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 (auto VT : { MVT::v16i8, MVT::v8i16, MVT::v4i32 }) {
862 setOperationAction(ISD::BUILD_VECTOR, VT, Custom);
863 setOperationAction(ISD::VECTOR_SHUFFLE, VT, Custom);
864 setOperationAction(ISD::VSELECT, VT, Custom);
865 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Custom);
868 // We support custom legalizing of sext and anyext loads for specific
869 // memory vector types which we can load as a scalar (or sequence of
870 // scalars) and extend in-register to a legal 128-bit vector type. For sext
871 // loads these must work with a single scalar load.
872 for (MVT VT : MVT::integer_vector_valuetypes()) {
873 setLoadExtAction(ISD::SEXTLOAD, VT, MVT::v4i8, Custom);
874 setLoadExtAction(ISD::SEXTLOAD, VT, MVT::v4i16, Custom);
875 setLoadExtAction(ISD::SEXTLOAD, VT, MVT::v8i8, Custom);
876 setLoadExtAction(ISD::EXTLOAD, VT, MVT::v2i8, Custom);
877 setLoadExtAction(ISD::EXTLOAD, VT, MVT::v2i16, Custom);
878 setLoadExtAction(ISD::EXTLOAD, VT, MVT::v2i32, Custom);
879 setLoadExtAction(ISD::EXTLOAD, VT, MVT::v4i8, Custom);
880 setLoadExtAction(ISD::EXTLOAD, VT, MVT::v4i16, Custom);
881 setLoadExtAction(ISD::EXTLOAD, VT, MVT::v8i8, Custom);
884 setOperationAction(ISD::BUILD_VECTOR, MVT::v2f64, Custom);
885 setOperationAction(ISD::BUILD_VECTOR, MVT::v2i64, Custom);
886 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2f64, Custom);
887 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2i64, Custom);
888 setOperationAction(ISD::VSELECT, MVT::v2f64, Custom);
889 setOperationAction(ISD::VSELECT, MVT::v2i64, Custom);
890 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2f64, Custom);
891 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2f64, Custom);
893 if (Subtarget->is64Bit()) {
894 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2i64, Custom);
895 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2i64, Custom);
898 // Promote v16i8, v8i16, v4i32 load, select, and, or, xor to v2i64.
899 for (auto VT : { MVT::v16i8, MVT::v8i16, MVT::v4i32 }) {
900 setOperationAction(ISD::AND, VT, Promote);
901 AddPromotedToType (ISD::AND, VT, MVT::v2i64);
902 setOperationAction(ISD::OR, VT, Promote);
903 AddPromotedToType (ISD::OR, VT, MVT::v2i64);
904 setOperationAction(ISD::XOR, VT, Promote);
905 AddPromotedToType (ISD::XOR, VT, MVT::v2i64);
906 setOperationAction(ISD::LOAD, VT, Promote);
907 AddPromotedToType (ISD::LOAD, VT, MVT::v2i64);
908 setOperationAction(ISD::SELECT, VT, Promote);
909 AddPromotedToType (ISD::SELECT, VT, MVT::v2i64);
912 // Custom lower v2i64 and v2f64 selects.
913 setOperationAction(ISD::LOAD, MVT::v2f64, Legal);
914 setOperationAction(ISD::LOAD, MVT::v2i64, Legal);
915 setOperationAction(ISD::SELECT, MVT::v2f64, Custom);
916 setOperationAction(ISD::SELECT, MVT::v2i64, Custom);
918 setOperationAction(ISD::FP_TO_SINT, MVT::v4i32, Legal);
919 setOperationAction(ISD::SINT_TO_FP, MVT::v4i32, Legal);
921 setOperationAction(ISD::SINT_TO_FP, MVT::v2i32, Custom);
923 setOperationAction(ISD::UINT_TO_FP, MVT::v4i8, Custom);
924 setOperationAction(ISD::UINT_TO_FP, MVT::v4i16, Custom);
925 // As there is no 64-bit GPR available, we need build a special custom
926 // sequence to convert from v2i32 to v2f32.
927 if (!Subtarget->is64Bit())
928 setOperationAction(ISD::UINT_TO_FP, MVT::v2f32, Custom);
930 setOperationAction(ISD::FP_EXTEND, MVT::v2f32, Custom);
931 setOperationAction(ISD::FP_ROUND, MVT::v2f32, Custom);
933 for (MVT VT : MVT::fp_vector_valuetypes())
934 setLoadExtAction(ISD::EXTLOAD, VT, MVT::v2f32, Legal);
936 setOperationAction(ISD::BITCAST, MVT::v2i32, Custom);
937 setOperationAction(ISD::BITCAST, MVT::v4i16, Custom);
938 setOperationAction(ISD::BITCAST, MVT::v8i8, Custom);
941 if (!Subtarget->useSoftFloat() && Subtarget->hasSSE41()) {
942 for (MVT RoundedTy : {MVT::f32, MVT::f64, MVT::v4f32, MVT::v2f64}) {
943 setOperationAction(ISD::FFLOOR, RoundedTy, Legal);
944 setOperationAction(ISD::FCEIL, RoundedTy, Legal);
945 setOperationAction(ISD::FTRUNC, RoundedTy, Legal);
946 setOperationAction(ISD::FRINT, RoundedTy, Legal);
947 setOperationAction(ISD::FNEARBYINT, RoundedTy, Legal);
950 setOperationAction(ISD::SMAX, MVT::v16i8, Legal);
951 setOperationAction(ISD::SMAX, MVT::v4i32, Legal);
952 setOperationAction(ISD::UMAX, MVT::v8i16, Legal);
953 setOperationAction(ISD::UMAX, MVT::v4i32, Legal);
954 setOperationAction(ISD::SMIN, MVT::v16i8, Legal);
955 setOperationAction(ISD::SMIN, MVT::v4i32, Legal);
956 setOperationAction(ISD::UMIN, MVT::v8i16, Legal);
957 setOperationAction(ISD::UMIN, MVT::v4i32, Legal);
959 // FIXME: Do we need to handle scalar-to-vector here?
960 setOperationAction(ISD::MUL, MVT::v4i32, Legal);
962 // We directly match byte blends in the backend as they match the VSELECT
964 setOperationAction(ISD::VSELECT, MVT::v16i8, Legal);
966 // SSE41 brings specific instructions for doing vector sign extend even in
967 // cases where we don't have SRA.
968 for (MVT VT : MVT::integer_vector_valuetypes()) {
969 setLoadExtAction(ISD::SEXTLOAD, VT, MVT::v2i8, Custom);
970 setLoadExtAction(ISD::SEXTLOAD, VT, MVT::v2i16, Custom);
971 setLoadExtAction(ISD::SEXTLOAD, VT, MVT::v2i32, Custom);
974 // SSE41 also has vector sign/zero extending loads, PMOV[SZ]X
975 setLoadExtAction(ISD::SEXTLOAD, MVT::v8i16, MVT::v8i8, Legal);
976 setLoadExtAction(ISD::SEXTLOAD, MVT::v4i32, MVT::v4i8, Legal);
977 setLoadExtAction(ISD::SEXTLOAD, MVT::v2i64, MVT::v2i8, Legal);
978 setLoadExtAction(ISD::SEXTLOAD, MVT::v4i32, MVT::v4i16, Legal);
979 setLoadExtAction(ISD::SEXTLOAD, MVT::v2i64, MVT::v2i16, Legal);
980 setLoadExtAction(ISD::SEXTLOAD, MVT::v2i64, MVT::v2i32, Legal);
982 setLoadExtAction(ISD::ZEXTLOAD, MVT::v8i16, MVT::v8i8, Legal);
983 setLoadExtAction(ISD::ZEXTLOAD, MVT::v4i32, MVT::v4i8, Legal);
984 setLoadExtAction(ISD::ZEXTLOAD, MVT::v2i64, MVT::v2i8, Legal);
985 setLoadExtAction(ISD::ZEXTLOAD, MVT::v4i32, MVT::v4i16, Legal);
986 setLoadExtAction(ISD::ZEXTLOAD, MVT::v2i64, MVT::v2i16, Legal);
987 setLoadExtAction(ISD::ZEXTLOAD, MVT::v2i64, MVT::v2i32, Legal);
989 // i8 and i16 vectors are custom because the source register and source
990 // source memory operand types are not the same width. f32 vectors are
991 // custom since the immediate controlling the insert encodes additional
993 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v16i8, Custom);
994 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i16, Custom);
995 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4i32, Custom);
996 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v4f32, Custom);
998 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v16i8, Custom);
999 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v8i16, Custom);
1000 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4i32, Custom);
1001 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v4f32, Custom);
1003 // FIXME: these should be Legal, but that's only for the case where
1004 // the index is constant. For now custom expand to deal with that.
1005 if (Subtarget->is64Bit()) {
1006 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v2i64, Custom);
1007 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v2i64, Custom);
1011 if (Subtarget->hasSSE2()) {
1012 setOperationAction(ISD::SIGN_EXTEND_VECTOR_INREG, MVT::v2i64, Custom);
1013 setOperationAction(ISD::SIGN_EXTEND_VECTOR_INREG, MVT::v4i32, Custom);
1014 setOperationAction(ISD::SIGN_EXTEND_VECTOR_INREG, MVT::v8i16, Custom);
1016 setOperationAction(ISD::SRL, MVT::v8i16, Custom);
1017 setOperationAction(ISD::SRL, MVT::v16i8, Custom);
1019 setOperationAction(ISD::SHL, MVT::v8i16, Custom);
1020 setOperationAction(ISD::SHL, MVT::v16i8, Custom);
1022 setOperationAction(ISD::SRA, MVT::v8i16, Custom);
1023 setOperationAction(ISD::SRA, MVT::v16i8, Custom);
1025 // In the customized shift lowering, the legal cases in AVX2 will be
1027 setOperationAction(ISD::SRL, MVT::v2i64, Custom);
1028 setOperationAction(ISD::SRL, MVT::v4i32, Custom);
1030 setOperationAction(ISD::SHL, MVT::v2i64, Custom);
1031 setOperationAction(ISD::SHL, MVT::v4i32, Custom);
1033 setOperationAction(ISD::SRA, MVT::v2i64, Custom);
1034 setOperationAction(ISD::SRA, MVT::v4i32, Custom);
1037 if (Subtarget->hasXOP()) {
1038 setOperationAction(ISD::ROTL, MVT::v16i8, Custom);
1039 setOperationAction(ISD::ROTL, MVT::v8i16, Custom);
1040 setOperationAction(ISD::ROTL, MVT::v4i32, Custom);
1041 setOperationAction(ISD::ROTL, MVT::v2i64, Custom);
1042 setOperationAction(ISD::ROTL, MVT::v32i8, Custom);
1043 setOperationAction(ISD::ROTL, MVT::v16i16, Custom);
1044 setOperationAction(ISD::ROTL, MVT::v8i32, Custom);
1045 setOperationAction(ISD::ROTL, MVT::v4i64, Custom);
1048 if (!Subtarget->useSoftFloat() && Subtarget->hasFp256()) {
1049 addRegisterClass(MVT::v32i8, &X86::VR256RegClass);
1050 addRegisterClass(MVT::v16i16, &X86::VR256RegClass);
1051 addRegisterClass(MVT::v8i32, &X86::VR256RegClass);
1052 addRegisterClass(MVT::v8f32, &X86::VR256RegClass);
1053 addRegisterClass(MVT::v4i64, &X86::VR256RegClass);
1054 addRegisterClass(MVT::v4f64, &X86::VR256RegClass);
1056 setOperationAction(ISD::LOAD, MVT::v8f32, Legal);
1057 setOperationAction(ISD::LOAD, MVT::v4f64, Legal);
1058 setOperationAction(ISD::LOAD, MVT::v4i64, Legal);
1060 setOperationAction(ISD::FADD, MVT::v8f32, Legal);
1061 setOperationAction(ISD::FSUB, MVT::v8f32, Legal);
1062 setOperationAction(ISD::FMUL, MVT::v8f32, Legal);
1063 setOperationAction(ISD::FDIV, MVT::v8f32, Legal);
1064 setOperationAction(ISD::FSQRT, MVT::v8f32, Legal);
1065 setOperationAction(ISD::FFLOOR, MVT::v8f32, Legal);
1066 setOperationAction(ISD::FCEIL, MVT::v8f32, Legal);
1067 setOperationAction(ISD::FTRUNC, MVT::v8f32, Legal);
1068 setOperationAction(ISD::FRINT, MVT::v8f32, Legal);
1069 setOperationAction(ISD::FNEARBYINT, MVT::v8f32, Legal);
1070 setOperationAction(ISD::FNEG, MVT::v8f32, Custom);
1071 setOperationAction(ISD::FABS, MVT::v8f32, Custom);
1073 setOperationAction(ISD::FADD, MVT::v4f64, Legal);
1074 setOperationAction(ISD::FSUB, MVT::v4f64, Legal);
1075 setOperationAction(ISD::FMUL, MVT::v4f64, Legal);
1076 setOperationAction(ISD::FDIV, MVT::v4f64, Legal);
1077 setOperationAction(ISD::FSQRT, MVT::v4f64, Legal);
1078 setOperationAction(ISD::FFLOOR, MVT::v4f64, Legal);
1079 setOperationAction(ISD::FCEIL, MVT::v4f64, Legal);
1080 setOperationAction(ISD::FTRUNC, MVT::v4f64, Legal);
1081 setOperationAction(ISD::FRINT, MVT::v4f64, Legal);
1082 setOperationAction(ISD::FNEARBYINT, MVT::v4f64, Legal);
1083 setOperationAction(ISD::FNEG, MVT::v4f64, Custom);
1084 setOperationAction(ISD::FABS, MVT::v4f64, Custom);
1086 // (fp_to_int:v8i16 (v8f32 ..)) requires the result type to be promoted
1087 // even though v8i16 is a legal type.
1088 setOperationAction(ISD::FP_TO_SINT, MVT::v8i16, Promote);
1089 setOperationAction(ISD::FP_TO_UINT, MVT::v8i16, Promote);
1090 setOperationAction(ISD::FP_TO_SINT, MVT::v8i32, Legal);
1092 setOperationAction(ISD::SINT_TO_FP, MVT::v8i16, Promote);
1093 setOperationAction(ISD::SINT_TO_FP, MVT::v8i32, Legal);
1094 setOperationAction(ISD::FP_ROUND, MVT::v4f32, Legal);
1096 setOperationAction(ISD::UINT_TO_FP, MVT::v8i8, Custom);
1097 setOperationAction(ISD::UINT_TO_FP, MVT::v8i16, Custom);
1099 for (MVT VT : MVT::fp_vector_valuetypes())
1100 setLoadExtAction(ISD::EXTLOAD, VT, MVT::v4f32, Legal);
1102 setOperationAction(ISD::SRL, MVT::v16i16, Custom);
1103 setOperationAction(ISD::SRL, MVT::v32i8, Custom);
1105 setOperationAction(ISD::SHL, MVT::v16i16, Custom);
1106 setOperationAction(ISD::SHL, MVT::v32i8, Custom);
1108 setOperationAction(ISD::SRA, MVT::v16i16, Custom);
1109 setOperationAction(ISD::SRA, MVT::v32i8, Custom);
1111 setOperationAction(ISD::SETCC, MVT::v32i8, Custom);
1112 setOperationAction(ISD::SETCC, MVT::v16i16, Custom);
1113 setOperationAction(ISD::SETCC, MVT::v8i32, Custom);
1114 setOperationAction(ISD::SETCC, MVT::v4i64, Custom);
1116 setOperationAction(ISD::SELECT, MVT::v4f64, Custom);
1117 setOperationAction(ISD::SELECT, MVT::v4i64, Custom);
1118 setOperationAction(ISD::SELECT, MVT::v8f32, Custom);
1120 setOperationAction(ISD::SIGN_EXTEND, MVT::v4i64, Custom);
1121 setOperationAction(ISD::SIGN_EXTEND, MVT::v8i32, Custom);
1122 setOperationAction(ISD::SIGN_EXTEND, MVT::v16i16, Custom);
1123 setOperationAction(ISD::ZERO_EXTEND, MVT::v4i64, Custom);
1124 setOperationAction(ISD::ZERO_EXTEND, MVT::v8i32, Custom);
1125 setOperationAction(ISD::ZERO_EXTEND, MVT::v16i16, Custom);
1126 setOperationAction(ISD::ANY_EXTEND, MVT::v4i64, Custom);
1127 setOperationAction(ISD::ANY_EXTEND, MVT::v8i32, Custom);
1128 setOperationAction(ISD::ANY_EXTEND, MVT::v16i16, Custom);
1129 setOperationAction(ISD::TRUNCATE, MVT::v16i8, Custom);
1130 setOperationAction(ISD::TRUNCATE, MVT::v8i16, Custom);
1131 setOperationAction(ISD::TRUNCATE, MVT::v4i32, Custom);
1133 setOperationAction(ISD::CTPOP, MVT::v32i8, Custom);
1134 setOperationAction(ISD::CTPOP, MVT::v16i16, Custom);
1135 setOperationAction(ISD::CTPOP, MVT::v8i32, Custom);
1136 setOperationAction(ISD::CTPOP, MVT::v4i64, Custom);
1138 setOperationAction(ISD::CTTZ, MVT::v32i8, Custom);
1139 setOperationAction(ISD::CTTZ, MVT::v16i16, Custom);
1140 setOperationAction(ISD::CTTZ, MVT::v8i32, Custom);
1141 setOperationAction(ISD::CTTZ, MVT::v4i64, Custom);
1142 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::v32i8, Custom);
1143 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::v16i16, Custom);
1144 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::v8i32, Custom);
1145 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::v4i64, Custom);
1147 if (Subtarget->hasFMA() || Subtarget->hasFMA4() || Subtarget->hasAVX512()) {
1148 setOperationAction(ISD::FMA, MVT::v8f32, Legal);
1149 setOperationAction(ISD::FMA, MVT::v4f64, Legal);
1150 setOperationAction(ISD::FMA, MVT::v4f32, Legal);
1151 setOperationAction(ISD::FMA, MVT::v2f64, Legal);
1152 setOperationAction(ISD::FMA, MVT::f32, Legal);
1153 setOperationAction(ISD::FMA, MVT::f64, Legal);
1156 if (Subtarget->hasInt256()) {
1157 setOperationAction(ISD::ADD, MVT::v4i64, Legal);
1158 setOperationAction(ISD::ADD, MVT::v8i32, Legal);
1159 setOperationAction(ISD::ADD, MVT::v16i16, Legal);
1160 setOperationAction(ISD::ADD, MVT::v32i8, Legal);
1162 setOperationAction(ISD::SUB, MVT::v4i64, Legal);
1163 setOperationAction(ISD::SUB, MVT::v8i32, Legal);
1164 setOperationAction(ISD::SUB, MVT::v16i16, Legal);
1165 setOperationAction(ISD::SUB, MVT::v32i8, Legal);
1167 setOperationAction(ISD::MUL, MVT::v4i64, Custom);
1168 setOperationAction(ISD::MUL, MVT::v8i32, Legal);
1169 setOperationAction(ISD::MUL, MVT::v16i16, Legal);
1170 setOperationAction(ISD::MUL, MVT::v32i8, Custom);
1172 setOperationAction(ISD::UMUL_LOHI, MVT::v8i32, Custom);
1173 setOperationAction(ISD::SMUL_LOHI, MVT::v8i32, Custom);
1174 setOperationAction(ISD::MULHU, MVT::v16i16, Legal);
1175 setOperationAction(ISD::MULHS, MVT::v16i16, Legal);
1177 setOperationAction(ISD::SMAX, MVT::v32i8, Legal);
1178 setOperationAction(ISD::SMAX, MVT::v16i16, Legal);
1179 setOperationAction(ISD::SMAX, MVT::v8i32, Legal);
1180 setOperationAction(ISD::UMAX, MVT::v32i8, Legal);
1181 setOperationAction(ISD::UMAX, MVT::v16i16, Legal);
1182 setOperationAction(ISD::UMAX, MVT::v8i32, Legal);
1183 setOperationAction(ISD::SMIN, MVT::v32i8, Legal);
1184 setOperationAction(ISD::SMIN, MVT::v16i16, Legal);
1185 setOperationAction(ISD::SMIN, MVT::v8i32, Legal);
1186 setOperationAction(ISD::UMIN, MVT::v32i8, Legal);
1187 setOperationAction(ISD::UMIN, MVT::v16i16, Legal);
1188 setOperationAction(ISD::UMIN, MVT::v8i32, Legal);
1190 // The custom lowering for UINT_TO_FP for v8i32 becomes interesting
1191 // when we have a 256bit-wide blend with immediate.
1192 setOperationAction(ISD::UINT_TO_FP, MVT::v8i32, Custom);
1194 // AVX2 also has wider vector sign/zero extending loads, VPMOV[SZ]X
1195 setLoadExtAction(ISD::SEXTLOAD, MVT::v16i16, MVT::v16i8, Legal);
1196 setLoadExtAction(ISD::SEXTLOAD, MVT::v8i32, MVT::v8i8, Legal);
1197 setLoadExtAction(ISD::SEXTLOAD, MVT::v4i64, MVT::v4i8, Legal);
1198 setLoadExtAction(ISD::SEXTLOAD, MVT::v8i32, MVT::v8i16, Legal);
1199 setLoadExtAction(ISD::SEXTLOAD, MVT::v4i64, MVT::v4i16, Legal);
1200 setLoadExtAction(ISD::SEXTLOAD, MVT::v4i64, MVT::v4i32, Legal);
1202 setLoadExtAction(ISD::ZEXTLOAD, MVT::v16i16, MVT::v16i8, Legal);
1203 setLoadExtAction(ISD::ZEXTLOAD, MVT::v8i32, MVT::v8i8, Legal);
1204 setLoadExtAction(ISD::ZEXTLOAD, MVT::v4i64, MVT::v4i8, Legal);
1205 setLoadExtAction(ISD::ZEXTLOAD, MVT::v8i32, MVT::v8i16, Legal);
1206 setLoadExtAction(ISD::ZEXTLOAD, MVT::v4i64, MVT::v4i16, Legal);
1207 setLoadExtAction(ISD::ZEXTLOAD, MVT::v4i64, MVT::v4i32, Legal);
1209 setOperationAction(ISD::ADD, MVT::v4i64, Custom);
1210 setOperationAction(ISD::ADD, MVT::v8i32, Custom);
1211 setOperationAction(ISD::ADD, MVT::v16i16, Custom);
1212 setOperationAction(ISD::ADD, MVT::v32i8, Custom);
1214 setOperationAction(ISD::SUB, MVT::v4i64, Custom);
1215 setOperationAction(ISD::SUB, MVT::v8i32, Custom);
1216 setOperationAction(ISD::SUB, MVT::v16i16, Custom);
1217 setOperationAction(ISD::SUB, MVT::v32i8, Custom);
1219 setOperationAction(ISD::MUL, MVT::v4i64, Custom);
1220 setOperationAction(ISD::MUL, MVT::v8i32, Custom);
1221 setOperationAction(ISD::MUL, MVT::v16i16, Custom);
1222 setOperationAction(ISD::MUL, MVT::v32i8, Custom);
1224 setOperationAction(ISD::SMAX, MVT::v32i8, Custom);
1225 setOperationAction(ISD::SMAX, MVT::v16i16, Custom);
1226 setOperationAction(ISD::SMAX, MVT::v8i32, Custom);
1227 setOperationAction(ISD::UMAX, MVT::v32i8, Custom);
1228 setOperationAction(ISD::UMAX, MVT::v16i16, Custom);
1229 setOperationAction(ISD::UMAX, MVT::v8i32, Custom);
1230 setOperationAction(ISD::SMIN, MVT::v32i8, Custom);
1231 setOperationAction(ISD::SMIN, MVT::v16i16, Custom);
1232 setOperationAction(ISD::SMIN, MVT::v8i32, Custom);
1233 setOperationAction(ISD::UMIN, MVT::v32i8, Custom);
1234 setOperationAction(ISD::UMIN, MVT::v16i16, Custom);
1235 setOperationAction(ISD::UMIN, MVT::v8i32, Custom);
1238 // In the customized shift lowering, the legal cases in AVX2 will be
1240 setOperationAction(ISD::SRL, MVT::v4i64, Custom);
1241 setOperationAction(ISD::SRL, MVT::v8i32, Custom);
1243 setOperationAction(ISD::SHL, MVT::v4i64, Custom);
1244 setOperationAction(ISD::SHL, MVT::v8i32, Custom);
1246 setOperationAction(ISD::SRA, MVT::v4i64, Custom);
1247 setOperationAction(ISD::SRA, MVT::v8i32, Custom);
1249 // Custom lower several nodes for 256-bit types.
1250 for (MVT VT : MVT::vector_valuetypes()) {
1251 if (VT.getScalarSizeInBits() >= 32) {
1252 setOperationAction(ISD::MLOAD, VT, Legal);
1253 setOperationAction(ISD::MSTORE, VT, Legal);
1255 // Extract subvector is special because the value type
1256 // (result) is 128-bit but the source is 256-bit wide.
1257 if (VT.is128BitVector()) {
1258 setOperationAction(ISD::EXTRACT_SUBVECTOR, VT, Custom);
1260 // Do not attempt to custom lower other non-256-bit vectors
1261 if (!VT.is256BitVector())
1264 setOperationAction(ISD::BUILD_VECTOR, VT, Custom);
1265 setOperationAction(ISD::VECTOR_SHUFFLE, VT, Custom);
1266 setOperationAction(ISD::VSELECT, VT, Custom);
1267 setOperationAction(ISD::INSERT_VECTOR_ELT, VT, Custom);
1268 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Custom);
1269 setOperationAction(ISD::SCALAR_TO_VECTOR, VT, Custom);
1270 setOperationAction(ISD::INSERT_SUBVECTOR, VT, Custom);
1271 setOperationAction(ISD::CONCAT_VECTORS, VT, Custom);
1274 if (Subtarget->hasInt256())
1275 setOperationAction(ISD::VSELECT, MVT::v32i8, Legal);
1277 // Promote v32i8, v16i16, v8i32 select, and, or, xor to v4i64.
1278 for (auto VT : { MVT::v32i8, MVT::v16i16, MVT::v8i32 }) {
1279 setOperationAction(ISD::AND, VT, Promote);
1280 AddPromotedToType (ISD::AND, VT, MVT::v4i64);
1281 setOperationAction(ISD::OR, VT, Promote);
1282 AddPromotedToType (ISD::OR, VT, MVT::v4i64);
1283 setOperationAction(ISD::XOR, VT, Promote);
1284 AddPromotedToType (ISD::XOR, VT, MVT::v4i64);
1285 setOperationAction(ISD::LOAD, VT, Promote);
1286 AddPromotedToType (ISD::LOAD, VT, MVT::v4i64);
1287 setOperationAction(ISD::SELECT, VT, Promote);
1288 AddPromotedToType (ISD::SELECT, VT, MVT::v4i64);
1292 if (!Subtarget->useSoftFloat() && Subtarget->hasAVX512()) {
1293 addRegisterClass(MVT::v16i32, &X86::VR512RegClass);
1294 addRegisterClass(MVT::v16f32, &X86::VR512RegClass);
1295 addRegisterClass(MVT::v8i64, &X86::VR512RegClass);
1296 addRegisterClass(MVT::v8f64, &X86::VR512RegClass);
1298 addRegisterClass(MVT::i1, &X86::VK1RegClass);
1299 addRegisterClass(MVT::v8i1, &X86::VK8RegClass);
1300 addRegisterClass(MVT::v16i1, &X86::VK16RegClass);
1302 for (MVT VT : MVT::fp_vector_valuetypes())
1303 setLoadExtAction(ISD::EXTLOAD, VT, MVT::v8f32, Legal);
1305 setLoadExtAction(ISD::ZEXTLOAD, MVT::v16i32, MVT::v16i8, Legal);
1306 setLoadExtAction(ISD::SEXTLOAD, MVT::v16i32, MVT::v16i8, Legal);
1307 setLoadExtAction(ISD::ZEXTLOAD, MVT::v16i32, MVT::v16i16, Legal);
1308 setLoadExtAction(ISD::SEXTLOAD, MVT::v16i32, MVT::v16i16, Legal);
1309 setLoadExtAction(ISD::ZEXTLOAD, MVT::v32i16, MVT::v32i8, Legal);
1310 setLoadExtAction(ISD::SEXTLOAD, MVT::v32i16, MVT::v32i8, Legal);
1311 setLoadExtAction(ISD::ZEXTLOAD, MVT::v8i64, MVT::v8i8, Legal);
1312 setLoadExtAction(ISD::SEXTLOAD, MVT::v8i64, MVT::v8i8, Legal);
1313 setLoadExtAction(ISD::ZEXTLOAD, MVT::v8i64, MVT::v8i16, Legal);
1314 setLoadExtAction(ISD::SEXTLOAD, MVT::v8i64, MVT::v8i16, Legal);
1315 setLoadExtAction(ISD::ZEXTLOAD, MVT::v8i64, MVT::v8i32, Legal);
1316 setLoadExtAction(ISD::SEXTLOAD, MVT::v8i64, MVT::v8i32, Legal);
1318 setOperationAction(ISD::BR_CC, MVT::i1, Expand);
1319 setOperationAction(ISD::SETCC, MVT::i1, Custom);
1320 setOperationAction(ISD::SELECT_CC, MVT::i1, Expand);
1321 setOperationAction(ISD::XOR, MVT::i1, Legal);
1322 setOperationAction(ISD::OR, MVT::i1, Legal);
1323 setOperationAction(ISD::AND, MVT::i1, Legal);
1324 setOperationAction(ISD::SUB, MVT::i1, Custom);
1325 setOperationAction(ISD::ADD, MVT::i1, Custom);
1326 setOperationAction(ISD::MUL, MVT::i1, Custom);
1327 setOperationAction(ISD::LOAD, MVT::v16f32, Legal);
1328 setOperationAction(ISD::LOAD, MVT::v8f64, Legal);
1329 setOperationAction(ISD::LOAD, MVT::v8i64, Legal);
1330 setOperationAction(ISD::LOAD, MVT::v16i32, Legal);
1331 setOperationAction(ISD::LOAD, MVT::v16i1, Legal);
1333 setOperationAction(ISD::FADD, MVT::v16f32, Legal);
1334 setOperationAction(ISD::FSUB, MVT::v16f32, Legal);
1335 setOperationAction(ISD::FMUL, MVT::v16f32, Legal);
1336 setOperationAction(ISD::FDIV, MVT::v16f32, Legal);
1337 setOperationAction(ISD::FSQRT, MVT::v16f32, Legal);
1338 setOperationAction(ISD::FNEG, MVT::v16f32, Custom);
1340 setOperationAction(ISD::FADD, MVT::v8f64, Legal);
1341 setOperationAction(ISD::FSUB, MVT::v8f64, Legal);
1342 setOperationAction(ISD::FMUL, MVT::v8f64, Legal);
1343 setOperationAction(ISD::FDIV, MVT::v8f64, Legal);
1344 setOperationAction(ISD::FSQRT, MVT::v8f64, Legal);
1345 setOperationAction(ISD::FNEG, MVT::v8f64, Custom);
1346 setOperationAction(ISD::FMA, MVT::v8f64, Legal);
1347 setOperationAction(ISD::FMA, MVT::v16f32, Legal);
1349 setOperationAction(ISD::FP_TO_SINT, MVT::v16i32, Legal);
1350 setOperationAction(ISD::FP_TO_UINT, MVT::v16i32, Legal);
1351 setOperationAction(ISD::FP_TO_UINT, MVT::v8i32, Legal);
1352 setOperationAction(ISD::FP_TO_UINT, MVT::v4i32, Legal);
1353 setOperationAction(ISD::SINT_TO_FP, MVT::v16i32, Legal);
1354 setOperationAction(ISD::SINT_TO_FP, MVT::v8i1, Custom);
1355 setOperationAction(ISD::SINT_TO_FP, MVT::v16i1, Custom);
1356 setOperationAction(ISD::SINT_TO_FP, MVT::v16i8, Promote);
1357 setOperationAction(ISD::SINT_TO_FP, MVT::v16i16, Promote);
1358 setOperationAction(ISD::UINT_TO_FP, MVT::v16i32, Legal);
1359 setOperationAction(ISD::UINT_TO_FP, MVT::v8i32, Legal);
1360 setOperationAction(ISD::UINT_TO_FP, MVT::v4i32, Legal);
1361 setOperationAction(ISD::UINT_TO_FP, MVT::v16i8, Custom);
1362 setOperationAction(ISD::UINT_TO_FP, MVT::v16i16, Custom);
1363 setOperationAction(ISD::FP_ROUND, MVT::v8f32, Legal);
1364 setOperationAction(ISD::FP_EXTEND, MVT::v8f32, Legal);
1366 setTruncStoreAction(MVT::v8i64, MVT::v8i8, Legal);
1367 setTruncStoreAction(MVT::v8i64, MVT::v8i16, Legal);
1368 setTruncStoreAction(MVT::v8i64, MVT::v8i32, Legal);
1369 setTruncStoreAction(MVT::v16i32, MVT::v16i8, Legal);
1370 setTruncStoreAction(MVT::v16i32, MVT::v16i16, Legal);
1371 if (Subtarget->hasVLX()){
1372 setTruncStoreAction(MVT::v4i64, MVT::v4i8, Legal);
1373 setTruncStoreAction(MVT::v4i64, MVT::v4i16, Legal);
1374 setTruncStoreAction(MVT::v4i64, MVT::v4i32, Legal);
1375 setTruncStoreAction(MVT::v8i32, MVT::v8i8, Legal);
1376 setTruncStoreAction(MVT::v8i32, MVT::v8i16, Legal);
1378 setTruncStoreAction(MVT::v2i64, MVT::v2i8, Legal);
1379 setTruncStoreAction(MVT::v2i64, MVT::v2i16, Legal);
1380 setTruncStoreAction(MVT::v2i64, MVT::v2i32, Legal);
1381 setTruncStoreAction(MVT::v4i32, MVT::v4i8, Legal);
1382 setTruncStoreAction(MVT::v4i32, MVT::v4i16, Legal);
1384 setOperationAction(ISD::TRUNCATE, MVT::i1, Custom);
1385 setOperationAction(ISD::TRUNCATE, MVT::v16i8, Custom);
1386 setOperationAction(ISD::TRUNCATE, MVT::v8i32, Custom);
1387 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v8i1, Custom);
1388 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v16i1, Custom);
1389 if (Subtarget->hasDQI()) {
1390 setOperationAction(ISD::TRUNCATE, MVT::v2i1, Custom);
1391 setOperationAction(ISD::TRUNCATE, MVT::v4i1, Custom);
1393 setOperationAction(ISD::SINT_TO_FP, MVT::v8i64, Legal);
1394 setOperationAction(ISD::UINT_TO_FP, MVT::v8i64, Legal);
1395 setOperationAction(ISD::FP_TO_SINT, MVT::v8i64, Legal);
1396 setOperationAction(ISD::FP_TO_UINT, MVT::v8i64, Legal);
1397 if (Subtarget->hasVLX()) {
1398 setOperationAction(ISD::SINT_TO_FP, MVT::v4i64, Legal);
1399 setOperationAction(ISD::SINT_TO_FP, MVT::v2i64, Legal);
1400 setOperationAction(ISD::UINT_TO_FP, MVT::v4i64, Legal);
1401 setOperationAction(ISD::UINT_TO_FP, MVT::v2i64, Legal);
1402 setOperationAction(ISD::FP_TO_SINT, MVT::v4i64, Legal);
1403 setOperationAction(ISD::FP_TO_SINT, MVT::v2i64, Legal);
1404 setOperationAction(ISD::FP_TO_UINT, MVT::v4i64, Legal);
1405 setOperationAction(ISD::FP_TO_UINT, MVT::v2i64, Legal);
1408 if (Subtarget->hasVLX()) {
1409 setOperationAction(ISD::SINT_TO_FP, MVT::v8i32, Legal);
1410 setOperationAction(ISD::UINT_TO_FP, MVT::v8i32, Legal);
1411 setOperationAction(ISD::FP_TO_SINT, MVT::v8i32, Legal);
1412 setOperationAction(ISD::FP_TO_UINT, MVT::v8i32, Legal);
1413 setOperationAction(ISD::SINT_TO_FP, MVT::v4i32, Legal);
1414 setOperationAction(ISD::UINT_TO_FP, MVT::v4i32, Legal);
1415 setOperationAction(ISD::FP_TO_SINT, MVT::v4i32, Legal);
1416 setOperationAction(ISD::FP_TO_UINT, MVT::v4i32, Legal);
1418 setOperationAction(ISD::TRUNCATE, MVT::v8i1, Custom);
1419 setOperationAction(ISD::TRUNCATE, MVT::v16i1, Custom);
1420 setOperationAction(ISD::TRUNCATE, MVT::v16i16, Custom);
1421 setOperationAction(ISD::ZERO_EXTEND, MVT::v16i32, Custom);
1422 setOperationAction(ISD::ZERO_EXTEND, MVT::v8i64, Custom);
1423 setOperationAction(ISD::ANY_EXTEND, MVT::v16i32, Custom);
1424 setOperationAction(ISD::ANY_EXTEND, MVT::v8i64, Custom);
1425 setOperationAction(ISD::SIGN_EXTEND, MVT::v16i32, Custom);
1426 setOperationAction(ISD::SIGN_EXTEND, MVT::v8i64, Custom);
1427 setOperationAction(ISD::SIGN_EXTEND, MVT::v16i8, Custom);
1428 setOperationAction(ISD::SIGN_EXTEND, MVT::v8i16, Custom);
1429 setOperationAction(ISD::SIGN_EXTEND, MVT::v16i16, Custom);
1430 if (Subtarget->hasDQI()) {
1431 setOperationAction(ISD::SIGN_EXTEND, MVT::v4i32, Custom);
1432 setOperationAction(ISD::SIGN_EXTEND, MVT::v2i64, Custom);
1434 setOperationAction(ISD::FFLOOR, MVT::v16f32, Legal);
1435 setOperationAction(ISD::FFLOOR, MVT::v8f64, Legal);
1436 setOperationAction(ISD::FCEIL, MVT::v16f32, Legal);
1437 setOperationAction(ISD::FCEIL, MVT::v8f64, Legal);
1438 setOperationAction(ISD::FTRUNC, MVT::v16f32, Legal);
1439 setOperationAction(ISD::FTRUNC, MVT::v8f64, Legal);
1440 setOperationAction(ISD::FRINT, MVT::v16f32, Legal);
1441 setOperationAction(ISD::FRINT, MVT::v8f64, Legal);
1442 setOperationAction(ISD::FNEARBYINT, MVT::v16f32, Legal);
1443 setOperationAction(ISD::FNEARBYINT, MVT::v8f64, Legal);
1445 setOperationAction(ISD::CONCAT_VECTORS, MVT::v8f64, Custom);
1446 setOperationAction(ISD::CONCAT_VECTORS, MVT::v8i64, Custom);
1447 setOperationAction(ISD::CONCAT_VECTORS, MVT::v16f32, Custom);
1448 setOperationAction(ISD::CONCAT_VECTORS, MVT::v16i32, Custom);
1449 setOperationAction(ISD::CONCAT_VECTORS, MVT::v16i1, Legal);
1451 setOperationAction(ISD::SETCC, MVT::v16i1, Custom);
1452 setOperationAction(ISD::SETCC, MVT::v8i1, Custom);
1454 setOperationAction(ISD::MUL, MVT::v8i64, Custom);
1456 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v8i1, Custom);
1457 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v16i1, Custom);
1458 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v16i1, Custom);
1459 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i1, Custom);
1460 setOperationAction(ISD::BUILD_VECTOR, MVT::v8i1, Custom);
1461 setOperationAction(ISD::BUILD_VECTOR, MVT::v16i1, Custom);
1462 setOperationAction(ISD::SELECT, MVT::v8f64, Custom);
1463 setOperationAction(ISD::SELECT, MVT::v8i64, Custom);
1464 setOperationAction(ISD::SELECT, MVT::v16f32, Custom);
1465 setOperationAction(ISD::SELECT, MVT::v16i1, Custom);
1466 setOperationAction(ISD::SELECT, MVT::v8i1, Custom);
1468 setOperationAction(ISD::SMAX, MVT::v16i32, Legal);
1469 setOperationAction(ISD::SMAX, MVT::v8i64, Legal);
1470 setOperationAction(ISD::UMAX, MVT::v16i32, Legal);
1471 setOperationAction(ISD::UMAX, MVT::v8i64, Legal);
1472 setOperationAction(ISD::SMIN, MVT::v16i32, Legal);
1473 setOperationAction(ISD::SMIN, MVT::v8i64, Legal);
1474 setOperationAction(ISD::UMIN, MVT::v16i32, Legal);
1475 setOperationAction(ISD::UMIN, MVT::v8i64, Legal);
1477 setOperationAction(ISD::ADD, MVT::v8i64, Legal);
1478 setOperationAction(ISD::ADD, MVT::v16i32, Legal);
1480 setOperationAction(ISD::SUB, MVT::v8i64, Legal);
1481 setOperationAction(ISD::SUB, MVT::v16i32, Legal);
1483 setOperationAction(ISD::MUL, MVT::v16i32, Legal);
1485 setOperationAction(ISD::SRL, MVT::v8i64, Custom);
1486 setOperationAction(ISD::SRL, MVT::v16i32, Custom);
1488 setOperationAction(ISD::SHL, MVT::v8i64, Custom);
1489 setOperationAction(ISD::SHL, MVT::v16i32, Custom);
1491 setOperationAction(ISD::SRA, MVT::v8i64, Custom);
1492 setOperationAction(ISD::SRA, MVT::v16i32, Custom);
1494 setOperationAction(ISD::AND, MVT::v8i64, Legal);
1495 setOperationAction(ISD::OR, MVT::v8i64, Legal);
1496 setOperationAction(ISD::XOR, MVT::v8i64, Legal);
1497 setOperationAction(ISD::AND, MVT::v16i32, Legal);
1498 setOperationAction(ISD::OR, MVT::v16i32, Legal);
1499 setOperationAction(ISD::XOR, MVT::v16i32, Legal);
1501 if (Subtarget->hasCDI()) {
1502 setOperationAction(ISD::CTLZ, MVT::v8i64, Legal);
1503 setOperationAction(ISD::CTLZ, MVT::v16i32, Legal);
1504 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::v8i64, Legal);
1505 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::v16i32, Legal);
1507 setOperationAction(ISD::CTLZ, MVT::v8i16, Custom);
1508 setOperationAction(ISD::CTLZ, MVT::v16i8, Custom);
1509 setOperationAction(ISD::CTLZ, MVT::v16i16, Custom);
1510 setOperationAction(ISD::CTLZ, MVT::v32i8, Custom);
1511 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::v8i16, Custom);
1512 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::v16i8, Custom);
1513 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::v16i16, Custom);
1514 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::v32i8, Custom);
1516 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::v8i64, Custom);
1517 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::v16i32, Custom);
1519 if (Subtarget->hasVLX()) {
1520 setOperationAction(ISD::CTLZ, MVT::v4i64, Legal);
1521 setOperationAction(ISD::CTLZ, MVT::v8i32, Legal);
1522 setOperationAction(ISD::CTLZ, MVT::v2i64, Legal);
1523 setOperationAction(ISD::CTLZ, MVT::v4i32, Legal);
1524 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::v4i64, Legal);
1525 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::v8i32, Legal);
1526 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::v2i64, Legal);
1527 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::v4i32, Legal);
1529 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::v4i64, Custom);
1530 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::v8i32, Custom);
1531 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::v2i64, Custom);
1532 setOperationAction(ISD::CTTZ_ZERO_UNDEF, MVT::v4i32, Custom);
1534 setOperationAction(ISD::CTLZ, MVT::v4i64, Custom);
1535 setOperationAction(ISD::CTLZ, MVT::v8i32, Custom);
1536 setOperationAction(ISD::CTLZ, MVT::v2i64, Custom);
1537 setOperationAction(ISD::CTLZ, MVT::v4i32, Custom);
1538 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::v4i64, Custom);
1539 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::v8i32, Custom);
1540 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::v2i64, Custom);
1541 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::v4i32, Custom);
1543 } // Subtarget->hasCDI()
1545 if (Subtarget->hasDQI()) {
1546 setOperationAction(ISD::MUL, MVT::v2i64, Legal);
1547 setOperationAction(ISD::MUL, MVT::v4i64, Legal);
1548 setOperationAction(ISD::MUL, MVT::v8i64, Legal);
1550 // Custom lower several nodes.
1551 for (MVT VT : MVT::vector_valuetypes()) {
1552 unsigned EltSize = VT.getVectorElementType().getSizeInBits();
1554 setOperationAction(ISD::AND, VT, Legal);
1555 setOperationAction(ISD::OR, VT, Legal);
1556 setOperationAction(ISD::XOR, VT, Legal);
1558 if (EltSize >= 32 && VT.getSizeInBits() <= 512) {
1559 setOperationAction(ISD::MGATHER, VT, Custom);
1560 setOperationAction(ISD::MSCATTER, VT, Custom);
1562 // Extract subvector is special because the value type
1563 // (result) is 256/128-bit but the source is 512-bit wide.
1564 if (VT.is128BitVector() || VT.is256BitVector()) {
1565 setOperationAction(ISD::EXTRACT_SUBVECTOR, VT, Custom);
1567 if (VT.getVectorElementType() == MVT::i1)
1568 setOperationAction(ISD::EXTRACT_SUBVECTOR, VT, Legal);
1570 // Do not attempt to custom lower other non-512-bit vectors
1571 if (!VT.is512BitVector())
1574 if (EltSize >= 32) {
1575 setOperationAction(ISD::VECTOR_SHUFFLE, VT, Custom);
1576 setOperationAction(ISD::INSERT_VECTOR_ELT, VT, Custom);
1577 setOperationAction(ISD::BUILD_VECTOR, VT, Custom);
1578 setOperationAction(ISD::VSELECT, VT, Legal);
1579 setOperationAction(ISD::EXTRACT_VECTOR_ELT, VT, Custom);
1580 setOperationAction(ISD::SCALAR_TO_VECTOR, VT, Custom);
1581 setOperationAction(ISD::INSERT_SUBVECTOR, VT, Custom);
1582 setOperationAction(ISD::MLOAD, VT, Legal);
1583 setOperationAction(ISD::MSTORE, VT, Legal);
1586 for (auto VT : { MVT::v64i8, MVT::v32i16, MVT::v16i32 }) {
1587 setOperationAction(ISD::SELECT, VT, Promote);
1588 AddPromotedToType (ISD::SELECT, VT, MVT::v8i64);
1592 if (!Subtarget->useSoftFloat() && Subtarget->hasBWI()) {
1593 addRegisterClass(MVT::v32i16, &X86::VR512RegClass);
1594 addRegisterClass(MVT::v64i8, &X86::VR512RegClass);
1596 addRegisterClass(MVT::v32i1, &X86::VK32RegClass);
1597 addRegisterClass(MVT::v64i1, &X86::VK64RegClass);
1599 setOperationAction(ISD::LOAD, MVT::v32i16, Legal);
1600 setOperationAction(ISD::LOAD, MVT::v64i8, Legal);
1601 setOperationAction(ISD::SETCC, MVT::v32i1, Custom);
1602 setOperationAction(ISD::SETCC, MVT::v64i1, Custom);
1603 setOperationAction(ISD::ADD, MVT::v32i16, Legal);
1604 setOperationAction(ISD::ADD, MVT::v64i8, Legal);
1605 setOperationAction(ISD::SUB, MVT::v32i16, Legal);
1606 setOperationAction(ISD::SUB, MVT::v64i8, Legal);
1607 setOperationAction(ISD::MUL, MVT::v32i16, Legal);
1608 setOperationAction(ISD::MULHS, MVT::v32i16, Legal);
1609 setOperationAction(ISD::MULHU, MVT::v32i16, Legal);
1610 setOperationAction(ISD::CONCAT_VECTORS, MVT::v32i1, Legal);
1611 setOperationAction(ISD::CONCAT_VECTORS, MVT::v64i1, Legal);
1612 setOperationAction(ISD::CONCAT_VECTORS, MVT::v32i16, Custom);
1613 setOperationAction(ISD::CONCAT_VECTORS, MVT::v64i8, Custom);
1614 setOperationAction(ISD::INSERT_SUBVECTOR, MVT::v32i1, Custom);
1615 setOperationAction(ISD::INSERT_SUBVECTOR, MVT::v64i1, Custom);
1616 setOperationAction(ISD::INSERT_SUBVECTOR, MVT::v32i16, Custom);
1617 setOperationAction(ISD::INSERT_SUBVECTOR, MVT::v64i8, Custom);
1618 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v32i16, Custom);
1619 setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v64i8, Custom);
1620 setOperationAction(ISD::SELECT, MVT::v32i1, Custom);
1621 setOperationAction(ISD::SELECT, MVT::v64i1, Custom);
1622 setOperationAction(ISD::SIGN_EXTEND, MVT::v32i8, Custom);
1623 setOperationAction(ISD::ZERO_EXTEND, MVT::v32i8, Custom);
1624 setOperationAction(ISD::SIGN_EXTEND, MVT::v32i16, Custom);
1625 setOperationAction(ISD::ZERO_EXTEND, MVT::v32i16, Custom);
1626 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v32i16, Custom);
1627 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v64i8, Custom);
1628 setOperationAction(ISD::SIGN_EXTEND, MVT::v64i8, Custom);
1629 setOperationAction(ISD::ZERO_EXTEND, MVT::v64i8, Custom);
1630 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v32i1, Custom);
1631 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v64i1, Custom);
1632 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v32i16, Custom);
1633 setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v64i8, Custom);
1634 setOperationAction(ISD::VSELECT, MVT::v32i16, Legal);
1635 setOperationAction(ISD::VSELECT, MVT::v64i8, Legal);
1636 setOperationAction(ISD::TRUNCATE, MVT::v32i1, Custom);
1637 setOperationAction(ISD::TRUNCATE, MVT::v64i1, Custom);
1638 setOperationAction(ISD::TRUNCATE, MVT::v32i8, Custom);
1639 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v32i1, Custom);
1640 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v64i1, Custom);
1642 setOperationAction(ISD::SMAX, MVT::v64i8, Legal);
1643 setOperationAction(ISD::SMAX, MVT::v32i16, Legal);
1644 setOperationAction(ISD::UMAX, MVT::v64i8, Legal);
1645 setOperationAction(ISD::UMAX, MVT::v32i16, Legal);
1646 setOperationAction(ISD::SMIN, MVT::v64i8, Legal);
1647 setOperationAction(ISD::SMIN, MVT::v32i16, Legal);
1648 setOperationAction(ISD::UMIN, MVT::v64i8, Legal);
1649 setOperationAction(ISD::UMIN, MVT::v32i16, Legal);
1651 setTruncStoreAction(MVT::v32i16, MVT::v32i8, Legal);
1652 setTruncStoreAction(MVT::v16i16, MVT::v16i8, Legal);
1653 if (Subtarget->hasVLX())
1654 setTruncStoreAction(MVT::v8i16, MVT::v8i8, Legal);
1656 if (Subtarget->hasCDI()) {
1657 setOperationAction(ISD::CTLZ, MVT::v32i16, Custom);
1658 setOperationAction(ISD::CTLZ, MVT::v64i8, Custom);
1659 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::v32i16, Custom);
1660 setOperationAction(ISD::CTLZ_ZERO_UNDEF, MVT::v64i8, Custom);
1663 for (auto VT : { MVT::v64i8, MVT::v32i16 }) {
1664 setOperationAction(ISD::BUILD_VECTOR, VT, Custom);
1665 setOperationAction(ISD::VSELECT, VT, Legal);
1669 if (!Subtarget->useSoftFloat() && Subtarget->hasVLX()) {
1670 addRegisterClass(MVT::v4i1, &X86::VK4RegClass);
1671 addRegisterClass(MVT::v2i1, &X86::VK2RegClass);
1673 setOperationAction(ISD::SETCC, MVT::v4i1, Custom);
1674 setOperationAction(ISD::SETCC, MVT::v2i1, Custom);
1675 setOperationAction(ISD::CONCAT_VECTORS, MVT::v4i1, Custom);
1676 setOperationAction(ISD::CONCAT_VECTORS, MVT::v8i1, Custom);
1677 setOperationAction(ISD::INSERT_SUBVECTOR, MVT::v8i1, Custom);
1678 setOperationAction(ISD::INSERT_SUBVECTOR, MVT::v4i1, Custom);
1679 setOperationAction(ISD::SELECT, MVT::v4i1, Custom);
1680 setOperationAction(ISD::SELECT, MVT::v2i1, Custom);
1681 setOperationAction(ISD::BUILD_VECTOR, MVT::v4i1, Custom);
1682 setOperationAction(ISD::BUILD_VECTOR, MVT::v2i1, Custom);
1683 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v2i1, Custom);
1684 setOperationAction(ISD::VECTOR_SHUFFLE, MVT::v4i1, Custom);
1686 setOperationAction(ISD::AND, MVT::v8i32, Legal);
1687 setOperationAction(ISD::OR, MVT::v8i32, Legal);
1688 setOperationAction(ISD::XOR, MVT::v8i32, Legal);
1689 setOperationAction(ISD::AND, MVT::v4i32, Legal);
1690 setOperationAction(ISD::OR, MVT::v4i32, Legal);
1691 setOperationAction(ISD::XOR, MVT::v4i32, Legal);
1692 setOperationAction(ISD::SRA, MVT::v2i64, Custom);
1693 setOperationAction(ISD::SRA, MVT::v4i64, Custom);
1695 setOperationAction(ISD::SMAX, MVT::v2i64, Legal);
1696 setOperationAction(ISD::SMAX, MVT::v4i64, Legal);
1697 setOperationAction(ISD::UMAX, MVT::v2i64, Legal);
1698 setOperationAction(ISD::UMAX, MVT::v4i64, Legal);
1699 setOperationAction(ISD::SMIN, MVT::v2i64, Legal);
1700 setOperationAction(ISD::SMIN, MVT::v4i64, Legal);
1701 setOperationAction(ISD::UMIN, MVT::v2i64, Legal);
1702 setOperationAction(ISD::UMIN, MVT::v4i64, Legal);
1705 // We want to custom lower some of our intrinsics.
1706 setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
1707 setOperationAction(ISD::INTRINSIC_W_CHAIN, MVT::Other, Custom);
1708 setOperationAction(ISD::INTRINSIC_VOID, MVT::Other, Custom);
1709 if (!Subtarget->is64Bit())
1710 setOperationAction(ISD::INTRINSIC_W_CHAIN, MVT::i64, Custom);
1712 // Only custom-lower 64-bit SADDO and friends on 64-bit because we don't
1713 // handle type legalization for these operations here.
1715 // FIXME: We really should do custom legalization for addition and
1716 // subtraction on x86-32 once PR3203 is fixed. We really can't do much better
1717 // than generic legalization for 64-bit multiplication-with-overflow, though.
1718 for (auto VT : { MVT::i8, MVT::i16, MVT::i32, MVT::i64 }) {
1719 if (VT == MVT::i64 && !Subtarget->is64Bit())
1721 // Add/Sub/Mul with overflow operations are custom lowered.
1722 setOperationAction(ISD::SADDO, VT, Custom);
1723 setOperationAction(ISD::UADDO, VT, Custom);
1724 setOperationAction(ISD::SSUBO, VT, Custom);
1725 setOperationAction(ISD::USUBO, VT, Custom);
1726 setOperationAction(ISD::SMULO, VT, Custom);
1727 setOperationAction(ISD::UMULO, VT, Custom);
1730 if (!Subtarget->is64Bit()) {
1731 // These libcalls are not available in 32-bit.
1732 setLibcallName(RTLIB::SHL_I128, nullptr);
1733 setLibcallName(RTLIB::SRL_I128, nullptr);
1734 setLibcallName(RTLIB::SRA_I128, nullptr);
1737 // Combine sin / cos into one node or libcall if possible.
1738 if (Subtarget->hasSinCos()) {
1739 setLibcallName(RTLIB::SINCOS_F32, "sincosf");
1740 setLibcallName(RTLIB::SINCOS_F64, "sincos");
1741 if (Subtarget->isTargetDarwin()) {
1742 // For MacOSX, we don't want the normal expansion of a libcall to sincos.
1743 // We want to issue a libcall to __sincos_stret to avoid memory traffic.
1744 setOperationAction(ISD::FSINCOS, MVT::f64, Custom);
1745 setOperationAction(ISD::FSINCOS, MVT::f32, Custom);
1749 if (Subtarget->isTargetWin64()) {
1750 setOperationAction(ISD::SDIV, MVT::i128, Custom);
1751 setOperationAction(ISD::UDIV, MVT::i128, Custom);
1752 setOperationAction(ISD::SREM, MVT::i128, Custom);
1753 setOperationAction(ISD::UREM, MVT::i128, Custom);
1754 setOperationAction(ISD::SDIVREM, MVT::i128, Custom);
1755 setOperationAction(ISD::UDIVREM, MVT::i128, Custom);
1758 // We have target-specific dag combine patterns for the following nodes:
1759 setTargetDAGCombine(ISD::VECTOR_SHUFFLE);
1760 setTargetDAGCombine(ISD::EXTRACT_VECTOR_ELT);
1761 setTargetDAGCombine(ISD::BITCAST);
1762 setTargetDAGCombine(ISD::VSELECT);
1763 setTargetDAGCombine(ISD::SELECT);
1764 setTargetDAGCombine(ISD::SHL);
1765 setTargetDAGCombine(ISD::SRA);
1766 setTargetDAGCombine(ISD::SRL);
1767 setTargetDAGCombine(ISD::OR);
1768 setTargetDAGCombine(ISD::AND);
1769 setTargetDAGCombine(ISD::ADD);
1770 setTargetDAGCombine(ISD::FADD);
1771 setTargetDAGCombine(ISD::FSUB);
1772 setTargetDAGCombine(ISD::FMA);
1773 setTargetDAGCombine(ISD::SUB);
1774 setTargetDAGCombine(ISD::LOAD);
1775 setTargetDAGCombine(ISD::MLOAD);
1776 setTargetDAGCombine(ISD::STORE);
1777 setTargetDAGCombine(ISD::MSTORE);
1778 setTargetDAGCombine(ISD::ZERO_EXTEND);
1779 setTargetDAGCombine(ISD::ANY_EXTEND);
1780 setTargetDAGCombine(ISD::SIGN_EXTEND);
1781 setTargetDAGCombine(ISD::SIGN_EXTEND_INREG);
1782 setTargetDAGCombine(ISD::SINT_TO_FP);
1783 setTargetDAGCombine(ISD::UINT_TO_FP);
1784 setTargetDAGCombine(ISD::SETCC);
1785 setTargetDAGCombine(ISD::BUILD_VECTOR);
1786 setTargetDAGCombine(ISD::MUL);
1787 setTargetDAGCombine(ISD::XOR);
1789 computeRegisterProperties(Subtarget->getRegisterInfo());
1791 MaxStoresPerMemset = 16; // For @llvm.memset -> sequence of stores
1792 MaxStoresPerMemsetOptSize = 8;
1793 MaxStoresPerMemcpy = 8; // For @llvm.memcpy -> sequence of stores
1794 MaxStoresPerMemcpyOptSize = 4;
1795 MaxStoresPerMemmove = 8; // For @llvm.memmove -> sequence of stores
1796 MaxStoresPerMemmoveOptSize = 4;
1797 setPrefLoopAlignment(4); // 2^4 bytes.
1799 // A predictable cmov does not hurt on an in-order CPU.
1800 // FIXME: Use a CPU attribute to trigger this, not a CPU model.
1801 PredictableSelectIsExpensive = !Subtarget->isAtom();
1802 EnableExtLdPromotion = true;
1803 setPrefFunctionAlignment(4); // 2^4 bytes.
1805 verifyIntrinsicTables();
1808 // This has so far only been implemented for 64-bit MachO.
1809 bool X86TargetLowering::useLoadStackGuardNode() const {
1810 return Subtarget->isTargetMachO() && Subtarget->is64Bit();
1813 TargetLoweringBase::LegalizeTypeAction
1814 X86TargetLowering::getPreferredVectorAction(EVT VT) const {
1815 if (ExperimentalVectorWideningLegalization &&
1816 VT.getVectorNumElements() != 1 &&
1817 VT.getVectorElementType().getSimpleVT() != MVT::i1)
1818 return TypeWidenVector;
1820 return TargetLoweringBase::getPreferredVectorAction(VT);
1823 EVT X86TargetLowering::getSetCCResultType(const DataLayout &DL, LLVMContext &,
1826 return Subtarget->hasAVX512() ? MVT::i1: MVT::i8;
1828 const unsigned NumElts = VT.getVectorNumElements();
1829 const EVT EltVT = VT.getVectorElementType();
1830 if (VT.is512BitVector()) {
1831 if (Subtarget->hasAVX512())
1832 if (EltVT == MVT::i32 || EltVT == MVT::i64 ||
1833 EltVT == MVT::f32 || EltVT == MVT::f64)
1835 case 8: return MVT::v8i1;
1836 case 16: return MVT::v16i1;
1838 if (Subtarget->hasBWI())
1839 if (EltVT == MVT::i8 || EltVT == MVT::i16)
1841 case 32: return MVT::v32i1;
1842 case 64: return MVT::v64i1;
1846 if (VT.is256BitVector() || VT.is128BitVector()) {
1847 if (Subtarget->hasVLX())
1848 if (EltVT == MVT::i32 || EltVT == MVT::i64 ||
1849 EltVT == MVT::f32 || EltVT == MVT::f64)
1851 case 2: return MVT::v2i1;
1852 case 4: return MVT::v4i1;
1853 case 8: return MVT::v8i1;
1855 if (Subtarget->hasBWI() && Subtarget->hasVLX())
1856 if (EltVT == MVT::i8 || EltVT == MVT::i16)
1858 case 8: return MVT::v8i1;
1859 case 16: return MVT::v16i1;
1860 case 32: return MVT::v32i1;
1864 return VT.changeVectorElementTypeToInteger();
1867 /// Helper for getByValTypeAlignment to determine
1868 /// the desired ByVal argument alignment.
1869 static void getMaxByValAlign(Type *Ty, unsigned &MaxAlign) {
1872 if (VectorType *VTy = dyn_cast<VectorType>(Ty)) {
1873 if (VTy->getBitWidth() == 128)
1875 } else if (ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
1876 unsigned EltAlign = 0;
1877 getMaxByValAlign(ATy->getElementType(), EltAlign);
1878 if (EltAlign > MaxAlign)
1879 MaxAlign = EltAlign;
1880 } else if (StructType *STy = dyn_cast<StructType>(Ty)) {
1881 for (auto *EltTy : STy->elements()) {
1882 unsigned EltAlign = 0;
1883 getMaxByValAlign(EltTy, EltAlign);
1884 if (EltAlign > MaxAlign)
1885 MaxAlign = EltAlign;
1892 /// Return the desired alignment for ByVal aggregate
1893 /// function arguments in the caller parameter area. For X86, aggregates
1894 /// that contain SSE vectors are placed at 16-byte boundaries while the rest
1895 /// are at 4-byte boundaries.
1896 unsigned X86TargetLowering::getByValTypeAlignment(Type *Ty,
1897 const DataLayout &DL) const {
1898 if (Subtarget->is64Bit()) {
1899 // Max of 8 and alignment of type.
1900 unsigned TyAlign = DL.getABITypeAlignment(Ty);
1907 if (Subtarget->hasSSE1())
1908 getMaxByValAlign(Ty, Align);
1912 /// Returns the target specific optimal type for load
1913 /// and store operations as a result of memset, memcpy, and memmove
1914 /// lowering. If DstAlign is zero that means it's safe to destination
1915 /// alignment can satisfy any constraint. Similarly if SrcAlign is zero it
1916 /// means there isn't a need to check it against alignment requirement,
1917 /// probably because the source does not need to be loaded. If 'IsMemset' is
1918 /// true, that means it's expanding a memset. If 'ZeroMemset' is true, that
1919 /// means it's a memset of zero. 'MemcpyStrSrc' indicates whether the memcpy
1920 /// source is constant so it does not need to be loaded.
1921 /// It returns EVT::Other if the type should be determined using generic
1922 /// target-independent logic.
1924 X86TargetLowering::getOptimalMemOpType(uint64_t Size,
1925 unsigned DstAlign, unsigned SrcAlign,
1926 bool IsMemset, bool ZeroMemset,
1928 MachineFunction &MF) const {
1929 const Function *F = MF.getFunction();
1930 if ((!IsMemset || ZeroMemset) &&
1931 !F->hasFnAttribute(Attribute::NoImplicitFloat)) {
1933 (!Subtarget->isUnalignedMem16Slow() ||
1934 ((DstAlign == 0 || DstAlign >= 16) &&
1935 (SrcAlign == 0 || SrcAlign >= 16)))) {
1937 // FIXME: Check if unaligned 32-byte accesses are slow.
1938 if (Subtarget->hasInt256())
1940 if (Subtarget->hasFp256())
1943 if (Subtarget->hasSSE2())
1945 if (Subtarget->hasSSE1())
1947 } else if (!MemcpyStrSrc && Size >= 8 &&
1948 !Subtarget->is64Bit() &&
1949 Subtarget->hasSSE2()) {
1950 // Do not use f64 to lower memcpy if source is string constant. It's
1951 // better to use i32 to avoid the loads.
1955 // This is a compromise. If we reach here, unaligned accesses may be slow on
1956 // this target. However, creating smaller, aligned accesses could be even
1957 // slower and would certainly be a lot more code.
1958 if (Subtarget->is64Bit() && Size >= 8)
1963 bool X86TargetLowering::isSafeMemOpType(MVT VT) const {
1965 return X86ScalarSSEf32;
1966 else if (VT == MVT::f64)
1967 return X86ScalarSSEf64;
1972 X86TargetLowering::allowsMisalignedMemoryAccesses(EVT VT,
1977 switch (VT.getSizeInBits()) {
1979 // 8-byte and under are always assumed to be fast.
1983 *Fast = !Subtarget->isUnalignedMem16Slow();
1986 *Fast = !Subtarget->isUnalignedMem32Slow();
1988 // TODO: What about AVX-512 (512-bit) accesses?
1991 // Misaligned accesses of any size are always allowed.
1995 /// Return the entry encoding for a jump table in the
1996 /// current function. The returned value is a member of the
1997 /// MachineJumpTableInfo::JTEntryKind enum.
1998 unsigned X86TargetLowering::getJumpTableEncoding() const {
1999 // In GOT pic mode, each entry in the jump table is emitted as a @GOTOFF
2001 if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
2002 Subtarget->isPICStyleGOT())
2003 return MachineJumpTableInfo::EK_Custom32;
2005 // Otherwise, use the normal jump table encoding heuristics.
2006 return TargetLowering::getJumpTableEncoding();
2009 bool X86TargetLowering::useSoftFloat() const {
2010 return Subtarget->useSoftFloat();
2014 X86TargetLowering::LowerCustomJumpTableEntry(const MachineJumpTableInfo *MJTI,
2015 const MachineBasicBlock *MBB,
2016 unsigned uid,MCContext &Ctx) const{
2017 assert(MBB->getParent()->getTarget().getRelocationModel() == Reloc::PIC_ &&
2018 Subtarget->isPICStyleGOT());
2019 // In 32-bit ELF systems, our jump table entries are formed with @GOTOFF
2021 return MCSymbolRefExpr::create(MBB->getSymbol(),
2022 MCSymbolRefExpr::VK_GOTOFF, Ctx);
2025 /// Returns relocation base for the given PIC jumptable.
2026 SDValue X86TargetLowering::getPICJumpTableRelocBase(SDValue Table,
2027 SelectionDAG &DAG) const {
2028 if (!Subtarget->is64Bit())
2029 // This doesn't have SDLoc associated with it, but is not really the
2030 // same as a Register.
2031 return DAG.getNode(X86ISD::GlobalBaseReg, SDLoc(),
2032 getPointerTy(DAG.getDataLayout()));
2036 /// This returns the relocation base for the given PIC jumptable,
2037 /// the same as getPICJumpTableRelocBase, but as an MCExpr.
2038 const MCExpr *X86TargetLowering::
2039 getPICJumpTableRelocBaseExpr(const MachineFunction *MF, unsigned JTI,
2040 MCContext &Ctx) const {
2041 // X86-64 uses RIP relative addressing based on the jump table label.
2042 if (Subtarget->isPICStyleRIPRel())
2043 return TargetLowering::getPICJumpTableRelocBaseExpr(MF, JTI, Ctx);
2045 // Otherwise, the reference is relative to the PIC base.
2046 return MCSymbolRefExpr::create(MF->getPICBaseSymbol(), Ctx);
2049 std::pair<const TargetRegisterClass *, uint8_t>
2050 X86TargetLowering::findRepresentativeClass(const TargetRegisterInfo *TRI,
2052 const TargetRegisterClass *RRC = nullptr;
2054 switch (VT.SimpleTy) {
2056 return TargetLowering::findRepresentativeClass(TRI, VT);
2057 case MVT::i8: case MVT::i16: case MVT::i32: case MVT::i64:
2058 RRC = Subtarget->is64Bit() ? &X86::GR64RegClass : &X86::GR32RegClass;
2061 RRC = &X86::VR64RegClass;
2063 case MVT::f32: case MVT::f64:
2064 case MVT::v16i8: case MVT::v8i16: case MVT::v4i32: case MVT::v2i64:
2065 case MVT::v4f32: case MVT::v2f64:
2066 case MVT::v32i8: case MVT::v8i32: case MVT::v4i64: case MVT::v8f32:
2068 RRC = &X86::VR128RegClass;
2071 return std::make_pair(RRC, Cost);
2074 bool X86TargetLowering::getStackCookieLocation(unsigned &AddressSpace,
2075 unsigned &Offset) const {
2076 if (!Subtarget->isTargetLinux())
2079 if (Subtarget->is64Bit()) {
2080 // %fs:0x28, unless we're using a Kernel code model, in which case it's %gs:
2082 if (getTargetMachine().getCodeModel() == CodeModel::Kernel)
2094 Value *X86TargetLowering::getSafeStackPointerLocation(IRBuilder<> &IRB) const {
2095 if (!Subtarget->isTargetAndroid())
2096 return TargetLowering::getSafeStackPointerLocation(IRB);
2098 // Android provides a fixed TLS slot for the SafeStack pointer. See the
2099 // definition of TLS_SLOT_SAFESTACK in
2100 // https://android.googlesource.com/platform/bionic/+/master/libc/private/bionic_tls.h
2101 unsigned AddressSpace, Offset;
2102 if (Subtarget->is64Bit()) {
2103 // %fs:0x48, unless we're using a Kernel code model, in which case it's %gs:
2105 if (getTargetMachine().getCodeModel() == CodeModel::Kernel)
2115 return ConstantExpr::getIntToPtr(
2116 ConstantInt::get(Type::getInt32Ty(IRB.getContext()), Offset),
2117 Type::getInt8PtrTy(IRB.getContext())->getPointerTo(AddressSpace));
2120 bool X86TargetLowering::isNoopAddrSpaceCast(unsigned SrcAS,
2121 unsigned DestAS) const {
2122 assert(SrcAS != DestAS && "Expected different address spaces!");
2124 return SrcAS < 256 && DestAS < 256;
2127 //===----------------------------------------------------------------------===//
2128 // Return Value Calling Convention Implementation
2129 //===----------------------------------------------------------------------===//
2131 #include "X86GenCallingConv.inc"
2133 bool X86TargetLowering::CanLowerReturn(
2134 CallingConv::ID CallConv, MachineFunction &MF, bool isVarArg,
2135 const SmallVectorImpl<ISD::OutputArg> &Outs, LLVMContext &Context) const {
2136 SmallVector<CCValAssign, 16> RVLocs;
2137 CCState CCInfo(CallConv, isVarArg, MF, RVLocs, Context);
2138 return CCInfo.CheckReturn(Outs, RetCC_X86);
2141 const MCPhysReg *X86TargetLowering::getScratchRegisters(CallingConv::ID) const {
2142 static const MCPhysReg ScratchRegs[] = { X86::R11, 0 };
2147 X86TargetLowering::LowerReturn(SDValue Chain,
2148 CallingConv::ID CallConv, bool isVarArg,
2149 const SmallVectorImpl<ISD::OutputArg> &Outs,
2150 const SmallVectorImpl<SDValue> &OutVals,
2151 SDLoc dl, SelectionDAG &DAG) const {
2152 MachineFunction &MF = DAG.getMachineFunction();
2153 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
2155 SmallVector<CCValAssign, 16> RVLocs;
2156 CCState CCInfo(CallConv, isVarArg, MF, RVLocs, *DAG.getContext());
2157 CCInfo.AnalyzeReturn(Outs, RetCC_X86);
2160 SmallVector<SDValue, 6> RetOps;
2161 RetOps.push_back(Chain); // Operand #0 = Chain (updated below)
2162 // Operand #1 = Bytes To Pop
2163 RetOps.push_back(DAG.getTargetConstant(FuncInfo->getBytesToPopOnReturn(), dl,
2166 // Copy the result values into the output registers.
2167 for (unsigned i = 0; i != RVLocs.size(); ++i) {
2168 CCValAssign &VA = RVLocs[i];
2169 assert(VA.isRegLoc() && "Can only return in registers!");
2170 SDValue ValToCopy = OutVals[i];
2171 EVT ValVT = ValToCopy.getValueType();
2173 // Promote values to the appropriate types.
2174 if (VA.getLocInfo() == CCValAssign::SExt)
2175 ValToCopy = DAG.getNode(ISD::SIGN_EXTEND, dl, VA.getLocVT(), ValToCopy);
2176 else if (VA.getLocInfo() == CCValAssign::ZExt)
2177 ValToCopy = DAG.getNode(ISD::ZERO_EXTEND, dl, VA.getLocVT(), ValToCopy);
2178 else if (VA.getLocInfo() == CCValAssign::AExt) {
2179 if (ValVT.isVector() && ValVT.getScalarType() == MVT::i1)
2180 ValToCopy = DAG.getNode(ISD::SIGN_EXTEND, dl, VA.getLocVT(), ValToCopy);
2182 ValToCopy = DAG.getNode(ISD::ANY_EXTEND, dl, VA.getLocVT(), ValToCopy);
2184 else if (VA.getLocInfo() == CCValAssign::BCvt)
2185 ValToCopy = DAG.getBitcast(VA.getLocVT(), ValToCopy);
2187 assert(VA.getLocInfo() != CCValAssign::FPExt &&
2188 "Unexpected FP-extend for return value.");
2190 // If this is x86-64, and we disabled SSE, we can't return FP values,
2191 // or SSE or MMX vectors.
2192 if ((ValVT == MVT::f32 || ValVT == MVT::f64 ||
2193 VA.getLocReg() == X86::XMM0 || VA.getLocReg() == X86::XMM1) &&
2194 (Subtarget->is64Bit() && !Subtarget->hasSSE1())) {
2195 report_fatal_error("SSE register return with SSE disabled");
2197 // Likewise we can't return F64 values with SSE1 only. gcc does so, but
2198 // llvm-gcc has never done it right and no one has noticed, so this
2199 // should be OK for now.
2200 if (ValVT == MVT::f64 &&
2201 (Subtarget->is64Bit() && !Subtarget->hasSSE2()))
2202 report_fatal_error("SSE2 register return with SSE2 disabled");
2204 // Returns in ST0/ST1 are handled specially: these are pushed as operands to
2205 // the RET instruction and handled by the FP Stackifier.
2206 if (VA.getLocReg() == X86::FP0 ||
2207 VA.getLocReg() == X86::FP1) {
2208 // If this is a copy from an xmm register to ST(0), use an FPExtend to
2209 // change the value to the FP stack register class.
2210 if (isScalarFPTypeInSSEReg(VA.getValVT()))
2211 ValToCopy = DAG.getNode(ISD::FP_EXTEND, dl, MVT::f80, ValToCopy);
2212 RetOps.push_back(ValToCopy);
2213 // Don't emit a copytoreg.
2217 // 64-bit vector (MMX) values are returned in XMM0 / XMM1 except for v1i64
2218 // which is returned in RAX / RDX.
2219 if (Subtarget->is64Bit()) {
2220 if (ValVT == MVT::x86mmx) {
2221 if (VA.getLocReg() == X86::XMM0 || VA.getLocReg() == X86::XMM1) {
2222 ValToCopy = DAG.getBitcast(MVT::i64, ValToCopy);
2223 ValToCopy = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i64,
2225 // If we don't have SSE2 available, convert to v4f32 so the generated
2226 // register is legal.
2227 if (!Subtarget->hasSSE2())
2228 ValToCopy = DAG.getBitcast(MVT::v4f32, ValToCopy);
2233 Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), ValToCopy, Flag);
2234 Flag = Chain.getValue(1);
2235 RetOps.push_back(DAG.getRegister(VA.getLocReg(), VA.getLocVT()));
2238 // All x86 ABIs require that for returning structs by value we copy
2239 // the sret argument into %rax/%eax (depending on ABI) for the return.
2240 // We saved the argument into a virtual register in the entry block,
2241 // so now we copy the value out and into %rax/%eax.
2243 // Checking Function.hasStructRetAttr() here is insufficient because the IR
2244 // may not have an explicit sret argument. If FuncInfo.CanLowerReturn is
2245 // false, then an sret argument may be implicitly inserted in the SelDAG. In
2246 // either case FuncInfo->setSRetReturnReg() will have been called.
2247 if (unsigned SRetReg = FuncInfo->getSRetReturnReg()) {
2248 SDValue Val = DAG.getCopyFromReg(Chain, dl, SRetReg,
2249 getPointerTy(MF.getDataLayout()));
2252 = (Subtarget->is64Bit() && !Subtarget->isTarget64BitILP32()) ?
2253 X86::RAX : X86::EAX;
2254 Chain = DAG.getCopyToReg(Chain, dl, RetValReg, Val, Flag);
2255 Flag = Chain.getValue(1);
2257 // RAX/EAX now acts like a return value.
2259 DAG.getRegister(RetValReg, getPointerTy(DAG.getDataLayout())));
2262 RetOps[0] = Chain; // Update chain.
2264 // Add the flag if we have it.
2266 RetOps.push_back(Flag);
2268 return DAG.getNode(X86ISD::RET_FLAG, dl, MVT::Other, RetOps);
2271 bool X86TargetLowering::isUsedByReturnOnly(SDNode *N, SDValue &Chain) const {
2272 if (N->getNumValues() != 1)
2274 if (!N->hasNUsesOfValue(1, 0))
2277 SDValue TCChain = Chain;
2278 SDNode *Copy = *N->use_begin();
2279 if (Copy->getOpcode() == ISD::CopyToReg) {
2280 // If the copy has a glue operand, we conservatively assume it isn't safe to
2281 // perform a tail call.
2282 if (Copy->getOperand(Copy->getNumOperands()-1).getValueType() == MVT::Glue)
2284 TCChain = Copy->getOperand(0);
2285 } else if (Copy->getOpcode() != ISD::FP_EXTEND)
2288 bool HasRet = false;
2289 for (SDNode::use_iterator UI = Copy->use_begin(), UE = Copy->use_end();
2291 if (UI->getOpcode() != X86ISD::RET_FLAG)
2293 // If we are returning more than one value, we can definitely
2294 // not make a tail call see PR19530
2295 if (UI->getNumOperands() > 4)
2297 if (UI->getNumOperands() == 4 &&
2298 UI->getOperand(UI->getNumOperands()-1).getValueType() != MVT::Glue)
2311 X86TargetLowering::getTypeForExtArgOrReturn(LLVMContext &Context, EVT VT,
2312 ISD::NodeType ExtendKind) const {
2314 // TODO: Is this also valid on 32-bit?
2315 if (Subtarget->is64Bit() && VT == MVT::i1 && ExtendKind == ISD::ZERO_EXTEND)
2316 ReturnMVT = MVT::i8;
2318 ReturnMVT = MVT::i32;
2320 EVT MinVT = getRegisterType(Context, ReturnMVT);
2321 return VT.bitsLT(MinVT) ? MinVT : VT;
2324 /// Lower the result values of a call into the
2325 /// appropriate copies out of appropriate physical registers.
2328 X86TargetLowering::LowerCallResult(SDValue Chain, SDValue InFlag,
2329 CallingConv::ID CallConv, bool isVarArg,
2330 const SmallVectorImpl<ISD::InputArg> &Ins,
2331 SDLoc dl, SelectionDAG &DAG,
2332 SmallVectorImpl<SDValue> &InVals) const {
2334 // Assign locations to each value returned by this call.
2335 SmallVector<CCValAssign, 16> RVLocs;
2336 bool Is64Bit = Subtarget->is64Bit();
2337 CCState CCInfo(CallConv, isVarArg, DAG.getMachineFunction(), RVLocs,
2339 CCInfo.AnalyzeCallResult(Ins, RetCC_X86);
2341 // Copy all of the result registers out of their specified physreg.
2342 for (unsigned i = 0, e = RVLocs.size(); i != e; ++i) {
2343 CCValAssign &VA = RVLocs[i];
2344 EVT CopyVT = VA.getLocVT();
2346 // If this is x86-64, and we disabled SSE, we can't return FP values
2347 if ((CopyVT == MVT::f32 || CopyVT == MVT::f64) &&
2348 ((Is64Bit || Ins[i].Flags.isInReg()) && !Subtarget->hasSSE1())) {
2349 report_fatal_error("SSE register return with SSE disabled");
2352 // If we prefer to use the value in xmm registers, copy it out as f80 and
2353 // use a truncate to move it from fp stack reg to xmm reg.
2354 bool RoundAfterCopy = false;
2355 if ((VA.getLocReg() == X86::FP0 || VA.getLocReg() == X86::FP1) &&
2356 isScalarFPTypeInSSEReg(VA.getValVT())) {
2358 RoundAfterCopy = (CopyVT != VA.getLocVT());
2361 Chain = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(),
2362 CopyVT, InFlag).getValue(1);
2363 SDValue Val = Chain.getValue(0);
2366 Val = DAG.getNode(ISD::FP_ROUND, dl, VA.getValVT(), Val,
2367 // This truncation won't change the value.
2368 DAG.getIntPtrConstant(1, dl));
2370 if (VA.isExtInLoc() && VA.getValVT().getScalarType() == MVT::i1)
2371 Val = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), Val);
2373 InFlag = Chain.getValue(2);
2374 InVals.push_back(Val);
2380 //===----------------------------------------------------------------------===//
2381 // C & StdCall & Fast Calling Convention implementation
2382 //===----------------------------------------------------------------------===//
2383 // StdCall calling convention seems to be standard for many Windows' API
2384 // routines and around. It differs from C calling convention just a little:
2385 // callee should clean up the stack, not caller. Symbols should be also
2386 // decorated in some fancy way :) It doesn't support any vector arguments.
2387 // For info on fast calling convention see Fast Calling Convention (tail call)
2388 // implementation LowerX86_32FastCCCallTo.
2390 /// CallIsStructReturn - Determines whether a call uses struct return
2392 enum StructReturnType {
2397 static StructReturnType
2398 callIsStructReturn(const SmallVectorImpl<ISD::OutputArg> &Outs) {
2400 return NotStructReturn;
2402 const ISD::ArgFlagsTy &Flags = Outs[0].Flags;
2403 if (!Flags.isSRet())
2404 return NotStructReturn;
2405 if (Flags.isInReg())
2406 return RegStructReturn;
2407 return StackStructReturn;
2410 /// Determines whether a function uses struct return semantics.
2411 static StructReturnType
2412 argsAreStructReturn(const SmallVectorImpl<ISD::InputArg> &Ins) {
2414 return NotStructReturn;
2416 const ISD::ArgFlagsTy &Flags = Ins[0].Flags;
2417 if (!Flags.isSRet())
2418 return NotStructReturn;
2419 if (Flags.isInReg())
2420 return RegStructReturn;
2421 return StackStructReturn;
2424 /// Make a copy of an aggregate at address specified by "Src" to address
2425 /// "Dst" with size and alignment information specified by the specific
2426 /// parameter attribute. The copy will be passed as a byval function parameter.
2428 CreateCopyOfByValArgument(SDValue Src, SDValue Dst, SDValue Chain,
2429 ISD::ArgFlagsTy Flags, SelectionDAG &DAG,
2431 SDValue SizeNode = DAG.getConstant(Flags.getByValSize(), dl, MVT::i32);
2433 return DAG.getMemcpy(Chain, dl, Dst, Src, SizeNode, Flags.getByValAlign(),
2434 /*isVolatile*/false, /*AlwaysInline=*/true,
2435 /*isTailCall*/false,
2436 MachinePointerInfo(), MachinePointerInfo());
2439 /// Return true if the calling convention is one that we can guarantee TCO for.
2440 static bool canGuaranteeTCO(CallingConv::ID CC) {
2441 return (CC == CallingConv::Fast || CC == CallingConv::GHC ||
2442 CC == CallingConv::HiPE || CC == CallingConv::HHVM);
2445 /// Return true if we might ever do TCO for calls with this calling convention.
2446 static bool mayTailCallThisCC(CallingConv::ID CC) {
2448 // C calling conventions:
2449 case CallingConv::C:
2450 case CallingConv::X86_64_Win64:
2451 case CallingConv::X86_64_SysV:
2452 // Callee pop conventions:
2453 case CallingConv::X86_ThisCall:
2454 case CallingConv::X86_StdCall:
2455 case CallingConv::X86_VectorCall:
2456 case CallingConv::X86_FastCall:
2459 return canGuaranteeTCO(CC);
2463 /// Return true if the function is being made into a tailcall target by
2464 /// changing its ABI.
2465 static bool shouldGuaranteeTCO(CallingConv::ID CC, bool GuaranteedTailCallOpt) {
2466 return GuaranteedTailCallOpt && canGuaranteeTCO(CC);
2469 bool X86TargetLowering::mayBeEmittedAsTailCall(CallInst *CI) const {
2471 CI->getParent()->getParent()->getFnAttribute("disable-tail-calls");
2472 if (!CI->isTailCall() || Attr.getValueAsString() == "true")
2476 CallingConv::ID CalleeCC = CS.getCallingConv();
2477 if (!mayTailCallThisCC(CalleeCC))
2484 X86TargetLowering::LowerMemArgument(SDValue Chain,
2485 CallingConv::ID CallConv,
2486 const SmallVectorImpl<ISD::InputArg> &Ins,
2487 SDLoc dl, SelectionDAG &DAG,
2488 const CCValAssign &VA,
2489 MachineFrameInfo *MFI,
2491 // Create the nodes corresponding to a load from this parameter slot.
2492 ISD::ArgFlagsTy Flags = Ins[i].Flags;
2493 bool AlwaysUseMutable = shouldGuaranteeTCO(
2494 CallConv, DAG.getTarget().Options.GuaranteedTailCallOpt);
2495 bool isImmutable = !AlwaysUseMutable && !Flags.isByVal();
2498 // If value is passed by pointer we have address passed instead of the value
2500 bool ExtendedInMem = VA.isExtInLoc() &&
2501 VA.getValVT().getScalarType() == MVT::i1;
2503 if (VA.getLocInfo() == CCValAssign::Indirect || ExtendedInMem)
2504 ValVT = VA.getLocVT();
2506 ValVT = VA.getValVT();
2508 // FIXME: For now, all byval parameter objects are marked mutable. This can be
2509 // changed with more analysis.
2510 // In case of tail call optimization mark all arguments mutable. Since they
2511 // could be overwritten by lowering of arguments in case of a tail call.
2512 if (Flags.isByVal()) {
2513 unsigned Bytes = Flags.getByValSize();
2514 if (Bytes == 0) Bytes = 1; // Don't create zero-sized stack objects.
2515 int FI = MFI->CreateFixedObject(Bytes, VA.getLocMemOffset(), isImmutable);
2516 return DAG.getFrameIndex(FI, getPointerTy(DAG.getDataLayout()));
2518 int FI = MFI->CreateFixedObject(ValVT.getSizeInBits()/8,
2519 VA.getLocMemOffset(), isImmutable);
2520 SDValue FIN = DAG.getFrameIndex(FI, getPointerTy(DAG.getDataLayout()));
2521 SDValue Val = DAG.getLoad(
2522 ValVT, dl, Chain, FIN,
2523 MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), FI), false,
2525 return ExtendedInMem ?
2526 DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), Val) : Val;
2530 // FIXME: Get this from tablegen.
2531 static ArrayRef<MCPhysReg> get64BitArgumentGPRs(CallingConv::ID CallConv,
2532 const X86Subtarget *Subtarget) {
2533 assert(Subtarget->is64Bit());
2535 if (Subtarget->isCallingConvWin64(CallConv)) {
2536 static const MCPhysReg GPR64ArgRegsWin64[] = {
2537 X86::RCX, X86::RDX, X86::R8, X86::R9
2539 return makeArrayRef(std::begin(GPR64ArgRegsWin64), std::end(GPR64ArgRegsWin64));
2542 static const MCPhysReg GPR64ArgRegs64Bit[] = {
2543 X86::RDI, X86::RSI, X86::RDX, X86::RCX, X86::R8, X86::R9
2545 return makeArrayRef(std::begin(GPR64ArgRegs64Bit), std::end(GPR64ArgRegs64Bit));
2548 // FIXME: Get this from tablegen.
2549 static ArrayRef<MCPhysReg> get64BitArgumentXMMs(MachineFunction &MF,
2550 CallingConv::ID CallConv,
2551 const X86Subtarget *Subtarget) {
2552 assert(Subtarget->is64Bit());
2553 if (Subtarget->isCallingConvWin64(CallConv)) {
2554 // The XMM registers which might contain var arg parameters are shadowed
2555 // in their paired GPR. So we only need to save the GPR to their home
2557 // TODO: __vectorcall will change this.
2561 const Function *Fn = MF.getFunction();
2562 bool NoImplicitFloatOps = Fn->hasFnAttribute(Attribute::NoImplicitFloat);
2563 bool isSoftFloat = Subtarget->useSoftFloat();
2564 assert(!(isSoftFloat && NoImplicitFloatOps) &&
2565 "SSE register cannot be used when SSE is disabled!");
2566 if (isSoftFloat || NoImplicitFloatOps || !Subtarget->hasSSE1())
2567 // Kernel mode asks for SSE to be disabled, so there are no XMM argument
2571 static const MCPhysReg XMMArgRegs64Bit[] = {
2572 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
2573 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
2575 return makeArrayRef(std::begin(XMMArgRegs64Bit), std::end(XMMArgRegs64Bit));
2578 SDValue X86TargetLowering::LowerFormalArguments(
2579 SDValue Chain, CallingConv::ID CallConv, bool isVarArg,
2580 const SmallVectorImpl<ISD::InputArg> &Ins, SDLoc dl, SelectionDAG &DAG,
2581 SmallVectorImpl<SDValue> &InVals) const {
2582 MachineFunction &MF = DAG.getMachineFunction();
2583 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
2584 const TargetFrameLowering &TFI = *Subtarget->getFrameLowering();
2586 const Function* Fn = MF.getFunction();
2587 if (Fn->hasExternalLinkage() &&
2588 Subtarget->isTargetCygMing() &&
2589 Fn->getName() == "main")
2590 FuncInfo->setForceFramePointer(true);
2592 MachineFrameInfo *MFI = MF.getFrameInfo();
2593 bool Is64Bit = Subtarget->is64Bit();
2594 bool IsWin64 = Subtarget->isCallingConvWin64(CallConv);
2596 assert(!(isVarArg && canGuaranteeTCO(CallConv)) &&
2597 "Var args not supported with calling convention fastcc, ghc or hipe");
2599 // Assign locations to all of the incoming arguments.
2600 SmallVector<CCValAssign, 16> ArgLocs;
2601 CCState CCInfo(CallConv, isVarArg, MF, ArgLocs, *DAG.getContext());
2603 // Allocate shadow area for Win64
2605 CCInfo.AllocateStack(32, 8);
2607 CCInfo.AnalyzeFormalArguments(Ins, CC_X86);
2609 unsigned LastVal = ~0U;
2611 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2612 CCValAssign &VA = ArgLocs[i];
2613 // TODO: If an arg is passed in two places (e.g. reg and stack), skip later
2615 assert(VA.getValNo() != LastVal &&
2616 "Don't support value assigned to multiple locs yet");
2618 LastVal = VA.getValNo();
2620 if (VA.isRegLoc()) {
2621 EVT RegVT = VA.getLocVT();
2622 const TargetRegisterClass *RC;
2623 if (RegVT == MVT::i32)
2624 RC = &X86::GR32RegClass;
2625 else if (Is64Bit && RegVT == MVT::i64)
2626 RC = &X86::GR64RegClass;
2627 else if (RegVT == MVT::f32)
2628 RC = &X86::FR32RegClass;
2629 else if (RegVT == MVT::f64)
2630 RC = &X86::FR64RegClass;
2631 else if (RegVT.is512BitVector())
2632 RC = &X86::VR512RegClass;
2633 else if (RegVT.is256BitVector())
2634 RC = &X86::VR256RegClass;
2635 else if (RegVT.is128BitVector())
2636 RC = &X86::VR128RegClass;
2637 else if (RegVT == MVT::x86mmx)
2638 RC = &X86::VR64RegClass;
2639 else if (RegVT == MVT::i1)
2640 RC = &X86::VK1RegClass;
2641 else if (RegVT == MVT::v8i1)
2642 RC = &X86::VK8RegClass;
2643 else if (RegVT == MVT::v16i1)
2644 RC = &X86::VK16RegClass;
2645 else if (RegVT == MVT::v32i1)
2646 RC = &X86::VK32RegClass;
2647 else if (RegVT == MVT::v64i1)
2648 RC = &X86::VK64RegClass;
2650 llvm_unreachable("Unknown argument type!");
2652 unsigned Reg = MF.addLiveIn(VA.getLocReg(), RC);
2653 ArgValue = DAG.getCopyFromReg(Chain, dl, Reg, RegVT);
2655 // If this is an 8 or 16-bit value, it is really passed promoted to 32
2656 // bits. Insert an assert[sz]ext to capture this, then truncate to the
2658 if (VA.getLocInfo() == CCValAssign::SExt)
2659 ArgValue = DAG.getNode(ISD::AssertSext, dl, RegVT, ArgValue,
2660 DAG.getValueType(VA.getValVT()));
2661 else if (VA.getLocInfo() == CCValAssign::ZExt)
2662 ArgValue = DAG.getNode(ISD::AssertZext, dl, RegVT, ArgValue,
2663 DAG.getValueType(VA.getValVT()));
2664 else if (VA.getLocInfo() == CCValAssign::BCvt)
2665 ArgValue = DAG.getBitcast(VA.getValVT(), ArgValue);
2667 if (VA.isExtInLoc()) {
2668 // Handle MMX values passed in XMM regs.
2669 if (RegVT.isVector() && VA.getValVT().getScalarType() != MVT::i1)
2670 ArgValue = DAG.getNode(X86ISD::MOVDQ2Q, dl, VA.getValVT(), ArgValue);
2672 ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
2675 assert(VA.isMemLoc());
2676 ArgValue = LowerMemArgument(Chain, CallConv, Ins, dl, DAG, VA, MFI, i);
2679 // If value is passed via pointer - do a load.
2680 if (VA.getLocInfo() == CCValAssign::Indirect)
2681 ArgValue = DAG.getLoad(VA.getValVT(), dl, Chain, ArgValue,
2682 MachinePointerInfo(), false, false, false, 0);
2684 InVals.push_back(ArgValue);
2687 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
2688 // All x86 ABIs require that for returning structs by value we copy the
2689 // sret argument into %rax/%eax (depending on ABI) for the return. Save
2690 // the argument into a virtual register so that we can access it from the
2692 if (Ins[i].Flags.isSRet()) {
2693 unsigned Reg = FuncInfo->getSRetReturnReg();
2695 MVT PtrTy = getPointerTy(DAG.getDataLayout());
2696 Reg = MF.getRegInfo().createVirtualRegister(getRegClassFor(PtrTy));
2697 FuncInfo->setSRetReturnReg(Reg);
2699 SDValue Copy = DAG.getCopyToReg(DAG.getEntryNode(), dl, Reg, InVals[i]);
2700 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Copy, Chain);
2705 unsigned StackSize = CCInfo.getNextStackOffset();
2706 // Align stack specially for tail calls.
2707 if (shouldGuaranteeTCO(CallConv,
2708 MF.getTarget().Options.GuaranteedTailCallOpt))
2709 StackSize = GetAlignedArgumentStackSize(StackSize, DAG);
2711 // If the function takes variable number of arguments, make a frame index for
2712 // the start of the first vararg value... for expansion of llvm.va_start. We
2713 // can skip this if there are no va_start calls.
2714 if (MFI->hasVAStart() &&
2715 (Is64Bit || (CallConv != CallingConv::X86_FastCall &&
2716 CallConv != CallingConv::X86_ThisCall))) {
2717 FuncInfo->setVarArgsFrameIndex(
2718 MFI->CreateFixedObject(1, StackSize, true));
2721 MachineModuleInfo &MMI = MF.getMMI();
2723 // Figure out if XMM registers are in use.
2724 assert(!(Subtarget->useSoftFloat() &&
2725 Fn->hasFnAttribute(Attribute::NoImplicitFloat)) &&
2726 "SSE register cannot be used when SSE is disabled!");
2728 // 64-bit calling conventions support varargs and register parameters, so we
2729 // have to do extra work to spill them in the prologue.
2730 if (Is64Bit && isVarArg && MFI->hasVAStart()) {
2731 // Find the first unallocated argument registers.
