (UseLoV2 ? SplitNumElements : 0);
9843 return DAG.getVectorShuffle(SplitVT, DL, V1Blend, V2Blend, BlendMask);
9845 SDValue Lo = HalfBlend(LoMask);
9846 SDValue Hi = HalfBlend(HiMask);
9847 return DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, Lo, Hi);
9850 /// \brief Either split a vector in halves or decompose the shuffles and the
9853 /// This is provided as a good fallback for many lowerings of non-single-input
9854 /// shuffles with more than one 128-bit lane. In those cases, we want to select
9855 /// between splitting the shuffle into 128-bit components and stitching those
9856 /// back together vs. extracting the single-input shuffles and blending those
9858 static SDValue lowerVectorShuffleAsSplitOrBlend(SDLoc DL, MVT VT, SDValue V1,
9859 SDValue V2, ArrayRef<int> Mask,
9860 SelectionDAG &DAG) {
9861 assert(!isSingleInputShuffleMask(Mask) && "This routine must not be used to "
9862 "lower single-input shuffles as it "
9863 "could then recurse on itself.");
9864 int Size = Mask.size();
9866 // If this can be modeled as a broadcast of two elements followed by a blend,
9867 // prefer that lowering. This is especially important because broadcasts can
9868 // often fold with memory operands.
9869 auto DoBothBroadcast = [&] {
9870 int V1BroadcastIdx = -1, V2BroadcastIdx = -1;
9873 if (V2BroadcastIdx == -1)
9874 V2BroadcastIdx = M - Size;
9875 else if (M - Size != V2BroadcastIdx)
9877 } else if (M >= 0) {
9878 if (V1BroadcastIdx == -1)
9880 else if (M != V1BroadcastIdx)
9885 if (DoBothBroadcast())
9886 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, VT, V1, V2, Mask,
9889 // If the inputs all stem from a single 128-bit lane of each input, then we
9890 // split them rather than blending because the split will decompose to
9891 // unusually few instructions.
9892 int LaneCount = VT.getSizeInBits() / 128;
9893 int LaneSize = Size / LaneCount;
9894 SmallBitVector LaneInputs[2];
9895 LaneInputs[0].resize(LaneCount, false);
9896 LaneInputs[1].resize(LaneCount, false);
9897 for (int i = 0; i < Size; ++i)
9899 LaneInputs[Mask[i] / Size][(Mask[i] % Size) / LaneSize] = true;
9900 if (LaneInputs[0].count() <= 1 && LaneInputs[1].count() <= 1)
9901 return splitAndLowerVectorShuffle(DL, VT, V1, V2, Mask, DAG);
9903 // Otherwise, just fall back to decomposed shuffles and a blend. This requires
9904 // that the decomposed single-input shuffles don't end up here.
9905 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, VT, V1, V2, Mask, DAG);
9908 /// \brief Lower a vector shuffle crossing multiple 128-bit lanes as
9909 /// a permutation and blend of those lanes.
9911 /// This essentially blends the out-of-lane inputs to each lane into the lane
9912 /// from a permuted copy of the vector. This lowering strategy results in four
9913 /// instructions in the worst case for a single-input cross lane shuffle which
9914 /// is lower than any other fully general cross-lane shuffle strategy I'm aware
9915 /// of. Special cases for each particular shuffle pattern should be handled
9916 /// prior to trying this lowering.
9917 static SDValue lowerVectorShuffleAsLanePermuteAndBlend(SDLoc DL, MVT VT,
9918 SDValue V1, SDValue V2,
9920 SelectionDAG &DAG) {
9921 // FIXME: This should probably be generalized for 512-bit vectors as well.
9922 assert(VT.getSizeInBits() == 256 && "Only for 256-bit vector shuffles!");
9923 int LaneSize = Mask.size() / 2;
9925 // If there are only inputs from one 128-bit lane, splitting will in fact be
9926 // less expensive. The flags track whether the given lane contains an element
9927 // that crosses to another lane.
9928 bool LaneCrossing[2] = {false, false};
9929 for (int i = 0, Size = Mask.size(); i < Size; ++i)
9930 if (Mask[i] >= 0 && (Mask[i] % Size) / LaneSize != i / LaneSize)
9931 LaneCrossing[(Mask[i] % Size) / LaneSize] = true;
9932 if (!LaneCrossing[0] || !LaneCrossing[1])
9933 return splitAndLowerVectorShuffle(DL, VT, V1, V2, Mask, DAG);
9935 if (isSingleInputShuffleMask(Mask)) {
9936 SmallVector<int, 32> FlippedBlendMask;
9937 for (int i = 0, Size = Mask.size(); i < Size; ++i)
9938 FlippedBlendMask.push_back(
9939 Mask[i] < 0 ? -1 : (((Mask[i] % Size) / LaneSize == i / LaneSize)
9941 : Mask[i] % LaneSize +
9942 (i / LaneSize) * LaneSize + Size));
9944 // Flip the vector, and blend the results which should now be in-lane. The
9945 // VPERM2X128 mask uses the low 2 bits for the low source and bits 4 and
9946 // 5 for the high source. The value 3 selects the high half of source 2 and
9947 // the value 2 selects the low half of source 2. We only use source 2 to
9948 // allow folding it into a memory operand.
9949 unsigned PERMMask = 3 | 2 << 4;
9950 SDValue Flipped = DAG.getNode(X86ISD::VPERM2X128, DL, VT, DAG.getUNDEF(VT),
9951 V1, DAG.getConstant(PERMMask, DL, MVT::i8));
9952 return DAG.getVectorShuffle(VT, DL, V1, Flipped, FlippedBlendMask);
9955 // This now reduces to two single-input shuffles of V1 and V2 which at worst
9956 // will be handled by the above logic and a blend of the results, much like
9957 // other patterns in AVX.
9958 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, VT, V1, V2, Mask, DAG);
9961 /// \brief Handle lowering 2-lane 128-bit shuffles.
9962 static SDValue lowerV2X128VectorShuffle(SDLoc DL, MVT VT, SDValue V1,
9963 SDValue V2, ArrayRef<int> Mask,
9964 const X86Subtarget *Subtarget,
9965 SelectionDAG &DAG) {
9966 // TODO: If minimizing size and one of the inputs is a zero vector and the
9967 // the zero vector has only one use, we could use a VPERM2X128 to save the
9968 // instruction bytes needed to explicitly generate the zero vector.
9970 // Blends are faster and handle all the non-lane-crossing cases.
9971 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, VT, V1, V2, Mask,
9975 bool IsV1Zero = ISD::isBuildVectorAllZeros(V1.getNode());
9976 bool IsV2Zero = ISD::isBuildVectorAllZeros(V2.getNode());
9978 // If either input operand is a zero vector, use VPERM2X128 because its mask
9979 // allows us to replace the zero input with an implicit zero.
9980 if (!IsV1Zero && !IsV2Zero) {
9981 // Check for patterns which can be matched with a single insert of a 128-bit
9983 bool OnlyUsesV1 = isShuffleEquivalent(V1, V2, Mask, {0, 1, 0, 1});
9984 if (OnlyUsesV1 || isShuffleEquivalent(V1, V2, Mask, {0, 1, 4, 5})) {
9985 MVT SubVT = MVT::getVectorVT(VT.getVectorElementType(),
9986 VT.getVectorNumElements() / 2);
9987 SDValue LoV = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, SubVT, V1,
9988 DAG.getIntPtrConstant(0, DL));
9989 SDValue HiV = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, SubVT,
9990 OnlyUsesV1 ? V1 : V2,
9991 DAG.getIntPtrConstant(0, DL));
9992 return DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, LoV, HiV);
9996 // Otherwise form a 128-bit permutation. After accounting for undefs,
9997 // convert the 64-bit shuffle mask selection values into 128-bit
9998 // selection bits by dividing the indexes by 2 and shifting into positions
9999 // defined by a vperm2*128 instruction's immediate control byte.
10001 // The immediate permute control byte looks like this:
10002 // [1:0] - select 128 bits from sources for low half of destination
10004 // [3] - zero low half of destination
10005 // [5:4] - select 128 bits from sources for high half of destination
10007 // [7] - zero high half of destination
10009 int MaskLO = Mask[0];
10010 if (MaskLO == SM_SentinelUndef)
10011 MaskLO = Mask[1] == SM_SentinelUndef ? 0 : Mask[1];
10013 int MaskHI = Mask[2];
10014 if (MaskHI == SM_SentinelUndef)
10015 MaskHI = Mask[3] == SM_SentinelUndef ? 0 : Mask[3];
10017 unsigned PermMask = MaskLO / 2 | (MaskHI / 2) << 4;
10019 // If either input is a zero vector, replace it with an undef input.
10020 // Shuffle mask values < 4 are selecting elements of V1.
10021 // Shuffle mask values >= 4 are selecting elements of V2.
10022 // Adjust each half of the permute mask by clearing the half that was
10023 // selecting the zero vector and setting the zero mask bit.
10025 V1 = DAG.getUNDEF(VT);
10027 PermMask = (PermMask & 0xf0) | 0x08;
10029 PermMask = (PermMask & 0x0f) | 0x80;
10032 V2 = DAG.getUNDEF(VT);
10034 PermMask = (PermMask & 0xf0) | 0x08;
10036 PermMask = (PermMask & 0x0f) | 0x80;
10039 return DAG.getNode(X86ISD::VPERM2X128, DL, VT, V1, V2,
10040 DAG.getConstant(PermMask, DL, MVT::i8));
10043 /// \brief Lower a vector shuffle by first fixing the 128-bit lanes and then
10044 /// shuffling each lane.
10046 /// This will only succeed when the result of fixing the 128-bit lanes results
10047 /// in a single-input non-lane-crossing shuffle with a repeating shuffle mask in
10048 /// each 128-bit lanes. This handles many cases where we can quickly blend away
10049 /// the lane crosses early and then use simpler shuffles within each lane.
10051 /// FIXME: It might be worthwhile at some point to support this without
10052 /// requiring the 128-bit lane-relative shuffles to be repeating, but currently
10053 /// in x86 only floating point has interesting non-repeating shuffles, and even
10054 /// those are still *marginally* more expensive.
10055 static SDValue lowerVectorShuffleByMerging128BitLanes(
10056 SDLoc DL, MVT VT, SDValue V1, SDValue V2, ArrayRef<int> Mask,
10057 const X86Subtarget *Subtarget, SelectionDAG &DAG) {
10058 assert(!isSingleInputShuffleMask(Mask) &&
10059 "This is only useful with multiple inputs.");
10061 int Size = Mask.size();
10062 int LaneSize = 128 / VT.getScalarSizeInBits();
10063 int NumLanes = Size / LaneSize;
10064 assert(NumLanes > 1 && "Only handles 256-bit and wider shuffles.");
10066 // See if we can build a hypothetical 128-bit lane-fixing shuffle mask. Also
10067 // check whether the in-128-bit lane shuffles share a repeating pattern.
10068 SmallVector<int, 4> Lanes;
10069 Lanes.resize(NumLanes, -1);
10070 SmallVector<int, 4> InLaneMask;
10071 InLaneMask.resize(LaneSize, -1);
10072 for (int i = 0; i < Size; ++i) {
10076 int j = i / LaneSize;
10078 if (Lanes[j] < 0) {
10079 // First entry we've seen for this lane.
10080 Lanes[j] = Mask[i] / LaneSize;
10081 } else if (Lanes[j] != Mask[i] / LaneSize) {
10082 // This doesn't match the lane selected previously!
10086 // Check that within each lane we have a consistent shuffle mask.
10087 int k = i % LaneSize;
10088 if (InLaneMask[k] < 0) {
10089 InLaneMask[k] = Mask[i] % LaneSize;
10090 } else if (InLaneMask[k] != Mask[i] % LaneSize) {
10091 // This doesn't fit a repeating in-lane mask.
10096 // First shuffle the lanes into place.
10097 MVT LaneVT = MVT::getVectorVT(VT.isFloatingPoint() ? MVT::f64 : MVT::i64,
10098 VT.getSizeInBits() / 64);
10099 SmallVector<int, 8> LaneMask;
10100 LaneMask.resize(NumLanes * 2, -1);
10101 for (int i = 0; i < NumLanes; ++i)
10102 if (Lanes[i] >= 0) {
10103 LaneMask[2 * i + 0] = 2*Lanes[i] + 0;
10104 LaneMask[2 * i + 1] = 2*Lanes[i] + 1;
10107 V1 = DAG.getBitcast(LaneVT, V1);
10108 V2 = DAG.getBitcast(LaneVT, V2);
10109 SDValue LaneShuffle = DAG.getVectorShuffle(LaneVT, DL, V1, V2, LaneMask);
10111 // Cast it back to the type we actually want.
10112 LaneShuffle = DAG.getBitcast(VT, LaneShuffle);
10114 // Now do a simple shuffle that isn't lane crossing.
10115 SmallVector<int, 8> NewMask;
10116 NewMask.resize(Size, -1);
10117 for (int i = 0; i < Size; ++i)
10119 NewMask[i] = (i / LaneSize) * LaneSize + Mask[i] % LaneSize;
10120 assert(!is128BitLaneCrossingShuffleMask(VT, NewMask) &&
10121 "Must not introduce lane crosses at this point!");
10123 return DAG.getVectorShuffle(VT, DL, LaneShuffle, DAG.getUNDEF(VT), NewMask);
10126 /// \brief Test whether the specified input (0 or 1) is in-place blended by the
10129 /// This returns true if the elements from a particular input are already in the
10130 /// slot required by the given mask and require no permutation.
10131 static bool isShuffleMaskInputInPlace(int Input, ArrayRef<int> Mask) {
10132 assert((Input == 0 || Input == 1) && "Only two inputs to shuffles.");
10133 int Size = Mask.size();
10134 for (int i = 0; i < Size; ++i)
10135 if (Mask[i] >= 0 && Mask[i] / Size == Input && Mask[i] % Size != i)
10141 static SDValue lowerVectorShuffleWithSHUFPD(SDLoc DL, MVT VT,
10142 ArrayRef<int> Mask, SDValue V1,
10143 SDValue V2, SelectionDAG &DAG) {
10145 // Mask for V8F64: 0/1, 8/9, 2/3, 10/11, 4/5, ..
10146 // Mask for V4F64; 0/1, 4/5, 2/3, 6/7..
10147 assert(VT.getScalarSizeInBits() == 64 && "Unexpected data type for VSHUFPD");
10148 int NumElts = VT.getVectorNumElements();
10149 bool ShufpdMask = true;
10150 bool CommutableMask = true;
10151 unsigned Immediate = 0;
10152 for (int i = 0; i < NumElts; ++i) {
10155 int Val = (i & 6) + NumElts * (i & 1);
10156 int CommutVal = (i & 0xe) + NumElts * ((i & 1)^1);
10157 if (Mask[i] < Val || Mask[i] > Val + 1)
10158 ShufpdMask = false;
10159 if (Mask[i] < CommutVal || Mask[i] > CommutVal + 1)
10160 CommutableMask = false;
10161 Immediate |= (Mask[i] % 2) << i;
10164 return DAG.getNode(X86ISD::SHUFP, DL, VT, V1, V2,
10165 DAG.getConstant(Immediate, DL, MVT::i8));
10166 if (CommutableMask)
10167 return DAG.getNode(X86ISD::SHUFP, DL, VT, V2, V1,
10168 DAG.getConstant(Immediate, DL, MVT::i8));
10172 /// \brief Handle lowering of 4-lane 64-bit floating point shuffles.
10174 /// Also ends up handling lowering of 4-lane 64-bit integer shuffles when AVX2
10175 /// isn't available.
10176 static SDValue lowerV4F64VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10177 const X86Subtarget *Subtarget,
10178 SelectionDAG &DAG) {
10180 assert(V1.getSimpleValueType() == MVT::v4f64 && "Bad operand type!");
10181 assert(V2.getSimpleValueType() == MVT::v4f64 && "Bad operand type!");
10182 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10183 ArrayRef<int> Mask = SVOp->getMask();
10184 assert(Mask.size() == 4 && "Unexpected mask size for v4 shuffle!");
10186 SmallVector<int, 4> WidenedMask;
10187 if (canWidenShuffleElements(Mask, WidenedMask))
10188 return lowerV2X128VectorShuffle(DL, MVT::v4f64, V1, V2, Mask, Subtarget,
10191 if (isSingleInputShuffleMask(Mask)) {
10192 // Check for being able to broadcast a single element.
10193 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v4f64, V1,
10194 Mask, Subtarget, DAG))
10197 // Use low duplicate instructions for masks that match their pattern.
10198 if (isShuffleEquivalent(V1, V2, Mask, {0, 0, 2, 2}))
10199 return DAG.getNode(X86ISD::MOVDDUP, DL, MVT::v4f64, V1);
10201 if (!is128BitLaneCrossingShuffleMask(MVT::v4f64, Mask)) {
10202 // Non-half-crossing single input shuffles can be lowerid with an
10203 // interleaved permutation.
10204 unsigned VPERMILPMask = (Mask[0] == 1) | ((Mask[1] == 1) << 1) |
10205 ((Mask[2] == 3) << 2) | ((Mask[3] == 3) << 3);
10206 return DAG.getNode(X86ISD::VPERMILPI, DL, MVT::v4f64, V1,
10207 DAG.getConstant(VPERMILPMask, DL, MVT::i8));
10210 // With AVX2 we have direct support for this permutation.
10211 if (Subtarget->hasAVX2())
10212 return DAG.getNode(X86ISD::VPERMI, DL, MVT::v4f64, V1,
10213 getV4X86ShuffleImm8ForMask(Mask, DL, DAG));
10215 // Otherwise, fall back.
10216 return lowerVectorShuffleAsLanePermuteAndBlend(DL, MVT::v4f64, V1, V2, Mask,
10220 // Use dedicated unpack instructions for masks that match their pattern.
10222 lowerVectorShuffleWithUNPCK(DL, MVT::v4f64, Mask, V1, V2, DAG))
10225 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v4f64, V1, V2, Mask,
10229 // Check if the blend happens to exactly fit that of SHUFPD.
10231 lowerVectorShuffleWithSHUFPD(DL, MVT::v4f64, Mask, V1, V2, DAG))
10234 // Try to simplify this by merging 128-bit lanes to enable a lane-based
10235 // shuffle. However, if we have AVX2 and either inputs are already in place,
10236 // we will be able to shuffle even across lanes the other input in a single
10237 // instruction so skip this pattern.
10238 if (!(Subtarget->hasAVX2() && (isShuffleMaskInputInPlace(0, Mask) ||
10239 isShuffleMaskInputInPlace(1, Mask))))
10240 if (SDValue Result = lowerVectorShuffleByMerging128BitLanes(
10241 DL, MVT::v4f64, V1, V2, Mask, Subtarget, DAG))
10244 // If we have AVX2 then we always want to lower with a blend because an v4 we
10245 // can fully permute the elements.
10246 if (Subtarget->hasAVX2())
10247 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v4f64, V1, V2,
10250 // Otherwise fall back on generic lowering.
10251 return lowerVectorShuffleAsSplitOrBlend(DL, MVT::v4f64, V1, V2, Mask, DAG);
10254 /// \brief Handle lowering of 4-lane 64-bit integer shuffles.
10256 /// This routine is only called when we have AVX2 and thus a reasonable
10257 /// instruction set for v4i64 shuffling..
10258 static SDValue lowerV4I64VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10259 const X86Subtarget *Subtarget,
10260 SelectionDAG &DAG) {
10262 assert(V1.getSimpleValueType() == MVT::v4i64 && "Bad operand type!");
10263 assert(V2.getSimpleValueType() == MVT::v4i64 && "Bad operand type!");
10264 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10265 ArrayRef<int> Mask = SVOp->getMask();
10266 assert(Mask.size() == 4 && "Unexpected mask size for v4 shuffle!");
10267 assert(Subtarget->hasAVX2() && "We can only lower v4i64 with AVX2!");
10269 SmallVector<int, 4> WidenedMask;
10270 if (canWidenShuffleElements(Mask, WidenedMask))
10271 return lowerV2X128VectorShuffle(DL, MVT::v4i64, V1, V2, Mask, Subtarget,
10274 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v4i64, V1, V2, Mask,
10278 // Check for being able to broadcast a single element.
10279 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v4i64, V1,
10280 Mask, Subtarget, DAG))
10283 // When the shuffle is mirrored between the 128-bit lanes of the unit, we can
10284 // use lower latency instructions that will operate on both 128-bit lanes.
10285 SmallVector<int, 2> RepeatedMask;
10286 if (is128BitLaneRepeatedShuffleMask(MVT::v4i64, Mask, RepeatedMask)) {
10287 if (isSingleInputShuffleMask(Mask)) {
10288 int PSHUFDMask[] = {-1, -1, -1, -1};
10289 for (int i = 0; i < 2; ++i)
10290 if (RepeatedMask[i] >= 0) {
10291 PSHUFDMask[2 * i] = 2 * RepeatedMask[i];
10292 PSHUFDMask[2 * i + 1] = 2 * RepeatedMask[i] + 1;
10294 return DAG.getBitcast(
10296 DAG.getNode(X86ISD::PSHUFD, DL, MVT::v8i32,
10297 DAG.getBitcast(MVT::v8i32, V1),
10298 getV4X86ShuffleImm8ForMask(PSHUFDMask, DL, DAG)));
10302 // AVX2 provides a direct instruction for permuting a single input across
10304 if (isSingleInputShuffleMask(Mask))
10305 return DAG.getNode(X86ISD::VPERMI, DL, MVT::v4i64, V1,
10306 getV4X86ShuffleImm8ForMask(Mask, DL, DAG));
10308 // Try to use shift instructions.
10309 if (SDValue Shift =
10310 lowerVectorShuffleAsShift(DL, MVT::v4i64, V1, V2, Mask, DAG))
10313 // Use dedicated unpack instructions for masks that match their pattern.
10315 lowerVectorShuffleWithUNPCK(DL, MVT::v4i64, Mask, V1, V2, DAG))
10318 // Try to simplify this by merging 128-bit lanes to enable a lane-based
10319 // shuffle. However, if we have AVX2 and either inputs are already in place,
10320 // we will be able to shuffle even across lanes the other input in a single
10321 // instruction so skip this pattern.
10322 if (!(Subtarget->hasAVX2() && (isShuffleMaskInputInPlace(0, Mask) ||
10323 isShuffleMaskInputInPlace(1, Mask))))
10324 if (SDValue Result = lowerVectorShuffleByMerging128BitLanes(
10325 DL, MVT::v4i64, V1, V2, Mask, Subtarget, DAG))
10328 // Otherwise fall back on generic blend lowering.
10329 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v4i64, V1, V2,
10333 /// \brief Handle lowering of 8-lane 32-bit floating point shuffles.
10335 /// Also ends up handling lowering of 8-lane 32-bit integer shuffles when AVX2
10336 /// isn't available.
10337 static SDValue lowerV8F32VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10338 const X86Subtarget *Subtarget,
10339 SelectionDAG &DAG) {
10341 assert(V1.getSimpleValueType() == MVT::v8f32 && "Bad operand type!");
10342 assert(V2.getSimpleValueType() == MVT::v8f32 && "Bad operand type!");
10343 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10344 ArrayRef<int> Mask = SVOp->getMask();
10345 assert(Mask.size() == 8 && "Unexpected mask size for v8 shuffle!");
10347 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v8f32, V1, V2, Mask,
10351 // Check for being able to broadcast a single element.
10352 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v8f32, V1,
10353 Mask, Subtarget, DAG))
10356 // If the shuffle mask is repeated in each 128-bit lane, we have many more
10357 // options to efficiently lower the shuffle.
10358 SmallVector<int, 4> RepeatedMask;
10359 if (is128BitLaneRepeatedShuffleMask(MVT::v8f32, Mask, RepeatedMask)) {
10360 assert(RepeatedMask.size() == 4 &&
10361 "Repeated masks must be half the mask width!");
10363 // Use even/odd duplicate instructions for masks that match their pattern.
10364 if (isShuffleEquivalent(V1, V2, Mask, {0, 0, 2, 2, 4, 4, 6, 6}))
10365 return DAG.getNode(X86ISD::MOVSLDUP, DL, MVT::v8f32, V1);
10366 if (isShuffleEquivalent(V1, V2, Mask, {1, 1, 3, 3, 5, 5, 7, 7}))
10367 return DAG.getNode(X86ISD::MOVSHDUP, DL, MVT::v8f32, V1);
10369 if (isSingleInputShuffleMask(Mask))
10370 return DAG.getNode(X86ISD::VPERMILPI, DL, MVT::v8f32, V1,
10371 getV4X86ShuffleImm8ForMask(RepeatedMask, DL, DAG));
10373 // Use dedicated unpack instructions for masks that match their pattern.
10375 lowerVectorShuffleWithUNPCK(DL, MVT::v8f32, Mask, V1, V2, DAG))
10378 // Otherwise, fall back to a SHUFPS sequence. Here it is important that we
10379 // have already handled any direct blends. We also need to squash the
10380 // repeated mask into a simulated v4f32 mask.
10381 for (int i = 0; i < 4; ++i)
10382 if (RepeatedMask[i] >= 8)
10383 RepeatedMask[i] -= 4;
10384 return lowerVectorShuffleWithSHUFPS(DL, MVT::v8f32, RepeatedMask, V1, V2, DAG);
10387 // If we have a single input shuffle with different shuffle patterns in the
10388 // two 128-bit lanes use the variable mask to VPERMILPS.
10389 if (isSingleInputShuffleMask(Mask)) {
10390 SDValue VPermMask[8];
10391 for (int i = 0; i < 8; ++i)
10392 VPermMask[i] = Mask[i] < 0 ? DAG.getUNDEF(MVT::i32)
10393 : DAG.getConstant(Mask[i], DL, MVT::i32);
10394 if (!is128BitLaneCrossingShuffleMask(MVT::v8f32, Mask))
10395 return DAG.getNode(
10396 X86ISD::VPERMILPV, DL, MVT::v8f32, V1,
10397 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v8i32, VPermMask));
10399 if (Subtarget->hasAVX2())
10400 return DAG.getNode(
10401 X86ISD::VPERMV, DL, MVT::v8f32,
10402 DAG.getBitcast(MVT::v8f32, DAG.getNode(ISD::BUILD_VECTOR, DL,
10403 MVT::v8i32, VPermMask)),
10406 // Otherwise, fall back.
10407 return lowerVectorShuffleAsLanePermuteAndBlend(DL, MVT::v8f32, V1, V2, Mask,
10411 // Try to simplify this by merging 128-bit lanes to enable a lane-based
10413 if (SDValue Result = lowerVectorShuffleByMerging128BitLanes(
10414 DL, MVT::v8f32, V1, V2, Mask, Subtarget, DAG))
10417 // If we have AVX2 then we always want to lower with a blend because at v8 we
10418 // can fully permute the elements.
10419 if (Subtarget->hasAVX2())
10420 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v8f32, V1, V2,
10423 // Otherwise fall back on generic lowering.
10424 return lowerVectorShuffleAsSplitOrBlend(DL, MVT::v8f32, V1, V2, Mask, DAG);
10427 /// \brief Handle lowering of 8-lane 32-bit integer shuffles.
10429 /// This routine is only called when we have AVX2 and thus a reasonable
10430 /// instruction set for v8i32 shuffling..
10431 static SDValue lowerV8I32VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10432 const X86Subtarget *Subtarget,
10433 SelectionDAG &DAG) {
10435 assert(V1.getSimpleValueType() == MVT::v8i32 && "Bad operand type!");
10436 assert(V2.getSimpleValueType() == MVT::v8i32 && "Bad operand type!");
10437 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10438 ArrayRef<int> Mask = SVOp->getMask();
10439 assert(Mask.size() == 8 && "Unexpected mask size for v8 shuffle!");
10440 assert(Subtarget->hasAVX2() && "We can only lower v8i32 with AVX2!");
10442 // Whenever we can lower this as a zext, that instruction is strictly faster
10443 // than any alternative. It also allows us to fold memory operands into the
10444 // shuffle in many cases.
10445 if (SDValue ZExt = lowerVectorShuffleAsZeroOrAnyExtend(DL, MVT::v8i32, V1, V2,
10446 Mask, Subtarget, DAG))
10449 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v8i32, V1, V2, Mask,
10453 // Check for being able to broadcast a single element.
10454 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v8i32, V1,
10455 Mask, Subtarget, DAG))
10458 // If the shuffle mask is repeated in each 128-bit lane we can use more
10459 // efficient instructions that mirror the shuffles across the two 128-bit
10461 SmallVector<int, 4> RepeatedMask;
10462 if (is128BitLaneRepeatedShuffleMask(MVT::v8i32, Mask, RepeatedMask)) {
10463 assert(RepeatedMask.size() == 4 && "Unexpected repeated mask size!");
10464 if (isSingleInputShuffleMask(Mask))
10465 return DAG.getNode(X86ISD::PSHUFD, DL, MVT::v8i32, V1,
10466 getV4X86ShuffleImm8ForMask(RepeatedMask, DL, DAG));
10468 // Use dedicated unpack instructions for masks that match their pattern.
10470 lowerVectorShuffleWithUNPCK(DL, MVT::v8i32, Mask, V1, V2, DAG))
10474 // Try to use shift instructions.
10475 if (SDValue Shift =
10476 lowerVectorShuffleAsShift(DL, MVT::v8i32, V1, V2, Mask, DAG))
10479 if (SDValue Rotate = lowerVectorShuffleAsByteRotate(
10480 DL, MVT::v8i32, V1, V2, Mask, Subtarget, DAG))
10483 // If the shuffle patterns aren't repeated but it is a single input, directly
10484 // generate a cross-lane VPERMD instruction.
10485 if (isSingleInputShuffleMask(Mask)) {
10486 SDValue VPermMask[8];
10487 for (int i = 0; i < 8; ++i)
10488 VPermMask[i] = Mask[i] < 0 ? DAG.getUNDEF(MVT::i32)
10489 : DAG.getConstant(Mask[i], DL, MVT::i32);
10490 return DAG.getNode(
10491 X86ISD::VPERMV, DL, MVT::v8i32,
10492 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v8i32, VPermMask), V1);
10495 // Try to simplify this by merging 128-bit lanes to enable a lane-based
10497 if (SDValue Result = lowerVectorShuffleByMerging128BitLanes(
10498 DL, MVT::v8i32, V1, V2, Mask, Subtarget, DAG))
10501 // Otherwise fall back on generic blend lowering.
10502 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v8i32, V1, V2,
10506 /// \brief Handle lowering of 16-lane 16-bit integer shuffles.
10508 /// This routine is only called when we have AVX2 and thus a reasonable
10509 /// instruction set for v16i16 shuffling..
10510 static SDValue lowerV16I16VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10511 const X86Subtarget *Subtarget,
10512 SelectionDAG &DAG) {
10514 assert(V1.getSimpleValueType() == MVT::v16i16 && "Bad operand type!");
10515 assert(V2.getSimpleValueType() == MVT::v16i16 && "Bad operand type!");
10516 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10517 ArrayRef<int> Mask = SVOp->getMask();
10518 assert(Mask.size() == 16 && "Unexpected mask size for v16 shuffle!");
10519 assert(Subtarget->hasAVX2() && "We can only lower v16i16 with AVX2!");
10521 // Whenever we can lower this as a zext, that instruction is strictly faster
10522 // than any alternative. It also allows us to fold memory operands into the
10523 // shuffle in many cases.
10524 if (SDValue ZExt = lowerVectorShuffleAsZeroOrAnyExtend(DL, MVT::v16i16, V1, V2,
10525 Mask, Subtarget, DAG))
10528 // Check for being able to broadcast a single element.
10529 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v16i16, V1,
10530 Mask, Subtarget, DAG))
10533 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v16i16, V1, V2, Mask,
10537 // Use dedicated unpack instructions for masks that match their pattern.
10539 lowerVectorShuffleWithUNPCK(DL, MVT::v16i16, Mask, V1, V2, DAG))
10542 // Try to use shift instructions.
10543 if (SDValue Shift =
10544 lowerVectorShuffleAsShift(DL, MVT::v16i16, V1, V2, Mask, DAG))
10547 // Try to use byte rotation instructions.
10548 if (SDValue Rotate = lowerVectorShuffleAsByteRotate(
10549 DL, MVT::v16i16, V1, V2, Mask, Subtarget, DAG))
10552 if (isSingleInputShuffleMask(Mask)) {
10553 // There are no generalized cross-lane shuffle operations available on i16
10555 if (is128BitLaneCrossingShuffleMask(MVT::v16i16, Mask))
10556 return lowerVectorShuffleAsLanePermuteAndBlend(DL, MVT::v16i16, V1, V2,
10559 SmallVector<int, 8> RepeatedMask;
10560 if (is128BitLaneRepeatedShuffleMask(MVT::v16i16, Mask, RepeatedMask)) {
10561 // As this is a single-input shuffle, the repeated mask should be
10562 // a strictly valid v8i16 mask that we can pass through to the v8i16
10563 // lowering to handle even the v16 case.
10564 return lowerV8I16GeneralSingleInputVectorShuffle(
10565 DL, MVT::v16i16, V1, RepeatedMask, Subtarget, DAG);
10568 SDValue PSHUFBMask[32];
10569 for (int i = 0; i < 16; ++i) {
10570 if (Mask[i] == -1) {
10571 PSHUFBMask[2 * i] = PSHUFBMask[2 * i + 1] = DAG.getUNDEF(MVT::i8);
10575 int M = i < 8 ? Mask[i] : Mask[i] - 8;
10576 assert(M >= 0 && M < 8 && "Invalid single-input mask!");
10577 PSHUFBMask[2 * i] = DAG.getConstant(2 * M, DL, MVT::i8);
10578 PSHUFBMask[2 * i + 1] = DAG.getConstant(2 * M + 1, DL, MVT::i8);
10580 return DAG.getBitcast(MVT::v16i16,
10581 DAG.getNode(X86ISD::PSHUFB, DL, MVT::v32i8,
10582 DAG.getBitcast(MVT::v32i8, V1),
10583 DAG.getNode(ISD::BUILD_VECTOR, DL,
10584 MVT::v32i8, PSHUFBMask)));
10587 // Try to simplify this by merging 128-bit lanes to enable a lane-based
10589 if (SDValue Result = lowerVectorShuffleByMerging128BitLanes(
10590 DL, MVT::v16i16, V1, V2, Mask, Subtarget, DAG))
10593 // Otherwise fall back on generic lowering.
10594 return lowerVectorShuffleAsSplitOrBlend(DL, MVT::v16i16, V1, V2, Mask, DAG);
10597 /// \brief Handle lowering of 32-lane 8-bit integer shuffles.
10599 /// This routine is only called when we have AVX2 and thus a reasonable
10600 /// instruction set for v32i8 shuffling..
10601 static SDValue lowerV32I8VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10602 const X86Subtarget *Subtarget,
10603 SelectionDAG &DAG) {
10605 assert(V1.getSimpleValueType() == MVT::v32i8 && "Bad operand type!");
10606 assert(V2.getSimpleValueType() == MVT::v32i8 && "Bad operand type!");
10607 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10608 ArrayRef<int> Mask = SVOp->getMask();
10609 assert(Mask.size() == 32 && "Unexpected mask size for v32 shuffle!");
10610 assert(Subtarget->hasAVX2() && "We can only lower v32i8 with AVX2!");
10612 // Whenever we can lower this as a zext, that instruction is strictly faster
10613 // than any alternative. It also allows us to fold memory operands into the
10614 // shuffle in many cases.
10615 if (SDValue ZExt = lowerVectorShuffleAsZeroOrAnyExtend(DL, MVT::v32i8, V1, V2,
10616 Mask, Subtarget, DAG))
10619 // Check for being able to broadcast a single element.
10620 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v32i8, V1,
10621 Mask, Subtarget, DAG))
10624 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v32i8, V1, V2, Mask,
10628 // Use dedicated unpack instructions for masks that match their pattern.
10630 lowerVectorShuffleWithUNPCK(DL, MVT::v32i8, Mask, V1, V2, DAG))
10633 // Try to use shift instructions.
10634 if (SDValue Shift =
10635 lowerVectorShuffleAsShift(DL, MVT::v32i8, V1, V2, Mask, DAG))
10638 // Try to use byte rotation instructions.
10639 if (SDValue Rotate = lowerVectorShuffleAsByteRotate(
10640 DL, MVT::v32i8, V1, V2, Mask, Subtarget, DAG))
10643 if (isSingleInputShuffleMask(Mask)) {
10644 // There are no generalized cross-lane shuffle operations available on i8
10646 if (is128BitLaneCrossingShuffleMask(MVT::v32i8, Mask))
10647 return lowerVectorShuffleAsLanePermuteAndBlend(DL, MVT::v32i8, V1, V2,
10650 SDValue PSHUFBMask[32];
10651 for (int i = 0; i < 32; ++i)
10654 ? DAG.getUNDEF(MVT::i8)
10655 : DAG.getConstant(Mask[i] < 16 ? Mask[i] : Mask[i] - 16, DL,
10658 return DAG.getNode(
10659 X86ISD::PSHUFB, DL, MVT::v32i8, V1,
10660 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v32i8, PSHUFBMask));
10663 // Try to simplify this by merging 128-bit lanes to enable a lane-based
10665 if (SDValue Result = lowerVectorShuffleByMerging128BitLanes(
10666 DL, MVT::v32i8, V1, V2, Mask, Subtarget, DAG))
10669 // Otherwise fall back on generic lowering.
10670 return lowerVectorShuffleAsSplitOrBlend(DL, MVT::v32i8, V1, V2, Mask, DAG);
10673 /// \brief High-level routine to lower various 256-bit x86 vector shuffles.
10675 /// This routine either breaks down the specific type of a 256-bit x86 vector
10676 /// shuffle or splits it into two 128-bit shuffles and fuses the results back
10677 /// together based on the available instructions.
10678 static SDValue lower256BitVectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10679 MVT VT, const X86Subtarget *Subtarget,
10680 SelectionDAG &DAG) {
10682 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10683 ArrayRef<int> Mask = SVOp->getMask();
10685 // If we have a single input to the zero element, insert that into V1 if we
10686 // can do so cheaply.
10687 int NumElts = VT.getVectorNumElements();
10688 int NumV2Elements = std::count_if(Mask.begin(), Mask.end(), [NumElts](int M) {
10689 return M >= NumElts;
10692 if (NumV2Elements == 1 && Mask[0] >= NumElts)
10693 if (SDValue Insertion = lowerVectorShuffleAsElementInsertion(
10694 DL, VT, V1, V2, Mask, Subtarget, DAG))
10697 // There is a really nice hard cut-over between AVX1 and AVX2 that means we
10698 // can check for those subtargets here and avoid much of the subtarget
10699 // querying in the per-vector-type lowering routines. With AVX1 we have
10700 // essentially *zero* ability to manipulate a 256-bit vector with integer
10701 // types. Since we'll use floating point types there eventually, just
10702 // immediately cast everything to a float and operate entirely in that domain.
10703 if (VT.isInteger() && !Subtarget->hasAVX2()) {
10704 int ElementBits = VT.getScalarSizeInBits();
10705 if (ElementBits < 32)
10706 // No floating point type available, decompose into 128-bit vectors.
10707 return splitAndLowerVectorShuffle(DL, VT, V1, V2, Mask, DAG);
10709 MVT FpVT = MVT::getVectorVT(MVT::getFloatingPointVT(ElementBits),
10710 VT.getVectorNumElements());
10711 V1 = DAG.getBitcast(FpVT, V1);
10712 V2 = DAG.getBitcast(FpVT, V2);
10713 return DAG.getBitcast(VT, DAG.getVectorShuffle(FpVT, DL, V1, V2, Mask));
10716 switch (VT.SimpleTy) {
10718 return lowerV4F64VectorShuffle(Op, V1, V2, Subtarget, DAG);
10720 return lowerV4I64VectorShuffle(Op, V1, V2, Subtarget, DAG);
10722 return lowerV8F32VectorShuffle(Op, V1, V2, Subtarget, DAG);
10724 return lowerV8I32VectorShuffle(Op, V1, V2, Subtarget, DAG);
10726 return lowerV16I16VectorShuffle(Op, V1, V2, Subtarget, DAG);
10728 return lowerV32I8VectorShuffle(Op, V1, V2, Subtarget, DAG);
10731 llvm_unreachable("Not a valid 256-bit x86 vector type!");
10735 /// \brief Try to lower a vector shuffle as a 128-bit shuffles.
10736 static SDValue lowerV4X128VectorShuffle(SDLoc DL, MVT VT,
10737 ArrayRef<int> Mask,
10738 SDValue V1, SDValue V2,
10739 SelectionDAG &DAG) {
10740 assert(VT.getScalarSizeInBits() == 64 &&
10741 "Unexpected element type size for 128bit shuffle.");
10743 // To handle 256 bit vector requires VLX and most probably
10744 // function lowerV2X128VectorShuffle() is better solution.
10745 assert(VT.getSizeInBits() == 512 &&
10746 "Unexpected vector size for 128bit shuffle.");
10748 SmallVector<int, 4> WidenedMask;
10749 if (!canWidenShuffleElements(Mask, WidenedMask))
10752 // Form a 128-bit permutation.
10753 // Convert the 64-bit shuffle mask selection values into 128-bit selection
10754 // bits defined by a vshuf64x2 instruction's immediate control byte.
10755 unsigned PermMask = 0, Imm = 0;
10756 unsigned ControlBitsNum = WidenedMask.size() / 2;
10758 for (int i = 0, Size = WidenedMask.size(); i < Size; ++i) {
10759 if (WidenedMask[i] == SM_SentinelZero)
10762 // Use first element in place of undef mask.
10763 Imm = (WidenedMask[i] == SM_SentinelUndef) ? 0 : WidenedMask[i];
10764 PermMask |= (Imm % WidenedMask.size()) << (i * ControlBitsNum);
10767 return DAG.getNode(X86ISD::SHUF128, DL, VT, V1, V2,
10768 DAG.getConstant(PermMask, DL, MVT::i8));
10771 static SDValue lowerVectorShuffleWithPERMV(SDLoc DL, MVT VT,
10772 ArrayRef<int> Mask, SDValue V1,
10773 SDValue V2, SelectionDAG &DAG) {
10775 assert(VT.getScalarSizeInBits() >= 16 && "Unexpected data type for PERMV");
10777 MVT MaskEltVT = MVT::getIntegerVT(VT.getScalarSizeInBits());
10778 MVT MaskVecVT = MVT::getVectorVT(MaskEltVT, VT.getVectorNumElements());
10780 SDValue MaskNode = getConstVector(Mask, MaskVecVT, DAG, DL, true);
10781 if (isSingleInputShuffleMask(Mask))
10782 return DAG.getNode(X86ISD::VPERMV, DL, VT, MaskNode, V1);
10784 return DAG.getNode(X86ISD::VPERMV3, DL, VT, V1, MaskNode, V2);
10787 /// \brief Handle lowering of 8-lane 64-bit floating point shuffles.
10788 static SDValue lowerV8F64VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10789 const X86Subtarget *Subtarget,
10790 SelectionDAG &DAG) {
10792 assert(V1.getSimpleValueType() == MVT::v8f64 && "Bad operand type!");
10793 assert(V2.getSimpleValueType() == MVT::v8f64 && "Bad operand type!");
10794 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10795 ArrayRef<int> Mask = SVOp->getMask();
10796 assert(Mask.size() == 8 && "Unexpected mask size for v8 shuffle!");
10798 if (SDValue Shuf128 =
10799 lowerV4X128VectorShuffle(DL, MVT::v8f64, Mask, V1, V2, DAG))
10802 if (SDValue Unpck =
10803 lowerVectorShuffleWithUNPCK(DL, MVT::v8f64, Mask, V1, V2, DAG))
10806 return lowerVectorShuffleWithPERMV(DL, MVT::v8f64, Mask, V1, V2, DAG);
10809 /// \brief Handle lowering of 16-lane 32-bit floating point shuffles.
10810 static SDValue lowerV16F32VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10811 const X86Subtarget *Subtarget,
10812 SelectionDAG &DAG) {
10814 assert(V1.getSimpleValueType() == MVT::v16f32 && "Bad operand type!");
10815 assert(V2.getSimpleValueType() == MVT::v16f32 && "Bad operand type!");
10816 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10817 ArrayRef<int> Mask = SVOp->getMask();
10818 assert(Mask.size() == 16 && "Unexpected mask size for v16 shuffle!");
10820 if (SDValue Unpck =
10821 lowerVectorShuffleWithUNPCK(DL, MVT::v16f32, Mask, V1, V2, DAG))
10824 return lowerVectorShuffleWithPERMV(DL, MVT::v16f32, Mask, V1, V2, DAG);
10827 /// \brief Handle lowering of 8-lane 64-bit integer shuffles.
10828 static SDValue lowerV8I64VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10829 const X86Subtarget *Subtarget,
10830 SelectionDAG &DAG) {
10832 assert(V1.getSimpleValueType() == MVT::v8i64 && "Bad operand type!");
10833 assert(V2.getSimpleValueType() == MVT::v8i64 && "Bad operand type!");
10834 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10835 ArrayRef<int> Mask = SVOp->getMask();
10836 assert(Mask.size() == 8 && "Unexpected mask size for v8 shuffle!");
10838 if (SDValue Shuf128 =
10839 lowerV4X128VectorShuffle(DL, MVT::v8i64, Mask, V1, V2, DAG))
10842 if (SDValue Unpck =
10843 lowerVectorShuffleWithUNPCK(DL, MVT::v8i64, Mask, V1, V2, DAG))
10846 return lowerVectorShuffleWithPERMV(DL, MVT::v8i64, Mask, V1, V2, DAG);
10849 /// \brief Handle lowering of 16-lane 32-bit integer shuffles.
10850 static SDValue lowerV16I32VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10851 const X86Subtarget *Subtarget,
10852 SelectionDAG &DAG) {
10854 assert(V1.getSimpleValueType() == MVT::v16i32 && "Bad operand type!");
10855 assert(V2.getSimpleValueType() == MVT::v16i32 && "Bad operand type!");
10856 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10857 ArrayRef<int> Mask = SVOp->getMask();
10858 assert(Mask.size() == 16 && "Unexpected mask size for v16 shuffle!");
10860 if (SDValue Unpck =
10861 lowerVectorShuffleWithUNPCK(DL, MVT::v16i32, Mask, V1, V2, DAG))
10864 return lowerVectorShuffleWithPERMV(DL, MVT::v16i32, Mask, V1, V2, DAG);
10867 /// \brief Handle lowering of 32-lane 16-bit integer shuffles.
10868 static SDValue lowerV32I16VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10869 const X86Subtarget *Subtarget,
10870 SelectionDAG &DAG) {
10872 assert(V1.getSimpleValueType() == MVT::v32i16 && "Bad operand type!");
10873 assert(V2.getSimpleValueType() == MVT::v32i16 && "Bad operand type!");
10874 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10875 ArrayRef<int> Mask = SVOp->getMask();
10876 assert(Mask.size() == 32 && "Unexpected mask size for v32 shuffle!");
10877 assert(Subtarget->hasBWI() && "We can only lower v32i16 with AVX-512-BWI!");
10879 return lowerVectorShuffleWithPERMV(DL, MVT::v32i16, Mask, V1, V2, DAG);
10882 /// \brief Handle lowering of 64-lane 8-bit integer shuffles.
10883 static SDValue lowerV64I8VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10884 const X86Subtarget *Subtarget,
10885 SelectionDAG &DAG) {
10887 assert(V1.getSimpleValueType() == MVT::v64i8 && "Bad operand type!");
10888 assert(V2.getSimpleValueType() == MVT::v64i8 && "Bad operand type!");
10889 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10890 ArrayRef<int> Mask = SVOp->getMask();
10891 assert(Mask.size() == 64 && "Unexpected mask size for v64 shuffle!");
10892 assert(Subtarget->hasBWI() && "We can only lower v64i8 with AVX-512-BWI!");
10894 // FIXME: Implement direct support for this type!
10895 return splitAndLowerVectorShuffle(DL, MVT::v64i8, V1, V2, Mask, DAG);
10898 /// \brief High-level routine to lower various 512-bit x86 vector shuffles.
10900 /// This routine either breaks down the specific type of a 512-bit x86 vector
10901 /// shuffle or splits it into two 256-bit shuffles and fuses the results back
10902 /// together based on the available instructions.
10903 static SDValue lower512BitVectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10904 MVT VT, const X86Subtarget *Subtarget,
10905 SelectionDAG &DAG) {
10907 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10908 ArrayRef<int> Mask = SVOp->getMask();
10909 assert(Subtarget->hasAVX512() &&
10910 "Cannot lower 512-bit vectors w/ basic ISA!");
10912 // Check for being able to broadcast a single element.
10913 if (SDValue Broadcast =
10914 lowerVectorShuffleAsBroadcast(DL, VT, V1, Mask, Subtarget, DAG))
10917 // Dispatch to each element type for lowering. If we don't have supprot for
10918 // specific element type shuffles at 512 bits, immediately split them and
10919 // lower them. Each lowering routine of a given type is allowed to assume that
10920 // the requisite ISA extensions for that element type are available.
10921 switch (VT.SimpleTy) {
10923 return lowerV8F64VectorShuffle(Op, V1, V2, Subtarget, DAG);
10925 return lowerV16F32VectorShuffle(Op, V1, V2, Subtarget, DAG);
10927 return lowerV8I64VectorShuffle(Op, V1, V2, Subtarget, DAG);
10929 return lowerV16I32VectorShuffle(Op, V1, V2, Subtarget, DAG);
10931 if (Subtarget->hasBWI())
10932 return lowerV32I16VectorShuffle(Op, V1, V2, Subtarget, DAG);
10935 if (Subtarget->hasBWI())
10936 return lowerV64I8VectorShuffle(Op, V1, V2, Subtarget, DAG);
10940 llvm_unreachable("Not a valid 512-bit x86 vector type!");
10943 // Otherwise fall back on splitting.
10944 return splitAndLowerVectorShuffle(DL, VT, V1, V2, Mask, DAG);
10947 // Lower vXi1 vector shuffles.
10948 // There is no a dedicated instruction on AVX-512 that shuffles the masks.
10949 // The only way to shuffle bits is to sign-extend the mask vector to SIMD
10950 // vector, shuffle and then truncate it back.
10951 static SDValue lower1BitVectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10952 MVT VT, const X86Subtarget *Subtarget,
10953 SelectionDAG &DAG) {
10955 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10956 ArrayRef<int> Mask = SVOp->getMask();
10957 assert(Subtarget->hasAVX512() &&
10958 "Cannot lower 512-bit vectors w/o basic ISA!");
10960 switch (VT.SimpleTy) {
10962 llvm_unreachable("Expected a vector of i1 elements");
10964 ExtVT = MVT::v2i64;
10967 ExtVT = MVT::v4i32;
10970 ExtVT = MVT::v8i64; // Take 512-bit type, more shuffles on KNL
10973 ExtVT = MVT::v16i32;
10976 ExtVT = MVT::v32i16;
10979 ExtVT = MVT::v64i8;
10983 if (ISD::isBuildVectorAllZeros(V1.getNode()))
10984 V1 = getZeroVector(ExtVT, Subtarget, DAG, DL);
10985 else if (ISD::isBuildVectorAllOnes(V1.getNode()))
10986 V1 = getOnesVector(ExtVT, Subtarget, DAG, DL);
10988 V1 = DAG.getNode(ISD::SIGN_EXTEND, DL, ExtVT, V1);
10991 V2 = DAG.getUNDEF(ExtVT);
10992 else if (ISD::isBuildVectorAllZeros(V2.getNode()))
10993 V2 = getZeroVector(ExtVT, Subtarget, DAG, DL);
10994 else if (ISD::isBuildVectorAllOnes(V2.getNode()))
10995 V2 = getOnesVector(ExtVT, Subtarget, DAG, DL);
10997 V2 = DAG.getNode(ISD::SIGN_EXTEND, DL, ExtVT, V2);
10998 return DAG.getNode(ISD::TRUNCATE, DL, VT,
10999 DAG.getVectorShuffle(ExtVT, DL, V1, V2, Mask));
11001 /// \brief Top-level lowering for x86 vector shuffles.
11003 /// This handles decomposition, canonicalization, and lowering of all x86
11004 /// vector shuffles. Most of the specific lowering strategies are encapsulated
11005 /// above in helper routines. The canonicalization attempts to widen shuffles
11006 /// to involve fewer lanes of wider elements, consolidate symmetric patterns
11007 /// s.t. only one of the two inputs needs to be tested, etc.
11008 static SDValue lowerVectorShuffle(SDValue Op, const X86Subtarget *Subtarget,
11009 SelectionDAG &DAG) {
11010 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
11011 ArrayRef<int> Mask = SVOp->getMask();
11012 SDValue V1 = Op.getOperand(0);
11013 SDValue V2 = Op.getOperand(1);
11014 MVT VT = Op.getSimpleValueType();
11015 int NumElements = VT.getVectorNumElements();
11017 bool Is1BitVector = (VT.getScalarType() == MVT::i1);
11019 assert((VT.getSizeInBits() != 64 || Is1BitVector) &&
11020 "Can't lower MMX shuffles");
11022 bool V1IsUndef = V1.getOpcode() == ISD::UNDEF;
11023 bool V2IsUndef = V2.getOpcode() == ISD::UNDEF;
11024 if (V1IsUndef && V2IsUndef)
11025 return DAG.getUNDEF(VT);
11027 // When we create a shuffle node we put the UNDEF node to second operand,
11028 // but in some cases the first operand may be transformed to UNDEF.
11029 // In this case we should just commute the node.
11031 return DAG.getCommutedVectorShuffle(*SVOp);
11033 // Check for non-undef masks pointing at an undef vector and make the masks
11034 // undef as well. This makes it easier to match the shuffle based solely on
11038 if (M >= NumElements) {
11039 SmallVector<int, 8> NewMask(Mask.begin(), Mask.end());
11040 for (int &M : NewMask)
11041 if (M >= NumElements)
11043 return DAG.getVectorShuffle(VT, dl, V1, V2, NewMask);
11046 // We actually see shuffles that are entirely re-arrangements of a set of
11047 // zero inputs. This mostly happens while decomposing complex shuffles into
11048 // simple ones. Directly lower these as a buildvector of zeros.
11049 SmallBitVector Zeroable = computeZeroableShuffleElements(Mask, V1, V2);
11050 if (Zeroable.all())
11051 return getZeroVector(VT, Subtarget, DAG, dl);
11053 // Try to collapse shuffles into using a vector type with fewer elements but
11054 // wider element types. We cap this to not form integers or floating point
11055 // elements wider than 64 bits, but it might be interesting to form i128
11056 // integers to handle flipping the low and high halves of AVX 256-bit vectors.
11057 SmallVector<int, 16> WidenedMask;
11058 if (VT.getScalarSizeInBits() < 64 && !Is1BitVector &&
11059 canWidenShuffleElements(Mask, WidenedMask)) {
11060 MVT NewEltVT = VT.isFloatingPoint()
11061 ? MVT::getFloatingPointVT(VT.getScalarSizeInBits() * 2)
11062 : MVT::getIntegerVT(VT.getScalarSizeInBits() * 2);
11063 MVT NewVT = MVT::getVectorVT(NewEltVT, VT.getVectorNumElements() / 2);
11064 // Make sure that the new vector type is legal. For example, v2f64 isn't
11066 if (DAG.getTargetLoweringInfo().isTypeLegal(NewVT)) {
11067 V1 = DAG.getBitcast(NewVT, V1);
11068 V2 = DAG.getBitcast(NewVT, V2);
11069 return DAG.getBitcast(
11070 VT, DAG.getVectorShuffle(NewVT, dl, V1, V2, WidenedMask));
11074 int NumV1Elements = 0, NumUndefElements = 0, NumV2Elements = 0;
11075 for (int M : SVOp->getMask())
11077 ++NumUndefElements;
11078 else if (M < NumElements)
11083 // Commute the shuffle as needed such that more elements come from V1 than
11084 // V2. This allows us to match the shuffle pattern strictly on how many
11085 // elements come from V1 without handling the symmetric cases.
11086 if (NumV2Elements > NumV1Elements)
11087 return DAG.getCommutedVectorShuffle(*SVOp);
11089 // When the number of V1 and V2 elements are the same, try to minimize the
11090 // number of uses of V2 in the low half of the vector. When that is tied,
11091 // ensure that the sum of indices for V1 is equal to or lower than the sum
11092 // indices for V2. When those are equal, try to ensure that the number of odd
11093 // indices for V1 is lower than the number of odd indices for V2.
11094 if (NumV1Elements == NumV2Elements) {
11095 int LowV1Elements = 0, LowV2Elements = 0;
11096 for (int M : SVOp->getMask().slice(0, NumElements / 2))
11097 if (M >= NumElements)
11101 if (LowV2Elements > LowV1Elements) {
11102 return DAG.getCommutedVectorShuffle(*SVOp);
11103 } else if (LowV2Elements == LowV1Elements) {
11104 int SumV1Indices = 0, SumV2Indices = 0;
11105 for (int i = 0, Size = SVOp->getMask().size(); i < Size; ++i)
11106 if (SVOp->getMask()[i] >= NumElements)
11108 else if (SVOp->getMask()[i] >= 0)
11110 if (SumV2Indices < SumV1Indices) {
11111 return DAG.getCommutedVectorShuffle(*SVOp);
11112 } else if (SumV2Indices == SumV1Indices) {
11113 int NumV1OddIndices = 0, NumV2OddIndices = 0;
11114 for (int i = 0, Size = SVOp->getMask().size(); i < Size; ++i)
11115 if (SVOp->getMask()[i] >= NumElements)
11116 NumV2OddIndices += i % 2;
11117 else if (SVOp->getMask()[i] >= 0)
11118 NumV1OddIndices += i % 2;
11119 if (NumV2OddIndices < NumV1OddIndices)
11120 return DAG.getCommutedVectorShuffle(*SVOp);
11125 // For each vector width, delegate to a specialized lowering routine.
11126 if (VT.getSizeInBits() == 128)
11127 return lower128BitVectorShuffle(Op, V1, V2, VT, Subtarget, DAG);
11129 if (VT.getSizeInBits() == 256)
11130 return lower256BitVectorShuffle(Op, V1, V2, VT, Subtarget, DAG);
11132 if (VT.getSizeInBits() == 512)
11133 return lower512BitVectorShuffle(Op, V1, V2, VT, Subtarget, DAG);
11136 return lower1BitVectorShuffle(Op, V1, V2, VT, Subtarget, DAG);
11137 llvm_unreachable("Unimplemented!");
11140 // This function assumes its argument is a BUILD_VECTOR of constants or
11141 // undef SDNodes. i.e: ISD::isBuildVectorOfConstantSDNodes(BuildVector) is
11143 static bool BUILD_VECTORtoBlendMask(BuildVectorSDNode *BuildVector,
11144 unsigned &MaskValue) {
11146 unsigned NumElems = BuildVector->getNumOperands();
11148 // There are 2 lanes if (NumElems > 8), and 1 lane otherwise.
11149 // We don't handle the >2 lanes case right now.
11150 unsigned NumLanes = (NumElems - 1) / 8 + 1;
11154 unsigned NumElemsInLane = NumElems / NumLanes;
11156 // Blend for v16i16 should be symmetric for the both lanes.
11157 for (unsigned i = 0; i < NumElemsInLane; ++i) {
11158 SDValue EltCond = BuildVector->getOperand(i);
11159 SDValue SndLaneEltCond =
11160 (NumLanes == 2) ? BuildVector->getOperand(i + NumElemsInLane) : EltCond;
11162 int Lane1Cond = -1, Lane2Cond = -1;
11163 if (isa<ConstantSDNode>(EltCond))
11164 Lane1Cond = !isZero(EltCond);
11165 if (isa<ConstantSDNode>(SndLaneEltCond))
11166 Lane2Cond = !isZero(SndLaneEltCond);
11168 unsigned LaneMask = 0;
11169 if (Lane1Cond == Lane2Cond || Lane2Cond < 0)
11170 // Lane1Cond != 0, means we want the first argument.
11171 // Lane1Cond == 0, means we want the second argument.
11172 // The encoding of this argument is 0 for the first argument, 1
11173 // for the second. Therefore, invert the condition.
11174 LaneMask = !Lane1Cond << i;
11175 else if (Lane1Cond < 0)
11176 LaneMask = !Lane2Cond << i;
11180 MaskValue |= LaneMask;
11182 MaskValue |= LaneMask << NumElemsInLane;
11187 /// \brief Try to lower a VSELECT instruction to a vector shuffle.
11188 static SDValue lowerVSELECTtoVectorShuffle(SDValue Op,
11189 const X86Subtarget *Subtarget,
11190 SelectionDAG &DAG) {
11191 SDValue Cond = Op.getOperand(0);
11192 SDValue LHS = Op.getOperand(1);
11193 SDValue RHS = Op.getOperand(2);
11195 MVT VT = Op.getSimpleValueType();
11197 if (!ISD::isBuildVectorOfConstantSDNodes(Cond.getNode()))
11199 auto *CondBV = cast<BuildVectorSDNode>(Cond);
11201 // Only non-legal VSELECTs reach this lowering, convert those into generic
11202 // shuffles and re-use the shuffle lowering path for blends.
11203 SmallVector<int, 32> Mask;
11204 for (int i = 0, Size = VT.getVectorNumElements(); i < Size; ++i) {
11205 SDValue CondElt = CondBV->getOperand(i);
11207 isa<ConstantSDNode>(CondElt) ? i + (isZero(CondElt) ? Size : 0) : -1);
11209 return DAG.getVectorShuffle(VT, dl, LHS, RHS, Mask);
11212 SDValue X86TargetLowering::LowerVSELECT(SDValue Op, SelectionDAG &DAG) const {
11213 // A vselect where all conditions and data are constants can be optimized into
11214 // a single vector load by SelectionDAGLegalize::ExpandBUILD_VECTOR().
11215 if (ISD::isBuildVectorOfConstantSDNodes(Op.getOperand(0).getNode()) &&
11216 ISD::isBuildVectorOfConstantSDNodes(Op.getOperand(1).getNode()) &&
11217 ISD::isBuildVectorOfConstantSDNodes(Op.getOperand(2).getNode()))
11220 // Try to lower this to a blend-style vector shuffle. This can handle all
11221 // constant condition cases.
11222 if (SDValue BlendOp = lowerVSELECTtoVectorShuffle(Op, Subtarget, DAG))
11225 // Variable blends are only legal from SSE4.1 onward.
11226 if (!Subtarget->hasSSE41())
11229 // Only some types will be legal on some subtargets. If we can emit a legal
11230 // VSELECT-matching blend, return Op, and but if we need to expand, return
11232 switch (Op.getSimpleValueType().SimpleTy) {
11234 // Most of the vector types have blends past SSE4.1.
11238 // The byte blends for AVX vectors were introduced only in AVX2.
11239 if (Subtarget->hasAVX2())
11246 // AVX-512 BWI and VLX features support VSELECT with i16 elements.
11247 if (Subtarget->hasBWI() && Subtarget->hasVLX())
11250 // FIXME: We should custom lower this by fixing the condition and using i8
11256 static SDValue LowerEXTRACT_VECTOR_ELT_SSE4(SDValue Op, SelectionDAG &DAG) {
11257 MVT VT = Op.getSimpleValueType();
11260 if (!Op.getOperand(0).getSimpleValueType().is128BitVector())
11263 if (VT.getSizeInBits() == 8) {
11264 SDValue Extract = DAG.getNode(X86ISD::PEXTRB, dl, MVT::i32,
11265 Op.getOperand(0), Op.getOperand(1));
11266 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Extract,
11267 DAG.getValueType(VT));
11268 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
11271 if (VT.getSizeInBits() == 16) {
11272 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
11273 // If Idx is 0, it's cheaper to do a move instead of a pextrw.
11275 return DAG.getNode(
11276 ISD::TRUNCATE, dl, MVT::i16,
11277 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
11278 DAG.getBitcast(MVT::v4i32, Op.getOperand(0)),
11279 Op.getOperand(1)));
11280 SDValue Extract = DAG.getNode(X86ISD::PEXTRW, dl, MVT::i32,
11281 Op.getOperand(0), Op.getOperand(1));
11282 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Extract,
11283 DAG.getValueType(VT));
11284 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
11287 if (VT == MVT::f32) {
11288 // EXTRACTPS outputs to a GPR32 register which will require a movd to copy
11289 // the result back to FR32 register. It's only worth matching if the
11290 // result has a single use which is a store or a bitcast to i32. And in
11291 // the case of a store, it's not worth it if the index is a constant 0,
11292 // because a MOVSSmr can be used instead, which is smaller and faster.
11293 if (!Op.hasOneUse())
11295 SDNode *User = *Op.getNode()->use_begin();
11296 if ((User->getOpcode() != ISD::STORE ||
11297 (isa<ConstantSDNode>(Op.getOperand(1)) &&
11298 cast<ConstantSDNode>(Op.getOperand(1))->isNullValue())) &&
11299 (User->getOpcode() != ISD::BITCAST ||
11300 User->getValueType(0) != MVT::i32))
11302 SDValue Extract = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
11303 DAG.getBitcast(MVT::v4i32, Op.getOperand(0)),
11305 return DAG.getBitcast(MVT::f32, Extract);
11308 if (VT == MVT::i32 || VT == MVT::i64) {
11309 // ExtractPS/pextrq works with constant index.
11310 if (isa<ConstantSDNode>(Op.getOperand(1)))
11316 /// Extract one bit from mask vector, like v16i1 or v8i1.
11317 /// AVX-512 feature.
11319 X86TargetLowering::ExtractBitFromMaskVector(SDValue Op, SelectionDAG &DAG) const {
11320 SDValue Vec = Op.getOperand(0);
11322 MVT VecVT = Vec.getSimpleValueType();
11323 SDValue Idx = Op.getOperand(1);
11324 MVT EltVT = Op.getSimpleValueType();
11326 assert((EltVT == MVT::i1) && "Unexpected operands in ExtractBitFromMaskVector");
11327 assert((VecVT.getVectorNumElements() <= 16 || Subtarget->hasBWI()) &&
11328 "Unexpected vector type in ExtractBitFromMaskVector");
11330 // variable index can't be handled in mask registers,
11331 // extend vector to VR512
11332 if (!isa<ConstantSDNode>(Idx)) {
11333 MVT ExtVT = (VecVT == MVT::v8i1 ? MVT::v8i64 : MVT::v16i32);
11334 SDValue Ext = DAG.getNode(ISD::ZERO_EXTEND, dl, ExtVT, Vec);
11335 SDValue Elt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl,
11336 ExtVT.getVectorElementType(), Ext, Idx);
11337 return DAG.getNode(ISD::TRUNCATE, dl, EltVT, Elt);
11340 unsigned IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
11341 const TargetRegisterClass* rc = getRegClassFor(VecVT);
11342 if (!Subtarget->hasDQI() && (VecVT.getVectorNumElements() <= 8))
11343 rc = getRegClassFor(MVT::v16i1);
11344 unsigned MaxSift = rc->getSize()*8 - 1;
11345 Vec = DAG.getNode(X86ISD::VSHLI, dl, VecVT, Vec,
11346 DAG.getConstant(MaxSift - IdxVal, dl, MVT::i8));
11347 Vec = DAG.getNode(X86ISD::VSRLI, dl, VecVT, Vec,
11348 DAG.getConstant(MaxSift, dl, MVT::i8));
11349 return DAG.getNode(X86ISD::VEXTRACT, dl, MVT::i1, Vec,
11350 DAG.getIntPtrConstant(0, dl));
11354 X86TargetLowering::LowerEXTRACT_VECTOR_ELT(SDValue Op,
11355 SelectionDAG &DAG) const {
11357 SDValue Vec = Op.getOperand(0);
11358 MVT VecVT = Vec.getSimpleValueType();
11359 SDValue Idx = Op.getOperand(1);
11361 if (Op.getSimpleValueType() == MVT::i1)
11362 return ExtractBitFromMaskVector(Op, DAG);
11364 if (!isa<ConstantSDNode>(Idx)) {
11365 if (VecVT.is512BitVector() ||
11366 (VecVT.is256BitVector() && Subtarget->hasInt256() &&
11367 VecVT.getVectorElementType().getSizeInBits() == 32)) {
11370 MVT::getIntegerVT(VecVT.getVectorElementType().getSizeInBits());
11371 MVT MaskVT = MVT::getVectorVT(MaskEltVT, VecVT.getSizeInBits() /
11372 MaskEltVT.getSizeInBits());
11374 Idx = DAG.getZExtOrTrunc(Idx, dl, MaskEltVT);
11375 auto PtrVT = getPointerTy(DAG.getDataLayout());
11376 SDValue Mask = DAG.getNode(X86ISD::VINSERT, dl, MaskVT,
11377 getZeroVector(MaskVT, Subtarget, DAG, dl), Idx,
11378 DAG.getConstant(0, dl, PtrVT));
11379 SDValue Perm = DAG.getNode(X86ISD::VPERMV, dl, VecVT, Mask, Vec);
11380 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, Op.getValueType(), Perm,
11381 DAG.getConstant(0, dl, PtrVT));
11386 // If this is a 256-bit vector result, first extract the 128-bit vector and
11387 // then extract the element from the 128-bit vector.
11388 if (VecVT.is256BitVector() || VecVT.is512BitVector()) {
11390 unsigned IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
11391 // Get the 128-bit vector.
11392 Vec = Extract128BitVector(Vec, IdxVal, DAG, dl);
11393 MVT EltVT = VecVT.getVectorElementType();
11395 unsigned ElemsPerChunk = 128 / EltVT.getSizeInBits();
11397 //if (IdxVal >= NumElems/2)
11398 // IdxVal -= NumElems/2;
11399 IdxVal -= (IdxVal/ElemsPerChunk)*ElemsPerChunk;
11400 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, Op.getValueType(), Vec,
11401 DAG.getConstant(IdxVal, dl, MVT::i32));
11404 assert(VecVT.is128BitVector() && "Unexpected vector length");
11406 if (Subtarget->hasSSE41())
11407 if (SDValue Res = LowerEXTRACT_VECTOR_ELT_SSE4(Op, DAG))
11410 MVT VT = Op.getSimpleValueType();
11411 // TODO: handle v16i8.
11412 if (VT.getSizeInBits() == 16) {
11413 SDValue Vec = Op.getOperand(0);
11414 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
11416 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i16,
11417 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
11418 DAG.getBitcast(MVT::v4i32, Vec),
11419 Op.getOperand(1)));
11420 // Transform it so it match pextrw which produces a 32-bit result.
11421 MVT EltVT = MVT::i32;
11422 SDValue Extract = DAG.getNode(X86ISD::PEXTRW, dl, EltVT,
11423 Op.getOperand(0), Op.getOperand(1));
11424 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, EltVT, Extract,
11425 DAG.getValueType(VT));
11426 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
11429 if (VT.getSizeInBits() == 32) {
11430 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
11434 // SHUFPS the element to the lowest double word, then movss.
11435 int Mask[4] = { static_cast<int>(Idx), -1, -1, -1 };
11436 MVT VVT = Op.getOperand(0).getSimpleValueType();
11437 SDValue Vec = DAG.getVectorShuffle(VVT, dl, Op.getOperand(0),
11438 DAG.getUNDEF(VVT), Mask);
11439 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, Vec,
11440 DAG.getIntPtrConstant(0, dl));
11443 if (VT.getSizeInBits() == 64) {
11444 // FIXME: .td only matches this for <2 x f64>, not <2 x i64> on 32b
11445 // FIXME: seems like this should be unnecessary if mov{h,l}pd were taught
11446 // to match extract_elt for f64.
11447 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
11451 // UNPCKHPD the element to the lowest double word, then movsd.
11452 // Note if the lower 64 bits of the result of the UNPCKHPD is then stored
11453 // to a f64mem, the whole operation is folded into a single MOVHPDmr.
11454 int Mask[2] = { 1, -1 };
11455 MVT VVT = Op.getOperand(0).getSimpleValueType();
11456 SDValue Vec = DAG.getVectorShuffle(VVT, dl, Op.getOperand(0),
11457 DAG.getUNDEF(VVT), Mask);
11458 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, Vec,
11459 DAG.getIntPtrConstant(0, dl));
11465 /// Insert one bit to mask vector, like v16i1 or v8i1.
11466 /// AVX-512 feature.
11468 X86TargetLowering::InsertBitToMaskVector(SDValue Op, SelectionDAG &DAG) const {
11470 SDValue Vec = Op.getOperand(0);
11471 SDValue Elt = Op.getOperand(1);
11472 SDValue Idx = Op.getOperand(2);
11473 MVT VecVT = Vec.getSimpleValueType();
11475 if (!isa<ConstantSDNode>(Idx)) {
11476 // Non constant index. Extend source and destination,
11477 // insert element and then truncate the result.
11478 MVT ExtVecVT = (VecVT == MVT::v8i1 ? MVT::v8i64 : MVT::v16i32);
11479 MVT ExtEltVT = (VecVT == MVT::v8i1 ? MVT::i64 : MVT::i32);
11480 SDValue ExtOp = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, ExtVecVT,
11481 DAG.getNode(ISD::ZERO_EXTEND, dl, ExtVecVT, Vec),
11482 DAG.getNode(ISD::ZERO_EXTEND, dl, ExtEltVT, Elt), Idx);
11483 return DAG.getNode(ISD::TRUNCATE, dl, VecVT, ExtOp);
11486 unsigned IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
11487 SDValue EltInVec = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VecVT, Elt);
11489 EltInVec = DAG.getNode(X86ISD::VSHLI, dl, VecVT, EltInVec,
11490 DAG.getConstant(IdxVal, dl, MVT::i8));
11491 if (Vec.getOpcode() == ISD::UNDEF)
11493 return DAG.getNode(ISD::OR, dl, VecVT, Vec, EltInVec);
11496 SDValue X86TargetLowering::LowerINSERT_VECTOR_ELT(SDValue Op,
11497 SelectionDAG &DAG) const {
11498 MVT VT = Op.getSimpleValueType();
11499 MVT EltVT = VT.getVectorElementType();
11501 if (EltVT == MVT::i1)
11502 return InsertBitToMaskVector(Op, DAG);
11505 SDValue N0 = Op.getOperand(0);
11506 SDValue N1 = Op.getOperand(1);
11507 SDValue N2 = Op.getOperand(2);
11508 if (!isa<ConstantSDNode>(N2))
11510 auto *N2C = cast<ConstantSDNode>(N2);
11511 unsigned IdxVal = N2C->getZExtValue();
11513 // If the vector is wider than 128 bits, extract the 128-bit subvector, insert
11514 // into that, and then insert the subvector back into the result.
11515 if (VT.is256BitVector() || VT.is512BitVector()) {
11516 // With a 256-bit vector, we can insert into the zero element efficiently
11517 // using a blend if we have AVX or AVX2 and the right data type.
11518 if (VT.is256BitVector() && IdxVal == 0) {
11519 // TODO: It is worthwhile to cast integer to floating point and back
11520 // and incur a domain crossing penalty if that's what we'll end up
11521 // doing anyway after extracting to a 128-bit vector.
11522 if ((Subtarget->hasAVX() && (EltVT == MVT::f64 || EltVT == MVT::f32)) ||
11523 (Subtarget->hasAVX2() && EltVT == MVT::i32)) {
11524 SDValue N1Vec = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, N1);
11525 N2 = DAG.getIntPtrConstant(1, dl);
11526 return DAG.getNode(X86ISD::BLENDI, dl, VT, N0, N1Vec, N2);
11530 // Get the desired 128-bit vector chunk.
11531 SDValue V = Extract128BitVector(N0, IdxVal, DAG, dl);
11533 // Insert the element into the desired chunk.
11534 unsigned NumEltsIn128 = 128 / EltVT.getSizeInBits();
11535 unsigned IdxIn128 = IdxVal - (IdxVal / NumEltsIn128) * NumEltsIn128;
11537 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, V.getValueType(), V, N1,
11538 DAG.getConstant(IdxIn128, dl, MVT::i32));
11540 // Insert the changed part back into the bigger vector
11541 return Insert128BitVector(N0, V, IdxVal, DAG, dl);
11543 assert(VT.is128BitVector() && "Only 128-bit vector types should be left!");
11545 if (Subtarget->hasSSE41()) {
11546 if (EltVT.getSizeInBits() == 8 || EltVT.getSizeInBits() == 16) {
11548 if (VT == MVT::v8i16) {
11549 Opc = X86ISD::PINSRW;
11551 assert(VT == MVT::v16i8);
11552 Opc = X86ISD::PINSRB;
11555 // Transform it so it match pinsr{b,w} which expects a GR32 as its second
11557 if (N1.getValueType() != MVT::i32)
11558 N1 = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, N1);
11559 if (N2.getValueType() != MVT::i32)
11560 N2 = DAG.getIntPtrConstant(IdxVal, dl);
11561 return DAG.getNode(Opc, dl, VT, N0, N1, N2);
11564 if (EltVT == MVT::f32) {
11565 // Bits [7:6] of the constant are the source select. This will always be
11566 // zero here. The DAG Combiner may combine an extract_elt index into
11567 // these bits. For example (insert (extract, 3), 2) could be matched by
11568 // putting the '3' into bits [7:6] of X86ISD::INSERTPS.
11569 // Bits [5:4] of the constant are the destination select. This is the
11570 // value of the incoming immediate.
11571 // Bits [3:0] of the constant are the zero mask. The DAG Combiner may
11572 // combine either bitwise AND or insert of float 0.0 to set these bits.
11574 bool MinSize = DAG.getMachineFunction().getFunction()->optForMinSize();
11575 if (IdxVal == 0 && (!MinSize || !MayFoldLoad(N1))) {
11576 // If this is an insertion of 32-bits into the low 32-bits of
11577 // a vector, we prefer to generate a blend with immediate rather
11578 // than an insertps. Blends are simpler operations in hardware and so
11579 // will always have equal or better performance than insertps.
11580 // But if optimizing for size and there's a load folding opportunity,
11581 // generate insertps because blendps does not have a 32-bit memory
11583 N2 = DAG.getIntPtrConstant(1, dl);
11584 N1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4f32, N1);
11585 return DAG.getNode(X86ISD::BLENDI, dl, VT, N0, N1, N2);
11587 N2 = DAG.getIntPtrConstant(IdxVal << 4, dl);
11588 // Create this as a scalar to vector..
11589 N1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4f32, N1);
11590 return DAG.getNode(X86ISD::INSERTPS, dl, VT, N0, N1, N2);
11593 if (EltVT == MVT::i32 || EltVT == MVT::i64) {
11594 // PINSR* works with constant index.
11599 if (EltVT == MVT::i8)
11602 if (EltVT.getSizeInBits() == 16) {
11603 // Transform it so it match pinsrw which expects a 16-bit value in a GR32
11604 // as its second argument.
11605 if (N1.getValueType() != MVT::i32)
11606 N1 = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, N1);
11607 if (N2.getValueType() != MVT::i32)
11608 N2 = DAG.getIntPtrConstant(IdxVal, dl);
11609 return DAG.getNode(X86ISD::PINSRW, dl, VT, N0, N1, N2);
11614 static SDValue LowerSCALAR_TO_VECTOR(SDValue Op, SelectionDAG &DAG) {
11616 MVT OpVT = Op.getSimpleValueType();
11618 // If this is a 256-bit vector result, first insert into a 128-bit
11619 // vector and then insert into the 256-bit vector.
11620 if (!OpVT.is128BitVector()) {
11621 // Insert into a 128-bit vector.
11622 unsigned SizeFactor = OpVT.getSizeInBits()/128;
11623 MVT VT128 = MVT::getVectorVT(OpVT.getVectorElementType(),
11624 OpVT.getVectorNumElements() / SizeFactor);
11626 Op = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT128, Op.getOperand(0));
11628 // Insert the 128-bit vector.
11629 return Insert128BitVector(DAG.getUNDEF(OpVT), Op, 0, DAG, dl);
11632 if (OpVT == MVT::v1i64 &&
11633 Op.getOperand(0).getValueType() == MVT::i64)
11634 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v1i64, Op.getOperand(0));
11636 SDValue AnyExt = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, Op.getOperand(0));
11637 assert(OpVT.is128BitVector() && "Expected an SSE type!");
11638 return DAG.getBitcast(
11639 OpVT, DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32, AnyExt));
11642 // Lower a node with an EXTRACT_SUBVECTOR opcode. This may result in
11643 // a simple subregister reference or explicit instructions to grab
11644 // upper bits of a vector.
11645 static SDValue LowerEXTRACT_SUBVECTOR(SDValue Op, const X86Subtarget *Subtarget,
11646 SelectionDAG &DAG) {
11648 SDValue In = Op.getOperand(0);
11649 SDValue Idx = Op.getOperand(1);
11650 unsigned IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
11651 MVT ResVT = Op.getSimpleValueType();
11652 MVT InVT = In.getSimpleValueType();
11654 if (Subtarget->hasFp256()) {
11655 if (ResVT.is128BitVector() &&
11656 (InVT.is256BitVector() || InVT.is512BitVector()) &&
11657 isa<ConstantSDNode>(Idx)) {
11658 return Extract128BitVector(In, IdxVal, DAG, dl);
11660 if (ResVT.is256BitVector() && InVT.is512BitVector() &&
11661 isa<ConstantSDNode>(Idx)) {
11662 return Extract256BitVector(In, IdxVal, DAG, dl);
11668 // Lower a node with an INSERT_SUBVECTOR opcode. This may result in a
11669 // simple superregister reference or explicit instructions to insert
11670 // the upper bits of a vector.
11671 static SDValue LowerINSERT_SUBVECTOR(SDValue Op, const X86Subtarget *Subtarget,
11672 SelectionDAG &DAG) {
11673 if (!Subtarget->hasAVX())
11677 SDValue Vec = Op.getOperand(0);
11678 SDValue SubVec = Op.getOperand(1);
11679 SDValue Idx = Op.getOperand(2);
11681 if (!isa<ConstantSDNode>(Idx))
11684 unsigned IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
11685 MVT OpVT = Op.getSimpleValueType();
11686 MVT SubVecVT = SubVec.getSimpleValueType();
11688 // Fold two 16-byte subvector loads into one 32-byte load:
11689 // (insert_subvector (insert_subvector undef, (load addr), 0),
11690 // (load addr + 16), Elts/2)
11692 if ((IdxVal == OpVT.getVectorNumElements() / 2) &&
11693 Vec.getOpcode() == ISD::INSERT_SUBVECTOR &&
11694 OpVT.is256BitVector() && SubVecVT.is128BitVector()) {
11695 auto *Idx2 = dyn_cast<ConstantSDNode>(Vec.getOperand(2));
11696 if (Idx2 && Idx2->getZExtValue() == 0) {
11697 SDValue SubVec2 = Vec.getOperand(1);
11698 // If needed, look through a bitcast to get to the load.
11699 if (SubVec2.getNode() && SubVec2.getOpcode() == ISD::BITCAST)
11700 SubVec2 = SubVec2.getOperand(0);
11702 if (auto *FirstLd = dyn_cast<LoadSDNode>(SubVec2)) {
11704 unsigned Alignment = FirstLd->getAlignment();
11705 unsigned AS = FirstLd->getAddressSpace();
11706 const X86TargetLowering *TLI = Subtarget->getTargetLowering();
11707 if (TLI->allowsMemoryAccess(*DAG.getContext(), DAG.getDataLayout(),
11708 OpVT, AS, Alignment, &Fast) && Fast) {
11709 SDValue Ops[] = { SubVec2, SubVec };
11710 if (SDValue Ld = EltsFromConsecutiveLoads(OpVT, Ops, dl, DAG, false))
11717 if ((OpVT.is256BitVector() || OpVT.is512BitVector()) &&
11718 SubVecVT.is128BitVector())
11719 return Insert128BitVector(Vec, SubVec, IdxVal, DAG, dl);
11721 if (OpVT.is512BitVector() && SubVecVT.is256BitVector())
11722 return Insert256BitVector(Vec, SubVec, IdxVal, DAG, dl);
11724 if (OpVT.getVectorElementType() == MVT::i1) {
11725 if (IdxVal == 0 && Vec.getOpcode() == ISD::UNDEF) // the operation is legal
11727 SDValue ZeroIdx = DAG.getIntPtrConstant(0, dl);
11728 SDValue Undef = DAG.getUNDEF(OpVT);
11729 unsigned NumElems = OpVT.getVectorNumElements();
11730 SDValue ShiftBits = DAG.getConstant(NumElems/2, dl, MVT::i8);
11732 if (IdxVal == OpVT.getVectorNumElements() / 2) {
11733 // Zero upper bits of the Vec
11734 Vec = DAG.getNode(X86ISD::VSHLI, dl, OpVT, Vec, ShiftBits);
11735 Vec = DAG.getNode(X86ISD::VSRLI, dl, OpVT, Vec, ShiftBits);
11737 SDValue Vec2 = DAG.getNode(ISD::INSERT_SUBVECTOR, dl, OpVT, Undef,
11739 Vec2 = DAG.getNode(X86ISD::VSHLI, dl, OpVT, Vec2, ShiftBits);
11740 return DAG.getNode(ISD::OR, dl, OpVT, Vec, Vec2);
11743 SDValue Vec2 = DAG.getNode(ISD::INSERT_SUBVECTOR, dl, OpVT, Undef,
11745 // Zero upper bits of the Vec2
11746 Vec2 = DAG.getNode(X86ISD::VSHLI, dl, OpVT, Vec2, ShiftBits);
11747 Vec2 = DAG.getNode(X86ISD::VSRLI, dl, OpVT, Vec2, ShiftBits);
11748 // Zero lower bits of the Vec
11749 Vec = DAG.getNode(X86ISD::VSRLI, dl, OpVT, Vec, ShiftBits);
11750 Vec = DAG.getNode(X86ISD::VSHLI, dl, OpVT, Vec, ShiftBits);
11751 // Merge them together
11752 return DAG.getNode(ISD::OR, dl, OpVT, Vec, Vec2);
11758 // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
11759 // their target countpart wrapped in the X86ISD::Wrapper node. Suppose N is
11760 // one of the above mentioned nodes. It has to be wrapped because otherwise
11761 // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
11762 // be used to form addressing mode. These wrapped nodes will be selected
11765 X86TargetLowering::LowerConstantPool(SDValue Op, SelectionDAG &DAG) const {
11766 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
11768 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
11769 // global base reg.
11770 unsigned char OpFlag = 0;
11771 unsigned WrapperKind = X86ISD::Wrapper;
11772 CodeModel::Model M = DAG.getTarget().getCodeModel();
11774 if (Subtarget->isPICStyleRIPRel() &&
11775 (M == CodeModel::Small || M == CodeModel::Kernel))
11776 WrapperKind = X86ISD::WrapperRIP;
11777 else if (Subtarget->isPICStyleGOT())
11778 OpFlag = X86II::MO_GOTOFF;
11779 else if (Subtarget->isPICStyleStubPIC())
11780 OpFlag = X86II::MO_PIC_BASE_OFFSET;
11782 auto PtrVT = getPointerTy(DAG.getDataLayout());
11783 SDValue Result = DAG.getTargetConstantPool(
11784 CP->getConstVal(), PtrVT, CP->getAlignment(), CP->getOffset(), OpFlag);
11786 Result = DAG.getNode(WrapperKind, DL, PtrVT, Result);
11787 // With PIC, the address is actually $g + Offset.
11790 DAG.getNode(ISD::ADD, DL, PtrVT,
11791 DAG.getNode(X86ISD::GlobalBaseReg, SDLoc(), PtrVT), Result);
11797 SDValue X86TargetLowering::LowerJumpTable(SDValue Op, SelectionDAG &DAG) const {
11798 JumpTableSDNode *JT = cast<JumpTableSDNode>(Op);
11800 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
11801 // global base reg.
11802 unsigned char OpFlag = 0;
11803 unsigned WrapperKind = X86ISD::Wrapper;
11804 CodeModel::Model M = DAG.getTarget().getCodeModel();
11806 if (Subtarget->isPICStyleRIPRel() &&
11807 (M == CodeModel::Small || M == CodeModel::Kernel))
11808 WrapperKind = X86ISD::WrapperRIP;
11809 else if (Subtarget->isPICStyleGOT())
11810 OpFlag = X86II::MO_GOTOFF;
11811 else if (Subtarget->isPICStyleStubPIC())
11812 OpFlag = X86II::MO_PIC_BASE_OFFSET;
11814 auto PtrVT = getPointerTy(DAG.getDataLayout());
11815 SDValue Result = DAG.getTargetJumpTable(JT->getIndex(), PtrVT, OpFlag);
11817 Result = DAG.getNode(WrapperKind, DL, PtrVT, Result);
11819 // With PIC, the address is actually $g + Offset.
11822 DAG.getNode(ISD::ADD, DL, PtrVT,
11823 DAG.getNode(X86ISD::GlobalBaseReg, SDLoc(), PtrVT), Result);
11829 X86TargetLowering::LowerExternalSymbol(SDValue Op, SelectionDAG &DAG) const {
11830 const char *Sym = cast<ExternalSymbolSDNode>(Op)->getSymbol();
11832 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
11833 // global base reg.
11834 unsigned char OpFlag = 0;
11835 unsigned WrapperKind = X86ISD::Wrapper;
11836 CodeModel::Model M = DAG.getTarget().getCodeModel();
11838 if (Subtarget->isPICStyleRIPRel() &&
11839 (M == CodeModel::Small || M == CodeModel::Kernel)) {
11840 if (Subtarget->isTargetDarwin() || Subtarget->isTargetELF())
11841 OpFlag = X86II::MO_GOTPCREL;
11842 WrapperKind = X86ISD::WrapperRIP;
11843 } else if (Subtarget->isPICStyleGOT()) {
11844 OpFlag = X86II::MO_GOT;
11845 } else if (Subtarget->isPICStyleStubPIC()) {
11846 OpFlag = X86II::MO_DARWIN_NONLAZY_PIC_BASE;
11847 } else if (Subtarget->isPICStyleStubNoDynamic()) {
11848 OpFlag = X86II::MO_DARWIN_NONLAZY;
11851 auto PtrVT = getPointerTy(DAG.getDataLayout());
11852 SDValue Result = DAG.getTargetExternalSymbol(Sym, PtrVT, OpFlag);
11855 Result = DAG.getNode(WrapperKind, DL, PtrVT, Result);
11857 // With PIC, the address is actually $g + Offset.
11858 if (DAG.getTarget().getRelocationModel() == Reloc::PIC_ &&
11859 !Subtarget->is64Bit()) {
11861 DAG.getNode(ISD::ADD, DL, PtrVT,
11862 DAG.getNode(X86ISD::GlobalBaseReg, SDLoc(), PtrVT), Result);
11865 // For symbols that require a load from a stub to get the address, emit the
11867 if (isGlobalStubReference(OpFlag))
11868 Result = DAG.getLoad(PtrVT, DL, DAG.getEntryNode(), Result,
11869 MachinePointerInfo::getGOT(DAG.getMachineFunction()),
11870 false, false, false, 0);
11876 X86TargetLowering::LowerBlockAddress(SDValue Op, SelectionDAG &DAG) const {
11877 // Create the TargetBlockAddressAddress node.
11878 unsigned char OpFlags =
11879 Subtarget->ClassifyBlockAddressReference();
11880 CodeModel::Model M = DAG.getTarget().getCodeModel();
11881 const BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress();
11882 int64_t Offset = cast<BlockAddressSDNode>(Op)->getOffset();
11884 auto PtrVT = getPointerTy(DAG.getDataLayout());
11885 SDValue Result = DAG.getTargetBlockAddress(BA, PtrVT, Offset, OpFlags);
11887 if (Subtarget->isPICStyleRIPRel() &&
11888 (M == CodeModel::Small || M == CodeModel::Kernel))
11889 Result = DAG.getNode(X86ISD::WrapperRIP, dl, PtrVT, Result);
11891 Result = DAG.getNode(X86ISD::Wrapper, dl, PtrVT, Result);
11893 // With PIC, the address is actually $g + Offset.
11894 if (isGlobalRelativeToPICBase(OpFlags)) {
11895 Result = DAG.getNode(ISD::ADD, dl, PtrVT,
11896 DAG.getNode(X86ISD::GlobalBaseReg, dl, PtrVT), Result);
11903 X86TargetLowering::LowerGlobalAddress(const GlobalValue *GV, SDLoc dl,
11904 int64_t Offset, SelectionDAG &DAG) const {
11905 // Create the TargetGlobalAddress node, folding in the constant
11906 // offset if it is legal.
11907 unsigned char OpFlags =
11908 Subtarget->ClassifyGlobalReference(GV, DAG.getTarget());
11909 CodeModel::Model M = DAG.getTarget().getCodeModel();
11910 auto PtrVT = getPointerTy(DAG.getDataLayout());
11912 if (OpFlags == X86II::MO_NO_FLAG &&
11913 X86::isOffsetSuitableForCodeModel(Offset, M)) {
11914 // A direct static reference to a global.
11915 Result = DAG.getTargetGlobalAddress(GV, dl, PtrVT, Offset);
11918 Result = DAG.getTargetGlobalAddress(GV, dl, PtrVT, 0, OpFlags);
11921 if (Subtarget->isPICStyleRIPRel() &&
11922 (M == CodeModel::Small || M == CodeModel::Kernel))
11923 Result = DAG.getNode(X86ISD::WrapperRIP, dl, PtrVT, Result);
11925 Result = DAG.getNode(X86ISD::Wrapper, dl, PtrVT, Result);
11927 // With PIC, the address is actually $g + Offset.
11928 if (isGlobalRelativeToPICBase(OpFlags)) {
11929 Result = DAG.getNode(ISD::ADD, dl, PtrVT,
11930 DAG.getNode(X86ISD::GlobalBaseReg, dl, PtrVT), Result);
11933 // For globals that require a load from a stub to get the address, emit the
11935 if (isGlobalStubReference(OpFlags))
11936 Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Result,
11937 MachinePointerInfo::getGOT(DAG.getMachineFunction()),
11938 false, false, false, 0);
11940 // If there was a non-zero offset that we didn't fold, create an explicit
11941 // addition for it.
11943 Result = DAG.getNode(ISD::ADD, dl, PtrVT, Result,
11944 DAG.getConstant(Offset, dl, PtrVT));
11950 X86TargetLowering::LowerGlobalAddress(SDValue Op, SelectionDAG &DAG) const {
11951 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
11952 int64_t Offset = cast<GlobalAddressSDNode>(Op)->getOffset();
11953 return LowerGlobalAddress(GV, SDLoc(Op), Offset, DAG);
11957 GetTLSADDR(SelectionDAG &DAG, SDValue Chain, GlobalAddressSDNode *GA,
11958 SDValue *InFlag, const EVT PtrVT, unsigned ReturnReg,
11959 unsigned char OperandFlags, bool LocalDynamic = false) {
11960 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
11961 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
11963 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(), dl,
11964 GA->getValueType(0),
11968 X86ISD::NodeType CallType = LocalDynamic ? X86ISD::TLSBASEADDR
11972 SDValue Ops[] = { Chain, TGA, *InFlag };
11973 Chain = DAG.getNode(CallType, dl, NodeTys, Ops);
11975 SDValue Ops[] = { Chain, TGA };
11976 Chain = DAG.getNode(CallType, dl, NodeTys, Ops);
11979 // TLSADDR will be codegen'ed as call. Inform MFI that function has calls.
11980 MFI->setAdjustsStack(true);
11981 MFI->setHasCalls(true);
11983 SDValue Flag = Chain.getValue(1);
11984 return DAG.getCopyFromReg(Chain, dl, ReturnReg, PtrVT, Flag);
11987 // Lower ISD::GlobalTLSAddress using the "general dynamic" model, 32 bit
11989 LowerToTLSGeneralDynamicModel32(GlobalAddressSDNode *GA, SelectionDAG &DAG,
11992 SDLoc dl(GA); // ? function entry point might be better
11993 SDValue Chain = DAG.getCopyToReg(DAG.getEntryNode(), dl, X86::EBX,
11994 DAG.getNode(X86ISD::GlobalBaseReg,
11995 SDLoc(), PtrVT), InFlag);
11996 InFlag = Chain.getValue(1);
11998 return GetTLSADDR(DAG, Chain, GA, &InFlag, PtrVT, X86::EAX, X86II::MO_TLSGD);
12001 // Lower ISD::GlobalTLSAddress using the "general dynamic" model, 64 bit
12003 LowerToTLSGeneralDynamicModel64(GlobalAddressSDNode *GA, SelectionDAG &DAG,
12005 return GetTLSADDR(DAG, DAG.getEntryNode(), GA, nullptr, PtrVT,
12006 X86::RAX, X86II::MO_TLSGD);
12009 static SDValue LowerToTLSLocalDynamicModel(GlobalAddressSDNode *GA,
12015 // Get the start address of the TLS block for this module.
12016 X86MachineFunctionInfo* MFI = DAG.getMachineFunction()
12017 .getInfo<X86MachineFunctionInfo>();
12018 MFI->incNumLocalDynamicTLSAccesses();
12022 Base = GetTLSADDR(DAG, DAG.getEntryNode(), GA, nullptr, PtrVT, X86::RAX,
12023 X86II::MO_TLSLD, /*LocalDynamic=*/true);
12026 SDValue Chain = DAG.getCopyToReg(DAG.getEntryNode(), dl, X86::EBX,
12027 DAG.getNode(X86ISD::GlobalBaseReg, SDLoc(), PtrVT), InFlag);
12028 InFlag = Chain.getValue(1);
12029 Base = GetTLSADDR(DAG, Chain, GA, &InFlag, PtrVT, X86::EAX,
12030 X86II::MO_TLSLDM, /*LocalDynamic=*/true);
12033 // Note: the CleanupLocalDynamicTLSPass will remove redundant computations
12037 unsigned char OperandFlags = X86II::MO_DTPOFF;
12038 unsigned WrapperKind = X86ISD::Wrapper;
12039 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(), dl,
12040 GA->getValueType(0),
12041 GA->getOffset(), OperandFlags);
12042 SDValue Offset = DAG.getNode(WrapperKind, dl, PtrVT, TGA);
12044 // Add x@dtpoff with the base.
12045 return DAG.getNode(ISD::ADD, dl, PtrVT, Offset, Base);
12048 // Lower ISD::GlobalTLSAddress using the "initial exec" or "local exec" model.
12049 static SDValue LowerToTLSExecModel(GlobalAddressSDNode *GA, SelectionDAG &DAG,
12050 const EVT PtrVT, TLSModel::Model model,
12051 bool is64Bit, bool isPIC) {
12054 // Get the Thread Pointer, which is %gs:0 (32-bit) or %fs:0 (64-bit).
12055 Value *Ptr = Constant::getNullValue(Type::getInt8PtrTy(*DAG.getContext(),
12056 is64Bit ? 257 : 256));
12058 SDValue ThreadPointer =
12059 DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), DAG.getIntPtrConstant(0, dl),
12060 MachinePointerInfo(Ptr), false, false, false, 0);
12062 unsigned char OperandFlags = 0;
12063 // Most TLS accesses are not RIP relative, even on x86-64. One exception is
12065 unsigned WrapperKind = X86ISD::Wrapper;
12066 if (model == TLSModel::LocalExec) {
12067 OperandFlags = is64Bit ? X86II::MO_TPOFF : X86II::MO_NTPOFF;
12068 } else if (model == TLSModel::InitialExec) {
12070 OperandFlags = X86II::MO_GOTTPOFF;
12071 WrapperKind = X86ISD::WrapperRIP;
12073 OperandFlags = isPIC ? X86II::MO_GOTNTPOFF : X86II::MO_INDNTPOFF;
12076 llvm_unreachable("Unexpected model");
12079 // emit "addl x@ntpoff,%eax" (local exec)
12080 // or "addl x@indntpoff,%eax" (initial exec)
12081 // or "addl x@gotntpoff(%ebx) ,%eax" (initial exec, 32-bit pic)
12083 DAG.getTargetGlobalAddress(GA->getGlobal(), dl, GA->getValueType(0),
12084 GA->getOffset(), OperandFlags);
12085 SDValue Offset = DAG.getNode(WrapperKind, dl, PtrVT, TGA);
12087 if (model == TLSModel::InitialExec) {
12088 if (isPIC && !is64Bit) {
12089 Offset = DAG.getNode(ISD::ADD, dl, PtrVT,
12090 DAG.getNode(X86ISD::GlobalBaseReg, SDLoc(), PtrVT),
12094 Offset = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Offset,
12095 MachinePointerInfo::getGOT(DAG.getMachineFunction()),
12096 false, false, false, 0);
12099 // The address of the thread local variable is the add of the thread
12100 // pointer with the offset of the variable.
12101 return DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, Offset);
12105 X86TargetLowering::LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const {
12107 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
12108 const GlobalValue *GV = GA->getGlobal();
12109 auto PtrVT = getPointerTy(DAG.getDataLayout());
12111 if (Subtarget->isTargetELF()) {
12112 if (DAG.getTarget().Options.EmulatedTLS)
12113 return LowerToTLSEmulatedModel(GA, DAG);
12114 TLSModel::Model model = DAG.getTarget().getTLSModel(GV);
12116 case TLSModel::GeneralDynamic:
12117 if (Subtarget->is64Bit())
12118 return LowerToTLSGeneralDynamicModel64(GA, DAG, PtrVT);
12119 return LowerToTLSGeneralDynamicModel32(GA, DAG, PtrVT);
12120 case TLSModel::LocalDynamic:
12121 return LowerToTLSLocalDynamicModel(GA, DAG, PtrVT,
12122 Subtarget->is64Bit());
12123 case TLSModel::InitialExec:
12124 case TLSModel::LocalExec:
12125 return LowerToTLSExecModel(GA, DAG, PtrVT, model, Subtarget->is64Bit(),
12126 DAG.getTarget().getRelocationModel() ==
12129 llvm_unreachable("Unknown TLS model.");
12132 if (Subtarget->isTargetDarwin()) {
12133 // Darwin only has one model of TLS. Lower to that.
12134 unsigned char OpFlag = 0;
12135 unsigned WrapperKind = Subtarget->isPICStyleRIPRel() ?
12136 X86ISD::WrapperRIP : X86ISD::Wrapper;
12138 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
12139 // global base reg.
12140 bool PIC32 = (DAG.getTarget().getRelocationModel() == Reloc::PIC_) &&
12141 !Subtarget->is64Bit();
12143 OpFlag = X86II::MO_TLVP_PIC_BASE;
12145 OpFlag = X86II::MO_TLVP;
12147 SDValue Result = DAG.getTargetGlobalAddress(GA->getGlobal(), DL,
12148 GA->getValueType(0),
12149 GA->getOffset(), OpFlag);
12150 SDValue Offset = DAG.getNode(WrapperKind, DL, PtrVT, Result);
12152 // With PIC32, the address is actually $g + Offset.
12154 Offset = DAG.getNode(ISD::ADD, DL, PtrVT,
12155 DAG.getNode(X86ISD::GlobalBaseReg, SDLoc(), PtrVT),
12158 // Lowering the machine isd will make sure everything is in the right
12160 SDValue Chain = DAG.getEntryNode();
12161 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
12162 SDValue Args[] = { Chain, Offset };
12163 Chain = DAG.getNode(X86ISD::TLSCALL, DL, NodeTys, Args);
12165 // TLSCALL will be codegen'ed as call. Inform MFI that function has calls.
12166 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
12167 MFI->setAdjustsStack(true);
12169 // And our return value (tls address) is in the standard call return value
12171 unsigned Reg = Subtarget->is64Bit() ? X86::RAX : X86::EAX;
12172 return DAG.getCopyFromReg(Chain, DL, Reg, PtrVT, Chain.getValue(1));
12175 if (Subtarget->isTargetKnownWindowsMSVC() ||
12176 Subtarget->isTargetWindowsGNU()) {
12177 // Just use the implicit TLS architecture
12178 // Need to generate someting similar to:
12179 // mov rdx, qword [gs:abs 58H]; Load pointer to ThreadLocalStorage
12181 // mov ecx, dword [rel _tls_index]: Load index (from C runtime)
12182 // mov rcx, qword [rdx+rcx*8]
12183 // mov eax, .tls$:tlsvar
12184 // [rax+rcx] contains the address
12185 // Windows 64bit: gs:0x58
12186 // Windows 32bit: fs:__tls_array
12189 SDValue Chain = DAG.getEntryNode();
12191 // Get the Thread Pointer, which is %fs:__tls_array (32-bit) or
12192 // %gs:0x58 (64-bit). On MinGW, __tls_array is not available, so directly
12193 // use its literal value of 0x2C.
12194 Value *Ptr = Constant::getNullValue(Subtarget->is64Bit()
12195 ? Type::getInt8PtrTy(*DAG.getContext(),
12197 : Type::getInt32PtrTy(*DAG.getContext(),
12200 SDValue TlsArray = Subtarget->is64Bit()
12201 ? DAG.getIntPtrConstant(0x58, dl)
12202 : (Subtarget->isTargetWindowsGNU()
12203 ? DAG.getIntPtrConstant(0x2C, dl)
12204 : DAG.getExternalSymbol("_tls_array", PtrVT));
12206 SDValue ThreadPointer =
12207 DAG.getLoad(PtrVT, dl, Chain, TlsArray, MachinePointerInfo(Ptr), false,
12211 if (GV->getThreadLocalMode() == GlobalVariable::LocalExecTLSModel) {
12212 res = ThreadPointer;
12214 // Load the _tls_index variable
12215 SDValue IDX = DAG.getExternalSymbol("_tls_index", PtrVT);
12216 if (Subtarget->is64Bit())
12217 IDX = DAG.getExtLoad(ISD::ZEXTLOAD, dl, PtrVT, Chain, IDX,
12218 MachinePointerInfo(), MVT::i32, false, false,
12221 IDX = DAG.getLoad(PtrVT, dl, Chain, IDX, MachinePointerInfo(), false,
12224 auto &DL = DAG.getDataLayout();
12226 DAG.getConstant(Log2_64_Ceil(DL.getPointerSize()), dl, PtrVT);
12227 IDX = DAG.getNode(ISD::SHL, dl, PtrVT, IDX, Scale);
12229 res = DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, IDX);
12232 res = DAG.getLoad(PtrVT, dl, Chain, res, MachinePointerInfo(), false, false,
12235 // Get the offset of start of .tls section
12236 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(), dl,
12237 GA->getValueType(0),
12238 GA->getOffset(), X86II::MO_SECREL);
12239 SDValue Offset = DAG.getNode(X86ISD::Wrapper, dl, PtrVT, TGA);
12241 // The address of the thread local variable is the add of the thread
12242 // pointer with the offset of the variable.
12243 return DAG.getNode(ISD::ADD, dl, PtrVT, res, Offset);
12246 llvm_unreachable("TLS not implemented for this target.");
12249 /// LowerShiftParts - Lower SRA_PARTS and friends, which return two i32 values
12250 /// and take a 2 x i32 value to shift plus a shift amount.
12251 static SDValue LowerShiftParts(SDValue Op, SelectionDAG &DAG) {
12252 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
12253 MVT VT = Op.getSimpleValueType();
12254 unsigned VTBits = VT.getSizeInBits();
12256 bool isSRA = Op.getOpcode() == ISD::SRA_PARTS;
12257 SDValue ShOpLo = Op.getOperand(0);
12258 SDValue ShOpHi = Op.getOperand(1);
12259 SDValue ShAmt = Op.getOperand(2);
12260 // X86ISD::SHLD and X86ISD::SHRD have defined overflow behavior but the
12261 // generic ISD nodes haven't. Insert an AND to be safe, it's optimized away
12263 SDValue SafeShAmt = DAG.getNode(ISD::AND, dl, MVT::i8, ShAmt,
12264 DAG.getConstant(VTBits - 1, dl, MVT::i8));
12265 SDValue Tmp1 = isSRA ? DAG.getNode(ISD::SRA, dl, VT, ShOpHi,
12266 DAG.getConstant(VTBits - 1, dl, MVT::i8))
12267 : DAG.getConstant(0, dl, VT);
12269 SDValue Tmp2, Tmp3;
12270 if (Op.getOpcode() == ISD::SHL_PARTS) {
12271 Tmp2 = DAG.getNode(X86ISD::SHLD, dl, VT, ShOpHi, ShOpLo, ShAmt);
12272 Tmp3 = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, SafeShAmt);
12274 Tmp2 = DAG.getNode(X86ISD::SHRD, dl, VT, ShOpLo, ShOpHi, ShAmt);
12275 Tmp3 = DAG.getNode(isSRA ? ISD::SRA : ISD::SRL, dl, VT, ShOpHi, SafeShAmt);
12278 // If the shift amount is larger or equal than the width of a part we can't
12279 // rely on the results of shld/shrd. Insert a test and select the appropriate
12280 // values for large shift amounts.
12281 SDValue AndNode = DAG.getNode(ISD::AND, dl, MVT::i8, ShAmt,
12282 DAG.getConstant(VTBits, dl, MVT::i8));
12283 SDValue Cond = DAG.getNode(X86ISD::CMP, dl, MVT::i32,
12284 AndNode, DAG.getConstant(0, dl, MVT::i8));
12287 SDValue CC = DAG.getConstant(X86::COND_NE, dl, MVT::i8);
12288 SDValue Ops0[4] = { Tmp2, Tmp3, CC, Cond };
12289 SDValue Ops1[4] = { Tmp3, Tmp1, CC, Cond };
12291 if (Op.getOpcode() == ISD::SHL_PARTS) {
12292 Hi = DAG.getNode(X86ISD::CMOV, dl, VT, Ops0);
12293 Lo = DAG.getNode(X86ISD::CMOV, dl, VT, Ops1);
12295 Lo = DAG.getNode(X86ISD::CMOV, dl, VT, Ops0);
12296 Hi = DAG.getNode(X86ISD::CMOV, dl, VT, Ops1);
12299 SDValue Ops[2] = { Lo, Hi };
12300 return DAG.getMergeValues(Ops, dl);
12303 SDValue X86TargetLowering::LowerSINT_TO_FP(SDValue Op,
12304 SelectionDAG &DAG) const {
12305 SDValue Src = Op.getOperand(0);
12306 MVT SrcVT = Src.getSimpleValueType();
12307 MVT VT = Op.getSimpleValueType();
12310 if (SrcVT.isVector()) {
12311 if (SrcVT == MVT::v2i32 && VT == MVT::v2f64) {
12312 return DAG.getNode(X86ISD::CVTDQ2PD, dl, VT,
12313 DAG.getNode(ISD::CONCAT_VECTORS, dl, MVT::v4i32, Src,
12314 DAG.getUNDEF(SrcVT)));
12316 if (SrcVT.getVectorElementType() == MVT::i1) {
12317 MVT IntegerVT = MVT::getVectorVT(MVT::i32, SrcVT.getVectorNumElements());
12318 return DAG.getNode(ISD::SINT_TO_FP, dl, Op.getValueType(),
12319 DAG.getNode(ISD::SIGN_EXTEND, dl, IntegerVT, Src));
12324 assert(SrcVT <= MVT::i64 && SrcVT >= MVT::i16 &&
12325 "Unknown SINT_TO_FP to lower!");
12327 // These are really Legal; return the operand so the caller accepts it as
12329 if (SrcVT == MVT::i32 && isScalarFPTypeInSSEReg(Op.getValueType()))
12331 if (SrcVT == MVT::i64 && isScalarFPTypeInSSEReg(Op.getValueType()) &&
12332 Subtarget->is64Bit()) {
12336 unsigned Size = SrcVT.getSizeInBits()/8;
12337 MachineFunction &MF = DAG.getMachineFunction();
12338 auto PtrVT = getPointerTy(MF.getDataLayout());
12339 int SSFI = MF.getFrameInfo()->CreateStackObject(Size, Size, false);
12340 SDValue StackSlot = DAG.getFrameIndex(SSFI, PtrVT);
12341 SDValue Chain = DAG.getStore(
12342 DAG.getEntryNode(), dl, Op.getOperand(0), StackSlot,
12343 MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), SSFI), false,
12345 return BuildFILD(Op, SrcVT, Chain, StackSlot, DAG);
12348 SDValue X86TargetLowering::BuildFILD(SDValue Op, EVT SrcVT, SDValue Chain,
12350 SelectionDAG &DAG) const {
12354 bool useSSE = isScalarFPTypeInSSEReg(Op.getValueType());
12356 Tys = DAG.getVTList(MVT::f64, MVT::Other, MVT::Glue);
12358 Tys = DAG.getVTList(Op.getValueType(), MVT::Other);
12360 unsigned ByteSize = SrcVT.getSizeInBits()/8;
12362 FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(StackSlot);
12363 MachineMemOperand *MMO;
12365 int SSFI = FI->getIndex();
12366 MMO = DAG.getMachineFunction().getMachineMemOperand(
12367 MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), SSFI),
12368 MachineMemOperand::MOLoad, ByteSize, ByteSize);
12370 MMO = cast<LoadSDNode>(StackSlot)->getMemOperand();
12371 StackSlot = StackSlot.getOperand(1);
12373 SDValue Ops[] = { Chain, StackSlot, DAG.getValueType(SrcVT) };
12374 SDValue Result = DAG.getMemIntrinsicNode(useSSE ? X86ISD::FILD_FLAG :
12376 Tys, Ops, SrcVT, MMO);
12379 Chain = Result.getValue(1);
12380 SDValue InFlag = Result.getValue(2);
12382 // FIXME: Currently the FST is flagged to the FILD_FLAG. This
12383 // shouldn't be necessary except that RFP cannot be live across
12384 // multiple blocks. When stackifier is fixed, they can be uncoupled.
12385 MachineFunction &MF = DAG.getMachineFunction();
12386 unsigned SSFISize = Op.getValueType().getSizeInBits()/8;
12387 int SSFI = MF.getFrameInfo()->CreateStackObject(SSFISize, SSFISize, false);
12388 auto PtrVT = getPointerTy(MF.getDataLayout());
12389 SDValue StackSlot = DAG.getFrameIndex(SSFI, PtrVT);
12390 Tys = DAG.getVTList(MVT::Other);
12392 Chain, Result, StackSlot, DAG.getValueType(Op.getValueType()), InFlag
12394 MachineMemOperand *MMO = DAG.getMachineFunction().getMachineMemOperand(
12395 MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), SSFI),
12396 MachineMemOperand::MOStore, SSFISize, SSFISize);
12398 Chain = DAG.getMemIntrinsicNode(X86ISD::FST, DL, Tys,
12399 Ops, Op.getValueType(), MMO);
12400 Result = DAG.getLoad(
12401 Op.getValueType(), DL, Chain, StackSlot,
12402 MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), SSFI),
12403 false, false, false, 0);
12409 // LowerUINT_TO_FP_i64 - 64-bit unsigned integer to double expansion.
12410 SDValue X86TargetLowering::LowerUINT_TO_FP_i64(SDValue Op,
12411 SelectionDAG &DAG) const {
12412 // This algorithm is not obvious. Here it is what we're trying to output:
12415 punpckldq (c0), %xmm0 // c0: (uint4){ 0x43300000U, 0x45300000U, 0U, 0U }
12416 subpd (c1), %xmm0 // c1: (double2){ 0x1.0p52, 0x1.0p52 * 0x1.0p32 }
12418 haddpd %xmm0, %xmm0
12420 pshufd $0x4e, %xmm0, %xmm1
12426 LLVMContext *Context = DAG.getContext();
12428 // Build some magic constants.
12429 static const uint32_t CV0[] = { 0x43300000, 0x45300000, 0, 0 };
12430 Constant *C0 = ConstantDataVector::get(*Context, CV0);
12431 auto PtrVT = getPointerTy(DAG.getDataLayout());
12432 SDValue CPIdx0 = DAG.getConstantPool(C0, PtrVT, 16);
12434 SmallVector<Constant*,2> CV1;
12436 ConstantFP::get(*Context, APFloat(APFloat::IEEEdouble,
12437 APInt(64, 0x4330000000000000ULL))));
12439 ConstantFP::get(*Context, APFloat(APFloat::IEEEdouble,
12440 APInt(64, 0x4530000000000000ULL))));
12441 Constant *C1 = ConstantVector::get(CV1);
12442 SDValue CPIdx1 = DAG.getConstantPool(C1, PtrVT, 16);
12444 // Load the 64-bit value into an XMM register.
12445 SDValue XR1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i64,
12448 DAG.getLoad(MVT::v4i32, dl, DAG.getEntryNode(), CPIdx0,
12449 MachinePointerInfo::getConstantPool(DAG.getMachineFunction()),
12450 false, false, false, 16);
12452 getUnpackl(DAG, dl, MVT::v4i32, DAG.getBitcast(MVT::v4i32, XR1), CLod0);
12455 DAG.getLoad(MVT::v2f64, dl, CLod0.getValue(1), CPIdx1,
12456 MachinePointerInfo::getConstantPool(DAG.getMachineFunction()),
12457 false, false, false, 16);
12458 SDValue XR2F = DAG.getBitcast(MVT::v2f64, Unpck1);
12459 // TODO: Are there any fast-math-flags to propagate here?
12460 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::v2f64, XR2F, CLod1);
12463 if (Subtarget->hasSSE3()) {
12464 // FIXME: The 'haddpd' instruction may be slower than 'movhlps + addsd'.
12465 Result = DAG.getNode(X86ISD::FHADD, dl, MVT::v2f64, Sub, Sub);
12467 SDValue S2F = DAG.getBitcast(MVT::v4i32, Sub);
12468 SDValue Shuffle = getTargetShuffleNode(X86ISD::PSHUFD, dl, MVT::v4i32,
12470 Result = DAG.getNode(ISD::FADD, dl, MVT::v2f64,
12471 DAG.getBitcast(MVT::v2f64, Shuffle), Sub);
12474 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Result,
12475 DAG.getIntPtrConstant(0, dl));
12478 // LowerUINT_TO_FP_i32 - 32-bit unsigned integer to float expansion.
12479 SDValue X86TargetLowering::LowerUINT_TO_FP_i32(SDValue Op,
12480 SelectionDAG &DAG) const {
12482 // FP constant to bias correct the final result.
12483 SDValue Bias = DAG.getConstantFP(BitsToDouble(0x4330000000000000ULL), dl,
12486 // Load the 32-bit value into an XMM register.
12487 SDValue Load = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
12490 // Zero out the upper parts of the register.
12491 Load = getShuffleVectorZeroOrUndef(Load, 0, true, Subtarget, DAG);
12493 Load = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
12494 DAG.getBitcast(MVT::v2f64, Load),
12495 DAG.getIntPtrConstant(0, dl));
12497 // Or the load with the bias.
12498 SDValue Or = DAG.getNode(
12499 ISD::OR, dl, MVT::v2i64,
12500 DAG.getBitcast(MVT::v2i64,
12501 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2f64, Load)),
12502 DAG.getBitcast(MVT::v2i64,
12503 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2f64, Bias)));
12505 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
12506 DAG.getBitcast(MVT::v2f64, Or), DAG.getIntPtrConstant(0, dl));
12508 // Subtract the bias.
12509 // TODO: Are there any fast-math-flags to propagate here?
12510 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::f64, Or, Bias);
12512 // Handle final rounding.
12513 EVT DestVT = Op.getValueType();
12515 if (DestVT.bitsLT(MVT::f64))
12516 return DAG.getNode(ISD::FP_ROUND, dl, DestVT, Sub,
12517 DAG.getIntPtrConstant(0, dl));
12518 if (DestVT.bitsGT(MVT::f64))
12519 return DAG.getNode(ISD::FP_EXTEND, dl, DestVT, Sub);
12521 // Handle final rounding.
12525 static SDValue lowerUINT_TO_FP_vXi32(SDValue Op, SelectionDAG &DAG,
12526 const X86Subtarget &Subtarget) {
12527 // The algorithm is the following:
12528 // #ifdef __SSE4_1__
12529 // uint4 lo = _mm_blend_epi16( v, (uint4) 0x4b000000, 0xaa);
12530 // uint4 hi = _mm_blend_epi16( _mm_srli_epi32(v,16),
12531 // (uint4) 0x53000000, 0xaa);
12533 // uint4 lo = (v & (uint4) 0xffff) | (uint4) 0x4b000000;
12534 // uint4 hi = (v >> 16) | (uint4) 0x53000000;
12536 // float4 fhi = (float4) hi - (0x1.0p39f + 0x1.0p23f);
12537 // return (float4) lo + fhi;
12539 // We shouldn't use it when unsafe-fp-math is enabled though: we might later
12540 // reassociate the two FADDs, and if we do that, the algorithm fails
12541 // spectacularly (PR24512).
12542 // FIXME: If we ever have some kind of Machine FMF, this should be marked
12543 // as non-fast and always be enabled. Why isn't SDAG FMF enough? Because
12544 // there's also the MachineCombiner reassociations happening on Machine IR.
12545 if (DAG.getTarget().Options.UnsafeFPMath)
12549 SDValue V = Op->getOperand(0);
12550 EVT VecIntVT = V.getValueType();
12551 bool Is128 = VecIntVT == MVT::v4i32;
12552 EVT VecFloatVT = Is128 ? MVT::v4f32 : MVT::v8f32;
12553 // If we convert to something else than the supported type, e.g., to v4f64,
12555 if (VecFloatVT != Op->getValueType(0))
12558 unsigned NumElts = VecIntVT.getVectorNumElements();
12559 assert((VecIntVT == MVT::v4i32 || VecIntVT == MVT::v8i32) &&
12560 "Unsupported custom type");
12561 assert(NumElts <= 8 && "The size of the constant array must be fixed");
12563 // In the #idef/#else code, we have in common:
12564 // - The vector of constants:
12570 // Create the splat vector for 0x4b000000.
12571 SDValue CstLow = DAG.getConstant(0x4b000000, DL, MVT::i32);
12572 SDValue CstLowArray[] = {CstLow, CstLow, CstLow, CstLow,
12573 CstLow, CstLow, CstLow, CstLow};
12574 SDValue VecCstLow = DAG.getNode(ISD::BUILD_VECTOR, DL, VecIntVT,
12575 makeArrayRef(&CstLowArray[0], NumElts));
12576 // Create the splat vector for 0x53000000.
12577 SDValue CstHigh = DAG.getConstant(0x53000000, DL, MVT::i32);
12578 SDValue CstHighArray[] = {CstHigh, CstHigh, CstHigh, CstHigh,
12579 CstHigh, CstHigh, CstHigh, CstHigh};
12580 SDValue VecCstHigh = DAG.getNode(ISD::BUILD_VECTOR, DL, VecIntVT,
12581 makeArrayRef(&CstHighArray[0], NumElts));
12583 // Create the right shift.
12584 SDValue CstShift = DAG.getConstant(16, DL, MVT::i32);
12585 SDValue CstShiftArray[] = {CstShift, CstShift, CstShift, CstShift,
12586 CstShift, CstShift, CstShift, CstShift};
12587 SDValue VecCstShift = DAG.getNode(ISD::BUILD_VECTOR, DL, VecIntVT,
12588 makeArrayRef(&CstShiftArray[0], NumElts));
12589 SDValue HighShift = DAG.getNode(ISD::SRL, DL, VecIntVT, V, VecCstShift);
12592 if (Subtarget.hasSSE41()) {
12593 EVT VecI16VT = Is128 ? MVT::v8i16 : MVT::v16i16;
12594 // uint4 lo = _mm_blend_epi16( v, (uint4) 0x4b000000, 0xaa);
12595 SDValue VecCstLowBitcast = DAG.getBitcast(VecI16VT, VecCstLow);
12596 SDValue VecBitcast = DAG.getBitcast(VecI16VT, V);
12597 // Low will be bitcasted right away, so do not bother bitcasting back to its
12599 Low = DAG.getNode(X86ISD::BLENDI, DL, VecI16VT, VecBitcast,
12600 VecCstLowBitcast, DAG.getConstant(0xaa, DL, MVT::i32));
12601 // uint4 hi = _mm_blend_epi16( _mm_srli_epi32(v,16),
12602 // (uint4) 0x53000000, 0xaa);
12603 SDValue VecCstHighBitcast = DAG.getBitcast(VecI16VT, VecCstHigh);
12604 SDValue VecShiftBitcast = DAG.getBitcast(VecI16VT, HighShift);
12605 // High will be bitcasted right away, so do not bother bitcasting back to
12606 // its original type.
12607 High = DAG.getNode(X86ISD::BLENDI, DL, VecI16VT, VecShiftBitcast,
12608 VecCstHighBitcast, DAG.getConstant(0xaa, DL, MVT::i32));
12610 SDValue CstMask = DAG.getConstant(0xffff, DL, MVT::i32);
12611 SDValue VecCstMask = DAG.getNode(ISD::BUILD_VECTOR, DL, VecIntVT, CstMask,
12612 CstMask, CstMask, CstMask);
12613 // uint4 lo = (v & (uint4) 0xffff) | (uint4) 0x4b000000;
12614 SDValue LowAnd = DAG.getNode(ISD::AND, DL, VecIntVT, V, VecCstMask);
12615 Low = DAG.getNode(ISD::OR, DL, VecIntVT, LowAnd, VecCstLow);
12617 // uint4 hi = (v >> 16) | (uint4) 0x53000000;
12618 High = DAG.getNode(ISD::OR, DL, VecIntVT, HighShift, VecCstHigh);
12621 // Create the vector constant for -(0x1.0p39f + 0x1.0p23f).
12622 SDValue CstFAdd = DAG.getConstantFP(
12623 APFloat(APFloat::IEEEsingle, APInt(32, 0xD3000080)), DL, MVT::f32);
12624 SDValue CstFAddArray[] = {CstFAdd, CstFAdd, CstFAdd, CstFAdd,
12625 CstFAdd, CstFAdd, CstFAdd, CstFAdd};
12626 SDValue VecCstFAdd = DAG.getNode(ISD::BUILD_VECTOR, DL, VecFloatVT,
12627 makeArrayRef(&CstFAddArray[0], NumElts));
12629 // float4 fhi = (float4) hi - (0x1.0p39f + 0x1.0p23f);
12630 SDValue HighBitcast = DAG.getBitcast(VecFloatVT, High);
12631 // TODO: Are there any fast-math-flags to propagate here?
12633 DAG.getNode(ISD::FADD, DL, VecFloatVT, HighBitcast, VecCstFAdd);
12634 // return (float4) lo + fhi;
12635 SDValue LowBitcast = DAG.getBitcast(VecFloatVT, Low);
12636 return DAG.getNode(ISD::FADD, DL, VecFloatVT, LowBitcast, FHigh);
12639 SDValue X86TargetLowering::lowerUINT_TO_FP_vec(SDValue Op,
12640 SelectionDAG &DAG) const {
12641 SDValue N0 = Op.getOperand(0);
12642 MVT SVT = N0.getSimpleValueType();
12645 switch (SVT.SimpleTy) {
12647 llvm_unreachable("Custom UINT_TO_FP is not supported!");
12652 MVT NVT = MVT::getVectorVT(MVT::i32, SVT.getVectorNumElements());
12653 return DAG.getNode(ISD::SINT_TO_FP, dl, Op.getValueType(),
12654 DAG.getNode(ISD::ZERO_EXTEND, dl, NVT, N0));
12658 return lowerUINT_TO_FP_vXi32(Op, DAG, *Subtarget);
12661 if (Subtarget->hasAVX512())
12662 return DAG.getNode(ISD::UINT_TO_FP, dl, Op.getValueType(),
12663 DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v16i32, N0));
12665 llvm_unreachable(nullptr);
12668 SDValue X86TargetLowering::LowerUINT_TO_FP(SDValue Op,
12669 SelectionDAG &DAG) const {
12670 SDValue N0 = Op.getOperand(0);
12672 auto PtrVT = getPointerTy(DAG.getDataLayout());
12674 if (Op.getValueType().isVector())
12675 return lowerUINT_TO_FP_vec(Op, DAG);
12677 // Since UINT_TO_FP is legal (it's marked custom), dag combiner won't
12678 // optimize it to a SINT_TO_FP when the sign bit is known zero. Perform
12679 // the optimization here.
12680 if (DAG.SignBitIsZero(N0))
12681 return DAG.getNode(ISD::SINT_TO_FP, dl, Op.getValueType(), N0);
12683 MVT SrcVT = N0.getSimpleValueType();
12684 MVT DstVT = Op.getSimpleValueType();
12686 if (Subtarget->hasAVX512() && isScalarFPTypeInSSEReg(DstVT) &&
12687 (SrcVT == MVT::i32 || (SrcVT == MVT::i64 && Subtarget->is64Bit()))) {
12688 // Conversions from unsigned i32 to f32/f64 are legal,
12689 // using VCVTUSI2SS/SD. Same for i64 in 64-bit mode.
12693 if (SrcVT == MVT::i64 && DstVT == MVT::f64 && X86ScalarSSEf64)
12694 return LowerUINT_TO_FP_i64(Op, DAG);
12695 if (SrcVT == MVT::i32 && X86ScalarSSEf64)
12696 return LowerUINT_TO_FP_i32(Op, DAG);
12697 if (Subtarget->is64Bit() && SrcVT == MVT::i64 && DstVT == MVT::f32)
12700 // Make a 64-bit buffer, and use it to build an FILD.
12701 SDValue StackSlot = DAG.CreateStackTemporary(MVT::i64);
12702 if (SrcVT == MVT::i32) {
12703 SDValue WordOff = DAG.getConstant(4, dl, PtrVT);
12704 SDValue OffsetSlot = DAG.getNode(ISD::ADD, dl, PtrVT, StackSlot, WordOff);
12705 SDValue Store1 = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
12706 StackSlot, MachinePointerInfo(),
12708 SDValue Store2 = DAG.getStore(Store1, dl, DAG.getConstant(0, dl, MVT::i32),
12709 OffsetSlot, MachinePointerInfo(),
12711 SDValue Fild = BuildFILD(Op, MVT::i64, Store2, StackSlot, DAG);
12715 assert(SrcVT == MVT::i64 && "Unexpected type in UINT_TO_FP");
12716 SDValue Store = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
12717 StackSlot, MachinePointerInfo(),
12719 // For i64 source, we need to add the appropriate power of 2 if the input
12720 // was negative. This is the same as the optimization in
12721 // DAGTypeLegalizer::ExpandIntOp_UNIT_TO_FP, and for it to be safe here,
12722 // we must be careful to do the computation in x87 extended precision, not
12723 // in SSE. (The generic code can't know it's OK to do this, or how to.)
12724 int SSFI = cast<FrameIndexSDNode>(StackSlot)->getIndex();
12725 MachineMemOperand *MMO = DAG.getMachineFunction().getMachineMemOperand(
12726 MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), SSFI),
12727 MachineMemOperand::MOLoad, 8, 8);
12729 SDVTList Tys = DAG.getVTList(MVT::f80, MVT::Other);
12730 SDValue Ops[] = { Store, StackSlot, DAG.getValueType(MVT::i64) };
12731 SDValue Fild = DAG.getMemIntrinsicNode(X86ISD::FILD, dl, Tys, Ops,
12734 APInt FF(32, 0x5F800000ULL);
12736 // Check whether the sign bit is set.
12737 SDValue SignSet = DAG.getSetCC(
12738 dl, getSetCCResultType(DAG.getDataLayout(), *DAG.getContext(), MVT::i64),
12739 Op.getOperand(0), DAG.getConstant(0, dl, MVT::i64), ISD::SETLT);
12741 // Build a 64 bit pair (0, FF) in the constant pool, with FF in the lo bits.
12742 SDValue FudgePtr = DAG.getConstantPool(
12743 ConstantInt::get(*DAG.getContext(), FF.zext(64)), PtrVT);
12745 // Get a pointer to FF if the sign bit was set, or to 0 otherwise.
12746 SDValue Zero = DAG.getIntPtrConstant(0, dl);
12747 SDValue Four = DAG.getIntPtrConstant(4, dl);
12748 SDValue Offset = DAG.getNode(ISD::SELECT, dl, Zero.getValueType(), SignSet,
12750 FudgePtr = DAG.getNode(ISD::ADD, dl, PtrVT, FudgePtr, Offset);
12752 // Load the value out, extending it from f32 to f80.
12753 // FIXME: Avoid the extend by constructing the right constant pool?
12754 SDValue Fudge = DAG.getExtLoad(
12755 ISD::EXTLOAD, dl, MVT::f80, DAG.getEntryNode(), FudgePtr,
12756 MachinePointerInfo::getConstantPool(DAG.getMachineFunction()), MVT::f32,
12757 false, false, false, 4);
12758 // Extend everything to 80 bits to force it to be done on x87.
12759 // TODO: Are there any fast-math-flags to propagate here?
12760 SDValue Add = DAG.getNode(ISD::FADD, dl, MVT::f80, Fild, Fudge);
12761 return DAG.getNode(ISD::FP_ROUND, dl, DstVT, Add,
12762 DAG.getIntPtrConstant(0, dl));
12765 // If the given FP_TO_SINT (IsSigned) or FP_TO_UINT (!IsSigned) operation
12766 // is legal, or has an fp128 or f16 source (which needs to be promoted to f32),
12767 // just return an <SDValue(), SDValue()> pair.
12768 // Otherwise it is assumed to be a conversion from one of f32, f64 or f80
12769 // to i16, i32 or i64, and we lower it to a legal sequence.
12770 // If lowered to the final integer result we return a <result, SDValue()> pair.
12771 // Otherwise we lower it to a sequence ending with a FIST, return a
12772 // <FIST, StackSlot> pair, and the caller is responsible for loading
12773 // the final integer result from StackSlot.
12774 std::pair<SDValue,SDValue>
12775 X86TargetLowering::FP_TO_INTHelper(SDValue Op, SelectionDAG &DAG,
12776 bool IsSigned, bool IsReplace) const {
12779 EVT DstTy = Op.getValueType();
12780 EVT TheVT = Op.getOperand(0).getValueType();
12781 auto PtrVT = getPointerTy(DAG.getDataLayout());
12783 if (TheVT != MVT::f32 && TheVT != MVT::f64 && TheVT != MVT::f80) {
12784 // f16 must be promoted before using the lowering in this routine.
12785 // fp128 does not use this lowering.
12786 return std::make_pair(SDValue(), SDValue());
12789 // If using FIST to compute an unsigned i64, we'll need some fixup
12790 // to handle values above the maximum signed i64. A FIST is always
12791 // used for the 32-bit subtarget, but also for f80 on a 64-bit target.
12792 bool UnsignedFixup = !IsSigned &&
12793 DstTy == MVT::i64 &&
12794 (!Subtarget->is64Bit() ||
12795 !isScalarFPTypeInSSEReg(TheVT));
12797 if (!IsSigned && DstTy != MVT::i64 && !Subtarget->hasAVX512()) {
12798 // Replace the fp-to-uint32 operation with an fp-to-sint64 FIST.
12799 // The low 32 bits of the fist result will have the correct uint32 result.
12800 assert(DstTy == MVT::i32 && "Unexpected FP_TO_UINT");
12804 assert(DstTy.getSimpleVT() <= MVT::i64 &&
12805 DstTy.getSimpleVT() >= MVT::i16 &&
12806 "Unknown FP_TO_INT to lower!");
12808 // These are really Legal.
12809 if (DstTy == MVT::i32 &&
12810 isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType()))
12811 return std::make_pair(SDValue(), SDValue());
12812 if (Subtarget->is64Bit() &&
12813 DstTy == MVT::i64 &&
12814 isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType()))
12815 return std::make_pair(SDValue(), SDValue());
12817 // We lower FP->int64 into FISTP64 followed by a load from a temporary
12819 MachineFunction &MF = DAG.getMachineFunction();
12820 unsigned MemSize = DstTy.getSizeInBits()/8;
12821 int SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize, false);
12822 SDValue StackSlot = DAG.getFrameIndex(SSFI, PtrVT);
12825 switch (DstTy.getSimpleVT().SimpleTy) {
12826 default: llvm_unreachable("Invalid FP_TO_SINT to lower!");
12827 case MVT::i16: Opc = X86ISD::FP_TO_INT16_IN_MEM; break;
12828 case MVT::i32: Opc = X86ISD::FP_TO_INT32_IN_MEM; break;
12829 case MVT::i64: Opc = X86ISD::FP_TO_INT64_IN_MEM; break;
12832 SDValue Chain = DAG.getEntryNode();
12833 SDValue Value = Op.getOperand(0);
12834 SDValue Adjust; // 0x0 or 0x80000000, for result sign bit adjustment.
12836 if (UnsignedFixup) {
12838 // Conversion to unsigned i64 is implemented with a select,
12839 // depending on whether the source value fits in the range
12840 // of a signed i64. Let Thresh be the FP equivalent of
12841 // 0x8000000000000000ULL.
12843 // Adjust i32 = (Value < Thresh) ? 0 : 0x80000000;
12844 // FistSrc = (Value < Thresh) ? Value : (Value - Thresh);
12845 // Fist-to-mem64 FistSrc
12846 // Add 0 or 0x800...0ULL to the 64-bit result, which is equivalent
12847 // to XOR'ing the high 32 bits with Adjust.
12849 // Being a power of 2, Thresh is exactly representable in all FP formats.
12850 // For X87 we'd like to use the smallest FP type for this constant, but
12851 // for DAG type consistency we have to match the FP operand type.
12853 APFloat Thresh(APFloat::IEEEsingle, APInt(32, 0x5f000000));
12854 LLVM_ATTRIBUTE_UNUSED APFloat::opStatus Status = APFloat::opOK;
12855 bool LosesInfo = false;
12856 if (TheVT == MVT::f64)
12857 // The rounding mode is irrelevant as the conversion should be exact.
12858 Status = Thresh.convert(APFloat::IEEEdouble, APFloat::rmNearestTiesToEven,
12860 else if (TheVT == MVT::f80)
12861 Status = Thresh.convert(APFloat::x87DoubleExtended,
12862 APFloat::rmNearestTiesToEven, &LosesInfo);
12864 assert(Status == APFloat::opOK && !LosesInfo &&
12865 "FP conversion should have been exact");
12867 SDValue ThreshVal = DAG.getConstantFP(Thresh, DL, TheVT);
12869 SDValue Cmp = DAG.getSetCC(DL,
12870 getSetCCResultType(DAG.getDataLayout(),
12871 *DAG.getContext(), TheVT),
12872 Value, ThreshVal, ISD::SETLT);
12873 Adjust = DAG.getSelect(DL, MVT::i32, Cmp,
12874 DAG.getConstant(0, DL, MVT::i32),
12875 DAG.getConstant(0x80000000, DL, MVT::i32));
12876 SDValue Sub = DAG.getNode(ISD::FSUB, DL, TheVT, Value, ThreshVal);
12877 Cmp = DAG.getSetCC(DL, getSetCCResultType(DAG.getDataLayout(),
12878 *DAG.getContext(), TheVT),
12879 Value, ThreshVal, ISD::SETLT);
12880 Value = DAG.getSelect(DL, TheVT, Cmp, Value, Sub);
12883 // FIXME This causes a redundant load/store if the SSE-class value is already
12884 // in memory, such as if it is on the callstack.
12885 if (isScalarFPTypeInSSEReg(TheVT)) {
12886 assert(DstTy == MVT::i64 && "Invalid FP_TO_SINT to lower!");
12887 Chain = DAG.getStore(Chain, DL, Value, StackSlot,
12888 MachinePointerInfo::getFixedStack(MF, SSFI), false,
12890 SDVTList Tys = DAG.getVTList(Op.getOperand(0).getValueType(), MVT::Other);
12892 Chain, StackSlot, DAG.getValueType(TheVT)
12895 MachineMemOperand *MMO =
12896 MF.getMachineMemOperand(MachinePointerInfo::getFixedStack(MF, SSFI),
12897 MachineMemOperand::MOLoad, MemSize, MemSize);
12898 Value = DAG.getMemIntrinsicNode(X86ISD::FLD, DL, Tys, Ops, DstTy, MMO);
12899 Chain = Value.getValue(1);
12900 SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize, false);
12901 StackSlot = DAG.getFrameIndex(SSFI, PtrVT);
12904 MachineMemOperand *MMO =
12905 MF.getMachineMemOperand(MachinePointerInfo::getFixedStack(MF, SSFI),
12906 MachineMemOperand::MOStore, MemSize, MemSize);
12908 if (UnsignedFixup) {
12910 // Insert the FIST, load its result as two i32's,
12911 // and XOR the high i32 with Adjust.
12913 SDValue FistOps[] = { Chain, Value, StackSlot };
12914 SDValue FIST = DAG.getMemIntrinsicNode(Opc, DL, DAG.getVTList(MVT::Other),
12915 FistOps, DstTy, MMO);
12917 SDValue Low32 = DAG.getLoad(MVT::i32, DL, FIST, StackSlot,
12918 MachinePointerInfo(),
12919 false, false, false, 0);
12920 SDValue HighAddr = DAG.getNode(ISD::ADD, DL, PtrVT, StackSlot,
12921 DAG.getConstant(4, DL, PtrVT));
12923 SDValue High32 = DAG.getLoad(MVT::i32, DL, FIST, HighAddr,
12924 MachinePointerInfo(),
12925 false, false, false, 0);
12926 High32 = DAG.getNode(ISD::XOR, DL, MVT::i32, High32, Adjust);
12928 if (Subtarget->is64Bit()) {
12929 // Join High32 and Low32 into a 64-bit result.
12930 // (High32 << 32) | Low32
12931 Low32 = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::i64, Low32);
12932 High32 = DAG.getNode(ISD::ANY_EXTEND, DL, MVT::i64, High32);
12933 High32 = DAG.getNode(ISD::SHL, DL, MVT::i64, High32,
12934 DAG.getConstant(32, DL, MVT::i8));
12935 SDValue Result = DAG.getNode(ISD::OR, DL, MVT::i64, High32, Low32);
12936 return std::make_pair(Result, SDValue());
12939 SDValue ResultOps[] = { Low32, High32 };
12941 SDValue pair = IsReplace
12942 ? DAG.getNode(ISD::BUILD_PAIR, DL, MVT::i64, ResultOps)
12943 : DAG.getMergeValues(ResultOps, DL);
12944 return std::make_pair(pair, SDValue());
12946 // Build the FP_TO_INT*_IN_MEM
12947 SDValue Ops[] = { Chain, Value, StackSlot };
12948 SDValue FIST = DAG.getMemIntrinsicNode(Opc, DL, DAG.getVTList(MVT::Other),
12950 return std::make_pair(FIST, StackSlot);
12954 static SDValue LowerAVXExtend(SDValue Op, SelectionDAG &DAG,
12955 const X86Subtarget *Subtarget) {
12956 MVT VT = Op->getSimpleValueType(0);
12957 SDValue In = Op->getOperand(0);
12958 MVT InVT = In.getSimpleValueType();
12961 if (VT.is512BitVector() || InVT.getScalarType() == MVT::i1)
12962 return DAG.getNode(ISD::ZERO_EXTEND, dl, VT, In);
12964 // Optimize vectors in AVX mode:
12967 // Use vpunpcklwd for 4 lower elements v8i16 -> v4i32.
12968 // Use vpunpckhwd for 4 upper elements v8i16 -> v4i32.
12969 // Concat upper and lower parts.
12972 // Use vpunpckldq for 4 lower elements v4i32 -> v2i64.
12973 // Use vpunpckhdq for 4 upper elements v4i32 -> v2i64.
12974 // Concat upper and lower parts.
12977 if (((VT != MVT::v16i16) || (InVT != MVT::v16i8)) &&
12978 ((VT != MVT::v8i32) || (InVT != MVT::v8i16)) &&
12979 ((VT != MVT::v4i64) || (InVT != MVT::v4i32)))
12982 if (Subtarget->hasInt256())
12983 return DAG.getNode(X86ISD::VZEXT, dl, VT, In);
12985 SDValue ZeroVec = getZeroVector(InVT, Subtarget, DAG, dl);
12986 SDValue Undef = DAG.getUNDEF(InVT);
12987 bool NeedZero = Op.getOpcode() == ISD::ZERO_EXTEND;
12988 SDValue OpLo = getUnpackl(DAG, dl, InVT, In, NeedZero ? ZeroVec : Undef);
12989 SDValue OpHi = getUnpackh(DAG, dl, InVT, In, NeedZero ? ZeroVec : Undef);
12991 MVT HVT = MVT::getVectorVT(VT.getVectorElementType(),
12992 VT.getVectorNumElements()/2);
12994 OpLo = DAG.getBitcast(HVT, OpLo);
12995 OpHi = DAG.getBitcast(HVT, OpHi);
12997 return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT, OpLo, OpHi);
13000 static SDValue LowerZERO_EXTEND_AVX512(SDValue Op,
13001 const X86Subtarget *Subtarget, SelectionDAG &DAG) {
13002 MVT VT = Op->getSimpleValueType(0);
13003 SDValue In = Op->getOperand(0);
13004 MVT InVT = In.getSimpleValueType();
13006 unsigned int NumElts = VT.getVectorNumElements();
13007 if (NumElts != 8 && NumElts != 16 && !Subtarget->hasBWI())
13010 if (VT.is512BitVector() && InVT.getVectorElementType() != MVT::i1)
13011 return DAG.getNode(X86ISD::VZEXT, DL, VT, In);
13013 assert(InVT.getVectorElementType() == MVT::i1);
13014 MVT ExtVT = NumElts == 8 ? MVT::v8i64 : MVT::v16i32;
13016 DAG.getConstant(APInt(ExtVT.getScalarSizeInBits(), 1), DL, ExtVT);
13018 DAG.getConstant(APInt::getNullValue(ExtVT.getScalarSizeInBits()), DL, ExtVT);
13020 SDValue V = DAG.getNode(ISD::VSELECT, DL, ExtVT, In, One, Zero);
13021 if (VT.is512BitVector())
13023 return DAG.getNode(X86ISD::VTRUNC, DL, VT, V);
13026 static SDValue LowerANY_EXTEND(SDValue Op, const X86Subtarget *Subtarget,
13027 SelectionDAG &DAG) {
13028 if (Subtarget->hasFp256())
13029 if (SDValue Res = LowerAVXExtend(Op, DAG, Subtarget))
13035 static SDValue LowerZERO_EXTEND(SDValue Op, const X86Subtarget *Subtarget,
13036 SelectionDAG &DAG) {
13038 MVT VT = Op.getSimpleValueType();
13039 SDValue In = Op.getOperand(0);
13040 MVT SVT = In.getSimpleValueType();
13042 if (VT.is512BitVector() || SVT.getVectorElementType() == MVT::i1)
13043 return LowerZERO_EXTEND_AVX512(Op, Subtarget, DAG);
13045 if (Subtarget->hasFp256())
13046 if (SDValue Res = LowerAVXExtend(Op, DAG, Subtarget))
13049 assert(!VT.is256BitVector() || !SVT.is128BitVector() ||
13050 VT.getVectorNumElements() != SVT.getVectorNumElements());
13054 SDValue X86TargetLowering::LowerTRUNCATE(SDValue Op, SelectionDAG &DAG) const {
13056 MVT VT = Op.getSimpleValueType();
13057 SDValue In = Op.getOperand(0);
13058 MVT InVT = In.getSimpleValueType();
13060 if (VT == MVT::i1) {
13061 assert((InVT.isInteger() && (InVT.getSizeInBits() <= 64)) &&
13062 "Invalid scalar TRUNCATE operation");
13063 if (InVT.getSizeInBits() >= 32)
13065 In = DAG.getNode(ISD::ANY_EXTEND, DL, MVT::i32, In);
13066 return DAG.getNode(ISD::TRUNCATE, DL, VT, In);
13068 assert(VT.getVectorNumElements() == InVT.getVectorNumElements() &&
13069 "Invalid TRUNCATE operation");
13071 // move vector to mask - truncate solution for SKX
13072 if (VT.getVectorElementType() == MVT::i1) {
13073 if (InVT.is512BitVector() && InVT.getScalarSizeInBits() <= 16 &&
13074 Subtarget->hasBWI())
13075 return Op; // legal, will go to VPMOVB2M, VPMOVW2M
13076 if ((InVT.is256BitVector() || InVT.is128BitVector())
13077 && InVT.getScalarSizeInBits() <= 16 &&
13078 Subtarget->hasBWI() && Subtarget->hasVLX())
13079 return Op; // legal, will go to VPMOVB2M, VPMOVW2M
13080 if (InVT.is512BitVector() && InVT.getScalarSizeInBits() >= 32 &&
13081 Subtarget->hasDQI())
13082 return Op; // legal, will go to VPMOVD2M, VPMOVQ2M
13083 if ((InVT.is256BitVector() || InVT.is128BitVector())
13084 && InVT.getScalarSizeInBits() >= 32 &&
13085 Subtarget->hasDQI() && Subtarget->hasVLX())
13086 return Op; // legal, will go to VPMOVB2M, VPMOVQ2M
13089 if (VT.getVectorElementType() == MVT::i1) {
13090 assert(VT.getVectorElementType() == MVT::i1 && "Unexpected vector type");
13091 unsigned NumElts = InVT.getVectorNumElements();
13092 assert ((NumElts == 8 || NumElts == 16) && "Unexpected vector type");
13093 if (InVT.getSizeInBits() < 512) {
13094 MVT ExtVT = (NumElts == 16)? MVT::v16i32 : MVT::v8i64;
13095 In = DAG.getNode(ISD::SIGN_EXTEND, DL, ExtVT, In);
13100 DAG.getConstant(APInt::getSignBit(InVT.getScalarSizeInBits()), DL, InVT);
13101 SDValue And = DAG.getNode(ISD::AND, DL, InVT, OneV, In);
13102 return DAG.getNode(X86ISD::TESTM, DL, VT, And, And);
13105 // vpmovqb/w/d, vpmovdb/w, vpmovwb
13106 if (((!InVT.is512BitVector() && Subtarget->hasVLX()) || InVT.is512BitVector()) &&
13107 (InVT.getVectorElementType() != MVT::i16 || Subtarget->hasBWI()))
13108 return DAG.getNode(X86ISD::VTRUNC, DL, VT, In);
13110 if ((VT == MVT::v4i32) && (InVT == MVT::v4i64)) {
13111 // On AVX2, v4i64 -> v4i32 becomes VPERMD.
13112 if (Subtarget->hasInt256()) {
13113 static const int ShufMask[] = {0, 2, 4, 6, -1, -1, -1, -1};
13114 In = DAG.getBitcast(MVT::v8i32, In);
13115 In = DAG.getVectorShuffle(MVT::v8i32, DL, In, DAG.getUNDEF(MVT::v8i32),
13117 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, VT, In,
13118 DAG.getIntPtrConstant(0, DL));
13121 SDValue OpLo = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v2i64, In,
13122 DAG.getIntPtrConstant(0, DL));
13123 SDValue OpHi = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v2i64, In,
13124 DAG.getIntPtrConstant(2, DL));
13125 OpLo = DAG.getBitcast(MVT::v4i32, OpLo);
13126 OpHi = DAG.getBitcast(MVT::v4i32, OpHi);
13127 static const int ShufMask[] = {0, 2, 4, 6};
13128 return DAG.getVectorShuffle(VT, DL, OpLo, OpHi, ShufMask);
13131 if ((VT == MVT::v8i16) && (InVT == MVT::v8i32)) {
13132 // On AVX2, v8i32 -> v8i16 becomed PSHUFB.
13133 if (Subtarget->hasInt256()) {
13134 In = DAG.getBitcast(MVT::v32i8, In);
13136 SmallVector<SDValue,32> pshufbMask;
13137 for (unsigned i = 0; i < 2; ++i) {
13138 pshufbMask.push_back(DAG.getConstant(0x0, DL, MVT::i8));
13139 pshufbMask.push_back(DAG.getConstant(0x1, DL, MVT::i8));
13140 pshufbMask.push_back(DAG.getConstant(0x4, DL, MVT::i8));
13141 pshufbMask.push_back(DAG.getConstant(0x5, DL, MVT::i8));
13142 pshufbMask.push_back(DAG.getConstant(0x8, DL, MVT::i8));
13143 pshufbMask.push_back(DAG.getConstant(0x9, DL, MVT::i8));
13144 pshufbMask.push_back(DAG.getConstant(0xc, DL, MVT::i8));
13145 pshufbMask.push_back(DAG.getConstant(0xd, DL, MVT::i8));
13146 for (unsigned j = 0; j < 8; ++j)
13147 pshufbMask.push_back(DAG.getConstant(0x80, DL, MVT::i8));
13149 SDValue BV = DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v32i8, pshufbMask);
13150 In = DAG.getNode(X86ISD::PSHUFB, DL, MVT::v32i8, In, BV);
13151 In = DAG.getBitcast(MVT::v4i64, In);
13153 static const int ShufMask[] = {0, 2, -1, -1};
13154 In = DAG.getVectorShuffle(MVT::v4i64, DL, In, DAG.getUNDEF(MVT::v4i64),
13156 In = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v2i64, In,
13157 DAG.getIntPtrConstant(0, DL));
13158 return DAG.getBitcast(VT, In);
13161 SDValue OpLo = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v4i32, In,
13162 DAG.getIntPtrConstant(0, DL));
13164 SDValue OpHi = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v4i32, In,
13165 DAG.getIntPtrConstant(4, DL));
13167 OpLo = DAG.getBitcast(MVT::v16i8, OpLo);
13168 OpHi = DAG.getBitcast(MVT::v16i8, OpHi);
13170 // The PSHUFB mask:
13171 static const int ShufMask1[] = {0, 1, 4, 5, 8, 9, 12, 13,
13172 -1, -1, -1, -1, -1, -1, -1, -1};
13174 SDValue Undef = DAG.getUNDEF(MVT::v16i8);
13175 OpLo = DAG.getVectorShuffle(MVT::v16i8, DL, OpLo, Undef, ShufMask1);
13176 OpHi = DAG.getVectorShuffle(MVT::v16i8, DL, OpHi, Undef, ShufMask1);
13178 OpLo = DAG.getBitcast(MVT::v4i32, OpLo);
13179 OpHi = DAG.getBitcast(MVT::v4i32, OpHi);
13181 // The MOVLHPS Mask:
13182 static const int ShufMask2[] = {0, 1, 4, 5};
13183 SDValue res = DAG.getVectorShuffle(MVT::v4i32, DL, OpLo, OpHi, ShufMask2);
13184 return DAG.getBitcast(MVT::v8i16, res);
13187 // Handle truncation of V256 to V128 using shuffles.
13188 if (!VT.is128BitVector() || !InVT.is256BitVector())
13191 assert(Subtarget->hasFp256() && "256-bit vector without AVX!");
13193 unsigned NumElems = VT.getVectorNumElements();
13194 MVT NVT = MVT::getVectorVT(VT.getVectorElementType(), NumElems * 2);
13196 SmallVector<int, 16> MaskVec(NumElems * 2, -1);
13197 // Prepare truncation shuffle mask
13198 for (unsigned i = 0; i != NumElems; ++i)
13199 MaskVec[i] = i * 2;
13200 SDValue V = DAG.getVectorShuffle(NVT, DL, DAG.getBitcast(NVT, In),
13201 DAG.getUNDEF(NVT), &MaskVec[0]);
13202 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, VT, V,
13203 DAG.getIntPtrConstant(0, DL));
13206 SDValue X86TargetLowering::LowerFP_TO_SINT(SDValue Op,
13207 SelectionDAG &DAG) const {
13208 assert(!Op.getSimpleValueType().isVector());
13210 std::pair<SDValue,SDValue> Vals = FP_TO_INTHelper(Op, DAG,
13211 /*IsSigned=*/ true, /*IsReplace=*/ false);
13212 SDValue FIST = Vals.first, StackSlot = Vals.second;
13213 // If FP_TO_INTHelper failed, the node is actually supposed to be Legal.
13214 if (!FIST.getNode())
13217 if (StackSlot.getNode())
13218 // Load the result.
13219 return DAG.getLoad(Op.getValueType(), SDLoc(Op),
13220 FIST, StackSlot, MachinePointerInfo(),
13221 false, false, false, 0);
13223 // The node is the result.
13227 SDValue X86TargetLowering::LowerFP_TO_UINT(SDValue Op,
13228 SelectionDAG &DAG) const {
13229 std::pair<SDValue,SDValue> Vals = FP_TO_INTHelper(Op, DAG,
13230 /*IsSigned=*/ false, /*IsReplace=*/ false);
13231 SDValue FIST = Vals.first, StackSlot = Vals.second;
13232 // If FP_TO_INTHelper failed, the node is actually supposed to be Legal.
13233 if (!FIST.getNode())
13236 if (StackSlot.getNode())
13237 // Load the result.
13238 return DAG.getLoad(Op.getValueType(), SDLoc(Op),
13239 FIST, StackSlot, MachinePointerInfo(),
13240 false, false, false, 0);
13242 // The node is the result.
13246 static SDValue LowerFP_EXTEND(SDValue Op, SelectionDAG &DAG) {
13248 MVT VT = Op.getSimpleValueType();
13249 SDValue In = Op.getOperand(0);
13250 MVT SVT = In.getSimpleValueType();
13252 assert(SVT == MVT::v2f32 && "Only customize MVT::v2f32 type legalization!");
13254 return DAG.getNode(X86ISD::VFPEXT, DL, VT,
13255 DAG.getNode(ISD::CONCAT_VECTORS, DL, MVT::v4f32,
13256 In, DAG.getUNDEF(SVT)));
13259 /// The only differences between FABS and FNEG are the mask and the logic op.
13260 /// FNEG also has a folding opportunity for FNEG(FABS(x)).
13261 static SDValue LowerFABSorFNEG(SDValue Op, SelectionDAG &DAG) {
13262 assert((Op.getOpcode() == ISD::FABS || Op.getOpcode() == ISD::FNEG) &&
13263 "Wrong opcode for lowering FABS or FNEG.");
13265 bool IsFABS = (Op.getOpcode() == ISD::FABS);
13267 // If this is a FABS and it has an FNEG user, bail out to fold the combination
13268 // into an FNABS. We'll lower the FABS after that if it is still in use.
13270 for (SDNode *User : Op->uses())
13271 if (User->getOpcode() == ISD::FNEG)
13275 MVT VT = Op.getSimpleValueType();
13277 // FIXME: Use function attribute "OptimizeForSize" and/or CodeGenOpt::Level to
13278 // decide if we should generate a 16-byte constant mask when we only need 4 or
13279 // 8 bytes for the scalar case.
13285 if (VT.isVector()) {
13287 EltVT = VT.getVectorElementType();
13288 NumElts = VT.getVectorNumElements();
13290 // There are no scalar bitwise logical SSE/AVX instructions, so we
13291 // generate a 16-byte vector constant and logic op even for the scalar case.
13292 // Using a 16-byte mask allows folding the load of the mask with
13293 // the logic op, so it can save (~4 bytes) on code size.
13294 LogicVT = (VT == MVT::f64) ? MVT::v2f64 : MVT::v4f32;
13296 NumElts = (VT == MVT::f64) ? 2 : 4;
13299 unsigned EltBits = EltVT.getSizeInBits();
13300 LLVMContext *Context = DAG.getContext();
13301 // For FABS, mask is 0x7f...; for FNEG, mask is 0x80...
13303 IsFABS ? APInt::getSignedMaxValue(EltBits) : APInt::getSignBit(EltBits);
13304 Constant *C = ConstantInt::get(*Context, MaskElt);
13305 C = ConstantVector::getSplat(NumElts, C);
13306 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
13307 SDValue CPIdx = DAG.getConstantPool(C, TLI.getPointerTy(DAG.getDataLayout()));
13308 unsigned Alignment = cast<ConstantPoolSDNode>(CPIdx)->getAlignment();
13310 DAG.getLoad(LogicVT, dl, DAG.getEntryNode(), CPIdx,
13311 MachinePointerInfo::getConstantPool(DAG.getMachineFunction()),
13312 false, false, false, Alignment);
13314 SDValue Op0 = Op.getOperand(0);
13315 bool IsFNABS = !IsFABS && (Op0.getOpcode() == ISD::FABS);
13317 IsFABS ? X86ISD::FAND : IsFNABS ? X86ISD::FOR : X86ISD::FXOR;
13318 SDValue Operand = IsFNABS ? Op0.getOperand(0) : Op0;
13321 return DAG.getNode(LogicOp, dl, LogicVT, Operand, Mask);
13323 // For the scalar case extend to a 128-bit vector, perform the logic op,
13324 // and extract the scalar result back out.
13325 Operand = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, LogicVT, Operand);
13326 SDValue LogicNode = DAG.getNode(LogicOp, dl, LogicVT, Operand, Mask);
13327 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, LogicNode,
13328 DAG.getIntPtrConstant(0, dl));
13331 static SDValue LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) {
13332 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
13333 LLVMContext *Context = DAG.getContext();
13334 SDValue Op0 = Op.getOperand(0);
13335 SDValue Op1 = Op.getOperand(1);
13337 MVT VT = Op.getSimpleValueType();
13338 MVT SrcVT = Op1.getSimpleValueType();
13340 // If second operand is smaller, extend it first.
13341 if (SrcVT.bitsLT(VT)) {
13342 Op1 = DAG.getNode(ISD::FP_EXTEND, dl, VT, Op1);
13345 // And if it is bigger, shrink it first.
13346 if (SrcVT.bitsGT(VT)) {
13347 Op1 = DAG.getNode(ISD::FP_ROUND, dl, VT, Op1, DAG.getIntPtrConstant(1, dl));
13351 // At this point the operands and the result should have the same
13352 // type, and that won't be f80 since that is not custom lowered.
13354 const fltSemantics &Sem =
13355 VT == MVT::f64 ? APFloat::IEEEdouble : APFloat::IEEEsingle;
13356 const unsigned SizeInBits = VT.getSizeInBits();
13358 SmallVector<Constant *, 4> CV(
13359 VT == MVT::f64 ? 2 : 4,
13360 ConstantFP::get(*Context, APFloat(Sem, APInt(SizeInBits, 0))));
13362 // First, clear all bits but the sign bit from the second operand (sign).
13363 CV[0] = ConstantFP::get(*Context,
13364 APFloat(Sem, APInt::getHighBitsSet(SizeInBits, 1)));
13365 Constant *C = ConstantVector::get(CV);
13366 auto PtrVT = TLI.getPointerTy(DAG.getDataLayout());
13367 SDValue CPIdx = DAG.getConstantPool(C, PtrVT, 16);
13369 // Perform all logic operations as 16-byte vectors because there are no
13370 // scalar FP logic instructions in SSE. This allows load folding of the
13371 // constants into the logic instructions.
13372 MVT LogicVT = (VT == MVT::f64) ? MVT::v2f64 : MVT::v4f32;
13374 DAG.getLoad(LogicVT, dl, DAG.getEntryNode(), CPIdx,
13375 MachinePointerInfo::getConstantPool(DAG.getMachineFunction()),
13376 false, false, false, 16);
13377 Op1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, LogicVT, Op1);
13378 SDValue SignBit = DAG.getNode(X86ISD::FAND, dl, LogicVT, Op1, Mask1);
13380 // Next, clear the sign bit from the first operand (magnitude).
13381 // If it's a constant, we can clear it here.
13382 if (ConstantFPSDNode *Op0CN = dyn_cast<ConstantFPSDNode>(Op0)) {
13383 APFloat APF = Op0CN->getValueAPF();
13384 // If the magnitude is a positive zero, the sign bit alone is enough.
13385 if (APF.isPosZero())
13386 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, SrcVT, SignBit,
13387 DAG.getIntPtrConstant(0, dl));
13389 CV[0] = ConstantFP::get(*Context, APF);
13391 CV[0] = ConstantFP::get(
13393 APFloat(Sem, APInt::getLowBitsSet(SizeInBits, SizeInBits - 1)));
13395 C = ConstantVector::get(CV);
13396 CPIdx = DAG.getConstantPool(C, PtrVT, 16);
13398 DAG.getLoad(LogicVT, dl, DAG.getEntryNode(), CPIdx,
13399 MachinePointerInfo::getConstantPool(DAG.getMachineFunction()),
13400 false, false, false, 16);
13401 // If the magnitude operand wasn't a constant, we need to AND out the sign.
13402 if (!isa<ConstantFPSDNode>(Op0)) {
13403 Op0 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, LogicVT, Op0);
13404 Val = DAG.getNode(X86ISD::FAND, dl, LogicVT, Op0, Val);
13406 // OR the magnitude value with the sign bit.
13407 Val = DAG.getNode(X86ISD::FOR, dl, LogicVT, Val, SignBit);
13408 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, SrcVT, Val,
13409 DAG.getIntPtrConstant(0, dl));
13412 static SDValue LowerFGETSIGN(SDValue Op, SelectionDAG &DAG) {
13413 SDValue N0 = Op.getOperand(0);
13415 MVT VT = Op.getSimpleValueType();
13417 // Lower ISD::FGETSIGN to (AND (X86ISD::FGETSIGNx86 ...) 1).
13418 SDValue xFGETSIGN = DAG.getNode(X86ISD::FGETSIGNx86, dl, VT, N0,
13419 DAG.getConstant(1, dl, VT));
13420 return DAG.getNode(ISD::AND, dl, VT, xFGETSIGN, DAG.getConstant(1, dl, VT));
13423 // Check whether an OR'd tree is PTEST-able.
13424 static SDValue LowerVectorAllZeroTest(SDValue Op, const X86Subtarget *Subtarget,
13425 SelectionDAG &DAG) {
13426 assert(Op.getOpcode() == ISD::OR && "Only check OR'd tree.");
13428 if (!Subtarget->hasSSE41())
13431 if (!Op->hasOneUse())
13434 SDNode *N = Op.getNode();
13437 SmallVector<SDValue, 8> Opnds;
13438 DenseMap<SDValue, unsigned> VecInMap;
13439 SmallVector<SDValue, 8> VecIns;
13440 EVT VT = MVT::Other;
13442 // Recognize a special case where a vector is casted into wide integer to
13444 Opnds.push_back(N->getOperand(0));
13445 Opnds.push_back(N->getOperand(1));
13447 for (unsigned Slot = 0, e = Opnds.size(); Slot < e; ++Slot) {
13448 SmallVectorImpl<SDValue>::const_iterator I = Opnds.begin() + Slot;
13449 // BFS traverse all OR'd operands.
13450 if (I->getOpcode() == ISD::OR) {
13451 Opnds.push_back(I->getOperand(0));
13452 Opnds.push_back(I->getOperand(1));
13453 // Re-evaluate the number of nodes to be traversed.
13454 e += 2; // 2 more nodes (LHS and RHS) are pushed.
13458 // Quit if a non-EXTRACT_VECTOR_ELT
13459 if (I->getOpcode() != ISD::EXTRACT_VECTOR_ELT)
13462 // Quit if without a constant index.
13463 SDValue Idx = I->getOperand(1);
13464 if (!isa<ConstantSDNode>(Idx))
13467 SDValue ExtractedFromVec = I->getOperand(0);
13468 DenseMap<SDValue, unsigned>::iterator M = VecInMap.find(ExtractedFromVec);
13469 if (M == VecInMap.end()) {
13470 VT = ExtractedFromVec.getValueType();
13471 // Quit if not 128/256-bit vector.
13472 if (!VT.is128BitVector() && !VT.is256BitVector())
13474 // Quit if not the same type.
13475 if (VecInMap.begin() != VecInMap.end() &&
13476 VT != VecInMap.begin()->first.getValueType())
13478 M = VecInMap.insert(std::make_pair(ExtractedFromVec, 0)).first;
13479 VecIns.push_back(ExtractedFromVec);
13481 M->second |= 1U << cast<ConstantSDNode>(Idx)->getZExtValue();
13484 assert((VT.is128BitVector() || VT.is256BitVector()) &&
13485 "Not extracted from 128-/256-bit vector.");
13487 unsigned FullMask = (1U << VT.getVectorNumElements()) - 1U;
13489 for (DenseMap<SDValue, unsigned>::const_iterator
13490 I = VecInMap.begin(), E = VecInMap.end(); I != E; ++I) {
13491 // Quit if not all elements are used.
13492 if (I->second != FullMask)
13496 EVT TestVT = VT.is128BitVector() ? MVT::v2i64 : MVT::v4i64;
13498 // Cast all vectors into TestVT for PTEST.
13499 for (unsigned i = 0, e = VecIns.size(); i < e; ++i)
13500 VecIns[i] = DAG.getBitcast(TestVT, VecIns[i]);
13502 // If more than one full vectors are evaluated, OR them first before PTEST.
13503 for (unsigned Slot = 0, e = VecIns.size(); e - Slot > 1; Slot += 2, e += 1) {
13504 // Each iteration will OR 2 nodes and append the result until there is only
13505 // 1 node left, i.e. the final OR'd value of all vectors.
13506 SDValue LHS = VecIns[Slot];
13507 SDValue RHS = VecIns[Slot + 1];
13508 VecIns.push_back(DAG.getNode(ISD::OR, DL, TestVT, LHS, RHS));
13511 return DAG.getNode(X86ISD::PTEST, DL, MVT::i32,
13512 VecIns.back(), VecIns.back());
13515 /// \brief return true if \c Op has a use that doesn't just read flags.
13516 static bool hasNonFlagsUse(SDValue Op) {
13517 for (SDNode::use_iterator UI = Op->use_begin(), UE = Op->use_end(); UI != UE;
13519 SDNode *User = *UI;
13520 unsigned UOpNo = UI.getOperandNo();
13521 if (User->getOpcode() == ISD::TRUNCATE && User->hasOneUse()) {
13522 // Look pass truncate.
13523 UOpNo = User->use_begin().getOperandNo();
13524 User = *User->use_begin();
13527 if (User->getOpcode() != ISD::BRCOND && User->getOpcode() != ISD::SETCC &&
13528 !(User->getOpcode() == ISD::SELECT && UOpNo == 0))
13534 /// Emit nodes that will be selected as "test Op0,Op0", or something
13536 SDValue X86TargetLowering::EmitTest(SDValue Op, unsigned X86CC, SDLoc dl,
13537 SelectionDAG &DAG) const {
13538 if (Op.getValueType() == MVT::i1) {
13539 SDValue ExtOp = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i8, Op);
13540 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, ExtOp,
13541 DAG.getConstant(0, dl, MVT::i8));
13543 // CF and OF aren't always set the way we want. Determine which
13544 // of these we need.
13545 bool NeedCF = false;
13546 bool NeedOF = false;
13549 case X86::COND_A: case X86::COND_AE:
13550 case X86::COND_B: case X86::COND_BE:
13553 case X86::COND_G: case X86::COND_GE:
13554 case X86::COND_L: case X86::COND_LE:
13555 case X86::COND_O: case X86::COND_NO: {
13556 // Check if we really need to set the
13557 // Overflow flag. If NoSignedWrap is present
13558 // that is not actually needed.
13559 switch (Op->getOpcode()) {
13564 const auto *BinNode = cast<BinaryWithFlagsSDNode>(Op.getNode());
13565 if (BinNode->Flags.hasNoSignedWrap())
13575 // See if we can use the EFLAGS value from the operand instead of
13576 // doing a separate TEST. TEST always sets OF and CF to 0, so unless
13577 // we prove that the arithmetic won't overflow, we can't use OF or CF.
13578 if (Op.getResNo() != 0 || NeedOF || NeedCF) {
13579 // Emit a CMP with 0, which is the TEST pattern.
13580 //if (Op.getValueType() == MVT::i1)
13581 // return DAG.getNode(X86ISD::CMP, dl, MVT::i1, Op,
13582 // DAG.getConstant(0, MVT::i1));
13583 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op,
13584 DAG.getConstant(0, dl, Op.getValueType()));
13586 unsigned Opcode = 0;
13587 unsigned NumOperands = 0;
13589 // Truncate operations may prevent the merge of the SETCC instruction
13590 // and the arithmetic instruction before it. Attempt to truncate the operands
13591 // of the arithmetic instruction and use a reduced bit-width instruction.
13592 bool NeedTruncation = false;
13593 SDValue ArithOp = Op;
13594 if (Op->getOpcode() == ISD::TRUNCATE && Op->hasOneUse()) {
13595 SDValue Arith = Op->getOperand(0);
13596 // Both the trunc and the arithmetic op need to have one user each.
13597 if (Arith->hasOneUse())
13598 switch (Arith.getOpcode()) {
13605 NeedTruncation = true;
13611 // NOTICE: In the code below we use ArithOp to hold the arithmetic operation
13612 // which may be the result of a CAST. We use the variable 'Op', which is the
13613 // non-casted variable when we check for possible users.
13614 switch (ArithOp.getOpcode()) {
13616 // Due to an isel shortcoming, be conservative if this add is likely to be
13617 // selected as part of a load-modify-store instruction. When the root node
13618 // in a match is a store, isel doesn't know how to remap non-chain non-flag
13619 // uses of other nodes in the match, such as the ADD in this case. This
13620 // leads to the ADD being left around and reselected, with the result being
13621 // two adds in the output. Alas, even if none our users are stores, that
13622 // doesn't prove we're O.K. Ergo, if we have any parents that aren't
13623 // CopyToReg or SETCC, eschew INC/DEC. A better fix seems to require
13624 // climbing the DAG back to the root, and it doesn't seem to be worth the
13626 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
13627 UE = Op.getNode()->use_end(); UI != UE; ++UI)
13628 if (UI->getOpcode() != ISD::CopyToReg &&
13629 UI->getOpcode() != ISD::SETCC &&
13630 UI->getOpcode() != ISD::STORE)
13633 if (ConstantSDNode *C =
13634 dyn_cast<ConstantSDNode>(ArithOp.getNode()->getOperand(1))) {
13635 // An add of one will be selected as an INC.
13636 if (C->getAPIntValue() == 1 && !Subtarget->slowIncDec()) {
13637 Opcode = X86ISD::INC;
13642 // An add of negative one (subtract of one) will be selected as a DEC.
13643 if (C->getAPIntValue().isAllOnesValue() && !Subtarget->slowIncDec()) {
13644 Opcode = X86ISD::DEC;
13650 // Otherwise use a regular EFLAGS-setting add.
13651 Opcode = X86ISD::ADD;
13656 // If we have a constant logical shift that's only used in a comparison
13657 // against zero turn it into an equivalent AND. This allows turning it into
13658 // a TEST instruction later.
13659 if ((X86CC == X86::COND_E || X86CC == X86::COND_NE) && Op->hasOneUse() &&
13660 isa<ConstantSDNode>(Op->getOperand(1)) && !hasNonFlagsUse(Op)) {
13661 EVT VT = Op.getValueType();
13662 unsigned BitWidth = VT.getSizeInBits();
13663 unsigned ShAmt = Op->getConstantOperandVal(1);
13664 if (ShAmt >= BitWidth) // Avoid undefined shifts.
13666 APInt Mask = ArithOp.getOpcode() == ISD::SRL
13667 ? APInt::getHighBitsSet(BitWidth, BitWidth - ShAmt)
13668 : APInt::getLowBitsSet(BitWidth, BitWidth - ShAmt);
13669 if (!Mask.isSignedIntN(32)) // Avoid large immediates.
13671 SDValue New = DAG.getNode(ISD::AND, dl, VT, Op->getOperand(0),
13672 DAG.getConstant(Mask, dl, VT));
13673 DAG.ReplaceAllUsesWith(Op, New);
13679 // If the primary and result isn't used, don't bother using X86ISD::AND,
13680 // because a TEST instruction will be better.
13681 if (!hasNonFlagsUse(Op))
13687 // Due to the ISEL shortcoming noted above, be conservative if this op is
13688 // likely to be selected as part of a load-modify-store instruction.
13689 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
13690 UE = Op.getNode()->use_end(); UI != UE; ++UI)
13691 if (UI->getOpcode() == ISD::STORE)
13694 // Otherwise use a regular EFLAGS-setting instruction.
13695 switch (ArithOp.getOpcode()) {
13696 default: llvm_unreachable("unexpected operator!");
13697 case ISD::SUB: Opcode = X86ISD::SUB; break;
13698 case ISD::XOR: Opcode = X86ISD::XOR; break;
13699 case ISD::AND: Opcode = X86ISD::AND; break;
13701 if (!NeedTruncation && (X86CC == X86::COND_E || X86CC == X86::COND_NE)) {
13702 SDValue EFLAGS = LowerVectorAllZeroTest(Op, Subtarget, DAG);
13703 if (EFLAGS.getNode())
13706 Opcode = X86ISD::OR;
13720 return SDValue(Op.getNode(), 1);
13726 // If we found that truncation is beneficial, perform the truncation and
13728 if (NeedTruncation) {
13729 EVT VT = Op.getValueType();
13730 SDValue WideVal = Op->getOperand(0);
13731 EVT WideVT = WideVal.getValueType();
13732 unsigned ConvertedOp = 0;
13733 // Use a target machine opcode to prevent further DAGCombine
13734 // optimizations that may separate the arithmetic operations
13735 // from the setcc node.
13736 switch (WideVal.getOpcode()) {
13738 case ISD::ADD: ConvertedOp = X86ISD::ADD; break;
13739 case ISD::SUB: ConvertedOp = X86ISD::SUB; break;
13740 case ISD::AND: ConvertedOp = X86ISD::AND; break;
13741 case ISD::OR: ConvertedOp = X86ISD::OR; break;
13742 case ISD::XOR: ConvertedOp = X86ISD::XOR; break;
13746 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
13747 if (TLI.isOperationLegal(WideVal.getOpcode(), WideVT)) {
13748 SDValue V0 = DAG.getNode(ISD::TRUNCATE, dl, VT, WideVal.getOperand(0));
13749 SDValue V1 = DAG.getNode(ISD::TRUNCATE, dl, VT, WideVal.getOperand(1));
13750 Op = DAG.getNode(ConvertedOp, dl, VT, V0, V1);
13756 // Emit a CMP with 0, which is the TEST pattern.
13757 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op,
13758 DAG.getConstant(0, dl, Op.getValueType()));
13760 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::i32);
13761 SmallVector<SDValue, 4> Ops(Op->op_begin(), Op->op_begin() + NumOperands);
13763 SDValue New = DAG.getNode(Opcode, dl, VTs, Ops);
13764 DAG.ReplaceAllUsesWith(Op, New);
13765 return SDValue(New.getNode(), 1);
13768 /// Emit nodes that will be selected as "cmp Op0,Op1", or something
13770 SDValue X86TargetLowering::EmitCmp(SDValue Op0, SDValue Op1, unsigned X86CC,
13771 SDLoc dl, SelectionDAG &DAG) const {
13772 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op1)) {
13773 if (C->getAPIntValue() == 0)
13774 return EmitTest(Op0, X86CC, dl, DAG);
13776 if (Op0.getValueType() == MVT::i1)
13777 llvm_unreachable("Unexpected comparison operation for MVT::i1 operands");
13780 if ((Op0.getValueType() == MVT::i8 || Op0.getValueType() == MVT::i16 ||
13781 Op0.getValueType() == MVT::i32 || Op0.getValueType() == MVT::i64)) {
13782 // Do the comparison at i32 if it's smaller, besides the Atom case.
13783 // This avoids subregister aliasing issues. Keep the smaller reference
13784 // if we're optimizing for size, however, as that'll allow better folding
13785 // of memory operations.
13786 if (Op0.getValueType() != MVT::i32 && Op0.getValueType() != MVT::i64 &&
13787 !DAG.getMachineFunction().getFunction()->optForMinSize() &&
13788 !Subtarget->isAtom()) {
13789 unsigned ExtendOp =
13790 isX86CCUnsigned(X86CC) ? ISD::ZERO_EXTEND : ISD::SIGN_EXTEND;
13791 Op0 = DAG.getNode(ExtendOp, dl, MVT::i32, Op0);
13792 Op1 = DAG.getNode(ExtendOp, dl, MVT::i32, Op1);
13794 // Use SUB instead of CMP to enable CSE between SUB and CMP.
13795 SDVTList VTs = DAG.getVTList(Op0.getValueType(), MVT::i32);
13796 SDValue Sub = DAG.getNode(X86ISD::SUB, dl, VTs,
13798 return SDValue(Sub.getNode(), 1);
13800 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op0, Op1);
13803 /// Convert a comparison if required by the subtarget.
13804 SDValue X86TargetLowering::ConvertCmpIfNecessary(SDValue Cmp,
13805 SelectionDAG &DAG) const {
13806 // If the subtarget does not support the FUCOMI instruction, floating-point
13807 // comparisons have to be converted.
13808 if (Subtarget->hasCMov() ||
13809 Cmp.getOpcode() != X86ISD::CMP ||
13810 !Cmp.getOperand(0).getValueType().isFloatingPoint() ||
13811 !Cmp.getOperand(1).getValueType().isFloatingPoint())
13814 // The instruction selector will select an FUCOM instruction instead of
13815 // FUCOMI, which writes the comparison result to FPSW instead of EFLAGS. Hence
13816 // build an SDNode sequence that transfers the result from FPSW into EFLAGS:
13817 // (X86sahf (trunc (srl (X86fp_stsw (trunc (X86cmp ...)), 8))))
13819 SDValue TruncFPSW = DAG.getNode(ISD::TRUNCATE, dl, MVT::i16, Cmp);
13820 SDValue FNStSW = DAG.getNode(X86ISD::FNSTSW16r, dl, MVT::i16, TruncFPSW);
13821 SDValue Srl = DAG.getNode(ISD::SRL, dl, MVT::i16, FNStSW,
13822 DAG.getConstant(8, dl, MVT::i8));
13823 SDValue TruncSrl = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Srl);
13824 return DAG.getNode(X86ISD::SAHF, dl, MVT::i32, TruncSrl);
13827 /// The minimum architected relative accuracy is 2^-12. We need one
13828 /// Newton-Raphson step to have a good float result (24 bits of precision).
13829 SDValue X86TargetLowering::getRsqrtEstimate(SDValue Op,
13830 DAGCombinerInfo &DCI,
13831 unsigned &RefinementSteps,
13832 bool &UseOneConstNR) const {
13833 EVT VT = Op.getValueType();
13834 const char *RecipOp;
13836 // SSE1 has rsqrtss and rsqrtps. AVX adds a 256-bit variant for rsqrtps.
13837 // TODO: Add support for AVX512 (v16f32).
13838 // It is likely not profitable to do this for f64 because a double-precision
13839 // rsqrt estimate with refinement on x86 prior to FMA requires at least 16
13840 // instructions: convert to single, rsqrtss, convert back to double, refine
13841 // (3 steps = at least 13 insts). If an 'rsqrtsd' variant was added to the ISA
13842 // along with FMA, this could be a throughput win.
13843 if (VT == MVT::f32 && Subtarget->hasSSE1())
13845 else if ((VT == MVT::v4f32 && Subtarget->hasSSE1()) ||
13846 (VT == MVT::v8f32 && Subtarget->hasAVX()))
13847 RecipOp = "vec-sqrtf";
13851 TargetRecip Recips = DCI.DAG.getTarget().Options.Reciprocals;
13852 if (!Recips.isEnabled(RecipOp))
13855 RefinementSteps = Recips.getRefinementSteps(RecipOp);
13856 UseOneConstNR = false;
13857 return DCI.DAG.getNode(X86ISD::FRSQRT, SDLoc(Op), VT, Op);
13860 /// The minimum architected relative accuracy is 2^-12. We need one
13861 /// Newton-Raphson step to have a good float result (24 bits of precision).
13862 SDValue X86TargetLowering::getRecipEstimate(SDValue Op,
13863 DAGCombinerInfo &DCI,
13864 unsigned &RefinementSteps) const {
13865 EVT VT = Op.getValueType();
13866 const char *RecipOp;
13868 // SSE1 has rcpss and rcpps. AVX adds a 256-bit variant for rcpps.
13869 // TODO: Add support for AVX512 (v16f32).
13870 // It is likely not profitable to do this for f64 because a double-precision
13871 // reciprocal estimate with refinement on x86 prior to FMA requires
13872 // 15 instructions: convert to single, rcpss, convert back to double, refine
13873 // (3 steps = 12 insts). If an 'rcpsd' variant was added to the ISA
13874 // along with FMA, this could be a throughput win.
13875 if (VT == MVT::f32 && Subtarget->hasSSE1())
13877 else if ((VT == MVT::v4f32 && Subtarget->hasSSE1()) ||
13878 (VT == MVT::v8f32 && Subtarget->hasAVX()))
13879 RecipOp = "vec-divf";
13883 TargetRecip Recips = DCI.DAG.getTarget().Options.Reciprocals;
13884 if (!Recips.isEnabled(RecipOp))
13887 RefinementSteps = Recips.getRefinementSteps(RecipOp);
13888 return DCI.DAG.getNode(X86ISD::FRCP, SDLoc(Op), VT, Op);
13891 /// If we have at least two divisions that use the same divisor, convert to
13892 /// multplication by a reciprocal. This may need to be adjusted for a given
13893 /// CPU if a division's cost is not at least twice the cost of a multiplication.
13894 /// This is because we still need one division to calculate the reciprocal and
13895 /// then we need two multiplies by that reciprocal as replacements for the
13896 /// original divisions.
13897 unsigned X86TargetLowering::combineRepeatedFPDivisors() const {
13901 static bool isAllOnes(SDValue V) {
13902 ConstantSDNode *C = dyn_cast<ConstantSDNode>(V);
13903 return C && C->isAllOnesValue();
13906 /// LowerToBT - Result of 'and' is compared against zero. Turn it into a BT node
13907 /// if it's possible.
13908 SDValue X86TargetLowering::LowerToBT(SDValue And, ISD::CondCode CC,
13909 SDLoc dl, SelectionDAG &DAG) const {
13910 SDValue Op0 = And.getOperand(0);
13911 SDValue Op1 = And.getOperand(1);
13912 if (Op0.getOpcode() == ISD::TRUNCATE)
13913 Op0 = Op0.getOperand(0);
13914 if (Op1.getOpcode() == ISD::TRUNCATE)
13915 Op1 = Op1.getOperand(0);
13918 if (Op1.getOpcode() == ISD::SHL)
13919 std::swap(Op0, Op1);
13920 if (Op0.getOpcode() == ISD::SHL) {
13921 if (ConstantSDNode *And00C = dyn_cast<ConstantSDNode>(Op0.getOperand(0)))
13922 if (And00C->getZExtValue() == 1) {
13923 // If we looked past a truncate, check that it's only truncating away
13925 unsigned BitWidth = Op0.getValueSizeInBits();
13926 unsigned AndBitWidth = And.getValueSizeInBits();
13927 if (BitWidth > AndBitWidth) {
13929 DAG.computeKnownBits(Op0, Zeros, Ones);
13930 if (Zeros.countLeadingOnes() < BitWidth - AndBitWidth)
13934 RHS = Op0.getOperand(1);
13936 } else if (Op1.getOpcode() == ISD::Constant) {
13937 ConstantSDNode *AndRHS = cast<ConstantSDNode>(Op1);
13938 uint64_t AndRHSVal = AndRHS->getZExtValue();
13939 SDValue AndLHS = Op0;
13941 if (AndRHSVal == 1 && AndLHS.getOpcode() == ISD::SRL) {
13942 LHS = AndLHS.getOperand(0);
13943 RHS = AndLHS.getOperand(1);
13946 // Use BT if the immediate can't be encoded in a TEST instruction.
13947 if (!isUInt<32>(AndRHSVal) && isPowerOf2_64(AndRHSVal)) {
13949 RHS = DAG.getConstant(Log2_64_Ceil(AndRHSVal), dl, LHS.getValueType());
13953 if (LHS.getNode()) {
13954 // If LHS is i8, promote it to i32 with any_extend. There is no i8 BT
13955 // instruction. Since the shift amount is in-range-or-undefined, we know
13956 // that doing a bittest on the i32 value is ok. We extend to i32 because
13957 // the encoding for the i16 version is larger than the i32 version.
13958 // Also promote i16 to i32 for performance / code size reason.
13959 if (LHS.getValueType() == MVT::i8 ||
13960 LHS.getValueType() == MVT::i16)
13961 LHS = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, LHS);
13963 // If the operand types disagree, extend the shift amount to match. Since
13964 // BT ignores high bits (like shifts) we can use anyextend.
13965 if (LHS.getValueType() != RHS.getValueType())
13966 RHS = DAG.getNode(ISD::ANY_EXTEND, dl, LHS.getValueType(), RHS);
13968 SDValue BT = DAG.getNode(X86ISD::BT, dl, MVT::i32, LHS, RHS);
13969 X86::CondCode Cond = CC == ISD::SETEQ ? X86::COND_AE : X86::COND_B;
13970 return DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
13971 DAG.getConstant(Cond, dl, MVT::i8), BT);
13977 /// \brief - Turns an ISD::CondCode into a value suitable for SSE floating point
13979 static int translateX86FSETCC(ISD::CondCode SetCCOpcode, SDValue &Op0,
13984 // SSE Condition code mapping:
13993 switch (SetCCOpcode) {
13994 default: llvm_unreachable("Unexpected SETCC condition");
13996 case ISD::SETEQ: SSECC = 0; break;
13998 case ISD::SETGT: Swap = true; // Fallthrough
14000 case ISD::SETOLT: SSECC = 1; break;
14002 case ISD::SETGE: Swap = true; // Fallthrough
14004 case ISD::SETOLE: SSECC = 2; break;
14005 case ISD::SETUO: SSECC = 3; break;
14007 case ISD::SETNE: SSECC = 4; break;
14008 case ISD::SETULE: Swap = true; // Fallthrough
14009 case ISD::SETUGE: SSECC = 5; break;
14010 case ISD::SETULT: Swap = true; // Fallthrough
14011 case ISD::SETUGT: SSECC = 6; break;
14012 case ISD::SETO: SSECC = 7; break;
14014 case ISD::SETONE: SSECC = 8; break;
14017 std::swap(Op0, Op1);
14022 // Lower256IntVSETCC - Break a VSETCC 256-bit integer VSETCC into two new 128
14023 // ones, and then concatenate the result back.
14024 static SDValue Lower256IntVSETCC(SDValue Op, SelectionDAG &DAG) {
14025 MVT VT = Op.getSimpleValueType();
14027 assert(VT.is256BitVector() && Op.getOpcode() == ISD::SETCC &&
14028 "Unsupported value type for operation");
14030 unsigned NumElems = VT.getVectorNumElements();
14032 SDValue CC = Op.getOperand(2);
14034 // Extract the LHS vectors
14035 SDValue LHS = Op.getOperand(0);
14036 SDValue LHS1 = Extract128BitVector(LHS, 0, DAG, dl);
14037 SDValue LHS2 = Extract128BitVector(LHS, NumElems/2, DAG, dl);
14039 // Extract the RHS vectors
14040 SDValue RHS = Op.getOperand(1);
14041 SDValue RHS1 = Extract128BitVector(RHS, 0, DAG, dl);
14042 SDValue RHS2 = Extract128BitVector(RHS, NumElems/2, DAG, dl);
14044 // Issue the operation on the smaller types and concatenate the result back
14045 MVT EltVT = VT.getVectorElementType();
14046 MVT NewVT = MVT::getVectorVT(EltVT, NumElems/2);
14047 return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT,
14048 DAG.getNode(Op.getOpcode(), dl, NewVT, LHS1, RHS1, CC),
14049 DAG.getNode(Op.getOpcode(), dl, NewVT, LHS2, RHS2, CC));
14052 static SDValue LowerBoolVSETCC_AVX512(SDValue Op, SelectionDAG &DAG) {
14053 SDValue Op0 = Op.getOperand(0);
14054 SDValue Op1 = Op.getOperand(1);
14055 SDValue CC = Op.getOperand(2);
14056 MVT VT = Op.getSimpleValueType();
14059 assert(Op0.getValueType().getVectorElementType() == MVT::i1 &&
14060 "Unexpected type for boolean compare operation");
14061 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
14062 SDValue NotOp0 = DAG.getNode(ISD::XOR, dl, VT, Op0,
14063 DAG.getConstant(-1, dl, VT));
14064 SDValue NotOp1 = DAG.getNode(ISD::XOR, dl, VT, Op1,
14065 DAG.getConstant(-1, dl, VT));
14066 switch (SetCCOpcode) {
14067 default: llvm_unreachable("Unexpected SETCC condition");
14069 // (x == y) -> ~(x ^ y)
14070 return DAG.getNode(ISD::XOR, dl, VT,
14071 DAG.getNode(ISD::XOR, dl, VT, Op0, Op1),
14072 DAG.getConstant(-1, dl, VT));
14074 // (x != y) -> (x ^ y)
14075 return DAG.getNode(ISD::XOR, dl, VT, Op0, Op1);
14078 // (x > y) -> (x & ~y)
14079 return DAG.getNode(ISD::AND, dl, VT, Op0, NotOp1);
14082 // (x < y) -> (~x & y)
14083 return DAG.getNode(ISD::AND, dl, VT, NotOp0, Op1);
14086 // (x <= y) -> (~x | y)
14087 return DAG.getNode(ISD::OR, dl, VT, NotOp0, Op1);
14090 // (x >=y) -> (x | ~y)
14091 return DAG.getNode(ISD::OR, dl, VT, Op0, NotOp1);
14095 static SDValue LowerIntVSETCC_AVX512(SDValue Op, SelectionDAG &DAG,
14096 const X86Subtarget *Subtarget) {
14097 SDValue Op0 = Op.getOperand(0);
14098 SDValue Op1 = Op.getOperand(1);
14099 SDValue CC = Op.getOperand(2);
14100 MVT VT = Op.getSimpleValueType();
14103 assert(Op0.getValueType().getVectorElementType().getSizeInBits() >= 8 &&
14104 Op.getValueType().getScalarType() == MVT::i1 &&
14105 "Cannot set masked compare for this operation");
14107 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
14109 bool Unsigned = false;
14112 switch (SetCCOpcode) {
14113 default: llvm_unreachable("Unexpected SETCC condition");
14114 case ISD::SETNE: SSECC = 4; break;
14115 case ISD::SETEQ: Opc = X86ISD::PCMPEQM; break;
14116 case ISD::SETUGT: SSECC = 6; Unsigned = true; break;
14117 case ISD::SETLT: Swap = true; //fall-through
14118 case ISD::SETGT: Opc = X86ISD::PCMPGTM; break;
14119 case ISD::SETULT: SSECC = 1; Unsigned = true; break;
14120 case ISD::SETUGE: SSECC = 5; Unsigned = true; break; //NLT
14121 case ISD::SETGE: Swap = true; SSECC = 2; break; // LE + swap
14122 case ISD::SETULE: Unsigned = true; //fall-through
14123 case ISD::SETLE: SSECC = 2; break;
14127 std::swap(Op0, Op1);
14129 return DAG.getNode(Opc, dl, VT, Op0, Op1);
14130 Opc = Unsigned ? X86ISD::CMPMU: X86ISD::CMPM;
14131 return DAG.getNode(Opc, dl, VT, Op0, Op1,
14132 DAG.getConstant(SSECC, dl, MVT::i8));
14135 /// \brief Try to turn a VSETULT into a VSETULE by modifying its second
14136 /// operand \p Op1. If non-trivial (for example because it's not constant)
14137 /// return an empty value.
14138 static SDValue ChangeVSETULTtoVSETULE(SDLoc dl, SDValue Op1, SelectionDAG &DAG)
14140 BuildVectorSDNode *BV = dyn_cast<BuildVectorSDNode>(Op1.getNode());
14144 MVT VT = Op1.getSimpleValueType();
14145 MVT EVT = VT.getVectorElementType();
14146 unsigned n = VT.getVectorNumElements();
14147 SmallVector<SDValue, 8> ULTOp1;
14149 for (unsigned i = 0; i < n; ++i) {
14150 ConstantSDNode *Elt = dyn_cast<ConstantSDNode>(BV->getOperand(i));
14151 if (!Elt || Elt->isOpaque() || Elt->getValueType(0) != EVT)
14154 // Avoid underflow.
14155 APInt Val = Elt->getAPIntValue();
14159 ULTOp1.push_back(DAG.getConstant(Val - 1, dl, EVT));
14162 return DAG.getNode(ISD::BUILD_VECTOR, dl, VT, ULTOp1);
14165 static SDValue LowerVSETCC(SDValue Op, const X86Subtarget *Subtarget,
14166 SelectionDAG &DAG) {
14167 SDValue Op0 = Op.getOperand(0);
14168 SDValue Op1 = Op.getOperand(1);
14169 SDValue CC = Op.getOperand(2);
14170 MVT VT = Op.getSimpleValueType();
14171 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
14172 bool isFP = Op.getOperand(1).getSimpleValueType().isFloatingPoint();
14177 MVT EltVT = Op0.getSimpleValueType().getVectorElementType();
14178 assert(EltVT == MVT::f32 || EltVT == MVT::f64);
14181 unsigned SSECC = translateX86FSETCC(SetCCOpcode, Op0, Op1);
14182 unsigned Opc = X86ISD::CMPP;
14183 if (Subtarget->hasAVX512() && VT.getVectorElementType() == MVT::i1) {
14184 assert(VT.getVectorNumElements() <= 16);
14185 Opc = X86ISD::CMPM;
14187 // In the two special cases we can't handle, emit two comparisons.
14190 unsigned CombineOpc;
14191 if (SetCCOpcode == ISD::SETUEQ) {
14192 CC0 = 3; CC1 = 0; CombineOpc = ISD::OR;
14194 assert(SetCCOpcode == ISD::SETONE);
14195 CC0 = 7; CC1 = 4; CombineOpc = ISD::AND;
14198 SDValue Cmp0 = DAG.getNode(Opc, dl, VT, Op0, Op1,
14199 DAG.getConstant(CC0, dl, MVT::i8));
14200 SDValue Cmp1 = DAG.getNode(Opc, dl, VT, Op0, Op1,
14201 DAG.getConstant(CC1, dl, MVT::i8));
14202 return DAG.getNode(CombineOpc, dl, VT, Cmp0, Cmp1);
14204 // Handle all other FP comparisons here.
14205 return DAG.getNode(Opc, dl, VT, Op0, Op1,
14206 DAG.getConstant(SSECC, dl, MVT::i8));
14209 MVT VTOp0 = Op0.getSimpleValueType();
14210 assert(VTOp0 == Op1.getSimpleValueType() &&
14211 "Expected operands with same type!");
14212 assert(VT.getVectorNumElements() == VTOp0.getVectorNumElements() &&
14213 "Invalid number of packed elements for source and destination!");
14215 if (VT.is128BitVector() && VTOp0.is256BitVector()) {
14216 // On non-AVX512 targets, a vector of MVT::i1 is promoted by the type
14217 // legalizer to a wider vector type. In the case of 'vsetcc' nodes, the
14218 // legalizer firstly checks if the first operand in input to the setcc has
14219 // a legal type. If so, then it promotes the return type to that same type.
14220 // Otherwise, the return type is promoted to the 'next legal type' which,
14221 // for a vector of MVT::i1 is always a 128-bit integer vector type.
14223 // We reach this code only if the following two conditions are met:
14224 // 1. Both return type and operand type have been promoted to wider types
14225 // by the type legalizer.
14226 // 2. The original operand type has been promoted to a 256-bit vector.
14228 // Note that condition 2. only applies for AVX targets.
14229 SDValue NewOp = DAG.getSetCC(dl, VTOp0, Op0, Op1, SetCCOpcode);
14230 return DAG.getZExtOrTrunc(NewOp, dl, VT);
14233 // The non-AVX512 code below works under the assumption that source and
14234 // destination types are the same.
14235 assert((Subtarget->hasAVX512() || (VT == VTOp0)) &&
14236 "Value types for source and destination must be the same!");
14238 // Break 256-bit integer vector compare into smaller ones.
14239 if (VT.is256BitVector() && !Subtarget->hasInt256())
14240 return Lower256IntVSETCC(Op, DAG);
14242 EVT OpVT = Op1.getValueType();
14243 if (OpVT.getVectorElementType() == MVT::i1)
14244 return LowerBoolVSETCC_AVX512(Op, DAG);
14246 bool MaskResult = (VT.getVectorElementType() == MVT::i1);
14247 if (Subtarget->hasAVX512()) {
14248 if (Op1.getValueType().is512BitVector() ||
14249 (Subtarget->hasBWI() && Subtarget->hasVLX()) ||
14250 (MaskResult && OpVT.getVectorElementType().getSizeInBits() >= 32))
14251 return LowerIntVSETCC_AVX512(Op, DAG, Subtarget);
14253 // In AVX-512 architecture setcc returns mask with i1 elements,
14254 // But there is no compare instruction for i8 and i16 elements in KNL.
14255 // We are not talking about 512-bit operands in this case, these
14256 // types are illegal.
14258 (OpVT.getVectorElementType().getSizeInBits() < 32 &&
14259 OpVT.getVectorElementType().getSizeInBits() >= 8))
14260 return DAG.getNode(ISD::TRUNCATE, dl, VT,
14261 DAG.getNode(ISD::SETCC, dl, OpVT, Op0, Op1, CC));
14264 // Lower using XOP integer comparisons.
14265 if ((VT == MVT::v16i8 || VT == MVT::v8i16 ||
14266 VT == MVT::v4i32 || VT == MVT::v2i64) && Subtarget->hasXOP()) {
14267 // Translate compare code to XOP PCOM compare mode.
14268 unsigned CmpMode = 0;
14269 switch (SetCCOpcode) {
14270 default: llvm_unreachable("Unexpected SETCC condition");
14272 case ISD::SETLT: CmpMode = 0x00; break;
14274 case ISD::SETLE: CmpMode = 0x01; break;
14276 case ISD::SETGT: CmpMode = 0x02; break;
14278 case ISD::SETGE: CmpMode = 0x03; break;
14279 case ISD::SETEQ: CmpMode = 0x04; break;
14280 case ISD::SETNE: CmpMode = 0x05; break;
14283 // Are we comparing unsigned or signed integers?
14284 unsigned Opc = ISD::isUnsignedIntSetCC(SetCCOpcode)
14285 ? X86ISD::VPCOMU : X86ISD::VPCOM;
14287 return DAG.getNode(Opc, dl, VT, Op0, Op1,
14288 DAG.getConstant(CmpMode, dl, MVT::i8));
14291 // We are handling one of the integer comparisons here. Since SSE only has
14292 // GT and EQ comparisons for integer, swapping operands and multiple
14293 // operations may be required for some comparisons.
14295 bool Swap = false, Invert = false, FlipSigns = false, MinMax = false;
14296 bool Subus = false;
14298 switch (SetCCOpcode) {
14299 default: llvm_unreachable("Unexpected SETCC condition");
14300 case ISD::SETNE: Invert = true;
14301 case ISD::SETEQ: Opc = X86ISD::PCMPEQ; break;
14302 case ISD::SETLT: Swap = true;
14303 case ISD::SETGT: Opc = X86ISD::PCMPGT; break;
14304 case ISD::SETGE: Swap = true;
14305 case ISD::SETLE: Opc = X86ISD::PCMPGT;
14306 Invert = true; break;
14307 case ISD::SETULT: Swap = true;
14308 case ISD::SETUGT: Opc = X86ISD::PCMPGT;
14309 FlipSigns = true; break;
14310 case ISD::SETUGE: Swap = true;
14311 case ISD::SETULE: Opc = X86ISD::PCMPGT;
14312 FlipSigns = true; Invert = true; break;
14315 // Special case: Use min/max operations for SETULE/SETUGE
14316 MVT VET = VT.getVectorElementType();
14318 (Subtarget->hasSSE41() && (VET >= MVT::i8 && VET <= MVT::i32))
14319 || (Subtarget->hasSSE2() && (VET == MVT::i8));
14322 switch (SetCCOpcode) {
14324 case ISD::SETULE: Opc = ISD::UMIN; MinMax = true; break;
14325 case ISD::SETUGE: Opc = ISD::UMAX; MinMax = true; break;
14328 if (MinMax) { Swap = false; Invert = false; FlipSigns = false; }
14331 bool hasSubus = Subtarget->hasSSE2() && (VET == MVT::i8 || VET == MVT::i16);
14332 if (!MinMax && hasSubus) {
14333 // As another special case, use PSUBUS[BW] when it's profitable. E.g. for
14335 // t = psubus Op0, Op1
14336 // pcmpeq t, <0..0>
14337 switch (SetCCOpcode) {
14339 case ISD::SETULT: {
14340 // If the comparison is against a constant we can turn this into a
14341 // setule. With psubus, setule does not require a swap. This is
14342 // beneficial because the constant in the register is no longer
14343 // destructed as the destination so it can be hoisted out of a loop.