2732 ArrayRef<MCPhysReg> ArgGPRs = get64BitArgumentGPRs(CallConv, Subtarget);
2733 ArrayRef<MCPhysReg> ArgXMMs = get64BitArgumentXMMs(MF, CallConv, Subtarget);
2734 unsigned NumIntRegs = CCInfo.getFirstUnallocated(ArgGPRs);
2735 unsigned NumXMMRegs = CCInfo.getFirstUnallocated(ArgXMMs);
2736 assert(!(NumXMMRegs && !Subtarget->hasSSE1()) &&
2737 "SSE register cannot be used when SSE is disabled!");
2739 // Gather all the live in physical registers.
2740 SmallVector<SDValue, 6> LiveGPRs;
2741 SmallVector<SDValue, 8> LiveXMMRegs;
2743 for (MCPhysReg Reg : ArgGPRs.slice(NumIntRegs)) {
2744 unsigned GPR = MF.addLiveIn(Reg, &X86::GR64RegClass);
2746 DAG.getCopyFromReg(Chain, dl, GPR, MVT::i64));
2748 if (!ArgXMMs.empty()) {
2749 unsigned AL = MF.addLiveIn(X86::AL, &X86::GR8RegClass);
2750 ALVal = DAG.getCopyFromReg(Chain, dl, AL, MVT::i8);
2751 for (MCPhysReg Reg : ArgXMMs.slice(NumXMMRegs)) {
2752 unsigned XMMReg = MF.addLiveIn(Reg, &X86::VR128RegClass);
2753 LiveXMMRegs.push_back(
2754 DAG.getCopyFromReg(Chain, dl, XMMReg, MVT::v4f32));
2759 // Get to the caller-allocated home save location. Add 8 to account
2760 // for the return address.
2761 int HomeOffset = TFI.getOffsetOfLocalArea() + 8;
2762 FuncInfo->setRegSaveFrameIndex(
2763 MFI->CreateFixedObject(1, NumIntRegs * 8 + HomeOffset, false));
2764 // Fixup to set vararg frame on shadow area (4 x i64).
2766 FuncInfo->setVarArgsFrameIndex(FuncInfo->getRegSaveFrameIndex());
2768 // For X86-64, if there are vararg parameters that are passed via
2769 // registers, then we must store them to their spots on the stack so
2770 // they may be loaded by deferencing the result of va_next.
2771 FuncInfo->setVarArgsGPOffset(NumIntRegs * 8);
2772 FuncInfo->setVarArgsFPOffset(ArgGPRs.size() * 8 + NumXMMRegs * 16);
2773 FuncInfo->setRegSaveFrameIndex(MFI->CreateStackObject(
2774 ArgGPRs.size() * 8 + ArgXMMs.size() * 16, 16, false));
2777 // Store the integer parameter registers.
2778 SmallVector<SDValue, 8> MemOps;
2779 SDValue RSFIN = DAG.getFrameIndex(FuncInfo->getRegSaveFrameIndex(),
2780 getPointerTy(DAG.getDataLayout()));
2781 unsigned Offset = FuncInfo->getVarArgsGPOffset();
2782 for (SDValue Val : LiveGPRs) {
2783 SDValue FIN = DAG.getNode(ISD::ADD, dl, getPointerTy(DAG.getDataLayout()),
2784 RSFIN, DAG.getIntPtrConstant(Offset, dl));
2786 DAG.getStore(Val.getValue(1), dl, Val, FIN,
2787 MachinePointerInfo::getFixedStack(
2788 DAG.getMachineFunction(),
2789 FuncInfo->getRegSaveFrameIndex(), Offset),
2791 MemOps.push_back(Store);
2795 if (!ArgXMMs.empty() && NumXMMRegs != ArgXMMs.size()) {
2796 // Now store the XMM (fp + vector) parameter registers.
2797 SmallVector<SDValue, 12> SaveXMMOps;
2798 SaveXMMOps.push_back(Chain);
2799 SaveXMMOps.push_back(ALVal);
2800 SaveXMMOps.push_back(DAG.getIntPtrConstant(
2801 FuncInfo->getRegSaveFrameIndex(), dl));
2802 SaveXMMOps.push_back(DAG.getIntPtrConstant(
2803 FuncInfo->getVarArgsFPOffset(), dl));
2804 SaveXMMOps.insert(SaveXMMOps.end(), LiveXMMRegs.begin(),
2806 MemOps.push_back(DAG.getNode(X86ISD::VASTART_SAVE_XMM_REGS, dl,
2807 MVT::Other, SaveXMMOps));
2810 if (!MemOps.empty())
2811 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, MemOps);
2814 if (isVarArg && MFI->hasMustTailInVarArgFunc()) {
2815 // Find the largest legal vector type.
2816 MVT VecVT = MVT::Other;
2817 // FIXME: Only some x86_32 calling conventions support AVX512.
2818 if (Subtarget->hasAVX512() &&
2819 (Is64Bit || (CallConv == CallingConv::X86_VectorCall ||
2820 CallConv == CallingConv::Intel_OCL_BI)))
2821 VecVT = MVT::v16f32;
2822 else if (Subtarget->hasAVX())
2824 else if (Subtarget->hasSSE2())
2827 // We forward some GPRs and some vector types.
2828 SmallVector<MVT, 2> RegParmTypes;
2829 MVT IntVT = Is64Bit ? MVT::i64 : MVT::i32;
2830 RegParmTypes.push_back(IntVT);
2831 if (VecVT != MVT::Other)
2832 RegParmTypes.push_back(VecVT);
2834 // Compute the set of forwarded registers. The rest are scratch.
2835 SmallVectorImpl<ForwardedRegister> &Forwards =
2836 FuncInfo->getForwardedMustTailRegParms();
2837 CCInfo.analyzeMustTailForwardedRegisters(Forwards, RegParmTypes, CC_X86);
2839 // Conservatively forward AL on x86_64, since it might be used for varargs.
2840 if (Is64Bit && !CCInfo.isAllocated(X86::AL)) {
2841 unsigned ALVReg = MF.addLiveIn(X86::AL, &X86::GR8RegClass);
2842 Forwards.push_back(ForwardedRegister(ALVReg, X86::AL, MVT::i8));
2845 // Copy all forwards from physical to virtual registers.
2846 for (ForwardedRegister &F : Forwards) {
2847 // FIXME: Can we use a less constrained schedule?
2848 SDValue RegVal = DAG.getCopyFromReg(Chain, dl, F.VReg, F.VT);
2849 F.VReg = MF.getRegInfo().createVirtualRegister(getRegClassFor(F.VT));
2850 Chain = DAG.getCopyToReg(Chain, dl, F.VReg, RegVal);
2854 // Some CCs need callee pop.
2855 if (X86::isCalleePop(CallConv, Is64Bit, isVarArg,
2856 MF.getTarget().Options.GuaranteedTailCallOpt)) {
2857 FuncInfo->setBytesToPopOnReturn(StackSize); // Callee pops everything.
2859 FuncInfo->setBytesToPopOnReturn(0); // Callee pops nothing.
2860 // If this is an sret function, the return should pop the hidden pointer.
2861 if (!Is64Bit && !canGuaranteeTCO(CallConv) &&
2862 !Subtarget->getTargetTriple().isOSMSVCRT() &&
2863 argsAreStructReturn(Ins) == StackStructReturn)
2864 FuncInfo->setBytesToPopOnReturn(4);
2868 // RegSaveFrameIndex is X86-64 only.
2869 FuncInfo->setRegSaveFrameIndex(0xAAAAAAA);
2870 if (CallConv == CallingConv::X86_FastCall ||
2871 CallConv == CallingConv::X86_ThisCall)
2872 // fastcc functions can't have varargs.
2873 FuncInfo->setVarArgsFrameIndex(0xAAAAAAA);
2876 FuncInfo->setArgumentStackSize(StackSize);
2878 if (MMI.hasWinEHFuncInfo(Fn)) {
2880 int UnwindHelpFI = MFI->CreateStackObject(8, 8, /*isSS=*/false);
2881 SDValue StackSlot = DAG.getFrameIndex(UnwindHelpFI, MVT::i64);
2882 MMI.getWinEHFuncInfo(MF.getFunction()).UnwindHelpFrameIdx = UnwindHelpFI;
2883 SDValue Neg2 = DAG.getConstant(-2, dl, MVT::i64);
2884 Chain = DAG.getStore(Chain, dl, Neg2, StackSlot,
2885 MachinePointerInfo::getFixedStack(
2886 DAG.getMachineFunction(), UnwindHelpFI),
2887 /*isVolatile=*/true,
2888 /*isNonTemporal=*/false, /*Alignment=*/0);
2890 // Functions using Win32 EH are considered to have opaque SP adjustments
2891 // to force local variables to be addressed from the frame or base
2893 MFI->setHasOpaqueSPAdjustment(true);
2901 X86TargetLowering::LowerMemOpCallTo(SDValue Chain,
2902 SDValue StackPtr, SDValue Arg,
2903 SDLoc dl, SelectionDAG &DAG,
2904 const CCValAssign &VA,
2905 ISD::ArgFlagsTy Flags) const {
2906 unsigned LocMemOffset = VA.getLocMemOffset();
2907 SDValue PtrOff = DAG.getIntPtrConstant(LocMemOffset, dl);
2908 PtrOff = DAG.getNode(ISD::ADD, dl, getPointerTy(DAG.getDataLayout()),
2910 if (Flags.isByVal())
2911 return CreateCopyOfByValArgument(Arg, PtrOff, Chain, Flags, DAG, dl);
2913 return DAG.getStore(
2914 Chain, dl, Arg, PtrOff,
2915 MachinePointerInfo::getStack(DAG.getMachineFunction(), LocMemOffset),
2919 /// Emit a load of return address if tail call
2920 /// optimization is performed and it is required.
2922 X86TargetLowering::EmitTailCallLoadRetAddr(SelectionDAG &DAG,
2923 SDValue &OutRetAddr, SDValue Chain,
2924 bool IsTailCall, bool Is64Bit,
2925 int FPDiff, SDLoc dl) const {
2926 // Adjust the Return address stack slot.
2927 EVT VT = getPointerTy(DAG.getDataLayout());
2928 OutRetAddr = getReturnAddressFrameIndex(DAG);
2930 // Load the "old" Return address.
2931 OutRetAddr = DAG.getLoad(VT, dl, Chain, OutRetAddr, MachinePointerInfo(),
2932 false, false, false, 0);
2933 return SDValue(OutRetAddr.getNode(), 1);
2936 /// Emit a store of the return address if tail call
2937 /// optimization is performed and it is required (FPDiff!=0).
2938 static SDValue EmitTailCallStoreRetAddr(SelectionDAG &DAG, MachineFunction &MF,
2939 SDValue Chain, SDValue RetAddrFrIdx,
2940 EVT PtrVT, unsigned SlotSize,
2941 int FPDiff, SDLoc dl) {
2942 // Store the return address to the appropriate stack slot.
2943 if (!FPDiff) return Chain;
2944 // Calculate the new stack slot for the return address.
2945 int NewReturnAddrFI =
2946 MF.getFrameInfo()->CreateFixedObject(SlotSize, (int64_t)FPDiff - SlotSize,
2948 SDValue NewRetAddrFrIdx = DAG.getFrameIndex(NewReturnAddrFI, PtrVT);
2949 Chain = DAG.getStore(Chain, dl, RetAddrFrIdx, NewRetAddrFrIdx,
2950 MachinePointerInfo::getFixedStack(
2951 DAG.getMachineFunction(), NewReturnAddrFI),
2956 /// Returns a vector_shuffle mask for an movs{s|d}, movd
2957 /// operation of specified width.
2958 static SDValue getMOVL(SelectionDAG &DAG, SDLoc dl, EVT VT, SDValue V1,
2960 unsigned NumElems = VT.getVectorNumElements();
2961 SmallVector<int, 8> Mask;
2962 Mask.push_back(NumElems);
2963 for (unsigned i = 1; i != NumElems; ++i)
2965 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
2969 X86TargetLowering::LowerCall(TargetLowering::CallLoweringInfo &CLI,
2970 SmallVectorImpl<SDValue> &InVals) const {
2971 SelectionDAG &DAG = CLI.DAG;
2973 SmallVectorImpl<ISD::OutputArg> &Outs = CLI.Outs;
2974 SmallVectorImpl<SDValue> &OutVals = CLI.OutVals;
2975 SmallVectorImpl<ISD::InputArg> &Ins = CLI.Ins;
2976 SDValue Chain = CLI.Chain;
2977 SDValue Callee = CLI.Callee;
2978 CallingConv::ID CallConv = CLI.CallConv;
2979 bool &isTailCall = CLI.IsTailCall;
2980 bool isVarArg = CLI.IsVarArg;
2982 MachineFunction &MF = DAG.getMachineFunction();
2983 bool Is64Bit = Subtarget->is64Bit();
2984 bool IsWin64 = Subtarget->isCallingConvWin64(CallConv);
2985 StructReturnType SR = callIsStructReturn(Outs);
2986 bool IsSibcall = false;
2987 X86MachineFunctionInfo *X86Info = MF.getInfo<X86MachineFunctionInfo>();
2988 auto Attr = MF.getFunction()->getFnAttribute("disable-tail-calls");
2990 if (Attr.getValueAsString() == "true")
2993 if (Subtarget->isPICStyleGOT() &&
2994 !MF.getTarget().Options.GuaranteedTailCallOpt) {
2995 // If we are using a GOT, disable tail calls to external symbols with
2996 // default visibility. Tail calling such a symbol requires using a GOT
2997 // relocation, which forces early binding of the symbol. This breaks code
2998 // that require lazy function symbol resolution. Using musttail or
2999 // GuaranteedTailCallOpt will override this.
3000 GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee);
3001 if (!G || (!G->getGlobal()->hasLocalLinkage() &&
3002 G->getGlobal()->hasDefaultVisibility()))
3006 bool IsMustTail = CLI.CS && CLI.CS->isMustTailCall();
3008 // Force this to be a tail call. The verifier rules are enough to ensure
3009 // that we can lower this successfully without moving the return address
3012 } else if (isTailCall) {
3013 // Check if it's really possible to do a tail call.
3014 isTailCall = IsEligibleForTailCallOptimization(Callee, CallConv,
3015 isVarArg, SR != NotStructReturn,
3016 MF.getFunction()->hasStructRetAttr(), CLI.RetTy,
3017 Outs, OutVals, Ins, DAG);
3019 // Sibcalls are automatically detected tailcalls which do not require
3021 if (!MF.getTarget().Options.GuaranteedTailCallOpt && isTailCall)
3028 assert(!(isVarArg && canGuaranteeTCO(CallConv)) &&
3029 "Var args not supported with calling convention fastcc, ghc or hipe");
3031 // Analyze operands of the call, assigning locations to each operand.
3032 SmallVector<CCValAssign, 16> ArgLocs;
3033 CCState CCInfo(CallConv, isVarArg, MF, ArgLocs, *DAG.getContext());
3035 // Allocate shadow area for Win64
3037 CCInfo.AllocateStack(32, 8);
3039 CCInfo.AnalyzeCallOperands(Outs, CC_X86);
3041 // Get a count of how many bytes are to be pushed on the stack.
3042 unsigned NumBytes = CCInfo.getAlignedCallFrameSize();
3044 // This is a sibcall. The memory operands are available in caller's
3045 // own caller's stack.
3047 else if (MF.getTarget().Options.GuaranteedTailCallOpt &&
3048 canGuaranteeTCO(CallConv))
3049 NumBytes = GetAlignedArgumentStackSize(NumBytes, DAG);
3052 if (isTailCall && !IsSibcall && !IsMustTail) {
3053 // Lower arguments at fp - stackoffset + fpdiff.
3054 unsigned NumBytesCallerPushed = X86Info->getBytesToPopOnReturn();
3056 FPDiff = NumBytesCallerPushed - NumBytes;
3058 // Set the delta of movement of the returnaddr stackslot.
3059 // But only set if delta is greater than previous delta.
3060 if (FPDiff < X86Info->getTCReturnAddrDelta())
3061 X86Info->setTCReturnAddrDelta(FPDiff);
3064 unsigned NumBytesToPush = NumBytes;
3065 unsigned NumBytesToPop = NumBytes;
3067 // If we have an inalloca argument, all stack space has already been allocated
3068 // for us and be right at the top of the stack. We don't support multiple
3069 // arguments passed in memory when using inalloca.
3070 if (!Outs.empty() && Outs.back().Flags.isInAlloca()) {
3072 if (!ArgLocs.back().isMemLoc())
3073 report_fatal_error("cannot use inalloca attribute on a register "
3075 if (ArgLocs.back().getLocMemOffset() != 0)
3076 report_fatal_error("any parameter with the inalloca attribute must be "
3077 "the only memory argument");
3081 Chain = DAG.getCALLSEQ_START(
3082 Chain, DAG.getIntPtrConstant(NumBytesToPush, dl, true), dl);
3084 SDValue RetAddrFrIdx;
3085 // Load return address for tail calls.
3086 if (isTailCall && FPDiff)
3087 Chain = EmitTailCallLoadRetAddr(DAG, RetAddrFrIdx, Chain, isTailCall,
3088 Is64Bit, FPDiff, dl);
3090 SmallVector<std::pair<unsigned, SDValue>, 8> RegsToPass;
3091 SmallVector<SDValue, 8> MemOpChains;
3094 // Walk the register/memloc assignments, inserting copies/loads. In the case
3095 // of tail call optimization arguments are handle later.
3096 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
3097 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
3098 // Skip inalloca arguments, they have already been written.
3099 ISD::ArgFlagsTy Flags = Outs[i].Flags;
3100 if (Flags.isInAlloca())
3103 CCValAssign &VA = ArgLocs[i];
3104 EVT RegVT = VA.getLocVT();
3105 SDValue Arg = OutVals[i];
3106 bool isByVal = Flags.isByVal();
3108 // Promote the value if needed.
3109 switch (VA.getLocInfo()) {
3110 default: llvm_unreachable("Unknown loc info!");
3111 case CCValAssign::Full: break;
3112 case CCValAssign::SExt:
3113 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, RegVT, Arg);
3115 case CCValAssign::ZExt:
3116 Arg = DAG.getNode(ISD::ZERO_EXTEND, dl, RegVT, Arg);
3118 case CCValAssign::AExt:
3119 if (Arg.getValueType().isVector() &&
3120 Arg.getValueType().getScalarType() == MVT::i1)
3121 Arg = DAG.getNode(ISD::SIGN_EXTEND, dl, RegVT, Arg);
3122 else if (RegVT.is128BitVector()) {
3123 // Special case: passing MMX values in XMM registers.
3124 Arg = DAG.getBitcast(MVT::i64, Arg);
3125 Arg = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i64, Arg);
3126 Arg = getMOVL(DAG, dl, MVT::v2i64, DAG.getUNDEF(MVT::v2i64), Arg);
3128 Arg = DAG.getNode(ISD::ANY_EXTEND, dl, RegVT, Arg);
3130 case CCValAssign::BCvt:
3131 Arg = DAG.getBitcast(RegVT, Arg);
3133 case CCValAssign::Indirect: {
3134 // Store the argument.
3135 SDValue SpillSlot = DAG.CreateStackTemporary(VA.getValVT());
3136 int FI = cast<FrameIndexSDNode>(SpillSlot)->getIndex();
3137 Chain = DAG.getStore(
3138 Chain, dl, Arg, SpillSlot,
3139 MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), FI),
3146 if (VA.isRegLoc()) {
3147 RegsToPass.push_back(std::make_pair(VA.getLocReg(), Arg));
3148 if (isVarArg && IsWin64) {
3149 // Win64 ABI requires argument XMM reg to be copied to the corresponding
3150 // shadow reg if callee is a varargs function.
3151 unsigned ShadowReg = 0;
3152 switch (VA.getLocReg()) {
3153 case X86::XMM0: ShadowReg = X86::RCX; break;
3154 case X86::XMM1: ShadowReg = X86::RDX; break;
3155 case X86::XMM2: ShadowReg = X86::R8; break;
3156 case X86::XMM3: ShadowReg = X86::R9; break;
3159 RegsToPass.push_back(std::make_pair(ShadowReg, Arg));
3161 } else if (!IsSibcall && (!isTailCall || isByVal)) {
3162 assert(VA.isMemLoc());
3163 if (!StackPtr.getNode())
3164 StackPtr = DAG.getCopyFromReg(Chain, dl, RegInfo->getStackRegister(),
3165 getPointerTy(DAG.getDataLayout()));
3166 MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Arg,
3167 dl, DAG, VA, Flags));
3171 if (!MemOpChains.empty())
3172 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, MemOpChains);
3174 if (Subtarget->isPICStyleGOT()) {
3175 // ELF / PIC requires GOT in the EBX register before function calls via PLT
3178 RegsToPass.push_back(std::make_pair(
3179 unsigned(X86::EBX), DAG.getNode(X86ISD::GlobalBaseReg, SDLoc(),
3180 getPointerTy(DAG.getDataLayout()))));
3182 // If we are tail calling and generating PIC/GOT style code load the
3183 // address of the callee into ECX. The value in ecx is used as target of
3184 // the tail jump. This is done to circumvent the ebx/callee-saved problem
3185 // for tail calls on PIC/GOT architectures. Normally we would just put the
3186 // address of GOT into ebx and then call target@PLT. But for tail calls
3187 // ebx would be restored (since ebx is callee saved) before jumping to the
3190 // Note: The actual moving to ECX is done further down.
3191 GlobalAddressSDNode *G = dyn_cast<GlobalAddressSDNode>(Callee);
3192 if (G && !G->getGlobal()->hasLocalLinkage() &&
3193 G->getGlobal()->hasDefaultVisibility())
3194 Callee = LowerGlobalAddress(Callee, DAG);
3195 else if (isa<ExternalSymbolSDNode>(Callee))
3196 Callee = LowerExternalSymbol(Callee, DAG);
3200 if (Is64Bit && isVarArg && !IsWin64 && !IsMustTail) {
3201 // From AMD64 ABI document:
3202 // For calls that may call functions that use varargs or stdargs
3203 // (prototype-less calls or calls to functions containing ellipsis (...) in
3204 // the declaration) %al is used as hidden argument to specify the number
3205 // of SSE registers used. The contents of %al do not need to match exactly
3206 // the number of registers, but must be an ubound on the number of SSE
3207 // registers used and is in the range 0 - 8 inclusive.
3209 // Count the number of XMM registers allocated.
3210 static const MCPhysReg XMMArgRegs[] = {
3211 X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
3212 X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
3214 unsigned NumXMMRegs = CCInfo.getFirstUnallocated(XMMArgRegs);
3215 assert((Subtarget->hasSSE1() || !NumXMMRegs)
3216 && "SSE registers cannot be used when SSE is disabled");
3218 RegsToPass.push_back(std::make_pair(unsigned(X86::AL),
3219 DAG.getConstant(NumXMMRegs, dl,
3223 if (isVarArg && IsMustTail) {
3224 const auto &Forwards = X86Info->getForwardedMustTailRegParms();
3225 for (const auto &F : Forwards) {
3226 SDValue Val = DAG.getCopyFromReg(Chain, dl, F.VReg, F.VT);
3227 RegsToPass.push_back(std::make_pair(unsigned(F.PReg), Val));
3231 // For tail calls lower the arguments to the 'real' stack slots. Sibcalls
3232 // don't need this because the eligibility check rejects calls that require
3233 // shuffling arguments passed in memory.
3234 if (!IsSibcall && isTailCall) {
3235 // Force all the incoming stack arguments to be loaded from the stack
3236 // before any new outgoing arguments are stored to the stack, because the
3237 // outgoing stack slots may alias the incoming argument stack slots, and
3238 // the alias isn't otherwise explicit. This is slightly more conservative
3239 // than necessary, because it means that each store effectively depends
3240 // on every argument instead of just those arguments it would clobber.
3241 SDValue ArgChain = DAG.getStackArgumentTokenFactor(Chain);
3243 SmallVector<SDValue, 8> MemOpChains2;
3246 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
3247 CCValAssign &VA = ArgLocs[i];
3250 assert(VA.isMemLoc());
3251 SDValue Arg = OutVals[i];
3252 ISD::ArgFlagsTy Flags = Outs[i].Flags;
3253 // Skip inalloca arguments. They don't require any work.
3254 if (Flags.isInAlloca())
3256 // Create frame index.
3257 int32_t Offset = VA.getLocMemOffset()+FPDiff;
3258 uint32_t OpSize = (VA.getLocVT().getSizeInBits()+7)/8;
3259 FI = MF.getFrameInfo()->CreateFixedObject(OpSize, Offset, true);
3260 FIN = DAG.getFrameIndex(FI, getPointerTy(DAG.getDataLayout()));
3262 if (Flags.isByVal()) {
3263 // Copy relative to framepointer.
3264 SDValue Source = DAG.getIntPtrConstant(VA.getLocMemOffset(), dl);
3265 if (!StackPtr.getNode())
3266 StackPtr = DAG.getCopyFromReg(Chain, dl, RegInfo->getStackRegister(),
3267 getPointerTy(DAG.getDataLayout()));
3268 Source = DAG.getNode(ISD::ADD, dl, getPointerTy(DAG.getDataLayout()),
3271 MemOpChains2.push_back(CreateCopyOfByValArgument(Source, FIN,
3275 // Store relative to framepointer.
3276 MemOpChains2.push_back(DAG.getStore(
3277 ArgChain, dl, Arg, FIN,
3278 MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), FI),
3283 if (!MemOpChains2.empty())
3284 Chain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, MemOpChains2);
3286 // Store the return address to the appropriate stack slot.
3287 Chain = EmitTailCallStoreRetAddr(DAG, MF, Chain, RetAddrFrIdx,
3288 getPointerTy(DAG.getDataLayout()),
3289 RegInfo->getSlotSize(), FPDiff, dl);
3292 // Build a sequence of copy-to-reg nodes chained together with token chain
3293 // and flag operands which copy the outgoing args into registers.
3295 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i) {
3296 Chain = DAG.getCopyToReg(Chain, dl, RegsToPass[i].first,
3297 RegsToPass[i].second, InFlag);
3298 InFlag = Chain.getValue(1);
3301 if (DAG.getTarget().getCodeModel() == CodeModel::Large) {
3302 assert(Is64Bit && "Large code model is only legal in 64-bit mode.");
3303 // In the 64-bit large code model, we have to make all calls
3304 // through a register, since the call instruction's 32-bit
3305 // pc-relative offset may not be large enough to hold the whole
3307 } else if (Callee->getOpcode() == ISD::GlobalAddress) {
3308 // If the callee is a GlobalAddress node (quite common, every direct call
3309 // is) turn it into a TargetGlobalAddress node so that legalize doesn't hack
3311 GlobalAddressSDNode* G = cast<GlobalAddressSDNode>(Callee);
3313 // We should use extra load for direct calls to dllimported functions in
3315 const GlobalValue *GV = G->getGlobal();
3316 if (!GV->hasDLLImportStorageClass()) {
3317 unsigned char OpFlags = 0;
3318 bool ExtraLoad = false;
3319 unsigned WrapperKind = ISD::DELETED_NODE;
3321 // On ELF targets, in both X86-64 and X86-32 mode, direct calls to
3322 // external symbols most go through the PLT in PIC mode. If the symbol
3323 // has hidden or protected visibility, or if it is static or local, then
3324 // we don't need to use the PLT - we can directly call it.
3325 if (Subtarget->isTargetELF() &&
3326 DAG.getTarget().getRelocationModel() == Reloc::PIC_ &&
3327 GV->hasDefaultVisibility() && !GV->hasLocalLinkage()) {
3328 OpFlags = X86II::MO_PLT;
3329 } else if (Subtarget->isPICStyleStubAny() &&
3330 !GV->isStrongDefinitionForLinker() &&
3331 (!Subtarget->getTargetTriple().isMacOSX() ||
3332 Subtarget->getTargetTriple().isMacOSXVersionLT(10, 5))) {
3333 // PC-relative references to external symbols should go through $stub,
3334 // unless we're building with the leopard linker or later, which
3335 // automatically synthesizes these stubs.
3336 OpFlags = X86II::MO_DARWIN_STUB;
3337 } else if (Subtarget->isPICStyleRIPRel() && isa<Function>(GV) &&
3338 cast<Function>(GV)->hasFnAttribute(Attribute::NonLazyBind)) {
3339 // If the function is marked as non-lazy, generate an indirect call
3340 // which loads from the GOT directly. This avoids runtime overhead
3341 // at the cost of eager binding (and one extra byte of encoding).
3342 OpFlags = X86II::MO_GOTPCREL;
3343 WrapperKind = X86ISD::WrapperRIP;
3347 Callee = DAG.getTargetGlobalAddress(
3348 GV, dl, getPointerTy(DAG.getDataLayout()), G->getOffset(), OpFlags);
3350 // Add a wrapper if needed.
3351 if (WrapperKind != ISD::DELETED_NODE)
3352 Callee = DAG.getNode(X86ISD::WrapperRIP, dl,
3353 getPointerTy(DAG.getDataLayout()), Callee);
3354 // Add extra indirection if needed.
3356 Callee = DAG.getLoad(
3357 getPointerTy(DAG.getDataLayout()), dl, DAG.getEntryNode(), Callee,
3358 MachinePointerInfo::getGOT(DAG.getMachineFunction()), false, false,
3361 } else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
3362 unsigned char OpFlags = 0;
3364 // On ELF targets, in either X86-64 or X86-32 mode, direct calls to
3365 // external symbols should go through the PLT.
3366 if (Subtarget->isTargetELF() &&
3367 DAG.getTarget().getRelocationModel() == Reloc::PIC_) {
3368 OpFlags = X86II::MO_PLT;
3369 } else if (Subtarget->isPICStyleStubAny() &&
3370 (!Subtarget->getTargetTriple().isMacOSX() ||
3371 Subtarget->getTargetTriple().isMacOSXVersionLT(10, 5))) {
3372 // PC-relative references to external symbols should go through $stub,
3373 // unless we're building with the leopard linker or later, which
3374 // automatically synthesizes these stubs.
3375 OpFlags = X86II::MO_DARWIN_STUB;
3378 Callee = DAG.getTargetExternalSymbol(
3379 S->getSymbol(), getPointerTy(DAG.getDataLayout()), OpFlags);
3380 } else if (Subtarget->isTarget64BitILP32() &&
3381 Callee->getValueType(0) == MVT::i32) {
3382 // Zero-extend the 32-bit Callee address into a 64-bit according to x32 ABI
3383 Callee = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i64, Callee);
3386 // Returns a chain & a flag for retval copy to use.
3387 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
3388 SmallVector<SDValue, 8> Ops;
3390 if (!IsSibcall && isTailCall) {
3391 Chain = DAG.getCALLSEQ_END(Chain,
3392 DAG.getIntPtrConstant(NumBytesToPop, dl, true),
3393 DAG.getIntPtrConstant(0, dl, true), InFlag, dl);
3394 InFlag = Chain.getValue(1);
3397 Ops.push_back(Chain);
3398 Ops.push_back(Callee);
3401 Ops.push_back(DAG.getConstant(FPDiff, dl, MVT::i32));
3403 // Add argument registers to the end of the list so that they are known live
3405 for (unsigned i = 0, e = RegsToPass.size(); i != e; ++i)
3406 Ops.push_back(DAG.getRegister(RegsToPass[i].first,
3407 RegsToPass[i].second.getValueType()));
3409 // Add a register mask operand representing the call-preserved registers.
3410 const uint32_t *Mask = RegInfo->getCallPreservedMask(MF, CallConv);
3411 assert(Mask && "Missing call preserved mask for calling convention");
3413 // If this is an invoke in a 32-bit function using a funclet-based
3414 // personality, assume the function clobbers all registers. If an exception
3415 // is thrown, the runtime will not restore CSRs.
3416 // FIXME: Model this more precisely so that we can register allocate across
3417 // the normal edge and spill and fill across the exceptional edge.
3418 if (!Is64Bit && CLI.CS && CLI.CS->isInvoke()) {
3419 const Function *CallerFn = MF.getFunction();
3420 EHPersonality Pers =
3421 CallerFn->hasPersonalityFn()
3422 ? classifyEHPersonality(CallerFn->getPersonalityFn())
3423 : EHPersonality::Unknown;
3424 if (isFuncletEHPersonality(Pers))
3425 Mask = RegInfo->getNoPreservedMask();
3428 Ops.push_back(DAG.getRegisterMask(Mask));
3430 if (InFlag.getNode())
3431 Ops.push_back(InFlag);
3435 //// If this is the first return lowered for this function, add the regs
3436 //// to the liveout set for the function.
3437 // This isn't right, although it's probably harmless on x86; liveouts
3438 // should be computed from returns not tail calls. Consider a void
3439 // function making a tail call to a function returning int.
3440 MF.getFrameInfo()->setHasTailCall();
3441 return DAG.getNode(X86ISD::TC_RETURN, dl, NodeTys, Ops);
3444 Chain = DAG.getNode(X86ISD::CALL, dl, NodeTys, Ops);
3445 InFlag = Chain.getValue(1);
3447 // Create the CALLSEQ_END node.
3448 unsigned NumBytesForCalleeToPop;
3449 if (X86::isCalleePop(CallConv, Is64Bit, isVarArg,
3450 DAG.getTarget().Options.GuaranteedTailCallOpt))
3451 NumBytesForCalleeToPop = NumBytes; // Callee pops everything
3452 else if (!Is64Bit && !canGuaranteeTCO(CallConv) &&
3453 !Subtarget->getTargetTriple().isOSMSVCRT() &&
3454 SR == StackStructReturn)
3455 // If this is a call to a struct-return function, the callee
3456 // pops the hidden struct pointer, so we have to push it back.
3457 // This is common for Darwin/X86, Linux & Mingw32 targets.
3458 // For MSVC Win32 targets, the caller pops the hidden struct pointer.
3459 NumBytesForCalleeToPop = 4;
3461 NumBytesForCalleeToPop = 0; // Callee pops nothing.
3463 // Returns a flag for retval copy to use.
3465 Chain = DAG.getCALLSEQ_END(Chain,
3466 DAG.getIntPtrConstant(NumBytesToPop, dl, true),
3467 DAG.getIntPtrConstant(NumBytesForCalleeToPop, dl,
3470 InFlag = Chain.getValue(1);
3473 // Handle result values, copying them out of physregs into vregs that we
3475 return LowerCallResult(Chain, InFlag, CallConv, isVarArg,
3476 Ins, dl, DAG, InVals);
3479 //===----------------------------------------------------------------------===//
3480 // Fast Calling Convention (tail call) implementation
3481 //===----------------------------------------------------------------------===//
3483 // Like std call, callee cleans arguments, convention except that ECX is
3484 // reserved for storing the tail called function address. Only 2 registers are
3485 // free for argument passing (inreg). Tail call optimization is performed
3487 // * tailcallopt is enabled
3488 // * caller/callee are fastcc
3489 // On X86_64 architecture with GOT-style position independent code only local
3490 // (within module) calls are supported at the moment.
3491 // To keep the stack aligned according to platform abi the function
3492 // GetAlignedArgumentStackSize ensures that argument delta is always multiples
3493 // of stack alignment. (Dynamic linkers need this - darwin's dyld for example)
3494 // If a tail called function callee has more arguments than the caller the
3495 // caller needs to make sure that there is room to move the RETADDR to. This is
3496 // achieved by reserving an area the size of the argument delta right after the
3497 // original RETADDR, but before the saved framepointer or the spilled registers
3498 // e.g. caller(arg1, arg2) calls callee(arg1, arg2,arg3,arg4)
3510 /// Make the stack size align e.g 16n + 12 aligned for a 16-byte align
3513 X86TargetLowering::GetAlignedArgumentStackSize(unsigned StackSize,
3514 SelectionDAG& DAG) const {
3515 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
3516 const TargetFrameLowering &TFI = *Subtarget->getFrameLowering();
3517 unsigned StackAlignment = TFI.getStackAlignment();
3518 uint64_t AlignMask = StackAlignment - 1;
3519 int64_t Offset = StackSize;
3520 unsigned SlotSize = RegInfo->getSlotSize();
3521 if ( (Offset & AlignMask) <= (StackAlignment - SlotSize) ) {
3522 // Number smaller than 12 so just add the difference.
3523 Offset += ((StackAlignment - SlotSize) - (Offset & AlignMask));
3525 // Mask out lower bits, add stackalignment once plus the 12 bytes.
3526 Offset = ((~AlignMask) & Offset) + StackAlignment +
3527 (StackAlignment-SlotSize);
3532 /// Return true if the given stack call argument is already available in the
3533 /// same position (relatively) of the caller's incoming argument stack.
3535 bool MatchingStackOffset(SDValue Arg, unsigned Offset, ISD::ArgFlagsTy Flags,
3536 MachineFrameInfo *MFI, const MachineRegisterInfo *MRI,
3537 const X86InstrInfo *TII) {
3538 unsigned Bytes = Arg.getValueType().getSizeInBits() / 8;
3540 if (Arg.getOpcode() == ISD::CopyFromReg) {
3541 unsigned VR = cast<RegisterSDNode>(Arg.getOperand(1))->getReg();
3542 if (!TargetRegisterInfo::isVirtualRegister(VR))
3544 MachineInstr *Def = MRI->getVRegDef(VR);
3547 if (!Flags.isByVal()) {
3548 if (!TII->isLoadFromStackSlot(Def, FI))
3551 unsigned Opcode = Def->getOpcode();
3552 if ((Opcode == X86::LEA32r || Opcode == X86::LEA64r ||
3553 Opcode == X86::LEA64_32r) &&
3554 Def->getOperand(1).isFI()) {
3555 FI = Def->getOperand(1).getIndex();
3556 Bytes = Flags.getByValSize();
3560 } else if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(Arg)) {
3561 if (Flags.isByVal())
3562 // ByVal argument is passed in as a pointer but it's now being
3563 // dereferenced. e.g.
3564 // define @foo(%struct.X* %A) {
3565 // tail call @bar(%struct.X* byval %A)
3568 SDValue Ptr = Ld->getBasePtr();
3569 FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr);
3572 FI = FINode->getIndex();
3573 } else if (Arg.getOpcode() == ISD::FrameIndex && Flags.isByVal()) {
3574 FrameIndexSDNode *FINode = cast<FrameIndexSDNode>(Arg);
3575 FI = FINode->getIndex();
3576 Bytes = Flags.getByValSize();
3580 assert(FI != INT_MAX);
3581 if (!MFI->isFixedObjectIndex(FI))
3583 return Offset == MFI->getObjectOffset(FI) && Bytes == MFI->getObjectSize(FI);
3586 /// Check whether the call is eligible for tail call optimization. Targets
3587 /// that want to do tail call optimization should implement this function.
3588 bool X86TargetLowering::IsEligibleForTailCallOptimization(
3589 SDValue Callee, CallingConv::ID CalleeCC, bool isVarArg,
3590 bool isCalleeStructRet, bool isCallerStructRet, Type *RetTy,
3591 const SmallVectorImpl<ISD::OutputArg> &Outs,
3592 const SmallVectorImpl<SDValue> &OutVals,
3593 const SmallVectorImpl<ISD::InputArg> &Ins, SelectionDAG &DAG) const {
3594 if (!mayTailCallThisCC(CalleeCC))
3597 // If -tailcallopt is specified, make fastcc functions tail-callable.
3598 MachineFunction &MF = DAG.getMachineFunction();
3599 const Function *CallerF = MF.getFunction();
3601 // If the function return type is x86_fp80 and the callee return type is not,
3602 // then the FP_EXTEND of the call result is not a nop. It's not safe to
3603 // perform a tailcall optimization here.
3604 if (CallerF->getReturnType()->isX86_FP80Ty() && !RetTy->isX86_FP80Ty())
3607 CallingConv::ID CallerCC = CallerF->getCallingConv();
3608 bool CCMatch = CallerCC == CalleeCC;
3609 bool IsCalleeWin64 = Subtarget->isCallingConvWin64(CalleeCC);
3610 bool IsCallerWin64 = Subtarget->isCallingConvWin64(CallerCC);
3612 // Win64 functions have extra shadow space for argument homing. Don't do the
3613 // sibcall if the caller and callee have mismatched expectations for this
3615 if (IsCalleeWin64 != IsCallerWin64)
3618 if (DAG.getTarget().Options.GuaranteedTailCallOpt) {
3619 if (canGuaranteeTCO(CalleeCC) && CCMatch)
3624 // Look for obvious safe cases to perform tail call optimization that do not
3625 // require ABI changes. This is what gcc calls sibcall.
3627 // Can't do sibcall if stack needs to be dynamically re-aligned. PEI needs to
3628 // emit a special epilogue.
3629 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
3630 if (RegInfo->needsStackRealignment(MF))
3633 // Also avoid sibcall optimization if either caller or callee uses struct
3634 // return semantics.
3635 if (isCalleeStructRet || isCallerStructRet)
3638 // Do not sibcall optimize vararg calls unless all arguments are passed via
3640 if (isVarArg && !Outs.empty()) {
3641 // Optimizing for varargs on Win64 is unlikely to be safe without
3642 // additional testing.
3643 if (IsCalleeWin64 || IsCallerWin64)
3646 SmallVector<CCValAssign, 16> ArgLocs;
3647 CCState CCInfo(CalleeCC, isVarArg, DAG.getMachineFunction(), ArgLocs,
3650 CCInfo.AnalyzeCallOperands(Outs, CC_X86);
3651 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i)
3652 if (!ArgLocs[i].isRegLoc())
3656 // If the call result is in ST0 / ST1, it needs to be popped off the x87
3657 // stack. Therefore, if it's not used by the call it is not safe to optimize
3658 // this into a sibcall.
3659 bool Unused = false;
3660 for (unsigned i = 0, e = Ins.size(); i != e; ++i) {
3667 SmallVector<CCValAssign, 16> RVLocs;
3668 CCState CCInfo(CalleeCC, false, DAG.getMachineFunction(), RVLocs,
3670 CCInfo.AnalyzeCallResult(Ins, RetCC_X86);
3671 for (unsigned i = 0, e = RVLocs.size(); i != e; ++i) {
3672 CCValAssign &VA = RVLocs[i];
3673 if (VA.getLocReg() == X86::FP0 || VA.getLocReg() == X86::FP1)
3678 // If the calling conventions do not match, then we'd better make sure the
3679 // results are returned in the same way as what the caller expects.
3681 SmallVector<CCValAssign, 16> RVLocs1;
3682 CCState CCInfo1(CalleeCC, false, DAG.getMachineFunction(), RVLocs1,
3684 CCInfo1.AnalyzeCallResult(Ins, RetCC_X86);
3686 SmallVector<CCValAssign, 16> RVLocs2;
3687 CCState CCInfo2(CallerCC, false, DAG.getMachineFunction(), RVLocs2,
3689 CCInfo2.AnalyzeCallResult(Ins, RetCC_X86);
3691 if (RVLocs1.size() != RVLocs2.size())
3693 for (unsigned i = 0, e = RVLocs1.size(); i != e; ++i) {
3694 if (RVLocs1[i].isRegLoc() != RVLocs2[i].isRegLoc())
3696 if (RVLocs1[i].getLocInfo() != RVLocs2[i].getLocInfo())
3698 if (RVLocs1[i].isRegLoc()) {
3699 if (RVLocs1[i].getLocReg() != RVLocs2[i].getLocReg())
3702 if (RVLocs1[i].getLocMemOffset() != RVLocs2[i].getLocMemOffset())
3708 unsigned StackArgsSize = 0;
3710 // If the callee takes no arguments then go on to check the results of the
3712 if (!Outs.empty()) {
3713 // Check if stack adjustment is needed. For now, do not do this if any
3714 // argument is passed on the stack.
3715 SmallVector<CCValAssign, 16> ArgLocs;
3716 CCState CCInfo(CalleeCC, isVarArg, DAG.getMachineFunction(), ArgLocs,
3719 // Allocate shadow area for Win64
3721 CCInfo.AllocateStack(32, 8);
3723 CCInfo.AnalyzeCallOperands(Outs, CC_X86);
3724 StackArgsSize = CCInfo.getNextStackOffset();
3726 if (CCInfo.getNextStackOffset()) {
3727 // Check if the arguments are already laid out in the right way as
3728 // the caller's fixed stack objects.
3729 MachineFrameInfo *MFI = MF.getFrameInfo();
3730 const MachineRegisterInfo *MRI = &MF.getRegInfo();
3731 const X86InstrInfo *TII = Subtarget->getInstrInfo();
3732 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
3733 CCValAssign &VA = ArgLocs[i];
3734 SDValue Arg = OutVals[i];
3735 ISD::ArgFlagsTy Flags = Outs[i].Flags;
3736 if (VA.getLocInfo() == CCValAssign::Indirect)
3738 if (!VA.isRegLoc()) {
3739 if (!MatchingStackOffset(Arg, VA.getLocMemOffset(), Flags,
3746 // If the tailcall address may be in a register, then make sure it's
3747 // possible to register allocate for it. In 32-bit, the call address can
3748 // only target EAX, EDX, or ECX since the tail call must be scheduled after
3749 // callee-saved registers are restored. These happen to be the same
3750 // registers used to pass 'inreg' arguments so watch out for those.
3751 if (!Subtarget->is64Bit() &&
3752 ((!isa<GlobalAddressSDNode>(Callee) &&
3753 !isa<ExternalSymbolSDNode>(Callee)) ||
3754 DAG.getTarget().getRelocationModel() == Reloc::PIC_)) {
3755 unsigned NumInRegs = 0;
3756 // In PIC we need an extra register to formulate the address computation
3758 unsigned MaxInRegs =
3759 (DAG.getTarget().getRelocationModel() == Reloc::PIC_) ? 2 : 3;
3761 for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
3762 CCValAssign &VA = ArgLocs[i];
3765 unsigned Reg = VA.getLocReg();
3768 case X86::EAX: case X86::EDX: case X86::ECX:
3769 if (++NumInRegs == MaxInRegs)
3777 bool CalleeWillPop =
3778 X86::isCalleePop(CalleeCC, Subtarget->is64Bit(), isVarArg,
3779 MF.getTarget().Options.GuaranteedTailCallOpt);
3781 if (unsigned BytesToPop =
3782 MF.getInfo<X86MachineFunctionInfo>()->getBytesToPopOnReturn()) {
3783 // If we have bytes to pop, the callee must pop them.
3784 bool CalleePopMatches = CalleeWillPop && BytesToPop == StackArgsSize;
3785 if (!CalleePopMatches)
3787 } else if (CalleeWillPop && StackArgsSize > 0) {
3788 // If we don't have bytes to pop, make sure the callee doesn't pop any.
3796 X86TargetLowering::createFastISel(FunctionLoweringInfo &funcInfo,
3797 const TargetLibraryInfo *libInfo) const {
3798 return X86::createFastISel(funcInfo, libInfo);
3801 //===----------------------------------------------------------------------===//
3802 // Other Lowering Hooks
3803 //===----------------------------------------------------------------------===//
3805 static bool MayFoldLoad(SDValue Op) {
3806 return Op.hasOneUse() && ISD::isNormalLoad(Op.getNode());
3809 static bool MayFoldIntoStore(SDValue Op) {
3810 return Op.hasOneUse() && ISD::isNormalStore(*Op.getNode()->use_begin());
3813 static bool isTargetShuffle(unsigned Opcode) {
3815 default: return false;
3816 case X86ISD::BLENDI:
3817 case X86ISD::PSHUFB:
3818 case X86ISD::PSHUFD:
3819 case X86ISD::PSHUFHW:
3820 case X86ISD::PSHUFLW:
3822 case X86ISD::PALIGNR:
3823 case X86ISD::MOVLHPS:
3824 case X86ISD::MOVLHPD:
3825 case X86ISD::MOVHLPS:
3826 case X86ISD::MOVLPS:
3827 case X86ISD::MOVLPD:
3828 case X86ISD::MOVSHDUP:
3829 case X86ISD::MOVSLDUP:
3830 case X86ISD::MOVDDUP:
3833 case X86ISD::UNPCKL:
3834 case X86ISD::UNPCKH:
3835 case X86ISD::VPERMILPI:
3836 case X86ISD::VPERM2X128:
3837 case X86ISD::VPERMI:
3838 case X86ISD::VPERMV:
3839 case X86ISD::VPERMV3:
3844 static SDValue getTargetShuffleNode(unsigned Opc, SDLoc dl, EVT VT,
3845 SDValue V1, unsigned TargetMask,
3846 SelectionDAG &DAG) {
3848 default: llvm_unreachable("Unknown x86 shuffle node");
3849 case X86ISD::PSHUFD:
3850 case X86ISD::PSHUFHW:
3851 case X86ISD::PSHUFLW:
3852 case X86ISD::VPERMILPI:
3853 case X86ISD::VPERMI:
3854 return DAG.getNode(Opc, dl, VT, V1,
3855 DAG.getConstant(TargetMask, dl, MVT::i8));
3859 static SDValue getTargetShuffleNode(unsigned Opc, SDLoc dl, EVT VT,
3860 SDValue V1, SDValue V2, SelectionDAG &DAG) {
3862 default: llvm_unreachable("Unknown x86 shuffle node");
3863 case X86ISD::MOVLHPS:
3864 case X86ISD::MOVLHPD:
3865 case X86ISD::MOVHLPS:
3866 case X86ISD::MOVLPS:
3867 case X86ISD::MOVLPD:
3870 case X86ISD::UNPCKL:
3871 case X86ISD::UNPCKH:
3872 return DAG.getNode(Opc, dl, VT, V1, V2);
3876 SDValue X86TargetLowering::getReturnAddressFrameIndex(SelectionDAG &DAG) const {
3877 MachineFunction &MF = DAG.getMachineFunction();
3878 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
3879 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
3880 int ReturnAddrIndex = FuncInfo->getRAIndex();
3882 if (ReturnAddrIndex == 0) {
3883 // Set up a frame object for the return address.
3884 unsigned SlotSize = RegInfo->getSlotSize();
3885 ReturnAddrIndex = MF.getFrameInfo()->CreateFixedObject(SlotSize,
3888 FuncInfo->setRAIndex(ReturnAddrIndex);
3891 return DAG.getFrameIndex(ReturnAddrIndex, getPointerTy(DAG.getDataLayout()));
3894 bool X86::isOffsetSuitableForCodeModel(int64_t Offset, CodeModel::Model M,
3895 bool hasSymbolicDisplacement) {
3896 // Offset should fit into 32 bit immediate field.
3897 if (!isInt<32>(Offset))
3900 // If we don't have a symbolic displacement - we don't have any extra
3902 if (!hasSymbolicDisplacement)
3905 // FIXME: Some tweaks might be needed for medium code model.
3906 if (M != CodeModel::Small && M != CodeModel::Kernel)
3909 // For small code model we assume that latest object is 16MB before end of 31
3910 // bits boundary. We may also accept pretty large negative constants knowing
3911 // that all objects are in the positive half of address space.
3912 if (M == CodeModel::Small && Offset < 16*1024*1024)
3915 // For kernel code model we know that all object resist in the negative half
3916 // of 32bits address space. We may not accept negative offsets, since they may
3917 // be just off and we may accept pretty large positive ones.
3918 if (M == CodeModel::Kernel && Offset >= 0)
3924 /// Determines whether the callee is required to pop its own arguments.
3925 /// Callee pop is necessary to support tail calls.
3926 bool X86::isCalleePop(CallingConv::ID CallingConv,
3927 bool is64Bit, bool IsVarArg, bool GuaranteeTCO) {
3928 // If GuaranteeTCO is true, we force some calls to be callee pop so that we
3929 // can guarantee TCO.
3930 if (!IsVarArg && shouldGuaranteeTCO(CallingConv, GuaranteeTCO))
3933 switch (CallingConv) {
3936 case CallingConv::X86_StdCall:
3937 case CallingConv::X86_FastCall:
3938 case CallingConv::X86_ThisCall:
3939 case CallingConv::X86_VectorCall:
3944 /// \brief Return true if the condition is an unsigned comparison operation.
3945 static bool isX86CCUnsigned(unsigned X86CC) {
3947 default: llvm_unreachable("Invalid integer condition!");
3948 case X86::COND_E: return true;
3949 case X86::COND_G: return false;
3950 case X86::COND_GE: return false;
3951 case X86::COND_L: return false;
3952 case X86::COND_LE: return false;
3953 case X86::COND_NE: return true;
3954 case X86::COND_B: return true;
3955 case X86::COND_A: return true;
3956 case X86::COND_BE: return true;
3957 case X86::COND_AE: return true;
3959 llvm_unreachable("covered switch fell through?!");
3962 /// Do a one-to-one translation of a ISD::CondCode to the X86-specific
3963 /// condition code, returning the condition code and the LHS/RHS of the
3964 /// comparison to make.
3965 static unsigned TranslateX86CC(ISD::CondCode SetCCOpcode, SDLoc DL, bool isFP,
3966 SDValue &LHS, SDValue &RHS, SelectionDAG &DAG) {
3968 if (ConstantSDNode *RHSC = dyn_cast<ConstantSDNode>(RHS)) {
3969 if (SetCCOpcode == ISD::SETGT && RHSC->isAllOnesValue()) {
3970 // X > -1 -> X == 0, jump !sign.
3971 RHS = DAG.getConstant(0, DL, RHS.getValueType());
3972 return X86::COND_NS;
3974 if (SetCCOpcode == ISD::SETLT && RHSC->isNullValue()) {
3975 // X < 0 -> X == 0, jump on sign.
3978 if (SetCCOpcode == ISD::SETLT && RHSC->getZExtValue() == 1) {
3980 RHS = DAG.getConstant(0, DL, RHS.getValueType());
3981 return X86::COND_LE;
3985 switch (SetCCOpcode) {
3986 default: llvm_unreachable("Invalid integer condition!");
3987 case ISD::SETEQ: return X86::COND_E;
3988 case ISD::SETGT: return X86::COND_G;
3989 case ISD::SETGE: return X86::COND_GE;
3990 case ISD::SETLT: return X86::COND_L;
3991 case ISD::SETLE: return X86::COND_LE;
3992 case ISD::SETNE: return X86::COND_NE;
3993 case ISD::SETULT: return X86::COND_B;
3994 case ISD::SETUGT: return X86::COND_A;
3995 case ISD::SETULE: return X86::COND_BE;
3996 case ISD::SETUGE: return X86::COND_AE;
4000 // First determine if it is required or is profitable to flip the operands.
4002 // If LHS is a foldable load, but RHS is not, flip the condition.
4003 if (ISD::isNON_EXTLoad(LHS.getNode()) &&
4004 !ISD::isNON_EXTLoad(RHS.getNode())) {
4005 SetCCOpcode = getSetCCSwappedOperands(SetCCOpcode);
4006 std::swap(LHS, RHS);
4009 switch (SetCCOpcode) {
4015 std::swap(LHS, RHS);
4019 // On a floating point condition, the flags are set as follows:
4021 // 0 | 0 | 0 | X > Y
4022 // 0 | 0 | 1 | X < Y
4023 // 1 | 0 | 0 | X == Y
4024 // 1 | 1 | 1 | unordered
4025 switch (SetCCOpcode) {
4026 default: llvm_unreachable("Condcode should be pre-legalized away");
4028 case ISD::SETEQ: return X86::COND_E;
4029 case ISD::SETOLT: // flipped
4031 case ISD::SETGT: return X86::COND_A;
4032 case ISD::SETOLE: // flipped
4034 case ISD::SETGE: return X86::COND_AE;
4035 case ISD::SETUGT: // flipped
4037 case ISD::SETLT: return X86::COND_B;
4038 case ISD::SETUGE: // flipped
4040 case ISD::SETLE: return X86::COND_BE;
4042 case ISD::SETNE: return X86::COND_NE;
4043 case ISD::SETUO: return X86::COND_P;
4044 case ISD::SETO: return X86::COND_NP;
4046 case ISD::SETUNE: return X86::COND_INVALID;
4050 /// Is there a floating point cmov for the specific X86 condition code?
4051 /// Current x86 isa includes the following FP cmov instructions:
4052 /// fcmovb, fcomvbe, fcomve, fcmovu, fcmovae, fcmova, fcmovne, fcmovnu.
4053 static bool hasFPCMov(unsigned X86CC) {
4069 /// Returns true if the target can instruction select the
4070 /// specified FP immediate natively. If false, the legalizer will
4071 /// materialize the FP immediate as a load from a constant pool.
4072 bool X86TargetLowering::isFPImmLegal(const APFloat &Imm, EVT VT) const {
4073 for (unsigned i = 0, e = LegalFPImmediates.size(); i != e; ++i) {
4074 if (Imm.bitwiseIsEqual(LegalFPImmediates[i]))
4080 bool X86TargetLowering::shouldReduceLoadWidth(SDNode *Load,
4081 ISD::LoadExtType ExtTy,
4083 // "ELF Handling for Thread-Local Storage" specifies that R_X86_64_GOTTPOFF
4084 // relocation target a movq or addq instruction: don't let the load shrink.
4085 SDValue BasePtr = cast<LoadSDNode>(Load)->getBasePtr();
4086 if (BasePtr.getOpcode() == X86ISD::WrapperRIP)
4087 if (const auto *GA = dyn_cast<GlobalAddressSDNode>(BasePtr.getOperand(0)))
4088 return GA->getTargetFlags() != X86II::MO_GOTTPOFF;
4092 /// \brief Returns true if it is beneficial to convert a load of a constant
4093 /// to just the constant itself.
4094 bool X86TargetLowering::shouldConvertConstantLoadToIntImm(const APInt &Imm,
4096 assert(Ty->isIntegerTy());
4098 unsigned BitSize = Ty->getPrimitiveSizeInBits();
4099 if (BitSize == 0 || BitSize > 64)
4104 bool X86TargetLowering::isExtractSubvectorCheap(EVT ResVT,
4105 unsigned Index) const {
4106 if (!isOperationLegalOrCustom(ISD::EXTRACT_SUBVECTOR, ResVT))
4109 return (Index == 0 || Index == ResVT.getVectorNumElements());
4112 bool X86TargetLowering::isCheapToSpeculateCttz() const {
4113 // Speculate cttz only if we can directly use TZCNT.
4114 return Subtarget->hasBMI();
4117 bool X86TargetLowering::isCheapToSpeculateCtlz() const {
4118 // Speculate ctlz only if we can directly use LZCNT.
4119 return Subtarget->hasLZCNT();
4122 /// Return true if every element in Mask, beginning
4123 /// from position Pos and ending in Pos+Size is undef.
4124 static bool isUndefInRange(ArrayRef<int> Mask, unsigned Pos, unsigned Size) {
4125 for (unsigned i = Pos, e = Pos + Size; i != e; ++i)
4131 /// Return true if Val is undef or if its value falls within the
4132 /// specified range (L, H].
4133 static bool isUndefOrInRange(int Val, int Low, int Hi) {
4134 return (Val < 0) || (Val >= Low && Val < Hi);
4137 /// Val is either less than zero (undef) or equal to the specified value.
4138 static bool isUndefOrEqual(int Val, int CmpVal) {
4139 return (Val < 0 || Val == CmpVal);
4142 /// Return true if every element in Mask, beginning
4143 /// from position Pos and ending in Pos+Size, falls within the specified
4144 /// sequential range (Low, Low+Size]. or is undef.
4145 static bool isSequentialOrUndefInRange(ArrayRef<int> Mask,
4146 unsigned Pos, unsigned Size, int Low) {
4147 for (unsigned i = Pos, e = Pos+Size; i != e; ++i, ++Low)
4148 if (!isUndefOrEqual(Mask[i], Low))
4153 /// Return true if the specified EXTRACT_SUBVECTOR operand specifies a vector
4154 /// extract that is suitable for instruction that extract 128 or 256 bit vectors
4155 static bool isVEXTRACTIndex(SDNode *N, unsigned vecWidth) {
4156 assert((vecWidth == 128 || vecWidth == 256) && "Unexpected vector width");
4157 if (!isa<ConstantSDNode>(N->getOperand(1).getNode()))
4160 // The index should be aligned on a vecWidth-bit boundary.
4162 cast<ConstantSDNode>(N->getOperand(1).getNode())->getZExtValue();
4164 MVT VT = N->getSimpleValueType(0);
4165 unsigned ElSize = VT.getVectorElementType().getSizeInBits();
4166 bool Result = (Index * ElSize) % vecWidth == 0;
4171 /// Return true if the specified INSERT_SUBVECTOR
4172 /// operand specifies a subvector insert that is suitable for input to
4173 /// insertion of 128 or 256-bit subvectors
4174 static bool isVINSERTIndex(SDNode *N, unsigned vecWidth) {
4175 assert((vecWidth == 128 || vecWidth == 256) && "Unexpected vector width");
4176 if (!isa<ConstantSDNode>(N->getOperand(2).getNode()))
4178 // The index should be aligned on a vecWidth-bit boundary.
4180 cast<ConstantSDNode>(N->getOperand(2).getNode())->getZExtValue();
4182 MVT VT = N->getSimpleValueType(0);
4183 unsigned ElSize = VT.getVectorElementType().getSizeInBits();
4184 bool Result = (Index * ElSize) % vecWidth == 0;
4189 bool X86::isVINSERT128Index(SDNode *N) {
4190 return isVINSERTIndex(N, 128);
4193 bool X86::isVINSERT256Index(SDNode *N) {
4194 return isVINSERTIndex(N, 256);
4197 bool X86::isVEXTRACT128Index(SDNode *N) {
4198 return isVEXTRACTIndex(N, 128);
4201 bool X86::isVEXTRACT256Index(SDNode *N) {
4202 return isVEXTRACTIndex(N, 256);
4205 static unsigned getExtractVEXTRACTImmediate(SDNode *N, unsigned vecWidth) {
4206 assert((vecWidth == 128 || vecWidth == 256) && "Unsupported vector width");
4207 if (!isa<ConstantSDNode>(N->getOperand(1).getNode()))
4208 llvm_unreachable("Illegal extract subvector for VEXTRACT");
4211 cast<ConstantSDNode>(N->getOperand(1).getNode())->getZExtValue();
4213 MVT VecVT = N->getOperand(0).getSimpleValueType();
4214 MVT ElVT = VecVT.getVectorElementType();
4216 unsigned NumElemsPerChunk = vecWidth / ElVT.getSizeInBits();
4217 return Index / NumElemsPerChunk;
4220 static unsigned getInsertVINSERTImmediate(SDNode *N, unsigned vecWidth) {
4221 assert((vecWidth == 128 || vecWidth == 256) && "Unsupported vector width");
4222 if (!isa<ConstantSDNode>(N->getOperand(2).getNode()))
4223 llvm_unreachable("Illegal insert subvector for VINSERT");
4226 cast<ConstantSDNode>(N->getOperand(2).getNode())->getZExtValue();
4228 MVT VecVT = N->getSimpleValueType(0);
4229 MVT ElVT = VecVT.getVectorElementType();
4231 unsigned NumElemsPerChunk = vecWidth / ElVT.getSizeInBits();
4232 return Index / NumElemsPerChunk;
4235 /// Return the appropriate immediate to extract the specified
4236 /// EXTRACT_SUBVECTOR index with VEXTRACTF128 and VINSERTI128 instructions.
4237 unsigned X86::getExtractVEXTRACT128Immediate(SDNode *N) {
4238 return getExtractVEXTRACTImmediate(N, 128);
4241 /// Return the appropriate immediate to extract the specified
4242 /// EXTRACT_SUBVECTOR index with VEXTRACTF64x4 and VINSERTI64x4 instructions.
4243 unsigned X86::getExtractVEXTRACT256Immediate(SDNode *N) {
4244 return getExtractVEXTRACTImmediate(N, 256);
4247 /// Return the appropriate immediate to insert at the specified
4248 /// INSERT_SUBVECTOR index with VINSERTF128 and VINSERTI128 instructions.
4249 unsigned X86::getInsertVINSERT128Immediate(SDNode *N) {
4250 return getInsertVINSERTImmediate(N, 128);
4253 /// Return the appropriate immediate to insert at the specified
4254 /// INSERT_SUBVECTOR index with VINSERTF46x4 and VINSERTI64x4 instructions.
4255 unsigned X86::getInsertVINSERT256Immediate(SDNode *N) {
4256 return getInsertVINSERTImmediate(N, 256);
4259 /// Returns true if V is a constant integer zero.
4260 static bool isZero(SDValue V) {
4261 ConstantSDNode *C = dyn_cast<ConstantSDNode>(V);
4262 return C && C->isNullValue();
4265 /// Returns true if Elt is a constant zero or a floating point constant +0.0.
4266 bool X86::isZeroNode(SDValue Elt) {
4269 if (ConstantFPSDNode *CFP = dyn_cast<ConstantFPSDNode>(Elt))
4270 return CFP->getValueAPF().isPosZero();
4274 // Build a vector of constants
4275 // Use an UNDEF node if MaskElt == -1.
4276 // Spilt 64-bit constants in the 32-bit mode.
4277 static SDValue getConstVector(ArrayRef<int> Values, EVT VT,
4279 SDLoc dl, bool IsMask = false) {
4281 SmallVector<SDValue, 32> Ops;
4284 EVT ConstVecVT = VT;
4285 unsigned NumElts = VT.getVectorNumElements();
4286 bool In64BitMode = DAG.getTargetLoweringInfo().isTypeLegal(MVT::i64);
4287 if (!In64BitMode && VT.getScalarType() == MVT::i64) {
4288 ConstVecVT = MVT::getVectorVT(MVT::i32, NumElts * 2);
4292 EVT EltVT = ConstVecVT.getScalarType();
4293 for (unsigned i = 0; i < NumElts; ++i) {
4294 bool IsUndef = Values[i] < 0 && IsMask;
4295 SDValue OpNode = IsUndef ? DAG.getUNDEF(EltVT) :
4296 DAG.getConstant(Values[i], dl, EltVT);
4297 Ops.push_back(OpNode);
4299 Ops.push_back(IsUndef ? DAG.getUNDEF(EltVT) :
4300 DAG.getConstant(0, dl, EltVT));
4302 SDValue ConstsNode = DAG.getNode(ISD::BUILD_VECTOR, dl, ConstVecVT, Ops);
4304 ConstsNode = DAG.getBitcast(VT, ConstsNode);
4308 /// Returns a vector of specified type with all zero elements.
4309 static SDValue getZeroVector(EVT VT, const X86Subtarget *Subtarget,
4310 SelectionDAG &DAG, SDLoc dl) {
4311 assert(VT.isVector() && "Expected a vector type");
4313 // Always build SSE zero vectors as <4 x i32> bitcasted
4314 // to their dest type. This ensures they get CSE'd.
4316 if (VT.is128BitVector()) { // SSE
4317 if (Subtarget->hasSSE2()) { // SSE2
4318 SDValue Cst = DAG.getConstant(0, dl, MVT::i32);
4319 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
4321 SDValue Cst = DAG.getConstantFP(+0.0, dl, MVT::f32);
4322 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4f32, Cst, Cst, Cst, Cst);
4324 } else if (VT.is256BitVector()) { // AVX
4325 if (Subtarget->hasInt256()) { // AVX2
4326 SDValue Cst = DAG.getConstant(0, dl, MVT::i32);
4327 SDValue Ops[] = { Cst, Cst, Cst, Cst, Cst, Cst, Cst, Cst };
4328 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v8i32, Ops);
4330 // 256-bit logic and arithmetic instructions in AVX are all
4331 // floating-point, no support for integer ops. Emit fp zeroed vectors.
4332 SDValue Cst = DAG.getConstantFP(+0.0, dl, MVT::f32);
4333 SDValue Ops[] = { Cst, Cst, Cst, Cst, Cst, Cst, Cst, Cst };
4334 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v8f32, Ops);
4336 } else if (VT.is512BitVector()) { // AVX-512
4337 SDValue Cst = DAG.getConstant(0, dl, MVT::i32);
4338 SDValue Ops[] = { Cst, Cst, Cst, Cst, Cst, Cst, Cst, Cst,
4339 Cst, Cst, Cst, Cst, Cst, Cst, Cst, Cst };
4340 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v16i32, Ops);
4341 } else if (VT.getScalarType() == MVT::i1) {
4343 assert((Subtarget->hasBWI() || VT.getVectorNumElements() <= 16)
4344 && "Unexpected vector type");
4345 assert((Subtarget->hasVLX() || VT.getVectorNumElements() >= 8)
4346 && "Unexpected vector type");
4347 SDValue Cst = DAG.getConstant(0, dl, MVT::i1);
4348 SmallVector<SDValue, 64> Ops(VT.getVectorNumElements(), Cst);
4349 return DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Ops);
4351 llvm_unreachable("Unexpected vector type");
4353 return DAG.getBitcast(VT, Vec);
4356 static SDValue ExtractSubVector(SDValue Vec, unsigned IdxVal,
4357 SelectionDAG &DAG, SDLoc dl,
4358 unsigned vectorWidth) {
4359 assert((vectorWidth == 128 || vectorWidth == 256) &&
4360 "Unsupported vector width");
4361 EVT VT = Vec.getValueType();
4362 EVT ElVT = VT.getVectorElementType();
4363 unsigned Factor = VT.getSizeInBits()/vectorWidth;
4364 EVT ResultVT = EVT::getVectorVT(*DAG.getContext(), ElVT,
4365 VT.getVectorNumElements()/Factor);
4367 // Extract from UNDEF is UNDEF.
4368 if (Vec.getOpcode() == ISD::UNDEF)
4369 return DAG.getUNDEF(ResultVT);
4371 // Extract the relevant vectorWidth bits. Generate an EXTRACT_SUBVECTOR
4372 unsigned ElemsPerChunk = vectorWidth / ElVT.getSizeInBits();
4374 // This is the index of the first element of the vectorWidth-bit chunk
4376 unsigned NormalizedIdxVal = (((IdxVal * ElVT.getSizeInBits()) / vectorWidth)
4379 // If the input is a buildvector just emit a smaller one.
4380 if (Vec.getOpcode() == ISD::BUILD_VECTOR)
4381 return DAG.getNode(ISD::BUILD_VECTOR, dl, ResultVT,
4382 makeArrayRef(Vec->op_begin() + NormalizedIdxVal,
4385 SDValue VecIdx = DAG.getIntPtrConstant(NormalizedIdxVal, dl);
4386 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, ResultVT, Vec, VecIdx);
4389 /// Generate a DAG to grab 128-bits from a vector > 128 bits. This
4390 /// sets things up to match to an AVX VEXTRACTF128 / VEXTRACTI128
4391 /// or AVX-512 VEXTRACTF32x4 / VEXTRACTI32x4
4392 /// instructions or a simple subregister reference. Idx is an index in the
4393 /// 128 bits we want. It need not be aligned to a 128-bit boundary. That makes
4394 /// lowering EXTRACT_VECTOR_ELT operations easier.
4395 static SDValue Extract128BitVector(SDValue Vec, unsigned IdxVal,
4396 SelectionDAG &DAG, SDLoc dl) {
4397 assert((Vec.getValueType().is256BitVector() ||
4398 Vec.getValueType().is512BitVector()) && "Unexpected vector size!");
4399 return ExtractSubVector(Vec, IdxVal, DAG, dl, 128);
4402 /// Generate a DAG to grab 256-bits from a 512-bit vector.
4403 static SDValue Extract256BitVector(SDValue Vec, unsigned IdxVal,
4404 SelectionDAG &DAG, SDLoc dl) {
4405 assert(Vec.getValueType().is512BitVector() && "Unexpected vector size!");
4406 return ExtractSubVector(Vec, IdxVal, DAG, dl, 256);
4409 static SDValue InsertSubVector(SDValue Result, SDValue Vec,
4410 unsigned IdxVal, SelectionDAG &DAG,
4411 SDLoc dl, unsigned vectorWidth) {
4412 assert((vectorWidth == 128 || vectorWidth == 256) &&
4413 "Unsupported vector width");
4414 // Inserting UNDEF is Result
4415 if (Vec.getOpcode() == ISD::UNDEF)
4417 EVT VT = Vec.getValueType();
4418 EVT ElVT = VT.getVectorElementType();
4419 EVT ResultVT = Result.getValueType();
4421 // Insert the relevant vectorWidth bits.
4422 unsigned ElemsPerChunk = vectorWidth/ElVT.getSizeInBits();
4424 // This is the index of the first element of the vectorWidth-bit chunk
4426 unsigned NormalizedIdxVal = (((IdxVal * ElVT.getSizeInBits())/vectorWidth)
4429 SDValue VecIdx = DAG.getIntPtrConstant(NormalizedIdxVal, dl);
4430 return DAG.getNode(ISD::INSERT_SUBVECTOR, dl, ResultVT, Result, Vec, VecIdx);
4433 /// Generate a DAG to put 128-bits into a vector > 128 bits. This
4434 /// sets things up to match to an AVX VINSERTF128/VINSERTI128 or
4435 /// AVX-512 VINSERTF32x4/VINSERTI32x4 instructions or a
4436 /// simple superregister reference. Idx is an index in the 128 bits
4437 /// we want. It need not be aligned to a 128-bit boundary. That makes
4438 /// lowering INSERT_VECTOR_ELT operations easier.
4439 static SDValue Insert128BitVector(SDValue Result, SDValue Vec, unsigned IdxVal,
4440 SelectionDAG &DAG, SDLoc dl) {
4441 assert(Vec.getValueType().is128BitVector() && "Unexpected vector size!");
4443 // For insertion into the zero index (low half) of a 256-bit vector, it is
4444 // more efficient to generate a blend with immediate instead of an insert*128.
4445 // We are still creating an INSERT_SUBVECTOR below with an undef node to
4446 // extend the subvector to the size of the result vector. Make sure that
4447 // we are not recursing on that node by checking for undef here.
4448 if (IdxVal == 0 && Result.getValueType().is256BitVector() &&
4449 Result.getOpcode() != ISD::UNDEF) {
4450 EVT ResultVT = Result.getValueType();
4451 SDValue ZeroIndex = DAG.getIntPtrConstant(0, dl);
4452 SDValue Undef = DAG.getUNDEF(ResultVT);
4453 SDValue Vec256 = DAG.getNode(ISD::INSERT_SUBVECTOR, dl, ResultVT, Undef,
4456 // The blend instruction, and therefore its mask, depend on the data type.
4457 MVT ScalarType = ResultVT.getScalarType().getSimpleVT();
4458 if (ScalarType.isFloatingPoint()) {
4459 // Choose either vblendps (float) or vblendpd (double).
4460 unsigned ScalarSize = ScalarType.getSizeInBits();
4461 assert((ScalarSize == 64 || ScalarSize == 32) && "Unknown float type");
4462 unsigned MaskVal = (ScalarSize == 64) ? 0x03 : 0x0f;
4463 SDValue Mask = DAG.getConstant(MaskVal, dl, MVT::i8);
4464 return DAG.getNode(X86ISD::BLENDI, dl, ResultVT, Result, Vec256, Mask);
4467 const X86Subtarget &Subtarget =
4468 static_cast<const X86Subtarget &>(DAG.getSubtarget());
4470 // AVX2 is needed for 256-bit integer blend support.
4471 // Integers must be cast to 32-bit because there is only vpblendd;
4472 // vpblendw can't be used for this because it has a handicapped mask.
4474 // If we don't have AVX2, then cast to float. Using a wrong domain blend
4475 // is still more efficient than using the wrong domain vinsertf128 that
4476 // will be created by InsertSubVector().
4477 MVT CastVT = Subtarget.hasAVX2() ? MVT::v8i32 : MVT::v8f32;
4479 SDValue Mask = DAG.getConstant(0x0f, dl, MVT::i8);
4480 Vec256 = DAG.getBitcast(CastVT, Vec256);
4481 Vec256 = DAG.getNode(X86ISD::BLENDI, dl, CastVT, Result, Vec256, Mask);
4482 return DAG.getBitcast(ResultVT, Vec256);
4485 return InsertSubVector(Result, Vec, IdxVal, DAG, dl, 128);
4488 static SDValue Insert256BitVector(SDValue Result, SDValue Vec, unsigned IdxVal,
4489 SelectionDAG &DAG, SDLoc dl) {
4490 assert(Vec.getValueType().is256BitVector() && "Unexpected vector size!");
4491 return InsertSubVector(Result, Vec, IdxVal, DAG, dl, 256);
4494 /// Concat two 128-bit vectors into a 256 bit vector using VINSERTF128
4495 /// instructions. This is used because creating CONCAT_VECTOR nodes of
4496 /// BUILD_VECTORS returns a larger BUILD_VECTOR while we're trying to lower
4497 /// large BUILD_VECTORS.
4498 static SDValue Concat128BitVectors(SDValue V1, SDValue V2, EVT VT,
4499 unsigned NumElems, SelectionDAG &DAG,
4501 SDValue V = Insert128BitVector(DAG.getUNDEF(VT), V1, 0, DAG, dl);
4502 return Insert128BitVector(V, V2, NumElems/2, DAG, dl);
4505 static SDValue Concat256BitVectors(SDValue V1, SDValue V2, EVT VT,
4506 unsigned NumElems, SelectionDAG &DAG,
4508 SDValue V = Insert256BitVector(DAG.getUNDEF(VT), V1, 0, DAG, dl);
4509 return Insert256BitVector(V, V2, NumElems/2, DAG, dl);
4512 /// Returns a vector of specified type with all bits set.
4513 /// Always build ones vectors as <4 x i32> or <8 x i32>. For 256-bit types with
4514 /// no AVX2 supprt, use two <4 x i32> inserted in a <8 x i32> appropriately.
4515 /// Then bitcast to their original type, ensuring they get CSE'd.
4516 static SDValue getOnesVector(EVT VT, const X86Subtarget *Subtarget,
4517 SelectionDAG &DAG, SDLoc dl) {
4518 assert(VT.isVector() && "Expected a vector type");
4520 SDValue Cst = DAG.getConstant(~0U, dl, MVT::i32);
4522 if (VT.is512BitVector()) {
4523 SDValue Ops[] = { Cst, Cst, Cst, Cst, Cst, Cst, Cst, Cst,
4524 Cst, Cst, Cst, Cst, Cst, Cst, Cst, Cst };
4525 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v16i32, Ops);
4526 } else if (VT.is256BitVector()) {
4527 if (Subtarget->hasInt256()) { // AVX2
4528 SDValue Ops[] = { Cst, Cst, Cst, Cst, Cst, Cst, Cst, Cst };
4529 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v8i32, Ops);
4531 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
4532 Vec = Concat128BitVectors(Vec, Vec, MVT::v8i32, 8, DAG, dl);
4534 } else if (VT.is128BitVector()) {
4535 Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
4537 llvm_unreachable("Unexpected vector type");
4539 return DAG.getBitcast(VT, Vec);
4542 /// Returns a vector_shuffle node for an unpackl operation.
4543 static SDValue getUnpackl(SelectionDAG &DAG, SDLoc dl, MVT VT, SDValue V1,
4545 unsigned NumElems = VT.getVectorNumElements();
4546 SmallVector<int, 8> Mask;
4547 for (unsigned i = 0, e = NumElems/2; i != e; ++i) {
4549 Mask.push_back(i + NumElems);
4551 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
4554 /// Returns a vector_shuffle node for an unpackh operation.
4555 static SDValue getUnpackh(SelectionDAG &DAG, SDLoc dl, MVT VT, SDValue V1,
4557 unsigned NumElems = VT.getVectorNumElements();
4558 SmallVector<int, 8> Mask;
4559 for (unsigned i = 0, Half = NumElems/2; i != Half; ++i) {
4560 Mask.push_back(i + Half);
4561 Mask.push_back(i + NumElems + Half);
4563 return DAG.getVectorShuffle(VT, dl, V1, V2, &Mask[0]);
4566 /// Return a vector_shuffle of the specified vector of zero or undef vector.
4567 /// This produces a shuffle where the low element of V2 is swizzled into the
4568 /// zero/undef vector, landing at element Idx.
4569 /// This produces a shuffle mask like 4,1,2,3 (idx=0) or 0,1,2,4 (idx=3).
4570 static SDValue getShuffleVectorZeroOrUndef(SDValue V2, unsigned Idx,
4572 const X86Subtarget *Subtarget,
4573 SelectionDAG &DAG) {
4574 MVT VT = V2.getSimpleValueType();
4576 ? getZeroVector(VT, Subtarget, DAG, SDLoc(V2)) : DAG.getUNDEF(VT);
4577 unsigned NumElems = VT.getVectorNumElements();
4578 SmallVector<int, 16> MaskVec;
4579 for (unsigned i = 0; i != NumElems; ++i)
4580 // If this is the insertion idx, put the low elt of V2 here.
4581 MaskVec.push_back(i == Idx ? NumElems : i);
4582 return DAG.getVectorShuffle(VT, SDLoc(V2), V1, V2, &MaskVec[0]);
4585 /// Calculates the shuffle mask corresponding to the target-specific opcode.
4586 /// Returns true if the Mask could be calculated. Sets IsUnary to true if only
4587 /// uses one source. Note that this will set IsUnary for shuffles which use a
4588 /// single input multiple times, and in those cases it will
4589 /// adjust the mask to only have indices within that single input.
4590 /// FIXME: Add support for Decode*Mask functions that return SM_SentinelZero.
4591 static bool getTargetShuffleMask(SDNode *N, MVT VT,
4592 SmallVectorImpl<int> &Mask, bool &IsUnary) {
4593 unsigned NumElems = VT.getVectorNumElements();
4597 bool IsFakeUnary = false;
4598 switch(N->getOpcode()) {
4599 case X86ISD::BLENDI:
4600 ImmN = N->getOperand(N->getNumOperands()-1);
4601 DecodeBLENDMask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
4604 ImmN = N->getOperand(N->getNumOperands()-1);
4605 DecodeSHUFPMask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
4606 IsUnary = IsFakeUnary = N->getOperand(0) == N->getOperand(1);
4608 case X86ISD::UNPCKH:
4609 DecodeUNPCKHMask(VT, Mask);
4610 IsUnary = IsFakeUnary = N->getOperand(0) == N->getOperand(1);
4612 case X86ISD::UNPCKL:
4613 DecodeUNPCKLMask(VT, Mask);
4614 IsUnary = IsFakeUnary = N->getOperand(0) == N->getOperand(1);
4616 case X86ISD::MOVHLPS:
4617 DecodeMOVHLPSMask(NumElems, Mask);
4618 IsUnary = IsFakeUnary = N->getOperand(0) == N->getOperand(1);
4620 case X86ISD::MOVLHPS:
4621 DecodeMOVLHPSMask(NumElems, Mask);
4622 IsUnary = IsFakeUnary = N->getOperand(0) == N->getOperand(1);
4624 case X86ISD::PALIGNR:
4625 ImmN = N->getOperand(N->getNumOperands()-1);
4626 DecodePALIGNRMask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
4628 case X86ISD::PSHUFD:
4629 case X86ISD::VPERMILPI:
4630 ImmN = N->getOperand(N->getNumOperands()-1);
4631 DecodePSHUFMask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
4634 case X86ISD::PSHUFHW:
4635 ImmN = N->getOperand(N->getNumOperands()-1);
4636 DecodePSHUFHWMask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
4639 case X86ISD::PSHUFLW:
4640 ImmN = N->getOperand(N->getNumOperands()-1);
4641 DecodePSHUFLWMask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
4644 case X86ISD::PSHUFB: {
4646 SDValue MaskNode = N->getOperand(1);
4647 while (MaskNode->getOpcode() == ISD::BITCAST)
4648 MaskNode = MaskNode->getOperand(0);
4650 if (MaskNode->getOpcode() == ISD::BUILD_VECTOR) {
4651 // If we have a build-vector, then things are easy.
4652 EVT VT = MaskNode.getValueType();
4653 assert(VT.isVector() &&
4654 "Can't produce a non-vector with a build_vector!");
4655 if (!VT.isInteger())
4658 int NumBytesPerElement = VT.getVectorElementType().getSizeInBits() / 8;
4660 SmallVector<uint64_t, 32> RawMask;
4661 for (int i = 0, e = MaskNode->getNumOperands(); i < e; ++i) {
4662 SDValue Op = MaskNode->getOperand(i);
4663 if (Op->getOpcode() == ISD::UNDEF) {
4664 RawMask.push_back((uint64_t)SM_SentinelUndef);
4667 auto *CN = dyn_cast<ConstantSDNode>(Op.getNode());
4670 APInt MaskElement = CN->getAPIntValue();
4672 // We now have to decode the element which could be any integer size and
4673 // extract each byte of it.
4674 for (int j = 0; j < NumBytesPerElement; ++j) {
4675 // Note that this is x86 and so always little endian: the low byte is
4676 // the first byte of the mask.
4677 RawMask.push_back(MaskElement.getLoBits(8).getZExtValue());
4678 MaskElement = MaskElement.lshr(8);
4681 DecodePSHUFBMask(RawMask, Mask);
4685 auto *MaskLoad = dyn_cast<LoadSDNode>(MaskNode);
4689 SDValue Ptr = MaskLoad->getBasePtr();
4690 if (Ptr->getOpcode() == X86ISD::Wrapper ||
4691 Ptr->getOpcode() == X86ISD::WrapperRIP)
4692 Ptr = Ptr->getOperand(0);
4694 auto *MaskCP = dyn_cast<ConstantPoolSDNode>(Ptr);
4695 if (!MaskCP || MaskCP->isMachineConstantPoolEntry())
4698 if (auto *C = dyn_cast<Constant>(MaskCP->getConstVal())) {
4699 DecodePSHUFBMask(C, Mask);
4707 case X86ISD::VPERMI:
4708 ImmN = N->getOperand(N->getNumOperands()-1);
4709 DecodeVPERMMask(cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
4714 DecodeScalarMoveMask(VT, /* IsLoad */ false, Mask);
4716 case X86ISD::VPERM2X128:
4717 ImmN = N->getOperand(N->getNumOperands()-1);
4718 DecodeVPERM2X128Mask(VT, cast<ConstantSDNode>(ImmN)->getZExtValue(), Mask);
4719 if (Mask.empty()) return false;
4720 // Mask only contains negative index if an element is zero.
4721 if (std::any_of(Mask.begin(), Mask.end(),
4722 [](int M){ return M == SM_SentinelZero; }))
4725 case X86ISD::MOVSLDUP:
4726 DecodeMOVSLDUPMask(VT, Mask);
4729 case X86ISD::MOVSHDUP:
4730 DecodeMOVSHDUPMask(VT, Mask);
4733 case X86ISD::MOVDDUP:
4734 DecodeMOVDDUPMask(VT, Mask);
4737 case X86ISD::MOVLHPD:
4738 case X86ISD::MOVLPD:
4739 case X86ISD::MOVLPS:
4740 // Not yet implemented
4742 case X86ISD::VPERMV: {
4744 SDValue MaskNode = N->getOperand(0);
4745 while (MaskNode->getOpcode() == ISD::BITCAST)
4746 MaskNode = MaskNode->getOperand(0);
4748 unsigned MaskLoBits = Log2_64(VT.getVectorNumElements());
4749 SmallVector<uint64_t, 32> RawMask;
4750 if (MaskNode->getOpcode() == ISD::BUILD_VECTOR) {
4751 // If we have a build-vector, then things are easy.
4752 assert(MaskNode.getValueType().isInteger() &&
4753 MaskNode.getValueType().getVectorNumElements() ==
4754 VT.getVectorNumElements());
4756 for (unsigned i = 0; i < MaskNode->getNumOperands(); ++i) {
4757 SDValue Op = MaskNode->getOperand(i);
4758 if (Op->getOpcode() == ISD::UNDEF)
4759 RawMask.push_back((uint64_t)SM_SentinelUndef);
4760 else if (isa<ConstantSDNode>(Op)) {
4761 APInt MaskElement = cast<ConstantSDNode>(Op)->getAPIntValue();
4762 RawMask.push_back(MaskElement.getLoBits(MaskLoBits).getZExtValue());
4766 DecodeVPERMVMask(RawMask, Mask);
4769 if (MaskNode->getOpcode() == X86ISD::VBROADCAST) {
4770 unsigned NumEltsInMask = MaskNode->getNumOperands();
4771 MaskNode = MaskNode->getOperand(0);
4772 auto *CN = dyn_cast<ConstantSDNode>(MaskNode);
4774 APInt MaskEltValue = CN->getAPIntValue();
4775 for (unsigned i = 0; i < NumEltsInMask; ++i)
4776 RawMask.push_back(MaskEltValue.getLoBits(MaskLoBits).getZExtValue());
4777 DecodeVPERMVMask(RawMask, Mask);
4780 // It may be a scalar load
4783 auto *MaskLoad = dyn_cast<LoadSDNode>(MaskNode);
4787 SDValue Ptr = MaskLoad->getBasePtr();
4788 if (Ptr->getOpcode() == X86ISD::Wrapper ||
4789 Ptr->getOpcode() == X86ISD::WrapperRIP)
4790 Ptr = Ptr->getOperand(0);
4792 auto *MaskCP = dyn_cast<ConstantPoolSDNode>(Ptr);
4793 if (!MaskCP || MaskCP->isMachineConstantPoolEntry())
4796 auto *C = dyn_cast<Constant>(MaskCP->getConstVal());
4798 DecodeVPERMVMask(C, VT, Mask);
4805 case X86ISD::VPERMV3: {
4807 SDValue MaskNode = N->getOperand(1);
4808 while (MaskNode->getOpcode() == ISD::BITCAST)
4809 MaskNode = MaskNode->getOperand(1);
4811 if (MaskNode->getOpcode() == ISD::BUILD_VECTOR) {
4812 // If we have a build-vector, then things are easy.
4813 assert(MaskNode.getValueType().isInteger() &&
4814 MaskNode.getValueType().getVectorNumElements() ==
4815 VT.getVectorNumElements());
4817 SmallVector<uint64_t, 32> RawMask;
4818 unsigned MaskLoBits = Log2_64(VT.getVectorNumElements()*2);
4820 for (unsigned i = 0; i < MaskNode->getNumOperands(); ++i) {
4821 SDValue Op = MaskNode->getOperand(i);
4822 if (Op->getOpcode() == ISD::UNDEF)
4823 RawMask.push_back((uint64_t)SM_SentinelUndef);
4825 auto *CN = dyn_cast<ConstantSDNode>(Op.getNode());
4828 APInt MaskElement = CN->getAPIntValue();
4829 RawMask.push_back(MaskElement.getLoBits(MaskLoBits).getZExtValue());
4832 DecodeVPERMV3Mask(RawMask, Mask);
4836 auto *MaskLoad = dyn_cast<LoadSDNode>(MaskNode);
4840 SDValue Ptr = MaskLoad->getBasePtr();
4841 if (Ptr->getOpcode() == X86ISD::Wrapper ||
4842 Ptr->getOpcode() == X86ISD::WrapperRIP)
4843 Ptr = Ptr->getOperand(0);
4845 auto *MaskCP = dyn_cast<ConstantPoolSDNode>(Ptr);
4846 if (!MaskCP || MaskCP->isMachineConstantPoolEntry())
4849 auto *C = dyn_cast<Constant>(MaskCP->getConstVal());
4851 DecodeVPERMV3Mask(C, VT, Mask);
4858 default: llvm_unreachable("unknown target shuffle node");
4861 // If we have a fake unary shuffle, the shuffle mask is spread across two
4862 // inputs that are actually the same node. Re-map the mask to always point
4863 // into the first input.
4866 if (M >= (int)Mask.size())
4872 /// Returns the scalar element that will make up the ith
4873 /// element of the result of the vector shuffle.
4874 static SDValue getShuffleScalarElt(SDNode *N, unsigned Index, SelectionDAG &DAG,
4877 return SDValue(); // Limit search depth.
4879 SDValue V = SDValue(N, 0);
4880 EVT VT = V.getValueType();
4881 unsigned Opcode = V.getOpcode();
4883 // Recurse into ISD::VECTOR_SHUFFLE node to find scalars.
4884 if (const ShuffleVectorSDNode *SV = dyn_cast<ShuffleVectorSDNode>(N)) {
4885 int Elt = SV->getMaskElt(Index);
4888 return DAG.getUNDEF(VT.getVectorElementType());
4890 unsigned NumElems = VT.getVectorNumElements();
4891 SDValue NewV = (Elt < (int)NumElems) ? SV->getOperand(0)
4892 : SV->getOperand(1);
4893 return getShuffleScalarElt(NewV.getNode(), Elt % NumElems, DAG, Depth+1);
4896 // Recurse into target specific vector shuffles to find scalars.
4897 if (isTargetShuffle(Opcode)) {
4898 MVT ShufVT = V.getSimpleValueType();
4899 unsigned NumElems = ShufVT.getVectorNumElements();
4900 SmallVector<int, 16> ShuffleMask;
4903 if (!getTargetShuffleMask(N, ShufVT, ShuffleMask, IsUnary))
4906 int Elt = ShuffleMask[Index];
4908 return DAG.getUNDEF(ShufVT.getVectorElementType());
4910 SDValue NewV = (Elt < (int)NumElems) ? N->getOperand(0)
4912 return getShuffleScalarElt(NewV.getNode(), Elt % NumElems, DAG,
4916 // Actual nodes that may contain scalar elements
4917 if (Opcode == ISD::BITCAST) {
4918 V = V.getOperand(0);
4919 EVT SrcVT = V.getValueType();
4920 unsigned NumElems = VT.getVectorNumElements();
4922 if (!SrcVT.isVector() || SrcVT.getVectorNumElements() != NumElems)
4926 if (V.getOpcode() == ISD::SCALAR_TO_VECTOR)
4927 return (Index == 0) ? V.getOperand(0)
4928 : DAG.getUNDEF(VT.getVectorElementType());
4930 if (V.getOpcode() == ISD::BUILD_VECTOR)
4931 return V.getOperand(Index);
4936 /// Custom lower build_vector of v16i8.
4937 static SDValue LowerBuildVectorv16i8(SDValue Op, unsigned NonZeros,
4938 unsigned NumNonZero, unsigned NumZero,
4940 const X86Subtarget* Subtarget,
4941 const TargetLowering &TLI) {
4949 // SSE4.1 - use PINSRB to insert each byte directly.
4950 if (Subtarget->hasSSE41()) {
4951 for (unsigned i = 0; i < 16; ++i) {
4952 bool isNonZero = (NonZeros & (1 << i)) != 0;
4956 V = getZeroVector(MVT::v16i8, Subtarget, DAG, dl);
4958 V = DAG.getUNDEF(MVT::v16i8);
4961 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl,
4962 MVT::v16i8, V, Op.getOperand(i),
4963 DAG.getIntPtrConstant(i, dl));
4970 // Pre-SSE4.1 - merge byte pairs and insert with PINSRW.
4971 for (unsigned i = 0; i < 16; ++i) {
4972 bool ThisIsNonZero = (NonZeros & (1 << i)) != 0;
4973 if (ThisIsNonZero && First) {
4975 V = getZeroVector(MVT::v8i16, Subtarget, DAG, dl);
4977 V = DAG.getUNDEF(MVT::v8i16);
4982 SDValue ThisElt, LastElt;
4983 bool LastIsNonZero = (NonZeros & (1 << (i-1))) != 0;
4984 if (LastIsNonZero) {
4985 LastElt = DAG.getNode(ISD::ZERO_EXTEND, dl,
4986 MVT::i16, Op.getOperand(i-1));
4988 if (ThisIsNonZero) {
4989 ThisElt = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i16, Op.getOperand(i));
4990 ThisElt = DAG.getNode(ISD::SHL, dl, MVT::i16,
4991 ThisElt, DAG.getConstant(8, dl, MVT::i8));
4993 ThisElt = DAG.getNode(ISD::OR, dl, MVT::i16, ThisElt, LastElt);
4997 if (ThisElt.getNode())
4998 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, V, ThisElt,
4999 DAG.getIntPtrConstant(i/2, dl));
5003 return DAG.getBitcast(MVT::v16i8, V);
5006 /// Custom lower build_vector of v8i16.
5007 static SDValue LowerBuildVectorv8i16(SDValue Op, unsigned NonZeros,
5008 unsigned NumNonZero, unsigned NumZero,
5010 const X86Subtarget* Subtarget,
5011 const TargetLowering &TLI) {
5018 for (unsigned i = 0; i < 8; ++i) {
5019 bool isNonZero = (NonZeros & (1 << i)) != 0;
5023 V = getZeroVector(MVT::v8i16, Subtarget, DAG, dl);
5025 V = DAG.getUNDEF(MVT::v8i16);
5028 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl,
5029 MVT::v8i16, V, Op.getOperand(i),
5030 DAG.getIntPtrConstant(i, dl));
5037 /// Custom lower build_vector of v4i32 or v4f32.
5038 static SDValue LowerBuildVectorv4x32(SDValue Op, SelectionDAG &DAG,
5039 const X86Subtarget *Subtarget,
5040 const TargetLowering &TLI) {
5041 // Find all zeroable elements.
5042 std::bitset<4> Zeroable;
5043 for (int i=0; i < 4; ++i) {
5044 SDValue Elt = Op->getOperand(i);
5045 Zeroable[i] = (Elt.getOpcode() == ISD::UNDEF || X86::isZeroNode(Elt));
5047 assert(Zeroable.size() - Zeroable.count() > 1 &&
5048 "We expect at least two non-zero elements!");
5050 // We only know how to deal with build_vector nodes where elements are either
5051 // zeroable or extract_vector_elt with constant index.
5052 SDValue FirstNonZero;
5053 unsigned FirstNonZeroIdx;
5054 for (unsigned i=0; i < 4; ++i) {
5057 SDValue Elt = Op->getOperand(i);
5058 if (Elt.getOpcode() != ISD::EXTRACT_VECTOR_ELT ||
5059 !isa<ConstantSDNode>(Elt.getOperand(1)))
5061 // Make sure that this node is extracting from a 128-bit vector.
5062 MVT VT = Elt.getOperand(0).getSimpleValueType();
5063 if (!VT.is128BitVector())
5065 if (!FirstNonZero.getNode()) {
5067 FirstNonZeroIdx = i;
5071 assert(FirstNonZero.getNode() && "Unexpected build vector of all zeros!");
5072 SDValue V1 = FirstNonZero.getOperand(0);
5073 MVT VT = V1.getSimpleValueType();
5075 // See if this build_vector can be lowered as a blend with zero.
5077 unsigned EltMaskIdx, EltIdx;
5079 for (EltIdx = 0; EltIdx < 4; ++EltIdx) {
5080 if (Zeroable[EltIdx]) {
5081 // The zero vector will be on the right hand side.
5082 Mask[EltIdx] = EltIdx+4;
5086 Elt = Op->getOperand(EltIdx);
5087 // By construction, Elt is a EXTRACT_VECTOR_ELT with constant index.
5088 EltMaskIdx = cast<ConstantSDNode>(Elt.getOperand(1))->getZExtValue();
5089 if (Elt.getOperand(0) != V1 || EltMaskIdx != EltIdx)
5091 Mask[EltIdx] = EltIdx;
5095 // Let the shuffle legalizer deal with blend operations.
5096 SDValue VZero = getZeroVector(VT, Subtarget, DAG, SDLoc(Op));
5097 if (V1.getSimpleValueType() != VT)
5098 V1 = DAG.getNode(ISD::BITCAST, SDLoc(V1), VT, V1);
5099 return DAG.getVectorShuffle(VT, SDLoc(V1), V1, VZero, &Mask[0]);
5102 // See if we can lower this build_vector to a INSERTPS.
5103 if (!Subtarget->hasSSE41())
5106 SDValue V2 = Elt.getOperand(0);
5107 if (Elt == FirstNonZero && EltIdx == FirstNonZeroIdx)
5110 bool CanFold = true;
5111 for (unsigned i = EltIdx + 1; i < 4 && CanFold; ++i) {
5115 SDValue Current = Op->getOperand(i);
5116 SDValue SrcVector = Current->getOperand(0);
5119 CanFold = SrcVector == V1 &&
5120 cast<ConstantSDNode>(Current.getOperand(1))->getZExtValue() == i;
5126 assert(V1.getNode() && "Expected at least two non-zero elements!");
5127 if (V1.getSimpleValueType() != MVT::v4f32)
5128 V1 = DAG.getNode(ISD::BITCAST, SDLoc(V1), MVT::v4f32, V1);
5129 if (V2.getSimpleValueType() != MVT::v4f32)
5130 V2 = DAG.getNode(ISD::BITCAST, SDLoc(V2), MVT::v4f32, V2);
5132 // Ok, we can emit an INSERTPS instruction.
5133 unsigned ZMask = Zeroable.to_ulong();
5135 unsigned InsertPSMask = EltMaskIdx << 6 | EltIdx << 4 | ZMask;
5136 assert((InsertPSMask & ~0xFFu) == 0 && "Invalid mask!");
5138 SDValue Result = DAG.getNode(X86ISD::INSERTPS, DL, MVT::v4f32, V1, V2,
5139 DAG.getIntPtrConstant(InsertPSMask, DL));
5140 return DAG.getBitcast(VT, Result);
5143 /// Return a vector logical shift node.
5144 static SDValue getVShift(bool isLeft, EVT VT, SDValue SrcOp,
5145 unsigned NumBits, SelectionDAG &DAG,
5146 const TargetLowering &TLI, SDLoc dl) {
5147 assert(VT.is128BitVector() && "Unknown type for VShift");
5148 MVT ShVT = MVT::v2i64;
5149 unsigned Opc = isLeft ? X86ISD::VSHLDQ : X86ISD::VSRLDQ;
5150 SrcOp = DAG.getBitcast(ShVT, SrcOp);
5151 MVT ScalarShiftTy = TLI.getScalarShiftAmountTy(DAG.getDataLayout(), VT);
5152 assert(NumBits % 8 == 0 && "Only support byte sized shifts");
5153 SDValue ShiftVal = DAG.getConstant(NumBits/8, dl, ScalarShiftTy);
5154 return DAG.getBitcast(VT, DAG.getNode(Opc, dl, ShVT, SrcOp, ShiftVal));
5158 LowerAsSplatVectorLoad(SDValue SrcOp, MVT VT, SDLoc dl, SelectionDAG &DAG) {
5160 // Check if the scalar load can be widened into a vector load. And if
5161 // the address is "base + cst" see if the cst can be "absorbed" into
5162 // the shuffle mask.
5163 if (LoadSDNode *LD = dyn_cast<LoadSDNode>(SrcOp)) {
5164 SDValue Ptr = LD->getBasePtr();
5165 if (!ISD::isNormalLoad(LD) || LD->isVolatile())
5167 EVT PVT = LD->getValueType(0);
5168 if (PVT != MVT::i32 && PVT != MVT::f32)
5173 if (FrameIndexSDNode *FINode = dyn_cast<FrameIndexSDNode>(Ptr)) {
5174 FI = FINode->getIndex();
5176 } else if (DAG.isBaseWithConstantOffset(Ptr) &&
5177 isa<FrameIndexSDNode>(Ptr.getOperand(0))) {
5178 FI = cast<FrameIndexSDNode>(Ptr.getOperand(0))->getIndex();
5179 Offset = Ptr.getConstantOperandVal(1);
5180 Ptr = Ptr.getOperand(0);
5185 // FIXME: 256-bit vector instructions don't require a strict alignment,
5186 // improve this code to support it better.
5187 unsigned RequiredAlign = VT.getSizeInBits()/8;
5188 SDValue Chain = LD->getChain();
5189 // Make sure the stack object alignment is at least 16 or 32.
5190 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
5191 if (DAG.InferPtrAlignment(Ptr) < RequiredAlign) {
5192 if (MFI->isFixedObjectIndex(FI)) {
5193 // Can't change the alignment. FIXME: It's possible to compute
5194 // the exact stack offset and reference FI + adjust offset instead.
5195 // If someone *really* cares about this. That's the way to implement it.
5198 MFI->setObjectAlignment(FI, RequiredAlign);
5202 // (Offset % 16 or 32) must be multiple of 4. Then address is then
5203 // Ptr + (Offset & ~15).
5206 if ((Offset % RequiredAlign) & 3)
5208 int64_t StartOffset = Offset & ~int64_t(RequiredAlign - 1);
5211 Ptr = DAG.getNode(ISD::ADD, DL, Ptr.getValueType(), Ptr,
5212 DAG.getConstant(StartOffset, DL, Ptr.getValueType()));
5215 int EltNo = (Offset - StartOffset) >> 2;
5216 unsigned NumElems = VT.getVectorNumElements();
5218 EVT NVT = EVT::getVectorVT(*DAG.getContext(), PVT, NumElems);
5219 SDValue V1 = DAG.getLoad(NVT, dl, Chain, Ptr,
5220 LD->getPointerInfo().getWithOffset(StartOffset),
5221 false, false, false, 0);
5223 SmallVector<int, 8> Mask(NumElems, EltNo);
5225 return DAG.getVectorShuffle(NVT, dl, V1, DAG.getUNDEF(NVT), &Mask[0]);
5231 /// Given the initializing elements 'Elts' of a vector of type 'VT', see if the
5232 /// elements can be replaced by a single large load which has the same value as
5233 /// a build_vector or insert_subvector whose loaded operands are 'Elts'.
5235 /// Example: <load i32 *a, load i32 *a+4, undef, undef> -> zextload a
5237 /// FIXME: we'd also like to handle the case where the last elements are zero
5238 /// rather than undef via VZEXT_LOAD, but we do not detect that case today.
5239 /// There's even a handy isZeroNode for that purpose.
5240 static SDValue EltsFromConsecutiveLoads(EVT VT, ArrayRef<SDValue> Elts,
5241 SDLoc &DL, SelectionDAG &DAG,
5242 bool isAfterLegalize) {
5243 unsigned NumElems = Elts.size();
5245 LoadSDNode *LDBase = nullptr;
5246 unsigned LastLoadedElt = -1U;
5248 // For each element in the initializer, see if we've found a load or an undef.
5249 // If we don't find an initial load element, or later load elements are
5250 // non-consecutive, bail out.
5251 for (unsigned i = 0; i < NumElems; ++i) {
5252 SDValue Elt = Elts[i];
5253 // Look through a bitcast.
5254 if (Elt.getNode() && Elt.getOpcode() == ISD::BITCAST)
5255 Elt = Elt.getOperand(0);
5256 if (!Elt.getNode() ||
5257 (Elt.getOpcode() != ISD::UNDEF && !ISD::isNON_EXTLoad(Elt.getNode())))
5260 if (Elt.getNode()->getOpcode() == ISD::UNDEF)
5262 LDBase = cast<LoadSDNode>(Elt.getNode());
5266 if (Elt.getOpcode() == ISD::UNDEF)
5269 LoadSDNode *LD = cast<LoadSDNode>(Elt);
5270 EVT LdVT = Elt.getValueType();
5271 // Each loaded element must be the correct fractional portion of the
5272 // requested vector load.
5273 if (LdVT.getSizeInBits() != VT.getSizeInBits() / NumElems)
5275 if (!DAG.isConsecutiveLoad(LD, LDBase, LdVT.getSizeInBits() / 8, i))
5280 // If we have found an entire vector of loads and undefs, then return a large
5281 // load of the entire vector width starting at the base pointer. If we found
5282 // consecutive loads for the low half, generate a vzext_load node.
5283 if (LastLoadedElt == NumElems - 1) {
5284 assert(LDBase && "Did not find base load for merging consecutive loads");
5285 EVT EltVT = LDBase->getValueType(0);
5286 // Ensure that the input vector size for the merged loads matches the
5287 // cumulative size of the input elements.
5288 if (VT.getSizeInBits() != EltVT.getSizeInBits() * NumElems)
5291 if (isAfterLegalize &&
5292 !DAG.getTargetLoweringInfo().isOperationLegal(ISD::LOAD, VT))
5295 SDValue NewLd = SDValue();
5297 NewLd = DAG.getLoad(VT, DL, LDBase->getChain(), LDBase->getBasePtr(),
5298 LDBase->getPointerInfo(), LDBase->isVolatile(),
5299 LDBase->isNonTemporal(), LDBase->isInvariant(),
5300 LDBase->getAlignment());
5302 if (LDBase->hasAnyUseOfValue(1)) {
5303 SDValue NewChain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other,
5305 SDValue(NewLd.getNode(), 1));
5306 DAG.ReplaceAllUsesOfValueWith(SDValue(LDBase, 1), NewChain);
5307 DAG.UpdateNodeOperands(NewChain.getNode(), SDValue(LDBase, 1),
5308 SDValue(NewLd.getNode(), 1));
5314 //TODO: The code below fires only for for loading the low v2i32 / v2f32
5315 //of a v4i32 / v4f32. It's probably worth generalizing.
5316 EVT EltVT = VT.getVectorElementType();
5317 if (NumElems == 4 && LastLoadedElt == 1 && (EltVT.getSizeInBits() == 32) &&
5318 DAG.getTargetLoweringInfo().isTypeLegal(MVT::v2i64)) {
5319 SDVTList Tys = DAG.getVTList(MVT::v2i64, MVT::Other);
5320 SDValue Ops[] = { LDBase->getChain(), LDBase->getBasePtr() };
5322 DAG.getMemIntrinsicNode(X86ISD::VZEXT_LOAD, DL, Tys, Ops, MVT::i64,
5323 LDBase->getPointerInfo(),
5324 LDBase->getAlignment(),
5325 false/*isVolatile*/, true/*ReadMem*/,
5328 // Make sure the newly-created LOAD is in the same position as LDBase in
5329 // terms of dependency. We create a TokenFactor for LDBase and ResNode, and
5330 // update uses of LDBase's output chain to use the TokenFactor.
5331 if (LDBase->hasAnyUseOfValue(1)) {
5332 SDValue NewChain = DAG.getNode(ISD::TokenFactor, DL, MVT::Other,
5333 SDValue(LDBase, 1), SDValue(ResNode.getNode(), 1));
5334 DAG.ReplaceAllUsesOfValueWith(SDValue(LDBase, 1), NewChain);
5335 DAG.UpdateNodeOperands(NewChain.getNode(), SDValue(LDBase, 1),
5336 SDValue(ResNode.getNode(), 1));
5339 return DAG.getBitcast(VT, ResNode);
5344 /// LowerVectorBroadcast - Attempt to use the vbroadcast instruction
5345 /// to generate a splat value for the following cases:
5346 /// 1. A splat BUILD_VECTOR which uses a single scalar load, or a constant.
5347 /// 2. A splat shuffle which uses a scalar_to_vector node which comes from
5348 /// a scalar load, or a constant.
5349 /// The VBROADCAST node is returned when a pattern is found,
5350 /// or SDValue() otherwise.
5351 static SDValue LowerVectorBroadcast(SDValue Op, const X86Subtarget* Subtarget,
5352 SelectionDAG &DAG) {
5353 // VBROADCAST requires AVX.
5354 // TODO: Splats could be generated for non-AVX CPUs using SSE
5355 // instructions, but there's less potential gain for only 128-bit vectors.
5356 if (!Subtarget->hasAVX())
5359 MVT VT = Op.getSimpleValueType();
5362 assert((VT.is128BitVector() || VT.is256BitVector() || VT.is512BitVector()) &&
5363 "Unsupported vector type for broadcast.");
5368 switch (Op.getOpcode()) {
5370 // Unknown pattern found.
5373 case ISD::BUILD_VECTOR: {
5374 auto *BVOp = cast<BuildVectorSDNode>(Op.getNode());
5375 BitVector UndefElements;
5376 SDValue Splat = BVOp->getSplatValue(&UndefElements);
5378 // We need a splat of a single value to use broadcast, and it doesn't
5379 // make any sense if the value is only in one element of the vector.
5380 if (!Splat || (VT.getVectorNumElements() - UndefElements.count()) <= 1)
5384 ConstSplatVal = (Ld.getOpcode() == ISD::Constant ||
5385 Ld.getOpcode() == ISD::ConstantFP);
5387 // Make sure that all of the users of a non-constant load are from the
5388 // BUILD_VECTOR node.
5389 if (!ConstSplatVal && !BVOp->isOnlyUserOf(Ld.getNode()))
5394 case ISD::VECTOR_SHUFFLE: {
5395 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
5397 // Shuffles must have a splat mask where the first element is
5399 if ((!SVOp->isSplat()) || SVOp->getMaskElt(0) != 0)
5402 SDValue Sc = Op.getOperand(0);
5403 if (Sc.getOpcode() != ISD::SCALAR_TO_VECTOR &&
5404 Sc.getOpcode() != ISD::BUILD_VECTOR) {
5406 if (!Subtarget->hasInt256())
5409 // Use the register form of the broadcast instruction available on AVX2.
5410 if (VT.getSizeInBits() >= 256)
5411 Sc = Extract128BitVector(Sc, 0, DAG, dl);
5412 return DAG.getNode(X86ISD::VBROADCAST, dl, VT, Sc);
5415 Ld = Sc.getOperand(0);
5416 ConstSplatVal = (Ld.getOpcode() == ISD::Constant ||
5417 Ld.getOpcode() == ISD::ConstantFP);
5419 // The scalar_to_vector node and the suspected
5420 // load node must have exactly one user.
5421 // Constants may have multiple users.
5423 // AVX-512 has register version of the broadcast
5424 bool hasRegVer = Subtarget->hasAVX512() && VT.is512BitVector() &&
5425 Ld.getValueType().getSizeInBits() >= 32;
5426 if (!ConstSplatVal && ((!Sc.hasOneUse() || !Ld.hasOneUse()) &&
5433 unsigned ScalarSize = Ld.getValueType().getSizeInBits();
5434 bool IsGE256 = (VT.getSizeInBits() >= 256);
5436 // When optimizing for size, generate up to 5 extra bytes for a broadcast
5437 // instruction to save 8 or more bytes of constant pool data.
5438 // TODO: If multiple splats are generated to load the same constant,
5439 // it may be detrimental to overall size. There needs to be a way to detect
5440 // that condition to know if this is truly a size win.
5441 bool OptForSize = DAG.getMachineFunction().getFunction()->optForSize();
5443 // Handle broadcasting a single constant scalar from the constant pool
5445 // On Sandybridge (no AVX2), it is still better to load a constant vector
5446 // from the constant pool and not to broadcast it from a scalar.
5447 // But override that restriction when optimizing for size.
5448 // TODO: Check if splatting is recommended for other AVX-capable CPUs.
5449 if (ConstSplatVal && (Subtarget->hasAVX2() || OptForSize)) {
5450 EVT CVT = Ld.getValueType();
5451 assert(!CVT.isVector() && "Must not broadcast a vector type");
5453 // Splat f32, i32, v4f64, v4i64 in all cases with AVX2.
5454 // For size optimization, also splat v2f64 and v2i64, and for size opt
5455 // with AVX2, also splat i8 and i16.
5456 // With pattern matching, the VBROADCAST node may become a VMOVDDUP.
5457 if (ScalarSize == 32 || (IsGE256 && ScalarSize == 64) ||
5458 (OptForSize && (ScalarSize == 64 || Subtarget->hasAVX2()))) {
5459 const Constant *C = nullptr;
5460 if (ConstantSDNode *CI = dyn_cast<ConstantSDNode>(Ld))
5461 C = CI->getConstantIntValue();
5462 else if (ConstantFPSDNode *CF = dyn_cast<ConstantFPSDNode>(Ld))
5463 C = CF->getConstantFPValue();
5465 assert(C && "Invalid constant type");
5467 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
5469 DAG.getConstantPool(C, TLI.getPointerTy(DAG.getDataLayout()));
5470 unsigned Alignment = cast<ConstantPoolSDNode>(CP)->getAlignment();
5472 CVT, dl, DAG.getEntryNode(), CP,
5473 MachinePointerInfo::getConstantPool(DAG.getMachineFunction()), false,
5474 false, false, Alignment);
5476 return DAG.getNode(X86ISD::VBROADCAST, dl, VT, Ld);
5480 bool IsLoad = ISD::isNormalLoad(Ld.getNode());
5482 // Handle AVX2 in-register broadcasts.
5483 if (!IsLoad && Subtarget->hasInt256() &&
5484 (ScalarSize == 32 || (IsGE256 && ScalarSize == 64)))
5485 return DAG.getNode(X86ISD::VBROADCAST, dl, VT, Ld);
5487 // The scalar source must be a normal load.
5491 if (ScalarSize == 32 || (IsGE256 && ScalarSize == 64) ||
5492 (Subtarget->hasVLX() && ScalarSize == 64))
5493 return DAG.getNode(X86ISD::VBROADCAST, dl, VT, Ld);
5495 // The integer check is needed for the 64-bit into 128-bit so it doesn't match
5496 // double since there is no vbroadcastsd xmm
5497 if (Subtarget->hasInt256() && Ld.getValueType().isInteger()) {
5498 if (ScalarSize == 8 || ScalarSize == 16 || ScalarSize == 64)
5499 return DAG.getNode(X86ISD::VBROADCAST, dl, VT, Ld);
5502 // Unsupported broadcast.
5506 /// \brief For an EXTRACT_VECTOR_ELT with a constant index return the real
5507 /// underlying vector and index.
5509 /// Modifies \p ExtractedFromVec to the real vector and returns the real
5511 static int getUnderlyingExtractedFromVec(SDValue &ExtractedFromVec,
5513 int Idx = cast<ConstantSDNode>(ExtIdx)->getZExtValue();
5514 if (!isa<ShuffleVectorSDNode>(ExtractedFromVec))
5517 // For 256-bit vectors, LowerEXTRACT_VECTOR_ELT_SSE4 may have already
5519 // (extract_vector_elt (v8f32 %vreg1), Constant<6>)
5521 // (extract_vector_elt (vector_shuffle<2,u,u,u>
5522 // (extract_subvector (v8f32 %vreg0), Constant<4>),
5525 // In this case the vector is the extract_subvector expression and the index
5526 // is 2, as specified by the shuffle.
5527 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(ExtractedFromVec);
5528 SDValue ShuffleVec = SVOp->getOperand(0);
5529 MVT ShuffleVecVT = ShuffleVec.getSimpleValueType();
5530 assert(ShuffleVecVT.getVectorElementType() ==
5531 ExtractedFromVec.getSimpleValueType().getVectorElementType());
5533 int ShuffleIdx = SVOp->getMaskElt(Idx);
5534 if (isUndefOrInRange(ShuffleIdx, 0, ShuffleVecVT.getVectorNumElements())) {
5535 ExtractedFromVec = ShuffleVec;
5541 static SDValue buildFromShuffleMostly(SDValue Op, SelectionDAG &DAG) {
5542 MVT VT = Op.getSimpleValueType();
5544 // Skip if insert_vec_elt is not supported.
5545 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
5546 if (!TLI.isOperationLegalOrCustom(ISD::INSERT_VECTOR_ELT, VT))
5550 unsigned NumElems = Op.getNumOperands();
5554 SmallVector<unsigned, 4> InsertIndices;
5555 SmallVector<int, 8> Mask(NumElems, -1);
5557 for (unsigned i = 0; i != NumElems; ++i) {
5558 unsigned Opc = Op.getOperand(i).getOpcode();
5560 if (Opc == ISD::UNDEF)
5563 if (Opc != ISD::EXTRACT_VECTOR_ELT) {
5564 // Quit if more than 1 elements need inserting.
5565 if (InsertIndices.size() > 1)
5568 InsertIndices.push_back(i);
5572 SDValue ExtractedFromVec = Op.getOperand(i).getOperand(0);
5573 SDValue ExtIdx = Op.getOperand(i).getOperand(1);
5574 // Quit if non-constant index.
5575 if (!isa<ConstantSDNode>(ExtIdx))
5577 int Idx = getUnderlyingExtractedFromVec(ExtractedFromVec, ExtIdx);
5579 // Quit if extracted from vector of different type.
5580 if (ExtractedFromVec.getValueType() != VT)
5583 if (!VecIn1.getNode())
5584 VecIn1 = ExtractedFromVec;
5585 else if (VecIn1 != ExtractedFromVec) {
5586 if (!VecIn2.getNode())
5587 VecIn2 = ExtractedFromVec;
5588 else if (VecIn2 != ExtractedFromVec)
5589 // Quit if more than 2 vectors to shuffle
5593 if (ExtractedFromVec == VecIn1)
5595 else if (ExtractedFromVec == VecIn2)
5596 Mask[i] = Idx + NumElems;
5599 if (!VecIn1.getNode())
5602 VecIn2 = VecIn2.getNode() ? VecIn2 : DAG.getUNDEF(VT);
5603 SDValue NV = DAG.getVectorShuffle(VT, DL, VecIn1, VecIn2, &Mask[0]);
5604 for (unsigned i = 0, e = InsertIndices.size(); i != e; ++i) {
5605 unsigned Idx = InsertIndices[i];
5606 NV = DAG.getNode(ISD::INSERT_VECTOR_ELT, DL, VT, NV, Op.getOperand(Idx),
5607 DAG.getIntPtrConstant(Idx, DL));
5613 static SDValue ConvertI1VectorToInteger(SDValue Op, SelectionDAG &DAG) {
5614 assert(ISD::isBuildVectorOfConstantSDNodes(Op.getNode()) &&
5615 Op.getScalarValueSizeInBits() == 1 &&
5616 "Can not convert non-constant vector");
5617 uint64_t Immediate = 0;
5618 for (unsigned idx = 0, e = Op.getNumOperands(); idx < e; ++idx) {
5619 SDValue In = Op.getOperand(idx);
5620 if (In.getOpcode() != ISD::UNDEF)
5621 Immediate |= cast<ConstantSDNode>(In)->getZExtValue() << idx;
5625 MVT::getIntegerVT(std::max((int)Op.getValueType().getSizeInBits(), 8));
5626 return DAG.getConstant(Immediate, dl, VT);
5628 // Lower BUILD_VECTOR operation for v8i1 and v16i1 types.
5630 X86TargetLowering::LowerBUILD_VECTORvXi1(SDValue Op, SelectionDAG &DAG) const {
5632 MVT VT = Op.getSimpleValueType();
5633 assert((VT.getVectorElementType() == MVT::i1) &&
5634 "Unexpected type in LowerBUILD_VECTORvXi1!");
5637 if (ISD::isBuildVectorAllZeros(Op.getNode())) {
5638 SDValue Cst = DAG.getTargetConstant(0, dl, MVT::i1);
5639 SmallVector<SDValue, 16> Ops(VT.getVectorNumElements(), Cst);
5640 return DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Ops);
5643 if (ISD::isBuildVectorAllOnes(Op.getNode())) {
5644 SDValue Cst = DAG.getTargetConstant(1, dl, MVT::i1);
5645 SmallVector<SDValue, 16> Ops(VT.getVectorNumElements(), Cst);
5646 return DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Ops);
5649 if (ISD::isBuildVectorOfConstantSDNodes(Op.getNode())) {
5650 SDValue Imm = ConvertI1VectorToInteger(Op, DAG);
5651 if (Imm.getValueSizeInBits() == VT.getSizeInBits())
5652 return DAG.getBitcast(VT, Imm);
5653 SDValue ExtVec = DAG.getBitcast(MVT::v8i1, Imm);
5654 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT, ExtVec,
5655 DAG.getIntPtrConstant(0, dl));
5658 // Vector has one or more non-const elements
5659 uint64_t Immediate = 0;
5660 SmallVector<unsigned, 16> NonConstIdx;
5661 bool IsSplat = true;
5662 bool HasConstElts = false;
5664 for (unsigned idx = 0, e = Op.getNumOperands(); idx < e; ++idx) {
5665 SDValue In = Op.getOperand(idx);
5666 if (In.getOpcode() == ISD::UNDEF)
5668 if (!isa<ConstantSDNode>(In))
5669 NonConstIdx.push_back(idx);
5671 Immediate |= cast<ConstantSDNode>(In)->getZExtValue() << idx;
5672 HasConstElts = true;
5676 else if (In != Op.getOperand(SplatIdx))
5680 // for splat use " (select i1 splat_elt, all-ones, all-zeroes)"
5682 return DAG.getNode(ISD::SELECT, dl, VT, Op.getOperand(SplatIdx),
5683 DAG.getConstant(1, dl, VT),
5684 DAG.getConstant(0, dl, VT));
5686 // insert elements one by one
5690 MVT ImmVT = MVT::getIntegerVT(std::max((int)VT.getSizeInBits(), 8));
5691 Imm = DAG.getConstant(Immediate, dl, ImmVT);
5693 else if (HasConstElts)
5694 Imm = DAG.getConstant(0, dl, VT);
5696 Imm = DAG.getUNDEF(VT);
5697 if (Imm.getValueSizeInBits() == VT.getSizeInBits())
5698 DstVec = DAG.getBitcast(VT, Imm);
5700 SDValue ExtVec = DAG.getBitcast(MVT::v8i1, Imm);
5701 DstVec = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT, ExtVec,
5702 DAG.getIntPtrConstant(0, dl));
5705 for (unsigned i = 0; i < NonConstIdx.size(); ++i) {
5706 unsigned InsertIdx = NonConstIdx[i];
5707 DstVec = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, DstVec,
5708 Op.getOperand(InsertIdx),
5709 DAG.getIntPtrConstant(InsertIdx, dl));
5714 /// \brief Return true if \p N implements a horizontal binop and return the
5715 /// operands for the horizontal binop into V0 and V1.
5717 /// This is a helper function of LowerToHorizontalOp().
5718 /// This function checks that the build_vector \p N in input implements a
5719 /// horizontal operation. Parameter \p Opcode defines the kind of horizontal
5720 /// operation to match.
5721 /// For example, if \p Opcode is equal to ISD::ADD, then this function
5722 /// checks if \p N implements a horizontal arithmetic add; if instead \p Opcode
5723 /// is equal to ISD::SUB, then this function checks if this is a horizontal
5726 /// This function only analyzes elements of \p N whose indices are
5727 /// in range [BaseIdx, LastIdx).
5728 static bool isHorizontalBinOp(const BuildVectorSDNode *N, unsigned Opcode,
5730 unsigned BaseIdx, unsigned LastIdx,
5731 SDValue &V0, SDValue &V1) {
5732 EVT VT = N->getValueType(0);
5734 assert(BaseIdx * 2 <= LastIdx && "Invalid Indices in input!");
5735 assert(VT.isVector() && VT.getVectorNumElements() >= LastIdx &&
5736 "Invalid Vector in input!");
5738 bool IsCommutable = (Opcode == ISD::ADD || Opcode == ISD::FADD);
5739 bool CanFold = true;
5740 unsigned ExpectedVExtractIdx = BaseIdx;
5741 unsigned NumElts = LastIdx - BaseIdx;
5742 V0 = DAG.getUNDEF(VT);
5743 V1 = DAG.getUNDEF(VT);
5745 // Check if N implements a horizontal binop.
5746 for (unsigned i = 0, e = NumElts; i != e && CanFold; ++i) {
5747 SDValue Op = N->getOperand(i + BaseIdx);
5750 if (Op->getOpcode() == ISD::UNDEF) {
5751 // Update the expected vector extract index.
5752 if (i * 2 == NumElts)
5753 ExpectedVExtractIdx = BaseIdx;
5754 ExpectedVExtractIdx += 2;
5758 CanFold = Op->getOpcode() == Opcode && Op->hasOneUse();
5763 SDValue Op0 = Op.getOperand(0);
5764 SDValue Op1 = Op.getOperand(1);
5766 // Try to match the following pattern:
5767 // (BINOP (extract_vector_elt A, I), (extract_vector_elt A, I+1))
5768 CanFold = (Op0.getOpcode() == ISD::EXTRACT_VECTOR_ELT &&
5769 Op1.getOpcode() == ISD::EXTRACT_VECTOR_ELT &&
5770 Op0.getOperand(0) == Op1.getOperand(0) &&
5771 isa<ConstantSDNode>(Op0.getOperand(1)) &&
5772 isa<ConstantSDNode>(Op1.getOperand(1)));
5776 unsigned I0 = cast<ConstantSDNode>(Op0.getOperand(1))->getZExtValue();
5777 unsigned I1 = cast<ConstantSDNode>(Op1.getOperand(1))->getZExtValue();
5779 if (i * 2 < NumElts) {
5780 if (V0.getOpcode() == ISD::UNDEF) {
5781 V0 = Op0.getOperand(0);
5782 if (V0.getValueType() != VT)
5786 if (V1.getOpcode() == ISD::UNDEF) {
5787 V1 = Op0.getOperand(0);
5788 if (V1.getValueType() != VT)
5791 if (i * 2 == NumElts)
5792 ExpectedVExtractIdx = BaseIdx;
5795 SDValue Expected = (i * 2 < NumElts) ? V0 : V1;
5796 if (I0 == ExpectedVExtractIdx)
5797 CanFold = I1 == I0 + 1 && Op0.getOperand(0) == Expected;
5798 else if (IsCommutable && I1 == ExpectedVExtractIdx) {
5799 // Try to match the following dag sequence:
5800 // (BINOP (extract_vector_elt A, I+1), (extract_vector_elt A, I))
5801 CanFold = I0 == I1 + 1 && Op1.getOperand(0) == Expected;
5805 ExpectedVExtractIdx += 2;
5811 /// \brief Emit a sequence of two 128-bit horizontal add/sub followed by
5812 /// a concat_vector.
5814 /// This is a helper function of LowerToHorizontalOp().
5815 /// This function expects two 256-bit vectors called V0 and V1.
5816 /// At first, each vector is split into two separate 128-bit vectors.
5817 /// Then, the resulting 128-bit vectors are used to implement two
5818 /// horizontal binary operations.
5820 /// The kind of horizontal binary operation is defined by \p X86Opcode.
5822 /// \p Mode specifies how the 128-bit parts of V0 and V1 are passed in input to
5823 /// the two new horizontal binop.
5824 /// When Mode is set, the first horizontal binop dag node would take as input
5825 /// the lower 128-bit of V0 and the upper 128-bit of V0. The second
5826 /// horizontal binop dag node would take as input the lower 128-bit of V1
5827 /// and the upper 128-bit of V1.
5829 /// HADD V0_LO, V0_HI
5830 /// HADD V1_LO, V1_HI
5832 /// Otherwise, the first horizontal binop dag node takes as input the lower
5833 /// 128-bit of V0 and the lower 128-bit of V1, and the second horizontal binop
5834 /// dag node takes the upper 128-bit of V0 and the upper 128-bit of V1.
5836 /// HADD V0_LO, V1_LO
5837 /// HADD V0_HI, V1_HI
5839 /// If \p isUndefLO is set, then the algorithm propagates UNDEF to the lower
5840 /// 128-bits of the result. If \p isUndefHI is set, then UNDEF is propagated to
5841 /// the upper 128-bits of the result.
5842 static SDValue ExpandHorizontalBinOp(const SDValue &V0, const SDValue &V1,
5843 SDLoc DL, SelectionDAG &DAG,
5844 unsigned X86Opcode, bool Mode,
5845 bool isUndefLO, bool isUndefHI) {
5846 EVT VT = V0.getValueType();
5847 assert(VT.is256BitVector() && VT == V1.getValueType() &&
5848 "Invalid nodes in input!");
5850 unsigned NumElts = VT.getVectorNumElements();
5851 SDValue V0_LO = Extract128BitVector(V0, 0, DAG, DL);
5852 SDValue V0_HI = Extract128BitVector(V0, NumElts/2, DAG, DL);
5853 SDValue V1_LO = Extract128BitVector(V1, 0, DAG, DL);
5854 SDValue V1_HI = Extract128BitVector(V1, NumElts/2, DAG, DL);
5855 EVT NewVT = V0_LO.getValueType();
5857 SDValue LO = DAG.getUNDEF(NewVT);
5858 SDValue HI = DAG.getUNDEF(NewVT);
5861 // Don't emit a horizontal binop if the result is expected to be UNDEF.
5862 if (!isUndefLO && V0->getOpcode() != ISD::UNDEF)
5863 LO = DAG.getNode(X86Opcode, DL, NewVT, V0_LO, V0_HI);
5864 if (!isUndefHI && V1->getOpcode() != ISD::UNDEF)
5865 HI = DAG.getNode(X86Opcode, DL, NewVT, V1_LO, V1_HI);
5867 // Don't emit a horizontal binop if the result is expected to be UNDEF.
5868 if (!isUndefLO && (V0_LO->getOpcode() != ISD::UNDEF ||
5869 V1_LO->getOpcode() != ISD::UNDEF))
5870 LO = DAG.getNode(X86Opcode, DL, NewVT, V0_LO, V1_LO);
5872 if (!isUndefHI && (V0_HI->getOpcode() != ISD::UNDEF ||
5873 V1_HI->getOpcode() != ISD::UNDEF))
5874 HI = DAG.getNode(X86Opcode, DL, NewVT, V0_HI, V1_HI);
5877 return DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, LO, HI);
5880 /// Try to fold a build_vector that performs an 'addsub' to an X86ISD::ADDSUB
5882 static SDValue LowerToAddSub(const BuildVectorSDNode *BV,
5883 const X86Subtarget *Subtarget, SelectionDAG &DAG) {
5884 EVT VT = BV->getValueType(0);
5885 if ((!Subtarget->hasSSE3() || (VT != MVT::v4f32 && VT != MVT::v2f64)) &&
5886 (!Subtarget->hasAVX() || (VT != MVT::v8f32 && VT != MVT::v4f64)))
5890 unsigned NumElts = VT.getVectorNumElements();
5891 SDValue InVec0 = DAG.getUNDEF(VT);
5892 SDValue InVec1 = DAG.getUNDEF(VT);
5894 assert((VT == MVT::v8f32 || VT == MVT::v4f64 || VT == MVT::v4f32 ||
5895 VT == MVT::v2f64) && "build_vector with an invalid type found!");
5897 // Odd-numbered elements in the input build vector are obtained from
5898 // adding two integer/float elements.
5899 // Even-numbered elements in the input build vector are obtained from
5900 // subtracting two integer/float elements.
5901 unsigned ExpectedOpcode = ISD::FSUB;
5902 unsigned NextExpectedOpcode = ISD::FADD;
5903 bool AddFound = false;
5904 bool SubFound = false;
5906 for (unsigned i = 0, e = NumElts; i != e; ++i) {
5907 SDValue Op = BV->getOperand(i);
5909 // Skip 'undef' values.
5910 unsigned Opcode = Op.getOpcode();
5911 if (Opcode == ISD::UNDEF) {
5912 std::swap(ExpectedOpcode, NextExpectedOpcode);
5916 // Early exit if we found an unexpected opcode.
5917 if (Opcode != ExpectedOpcode)
5920 SDValue Op0 = Op.getOperand(0);
5921 SDValue Op1 = Op.getOperand(1);
5923 // Try to match the following pattern:
5924 // (BINOP (extract_vector_elt A, i), (extract_vector_elt B, i))
5925 // Early exit if we cannot match that sequence.
5926 if (Op0.getOpcode() != ISD::EXTRACT_VECTOR_ELT ||
5927 Op1.getOpcode() != ISD::EXTRACT_VECTOR_ELT ||
5928 !isa<ConstantSDNode>(Op0.getOperand(1)) ||
5929 !isa<ConstantSDNode>(Op1.getOperand(1)) ||
5930 Op0.getOperand(1) != Op1.getOperand(1))
5933 unsigned I0 = cast<ConstantSDNode>(Op0.getOperand(1))->getZExtValue();
5937 // We found a valid add/sub node. Update the information accordingly.
5943 // Update InVec0 and InVec1.
5944 if (InVec0.getOpcode() == ISD::UNDEF) {
5945 InVec0 = Op0.getOperand(0);
5946 if (InVec0.getValueType() != VT)
5949 if (InVec1.getOpcode() == ISD::UNDEF) {
5950 InVec1 = Op1.getOperand(0);
5951 if (InVec1.getValueType() != VT)
5955 // Make sure that operands in input to each add/sub node always
5956 // come from a same pair of vectors.
5957 if (InVec0 != Op0.getOperand(0)) {
5958 if (ExpectedOpcode == ISD::FSUB)
5961 // FADD is commutable. Try to commute the operands
5962 // and then test again.
5963 std::swap(Op0, Op1);
5964 if (InVec0 != Op0.getOperand(0))
5968 if (InVec1 != Op1.getOperand(0))
5971 // Update the pair of expected opcodes.
5972 std::swap(ExpectedOpcode, NextExpectedOpcode);
5975 // Don't try to fold this build_vector into an ADDSUB if the inputs are undef.
5976 if (AddFound && SubFound && InVec0.getOpcode() != ISD::UNDEF &&
5977 InVec1.getOpcode() != ISD::UNDEF)
5978 return DAG.getNode(X86ISD::ADDSUB, DL, VT, InVec0, InVec1);
5983 /// Lower BUILD_VECTOR to a horizontal add/sub operation if possible.
5984 static SDValue LowerToHorizontalOp(const BuildVectorSDNode *BV,
5985 const X86Subtarget *Subtarget,
5986 SelectionDAG &DAG) {
5987 EVT VT = BV->getValueType(0);
5988 unsigned NumElts = VT.getVectorNumElements();
5989 unsigned NumUndefsLO = 0;
5990 unsigned NumUndefsHI = 0;
5991 unsigned Half = NumElts/2;
5993 // Count the number of UNDEF operands in the build_vector in input.
5994 for (unsigned i = 0, e = Half; i != e; ++i)
5995 if (BV->getOperand(i)->getOpcode() == ISD::UNDEF)
5998 for (unsigned i = Half, e = NumElts; i != e; ++i)
5999 if (BV->getOperand(i)->getOpcode() == ISD::UNDEF)
6002 // Early exit if this is either a build_vector of all UNDEFs or all the
6003 // operands but one are UNDEF.
6004 if (NumUndefsLO + NumUndefsHI + 1 >= NumElts)
6008 SDValue InVec0, InVec1;
6009 if ((VT == MVT::v4f32 || VT == MVT::v2f64) && Subtarget->hasSSE3()) {
6010 // Try to match an SSE3 float HADD/HSUB.
6011 if (isHorizontalBinOp(BV, ISD::FADD, DAG, 0, NumElts, InVec0, InVec1))
6012 return DAG.getNode(X86ISD::FHADD, DL, VT, InVec0, InVec1);
6014 if (isHorizontalBinOp(BV, ISD::FSUB, DAG, 0, NumElts, InVec0, InVec1))
6015 return DAG.getNode(X86ISD::FHSUB, DL, VT, InVec0, InVec1);
6016 } else if ((VT == MVT::v4i32 || VT == MVT::v8i16) && Subtarget->hasSSSE3()) {
6017 // Try to match an SSSE3 integer HADD/HSUB.
6018 if (isHorizontalBinOp(BV, ISD::ADD, DAG, 0, NumElts, InVec0, InVec1))
6019 return DAG.getNode(X86ISD::HADD, DL, VT, InVec0, InVec1);
6021 if (isHorizontalBinOp(BV, ISD::SUB, DAG, 0, NumElts, InVec0, InVec1))
6022 return DAG.getNode(X86ISD::HSUB, DL, VT, InVec0, InVec1);
6025 if (!Subtarget->hasAVX())
6028 if ((VT == MVT::v8f32 || VT == MVT::v4f64)) {
6029 // Try to match an AVX horizontal add/sub of packed single/double
6030 // precision floating point values from 256-bit vectors.
6031 SDValue InVec2, InVec3;
6032 if (isHorizontalBinOp(BV, ISD::FADD, DAG, 0, Half, InVec0, InVec1) &&
6033 isHorizontalBinOp(BV, ISD::FADD, DAG, Half, NumElts, InVec2, InVec3) &&
6034 ((InVec0.getOpcode() == ISD::UNDEF ||
6035 InVec2.getOpcode() == ISD::UNDEF) || InVec0 == InVec2) &&
6036 ((InVec1.getOpcode() == ISD::UNDEF ||
6037 InVec3.getOpcode() == ISD::UNDEF) || InVec1 == InVec3))
6038 return DAG.getNode(X86ISD::FHADD, DL, VT, InVec0, InVec1);
6040 if (isHorizontalBinOp(BV, ISD::FSUB, DAG, 0, Half, InVec0, InVec1) &&
6041 isHorizontalBinOp(BV, ISD::FSUB, DAG, Half, NumElts, InVec2, InVec3) &&
6042 ((InVec0.getOpcode() == ISD::UNDEF ||
6043 InVec2.getOpcode() == ISD::UNDEF) || InVec0 == InVec2) &&
6044 ((InVec1.getOpcode() == ISD::UNDEF ||
6045 InVec3.getOpcode() == ISD::UNDEF) || InVec1 == InVec3))
6046 return DAG.getNode(X86ISD::FHSUB, DL, VT, InVec0, InVec1);
6047 } else if (VT == MVT::v8i32 || VT == MVT::v16i16) {
6048 // Try to match an AVX2 horizontal add/sub of signed integers.
6049 SDValue InVec2, InVec3;
6051 bool CanFold = true;
6053 if (isHorizontalBinOp(BV, ISD::ADD, DAG, 0, Half, InVec0, InVec1) &&
6054 isHorizontalBinOp(BV, ISD::ADD, DAG, Half, NumElts, InVec2, InVec3) &&
6055 ((InVec0.getOpcode() == ISD::UNDEF ||
6056 InVec2.getOpcode() == ISD::UNDEF) || InVec0 == InVec2) &&
6057 ((InVec1.getOpcode() == ISD::UNDEF ||
6058 InVec3.getOpcode() == ISD::UNDEF) || InVec1 == InVec3))
6059 X86Opcode = X86ISD::HADD;
6060 else if (isHorizontalBinOp(BV, ISD::SUB, DAG, 0, Half, InVec0, InVec1) &&
6061 isHorizontalBinOp(BV, ISD::SUB, DAG, Half, NumElts, InVec2, InVec3) &&
6062 ((InVec0.getOpcode() == ISD::UNDEF ||
6063 InVec2.getOpcode() == ISD::UNDEF) || InVec0 == InVec2) &&
6064 ((InVec1.getOpcode() == ISD::UNDEF ||
6065 InVec3.getOpcode() == ISD::UNDEF) || InVec1 == InVec3))
6066 X86Opcode = X86ISD::HSUB;
6071 // Fold this build_vector into a single horizontal add/sub.
6072 // Do this only if the target has AVX2.
6073 if (Subtarget->hasAVX2())
6074 return DAG.getNode(X86Opcode, DL, VT, InVec0, InVec1);
6076 // Do not try to expand this build_vector into a pair of horizontal
6077 // add/sub if we can emit a pair of scalar add/sub.
6078 if (NumUndefsLO + 1 == Half || NumUndefsHI + 1 == Half)
6081 // Convert this build_vector into a pair of horizontal binop followed by
6083 bool isUndefLO = NumUndefsLO == Half;
6084 bool isUndefHI = NumUndefsHI == Half;
6085 return ExpandHorizontalBinOp(InVec0, InVec1, DL, DAG, X86Opcode, false,
6086 isUndefLO, isUndefHI);
6090 if ((VT == MVT::v8f32 || VT == MVT::v4f64 || VT == MVT::v8i32 ||
6091 VT == MVT::v16i16) && Subtarget->hasAVX()) {
6093 if (isHorizontalBinOp(BV, ISD::ADD, DAG, 0, NumElts, InVec0, InVec1))
6094 X86Opcode = X86ISD::HADD;
6095 else if (isHorizontalBinOp(BV, ISD::SUB, DAG, 0, NumElts, InVec0, InVec1))
6096 X86Opcode = X86ISD::HSUB;
6097 else if (isHorizontalBinOp(BV, ISD::FADD, DAG, 0, NumElts, InVec0, InVec1))
6098 X86Opcode = X86ISD::FHADD;
6099 else if (isHorizontalBinOp(BV, ISD::FSUB, DAG, 0, NumElts, InVec0, InVec1))
6100 X86Opcode = X86ISD::FHSUB;
6104 // Don't try to expand this build_vector into a pair of horizontal add/sub
6105 // if we can simply emit a pair of scalar add/sub.
6106 if (NumUndefsLO + 1 == Half || NumUndefsHI + 1 == Half)
6109 // Convert this build_vector into two horizontal add/sub followed by
6111 bool isUndefLO = NumUndefsLO == Half;
6112 bool isUndefHI = NumUndefsHI == Half;
6113 return ExpandHorizontalBinOp(InVec0, InVec1, DL, DAG, X86Opcode, true,
6114 isUndefLO, isUndefHI);
6121 X86TargetLowering::LowerBUILD_VECTOR(SDValue Op, SelectionDAG &DAG) const {
6124 MVT VT = Op.getSimpleValueType();
6125 MVT ExtVT = VT.getVectorElementType();
6126 unsigned NumElems = Op.getNumOperands();
6128 // Generate vectors for predicate vectors.
6129 if (VT.getScalarType() == MVT::i1 && Subtarget->hasAVX512())
6130 return LowerBUILD_VECTORvXi1(Op, DAG);
6132 // Vectors containing all zeros can be matched by pxor and xorps later
6133 if (ISD::isBuildVectorAllZeros(Op.getNode())) {
6134 // Canonicalize this to <4 x i32> to 1) ensure the zero vectors are CSE'd
6135 // and 2) ensure that i64 scalars are eliminated on x86-32 hosts.
6136 if (VT == MVT::v4i32 || VT == MVT::v8i32 || VT == MVT::v16i32)
6139 return getZeroVector(VT, Subtarget, DAG, dl);
6142 // Vectors containing all ones can be matched by pcmpeqd on 128-bit width
6143 // vectors or broken into v4i32 operations on 256-bit vectors. AVX2 can use
6144 // vpcmpeqd on 256-bit vectors.
6145 if (Subtarget->hasSSE2() && ISD::isBuildVectorAllOnes(Op.getNode())) {
6146 if (VT == MVT::v4i32 || (VT == MVT::v8i32 && Subtarget->hasInt256()))
6149 if (!VT.is512BitVector())
6150 return getOnesVector(VT, Subtarget, DAG, dl);
6153 BuildVectorSDNode *BV = cast<BuildVectorSDNode>(Op.getNode());
6154 if (SDValue AddSub = LowerToAddSub(BV, Subtarget, DAG))
6156 if (SDValue HorizontalOp = LowerToHorizontalOp(BV, Subtarget, DAG))
6157 return HorizontalOp;
6158 if (SDValue Broadcast = LowerVectorBroadcast(Op, Subtarget, DAG))
6161 unsigned EVTBits = ExtVT.getSizeInBits();
6163 unsigned NumZero = 0;
6164 unsigned NumNonZero = 0;
6165 unsigned NonZeros = 0;
6166 bool IsAllConstants = true;
6167 SmallSet<SDValue, 8> Values;
6168 for (unsigned i = 0; i < NumElems; ++i) {
6169 SDValue Elt = Op.getOperand(i);
6170 if (Elt.getOpcode() == ISD::UNDEF)
6173 if (Elt.getOpcode() != ISD::Constant &&
6174 Elt.getOpcode() != ISD::ConstantFP)
6175 IsAllConstants = false;
6176 if (X86::isZeroNode(Elt))
6179 NonZeros |= (1 << i);
6184 // All undef vector. Return an UNDEF. All zero vectors were handled above.
6185 if (NumNonZero == 0)
6186 return DAG.getUNDEF(VT);
6188 // Special case for single non-zero, non-undef, element.
6189 if (NumNonZero == 1) {
6190 unsigned Idx = countTrailingZeros(NonZeros);
6191 SDValue Item = Op.getOperand(Idx);
6193 // If this is an insertion of an i64 value on x86-32, and if the top bits of
6194 // the value are obviously zero, truncate the value to i32 and do the
6195 // insertion that way. Only do this if the value is non-constant or if the
6196 // value is a constant being inserted into element 0. It is cheaper to do
6197 // a constant pool load than it is to do a movd + shuffle.
6198 if (ExtVT == MVT::i64 && !Subtarget->is64Bit() &&
6199 (!IsAllConstants || Idx == 0)) {
6200 if (DAG.MaskedValueIsZero(Item, APInt::getBitsSet(64, 32, 64))) {
6202 assert(VT == MVT::v2i64 && "Expected an SSE value type!");
6203 EVT VecVT = MVT::v4i32;
6205 // Truncate the value (which may itself be a constant) to i32, and
6206 // convert it to a vector with movd (S2V+shuffle to zero extend).
6207 Item = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, Item);
6208 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VecVT, Item);
6209 return DAG.getBitcast(VT, getShuffleVectorZeroOrUndef(
6210 Item, Idx * 2, true, Subtarget, DAG));
6214 // If we have a constant or non-constant insertion into the low element of
6215 // a vector, we can do this with SCALAR_TO_VECTOR + shuffle of zero into
6216 // the rest of the elements. This will be matched as movd/movq/movss/movsd
6217 // depending on what the source datatype is.
6220 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
6222 if (ExtVT == MVT::i32 || ExtVT == MVT::f32 || ExtVT == MVT::f64 ||
6223 (ExtVT == MVT::i64 && Subtarget->is64Bit())) {
6224 if (VT.is512BitVector()) {
6225 SDValue ZeroVec = getZeroVector(VT, Subtarget, DAG, dl);
6226 return DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, ZeroVec,
6227 Item, DAG.getIntPtrConstant(0, dl));
6229 assert((VT.is128BitVector() || VT.is256BitVector()) &&
6230 "Expected an SSE value type!");
6231 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
6232 // Turn it into a MOVL (i.e. movss, movsd, or movd) to a zero vector.
6233 return getShuffleVectorZeroOrUndef(Item, 0, true, Subtarget, DAG);
6236 // We can't directly insert an i8 or i16 into a vector, so zero extend
6238 if (ExtVT == MVT::i16 || ExtVT == MVT::i8) {
6239 Item = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, Item);
6240 if (VT.is256BitVector()) {
6241 if (Subtarget->hasAVX()) {
6242 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v8i32, Item);
6243 Item = getShuffleVectorZeroOrUndef(Item, 0, true, Subtarget, DAG);
6245 // Without AVX, we need to extend to a 128-bit vector and then
6246 // insert into the 256-bit vector.
6247 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32, Item);
6248 SDValue ZeroVec = getZeroVector(MVT::v8i32, Subtarget, DAG, dl);
6249 Item = Insert128BitVector(ZeroVec, Item, 0, DAG, dl);
6252 assert(VT.is128BitVector() && "Expected an SSE value type!");
6253 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32, Item);
6254 Item = getShuffleVectorZeroOrUndef(Item, 0, true, Subtarget, DAG);
6256 return DAG.getBitcast(VT, Item);
6260 // Is it a vector logical left shift?
6261 if (NumElems == 2 && Idx == 1 &&
6262 X86::isZeroNode(Op.getOperand(0)) &&
6263 !X86::isZeroNode(Op.getOperand(1))) {
6264 unsigned NumBits = VT.getSizeInBits();
6265 return getVShift(true, VT,
6266 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
6267 VT, Op.getOperand(1)),
6268 NumBits/2, DAG, *this, dl);
6271 if (IsAllConstants) // Otherwise, it's better to do a constpool load.
6274 // Otherwise, if this is a vector with i32 or f32 elements, and the element
6275 // is a non-constant being inserted into an element other than the low one,
6276 // we can't use a constant pool load. Instead, use SCALAR_TO_VECTOR (aka
6277 // movd/movss) to move this into the low element, then shuffle it into
6279 if (EVTBits == 32) {
6280 Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Item);
6281 return getShuffleVectorZeroOrUndef(Item, Idx, NumZero > 0, Subtarget, DAG);
6285 // Splat is obviously ok. Let legalizer expand it to a shuffle.
6286 if (Values.size() == 1) {
6287 if (EVTBits == 32) {
6288 // Instead of a shuffle like this:
6289 // shuffle (scalar_to_vector (load (ptr + 4))), undef, <0, 0, 0, 0>
6290 // Check if it's possible to issue this instead.
6291 // shuffle (vload ptr)), undef, <1, 1, 1, 1>
6292 unsigned Idx = countTrailingZeros(NonZeros);
6293 SDValue Item = Op.getOperand(Idx);
6294 if (Op.getNode()->isOnlyUserOf(Item.getNode()))
6295 return LowerAsSplatVectorLoad(Item, VT, dl, DAG);
6300 // A vector full of immediates; various special cases are already
6301 // handled, so this is best done with a single constant-pool load.
6305 // For AVX-length vectors, see if we can use a vector load to get all of the
6306 // elements, otherwise build the individual 128-bit pieces and use
6307 // shuffles to put them in place.
6308 if (VT.is256BitVector() || VT.is512BitVector()) {
6309 SmallVector<SDValue, 64> V(Op->op_begin(), Op->op_begin() + NumElems);
6311 // Check for a build vector of consecutive loads.
6312 if (SDValue LD = EltsFromConsecutiveLoads(VT, V, dl, DAG, false))
6315 EVT HVT = EVT::getVectorVT(*DAG.getContext(), ExtVT, NumElems/2);
6317 // Build both the lower and upper subvector.
6318 SDValue Lower = DAG.getNode(ISD::BUILD_VECTOR, dl, HVT,
6319 makeArrayRef(&V[0], NumElems/2));
6320 SDValue Upper = DAG.getNode(ISD::BUILD_VECTOR, dl, HVT,
6321 makeArrayRef(&V[NumElems / 2], NumElems/2));
6323 // Recreate the wider vector with the lower and upper part.
6324 if (VT.is256BitVector())
6325 return Concat128BitVectors(Lower, Upper, VT, NumElems, DAG, dl);
6326 return Concat256BitVectors(Lower, Upper, VT, NumElems, DAG, dl);
6329 // Let legalizer expand 2-wide build_vectors.
6330 if (EVTBits == 64) {
6331 if (NumNonZero == 1) {
6332 // One half is zero or undef.
6333 unsigned Idx = countTrailingZeros(NonZeros);
6334 SDValue V2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT,
6335 Op.getOperand(Idx));
6336 return getShuffleVectorZeroOrUndef(V2, Idx, true, Subtarget, DAG);
6341 // If element VT is < 32 bits, convert it to inserts into a zero vector.
6342 if (EVTBits == 8 && NumElems == 16)
6343 if (SDValue V = LowerBuildVectorv16i8(Op, NonZeros,NumNonZero,NumZero, DAG,
6347 if (EVTBits == 16 && NumElems == 8)
6348 if (SDValue V = LowerBuildVectorv8i16(Op, NonZeros,NumNonZero,NumZero, DAG,
6352 // If element VT is == 32 bits and has 4 elems, try to generate an INSERTPS
6353 if (EVTBits == 32 && NumElems == 4)
6354 if (SDValue V = LowerBuildVectorv4x32(Op, DAG, Subtarget, *this))
6357 // If element VT is == 32 bits, turn it into a number of shuffles.
6358 SmallVector<SDValue, 8> V(NumElems);
6359 if (NumElems == 4 && NumZero > 0) {
6360 for (unsigned i = 0; i < 4; ++i) {
6361 bool isZero = !(NonZeros & (1 << i));
6363 V[i] = getZeroVector(VT, Subtarget, DAG, dl);
6365 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(i));
6368 for (unsigned i = 0; i < 2; ++i) {
6369 switch ((NonZeros & (0x3 << i*2)) >> (i*2)) {
6372 V[i] = V[i*2]; // Must be a zero vector.
6375 V[i] = getMOVL(DAG, dl, VT, V[i*2+1], V[i*2]);
6378 V[i] = getMOVL(DAG, dl, VT, V[i*2], V[i*2+1]);
6381 V[i] = getUnpackl(DAG, dl, VT, V[i*2], V[i*2+1]);
6386 bool Reverse1 = (NonZeros & 0x3) == 2;
6387 bool Reverse2 = ((NonZeros & (0x3 << 2)) >> 2) == 2;
6391 static_cast<int>(Reverse2 ? NumElems+1 : NumElems),
6392 static_cast<int>(Reverse2 ? NumElems : NumElems+1)
6394 return DAG.getVectorShuffle(VT, dl, V[0], V[1], &MaskVec[0]);
6397 if (Values.size() > 1 && VT.is128BitVector()) {
6398 // Check for a build vector of consecutive loads.
6399 for (unsigned i = 0; i < NumElems; ++i)
6400 V[i] = Op.getOperand(i);
6402 // Check for elements which are consecutive loads.
6403 if (SDValue LD = EltsFromConsecutiveLoads(VT, V, dl, DAG, false))
6406 // Check for a build vector from mostly shuffle plus few inserting.
6407 if (SDValue Sh = buildFromShuffleMostly(Op, DAG))
6410 // For SSE 4.1, use insertps to put the high elements into the low element.
6411 if (Subtarget->hasSSE41()) {
6413 if (Op.getOperand(0).getOpcode() != ISD::UNDEF)
6414 Result = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(0));
6416 Result = DAG.getUNDEF(VT);
6418 for (unsigned i = 1; i < NumElems; ++i) {
6419 if (Op.getOperand(i).getOpcode() == ISD::UNDEF) continue;
6420 Result = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, Result,
6421 Op.getOperand(i), DAG.getIntPtrConstant(i, dl));
6426 // Otherwise, expand into a number of unpckl*, start by extending each of
6427 // our (non-undef) elements to the full vector width with the element in the
6428 // bottom slot of the vector (which generates no code for SSE).
6429 for (unsigned i = 0; i < NumElems; ++i) {
6430 if (Op.getOperand(i).getOpcode() != ISD::UNDEF)
6431 V[i] = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(i));
6433 V[i] = DAG.getUNDEF(VT);
6436 // Next, we iteratively mix elements, e.g. for v4f32:
6437 // Step 1: unpcklps 0, 2 ==> X: <?, ?, 2, 0>
6438 // : unpcklps 1, 3 ==> Y: <?, ?, 3, 1>
6439 // Step 2: unpcklps X, Y ==> <3, 2, 1, 0>
6440 unsigned EltStride = NumElems >> 1;
6441 while (EltStride != 0) {
6442 for (unsigned i = 0; i < EltStride; ++i) {
6443 // If V[i+EltStride] is undef and this is the first round of mixing,
6444 // then it is safe to just drop this shuffle: V[i] is already in the
6445 // right place, the one element (since it's the first round) being
6446 // inserted as undef can be dropped. This isn't safe for successive
6447 // rounds because they will permute elements within both vectors.
6448 if (V[i+EltStride].getOpcode() == ISD::UNDEF &&
6449 EltStride == NumElems/2)
6452 V[i] = getUnpackl(DAG, dl, VT, V[i], V[i + EltStride]);
6461 // 256-bit AVX can use the vinsertf128 instruction
6462 // to create 256-bit vectors from two other 128-bit ones.
6463 static SDValue LowerAVXCONCAT_VECTORS(SDValue Op, SelectionDAG &DAG) {
6465 MVT ResVT = Op.getSimpleValueType();
6467 assert((ResVT.is256BitVector() ||
6468 ResVT.is512BitVector()) && "Value type must be 256-/512-bit wide");
6470 SDValue V1 = Op.getOperand(0);
6471 SDValue V2 = Op.getOperand(1);
6472 unsigned NumElems = ResVT.getVectorNumElements();
6473 if (ResVT.is256BitVector())
6474 return Concat128BitVectors(V1, V2, ResVT, NumElems, DAG, dl);
6476 if (Op.getNumOperands() == 4) {
6477 MVT HalfVT = MVT::getVectorVT(ResVT.getScalarType(),
6478 ResVT.getVectorNumElements()/2);
6479 SDValue V3 = Op.getOperand(2);
6480 SDValue V4 = Op.getOperand(3);
6481 return Concat256BitVectors(Concat128BitVectors(V1, V2, HalfVT, NumElems/2, DAG, dl),
6482 Concat128BitVectors(V3, V4, HalfVT, NumElems/2, DAG, dl), ResVT, NumElems, DAG, dl);
6484 return Concat256BitVectors(V1, V2, ResVT, NumElems, DAG, dl);
6487 static SDValue LowerCONCAT_VECTORSvXi1(SDValue Op,
6488 const X86Subtarget *Subtarget,
6489 SelectionDAG & DAG) {
6491 MVT ResVT = Op.getSimpleValueType();
6492 unsigned NumOfOperands = Op.getNumOperands();
6494 assert(isPowerOf2_32(NumOfOperands) &&
6495 "Unexpected number of operands in CONCAT_VECTORS");
6497 if (NumOfOperands > 2) {
6498 MVT HalfVT = MVT::getVectorVT(ResVT.getScalarType(),
6499 ResVT.getVectorNumElements()/2);
6500 SmallVector<SDValue, 2> Ops;
6501 for (unsigned i = 0; i < NumOfOperands/2; i++)
6502 Ops.push_back(Op.getOperand(i));
6503 SDValue Lo = DAG.getNode(ISD::CONCAT_VECTORS, dl, HalfVT, Ops);
6505 for (unsigned i = NumOfOperands/2; i < NumOfOperands; i++)
6506 Ops.push_back(Op.getOperand(i));
6507 SDValue Hi = DAG.getNode(ISD::CONCAT_VECTORS, dl, HalfVT, Ops);
6508 return DAG.getNode(ISD::CONCAT_VECTORS, dl, ResVT, Lo, Hi);
6511 SDValue V1 = Op.getOperand(0);
6512 SDValue V2 = Op.getOperand(1);
6513 bool IsZeroV1 = ISD::isBuildVectorAllZeros(V1.getNode());
6514 bool IsZeroV2 = ISD::isBuildVectorAllZeros(V2.getNode());
6516 if (IsZeroV1 && IsZeroV2)
6517 return getZeroVector(ResVT, Subtarget, DAG, dl);
6519 SDValue ZeroIdx = DAG.getIntPtrConstant(0, dl);
6520 SDValue Undef = DAG.getUNDEF(ResVT);
6521 unsigned NumElems = ResVT.getVectorNumElements();
6522 SDValue ShiftBits = DAG.getConstant(NumElems/2, dl, MVT::i8);
6524 V2 = DAG.getNode(ISD::INSERT_SUBVECTOR, dl, ResVT, Undef, V2, ZeroIdx);
6525 V2 = DAG.getNode(X86ISD::VSHLI, dl, ResVT, V2, ShiftBits);
6529 V1 = DAG.getNode(ISD::INSERT_SUBVECTOR, dl, ResVT, Undef, V1, ZeroIdx);
6530 // Zero the upper bits of V1
6531 V1 = DAG.getNode(X86ISD::VSHLI, dl, ResVT, V1, ShiftBits);
6532 V1 = DAG.getNode(X86ISD::VSRLI, dl, ResVT, V1, ShiftBits);
6535 return DAG.getNode(ISD::OR, dl, ResVT, V1, V2);
6538 static SDValue LowerCONCAT_VECTORS(SDValue Op,
6539 const X86Subtarget *Subtarget,
6540 SelectionDAG &DAG) {
6541 MVT VT = Op.getSimpleValueType();
6542 if (VT.getVectorElementType() == MVT::i1)
6543 return LowerCONCAT_VECTORSvXi1(Op, Subtarget, DAG);
6545 assert((VT.is256BitVector() && Op.getNumOperands() == 2) ||
6546 (VT.is512BitVector() && (Op.getNumOperands() == 2 ||
6547 Op.getNumOperands() == 4)));
6549 // AVX can use the vinsertf128 instruction to create 256-bit vectors
6550 // from two other 128-bit ones.
6552 // 512-bit vector may contain 2 256-bit vectors or 4 128-bit vectors
6553 return LowerAVXCONCAT_VECTORS(Op, DAG);
6556 //===----------------------------------------------------------------------===//
6557 // Vector shuffle lowering
6559 // This is an experimental code path for lowering vector shuffles on x86. It is
6560 // designed to handle arbitrary vector shuffles and blends, gracefully
6561 // degrading performance as necessary. It works hard to recognize idiomatic
6562 // shuffles and lower them to optimal instruction patterns without leaving
6563 // a framework that allows reasonably efficient handling of all vector shuffle
6565 //===----------------------------------------------------------------------===//
6567 /// \brief Tiny helper function to identify a no-op mask.
6569 /// This is a somewhat boring predicate function. It checks whether the mask
6570 /// array input, which is assumed to be a single-input shuffle mask of the kind
6571 /// used by the X86 shuffle instructions (not a fully general
6572 /// ShuffleVectorSDNode mask) requires any shuffles to occur. Both undef and an
6573 /// in-place shuffle are 'no-op's.
6574 static bool isNoopShuffleMask(ArrayRef<int> Mask) {
6575 for (int i = 0, Size = Mask.size(); i < Size; ++i)
6576 if (Mask[i] != -1 && Mask[i] != i)
6581 /// \brief Helper function to classify a mask as a single-input mask.
6583 /// This isn't a generic single-input test because in the vector shuffle
6584 /// lowering we canonicalize single inputs to be the first input operand. This
6585 /// means we can more quickly test for a single input by only checking whether
6586 /// an input from the second operand exists. We also assume that the size of
6587 /// mask corresponds to the size of the input vectors which isn't true in the
6588 /// fully general case.
6589 static bool isSingleInputShuffleMask(ArrayRef<int> Mask) {
6591 if (M >= (int)Mask.size())
6596 /// \brief Test whether there are elements crossing 128-bit lanes in this
6599 /// X86 divides up its shuffles into in-lane and cross-lane shuffle operations
6600 /// and we routinely test for these.
6601 static bool is128BitLaneCrossingShuffleMask(MVT VT, ArrayRef<int> Mask) {
6602 int LaneSize = 128 / VT.getScalarSizeInBits();
6603 int Size = Mask.size();
6604 for (int i = 0; i < Size; ++i)
6605 if (Mask[i] >= 0 && (Mask[i] % Size) / LaneSize != i / LaneSize)
6610 /// \brief Test whether a shuffle mask is equivalent within each 128-bit lane.
6612 /// This checks a shuffle mask to see if it is performing the same
6613 /// 128-bit lane-relative shuffle in each 128-bit lane. This trivially implies
6614 /// that it is also not lane-crossing. It may however involve a blend from the
6615 /// same lane of a second vector.
6617 /// The specific repeated shuffle mask is populated in \p RepeatedMask, as it is
6618 /// non-trivial to compute in the face of undef lanes. The representation is
6619 /// *not* suitable for use with existing 128-bit shuffles as it will contain
6620 /// entries from both V1 and V2 inputs to the wider mask.
6622 is128BitLaneRepeatedShuffleMask(MVT VT, ArrayRef<int> Mask,
6623 SmallVectorImpl<int> &RepeatedMask) {
6624 int LaneSize = 128 / VT.getScalarSizeInBits();
6625 RepeatedMask.resize(LaneSize, -1);
6626 int Size = Mask.size();
6627 for (int i = 0; i < Size; ++i) {
6630 if ((Mask[i] % Size) / LaneSize != i / LaneSize)
6631 // This entry crosses lanes, so there is no way to model this shuffle.
6634 // Ok, handle the in-lane shuffles by detecting if and when they repeat.
6635 if (RepeatedMask[i % LaneSize] == -1)
6636 // This is the first non-undef entry in this slot of a 128-bit lane.
6637 RepeatedMask[i % LaneSize] =
6638 Mask[i] < Size ? Mask[i] % LaneSize : Mask[i] % LaneSize + Size;
6639 else if (RepeatedMask[i % LaneSize] + (i / LaneSize) * LaneSize != Mask[i])
6640 // Found a mismatch with the repeated mask.
6646 /// \brief Checks whether a shuffle mask is equivalent to an explicit list of
6649 /// This is a fast way to test a shuffle mask against a fixed pattern:
6651 /// if (isShuffleEquivalent(Mask, 3, 2, {1, 0})) { ... }
6653 /// It returns true if the mask is exactly as wide as the argument list, and
6654 /// each element of the mask is either -1 (signifying undef) or the value given
6655 /// in the argument.
6656 static bool isShuffleEquivalent(SDValue V1, SDValue V2, ArrayRef<int> Mask,
6657 ArrayRef<int> ExpectedMask) {
6658 if (Mask.size() != ExpectedMask.size())
6661 int Size = Mask.size();
6663 // If the values are build vectors, we can look through them to find
6664 // equivalent inputs that make the shuffles equivalent.
6665 auto *BV1 = dyn_cast<BuildVectorSDNode>(V1);
6666 auto *BV2 = dyn_cast<BuildVectorSDNode>(V2);
6668 for (int i = 0; i < Size; ++i)
6669 if (Mask[i] != -1 && Mask[i] != ExpectedMask[i]) {
6670 auto *MaskBV = Mask[i] < Size ? BV1 : BV2;
6671 auto *ExpectedBV = ExpectedMask[i] < Size ? BV1 : BV2;
6672 if (!MaskBV || !ExpectedBV ||
6673 MaskBV->getOperand(Mask[i] % Size) !=
6674 ExpectedBV->getOperand(ExpectedMask[i] % Size))
6681 /// \brief Get a 4-lane 8-bit shuffle immediate for a mask.
6683 /// This helper function produces an 8-bit shuffle immediate corresponding to
6684 /// the ubiquitous shuffle encoding scheme used in x86 instructions for
6685 /// shuffling 4 lanes. It can be used with most of the PSHUF instructions for
6688 /// NB: We rely heavily on "undef" masks preserving the input lane.
6689 static SDValue getV4X86ShuffleImm8ForMask(ArrayRef<int> Mask, SDLoc DL,
6690 SelectionDAG &DAG) {
6691 assert(Mask.size() == 4 && "Only 4-lane shuffle masks");
6692 assert(Mask[0] >= -1 && Mask[0] < 4 && "Out of bound mask element!");
6693 assert(Mask[1] >= -1 && Mask[1] < 4 && "Out of bound mask element!");
6694 assert(Mask[2] >= -1 && Mask[2] < 4 && "Out of bound mask element!");
6695 assert(Mask[3] >= -1 && Mask[3] < 4 && "Out of bound mask element!");
6698 Imm |= (Mask[0] == -1 ? 0 : Mask[0]) << 0;
6699 Imm |= (Mask[1] == -1 ? 1 : Mask[1]) << 2;
6700 Imm |= (Mask[2] == -1 ? 2 : Mask[2]) << 4;
6701 Imm |= (Mask[3] == -1 ? 3 : Mask[3]) << 6;
6702 return DAG.getConstant(Imm, DL, MVT::i8);
6705 /// \brief Compute whether each element of a shuffle is zeroable.
6707 /// A "zeroable" vector shuffle element is one which can be lowered to zero.
6708 /// Either it is an undef element in the shuffle mask, the element of the input
6709 /// referenced is undef, or the element of the input referenced is known to be
6710 /// zero. Many x86 shuffles can zero lanes cheaply and we often want to handle
6711 /// as many lanes with this technique as possible to simplify the remaining
6713 static SmallBitVector computeZeroableShuffleElements(ArrayRef<int> Mask,
6714 SDValue V1, SDValue V2) {
6715 SmallBitVector Zeroable(Mask.size(), false);
6717 while (V1.getOpcode() == ISD::BITCAST)
6718 V1 = V1->getOperand(0);
6719 while (V2.getOpcode() == ISD::BITCAST)
6720 V2 = V2->getOperand(0);
6722 bool V1IsZero = ISD::isBuildVectorAllZeros(V1.getNode());
6723 bool V2IsZero = ISD::isBuildVectorAllZeros(V2.getNode());
6725 for (int i = 0, Size = Mask.size(); i < Size; ++i) {
6727 // Handle the easy cases.
6728 if (M < 0 || (M >= 0 && M < Size && V1IsZero) || (M >= Size && V2IsZero)) {
6733 // If this is an index into a build_vector node (which has the same number
6734 // of elements), dig out the input value and use it.
6735 SDValue V = M < Size ? V1 : V2;
6736 if (V.getOpcode() != ISD::BUILD_VECTOR || Size != (int)V.getNumOperands())
6739 SDValue Input = V.getOperand(M % Size);
6740 // The UNDEF opcode check really should be dead code here, but not quite
6741 // worth asserting on (it isn't invalid, just unexpected).
6742 if (Input.getOpcode() == ISD::UNDEF || X86::isZeroNode(Input))
6749 // X86 has dedicated unpack instructions that can handle specific blend
6750 // operations: UNPCKH and UNPCKL.
6751 static SDValue lowerVectorShuffleWithUNPCK(SDLoc DL, MVT VT, ArrayRef<int> Mask,
6752 SDValue V1, SDValue V2,
6753 SelectionDAG &DAG) {
6754 int NumElts = VT.getVectorNumElements();
6755 int NumEltsInLane = 128 / VT.getScalarSizeInBits();
6756 SmallVector<int, 8> Unpckl;
6757 SmallVector<int, 8> Unpckh;
6759 for (int i = 0; i < NumElts; ++i) {
6760 unsigned LaneStart = (i / NumEltsInLane) * NumEltsInLane;
6761 int LoPos = (i % NumEltsInLane) / 2 + LaneStart + NumElts * (i % 2);
6762 int HiPos = LoPos + NumEltsInLane / 2;
6763 Unpckl.push_back(LoPos);
6764 Unpckh.push_back(HiPos);
6767 if (isShuffleEquivalent(V1, V2, Mask, Unpckl))
6768 return DAG.getNode(X86ISD::UNPCKL, DL, VT, V1, V2);
6769 if (isShuffleEquivalent(V1, V2, Mask, Unpckh))
6770 return DAG.getNode(X86ISD::UNPCKH, DL, VT, V1, V2);
6772 // Commute and try again.
6773 ShuffleVectorSDNode::commuteMask(Unpckl);
6774 if (isShuffleEquivalent(V1, V2, Mask, Unpckl))
6775 return DAG.getNode(X86ISD::UNPCKL, DL, VT, V2, V1);
6777 ShuffleVectorSDNode::commuteMask(Unpckh);
6778 if (isShuffleEquivalent(V1, V2, Mask, Unpckh))
6779 return DAG.getNode(X86ISD::UNPCKH, DL, VT, V2, V1);
6784 /// \brief Try to emit a bitmask instruction for a shuffle.
6786 /// This handles cases where we can model a blend exactly as a bitmask due to
6787 /// one of the inputs being zeroable.
6788 static SDValue lowerVectorShuffleAsBitMask(SDLoc DL, MVT VT, SDValue V1,
6789 SDValue V2, ArrayRef<int> Mask,
6790 SelectionDAG &DAG) {
6791 MVT EltVT = VT.getScalarType();
6792 int NumEltBits = EltVT.getSizeInBits();
6793 MVT IntEltVT = MVT::getIntegerVT(NumEltBits);
6794 SDValue Zero = DAG.getConstant(0, DL, IntEltVT);
6795 SDValue AllOnes = DAG.getConstant(APInt::getAllOnesValue(NumEltBits), DL,
6797 if (EltVT.isFloatingPoint()) {
6798 Zero = DAG.getBitcast(EltVT, Zero);
6799 AllOnes = DAG.getBitcast(EltVT, AllOnes);
6801 SmallVector<SDValue, 16> VMaskOps(Mask.size(), Zero);
6802 SmallBitVector Zeroable = computeZeroableShuffleElements(Mask, V1, V2);
6804 for (int i = 0, Size = Mask.size(); i < Size; ++i) {
6807 if (Mask[i] % Size != i)
6808 return SDValue(); // Not a blend.
6810 V = Mask[i] < Size ? V1 : V2;
6811 else if (V != (Mask[i] < Size ? V1 : V2))
6812 return SDValue(); // Can only let one input through the mask.
6814 VMaskOps[i] = AllOnes;
6817 return SDValue(); // No non-zeroable elements!
6819 SDValue VMask = DAG.getNode(ISD::BUILD_VECTOR, DL, VT, VMaskOps);
6820 V = DAG.getNode(VT.isFloatingPoint()
6821 ? (unsigned) X86ISD::FAND : (unsigned) ISD::AND,
6826 /// \brief Try to emit a blend instruction for a shuffle using bit math.
6828 /// This is used as a fallback approach when first class blend instructions are
6829 /// unavailable. Currently it is only suitable for integer vectors, but could
6830 /// be generalized for floating point vectors if desirable.
6831 static SDValue lowerVectorShuffleAsBitBlend(SDLoc DL, MVT VT, SDValue V1,
6832 SDValue V2, ArrayRef<int> Mask,
6833 SelectionDAG &DAG) {
6834 assert(VT.isInteger() && "Only supports integer vector types!");
6835 MVT EltVT = VT.getScalarType();
6836 int NumEltBits = EltVT.getSizeInBits();
6837 SDValue Zero = DAG.getConstant(0, DL, EltVT);
6838 SDValue AllOnes = DAG.getConstant(APInt::getAllOnesValue(NumEltBits), DL,
6840 SmallVector<SDValue, 16> MaskOps;
6841 for (int i = 0, Size = Mask.size(); i < Size; ++i) {
6842 if (Mask[i] != -1 && Mask[i] != i && Mask[i] != i + Size)
6843 return SDValue(); // Shuffled input!
6844 MaskOps.push_back(Mask[i] < Size ? AllOnes : Zero);
6847 SDValue V1Mask = DAG.getNode(ISD::BUILD_VECTOR, DL, VT, MaskOps);
6848 V1 = DAG.getNode(ISD::AND, DL, VT, V1, V1Mask);
6849 // We have to cast V2 around.
6850 MVT MaskVT = MVT::getVectorVT(MVT::i64, VT.getSizeInBits() / 64);
6851 V2 = DAG.getBitcast(VT, DAG.getNode(X86ISD::ANDNP, DL, MaskVT,
6852 DAG.getBitcast(MaskVT, V1Mask),
6853 DAG.getBitcast(MaskVT, V2)));
6854 return DAG.getNode(ISD::OR, DL, VT, V1, V2);
6857 /// \brief Try to emit a blend instruction for a shuffle.
6859 /// This doesn't do any checks for the availability of instructions for blending
6860 /// these values. It relies on the availability of the X86ISD::BLENDI pattern to
6861 /// be matched in the backend with the type given. What it does check for is
6862 /// that the shuffle mask is a blend, or convertible into a blend with zero.
6863 static SDValue lowerVectorShuffleAsBlend(SDLoc DL, MVT VT, SDValue V1,
6864 SDValue V2, ArrayRef<int> Original,
6865 const X86Subtarget *Subtarget,
6866 SelectionDAG &DAG) {
6867 bool V1IsZero = ISD::isBuildVectorAllZeros(V1.getNode());
6868 bool V2IsZero = ISD::isBuildVectorAllZeros(V2.getNode());
6869 SmallVector<int, 8> Mask(Original.begin(), Original.end());
6870 SmallBitVector Zeroable = computeZeroableShuffleElements(Mask, V1, V2);
6871 bool ForceV1Zero = false, ForceV2Zero = false;
6873 // Attempt to generate the binary blend mask. If an input is zero then
6874 // we can use any lane.
6875 // TODO: generalize the zero matching to any scalar like isShuffleEquivalent.
6876 unsigned BlendMask = 0;
6877 for (int i = 0, Size = Mask.size(); i < Size; ++i) {
6883 if (M == i + Size) {
6884 BlendMask |= 1u << i;
6895 BlendMask |= 1u << i;
6900 return SDValue(); // Shuffled input!
6903 // Create a REAL zero vector - ISD::isBuildVectorAllZeros allows UNDEFs.
6905 V1 = getZeroVector(VT, Subtarget, DAG, DL);
6907 V2 = getZeroVector(VT, Subtarget, DAG, DL);
6909 auto ScaleBlendMask = [](unsigned BlendMask, int Size, int Scale) {
6910 unsigned ScaledMask = 0;
6911 for (int i = 0; i != Size; ++i)
6912 if (BlendMask & (1u << i))
6913 for (int j = 0; j != Scale; ++j)
6914 ScaledMask |= 1u << (i * Scale + j);
6918 switch (VT.SimpleTy) {
6923 return DAG.getNode(X86ISD::BLENDI, DL, VT, V1, V2,
6924 DAG.getConstant(BlendMask, DL, MVT::i8));
6928 assert(Subtarget->hasAVX2() && "256-bit integer blends require AVX2!");
6932 // If we have AVX2 it is faster to use VPBLENDD when the shuffle fits into
6933 // that instruction.
6934 if (Subtarget->hasAVX2()) {
6935 // Scale the blend by the number of 32-bit dwords per element.
6936 int Scale = VT.getScalarSizeInBits() / 32;
6937 BlendMask = ScaleBlendMask(BlendMask, Mask.size(), Scale);
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();
6950 BlendMask = ScaleBlendMask(BlendMask, Mask.size(), Scale);
6951 V1 = DAG.getBitcast(MVT::v8i16, V1);
6952 V2 = DAG.getBitcast(MVT::v8i16, V2);
6953 return DAG.getBitcast(VT,
6954 DAG.getNode(X86ISD::BLENDI, DL, MVT::v8i16, V1, V2,
6955 DAG.getConstant(BlendMask, DL, MVT::i8)));
6959 assert(Subtarget->hasAVX2() && "256-bit integer blends require AVX2!");
6960 SmallVector<int, 8> RepeatedMask;
6961 if (is128BitLaneRepeatedShuffleMask(MVT::v16i16, Mask, RepeatedMask)) {
6962 // We can lower these with PBLENDW which is mirrored across 128-bit lanes.
6963 assert(RepeatedMask.size() == 8 && "Repeated mask size doesn't match!");
6965 for (int i = 0; i < 8; ++i)
6966 if (RepeatedMask[i] >= 16)
6967 BlendMask |= 1u << i;
6968 return DAG.getNode(X86ISD::BLENDI, DL, MVT::v16i16, V1, V2,
6969 DAG.getConstant(BlendMask, DL, MVT::i8));
6975 assert((VT.getSizeInBits() == 128 || Subtarget->hasAVX2()) &&
6976 "256-bit byte-blends require AVX2 support!");
6978 // Attempt to lower to a bitmask if we can. VPAND is faster than VPBLENDVB.
6979 if (SDValue Masked = lowerVectorShuffleAsBitMask(DL, VT, V1, V2, Mask, DAG))
6982 // Scale the blend by the number of bytes per element.
6983 int Scale = VT.getScalarSizeInBits() / 8;
6985 // This form of blend is always done on bytes. Compute the byte vector
6987 MVT BlendVT = MVT::getVectorVT(MVT::i8, VT.getSizeInBits() / 8);
6989 // Compute the VSELECT mask. Note that VSELECT is really confusing in the
6990 // mix of LLVM's code generator and the x86 backend. We tell the code
6991 // generator that boolean values in the elements of an x86 vector register
6992 // are -1 for true and 0 for false. We then use the LLVM semantics of 'true'
6993 // mapping a select to operand #1, and 'false' mapping to operand #2. The
6994 // reality in x86 is that vector masks (pre-AVX-512) use only the high bit
6995 // of the element (the remaining are ignored) and 0 in that high bit would
6996 // mean operand #1 while 1 in the high bit would mean operand #2. So while
6997 // the LLVM model for boolean values in vector elements gets the relevant
6998 // bit set, it is set backwards and over constrained relative to x86's
7000 SmallVector<SDValue, 32> VSELECTMask;
7001 for (int i = 0, Size = Mask.size(); i < Size; ++i)
7002 for (int j = 0; j < Scale; ++j)
7003 VSELECTMask.push_back(
7004 Mask[i] < 0 ? DAG.getUNDEF(MVT::i8)
7005 : DAG.getConstant(Mask[i] < Size ? -1 : 0, DL,
7008 V1 = DAG.getBitcast(BlendVT, V1);
7009 V2 = DAG.getBitcast(BlendVT, V2);
7010 return DAG.getBitcast(VT, DAG.getNode(ISD::VSELECT, DL, BlendVT,
7011 DAG.getNode(ISD::BUILD_VECTOR, DL,
7012 BlendVT, VSELECTMask),
7017 llvm_unreachable("Not a supported integer vector type!");
7021 /// \brief Try to lower as a blend of elements from two inputs followed by
7022 /// a single-input permutation.
7024 /// This matches the pattern where we can blend elements from two inputs and
7025 /// then reduce the shuffle to a single-input permutation.
7026 static SDValue lowerVectorShuffleAsBlendAndPermute(SDLoc DL, MVT VT, SDValue V1,
7029 SelectionDAG &DAG) {
7030 // We build up the blend mask while checking whether a blend is a viable way
7031 // to reduce the shuffle.
7032 SmallVector<int, 32> BlendMask(Mask.size(), -1);
7033 SmallVector<int, 32> PermuteMask(Mask.size(), -1);
7035 for (int i = 0, Size = Mask.size(); i < Size; ++i) {
7039 assert(Mask[i] < Size * 2 && "Shuffle input is out of bounds.");
7041 if (BlendMask[Mask[i] % Size] == -1)
7042 BlendMask[Mask[i] % Size] = Mask[i];
7043 else if (BlendMask[Mask[i] % Size] != Mask[i])
7044 return SDValue(); // Can't blend in the needed input!
7046 PermuteMask[i] = Mask[i] % Size;
7049 SDValue V = DAG.getVectorShuffle(VT, DL, V1, V2, BlendMask);
7050 return DAG.getVectorShuffle(VT, DL, V, DAG.getUNDEF(VT), PermuteMask);
7053 /// \brief Generic routine to decompose a shuffle and blend into indepndent
7054 /// blends and permutes.
7056 /// This matches the extremely common pattern for handling combined
7057 /// shuffle+blend operations on newer X86 ISAs where we have very fast blend
7058 /// operations. It will try to pick the best arrangement of shuffles and
7060 static SDValue lowerVectorShuffleAsDecomposedShuffleBlend(SDLoc DL, MVT VT,
7064 SelectionDAG &DAG) {
7065 // Shuffle the input elements into the desired positions in V1 and V2 and
7066 // blend them together.
7067 SmallVector<int, 32> V1Mask(Mask.size(), -1);
7068 SmallVector<int, 32> V2Mask(Mask.size(), -1);
7069 SmallVector<int, 32> BlendMask(Mask.size(), -1);
7070 for (int i = 0, Size = Mask.size(); i < Size; ++i)
7071 if (Mask[i] >= 0 && Mask[i] < Size) {
7072 V1Mask[i] = Mask[i];
7074 } else if (Mask[i] >= Size) {
7075 V2Mask[i] = Mask[i] - Size;
7076 BlendMask[i] = i + Size;
7079 // Try to lower with the simpler initial blend strategy unless one of the
7080 // input shuffles would be a no-op. We prefer to shuffle inputs as the
7081 // shuffle may be able to fold with a load or other benefit. However, when
7082 // we'll have to do 2x as many shuffles in order to achieve this, blending
7083 // first is a better strategy.
7084 if (!isNoopShuffleMask(V1Mask) && !isNoopShuffleMask(V2Mask))
7085 if (SDValue BlendPerm =
7086 lowerVectorShuffleAsBlendAndPermute(DL, VT, V1, V2, Mask, DAG))
7089 V1 = DAG.getVectorShuffle(VT, DL, V1, DAG.getUNDEF(VT), V1Mask);
7090 V2 = DAG.getVectorShuffle(VT, DL, V2, DAG.getUNDEF(VT), V2Mask);
7091 return DAG.getVectorShuffle(VT, DL, V1, V2, BlendMask);
7094 /// \brief Try to lower a vector shuffle as a byte rotation.
7096 /// SSSE3 has a generic PALIGNR instruction in x86 that will do an arbitrary
7097 /// byte-rotation of the concatenation of two vectors; pre-SSSE3 can use
7098 /// a PSRLDQ/PSLLDQ/POR pattern to get a similar effect. This routine will
7099 /// try to generically lower a vector shuffle through such an pattern. It
7100 /// does not check for the profitability of lowering either as PALIGNR or
7101 /// PSRLDQ/PSLLDQ/POR, only whether the mask is valid to lower in that form.
7102 /// This matches shuffle vectors that look like:
7104 /// v8i16 [11, 12, 13, 14, 15, 0, 1, 2]
7106 /// Essentially it concatenates V1 and V2, shifts right by some number of
7107 /// elements, and takes the low elements as the result. Note that while this is
7108 /// specified as a *right shift* because x86 is little-endian, it is a *left
7109 /// rotate* of the vector lanes.
7110 static SDValue lowerVectorShuffleAsByteRotate(SDLoc DL, MVT VT, SDValue V1,
7113 const X86Subtarget *Subtarget,
7114 SelectionDAG &DAG) {
7115 assert(!isNoopShuffleMask(Mask) && "We shouldn't lower no-op shuffles!");
7117 int NumElts = Mask.size();
7118 int NumLanes = VT.getSizeInBits() / 128;
7119 int NumLaneElts = NumElts / NumLanes;
7121 // We need to detect various ways of spelling a rotation:
7122 // [11, 12, 13, 14, 15, 0, 1, 2]
7123 // [-1, 12, 13, 14, -1, -1, 1, -1]
7124 // [-1, -1, -1, -1, -1, -1, 1, 2]
7125 // [ 3, 4, 5, 6, 7, 8, 9, 10]
7126 // [-1, 4, 5, 6, -1, -1, 9, -1]
7127 // [-1, 4, 5, 6, -1, -1, -1, -1]
7130 for (int l = 0; l < NumElts; l += NumLaneElts) {
7131 for (int i = 0; i < NumLaneElts; ++i) {
7132 if (Mask[l + i] == -1)
7134 assert(Mask[l + i] >= 0 && "Only -1 is a valid negative mask element!");
7136 // Get the mod-Size index and lane correct it.
7137 int LaneIdx = (Mask[l + i] % NumElts) - l;
7138 // Make sure it was in this lane.
7139 if (LaneIdx < 0 || LaneIdx >= NumLaneElts)
7142 // Determine where a rotated vector would have started.
7143 int StartIdx = i - LaneIdx;
7145 // The identity rotation isn't interesting, stop.
7148 // If we found the tail of a vector the rotation must be the missing
7149 // front. If we found the head of a vector, it must be how much of the
7151 int CandidateRotation = StartIdx < 0 ? -StartIdx : NumLaneElts - StartIdx;
7154 Rotation = CandidateRotation;
7155 else if (Rotation != CandidateRotation)
7156 // The rotations don't match, so we can't match this mask.
7159 // Compute which value this mask is pointing at.
7160 SDValue MaskV = Mask[l + i] < NumElts ? V1 : V2;
7162 // Compute which of the two target values this index should be assigned
7163 // to. This reflects whether the high elements are remaining or the low
7164 // elements are remaining.
7165 SDValue &TargetV = StartIdx < 0 ? Hi : Lo;
7167 // Either set up this value if we've not encountered it before, or check
7168 // that it remains consistent.
7171 else if (TargetV != MaskV)
7172 // This may be a rotation, but it pulls from the inputs in some
7173 // unsupported interleaving.
7178 // Check that we successfully analyzed the mask, and normalize the results.
7179 assert(Rotation != 0 && "Failed to locate a viable rotation!");
7180 assert((Lo || Hi) && "Failed to find a rotated input vector!");
7186 // The actual rotate instruction rotates bytes, so we need to scale the
7187 // rotation based on how many bytes are in the vector lane.
7188 int Scale = 16 / NumLaneElts;
7190 // SSSE3 targets can use the palignr instruction.
7191 if (Subtarget->hasSSSE3()) {
7192 // Cast the inputs to i8 vector of correct length to match PALIGNR.
7193 MVT AlignVT = MVT::getVectorVT(MVT::i8, 16 * NumLanes);
7194 Lo = DAG.getBitcast(AlignVT, Lo);
7195 Hi = DAG.getBitcast(AlignVT, Hi);
7197 return DAG.getBitcast(
7198 VT, DAG.getNode(X86ISD::PALIGNR, DL, AlignVT, Lo, Hi,
7199 DAG.getConstant(Rotation * Scale, DL, MVT::i8)));
7202 assert(VT.getSizeInBits() == 128 &&
7203 "Rotate-based lowering only supports 128-bit lowering!");
7204 assert(Mask.size() <= 16 &&
7205 "Can shuffle at most 16 bytes in a 128-bit vector!");
7207 // Default SSE2 implementation
7208 int LoByteShift = 16 - Rotation * Scale;
7209 int HiByteShift = Rotation * Scale;
7211 // Cast the inputs to v2i64 to match PSLLDQ/PSRLDQ.
7212 Lo = DAG.getBitcast(MVT::v2i64, Lo);
7213 Hi = DAG.getBitcast(MVT::v2i64, Hi);
7215 SDValue LoShift = DAG.getNode(X86ISD::VSHLDQ, DL, MVT::v2i64, Lo,
7216 DAG.getConstant(LoByteShift, DL, MVT::i8));
7217 SDValue HiShift = DAG.getNode(X86ISD::VSRLDQ, DL, MVT::v2i64, Hi,
7218 DAG.getConstant(HiByteShift, DL, MVT::i8));
7219 return DAG.getBitcast(VT,
7220 DAG.getNode(ISD::OR, DL, MVT::v2i64, LoShift, HiShift));
7223 /// \brief Try to lower a vector shuffle as a bit shift (shifts in zeros).
7225 /// Attempts to match a shuffle mask against the PSLL(W/D/Q/DQ) and
7226 /// PSRL(W/D/Q/DQ) SSE2 and AVX2 logical bit-shift instructions. The function
7227 /// matches elements from one of the input vectors shuffled to the left or
7228 /// right with zeroable elements 'shifted in'. It handles both the strictly
7229 /// bit-wise element shifts and the byte shift across an entire 128-bit double
7232 /// PSHL : (little-endian) left bit shift.
7233 /// [ zz, 0, zz, 2 ]
7234 /// [ -1, 4, zz, -1 ]
7235 /// PSRL : (little-endian) right bit shift.
7237 /// [ -1, -1, 7, zz]
7238 /// PSLLDQ : (little-endian) left byte shift
7239 /// [ zz, 0, 1, 2, 3, 4, 5, 6]
7240 /// [ zz, zz, -1, -1, 2, 3, 4, -1]
7241 /// [ zz, zz, zz, zz, zz, zz, -1, 1]
7242 /// PSRLDQ : (little-endian) right byte shift
7243 /// [ 5, 6, 7, zz, zz, zz, zz, zz]
7244 /// [ -1, 5, 6, 7, zz, zz, zz, zz]
7245 /// [ 1, 2, -1, -1, -1, -1, zz, zz]
7246 static SDValue lowerVectorShuffleAsShift(SDLoc DL, MVT VT, SDValue V1,
7247 SDValue V2, ArrayRef<int> Mask,
7248 SelectionDAG &DAG) {
7249 SmallBitVector Zeroable = computeZeroableShuffleElements(Mask, V1, V2);
7251 int Size = Mask.size();
7252 assert(Size == (int)VT.getVectorNumElements() && "Unexpected mask size");
7254 auto CheckZeros = [&](int Shift, int Scale, bool Left) {
7255 for (int i = 0; i < Size; i += Scale)
7256 for (int j = 0; j < Shift; ++j)
7257 if (!Zeroable[i + j + (Left ? 0 : (Scale - Shift))])
7263 auto MatchShift = [&](int Shift, int Scale, bool Left, SDValue V) {
7264 for (int i = 0; i != Size; i += Scale) {
7265 unsigned Pos = Left ? i + Shift : i;
7266 unsigned Low = Left ? i : i + Shift;
7267 unsigned Len = Scale - Shift;
7268 if (!isSequentialOrUndefInRange(Mask, Pos, Len,
7269 Low + (V == V1 ? 0 : Size)))
7273 int ShiftEltBits = VT.getScalarSizeInBits() * Scale;
7274 bool ByteShift = ShiftEltBits > 64;
7275 unsigned OpCode = Left ? (ByteShift ? X86ISD::VSHLDQ : X86ISD::VSHLI)
7276 : (ByteShift ? X86ISD::VSRLDQ : X86ISD::VSRLI);
7277 int ShiftAmt = Shift * VT.getScalarSizeInBits() / (ByteShift ? 8 : 1);
7279 // Normalize the scale for byte shifts to still produce an i64 element
7281 Scale = ByteShift ? Scale / 2 : Scale;
7283 // We need to round trip through the appropriate type for the shift.