14344 // Only do this pre-AVX since vpcmp* is no longer destructive.
14345 if (Subtarget->hasAVX())
14347 SDValue ULEOp1 = ChangeVSETULTtoVSETULE(dl, Op1, DAG);
14348 if (ULEOp1.getNode()) {
14350 Subus = true; Invert = false; Swap = false;
14354 // Psubus is better than flip-sign because it requires no inversion.
14355 case ISD::SETUGE: Subus = true; Invert = false; Swap = true; break;
14356 case ISD::SETULE: Subus = true; Invert = false; Swap = false; break;
14360 Opc = X86ISD::SUBUS;
14366 std::swap(Op0, Op1);
14368 // Check that the operation in question is available (most are plain SSE2,
14369 // but PCMPGTQ and PCMPEQQ have different requirements).
14370 if (VT == MVT::v2i64) {
14371 if (Opc == X86ISD::PCMPGT && !Subtarget->hasSSE42()) {
14372 assert(Subtarget->hasSSE2() && "Don't know how to lower!");
14374 // First cast everything to the right type.
14375 Op0 = DAG.getBitcast(MVT::v4i32, Op0);
14376 Op1 = DAG.getBitcast(MVT::v4i32, Op1);
14378 // Since SSE has no unsigned integer comparisons, we need to flip the sign
14379 // bits of the inputs before performing those operations. The lower
14380 // compare is always unsigned.
14383 SB = DAG.getConstant(0x80000000U, dl, MVT::v4i32);
14385 SDValue Sign = DAG.getConstant(0x80000000U, dl, MVT::i32);
14386 SDValue Zero = DAG.getConstant(0x00000000U, dl, MVT::i32);
14387 SB = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32,
14388 Sign, Zero, Sign, Zero);
14390 Op0 = DAG.getNode(ISD::XOR, dl, MVT::v4i32, Op0, SB);
14391 Op1 = DAG.getNode(ISD::XOR, dl, MVT::v4i32, Op1, SB);
14393 // Emulate PCMPGTQ with (hi1 > hi2) | ((hi1 == hi2) & (lo1 > lo2))
14394 SDValue GT = DAG.getNode(X86ISD::PCMPGT, dl, MVT::v4i32, Op0, Op1);
14395 SDValue EQ = DAG.getNode(X86ISD::PCMPEQ, dl, MVT::v4i32, Op0, Op1);
14397 // Create masks for only the low parts/high parts of the 64 bit integers.
14398 static const int MaskHi[] = { 1, 1, 3, 3 };
14399 static const int MaskLo[] = { 0, 0, 2, 2 };
14400 SDValue EQHi = DAG.getVectorShuffle(MVT::v4i32, dl, EQ, EQ, MaskHi);
14401 SDValue GTLo = DAG.getVectorShuffle(MVT::v4i32, dl, GT, GT, MaskLo);
14402 SDValue GTHi = DAG.getVectorShuffle(MVT::v4i32, dl, GT, GT, MaskHi);
14404 SDValue Result = DAG.getNode(ISD::AND, dl, MVT::v4i32, EQHi, GTLo);
14405 Result = DAG.getNode(ISD::OR, dl, MVT::v4i32, Result, GTHi);
14408 Result = DAG.getNOT(dl, Result, MVT::v4i32);
14410 return DAG.getBitcast(VT, Result);
14413 if (Opc == X86ISD::PCMPEQ && !Subtarget->hasSSE41()) {
14414 // If pcmpeqq is missing but pcmpeqd is available synthesize pcmpeqq with
14415 // pcmpeqd + pshufd + pand.
14416 assert(Subtarget->hasSSE2() && !FlipSigns && "Don't know how to lower!");
14418 // First cast everything to the right type.
14419 Op0 = DAG.getBitcast(MVT::v4i32, Op0);
14420 Op1 = DAG.getBitcast(MVT::v4i32, Op1);
14423 SDValue Result = DAG.getNode(Opc, dl, MVT::v4i32, Op0, Op1);
14425 // Make sure the lower and upper halves are both all-ones.
14426 static const int Mask[] = { 1, 0, 3, 2 };
14427 SDValue Shuf = DAG.getVectorShuffle(MVT::v4i32, dl, Result, Result, Mask);
14428 Result = DAG.getNode(ISD::AND, dl, MVT::v4i32, Result, Shuf);
14431 Result = DAG.getNOT(dl, Result, MVT::v4i32);
14433 return DAG.getBitcast(VT, Result);
14437 // Since SSE has no unsigned integer comparisons, we need to flip the sign
14438 // bits of the inputs before performing those operations.
14440 EVT EltVT = VT.getVectorElementType();
14441 SDValue SB = DAG.getConstant(APInt::getSignBit(EltVT.getSizeInBits()), dl,
14443 Op0 = DAG.getNode(ISD::XOR, dl, VT, Op0, SB);
14444 Op1 = DAG.getNode(ISD::XOR, dl, VT, Op1, SB);
14447 SDValue Result = DAG.getNode(Opc, dl, VT, Op0, Op1);
14449 // If the logical-not of the result is required, perform that now.
14451 Result = DAG.getNOT(dl, Result, VT);
14454 Result = DAG.getNode(X86ISD::PCMPEQ, dl, VT, Op0, Result);
14457 Result = DAG.getNode(X86ISD::PCMPEQ, dl, VT, Result,
14458 getZeroVector(VT, Subtarget, DAG, dl));
14463 SDValue X86TargetLowering::LowerSETCC(SDValue Op, SelectionDAG &DAG) const {
14465 MVT VT = Op.getSimpleValueType();
14467 if (VT.isVector()) return LowerVSETCC(Op, Subtarget, DAG);
14469 assert(((!Subtarget->hasAVX512() && VT == MVT::i8) || (VT == MVT::i1))
14470 && "SetCC type must be 8-bit or 1-bit integer");
14471 SDValue Op0 = Op.getOperand(0);
14472 SDValue Op1 = Op.getOperand(1);
14474 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
14476 // Optimize to BT if possible.
14477 // Lower (X & (1 << N)) == 0 to BT(X, N).
14478 // Lower ((X >>u N) & 1) != 0 to BT(X, N).
14479 // Lower ((X >>s N) & 1) != 0 to BT(X, N).
14480 if (Op0.getOpcode() == ISD::AND && Op0.hasOneUse() &&
14481 Op1.getOpcode() == ISD::Constant &&
14482 cast<ConstantSDNode>(Op1)->isNullValue() &&
14483 (CC == ISD::SETEQ || CC == ISD::SETNE)) {
14484 SDValue NewSetCC = LowerToBT(Op0, CC, dl, DAG);
14485 if (NewSetCC.getNode()) {
14487 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i1, NewSetCC);
14492 // Look for X == 0, X == 1, X != 0, or X != 1. We can simplify some forms of
14494 if (Op1.getOpcode() == ISD::Constant &&
14495 (cast<ConstantSDNode>(Op1)->getZExtValue() == 1 ||
14496 cast<ConstantSDNode>(Op1)->isNullValue()) &&
14497 (CC == ISD::SETEQ || CC == ISD::SETNE)) {
14499 // If the input is a setcc, then reuse the input setcc or use a new one with
14500 // the inverted condition.
14501 if (Op0.getOpcode() == X86ISD::SETCC) {
14502 X86::CondCode CCode = (X86::CondCode)Op0.getConstantOperandVal(0);
14503 bool Invert = (CC == ISD::SETNE) ^
14504 cast<ConstantSDNode>(Op1)->isNullValue();
14508 CCode = X86::GetOppositeBranchCondition(CCode);
14509 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
14510 DAG.getConstant(CCode, dl, MVT::i8),
14511 Op0.getOperand(1));
14513 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i1, SetCC);
14517 if ((Op0.getValueType() == MVT::i1) && (Op1.getOpcode() == ISD::Constant) &&
14518 (cast<ConstantSDNode>(Op1)->getZExtValue() == 1) &&
14519 (CC == ISD::SETEQ || CC == ISD::SETNE)) {
14521 ISD::CondCode NewCC = ISD::getSetCCInverse(CC, true);
14522 return DAG.getSetCC(dl, VT, Op0, DAG.getConstant(0, dl, MVT::i1), NewCC);
14525 bool isFP = Op1.getSimpleValueType().isFloatingPoint();
14526 unsigned X86CC = TranslateX86CC(CC, dl, isFP, Op0, Op1, DAG);
14527 if (X86CC == X86::COND_INVALID)
14530 SDValue EFLAGS = EmitCmp(Op0, Op1, X86CC, dl, DAG);
14531 EFLAGS = ConvertCmpIfNecessary(EFLAGS, DAG);
14532 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
14533 DAG.getConstant(X86CC, dl, MVT::i8), EFLAGS);
14535 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i1, SetCC);
14539 // isX86LogicalCmp - Return true if opcode is a X86 logical comparison.
14540 static bool isX86LogicalCmp(SDValue Op) {
14541 unsigned Opc = Op.getNode()->getOpcode();
14542 if (Opc == X86ISD::CMP || Opc == X86ISD::COMI || Opc == X86ISD::UCOMI ||
14543 Opc == X86ISD::SAHF)
14545 if (Op.getResNo() == 1 &&
14546 (Opc == X86ISD::ADD ||
14547 Opc == X86ISD::SUB ||
14548 Opc == X86ISD::ADC ||
14549 Opc == X86ISD::SBB ||
14550 Opc == X86ISD::SMUL ||
14551 Opc == X86ISD::UMUL ||
14552 Opc == X86ISD::INC ||
14553 Opc == X86ISD::DEC ||
14554 Opc == X86ISD::OR ||
14555 Opc == X86ISD::XOR ||
14556 Opc == X86ISD::AND))
14559 if (Op.getResNo() == 2 && Opc == X86ISD::UMUL)
14565 static bool isTruncWithZeroHighBitsInput(SDValue V, SelectionDAG &DAG) {
14566 if (V.getOpcode() != ISD::TRUNCATE)
14569 SDValue VOp0 = V.getOperand(0);
14570 unsigned InBits = VOp0.getValueSizeInBits();
14571 unsigned Bits = V.getValueSizeInBits();
14572 return DAG.MaskedValueIsZero(VOp0, APInt::getHighBitsSet(InBits,InBits-Bits));
14575 SDValue X86TargetLowering::LowerSELECT(SDValue Op, SelectionDAG &DAG) const {
14576 bool addTest = true;
14577 SDValue Cond = Op.getOperand(0);
14578 SDValue Op1 = Op.getOperand(1);
14579 SDValue Op2 = Op.getOperand(2);
14581 EVT VT = Op1.getValueType();
14584 // Lower FP selects into a CMP/AND/ANDN/OR sequence when the necessary SSE ops
14585 // are available or VBLENDV if AVX is available.
14586 // Otherwise FP cmovs get lowered into a less efficient branch sequence later.
14587 if (Cond.getOpcode() == ISD::SETCC &&
14588 ((Subtarget->hasSSE2() && (VT == MVT::f32 || VT == MVT::f64)) ||
14589 (Subtarget->hasSSE1() && VT == MVT::f32)) &&
14590 VT == Cond.getOperand(0).getValueType() && Cond->hasOneUse()) {
14591 SDValue CondOp0 = Cond.getOperand(0), CondOp1 = Cond.getOperand(1);
14592 int SSECC = translateX86FSETCC(
14593 cast<CondCodeSDNode>(Cond.getOperand(2))->get(), CondOp0, CondOp1);
14596 if (Subtarget->hasAVX512()) {
14597 SDValue Cmp = DAG.getNode(X86ISD::FSETCC, DL, MVT::i1, CondOp0, CondOp1,
14598 DAG.getConstant(SSECC, DL, MVT::i8));
14599 return DAG.getNode(X86ISD::SELECT, DL, VT, Cmp, Op1, Op2);
14602 SDValue Cmp = DAG.getNode(X86ISD::FSETCC, DL, VT, CondOp0, CondOp1,
14603 DAG.getConstant(SSECC, DL, MVT::i8));
14605 // If we have AVX, we can use a variable vector select (VBLENDV) instead
14606 // of 3 logic instructions for size savings and potentially speed.
14607 // Unfortunately, there is no scalar form of VBLENDV.
14609 // If either operand is a constant, don't try this. We can expect to
14610 // optimize away at least one of the logic instructions later in that
14611 // case, so that sequence would be faster than a variable blend.
14613 // BLENDV was introduced with SSE 4.1, but the 2 register form implicitly
14614 // uses XMM0 as the selection register. That may need just as many
14615 // instructions as the AND/ANDN/OR sequence due to register moves, so
14618 if (Subtarget->hasAVX() &&
14619 !isa<ConstantFPSDNode>(Op1) && !isa<ConstantFPSDNode>(Op2)) {
14621 // Convert to vectors, do a VSELECT, and convert back to scalar.
14622 // All of the conversions should be optimized away.
14624 EVT VecVT = VT == MVT::f32 ? MVT::v4f32 : MVT::v2f64;
14625 SDValue VOp1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, VecVT, Op1);
14626 SDValue VOp2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, VecVT, Op2);
14627 SDValue VCmp = DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, VecVT, Cmp);
14629 EVT VCmpVT = VT == MVT::f32 ? MVT::v4i32 : MVT::v2i64;
14630 VCmp = DAG.getBitcast(VCmpVT, VCmp);
14632 SDValue VSel = DAG.getNode(ISD::VSELECT, DL, VecVT, VCmp, VOp1, VOp2);
14634 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, VT,
14635 VSel, DAG.getIntPtrConstant(0, DL));
14637 SDValue AndN = DAG.getNode(X86ISD::FANDN, DL, VT, Cmp, Op2);
14638 SDValue And = DAG.getNode(X86ISD::FAND, DL, VT, Cmp, Op1);
14639 return DAG.getNode(X86ISD::FOR, DL, VT, AndN, And);
14643 if (VT.isVector() && VT.getScalarType() == MVT::i1) {
14645 if (ISD::isBuildVectorOfConstantSDNodes(Op1.getNode()))
14646 Op1Scalar = ConvertI1VectorToInteger(Op1, DAG);
14647 else if (Op1.getOpcode() == ISD::BITCAST && Op1.getOperand(0))
14648 Op1Scalar = Op1.getOperand(0);
14650 if (ISD::isBuildVectorOfConstantSDNodes(Op2.getNode()))
14651 Op2Scalar = ConvertI1VectorToInteger(Op2, DAG);
14652 else if (Op2.getOpcode() == ISD::BITCAST && Op2.getOperand(0))
14653 Op2Scalar = Op2.getOperand(0);
14654 if (Op1Scalar.getNode() && Op2Scalar.getNode()) {
14655 SDValue newSelect = DAG.getNode(ISD::SELECT, DL,
14656 Op1Scalar.getValueType(),
14657 Cond, Op1Scalar, Op2Scalar);
14658 if (newSelect.getValueSizeInBits() == VT.getSizeInBits())
14659 return DAG.getBitcast(VT, newSelect);
14660 SDValue ExtVec = DAG.getBitcast(MVT::v8i1, newSelect);
14661 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, VT, ExtVec,
14662 DAG.getIntPtrConstant(0, DL));
14666 if (VT == MVT::v4i1 || VT == MVT::v2i1) {
14667 SDValue zeroConst = DAG.getIntPtrConstant(0, DL);
14668 Op1 = DAG.getNode(ISD::INSERT_SUBVECTOR, DL, MVT::v8i1,
14669 DAG.getUNDEF(MVT::v8i1), Op1, zeroConst);
14670 Op2 = DAG.getNode(ISD::INSERT_SUBVECTOR, DL, MVT::v8i1,
14671 DAG.getUNDEF(MVT::v8i1), Op2, zeroConst);
14672 SDValue newSelect = DAG.getNode(ISD::SELECT, DL, MVT::v8i1,
14674 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, VT, newSelect, zeroConst);
14677 if (Cond.getOpcode() == ISD::SETCC) {
14678 SDValue NewCond = LowerSETCC(Cond, DAG);
14679 if (NewCond.getNode())
14683 // (select (x == 0), -1, y) -> (sign_bit (x - 1)) | y
14684 // (select (x == 0), y, -1) -> ~(sign_bit (x - 1)) | y
14685 // (select (x != 0), y, -1) -> (sign_bit (x - 1)) | y
14686 // (select (x != 0), -1, y) -> ~(sign_bit (x - 1)) | y
14687 if (Cond.getOpcode() == X86ISD::SETCC &&
14688 Cond.getOperand(1).getOpcode() == X86ISD::CMP &&
14689 isZero(Cond.getOperand(1).getOperand(1))) {
14690 SDValue Cmp = Cond.getOperand(1);
14692 unsigned CondCode =cast<ConstantSDNode>(Cond.getOperand(0))->getZExtValue();
14694 if ((isAllOnes(Op1) || isAllOnes(Op2)) &&
14695 (CondCode == X86::COND_E || CondCode == X86::COND_NE)) {
14696 SDValue Y = isAllOnes(Op2) ? Op1 : Op2;
14698 SDValue CmpOp0 = Cmp.getOperand(0);
14699 // Apply further optimizations for special cases
14700 // (select (x != 0), -1, 0) -> neg & sbb
14701 // (select (x == 0), 0, -1) -> neg & sbb
14702 if (ConstantSDNode *YC = dyn_cast<ConstantSDNode>(Y))
14703 if (YC->isNullValue() &&
14704 (isAllOnes(Op1) == (CondCode == X86::COND_NE))) {
14705 SDVTList VTs = DAG.getVTList(CmpOp0.getValueType(), MVT::i32);
14706 SDValue Neg = DAG.getNode(X86ISD::SUB, DL, VTs,
14707 DAG.getConstant(0, DL,
14708 CmpOp0.getValueType()),
14710 SDValue Res = DAG.getNode(X86ISD::SETCC_CARRY, DL, Op.getValueType(),
14711 DAG.getConstant(X86::COND_B, DL, MVT::i8),
14712 SDValue(Neg.getNode(), 1));
14716 Cmp = DAG.getNode(X86ISD::CMP, DL, MVT::i32,
14717 CmpOp0, DAG.getConstant(1, DL, CmpOp0.getValueType()));
14718 Cmp = ConvertCmpIfNecessary(Cmp, DAG);
14720 SDValue Res = // Res = 0 or -1.
14721 DAG.getNode(X86ISD::SETCC_CARRY, DL, Op.getValueType(),
14722 DAG.getConstant(X86::COND_B, DL, MVT::i8), Cmp);
14724 if (isAllOnes(Op1) != (CondCode == X86::COND_E))
14725 Res = DAG.getNOT(DL, Res, Res.getValueType());
14727 ConstantSDNode *N2C = dyn_cast<ConstantSDNode>(Op2);
14728 if (!N2C || !N2C->isNullValue())
14729 Res = DAG.getNode(ISD::OR, DL, Res.getValueType(), Res, Y);
14734 // Look past (and (setcc_carry (cmp ...)), 1).
14735 if (Cond.getOpcode() == ISD::AND &&
14736 Cond.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY) {
14737 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Cond.getOperand(1));
14738 if (C && C->getAPIntValue() == 1)
14739 Cond = Cond.getOperand(0);
14742 // If condition flag is set by a X86ISD::CMP, then use it as the condition
14743 // setting operand in place of the X86ISD::SETCC.
14744 unsigned CondOpcode = Cond.getOpcode();
14745 if (CondOpcode == X86ISD::SETCC ||
14746 CondOpcode == X86ISD::SETCC_CARRY) {
14747 CC = Cond.getOperand(0);
14749 SDValue Cmp = Cond.getOperand(1);
14750 unsigned Opc = Cmp.getOpcode();
14751 MVT VT = Op.getSimpleValueType();
14753 bool IllegalFPCMov = false;
14754 if (VT.isFloatingPoint() && !VT.isVector() &&
14755 !isScalarFPTypeInSSEReg(VT)) // FPStack?
14756 IllegalFPCMov = !hasFPCMov(cast<ConstantSDNode>(CC)->getSExtValue());
14758 if ((isX86LogicalCmp(Cmp) && !IllegalFPCMov) ||
14759 Opc == X86ISD::BT) { // FIXME
14763 } else if (CondOpcode == ISD::USUBO || CondOpcode == ISD::SSUBO ||
14764 CondOpcode == ISD::UADDO || CondOpcode == ISD::SADDO ||
14765 ((CondOpcode == ISD::UMULO || CondOpcode == ISD::SMULO) &&
14766 Cond.getOperand(0).getValueType() != MVT::i8)) {
14767 SDValue LHS = Cond.getOperand(0);
14768 SDValue RHS = Cond.getOperand(1);
14769 unsigned X86Opcode;
14772 switch (CondOpcode) {
14773 case ISD::UADDO: X86Opcode = X86ISD::ADD; X86Cond = X86::COND_B; break;
14774 case ISD::SADDO: X86Opcode = X86ISD::ADD; X86Cond = X86::COND_O; break;
14775 case ISD::USUBO: X86Opcode = X86ISD::SUB; X86Cond = X86::COND_B; break;
14776 case ISD::SSUBO: X86Opcode = X86ISD::SUB; X86Cond = X86::COND_O; break;
14777 case ISD::UMULO: X86Opcode = X86ISD::UMUL; X86Cond = X86::COND_O; break;
14778 case ISD::SMULO: X86Opcode = X86ISD::SMUL; X86Cond = X86::COND_O; break;
14779 default: llvm_unreachable("unexpected overflowing operator");
14781 if (CondOpcode == ISD::UMULO)
14782 VTs = DAG.getVTList(LHS.getValueType(), LHS.getValueType(),
14785 VTs = DAG.getVTList(LHS.getValueType(), MVT::i32);
14787 SDValue X86Op = DAG.getNode(X86Opcode, DL, VTs, LHS, RHS);
14789 if (CondOpcode == ISD::UMULO)
14790 Cond = X86Op.getValue(2);
14792 Cond = X86Op.getValue(1);
14794 CC = DAG.getConstant(X86Cond, DL, MVT::i8);
14799 // Look past the truncate if the high bits are known zero.
14800 if (isTruncWithZeroHighBitsInput(Cond, DAG))
14801 Cond = Cond.getOperand(0);
14803 // We know the result of AND is compared against zero. Try to match
14805 if (Cond.getOpcode() == ISD::AND && Cond.hasOneUse()) {
14806 SDValue NewSetCC = LowerToBT(Cond, ISD::SETNE, DL, DAG);
14807 if (NewSetCC.getNode()) {
14808 CC = NewSetCC.getOperand(0);
14809 Cond = NewSetCC.getOperand(1);
14816 CC = DAG.getConstant(X86::COND_NE, DL, MVT::i8);
14817 Cond = EmitTest(Cond, X86::COND_NE, DL, DAG);
14820 // a < b ? -1 : 0 -> RES = ~setcc_carry
14821 // a < b ? 0 : -1 -> RES = setcc_carry
14822 // a >= b ? -1 : 0 -> RES = setcc_carry
14823 // a >= b ? 0 : -1 -> RES = ~setcc_carry
14824 if (Cond.getOpcode() == X86ISD::SUB) {
14825 Cond = ConvertCmpIfNecessary(Cond, DAG);
14826 unsigned CondCode = cast<ConstantSDNode>(CC)->getZExtValue();
14828 if ((CondCode == X86::COND_AE || CondCode == X86::COND_B) &&
14829 (isAllOnes(Op1) || isAllOnes(Op2)) && (isZero(Op1) || isZero(Op2))) {
14830 SDValue Res = DAG.getNode(X86ISD::SETCC_CARRY, DL, Op.getValueType(),
14831 DAG.getConstant(X86::COND_B, DL, MVT::i8),
14833 if (isAllOnes(Op1) != (CondCode == X86::COND_B))
14834 return DAG.getNOT(DL, Res, Res.getValueType());
14839 // X86 doesn't have an i8 cmov. If both operands are the result of a truncate
14840 // widen the cmov and push the truncate through. This avoids introducing a new
14841 // branch during isel and doesn't add any extensions.
14842 if (Op.getValueType() == MVT::i8 &&
14843 Op1.getOpcode() == ISD::TRUNCATE && Op2.getOpcode() == ISD::TRUNCATE) {
14844 SDValue T1 = Op1.getOperand(0), T2 = Op2.getOperand(0);
14845 if (T1.getValueType() == T2.getValueType() &&
14846 // Blacklist CopyFromReg to avoid partial register stalls.
14847 T1.getOpcode() != ISD::CopyFromReg && T2.getOpcode()!=ISD::CopyFromReg){
14848 SDVTList VTs = DAG.getVTList(T1.getValueType(), MVT::Glue);
14849 SDValue Cmov = DAG.getNode(X86ISD::CMOV, DL, VTs, T2, T1, CC, Cond);
14850 return DAG.getNode(ISD::TRUNCATE, DL, Op.getValueType(), Cmov);
14854 // X86ISD::CMOV means set the result (which is operand 1) to the RHS if
14855 // condition is true.
14856 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::Glue);
14857 SDValue Ops[] = { Op2, Op1, CC, Cond };
14858 return DAG.getNode(X86ISD::CMOV, DL, VTs, Ops);
14861 static SDValue LowerSIGN_EXTEND_AVX512(SDValue Op,
14862 const X86Subtarget *Subtarget,
14863 SelectionDAG &DAG) {
14864 MVT VT = Op->getSimpleValueType(0);
14865 SDValue In = Op->getOperand(0);
14866 MVT InVT = In.getSimpleValueType();
14867 MVT VTElt = VT.getVectorElementType();
14868 MVT InVTElt = InVT.getVectorElementType();
14872 if ((InVTElt == MVT::i1) &&
14873 (((Subtarget->hasBWI() && Subtarget->hasVLX() &&
14874 VT.getSizeInBits() <= 256 && VTElt.getSizeInBits() <= 16)) ||
14876 ((Subtarget->hasBWI() && VT.is512BitVector() &&
14877 VTElt.getSizeInBits() <= 16)) ||
14879 ((Subtarget->hasDQI() && Subtarget->hasVLX() &&
14880 VT.getSizeInBits() <= 256 && VTElt.getSizeInBits() >= 32)) ||
14882 ((Subtarget->hasDQI() && VT.is512BitVector() &&
14883 VTElt.getSizeInBits() >= 32))))
14884 return DAG.getNode(X86ISD::VSEXT, dl, VT, In);
14886 unsigned int NumElts = VT.getVectorNumElements();
14888 if (NumElts != 8 && NumElts != 16 && !Subtarget->hasBWI())
14891 if (VT.is512BitVector() && InVT.getVectorElementType() != MVT::i1) {
14892 if (In.getOpcode() == X86ISD::VSEXT || In.getOpcode() == X86ISD::VZEXT)
14893 return DAG.getNode(In.getOpcode(), dl, VT, In.getOperand(0));
14894 return DAG.getNode(X86ISD::VSEXT, dl, VT, In);
14897 assert (InVT.getVectorElementType() == MVT::i1 && "Unexpected vector type");
14898 MVT ExtVT = NumElts == 8 ? MVT::v8i64 : MVT::v16i32;
14900 DAG.getConstant(APInt::getAllOnesValue(ExtVT.getScalarSizeInBits()), dl,
14903 DAG.getConstant(APInt::getNullValue(ExtVT.getScalarSizeInBits()), dl, ExtVT);
14905 SDValue V = DAG.getNode(ISD::VSELECT, dl, ExtVT, In, NegOne, Zero);
14906 if (VT.is512BitVector())
14908 return DAG.getNode(X86ISD::VTRUNC, dl, VT, V);
14911 static SDValue LowerSIGN_EXTEND_VECTOR_INREG(SDValue Op,
14912 const X86Subtarget *Subtarget,
14913 SelectionDAG &DAG) {
14914 SDValue In = Op->getOperand(0);
14915 MVT VT = Op->getSimpleValueType(0);
14916 MVT InVT = In.getSimpleValueType();
14917 assert(VT.getSizeInBits() == InVT.getSizeInBits());
14919 MVT InSVT = InVT.getScalarType();
14920 assert(VT.getScalarType().getScalarSizeInBits() > InSVT.getScalarSizeInBits());
14922 if (VT != MVT::v2i64 && VT != MVT::v4i32 && VT != MVT::v8i16)
14924 if (InSVT != MVT::i32 && InSVT != MVT::i16 && InSVT != MVT::i8)
14929 // SSE41 targets can use the pmovsx* instructions directly.
14930 if (Subtarget->hasSSE41())
14931 return DAG.getNode(X86ISD::VSEXT, dl, VT, In);
14933 // pre-SSE41 targets unpack lower lanes and then sign-extend using SRAI.
14937 // As SRAI is only available on i16/i32 types, we expand only up to i32
14938 // and handle i64 separately.
14939 while (CurrVT != VT && CurrVT.getScalarType() != MVT::i32) {
14940 Curr = DAG.getNode(X86ISD::UNPCKL, dl, CurrVT, DAG.getUNDEF(CurrVT), Curr);
14941 MVT CurrSVT = MVT::getIntegerVT(CurrVT.getScalarSizeInBits() * 2);
14942 CurrVT = MVT::getVectorVT(CurrSVT, CurrVT.getVectorNumElements() / 2);
14943 Curr = DAG.getBitcast(CurrVT, Curr);
14946 SDValue SignExt = Curr;
14947 if (CurrVT != InVT) {
14948 unsigned SignExtShift =
14949 CurrVT.getScalarSizeInBits() - InSVT.getScalarSizeInBits();
14950 SignExt = DAG.getNode(X86ISD::VSRAI, dl, CurrVT, Curr,
14951 DAG.getConstant(SignExtShift, dl, MVT::i8));
14957 if (VT == MVT::v2i64 && CurrVT == MVT::v4i32) {
14958 SDValue Sign = DAG.getNode(X86ISD::VSRAI, dl, CurrVT, Curr,
14959 DAG.getConstant(31, dl, MVT::i8));
14960 SDValue Ext = DAG.getVectorShuffle(CurrVT, dl, SignExt, Sign, {0, 4, 1, 5});
14961 return DAG.getBitcast(VT, Ext);
14967 static SDValue LowerSIGN_EXTEND(SDValue Op, const X86Subtarget *Subtarget,
14968 SelectionDAG &DAG) {
14969 MVT VT = Op->getSimpleValueType(0);
14970 SDValue In = Op->getOperand(0);
14971 MVT InVT = In.getSimpleValueType();
14974 if (VT.is512BitVector() || InVT.getVectorElementType() == MVT::i1)
14975 return LowerSIGN_EXTEND_AVX512(Op, Subtarget, DAG);
14977 if ((VT != MVT::v4i64 || InVT != MVT::v4i32) &&
14978 (VT != MVT::v8i32 || InVT != MVT::v8i16) &&
14979 (VT != MVT::v16i16 || InVT != MVT::v16i8))
14982 if (Subtarget->hasInt256())
14983 return DAG.getNode(X86ISD::VSEXT, dl, VT, In);
14985 // Optimize vectors in AVX mode
14986 // Sign extend v8i16 to v8i32 and
14989 // Divide input vector into two parts
14990 // for v4i32 the shuffle mask will be { 0, 1, -1, -1} {2, 3, -1, -1}
14991 // use vpmovsx instruction to extend v4i32 -> v2i64; v8i16 -> v4i32
14992 // concat the vectors to original VT
14994 unsigned NumElems = InVT.getVectorNumElements();
14995 SDValue Undef = DAG.getUNDEF(InVT);
14997 SmallVector<int,8> ShufMask1(NumElems, -1);
14998 for (unsigned i = 0; i != NumElems/2; ++i)
15001 SDValue OpLo = DAG.getVectorShuffle(InVT, dl, In, Undef, &ShufMask1[0]);
15003 SmallVector<int,8> ShufMask2(NumElems, -1);
15004 for (unsigned i = 0; i != NumElems/2; ++i)
15005 ShufMask2[i] = i + NumElems/2;
15007 SDValue OpHi = DAG.getVectorShuffle(InVT, dl, In, Undef, &ShufMask2[0]);
15009 MVT HalfVT = MVT::getVectorVT(VT.getScalarType(),
15010 VT.getVectorNumElements()/2);
15012 OpLo = DAG.getNode(X86ISD::VSEXT, dl, HalfVT, OpLo);
15013 OpHi = DAG.getNode(X86ISD::VSEXT, dl, HalfVT, OpHi);
15015 return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT, OpLo, OpHi);
15018 // Lower vector extended loads using a shuffle. If SSSE3 is not available we
15019 // may emit an illegal shuffle but the expansion is still better than scalar
15020 // code. We generate X86ISD::VSEXT for SEXTLOADs if it's available, otherwise
15021 // we'll emit a shuffle and a arithmetic shift.
15022 // FIXME: Is the expansion actually better than scalar code? It doesn't seem so.
15023 // TODO: It is possible to support ZExt by zeroing the undef values during
15024 // the shuffle phase or after the shuffle.
15025 static SDValue LowerExtendedLoad(SDValue Op, const X86Subtarget *Subtarget,
15026 SelectionDAG &DAG) {
15027 MVT RegVT = Op.getSimpleValueType();
15028 assert(RegVT.isVector() && "We only custom lower vector sext loads.");
15029 assert(RegVT.isInteger() &&
15030 "We only custom lower integer vector sext loads.");
15032 // Nothing useful we can do without SSE2 shuffles.
15033 assert(Subtarget->hasSSE2() && "We only custom lower sext loads with SSE2.");
15035 LoadSDNode *Ld = cast<LoadSDNode>(Op.getNode());
15037 EVT MemVT = Ld->getMemoryVT();
15038 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
15039 unsigned RegSz = RegVT.getSizeInBits();
15041 ISD::LoadExtType Ext = Ld->getExtensionType();
15043 assert((Ext == ISD::EXTLOAD || Ext == ISD::SEXTLOAD)
15044 && "Only anyext and sext are currently implemented.");
15045 assert(MemVT != RegVT && "Cannot extend to the same type");
15046 assert(MemVT.isVector() && "Must load a vector from memory");
15048 unsigned NumElems = RegVT.getVectorNumElements();
15049 unsigned MemSz = MemVT.getSizeInBits();
15050 assert(RegSz > MemSz && "Register size must be greater than the mem size");
15052 if (Ext == ISD::SEXTLOAD && RegSz == 256 && !Subtarget->hasInt256()) {
15053 // The only way in which we have a legal 256-bit vector result but not the
15054 // integer 256-bit operations needed to directly lower a sextload is if we
15055 // have AVX1 but not AVX2. In that case, we can always emit a sextload to
15056 // a 128-bit vector and a normal sign_extend to 256-bits that should get
15057 // correctly legalized. We do this late to allow the canonical form of
15058 // sextload to persist throughout the rest of the DAG combiner -- it wants
15059 // to fold together any extensions it can, and so will fuse a sign_extend
15060 // of an sextload into a sextload targeting a wider value.
15062 if (MemSz == 128) {
15063 // Just switch this to a normal load.
15064 assert(TLI.isTypeLegal(MemVT) && "If the memory type is a 128-bit type, "
15065 "it must be a legal 128-bit vector "
15067 Load = DAG.getLoad(MemVT, dl, Ld->getChain(), Ld->getBasePtr(),
15068 Ld->getPointerInfo(), Ld->isVolatile(), Ld->isNonTemporal(),
15069 Ld->isInvariant(), Ld->getAlignment());
15071 assert(MemSz < 128 &&
15072 "Can't extend a type wider than 128 bits to a 256 bit vector!");
15073 // Do an sext load to a 128-bit vector type. We want to use the same
15074 // number of elements, but elements half as wide. This will end up being
15075 // recursively lowered by this routine, but will succeed as we definitely
15076 // have all the necessary features if we're using AVX1.
15078 EVT::getIntegerVT(*DAG.getContext(), RegVT.getScalarSizeInBits() / 2);
15079 EVT HalfVecVT = EVT::getVectorVT(*DAG.getContext(), HalfEltVT, NumElems);
15081 DAG.getExtLoad(Ext, dl, HalfVecVT, Ld->getChain(), Ld->getBasePtr(),
15082 Ld->getPointerInfo(), MemVT, Ld->isVolatile(),
15083 Ld->isNonTemporal(), Ld->isInvariant(),
15084 Ld->getAlignment());
15087 // Replace chain users with the new chain.
15088 assert(Load->getNumValues() == 2 && "Loads must carry a chain!");
15089 DAG.ReplaceAllUsesOfValueWith(SDValue(Ld, 1), Load.getValue(1));
15091 // Finally, do a normal sign-extend to the desired register.
15092 return DAG.getSExtOrTrunc(Load, dl, RegVT);
15095 // All sizes must be a power of two.
15096 assert(isPowerOf2_32(RegSz * MemSz * NumElems) &&
15097 "Non-power-of-two elements are not custom lowered!");
15099 // Attempt to load the original value using scalar loads.
15100 // Find the largest scalar type that divides the total loaded size.
15101 MVT SclrLoadTy = MVT::i8;
15102 for (MVT Tp : MVT::integer_valuetypes()) {
15103 if (TLI.isTypeLegal(Tp) && ((MemSz % Tp.getSizeInBits()) == 0)) {
15108 // On 32bit systems, we can't save 64bit integers. Try bitcasting to F64.
15109 if (TLI.isTypeLegal(MVT::f64) && SclrLoadTy.getSizeInBits() < 64 &&
15111 SclrLoadTy = MVT::f64;
15113 // Calculate the number of scalar loads that we need to perform
15114 // in order to load our vector from memory.
15115 unsigned NumLoads = MemSz / SclrLoadTy.getSizeInBits();
15117 assert((Ext != ISD::SEXTLOAD || NumLoads == 1) &&
15118 "Can only lower sext loads with a single scalar load!");
15120 unsigned loadRegZize = RegSz;
15121 if (Ext == ISD::SEXTLOAD && RegSz >= 256)
15124 // Represent our vector as a sequence of elements which are the
15125 // largest scalar that we can load.
15126 EVT LoadUnitVecVT = EVT::getVectorVT(
15127 *DAG.getContext(), SclrLoadTy, loadRegZize / SclrLoadTy.getSizeInBits());
15129 // Represent the data using the same element type that is stored in
15130 // memory. In practice, we ''widen'' MemVT.
15132 EVT::getVectorVT(*DAG.getContext(), MemVT.getScalarType(),
15133 loadRegZize / MemVT.getScalarType().getSizeInBits());
15135 assert(WideVecVT.getSizeInBits() == LoadUnitVecVT.getSizeInBits() &&
15136 "Invalid vector type");
15138 // We can't shuffle using an illegal type.
15139 assert(TLI.isTypeLegal(WideVecVT) &&
15140 "We only lower types that form legal widened vector types");
15142 SmallVector<SDValue, 8> Chains;
15143 SDValue Ptr = Ld->getBasePtr();
15144 SDValue Increment = DAG.getConstant(SclrLoadTy.getSizeInBits() / 8, dl,
15145 TLI.getPointerTy(DAG.getDataLayout()));
15146 SDValue Res = DAG.getUNDEF(LoadUnitVecVT);
15148 for (unsigned i = 0; i < NumLoads; ++i) {
15149 // Perform a single load.
15150 SDValue ScalarLoad =
15151 DAG.getLoad(SclrLoadTy, dl, Ld->getChain(), Ptr, Ld->getPointerInfo(),
15152 Ld->isVolatile(), Ld->isNonTemporal(), Ld->isInvariant(),
15153 Ld->getAlignment());
15154 Chains.push_back(ScalarLoad.getValue(1));
15155 // Create the first element type using SCALAR_TO_VECTOR in order to avoid
15156 // another round of DAGCombining.
15158 Res = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, LoadUnitVecVT, ScalarLoad);
15160 Res = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, LoadUnitVecVT, Res,
15161 ScalarLoad, DAG.getIntPtrConstant(i, dl));
15163 Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr, Increment);
15166 SDValue TF = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Chains);
15168 // Bitcast the loaded value to a vector of the original element type, in
15169 // the size of the target vector type.
15170 SDValue SlicedVec = DAG.getBitcast(WideVecVT, Res);
15171 unsigned SizeRatio = RegSz / MemSz;
15173 if (Ext == ISD::SEXTLOAD) {
15174 // If we have SSE4.1, we can directly emit a VSEXT node.
15175 if (Subtarget->hasSSE41()) {
15176 SDValue Sext = DAG.getNode(X86ISD::VSEXT, dl, RegVT, SlicedVec);
15177 DAG.ReplaceAllUsesOfValueWith(SDValue(Ld, 1), TF);
15181 // Otherwise we'll use SIGN_EXTEND_VECTOR_INREG to sign extend the lowest
15183 assert(TLI.isOperationLegalOrCustom(ISD::SIGN_EXTEND_VECTOR_INREG, RegVT) &&
15184 "We can't implement a sext load without SIGN_EXTEND_VECTOR_INREG!");
15186 SDValue Shuff = DAG.getSignExtendVectorInReg(SlicedVec, dl, RegVT);
15187 DAG.ReplaceAllUsesOfValueWith(SDValue(Ld, 1), TF);
15191 // Redistribute the loaded elements into the different locations.
15192 SmallVector<int, 16> ShuffleVec(NumElems * SizeRatio, -1);
15193 for (unsigned i = 0; i != NumElems; ++i)
15194 ShuffleVec[i * SizeRatio] = i;
15196 SDValue Shuff = DAG.getVectorShuffle(WideVecVT, dl, SlicedVec,
15197 DAG.getUNDEF(WideVecVT), &ShuffleVec[0]);
15199 // Bitcast to the requested type.
15200 Shuff = DAG.getBitcast(RegVT, Shuff);
15201 DAG.ReplaceAllUsesOfValueWith(SDValue(Ld, 1), TF);
15205 // isAndOrOfSingleUseSetCCs - Return true if node is an ISD::AND or
15206 // ISD::OR of two X86ISD::SETCC nodes each of which has no other use apart
15207 // from the AND / OR.
15208 static bool isAndOrOfSetCCs(SDValue Op, unsigned &Opc) {
15209 Opc = Op.getOpcode();
15210 if (Opc != ISD::OR && Opc != ISD::AND)
15212 return (Op.getOperand(0).getOpcode() == X86ISD::SETCC &&
15213 Op.getOperand(0).hasOneUse() &&
15214 Op.getOperand(1).getOpcode() == X86ISD::SETCC &&
15215 Op.getOperand(1).hasOneUse());
15218 // isXor1OfSetCC - Return true if node is an ISD::XOR of a X86ISD::SETCC and
15219 // 1 and that the SETCC node has a single use.
15220 static bool isXor1OfSetCC(SDValue Op) {
15221 if (Op.getOpcode() != ISD::XOR)
15223 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(Op.getOperand(1));
15224 if (N1C && N1C->getAPIntValue() == 1) {
15225 return Op.getOperand(0).getOpcode() == X86ISD::SETCC &&
15226 Op.getOperand(0).hasOneUse();
15231 SDValue X86TargetLowering::LowerBRCOND(SDValue Op, SelectionDAG &DAG) const {
15232 bool addTest = true;
15233 SDValue Chain = Op.getOperand(0);
15234 SDValue Cond = Op.getOperand(1);
15235 SDValue Dest = Op.getOperand(2);
15238 bool Inverted = false;
15240 if (Cond.getOpcode() == ISD::SETCC) {
15241 // Check for setcc([su]{add,sub,mul}o == 0).
15242 if (cast<CondCodeSDNode>(Cond.getOperand(2))->get() == ISD::SETEQ &&
15243 isa<ConstantSDNode>(Cond.getOperand(1)) &&
15244 cast<ConstantSDNode>(Cond.getOperand(1))->isNullValue() &&
15245 Cond.getOperand(0).getResNo() == 1 &&
15246 (Cond.getOperand(0).getOpcode() == ISD::SADDO ||
15247 Cond.getOperand(0).getOpcode() == ISD::UADDO ||
15248 Cond.getOperand(0).getOpcode() == ISD::SSUBO ||
15249 Cond.getOperand(0).getOpcode() == ISD::USUBO ||
15250 Cond.getOperand(0).getOpcode() == ISD::SMULO ||
15251 Cond.getOperand(0).getOpcode() == ISD::UMULO)) {
15253 Cond = Cond.getOperand(0);
15255 SDValue NewCond = LowerSETCC(Cond, DAG);
15256 if (NewCond.getNode())
15261 // FIXME: LowerXALUO doesn't handle these!!
15262 else if (Cond.getOpcode() == X86ISD::ADD ||
15263 Cond.getOpcode() == X86ISD::SUB ||
15264 Cond.getOpcode() == X86ISD::SMUL ||
15265 Cond.getOpcode() == X86ISD::UMUL)
15266 Cond = LowerXALUO(Cond, DAG);
15269 // Look pass (and (setcc_carry (cmp ...)), 1).
15270 if (Cond.getOpcode() == ISD::AND &&
15271 Cond.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY) {
15272 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Cond.getOperand(1));
15273 if (C && C->getAPIntValue() == 1)
15274 Cond = Cond.getOperand(0);
15277 // If condition flag is set by a X86ISD::CMP, then use it as the condition
15278 // setting operand in place of the X86ISD::SETCC.
15279 unsigned CondOpcode = Cond.getOpcode();
15280 if (CondOpcode == X86ISD::SETCC ||
15281 CondOpcode == X86ISD::SETCC_CARRY) {
15282 CC = Cond.getOperand(0);
15284 SDValue Cmp = Cond.getOperand(1);
15285 unsigned Opc = Cmp.getOpcode();
15286 // FIXME: WHY THE SPECIAL CASING OF LogicalCmp??
15287 if (isX86LogicalCmp(Cmp) || Opc == X86ISD::BT) {
15291 switch (cast<ConstantSDNode>(CC)->getZExtValue()) {
15295 // These can only come from an arithmetic instruction with overflow,
15296 // e.g. SADDO, UADDO.
15297 Cond = Cond.getNode()->getOperand(1);
15303 CondOpcode = Cond.getOpcode();
15304 if (CondOpcode == ISD::UADDO || CondOpcode == ISD::SADDO ||
15305 CondOpcode == ISD::USUBO || CondOpcode == ISD::SSUBO ||
15306 ((CondOpcode == ISD::UMULO || CondOpcode == ISD::SMULO) &&
15307 Cond.getOperand(0).getValueType() != MVT::i8)) {
15308 SDValue LHS = Cond.getOperand(0);
15309 SDValue RHS = Cond.getOperand(1);
15310 unsigned X86Opcode;
15313 // Keep this in sync with LowerXALUO, otherwise we might create redundant
15314 // instructions that can't be removed afterwards (i.e. X86ISD::ADD and
15316 switch (CondOpcode) {
15317 case ISD::UADDO: X86Opcode = X86ISD::ADD; X86Cond = X86::COND_B; break;
15319 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(RHS))
15321 X86Opcode = X86ISD::INC; X86Cond = X86::COND_O;
15324 X86Opcode = X86ISD::ADD; X86Cond = X86::COND_O; break;
15325 case ISD::USUBO: X86Opcode = X86ISD::SUB; X86Cond = X86::COND_B; break;
15327 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(RHS))
15329 X86Opcode = X86ISD::DEC; X86Cond = X86::COND_O;
15332 X86Opcode = X86ISD::SUB; X86Cond = X86::COND_O; break;
15333 case ISD::UMULO: X86Opcode = X86ISD::UMUL; X86Cond = X86::COND_O; break;
15334 case ISD::SMULO: X86Opcode = X86ISD::SMUL; X86Cond = X86::COND_O; break;
15335 default: llvm_unreachable("unexpected overflowing operator");
15338 X86Cond = X86::GetOppositeBranchCondition((X86::CondCode)X86Cond);
15339 if (CondOpcode == ISD::UMULO)
15340 VTs = DAG.getVTList(LHS.getValueType(), LHS.getValueType(),
15343 VTs = DAG.getVTList(LHS.getValueType(), MVT::i32);
15345 SDValue X86Op = DAG.getNode(X86Opcode, dl, VTs, LHS, RHS);
15347 if (CondOpcode == ISD::UMULO)
15348 Cond = X86Op.getValue(2);
15350 Cond = X86Op.getValue(1);
15352 CC = DAG.getConstant(X86Cond, dl, MVT::i8);
15356 if (Cond.hasOneUse() && isAndOrOfSetCCs(Cond, CondOpc)) {
15357 SDValue Cmp = Cond.getOperand(0).getOperand(1);
15358 if (CondOpc == ISD::OR) {
15359 // Also, recognize the pattern generated by an FCMP_UNE. We can emit
15360 // two branches instead of an explicit OR instruction with a
15362 if (Cmp == Cond.getOperand(1).getOperand(1) &&
15363 isX86LogicalCmp(Cmp)) {
15364 CC = Cond.getOperand(0).getOperand(0);
15365 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
15366 Chain, Dest, CC, Cmp);
15367 CC = Cond.getOperand(1).getOperand(0);
15371 } else { // ISD::AND
15372 // Also, recognize the pattern generated by an FCMP_OEQ. We can emit
15373 // two branches instead of an explicit AND instruction with a
15374 // separate test. However, we only do this if this block doesn't
15375 // have a fall-through edge, because this requires an explicit
15376 // jmp when the condition is false.
15377 if (Cmp == Cond.getOperand(1).getOperand(1) &&
15378 isX86LogicalCmp(Cmp) &&
15379 Op.getNode()->hasOneUse()) {
15380 X86::CondCode CCode =
15381 (X86::CondCode)Cond.getOperand(0).getConstantOperandVal(0);
15382 CCode = X86::GetOppositeBranchCondition(CCode);
15383 CC = DAG.getConstant(CCode, dl, MVT::i8);
15384 SDNode *User = *Op.getNode()->use_begin();
15385 // Look for an unconditional branch following this conditional branch.
15386 // We need this because we need to reverse the successors in order
15387 // to implement FCMP_OEQ.
15388 if (User->getOpcode() == ISD::BR) {
15389 SDValue FalseBB = User->getOperand(1);
15391 DAG.UpdateNodeOperands(User, User->getOperand(0), Dest);
15392 assert(NewBR == User);
15396 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
15397 Chain, Dest, CC, Cmp);
15398 X86::CondCode CCode =
15399 (X86::CondCode)Cond.getOperand(1).getConstantOperandVal(0);
15400 CCode = X86::GetOppositeBranchCondition(CCode);
15401 CC = DAG.getConstant(CCode, dl, MVT::i8);
15407 } else if (Cond.hasOneUse() && isXor1OfSetCC(Cond)) {
15408 // Recognize for xorb (setcc), 1 patterns. The xor inverts the condition.
15409 // It should be transformed during dag combiner except when the condition
15410 // is set by a arithmetics with overflow node.
15411 X86::CondCode CCode =
15412 (X86::CondCode)Cond.getOperand(0).getConstantOperandVal(0);
15413 CCode = X86::GetOppositeBranchCondition(CCode);
15414 CC = DAG.getConstant(CCode, dl, MVT::i8);
15415 Cond = Cond.getOperand(0).getOperand(1);
15417 } else if (Cond.getOpcode() == ISD::SETCC &&
15418 cast<CondCodeSDNode>(Cond.getOperand(2))->get() == ISD::SETOEQ) {
15419 // For FCMP_OEQ, we can emit
15420 // two branches instead of an explicit AND instruction with a
15421 // separate test. However, we only do this if this block doesn't
15422 // have a fall-through edge, because this requires an explicit
15423 // jmp when the condition is false.
15424 if (Op.getNode()->hasOneUse()) {
15425 SDNode *User = *Op.getNode()->use_begin();
15426 // Look for an unconditional branch following this conditional branch.
15427 // We need this because we need to reverse the successors in order
15428 // to implement FCMP_OEQ.
15429 if (User->getOpcode() == ISD::BR) {
15430 SDValue FalseBB = User->getOperand(1);
15432 DAG.UpdateNodeOperands(User, User->getOperand(0), Dest);
15433 assert(NewBR == User);
15437 SDValue Cmp = DAG.getNode(X86ISD::CMP, dl, MVT::i32,
15438 Cond.getOperand(0), Cond.getOperand(1));
15439 Cmp = ConvertCmpIfNecessary(Cmp, DAG);
15440 CC = DAG.getConstant(X86::COND_NE, dl, MVT::i8);
15441 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
15442 Chain, Dest, CC, Cmp);
15443 CC = DAG.getConstant(X86::COND_P, dl, MVT::i8);
15448 } else if (Cond.getOpcode() == ISD::SETCC &&
15449 cast<CondCodeSDNode>(Cond.getOperand(2))->get() == ISD::SETUNE) {
15450 // For FCMP_UNE, we can emit
15451 // two branches instead of an explicit AND instruction with a
15452 // separate test. However, we only do this if this block doesn't
15453 // have a fall-through edge, because this requires an explicit
15454 // jmp when the condition is false.
15455 if (Op.getNode()->hasOneUse()) {
15456 SDNode *User = *Op.getNode()->use_begin();
15457 // Look for an unconditional branch following this conditional branch.
15458 // We need this because we need to reverse the successors in order
15459 // to implement FCMP_UNE.
15460 if (User->getOpcode() == ISD::BR) {
15461 SDValue FalseBB = User->getOperand(1);
15463 DAG.UpdateNodeOperands(User, User->getOperand(0), Dest);
15464 assert(NewBR == User);
15467 SDValue Cmp = DAG.getNode(X86ISD::CMP, dl, MVT::i32,
15468 Cond.getOperand(0), Cond.getOperand(1));
15469 Cmp = ConvertCmpIfNecessary(Cmp, DAG);
15470 CC = DAG.getConstant(X86::COND_NE, dl, MVT::i8);
15471 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
15472 Chain, Dest, CC, Cmp);
15473 CC = DAG.getConstant(X86::COND_NP, dl, MVT::i8);
15483 // Look pass the truncate if the high bits are known zero.
15484 if (isTruncWithZeroHighBitsInput(Cond, DAG))
15485 Cond = Cond.getOperand(0);
15487 // We know the result of AND is compared against zero. Try to match
15489 if (Cond.getOpcode() == ISD::AND && Cond.hasOneUse()) {
15490 SDValue NewSetCC = LowerToBT(Cond, ISD::SETNE, dl, DAG);
15491 if (NewSetCC.getNode()) {
15492 CC = NewSetCC.getOperand(0);
15493 Cond = NewSetCC.getOperand(1);
15500 X86::CondCode X86Cond = Inverted ? X86::COND_E : X86::COND_NE;
15501 CC = DAG.getConstant(X86Cond, dl, MVT::i8);
15502 Cond = EmitTest(Cond, X86Cond, dl, DAG);
15504 Cond = ConvertCmpIfNecessary(Cond, DAG);
15505 return DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
15506 Chain, Dest, CC, Cond);
15509 // Lower dynamic stack allocation to _alloca call for Cygwin/Mingw targets.
15510 // Calls to _alloca are needed to probe the stack when allocating more than 4k
15511 // bytes in one go. Touching the stack at 4K increments is necessary to ensure
15512 // that the guard pages used by the OS virtual memory manager are allocated in
15513 // correct sequence.
15515 X86TargetLowering::LowerDYNAMIC_STACKALLOC(SDValue Op,
15516 SelectionDAG &DAG) const {
15517 MachineFunction &MF = DAG.getMachineFunction();
15518 bool SplitStack = MF.shouldSplitStack();
15519 bool Lower = (Subtarget->isOSWindows() && !Subtarget->isTargetMachO()) ||
15524 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
15525 SDNode* Node = Op.getNode();
15527 unsigned SPReg = TLI.getStackPointerRegisterToSaveRestore();
15528 assert(SPReg && "Target cannot require DYNAMIC_STACKALLOC expansion and"
15529 " not tell us which reg is the stack pointer!");
15530 EVT VT = Node->getValueType(0);
15531 SDValue Tmp1 = SDValue(Node, 0);
15532 SDValue Tmp2 = SDValue(Node, 1);
15533 SDValue Tmp3 = Node->getOperand(2);
15534 SDValue Chain = Tmp1.getOperand(0);
15536 // Chain the dynamic stack allocation so that it doesn't modify the stack
15537 // pointer when other instructions are using the stack.
15538 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(0, dl, true),
15541 SDValue Size = Tmp2.getOperand(1);
15542 SDValue SP = DAG.getCopyFromReg(Chain, dl, SPReg, VT);
15543 Chain = SP.getValue(1);
15544 unsigned Align = cast<ConstantSDNode>(Tmp3)->getZExtValue();
15545 const TargetFrameLowering &TFI = *Subtarget->getFrameLowering();
15546 unsigned StackAlign = TFI.getStackAlignment();
15547 Tmp1 = DAG.getNode(ISD::SUB, dl, VT, SP, Size); // Value
15548 if (Align > StackAlign)
15549 Tmp1 = DAG.getNode(ISD::AND, dl, VT, Tmp1,
15550 DAG.getConstant(-(uint64_t)Align, dl, VT));
15551 Chain = DAG.getCopyToReg(Chain, dl, SPReg, Tmp1); // Output chain
15553 Tmp2 = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(0, dl, true),
15554 DAG.getIntPtrConstant(0, dl, true), SDValue(),
15557 SDValue Ops[2] = { Tmp1, Tmp2 };
15558 return DAG.getMergeValues(Ops, dl);
15562 SDValue Chain = Op.getOperand(0);
15563 SDValue Size = Op.getOperand(1);
15564 unsigned Align = cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue();
15565 EVT VT = Op.getNode()->getValueType(0);
15567 bool Is64Bit = Subtarget->is64Bit();
15568 MVT SPTy = getPointerTy(DAG.getDataLayout());
15571 MachineRegisterInfo &MRI = MF.getRegInfo();
15574 // The 64 bit implementation of segmented stacks needs to clobber both r10
15575 // r11. This makes it impossible to use it along with nested parameters.
15576 const Function *F = MF.getFunction();
15578 for (Function::const_arg_iterator I = F->arg_begin(), E = F->arg_end();
15580 if (I->hasNestAttr())
15581 report_fatal_error("Cannot use segmented stacks with functions that "
15582 "have nested arguments.");
15585 const TargetRegisterClass *AddrRegClass = getRegClassFor(SPTy);
15586 unsigned Vreg = MRI.createVirtualRegister(AddrRegClass);
15587 Chain = DAG.getCopyToReg(Chain, dl, Vreg, Size);
15588 SDValue Value = DAG.getNode(X86ISD::SEG_ALLOCA, dl, SPTy, Chain,
15589 DAG.getRegister(Vreg, SPTy));
15590 SDValue Ops1[2] = { Value, Chain };
15591 return DAG.getMergeValues(Ops1, dl);
15594 const unsigned Reg = (Subtarget->isTarget64BitLP64() ? X86::RAX : X86::EAX);
15596 Chain = DAG.getCopyToReg(Chain, dl, Reg, Size, Flag);
15597 Flag = Chain.getValue(1);
15598 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
15600 Chain = DAG.getNode(X86ISD::WIN_ALLOCA, dl, NodeTys, Chain, Flag);
15602 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
15603 unsigned SPReg = RegInfo->getStackRegister();
15604 SDValue SP = DAG.getCopyFromReg(Chain, dl, SPReg, SPTy);
15605 Chain = SP.getValue(1);
15608 SP = DAG.getNode(ISD::AND, dl, VT, SP.getValue(0),
15609 DAG.getConstant(-(uint64_t)Align, dl, VT));
15610 Chain = DAG.getCopyToReg(Chain, dl, SPReg, SP);
15613 SDValue Ops1[2] = { SP, Chain };
15614 return DAG.getMergeValues(Ops1, dl);
15618 SDValue X86TargetLowering::LowerVASTART(SDValue Op, SelectionDAG &DAG) const {
15619 MachineFunction &MF = DAG.getMachineFunction();
15620 auto PtrVT = getPointerTy(MF.getDataLayout());
15621 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
15623 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
15626 if (!Subtarget->is64Bit() ||
15627 Subtarget->isCallingConvWin64(MF.getFunction()->getCallingConv())) {
15628 // vastart just stores the address of the VarArgsFrameIndex slot into the
15629 // memory location argument.
15630 SDValue FR = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(), PtrVT);
15631 return DAG.getStore(Op.getOperand(0), DL, FR, Op.getOperand(1),
15632 MachinePointerInfo(SV), false, false, 0);
15636 // gp_offset (0 - 6 * 8)
15637 // fp_offset (48 - 48 + 8 * 16)
15638 // overflow_arg_area (point to parameters coming in memory).
15640 SmallVector<SDValue, 8> MemOps;
15641 SDValue FIN = Op.getOperand(1);
15643 SDValue Store = DAG.getStore(Op.getOperand(0), DL,
15644 DAG.getConstant(FuncInfo->getVarArgsGPOffset(),
15646 FIN, MachinePointerInfo(SV), false, false, 0);
15647 MemOps.push_back(Store);
15650 FIN = DAG.getNode(ISD::ADD, DL, PtrVT, FIN, DAG.getIntPtrConstant(4, DL));
15651 Store = DAG.getStore(Op.getOperand(0), DL,
15652 DAG.getConstant(FuncInfo->getVarArgsFPOffset(), DL,
15654 FIN, MachinePointerInfo(SV, 4), false, false, 0);
15655 MemOps.push_back(Store);
15657 // Store ptr to overflow_arg_area
15658 FIN = DAG.getNode(ISD::ADD, DL, PtrVT, FIN, DAG.getIntPtrConstant(4, DL));
15659 SDValue OVFIN = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(), PtrVT);
15660 Store = DAG.getStore(Op.getOperand(0), DL, OVFIN, FIN,
15661 MachinePointerInfo(SV, 8),
15663 MemOps.push_back(Store);
15665 // Store ptr to reg_save_area.
15666 FIN = DAG.getNode(ISD::ADD, DL, PtrVT, FIN, DAG.getIntPtrConstant(
15667 Subtarget->isTarget64BitLP64() ? 8 : 4, DL));
15668 SDValue RSFIN = DAG.getFrameIndex(FuncInfo->getRegSaveFrameIndex(), PtrVT);
15669 Store = DAG.getStore(Op.getOperand(0), DL, RSFIN, FIN, MachinePointerInfo(
15670 SV, Subtarget->isTarget64BitLP64() ? 16 : 12), false, false, 0);
15671 MemOps.push_back(Store);
15672 return DAG.getNode(ISD::TokenFactor, DL, MVT::Other, MemOps);
15675 SDValue X86TargetLowering::LowerVAARG(SDValue Op, SelectionDAG &DAG) const {
15676 assert(Subtarget->is64Bit() &&
15677 "LowerVAARG only handles 64-bit va_arg!");
15678 assert(Op.getNode()->getNumOperands() == 4);
15680 MachineFunction &MF = DAG.getMachineFunction();
15681 if (Subtarget->isCallingConvWin64(MF.getFunction()->getCallingConv()))
15682 // The Win64 ABI uses char* instead of a structure.
15683 return DAG.expandVAArg(Op.getNode());
15685 SDValue Chain = Op.getOperand(0);
15686 SDValue SrcPtr = Op.getOperand(1);
15687 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
15688 unsigned Align = Op.getConstantOperandVal(3);
15691 EVT ArgVT = Op.getNode()->getValueType(0);
15692 Type *ArgTy = ArgVT.getTypeForEVT(*DAG.getContext());
15693 uint32_t ArgSize = DAG.getDataLayout().getTypeAllocSize(ArgTy);
15696 // Decide which area this value should be read from.
15697 // TODO: Implement the AMD64 ABI in its entirety. This simple
15698 // selection mechanism works only for the basic types.
15699 if (ArgVT == MVT::f80) {
15700 llvm_unreachable("va_arg for f80 not yet implemented");
15701 } else if (ArgVT.isFloatingPoint() && ArgSize <= 16 /*bytes*/) {
15702 ArgMode = 2; // Argument passed in XMM register. Use fp_offset.
15703 } else if (ArgVT.isInteger() && ArgSize <= 32 /*bytes*/) {
15704 ArgMode = 1; // Argument passed in GPR64 register(s). Use gp_offset.
15706 llvm_unreachable("Unhandled argument type in LowerVAARG");
15709 if (ArgMode == 2) {
15710 // Sanity Check: Make sure using fp_offset makes sense.
15711 assert(!Subtarget->useSoftFloat() &&
15712 !(MF.getFunction()->hasFnAttribute(Attribute::NoImplicitFloat)) &&
15713 Subtarget->hasSSE1());
15716 // Insert VAARG_64 node into the DAG
15717 // VAARG_64 returns two values: Variable Argument Address, Chain
15718 SDValue InstOps[] = {Chain, SrcPtr, DAG.getConstant(ArgSize, dl, MVT::i32),
15719 DAG.getConstant(ArgMode, dl, MVT::i8),
15720 DAG.getConstant(Align, dl, MVT::i32)};
15721 SDVTList VTs = DAG.getVTList(getPointerTy(DAG.getDataLayout()), MVT::Other);
15722 SDValue VAARG = DAG.getMemIntrinsicNode(X86ISD::VAARG_64, dl,
15723 VTs, InstOps, MVT::i64,
15724 MachinePointerInfo(SV),
15726 /*Volatile=*/false,
15728 /*WriteMem=*/true);
15729 Chain = VAARG.getValue(1);
15731 // Load the next argument and return it
15732 return DAG.getLoad(ArgVT, dl,
15735 MachinePointerInfo(),
15736 false, false, false, 0);
15739 static SDValue LowerVACOPY(SDValue Op, const X86Subtarget *Subtarget,
15740 SelectionDAG &DAG) {
15741 // X86-64 va_list is a struct { i32, i32, i8*, i8* }, except on Windows,
15742 // where a va_list is still an i8*.
15743 assert(Subtarget->is64Bit() && "This code only handles 64-bit va_copy!");
15744 if (Subtarget->isCallingConvWin64(
15745 DAG.getMachineFunction().getFunction()->getCallingConv()))
15746 // Probably a Win64 va_copy.
15747 return DAG.expandVACopy(Op.getNode());
15749 SDValue Chain = Op.getOperand(0);
15750 SDValue DstPtr = Op.getOperand(1);
15751 SDValue SrcPtr = Op.getOperand(2);
15752 const Value *DstSV = cast<SrcValueSDNode>(Op.getOperand(3))->getValue();
15753 const Value *SrcSV = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
15756 return DAG.getMemcpy(Chain, DL, DstPtr, SrcPtr,
15757 DAG.getIntPtrConstant(24, DL), 8, /*isVolatile*/false,
15759 MachinePointerInfo(DstSV), MachinePointerInfo(SrcSV));
15762 // getTargetVShiftByConstNode - Handle vector element shifts where the shift
15763 // amount is a constant. Takes immediate version of shift as input.
15764 static SDValue getTargetVShiftByConstNode(unsigned Opc, SDLoc dl, MVT VT,
15765 SDValue SrcOp, uint64_t ShiftAmt,
15766 SelectionDAG &DAG) {
15767 MVT ElementType = VT.getVectorElementType();
15769 // Fold this packed shift into its first operand if ShiftAmt is 0.
15773 // Check for ShiftAmt >= element width
15774 if (ShiftAmt >= ElementType.getSizeInBits()) {
15775 if (Opc == X86ISD::VSRAI)
15776 ShiftAmt = ElementType.getSizeInBits() - 1;
15778 return DAG.getConstant(0, dl, VT);
15781 assert((Opc == X86ISD::VSHLI || Opc == X86ISD::VSRLI || Opc == X86ISD::VSRAI)
15782 && "Unknown target vector shift-by-constant node");
15784 // Fold this packed vector shift into a build vector if SrcOp is a
15785 // vector of Constants or UNDEFs, and SrcOp valuetype is the same as VT.
15786 if (VT == SrcOp.getSimpleValueType() &&
15787 ISD::isBuildVectorOfConstantSDNodes(SrcOp.getNode())) {
15788 SmallVector<SDValue, 8> Elts;
15789 unsigned NumElts = SrcOp->getNumOperands();
15790 ConstantSDNode *ND;
15793 default: llvm_unreachable(nullptr);
15794 case X86ISD::VSHLI:
15795 for (unsigned i=0; i!=NumElts; ++i) {
15796 SDValue CurrentOp = SrcOp->getOperand(i);
15797 if (CurrentOp->getOpcode() == ISD::UNDEF) {
15798 Elts.push_back(CurrentOp);
15801 ND = cast<ConstantSDNode>(CurrentOp);
15802 const APInt &C = ND->getAPIntValue();
15803 Elts.push_back(DAG.getConstant(C.shl(ShiftAmt), dl, ElementType));
15806 case X86ISD::VSRLI:
15807 for (unsigned i=0; i!=NumElts; ++i) {
15808 SDValue CurrentOp = SrcOp->getOperand(i);
15809 if (CurrentOp->getOpcode() == ISD::UNDEF) {
15810 Elts.push_back(CurrentOp);
15813 ND = cast<ConstantSDNode>(CurrentOp);
15814 const APInt &C = ND->getAPIntValue();
15815 Elts.push_back(DAG.getConstant(C.lshr(ShiftAmt), dl, ElementType));
15818 case X86ISD::VSRAI:
15819 for (unsigned i=0; i!=NumElts; ++i) {
15820 SDValue CurrentOp = SrcOp->getOperand(i);
15821 if (CurrentOp->getOpcode() == ISD::UNDEF) {
15822 Elts.push_back(CurrentOp);
15825 ND = cast<ConstantSDNode>(CurrentOp);
15826 const APInt &C = ND->getAPIntValue();
15827 Elts.push_back(DAG.getConstant(C.ashr(ShiftAmt), dl, ElementType));
15832 return DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Elts);
15835 return DAG.getNode(Opc, dl, VT, SrcOp,
15836 DAG.getConstant(ShiftAmt, dl, MVT::i8));
15839 // getTargetVShiftNode - Handle vector element shifts where the shift amount
15840 // may or may not be a constant. Takes immediate version of shift as input.
15841 static SDValue getTargetVShiftNode(unsigned Opc, SDLoc dl, MVT VT,
15842 SDValue SrcOp, SDValue ShAmt,
15843 SelectionDAG &DAG) {
15844 MVT SVT = ShAmt.getSimpleValueType();
15845 assert((SVT == MVT::i32 || SVT == MVT::i64) && "Unexpected value type!");
15847 // Catch shift-by-constant.
15848 if (ConstantSDNode *CShAmt = dyn_cast<ConstantSDNode>(ShAmt))
15849 return getTargetVShiftByConstNode(Opc, dl, VT, SrcOp,
15850 CShAmt->getZExtValue(), DAG);
15852 // Change opcode to non-immediate version
15854 default: llvm_unreachable("Unknown target vector shift node");
15855 case X86ISD::VSHLI: Opc = X86ISD::VSHL; break;
15856 case X86ISD::VSRLI: Opc = X86ISD::VSRL; break;
15857 case X86ISD::VSRAI: Opc = X86ISD::VSRA; break;
15860 const X86Subtarget &Subtarget =
15861 static_cast<const X86Subtarget &>(DAG.getSubtarget());
15862 if (Subtarget.hasSSE41() && ShAmt.getOpcode() == ISD::ZERO_EXTEND &&
15863 ShAmt.getOperand(0).getSimpleValueType() == MVT::i16) {
15864 // Let the shuffle legalizer expand this shift amount node.
15865 SDValue Op0 = ShAmt.getOperand(0);
15866 Op0 = DAG.getNode(ISD::SCALAR_TO_VECTOR, SDLoc(Op0), MVT::v8i16, Op0);
15867 ShAmt = getShuffleVectorZeroOrUndef(Op0, 0, true, &Subtarget, DAG);
15869 // Need to build a vector containing shift amount.
15870 // SSE/AVX packed shifts only use the lower 64-bit of the shift count.
15871 SmallVector<SDValue, 4> ShOps;
15872 ShOps.push_back(ShAmt);
15873 if (SVT == MVT::i32) {
15874 ShOps.push_back(DAG.getConstant(0, dl, SVT));
15875 ShOps.push_back(DAG.getUNDEF(SVT));
15877 ShOps.push_back(DAG.getUNDEF(SVT));
15879 MVT BVT = SVT == MVT::i32 ? MVT::v4i32 : MVT::v2i64;
15880 ShAmt = DAG.getNode(ISD::BUILD_VECTOR, dl, BVT, ShOps);
15883 // The return type has to be a 128-bit type with the same element
15884 // type as the input type.
15885 MVT EltVT = VT.getVectorElementType();
15886 EVT ShVT = MVT::getVectorVT(EltVT, 128/EltVT.getSizeInBits());
15888 ShAmt = DAG.getBitcast(ShVT, ShAmt);
15889 return DAG.getNode(Opc, dl, VT, SrcOp, ShAmt);
15892 /// \brief Return (and \p Op, \p Mask) for compare instructions or
15893 /// (vselect \p Mask, \p Op, \p PreservedSrc) for others along with the
15894 /// necessary casting or extending for \p Mask when lowering masking intrinsics
15895 static SDValue getVectorMaskingNode(SDValue Op, SDValue Mask,
15896 SDValue PreservedSrc,
15897 const X86Subtarget *Subtarget,
15898 SelectionDAG &DAG) {
15899 EVT VT = Op.getValueType();
15900 EVT MaskVT = EVT::getVectorVT(*DAG.getContext(),
15901 MVT::i1, VT.getVectorNumElements());
15902 SDValue VMask = SDValue();
15903 unsigned OpcodeSelect = ISD::VSELECT;
15906 assert(MaskVT.isSimple() && "invalid mask type");
15908 if (isAllOnes(Mask))
15911 if (MaskVT.bitsGT(Mask.getValueType())) {
15912 EVT newMaskVT = EVT::getIntegerVT(*DAG.getContext(),
15913 MaskVT.getSizeInBits());
15914 VMask = DAG.getBitcast(MaskVT,
15915 DAG.getNode(ISD::ANY_EXTEND, dl, newMaskVT, Mask));
15917 EVT BitcastVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
15918 Mask.getValueType().getSizeInBits());
15919 // In case when MaskVT equals v2i1 or v4i1, low 2 or 4 elements
15920 // are extracted by EXTRACT_SUBVECTOR.
15921 VMask = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MaskVT,
15922 DAG.getBitcast(BitcastVT, Mask),
15923 DAG.getIntPtrConstant(0, dl));
15926 switch (Op.getOpcode()) {
15928 case X86ISD::PCMPEQM:
15929 case X86ISD::PCMPGTM:
15931 case X86ISD::CMPMU:
15932 return DAG.getNode(ISD::AND, dl, VT, Op, VMask);
15933 case X86ISD::VFPCLASS:
15934 return DAG.getNode(ISD::OR, dl, VT, Op, VMask);
15935 case X86ISD::VTRUNC:
15936 case X86ISD::VTRUNCS:
15937 case X86ISD::VTRUNCUS:
15938 // We can't use ISD::VSELECT here because it is not always "Legal"
15939 // for the destination type. For example vpmovqb require only AVX512
15940 // and vselect that can operate on byte element type require BWI
15941 OpcodeSelect = X86ISD::SELECT;
15944 if (PreservedSrc.getOpcode() == ISD::UNDEF)
15945 PreservedSrc = getZeroVector(VT, Subtarget, DAG, dl);
15946 return DAG.getNode(OpcodeSelect, dl, VT, VMask, Op, PreservedSrc);
15949 /// \brief Creates an SDNode for a predicated scalar operation.
15950 /// \returns (X86vselect \p Mask, \p Op, \p PreservedSrc).
15951 /// The mask is coming as MVT::i8 and it should be truncated
15952 /// to MVT::i1 while lowering masking intrinsics.
15953 /// The main difference between ScalarMaskingNode and VectorMaskingNode is using
15954 /// "X86select" instead of "vselect". We just can't create the "vselect" node
15955 /// for a scalar instruction.
15956 static SDValue getScalarMaskingNode(SDValue Op, SDValue Mask,
15957 SDValue PreservedSrc,
15958 const X86Subtarget *Subtarget,
15959 SelectionDAG &DAG) {
15960 if (isAllOnes(Mask))
15963 EVT VT = Op.getValueType();
15965 // The mask should be of type MVT::i1
15966 SDValue IMask = DAG.getNode(ISD::TRUNCATE, dl, MVT::i1, Mask);
15968 if (Op.getOpcode() == X86ISD::FSETCC)
15969 return DAG.getNode(ISD::AND, dl, VT, Op, IMask);
15970 if (Op.getOpcode() == X86ISD::VFPCLASS)
15971 return DAG.getNode(ISD::OR, dl, VT, Op, IMask);
15973 if (PreservedSrc.getOpcode() == ISD::UNDEF)
15974 PreservedSrc = getZeroVector(VT, Subtarget, DAG, dl);
15975 return DAG.getNode(X86ISD::SELECT, dl, VT, IMask, Op, PreservedSrc);
15978 static int getSEHRegistrationNodeSize(const Function *Fn) {
15979 if (!Fn->hasPersonalityFn())
15980 report_fatal_error(
15981 "querying registration node size for function without personality");
15982 // The RegNodeSize is 6 32-bit words for SEH and 4 for C++ EH. See
15983 // WinEHStatePass for the full struct definition.
15984 switch (classifyEHPersonality(Fn->getPersonalityFn())) {
15985 case EHPersonality::MSVC_X86SEH: return 24;
15986 case EHPersonality::MSVC_CXX: return 16;
15989 report_fatal_error("can only recover FP for MSVC EH personality functions");
15992 /// When the 32-bit MSVC runtime transfers control to us, either to an outlined
15993 /// function or when returning to a parent frame after catching an exception, we
15994 /// recover the parent frame pointer by doing arithmetic on the incoming EBP.
15995 /// Here's the math:
15996 /// RegNodeBase = EntryEBP - RegNodeSize
15997 /// ParentFP = RegNodeBase - RegNodeFrameOffset
15998 /// Subtracting RegNodeSize takes us to the offset of the registration node, and
15999 /// subtracting the offset (negative on x86) takes us back to the parent FP.
16000 static SDValue recoverFramePointer(SelectionDAG &DAG, const Function *Fn,
16001 SDValue EntryEBP) {
16002 MachineFunction &MF = DAG.getMachineFunction();
16005 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
16006 MVT PtrVT = TLI.getPointerTy(DAG.getDataLayout());
16008 // It's possible that the parent function no longer has a personality function
16009 // if the exceptional code was optimized away, in which case we just return
16010 // the incoming EBP.
16011 if (!Fn->hasPersonalityFn())
16014 int RegNodeSize = getSEHRegistrationNodeSize(Fn);
16016 // Get an MCSymbol that will ultimately resolve to the frame offset of the EH
16018 MCSymbol *OffsetSym =
16019 MF.getMMI().getContext().getOrCreateParentFrameOffsetSymbol(
16020 GlobalValue::getRealLinkageName(Fn->getName()));
16021 SDValue OffsetSymVal = DAG.getMCSymbol(OffsetSym, PtrVT);
16022 SDValue RegNodeFrameOffset =
16023 DAG.getNode(ISD::LOCAL_RECOVER, dl, PtrVT, OffsetSymVal);
16025 // RegNodeBase = EntryEBP - RegNodeSize
16026 // ParentFP = RegNodeBase - RegNodeFrameOffset
16027 SDValue RegNodeBase = DAG.getNode(ISD::SUB, dl, PtrVT, EntryEBP,
16028 DAG.getConstant(RegNodeSize, dl, PtrVT));
16029 return DAG.getNode(ISD::SUB, dl, PtrVT, RegNodeBase, RegNodeFrameOffset);
16032 static SDValue LowerINTRINSIC_WO_CHAIN(SDValue Op, const X86Subtarget *Subtarget,
16033 SelectionDAG &DAG) {
16035 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
16036 EVT VT = Op.getValueType();
16037 const IntrinsicData* IntrData = getIntrinsicWithoutChain(IntNo);
16039 switch(IntrData->Type) {
16040 case INTR_TYPE_1OP:
16041 return DAG.getNode(IntrData->Opc0, dl, Op.getValueType(), Op.getOperand(1));
16042 case INTR_TYPE_2OP:
16043 return DAG.getNode(IntrData->Opc0, dl, Op.getValueType(), Op.getOperand(1),
16045 case INTR_TYPE_2OP_IMM8:
16046 return DAG.getNode(IntrData->Opc0, dl, Op.getValueType(), Op.getOperand(1),
16047 DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op.getOperand(2)));
16048 case INTR_TYPE_3OP:
16049 return DAG.getNode(IntrData->Opc0, dl, Op.getValueType(), Op.getOperand(1),
16050 Op.getOperand(2), Op.getOperand(3));
16051 case INTR_TYPE_4OP:
16052 return DAG.getNode(IntrData->Opc0, dl, Op.getValueType(), Op.getOperand(1),
16053 Op.getOperand(2), Op.getOperand(3), Op.getOperand(4));
16054 case INTR_TYPE_1OP_MASK_RM: {
16055 SDValue Src = Op.getOperand(1);
16056 SDValue PassThru = Op.getOperand(2);
16057 SDValue Mask = Op.getOperand(3);
16058 SDValue RoundingMode;
16059 // We allways add rounding mode to the Node.
16060 // If the rounding mode is not specified, we add the
16061 // "current direction" mode.
16062 if (Op.getNumOperands() == 4)
16064 DAG.getConstant(X86::STATIC_ROUNDING::CUR_DIRECTION, dl, MVT::i32);
16066 RoundingMode = Op.getOperand(4);
16067 unsigned IntrWithRoundingModeOpcode = IntrData->Opc1;
16068 if (IntrWithRoundingModeOpcode != 0)
16069 if (cast<ConstantSDNode>(RoundingMode)->getZExtValue() !=
16070 X86::STATIC_ROUNDING::CUR_DIRECTION)
16071 return getVectorMaskingNode(DAG.getNode(IntrWithRoundingModeOpcode,
16072 dl, Op.getValueType(), Src, RoundingMode),
16073 Mask, PassThru, Subtarget, DAG);
16074 return getVectorMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT, Src,
16076 Mask, PassThru, Subtarget, DAG);
16078 case INTR_TYPE_1OP_MASK: {
16079 SDValue Src = Op.getOperand(1);
16080 SDValue PassThru = Op.getOperand(2);
16081 SDValue Mask = Op.getOperand(3);
16082 // We add rounding mode to the Node when
16083 // - RM Opcode is specified and
16084 // - RM is not "current direction".
16085 unsigned IntrWithRoundingModeOpcode = IntrData->Opc1;
16086 if (IntrWithRoundingModeOpcode != 0) {
16087 SDValue Rnd = Op.getOperand(4);
16088 unsigned Round = cast<ConstantSDNode>(Rnd)->getZExtValue();
16089 if (Round != X86::STATIC_ROUNDING::CUR_DIRECTION) {
16090 return getVectorMaskingNode(DAG.getNode(IntrWithRoundingModeOpcode,
16091 dl, Op.getValueType(),
16093 Mask, PassThru, Subtarget, DAG);
16096 return getVectorMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT, Src),
16097 Mask, PassThru, Subtarget, DAG);
16099 case INTR_TYPE_SCALAR_MASK: {
16100 SDValue Src1 = Op.getOperand(1);
16101 SDValue Src2 = Op.getOperand(2);
16102 SDValue passThru = Op.getOperand(3);
16103 SDValue Mask = Op.getOperand(4);
16104 return getScalarMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT, Src1, Src2),
16105 Mask, passThru, Subtarget, DAG);
16107 case INTR_TYPE_SCALAR_MASK_RM: {
16108 SDValue Src1 = Op.getOperand(1);
16109 SDValue Src2 = Op.getOperand(2);
16110 SDValue Src0 = Op.getOperand(3);
16111 SDValue Mask = Op.getOperand(4);
16112 // There are 2 kinds of intrinsics in this group:
16113 // (1) With suppress-all-exceptions (sae) or rounding mode- 6 operands
16114 // (2) With rounding mode and sae - 7 operands.
16115 if (Op.getNumOperands() == 6) {
16116 SDValue Sae = Op.getOperand(5);
16117 unsigned Opc = IntrData->Opc1 ? IntrData->Opc1 : IntrData->Opc0;
16118 return getScalarMaskingNode(DAG.getNode(Opc, dl, VT, Src1, Src2,
16120 Mask, Src0, Subtarget, DAG);
16122 assert(Op.getNumOperands() == 7 && "Unexpected intrinsic form");
16123 SDValue RoundingMode = Op.getOperand(5);
16124 SDValue Sae = Op.getOperand(6);
16125 return getScalarMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT, Src1, Src2,
16126 RoundingMode, Sae),
16127 Mask, Src0, Subtarget, DAG);
16129 case INTR_TYPE_2OP_MASK:
16130 case INTR_TYPE_2OP_IMM8_MASK: {
16131 SDValue Src1 = Op.getOperand(1);
16132 SDValue Src2 = Op.getOperand(2);
16133 SDValue PassThru = Op.getOperand(3);
16134 SDValue Mask = Op.getOperand(4);
16136 if (IntrData->Type == INTR_TYPE_2OP_IMM8_MASK)
16137 Src2 = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Src2);
16139 // We specify 2 possible opcodes for intrinsics with rounding modes.
16140 // First, we check if the intrinsic may have non-default rounding mode,
16141 // (IntrData->Opc1 != 0), then we check the rounding mode operand.
16142 unsigned IntrWithRoundingModeOpcode = IntrData->Opc1;
16143 if (IntrWithRoundingModeOpcode != 0) {
16144 SDValue Rnd = Op.getOperand(5);
16145 unsigned Round = cast<ConstantSDNode>(Rnd)->getZExtValue();
16146 if (Round != X86::STATIC_ROUNDING::CUR_DIRECTION) {
16147 return getVectorMaskingNode(DAG.getNode(IntrWithRoundingModeOpcode,
16148 dl, Op.getValueType(),
16150 Mask, PassThru, Subtarget, DAG);
16153 // TODO: Intrinsics should have fast-math-flags to propagate.
16154 return getVectorMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT,Src1,Src2),
16155 Mask, PassThru, Subtarget, DAG);
16157 case INTR_TYPE_2OP_MASK_RM: {
16158 SDValue Src1 = Op.getOperand(1);
16159 SDValue Src2 = Op.getOperand(2);
16160 SDValue PassThru = Op.getOperand(3);
16161 SDValue Mask = Op.getOperand(4);
16162 // We specify 2 possible modes for intrinsics, with/without rounding
16164 // First, we check if the intrinsic have rounding mode (6 operands),
16165 // if not, we set rounding mode to "current".
16167 if (Op.getNumOperands() == 6)
16168 Rnd = Op.getOperand(5);
16170 Rnd = DAG.getConstant(X86::STATIC_ROUNDING::CUR_DIRECTION, dl, MVT::i32);
16171 return getVectorMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT,
16173 Mask, PassThru, Subtarget, DAG);
16175 case INTR_TYPE_3OP_SCALAR_MASK_RM: {
16176 SDValue Src1 = Op.getOperand(1);
16177 SDValue Src2 = Op.getOperand(2);
16178 SDValue Src3 = Op.getOperand(3);
16179 SDValue PassThru = Op.getOperand(4);
16180 SDValue Mask = Op.getOperand(5);
16181 SDValue Sae = Op.getOperand(6);
16183 return getScalarMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT, Src1,
16185 Mask, PassThru, Subtarget, DAG);
16187 case INTR_TYPE_3OP_MASK_RM: {
16188 SDValue Src1 = Op.getOperand(1);
16189 SDValue Src2 = Op.getOperand(2);
16190 SDValue Imm = Op.getOperand(3);
16191 SDValue PassThru = Op.getOperand(4);
16192 SDValue Mask = Op.getOperand(5);
16193 // We specify 2 possible modes for intrinsics, with/without rounding
16195 // First, we check if the intrinsic have rounding mode (7 operands),
16196 // if not, we set rounding mode to "current".
16198 if (Op.getNumOperands() == 7)
16199 Rnd = Op.getOperand(6);
16201 Rnd = DAG.getConstant(X86::STATIC_ROUNDING::CUR_DIRECTION, dl, MVT::i32);
16202 return getVectorMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT,
16203 Src1, Src2, Imm, Rnd),
16204 Mask, PassThru, Subtarget, DAG);
16206 case INTR_TYPE_3OP_IMM8_MASK:
16207 case INTR_TYPE_3OP_MASK:
16208 case INSERT_SUBVEC: {
16209 SDValue Src1 = Op.getOperand(1);
16210 SDValue Src2 = Op.getOperand(2);
16211 SDValue Src3 = Op.getOperand(3);
16212 SDValue PassThru = Op.getOperand(4);
16213 SDValue Mask = Op.getOperand(5);
16215 if (IntrData->Type == INTR_TYPE_3OP_IMM8_MASK)
16216 Src3 = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Src3);
16217 else if (IntrData->Type == INSERT_SUBVEC) {
16218 // imm should be adapted to ISD::INSERT_SUBVECTOR behavior
16219 assert(isa<ConstantSDNode>(Src3) && "Expected a ConstantSDNode here!");
16220 unsigned Imm = cast<ConstantSDNode>(Src3)->getZExtValue();
16221 Imm *= Src2.getValueType().getVectorNumElements();
16222 Src3 = DAG.getTargetConstant(Imm, dl, MVT::i32);
16225 // We specify 2 possible opcodes for intrinsics with rounding modes.
16226 // First, we check if the intrinsic may have non-default rounding mode,
16227 // (IntrData->Opc1 != 0), then we check the rounding mode operand.
16228 unsigned IntrWithRoundingModeOpcode = IntrData->Opc1;
16229 if (IntrWithRoundingModeOpcode != 0) {
16230 SDValue Rnd = Op.getOperand(6);
16231 unsigned Round = cast<ConstantSDNode>(Rnd)->getZExtValue();
16232 if (Round != X86::STATIC_ROUNDING::CUR_DIRECTION) {
16233 return getVectorMaskingNode(DAG.getNode(IntrWithRoundingModeOpcode,
16234 dl, Op.getValueType(),
16235 Src1, Src2, Src3, Rnd),
16236 Mask, PassThru, Subtarget, DAG);
16239 return getVectorMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT,
16241 Mask, PassThru, Subtarget, DAG);
16243 case VPERM_3OP_MASKZ:
16244 case VPERM_3OP_MASK:
16247 case FMA_OP_MASK: {
16248 SDValue Src1 = Op.getOperand(1);
16249 SDValue Src2 = Op.getOperand(2);
16250 SDValue Src3 = Op.getOperand(3);
16251 SDValue Mask = Op.getOperand(4);
16252 EVT VT = Op.getValueType();
16253 SDValue PassThru = SDValue();
16255 // set PassThru element
16256 if (IntrData->Type == VPERM_3OP_MASKZ || IntrData->Type == FMA_OP_MASKZ)
16257 PassThru = getZeroVector(VT, Subtarget, DAG, dl);
16258 else if (IntrData->Type == FMA_OP_MASK3)
16263 // We specify 2 possible opcodes for intrinsics with rounding modes.
16264 // First, we check if the intrinsic may have non-default rounding mode,
16265 // (IntrData->Opc1 != 0), then we check the rounding mode operand.
16266 unsigned IntrWithRoundingModeOpcode = IntrData->Opc1;
16267 if (IntrWithRoundingModeOpcode != 0) {
16268 SDValue Rnd = Op.getOperand(5);
16269 if (cast<ConstantSDNode>(Rnd)->getZExtValue() !=
16270 X86::STATIC_ROUNDING::CUR_DIRECTION)
16271 return getVectorMaskingNode(DAG.getNode(IntrWithRoundingModeOpcode,
16272 dl, Op.getValueType(),
16273 Src1, Src2, Src3, Rnd),
16274 Mask, PassThru, Subtarget, DAG);
16276 return getVectorMaskingNode(DAG.getNode(IntrData->Opc0,
16277 dl, Op.getValueType(),
16279 Mask, PassThru, Subtarget, DAG);
16281 case TERLOG_OP_MASK:
16282 case TERLOG_OP_MASKZ: {
16283 SDValue Src1 = Op.getOperand(1);
16284 SDValue Src2 = Op.getOperand(2);
16285 SDValue Src3 = Op.getOperand(3);
16286 SDValue Src4 = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op.getOperand(4));
16287 SDValue Mask = Op.getOperand(5);
16288 EVT VT = Op.getValueType();
16289 SDValue PassThru = Src1;
16290 // Set PassThru element.
16291 if (IntrData->Type == TERLOG_OP_MASKZ)
16292 PassThru = getZeroVector(VT, Subtarget, DAG, dl);
16294 return getVectorMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT,
16295 Src1, Src2, Src3, Src4),
16296 Mask, PassThru, Subtarget, DAG);
16299 // FPclass intrinsics with mask
16300 SDValue Src1 = Op.getOperand(1);
16301 EVT VT = Src1.getValueType();
16302 EVT MaskVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
16303 VT.getVectorNumElements());
16304 SDValue Imm = Op.getOperand(2);
16305 SDValue Mask = Op.getOperand(3);
16306 EVT BitcastVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
16307 Mask.getValueType().getSizeInBits());
16308 SDValue FPclass = DAG.getNode(IntrData->Opc0, dl, MaskVT, Src1, Imm);
16309 SDValue FPclassMask = getVectorMaskingNode(FPclass, Mask,
16310 DAG.getTargetConstant(0, dl, MaskVT),
16312 SDValue Res = DAG.getNode(ISD::INSERT_SUBVECTOR, dl, BitcastVT,
16313 DAG.getUNDEF(BitcastVT), FPclassMask,
16314 DAG.getIntPtrConstant(0, dl));
16315 return DAG.getBitcast(Op.getValueType(), Res);
16318 SDValue Src1 = Op.getOperand(1);
16319 SDValue Imm = Op.getOperand(2);
16320 SDValue Mask = Op.getOperand(3);
16321 SDValue FPclass = DAG.getNode(IntrData->Opc0, dl, MVT::i1, Src1, Imm);
16322 SDValue FPclassMask = getScalarMaskingNode(FPclass, Mask,
16323 DAG.getTargetConstant(0, dl, MVT::i1), Subtarget, DAG);
16324 return DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::i8, FPclassMask);
16327 case CMP_MASK_CC: {
16328 // Comparison intrinsics with masks.
16329 // Example of transformation:
16330 // (i8 (int_x86_avx512_mask_pcmpeq_q_128
16331 // (v2i64 %a), (v2i64 %b), (i8 %mask))) ->
16333 // (v8i1 (insert_subvector undef,
16334 // (v2i1 (and (PCMPEQM %a, %b),
16335 // (extract_subvector
16336 // (v8i1 (bitcast %mask)), 0))), 0))))
16337 EVT VT = Op.getOperand(1).getValueType();
16338 EVT MaskVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
16339 VT.getVectorNumElements());
16340 SDValue Mask = Op.getOperand((IntrData->Type == CMP_MASK_CC) ? 4 : 3);
16341 EVT BitcastVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
16342 Mask.getValueType().getSizeInBits());
16344 if (IntrData->Type == CMP_MASK_CC) {
16345 SDValue CC = Op.getOperand(3);
16346 CC = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, CC);
16347 // We specify 2 possible opcodes for intrinsics with rounding modes.
16348 // First, we check if the intrinsic may have non-default rounding mode,
16349 // (IntrData->Opc1 != 0), then we check the rounding mode operand.
16350 if (IntrData->Opc1 != 0) {
16351 SDValue Rnd = Op.getOperand(5);
16352 if (cast<ConstantSDNode>(Rnd)->getZExtValue() !=
16353 X86::STATIC_ROUNDING::CUR_DIRECTION)
16354 Cmp = DAG.getNode(IntrData->Opc1, dl, MaskVT, Op.getOperand(1),
16355 Op.getOperand(2), CC, Rnd);
16357 //default rounding mode
16359 Cmp = DAG.getNode(IntrData->Opc0, dl, MaskVT, Op.getOperand(1),
16360 Op.getOperand(2), CC);
16363 assert(IntrData->Type == CMP_MASK && "Unexpected intrinsic type!");
16364 Cmp = DAG.getNode(IntrData->Opc0, dl, MaskVT, Op.getOperand(1),
16367 SDValue CmpMask = getVectorMaskingNode(Cmp, Mask,
16368 DAG.getTargetConstant(0, dl,
16371 SDValue Res = DAG.getNode(ISD::INSERT_SUBVECTOR, dl, BitcastVT,
16372 DAG.getUNDEF(BitcastVT), CmpMask,
16373 DAG.getIntPtrConstant(0, dl));
16374 return DAG.getBitcast(Op.getValueType(), Res);
16376 case CMP_MASK_SCALAR_CC: {
16377 SDValue Src1 = Op.getOperand(1);
16378 SDValue Src2 = Op.getOperand(2);
16379 SDValue CC = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op.getOperand(3));
16380 SDValue Mask = Op.getOperand(4);
16383 if (IntrData->Opc1 != 0) {
16384 SDValue Rnd = Op.getOperand(5);
16385 if (cast<ConstantSDNode>(Rnd)->getZExtValue() !=
16386 X86::STATIC_ROUNDING::CUR_DIRECTION)
16387 Cmp = DAG.getNode(IntrData->Opc1, dl, MVT::i1, Src1, Src2, CC, Rnd);
16389 //default rounding mode
16391 Cmp = DAG.getNode(IntrData->Opc0, dl, MVT::i1, Src1, Src2, CC);
16393 SDValue CmpMask = getScalarMaskingNode(Cmp, Mask,
16394 DAG.getTargetConstant(0, dl,
16398 return DAG.getNode(ISD::SIGN_EXTEND_INREG, dl, MVT::i8,
16399 DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i8, CmpMask),
16400 DAG.getValueType(MVT::i1));
16402 case COMI: { // Comparison intrinsics
16403 ISD::CondCode CC = (ISD::CondCode)IntrData->Opc1;
16404 SDValue LHS = Op.getOperand(1);
16405 SDValue RHS = Op.getOperand(2);
16406 unsigned X86CC = TranslateX86CC(CC, dl, true, LHS, RHS, DAG);
16407 assert(X86CC != X86::COND_INVALID && "Unexpected illegal condition!");
16408 SDValue Cond = DAG.getNode(IntrData->Opc0, dl, MVT::i32, LHS, RHS);
16409 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
16410 DAG.getConstant(X86CC, dl, MVT::i8), Cond);
16411 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
16414 return getTargetVShiftNode(IntrData->Opc0, dl, Op.getSimpleValueType(),
16415 Op.getOperand(1), Op.getOperand(2), DAG);
16417 return getVectorMaskingNode(getTargetVShiftNode(IntrData->Opc0, dl,
16418 Op.getSimpleValueType(),
16420 Op.getOperand(2), DAG),
16421 Op.getOperand(4), Op.getOperand(3), Subtarget,
16423 case COMPRESS_EXPAND_IN_REG: {
16424 SDValue Mask = Op.getOperand(3);
16425 SDValue DataToCompress = Op.getOperand(1);
16426 SDValue PassThru = Op.getOperand(2);
16427 if (isAllOnes(Mask)) // return data as is
16428 return Op.getOperand(1);
16430 return getVectorMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT,
16432 Mask, PassThru, Subtarget, DAG);
16435 SDValue Mask = Op.getOperand(3);
16436 EVT VT = Op.getValueType();
16437 EVT MaskVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
16438 VT.getVectorNumElements());
16439 EVT BitcastVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
16440 Mask.getValueType().getSizeInBits());
16442 SDValue VMask = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MaskVT,
16443 DAG.getBitcast(BitcastVT, Mask),
16444 DAG.getIntPtrConstant(0, dl));
16445 return DAG.getNode(IntrData->Opc0, dl, VT, VMask, Op.getOperand(1),
16454 default: return SDValue(); // Don't custom lower most intrinsics.
16456 case Intrinsic::x86_avx2_permd:
16457 case Intrinsic::x86_avx2_permps:
16458 // Operands intentionally swapped. Mask is last operand to intrinsic,
16459 // but second operand for node/instruction.
16460 return DAG.getNode(X86ISD::VPERMV, dl, Op.getValueType(),
16461 Op.getOperand(2), Op.getOperand(1));
16463 // ptest and testp intrinsics. The intrinsic these come from are designed to
16464 // return an integer value, not just an instruction so lower it to the ptest
16465 // or testp pattern and a setcc for the result.