7284 MVT ShiftSVT = MVT::getIntegerVT(VT.getScalarSizeInBits() * Scale);
7285 MVT ShiftVT = MVT::getVectorVT(ShiftSVT, Size / Scale);
7286 assert(DAG.getTargetLoweringInfo().isTypeLegal(ShiftVT) &&
7287 "Illegal integer vector type");
7288 V = DAG.getBitcast(ShiftVT, V);
7290 V = DAG.getNode(OpCode, DL, ShiftVT, V,
7291 DAG.getConstant(ShiftAmt, DL, MVT::i8));
7292 return DAG.getBitcast(VT, V);
7295 // SSE/AVX supports logical shifts up to 64-bit integers - so we can just
7296 // keep doubling the size of the integer elements up to that. We can
7297 // then shift the elements of the integer vector by whole multiples of
7298 // their width within the elements of the larger integer vector. Test each
7299 // multiple to see if we can find a match with the moved element indices
7300 // and that the shifted in elements are all zeroable.
7301 for (int Scale = 2; Scale * VT.getScalarSizeInBits() <= 128; Scale *= 2)
7302 for (int Shift = 1; Shift != Scale; ++Shift)
7303 for (bool Left : {true, false})
7304 if (CheckZeros(Shift, Scale, Left))
7305 for (SDValue V : {V1, V2})
7306 if (SDValue Match = MatchShift(Shift, Scale, Left, V))
7313 /// \brief Try to lower a vector shuffle using SSE4a EXTRQ/INSERTQ.
7314 static SDValue lowerVectorShuffleWithSSE4A(SDLoc DL, MVT VT, SDValue V1,
7315 SDValue V2, ArrayRef<int> Mask,
7316 SelectionDAG &DAG) {
7317 SmallBitVector Zeroable = computeZeroableShuffleElements(Mask, V1, V2);
7318 assert(!Zeroable.all() && "Fully zeroable shuffle mask");
7320 int Size = Mask.size();
7321 int HalfSize = Size / 2;
7322 assert(Size == (int)VT.getVectorNumElements() && "Unexpected mask size");
7324 // Upper half must be undefined.
7325 if (!isUndefInRange(Mask, HalfSize, HalfSize))
7328 // EXTRQ: Extract Len elements from lower half of source, starting at Idx.
7329 // Remainder of lower half result is zero and upper half is all undef.
7330 auto LowerAsEXTRQ = [&]() {
7331 // Determine the extraction length from the part of the
7332 // lower half that isn't zeroable.
7334 for (; Len > 0; --Len)
7335 if (!Zeroable[Len - 1])
7337 assert(Len > 0 && "Zeroable shuffle mask");
7339 // Attempt to match first Len sequential elements from the lower half.
7342 for (int i = 0; i != Len; ++i) {
7346 SDValue &V = (M < Size ? V1 : V2);
7349 // All mask elements must be in the lower half.
7353 if (Idx < 0 || (Src == V && Idx == (M - i))) {
7364 assert((Idx + Len) <= HalfSize && "Illegal extraction mask");
7365 int BitLen = (Len * VT.getScalarSizeInBits()) & 0x3f;
7366 int BitIdx = (Idx * VT.getScalarSizeInBits()) & 0x3f;
7367 return DAG.getNode(X86ISD::EXTRQI, DL, VT, Src,
7368 DAG.getConstant(BitLen, DL, MVT::i8),
7369 DAG.getConstant(BitIdx, DL, MVT::i8));
7372 if (SDValue ExtrQ = LowerAsEXTRQ())
7375 // INSERTQ: Extract lowest Len elements from lower half of second source and
7376 // insert over first source, starting at Idx.
7377 // { A[0], .., A[Idx-1], B[0], .., B[Len-1], A[Idx+Len], .., UNDEF, ... }
7378 auto LowerAsInsertQ = [&]() {
7379 for (int Idx = 0; Idx != HalfSize; ++Idx) {
7382 // Attempt to match first source from mask before insertion point.
7383 if (isUndefInRange(Mask, 0, Idx)) {
7385 } else if (isSequentialOrUndefInRange(Mask, 0, Idx, 0)) {
7387 } else if (isSequentialOrUndefInRange(Mask, 0, Idx, Size)) {
7393 // Extend the extraction length looking to match both the insertion of
7394 // the second source and the remaining elements of the first.
7395 for (int Hi = Idx + 1; Hi <= HalfSize; ++Hi) {
7400 if (isSequentialOrUndefInRange(Mask, Idx, Len, 0)) {
7402 } else if (isSequentialOrUndefInRange(Mask, Idx, Len, Size)) {
7408 // Match the remaining elements of the lower half.
7409 if (isUndefInRange(Mask, Hi, HalfSize - Hi)) {
7411 } else if ((!Base || (Base == V1)) &&
7412 isSequentialOrUndefInRange(Mask, Hi, HalfSize - Hi, Hi)) {
7414 } else if ((!Base || (Base == V2)) &&
7415 isSequentialOrUndefInRange(Mask, Hi, HalfSize - Hi,
7422 // We may not have a base (first source) - this can safely be undefined.
7424 Base = DAG.getUNDEF(VT);
7426 int BitLen = (Len * VT.getScalarSizeInBits()) & 0x3f;
7427 int BitIdx = (Idx * VT.getScalarSizeInBits()) & 0x3f;
7428 return DAG.getNode(X86ISD::INSERTQI, DL, VT, Base, Insert,
7429 DAG.getConstant(BitLen, DL, MVT::i8),
7430 DAG.getConstant(BitIdx, DL, MVT::i8));
7437 if (SDValue InsertQ = LowerAsInsertQ())
7443 /// \brief Lower a vector shuffle as a zero or any extension.
7445 /// Given a specific number of elements, element bit width, and extension
7446 /// stride, produce either a zero or any extension based on the available
7447 /// features of the subtarget. The extended elements are consecutive and
7448 /// begin and can start from an offseted element index in the input; to
7449 /// avoid excess shuffling the offset must either being in the bottom lane
7450 /// or at the start of a higher lane. All extended elements must be from
7452 static SDValue lowerVectorShuffleAsSpecificZeroOrAnyExtend(
7453 SDLoc DL, MVT VT, int Scale, int Offset, bool AnyExt, SDValue InputV,
7454 ArrayRef<int> Mask, const X86Subtarget *Subtarget, SelectionDAG &DAG) {
7455 assert(Scale > 1 && "Need a scale to extend.");
7456 int EltBits = VT.getScalarSizeInBits();
7457 int NumElements = VT.getVectorNumElements();
7458 int NumEltsPerLane = 128 / EltBits;
7459 int OffsetLane = Offset / NumEltsPerLane;
7460 assert((EltBits == 8 || EltBits == 16 || EltBits == 32) &&
7461 "Only 8, 16, and 32 bit elements can be extended.");
7462 assert(Scale * EltBits <= 64 && "Cannot zero extend past 64 bits.");
7463 assert(0 <= Offset && "Extension offset must be positive.");
7464 assert((Offset < NumEltsPerLane || Offset % NumEltsPerLane == 0) &&
7465 "Extension offset must be in the first lane or start an upper lane.");
7467 // Check that an index is in same lane as the base offset.
7468 auto SafeOffset = [&](int Idx) {
7469 return OffsetLane == (Idx / NumEltsPerLane);
7472 // Shift along an input so that the offset base moves to the first element.
7473 auto ShuffleOffset = [&](SDValue V) {
7477 SmallVector<int, 8> ShMask((unsigned)NumElements, -1);
7478 for (int i = 0; i * Scale < NumElements; ++i) {
7479 int SrcIdx = i + Offset;
7480 ShMask[i] = SafeOffset(SrcIdx) ? SrcIdx : -1;
7482 return DAG.getVectorShuffle(VT, DL, V, DAG.getUNDEF(VT), ShMask);
7485 // Found a valid zext mask! Try various lowering strategies based on the
7486 // input type and available ISA extensions.
7487 if (Subtarget->hasSSE41()) {
7488 // Not worth offseting 128-bit vectors if scale == 2, a pattern using
7489 // PUNPCK will catch this in a later shuffle match.
7490 if (Offset && Scale == 2 && VT.getSizeInBits() == 128)
7492 MVT ExtVT = MVT::getVectorVT(MVT::getIntegerVT(EltBits * Scale),
7493 NumElements / Scale);
7494 InputV = DAG.getNode(X86ISD::VZEXT, DL, ExtVT, ShuffleOffset(InputV));
7495 return DAG.getBitcast(VT, InputV);
7498 assert(VT.getSizeInBits() == 128 && "Only 128-bit vectors can be extended.");
7500 // For any extends we can cheat for larger element sizes and use shuffle
7501 // instructions that can fold with a load and/or copy.
7502 if (AnyExt && EltBits == 32) {
7503 int PSHUFDMask[4] = {Offset, -1, SafeOffset(Offset + 1) ? Offset + 1 : -1,
7505 return DAG.getBitcast(
7506 VT, DAG.getNode(X86ISD::PSHUFD, DL, MVT::v4i32,
7507 DAG.getBitcast(MVT::v4i32, InputV),
7508 getV4X86ShuffleImm8ForMask(PSHUFDMask, DL, DAG)));
7510 if (AnyExt && EltBits == 16 && Scale > 2) {
7511 int PSHUFDMask[4] = {Offset / 2, -1,
7512 SafeOffset(Offset + 1) ? (Offset + 1) / 2 : -1, -1};
7513 InputV = DAG.getNode(X86ISD::PSHUFD, DL, MVT::v4i32,
7514 DAG.getBitcast(MVT::v4i32, InputV),
7515 getV4X86ShuffleImm8ForMask(PSHUFDMask, DL, DAG));
7516 int PSHUFWMask[4] = {1, -1, -1, -1};
7517 unsigned OddEvenOp = (Offset & 1 ? X86ISD::PSHUFLW : X86ISD::PSHUFHW);
7518 return DAG.getBitcast(
7519 VT, DAG.getNode(OddEvenOp, DL, MVT::v8i16,
7520 DAG.getBitcast(MVT::v8i16, InputV),
7521 getV4X86ShuffleImm8ForMask(PSHUFWMask, DL, DAG)));
7524 // The SSE4A EXTRQ instruction can efficiently extend the first 2 lanes
7526 if ((Scale * EltBits) == 64 && EltBits < 32 && Subtarget->hasSSE4A()) {
7527 assert(NumElements == (int)Mask.size() && "Unexpected shuffle mask size!");
7528 assert(VT.getSizeInBits() == 128 && "Unexpected vector width!");
7530 int LoIdx = Offset * EltBits;
7531 SDValue Lo = DAG.getNode(ISD::BITCAST, DL, MVT::v2i64,
7532 DAG.getNode(X86ISD::EXTRQI, DL, VT, InputV,
7533 DAG.getConstant(EltBits, DL, MVT::i8),
7534 DAG.getConstant(LoIdx, DL, MVT::i8)));
7536 if (isUndefInRange(Mask, NumElements / 2, NumElements / 2) ||
7537 !SafeOffset(Offset + 1))
7538 return DAG.getNode(ISD::BITCAST, DL, VT, Lo);
7540 int HiIdx = (Offset + 1) * EltBits;
7541 SDValue Hi = DAG.getNode(ISD::BITCAST, DL, MVT::v2i64,
7542 DAG.getNode(X86ISD::EXTRQI, DL, VT, InputV,
7543 DAG.getConstant(EltBits, DL, MVT::i8),
7544 DAG.getConstant(HiIdx, DL, MVT::i8)));
7545 return DAG.getNode(ISD::BITCAST, DL, VT,
7546 DAG.getNode(X86ISD::UNPCKL, DL, MVT::v2i64, Lo, Hi));
7549 // If this would require more than 2 unpack instructions to expand, use
7550 // pshufb when available. We can only use more than 2 unpack instructions
7551 // when zero extending i8 elements which also makes it easier to use pshufb.
7552 if (Scale > 4 && EltBits == 8 && Subtarget->hasSSSE3()) {
7553 assert(NumElements == 16 && "Unexpected byte vector width!");
7554 SDValue PSHUFBMask[16];
7555 for (int i = 0; i < 16; ++i) {
7556 int Idx = Offset + (i / Scale);
7557 PSHUFBMask[i] = DAG.getConstant(
7558 (i % Scale == 0 && SafeOffset(Idx)) ? Idx : 0x80, DL, MVT::i8);
7560 InputV = DAG.getBitcast(MVT::v16i8, InputV);
7561 return DAG.getBitcast(VT,
7562 DAG.getNode(X86ISD::PSHUFB, DL, MVT::v16i8, InputV,
7563 DAG.getNode(ISD::BUILD_VECTOR, DL,
7564 MVT::v16i8, PSHUFBMask)));
7567 // If we are extending from an offset, ensure we start on a boundary that
7568 // we can unpack from.
7569 int AlignToUnpack = Offset % (NumElements / Scale);
7570 if (AlignToUnpack) {
7571 SmallVector<int, 8> ShMask((unsigned)NumElements, -1);
7572 for (int i = AlignToUnpack; i < NumElements; ++i)
7573 ShMask[i - AlignToUnpack] = i;
7574 InputV = DAG.getVectorShuffle(VT, DL, InputV, DAG.getUNDEF(VT), ShMask);
7575 Offset -= AlignToUnpack;
7578 // Otherwise emit a sequence of unpacks.
7580 unsigned UnpackLoHi = X86ISD::UNPCKL;
7581 if (Offset >= (NumElements / 2)) {
7582 UnpackLoHi = X86ISD::UNPCKH;
7583 Offset -= (NumElements / 2);
7586 MVT InputVT = MVT::getVectorVT(MVT::getIntegerVT(EltBits), NumElements);
7587 SDValue Ext = AnyExt ? DAG.getUNDEF(InputVT)
7588 : getZeroVector(InputVT, Subtarget, DAG, DL);
7589 InputV = DAG.getBitcast(InputVT, InputV);
7590 InputV = DAG.getNode(UnpackLoHi, DL, InputVT, InputV, Ext);
7594 } while (Scale > 1);
7595 return DAG.getBitcast(VT, InputV);
7598 /// \brief Try to lower a vector shuffle as a zero extension on any microarch.
7600 /// This routine will try to do everything in its power to cleverly lower
7601 /// a shuffle which happens to match the pattern of a zero extend. It doesn't
7602 /// check for the profitability of this lowering, it tries to aggressively
7603 /// match this pattern. It will use all of the micro-architectural details it
7604 /// can to emit an efficient lowering. It handles both blends with all-zero
7605 /// inputs to explicitly zero-extend and undef-lanes (sometimes undef due to
7606 /// masking out later).
7608 /// The reason we have dedicated lowering for zext-style shuffles is that they
7609 /// are both incredibly common and often quite performance sensitive.
7610 static SDValue lowerVectorShuffleAsZeroOrAnyExtend(
7611 SDLoc DL, MVT VT, SDValue V1, SDValue V2, ArrayRef<int> Mask,
7612 const X86Subtarget *Subtarget, SelectionDAG &DAG) {
7613 SmallBitVector Zeroable = computeZeroableShuffleElements(Mask, V1, V2);
7615 int Bits = VT.getSizeInBits();
7616 int NumLanes = Bits / 128;
7617 int NumElements = VT.getVectorNumElements();
7618 int NumEltsPerLane = NumElements / NumLanes;
7619 assert(VT.getScalarSizeInBits() <= 32 &&
7620 "Exceeds 32-bit integer zero extension limit");
7621 assert((int)Mask.size() == NumElements && "Unexpected shuffle mask size");
7623 // Define a helper function to check a particular ext-scale and lower to it if
7625 auto Lower = [&](int Scale) -> SDValue {
7630 for (int i = 0; i < NumElements; ++i) {
7633 continue; // Valid anywhere but doesn't tell us anything.
7634 if (i % Scale != 0) {
7635 // Each of the extended elements need to be zeroable.
7639 // We no longer are in the anyext case.
7644 // Each of the base elements needs to be consecutive indices into the
7645 // same input vector.
7646 SDValue V = M < NumElements ? V1 : V2;
7647 M = M % NumElements;
7650 Offset = M - (i / Scale);
7651 } else if (InputV != V)
7652 return SDValue(); // Flip-flopping inputs.
7654 // Offset must start in the lowest 128-bit lane or at the start of an
7656 // FIXME: Is it ever worth allowing a negative base offset?
7657 if (!((0 <= Offset && Offset < NumEltsPerLane) ||
7658 (Offset % NumEltsPerLane) == 0))
7661 // If we are offsetting, all referenced entries must come from the same
7663 if (Offset && (Offset / NumEltsPerLane) != (M / NumEltsPerLane))
7666 if ((M % NumElements) != (Offset + (i / Scale)))
7667 return SDValue(); // Non-consecutive strided elements.
7671 // If we fail to find an input, we have a zero-shuffle which should always
7672 // have already been handled.
7673 // FIXME: Maybe handle this here in case during blending we end up with one?
7677 // If we are offsetting, don't extend if we only match a single input, we
7678 // can always do better by using a basic PSHUF or PUNPCK.
7679 if (Offset != 0 && Matches < 2)
7682 return lowerVectorShuffleAsSpecificZeroOrAnyExtend(
7683 DL, VT, Scale, Offset, AnyExt, InputV, Mask, Subtarget, DAG);
7686 // The widest scale possible for extending is to a 64-bit integer.
7687 assert(Bits % 64 == 0 &&
7688 "The number of bits in a vector must be divisible by 64 on x86!");
7689 int NumExtElements = Bits / 64;
7691 // Each iteration, try extending the elements half as much, but into twice as
7693 for (; NumExtElements < NumElements; NumExtElements *= 2) {
7694 assert(NumElements % NumExtElements == 0 &&
7695 "The input vector size must be divisible by the extended size.");
7696 if (SDValue V = Lower(NumElements / NumExtElements))
7700 // General extends failed, but 128-bit vectors may be able to use MOVQ.
7704 // Returns one of the source operands if the shuffle can be reduced to a
7705 // MOVQ, copying the lower 64-bits and zero-extending to the upper 64-bits.
7706 auto CanZExtLowHalf = [&]() {
7707 for (int i = NumElements / 2; i != NumElements; ++i)
7710 if (isSequentialOrUndefInRange(Mask, 0, NumElements / 2, 0))
7712 if (isSequentialOrUndefInRange(Mask, 0, NumElements / 2, NumElements))
7717 if (SDValue V = CanZExtLowHalf()) {
7718 V = DAG.getBitcast(MVT::v2i64, V);
7719 V = DAG.getNode(X86ISD::VZEXT_MOVL, DL, MVT::v2i64, V);
7720 return DAG.getBitcast(VT, V);
7723 // No viable ext lowering found.
7727 /// \brief Try to get a scalar value for a specific element of a vector.
7729 /// Looks through BUILD_VECTOR and SCALAR_TO_VECTOR nodes to find a scalar.
7730 static SDValue getScalarValueForVectorElement(SDValue V, int Idx,
7731 SelectionDAG &DAG) {
7732 MVT VT = V.getSimpleValueType();
7733 MVT EltVT = VT.getVectorElementType();
7734 while (V.getOpcode() == ISD::BITCAST)
7735 V = V.getOperand(0);
7736 // If the bitcasts shift the element size, we can't extract an equivalent
7738 MVT NewVT = V.getSimpleValueType();
7739 if (!NewVT.isVector() || NewVT.getScalarSizeInBits() != VT.getScalarSizeInBits())
7742 if (V.getOpcode() == ISD::BUILD_VECTOR ||
7743 (Idx == 0 && V.getOpcode() == ISD::SCALAR_TO_VECTOR)) {
7744 // Ensure the scalar operand is the same size as the destination.
7745 // FIXME: Add support for scalar truncation where possible.
7746 SDValue S = V.getOperand(Idx);
7747 if (EltVT.getSizeInBits() == S.getSimpleValueType().getSizeInBits())
7748 return DAG.getNode(ISD::BITCAST, SDLoc(V), EltVT, S);
7754 /// \brief Helper to test for a load that can be folded with x86 shuffles.
7756 /// This is particularly important because the set of instructions varies
7757 /// significantly based on whether the operand is a load or not.
7758 static bool isShuffleFoldableLoad(SDValue V) {
7759 while (V.getOpcode() == ISD::BITCAST)
7760 V = V.getOperand(0);
7762 return ISD::isNON_EXTLoad(V.getNode());
7765 /// \brief Try to lower insertion of a single element into a zero vector.
7767 /// This is a common pattern that we have especially efficient patterns to lower
7768 /// across all subtarget feature sets.
7769 static SDValue lowerVectorShuffleAsElementInsertion(
7770 SDLoc DL, MVT VT, SDValue V1, SDValue V2, ArrayRef<int> Mask,
7771 const X86Subtarget *Subtarget, SelectionDAG &DAG) {
7772 SmallBitVector Zeroable = computeZeroableShuffleElements(Mask, V1, V2);
7774 MVT EltVT = VT.getVectorElementType();
7776 int V2Index = std::find_if(Mask.begin(), Mask.end(),
7777 [&Mask](int M) { return M >= (int)Mask.size(); }) -
7779 bool IsV1Zeroable = true;
7780 for (int i = 0, Size = Mask.size(); i < Size; ++i)
7781 if (i != V2Index && !Zeroable[i]) {
7782 IsV1Zeroable = false;
7786 // Check for a single input from a SCALAR_TO_VECTOR node.
7787 // FIXME: All of this should be canonicalized into INSERT_VECTOR_ELT and
7788 // all the smarts here sunk into that routine. However, the current
7789 // lowering of BUILD_VECTOR makes that nearly impossible until the old
7790 // vector shuffle lowering is dead.
7791 SDValue V2S = getScalarValueForVectorElement(V2, Mask[V2Index] - Mask.size(),
7793 if (V2S && DAG.getTargetLoweringInfo().isTypeLegal(V2S.getValueType())) {
7794 // We need to zext the scalar if it is smaller than an i32.
7795 V2S = DAG.getBitcast(EltVT, V2S);
7796 if (EltVT == MVT::i8 || EltVT == MVT::i16) {
7797 // Using zext to expand a narrow element won't work for non-zero
7802 // Zero-extend directly to i32.
7804 V2S = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::i32, V2S);
7806 V2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, ExtVT, V2S);
7807 } else if (Mask[V2Index] != (int)Mask.size() || EltVT == MVT::i8 ||
7808 EltVT == MVT::i16) {
7809 // Either not inserting from the low element of the input or the input
7810 // element size is too small to use VZEXT_MOVL to clear the high bits.
7814 if (!IsV1Zeroable) {
7815 // If V1 can't be treated as a zero vector we have fewer options to lower
7816 // this. We can't support integer vectors or non-zero targets cheaply, and
7817 // the V1 elements can't be permuted in any way.
7818 assert(VT == ExtVT && "Cannot change extended type when non-zeroable!");
7819 if (!VT.isFloatingPoint() || V2Index != 0)
7821 SmallVector<int, 8> V1Mask(Mask.begin(), Mask.end());
7822 V1Mask[V2Index] = -1;
7823 if (!isNoopShuffleMask(V1Mask))
7825 // This is essentially a special case blend operation, but if we have
7826 // general purpose blend operations, they are always faster. Bail and let
7827 // the rest of the lowering handle these as blends.
7828 if (Subtarget->hasSSE41())
7831 // Otherwise, use MOVSD or MOVSS.
7832 assert((EltVT == MVT::f32 || EltVT == MVT::f64) &&
7833 "Only two types of floating point element types to handle!");
7834 return DAG.getNode(EltVT == MVT::f32 ? X86ISD::MOVSS : X86ISD::MOVSD, DL,
7838 // This lowering only works for the low element with floating point vectors.
7839 if (VT.isFloatingPoint() && V2Index != 0)
7842 V2 = DAG.getNode(X86ISD::VZEXT_MOVL, DL, ExtVT, V2);
7844 V2 = DAG.getBitcast(VT, V2);
7847 // If we have 4 or fewer lanes we can cheaply shuffle the element into
7848 // the desired position. Otherwise it is more efficient to do a vector
7849 // shift left. We know that we can do a vector shift left because all
7850 // the inputs are zero.
7851 if (VT.isFloatingPoint() || VT.getVectorNumElements() <= 4) {
7852 SmallVector<int, 4> V2Shuffle(Mask.size(), 1);
7853 V2Shuffle[V2Index] = 0;
7854 V2 = DAG.getVectorShuffle(VT, DL, V2, DAG.getUNDEF(VT), V2Shuffle);
7856 V2 = DAG.getBitcast(MVT::v2i64, V2);
7858 X86ISD::VSHLDQ, DL, MVT::v2i64, V2,
7859 DAG.getConstant(V2Index * EltVT.getSizeInBits() / 8, DL,
7860 DAG.getTargetLoweringInfo().getScalarShiftAmountTy(
7861 DAG.getDataLayout(), VT)));
7862 V2 = DAG.getBitcast(VT, V2);
7868 /// \brief Try to lower broadcast of a single element.
7870 /// For convenience, this code also bundles all of the subtarget feature set
7871 /// filtering. While a little annoying to re-dispatch on type here, there isn't
7872 /// a convenient way to factor it out.
7873 static SDValue lowerVectorShuffleAsBroadcast(SDLoc DL, MVT VT, SDValue V,
7875 const X86Subtarget *Subtarget,
7876 SelectionDAG &DAG) {
7877 if (!Subtarget->hasAVX())
7879 if (VT.isInteger() && !Subtarget->hasAVX2())
7882 // Check that the mask is a broadcast.
7883 int BroadcastIdx = -1;
7885 if (M >= 0 && BroadcastIdx == -1)
7887 else if (M >= 0 && M != BroadcastIdx)
7890 assert(BroadcastIdx < (int)Mask.size() && "We only expect to be called with "
7891 "a sorted mask where the broadcast "
7894 // Go up the chain of (vector) values to find a scalar load that we can
7895 // combine with the broadcast.
7897 switch (V.getOpcode()) {
7898 case ISD::CONCAT_VECTORS: {
7899 int OperandSize = Mask.size() / V.getNumOperands();
7900 V = V.getOperand(BroadcastIdx / OperandSize);
7901 BroadcastIdx %= OperandSize;
7905 case ISD::INSERT_SUBVECTOR: {
7906 SDValue VOuter = V.getOperand(0), VInner = V.getOperand(1);
7907 auto ConstantIdx = dyn_cast<ConstantSDNode>(V.getOperand(2));
7911 int BeginIdx = (int)ConstantIdx->getZExtValue();
7913 BeginIdx + (int)VInner.getValueType().getVectorNumElements();
7914 if (BroadcastIdx >= BeginIdx && BroadcastIdx < EndIdx) {
7915 BroadcastIdx -= BeginIdx;
7926 // Check if this is a broadcast of a scalar. We special case lowering
7927 // for scalars so that we can more effectively fold with loads.
7928 // First, look through bitcast: if the original value has a larger element
7929 // type than the shuffle, the broadcast element is in essence truncated.
7930 // Make that explicit to ease folding.
7931 if (V.getOpcode() == ISD::BITCAST && VT.isInteger()) {
7932 EVT EltVT = VT.getVectorElementType();
7933 SDValue V0 = V.getOperand(0);
7934 EVT V0VT = V0.getValueType();
7936 if (V0VT.isInteger() && V0VT.getVectorElementType().bitsGT(EltVT) &&
7937 ((V0.getOpcode() == ISD::BUILD_VECTOR ||
7938 (V0.getOpcode() == ISD::SCALAR_TO_VECTOR && BroadcastIdx == 0)))) {
7939 V = DAG.getNode(ISD::TRUNCATE, DL, EltVT, V0.getOperand(BroadcastIdx));
7944 // Also check the simpler case, where we can directly reuse the scalar.
7945 if (V.getOpcode() == ISD::BUILD_VECTOR ||
7946 (V.getOpcode() == ISD::SCALAR_TO_VECTOR && BroadcastIdx == 0)) {
7947 V = V.getOperand(BroadcastIdx);
7949 // If the scalar isn't a load, we can't broadcast from it in AVX1.
7950 // Only AVX2 has register broadcasts.
7951 if (!Subtarget->hasAVX2() && !isShuffleFoldableLoad(V))
7953 } else if (BroadcastIdx != 0 || !Subtarget->hasAVX2()) {
7954 // We can't broadcast from a vector register without AVX2, and we can only
7955 // broadcast from the zero-element of a vector register.
7959 return DAG.getNode(X86ISD::VBROADCAST, DL, VT, V);
7962 // Check for whether we can use INSERTPS to perform the shuffle. We only use
7963 // INSERTPS when the V1 elements are already in the correct locations
7964 // because otherwise we can just always use two SHUFPS instructions which
7965 // are much smaller to encode than a SHUFPS and an INSERTPS. We can also
7966 // perform INSERTPS if a single V1 element is out of place and all V2
7967 // elements are zeroable.
7968 static SDValue lowerVectorShuffleAsInsertPS(SDValue Op, SDValue V1, SDValue V2,
7970 SelectionDAG &DAG) {
7971 assert(Op.getSimpleValueType() == MVT::v4f32 && "Bad shuffle type!");
7972 assert(V1.getSimpleValueType() == MVT::v4f32 && "Bad operand type!");
7973 assert(V2.getSimpleValueType() == MVT::v4f32 && "Bad operand type!");
7974 assert(Mask.size() == 4 && "Unexpected mask size for v4 shuffle!");
7976 SmallBitVector Zeroable = computeZeroableShuffleElements(Mask, V1, V2);
7979 int V1DstIndex = -1;
7980 int V2DstIndex = -1;
7981 bool V1UsedInPlace = false;
7983 for (int i = 0; i < 4; ++i) {
7984 // Synthesize a zero mask from the zeroable elements (includes undefs).
7990 // Flag if we use any V1 inputs in place.
7992 V1UsedInPlace = true;
7996 // We can only insert a single non-zeroable element.
7997 if (V1DstIndex != -1 || V2DstIndex != -1)
8001 // V1 input out of place for insertion.
8004 // V2 input for insertion.
8009 // Don't bother if we have no (non-zeroable) element for insertion.
8010 if (V1DstIndex == -1 && V2DstIndex == -1)
8013 // Determine element insertion src/dst indices. The src index is from the
8014 // start of the inserted vector, not the start of the concatenated vector.
8015 unsigned V2SrcIndex = 0;
8016 if (V1DstIndex != -1) {
8017 // If we have a V1 input out of place, we use V1 as the V2 element insertion
8018 // and don't use the original V2 at all.
8019 V2SrcIndex = Mask[V1DstIndex];
8020 V2DstIndex = V1DstIndex;
8023 V2SrcIndex = Mask[V2DstIndex] - 4;
8026 // If no V1 inputs are used in place, then the result is created only from
8027 // the zero mask and the V2 insertion - so remove V1 dependency.
8029 V1 = DAG.getUNDEF(MVT::v4f32);
8031 unsigned InsertPSMask = V2SrcIndex << 6 | V2DstIndex << 4 | ZMask;
8032 assert((InsertPSMask & ~0xFFu) == 0 && "Invalid mask!");
8034 // Insert the V2 element into the desired position.
8036 return DAG.getNode(X86ISD::INSERTPS, DL, MVT::v4f32, V1, V2,
8037 DAG.getConstant(InsertPSMask, DL, MVT::i8));
8040 /// \brief Try to lower a shuffle as a permute of the inputs followed by an
8041 /// UNPCK instruction.
8043 /// This specifically targets cases where we end up with alternating between
8044 /// the two inputs, and so can permute them into something that feeds a single
8045 /// UNPCK instruction. Note that this routine only targets integer vectors
8046 /// because for floating point vectors we have a generalized SHUFPS lowering
8047 /// strategy that handles everything that doesn't *exactly* match an unpack,
8048 /// making this clever lowering unnecessary.
8049 static SDValue lowerVectorShuffleAsPermuteAndUnpack(SDLoc DL, MVT VT,
8050 SDValue V1, SDValue V2,
8052 SelectionDAG &DAG) {
8053 assert(!VT.isFloatingPoint() &&
8054 "This routine only supports integer vectors.");
8055 assert(!isSingleInputShuffleMask(Mask) &&
8056 "This routine should only be used when blending two inputs.");
8057 assert(Mask.size() >= 2 && "Single element masks are invalid.");
8059 int Size = Mask.size();
8061 int NumLoInputs = std::count_if(Mask.begin(), Mask.end(), [Size](int M) {
8062 return M >= 0 && M % Size < Size / 2;
8064 int NumHiInputs = std::count_if(
8065 Mask.begin(), Mask.end(), [Size](int M) { return M % Size >= Size / 2; });
8067 bool UnpackLo = NumLoInputs >= NumHiInputs;
8069 auto TryUnpack = [&](MVT UnpackVT, int Scale) {
8070 SmallVector<int, 32> V1Mask(Mask.size(), -1);
8071 SmallVector<int, 32> V2Mask(Mask.size(), -1);
8073 for (int i = 0; i < Size; ++i) {
8077 // Each element of the unpack contains Scale elements from this mask.
8078 int UnpackIdx = i / Scale;
8080 // We only handle the case where V1 feeds the first slots of the unpack.
8081 // We rely on canonicalization to ensure this is the case.
8082 if ((UnpackIdx % 2 == 0) != (Mask[i] < Size))
8085 // Setup the mask for this input. The indexing is tricky as we have to
8086 // handle the unpack stride.
8087 SmallVectorImpl<int> &VMask = (UnpackIdx % 2 == 0) ? V1Mask : V2Mask;
8088 VMask[(UnpackIdx / 2) * Scale + i % Scale + (UnpackLo ? 0 : Size / 2)] =
8092 // If we will have to shuffle both inputs to use the unpack, check whether
8093 // we can just unpack first and shuffle the result. If so, skip this unpack.
8094 if ((NumLoInputs == 0 || NumHiInputs == 0) && !isNoopShuffleMask(V1Mask) &&
8095 !isNoopShuffleMask(V2Mask))
8098 // Shuffle the inputs into place.
8099 V1 = DAG.getVectorShuffle(VT, DL, V1, DAG.getUNDEF(VT), V1Mask);
8100 V2 = DAG.getVectorShuffle(VT, DL, V2, DAG.getUNDEF(VT), V2Mask);
8102 // Cast the inputs to the type we will use to unpack them.
8103 V1 = DAG.getBitcast(UnpackVT, V1);
8104 V2 = DAG.getBitcast(UnpackVT, V2);
8106 // Unpack the inputs and cast the result back to the desired type.
8107 return DAG.getBitcast(
8108 VT, DAG.getNode(UnpackLo ? X86ISD::UNPCKL : X86ISD::UNPCKH, DL,
8112 // We try each unpack from the largest to the smallest to try and find one
8113 // that fits this mask.
8114 int OrigNumElements = VT.getVectorNumElements();
8115 int OrigScalarSize = VT.getScalarSizeInBits();
8116 for (int ScalarSize = 64; ScalarSize >= OrigScalarSize; ScalarSize /= 2) {
8117 int Scale = ScalarSize / OrigScalarSize;
8118 int NumElements = OrigNumElements / Scale;
8119 MVT UnpackVT = MVT::getVectorVT(MVT::getIntegerVT(ScalarSize), NumElements);
8120 if (SDValue Unpack = TryUnpack(UnpackVT, Scale))
8124 // If none of the unpack-rooted lowerings worked (or were profitable) try an
8126 if (NumLoInputs == 0 || NumHiInputs == 0) {
8127 assert((NumLoInputs > 0 || NumHiInputs > 0) &&
8128 "We have to have *some* inputs!");
8129 int HalfOffset = NumLoInputs == 0 ? Size / 2 : 0;
8131 // FIXME: We could consider the total complexity of the permute of each
8132 // possible unpacking. Or at the least we should consider how many
8133 // half-crossings are created.
8134 // FIXME: We could consider commuting the unpacks.
8136 SmallVector<int, 32> PermMask;
8137 PermMask.assign(Size, -1);
8138 for (int i = 0; i < Size; ++i) {
8142 assert(Mask[i] % Size >= HalfOffset && "Found input from wrong half!");
8145 2 * ((Mask[i] % Size) - HalfOffset) + (Mask[i] < Size ? 0 : 1);
8147 return DAG.getVectorShuffle(
8148 VT, DL, DAG.getNode(NumLoInputs == 0 ? X86ISD::UNPCKH : X86ISD::UNPCKL,
8150 DAG.getUNDEF(VT), PermMask);
8156 /// \brief Handle lowering of 2-lane 64-bit floating point shuffles.
8158 /// This is the basis function for the 2-lane 64-bit shuffles as we have full
8159 /// support for floating point shuffles but not integer shuffles. These
8160 /// instructions will incur a domain crossing penalty on some chips though so
8161 /// it is better to avoid lowering through this for integer vectors where
8163 static SDValue lowerV2F64VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
8164 const X86Subtarget *Subtarget,
8165 SelectionDAG &DAG) {
8167 assert(Op.getSimpleValueType() == MVT::v2f64 && "Bad shuffle type!");
8168 assert(V1.getSimpleValueType() == MVT::v2f64 && "Bad operand type!");
8169 assert(V2.getSimpleValueType() == MVT::v2f64 && "Bad operand type!");
8170 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
8171 ArrayRef<int> Mask = SVOp->getMask();
8172 assert(Mask.size() == 2 && "Unexpected mask size for v2 shuffle!");
8174 if (isSingleInputShuffleMask(Mask)) {
8175 // Use low duplicate instructions for masks that match their pattern.
8176 if (Subtarget->hasSSE3())
8177 if (isShuffleEquivalent(V1, V2, Mask, {0, 0}))
8178 return DAG.getNode(X86ISD::MOVDDUP, DL, MVT::v2f64, V1);
8180 // Straight shuffle of a single input vector. Simulate this by using the
8181 // single input as both of the "inputs" to this instruction..
8182 unsigned SHUFPDMask = (Mask[0] == 1) | ((Mask[1] == 1) << 1);
8184 if (Subtarget->hasAVX()) {
8185 // If we have AVX, we can use VPERMILPS which will allow folding a load
8186 // into the shuffle.
8187 return DAG.getNode(X86ISD::VPERMILPI, DL, MVT::v2f64, V1,
8188 DAG.getConstant(SHUFPDMask, DL, MVT::i8));
8191 return DAG.getNode(X86ISD::SHUFP, DL, MVT::v2f64, V1, V1,
8192 DAG.getConstant(SHUFPDMask, DL, MVT::i8));
8194 assert(Mask[0] >= 0 && Mask[0] < 2 && "Non-canonicalized blend!");
8195 assert(Mask[1] >= 2 && "Non-canonicalized blend!");
8197 // If we have a single input, insert that into V1 if we can do so cheaply.
8198 if ((Mask[0] >= 2) + (Mask[1] >= 2) == 1) {
8199 if (SDValue Insertion = lowerVectorShuffleAsElementInsertion(
8200 DL, MVT::v2f64, V1, V2, Mask, Subtarget, DAG))
8202 // Try inverting the insertion since for v2 masks it is easy to do and we
8203 // can't reliably sort the mask one way or the other.
8204 int InverseMask[2] = {Mask[0] < 0 ? -1 : (Mask[0] ^ 2),
8205 Mask[1] < 0 ? -1 : (Mask[1] ^ 2)};
8206 if (SDValue Insertion = lowerVectorShuffleAsElementInsertion(
8207 DL, MVT::v2f64, V2, V1, InverseMask, Subtarget, DAG))
8211 // Try to use one of the special instruction patterns to handle two common
8212 // blend patterns if a zero-blend above didn't work.
8213 if (isShuffleEquivalent(V1, V2, Mask, {0, 3}) ||
8214 isShuffleEquivalent(V1, V2, Mask, {1, 3}))
8215 if (SDValue V1S = getScalarValueForVectorElement(V1, Mask[0], DAG))
8216 // We can either use a special instruction to load over the low double or
8217 // to move just the low double.
8219 isShuffleFoldableLoad(V1S) ? X86ISD::MOVLPD : X86ISD::MOVSD,
8221 DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, MVT::v2f64, V1S));
8223 if (Subtarget->hasSSE41())
8224 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v2f64, V1, V2, Mask,
8228 // Use dedicated unpack instructions for masks that match their pattern.
8230 lowerVectorShuffleWithUNPCK(DL, MVT::v2f64, Mask, V1, V2, DAG))
8233 unsigned SHUFPDMask = (Mask[0] == 1) | (((Mask[1] - 2) == 1) << 1);
8234 return DAG.getNode(X86ISD::SHUFP, DL, MVT::v2f64, V1, V2,
8235 DAG.getConstant(SHUFPDMask, DL, MVT::i8));
8238 /// \brief Handle lowering of 2-lane 64-bit integer shuffles.
8240 /// Tries to lower a 2-lane 64-bit shuffle using shuffle operations provided by
8241 /// the integer unit to minimize domain crossing penalties. However, for blends
8242 /// it falls back to the floating point shuffle operation with appropriate bit
8244 static SDValue lowerV2I64VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
8245 const X86Subtarget *Subtarget,
8246 SelectionDAG &DAG) {
8248 assert(Op.getSimpleValueType() == MVT::v2i64 && "Bad shuffle type!");
8249 assert(V1.getSimpleValueType() == MVT::v2i64 && "Bad operand type!");
8250 assert(V2.getSimpleValueType() == MVT::v2i64 && "Bad operand type!");
8251 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
8252 ArrayRef<int> Mask = SVOp->getMask();
8253 assert(Mask.size() == 2 && "Unexpected mask size for v2 shuffle!");
8255 if (isSingleInputShuffleMask(Mask)) {
8256 // Check for being able to broadcast a single element.
8257 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v2i64, V1,
8258 Mask, Subtarget, DAG))
8261 // Straight shuffle of a single input vector. For everything from SSE2
8262 // onward this has a single fast instruction with no scary immediates.
8263 // We have to map the mask as it is actually a v4i32 shuffle instruction.
8264 V1 = DAG.getBitcast(MVT::v4i32, V1);
8265 int WidenedMask[4] = {
8266 std::max(Mask[0], 0) * 2, std::max(Mask[0], 0) * 2 + 1,
8267 std::max(Mask[1], 0) * 2, std::max(Mask[1], 0) * 2 + 1};
8268 return DAG.getBitcast(
8270 DAG.getNode(X86ISD::PSHUFD, DL, MVT::v4i32, V1,
8271 getV4X86ShuffleImm8ForMask(WidenedMask, DL, DAG)));
8273 assert(Mask[0] != -1 && "No undef lanes in multi-input v2 shuffles!");
8274 assert(Mask[1] != -1 && "No undef lanes in multi-input v2 shuffles!");
8275 assert(Mask[0] < 2 && "We sort V1 to be the first input.");
8276 assert(Mask[1] >= 2 && "We sort V2 to be the second input.");
8278 // If we have a blend of two PACKUS operations an the blend aligns with the
8279 // low and half halves, we can just merge the PACKUS operations. This is
8280 // particularly important as it lets us merge shuffles that this routine itself
8282 auto GetPackNode = [](SDValue V) {
8283 while (V.getOpcode() == ISD::BITCAST)
8284 V = V.getOperand(0);
8286 return V.getOpcode() == X86ISD::PACKUS ? V : SDValue();
8288 if (SDValue V1Pack = GetPackNode(V1))
8289 if (SDValue V2Pack = GetPackNode(V2))
8290 return DAG.getBitcast(MVT::v2i64,
8291 DAG.getNode(X86ISD::PACKUS, DL, MVT::v16i8,
8292 Mask[0] == 0 ? V1Pack.getOperand(0)
8293 : V1Pack.getOperand(1),
8294 Mask[1] == 2 ? V2Pack.getOperand(0)
8295 : V2Pack.getOperand(1)));
8297 // Try to use shift instructions.
8299 lowerVectorShuffleAsShift(DL, MVT::v2i64, V1, V2, Mask, DAG))
8302 // When loading a scalar and then shuffling it into a vector we can often do
8303 // the insertion cheaply.
8304 if (SDValue Insertion = lowerVectorShuffleAsElementInsertion(
8305 DL, MVT::v2i64, V1, V2, Mask, Subtarget, DAG))
8307 // Try inverting the insertion since for v2 masks it is easy to do and we
8308 // can't reliably sort the mask one way or the other.
8309 int InverseMask[2] = {Mask[0] ^ 2, Mask[1] ^ 2};
8310 if (SDValue Insertion = lowerVectorShuffleAsElementInsertion(
8311 DL, MVT::v2i64, V2, V1, InverseMask, Subtarget, DAG))
8314 // We have different paths for blend lowering, but they all must use the
8315 // *exact* same predicate.
8316 bool IsBlendSupported = Subtarget->hasSSE41();
8317 if (IsBlendSupported)
8318 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v2i64, V1, V2, Mask,
8322 // Use dedicated unpack instructions for masks that match their pattern.
8324 lowerVectorShuffleWithUNPCK(DL, MVT::v2i64, Mask, V1, V2, DAG))
8327 // Try to use byte rotation instructions.
8328 // Its more profitable for pre-SSSE3 to use shuffles/unpacks.
8329 if (Subtarget->hasSSSE3())
8330 if (SDValue Rotate = lowerVectorShuffleAsByteRotate(
8331 DL, MVT::v2i64, V1, V2, Mask, Subtarget, DAG))
8334 // If we have direct support for blends, we should lower by decomposing into
8335 // a permute. That will be faster than the domain cross.
8336 if (IsBlendSupported)
8337 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v2i64, V1, V2,
8340 // We implement this with SHUFPD which is pretty lame because it will likely
8341 // incur 2 cycles of stall for integer vectors on Nehalem and older chips.
8342 // However, all the alternatives are still more cycles and newer chips don't
8343 // have this problem. It would be really nice if x86 had better shuffles here.
8344 V1 = DAG.getBitcast(MVT::v2f64, V1);
8345 V2 = DAG.getBitcast(MVT::v2f64, V2);
8346 return DAG.getBitcast(MVT::v2i64,
8347 DAG.getVectorShuffle(MVT::v2f64, DL, V1, V2, Mask));
8350 /// \brief Test whether this can be lowered with a single SHUFPS instruction.
8352 /// This is used to disable more specialized lowerings when the shufps lowering
8353 /// will happen to be efficient.
8354 static bool isSingleSHUFPSMask(ArrayRef<int> Mask) {
8355 // This routine only handles 128-bit shufps.
8356 assert(Mask.size() == 4 && "Unsupported mask size!");
8358 // To lower with a single SHUFPS we need to have the low half and high half
8359 // each requiring a single input.
8360 if (Mask[0] != -1 && Mask[1] != -1 && (Mask[0] < 4) != (Mask[1] < 4))
8362 if (Mask[2] != -1 && Mask[3] != -1 && (Mask[2] < 4) != (Mask[3] < 4))
8368 /// \brief Lower a vector shuffle using the SHUFPS instruction.
8370 /// This is a helper routine dedicated to lowering vector shuffles using SHUFPS.
8371 /// It makes no assumptions about whether this is the *best* lowering, it simply
8373 static SDValue lowerVectorShuffleWithSHUFPS(SDLoc DL, MVT VT,
8374 ArrayRef<int> Mask, SDValue V1,
8375 SDValue V2, SelectionDAG &DAG) {
8376 SDValue LowV = V1, HighV = V2;
8377 int NewMask[4] = {Mask[0], Mask[1], Mask[2], Mask[3]};
8380 std::count_if(Mask.begin(), Mask.end(), [](int M) { return M >= 4; });
8382 if (NumV2Elements == 1) {
8384 std::find_if(Mask.begin(), Mask.end(), [](int M) { return M >= 4; }) -
8387 // Compute the index adjacent to V2Index and in the same half by toggling
8389 int V2AdjIndex = V2Index ^ 1;
8391 if (Mask[V2AdjIndex] == -1) {
8392 // Handles all the cases where we have a single V2 element and an undef.
8393 // This will only ever happen in the high lanes because we commute the
8394 // vector otherwise.
8396 std::swap(LowV, HighV);
8397 NewMask[V2Index] -= 4;
8399 // Handle the case where the V2 element ends up adjacent to a V1 element.
8400 // To make this work, blend them together as the first step.
8401 int V1Index = V2AdjIndex;
8402 int BlendMask[4] = {Mask[V2Index] - 4, 0, Mask[V1Index], 0};
8403 V2 = DAG.getNode(X86ISD::SHUFP, DL, VT, V2, V1,
8404 getV4X86ShuffleImm8ForMask(BlendMask, DL, DAG));
8406 // Now proceed to reconstruct the final blend as we have the necessary
8407 // high or low half formed.
8414 NewMask[V1Index] = 2; // We put the V1 element in V2[2].
8415 NewMask[V2Index] = 0; // We shifted the V2 element into V2[0].
8417 } else if (NumV2Elements == 2) {
8418 if (Mask[0] < 4 && Mask[1] < 4) {
8419 // Handle the easy case where we have V1 in the low lanes and V2 in the
8423 } else if (Mask[2] < 4 && Mask[3] < 4) {
8424 // We also handle the reversed case because this utility may get called
8425 // when we detect a SHUFPS pattern but can't easily commute the shuffle to
8426 // arrange things in the right direction.
8432 // We have a mixture of V1 and V2 in both low and high lanes. Rather than
8433 // trying to place elements directly, just blend them and set up the final
8434 // shuffle to place them.
8436 // The first two blend mask elements are for V1, the second two are for
8438 int BlendMask[4] = {Mask[0] < 4 ? Mask[0] : Mask[1],
8439 Mask[2] < 4 ? Mask[2] : Mask[3],
8440 (Mask[0] >= 4 ? Mask[0] : Mask[1]) - 4,
8441 (Mask[2] >= 4 ? Mask[2] : Mask[3]) - 4};
8442 V1 = DAG.getNode(X86ISD::SHUFP, DL, VT, V1, V2,
8443 getV4X86ShuffleImm8ForMask(BlendMask, DL, DAG));
8445 // Now we do a normal shuffle of V1 by giving V1 as both operands to
8448 NewMask[0] = Mask[0] < 4 ? 0 : 2;
8449 NewMask[1] = Mask[0] < 4 ? 2 : 0;
8450 NewMask[2] = Mask[2] < 4 ? 1 : 3;
8451 NewMask[3] = Mask[2] < 4 ? 3 : 1;
8454 return DAG.getNode(X86ISD::SHUFP, DL, VT, LowV, HighV,
8455 getV4X86ShuffleImm8ForMask(NewMask, DL, DAG));
8458 /// \brief Lower 4-lane 32-bit floating point shuffles.
8460 /// Uses instructions exclusively from the floating point unit to minimize
8461 /// domain crossing penalties, as these are sufficient to implement all v4f32
8463 static SDValue lowerV4F32VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
8464 const X86Subtarget *Subtarget,
8465 SelectionDAG &DAG) {
8467 assert(Op.getSimpleValueType() == MVT::v4f32 && "Bad shuffle type!");
8468 assert(V1.getSimpleValueType() == MVT::v4f32 && "Bad operand type!");
8469 assert(V2.getSimpleValueType() == MVT::v4f32 && "Bad operand type!");
8470 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
8471 ArrayRef<int> Mask = SVOp->getMask();
8472 assert(Mask.size() == 4 && "Unexpected mask size for v4 shuffle!");
8475 std::count_if(Mask.begin(), Mask.end(), [](int M) { return M >= 4; });
8477 if (NumV2Elements == 0) {
8478 // Check for being able to broadcast a single element.
8479 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v4f32, V1,
8480 Mask, Subtarget, DAG))
8483 // Use even/odd duplicate instructions for masks that match their pattern.
8484 if (Subtarget->hasSSE3()) {
8485 if (isShuffleEquivalent(V1, V2, Mask, {0, 0, 2, 2}))
8486 return DAG.getNode(X86ISD::MOVSLDUP, DL, MVT::v4f32, V1);
8487 if (isShuffleEquivalent(V1, V2, Mask, {1, 1, 3, 3}))
8488 return DAG.getNode(X86ISD::MOVSHDUP, DL, MVT::v4f32, V1);
8491 if (Subtarget->hasAVX()) {
8492 // If we have AVX, we can use VPERMILPS which will allow folding a load
8493 // into the shuffle.
8494 return DAG.getNode(X86ISD::VPERMILPI, DL, MVT::v4f32, V1,
8495 getV4X86ShuffleImm8ForMask(Mask, DL, DAG));
8498 // Otherwise, use a straight shuffle of a single input vector. We pass the
8499 // input vector to both operands to simulate this with a SHUFPS.
8500 return DAG.getNode(X86ISD::SHUFP, DL, MVT::v4f32, V1, V1,
8501 getV4X86ShuffleImm8ForMask(Mask, DL, DAG));
8504 // There are special ways we can lower some single-element blends. However, we
8505 // have custom ways we can lower more complex single-element blends below that
8506 // we defer to if both this and BLENDPS fail to match, so restrict this to
8507 // when the V2 input is targeting element 0 of the mask -- that is the fast
8509 if (NumV2Elements == 1 && Mask[0] >= 4)
8510 if (SDValue V = lowerVectorShuffleAsElementInsertion(DL, MVT::v4f32, V1, V2,
8511 Mask, Subtarget, DAG))
8514 if (Subtarget->hasSSE41()) {
8515 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v4f32, V1, V2, Mask,
8519 // Use INSERTPS if we can complete the shuffle efficiently.
8520 if (SDValue V = lowerVectorShuffleAsInsertPS(Op, V1, V2, Mask, DAG))
8523 if (!isSingleSHUFPSMask(Mask))
8524 if (SDValue BlendPerm = lowerVectorShuffleAsBlendAndPermute(
8525 DL, MVT::v4f32, V1, V2, Mask, DAG))
8529 // Use dedicated unpack instructions for masks that match their pattern.
8531 lowerVectorShuffleWithUNPCK(DL, MVT::v4f32, Mask, V1, V2, DAG))
8534 // Otherwise fall back to a SHUFPS lowering strategy.
8535 return lowerVectorShuffleWithSHUFPS(DL, MVT::v4f32, Mask, V1, V2, DAG);
8538 /// \brief Lower 4-lane i32 vector shuffles.
8540 /// We try to handle these with integer-domain shuffles where we can, but for
8541 /// blends we use the floating point domain blend instructions.
8542 static SDValue lowerV4I32VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
8543 const X86Subtarget *Subtarget,
8544 SelectionDAG &DAG) {
8546 assert(Op.getSimpleValueType() == MVT::v4i32 && "Bad shuffle type!");
8547 assert(V1.getSimpleValueType() == MVT::v4i32 && "Bad operand type!");
8548 assert(V2.getSimpleValueType() == MVT::v4i32 && "Bad operand type!");
8549 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
8550 ArrayRef<int> Mask = SVOp->getMask();
8551 assert(Mask.size() == 4 && "Unexpected mask size for v4 shuffle!");
8553 // Whenever we can lower this as a zext, that instruction is strictly faster
8554 // than any alternative. It also allows us to fold memory operands into the
8555 // shuffle in many cases.
8556 if (SDValue ZExt = lowerVectorShuffleAsZeroOrAnyExtend(DL, MVT::v4i32, V1, V2,
8557 Mask, Subtarget, DAG))
8561 std::count_if(Mask.begin(), Mask.end(), [](int M) { return M >= 4; });
8563 if (NumV2Elements == 0) {
8564 // Check for being able to broadcast a single element.
8565 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v4i32, V1,
8566 Mask, Subtarget, DAG))
8569 // Straight shuffle of a single input vector. For everything from SSE2
8570 // onward this has a single fast instruction with no scary immediates.
8571 // We coerce the shuffle pattern to be compatible with UNPCK instructions
8572 // but we aren't actually going to use the UNPCK instruction because doing
8573 // so prevents folding a load into this instruction or making a copy.
8574 const int UnpackLoMask[] = {0, 0, 1, 1};
8575 const int UnpackHiMask[] = {2, 2, 3, 3};
8576 if (isShuffleEquivalent(V1, V2, Mask, {0, 0, 1, 1}))
8577 Mask = UnpackLoMask;
8578 else if (isShuffleEquivalent(V1, V2, Mask, {2, 2, 3, 3}))
8579 Mask = UnpackHiMask;
8581 return DAG.getNode(X86ISD::PSHUFD, DL, MVT::v4i32, V1,
8582 getV4X86ShuffleImm8ForMask(Mask, DL, DAG));
8585 // Try to use shift instructions.
8587 lowerVectorShuffleAsShift(DL, MVT::v4i32, V1, V2, Mask, DAG))
8590 // There are special ways we can lower some single-element blends.
8591 if (NumV2Elements == 1)
8592 if (SDValue V = lowerVectorShuffleAsElementInsertion(DL, MVT::v4i32, V1, V2,
8593 Mask, Subtarget, DAG))
8596 // We have different paths for blend lowering, but they all must use the
8597 // *exact* same predicate.
8598 bool IsBlendSupported = Subtarget->hasSSE41();
8599 if (IsBlendSupported)
8600 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v4i32, V1, V2, Mask,
8604 if (SDValue Masked =
8605 lowerVectorShuffleAsBitMask(DL, MVT::v4i32, V1, V2, Mask, DAG))
8608 // Use dedicated unpack instructions for masks that match their pattern.
8610 lowerVectorShuffleWithUNPCK(DL, MVT::v4i32, Mask, V1, V2, DAG))
8613 // Try to use byte rotation instructions.
8614 // Its more profitable for pre-SSSE3 to use shuffles/unpacks.
8615 if (Subtarget->hasSSSE3())
8616 if (SDValue Rotate = lowerVectorShuffleAsByteRotate(
8617 DL, MVT::v4i32, V1, V2, Mask, Subtarget, DAG))
8620 // If we have direct support for blends, we should lower by decomposing into
8621 // a permute. That will be faster than the domain cross.
8622 if (IsBlendSupported)
8623 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v4i32, V1, V2,
8626 // Try to lower by permuting the inputs into an unpack instruction.
8627 if (SDValue Unpack = lowerVectorShuffleAsPermuteAndUnpack(DL, MVT::v4i32, V1,
8631 // We implement this with SHUFPS because it can blend from two vectors.
8632 // Because we're going to eventually use SHUFPS, we use SHUFPS even to build
8633 // up the inputs, bypassing domain shift penalties that we would encur if we
8634 // directly used PSHUFD on Nehalem and older. For newer chips, this isn't
8636 return DAG.getBitcast(
8638 DAG.getVectorShuffle(MVT::v4f32, DL, DAG.getBitcast(MVT::v4f32, V1),
8639 DAG.getBitcast(MVT::v4f32, V2), Mask));
8642 /// \brief Lowering of single-input v8i16 shuffles is the cornerstone of SSE2
8643 /// shuffle lowering, and the most complex part.
8645 /// The lowering strategy is to try to form pairs of input lanes which are
8646 /// targeted at the same half of the final vector, and then use a dword shuffle
8647 /// to place them onto the right half, and finally unpack the paired lanes into
8648 /// their final position.
8650 /// The exact breakdown of how to form these dword pairs and align them on the
8651 /// correct sides is really tricky. See the comments within the function for
8652 /// more of the details.
8654 /// This code also handles repeated 128-bit lanes of v8i16 shuffles, but each
8655 /// lane must shuffle the *exact* same way. In fact, you must pass a v8 Mask to
8656 /// this routine for it to work correctly. To shuffle a 256-bit or 512-bit i16
8657 /// vector, form the analogous 128-bit 8-element Mask.
8658 static SDValue lowerV8I16GeneralSingleInputVectorShuffle(
8659 SDLoc DL, MVT VT, SDValue V, MutableArrayRef<int> Mask,
8660 const X86Subtarget *Subtarget, SelectionDAG &DAG) {
8661 assert(VT.getScalarType() == MVT::i16 && "Bad input type!");
8662 MVT PSHUFDVT = MVT::getVectorVT(MVT::i32, VT.getVectorNumElements() / 2);
8664 assert(Mask.size() == 8 && "Shuffle mask length doen't match!");
8665 MutableArrayRef<int> LoMask = Mask.slice(0, 4);
8666 MutableArrayRef<int> HiMask = Mask.slice(4, 4);
8668 SmallVector<int, 4> LoInputs;
8669 std::copy_if(LoMask.begin(), LoMask.end(), std::back_inserter(LoInputs),
8670 [](int M) { return M >= 0; });
8671 std::sort(LoInputs.begin(), LoInputs.end());
8672 LoInputs.erase(std::unique(LoInputs.begin(), LoInputs.end()), LoInputs.end());
8673 SmallVector<int, 4> HiInputs;
8674 std::copy_if(HiMask.begin(), HiMask.end(), std::back_inserter(HiInputs),
8675 [](int M) { return M >= 0; });
8676 std::sort(HiInputs.begin(), HiInputs.end());
8677 HiInputs.erase(std::unique(HiInputs.begin(), HiInputs.end()), HiInputs.end());
8679 std::lower_bound(LoInputs.begin(), LoInputs.end(), 4) - LoInputs.begin();
8680 int NumHToL = LoInputs.size() - NumLToL;
8682 std::lower_bound(HiInputs.begin(), HiInputs.end(), 4) - HiInputs.begin();
8683 int NumHToH = HiInputs.size() - NumLToH;
8684 MutableArrayRef<int> LToLInputs(LoInputs.data(), NumLToL);
8685 MutableArrayRef<int> LToHInputs(HiInputs.data(), NumLToH);
8686 MutableArrayRef<int> HToLInputs(LoInputs.data() + NumLToL, NumHToL);
8687 MutableArrayRef<int> HToHInputs(HiInputs.data() + NumLToH, NumHToH);
8689 // Simplify the 1-into-3 and 3-into-1 cases with a single pshufd. For all
8690 // such inputs we can swap two of the dwords across the half mark and end up
8691 // with <=2 inputs to each half in each half. Once there, we can fall through
8692 // to the generic code below. For example:
8694 // Input: [a, b, c, d, e, f, g, h] -PSHUFD[0,2,1,3]-> [a, b, e, f, c, d, g, h]
8695 // Mask: [0, 1, 2, 7, 4, 5, 6, 3] -----------------> [0, 1, 4, 7, 2, 3, 6, 5]
8697 // However in some very rare cases we have a 1-into-3 or 3-into-1 on one half
8698 // and an existing 2-into-2 on the other half. In this case we may have to
8699 // pre-shuffle the 2-into-2 half to avoid turning it into a 3-into-1 or
8700 // 1-into-3 which could cause us to cycle endlessly fixing each side in turn.
8701 // Fortunately, we don't have to handle anything but a 2-into-2 pattern
8702 // because any other situation (including a 3-into-1 or 1-into-3 in the other
8703 // half than the one we target for fixing) will be fixed when we re-enter this
8704 // path. We will also combine away any sequence of PSHUFD instructions that
8705 // result into a single instruction. Here is an example of the tricky case:
8707 // Input: [a, b, c, d, e, f, g, h] -PSHUFD[0,2,1,3]-> [a, b, e, f, c, d, g, h]
8708 // Mask: [3, 7, 1, 0, 2, 7, 3, 5] -THIS-IS-BAD!!!!-> [5, 7, 1, 0, 4, 7, 5, 3]
8710 // This now has a 1-into-3 in the high half! Instead, we do two shuffles:
8712 // Input: [a, b, c, d, e, f, g, h] PSHUFHW[0,2,1,3]-> [a, b, c, d, e, g, f, h]
8713 // Mask: [3, 7, 1, 0, 2, 7, 3, 5] -----------------> [3, 7, 1, 0, 2, 7, 3, 6]
8715 // Input: [a, b, c, d, e, g, f, h] -PSHUFD[0,2,1,3]-> [a, b, e, g, c, d, f, h]
8716 // Mask: [3, 7, 1, 0, 2, 7, 3, 6] -----------------> [5, 7, 1, 0, 4, 7, 5, 6]
8718 // The result is fine to be handled by the generic logic.
8719 auto balanceSides = [&](ArrayRef<int> AToAInputs, ArrayRef<int> BToAInputs,
8720 ArrayRef<int> BToBInputs, ArrayRef<int> AToBInputs,
8721 int AOffset, int BOffset) {
8722 assert((AToAInputs.size() == 3 || AToAInputs.size() == 1) &&
8723 "Must call this with A having 3 or 1 inputs from the A half.");
8724 assert((BToAInputs.size() == 1 || BToAInputs.size() == 3) &&
8725 "Must call this with B having 1 or 3 inputs from the B half.");
8726 assert(AToAInputs.size() + BToAInputs.size() == 4 &&
8727 "Must call this with either 3:1 or 1:3 inputs (summing to 4).");
8729 bool ThreeAInputs = AToAInputs.size() == 3;
8731 // Compute the index of dword with only one word among the three inputs in
8732 // a half by taking the sum of the half with three inputs and subtracting
8733 // the sum of the actual three inputs. The difference is the remaining
8736 int &TripleDWord = ThreeAInputs ? ADWord : BDWord;
8737 int &OneInputDWord = ThreeAInputs ? BDWord : ADWord;
8738 int TripleInputOffset = ThreeAInputs ? AOffset : BOffset;
8739 ArrayRef<int> TripleInputs = ThreeAInputs ? AToAInputs : BToAInputs;
8740 int OneInput = ThreeAInputs ? BToAInputs[0] : AToAInputs[0];
8741 int TripleInputSum = 0 + 1 + 2 + 3 + (4 * TripleInputOffset);
8742 int TripleNonInputIdx =
8743 TripleInputSum - std::accumulate(TripleInputs.begin(), TripleInputs.end(), 0);
8744 TripleDWord = TripleNonInputIdx / 2;
8746 // We use xor with one to compute the adjacent DWord to whichever one the
8748 OneInputDWord = (OneInput / 2) ^ 1;
8750 // Check for one tricky case: We're fixing a 3<-1 or a 1<-3 shuffle for AToA
8751 // and BToA inputs. If there is also such a problem with the BToB and AToB
8752 // inputs, we don't try to fix it necessarily -- we'll recurse and see it in
8753 // the next pass. However, if we have a 2<-2 in the BToB and AToB inputs, it
8754 // is essential that we don't *create* a 3<-1 as then we might oscillate.
8755 if (BToBInputs.size() == 2 && AToBInputs.size() == 2) {
8756 // Compute how many inputs will be flipped by swapping these DWords. We
8758 // to balance this to ensure we don't form a 3-1 shuffle in the other
8760 int NumFlippedAToBInputs =
8761 std::count(AToBInputs.begin(), AToBInputs.end(), 2 * ADWord) +
8762 std::count(AToBInputs.begin(), AToBInputs.end(), 2 * ADWord + 1);
8763 int NumFlippedBToBInputs =
8764 std::count(BToBInputs.begin(), BToBInputs.end(), 2 * BDWord) +
8765 std::count(BToBInputs.begin(), BToBInputs.end(), 2 * BDWord + 1);
8766 if ((NumFlippedAToBInputs == 1 &&
8767 (NumFlippedBToBInputs == 0 || NumFlippedBToBInputs == 2)) ||
8768 (NumFlippedBToBInputs == 1 &&
8769 (NumFlippedAToBInputs == 0 || NumFlippedAToBInputs == 2))) {
8770 // We choose whether to fix the A half or B half based on whether that
8771 // half has zero flipped inputs. At zero, we may not be able to fix it
8772 // with that half. We also bias towards fixing the B half because that
8773 // will more commonly be the high half, and we have to bias one way.
8774 auto FixFlippedInputs = [&V, &DL, &Mask, &DAG](int PinnedIdx, int DWord,
8775 ArrayRef<int> Inputs) {
8776 int FixIdx = PinnedIdx ^ 1; // The adjacent slot to the pinned slot.
8777 bool IsFixIdxInput = std::find(Inputs.begin(), Inputs.end(),
8778 PinnedIdx ^ 1) != Inputs.end();
8779 // Determine whether the free index is in the flipped dword or the
8780 // unflipped dword based on where the pinned index is. We use this bit
8781 // in an xor to conditionally select the adjacent dword.
8782 int FixFreeIdx = 2 * (DWord ^ (PinnedIdx / 2 == DWord));
8783 bool IsFixFreeIdxInput = std::find(Inputs.begin(), Inputs.end(),
8784 FixFreeIdx) != Inputs.end();
8785 if (IsFixIdxInput == IsFixFreeIdxInput)
8787 IsFixFreeIdxInput = std::find(Inputs.begin(), Inputs.end(),
8788 FixFreeIdx) != Inputs.end();
8789 assert(IsFixIdxInput != IsFixFreeIdxInput &&
8790 "We need to be changing the number of flipped inputs!");
8791 int PSHUFHalfMask[] = {0, 1, 2, 3};
8792 std::swap(PSHUFHalfMask[FixFreeIdx % 4], PSHUFHalfMask[FixIdx % 4]);
8793 V = DAG.getNode(FixIdx < 4 ? X86ISD::PSHUFLW : X86ISD::PSHUFHW, DL,
8795 getV4X86ShuffleImm8ForMask(PSHUFHalfMask, DL, DAG));
8798 if (M != -1 && M == FixIdx)
8800 else if (M != -1 && M == FixFreeIdx)
8803 if (NumFlippedBToBInputs != 0) {
8805 BToAInputs.size() == 3 ? TripleNonInputIdx : OneInput;
8806 FixFlippedInputs(BPinnedIdx, BDWord, BToBInputs);
8808 assert(NumFlippedAToBInputs != 0 && "Impossible given predicates!");
8809 int APinnedIdx = ThreeAInputs ? TripleNonInputIdx : OneInput;
8810 FixFlippedInputs(APinnedIdx, ADWord, AToBInputs);
8815 int PSHUFDMask[] = {0, 1, 2, 3};
8816 PSHUFDMask[ADWord] = BDWord;
8817 PSHUFDMask[BDWord] = ADWord;
8820 DAG.getNode(X86ISD::PSHUFD, DL, PSHUFDVT, DAG.getBitcast(PSHUFDVT, V),
8821 getV4X86ShuffleImm8ForMask(PSHUFDMask, DL, DAG)));
8823 // Adjust the mask to match the new locations of A and B.
8825 if (M != -1 && M/2 == ADWord)
8826 M = 2 * BDWord + M % 2;
8827 else if (M != -1 && M/2 == BDWord)
8828 M = 2 * ADWord + M % 2;
8830 // Recurse back into this routine to re-compute state now that this isn't
8831 // a 3 and 1 problem.
8832 return lowerV8I16GeneralSingleInputVectorShuffle(DL, VT, V, Mask, Subtarget,
8835 if ((NumLToL == 3 && NumHToL == 1) || (NumLToL == 1 && NumHToL == 3))
8836 return balanceSides(LToLInputs, HToLInputs, HToHInputs, LToHInputs, 0, 4);
8837 else if ((NumHToH == 3 && NumLToH == 1) || (NumHToH == 1 && NumLToH == 3))
8838 return balanceSides(HToHInputs, LToHInputs, LToLInputs, HToLInputs, 4, 0);
8840 // At this point there are at most two inputs to the low and high halves from
8841 // each half. That means the inputs can always be grouped into dwords and
8842 // those dwords can then be moved to the correct half with a dword shuffle.
8843 // We use at most one low and one high word shuffle to collect these paired
8844 // inputs into dwords, and finally a dword shuffle to place them.
8845 int PSHUFLMask[4] = {-1, -1, -1, -1};
8846 int PSHUFHMask[4] = {-1, -1, -1, -1};
8847 int PSHUFDMask[4] = {-1, -1, -1, -1};
8849 // First fix the masks for all the inputs that are staying in their
8850 // original halves. This will then dictate the targets of the cross-half
8852 auto fixInPlaceInputs =
8853 [&PSHUFDMask](ArrayRef<int> InPlaceInputs, ArrayRef<int> IncomingInputs,
8854 MutableArrayRef<int> SourceHalfMask,
8855 MutableArrayRef<int> HalfMask, int HalfOffset) {
8856 if (InPlaceInputs.empty())
8858 if (InPlaceInputs.size() == 1) {
8859 SourceHalfMask[InPlaceInputs[0] - HalfOffset] =
8860 InPlaceInputs[0] - HalfOffset;
8861 PSHUFDMask[InPlaceInputs[0] / 2] = InPlaceInputs[0] / 2;
8864 if (IncomingInputs.empty()) {
8865 // Just fix all of the in place inputs.
8866 for (int Input : InPlaceInputs) {
8867 SourceHalfMask[Input - HalfOffset] = Input - HalfOffset;
8868 PSHUFDMask[Input / 2] = Input / 2;
8873 assert(InPlaceInputs.size() == 2 && "Cannot handle 3 or 4 inputs!");
8874 SourceHalfMask[InPlaceInputs[0] - HalfOffset] =
8875 InPlaceInputs[0] - HalfOffset;
8876 // Put the second input next to the first so that they are packed into
8877 // a dword. We find the adjacent index by toggling the low bit.
8878 int AdjIndex = InPlaceInputs[0] ^ 1;
8879 SourceHalfMask[AdjIndex - HalfOffset] = InPlaceInputs[1] - HalfOffset;
8880 std::replace(HalfMask.begin(), HalfMask.end(), InPlaceInputs[1], AdjIndex);
8881 PSHUFDMask[AdjIndex / 2] = AdjIndex / 2;
8883 fixInPlaceInputs(LToLInputs, HToLInputs, PSHUFLMask, LoMask, 0);
8884 fixInPlaceInputs(HToHInputs, LToHInputs, PSHUFHMask, HiMask, 4);
8886 // Now gather the cross-half inputs and place them into a free dword of
8887 // their target half.
8888 // FIXME: This operation could almost certainly be simplified dramatically to
8889 // look more like the 3-1 fixing operation.
8890 auto moveInputsToRightHalf = [&PSHUFDMask](
8891 MutableArrayRef<int> IncomingInputs, ArrayRef<int> ExistingInputs,
8892 MutableArrayRef<int> SourceHalfMask, MutableArrayRef<int> HalfMask,
8893 MutableArrayRef<int> FinalSourceHalfMask, int SourceOffset,
8895 auto isWordClobbered = [](ArrayRef<int> SourceHalfMask, int Word) {
8896 return SourceHalfMask[Word] != -1 && SourceHalfMask[Word] != Word;
8898 auto isDWordClobbered = [&isWordClobbered](ArrayRef<int> SourceHalfMask,
8900 int LowWord = Word & ~1;
8901 int HighWord = Word | 1;
8902 return isWordClobbered(SourceHalfMask, LowWord) ||
8903 isWordClobbered(SourceHalfMask, HighWord);
8906 if (IncomingInputs.empty())
8909 if (ExistingInputs.empty()) {
8910 // Map any dwords with inputs from them into the right half.
8911 for (int Input : IncomingInputs) {
8912 // If the source half mask maps over the inputs, turn those into
8913 // swaps and use the swapped lane.
8914 if (isWordClobbered(SourceHalfMask, Input - SourceOffset)) {
8915 if (SourceHalfMask[SourceHalfMask[Input - SourceOffset]] == -1) {
8916 SourceHalfMask[SourceHalfMask[Input - SourceOffset]] =
8917 Input - SourceOffset;
8918 // We have to swap the uses in our half mask in one sweep.
8919 for (int &M : HalfMask)
8920 if (M == SourceHalfMask[Input - SourceOffset] + SourceOffset)
8922 else if (M == Input)
8923 M = SourceHalfMask[Input - SourceOffset] + SourceOffset;
8925 assert(SourceHalfMask[SourceHalfMask[Input - SourceOffset]] ==
8926 Input - SourceOffset &&
8927 "Previous placement doesn't match!");
8929 // Note that this correctly re-maps both when we do a swap and when
8930 // we observe the other side of the swap above. We rely on that to
8931 // avoid swapping the members of the input list directly.
8932 Input = SourceHalfMask[Input - SourceOffset] + SourceOffset;
8935 // Map the input's dword into the correct half.
8936 if (PSHUFDMask[(Input - SourceOffset + DestOffset) / 2] == -1)
8937 PSHUFDMask[(Input - SourceOffset + DestOffset) / 2] = Input / 2;
8939 assert(PSHUFDMask[(Input - SourceOffset + DestOffset) / 2] ==
8941 "Previous placement doesn't match!");
8944 // And just directly shift any other-half mask elements to be same-half
8945 // as we will have mirrored the dword containing the element into the
8946 // same position within that half.
8947 for (int &M : HalfMask)
8948 if (M >= SourceOffset && M < SourceOffset + 4) {
8949 M = M - SourceOffset + DestOffset;
8950 assert(M >= 0 && "This should never wrap below zero!");
8955 // Ensure we have the input in a viable dword of its current half. This
8956 // is particularly tricky because the original position may be clobbered
8957 // by inputs being moved and *staying* in that half.
8958 if (IncomingInputs.size() == 1) {
8959 if (isWordClobbered(SourceHalfMask, IncomingInputs[0] - SourceOffset)) {
8960 int InputFixed = std::find(std::begin(SourceHalfMask),
8961 std::end(SourceHalfMask), -1) -
8962 std::begin(SourceHalfMask) + SourceOffset;
8963 SourceHalfMask[InputFixed - SourceOffset] =
8964 IncomingInputs[0] - SourceOffset;
8965 std::replace(HalfMask.begin(), HalfMask.end(), IncomingInputs[0],
8967 IncomingInputs[0] = InputFixed;
8969 } else if (IncomingInputs.size() == 2) {
8970 if (IncomingInputs[0] / 2 != IncomingInputs[1] / 2 ||
8971 isDWordClobbered(SourceHalfMask, IncomingInputs[0] - SourceOffset)) {
8972 // We have two non-adjacent or clobbered inputs we need to extract from
8973 // the source half. To do this, we need to map them into some adjacent
8974 // dword slot in the source mask.
8975 int InputsFixed[2] = {IncomingInputs[0] - SourceOffset,
8976 IncomingInputs[1] - SourceOffset};
8978 // If there is a free slot in the source half mask adjacent to one of
8979 // the inputs, place the other input in it. We use (Index XOR 1) to
8980 // compute an adjacent index.
8981 if (!isWordClobbered(SourceHalfMask, InputsFixed[0]) &&
8982 SourceHalfMask[InputsFixed[0] ^ 1] == -1) {
8983 SourceHalfMask[InputsFixed[0]] = InputsFixed[0];
8984 SourceHalfMask[InputsFixed[0] ^ 1] = InputsFixed[1];
8985 InputsFixed[1] = InputsFixed[0] ^ 1;
8986 } else if (!isWordClobbered(SourceHalfMask, InputsFixed[1]) &&
8987 SourceHalfMask[InputsFixed[1] ^ 1] == -1) {
8988 SourceHalfMask[InputsFixed[1]] = InputsFixed[1];
8989 SourceHalfMask[InputsFixed[1] ^ 1] = InputsFixed[0];
8990 InputsFixed[0] = InputsFixed[1] ^ 1;
8991 } else if (SourceHalfMask[2 * ((InputsFixed[0] / 2) ^ 1)] == -1 &&
8992 SourceHalfMask[2 * ((InputsFixed[0] / 2) ^ 1) + 1] == -1) {
8993 // The two inputs are in the same DWord but it is clobbered and the
8994 // adjacent DWord isn't used at all. Move both inputs to the free
8996 SourceHalfMask[2 * ((InputsFixed[0] / 2) ^ 1)] = InputsFixed[0];
8997 SourceHalfMask[2 * ((InputsFixed[0] / 2) ^ 1) + 1] = InputsFixed[1];
8998 InputsFixed[0] = 2 * ((InputsFixed[0] / 2) ^ 1);
8999 InputsFixed[1] = 2 * ((InputsFixed[0] / 2) ^ 1) + 1;
9001 // The only way we hit this point is if there is no clobbering
9002 // (because there are no off-half inputs to this half) and there is no
9003 // free slot adjacent to one of the inputs. In this case, we have to
9004 // swap an input with a non-input.
9005 for (int i = 0; i < 4; ++i)
9006 assert((SourceHalfMask[i] == -1 || SourceHalfMask[i] == i) &&
9007 "We can't handle any clobbers here!");
9008 assert(InputsFixed[1] != (InputsFixed[0] ^ 1) &&
9009 "Cannot have adjacent inputs here!");
9011 SourceHalfMask[InputsFixed[0] ^ 1] = InputsFixed[1];
9012 SourceHalfMask[InputsFixed[1]] = InputsFixed[0] ^ 1;
9014 // We also have to update the final source mask in this case because
9015 // it may need to undo the above swap.
9016 for (int &M : FinalSourceHalfMask)
9017 if (M == (InputsFixed[0] ^ 1) + SourceOffset)
9018 M = InputsFixed[1] + SourceOffset;
9019 else if (M == InputsFixed[1] + SourceOffset)
9020 M = (InputsFixed[0] ^ 1) + SourceOffset;
9022 InputsFixed[1] = InputsFixed[0] ^ 1;
9025 // Point everything at the fixed inputs.
9026 for (int &M : HalfMask)
9027 if (M == IncomingInputs[0])
9028 M = InputsFixed[0] + SourceOffset;
9029 else if (M == IncomingInputs[1])
9030 M = InputsFixed[1] + SourceOffset;
9032 IncomingInputs[0] = InputsFixed[0] + SourceOffset;
9033 IncomingInputs[1] = InputsFixed[1] + SourceOffset;
9036 llvm_unreachable("Unhandled input size!");
9039 // Now hoist the DWord down to the right half.
9040 int FreeDWord = (PSHUFDMask[DestOffset / 2] == -1 ? 0 : 1) + DestOffset / 2;
9041 assert(PSHUFDMask[FreeDWord] == -1 && "DWord not free");
9042 PSHUFDMask[FreeDWord] = IncomingInputs[0] / 2;
9043 for (int &M : HalfMask)
9044 for (int Input : IncomingInputs)
9046 M = FreeDWord * 2 + Input % 2;
9048 moveInputsToRightHalf(HToLInputs, LToLInputs, PSHUFHMask, LoMask, HiMask,
9049 /*SourceOffset*/ 4, /*DestOffset*/ 0);
9050 moveInputsToRightHalf(LToHInputs, HToHInputs, PSHUFLMask, HiMask, LoMask,
9051 /*SourceOffset*/ 0, /*DestOffset*/ 4);
9053 // Now enact all the shuffles we've computed to move the inputs into their
9055 if (!isNoopShuffleMask(PSHUFLMask))
9056 V = DAG.getNode(X86ISD::PSHUFLW, DL, VT, V,
9057 getV4X86ShuffleImm8ForMask(PSHUFLMask, DL, DAG));
9058 if (!isNoopShuffleMask(PSHUFHMask))
9059 V = DAG.getNode(X86ISD::PSHUFHW, DL, VT, V,
9060 getV4X86ShuffleImm8ForMask(PSHUFHMask, DL, DAG));
9061 if (!isNoopShuffleMask(PSHUFDMask))
9064 DAG.getNode(X86ISD::PSHUFD, DL, PSHUFDVT, DAG.getBitcast(PSHUFDVT, V),
9065 getV4X86ShuffleImm8ForMask(PSHUFDMask, DL, DAG)));
9067 // At this point, each half should contain all its inputs, and we can then
9068 // just shuffle them into their final position.
9069 assert(std::count_if(LoMask.begin(), LoMask.end(),
9070 [](int M) { return M >= 4; }) == 0 &&
9071 "Failed to lift all the high half inputs to the low mask!");
9072 assert(std::count_if(HiMask.begin(), HiMask.end(),
9073 [](int M) { return M >= 0 && M < 4; }) == 0 &&
9074 "Failed to lift all the low half inputs to the high mask!");
9076 // Do a half shuffle for the low mask.
9077 if (!isNoopShuffleMask(LoMask))
9078 V = DAG.getNode(X86ISD::PSHUFLW, DL, VT, V,
9079 getV4X86ShuffleImm8ForMask(LoMask, DL, DAG));
9081 // Do a half shuffle with the high mask after shifting its values down.
9082 for (int &M : HiMask)
9085 if (!isNoopShuffleMask(HiMask))
9086 V = DAG.getNode(X86ISD::PSHUFHW, DL, VT, V,
9087 getV4X86ShuffleImm8ForMask(HiMask, DL, DAG));
9092 /// \brief Helper to form a PSHUFB-based shuffle+blend.
9093 static SDValue lowerVectorShuffleAsPSHUFB(SDLoc DL, MVT VT, SDValue V1,
9094 SDValue V2, ArrayRef<int> Mask,
9095 SelectionDAG &DAG, bool &V1InUse,
9097 SmallBitVector Zeroable = computeZeroableShuffleElements(Mask, V1, V2);
9103 int Size = Mask.size();
9104 int Scale = 16 / Size;
9105 for (int i = 0; i < 16; ++i) {
9106 if (Mask[i / Scale] == -1) {
9107 V1Mask[i] = V2Mask[i] = DAG.getUNDEF(MVT::i8);
9109 const int ZeroMask = 0x80;
9110 int V1Idx = Mask[i / Scale] < Size ? Mask[i / Scale] * Scale + i % Scale
9112 int V2Idx = Mask[i / Scale] < Size
9114 : (Mask[i / Scale] - Size) * Scale + i % Scale;
9115 if (Zeroable[i / Scale])
9116 V1Idx = V2Idx = ZeroMask;
9117 V1Mask[i] = DAG.getConstant(V1Idx, DL, MVT::i8);
9118 V2Mask[i] = DAG.getConstant(V2Idx, DL, MVT::i8);
9119 V1InUse |= (ZeroMask != V1Idx);
9120 V2InUse |= (ZeroMask != V2Idx);
9125 V1 = DAG.getNode(X86ISD::PSHUFB, DL, MVT::v16i8,
9126 DAG.getBitcast(MVT::v16i8, V1),
9127 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v16i8, V1Mask));
9129 V2 = DAG.getNode(X86ISD::PSHUFB, DL, MVT::v16i8,
9130 DAG.getBitcast(MVT::v16i8, V2),
9131 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v16i8, V2Mask));
9133 // If we need shuffled inputs from both, blend the two.
9135 if (V1InUse && V2InUse)
9136 V = DAG.getNode(ISD::OR, DL, MVT::v16i8, V1, V2);
9138 V = V1InUse ? V1 : V2;
9140 // Cast the result back to the correct type.
9141 return DAG.getBitcast(VT, V);
9144 /// \brief Generic lowering of 8-lane i16 shuffles.
9146 /// This handles both single-input shuffles and combined shuffle/blends with
9147 /// two inputs. The single input shuffles are immediately delegated to
9148 /// a dedicated lowering routine.
9150 /// The blends are lowered in one of three fundamental ways. If there are few
9151 /// enough inputs, it delegates to a basic UNPCK-based strategy. If the shuffle
9152 /// of the input is significantly cheaper when lowered as an interleaving of
9153 /// the two inputs, try to interleave them. Otherwise, blend the low and high
9154 /// halves of the inputs separately (making them have relatively few inputs)
9155 /// and then concatenate them.
9156 static SDValue lowerV8I16VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
9157 const X86Subtarget *Subtarget,
9158 SelectionDAG &DAG) {
9160 assert(Op.getSimpleValueType() == MVT::v8i16 && "Bad shuffle type!");
9161 assert(V1.getSimpleValueType() == MVT::v8i16 && "Bad operand type!");
9162 assert(V2.getSimpleValueType() == MVT::v8i16 && "Bad operand type!");
9163 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
9164 ArrayRef<int> OrigMask = SVOp->getMask();
9165 int MaskStorage[8] = {OrigMask[0], OrigMask[1], OrigMask[2], OrigMask[3],
9166 OrigMask[4], OrigMask[5], OrigMask[6], OrigMask[7]};
9167 MutableArrayRef<int> Mask(MaskStorage);
9169 assert(Mask.size() == 8 && "Unexpected mask size for v8 shuffle!");
9171 // Whenever we can lower this as a zext, that instruction is strictly faster
9172 // than any alternative.
9173 if (SDValue ZExt = lowerVectorShuffleAsZeroOrAnyExtend(
9174 DL, MVT::v8i16, V1, V2, OrigMask, Subtarget, DAG))
9177 auto isV1 = [](int M) { return M >= 0 && M < 8; };
9179 auto isV2 = [](int M) { return M >= 8; };
9181 int NumV2Inputs = std::count_if(Mask.begin(), Mask.end(), isV2);
9183 if (NumV2Inputs == 0) {
9184 // Check for being able to broadcast a single element.
9185 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v8i16, V1,
9186 Mask, Subtarget, DAG))
9189 // Try to use shift instructions.
9191 lowerVectorShuffleAsShift(DL, MVT::v8i16, V1, V1, Mask, DAG))
9194 // Use dedicated unpack instructions for masks that match their pattern.
9196 lowerVectorShuffleWithUNPCK(DL, MVT::v8i16, Mask, V1, V2, DAG))
9199 // Try to use byte rotation instructions.
9200 if (SDValue Rotate = lowerVectorShuffleAsByteRotate(DL, MVT::v8i16, V1, V1,
9201 Mask, Subtarget, DAG))
9204 return lowerV8I16GeneralSingleInputVectorShuffle(DL, MVT::v8i16, V1, Mask,
9208 assert(std::any_of(Mask.begin(), Mask.end(), isV1) &&
9209 "All single-input shuffles should be canonicalized to be V1-input "
9212 // Try to use shift instructions.
9214 lowerVectorShuffleAsShift(DL, MVT::v8i16, V1, V2, Mask, DAG))
9217 // See if we can use SSE4A Extraction / Insertion.
9218 if (Subtarget->hasSSE4A())
9219 if (SDValue V = lowerVectorShuffleWithSSE4A(DL, MVT::v8i16, V1, V2, Mask, DAG))
9222 // There are special ways we can lower some single-element blends.
9223 if (NumV2Inputs == 1)
9224 if (SDValue V = lowerVectorShuffleAsElementInsertion(DL, MVT::v8i16, V1, V2,
9225 Mask, Subtarget, DAG))
9228 // We have different paths for blend lowering, but they all must use the
9229 // *exact* same predicate.
9230 bool IsBlendSupported = Subtarget->hasSSE41();
9231 if (IsBlendSupported)
9232 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v8i16, V1, V2, Mask,
9236 if (SDValue Masked =
9237 lowerVectorShuffleAsBitMask(DL, MVT::v8i16, V1, V2, Mask, DAG))
9240 // Use dedicated unpack instructions for masks that match their pattern.
9242 lowerVectorShuffleWithUNPCK(DL, MVT::v8i16, Mask, V1, V2, DAG))
9245 // Try to use byte rotation instructions.
9246 if (SDValue Rotate = lowerVectorShuffleAsByteRotate(
9247 DL, MVT::v8i16, V1, V2, Mask, Subtarget, DAG))
9250 if (SDValue BitBlend =
9251 lowerVectorShuffleAsBitBlend(DL, MVT::v8i16, V1, V2, Mask, DAG))
9254 if (SDValue Unpack = lowerVectorShuffleAsPermuteAndUnpack(DL, MVT::v8i16, V1,
9258 // If we can't directly blend but can use PSHUFB, that will be better as it
9259 // can both shuffle and set up the inefficient blend.
9260 if (!IsBlendSupported && Subtarget->hasSSSE3()) {
9261 bool V1InUse, V2InUse;
9262 return lowerVectorShuffleAsPSHUFB(DL, MVT::v8i16, V1, V2, Mask, DAG,
9266 // We can always bit-blend if we have to so the fallback strategy is to
9267 // decompose into single-input permutes and blends.
9268 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v8i16, V1, V2,
9272 /// \brief Check whether a compaction lowering can be done by dropping even
9273 /// elements and compute how many times even elements must be dropped.
9275 /// This handles shuffles which take every Nth element where N is a power of
9276 /// two. Example shuffle masks:
9278 /// N = 1: 0, 2, 4, 6, 8, 10, 12, 14, 0, 2, 4, 6, 8, 10, 12, 14
9279 /// N = 1: 0, 2, 4, 6, 8, 10, 12, 14, 16, 18, 20, 22, 24, 26, 28, 30
9280 /// N = 2: 0, 4, 8, 12, 0, 4, 8, 12, 0, 4, 8, 12, 0, 4, 8, 12
9281 /// N = 2: 0, 4, 8, 12, 16, 20, 24, 28, 0, 4, 8, 12, 16, 20, 24, 28
9282 /// N = 3: 0, 8, 0, 8, 0, 8, 0, 8, 0, 8, 0, 8, 0, 8, 0, 8
9283 /// N = 3: 0, 8, 16, 24, 0, 8, 16, 24, 0, 8, 16, 24, 0, 8, 16, 24
9285 /// Any of these lanes can of course be undef.
9287 /// This routine only supports N <= 3.
9288 /// FIXME: Evaluate whether either AVX or AVX-512 have any opportunities here
9291 /// \returns N above, or the number of times even elements must be dropped if
9292 /// there is such a number. Otherwise returns zero.
9293 static int canLowerByDroppingEvenElements(ArrayRef<int> Mask) {
9294 // Figure out whether we're looping over two inputs or just one.
9295 bool IsSingleInput = isSingleInputShuffleMask(Mask);
9297 // The modulus for the shuffle vector entries is based on whether this is
9298 // a single input or not.
9299 int ShuffleModulus = Mask.size() * (IsSingleInput ? 1 : 2);
9300 assert(isPowerOf2_32((uint32_t)ShuffleModulus) &&
9301 "We should only be called with masks with a power-of-2 size!");
9303 uint64_t ModMask = (uint64_t)ShuffleModulus - 1;
9305 // We track whether the input is viable for all power-of-2 strides 2^1, 2^2,
9306 // and 2^3 simultaneously. This is because we may have ambiguity with
9307 // partially undef inputs.
9308 bool ViableForN[3] = {true, true, true};
9310 for (int i = 0, e = Mask.size(); i < e; ++i) {
9311 // Ignore undef lanes, we'll optimistically collapse them to the pattern we
9316 bool IsAnyViable = false;
9317 for (unsigned j = 0; j != array_lengthof(ViableForN); ++j)
9318 if (ViableForN[j]) {
9321 // The shuffle mask must be equal to (i * 2^N) % M.
9322 if ((uint64_t)Mask[i] == (((uint64_t)i << N) & ModMask))
9325 ViableForN[j] = false;
9327 // Early exit if we exhaust the possible powers of two.
9332 for (unsigned j = 0; j != array_lengthof(ViableForN); ++j)
9336 // Return 0 as there is no viable power of two.
9340 /// \brief Generic lowering of v16i8 shuffles.
9342 /// This is a hybrid strategy to lower v16i8 vectors. It first attempts to
9343 /// detect any complexity reducing interleaving. If that doesn't help, it uses
9344 /// UNPCK to spread the i8 elements across two i16-element vectors, and uses
9345 /// the existing lowering for v8i16 blends on each half, finally PACK-ing them
9347 static SDValue lowerV16I8VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
9348 const X86Subtarget *Subtarget,
9349 SelectionDAG &DAG) {
9351 assert(Op.getSimpleValueType() == MVT::v16i8 && "Bad shuffle type!");
9352 assert(V1.getSimpleValueType() == MVT::v16i8 && "Bad operand type!");
9353 assert(V2.getSimpleValueType() == MVT::v16i8 && "Bad operand type!");
9354 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
9355 ArrayRef<int> Mask = SVOp->getMask();
9356 assert(Mask.size() == 16 && "Unexpected mask size for v16 shuffle!");
9358 // Try to use shift instructions.
9360 lowerVectorShuffleAsShift(DL, MVT::v16i8, V1, V2, Mask, DAG))
9363 // Try to use byte rotation instructions.
9364 if (SDValue Rotate = lowerVectorShuffleAsByteRotate(
9365 DL, MVT::v16i8, V1, V2, Mask, Subtarget, DAG))
9368 // Try to use a zext lowering.
9369 if (SDValue ZExt = lowerVectorShuffleAsZeroOrAnyExtend(
9370 DL, MVT::v16i8, V1, V2, Mask, Subtarget, DAG))
9373 // See if we can use SSE4A Extraction / Insertion.
9374 if (Subtarget->hasSSE4A())
9375 if (SDValue V = lowerVectorShuffleWithSSE4A(DL, MVT::v16i8, V1, V2, Mask, DAG))
9379 std::count_if(Mask.begin(), Mask.end(), [](int M) { return M >= 16; });
9381 // For single-input shuffles, there are some nicer lowering tricks we can use.
9382 if (NumV2Elements == 0) {
9383 // Check for being able to broadcast a single element.
9384 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v16i8, V1,
9385 Mask, Subtarget, DAG))
9388 // Check whether we can widen this to an i16 shuffle by duplicating bytes.
9389 // Notably, this handles splat and partial-splat shuffles more efficiently.
9390 // However, it only makes sense if the pre-duplication shuffle simplifies
9391 // things significantly. Currently, this means we need to be able to
9392 // express the pre-duplication shuffle as an i16 shuffle.
9394 // FIXME: We should check for other patterns which can be widened into an
9395 // i16 shuffle as well.
9396 auto canWidenViaDuplication = [](ArrayRef<int> Mask) {
9397 for (int i = 0; i < 16; i += 2)
9398 if (Mask[i] != -1 && Mask[i + 1] != -1 && Mask[i] != Mask[i + 1])
9403 auto tryToWidenViaDuplication = [&]() -> SDValue {
9404 if (!canWidenViaDuplication(Mask))
9406 SmallVector<int, 4> LoInputs;
9407 std::copy_if(Mask.begin(), Mask.end(), std::back_inserter(LoInputs),
9408 [](int M) { return M >= 0 && M < 8; });
9409 std::sort(LoInputs.begin(), LoInputs.end());
9410 LoInputs.erase(std::unique(LoInputs.begin(), LoInputs.end()),
9412 SmallVector<int, 4> HiInputs;
9413 std::copy_if(Mask.begin(), Mask.end(), std::back_inserter(HiInputs),
9414 [](int M) { return M >= 8; });
9415 std::sort(HiInputs.begin(), HiInputs.end());
9416 HiInputs.erase(std::unique(HiInputs.begin(), HiInputs.end()),
9419 bool TargetLo = LoInputs.size() >= HiInputs.size();
9420 ArrayRef<int> InPlaceInputs = TargetLo ? LoInputs : HiInputs;
9421 ArrayRef<int> MovingInputs = TargetLo ? HiInputs : LoInputs;
9423 int PreDupI16Shuffle[] = {-1, -1, -1, -1, -1, -1, -1, -1};
9424 SmallDenseMap<int, int, 8> LaneMap;
9425 for (int I : InPlaceInputs) {
9426 PreDupI16Shuffle[I/2] = I/2;
9429 int j = TargetLo ? 0 : 4, je = j + 4;
9430 for (int i = 0, ie = MovingInputs.size(); i < ie; ++i) {
9431 // Check if j is already a shuffle of this input. This happens when
9432 // there are two adjacent bytes after we move the low one.