16466 case Intrinsic::x86_sse41_ptestz:
16467 case Intrinsic::x86_sse41_ptestc:
16468 case Intrinsic::x86_sse41_ptestnzc:
16469 case Intrinsic::x86_avx_ptestz_256:
16470 case Intrinsic::x86_avx_ptestc_256:
16471 case Intrinsic::x86_avx_ptestnzc_256:
16472 case Intrinsic::x86_avx_vtestz_ps:
16473 case Intrinsic::x86_avx_vtestc_ps:
16474 case Intrinsic::x86_avx_vtestnzc_ps:
16475 case Intrinsic::x86_avx_vtestz_pd:
16476 case Intrinsic::x86_avx_vtestc_pd:
16477 case Intrinsic::x86_avx_vtestnzc_pd:
16478 case Intrinsic::x86_avx_vtestz_ps_256:
16479 case Intrinsic::x86_avx_vtestc_ps_256:
16480 case Intrinsic::x86_avx_vtestnzc_ps_256:
16481 case Intrinsic::x86_avx_vtestz_pd_256:
16482 case Intrinsic::x86_avx_vtestc_pd_256:
16483 case Intrinsic::x86_avx_vtestnzc_pd_256: {
16484 bool IsTestPacked = false;
16487 default: llvm_unreachable("Bad fallthrough in Intrinsic lowering.");
16488 case Intrinsic::x86_avx_vtestz_ps:
16489 case Intrinsic::x86_avx_vtestz_pd:
16490 case Intrinsic::x86_avx_vtestz_ps_256:
16491 case Intrinsic::x86_avx_vtestz_pd_256:
16492 IsTestPacked = true; // Fallthrough
16493 case Intrinsic::x86_sse41_ptestz:
16494 case Intrinsic::x86_avx_ptestz_256:
16496 X86CC = X86::COND_E;
16498 case Intrinsic::x86_avx_vtestc_ps:
16499 case Intrinsic::x86_avx_vtestc_pd:
16500 case Intrinsic::x86_avx_vtestc_ps_256:
16501 case Intrinsic::x86_avx_vtestc_pd_256:
16502 IsTestPacked = true; // Fallthrough
16503 case Intrinsic::x86_sse41_ptestc:
16504 case Intrinsic::x86_avx_ptestc_256:
16506 X86CC = X86::COND_B;
16508 case Intrinsic::x86_avx_vtestnzc_ps:
16509 case Intrinsic::x86_avx_vtestnzc_pd:
16510 case Intrinsic::x86_avx_vtestnzc_ps_256:
16511 case Intrinsic::x86_avx_vtestnzc_pd_256:
16512 IsTestPacked = true; // Fallthrough
16513 case Intrinsic::x86_sse41_ptestnzc:
16514 case Intrinsic::x86_avx_ptestnzc_256:
16516 X86CC = X86::COND_A;
16520 SDValue LHS = Op.getOperand(1);
16521 SDValue RHS = Op.getOperand(2);
16522 unsigned TestOpc = IsTestPacked ? X86ISD::TESTP : X86ISD::PTEST;
16523 SDValue Test = DAG.getNode(TestOpc, dl, MVT::i32, LHS, RHS);
16524 SDValue CC = DAG.getConstant(X86CC, dl, MVT::i8);
16525 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8, CC, Test);
16526 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
16528 case Intrinsic::x86_avx512_kortestz_w:
16529 case Intrinsic::x86_avx512_kortestc_w: {
16530 unsigned X86CC = (IntNo == Intrinsic::x86_avx512_kortestz_w)? X86::COND_E: X86::COND_B;
16531 SDValue LHS = DAG.getBitcast(MVT::v16i1, Op.getOperand(1));
16532 SDValue RHS = DAG.getBitcast(MVT::v16i1, Op.getOperand(2));
16533 SDValue CC = DAG.getConstant(X86CC, dl, MVT::i8);
16534 SDValue Test = DAG.getNode(X86ISD::KORTEST, dl, MVT::i32, LHS, RHS);
16535 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i1, CC, Test);
16536 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
16539 case Intrinsic::x86_sse42_pcmpistria128:
16540 case Intrinsic::x86_sse42_pcmpestria128:
16541 case Intrinsic::x86_sse42_pcmpistric128:
16542 case Intrinsic::x86_sse42_pcmpestric128:
16543 case Intrinsic::x86_sse42_pcmpistrio128:
16544 case Intrinsic::x86_sse42_pcmpestrio128:
16545 case Intrinsic::x86_sse42_pcmpistris128:
16546 case Intrinsic::x86_sse42_pcmpestris128:
16547 case Intrinsic::x86_sse42_pcmpistriz128:
16548 case Intrinsic::x86_sse42_pcmpestriz128: {
16552 default: llvm_unreachable("Impossible intrinsic"); // Can't reach here.
16553 case Intrinsic::x86_sse42_pcmpistria128:
16554 Opcode = X86ISD::PCMPISTRI;
16555 X86CC = X86::COND_A;
16557 case Intrinsic::x86_sse42_pcmpestria128:
16558 Opcode = X86ISD::PCMPESTRI;
16559 X86CC = X86::COND_A;
16561 case Intrinsic::x86_sse42_pcmpistric128:
16562 Opcode = X86ISD::PCMPISTRI;
16563 X86CC = X86::COND_B;
16565 case Intrinsic::x86_sse42_pcmpestric128:
16566 Opcode = X86ISD::PCMPESTRI;
16567 X86CC = X86::COND_B;
16569 case Intrinsic::x86_sse42_pcmpistrio128:
16570 Opcode = X86ISD::PCMPISTRI;
16571 X86CC = X86::COND_O;
16573 case Intrinsic::x86_sse42_pcmpestrio128:
16574 Opcode = X86ISD::PCMPESTRI;
16575 X86CC = X86::COND_O;
16577 case Intrinsic::x86_sse42_pcmpistris128:
16578 Opcode = X86ISD::PCMPISTRI;
16579 X86CC = X86::COND_S;
16581 case Intrinsic::x86_sse42_pcmpestris128:
16582 Opcode = X86ISD::PCMPESTRI;
16583 X86CC = X86::COND_S;
16585 case Intrinsic::x86_sse42_pcmpistriz128:
16586 Opcode = X86ISD::PCMPISTRI;
16587 X86CC = X86::COND_E;
16589 case Intrinsic::x86_sse42_pcmpestriz128:
16590 Opcode = X86ISD::PCMPESTRI;
16591 X86CC = X86::COND_E;
16594 SmallVector<SDValue, 5> NewOps(Op->op_begin()+1, Op->op_end());
16595 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::i32);
16596 SDValue PCMP = DAG.getNode(Opcode, dl, VTs, NewOps);
16597 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
16598 DAG.getConstant(X86CC, dl, MVT::i8),
16599 SDValue(PCMP.getNode(), 1));
16600 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
16603 case Intrinsic::x86_sse42_pcmpistri128:
16604 case Intrinsic::x86_sse42_pcmpestri128: {
16606 if (IntNo == Intrinsic::x86_sse42_pcmpistri128)
16607 Opcode = X86ISD::PCMPISTRI;
16609 Opcode = X86ISD::PCMPESTRI;
16611 SmallVector<SDValue, 5> NewOps(Op->op_begin()+1, Op->op_end());
16612 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::i32);
16613 return DAG.getNode(Opcode, dl, VTs, NewOps);
16616 case Intrinsic::x86_seh_lsda: {
16617 // Compute the symbol for the LSDA. We know it'll get emitted later.
16618 MachineFunction &MF = DAG.getMachineFunction();
16619 SDValue Op1 = Op.getOperand(1);
16620 auto *Fn = cast<Function>(cast<GlobalAddressSDNode>(Op1)->getGlobal());
16621 MCSymbol *LSDASym = MF.getMMI().getContext().getOrCreateLSDASymbol(
16622 GlobalValue::getRealLinkageName(Fn->getName()));
16624 // Generate a simple absolute symbol reference. This intrinsic is only
16625 // supported on 32-bit Windows, which isn't PIC.
16626 SDValue Result = DAG.getMCSymbol(LSDASym, VT);
16627 return DAG.getNode(X86ISD::Wrapper, dl, VT, Result);
16630 case Intrinsic::x86_seh_recoverfp: {
16631 SDValue FnOp = Op.getOperand(1);
16632 SDValue IncomingFPOp = Op.getOperand(2);
16633 GlobalAddressSDNode *GSD = dyn_cast<GlobalAddressSDNode>(FnOp);
16634 auto *Fn = dyn_cast_or_null<Function>(GSD ? GSD->getGlobal() : nullptr);
16636 report_fatal_error(
16637 "llvm.x86.seh.recoverfp must take a function as the first argument");
16638 return recoverFramePointer(DAG, Fn, IncomingFPOp);
16641 case Intrinsic::localaddress: {
16642 // Returns one of the stack, base, or frame pointer registers, depending on
16643 // which is used to reference local variables.
16644 MachineFunction &MF = DAG.getMachineFunction();
16645 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
16647 if (RegInfo->hasBasePointer(MF))
16648 Reg = RegInfo->getBaseRegister();
16649 else // This function handles the SP or FP case.
16650 Reg = RegInfo->getPtrSizedFrameRegister(MF);
16651 return DAG.getCopyFromReg(DAG.getEntryNode(), dl, Reg, VT);
16656 static SDValue getGatherNode(unsigned Opc, SDValue Op, SelectionDAG &DAG,
16657 SDValue Src, SDValue Mask, SDValue Base,
16658 SDValue Index, SDValue ScaleOp, SDValue Chain,
16659 const X86Subtarget * Subtarget) {
16661 ConstantSDNode *C = dyn_cast<ConstantSDNode>(ScaleOp);
16663 llvm_unreachable("Invalid scale type");
16664 unsigned ScaleVal = C->getZExtValue();
16665 if (ScaleVal > 2 && ScaleVal != 4 && ScaleVal != 8)
16666 llvm_unreachable("Valid scale values are 1, 2, 4, 8");
16668 SDValue Scale = DAG.getTargetConstant(C->getZExtValue(), dl, MVT::i8);
16669 EVT MaskVT = MVT::getVectorVT(MVT::i1,
16670 Index.getSimpleValueType().getVectorNumElements());
16672 ConstantSDNode *MaskC = dyn_cast<ConstantSDNode>(Mask);
16674 MaskInReg = DAG.getTargetConstant(MaskC->getSExtValue(), dl, MaskVT);
16676 EVT BitcastVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
16677 Mask.getValueType().getSizeInBits());
16679 // In case when MaskVT equals v2i1 or v4i1, low 2 or 4 elements
16680 // are extracted by EXTRACT_SUBVECTOR.
16681 MaskInReg = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MaskVT,
16682 DAG.getBitcast(BitcastVT, Mask),
16683 DAG.getIntPtrConstant(0, dl));
16685 SDVTList VTs = DAG.getVTList(Op.getValueType(), MaskVT, MVT::Other);
16686 SDValue Disp = DAG.getTargetConstant(0, dl, MVT::i32);
16687 SDValue Segment = DAG.getRegister(0, MVT::i32);
16688 if (Src.getOpcode() == ISD::UNDEF)
16689 Src = getZeroVector(Op.getValueType(), Subtarget, DAG, dl);
16690 SDValue Ops[] = {Src, MaskInReg, Base, Scale, Index, Disp, Segment, Chain};
16691 SDNode *Res = DAG.getMachineNode(Opc, dl, VTs, Ops);
16692 SDValue RetOps[] = { SDValue(Res, 0), SDValue(Res, 2) };
16693 return DAG.getMergeValues(RetOps, dl);
16696 static SDValue getScatterNode(unsigned Opc, SDValue Op, SelectionDAG &DAG,
16697 SDValue Src, SDValue Mask, SDValue Base,
16698 SDValue Index, SDValue ScaleOp, SDValue Chain) {
16700 ConstantSDNode *C = dyn_cast<ConstantSDNode>(ScaleOp);
16702 llvm_unreachable("Invalid scale type");
16703 unsigned ScaleVal = C->getZExtValue();
16704 if (ScaleVal > 2 && ScaleVal != 4 && ScaleVal != 8)
16705 llvm_unreachable("Valid scale values are 1, 2, 4, 8");
16707 SDValue Scale = DAG.getTargetConstant(C->getZExtValue(), dl, MVT::i8);
16708 SDValue Disp = DAG.getTargetConstant(0, dl, MVT::i32);
16709 SDValue Segment = DAG.getRegister(0, MVT::i32);
16710 EVT MaskVT = MVT::getVectorVT(MVT::i1,
16711 Index.getSimpleValueType().getVectorNumElements());
16713 ConstantSDNode *MaskC = dyn_cast<ConstantSDNode>(Mask);
16715 MaskInReg = DAG.getTargetConstant(MaskC->getSExtValue(), dl, MaskVT);
16717 EVT BitcastVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
16718 Mask.getValueType().getSizeInBits());
16720 // In case when MaskVT equals v2i1 or v4i1, low 2 or 4 elements
16721 // are extracted by EXTRACT_SUBVECTOR.
16722 MaskInReg = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MaskVT,
16723 DAG.getBitcast(BitcastVT, Mask),
16724 DAG.getIntPtrConstant(0, dl));
16726 SDVTList VTs = DAG.getVTList(MaskVT, MVT::Other);
16727 SDValue Ops[] = {Base, Scale, Index, Disp, Segment, MaskInReg, Src, Chain};
16728 SDNode *Res = DAG.getMachineNode(Opc, dl, VTs, Ops);
16729 return SDValue(Res, 1);
16732 static SDValue getPrefetchNode(unsigned Opc, SDValue Op, SelectionDAG &DAG,
16733 SDValue Mask, SDValue Base, SDValue Index,
16734 SDValue ScaleOp, SDValue Chain) {
16736 ConstantSDNode *C = dyn_cast<ConstantSDNode>(ScaleOp);
16737 assert(C && "Invalid scale type");
16738 SDValue Scale = DAG.getTargetConstant(C->getZExtValue(), dl, MVT::i8);
16739 SDValue Disp = DAG.getTargetConstant(0, dl, MVT::i32);
16740 SDValue Segment = DAG.getRegister(0, MVT::i32);
16742 MVT::getVectorVT(MVT::i1, Index.getSimpleValueType().getVectorNumElements());
16744 ConstantSDNode *MaskC = dyn_cast<ConstantSDNode>(Mask);
16746 MaskInReg = DAG.getTargetConstant(MaskC->getSExtValue(), dl, MaskVT);
16748 MaskInReg = DAG.getBitcast(MaskVT, Mask);
16749 //SDVTList VTs = DAG.getVTList(MVT::Other);
16750 SDValue Ops[] = {MaskInReg, Base, Scale, Index, Disp, Segment, Chain};
16751 SDNode *Res = DAG.getMachineNode(Opc, dl, MVT::Other, Ops);
16752 return SDValue(Res, 0);
16755 // getReadPerformanceCounter - Handles the lowering of builtin intrinsics that
16756 // read performance monitor counters (x86_rdpmc).
16757 static void getReadPerformanceCounter(SDNode *N, SDLoc DL,
16758 SelectionDAG &DAG, const X86Subtarget *Subtarget,
16759 SmallVectorImpl<SDValue> &Results) {
16760 assert(N->getNumOperands() == 3 && "Unexpected number of operands!");
16761 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Glue);
16764 // The ECX register is used to select the index of the performance counter
16766 SDValue Chain = DAG.getCopyToReg(N->getOperand(0), DL, X86::ECX,
16768 SDValue rd = DAG.getNode(X86ISD::RDPMC_DAG, DL, Tys, Chain);
16770 // Reads the content of a 64-bit performance counter and returns it in the
16771 // registers EDX:EAX.
16772 if (Subtarget->is64Bit()) {
16773 LO = DAG.getCopyFromReg(rd, DL, X86::RAX, MVT::i64, rd.getValue(1));
16774 HI = DAG.getCopyFromReg(LO.getValue(1), DL, X86::RDX, MVT::i64,
16777 LO = DAG.getCopyFromReg(rd, DL, X86::EAX, MVT::i32, rd.getValue(1));
16778 HI = DAG.getCopyFromReg(LO.getValue(1), DL, X86::EDX, MVT::i32,
16781 Chain = HI.getValue(1);
16783 if (Subtarget->is64Bit()) {
16784 // The EAX register is loaded with the low-order 32 bits. The EDX register
16785 // is loaded with the supported high-order bits of the counter.
16786 SDValue Tmp = DAG.getNode(ISD::SHL, DL, MVT::i64, HI,
16787 DAG.getConstant(32, DL, MVT::i8));
16788 Results.push_back(DAG.getNode(ISD::OR, DL, MVT::i64, LO, Tmp));
16789 Results.push_back(Chain);
16793 // Use a buildpair to merge the two 32-bit values into a 64-bit one.
16794 SDValue Ops[] = { LO, HI };
16795 SDValue Pair = DAG.getNode(ISD::BUILD_PAIR, DL, MVT::i64, Ops);
16796 Results.push_back(Pair);
16797 Results.push_back(Chain);
16800 // getReadTimeStampCounter - Handles the lowering of builtin intrinsics that
16801 // read the time stamp counter (x86_rdtsc and x86_rdtscp). This function is
16802 // also used to custom lower READCYCLECOUNTER nodes.
16803 static void getReadTimeStampCounter(SDNode *N, SDLoc DL, unsigned Opcode,
16804 SelectionDAG &DAG, const X86Subtarget *Subtarget,
16805 SmallVectorImpl<SDValue> &Results) {
16806 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Glue);
16807 SDValue rd = DAG.getNode(Opcode, DL, Tys, N->getOperand(0));
16810 // The processor's time-stamp counter (a 64-bit MSR) is stored into the
16811 // EDX:EAX registers. EDX is loaded with the high-order 32 bits of the MSR
16812 // and the EAX register is loaded with the low-order 32 bits.
16813 if (Subtarget->is64Bit()) {
16814 LO = DAG.getCopyFromReg(rd, DL, X86::RAX, MVT::i64, rd.getValue(1));
16815 HI = DAG.getCopyFromReg(LO.getValue(1), DL, X86::RDX, MVT::i64,
16818 LO = DAG.getCopyFromReg(rd, DL, X86::EAX, MVT::i32, rd.getValue(1));
16819 HI = DAG.getCopyFromReg(LO.getValue(1), DL, X86::EDX, MVT::i32,
16822 SDValue Chain = HI.getValue(1);
16824 if (Opcode == X86ISD::RDTSCP_DAG) {
16825 assert(N->getNumOperands() == 3 && "Unexpected number of operands!");
16827 // Instruction RDTSCP loads the IA32:TSC_AUX_MSR (address C000_0103H) into
16828 // the ECX register. Add 'ecx' explicitly to the chain.
16829 SDValue ecx = DAG.getCopyFromReg(Chain, DL, X86::ECX, MVT::i32,
16831 // Explicitly store the content of ECX at the location passed in input
16832 // to the 'rdtscp' intrinsic.
16833 Chain = DAG.getStore(ecx.getValue(1), DL, ecx, N->getOperand(2),
16834 MachinePointerInfo(), false, false, 0);
16837 if (Subtarget->is64Bit()) {
16838 // The EDX register is loaded with the high-order 32 bits of the MSR, and
16839 // the EAX register is loaded with the low-order 32 bits.
16840 SDValue Tmp = DAG.getNode(ISD::SHL, DL, MVT::i64, HI,
16841 DAG.getConstant(32, DL, MVT::i8));
16842 Results.push_back(DAG.getNode(ISD::OR, DL, MVT::i64, LO, Tmp));
16843 Results.push_back(Chain);
16847 // Use a buildpair to merge the two 32-bit values into a 64-bit one.
16848 SDValue Ops[] = { LO, HI };
16849 SDValue Pair = DAG.getNode(ISD::BUILD_PAIR, DL, MVT::i64, Ops);
16850 Results.push_back(Pair);
16851 Results.push_back(Chain);
16854 static SDValue LowerREADCYCLECOUNTER(SDValue Op, const X86Subtarget *Subtarget,
16855 SelectionDAG &DAG) {
16856 SmallVector<SDValue, 2> Results;
16858 getReadTimeStampCounter(Op.getNode(), DL, X86ISD::RDTSC_DAG, DAG, Subtarget,
16860 return DAG.getMergeValues(Results, DL);
16863 static SDValue LowerSEHRESTOREFRAME(SDValue Op, const X86Subtarget *Subtarget,
16864 SelectionDAG &DAG) {
16865 MachineFunction &MF = DAG.getMachineFunction();
16866 const Function *Fn = MF.getFunction();
16868 SDValue Chain = Op.getOperand(0);
16870 assert(Subtarget->getFrameLowering()->hasFP(MF) &&
16871 "using llvm.x86.seh.restoreframe requires a frame pointer");
16873 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
16874 MVT VT = TLI.getPointerTy(DAG.getDataLayout());
16876 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
16877 unsigned FrameReg =
16878 RegInfo->getPtrSizedFrameRegister(DAG.getMachineFunction());
16879 unsigned SPReg = RegInfo->getStackRegister();
16880 unsigned SlotSize = RegInfo->getSlotSize();
16882 // Get incoming EBP.
16883 SDValue IncomingEBP =
16884 DAG.getCopyFromReg(Chain, dl, FrameReg, VT);
16886 // SP is saved in the first field of every registration node, so load
16887 // [EBP-RegNodeSize] into SP.
16888 int RegNodeSize = getSEHRegistrationNodeSize(Fn);
16889 SDValue SPAddr = DAG.getNode(ISD::ADD, dl, VT, IncomingEBP,
16890 DAG.getConstant(-RegNodeSize, dl, VT));
16892 DAG.getLoad(VT, dl, Chain, SPAddr, MachinePointerInfo(), false, false,
16893 false, VT.getScalarSizeInBits() / 8);
16894 Chain = DAG.getCopyToReg(Chain, dl, SPReg, NewSP);
16896 if (!RegInfo->needsStackRealignment(MF)) {
16897 // Adjust EBP to point back to the original frame position.
16898 SDValue NewFP = recoverFramePointer(DAG, Fn, IncomingEBP);
16899 Chain = DAG.getCopyToReg(Chain, dl, FrameReg, NewFP);
16901 assert(RegInfo->hasBasePointer(MF) &&
16902 "functions with Win32 EH must use frame or base pointer register");
16904 // Reload the base pointer (ESI) with the adjusted incoming EBP.
16905 SDValue NewBP = recoverFramePointer(DAG, Fn, IncomingEBP);
16906 Chain = DAG.getCopyToReg(Chain, dl, RegInfo->getBaseRegister(), NewBP);
16908 // Reload the spilled EBP value, now that the stack and base pointers are
16910 X86MachineFunctionInfo *X86FI = MF.getInfo<X86MachineFunctionInfo>();
16911 X86FI->setHasSEHFramePtrSave(true);
16912 int FI = MF.getFrameInfo()->CreateSpillStackObject(SlotSize, SlotSize);
16913 X86FI->setSEHFramePtrSaveIndex(FI);
16914 SDValue NewFP = DAG.getLoad(VT, dl, Chain, DAG.getFrameIndex(FI, VT),
16915 MachinePointerInfo(), false, false, false,
16916 VT.getScalarSizeInBits() / 8);
16917 Chain = DAG.getCopyToReg(NewFP, dl, FrameReg, NewFP);
16923 /// \brief Lower intrinsics for TRUNCATE_TO_MEM case
16924 /// return truncate Store/MaskedStore Node
16925 static SDValue LowerINTRINSIC_TRUNCATE_TO_MEM(const SDValue & Op,
16929 SDValue Mask = Op.getOperand(4);
16930 SDValue DataToTruncate = Op.getOperand(3);
16931 SDValue Addr = Op.getOperand(2);
16932 SDValue Chain = Op.getOperand(0);
16934 EVT VT = DataToTruncate.getValueType();
16935 EVT SVT = EVT::getVectorVT(*DAG.getContext(),
16936 ElementType, VT.getVectorNumElements());
16938 if (isAllOnes(Mask)) // return just a truncate store
16939 return DAG.getTruncStore(Chain, dl, DataToTruncate, Addr,
16940 MachinePointerInfo(), SVT, false, false,
16941 SVT.getScalarSizeInBits()/8);
16943 EVT MaskVT = EVT::getVectorVT(*DAG.getContext(),
16944 MVT::i1, VT.getVectorNumElements());
16945 EVT BitcastVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
16946 Mask.getValueType().getSizeInBits());
16947 // In case when MaskVT equals v2i1 or v4i1, low 2 or 4 elements
16948 // are extracted by EXTRACT_SUBVECTOR.
16949 SDValue VMask = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MaskVT,
16950 DAG.getBitcast(BitcastVT, Mask),
16951 DAG.getIntPtrConstant(0, dl));
16953 MachineMemOperand *MMO = DAG.getMachineFunction().
16954 getMachineMemOperand(MachinePointerInfo(),
16955 MachineMemOperand::MOStore, SVT.getStoreSize(),
16956 SVT.getScalarSizeInBits()/8);
16958 return DAG.getMaskedStore(Chain, dl, DataToTruncate, Addr,
16959 VMask, SVT, MMO, true);
16962 static SDValue LowerINTRINSIC_W_CHAIN(SDValue Op, const X86Subtarget *Subtarget,
16963 SelectionDAG &DAG) {
16964 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
16966 const IntrinsicData* IntrData = getIntrinsicWithChain(IntNo);
16968 if (IntNo == llvm::Intrinsic::x86_seh_restoreframe)
16969 return LowerSEHRESTOREFRAME(Op, Subtarget, DAG);
16974 switch(IntrData->Type) {
16976 llvm_unreachable("Unknown Intrinsic Type");
16980 // Emit the node with the right value type.
16981 SDVTList VTs = DAG.getVTList(Op->getValueType(0), MVT::Glue, MVT::Other);
16982 SDValue Result = DAG.getNode(IntrData->Opc0, dl, VTs, Op.getOperand(0));
16984 // If the value returned by RDRAND/RDSEED was valid (CF=1), return 1.
16985 // Otherwise return the value from Rand, which is always 0, casted to i32.
16986 SDValue Ops[] = { DAG.getZExtOrTrunc(Result, dl, Op->getValueType(1)),
16987 DAG.getConstant(1, dl, Op->getValueType(1)),
16988 DAG.getConstant(X86::COND_B, dl, MVT::i32),
16989 SDValue(Result.getNode(), 1) };
16990 SDValue isValid = DAG.getNode(X86ISD::CMOV, dl,
16991 DAG.getVTList(Op->getValueType(1), MVT::Glue),
16994 // Return { result, isValid, chain }.
16995 return DAG.getNode(ISD::MERGE_VALUES, dl, Op->getVTList(), Result, isValid,
16996 SDValue(Result.getNode(), 2));
16999 //gather(v1, mask, index, base, scale);
17000 SDValue Chain = Op.getOperand(0);
17001 SDValue Src = Op.getOperand(2);
17002 SDValue Base = Op.getOperand(3);
17003 SDValue Index = Op.getOperand(4);
17004 SDValue Mask = Op.getOperand(5);
17005 SDValue Scale = Op.getOperand(6);
17006 return getGatherNode(IntrData->Opc0, Op, DAG, Src, Mask, Base, Index, Scale,
17010 //scatter(base, mask, index, v1, scale);
17011 SDValue Chain = Op.getOperand(0);
17012 SDValue Base = Op.getOperand(2);
17013 SDValue Mask = Op.getOperand(3);
17014 SDValue Index = Op.getOperand(4);
17015 SDValue Src = Op.getOperand(5);
17016 SDValue Scale = Op.getOperand(6);
17017 return getScatterNode(IntrData->Opc0, Op, DAG, Src, Mask, Base, Index,
17021 SDValue Hint = Op.getOperand(6);
17022 unsigned HintVal = cast<ConstantSDNode>(Hint)->getZExtValue();
17023 assert(HintVal < 2 && "Wrong prefetch hint in intrinsic: should be 0 or 1");
17024 unsigned Opcode = (HintVal ? IntrData->Opc1 : IntrData->Opc0);
17025 SDValue Chain = Op.getOperand(0);
17026 SDValue Mask = Op.getOperand(2);
17027 SDValue Index = Op.getOperand(3);
17028 SDValue Base = Op.getOperand(4);
17029 SDValue Scale = Op.getOperand(5);
17030 return getPrefetchNode(Opcode, Op, DAG, Mask, Base, Index, Scale, Chain);
17032 // Read Time Stamp Counter (RDTSC) and Processor ID (RDTSCP).
17034 SmallVector<SDValue, 2> Results;
17035 getReadTimeStampCounter(Op.getNode(), dl, IntrData->Opc0, DAG, Subtarget,
17037 return DAG.getMergeValues(Results, dl);
17039 // Read Performance Monitoring Counters.
17041 SmallVector<SDValue, 2> Results;
17042 getReadPerformanceCounter(Op.getNode(), dl, DAG, Subtarget, Results);
17043 return DAG.getMergeValues(Results, dl);
17045 // XTEST intrinsics.
17047 SDVTList VTs = DAG.getVTList(Op->getValueType(0), MVT::Other);
17048 SDValue InTrans = DAG.getNode(IntrData->Opc0, dl, VTs, Op.getOperand(0));
17049 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
17050 DAG.getConstant(X86::COND_NE, dl, MVT::i8),
17052 SDValue Ret = DAG.getNode(ISD::ZERO_EXTEND, dl, Op->getValueType(0), SetCC);
17053 return DAG.getNode(ISD::MERGE_VALUES, dl, Op->getVTList(),
17054 Ret, SDValue(InTrans.getNode(), 1));
17058 SmallVector<SDValue, 2> Results;
17059 SDVTList CFVTs = DAG.getVTList(Op->getValueType(0), MVT::Other);
17060 SDVTList VTs = DAG.getVTList(Op.getOperand(3)->getValueType(0), MVT::Other);
17061 SDValue GenCF = DAG.getNode(X86ISD::ADD, dl, CFVTs, Op.getOperand(2),
17062 DAG.getConstant(-1, dl, MVT::i8));
17063 SDValue Res = DAG.getNode(IntrData->Opc0, dl, VTs, Op.getOperand(3),
17064 Op.getOperand(4), GenCF.getValue(1));
17065 SDValue Store = DAG.getStore(Op.getOperand(0), dl, Res.getValue(0),
17066 Op.getOperand(5), MachinePointerInfo(),
17068 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
17069 DAG.getConstant(X86::COND_B, dl, MVT::i8),
17071 Results.push_back(SetCC);
17072 Results.push_back(Store);
17073 return DAG.getMergeValues(Results, dl);
17075 case COMPRESS_TO_MEM: {
17077 SDValue Mask = Op.getOperand(4);
17078 SDValue DataToCompress = Op.getOperand(3);
17079 SDValue Addr = Op.getOperand(2);
17080 SDValue Chain = Op.getOperand(0);
17082 EVT VT = DataToCompress.getValueType();
17083 if (isAllOnes(Mask)) // return just a store
17084 return DAG.getStore(Chain, dl, DataToCompress, Addr,
17085 MachinePointerInfo(), false, false,
17086 VT.getScalarSizeInBits()/8);
17088 SDValue Compressed =
17089 getVectorMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT, DataToCompress),
17090 Mask, DAG.getUNDEF(VT), Subtarget, DAG);
17091 return DAG.getStore(Chain, dl, Compressed, Addr,
17092 MachinePointerInfo(), false, false,
17093 VT.getScalarSizeInBits()/8);
17095 case TRUNCATE_TO_MEM_VI8:
17096 return LowerINTRINSIC_TRUNCATE_TO_MEM(Op, DAG, MVT::i8);
17097 case TRUNCATE_TO_MEM_VI16:
17098 return LowerINTRINSIC_TRUNCATE_TO_MEM(Op, DAG, MVT::i16);
17099 case TRUNCATE_TO_MEM_VI32:
17100 return LowerINTRINSIC_TRUNCATE_TO_MEM(Op, DAG, MVT::i32);
17101 case EXPAND_FROM_MEM: {
17103 SDValue Mask = Op.getOperand(4);
17104 SDValue PassThru = Op.getOperand(3);
17105 SDValue Addr = Op.getOperand(2);
17106 SDValue Chain = Op.getOperand(0);
17107 EVT VT = Op.getValueType();
17109 if (isAllOnes(Mask)) // return just a load
17110 return DAG.getLoad(VT, dl, Chain, Addr, MachinePointerInfo(), false, false,
17111 false, VT.getScalarSizeInBits()/8);
17113 SDValue DataToExpand = DAG.getLoad(VT, dl, Chain, Addr, MachinePointerInfo(),
17114 false, false, false,
17115 VT.getScalarSizeInBits()/8);
17117 SDValue Results[] = {
17118 getVectorMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT, DataToExpand),
17119 Mask, PassThru, Subtarget, DAG), Chain};
17120 return DAG.getMergeValues(Results, dl);
17125 SDValue X86TargetLowering::LowerRETURNADDR(SDValue Op,
17126 SelectionDAG &DAG) const {
17127 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
17128 MFI->setReturnAddressIsTaken(true);
17130 if (verifyReturnAddressArgumentIsConstant(Op, DAG))
17133 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
17135 EVT PtrVT = getPointerTy(DAG.getDataLayout());
17138 SDValue FrameAddr = LowerFRAMEADDR(Op, DAG);
17139 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
17140 SDValue Offset = DAG.getConstant(RegInfo->getSlotSize(), dl, PtrVT);
17141 return DAG.getLoad(PtrVT, dl, DAG.getEntryNode(),
17142 DAG.getNode(ISD::ADD, dl, PtrVT,
17143 FrameAddr, Offset),
17144 MachinePointerInfo(), false, false, false, 0);
17147 // Just load the return address.
17148 SDValue RetAddrFI = getReturnAddressFrameIndex(DAG);
17149 return DAG.getLoad(PtrVT, dl, DAG.getEntryNode(),
17150 RetAddrFI, MachinePointerInfo(), false, false, false, 0);
17153 SDValue X86TargetLowering::LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const {
17154 MachineFunction &MF = DAG.getMachineFunction();
17155 MachineFrameInfo *MFI = MF.getFrameInfo();
17156 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
17157 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
17158 EVT VT = Op.getValueType();
17160 MFI->setFrameAddressIsTaken(true);
17162 if (MF.getTarget().getMCAsmInfo()->usesWindowsCFI()) {
17163 // Depth > 0 makes no sense on targets which use Windows unwind codes. It
17164 // is not possible to crawl up the stack without looking at the unwind codes
17166 int FrameAddrIndex = FuncInfo->getFAIndex();
17167 if (!FrameAddrIndex) {
17168 // Set up a frame object for the return address.
17169 unsigned SlotSize = RegInfo->getSlotSize();
17170 FrameAddrIndex = MF.getFrameInfo()->CreateFixedObject(
17171 SlotSize, /*Offset=*/0, /*IsImmutable=*/false);
17172 FuncInfo->setFAIndex(FrameAddrIndex);
17174 return DAG.getFrameIndex(FrameAddrIndex, VT);
17177 unsigned FrameReg =
17178 RegInfo->getPtrSizedFrameRegister(DAG.getMachineFunction());
17179 SDLoc dl(Op); // FIXME probably not meaningful
17180 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
17181 assert(((FrameReg == X86::RBP && VT == MVT::i64) ||
17182 (FrameReg == X86::EBP && VT == MVT::i32)) &&
17183 "Invalid Frame Register!");
17184 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, VT);
17186 FrameAddr = DAG.getLoad(VT, dl, DAG.getEntryNode(), FrameAddr,
17187 MachinePointerInfo(),
17188 false, false, false, 0);
17192 // FIXME? Maybe this could be a TableGen attribute on some registers and
17193 // this table could be generated automatically from RegInfo.
17194 unsigned X86TargetLowering::getRegisterByName(const char* RegName, EVT VT,
17195 SelectionDAG &DAG) const {
17196 const TargetFrameLowering &TFI = *Subtarget->getFrameLowering();
17197 const MachineFunction &MF = DAG.getMachineFunction();
17199 unsigned Reg = StringSwitch<unsigned>(RegName)
17200 .Case("esp", X86::ESP)
17201 .Case("rsp", X86::RSP)
17202 .Case("ebp", X86::EBP)
17203 .Case("rbp", X86::RBP)
17206 if (Reg == X86::EBP || Reg == X86::RBP) {
17207 if (!TFI.hasFP(MF))
17208 report_fatal_error("register " + StringRef(RegName) +
17209 " is allocatable: function has no frame pointer");
17212 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
17213 unsigned FrameReg =
17214 RegInfo->getPtrSizedFrameRegister(DAG.getMachineFunction());
17215 assert((FrameReg == X86::EBP || FrameReg == X86::RBP) &&
17216 "Invalid Frame Register!");
17224 report_fatal_error("Invalid register name global variable");
17227 SDValue X86TargetLowering::LowerFRAME_TO_ARGS_OFFSET(SDValue Op,
17228 SelectionDAG &DAG) const {
17229 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
17230 return DAG.getIntPtrConstant(2 * RegInfo->getSlotSize(), SDLoc(Op));
17233 SDValue X86TargetLowering::LowerEH_RETURN(SDValue Op, SelectionDAG &DAG) const {
17234 SDValue Chain = Op.getOperand(0);
17235 SDValue Offset = Op.getOperand(1);
17236 SDValue Handler = Op.getOperand(2);
17239 EVT PtrVT = getPointerTy(DAG.getDataLayout());
17240 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
17241 unsigned FrameReg = RegInfo->getFrameRegister(DAG.getMachineFunction());
17242 assert(((FrameReg == X86::RBP && PtrVT == MVT::i64) ||
17243 (FrameReg == X86::EBP && PtrVT == MVT::i32)) &&
17244 "Invalid Frame Register!");
17245 SDValue Frame = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, PtrVT);
17246 unsigned StoreAddrReg = (PtrVT == MVT::i64) ? X86::RCX : X86::ECX;
17248 SDValue StoreAddr = DAG.getNode(ISD::ADD, dl, PtrVT, Frame,
17249 DAG.getIntPtrConstant(RegInfo->getSlotSize(),
17251 StoreAddr = DAG.getNode(ISD::ADD, dl, PtrVT, StoreAddr, Offset);
17252 Chain = DAG.getStore(Chain, dl, Handler, StoreAddr, MachinePointerInfo(),
17254 Chain = DAG.getCopyToReg(Chain, dl, StoreAddrReg, StoreAddr);
17256 return DAG.getNode(X86ISD::EH_RETURN, dl, MVT::Other, Chain,
17257 DAG.getRegister(StoreAddrReg, PtrVT));
17260 SDValue X86TargetLowering::lowerEH_SJLJ_SETJMP(SDValue Op,
17261 SelectionDAG &DAG) const {
17263 return DAG.getNode(X86ISD::EH_SJLJ_SETJMP, DL,
17264 DAG.getVTList(MVT::i32, MVT::Other),
17265 Op.getOperand(0), Op.getOperand(1));
17268 SDValue X86TargetLowering::lowerEH_SJLJ_LONGJMP(SDValue Op,
17269 SelectionDAG &DAG) const {
17271 return DAG.getNode(X86ISD::EH_SJLJ_LONGJMP, DL, MVT::Other,
17272 Op.getOperand(0), Op.getOperand(1));
17275 static SDValue LowerADJUST_TRAMPOLINE(SDValue Op, SelectionDAG &DAG) {
17276 return Op.getOperand(0);
17279 SDValue X86TargetLowering::LowerINIT_TRAMPOLINE(SDValue Op,
17280 SelectionDAG &DAG) const {
17281 SDValue Root = Op.getOperand(0);
17282 SDValue Trmp = Op.getOperand(1); // trampoline
17283 SDValue FPtr = Op.getOperand(2); // nested function
17284 SDValue Nest = Op.getOperand(3); // 'nest' parameter value
17287 const Value *TrmpAddr = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
17288 const TargetRegisterInfo *TRI = Subtarget->getRegisterInfo();
17290 if (Subtarget->is64Bit()) {
17291 SDValue OutChains[6];
17293 // Large code-model.
17294 const unsigned char JMP64r = 0xFF; // 64-bit jmp through register opcode.
17295 const unsigned char MOV64ri = 0xB8; // X86::MOV64ri opcode.
17297 const unsigned char N86R10 = TRI->getEncodingValue(X86::R10) & 0x7;
17298 const unsigned char N86R11 = TRI->getEncodingValue(X86::R11) & 0x7;
17300 const unsigned char REX_WB = 0x40 | 0x08 | 0x01; // REX prefix
17302 // Load the pointer to the nested function into R11.
17303 unsigned OpCode = ((MOV64ri | N86R11) << 8) | REX_WB; // movabsq r11
17304 SDValue Addr = Trmp;
17305 OutChains[0] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, dl, MVT::i16),
17306 Addr, MachinePointerInfo(TrmpAddr),
17309 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
17310 DAG.getConstant(2, dl, MVT::i64));
17311 OutChains[1] = DAG.getStore(Root, dl, FPtr, Addr,
17312 MachinePointerInfo(TrmpAddr, 2),
17315 // Load the 'nest' parameter value into R10.
17316 // R10 is specified in X86CallingConv.td
17317 OpCode = ((MOV64ri | N86R10) << 8) | REX_WB; // movabsq r10
17318 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
17319 DAG.getConstant(10, dl, MVT::i64));
17320 OutChains[2] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, dl, MVT::i16),
17321 Addr, MachinePointerInfo(TrmpAddr, 10),
17324 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
17325 DAG.getConstant(12, dl, MVT::i64));
17326 OutChains[3] = DAG.getStore(Root, dl, Nest, Addr,
17327 MachinePointerInfo(TrmpAddr, 12),
17330 // Jump to the nested function.
17331 OpCode = (JMP64r << 8) | REX_WB; // jmpq *...
17332 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
17333 DAG.getConstant(20, dl, MVT::i64));
17334 OutChains[4] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, dl, MVT::i16),
17335 Addr, MachinePointerInfo(TrmpAddr, 20),
17338 unsigned char ModRM = N86R11 | (4 << 3) | (3 << 6); // ...r11
17339 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
17340 DAG.getConstant(22, dl, MVT::i64));
17341 OutChains[5] = DAG.getStore(Root, dl, DAG.getConstant(ModRM, dl, MVT::i8),
17342 Addr, MachinePointerInfo(TrmpAddr, 22),
17345 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other, OutChains);
17347 const Function *Func =
17348 cast<Function>(cast<SrcValueSDNode>(Op.getOperand(5))->getValue());
17349 CallingConv::ID CC = Func->getCallingConv();
17354 llvm_unreachable("Unsupported calling convention");
17355 case CallingConv::C:
17356 case CallingConv::X86_StdCall: {
17357 // Pass 'nest' parameter in ECX.
17358 // Must be kept in sync with X86CallingConv.td
17359 NestReg = X86::ECX;
17361 // Check that ECX wasn't needed by an 'inreg' parameter.
17362 FunctionType *FTy = Func->getFunctionType();
17363 const AttributeSet &Attrs = Func->getAttributes();
17365 if (!Attrs.isEmpty() && !Func->isVarArg()) {
17366 unsigned InRegCount = 0;
17369 for (FunctionType::param_iterator I = FTy->param_begin(),
17370 E = FTy->param_end(); I != E; ++I, ++Idx)
17371 if (Attrs.hasAttribute(Idx, Attribute::InReg)) {
17372 auto &DL = DAG.getDataLayout();
17373 // FIXME: should only count parameters that are lowered to integers.
17374 InRegCount += (DL.getTypeSizeInBits(*I) + 31) / 32;
17377 if (InRegCount > 2) {
17378 report_fatal_error("Nest register in use - reduce number of inreg"
17384 case CallingConv::X86_FastCall:
17385 case CallingConv::X86_ThisCall:
17386 case CallingConv::Fast:
17387 // Pass 'nest' parameter in EAX.
17388 // Must be kept in sync with X86CallingConv.td
17389 NestReg = X86::EAX;
17393 SDValue OutChains[4];
17394 SDValue Addr, Disp;
17396 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
17397 DAG.getConstant(10, dl, MVT::i32));
17398 Disp = DAG.getNode(ISD::SUB, dl, MVT::i32, FPtr, Addr);
17400 // This is storing the opcode for MOV32ri.
17401 const unsigned char MOV32ri = 0xB8; // X86::MOV32ri's opcode byte.
17402 const unsigned char N86Reg = TRI->getEncodingValue(NestReg) & 0x7;
17403 OutChains[0] = DAG.getStore(Root, dl,
17404 DAG.getConstant(MOV32ri|N86Reg, dl, MVT::i8),
17405 Trmp, MachinePointerInfo(TrmpAddr),
17408 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
17409 DAG.getConstant(1, dl, MVT::i32));
17410 OutChains[1] = DAG.getStore(Root, dl, Nest, Addr,
17411 MachinePointerInfo(TrmpAddr, 1),
17414 const unsigned char JMP = 0xE9; // jmp <32bit dst> opcode.
17415 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
17416 DAG.getConstant(5, dl, MVT::i32));
17417 OutChains[2] = DAG.getStore(Root, dl, DAG.getConstant(JMP, dl, MVT::i8),
17418 Addr, MachinePointerInfo(TrmpAddr, 5),
17421 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
17422 DAG.getConstant(6, dl, MVT::i32));
17423 OutChains[3] = DAG.getStore(Root, dl, Disp, Addr,
17424 MachinePointerInfo(TrmpAddr, 6),
17427 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other, OutChains);
17431 SDValue X86TargetLowering::LowerFLT_ROUNDS_(SDValue Op,
17432 SelectionDAG &DAG) const {
17434 The rounding mode is in bits 11:10 of FPSR, and has the following
17436 00 Round to nearest
17441 FLT_ROUNDS, on the other hand, expects the following:
17448 To perform the conversion, we do:
17449 (((((FPSR & 0x800) >> 11) | ((FPSR & 0x400) >> 9)) + 1) & 3)
17452 MachineFunction &MF = DAG.getMachineFunction();
17453 const TargetFrameLowering &TFI = *Subtarget->getFrameLowering();
17454 unsigned StackAlignment = TFI.getStackAlignment();
17455 MVT VT = Op.getSimpleValueType();
17458 // Save FP Control Word to stack slot
17459 int SSFI = MF.getFrameInfo()->CreateStackObject(2, StackAlignment, false);
17460 SDValue StackSlot =
17461 DAG.getFrameIndex(SSFI, getPointerTy(DAG.getDataLayout()));
17463 MachineMemOperand *MMO =
17464 MF.getMachineMemOperand(MachinePointerInfo::getFixedStack(MF, SSFI),
17465 MachineMemOperand::MOStore, 2, 2);
17467 SDValue Ops[] = { DAG.getEntryNode(), StackSlot };
17468 SDValue Chain = DAG.getMemIntrinsicNode(X86ISD::FNSTCW16m, DL,
17469 DAG.getVTList(MVT::Other),
17470 Ops, MVT::i16, MMO);
17472 // Load FP Control Word from stack slot
17473 SDValue CWD = DAG.getLoad(MVT::i16, DL, Chain, StackSlot,
17474 MachinePointerInfo(), false, false, false, 0);
17476 // Transform as necessary
17478 DAG.getNode(ISD::SRL, DL, MVT::i16,
17479 DAG.getNode(ISD::AND, DL, MVT::i16,
17480 CWD, DAG.getConstant(0x800, DL, MVT::i16)),
17481 DAG.getConstant(11, DL, MVT::i8));
17483 DAG.getNode(ISD::SRL, DL, MVT::i16,
17484 DAG.getNode(ISD::AND, DL, MVT::i16,
17485 CWD, DAG.getConstant(0x400, DL, MVT::i16)),
17486 DAG.getConstant(9, DL, MVT::i8));
17489 DAG.getNode(ISD::AND, DL, MVT::i16,
17490 DAG.getNode(ISD::ADD, DL, MVT::i16,
17491 DAG.getNode(ISD::OR, DL, MVT::i16, CWD1, CWD2),
17492 DAG.getConstant(1, DL, MVT::i16)),
17493 DAG.getConstant(3, DL, MVT::i16));
17495 return DAG.getNode((VT.getSizeInBits() < 16 ?
17496 ISD::TRUNCATE : ISD::ZERO_EXTEND), DL, VT, RetVal);
17499 /// \brief Lower a vector CTLZ using native supported vector CTLZ instruction.
17501 // 1. i32/i64 128/256-bit vector (native support require VLX) are expended
17502 // to 512-bit vector.
17503 // 2. i8/i16 vector implemented using dword LZCNT vector instruction
17504 // ( sub(trunc(lzcnt(zext32(x)))) ). In case zext32(x) is illegal,
17505 // split the vector, perform operation on it's Lo a Hi part and
17506 // concatenate the results.
17507 static SDValue LowerVectorCTLZ_AVX512(SDValue Op, SelectionDAG &DAG) {
17509 MVT VT = Op.getSimpleValueType();
17510 MVT EltVT = VT.getVectorElementType();
17511 unsigned NumElems = VT.getVectorNumElements();
17513 if (EltVT == MVT::i64 || EltVT == MVT::i32) {
17514 // Extend to 512 bit vector.
17515 assert((VT.is256BitVector() || VT.is128BitVector()) &&
17516 "Unsupported value type for operation");
17518 MVT NewVT = MVT::getVectorVT(EltVT, 512 / VT.getScalarSizeInBits());
17519 SDValue Vec512 = DAG.getNode(ISD::INSERT_SUBVECTOR, dl, NewVT,
17520 DAG.getUNDEF(NewVT),
17522 DAG.getIntPtrConstant(0, dl));
17523 SDValue CtlzNode = DAG.getNode(ISD::CTLZ, dl, NewVT, Vec512);
17525 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT, CtlzNode,
17526 DAG.getIntPtrConstant(0, dl));
17529 assert((EltVT == MVT::i8 || EltVT == MVT::i16) &&
17530 "Unsupported element type");
17532 if (16 < NumElems) {
17533 // Split vector, it's Lo and Hi parts will be handled in next iteration.
17535 std::tie(Lo, Hi) = DAG.SplitVector(Op.getOperand(0), dl);
17536 MVT OutVT = MVT::getVectorVT(EltVT, NumElems/2);
17538 Lo = DAG.getNode(Op.getOpcode(), dl, OutVT, Lo);
17539 Hi = DAG.getNode(Op.getOpcode(), dl, OutVT, Hi);
17541 return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT, Lo, Hi);
17544 MVT NewVT = MVT::getVectorVT(MVT::i32, NumElems);
17546 assert((NewVT.is256BitVector() || NewVT.is512BitVector()) &&
17547 "Unsupported value type for operation");
17549 // Use native supported vector instruction vplzcntd.
17550 Op = DAG.getNode(ISD::ZERO_EXTEND, dl, NewVT, Op.getOperand(0));
17551 SDValue CtlzNode = DAG.getNode(ISD::CTLZ, dl, NewVT, Op);
17552 SDValue TruncNode = DAG.getNode(ISD::TRUNCATE, dl, VT, CtlzNode);
17553 SDValue Delta = DAG.getConstant(32 - EltVT.getSizeInBits(), dl, VT);
17555 return DAG.getNode(ISD::SUB, dl, VT, TruncNode, Delta);
17558 static SDValue LowerCTLZ(SDValue Op, const X86Subtarget *Subtarget,
17559 SelectionDAG &DAG) {
17560 MVT VT = Op.getSimpleValueType();
17562 unsigned NumBits = VT.getSizeInBits();
17565 if (VT.isVector() && Subtarget->hasAVX512())
17566 return LowerVectorCTLZ_AVX512(Op, DAG);
17568 Op = Op.getOperand(0);
17569 if (VT == MVT::i8) {
17570 // Zero extend to i32 since there is not an i8 bsr.
17572 Op = DAG.getNode(ISD::ZERO_EXTEND, dl, OpVT, Op);
17575 // Issue a bsr (scan bits in reverse) which also sets EFLAGS.
17576 SDVTList VTs = DAG.getVTList(OpVT, MVT::i32);
17577 Op = DAG.getNode(X86ISD::BSR, dl, VTs, Op);
17579 // If src is zero (i.e. bsr sets ZF), returns NumBits.
17582 DAG.getConstant(NumBits + NumBits - 1, dl, OpVT),
17583 DAG.getConstant(X86::COND_E, dl, MVT::i8),
17586 Op = DAG.getNode(X86ISD::CMOV, dl, OpVT, Ops);
17588 // Finally xor with NumBits-1.
17589 Op = DAG.getNode(ISD::XOR, dl, OpVT, Op,
17590 DAG.getConstant(NumBits - 1, dl, OpVT));
17593 Op = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op);
17597 static SDValue LowerCTLZ_ZERO_UNDEF(SDValue Op, const X86Subtarget *Subtarget,
17598 SelectionDAG &DAG) {
17599 MVT VT = Op.getSimpleValueType();
17601 unsigned NumBits = VT.getSizeInBits();
17604 if (VT.isVector() && Subtarget->hasAVX512())
17605 return LowerVectorCTLZ_AVX512(Op, DAG);
17607 Op = Op.getOperand(0);
17608 if (VT == MVT::i8) {
17609 // Zero extend to i32 since there is not an i8 bsr.
17611 Op = DAG.getNode(ISD::ZERO_EXTEND, dl, OpVT, Op);
17614 // Issue a bsr (scan bits in reverse).
17615 SDVTList VTs = DAG.getVTList(OpVT, MVT::i32);
17616 Op = DAG.getNode(X86ISD::BSR, dl, VTs, Op);
17618 // And xor with NumBits-1.
17619 Op = DAG.getNode(ISD::XOR, dl, OpVT, Op,
17620 DAG.getConstant(NumBits - 1, dl, OpVT));
17623 Op = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op);
17627 static SDValue LowerCTTZ(SDValue Op, SelectionDAG &DAG) {
17628 MVT VT = Op.getSimpleValueType();
17629 unsigned NumBits = VT.getScalarSizeInBits();
17632 if (VT.isVector()) {
17633 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
17635 SDValue N0 = Op.getOperand(0);
17636 SDValue Zero = DAG.getConstant(0, dl, VT);
17638 // lsb(x) = (x & -x)
17639 SDValue LSB = DAG.getNode(ISD::AND, dl, VT, N0,
17640 DAG.getNode(ISD::SUB, dl, VT, Zero, N0));
17642 // cttz_undef(x) = (width - 1) - ctlz(lsb)
17643 if (Op.getOpcode() == ISD::CTTZ_ZERO_UNDEF &&
17644 TLI.isOperationLegal(ISD::CTLZ, VT)) {
17645 SDValue WidthMinusOne = DAG.getConstant(NumBits - 1, dl, VT);
17646 return DAG.getNode(ISD::SUB, dl, VT, WidthMinusOne,
17647 DAG.getNode(ISD::CTLZ, dl, VT, LSB));
17650 // cttz(x) = ctpop(lsb - 1)
17651 SDValue One = DAG.getConstant(1, dl, VT);
17652 return DAG.getNode(ISD::CTPOP, dl, VT,
17653 DAG.getNode(ISD::SUB, dl, VT, LSB, One));
17656 assert(Op.getOpcode() == ISD::CTTZ &&
17657 "Only scalar CTTZ requires custom lowering");
17659 // Issue a bsf (scan bits forward) which also sets EFLAGS.
17660 SDVTList VTs = DAG.getVTList(VT, MVT::i32);
17661 Op = DAG.getNode(X86ISD::BSF, dl, VTs, Op.getOperand(0));
17663 // If src is zero (i.e. bsf sets ZF), returns NumBits.
17666 DAG.getConstant(NumBits, dl, VT),
17667 DAG.getConstant(X86::COND_E, dl, MVT::i8),
17670 return DAG.getNode(X86ISD::CMOV, dl, VT, Ops);
17673 // Lower256IntArith - Break a 256-bit integer operation into two new 128-bit
17674 // ones, and then concatenate the result back.
17675 static SDValue Lower256IntArith(SDValue Op, SelectionDAG &DAG) {
17676 MVT VT = Op.getSimpleValueType();
17678 assert(VT.is256BitVector() && VT.isInteger() &&
17679 "Unsupported value type for operation");
17681 unsigned NumElems = VT.getVectorNumElements();
17684 // Extract the LHS vectors
17685 SDValue LHS = Op.getOperand(0);
17686 SDValue LHS1 = Extract128BitVector(LHS, 0, DAG, dl);
17687 SDValue LHS2 = Extract128BitVector(LHS, NumElems/2, DAG, dl);
17689 // Extract the RHS vectors
17690 SDValue RHS = Op.getOperand(1);
17691 SDValue RHS1 = Extract128BitVector(RHS, 0, DAG, dl);
17692 SDValue RHS2 = Extract128BitVector(RHS, NumElems/2, DAG, dl);
17694 MVT EltVT = VT.getVectorElementType();
17695 MVT NewVT = MVT::getVectorVT(EltVT, NumElems/2);
17697 return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT,
17698 DAG.getNode(Op.getOpcode(), dl, NewVT, LHS1, RHS1),
17699 DAG.getNode(Op.getOpcode(), dl, NewVT, LHS2, RHS2));
17702 static SDValue LowerADD(SDValue Op, SelectionDAG &DAG) {
17703 if (Op.getValueType() == MVT::i1)
17704 return DAG.getNode(ISD::XOR, SDLoc(Op), Op.getValueType(),
17705 Op.getOperand(0), Op.getOperand(1));
17706 assert(Op.getSimpleValueType().is256BitVector() &&
17707 Op.getSimpleValueType().isInteger() &&
17708 "Only handle AVX 256-bit vector integer operation");
17709 return Lower256IntArith(Op, DAG);
17712 static SDValue LowerSUB(SDValue Op, SelectionDAG &DAG) {
17713 if (Op.getValueType() == MVT::i1)
17714 return DAG.getNode(ISD::XOR, SDLoc(Op), Op.getValueType(),
17715 Op.getOperand(0), Op.getOperand(1));
17716 assert(Op.getSimpleValueType().is256BitVector() &&
17717 Op.getSimpleValueType().isInteger() &&
17718 "Only handle AVX 256-bit vector integer operation");
17719 return Lower256IntArith(Op, DAG);
17722 static SDValue LowerMINMAX(SDValue Op, SelectionDAG &DAG) {
17723 assert(Op.getSimpleValueType().is256BitVector() &&
17724 Op.getSimpleValueType().isInteger() &&
17725 "Only handle AVX 256-bit vector integer operation");
17726 return Lower256IntArith(Op, DAG);
17729 static SDValue LowerMUL(SDValue Op, const X86Subtarget *Subtarget,
17730 SelectionDAG &DAG) {
17732 MVT VT = Op.getSimpleValueType();
17735 return DAG.getNode(ISD::AND, dl, VT, Op.getOperand(0), Op.getOperand(1));
17737 // Decompose 256-bit ops into smaller 128-bit ops.
17738 if (VT.is256BitVector() && !Subtarget->hasInt256())
17739 return Lower256IntArith(Op, DAG);
17741 SDValue A = Op.getOperand(0);
17742 SDValue B = Op.getOperand(1);
17744 // Lower v16i8/v32i8 mul as promotion to v8i16/v16i16 vector
17745 // pairs, multiply and truncate.
17746 if (VT == MVT::v16i8 || VT == MVT::v32i8) {
17747 if (Subtarget->hasInt256()) {
17748 if (VT == MVT::v32i8) {
17749 MVT SubVT = MVT::getVectorVT(MVT::i8, VT.getVectorNumElements() / 2);
17750 SDValue Lo = DAG.getIntPtrConstant(0, dl);
17751 SDValue Hi = DAG.getIntPtrConstant(VT.getVectorNumElements() / 2, dl);
17752 SDValue ALo = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, SubVT, A, Lo);
17753 SDValue BLo = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, SubVT, B, Lo);
17754 SDValue AHi = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, SubVT, A, Hi);
17755 SDValue BHi = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, SubVT, B, Hi);
17756 return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT,
17757 DAG.getNode(ISD::MUL, dl, SubVT, ALo, BLo),
17758 DAG.getNode(ISD::MUL, dl, SubVT, AHi, BHi));
17761 MVT ExVT = MVT::getVectorVT(MVT::i16, VT.getVectorNumElements());
17762 return DAG.getNode(
17763 ISD::TRUNCATE, dl, VT,
17764 DAG.getNode(ISD::MUL, dl, ExVT,
17765 DAG.getNode(ISD::SIGN_EXTEND, dl, ExVT, A),
17766 DAG.getNode(ISD::SIGN_EXTEND, dl, ExVT, B)));
17769 assert(VT == MVT::v16i8 &&
17770 "Pre-AVX2 support only supports v16i8 multiplication");
17771 MVT ExVT = MVT::v8i16;
17773 // Extract the lo parts and sign extend to i16
17775 if (Subtarget->hasSSE41()) {
17776 ALo = DAG.getNode(X86ISD::VSEXT, dl, ExVT, A);
17777 BLo = DAG.getNode(X86ISD::VSEXT, dl, ExVT, B);
17779 const int ShufMask[] = {-1, 0, -1, 1, -1, 2, -1, 3,
17780 -1, 4, -1, 5, -1, 6, -1, 7};
17781 ALo = DAG.getVectorShuffle(VT, dl, A, A, ShufMask);
17782 BLo = DAG.getVectorShuffle(VT, dl, B, B, ShufMask);
17783 ALo = DAG.getBitcast(ExVT, ALo);
17784 BLo = DAG.getBitcast(ExVT, BLo);
17785 ALo = DAG.getNode(ISD::SRA, dl, ExVT, ALo, DAG.getConstant(8, dl, ExVT));
17786 BLo = DAG.getNode(ISD::SRA, dl, ExVT, BLo, DAG.getConstant(8, dl, ExVT));
17789 // Extract the hi parts and sign extend to i16
17791 if (Subtarget->hasSSE41()) {
17792 const int ShufMask[] = {8, 9, 10, 11, 12, 13, 14, 15,
17793 -1, -1, -1, -1, -1, -1, -1, -1};
17794 AHi = DAG.getVectorShuffle(VT, dl, A, A, ShufMask);
17795 BHi = DAG.getVectorShuffle(VT, dl, B, B, ShufMask);
17796 AHi = DAG.getNode(X86ISD::VSEXT, dl, ExVT, AHi);
17797 BHi = DAG.getNode(X86ISD::VSEXT, dl, ExVT, BHi);
17799 const int ShufMask[] = {-1, 8, -1, 9, -1, 10, -1, 11,
17800 -1, 12, -1, 13, -1, 14, -1, 15};
17801 AHi = DAG.getVectorShuffle(VT, dl, A, A, ShufMask);
17802 BHi = DAG.getVectorShuffle(VT, dl, B, B, ShufMask);
17803 AHi = DAG.getBitcast(ExVT, AHi);
17804 BHi = DAG.getBitcast(ExVT, BHi);
17805 AHi = DAG.getNode(ISD::SRA, dl, ExVT, AHi, DAG.getConstant(8, dl, ExVT));
17806 BHi = DAG.getNode(ISD::SRA, dl, ExVT, BHi, DAG.getConstant(8, dl, ExVT));
17809 // Multiply, mask the lower 8bits of the lo/hi results and pack
17810 SDValue RLo = DAG.getNode(ISD::MUL, dl, ExVT, ALo, BLo);
17811 SDValue RHi = DAG.getNode(ISD::MUL, dl, ExVT, AHi, BHi);
17812 RLo = DAG.getNode(ISD::AND, dl, ExVT, RLo, DAG.getConstant(255, dl, ExVT));
17813 RHi = DAG.getNode(ISD::AND, dl, ExVT, RHi, DAG.getConstant(255, dl, ExVT));
17814 return DAG.getNode(X86ISD::PACKUS, dl, VT, RLo, RHi);
17817 // Lower v4i32 mul as 2x shuffle, 2x pmuludq, 2x shuffle.
17818 if (VT == MVT::v4i32) {
17819 assert(Subtarget->hasSSE2() && !Subtarget->hasSSE41() &&
17820 "Should not custom lower when pmuldq is available!");
17822 // Extract the odd parts.
17823 static const int UnpackMask[] = { 1, -1, 3, -1 };
17824 SDValue Aodds = DAG.getVectorShuffle(VT, dl, A, A, UnpackMask);
17825 SDValue Bodds = DAG.getVectorShuffle(VT, dl, B, B, UnpackMask);
17827 // Multiply the even parts.
17828 SDValue Evens = DAG.getNode(X86ISD::PMULUDQ, dl, MVT::v2i64, A, B);
17829 // Now multiply odd parts.
17830 SDValue Odds = DAG.getNode(X86ISD::PMULUDQ, dl, MVT::v2i64, Aodds, Bodds);
17832 Evens = DAG.getBitcast(VT, Evens);
17833 Odds = DAG.getBitcast(VT, Odds);
17835 // Merge the two vectors back together with a shuffle. This expands into 2
17837 static const int ShufMask[] = { 0, 4, 2, 6 };
17838 return DAG.getVectorShuffle(VT, dl, Evens, Odds, ShufMask);
17841 assert((VT == MVT::v2i64 || VT == MVT::v4i64 || VT == MVT::v8i64) &&
17842 "Only know how to lower V2I64/V4I64/V8I64 multiply");
17844 // Ahi = psrlqi(a, 32);
17845 // Bhi = psrlqi(b, 32);
17847 // AloBlo = pmuludq(a, b);
17848 // AloBhi = pmuludq(a, Bhi);
17849 // AhiBlo = pmuludq(Ahi, b);
17851 // AloBhi = psllqi(AloBhi, 32);
17852 // AhiBlo = psllqi(AhiBlo, 32);
17853 // return AloBlo + AloBhi + AhiBlo;
17855 SDValue Ahi = getTargetVShiftByConstNode(X86ISD::VSRLI, dl, VT, A, 32, DAG);
17856 SDValue Bhi = getTargetVShiftByConstNode(X86ISD::VSRLI, dl, VT, B, 32, DAG);
17858 SDValue AhiBlo = Ahi;
17859 SDValue AloBhi = Bhi;
17860 // Bit cast to 32-bit vectors for MULUDQ
17861 EVT MulVT = (VT == MVT::v2i64) ? MVT::v4i32 :
17862 (VT == MVT::v4i64) ? MVT::v8i32 : MVT::v16i32;
17863 A = DAG.getBitcast(MulVT, A);
17864 B = DAG.getBitcast(MulVT, B);
17865 Ahi = DAG.getBitcast(MulVT, Ahi);
17866 Bhi = DAG.getBitcast(MulVT, Bhi);
17868 SDValue AloBlo = DAG.getNode(X86ISD::PMULUDQ, dl, VT, A, B);
17869 // After shifting right const values the result may be all-zero.
17870 if (!ISD::isBuildVectorAllZeros(Ahi.getNode())) {
17871 AhiBlo = DAG.getNode(X86ISD::PMULUDQ, dl, VT, Ahi, B);
17872 AhiBlo = getTargetVShiftByConstNode(X86ISD::VSHLI, dl, VT, AhiBlo, 32, DAG);
17874 if (!ISD::isBuildVectorAllZeros(Bhi.getNode())) {
17875 AloBhi = DAG.getNode(X86ISD::PMULUDQ, dl, VT, A, Bhi);
17876 AloBhi = getTargetVShiftByConstNode(X86ISD::VSHLI, dl, VT, AloBhi, 32, DAG);
17879 SDValue Res = DAG.getNode(ISD::ADD, dl, VT, AloBlo, AloBhi);
17880 return DAG.getNode(ISD::ADD, dl, VT, Res, AhiBlo);
17883 SDValue X86TargetLowering::LowerWin64_i128OP(SDValue Op, SelectionDAG &DAG) const {
17884 assert(Subtarget->isTargetWin64() && "Unexpected target");
17885 EVT VT = Op.getValueType();
17886 assert(VT.isInteger() && VT.getSizeInBits() == 128 &&
17887 "Unexpected return type for lowering");
17891 switch (Op->getOpcode()) {
17892 default: llvm_unreachable("Unexpected request for libcall!");
17893 case ISD::SDIV: isSigned = true; LC = RTLIB::SDIV_I128; break;
17894 case ISD::UDIV: isSigned = false; LC = RTLIB::UDIV_I128; break;
17895 case ISD::SREM: isSigned = true; LC = RTLIB::SREM_I128; break;
17896 case ISD::UREM: isSigned = false; LC = RTLIB::UREM_I128; break;
17897 case ISD::SDIVREM: isSigned = true; LC = RTLIB::SDIVREM_I128; break;
17898 case ISD::UDIVREM: isSigned = false; LC = RTLIB::UDIVREM_I128; break;
17902 SDValue InChain = DAG.getEntryNode();
17904 TargetLowering::ArgListTy Args;
17905 TargetLowering::ArgListEntry Entry;
17906 for (unsigned i = 0, e = Op->getNumOperands(); i != e; ++i) {
17907 EVT ArgVT = Op->getOperand(i).getValueType();
17908 assert(ArgVT.isInteger() && ArgVT.getSizeInBits() == 128 &&
17909 "Unexpected argument type for lowering");
17910 SDValue StackPtr = DAG.CreateStackTemporary(ArgVT, 16);
17911 Entry.Node = StackPtr;
17912 InChain = DAG.getStore(InChain, dl, Op->getOperand(i), StackPtr, MachinePointerInfo(),
17914 Type *ArgTy = ArgVT.getTypeForEVT(*DAG.getContext());
17915 Entry.Ty = PointerType::get(ArgTy,0);
17916 Entry.isSExt = false;
17917 Entry.isZExt = false;
17918 Args.push_back(Entry);
17921 SDValue Callee = DAG.getExternalSymbol(getLibcallName(LC),
17922 getPointerTy(DAG.getDataLayout()));
17924 TargetLowering::CallLoweringInfo CLI(DAG);
17925 CLI.setDebugLoc(dl).setChain(InChain)
17926 .setCallee(getLibcallCallingConv(LC),
17927 static_cast<EVT>(MVT::v2i64).getTypeForEVT(*DAG.getContext()),
17928 Callee, std::move(Args), 0)
17929 .setInRegister().setSExtResult(isSigned).setZExtResult(!isSigned);
17931 std::pair<SDValue, SDValue> CallInfo = LowerCallTo(CLI);
17932 return DAG.getBitcast(VT, CallInfo.first);
17935 static SDValue LowerMUL_LOHI(SDValue Op, const X86Subtarget *Subtarget,
17936 SelectionDAG &DAG) {
17937 SDValue Op0 = Op.getOperand(0), Op1 = Op.getOperand(1);
17938 EVT VT = Op0.getValueType();
17941 assert((VT == MVT::v4i32 && Subtarget->hasSSE2()) ||
17942 (VT == MVT::v8i32 && Subtarget->hasInt256()));
17944 // PMULxD operations multiply each even value (starting at 0) of LHS with
17945 // the related value of RHS and produce a widen result.
17946 // E.g., PMULUDQ <4 x i32> <a|b|c|d>, <4 x i32> <e|f|g|h>
17947 // => <2 x i64> <ae|cg>
17949 // In other word, to have all the results, we need to perform two PMULxD:
17950 // 1. one with the even values.
17951 // 2. one with the odd values.
17952 // To achieve #2, with need to place the odd values at an even position.
17954 // Place the odd value at an even position (basically, shift all values 1
17955 // step to the left):
17956 const int Mask[] = {1, -1, 3, -1, 5, -1, 7, -1};
17957 // <a|b|c|d> => <b|undef|d|undef>
17958 SDValue Odd0 = DAG.getVectorShuffle(VT, dl, Op0, Op0, Mask);
17959 // <e|f|g|h> => <f|undef|h|undef>
17960 SDValue Odd1 = DAG.getVectorShuffle(VT, dl, Op1, Op1, Mask);
17962 // Emit two multiplies, one for the lower 2 ints and one for the higher 2
17964 MVT MulVT = VT == MVT::v4i32 ? MVT::v2i64 : MVT::v4i64;
17965 bool IsSigned = Op->getOpcode() == ISD::SMUL_LOHI;
17967 (!IsSigned || !Subtarget->hasSSE41()) ? X86ISD::PMULUDQ : X86ISD::PMULDQ;
17968 // PMULUDQ <4 x i32> <a|b|c|d>, <4 x i32> <e|f|g|h>
17969 // => <2 x i64> <ae|cg>
17970 SDValue Mul1 = DAG.getBitcast(VT, DAG.getNode(Opcode, dl, MulVT, Op0, Op1));
17971 // PMULUDQ <4 x i32> <b|undef|d|undef>, <4 x i32> <f|undef|h|undef>
17972 // => <2 x i64> <bf|dh>
17973 SDValue Mul2 = DAG.getBitcast(VT, DAG.getNode(Opcode, dl, MulVT, Odd0, Odd1));
17975 // Shuffle it back into the right order.
17976 SDValue Highs, Lows;
17977 if (VT == MVT::v8i32) {
17978 const int HighMask[] = {1, 9, 3, 11, 5, 13, 7, 15};
17979 Highs = DAG.getVectorShuffle(VT, dl, Mul1, Mul2, HighMask);
17980 const int LowMask[] = {0, 8, 2, 10, 4, 12, 6, 14};
17981 Lows = DAG.getVectorShuffle(VT, dl, Mul1, Mul2, LowMask);
17983 const int HighMask[] = {1, 5, 3, 7};
17984 Highs = DAG.getVectorShuffle(VT, dl, Mul1, Mul2, HighMask);
17985 const int LowMask[] = {0, 4, 2, 6};
17986 Lows = DAG.getVectorShuffle(VT, dl, Mul1, Mul2, LowMask);
17989 // If we have a signed multiply but no PMULDQ fix up the high parts of a
17990 // unsigned multiply.
17991 if (IsSigned && !Subtarget->hasSSE41()) {
17992 SDValue ShAmt = DAG.getConstant(
17994 DAG.getTargetLoweringInfo().getShiftAmountTy(VT, DAG.getDataLayout()));
17995 SDValue T1 = DAG.getNode(ISD::AND, dl, VT,
17996 DAG.getNode(ISD::SRA, dl, VT, Op0, ShAmt), Op1);
17997 SDValue T2 = DAG.getNode(ISD::AND, dl, VT,
17998 DAG.getNode(ISD::SRA, dl, VT, Op1, ShAmt), Op0);
18000 SDValue Fixup = DAG.getNode(ISD::ADD, dl, VT, T1, T2);
18001 Highs = DAG.getNode(ISD::SUB, dl, VT, Highs, Fixup);
18004 // The first result of MUL_LOHI is actually the low value, followed by the
18006 SDValue Ops[] = {Lows, Highs};
18007 return DAG.getMergeValues(Ops, dl);
18010 // Return true if the required (according to Opcode) shift-imm form is natively
18011 // supported by the Subtarget
18012 static bool SupportedVectorShiftWithImm(MVT VT, const X86Subtarget *Subtarget,
18014 if (VT.getScalarSizeInBits() < 16)
18017 if (VT.is512BitVector() &&
18018 (VT.getScalarSizeInBits() > 16 || Subtarget->hasBWI()))
18021 bool LShift = VT.is128BitVector() ||
18022 (VT.is256BitVector() && Subtarget->hasInt256());
18024 bool AShift = LShift && (Subtarget->hasVLX() ||
18025 (VT != MVT::v2i64 && VT != MVT::v4i64));
18026 return (Opcode == ISD::SRA) ? AShift : LShift;
18029 // The shift amount is a variable, but it is the same for all vector lanes.
18030 // These instructions are defined together with shift-immediate.
18032 bool SupportedVectorShiftWithBaseAmnt(MVT VT, const X86Subtarget *Subtarget,
18034 return SupportedVectorShiftWithImm(VT, Subtarget, Opcode);
18037 // Return true if the required (according to Opcode) variable-shift form is
18038 // natively supported by the Subtarget
18039 static bool SupportedVectorVarShift(MVT VT, const X86Subtarget *Subtarget,
18042 if (!Subtarget->hasInt256() || VT.getScalarSizeInBits() < 16)
18045 // vXi16 supported only on AVX-512, BWI
18046 if (VT.getScalarSizeInBits() == 16 && !Subtarget->hasBWI())
18049 if (VT.is512BitVector() || Subtarget->hasVLX())
18052 bool LShift = VT.is128BitVector() || VT.is256BitVector();
18053 bool AShift = LShift && VT != MVT::v2i64 && VT != MVT::v4i64;
18054 return (Opcode == ISD::SRA) ? AShift : LShift;
18057 static SDValue LowerScalarImmediateShift(SDValue Op, SelectionDAG &DAG,
18058 const X86Subtarget *Subtarget) {
18059 MVT VT = Op.getSimpleValueType();
18061 SDValue R = Op.getOperand(0);
18062 SDValue Amt = Op.getOperand(1);
18064 unsigned X86Opc = (Op.getOpcode() == ISD::SHL) ? X86ISD::VSHLI :
18065 (Op.getOpcode() == ISD::SRL) ? X86ISD::VSRLI : X86ISD::VSRAI;
18067 auto ArithmeticShiftRight64 = [&](uint64_t ShiftAmt) {
18068 assert((VT == MVT::v2i64 || VT == MVT::v4i64) && "Unexpected SRA type");
18069 MVT ExVT = MVT::getVectorVT(MVT::i32, VT.getVectorNumElements() * 2);
18070 SDValue Ex = DAG.getBitcast(ExVT, R);
18072 if (ShiftAmt >= 32) {
18073 // Splat sign to upper i32 dst, and SRA upper i32 src to lower i32.
18075 getTargetVShiftByConstNode(X86ISD::VSRAI, dl, ExVT, Ex, 31, DAG);
18076 SDValue Lower = getTargetVShiftByConstNode(X86ISD::VSRAI, dl, ExVT, Ex,
18077 ShiftAmt - 32, DAG);
18078 if (VT == MVT::v2i64)
18079 Ex = DAG.getVectorShuffle(ExVT, dl, Upper, Lower, {5, 1, 7, 3});
18080 if (VT == MVT::v4i64)
18081 Ex = DAG.getVectorShuffle(ExVT, dl, Upper, Lower,
18082 {9, 1, 11, 3, 13, 5, 15, 7});
18084 // SRA upper i32, SHL whole i64 and select lower i32.
18085 SDValue Upper = getTargetVShiftByConstNode(X86ISD::VSRAI, dl, ExVT, Ex,
18088 getTargetVShiftByConstNode(X86ISD::VSRLI, dl, VT, R, ShiftAmt, DAG);
18089 Lower = DAG.getBitcast(ExVT, Lower);
18090 if (VT == MVT::v2i64)
18091 Ex = DAG.getVectorShuffle(ExVT, dl, Upper, Lower, {4, 1, 6, 3});
18092 if (VT == MVT::v4i64)
18093 Ex = DAG.getVectorShuffle(ExVT, dl, Upper, Lower,
18094 {8, 1, 10, 3, 12, 5, 14, 7});
18096 return DAG.getBitcast(VT, Ex);
18099 // Optimize shl/srl/sra with constant shift amount.
18100 if (auto *BVAmt = dyn_cast<BuildVectorSDNode>(Amt)) {
18101 if (auto *ShiftConst = BVAmt->getConstantSplatNode()) {
18102 uint64_t ShiftAmt = ShiftConst->getZExtValue();
18104 if (SupportedVectorShiftWithImm(VT, Subtarget, Op.getOpcode()))
18105 return getTargetVShiftByConstNode(X86Opc, dl, VT, R, ShiftAmt, DAG);
18107 // i64 SRA needs to be performed as partial shifts.
18108 if ((VT == MVT::v2i64 || (Subtarget->hasInt256() && VT == MVT::v4i64)) &&
18109 Op.getOpcode() == ISD::SRA && !Subtarget->hasXOP())
18110 return ArithmeticShiftRight64(ShiftAmt);
18112 if (VT == MVT::v16i8 || (Subtarget->hasInt256() && VT == MVT::v32i8)) {
18113 unsigned NumElts = VT.getVectorNumElements();
18114 MVT ShiftVT = MVT::getVectorVT(MVT::i16, NumElts / 2);
18116 // Simple i8 add case
18117 if (Op.getOpcode() == ISD::SHL && ShiftAmt == 1)
18118 return DAG.getNode(ISD::ADD, dl, VT, R, R);
18120 // ashr(R, 7) === cmp_slt(R, 0)
18121 if (Op.getOpcode() == ISD::SRA && ShiftAmt == 7) {
18122 SDValue Zeros = getZeroVector(VT, Subtarget, DAG, dl);
18123 return DAG.getNode(X86ISD::PCMPGT, dl, VT, Zeros, R);
18126 // XOP can shift v16i8 directly instead of as shift v8i16 + mask.
18127 if (VT == MVT::v16i8 && Subtarget->hasXOP())
18130 if (Op.getOpcode() == ISD::SHL) {
18131 // Make a large shift.
18132 SDValue SHL = getTargetVShiftByConstNode(X86ISD::VSHLI, dl, ShiftVT,
18134 SHL = DAG.getBitcast(VT, SHL);
18135 // Zero out the rightmost bits.
18136 SmallVector<SDValue, 32> V(
18137 NumElts, DAG.getConstant(uint8_t(-1U << ShiftAmt), dl, MVT::i8));
18138 return DAG.getNode(ISD::AND, dl, VT, SHL,
18139 DAG.getNode(ISD::BUILD_VECTOR, dl, VT, V));
18141 if (Op.getOpcode() == ISD::SRL) {
18142 // Make a large shift.
18143 SDValue SRL = getTargetVShiftByConstNode(X86ISD::VSRLI, dl, ShiftVT,
18145 SRL = DAG.getBitcast(VT, SRL);
18146 // Zero out the leftmost bits.
18147 SmallVector<SDValue, 32> V(
18148 NumElts, DAG.getConstant(uint8_t(-1U) >> ShiftAmt, dl, MVT::i8));
18149 return DAG.getNode(ISD::AND, dl, VT, SRL,
18150 DAG.getNode(ISD::BUILD_VECTOR, dl, VT, V));
18152 if (Op.getOpcode() == ISD::SRA) {
18153 // ashr(R, Amt) === sub(xor(lshr(R, Amt), Mask), Mask)
18154 SDValue Res = DAG.getNode(ISD::SRL, dl, VT, R, Amt);
18155 SmallVector<SDValue, 32> V(NumElts,
18156 DAG.getConstant(128 >> ShiftAmt, dl,
18158 SDValue Mask = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, V);
18159 Res = DAG.getNode(ISD::XOR, dl, VT, Res, Mask);
18160 Res = DAG.getNode(ISD::SUB, dl, VT, Res, Mask);
18163 llvm_unreachable("Unknown shift opcode.");
18168 // Special case in 32-bit mode, where i64 is expanded into high and low parts.
18169 if (!Subtarget->is64Bit() && !Subtarget->hasXOP() &&
18170 (VT == MVT::v2i64 || (Subtarget->hasInt256() && VT == MVT::v4i64))) {
18172 // Peek through any splat that was introduced for i64 shift vectorization.
18173 int SplatIndex = -1;
18174 if (ShuffleVectorSDNode *SVN = dyn_cast<ShuffleVectorSDNode>(Amt.getNode()))
18175 if (SVN->isSplat()) {
18176 SplatIndex = SVN->getSplatIndex();
18177 Amt = Amt.getOperand(0);
18178 assert(SplatIndex < (int)VT.getVectorNumElements() &&
18179 "Splat shuffle referencing second operand");
18182 if (Amt.getOpcode() != ISD::BITCAST ||
18183 Amt.getOperand(0).getOpcode() != ISD::BUILD_VECTOR)
18186 Amt = Amt.getOperand(0);
18187 unsigned Ratio = Amt.getSimpleValueType().getVectorNumElements() /
18188 VT.getVectorNumElements();
18189 unsigned RatioInLog2 = Log2_32_Ceil(Ratio);
18190 uint64_t ShiftAmt = 0;
18191 unsigned BaseOp = (SplatIndex < 0 ? 0 : SplatIndex * Ratio);
18192 for (unsigned i = 0; i != Ratio; ++i) {
18193 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Amt.getOperand(i + BaseOp));
18197 ShiftAmt |= C->getZExtValue() << (i * (1 << (6 - RatioInLog2)));
18200 // Check remaining shift amounts (if not a splat).
18201 if (SplatIndex < 0) {
18202 for (unsigned i = Ratio; i != Amt.getNumOperands(); i += Ratio) {
18203 uint64_t ShAmt = 0;
18204 for (unsigned j = 0; j != Ratio; ++j) {
18205 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Amt.getOperand(i + j));
18209 ShAmt |= C->getZExtValue() << (j * (1 << (6 - RatioInLog2)));
18211 if (ShAmt != ShiftAmt)
18216 if (SupportedVectorShiftWithImm(VT, Subtarget, Op.getOpcode()))
18217 return getTargetVShiftByConstNode(X86Opc, dl, VT, R, ShiftAmt, DAG);
18219 if (Op.getOpcode() == ISD::SRA)
18220 return ArithmeticShiftRight64(ShiftAmt);
18226 static SDValue LowerScalarVariableShift(SDValue Op, SelectionDAG &DAG,
18227 const X86Subtarget* Subtarget) {
18228 MVT VT = Op.getSimpleValueType();
18230 SDValue R = Op.getOperand(0);
18231 SDValue Amt = Op.getOperand(1);
18233 unsigned X86OpcI = (Op.getOpcode() == ISD::SHL) ? X86ISD::VSHLI :
18234 (Op.getOpcode() == ISD::SRL) ? X86ISD::VSRLI : X86ISD::VSRAI;
18236 unsigned X86OpcV = (Op.getOpcode() == ISD::SHL) ? X86ISD::VSHL :
18237 (Op.getOpcode() == ISD::SRL) ? X86ISD::VSRL : X86ISD::VSRA;
18239 if (SupportedVectorShiftWithBaseAmnt(VT, Subtarget, Op.getOpcode())) {
18241 EVT EltVT = VT.getVectorElementType();
18243 if (BuildVectorSDNode *BV = dyn_cast<BuildVectorSDNode>(Amt)) {
18244 // Check if this build_vector node is doing a splat.
18245 // If so, then set BaseShAmt equal to the splat value.
18246 BaseShAmt = BV->getSplatValue();
18247 if (BaseShAmt && BaseShAmt.getOpcode() == ISD::UNDEF)
18248 BaseShAmt = SDValue();
18250 if (Amt.getOpcode() == ISD::EXTRACT_SUBVECTOR)
18251 Amt = Amt.getOperand(0);
18253 ShuffleVectorSDNode *SVN = dyn_cast<ShuffleVectorSDNode>(Amt);
18254 if (SVN && SVN->isSplat()) {
18255 unsigned SplatIdx = (unsigned)SVN->getSplatIndex();
18256 SDValue InVec = Amt.getOperand(0);
18257 if (InVec.getOpcode() == ISD::BUILD_VECTOR) {
18258 assert((SplatIdx < InVec.getValueType().getVectorNumElements()) &&
18259 "Unexpected shuffle index found!");
18260 BaseShAmt = InVec.getOperand(SplatIdx);
18261 } else if (InVec.getOpcode() == ISD::INSERT_VECTOR_ELT) {
18262 if (ConstantSDNode *C =
18263 dyn_cast<ConstantSDNode>(InVec.getOperand(2))) {
18264 if (C->getZExtValue() == SplatIdx)
18265 BaseShAmt = InVec.getOperand(1);
18270 // Avoid introducing an extract element from a shuffle.
18271 BaseShAmt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, EltVT, InVec,
18272 DAG.getIntPtrConstant(SplatIdx, dl));
18276 if (BaseShAmt.getNode()) {
18277 assert(EltVT.bitsLE(MVT::i64) && "Unexpected element type!");
18278 if (EltVT != MVT::i64 && EltVT.bitsGT(MVT::i32))
18279 BaseShAmt = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i64, BaseShAmt);
18280 else if (EltVT.bitsLT(MVT::i32))
18281 BaseShAmt = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, BaseShAmt);
18283 return getTargetVShiftNode(X86OpcI, dl, VT, R, BaseShAmt, DAG);
18287 // Special case in 32-bit mode, where i64 is expanded into high and low parts.
18288 if (!Subtarget->is64Bit() && VT == MVT::v2i64 &&
18289 Amt.getOpcode() == ISD::BITCAST &&
18290 Amt.getOperand(0).getOpcode() == ISD::BUILD_VECTOR) {
18291 Amt = Amt.getOperand(0);
18292 unsigned Ratio = Amt.getSimpleValueType().getVectorNumElements() /
18293 VT.getVectorNumElements();
18294 std::vector<SDValue> Vals(Ratio);
18295 for (unsigned i = 0; i != Ratio; ++i)
18296 Vals[i] = Amt.getOperand(i);
18297 for (unsigned i = Ratio; i != Amt.getNumOperands(); i += Ratio) {
18298 for (unsigned j = 0; j != Ratio; ++j)
18299 if (Vals[j] != Amt.getOperand(i + j))
18303 if (SupportedVectorShiftWithBaseAmnt(VT, Subtarget, Op.getOpcode()))
18304 return DAG.getNode(X86OpcV, dl, VT, R, Op.getOperand(1));
18309 static SDValue LowerShift(SDValue Op, const X86Subtarget* Subtarget,
18310 SelectionDAG &DAG) {
18311 MVT VT = Op.getSimpleValueType();
18313 SDValue R = Op.getOperand(0);
18314 SDValue Amt = Op.getOperand(1);
18316 assert(VT.isVector() && "Custom lowering only for vector shifts!");
18317 assert(Subtarget->hasSSE2() && "Only custom lower when we have SSE2!");
18319 if (SDValue V = LowerScalarImmediateShift(Op, DAG, Subtarget))
18322 if (SDValue V = LowerScalarVariableShift(Op, DAG, Subtarget))
18325 if (SupportedVectorVarShift(VT, Subtarget, Op.getOpcode()))
18328 // XOP has 128-bit variable logical/arithmetic shifts.
18329 // +ve/-ve Amt = shift left/right.
18330 if (Subtarget->hasXOP() &&
18331 (VT == MVT::v2i64 || VT == MVT::v4i32 ||
18332 VT == MVT::v8i16 || VT == MVT::v16i8)) {
18333 if (Op.getOpcode() == ISD::SRL || Op.getOpcode() == ISD::SRA) {
18334 SDValue Zero = getZeroVector(VT, Subtarget, DAG, dl);
18335 Amt = DAG.getNode(ISD::SUB, dl, VT, Zero, Amt);
18337 if (Op.getOpcode() == ISD::SHL || Op.getOpcode() == ISD::SRL)
18338 return DAG.getNode(X86ISD::VPSHL, dl, VT, R, Amt);
18339 if (Op.getOpcode() == ISD::SRA)
18340 return DAG.getNode(X86ISD::VPSHA, dl, VT, R, Amt);
18343 // 2i64 vector logical shifts can efficiently avoid scalarization - do the
18344 // shifts per-lane and then shuffle the partial results back together.
18345 if (VT == MVT::v2i64 && Op.getOpcode() != ISD::SRA) {
18346 // Splat the shift amounts so the scalar shifts above will catch it.
18347 SDValue Amt0 = DAG.getVectorShuffle(VT, dl, Amt, Amt, {0, 0});
18348 SDValue Amt1 = DAG.getVectorShuffle(VT, dl, Amt, Amt, {1, 1});
18349 SDValue R0 = DAG.getNode(Op->getOpcode(), dl, VT, R, Amt0);
18350 SDValue R1 = DAG.getNode(Op->getOpcode(), dl, VT, R, Amt1);
18351 return DAG.getVectorShuffle(VT, dl, R0, R1, {0, 3});
18354 // i64 vector arithmetic shift can be emulated with the transform:
18355 // M = lshr(SIGN_BIT, Amt)
18356 // ashr(R, Amt) === sub(xor(lshr(R, Amt), M), M)
18357 if ((VT == MVT::v2i64 || (VT == MVT::v4i64 && Subtarget->hasInt256())) &&
18358 Op.getOpcode() == ISD::SRA) {
18359 SDValue S = DAG.getConstant(APInt::getSignBit(64), dl, VT);
18360 SDValue M = DAG.getNode(ISD::SRL, dl, VT, S, Amt);
18361 R = DAG.getNode(ISD::SRL, dl, VT, R, Amt);
18362 R = DAG.getNode(ISD::XOR, dl, VT, R, M);
18363 R = DAG.getNode(ISD::SUB, dl, VT, R, M);
18367 // If possible, lower this packed shift into a vector multiply instead of
18368 // expanding it into a sequence of scalar shifts.
18369 // Do this only if the vector shift count is a constant build_vector.
18370 if (Op.getOpcode() == ISD::SHL &&
18371 (VT == MVT::v8i16 || VT == MVT::v4i32 ||
18372 (Subtarget->hasInt256() && VT == MVT::v16i16)) &&
18373 ISD::isBuildVectorOfConstantSDNodes(Amt.getNode())) {
18374 SmallVector<SDValue, 8> Elts;
18375 EVT SVT = VT.getScalarType();
18376 unsigned SVTBits = SVT.getSizeInBits();
18377 const APInt &One = APInt(SVTBits, 1);
18378 unsigned NumElems = VT.getVectorNumElements();
18380 for (unsigned i=0; i !=NumElems; ++i) {
18381 SDValue Op = Amt->getOperand(i);
18382 if (Op->getOpcode() == ISD::UNDEF) {
18383 Elts.push_back(Op);
18387 ConstantSDNode *ND = cast<ConstantSDNode>(Op);
18388 const APInt &C = APInt(SVTBits, ND->getAPIntValue().getZExtValue());
18389 uint64_t ShAmt = C.getZExtValue();
18390 if (ShAmt >= SVTBits) {
18391 Elts.push_back(DAG.getUNDEF(SVT));
18394 Elts.push_back(DAG.getConstant(One.shl(ShAmt), dl, SVT));
18396 SDValue BV = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Elts);
18397 return DAG.getNode(ISD::MUL, dl, VT, R, BV);
18400 // Lower SHL with variable shift amount.
18401 if (VT == MVT::v4i32 && Op->getOpcode() == ISD::SHL) {
18402 Op = DAG.getNode(ISD::SHL, dl, VT, Amt, DAG.getConstant(23, dl, VT));
18404 Op = DAG.getNode(ISD::ADD, dl, VT, Op,
18405 DAG.getConstant(0x3f800000U, dl, VT));
18406 Op = DAG.getBitcast(MVT::v4f32, Op);
18407 Op = DAG.getNode(ISD::FP_TO_SINT, dl, VT, Op);
18408 return DAG.getNode(ISD::MUL, dl, VT, Op, R);
18411 // If possible, lower this shift as a sequence of two shifts by
18412 // constant plus a MOVSS/MOVSD instead of scalarizing it.
18414 // (v4i32 (srl A, (build_vector < X, Y, Y, Y>)))
18416 // Could be rewritten as:
18417 // (v4i32 (MOVSS (srl A, <Y,Y,Y,Y>), (srl A, <X,X,X,X>)))
18419 // The advantage is that the two shifts from the example would be
18420 // lowered as X86ISD::VSRLI nodes. This would be cheaper than scalarizing
18421 // the vector shift into four scalar shifts plus four pairs of vector
18423 if ((VT == MVT::v8i16 || VT == MVT::v4i32) &&
18424 ISD::isBuildVectorOfConstantSDNodes(Amt.getNode())) {
18425 unsigned TargetOpcode = X86ISD::MOVSS;
18426 bool CanBeSimplified;
18427 // The splat value for the first packed shift (the 'X' from the example).
18428 SDValue Amt1 = Amt->getOperand(0);
18429 // The splat value for the second packed shift (the 'Y' from the example).
18430 SDValue Amt2 = (VT == MVT::v4i32) ? Amt->getOperand(1) :
18431 Amt->getOperand(2);
18433 // See if it is possible to replace this node with a sequence of
18434 // two shifts followed by a MOVSS/MOVSD
18435 if (VT == MVT::v4i32) {
18436 // Check if it is legal to use a MOVSS.
18437 CanBeSimplified = Amt2 == Amt->getOperand(2) &&
18438 Amt2 == Amt->getOperand(3);
18439 if (!CanBeSimplified) {
18440 // Otherwise, check if we can still simplify this node using a MOVSD.
18441 CanBeSimplified = Amt1 == Amt->getOperand(1) &&
18442 Amt->getOperand(2) == Amt->getOperand(3);
18443 TargetOpcode = X86ISD::MOVSD;
18444 Amt2 = Amt->getOperand(2);
18447 // Do similar checks for the case where the machine value type
18449 CanBeSimplified = Amt1 == Amt->getOperand(1);
18450 for (unsigned i=3; i != 8 && CanBeSimplified; ++i)
18451 CanBeSimplified = Amt2 == Amt->getOperand(i);
18453 if (!CanBeSimplified) {
18454 TargetOpcode = X86ISD::MOVSD;
18455 CanBeSimplified = true;
18456 Amt2 = Amt->getOperand(4);
18457 for (unsigned i=0; i != 4 && CanBeSimplified; ++i)
18458 CanBeSimplified = Amt1 == Amt->getOperand(i);
18459 for (unsigned j=4; j != 8 && CanBeSimplified; ++j)
18460 CanBeSimplified = Amt2 == Amt->getOperand(j);
18464 if (CanBeSimplified && isa<ConstantSDNode>(Amt1) &&
18465 isa<ConstantSDNode>(Amt2)) {
18466 // Replace this node with two shifts followed by a MOVSS/MOVSD.
18467 EVT CastVT = MVT::v4i32;
18469 DAG.getConstant(cast<ConstantSDNode>(Amt1)->getAPIntValue(), dl, VT);
18470 SDValue Shift1 = DAG.getNode(Op->getOpcode(), dl, VT, R, Splat1);
18472 DAG.getConstant(cast<ConstantSDNode>(Amt2)->getAPIntValue(), dl, VT);
18473 SDValue Shift2 = DAG.getNode(Op->getOpcode(), dl, VT, R, Splat2);
18474 if (TargetOpcode == X86ISD::MOVSD)
18475 CastVT = MVT::v2i64;
18476 SDValue BitCast1 = DAG.getBitcast(CastVT, Shift1);
18477 SDValue BitCast2 = DAG.getBitcast(CastVT, Shift2);
18478 SDValue Result = getTargetShuffleNode(TargetOpcode, dl, CastVT, BitCast2,
18480 return DAG.getBitcast(VT, Result);
18484 // v4i32 Non Uniform Shifts.
18485 // If the shift amount is constant we can shift each lane using the SSE2
18486 // immediate shifts, else we need to zero-extend each lane to the lower i64
18487 // and shift using the SSE2 variable shifts.
18488 // The separate results can then be blended together.
18489 if (VT == MVT::v4i32) {
18490 unsigned Opc = Op.getOpcode();
18491 SDValue Amt0, Amt1, Amt2, Amt3;
18492 if (ISD::isBuildVectorOfConstantSDNodes(Amt.getNode())) {
18493 Amt0 = DAG.getVectorShuffle(VT, dl, Amt, DAG.getUNDEF(VT), {0, 0, 0, 0});
18494 Amt1 = DAG.getVectorShuffle(VT, dl, Amt, DAG.getUNDEF(VT), {1, 1, 1, 1});
18495 Amt2 = DAG.getVectorShuffle(VT, dl, Amt, DAG.getUNDEF(VT), {2, 2, 2, 2});
18496 Amt3 = DAG.getVectorShuffle(VT, dl, Amt, DAG.getUNDEF(VT), {3, 3, 3, 3});
18498 // ISD::SHL is handled above but we include it here for completeness.
18501 llvm_unreachable("Unknown target vector shift node");
18503 Opc = X86ISD::VSHL;
18506 Opc = X86ISD::VSRL;
18509 Opc = X86ISD::VSRA;
18512 // The SSE2 shifts use the lower i64 as the same shift amount for
18513 // all lanes and the upper i64 is ignored. These shuffle masks
18514 // optimally zero-extend each lanes on SSE2/SSE41/AVX targets.
18515 SDValue Z = getZeroVector(VT, Subtarget, DAG, dl);
18516 Amt0 = DAG.getVectorShuffle(VT, dl, Amt, Z, {0, 4, -1, -1});
18517 Amt1 = DAG.getVectorShuffle(VT, dl, Amt, Z, {1, 5, -1, -1});
18518 Amt2 = DAG.getVectorShuffle(VT, dl, Amt, Z, {2, 6, -1, -1});
18519 Amt3 = DAG.getVectorShuffle(VT, dl, Amt, Z, {3, 7, -1, -1});
18522 SDValue R0 = DAG.getNode(Opc, dl, VT, R, Amt0);
18523 SDValue R1 = DAG.getNode(Opc, dl, VT, R, Amt1);
18524 SDValue R2 = DAG.getNode(Opc, dl, VT, R, Amt2);
18525 SDValue R3 = DAG.getNode(Opc, dl, VT, R, Amt3);
18526 SDValue R02 = DAG.getVectorShuffle(VT, dl, R0, R2, {0, -1, 6, -1});
18527 SDValue R13 = DAG.getVectorShuffle(VT, dl, R1, R3, {-1, 1, -1, 7});
18528 return DAG.getVectorShuffle(VT, dl, R02, R13, {0, 5, 2, 7});
18531 if (VT == MVT::v16i8 ||
18532 (VT == MVT::v32i8 && Subtarget->hasInt256() && !Subtarget->hasXOP())) {
18533 MVT ExtVT = MVT::getVectorVT(MVT::i16, VT.getVectorNumElements() / 2);
18534 unsigned ShiftOpcode = Op->getOpcode();
18536 auto SignBitSelect = [&](MVT SelVT, SDValue Sel, SDValue V0, SDValue V1) {
18537 // On SSE41 targets we make use of the fact that VSELECT lowers
18538 // to PBLENDVB which selects bytes based just on the sign bit.
18539 if (Subtarget->hasSSE41()) {
18540 V0 = DAG.getBitcast(VT, V0);
18541 V1 = DAG.getBitcast(VT, V1);
18542 Sel = DAG.getBitcast(VT, Sel);
18543 return DAG.getBitcast(SelVT,
18544 DAG.getNode(ISD::VSELECT, dl, VT, Sel, V0, V1));
18546 // On pre-SSE41 targets we test for the sign bit by comparing to
18547 // zero - a negative value will set all bits of the lanes to true
18548 // and VSELECT uses that in its OR(AND(V0,C),AND(V1,~C)) lowering.
18549 SDValue Z = getZeroVector(SelVT, Subtarget, DAG, dl);
18550 SDValue C = DAG.getNode(X86ISD::PCMPGT, dl, SelVT, Z, Sel);
18551 return DAG.getNode(ISD::VSELECT, dl, SelVT, C, V0, V1);
18554 // Turn 'a' into a mask suitable for VSELECT: a = a << 5;
18555 // We can safely do this using i16 shifts as we're only interested in
18556 // the 3 lower bits of each byte.