9433 if (PreDupI16Shuffle[j] != MovingInputs[i] / 2) {
9434 // If we haven't yet mapped the input, search for a slot into which
9436 while (j < je && PreDupI16Shuffle[j] != -1)
9440 // We can't place the inputs into a single half with a simple i16 shuffle, so bail.
9443 // Map this input with the i16 shuffle.
9444 PreDupI16Shuffle[j] = MovingInputs[i] / 2;
9447 // Update the lane map based on the mapping we ended up with.
9448 LaneMap[MovingInputs[i]] = 2 * j + MovingInputs[i] % 2;
9450 V1 = DAG.getBitcast(
9452 DAG.getVectorShuffle(MVT::v8i16, DL, DAG.getBitcast(MVT::v8i16, V1),
9453 DAG.getUNDEF(MVT::v8i16), PreDupI16Shuffle));
9455 // Unpack the bytes to form the i16s that will be shuffled into place.
9456 V1 = DAG.getNode(TargetLo ? X86ISD::UNPCKL : X86ISD::UNPCKH, DL,
9457 MVT::v16i8, V1, V1);
9459 int PostDupI16Shuffle[8] = {-1, -1, -1, -1, -1, -1, -1, -1};
9460 for (int i = 0; i < 16; ++i)
9461 if (Mask[i] != -1) {
9462 int MappedMask = LaneMap[Mask[i]] - (TargetLo ? 0 : 8);
9463 assert(MappedMask < 8 && "Invalid v8 shuffle mask!");
9464 if (PostDupI16Shuffle[i / 2] == -1)
9465 PostDupI16Shuffle[i / 2] = MappedMask;
9467 assert(PostDupI16Shuffle[i / 2] == MappedMask &&
9468 "Conflicting entrties in the original shuffle!");
9470 return DAG.getBitcast(
9472 DAG.getVectorShuffle(MVT::v8i16, DL, DAG.getBitcast(MVT::v8i16, V1),
9473 DAG.getUNDEF(MVT::v8i16), PostDupI16Shuffle));
9475 if (SDValue V = tryToWidenViaDuplication())
9479 if (SDValue Masked =
9480 lowerVectorShuffleAsBitMask(DL, MVT::v16i8, V1, V2, Mask, DAG))
9483 // Use dedicated unpack instructions for masks that match their pattern.
9485 lowerVectorShuffleWithUNPCK(DL, MVT::v16i8, Mask, V1, V2, DAG))
9488 // Check for SSSE3 which lets us lower all v16i8 shuffles much more directly
9489 // with PSHUFB. It is important to do this before we attempt to generate any
9490 // blends but after all of the single-input lowerings. If the single input
9491 // lowerings can find an instruction sequence that is faster than a PSHUFB, we
9492 // want to preserve that and we can DAG combine any longer sequences into
9493 // a PSHUFB in the end. But once we start blending from multiple inputs,
9494 // the complexity of DAG combining bad patterns back into PSHUFB is too high,
9495 // and there are *very* few patterns that would actually be faster than the
9496 // PSHUFB approach because of its ability to zero lanes.
9498 // FIXME: The only exceptions to the above are blends which are exact
9499 // interleavings with direct instructions supporting them. We currently don't
9500 // handle those well here.
9501 if (Subtarget->hasSSSE3()) {
9502 bool V1InUse = false;
9503 bool V2InUse = false;
9505 SDValue PSHUFB = lowerVectorShuffleAsPSHUFB(DL, MVT::v16i8, V1, V2, Mask,
9506 DAG, V1InUse, V2InUse);
9508 // If both V1 and V2 are in use and we can use a direct blend or an unpack,
9509 // do so. This avoids using them to handle blends-with-zero which is
9510 // important as a single pshufb is significantly faster for that.
9511 if (V1InUse && V2InUse) {
9512 if (Subtarget->hasSSE41())
9513 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v16i8, V1, V2,
9514 Mask, Subtarget, DAG))
9517 // We can use an unpack to do the blending rather than an or in some
9518 // cases. Even though the or may be (very minorly) more efficient, we
9519 // preference this lowering because there are common cases where part of
9520 // the complexity of the shuffles goes away when we do the final blend as
9522 // FIXME: It might be worth trying to detect if the unpack-feeding
9523 // shuffles will both be pshufb, in which case we shouldn't bother with
9525 if (SDValue Unpack = lowerVectorShuffleAsPermuteAndUnpack(
9526 DL, MVT::v16i8, V1, V2, Mask, DAG))
9533 // There are special ways we can lower some single-element blends.
9534 if (NumV2Elements == 1)
9535 if (SDValue V = lowerVectorShuffleAsElementInsertion(DL, MVT::v16i8, V1, V2,
9536 Mask, Subtarget, DAG))
9539 if (SDValue BitBlend =
9540 lowerVectorShuffleAsBitBlend(DL, MVT::v16i8, V1, V2, Mask, DAG))
9543 // Check whether a compaction lowering can be done. This handles shuffles
9544 // which take every Nth element for some even N. See the helper function for
9547 // We special case these as they can be particularly efficiently handled with
9548 // the PACKUSB instruction on x86 and they show up in common patterns of
9549 // rearranging bytes to truncate wide elements.
9550 if (int NumEvenDrops = canLowerByDroppingEvenElements(Mask)) {
9551 // NumEvenDrops is the power of two stride of the elements. Another way of
9552 // thinking about it is that we need to drop the even elements this many
9553 // times to get the original input.
9554 bool IsSingleInput = isSingleInputShuffleMask(Mask);
9556 // First we need to zero all the dropped bytes.
9557 assert(NumEvenDrops <= 3 &&
9558 "No support for dropping even elements more than 3 times.");
9559 // We use the mask type to pick which bytes are preserved based on how many
9560 // elements are dropped.
9561 MVT MaskVTs[] = { MVT::v8i16, MVT::v4i32, MVT::v2i64 };
9562 SDValue ByteClearMask = DAG.getBitcast(
9563 MVT::v16i8, DAG.getConstant(0xFF, DL, MaskVTs[NumEvenDrops - 1]));
9564 V1 = DAG.getNode(ISD::AND, DL, MVT::v16i8, V1, ByteClearMask);
9566 V2 = DAG.getNode(ISD::AND, DL, MVT::v16i8, V2, ByteClearMask);
9568 // Now pack things back together.
9569 V1 = DAG.getBitcast(MVT::v8i16, V1);
9570 V2 = IsSingleInput ? V1 : DAG.getBitcast(MVT::v8i16, V2);
9571 SDValue Result = DAG.getNode(X86ISD::PACKUS, DL, MVT::v16i8, V1, V2);
9572 for (int i = 1; i < NumEvenDrops; ++i) {
9573 Result = DAG.getBitcast(MVT::v8i16, Result);
9574 Result = DAG.getNode(X86ISD::PACKUS, DL, MVT::v16i8, Result, Result);
9580 // Handle multi-input cases by blending single-input shuffles.
9581 if (NumV2Elements > 0)
9582 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v16i8, V1, V2,
9585 // The fallback path for single-input shuffles widens this into two v8i16
9586 // vectors with unpacks, shuffles those, and then pulls them back together
9590 int LoBlendMask[8] = {-1, -1, -1, -1, -1, -1, -1, -1};
9591 int HiBlendMask[8] = {-1, -1, -1, -1, -1, -1, -1, -1};
9592 for (int i = 0; i < 16; ++i)
9594 (i < 8 ? LoBlendMask[i] : HiBlendMask[i % 8]) = Mask[i];
9596 SDValue Zero = getZeroVector(MVT::v8i16, Subtarget, DAG, DL);
9598 SDValue VLoHalf, VHiHalf;
9599 // Check if any of the odd lanes in the v16i8 are used. If not, we can mask
9600 // them out and avoid using UNPCK{L,H} to extract the elements of V as
9602 if (std::none_of(std::begin(LoBlendMask), std::end(LoBlendMask),
9603 [](int M) { return M >= 0 && M % 2 == 1; }) &&
9604 std::none_of(std::begin(HiBlendMask), std::end(HiBlendMask),
9605 [](int M) { return M >= 0 && M % 2 == 1; })) {
9606 // Use a mask to drop the high bytes.
9607 VLoHalf = DAG.getBitcast(MVT::v8i16, V);
9608 VLoHalf = DAG.getNode(ISD::AND, DL, MVT::v8i16, VLoHalf,
9609 DAG.getConstant(0x00FF, DL, MVT::v8i16));
9611 // This will be a single vector shuffle instead of a blend so nuke VHiHalf.
9612 VHiHalf = DAG.getUNDEF(MVT::v8i16);
9614 // Squash the masks to point directly into VLoHalf.
9615 for (int &M : LoBlendMask)
9618 for (int &M : HiBlendMask)
9622 // Otherwise just unpack the low half of V into VLoHalf and the high half into
9623 // VHiHalf so that we can blend them as i16s.
9624 VLoHalf = DAG.getBitcast(
9625 MVT::v8i16, DAG.getNode(X86ISD::UNPCKL, DL, MVT::v16i8, V, Zero));
9626 VHiHalf = DAG.getBitcast(
9627 MVT::v8i16, DAG.getNode(X86ISD::UNPCKH, DL, MVT::v16i8, V, Zero));
9630 SDValue LoV = DAG.getVectorShuffle(MVT::v8i16, DL, VLoHalf, VHiHalf, LoBlendMask);
9631 SDValue HiV = DAG.getVectorShuffle(MVT::v8i16, DL, VLoHalf, VHiHalf, HiBlendMask);
9633 return DAG.getNode(X86ISD::PACKUS, DL, MVT::v16i8, LoV, HiV);
9636 /// \brief Dispatching routine to lower various 128-bit x86 vector shuffles.
9638 /// This routine breaks down the specific type of 128-bit shuffle and
9639 /// dispatches to the lowering routines accordingly.
9640 static SDValue lower128BitVectorShuffle(SDValue Op, SDValue V1, SDValue V2,
9641 MVT VT, const X86Subtarget *Subtarget,
9642 SelectionDAG &DAG) {
9643 switch (VT.SimpleTy) {
9645 return lowerV2I64VectorShuffle(Op, V1, V2, Subtarget, DAG);
9647 return lowerV2F64VectorShuffle(Op, V1, V2, Subtarget, DAG);
9649 return lowerV4I32VectorShuffle(Op, V1, V2, Subtarget, DAG);
9651 return lowerV4F32VectorShuffle(Op, V1, V2, Subtarget, DAG);
9653 return lowerV8I16VectorShuffle(Op, V1, V2, Subtarget, DAG);
9655 return lowerV16I8VectorShuffle(Op, V1, V2, Subtarget, DAG);
9658 llvm_unreachable("Unimplemented!");
9662 /// \brief Helper function to test whether a shuffle mask could be
9663 /// simplified by widening the elements being shuffled.
9665 /// Appends the mask for wider elements in WidenedMask if valid. Otherwise
9666 /// leaves it in an unspecified state.
9668 /// NOTE: This must handle normal vector shuffle masks and *target* vector
9669 /// shuffle masks. The latter have the special property of a '-2' representing
9670 /// a zero-ed lane of a vector.
9671 static bool canWidenShuffleElements(ArrayRef<int> Mask,
9672 SmallVectorImpl<int> &WidenedMask) {
9673 for (int i = 0, Size = Mask.size(); i < Size; i += 2) {
9674 // If both elements are undef, its trivial.
9675 if (Mask[i] == SM_SentinelUndef && Mask[i + 1] == SM_SentinelUndef) {
9676 WidenedMask.push_back(SM_SentinelUndef);
9680 // Check for an undef mask and a mask value properly aligned to fit with
9681 // a pair of values. If we find such a case, use the non-undef mask's value.
9682 if (Mask[i] == SM_SentinelUndef && Mask[i + 1] >= 0 && Mask[i + 1] % 2 == 1) {
9683 WidenedMask.push_back(Mask[i + 1] / 2);
9686 if (Mask[i + 1] == SM_SentinelUndef && Mask[i] >= 0 && Mask[i] % 2 == 0) {
9687 WidenedMask.push_back(Mask[i] / 2);
9691 // When zeroing, we need to spread the zeroing across both lanes to widen.
9692 if (Mask[i] == SM_SentinelZero || Mask[i + 1] == SM_SentinelZero) {
9693 if ((Mask[i] == SM_SentinelZero || Mask[i] == SM_SentinelUndef) &&
9694 (Mask[i + 1] == SM_SentinelZero || Mask[i + 1] == SM_SentinelUndef)) {
9695 WidenedMask.push_back(SM_SentinelZero);
9701 // Finally check if the two mask values are adjacent and aligned with
9703 if (Mask[i] != SM_SentinelUndef && Mask[i] % 2 == 0 && Mask[i] + 1 == Mask[i + 1]) {
9704 WidenedMask.push_back(Mask[i] / 2);
9708 // Otherwise we can't safely widen the elements used in this shuffle.
9711 assert(WidenedMask.size() == Mask.size() / 2 &&
9712 "Incorrect size of mask after widening the elements!");
9717 /// \brief Generic routine to split vector shuffle into half-sized shuffles.
9719 /// This routine just extracts two subvectors, shuffles them independently, and
9720 /// then concatenates them back together. This should work effectively with all
9721 /// AVX vector shuffle types.
9722 static SDValue splitAndLowerVectorShuffle(SDLoc DL, MVT VT, SDValue V1,
9723 SDValue V2, ArrayRef<int> Mask,
9724 SelectionDAG &DAG) {
9725 assert(VT.getSizeInBits() >= 256 &&
9726 "Only for 256-bit or wider vector shuffles!");
9727 assert(V1.getSimpleValueType() == VT && "Bad operand type!");
9728 assert(V2.getSimpleValueType() == VT && "Bad operand type!");
9730 ArrayRef<int> LoMask = Mask.slice(0, Mask.size() / 2);
9731 ArrayRef<int> HiMask = Mask.slice(Mask.size() / 2);
9733 int NumElements = VT.getVectorNumElements();
9734 int SplitNumElements = NumElements / 2;
9735 MVT ScalarVT = VT.getScalarType();
9736 MVT SplitVT = MVT::getVectorVT(ScalarVT, NumElements / 2);
9738 // Rather than splitting build-vectors, just build two narrower build
9739 // vectors. This helps shuffling with splats and zeros.
9740 auto SplitVector = [&](SDValue V) {
9741 while (V.getOpcode() == ISD::BITCAST)
9742 V = V->getOperand(0);
9744 MVT OrigVT = V.getSimpleValueType();
9745 int OrigNumElements = OrigVT.getVectorNumElements();
9746 int OrigSplitNumElements = OrigNumElements / 2;
9747 MVT OrigScalarVT = OrigVT.getScalarType();
9748 MVT OrigSplitVT = MVT::getVectorVT(OrigScalarVT, OrigNumElements / 2);
9752 auto *BV = dyn_cast<BuildVectorSDNode>(V);
9754 LoV = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, OrigSplitVT, V,
9755 DAG.getIntPtrConstant(0, DL));
9756 HiV = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, OrigSplitVT, V,
9757 DAG.getIntPtrConstant(OrigSplitNumElements, DL));
9760 SmallVector<SDValue, 16> LoOps, HiOps;
9761 for (int i = 0; i < OrigSplitNumElements; ++i) {
9762 LoOps.push_back(BV->getOperand(i));
9763 HiOps.push_back(BV->getOperand(i + OrigSplitNumElements));
9765 LoV = DAG.getNode(ISD::BUILD_VECTOR, DL, OrigSplitVT, LoOps);
9766 HiV = DAG.getNode(ISD::BUILD_VECTOR, DL, OrigSplitVT, HiOps);
9768 return std::make_pair(DAG.getBitcast(SplitVT, LoV),
9769 DAG.getBitcast(SplitVT, HiV));
9772 SDValue LoV1, HiV1, LoV2, HiV2;
9773 std::tie(LoV1, HiV1) = SplitVector(V1);
9774 std::tie(LoV2, HiV2) = SplitVector(V2);
9776 // Now create two 4-way blends of these half-width vectors.
9777 auto HalfBlend = [&](ArrayRef<int> HalfMask) {
9778 bool UseLoV1 = false, UseHiV1 = false, UseLoV2 = false, UseHiV2 = false;
9779 SmallVector<int, 32> V1BlendMask, V2BlendMask, BlendMask;
9780 for (int i = 0; i < SplitNumElements; ++i) {
9781 int M = HalfMask[i];
9782 if (M >= NumElements) {
9783 if (M >= NumElements + SplitNumElements)
9787 V2BlendMask.push_back(M - NumElements);
9788 V1BlendMask.push_back(-1);
9789 BlendMask.push_back(SplitNumElements + i);
9790 } else if (M >= 0) {
9791 if (M >= SplitNumElements)
9795 V2BlendMask.push_back(-1);
9796 V1BlendMask.push_back(M);
9797 BlendMask.push_back(i);
9799 V2BlendMask.push_back(-1);
9800 V1BlendMask.push_back(-1);
9801 BlendMask.push_back(-1);
9805 // Because the lowering happens after all combining takes place, we need to
9806 // manually combine these blend masks as much as possible so that we create
9807 // a minimal number of high-level vector shuffle nodes.
9809 // First try just blending the halves of V1 or V2.
9810 if (!UseLoV1 && !UseHiV1 && !UseLoV2 && !UseHiV2)
9811 return DAG.getUNDEF(SplitVT);
9812 if (!UseLoV2 && !UseHiV2)
9813 return DAG.getVectorShuffle(SplitVT, DL, LoV1, HiV1, V1BlendMask);
9814 if (!UseLoV1 && !UseHiV1)
9815 return DAG.getVectorShuffle(SplitVT, DL, LoV2, HiV2, V2BlendMask);
9817 SDValue V1Blend, V2Blend;
9818 if (UseLoV1 && UseHiV1) {
9820 DAG.getVectorShuffle(SplitVT, DL, LoV1, HiV1, V1BlendMask);
9822 // We only use half of V1 so map the usage down into the final blend mask.
9823 V1Blend = UseLoV1 ? LoV1 : HiV1;
9824 for (int i = 0; i < SplitNumElements; ++i)
9825 if (BlendMask[i] >= 0 && BlendMask[i] < SplitNumElements)
9826 BlendMask[i] = V1BlendMask[i] - (UseLoV1 ? 0 : SplitNumElements);
9828 if (UseLoV2 && UseHiV2) {
9830 DAG.getVectorShuffle(SplitVT, DL, LoV2, HiV2, V2BlendMask);
9832 // We only use half of V2 so map the usage down into the final blend mask.
9833 V2Blend = UseLoV2 ? LoV2 : HiV2;
9834 for (int i = 0; i < SplitNumElements; ++i)
9835 if (BlendMask[i] >= SplitNumElements)
9836 BlendMask[i] = V2BlendMask[i] + (UseLoV2 ? SplitNumElements : 0);
9838 return DAG.getVectorShuffle(SplitVT, DL, V1Blend, V2Blend, BlendMask);
9840 SDValue Lo = HalfBlend(LoMask);
9841 SDValue Hi = HalfBlend(HiMask);
9842 return DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, Lo, Hi);
9845 /// \brief Either split a vector in halves or decompose the shuffles and the
9848 /// This is provided as a good fallback for many lowerings of non-single-input
9849 /// shuffles with more than one 128-bit lane. In those cases, we want to select
9850 /// between splitting the shuffle into 128-bit components and stitching those
9851 /// back together vs. extracting the single-input shuffles and blending those
9853 static SDValue lowerVectorShuffleAsSplitOrBlend(SDLoc DL, MVT VT, SDValue V1,
9854 SDValue V2, ArrayRef<int> Mask,
9855 SelectionDAG &DAG) {
9856 assert(!isSingleInputShuffleMask(Mask) && "This routine must not be used to "
9857 "lower single-input shuffles as it "
9858 "could then recurse on itself.");
9859 int Size = Mask.size();
9861 // If this can be modeled as a broadcast of two elements followed by a blend,
9862 // prefer that lowering. This is especially important because broadcasts can
9863 // often fold with memory operands.
9864 auto DoBothBroadcast = [&] {
9865 int V1BroadcastIdx = -1, V2BroadcastIdx = -1;
9868 if (V2BroadcastIdx == -1)
9869 V2BroadcastIdx = M - Size;
9870 else if (M - Size != V2BroadcastIdx)
9872 } else if (M >= 0) {
9873 if (V1BroadcastIdx == -1)
9875 else if (M != V1BroadcastIdx)
9880 if (DoBothBroadcast())
9881 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, VT, V1, V2, Mask,
9884 // If the inputs all stem from a single 128-bit lane of each input, then we
9885 // split them rather than blending because the split will decompose to
9886 // unusually few instructions.
9887 int LaneCount = VT.getSizeInBits() / 128;
9888 int LaneSize = Size / LaneCount;
9889 SmallBitVector LaneInputs[2];
9890 LaneInputs[0].resize(LaneCount, false);
9891 LaneInputs[1].resize(LaneCount, false);
9892 for (int i = 0; i < Size; ++i)
9894 LaneInputs[Mask[i] / Size][(Mask[i] % Size) / LaneSize] = true;
9895 if (LaneInputs[0].count() <= 1 && LaneInputs[1].count() <= 1)
9896 return splitAndLowerVectorShuffle(DL, VT, V1, V2, Mask, DAG);
9898 // Otherwise, just fall back to decomposed shuffles and a blend. This requires
9899 // that the decomposed single-input shuffles don't end up here.
9900 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, VT, V1, V2, Mask, DAG);
9903 /// \brief Lower a vector shuffle crossing multiple 128-bit lanes as
9904 /// a permutation and blend of those lanes.
9906 /// This essentially blends the out-of-lane inputs to each lane into the lane
9907 /// from a permuted copy of the vector. This lowering strategy results in four
9908 /// instructions in the worst case for a single-input cross lane shuffle which
9909 /// is lower than any other fully general cross-lane shuffle strategy I'm aware
9910 /// of. Special cases for each particular shuffle pattern should be handled
9911 /// prior to trying this lowering.
9912 static SDValue lowerVectorShuffleAsLanePermuteAndBlend(SDLoc DL, MVT VT,
9913 SDValue V1, SDValue V2,
9915 SelectionDAG &DAG) {
9916 // FIXME: This should probably be generalized for 512-bit vectors as well.
9917 assert(VT.getSizeInBits() == 256 && "Only for 256-bit vector shuffles!");
9918 int LaneSize = Mask.size() / 2;
9920 // If there are only inputs from one 128-bit lane, splitting will in fact be
9921 // less expensive. The flags track whether the given lane contains an element
9922 // that crosses to another lane.
9923 bool LaneCrossing[2] = {false, false};
9924 for (int i = 0, Size = Mask.size(); i < Size; ++i)
9925 if (Mask[i] >= 0 && (Mask[i] % Size) / LaneSize != i / LaneSize)
9926 LaneCrossing[(Mask[i] % Size) / LaneSize] = true;
9927 if (!LaneCrossing[0] || !LaneCrossing[1])
9928 return splitAndLowerVectorShuffle(DL, VT, V1, V2, Mask, DAG);
9930 if (isSingleInputShuffleMask(Mask)) {
9931 SmallVector<int, 32> FlippedBlendMask;
9932 for (int i = 0, Size = Mask.size(); i < Size; ++i)
9933 FlippedBlendMask.push_back(
9934 Mask[i] < 0 ? -1 : (((Mask[i] % Size) / LaneSize == i / LaneSize)
9936 : Mask[i] % LaneSize +
9937 (i / LaneSize) * LaneSize + Size));
9939 // Flip the vector, and blend the results which should now be in-lane. The
9940 // VPERM2X128 mask uses the low 2 bits for the low source and bits 4 and
9941 // 5 for the high source. The value 3 selects the high half of source 2 and
9942 // the value 2 selects the low half of source 2. We only use source 2 to
9943 // allow folding it into a memory operand.
9944 unsigned PERMMask = 3 | 2 << 4;
9945 SDValue Flipped = DAG.getNode(X86ISD::VPERM2X128, DL, VT, DAG.getUNDEF(VT),
9946 V1, DAG.getConstant(PERMMask, DL, MVT::i8));
9947 return DAG.getVectorShuffle(VT, DL, V1, Flipped, FlippedBlendMask);
9950 // This now reduces to two single-input shuffles of V1 and V2 which at worst
9951 // will be handled by the above logic and a blend of the results, much like
9952 // other patterns in AVX.
9953 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, VT, V1, V2, Mask, DAG);
9956 /// \brief Handle lowering 2-lane 128-bit shuffles.
9957 static SDValue lowerV2X128VectorShuffle(SDLoc DL, MVT VT, SDValue V1,
9958 SDValue V2, ArrayRef<int> Mask,
9959 const X86Subtarget *Subtarget,
9960 SelectionDAG &DAG) {
9961 // TODO: If minimizing size and one of the inputs is a zero vector and the
9962 // the zero vector has only one use, we could use a VPERM2X128 to save the
9963 // instruction bytes needed to explicitly generate the zero vector.
9965 // Blends are faster and handle all the non-lane-crossing cases.
9966 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, VT, V1, V2, Mask,
9970 bool IsV1Zero = ISD::isBuildVectorAllZeros(V1.getNode());
9971 bool IsV2Zero = ISD::isBuildVectorAllZeros(V2.getNode());
9973 // If either input operand is a zero vector, use VPERM2X128 because its mask
9974 // allows us to replace the zero input with an implicit zero.
9975 if (!IsV1Zero && !IsV2Zero) {
9976 // Check for patterns which can be matched with a single insert of a 128-bit
9978 bool OnlyUsesV1 = isShuffleEquivalent(V1, V2, Mask, {0, 1, 0, 1});
9979 if (OnlyUsesV1 || isShuffleEquivalent(V1, V2, Mask, {0, 1, 4, 5})) {
9980 MVT SubVT = MVT::getVectorVT(VT.getVectorElementType(),
9981 VT.getVectorNumElements() / 2);
9982 SDValue LoV = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, SubVT, V1,
9983 DAG.getIntPtrConstant(0, DL));
9984 SDValue HiV = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, SubVT,
9985 OnlyUsesV1 ? V1 : V2,
9986 DAG.getIntPtrConstant(0, DL));
9987 return DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, LoV, HiV);
9991 // Otherwise form a 128-bit permutation. After accounting for undefs,
9992 // convert the 64-bit shuffle mask selection values into 128-bit
9993 // selection bits by dividing the indexes by 2 and shifting into positions
9994 // defined by a vperm2*128 instruction's immediate control byte.
9996 // The immediate permute control byte looks like this:
9997 // [1:0] - select 128 bits from sources for low half of destination
9999 // [3] - zero low half of destination
10000 // [5:4] - select 128 bits from sources for high half of destination
10002 // [7] - zero high half of destination
10004 int MaskLO = Mask[0];
10005 if (MaskLO == SM_SentinelUndef)
10006 MaskLO = Mask[1] == SM_SentinelUndef ? 0 : Mask[1];
10008 int MaskHI = Mask[2];
10009 if (MaskHI == SM_SentinelUndef)
10010 MaskHI = Mask[3] == SM_SentinelUndef ? 0 : Mask[3];
10012 unsigned PermMask = MaskLO / 2 | (MaskHI / 2) << 4;
10014 // If either input is a zero vector, replace it with an undef input.
10015 // Shuffle mask values < 4 are selecting elements of V1.
10016 // Shuffle mask values >= 4 are selecting elements of V2.
10017 // Adjust each half of the permute mask by clearing the half that was
10018 // selecting the zero vector and setting the zero mask bit.
10020 V1 = DAG.getUNDEF(VT);
10022 PermMask = (PermMask & 0xf0) | 0x08;
10024 PermMask = (PermMask & 0x0f) | 0x80;
10027 V2 = DAG.getUNDEF(VT);
10029 PermMask = (PermMask & 0xf0) | 0x08;
10031 PermMask = (PermMask & 0x0f) | 0x80;
10034 return DAG.getNode(X86ISD::VPERM2X128, DL, VT, V1, V2,
10035 DAG.getConstant(PermMask, DL, MVT::i8));
10038 /// \brief Lower a vector shuffle by first fixing the 128-bit lanes and then
10039 /// shuffling each lane.
10041 /// This will only succeed when the result of fixing the 128-bit lanes results
10042 /// in a single-input non-lane-crossing shuffle with a repeating shuffle mask in
10043 /// each 128-bit lanes. This handles many cases where we can quickly blend away
10044 /// the lane crosses early and then use simpler shuffles within each lane.
10046 /// FIXME: It might be worthwhile at some point to support this without
10047 /// requiring the 128-bit lane-relative shuffles to be repeating, but currently
10048 /// in x86 only floating point has interesting non-repeating shuffles, and even
10049 /// those are still *marginally* more expensive.
10050 static SDValue lowerVectorShuffleByMerging128BitLanes(
10051 SDLoc DL, MVT VT, SDValue V1, SDValue V2, ArrayRef<int> Mask,
10052 const X86Subtarget *Subtarget, SelectionDAG &DAG) {
10053 assert(!isSingleInputShuffleMask(Mask) &&
10054 "This is only useful with multiple inputs.");
10056 int Size = Mask.size();
10057 int LaneSize = 128 / VT.getScalarSizeInBits();
10058 int NumLanes = Size / LaneSize;
10059 assert(NumLanes > 1 && "Only handles 256-bit and wider shuffles.");
10061 // See if we can build a hypothetical 128-bit lane-fixing shuffle mask. Also
10062 // check whether the in-128-bit lane shuffles share a repeating pattern.
10063 SmallVector<int, 4> Lanes;
10064 Lanes.resize(NumLanes, -1);
10065 SmallVector<int, 4> InLaneMask;
10066 InLaneMask.resize(LaneSize, -1);
10067 for (int i = 0; i < Size; ++i) {
10071 int j = i / LaneSize;
10073 if (Lanes[j] < 0) {
10074 // First entry we've seen for this lane.
10075 Lanes[j] = Mask[i] / LaneSize;
10076 } else if (Lanes[j] != Mask[i] / LaneSize) {
10077 // This doesn't match the lane selected previously!
10081 // Check that within each lane we have a consistent shuffle mask.
10082 int k = i % LaneSize;
10083 if (InLaneMask[k] < 0) {
10084 InLaneMask[k] = Mask[i] % LaneSize;
10085 } else if (InLaneMask[k] != Mask[i] % LaneSize) {
10086 // This doesn't fit a repeating in-lane mask.
10091 // First shuffle the lanes into place.
10092 MVT LaneVT = MVT::getVectorVT(VT.isFloatingPoint() ? MVT::f64 : MVT::i64,
10093 VT.getSizeInBits() / 64);
10094 SmallVector<int, 8> LaneMask;
10095 LaneMask.resize(NumLanes * 2, -1);
10096 for (int i = 0; i < NumLanes; ++i)
10097 if (Lanes[i] >= 0) {
10098 LaneMask[2 * i + 0] = 2*Lanes[i] + 0;
10099 LaneMask[2 * i + 1] = 2*Lanes[i] + 1;
10102 V1 = DAG.getBitcast(LaneVT, V1);
10103 V2 = DAG.getBitcast(LaneVT, V2);
10104 SDValue LaneShuffle = DAG.getVectorShuffle(LaneVT, DL, V1, V2, LaneMask);
10106 // Cast it back to the type we actually want.
10107 LaneShuffle = DAG.getBitcast(VT, LaneShuffle);
10109 // Now do a simple shuffle that isn't lane crossing.
10110 SmallVector<int, 8> NewMask;
10111 NewMask.resize(Size, -1);
10112 for (int i = 0; i < Size; ++i)
10114 NewMask[i] = (i / LaneSize) * LaneSize + Mask[i] % LaneSize;
10115 assert(!is128BitLaneCrossingShuffleMask(VT, NewMask) &&
10116 "Must not introduce lane crosses at this point!");
10118 return DAG.getVectorShuffle(VT, DL, LaneShuffle, DAG.getUNDEF(VT), NewMask);
10121 /// \brief Test whether the specified input (0 or 1) is in-place blended by the
10124 /// This returns true if the elements from a particular input are already in the
10125 /// slot required by the given mask and require no permutation.
10126 static bool isShuffleMaskInputInPlace(int Input, ArrayRef<int> Mask) {
10127 assert((Input == 0 || Input == 1) && "Only two inputs to shuffles.");
10128 int Size = Mask.size();
10129 for (int i = 0; i < Size; ++i)
10130 if (Mask[i] >= 0 && Mask[i] / Size == Input && Mask[i] % Size != i)
10136 static SDValue lowerVectorShuffleWithSHUFPD(SDLoc DL, MVT VT,
10137 ArrayRef<int> Mask, SDValue V1,
10138 SDValue V2, SelectionDAG &DAG) {
10140 // Mask for V8F64: 0/1, 8/9, 2/3, 10/11, 4/5, ..
10141 // Mask for V4F64; 0/1, 4/5, 2/3, 6/7..
10142 assert(VT.getScalarSizeInBits() == 64 && "Unexpected data type for VSHUFPD");
10143 int NumElts = VT.getVectorNumElements();
10144 bool ShufpdMask = true;
10145 bool CommutableMask = true;
10146 unsigned Immediate = 0;
10147 for (int i = 0; i < NumElts; ++i) {
10150 int Val = (i & 6) + NumElts * (i & 1);
10151 int CommutVal = (i & 0xe) + NumElts * ((i & 1)^1);
10152 if (Mask[i] < Val || Mask[i] > Val + 1)
10153 ShufpdMask = false;
10154 if (Mask[i] < CommutVal || Mask[i] > CommutVal + 1)
10155 CommutableMask = false;
10156 Immediate |= (Mask[i] % 2) << i;
10159 return DAG.getNode(X86ISD::SHUFP, DL, VT, V1, V2,
10160 DAG.getConstant(Immediate, DL, MVT::i8));
10161 if (CommutableMask)
10162 return DAG.getNode(X86ISD::SHUFP, DL, VT, V2, V1,
10163 DAG.getConstant(Immediate, DL, MVT::i8));
10167 /// \brief Handle lowering of 4-lane 64-bit floating point shuffles.
10169 /// Also ends up handling lowering of 4-lane 64-bit integer shuffles when AVX2
10170 /// isn't available.
10171 static SDValue lowerV4F64VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10172 const X86Subtarget *Subtarget,
10173 SelectionDAG &DAG) {
10175 assert(V1.getSimpleValueType() == MVT::v4f64 && "Bad operand type!");
10176 assert(V2.getSimpleValueType() == MVT::v4f64 && "Bad operand type!");
10177 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10178 ArrayRef<int> Mask = SVOp->getMask();
10179 assert(Mask.size() == 4 && "Unexpected mask size for v4 shuffle!");
10181 SmallVector<int, 4> WidenedMask;
10182 if (canWidenShuffleElements(Mask, WidenedMask))
10183 return lowerV2X128VectorShuffle(DL, MVT::v4f64, V1, V2, Mask, Subtarget,
10186 if (isSingleInputShuffleMask(Mask)) {
10187 // Check for being able to broadcast a single element.
10188 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v4f64, V1,
10189 Mask, Subtarget, DAG))
10192 // Use low duplicate instructions for masks that match their pattern.
10193 if (isShuffleEquivalent(V1, V2, Mask, {0, 0, 2, 2}))
10194 return DAG.getNode(X86ISD::MOVDDUP, DL, MVT::v4f64, V1);
10196 if (!is128BitLaneCrossingShuffleMask(MVT::v4f64, Mask)) {
10197 // Non-half-crossing single input shuffles can be lowerid with an
10198 // interleaved permutation.
10199 unsigned VPERMILPMask = (Mask[0] == 1) | ((Mask[1] == 1) << 1) |
10200 ((Mask[2] == 3) << 2) | ((Mask[3] == 3) << 3);
10201 return DAG.getNode(X86ISD::VPERMILPI, DL, MVT::v4f64, V1,
10202 DAG.getConstant(VPERMILPMask, DL, MVT::i8));
10205 // With AVX2 we have direct support for this permutation.
10206 if (Subtarget->hasAVX2())
10207 return DAG.getNode(X86ISD::VPERMI, DL, MVT::v4f64, V1,
10208 getV4X86ShuffleImm8ForMask(Mask, DL, DAG));
10210 // Otherwise, fall back.
10211 return lowerVectorShuffleAsLanePermuteAndBlend(DL, MVT::v4f64, V1, V2, Mask,
10215 // Use dedicated unpack instructions for masks that match their pattern.
10217 lowerVectorShuffleWithUNPCK(DL, MVT::v4f64, Mask, V1, V2, DAG))
10220 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v4f64, V1, V2, Mask,
10224 // Check if the blend happens to exactly fit that of SHUFPD.
10226 lowerVectorShuffleWithSHUFPD(DL, MVT::v4f64, Mask, V1, V2, DAG))
10229 // Try to simplify this by merging 128-bit lanes to enable a lane-based
10230 // shuffle. However, if we have AVX2 and either inputs are already in place,
10231 // we will be able to shuffle even across lanes the other input in a single
10232 // instruction so skip this pattern.
10233 if (!(Subtarget->hasAVX2() && (isShuffleMaskInputInPlace(0, Mask) ||
10234 isShuffleMaskInputInPlace(1, Mask))))
10235 if (SDValue Result = lowerVectorShuffleByMerging128BitLanes(
10236 DL, MVT::v4f64, V1, V2, Mask, Subtarget, DAG))
10239 // If we have AVX2 then we always want to lower with a blend because an v4 we
10240 // can fully permute the elements.
10241 if (Subtarget->hasAVX2())
10242 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v4f64, V1, V2,
10245 // Otherwise fall back on generic lowering.
10246 return lowerVectorShuffleAsSplitOrBlend(DL, MVT::v4f64, V1, V2, Mask, DAG);
10249 /// \brief Handle lowering of 4-lane 64-bit integer shuffles.
10251 /// This routine is only called when we have AVX2 and thus a reasonable
10252 /// instruction set for v4i64 shuffling..
10253 static SDValue lowerV4I64VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10254 const X86Subtarget *Subtarget,
10255 SelectionDAG &DAG) {
10257 assert(V1.getSimpleValueType() == MVT::v4i64 && "Bad operand type!");
10258 assert(V2.getSimpleValueType() == MVT::v4i64 && "Bad operand type!");
10259 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10260 ArrayRef<int> Mask = SVOp->getMask();
10261 assert(Mask.size() == 4 && "Unexpected mask size for v4 shuffle!");
10262 assert(Subtarget->hasAVX2() && "We can only lower v4i64 with AVX2!");
10264 SmallVector<int, 4> WidenedMask;
10265 if (canWidenShuffleElements(Mask, WidenedMask))
10266 return lowerV2X128VectorShuffle(DL, MVT::v4i64, V1, V2, Mask, Subtarget,
10269 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v4i64, V1, V2, Mask,
10273 // Check for being able to broadcast a single element.
10274 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v4i64, V1,
10275 Mask, Subtarget, DAG))
10278 // When the shuffle is mirrored between the 128-bit lanes of the unit, we can
10279 // use lower latency instructions that will operate on both 128-bit lanes.
10280 SmallVector<int, 2> RepeatedMask;
10281 if (is128BitLaneRepeatedShuffleMask(MVT::v4i64, Mask, RepeatedMask)) {
10282 if (isSingleInputShuffleMask(Mask)) {
10283 int PSHUFDMask[] = {-1, -1, -1, -1};
10284 for (int i = 0; i < 2; ++i)
10285 if (RepeatedMask[i] >= 0) {
10286 PSHUFDMask[2 * i] = 2 * RepeatedMask[i];
10287 PSHUFDMask[2 * i + 1] = 2 * RepeatedMask[i] + 1;
10289 return DAG.getBitcast(
10291 DAG.getNode(X86ISD::PSHUFD, DL, MVT::v8i32,
10292 DAG.getBitcast(MVT::v8i32, V1),
10293 getV4X86ShuffleImm8ForMask(PSHUFDMask, DL, DAG)));
10297 // AVX2 provides a direct instruction for permuting a single input across
10299 if (isSingleInputShuffleMask(Mask))
10300 return DAG.getNode(X86ISD::VPERMI, DL, MVT::v4i64, V1,
10301 getV4X86ShuffleImm8ForMask(Mask, DL, DAG));
10303 // Try to use shift instructions.
10304 if (SDValue Shift =
10305 lowerVectorShuffleAsShift(DL, MVT::v4i64, V1, V2, Mask, DAG))
10308 // Use dedicated unpack instructions for masks that match their pattern.
10310 lowerVectorShuffleWithUNPCK(DL, MVT::v4i64, Mask, V1, V2, DAG))
10313 // Try to simplify this by merging 128-bit lanes to enable a lane-based
10314 // shuffle. However, if we have AVX2 and either inputs are already in place,
10315 // we will be able to shuffle even across lanes the other input in a single
10316 // instruction so skip this pattern.
10317 if (!(Subtarget->hasAVX2() && (isShuffleMaskInputInPlace(0, Mask) ||
10318 isShuffleMaskInputInPlace(1, Mask))))
10319 if (SDValue Result = lowerVectorShuffleByMerging128BitLanes(
10320 DL, MVT::v4i64, V1, V2, Mask, Subtarget, DAG))
10323 // Otherwise fall back on generic blend lowering.
10324 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v4i64, V1, V2,
10328 /// \brief Handle lowering of 8-lane 32-bit floating point shuffles.
10330 /// Also ends up handling lowering of 8-lane 32-bit integer shuffles when AVX2
10331 /// isn't available.
10332 static SDValue lowerV8F32VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10333 const X86Subtarget *Subtarget,
10334 SelectionDAG &DAG) {
10336 assert(V1.getSimpleValueType() == MVT::v8f32 && "Bad operand type!");
10337 assert(V2.getSimpleValueType() == MVT::v8f32 && "Bad operand type!");
10338 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10339 ArrayRef<int> Mask = SVOp->getMask();
10340 assert(Mask.size() == 8 && "Unexpected mask size for v8 shuffle!");
10342 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v8f32, V1, V2, Mask,
10346 // Check for being able to broadcast a single element.
10347 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v8f32, V1,
10348 Mask, Subtarget, DAG))
10351 // If the shuffle mask is repeated in each 128-bit lane, we have many more
10352 // options to efficiently lower the shuffle.
10353 SmallVector<int, 4> RepeatedMask;
10354 if (is128BitLaneRepeatedShuffleMask(MVT::v8f32, Mask, RepeatedMask)) {
10355 assert(RepeatedMask.size() == 4 &&
10356 "Repeated masks must be half the mask width!");
10358 // Use even/odd duplicate instructions for masks that match their pattern.
10359 if (isShuffleEquivalent(V1, V2, Mask, {0, 0, 2, 2, 4, 4, 6, 6}))
10360 return DAG.getNode(X86ISD::MOVSLDUP, DL, MVT::v8f32, V1);
10361 if (isShuffleEquivalent(V1, V2, Mask, {1, 1, 3, 3, 5, 5, 7, 7}))
10362 return DAG.getNode(X86ISD::MOVSHDUP, DL, MVT::v8f32, V1);
10364 if (isSingleInputShuffleMask(Mask))
10365 return DAG.getNode(X86ISD::VPERMILPI, DL, MVT::v8f32, V1,
10366 getV4X86ShuffleImm8ForMask(RepeatedMask, DL, DAG));
10368 // Use dedicated unpack instructions for masks that match their pattern.
10370 lowerVectorShuffleWithUNPCK(DL, MVT::v8f32, Mask, V1, V2, DAG))
10373 // Otherwise, fall back to a SHUFPS sequence. Here it is important that we
10374 // have already handled any direct blends. We also need to squash the
10375 // repeated mask into a simulated v4f32 mask.
10376 for (int i = 0; i < 4; ++i)
10377 if (RepeatedMask[i] >= 8)
10378 RepeatedMask[i] -= 4;
10379 return lowerVectorShuffleWithSHUFPS(DL, MVT::v8f32, RepeatedMask, V1, V2, DAG);
10382 // If we have a single input shuffle with different shuffle patterns in the
10383 // two 128-bit lanes use the variable mask to VPERMILPS.
10384 if (isSingleInputShuffleMask(Mask)) {
10385 SDValue VPermMask[8];
10386 for (int i = 0; i < 8; ++i)
10387 VPermMask[i] = Mask[i] < 0 ? DAG.getUNDEF(MVT::i32)
10388 : DAG.getConstant(Mask[i], DL, MVT::i32);
10389 if (!is128BitLaneCrossingShuffleMask(MVT::v8f32, Mask))
10390 return DAG.getNode(
10391 X86ISD::VPERMILPV, DL, MVT::v8f32, V1,
10392 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v8i32, VPermMask));
10394 if (Subtarget->hasAVX2())
10395 return DAG.getNode(
10396 X86ISD::VPERMV, DL, MVT::v8f32,
10397 DAG.getBitcast(MVT::v8f32, DAG.getNode(ISD::BUILD_VECTOR, DL,
10398 MVT::v8i32, VPermMask)),
10401 // Otherwise, fall back.
10402 return lowerVectorShuffleAsLanePermuteAndBlend(DL, MVT::v8f32, V1, V2, Mask,
10406 // Try to simplify this by merging 128-bit lanes to enable a lane-based
10408 if (SDValue Result = lowerVectorShuffleByMerging128BitLanes(
10409 DL, MVT::v8f32, V1, V2, Mask, Subtarget, DAG))
10412 // If we have AVX2 then we always want to lower with a blend because at v8 we
10413 // can fully permute the elements.
10414 if (Subtarget->hasAVX2())
10415 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v8f32, V1, V2,
10418 // Otherwise fall back on generic lowering.
10419 return lowerVectorShuffleAsSplitOrBlend(DL, MVT::v8f32, V1, V2, Mask, DAG);
10422 /// \brief Handle lowering of 8-lane 32-bit integer shuffles.
10424 /// This routine is only called when we have AVX2 and thus a reasonable
10425 /// instruction set for v8i32 shuffling..
10426 static SDValue lowerV8I32VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10427 const X86Subtarget *Subtarget,
10428 SelectionDAG &DAG) {
10430 assert(V1.getSimpleValueType() == MVT::v8i32 && "Bad operand type!");
10431 assert(V2.getSimpleValueType() == MVT::v8i32 && "Bad operand type!");
10432 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10433 ArrayRef<int> Mask = SVOp->getMask();
10434 assert(Mask.size() == 8 && "Unexpected mask size for v8 shuffle!");
10435 assert(Subtarget->hasAVX2() && "We can only lower v8i32 with AVX2!");
10437 // Whenever we can lower this as a zext, that instruction is strictly faster
10438 // than any alternative. It also allows us to fold memory operands into the
10439 // shuffle in many cases.
10440 if (SDValue ZExt = lowerVectorShuffleAsZeroOrAnyExtend(DL, MVT::v8i32, V1, V2,
10441 Mask, Subtarget, DAG))
10444 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v8i32, V1, V2, Mask,
10448 // Check for being able to broadcast a single element.
10449 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v8i32, V1,
10450 Mask, Subtarget, DAG))
10453 // If the shuffle mask is repeated in each 128-bit lane we can use more
10454 // efficient instructions that mirror the shuffles across the two 128-bit
10456 SmallVector<int, 4> RepeatedMask;
10457 if (is128BitLaneRepeatedShuffleMask(MVT::v8i32, Mask, RepeatedMask)) {
10458 assert(RepeatedMask.size() == 4 && "Unexpected repeated mask size!");
10459 if (isSingleInputShuffleMask(Mask))
10460 return DAG.getNode(X86ISD::PSHUFD, DL, MVT::v8i32, V1,
10461 getV4X86ShuffleImm8ForMask(RepeatedMask, DL, DAG));
10463 // Use dedicated unpack instructions for masks that match their pattern.
10465 lowerVectorShuffleWithUNPCK(DL, MVT::v8i32, Mask, V1, V2, DAG))
10469 // Try to use shift instructions.
10470 if (SDValue Shift =
10471 lowerVectorShuffleAsShift(DL, MVT::v8i32, V1, V2, Mask, DAG))
10474 if (SDValue Rotate = lowerVectorShuffleAsByteRotate(
10475 DL, MVT::v8i32, V1, V2, Mask, Subtarget, DAG))
10478 // If the shuffle patterns aren't repeated but it is a single input, directly
10479 // generate a cross-lane VPERMD instruction.
10480 if (isSingleInputShuffleMask(Mask)) {
10481 SDValue VPermMask[8];
10482 for (int i = 0; i < 8; ++i)
10483 VPermMask[i] = Mask[i] < 0 ? DAG.getUNDEF(MVT::i32)
10484 : DAG.getConstant(Mask[i], DL, MVT::i32);
10485 return DAG.getNode(
10486 X86ISD::VPERMV, DL, MVT::v8i32,
10487 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v8i32, VPermMask), V1);
10490 // Try to simplify this by merging 128-bit lanes to enable a lane-based
10492 if (SDValue Result = lowerVectorShuffleByMerging128BitLanes(
10493 DL, MVT::v8i32, V1, V2, Mask, Subtarget, DAG))
10496 // Otherwise fall back on generic blend lowering.
10497 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v8i32, V1, V2,
10501 /// \brief Handle lowering of 16-lane 16-bit integer shuffles.
10503 /// This routine is only called when we have AVX2 and thus a reasonable
10504 /// instruction set for v16i16 shuffling..
10505 static SDValue lowerV16I16VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10506 const X86Subtarget *Subtarget,
10507 SelectionDAG &DAG) {
10509 assert(V1.getSimpleValueType() == MVT::v16i16 && "Bad operand type!");
10510 assert(V2.getSimpleValueType() == MVT::v16i16 && "Bad operand type!");
10511 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10512 ArrayRef<int> Mask = SVOp->getMask();
10513 assert(Mask.size() == 16 && "Unexpected mask size for v16 shuffle!");
10514 assert(Subtarget->hasAVX2() && "We can only lower v16i16 with AVX2!");
10516 // Whenever we can lower this as a zext, that instruction is strictly faster
10517 // than any alternative. It also allows us to fold memory operands into the
10518 // shuffle in many cases.
10519 if (SDValue ZExt = lowerVectorShuffleAsZeroOrAnyExtend(DL, MVT::v16i16, V1, V2,
10520 Mask, Subtarget, DAG))
10523 // Check for being able to broadcast a single element.
10524 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v16i16, V1,
10525 Mask, Subtarget, DAG))
10528 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v16i16, V1, V2, Mask,
10532 // Use dedicated unpack instructions for masks that match their pattern.
10534 lowerVectorShuffleWithUNPCK(DL, MVT::v16i16, Mask, V1, V2, DAG))
10537 // Try to use shift instructions.
10538 if (SDValue Shift =
10539 lowerVectorShuffleAsShift(DL, MVT::v16i16, V1, V2, Mask, DAG))
10542 // Try to use byte rotation instructions.
10543 if (SDValue Rotate = lowerVectorShuffleAsByteRotate(
10544 DL, MVT::v16i16, V1, V2, Mask, Subtarget, DAG))
10547 if (isSingleInputShuffleMask(Mask)) {
10548 // There are no generalized cross-lane shuffle operations available on i16
10550 if (is128BitLaneCrossingShuffleMask(MVT::v16i16, Mask))
10551 return lowerVectorShuffleAsLanePermuteAndBlend(DL, MVT::v16i16, V1, V2,
10554 SmallVector<int, 8> RepeatedMask;
10555 if (is128BitLaneRepeatedShuffleMask(MVT::v16i16, Mask, RepeatedMask)) {
10556 // As this is a single-input shuffle, the repeated mask should be
10557 // a strictly valid v8i16 mask that we can pass through to the v8i16
10558 // lowering to handle even the v16 case.
10559 return lowerV8I16GeneralSingleInputVectorShuffle(
10560 DL, MVT::v16i16, V1, RepeatedMask, Subtarget, DAG);
10563 SDValue PSHUFBMask[32];
10564 for (int i = 0; i < 16; ++i) {
10565 if (Mask[i] == -1) {
10566 PSHUFBMask[2 * i] = PSHUFBMask[2 * i + 1] = DAG.getUNDEF(MVT::i8);
10570 int M = i < 8 ? Mask[i] : Mask[i] - 8;
10571 assert(M >= 0 && M < 8 && "Invalid single-input mask!");
10572 PSHUFBMask[2 * i] = DAG.getConstant(2 * M, DL, MVT::i8);
10573 PSHUFBMask[2 * i + 1] = DAG.getConstant(2 * M + 1, DL, MVT::i8);
10575 return DAG.getBitcast(MVT::v16i16,
10576 DAG.getNode(X86ISD::PSHUFB, DL, MVT::v32i8,
10577 DAG.getBitcast(MVT::v32i8, V1),
10578 DAG.getNode(ISD::BUILD_VECTOR, DL,
10579 MVT::v32i8, PSHUFBMask)));
10582 // Try to simplify this by merging 128-bit lanes to enable a lane-based
10584 if (SDValue Result = lowerVectorShuffleByMerging128BitLanes(
10585 DL, MVT::v16i16, V1, V2, Mask, Subtarget, DAG))
10588 // Otherwise fall back on generic lowering.
10589 return lowerVectorShuffleAsSplitOrBlend(DL, MVT::v16i16, V1, V2, Mask, DAG);
10592 /// \brief Handle lowering of 32-lane 8-bit integer shuffles.
10594 /// This routine is only called when we have AVX2 and thus a reasonable
10595 /// instruction set for v32i8 shuffling..
10596 static SDValue lowerV32I8VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10597 const X86Subtarget *Subtarget,
10598 SelectionDAG &DAG) {
10600 assert(V1.getSimpleValueType() == MVT::v32i8 && "Bad operand type!");
10601 assert(V2.getSimpleValueType() == MVT::v32i8 && "Bad operand type!");
10602 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10603 ArrayRef<int> Mask = SVOp->getMask();
10604 assert(Mask.size() == 32 && "Unexpected mask size for v32 shuffle!");
10605 assert(Subtarget->hasAVX2() && "We can only lower v32i8 with AVX2!");
10607 // Whenever we can lower this as a zext, that instruction is strictly faster
10608 // than any alternative. It also allows us to fold memory operands into the
10609 // shuffle in many cases.
10610 if (SDValue ZExt = lowerVectorShuffleAsZeroOrAnyExtend(DL, MVT::v32i8, V1, V2,
10611 Mask, Subtarget, DAG))
10614 // Check for being able to broadcast a single element.
10615 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v32i8, V1,
10616 Mask, Subtarget, DAG))
10619 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v32i8, V1, V2, Mask,
10623 // Use dedicated unpack instructions for masks that match their pattern.
10625 lowerVectorShuffleWithUNPCK(DL, MVT::v32i8, Mask, V1, V2, DAG))
10628 // Try to use shift instructions.
10629 if (SDValue Shift =
10630 lowerVectorShuffleAsShift(DL, MVT::v32i8, V1, V2, Mask, DAG))
10633 // Try to use byte rotation instructions.
10634 if (SDValue Rotate = lowerVectorShuffleAsByteRotate(
10635 DL, MVT::v32i8, V1, V2, Mask, Subtarget, DAG))
10638 if (isSingleInputShuffleMask(Mask)) {
10639 // There are no generalized cross-lane shuffle operations available on i8
10641 if (is128BitLaneCrossingShuffleMask(MVT::v32i8, Mask))
10642 return lowerVectorShuffleAsLanePermuteAndBlend(DL, MVT::v32i8, V1, V2,
10645 SDValue PSHUFBMask[32];
10646 for (int i = 0; i < 32; ++i)
10649 ? DAG.getUNDEF(MVT::i8)
10650 : DAG.getConstant(Mask[i] < 16 ? Mask[i] : Mask[i] - 16, DL,
10653 return DAG.getNode(
10654 X86ISD::PSHUFB, DL, MVT::v32i8, V1,
10655 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v32i8, PSHUFBMask));
10658 // Try to simplify this by merging 128-bit lanes to enable a lane-based
10660 if (SDValue Result = lowerVectorShuffleByMerging128BitLanes(
10661 DL, MVT::v32i8, V1, V2, Mask, Subtarget, DAG))
10664 // Otherwise fall back on generic lowering.
10665 return lowerVectorShuffleAsSplitOrBlend(DL, MVT::v32i8, V1, V2, Mask, DAG);
10668 /// \brief High-level routine to lower various 256-bit x86 vector shuffles.
10670 /// This routine either breaks down the specific type of a 256-bit x86 vector
10671 /// shuffle or splits it into two 128-bit shuffles and fuses the results back
10672 /// together based on the available instructions.
10673 static SDValue lower256BitVectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10674 MVT VT, const X86Subtarget *Subtarget,
10675 SelectionDAG &DAG) {
10677 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10678 ArrayRef<int> Mask = SVOp->getMask();
10680 // If we have a single input to the zero element, insert that into V1 if we
10681 // can do so cheaply.
10682 int NumElts = VT.getVectorNumElements();
10683 int NumV2Elements = std::count_if(Mask.begin(), Mask.end(), [NumElts](int M) {
10684 return M >= NumElts;
10687 if (NumV2Elements == 1 && Mask[0] >= NumElts)
10688 if (SDValue Insertion = lowerVectorShuffleAsElementInsertion(
10689 DL, VT, V1, V2, Mask, Subtarget, DAG))
10692 // There is a really nice hard cut-over between AVX1 and AVX2 that means we
10693 // can check for those subtargets here and avoid much of the subtarget
10694 // querying in the per-vector-type lowering routines. With AVX1 we have
10695 // essentially *zero* ability to manipulate a 256-bit vector with integer
10696 // types. Since we'll use floating point types there eventually, just
10697 // immediately cast everything to a float and operate entirely in that domain.
10698 if (VT.isInteger() && !Subtarget->hasAVX2()) {
10699 int ElementBits = VT.getScalarSizeInBits();
10700 if (ElementBits < 32)
10701 // No floating point type available, decompose into 128-bit vectors.
10702 return splitAndLowerVectorShuffle(DL, VT, V1, V2, Mask, DAG);
10704 MVT FpVT = MVT::getVectorVT(MVT::getFloatingPointVT(ElementBits),
10705 VT.getVectorNumElements());
10706 V1 = DAG.getBitcast(FpVT, V1);
10707 V2 = DAG.getBitcast(FpVT, V2);
10708 return DAG.getBitcast(VT, DAG.getVectorShuffle(FpVT, DL, V1, V2, Mask));
10711 switch (VT.SimpleTy) {
10713 return lowerV4F64VectorShuffle(Op, V1, V2, Subtarget, DAG);
10715 return lowerV4I64VectorShuffle(Op, V1, V2, Subtarget, DAG);
10717 return lowerV8F32VectorShuffle(Op, V1, V2, Subtarget, DAG);
10719 return lowerV8I32VectorShuffle(Op, V1, V2, Subtarget, DAG);
10721 return lowerV16I16VectorShuffle(Op, V1, V2, Subtarget, DAG);
10723 return lowerV32I8VectorShuffle(Op, V1, V2, Subtarget, DAG);
10726 llvm_unreachable("Not a valid 256-bit x86 vector type!");
10730 /// \brief Try to lower a vector shuffle as a 128-bit shuffles.
10731 static SDValue lowerV4X128VectorShuffle(SDLoc DL, MVT VT,
10732 ArrayRef<int> Mask,
10733 SDValue V1, SDValue V2,
10734 SelectionDAG &DAG) {
10735 assert(VT.getScalarSizeInBits() == 64 &&
10736 "Unexpected element type size for 128bit shuffle.");
10738 // To handle 256 bit vector requires VLX and most probably
10739 // function lowerV2X128VectorShuffle() is better solution.
10740 assert(VT.getSizeInBits() == 512 &&
10741 "Unexpected vector size for 128bit shuffle.");
10743 SmallVector<int, 4> WidenedMask;
10744 if (!canWidenShuffleElements(Mask, WidenedMask))
10747 // Form a 128-bit permutation.
10748 // Convert the 64-bit shuffle mask selection values into 128-bit selection
10749 // bits defined by a vshuf64x2 instruction's immediate control byte.
10750 unsigned PermMask = 0, Imm = 0;
10751 unsigned ControlBitsNum = WidenedMask.size() / 2;
10753 for (int i = 0, Size = WidenedMask.size(); i < Size; ++i) {
10754 if (WidenedMask[i] == SM_SentinelZero)
10757 // Use first element in place of undef mask.
10758 Imm = (WidenedMask[i] == SM_SentinelUndef) ? 0 : WidenedMask[i];
10759 PermMask |= (Imm % WidenedMask.size()) << (i * ControlBitsNum);
10762 return DAG.getNode(X86ISD::SHUF128, DL, VT, V1, V2,
10763 DAG.getConstant(PermMask, DL, MVT::i8));
10766 static SDValue lowerVectorShuffleWithPERMV(SDLoc DL, MVT VT,
10767 ArrayRef<int> Mask, SDValue V1,
10768 SDValue V2, SelectionDAG &DAG) {
10770 assert(VT.getScalarSizeInBits() >= 16 && "Unexpected data type for PERMV");
10772 MVT MaskEltVT = MVT::getIntegerVT(VT.getScalarSizeInBits());
10773 MVT MaskVecVT = MVT::getVectorVT(MaskEltVT, VT.getVectorNumElements());
10775 SDValue MaskNode = getConstVector(Mask, MaskVecVT, DAG, DL, true);
10776 if (isSingleInputShuffleMask(Mask))
10777 return DAG.getNode(X86ISD::VPERMV, DL, VT, MaskNode, V1);
10779 return DAG.getNode(X86ISD::VPERMV3, DL, VT, V1, MaskNode, V2);
10782 /// \brief Handle lowering of 8-lane 64-bit floating point shuffles.
10783 static SDValue lowerV8F64VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10784 const X86Subtarget *Subtarget,
10785 SelectionDAG &DAG) {
10787 assert(V1.getSimpleValueType() == MVT::v8f64 && "Bad operand type!");
10788 assert(V2.getSimpleValueType() == MVT::v8f64 && "Bad operand type!");
10789 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10790 ArrayRef<int> Mask = SVOp->getMask();
10791 assert(Mask.size() == 8 && "Unexpected mask size for v8 shuffle!");
10793 if (SDValue Shuf128 =
10794 lowerV4X128VectorShuffle(DL, MVT::v8f64, Mask, V1, V2, DAG))
10797 if (SDValue Unpck =
10798 lowerVectorShuffleWithUNPCK(DL, MVT::v8f64, Mask, V1, V2, DAG))
10801 return lowerVectorShuffleWithPERMV(DL, MVT::v8f64, Mask, V1, V2, DAG);
10804 /// \brief Handle lowering of 16-lane 32-bit floating point shuffles.
10805 static SDValue lowerV16F32VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10806 const X86Subtarget *Subtarget,
10807 SelectionDAG &DAG) {
10809 assert(V1.getSimpleValueType() == MVT::v16f32 && "Bad operand type!");
10810 assert(V2.getSimpleValueType() == MVT::v16f32 && "Bad operand type!");
10811 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10812 ArrayRef<int> Mask = SVOp->getMask();
10813 assert(Mask.size() == 16 && "Unexpected mask size for v16 shuffle!");
10815 if (SDValue Unpck =
10816 lowerVectorShuffleWithUNPCK(DL, MVT::v16f32, Mask, V1, V2, DAG))
10819 return lowerVectorShuffleWithPERMV(DL, MVT::v16f32, Mask, V1, V2, DAG);
10822 /// \brief Handle lowering of 8-lane 64-bit integer shuffles.
10823 static SDValue lowerV8I64VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10824 const X86Subtarget *Subtarget,
10825 SelectionDAG &DAG) {
10827 assert(V1.getSimpleValueType() == MVT::v8i64 && "Bad operand type!");
10828 assert(V2.getSimpleValueType() == MVT::v8i64 && "Bad operand type!");
10829 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10830 ArrayRef<int> Mask = SVOp->getMask();
10831 assert(Mask.size() == 8 && "Unexpected mask size for v8 shuffle!");
10833 if (SDValue Shuf128 =
10834 lowerV4X128VectorShuffle(DL, MVT::v8i64, Mask, V1, V2, DAG))
10837 if (SDValue Unpck =
10838 lowerVectorShuffleWithUNPCK(DL, MVT::v8i64, Mask, V1, V2, DAG))
10841 return lowerVectorShuffleWithPERMV(DL, MVT::v8i64, Mask, V1, V2, DAG);
10844 /// \brief Handle lowering of 16-lane 32-bit integer shuffles.
10845 static SDValue lowerV16I32VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10846 const X86Subtarget *Subtarget,
10847 SelectionDAG &DAG) {
10849 assert(V1.getSimpleValueType() == MVT::v16i32 && "Bad operand type!");
10850 assert(V2.getSimpleValueType() == MVT::v16i32 && "Bad operand type!");
10851 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10852 ArrayRef<int> Mask = SVOp->getMask();
10853 assert(Mask.size() == 16 && "Unexpected mask size for v16 shuffle!");
10855 if (SDValue Unpck =
10856 lowerVectorShuffleWithUNPCK(DL, MVT::v16i32, Mask, V1, V2, DAG))
10859 return lowerVectorShuffleWithPERMV(DL, MVT::v16i32, Mask, V1, V2, DAG);
10862 /// \brief Handle lowering of 32-lane 16-bit integer shuffles.
10863 static SDValue lowerV32I16VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10864 const X86Subtarget *Subtarget,
10865 SelectionDAG &DAG) {
10867 assert(V1.getSimpleValueType() == MVT::v32i16 && "Bad operand type!");
10868 assert(V2.getSimpleValueType() == MVT::v32i16 && "Bad operand type!");
10869 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10870 ArrayRef<int> Mask = SVOp->getMask();
10871 assert(Mask.size() == 32 && "Unexpected mask size for v32 shuffle!");
10872 assert(Subtarget->hasBWI() && "We can only lower v32i16 with AVX-512-BWI!");
10874 return lowerVectorShuffleWithPERMV(DL, MVT::v32i16, Mask, V1, V2, DAG);
10877 /// \brief Handle lowering of 64-lane 8-bit integer shuffles.
10878 static SDValue lowerV64I8VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10879 const X86Subtarget *Subtarget,
10880 SelectionDAG &DAG) {
10882 assert(V1.getSimpleValueType() == MVT::v64i8 && "Bad operand type!");
10883 assert(V2.getSimpleValueType() == MVT::v64i8 && "Bad operand type!");
10884 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10885 ArrayRef<int> Mask = SVOp->getMask();
10886 assert(Mask.size() == 64 && "Unexpected mask size for v64 shuffle!");
10887 assert(Subtarget->hasBWI() && "We can only lower v64i8 with AVX-512-BWI!");
10889 // FIXME: Implement direct support for this type!
10890 return splitAndLowerVectorShuffle(DL, MVT::v64i8, V1, V2, Mask, DAG);
10893 /// \brief High-level routine to lower various 512-bit x86 vector shuffles.
10895 /// This routine either breaks down the specific type of a 512-bit x86 vector
10896 /// shuffle or splits it into two 256-bit shuffles and fuses the results back
10897 /// together based on the available instructions.
10898 static SDValue lower512BitVectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10899 MVT VT, const X86Subtarget *Subtarget,
10900 SelectionDAG &DAG) {
10902 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10903 ArrayRef<int> Mask = SVOp->getMask();
10904 assert(Subtarget->hasAVX512() &&
10905 "Cannot lower 512-bit vectors w/ basic ISA!");
10907 // Check for being able to broadcast a single element.
10908 if (SDValue Broadcast =
10909 lowerVectorShuffleAsBroadcast(DL, VT, V1, Mask, Subtarget, DAG))
10912 // Dispatch to each element type for lowering. If we don't have supprot for
10913 // specific element type shuffles at 512 bits, immediately split them and
10914 // lower them. Each lowering routine of a given type is allowed to assume that
10915 // the requisite ISA extensions for that element type are available.
10916 switch (VT.SimpleTy) {
10918 return lowerV8F64VectorShuffle(Op, V1, V2, Subtarget, DAG);
10920 return lowerV16F32VectorShuffle(Op, V1, V2, Subtarget, DAG);
10922 return lowerV8I64VectorShuffle(Op, V1, V2, Subtarget, DAG);
10924 return lowerV16I32VectorShuffle(Op, V1, V2, Subtarget, DAG);
10926 if (Subtarget->hasBWI())
10927 return lowerV32I16VectorShuffle(Op, V1, V2, Subtarget, DAG);
10930 if (Subtarget->hasBWI())
10931 return lowerV64I8VectorShuffle(Op, V1, V2, Subtarget, DAG);
10935 llvm_unreachable("Not a valid 512-bit x86 vector type!");
10938 // Otherwise fall back on splitting.
10939 return splitAndLowerVectorShuffle(DL, VT, V1, V2, Mask, DAG);
10942 // Lower vXi1 vector shuffles.
10943 // There is no a dedicated instruction on AVX-512 that shuffles the masks.
10944 // The only way to shuffle bits is to sign-extend the mask vector to SIMD
10945 // vector, shuffle and then truncate it back.
10946 static SDValue lower1BitVectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10947 MVT VT, const X86Subtarget *Subtarget,
10948 SelectionDAG &DAG) {
10950 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10951 ArrayRef<int> Mask = SVOp->getMask();
10952 assert(Subtarget->hasAVX512() &&
10953 "Cannot lower 512-bit vectors w/o basic ISA!");
10955 switch (VT.SimpleTy) {
10957 llvm_unreachable("Expected a vector of i1 elements");
10959 ExtVT = MVT::v2i64;
10962 ExtVT = MVT::v4i32;
10965 ExtVT = MVT::v8i64; // Take 512-bit type, more shuffles on KNL
10968 ExtVT = MVT::v16i32;
10971 ExtVT = MVT::v32i16;
10974 ExtVT = MVT::v64i8;
10978 if (ISD::isBuildVectorAllZeros(V1.getNode()))
10979 V1 = getZeroVector(ExtVT, Subtarget, DAG, DL);
10980 else if (ISD::isBuildVectorAllOnes(V1.getNode()))
10981 V1 = getOnesVector(ExtVT, Subtarget, DAG, DL);
10983 V1 = DAG.getNode(ISD::SIGN_EXTEND, DL, ExtVT, V1);
10986 V2 = DAG.getUNDEF(ExtVT);
10987 else if (ISD::isBuildVectorAllZeros(V2.getNode()))
10988 V2 = getZeroVector(ExtVT, Subtarget, DAG, DL);
10989 else if (ISD::isBuildVectorAllOnes(V2.getNode()))
10990 V2 = getOnesVector(ExtVT, Subtarget, DAG, DL);
10992 V2 = DAG.getNode(ISD::SIGN_EXTEND, DL, ExtVT, V2);
10993 return DAG.getNode(ISD::TRUNCATE, DL, VT,
10994 DAG.getVectorShuffle(ExtVT, DL, V1, V2, Mask));
10996 /// \brief Top-level lowering for x86 vector shuffles.
10998 /// This handles decomposition, canonicalization, and lowering of all x86
10999 /// vector shuffles. Most of the specific lowering strategies are encapsulated
11000 /// above in helper routines. The canonicalization attempts to widen shuffles
11001 /// to involve fewer lanes of wider elements, consolidate symmetric patterns
11002 /// s.t. only one of the two inputs needs to be tested, etc.
11003 static SDValue lowerVectorShuffle(SDValue Op, const X86Subtarget *Subtarget,
11004 SelectionDAG &DAG) {
11005 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
11006 ArrayRef<int> Mask = SVOp->getMask();
11007 SDValue V1 = Op.getOperand(0);
11008 SDValue V2 = Op.getOperand(1);
11009 MVT VT = Op.getSimpleValueType();
11010 int NumElements = VT.getVectorNumElements();
11012 bool Is1BitVector = (VT.getScalarType() == MVT::i1);
11014 assert((VT.getSizeInBits() != 64 || Is1BitVector) &&
11015 "Can't lower MMX shuffles");
11017 bool V1IsUndef = V1.getOpcode() == ISD::UNDEF;
11018 bool V2IsUndef = V2.getOpcode() == ISD::UNDEF;
11019 if (V1IsUndef && V2IsUndef)
11020 return DAG.getUNDEF(VT);
11022 // When we create a shuffle node we put the UNDEF node to second operand,
11023 // but in some cases the first operand may be transformed to UNDEF.
11024 // In this case we should just commute the node.
11026 return DAG.getCommutedVectorShuffle(*SVOp);
11028 // Check for non-undef masks pointing at an undef vector and make the masks
11029 // undef as well. This makes it easier to match the shuffle based solely on
11033 if (M >= NumElements) {
11034 SmallVector<int, 8> NewMask(Mask.begin(), Mask.end());
11035 for (int &M : NewMask)
11036 if (M >= NumElements)
11038 return DAG.getVectorShuffle(VT, dl, V1, V2, NewMask);
11041 // We actually see shuffles that are entirely re-arrangements of a set of
11042 // zero inputs. This mostly happens while decomposing complex shuffles into
11043 // simple ones. Directly lower these as a buildvector of zeros.
11044 SmallBitVector Zeroable = computeZeroableShuffleElements(Mask, V1, V2);
11045 if (Zeroable.all())
11046 return getZeroVector(VT, Subtarget, DAG, dl);
11048 // Try to collapse shuffles into using a vector type with fewer elements but
11049 // wider element types. We cap this to not form integers or floating point
11050 // elements wider than 64 bits, but it might be interesting to form i128
11051 // integers to handle flipping the low and high halves of AVX 256-bit vectors.
11052 SmallVector<int, 16> WidenedMask;
11053 if (VT.getScalarSizeInBits() < 64 && !Is1BitVector &&
11054 canWidenShuffleElements(Mask, WidenedMask)) {
11055 MVT NewEltVT = VT.isFloatingPoint()
11056 ? MVT::getFloatingPointVT(VT.getScalarSizeInBits() * 2)
11057 : MVT::getIntegerVT(VT.getScalarSizeInBits() * 2);
11058 MVT NewVT = MVT::getVectorVT(NewEltVT, VT.getVectorNumElements() / 2);
11059 // Make sure that the new vector type is legal. For example, v2f64 isn't
11061 if (DAG.getTargetLoweringInfo().isTypeLegal(NewVT)) {
11062 V1 = DAG.getBitcast(NewVT, V1);
11063 V2 = DAG.getBitcast(NewVT, V2);
11064 return DAG.getBitcast(
11065 VT, DAG.getVectorShuffle(NewVT, dl, V1, V2, WidenedMask));
11069 int NumV1Elements = 0, NumUndefElements = 0, NumV2Elements = 0;
11070 for (int M : SVOp->getMask())
11072 ++NumUndefElements;
11073 else if (M < NumElements)
11078 // Commute the shuffle as needed such that more elements come from V1 than
11079 // V2. This allows us to match the shuffle pattern strictly on how many
11080 // elements come from V1 without handling the symmetric cases.
11081 if (NumV2Elements > NumV1Elements)
11082 return DAG.getCommutedVectorShuffle(*SVOp);
11084 // When the number of V1 and V2 elements are the same, try to minimize the
11085 // number of uses of V2 in the low half of the vector. When that is tied,
11086 // ensure that the sum of indices for V1 is equal to or lower than the sum
11087 // indices for V2. When those are equal, try to ensure that the number of odd
11088 // indices for V1 is lower than the number of odd indices for V2.
11089 if (NumV1Elements == NumV2Elements) {
11090 int LowV1Elements = 0, LowV2Elements = 0;
11091 for (int M : SVOp->getMask().slice(0, NumElements / 2))
11092 if (M >= NumElements)
11096 if (LowV2Elements > LowV1Elements) {
11097 return DAG.getCommutedVectorShuffle(*SVOp);
11098 } else if (LowV2Elements == LowV1Elements) {
11099 int SumV1Indices = 0, SumV2Indices = 0;
11100 for (int i = 0, Size = SVOp->getMask().size(); i < Size; ++i)
11101 if (SVOp->getMask()[i] >= NumElements)
11103 else if (SVOp->getMask()[i] >= 0)
11105 if (SumV2Indices < SumV1Indices) {
11106 return DAG.getCommutedVectorShuffle(*SVOp);
11107 } else if (SumV2Indices == SumV1Indices) {
11108 int NumV1OddIndices = 0, NumV2OddIndices = 0;
11109 for (int i = 0, Size = SVOp->getMask().size(); i < Size; ++i)
11110 if (SVOp->getMask()[i] >= NumElements)
11111 NumV2OddIndices += i % 2;
11112 else if (SVOp->getMask()[i] >= 0)
11113 NumV1OddIndices += i % 2;
11114 if (NumV2OddIndices < NumV1OddIndices)
11115 return DAG.getCommutedVectorShuffle(*SVOp);
11120 // For each vector width, delegate to a specialized lowering routine.
11121 if (VT.getSizeInBits() == 128)
11122 return lower128BitVectorShuffle(Op, V1, V2, VT, Subtarget, DAG);
11124 if (VT.getSizeInBits() == 256)
11125 return lower256BitVectorShuffle(Op, V1, V2, VT, Subtarget, DAG);
11127 if (VT.getSizeInBits() == 512)
11128 return lower512BitVectorShuffle(Op, V1, V2, VT, Subtarget, DAG);
11131 return lower1BitVectorShuffle(Op, V1, V2, VT, Subtarget, DAG);
11132 llvm_unreachable("Unimplemented!");
11135 // This function assumes its argument is a BUILD_VECTOR of constants or
11136 // undef SDNodes. i.e: ISD::isBuildVectorOfConstantSDNodes(BuildVector) is
11138 static bool BUILD_VECTORtoBlendMask(BuildVectorSDNode *BuildVector,
11139 unsigned &MaskValue) {
11141 unsigned NumElems = BuildVector->getNumOperands();
11143 // There are 2 lanes if (NumElems > 8), and 1 lane otherwise.
11144 // We don't handle the >2 lanes case right now.
11145 unsigned NumLanes = (NumElems - 1) / 8 + 1;
11149 unsigned NumElemsInLane = NumElems / NumLanes;
11151 // Blend for v16i16 should be symmetric for the both lanes.
11152 for (unsigned i = 0; i < NumElemsInLane; ++i) {
11153 SDValue EltCond = BuildVector->getOperand(i);
11154 SDValue SndLaneEltCond =
11155 (NumLanes == 2) ? BuildVector->getOperand(i + NumElemsInLane) : EltCond;
11157 int Lane1Cond = -1, Lane2Cond = -1;
11158 if (isa<ConstantSDNode>(EltCond))
11159 Lane1Cond = !isZero(EltCond);
11160 if (isa<ConstantSDNode>(SndLaneEltCond))
11161 Lane2Cond = !isZero(SndLaneEltCond);
11163 unsigned LaneMask = 0;
11164 if (Lane1Cond == Lane2Cond || Lane2Cond < 0)
11165 // Lane1Cond != 0, means we want the first argument.
11166 // Lane1Cond == 0, means we want the second argument.
11167 // The encoding of this argument is 0 for the first argument, 1
11168 // for the second. Therefore, invert the condition.
11169 LaneMask = !Lane1Cond << i;
11170 else if (Lane1Cond < 0)
11171 LaneMask = !Lane2Cond << i;
11175 MaskValue |= LaneMask;
11177 MaskValue |= LaneMask << NumElemsInLane;
11182 /// \brief Try to lower a VSELECT instruction to a vector shuffle.
11183 static SDValue lowerVSELECTtoVectorShuffle(SDValue Op,
11184 const X86Subtarget *Subtarget,
11185 SelectionDAG &DAG) {
11186 SDValue Cond = Op.getOperand(0);
11187 SDValue LHS = Op.getOperand(1);
11188 SDValue RHS = Op.getOperand(2);
11190 MVT VT = Op.getSimpleValueType();
11192 if (!ISD::isBuildVectorOfConstantSDNodes(Cond.getNode()))
11194 auto *CondBV = cast<BuildVectorSDNode>(Cond);
11196 // Only non-legal VSELECTs reach this lowering, convert those into generic
11197 // shuffles and re-use the shuffle lowering path for blends.
11198 SmallVector<int, 32> Mask;
11199 for (int i = 0, Size = VT.getVectorNumElements(); i < Size; ++i) {
11200 SDValue CondElt = CondBV->getOperand(i);
11202 isa<ConstantSDNode>(CondElt) ? i + (isZero(CondElt) ? Size : 0) : -1);
11204 return DAG.getVectorShuffle(VT, dl, LHS, RHS, Mask);
11207 SDValue X86TargetLowering::LowerVSELECT(SDValue Op, SelectionDAG &DAG) const {
11208 // A vselect where all conditions and data are constants can be optimized into
11209 // a single vector load by SelectionDAGLegalize::ExpandBUILD_VECTOR().
11210 if (ISD::isBuildVectorOfConstantSDNodes(Op.getOperand(0).getNode()) &&
11211 ISD::isBuildVectorOfConstantSDNodes(Op.getOperand(1).getNode()) &&
11212 ISD::isBuildVectorOfConstantSDNodes(Op.getOperand(2).getNode()))
11215 // Try to lower this to a blend-style vector shuffle. This can handle all
11216 // constant condition cases.
11217 if (SDValue BlendOp = lowerVSELECTtoVectorShuffle(Op, Subtarget, DAG))
11220 // Variable blends are only legal from SSE4.1 onward.
11221 if (!Subtarget->hasSSE41())
11224 // Only some types will be legal on some subtargets. If we can emit a legal
11225 // VSELECT-matching blend, return Op, and but if we need to expand, return
11227 switch (Op.getSimpleValueType().SimpleTy) {
11229 // Most of the vector types have blends past SSE4.1.
11233 // The byte blends for AVX vectors were introduced only in AVX2.
11234 if (Subtarget->hasAVX2())
11241 // AVX-512 BWI and VLX features support VSELECT with i16 elements.
11242 if (Subtarget->hasBWI() && Subtarget->hasVLX())
11245 // FIXME: We should custom lower this by fixing the condition and using i8
11251 static SDValue LowerEXTRACT_VECTOR_ELT_SSE4(SDValue Op, SelectionDAG &DAG) {
11252 MVT VT = Op.getSimpleValueType();
11255 if (!Op.getOperand(0).getSimpleValueType().is128BitVector())
11258 if (VT.getSizeInBits() == 8) {
11259 SDValue Extract = DAG.getNode(X86ISD::PEXTRB, dl, MVT::i32,
11260 Op.getOperand(0), Op.getOperand(1));
11261 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Extract,
11262 DAG.getValueType(VT));
11263 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
11266 if (VT.getSizeInBits() == 16) {
11267 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
11268 // If Idx is 0, it's cheaper to do a move instead of a pextrw.
11270 return DAG.getNode(
11271 ISD::TRUNCATE, dl, MVT::i16,
11272 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
11273 DAG.getBitcast(MVT::v4i32, Op.getOperand(0)),
11274 Op.getOperand(1)));
11275 SDValue Extract = DAG.getNode(X86ISD::PEXTRW, dl, MVT::i32,
11276 Op.getOperand(0), Op.getOperand(1));
11277 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Extract,
11278 DAG.getValueType(VT));
11279 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
11282 if (VT == MVT::f32) {
11283 // EXTRACTPS outputs to a GPR32 register which will require a movd to copy
11284 // the result back to FR32 register. It's only worth matching if the
11285 // result has a single use which is a store or a bitcast to i32. And in
11286 // the case of a store, it's not worth it if the index is a constant 0,
11287 // because a MOVSSmr can be used instead, which is smaller and faster.
11288 if (!Op.hasOneUse())
11290 SDNode *User = *Op.getNode()->use_begin();
11291 if ((User->getOpcode() != ISD::STORE ||
11292 (isa<ConstantSDNode>(Op.getOperand(1)) &&
11293 cast<ConstantSDNode>(Op.getOperand(1))->isNullValue())) &&
11294 (User->getOpcode() != ISD::BITCAST ||
11295 User->getValueType(0) != MVT::i32))
11297 SDValue Extract = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
11298 DAG.getBitcast(MVT::v4i32, Op.getOperand(0)),
11300 return DAG.getBitcast(MVT::f32, Extract);
11303 if (VT == MVT::i32 || VT == MVT::i64) {
11304 // ExtractPS/pextrq works with constant index.
11305 if (isa<ConstantSDNode>(Op.getOperand(1)))
11311 /// Extract one bit from mask vector, like v16i1 or v8i1.
11312 /// AVX-512 feature.
11314 X86TargetLowering::ExtractBitFromMaskVector(SDValue Op, SelectionDAG &DAG) const {
11315 SDValue Vec = Op.getOperand(0);
11317 MVT VecVT = Vec.getSimpleValueType();
11318 SDValue Idx = Op.getOperand(1);
11319 MVT EltVT = Op.getSimpleValueType();
11321 assert((EltVT == MVT::i1) && "Unexpected operands in ExtractBitFromMaskVector");
11322 assert((VecVT.getVectorNumElements() <= 16 || Subtarget->hasBWI()) &&
11323 "Unexpected vector type in ExtractBitFromMaskVector");
11325 // variable index can't be handled in mask registers,
11326 // extend vector to VR512
11327 if (!isa<ConstantSDNode>(Idx)) {
11328 MVT ExtVT = (VecVT == MVT::v8i1 ? MVT::v8i64 : MVT::v16i32);
11329 SDValue Ext = DAG.getNode(ISD::ZERO_EXTEND, dl, ExtVT, Vec);
11330 SDValue Elt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl,
11331 ExtVT.getVectorElementType(), Ext, Idx);
11332 return DAG.getNode(ISD::TRUNCATE, dl, EltVT, Elt);
11335 unsigned IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
11336 const TargetRegisterClass* rc = getRegClassFor(VecVT);
11337 if (!Subtarget->hasDQI() && (VecVT.getVectorNumElements() <= 8))
11338 rc = getRegClassFor(MVT::v16i1);
11339 unsigned MaxSift = rc->getSize()*8 - 1;
11340 Vec = DAG.getNode(X86ISD::VSHLI, dl, VecVT, Vec,
11341 DAG.getConstant(MaxSift - IdxVal, dl, MVT::i8));
11342 Vec = DAG.getNode(X86ISD::VSRLI, dl, VecVT, Vec,
11343 DAG.getConstant(MaxSift, dl, MVT::i8));
11344 return DAG.getNode(X86ISD::VEXTRACT, dl, MVT::i1, Vec,
11345 DAG.getIntPtrConstant(0, dl));
11349 X86TargetLowering::LowerEXTRACT_VECTOR_ELT(SDValue Op,
11350 SelectionDAG &DAG) const {
11352 SDValue Vec = Op.getOperand(0);
11353 MVT VecVT = Vec.getSimpleValueType();
11354 SDValue Idx = Op.getOperand(1);
11356 if (Op.getSimpleValueType() == MVT::i1)
11357 return ExtractBitFromMaskVector(Op, DAG);
11359 if (!isa<ConstantSDNode>(Idx)) {
11360 if (VecVT.is512BitVector() ||
11361 (VecVT.is256BitVector() && Subtarget->hasInt256() &&
11362 VecVT.getVectorElementType().getSizeInBits() == 32)) {
11365 MVT::getIntegerVT(VecVT.getVectorElementType().getSizeInBits());
11366 MVT MaskVT = MVT::getVectorVT(MaskEltVT, VecVT.getSizeInBits() /
11367 MaskEltVT.getSizeInBits());
11369 Idx = DAG.getZExtOrTrunc(Idx, dl, MaskEltVT);
11370 auto PtrVT = getPointerTy(DAG.getDataLayout());
11371 SDValue Mask = DAG.getNode(X86ISD::VINSERT, dl, MaskVT,
11372 getZeroVector(MaskVT, Subtarget, DAG, dl), Idx,
11373 DAG.getConstant(0, dl, PtrVT));
11374 SDValue Perm = DAG.getNode(X86ISD::VPERMV, dl, VecVT, Mask, Vec);
11375 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, Op.getValueType(), Perm,
11376 DAG.getConstant(0, dl, PtrVT));
11381 // If this is a 256-bit vector result, first extract the 128-bit vector and
11382 // then extract the element from the 128-bit vector.
11383 if (VecVT.is256BitVector() || VecVT.is512BitVector()) {
11385 unsigned IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
11386 // Get the 128-bit vector.
11387 Vec = Extract128BitVector(Vec, IdxVal, DAG, dl);
11388 MVT EltVT = VecVT.getVectorElementType();
11390 unsigned ElemsPerChunk = 128 / EltVT.getSizeInBits();
11392 //if (IdxVal >= NumElems/2)
11393 // IdxVal -= NumElems/2;
11394 IdxVal -= (IdxVal/ElemsPerChunk)*ElemsPerChunk;
11395 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, Op.getValueType(), Vec,
11396 DAG.getConstant(IdxVal, dl, MVT::i32));
11399 assert(VecVT.is128BitVector() && "Unexpected vector length");
11401 if (Subtarget->hasSSE41())
11402 if (SDValue Res = LowerEXTRACT_VECTOR_ELT_SSE4(Op, DAG))
11405 MVT VT = Op.getSimpleValueType();
11406 // TODO: handle v16i8.
11407 if (VT.getSizeInBits() == 16) {
11408 SDValue Vec = Op.getOperand(0);
11409 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
11411 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i16,
11412 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
11413 DAG.getBitcast(MVT::v4i32, Vec),
11414 Op.getOperand(1)));
11415 // Transform it so it match pextrw which produces a 32-bit result.
11416 MVT EltVT = MVT::i32;
11417 SDValue Extract = DAG.getNode(X86ISD::PEXTRW, dl, EltVT,
11418 Op.getOperand(0), Op.getOperand(1));
11419 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, EltVT, Extract,
11420 DAG.getValueType(VT));
11421 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
11424 if (VT.getSizeInBits() == 32) {
11425 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
11429 // SHUFPS the element to the lowest double word, then movss.
11430 int Mask[4] = { static_cast<int>(Idx), -1, -1, -1 };
11431 MVT VVT = Op.getOperand(0).getSimpleValueType();
11432 SDValue Vec = DAG.getVectorShuffle(VVT, dl, Op.getOperand(0),
11433 DAG.getUNDEF(VVT), Mask);
11434 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, Vec,
11435 DAG.getIntPtrConstant(0, dl));
11438 if (VT.getSizeInBits() == 64) {
11439 // FIXME: .td only matches this for <2 x f64>, not <2 x i64> on 32b
11440 // FIXME: seems like this should be unnecessary if mov{h,l}pd were taught
11441 // to match extract_elt for f64.
11442 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
11446 // UNPCKHPD the element to the lowest double word, then movsd.
11447 // Note if the lower 64 bits of the result of the UNPCKHPD is then stored
11448 // to a f64mem, the whole operation is folded into a single MOVHPDmr.
11449 int Mask[2] = { 1, -1 };
11450 MVT VVT = Op.getOperand(0).getSimpleValueType();
11451 SDValue Vec = DAG.getVectorShuffle(VVT, dl, Op.getOperand(0),
11452 DAG.getUNDEF(VVT), Mask);
11453 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, Vec,
11454 DAG.getIntPtrConstant(0, dl));
11460 /// Insert one bit to mask vector, like v16i1 or v8i1.
11461 /// AVX-512 feature.
11463 X86TargetLowering::InsertBitToMaskVector(SDValue Op, SelectionDAG &DAG) const {
11465 SDValue Vec = Op.getOperand(0);
11466 SDValue Elt = Op.getOperand(1);
11467 SDValue Idx = Op.getOperand(2);
11468 MVT VecVT = Vec.getSimpleValueType();
11470 if (!isa<ConstantSDNode>(Idx)) {
11471 // Non constant index. Extend source and destination,
11472 // insert element and then truncate the result.
11473 MVT ExtVecVT = (VecVT == MVT::v8i1 ? MVT::v8i64 : MVT::v16i32);
11474 MVT ExtEltVT = (VecVT == MVT::v8i1 ? MVT::i64 : MVT::i32);
11475 SDValue ExtOp = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, ExtVecVT,
11476 DAG.getNode(ISD::ZERO_EXTEND, dl, ExtVecVT, Vec),
11477 DAG.getNode(ISD::ZERO_EXTEND, dl, ExtEltVT, Elt), Idx);
11478 return DAG.getNode(ISD::TRUNCATE, dl, VecVT, ExtOp);
11481 unsigned IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
11482 SDValue EltInVec = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VecVT, Elt);
11484 EltInVec = DAG.getNode(X86ISD::VSHLI, dl, VecVT, EltInVec,
11485 DAG.getConstant(IdxVal, dl, MVT::i8));
11486 if (Vec.getOpcode() == ISD::UNDEF)
11488 return DAG.getNode(ISD::OR, dl, VecVT, Vec, EltInVec);
11491 SDValue X86TargetLowering::LowerINSERT_VECTOR_ELT(SDValue Op,
11492 SelectionDAG &DAG) const {
11493 MVT VT = Op.getSimpleValueType();
11494 MVT EltVT = VT.getVectorElementType();
11496 if (EltVT == MVT::i1)
11497 return InsertBitToMaskVector(Op, DAG);
11500 SDValue N0 = Op.getOperand(0);
11501 SDValue N1 = Op.getOperand(1);
11502 SDValue N2 = Op.getOperand(2);
11503 if (!isa<ConstantSDNode>(N2))
11505 auto *N2C = cast<ConstantSDNode>(N2);
11506 unsigned IdxVal = N2C->getZExtValue();
11508 // If the vector is wider than 128 bits, extract the 128-bit subvector, insert
11509 // into that, and then insert the subvector back into the result.
11510 if (VT.is256BitVector() || VT.is512BitVector()) {
11511 // With a 256-bit vector, we can insert into the zero element efficiently
11512 // using a blend if we have AVX or AVX2 and the right data type.
11513 if (VT.is256BitVector() && IdxVal == 0) {
11514 // TODO: It is worthwhile to cast integer to floating point and back
11515 // and incur a domain crossing penalty if that's what we'll end up
11516 // doing anyway after extracting to a 128-bit vector.
11517 if ((Subtarget->hasAVX() && (EltVT == MVT::f64 || EltVT == MVT::f32)) ||
11518 (Subtarget->hasAVX2() && EltVT == MVT::i32)) {
11519 SDValue N1Vec = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, N1);
11520 N2 = DAG.getIntPtrConstant(1, dl);
11521 return DAG.getNode(X86ISD::BLENDI, dl, VT, N0, N1Vec, N2);
11525 // Get the desired 128-bit vector chunk.
11526 SDValue V = Extract128BitVector(N0, IdxVal, DAG, dl);
11528 // Insert the element into the desired chunk.
11529 unsigned NumEltsIn128 = 128 / EltVT.getSizeInBits();
11530 unsigned IdxIn128 = IdxVal - (IdxVal / NumEltsIn128) * NumEltsIn128;
11532 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, V.getValueType(), V, N1,
11533 DAG.getConstant(IdxIn128, dl, MVT::i32));
11535 // Insert the changed part back into the bigger vector
11536 return Insert128BitVector(N0, V, IdxVal, DAG, dl);
11538 assert(VT.is128BitVector() && "Only 128-bit vector types should be left!");
11540 if (Subtarget->hasSSE41()) {
11541 if (EltVT.getSizeInBits() == 8 || EltVT.getSizeInBits() == 16) {
11543 if (VT == MVT::v8i16) {
11544 Opc = X86ISD::PINSRW;
11546 assert(VT == MVT::v16i8);
11547 Opc = X86ISD::PINSRB;
11550 // Transform it so it match pinsr{b,w} which expects a GR32 as its second
11552 if (N1.getValueType() != MVT::i32)
11553 N1 = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, N1);
11554 if (N2.getValueType() != MVT::i32)
11555 N2 = DAG.getIntPtrConstant(IdxVal, dl);
11556 return DAG.getNode(Opc, dl, VT, N0, N1, N2);
11559 if (EltVT == MVT::f32) {
11560 // Bits [7:6] of the constant are the source select. This will always be
11561 // zero here. The DAG Combiner may combine an extract_elt index into
11562 // these bits. For example (insert (extract, 3), 2) could be matched by
11563 // putting the '3' into bits [7:6] of X86ISD::INSERTPS.
11564 // Bits [5:4] of the constant are the destination select. This is the
11565 // value of the incoming immediate.
11566 // Bits [3:0] of the constant are the zero mask. The DAG Combiner may
11567 // combine either bitwise AND or insert of float 0.0 to set these bits.
11569 bool MinSize = DAG.getMachineFunction().getFunction()->optForMinSize();
11570 if (IdxVal == 0 && (!MinSize || !MayFoldLoad(N1))) {
11571 // If this is an insertion of 32-bits into the low 32-bits of
11572 // a vector, we prefer to generate a blend with immediate rather
11573 // than an insertps. Blends are simpler operations in hardware and so
11574 // will always have equal or better performance than insertps.
11575 // But if optimizing for size and there's a load folding opportunity,
11576 // generate insertps because blendps does not have a 32-bit memory
11578 N2 = DAG.getIntPtrConstant(1, dl);
11579 N1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4f32, N1);
11580 return DAG.getNode(X86ISD::BLENDI, dl, VT, N0, N1, N2);
11582 N2 = DAG.getIntPtrConstant(IdxVal << 4, dl);
11583 // Create this as a scalar to vector..