18557 Amt = DAG.getBitcast(ExtVT, Amt);
18558 Amt = DAG.getNode(ISD::SHL, dl, ExtVT, Amt, DAG.getConstant(5, dl, ExtVT));
18559 Amt = DAG.getBitcast(VT, Amt);
18561 if (Op->getOpcode() == ISD::SHL || Op->getOpcode() == ISD::SRL) {
18562 // r = VSELECT(r, shift(r, 4), a);
18564 DAG.getNode(ShiftOpcode, dl, VT, R, DAG.getConstant(4, dl, VT));
18565 R = SignBitSelect(VT, Amt, M, R);
18568 Amt = DAG.getNode(ISD::ADD, dl, VT, Amt, Amt);
18570 // r = VSELECT(r, shift(r, 2), a);
18571 M = DAG.getNode(ShiftOpcode, dl, VT, R, DAG.getConstant(2, dl, VT));
18572 R = SignBitSelect(VT, Amt, M, R);
18575 Amt = DAG.getNode(ISD::ADD, dl, VT, Amt, Amt);
18577 // return VSELECT(r, shift(r, 1), a);
18578 M = DAG.getNode(ShiftOpcode, dl, VT, R, DAG.getConstant(1, dl, VT));
18579 R = SignBitSelect(VT, Amt, M, R);
18583 if (Op->getOpcode() == ISD::SRA) {
18584 // For SRA we need to unpack each byte to the higher byte of a i16 vector
18585 // so we can correctly sign extend. We don't care what happens to the
18587 SDValue ALo = DAG.getNode(X86ISD::UNPCKL, dl, VT, DAG.getUNDEF(VT), Amt);
18588 SDValue AHi = DAG.getNode(X86ISD::UNPCKH, dl, VT, DAG.getUNDEF(VT), Amt);
18589 SDValue RLo = DAG.getNode(X86ISD::UNPCKL, dl, VT, DAG.getUNDEF(VT), R);
18590 SDValue RHi = DAG.getNode(X86ISD::UNPCKH, dl, VT, DAG.getUNDEF(VT), R);
18591 ALo = DAG.getBitcast(ExtVT, ALo);
18592 AHi = DAG.getBitcast(ExtVT, AHi);
18593 RLo = DAG.getBitcast(ExtVT, RLo);
18594 RHi = DAG.getBitcast(ExtVT, RHi);
18596 // r = VSELECT(r, shift(r, 4), a);
18597 SDValue MLo = DAG.getNode(ShiftOpcode, dl, ExtVT, RLo,
18598 DAG.getConstant(4, dl, ExtVT));
18599 SDValue MHi = DAG.getNode(ShiftOpcode, dl, ExtVT, RHi,
18600 DAG.getConstant(4, dl, ExtVT));
18601 RLo = SignBitSelect(ExtVT, ALo, MLo, RLo);
18602 RHi = SignBitSelect(ExtVT, AHi, MHi, RHi);
18605 ALo = DAG.getNode(ISD::ADD, dl, ExtVT, ALo, ALo);
18606 AHi = DAG.getNode(ISD::ADD, dl, ExtVT, AHi, AHi);
18608 // r = VSELECT(r, shift(r, 2), a);
18609 MLo = DAG.getNode(ShiftOpcode, dl, ExtVT, RLo,
18610 DAG.getConstant(2, dl, ExtVT));
18611 MHi = DAG.getNode(ShiftOpcode, dl, ExtVT, RHi,
18612 DAG.getConstant(2, dl, ExtVT));
18613 RLo = SignBitSelect(ExtVT, ALo, MLo, RLo);
18614 RHi = SignBitSelect(ExtVT, AHi, MHi, RHi);
18617 ALo = DAG.getNode(ISD::ADD, dl, ExtVT, ALo, ALo);
18618 AHi = DAG.getNode(ISD::ADD, dl, ExtVT, AHi, AHi);
18620 // r = VSELECT(r, shift(r, 1), a);
18621 MLo = DAG.getNode(ShiftOpcode, dl, ExtVT, RLo,
18622 DAG.getConstant(1, dl, ExtVT));
18623 MHi = DAG.getNode(ShiftOpcode, dl, ExtVT, RHi,
18624 DAG.getConstant(1, dl, ExtVT));
18625 RLo = SignBitSelect(ExtVT, ALo, MLo, RLo);
18626 RHi = SignBitSelect(ExtVT, AHi, MHi, RHi);
18628 // Logical shift the result back to the lower byte, leaving a zero upper
18630 // meaning that we can safely pack with PACKUSWB.
18632 DAG.getNode(ISD::SRL, dl, ExtVT, RLo, DAG.getConstant(8, dl, ExtVT));
18634 DAG.getNode(ISD::SRL, dl, ExtVT, RHi, DAG.getConstant(8, dl, ExtVT));
18635 return DAG.getNode(X86ISD::PACKUS, dl, VT, RLo, RHi);
18639 // It's worth extending once and using the v8i32 shifts for 16-bit types, but
18640 // the extra overheads to get from v16i8 to v8i32 make the existing SSE
18641 // solution better.
18642 if (Subtarget->hasInt256() && VT == MVT::v8i16) {
18643 MVT ExtVT = MVT::v8i32;
18645 Op.getOpcode() == ISD::SRA ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND;
18646 R = DAG.getNode(ExtOpc, dl, ExtVT, R);
18647 Amt = DAG.getNode(ISD::ANY_EXTEND, dl, ExtVT, Amt);
18648 return DAG.getNode(ISD::TRUNCATE, dl, VT,
18649 DAG.getNode(Op.getOpcode(), dl, ExtVT, R, Amt));
18652 if (Subtarget->hasInt256() && !Subtarget->hasXOP() && VT == MVT::v16i16) {
18653 MVT ExtVT = MVT::v8i32;
18654 SDValue Z = getZeroVector(VT, Subtarget, DAG, dl);
18655 SDValue ALo = DAG.getNode(X86ISD::UNPCKL, dl, VT, Amt, Z);
18656 SDValue AHi = DAG.getNode(X86ISD::UNPCKH, dl, VT, Amt, Z);
18657 SDValue RLo = DAG.getNode(X86ISD::UNPCKL, dl, VT, R, R);
18658 SDValue RHi = DAG.getNode(X86ISD::UNPCKH, dl, VT, R, R);
18659 ALo = DAG.getBitcast(ExtVT, ALo);
18660 AHi = DAG.getBitcast(ExtVT, AHi);
18661 RLo = DAG.getBitcast(ExtVT, RLo);
18662 RHi = DAG.getBitcast(ExtVT, RHi);
18663 SDValue Lo = DAG.getNode(Op.getOpcode(), dl, ExtVT, RLo, ALo);
18664 SDValue Hi = DAG.getNode(Op.getOpcode(), dl, ExtVT, RHi, AHi);
18665 Lo = DAG.getNode(ISD::SRL, dl, ExtVT, Lo, DAG.getConstant(16, dl, ExtVT));
18666 Hi = DAG.getNode(ISD::SRL, dl, ExtVT, Hi, DAG.getConstant(16, dl, ExtVT));
18667 return DAG.getNode(X86ISD::PACKUS, dl, VT, Lo, Hi);
18670 if (VT == MVT::v8i16) {
18671 unsigned ShiftOpcode = Op->getOpcode();
18673 auto SignBitSelect = [&](SDValue Sel, SDValue V0, SDValue V1) {
18674 // On SSE41 targets we make use of the fact that VSELECT lowers
18675 // to PBLENDVB which selects bytes based just on the sign bit.
18676 if (Subtarget->hasSSE41()) {
18677 MVT ExtVT = MVT::getVectorVT(MVT::i8, VT.getVectorNumElements() * 2);
18678 V0 = DAG.getBitcast(ExtVT, V0);
18679 V1 = DAG.getBitcast(ExtVT, V1);
18680 Sel = DAG.getBitcast(ExtVT, Sel);
18681 return DAG.getBitcast(
18682 VT, DAG.getNode(ISD::VSELECT, dl, ExtVT, Sel, V0, V1));
18684 // On pre-SSE41 targets we splat the sign bit - a negative value will
18685 // set all bits of the lanes to true and VSELECT uses that in
18686 // its OR(AND(V0,C),AND(V1,~C)) lowering.
18688 DAG.getNode(ISD::SRA, dl, VT, Sel, DAG.getConstant(15, dl, VT));
18689 return DAG.getNode(ISD::VSELECT, dl, VT, C, V0, V1);
18692 // Turn 'a' into a mask suitable for VSELECT: a = a << 12;
18693 if (Subtarget->hasSSE41()) {
18694 // On SSE41 targets we need to replicate the shift mask in both
18695 // bytes for PBLENDVB.
18698 DAG.getNode(ISD::SHL, dl, VT, Amt, DAG.getConstant(4, dl, VT)),
18699 DAG.getNode(ISD::SHL, dl, VT, Amt, DAG.getConstant(12, dl, VT)));
18701 Amt = DAG.getNode(ISD::SHL, dl, VT, Amt, DAG.getConstant(12, dl, VT));
18704 // r = VSELECT(r, shift(r, 8), a);
18705 SDValue M = DAG.getNode(ShiftOpcode, dl, VT, R, DAG.getConstant(8, dl, VT));
18706 R = SignBitSelect(Amt, M, R);
18709 Amt = DAG.getNode(ISD::ADD, dl, VT, Amt, Amt);
18711 // r = VSELECT(r, shift(r, 4), a);
18712 M = DAG.getNode(ShiftOpcode, dl, VT, R, DAG.getConstant(4, dl, VT));
18713 R = SignBitSelect(Amt, M, R);
18716 Amt = DAG.getNode(ISD::ADD, dl, VT, Amt, Amt);
18718 // r = VSELECT(r, shift(r, 2), a);
18719 M = DAG.getNode(ShiftOpcode, dl, VT, R, DAG.getConstant(2, dl, VT));
18720 R = SignBitSelect(Amt, M, R);
18723 Amt = DAG.getNode(ISD::ADD, dl, VT, Amt, Amt);
18725 // return VSELECT(r, shift(r, 1), a);
18726 M = DAG.getNode(ShiftOpcode, dl, VT, R, DAG.getConstant(1, dl, VT));
18727 R = SignBitSelect(Amt, M, R);
18731 // Decompose 256-bit shifts into smaller 128-bit shifts.
18732 if (VT.is256BitVector()) {
18733 unsigned NumElems = VT.getVectorNumElements();
18734 MVT EltVT = VT.getVectorElementType();
18735 EVT NewVT = MVT::getVectorVT(EltVT, NumElems/2);
18737 // Extract the two vectors
18738 SDValue V1 = Extract128BitVector(R, 0, DAG, dl);
18739 SDValue V2 = Extract128BitVector(R, NumElems/2, DAG, dl);
18741 // Recreate the shift amount vectors
18742 SDValue Amt1, Amt2;
18743 if (Amt.getOpcode() == ISD::BUILD_VECTOR) {
18744 // Constant shift amount
18745 SmallVector<SDValue, 8> Ops(Amt->op_begin(), Amt->op_begin() + NumElems);
18746 ArrayRef<SDValue> Amt1Csts = makeArrayRef(Ops).slice(0, NumElems / 2);
18747 ArrayRef<SDValue> Amt2Csts = makeArrayRef(Ops).slice(NumElems / 2);
18749 Amt1 = DAG.getNode(ISD::BUILD_VECTOR, dl, NewVT, Amt1Csts);
18750 Amt2 = DAG.getNode(ISD::BUILD_VECTOR, dl, NewVT, Amt2Csts);
18752 // Variable shift amount
18753 Amt1 = Extract128BitVector(Amt, 0, DAG, dl);
18754 Amt2 = Extract128BitVector(Amt, NumElems/2, DAG, dl);
18757 // Issue new vector shifts for the smaller types
18758 V1 = DAG.getNode(Op.getOpcode(), dl, NewVT, V1, Amt1);
18759 V2 = DAG.getNode(Op.getOpcode(), dl, NewVT, V2, Amt2);
18761 // Concatenate the result back
18762 return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT, V1, V2);
18768 static SDValue LowerRotate(SDValue Op, const X86Subtarget *Subtarget,
18769 SelectionDAG &DAG) {
18770 MVT VT = Op.getSimpleValueType();
18772 SDValue R = Op.getOperand(0);
18773 SDValue Amt = Op.getOperand(1);
18775 assert(VT.isVector() && "Custom lowering only for vector rotates!");
18776 assert(Subtarget->hasXOP() && "XOP support required for vector rotates!");
18777 assert((Op.getOpcode() == ISD::ROTL) && "Only ROTL supported");
18779 // XOP has 128-bit vector variable + immediate rotates.
18780 // +ve/-ve Amt = rotate left/right.
18782 // Split 256-bit integers.
18783 if (VT.getSizeInBits() == 256)
18784 return Lower256IntArith(Op, DAG);
18786 assert(VT.getSizeInBits() == 128 && "Only rotate 128-bit vectors!");
18788 // Attempt to rotate by immediate.
18789 if (auto *BVAmt = dyn_cast<BuildVectorSDNode>(Amt)) {
18790 if (auto *RotateConst = BVAmt->getConstantSplatNode()) {
18791 uint64_t RotateAmt = RotateConst->getAPIntValue().getZExtValue();
18792 assert(RotateAmt < VT.getScalarSizeInBits() && "Rotation out of range");
18793 return DAG.getNode(X86ISD::VPROTI, DL, VT, R,
18794 DAG.getConstant(RotateAmt, DL, MVT::i8));
18798 // Use general rotate by variable (per-element).
18799 return DAG.getNode(X86ISD::VPROT, DL, VT, R, Amt);
18802 static SDValue LowerXALUO(SDValue Op, SelectionDAG &DAG) {
18803 // Lower the "add/sub/mul with overflow" instruction into a regular ins plus
18804 // a "setcc" instruction that checks the overflow flag. The "brcond" lowering
18805 // looks for this combo and may remove the "setcc" instruction if the "setcc"
18806 // has only one use.
18807 SDNode *N = Op.getNode();
18808 SDValue LHS = N->getOperand(0);
18809 SDValue RHS = N->getOperand(1);
18810 unsigned BaseOp = 0;
18813 switch (Op.getOpcode()) {
18814 default: llvm_unreachable("Unknown ovf instruction!");
18816 // A subtract of one will be selected as a INC. Note that INC doesn't
18817 // set CF, so we can't do this for UADDO.
18818 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(RHS))
18820 BaseOp = X86ISD::INC;
18821 Cond = X86::COND_O;
18824 BaseOp = X86ISD::ADD;
18825 Cond = X86::COND_O;
18828 BaseOp = X86ISD::ADD;
18829 Cond = X86::COND_B;
18832 // A subtract of one will be selected as a DEC. Note that DEC doesn't
18833 // set CF, so we can't do this for USUBO.
18834 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(RHS))
18836 BaseOp = X86ISD::DEC;
18837 Cond = X86::COND_O;
18840 BaseOp = X86ISD::SUB;
18841 Cond = X86::COND_O;
18844 BaseOp = X86ISD::SUB;
18845 Cond = X86::COND_B;
18848 BaseOp = N->getValueType(0) == MVT::i8 ? X86ISD::SMUL8 : X86ISD::SMUL;
18849 Cond = X86::COND_O;
18851 case ISD::UMULO: { // i64, i8 = umulo lhs, rhs --> i64, i64, i32 umul lhs,rhs
18852 if (N->getValueType(0) == MVT::i8) {
18853 BaseOp = X86ISD::UMUL8;
18854 Cond = X86::COND_O;
18857 SDVTList VTs = DAG.getVTList(N->getValueType(0), N->getValueType(0),
18859 SDValue Sum = DAG.getNode(X86ISD::UMUL, DL, VTs, LHS, RHS);
18862 DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
18863 DAG.getConstant(X86::COND_O, DL, MVT::i32),
18864 SDValue(Sum.getNode(), 2));
18866 return DAG.getNode(ISD::MERGE_VALUES, DL, N->getVTList(), Sum, SetCC);
18870 // Also sets EFLAGS.
18871 SDVTList VTs = DAG.getVTList(N->getValueType(0), MVT::i32);
18872 SDValue Sum = DAG.getNode(BaseOp, DL, VTs, LHS, RHS);
18875 DAG.getNode(X86ISD::SETCC, DL, N->getValueType(1),
18876 DAG.getConstant(Cond, DL, MVT::i32),
18877 SDValue(Sum.getNode(), 1));
18879 return DAG.getNode(ISD::MERGE_VALUES, DL, N->getVTList(), Sum, SetCC);
18882 /// Returns true if the operand type is exactly twice the native width, and
18883 /// the corresponding cmpxchg8b or cmpxchg16b instruction is available.
18884 /// Used to know whether to use cmpxchg8/16b when expanding atomic operations
18885 /// (otherwise we leave them alone to become __sync_fetch_and_... calls).
18886 bool X86TargetLowering::needsCmpXchgNb(Type *MemType) const {
18887 unsigned OpWidth = MemType->getPrimitiveSizeInBits();
18890 return !Subtarget->is64Bit(); // FIXME this should be Subtarget.hasCmpxchg8b
18891 else if (OpWidth == 128)
18892 return Subtarget->hasCmpxchg16b();
18897 bool X86TargetLowering::shouldExpandAtomicStoreInIR(StoreInst *SI) const {
18898 return needsCmpXchgNb(SI->getValueOperand()->getType());
18901 // Note: this turns large loads into lock cmpxchg8b/16b.
18902 // FIXME: On 32 bits x86, fild/movq might be faster than lock cmpxchg8b.
18903 TargetLowering::AtomicExpansionKind
18904 X86TargetLowering::shouldExpandAtomicLoadInIR(LoadInst *LI) const {
18905 auto PTy = cast<PointerType>(LI->getPointerOperand()->getType());
18906 return needsCmpXchgNb(PTy->getElementType()) ? AtomicExpansionKind::CmpXChg
18907 : AtomicExpansionKind::None;
18910 TargetLowering::AtomicExpansionKind
18911 X86TargetLowering::shouldExpandAtomicRMWInIR(AtomicRMWInst *AI) const {
18912 unsigned NativeWidth = Subtarget->is64Bit() ? 64 : 32;
18913 Type *MemType = AI->getType();
18915 // If the operand is too big, we must see if cmpxchg8/16b is available
18916 // and default to library calls otherwise.
18917 if (MemType->getPrimitiveSizeInBits() > NativeWidth) {
18918 return needsCmpXchgNb(MemType) ? AtomicExpansionKind::CmpXChg
18919 : AtomicExpansionKind::None;
18922 AtomicRMWInst::BinOp Op = AI->getOperation();
18925 llvm_unreachable("Unknown atomic operation");
18926 case AtomicRMWInst::Xchg:
18927 case AtomicRMWInst::Add:
18928 case AtomicRMWInst::Sub:
18929 // It's better to use xadd, xsub or xchg for these in all cases.
18930 return AtomicExpansionKind::None;
18931 case AtomicRMWInst::Or:
18932 case AtomicRMWInst::And:
18933 case AtomicRMWInst::Xor:
18934 // If the atomicrmw's result isn't actually used, we can just add a "lock"
18935 // prefix to a normal instruction for these operations.
18936 return !AI->use_empty() ? AtomicExpansionKind::CmpXChg
18937 : AtomicExpansionKind::None;
18938 case AtomicRMWInst::Nand:
18939 case AtomicRMWInst::Max:
18940 case AtomicRMWInst::Min:
18941 case AtomicRMWInst::UMax:
18942 case AtomicRMWInst::UMin:
18943 // These always require a non-trivial set of data operations on x86. We must
18944 // use a cmpxchg loop.
18945 return AtomicExpansionKind::CmpXChg;
18949 static bool hasMFENCE(const X86Subtarget& Subtarget) {
18950 // Use mfence if we have SSE2 or we're on x86-64 (even if we asked for
18951 // no-sse2). There isn't any reason to disable it if the target processor
18953 return Subtarget.hasSSE2() || Subtarget.is64Bit();
18957 X86TargetLowering::lowerIdempotentRMWIntoFencedLoad(AtomicRMWInst *AI) const {
18958 unsigned NativeWidth = Subtarget->is64Bit() ? 64 : 32;
18959 Type *MemType = AI->getType();
18960 // Accesses larger than the native width are turned into cmpxchg/libcalls, so
18961 // there is no benefit in turning such RMWs into loads, and it is actually
18962 // harmful as it introduces a mfence.
18963 if (MemType->getPrimitiveSizeInBits() > NativeWidth)
18966 auto Builder = IRBuilder<>(AI);
18967 Module *M = Builder.GetInsertBlock()->getParent()->getParent();
18968 auto SynchScope = AI->getSynchScope();
18969 // We must restrict the ordering to avoid generating loads with Release or
18970 // ReleaseAcquire orderings.
18971 auto Order = AtomicCmpXchgInst::getStrongestFailureOrdering(AI->getOrdering());
18972 auto Ptr = AI->getPointerOperand();
18974 // Before the load we need a fence. Here is an example lifted from
18975 // http://www.hpl.hp.com/techreports/2012/HPL-2012-68.pdf showing why a fence
18978 // x.store(1, relaxed);
18979 // r1 = y.fetch_add(0, release);
18981 // y.fetch_add(42, acquire);
18982 // r2 = x.load(relaxed);
18983 // r1 = r2 = 0 is impossible, but becomes possible if the idempotent rmw is
18984 // lowered to just a load without a fence. A mfence flushes the store buffer,
18985 // making the optimization clearly correct.
18986 // FIXME: it is required if isAtLeastRelease(Order) but it is not clear
18987 // otherwise, we might be able to be more aggressive on relaxed idempotent
18988 // rmw. In practice, they do not look useful, so we don't try to be
18989 // especially clever.
18990 if (SynchScope == SingleThread)
18991 // FIXME: we could just insert an X86ISD::MEMBARRIER here, except we are at
18992 // the IR level, so we must wrap it in an intrinsic.
18995 if (!hasMFENCE(*Subtarget))
18996 // FIXME: it might make sense to use a locked operation here but on a
18997 // different cache-line to prevent cache-line bouncing. In practice it
18998 // is probably a small win, and x86 processors without mfence are rare
18999 // enough that we do not bother.
19003 llvm::Intrinsic::getDeclaration(M, Intrinsic::x86_sse2_mfence);
19004 Builder.CreateCall(MFence, {});
19006 // Finally we can emit the atomic load.
19007 LoadInst *Loaded = Builder.CreateAlignedLoad(Ptr,
19008 AI->getType()->getPrimitiveSizeInBits());
19009 Loaded->setAtomic(Order, SynchScope);
19010 AI->replaceAllUsesWith(Loaded);
19011 AI->eraseFromParent();
19015 static SDValue LowerATOMIC_FENCE(SDValue Op, const X86Subtarget *Subtarget,
19016 SelectionDAG &DAG) {
19018 AtomicOrdering FenceOrdering = static_cast<AtomicOrdering>(
19019 cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue());
19020 SynchronizationScope FenceScope = static_cast<SynchronizationScope>(
19021 cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue());
19023 // The only fence that needs an instruction is a sequentially-consistent
19024 // cross-thread fence.
19025 if (FenceOrdering == SequentiallyConsistent && FenceScope == CrossThread) {
19026 if (hasMFENCE(*Subtarget))
19027 return DAG.getNode(X86ISD::MFENCE, dl, MVT::Other, Op.getOperand(0));
19029 SDValue Chain = Op.getOperand(0);
19030 SDValue Zero = DAG.getConstant(0, dl, MVT::i32);
19032 DAG.getRegister(X86::ESP, MVT::i32), // Base
19033 DAG.getTargetConstant(1, dl, MVT::i8), // Scale
19034 DAG.getRegister(0, MVT::i32), // Index
19035 DAG.getTargetConstant(0, dl, MVT::i32), // Disp
19036 DAG.getRegister(0, MVT::i32), // Segment.
19040 SDNode *Res = DAG.getMachineNode(X86::OR32mrLocked, dl, MVT::Other, Ops);
19041 return SDValue(Res, 0);
19044 // MEMBARRIER is a compiler barrier; it codegens to a no-op.
19045 return DAG.getNode(X86ISD::MEMBARRIER, dl, MVT::Other, Op.getOperand(0));
19048 static SDValue LowerCMP_SWAP(SDValue Op, const X86Subtarget *Subtarget,
19049 SelectionDAG &DAG) {
19050 MVT T = Op.getSimpleValueType();
19054 switch(T.SimpleTy) {
19055 default: llvm_unreachable("Invalid value type!");
19056 case MVT::i8: Reg = X86::AL; size = 1; break;
19057 case MVT::i16: Reg = X86::AX; size = 2; break;
19058 case MVT::i32: Reg = X86::EAX; size = 4; break;
19060 assert(Subtarget->is64Bit() && "Node not type legal!");
19061 Reg = X86::RAX; size = 8;
19064 SDValue cpIn = DAG.getCopyToReg(Op.getOperand(0), DL, Reg,
19065 Op.getOperand(2), SDValue());
19066 SDValue Ops[] = { cpIn.getValue(0),
19069 DAG.getTargetConstant(size, DL, MVT::i8),
19070 cpIn.getValue(1) };
19071 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Glue);
19072 MachineMemOperand *MMO = cast<AtomicSDNode>(Op)->getMemOperand();
19073 SDValue Result = DAG.getMemIntrinsicNode(X86ISD::LCMPXCHG_DAG, DL, Tys,
19077 DAG.getCopyFromReg(Result.getValue(0), DL, Reg, T, Result.getValue(1));
19078 SDValue EFLAGS = DAG.getCopyFromReg(cpOut.getValue(1), DL, X86::EFLAGS,
19079 MVT::i32, cpOut.getValue(2));
19080 SDValue Success = DAG.getNode(X86ISD::SETCC, DL, Op->getValueType(1),
19081 DAG.getConstant(X86::COND_E, DL, MVT::i8),
19084 DAG.ReplaceAllUsesOfValueWith(Op.getValue(0), cpOut);
19085 DAG.ReplaceAllUsesOfValueWith(Op.getValue(1), Success);
19086 DAG.ReplaceAllUsesOfValueWith(Op.getValue(2), EFLAGS.getValue(1));
19090 static SDValue LowerBITCAST(SDValue Op, const X86Subtarget *Subtarget,
19091 SelectionDAG &DAG) {
19092 MVT SrcVT = Op.getOperand(0).getSimpleValueType();
19093 MVT DstVT = Op.getSimpleValueType();
19095 if (SrcVT == MVT::v2i32 || SrcVT == MVT::v4i16 || SrcVT == MVT::v8i8) {
19096 assert(Subtarget->hasSSE2() && "Requires at least SSE2!");
19097 if (DstVT != MVT::f64)
19098 // This conversion needs to be expanded.
19101 SDValue InVec = Op->getOperand(0);
19103 unsigned NumElts = SrcVT.getVectorNumElements();
19104 EVT SVT = SrcVT.getVectorElementType();
19106 // Widen the vector in input in the case of MVT::v2i32.
19107 // Example: from MVT::v2i32 to MVT::v4i32.
19108 SmallVector<SDValue, 16> Elts;
19109 for (unsigned i = 0, e = NumElts; i != e; ++i)
19110 Elts.push_back(DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, SVT, InVec,
19111 DAG.getIntPtrConstant(i, dl)));
19113 // Explicitly mark the extra elements as Undef.
19114 Elts.append(NumElts, DAG.getUNDEF(SVT));
19116 EVT NewVT = EVT::getVectorVT(*DAG.getContext(), SVT, NumElts * 2);
19117 SDValue BV = DAG.getNode(ISD::BUILD_VECTOR, dl, NewVT, Elts);
19118 SDValue ToV2F64 = DAG.getBitcast(MVT::v2f64, BV);
19119 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, ToV2F64,
19120 DAG.getIntPtrConstant(0, dl));
19123 assert(Subtarget->is64Bit() && !Subtarget->hasSSE2() &&
19124 Subtarget->hasMMX() && "Unexpected custom BITCAST");
19125 assert((DstVT == MVT::i64 ||
19126 (DstVT.isVector() && DstVT.getSizeInBits()==64)) &&
19127 "Unexpected custom BITCAST");
19128 // i64 <=> MMX conversions are Legal.
19129 if (SrcVT==MVT::i64 && DstVT.isVector())
19131 if (DstVT==MVT::i64 && SrcVT.isVector())
19133 // MMX <=> MMX conversions are Legal.
19134 if (SrcVT.isVector() && DstVT.isVector())
19136 // All other conversions need to be expanded.
19140 /// Compute the horizontal sum of bytes in V for the elements of VT.
19142 /// Requires V to be a byte vector and VT to be an integer vector type with
19143 /// wider elements than V's type. The width of the elements of VT determines
19144 /// how many bytes of V are summed horizontally to produce each element of the
19146 static SDValue LowerHorizontalByteSum(SDValue V, MVT VT,
19147 const X86Subtarget *Subtarget,
19148 SelectionDAG &DAG) {
19150 MVT ByteVecVT = V.getSimpleValueType();
19151 MVT EltVT = VT.getVectorElementType();
19152 int NumElts = VT.getVectorNumElements();
19153 assert(ByteVecVT.getVectorElementType() == MVT::i8 &&
19154 "Expected value to have byte element type.");
19155 assert(EltVT != MVT::i8 &&
19156 "Horizontal byte sum only makes sense for wider elements!");
19157 unsigned VecSize = VT.getSizeInBits();
19158 assert(ByteVecVT.getSizeInBits() == VecSize && "Cannot change vector size!");
19160 // PSADBW instruction horizontally add all bytes and leave the result in i64
19161 // chunks, thus directly computes the pop count for v2i64 and v4i64.
19162 if (EltVT == MVT::i64) {
19163 SDValue Zeros = getZeroVector(ByteVecVT, Subtarget, DAG, DL);
19164 V = DAG.getNode(X86ISD::PSADBW, DL, ByteVecVT, V, Zeros);
19165 return DAG.getBitcast(VT, V);
19168 if (EltVT == MVT::i32) {
19169 // We unpack the low half and high half into i32s interleaved with zeros so
19170 // that we can use PSADBW to horizontally sum them. The most useful part of
19171 // this is that it lines up the results of two PSADBW instructions to be
19172 // two v2i64 vectors which concatenated are the 4 population counts. We can
19173 // then use PACKUSWB to shrink and concatenate them into a v4i32 again.
19174 SDValue Zeros = getZeroVector(VT, Subtarget, DAG, DL);
19175 SDValue Low = DAG.getNode(X86ISD::UNPCKL, DL, VT, V, Zeros);
19176 SDValue High = DAG.getNode(X86ISD::UNPCKH, DL, VT, V, Zeros);
19178 // Do the horizontal sums into two v2i64s.
19179 Zeros = getZeroVector(ByteVecVT, Subtarget, DAG, DL);
19180 Low = DAG.getNode(X86ISD::PSADBW, DL, ByteVecVT,
19181 DAG.getBitcast(ByteVecVT, Low), Zeros);
19182 High = DAG.getNode(X86ISD::PSADBW, DL, ByteVecVT,
19183 DAG.getBitcast(ByteVecVT, High), Zeros);
19185 // Merge them together.
19186 MVT ShortVecVT = MVT::getVectorVT(MVT::i16, VecSize / 16);
19187 V = DAG.getNode(X86ISD::PACKUS, DL, ByteVecVT,
19188 DAG.getBitcast(ShortVecVT, Low),
19189 DAG.getBitcast(ShortVecVT, High));
19191 return DAG.getBitcast(VT, V);
19194 // The only element type left is i16.
19195 assert(EltVT == MVT::i16 && "Unknown how to handle type");
19197 // To obtain pop count for each i16 element starting from the pop count for
19198 // i8 elements, shift the i16s left by 8, sum as i8s, and then shift as i16s
19199 // right by 8. It is important to shift as i16s as i8 vector shift isn't
19200 // directly supported.
19201 SmallVector<SDValue, 16> Shifters(NumElts, DAG.getConstant(8, DL, EltVT));
19202 SDValue Shifter = DAG.getNode(ISD::BUILD_VECTOR, DL, VT, Shifters);
19203 SDValue Shl = DAG.getNode(ISD::SHL, DL, VT, DAG.getBitcast(VT, V), Shifter);
19204 V = DAG.getNode(ISD::ADD, DL, ByteVecVT, DAG.getBitcast(ByteVecVT, Shl),
19205 DAG.getBitcast(ByteVecVT, V));
19206 return DAG.getNode(ISD::SRL, DL, VT, DAG.getBitcast(VT, V), Shifter);
19209 static SDValue LowerVectorCTPOPInRegLUT(SDValue Op, SDLoc DL,
19210 const X86Subtarget *Subtarget,
19211 SelectionDAG &DAG) {
19212 MVT VT = Op.getSimpleValueType();
19213 MVT EltVT = VT.getVectorElementType();
19214 unsigned VecSize = VT.getSizeInBits();
19216 // Implement a lookup table in register by using an algorithm based on:
19217 // http://wm.ite.pl/articles/sse-popcount.html
19219 // The general idea is that every lower byte nibble in the input vector is an
19220 // index into a in-register pre-computed pop count table. We then split up the
19221 // input vector in two new ones: (1) a vector with only the shifted-right
19222 // higher nibbles for each byte and (2) a vector with the lower nibbles (and
19223 // masked out higher ones) for each byte. PSHUB is used separately with both
19224 // to index the in-register table. Next, both are added and the result is a
19225 // i8 vector where each element contains the pop count for input byte.
19227 // To obtain the pop count for elements != i8, we follow up with the same
19228 // approach and use additional tricks as described below.
19230 const int LUT[16] = {/* 0 */ 0, /* 1 */ 1, /* 2 */ 1, /* 3 */ 2,
19231 /* 4 */ 1, /* 5 */ 2, /* 6 */ 2, /* 7 */ 3,
19232 /* 8 */ 1, /* 9 */ 2, /* a */ 2, /* b */ 3,
19233 /* c */ 2, /* d */ 3, /* e */ 3, /* f */ 4};
19235 int NumByteElts = VecSize / 8;
19236 MVT ByteVecVT = MVT::getVectorVT(MVT::i8, NumByteElts);
19237 SDValue In = DAG.getBitcast(ByteVecVT, Op);
19238 SmallVector<SDValue, 16> LUTVec;
19239 for (int i = 0; i < NumByteElts; ++i)
19240 LUTVec.push_back(DAG.getConstant(LUT[i % 16], DL, MVT::i8));
19241 SDValue InRegLUT = DAG.getNode(ISD::BUILD_VECTOR, DL, ByteVecVT, LUTVec);
19242 SmallVector<SDValue, 16> Mask0F(NumByteElts,
19243 DAG.getConstant(0x0F, DL, MVT::i8));
19244 SDValue M0F = DAG.getNode(ISD::BUILD_VECTOR, DL, ByteVecVT, Mask0F);
19247 SmallVector<SDValue, 16> Four(NumByteElts, DAG.getConstant(4, DL, MVT::i8));
19248 SDValue FourV = DAG.getNode(ISD::BUILD_VECTOR, DL, ByteVecVT, Four);
19249 SDValue HighNibbles = DAG.getNode(ISD::SRL, DL, ByteVecVT, In, FourV);
19252 SDValue LowNibbles = DAG.getNode(ISD::AND, DL, ByteVecVT, In, M0F);
19254 // The input vector is used as the shuffle mask that index elements into the
19255 // LUT. After counting low and high nibbles, add the vector to obtain the
19256 // final pop count per i8 element.
19257 SDValue HighPopCnt =
19258 DAG.getNode(X86ISD::PSHUFB, DL, ByteVecVT, InRegLUT, HighNibbles);
19259 SDValue LowPopCnt =
19260 DAG.getNode(X86ISD::PSHUFB, DL, ByteVecVT, InRegLUT, LowNibbles);
19261 SDValue PopCnt = DAG.getNode(ISD::ADD, DL, ByteVecVT, HighPopCnt, LowPopCnt);
19263 if (EltVT == MVT::i8)
19266 return LowerHorizontalByteSum(PopCnt, VT, Subtarget, DAG);
19269 static SDValue LowerVectorCTPOPBitmath(SDValue Op, SDLoc DL,
19270 const X86Subtarget *Subtarget,
19271 SelectionDAG &DAG) {
19272 MVT VT = Op.getSimpleValueType();
19273 assert(VT.is128BitVector() &&
19274 "Only 128-bit vector bitmath lowering supported.");
19276 int VecSize = VT.getSizeInBits();
19277 MVT EltVT = VT.getVectorElementType();
19278 int Len = EltVT.getSizeInBits();
19280 // This is the vectorized version of the "best" algorithm from
19281 // http://graphics.stanford.edu/~seander/bithacks.html#CountBitsSetParallel
19282 // with a minor tweak to use a series of adds + shifts instead of vector
19283 // multiplications. Implemented for all integer vector types. We only use
19284 // this when we don't have SSSE3 which allows a LUT-based lowering that is
19285 // much faster, even faster than using native popcnt instructions.
19287 auto GetShift = [&](unsigned OpCode, SDValue V, int Shifter) {
19288 MVT VT = V.getSimpleValueType();
19289 SmallVector<SDValue, 32> Shifters(
19290 VT.getVectorNumElements(),
19291 DAG.getConstant(Shifter, DL, VT.getVectorElementType()));
19292 return DAG.getNode(OpCode, DL, VT, V,
19293 DAG.getNode(ISD::BUILD_VECTOR, DL, VT, Shifters));
19295 auto GetMask = [&](SDValue V, APInt Mask) {
19296 MVT VT = V.getSimpleValueType();
19297 SmallVector<SDValue, 32> Masks(
19298 VT.getVectorNumElements(),
19299 DAG.getConstant(Mask, DL, VT.getVectorElementType()));
19300 return DAG.getNode(ISD::AND, DL, VT, V,
19301 DAG.getNode(ISD::BUILD_VECTOR, DL, VT, Masks));
19304 // We don't want to incur the implicit masks required to SRL vNi8 vectors on
19305 // x86, so set the SRL type to have elements at least i16 wide. This is
19306 // correct because all of our SRLs are followed immediately by a mask anyways
19307 // that handles any bits that sneak into the high bits of the byte elements.
19308 MVT SrlVT = Len > 8 ? VT : MVT::getVectorVT(MVT::i16, VecSize / 16);
19312 // v = v - ((v >> 1) & 0x55555555...)
19314 DAG.getBitcast(VT, GetShift(ISD::SRL, DAG.getBitcast(SrlVT, V), 1));
19315 SDValue And = GetMask(Srl, APInt::getSplat(Len, APInt(8, 0x55)));
19316 V = DAG.getNode(ISD::SUB, DL, VT, V, And);
19318 // v = (v & 0x33333333...) + ((v >> 2) & 0x33333333...)
19319 SDValue AndLHS = GetMask(V, APInt::getSplat(Len, APInt(8, 0x33)));
19320 Srl = DAG.getBitcast(VT, GetShift(ISD::SRL, DAG.getBitcast(SrlVT, V), 2));
19321 SDValue AndRHS = GetMask(Srl, APInt::getSplat(Len, APInt(8, 0x33)));
19322 V = DAG.getNode(ISD::ADD, DL, VT, AndLHS, AndRHS);
19324 // v = (v + (v >> 4)) & 0x0F0F0F0F...
19325 Srl = DAG.getBitcast(VT, GetShift(ISD::SRL, DAG.getBitcast(SrlVT, V), 4));
19326 SDValue Add = DAG.getNode(ISD::ADD, DL, VT, V, Srl);
19327 V = GetMask(Add, APInt::getSplat(Len, APInt(8, 0x0F)));
19329 // At this point, V contains the byte-wise population count, and we are
19330 // merely doing a horizontal sum if necessary to get the wider element
19332 if (EltVT == MVT::i8)
19335 return LowerHorizontalByteSum(
19336 DAG.getBitcast(MVT::getVectorVT(MVT::i8, VecSize / 8), V), VT, Subtarget,
19340 static SDValue LowerVectorCTPOP(SDValue Op, const X86Subtarget *Subtarget,
19341 SelectionDAG &DAG) {
19342 MVT VT = Op.getSimpleValueType();
19343 // FIXME: Need to add AVX-512 support here!
19344 assert((VT.is256BitVector() || VT.is128BitVector()) &&
19345 "Unknown CTPOP type to handle");
19346 SDLoc DL(Op.getNode());
19347 SDValue Op0 = Op.getOperand(0);
19349 if (!Subtarget->hasSSSE3()) {
19350 // We can't use the fast LUT approach, so fall back on vectorized bitmath.
19351 assert(VT.is128BitVector() && "Only 128-bit vectors supported in SSE!");
19352 return LowerVectorCTPOPBitmath(Op0, DL, Subtarget, DAG);
19355 if (VT.is256BitVector() && !Subtarget->hasInt256()) {
19356 unsigned NumElems = VT.getVectorNumElements();
19358 // Extract each 128-bit vector, compute pop count and concat the result.
19359 SDValue LHS = Extract128BitVector(Op0, 0, DAG, DL);
19360 SDValue RHS = Extract128BitVector(Op0, NumElems/2, DAG, DL);
19362 return DAG.getNode(ISD::CONCAT_VECTORS, DL, VT,
19363 LowerVectorCTPOPInRegLUT(LHS, DL, Subtarget, DAG),
19364 LowerVectorCTPOPInRegLUT(RHS, DL, Subtarget, DAG));
19367 return LowerVectorCTPOPInRegLUT(Op0, DL, Subtarget, DAG);
19370 static SDValue LowerCTPOP(SDValue Op, const X86Subtarget *Subtarget,
19371 SelectionDAG &DAG) {
19372 assert(Op.getValueType().isVector() &&
19373 "We only do custom lowering for vector population count.");
19374 return LowerVectorCTPOP(Op, Subtarget, DAG);
19377 static SDValue LowerLOAD_SUB(SDValue Op, SelectionDAG &DAG) {
19378 SDNode *Node = Op.getNode();
19380 EVT T = Node->getValueType(0);
19381 SDValue negOp = DAG.getNode(ISD::SUB, dl, T,
19382 DAG.getConstant(0, dl, T), Node->getOperand(2));
19383 return DAG.getAtomic(ISD::ATOMIC_LOAD_ADD, dl,
19384 cast<AtomicSDNode>(Node)->getMemoryVT(),
19385 Node->getOperand(0),
19386 Node->getOperand(1), negOp,
19387 cast<AtomicSDNode>(Node)->getMemOperand(),
19388 cast<AtomicSDNode>(Node)->getOrdering(),
19389 cast<AtomicSDNode>(Node)->getSynchScope());
19392 static SDValue LowerATOMIC_STORE(SDValue Op, SelectionDAG &DAG) {
19393 SDNode *Node = Op.getNode();
19395 EVT VT = cast<AtomicSDNode>(Node)->getMemoryVT();
19397 // Convert seq_cst store -> xchg
19398 // Convert wide store -> swap (-> cmpxchg8b/cmpxchg16b)
19399 // FIXME: On 32-bit, store -> fist or movq would be more efficient
19400 // (The only way to get a 16-byte store is cmpxchg16b)
19401 // FIXME: 16-byte ATOMIC_SWAP isn't actually hooked up at the moment.
19402 if (cast<AtomicSDNode>(Node)->getOrdering() == SequentiallyConsistent ||
19403 !DAG.getTargetLoweringInfo().isTypeLegal(VT)) {
19404 SDValue Swap = DAG.getAtomic(ISD::ATOMIC_SWAP, dl,
19405 cast<AtomicSDNode>(Node)->getMemoryVT(),
19406 Node->getOperand(0),
19407 Node->getOperand(1), Node->getOperand(2),
19408 cast<AtomicSDNode>(Node)->getMemOperand(),
19409 cast<AtomicSDNode>(Node)->getOrdering(),
19410 cast<AtomicSDNode>(Node)->getSynchScope());
19411 return Swap.getValue(1);
19413 // Other atomic stores have a simple pattern.
19417 static SDValue LowerADDC_ADDE_SUBC_SUBE(SDValue Op, SelectionDAG &DAG) {
19418 EVT VT = Op.getNode()->getSimpleValueType(0);
19420 // Let legalize expand this if it isn't a legal type yet.
19421 if (!DAG.getTargetLoweringInfo().isTypeLegal(VT))
19424 SDVTList VTs = DAG.getVTList(VT, MVT::i32);
19427 bool ExtraOp = false;
19428 switch (Op.getOpcode()) {
19429 default: llvm_unreachable("Invalid code");
19430 case ISD::ADDC: Opc = X86ISD::ADD; break;
19431 case ISD::ADDE: Opc = X86ISD::ADC; ExtraOp = true; break;
19432 case ISD::SUBC: Opc = X86ISD::SUB; break;
19433 case ISD::SUBE: Opc = X86ISD::SBB; ExtraOp = true; break;
19437 return DAG.getNode(Opc, SDLoc(Op), VTs, Op.getOperand(0),
19439 return DAG.getNode(Opc, SDLoc(Op), VTs, Op.getOperand(0),
19440 Op.getOperand(1), Op.getOperand(2));
19443 static SDValue LowerFSINCOS(SDValue Op, const X86Subtarget *Subtarget,
19444 SelectionDAG &DAG) {
19445 assert(Subtarget->isTargetDarwin() && Subtarget->is64Bit());
19447 // For MacOSX, we want to call an alternative entry point: __sincos_stret,
19448 // which returns the values as { float, float } (in XMM0) or
19449 // { double, double } (which is returned in XMM0, XMM1).
19451 SDValue Arg = Op.getOperand(0);
19452 EVT ArgVT = Arg.getValueType();
19453 Type *ArgTy = ArgVT.getTypeForEVT(*DAG.getContext());
19455 TargetLowering::ArgListTy Args;
19456 TargetLowering::ArgListEntry Entry;
19460 Entry.isSExt = false;
19461 Entry.isZExt = false;
19462 Args.push_back(Entry);
19464 bool isF64 = ArgVT == MVT::f64;
19465 // Only optimize x86_64 for now. i386 is a bit messy. For f32,
19466 // the small struct {f32, f32} is returned in (eax, edx). For f64,
19467 // the results are returned via SRet in memory.
19468 const char *LibcallName = isF64 ? "__sincos_stret" : "__sincosf_stret";
19469 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
19471 DAG.getExternalSymbol(LibcallName, TLI.getPointerTy(DAG.getDataLayout()));
19473 Type *RetTy = isF64
19474 ? (Type*)StructType::get(ArgTy, ArgTy, nullptr)
19475 : (Type*)VectorType::get(ArgTy, 4);
19477 TargetLowering::CallLoweringInfo CLI(DAG);
19478 CLI.setDebugLoc(dl).setChain(DAG.getEntryNode())
19479 .setCallee(CallingConv::C, RetTy, Callee, std::move(Args), 0);
19481 std::pair<SDValue, SDValue> CallResult = TLI.LowerCallTo(CLI);
19484 // Returned in xmm0 and xmm1.
19485 return CallResult.first;
19487 // Returned in bits 0:31 and 32:64 xmm0.
19488 SDValue SinVal = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, ArgVT,
19489 CallResult.first, DAG.getIntPtrConstant(0, dl));
19490 SDValue CosVal = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, ArgVT,
19491 CallResult.first, DAG.getIntPtrConstant(1, dl));
19492 SDVTList Tys = DAG.getVTList(ArgVT, ArgVT);
19493 return DAG.getNode(ISD::MERGE_VALUES, dl, Tys, SinVal, CosVal);
19496 static SDValue LowerMSCATTER(SDValue Op, const X86Subtarget *Subtarget,
19497 SelectionDAG &DAG) {
19498 assert(Subtarget->hasAVX512() &&
19499 "MGATHER/MSCATTER are supported on AVX-512 arch only");
19501 MaskedScatterSDNode *N = cast<MaskedScatterSDNode>(Op.getNode());
19502 EVT VT = N->getValue().getValueType();
19503 assert(VT.getScalarSizeInBits() >= 32 && "Unsupported scatter op");
19506 // X86 scatter kills mask register, so its type should be added to
19507 // the list of return values
19508 if (N->getNumValues() == 1) {
19509 SDValue Index = N->getIndex();
19510 if (!Subtarget->hasVLX() && !VT.is512BitVector() &&
19511 !Index.getValueType().is512BitVector())
19512 Index = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v8i64, Index);
19514 SDVTList VTs = DAG.getVTList(N->getMask().getValueType(), MVT::Other);
19515 SDValue Ops[] = { N->getOperand(0), N->getOperand(1), N->getOperand(2),
19516 N->getOperand(3), Index };
19518 SDValue NewScatter = DAG.getMaskedScatter(VTs, VT, dl, Ops, N->getMemOperand());
19519 DAG.ReplaceAllUsesWith(Op, SDValue(NewScatter.getNode(), 1));
19520 return SDValue(NewScatter.getNode(), 0);
19525 static SDValue LowerMGATHER(SDValue Op, const X86Subtarget *Subtarget,
19526 SelectionDAG &DAG) {
19527 assert(Subtarget->hasAVX512() &&
19528 "MGATHER/MSCATTER are supported on AVX-512 arch only");
19530 MaskedGatherSDNode *N = cast<MaskedGatherSDNode>(Op.getNode());
19531 EVT VT = Op.getValueType();
19532 assert(VT.getScalarSizeInBits() >= 32 && "Unsupported gather op");
19535 SDValue Index = N->getIndex();
19536 if (!Subtarget->hasVLX() && !VT.is512BitVector() &&
19537 !Index.getValueType().is512BitVector()) {
19538 Index = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v8i64, Index);
19539 SDValue Ops[] = { N->getOperand(0), N->getOperand(1), N->getOperand(2),
19540 N->getOperand(3), Index };
19541 DAG.UpdateNodeOperands(N, Ops);
19546 SDValue X86TargetLowering::LowerGC_TRANSITION_START(SDValue Op,
19547 SelectionDAG &DAG) const {
19548 // TODO: Eventually, the lowering of these nodes should be informed by or
19549 // deferred to the GC strategy for the function in which they appear. For
19550 // now, however, they must be lowered to something. Since they are logically
19551 // no-ops in the case of a null GC strategy (or a GC strategy which does not
19552 // require special handling for these nodes), lower them as literal NOOPs for
19554 SmallVector<SDValue, 2> Ops;
19556 Ops.push_back(Op.getOperand(0));
19557 if (Op->getGluedNode())
19558 Ops.push_back(Op->getOperand(Op->getNumOperands() - 1));
19561 SDVTList VTs = DAG.getVTList(MVT::Other, MVT::Glue);
19562 SDValue NOOP(DAG.getMachineNode(X86::NOOP, SDLoc(Op), VTs, Ops), 0);
19567 SDValue X86TargetLowering::LowerGC_TRANSITION_END(SDValue Op,
19568 SelectionDAG &DAG) const {
19569 // TODO: Eventually, the lowering of these nodes should be informed by or
19570 // deferred to the GC strategy for the function in which they appear. For
19571 // now, however, they must be lowered to something. Since they are logically
19572 // no-ops in the case of a null GC strategy (or a GC strategy which does not
19573 // require special handling for these nodes), lower them as literal NOOPs for
19575 SmallVector<SDValue, 2> Ops;
19577 Ops.push_back(Op.getOperand(0));
19578 if (Op->getGluedNode())
19579 Ops.push_back(Op->getOperand(Op->getNumOperands() - 1));
19582 SDVTList VTs = DAG.getVTList(MVT::Other, MVT::Glue);
19583 SDValue NOOP(DAG.getMachineNode(X86::NOOP, SDLoc(Op), VTs, Ops), 0);
19588 /// LowerOperation - Provide custom lowering hooks for some operations.
19590 SDValue X86TargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
19591 switch (Op.getOpcode()) {
19592 default: llvm_unreachable("Should not custom lower this!");
19593 case ISD::ATOMIC_FENCE: return LowerATOMIC_FENCE(Op, Subtarget, DAG);
19594 case ISD::ATOMIC_CMP_SWAP_WITH_SUCCESS:
19595 return LowerCMP_SWAP(Op, Subtarget, DAG);
19596 case ISD::CTPOP: return LowerCTPOP(Op, Subtarget, DAG);
19597 case ISD::ATOMIC_LOAD_SUB: return LowerLOAD_SUB(Op,DAG);
19598 case ISD::ATOMIC_STORE: return LowerATOMIC_STORE(Op,DAG);
19599 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG);
19600 case ISD::CONCAT_VECTORS: return LowerCONCAT_VECTORS(Op, Subtarget, DAG);
19601 case ISD::VECTOR_SHUFFLE: return lowerVectorShuffle(Op, Subtarget, DAG);
19602 case ISD::VSELECT: return LowerVSELECT(Op, DAG);
19603 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
19604 case ISD::INSERT_VECTOR_ELT: return LowerINSERT_VECTOR_ELT(Op, DAG);
19605 case ISD::EXTRACT_SUBVECTOR: return LowerEXTRACT_SUBVECTOR(Op,Subtarget,DAG);
19606 case ISD::INSERT_SUBVECTOR: return LowerINSERT_SUBVECTOR(Op, Subtarget,DAG);
19607 case ISD::SCALAR_TO_VECTOR: return LowerSCALAR_TO_VECTOR(Op, DAG);
19608 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
19609 case ISD::GlobalAddress: return LowerGlobalAddress(Op, DAG);
19610 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
19611 case ISD::ExternalSymbol: return LowerExternalSymbol(Op, DAG);
19612 case ISD::BlockAddress: return LowerBlockAddress(Op, DAG);
19613 case ISD::SHL_PARTS:
19614 case ISD::SRA_PARTS:
19615 case ISD::SRL_PARTS: return LowerShiftParts(Op, DAG);
19616 case ISD::SINT_TO_FP: return LowerSINT_TO_FP(Op, DAG);
19617 case ISD::UINT_TO_FP: return LowerUINT_TO_FP(Op, DAG);
19618 case ISD::TRUNCATE: return LowerTRUNCATE(Op, DAG);
19619 case ISD::ZERO_EXTEND: return LowerZERO_EXTEND(Op, Subtarget, DAG);
19620 case ISD::SIGN_EXTEND: return LowerSIGN_EXTEND(Op, Subtarget, DAG);
19621 case ISD::ANY_EXTEND: return LowerANY_EXTEND(Op, Subtarget, DAG);
19622 case ISD::SIGN_EXTEND_VECTOR_INREG:
19623 return LowerSIGN_EXTEND_VECTOR_INREG(Op, Subtarget, DAG);
19624 case ISD::FP_TO_SINT: return LowerFP_TO_SINT(Op, DAG);
19625 case ISD::FP_TO_UINT: return LowerFP_TO_UINT(Op, DAG);
19626 case ISD::FP_EXTEND: return LowerFP_EXTEND(Op, DAG);
19627 case ISD::LOAD: return LowerExtendedLoad(Op, Subtarget, DAG);
19629 case ISD::FNEG: return LowerFABSorFNEG(Op, DAG);
19630 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
19631 case ISD::FGETSIGN: return LowerFGETSIGN(Op, DAG);
19632 case ISD::SETCC: return LowerSETCC(Op, DAG);
19633 case ISD::SELECT: return LowerSELECT(Op, DAG);
19634 case ISD::BRCOND: return LowerBRCOND(Op, DAG);
19635 case ISD::JumpTable: return LowerJumpTable(Op, DAG);
19636 case ISD::VASTART: return LowerVASTART(Op, DAG);
19637 case ISD::VAARG: return LowerVAARG(Op, DAG);
19638 case ISD::VACOPY: return LowerVACOPY(Op, Subtarget, DAG);
19639 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, Subtarget, DAG);
19640 case ISD::INTRINSIC_VOID:
19641 case ISD::INTRINSIC_W_CHAIN: return LowerINTRINSIC_W_CHAIN(Op, Subtarget, DAG);
19642 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
19643 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
19644 case ISD::FRAME_TO_ARGS_OFFSET:
19645 return LowerFRAME_TO_ARGS_OFFSET(Op, DAG);
19646 case ISD::DYNAMIC_STACKALLOC: return LowerDYNAMIC_STACKALLOC(Op, DAG);
19647 case ISD::EH_RETURN: return LowerEH_RETURN(Op, DAG);
19648 case ISD::EH_SJLJ_SETJMP: return lowerEH_SJLJ_SETJMP(Op, DAG);
19649 case ISD::EH_SJLJ_LONGJMP: return lowerEH_SJLJ_LONGJMP(Op, DAG);
19650 case ISD::INIT_TRAMPOLINE: return LowerINIT_TRAMPOLINE(Op, DAG);
19651 case ISD::ADJUST_TRAMPOLINE: return LowerADJUST_TRAMPOLINE(Op, DAG);
19652 case ISD::FLT_ROUNDS_: return LowerFLT_ROUNDS_(Op, DAG);
19653 case ISD::CTLZ: return LowerCTLZ(Op, Subtarget, DAG);
19654 case ISD::CTLZ_ZERO_UNDEF: return LowerCTLZ_ZERO_UNDEF(Op, Subtarget, DAG);
19656 case ISD::CTTZ_ZERO_UNDEF: return LowerCTTZ(Op, DAG);
19657 case ISD::MUL: return LowerMUL(Op, Subtarget, DAG);
19658 case ISD::UMUL_LOHI:
19659 case ISD::SMUL_LOHI: return LowerMUL_LOHI(Op, Subtarget, DAG);
19660 case ISD::ROTL: return LowerRotate(Op, Subtarget, DAG);
19663 case ISD::SHL: return LowerShift(Op, Subtarget, DAG);
19669 case ISD::UMULO: return LowerXALUO(Op, DAG);
19670 case ISD::READCYCLECOUNTER: return LowerREADCYCLECOUNTER(Op, Subtarget,DAG);
19671 case ISD::BITCAST: return LowerBITCAST(Op, Subtarget, DAG);
19675 case ISD::SUBE: return LowerADDC_ADDE_SUBC_SUBE(Op, DAG);
19676 case ISD::ADD: return LowerADD(Op, DAG);
19677 case ISD::SUB: return LowerSUB(Op, DAG);
19681 case ISD::UMIN: return LowerMINMAX(Op, DAG);
19682 case ISD::FSINCOS: return LowerFSINCOS(Op, Subtarget, DAG);
19683 case ISD::MGATHER: return LowerMGATHER(Op, Subtarget, DAG);
19684 case ISD::MSCATTER: return LowerMSCATTER(Op, Subtarget, DAG);
19685 case ISD::GC_TRANSITION_START:
19686 return LowerGC_TRANSITION_START(Op, DAG);
19687 case ISD::GC_TRANSITION_END: return LowerGC_TRANSITION_END(Op, DAG);
19691 /// ReplaceNodeResults - Replace a node with an illegal result type
19692 /// with a new node built out of custom code.
19693 void X86TargetLowering::ReplaceNodeResults(SDNode *N,
19694 SmallVectorImpl<SDValue>&Results,
19695 SelectionDAG &DAG) const {
19697 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
19698 switch (N->getOpcode()) {
19700 llvm_unreachable("Do not know how to custom type legalize this operation!");
19701 // We might have generated v2f32 FMIN/FMAX operations. Widen them to v4f32.
19702 case X86ISD::FMINC:
19704 case X86ISD::FMAXC:
19705 case X86ISD::FMAX: {
19706 EVT VT = N->getValueType(0);
19707 if (VT != MVT::v2f32)
19708 llvm_unreachable("Unexpected type (!= v2f32) on FMIN/FMAX.");
19709 SDValue UNDEF = DAG.getUNDEF(VT);
19710 SDValue LHS = DAG.getNode(ISD::CONCAT_VECTORS, dl, MVT::v4f32,
19711 N->getOperand(0), UNDEF);
19712 SDValue RHS = DAG.getNode(ISD::CONCAT_VECTORS, dl, MVT::v4f32,
19713 N->getOperand(1), UNDEF);
19714 Results.push_back(DAG.getNode(N->getOpcode(), dl, MVT::v4f32, LHS, RHS));
19717 case ISD::SIGN_EXTEND_INREG:
19722 // We don't want to expand or promote these.
19729 case ISD::UDIVREM: {
19730 SDValue V = LowerWin64_i128OP(SDValue(N,0), DAG);
19731 Results.push_back(V);
19734 case ISD::FP_TO_SINT:
19735 case ISD::FP_TO_UINT: {
19736 bool IsSigned = N->getOpcode() == ISD::FP_TO_SINT;
19738 std::pair<SDValue,SDValue> Vals =
19739 FP_TO_INTHelper(SDValue(N, 0), DAG, IsSigned, /*IsReplace=*/ true);
19740 SDValue FIST = Vals.first, StackSlot = Vals.second;
19741 if (FIST.getNode()) {
19742 EVT VT = N->getValueType(0);
19743 // Return a load from the stack slot.
19744 if (StackSlot.getNode())
19745 Results.push_back(DAG.getLoad(VT, dl, FIST, StackSlot,
19746 MachinePointerInfo(),
19747 false, false, false, 0));
19749 Results.push_back(FIST);
19753 case ISD::UINT_TO_FP: {
19754 assert(Subtarget->hasSSE2() && "Requires at least SSE2!");
19755 if (N->getOperand(0).getValueType() != MVT::v2i32 ||
19756 N->getValueType(0) != MVT::v2f32)
19758 SDValue ZExtIn = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v2i64,
19760 SDValue Bias = DAG.getConstantFP(BitsToDouble(0x4330000000000000ULL), dl,
19762 SDValue VBias = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v2f64, Bias, Bias);
19763 SDValue Or = DAG.getNode(ISD::OR, dl, MVT::v2i64, ZExtIn,
19764 DAG.getBitcast(MVT::v2i64, VBias));
19765 Or = DAG.getBitcast(MVT::v2f64, Or);
19766 // TODO: Are there any fast-math-flags to propagate here?
19767 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::v2f64, Or, VBias);
19768 Results.push_back(DAG.getNode(X86ISD::VFPROUND, dl, MVT::v4f32, Sub));
19771 case ISD::FP_ROUND: {
19772 if (!TLI.isTypeLegal(N->getOperand(0).getValueType()))
19774 SDValue V = DAG.getNode(X86ISD::VFPROUND, dl, MVT::v4f32, N->getOperand(0));
19775 Results.push_back(V);
19778 case ISD::FP_EXTEND: {
19779 // Right now, only MVT::v2f32 has OperationAction for FP_EXTEND.
19780 // No other ValueType for FP_EXTEND should reach this point.
19781 assert(N->getValueType(0) == MVT::v2f32 &&
19782 "Do not know how to legalize this Node");
19785 case ISD::INTRINSIC_W_CHAIN: {
19786 unsigned IntNo = cast<ConstantSDNode>(N->getOperand(1))->getZExtValue();
19788 default : llvm_unreachable("Do not know how to custom type "
19789 "legalize this intrinsic operation!");
19790 case Intrinsic::x86_rdtsc:
19791 return getReadTimeStampCounter(N, dl, X86ISD::RDTSC_DAG, DAG, Subtarget,
19793 case Intrinsic::x86_rdtscp:
19794 return getReadTimeStampCounter(N, dl, X86ISD::RDTSCP_DAG, DAG, Subtarget,
19796 case Intrinsic::x86_rdpmc:
19797 return getReadPerformanceCounter(N, dl, DAG, Subtarget, Results);
19800 case ISD::READCYCLECOUNTER: {
19801 return getReadTimeStampCounter(N, dl, X86ISD::RDTSC_DAG, DAG, Subtarget,
19804 case ISD::ATOMIC_CMP_SWAP_WITH_SUCCESS: {
19805 EVT T = N->getValueType(0);
19806 assert((T == MVT::i64 || T == MVT::i128) && "can only expand cmpxchg pair");
19807 bool Regs64bit = T == MVT::i128;
19808 EVT HalfT = Regs64bit ? MVT::i64 : MVT::i32;
19809 SDValue cpInL, cpInH;
19810 cpInL = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, HalfT, N->getOperand(2),
19811 DAG.getConstant(0, dl, HalfT));
19812 cpInH = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, HalfT, N->getOperand(2),
19813 DAG.getConstant(1, dl, HalfT));
19814 cpInL = DAG.getCopyToReg(N->getOperand(0), dl,
19815 Regs64bit ? X86::RAX : X86::EAX,
19817 cpInH = DAG.getCopyToReg(cpInL.getValue(0), dl,
19818 Regs64bit ? X86::RDX : X86::EDX,
19819 cpInH, cpInL.getValue(1));
19820 SDValue swapInL, swapInH;
19821 swapInL = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, HalfT, N->getOperand(3),
19822 DAG.getConstant(0, dl, HalfT));
19823 swapInH = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, HalfT, N->getOperand(3),
19824 DAG.getConstant(1, dl, HalfT));
19825 swapInL = DAG.getCopyToReg(cpInH.getValue(0), dl,
19826 Regs64bit ? X86::RBX : X86::EBX,
19827 swapInL, cpInH.getValue(1));
19828 swapInH = DAG.getCopyToReg(swapInL.getValue(0), dl,
19829 Regs64bit ? X86::RCX : X86::ECX,
19830 swapInH, swapInL.getValue(1));
19831 SDValue Ops[] = { swapInH.getValue(0),
19833 swapInH.getValue(1) };
19834 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Glue);
19835 MachineMemOperand *MMO = cast<AtomicSDNode>(N)->getMemOperand();
19836 unsigned Opcode = Regs64bit ? X86ISD::LCMPXCHG16_DAG :
19837 X86ISD::LCMPXCHG8_DAG;
19838 SDValue Result = DAG.getMemIntrinsicNode(Opcode, dl, Tys, Ops, T, MMO);
19839 SDValue cpOutL = DAG.getCopyFromReg(Result.getValue(0), dl,
19840 Regs64bit ? X86::RAX : X86::EAX,
19841 HalfT, Result.getValue(1));
19842 SDValue cpOutH = DAG.getCopyFromReg(cpOutL.getValue(1), dl,
19843 Regs64bit ? X86::RDX : X86::EDX,
19844 HalfT, cpOutL.getValue(2));
19845 SDValue OpsF[] = { cpOutL.getValue(0), cpOutH.getValue(0)};
19847 SDValue EFLAGS = DAG.getCopyFromReg(cpOutH.getValue(1), dl, X86::EFLAGS,
19848 MVT::i32, cpOutH.getValue(2));
19850 DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
19851 DAG.getConstant(X86::COND_E, dl, MVT::i8), EFLAGS);
19852 Success = DAG.getZExtOrTrunc(Success, dl, N->getValueType(1));
19854 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, T, OpsF));
19855 Results.push_back(Success);
19856 Results.push_back(EFLAGS.getValue(1));
19859 case ISD::ATOMIC_SWAP:
19860 case ISD::ATOMIC_LOAD_ADD:
19861 case ISD::ATOMIC_LOAD_SUB:
19862 case ISD::ATOMIC_LOAD_AND:
19863 case ISD::ATOMIC_LOAD_OR:
19864 case ISD::ATOMIC_LOAD_XOR:
19865 case ISD::ATOMIC_LOAD_NAND:
19866 case ISD::ATOMIC_LOAD_MIN:
19867 case ISD::ATOMIC_LOAD_MAX:
19868 case ISD::ATOMIC_LOAD_UMIN:
19869 case ISD::ATOMIC_LOAD_UMAX:
19870 case ISD::ATOMIC_LOAD: {
19871 // Delegate to generic TypeLegalization. Situations we can really handle
19872 // should have already been dealt with by AtomicExpandPass.cpp.
19875 case ISD::BITCAST: {
19876 assert(Subtarget->hasSSE2() && "Requires at least SSE2!");
19877 EVT DstVT = N->getValueType(0);
19878 EVT SrcVT = N->getOperand(0)->getValueType(0);
19880 if (SrcVT != MVT::f64 ||
19881 (DstVT != MVT::v2i32 && DstVT != MVT::v4i16 && DstVT != MVT::v8i8))
19884 unsigned NumElts = DstVT.getVectorNumElements();
19885 EVT SVT = DstVT.getVectorElementType();
19886 EVT WiderVT = EVT::getVectorVT(*DAG.getContext(), SVT, NumElts * 2);
19887 SDValue Expanded = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
19888 MVT::v2f64, N->getOperand(0));
19889 SDValue ToVecInt = DAG.getBitcast(WiderVT, Expanded);
19891 if (ExperimentalVectorWideningLegalization) {
19892 // If we are legalizing vectors by widening, we already have the desired
19893 // legal vector type, just return it.
19894 Results.push_back(ToVecInt);
19898 SmallVector<SDValue, 8> Elts;
19899 for (unsigned i = 0, e = NumElts; i != e; ++i)
19900 Elts.push_back(DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, SVT,
19901 ToVecInt, DAG.getIntPtrConstant(i, dl)));
19903 Results.push_back(DAG.getNode(ISD::BUILD_VECTOR, dl, DstVT, Elts));
19908 const char *X86TargetLowering::getTargetNodeName(unsigned Opcode) const {
19909 switch ((X86ISD::NodeType)Opcode) {
19910 case X86ISD::FIRST_NUMBER: break;
19911 case X86ISD::BSF: return "X86ISD::BSF";
19912 case X86ISD::BSR: return "X86ISD::BSR";
19913 case X86ISD::SHLD: return "X86ISD::SHLD";
19914 case X86ISD::SHRD: return "X86ISD::SHRD";
19915 case X86ISD::FAND: return "X86ISD::FAND";
19916 case X86ISD::FANDN: return "X86ISD::FANDN";
19917 case X86ISD::FOR: return "X86ISD::FOR";
19918 case X86ISD::FXOR: return "X86ISD::FXOR";
19919 case X86ISD::FILD: return "X86ISD::FILD";
19920 case X86ISD::FILD_FLAG: return "X86ISD::FILD_FLAG";
19921 case X86ISD::FP_TO_INT16_IN_MEM: return "X86ISD::FP_TO_INT16_IN_MEM";
19922 case X86ISD::FP_TO_INT32_IN_MEM: return "X86ISD::FP_TO_INT32_IN_MEM";
19923 case X86ISD::FP_TO_INT64_IN_MEM: return "X86ISD::FP_TO_INT64_IN_MEM";
19924 case X86ISD::FLD: return "X86ISD::FLD";
19925 case X86ISD::FST: return "X86ISD::FST";
19926 case X86ISD::CALL: return "X86ISD::CALL";
19927 case X86ISD::RDTSC_DAG: return "X86ISD::RDTSC_DAG";
19928 case X86ISD::RDTSCP_DAG: return "X86ISD::RDTSCP_DAG";
19929 case X86ISD::RDPMC_DAG: return "X86ISD::RDPMC_DAG";
19930 case X86ISD::BT: return "X86ISD::BT";
19931 case X86ISD::CMP: return "X86ISD::CMP";
19932 case X86ISD::COMI: return "X86ISD::COMI";
19933 case X86ISD::UCOMI: return "X86ISD::UCOMI";
19934 case X86ISD::CMPM: return "X86ISD::CMPM";
19935 case X86ISD::CMPMU: return "X86ISD::CMPMU";
19936 case X86ISD::CMPM_RND: return "X86ISD::CMPM_RND";
19937 case X86ISD::SETCC: return "X86ISD::SETCC";
19938 case X86ISD::SETCC_CARRY: return "X86ISD::SETCC_CARRY";
19939 case X86ISD::FSETCC: return "X86ISD::FSETCC";
19940 case X86ISD::FGETSIGNx86: return "X86ISD::FGETSIGNx86";
19941 case X86ISD::CMOV: return "X86ISD::CMOV";
19942 case X86ISD::BRCOND: return "X86ISD::BRCOND";
19943 case X86ISD::RET_FLAG: return "X86ISD::RET_FLAG";
19944 case X86ISD::REP_STOS: return "X86ISD::REP_STOS";
19945 case X86ISD::REP_MOVS: return "X86ISD::REP_MOVS";
19946 case X86ISD::GlobalBaseReg: return "X86ISD::GlobalBaseReg";
19947 case X86ISD::Wrapper: return "X86ISD::Wrapper";
19948 case X86ISD::WrapperRIP: return "X86ISD::WrapperRIP";
19949 case X86ISD::MOVDQ2Q: return "X86ISD::MOVDQ2Q";
19950 case X86ISD::MMX_MOVD2W: return "X86ISD::MMX_MOVD2W";
19951 case X86ISD::MMX_MOVW2D: return "X86ISD::MMX_MOVW2D";
19952 case X86ISD::PEXTRB: return "X86ISD::PEXTRB";
19953 case X86ISD::PEXTRW: return "X86ISD::PEXTRW";
19954 case X86ISD::INSERTPS: return "X86ISD::INSERTPS";
19955 case X86ISD::PINSRB: return "X86ISD::PINSRB";
19956 case X86ISD::PINSRW: return "X86ISD::PINSRW";
19957 case X86ISD::MMX_PINSRW: return "X86ISD::MMX_PINSRW";
19958 case X86ISD::PSHUFB: return "X86ISD::PSHUFB";
19959 case X86ISD::ANDNP: return "X86ISD::ANDNP";
19960 case X86ISD::PSIGN: return "X86ISD::PSIGN";
19961 case X86ISD::BLENDI: return "X86ISD::BLENDI";
19962 case X86ISD::SHRUNKBLEND: return "X86ISD::SHRUNKBLEND";
19963 case X86ISD::ADDUS: return "X86ISD::ADDUS";
19964 case X86ISD::SUBUS: return "X86ISD::SUBUS";
19965 case X86ISD::HADD: return "X86ISD::HADD";
19966 case X86ISD::HSUB: return "X86ISD::HSUB";
19967 case X86ISD::FHADD: return "X86ISD::FHADD";
19968 case X86ISD::FHSUB: return "X86ISD::FHSUB";
19969 case X86ISD::ABS: return "X86ISD::ABS";
19970 case X86ISD::CONFLICT: return "X86ISD::CONFLICT";
19971 case X86ISD::FMAX: return "X86ISD::FMAX";
19972 case X86ISD::FMAX_RND: return "X86ISD::FMAX_RND";
19973 case X86ISD::FMIN: return "X86ISD::FMIN";
19974 case X86ISD::FMIN_RND: return "X86ISD::FMIN_RND";
19975 case X86ISD::FMAXC: return "X86ISD::FMAXC";
19976 case X86ISD::FMINC: return "X86ISD::FMINC";
19977 case X86ISD::FRSQRT: return "X86ISD::FRSQRT";
19978 case X86ISD::FRCP: return "X86ISD::FRCP";
19979 case X86ISD::EXTRQI: return "X86ISD::EXTRQI";
19980 case X86ISD::INSERTQI: return "X86ISD::INSERTQI";
19981 case X86ISD::TLSADDR: return "X86ISD::TLSADDR";
19982 case X86ISD::TLSBASEADDR: return "X86ISD::TLSBASEADDR";
19983 case X86ISD::TLSCALL: return "X86ISD::TLSCALL";
19984 case X86ISD::EH_SJLJ_SETJMP: return "X86ISD::EH_SJLJ_SETJMP";
19985 case X86ISD::EH_SJLJ_LONGJMP: return "X86ISD::EH_SJLJ_LONGJMP";
19986 case X86ISD::EH_RETURN: return "X86ISD::EH_RETURN";
19987 case X86ISD::TC_RETURN: return "X86ISD::TC_RETURN";
19988 case X86ISD::FNSTCW16m: return "X86ISD::FNSTCW16m";
19989 case X86ISD::FNSTSW16r: return "X86ISD::FNSTSW16r";
19990 case X86ISD::LCMPXCHG_DAG: return "X86ISD::LCMPXCHG_DAG";
19991 case X86ISD::LCMPXCHG8_DAG: return "X86ISD::LCMPXCHG8_DAG";
19992 case X86ISD::LCMPXCHG16_DAG: return "X86ISD::LCMPXCHG16_DAG";
19993 case X86ISD::VZEXT_MOVL: return "X86ISD::VZEXT_MOVL";
19994 case X86ISD::VZEXT_LOAD: return "X86ISD::VZEXT_LOAD";
19995 case X86ISD::VZEXT: return "X86ISD::VZEXT";
19996 case X86ISD::VSEXT: return "X86ISD::VSEXT";
19997 case X86ISD::VTRUNC: return "X86ISD::VTRUNC";
19998 case X86ISD::VTRUNCS: return "X86ISD::VTRUNCS";
19999 case X86ISD::VTRUNCUS: return "X86ISD::VTRUNCUS";
20000 case X86ISD::VINSERT: return "X86ISD::VINSERT";
20001 case X86ISD::VFPEXT: return "X86ISD::VFPEXT";
20002 case X86ISD::VFPROUND: return "X86ISD::VFPROUND";
20003 case X86ISD::CVTDQ2PD: return "X86ISD::CVTDQ2PD";
20004 case X86ISD::CVTUDQ2PD: return "X86ISD::CVTUDQ2PD";
20005 case X86ISD::VSHLDQ: return "X86ISD::VSHLDQ";
20006 case X86ISD::VSRLDQ: return "X86ISD::VSRLDQ";
20007 case X86ISD::VSHL: return "X86ISD::VSHL";
20008 case X86ISD::VSRL: return "X86ISD::VSRL";
20009 case X86ISD::VSRA: return "X86ISD::VSRA";
20010 case X86ISD::VSHLI: return "X86ISD::VSHLI";
20011 case X86ISD::VSRLI: return "X86ISD::VSRLI";
20012 case X86ISD::VSRAI: return "X86ISD::VSRAI";
20013 case X86ISD::CMPP: return "X86ISD::CMPP";
20014 case X86ISD::PCMPEQ: return "X86ISD::PCMPEQ";
20015 case X86ISD::PCMPGT: return "X86ISD::PCMPGT";
20016 case X86ISD::PCMPEQM: return "X86ISD::PCMPEQM";
20017 case X86ISD::PCMPGTM: return "X86ISD::PCMPGTM";
20018 case X86ISD::ADD: return "X86ISD::ADD";
20019 case X86ISD::SUB: return "X86ISD::SUB";
20020 case X86ISD::ADC: return "X86ISD::ADC";
20021 case X86ISD::SBB: return "X86ISD::SBB";
20022 case X86ISD::SMUL: return "X86ISD::SMUL";
20023 case X86ISD::UMUL: return "X86ISD::UMUL";
20024 case X86ISD::SMUL8: return "X86ISD::SMUL8";
20025 case X86ISD::UMUL8: return "X86ISD::UMUL8";
20026 case X86ISD::SDIVREM8_SEXT_HREG: return "X86ISD::SDIVREM8_SEXT_HREG";
20027 case X86ISD::UDIVREM8_ZEXT_HREG: return "X86ISD::UDIVREM8_ZEXT_HREG";
20028 case X86ISD::INC: return "X86ISD::INC";
20029 case X86ISD::DEC: return "X86ISD::DEC";
20030 case X86ISD::OR: return "X86ISD::OR";
20031 case X86ISD::XOR: return "X86ISD::XOR";
20032 case X86ISD::AND: return "X86ISD::AND";
20033 case X86ISD::BEXTR: return "X86ISD::BEXTR";
20034 case X86ISD::MUL_IMM: return "X86ISD::MUL_IMM";
20035 case X86ISD::PTEST: return "X86ISD::PTEST";
20036 case X86ISD::TESTP: return "X86ISD::TESTP";
20037 case X86ISD::TESTM: return "X86ISD::TESTM";
20038 case X86ISD::TESTNM: return "X86ISD::TESTNM";
20039 case X86ISD::KORTEST: return "X86ISD::KORTEST";
20040 case X86ISD::KTEST: return "X86ISD::KTEST";
20041 case X86ISD::PACKSS: return "X86ISD::PACKSS";
20042 case X86ISD::PACKUS: return "X86ISD::PACKUS";
20043 case X86ISD::PALIGNR: return "X86ISD::PALIGNR";
20044 case X86ISD::VALIGN: return "X86ISD::VALIGN";
20045 case X86ISD::PSHUFD: return "X86ISD::PSHUFD";
20046 case X86ISD::PSHUFHW: return "X86ISD::PSHUFHW";
20047 case X86ISD::PSHUFLW: return "X86ISD::PSHUFLW";
20048 case X86ISD::SHUFP: return "X86ISD::SHUFP";
20049 case X86ISD::SHUF128: return "X86ISD::SHUF128";
20050 case X86ISD::MOVLHPS: return "X86ISD::MOVLHPS";
20051 case X86ISD::MOVLHPD: return "X86ISD::MOVLHPD";
20052 case X86ISD::MOVHLPS: return "X86ISD::MOVHLPS";
20053 case X86ISD::MOVLPS: return "X86ISD::MOVLPS";
20054 case X86ISD::MOVLPD: return "X86ISD::MOVLPD";
20055 case X86ISD::MOVDDUP: return "X86ISD::MOVDDUP";
20056 case X86ISD::MOVSHDUP: return "X86ISD::MOVSHDUP";
20057 case X86ISD::MOVSLDUP: return "X86ISD::MOVSLDUP";
20058 case X86ISD::MOVSD: return "X86ISD::MOVSD";
20059 case X86ISD::MOVSS: return "X86ISD::MOVSS";
20060 case X86ISD::UNPCKL: return "X86ISD::UNPCKL";
20061 case X86ISD::UNPCKH: return "X86ISD::UNPCKH";
20062 case X86ISD::VBROADCAST: return "X86ISD::VBROADCAST";
20063 case X86ISD::SUBV_BROADCAST: return "X86ISD::SUBV_BROADCAST";
20064 case X86ISD::VEXTRACT: return "X86ISD::VEXTRACT";
20065 case X86ISD::VPERMILPV: return "X86ISD::VPERMILPV";
20066 case X86ISD::VPERMILPI: return "X86ISD::VPERMILPI";
20067 case X86ISD::VPERM2X128: return "X86ISD::VPERM2X128";
20068 case X86ISD::VPERMV: return "X86ISD::VPERMV";
20069 case X86ISD::VPERMV3: return "X86ISD::VPERMV3";
20070 case X86ISD::VPERMIV3: return "X86ISD::VPERMIV3";
20071 case X86ISD::VPERMI: return "X86ISD::VPERMI";
20072 case X86ISD::VPTERNLOG: return "X86ISD::VPTERNLOG";
20073 case X86ISD::VFIXUPIMM: return "X86ISD::VFIXUPIMM";
20074 case X86ISD::VRANGE: return "X86ISD::VRANGE";
20075 case X86ISD::PMULUDQ: return "X86ISD::PMULUDQ";
20076 case X86ISD::PMULDQ: return "X86ISD::PMULDQ";
20077 case X86ISD::PSADBW: return "X86ISD::PSADBW";
20078 case X86ISD::DBPSADBW: return "X86ISD::DBPSADBW";
20079 case X86ISD::VASTART_SAVE_XMM_REGS: return "X86ISD::VASTART_SAVE_XMM_REGS";
20080 case X86ISD::VAARG_64: return "X86ISD::VAARG_64";
20081 case X86ISD::WIN_ALLOCA: return "X86ISD::WIN_ALLOCA";
20082 case X86ISD::MEMBARRIER: return "X86ISD::MEMBARRIER";
20083 case X86ISD::MFENCE: return "X86ISD::MFENCE";
20084 case X86ISD::SFENCE: return "X86ISD::SFENCE";
20085 case X86ISD::LFENCE: return "X86ISD::LFENCE";
20086 case X86ISD::SEG_ALLOCA: return "X86ISD::SEG_ALLOCA";
20087 case X86ISD::SAHF: return "X86ISD::SAHF";
20088 case X86ISD::RDRAND: return "X86ISD::RDRAND";
20089 case X86ISD::RDSEED: return "X86ISD::RDSEED";
20090 case X86ISD::VPMADDUBSW: return "X86ISD::VPMADDUBSW";
20091 case X86ISD::VPMADDWD: return "X86ISD::VPMADDWD";
20092 case X86ISD::VPROT: return "X86ISD::VPROT";
20093 case X86ISD::VPROTI: return "X86ISD::VPROTI";
20094 case X86ISD::VPSHA: return "X86ISD::VPSHA";
20095 case X86ISD::VPSHL: return "X86ISD::VPSHL";
20096 case X86ISD::VPCOM: return "X86ISD::VPCOM";
20097 case X86ISD::VPCOMU: return "X86ISD::VPCOMU";
20098 case X86ISD::FMADD: return "X86ISD::FMADD";
20099 case X86ISD::FMSUB: return "X86ISD::FMSUB";
20100 case X86ISD::FNMADD: return "X86ISD::FNMADD";
20101 case X86ISD::FNMSUB: return "X86ISD::FNMSUB";
20102 case X86ISD::FMADDSUB: return "X86ISD::FMADDSUB";
20103 case X86ISD::FMSUBADD: return "X86ISD::FMSUBADD";
20104 case X86ISD::FMADD_RND: return "X86ISD::FMADD_RND";
20105 case X86ISD::FNMADD_RND: return "X86ISD::FNMADD_RND";
20106 case X86ISD::FMSUB_RND: return "X86ISD::FMSUB_RND";
20107 case X86ISD::FNMSUB_RND: return "X86ISD::FNMSUB_RND";
20108 case X86ISD::FMADDSUB_RND: return "X86ISD::FMADDSUB_RND";
20109 case X86ISD::FMSUBADD_RND: return "X86ISD::FMSUBADD_RND";
20110 case X86ISD::VRNDSCALE: return "X86ISD::VRNDSCALE";
20111 case X86ISD::VREDUCE: return "X86ISD::VREDUCE";
20112 case X86ISD::VGETMANT: return "X86ISD::VGETMANT";
20113 case X86ISD::PCMPESTRI: return "X86ISD::PCMPESTRI";
20114 case X86ISD::PCMPISTRI: return "X86ISD::PCMPISTRI";
20115 case X86ISD::XTEST: return "X86ISD::XTEST";
20116 case X86ISD::COMPRESS: return "X86ISD::COMPRESS";
20117 case X86ISD::EXPAND: return "X86ISD::EXPAND";
20118 case X86ISD::SELECT: return "X86ISD::SELECT";
20119 case X86ISD::ADDSUB: return "X86ISD::ADDSUB";
20120 case X86ISD::RCP28: return "X86ISD::RCP28";
20121 case X86ISD::EXP2: return "X86ISD::EXP2";
20122 case X86ISD::RSQRT28: return "X86ISD::RSQRT28";
20123 case X86ISD::FADD_RND: return "X86ISD::FADD_RND";
20124 case X86ISD::FSUB_RND: return "X86ISD::FSUB_RND";
20125 case X86ISD::FMUL_RND: return "X86ISD::FMUL_RND";
20126 case X86ISD::FDIV_RND: return "X86ISD::FDIV_RND";
20127 case X86ISD::FSQRT_RND: return "X86ISD::FSQRT_RND";
20128 case X86ISD::FGETEXP_RND: return "X86ISD::FGETEXP_RND";
20129 case X86ISD::SCALEF: return "X86ISD::SCALEF";
20130 case X86ISD::ADDS: return "X86ISD::ADDS";
20131 case X86ISD::SUBS: return "X86ISD::SUBS";
20132 case X86ISD::AVG: return "X86ISD::AVG";
20133 case X86ISD::MULHRS: return "X86ISD::MULHRS";
20134 case X86ISD::SINT_TO_FP_RND: return "X86ISD::SINT_TO_FP_RND";
20135 case X86ISD::UINT_TO_FP_RND: return "X86ISD::UINT_TO_FP_RND";
20136 case X86ISD::FP_TO_SINT_RND: return "X86ISD::FP_TO_SINT_RND";
20137 case X86ISD::FP_TO_UINT_RND: return "X86ISD::FP_TO_UINT_RND";
20138 case X86ISD::VFPCLASS: return "X86ISD::VFPCLASS";
20143 // isLegalAddressingMode - Return true if the addressing mode represented
20144 // by AM is legal for this target, for a load/store of the specified type.
20145 bool X86TargetLowering::isLegalAddressingMode(const DataLayout &DL,
20146 const AddrMode &AM, Type *Ty,
20147 unsigned AS) const {
20148 // X86 supports extremely general addressing modes.
20149 CodeModel::Model M = getTargetMachine().getCodeModel();
20150 Reloc::Model R = getTargetMachine().getRelocationModel();
20152 // X86 allows a sign-extended 32-bit immediate field as a displacement.
20153 if (!X86::isOffsetSuitableForCodeModel(AM.BaseOffs, M, AM.BaseGV != nullptr))
20158 Subtarget->ClassifyGlobalReference(AM.BaseGV, getTargetMachine());
20160 // If a reference to this global requires an extra load, we can't fold it.
20161 if (isGlobalStubReference(GVFlags))
20164 // If BaseGV requires a register for the PIC base, we cannot also have a
20165 // BaseReg specified.
20166 if (AM.HasBaseReg && isGlobalRelativeToPICBase(GVFlags))
20169 // If lower 4G is not available, then we must use rip-relative addressing.
20170 if ((M != CodeModel::Small || R != Reloc::Static) &&
20171 Subtarget->is64Bit() && (AM.BaseOffs || AM.Scale > 1))
20175 switch (AM.Scale) {
20181 // These scales always work.
20186 // These scales are formed with basereg+scalereg. Only accept if there is
20191 default: // Other stuff never works.
20198 bool X86TargetLowering::isVectorShiftByScalarCheap(Type *Ty) const {
20199 unsigned Bits = Ty->getScalarSizeInBits();
20201 // 8-bit shifts are always expensive, but versions with a scalar amount aren't
20202 // particularly cheaper than those without.
20206 // On AVX2 there are new vpsllv[dq] instructions (and other shifts), that make
20207 // variable shifts just as cheap as scalar ones.
20208 if (Subtarget->hasInt256() && (Bits == 32 || Bits == 64))
20211 // Otherwise, it's significantly cheaper to shift by a scalar amount than by a
20212 // fully general vector.
20216 bool X86TargetLowering::isTruncateFree(Type *Ty1, Type *Ty2) const {
20217 if (!Ty1->isIntegerTy() || !Ty2->isIntegerTy())
20219 unsigned NumBits1 = Ty1->getPrimitiveSizeInBits();
20220 unsigned NumBits2 = Ty2->getPrimitiveSizeInBits();
20221 return NumBits1 > NumBits2;
20224 bool X86TargetLowering::allowTruncateForTailCall(Type *Ty1, Type *Ty2) const {
20225 if (!Ty1->isIntegerTy() || !Ty2->isIntegerTy())
20228 if (!isTypeLegal(EVT::getEVT(Ty1)))
20231 assert(Ty1->getPrimitiveSizeInBits() <= 64 && "i128 is probably not a noop");
20233 // Assuming the caller doesn't have a zeroext or signext return parameter,
20234 // truncation all the way down to i1 is valid.
20238 bool X86TargetLowering::isLegalICmpImmediate(int64_t Imm) const {
20239 return isInt<32>(Imm);
20242 bool X86TargetLowering::isLegalAddImmediate(int64_t Imm) const {
20243 // Can also use sub to handle negated immediates.
20244 return isInt<32>(Imm);
20247 bool X86TargetLowering::isTruncateFree(EVT VT1, EVT VT2) const {
20248 if (!VT1.isInteger() || !VT2.isInteger())
20250 unsigned NumBits1 = VT1.getSizeInBits();
20251 unsigned NumBits2 = VT2.getSizeInBits();
20252 return NumBits1 > NumBits2;
20255 bool X86TargetLowering::isZExtFree(Type *Ty1, Type *Ty2) const {
20256 // x86-64 implicitly zero-extends 32-bit results in 64-bit registers.
20257 return Ty1->isIntegerTy(32) && Ty2->isIntegerTy(64) && Subtarget->is64Bit();
20260 bool X86TargetLowering::isZExtFree(EVT VT1, EVT VT2) const {
20261 // x86-64 implicitly zero-extends 32-bit results in 64-bit registers.
20262 return VT1 == MVT::i32 && VT2 == MVT::i64 && Subtarget->is64Bit();
20265 bool X86TargetLowering::isZExtFree(SDValue Val, EVT VT2) const {
20266 EVT VT1 = Val.getValueType();
20267 if (isZExtFree(VT1, VT2))
20270 if (Val.getOpcode() != ISD::LOAD)
20273 if (!VT1.isSimple() || !VT1.isInteger() ||
20274 !VT2.isSimple() || !VT2.isInteger())
20277 switch (VT1.getSimpleVT().SimpleTy) {
20282 // X86 has 8, 16, and 32-bit zero-extending loads.
20289 bool X86TargetLowering::isVectorLoadExtDesirable(SDValue) const { return true; }
20292 X86TargetLowering::isFMAFasterThanFMulAndFAdd(EVT VT) const {
20293 if (!(Subtarget->hasFMA() || Subtarget->hasFMA4() || Subtarget->hasAVX512()))
20296 VT = VT.getScalarType();
20298 if (!VT.isSimple())
20301 switch (VT.getSimpleVT().SimpleTy) {
20312 bool X86TargetLowering::isNarrowingProfitable(EVT VT1, EVT VT2) const {
20313 // i16 instructions are longer (0x66 prefix) and potentially slower.
20314 return !(VT1 == MVT::i32 && VT2 == MVT::i16);
20317 /// isShuffleMaskLegal - Targets can use this to indicate that they only
20318 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
20319 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
20320 /// are assumed to be legal.
20322 X86TargetLowering::isShuffleMaskLegal(const SmallVectorImpl<int> &M,
20324 if (!VT.isSimple())
20327 // Not for i1 vectors
20328 if (VT.getScalarType() == MVT::i1)
20331 // Very little shuffling can be done for 64-bit vectors right now.
20332 if (VT.getSizeInBits() == 64)
20335 // We only care that the types being shuffled are legal. The lowering can
20336 // handle any possible shuffle mask that results.
20337 return isTypeLegal(VT.getSimpleVT());
20341 X86TargetLowering::isVectorClearMaskLegal(const SmallVectorImpl<int> &Mask,
20343 // Just delegate to the generic legality, clear masks aren't special.
20344 return isShuffleMaskLegal(Mask, VT);
20347 //===----------------------------------------------------------------------===//
20348 // X86 Scheduler Hooks
20349 //===----------------------------------------------------------------------===//
20351 /// Utility function to emit xbegin specifying the start of an RTM region.
20352 static MachineBasicBlock *EmitXBegin(MachineInstr *MI, MachineBasicBlock *MBB,
20353 const TargetInstrInfo *TII) {
20354 DebugLoc DL = MI->getDebugLoc();
20356 const BasicBlock *BB = MBB->getBasicBlock();
20357 MachineFunction::iterator I = ++MBB->getIterator();
20359 // For the v = xbegin(), we generate
20370 MachineBasicBlock *thisMBB = MBB;
20371 MachineFunction *MF = MBB->getParent();
20372 MachineBasicBlock *mainMBB = MF->CreateMachineBasicBlock(BB);
20373 MachineBasicBlock *sinkMBB = MF->CreateMachineBasicBlock(BB);
20374 MF->insert(I, mainMBB);
20375 MF->insert(I, sinkMBB);
20377 // Transfer the remainder of BB and its successor edges to sinkMBB.
20378 sinkMBB->splice(sinkMBB->begin(), MBB,
20379 std::next(MachineBasicBlock::iterator(MI)), MBB->end());
20380 sinkMBB->transferSuccessorsAndUpdatePHIs(MBB);
20384 // # fallthrough to mainMBB
20385 // # abortion to sinkMBB
20386 BuildMI(thisMBB, DL, TII->get(X86::XBEGIN_4)).addMBB(sinkMBB);
20387 thisMBB->addSuccessor(mainMBB);
20388 thisMBB->addSuccessor(sinkMBB);
20392 BuildMI(mainMBB, DL, TII->get(X86::MOV32ri), X86::EAX).addImm(-1);
20393 mainMBB->addSuccessor(sinkMBB);
20396 // EAX is live into the sinkMBB
20397 sinkMBB->addLiveIn(X86::EAX);
20398 BuildMI(*sinkMBB, sinkMBB->begin(), DL,
20399 TII->get(TargetOpcode::COPY), MI->getOperand(0).getReg())
20402 MI->eraseFromParent();
20406 // FIXME: When we get size specific XMM0 registers, i.e. XMM0_V16I8
20407 // or XMM0_V32I8 in AVX all of this code can be replaced with that
20408 // in the .td file.
20409 static MachineBasicBlock *EmitPCMPSTRM(MachineInstr *MI, MachineBasicBlock *BB,
20410 const TargetInstrInfo *TII) {
20412 switch (MI->getOpcode()) {
20413 default: llvm_unreachable("illegal opcode!");
20414 case X86::PCMPISTRM128REG: Opc = X86::PCMPISTRM128rr; break;
20415 case X86::VPCMPISTRM128REG: Opc = X86::VPCMPISTRM128rr; break;
20416 case X86::PCMPISTRM128MEM: Opc = X86::PCMPISTRM128rm; break;
20417 case X86::VPCMPISTRM128MEM: Opc = X86::VPCMPISTRM128rm; break;
20418 case X86::PCMPESTRM128REG: Opc = X86::PCMPESTRM128rr; break;
20419 case X86::VPCMPESTRM128REG: Opc = X86::VPCMPESTRM128rr; break;
20420 case X86::PCMPESTRM128MEM: Opc = X86::PCMPESTRM128rm; break;
20421 case X86::VPCMPESTRM128MEM: Opc = X86::VPCMPESTRM128rm; break;
20424 DebugLoc dl = MI->getDebugLoc();
20425 MachineInstrBuilder MIB = BuildMI(*BB, MI, dl, TII->get(Opc));
20427 unsigned NumArgs = MI->getNumOperands();
20428 for (unsigned i = 1; i < NumArgs; ++i) {
20429 MachineOperand &Op = MI->getOperand(i);
20430 if (!(Op.isReg() && Op.isImplicit()))
20431 MIB.addOperand(Op);
20433 if (MI->hasOneMemOperand())
20434 MIB->setMemRefs(MI->memoperands_begin(), MI->memoperands_end());
20436 BuildMI(*BB, MI, dl,
20437 TII->get(TargetOpcode::COPY), MI->getOperand(0).getReg())
20438 .addReg(X86::XMM0);
20440 MI->eraseFromParent();
20444 // FIXME: Custom handling because TableGen doesn't support multiple implicit
20445 // defs in an instruction pattern
20446 static MachineBasicBlock *EmitPCMPSTRI(MachineInstr *MI, MachineBasicBlock *BB,
20447 const TargetInstrInfo *TII) {
20449 switch (MI->getOpcode()) {
20450 default: llvm_unreachable("illegal opcode!");
20451 case X86::PCMPISTRIREG: Opc = X86::PCMPISTRIrr; break;
20452 case X86::VPCMPISTRIREG: Opc = X86::VPCMPISTRIrr; break;
20453 case X86::PCMPISTRIMEM: Opc = X86::PCMPISTRIrm; break;
20454 case X86::VPCMPISTRIMEM: Opc = X86::VPCMPISTRIrm; break;
20455 case X86::PCMPESTRIREG: Opc = X86::PCMPESTRIrr; break;
20456 case X86::VPCMPESTRIREG: Opc = X86::VPCMPESTRIrr; break;
20457 case X86::PCMPESTRIMEM: Opc = X86::PCMPESTRIrm; break;
20458 case X86::VPCMPESTRIMEM: Opc = X86::VPCMPESTRIrm; break;
20461 DebugLoc dl = MI->getDebugLoc();
20462 MachineInstrBuilder MIB = BuildMI(*BB, MI, dl, TII->get(Opc));
20464 unsigned NumArgs = MI->getNumOperands(); // remove the results
20465 for (unsigned i = 1; i < NumArgs; ++i) {
20466 MachineOperand &Op = MI->getOperand(i);
20467 if (!(Op.isReg() && Op.isImplicit()))
20468 MIB.addOperand(Op);
20470 if (MI->hasOneMemOperand())
20471 MIB->setMemRefs(MI->memoperands_begin(), MI->memoperands_end());
20473 BuildMI(*BB, MI, dl,
20474 TII->get(TargetOpcode::COPY), MI->getOperand(0).getReg())
20477 MI->eraseFromParent();
20481 static MachineBasicBlock *EmitMonitor(MachineInstr *MI, MachineBasicBlock *BB,
20482 const X86Subtarget *Subtarget) {
20483 DebugLoc dl = MI->getDebugLoc();
20484 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
20485 // Address into RAX/EAX, other two args into ECX, EDX.
20486 unsigned MemOpc = Subtarget->is64Bit() ? X86::LEA64r : X86::LEA32r;
20487 unsigned MemReg = Subtarget->is64Bit() ? X86::RAX : X86::EAX;
20488 MachineInstrBuilder MIB = BuildMI(*BB, MI, dl, TII->get(MemOpc), MemReg);
20489 for (int i = 0; i < X86::AddrNumOperands; ++i)
20490 MIB.addOperand(MI->getOperand(i));
20492 unsigned ValOps = X86::AddrNumOperands;
20493 BuildMI(*BB, MI, dl, TII->get(TargetOpcode::COPY), X86::ECX)
20494 .addReg(MI->getOperand(ValOps).getReg());
20495 BuildMI(*BB, MI, dl, TII->get(TargetOpcode::COPY), X86::EDX)
20496 .addReg(MI->getOperand(ValOps+1).getReg());
20498 // The instruction doesn't actually take any operands though.
20499 BuildMI(*BB, MI, dl, TII->get(X86::MONITORrrr));
20501 MI->eraseFromParent(); // The pseudo is gone now.
20505 MachineBasicBlock *
20506 X86TargetLowering::EmitVAARG64WithCustomInserter(MachineInstr *MI,
20507 MachineBasicBlock *MBB) const {
20508 // Emit va_arg instruction on X86-64.