11584 N1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4f32, N1);
11585 return DAG.getNode(X86ISD::INSERTPS, dl, VT, N0, N1, N2);
11588 if (EltVT == MVT::i32 || EltVT == MVT::i64) {
11589 // PINSR* works with constant index.
11594 if (EltVT == MVT::i8)
11597 if (EltVT.getSizeInBits() == 16) {
11598 // Transform it so it match pinsrw which expects a 16-bit value in a GR32
11599 // as its second argument.
11600 if (N1.getValueType() != MVT::i32)
11601 N1 = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, N1);
11602 if (N2.getValueType() != MVT::i32)
11603 N2 = DAG.getIntPtrConstant(IdxVal, dl);
11604 return DAG.getNode(X86ISD::PINSRW, dl, VT, N0, N1, N2);
11609 static SDValue LowerSCALAR_TO_VECTOR(SDValue Op, SelectionDAG &DAG) {
11611 MVT OpVT = Op.getSimpleValueType();
11613 // If this is a 256-bit vector result, first insert into a 128-bit
11614 // vector and then insert into the 256-bit vector.
11615 if (!OpVT.is128BitVector()) {
11616 // Insert into a 128-bit vector.
11617 unsigned SizeFactor = OpVT.getSizeInBits()/128;
11618 MVT VT128 = MVT::getVectorVT(OpVT.getVectorElementType(),
11619 OpVT.getVectorNumElements() / SizeFactor);
11621 Op = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT128, Op.getOperand(0));
11623 // Insert the 128-bit vector.
11624 return Insert128BitVector(DAG.getUNDEF(OpVT), Op, 0, DAG, dl);
11627 if (OpVT == MVT::v1i64 &&
11628 Op.getOperand(0).getValueType() == MVT::i64)
11629 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v1i64, Op.getOperand(0));
11631 SDValue AnyExt = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, Op.getOperand(0));
11632 assert(OpVT.is128BitVector() && "Expected an SSE type!");
11633 return DAG.getBitcast(
11634 OpVT, DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32, AnyExt));
11637 // Lower a node with an EXTRACT_SUBVECTOR opcode. This may result in
11638 // a simple subregister reference or explicit instructions to grab
11639 // upper bits of a vector.
11640 static SDValue LowerEXTRACT_SUBVECTOR(SDValue Op, const X86Subtarget *Subtarget,
11641 SelectionDAG &DAG) {
11643 SDValue In = Op.getOperand(0);
11644 SDValue Idx = Op.getOperand(1);
11645 unsigned IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
11646 MVT ResVT = Op.getSimpleValueType();
11647 MVT InVT = In.getSimpleValueType();
11649 if (Subtarget->hasFp256()) {
11650 if (ResVT.is128BitVector() &&
11651 (InVT.is256BitVector() || InVT.is512BitVector()) &&
11652 isa<ConstantSDNode>(Idx)) {
11653 return Extract128BitVector(In, IdxVal, DAG, dl);
11655 if (ResVT.is256BitVector() && InVT.is512BitVector() &&
11656 isa<ConstantSDNode>(Idx)) {
11657 return Extract256BitVector(In, IdxVal, DAG, dl);
11663 // Lower a node with an INSERT_SUBVECTOR opcode. This may result in a
11664 // simple superregister reference or explicit instructions to insert
11665 // the upper bits of a vector.
11666 static SDValue LowerINSERT_SUBVECTOR(SDValue Op, const X86Subtarget *Subtarget,
11667 SelectionDAG &DAG) {
11668 if (!Subtarget->hasAVX())
11672 SDValue Vec = Op.getOperand(0);
11673 SDValue SubVec = Op.getOperand(1);
11674 SDValue Idx = Op.getOperand(2);
11676 if (!isa<ConstantSDNode>(Idx))
11679 unsigned IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
11680 MVT OpVT = Op.getSimpleValueType();
11681 MVT SubVecVT = SubVec.getSimpleValueType();
11683 // Fold two 16-byte subvector loads into one 32-byte load:
11684 // (insert_subvector (insert_subvector undef, (load addr), 0),
11685 // (load addr + 16), Elts/2)
11687 if ((IdxVal == OpVT.getVectorNumElements() / 2) &&
11688 Vec.getOpcode() == ISD::INSERT_SUBVECTOR &&
11689 OpVT.is256BitVector() && SubVecVT.is128BitVector()) {
11690 auto *Idx2 = dyn_cast<ConstantSDNode>(Vec.getOperand(2));
11691 if (Idx2 && Idx2->getZExtValue() == 0) {
11692 SDValue SubVec2 = Vec.getOperand(1);
11693 // If needed, look through a bitcast to get to the load.
11694 if (SubVec2.getNode() && SubVec2.getOpcode() == ISD::BITCAST)
11695 SubVec2 = SubVec2.getOperand(0);
11697 if (auto *FirstLd = dyn_cast<LoadSDNode>(SubVec2)) {
11699 unsigned Alignment = FirstLd->getAlignment();
11700 unsigned AS = FirstLd->getAddressSpace();
11701 const X86TargetLowering *TLI = Subtarget->getTargetLowering();
11702 if (TLI->allowsMemoryAccess(*DAG.getContext(), DAG.getDataLayout(),
11703 OpVT, AS, Alignment, &Fast) && Fast) {
11704 SDValue Ops[] = { SubVec2, SubVec };
11705 if (SDValue Ld = EltsFromConsecutiveLoads(OpVT, Ops, dl, DAG, false))
11712 if ((OpVT.is256BitVector() || OpVT.is512BitVector()) &&
11713 SubVecVT.is128BitVector())
11714 return Insert128BitVector(Vec, SubVec, IdxVal, DAG, dl);
11716 if (OpVT.is512BitVector() && SubVecVT.is256BitVector())
11717 return Insert256BitVector(Vec, SubVec, IdxVal, DAG, dl);
11719 if (OpVT.getVectorElementType() == MVT::i1) {
11720 if (IdxVal == 0 && Vec.getOpcode() == ISD::UNDEF) // the operation is legal
11722 SDValue ZeroIdx = DAG.getIntPtrConstant(0, dl);
11723 SDValue Undef = DAG.getUNDEF(OpVT);
11724 unsigned NumElems = OpVT.getVectorNumElements();
11725 SDValue ShiftBits = DAG.getConstant(NumElems/2, dl, MVT::i8);
11727 if (IdxVal == OpVT.getVectorNumElements() / 2) {
11728 // Zero upper bits of the Vec
11729 Vec = DAG.getNode(X86ISD::VSHLI, dl, OpVT, Vec, ShiftBits);
11730 Vec = DAG.getNode(X86ISD::VSRLI, dl, OpVT, Vec, ShiftBits);
11732 SDValue Vec2 = DAG.getNode(ISD::INSERT_SUBVECTOR, dl, OpVT, Undef,
11734 Vec2 = DAG.getNode(X86ISD::VSHLI, dl, OpVT, Vec2, ShiftBits);
11735 return DAG.getNode(ISD::OR, dl, OpVT, Vec, Vec2);
11738 SDValue Vec2 = DAG.getNode(ISD::INSERT_SUBVECTOR, dl, OpVT, Undef,
11740 // Zero upper bits of the Vec2
11741 Vec2 = DAG.getNode(X86ISD::VSHLI, dl, OpVT, Vec2, ShiftBits);
11742 Vec2 = DAG.getNode(X86ISD::VSRLI, dl, OpVT, Vec2, ShiftBits);
11743 // Zero lower bits of the Vec
11744 Vec = DAG.getNode(X86ISD::VSRLI, dl, OpVT, Vec, ShiftBits);
11745 Vec = DAG.getNode(X86ISD::VSHLI, dl, OpVT, Vec, ShiftBits);
11746 // Merge them together
11747 return DAG.getNode(ISD::OR, dl, OpVT, Vec, Vec2);
11753 // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
11754 // their target countpart wrapped in the X86ISD::Wrapper node. Suppose N is
11755 // one of the above mentioned nodes. It has to be wrapped because otherwise
11756 // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
11757 // be used to form addressing mode. These wrapped nodes will be selected
11760 X86TargetLowering::LowerConstantPool(SDValue Op, SelectionDAG &DAG) const {
11761 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
11763 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
11764 // global base reg.
11765 unsigned char OpFlag = 0;
11766 unsigned WrapperKind = X86ISD::Wrapper;
11767 CodeModel::Model M = DAG.getTarget().getCodeModel();
11769 if (Subtarget->isPICStyleRIPRel() &&
11770 (M == CodeModel::Small || M == CodeModel::Kernel))
11771 WrapperKind = X86ISD::WrapperRIP;
11772 else if (Subtarget->isPICStyleGOT())
11773 OpFlag = X86II::MO_GOTOFF;
11774 else if (Subtarget->isPICStyleStubPIC())
11775 OpFlag = X86II::MO_PIC_BASE_OFFSET;
11777 auto PtrVT = getPointerTy(DAG.getDataLayout());
11778 SDValue Result = DAG.getTargetConstantPool(
11779 CP->getConstVal(), PtrVT, CP->getAlignment(), CP->getOffset(), OpFlag);
11781 Result = DAG.getNode(WrapperKind, DL, PtrVT, Result);
11782 // With PIC, the address is actually $g + Offset.
11785 DAG.getNode(ISD::ADD, DL, PtrVT,
11786 DAG.getNode(X86ISD::GlobalBaseReg, SDLoc(), PtrVT), Result);
11792 SDValue X86TargetLowering::LowerJumpTable(SDValue Op, SelectionDAG &DAG) const {
11793 JumpTableSDNode *JT = cast<JumpTableSDNode>(Op);
11795 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
11796 // global base reg.
11797 unsigned char OpFlag = 0;
11798 unsigned WrapperKind = X86ISD::Wrapper;
11799 CodeModel::Model M = DAG.getTarget().getCodeModel();
11801 if (Subtarget->isPICStyleRIPRel() &&
11802 (M == CodeModel::Small || M == CodeModel::Kernel))
11803 WrapperKind = X86ISD::WrapperRIP;
11804 else if (Subtarget->isPICStyleGOT())
11805 OpFlag = X86II::MO_GOTOFF;
11806 else if (Subtarget->isPICStyleStubPIC())
11807 OpFlag = X86II::MO_PIC_BASE_OFFSET;
11809 auto PtrVT = getPointerTy(DAG.getDataLayout());
11810 SDValue Result = DAG.getTargetJumpTable(JT->getIndex(), PtrVT, OpFlag);
11812 Result = DAG.getNode(WrapperKind, DL, PtrVT, Result);
11814 // With PIC, the address is actually $g + Offset.
11817 DAG.getNode(ISD::ADD, DL, PtrVT,
11818 DAG.getNode(X86ISD::GlobalBaseReg, SDLoc(), PtrVT), Result);
11824 X86TargetLowering::LowerExternalSymbol(SDValue Op, SelectionDAG &DAG) const {
11825 const char *Sym = cast<ExternalSymbolSDNode>(Op)->getSymbol();
11827 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
11828 // global base reg.
11829 unsigned char OpFlag = 0;
11830 unsigned WrapperKind = X86ISD::Wrapper;
11831 CodeModel::Model M = DAG.getTarget().getCodeModel();
11833 if (Subtarget->isPICStyleRIPRel() &&
11834 (M == CodeModel::Small || M == CodeModel::Kernel)) {
11835 if (Subtarget->isTargetDarwin() || Subtarget->isTargetELF())
11836 OpFlag = X86II::MO_GOTPCREL;
11837 WrapperKind = X86ISD::WrapperRIP;
11838 } else if (Subtarget->isPICStyleGOT()) {
11839 OpFlag = X86II::MO_GOT;
11840 } else if (Subtarget->isPICStyleStubPIC()) {
11841 OpFlag = X86II::MO_DARWIN_NONLAZY_PIC_BASE;
11842 } else if (Subtarget->isPICStyleStubNoDynamic()) {
11843 OpFlag = X86II::MO_DARWIN_NONLAZY;
11846 auto PtrVT = getPointerTy(DAG.getDataLayout());
11847 SDValue Result = DAG.getTargetExternalSymbol(Sym, PtrVT, OpFlag);
11850 Result = DAG.getNode(WrapperKind, DL, PtrVT, Result);
11852 // With PIC, the address is actually $g + Offset.
11853 if (DAG.getTarget().getRelocationModel() == Reloc::PIC_ &&
11854 !Subtarget->is64Bit()) {
11856 DAG.getNode(ISD::ADD, DL, PtrVT,
11857 DAG.getNode(X86ISD::GlobalBaseReg, SDLoc(), PtrVT), Result);
11860 // For symbols that require a load from a stub to get the address, emit the
11862 if (isGlobalStubReference(OpFlag))
11863 Result = DAG.getLoad(PtrVT, DL, DAG.getEntryNode(), Result,
11864 MachinePointerInfo::getGOT(DAG.getMachineFunction()),
11865 false, false, false, 0);
11871 X86TargetLowering::LowerBlockAddress(SDValue Op, SelectionDAG &DAG) const {
11872 // Create the TargetBlockAddressAddress node.
11873 unsigned char OpFlags =
11874 Subtarget->ClassifyBlockAddressReference();
11875 CodeModel::Model M = DAG.getTarget().getCodeModel();
11876 const BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress();
11877 int64_t Offset = cast<BlockAddressSDNode>(Op)->getOffset();
11879 auto PtrVT = getPointerTy(DAG.getDataLayout());
11880 SDValue Result = DAG.getTargetBlockAddress(BA, PtrVT, Offset, OpFlags);
11882 if (Subtarget->isPICStyleRIPRel() &&
11883 (M == CodeModel::Small || M == CodeModel::Kernel))
11884 Result = DAG.getNode(X86ISD::WrapperRIP, dl, PtrVT, Result);
11886 Result = DAG.getNode(X86ISD::Wrapper, dl, PtrVT, Result);
11888 // With PIC, the address is actually $g + Offset.
11889 if (isGlobalRelativeToPICBase(OpFlags)) {
11890 Result = DAG.getNode(ISD::ADD, dl, PtrVT,
11891 DAG.getNode(X86ISD::GlobalBaseReg, dl, PtrVT), Result);
11898 X86TargetLowering::LowerGlobalAddress(const GlobalValue *GV, SDLoc dl,
11899 int64_t Offset, SelectionDAG &DAG) const {
11900 // Create the TargetGlobalAddress node, folding in the constant
11901 // offset if it is legal.
11902 unsigned char OpFlags =
11903 Subtarget->ClassifyGlobalReference(GV, DAG.getTarget());
11904 CodeModel::Model M = DAG.getTarget().getCodeModel();
11905 auto PtrVT = getPointerTy(DAG.getDataLayout());
11907 if (OpFlags == X86II::MO_NO_FLAG &&
11908 X86::isOffsetSuitableForCodeModel(Offset, M)) {
11909 // A direct static reference to a global.
11910 Result = DAG.getTargetGlobalAddress(GV, dl, PtrVT, Offset);
11913 Result = DAG.getTargetGlobalAddress(GV, dl, PtrVT, 0, OpFlags);
11916 if (Subtarget->isPICStyleRIPRel() &&
11917 (M == CodeModel::Small || M == CodeModel::Kernel))
11918 Result = DAG.getNode(X86ISD::WrapperRIP, dl, PtrVT, Result);
11920 Result = DAG.getNode(X86ISD::Wrapper, dl, PtrVT, Result);
11922 // With PIC, the address is actually $g + Offset.
11923 if (isGlobalRelativeToPICBase(OpFlags)) {
11924 Result = DAG.getNode(ISD::ADD, dl, PtrVT,
11925 DAG.getNode(X86ISD::GlobalBaseReg, dl, PtrVT), Result);
11928 // For globals that require a load from a stub to get the address, emit the
11930 if (isGlobalStubReference(OpFlags))
11931 Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Result,
11932 MachinePointerInfo::getGOT(DAG.getMachineFunction()),
11933 false, false, false, 0);
11935 // If there was a non-zero offset that we didn't fold, create an explicit
11936 // addition for it.
11938 Result = DAG.getNode(ISD::ADD, dl, PtrVT, Result,
11939 DAG.getConstant(Offset, dl, PtrVT));
11945 X86TargetLowering::LowerGlobalAddress(SDValue Op, SelectionDAG &DAG) const {
11946 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
11947 int64_t Offset = cast<GlobalAddressSDNode>(Op)->getOffset();
11948 return LowerGlobalAddress(GV, SDLoc(Op), Offset, DAG);
11952 GetTLSADDR(SelectionDAG &DAG, SDValue Chain, GlobalAddressSDNode *GA,
11953 SDValue *InFlag, const EVT PtrVT, unsigned ReturnReg,
11954 unsigned char OperandFlags, bool LocalDynamic = false) {
11955 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
11956 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
11958 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(), dl,
11959 GA->getValueType(0),
11963 X86ISD::NodeType CallType = LocalDynamic ? X86ISD::TLSBASEADDR
11967 SDValue Ops[] = { Chain, TGA, *InFlag };
11968 Chain = DAG.getNode(CallType, dl, NodeTys, Ops);
11970 SDValue Ops[] = { Chain, TGA };
11971 Chain = DAG.getNode(CallType, dl, NodeTys, Ops);
11974 // TLSADDR will be codegen'ed as call. Inform MFI that function has calls.
11975 MFI->setAdjustsStack(true);
11976 MFI->setHasCalls(true);
11978 SDValue Flag = Chain.getValue(1);
11979 return DAG.getCopyFromReg(Chain, dl, ReturnReg, PtrVT, Flag);
11982 // Lower ISD::GlobalTLSAddress using the "general dynamic" model, 32 bit
11984 LowerToTLSGeneralDynamicModel32(GlobalAddressSDNode *GA, SelectionDAG &DAG,
11987 SDLoc dl(GA); // ? function entry point might be better
11988 SDValue Chain = DAG.getCopyToReg(DAG.getEntryNode(), dl, X86::EBX,
11989 DAG.getNode(X86ISD::GlobalBaseReg,
11990 SDLoc(), PtrVT), InFlag);
11991 InFlag = Chain.getValue(1);
11993 return GetTLSADDR(DAG, Chain, GA, &InFlag, PtrVT, X86::EAX, X86II::MO_TLSGD);
11996 // Lower ISD::GlobalTLSAddress using the "general dynamic" model, 64 bit
11998 LowerToTLSGeneralDynamicModel64(GlobalAddressSDNode *GA, SelectionDAG &DAG,
12000 return GetTLSADDR(DAG, DAG.getEntryNode(), GA, nullptr, PtrVT,
12001 X86::RAX, X86II::MO_TLSGD);
12004 static SDValue LowerToTLSLocalDynamicModel(GlobalAddressSDNode *GA,
12010 // Get the start address of the TLS block for this module.
12011 X86MachineFunctionInfo* MFI = DAG.getMachineFunction()
12012 .getInfo<X86MachineFunctionInfo>();
12013 MFI->incNumLocalDynamicTLSAccesses();
12017 Base = GetTLSADDR(DAG, DAG.getEntryNode(), GA, nullptr, PtrVT, X86::RAX,
12018 X86II::MO_TLSLD, /*LocalDynamic=*/true);
12021 SDValue Chain = DAG.getCopyToReg(DAG.getEntryNode(), dl, X86::EBX,
12022 DAG.getNode(X86ISD::GlobalBaseReg, SDLoc(), PtrVT), InFlag);
12023 InFlag = Chain.getValue(1);
12024 Base = GetTLSADDR(DAG, Chain, GA, &InFlag, PtrVT, X86::EAX,
12025 X86II::MO_TLSLDM, /*LocalDynamic=*/true);
12028 // Note: the CleanupLocalDynamicTLSPass will remove redundant computations
12032 unsigned char OperandFlags = X86II::MO_DTPOFF;
12033 unsigned WrapperKind = X86ISD::Wrapper;
12034 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(), dl,
12035 GA->getValueType(0),
12036 GA->getOffset(), OperandFlags);
12037 SDValue Offset = DAG.getNode(WrapperKind, dl, PtrVT, TGA);
12039 // Add x@dtpoff with the base.
12040 return DAG.getNode(ISD::ADD, dl, PtrVT, Offset, Base);
12043 // Lower ISD::GlobalTLSAddress using the "initial exec" or "local exec" model.
12044 static SDValue LowerToTLSExecModel(GlobalAddressSDNode *GA, SelectionDAG &DAG,
12045 const EVT PtrVT, TLSModel::Model model,
12046 bool is64Bit, bool isPIC) {
12049 // Get the Thread Pointer, which is %gs:0 (32-bit) or %fs:0 (64-bit).
12050 Value *Ptr = Constant::getNullValue(Type::getInt8PtrTy(*DAG.getContext(),
12051 is64Bit ? 257 : 256));
12053 SDValue ThreadPointer =
12054 DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), DAG.getIntPtrConstant(0, dl),
12055 MachinePointerInfo(Ptr), false, false, false, 0);
12057 unsigned char OperandFlags = 0;
12058 // Most TLS accesses are not RIP relative, even on x86-64. One exception is
12060 unsigned WrapperKind = X86ISD::Wrapper;
12061 if (model == TLSModel::LocalExec) {
12062 OperandFlags = is64Bit ? X86II::MO_TPOFF : X86II::MO_NTPOFF;
12063 } else if (model == TLSModel::InitialExec) {
12065 OperandFlags = X86II::MO_GOTTPOFF;
12066 WrapperKind = X86ISD::WrapperRIP;
12068 OperandFlags = isPIC ? X86II::MO_GOTNTPOFF : X86II::MO_INDNTPOFF;
12071 llvm_unreachable("Unexpected model");
12074 // emit "addl x@ntpoff,%eax" (local exec)
12075 // or "addl x@indntpoff,%eax" (initial exec)
12076 // or "addl x@gotntpoff(%ebx) ,%eax" (initial exec, 32-bit pic)
12078 DAG.getTargetGlobalAddress(GA->getGlobal(), dl, GA->getValueType(0),
12079 GA->getOffset(), OperandFlags);
12080 SDValue Offset = DAG.getNode(WrapperKind, dl, PtrVT, TGA);
12082 if (model == TLSModel::InitialExec) {
12083 if (isPIC && !is64Bit) {
12084 Offset = DAG.getNode(ISD::ADD, dl, PtrVT,
12085 DAG.getNode(X86ISD::GlobalBaseReg, SDLoc(), PtrVT),
12089 Offset = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Offset,
12090 MachinePointerInfo::getGOT(DAG.getMachineFunction()),
12091 false, false, false, 0);
12094 // The address of the thread local variable is the add of the thread
12095 // pointer with the offset of the variable.
12096 return DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, Offset);
12100 X86TargetLowering::LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const {
12102 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
12103 const GlobalValue *GV = GA->getGlobal();
12104 auto PtrVT = getPointerTy(DAG.getDataLayout());
12106 if (Subtarget->isTargetELF()) {
12107 if (DAG.getTarget().Options.EmulatedTLS)
12108 return LowerToTLSEmulatedModel(GA, DAG);
12109 TLSModel::Model model = DAG.getTarget().getTLSModel(GV);
12111 case TLSModel::GeneralDynamic:
12112 if (Subtarget->is64Bit())
12113 return LowerToTLSGeneralDynamicModel64(GA, DAG, PtrVT);
12114 return LowerToTLSGeneralDynamicModel32(GA, DAG, PtrVT);
12115 case TLSModel::LocalDynamic:
12116 return LowerToTLSLocalDynamicModel(GA, DAG, PtrVT,
12117 Subtarget->is64Bit());
12118 case TLSModel::InitialExec:
12119 case TLSModel::LocalExec:
12120 return LowerToTLSExecModel(GA, DAG, PtrVT, model, Subtarget->is64Bit(),
12121 DAG.getTarget().getRelocationModel() ==
12124 llvm_unreachable("Unknown TLS model.");
12127 if (Subtarget->isTargetDarwin()) {
12128 // Darwin only has one model of TLS. Lower to that.
12129 unsigned char OpFlag = 0;
12130 unsigned WrapperKind = Subtarget->isPICStyleRIPRel() ?
12131 X86ISD::WrapperRIP : X86ISD::Wrapper;
12133 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
12134 // global base reg.
12135 bool PIC32 = (DAG.getTarget().getRelocationModel() == Reloc::PIC_) &&
12136 !Subtarget->is64Bit();
12138 OpFlag = X86II::MO_TLVP_PIC_BASE;
12140 OpFlag = X86II::MO_TLVP;
12142 SDValue Result = DAG.getTargetGlobalAddress(GA->getGlobal(), DL,
12143 GA->getValueType(0),
12144 GA->getOffset(), OpFlag);
12145 SDValue Offset = DAG.getNode(WrapperKind, DL, PtrVT, Result);
12147 // With PIC32, the address is actually $g + Offset.
12149 Offset = DAG.getNode(ISD::ADD, DL, PtrVT,
12150 DAG.getNode(X86ISD::GlobalBaseReg, SDLoc(), PtrVT),
12153 // Lowering the machine isd will make sure everything is in the right
12155 SDValue Chain = DAG.getEntryNode();
12156 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
12157 SDValue Args[] = { Chain, Offset };
12158 Chain = DAG.getNode(X86ISD::TLSCALL, DL, NodeTys, Args);
12160 // TLSCALL will be codegen'ed as call. Inform MFI that function has calls.
12161 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
12162 MFI->setAdjustsStack(true);
12164 // And our return value (tls address) is in the standard call return value
12166 unsigned Reg = Subtarget->is64Bit() ? X86::RAX : X86::EAX;
12167 return DAG.getCopyFromReg(Chain, DL, Reg, PtrVT, Chain.getValue(1));
12170 if (Subtarget->isTargetKnownWindowsMSVC() ||
12171 Subtarget->isTargetWindowsGNU()) {
12172 // Just use the implicit TLS architecture
12173 // Need to generate someting similar to:
12174 // mov rdx, qword [gs:abs 58H]; Load pointer to ThreadLocalStorage
12176 // mov ecx, dword [rel _tls_index]: Load index (from C runtime)
12177 // mov rcx, qword [rdx+rcx*8]
12178 // mov eax, .tls$:tlsvar
12179 // [rax+rcx] contains the address
12180 // Windows 64bit: gs:0x58
12181 // Windows 32bit: fs:__tls_array
12184 SDValue Chain = DAG.getEntryNode();
12186 // Get the Thread Pointer, which is %fs:__tls_array (32-bit) or
12187 // %gs:0x58 (64-bit). On MinGW, __tls_array is not available, so directly
12188 // use its literal value of 0x2C.
12189 Value *Ptr = Constant::getNullValue(Subtarget->is64Bit()
12190 ? Type::getInt8PtrTy(*DAG.getContext(),
12192 : Type::getInt32PtrTy(*DAG.getContext(),
12195 SDValue TlsArray = Subtarget->is64Bit()
12196 ? DAG.getIntPtrConstant(0x58, dl)
12197 : (Subtarget->isTargetWindowsGNU()
12198 ? DAG.getIntPtrConstant(0x2C, dl)
12199 : DAG.getExternalSymbol("_tls_array", PtrVT));
12201 SDValue ThreadPointer =
12202 DAG.getLoad(PtrVT, dl, Chain, TlsArray, MachinePointerInfo(Ptr), false,
12206 if (GV->getThreadLocalMode() == GlobalVariable::LocalExecTLSModel) {
12207 res = ThreadPointer;
12209 // Load the _tls_index variable
12210 SDValue IDX = DAG.getExternalSymbol("_tls_index", PtrVT);
12211 if (Subtarget->is64Bit())
12212 IDX = DAG.getExtLoad(ISD::ZEXTLOAD, dl, PtrVT, Chain, IDX,
12213 MachinePointerInfo(), MVT::i32, false, false,
12216 IDX = DAG.getLoad(PtrVT, dl, Chain, IDX, MachinePointerInfo(), false,
12219 auto &DL = DAG.getDataLayout();
12221 DAG.getConstant(Log2_64_Ceil(DL.getPointerSize()), dl, PtrVT);
12222 IDX = DAG.getNode(ISD::SHL, dl, PtrVT, IDX, Scale);
12224 res = DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, IDX);
12227 res = DAG.getLoad(PtrVT, dl, Chain, res, MachinePointerInfo(), false, false,
12230 // Get the offset of start of .tls section
12231 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(), dl,
12232 GA->getValueType(0),
12233 GA->getOffset(), X86II::MO_SECREL);
12234 SDValue Offset = DAG.getNode(X86ISD::Wrapper, dl, PtrVT, TGA);
12236 // The address of the thread local variable is the add of the thread
12237 // pointer with the offset of the variable.
12238 return DAG.getNode(ISD::ADD, dl, PtrVT, res, Offset);
12241 llvm_unreachable("TLS not implemented for this target.");
12244 /// LowerShiftParts - Lower SRA_PARTS and friends, which return two i32 values
12245 /// and take a 2 x i32 value to shift plus a shift amount.
12246 static SDValue LowerShiftParts(SDValue Op, SelectionDAG &DAG) {
12247 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
12248 MVT VT = Op.getSimpleValueType();
12249 unsigned VTBits = VT.getSizeInBits();
12251 bool isSRA = Op.getOpcode() == ISD::SRA_PARTS;
12252 SDValue ShOpLo = Op.getOperand(0);
12253 SDValue ShOpHi = Op.getOperand(1);
12254 SDValue ShAmt = Op.getOperand(2);
12255 // X86ISD::SHLD and X86ISD::SHRD have defined overflow behavior but the
12256 // generic ISD nodes haven't. Insert an AND to be safe, it's optimized away
12258 SDValue SafeShAmt = DAG.getNode(ISD::AND, dl, MVT::i8, ShAmt,
12259 DAG.getConstant(VTBits - 1, dl, MVT::i8));
12260 SDValue Tmp1 = isSRA ? DAG.getNode(ISD::SRA, dl, VT, ShOpHi,
12261 DAG.getConstant(VTBits - 1, dl, MVT::i8))
12262 : DAG.getConstant(0, dl, VT);
12264 SDValue Tmp2, Tmp3;
12265 if (Op.getOpcode() == ISD::SHL_PARTS) {
12266 Tmp2 = DAG.getNode(X86ISD::SHLD, dl, VT, ShOpHi, ShOpLo, ShAmt);
12267 Tmp3 = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, SafeShAmt);
12269 Tmp2 = DAG.getNode(X86ISD::SHRD, dl, VT, ShOpLo, ShOpHi, ShAmt);
12270 Tmp3 = DAG.getNode(isSRA ? ISD::SRA : ISD::SRL, dl, VT, ShOpHi, SafeShAmt);
12273 // If the shift amount is larger or equal than the width of a part we can't
12274 // rely on the results of shld/shrd. Insert a test and select the appropriate
12275 // values for large shift amounts.
12276 SDValue AndNode = DAG.getNode(ISD::AND, dl, MVT::i8, ShAmt,
12277 DAG.getConstant(VTBits, dl, MVT::i8));
12278 SDValue Cond = DAG.getNode(X86ISD::CMP, dl, MVT::i32,
12279 AndNode, DAG.getConstant(0, dl, MVT::i8));
12282 SDValue CC = DAG.getConstant(X86::COND_NE, dl, MVT::i8);
12283 SDValue Ops0[4] = { Tmp2, Tmp3, CC, Cond };
12284 SDValue Ops1[4] = { Tmp3, Tmp1, CC, Cond };
12286 if (Op.getOpcode() == ISD::SHL_PARTS) {
12287 Hi = DAG.getNode(X86ISD::CMOV, dl, VT, Ops0);
12288 Lo = DAG.getNode(X86ISD::CMOV, dl, VT, Ops1);
12290 Lo = DAG.getNode(X86ISD::CMOV, dl, VT, Ops0);
12291 Hi = DAG.getNode(X86ISD::CMOV, dl, VT, Ops1);
12294 SDValue Ops[2] = { Lo, Hi };
12295 return DAG.getMergeValues(Ops, dl);
12298 SDValue X86TargetLowering::LowerSINT_TO_FP(SDValue Op,
12299 SelectionDAG &DAG) const {
12300 SDValue Src = Op.getOperand(0);
12301 MVT SrcVT = Src.getSimpleValueType();
12302 MVT VT = Op.getSimpleValueType();
12305 if (SrcVT.isVector()) {
12306 if (SrcVT == MVT::v2i32 && VT == MVT::v2f64) {
12307 return DAG.getNode(X86ISD::CVTDQ2PD, dl, VT,
12308 DAG.getNode(ISD::CONCAT_VECTORS, dl, MVT::v4i32, Src,
12309 DAG.getUNDEF(SrcVT)));
12311 if (SrcVT.getVectorElementType() == MVT::i1) {
12312 MVT IntegerVT = MVT::getVectorVT(MVT::i32, SrcVT.getVectorNumElements());
12313 return DAG.getNode(ISD::SINT_TO_FP, dl, Op.getValueType(),
12314 DAG.getNode(ISD::SIGN_EXTEND, dl, IntegerVT, Src));
12319 assert(SrcVT <= MVT::i64 && SrcVT >= MVT::i16 &&
12320 "Unknown SINT_TO_FP to lower!");
12322 // These are really Legal; return the operand so the caller accepts it as
12324 if (SrcVT == MVT::i32 && isScalarFPTypeInSSEReg(Op.getValueType()))
12326 if (SrcVT == MVT::i64 && isScalarFPTypeInSSEReg(Op.getValueType()) &&
12327 Subtarget->is64Bit()) {
12331 unsigned Size = SrcVT.getSizeInBits()/8;
12332 MachineFunction &MF = DAG.getMachineFunction();
12333 auto PtrVT = getPointerTy(MF.getDataLayout());
12334 int SSFI = MF.getFrameInfo()->CreateStackObject(Size, Size, false);
12335 SDValue StackSlot = DAG.getFrameIndex(SSFI, PtrVT);
12336 SDValue Chain = DAG.getStore(
12337 DAG.getEntryNode(), dl, Op.getOperand(0), StackSlot,
12338 MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), SSFI), false,
12340 return BuildFILD(Op, SrcVT, Chain, StackSlot, DAG);
12343 SDValue X86TargetLowering::BuildFILD(SDValue Op, EVT SrcVT, SDValue Chain,
12345 SelectionDAG &DAG) const {
12349 bool useSSE = isScalarFPTypeInSSEReg(Op.getValueType());
12351 Tys = DAG.getVTList(MVT::f64, MVT::Other, MVT::Glue);
12353 Tys = DAG.getVTList(Op.getValueType(), MVT::Other);
12355 unsigned ByteSize = SrcVT.getSizeInBits()/8;
12357 FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(StackSlot);
12358 MachineMemOperand *MMO;
12360 int SSFI = FI->getIndex();
12361 MMO = DAG.getMachineFunction().getMachineMemOperand(
12362 MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), SSFI),
12363 MachineMemOperand::MOLoad, ByteSize, ByteSize);
12365 MMO = cast<LoadSDNode>(StackSlot)->getMemOperand();
12366 StackSlot = StackSlot.getOperand(1);
12368 SDValue Ops[] = { Chain, StackSlot, DAG.getValueType(SrcVT) };
12369 SDValue Result = DAG.getMemIntrinsicNode(useSSE ? X86ISD::FILD_FLAG :
12371 Tys, Ops, SrcVT, MMO);
12374 Chain = Result.getValue(1);
12375 SDValue InFlag = Result.getValue(2);
12377 // FIXME: Currently the FST is flagged to the FILD_FLAG. This
12378 // shouldn't be necessary except that RFP cannot be live across
12379 // multiple blocks. When stackifier is fixed, they can be uncoupled.
12380 MachineFunction &MF = DAG.getMachineFunction();
12381 unsigned SSFISize = Op.getValueType().getSizeInBits()/8;
12382 int SSFI = MF.getFrameInfo()->CreateStackObject(SSFISize, SSFISize, false);
12383 auto PtrVT = getPointerTy(MF.getDataLayout());
12384 SDValue StackSlot = DAG.getFrameIndex(SSFI, PtrVT);
12385 Tys = DAG.getVTList(MVT::Other);
12387 Chain, Result, StackSlot, DAG.getValueType(Op.getValueType()), InFlag
12389 MachineMemOperand *MMO = DAG.getMachineFunction().getMachineMemOperand(
12390 MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), SSFI),
12391 MachineMemOperand::MOStore, SSFISize, SSFISize);
12393 Chain = DAG.getMemIntrinsicNode(X86ISD::FST, DL, Tys,
12394 Ops, Op.getValueType(), MMO);
12395 Result = DAG.getLoad(
12396 Op.getValueType(), DL, Chain, StackSlot,
12397 MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), SSFI),
12398 false, false, false, 0);
12404 // LowerUINT_TO_FP_i64 - 64-bit unsigned integer to double expansion.
12405 SDValue X86TargetLowering::LowerUINT_TO_FP_i64(SDValue Op,
12406 SelectionDAG &DAG) const {
12407 // This algorithm is not obvious. Here it is what we're trying to output:
12410 punpckldq (c0), %xmm0 // c0: (uint4){ 0x43300000U, 0x45300000U, 0U, 0U }
12411 subpd (c1), %xmm0 // c1: (double2){ 0x1.0p52, 0x1.0p52 * 0x1.0p32 }
12413 haddpd %xmm0, %xmm0
12415 pshufd $0x4e, %xmm0, %xmm1
12421 LLVMContext *Context = DAG.getContext();
12423 // Build some magic constants.
12424 static const uint32_t CV0[] = { 0x43300000, 0x45300000, 0, 0 };
12425 Constant *C0 = ConstantDataVector::get(*Context, CV0);
12426 auto PtrVT = getPointerTy(DAG.getDataLayout());
12427 SDValue CPIdx0 = DAG.getConstantPool(C0, PtrVT, 16);
12429 SmallVector<Constant*,2> CV1;
12431 ConstantFP::get(*Context, APFloat(APFloat::IEEEdouble,
12432 APInt(64, 0x4330000000000000ULL))));
12434 ConstantFP::get(*Context, APFloat(APFloat::IEEEdouble,
12435 APInt(64, 0x4530000000000000ULL))));
12436 Constant *C1 = ConstantVector::get(CV1);
12437 SDValue CPIdx1 = DAG.getConstantPool(C1, PtrVT, 16);
12439 // Load the 64-bit value into an XMM register.
12440 SDValue XR1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i64,
12443 DAG.getLoad(MVT::v4i32, dl, DAG.getEntryNode(), CPIdx0,
12444 MachinePointerInfo::getConstantPool(DAG.getMachineFunction()),
12445 false, false, false, 16);
12447 getUnpackl(DAG, dl, MVT::v4i32, DAG.getBitcast(MVT::v4i32, XR1), CLod0);
12450 DAG.getLoad(MVT::v2f64, dl, CLod0.getValue(1), CPIdx1,
12451 MachinePointerInfo::getConstantPool(DAG.getMachineFunction()),
12452 false, false, false, 16);
12453 SDValue XR2F = DAG.getBitcast(MVT::v2f64, Unpck1);
12454 // TODO: Are there any fast-math-flags to propagate here?
12455 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::v2f64, XR2F, CLod1);
12458 if (Subtarget->hasSSE3()) {
12459 // FIXME: The 'haddpd' instruction may be slower than 'movhlps + addsd'.
12460 Result = DAG.getNode(X86ISD::FHADD, dl, MVT::v2f64, Sub, Sub);
12462 SDValue S2F = DAG.getBitcast(MVT::v4i32, Sub);
12463 SDValue Shuffle = getTargetShuffleNode(X86ISD::PSHUFD, dl, MVT::v4i32,
12465 Result = DAG.getNode(ISD::FADD, dl, MVT::v2f64,
12466 DAG.getBitcast(MVT::v2f64, Shuffle), Sub);
12469 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Result,
12470 DAG.getIntPtrConstant(0, dl));
12473 // LowerUINT_TO_FP_i32 - 32-bit unsigned integer to float expansion.
12474 SDValue X86TargetLowering::LowerUINT_TO_FP_i32(SDValue Op,
12475 SelectionDAG &DAG) const {
12477 // FP constant to bias correct the final result.
12478 SDValue Bias = DAG.getConstantFP(BitsToDouble(0x4330000000000000ULL), dl,
12481 // Load the 32-bit value into an XMM register.
12482 SDValue Load = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
12485 // Zero out the upper parts of the register.
12486 Load = getShuffleVectorZeroOrUndef(Load, 0, true, Subtarget, DAG);
12488 Load = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
12489 DAG.getBitcast(MVT::v2f64, Load),
12490 DAG.getIntPtrConstant(0, dl));
12492 // Or the load with the bias.
12493 SDValue Or = DAG.getNode(
12494 ISD::OR, dl, MVT::v2i64,
12495 DAG.getBitcast(MVT::v2i64,
12496 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2f64, Load)),
12497 DAG.getBitcast(MVT::v2i64,
12498 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2f64, Bias)));
12500 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
12501 DAG.getBitcast(MVT::v2f64, Or), DAG.getIntPtrConstant(0, dl));
12503 // Subtract the bias.
12504 // TODO: Are there any fast-math-flags to propagate here?
12505 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::f64, Or, Bias);
12507 // Handle final rounding.
12508 EVT DestVT = Op.getValueType();
12510 if (DestVT.bitsLT(MVT::f64))
12511 return DAG.getNode(ISD::FP_ROUND, dl, DestVT, Sub,
12512 DAG.getIntPtrConstant(0, dl));
12513 if (DestVT.bitsGT(MVT::f64))
12514 return DAG.getNode(ISD::FP_EXTEND, dl, DestVT, Sub);
12516 // Handle final rounding.
12520 static SDValue lowerUINT_TO_FP_vXi32(SDValue Op, SelectionDAG &DAG,
12521 const X86Subtarget &Subtarget) {
12522 // The algorithm is the following:
12523 // #ifdef __SSE4_1__
12524 // uint4 lo = _mm_blend_epi16( v, (uint4) 0x4b000000, 0xaa);
12525 // uint4 hi = _mm_blend_epi16( _mm_srli_epi32(v,16),
12526 // (uint4) 0x53000000, 0xaa);
12528 // uint4 lo = (v & (uint4) 0xffff) | (uint4) 0x4b000000;
12529 // uint4 hi = (v >> 16) | (uint4) 0x53000000;
12531 // float4 fhi = (float4) hi - (0x1.0p39f + 0x1.0p23f);
12532 // return (float4) lo + fhi;
12534 // We shouldn't use it when unsafe-fp-math is enabled though: we might later
12535 // reassociate the two FADDs, and if we do that, the algorithm fails
12536 // spectacularly (PR24512).
12537 // FIXME: If we ever have some kind of Machine FMF, this should be marked
12538 // as non-fast and always be enabled. Why isn't SDAG FMF enough? Because
12539 // there's also the MachineCombiner reassociations happening on Machine IR.
12540 if (DAG.getTarget().Options.UnsafeFPMath)
12544 SDValue V = Op->getOperand(0);
12545 EVT VecIntVT = V.getValueType();
12546 bool Is128 = VecIntVT == MVT::v4i32;
12547 EVT VecFloatVT = Is128 ? MVT::v4f32 : MVT::v8f32;
12548 // If we convert to something else than the supported type, e.g., to v4f64,
12550 if (VecFloatVT != Op->getValueType(0))
12553 unsigned NumElts = VecIntVT.getVectorNumElements();
12554 assert((VecIntVT == MVT::v4i32 || VecIntVT == MVT::v8i32) &&
12555 "Unsupported custom type");
12556 assert(NumElts <= 8 && "The size of the constant array must be fixed");
12558 // In the #idef/#else code, we have in common:
12559 // - The vector of constants:
12565 // Create the splat vector for 0x4b000000.
12566 SDValue CstLow = DAG.getConstant(0x4b000000, DL, MVT::i32);
12567 SDValue CstLowArray[] = {CstLow, CstLow, CstLow, CstLow,
12568 CstLow, CstLow, CstLow, CstLow};
12569 SDValue VecCstLow = DAG.getNode(ISD::BUILD_VECTOR, DL, VecIntVT,
12570 makeArrayRef(&CstLowArray[0], NumElts));
12571 // Create the splat vector for 0x53000000.
12572 SDValue CstHigh = DAG.getConstant(0x53000000, DL, MVT::i32);
12573 SDValue CstHighArray[] = {CstHigh, CstHigh, CstHigh, CstHigh,
12574 CstHigh, CstHigh, CstHigh, CstHigh};
12575 SDValue VecCstHigh = DAG.getNode(ISD::BUILD_VECTOR, DL, VecIntVT,
12576 makeArrayRef(&CstHighArray[0], NumElts));
12578 // Create the right shift.
12579 SDValue CstShift = DAG.getConstant(16, DL, MVT::i32);
12580 SDValue CstShiftArray[] = {CstShift, CstShift, CstShift, CstShift,
12581 CstShift, CstShift, CstShift, CstShift};
12582 SDValue VecCstShift = DAG.getNode(ISD::BUILD_VECTOR, DL, VecIntVT,
12583 makeArrayRef(&CstShiftArray[0], NumElts));
12584 SDValue HighShift = DAG.getNode(ISD::SRL, DL, VecIntVT, V, VecCstShift);
12587 if (Subtarget.hasSSE41()) {
12588 EVT VecI16VT = Is128 ? MVT::v8i16 : MVT::v16i16;
12589 // uint4 lo = _mm_blend_epi16( v, (uint4) 0x4b000000, 0xaa);
12590 SDValue VecCstLowBitcast = DAG.getBitcast(VecI16VT, VecCstLow);
12591 SDValue VecBitcast = DAG.getBitcast(VecI16VT, V);
12592 // Low will be bitcasted right away, so do not bother bitcasting back to its
12594 Low = DAG.getNode(X86ISD::BLENDI, DL, VecI16VT, VecBitcast,
12595 VecCstLowBitcast, DAG.getConstant(0xaa, DL, MVT::i32));
12596 // uint4 hi = _mm_blend_epi16( _mm_srli_epi32(v,16),
12597 // (uint4) 0x53000000, 0xaa);
12598 SDValue VecCstHighBitcast = DAG.getBitcast(VecI16VT, VecCstHigh);
12599 SDValue VecShiftBitcast = DAG.getBitcast(VecI16VT, HighShift);
12600 // High will be bitcasted right away, so do not bother bitcasting back to
12601 // its original type.
12602 High = DAG.getNode(X86ISD::BLENDI, DL, VecI16VT, VecShiftBitcast,
12603 VecCstHighBitcast, DAG.getConstant(0xaa, DL, MVT::i32));
12605 SDValue CstMask = DAG.getConstant(0xffff, DL, MVT::i32);
12606 SDValue VecCstMask = DAG.getNode(ISD::BUILD_VECTOR, DL, VecIntVT, CstMask,
12607 CstMask, CstMask, CstMask);
12608 // uint4 lo = (v & (uint4) 0xffff) | (uint4) 0x4b000000;
12609 SDValue LowAnd = DAG.getNode(ISD::AND, DL, VecIntVT, V, VecCstMask);
12610 Low = DAG.getNode(ISD::OR, DL, VecIntVT, LowAnd, VecCstLow);
12612 // uint4 hi = (v >> 16) | (uint4) 0x53000000;
12613 High = DAG.getNode(ISD::OR, DL, VecIntVT, HighShift, VecCstHigh);
12616 // Create the vector constant for -(0x1.0p39f + 0x1.0p23f).
12617 SDValue CstFAdd = DAG.getConstantFP(
12618 APFloat(APFloat::IEEEsingle, APInt(32, 0xD3000080)), DL, MVT::f32);
12619 SDValue CstFAddArray[] = {CstFAdd, CstFAdd, CstFAdd, CstFAdd,
12620 CstFAdd, CstFAdd, CstFAdd, CstFAdd};
12621 SDValue VecCstFAdd = DAG.getNode(ISD::BUILD_VECTOR, DL, VecFloatVT,
12622 makeArrayRef(&CstFAddArray[0], NumElts));
12624 // float4 fhi = (float4) hi - (0x1.0p39f + 0x1.0p23f);
12625 SDValue HighBitcast = DAG.getBitcast(VecFloatVT, High);
12626 // TODO: Are there any fast-math-flags to propagate here?
12628 DAG.getNode(ISD::FADD, DL, VecFloatVT, HighBitcast, VecCstFAdd);
12629 // return (float4) lo + fhi;
12630 SDValue LowBitcast = DAG.getBitcast(VecFloatVT, Low);
12631 return DAG.getNode(ISD::FADD, DL, VecFloatVT, LowBitcast, FHigh);
12634 SDValue X86TargetLowering::lowerUINT_TO_FP_vec(SDValue Op,
12635 SelectionDAG &DAG) const {
12636 SDValue N0 = Op.getOperand(0);
12637 MVT SVT = N0.getSimpleValueType();
12640 switch (SVT.SimpleTy) {
12642 llvm_unreachable("Custom UINT_TO_FP is not supported!");
12647 MVT NVT = MVT::getVectorVT(MVT::i32, SVT.getVectorNumElements());
12648 return DAG.getNode(ISD::SINT_TO_FP, dl, Op.getValueType(),
12649 DAG.getNode(ISD::ZERO_EXTEND, dl, NVT, N0));
12653 return lowerUINT_TO_FP_vXi32(Op, DAG, *Subtarget);
12656 if (Subtarget->hasAVX512())
12657 return DAG.getNode(ISD::UINT_TO_FP, dl, Op.getValueType(),
12658 DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v16i32, N0));
12660 llvm_unreachable(nullptr);
12663 SDValue X86TargetLowering::LowerUINT_TO_FP(SDValue Op,
12664 SelectionDAG &DAG) const {
12665 SDValue N0 = Op.getOperand(0);
12667 auto PtrVT = getPointerTy(DAG.getDataLayout());
12669 if (Op.getValueType().isVector())
12670 return lowerUINT_TO_FP_vec(Op, DAG);
12672 // Since UINT_TO_FP is legal (it's marked custom), dag combiner won't
12673 // optimize it to a SINT_TO_FP when the sign bit is known zero. Perform
12674 // the optimization here.
12675 if (DAG.SignBitIsZero(N0))
12676 return DAG.getNode(ISD::SINT_TO_FP, dl, Op.getValueType(), N0);
12678 MVT SrcVT = N0.getSimpleValueType();
12679 MVT DstVT = Op.getSimpleValueType();
12681 if (Subtarget->hasAVX512() && isScalarFPTypeInSSEReg(DstVT) &&
12682 (SrcVT == MVT::i32 || (SrcVT == MVT::i64 && Subtarget->is64Bit()))) {
12683 // Conversions from unsigned i32 to f32/f64 are legal,
12684 // using VCVTUSI2SS/SD. Same for i64 in 64-bit mode.
12688 if (SrcVT == MVT::i64 && DstVT == MVT::f64 && X86ScalarSSEf64)
12689 return LowerUINT_TO_FP_i64(Op, DAG);
12690 if (SrcVT == MVT::i32 && X86ScalarSSEf64)
12691 return LowerUINT_TO_FP_i32(Op, DAG);
12692 if (Subtarget->is64Bit() && SrcVT == MVT::i64 && DstVT == MVT::f32)
12695 // Make a 64-bit buffer, and use it to build an FILD.
12696 SDValue StackSlot = DAG.CreateStackTemporary(MVT::i64);
12697 if (SrcVT == MVT::i32) {
12698 SDValue WordOff = DAG.getConstant(4, dl, PtrVT);
12699 SDValue OffsetSlot = DAG.getNode(ISD::ADD, dl, PtrVT, StackSlot, WordOff);
12700 SDValue Store1 = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
12701 StackSlot, MachinePointerInfo(),
12703 SDValue Store2 = DAG.getStore(Store1, dl, DAG.getConstant(0, dl, MVT::i32),
12704 OffsetSlot, MachinePointerInfo(),
12706 SDValue Fild = BuildFILD(Op, MVT::i64, Store2, StackSlot, DAG);
12710 assert(SrcVT == MVT::i64 && "Unexpected type in UINT_TO_FP");
12711 SDValue Store = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
12712 StackSlot, MachinePointerInfo(),
12714 // For i64 source, we need to add the appropriate power of 2 if the input
12715 // was negative. This is the same as the optimization in
12716 // DAGTypeLegalizer::ExpandIntOp_UNIT_TO_FP, and for it to be safe here,
12717 // we must be careful to do the computation in x87 extended precision, not
12718 // in SSE. (The generic code can't know it's OK to do this, or how to.)
12719 int SSFI = cast<FrameIndexSDNode>(StackSlot)->getIndex();
12720 MachineMemOperand *MMO = DAG.getMachineFunction().getMachineMemOperand(
12721 MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), SSFI),
12722 MachineMemOperand::MOLoad, 8, 8);
12724 SDVTList Tys = DAG.getVTList(MVT::f80, MVT::Other);
12725 SDValue Ops[] = { Store, StackSlot, DAG.getValueType(MVT::i64) };
12726 SDValue Fild = DAG.getMemIntrinsicNode(X86ISD::FILD, dl, Tys, Ops,
12729 APInt FF(32, 0x5F800000ULL);
12731 // Check whether the sign bit is set.
12732 SDValue SignSet = DAG.getSetCC(
12733 dl, getSetCCResultType(DAG.getDataLayout(), *DAG.getContext(), MVT::i64),
12734 Op.getOperand(0), DAG.getConstant(0, dl, MVT::i64), ISD::SETLT);
12736 // Build a 64 bit pair (0, FF) in the constant pool, with FF in the lo bits.
12737 SDValue FudgePtr = DAG.getConstantPool(
12738 ConstantInt::get(*DAG.getContext(), FF.zext(64)), PtrVT);
12740 // Get a pointer to FF if the sign bit was set, or to 0 otherwise.
12741 SDValue Zero = DAG.getIntPtrConstant(0, dl);
12742 SDValue Four = DAG.getIntPtrConstant(4, dl);
12743 SDValue Offset = DAG.getNode(ISD::SELECT, dl, Zero.getValueType(), SignSet,
12745 FudgePtr = DAG.getNode(ISD::ADD, dl, PtrVT, FudgePtr, Offset);
12747 // Load the value out, extending it from f32 to f80.
12748 // FIXME: Avoid the extend by constructing the right constant pool?
12749 SDValue Fudge = DAG.getExtLoad(
12750 ISD::EXTLOAD, dl, MVT::f80, DAG.getEntryNode(), FudgePtr,
12751 MachinePointerInfo::getConstantPool(DAG.getMachineFunction()), MVT::f32,
12752 false, false, false, 4);
12753 // Extend everything to 80 bits to force it to be done on x87.
12754 // TODO: Are there any fast-math-flags to propagate here?
12755 SDValue Add = DAG.getNode(ISD::FADD, dl, MVT::f80, Fild, Fudge);
12756 return DAG.getNode(ISD::FP_ROUND, dl, DstVT, Add,
12757 DAG.getIntPtrConstant(0, dl));
12760 // If the given FP_TO_SINT (IsSigned) or FP_TO_UINT (!IsSigned) operation
12761 // is legal, or has an fp128 or f16 source (which needs to be promoted to f32),
12762 // just return an <SDValue(), SDValue()> pair.
12763 // Otherwise it is assumed to be a conversion from one of f32, f64 or f80
12764 // to i16, i32 or i64, and we lower it to a legal sequence.
12765 // If lowered to the final integer result we return a <result, SDValue()> pair.
12766 // Otherwise we lower it to a sequence ending with a FIST, return a
12767 // <FIST, StackSlot> pair, and the caller is responsible for loading
12768 // the final integer result from StackSlot.
12769 std::pair<SDValue,SDValue>
12770 X86TargetLowering::FP_TO_INTHelper(SDValue Op, SelectionDAG &DAG,
12771 bool IsSigned, bool IsReplace) const {
12774 EVT DstTy = Op.getValueType();
12775 EVT TheVT = Op.getOperand(0).getValueType();
12776 auto PtrVT = getPointerTy(DAG.getDataLayout());
12778 if (TheVT != MVT::f32 && TheVT != MVT::f64 && TheVT != MVT::f80) {
12779 // f16 must be promoted before using the lowering in this routine.
12780 // fp128 does not use this lowering.
12781 return std::make_pair(SDValue(), SDValue());
12784 // If using FIST to compute an unsigned i64, we'll need some fixup
12785 // to handle values above the maximum signed i64. A FIST is always
12786 // used for the 32-bit subtarget, but also for f80 on a 64-bit target.
12787 bool UnsignedFixup = !IsSigned &&
12788 DstTy == MVT::i64 &&
12789 (!Subtarget->is64Bit() ||
12790 !isScalarFPTypeInSSEReg(TheVT));
12792 if (!IsSigned && DstTy != MVT::i64 && !Subtarget->hasAVX512()) {
12793 // Replace the fp-to-uint32 operation with an fp-to-sint64 FIST.
12794 // The low 32 bits of the fist result will have the correct uint32 result.
12795 assert(DstTy == MVT::i32 && "Unexpected FP_TO_UINT");
12799 assert(DstTy.getSimpleVT() <= MVT::i64 &&
12800 DstTy.getSimpleVT() >= MVT::i16 &&
12801 "Unknown FP_TO_INT to lower!");
12803 // These are really Legal.
12804 if (DstTy == MVT::i32 &&
12805 isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType()))
12806 return std::make_pair(SDValue(), SDValue());
12807 if (Subtarget->is64Bit() &&
12808 DstTy == MVT::i64 &&
12809 isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType()))
12810 return std::make_pair(SDValue(), SDValue());
12812 // We lower FP->int64 into FISTP64 followed by a load from a temporary
12814 MachineFunction &MF = DAG.getMachineFunction();
12815 unsigned MemSize = DstTy.getSizeInBits()/8;
12816 int SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize, false);
12817 SDValue StackSlot = DAG.getFrameIndex(SSFI, PtrVT);
12820 switch (DstTy.getSimpleVT().SimpleTy) {
12821 default: llvm_unreachable("Invalid FP_TO_SINT to lower!");
12822 case MVT::i16: Opc = X86ISD::FP_TO_INT16_IN_MEM; break;
12823 case MVT::i32: Opc = X86ISD::FP_TO_INT32_IN_MEM; break;
12824 case MVT::i64: Opc = X86ISD::FP_TO_INT64_IN_MEM; break;
12827 SDValue Chain = DAG.getEntryNode();
12828 SDValue Value = Op.getOperand(0);
12829 SDValue Adjust; // 0x0 or 0x80000000, for result sign bit adjustment.
12831 if (UnsignedFixup) {
12833 // Conversion to unsigned i64 is implemented with a select,
12834 // depending on whether the source value fits in the range
12835 // of a signed i64. Let Thresh be the FP equivalent of
12836 // 0x8000000000000000ULL.
12838 // Adjust i32 = (Value < Thresh) ? 0 : 0x80000000;
12839 // FistSrc = (Value < Thresh) ? Value : (Value - Thresh);
12840 // Fist-to-mem64 FistSrc
12841 // Add 0 or 0x800...0ULL to the 64-bit result, which is equivalent
12842 // to XOR'ing the high 32 bits with Adjust.
12844 // Being a power of 2, Thresh is exactly representable in all FP formats.
12845 // For X87 we'd like to use the smallest FP type for this constant, but
12846 // for DAG type consistency we have to match the FP operand type.
12848 APFloat Thresh(APFloat::IEEEsingle, APInt(32, 0x5f000000));
12849 LLVM_ATTRIBUTE_UNUSED APFloat::opStatus Status = APFloat::opOK;
12850 bool LosesInfo = false;
12851 if (TheVT == MVT::f64)
12852 // The rounding mode is irrelevant as the conversion should be exact.
12853 Status = Thresh.convert(APFloat::IEEEdouble, APFloat::rmNearestTiesToEven,
12855 else if (TheVT == MVT::f80)
12856 Status = Thresh.convert(APFloat::x87DoubleExtended,
12857 APFloat::rmNearestTiesToEven, &LosesInfo);
12859 assert(Status == APFloat::opOK && !LosesInfo &&
12860 "FP conversion should have been exact");
12862 SDValue ThreshVal = DAG.getConstantFP(Thresh, DL, TheVT);
12864 SDValue Cmp = DAG.getSetCC(DL,
12865 getSetCCResultType(DAG.getDataLayout(),
12866 *DAG.getContext(), TheVT),
12867 Value, ThreshVal, ISD::SETLT);
12868 Adjust = DAG.getSelect(DL, MVT::i32, Cmp,
12869 DAG.getConstant(0, DL, MVT::i32),
12870 DAG.getConstant(0x80000000, DL, MVT::i32));
12871 SDValue Sub = DAG.getNode(ISD::FSUB, DL, TheVT, Value, ThreshVal);
12872 Cmp = DAG.getSetCC(DL, getSetCCResultType(DAG.getDataLayout(),
12873 *DAG.getContext(), TheVT),
12874 Value, ThreshVal, ISD::SETLT);
12875 Value = DAG.getSelect(DL, TheVT, Cmp, Value, Sub);
12878 // FIXME This causes a redundant load/store if the SSE-class value is already
12879 // in memory, such as if it is on the callstack.
12880 if (isScalarFPTypeInSSEReg(TheVT)) {
12881 assert(DstTy == MVT::i64 && "Invalid FP_TO_SINT to lower!");
12882 Chain = DAG.getStore(Chain, DL, Value, StackSlot,
12883 MachinePointerInfo::getFixedStack(MF, SSFI), false,
12885 SDVTList Tys = DAG.getVTList(Op.getOperand(0).getValueType(), MVT::Other);
12887 Chain, StackSlot, DAG.getValueType(TheVT)
12890 MachineMemOperand *MMO =
12891 MF.getMachineMemOperand(MachinePointerInfo::getFixedStack(MF, SSFI),
12892 MachineMemOperand::MOLoad, MemSize, MemSize);
12893 Value = DAG.getMemIntrinsicNode(X86ISD::FLD, DL, Tys, Ops, DstTy, MMO);
12894 Chain = Value.getValue(1);
12895 SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize, false);
12896 StackSlot = DAG.getFrameIndex(SSFI, PtrVT);
12899 MachineMemOperand *MMO =
12900 MF.getMachineMemOperand(MachinePointerInfo::getFixedStack(MF, SSFI),
12901 MachineMemOperand::MOStore, MemSize, MemSize);
12903 if (UnsignedFixup) {
12905 // Insert the FIST, load its result as two i32's,
12906 // and XOR the high i32 with Adjust.
12908 SDValue FistOps[] = { Chain, Value, StackSlot };
12909 SDValue FIST = DAG.getMemIntrinsicNode(Opc, DL, DAG.getVTList(MVT::Other),
12910 FistOps, DstTy, MMO);
12912 SDValue Low32 = DAG.getLoad(MVT::i32, DL, FIST, StackSlot,
12913 MachinePointerInfo(),
12914 false, false, false, 0);
12915 SDValue HighAddr = DAG.getNode(ISD::ADD, DL, PtrVT, StackSlot,
12916 DAG.getConstant(4, DL, PtrVT));
12918 SDValue High32 = DAG.getLoad(MVT::i32, DL, FIST, HighAddr,
12919 MachinePointerInfo(),
12920 false, false, false, 0);
12921 High32 = DAG.getNode(ISD::XOR, DL, MVT::i32, High32, Adjust);
12923 if (Subtarget->is64Bit()) {
12924 // Join High32 and Low32 into a 64-bit result.
12925 // (High32 << 32) | Low32
12926 Low32 = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::i64, Low32);
12927 High32 = DAG.getNode(ISD::ANY_EXTEND, DL, MVT::i64, High32);
12928 High32 = DAG.getNode(ISD::SHL, DL, MVT::i64, High32,
12929 DAG.getConstant(32, DL, MVT::i8));
12930 SDValue Result = DAG.getNode(ISD::OR, DL, MVT::i64, High32, Low32);
12931 return std::make_pair(Result, SDValue());
12934 SDValue ResultOps[] = { Low32, High32 };
12936 SDValue pair = IsReplace
12937 ? DAG.getNode(ISD::BUILD_PAIR, DL, MVT::i64, ResultOps)
12938 : DAG.getMergeValues(ResultOps, DL);
12939 return std::make_pair(pair, SDValue());
12941 // Build the FP_TO_INT*_IN_MEM
12942 SDValue Ops[] = { Chain, Value, StackSlot };
12943 SDValue FIST = DAG.getMemIntrinsicNode(Opc, DL, DAG.getVTList(MVT::Other),
12945 return std::make_pair(FIST, StackSlot);
12949 static SDValue LowerAVXExtend(SDValue Op, SelectionDAG &DAG,
12950 const X86Subtarget *Subtarget) {
12951 MVT VT = Op->getSimpleValueType(0);
12952 SDValue In = Op->getOperand(0);
12953 MVT InVT = In.getSimpleValueType();
12956 if (VT.is512BitVector() || InVT.getScalarType() == MVT::i1)
12957 return DAG.getNode(ISD::ZERO_EXTEND, dl, VT, In);
12959 // Optimize vectors in AVX mode:
12962 // Use vpunpcklwd for 4 lower elements v8i16 -> v4i32.
12963 // Use vpunpckhwd for 4 upper elements v8i16 -> v4i32.
12964 // Concat upper and lower parts.
12967 // Use vpunpckldq for 4 lower elements v4i32 -> v2i64.
12968 // Use vpunpckhdq for 4 upper elements v4i32 -> v2i64.
12969 // Concat upper and lower parts.
12972 if (((VT != MVT::v16i16) || (InVT != MVT::v16i8)) &&
12973 ((VT != MVT::v8i32) || (InVT != MVT::v8i16)) &&
12974 ((VT != MVT::v4i64) || (InVT != MVT::v4i32)))
12977 if (Subtarget->hasInt256())
12978 return DAG.getNode(X86ISD::VZEXT, dl, VT, In);
12980 SDValue ZeroVec = getZeroVector(InVT, Subtarget, DAG, dl);
12981 SDValue Undef = DAG.getUNDEF(InVT);
12982 bool NeedZero = Op.getOpcode() == ISD::ZERO_EXTEND;
12983 SDValue OpLo = getUnpackl(DAG, dl, InVT, In, NeedZero ? ZeroVec : Undef);
12984 SDValue OpHi = getUnpackh(DAG, dl, InVT, In, NeedZero ? ZeroVec : Undef);
12986 MVT HVT = MVT::getVectorVT(VT.getVectorElementType(),
12987 VT.getVectorNumElements()/2);
12989 OpLo = DAG.getBitcast(HVT, OpLo);
12990 OpHi = DAG.getBitcast(HVT, OpHi);
12992 return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT, OpLo, OpHi);
12995 static SDValue LowerZERO_EXTEND_AVX512(SDValue Op,
12996 const X86Subtarget *Subtarget, SelectionDAG &DAG) {
12997 MVT VT = Op->getSimpleValueType(0);
12998 SDValue In = Op->getOperand(0);
12999 MVT InVT = In.getSimpleValueType();
13001 unsigned int NumElts = VT.getVectorNumElements();
13002 if (NumElts != 8 && NumElts != 16 && !Subtarget->hasBWI())
13005 if (VT.is512BitVector() && InVT.getVectorElementType() != MVT::i1)
13006 return DAG.getNode(X86ISD::VZEXT, DL, VT, In);
13008 assert(InVT.getVectorElementType() == MVT::i1);
13009 MVT ExtVT = NumElts == 8 ? MVT::v8i64 : MVT::v16i32;
13011 DAG.getConstant(APInt(ExtVT.getScalarSizeInBits(), 1), DL, ExtVT);
13013 DAG.getConstant(APInt::getNullValue(ExtVT.getScalarSizeInBits()), DL, ExtVT);
13015 SDValue V = DAG.getNode(ISD::VSELECT, DL, ExtVT, In, One, Zero);
13016 if (VT.is512BitVector())
13018 return DAG.getNode(X86ISD::VTRUNC, DL, VT, V);
13021 static SDValue LowerANY_EXTEND(SDValue Op, const X86Subtarget *Subtarget,
13022 SelectionDAG &DAG) {
13023 if (Subtarget->hasFp256())
13024 if (SDValue Res = LowerAVXExtend(Op, DAG, Subtarget))
13030 static SDValue LowerZERO_EXTEND(SDValue Op, const X86Subtarget *Subtarget,
13031 SelectionDAG &DAG) {
13033 MVT VT = Op.getSimpleValueType();
13034 SDValue In = Op.getOperand(0);
13035 MVT SVT = In.getSimpleValueType();
13037 if (VT.is512BitVector() || SVT.getVectorElementType() == MVT::i1)
13038 return LowerZERO_EXTEND_AVX512(Op, Subtarget, DAG);
13040 if (Subtarget->hasFp256())
13041 if (SDValue Res = LowerAVXExtend(Op, DAG, Subtarget))
13044 assert(!VT.is256BitVector() || !SVT.is128BitVector() ||
13045 VT.getVectorNumElements() != SVT.getVectorNumElements());
13049 SDValue X86TargetLowering::LowerTRUNCATE(SDValue Op, SelectionDAG &DAG) const {
13051 MVT VT = Op.getSimpleValueType();
13052 SDValue In = Op.getOperand(0);
13053 MVT InVT = In.getSimpleValueType();
13055 if (VT == MVT::i1) {
13056 assert((InVT.isInteger() && (InVT.getSizeInBits() <= 64)) &&
13057 "Invalid scalar TRUNCATE operation");
13058 if (InVT.getSizeInBits() >= 32)
13060 In = DAG.getNode(ISD::ANY_EXTEND, DL, MVT::i32, In);
13061 return DAG.getNode(ISD::TRUNCATE, DL, VT, In);
13063 assert(VT.getVectorNumElements() == InVT.getVectorNumElements() &&
13064 "Invalid TRUNCATE operation");
13066 // move vector to mask - truncate solution for SKX
13067 if (VT.getVectorElementType() == MVT::i1) {
13068 if (InVT.is512BitVector() && InVT.getScalarSizeInBits() <= 16 &&
13069 Subtarget->hasBWI())
13070 return Op; // legal, will go to VPMOVB2M, VPMOVW2M
13071 if ((InVT.is256BitVector() || InVT.is128BitVector())
13072 && InVT.getScalarSizeInBits() <= 16 &&
13073 Subtarget->hasBWI() && Subtarget->hasVLX())
13074 return Op; // legal, will go to VPMOVB2M, VPMOVW2M
13075 if (InVT.is512BitVector() && InVT.getScalarSizeInBits() >= 32 &&
13076 Subtarget->hasDQI())
13077 return Op; // legal, will go to VPMOVD2M, VPMOVQ2M
13078 if ((InVT.is256BitVector() || InVT.is128BitVector())
13079 && InVT.getScalarSizeInBits() >= 32 &&
13080 Subtarget->hasDQI() && Subtarget->hasVLX())
13081 return Op; // legal, will go to VPMOVB2M, VPMOVQ2M
13084 if (VT.getVectorElementType() == MVT::i1) {
13085 assert(VT.getVectorElementType() == MVT::i1 && "Unexpected vector type");
13086 unsigned NumElts = InVT.getVectorNumElements();
13087 assert ((NumElts == 8 || NumElts == 16) && "Unexpected vector type");
13088 if (InVT.getSizeInBits() < 512) {
13089 MVT ExtVT = (NumElts == 16)? MVT::v16i32 : MVT::v8i64;
13090 In = DAG.getNode(ISD::SIGN_EXTEND, DL, ExtVT, In);
13095 DAG.getConstant(APInt::getSignBit(InVT.getScalarSizeInBits()), DL, InVT);
13096 SDValue And = DAG.getNode(ISD::AND, DL, InVT, OneV, In);
13097 return DAG.getNode(X86ISD::TESTM, DL, VT, And, And);
13100 // vpmovqb/w/d, vpmovdb/w, vpmovwb
13101 if (((!InVT.is512BitVector() && Subtarget->hasVLX()) || InVT.is512BitVector()) &&
13102 (InVT.getVectorElementType() != MVT::i16 || Subtarget->hasBWI()))
13103 return DAG.getNode(X86ISD::VTRUNC, DL, VT, In);
13105 if ((VT == MVT::v4i32) && (InVT == MVT::v4i64)) {
13106 // On AVX2, v4i64 -> v4i32 becomes VPERMD.
13107 if (Subtarget->hasInt256()) {
13108 static const int ShufMask[] = {0, 2, 4, 6, -1, -1, -1, -1};
13109 In = DAG.getBitcast(MVT::v8i32, In);
13110 In = DAG.getVectorShuffle(MVT::v8i32, DL, In, DAG.getUNDEF(MVT::v8i32),
13112 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, VT, In,
13113 DAG.getIntPtrConstant(0, DL));
13116 SDValue OpLo = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v2i64, In,
13117 DAG.getIntPtrConstant(0, DL));
13118 SDValue OpHi = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v2i64, In,
13119 DAG.getIntPtrConstant(2, DL));
13120 OpLo = DAG.getBitcast(MVT::v4i32, OpLo);
13121 OpHi = DAG.getBitcast(MVT::v4i32, OpHi);
13122 static const int ShufMask[] = {0, 2, 4, 6};
13123 return DAG.getVectorShuffle(VT, DL, OpLo, OpHi, ShufMask);
13126 if ((VT == MVT::v8i16) && (InVT == MVT::v8i32)) {
13127 // On AVX2, v8i32 -> v8i16 becomed PSHUFB.
13128 if (Subtarget->hasInt256()) {
13129 In = DAG.getBitcast(MVT::v32i8, In);
13131 SmallVector<SDValue,32> pshufbMask;
13132 for (unsigned i = 0; i < 2; ++i) {
13133 pshufbMask.push_back(DAG.getConstant(0x0, DL, MVT::i8));
13134 pshufbMask.push_back(DAG.getConstant(0x1, DL, MVT::i8));
13135 pshufbMask.push_back(DAG.getConstant(0x4, DL, MVT::i8));
13136 pshufbMask.push_back(DAG.getConstant(0x5, DL, MVT::i8));
13137 pshufbMask.push_back(DAG.getConstant(0x8, DL, MVT::i8));
13138 pshufbMask.push_back(DAG.getConstant(0x9, DL, MVT::i8));
13139 pshufbMask.push_back(DAG.getConstant(0xc, DL, MVT::i8));
13140 pshufbMask.push_back(DAG.getConstant(0xd, DL, MVT::i8));
13141 for (unsigned j = 0; j < 8; ++j)
13142 pshufbMask.push_back(DAG.getConstant(0x80, DL, MVT::i8));
13144 SDValue BV = DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v32i8, pshufbMask);
13145 In = DAG.getNode(X86ISD::PSHUFB, DL, MVT::v32i8, In, BV);
13146 In = DAG.getBitcast(MVT::v4i64, In);
13148 static const int ShufMask[] = {0, 2, -1, -1};
13149 In = DAG.getVectorShuffle(MVT::v4i64, DL, In, DAG.getUNDEF(MVT::v4i64),
13151 In = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v2i64, In,
13152 DAG.getIntPtrConstant(0, DL));
13153 return DAG.getBitcast(VT, In);
13156 SDValue OpLo = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v4i32, In,
13157 DAG.getIntPtrConstant(0, DL));
13159 SDValue OpHi = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v4i32, In,
13160 DAG.getIntPtrConstant(4, DL));
13162 OpLo = DAG.getBitcast(MVT::v16i8, OpLo);
13163 OpHi = DAG.getBitcast(MVT::v16i8, OpHi);
13165 // The PSHUFB mask:
13166 static const int ShufMask1[] = {0, 1, 4, 5, 8, 9, 12, 13,
13167 -1, -1, -1, -1, -1, -1, -1, -1};
13169 SDValue Undef = DAG.getUNDEF(MVT::v16i8);
13170 OpLo = DAG.getVectorShuffle(MVT::v16i8, DL, OpLo, Undef, ShufMask1);
13171 OpHi = DAG.getVectorShuffle(MVT::v16i8, DL, OpHi, Undef, ShufMask1);
13173 OpLo = DAG.getBitcast(MVT::v4i32, OpLo);
13174 OpHi = DAG.getBitcast(MVT::v4i32, OpHi);
13176 // The MOVLHPS Mask:
13177 static const int ShufMask2[] = {0, 1, 4, 5};
13178 SDValue res = DAG.getVectorShuffle(MVT::v4i32, DL, OpLo, OpHi, ShufMask2);
13179 return DAG.getBitcast(MVT::v8i16, res);
13182 // Handle truncation of V256 to V128 using shuffles.
13183 if (!VT.is128BitVector() || !InVT.is256BitVector())
13186 assert(Subtarget->hasFp256() && "256-bit vector without AVX!");
13188 unsigned NumElems = VT.getVectorNumElements();
13189 MVT NVT = MVT::getVectorVT(VT.getVectorElementType(), NumElems * 2);
13191 SmallVector<int, 16> MaskVec(NumElems * 2, -1);
13192 // Prepare truncation shuffle mask
13193 for (unsigned i = 0; i != NumElems; ++i)
13194 MaskVec[i] = i * 2;
13195 SDValue V = DAG.getVectorShuffle(NVT, DL, DAG.getBitcast(NVT, In),
13196 DAG.getUNDEF(NVT), &MaskVec[0]);
13197 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, VT, V,
13198 DAG.getIntPtrConstant(0, DL));
13201 SDValue X86TargetLowering::LowerFP_TO_SINT(SDValue Op,
13202 SelectionDAG &DAG) const {
13203 assert(!Op.getSimpleValueType().isVector());
13205 std::pair<SDValue,SDValue> Vals = FP_TO_INTHelper(Op, DAG,
13206 /*IsSigned=*/ true, /*IsReplace=*/ false);
13207 SDValue FIST = Vals.first, StackSlot = Vals.second;
13208 // If FP_TO_INTHelper failed, the node is actually supposed to be Legal.
13209 if (!FIST.getNode())
13212 if (StackSlot.getNode())
13213 // Load the result.
13214 return DAG.getLoad(Op.getValueType(), SDLoc(Op),
13215 FIST, StackSlot, MachinePointerInfo(),
13216 false, false, false, 0);
13218 // The node is the result.
13222 SDValue X86TargetLowering::LowerFP_TO_UINT(SDValue Op,
13223 SelectionDAG &DAG) const {
13224 std::pair<SDValue,SDValue> Vals = FP_TO_INTHelper(Op, DAG,
13225 /*IsSigned=*/ false, /*IsReplace=*/ false);
13226 SDValue FIST = Vals.first, StackSlot = Vals.second;
13227 // If FP_TO_INTHelper failed, the node is actually supposed to be Legal.
13228 if (!FIST.getNode())
13231 if (StackSlot.getNode())
13232 // Load the result.
13233 return DAG.getLoad(Op.getValueType(), SDLoc(Op),
13234 FIST, StackSlot, MachinePointerInfo(),
13235 false, false, false, 0);
13237 // The node is the result.
13241 static SDValue LowerFP_EXTEND(SDValue Op, SelectionDAG &DAG) {
13243 MVT VT = Op.getSimpleValueType();
13244 SDValue In = Op.getOperand(0);
13245 MVT SVT = In.getSimpleValueType();
13247 assert(SVT == MVT::v2f32 && "Only customize MVT::v2f32 type legalization!");
13249 return DAG.getNode(X86ISD::VFPEXT, DL, VT,
13250 DAG.getNode(ISD::CONCAT_VECTORS, DL, MVT::v4f32,
13251 In, DAG.getUNDEF(SVT)));
13254 /// The only differences between FABS and FNEG are the mask and the logic op.
13255 /// FNEG also has a folding opportunity for FNEG(FABS(x)).
13256 static SDValue LowerFABSorFNEG(SDValue Op, SelectionDAG &DAG) {
13257 assert((Op.getOpcode() == ISD::FABS || Op.getOpcode() == ISD::FNEG) &&
13258 "Wrong opcode for lowering FABS or FNEG.");
13260 bool IsFABS = (Op.getOpcode() == ISD::FABS);
13262 // If this is a FABS and it has an FNEG user, bail out to fold the combination
13263 // into an FNABS. We'll lower the FABS after that if it is still in use.
13265 for (SDNode *User : Op->uses())
13266 if (User->getOpcode() == ISD::FNEG)
13270 MVT VT = Op.getSimpleValueType();
13272 // FIXME: Use function attribute "OptimizeForSize" and/or CodeGenOpt::Level to
13273 // decide if we should generate a 16-byte constant mask when we only need 4 or
13274 // 8 bytes for the scalar case.
13280 if (VT.isVector()) {
13282 EltVT = VT.getVectorElementType();
13283 NumElts = VT.getVectorNumElements();
13285 // There are no scalar bitwise logical SSE/AVX instructions, so we
13286 // generate a 16-byte vector constant and logic op even for the scalar case.
13287 // Using a 16-byte mask allows folding the load of the mask with
13288 // the logic op, so it can save (~4 bytes) on code size.
13289 LogicVT = (VT == MVT::f64) ? MVT::v2f64 : MVT::v4f32;
13291 NumElts = (VT == MVT::f64) ? 2 : 4;
13294 unsigned EltBits = EltVT.getSizeInBits();
13295 LLVMContext *Context = DAG.getContext();
13296 // For FABS, mask is 0x7f...; for FNEG, mask is 0x80...
13298 IsFABS ? APInt::getSignedMaxValue(EltBits) : APInt::getSignBit(EltBits);
13299 Constant *C = ConstantInt::get(*Context, MaskElt);
13300 C = ConstantVector::getSplat(NumElts, C);
13301 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
13302 SDValue CPIdx = DAG.getConstantPool(C, TLI.getPointerTy(DAG.getDataLayout()));
13303 unsigned Alignment = cast<ConstantPoolSDNode>(CPIdx)->getAlignment();
13305 DAG.getLoad(LogicVT, dl, DAG.getEntryNode(), CPIdx,
13306 MachinePointerInfo::getConstantPool(DAG.getMachineFunction()),
13307 false, false, false, Alignment);
13309 SDValue Op0 = Op.getOperand(0);
13310 bool IsFNABS = !IsFABS && (Op0.getOpcode() == ISD::FABS);
13312 IsFABS ? X86ISD::FAND : IsFNABS ? X86ISD::FOR : X86ISD::FXOR;
13313 SDValue Operand = IsFNABS ? Op0.getOperand(0) : Op0;
13316 return DAG.getNode(LogicOp, dl, LogicVT, Operand, Mask);
13318 // For the scalar case extend to a 128-bit vector, perform the logic op,
13319 // and extract the scalar result back out.
13320 Operand = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, LogicVT, Operand);
13321 SDValue LogicNode = DAG.getNode(LogicOp, dl, LogicVT, Operand, Mask);
13322 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, LogicNode,
13323 DAG.getIntPtrConstant(0, dl));
13326 static SDValue LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) {
13327 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
13328 LLVMContext *Context = DAG.getContext();
13329 SDValue Op0 = Op.getOperand(0);
13330 SDValue Op1 = Op.getOperand(1);
13332 MVT VT = Op.getSimpleValueType();
13333 MVT SrcVT = Op1.getSimpleValueType();
13335 // If second operand is smaller, extend it first.
13336 if (SrcVT.bitsLT(VT)) {
13337 Op1 = DAG.getNode(ISD::FP_EXTEND, dl, VT, Op1);
13340 // And if it is bigger, shrink it first.
13341 if (SrcVT.bitsGT(VT)) {
13342 Op1 = DAG.getNode(ISD::FP_ROUND, dl, VT, Op1, DAG.getIntPtrConstant(1, dl));
13346 // At this point the operands and the result should have the same
13347 // type, and that won't be f80 since that is not custom lowered.
13349 const fltSemantics &Sem =
13350 VT == MVT::f64 ? APFloat::IEEEdouble : APFloat::IEEEsingle;
13351 const unsigned SizeInBits = VT.getSizeInBits();
13353 SmallVector<Constant *, 4> CV(
13354 VT == MVT::f64 ? 2 : 4,
13355 ConstantFP::get(*Context, APFloat(Sem, APInt(SizeInBits, 0))));
13357 // First, clear all bits but the sign bit from the second operand (sign).
13358 CV[0] = ConstantFP::get(*Context,
13359 APFloat(Sem, APInt::getHighBitsSet(SizeInBits, 1)));
13360 Constant *C = ConstantVector::get(CV);
13361 auto PtrVT = TLI.getPointerTy(DAG.getDataLayout());
13362 SDValue CPIdx = DAG.getConstantPool(C, PtrVT, 16);
13364 // Perform all logic operations as 16-byte vectors because there are no
13365 // scalar FP logic instructions in SSE. This allows load folding of the
13366 // constants into the logic instructions.
13367 MVT LogicVT = (VT == MVT::f64) ? MVT::v2f64 : MVT::v4f32;
13369 DAG.getLoad(LogicVT, dl, DAG.getEntryNode(), CPIdx,
13370 MachinePointerInfo::getConstantPool(DAG.getMachineFunction()),
13371 false, false, false, 16);
13372 Op1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, LogicVT, Op1);
13373 SDValue SignBit = DAG.getNode(X86ISD::FAND, dl, LogicVT, Op1, Mask1);
13375 // Next, clear the sign bit from the first operand (magnitude).
13376 // If it's a constant, we can clear it here.
13377 if (ConstantFPSDNode *Op0CN = dyn_cast<ConstantFPSDNode>(Op0)) {
13378 APFloat APF = Op0CN->getValueAPF();
13379 // If the magnitude is a positive zero, the sign bit alone is enough.
13380 if (APF.isPosZero())
13381 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, SrcVT, SignBit,
13382 DAG.getIntPtrConstant(0, dl));
13384 CV[0] = ConstantFP::get(*Context, APF);
13386 CV[0] = ConstantFP::get(
13388 APFloat(Sem, APInt::getLowBitsSet(SizeInBits, SizeInBits - 1)));
13390 C = ConstantVector::get(CV);
13391 CPIdx = DAG.getConstantPool(C, PtrVT, 16);
13393 DAG.getLoad(LogicVT, dl, DAG.getEntryNode(), CPIdx,
13394 MachinePointerInfo::getConstantPool(DAG.getMachineFunction()),
13395 false, false, false, 16);
13396 // If the magnitude operand wasn't a constant, we need to AND out the sign.
13397 if (!isa<ConstantFPSDNode>(Op0)) {
13398 Op0 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, LogicVT, Op0);
13399 Val = DAG.getNode(X86ISD::FAND, dl, LogicVT, Op0, Val);
13401 // OR the magnitude value with the sign bit.
13402 Val = DAG.getNode(X86ISD::FOR, dl, LogicVT, Val, SignBit);
13403 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, SrcVT, Val,
13404 DAG.getIntPtrConstant(0, dl));
13407 static SDValue LowerFGETSIGN(SDValue Op, SelectionDAG &DAG) {
13408 SDValue N0 = Op.getOperand(0);
13410 MVT VT = Op.getSimpleValueType();
13412 // Lower ISD::FGETSIGN to (AND (X86ISD::FGETSIGNx86 ...) 1).
13413 SDValue xFGETSIGN = DAG.getNode(X86ISD::FGETSIGNx86, dl, VT, N0,
13414 DAG.getConstant(1, dl, VT));
13415 return DAG.getNode(ISD::AND, dl, VT, xFGETSIGN, DAG.getConstant(1, dl, VT));
13418 // Check whether an OR'd tree is PTEST-able.
13419 static SDValue LowerVectorAllZeroTest(SDValue Op, const X86Subtarget *Subtarget,
13420 SelectionDAG &DAG) {
13421 assert(Op.getOpcode() == ISD::OR && "Only check OR'd tree.");
13423 if (!Subtarget->hasSSE41())
13426 if (!Op->hasOneUse())
13429 SDNode *N = Op.getNode();
13432 SmallVector<SDValue, 8> Opnds;
13433 DenseMap<SDValue, unsigned> VecInMap;
13434 SmallVector<SDValue, 8> VecIns;
13435 EVT VT = MVT::Other;
13437 // Recognize a special case where a vector is casted into wide integer to
13439 Opnds.push_back(N->getOperand(0));
13440 Opnds.push_back(N->getOperand(1));
13442 for (unsigned Slot = 0, e = Opnds.size(); Slot < e; ++Slot) {
13443 SmallVectorImpl<SDValue>::const_iterator I = Opnds.begin() + Slot;
13444 // BFS traverse all OR'd operands.
13445 if (I->getOpcode() == ISD::OR) {
13446 Opnds.push_back(I->getOperand(0));
13447 Opnds.push_back(I->getOperand(1));
13448 // Re-evaluate the number of nodes to be traversed.
13449 e += 2; // 2 more nodes (LHS and RHS) are pushed.
13453 // Quit if a non-EXTRACT_VECTOR_ELT
13454 if (I->getOpcode() != ISD::EXTRACT_VECTOR_ELT)
13457 // Quit if without a constant index.
13458 SDValue Idx = I->getOperand(1);
13459 if (!isa<ConstantSDNode>(Idx))
13462 SDValue ExtractedFromVec = I->getOperand(0);
13463 DenseMap<SDValue, unsigned>::iterator M = VecInMap.find(ExtractedFromVec);
13464 if (M == VecInMap.end()) {
13465 VT = ExtractedFromVec.getValueType();
13466 // Quit if not 128/256-bit vector.
13467 if (!VT.is128BitVector() && !VT.is256BitVector())
13469 // Quit if not the same type.
13470 if (VecInMap.begin() != VecInMap.end() &&
13471 VT != VecInMap.begin()->first.getValueType())
13473 M = VecInMap.insert(std::make_pair(ExtractedFromVec, 0)).first;
13474 VecIns.push_back(ExtractedFromVec);
13476 M->second |= 1U << cast<ConstantSDNode>(Idx)->getZExtValue();
13479 assert((VT.is128BitVector() || VT.is256BitVector()) &&
13480 "Not extracted from 128-/256-bit vector.");
13482 unsigned FullMask = (1U << VT.getVectorNumElements()) - 1U;
13484 for (DenseMap<SDValue, unsigned>::const_iterator
13485 I = VecInMap.begin(), E = VecInMap.end(); I != E; ++I) {
13486 // Quit if not all elements are used.
13487 if (I->second != FullMask)
13491 EVT TestVT = VT.is128BitVector() ? MVT::v2i64 : MVT::v4i64;
13493 // Cast all vectors into TestVT for PTEST.
13494 for (unsigned i = 0, e = VecIns.size(); i < e; ++i)
13495 VecIns[i] = DAG.getBitcast(TestVT, VecIns[i]);
13497 // If more than one full vectors are evaluated, OR them first before PTEST.
13498 for (unsigned Slot = 0, e = VecIns.size(); e - Slot > 1; Slot += 2, e += 1) {
13499 // Each iteration will OR 2 nodes and append the result until there is only
13500 // 1 node left, i.e. the final OR'd value of all vectors.
13501 SDValue LHS = VecIns[Slot];
13502 SDValue RHS = VecIns[Slot + 1];
13503 VecIns.push_back(DAG.getNode(ISD::OR, DL, TestVT, LHS, RHS));
13506 return DAG.getNode(X86ISD::PTEST, DL, MVT::i32,
13507 VecIns.back(), VecIns.back());
13510 /// \brief return true if \c Op has a use that doesn't just read flags.
13511 static bool hasNonFlagsUse(SDValue Op) {
13512 for (SDNode::use_iterator UI = Op->use_begin(), UE = Op->use_end(); UI != UE;
13514 SDNode *User = *UI;
13515 unsigned UOpNo = UI.getOperandNo();
13516 if (User->getOpcode() == ISD::TRUNCATE && User->hasOneUse()) {
13517 // Look pass truncate.
13518 UOpNo = User->use_begin().getOperandNo();
13519 User = *User->use_begin();
13522 if (User->getOpcode() != ISD::BRCOND && User->getOpcode() != ISD::SETCC &&
13523 !(User->getOpcode() == ISD::SELECT && UOpNo == 0))
13529 /// Emit nodes that will be selected as "test Op0,Op0", or something
13531 SDValue X86TargetLowering::EmitTest(SDValue Op, unsigned X86CC, SDLoc dl,
13532 SelectionDAG &DAG) const {
13533 if (Op.getValueType() == MVT::i1) {
13534 SDValue ExtOp = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i8, Op);
13535 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, ExtOp,
13536 DAG.getConstant(0, dl, MVT::i8));
13538 // CF and OF aren't always set the way we want. Determine which
13539 // of these we need.
13540 bool NeedCF = false;
13541 bool NeedOF = false;
13544 case X86::COND_A: case X86::COND_AE:
13545 case X86::COND_B: case X86::COND_BE:
13548 case X86::COND_G: case X86::COND_GE:
13549 case X86::COND_L: case X86::COND_LE:
13550 case X86::COND_O: case X86::COND_NO: {
13551 // Check if we really need to set the
13552 // Overflow flag. If NoSignedWrap is present
13553 // that is not actually needed.
13554 switch (Op->getOpcode()) {
13559 const auto *BinNode = cast<BinaryWithFlagsSDNode>(Op.getNode());
13560 if (BinNode->Flags.hasNoSignedWrap())
13570 // See if we can use the EFLAGS value from the operand instead of
13571 // doing a separate TEST. TEST always sets OF and CF to 0, so unless
13572 // we prove that the arithmetic won't overflow, we can't use OF or CF.
13573 if (Op.getResNo() != 0 || NeedOF || NeedCF) {
13574 // Emit a CMP with 0, which is the TEST pattern.
13575 //if (Op.getValueType() == MVT::i1)
13576 // return DAG.getNode(X86ISD::CMP, dl, MVT::i1, Op,
13577 // DAG.getConstant(0, MVT::i1));
13578 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op,
13579 DAG.getConstant(0, dl, Op.getValueType()));
13581 unsigned Opcode = 0;
13582 unsigned NumOperands = 0;
13584 // Truncate operations may prevent the merge of the SETCC instruction
13585 // and the arithmetic instruction before it. Attempt to truncate the operands
13586 // of the arithmetic instruction and use a reduced bit-width instruction.
13587 bool NeedTruncation = false;
13588 SDValue ArithOp = Op;
13589 if (Op->getOpcode() == ISD::TRUNCATE && Op->hasOneUse()) {
13590 SDValue Arith = Op->getOperand(0);
13591 // Both the trunc and the arithmetic op need to have one user each.
13592 if (Arith->hasOneUse())
13593 switch (Arith.getOpcode()) {
13600 NeedTruncation = true;
13606 // NOTICE: In the code below we use ArithOp to hold the arithmetic operation
13607 // which may be the result of a CAST. We use the variable 'Op', which is the
13608 // non-casted variable when we check for possible users.
13609 switch (ArithOp.getOpcode()) {
13611 // Due to an isel shortcoming, be conservative if this add is likely to be
13612 // selected as part of a load-modify-store instruction. When the root node
13613 // in a match is a store, isel doesn't know how to remap non-chain non-flag
13614 // uses of other nodes in the match, such as the ADD in this case. This
13615 // leads to the ADD being left around and reselected, with the result being
13616 // two adds in the output. Alas, even if none our users are stores, that
13617 // doesn't prove we're O.K. Ergo, if we have any parents that aren't
13618 // CopyToReg or SETCC, eschew INC/DEC. A better fix seems to require
13619 // climbing the DAG back to the root, and it doesn't seem to be worth the
13621 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
13622 UE = Op.getNode()->use_end(); UI != UE; ++UI)
13623 if (UI->getOpcode() != ISD::CopyToReg &&
13624 UI->getOpcode() != ISD::SETCC &&
13625 UI->getOpcode() != ISD::STORE)
13628 if (ConstantSDNode *C =
13629 dyn_cast<ConstantSDNode>(ArithOp.getNode()->getOperand(1))) {
13630 // An add of one will be selected as an INC.
13631 if (C->getAPIntValue() == 1 && !Subtarget->slowIncDec()) {
13632 Opcode = X86ISD::INC;
13637 // An add of negative one (subtract of one) will be selected as a DEC.
13638 if (C->getAPIntValue().isAllOnesValue() && !Subtarget->slowIncDec()) {
13639 Opcode = X86ISD::DEC;
13645 // Otherwise use a regular EFLAGS-setting add.
13646 Opcode = X86ISD::ADD;
13651 // If we have a constant logical shift that's only used in a comparison
13652 // against zero turn it into an equivalent AND. This allows turning it into
13653 // a TEST instruction later.
13654 if ((X86CC == X86::COND_E || X86CC == X86::COND_NE) && Op->hasOneUse() &&
13655 isa<ConstantSDNode>(Op->getOperand(1)) && !hasNonFlagsUse(Op)) {
13656 EVT VT = Op.getValueType();
13657 unsigned BitWidth = VT.getSizeInBits();
13658 unsigned ShAmt = Op->getConstantOperandVal(1);
13659 if (ShAmt >= BitWidth) // Avoid undefined shifts.
13661 APInt Mask = ArithOp.getOpcode() == ISD::SRL
13662 ? APInt::getHighBitsSet(BitWidth, BitWidth - ShAmt)
13663 : APInt::getLowBitsSet(BitWidth, BitWidth - ShAmt);
13664 if (!Mask.isSignedIntN(32)) // Avoid large immediates.
13666 SDValue New = DAG.getNode(ISD::AND, dl, VT, Op->getOperand(0),
13667 DAG.getConstant(Mask, dl, VT));
13668 DAG.ReplaceAllUsesWith(Op, New);
13674 // If the primary and result isn't used, don't bother using X86ISD::AND,
13675 // because a TEST instruction will be better.
13676 if (!hasNonFlagsUse(Op))
13682 // Due to the ISEL shortcoming noted above, be conservative if this op is
13683 // likely to be selected as part of a load-modify-store instruction.
13684 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
13685 UE = Op.getNode()->use_end(); UI != UE; ++UI)
13686 if (UI->getOpcode() == ISD::STORE)
13689 // Otherwise use a regular EFLAGS-setting instruction.