20510 // Operands to this pseudo-instruction:
20511 // 0 ) Output : destination address (reg)
20512 // 1-5) Input : va_list address (addr, i64mem)
20513 // 6 ) ArgSize : Size (in bytes) of vararg type
20514 // 7 ) ArgMode : 0=overflow only, 1=use gp_offset, 2=use fp_offset
20515 // 8 ) Align : Alignment of type
20516 // 9 ) EFLAGS (implicit-def)
20518 assert(MI->getNumOperands() == 10 && "VAARG_64 should have 10 operands!");
20519 static_assert(X86::AddrNumOperands == 5,
20520 "VAARG_64 assumes 5 address operands");
20522 unsigned DestReg = MI->getOperand(0).getReg();
20523 MachineOperand &Base = MI->getOperand(1);
20524 MachineOperand &Scale = MI->getOperand(2);
20525 MachineOperand &Index = MI->getOperand(3);
20526 MachineOperand &Disp = MI->getOperand(4);
20527 MachineOperand &Segment = MI->getOperand(5);
20528 unsigned ArgSize = MI->getOperand(6).getImm();
20529 unsigned ArgMode = MI->getOperand(7).getImm();
20530 unsigned Align = MI->getOperand(8).getImm();
20532 // Memory Reference
20533 assert(MI->hasOneMemOperand() && "Expected VAARG_64 to have one memoperand");
20534 MachineInstr::mmo_iterator MMOBegin = MI->memoperands_begin();
20535 MachineInstr::mmo_iterator MMOEnd = MI->memoperands_end();
20537 // Machine Information
20538 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
20539 MachineRegisterInfo &MRI = MBB->getParent()->getRegInfo();
20540 const TargetRegisterClass *AddrRegClass = getRegClassFor(MVT::i64);
20541 const TargetRegisterClass *OffsetRegClass = getRegClassFor(MVT::i32);
20542 DebugLoc DL = MI->getDebugLoc();
20544 // struct va_list {
20547 // i64 overflow_area (address)
20548 // i64 reg_save_area (address)
20550 // sizeof(va_list) = 24
20551 // alignment(va_list) = 8
20553 unsigned TotalNumIntRegs = 6;
20554 unsigned TotalNumXMMRegs = 8;
20555 bool UseGPOffset = (ArgMode == 1);
20556 bool UseFPOffset = (ArgMode == 2);
20557 unsigned MaxOffset = TotalNumIntRegs * 8 +
20558 (UseFPOffset ? TotalNumXMMRegs * 16 : 0);
20560 /* Align ArgSize to a multiple of 8 */
20561 unsigned ArgSizeA8 = (ArgSize + 7) & ~7;
20562 bool NeedsAlign = (Align > 8);
20564 MachineBasicBlock *thisMBB = MBB;
20565 MachineBasicBlock *overflowMBB;
20566 MachineBasicBlock *offsetMBB;
20567 MachineBasicBlock *endMBB;
20569 unsigned OffsetDestReg = 0; // Argument address computed by offsetMBB
20570 unsigned OverflowDestReg = 0; // Argument address computed by overflowMBB
20571 unsigned OffsetReg = 0;
20573 if (!UseGPOffset && !UseFPOffset) {
20574 // If we only pull from the overflow region, we don't create a branch.
20575 // We don't need to alter control flow.
20576 OffsetDestReg = 0; // unused
20577 OverflowDestReg = DestReg;
20579 offsetMBB = nullptr;
20580 overflowMBB = thisMBB;
20583 // First emit code to check if gp_offset (or fp_offset) is below the bound.
20584 // If so, pull the argument from reg_save_area. (branch to offsetMBB)
20585 // If not, pull from overflow_area. (branch to overflowMBB)
20590 // offsetMBB overflowMBB
20595 // Registers for the PHI in endMBB
20596 OffsetDestReg = MRI.createVirtualRegister(AddrRegClass);
20597 OverflowDestReg = MRI.createVirtualRegister(AddrRegClass);
20599 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
20600 MachineFunction *MF = MBB->getParent();
20601 overflowMBB = MF->CreateMachineBasicBlock(LLVM_BB);
20602 offsetMBB = MF->CreateMachineBasicBlock(LLVM_BB);
20603 endMBB = MF->CreateMachineBasicBlock(LLVM_BB);
20605 MachineFunction::iterator MBBIter = ++MBB->getIterator();
20607 // Insert the new basic blocks
20608 MF->insert(MBBIter, offsetMBB);
20609 MF->insert(MBBIter, overflowMBB);
20610 MF->insert(MBBIter, endMBB);
20612 // Transfer the remainder of MBB and its successor edges to endMBB.
20613 endMBB->splice(endMBB->begin(), thisMBB,
20614 std::next(MachineBasicBlock::iterator(MI)), thisMBB->end());
20615 endMBB->transferSuccessorsAndUpdatePHIs(thisMBB);
20617 // Make offsetMBB and overflowMBB successors of thisMBB
20618 thisMBB->addSuccessor(offsetMBB);
20619 thisMBB->addSuccessor(overflowMBB);
20621 // endMBB is a successor of both offsetMBB and overflowMBB
20622 offsetMBB->addSuccessor(endMBB);
20623 overflowMBB->addSuccessor(endMBB);
20625 // Load the offset value into a register
20626 OffsetReg = MRI.createVirtualRegister(OffsetRegClass);
20627 BuildMI(thisMBB, DL, TII->get(X86::MOV32rm), OffsetReg)
20631 .addDisp(Disp, UseFPOffset ? 4 : 0)
20632 .addOperand(Segment)
20633 .setMemRefs(MMOBegin, MMOEnd);
20635 // Check if there is enough room left to pull this argument.
20636 BuildMI(thisMBB, DL, TII->get(X86::CMP32ri))
20638 .addImm(MaxOffset + 8 - ArgSizeA8);
20640 // Branch to "overflowMBB" if offset >= max
20641 // Fall through to "offsetMBB" otherwise
20642 BuildMI(thisMBB, DL, TII->get(X86::GetCondBranchFromCond(X86::COND_AE)))
20643 .addMBB(overflowMBB);
20646 // In offsetMBB, emit code to use the reg_save_area.
20648 assert(OffsetReg != 0);
20650 // Read the reg_save_area address.
20651 unsigned RegSaveReg = MRI.createVirtualRegister(AddrRegClass);
20652 BuildMI(offsetMBB, DL, TII->get(X86::MOV64rm), RegSaveReg)
20657 .addOperand(Segment)
20658 .setMemRefs(MMOBegin, MMOEnd);
20660 // Zero-extend the offset
20661 unsigned OffsetReg64 = MRI.createVirtualRegister(AddrRegClass);
20662 BuildMI(offsetMBB, DL, TII->get(X86::SUBREG_TO_REG), OffsetReg64)
20665 .addImm(X86::sub_32bit);
20667 // Add the offset to the reg_save_area to get the final address.
20668 BuildMI(offsetMBB, DL, TII->get(X86::ADD64rr), OffsetDestReg)
20669 .addReg(OffsetReg64)
20670 .addReg(RegSaveReg);
20672 // Compute the offset for the next argument
20673 unsigned NextOffsetReg = MRI.createVirtualRegister(OffsetRegClass);
20674 BuildMI(offsetMBB, DL, TII->get(X86::ADD32ri), NextOffsetReg)
20676 .addImm(UseFPOffset ? 16 : 8);
20678 // Store it back into the va_list.
20679 BuildMI(offsetMBB, DL, TII->get(X86::MOV32mr))
20683 .addDisp(Disp, UseFPOffset ? 4 : 0)
20684 .addOperand(Segment)
20685 .addReg(NextOffsetReg)
20686 .setMemRefs(MMOBegin, MMOEnd);
20689 BuildMI(offsetMBB, DL, TII->get(X86::JMP_1))
20694 // Emit code to use overflow area
20697 // Load the overflow_area address into a register.
20698 unsigned OverflowAddrReg = MRI.createVirtualRegister(AddrRegClass);
20699 BuildMI(overflowMBB, DL, TII->get(X86::MOV64rm), OverflowAddrReg)
20704 .addOperand(Segment)
20705 .setMemRefs(MMOBegin, MMOEnd);
20707 // If we need to align it, do so. Otherwise, just copy the address
20708 // to OverflowDestReg.
20710 // Align the overflow address
20711 assert((Align & (Align-1)) == 0 && "Alignment must be a power of 2");
20712 unsigned TmpReg = MRI.createVirtualRegister(AddrRegClass);
20714 // aligned_addr = (addr + (align-1)) & ~(align-1)
20715 BuildMI(overflowMBB, DL, TII->get(X86::ADD64ri32), TmpReg)
20716 .addReg(OverflowAddrReg)
20719 BuildMI(overflowMBB, DL, TII->get(X86::AND64ri32), OverflowDestReg)
20721 .addImm(~(uint64_t)(Align-1));
20723 BuildMI(overflowMBB, DL, TII->get(TargetOpcode::COPY), OverflowDestReg)
20724 .addReg(OverflowAddrReg);
20727 // Compute the next overflow address after this argument.
20728 // (the overflow address should be kept 8-byte aligned)
20729 unsigned NextAddrReg = MRI.createVirtualRegister(AddrRegClass);
20730 BuildMI(overflowMBB, DL, TII->get(X86::ADD64ri32), NextAddrReg)
20731 .addReg(OverflowDestReg)
20732 .addImm(ArgSizeA8);
20734 // Store the new overflow address.
20735 BuildMI(overflowMBB, DL, TII->get(X86::MOV64mr))
20740 .addOperand(Segment)
20741 .addReg(NextAddrReg)
20742 .setMemRefs(MMOBegin, MMOEnd);
20744 // If we branched, emit the PHI to the front of endMBB.
20746 BuildMI(*endMBB, endMBB->begin(), DL,
20747 TII->get(X86::PHI), DestReg)
20748 .addReg(OffsetDestReg).addMBB(offsetMBB)
20749 .addReg(OverflowDestReg).addMBB(overflowMBB);
20752 // Erase the pseudo instruction
20753 MI->eraseFromParent();
20758 MachineBasicBlock *
20759 X86TargetLowering::EmitVAStartSaveXMMRegsWithCustomInserter(
20761 MachineBasicBlock *MBB) const {
20762 // Emit code to save XMM registers to the stack. The ABI says that the
20763 // number of registers to save is given in %al, so it's theoretically
20764 // possible to do an indirect jump trick to avoid saving all of them,
20765 // however this code takes a simpler approach and just executes all
20766 // of the stores if %al is non-zero. It's less code, and it's probably
20767 // easier on the hardware branch predictor, and stores aren't all that
20768 // expensive anyway.
20770 // Create the new basic blocks. One block contains all the XMM stores,
20771 // and one block is the final destination regardless of whether any
20772 // stores were performed.
20773 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
20774 MachineFunction *F = MBB->getParent();
20775 MachineFunction::iterator MBBIter = ++MBB->getIterator();
20776 MachineBasicBlock *XMMSaveMBB = F->CreateMachineBasicBlock(LLVM_BB);
20777 MachineBasicBlock *EndMBB = F->CreateMachineBasicBlock(LLVM_BB);
20778 F->insert(MBBIter, XMMSaveMBB);
20779 F->insert(MBBIter, EndMBB);
20781 // Transfer the remainder of MBB and its successor edges to EndMBB.
20782 EndMBB->splice(EndMBB->begin(), MBB,
20783 std::next(MachineBasicBlock::iterator(MI)), MBB->end());
20784 EndMBB->transferSuccessorsAndUpdatePHIs(MBB);
20786 // The original block will now fall through to the XMM save block.
20787 MBB->addSuccessor(XMMSaveMBB);
20788 // The XMMSaveMBB will fall through to the end block.
20789 XMMSaveMBB->addSuccessor(EndMBB);
20791 // Now add the instructions.
20792 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
20793 DebugLoc DL = MI->getDebugLoc();
20795 unsigned CountReg = MI->getOperand(0).getReg();
20796 int64_t RegSaveFrameIndex = MI->getOperand(1).getImm();
20797 int64_t VarArgsFPOffset = MI->getOperand(2).getImm();
20799 if (!Subtarget->isCallingConvWin64(F->getFunction()->getCallingConv())) {
20800 // If %al is 0, branch around the XMM save block.
20801 BuildMI(MBB, DL, TII->get(X86::TEST8rr)).addReg(CountReg).addReg(CountReg);
20802 BuildMI(MBB, DL, TII->get(X86::JE_1)).addMBB(EndMBB);
20803 MBB->addSuccessor(EndMBB);
20806 // Make sure the last operand is EFLAGS, which gets clobbered by the branch
20807 // that was just emitted, but clearly shouldn't be "saved".
20808 assert((MI->getNumOperands() <= 3 ||
20809 !MI->getOperand(MI->getNumOperands() - 1).isReg() ||
20810 MI->getOperand(MI->getNumOperands() - 1).getReg() == X86::EFLAGS)
20811 && "Expected last argument to be EFLAGS");
20812 unsigned MOVOpc = Subtarget->hasFp256() ? X86::VMOVAPSmr : X86::MOVAPSmr;
20813 // In the XMM save block, save all the XMM argument registers.
20814 for (int i = 3, e = MI->getNumOperands() - 1; i != e; ++i) {
20815 int64_t Offset = (i - 3) * 16 + VarArgsFPOffset;
20816 MachineMemOperand *MMO = F->getMachineMemOperand(
20817 MachinePointerInfo::getFixedStack(*F, RegSaveFrameIndex, Offset),
20818 MachineMemOperand::MOStore,
20819 /*Size=*/16, /*Align=*/16);
20820 BuildMI(XMMSaveMBB, DL, TII->get(MOVOpc))
20821 .addFrameIndex(RegSaveFrameIndex)
20822 .addImm(/*Scale=*/1)
20823 .addReg(/*IndexReg=*/0)
20824 .addImm(/*Disp=*/Offset)
20825 .addReg(/*Segment=*/0)
20826 .addReg(MI->getOperand(i).getReg())
20827 .addMemOperand(MMO);
20830 MI->eraseFromParent(); // The pseudo instruction is gone now.
20835 // The EFLAGS operand of SelectItr might be missing a kill marker
20836 // because there were multiple uses of EFLAGS, and ISel didn't know
20837 // which to mark. Figure out whether SelectItr should have had a
20838 // kill marker, and set it if it should. Returns the correct kill
20840 static bool checkAndUpdateEFLAGSKill(MachineBasicBlock::iterator SelectItr,
20841 MachineBasicBlock* BB,
20842 const TargetRegisterInfo* TRI) {
20843 // Scan forward through BB for a use/def of EFLAGS.
20844 MachineBasicBlock::iterator miI(std::next(SelectItr));
20845 for (MachineBasicBlock::iterator miE = BB->end(); miI != miE; ++miI) {
20846 const MachineInstr& mi = *miI;
20847 if (mi.readsRegister(X86::EFLAGS))
20849 if (mi.definesRegister(X86::EFLAGS))
20850 break; // Should have kill-flag - update below.
20853 // If we hit the end of the block, check whether EFLAGS is live into a
20855 if (miI == BB->end()) {
20856 for (MachineBasicBlock::succ_iterator sItr = BB->succ_begin(),
20857 sEnd = BB->succ_end();
20858 sItr != sEnd; ++sItr) {
20859 MachineBasicBlock* succ = *sItr;
20860 if (succ->isLiveIn(X86::EFLAGS))
20865 // We found a def, or hit the end of the basic block and EFLAGS wasn't live
20866 // out. SelectMI should have a kill flag on EFLAGS.
20867 SelectItr->addRegisterKilled(X86::EFLAGS, TRI);
20871 // Return true if it is OK for this CMOV pseudo-opcode to be cascaded
20872 // together with other CMOV pseudo-opcodes into a single basic-block with
20873 // conditional jump around it.
20874 static bool isCMOVPseudo(MachineInstr *MI) {
20875 switch (MI->getOpcode()) {
20876 case X86::CMOV_FR32:
20877 case X86::CMOV_FR64:
20878 case X86::CMOV_GR8:
20879 case X86::CMOV_GR16:
20880 case X86::CMOV_GR32:
20881 case X86::CMOV_RFP32:
20882 case X86::CMOV_RFP64:
20883 case X86::CMOV_RFP80:
20884 case X86::CMOV_V2F64:
20885 case X86::CMOV_V2I64:
20886 case X86::CMOV_V4F32:
20887 case X86::CMOV_V4F64:
20888 case X86::CMOV_V4I64:
20889 case X86::CMOV_V16F32:
20890 case X86::CMOV_V8F32:
20891 case X86::CMOV_V8F64:
20892 case X86::CMOV_V8I64:
20893 case X86::CMOV_V8I1:
20894 case X86::CMOV_V16I1:
20895 case X86::CMOV_V32I1:
20896 case X86::CMOV_V64I1:
20904 MachineBasicBlock *
20905 X86TargetLowering::EmitLoweredSelect(MachineInstr *MI,
20906 MachineBasicBlock *BB) const {
20907 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
20908 DebugLoc DL = MI->getDebugLoc();
20910 // To "insert" a SELECT_CC instruction, we actually have to insert the
20911 // diamond control-flow pattern. The incoming instruction knows the
20912 // destination vreg to set, the condition code register to branch on, the
20913 // true/false values to select between, and a branch opcode to use.
20914 const BasicBlock *LLVM_BB = BB->getBasicBlock();
20915 MachineFunction::iterator It = ++BB->getIterator();
20920 // cmpTY ccX, r1, r2
20922 // fallthrough --> copy0MBB
20923 MachineBasicBlock *thisMBB = BB;
20924 MachineFunction *F = BB->getParent();
20926 // This code lowers all pseudo-CMOV instructions. Generally it lowers these
20927 // as described above, by inserting a BB, and then making a PHI at the join
20928 // point to select the true and false operands of the CMOV in the PHI.
20930 // The code also handles two different cases of multiple CMOV opcodes
20934 // In this case, there are multiple CMOVs in a row, all which are based on
20935 // the same condition setting (or the exact opposite condition setting).
20936 // In this case we can lower all the CMOVs using a single inserted BB, and
20937 // then make a number of PHIs at the join point to model the CMOVs. The only
20938 // trickiness here, is that in a case like:
20940 // t2 = CMOV cond1 t1, f1
20941 // t3 = CMOV cond1 t2, f2
20943 // when rewriting this into PHIs, we have to perform some renaming on the
20944 // temps since you cannot have a PHI operand refer to a PHI result earlier
20945 // in the same block. The "simple" but wrong lowering would be:
20947 // t2 = PHI t1(BB1), f1(BB2)
20948 // t3 = PHI t2(BB1), f2(BB2)
20950 // but clearly t2 is not defined in BB1, so that is incorrect. The proper
20951 // renaming is to note that on the path through BB1, t2 is really just a
20952 // copy of t1, and do that renaming, properly generating:
20954 // t2 = PHI t1(BB1), f1(BB2)
20955 // t3 = PHI t1(BB1), f2(BB2)
20957 // Case 2, we lower cascaded CMOVs such as
20959 // (CMOV (CMOV F, T, cc1), T, cc2)
20961 // to two successives branches. For that, we look for another CMOV as the
20962 // following instruction.
20964 // Without this, we would add a PHI between the two jumps, which ends up
20965 // creating a few copies all around. For instance, for
20967 // (sitofp (zext (fcmp une)))
20969 // we would generate:
20971 // ucomiss %xmm1, %xmm0
20972 // movss <1.0f>, %xmm0
20973 // movaps %xmm0, %xmm1
20975 // xorps %xmm1, %xmm1
20978 // movaps %xmm1, %xmm0
20982 // because this custom-inserter would have generated:
20994 // A: X = ...; Y = ...
20996 // C: Z = PHI [X, A], [Y, B]
20998 // E: PHI [X, C], [Z, D]
21000 // If we lower both CMOVs in a single step, we can instead generate:
21012 // A: X = ...; Y = ...
21014 // E: PHI [X, A], [X, C], [Y, D]
21016 // Which, in our sitofp/fcmp example, gives us something like:
21018 // ucomiss %xmm1, %xmm0
21019 // movss <1.0f>, %xmm0
21022 // xorps %xmm0, %xmm0
21026 MachineInstr *CascadedCMOV = nullptr;
21027 MachineInstr *LastCMOV = MI;
21028 X86::CondCode CC = X86::CondCode(MI->getOperand(3).getImm());
21029 X86::CondCode OppCC = X86::GetOppositeBranchCondition(CC);
21030 MachineBasicBlock::iterator NextMIIt =
21031 std::next(MachineBasicBlock::iterator(MI));
21033 // Check for case 1, where there are multiple CMOVs with the same condition
21034 // first. Of the two cases of multiple CMOV lowerings, case 1 reduces the
21035 // number of jumps the most.
21037 if (isCMOVPseudo(MI)) {
21038 // See if we have a string of CMOVS with the same condition.
21039 while (NextMIIt != BB->end() &&
21040 isCMOVPseudo(NextMIIt) &&
21041 (NextMIIt->getOperand(3).getImm() == CC ||
21042 NextMIIt->getOperand(3).getImm() == OppCC)) {
21043 LastCMOV = &*NextMIIt;
21048 // This checks for case 2, but only do this if we didn't already find
21049 // case 1, as indicated by LastCMOV == MI.
21050 if (LastCMOV == MI &&
21051 NextMIIt != BB->end() && NextMIIt->getOpcode() == MI->getOpcode() &&
21052 NextMIIt->getOperand(2).getReg() == MI->getOperand(2).getReg() &&
21053 NextMIIt->getOperand(1).getReg() == MI->getOperand(0).getReg()) {
21054 CascadedCMOV = &*NextMIIt;
21057 MachineBasicBlock *jcc1MBB = nullptr;
21059 // If we have a cascaded CMOV, we lower it to two successive branches to
21060 // the same block. EFLAGS is used by both, so mark it as live in the second.
21061 if (CascadedCMOV) {
21062 jcc1MBB = F->CreateMachineBasicBlock(LLVM_BB);
21063 F->insert(It, jcc1MBB);
21064 jcc1MBB->addLiveIn(X86::EFLAGS);
21067 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
21068 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
21069 F->insert(It, copy0MBB);
21070 F->insert(It, sinkMBB);
21072 // If the EFLAGS register isn't dead in the terminator, then claim that it's
21073 // live into the sink and copy blocks.
21074 const TargetRegisterInfo *TRI = Subtarget->getRegisterInfo();
21076 MachineInstr *LastEFLAGSUser = CascadedCMOV ? CascadedCMOV : LastCMOV;
21077 if (!LastEFLAGSUser->killsRegister(X86::EFLAGS) &&
21078 !checkAndUpdateEFLAGSKill(LastEFLAGSUser, BB, TRI)) {
21079 copy0MBB->addLiveIn(X86::EFLAGS);
21080 sinkMBB->addLiveIn(X86::EFLAGS);
21083 // Transfer the remainder of BB and its successor edges to sinkMBB.
21084 sinkMBB->splice(sinkMBB->begin(), BB,
21085 std::next(MachineBasicBlock::iterator(LastCMOV)), BB->end());
21086 sinkMBB->transferSuccessorsAndUpdatePHIs(BB);
21088 // Add the true and fallthrough blocks as its successors.
21089 if (CascadedCMOV) {
21090 // The fallthrough block may be jcc1MBB, if we have a cascaded CMOV.
21091 BB->addSuccessor(jcc1MBB);
21093 // In that case, jcc1MBB will itself fallthrough the copy0MBB, and
21094 // jump to the sinkMBB.
21095 jcc1MBB->addSuccessor(copy0MBB);
21096 jcc1MBB->addSuccessor(sinkMBB);
21098 BB->addSuccessor(copy0MBB);
21101 // The true block target of the first (or only) branch is always sinkMBB.
21102 BB->addSuccessor(sinkMBB);
21104 // Create the conditional branch instruction.
21105 unsigned Opc = X86::GetCondBranchFromCond(CC);
21106 BuildMI(BB, DL, TII->get(Opc)).addMBB(sinkMBB);
21108 if (CascadedCMOV) {
21109 unsigned Opc2 = X86::GetCondBranchFromCond(
21110 (X86::CondCode)CascadedCMOV->getOperand(3).getImm());
21111 BuildMI(jcc1MBB, DL, TII->get(Opc2)).addMBB(sinkMBB);
21115 // %FalseValue = ...
21116 // # fallthrough to sinkMBB
21117 copy0MBB->addSuccessor(sinkMBB);
21120 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
21122 MachineBasicBlock::iterator MIItBegin = MachineBasicBlock::iterator(MI);
21123 MachineBasicBlock::iterator MIItEnd =
21124 std::next(MachineBasicBlock::iterator(LastCMOV));
21125 MachineBasicBlock::iterator SinkInsertionPoint = sinkMBB->begin();
21126 DenseMap<unsigned, std::pair<unsigned, unsigned>> RegRewriteTable;
21127 MachineInstrBuilder MIB;
21129 // As we are creating the PHIs, we have to be careful if there is more than
21130 // one. Later CMOVs may reference the results of earlier CMOVs, but later
21131 // PHIs have to reference the individual true/false inputs from earlier PHIs.
21132 // That also means that PHI construction must work forward from earlier to
21133 // later, and that the code must maintain a mapping from earlier PHI's
21134 // destination registers, and the registers that went into the PHI.
21136 for (MachineBasicBlock::iterator MIIt = MIItBegin; MIIt != MIItEnd; ++MIIt) {
21137 unsigned DestReg = MIIt->getOperand(0).getReg();
21138 unsigned Op1Reg = MIIt->getOperand(1).getReg();
21139 unsigned Op2Reg = MIIt->getOperand(2).getReg();
21141 // If this CMOV we are generating is the opposite condition from
21142 // the jump we generated, then we have to swap the operands for the
21143 // PHI that is going to be generated.
21144 if (MIIt->getOperand(3).getImm() == OppCC)
21145 std::swap(Op1Reg, Op2Reg);
21147 if (RegRewriteTable.find(Op1Reg) != RegRewriteTable.end())
21148 Op1Reg = RegRewriteTable[Op1Reg].first;
21150 if (RegRewriteTable.find(Op2Reg) != RegRewriteTable.end())
21151 Op2Reg = RegRewriteTable[Op2Reg].second;
21153 MIB = BuildMI(*sinkMBB, SinkInsertionPoint, DL,
21154 TII->get(X86::PHI), DestReg)
21155 .addReg(Op1Reg).addMBB(copy0MBB)
21156 .addReg(Op2Reg).addMBB(thisMBB);
21158 // Add this PHI to the rewrite table.
21159 RegRewriteTable[DestReg] = std::make_pair(Op1Reg, Op2Reg);
21162 // If we have a cascaded CMOV, the second Jcc provides the same incoming
21163 // value as the first Jcc (the True operand of the SELECT_CC/CMOV nodes).
21164 if (CascadedCMOV) {
21165 MIB.addReg(MI->getOperand(2).getReg()).addMBB(jcc1MBB);
21166 // Copy the PHI result to the register defined by the second CMOV.
21167 BuildMI(*sinkMBB, std::next(MachineBasicBlock::iterator(MIB.getInstr())),
21168 DL, TII->get(TargetOpcode::COPY),
21169 CascadedCMOV->getOperand(0).getReg())
21170 .addReg(MI->getOperand(0).getReg());
21171 CascadedCMOV->eraseFromParent();
21174 // Now remove the CMOV(s).
21175 for (MachineBasicBlock::iterator MIIt = MIItBegin; MIIt != MIItEnd; )
21176 (MIIt++)->eraseFromParent();
21181 MachineBasicBlock *
21182 X86TargetLowering::EmitLoweredAtomicFP(MachineInstr *MI,
21183 MachineBasicBlock *BB) const {
21184 // Combine the following atomic floating-point modification pattern:
21185 // a.store(reg OP a.load(acquire), release)
21186 // Transform them into:
21187 // OPss (%gpr), %xmm
21188 // movss %xmm, (%gpr)
21189 // Or sd equivalent for 64-bit operations.
21191 switch (MI->getOpcode()) {
21192 default: llvm_unreachable("unexpected instr type for EmitLoweredAtomicFP");
21193 case X86::RELEASE_FADD32mr: MOp = X86::MOVSSmr; FOp = X86::ADDSSrm; break;
21194 case X86::RELEASE_FADD64mr: MOp = X86::MOVSDmr; FOp = X86::ADDSDrm; break;
21196 const X86InstrInfo *TII = Subtarget->getInstrInfo();
21197 DebugLoc DL = MI->getDebugLoc();
21198 MachineRegisterInfo &MRI = BB->getParent()->getRegInfo();
21199 MachineOperand MSrc = MI->getOperand(0);
21200 unsigned VSrc = MI->getOperand(5).getReg();
21201 const MachineOperand &Disp = MI->getOperand(3);
21202 MachineOperand ZeroDisp = MachineOperand::CreateImm(0);
21203 bool hasDisp = Disp.isGlobal() || Disp.isImm();
21204 if (hasDisp && MSrc.isReg())
21205 MSrc.setIsKill(false);
21206 MachineInstrBuilder MIM = BuildMI(*BB, MI, DL, TII->get(MOp))
21207 .addOperand(/*Base=*/MSrc)
21208 .addImm(/*Scale=*/1)
21209 .addReg(/*Index=*/0)
21210 .addDisp(hasDisp ? Disp : ZeroDisp, /*off=*/0)
21212 MachineInstr *MIO = BuildMI(*BB, (MachineInstr *)MIM, DL, TII->get(FOp),
21213 MRI.createVirtualRegister(MRI.getRegClass(VSrc)))
21215 .addOperand(/*Base=*/MSrc)
21216 .addImm(/*Scale=*/1)
21217 .addReg(/*Index=*/0)
21218 .addDisp(hasDisp ? Disp : ZeroDisp, /*off=*/0)
21219 .addReg(/*Segment=*/0);
21220 MIM.addReg(MIO->getOperand(0).getReg(), RegState::Kill);
21221 MI->eraseFromParent(); // The pseudo instruction is gone now.
21225 MachineBasicBlock *
21226 X86TargetLowering::EmitLoweredSegAlloca(MachineInstr *MI,
21227 MachineBasicBlock *BB) const {
21228 MachineFunction *MF = BB->getParent();
21229 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
21230 DebugLoc DL = MI->getDebugLoc();
21231 const BasicBlock *LLVM_BB = BB->getBasicBlock();
21233 assert(MF->shouldSplitStack());
21235 const bool Is64Bit = Subtarget->is64Bit();
21236 const bool IsLP64 = Subtarget->isTarget64BitLP64();
21238 const unsigned TlsReg = Is64Bit ? X86::FS : X86::GS;
21239 const unsigned TlsOffset = IsLP64 ? 0x70 : Is64Bit ? 0x40 : 0x30;
21242 // ... [Till the alloca]
21243 // If stacklet is not large enough, jump to mallocMBB
21246 // Allocate by subtracting from RSP
21247 // Jump to continueMBB
21250 // Allocate by call to runtime
21254 // [rest of original BB]
21257 MachineBasicBlock *mallocMBB = MF->CreateMachineBasicBlock(LLVM_BB);
21258 MachineBasicBlock *bumpMBB = MF->CreateMachineBasicBlock(LLVM_BB);
21259 MachineBasicBlock *continueMBB = MF->CreateMachineBasicBlock(LLVM_BB);
21261 MachineRegisterInfo &MRI = MF->getRegInfo();
21262 const TargetRegisterClass *AddrRegClass =
21263 getRegClassFor(getPointerTy(MF->getDataLayout()));
21265 unsigned mallocPtrVReg = MRI.createVirtualRegister(AddrRegClass),
21266 bumpSPPtrVReg = MRI.createVirtualRegister(AddrRegClass),
21267 tmpSPVReg = MRI.createVirtualRegister(AddrRegClass),
21268 SPLimitVReg = MRI.createVirtualRegister(AddrRegClass),
21269 sizeVReg = MI->getOperand(1).getReg(),
21270 physSPReg = IsLP64 || Subtarget->isTargetNaCl64() ? X86::RSP : X86::ESP;
21272 MachineFunction::iterator MBBIter = ++BB->getIterator();
21274 MF->insert(MBBIter, bumpMBB);
21275 MF->insert(MBBIter, mallocMBB);
21276 MF->insert(MBBIter, continueMBB);
21278 continueMBB->splice(continueMBB->begin(), BB,
21279 std::next(MachineBasicBlock::iterator(MI)), BB->end());
21280 continueMBB->transferSuccessorsAndUpdatePHIs(BB);
21282 // Add code to the main basic block to check if the stack limit has been hit,
21283 // and if so, jump to mallocMBB otherwise to bumpMBB.
21284 BuildMI(BB, DL, TII->get(TargetOpcode::COPY), tmpSPVReg).addReg(physSPReg);
21285 BuildMI(BB, DL, TII->get(IsLP64 ? X86::SUB64rr:X86::SUB32rr), SPLimitVReg)
21286 .addReg(tmpSPVReg).addReg(sizeVReg);
21287 BuildMI(BB, DL, TII->get(IsLP64 ? X86::CMP64mr:X86::CMP32mr))
21288 .addReg(0).addImm(1).addReg(0).addImm(TlsOffset).addReg(TlsReg)
21289 .addReg(SPLimitVReg);
21290 BuildMI(BB, DL, TII->get(X86::JG_1)).addMBB(mallocMBB);
21292 // bumpMBB simply decreases the stack pointer, since we know the current
21293 // stacklet has enough space.
21294 BuildMI(bumpMBB, DL, TII->get(TargetOpcode::COPY), physSPReg)
21295 .addReg(SPLimitVReg);
21296 BuildMI(bumpMBB, DL, TII->get(TargetOpcode::COPY), bumpSPPtrVReg)
21297 .addReg(SPLimitVReg);
21298 BuildMI(bumpMBB, DL, TII->get(X86::JMP_1)).addMBB(continueMBB);
21300 // Calls into a routine in libgcc to allocate more space from the heap.
21301 const uint32_t *RegMask =
21302 Subtarget->getRegisterInfo()->getCallPreservedMask(*MF, CallingConv::C);
21304 BuildMI(mallocMBB, DL, TII->get(X86::MOV64rr), X86::RDI)
21306 BuildMI(mallocMBB, DL, TII->get(X86::CALL64pcrel32))
21307 .addExternalSymbol("__morestack_allocate_stack_space")
21308 .addRegMask(RegMask)
21309 .addReg(X86::RDI, RegState::Implicit)
21310 .addReg(X86::RAX, RegState::ImplicitDefine);
21311 } else if (Is64Bit) {
21312 BuildMI(mallocMBB, DL, TII->get(X86::MOV32rr), X86::EDI)
21314 BuildMI(mallocMBB, DL, TII->get(X86::CALL64pcrel32))
21315 .addExternalSymbol("__morestack_allocate_stack_space")
21316 .addRegMask(RegMask)
21317 .addReg(X86::EDI, RegState::Implicit)
21318 .addReg(X86::EAX, RegState::ImplicitDefine);
21320 BuildMI(mallocMBB, DL, TII->get(X86::SUB32ri), physSPReg).addReg(physSPReg)
21322 BuildMI(mallocMBB, DL, TII->get(X86::PUSH32r)).addReg(sizeVReg);
21323 BuildMI(mallocMBB, DL, TII->get(X86::CALLpcrel32))
21324 .addExternalSymbol("__morestack_allocate_stack_space")
21325 .addRegMask(RegMask)
21326 .addReg(X86::EAX, RegState::ImplicitDefine);
21330 BuildMI(mallocMBB, DL, TII->get(X86::ADD32ri), physSPReg).addReg(physSPReg)
21333 BuildMI(mallocMBB, DL, TII->get(TargetOpcode::COPY), mallocPtrVReg)
21334 .addReg(IsLP64 ? X86::RAX : X86::EAX);
21335 BuildMI(mallocMBB, DL, TII->get(X86::JMP_1)).addMBB(continueMBB);
21337 // Set up the CFG correctly.
21338 BB->addSuccessor(bumpMBB);
21339 BB->addSuccessor(mallocMBB);
21340 mallocMBB->addSuccessor(continueMBB);
21341 bumpMBB->addSuccessor(continueMBB);
21343 // Take care of the PHI nodes.
21344 BuildMI(*continueMBB, continueMBB->begin(), DL, TII->get(X86::PHI),
21345 MI->getOperand(0).getReg())
21346 .addReg(mallocPtrVReg).addMBB(mallocMBB)
21347 .addReg(bumpSPPtrVReg).addMBB(bumpMBB);
21349 // Delete the original pseudo instruction.
21350 MI->eraseFromParent();
21353 return continueMBB;
21356 MachineBasicBlock *
21357 X86TargetLowering::EmitLoweredWinAlloca(MachineInstr *MI,
21358 MachineBasicBlock *BB) const {
21359 DebugLoc DL = MI->getDebugLoc();
21361 assert(!Subtarget->isTargetMachO());
21363 Subtarget->getFrameLowering()->emitStackProbeCall(*BB->getParent(), *BB, MI,
21366 MI->eraseFromParent(); // The pseudo instruction is gone now.
21370 MachineBasicBlock *
21371 X86TargetLowering::EmitLoweredTLSCall(MachineInstr *MI,
21372 MachineBasicBlock *BB) const {
21373 // This is pretty easy. We're taking the value that we received from
21374 // our load from the relocation, sticking it in either RDI (x86-64)
21375 // or EAX and doing an indirect call. The return value will then
21376 // be in the normal return register.
21377 MachineFunction *F = BB->getParent();
21378 const X86InstrInfo *TII = Subtarget->getInstrInfo();
21379 DebugLoc DL = MI->getDebugLoc();
21381 assert(Subtarget->isTargetDarwin() && "Darwin only instr emitted?");
21382 assert(MI->getOperand(3).isGlobal() && "This should be a global");
21384 // Get a register mask for the lowered call.
21385 // FIXME: The 32-bit calls have non-standard calling conventions. Use a
21386 // proper register mask.
21387 const uint32_t *RegMask =
21388 Subtarget->getRegisterInfo()->getCallPreservedMask(*F, CallingConv::C);
21389 if (Subtarget->is64Bit()) {
21390 MachineInstrBuilder MIB = BuildMI(*BB, MI, DL,
21391 TII->get(X86::MOV64rm), X86::RDI)
21393 .addImm(0).addReg(0)
21394 .addGlobalAddress(MI->getOperand(3).getGlobal(), 0,
21395 MI->getOperand(3).getTargetFlags())
21397 MIB = BuildMI(*BB, MI, DL, TII->get(X86::CALL64m));
21398 addDirectMem(MIB, X86::RDI);
21399 MIB.addReg(X86::RAX, RegState::ImplicitDefine).addRegMask(RegMask);
21400 } else if (F->getTarget().getRelocationModel() != Reloc::PIC_) {
21401 MachineInstrBuilder MIB = BuildMI(*BB, MI, DL,
21402 TII->get(X86::MOV32rm), X86::EAX)
21404 .addImm(0).addReg(0)
21405 .addGlobalAddress(MI->getOperand(3).getGlobal(), 0,
21406 MI->getOperand(3).getTargetFlags())
21408 MIB = BuildMI(*BB, MI, DL, TII->get(X86::CALL32m));
21409 addDirectMem(MIB, X86::EAX);
21410 MIB.addReg(X86::EAX, RegState::ImplicitDefine).addRegMask(RegMask);
21412 MachineInstrBuilder MIB = BuildMI(*BB, MI, DL,
21413 TII->get(X86::MOV32rm), X86::EAX)
21414 .addReg(TII->getGlobalBaseReg(F))
21415 .addImm(0).addReg(0)
21416 .addGlobalAddress(MI->getOperand(3).getGlobal(), 0,
21417 MI->getOperand(3).getTargetFlags())
21419 MIB = BuildMI(*BB, MI, DL, TII->get(X86::CALL32m));
21420 addDirectMem(MIB, X86::EAX);
21421 MIB.addReg(X86::EAX, RegState::ImplicitDefine).addRegMask(RegMask);
21424 MI->eraseFromParent(); // The pseudo instruction is gone now.
21428 MachineBasicBlock *
21429 X86TargetLowering::emitEHSjLjSetJmp(MachineInstr *MI,
21430 MachineBasicBlock *MBB) const {
21431 DebugLoc DL = MI->getDebugLoc();
21432 MachineFunction *MF = MBB->getParent();
21433 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
21434 MachineRegisterInfo &MRI = MF->getRegInfo();
21436 const BasicBlock *BB = MBB->getBasicBlock();
21437 MachineFunction::iterator I = ++MBB->getIterator();
21439 // Memory Reference
21440 MachineInstr::mmo_iterator MMOBegin = MI->memoperands_begin();
21441 MachineInstr::mmo_iterator MMOEnd = MI->memoperands_end();
21444 unsigned MemOpndSlot = 0;
21446 unsigned CurOp = 0;
21448 DstReg = MI->getOperand(CurOp++).getReg();
21449 const TargetRegisterClass *RC = MRI.getRegClass(DstReg);
21450 assert(RC->hasType(MVT::i32) && "Invalid destination!");
21451 unsigned mainDstReg = MRI.createVirtualRegister(RC);
21452 unsigned restoreDstReg = MRI.createVirtualRegister(RC);
21454 MemOpndSlot = CurOp;
21456 MVT PVT = getPointerTy(MF->getDataLayout());
21457 assert((PVT == MVT::i64 || PVT == MVT::i32) &&
21458 "Invalid Pointer Size!");
21460 // For v = setjmp(buf), we generate
21463 // buf[LabelOffset] = restoreMBB <-- takes address of restoreMBB
21464 // SjLjSetup restoreMBB
21470 // v = phi(main, restore)
21473 // if base pointer being used, load it from frame
21476 MachineBasicBlock *thisMBB = MBB;
21477 MachineBasicBlock *mainMBB = MF->CreateMachineBasicBlock(BB);
21478 MachineBasicBlock *sinkMBB = MF->CreateMachineBasicBlock(BB);
21479 MachineBasicBlock *restoreMBB = MF->CreateMachineBasicBlock(BB);
21480 MF->insert(I, mainMBB);
21481 MF->insert(I, sinkMBB);
21482 MF->push_back(restoreMBB);
21483 restoreMBB->setHasAddressTaken();
21485 MachineInstrBuilder MIB;
21487 // Transfer the remainder of BB and its successor edges to sinkMBB.
21488 sinkMBB->splice(sinkMBB->begin(), MBB,
21489 std::next(MachineBasicBlock::iterator(MI)), MBB->end());
21490 sinkMBB->transferSuccessorsAndUpdatePHIs(MBB);
21493 unsigned PtrStoreOpc = 0;
21494 unsigned LabelReg = 0;
21495 const int64_t LabelOffset = 1 * PVT.getStoreSize();
21496 Reloc::Model RM = MF->getTarget().getRelocationModel();
21497 bool UseImmLabel = (MF->getTarget().getCodeModel() == CodeModel::Small) &&
21498 (RM == Reloc::Static || RM == Reloc::DynamicNoPIC);
21500 // Prepare IP either in reg or imm.
21501 if (!UseImmLabel) {
21502 PtrStoreOpc = (PVT == MVT::i64) ? X86::MOV64mr : X86::MOV32mr;
21503 const TargetRegisterClass *PtrRC = getRegClassFor(PVT);
21504 LabelReg = MRI.createVirtualRegister(PtrRC);
21505 if (Subtarget->is64Bit()) {
21506 MIB = BuildMI(*thisMBB, MI, DL, TII->get(X86::LEA64r), LabelReg)
21510 .addMBB(restoreMBB)
21513 const X86InstrInfo *XII = static_cast<const X86InstrInfo*>(TII);
21514 MIB = BuildMI(*thisMBB, MI, DL, TII->get(X86::LEA32r), LabelReg)
21515 .addReg(XII->getGlobalBaseReg(MF))
21518 .addMBB(restoreMBB, Subtarget->ClassifyBlockAddressReference())
21522 PtrStoreOpc = (PVT == MVT::i64) ? X86::MOV64mi32 : X86::MOV32mi;
21524 MIB = BuildMI(*thisMBB, MI, DL, TII->get(PtrStoreOpc));
21525 for (unsigned i = 0; i < X86::AddrNumOperands; ++i) {
21526 if (i == X86::AddrDisp)
21527 MIB.addDisp(MI->getOperand(MemOpndSlot + i), LabelOffset);
21529 MIB.addOperand(MI->getOperand(MemOpndSlot + i));
21532 MIB.addReg(LabelReg);
21534 MIB.addMBB(restoreMBB);
21535 MIB.setMemRefs(MMOBegin, MMOEnd);
21537 MIB = BuildMI(*thisMBB, MI, DL, TII->get(X86::EH_SjLj_Setup))
21538 .addMBB(restoreMBB);
21540 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
21541 MIB.addRegMask(RegInfo->getNoPreservedMask());
21542 thisMBB->addSuccessor(mainMBB);
21543 thisMBB->addSuccessor(restoreMBB);
21547 BuildMI(mainMBB, DL, TII->get(X86::MOV32r0), mainDstReg);
21548 mainMBB->addSuccessor(sinkMBB);
21551 BuildMI(*sinkMBB, sinkMBB->begin(), DL,
21552 TII->get(X86::PHI), DstReg)
21553 .addReg(mainDstReg).addMBB(mainMBB)
21554 .addReg(restoreDstReg).addMBB(restoreMBB);
21557 if (RegInfo->hasBasePointer(*MF)) {
21558 const bool Uses64BitFramePtr =
21559 Subtarget->isTarget64BitLP64() || Subtarget->isTargetNaCl64();
21560 X86MachineFunctionInfo *X86FI = MF->getInfo<X86MachineFunctionInfo>();
21561 X86FI->setRestoreBasePointer(MF);
21562 unsigned FramePtr = RegInfo->getFrameRegister(*MF);
21563 unsigned BasePtr = RegInfo->getBaseRegister();
21564 unsigned Opm = Uses64BitFramePtr ? X86::MOV64rm : X86::MOV32rm;
21565 addRegOffset(BuildMI(restoreMBB, DL, TII->get(Opm), BasePtr),
21566 FramePtr, true, X86FI->getRestoreBasePointerOffset())
21567 .setMIFlag(MachineInstr::FrameSetup);
21569 BuildMI(restoreMBB, DL, TII->get(X86::MOV32ri), restoreDstReg).addImm(1);
21570 BuildMI(restoreMBB, DL, TII->get(X86::JMP_1)).addMBB(sinkMBB);
21571 restoreMBB->addSuccessor(sinkMBB);
21573 MI->eraseFromParent();
21577 MachineBasicBlock *
21578 X86TargetLowering::emitEHSjLjLongJmp(MachineInstr *MI,
21579 MachineBasicBlock *MBB) const {
21580 DebugLoc DL = MI->getDebugLoc();
21581 MachineFunction *MF = MBB->getParent();
21582 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
21583 MachineRegisterInfo &MRI = MF->getRegInfo();
21585 // Memory Reference
21586 MachineInstr::mmo_iterator MMOBegin = MI->memoperands_begin();
21587 MachineInstr::mmo_iterator MMOEnd = MI->memoperands_end();
21589 MVT PVT = getPointerTy(MF->getDataLayout());
21590 assert((PVT == MVT::i64 || PVT == MVT::i32) &&
21591 "Invalid Pointer Size!");
21593 const TargetRegisterClass *RC =
21594 (PVT == MVT::i64) ? &X86::GR64RegClass : &X86::GR32RegClass;
21595 unsigned Tmp = MRI.createVirtualRegister(RC);
21596 // Since FP is only updated here but NOT referenced, it's treated as GPR.
21597 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
21598 unsigned FP = (PVT == MVT::i64) ? X86::RBP : X86::EBP;
21599 unsigned SP = RegInfo->getStackRegister();
21601 MachineInstrBuilder MIB;
21603 const int64_t LabelOffset = 1 * PVT.getStoreSize();
21604 const int64_t SPOffset = 2 * PVT.getStoreSize();
21606 unsigned PtrLoadOpc = (PVT == MVT::i64) ? X86::MOV64rm : X86::MOV32rm;
21607 unsigned IJmpOpc = (PVT == MVT::i64) ? X86::JMP64r : X86::JMP32r;
21610 MIB = BuildMI(*MBB, MI, DL, TII->get(PtrLoadOpc), FP);
21611 for (unsigned i = 0; i < X86::AddrNumOperands; ++i)
21612 MIB.addOperand(MI->getOperand(i));
21613 MIB.setMemRefs(MMOBegin, MMOEnd);
21615 MIB = BuildMI(*MBB, MI, DL, TII->get(PtrLoadOpc), Tmp);
21616 for (unsigned i = 0; i < X86::AddrNumOperands; ++i) {
21617 if (i == X86::AddrDisp)
21618 MIB.addDisp(MI->getOperand(i), LabelOffset);
21620 MIB.addOperand(MI->getOperand(i));
21622 MIB.setMemRefs(MMOBegin, MMOEnd);
21624 MIB = BuildMI(*MBB, MI, DL, TII->get(PtrLoadOpc), SP);
21625 for (unsigned i = 0; i < X86::AddrNumOperands; ++i) {
21626 if (i == X86::AddrDisp)
21627 MIB.addDisp(MI->getOperand(i), SPOffset);
21629 MIB.addOperand(MI->getOperand(i));
21631 MIB.setMemRefs(MMOBegin, MMOEnd);
21633 BuildMI(*MBB, MI, DL, TII->get(IJmpOpc)).addReg(Tmp);
21635 MI->eraseFromParent();
21639 // Replace 213-type (isel default) FMA3 instructions with 231-type for
21640 // accumulator loops. Writing back to the accumulator allows the coalescer
21641 // to remove extra copies in the loop.
21642 // FIXME: Do this on AVX512. We don't support 231 variants yet (PR23937).
21643 MachineBasicBlock *
21644 X86TargetLowering::emitFMA3Instr(MachineInstr *MI,
21645 MachineBasicBlock *MBB) const {
21646 MachineOperand &AddendOp = MI->getOperand(3);
21648 // Bail out early if the addend isn't a register - we can't switch these.
21649 if (!AddendOp.isReg())
21652 MachineFunction &MF = *MBB->getParent();
21653 MachineRegisterInfo &MRI = MF.getRegInfo();
21655 // Check whether the addend is defined by a PHI:
21656 assert(MRI.hasOneDef(AddendOp.getReg()) && "Multiple defs in SSA?");
21657 MachineInstr &AddendDef = *MRI.def_instr_begin(AddendOp.getReg());
21658 if (!AddendDef.isPHI())
21661 // Look for the following pattern:
21663 // %addend = phi [%entry, 0], [%loop, %result]
21665 // %result<tied1> = FMA213 %m2<tied0>, %m1, %addend
21669 // %addend = phi [%entry, 0], [%loop, %result]
21671 // %result<tied1> = FMA231 %addend<tied0>, %m1, %m2
21673 for (unsigned i = 1, e = AddendDef.getNumOperands(); i < e; i += 2) {
21674 assert(AddendDef.getOperand(i).isReg());
21675 MachineOperand PHISrcOp = AddendDef.getOperand(i);
21676 MachineInstr &PHISrcInst = *MRI.def_instr_begin(PHISrcOp.getReg());
21677 if (&PHISrcInst == MI) {
21678 // Found a matching instruction.
21679 unsigned NewFMAOpc = 0;
21680 switch (MI->getOpcode()) {
21681 case X86::VFMADDPDr213r: NewFMAOpc = X86::VFMADDPDr231r; break;
21682 case X86::VFMADDPSr213r: NewFMAOpc = X86::VFMADDPSr231r; break;
21683 case X86::VFMADDSDr213r: NewFMAOpc = X86::VFMADDSDr231r; break;
21684 case X86::VFMADDSSr213r: NewFMAOpc = X86::VFMADDSSr231r; break;
21685 case X86::VFMSUBPDr213r: NewFMAOpc = X86::VFMSUBPDr231r; break;
21686 case X86::VFMSUBPSr213r: NewFMAOpc = X86::VFMSUBPSr231r; break;
21687 case X86::VFMSUBSDr213r: NewFMAOpc = X86::VFMSUBSDr231r; break;
21688 case X86::VFMSUBSSr213r: NewFMAOpc = X86::VFMSUBSSr231r; break;
21689 case X86::VFNMADDPDr213r: NewFMAOpc = X86::VFNMADDPDr231r; break;
21690 case X86::VFNMADDPSr213r: NewFMAOpc = X86::VFNMADDPSr231r; break;
21691 case X86::VFNMADDSDr213r: NewFMAOpc = X86::VFNMADDSDr231r; break;
21692 case X86::VFNMADDSSr213r: NewFMAOpc = X86::VFNMADDSSr231r; break;
21693 case X86::VFNMSUBPDr213r: NewFMAOpc = X86::VFNMSUBPDr231r; break;
21694 case X86::VFNMSUBPSr213r: NewFMAOpc = X86::VFNMSUBPSr231r; break;
21695 case X86::VFNMSUBSDr213r: NewFMAOpc = X86::VFNMSUBSDr231r; break;
21696 case X86::VFNMSUBSSr213r: NewFMAOpc = X86::VFNMSUBSSr231r; break;
21697 case X86::VFMADDSUBPDr213r: NewFMAOpc = X86::VFMADDSUBPDr231r; break;
21698 case X86::VFMADDSUBPSr213r: NewFMAOpc = X86::VFMADDSUBPSr231r; break;
21699 case X86::VFMSUBADDPDr213r: NewFMAOpc = X86::VFMSUBADDPDr231r; break;
21700 case X86::VFMSUBADDPSr213r: NewFMAOpc = X86::VFMSUBADDPSr231r; break;
21702 case X86::VFMADDPDr213rY: NewFMAOpc = X86::VFMADDPDr231rY; break;
21703 case X86::VFMADDPSr213rY: NewFMAOpc = X86::VFMADDPSr231rY; break;
21704 case X86::VFMSUBPDr213rY: NewFMAOpc = X86::VFMSUBPDr231rY; break;
21705 case X86::VFMSUBPSr213rY: NewFMAOpc = X86::VFMSUBPSr231rY; break;
21706 case X86::VFNMADDPDr213rY: NewFMAOpc = X86::VFNMADDPDr231rY; break;
21707 case X86::VFNMADDPSr213rY: NewFMAOpc = X86::VFNMADDPSr231rY; break;
21708 case X86::VFNMSUBPDr213rY: NewFMAOpc = X86::VFNMSUBPDr231rY; break;
21709 case X86::VFNMSUBPSr213rY: NewFMAOpc = X86::VFNMSUBPSr231rY; break;
21710 case X86::VFMADDSUBPDr213rY: NewFMAOpc = X86::VFMADDSUBPDr231rY; break;
21711 case X86::VFMADDSUBPSr213rY: NewFMAOpc = X86::VFMADDSUBPSr231rY; break;
21712 case X86::VFMSUBADDPDr213rY: NewFMAOpc = X86::VFMSUBADDPDr231rY; break;
21713 case X86::VFMSUBADDPSr213rY: NewFMAOpc = X86::VFMSUBADDPSr231rY; break;
21714 default: llvm_unreachable("Unrecognized FMA variant.");
21717 const TargetInstrInfo &TII = *Subtarget->getInstrInfo();
21718 MachineInstrBuilder MIB =
21719 BuildMI(MF, MI->getDebugLoc(), TII.get(NewFMAOpc))
21720 .addOperand(MI->getOperand(0))
21721 .addOperand(MI->getOperand(3))
21722 .addOperand(MI->getOperand(2))
21723 .addOperand(MI->getOperand(1));
21724 MBB->insert(MachineBasicBlock::iterator(MI), MIB);
21725 MI->eraseFromParent();
21732 MachineBasicBlock *
21733 X86TargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
21734 MachineBasicBlock *BB) const {
21735 switch (MI->getOpcode()) {
21736 default: llvm_unreachable("Unexpected instr type to insert");
21737 case X86::TAILJMPd64:
21738 case X86::TAILJMPr64:
21739 case X86::TAILJMPm64:
21740 case X86::TAILJMPd64_REX:
21741 case X86::TAILJMPr64_REX:
21742 case X86::TAILJMPm64_REX:
21743 llvm_unreachable("TAILJMP64 would not be touched here.");
21744 case X86::TCRETURNdi64:
21745 case X86::TCRETURNri64:
21746 case X86::TCRETURNmi64:
21748 case X86::WIN_ALLOCA:
21749 return EmitLoweredWinAlloca(MI, BB);
21750 case X86::SEG_ALLOCA_32:
21751 case X86::SEG_ALLOCA_64:
21752 return EmitLoweredSegAlloca(MI, BB);
21753 case X86::TLSCall_32:
21754 case X86::TLSCall_64:
21755 return EmitLoweredTLSCall(MI, BB);
21756 case X86::CMOV_FR32:
21757 case X86::CMOV_FR64:
21758 case X86::CMOV_GR8:
21759 case X86::CMOV_GR16:
21760 case X86::CMOV_GR32:
21761 case X86::CMOV_RFP32:
21762 case X86::CMOV_RFP64:
21763 case X86::CMOV_RFP80:
21764 case X86::CMOV_V2F64:
21765 case X86::CMOV_V2I64:
21766 case X86::CMOV_V4F32:
21767 case X86::CMOV_V4F64:
21768 case X86::CMOV_V4I64:
21769 case X86::CMOV_V16F32:
21770 case X86::CMOV_V8F32:
21771 case X86::CMOV_V8F64:
21772 case X86::CMOV_V8I64:
21773 case X86::CMOV_V8I1:
21774 case X86::CMOV_V16I1:
21775 case X86::CMOV_V32I1:
21776 case X86::CMOV_V64I1:
21777 return EmitLoweredSelect(MI, BB);
21779 case X86::RELEASE_FADD32mr:
21780 case X86::RELEASE_FADD64mr:
21781 return EmitLoweredAtomicFP(MI, BB);
21783 case X86::FP32_TO_INT16_IN_MEM:
21784 case X86::FP32_TO_INT32_IN_MEM:
21785 case X86::FP32_TO_INT64_IN_MEM:
21786 case X86::FP64_TO_INT16_IN_MEM:
21787 case X86::FP64_TO_INT32_IN_MEM:
21788 case X86::FP64_TO_INT64_IN_MEM:
21789 case X86::FP80_TO_INT16_IN_MEM:
21790 case X86::FP80_TO_INT32_IN_MEM:
21791 case X86::FP80_TO_INT64_IN_MEM: {
21792 MachineFunction *F = BB->getParent();
21793 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
21794 DebugLoc DL = MI->getDebugLoc();
21796 // Change the floating point control register to use "round towards zero"
21797 // mode when truncating to an integer value.
21798 int CWFrameIdx = F->getFrameInfo()->CreateStackObject(2, 2, false);
21799 addFrameReference(BuildMI(*BB, MI, DL,
21800 TII->get(X86::FNSTCW16m)), CWFrameIdx);
21802 // Load the old value of the high byte of the control word...
21804 F->getRegInfo().createVirtualRegister(&X86::GR16RegClass);
21805 addFrameReference(BuildMI(*BB, MI, DL, TII->get(X86::MOV16rm), OldCW),
21808 // Set the high part to be round to zero...
21809 addFrameReference(BuildMI(*BB, MI, DL, TII->get(X86::MOV16mi)), CWFrameIdx)
21812 // Reload the modified control word now...
21813 addFrameReference(BuildMI(*BB, MI, DL,
21814 TII->get(X86::FLDCW16m)), CWFrameIdx);
21816 // Restore the memory image of control word to original value
21817 addFrameReference(BuildMI(*BB, MI, DL, TII->get(X86::MOV16mr)), CWFrameIdx)
21820 // Get the X86 opcode to use.
21822 switch (MI->getOpcode()) {
21823 default: llvm_unreachable("illegal opcode!");
21824 case X86::FP32_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m32; break;
21825 case X86::FP32_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m32; break;
21826 case X86::FP32_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m32; break;
21827 case X86::FP64_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m64; break;
21828 case X86::FP64_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m64; break;
21829 case X86::FP64_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m64; break;
21830 case X86::FP80_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m80; break;
21831 case X86::FP80_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m80; break;
21832 case X86::FP80_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m80; break;
21836 MachineOperand &Op = MI->getOperand(0);
21838 AM.BaseType = X86AddressMode::RegBase;
21839 AM.Base.Reg = Op.getReg();
21841 AM.BaseType = X86AddressMode::FrameIndexBase;
21842 AM.Base.FrameIndex = Op.getIndex();
21844 Op = MI->getOperand(1);
21846 AM.Scale = Op.getImm();
21847 Op = MI->getOperand(2);
21849 AM.IndexReg = Op.getImm();
21850 Op = MI->getOperand(3);
21851 if (Op.isGlobal()) {
21852 AM.GV = Op.getGlobal();
21854 AM.Disp = Op.getImm();
21856 addFullAddress(BuildMI(*BB, MI, DL, TII->get(Opc)), AM)
21857 .addReg(MI->getOperand(X86::AddrNumOperands).getReg());
21859 // Reload the original control word now.
21860 addFrameReference(BuildMI(*BB, MI, DL,
21861 TII->get(X86::FLDCW16m)), CWFrameIdx);
21863 MI->eraseFromParent(); // The pseudo instruction is gone now.
21866 // String/text processing lowering.
21867 case X86::PCMPISTRM128REG:
21868 case X86::VPCMPISTRM128REG:
21869 case X86::PCMPISTRM128MEM:
21870 case X86::VPCMPISTRM128MEM:
21871 case X86::PCMPESTRM128REG:
21872 case X86::VPCMPESTRM128REG:
21873 case X86::PCMPESTRM128MEM:
21874 case X86::VPCMPESTRM128MEM:
21875 assert(Subtarget->hasSSE42() &&
21876 "Target must have SSE4.2 or AVX features enabled");
21877 return EmitPCMPSTRM(MI, BB, Subtarget->getInstrInfo());
21879 // String/text processing lowering.
21880 case X86::PCMPISTRIREG:
21881 case X86::VPCMPISTRIREG:
21882 case X86::PCMPISTRIMEM:
21883 case X86::VPCMPISTRIMEM:
21884 case X86::PCMPESTRIREG:
21885 case X86::VPCMPESTRIREG:
21886 case X86::PCMPESTRIMEM:
21887 case X86::VPCMPESTRIMEM:
21888 assert(Subtarget->hasSSE42() &&
21889 "Target must have SSE4.2 or AVX features enabled");
21890 return EmitPCMPSTRI(MI, BB, Subtarget->getInstrInfo());
21892 // Thread synchronization.
21894 return EmitMonitor(MI, BB, Subtarget);
21898 return EmitXBegin(MI, BB, Subtarget->getInstrInfo());
21900 case X86::VASTART_SAVE_XMM_REGS:
21901 return EmitVAStartSaveXMMRegsWithCustomInserter(MI, BB);
21903 case X86::VAARG_64:
21904 return EmitVAARG64WithCustomInserter(MI, BB);
21906 case X86::EH_SjLj_SetJmp32:
21907 case X86::EH_SjLj_SetJmp64:
21908 return emitEHSjLjSetJmp(MI, BB);
21910 case X86::EH_SjLj_LongJmp32:
21911 case X86::EH_SjLj_LongJmp64:
21912 return emitEHSjLjLongJmp(MI, BB);
21914 case TargetOpcode::STATEPOINT:
21915 // As an implementation detail, STATEPOINT shares the STACKMAP format at
21916 // this point in the process. We diverge later.
21917 return emitPatchPoint(MI, BB);
21919 case TargetOpcode::STACKMAP:
21920 case TargetOpcode::PATCHPOINT:
21921 return emitPatchPoint(MI, BB);
21923 case X86::VFMADDPDr213r:
21924 case X86::VFMADDPSr213r:
21925 case X86::VFMADDSDr213r:
21926 case X86::VFMADDSSr213r:
21927 case X86::VFMSUBPDr213r:
21928 case X86::VFMSUBPSr213r:
21929 case X86::VFMSUBSDr213r:
21930 case X86::VFMSUBSSr213r:
21931 case X86::VFNMADDPDr213r:
21932 case X86::VFNMADDPSr213r:
21933 case X86::VFNMADDSDr213r:
21934 case X86::VFNMADDSSr213r:
21935 case X86::VFNMSUBPDr213r:
21936 case X86::VFNMSUBPSr213r:
21937 case X86::VFNMSUBSDr213r:
21938 case X86::VFNMSUBSSr213r:
21939 case X86::VFMADDSUBPDr213r:
21940 case X86::VFMADDSUBPSr213r:
21941 case X86::VFMSUBADDPDr213r:
21942 case X86::VFMSUBADDPSr213r:
21943 case X86::VFMADDPDr213rY:
21944 case X86::VFMADDPSr213rY:
21945 case X86::VFMSUBPDr213rY:
21946 case X86::VFMSUBPSr213rY:
21947 case X86::VFNMADDPDr213rY:
21948 case X86::VFNMADDPSr213rY:
21949 case X86::VFNMSUBPDr213rY:
21950 case X86::VFNMSUBPSr213rY:
21951 case X86::VFMADDSUBPDr213rY:
21952 case X86::VFMADDSUBPSr213rY:
21953 case X86::VFMSUBADDPDr213rY:
21954 case X86::VFMSUBADDPSr213rY:
21955 return emitFMA3Instr(MI, BB);
21959 //===----------------------------------------------------------------------===//
21960 // X86 Optimization Hooks
21961 //===----------------------------------------------------------------------===//
21963 void X86TargetLowering::computeKnownBitsForTargetNode(const SDValue Op,
21966 const SelectionDAG &DAG,
21967 unsigned Depth) const {
21968 unsigned BitWidth = KnownZero.getBitWidth();
21969 unsigned Opc = Op.getOpcode();
21970 assert((Opc >= ISD::BUILTIN_OP_END ||
21971 Opc == ISD::INTRINSIC_WO_CHAIN ||
21972 Opc == ISD::INTRINSIC_W_CHAIN ||
21973 Opc == ISD::INTRINSIC_VOID) &&
21974 "Should use MaskedValueIsZero if you don't know whether Op"
21975 " is a target node!");
21977 KnownZero = KnownOne = APInt(BitWidth, 0); // Don't know anything.
21991 // These nodes' second result is a boolean.
21992 if (Op.getResNo() == 0)
21995 case X86ISD::SETCC:
21996 KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - 1);
21998 case ISD::INTRINSIC_WO_CHAIN: {
21999 unsigned IntId = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
22000 unsigned NumLoBits = 0;
22003 case Intrinsic::x86_sse_movmsk_ps:
22004 case Intrinsic::x86_avx_movmsk_ps_256:
22005 case Intrinsic::x86_sse2_movmsk_pd:
22006 case Intrinsic::x86_avx_movmsk_pd_256:
22007 case Intrinsic::x86_mmx_pmovmskb:
22008 case Intrinsic::x86_sse2_pmovmskb_128:
22009 case Intrinsic::x86_avx2_pmovmskb: {
22010 // High bits of movmskp{s|d}, pmovmskb are known zero.
22012 default: llvm_unreachable("Impossible intrinsic"); // Can't reach here.
22013 case Intrinsic::x86_sse_movmsk_ps: NumLoBits = 4; break;
22014 case Intrinsic::x86_avx_movmsk_ps_256: NumLoBits = 8; break;
22015 case Intrinsic::x86_sse2_movmsk_pd: NumLoBits = 2; break;
22016 case Intrinsic::x86_avx_movmsk_pd_256: NumLoBits = 4; break;
22017 case Intrinsic::x86_mmx_pmovmskb: NumLoBits = 8; break;
22018 case Intrinsic::x86_sse2_pmovmskb_128: NumLoBits = 16; break;
22019 case Intrinsic::x86_avx2_pmovmskb: NumLoBits = 32; break;
22021 KnownZero = APInt::getHighBitsSet(BitWidth, BitWidth - NumLoBits);
22030 unsigned X86TargetLowering::ComputeNumSignBitsForTargetNode(
22032 const SelectionDAG &,
22033 unsigned Depth) const {
22034 // SETCC_CARRY sets the dest to ~0 for true or 0 for false.
22035 if (Op.getOpcode() == X86ISD::SETCC_CARRY)
22036 return Op.getValueType().getScalarType().getSizeInBits();
22042 /// isGAPlusOffset - Returns true (and the GlobalValue and the offset) if the
22043 /// node is a GlobalAddress + offset.
22044 bool X86TargetLowering::isGAPlusOffset(SDNode *N,
22045 const GlobalValue* &GA,
22046 int64_t &Offset) const {
22047 if (N->getOpcode() == X86ISD::Wrapper) {
22048 if (isa<GlobalAddressSDNode>(N->getOperand(0))) {
22049 GA = cast<GlobalAddressSDNode>(N->getOperand(0))->getGlobal();
22050 Offset = cast<GlobalAddressSDNode>(N->getOperand(0))->getOffset();
22054 return TargetLowering::isGAPlusOffset(N, GA, Offset);
22057 /// isShuffleHigh128VectorInsertLow - Checks whether the shuffle node is the
22058 /// same as extracting the high 128-bit part of 256-bit vector and then
22059 /// inserting the result into the low part of a new 256-bit vector
22060 static bool isShuffleHigh128VectorInsertLow(ShuffleVectorSDNode *SVOp) {
22061 EVT VT = SVOp->getValueType(0);
22062 unsigned NumElems = VT.getVectorNumElements();
22064 // vector_shuffle <4, 5, 6, 7, u, u, u, u> or <2, 3, u, u>
22065 for (unsigned i = 0, j = NumElems/2; i != NumElems/2; ++i, ++j)
22066 if (!isUndefOrEqual(SVOp->getMaskElt(i), j) ||
22067 SVOp->getMaskElt(j) >= 0)
22073 /// isShuffleLow128VectorInsertHigh - Checks whether the shuffle node is the
22074 /// same as extracting the low 128-bit part of 256-bit vector and then
22075 /// inserting the result into the high part of a new 256-bit vector
22076 static bool isShuffleLow128VectorInsertHigh(ShuffleVectorSDNode *SVOp) {
22077 EVT VT = SVOp->getValueType(0);
22078 unsigned NumElems = VT.getVectorNumElements();
22080 // vector_shuffle <u, u, u, u, 0, 1, 2, 3> or <u, u, 0, 1>
22081 for (unsigned i = NumElems/2, j = 0; i != NumElems; ++i, ++j)
22082 if (!isUndefOrEqual(SVOp->getMaskElt(i), j) ||
22083 SVOp->getMaskElt(j) >= 0)
22089 /// PerformShuffleCombine256 - Performs shuffle combines for 256-bit vectors.
22090 static SDValue PerformShuffleCombine256(SDNode *N, SelectionDAG &DAG,
22091 TargetLowering::DAGCombinerInfo &DCI,
22092 const X86Subtarget* Subtarget) {
22094 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
22095 SDValue V1 = SVOp->getOperand(0);
22096 SDValue V2 = SVOp->getOperand(1);
22097 EVT VT = SVOp->getValueType(0);
22098 unsigned NumElems = VT.getVectorNumElements();
22100 if (V1.getOpcode() == ISD::CONCAT_VECTORS &&
22101 V2.getOpcode() == ISD::CONCAT_VECTORS) {
22105 // V UNDEF BUILD_VECTOR UNDEF
22107 // CONCAT_VECTOR CONCAT_VECTOR
22110 // RESULT: V + zero extended
22112 if (V2.getOperand(0).getOpcode() != ISD::BUILD_VECTOR ||
22113 V2.getOperand(1).getOpcode() != ISD::UNDEF ||
22114 V1.getOperand(1).getOpcode() != ISD::UNDEF)
22117 if (!ISD::isBuildVectorAllZeros(V2.getOperand(0).getNode()))
22120 // To match the shuffle mask, the first half of the mask should
22121 // be exactly the first vector, and all the rest a splat with the
22122 // first element of the second one.
22123 for (unsigned i = 0; i != NumElems/2; ++i)
22124 if (!isUndefOrEqual(SVOp->getMaskElt(i), i) ||
22125 !isUndefOrEqual(SVOp->getMaskElt(i+NumElems/2), NumElems))
22128 // If V1 is coming from a vector load then just fold to a VZEXT_LOAD.
22129 if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(V1.getOperand(0))) {
22130 if (Ld->hasNUsesOfValue(1, 0)) {
22131 SDVTList Tys = DAG.getVTList(MVT::v4i64, MVT::Other);
22132 SDValue Ops[] = { Ld->getChain(), Ld->getBasePtr() };
22134 DAG.getMemIntrinsicNode(X86ISD::VZEXT_LOAD, dl, Tys, Ops,
22136 Ld->getPointerInfo(),
22137 Ld->getAlignment(),
22138 false/*isVolatile*/, true/*ReadMem*/,
22139 false/*WriteMem*/);
22141 // Make sure the newly-created LOAD is in the same position as Ld in
22142 // terms of dependency. We create a TokenFactor for Ld and ResNode,
22143 // and update uses of Ld's output chain to use the TokenFactor.
22144 if (Ld->hasAnyUseOfValue(1)) {
22145 SDValue NewChain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
22146 SDValue(Ld, 1), SDValue(ResNode.getNode(), 1));
22147 DAG.ReplaceAllUsesOfValueWith(SDValue(Ld, 1), NewChain);
22148 DAG.UpdateNodeOperands(NewChain.getNode(), SDValue(Ld, 1),
22149 SDValue(ResNode.getNode(), 1));
22152 return DAG.getBitcast(VT, ResNode);
22156 // Emit a zeroed vector and insert the desired subvector on its
22158 SDValue Zeros = getZeroVector(VT, Subtarget, DAG, dl);
22159 SDValue InsV = Insert128BitVector(Zeros, V1.getOperand(0), 0, DAG, dl);
22160 return DCI.CombineTo(N, InsV);
22163 //===--------------------------------------------------------------------===//
22164 // Combine some shuffles into subvector extracts and inserts:
22167 // vector_shuffle <4, 5, 6, 7, u, u, u, u> or <2, 3, u, u>
22168 if (isShuffleHigh128VectorInsertLow(SVOp)) {
22169 SDValue V = Extract128BitVector(V1, NumElems/2, DAG, dl);
22170 SDValue InsV = Insert128BitVector(DAG.getUNDEF(VT), V, 0, DAG, dl);
22171 return DCI.CombineTo(N, InsV);
22174 // vector_shuffle <u, u, u, u, 0, 1, 2, 3> or <u, u, 0, 1>
22175 if (isShuffleLow128VectorInsertHigh(SVOp)) {
22176 SDValue V = Extract128BitVector(V1, 0, DAG, dl);
22177 SDValue InsV = Insert128BitVector(DAG.getUNDEF(VT), V, NumElems/2, DAG, dl);
22178 return DCI.CombineTo(N, InsV);
22184 /// \brief Combine an arbitrary chain of shuffles into a single instruction if
22187 /// This is the leaf of the recursive combinine below. When we have found some
22188 /// chain of single-use x86 shuffle instructions and accumulated the combined
22189 /// shuffle mask represented by them, this will try to pattern match that mask
22190 /// into either a single instruction if there is a special purpose instruction
22191 /// for this operation, or into a PSHUFB instruction which is a fully general
22192 /// instruction but should only be used to replace chains over a certain depth.
22193 static bool combineX86ShuffleChain(SDValue Op, SDValue Root, ArrayRef<int> Mask,
22194 int Depth, bool HasPSHUFB, SelectionDAG &DAG,
22195 TargetLowering::DAGCombinerInfo &DCI,
22196 const X86Subtarget *Subtarget) {
22197 assert(!Mask.empty() && "Cannot combine an empty shuffle mask!");
22199 // Find the operand that enters the chain. Note that multiple uses are OK
22200 // here, we're not going to remove the operand we find.
22201 SDValue Input = Op.getOperand(0);
22202 while (Input.getOpcode() == ISD::BITCAST)
22203 Input = Input.getOperand(0);
22205 MVT VT = Input.getSimpleValueType();
22206 MVT RootVT = Root.getSimpleValueType();
22209 if (Mask.size() == 1) {
22210 int Index = Mask[0];
22211 assert((Index >= 0 || Index == SM_SentinelUndef ||
22212 Index == SM_SentinelZero) &&
22213 "Invalid shuffle index found!");
22215 // We may end up with an accumulated mask of size 1 as a result of
22216 // widening of shuffle operands (see function canWidenShuffleElements).
22217 // If the only shuffle index is equal to SM_SentinelZero then propagate
22218 // a zero vector. Otherwise, the combine shuffle mask is a no-op shuffle
22219 // mask, and therefore the entire chain of shuffles can be folded away.
22220 if (Index == SM_SentinelZero)
22221 DCI.CombineTo(Root.getNode(), getZeroVector(RootVT, Subtarget, DAG, DL));
22223 DCI.CombineTo(Root.getNode(), DAG.getBitcast(RootVT, Input),
22228 // Use the float domain if the operand type is a floating point type.
22229 bool FloatDomain = VT.isFloatingPoint();
22231 // For floating point shuffles, we don't have free copies in the shuffle
22232 // instructions or the ability to load as part of the instruction, so
22233 // canonicalize their shuffles to UNPCK or MOV variants.
22235 // Note that even with AVX we prefer the PSHUFD form of shuffle for integer
22236 // vectors because it can have a load folded into it that UNPCK cannot. This
22237 // doesn't preclude something switching to the shorter encoding post-RA.
22239 // FIXME: Should teach these routines about AVX vector widths.
22240 if (FloatDomain && VT.getSizeInBits() == 128) {
22241 if (Mask.equals({0, 0}) || Mask.equals({1, 1})) {
22242 bool Lo = Mask.equals({0, 0});
22245 // Check if we have SSE3 which will let us use MOVDDUP. That instruction
22246 // is no slower than UNPCKLPD but has the option to fold the input operand
22247 // into even an unaligned memory load.
22248 if (Lo && Subtarget->hasSSE3()) {
22249 Shuffle = X86ISD::MOVDDUP;
22250 ShuffleVT = MVT::v2f64;
22252 // We have MOVLHPS and MOVHLPS throughout SSE and they encode smaller
22253 // than the UNPCK variants.
22254 Shuffle = Lo ? X86ISD::MOVLHPS : X86ISD::MOVHLPS;
22255 ShuffleVT = MVT::v4f32;
22257 if (Depth == 1 && Root->getOpcode() == Shuffle)
22258 return false; // Nothing to do!
22259 Op = DAG.getBitcast(ShuffleVT, Input);
22260 DCI.AddToWorklist(Op.getNode());
22261 if (Shuffle == X86ISD::MOVDDUP)
22262 Op = DAG.getNode(Shuffle, DL, ShuffleVT, Op);
22264 Op = DAG.getNode(Shuffle, DL, ShuffleVT, Op, Op);
22265 DCI.AddToWorklist(Op.getNode());
22266 DCI.CombineTo(Root.getNode(), DAG.getBitcast(RootVT, Op),
22270 if (Subtarget->hasSSE3() &&
22271 (Mask.equals({0, 0, 2, 2}) || Mask.equals({1, 1, 3, 3}))) {
22272 bool Lo = Mask.equals({0, 0, 2, 2});
22273 unsigned Shuffle = Lo ? X86ISD::MOVSLDUP : X86ISD::MOVSHDUP;
22274 MVT ShuffleVT = MVT::v4f32;
22275 if (Depth == 1 && Root->getOpcode() == Shuffle)
22276 return false; // Nothing to do!
22277 Op = DAG.getBitcast(ShuffleVT, Input);
22278 DCI.AddToWorklist(Op.getNode());
22279 Op = DAG.getNode(Shuffle, DL, ShuffleVT, Op);
22280 DCI.AddToWorklist(Op.getNode());
22281 DCI.CombineTo(Root.getNode(), DAG.getBitcast(RootVT, Op),
22285 if (Mask.equals({0, 0, 1, 1}) || Mask.equals({2, 2, 3, 3})) {
22286 bool Lo = Mask.equals({0, 0, 1, 1});
22287 unsigned Shuffle = Lo ? X86ISD::UNPCKL : X86ISD::UNPCKH;
22288 MVT ShuffleVT = MVT::v4f32;
22289 if (Depth == 1 && Root->getOpcode() == Shuffle)
22290 return false; // Nothing to do!
22291 Op = DAG.getBitcast(ShuffleVT, Input);
22292 DCI.AddToWorklist(Op.getNode());
22293 Op = DAG.getNode(Shuffle, DL, ShuffleVT, Op, Op);
22294 DCI.AddToWorklist(Op.getNode());
22295 DCI.CombineTo(Root.getNode(), DAG.getBitcast(RootVT, Op),
22301 // We always canonicalize the 8 x i16 and 16 x i8 shuffles into their UNPCK
22302 // variants as none of these have single-instruction variants that are
22303 // superior to the UNPCK formulation.
22304 if (!FloatDomain && VT.getSizeInBits() == 128 &&
22305 (Mask.equals({0, 0, 1, 1, 2, 2, 3, 3}) ||
22306 Mask.equals({4, 4, 5, 5, 6, 6, 7, 7}) ||
22307 Mask.equals({0, 0, 1, 1, 2, 2, 3, 3, 4, 4, 5, 5, 6, 6, 7, 7}) ||
22309 {8, 8, 9, 9, 10, 10, 11, 11, 12, 12, 13, 13, 14, 14, 15, 15}))) {
22310 bool Lo = Mask[0] == 0;
22311 unsigned Shuffle = Lo ? X86ISD::UNPCKL : X86ISD::UNPCKH;
22312 if (Depth == 1 && Root->getOpcode() == Shuffle)
22313 return false; // Nothing to do!
22315 switch (Mask.size()) {
22317 ShuffleVT = MVT::v8i16;
22320 ShuffleVT = MVT::v16i8;
22323 llvm_unreachable("Impossible mask size!");
22325 Op = DAG.getBitcast(ShuffleVT, Input);
22326 DCI.AddToWorklist(Op.getNode());
22327 Op = DAG.getNode(Shuffle, DL, ShuffleVT, Op, Op);
22328 DCI.AddToWorklist(Op.getNode());
22329 DCI.CombineTo(Root.getNode(), DAG.getBitcast(RootVT, Op),
22334 // Don't try to re-form single instruction chains under any circumstances now
22335 // that we've done encoding canonicalization for them.
22339 // If we have 3 or more shuffle instructions or a chain involving PSHUFB, we
22340 // can replace them with a single PSHUFB instruction profitably. Intel's
22341 // manuals suggest only using PSHUFB if doing so replacing 5 instructions, but
22342 // in practice PSHUFB tends to be *very* fast so we're more aggressive.
22343 if ((Depth >= 3 || HasPSHUFB) && Subtarget->hasSSSE3()) {
22344 SmallVector<SDValue, 16> PSHUFBMask;
22345 int NumBytes = VT.getSizeInBits() / 8;
22346 int Ratio = NumBytes / Mask.size();
22347 for (int i = 0; i < NumBytes; ++i) {
22348 if (Mask[i / Ratio] == SM_SentinelUndef) {
22349 PSHUFBMask.push_back(DAG.getUNDEF(MVT::i8));
22352 int M = Mask[i / Ratio] != SM_SentinelZero
22353 ? Ratio * Mask[i / Ratio] + i % Ratio
22355 PSHUFBMask.push_back(DAG.getConstant(M, DL, MVT::i8));
22357 MVT ByteVT = MVT::getVectorVT(MVT::i8, NumBytes);
22358 Op = DAG.getBitcast(ByteVT, Input);
22359 DCI.AddToWorklist(Op.getNode());
22360 SDValue PSHUFBMaskOp =
22361 DAG.getNode(ISD::BUILD_VECTOR, DL, ByteVT, PSHUFBMask);
22362 DCI.AddToWorklist(PSHUFBMaskOp.getNode());
22363 Op = DAG.getNode(X86ISD::PSHUFB, DL, ByteVT, Op, PSHUFBMaskOp);
22364 DCI.AddToWorklist(Op.getNode());
22365 DCI.CombineTo(Root.getNode(), DAG.getBitcast(RootVT, Op),
22370 // Failed to find any combines.
22374 /// \brief Fully generic combining of x86 shuffle instructions.
22376 /// This should be the last combine run over the x86 shuffle instructions. Once
22377 /// they have been fully optimized, this will recursively consider all chains
22378 /// of single-use shuffle instructions, build a generic model of the cumulative
22379 /// shuffle operation, and check for simpler instructions which implement this
22380 /// operation. We use this primarily for two purposes:
22382 /// 1) Collapse generic shuffles to specialized single instructions when
22383 /// equivalent. In most cases, this is just an encoding size win, but
22384 /// sometimes we will collapse multiple generic shuffles into a single
22385 /// special-purpose shuffle.
22386 /// 2) Look for sequences of shuffle instructions with 3 or more total
22387 /// instructions, and replace them with the slightly more expensive SSSE3
22388 /// PSHUFB instruction if available. We do this as the last combining step
22389 /// to ensure we avoid using PSHUFB if we can implement the shuffle with
22390 /// a suitable short sequence of other instructions. The PHUFB will either
22391 /// use a register or have to read from memory and so is slightly (but only
22392 /// slightly) more expensive than the other shuffle instructions.
22394 /// Because this is inherently a quadratic operation (for each shuffle in
22395 /// a chain, we recurse up the chain), the depth is limited to 8 instructions.
22396 /// This should never be an issue in practice as the shuffle lowering doesn't
22397 /// produce sequences of more than 8 instructions.
22399 /// FIXME: We will currently miss some cases where the redundant shuffling
22400 /// would simplify under the threshold for PSHUFB formation because of
22401 /// combine-ordering. To fix this, we should do the redundant instruction
22402 /// combining in this recursive walk.
22403 static bool combineX86ShufflesRecursively(SDValue Op, SDValue Root,
22404 ArrayRef<int> RootMask,
22405 int Depth, bool HasPSHUFB,
22407 TargetLowering::DAGCombinerInfo &DCI,
22408 const X86Subtarget *Subtarget) {
22409 // Bound the depth of our recursive combine because this is ultimately
22410 // quadratic in nature.
22414 // Directly rip through bitcasts to find the underlying operand.
22415 while (Op.getOpcode() == ISD::BITCAST && Op.getOperand(0).hasOneUse())
22416 Op = Op.getOperand(0);
22418 MVT VT = Op.getSimpleValueType();
22419 if (!VT.isVector())
22420 return false; // Bail if we hit a non-vector.
22422 assert(Root.getSimpleValueType().isVector() &&
22423 "Shuffles operate on vector types!");
22424 assert(VT.getSizeInBits() == Root.getSimpleValueType().getSizeInBits() &&
22425 "Can only combine shuffles of the same vector register size.");
22427 if (!isTargetShuffle(Op.getOpcode()))
22429 SmallVector<int, 16> OpMask;
22431 bool HaveMask = getTargetShuffleMask(Op.getNode(), VT, OpMask, IsUnary);
22432 // We only can combine unary shuffles which we can decode the mask for.
22433 if (!HaveMask || !IsUnary)
22436 assert(VT.getVectorNumElements() == OpMask.size() &&
22437 "Different mask size from vector size!");
22438 assert(((RootMask.size() > OpMask.size() &&
22439 RootMask.size() % OpMask.size() == 0) ||
22440 (OpMask.size() > RootMask.size() &&
22441 OpMask.size() % RootMask.size() == 0) ||
22442 OpMask.size() == RootMask.size()) &&
22443 "The smaller number of elements must divide the larger.");
22444 int RootRatio = std::max<int>(1, OpMask.size() / RootMask.size());
22445 int OpRatio = std::max<int>(1, RootMask.size() / OpMask.size());
22446 assert(((RootRatio == 1 && OpRatio == 1) ||
22447 (RootRatio == 1) != (OpRatio == 1)) &&
22448 "Must not have a ratio for both incoming and op masks!");
22450 SmallVector<int, 16> Mask;
22451 Mask.reserve(std::max(OpMask.size(), RootMask.size()));
22453 // Merge this shuffle operation's mask into our accumulated mask. Note that
22454 // this shuffle's mask will be the first applied to the input, followed by the
22455 // root mask to get us all the way to the root value arrangement. The reason
22456 // for this order is that we are recursing up the operation chain.
22457 for (int i = 0, e = std::max(OpMask.size(), RootMask.size()); i < e; ++i) {
22458 int RootIdx = i / RootRatio;
22459 if (RootMask[RootIdx] < 0) {
22460 // This is a zero or undef lane, we're done.
22461 Mask.push_back(RootMask[RootIdx]);
22465 int RootMaskedIdx = RootMask[RootIdx] * RootRatio + i % RootRatio;
22466 int OpIdx = RootMaskedIdx / OpRatio;
22467 if (OpMask[OpIdx] < 0) {
22468 // The incoming lanes are zero or undef, it doesn't matter which ones we
22470 Mask.push_back(OpMask[OpIdx]);
22474 // Ok, we have non-zero lanes, map them through.
22475 Mask.push_back(OpMask[OpIdx] * OpRatio +
22476 RootMaskedIdx % OpRatio);
22479 // See if we can recurse into the operand to combine more things.
22480 switch (Op.getOpcode()) {
22481 case X86ISD::PSHUFB:
22483 case X86ISD::PSHUFD:
22484 case X86ISD::PSHUFHW:
22485 case X86ISD::PSHUFLW:
22486 if (Op.getOperand(0).hasOneUse() &&
22487 combineX86ShufflesRecursively(Op.getOperand(0), Root, Mask, Depth + 1,
22488 HasPSHUFB, DAG, DCI, Subtarget))
22492 case X86ISD::UNPCKL:
22493 case X86ISD::UNPCKH:
22494 assert(Op.getOperand(0) == Op.getOperand(1) &&
22495 "We only combine unary shuffles!");
22496 // We can't check for single use, we have to check that this shuffle is the
22498 if (Op->isOnlyUserOf(Op.getOperand(0).getNode()) &&
22499 combineX86ShufflesRecursively(Op.getOperand(0), Root, Mask, Depth + 1,
22500 HasPSHUFB, DAG, DCI, Subtarget))
22505 // Minor canonicalization of the accumulated shuffle mask to make it easier
22506 // to match below. All this does is detect masks with squential pairs of
22507 // elements, and shrink them to the half-width mask. It does this in a loop
22508 // so it will reduce the size of the mask to the minimal width mask which
22509 // performs an equivalent shuffle.
22510 SmallVector<int, 16> WidenedMask;
22511 while (Mask.size() > 1 && canWidenShuffleElements(Mask, WidenedMask)) {
22512 Mask = std::move(WidenedMask);
22513 WidenedMask.clear();
22516 return combineX86ShuffleChain(Op, Root, Mask, Depth, HasPSHUFB, DAG, DCI,
22520 /// \brief Get the PSHUF-style mask from PSHUF node.
22522 /// This is a very minor wrapper around getTargetShuffleMask to easy forming v4
22523 /// PSHUF-style masks that can be reused with such instructions.
22524 static SmallVector<int, 4> getPSHUFShuffleMask(SDValue N) {
22525 MVT VT = N.getSimpleValueType();
22526 SmallVector<int, 4> Mask;
22528 bool HaveMask = getTargetShuffleMask(N.getNode(), VT, Mask, IsUnary);
22532 // If we have more than 128-bits, only the low 128-bits of shuffle mask
22533 // matter. Check that the upper masks are repeats and remove them.
22534 if (VT.getSizeInBits() > 128) {
22535 int LaneElts = 128 / VT.getScalarSizeInBits();
22537 for (int i = 1, NumLanes = VT.getSizeInBits() / 128; i < NumLanes; ++i)
22538 for (int j = 0; j < LaneElts; ++j)
22539 assert(Mask[j] == Mask[i * LaneElts + j] - (LaneElts * i) &&
22540 "Mask doesn't repeat in high 128-bit lanes!");
22542 Mask.resize(LaneElts);
22545 switch (N.getOpcode()) {
22546 case X86ISD::PSHUFD:
22548 case X86ISD::PSHUFLW:
22551 case X86ISD::PSHUFHW:
22552 Mask.erase(Mask.begin(), Mask.begin() + 4);
22553 for (int &M : Mask)
22557 llvm_unreachable("No valid shuffle instruction found!");
22561 /// \brief Search for a combinable shuffle across a chain ending in pshufd.
22563 /// We walk up the chain and look for a combinable shuffle, skipping over
22564 /// shuffles that we could hoist this shuffle's transformation past without
22565 /// altering anything.
22567 combineRedundantDWordShuffle(SDValue N, MutableArrayRef<int> Mask,
22569 TargetLowering::DAGCombinerInfo &DCI) {
22570 assert(N.getOpcode() == X86ISD::PSHUFD &&
22571 "Called with something other than an x86 128-bit half shuffle!");
22574 // Walk up a single-use chain looking for a combinable shuffle. Keep a stack
22575 // of the shuffles in the chain so that we can form a fresh chain to replace
22577 SmallVector<SDValue, 8> Chain;
22578 SDValue V = N.getOperand(0);
22579 for (; V.hasOneUse(); V = V.getOperand(0)) {
22580 switch (V.getOpcode()) {
22582 return SDValue(); // Nothing combined!
22585 // Skip bitcasts as we always know the type for the target specific
22589 case X86ISD::PSHUFD:
22590 // Found another dword shuffle.
22593 case X86ISD::PSHUFLW:
22594 // Check that the low words (being shuffled) are the identity in the
22595 // dword shuffle, and the high words are self-contained.
22596 if (Mask[0] != 0 || Mask[1] != 1 ||
22597 !(Mask[2] >= 2 && Mask[2] < 4 && Mask[3] >= 2 && Mask[3] < 4))
22600 Chain.push_back(V);
22603 case X86ISD::PSHUFHW:
22604 // Check that the high words (being shuffled) are the identity in the
22605 // dword shuffle, and the low words are self-contained.
22606 if (Mask[2] != 2 || Mask[3] != 3 ||
22607 !(Mask[0] >= 0 && Mask[0] < 2 && Mask[1] >= 0 && Mask[1] < 2))
22610 Chain.push_back(V);
22613 case X86ISD::UNPCKL:
22614 case X86ISD::UNPCKH:
22615 // For either i8 -> i16 or i16 -> i32 unpacks, we can combine a dword
22616 // shuffle into a preceding word shuffle.
22617 if (V.getSimpleValueType().getScalarType() != MVT::i8 &&
22618 V.getSimpleValueType().getScalarType() != MVT::i16)
22621 // Search for a half-shuffle which we can combine with.
22622 unsigned CombineOp =
22623 V.getOpcode() == X86ISD::UNPCKL ? X86ISD::PSHUFLW : X86ISD::PSHUFHW;
22624 if (V.getOperand(0) != V.getOperand(1) ||
22625 !V->isOnlyUserOf(V.getOperand(0).getNode()))
22627 Chain.push_back(V);
22628 V = V.getOperand(0);
22630 switch (V.getOpcode()) {
22632 return SDValue(); // Nothing to combine.
22634 case X86ISD::PSHUFLW:
22635 case X86ISD::PSHUFHW:
22636 if (V.getOpcode() == CombineOp)
22639 Chain.push_back(V);
22643 V = V.getOperand(0);
22647 } while (V.hasOneUse());
22650 // Break out of the loop if we break out of the switch.
22654 if (!V.hasOneUse())
22655 // We fell out of the loop without finding a viable combining instruction.
22658 // Merge this node's mask and our incoming mask.
22659 SmallVector<int, 4> VMask = getPSHUFShuffleMask(V);
22660 for (int &M : Mask)
22662 V = DAG.getNode(V.getOpcode(), DL, V.getValueType(), V.getOperand(0),
22663 getV4X86ShuffleImm8ForMask(Mask, DL, DAG));
22665 // Rebuild the chain around this new shuffle.
22666 while (!Chain.empty()) {
22667 SDValue W = Chain.pop_back_val();
22669 if (V.getValueType() != W.getOperand(0).getValueType())
22670 V = DAG.getBitcast(W.getOperand(0).getValueType(), V);
22672 switch (W.getOpcode()) {
22674 llvm_unreachable("Only PSHUF and UNPCK instructions get here!");
22676 case X86ISD::UNPCKL:
22677 case X86ISD::UNPCKH:
22678 V = DAG.getNode(W.getOpcode(), DL, W.getValueType(), V, V);
22681 case X86ISD::PSHUFD:
22682 case X86ISD::PSHUFLW:
22683 case X86ISD::PSHUFHW:
22684 V = DAG.getNode(W.getOpcode(), DL, W.getValueType(), V, W.getOperand(1));
22688 if (V.getValueType() != N.getValueType())
22689 V = DAG.getBitcast(N.getValueType(), V);
22691 // Return the new chain to replace N.
22695 /// \brief Search for a combinable shuffle across a chain ending in pshuflw or
22698 /// We walk up the chain, skipping shuffles of the other half and looking
22699 /// through shuffles which switch halves trying to find a shuffle of the same
22700 /// pair of dwords.
22701 static bool combineRedundantHalfShuffle(SDValue N, MutableArrayRef<int> Mask,
22703 TargetLowering::DAGCombinerInfo &DCI) {
22705 (N.getOpcode() == X86ISD::PSHUFLW || N.getOpcode() == X86ISD::PSHUFHW) &&
22706 "Called with something other than an x86 128-bit half shuffle!");
22708 unsigned CombineOpcode = N.getOpcode();
22710 // Walk up a single-use chain looking for a combinable shuffle.
22711 SDValue V = N.getOperand(0);
22712 for (; V.hasOneUse(); V = V.getOperand(0)) {
22713 switch (V.getOpcode()) {
22715 return false; // Nothing combined!
22718 // Skip bitcasts as we always know the type for the target specific
22722 case X86ISD::PSHUFLW:
22723 case X86ISD::PSHUFHW:
22724 if (V.getOpcode() == CombineOpcode)
22727 // Other-half shuffles are no-ops.
22730 // Break out of the loop if we break out of the switch.
22734 if (!V.hasOneUse())
22735 // We fell out of the loop without finding a viable combining instruction.
22738 // Combine away the bottom node as its shuffle will be accumulated into
22739 // a preceding shuffle.
22740 DCI.CombineTo(N.getNode(), N.getOperand(0), /*AddTo*/ true);
22742 // Record the old value.
22745 // Merge this node's mask and our incoming mask (adjusted to account for all
22746 // the pshufd instructions encountered).
22747 SmallVector<int, 4> VMask = getPSHUFShuffleMask(V);
22748 for (int &M : Mask)
22750 V = DAG.getNode(V.getOpcode(), DL, MVT::v8i16, V.getOperand(0),
22751 getV4X86ShuffleImm8ForMask(Mask, DL, DAG));
22753 // Check that the shuffles didn't cancel each other out. If not, we need to
22754 // combine to the new one.
22756 // Replace the combinable shuffle with the combined one, updating all users
22757 // so that we re-evaluate the chain here.
22758 DCI.CombineTo(Old.getNode(), V, /*AddTo*/ true);
22763 /// \brief Try to combine x86 target specific shuffles.
22764 static SDValue PerformTargetShuffleCombine(SDValue N, SelectionDAG &DAG,
22765 TargetLowering::DAGCombinerInfo &DCI,
22766 const X86Subtarget *Subtarget) {
22768 MVT VT = N.getSimpleValueType();
22769 SmallVector<int, 4> Mask;
22771 switch (N.getOpcode()) {
22772 case X86ISD::PSHUFD:
22773 case X86ISD::PSHUFLW:
22774 case X86ISD::PSHUFHW:
22775 Mask = getPSHUFShuffleMask(N);
22776 assert(Mask.size() == 4);
22782 // Nuke no-op shuffles that show up after combining.
22783 if (isNoopShuffleMask(Mask))
22784 return DCI.CombineTo(N.getNode(), N.getOperand(0), /*AddTo*/ true);
22786 // Look for simplifications involving one or two shuffle instructions.
22787 SDValue V = N.getOperand(0);
22788 switch (N.getOpcode()) {
22791 case X86ISD::PSHUFLW:
22792 case X86ISD::PSHUFHW:
22793 assert(VT.getScalarType() == MVT::i16 && "Bad word shuffle type!");
22795 if (combineRedundantHalfShuffle(N, Mask, DAG, DCI))
22796 return SDValue(); // We combined away this shuffle, so we're done.
22798 // See if this reduces to a PSHUFD which is no more expensive and can
22799 // combine with more operations. Note that it has to at least flip the
22800 // dwords as otherwise it would have been removed as a no-op.
22801 if (makeArrayRef(Mask).equals({2, 3, 0, 1})) {
22802 int DMask[] = {0, 1, 2, 3};
22803 int DOffset = N.getOpcode() == X86ISD::PSHUFLW ? 0 : 2;
22804 DMask[DOffset + 0] = DOffset + 1;
22805 DMask[DOffset + 1] = DOffset + 0;
22806 MVT DVT = MVT::getVectorVT(MVT::i32, VT.getVectorNumElements() / 2);
22807 V = DAG.getBitcast(DVT, V);
22808 DCI.AddToWorklist(V.getNode());
22809 V = DAG.getNode(X86ISD::PSHUFD, DL, DVT, V,
22810 getV4X86ShuffleImm8ForMask(DMask, DL, DAG));
22811 DCI.AddToWorklist(V.getNode());
22812 return DAG.getBitcast(VT, V);
22815 // Look for shuffle patterns which can be implemented as a single unpack.
22816 // FIXME: This doesn't handle the location of the PSHUFD generically, and
22817 // only works when we have a PSHUFD followed by two half-shuffles.
22818 if (Mask[0] == Mask[1] && Mask[2] == Mask[3] &&
22819 (V.getOpcode() == X86ISD::PSHUFLW ||
22820 V.getOpcode() == X86ISD::PSHUFHW) &&
22821 V.getOpcode() != N.getOpcode() &&
22823 SDValue D = V.getOperand(0);
22824 while (D.getOpcode() == ISD::BITCAST && D.hasOneUse())
22825 D = D.getOperand(0);
22826 if (D.getOpcode() == X86ISD::PSHUFD && D.hasOneUse()) {
22827 SmallVector<int, 4> VMask = getPSHUFShuffleMask(V);
22828 SmallVector<int, 4> DMask = getPSHUFShuffleMask(D);
22829 int NOffset = N.getOpcode() == X86ISD::PSHUFLW ? 0 : 4;
22830 int VOffset = V.getOpcode() == X86ISD::PSHUFLW ? 0 : 4;
22832 for (int i = 0; i < 4; ++i) {
22833 WordMask[i + NOffset] = Mask[i] + NOffset;
22834 WordMask[i + VOffset] = VMask[i] + VOffset;
22836 // Map the word mask through the DWord mask.
22838 for (int i = 0; i < 8; ++i)
22839 MappedMask[i] = 2 * DMask[WordMask[i] / 2] + WordMask[i] % 2;
22840 if (makeArrayRef(MappedMask).equals({0, 0, 1, 1, 2, 2, 3, 3}) ||
22841 makeArrayRef(MappedMask).equals({4, 4, 5, 5, 6, 6, 7, 7})) {
22842 // We can replace all three shuffles with an unpack.
22843 V = DAG.getBitcast(VT, D.getOperand(0));
22844 DCI.AddToWorklist(V.getNode());
22845 return DAG.getNode(MappedMask[0] == 0 ? X86ISD::UNPCKL
22854 case X86ISD::PSHUFD:
22855 if (SDValue NewN = combineRedundantDWordShuffle(N, Mask, DAG, DCI))
22864 /// \brief Try to combine a shuffle into a target-specific add-sub node.
22866 /// We combine this directly on the abstract vector shuffle nodes so it is
22867 /// easier to generically match. We also insert dummy vector shuffle nodes for
22868 /// the operands which explicitly discard the lanes which are unused by this
22869 /// operation to try to flow through the rest of the combiner the fact that
22870 /// they're unused.
22871 static SDValue combineShuffleToAddSub(SDNode *N, SelectionDAG &DAG) {
22873 EVT VT = N->getValueType(0);
22875 // We only handle target-independent shuffles.
22876 // FIXME: It would be easy and harmless to use the target shuffle mask
22877 // extraction tool to support more.
22878 if (N->getOpcode() != ISD::VECTOR_SHUFFLE)
22881 auto *SVN = cast<ShuffleVectorSDNode>(N);
22882 ArrayRef<int> Mask = SVN->getMask();
22883 SDValue V1 = N->getOperand(0);
22884 SDValue V2 = N->getOperand(1);
22886 // We require the first shuffle operand to be the SUB node, and the second to
22887 // be the ADD node.
22888 // FIXME: We should support the commuted patterns.
22889 if (V1->getOpcode() != ISD::FSUB || V2->getOpcode() != ISD::FADD)
22892 // If there are other uses of these operations we can't fold them.
22893 if (!V1->hasOneUse() || !V2->hasOneUse())
22896 // Ensure that both operations have the same operands. Note that we can
22897 // commute the FADD operands.
22898 SDValue LHS = V1->getOperand(0), RHS = V1->getOperand(1);
22899 if ((V2->getOperand(0) != LHS || V2->getOperand(1) != RHS) &&
22900 (V2->getOperand(0) != RHS || V2->getOperand(1) != LHS))
22903 // We're looking for blends between FADD and FSUB nodes. We insist on these
22904 // nodes being lined up in a specific expected pattern.
22905 if (!(isShuffleEquivalent(V1, V2, Mask, {0, 3}) ||
22906 isShuffleEquivalent(V1, V2, Mask, {0, 5, 2, 7}) ||
22907 isShuffleEquivalent(V1, V2, Mask, {0, 9, 2, 11, 4, 13, 6, 15})))
22910 // Only specific types are legal at this point, assert so we notice if and
22911 // when these change.
22912 assert((VT == MVT::v4f32 || VT == MVT::v2f64 || VT == MVT::v8f32 ||
22913 VT == MVT::v4f64) &&
22914 "Unknown vector type encountered!");
22916 return DAG.getNode(X86ISD::ADDSUB, DL, VT, LHS, RHS);
22919 /// PerformShuffleCombine - Performs several different shuffle combines.
22920 static SDValue PerformShuffleCombine(SDNode *N, SelectionDAG &DAG,
22921 TargetLowering::DAGCombinerInfo &DCI,
22922 const X86Subtarget *Subtarget) {
22924 SDValue N0 = N->getOperand(0);
22925 SDValue N1 = N->getOperand(1);
22926 EVT VT = N->getValueType(0);
22928 // Don't create instructions with illegal types after legalize types has run.
22929 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
22930 if (!DCI.isBeforeLegalize() && !TLI.isTypeLegal(VT.getVectorElementType()))
22933 // If we have legalized the vector types, look for blends of FADD and FSUB
22934 // nodes that we can fuse into an ADDSUB node.
22935 if (TLI.isTypeLegal(VT) && Subtarget->hasSSE3())
22936 if (SDValue AddSub = combineShuffleToAddSub(N, DAG))
22939 // Combine 256-bit vector shuffles. This is only profitable when in AVX mode
22940 if (Subtarget->hasFp256() && VT.is256BitVector() &&
22941 N->getOpcode() == ISD::VECTOR_SHUFFLE)
22942 return PerformShuffleCombine256(N, DAG, DCI, Subtarget);
22944 // During Type Legalization, when promoting illegal vector types,
22945 // the backend might introduce new shuffle dag nodes and bitcasts.
22947 // This code performs the following transformation:
22948 // fold: (shuffle (bitcast (BINOP A, B)), Undef, <Mask>) ->
22949 // (shuffle (BINOP (bitcast A), (bitcast B)), Undef, <Mask>)
22951 // We do this only if both the bitcast and the BINOP dag nodes have
22952 // one use. Also, perform this transformation only if the new binary
22953 // operation is legal. This is to avoid introducing dag nodes that
22954 // potentially need to be further expanded (or custom lowered) into a
22955 // less optimal sequence of dag nodes.
22956 if (!DCI.isBeforeLegalize() && DCI.isBeforeLegalizeOps() &&
22957 N1.getOpcode() == ISD::UNDEF && N0.hasOneUse() &&
22958 N0.getOpcode() == ISD::BITCAST) {
22959 SDValue BC0 = N0.getOperand(0);
22960 EVT SVT = BC0.getValueType();
22961 unsigned Opcode = BC0.getOpcode();
22962 unsigned NumElts = VT.getVectorNumElements();
22964 if (BC0.hasOneUse() && SVT.isVector() &&
22965 SVT.getVectorNumElements() * 2 == NumElts &&
22966 TLI.isOperationLegal(Opcode, VT)) {
22967 bool CanFold = false;
22979 unsigned SVTNumElts = SVT.getVectorNumElements();
22980 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
22981 for (unsigned i = 0, e = SVTNumElts; i != e && CanFold; ++i)
22982 CanFold = SVOp->getMaskElt(i) == (int)(i * 2);
22983 for (unsigned i = SVTNumElts, e = NumElts; i != e && CanFold; ++i)
22984 CanFold = SVOp->getMaskElt(i) < 0;
22987 SDValue BC00 = DAG.getBitcast(VT, BC0.getOperand(0));
22988 SDValue BC01 = DAG.getBitcast(VT, BC0.getOperand(1));
22989 SDValue NewBinOp = DAG.getNode(BC0.getOpcode(), dl, VT, BC00, BC01);
22990 return DAG.getVectorShuffle(VT, dl, NewBinOp, N1, &SVOp->getMask()[0]);
22995 // Combine a vector_shuffle that is equal to build_vector load1, load2, load3,
22996 // load4, <0, 1, 2, 3> into a 128-bit load if the load addresses are
22997 // consecutive, non-overlapping, and in the right order.
22998 SmallVector<SDValue, 16> Elts;
22999 for (unsigned i = 0, e = VT.getVectorNumElements(); i != e; ++i)
23000 Elts.push_back(getShuffleScalarElt(N, i, DAG, 0));
23002 if (SDValue LD = EltsFromConsecutiveLoads(VT, Elts, dl, DAG, true))
23005 if (isTargetShuffle(N->getOpcode())) {
23007 PerformTargetShuffleCombine(SDValue(N, 0), DAG, DCI, Subtarget);
23008 if (Shuffle.getNode())
23011 // Try recursively combining arbitrary sequences of x86 shuffle
23012 // instructions into higher-order shuffles. We do this after combining
23013 // specific PSHUF instruction sequences into their minimal form so that we
23014 // can evaluate how many specialized shuffle instructions are involved in
23015 // a particular chain.
23016 SmallVector<int, 1> NonceMask; // Just a placeholder.
23017 NonceMask.push_back(0);
23018 if (combineX86ShufflesRecursively(SDValue(N, 0), SDValue(N, 0), NonceMask,
23019 /*Depth*/ 1, /*HasPSHUFB*/ false, DAG,
23021 return SDValue(); // This routine will use CombineTo to replace N.
23027 /// XFormVExtractWithShuffleIntoLoad - Check if a vector extract from a target
23028 /// specific shuffle of a load can be folded into a single element load.
23029 /// Similar handling for VECTOR_SHUFFLE is performed by DAGCombiner, but
23030 /// shuffles have been custom lowered so we need to handle those here.
23031 static SDValue XFormVExtractWithShuffleIntoLoad(SDNode *N, SelectionDAG &DAG,
23032 TargetLowering::DAGCombinerInfo &DCI) {
23033 if (DCI.isBeforeLegalizeOps())
23036 SDValue InVec = N->getOperand(0);
23037 SDValue EltNo = N->getOperand(1);
23039 if (!isa<ConstantSDNode>(EltNo))
23042 EVT OriginalVT = InVec.getValueType();
23044 if (InVec.getOpcode() == ISD::BITCAST) {
23045 // Don't duplicate a load with other uses.
23046 if (!InVec.hasOneUse())
23048 EVT BCVT = InVec.getOperand(0).getValueType();
23049 if (!BCVT.isVector() ||
23050 BCVT.getVectorNumElements() != OriginalVT.getVectorNumElements())
23052 InVec = InVec.getOperand(0);
23055 EVT CurrentVT = InVec.getValueType();
23057 if (!isTargetShuffle(InVec.getOpcode()))
23060 // Don't duplicate a load with other uses.
23061 if (!InVec.hasOneUse())
23064 SmallVector<int, 16> ShuffleMask;
23066 if (!getTargetShuffleMask(InVec.getNode(), CurrentVT.getSimpleVT(),
23067 ShuffleMask, UnaryShuffle))
23070 // Select the input vector, guarding against out of range extract vector.
23071 unsigned NumElems = CurrentVT.getVectorNumElements();
23072 int Elt = cast<ConstantSDNode>(EltNo)->getZExtValue();
23073 int Idx = (Elt > (int)NumElems) ? -1 : ShuffleMask[Elt];
23074 SDValue LdNode = (Idx < (int)NumElems) ? InVec.getOperand(0)
23075 : InVec.getOperand(1);
23077 // If inputs to shuffle are the same for both ops, then allow 2 uses
23078 unsigned AllowedUses = InVec.getNumOperands() > 1 &&
23079 InVec.getOperand(0) == InVec.getOperand(1) ? 2 : 1;
23081 if (LdNode.getOpcode() == ISD::BITCAST) {
23082 // Don't duplicate a load with other uses.
23083 if (!LdNode.getNode()->hasNUsesOfValue(AllowedUses, 0))
23086 AllowedUses = 1; // only allow 1 load use if we have a bitcast
23087 LdNode = LdNode.getOperand(0);
23090 if (!ISD::isNormalLoad(LdNode.getNode()))
23093 LoadSDNode *LN0 = cast<LoadSDNode>(LdNode);
23095 if (!LN0 ||!LN0->hasNUsesOfValue(AllowedUses, 0) || LN0->isVolatile())
23098 EVT EltVT = N->getValueType(0);
23099 // If there's a bitcast before the shuffle, check if the load type and
23100 // alignment is valid.
23101 unsigned Align = LN0->getAlignment();
23102 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
23103 unsigned NewAlign = DAG.getDataLayout().getABITypeAlignment(
23104 EltVT.getTypeForEVT(*DAG.getContext()));
23106 if (NewAlign > Align || !TLI.isOperationLegalOrCustom(ISD::LOAD, EltVT))
23109 // All checks match so transform back to vector_shuffle so that DAG combiner
23110 // can finish the job
23113 // Create shuffle node taking into account the case that its a unary shuffle
23114 SDValue Shuffle = (UnaryShuffle) ? DAG.getUNDEF(CurrentVT)
23115 : InVec.getOperand(1);
23116 Shuffle = DAG.getVectorShuffle(CurrentVT, dl,
23117 InVec.getOperand(0), Shuffle,
23119 Shuffle = DAG.getBitcast(OriginalVT, Shuffle);
23120 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, N->getValueType(0), Shuffle,
23124 static SDValue PerformBITCASTCombine(SDNode *N, SelectionDAG &DAG,
23125 const X86Subtarget *Subtarget) {
23126 SDValue N0 = N->getOperand(0);
23127 EVT VT = N->getValueType(0);
23129 // Detect bitcasts between i32 to x86mmx low word. Since MMX types are
23130 // special and don't usually play with other vector types, it's better to
23131 // handle them early to be sure we emit efficient code by avoiding
23132 // store-load conversions.
23133 if (VT == MVT::x86mmx && N0.getOpcode() == ISD::BUILD_VECTOR &&
23134 N0.getValueType() == MVT::v2i32 &&
23135 isa<ConstantSDNode>(N0.getOperand(1))) {
23136 SDValue N00 = N0->getOperand(0);
23137 if (N0.getConstantOperandVal(1) == 0 && N00.getValueType() == MVT::i32)
23138 return DAG.getNode(X86ISD::MMX_MOVW2D, SDLoc(N00), VT, N00);
23141 // Convert a bitcasted integer logic operation that has one bitcasted
23142 // floating-point operand and one constant operand into a floating-point
23143 // logic operation. This may create a load of the constant, but that is
23144 // cheaper than materializing the constant in an integer register and
23145 // transferring it to an SSE register or transferring the SSE operand to
23146 // integer register and back.
23148 switch (N0.getOpcode()) {
23149 case ISD::AND: FPOpcode = X86ISD::FAND; break;
23150 case ISD::OR: FPOpcode = X86ISD::FOR; break;
23151 case ISD::XOR: FPOpcode = X86ISD::FXOR; break;
23152 default: return SDValue();
23154 if (((Subtarget->hasSSE1() && VT == MVT::f32) ||
23155 (Subtarget->hasSSE2() && VT == MVT::f64)) &&
23156 isa<ConstantSDNode>(N0.getOperand(1)) &&
23157 N0.getOperand(0).getOpcode() == ISD::BITCAST &&
23158 N0.getOperand(0).getOperand(0).getValueType() == VT) {
23159 SDValue N000 = N0.getOperand(0).getOperand(0);
23160 SDValue FPConst = DAG.getBitcast(VT, N0.getOperand(1));
23161 return DAG.getNode(FPOpcode, SDLoc(N0), VT, N000, FPConst);
23167 /// PerformEXTRACT_VECTOR_ELTCombine - Detect vector gather/scatter index
23168 /// generation and convert it from being a bunch of shuffles and extracts
23169 /// into a somewhat faster sequence. For i686, the best sequence is apparently
23170 /// storing the value and loading scalars back, while for x64 we should
23171 /// use 64-bit extracts and shifts.
23172 static SDValue PerformEXTRACT_VECTOR_ELTCombine(SDNode *N, SelectionDAG &DAG,
23173 TargetLowering::DAGCombinerInfo &DCI) {
23174 if (SDValue NewOp = XFormVExtractWithShuffleIntoLoad(N, DAG, DCI))
23177 SDValue InputVector = N->getOperand(0);
23178 SDLoc dl(InputVector);
23179 // Detect mmx to i32 conversion through a v2i32 elt extract.
23180 if (InputVector.getOpcode() == ISD::BITCAST && InputVector.hasOneUse() &&
23181 N->getValueType(0) == MVT::i32 &&
23182 InputVector.getValueType() == MVT::v2i32) {
23184 // The bitcast source is a direct mmx result.
23185 SDValue MMXSrc = InputVector.getNode()->getOperand(0);
23186 if (MMXSrc.getValueType() == MVT::x86mmx)
23187 return DAG.getNode(X86ISD::MMX_MOVD2W, SDLoc(InputVector),
23188 N->getValueType(0),
23189 InputVector.getNode()->getOperand(0));
23191 // The mmx is indirect: (i64 extract_elt (v1i64 bitcast (x86mmx ...))).
23192 if (MMXSrc.getOpcode() == ISD::EXTRACT_VECTOR_ELT && MMXSrc.hasOneUse() &&
23193 MMXSrc.getValueType() == MVT::i64) {
23194 SDValue MMXSrcOp = MMXSrc.getOperand(0);
23195 if (MMXSrcOp.hasOneUse() && MMXSrcOp.getOpcode() == ISD::BITCAST &&
23196 MMXSrcOp.getValueType() == MVT::v1i64 &&
23197 MMXSrcOp.getOperand(0).getValueType() == MVT::x86mmx)
23198 return DAG.getNode(X86ISD::MMX_MOVD2W, SDLoc(InputVector),
23199 N->getValueType(0), MMXSrcOp.getOperand(0));
23203 EVT VT = N->getValueType(0);
23205 if (VT == MVT::i1 && dyn_cast<ConstantSDNode>(N->getOperand(1)) &&
23206 InputVector.getOpcode() == ISD::BITCAST &&
23207 dyn_cast<ConstantSDNode>(InputVector.getOperand(0))) {
23208 uint64_t ExtractedElt =
23209 cast<ConstantSDNode>(N->getOperand(1))->getZExtValue();
23210 uint64_t InputValue =
23211 cast<ConstantSDNode>(InputVector.getOperand(0))->getZExtValue();
23212 uint64_t Res = (InputValue >> ExtractedElt) & 1;
23213 return DAG.getConstant(Res, dl, MVT::i1);
23215 // Only operate on vectors of 4 elements, where the alternative shuffling
23216 // gets to be more expensive.
23217 if (InputVector.getValueType() != MVT::v4i32)
23220 // Check whether every use of InputVector is an EXTRACT_VECTOR_ELT with a
23221 // single use which is a sign-extend or zero-extend, and all elements are
23223 SmallVector<SDNode *, 4> Uses;
23224 unsigned ExtractedElements = 0;
23225 for (SDNode::use_iterator UI = InputVector.getNode()->use_begin(),
23226 UE = InputVector.getNode()->use_end(); UI != UE; ++UI) {
23227 if (UI.getUse().getResNo() != InputVector.getResNo())
23230 SDNode *Extract = *UI;
23231 if (Extract->getOpcode() != ISD::EXTRACT_VECTOR_ELT)
23234 if (Extract->getValueType(0) != MVT::i32)
23236 if (!Extract->hasOneUse())
23238 if (Extract->use_begin()->getOpcode() != ISD::SIGN_EXTEND &&
23239 Extract->use_begin()->getOpcode() != ISD::ZERO_EXTEND)
23241 if (!isa<ConstantSDNode>(Extract->getOperand(1)))
23244 // Record which element was extracted.
23245 ExtractedElements |=
23246 1 << cast<ConstantSDNode>(Extract->getOperand(1))->getZExtValue();
23248 Uses.push_back(Extract);
23251 // If not all the elements were used, this may not be worthwhile.
23252 if (ExtractedElements != 15)
23255 // Ok, we've now decided to do the transformation.
23256 // If 64-bit shifts are legal, use the extract-shift sequence,
23257 // otherwise bounce the vector off the cache.
23258 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
23261 if (TLI.isOperationLegal(ISD::SRA, MVT::i64)) {
23262 SDValue Cst = DAG.getBitcast(MVT::v2i64, InputVector);
23263 auto &DL = DAG.getDataLayout();
23264 EVT VecIdxTy = DAG.getTargetLoweringInfo().getVectorIdxTy(DL);
23265 SDValue BottomHalf = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i64, Cst,
23266 DAG.getConstant(0, dl, VecIdxTy));
23267 SDValue TopHalf = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i64, Cst,
23268 DAG.getConstant(1, dl, VecIdxTy));
23270 SDValue ShAmt = DAG.getConstant(
23271 32, dl, DAG.getTargetLoweringInfo().getShiftAmountTy(MVT::i64, DL));
23272 Vals[0] = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, BottomHalf);
23273 Vals[1] = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32,
23274 DAG.getNode(ISD::SRA, dl, MVT::i64, BottomHalf, ShAmt));
23275 Vals[2] = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, TopHalf);
23276 Vals[3] = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32,
23277 DAG.getNode(ISD::SRA, dl, MVT::i64, TopHalf, ShAmt));
23279 // Store the value to a temporary stack slot.
23280 SDValue StackPtr = DAG.CreateStackTemporary(InputVector.getValueType());
23281 SDValue Ch = DAG.getStore(DAG.getEntryNode(), dl, InputVector, StackPtr,
23282 MachinePointerInfo(), false, false, 0);
23284 EVT ElementType = InputVector.getValueType().getVectorElementType();
23285 unsigned EltSize = ElementType.getSizeInBits() / 8;
23287 // Replace each use (extract) with a load of the appropriate element.
23288 for (unsigned i = 0; i < 4; ++i) {
23289 uint64_t Offset = EltSize * i;
23290 auto PtrVT = TLI.getPointerTy(DAG.getDataLayout());
23291 SDValue OffsetVal = DAG.getConstant(Offset, dl, PtrVT);
23293 SDValue ScalarAddr =
23294 DAG.getNode(ISD::ADD, dl, PtrVT, StackPtr, OffsetVal);
23296 // Load the scalar.
23297 Vals[i] = DAG.getLoad(ElementType, dl, Ch,
23298 ScalarAddr, MachinePointerInfo(),
23299 false, false, false, 0);
23304 // Replace the extracts
23305 for (SmallVectorImpl<SDNode *>::iterator UI = Uses.begin(),
23306 UE = Uses.end(); UI != UE; ++UI) {
23307 SDNode *Extract = *UI;
23309 SDValue Idx = Extract->getOperand(1);
23310 uint64_t IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
23311 DAG.ReplaceAllUsesOfValueWith(SDValue(Extract, 0), Vals[IdxVal]);
23314 // The replacement was made in place; don't return anything.
23319 transformVSELECTtoBlendVECTOR_SHUFFLE(SDNode *N, SelectionDAG &DAG,
23320 const X86Subtarget *Subtarget) {
23322 SDValue Cond = N->getOperand(0);
23323 SDValue LHS = N->getOperand(1);
23324 SDValue RHS = N->getOperand(2);
23326 if (Cond.getOpcode() == ISD::SIGN_EXTEND) {
23327 SDValue CondSrc = Cond->getOperand(0);
23328 if (CondSrc->getOpcode() == ISD::SIGN_EXTEND_INREG)
23329 Cond = CondSrc->getOperand(0);
23332 if (!ISD::isBuildVectorOfConstantSDNodes(Cond.getNode()))
23335 // A vselect where all conditions and data are constants can be optimized into
23336 // a single vector load by SelectionDAGLegalize::ExpandBUILD_VECTOR().
23337 if (ISD::isBuildVectorOfConstantSDNodes(LHS.getNode()) &&
23338 ISD::isBuildVectorOfConstantSDNodes(RHS.getNode()))
23341 unsigned MaskValue = 0;
23342 if (!BUILD_VECTORtoBlendMask(cast<BuildVectorSDNode>(Cond), MaskValue))
23345 MVT VT = N->getSimpleValueType(0);
23346 unsigned NumElems = VT.getVectorNumElements();
23347 SmallVector<int, 8> ShuffleMask(NumElems, -1);
23348 for (unsigned i = 0; i < NumElems; ++i) {
23349 // Be sure we emit undef where we can.
23350 if (Cond.getOperand(i)->getOpcode() == ISD::UNDEF)
23351 ShuffleMask[i] = -1;
23353 ShuffleMask[i] = i + NumElems * ((MaskValue >> i) & 1);
23356 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
23357 if (!TLI.isShuffleMaskLegal(ShuffleMask, VT))
23359 return DAG.getVectorShuffle(VT, dl, LHS, RHS, &ShuffleMask[0]);
23362 /// PerformSELECTCombine - Do target-specific dag combines on SELECT and VSELECT
23364 static SDValue PerformSELECTCombine(SDNode *N, SelectionDAG &DAG,
23365 TargetLowering::DAGCombinerInfo &DCI,
23366 const X86Subtarget *Subtarget) {
23368 SDValue Cond = N->getOperand(0);
23369 // Get the LHS/RHS of the select.
23370 SDValue LHS = N->getOperand(1);
23371 SDValue RHS = N->getOperand(2);
23372 EVT VT = LHS.getValueType();
23373 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
23375 // If we have SSE[12] support, try to form min/max nodes. SSE min/max
23376 // instructions match the semantics of the common C idiom x<y?x:y but not
23377 // x<=y?x:y, because of how they handle negative zero (which can be
23378 // ignored in unsafe-math mode).
23379 // We also try to create v2f32 min/max nodes, which we later widen to v4f32.
23380 if (Cond.getOpcode() == ISD::SETCC && VT.isFloatingPoint() &&
23381 VT != MVT::f80 && (TLI.isTypeLegal(VT) || VT == MVT::v2f32) &&
23382 (Subtarget->hasSSE2() ||
23383 (Subtarget->hasSSE1() && VT.getScalarType() == MVT::f32))) {
23384 ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
23386 unsigned Opcode = 0;
23387 // Check for x CC y ? x : y.
23388 if (DAG.isEqualTo(LHS, Cond.getOperand(0)) &&
23389 DAG.isEqualTo(RHS, Cond.getOperand(1))) {
23393 // Converting this to a min would handle NaNs incorrectly, and swapping
23394 // the operands would cause it to handle comparisons between positive
23395 // and negative zero incorrectly.
23396 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)) {
23397 if (!DAG.getTarget().Options.UnsafeFPMath &&
23398 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
23400 std::swap(LHS, RHS);
23402 Opcode = X86ISD::FMIN;
23405 // Converting this to a min would handle comparisons between positive
23406 // and negative zero incorrectly.
23407 if (!DAG.getTarget().Options.UnsafeFPMath &&
23408 !DAG.isKnownNeverZero(LHS) && !DAG.isKnownNeverZero(RHS))
23410 Opcode = X86ISD::FMIN;
23413 // Converting this to a min would handle both negative zeros and NaNs
23414 // incorrectly, but we can swap the operands to fix both.
23415 std::swap(LHS, RHS);
23419 Opcode = X86ISD::FMIN;
23423 // Converting this to a max would handle comparisons between positive
23424 // and negative zero incorrectly.
23425 if (!DAG.getTarget().Options.UnsafeFPMath &&
23426 !DAG.isKnownNeverZero(LHS) && !DAG.isKnownNeverZero(RHS))
23428 Opcode = X86ISD::FMAX;
23431 // Converting this to a max would handle NaNs incorrectly, and swapping
23432 // the operands would cause it to handle comparisons between positive
23433 // and negative zero incorrectly.
23434 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)) {
23435 if (!DAG.getTarget().Options.UnsafeFPMath &&
23436 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
23438 std::swap(LHS, RHS);
23440 Opcode = X86ISD::FMAX;
23443 // Converting this to a max would handle both negative zeros and NaNs
23444 // incorrectly, but we can swap the operands to fix both.
23445 std::swap(LHS, RHS);
23449 Opcode = X86ISD::FMAX;
23452 // Check for x CC y ? y : x -- a min/max with reversed arms.
23453 } else if (DAG.isEqualTo(LHS, Cond.getOperand(1)) &&
23454 DAG.isEqualTo(RHS, Cond.getOperand(0))) {
23458 // Converting this to a min would handle comparisons between positive
23459 // and negative zero incorrectly, and swapping the operands would
23460 // cause it to handle NaNs incorrectly.
23461 if (!DAG.getTarget().Options.UnsafeFPMath &&
23462 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS))) {
23463 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS))
23465 std::swap(LHS, RHS);
23467 Opcode = X86ISD::FMIN;
23470 // Converting this to a min would handle NaNs incorrectly.
23471 if (!DAG.getTarget().Options.UnsafeFPMath &&
23472 (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)))
23474 Opcode = X86ISD::FMIN;
23477 // Converting this to a min would handle both negative zeros and NaNs
23478 // incorrectly, but we can swap the operands to fix both.
23479 std::swap(LHS, RHS);
23483 Opcode = X86ISD::FMIN;
23487 // Converting this to a max would handle NaNs incorrectly.
23488 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS))
23490 Opcode = X86ISD::FMAX;
23493 // Converting this to a max would handle comparisons between positive
23494 // and negative zero incorrectly, and swapping the operands would
23495 // cause it to handle NaNs incorrectly.
23496 if (!DAG.getTarget().Options.UnsafeFPMath &&
23497 !DAG.isKnownNeverZero(LHS) && !DAG.isKnownNeverZero(RHS)) {
23498 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS))
23500 std::swap(LHS, RHS);
23502 Opcode = X86ISD::FMAX;
23505 // Converting this to a max would handle both negative zeros and NaNs
23506 // incorrectly, but we can swap the operands to fix both.
23507 std::swap(LHS, RHS);
23511 Opcode = X86ISD::FMAX;
23517 return DAG.getNode(Opcode, DL, N->getValueType(0), LHS, RHS);
23520 EVT CondVT = Cond.getValueType();
23521 if (Subtarget->hasAVX512() && VT.isVector() && CondVT.isVector() &&
23522 CondVT.getVectorElementType() == MVT::i1) {
23523 // v16i8 (select v16i1, v16i8, v16i8) does not have a proper
23524 // lowering on KNL. In this case we convert it to
23525 // v16i8 (select v16i8, v16i8, v16i8) and use AVX instruction.
23526 // The same situation for all 128 and 256-bit vectors of i8 and i16.
23527 // Since SKX these selects have a proper lowering.
23528 EVT OpVT = LHS.getValueType();
23529 if ((OpVT.is128BitVector() || OpVT.is256BitVector()) &&
23530 (OpVT.getVectorElementType() == MVT::i8 ||
23531 OpVT.getVectorElementType() == MVT::i16) &&
23532 !(Subtarget->hasBWI() && Subtarget->hasVLX())) {
23533 Cond = DAG.getNode(ISD::SIGN_EXTEND, DL, OpVT, Cond);
23534 DCI.AddToWorklist(Cond.getNode());
23535 return DAG.getNode(N->getOpcode(), DL, OpVT, Cond, LHS, RHS);
23538 // If this is a select between two integer constants, try to do some
23540 if (ConstantSDNode *TrueC = dyn_cast<ConstantSDNode>(LHS)) {
23541 if (ConstantSDNode *FalseC = dyn_cast<ConstantSDNode>(RHS))
23542 // Don't do this for crazy integer types.
23543 if (DAG.getTargetLoweringInfo().isTypeLegal(LHS.getValueType())) {
23544 // If this is efficiently invertible, canonicalize the LHSC/RHSC values
23545 // so that TrueC (the true value) is larger than FalseC.
23546 bool NeedsCondInvert = false;
23548 if (TrueC->getAPIntValue().ult(FalseC->getAPIntValue()) &&
23549 // Efficiently invertible.
23550 (Cond.getOpcode() == ISD::SETCC || // setcc -> invertible.
23551 (Cond.getOpcode() == ISD::XOR && // xor(X, C) -> invertible.
23552 isa<ConstantSDNode>(Cond.getOperand(1))))) {
23553 NeedsCondInvert = true;
23554 std::swap(TrueC, FalseC);
23557 // Optimize C ? 8 : 0 -> zext(C) << 3. Likewise for any pow2/0.
23558 if (FalseC->getAPIntValue() == 0 &&
23559 TrueC->getAPIntValue().isPowerOf2()) {
23560 if (NeedsCondInvert) // Invert the condition if needed.
23561 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
23562 DAG.getConstant(1, DL, Cond.getValueType()));
23564 // Zero extend the condition if needed.
23565 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, LHS.getValueType(), Cond);
23567 unsigned ShAmt = TrueC->getAPIntValue().logBase2();
23568 return DAG.getNode(ISD::SHL, DL, LHS.getValueType(), Cond,
23569 DAG.getConstant(ShAmt, DL, MVT::i8));
23572 // Optimize Cond ? cst+1 : cst -> zext(setcc(C)+cst.
23573 if (FalseC->getAPIntValue()+1 == TrueC->getAPIntValue()) {
23574 if (NeedsCondInvert) // Invert the condition if needed.
23575 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
23576 DAG.getConstant(1, DL, Cond.getValueType()));
23578 // Zero extend the condition if needed.
23579 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL,
23580 FalseC->getValueType(0), Cond);
23581 return DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
23582 SDValue(FalseC, 0));
23585 // Optimize cases that will turn into an LEA instruction. This requires
23586 // an i32 or i64 and an efficient multiplier (1, 2, 3, 4, 5, 8, 9).
23587 if (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i64) {
23588 uint64_t Diff = TrueC->getZExtValue()-FalseC->getZExtValue();
23589 if (N->getValueType(0) == MVT::i32) Diff = (unsigned)Diff;
23591 bool isFastMultiplier = false;
23593 switch ((unsigned char)Diff) {
23595 case 1: // result = add base, cond
23596 case 2: // result = lea base( , cond*2)
23597 case 3: // result = lea base(cond, cond*2)
23598 case 4: // result = lea base( , cond*4)
23599 case 5: // result = lea base(cond, cond*4)
23600 case 8: // result = lea base( , cond*8)
23601 case 9: // result = lea base(cond, cond*8)
23602 isFastMultiplier = true;
23607 if (isFastMultiplier) {
23608 APInt Diff = TrueC->getAPIntValue()-FalseC->getAPIntValue();
23609 if (NeedsCondInvert) // Invert the condition if needed.
23610 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
23611 DAG.getConstant(1, DL, Cond.getValueType()));
23613 // Zero extend the condition if needed.
23614 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, FalseC->getValueType(0),
23616 // Scale the condition by the difference.
23618 Cond = DAG.getNode(ISD::MUL, DL, Cond.getValueType(), Cond,
23619 DAG.getConstant(Diff, DL,
23620 Cond.getValueType()));
23622 // Add the base if non-zero.
23623 if (FalseC->getAPIntValue() != 0)
23624 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
23625 SDValue(FalseC, 0));
23632 // Canonicalize max and min:
23633 // (x > y) ? x : y -> (x >= y) ? x : y
23634 // (x < y) ? x : y -> (x <= y) ? x : y
23635 // This allows use of COND_S / COND_NS (see TranslateX86CC) which eliminates
23636 // the need for an extra compare
23637 // against zero. e.g.
23638 // (x - y) > 0 : (x - y) ? 0 -> (x - y) >= 0 : (x - y) ? 0
23640 // testl %edi, %edi
23642 // cmovgl %edi, %eax
23646 // cmovsl %eax, %edi
23647 if (N->getOpcode() == ISD::SELECT && Cond.getOpcode() == ISD::SETCC &&
23648 DAG.isEqualTo(LHS, Cond.getOperand(0)) &&
23649 DAG.isEqualTo(RHS, Cond.getOperand(1))) {
23650 ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
23655 ISD::CondCode NewCC = (CC == ISD::SETLT) ? ISD::SETLE : ISD::SETGE;
23656 Cond = DAG.getSetCC(SDLoc(Cond), Cond.getValueType(),
23657 Cond.getOperand(0), Cond.getOperand(1), NewCC);
23658 return DAG.getNode(ISD::SELECT, DL, VT, Cond, LHS, RHS);
23663 // Early exit check
23664 if (!TLI.isTypeLegal(VT))
23667 // Match VSELECTs into subs with unsigned saturation.
23668 if (N->getOpcode() == ISD::VSELECT && Cond.getOpcode() == ISD::SETCC &&
23669 // psubus is available in SSE2 and AVX2 for i8 and i16 vectors.
23670 ((Subtarget->hasSSE2() && (VT == MVT::v16i8 || VT == MVT::v8i16)) ||
23671 (Subtarget->hasAVX2() && (VT == MVT::v32i8 || VT == MVT::v16i16)))) {
23672 ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
23674 // Check if one of the arms of the VSELECT is a zero vector. If it's on the
23675 // left side invert the predicate to simplify logic below.
23677 if (ISD::isBuildVectorAllZeros(LHS.getNode())) {
23679 CC = ISD::getSetCCInverse(CC, true);
23680 } else if (ISD::isBuildVectorAllZeros(RHS.getNode())) {
23684 if (Other.getNode() && Other->getNumOperands() == 2 &&
23685 DAG.isEqualTo(Other->getOperand(0), Cond.getOperand(0))) {
23686 SDValue OpLHS = Other->getOperand(0), OpRHS = Other->getOperand(1);
23687 SDValue CondRHS = Cond->getOperand(1);
23689 // Look for a general sub with unsigned saturation first.
23690 // x >= y ? x-y : 0 --> subus x, y
23691 // x > y ? x-y : 0 --> subus x, y
23692 if ((CC == ISD::SETUGE || CC == ISD::SETUGT) &&
23693 Other->getOpcode() == ISD::SUB && DAG.isEqualTo(OpRHS, CondRHS))
23694 return DAG.getNode(X86ISD::SUBUS, DL, VT, OpLHS, OpRHS);
23696 if (auto *OpRHSBV = dyn_cast<BuildVectorSDNode>(OpRHS))
23697 if (auto *OpRHSConst = OpRHSBV->getConstantSplatNode()) {
23698 if (auto *CondRHSBV = dyn_cast<BuildVectorSDNode>(CondRHS))
23699 if (auto *CondRHSConst = CondRHSBV->getConstantSplatNode())
23700 // If the RHS is a constant we have to reverse the const
23701 // canonicalization.
23702 // x > C-1 ? x+-C : 0 --> subus x, C
23703 if (CC == ISD::SETUGT && Other->getOpcode() == ISD::ADD &&
23704 CondRHSConst->getAPIntValue() ==
23705 (-OpRHSConst->getAPIntValue() - 1))
23706 return DAG.getNode(
23707 X86ISD::SUBUS, DL, VT, OpLHS,
23708 DAG.getConstant(-OpRHSConst->getAPIntValue(), DL, VT));
23710 // Another special case: If C was a sign bit, the sub has been
23711 // canonicalized into a xor.
23712 // FIXME: Would it be better to use computeKnownBits to determine
23713 // whether it's safe to decanonicalize the xor?
23714 // x s< 0 ? x^C : 0 --> subus x, C
23715 if (CC == ISD::SETLT && Other->getOpcode() == ISD::XOR &&
23716 ISD::isBuildVectorAllZeros(CondRHS.getNode()) &&
23717 OpRHSConst->getAPIntValue().isSignBit())
23718 // Note that we have to rebuild the RHS constant here to ensure we
23719 // don't rely on particular values of undef lanes.
23720 return DAG.getNode(
23721 X86ISD::SUBUS, DL, VT, OpLHS,
23722 DAG.getConstant(OpRHSConst->getAPIntValue(), DL, VT));
23727 // Simplify vector selection if condition value type matches vselect
23729 if (N->getOpcode() == ISD::VSELECT && CondVT == VT) {
23730 assert(Cond.getValueType().isVector() &&
23731 "vector select expects a vector selector!");
23733 bool TValIsAllOnes = ISD::isBuildVectorAllOnes(LHS.getNode());
23734 bool FValIsAllZeros = ISD::isBuildVectorAllZeros(RHS.getNode());
23736 // Try invert the condition if true value is not all 1s and false value
23738 if (!TValIsAllOnes && !FValIsAllZeros &&
23739 // Check if the selector will be produced by CMPP*/PCMP*
23740 Cond.getOpcode() == ISD::SETCC &&
23741 // Check if SETCC has already been promoted
23742 TLI.getSetCCResultType(DAG.getDataLayout(), *DAG.getContext(), VT) ==
23744 bool TValIsAllZeros = ISD::isBuildVectorAllZeros(LHS.getNode());
23745 bool FValIsAllOnes = ISD::isBuildVectorAllOnes(RHS.getNode());
23747 if (TValIsAllZeros || FValIsAllOnes) {
23748 SDValue CC = Cond.getOperand(2);
23749 ISD::CondCode NewCC =
23750 ISD::getSetCCInverse(cast<CondCodeSDNode>(CC)->get(),
23751 Cond.getOperand(0).getValueType().isInteger());
23752 Cond = DAG.getSetCC(DL, CondVT, Cond.getOperand(0), Cond.getOperand(1), NewCC);
23753 std::swap(LHS, RHS);
23754 TValIsAllOnes = FValIsAllOnes;
23755 FValIsAllZeros = TValIsAllZeros;
23759 if (TValIsAllOnes || FValIsAllZeros) {
23762 if (TValIsAllOnes && FValIsAllZeros)
23764 else if (TValIsAllOnes)
23766 DAG.getNode(ISD::OR, DL, CondVT, Cond, DAG.getBitcast(CondVT, RHS));
23767 else if (FValIsAllZeros)
23768 Ret = DAG.getNode(ISD::AND, DL, CondVT, Cond,
23769 DAG.getBitcast(CondVT, LHS));
23771 return DAG.getBitcast(VT, Ret);
23775 // We should generate an X86ISD::BLENDI from a vselect if its argument
23776 // is a sign_extend_inreg of an any_extend of a BUILD_VECTOR of
23777 // constants. This specific pattern gets generated when we split a
23778 // selector for a 512 bit vector in a machine without AVX512 (but with
23779 // 256-bit vectors), during legalization:
23781 // (vselect (sign_extend (any_extend (BUILD_VECTOR)) i1) LHS RHS)
23783 // Iff we find this pattern and the build_vectors are built from
23784 // constants, we translate the vselect into a shuffle_vector that we
23785 // know will be matched by LowerVECTOR_SHUFFLEtoBlend.
23786 if ((N->getOpcode() == ISD::VSELECT ||
23787 N->getOpcode() == X86ISD::SHRUNKBLEND) &&
23788 !DCI.isBeforeLegalize() && !VT.is512BitVector()) {
23789 SDValue Shuffle = transformVSELECTtoBlendVECTOR_SHUFFLE(N, DAG, Subtarget);
23790 if (Shuffle.getNode())
23794 // If this is a *dynamic* select (non-constant condition) and we can match
23795 // this node with one of the variable blend instructions, restructure the
23796 // condition so that the blends can use the high bit of each element and use
23797 // SimplifyDemandedBits to simplify the condition operand.
23798 if (N->getOpcode() == ISD::VSELECT && DCI.isBeforeLegalizeOps() &&
23799 !DCI.isBeforeLegalize() &&
23800 !ISD::isBuildVectorOfConstantSDNodes(Cond.getNode())) {
23801 unsigned BitWidth = Cond.getValueType().getScalarType().getSizeInBits();
23803 // Don't optimize vector selects that map to mask-registers.
23807 // We can only handle the cases where VSELECT is directly legal on the
23808 // subtarget. We custom lower VSELECT nodes with constant conditions and
23809 // this makes it hard to see whether a dynamic VSELECT will correctly
23810 // lower, so we both check the operation's status and explicitly handle the
23811 // cases where a *dynamic* blend will fail even though a constant-condition
23812 // blend could be custom lowered.
23813 // FIXME: We should find a better way to handle this class of problems.
23814 // Potentially, we should combine constant-condition vselect nodes
23815 // pre-legalization into shuffles and not mark as many types as custom
23817 if (!TLI.isOperationLegalOrCustom(ISD::VSELECT, VT))
23819 // FIXME: We don't support i16-element blends currently. We could and
23820 // should support them by making *all* the bits in the condition be set
23821 // rather than just the high bit and using an i8-element blend.
23822 if (VT.getScalarType() == MVT::i16)
23824 // Dynamic blending was only available from SSE4.1 onward.
23825 if (VT.getSizeInBits() == 128 && !Subtarget->hasSSE41())
23827 // Byte blends are only available in AVX2
23828 if (VT.getSizeInBits() == 256 && VT.getScalarType() == MVT::i8 &&
23829 !Subtarget->hasAVX2())
23832 assert(BitWidth >= 8 && BitWidth <= 64 && "Invalid mask size");
23833 APInt DemandedMask = APInt::getHighBitsSet(BitWidth, 1);
23835 APInt KnownZero, KnownOne;
23836 TargetLowering::TargetLoweringOpt TLO(DAG, DCI.isBeforeLegalize(),
23837 DCI.isBeforeLegalizeOps());
23838 if (TLO.ShrinkDemandedConstant(Cond, DemandedMask) ||
23839 TLI.SimplifyDemandedBits(Cond, DemandedMask, KnownZero, KnownOne,
23841 // If we changed the computation somewhere in the DAG, this change
23842 // will affect all users of Cond.
23843 // Make sure it is fine and update all the nodes so that we do not
23844 // use the generic VSELECT anymore. Otherwise, we may perform
23845 // wrong optimizations as we messed up with the actual expectation
23846 // for the vector boolean values.
23847 if (Cond != TLO.Old) {
23848 // Check all uses of that condition operand to check whether it will be
23849 // consumed by non-BLEND instructions, which may depend on all bits are
23851 for (SDNode::use_iterator I = Cond->use_begin(), E = Cond->use_end();
23853 if (I->getOpcode() != ISD::VSELECT)
23854 // TODO: Add other opcodes eventually lowered into BLEND.
23857 // Update all the users of the condition, before committing the change,
23858 // so that the VSELECT optimizations that expect the correct vector
23859 // boolean value will not be triggered.
23860 for (SDNode::use_iterator I = Cond->use_begin(), E = Cond->use_end();
23862 DAG.ReplaceAllUsesOfValueWith(
23864 DAG.getNode(X86ISD::SHRUNKBLEND, SDLoc(*I), I->getValueType(0),
23865 Cond, I->getOperand(1), I->getOperand(2)));
23866 DCI.CommitTargetLoweringOpt(TLO);
23869 // At this point, only Cond is changed. Change the condition
23870 // just for N to keep the opportunity to optimize all other
23871 // users their own way.
23872 DAG.ReplaceAllUsesOfValueWith(
23874 DAG.getNode(X86ISD::SHRUNKBLEND, SDLoc(N), N->getValueType(0),
23875 TLO.New, N->getOperand(1), N->getOperand(2)));
23883 // Check whether a boolean test is testing a boolean value generated by
23884 // X86ISD::SETCC. If so, return the operand of that SETCC and proper condition
23887 // Simplify the following patterns:
23888 // (Op (CMP (SETCC Cond EFLAGS) 1) EQ) or
23889 // (Op (CMP (SETCC Cond EFLAGS) 0) NEQ)
23890 // to (Op EFLAGS Cond)
23892 // (Op (CMP (SETCC Cond EFLAGS) 0) EQ) or
23893 // (Op (CMP (SETCC Cond EFLAGS) 1) NEQ)
23894 // to (Op EFLAGS !Cond)
23896 // where Op could be BRCOND or CMOV.
23898 static SDValue checkBoolTestSetCCCombine(SDValue Cmp, X86::CondCode &CC) {
23899 // Quit if not CMP and SUB with its value result used.
23900 if (Cmp.getOpcode() != X86ISD::CMP &&
23901 (Cmp.getOpcode() != X86ISD::SUB || Cmp.getNode()->hasAnyUseOfValue(0)))
23904 // Quit if not used as a boolean value.
23905 if (CC != X86::COND_E && CC != X86::COND_NE)
23908 // Check CMP operands. One of them should be 0 or 1 and the other should be
23909 // an SetCC or extended from it.
23910 SDValue Op1 = Cmp.getOperand(0);
23911 SDValue Op2 = Cmp.getOperand(1);
23914 const ConstantSDNode* C = nullptr;
23915 bool needOppositeCond = (CC == X86::COND_E);
23916 bool checkAgainstTrue = false; // Is it a comparison against 1?
23918 if ((C = dyn_cast<ConstantSDNode>(Op1)))
23920 else if ((C = dyn_cast<ConstantSDNode>(Op2)))
23922 else // Quit if all operands are not constants.
23925 if (C->getZExtValue() == 1) {
23926 needOppositeCond = !needOppositeCond;
23927 checkAgainstTrue = true;
23928 } else if (C->getZExtValue() != 0)
23929 // Quit if the constant is neither 0 or 1.
23932 bool truncatedToBoolWithAnd = false;
23933 // Skip (zext $x), (trunc $x), or (and $x, 1) node.
23934 while (SetCC.getOpcode() == ISD::ZERO_EXTEND ||
23935 SetCC.getOpcode() == ISD::TRUNCATE ||
23936 SetCC.getOpcode() == ISD::AND) {
23937 if (SetCC.getOpcode() == ISD::AND) {
23939 ConstantSDNode *CS;
23940 if ((CS = dyn_cast<ConstantSDNode>(SetCC.getOperand(0))) &&
23941 CS->getZExtValue() == 1)
23943 if ((CS = dyn_cast<ConstantSDNode>(SetCC.getOperand(1))) &&
23944 CS->getZExtValue() == 1)
23948 SetCC = SetCC.getOperand(OpIdx);
23949 truncatedToBoolWithAnd = true;
23951 SetCC = SetCC.getOperand(0);
23954 switch (SetCC.getOpcode()) {
23955 case X86ISD::SETCC_CARRY:
23956 // Since SETCC_CARRY gives output based on R = CF ? ~0 : 0, it's unsafe to
23957 // simplify it if the result of SETCC_CARRY is not canonicalized to 0 or 1,
23958 // i.e. it's a comparison against true but the result of SETCC_CARRY is not
23959 // truncated to i1 using 'and'.
23960 if (checkAgainstTrue && !truncatedToBoolWithAnd)
23962 assert(X86::CondCode(SetCC.getConstantOperandVal(0)) == X86::COND_B &&
23963 "Invalid use of SETCC_CARRY!");
23965 case X86ISD::SETCC:
23966 // Set the condition code or opposite one if necessary.
23967 CC = X86::CondCode(SetCC.getConstantOperandVal(0));
23968 if (needOppositeCond)
23969 CC = X86::GetOppositeBranchCondition(CC);
23970 return SetCC.getOperand(1);
23971 case X86ISD::CMOV: {
23972 // Check whether false/true value has canonical one, i.e. 0 or 1.
23973 ConstantSDNode *FVal = dyn_cast<ConstantSDNode>(SetCC.getOperand(0));
23974 ConstantSDNode *TVal = dyn_cast<ConstantSDNode>(SetCC.getOperand(1));
23975 // Quit if true value is not a constant.
23978 // Quit if false value is not a constant.
23980 SDValue Op = SetCC.getOperand(0);
23981 // Skip 'zext' or 'trunc' node.
23982 if (Op.getOpcode() == ISD::ZERO_EXTEND ||
23983 Op.getOpcode() == ISD::TRUNCATE)
23984 Op = Op.getOperand(0);
23985 // A special case for rdrand/rdseed, where 0 is set if false cond is
23987 if ((Op.getOpcode() != X86ISD::RDRAND &&
23988 Op.getOpcode() != X86ISD::RDSEED) || Op.getResNo() != 0)
23991 // Quit if false value is not the constant 0 or 1.
23992 bool FValIsFalse = true;
23993 if (FVal && FVal->getZExtValue() != 0) {
23994 if (FVal->getZExtValue() != 1)
23996 // If FVal is 1, opposite cond is needed.
23997 needOppositeCond = !needOppositeCond;
23998 FValIsFalse = false;
24000 // Quit if TVal is not the constant opposite of FVal.
24001 if (FValIsFalse && TVal->getZExtValue() != 1)
24003 if (!FValIsFalse && TVal->getZExtValue() != 0)
24005 CC = X86::CondCode(SetCC.getConstantOperandVal(2));
24006 if (needOppositeCond)
24007 CC = X86::GetOppositeBranchCondition(CC);
24008 return SetCC.getOperand(3);
24015 /// Check whether Cond is an AND/OR of SETCCs off of the same EFLAGS.
24017 /// (X86or (X86setcc) (X86setcc))
24018 /// (X86cmp (and (X86setcc) (X86setcc)), 0)
24019 static bool checkBoolTestAndOrSetCCCombine(SDValue Cond, X86::CondCode &CC0,
24020 X86::CondCode &CC1, SDValue &Flags,
24022 if (Cond->getOpcode() == X86ISD::CMP) {
24023 ConstantSDNode *CondOp1C = dyn_cast<ConstantSDNode>(Cond->getOperand(1));
24024 if (!CondOp1C || !CondOp1C->isNullValue())
24027 Cond = Cond->getOperand(0);
24032 SDValue SetCC0, SetCC1;
24033 switch (Cond->getOpcode()) {
24034 default: return false;
24041 SetCC0 = Cond->getOperand(0);
24042 SetCC1 = Cond->getOperand(1);
24046 // Make sure we have SETCC nodes, using the same flags value.
24047 if (SetCC0.getOpcode() != X86ISD::SETCC ||
24048 SetCC1.getOpcode() != X86ISD::SETCC ||
24049 SetCC0->getOperand(1) != SetCC1->getOperand(1))
24052 CC0 = (X86::CondCode)SetCC0->getConstantOperandVal(0);
24053 CC1 = (X86::CondCode)SetCC1->getConstantOperandVal(0);
24054 Flags = SetCC0->getOperand(1);
24058 /// Optimize X86ISD::CMOV [LHS, RHS, CONDCODE (e.g. X86::COND_NE), CONDVAL]
24059 static SDValue PerformCMOVCombine(SDNode *N, SelectionDAG &DAG,
24060 TargetLowering::DAGCombinerInfo &DCI,
24061 const X86Subtarget *Subtarget) {
24064 // If the flag operand isn't dead, don't touch this CMOV.
24065 if (N->getNumValues() == 2 && !SDValue(N, 1).use_empty())
24068 SDValue FalseOp = N->getOperand(0);
24069 SDValue TrueOp = N->getOperand(1);
24070 X86::CondCode CC = (X86::CondCode)N->getConstantOperandVal(2);
24071 SDValue Cond = N->getOperand(3);
24073 if (CC == X86::COND_E || CC == X86::COND_NE) {
24074 switch (Cond.getOpcode()) {
24078 // If operand of BSR / BSF are proven never zero, then ZF cannot be set.
24079 if (DAG.isKnownNeverZero(Cond.getOperand(0)))
24080 return (CC == X86::COND_E) ? FalseOp : TrueOp;
24086 Flags = checkBoolTestSetCCCombine(Cond, CC);
24087 if (Flags.getNode() &&
24088 // Extra check as FCMOV only supports a subset of X86 cond.
24089 (FalseOp.getValueType() != MVT::f80 || hasFPCMov(CC))) {
24090 SDValue Ops[] = { FalseOp, TrueOp,
24091 DAG.getConstant(CC, DL, MVT::i8), Flags };
24092 return DAG.getNode(X86ISD::CMOV, DL, N->getVTList(), Ops);
24095 // If this is a select between two integer constants, try to do some
24096 // optimizations. Note that the operands are ordered the opposite of SELECT
24098 if (ConstantSDNode *TrueC = dyn_cast<ConstantSDNode>(TrueOp)) {
24099 if (ConstantSDNode *FalseC = dyn_cast<ConstantSDNode>(FalseOp)) {
24100 // Canonicalize the TrueC/FalseC values so that TrueC (the true value) is
24101 // larger than FalseC (the false value).
24102 if (TrueC->getAPIntValue().ult(FalseC->getAPIntValue())) {
24103 CC = X86::GetOppositeBranchCondition(CC);
24104 std::swap(TrueC, FalseC);
24105 std::swap(TrueOp, FalseOp);
24108 // Optimize C ? 8 : 0 -> zext(setcc(C)) << 3. Likewise for any pow2/0.
24109 // This is efficient for any integer data type (including i8/i16) and
24111 if (FalseC->getAPIntValue() == 0 && TrueC->getAPIntValue().isPowerOf2()) {
24112 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
24113 DAG.getConstant(CC, DL, MVT::i8), Cond);
24115 // Zero extend the condition if needed.
24116 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, TrueC->getValueType(0), Cond);
24118 unsigned ShAmt = TrueC->getAPIntValue().logBase2();
24119 Cond = DAG.getNode(ISD::SHL, DL, Cond.getValueType(), Cond,
24120 DAG.getConstant(ShAmt, DL, MVT::i8));
24121 if (N->getNumValues() == 2) // Dead flag value?
24122 return DCI.CombineTo(N, Cond, SDValue());
24126 // Optimize Cond ? cst+1 : cst -> zext(setcc(C)+cst. This is efficient
24127 // for any integer data type, including i8/i16.
24128 if (FalseC->getAPIntValue()+1 == TrueC->getAPIntValue()) {
24129 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
24130 DAG.getConstant(CC, DL, MVT::i8), Cond);
24132 // Zero extend the condition if needed.
24133 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL,
24134 FalseC->getValueType(0), Cond);
24135 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
24136 SDValue(FalseC, 0));
24138 if (N->getNumValues() == 2) // Dead flag value?
24139 return DCI.CombineTo(N, Cond, SDValue());
24143 // Optimize cases that will turn into an LEA instruction. This requires
24144 // an i32 or i64 and an efficient multiplier (1, 2, 3, 4, 5, 8, 9).
24145 if (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i64) {
24146 uint64_t Diff = TrueC->getZExtValue()-FalseC->getZExtValue();
24147 if (N->getValueType(0) == MVT::i32) Diff = (unsigned)Diff;
24149 bool isFastMultiplier = false;
24151 switch ((unsigned char)Diff) {
24153 case 1: // result = add base, cond
24154 case 2: // result = lea base( , cond*2)
24155 case 3: // result = lea base(cond, cond*2)
24156 case 4: // result = lea base( , cond*4)
24157 case 5: // result = lea base(cond, cond*4)
24158 case 8: // result = lea base( , cond*8)
24159 case 9: // result = lea base(cond, cond*8)
24160 isFastMultiplier = true;
24165 if (isFastMultiplier) {
24166 APInt Diff = TrueC->getAPIntValue()-FalseC->getAPIntValue();
24167 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
24168 DAG.getConstant(CC, DL, MVT::i8), Cond);
24169 // Zero extend the condition if needed.
24170 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, FalseC->getValueType(0),
24172 // Scale the condition by the difference.
24174 Cond = DAG.getNode(ISD::MUL, DL, Cond.getValueType(), Cond,
24175 DAG.getConstant(Diff, DL, Cond.getValueType()));
24177 // Add the base if non-zero.
24178 if (FalseC->getAPIntValue() != 0)
24179 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
24180 SDValue(FalseC, 0));
24181 if (N->getNumValues() == 2) // Dead flag value?
24182 return DCI.CombineTo(N, Cond, SDValue());
24189 // Handle these cases:
24190 // (select (x != c), e, c) -> select (x != c), e, x),
24191 // (select (x == c), c, e) -> select (x == c), x, e)
24192 // where the c is an integer constant, and the "select" is the combination
24193 // of CMOV and CMP.
24195 // The rationale for this change is that the conditional-move from a constant
24196 // needs two instructions, however, conditional-move from a register needs
24197 // only one instruction.
24199 // CAVEAT: By replacing a constant with a symbolic value, it may obscure
24200 // some instruction-combining opportunities. This opt needs to be
24201 // postponed as late as possible.
24203 if (!DCI.isBeforeLegalize() && !DCI.isBeforeLegalizeOps()) {
24204 // the DCI.xxxx conditions are provided to postpone the optimization as
24205 // late as possible.
24207 ConstantSDNode *CmpAgainst = nullptr;
24208 if ((Cond.getOpcode() == X86ISD::CMP || Cond.getOpcode() == X86ISD::SUB) &&
24209 (CmpAgainst = dyn_cast<ConstantSDNode>(Cond.getOperand(1))) &&
24210 !isa<ConstantSDNode>(Cond.getOperand(0))) {
24212 if (CC == X86::COND_NE &&
24213 CmpAgainst == dyn_cast<ConstantSDNode>(FalseOp)) {
24214 CC = X86::GetOppositeBranchCondition(CC);
24215 std::swap(TrueOp, FalseOp);
24218 if (CC == X86::COND_E &&
24219 CmpAgainst == dyn_cast<ConstantSDNode>(TrueOp)) {
24220 SDValue Ops[] = { FalseOp, Cond.getOperand(0),
24221 DAG.getConstant(CC, DL, MVT::i8), Cond };
24222 return DAG.getNode(X86ISD::CMOV, DL, N->getVTList (), Ops);
24227 // Fold and/or of setcc's to double CMOV:
24228 // (CMOV F, T, ((cc1 | cc2) != 0)) -> (CMOV (CMOV F, T, cc1), T, cc2)
24229 // (CMOV F, T, ((cc1 & cc2) != 0)) -> (CMOV (CMOV T, F, !cc1), F, !cc2)
24231 // This combine lets us generate:
24232 // cmovcc1 (jcc1 if we don't have CMOV)
24238 // cmovne (jne if we don't have CMOV)
24239 // When we can't use the CMOV instruction, it might increase branch
24241 // When we can use CMOV, or when there is no mispredict, this improves
24242 // throughput and reduces register pressure.
24244 if (CC == X86::COND_NE) {
24246 X86::CondCode CC0, CC1;
24248 if (checkBoolTestAndOrSetCCCombine(Cond, CC0, CC1, Flags, isAndSetCC)) {
24250 std::swap(FalseOp, TrueOp);
24251 CC0 = X86::GetOppositeBranchCondition(CC0);
24252 CC1 = X86::GetOppositeBranchCondition(CC1);
24255 SDValue LOps[] = {FalseOp, TrueOp, DAG.getConstant(CC0, DL, MVT::i8),
24257 SDValue LCMOV = DAG.getNode(X86ISD::CMOV, DL, N->getVTList(), LOps);
24258 SDValue Ops[] = {LCMOV, TrueOp, DAG.getConstant(CC1, DL, MVT::i8), Flags};
24259 SDValue CMOV = DAG.getNode(X86ISD::CMOV, DL, N->getVTList(), Ops);
24260 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), SDValue(CMOV.getNode(), 1));
24268 /// PerformMulCombine - Optimize a single multiply with constant into two
24269 /// in order to implement it with two cheaper instructions, e.g.
24270 /// LEA + SHL, LEA + LEA.
24271 static SDValue PerformMulCombine(SDNode *N, SelectionDAG &DAG,
24272 TargetLowering::DAGCombinerInfo &DCI) {
24273 // An imul is usually smaller than the alternative sequence.
24274 if (DAG.getMachineFunction().getFunction()->optForMinSize())
24277 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
24280 EVT VT = N->getValueType(0);
24281 if (VT != MVT::i64 && VT != MVT::i32)
24284 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1));
24287 uint64_t MulAmt = C->getZExtValue();
24288 if (isPowerOf2_64(MulAmt) || MulAmt == 3 || MulAmt == 5 || MulAmt == 9)
24291 uint64_t MulAmt1 = 0;
24292 uint64_t MulAmt2 = 0;
24293 if ((MulAmt % 9) == 0) {
24295 MulAmt2 = MulAmt / 9;
24296 } else if ((MulAmt % 5) == 0) {
24298 MulAmt2 = MulAmt / 5;
24299 } else if ((MulAmt % 3) == 0) {
24301 MulAmt2 = MulAmt / 3;
24304 (isPowerOf2_64(MulAmt2) || MulAmt2 == 3 || MulAmt2 == 5 || MulAmt2 == 9)){
24307 if (isPowerOf2_64(MulAmt2) &&
24308 !(N->hasOneUse() && N->use_begin()->getOpcode() == ISD::ADD))
24309 // If second multiplifer is pow2, issue it first. We want the multiply by
24310 // 3, 5, or 9 to be folded into the addressing mode unless the lone use
24312 std::swap(MulAmt1, MulAmt2);
24315 if (isPowerOf2_64(MulAmt1))
24316 NewMul = DAG.getNode(ISD::SHL, DL, VT, N->getOperand(0),
24317 DAG.getConstant(Log2_64(MulAmt1), DL, MVT::i8));
24319 NewMul = DAG.getNode(X86ISD::MUL_IMM, DL, VT, N->getOperand(0),
24320 DAG.getConstant(MulAmt1, DL, VT));
24322 if (isPowerOf2_64(MulAmt2))
24323 NewMul = DAG.getNode(ISD::SHL, DL, VT, NewMul,
24324 DAG.getConstant(Log2_64(MulAmt2), DL, MVT::i8));
24326 NewMul = DAG.getNode(X86ISD::MUL_IMM, DL, VT, NewMul,
24327 DAG.getConstant(MulAmt2, DL, VT));
24329 // Do not add new nodes to DAG combiner worklist.
24330 DCI.CombineTo(N, NewMul, false);
24335 static SDValue PerformSHLCombine(SDNode *N, SelectionDAG &DAG) {
24336 SDValue N0 = N->getOperand(0);
24337 SDValue N1 = N->getOperand(1);
24338 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
24339 EVT VT = N0.getValueType();
24341 // fold (shl (and (setcc_c), c1), c2) -> (and setcc_c, (c1 << c2))
24342 // since the result of setcc_c is all zero's or all ones.
24343 if (VT.isInteger() && !VT.isVector() &&
24344 N1C && N0.getOpcode() == ISD::AND &&
24345 N0.getOperand(1).getOpcode() == ISD::Constant) {
24346 SDValue N00 = N0.getOperand(0);
24347 APInt Mask = cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
24348 APInt ShAmt = N1C->getAPIntValue();
24349 Mask = Mask.shl(ShAmt);
24350 bool MaskOK = false;
24351 // We can handle cases concerning bit-widening nodes containing setcc_c if
24352 // we carefully interrogate the mask to make sure we are semantics
24354 // The transform is not safe if the result of C1 << C2 exceeds the bitwidth
24355 // of the underlying setcc_c operation if the setcc_c was zero extended.
24356 // Consider the following example:
24357 // zext(setcc_c) -> i32 0x0000FFFF
24358 // c1 -> i32 0x0000FFFF
24359 // c2 -> i32 0x00000001
24360 // (shl (and (setcc_c), c1), c2) -> i32 0x0001FFFE
24361 // (and setcc_c, (c1 << c2)) -> i32 0x0000FFFE
24362 if (N00.getOpcode() == X86ISD::SETCC_CARRY) {
24364 } else if (N00.getOpcode() == ISD::SIGN_EXTEND &&
24365 N00.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY) {
24367 } else if ((N00.getOpcode() == ISD::ZERO_EXTEND ||
24368 N00.getOpcode() == ISD::ANY_EXTEND) &&
24369 N00.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY) {
24370 MaskOK = Mask.isIntN(N00.getOperand(0).getValueSizeInBits());
24372 if (MaskOK && Mask != 0) {
24374 return DAG.getNode(ISD::AND, DL, VT, N00, DAG.getConstant(Mask, DL, VT));
24378 // Hardware support for vector shifts is sparse which makes us scalarize the
24379 // vector operations in many cases. Also, on sandybridge ADD is faster than
24381 // (shl V, 1) -> add V,V
24382 if (auto *N1BV = dyn_cast<BuildVectorSDNode>(N1))
24383 if (auto *N1SplatC = N1BV->getConstantSplatNode()) {
24384 assert(N0.getValueType().isVector() && "Invalid vector shift type");
24385 // We shift all of the values by one. In many cases we do not have
24386 // hardware support for this operation. This is better expressed as an ADD
24388 if (N1SplatC->getAPIntValue() == 1)
24389 return DAG.getNode(ISD::ADD, SDLoc(N), VT, N0, N0);
24395 /// \brief Returns a vector of 0s if the node in input is a vector logical
24396 /// shift by a constant amount which is known to be bigger than or equal
24397 /// to the vector element size in bits.
24398 static SDValue performShiftToAllZeros(SDNode *N, SelectionDAG &DAG,
24399 const X86Subtarget *Subtarget) {
24400 EVT VT = N->getValueType(0);
24402 if (VT != MVT::v2i64 && VT != MVT::v4i32 && VT != MVT::v8i16 &&
24403 (!Subtarget->hasInt256() ||
24404 (VT != MVT::v4i64 && VT != MVT::v8i32 && VT != MVT::v16i16)))
24407 SDValue Amt = N->getOperand(1);
24409 if (auto *AmtBV = dyn_cast<BuildVectorSDNode>(Amt))
24410 if (auto *AmtSplat = AmtBV->getConstantSplatNode()) {
24411 APInt ShiftAmt = AmtSplat->getAPIntValue();
24412 unsigned MaxAmount = VT.getVectorElementType().getSizeInBits();
24414 // SSE2/AVX2 logical shifts always return a vector of 0s
24415 // if the shift amount is bigger than or equal to
24416 // the element size. The constant shift amount will be
24417 // encoded as a 8-bit immediate.
24418 if (ShiftAmt.trunc(8).uge(MaxAmount))
24419 return getZeroVector(VT, Subtarget, DAG, DL);
24425 /// PerformShiftCombine - Combine shifts.
24426 static SDValue PerformShiftCombine(SDNode* N, SelectionDAG &DAG,
24427 TargetLowering::DAGCombinerInfo &DCI,
24428 const X86Subtarget *Subtarget) {
24429 if (N->getOpcode() == ISD::SHL)
24430 if (SDValue V = PerformSHLCombine(N, DAG))
24433 // Try to fold this logical shift into a zero vector.
24434 if (N->getOpcode() != ISD::SRA)
24435 if (SDValue V = performShiftToAllZeros(N, DAG, Subtarget))
24441 // CMPEQCombine - Recognize the distinctive (AND (setcc ...) (setcc ..))
24442 // where both setccs reference the same FP CMP, and rewrite for CMPEQSS
24443 // and friends. Likewise for OR -> CMPNEQSS.
24444 static SDValue CMPEQCombine(SDNode *N, SelectionDAG &DAG,
24445 TargetLowering::DAGCombinerInfo &DCI,
24446 const X86Subtarget *Subtarget) {
24449 // SSE1 supports CMP{eq|ne}SS, and SSE2 added CMP{eq|ne}SD, but
24450 // we're requiring SSE2 for both.
24451 if (Subtarget->hasSSE2() && isAndOrOfSetCCs(SDValue(N, 0U), opcode)) {
24452 SDValue N0 = N->getOperand(0);
24453 SDValue N1 = N->getOperand(1);
24454 SDValue CMP0 = N0->getOperand(1);
24455 SDValue CMP1 = N1->getOperand(1);
24458 // The SETCCs should both refer to the same CMP.
24459 if (CMP0.getOpcode() != X86ISD::CMP || CMP0 != CMP1)
24462 SDValue CMP00 = CMP0->getOperand(0);
24463 SDValue CMP01 = CMP0->getOperand(1);
24464 EVT VT = CMP00.getValueType();
24466 if (VT == MVT::f32 || VT == MVT::f64) {
24467 bool ExpectingFlags = false;
24468 // Check for any users that want flags:
24469 for (SDNode::use_iterator UI = N->use_begin(), UE = N->use_end();
24470 !ExpectingFlags && UI != UE; ++UI)
24471 switch (UI->getOpcode()) {
24476 ExpectingFlags = true;
24478 case ISD::CopyToReg:
24479 case ISD::SIGN_EXTEND:
24480 case ISD::ZERO_EXTEND:
24481 case ISD::ANY_EXTEND:
24485 if (!ExpectingFlags) {
24486 enum X86::CondCode cc0 = (enum X86::CondCode)N0.getConstantOperandVal(0);
24487 enum X86::CondCode cc1 = (enum X86::CondCode)N1.getConstantOperandVal(0);
24489 if (cc1 == X86::COND_E || cc1 == X86::COND_NE) {
24490 X86::CondCode tmp = cc0;
24495 if ((cc0 == X86::COND_E && cc1 == X86::COND_NP) ||
24496 (cc0 == X86::COND_NE && cc1 == X86::COND_P)) {
24497 // FIXME: need symbolic constants for these magic numbers.
24498 // See X86ATTInstPrinter.cpp:printSSECC().
24499 unsigned x86cc = (cc0 == X86::COND_E) ? 0 : 4;
24500 if (Subtarget->hasAVX512()) {
24501 SDValue FSetCC = DAG.getNode(X86ISD::FSETCC, DL, MVT::i1, CMP00,
24503 DAG.getConstant(x86cc, DL, MVT::i8));
24504 if (N->getValueType(0) != MVT::i1)
24505 return DAG.getNode(ISD::ZERO_EXTEND, DL, N->getValueType(0),
24509 SDValue OnesOrZeroesF = DAG.getNode(X86ISD::FSETCC, DL,
24510 CMP00.getValueType(), CMP00, CMP01,
24511 DAG.getConstant(x86cc, DL,
24514 bool is64BitFP = (CMP00.getValueType() == MVT::f64);
24515 MVT IntVT = is64BitFP ? MVT::i64 : MVT::i32;
24517 if (is64BitFP && !Subtarget->is64Bit()) {
24518 // On a 32-bit target, we cannot bitcast the 64-bit float to a
24519 // 64-bit integer, since that's not a legal type. Since
24520 // OnesOrZeroesF is all ones of all zeroes, we don't need all the
24521 // bits, but can do this little dance to extract the lowest 32 bits
24522 // and work with those going forward.
24523 SDValue Vector64 = DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, MVT::v2f64,
24525 SDValue Vector32 = DAG.getBitcast(MVT::v4f32, Vector64);
24526 OnesOrZeroesF = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::f32,
24527 Vector32, DAG.getIntPtrConstant(0, DL));
24531 SDValue OnesOrZeroesI = DAG.getBitcast(IntVT, OnesOrZeroesF);
24532 SDValue ANDed = DAG.getNode(ISD::AND, DL, IntVT, OnesOrZeroesI,
24533 DAG.getConstant(1, DL, IntVT));
24534 SDValue OneBitOfTruth = DAG.getNode(ISD::TRUNCATE, DL, MVT::i8,
24536 return OneBitOfTruth;
24544 /// CanFoldXORWithAllOnes - Test whether the XOR operand is a AllOnes vector
24545 /// so it can be folded inside ANDNP.
24546 static bool CanFoldXORWithAllOnes(const SDNode *N) {
24547 EVT VT = N->getValueType(0);
24549 // Match direct AllOnes for 128 and 256-bit vectors
24550 if (ISD::isBuildVectorAllOnes(N))
24553 // Look through a bit convert.
24554 if (N->getOpcode() == ISD::BITCAST)
24555 N = N->getOperand(0).getNode();
24557 // Sometimes the operand may come from a insert_subvector building a 256-bit
24559 if (VT.is256BitVector() &&
24560 N->getOpcode() == ISD::INSERT_SUBVECTOR) {
24561 SDValue V1 = N->getOperand(0);
24562 SDValue V2 = N->getOperand(1);
24564 if (V1.getOpcode() == ISD::INSERT_SUBVECTOR &&
24565 V1.getOperand(0).getOpcode() == ISD::UNDEF &&
24566 ISD::isBuildVectorAllOnes(V1.getOperand(1).getNode()) &&
24567 ISD::isBuildVectorAllOnes(V2.getNode()))
24574 // On AVX/AVX2 the type v8i1 is legalized to v8i16, which is an XMM sized
24575 // register. In most cases we actually compare or select YMM-sized registers
24576 // and mixing the two types creates horrible code. This method optimizes
24577 // some of the transition sequences.
24578 static SDValue WidenMaskArithmetic(SDNode *N, SelectionDAG &DAG,
24579 TargetLowering::DAGCombinerInfo &DCI,
24580 const X86Subtarget *Subtarget) {
24581 EVT VT = N->getValueType(0);
24582 if (!VT.is256BitVector())
24585 assert((N->getOpcode() == ISD::ANY_EXTEND ||
24586 N->getOpcode() == ISD::ZERO_EXTEND ||
24587 N->getOpcode() == ISD::SIGN_EXTEND) && "Invalid Node");
24589 SDValue Narrow = N->getOperand(0);
24590 EVT NarrowVT = Narrow->getValueType(0);
24591 if (!NarrowVT.is128BitVector())
24594 if (Narrow->getOpcode() != ISD::XOR &&
24595 Narrow->getOpcode() != ISD::AND &&
24596 Narrow->getOpcode() != ISD::OR)
24599 SDValue N0 = Narrow->getOperand(0);
24600 SDValue N1 = Narrow->getOperand(1);
24603 // The Left side has to be a trunc.
24604 if (N0.getOpcode() != ISD::TRUNCATE)
24607 // The type of the truncated inputs.
24608 EVT WideVT = N0->getOperand(0)->getValueType(0);
24612 // The right side has to be a 'trunc' or a constant vector.
24613 bool RHSTrunc = N1.getOpcode() == ISD::TRUNCATE;
24614 ConstantSDNode *RHSConstSplat = nullptr;
24615 if (auto *RHSBV = dyn_cast<BuildVectorSDNode>(N1))
24616 RHSConstSplat = RHSBV->getConstantSplatNode();
24617 if (!RHSTrunc && !RHSConstSplat)
24620 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
24622 if (!TLI.isOperationLegalOrPromote(Narrow->getOpcode(), WideVT))
24625 // Set N0 and N1 to hold the inputs to the new wide operation.
24626 N0 = N0->getOperand(0);
24627 if (RHSConstSplat) {
24628 N1 = DAG.getNode(ISD::ZERO_EXTEND, DL, WideVT.getScalarType(),
24629 SDValue(RHSConstSplat, 0));
24630 SmallVector<SDValue, 8> C(WideVT.getVectorNumElements(), N1);
24631 N1 = DAG.getNode(ISD::BUILD_VECTOR, DL, WideVT, C);
24632 } else if (RHSTrunc) {
24633 N1 = N1->getOperand(0);
24636 // Generate the wide operation.
24637 SDValue Op = DAG.getNode(Narrow->getOpcode(), DL, WideVT, N0, N1);
24638 unsigned Opcode = N->getOpcode();
24640 case ISD::ANY_EXTEND:
24642 case ISD::ZERO_EXTEND: {
24643 unsigned InBits = NarrowVT.getScalarType().getSizeInBits();
24644 APInt Mask = APInt::getAllOnesValue(InBits);
24645 Mask = Mask.zext(VT.getScalarType().getSizeInBits());
24646 return DAG.getNode(ISD::AND, DL, VT,
24647 Op, DAG.getConstant(Mask, DL, VT));
24649 case ISD::SIGN_EXTEND:
24650 return DAG.getNode(ISD::SIGN_EXTEND_INREG, DL, VT,
24651 Op, DAG.getValueType(NarrowVT));
24653 llvm_unreachable("Unexpected opcode");
24657 static SDValue VectorZextCombine(SDNode *N, SelectionDAG &DAG,
24658 TargetLowering::DAGCombinerInfo &DCI,
24659 const X86Subtarget *Subtarget) {
24660 SDValue N0 = N->getOperand(0);
24661 SDValue N1 = N->getOperand(1);
24664 // A vector zext_in_reg may be represented as a shuffle,
24665 // feeding into a bitcast (this represents anyext) feeding into
24666 // an and with a mask.
24667 // We'd like to try to combine that into a shuffle with zero
24668 // plus a bitcast, removing the and.
24669 if (N0.getOpcode() != ISD::BITCAST ||
24670 N0.getOperand(0).getOpcode() != ISD::VECTOR_SHUFFLE)
24673 // The other side of the AND should be a splat of 2^C, where C
24674 // is the number of bits in the source type.
24675 if (N1.getOpcode() == ISD::BITCAST)
24676 N1 = N1.getOperand(0);
24677 if (N1.getOpcode() != ISD::BUILD_VECTOR)
24679 BuildVectorSDNode *Vector = cast<BuildVectorSDNode>(N1);
24681 ShuffleVectorSDNode *Shuffle = cast<ShuffleVectorSDNode>(N0.getOperand(0));
24682 EVT SrcType = Shuffle->getValueType(0);
24684 // We expect a single-source shuffle
24685 if (Shuffle->getOperand(1)->getOpcode() != ISD::UNDEF)
24688 unsigned SrcSize = SrcType.getScalarSizeInBits();
24690 APInt SplatValue, SplatUndef;
24691 unsigned SplatBitSize;
24693 if (!Vector->isConstantSplat(SplatValue, SplatUndef,
24694 SplatBitSize, HasAnyUndefs))
24697 unsigned ResSize = N1.getValueType().getScalarSizeInBits();
24698 // Make sure the splat matches the mask we expect
24699 if (SplatBitSize > ResSize ||
24700 (SplatValue + 1).exactLogBase2() != (int)SrcSize)
24703 // Make sure the input and output size make sense
24704 if (SrcSize >= ResSize || ResSize % SrcSize)
24707 // We expect a shuffle of the form <0, u, u, u, 1, u, u, u...>
24708 // The number of u's between each two values depends on the ratio between
24709 // the source and dest type.
24710 unsigned ZextRatio = ResSize / SrcSize;
24711 bool IsZext = true;
24712 for (unsigned i = 0; i < SrcType.getVectorNumElements(); ++i) {
24713 if (i % ZextRatio) {
24714 if (Shuffle->getMaskElt(i) > 0) {
24720 if (Shuffle->getMaskElt(i) != (int)(i / ZextRatio)) {
24721 // Expected element number
24731 // Ok, perform the transformation - replace the shuffle with
24732 // a shuffle of the form <0, k, k, k, 1, k, k, k> with zero
24733 // (instead of undef) where the k elements come from the zero vector.
24734 SmallVector<int, 8> Mask;
24735 unsigned NumElems = SrcType.getVectorNumElements();
24736 for (unsigned i = 0; i < NumElems; ++i)
24738 Mask.push_back(NumElems);
24740 Mask.push_back(i / ZextRatio);
24742 SDValue NewShuffle = DAG.getVectorShuffle(Shuffle->getValueType(0), DL,
24743 Shuffle->getOperand(0), DAG.getConstant(0, DL, SrcType), Mask);
24744 return DAG.getBitcast(N0.getValueType(), NewShuffle);
24747 /// If both input operands of a logic op are being cast from floating point
24748 /// types, try to convert this into a floating point logic node to avoid
24749 /// unnecessary moves from SSE to integer registers.
24750 static SDValue convertIntLogicToFPLogic(SDNode *N, SelectionDAG &DAG,
24751 const X86Subtarget *Subtarget) {
24752 unsigned FPOpcode = ISD::DELETED_NODE;
24753 if (N->getOpcode() == ISD::AND)
24754 FPOpcode = X86ISD::FAND;
24755 else if (N->getOpcode() == ISD::OR)
24756 FPOpcode = X86ISD::FOR;
24757 else if (N->getOpcode() == ISD::XOR)
24758 FPOpcode = X86ISD::FXOR;
24760 assert(FPOpcode != ISD::DELETED_NODE &&
24761 "Unexpected input node for FP logic conversion");
24763 EVT VT = N->getValueType(0);
24764 SDValue N0 = N->getOperand(0);
24765 SDValue N1 = N->getOperand(1);
24767 if (N0.getOpcode() == ISD::BITCAST && N1.getOpcode() == ISD::BITCAST &&
24768 ((Subtarget->hasSSE1() && VT == MVT::i32) ||
24769 (Subtarget->hasSSE2() && VT == MVT::i64))) {
24770 SDValue N00 = N0.getOperand(0);
24771 SDValue N10 = N1.getOperand(0);
24772 EVT N00Type = N00.getValueType();
24773 EVT N10Type = N10.getValueType();
24774 if (N00Type.isFloatingPoint() && N10Type.isFloatingPoint()) {
24775 SDValue FPLogic = DAG.getNode(FPOpcode, DL, N00Type, N00, N10);
24776 return DAG.getBitcast(VT, FPLogic);
24782 static SDValue PerformAndCombine(SDNode *N, SelectionDAG &DAG,
24783 TargetLowering::DAGCombinerInfo &DCI,
24784 const X86Subtarget *Subtarget) {
24785 if (DCI.isBeforeLegalizeOps())
24788 if (SDValue Zext = VectorZextCombine(N, DAG, DCI, Subtarget))
24791 if (SDValue R = CMPEQCombine(N, DAG, DCI, Subtarget))
24794 if (SDValue FPLogic = convertIntLogicToFPLogic(N, DAG, Subtarget))
24797 EVT VT = N->getValueType(0);
24798 SDValue N0 = N->getOperand(0);
24799 SDValue N1 = N->getOperand(1);
24802 // Create BEXTR instructions
24803 // BEXTR is ((X >> imm) & (2**size-1))
24804 if (VT == MVT::i32 || VT == MVT::i64) {
24805 // Check for BEXTR.
24806 if ((Subtarget->hasBMI() || Subtarget->hasTBM()) &&
24807 (N0.getOpcode() == ISD::SRA || N0.getOpcode() == ISD::SRL)) {
24808 ConstantSDNode *MaskNode = dyn_cast<ConstantSDNode>(N1);
24809 ConstantSDNode *ShiftNode = dyn_cast<ConstantSDNode>(N0.getOperand(1));
24810 if (MaskNode && ShiftNode) {
24811 uint64_t Mask = MaskNode->getZExtValue();
24812 uint64_t Shift = ShiftNode->getZExtValue();
24813 if (isMask_64(Mask)) {
24814 uint64_t MaskSize = countPopulation(Mask);
24815 if (Shift + MaskSize <= VT.getSizeInBits())
24816 return DAG.getNode(X86ISD::BEXTR, DL, VT, N0.getOperand(0),
24817 DAG.getConstant(Shift | (MaskSize << 8), DL,
24826 // Want to form ANDNP nodes:
24827 // 1) In the hopes of then easily combining them with OR and AND nodes
24828 // to form PBLEND/PSIGN.
24829 // 2) To match ANDN packed intrinsics
24830 if (VT != MVT::v2i64 && VT != MVT::v4i64)
24833 // Check LHS for vnot
24834 if (N0.getOpcode() == ISD::XOR &&
24835 //ISD::isBuildVectorAllOnes(N0.getOperand(1).getNode()))
24836 CanFoldXORWithAllOnes(N0.getOperand(1).getNode()))
24837 return DAG.getNode(X86ISD::ANDNP, DL, VT, N0.getOperand(0), N1);
24839 // Check RHS for vnot
24840 if (N1.getOpcode() == ISD::XOR &&
24841 //ISD::isBuildVectorAllOnes(N1.getOperand(1).getNode()))
24842 CanFoldXORWithAllOnes(N1.getOperand(1).getNode()))
24843 return DAG.getNode(X86ISD::ANDNP, DL, VT, N1.getOperand(0), N0);
24848 static SDValue PerformOrCombine(SDNode *N, SelectionDAG &DAG,
24849 TargetLowering::DAGCombinerInfo &DCI,
24850 const X86Subtarget *Subtarget) {
24851 if (DCI.isBeforeLegalizeOps())
24854 if (SDValue R = CMPEQCombine(N, DAG, DCI, Subtarget))
24857 if (SDValue FPLogic = convertIntLogicToFPLogic(N, DAG, Subtarget))
24860 SDValue N0 = N->getOperand(0);
24861 SDValue N1 = N->getOperand(1);
24862 EVT VT = N->getValueType(0);
24864 // look for psign/blend
24865 if (VT == MVT::v2i64 || VT == MVT::v4i64) {
24866 if (!Subtarget->hasSSSE3() ||
24867 (VT == MVT::v4i64 && !Subtarget->hasInt256()))
24870 // Canonicalize pandn to RHS
24871 if (N0.getOpcode() == X86ISD::ANDNP)
24873 // or (and (m, y), (pandn m, x))
24874 if (N0.getOpcode() == ISD::AND && N1.getOpcode() == X86ISD::ANDNP) {
24875 SDValue Mask = N1.getOperand(0);
24876 SDValue X = N1.getOperand(1);
24878 if (N0.getOperand(0) == Mask)
24879 Y = N0.getOperand(1);
24880 if (N0.getOperand(1) == Mask)
24881 Y = N0.getOperand(0);
24883 // Check to see if the mask appeared in both the AND and ANDNP and
24887 // Validate that X, Y, and Mask are BIT_CONVERTS, and see through them.
24888 // Look through mask bitcast.
24889 if (Mask.getOpcode() == ISD::BITCAST)
24890 Mask = Mask.getOperand(0);
24891 if (X.getOpcode() == ISD::BITCAST)
24892 X = X.getOperand(0);
24893 if (Y.getOpcode() == ISD::BITCAST)
24894 Y = Y.getOperand(0);
24896 EVT MaskVT = Mask.getValueType();
24898 // Validate that the Mask operand is a vector sra node.
24899 // FIXME: what to do for bytes, since there is a psignb/pblendvb, but
24900 // there is no psrai.b
24901 unsigned EltBits = MaskVT.getVectorElementType().getSizeInBits();
24902 unsigned SraAmt = ~0;
24903 if (Mask.getOpcode() == ISD::SRA) {
24904 if (auto *AmtBV = dyn_cast<BuildVectorSDNode>(Mask.getOperand(1)))
24905 if (auto *AmtConst = AmtBV->getConstantSplatNode())
24906 SraAmt = AmtConst->getZExtValue();
24907 } else if (Mask.getOpcode() == X86ISD::VSRAI) {
24908 SDValue SraC = Mask.getOperand(1);
24909 SraAmt = cast<ConstantSDNode>(SraC)->getZExtValue();
24911 if ((SraAmt + 1) != EltBits)
24916 // Now we know we at least have a plendvb with the mask val. See if
24917 // we can form a psignb/w/d.
24918 // psign = x.type == y.type == mask.type && y = sub(0, x);
24919 if (Y.getOpcode() == ISD::SUB && Y.getOperand(1) == X &&
24920 ISD::isBuildVectorAllZeros(Y.getOperand(0).getNode()) &&
24921 X.getValueType() == MaskVT && Y.getValueType() == MaskVT) {
24922 assert((EltBits == 8 || EltBits == 16 || EltBits == 32) &&
24923 "Unsupported VT for PSIGN");
24924 Mask = DAG.getNode(X86ISD::PSIGN, DL, MaskVT, X, Mask.getOperand(0));
24925 return DAG.getBitcast(VT, Mask);
24927 // PBLENDVB only available on SSE 4.1
24928 if (!Subtarget->hasSSE41())
24931 EVT BlendVT = (VT == MVT::v4i64) ? MVT::v32i8 : MVT::v16i8;
24933 X = DAG.getBitcast(BlendVT, X);
24934 Y = DAG.getBitcast(BlendVT, Y);
24935 Mask = DAG.getBitcast(BlendVT, Mask);
24936 Mask = DAG.getNode(ISD::VSELECT, DL, BlendVT, Mask, Y, X);
24937 return DAG.getBitcast(VT, Mask);
24941 if (VT != MVT::i16 && VT != MVT::i32 && VT != MVT::i64)
24944 // fold (or (x << c) | (y >> (64 - c))) ==> (shld64 x, y, c)
24945 bool OptForSize = DAG.getMachineFunction().getFunction()->optForSize();
24947 // SHLD/SHRD instructions have lower register pressure, but on some
24948 // platforms they have higher latency than the equivalent
24949 // series of shifts/or that would otherwise be generated.
24950 // Don't fold (or (x << c) | (y >> (64 - c))) if SHLD/SHRD instructions
24951 // have higher latencies and we are not optimizing for size.
24952 if (!OptForSize && Subtarget->isSHLDSlow())
24955 if (N0.getOpcode() == ISD::SRL && N1.getOpcode() == ISD::SHL)
24957 if (N0.getOpcode() != ISD::SHL || N1.getOpcode() != ISD::SRL)
24959 if (!N0.hasOneUse() || !N1.hasOneUse())
24962 SDValue ShAmt0 = N0.getOperand(1);
24963 if (ShAmt0.getValueType() != MVT::i8)
24965 SDValue ShAmt1 = N1.getOperand(1);
24966 if (ShAmt1.getValueType() != MVT::i8)
24968 if (ShAmt0.getOpcode() == ISD::TRUNCATE)
24969 ShAmt0 = ShAmt0.getOperand(0);
24970 if (ShAmt1.getOpcode() == ISD::TRUNCATE)
24971 ShAmt1 = ShAmt1.getOperand(0);
24974 unsigned Opc = X86ISD::SHLD;
24975 SDValue Op0 = N0.getOperand(0);
24976 SDValue Op1 = N1.getOperand(0);
24977 if (ShAmt0.getOpcode() == ISD::SUB) {
24978 Opc = X86ISD::SHRD;
24979 std::swap(Op0, Op1);
24980 std::swap(ShAmt0, ShAmt1);
24983 unsigned Bits = VT.getSizeInBits();
24984 if (ShAmt1.getOpcode() == ISD::SUB) {
24985 SDValue Sum = ShAmt1.getOperand(0);
24986 if (ConstantSDNode *SumC = dyn_cast<ConstantSDNode>(Sum)) {
24987 SDValue ShAmt1Op1 = ShAmt1.getOperand(1);
24988 if (ShAmt1Op1.getNode()->getOpcode() == ISD::TRUNCATE)
24989 ShAmt1Op1 = ShAmt1Op1.getOperand(0);
24990 if (SumC->getSExtValue() == Bits && ShAmt1Op1 == ShAmt0)
24991 return DAG.getNode(Opc, DL, VT,
24993 DAG.getNode(ISD::TRUNCATE, DL,
24996 } else if (ConstantSDNode *ShAmt1C = dyn_cast<ConstantSDNode>(ShAmt1)) {
24997 ConstantSDNode *ShAmt0C = dyn_cast<ConstantSDNode>(ShAmt0);
24999 ShAmt0C->getSExtValue() + ShAmt1C->getSExtValue() == Bits)
25000 return DAG.getNode(Opc, DL, VT,
25001 N0.getOperand(0), N1.getOperand(0),
25002 DAG.getNode(ISD::TRUNCATE, DL,
25009 // Generate NEG and CMOV for integer abs.
25010 static SDValue performIntegerAbsCombine(SDNode *N, SelectionDAG &DAG) {
25011 EVT VT = N->getValueType(0);
25013 // Since X86 does not have CMOV for 8-bit integer, we don't convert
25014 // 8-bit integer abs to NEG and CMOV.
25015 if (VT.isInteger() && VT.getSizeInBits() == 8)
25018 SDValue N0 = N->getOperand(0);
25019 SDValue N1 = N->getOperand(1);
25022 // Check pattern of XOR(ADD(X,Y), Y) where Y is SRA(X, size(X)-1)
25023 // and change it to SUB and CMOV.
25024 if (VT.isInteger() && N->getOpcode() == ISD::XOR &&
25025 N0.getOpcode() == ISD::ADD &&
25026 N0.getOperand(1) == N1 &&
25027 N1.getOpcode() == ISD::SRA &&
25028 N1.getOperand(0) == N0.getOperand(0))
25029 if (ConstantSDNode *Y1C = dyn_cast<ConstantSDNode>(N1.getOperand(1)))
25030 if (Y1C->getAPIntValue() == VT.getSizeInBits()-1) {
25031 // Generate SUB & CMOV.
25032 SDValue Neg = DAG.getNode(X86ISD::SUB, DL, DAG.getVTList(VT, MVT::i32),
25033 DAG.getConstant(0, DL, VT), N0.getOperand(0));
25035 SDValue Ops[] = { N0.getOperand(0), Neg,
25036 DAG.getConstant(X86::COND_GE, DL, MVT::i8),
25037 SDValue(Neg.getNode(), 1) };
25038 return DAG.getNode(X86ISD::CMOV, DL, DAG.getVTList(VT, MVT::Glue), Ops);
25043 // Try to turn tests against the signbit in the form of:
25044 // XOR(TRUNCATE(SRL(X, size(X)-1)), 1)
25047 static SDValue foldXorTruncShiftIntoCmp(SDNode *N, SelectionDAG &DAG) {
25048 // This is only worth doing if the output type is i8.
25049 if (N->getValueType(0) != MVT::i8)
25052 SDValue N0 = N->getOperand(0);
25053 SDValue N1 = N->getOperand(1);
25055 // We should be performing an xor against a truncated shift.
25056 if (N0.getOpcode() != ISD::TRUNCATE || !N0.hasOneUse())
25059 // Make sure we are performing an xor against one.
25060 if (!isa<ConstantSDNode>(N1) || !cast<ConstantSDNode>(N1)->isOne())
25063 // SetCC on x86 zero extends so only act on this if it's a logical shift.
25064 SDValue Shift = N0.getOperand(0);
25065 if (Shift.getOpcode() != ISD::SRL || !Shift.hasOneUse())
25068 // Make sure we are truncating from one of i16, i32 or i64.
25069 EVT ShiftTy = Shift.getValueType();
25070 if (ShiftTy != MVT::i16 && ShiftTy != MVT::i32 && ShiftTy != MVT::i64)
25073 // Make sure the shift amount extracts the sign bit.
25074 if (!isa<ConstantSDNode>(Shift.getOperand(1)) ||
25075 Shift.getConstantOperandVal(1) != ShiftTy.getSizeInBits() - 1)
25078 // Create a greater-than comparison against -1.
25079 // N.B. Using SETGE against 0 works but we want a canonical looking
25080 // comparison, using SETGT matches up with what TranslateX86CC.
25082 SDValue ShiftOp = Shift.getOperand(0);
25083 EVT ShiftOpTy = ShiftOp.getValueType();
25084 SDValue Cond = DAG.getSetCC(DL, MVT::i8, ShiftOp,
25085 DAG.getConstant(-1, DL, ShiftOpTy), ISD::SETGT);
25089 static SDValue PerformXorCombine(SDNode *N, SelectionDAG &DAG,
25090 TargetLowering::DAGCombinerInfo &DCI,
25091 const X86Subtarget *Subtarget) {
25092 if (DCI.isBeforeLegalizeOps())
25095 if (SDValue RV = foldXorTruncShiftIntoCmp(N, DAG))
25098 if (Subtarget->hasCMov())
25099 if (SDValue RV = performIntegerAbsCombine(N, DAG))
25102 if (SDValue FPLogic = convertIntLogicToFPLogic(N, DAG, Subtarget))
25108 /// PerformLOADCombine - Do target-specific dag combines on LOAD nodes.
25109 static SDValue PerformLOADCombine(SDNode *N, SelectionDAG &DAG,
25110 TargetLowering::DAGCombinerInfo &DCI,
25111 const X86Subtarget *Subtarget) {
25112 LoadSDNode *Ld = cast<LoadSDNode>(N);
25113 EVT RegVT = Ld->getValueType(0);
25114 EVT MemVT = Ld->getMemoryVT();
25116 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
25118 // For chips with slow 32-byte unaligned loads, break the 32-byte operation
25119 // into two 16-byte operations.
25120 ISD::LoadExtType Ext = Ld->getExtensionType();
25122 unsigned AddressSpace = Ld->getAddressSpace();
25123 unsigned Alignment = Ld->getAlignment();
25124 if (RegVT.is256BitVector() && !DCI.isBeforeLegalizeOps() &&
25125 Ext == ISD::NON_EXTLOAD &&
25126 TLI.allowsMemoryAccess(*DAG.getContext(), DAG.getDataLayout(), RegVT,
25127 AddressSpace, Alignment, &Fast) && !Fast) {
25128 unsigned NumElems = RegVT.getVectorNumElements();
25132 SDValue Ptr = Ld->getBasePtr();
25133 SDValue Increment =
25134 DAG.getConstant(16, dl, TLI.getPointerTy(DAG.getDataLayout()));
25136 EVT HalfVT = EVT::getVectorVT(*DAG.getContext(), MemVT.getScalarType(),
25138 SDValue Load1 = DAG.getLoad(HalfVT, dl, Ld->getChain(), Ptr,
25139 Ld->getPointerInfo(), Ld->isVolatile(),
25140 Ld->isNonTemporal(), Ld->isInvariant(),
25142 Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr, Increment);
25143 SDValue Load2 = DAG.getLoad(HalfVT, dl, Ld->getChain(), Ptr,
25144 Ld->getPointerInfo(), Ld->isVolatile(),
25145 Ld->isNonTemporal(), Ld->isInvariant(),
25146 std::min(16U, Alignment));
25147 SDValue TF = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
25149 Load2.getValue(1));
25151 SDValue NewVec = DAG.getUNDEF(RegVT);
25152 NewVec = Insert128BitVector(NewVec, Load1, 0, DAG, dl);
25153 NewVec = Insert128BitVector(NewVec, Load2, NumElems/2, DAG, dl);
25154 return DCI.CombineTo(N, NewVec, TF, true);
25160 /// PerformMLOADCombine - Resolve extending loads
25161 static SDValue PerformMLOADCombine(SDNode *N, SelectionDAG &DAG,
25162 TargetLowering::DAGCombinerInfo &DCI,
25163 const X86Subtarget *Subtarget) {
25164 MaskedLoadSDNode *Mld = cast<MaskedLoadSDNode>(N);
25165 if (Mld->getExtensionType() != ISD::SEXTLOAD)
25168 EVT VT = Mld->getValueType(0);
25169 unsigned NumElems = VT.getVectorNumElements();
25170 EVT LdVT = Mld->getMemoryVT();
25173 assert(LdVT != VT && "Cannot extend to the same type");
25174 unsigned ToSz = VT.getVectorElementType().getSizeInBits();
25175 unsigned FromSz = LdVT.getVectorElementType().getSizeInBits();
25176 // From, To sizes and ElemCount must be pow of two
25177 assert (isPowerOf2_32(NumElems * FromSz * ToSz) &&
25178 "Unexpected size for extending masked load");
25180 unsigned SizeRatio = ToSz / FromSz;
25181 assert(SizeRatio * NumElems * FromSz == VT.getSizeInBits());
25183 // Create a type on which we perform the shuffle
25184 EVT WideVecVT = EVT::getVectorVT(*DAG.getContext(),
25185 LdVT.getScalarType(), NumElems*SizeRatio);
25186 assert(WideVecVT.getSizeInBits() == VT.getSizeInBits());
25188 // Convert Src0 value
25189 SDValue WideSrc0 = DAG.getBitcast(WideVecVT, Mld->getSrc0());
25190 if (Mld->getSrc0().getOpcode() != ISD::UNDEF) {
25191 SmallVector<int, 16> ShuffleVec(NumElems * SizeRatio, -1);
25192 for (unsigned i = 0; i != NumElems; ++i)
25193 ShuffleVec[i] = i * SizeRatio;
25195 // Can't shuffle using an illegal type.