13690 switch (ArithOp.getOpcode()) {
13691 default: llvm_unreachable("unexpected operator!");
13692 case ISD::SUB: Opcode = X86ISD::SUB; break;
13693 case ISD::XOR: Opcode = X86ISD::XOR; break;
13694 case ISD::AND: Opcode = X86ISD::AND; break;
13696 if (!NeedTruncation && (X86CC == X86::COND_E || X86CC == X86::COND_NE)) {
13697 SDValue EFLAGS = LowerVectorAllZeroTest(Op, Subtarget, DAG);
13698 if (EFLAGS.getNode())
13701 Opcode = X86ISD::OR;
13715 return SDValue(Op.getNode(), 1);
13721 // If we found that truncation is beneficial, perform the truncation and
13723 if (NeedTruncation) {
13724 EVT VT = Op.getValueType();
13725 SDValue WideVal = Op->getOperand(0);
13726 EVT WideVT = WideVal.getValueType();
13727 unsigned ConvertedOp = 0;
13728 // Use a target machine opcode to prevent further DAGCombine
13729 // optimizations that may separate the arithmetic operations
13730 // from the setcc node.
13731 switch (WideVal.getOpcode()) {
13733 case ISD::ADD: ConvertedOp = X86ISD::ADD; break;
13734 case ISD::SUB: ConvertedOp = X86ISD::SUB; break;
13735 case ISD::AND: ConvertedOp = X86ISD::AND; break;
13736 case ISD::OR: ConvertedOp = X86ISD::OR; break;
13737 case ISD::XOR: ConvertedOp = X86ISD::XOR; break;
13741 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
13742 if (TLI.isOperationLegal(WideVal.getOpcode(), WideVT)) {
13743 SDValue V0 = DAG.getNode(ISD::TRUNCATE, dl, VT, WideVal.getOperand(0));
13744 SDValue V1 = DAG.getNode(ISD::TRUNCATE, dl, VT, WideVal.getOperand(1));
13745 Op = DAG.getNode(ConvertedOp, dl, VT, V0, V1);
13751 // Emit a CMP with 0, which is the TEST pattern.
13752 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op,
13753 DAG.getConstant(0, dl, Op.getValueType()));
13755 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::i32);
13756 SmallVector<SDValue, 4> Ops(Op->op_begin(), Op->op_begin() + NumOperands);
13758 SDValue New = DAG.getNode(Opcode, dl, VTs, Ops);
13759 DAG.ReplaceAllUsesWith(Op, New);
13760 return SDValue(New.getNode(), 1);
13763 /// Emit nodes that will be selected as "cmp Op0,Op1", or something
13765 SDValue X86TargetLowering::EmitCmp(SDValue Op0, SDValue Op1, unsigned X86CC,
13766 SDLoc dl, SelectionDAG &DAG) const {
13767 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op1)) {
13768 if (C->getAPIntValue() == 0)
13769 return EmitTest(Op0, X86CC, dl, DAG);
13771 if (Op0.getValueType() == MVT::i1)
13772 llvm_unreachable("Unexpected comparison operation for MVT::i1 operands");
13775 if ((Op0.getValueType() == MVT::i8 || Op0.getValueType() == MVT::i16 ||
13776 Op0.getValueType() == MVT::i32 || Op0.getValueType() == MVT::i64)) {
13777 // Do the comparison at i32 if it's smaller, besides the Atom case.
13778 // This avoids subregister aliasing issues. Keep the smaller reference
13779 // if we're optimizing for size, however, as that'll allow better folding
13780 // of memory operations.
13781 if (Op0.getValueType() != MVT::i32 && Op0.getValueType() != MVT::i64 &&
13782 !DAG.getMachineFunction().getFunction()->optForMinSize() &&
13783 !Subtarget->isAtom()) {
13784 unsigned ExtendOp =
13785 isX86CCUnsigned(X86CC) ? ISD::ZERO_EXTEND : ISD::SIGN_EXTEND;
13786 Op0 = DAG.getNode(ExtendOp, dl, MVT::i32, Op0);
13787 Op1 = DAG.getNode(ExtendOp, dl, MVT::i32, Op1);
13789 // Use SUB instead of CMP to enable CSE between SUB and CMP.
13790 SDVTList VTs = DAG.getVTList(Op0.getValueType(), MVT::i32);
13791 SDValue Sub = DAG.getNode(X86ISD::SUB, dl, VTs,
13793 return SDValue(Sub.getNode(), 1);
13795 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op0, Op1);
13798 /// Convert a comparison if required by the subtarget.
13799 SDValue X86TargetLowering::ConvertCmpIfNecessary(SDValue Cmp,
13800 SelectionDAG &DAG) const {
13801 // If the subtarget does not support the FUCOMI instruction, floating-point
13802 // comparisons have to be converted.
13803 if (Subtarget->hasCMov() ||
13804 Cmp.getOpcode() != X86ISD::CMP ||
13805 !Cmp.getOperand(0).getValueType().isFloatingPoint() ||
13806 !Cmp.getOperand(1).getValueType().isFloatingPoint())
13809 // The instruction selector will select an FUCOM instruction instead of
13810 // FUCOMI, which writes the comparison result to FPSW instead of EFLAGS. Hence
13811 // build an SDNode sequence that transfers the result from FPSW into EFLAGS:
13812 // (X86sahf (trunc (srl (X86fp_stsw (trunc (X86cmp ...)), 8))))
13814 SDValue TruncFPSW = DAG.getNode(ISD::TRUNCATE, dl, MVT::i16, Cmp);
13815 SDValue FNStSW = DAG.getNode(X86ISD::FNSTSW16r, dl, MVT::i16, TruncFPSW);
13816 SDValue Srl = DAG.getNode(ISD::SRL, dl, MVT::i16, FNStSW,
13817 DAG.getConstant(8, dl, MVT::i8));
13818 SDValue TruncSrl = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Srl);
13819 return DAG.getNode(X86ISD::SAHF, dl, MVT::i32, TruncSrl);
13822 /// The minimum architected relative accuracy is 2^-12. We need one
13823 /// Newton-Raphson step to have a good float result (24 bits of precision).
13824 SDValue X86TargetLowering::getRsqrtEstimate(SDValue Op,
13825 DAGCombinerInfo &DCI,
13826 unsigned &RefinementSteps,
13827 bool &UseOneConstNR) const {
13828 EVT VT = Op.getValueType();
13829 const char *RecipOp;
13831 // SSE1 has rsqrtss and rsqrtps. AVX adds a 256-bit variant for rsqrtps.
13832 // TODO: Add support for AVX512 (v16f32).
13833 // It is likely not profitable to do this for f64 because a double-precision
13834 // rsqrt estimate with refinement on x86 prior to FMA requires at least 16
13835 // instructions: convert to single, rsqrtss, convert back to double, refine
13836 // (3 steps = at least 13 insts). If an 'rsqrtsd' variant was added to the ISA
13837 // along with FMA, this could be a throughput win.
13838 if (VT == MVT::f32 && Subtarget->hasSSE1())
13840 else if ((VT == MVT::v4f32 && Subtarget->hasSSE1()) ||
13841 (VT == MVT::v8f32 && Subtarget->hasAVX()))
13842 RecipOp = "vec-sqrtf";
13846 TargetRecip Recips = DCI.DAG.getTarget().Options.Reciprocals;
13847 if (!Recips.isEnabled(RecipOp))
13850 RefinementSteps = Recips.getRefinementSteps(RecipOp);
13851 UseOneConstNR = false;
13852 return DCI.DAG.getNode(X86ISD::FRSQRT, SDLoc(Op), VT, Op);
13855 /// The minimum architected relative accuracy is 2^-12. We need one
13856 /// Newton-Raphson step to have a good float result (24 bits of precision).
13857 SDValue X86TargetLowering::getRecipEstimate(SDValue Op,
13858 DAGCombinerInfo &DCI,
13859 unsigned &RefinementSteps) const {
13860 EVT VT = Op.getValueType();
13861 const char *RecipOp;
13863 // SSE1 has rcpss and rcpps. AVX adds a 256-bit variant for rcpps.
13864 // TODO: Add support for AVX512 (v16f32).
13865 // It is likely not profitable to do this for f64 because a double-precision
13866 // reciprocal estimate with refinement on x86 prior to FMA requires
13867 // 15 instructions: convert to single, rcpss, convert back to double, refine
13868 // (3 steps = 12 insts). If an 'rcpsd' variant was added to the ISA
13869 // along with FMA, this could be a throughput win.
13870 if (VT == MVT::f32 && Subtarget->hasSSE1())
13872 else if ((VT == MVT::v4f32 && Subtarget->hasSSE1()) ||
13873 (VT == MVT::v8f32 && Subtarget->hasAVX()))
13874 RecipOp = "vec-divf";
13878 TargetRecip Recips = DCI.DAG.getTarget().Options.Reciprocals;
13879 if (!Recips.isEnabled(RecipOp))
13882 RefinementSteps = Recips.getRefinementSteps(RecipOp);
13883 return DCI.DAG.getNode(X86ISD::FRCP, SDLoc(Op), VT, Op);
13886 /// If we have at least two divisions that use the same divisor, convert to
13887 /// multplication by a reciprocal. This may need to be adjusted for a given
13888 /// CPU if a division's cost is not at least twice the cost of a multiplication.
13889 /// This is because we still need one division to calculate the reciprocal and
13890 /// then we need two multiplies by that reciprocal as replacements for the
13891 /// original divisions.
13892 unsigned X86TargetLowering::combineRepeatedFPDivisors() const {
13896 static bool isAllOnes(SDValue V) {
13897 ConstantSDNode *C = dyn_cast<ConstantSDNode>(V);
13898 return C && C->isAllOnesValue();
13901 /// LowerToBT - Result of 'and' is compared against zero. Turn it into a BT node
13902 /// if it's possible.
13903 SDValue X86TargetLowering::LowerToBT(SDValue And, ISD::CondCode CC,
13904 SDLoc dl, SelectionDAG &DAG) const {
13905 SDValue Op0 = And.getOperand(0);
13906 SDValue Op1 = And.getOperand(1);
13907 if (Op0.getOpcode() == ISD::TRUNCATE)
13908 Op0 = Op0.getOperand(0);
13909 if (Op1.getOpcode() == ISD::TRUNCATE)
13910 Op1 = Op1.getOperand(0);
13913 if (Op1.getOpcode() == ISD::SHL)
13914 std::swap(Op0, Op1);
13915 if (Op0.getOpcode() == ISD::SHL) {
13916 if (ConstantSDNode *And00C = dyn_cast<ConstantSDNode>(Op0.getOperand(0)))
13917 if (And00C->getZExtValue() == 1) {
13918 // If we looked past a truncate, check that it's only truncating away
13920 unsigned BitWidth = Op0.getValueSizeInBits();
13921 unsigned AndBitWidth = And.getValueSizeInBits();
13922 if (BitWidth > AndBitWidth) {
13924 DAG.computeKnownBits(Op0, Zeros, Ones);
13925 if (Zeros.countLeadingOnes() < BitWidth - AndBitWidth)
13929 RHS = Op0.getOperand(1);
13931 } else if (Op1.getOpcode() == ISD::Constant) {
13932 ConstantSDNode *AndRHS = cast<ConstantSDNode>(Op1);
13933 uint64_t AndRHSVal = AndRHS->getZExtValue();
13934 SDValue AndLHS = Op0;
13936 if (AndRHSVal == 1 && AndLHS.getOpcode() == ISD::SRL) {
13937 LHS = AndLHS.getOperand(0);
13938 RHS = AndLHS.getOperand(1);
13941 // Use BT if the immediate can't be encoded in a TEST instruction.
13942 if (!isUInt<32>(AndRHSVal) && isPowerOf2_64(AndRHSVal)) {
13944 RHS = DAG.getConstant(Log2_64_Ceil(AndRHSVal), dl, LHS.getValueType());
13948 if (LHS.getNode()) {
13949 // If LHS is i8, promote it to i32 with any_extend. There is no i8 BT
13950 // instruction. Since the shift amount is in-range-or-undefined, we know
13951 // that doing a bittest on the i32 value is ok. We extend to i32 because
13952 // the encoding for the i16 version is larger than the i32 version.
13953 // Also promote i16 to i32 for performance / code size reason.
13954 if (LHS.getValueType() == MVT::i8 ||
13955 LHS.getValueType() == MVT::i16)
13956 LHS = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, LHS);
13958 // If the operand types disagree, extend the shift amount to match. Since
13959 // BT ignores high bits (like shifts) we can use anyextend.
13960 if (LHS.getValueType() != RHS.getValueType())
13961 RHS = DAG.getNode(ISD::ANY_EXTEND, dl, LHS.getValueType(), RHS);
13963 SDValue BT = DAG.getNode(X86ISD::BT, dl, MVT::i32, LHS, RHS);
13964 X86::CondCode Cond = CC == ISD::SETEQ ? X86::COND_AE : X86::COND_B;
13965 return DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
13966 DAG.getConstant(Cond, dl, MVT::i8), BT);
13972 /// \brief - Turns an ISD::CondCode into a value suitable for SSE floating point
13974 static int translateX86FSETCC(ISD::CondCode SetCCOpcode, SDValue &Op0,
13979 // SSE Condition code mapping:
13988 switch (SetCCOpcode) {
13989 default: llvm_unreachable("Unexpected SETCC condition");
13991 case ISD::SETEQ: SSECC = 0; break;
13993 case ISD::SETGT: Swap = true; // Fallthrough
13995 case ISD::SETOLT: SSECC = 1; break;
13997 case ISD::SETGE: Swap = true; // Fallthrough
13999 case ISD::SETOLE: SSECC = 2; break;
14000 case ISD::SETUO: SSECC = 3; break;
14002 case ISD::SETNE: SSECC = 4; break;
14003 case ISD::SETULE: Swap = true; // Fallthrough
14004 case ISD::SETUGE: SSECC = 5; break;
14005 case ISD::SETULT: Swap = true; // Fallthrough
14006 case ISD::SETUGT: SSECC = 6; break;
14007 case ISD::SETO: SSECC = 7; break;
14009 case ISD::SETONE: SSECC = 8; break;
14012 std::swap(Op0, Op1);
14017 // Lower256IntVSETCC - Break a VSETCC 256-bit integer VSETCC into two new 128
14018 // ones, and then concatenate the result back.
14019 static SDValue Lower256IntVSETCC(SDValue Op, SelectionDAG &DAG) {
14020 MVT VT = Op.getSimpleValueType();
14022 assert(VT.is256BitVector() && Op.getOpcode() == ISD::SETCC &&
14023 "Unsupported value type for operation");
14025 unsigned NumElems = VT.getVectorNumElements();
14027 SDValue CC = Op.getOperand(2);
14029 // Extract the LHS vectors
14030 SDValue LHS = Op.getOperand(0);
14031 SDValue LHS1 = Extract128BitVector(LHS, 0, DAG, dl);
14032 SDValue LHS2 = Extract128BitVector(LHS, NumElems/2, DAG, dl);
14034 // Extract the RHS vectors
14035 SDValue RHS = Op.getOperand(1);
14036 SDValue RHS1 = Extract128BitVector(RHS, 0, DAG, dl);
14037 SDValue RHS2 = Extract128BitVector(RHS, NumElems/2, DAG, dl);
14039 // Issue the operation on the smaller types and concatenate the result back
14040 MVT EltVT = VT.getVectorElementType();
14041 MVT NewVT = MVT::getVectorVT(EltVT, NumElems/2);
14042 return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT,
14043 DAG.getNode(Op.getOpcode(), dl, NewVT, LHS1, RHS1, CC),
14044 DAG.getNode(Op.getOpcode(), dl, NewVT, LHS2, RHS2, CC));
14047 static SDValue LowerBoolVSETCC_AVX512(SDValue Op, SelectionDAG &DAG) {
14048 SDValue Op0 = Op.getOperand(0);
14049 SDValue Op1 = Op.getOperand(1);
14050 SDValue CC = Op.getOperand(2);
14051 MVT VT = Op.getSimpleValueType();
14054 assert(Op0.getValueType().getVectorElementType() == MVT::i1 &&
14055 "Unexpected type for boolean compare operation");
14056 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
14057 SDValue NotOp0 = DAG.getNode(ISD::XOR, dl, VT, Op0,
14058 DAG.getConstant(-1, dl, VT));
14059 SDValue NotOp1 = DAG.getNode(ISD::XOR, dl, VT, Op1,
14060 DAG.getConstant(-1, dl, VT));
14061 switch (SetCCOpcode) {
14062 default: llvm_unreachable("Unexpected SETCC condition");
14064 // (x == y) -> ~(x ^ y)
14065 return DAG.getNode(ISD::XOR, dl, VT,
14066 DAG.getNode(ISD::XOR, dl, VT, Op0, Op1),
14067 DAG.getConstant(-1, dl, VT));
14069 // (x != y) -> (x ^ y)
14070 return DAG.getNode(ISD::XOR, dl, VT, Op0, Op1);
14073 // (x > y) -> (x & ~y)
14074 return DAG.getNode(ISD::AND, dl, VT, Op0, NotOp1);
14077 // (x < y) -> (~x & y)
14078 return DAG.getNode(ISD::AND, dl, VT, NotOp0, Op1);
14081 // (x <= y) -> (~x | y)
14082 return DAG.getNode(ISD::OR, dl, VT, NotOp0, Op1);
14085 // (x >=y) -> (x | ~y)
14086 return DAG.getNode(ISD::OR, dl, VT, Op0, NotOp1);
14090 static SDValue LowerIntVSETCC_AVX512(SDValue Op, SelectionDAG &DAG,
14091 const X86Subtarget *Subtarget) {
14092 SDValue Op0 = Op.getOperand(0);
14093 SDValue Op1 = Op.getOperand(1);
14094 SDValue CC = Op.getOperand(2);
14095 MVT VT = Op.getSimpleValueType();
14098 assert(Op0.getValueType().getVectorElementType().getSizeInBits() >= 8 &&
14099 Op.getValueType().getScalarType() == MVT::i1 &&
14100 "Cannot set masked compare for this operation");
14102 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
14104 bool Unsigned = false;
14107 switch (SetCCOpcode) {
14108 default: llvm_unreachable("Unexpected SETCC condition");
14109 case ISD::SETNE: SSECC = 4; break;
14110 case ISD::SETEQ: Opc = X86ISD::PCMPEQM; break;
14111 case ISD::SETUGT: SSECC = 6; Unsigned = true; break;
14112 case ISD::SETLT: Swap = true; //fall-through
14113 case ISD::SETGT: Opc = X86ISD::PCMPGTM; break;
14114 case ISD::SETULT: SSECC = 1; Unsigned = true; break;
14115 case ISD::SETUGE: SSECC = 5; Unsigned = true; break; //NLT
14116 case ISD::SETGE: Swap = true; SSECC = 2; break; // LE + swap
14117 case ISD::SETULE: Unsigned = true; //fall-through
14118 case ISD::SETLE: SSECC = 2; break;
14122 std::swap(Op0, Op1);
14124 return DAG.getNode(Opc, dl, VT, Op0, Op1);
14125 Opc = Unsigned ? X86ISD::CMPMU: X86ISD::CMPM;
14126 return DAG.getNode(Opc, dl, VT, Op0, Op1,
14127 DAG.getConstant(SSECC, dl, MVT::i8));
14130 /// \brief Try to turn a VSETULT into a VSETULE by modifying its second
14131 /// operand \p Op1. If non-trivial (for example because it's not constant)
14132 /// return an empty value.
14133 static SDValue ChangeVSETULTtoVSETULE(SDLoc dl, SDValue Op1, SelectionDAG &DAG)
14135 BuildVectorSDNode *BV = dyn_cast<BuildVectorSDNode>(Op1.getNode());
14139 MVT VT = Op1.getSimpleValueType();
14140 MVT EVT = VT.getVectorElementType();
14141 unsigned n = VT.getVectorNumElements();
14142 SmallVector<SDValue, 8> ULTOp1;
14144 for (unsigned i = 0; i < n; ++i) {
14145 ConstantSDNode *Elt = dyn_cast<ConstantSDNode>(BV->getOperand(i));
14146 if (!Elt || Elt->isOpaque() || Elt->getValueType(0) != EVT)
14149 // Avoid underflow.
14150 APInt Val = Elt->getAPIntValue();
14154 ULTOp1.push_back(DAG.getConstant(Val - 1, dl, EVT));
14157 return DAG.getNode(ISD::BUILD_VECTOR, dl, VT, ULTOp1);
14160 static SDValue LowerVSETCC(SDValue Op, const X86Subtarget *Subtarget,
14161 SelectionDAG &DAG) {
14162 SDValue Op0 = Op.getOperand(0);
14163 SDValue Op1 = Op.getOperand(1);
14164 SDValue CC = Op.getOperand(2);
14165 MVT VT = Op.getSimpleValueType();
14166 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
14167 bool isFP = Op.getOperand(1).getSimpleValueType().isFloatingPoint();
14172 MVT EltVT = Op0.getSimpleValueType().getVectorElementType();
14173 assert(EltVT == MVT::f32 || EltVT == MVT::f64);
14176 unsigned SSECC = translateX86FSETCC(SetCCOpcode, Op0, Op1);
14177 unsigned Opc = X86ISD::CMPP;
14178 if (Subtarget->hasAVX512() && VT.getVectorElementType() == MVT::i1) {
14179 assert(VT.getVectorNumElements() <= 16);
14180 Opc = X86ISD::CMPM;
14182 // In the two special cases we can't handle, emit two comparisons.
14185 unsigned CombineOpc;
14186 if (SetCCOpcode == ISD::SETUEQ) {
14187 CC0 = 3; CC1 = 0; CombineOpc = ISD::OR;
14189 assert(SetCCOpcode == ISD::SETONE);
14190 CC0 = 7; CC1 = 4; CombineOpc = ISD::AND;
14193 SDValue Cmp0 = DAG.getNode(Opc, dl, VT, Op0, Op1,
14194 DAG.getConstant(CC0, dl, MVT::i8));
14195 SDValue Cmp1 = DAG.getNode(Opc, dl, VT, Op0, Op1,
14196 DAG.getConstant(CC1, dl, MVT::i8));
14197 return DAG.getNode(CombineOpc, dl, VT, Cmp0, Cmp1);
14199 // Handle all other FP comparisons here.
14200 return DAG.getNode(Opc, dl, VT, Op0, Op1,
14201 DAG.getConstant(SSECC, dl, MVT::i8));
14204 MVT VTOp0 = Op0.getSimpleValueType();
14205 assert(VTOp0 == Op1.getSimpleValueType() &&
14206 "Expected operands with same type!");
14207 assert(VT.getVectorNumElements() == VTOp0.getVectorNumElements() &&
14208 "Invalid number of packed elements for source and destination!");
14210 if (VT.is128BitVector() && VTOp0.is256BitVector()) {
14211 // On non-AVX512 targets, a vector of MVT::i1 is promoted by the type
14212 // legalizer to a wider vector type. In the case of 'vsetcc' nodes, the
14213 // legalizer firstly checks if the first operand in input to the setcc has
14214 // a legal type. If so, then it promotes the return type to that same type.
14215 // Otherwise, the return type is promoted to the 'next legal type' which,
14216 // for a vector of MVT::i1 is always a 128-bit integer vector type.
14218 // We reach this code only if the following two conditions are met:
14219 // 1. Both return type and operand type have been promoted to wider types
14220 // by the type legalizer.
14221 // 2. The original operand type has been promoted to a 256-bit vector.
14223 // Note that condition 2. only applies for AVX targets.
14224 SDValue NewOp = DAG.getSetCC(dl, VTOp0, Op0, Op1, SetCCOpcode);
14225 return DAG.getZExtOrTrunc(NewOp, dl, VT);
14228 // The non-AVX512 code below works under the assumption that source and
14229 // destination types are the same.
14230 assert((Subtarget->hasAVX512() || (VT == VTOp0)) &&
14231 "Value types for source and destination must be the same!");
14233 // Break 256-bit integer vector compare into smaller ones.
14234 if (VT.is256BitVector() && !Subtarget->hasInt256())
14235 return Lower256IntVSETCC(Op, DAG);
14237 EVT OpVT = Op1.getValueType();
14238 if (OpVT.getVectorElementType() == MVT::i1)
14239 return LowerBoolVSETCC_AVX512(Op, DAG);
14241 bool MaskResult = (VT.getVectorElementType() == MVT::i1);
14242 if (Subtarget->hasAVX512()) {
14243 if (Op1.getValueType().is512BitVector() ||
14244 (Subtarget->hasBWI() && Subtarget->hasVLX()) ||
14245 (MaskResult && OpVT.getVectorElementType().getSizeInBits() >= 32))
14246 return LowerIntVSETCC_AVX512(Op, DAG, Subtarget);
14248 // In AVX-512 architecture setcc returns mask with i1 elements,
14249 // But there is no compare instruction for i8 and i16 elements in KNL.
14250 // We are not talking about 512-bit operands in this case, these
14251 // types are illegal.
14253 (OpVT.getVectorElementType().getSizeInBits() < 32 &&
14254 OpVT.getVectorElementType().getSizeInBits() >= 8))
14255 return DAG.getNode(ISD::TRUNCATE, dl, VT,
14256 DAG.getNode(ISD::SETCC, dl, OpVT, Op0, Op1, CC));
14259 // Lower using XOP integer comparisons.
14260 if ((VT == MVT::v16i8 || VT == MVT::v8i16 ||
14261 VT == MVT::v4i32 || VT == MVT::v2i64) && Subtarget->hasXOP()) {
14262 // Translate compare code to XOP PCOM compare mode.
14263 unsigned CmpMode = 0;
14264 switch (SetCCOpcode) {
14265 default: llvm_unreachable("Unexpected SETCC condition");
14267 case ISD::SETLT: CmpMode = 0x00; break;
14269 case ISD::SETLE: CmpMode = 0x01; break;
14271 case ISD::SETGT: CmpMode = 0x02; break;
14273 case ISD::SETGE: CmpMode = 0x03; break;
14274 case ISD::SETEQ: CmpMode = 0x04; break;
14275 case ISD::SETNE: CmpMode = 0x05; break;
14278 // Are we comparing unsigned or signed integers?
14279 unsigned Opc = ISD::isUnsignedIntSetCC(SetCCOpcode)
14280 ? X86ISD::VPCOMU : X86ISD::VPCOM;
14282 return DAG.getNode(Opc, dl, VT, Op0, Op1,
14283 DAG.getConstant(CmpMode, dl, MVT::i8));
14286 // We are handling one of the integer comparisons here. Since SSE only has
14287 // GT and EQ comparisons for integer, swapping operands and multiple
14288 // operations may be required for some comparisons.
14290 bool Swap = false, Invert = false, FlipSigns = false, MinMax = false;
14291 bool Subus = false;
14293 switch (SetCCOpcode) {
14294 default: llvm_unreachable("Unexpected SETCC condition");
14295 case ISD::SETNE: Invert = true;
14296 case ISD::SETEQ: Opc = X86ISD::PCMPEQ; break;
14297 case ISD::SETLT: Swap = true;
14298 case ISD::SETGT: Opc = X86ISD::PCMPGT; break;
14299 case ISD::SETGE: Swap = true;
14300 case ISD::SETLE: Opc = X86ISD::PCMPGT;
14301 Invert = true; break;
14302 case ISD::SETULT: Swap = true;
14303 case ISD::SETUGT: Opc = X86ISD::PCMPGT;
14304 FlipSigns = true; break;
14305 case ISD::SETUGE: Swap = true;
14306 case ISD::SETULE: Opc = X86ISD::PCMPGT;
14307 FlipSigns = true; Invert = true; break;
14310 // Special case: Use min/max operations for SETULE/SETUGE
14311 MVT VET = VT.getVectorElementType();
14313 (Subtarget->hasSSE41() && (VET >= MVT::i8 && VET <= MVT::i32))
14314 || (Subtarget->hasSSE2() && (VET == MVT::i8));
14317 switch (SetCCOpcode) {
14319 case ISD::SETULE: Opc = ISD::UMIN; MinMax = true; break;
14320 case ISD::SETUGE: Opc = ISD::UMAX; MinMax = true; break;
14323 if (MinMax) { Swap = false; Invert = false; FlipSigns = false; }
14326 bool hasSubus = Subtarget->hasSSE2() && (VET == MVT::i8 || VET == MVT::i16);
14327 if (!MinMax && hasSubus) {
14328 // As another special case, use PSUBUS[BW] when it's profitable. E.g. for
14330 // t = psubus Op0, Op1
14331 // pcmpeq t, <0..0>
14332 switch (SetCCOpcode) {
14334 case ISD::SETULT: {
14335 // If the comparison is against a constant we can turn this into a
14336 // setule. With psubus, setule does not require a swap. This is
14337 // beneficial because the constant in the register is no longer
14338 // destructed as the destination so it can be hoisted out of a loop.
14339 // Only do this pre-AVX since vpcmp* is no longer destructive.
14340 if (Subtarget->hasAVX())
14342 SDValue ULEOp1 = ChangeVSETULTtoVSETULE(dl, Op1, DAG);
14343 if (ULEOp1.getNode()) {
14345 Subus = true; Invert = false; Swap = false;
14349 // Psubus is better than flip-sign because it requires no inversion.
14350 case ISD::SETUGE: Subus = true; Invert = false; Swap = true; break;
14351 case ISD::SETULE: Subus = true; Invert = false; Swap = false; break;
14355 Opc = X86ISD::SUBUS;
14361 std::swap(Op0, Op1);
14363 // Check that the operation in question is available (most are plain SSE2,
14364 // but PCMPGTQ and PCMPEQQ have different requirements).
14365 if (VT == MVT::v2i64) {
14366 if (Opc == X86ISD::PCMPGT && !Subtarget->hasSSE42()) {
14367 assert(Subtarget->hasSSE2() && "Don't know how to lower!");
14369 // First cast everything to the right type.
14370 Op0 = DAG.getBitcast(MVT::v4i32, Op0);
14371 Op1 = DAG.getBitcast(MVT::v4i32, Op1);
14373 // Since SSE has no unsigned integer comparisons, we need to flip the sign
14374 // bits of the inputs before performing those operations. The lower
14375 // compare is always unsigned.
14378 SB = DAG.getConstant(0x80000000U, dl, MVT::v4i32);
14380 SDValue Sign = DAG.getConstant(0x80000000U, dl, MVT::i32);
14381 SDValue Zero = DAG.getConstant(0x00000000U, dl, MVT::i32);
14382 SB = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32,
14383 Sign, Zero, Sign, Zero);
14385 Op0 = DAG.getNode(ISD::XOR, dl, MVT::v4i32, Op0, SB);
14386 Op1 = DAG.getNode(ISD::XOR, dl, MVT::v4i32, Op1, SB);
14388 // Emulate PCMPGTQ with (hi1 > hi2) | ((hi1 == hi2) & (lo1 > lo2))
14389 SDValue GT = DAG.getNode(X86ISD::PCMPGT, dl, MVT::v4i32, Op0, Op1);
14390 SDValue EQ = DAG.getNode(X86ISD::PCMPEQ, dl, MVT::v4i32, Op0, Op1);
14392 // Create masks for only the low parts/high parts of the 64 bit integers.
14393 static const int MaskHi[] = { 1, 1, 3, 3 };
14394 static const int MaskLo[] = { 0, 0, 2, 2 };
14395 SDValue EQHi = DAG.getVectorShuffle(MVT::v4i32, dl, EQ, EQ, MaskHi);
14396 SDValue GTLo = DAG.getVectorShuffle(MVT::v4i32, dl, GT, GT, MaskLo);
14397 SDValue GTHi = DAG.getVectorShuffle(MVT::v4i32, dl, GT, GT, MaskHi);
14399 SDValue Result = DAG.getNode(ISD::AND, dl, MVT::v4i32, EQHi, GTLo);
14400 Result = DAG.getNode(ISD::OR, dl, MVT::v4i32, Result, GTHi);
14403 Result = DAG.getNOT(dl, Result, MVT::v4i32);
14405 return DAG.getBitcast(VT, Result);
14408 if (Opc == X86ISD::PCMPEQ && !Subtarget->hasSSE41()) {
14409 // If pcmpeqq is missing but pcmpeqd is available synthesize pcmpeqq with
14410 // pcmpeqd + pshufd + pand.
14411 assert(Subtarget->hasSSE2() && !FlipSigns && "Don't know how to lower!");
14413 // First cast everything to the right type.
14414 Op0 = DAG.getBitcast(MVT::v4i32, Op0);
14415 Op1 = DAG.getBitcast(MVT::v4i32, Op1);
14418 SDValue Result = DAG.getNode(Opc, dl, MVT::v4i32, Op0, Op1);
14420 // Make sure the lower and upper halves are both all-ones.
14421 static const int Mask[] = { 1, 0, 3, 2 };
14422 SDValue Shuf = DAG.getVectorShuffle(MVT::v4i32, dl, Result, Result, Mask);
14423 Result = DAG.getNode(ISD::AND, dl, MVT::v4i32, Result, Shuf);
14426 Result = DAG.getNOT(dl, Result, MVT::v4i32);
14428 return DAG.getBitcast(VT, Result);
14432 // Since SSE has no unsigned integer comparisons, we need to flip the sign
14433 // bits of the inputs before performing those operations.
14435 EVT EltVT = VT.getVectorElementType();
14436 SDValue SB = DAG.getConstant(APInt::getSignBit(EltVT.getSizeInBits()), dl,
14438 Op0 = DAG.getNode(ISD::XOR, dl, VT, Op0, SB);
14439 Op1 = DAG.getNode(ISD::XOR, dl, VT, Op1, SB);
14442 SDValue Result = DAG.getNode(Opc, dl, VT, Op0, Op1);
14444 // If the logical-not of the result is required, perform that now.
14446 Result = DAG.getNOT(dl, Result, VT);
14449 Result = DAG.getNode(X86ISD::PCMPEQ, dl, VT, Op0, Result);
14452 Result = DAG.getNode(X86ISD::PCMPEQ, dl, VT, Result,
14453 getZeroVector(VT, Subtarget, DAG, dl));
14458 SDValue X86TargetLowering::LowerSETCC(SDValue Op, SelectionDAG &DAG) const {
14460 MVT VT = Op.getSimpleValueType();
14462 if (VT.isVector()) return LowerVSETCC(Op, Subtarget, DAG);
14464 assert(((!Subtarget->hasAVX512() && VT == MVT::i8) || (VT == MVT::i1))
14465 && "SetCC type must be 8-bit or 1-bit integer");
14466 SDValue Op0 = Op.getOperand(0);
14467 SDValue Op1 = Op.getOperand(1);
14469 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
14471 // Optimize to BT if possible.
14472 // Lower (X & (1 << N)) == 0 to BT(X, N).
14473 // Lower ((X >>u N) & 1) != 0 to BT(X, N).
14474 // Lower ((X >>s N) & 1) != 0 to BT(X, N).
14475 if (Op0.getOpcode() == ISD::AND && Op0.hasOneUse() &&
14476 Op1.getOpcode() == ISD::Constant &&
14477 cast<ConstantSDNode>(Op1)->isNullValue() &&
14478 (CC == ISD::SETEQ || CC == ISD::SETNE)) {
14479 SDValue NewSetCC = LowerToBT(Op0, CC, dl, DAG);
14480 if (NewSetCC.getNode()) {
14482 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i1, NewSetCC);
14487 // Look for X == 0, X == 1, X != 0, or X != 1. We can simplify some forms of
14489 if (Op1.getOpcode() == ISD::Constant &&
14490 (cast<ConstantSDNode>(Op1)->getZExtValue() == 1 ||
14491 cast<ConstantSDNode>(Op1)->isNullValue()) &&
14492 (CC == ISD::SETEQ || CC == ISD::SETNE)) {
14494 // If the input is a setcc, then reuse the input setcc or use a new one with
14495 // the inverted condition.
14496 if (Op0.getOpcode() == X86ISD::SETCC) {
14497 X86::CondCode CCode = (X86::CondCode)Op0.getConstantOperandVal(0);
14498 bool Invert = (CC == ISD::SETNE) ^
14499 cast<ConstantSDNode>(Op1)->isNullValue();
14503 CCode = X86::GetOppositeBranchCondition(CCode);
14504 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
14505 DAG.getConstant(CCode, dl, MVT::i8),
14506 Op0.getOperand(1));
14508 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i1, SetCC);
14512 if ((Op0.getValueType() == MVT::i1) && (Op1.getOpcode() == ISD::Constant) &&
14513 (cast<ConstantSDNode>(Op1)->getZExtValue() == 1) &&
14514 (CC == ISD::SETEQ || CC == ISD::SETNE)) {
14516 ISD::CondCode NewCC = ISD::getSetCCInverse(CC, true);
14517 return DAG.getSetCC(dl, VT, Op0, DAG.getConstant(0, dl, MVT::i1), NewCC);
14520 bool isFP = Op1.getSimpleValueType().isFloatingPoint();
14521 unsigned X86CC = TranslateX86CC(CC, dl, isFP, Op0, Op1, DAG);
14522 if (X86CC == X86::COND_INVALID)
14525 SDValue EFLAGS = EmitCmp(Op0, Op1, X86CC, dl, DAG);
14526 EFLAGS = ConvertCmpIfNecessary(EFLAGS, DAG);
14527 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
14528 DAG.getConstant(X86CC, dl, MVT::i8), EFLAGS);
14530 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i1, SetCC);
14534 // isX86LogicalCmp - Return true if opcode is a X86 logical comparison.
14535 static bool isX86LogicalCmp(SDValue Op) {
14536 unsigned Opc = Op.getNode()->getOpcode();
14537 if (Opc == X86ISD::CMP || Opc == X86ISD::COMI || Opc == X86ISD::UCOMI ||
14538 Opc == X86ISD::SAHF)
14540 if (Op.getResNo() == 1 &&
14541 (Opc == X86ISD::ADD ||
14542 Opc == X86ISD::SUB ||
14543 Opc == X86ISD::ADC ||
14544 Opc == X86ISD::SBB ||
14545 Opc == X86ISD::SMUL ||
14546 Opc == X86ISD::UMUL ||
14547 Opc == X86ISD::INC ||
14548 Opc == X86ISD::DEC ||
14549 Opc == X86ISD::OR ||
14550 Opc == X86ISD::XOR ||
14551 Opc == X86ISD::AND))
14554 if (Op.getResNo() == 2 && Opc == X86ISD::UMUL)
14560 static bool isTruncWithZeroHighBitsInput(SDValue V, SelectionDAG &DAG) {
14561 if (V.getOpcode() != ISD::TRUNCATE)
14564 SDValue VOp0 = V.getOperand(0);
14565 unsigned InBits = VOp0.getValueSizeInBits();
14566 unsigned Bits = V.getValueSizeInBits();
14567 return DAG.MaskedValueIsZero(VOp0, APInt::getHighBitsSet(InBits,InBits-Bits));
14570 SDValue X86TargetLowering::LowerSELECT(SDValue Op, SelectionDAG &DAG) const {
14571 bool addTest = true;
14572 SDValue Cond = Op.getOperand(0);
14573 SDValue Op1 = Op.getOperand(1);
14574 SDValue Op2 = Op.getOperand(2);
14576 EVT VT = Op1.getValueType();
14579 // Lower FP selects into a CMP/AND/ANDN/OR sequence when the necessary SSE ops
14580 // are available or VBLENDV if AVX is available.
14581 // Otherwise FP cmovs get lowered into a less efficient branch sequence later.
14582 if (Cond.getOpcode() == ISD::SETCC &&
14583 ((Subtarget->hasSSE2() && (VT == MVT::f32 || VT == MVT::f64)) ||
14584 (Subtarget->hasSSE1() && VT == MVT::f32)) &&
14585 VT == Cond.getOperand(0).getValueType() && Cond->hasOneUse()) {
14586 SDValue CondOp0 = Cond.getOperand(0), CondOp1 = Cond.getOperand(1);
14587 int SSECC = translateX86FSETCC(
14588 cast<CondCodeSDNode>(Cond.getOperand(2))->get(), CondOp0, CondOp1);
14591 if (Subtarget->hasAVX512()) {
14592 SDValue Cmp = DAG.getNode(X86ISD::FSETCC, DL, MVT::i1, CondOp0, CondOp1,
14593 DAG.getConstant(SSECC, DL, MVT::i8));
14594 return DAG.getNode(X86ISD::SELECT, DL, VT, Cmp, Op1, Op2);
14597 SDValue Cmp = DAG.getNode(X86ISD::FSETCC, DL, VT, CondOp0, CondOp1,
14598 DAG.getConstant(SSECC, DL, MVT::i8));
14600 // If we have AVX, we can use a variable vector select (VBLENDV) instead
14601 // of 3 logic instructions for size savings and potentially speed.
14602 // Unfortunately, there is no scalar form of VBLENDV.
14604 // If either operand is a constant, don't try this. We can expect to
14605 // optimize away at least one of the logic instructions later in that
14606 // case, so that sequence would be faster than a variable blend.
14608 // BLENDV was introduced with SSE 4.1, but the 2 register form implicitly
14609 // uses XMM0 as the selection register. That may need just as many
14610 // instructions as the AND/ANDN/OR sequence due to register moves, so
14613 if (Subtarget->hasAVX() &&
14614 !isa<ConstantFPSDNode>(Op1) && !isa<ConstantFPSDNode>(Op2)) {
14616 // Convert to vectors, do a VSELECT, and convert back to scalar.
14617 // All of the conversions should be optimized away.
14619 EVT VecVT = VT == MVT::f32 ? MVT::v4f32 : MVT::v2f64;
14620 SDValue VOp1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, VecVT, Op1);
14621 SDValue VOp2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, VecVT, Op2);
14622 SDValue VCmp = DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, VecVT, Cmp);
14624 EVT VCmpVT = VT == MVT::f32 ? MVT::v4i32 : MVT::v2i64;
14625 VCmp = DAG.getBitcast(VCmpVT, VCmp);
14627 SDValue VSel = DAG.getNode(ISD::VSELECT, DL, VecVT, VCmp, VOp1, VOp2);
14629 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, VT,
14630 VSel, DAG.getIntPtrConstant(0, DL));
14632 SDValue AndN = DAG.getNode(X86ISD::FANDN, DL, VT, Cmp, Op2);
14633 SDValue And = DAG.getNode(X86ISD::FAND, DL, VT, Cmp, Op1);
14634 return DAG.getNode(X86ISD::FOR, DL, VT, AndN, And);
14638 if (VT.isVector() && VT.getScalarType() == MVT::i1) {
14640 if (ISD::isBuildVectorOfConstantSDNodes(Op1.getNode()))
14641 Op1Scalar = ConvertI1VectorToInteger(Op1, DAG);
14642 else if (Op1.getOpcode() == ISD::BITCAST && Op1.getOperand(0))
14643 Op1Scalar = Op1.getOperand(0);
14645 if (ISD::isBuildVectorOfConstantSDNodes(Op2.getNode()))
14646 Op2Scalar = ConvertI1VectorToInteger(Op2, DAG);
14647 else if (Op2.getOpcode() == ISD::BITCAST && Op2.getOperand(0))
14648 Op2Scalar = Op2.getOperand(0);
14649 if (Op1Scalar.getNode() && Op2Scalar.getNode()) {
14650 SDValue newSelect = DAG.getNode(ISD::SELECT, DL,
14651 Op1Scalar.getValueType(),
14652 Cond, Op1Scalar, Op2Scalar);
14653 if (newSelect.getValueSizeInBits() == VT.getSizeInBits())
14654 return DAG.getBitcast(VT, newSelect);
14655 SDValue ExtVec = DAG.getBitcast(MVT::v8i1, newSelect);
14656 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, VT, ExtVec,
14657 DAG.getIntPtrConstant(0, DL));
14661 if (VT == MVT::v4i1 || VT == MVT::v2i1) {
14662 SDValue zeroConst = DAG.getIntPtrConstant(0, DL);
14663 Op1 = DAG.getNode(ISD::INSERT_SUBVECTOR, DL, MVT::v8i1,
14664 DAG.getUNDEF(MVT::v8i1), Op1, zeroConst);
14665 Op2 = DAG.getNode(ISD::INSERT_SUBVECTOR, DL, MVT::v8i1,
14666 DAG.getUNDEF(MVT::v8i1), Op2, zeroConst);
14667 SDValue newSelect = DAG.getNode(ISD::SELECT, DL, MVT::v8i1,
14669 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, VT, newSelect, zeroConst);
14672 if (Cond.getOpcode() == ISD::SETCC) {
14673 SDValue NewCond = LowerSETCC(Cond, DAG);
14674 if (NewCond.getNode())
14678 // (select (x == 0), -1, y) -> (sign_bit (x - 1)) | y
14679 // (select (x == 0), y, -1) -> ~(sign_bit (x - 1)) | y
14680 // (select (x != 0), y, -1) -> (sign_bit (x - 1)) | y
14681 // (select (x != 0), -1, y) -> ~(sign_bit (x - 1)) | y
14682 if (Cond.getOpcode() == X86ISD::SETCC &&
14683 Cond.getOperand(1).getOpcode() == X86ISD::CMP &&
14684 isZero(Cond.getOperand(1).getOperand(1))) {
14685 SDValue Cmp = Cond.getOperand(1);
14687 unsigned CondCode =cast<ConstantSDNode>(Cond.getOperand(0))->getZExtValue();
14689 if ((isAllOnes(Op1) || isAllOnes(Op2)) &&
14690 (CondCode == X86::COND_E || CondCode == X86::COND_NE)) {
14691 SDValue Y = isAllOnes(Op2) ? Op1 : Op2;
14693 SDValue CmpOp0 = Cmp.getOperand(0);
14694 // Apply further optimizations for special cases
14695 // (select (x != 0), -1, 0) -> neg & sbb
14696 // (select (x == 0), 0, -1) -> neg & sbb
14697 if (ConstantSDNode *YC = dyn_cast<ConstantSDNode>(Y))
14698 if (YC->isNullValue() &&
14699 (isAllOnes(Op1) == (CondCode == X86::COND_NE))) {
14700 SDVTList VTs = DAG.getVTList(CmpOp0.getValueType(), MVT::i32);
14701 SDValue Neg = DAG.getNode(X86ISD::SUB, DL, VTs,
14702 DAG.getConstant(0, DL,
14703 CmpOp0.getValueType()),
14705 SDValue Res = DAG.getNode(X86ISD::SETCC_CARRY, DL, Op.getValueType(),
14706 DAG.getConstant(X86::COND_B, DL, MVT::i8),
14707 SDValue(Neg.getNode(), 1));
14711 Cmp = DAG.getNode(X86ISD::CMP, DL, MVT::i32,
14712 CmpOp0, DAG.getConstant(1, DL, CmpOp0.getValueType()));
14713 Cmp = ConvertCmpIfNecessary(Cmp, DAG);
14715 SDValue Res = // Res = 0 or -1.
14716 DAG.getNode(X86ISD::SETCC_CARRY, DL, Op.getValueType(),
14717 DAG.getConstant(X86::COND_B, DL, MVT::i8), Cmp);
14719 if (isAllOnes(Op1) != (CondCode == X86::COND_E))
14720 Res = DAG.getNOT(DL, Res, Res.getValueType());
14722 ConstantSDNode *N2C = dyn_cast<ConstantSDNode>(Op2);
14723 if (!N2C || !N2C->isNullValue())
14724 Res = DAG.getNode(ISD::OR, DL, Res.getValueType(), Res, Y);
14729 // Look past (and (setcc_carry (cmp ...)), 1).
14730 if (Cond.getOpcode() == ISD::AND &&
14731 Cond.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY) {
14732 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Cond.getOperand(1));
14733 if (C && C->getAPIntValue() == 1)
14734 Cond = Cond.getOperand(0);
14737 // If condition flag is set by a X86ISD::CMP, then use it as the condition
14738 // setting operand in place of the X86ISD::SETCC.
14739 unsigned CondOpcode = Cond.getOpcode();
14740 if (CondOpcode == X86ISD::SETCC ||
14741 CondOpcode == X86ISD::SETCC_CARRY) {
14742 CC = Cond.getOperand(0);
14744 SDValue Cmp = Cond.getOperand(1);
14745 unsigned Opc = Cmp.getOpcode();
14746 MVT VT = Op.getSimpleValueType();
14748 bool IllegalFPCMov = false;
14749 if (VT.isFloatingPoint() && !VT.isVector() &&
14750 !isScalarFPTypeInSSEReg(VT)) // FPStack?
14751 IllegalFPCMov = !hasFPCMov(cast<ConstantSDNode>(CC)->getSExtValue());
14753 if ((isX86LogicalCmp(Cmp) && !IllegalFPCMov) ||
14754 Opc == X86ISD::BT) { // FIXME
14758 } else if (CondOpcode == ISD::USUBO || CondOpcode == ISD::SSUBO ||
14759 CondOpcode == ISD::UADDO || CondOpcode == ISD::SADDO ||
14760 ((CondOpcode == ISD::UMULO || CondOpcode == ISD::SMULO) &&
14761 Cond.getOperand(0).getValueType() != MVT::i8)) {
14762 SDValue LHS = Cond.getOperand(0);
14763 SDValue RHS = Cond.getOperand(1);
14764 unsigned X86Opcode;
14767 switch (CondOpcode) {
14768 case ISD::UADDO: X86Opcode = X86ISD::ADD; X86Cond = X86::COND_B; break;
14769 case ISD::SADDO: X86Opcode = X86ISD::ADD; X86Cond = X86::COND_O; break;
14770 case ISD::USUBO: X86Opcode = X86ISD::SUB; X86Cond = X86::COND_B; break;
14771 case ISD::SSUBO: X86Opcode = X86ISD::SUB; X86Cond = X86::COND_O; break;
14772 case ISD::UMULO: X86Opcode = X86ISD::UMUL; X86Cond = X86::COND_O; break;
14773 case ISD::SMULO: X86Opcode = X86ISD::SMUL; X86Cond = X86::COND_O; break;
14774 default: llvm_unreachable("unexpected overflowing operator");
14776 if (CondOpcode == ISD::UMULO)
14777 VTs = DAG.getVTList(LHS.getValueType(), LHS.getValueType(),
14780 VTs = DAG.getVTList(LHS.getValueType(), MVT::i32);
14782 SDValue X86Op = DAG.getNode(X86Opcode, DL, VTs, LHS, RHS);
14784 if (CondOpcode == ISD::UMULO)
14785 Cond = X86Op.getValue(2);
14787 Cond = X86Op.getValue(1);
14789 CC = DAG.getConstant(X86Cond, DL, MVT::i8);
14794 // Look past the truncate if the high bits are known zero.
14795 if (isTruncWithZeroHighBitsInput(Cond, DAG))
14796 Cond = Cond.getOperand(0);
14798 // We know the result of AND is compared against zero. Try to match
14800 if (Cond.getOpcode() == ISD::AND && Cond.hasOneUse()) {
14801 SDValue NewSetCC = LowerToBT(Cond, ISD::SETNE, DL, DAG);
14802 if (NewSetCC.getNode()) {
14803 CC = NewSetCC.getOperand(0);
14804 Cond = NewSetCC.getOperand(1);
14811 CC = DAG.getConstant(X86::COND_NE, DL, MVT::i8);
14812 Cond = EmitTest(Cond, X86::COND_NE, DL, DAG);
14815 // a < b ? -1 : 0 -> RES = ~setcc_carry
14816 // a < b ? 0 : -1 -> RES = setcc_carry
14817 // a >= b ? -1 : 0 -> RES = setcc_carry
14818 // a >= b ? 0 : -1 -> RES = ~setcc_carry
14819 if (Cond.getOpcode() == X86ISD::SUB) {
14820 Cond = ConvertCmpIfNecessary(Cond, DAG);
14821 unsigned CondCode = cast<ConstantSDNode>(CC)->getZExtValue();
14823 if ((CondCode == X86::COND_AE || CondCode == X86::COND_B) &&
14824 (isAllOnes(Op1) || isAllOnes(Op2)) && (isZero(Op1) || isZero(Op2))) {
14825 SDValue Res = DAG.getNode(X86ISD::SETCC_CARRY, DL, Op.getValueType(),
14826 DAG.getConstant(X86::COND_B, DL, MVT::i8),
14828 if (isAllOnes(Op1) != (CondCode == X86::COND_B))
14829 return DAG.getNOT(DL, Res, Res.getValueType());
14834 // X86 doesn't have an i8 cmov. If both operands are the result of a truncate
14835 // widen the cmov and push the truncate through. This avoids introducing a new
14836 // branch during isel and doesn't add any extensions.
14837 if (Op.getValueType() == MVT::i8 &&
14838 Op1.getOpcode() == ISD::TRUNCATE && Op2.getOpcode() == ISD::TRUNCATE) {
14839 SDValue T1 = Op1.getOperand(0), T2 = Op2.getOperand(0);
14840 if (T1.getValueType() == T2.getValueType() &&
14841 // Blacklist CopyFromReg to avoid partial register stalls.
14842 T1.getOpcode() != ISD::CopyFromReg && T2.getOpcode()!=ISD::CopyFromReg){
14843 SDVTList VTs = DAG.getVTList(T1.getValueType(), MVT::Glue);
14844 SDValue Cmov = DAG.getNode(X86ISD::CMOV, DL, VTs, T2, T1, CC, Cond);
14845 return DAG.getNode(ISD::TRUNCATE, DL, Op.getValueType(), Cmov);
14849 // X86ISD::CMOV means set the result (which is operand 1) to the RHS if
14850 // condition is true.
14851 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::Glue);
14852 SDValue Ops[] = { Op2, Op1, CC, Cond };
14853 return DAG.getNode(X86ISD::CMOV, DL, VTs, Ops);
14856 static SDValue LowerSIGN_EXTEND_AVX512(SDValue Op,
14857 const X86Subtarget *Subtarget,
14858 SelectionDAG &DAG) {
14859 MVT VT = Op->getSimpleValueType(0);
14860 SDValue In = Op->getOperand(0);
14861 MVT InVT = In.getSimpleValueType();
14862 MVT VTElt = VT.getVectorElementType();
14863 MVT InVTElt = InVT.getVectorElementType();
14867 if ((InVTElt == MVT::i1) &&
14868 (((Subtarget->hasBWI() && Subtarget->hasVLX() &&
14869 VT.getSizeInBits() <= 256 && VTElt.getSizeInBits() <= 16)) ||
14871 ((Subtarget->hasBWI() && VT.is512BitVector() &&
14872 VTElt.getSizeInBits() <= 16)) ||
14874 ((Subtarget->hasDQI() && Subtarget->hasVLX() &&
14875 VT.getSizeInBits() <= 256 && VTElt.getSizeInBits() >= 32)) ||
14877 ((Subtarget->hasDQI() && VT.is512BitVector() &&
14878 VTElt.getSizeInBits() >= 32))))
14879 return DAG.getNode(X86ISD::VSEXT, dl, VT, In);
14881 unsigned int NumElts = VT.getVectorNumElements();
14883 if (NumElts != 8 && NumElts != 16 && !Subtarget->hasBWI())
14886 if (VT.is512BitVector() && InVT.getVectorElementType() != MVT::i1) {
14887 if (In.getOpcode() == X86ISD::VSEXT || In.getOpcode() == X86ISD::VZEXT)
14888 return DAG.getNode(In.getOpcode(), dl, VT, In.getOperand(0));
14889 return DAG.getNode(X86ISD::VSEXT, dl, VT, In);
14892 assert (InVT.getVectorElementType() == MVT::i1 && "Unexpected vector type");
14893 MVT ExtVT = NumElts == 8 ? MVT::v8i64 : MVT::v16i32;
14895 DAG.getConstant(APInt::getAllOnesValue(ExtVT.getScalarSizeInBits()), dl,
14898 DAG.getConstant(APInt::getNullValue(ExtVT.getScalarSizeInBits()), dl, ExtVT);
14900 SDValue V = DAG.getNode(ISD::VSELECT, dl, ExtVT, In, NegOne, Zero);
14901 if (VT.is512BitVector())
14903 return DAG.getNode(X86ISD::VTRUNC, dl, VT, V);
14906 static SDValue LowerSIGN_EXTEND_VECTOR_INREG(SDValue Op,
14907 const X86Subtarget *Subtarget,
14908 SelectionDAG &DAG) {
14909 SDValue In = Op->getOperand(0);
14910 MVT VT = Op->getSimpleValueType(0);
14911 MVT InVT = In.getSimpleValueType();
14912 assert(VT.getSizeInBits() == InVT.getSizeInBits());
14914 MVT InSVT = InVT.getScalarType();
14915 assert(VT.getScalarType().getScalarSizeInBits() > InSVT.getScalarSizeInBits());
14917 if (VT != MVT::v2i64 && VT != MVT::v4i32 && VT != MVT::v8i16)
14919 if (InSVT != MVT::i32 && InSVT != MVT::i16 && InSVT != MVT::i8)
14924 // SSE41 targets can use the pmovsx* instructions directly.
14925 if (Subtarget->hasSSE41())
14926 return DAG.getNode(X86ISD::VSEXT, dl, VT, In);
14928 // pre-SSE41 targets unpack lower lanes and then sign-extend using SRAI.
14932 // As SRAI is only available on i16/i32 types, we expand only up to i32
14933 // and handle i64 separately.
14934 while (CurrVT != VT && CurrVT.getScalarType() != MVT::i32) {
14935 Curr = DAG.getNode(X86ISD::UNPCKL, dl, CurrVT, DAG.getUNDEF(CurrVT), Curr);
14936 MVT CurrSVT = MVT::getIntegerVT(CurrVT.getScalarSizeInBits() * 2);
14937 CurrVT = MVT::getVectorVT(CurrSVT, CurrVT.getVectorNumElements() / 2);
14938 Curr = DAG.getBitcast(CurrVT, Curr);
14941 SDValue SignExt = Curr;
14942 if (CurrVT != InVT) {
14943 unsigned SignExtShift =
14944 CurrVT.getScalarSizeInBits() - InSVT.getScalarSizeInBits();
14945 SignExt = DAG.getNode(X86ISD::VSRAI, dl, CurrVT, Curr,
14946 DAG.getConstant(SignExtShift, dl, MVT::i8));
14952 if (VT == MVT::v2i64 && CurrVT == MVT::v4i32) {
14953 SDValue Sign = DAG.getNode(X86ISD::VSRAI, dl, CurrVT, Curr,
14954 DAG.getConstant(31, dl, MVT::i8));
14955 SDValue Ext = DAG.getVectorShuffle(CurrVT, dl, SignExt, Sign, {0, 4, 1, 5});
14956 return DAG.getBitcast(VT, Ext);
14962 static SDValue LowerSIGN_EXTEND(SDValue Op, const X86Subtarget *Subtarget,
14963 SelectionDAG &DAG) {
14964 MVT VT = Op->getSimpleValueType(0);
14965 SDValue In = Op->getOperand(0);
14966 MVT InVT = In.getSimpleValueType();
14969 if (VT.is512BitVector() || InVT.getVectorElementType() == MVT::i1)
14970 return LowerSIGN_EXTEND_AVX512(Op, Subtarget, DAG);
14972 if ((VT != MVT::v4i64 || InVT != MVT::v4i32) &&
14973 (VT != MVT::v8i32 || InVT != MVT::v8i16) &&
14974 (VT != MVT::v16i16 || InVT != MVT::v16i8))
14977 if (Subtarget->hasInt256())
14978 return DAG.getNode(X86ISD::VSEXT, dl, VT, In);
14980 // Optimize vectors in AVX mode
14981 // Sign extend v8i16 to v8i32 and
14984 // Divide input vector into two parts
14985 // for v4i32 the shuffle mask will be { 0, 1, -1, -1} {2, 3, -1, -1}
14986 // use vpmovsx instruction to extend v4i32 -> v2i64; v8i16 -> v4i32
14987 // concat the vectors to original VT
14989 unsigned NumElems = InVT.getVectorNumElements();
14990 SDValue Undef = DAG.getUNDEF(InVT);
14992 SmallVector<int,8> ShufMask1(NumElems, -1);
14993 for (unsigned i = 0; i != NumElems/2; ++i)
14996 SDValue OpLo = DAG.getVectorShuffle(InVT, dl, In, Undef, &ShufMask1[0]);
14998 SmallVector<int,8> ShufMask2(NumElems, -1);
14999 for (unsigned i = 0; i != NumElems/2; ++i)
15000 ShufMask2[i] = i + NumElems/2;
15002 SDValue OpHi = DAG.getVectorShuffle(InVT, dl, In, Undef, &ShufMask2[0]);
15004 MVT HalfVT = MVT::getVectorVT(VT.getScalarType(),
15005 VT.getVectorNumElements()/2);
15007 OpLo = DAG.getNode(X86ISD::VSEXT, dl, HalfVT, OpLo);
15008 OpHi = DAG.getNode(X86ISD::VSEXT, dl, HalfVT, OpHi);
15010 return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT, OpLo, OpHi);
15013 // Lower vector extended loads using a shuffle. If SSSE3 is not available we
15014 // may emit an illegal shuffle but the expansion is still better than scalar
15015 // code. We generate X86ISD::VSEXT for SEXTLOADs if it's available, otherwise
15016 // we'll emit a shuffle and a arithmetic shift.
15017 // FIXME: Is the expansion actually better than scalar code? It doesn't seem so.
15018 // TODO: It is possible to support ZExt by zeroing the undef values during
15019 // the shuffle phase or after the shuffle.
15020 static SDValue LowerExtendedLoad(SDValue Op, const X86Subtarget *Subtarget,
15021 SelectionDAG &DAG) {
15022 MVT RegVT = Op.getSimpleValueType();
15023 assert(RegVT.isVector() && "We only custom lower vector sext loads.");
15024 assert(RegVT.isInteger() &&
15025 "We only custom lower integer vector sext loads.");
15027 // Nothing useful we can do without SSE2 shuffles.
15028 assert(Subtarget->hasSSE2() && "We only custom lower sext loads with SSE2.");
15030 LoadSDNode *Ld = cast<LoadSDNode>(Op.getNode());
15032 EVT MemVT = Ld->getMemoryVT();
15033 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
15034 unsigned RegSz = RegVT.getSizeInBits();
15036 ISD::LoadExtType Ext = Ld->getExtensionType();
15038 assert((Ext == ISD::EXTLOAD || Ext == ISD::SEXTLOAD)
15039 && "Only anyext and sext are currently implemented.");
15040 assert(MemVT != RegVT && "Cannot extend to the same type");
15041 assert(MemVT.isVector() && "Must load a vector from memory");
15043 unsigned NumElems = RegVT.getVectorNumElements();
15044 unsigned MemSz = MemVT.getSizeInBits();
15045 assert(RegSz > MemSz && "Register size must be greater than the mem size");
15047 if (Ext == ISD::SEXTLOAD && RegSz == 256 && !Subtarget->hasInt256()) {
15048 // The only way in which we have a legal 256-bit vector result but not the
15049 // integer 256-bit operations needed to directly lower a sextload is if we
15050 // have AVX1 but not AVX2. In that case, we can always emit a sextload to
15051 // a 128-bit vector and a normal sign_extend to 256-bits that should get
15052 // correctly legalized. We do this late to allow the canonical form of
15053 // sextload to persist throughout the rest of the DAG combiner -- it wants
15054 // to fold together any extensions it can, and so will fuse a sign_extend
15055 // of an sextload into a sextload targeting a wider value.
15057 if (MemSz == 128) {
15058 // Just switch this to a normal load.
15059 assert(TLI.isTypeLegal(MemVT) && "If the memory type is a 128-bit type, "
15060 "it must be a legal 128-bit vector "
15062 Load = DAG.getLoad(MemVT, dl, Ld->getChain(), Ld->getBasePtr(),
15063 Ld->getPointerInfo(), Ld->isVolatile(), Ld->isNonTemporal(),
15064 Ld->isInvariant(), Ld->getAlignment());
15066 assert(MemSz < 128 &&
15067 "Can't extend a type wider than 128 bits to a 256 bit vector!");
15068 // Do an sext load to a 128-bit vector type. We want to use the same
15069 // number of elements, but elements half as wide. This will end up being
15070 // recursively lowered by this routine, but will succeed as we definitely
15071 // have all the necessary features if we're using AVX1.
15073 EVT::getIntegerVT(*DAG.getContext(), RegVT.getScalarSizeInBits() / 2);
15074 EVT HalfVecVT = EVT::getVectorVT(*DAG.getContext(), HalfEltVT, NumElems);
15076 DAG.getExtLoad(Ext, dl, HalfVecVT, Ld->getChain(), Ld->getBasePtr(),
15077 Ld->getPointerInfo(), MemVT, Ld->isVolatile(),
15078 Ld->isNonTemporal(), Ld->isInvariant(),
15079 Ld->getAlignment());
15082 // Replace chain users with the new chain.
15083 assert(Load->getNumValues() == 2 && "Loads must carry a chain!");
15084 DAG.ReplaceAllUsesOfValueWith(SDValue(Ld, 1), Load.getValue(1));
15086 // Finally, do a normal sign-extend to the desired register.
15087 return DAG.getSExtOrTrunc(Load, dl, RegVT);
15090 // All sizes must be a power of two.
15091 assert(isPowerOf2_32(RegSz * MemSz * NumElems) &&
15092 "Non-power-of-two elements are not custom lowered!");
15094 // Attempt to load the original value using scalar loads.
15095 // Find the largest scalar type that divides the total loaded size.
15096 MVT SclrLoadTy = MVT::i8;
15097 for (MVT Tp : MVT::integer_valuetypes()) {
15098 if (TLI.isTypeLegal(Tp) && ((MemSz % Tp.getSizeInBits()) == 0)) {
15103 // On 32bit systems, we can't save 64bit integers. Try bitcasting to F64.
15104 if (TLI.isTypeLegal(MVT::f64) && SclrLoadTy.getSizeInBits() < 64 &&
15106 SclrLoadTy = MVT::f64;
15108 // Calculate the number of scalar loads that we need to perform
15109 // in order to load our vector from memory.
15110 unsigned NumLoads = MemSz / SclrLoadTy.getSizeInBits();
15112 assert((Ext != ISD::SEXTLOAD || NumLoads == 1) &&
15113 "Can only lower sext loads with a single scalar load!");
15115 unsigned loadRegZize = RegSz;
15116 if (Ext == ISD::SEXTLOAD && RegSz >= 256)
15119 // Represent our vector as a sequence of elements which are the
15120 // largest scalar that we can load.
15121 EVT LoadUnitVecVT = EVT::getVectorVT(
15122 *DAG.getContext(), SclrLoadTy, loadRegZize / SclrLoadTy.getSizeInBits());
15124 // Represent the data using the same element type that is stored in
15125 // memory. In practice, we ''widen'' MemVT.
15127 EVT::getVectorVT(*DAG.getContext(), MemVT.getScalarType(),
15128 loadRegZize / MemVT.getScalarType().getSizeInBits());
15130 assert(WideVecVT.getSizeInBits() == LoadUnitVecVT.getSizeInBits() &&
15131 "Invalid vector type");
15133 // We can't shuffle using an illegal type.
15134 assert(TLI.isTypeLegal(WideVecVT) &&
15135 "We only lower types that form legal widened vector types");
15137 SmallVector<SDValue, 8> Chains;
15138 SDValue Ptr = Ld->getBasePtr();
15139 SDValue Increment = DAG.getConstant(SclrLoadTy.getSizeInBits() / 8, dl,
15140 TLI.getPointerTy(DAG.getDataLayout()));
15141 SDValue Res = DAG.getUNDEF(LoadUnitVecVT);
15143 for (unsigned i = 0; i < NumLoads; ++i) {
15144 // Perform a single load.
15145 SDValue ScalarLoad =
15146 DAG.getLoad(SclrLoadTy, dl, Ld->getChain(), Ptr, Ld->getPointerInfo(),
15147 Ld->isVolatile(), Ld->isNonTemporal(), Ld->isInvariant(),
15148 Ld->getAlignment());
15149 Chains.push_back(ScalarLoad.getValue(1));
15150 // Create the first element type using SCALAR_TO_VECTOR in order to avoid
15151 // another round of DAGCombining.
15153 Res = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, LoadUnitVecVT, ScalarLoad);
15155 Res = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, LoadUnitVecVT, Res,
15156 ScalarLoad, DAG.getIntPtrConstant(i, dl));
15158 Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr, Increment);
15161 SDValue TF = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Chains);
15163 // Bitcast the loaded value to a vector of the original element type, in
15164 // the size of the target vector type.
15165 SDValue SlicedVec = DAG.getBitcast(WideVecVT, Res);
15166 unsigned SizeRatio = RegSz / MemSz;
15168 if (Ext == ISD::SEXTLOAD) {
15169 // If we have SSE4.1, we can directly emit a VSEXT node.
15170 if (Subtarget->hasSSE41()) {
15171 SDValue Sext = DAG.getNode(X86ISD::VSEXT, dl, RegVT, SlicedVec);
15172 DAG.ReplaceAllUsesOfValueWith(SDValue(Ld, 1), TF);
15176 // Otherwise we'll use SIGN_EXTEND_VECTOR_INREG to sign extend the lowest
15178 assert(TLI.isOperationLegalOrCustom(ISD::SIGN_EXTEND_VECTOR_INREG, RegVT) &&
15179 "We can't implement a sext load without SIGN_EXTEND_VECTOR_INREG!");
15181 SDValue Shuff = DAG.getSignExtendVectorInReg(SlicedVec, dl, RegVT);
15182 DAG.ReplaceAllUsesOfValueWith(SDValue(Ld, 1), TF);
15186 // Redistribute the loaded elements into the different locations.
15187 SmallVector<int, 16> ShuffleVec(NumElems * SizeRatio, -1);
15188 for (unsigned i = 0; i != NumElems; ++i)
15189 ShuffleVec[i * SizeRatio] = i;
15191 SDValue Shuff = DAG.getVectorShuffle(WideVecVT, dl, SlicedVec,
15192 DAG.getUNDEF(WideVecVT), &ShuffleVec[0]);
15194 // Bitcast to the requested type.
15195 Shuff = DAG.getBitcast(RegVT, Shuff);
15196 DAG.ReplaceAllUsesOfValueWith(SDValue(Ld, 1), TF);
15200 // isAndOrOfSingleUseSetCCs - Return true if node is an ISD::AND or
15201 // ISD::OR of two X86ISD::SETCC nodes each of which has no other use apart
15202 // from the AND / OR.
15203 static bool isAndOrOfSetCCs(SDValue Op, unsigned &Opc) {
15204 Opc = Op.getOpcode();
15205 if (Opc != ISD::OR && Opc != ISD::AND)
15207 return (Op.getOperand(0).getOpcode() == X86ISD::SETCC &&
15208 Op.getOperand(0).hasOneUse() &&
15209 Op.getOperand(1).getOpcode() == X86ISD::SETCC &&
15210 Op.getOperand(1).hasOneUse());
15213 // isXor1OfSetCC - Return true if node is an ISD::XOR of a X86ISD::SETCC and
15214 // 1 and that the SETCC node has a single use.
15215 static bool isXor1OfSetCC(SDValue Op) {
15216 if (Op.getOpcode() != ISD::XOR)
15218 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(Op.getOperand(1));
15219 if (N1C && N1C->getAPIntValue() == 1) {
15220 return Op.getOperand(0).getOpcode() == X86ISD::SETCC &&
15221 Op.getOperand(0).hasOneUse();
15226 SDValue X86TargetLowering::LowerBRCOND(SDValue Op, SelectionDAG &DAG) const {
15227 bool addTest = true;
15228 SDValue Chain = Op.getOperand(0);
15229 SDValue Cond = Op.getOperand(1);
15230 SDValue Dest = Op.getOperand(2);
15233 bool Inverted = false;
15235 if (Cond.getOpcode() == ISD::SETCC) {
15236 // Check for setcc([su]{add,sub,mul}o == 0).
15237 if (cast<CondCodeSDNode>(Cond.getOperand(2))->get() == ISD::SETEQ &&
15238 isa<ConstantSDNode>(Cond.getOperand(1)) &&
15239 cast<ConstantSDNode>(Cond.getOperand(1))->isNullValue() &&
15240 Cond.getOperand(0).getResNo() == 1 &&
15241 (Cond.getOperand(0).getOpcode() == ISD::SADDO ||
15242 Cond.getOperand(0).getOpcode() == ISD::UADDO ||
15243 Cond.getOperand(0).getOpcode() == ISD::SSUBO ||
15244 Cond.getOperand(0).getOpcode() == ISD::USUBO ||
15245 Cond.getOperand(0).getOpcode() == ISD::SMULO ||
15246 Cond.getOperand(0).getOpcode() == ISD::UMULO)) {
15248 Cond = Cond.getOperand(0);
15250 SDValue NewCond = LowerSETCC(Cond, DAG);
15251 if (NewCond.getNode())
15256 // FIXME: LowerXALUO doesn't handle these!!
15257 else if (Cond.getOpcode() == X86ISD::ADD ||
15258 Cond.getOpcode() == X86ISD::SUB ||
15259 Cond.getOpcode() == X86ISD::SMUL ||
15260 Cond.getOpcode() == X86ISD::UMUL)
15261 Cond = LowerXALUO(Cond, DAG);
15264 // Look pass (and (setcc_carry (cmp ...)), 1).
15265 if (Cond.getOpcode() == ISD::AND &&
15266 Cond.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY) {
15267 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Cond.getOperand(1));
15268 if (C && C->getAPIntValue() == 1)
15269 Cond = Cond.getOperand(0);
15272 // If condition flag is set by a X86ISD::CMP, then use it as the condition
15273 // setting operand in place of the X86ISD::SETCC.
15274 unsigned CondOpcode = Cond.getOpcode();
15275 if (CondOpcode == X86ISD::SETCC ||
15276 CondOpcode == X86ISD::SETCC_CARRY) {
15277 CC = Cond.getOperand(0);
15279 SDValue Cmp = Cond.getOperand(1);
15280 unsigned Opc = Cmp.getOpcode();
15281 // FIXME: WHY THE SPECIAL CASING OF LogicalCmp??
15282 if (isX86LogicalCmp(Cmp) || Opc == X86ISD::BT) {
15286 switch (cast<ConstantSDNode>(CC)->getZExtValue()) {
15290 // These can only come from an arithmetic instruction with overflow,
15291 // e.g. SADDO, UADDO.
15292 Cond = Cond.getNode()->getOperand(1);
15298 CondOpcode = Cond.getOpcode();
15299 if (CondOpcode == ISD::UADDO || CondOpcode == ISD::SADDO ||
15300 CondOpcode == ISD::USUBO || CondOpcode == ISD::SSUBO ||
15301 ((CondOpcode == ISD::UMULO || CondOpcode == ISD::SMULO) &&
15302 Cond.getOperand(0).getValueType() != MVT::i8)) {
15303 SDValue LHS = Cond.getOperand(0);
15304 SDValue RHS = Cond.getOperand(1);
15305 unsigned X86Opcode;
15308 // Keep this in sync with LowerXALUO, otherwise we might create redundant
15309 // instructions that can't be removed afterwards (i.e. X86ISD::ADD and
15311 switch (CondOpcode) {
15312 case ISD::UADDO: X86Opcode = X86ISD::ADD; X86Cond = X86::COND_B; break;
15314 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(RHS))
15316 X86Opcode = X86ISD::INC; X86Cond = X86::COND_O;
15319 X86Opcode = X86ISD::ADD; X86Cond = X86::COND_O; break;
15320 case ISD::USUBO: X86Opcode = X86ISD::SUB; X86Cond = X86::COND_B; break;
15322 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(RHS))
15324 X86Opcode = X86ISD::DEC; X86Cond = X86::COND_O;
15327 X86Opcode = X86ISD::SUB; X86Cond = X86::COND_O; break;
15328 case ISD::UMULO: X86Opcode = X86ISD::UMUL; X86Cond = X86::COND_O; break;
15329 case ISD::SMULO: X86Opcode = X86ISD::SMUL; X86Cond = X86::COND_O; break;
15330 default: llvm_unreachable("unexpected overflowing operator");
15333 X86Cond = X86::GetOppositeBranchCondition((X86::CondCode)X86Cond);
15334 if (CondOpcode == ISD::UMULO)
15335 VTs = DAG.getVTList(LHS.getValueType(), LHS.getValueType(),
15338 VTs = DAG.getVTList(LHS.getValueType(), MVT::i32);
15340 SDValue X86Op = DAG.getNode(X86Opcode, dl, VTs, LHS, RHS);
15342 if (CondOpcode == ISD::UMULO)
15343 Cond = X86Op.getValue(2);
15345 Cond = X86Op.getValue(1);
15347 CC = DAG.getConstant(X86Cond, dl, MVT::i8);
15351 if (Cond.hasOneUse() && isAndOrOfSetCCs(Cond, CondOpc)) {
15352 SDValue Cmp = Cond.getOperand(0).getOperand(1);
15353 if (CondOpc == ISD::OR) {
15354 // Also, recognize the pattern generated by an FCMP_UNE. We can emit
15355 // two branches instead of an explicit OR instruction with a
15357 if (Cmp == Cond.getOperand(1).getOperand(1) &&
15358 isX86LogicalCmp(Cmp)) {
15359 CC = Cond.getOperand(0).getOperand(0);
15360 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
15361 Chain, Dest, CC, Cmp);
15362 CC = Cond.getOperand(1).getOperand(0);
15366 } else { // ISD::AND
15367 // Also, recognize the pattern generated by an FCMP_OEQ. We can emit
15368 // two branches instead of an explicit AND instruction with a
15369 // separate test. However, we only do this if this block doesn't
15370 // have a fall-through edge, because this requires an explicit
15371 // jmp when the condition is false.
15372 if (Cmp == Cond.getOperand(1).getOperand(1) &&
15373 isX86LogicalCmp(Cmp) &&
15374 Op.getNode()->hasOneUse()) {
15375 X86::CondCode CCode =
15376 (X86::CondCode)Cond.getOperand(0).getConstantOperandVal(0);
15377 CCode = X86::GetOppositeBranchCondition(CCode);
15378 CC = DAG.getConstant(CCode, dl, MVT::i8);
15379 SDNode *User = *Op.getNode()->use_begin();
15380 // Look for an unconditional branch following this conditional branch.
15381 // We need this because we need to reverse the successors in order
15382 // to implement FCMP_OEQ.
15383 if (User->getOpcode() == ISD::BR) {
15384 SDValue FalseBB = User->getOperand(1);
15386 DAG.UpdateNodeOperands(User, User->getOperand(0), Dest);
15387 assert(NewBR == User);
15391 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
15392 Chain, Dest, CC, Cmp);
15393 X86::CondCode CCode =
15394 (X86::CondCode)Cond.getOperand(1).getConstantOperandVal(0);
15395 CCode = X86::GetOppositeBranchCondition(CCode);
15396 CC = DAG.getConstant(CCode, dl, MVT::i8);
15402 } else if (Cond.hasOneUse() && isXor1OfSetCC(Cond)) {
15403 // Recognize for xorb (setcc), 1 patterns. The xor inverts the condition.
15404 // It should be transformed during dag combiner except when the condition
15405 // is set by a arithmetics with overflow node.
15406 X86::CondCode CCode =
15407 (X86::CondCode)Cond.getOperand(0).getConstantOperandVal(0);
15408 CCode = X86::GetOppositeBranchCondition(CCode);
15409 CC = DAG.getConstant(CCode, dl, MVT::i8);
15410 Cond = Cond.getOperand(0).getOperand(1);
15412 } else if (Cond.getOpcode() == ISD::SETCC &&
15413 cast<CondCodeSDNode>(Cond.getOperand(2))->get() == ISD::SETOEQ) {
15414 // For FCMP_OEQ, we can emit
15415 // two branches instead of an explicit AND instruction with a
15416 // separate test. However, we only do this if this block doesn't
15417 // have a fall-through edge, because this requires an explicit
15418 // jmp when the condition is false.
15419 if (Op.getNode()->hasOneUse()) {
15420 SDNode *User = *Op.getNode()->use_begin();
15421 // Look for an unconditional branch following this conditional branch.
15422 // We need this because we need to reverse the successors in order
15423 // to implement FCMP_OEQ.
15424 if (User->getOpcode() == ISD::BR) {
15425 SDValue FalseBB = User->getOperand(1);
15427 DAG.UpdateNodeOperands(User, User->getOperand(0), Dest);
15428 assert(NewBR == User);
15432 SDValue Cmp = DAG.getNode(X86ISD::CMP, dl, MVT::i32,
15433 Cond.getOperand(0), Cond.getOperand(1));
15434 Cmp = ConvertCmpIfNecessary(Cmp, DAG);
15435 CC = DAG.getConstant(X86::COND_NE, dl, MVT::i8);
15436 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
15437 Chain, Dest, CC, Cmp);
15438 CC = DAG.getConstant(X86::COND_P, dl, MVT::i8);
15443 } else if (Cond.getOpcode() == ISD::SETCC &&
15444 cast<CondCodeSDNode>(Cond.getOperand(2))->get() == ISD::SETUNE) {
15445 // For FCMP_UNE, we can emit
15446 // two branches instead of an explicit AND instruction with a
15447 // separate test. However, we only do this if this block doesn't
15448 // have a fall-through edge, because this requires an explicit
15449 // jmp when the condition is false.
15450 if (Op.getNode()->hasOneUse()) {
15451 SDNode *User = *Op.getNode()->use_begin();
15452 // Look for an unconditional branch following this conditional branch.
15453 // We need this because we need to reverse the successors in order
15454 // to implement FCMP_UNE.
15455 if (User->getOpcode() == ISD::BR) {
15456 SDValue FalseBB = User->getOperand(1);
15458 DAG.UpdateNodeOperands(User, User->getOperand(0), Dest);
15459 assert(NewBR == User);
15462 SDValue Cmp = DAG.getNode(X86ISD::CMP, dl, MVT::i32,
15463 Cond.getOperand(0), Cond.getOperand(1));
15464 Cmp = ConvertCmpIfNecessary(Cmp, DAG);
15465 CC = DAG.getConstant(X86::COND_NE, dl, MVT::i8);
15466 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
15467 Chain, Dest, CC, Cmp);
15468 CC = DAG.getConstant(X86::COND_NP, dl, MVT::i8);
15478 // Look pass the truncate if the high bits are known zero.
15479 if (isTruncWithZeroHighBitsInput(Cond, DAG))
15480 Cond = Cond.getOperand(0);
15482 // We know the result of AND is compared against zero. Try to match
15484 if (Cond.getOpcode() == ISD::AND && Cond.hasOneUse()) {
15485 SDValue NewSetCC = LowerToBT(Cond, ISD::SETNE, dl, DAG);
15486 if (NewSetCC.getNode()) {
15487 CC = NewSetCC.getOperand(0);
15488 Cond = NewSetCC.getOperand(1);
15495 X86::CondCode X86Cond = Inverted ? X86::COND_E : X86::COND_NE;
15496 CC = DAG.getConstant(X86Cond, dl, MVT::i8);
15497 Cond = EmitTest(Cond, X86Cond, dl, DAG);
15499 Cond = ConvertCmpIfNecessary(Cond, DAG);
15500 return DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
15501 Chain, Dest, CC, Cond);
15504 // Lower dynamic stack allocation to _alloca call for Cygwin/Mingw targets.
15505 // Calls to _alloca are needed to probe the stack when allocating more than 4k
15506 // bytes in one go. Touching the stack at 4K increments is necessary to ensure
15507 // that the guard pages used by the OS virtual memory manager are allocated in
15508 // correct sequence.
15510 X86TargetLowering::LowerDYNAMIC_STACKALLOC(SDValue Op,
15511 SelectionDAG &DAG) const {
15512 MachineFunction &MF = DAG.getMachineFunction();
15513 bool SplitStack = MF.shouldSplitStack();
15514 bool Lower = (Subtarget->isOSWindows() && !Subtarget->isTargetMachO()) ||
15519 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
15520 SDNode* Node = Op.getNode();
15522 unsigned SPReg = TLI.getStackPointerRegisterToSaveRestore();
15523 assert(SPReg && "Target cannot require DYNAMIC_STACKALLOC expansion and"
15524 " not tell us which reg is the stack pointer!");
15525 EVT VT = Node->getValueType(0);
15526 SDValue Tmp1 = SDValue(Node, 0);
15527 SDValue Tmp2 = SDValue(Node, 1);
15528 SDValue Tmp3 = Node->getOperand(2);
15529 SDValue Chain = Tmp1.getOperand(0);
15531 // Chain the dynamic stack allocation so that it doesn't modify the stack
15532 // pointer when other instructions are using the stack.
15533 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(0, dl, true),
15536 SDValue Size = Tmp2.getOperand(1);
15537 SDValue SP = DAG.getCopyFromReg(Chain, dl, SPReg, VT);
15538 Chain = SP.getValue(1);
15539 unsigned Align = cast<ConstantSDNode>(Tmp3)->getZExtValue();
15540 const TargetFrameLowering &TFI = *Subtarget->getFrameLowering();
15541 unsigned StackAlign = TFI.getStackAlignment();
15542 Tmp1 = DAG.getNode(ISD::SUB, dl, VT, SP, Size); // Value
15543 if (Align > StackAlign)
15544 Tmp1 = DAG.getNode(ISD::AND, dl, VT, Tmp1,
15545 DAG.getConstant(-(uint64_t)Align, dl, VT));
15546 Chain = DAG.getCopyToReg(Chain, dl, SPReg, Tmp1); // Output chain
15548 Tmp2 = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(0, dl, true),
15549 DAG.getIntPtrConstant(0, dl, true), SDValue(),
15552 SDValue Ops[2] = { Tmp1, Tmp2 };
15553 return DAG.getMergeValues(Ops, dl);
15557 SDValue Chain = Op.getOperand(0);
15558 SDValue Size = Op.getOperand(1);
15559 unsigned Align = cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue();
15560 EVT VT = Op.getNode()->getValueType(0);
15562 bool Is64Bit = Subtarget->is64Bit();
15563 MVT SPTy = getPointerTy(DAG.getDataLayout());
15566 MachineRegisterInfo &MRI = MF.getRegInfo();
15569 // The 64 bit implementation of segmented stacks needs to clobber both r10
15570 // r11. This makes it impossible to use it along with nested parameters.
15571 const Function *F = MF.getFunction();
15573 for (Function::const_arg_iterator I = F->arg_begin(), E = F->arg_end();
15575 if (I->hasNestAttr())
15576 report_fatal_error("Cannot use segmented stacks with functions that "
15577 "have nested arguments.");
15580 const TargetRegisterClass *AddrRegClass = getRegClassFor(SPTy);
15581 unsigned Vreg = MRI.createVirtualRegister(AddrRegClass);
15582 Chain = DAG.getCopyToReg(Chain, dl, Vreg, Size);
15583 SDValue Value = DAG.getNode(X86ISD::SEG_ALLOCA, dl, SPTy, Chain,
15584 DAG.getRegister(Vreg, SPTy));
15585 SDValue Ops1[2] = { Value, Chain };
15586 return DAG.getMergeValues(Ops1, dl);
15589 const unsigned Reg = (Subtarget->isTarget64BitLP64() ? X86::RAX : X86::EAX);
15591 Chain = DAG.getCopyToReg(Chain, dl, Reg, Size, Flag);
15592 Flag = Chain.getValue(1);
15593 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
15595 Chain = DAG.getNode(X86ISD::WIN_ALLOCA, dl, NodeTys, Chain, Flag);
15597 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
15598 unsigned SPReg = RegInfo->getStackRegister();
15599 SDValue SP = DAG.getCopyFromReg(Chain, dl, SPReg, SPTy);
15600 Chain = SP.getValue(1);
15603 SP = DAG.getNode(ISD::AND, dl, VT, SP.getValue(0),
15604 DAG.getConstant(-(uint64_t)Align, dl, VT));
15605 Chain = DAG.getCopyToReg(Chain, dl, SPReg, SP);
15608 SDValue Ops1[2] = { SP, Chain };
15609 return DAG.getMergeValues(Ops1, dl);
15613 SDValue X86TargetLowering::LowerVASTART(SDValue Op, SelectionDAG &DAG) const {
15614 MachineFunction &MF = DAG.getMachineFunction();
15615 auto PtrVT = getPointerTy(MF.getDataLayout());
15616 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
15618 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
15621 if (!Subtarget->is64Bit() ||
15622 Subtarget->isCallingConvWin64(MF.getFunction()->getCallingConv())) {
15623 // vastart just stores the address of the VarArgsFrameIndex slot into the
15624 // memory location argument.
15625 SDValue FR = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(), PtrVT);
15626 return DAG.getStore(Op.getOperand(0), DL, FR, Op.getOperand(1),
15627 MachinePointerInfo(SV), false, false, 0);
15631 // gp_offset (0 - 6 * 8)
15632 // fp_offset (48 - 48 + 8 * 16)
15633 // overflow_arg_area (point to parameters coming in memory).
15635 SmallVector<SDValue, 8> MemOps;
15636 SDValue FIN = Op.getOperand(1);
15638 SDValue Store = DAG.getStore(Op.getOperand(0), DL,
15639 DAG.getConstant(FuncInfo->getVarArgsGPOffset(),
15641 FIN, MachinePointerInfo(SV), false, false, 0);
15642 MemOps.push_back(Store);
15645 FIN = DAG.getNode(ISD::ADD, DL, PtrVT, FIN, DAG.getIntPtrConstant(4, DL));
15646 Store = DAG.getStore(Op.getOperand(0), DL,
15647 DAG.getConstant(FuncInfo->getVarArgsFPOffset(), DL,
15649 FIN, MachinePointerInfo(SV, 4), false, false, 0);
15650 MemOps.push_back(Store);
15652 // Store ptr to overflow_arg_area
15653 FIN = DAG.getNode(ISD::ADD, DL, PtrVT, FIN, DAG.getIntPtrConstant(4, DL));
15654 SDValue OVFIN = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(), PtrVT);
15655 Store = DAG.getStore(Op.getOperand(0), DL, OVFIN, FIN,
15656 MachinePointerInfo(SV, 8),
15658 MemOps.push_back(Store);
15660 // Store ptr to reg_save_area.
15661 FIN = DAG.getNode(ISD::ADD, DL, PtrVT, FIN, DAG.getIntPtrConstant(
15662 Subtarget->isTarget64BitLP64() ? 8 : 4, DL));
15663 SDValue RSFIN = DAG.getFrameIndex(FuncInfo->getRegSaveFrameIndex(), PtrVT);
15664 Store = DAG.getStore(Op.getOperand(0), DL, RSFIN, FIN, MachinePointerInfo(
15665 SV, Subtarget->isTarget64BitLP64() ? 16 : 12), false, false, 0);
15666 MemOps.push_back(Store);
15667 return DAG.getNode(ISD::TokenFactor, DL, MVT::Other, MemOps);
15670 SDValue X86TargetLowering::LowerVAARG(SDValue Op, SelectionDAG &DAG) const {
15671 assert(Subtarget->is64Bit() &&
15672 "LowerVAARG only handles 64-bit va_arg!");
15673 assert(Op.getNode()->getNumOperands() == 4);
15675 MachineFunction &MF = DAG.getMachineFunction();
15676 if (Subtarget->isCallingConvWin64(MF.getFunction()->getCallingConv()))
15677 // The Win64 ABI uses char* instead of a structure.
15678 return DAG.expandVAArg(Op.getNode());
15680 SDValue Chain = Op.getOperand(0);
15681 SDValue SrcPtr = Op.getOperand(1);
15682 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
15683 unsigned Align = Op.getConstantOperandVal(3);
15686 EVT ArgVT = Op.getNode()->getValueType(0);
15687 Type *ArgTy = ArgVT.getTypeForEVT(*DAG.getContext());
15688 uint32_t ArgSize = DAG.getDataLayout().getTypeAllocSize(ArgTy);
15691 // Decide which area this value should be read from.
15692 // TODO: Implement the AMD64 ABI in its entirety. This simple
15693 // selection mechanism works only for the basic types.
15694 if (ArgVT == MVT::f80) {
15695 llvm_unreachable("va_arg for f80 not yet implemented");
15696 } else if (ArgVT.isFloatingPoint() && ArgSize <= 16 /*bytes*/) {
15697 ArgMode = 2; // Argument passed in XMM register. Use fp_offset.
15698 } else if (ArgVT.isInteger() && ArgSize <= 32 /*bytes*/) {
15699 ArgMode = 1; // Argument passed in GPR64 register(s). Use gp_offset.
15701 llvm_unreachable("Unhandled argument type in LowerVAARG");
15704 if (ArgMode == 2) {
15705 // Sanity Check: Make sure using fp_offset makes sense.
15706 assert(!Subtarget->useSoftFloat() &&
15707 !(MF.getFunction()->hasFnAttribute(Attribute::NoImplicitFloat)) &&
15708 Subtarget->hasSSE1());
15711 // Insert VAARG_64 node into the DAG
15712 // VAARG_64 returns two values: Variable Argument Address, Chain
15713 SDValue InstOps[] = {Chain, SrcPtr, DAG.getConstant(ArgSize, dl, MVT::i32),
15714 DAG.getConstant(ArgMode, dl, MVT::i8),
15715 DAG.getConstant(Align, dl, MVT::i32)};
15716 SDVTList VTs = DAG.getVTList(getPointerTy(DAG.getDataLayout()), MVT::Other);
15717 SDValue VAARG = DAG.getMemIntrinsicNode(X86ISD::VAARG_64, dl,
15718 VTs, InstOps, MVT::i64,
15719 MachinePointerInfo(SV),
15721 /*Volatile=*/false,
15723 /*WriteMem=*/true);
15724 Chain = VAARG.getValue(1);
15726 // Load the next argument and return it
15727 return DAG.getLoad(ArgVT, dl,
15730 MachinePointerInfo(),
15731 false, false, false, 0);
15734 static SDValue LowerVACOPY(SDValue Op, const X86Subtarget *Subtarget,
15735 SelectionDAG &DAG) {
15736 // X86-64 va_list is a struct { i32, i32, i8*, i8* }, except on Windows,
15737 // where a va_list is still an i8*.
15738 assert(Subtarget->is64Bit() && "This code only handles 64-bit va_copy!");
15739 if (Subtarget->isCallingConvWin64(
15740 DAG.getMachineFunction().getFunction()->getCallingConv()))
15741 // Probably a Win64 va_copy.
15742 return DAG.expandVACopy(Op.getNode());
15744 SDValue Chain = Op.getOperand(0);
15745 SDValue DstPtr = Op.getOperand(1);
15746 SDValue SrcPtr = Op.getOperand(2);
15747 const Value *DstSV = cast<SrcValueSDNode>(Op.getOperand(3))->getValue();
15748 const Value *SrcSV = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
15751 return DAG.getMemcpy(Chain, DL, DstPtr, SrcPtr,
15752 DAG.getIntPtrConstant(24, DL), 8, /*isVolatile*/false,
15754 MachinePointerInfo(DstSV), MachinePointerInfo(SrcSV));
15757 // getTargetVShiftByConstNode - Handle vector element shifts where the shift
15758 // amount is a constant. Takes immediate version of shift as input.
15759 static SDValue getTargetVShiftByConstNode(unsigned Opc, SDLoc dl, MVT VT,
15760 SDValue SrcOp, uint64_t ShiftAmt,
15761 SelectionDAG &DAG) {
15762 MVT ElementType = VT.getVectorElementType();
15764 // Fold this packed shift into its first operand if ShiftAmt is 0.
15768 // Check for ShiftAmt >= element width
15769 if (ShiftAmt >= ElementType.getSizeInBits()) {
15770 if (Opc == X86ISD::VSRAI)
15771 ShiftAmt = ElementType.getSizeInBits() - 1;
15773 return DAG.getConstant(0, dl, VT);
15776 assert((Opc == X86ISD::VSHLI || Opc == X86ISD::VSRLI || Opc == X86ISD::VSRAI)
15777 && "Unknown target vector shift-by-constant node");
15779 // Fold this packed vector shift into a build vector if SrcOp is a
15780 // vector of Constants or UNDEFs, and SrcOp valuetype is the same as VT.
15781 if (VT == SrcOp.getSimpleValueType() &&
15782 ISD::isBuildVectorOfConstantSDNodes(SrcOp.getNode())) {
15783 SmallVector<SDValue, 8> Elts;
15784 unsigned NumElts = SrcOp->getNumOperands();
15785 ConstantSDNode *ND;
15788 default: llvm_unreachable(nullptr);
15789 case X86ISD::VSHLI:
15790 for (unsigned i=0; i!=NumElts; ++i) {
15791 SDValue CurrentOp = SrcOp->getOperand(i);
15792 if (CurrentOp->getOpcode() == ISD::UNDEF) {
15793 Elts.push_back(CurrentOp);
15796 ND = cast<ConstantSDNode>(CurrentOp);
15797 const APInt &C = ND->getAPIntValue();
15798 Elts.push_back(DAG.getConstant(C.shl(ShiftAmt), dl, ElementType));
15801 case X86ISD::VSRLI:
15802 for (unsigned i=0; i!=NumElts; ++i) {
15803 SDValue CurrentOp = SrcOp->getOperand(i);
15804 if (CurrentOp->getOpcode() == ISD::UNDEF) {
15805 Elts.push_back(CurrentOp);
15808 ND = cast<ConstantSDNode>(CurrentOp);
15809 const APInt &C = ND->getAPIntValue();
15810 Elts.push_back(DAG.getConstant(C.lshr(ShiftAmt), dl, ElementType));
15813 case X86ISD::VSRAI:
15814 for (unsigned i=0; i!=NumElts; ++i) {
15815 SDValue CurrentOp = SrcOp->getOperand(i);
15816 if (CurrentOp->getOpcode() == ISD::UNDEF) {
15817 Elts.push_back(CurrentOp);
15820 ND = cast<ConstantSDNode>(CurrentOp);
15821 const APInt &C = ND->getAPIntValue();
15822 Elts.push_back(DAG.getConstant(C.ashr(ShiftAmt), dl, ElementType));
15827 return DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Elts);
15830 return DAG.getNode(Opc, dl, VT, SrcOp,
15831 DAG.getConstant(ShiftAmt, dl, MVT::i8));
15834 // getTargetVShiftNode - Handle vector element shifts where the shift amount
15835 // may or may not be a constant. Takes immediate version of shift as input.