25196 assert(DAG.getTargetLoweringInfo().isTypeLegal(WideVecVT) &&
25197 "WideVecVT should be legal");
25198 WideSrc0 = DAG.getVectorShuffle(WideVecVT, dl, WideSrc0,
25199 DAG.getUNDEF(WideVecVT), &ShuffleVec[0]);
25201 // Prepare the new mask
25203 SDValue Mask = Mld->getMask();
25204 if (Mask.getValueType() == VT) {
25205 // Mask and original value have the same type
25206 NewMask = DAG.getBitcast(WideVecVT, Mask);
25207 SmallVector<int, 16> ShuffleVec(NumElems * SizeRatio, -1);
25208 for (unsigned i = 0; i != NumElems; ++i)
25209 ShuffleVec[i] = i * SizeRatio;
25210 for (unsigned i = NumElems; i != NumElems*SizeRatio; ++i)
25211 ShuffleVec[i] = NumElems*SizeRatio;
25212 NewMask = DAG.getVectorShuffle(WideVecVT, dl, NewMask,
25213 DAG.getConstant(0, dl, WideVecVT),
25217 assert(Mask.getValueType().getVectorElementType() == MVT::i1);
25218 unsigned WidenNumElts = NumElems*SizeRatio;
25219 unsigned MaskNumElts = VT.getVectorNumElements();
25220 EVT NewMaskVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
25223 unsigned NumConcat = WidenNumElts / MaskNumElts;
25224 SmallVector<SDValue, 16> Ops(NumConcat);
25225 SDValue ZeroVal = DAG.getConstant(0, dl, Mask.getValueType());
25227 for (unsigned i = 1; i != NumConcat; ++i)
25230 NewMask = DAG.getNode(ISD::CONCAT_VECTORS, dl, NewMaskVT, Ops);
25233 SDValue WideLd = DAG.getMaskedLoad(WideVecVT, dl, Mld->getChain(),
25234 Mld->getBasePtr(), NewMask, WideSrc0,
25235 Mld->getMemoryVT(), Mld->getMemOperand(),
25237 SDValue NewVec = DAG.getNode(X86ISD::VSEXT, dl, VT, WideLd);
25238 return DCI.CombineTo(N, NewVec, WideLd.getValue(1), true);
25240 /// PerformMSTORECombine - Resolve truncating stores
25241 static SDValue PerformMSTORECombine(SDNode *N, SelectionDAG &DAG,
25242 const X86Subtarget *Subtarget) {
25243 MaskedStoreSDNode *Mst = cast<MaskedStoreSDNode>(N);
25244 if (!Mst->isTruncatingStore())
25247 EVT VT = Mst->getValue().getValueType();
25248 unsigned NumElems = VT.getVectorNumElements();
25249 EVT StVT = Mst->getMemoryVT();
25252 assert(StVT != VT && "Cannot truncate to the same type");
25253 unsigned FromSz = VT.getVectorElementType().getSizeInBits();
25254 unsigned ToSz = StVT.getVectorElementType().getSizeInBits();
25256 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
25258 // The truncating store is legal in some cases. For example
25259 // vpmovqb, vpmovqw, vpmovqd, vpmovdb, vpmovdw
25260 // are designated for truncate store.
25261 // In this case we don't need any further transformations.
25262 if (TLI.isTruncStoreLegal(VT, StVT))
25265 // From, To sizes and ElemCount must be pow of two
25266 assert (isPowerOf2_32(NumElems * FromSz * ToSz) &&
25267 "Unexpected size for truncating masked store");
25268 // We are going to use the original vector elt for storing.
25269 // Accumulated smaller vector elements must be a multiple of the store size.
25270 assert (((NumElems * FromSz) % ToSz) == 0 &&
25271 "Unexpected ratio for truncating masked store");
25273 unsigned SizeRatio = FromSz / ToSz;
25274 assert(SizeRatio * NumElems * ToSz == VT.getSizeInBits());
25276 // Create a type on which we perform the shuffle
25277 EVT WideVecVT = EVT::getVectorVT(*DAG.getContext(),
25278 StVT.getScalarType(), NumElems*SizeRatio);
25280 assert(WideVecVT.getSizeInBits() == VT.getSizeInBits());
25282 SDValue WideVec = DAG.getBitcast(WideVecVT, Mst->getValue());
25283 SmallVector<int, 16> ShuffleVec(NumElems * SizeRatio, -1);
25284 for (unsigned i = 0; i != NumElems; ++i)
25285 ShuffleVec[i] = i * SizeRatio;
25287 // Can't shuffle using an illegal type.
25288 assert(DAG.getTargetLoweringInfo().isTypeLegal(WideVecVT) &&
25289 "WideVecVT should be legal");
25291 SDValue TruncatedVal = DAG.getVectorShuffle(WideVecVT, dl, WideVec,
25292 DAG.getUNDEF(WideVecVT),
25296 SDValue Mask = Mst->getMask();
25297 if (Mask.getValueType() == VT) {
25298 // Mask and original value have the same type
25299 NewMask = DAG.getBitcast(WideVecVT, Mask);
25300 for (unsigned i = 0; i != NumElems; ++i)
25301 ShuffleVec[i] = i * SizeRatio;
25302 for (unsigned i = NumElems; i != NumElems*SizeRatio; ++i)
25303 ShuffleVec[i] = NumElems*SizeRatio;
25304 NewMask = DAG.getVectorShuffle(WideVecVT, dl, NewMask,
25305 DAG.getConstant(0, dl, WideVecVT),
25309 assert(Mask.getValueType().getVectorElementType() == MVT::i1);
25310 unsigned WidenNumElts = NumElems*SizeRatio;
25311 unsigned MaskNumElts = VT.getVectorNumElements();
25312 EVT NewMaskVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
25315 unsigned NumConcat = WidenNumElts / MaskNumElts;
25316 SmallVector<SDValue, 16> Ops(NumConcat);
25317 SDValue ZeroVal = DAG.getConstant(0, dl, Mask.getValueType());
25319 for (unsigned i = 1; i != NumConcat; ++i)
25322 NewMask = DAG.getNode(ISD::CONCAT_VECTORS, dl, NewMaskVT, Ops);
25325 return DAG.getMaskedStore(Mst->getChain(), dl, TruncatedVal, Mst->getBasePtr(),
25326 NewMask, StVT, Mst->getMemOperand(), false);
25328 /// PerformSTORECombine - Do target-specific dag combines on STORE nodes.
25329 static SDValue PerformSTORECombine(SDNode *N, SelectionDAG &DAG,
25330 const X86Subtarget *Subtarget) {
25331 StoreSDNode *St = cast<StoreSDNode>(N);
25332 EVT VT = St->getValue().getValueType();
25333 EVT StVT = St->getMemoryVT();
25335 SDValue StoredVal = St->getOperand(1);
25336 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
25338 // If we are saving a concatenation of two XMM registers and 32-byte stores
25339 // are slow, such as on Sandy Bridge, perform two 16-byte stores.
25341 unsigned AddressSpace = St->getAddressSpace();
25342 unsigned Alignment = St->getAlignment();
25343 if (VT.is256BitVector() && StVT == VT &&
25344 TLI.allowsMemoryAccess(*DAG.getContext(), DAG.getDataLayout(), VT,
25345 AddressSpace, Alignment, &Fast) && !Fast) {
25346 unsigned NumElems = VT.getVectorNumElements();
25350 SDValue Value0 = Extract128BitVector(StoredVal, 0, DAG, dl);
25351 SDValue Value1 = Extract128BitVector(StoredVal, NumElems/2, DAG, dl);
25354 DAG.getConstant(16, dl, TLI.getPointerTy(DAG.getDataLayout()));
25355 SDValue Ptr0 = St->getBasePtr();
25356 SDValue Ptr1 = DAG.getNode(ISD::ADD, dl, Ptr0.getValueType(), Ptr0, Stride);
25358 SDValue Ch0 = DAG.getStore(St->getChain(), dl, Value0, Ptr0,
25359 St->getPointerInfo(), St->isVolatile(),
25360 St->isNonTemporal(), Alignment);
25361 SDValue Ch1 = DAG.getStore(St->getChain(), dl, Value1, Ptr1,
25362 St->getPointerInfo(), St->isVolatile(),
25363 St->isNonTemporal(),
25364 std::min(16U, Alignment));
25365 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Ch0, Ch1);
25368 // Optimize trunc store (of multiple scalars) to shuffle and store.
25369 // First, pack all of the elements in one place. Next, store to memory
25370 // in fewer chunks.
25371 if (St->isTruncatingStore() && VT.isVector()) {
25372 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
25373 unsigned NumElems = VT.getVectorNumElements();
25374 assert(StVT != VT && "Cannot truncate to the same type");
25375 unsigned FromSz = VT.getVectorElementType().getSizeInBits();
25376 unsigned ToSz = StVT.getVectorElementType().getSizeInBits();
25378 // The truncating store is legal in some cases. For example
25379 // vpmovqb, vpmovqw, vpmovqd, vpmovdb, vpmovdw
25380 // are designated for truncate store.
25381 // In this case we don't need any further transformations.
25382 if (TLI.isTruncStoreLegal(VT, StVT))
25385 // From, To sizes and ElemCount must be pow of two
25386 if (!isPowerOf2_32(NumElems * FromSz * ToSz)) return SDValue();
25387 // We are going to use the original vector elt for storing.
25388 // Accumulated smaller vector elements must be a multiple of the store size.
25389 if (0 != (NumElems * FromSz) % ToSz) return SDValue();
25391 unsigned SizeRatio = FromSz / ToSz;
25393 assert(SizeRatio * NumElems * ToSz == VT.getSizeInBits());
25395 // Create a type on which we perform the shuffle
25396 EVT WideVecVT = EVT::getVectorVT(*DAG.getContext(),
25397 StVT.getScalarType(), NumElems*SizeRatio);
25399 assert(WideVecVT.getSizeInBits() == VT.getSizeInBits());
25401 SDValue WideVec = DAG.getBitcast(WideVecVT, St->getValue());
25402 SmallVector<int, 8> ShuffleVec(NumElems * SizeRatio, -1);
25403 for (unsigned i = 0; i != NumElems; ++i)
25404 ShuffleVec[i] = i * SizeRatio;
25406 // Can't shuffle using an illegal type.
25407 if (!TLI.isTypeLegal(WideVecVT))
25410 SDValue Shuff = DAG.getVectorShuffle(WideVecVT, dl, WideVec,
25411 DAG.getUNDEF(WideVecVT),
25413 // At this point all of the data is stored at the bottom of the
25414 // register. We now need to save it to mem.
25416 // Find the largest store unit
25417 MVT StoreType = MVT::i8;
25418 for (MVT Tp : MVT::integer_valuetypes()) {
25419 if (TLI.isTypeLegal(Tp) && Tp.getSizeInBits() <= NumElems * ToSz)
25423 // On 32bit systems, we can't save 64bit integers. Try bitcasting to F64.
25424 if (TLI.isTypeLegal(MVT::f64) && StoreType.getSizeInBits() < 64 &&
25425 (64 <= NumElems * ToSz))
25426 StoreType = MVT::f64;
25428 // Bitcast the original vector into a vector of store-size units
25429 EVT StoreVecVT = EVT::getVectorVT(*DAG.getContext(),
25430 StoreType, VT.getSizeInBits()/StoreType.getSizeInBits());
25431 assert(StoreVecVT.getSizeInBits() == VT.getSizeInBits());
25432 SDValue ShuffWide = DAG.getBitcast(StoreVecVT, Shuff);
25433 SmallVector<SDValue, 8> Chains;
25434 SDValue Increment = DAG.getConstant(StoreType.getSizeInBits() / 8, dl,
25435 TLI.getPointerTy(DAG.getDataLayout()));
25436 SDValue Ptr = St->getBasePtr();
25438 // Perform one or more big stores into memory.
25439 for (unsigned i=0, e=(ToSz*NumElems)/StoreType.getSizeInBits(); i!=e; ++i) {
25440 SDValue SubVec = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl,
25441 StoreType, ShuffWide,
25442 DAG.getIntPtrConstant(i, dl));
25443 SDValue Ch = DAG.getStore(St->getChain(), dl, SubVec, Ptr,
25444 St->getPointerInfo(), St->isVolatile(),
25445 St->isNonTemporal(), St->getAlignment());
25446 Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr, Increment);
25447 Chains.push_back(Ch);
25450 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Chains);
25453 // Turn load->store of MMX types into GPR load/stores. This avoids clobbering
25454 // the FP state in cases where an emms may be missing.
25455 // A preferable solution to the general problem is to figure out the right
25456 // places to insert EMMS. This qualifies as a quick hack.
25458 // Similarly, turn load->store of i64 into double load/stores in 32-bit mode.
25459 if (VT.getSizeInBits() != 64)
25462 const Function *F = DAG.getMachineFunction().getFunction();
25463 bool NoImplicitFloatOps = F->hasFnAttribute(Attribute::NoImplicitFloat);
25465 !Subtarget->useSoftFloat() && !NoImplicitFloatOps && Subtarget->hasSSE2();
25466 if ((VT.isVector() ||
25467 (VT == MVT::i64 && F64IsLegal && !Subtarget->is64Bit())) &&
25468 isa<LoadSDNode>(St->getValue()) &&
25469 !cast<LoadSDNode>(St->getValue())->isVolatile() &&
25470 St->getChain().hasOneUse() && !St->isVolatile()) {
25471 SDNode* LdVal = St->getValue().getNode();
25472 LoadSDNode *Ld = nullptr;
25473 int TokenFactorIndex = -1;
25474 SmallVector<SDValue, 8> Ops;
25475 SDNode* ChainVal = St->getChain().getNode();
25476 // Must be a store of a load. We currently handle two cases: the load
25477 // is a direct child, and it's under an intervening TokenFactor. It is
25478 // possible to dig deeper under nested TokenFactors.
25479 if (ChainVal == LdVal)
25480 Ld = cast<LoadSDNode>(St->getChain());
25481 else if (St->getValue().hasOneUse() &&
25482 ChainVal->getOpcode() == ISD::TokenFactor) {
25483 for (unsigned i = 0, e = ChainVal->getNumOperands(); i != e; ++i) {
25484 if (ChainVal->getOperand(i).getNode() == LdVal) {
25485 TokenFactorIndex = i;
25486 Ld = cast<LoadSDNode>(St->getValue());
25488 Ops.push_back(ChainVal->getOperand(i));
25492 if (!Ld || !ISD::isNormalLoad(Ld))
25495 // If this is not the MMX case, i.e. we are just turning i64 load/store
25496 // into f64 load/store, avoid the transformation if there are multiple
25497 // uses of the loaded value.
25498 if (!VT.isVector() && !Ld->hasNUsesOfValue(1, 0))
25503 // If we are a 64-bit capable x86, lower to a single movq load/store pair.
25504 // Otherwise, if it's legal to use f64 SSE instructions, use f64 load/store
25506 if (Subtarget->is64Bit() || F64IsLegal) {
25507 EVT LdVT = Subtarget->is64Bit() ? MVT::i64 : MVT::f64;
25508 SDValue NewLd = DAG.getLoad(LdVT, LdDL, Ld->getChain(), Ld->getBasePtr(),
25509 Ld->getPointerInfo(), Ld->isVolatile(),
25510 Ld->isNonTemporal(), Ld->isInvariant(),
25511 Ld->getAlignment());
25512 SDValue NewChain = NewLd.getValue(1);
25513 if (TokenFactorIndex != -1) {
25514 Ops.push_back(NewChain);
25515 NewChain = DAG.getNode(ISD::TokenFactor, LdDL, MVT::Other, Ops);
25517 return DAG.getStore(NewChain, StDL, NewLd, St->getBasePtr(),
25518 St->getPointerInfo(),
25519 St->isVolatile(), St->isNonTemporal(),
25520 St->getAlignment());
25523 // Otherwise, lower to two pairs of 32-bit loads / stores.
25524 SDValue LoAddr = Ld->getBasePtr();
25525 SDValue HiAddr = DAG.getNode(ISD::ADD, LdDL, MVT::i32, LoAddr,
25526 DAG.getConstant(4, LdDL, MVT::i32));
25528 SDValue LoLd = DAG.getLoad(MVT::i32, LdDL, Ld->getChain(), LoAddr,
25529 Ld->getPointerInfo(),
25530 Ld->isVolatile(), Ld->isNonTemporal(),
25531 Ld->isInvariant(), Ld->getAlignment());
25532 SDValue HiLd = DAG.getLoad(MVT::i32, LdDL, Ld->getChain(), HiAddr,
25533 Ld->getPointerInfo().getWithOffset(4),
25534 Ld->isVolatile(), Ld->isNonTemporal(),
25536 MinAlign(Ld->getAlignment(), 4));
25538 SDValue NewChain = LoLd.getValue(1);
25539 if (TokenFactorIndex != -1) {
25540 Ops.push_back(LoLd);
25541 Ops.push_back(HiLd);
25542 NewChain = DAG.getNode(ISD::TokenFactor, LdDL, MVT::Other, Ops);
25545 LoAddr = St->getBasePtr();
25546 HiAddr = DAG.getNode(ISD::ADD, StDL, MVT::i32, LoAddr,
25547 DAG.getConstant(4, StDL, MVT::i32));
25549 SDValue LoSt = DAG.getStore(NewChain, StDL, LoLd, LoAddr,
25550 St->getPointerInfo(),
25551 St->isVolatile(), St->isNonTemporal(),
25552 St->getAlignment());
25553 SDValue HiSt = DAG.getStore(NewChain, StDL, HiLd, HiAddr,
25554 St->getPointerInfo().getWithOffset(4),
25556 St->isNonTemporal(),
25557 MinAlign(St->getAlignment(), 4));
25558 return DAG.getNode(ISD::TokenFactor, StDL, MVT::Other, LoSt, HiSt);
25561 // This is similar to the above case, but here we handle a scalar 64-bit
25562 // integer store that is extracted from a vector on a 32-bit target.
25563 // If we have SSE2, then we can treat it like a floating-point double
25564 // to get past legalization. The execution dependencies fixup pass will
25565 // choose the optimal machine instruction for the store if this really is
25566 // an integer or v2f32 rather than an f64.
25567 if (VT == MVT::i64 && F64IsLegal && !Subtarget->is64Bit() &&
25568 St->getOperand(1).getOpcode() == ISD::EXTRACT_VECTOR_ELT) {
25569 SDValue OldExtract = St->getOperand(1);
25570 SDValue ExtOp0 = OldExtract.getOperand(0);
25571 unsigned VecSize = ExtOp0.getValueSizeInBits();
25572 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), MVT::f64, VecSize / 64);
25573 SDValue BitCast = DAG.getBitcast(VecVT, ExtOp0);
25574 SDValue NewExtract = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
25575 BitCast, OldExtract.getOperand(1));
25576 return DAG.getStore(St->getChain(), dl, NewExtract, St->getBasePtr(),
25577 St->getPointerInfo(), St->isVolatile(),
25578 St->isNonTemporal(), St->getAlignment());
25584 /// Return 'true' if this vector operation is "horizontal"
25585 /// and return the operands for the horizontal operation in LHS and RHS. A
25586 /// horizontal operation performs the binary operation on successive elements
25587 /// of its first operand, then on successive elements of its second operand,
25588 /// returning the resulting values in a vector. For example, if
25589 /// A = < float a0, float a1, float a2, float a3 >
25591 /// B = < float b0, float b1, float b2, float b3 >
25592 /// then the result of doing a horizontal operation on A and B is
25593 /// A horizontal-op B = < a0 op a1, a2 op a3, b0 op b1, b2 op b3 >.
25594 /// In short, LHS and RHS are inspected to see if LHS op RHS is of the form
25595 /// A horizontal-op B, for some already available A and B, and if so then LHS is
25596 /// set to A, RHS to B, and the routine returns 'true'.
25597 /// Note that the binary operation should have the property that if one of the
25598 /// operands is UNDEF then the result is UNDEF.
25599 static bool isHorizontalBinOp(SDValue &LHS, SDValue &RHS, bool IsCommutative) {
25600 // Look for the following pattern: if
25601 // A = < float a0, float a1, float a2, float a3 >
25602 // B = < float b0, float b1, float b2, float b3 >
25604 // LHS = VECTOR_SHUFFLE A, B, <0, 2, 4, 6>
25605 // RHS = VECTOR_SHUFFLE A, B, <1, 3, 5, 7>
25606 // then LHS op RHS = < a0 op a1, a2 op a3, b0 op b1, b2 op b3 >
25607 // which is A horizontal-op B.
25609 // At least one of the operands should be a vector shuffle.
25610 if (LHS.getOpcode() != ISD::VECTOR_SHUFFLE &&
25611 RHS.getOpcode() != ISD::VECTOR_SHUFFLE)
25614 MVT VT = LHS.getSimpleValueType();
25616 assert((VT.is128BitVector() || VT.is256BitVector()) &&
25617 "Unsupported vector type for horizontal add/sub");
25619 // Handle 128 and 256-bit vector lengths. AVX defines horizontal add/sub to
25620 // operate independently on 128-bit lanes.
25621 unsigned NumElts = VT.getVectorNumElements();
25622 unsigned NumLanes = VT.getSizeInBits()/128;
25623 unsigned NumLaneElts = NumElts / NumLanes;
25624 assert((NumLaneElts % 2 == 0) &&
25625 "Vector type should have an even number of elements in each lane");
25626 unsigned HalfLaneElts = NumLaneElts/2;
25628 // View LHS in the form
25629 // LHS = VECTOR_SHUFFLE A, B, LMask
25630 // If LHS is not a shuffle then pretend it is the shuffle
25631 // LHS = VECTOR_SHUFFLE LHS, undef, <0, 1, ..., N-1>
25632 // NOTE: in what follows a default initialized SDValue represents an UNDEF of
25635 SmallVector<int, 16> LMask(NumElts);
25636 if (LHS.getOpcode() == ISD::VECTOR_SHUFFLE) {
25637 if (LHS.getOperand(0).getOpcode() != ISD::UNDEF)
25638 A = LHS.getOperand(0);
25639 if (LHS.getOperand(1).getOpcode() != ISD::UNDEF)
25640 B = LHS.getOperand(1);
25641 ArrayRef<int> Mask = cast<ShuffleVectorSDNode>(LHS.getNode())->getMask();
25642 std::copy(Mask.begin(), Mask.end(), LMask.begin());
25644 if (LHS.getOpcode() != ISD::UNDEF)
25646 for (unsigned i = 0; i != NumElts; ++i)
25650 // Likewise, view RHS in the form
25651 // RHS = VECTOR_SHUFFLE C, D, RMask
25653 SmallVector<int, 16> RMask(NumElts);
25654 if (RHS.getOpcode() == ISD::VECTOR_SHUFFLE) {
25655 if (RHS.getOperand(0).getOpcode() != ISD::UNDEF)
25656 C = RHS.getOperand(0);
25657 if (RHS.getOperand(1).getOpcode() != ISD::UNDEF)
25658 D = RHS.getOperand(1);
25659 ArrayRef<int> Mask = cast<ShuffleVectorSDNode>(RHS.getNode())->getMask();
25660 std::copy(Mask.begin(), Mask.end(), RMask.begin());
25662 if (RHS.getOpcode() != ISD::UNDEF)
25664 for (unsigned i = 0; i != NumElts; ++i)
25668 // Check that the shuffles are both shuffling the same vectors.
25669 if (!(A == C && B == D) && !(A == D && B == C))
25672 // If everything is UNDEF then bail out: it would be better to fold to UNDEF.
25673 if (!A.getNode() && !B.getNode())
25676 // If A and B occur in reverse order in RHS, then "swap" them (which means
25677 // rewriting the mask).
25679 ShuffleVectorSDNode::commuteMask(RMask);
25681 // At this point LHS and RHS are equivalent to
25682 // LHS = VECTOR_SHUFFLE A, B, LMask
25683 // RHS = VECTOR_SHUFFLE A, B, RMask
25684 // Check that the masks correspond to performing a horizontal operation.
25685 for (unsigned l = 0; l != NumElts; l += NumLaneElts) {
25686 for (unsigned i = 0; i != NumLaneElts; ++i) {
25687 int LIdx = LMask[i+l], RIdx = RMask[i+l];
25689 // Ignore any UNDEF components.
25690 if (LIdx < 0 || RIdx < 0 ||
25691 (!A.getNode() && (LIdx < (int)NumElts || RIdx < (int)NumElts)) ||
25692 (!B.getNode() && (LIdx >= (int)NumElts || RIdx >= (int)NumElts)))
25695 // Check that successive elements are being operated on. If not, this is
25696 // not a horizontal operation.
25697 unsigned Src = (i/HalfLaneElts); // each lane is split between srcs
25698 int Index = 2*(i%HalfLaneElts) + NumElts*Src + l;
25699 if (!(LIdx == Index && RIdx == Index + 1) &&
25700 !(IsCommutative && LIdx == Index + 1 && RIdx == Index))
25705 LHS = A.getNode() ? A : B; // If A is 'UNDEF', use B for it.
25706 RHS = B.getNode() ? B : A; // If B is 'UNDEF', use A for it.
25710 /// Do target-specific dag combines on floating point adds.
25711 static SDValue PerformFADDCombine(SDNode *N, SelectionDAG &DAG,
25712 const X86Subtarget *Subtarget) {
25713 EVT VT = N->getValueType(0);
25714 SDValue LHS = N->getOperand(0);
25715 SDValue RHS = N->getOperand(1);
25717 // Try to synthesize horizontal adds from adds of shuffles.
25718 if (((Subtarget->hasSSE3() && (VT == MVT::v4f32 || VT == MVT::v2f64)) ||
25719 (Subtarget->hasFp256() && (VT == MVT::v8f32 || VT == MVT::v4f64))) &&
25720 isHorizontalBinOp(LHS, RHS, true))
25721 return DAG.getNode(X86ISD::FHADD, SDLoc(N), VT, LHS, RHS);
25725 /// Do target-specific dag combines on floating point subs.
25726 static SDValue PerformFSUBCombine(SDNode *N, SelectionDAG &DAG,
25727 const X86Subtarget *Subtarget) {
25728 EVT VT = N->getValueType(0);
25729 SDValue LHS = N->getOperand(0);
25730 SDValue RHS = N->getOperand(1);
25732 // Try to synthesize horizontal subs from subs of shuffles.
25733 if (((Subtarget->hasSSE3() && (VT == MVT::v4f32 || VT == MVT::v2f64)) ||
25734 (Subtarget->hasFp256() && (VT == MVT::v8f32 || VT == MVT::v4f64))) &&
25735 isHorizontalBinOp(LHS, RHS, false))
25736 return DAG.getNode(X86ISD::FHSUB, SDLoc(N), VT, LHS, RHS);
25740 /// Do target-specific dag combines on X86ISD::FOR and X86ISD::FXOR nodes.
25741 static SDValue PerformFORCombine(SDNode *N, SelectionDAG &DAG,
25742 const X86Subtarget *Subtarget) {
25743 assert(N->getOpcode() == X86ISD::FOR || N->getOpcode() == X86ISD::FXOR);
25745 // F[X]OR(0.0, x) -> x
25746 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
25747 if (C->getValueAPF().isPosZero())
25748 return N->getOperand(1);
25750 // F[X]OR(x, 0.0) -> x
25751 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
25752 if (C->getValueAPF().isPosZero())
25753 return N->getOperand(0);
25755 EVT VT = N->getValueType(0);
25756 if (VT.is512BitVector() && !Subtarget->hasDQI()) {
25758 MVT IntScalar = MVT::getIntegerVT(VT.getScalarSizeInBits());
25759 MVT IntVT = MVT::getVectorVT(IntScalar, VT.getVectorNumElements());
25761 SDValue Op0 = DAG.getNode(ISD::BITCAST, dl, IntVT, N->getOperand(0));
25762 SDValue Op1 = DAG.getNode(ISD::BITCAST, dl, IntVT, N->getOperand(1));
25763 unsigned IntOpcode = (N->getOpcode() == X86ISD::FOR) ? ISD::OR : ISD::XOR;
25764 SDValue IntOp = DAG.getNode(IntOpcode, dl, IntVT, Op0, Op1);
25765 return DAG.getNode(ISD::BITCAST, dl, VT, IntOp);
25770 /// Do target-specific dag combines on X86ISD::FMIN and X86ISD::FMAX nodes.
25771 static SDValue PerformFMinFMaxCombine(SDNode *N, SelectionDAG &DAG) {
25772 assert(N->getOpcode() == X86ISD::FMIN || N->getOpcode() == X86ISD::FMAX);
25774 // Only perform optimizations if UnsafeMath is used.
25775 if (!DAG.getTarget().Options.UnsafeFPMath)
25778 // If we run in unsafe-math mode, then convert the FMAX and FMIN nodes
25779 // into FMINC and FMAXC, which are Commutative operations.
25780 unsigned NewOp = 0;
25781 switch (N->getOpcode()) {
25782 default: llvm_unreachable("unknown opcode");
25783 case X86ISD::FMIN: NewOp = X86ISD::FMINC; break;
25784 case X86ISD::FMAX: NewOp = X86ISD::FMAXC; break;
25787 return DAG.getNode(NewOp, SDLoc(N), N->getValueType(0),
25788 N->getOperand(0), N->getOperand(1));
25791 /// Do target-specific dag combines on X86ISD::FAND nodes.
25792 static SDValue PerformFANDCombine(SDNode *N, SelectionDAG &DAG) {
25793 // FAND(0.0, x) -> 0.0
25794 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
25795 if (C->getValueAPF().isPosZero())
25796 return N->getOperand(0);
25798 // FAND(x, 0.0) -> 0.0
25799 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
25800 if (C->getValueAPF().isPosZero())
25801 return N->getOperand(1);
25806 /// Do target-specific dag combines on X86ISD::FANDN nodes
25807 static SDValue PerformFANDNCombine(SDNode *N, SelectionDAG &DAG) {
25808 // FANDN(0.0, x) -> x
25809 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
25810 if (C->getValueAPF().isPosZero())
25811 return N->getOperand(1);
25813 // FANDN(x, 0.0) -> 0.0
25814 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
25815 if (C->getValueAPF().isPosZero())
25816 return N->getOperand(1);
25821 static SDValue PerformBTCombine(SDNode *N,
25823 TargetLowering::DAGCombinerInfo &DCI) {
25824 // BT ignores high bits in the bit index operand.
25825 SDValue Op1 = N->getOperand(1);
25826 if (Op1.hasOneUse()) {
25827 unsigned BitWidth = Op1.getValueSizeInBits();
25828 APInt DemandedMask = APInt::getLowBitsSet(BitWidth, Log2_32(BitWidth));
25829 APInt KnownZero, KnownOne;
25830 TargetLowering::TargetLoweringOpt TLO(DAG, !DCI.isBeforeLegalize(),
25831 !DCI.isBeforeLegalizeOps());
25832 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
25833 if (TLO.ShrinkDemandedConstant(Op1, DemandedMask) ||
25834 TLI.SimplifyDemandedBits(Op1, DemandedMask, KnownZero, KnownOne, TLO))
25835 DCI.CommitTargetLoweringOpt(TLO);
25840 static SDValue PerformVZEXT_MOVLCombine(SDNode *N, SelectionDAG &DAG) {
25841 SDValue Op = N->getOperand(0);
25842 if (Op.getOpcode() == ISD::BITCAST)
25843 Op = Op.getOperand(0);
25844 EVT VT = N->getValueType(0), OpVT = Op.getValueType();
25845 if (Op.getOpcode() == X86ISD::VZEXT_LOAD &&
25846 VT.getVectorElementType().getSizeInBits() ==
25847 OpVT.getVectorElementType().getSizeInBits()) {
25848 return DAG.getNode(ISD::BITCAST, SDLoc(N), VT, Op);
25853 static SDValue PerformSIGN_EXTEND_INREGCombine(SDNode *N, SelectionDAG &DAG,
25854 const X86Subtarget *Subtarget) {
25855 EVT VT = N->getValueType(0);
25856 if (!VT.isVector())
25859 SDValue N0 = N->getOperand(0);
25860 SDValue N1 = N->getOperand(1);
25861 EVT ExtraVT = cast<VTSDNode>(N1)->getVT();
25864 // The SIGN_EXTEND_INREG to v4i64 is expensive operation on the
25865 // both SSE and AVX2 since there is no sign-extended shift right
25866 // operation on a vector with 64-bit elements.
25867 //(sext_in_reg (v4i64 anyext (v4i32 x )), ExtraVT) ->
25868 // (v4i64 sext (v4i32 sext_in_reg (v4i32 x , ExtraVT)))
25869 if (VT == MVT::v4i64 && (N0.getOpcode() == ISD::ANY_EXTEND ||
25870 N0.getOpcode() == ISD::SIGN_EXTEND)) {
25871 SDValue N00 = N0.getOperand(0);
25873 // EXTLOAD has a better solution on AVX2,
25874 // it may be replaced with X86ISD::VSEXT node.
25875 if (N00.getOpcode() == ISD::LOAD && Subtarget->hasInt256())
25876 if (!ISD::isNormalLoad(N00.getNode()))
25879 if (N00.getValueType() == MVT::v4i32 && ExtraVT.getSizeInBits() < 128) {
25880 SDValue Tmp = DAG.getNode(ISD::SIGN_EXTEND_INREG, dl, MVT::v4i32,
25882 return DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i64, Tmp);
25888 /// sext(add_nsw(x, C)) --> add(sext(x), C_sext)
25889 /// Promoting a sign extension ahead of an 'add nsw' exposes opportunities
25890 /// to combine math ops, use an LEA, or use a complex addressing mode. This can
25891 /// eliminate extend, add, and shift instructions.
25892 static SDValue promoteSextBeforeAddNSW(SDNode *Sext, SelectionDAG &DAG,
25893 const X86Subtarget *Subtarget) {
25894 // TODO: This should be valid for other integer types.
25895 EVT VT = Sext->getValueType(0);
25896 if (VT != MVT::i64)
25899 // We need an 'add nsw' feeding into the 'sext'.
25900 SDValue Add = Sext->getOperand(0);
25901 if (Add.getOpcode() != ISD::ADD || !Add->getFlags()->hasNoSignedWrap())
25904 // Having a constant operand to the 'add' ensures that we are not increasing
25905 // the instruction count because the constant is extended for free below.
25906 // A constant operand can also become the displacement field of an LEA.
25907 auto *AddOp1 = dyn_cast<ConstantSDNode>(Add.getOperand(1));
25911 // Don't make the 'add' bigger if there's no hope of combining it with some
25912 // other 'add' or 'shl' instruction.
25913 // TODO: It may be profitable to generate simpler LEA instructions in place
25914 // of single 'add' instructions, but the cost model for selecting an LEA
25915 // currently has a high threshold.
25916 bool HasLEAPotential = false;
25917 for (auto *User : Sext->uses()) {
25918 if (User->getOpcode() == ISD::ADD || User->getOpcode() == ISD::SHL) {
25919 HasLEAPotential = true;
25923 if (!HasLEAPotential)
25926 // Everything looks good, so pull the 'sext' ahead of the 'add'.
25927 int64_t AddConstant = AddOp1->getSExtValue();
25928 SDValue AddOp0 = Add.getOperand(0);
25929 SDValue NewSext = DAG.getNode(ISD::SIGN_EXTEND, SDLoc(Sext), VT, AddOp0);
25930 SDValue NewConstant = DAG.getConstant(AddConstant, SDLoc(Add), VT);
25932 // The wider add is guaranteed to not wrap because both operands are
25935 Flags.setNoSignedWrap(true);
25936 return DAG.getNode(ISD::ADD, SDLoc(Add), VT, NewSext, NewConstant, &Flags);
25939 static SDValue PerformSExtCombine(SDNode *N, SelectionDAG &DAG,
25940 TargetLowering::DAGCombinerInfo &DCI,
25941 const X86Subtarget *Subtarget) {
25942 SDValue N0 = N->getOperand(0);
25943 EVT VT = N->getValueType(0);
25944 EVT SVT = VT.getScalarType();
25945 EVT InVT = N0.getValueType();
25946 EVT InSVT = InVT.getScalarType();
25949 // (i8,i32 sext (sdivrem (i8 x, i8 y)) ->
25950 // (i8,i32 (sdivrem_sext_hreg (i8 x, i8 y)
25951 // This exposes the sext to the sdivrem lowering, so that it directly extends
25952 // from AH (which we otherwise need to do contortions to access).
25953 if (N0.getOpcode() == ISD::SDIVREM && N0.getResNo() == 1 &&
25954 InVT == MVT::i8 && VT == MVT::i32) {
25955 SDVTList NodeTys = DAG.getVTList(MVT::i8, VT);
25956 SDValue R = DAG.getNode(X86ISD::SDIVREM8_SEXT_HREG, DL, NodeTys,
25957 N0.getOperand(0), N0.getOperand(1));
25958 DAG.ReplaceAllUsesOfValueWith(N0.getValue(0), R.getValue(0));
25959 return R.getValue(1);
25962 if (!DCI.isBeforeLegalizeOps()) {
25963 if (InVT == MVT::i1) {
25964 SDValue Zero = DAG.getConstant(0, DL, VT);
25966 DAG.getConstant(APInt::getAllOnesValue(VT.getSizeInBits()), DL, VT);
25967 return DAG.getNode(ISD::SELECT, DL, VT, N0, AllOnes, Zero);
25972 if (VT.isVector() && Subtarget->hasSSE2()) {
25973 auto ExtendVecSize = [&DAG](SDLoc DL, SDValue N, unsigned Size) {
25974 EVT InVT = N.getValueType();
25975 EVT OutVT = EVT::getVectorVT(*DAG.getContext(), InVT.getScalarType(),
25976 Size / InVT.getScalarSizeInBits());
25977 SmallVector<SDValue, 8> Opnds(Size / InVT.getSizeInBits(),
25978 DAG.getUNDEF(InVT));
25980 return DAG.getNode(ISD::CONCAT_VECTORS, DL, OutVT, Opnds);
25983 // If target-size is less than 128-bits, extend to a type that would extend
25984 // to 128 bits, extend that and extract the original target vector.
25985 if (VT.getSizeInBits() < 128 && !(128 % VT.getSizeInBits()) &&
25986 (SVT == MVT::i64 || SVT == MVT::i32 || SVT == MVT::i16) &&
25987 (InSVT == MVT::i32 || InSVT == MVT::i16 || InSVT == MVT::i8)) {
25988 unsigned Scale = 128 / VT.getSizeInBits();
25990 EVT::getVectorVT(*DAG.getContext(), SVT, 128 / SVT.getSizeInBits());
25991 SDValue Ex = ExtendVecSize(DL, N0, Scale * InVT.getSizeInBits());
25992 SDValue SExt = DAG.getNode(ISD::SIGN_EXTEND, DL, ExVT, Ex);
25993 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, VT, SExt,
25994 DAG.getIntPtrConstant(0, DL));
25997 // If target-size is 128-bits, then convert to ISD::SIGN_EXTEND_VECTOR_INREG
25998 // which ensures lowering to X86ISD::VSEXT (pmovsx*).
25999 if (VT.getSizeInBits() == 128 &&
26000 (SVT == MVT::i64 || SVT == MVT::i32 || SVT == MVT::i16) &&
26001 (InSVT == MVT::i32 || InSVT == MVT::i16 || InSVT == MVT::i8)) {
26002 SDValue ExOp = ExtendVecSize(DL, N0, 128);
26003 return DAG.getSignExtendVectorInReg(ExOp, DL, VT);
26006 // On pre-AVX2 targets, split into 128-bit nodes of
26007 // ISD::SIGN_EXTEND_VECTOR_INREG.
26008 if (!Subtarget->hasInt256() && !(VT.getSizeInBits() % 128) &&
26009 (SVT == MVT::i64 || SVT == MVT::i32 || SVT == MVT::i16) &&
26010 (InSVT == MVT::i32 || InSVT == MVT::i16 || InSVT == MVT::i8)) {
26011 unsigned NumVecs = VT.getSizeInBits() / 128;
26012 unsigned NumSubElts = 128 / SVT.getSizeInBits();
26013 EVT SubVT = EVT::getVectorVT(*DAG.getContext(), SVT, NumSubElts);
26014 EVT InSubVT = EVT::getVectorVT(*DAG.getContext(), InSVT, NumSubElts);
26016 SmallVector<SDValue, 8> Opnds;
26017 for (unsigned i = 0, Offset = 0; i != NumVecs;
26018 ++i, Offset += NumSubElts) {
26019 SDValue SrcVec = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, InSubVT, N0,
26020 DAG.getIntPtrConstant(Offset, DL));
26021 SrcVec = ExtendVecSize(DL, SrcVec, 128);
26022 SrcVec = DAG.getSignExtendVectorInReg(SrcVec, DL, SubVT);
26023 Opnds.push_back(SrcVec);
26025 return DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, Opnds);
26029 if (Subtarget->hasAVX() && VT.isVector() && VT.getSizeInBits() == 256)
26030 if (SDValue R = WidenMaskArithmetic(N, DAG, DCI, Subtarget))
26033 if (SDValue NewAdd = promoteSextBeforeAddNSW(N, DAG, Subtarget))
26039 static SDValue PerformFMACombine(SDNode *N, SelectionDAG &DAG,
26040 const X86Subtarget* Subtarget) {
26042 EVT VT = N->getValueType(0);
26044 // Let legalize expand this if it isn't a legal type yet.
26045 if (!DAG.getTargetLoweringInfo().isTypeLegal(VT))
26048 EVT ScalarVT = VT.getScalarType();
26049 if ((ScalarVT != MVT::f32 && ScalarVT != MVT::f64) ||
26050 (!Subtarget->hasFMA() && !Subtarget->hasFMA4() &&
26051 !Subtarget->hasAVX512()))
26054 SDValue A = N->getOperand(0);
26055 SDValue B = N->getOperand(1);
26056 SDValue C = N->getOperand(2);
26058 bool NegA = (A.getOpcode() == ISD::FNEG);
26059 bool NegB = (B.getOpcode() == ISD::FNEG);
26060 bool NegC = (C.getOpcode() == ISD::FNEG);
26062 // Negative multiplication when NegA xor NegB
26063 bool NegMul = (NegA != NegB);
26065 A = A.getOperand(0);
26067 B = B.getOperand(0);
26069 C = C.getOperand(0);
26073 Opcode = (!NegC) ? X86ISD::FMADD : X86ISD::FMSUB;
26075 Opcode = (!NegC) ? X86ISD::FNMADD : X86ISD::FNMSUB;
26077 return DAG.getNode(Opcode, dl, VT, A, B, C);
26080 static SDValue PerformZExtCombine(SDNode *N, SelectionDAG &DAG,
26081 TargetLowering::DAGCombinerInfo &DCI,
26082 const X86Subtarget *Subtarget) {
26083 // (i32 zext (and (i8 x86isd::setcc_carry), 1)) ->
26084 // (and (i32 x86isd::setcc_carry), 1)
26085 // This eliminates the zext. This transformation is necessary because
26086 // ISD::SETCC is always legalized to i8.
26088 SDValue N0 = N->getOperand(0);
26089 EVT VT = N->getValueType(0);
26091 if (N0.getOpcode() == ISD::AND &&
26093 N0.getOperand(0).hasOneUse()) {
26094 SDValue N00 = N0.getOperand(0);
26095 if (N00.getOpcode() == X86ISD::SETCC_CARRY) {
26096 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
26097 if (!C || C->getZExtValue() != 1)
26099 return DAG.getNode(ISD::AND, dl, VT,
26100 DAG.getNode(X86ISD::SETCC_CARRY, dl, VT,
26101 N00.getOperand(0), N00.getOperand(1)),
26102 DAG.getConstant(1, dl, VT));
26106 if (N0.getOpcode() == ISD::TRUNCATE &&
26108 N0.getOperand(0).hasOneUse()) {
26109 SDValue N00 = N0.getOperand(0);
26110 if (N00.getOpcode() == X86ISD::SETCC_CARRY) {
26111 return DAG.getNode(ISD::AND, dl, VT,
26112 DAG.getNode(X86ISD::SETCC_CARRY, dl, VT,
26113 N00.getOperand(0), N00.getOperand(1)),
26114 DAG.getConstant(1, dl, VT));
26118 if (VT.is256BitVector())
26119 if (SDValue R = WidenMaskArithmetic(N, DAG, DCI, Subtarget))
26122 // (i8,i32 zext (udivrem (i8 x, i8 y)) ->
26123 // (i8,i32 (udivrem_zext_hreg (i8 x, i8 y)
26124 // This exposes the zext to the udivrem lowering, so that it directly extends
26125 // from AH (which we otherwise need to do contortions to access).
26126 if (N0.getOpcode() == ISD::UDIVREM &&
26127 N0.getResNo() == 1 && N0.getValueType() == MVT::i8 &&
26128 (VT == MVT::i32 || VT == MVT::i64)) {
26129 SDVTList NodeTys = DAG.getVTList(MVT::i8, VT);
26130 SDValue R = DAG.getNode(X86ISD::UDIVREM8_ZEXT_HREG, dl, NodeTys,
26131 N0.getOperand(0), N0.getOperand(1));
26132 DAG.ReplaceAllUsesOfValueWith(N0.getValue(0), R.getValue(0));
26133 return R.getValue(1);
26139 // Optimize x == -y --> x+y == 0
26140 // x != -y --> x+y != 0
26141 static SDValue PerformISDSETCCCombine(SDNode *N, SelectionDAG &DAG,
26142 const X86Subtarget* Subtarget) {
26143 ISD::CondCode CC = cast<CondCodeSDNode>(N->getOperand(2))->get();
26144 SDValue LHS = N->getOperand(0);
26145 SDValue RHS = N->getOperand(1);
26146 EVT VT = N->getValueType(0);
26149 if ((CC == ISD::SETNE || CC == ISD::SETEQ) && LHS.getOpcode() == ISD::SUB)
26150 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(LHS.getOperand(0)))
26151 if (C->getAPIntValue() == 0 && LHS.hasOneUse()) {
26152 SDValue addV = DAG.getNode(ISD::ADD, DL, LHS.getValueType(), RHS,
26153 LHS.getOperand(1));
26154 return DAG.getSetCC(DL, N->getValueType(0), addV,
26155 DAG.getConstant(0, DL, addV.getValueType()), CC);
26157 if ((CC == ISD::SETNE || CC == ISD::SETEQ) && RHS.getOpcode() == ISD::SUB)
26158 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(RHS.getOperand(0)))
26159 if (C->getAPIntValue() == 0 && RHS.hasOneUse()) {
26160 SDValue addV = DAG.getNode(ISD::ADD, DL, RHS.getValueType(), LHS,
26161 RHS.getOperand(1));
26162 return DAG.getSetCC(DL, N->getValueType(0), addV,
26163 DAG.getConstant(0, DL, addV.getValueType()), CC);
26166 if (VT.getScalarType() == MVT::i1 &&
26167 (CC == ISD::SETNE || CC == ISD::SETEQ || ISD::isSignedIntSetCC(CC))) {
26169 (LHS.getOpcode() == ISD::SIGN_EXTEND) &&
26170 (LHS.getOperand(0).getValueType().getScalarType() == MVT::i1);
26171 bool IsVZero1 = ISD::isBuildVectorAllZeros(RHS.getNode());
26173 if (!IsSEXT0 || !IsVZero1) {
26174 // Swap the operands and update the condition code.
26175 std::swap(LHS, RHS);
26176 CC = ISD::getSetCCSwappedOperands(CC);
26178 IsSEXT0 = (LHS.getOpcode() == ISD::SIGN_EXTEND) &&
26179 (LHS.getOperand(0).getValueType().getScalarType() == MVT::i1);
26180 IsVZero1 = ISD::isBuildVectorAllZeros(RHS.getNode());
26183 if (IsSEXT0 && IsVZero1) {
26184 assert(VT == LHS.getOperand(0).getValueType() &&
26185 "Uexpected operand type");
26186 if (CC == ISD::SETGT)
26187 return DAG.getConstant(0, DL, VT);
26188 if (CC == ISD::SETLE)
26189 return DAG.getConstant(1, DL, VT);
26190 if (CC == ISD::SETEQ || CC == ISD::SETGE)
26191 return DAG.getNOT(DL, LHS.getOperand(0), VT);
26193 assert((CC == ISD::SETNE || CC == ISD::SETLT) &&
26194 "Unexpected condition code!");
26195 return LHS.getOperand(0);
26202 static SDValue NarrowVectorLoadToElement(LoadSDNode *Load, unsigned Index,
26203 SelectionDAG &DAG) {
26205 MVT VT = Load->getSimpleValueType(0);
26206 MVT EVT = VT.getVectorElementType();
26207 SDValue Addr = Load->getOperand(1);
26208 SDValue NewAddr = DAG.getNode(
26209 ISD::ADD, dl, Addr.getSimpleValueType(), Addr,
26210 DAG.getConstant(Index * EVT.getStoreSize(), dl,
26211 Addr.getSimpleValueType()));
26214 DAG.getLoad(EVT, dl, Load->getChain(), NewAddr,
26215 DAG.getMachineFunction().getMachineMemOperand(
26216 Load->getMemOperand(), 0, EVT.getStoreSize()));
26220 static SDValue PerformINSERTPSCombine(SDNode *N, SelectionDAG &DAG,
26221 const X86Subtarget *Subtarget) {
26223 MVT VT = N->getOperand(1)->getSimpleValueType(0);
26224 assert((VT == MVT::v4f32 || VT == MVT::v4i32) &&
26225 "X86insertps is only defined for v4x32");
26227 SDValue Ld = N->getOperand(1);
26228 if (MayFoldLoad(Ld)) {
26229 // Extract the countS bits from the immediate so we can get the proper
26230 // address when narrowing the vector load to a specific element.
26231 // When the second source op is a memory address, insertps doesn't use
26232 // countS and just gets an f32 from that address.
26233 unsigned DestIndex =
26234 cast<ConstantSDNode>(N->getOperand(2))->getZExtValue() >> 6;
26236 Ld = NarrowVectorLoadToElement(cast<LoadSDNode>(Ld), DestIndex, DAG);
26238 // Create this as a scalar to vector to match the instruction pattern.
26239 SDValue LoadScalarToVector = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Ld);
26240 // countS bits are ignored when loading from memory on insertps, which
26241 // means we don't need to explicitly set them to 0.
26242 return DAG.getNode(X86ISD::INSERTPS, dl, VT, N->getOperand(0),
26243 LoadScalarToVector, N->getOperand(2));
26248 static SDValue PerformBLENDICombine(SDNode *N, SelectionDAG &DAG) {
26249 SDValue V0 = N->getOperand(0);
26250 SDValue V1 = N->getOperand(1);
26252 EVT VT = N->getValueType(0);
26254 // Canonicalize a v2f64 blend with a mask of 2 by swapping the vector
26255 // operands and changing the mask to 1. This saves us a bunch of
26256 // pattern-matching possibilities related to scalar math ops in SSE/AVX.
26257 // x86InstrInfo knows how to commute this back after instruction selection
26258 // if it would help register allocation.
26260 // TODO: If optimizing for size or a processor that doesn't suffer from
26261 // partial register update stalls, this should be transformed into a MOVSD
26262 // instruction because a MOVSD is 1-2 bytes smaller than a BLENDPD.
26264 if (VT == MVT::v2f64)
26265 if (auto *Mask = dyn_cast<ConstantSDNode>(N->getOperand(2)))
26266 if (Mask->getZExtValue() == 2 && !isShuffleFoldableLoad(V0)) {
26267 SDValue NewMask = DAG.getConstant(1, DL, MVT::i8);
26268 return DAG.getNode(X86ISD::BLENDI, DL, VT, V1, V0, NewMask);
26274 // Helper function of PerformSETCCCombine. It is to materialize "setb reg"
26275 // as "sbb reg,reg", since it can be extended without zext and produces
26276 // an all-ones bit which is more useful than 0/1 in some cases.
26277 static SDValue MaterializeSETB(SDLoc DL, SDValue EFLAGS, SelectionDAG &DAG,
26280 return DAG.getNode(ISD::AND, DL, VT,
26281 DAG.getNode(X86ISD::SETCC_CARRY, DL, MVT::i8,
26282 DAG.getConstant(X86::COND_B, DL, MVT::i8),
26284 DAG.getConstant(1, DL, VT));
26285 assert (VT == MVT::i1 && "Unexpected type for SECCC node");
26286 return DAG.getNode(ISD::TRUNCATE, DL, MVT::i1,
26287 DAG.getNode(X86ISD::SETCC_CARRY, DL, MVT::i8,
26288 DAG.getConstant(X86::COND_B, DL, MVT::i8),
26292 // Optimize RES = X86ISD::SETCC CONDCODE, EFLAG_INPUT
26293 static SDValue PerformSETCCCombine(SDNode *N, SelectionDAG &DAG,
26294 TargetLowering::DAGCombinerInfo &DCI,
26295 const X86Subtarget *Subtarget) {
26297 X86::CondCode CC = X86::CondCode(N->getConstantOperandVal(0));
26298 SDValue EFLAGS = N->getOperand(1);
26300 if (CC == X86::COND_A) {
26301 // Try to convert COND_A into COND_B in an attempt to facilitate
26302 // materializing "setb reg".
26304 // Do not flip "e > c", where "c" is a constant, because Cmp instruction
26305 // cannot take an immediate as its first operand.
26307 if (EFLAGS.getOpcode() == X86ISD::SUB && EFLAGS.hasOneUse() &&
26308 EFLAGS.getValueType().isInteger() &&
26309 !isa<ConstantSDNode>(EFLAGS.getOperand(1))) {
26310 SDValue NewSub = DAG.getNode(X86ISD::SUB, SDLoc(EFLAGS),
26311 EFLAGS.getNode()->getVTList(),
26312 EFLAGS.getOperand(1), EFLAGS.getOperand(0));
26313 SDValue NewEFLAGS = SDValue(NewSub.getNode(), EFLAGS.getResNo());
26314 return MaterializeSETB(DL, NewEFLAGS, DAG, N->getSimpleValueType(0));
26318 // Materialize "setb reg" as "sbb reg,reg", since it can be extended without
26319 // a zext and produces an all-ones bit which is more useful than 0/1 in some
26321 if (CC == X86::COND_B)
26322 return MaterializeSETB(DL, EFLAGS, DAG, N->getSimpleValueType(0));
26324 if (SDValue Flags = checkBoolTestSetCCCombine(EFLAGS, CC)) {
26325 SDValue Cond = DAG.getConstant(CC, DL, MVT::i8);
26326 return DAG.getNode(X86ISD::SETCC, DL, N->getVTList(), Cond, Flags);
26332 // Optimize branch condition evaluation.
26334 static SDValue PerformBrCondCombine(SDNode *N, SelectionDAG &DAG,
26335 TargetLowering::DAGCombinerInfo &DCI,
26336 const X86Subtarget *Subtarget) {
26338 SDValue Chain = N->getOperand(0);
26339 SDValue Dest = N->getOperand(1);
26340 SDValue EFLAGS = N->getOperand(3);
26341 X86::CondCode CC = X86::CondCode(N->getConstantOperandVal(2));
26343 if (SDValue Flags = checkBoolTestSetCCCombine(EFLAGS, CC)) {
26344 SDValue Cond = DAG.getConstant(CC, DL, MVT::i8);
26345 return DAG.getNode(X86ISD::BRCOND, DL, N->getVTList(), Chain, Dest, Cond,
26352 static SDValue performVectorCompareAndMaskUnaryOpCombine(SDNode *N,
26353 SelectionDAG &DAG) {
26354 // Take advantage of vector comparisons producing 0 or -1 in each lane to
26355 // optimize away operation when it's from a constant.
26357 // The general transformation is:
26358 // UNARYOP(AND(VECTOR_CMP(x,y), constant)) -->
26359 // AND(VECTOR_CMP(x,y), constant2)
26360 // constant2 = UNARYOP(constant)
26362 // Early exit if this isn't a vector operation, the operand of the
26363 // unary operation isn't a bitwise AND, or if the sizes of the operations
26364 // aren't the same.
26365 EVT VT = N->getValueType(0);
26366 if (!VT.isVector() || N->getOperand(0)->getOpcode() != ISD::AND ||
26367 N->getOperand(0)->getOperand(0)->getOpcode() != ISD::SETCC ||
26368 VT.getSizeInBits() != N->getOperand(0)->getValueType(0).getSizeInBits())
26371 // Now check that the other operand of the AND is a constant. We could
26372 // make the transformation for non-constant splats as well, but it's unclear
26373 // that would be a benefit as it would not eliminate any operations, just
26374 // perform one more step in scalar code before moving to the vector unit.
26375 if (BuildVectorSDNode *BV =
26376 dyn_cast<BuildVectorSDNode>(N->getOperand(0)->getOperand(1))) {
26377 // Bail out if the vector isn't a constant.
26378 if (!BV->isConstant())
26381 // Everything checks out. Build up the new and improved node.
26383 EVT IntVT = BV->getValueType(0);
26384 // Create a new constant of the appropriate type for the transformed
26386 SDValue SourceConst = DAG.getNode(N->getOpcode(), DL, VT, SDValue(BV, 0));
26387 // The AND node needs bitcasts to/from an integer vector type around it.
26388 SDValue MaskConst = DAG.getBitcast(IntVT, SourceConst);
26389 SDValue NewAnd = DAG.getNode(ISD::AND, DL, IntVT,
26390 N->getOperand(0)->getOperand(0), MaskConst);
26391 SDValue Res = DAG.getBitcast(VT, NewAnd);
26398 static SDValue PerformUINT_TO_FPCombine(SDNode *N, SelectionDAG &DAG,
26399 const X86Subtarget *Subtarget) {
26400 SDValue Op0 = N->getOperand(0);
26401 EVT VT = N->getValueType(0);
26402 EVT InVT = Op0.getValueType();
26403 EVT InSVT = InVT.getScalarType();
26404 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
26406 // UINT_TO_FP(vXi8) -> SINT_TO_FP(ZEXT(vXi8 to vXi32))
26407 // UINT_TO_FP(vXi16) -> SINT_TO_FP(ZEXT(vXi16 to vXi32))
26408 if (InVT.isVector() && (InSVT == MVT::i8 || InSVT == MVT::i16)) {
26410 EVT DstVT = EVT::getVectorVT(*DAG.getContext(), MVT::i32,
26411 InVT.getVectorNumElements());
26412 SDValue P = DAG.getNode(ISD::ZERO_EXTEND, dl, DstVT, Op0);
26414 if (TLI.isOperationLegal(ISD::UINT_TO_FP, DstVT))
26415 return DAG.getNode(ISD::UINT_TO_FP, dl, VT, P);
26417 return DAG.getNode(ISD::SINT_TO_FP, dl, VT, P);
26423 static SDValue PerformSINT_TO_FPCombine(SDNode *N, SelectionDAG &DAG,
26424 const X86Subtarget *Subtarget) {
26425 // First try to optimize away the conversion entirely when it's
26426 // conditionally from a constant. Vectors only.
26427 if (SDValue Res = performVectorCompareAndMaskUnaryOpCombine(N, DAG))
26430 // Now move on to more general possibilities.
26431 SDValue Op0 = N->getOperand(0);
26432 EVT VT = N->getValueType(0);
26433 EVT InVT = Op0.getValueType();
26434 EVT InSVT = InVT.getScalarType();
26436 // SINT_TO_FP(vXi8) -> SINT_TO_FP(SEXT(vXi8 to vXi32))
26437 // SINT_TO_FP(vXi16) -> SINT_TO_FP(SEXT(vXi16 to vXi32))
26438 if (InVT.isVector() && (InSVT == MVT::i8 || InSVT == MVT::i16)) {
26440 EVT DstVT = EVT::getVectorVT(*DAG.getContext(), MVT::i32,
26441 InVT.getVectorNumElements());
26442 SDValue P = DAG.getNode(ISD::SIGN_EXTEND, dl, DstVT, Op0);
26443 return DAG.getNode(ISD::SINT_TO_FP, dl, VT, P);
26446 // Transform (SINT_TO_FP (i64 ...)) into an x87 operation if we have
26447 // a 32-bit target where SSE doesn't support i64->FP operations.
26448 if (Op0.getOpcode() == ISD::LOAD) {
26449 LoadSDNode *Ld = cast<LoadSDNode>(Op0.getNode());
26450 EVT LdVT = Ld->getValueType(0);
26452 // This transformation is not supported if the result type is f16
26453 if (VT == MVT::f16)
26456 if (!Ld->isVolatile() && !VT.isVector() &&
26457 ISD::isNON_EXTLoad(Op0.getNode()) && Op0.hasOneUse() &&
26458 !Subtarget->is64Bit() && LdVT == MVT::i64) {
26459 SDValue FILDChain = Subtarget->getTargetLowering()->BuildFILD(
26460 SDValue(N, 0), LdVT, Ld->getChain(), Op0, DAG);
26461 DAG.ReplaceAllUsesOfValueWith(Op0.getValue(1), FILDChain.getValue(1));
26468 // Optimize RES, EFLAGS = X86ISD::ADC LHS, RHS, EFLAGS
26469 static SDValue PerformADCCombine(SDNode *N, SelectionDAG &DAG,
26470 X86TargetLowering::DAGCombinerInfo &DCI) {
26471 // If the LHS and RHS of the ADC node are zero, then it can't overflow and
26472 // the result is either zero or one (depending on the input carry bit).
26473 // Strength reduce this down to a "set on carry" aka SETCC_CARRY&1.
26474 if (X86::isZeroNode(N->getOperand(0)) &&
26475 X86::isZeroNode(N->getOperand(1)) &&
26476 // We don't have a good way to replace an EFLAGS use, so only do this when
26478 SDValue(N, 1).use_empty()) {
26480 EVT VT = N->getValueType(0);
26481 SDValue CarryOut = DAG.getConstant(0, DL, N->getValueType(1));
26482 SDValue Res1 = DAG.getNode(ISD::AND, DL, VT,
26483 DAG.getNode(X86ISD::SETCC_CARRY, DL, VT,
26484 DAG.getConstant(X86::COND_B, DL,
26487 DAG.getConstant(1, DL, VT));
26488 return DCI.CombineTo(N, Res1, CarryOut);
26494 // fold (add Y, (sete X, 0)) -> adc 0, Y
26495 // (add Y, (setne X, 0)) -> sbb -1, Y
26496 // (sub (sete X, 0), Y) -> sbb 0, Y
26497 // (sub (setne X, 0), Y) -> adc -1, Y
26498 static SDValue OptimizeConditionalInDecrement(SDNode *N, SelectionDAG &DAG) {
26501 // Look through ZExts.
26502 SDValue Ext = N->getOperand(N->getOpcode() == ISD::SUB ? 1 : 0);
26503 if (Ext.getOpcode() != ISD::ZERO_EXTEND || !Ext.hasOneUse())
26506 SDValue SetCC = Ext.getOperand(0);
26507 if (SetCC.getOpcode() != X86ISD::SETCC || !SetCC.hasOneUse())
26510 X86::CondCode CC = (X86::CondCode)SetCC.getConstantOperandVal(0);
26511 if (CC != X86::COND_E && CC != X86::COND_NE)
26514 SDValue Cmp = SetCC.getOperand(1);
26515 if (Cmp.getOpcode() != X86ISD::CMP || !Cmp.hasOneUse() ||
26516 !X86::isZeroNode(Cmp.getOperand(1)) ||
26517 !Cmp.getOperand(0).getValueType().isInteger())
26520 SDValue CmpOp0 = Cmp.getOperand(0);
26521 SDValue NewCmp = DAG.getNode(X86ISD::CMP, DL, MVT::i32, CmpOp0,
26522 DAG.getConstant(1, DL, CmpOp0.getValueType()));
26524 SDValue OtherVal = N->getOperand(N->getOpcode() == ISD::SUB ? 0 : 1);
26525 if (CC == X86::COND_NE)
26526 return DAG.getNode(N->getOpcode() == ISD::SUB ? X86ISD::ADC : X86ISD::SBB,
26527 DL, OtherVal.getValueType(), OtherVal,
26528 DAG.getConstant(-1ULL, DL, OtherVal.getValueType()),
26530 return DAG.getNode(N->getOpcode() == ISD::SUB ? X86ISD::SBB : X86ISD::ADC,
26531 DL, OtherVal.getValueType(), OtherVal,
26532 DAG.getConstant(0, DL, OtherVal.getValueType()), NewCmp);
26535 /// PerformADDCombine - Do target-specific dag combines on integer adds.
26536 static SDValue PerformAddCombine(SDNode *N, SelectionDAG &DAG,
26537 const X86Subtarget *Subtarget) {
26538 EVT VT = N->getValueType(0);
26539 SDValue Op0 = N->getOperand(0);
26540 SDValue Op1 = N->getOperand(1);
26542 // Try to synthesize horizontal adds from adds of shuffles.
26543 if (((Subtarget->hasSSSE3() && (VT == MVT::v8i16 || VT == MVT::v4i32)) ||
26544 (Subtarget->hasInt256() && (VT == MVT::v16i16 || VT == MVT::v8i32))) &&
26545 isHorizontalBinOp(Op0, Op1, true))
26546 return DAG.getNode(X86ISD::HADD, SDLoc(N), VT, Op0, Op1);
26548 return OptimizeConditionalInDecrement(N, DAG);
26551 static SDValue PerformSubCombine(SDNode *N, SelectionDAG &DAG,
26552 const X86Subtarget *Subtarget) {
26553 SDValue Op0 = N->getOperand(0);
26554 SDValue Op1 = N->getOperand(1);
26556 // X86 can't encode an immediate LHS of a sub. See if we can push the
26557 // negation into a preceding instruction.
26558 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op0)) {
26559 // If the RHS of the sub is a XOR with one use and a constant, invert the
26560 // immediate. Then add one to the LHS of the sub so we can turn
26561 // X-Y -> X+~Y+1, saving one register.
26562 if (Op1->hasOneUse() && Op1.getOpcode() == ISD::XOR &&
26563 isa<ConstantSDNode>(Op1.getOperand(1))) {
26564 APInt XorC = cast<ConstantSDNode>(Op1.getOperand(1))->getAPIntValue();
26565 EVT VT = Op0.getValueType();
26566 SDValue NewXor = DAG.getNode(ISD::XOR, SDLoc(Op1), VT,
26568 DAG.getConstant(~XorC, SDLoc(Op1), VT));
26569 return DAG.getNode(ISD::ADD, SDLoc(N), VT, NewXor,
26570 DAG.getConstant(C->getAPIntValue() + 1, SDLoc(N), VT));
26574 // Try to synthesize horizontal adds from adds of shuffles.
26575 EVT VT = N->getValueType(0);
26576 if (((Subtarget->hasSSSE3() && (VT == MVT::v8i16 || VT == MVT::v4i32)) ||
26577 (Subtarget->hasInt256() && (VT == MVT::v16i16 || VT == MVT::v8i32))) &&
26578 isHorizontalBinOp(Op0, Op1, true))
26579 return DAG.getNode(X86ISD::HSUB, SDLoc(N), VT, Op0, Op1);
26581 return OptimizeConditionalInDecrement(N, DAG);
26584 /// performVZEXTCombine - Performs build vector combines
26585 static SDValue performVZEXTCombine(SDNode *N, SelectionDAG &DAG,
26586 TargetLowering::DAGCombinerInfo &DCI,
26587 const X86Subtarget *Subtarget) {
26589 MVT VT = N->getSimpleValueType(0);
26590 SDValue Op = N->getOperand(0);
26591 MVT OpVT = Op.getSimpleValueType();
26592 MVT OpEltVT = OpVT.getVectorElementType();
26593 unsigned InputBits = OpEltVT.getSizeInBits() * VT.getVectorNumElements();
26595 // (vzext (bitcast (vzext (x)) -> (vzext x)
26597 while (V.getOpcode() == ISD::BITCAST)
26598 V = V.getOperand(0);
26600 if (V != Op && V.getOpcode() == X86ISD::VZEXT) {
26601 MVT InnerVT = V.getSimpleValueType();
26602 MVT InnerEltVT = InnerVT.getVectorElementType();
26604 // If the element sizes match exactly, we can just do one larger vzext. This
26605 // is always an exact type match as vzext operates on integer types.
26606 if (OpEltVT == InnerEltVT) {
26607 assert(OpVT == InnerVT && "Types must match for vzext!");
26608 return DAG.getNode(X86ISD::VZEXT, DL, VT, V.getOperand(0));
26611 // The only other way we can combine them is if only a single element of the
26612 // inner vzext is used in the input to the outer vzext.
26613 if (InnerEltVT.getSizeInBits() < InputBits)
26616 // In this case, the inner vzext is completely dead because we're going to
26617 // only look at bits inside of the low element. Just do the outer vzext on
26618 // a bitcast of the input to the inner.
26619 return DAG.getNode(X86ISD::VZEXT, DL, VT, DAG.getBitcast(OpVT, V));
26622 // Check if we can bypass extracting and re-inserting an element of an input
26623 // vector. Essentially:
26624 // (bitcast (sclr2vec (ext_vec_elt x))) -> (bitcast x)
26625 if (V.getOpcode() == ISD::SCALAR_TO_VECTOR &&
26626 V.getOperand(0).getOpcode() == ISD::EXTRACT_VECTOR_ELT &&
26627 V.getOperand(0).getSimpleValueType().getSizeInBits() == InputBits) {
26628 SDValue ExtractedV = V.getOperand(0);
26629 SDValue OrigV = ExtractedV.getOperand(0);
26630 if (auto *ExtractIdx = dyn_cast<ConstantSDNode>(ExtractedV.getOperand(1)))
26631 if (ExtractIdx->getZExtValue() == 0) {
26632 MVT OrigVT = OrigV.getSimpleValueType();
26633 // Extract a subvector if necessary...
26634 if (OrigVT.getSizeInBits() > OpVT.getSizeInBits()) {
26635 int Ratio = OrigVT.getSizeInBits() / OpVT.getSizeInBits();
26636 OrigVT = MVT::getVectorVT(OrigVT.getVectorElementType(),
26637 OrigVT.getVectorNumElements() / Ratio);
26638 OrigV = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, OrigVT, OrigV,
26639 DAG.getIntPtrConstant(0, DL));
26641 Op = DAG.getBitcast(OpVT, OrigV);
26642 return DAG.getNode(X86ISD::VZEXT, DL, VT, Op);
26649 SDValue X86TargetLowering::PerformDAGCombine(SDNode *N,
26650 DAGCombinerInfo &DCI) const {
26651 SelectionDAG &DAG = DCI.DAG;
26652 switch (N->getOpcode()) {
26654 case ISD::EXTRACT_VECTOR_ELT:
26655 return PerformEXTRACT_VECTOR_ELTCombine(N, DAG, DCI);
26658 case X86ISD::SHRUNKBLEND:
26659 return PerformSELECTCombine(N, DAG, DCI, Subtarget);
26660 case ISD::BITCAST: return PerformBITCASTCombine(N, DAG, Subtarget);
26661 case X86ISD::CMOV: return PerformCMOVCombine(N, DAG, DCI, Subtarget);
26662 case ISD::ADD: return PerformAddCombine(N, DAG, Subtarget);
26663 case ISD::SUB: return PerformSubCombine(N, DAG, Subtarget);
26664 case X86ISD::ADC: return PerformADCCombine(N, DAG, DCI);
26665 case ISD::MUL: return PerformMulCombine(N, DAG, DCI);
26668 case ISD::SRL: return PerformShiftCombine(N, DAG, DCI, Subtarget);
26669 case ISD::AND: return PerformAndCombine(N, DAG, DCI, Subtarget);
26670 case ISD::OR: return PerformOrCombine(N, DAG, DCI, Subtarget);
26671 case ISD::XOR: return PerformXorCombine(N, DAG, DCI, Subtarget);
26672 case ISD::LOAD: return PerformLOADCombine(N, DAG, DCI, Subtarget);
26673 case ISD::MLOAD: return PerformMLOADCombine(N, DAG, DCI, Subtarget);
26674 case ISD::STORE: return PerformSTORECombine(N, DAG, Subtarget);
26675 case ISD::MSTORE: return PerformMSTORECombine(N, DAG, Subtarget);
26676 case ISD::SINT_TO_FP: return PerformSINT_TO_FPCombine(N, DAG, Subtarget);
26677 case ISD::UINT_TO_FP: return PerformUINT_TO_FPCombine(N, DAG, Subtarget);
26678 case ISD::FADD: return PerformFADDCombine(N, DAG, Subtarget);
26679 case ISD::FSUB: return PerformFSUBCombine(N, DAG, Subtarget);
26681 case X86ISD::FOR: return PerformFORCombine(N, DAG, Subtarget);
26683 case X86ISD::FMAX: return PerformFMinFMaxCombine(N, DAG);
26684 case X86ISD::FAND: return PerformFANDCombine(N, DAG);
26685 case X86ISD::FANDN: return PerformFANDNCombine(N, DAG);
26686 case X86ISD::BT: return PerformBTCombine(N, DAG, DCI);
26687 case X86ISD::VZEXT_MOVL: return PerformVZEXT_MOVLCombine(N, DAG);
26688 case ISD::ANY_EXTEND:
26689 case ISD::ZERO_EXTEND: return PerformZExtCombine(N, DAG, DCI, Subtarget);
26690 case ISD::SIGN_EXTEND: return PerformSExtCombine(N, DAG, DCI, Subtarget);
26691 case ISD::SIGN_EXTEND_INREG:
26692 return PerformSIGN_EXTEND_INREGCombine(N, DAG, Subtarget);
26693 case ISD::SETCC: return PerformISDSETCCCombine(N, DAG, Subtarget);
26694 case X86ISD::SETCC: return PerformSETCCCombine(N, DAG, DCI, Subtarget);
26695 case X86ISD::BRCOND: return PerformBrCondCombine(N, DAG, DCI, Subtarget);
26696 case X86ISD::VZEXT: return performVZEXTCombine(N, DAG, DCI, Subtarget);
26697 case X86ISD::SHUFP: // Handle all target specific shuffles
26698 case X86ISD::PALIGNR:
26699 case X86ISD::UNPCKH:
26700 case X86ISD::UNPCKL:
26701 case X86ISD::MOVHLPS:
26702 case X86ISD::MOVLHPS:
26703 case X86ISD::PSHUFB:
26704 case X86ISD::PSHUFD:
26705 case X86ISD::PSHUFHW:
26706 case X86ISD::PSHUFLW:
26707 case X86ISD::MOVSS:
26708 case X86ISD::MOVSD:
26709 case X86ISD::VPERMILPI:
26710 case X86ISD::VPERM2X128:
26711 case ISD::VECTOR_SHUFFLE: return PerformShuffleCombine(N, DAG, DCI,Subtarget);
26712 case ISD::FMA: return PerformFMACombine(N, DAG, Subtarget);
26713 case X86ISD::INSERTPS: {
26714 if (getTargetMachine().getOptLevel() > CodeGenOpt::None)
26715 return PerformINSERTPSCombine(N, DAG, Subtarget);
26718 case X86ISD::BLENDI: return PerformBLENDICombine(N, DAG);
26724 /// isTypeDesirableForOp - Return true if the target has native support for
26725 /// the specified value type and it is 'desirable' to use the type for the
26726 /// given node type. e.g. On x86 i16 is legal, but undesirable since i16
26727 /// instruction encodings are longer and some i16 instructions are slow.
26728 bool X86TargetLowering::isTypeDesirableForOp(unsigned Opc, EVT VT) const {
26729 if (!isTypeLegal(VT))
26731 if (VT != MVT::i16)
26738 case ISD::SIGN_EXTEND:
26739 case ISD::ZERO_EXTEND:
26740 case ISD::ANY_EXTEND:
26753 /// IsDesirableToPromoteOp - This method query the target whether it is
26754 /// beneficial for dag combiner to promote the specified node. If true, it
26755 /// should return the desired promotion type by reference.
26756 bool X86TargetLowering::IsDesirableToPromoteOp(SDValue Op, EVT &PVT) const {
26757 EVT VT = Op.getValueType();
26758 if (VT != MVT::i16)
26761 bool Promote = false;
26762 bool Commute = false;
26763 switch (Op.getOpcode()) {
26766 LoadSDNode *LD = cast<LoadSDNode>(Op);
26767 // If the non-extending load has a single use and it's not live out, then it
26768 // might be folded.
26769 if (LD->getExtensionType() == ISD::NON_EXTLOAD /*&&
26770 Op.hasOneUse()*/) {
26771 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
26772 UE = Op.getNode()->use_end(); UI != UE; ++UI) {
26773 // The only case where we'd want to promote LOAD (rather then it being
26774 // promoted as an operand is when it's only use is liveout.
26775 if (UI->getOpcode() != ISD::CopyToReg)
26782 case ISD::SIGN_EXTEND:
26783 case ISD::ZERO_EXTEND:
26784 case ISD::ANY_EXTEND:
26789 SDValue N0 = Op.getOperand(0);
26790 // Look out for (store (shl (load), x)).
26791 if (MayFoldLoad(N0) && MayFoldIntoStore(Op))
26804 SDValue N0 = Op.getOperand(0);
26805 SDValue N1 = Op.getOperand(1);
26806 if (!Commute && MayFoldLoad(N1))
26808 // Avoid disabling potential load folding opportunities.
26809 if (MayFoldLoad(N0) && (!isa<ConstantSDNode>(N1) || MayFoldIntoStore(Op)))
26811 if (MayFoldLoad(N1) && (!isa<ConstantSDNode>(N0) || MayFoldIntoStore(Op)))
26821 //===----------------------------------------------------------------------===//
26822 // X86 Inline Assembly Support
26823 //===----------------------------------------------------------------------===//
26825 // Helper to match a string separated by whitespace.
26826 static bool matchAsm(StringRef S, ArrayRef<const char *> Pieces) {
26827 S = S.substr(S.find_first_not_of(" \t")); // Skip leading whitespace.
26829 for (StringRef Piece : Pieces) {
26830 if (!S.startswith(Piece)) // Check if the piece matches.
26833 S = S.substr(Piece.size());
26834 StringRef::size_type Pos = S.find_first_not_of(" \t");
26835 if (Pos == 0) // We matched a prefix.
26844 static bool clobbersFlagRegisters(const SmallVector<StringRef, 4> &AsmPieces) {
26846 if (AsmPieces.size() == 3 || AsmPieces.size() == 4) {
26847 if (std::count(AsmPieces.begin(), AsmPieces.end(), "~{cc}") &&
26848 std::count(AsmPieces.begin(), AsmPieces.end(), "~{flags}") &&
26849 std::count(AsmPieces.begin(), AsmPieces.end(), "~{fpsr}")) {
26851 if (AsmPieces.size() == 3)
26853 else if (std::count(AsmPieces.begin(), AsmPieces.end(), "~{dirflag}"))
26860 bool X86TargetLowering::ExpandInlineAsm(CallInst *CI) const {
26861 InlineAsm *IA = cast<InlineAsm>(CI->getCalledValue());
26863 std::string AsmStr = IA->getAsmString();
26865 IntegerType *Ty = dyn_cast<IntegerType>(CI->getType());
26866 if (!Ty || Ty->getBitWidth() % 16 != 0)
26869 // TODO: should remove alternatives from the asmstring: "foo {a|b}" -> "foo a"
26870 SmallVector<StringRef, 4> AsmPieces;
26871 SplitString(AsmStr, AsmPieces, ";\n");
26873 switch (AsmPieces.size()) {
26874 default: return false;
26876 // FIXME: this should verify that we are targeting a 486 or better. If not,
26877 // we will turn this bswap into something that will be lowered to logical
26878 // ops instead of emitting the bswap asm. For now, we don't support 486 or
26879 // lower so don't worry about this.
26881 if (matchAsm(AsmPieces[0], {"bswap", "$0"}) ||
26882 matchAsm(AsmPieces[0], {"bswapl", "$0"}) ||
26883 matchAsm(AsmPieces[0], {"bswapq", "$0"}) ||
26884 matchAsm(AsmPieces[0], {"bswap", "${0:q}"}) ||
26885 matchAsm(AsmPieces[0], {"bswapl", "${0:q}"}) ||
26886 matchAsm(AsmPieces[0], {"bswapq", "${0:q}"})) {
26887 // No need to check constraints, nothing other than the equivalent of
26888 // "=r,0" would be valid here.
26889 return IntrinsicLowering::LowerToByteSwap(CI);
26892 // rorw $$8, ${0:w} --> llvm.bswap.i16
26893 if (CI->getType()->isIntegerTy(16) &&
26894 IA->getConstraintString().compare(0, 5, "=r,0,") == 0 &&
26895 (matchAsm(AsmPieces[0], {"rorw", "$$8,", "${0:w}"}) ||
26896 matchAsm(AsmPieces[0], {"rolw", "$$8,", "${0:w}"}))) {
26898 StringRef ConstraintsStr = IA->getConstraintString();
26899 SplitString(StringRef(ConstraintsStr).substr(5), AsmPieces, ",");
26900 array_pod_sort(AsmPieces.begin(), AsmPieces.end());
26901 if (clobbersFlagRegisters(AsmPieces))
26902 return IntrinsicLowering::LowerToByteSwap(CI);
26906 if (CI->getType()->isIntegerTy(32) &&
26907 IA->getConstraintString().compare(0, 5, "=r,0,") == 0 &&
26908 matchAsm(AsmPieces[0], {"rorw", "$$8,", "${0:w}"}) &&
26909 matchAsm(AsmPieces[1], {"rorl", "$$16,", "$0"}) &&
26910 matchAsm(AsmPieces[2], {"rorw", "$$8,", "${0:w}"})) {
26912 StringRef ConstraintsStr = IA->getConstraintString();
26913 SplitString(StringRef(ConstraintsStr).substr(5), AsmPieces, ",");
26914 array_pod_sort(AsmPieces.begin(), AsmPieces.end());
26915 if (clobbersFlagRegisters(AsmPieces))
26916 return IntrinsicLowering::LowerToByteSwap(CI);
26919 if (CI->getType()->isIntegerTy(64)) {
26920 InlineAsm::ConstraintInfoVector Constraints = IA->ParseConstraints();
26921 if (Constraints.size() >= 2 &&
26922 Constraints[0].Codes.size() == 1 && Constraints[0].Codes[0] == "A" &&
26923 Constraints[1].Codes.size() == 1 && Constraints[1].Codes[0] == "0") {
26924 // bswap %eax / bswap %edx / xchgl %eax, %edx -> llvm.bswap.i64
26925 if (matchAsm(AsmPieces[0], {"bswap", "%eax"}) &&
26926 matchAsm(AsmPieces[1], {"bswap", "%edx"}) &&
26927 matchAsm(AsmPieces[2], {"xchgl", "%eax,", "%edx"}))
26928 return IntrinsicLowering::LowerToByteSwap(CI);
26936 /// getConstraintType - Given a constraint letter, return the type of
26937 /// constraint it is for this target.
26938 X86TargetLowering::ConstraintType
26939 X86TargetLowering::getConstraintType(StringRef Constraint) const {
26940 if (Constraint.size() == 1) {
26941 switch (Constraint[0]) {
26952 return C_RegisterClass;
26976 return TargetLowering::getConstraintType(Constraint);
26979 /// Examine constraint type and operand type and determine a weight value.
26980 /// This object must already have been set up with the operand type
26981 /// and the current alternative constraint selected.
26982 TargetLowering::ConstraintWeight
26983 X86TargetLowering::getSingleConstraintMatchWeight(
26984 AsmOperandInfo &info, const char *constraint) const {
26985 ConstraintWeight weight = CW_Invalid;
26986 Value *CallOperandVal = info.CallOperandVal;
26987 // If we don't have a value, we can't do a match,
26988 // but allow it at the lowest weight.
26989 if (!CallOperandVal)
26991 Type *type = CallOperandVal->getType();
26992 // Look at the constraint type.
26993 switch (*constraint) {
26995 weight = TargetLowering::getSingleConstraintMatchWeight(info, constraint);
27006 if (CallOperandVal->getType()->isIntegerTy())
27007 weight = CW_SpecificReg;
27012 if (type->isFloatingPointTy())
27013 weight = CW_SpecificReg;
27016 if (type->isX86_MMXTy() && Subtarget->hasMMX())
27017 weight = CW_SpecificReg;
27021 if (((type->getPrimitiveSizeInBits() == 128) && Subtarget->hasSSE1()) ||
27022 ((type->getPrimitiveSizeInBits() == 256) && Subtarget->hasFp256()))
27023 weight = CW_Register;
27026 if (ConstantInt *C = dyn_cast<ConstantInt>(info.CallOperandVal)) {
27027 if (C->getZExtValue() <= 31)
27028 weight = CW_Constant;
27032 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
27033 if (C->getZExtValue() <= 63)
27034 weight = CW_Constant;
27038 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
27039 if ((C->getSExtValue() >= -0x80) && (C->getSExtValue() <= 0x7f))
27040 weight = CW_Constant;
27044 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
27045 if ((C->getZExtValue() == 0xff) || (C->getZExtValue() == 0xffff))
27046 weight = CW_Constant;
27050 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
27051 if (C->getZExtValue() <= 3)
27052 weight = CW_Constant;
27056 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
27057 if (C->getZExtValue() <= 0xff)
27058 weight = CW_Constant;
27063 if (isa<ConstantFP>(CallOperandVal)) {
27064 weight = CW_Constant;
27068 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
27069 if ((C->getSExtValue() >= -0x80000000LL) &&
27070 (C->getSExtValue() <= 0x7fffffffLL))
27071 weight = CW_Constant;
27075 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
27076 if (C->getZExtValue() <= 0xffffffff)
27077 weight = CW_Constant;
27084 /// LowerXConstraint - try to replace an X constraint, which matches anything,
27085 /// with another that has more specific requirements based on the type of the
27086 /// corresponding operand.
27087 const char *X86TargetLowering::
27088 LowerXConstraint(EVT ConstraintVT) const {
27089 // FP X constraints get lowered to SSE1/2 registers if available, otherwise
27090 // 'f' like normal targets.
27091 if (ConstraintVT.isFloatingPoint()) {
27092 if (Subtarget->hasSSE2())
27094 if (Subtarget->hasSSE1())
27098 return TargetLowering::LowerXConstraint(ConstraintVT);
27101 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
27102 /// vector. If it is invalid, don't add anything to Ops.
27103 void X86TargetLowering::LowerAsmOperandForConstraint(SDValue Op,
27104 std::string &Constraint,
27105 std::vector<SDValue>&Ops,
27106 SelectionDAG &DAG) const {
27109 // Only support length 1 constraints for now.
27110 if (Constraint.length() > 1) return;
27112 char ConstraintLetter = Constraint[0];
27113 switch (ConstraintLetter) {
27116 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
27117 if (C->getZExtValue() <= 31) {
27118 Result = DAG.getTargetConstant(C->getZExtValue(), SDLoc(Op),
27119 Op.getValueType());
27125 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
27126 if (C->getZExtValue() <= 63) {
27127 Result = DAG.getTargetConstant(C->getZExtValue(), SDLoc(Op),
27128 Op.getValueType());
27134 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
27135 if (isInt<8>(C->getSExtValue())) {
27136 Result = DAG.getTargetConstant(C->getZExtValue(), SDLoc(Op),
27137 Op.getValueType());
27143 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
27144 if (C->getZExtValue() == 0xff || C->getZExtValue() == 0xffff ||
27145 (Subtarget->is64Bit() && C->getZExtValue() == 0xffffffff)) {
27146 Result = DAG.getTargetConstant(C->getSExtValue(), SDLoc(Op),
27147 Op.getValueType());
27153 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
27154 if (C->getZExtValue() <= 3) {
27155 Result = DAG.getTargetConstant(C->getZExtValue(), SDLoc(Op),
27156 Op.getValueType());
27162 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
27163 if (C->getZExtValue() <= 255) {
27164 Result = DAG.getTargetConstant(C->getZExtValue(), SDLoc(Op),
27165 Op.getValueType());
27171 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
27172 if (C->getZExtValue() <= 127) {
27173 Result = DAG.getTargetConstant(C->getZExtValue(), SDLoc(Op),
27174 Op.getValueType());
27180 // 32-bit signed value
27181 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
27182 if (ConstantInt::isValueValidForType(Type::getInt32Ty(*DAG.getContext()),
27183 C->getSExtValue())) {
27184 // Widen to 64 bits here to get it sign extended.
27185 Result = DAG.getTargetConstant(C->getSExtValue(), SDLoc(Op), MVT::i64);
27188 // FIXME gcc accepts some relocatable values here too, but only in certain
27189 // memory models; it's complicated.
27194 // 32-bit unsigned value
27195 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
27196 if (ConstantInt::isValueValidForType(Type::getInt32Ty(*DAG.getContext()),
27197 C->getZExtValue())) {
27198 Result = DAG.getTargetConstant(C->getZExtValue(), SDLoc(Op),
27199 Op.getValueType());
27203 // FIXME gcc accepts some relocatable values here too, but only in certain
27204 // memory models; it's complicated.
27208 // Literal immediates are always ok.
27209 if (ConstantSDNode *CST = dyn_cast<ConstantSDNode>(Op)) {
27210 // Widen to 64 bits here to get it sign extended.
27211 Result = DAG.getTargetConstant(CST->getSExtValue(), SDLoc(Op), MVT::i64);
27215 // In any sort of PIC mode addresses need to be computed at runtime by
27216 // adding in a register or some sort of table lookup. These can't
27217 // be used as immediates.
27218 if (Subtarget->isPICStyleGOT() || Subtarget->isPICStyleStubPIC())
27221 // If we are in non-pic codegen mode, we allow the address of a global (with
27222 // an optional displacement) to be used with 'i'.
27223 GlobalAddressSDNode *GA = nullptr;
27224 int64_t Offset = 0;
27226 // Match either (GA), (GA+C), (GA+C1+C2), etc.
27228 if ((GA = dyn_cast<GlobalAddressSDNode>(Op))) {
27229 Offset += GA->getOffset();
27231 } else if (Op.getOpcode() == ISD::ADD) {
27232 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
27233 Offset += C->getZExtValue();
27234 Op = Op.getOperand(0);
27237 } else if (Op.getOpcode() == ISD::SUB) {
27238 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
27239 Offset += -C->getZExtValue();
27240 Op = Op.getOperand(0);
27245 // Otherwise, this isn't something we can handle, reject it.
27249 const GlobalValue *GV = GA->getGlobal();
27250 // If we require an extra load to get this address, as in PIC mode, we
27251 // can't accept it.
27252 if (isGlobalStubReference(
27253 Subtarget->ClassifyGlobalReference(GV, DAG.getTarget())))
27256 Result = DAG.getTargetGlobalAddress(GV, SDLoc(Op),
27257 GA->getValueType(0), Offset);
27262 if (Result.getNode()) {
27263 Ops.push_back(Result);
27266 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG);
27269 std::pair<unsigned, const TargetRegisterClass *>
27270 X86TargetLowering::getRegForInlineAsmConstraint(const TargetRegisterInfo *TRI,
27271 StringRef Constraint,
27273 // First, see if this is a constraint that directly corresponds to an LLVM
27275 if (Constraint.size() == 1) {
27276 // GCC Constraint Letters
27277 switch (Constraint[0]) {
27279 // TODO: Slight differences here in allocation order and leaving
27280 // RIP in the class. Do they matter any more here than they do
27281 // in the normal allocation?
27282 case 'q': // GENERAL_REGS in 64-bit mode, Q_REGS in 32-bit mode.
27283 if (Subtarget->is64Bit()) {
27284 if (VT == MVT::i32 || VT == MVT::f32)
27285 return std::make_pair(0U, &X86::GR32RegClass);
27286 if (VT == MVT::i16)
27287 return std::make_pair(0U, &X86::GR16RegClass);
27288 if (VT == MVT::i8 || VT == MVT::i1)
27289 return std::make_pair(0U, &X86::GR8RegClass);
27290 if (VT == MVT::i64 || VT == MVT::f64)
27291 return std::make_pair(0U, &X86::GR64RegClass);
27294 // 32-bit fallthrough
27295 case 'Q': // Q_REGS
27296 if (VT == MVT::i32 || VT == MVT::f32)
27297 return std::make_pair(0U, &X86::GR32_ABCDRegClass);
27298 if (VT == MVT::i16)
27299 return std::make_pair(0U, &X86::GR16_ABCDRegClass);
27300 if (VT == MVT::i8 || VT == MVT::i1)
27301 return std::make_pair(0U, &X86::GR8_ABCD_LRegClass);
27302 if (VT == MVT::i64)
27303 return std::make_pair(0U, &X86::GR64_ABCDRegClass);
27305 case 'r': // GENERAL_REGS
27306 case 'l': // INDEX_REGS
27307 if (VT == MVT::i8 || VT == MVT::i1)
27308 return std::make_pair(0U, &X86::GR8RegClass);
27309 if (VT == MVT::i16)
27310 return std::make_pair(0U, &X86::GR16RegClass);
27311 if (VT == MVT::i32 || VT == MVT::f32 || !Subtarget->is64Bit())
27312 return std::make_pair(0U, &X86::GR32RegClass);
27313 return std::make_pair(0U, &X86::GR64RegClass);
27314 case 'R': // LEGACY_REGS
27315 if (VT == MVT::i8 || VT == MVT::i1)
27316 return std::make_pair(0U, &X86::GR8_NOREXRegClass);
27317 if (VT == MVT::i16)
27318 return std::make_pair(0U, &X86::GR16_NOREXRegClass);
27319 if (VT == MVT::i32 || !Subtarget->is64Bit())
27320 return std::make_pair(0U, &X86::GR32_NOREXRegClass);
27321 return std::make_pair(0U, &X86::GR64_NOREXRegClass);
27322 case 'f': // FP Stack registers.
27323 // If SSE is enabled for this VT, use f80 to ensure the isel moves the
27324 // value to the correct fpstack register class.
27325 if (VT == MVT::f32 && !isScalarFPTypeInSSEReg(VT))
27326 return std::make_pair(0U, &X86::RFP32RegClass);
27327 if (VT == MVT::f64 && !isScalarFPTypeInSSEReg(VT))
27328 return std::make_pair(0U, &X86::RFP64RegClass);
27329 return std::make_pair(0U, &X86::RFP80RegClass);
27330 case 'y': // MMX_REGS if MMX allowed.
27331 if (!Subtarget->hasMMX()) break;
27332 return std::make_pair(0U, &X86::VR64RegClass);
27333 case 'Y': // SSE_REGS if SSE2 allowed
27334 if (!Subtarget->hasSSE2()) break;
27336 case 'x': // SSE_REGS if SSE1 allowed or AVX_REGS if AVX allowed
27337 if (!Subtarget->hasSSE1()) break;
27339 switch (VT.SimpleTy) {
27341 // Scalar SSE types.
27344 return std::make_pair(0U, &X86::FR32RegClass);
27347 return std::make_pair(0U, &X86::FR64RegClass);
27355 return std::make_pair(0U, &X86::VR128RegClass);
27363 return std::make_pair(0U, &X86::VR256RegClass);
27368 return std::make_pair(0U, &X86::VR512RegClass);
27374 // Use the default implementation in TargetLowering to convert the register
27375 // constraint into a member of a register class.
27376 std::pair<unsigned, const TargetRegisterClass*> Res;
27377 Res = TargetLowering::getRegForInlineAsmConstraint(TRI, Constraint, VT);
27379 // Not found as a standard register?
27381 // Map st(0) -> st(7) -> ST0
27382 if (Constraint.size() == 7 && Constraint[0] == '{' &&
27383 tolower(Constraint[1]) == 's' &&
27384 tolower(Constraint[2]) == 't' &&
27385 Constraint[3] == '(' &&
27386 (Constraint[4] >= '0' && Constraint[4] <= '7') &&
27387 Constraint[5] == ')' &&
27388 Constraint[6] == '}') {
27390 Res.first = X86::FP0+Constraint[4]-'0';
27391 Res.second = &X86::RFP80RegClass;
27395 // GCC allows "st(0)" to be called just plain "st".
27396 if (StringRef("{st}").equals_lower(Constraint)) {
27397 Res.first = X86::FP0;
27398 Res.second = &X86::RFP80RegClass;
27403 if (StringRef("{flags}").equals_lower(Constraint)) {
27404 Res.first = X86::EFLAGS;
27405 Res.second = &X86::CCRRegClass;
27409 // 'A' means EAX + EDX.
27410 if (Constraint == "A") {
27411 Res.first = X86::EAX;
27412 Res.second = &X86::GR32_ADRegClass;
27418 // Otherwise, check to see if this is a register class of the wrong value
27419 // type. For example, we want to map "{ax},i32" -> {eax}, we don't want it to
27420 // turn into {ax},{dx}.
27421 // MVT::Other is used to specify clobber names.
27422 if (Res.second->hasType(VT) || VT == MVT::Other)
27423 return Res; // Correct type already, nothing to do.
27425 // Get a matching integer of the correct size. i.e. "ax" with MVT::32 should
27426 // return "eax". This should even work for things like getting 64bit integer
27427 // registers when given an f64 type.
27428 const TargetRegisterClass *Class = Res.second;
27429 if (Class == &X86::GR8RegClass || Class == &X86::GR16RegClass ||
27430 Class == &X86::GR32RegClass || Class == &X86::GR64RegClass) {
27431 unsigned Size = VT.getSizeInBits();
27432 MVT::SimpleValueType SimpleTy = Size == 1 || Size == 8 ? MVT::i8
27433 : Size == 16 ? MVT::i16
27434 : Size == 32 ? MVT::i32
27435 : Size == 64 ? MVT::i64
27437 unsigned DestReg = getX86SubSuperRegisterOrZero(Res.first, SimpleTy);
27439 Res.first = DestReg;
27440 Res.second = SimpleTy == MVT::i8 ? &X86::GR8RegClass
27441 : SimpleTy == MVT::i16 ? &X86::GR16RegClass
27442 : SimpleTy == MVT::i32 ? &X86::GR32RegClass
27443 : &X86::GR64RegClass;
27444 assert(Res.second->contains(Res.first) && "Register in register class");
27446 // No register found/type mismatch.
27448 Res.second = nullptr;
27450 } else if (Class == &X86::FR32RegClass || Class == &X86::FR64RegClass ||
27451 Class == &X86::VR128RegClass || Class == &X86::VR256RegClass ||
27452 Class == &X86::FR32XRegClass || Class == &X86::FR64XRegClass ||
27453 Class == &X86::VR128XRegClass || Class == &X86::VR256XRegClass ||
27454 Class == &X86::VR512RegClass) {
27455 // Handle references to XMM physical registers that got mapped into the
27456 // wrong class. This can happen with constraints like {xmm0} where the
27457 // target independent register mapper will just pick the first match it can
27458 // find, ignoring the required type.
27460 if (VT == MVT::f32 || VT == MVT::i32)
27461 Res.second = &X86::FR32RegClass;
27462 else if (VT == MVT::f64 || VT == MVT::i64)
27463 Res.second = &X86::FR64RegClass;
27464 else if (X86::VR128RegClass.hasType(VT))
27465 Res.second = &X86::VR128RegClass;
27466 else if (X86::VR256RegClass.hasType(VT))
27467 Res.second = &X86::VR256RegClass;
27468 else if (X86::VR512RegClass.hasType(VT))
27469 Res.second = &X86::VR512RegClass;
27471 // Type mismatch and not a clobber: Return an error;
27473 Res.second = nullptr;
27480 int X86TargetLowering::getScalingFactorCost(const DataLayout &DL,
27481 const AddrMode &AM, Type *Ty,
27482 unsigned AS) const {
27483 // Scaling factors are not free at all.
27484 // An indexed folded instruction, i.e., inst (reg1, reg2, scale),
27485 // will take 2 allocations in the out of order engine instead of 1
27486 // for plain addressing mode, i.e. inst (reg1).
27488 // vaddps (%rsi,%drx), %ymm0, %ymm1
27489 // Requires two allocations (one for the load, one for the computation)
27491 // vaddps (%rsi), %ymm0, %ymm1
27492 // Requires just 1 allocation, i.e., freeing allocations for other operations
27493 // and having less micro operations to execute.
27495 // For some X86 architectures, this is even worse because for instance for
27496 // stores, the complex addressing mode forces the instruction to use the
27497 // "load" ports instead of the dedicated "store" port.
27498 // E.g., on Haswell:
27499 // vmovaps %ymm1, (%r8, %rdi) can use port 2 or 3.
27500 // vmovaps %ymm1, (%r8) can use port 2, 3, or 7.
27501 if (isLegalAddressingMode(DL, AM, Ty, AS))
27502 // Scale represents reg2 * scale, thus account for 1
27503 // as soon as we use a second register.
27504 return AM.Scale != 0;
27508 bool X86TargetLowering::isIntDivCheap(EVT VT, AttributeSet Attr) const {
27509 // Integer division on x86 is expensive. However, when aggressively optimizing
27510 // for code size, we prefer to use a div instruction, as it is usually smaller
27511 // than the alternative sequence.
27512 // The exception to this is vector division. Since x86 doesn't have vector
27513 // integer division, leaving the division as-is is a loss even in terms of
27514 // size, because it will have to be scalarized, while the alternative code
27515 // sequence can be performed in vector form.
27516 bool OptSize = Attr.hasAttribute(AttributeSet::FunctionIndex,
27517 Attribute::MinSize);
27518 return OptSize && !VT.isVector();
27521 void X86TargetLowering::markInRegArguments(SelectionDAG &DAG,
27522 TargetLowering::ArgListTy& Args) const {
27523 // The MCU psABI requires some arguments to be passed in-register.
27524 // For regular calls, the inreg arguments are marked by the front-end.
27525 // However, for compiler generated library calls, we have to patch this
27527 if (!Subtarget->isTargetMCU() || !Args.size())
27530 unsigned FreeRegs = 3;
27531 for (auto &Arg : Args) {
27532 // For library functions, we do not expect any fancy types.
27533 unsigned Size = DAG.getDataLayout().getTypeSizeInBits(Arg.Ty);
27534 unsigned SizeInRegs = (Size + 31) / 32;
27535 if (SizeInRegs > 2 || SizeInRegs > FreeRegs)
27538 Arg.isInReg = true;
27539 FreeRegs -= SizeInRegs;