15836 static SDValue getTargetVShiftNode(unsigned Opc, SDLoc dl, MVT VT,
15837 SDValue SrcOp, SDValue ShAmt,
15838 SelectionDAG &DAG) {
15839 MVT SVT = ShAmt.getSimpleValueType();
15840 assert((SVT == MVT::i32 || SVT == MVT::i64) && "Unexpected value type!");
15842 // Catch shift-by-constant.
15843 if (ConstantSDNode *CShAmt = dyn_cast<ConstantSDNode>(ShAmt))
15844 return getTargetVShiftByConstNode(Opc, dl, VT, SrcOp,
15845 CShAmt->getZExtValue(), DAG);
15847 // Change opcode to non-immediate version
15849 default: llvm_unreachable("Unknown target vector shift node");
15850 case X86ISD::VSHLI: Opc = X86ISD::VSHL; break;
15851 case X86ISD::VSRLI: Opc = X86ISD::VSRL; break;
15852 case X86ISD::VSRAI: Opc = X86ISD::VSRA; break;
15855 const X86Subtarget &Subtarget =
15856 static_cast<const X86Subtarget &>(DAG.getSubtarget());
15857 if (Subtarget.hasSSE41() && ShAmt.getOpcode() == ISD::ZERO_EXTEND &&
15858 ShAmt.getOperand(0).getSimpleValueType() == MVT::i16) {
15859 // Let the shuffle legalizer expand this shift amount node.
15860 SDValue Op0 = ShAmt.getOperand(0);
15861 Op0 = DAG.getNode(ISD::SCALAR_TO_VECTOR, SDLoc(Op0), MVT::v8i16, Op0);
15862 ShAmt = getShuffleVectorZeroOrUndef(Op0, 0, true, &Subtarget, DAG);
15864 // Need to build a vector containing shift amount.
15865 // SSE/AVX packed shifts only use the lower 64-bit of the shift count.
15866 SmallVector<SDValue, 4> ShOps;
15867 ShOps.push_back(ShAmt);
15868 if (SVT == MVT::i32) {
15869 ShOps.push_back(DAG.getConstant(0, dl, SVT));
15870 ShOps.push_back(DAG.getUNDEF(SVT));
15872 ShOps.push_back(DAG.getUNDEF(SVT));
15874 MVT BVT = SVT == MVT::i32 ? MVT::v4i32 : MVT::v2i64;
15875 ShAmt = DAG.getNode(ISD::BUILD_VECTOR, dl, BVT, ShOps);
15878 // The return type has to be a 128-bit type with the same element
15879 // type as the input type.
15880 MVT EltVT = VT.getVectorElementType();
15881 EVT ShVT = MVT::getVectorVT(EltVT, 128/EltVT.getSizeInBits());
15883 ShAmt = DAG.getBitcast(ShVT, ShAmt);
15884 return DAG.getNode(Opc, dl, VT, SrcOp, ShAmt);
15887 /// \brief Return (and \p Op, \p Mask) for compare instructions or
15888 /// (vselect \p Mask, \p Op, \p PreservedSrc) for others along with the
15889 /// necessary casting or extending for \p Mask when lowering masking intrinsics
15890 static SDValue getVectorMaskingNode(SDValue Op, SDValue Mask,
15891 SDValue PreservedSrc,
15892 const X86Subtarget *Subtarget,
15893 SelectionDAG &DAG) {
15894 EVT VT = Op.getValueType();
15895 EVT MaskVT = EVT::getVectorVT(*DAG.getContext(),
15896 MVT::i1, VT.getVectorNumElements());
15897 SDValue VMask = SDValue();
15898 unsigned OpcodeSelect = ISD::VSELECT;
15901 assert(MaskVT.isSimple() && "invalid mask type");
15903 if (isAllOnes(Mask))
15906 if (MaskVT.bitsGT(Mask.getValueType())) {
15907 EVT newMaskVT = EVT::getIntegerVT(*DAG.getContext(),
15908 MaskVT.getSizeInBits());
15909 VMask = DAG.getBitcast(MaskVT,
15910 DAG.getNode(ISD::ANY_EXTEND, dl, newMaskVT, Mask));
15912 EVT BitcastVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
15913 Mask.getValueType().getSizeInBits());
15914 // In case when MaskVT equals v2i1 or v4i1, low 2 or 4 elements
15915 // are extracted by EXTRACT_SUBVECTOR.
15916 VMask = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MaskVT,
15917 DAG.getBitcast(BitcastVT, Mask),
15918 DAG.getIntPtrConstant(0, dl));
15921 switch (Op.getOpcode()) {
15923 case X86ISD::PCMPEQM:
15924 case X86ISD::PCMPGTM:
15926 case X86ISD::CMPMU:
15927 return DAG.getNode(ISD::AND, dl, VT, Op, VMask);
15928 case X86ISD::VFPCLASS:
15929 return DAG.getNode(ISD::OR, dl, VT, Op, VMask);
15930 case X86ISD::VTRUNC:
15931 case X86ISD::VTRUNCS:
15932 case X86ISD::VTRUNCUS:
15933 // We can't use ISD::VSELECT here because it is not always "Legal"
15934 // for the destination type. For example vpmovqb require only AVX512
15935 // and vselect that can operate on byte element type require BWI
15936 OpcodeSelect = X86ISD::SELECT;
15939 if (PreservedSrc.getOpcode() == ISD::UNDEF)
15940 PreservedSrc = getZeroVector(VT, Subtarget, DAG, dl);
15941 return DAG.getNode(OpcodeSelect, dl, VT, VMask, Op, PreservedSrc);
15944 /// \brief Creates an SDNode for a predicated scalar operation.
15945 /// \returns (X86vselect \p Mask, \p Op, \p PreservedSrc).
15946 /// The mask is coming as MVT::i8 and it should be truncated
15947 /// to MVT::i1 while lowering masking intrinsics.
15948 /// The main difference between ScalarMaskingNode and VectorMaskingNode is using
15949 /// "X86select" instead of "vselect". We just can't create the "vselect" node
15950 /// for a scalar instruction.
15951 static SDValue getScalarMaskingNode(SDValue Op, SDValue Mask,
15952 SDValue PreservedSrc,
15953 const X86Subtarget *Subtarget,
15954 SelectionDAG &DAG) {
15955 if (isAllOnes(Mask))
15958 EVT VT = Op.getValueType();
15960 // The mask should be of type MVT::i1
15961 SDValue IMask = DAG.getNode(ISD::TRUNCATE, dl, MVT::i1, Mask);
15963 if (Op.getOpcode() == X86ISD::FSETCC)
15964 return DAG.getNode(ISD::AND, dl, VT, Op, IMask);
15965 if (Op.getOpcode() == X86ISD::VFPCLASS)
15966 return DAG.getNode(ISD::OR, dl, VT, Op, IMask);
15968 if (PreservedSrc.getOpcode() == ISD::UNDEF)
15969 PreservedSrc = getZeroVector(VT, Subtarget, DAG, dl);
15970 return DAG.getNode(X86ISD::SELECT, dl, VT, IMask, Op, PreservedSrc);
15973 static int getSEHRegistrationNodeSize(const Function *Fn) {
15974 if (!Fn->hasPersonalityFn())
15975 report_fatal_error(
15976 "querying registration node size for function without personality");
15977 // The RegNodeSize is 6 32-bit words for SEH and 4 for C++ EH. See
15978 // WinEHStatePass for the full struct definition.
15979 switch (classifyEHPersonality(Fn->getPersonalityFn())) {
15980 case EHPersonality::MSVC_X86SEH: return 24;
15981 case EHPersonality::MSVC_CXX: return 16;
15984 report_fatal_error("can only recover FP for MSVC EH personality functions");
15987 /// When the 32-bit MSVC runtime transfers control to us, either to an outlined
15988 /// function or when returning to a parent frame after catching an exception, we
15989 /// recover the parent frame pointer by doing arithmetic on the incoming EBP.
15990 /// Here's the math:
15991 /// RegNodeBase = EntryEBP - RegNodeSize
15992 /// ParentFP = RegNodeBase - RegNodeFrameOffset
15993 /// Subtracting RegNodeSize takes us to the offset of the registration node, and
15994 /// subtracting the offset (negative on x86) takes us back to the parent FP.
15995 static SDValue recoverFramePointer(SelectionDAG &DAG, const Function *Fn,
15996 SDValue EntryEBP) {
15997 MachineFunction &MF = DAG.getMachineFunction();
16000 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
16001 MVT PtrVT = TLI.getPointerTy(DAG.getDataLayout());
16003 // It's possible that the parent function no longer has a personality function
16004 // if the exceptional code was optimized away, in which case we just return
16005 // the incoming EBP.
16006 if (!Fn->hasPersonalityFn())
16009 int RegNodeSize = getSEHRegistrationNodeSize(Fn);
16011 // Get an MCSymbol that will ultimately resolve to the frame offset of the EH
16013 MCSymbol *OffsetSym =
16014 MF.getMMI().getContext().getOrCreateParentFrameOffsetSymbol(
16015 GlobalValue::getRealLinkageName(Fn->getName()));
16016 SDValue OffsetSymVal = DAG.getMCSymbol(OffsetSym, PtrVT);
16017 SDValue RegNodeFrameOffset =
16018 DAG.getNode(ISD::LOCAL_RECOVER, dl, PtrVT, OffsetSymVal);
16020 // RegNodeBase = EntryEBP - RegNodeSize
16021 // ParentFP = RegNodeBase - RegNodeFrameOffset
16022 SDValue RegNodeBase = DAG.getNode(ISD::SUB, dl, PtrVT, EntryEBP,
16023 DAG.getConstant(RegNodeSize, dl, PtrVT));
16024 return DAG.getNode(ISD::SUB, dl, PtrVT, RegNodeBase, RegNodeFrameOffset);
16027 static SDValue LowerINTRINSIC_WO_CHAIN(SDValue Op, const X86Subtarget *Subtarget,
16028 SelectionDAG &DAG) {
16030 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
16031 EVT VT = Op.getValueType();
16032 const IntrinsicData* IntrData = getIntrinsicWithoutChain(IntNo);
16034 switch(IntrData->Type) {
16035 case INTR_TYPE_1OP:
16036 return DAG.getNode(IntrData->Opc0, dl, Op.getValueType(), Op.getOperand(1));
16037 case INTR_TYPE_2OP:
16038 return DAG.getNode(IntrData->Opc0, dl, Op.getValueType(), Op.getOperand(1),
16040 case INTR_TYPE_2OP_IMM8:
16041 return DAG.getNode(IntrData->Opc0, dl, Op.getValueType(), Op.getOperand(1),
16042 DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op.getOperand(2)));
16043 case INTR_TYPE_3OP:
16044 return DAG.getNode(IntrData->Opc0, dl, Op.getValueType(), Op.getOperand(1),
16045 Op.getOperand(2), Op.getOperand(3));
16046 case INTR_TYPE_4OP:
16047 return DAG.getNode(IntrData->Opc0, dl, Op.getValueType(), Op.getOperand(1),
16048 Op.getOperand(2), Op.getOperand(3), Op.getOperand(4));
16049 case INTR_TYPE_1OP_MASK_RM: {
16050 SDValue Src = Op.getOperand(1);
16051 SDValue PassThru = Op.getOperand(2);
16052 SDValue Mask = Op.getOperand(3);
16053 SDValue RoundingMode;
16054 // We allways add rounding mode to the Node.
16055 // If the rounding mode is not specified, we add the
16056 // "current direction" mode.
16057 if (Op.getNumOperands() == 4)
16059 DAG.getConstant(X86::STATIC_ROUNDING::CUR_DIRECTION, dl, MVT::i32);
16061 RoundingMode = Op.getOperand(4);
16062 unsigned IntrWithRoundingModeOpcode = IntrData->Opc1;
16063 if (IntrWithRoundingModeOpcode != 0)
16064 if (cast<ConstantSDNode>(RoundingMode)->getZExtValue() !=
16065 X86::STATIC_ROUNDING::CUR_DIRECTION)
16066 return getVectorMaskingNode(DAG.getNode(IntrWithRoundingModeOpcode,
16067 dl, Op.getValueType(), Src, RoundingMode),
16068 Mask, PassThru, Subtarget, DAG);
16069 return getVectorMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT, Src,
16071 Mask, PassThru, Subtarget, DAG);
16073 case INTR_TYPE_1OP_MASK: {
16074 SDValue Src = Op.getOperand(1);
16075 SDValue PassThru = Op.getOperand(2);
16076 SDValue Mask = Op.getOperand(3);
16077 // We add rounding mode to the Node when
16078 // - RM Opcode is specified and
16079 // - RM is not "current direction".
16080 unsigned IntrWithRoundingModeOpcode = IntrData->Opc1;
16081 if (IntrWithRoundingModeOpcode != 0) {
16082 SDValue Rnd = Op.getOperand(4);
16083 unsigned Round = cast<ConstantSDNode>(Rnd)->getZExtValue();
16084 if (Round != X86::STATIC_ROUNDING::CUR_DIRECTION) {
16085 return getVectorMaskingNode(DAG.getNode(IntrWithRoundingModeOpcode,
16086 dl, Op.getValueType(),
16088 Mask, PassThru, Subtarget, DAG);
16091 return getVectorMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT, Src),
16092 Mask, PassThru, Subtarget, DAG);
16094 case INTR_TYPE_SCALAR_MASK: {
16095 SDValue Src1 = Op.getOperand(1);
16096 SDValue Src2 = Op.getOperand(2);
16097 SDValue passThru = Op.getOperand(3);
16098 SDValue Mask = Op.getOperand(4);
16099 return getScalarMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT, Src1, Src2),
16100 Mask, passThru, Subtarget, DAG);
16102 case INTR_TYPE_SCALAR_MASK_RM: {
16103 SDValue Src1 = Op.getOperand(1);
16104 SDValue Src2 = Op.getOperand(2);
16105 SDValue Src0 = Op.getOperand(3);
16106 SDValue Mask = Op.getOperand(4);
16107 // There are 2 kinds of intrinsics in this group:
16108 // (1) With suppress-all-exceptions (sae) or rounding mode- 6 operands
16109 // (2) With rounding mode and sae - 7 operands.
16110 if (Op.getNumOperands() == 6) {
16111 SDValue Sae = Op.getOperand(5);
16112 unsigned Opc = IntrData->Opc1 ? IntrData->Opc1 : IntrData->Opc0;
16113 return getScalarMaskingNode(DAG.getNode(Opc, dl, VT, Src1, Src2,
16115 Mask, Src0, Subtarget, DAG);
16117 assert(Op.getNumOperands() == 7 && "Unexpected intrinsic form");
16118 SDValue RoundingMode = Op.getOperand(5);
16119 SDValue Sae = Op.getOperand(6);
16120 return getScalarMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT, Src1, Src2,
16121 RoundingMode, Sae),
16122 Mask, Src0, Subtarget, DAG);
16124 case INTR_TYPE_2OP_MASK:
16125 case INTR_TYPE_2OP_IMM8_MASK: {
16126 SDValue Src1 = Op.getOperand(1);
16127 SDValue Src2 = Op.getOperand(2);
16128 SDValue PassThru = Op.getOperand(3);
16129 SDValue Mask = Op.getOperand(4);
16131 if (IntrData->Type == INTR_TYPE_2OP_IMM8_MASK)
16132 Src2 = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Src2);
16134 // We specify 2 possible opcodes for intrinsics with rounding modes.
16135 // First, we check if the intrinsic may have non-default rounding mode,
16136 // (IntrData->Opc1 != 0), then we check the rounding mode operand.
16137 unsigned IntrWithRoundingModeOpcode = IntrData->Opc1;
16138 if (IntrWithRoundingModeOpcode != 0) {
16139 SDValue Rnd = Op.getOperand(5);
16140 unsigned Round = cast<ConstantSDNode>(Rnd)->getZExtValue();
16141 if (Round != X86::STATIC_ROUNDING::CUR_DIRECTION) {
16142 return getVectorMaskingNode(DAG.getNode(IntrWithRoundingModeOpcode,
16143 dl, Op.getValueType(),
16145 Mask, PassThru, Subtarget, DAG);
16148 // TODO: Intrinsics should have fast-math-flags to propagate.
16149 return getVectorMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT,Src1,Src2),
16150 Mask, PassThru, Subtarget, DAG);
16152 case INTR_TYPE_2OP_MASK_RM: {
16153 SDValue Src1 = Op.getOperand(1);
16154 SDValue Src2 = Op.getOperand(2);
16155 SDValue PassThru = Op.getOperand(3);
16156 SDValue Mask = Op.getOperand(4);
16157 // We specify 2 possible modes for intrinsics, with/without rounding
16159 // First, we check if the intrinsic have rounding mode (6 operands),
16160 // if not, we set rounding mode to "current".
16162 if (Op.getNumOperands() == 6)
16163 Rnd = Op.getOperand(5);
16165 Rnd = DAG.getConstant(X86::STATIC_ROUNDING::CUR_DIRECTION, dl, MVT::i32);
16166 return getVectorMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT,
16168 Mask, PassThru, Subtarget, DAG);
16170 case INTR_TYPE_3OP_SCALAR_MASK_RM: {
16171 SDValue Src1 = Op.getOperand(1);
16172 SDValue Src2 = Op.getOperand(2);
16173 SDValue Src3 = Op.getOperand(3);
16174 SDValue PassThru = Op.getOperand(4);
16175 SDValue Mask = Op.getOperand(5);
16176 SDValue Sae = Op.getOperand(6);
16178 return getScalarMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT, Src1,
16180 Mask, PassThru, Subtarget, DAG);
16182 case INTR_TYPE_3OP_MASK_RM: {
16183 SDValue Src1 = Op.getOperand(1);
16184 SDValue Src2 = Op.getOperand(2);
16185 SDValue Imm = Op.getOperand(3);
16186 SDValue PassThru = Op.getOperand(4);
16187 SDValue Mask = Op.getOperand(5);
16188 // We specify 2 possible modes for intrinsics, with/without rounding
16190 // First, we check if the intrinsic have rounding mode (7 operands),
16191 // if not, we set rounding mode to "current".
16193 if (Op.getNumOperands() == 7)
16194 Rnd = Op.getOperand(6);
16196 Rnd = DAG.getConstant(X86::STATIC_ROUNDING::CUR_DIRECTION, dl, MVT::i32);
16197 return getVectorMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT,
16198 Src1, Src2, Imm, Rnd),
16199 Mask, PassThru, Subtarget, DAG);
16201 case INTR_TYPE_3OP_IMM8_MASK:
16202 case INTR_TYPE_3OP_MASK:
16203 case INSERT_SUBVEC: {
16204 SDValue Src1 = Op.getOperand(1);
16205 SDValue Src2 = Op.getOperand(2);
16206 SDValue Src3 = Op.getOperand(3);
16207 SDValue PassThru = Op.getOperand(4);
16208 SDValue Mask = Op.getOperand(5);
16210 if (IntrData->Type == INTR_TYPE_3OP_IMM8_MASK)
16211 Src3 = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Src3);
16212 else if (IntrData->Type == INSERT_SUBVEC) {
16213 // imm should be adapted to ISD::INSERT_SUBVECTOR behavior
16214 assert(isa<ConstantSDNode>(Src3) && "Expected a ConstantSDNode here!");
16215 unsigned Imm = cast<ConstantSDNode>(Src3)->getZExtValue();
16216 Imm *= Src2.getValueType().getVectorNumElements();
16217 Src3 = DAG.getTargetConstant(Imm, dl, MVT::i32);
16220 // We specify 2 possible opcodes for intrinsics with rounding modes.
16221 // First, we check if the intrinsic may have non-default rounding mode,
16222 // (IntrData->Opc1 != 0), then we check the rounding mode operand.
16223 unsigned IntrWithRoundingModeOpcode = IntrData->Opc1;
16224 if (IntrWithRoundingModeOpcode != 0) {
16225 SDValue Rnd = Op.getOperand(6);
16226 unsigned Round = cast<ConstantSDNode>(Rnd)->getZExtValue();
16227 if (Round != X86::STATIC_ROUNDING::CUR_DIRECTION) {
16228 return getVectorMaskingNode(DAG.getNode(IntrWithRoundingModeOpcode,
16229 dl, Op.getValueType(),
16230 Src1, Src2, Src3, Rnd),
16231 Mask, PassThru, Subtarget, DAG);
16234 return getVectorMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT,
16236 Mask, PassThru, Subtarget, DAG);
16238 case VPERM_3OP_MASKZ:
16239 case VPERM_3OP_MASK:
16242 case FMA_OP_MASK: {
16243 SDValue Src1 = Op.getOperand(1);
16244 SDValue Src2 = Op.getOperand(2);
16245 SDValue Src3 = Op.getOperand(3);
16246 SDValue Mask = Op.getOperand(4);
16247 EVT VT = Op.getValueType();
16248 SDValue PassThru = SDValue();
16250 // set PassThru element
16251 if (IntrData->Type == VPERM_3OP_MASKZ || IntrData->Type == FMA_OP_MASKZ)
16252 PassThru = getZeroVector(VT, Subtarget, DAG, dl);
16253 else if (IntrData->Type == FMA_OP_MASK3)
16258 // We specify 2 possible opcodes for intrinsics with rounding modes.
16259 // First, we check if the intrinsic may have non-default rounding mode,
16260 // (IntrData->Opc1 != 0), then we check the rounding mode operand.
16261 unsigned IntrWithRoundingModeOpcode = IntrData->Opc1;
16262 if (IntrWithRoundingModeOpcode != 0) {
16263 SDValue Rnd = Op.getOperand(5);
16264 if (cast<ConstantSDNode>(Rnd)->getZExtValue() !=
16265 X86::STATIC_ROUNDING::CUR_DIRECTION)
16266 return getVectorMaskingNode(DAG.getNode(IntrWithRoundingModeOpcode,
16267 dl, Op.getValueType(),
16268 Src1, Src2, Src3, Rnd),
16269 Mask, PassThru, Subtarget, DAG);
16271 return getVectorMaskingNode(DAG.getNode(IntrData->Opc0,
16272 dl, Op.getValueType(),
16274 Mask, PassThru, Subtarget, DAG);
16276 case TERLOG_OP_MASK:
16277 case TERLOG_OP_MASKZ: {
16278 SDValue Src1 = Op.getOperand(1);
16279 SDValue Src2 = Op.getOperand(2);
16280 SDValue Src3 = Op.getOperand(3);
16281 SDValue Src4 = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op.getOperand(4));
16282 SDValue Mask = Op.getOperand(5);
16283 EVT VT = Op.getValueType();
16284 SDValue PassThru = Src1;
16285 // Set PassThru element.
16286 if (IntrData->Type == TERLOG_OP_MASKZ)
16287 PassThru = getZeroVector(VT, Subtarget, DAG, dl);
16289 return getVectorMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT,
16290 Src1, Src2, Src3, Src4),
16291 Mask, PassThru, Subtarget, DAG);
16294 // FPclass intrinsics with mask
16295 SDValue Src1 = Op.getOperand(1);
16296 EVT VT = Src1.getValueType();
16297 EVT MaskVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
16298 VT.getVectorNumElements());
16299 SDValue Imm = Op.getOperand(2);
16300 SDValue Mask = Op.getOperand(3);
16301 EVT BitcastVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
16302 Mask.getValueType().getSizeInBits());
16303 SDValue FPclass = DAG.getNode(IntrData->Opc0, dl, MaskVT, Src1, Imm);
16304 SDValue FPclassMask = getVectorMaskingNode(FPclass, Mask,
16305 DAG.getTargetConstant(0, dl, MaskVT),
16307 SDValue Res = DAG.getNode(ISD::INSERT_SUBVECTOR, dl, BitcastVT,
16308 DAG.getUNDEF(BitcastVT), FPclassMask,
16309 DAG.getIntPtrConstant(0, dl));
16310 return DAG.getBitcast(Op.getValueType(), Res);
16313 SDValue Src1 = Op.getOperand(1);
16314 SDValue Imm = Op.getOperand(2);
16315 SDValue Mask = Op.getOperand(3);
16316 SDValue FPclass = DAG.getNode(IntrData->Opc0, dl, MVT::i1, Src1, Imm);
16317 SDValue FPclassMask = getScalarMaskingNode(FPclass, Mask,
16318 DAG.getTargetConstant(0, dl, MVT::i1), Subtarget, DAG);
16319 return DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::i8, FPclassMask);
16322 case CMP_MASK_CC: {
16323 // Comparison intrinsics with masks.
16324 // Example of transformation:
16325 // (i8 (int_x86_avx512_mask_pcmpeq_q_128
16326 // (v2i64 %a), (v2i64 %b), (i8 %mask))) ->
16328 // (v8i1 (insert_subvector undef,
16329 // (v2i1 (and (PCMPEQM %a, %b),
16330 // (extract_subvector
16331 // (v8i1 (bitcast %mask)), 0))), 0))))
16332 EVT VT = Op.getOperand(1).getValueType();
16333 EVT MaskVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
16334 VT.getVectorNumElements());
16335 SDValue Mask = Op.getOperand((IntrData->Type == CMP_MASK_CC) ? 4 : 3);
16336 EVT BitcastVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
16337 Mask.getValueType().getSizeInBits());
16339 if (IntrData->Type == CMP_MASK_CC) {
16340 SDValue CC = Op.getOperand(3);
16341 CC = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, CC);
16342 // We specify 2 possible opcodes for intrinsics with rounding modes.
16343 // First, we check if the intrinsic may have non-default rounding mode,
16344 // (IntrData->Opc1 != 0), then we check the rounding mode operand.
16345 if (IntrData->Opc1 != 0) {
16346 SDValue Rnd = Op.getOperand(5);
16347 if (cast<ConstantSDNode>(Rnd)->getZExtValue() !=
16348 X86::STATIC_ROUNDING::CUR_DIRECTION)
16349 Cmp = DAG.getNode(IntrData->Opc1, dl, MaskVT, Op.getOperand(1),
16350 Op.getOperand(2), CC, Rnd);
16352 //default rounding mode
16354 Cmp = DAG.getNode(IntrData->Opc0, dl, MaskVT, Op.getOperand(1),
16355 Op.getOperand(2), CC);
16358 assert(IntrData->Type == CMP_MASK && "Unexpected intrinsic type!");
16359 Cmp = DAG.getNode(IntrData->Opc0, dl, MaskVT, Op.getOperand(1),
16362 SDValue CmpMask = getVectorMaskingNode(Cmp, Mask,
16363 DAG.getTargetConstant(0, dl,
16366 SDValue Res = DAG.getNode(ISD::INSERT_SUBVECTOR, dl, BitcastVT,
16367 DAG.getUNDEF(BitcastVT), CmpMask,
16368 DAG.getIntPtrConstant(0, dl));
16369 return DAG.getBitcast(Op.getValueType(), Res);
16371 case CMP_MASK_SCALAR_CC: {
16372 SDValue Src1 = Op.getOperand(1);
16373 SDValue Src2 = Op.getOperand(2);
16374 SDValue CC = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op.getOperand(3));
16375 SDValue Mask = Op.getOperand(4);
16378 if (IntrData->Opc1 != 0) {
16379 SDValue Rnd = Op.getOperand(5);
16380 if (cast<ConstantSDNode>(Rnd)->getZExtValue() !=
16381 X86::STATIC_ROUNDING::CUR_DIRECTION)
16382 Cmp = DAG.getNode(IntrData->Opc1, dl, MVT::i1, Src1, Src2, CC, Rnd);
16384 //default rounding mode
16386 Cmp = DAG.getNode(IntrData->Opc0, dl, MVT::i1, Src1, Src2, CC);
16388 SDValue CmpMask = getScalarMaskingNode(Cmp, Mask,
16389 DAG.getTargetConstant(0, dl,
16393 return DAG.getNode(ISD::SIGN_EXTEND_INREG, dl, MVT::i8,
16394 DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i8, CmpMask),
16395 DAG.getValueType(MVT::i1));
16397 case COMI: { // Comparison intrinsics
16398 ISD::CondCode CC = (ISD::CondCode)IntrData->Opc1;
16399 SDValue LHS = Op.getOperand(1);
16400 SDValue RHS = Op.getOperand(2);
16401 unsigned X86CC = TranslateX86CC(CC, dl, true, LHS, RHS, DAG);
16402 assert(X86CC != X86::COND_INVALID && "Unexpected illegal condition!");
16403 SDValue Cond = DAG.getNode(IntrData->Opc0, dl, MVT::i32, LHS, RHS);
16404 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
16405 DAG.getConstant(X86CC, dl, MVT::i8), Cond);
16406 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
16409 return getTargetVShiftNode(IntrData->Opc0, dl, Op.getSimpleValueType(),
16410 Op.getOperand(1), Op.getOperand(2), DAG);
16412 return getVectorMaskingNode(getTargetVShiftNode(IntrData->Opc0, dl,
16413 Op.getSimpleValueType(),
16415 Op.getOperand(2), DAG),
16416 Op.getOperand(4), Op.getOperand(3), Subtarget,
16418 case COMPRESS_EXPAND_IN_REG: {
16419 SDValue Mask = Op.getOperand(3);
16420 SDValue DataToCompress = Op.getOperand(1);
16421 SDValue PassThru = Op.getOperand(2);
16422 if (isAllOnes(Mask)) // return data as is
16423 return Op.getOperand(1);
16425 return getVectorMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT,
16427 Mask, PassThru, Subtarget, DAG);
16430 SDValue Mask = Op.getOperand(3);
16431 EVT VT = Op.getValueType();
16432 EVT MaskVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
16433 VT.getVectorNumElements());
16434 EVT BitcastVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
16435 Mask.getValueType().getSizeInBits());
16437 SDValue VMask = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MaskVT,
16438 DAG.getBitcast(BitcastVT, Mask),
16439 DAG.getIntPtrConstant(0, dl));
16440 return DAG.getNode(IntrData->Opc0, dl, VT, VMask, Op.getOperand(1),
16449 default: return SDValue(); // Don't custom lower most intrinsics.
16451 case Intrinsic::x86_avx2_permd:
16452 case Intrinsic::x86_avx2_permps:
16453 // Operands intentionally swapped. Mask is last operand to intrinsic,
16454 // but second operand for node/instruction.
16455 return DAG.getNode(X86ISD::VPERMV, dl, Op.getValueType(),
16456 Op.getOperand(2), Op.getOperand(1));
16458 // ptest and testp intrinsics. The intrinsic these come from are designed to
16459 // return an integer value, not just an instruction so lower it to the ptest
16460 // or testp pattern and a setcc for the result.
16461 case Intrinsic::x86_sse41_ptestz:
16462 case Intrinsic::x86_sse41_ptestc:
16463 case Intrinsic::x86_sse41_ptestnzc:
16464 case Intrinsic::x86_avx_ptestz_256:
16465 case Intrinsic::x86_avx_ptestc_256:
16466 case Intrinsic::x86_avx_ptestnzc_256:
16467 case Intrinsic::x86_avx_vtestz_ps:
16468 case Intrinsic::x86_avx_vtestc_ps:
16469 case Intrinsic::x86_avx_vtestnzc_ps:
16470 case Intrinsic::x86_avx_vtestz_pd:
16471 case Intrinsic::x86_avx_vtestc_pd:
16472 case Intrinsic::x86_avx_vtestnzc_pd:
16473 case Intrinsic::x86_avx_vtestz_ps_256:
16474 case Intrinsic::x86_avx_vtestc_ps_256:
16475 case Intrinsic::x86_avx_vtestnzc_ps_256:
16476 case Intrinsic::x86_avx_vtestz_pd_256:
16477 case Intrinsic::x86_avx_vtestc_pd_256:
16478 case Intrinsic::x86_avx_vtestnzc_pd_256: {
16479 bool IsTestPacked = false;
16482 default: llvm_unreachable("Bad fallthrough in Intrinsic lowering.");
16483 case Intrinsic::x86_avx_vtestz_ps:
16484 case Intrinsic::x86_avx_vtestz_pd:
16485 case Intrinsic::x86_avx_vtestz_ps_256:
16486 case Intrinsic::x86_avx_vtestz_pd_256:
16487 IsTestPacked = true; // Fallthrough
16488 case Intrinsic::x86_sse41_ptestz:
16489 case Intrinsic::x86_avx_ptestz_256:
16491 X86CC = X86::COND_E;
16493 case Intrinsic::x86_avx_vtestc_ps:
16494 case Intrinsic::x86_avx_vtestc_pd:
16495 case Intrinsic::x86_avx_vtestc_ps_256:
16496 case Intrinsic::x86_avx_vtestc_pd_256:
16497 IsTestPacked = true; // Fallthrough
16498 case Intrinsic::x86_sse41_ptestc:
16499 case Intrinsic::x86_avx_ptestc_256:
16501 X86CC = X86::COND_B;
16503 case Intrinsic::x86_avx_vtestnzc_ps:
16504 case Intrinsic::x86_avx_vtestnzc_pd:
16505 case Intrinsic::x86_avx_vtestnzc_ps_256:
16506 case Intrinsic::x86_avx_vtestnzc_pd_256:
16507 IsTestPacked = true; // Fallthrough
16508 case Intrinsic::x86_sse41_ptestnzc:
16509 case Intrinsic::x86_avx_ptestnzc_256:
16511 X86CC = X86::COND_A;
16515 SDValue LHS = Op.getOperand(1);
16516 SDValue RHS = Op.getOperand(2);
16517 unsigned TestOpc = IsTestPacked ? X86ISD::TESTP : X86ISD::PTEST;
16518 SDValue Test = DAG.getNode(TestOpc, dl, MVT::i32, LHS, RHS);
16519 SDValue CC = DAG.getConstant(X86CC, dl, MVT::i8);
16520 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8, CC, Test);
16521 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
16523 case Intrinsic::x86_avx512_kortestz_w:
16524 case Intrinsic::x86_avx512_kortestc_w: {
16525 unsigned X86CC = (IntNo == Intrinsic::x86_avx512_kortestz_w)? X86::COND_E: X86::COND_B;
16526 SDValue LHS = DAG.getBitcast(MVT::v16i1, Op.getOperand(1));
16527 SDValue RHS = DAG.getBitcast(MVT::v16i1, Op.getOperand(2));
16528 SDValue CC = DAG.getConstant(X86CC, dl, MVT::i8);
16529 SDValue Test = DAG.getNode(X86ISD::KORTEST, dl, MVT::i32, LHS, RHS);
16530 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i1, CC, Test);
16531 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
16534 case Intrinsic::x86_sse42_pcmpistria128:
16535 case Intrinsic::x86_sse42_pcmpestria128:
16536 case Intrinsic::x86_sse42_pcmpistric128:
16537 case Intrinsic::x86_sse42_pcmpestric128:
16538 case Intrinsic::x86_sse42_pcmpistrio128:
16539 case Intrinsic::x86_sse42_pcmpestrio128:
16540 case Intrinsic::x86_sse42_pcmpistris128:
16541 case Intrinsic::x86_sse42_pcmpestris128:
16542 case Intrinsic::x86_sse42_pcmpistriz128:
16543 case Intrinsic::x86_sse42_pcmpestriz128: {
16547 default: llvm_unreachable("Impossible intrinsic"); // Can't reach here.
16548 case Intrinsic::x86_sse42_pcmpistria128:
16549 Opcode = X86ISD::PCMPISTRI;
16550 X86CC = X86::COND_A;
16552 case Intrinsic::x86_sse42_pcmpestria128:
16553 Opcode = X86ISD::PCMPESTRI;
16554 X86CC = X86::COND_A;
16556 case Intrinsic::x86_sse42_pcmpistric128:
16557 Opcode = X86ISD::PCMPISTRI;
16558 X86CC = X86::COND_B;
16560 case Intrinsic::x86_sse42_pcmpestric128:
16561 Opcode = X86ISD::PCMPESTRI;
16562 X86CC = X86::COND_B;
16564 case Intrinsic::x86_sse42_pcmpistrio128:
16565 Opcode = X86ISD::PCMPISTRI;
16566 X86CC = X86::COND_O;
16568 case Intrinsic::x86_sse42_pcmpestrio128:
16569 Opcode = X86ISD::PCMPESTRI;
16570 X86CC = X86::COND_O;
16572 case Intrinsic::x86_sse42_pcmpistris128:
16573 Opcode = X86ISD::PCMPISTRI;
16574 X86CC = X86::COND_S;
16576 case Intrinsic::x86_sse42_pcmpestris128:
16577 Opcode = X86ISD::PCMPESTRI;
16578 X86CC = X86::COND_S;
16580 case Intrinsic::x86_sse42_pcmpistriz128:
16581 Opcode = X86ISD::PCMPISTRI;
16582 X86CC = X86::COND_E;
16584 case Intrinsic::x86_sse42_pcmpestriz128:
16585 Opcode = X86ISD::PCMPESTRI;
16586 X86CC = X86::COND_E;
16589 SmallVector<SDValue, 5> NewOps(Op->op_begin()+1, Op->op_end());
16590 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::i32);
16591 SDValue PCMP = DAG.getNode(Opcode, dl, VTs, NewOps);
16592 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
16593 DAG.getConstant(X86CC, dl, MVT::i8),
16594 SDValue(PCMP.getNode(), 1));
16595 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
16598 case Intrinsic::x86_sse42_pcmpistri128:
16599 case Intrinsic::x86_sse42_pcmpestri128: {
16601 if (IntNo == Intrinsic::x86_sse42_pcmpistri128)
16602 Opcode = X86ISD::PCMPISTRI;
16604 Opcode = X86ISD::PCMPESTRI;
16606 SmallVector<SDValue, 5> NewOps(Op->op_begin()+1, Op->op_end());
16607 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::i32);
16608 return DAG.getNode(Opcode, dl, VTs, NewOps);
16611 case Intrinsic::x86_seh_lsda: {
16612 // Compute the symbol for the LSDA. We know it'll get emitted later.
16613 MachineFunction &MF = DAG.getMachineFunction();
16614 SDValue Op1 = Op.getOperand(1);
16615 auto *Fn = cast<Function>(cast<GlobalAddressSDNode>(Op1)->getGlobal());
16616 MCSymbol *LSDASym = MF.getMMI().getContext().getOrCreateLSDASymbol(
16617 GlobalValue::getRealLinkageName(Fn->getName()));
16619 // Generate a simple absolute symbol reference. This intrinsic is only
16620 // supported on 32-bit Windows, which isn't PIC.
16621 SDValue Result = DAG.getMCSymbol(LSDASym, VT);
16622 return DAG.getNode(X86ISD::Wrapper, dl, VT, Result);
16625 case Intrinsic::x86_seh_recoverfp: {
16626 SDValue FnOp = Op.getOperand(1);
16627 SDValue IncomingFPOp = Op.getOperand(2);
16628 GlobalAddressSDNode *GSD = dyn_cast<GlobalAddressSDNode>(FnOp);
16629 auto *Fn = dyn_cast_or_null<Function>(GSD ? GSD->getGlobal() : nullptr);
16631 report_fatal_error(
16632 "llvm.x86.seh.recoverfp must take a function as the first argument");
16633 return recoverFramePointer(DAG, Fn, IncomingFPOp);
16636 case Intrinsic::localaddress: {
16637 // Returns one of the stack, base, or frame pointer registers, depending on
16638 // which is used to reference local variables.
16639 MachineFunction &MF = DAG.getMachineFunction();
16640 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
16642 if (RegInfo->hasBasePointer(MF))
16643 Reg = RegInfo->getBaseRegister();
16644 else // This function handles the SP or FP case.
16645 Reg = RegInfo->getPtrSizedFrameRegister(MF);
16646 return DAG.getCopyFromReg(DAG.getEntryNode(), dl, Reg, VT);
16651 static SDValue getGatherNode(unsigned Opc, SDValue Op, SelectionDAG &DAG,
16652 SDValue Src, SDValue Mask, SDValue Base,
16653 SDValue Index, SDValue ScaleOp, SDValue Chain,
16654 const X86Subtarget * Subtarget) {
16656 ConstantSDNode *C = dyn_cast<ConstantSDNode>(ScaleOp);
16658 llvm_unreachable("Invalid scale type");
16659 unsigned ScaleVal = C->getZExtValue();
16660 if (ScaleVal > 2 && ScaleVal != 4 && ScaleVal != 8)
16661 llvm_unreachable("Valid scale values are 1, 2, 4, 8");
16663 SDValue Scale = DAG.getTargetConstant(C->getZExtValue(), dl, MVT::i8);
16664 EVT MaskVT = MVT::getVectorVT(MVT::i1,
16665 Index.getSimpleValueType().getVectorNumElements());
16667 ConstantSDNode *MaskC = dyn_cast<ConstantSDNode>(Mask);
16669 MaskInReg = DAG.getTargetConstant(MaskC->getSExtValue(), dl, MaskVT);
16671 EVT BitcastVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
16672 Mask.getValueType().getSizeInBits());
16674 // In case when MaskVT equals v2i1 or v4i1, low 2 or 4 elements
16675 // are extracted by EXTRACT_SUBVECTOR.
16676 MaskInReg = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MaskVT,
16677 DAG.getBitcast(BitcastVT, Mask),
16678 DAG.getIntPtrConstant(0, dl));
16680 SDVTList VTs = DAG.getVTList(Op.getValueType(), MaskVT, MVT::Other);
16681 SDValue Disp = DAG.getTargetConstant(0, dl, MVT::i32);
16682 SDValue Segment = DAG.getRegister(0, MVT::i32);
16683 if (Src.getOpcode() == ISD::UNDEF)
16684 Src = getZeroVector(Op.getValueType(), Subtarget, DAG, dl);
16685 SDValue Ops[] = {Src, MaskInReg, Base, Scale, Index, Disp, Segment, Chain};
16686 SDNode *Res = DAG.getMachineNode(Opc, dl, VTs, Ops);
16687 SDValue RetOps[] = { SDValue(Res, 0), SDValue(Res, 2) };
16688 return DAG.getMergeValues(RetOps, dl);
16691 static SDValue getScatterNode(unsigned Opc, SDValue Op, SelectionDAG &DAG,
16692 SDValue Src, SDValue Mask, SDValue Base,
16693 SDValue Index, SDValue ScaleOp, SDValue Chain) {
16695 ConstantSDNode *C = dyn_cast<ConstantSDNode>(ScaleOp);
16697 llvm_unreachable("Invalid scale type");
16698 unsigned ScaleVal = C->getZExtValue();
16699 if (ScaleVal > 2 && ScaleVal != 4 && ScaleVal != 8)
16700 llvm_unreachable("Valid scale values are 1, 2, 4, 8");
16702 SDValue Scale = DAG.getTargetConstant(C->getZExtValue(), dl, MVT::i8);
16703 SDValue Disp = DAG.getTargetConstant(0, dl, MVT::i32);
16704 SDValue Segment = DAG.getRegister(0, MVT::i32);
16705 EVT MaskVT = MVT::getVectorVT(MVT::i1,
16706 Index.getSimpleValueType().getVectorNumElements());
16708 ConstantSDNode *MaskC = dyn_cast<ConstantSDNode>(Mask);
16710 MaskInReg = DAG.getTargetConstant(MaskC->getSExtValue(), dl, MaskVT);
16712 EVT BitcastVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
16713 Mask.getValueType().getSizeInBits());
16715 // In case when MaskVT equals v2i1 or v4i1, low 2 or 4 elements
16716 // are extracted by EXTRACT_SUBVECTOR.
16717 MaskInReg = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MaskVT,
16718 DAG.getBitcast(BitcastVT, Mask),
16719 DAG.getIntPtrConstant(0, dl));
16721 SDVTList VTs = DAG.getVTList(MaskVT, MVT::Other);
16722 SDValue Ops[] = {Base, Scale, Index, Disp, Segment, MaskInReg, Src, Chain};
16723 SDNode *Res = DAG.getMachineNode(Opc, dl, VTs, Ops);
16724 return SDValue(Res, 1);
16727 static SDValue getPrefetchNode(unsigned Opc, SDValue Op, SelectionDAG &DAG,
16728 SDValue Mask, SDValue Base, SDValue Index,
16729 SDValue ScaleOp, SDValue Chain) {
16731 ConstantSDNode *C = dyn_cast<ConstantSDNode>(ScaleOp);
16732 assert(C && "Invalid scale type");
16733 SDValue Scale = DAG.getTargetConstant(C->getZExtValue(), dl, MVT::i8);
16734 SDValue Disp = DAG.getTargetConstant(0, dl, MVT::i32);
16735 SDValue Segment = DAG.getRegister(0, MVT::i32);
16737 MVT::getVectorVT(MVT::i1, Index.getSimpleValueType().getVectorNumElements());
16739 ConstantSDNode *MaskC = dyn_cast<ConstantSDNode>(Mask);
16741 MaskInReg = DAG.getTargetConstant(MaskC->getSExtValue(), dl, MaskVT);
16743 MaskInReg = DAG.getBitcast(MaskVT, Mask);
16744 //SDVTList VTs = DAG.getVTList(MVT::Other);
16745 SDValue Ops[] = {MaskInReg, Base, Scale, Index, Disp, Segment, Chain};
16746 SDNode *Res = DAG.getMachineNode(Opc, dl, MVT::Other, Ops);
16747 return SDValue(Res, 0);
16750 // getReadPerformanceCounter - Handles the lowering of builtin intrinsics that
16751 // read performance monitor counters (x86_rdpmc).
16752 static void getReadPerformanceCounter(SDNode *N, SDLoc DL,
16753 SelectionDAG &DAG, const X86Subtarget *Subtarget,
16754 SmallVectorImpl<SDValue> &Results) {
16755 assert(N->getNumOperands() == 3 && "Unexpected number of operands!");
16756 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Glue);
16759 // The ECX register is used to select the index of the performance counter
16761 SDValue Chain = DAG.getCopyToReg(N->getOperand(0), DL, X86::ECX,
16763 SDValue rd = DAG.getNode(X86ISD::RDPMC_DAG, DL, Tys, Chain);
16765 // Reads the content of a 64-bit performance counter and returns it in the
16766 // registers EDX:EAX.
16767 if (Subtarget->is64Bit()) {
16768 LO = DAG.getCopyFromReg(rd, DL, X86::RAX, MVT::i64, rd.getValue(1));
16769 HI = DAG.getCopyFromReg(LO.getValue(1), DL, X86::RDX, MVT::i64,
16772 LO = DAG.getCopyFromReg(rd, DL, X86::EAX, MVT::i32, rd.getValue(1));
16773 HI = DAG.getCopyFromReg(LO.getValue(1), DL, X86::EDX, MVT::i32,
16776 Chain = HI.getValue(1);
16778 if (Subtarget->is64Bit()) {
16779 // The EAX register is loaded with the low-order 32 bits. The EDX register
16780 // is loaded with the supported high-order bits of the counter.
16781 SDValue Tmp = DAG.getNode(ISD::SHL, DL, MVT::i64, HI,
16782 DAG.getConstant(32, DL, MVT::i8));
16783 Results.push_back(DAG.getNode(ISD::OR, DL, MVT::i64, LO, Tmp));
16784 Results.push_back(Chain);
16788 // Use a buildpair to merge the two 32-bit values into a 64-bit one.
16789 SDValue Ops[] = { LO, HI };
16790 SDValue Pair = DAG.getNode(ISD::BUILD_PAIR, DL, MVT::i64, Ops);
16791 Results.push_back(Pair);
16792 Results.push_back(Chain);
16795 // getReadTimeStampCounter - Handles the lowering of builtin intrinsics that
16796 // read the time stamp counter (x86_rdtsc and x86_rdtscp). This function is
16797 // also used to custom lower READCYCLECOUNTER nodes.
16798 static void getReadTimeStampCounter(SDNode *N, SDLoc DL, unsigned Opcode,
16799 SelectionDAG &DAG, const X86Subtarget *Subtarget,
16800 SmallVectorImpl<SDValue> &Results) {
16801 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Glue);
16802 SDValue rd = DAG.getNode(Opcode, DL, Tys, N->getOperand(0));
16805 // The processor's time-stamp counter (a 64-bit MSR) is stored into the
16806 // EDX:EAX registers. EDX is loaded with the high-order 32 bits of the MSR
16807 // and the EAX register is loaded with the low-order 32 bits.
16808 if (Subtarget->is64Bit()) {
16809 LO = DAG.getCopyFromReg(rd, DL, X86::RAX, MVT::i64, rd.getValue(1));
16810 HI = DAG.getCopyFromReg(LO.getValue(1), DL, X86::RDX, MVT::i64,
16813 LO = DAG.getCopyFromReg(rd, DL, X86::EAX, MVT::i32, rd.getValue(1));
16814 HI = DAG.getCopyFromReg(LO.getValue(1), DL, X86::EDX, MVT::i32,
16817 SDValue Chain = HI.getValue(1);
16819 if (Opcode == X86ISD::RDTSCP_DAG) {
16820 assert(N->getNumOperands() == 3 && "Unexpected number of operands!");
16822 // Instruction RDTSCP loads the IA32:TSC_AUX_MSR (address C000_0103H) into
16823 // the ECX register. Add 'ecx' explicitly to the chain.
16824 SDValue ecx = DAG.getCopyFromReg(Chain, DL, X86::ECX, MVT::i32,
16826 // Explicitly store the content of ECX at the location passed in input
16827 // to the 'rdtscp' intrinsic.
16828 Chain = DAG.getStore(ecx.getValue(1), DL, ecx, N->getOperand(2),
16829 MachinePointerInfo(), false, false, 0);
16832 if (Subtarget->is64Bit()) {
16833 // The EDX register is loaded with the high-order 32 bits of the MSR, and
16834 // the EAX register is loaded with the low-order 32 bits.
16835 SDValue Tmp = DAG.getNode(ISD::SHL, DL, MVT::i64, HI,
16836 DAG.getConstant(32, DL, MVT::i8));
16837 Results.push_back(DAG.getNode(ISD::OR, DL, MVT::i64, LO, Tmp));
16838 Results.push_back(Chain);
16842 // Use a buildpair to merge the two 32-bit values into a 64-bit one.
16843 SDValue Ops[] = { LO, HI };
16844 SDValue Pair = DAG.getNode(ISD::BUILD_PAIR, DL, MVT::i64, Ops);
16845 Results.push_back(Pair);
16846 Results.push_back(Chain);
16849 static SDValue LowerREADCYCLECOUNTER(SDValue Op, const X86Subtarget *Subtarget,
16850 SelectionDAG &DAG) {
16851 SmallVector<SDValue, 2> Results;
16853 getReadTimeStampCounter(Op.getNode(), DL, X86ISD::RDTSC_DAG, DAG, Subtarget,
16855 return DAG.getMergeValues(Results, DL);
16858 static SDValue LowerSEHRESTOREFRAME(SDValue Op, const X86Subtarget *Subtarget,
16859 SelectionDAG &DAG) {
16860 MachineFunction &MF = DAG.getMachineFunction();
16861 const Function *Fn = MF.getFunction();
16863 SDValue Chain = Op.getOperand(0);
16865 assert(Subtarget->getFrameLowering()->hasFP(MF) &&
16866 "using llvm.x86.seh.restoreframe requires a frame pointer");
16868 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
16869 MVT VT = TLI.getPointerTy(DAG.getDataLayout());
16871 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
16872 unsigned FrameReg =
16873 RegInfo->getPtrSizedFrameRegister(DAG.getMachineFunction());
16874 unsigned SPReg = RegInfo->getStackRegister();
16875 unsigned SlotSize = RegInfo->getSlotSize();
16877 // Get incoming EBP.
16878 SDValue IncomingEBP =
16879 DAG.getCopyFromReg(Chain, dl, FrameReg, VT);
16881 // SP is saved in the first field of every registration node, so load
16882 // [EBP-RegNodeSize] into SP.
16883 int RegNodeSize = getSEHRegistrationNodeSize(Fn);
16884 SDValue SPAddr = DAG.getNode(ISD::ADD, dl, VT, IncomingEBP,
16885 DAG.getConstant(-RegNodeSize, dl, VT));
16887 DAG.getLoad(VT, dl, Chain, SPAddr, MachinePointerInfo(), false, false,
16888 false, VT.getScalarSizeInBits() / 8);
16889 Chain = DAG.getCopyToReg(Chain, dl, SPReg, NewSP);
16891 if (!RegInfo->needsStackRealignment(MF)) {
16892 // Adjust EBP to point back to the original frame position.
16893 SDValue NewFP = recoverFramePointer(DAG, Fn, IncomingEBP);
16894 Chain = DAG.getCopyToReg(Chain, dl, FrameReg, NewFP);
16896 assert(RegInfo->hasBasePointer(MF) &&
16897 "functions with Win32 EH must use frame or base pointer register");
16899 // Reload the base pointer (ESI) with the adjusted incoming EBP.
16900 SDValue NewBP = recoverFramePointer(DAG, Fn, IncomingEBP);
16901 Chain = DAG.getCopyToReg(Chain, dl, RegInfo->getBaseRegister(), NewBP);
16903 // Reload the spilled EBP value, now that the stack and base pointers are
16905 X86MachineFunctionInfo *X86FI = MF.getInfo<X86MachineFunctionInfo>();
16906 X86FI->setHasSEHFramePtrSave(true);
16907 int FI = MF.getFrameInfo()->CreateSpillStackObject(SlotSize, SlotSize);
16908 X86FI->setSEHFramePtrSaveIndex(FI);
16909 SDValue NewFP = DAG.getLoad(VT, dl, Chain, DAG.getFrameIndex(FI, VT),
16910 MachinePointerInfo(), false, false, false,
16911 VT.getScalarSizeInBits() / 8);
16912 Chain = DAG.getCopyToReg(NewFP, dl, FrameReg, NewFP);
16918 /// \brief Lower intrinsics for TRUNCATE_TO_MEM case
16919 /// return truncate Store/MaskedStore Node
16920 static SDValue LowerINTRINSIC_TRUNCATE_TO_MEM(const SDValue & Op,
16924 SDValue Mask = Op.getOperand(4);
16925 SDValue DataToTruncate = Op.getOperand(3);
16926 SDValue Addr = Op.getOperand(2);
16927 SDValue Chain = Op.getOperand(0);
16929 EVT VT = DataToTruncate.getValueType();
16930 EVT SVT = EVT::getVectorVT(*DAG.getContext(),
16931 ElementType, VT.getVectorNumElements());
16933 if (isAllOnes(Mask)) // return just a truncate store
16934 return DAG.getTruncStore(Chain, dl, DataToTruncate, Addr,
16935 MachinePointerInfo(), SVT, false, false,
16936 SVT.getScalarSizeInBits()/8);
16938 EVT MaskVT = EVT::getVectorVT(*DAG.getContext(),
16939 MVT::i1, VT.getVectorNumElements());
16940 EVT BitcastVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
16941 Mask.getValueType().getSizeInBits());
16942 // In case when MaskVT equals v2i1 or v4i1, low 2 or 4 elements
16943 // are extracted by EXTRACT_SUBVECTOR.
16944 SDValue VMask = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MaskVT,
16945 DAG.getBitcast(BitcastVT, Mask),
16946 DAG.getIntPtrConstant(0, dl));
16948 MachineMemOperand *MMO = DAG.getMachineFunction().
16949 getMachineMemOperand(MachinePointerInfo(),
16950 MachineMemOperand::MOStore, SVT.getStoreSize(),
16951 SVT.getScalarSizeInBits()/8);
16953 return DAG.getMaskedStore(Chain, dl, DataToTruncate, Addr,
16954 VMask, SVT, MMO, true);
16957 static SDValue LowerINTRINSIC_W_CHAIN(SDValue Op, const X86Subtarget *Subtarget,
16958 SelectionDAG &DAG) {
16959 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
16961 const IntrinsicData* IntrData = getIntrinsicWithChain(IntNo);
16963 if (IntNo == llvm::Intrinsic::x86_seh_restoreframe)
16964 return LowerSEHRESTOREFRAME(Op, Subtarget, DAG);
16969 switch(IntrData->Type) {
16971 llvm_unreachable("Unknown Intrinsic Type");
16975 // Emit the node with the right value type.
16976 SDVTList VTs = DAG.getVTList(Op->getValueType(0), MVT::Glue, MVT::Other);
16977 SDValue Result = DAG.getNode(IntrData->Opc0, dl, VTs, Op.getOperand(0));
16979 // If the value returned by RDRAND/RDSEED was valid (CF=1), return 1.
16980 // Otherwise return the value from Rand, which is always 0, casted to i32.
16981 SDValue Ops[] = { DAG.getZExtOrTrunc(Result, dl, Op->getValueType(1)),
16982 DAG.getConstant(1, dl, Op->getValueType(1)),
16983 DAG.getConstant(X86::COND_B, dl, MVT::i32),
16984 SDValue(Result.getNode(), 1) };
16985 SDValue isValid = DAG.getNode(X86ISD::CMOV, dl,
16986 DAG.getVTList(Op->getValueType(1), MVT::Glue),
16989 // Return { result, isValid, chain }.
16990 return DAG.getNode(ISD::MERGE_VALUES, dl, Op->getVTList(), Result, isValid,
16991 SDValue(Result.getNode(), 2));
16994 //gather(v1, mask, index, base, scale);
16995 SDValue Chain = Op.getOperand(0);
16996 SDValue Src = Op.getOperand(2);
16997 SDValue Base = Op.getOperand(3);
16998 SDValue Index = Op.getOperand(4);
16999 SDValue Mask = Op.getOperand(5);
17000 SDValue Scale = Op.getOperand(6);
17001 return getGatherNode(IntrData->Opc0, Op, DAG, Src, Mask, Base, Index, Scale,
17005 //scatter(base, mask, index, v1, scale);
17006 SDValue Chain = Op.getOperand(0);
17007 SDValue Base = Op.getOperand(2);
17008 SDValue Mask = Op.getOperand(3);
17009 SDValue Index = Op.getOperand(4);
17010 SDValue Src = Op.getOperand(5);
17011 SDValue Scale = Op.getOperand(6);
17012 return getScatterNode(IntrData->Opc0, Op, DAG, Src, Mask, Base, Index,
17016 SDValue Hint = Op.getOperand(6);
17017 unsigned HintVal = cast<ConstantSDNode>(Hint)->getZExtValue();
17018 assert(HintVal < 2 && "Wrong prefetch hint in intrinsic: should be 0 or 1");
17019 unsigned Opcode = (HintVal ? IntrData->Opc1 : IntrData->Opc0);
17020 SDValue Chain = Op.getOperand(0);
17021 SDValue Mask = Op.getOperand(2);
17022 SDValue Index = Op.getOperand(3);
17023 SDValue Base = Op.getOperand(4);
17024 SDValue Scale = Op.getOperand(5);
17025 return getPrefetchNode(Opcode, Op, DAG, Mask, Base, Index, Scale, Chain);
17027 // Read Time Stamp Counter (RDTSC) and Processor ID (RDTSCP).
17029 SmallVector<SDValue, 2> Results;
17030 getReadTimeStampCounter(Op.getNode(), dl, IntrData->Opc0, DAG, Subtarget,
17032 return DAG.getMergeValues(Results, dl);
17034 // Read Performance Monitoring Counters.
17036 SmallVector<SDValue, 2> Results;
17037 getReadPerformanceCounter(Op.getNode(), dl, DAG, Subtarget, Results);
17038 return DAG.getMergeValues(Results, dl);
17040 // XTEST intrinsics.
17042 SDVTList VTs = DAG.getVTList(Op->getValueType(0), MVT::Other);
17043 SDValue InTrans = DAG.getNode(IntrData->Opc0, dl, VTs, Op.getOperand(0));
17044 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
17045 DAG.getConstant(X86::COND_NE, dl, MVT::i8),
17047 SDValue Ret = DAG.getNode(ISD::ZERO_EXTEND, dl, Op->getValueType(0), SetCC);
17048 return DAG.getNode(ISD::MERGE_VALUES, dl, Op->getVTList(),
17049 Ret, SDValue(InTrans.getNode(), 1));
17053 SmallVector<SDValue, 2> Results;
17054 SDVTList CFVTs = DAG.getVTList(Op->getValueType(0), MVT::Other);
17055 SDVTList VTs = DAG.getVTList(Op.getOperand(3)->getValueType(0), MVT::Other);
17056 SDValue GenCF = DAG.getNode(X86ISD::ADD, dl, CFVTs, Op.getOperand(2),
17057 DAG.getConstant(-1, dl, MVT::i8));
17058 SDValue Res = DAG.getNode(IntrData->Opc0, dl, VTs, Op.getOperand(3),
17059 Op.getOperand(4), GenCF.getValue(1));
17060 SDValue Store = DAG.getStore(Op.getOperand(0), dl, Res.getValue(0),
17061 Op.getOperand(5), MachinePointerInfo(),
17063 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
17064 DAG.getConstant(X86::COND_B, dl, MVT::i8),
17066 Results.push_back(SetCC);
17067 Results.push_back(Store);
17068 return DAG.getMergeValues(Results, dl);
17070 case COMPRESS_TO_MEM: {
17072 SDValue Mask = Op.getOperand(4);
17073 SDValue DataToCompress = Op.getOperand(3);
17074 SDValue Addr = Op.getOperand(2);
17075 SDValue Chain = Op.getOperand(0);
17077 EVT VT = DataToCompress.getValueType();
17078 if (isAllOnes(Mask)) // return just a store
17079 return DAG.getStore(Chain, dl, DataToCompress, Addr,
17080 MachinePointerInfo(), false, false,
17081 VT.getScalarSizeInBits()/8);
17083 SDValue Compressed =
17084 getVectorMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT, DataToCompress),
17085 Mask, DAG.getUNDEF(VT), Subtarget, DAG);
17086 return DAG.getStore(Chain, dl, Compressed, Addr,
17087 MachinePointerInfo(), false, false,
17088 VT.getScalarSizeInBits()/8);
17090 case TRUNCATE_TO_MEM_VI8:
17091 return LowerINTRINSIC_TRUNCATE_TO_MEM(Op, DAG, MVT::i8);
17092 case TRUNCATE_TO_MEM_VI16:
17093 return LowerINTRINSIC_TRUNCATE_TO_MEM(Op, DAG, MVT::i16);
17094 case TRUNCATE_TO_MEM_VI32:
17095 return LowerINTRINSIC_TRUNCATE_TO_MEM(Op, DAG, MVT::i32);
17096 case EXPAND_FROM_MEM: {
17098 SDValue Mask = Op.getOperand(4);
17099 SDValue PassThru = Op.getOperand(3);
17100 SDValue Addr = Op.getOperand(2);
17101 SDValue Chain = Op.getOperand(0);
17102 EVT VT = Op.getValueType();
17104 if (isAllOnes(Mask)) // return just a load
17105 return DAG.getLoad(VT, dl, Chain, Addr, MachinePointerInfo(), false, false,
17106 false, VT.getScalarSizeInBits()/8);
17108 SDValue DataToExpand = DAG.getLoad(VT, dl, Chain, Addr, MachinePointerInfo(),
17109 false, false, false,
17110 VT.getScalarSizeInBits()/8);
17112 SDValue Results[] = {
17113 getVectorMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT, DataToExpand),
17114 Mask, PassThru, Subtarget, DAG), Chain};
17115 return DAG.getMergeValues(Results, dl);
17120 SDValue X86TargetLowering::LowerRETURNADDR(SDValue Op,
17121 SelectionDAG &DAG) const {
17122 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
17123 MFI->setReturnAddressIsTaken(true);
17125 if (verifyReturnAddressArgumentIsConstant(Op, DAG))
17128 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
17130 EVT PtrVT = getPointerTy(DAG.getDataLayout());
17133 SDValue FrameAddr = LowerFRAMEADDR(Op, DAG);
17134 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
17135 SDValue Offset = DAG.getConstant(RegInfo->getSlotSize(), dl, PtrVT);
17136 return DAG.getLoad(PtrVT, dl, DAG.getEntryNode(),
17137 DAG.getNode(ISD::ADD, dl, PtrVT,
17138 FrameAddr, Offset),
17139 MachinePointerInfo(), false, false, false, 0);
17142 // Just load the return address.
17143 SDValue RetAddrFI = getReturnAddressFrameIndex(DAG);
17144 return DAG.getLoad(PtrVT, dl, DAG.getEntryNode(),
17145 RetAddrFI, MachinePointerInfo(), false, false, false, 0);
17148 SDValue X86TargetLowering::LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const {
17149 MachineFunction &MF = DAG.getMachineFunction();
17150 MachineFrameInfo *MFI = MF.getFrameInfo();
17151 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
17152 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
17153 EVT VT = Op.getValueType();
17155 MFI->setFrameAddressIsTaken(true);
17157 if (MF.getTarget().getMCAsmInfo()->usesWindowsCFI()) {
17158 // Depth > 0 makes no sense on targets which use Windows unwind codes. It
17159 // is not possible to crawl up the stack without looking at the unwind codes
17161 int FrameAddrIndex = FuncInfo->getFAIndex();
17162 if (!FrameAddrIndex) {
17163 // Set up a frame object for the return address.
17164 unsigned SlotSize = RegInfo->getSlotSize();
17165 FrameAddrIndex = MF.getFrameInfo()->CreateFixedObject(
17166 SlotSize, /*Offset=*/0, /*IsImmutable=*/false);
17167 FuncInfo->setFAIndex(FrameAddrIndex);
17169 return DAG.getFrameIndex(FrameAddrIndex, VT);
17172 unsigned FrameReg =
17173 RegInfo->getPtrSizedFrameRegister(DAG.getMachineFunction());
17174 SDLoc dl(Op); // FIXME probably not meaningful
17175 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
17176 assert(((FrameReg == X86::RBP && VT == MVT::i64) ||
17177 (FrameReg == X86::EBP && VT == MVT::i32)) &&
17178 "Invalid Frame Register!");
17179 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, VT);
17181 FrameAddr = DAG.getLoad(VT, dl, DAG.getEntryNode(), FrameAddr,
17182 MachinePointerInfo(),
17183 false, false, false, 0);
17187 // FIXME? Maybe this could be a TableGen attribute on some registers and
17188 // this table could be generated automatically from RegInfo.
17189 unsigned X86TargetLowering::getRegisterByName(const char* RegName, EVT VT,
17190 SelectionDAG &DAG) const {
17191 const TargetFrameLowering &TFI = *Subtarget->getFrameLowering();
17192 const MachineFunction &MF = DAG.getMachineFunction();
17194 unsigned Reg = StringSwitch<unsigned>(RegName)
17195 .Case("esp", X86::ESP)
17196 .Case("rsp", X86::RSP)
17197 .Case("ebp", X86::EBP)
17198 .Case("rbp", X86::RBP)
17201 if (Reg == X86::EBP || Reg == X86::RBP) {
17202 if (!TFI.hasFP(MF))
17203 report_fatal_error("register " + StringRef(RegName) +
17204 " is allocatable: function has no frame pointer");
17207 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
17208 unsigned FrameReg =
17209 RegInfo->getPtrSizedFrameRegister(DAG.getMachineFunction());
17210 assert((FrameReg == X86::EBP || FrameReg == X86::RBP) &&
17211 "Invalid Frame Register!");
17219 report_fatal_error("Invalid register name global variable");
17222 SDValue X86TargetLowering::LowerFRAME_TO_ARGS_OFFSET(SDValue Op,
17223 SelectionDAG &DAG) const {
17224 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
17225 return DAG.getIntPtrConstant(2 * RegInfo->getSlotSize(), SDLoc(Op));
17228 SDValue X86TargetLowering::LowerEH_RETURN(SDValue Op, SelectionDAG &DAG) const {
17229 SDValue Chain = Op.getOperand(0);
17230 SDValue Offset = Op.getOperand(1);
17231 SDValue Handler = Op.getOperand(2);
17234 EVT PtrVT = getPointerTy(DAG.getDataLayout());
17235 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
17236 unsigned FrameReg = RegInfo->getFrameRegister(DAG.getMachineFunction());
17237 assert(((FrameReg == X86::RBP && PtrVT == MVT::i64) ||
17238 (FrameReg == X86::EBP && PtrVT == MVT::i32)) &&
17239 "Invalid Frame Register!");
17240 SDValue Frame = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, PtrVT);
17241 unsigned StoreAddrReg = (PtrVT == MVT::i64) ? X86::RCX : X86::ECX;
17243 SDValue StoreAddr = DAG.getNode(ISD::ADD, dl, PtrVT, Frame,
17244 DAG.getIntPtrConstant(RegInfo->getSlotSize(),
17246 StoreAddr = DAG.getNode(ISD::ADD, dl, PtrVT, StoreAddr, Offset);
17247 Chain = DAG.getStore(Chain, dl, Handler, StoreAddr, MachinePointerInfo(),
17249 Chain = DAG.getCopyToReg(Chain, dl, StoreAddrReg, StoreAddr);
17251 return DAG.getNode(X86ISD::EH_RETURN, dl, MVT::Other, Chain,
17252 DAG.getRegister(StoreAddrReg, PtrVT));
17255 SDValue X86TargetLowering::lowerEH_SJLJ_SETJMP(SDValue Op,
17256 SelectionDAG &DAG) const {
17258 return DAG.getNode(X86ISD::EH_SJLJ_SETJMP, DL,
17259 DAG.getVTList(MVT::i32, MVT::Other),
17260 Op.getOperand(0), Op.getOperand(1));
17263 SDValue X86TargetLowering::lowerEH_SJLJ_LONGJMP(SDValue Op,
17264 SelectionDAG &DAG) const {
17266 return DAG.getNode(X86ISD::EH_SJLJ_LONGJMP, DL, MVT::Other,
17267 Op.getOperand(0), Op.getOperand(1));
17270 static SDValue LowerADJUST_TRAMPOLINE(SDValue Op, SelectionDAG &DAG) {
17271 return Op.getOperand(0);
17274 SDValue X86TargetLowering::LowerINIT_TRAMPOLINE(SDValue Op,
17275 SelectionDAG &DAG) const {
17276 SDValue Root = Op.getOperand(0);
17277 SDValue Trmp = Op.getOperand(1); // trampoline
17278 SDValue FPtr = Op.getOperand(2); // nested function
17279 SDValue Nest = Op.getOperand(3); // 'nest' parameter value
17282 const Value *TrmpAddr = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
17283 const TargetRegisterInfo *TRI = Subtarget->getRegisterInfo();
17285 if (Subtarget->is64Bit()) {
17286 SDValue OutChains[6];
17288 // Large code-model.
17289 const unsigned char JMP64r = 0xFF; // 64-bit jmp through register opcode.
17290 const unsigned char MOV64ri = 0xB8; // X86::MOV64ri opcode.
17292 const unsigned char N86R10 = TRI->getEncodingValue(X86::R10) & 0x7;
17293 const unsigned char N86R11 = TRI->getEncodingValue(X86::R11) & 0x7;
17295 const unsigned char REX_WB = 0x40 | 0x08 | 0x01; // REX prefix
17297 // Load the pointer to the nested function into R11.
17298 unsigned OpCode = ((MOV64ri | N86R11) << 8) | REX_WB; // movabsq r11
17299 SDValue Addr = Trmp;
17300 OutChains[0] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, dl, MVT::i16),
17301 Addr, MachinePointerInfo(TrmpAddr),
17304 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
17305 DAG.getConstant(2, dl, MVT::i64));
17306 OutChains[1] = DAG.getStore(Root, dl, FPtr, Addr,
17307 MachinePointerInfo(TrmpAddr, 2),
17310 // Load the 'nest' parameter value into R10.
17311 // R10 is specified in X86CallingConv.td
17312 OpCode = ((MOV64ri | N86R10) << 8) | REX_WB; // movabsq r10
17313 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
17314 DAG.getConstant(10, dl, MVT::i64));
17315 OutChains[2] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, dl, MVT::i16),
17316 Addr, MachinePointerInfo(TrmpAddr, 10),
17319 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
17320 DAG.getConstant(12, dl, MVT::i64));
17321 OutChains[3] = DAG.getStore(Root, dl, Nest, Addr,
17322 MachinePointerInfo(TrmpAddr, 12),
17325 // Jump to the nested function.
17326 OpCode = (JMP64r << 8) | REX_WB; // jmpq *...
17327 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
17328 DAG.getConstant(20, dl, MVT::i64));
17329 OutChains[4] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, dl, MVT::i16),
17330 Addr, MachinePointerInfo(TrmpAddr, 20),
17333 unsigned char ModRM = N86R11 | (4 << 3) | (3 << 6); // ...r11
17334 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
17335 DAG.getConstant(22, dl, MVT::i64));
17336 OutChains[5] = DAG.getStore(Root, dl, DAG.getConstant(ModRM, dl, MVT::i8),
17337 Addr, MachinePointerInfo(TrmpAddr, 22),
17340 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other, OutChains);
17342 const Function *Func =
17343 cast<Function>(cast<SrcValueSDNode>(Op.getOperand(5))->getValue());
17344 CallingConv::ID CC = Func->getCallingConv();
17349 llvm_unreachable("Unsupported calling convention");
17350 case CallingConv::C:
17351 case CallingConv::X86_StdCall: {
17352 // Pass 'nest' parameter in ECX.
17353 // Must be kept in sync with X86CallingConv.td
17354 NestReg = X86::ECX;
17356 // Check that ECX wasn't needed by an 'inreg' parameter.
17357 FunctionType *FTy = Func->getFunctionType();
17358 const AttributeSet &Attrs = Func->getAttributes();
17360 if (!Attrs.isEmpty() && !Func->isVarArg()) {
17361 unsigned InRegCount = 0;
17364 for (FunctionType::param_iterator I = FTy->param_begin(),
17365 E = FTy->param_end(); I != E; ++I, ++Idx)
17366 if (Attrs.hasAttribute(Idx, Attribute::InReg)) {
17367 auto &DL = DAG.getDataLayout();
17368 // FIXME: should only count parameters that are lowered to integers.
17369 InRegCount += (DL.getTypeSizeInBits(*I) + 31) / 32;
17372 if (InRegCount > 2) {
17373 report_fatal_error("Nest register in use - reduce number of inreg"
17379 case CallingConv::X86_FastCall:
17380 case CallingConv::X86_ThisCall:
17381 case CallingConv::Fast:
17382 // Pass 'nest' parameter in EAX.
17383 // Must be kept in sync with X86CallingConv.td
17384 NestReg = X86::EAX;
17388 SDValue OutChains[4];
17389 SDValue Addr, Disp;
17391 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
17392 DAG.getConstant(10, dl, MVT::i32));
17393 Disp = DAG.getNode(ISD::SUB, dl, MVT::i32, FPtr, Addr);
17395 // This is storing the opcode for MOV32ri.
17396 const unsigned char MOV32ri = 0xB8; // X86::MOV32ri's opcode byte.
17397 const unsigned char N86Reg = TRI->getEncodingValue(NestReg) & 0x7;
17398 OutChains[0] = DAG.getStore(Root, dl,
17399 DAG.getConstant(MOV32ri|N86Reg, dl, MVT::i8),
17400 Trmp, MachinePointerInfo(TrmpAddr),
17403 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
17404 DAG.getConstant(1, dl, MVT::i32));
17405 OutChains[1] = DAG.getStore(Root, dl, Nest, Addr,
17406 MachinePointerInfo(TrmpAddr, 1),
17409 const unsigned char JMP = 0xE9; // jmp <32bit dst> opcode.
17410 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
17411 DAG.getConstant(5, dl, MVT::i32));
17412 OutChains[2] = DAG.getStore(Root, dl, DAG.getConstant(JMP, dl, MVT::i8),
17413 Addr, MachinePointerInfo(TrmpAddr, 5),
17416 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
17417 DAG.getConstant(6, dl, MVT::i32));
17418 OutChains[3] = DAG.getStore(Root, dl, Disp, Addr,
17419 MachinePointerInfo(TrmpAddr, 6),
17422 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other, OutChains);
17426 SDValue X86TargetLowering::LowerFLT_ROUNDS_(SDValue Op,
17427 SelectionDAG &DAG) const {
17429 The rounding mode is in bits 11:10 of FPSR, and has the following
17431 00 Round to nearest
17436 FLT_ROUNDS, on the other hand, expects the following:
17443 To perform the conversion, we do:
17444 (((((FPSR & 0x800) >> 11) | ((FPSR & 0x400) >> 9)) + 1) & 3)
17447 MachineFunction &MF = DAG.getMachineFunction();
17448 const TargetFrameLowering &TFI = *Subtarget->getFrameLowering();
17449 unsigned StackAlignment = TFI.getStackAlignment();
17450 MVT VT = Op.getSimpleValueType();
17453 // Save FP Control Word to stack slot
17454 int SSFI = MF.getFrameInfo()->CreateStackObject(2, StackAlignment, false);
17455 SDValue StackSlot =
17456 DAG.getFrameIndex(SSFI, getPointerTy(DAG.getDataLayout()));
17458 MachineMemOperand *MMO =
17459 MF.getMachineMemOperand(MachinePointerInfo::getFixedStack(MF, SSFI),
17460 MachineMemOperand::MOStore, 2, 2);
17462 SDValue Ops[] = { DAG.getEntryNode(), StackSlot };
17463 SDValue Chain = DAG.getMemIntrinsicNode(X86ISD::FNSTCW16m, DL,
17464 DAG.getVTList(MVT::Other),
17465 Ops, MVT::i16, MMO);
17467 // Load FP Control Word from stack slot
17468 SDValue CWD = DAG.getLoad(MVT::i16, DL, Chain, StackSlot,
17469 MachinePointerInfo(), false, false, false, 0);
17471 // Transform as necessary
17473 DAG.getNode(ISD::SRL, DL, MVT::i16,
17474 DAG.getNode(ISD::AND, DL, MVT::i16,
17475 CWD, DAG.getConstant(0x800, DL, MVT::i16)),
17476 DAG.getConstant(11, DL, MVT::i8));
17478 DAG.getNode(ISD::SRL, DL, MVT::i16,
17479 DAG.getNode(ISD::AND, DL, MVT::i16,
17480 CWD, DAG.getConstant(0x400, DL, MVT::i16)),
17481 DAG.getConstant(9, DL, MVT::i8));
17484 DAG.getNode(ISD::AND, DL, MVT::i16,
17485 DAG.getNode(ISD::ADD, DL, MVT::i16,
17486 DAG.getNode(ISD::OR, DL, MVT::i16, CWD1, CWD2),
17487 DAG.getConstant(1, DL, MVT::i16)),
17488 DAG.getConstant(3, DL, MVT::i16));
17490 return DAG.getNode((VT.getSizeInBits() < 16 ?
17491 ISD::TRUNCATE : ISD::ZERO_EXTEND), DL, VT, RetVal);
17494 /// \brief Lower a vector CTLZ using native supported vector CTLZ instruction.
17496 // 1. i32/i64 128/256-bit vector (native support require VLX) are expended
17497 // to 512-bit vector.
17498 // 2. i8/i16 vector implemented using dword LZCNT vector instruction
17499 // ( sub(trunc(lzcnt(zext32(x)))) ). In case zext32(x) is illegal,
17500 // split the vector, perform operation on it's Lo a Hi part and
17501 // concatenate the results.
17502 static SDValue LowerVectorCTLZ_AVX512(SDValue Op, SelectionDAG &DAG) {
17504 MVT VT = Op.getSimpleValueType();
17505 MVT EltVT = VT.getVectorElementType();
17506 unsigned NumElems = VT.getVectorNumElements();
17508 if (EltVT == MVT::i64 || EltVT == MVT::i32) {
17509 // Extend to 512 bit vector.
17510 assert((VT.is256BitVector() || VT.is128BitVector()) &&
17511 "Unsupported value type for operation");
17513 MVT NewVT = MVT::getVectorVT(EltVT, 512 / VT.getScalarSizeInBits());
17514 SDValue Vec512 = DAG.getNode(ISD::INSERT_SUBVECTOR, dl, NewVT,
17515 DAG.getUNDEF(NewVT),
17517 DAG.getIntPtrConstant(0, dl));
17518 SDValue CtlzNode = DAG.getNode(ISD::CTLZ, dl, NewVT, Vec512);
17520 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT, CtlzNode,
17521 DAG.getIntPtrConstant(0, dl));
17524 assert((EltVT == MVT::i8 || EltVT == MVT::i16) &&
17525 "Unsupported element type");
17527 if (16 < NumElems) {
17528 // Split vector, it's Lo and Hi parts will be handled in next iteration.
17530 std::tie(Lo, Hi) = DAG.SplitVector(Op.getOperand(0), dl);
17531 MVT OutVT = MVT::getVectorVT(EltVT, NumElems/2);
17533 Lo = DAG.getNode(Op.getOpcode(), dl, OutVT, Lo);
17534 Hi = DAG.getNode(Op.getOpcode(), dl, OutVT, Hi);
17536 return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT, Lo, Hi);
17539 MVT NewVT = MVT::getVectorVT(MVT::i32, NumElems);
17541 assert((NewVT.is256BitVector() || NewVT.is512BitVector()) &&
17542 "Unsupported value type for operation");
17544 // Use native supported vector instruction vplzcntd.
17545 Op = DAG.getNode(ISD::ZERO_EXTEND, dl, NewVT, Op.getOperand(0));
17546 SDValue CtlzNode = DAG.getNode(ISD::CTLZ, dl, NewVT, Op);
17547 SDValue TruncNode = DAG.getNode(ISD::TRUNCATE, dl, VT, CtlzNode);
17548 SDValue Delta = DAG.getConstant(32 - EltVT.getSizeInBits(), dl, VT);
17550 return DAG.getNode(ISD::SUB, dl, VT, TruncNode, Delta);
17553 static SDValue LowerCTLZ(SDValue Op, const X86Subtarget *Subtarget,
17554 SelectionDAG &DAG) {
17555 MVT VT = Op.getSimpleValueType();
17557 unsigned NumBits = VT.getSizeInBits();
17560 if (VT.isVector() && Subtarget->hasAVX512())
17561 return LowerVectorCTLZ_AVX512(Op, DAG);
17563 Op = Op.getOperand(0);
17564 if (VT == MVT::i8) {
17565 // Zero extend to i32 since there is not an i8 bsr.
17567 Op = DAG.getNode(ISD::ZERO_EXTEND, dl, OpVT, Op);
17570 // Issue a bsr (scan bits in reverse) which also sets EFLAGS.
17571 SDVTList VTs = DAG.getVTList(OpVT, MVT::i32);
17572 Op = DAG.getNode(X86ISD::BSR, dl, VTs, Op);
17574 // If src is zero (i.e. bsr sets ZF), returns NumBits.
17577 DAG.getConstant(NumBits + NumBits - 1, dl, OpVT),
17578 DAG.getConstant(X86::COND_E, dl, MVT::i8),
17581 Op = DAG.getNode(X86ISD::CMOV, dl, OpVT, Ops);
17583 // Finally xor with NumBits-1.
17584 Op = DAG.getNode(ISD::XOR, dl, OpVT, Op,
17585 DAG.getConstant(NumBits - 1, dl, OpVT));
17588 Op = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op);
17592 static SDValue LowerCTLZ_ZERO_UNDEF(SDValue Op, const X86Subtarget *Subtarget,
17593 SelectionDAG &DAG) {
17594 MVT VT = Op.getSimpleValueType();
17596 unsigned NumBits = VT.getSizeInBits();
17599 if (VT.isVector() && Subtarget->hasAVX512())
17600 return LowerVectorCTLZ_AVX512(Op, DAG);
17602 Op = Op.getOperand(0);
17603 if (VT == MVT::i8) {
17604 // Zero extend to i32 since there is not an i8 bsr.
17606 Op = DAG.getNode(ISD::ZERO_EXTEND, dl, OpVT, Op);
17609 // Issue a bsr (scan bits in reverse).
17610 SDVTList VTs = DAG.getVTList(OpVT, MVT::i32);
17611 Op = DAG.getNode(X86ISD::BSR, dl, VTs, Op);
17613 // And xor with NumBits-1.
17614 Op = DAG.getNode(ISD::XOR, dl, OpVT, Op,
17615 DAG.getConstant(NumBits - 1, dl, OpVT));
17618 Op = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op);
17622 static SDValue LowerCTTZ(SDValue Op, SelectionDAG &DAG) {
17623 MVT VT = Op.getSimpleValueType();
17624 unsigned NumBits = VT.getScalarSizeInBits();
17627 if (VT.isVector()) {
17628 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
17630 SDValue N0 = Op.getOperand(0);
17631 SDValue Zero = DAG.getConstant(0, dl, VT);
17633 // lsb(x) = (x & -x)
17634 SDValue LSB = DAG.getNode(ISD::AND, dl, VT, N0,
17635 DAG.getNode(ISD::SUB, dl, VT, Zero, N0));
17637 // cttz_undef(x) = (width - 1) - ctlz(lsb)
17638 if (Op.getOpcode() == ISD::CTTZ_ZERO_UNDEF &&
17639 TLI.isOperationLegal(ISD::CTLZ, VT)) {
17640 SDValue WidthMinusOne = DAG.getConstant(NumBits - 1, dl, VT);
17641 return DAG.getNode(ISD::SUB, dl, VT, WidthMinusOne,
17642 DAG.getNode(ISD::CTLZ, dl, VT, LSB));
17645 // cttz(x) = ctpop(lsb - 1)
17646 SDValue One = DAG.getConstant(1, dl, VT);
17647 return DAG.getNode(ISD::CTPOP, dl, VT,
17648 DAG.getNode(ISD::SUB, dl, VT, LSB, One));
17651 assert(Op.getOpcode() == ISD::CTTZ &&
17652 "Only scalar CTTZ requires custom lowering");
17654 // Issue a bsf (scan bits forward) which also sets EFLAGS.
17655 SDVTList VTs = DAG.getVTList(VT, MVT::i32);
17656 Op = DAG.getNode(X86ISD::BSF, dl, VTs, Op.getOperand(0));
17658 // If src is zero (i.e. bsf sets ZF), returns NumBits.
17661 DAG.getConstant(NumBits, dl, VT),
17662 DAG.getConstant(X86::COND_E, dl, MVT::i8),
17665 return DAG.getNode(X86ISD::CMOV, dl, VT, Ops);
17668 // Lower256IntArith - Break a 256-bit integer operation into two new 128-bit
17669 // ones, and then concatenate the result back.
17670 static SDValue Lower256IntArith(SDValue Op, SelectionDAG &DAG) {
17671 MVT VT = Op.getSimpleValueType();
17673 assert(VT.is256BitVector() && VT.isInteger() &&
17674 "Unsupported value type for operation");
17676 unsigned NumElems = VT.getVectorNumElements();
17679 // Extract the LHS vectors
17680 SDValue LHS = Op.getOperand(0);
17681 SDValue LHS1 = Extract128BitVector(LHS, 0, DAG, dl);
17682 SDValue LHS2 = Extract128BitVector(LHS, NumElems/2, DAG, dl);
17684 // Extract the RHS vectors
17685 SDValue RHS = Op.getOperand(1);
17686 SDValue RHS1 = Extract128BitVector(RHS, 0, DAG, dl);
17687 SDValue RHS2 = Extract128BitVector(RHS, NumElems/2, DAG, dl);
17689 MVT EltVT = VT.getVectorElementType();
17690 MVT NewVT = MVT::getVectorVT(EltVT, NumElems/2);
17692 return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT,
17693 DAG.getNode(Op.getOpcode(), dl, NewVT, LHS1, RHS1),
17694 DAG.getNode(Op.getOpcode(), dl, NewVT, LHS2, RHS2));
17697 static SDValue LowerADD(SDValue Op, SelectionDAG &DAG) {
17698 if (Op.getValueType() == MVT::i1)
17699 return DAG.getNode(ISD::XOR, SDLoc(Op), Op.getValueType(),
17700 Op.getOperand(0), Op.getOperand(1));
17701 assert(Op.getSimpleValueType().is256BitVector() &&
17702 Op.getSimpleValueType().isInteger() &&
17703 "Only handle AVX 256-bit vector integer operation");
17704 return Lower256IntArith(Op, DAG);
17707 static SDValue LowerSUB(SDValue Op, SelectionDAG &DAG) {
17708 if (Op.getValueType() == MVT::i1)
17709 return DAG.getNode(ISD::XOR, SDLoc(Op), Op.getValueType(),
17710 Op.getOperand(0), Op.getOperand(1));
17711 assert(Op.getSimpleValueType().is256BitVector() &&
17712 Op.getSimpleValueType().isInteger() &&
17713 "Only handle AVX 256-bit vector integer operation");
17714 return Lower256IntArith(Op, DAG);
17717 static SDValue LowerMINMAX(SDValue Op, SelectionDAG &DAG) {
17718 assert(Op.getSimpleValueType().is256BitVector() &&
17719 Op.getSimpleValueType().isInteger() &&
17720 "Only handle AVX 256-bit vector integer operation");
17721 return Lower256IntArith(Op, DAG);
17724 static SDValue LowerMUL(SDValue Op, const X86Subtarget *Subtarget,
17725 SelectionDAG &DAG) {
17727 MVT VT = Op.getSimpleValueType();
17730 return DAG.getNode(ISD::AND, dl, VT, Op.getOperand(0), Op.getOperand(1));
17732 // Decompose 256-bit ops into smaller 128-bit ops.
17733 if (VT.is256BitVector() && !Subtarget->hasInt256())
17734 return Lower256IntArith(Op, DAG);
17736 SDValue A = Op.getOperand(0);
17737 SDValue B = Op.getOperand(1);
17739 // Lower v16i8/v32i8 mul as promotion to v8i16/v16i16 vector
17740 // pairs, multiply and truncate.
17741 if (VT == MVT::v16i8 || VT == MVT::v32i8) {
17742 if (Subtarget->hasInt256()) {
17743 if (VT == MVT::v32i8) {
17744 MVT SubVT = MVT::getVectorVT(MVT::i8, VT.getVectorNumElements() / 2);
17745 SDValue Lo = DAG.getIntPtrConstant(0, dl);
17746 SDValue Hi = DAG.getIntPtrConstant(VT.getVectorNumElements() / 2, dl);
17747 SDValue ALo = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, SubVT, A, Lo);
17748 SDValue BLo = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, SubVT, B, Lo);
17749 SDValue AHi = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, SubVT, A, Hi);
17750 SDValue BHi = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, SubVT, B, Hi);
17751 return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT,
17752 DAG.getNode(ISD::MUL, dl, SubVT, ALo, BLo),
17753 DAG.getNode(ISD::MUL, dl, SubVT, AHi, BHi));
17756 MVT ExVT = MVT::getVectorVT(MVT::i16, VT.getVectorNumElements());
17757 return DAG.getNode(
17758 ISD::TRUNCATE, dl, VT,
17759 DAG.getNode(ISD::MUL, dl, ExVT,
17760 DAG.getNode(ISD::SIGN_EXTEND, dl, ExVT, A),
17761 DAG.getNode(ISD::SIGN_EXTEND, dl, ExVT, B)));
17764 assert(VT == MVT::v16i8 &&
17765 "Pre-AVX2 support only supports v16i8 multiplication");
17766 MVT ExVT = MVT::v8i16;
17768 // Extract the lo parts and sign extend to i16
17770 if (Subtarget->hasSSE41()) {
17771 ALo = DAG.getNode(X86ISD::VSEXT, dl, ExVT, A);
17772 BLo = DAG.getNode(X86ISD::VSEXT, dl, ExVT, B);
17774 const int ShufMask[] = {-1, 0, -1, 1, -1, 2, -1, 3,
17775 -1, 4, -1, 5, -1, 6, -1, 7};
17776 ALo = DAG.getVectorShuffle(VT, dl, A, A, ShufMask);
17777 BLo = DAG.getVectorShuffle(VT, dl, B, B, ShufMask);
17778 ALo = DAG.getBitcast(ExVT, ALo);
17779 BLo = DAG.getBitcast(ExVT, BLo);
17780 ALo = DAG.getNode(ISD::SRA, dl, ExVT, ALo, DAG.getConstant(8, dl, ExVT));
17781 BLo = DAG.getNode(ISD::SRA, dl, ExVT, BLo, DAG.getConstant(8, dl, ExVT));
17784 // Extract the hi parts and sign extend to i16
17786 if (Subtarget->hasSSE41()) {
17787 const int ShufMask[] = {8, 9, 10, 11, 12, 13, 14, 15,
17788 -1, -1, -1, -1, -1, -1, -1, -1};
17789 AHi = DAG.getVectorShuffle(VT, dl, A, A, ShufMask);
17790 BHi = DAG.getVectorShuffle(VT, dl, B, B, ShufMask);
17791 AHi = DAG.getNode(X86ISD::VSEXT, dl, ExVT, AHi);
17792 BHi = DAG.getNode(X86ISD::VSEXT, dl, ExVT, BHi);
17794 const int ShufMask[] = {-1, 8, -1, 9, -1, 10, -1, 11,
17795 -1, 12, -1, 13, -1, 14, -1, 15};
17796 AHi = DAG.getVectorShuffle(VT, dl, A, A, ShufMask);
17797 BHi = DAG.getVectorShuffle(VT, dl, B, B, ShufMask);
17798 AHi = DAG.getBitcast(ExVT, AHi);
17799 BHi = DAG.getBitcast(ExVT, BHi);
17800 AHi = DAG.getNode(ISD::SRA, dl, ExVT, AHi, DAG.getConstant(8, dl, ExVT));
17801 BHi = DAG.getNode(ISD::SRA, dl, ExVT, BHi, DAG.getConstant(8, dl, ExVT));
17804 // Multiply, mask the lower 8bits of the lo/hi results and pack
17805 SDValue RLo = DAG.getNode(ISD::MUL, dl, ExVT, ALo, BLo);
17806 SDValue RHi = DAG.getNode(ISD::MUL, dl, ExVT, AHi, BHi);
17807 RLo = DAG.getNode(ISD::AND, dl, ExVT, RLo, DAG.getConstant(255, dl, ExVT));
17808 RHi = DAG.getNode(ISD::AND, dl, ExVT, RHi, DAG.getConstant(255, dl, ExVT));
17809 return DAG.getNode(X86ISD::PACKUS, dl, VT, RLo, RHi);
17812 // Lower v4i32 mul as 2x shuffle, 2x pmuludq, 2x shuffle.
17813 if (VT == MVT::v4i32) {
17814 assert(Subtarget->hasSSE2() && !Subtarget->hasSSE41() &&
17815 "Should not custom lower when pmuldq is available!");
17817 // Extract the odd parts.
17818 static const int UnpackMask[] = { 1, -1, 3, -1 };
17819 SDValue Aodds = DAG.getVectorShuffle(VT, dl, A, A, UnpackMask);
17820 SDValue Bodds = DAG.getVectorShuffle(VT, dl, B, B, UnpackMask);
17822 // Multiply the even parts.
17823 SDValue Evens = DAG.getNode(X86ISD::PMULUDQ, dl, MVT::v2i64, A, B);
17824 // Now multiply odd parts.
17825 SDValue Odds = DAG.getNode(X86ISD::PMULUDQ, dl, MVT::v2i64, Aodds, Bodds);
17827 Evens = DAG.getBitcast(VT, Evens);
17828 Odds = DAG.getBitcast(VT, Odds);
17830 // Merge the two vectors back together with a shuffle. This expands into 2
17832 static const int ShufMask[] = { 0, 4, 2, 6 };
17833 return DAG.getVectorShuffle(VT, dl, Evens, Odds, ShufMask);
17836 assert((VT == MVT::v2i64 || VT == MVT::v4i64 || VT == MVT::v8i64) &&
17837 "Only know how to lower V2I64/V4I64/V8I64 multiply");
17839 // Ahi = psrlqi(a, 32);
17840 // Bhi = psrlqi(b, 32);
17842 // AloBlo = pmuludq(a, b);
17843 // AloBhi = pmuludq(a, Bhi);
17844 // AhiBlo = pmuludq(Ahi, b);
17846 // AloBhi = psllqi(AloBhi, 32);
17847 // AhiBlo = psllqi(AhiBlo, 32);
17848 // return AloBlo + AloBhi + AhiBlo;
17850 SDValue Ahi = getTargetVShiftByConstNode(X86ISD::VSRLI, dl, VT, A, 32, DAG);
17851 SDValue Bhi = getTargetVShiftByConstNode(X86ISD::VSRLI, dl, VT, B, 32, DAG);
17853 SDValue AhiBlo = Ahi;
17854 SDValue AloBhi = Bhi;
17855 // Bit cast to 32-bit vectors for MULUDQ
17856 EVT MulVT = (VT == MVT::v2i64) ? MVT::v4i32 :
17857 (VT == MVT::v4i64) ? MVT::v8i32 : MVT::v16i32;
17858 A = DAG.getBitcast(MulVT, A);
17859 B = DAG.getBitcast(MulVT, B);
17860 Ahi = DAG.getBitcast(MulVT, Ahi);
17861 Bhi = DAG.getBitcast(MulVT, Bhi);
17863 SDValue AloBlo = DAG.getNode(X86ISD::PMULUDQ, dl, VT, A, B);
17864 // After shifting right const values the result may be all-zero.
17865 if (!ISD::isBuildVectorAllZeros(Ahi.getNode())) {
17866 AhiBlo = DAG.getNode(X86ISD::PMULUDQ, dl, VT, Ahi, B);
17867 AhiBlo = getTargetVShiftByConstNode(X86ISD::VSHLI, dl, VT, AhiBlo, 32, DAG);
17869 if (!ISD::isBuildVectorAllZeros(Bhi.getNode())) {
17870 AloBhi = DAG.getNode(X86ISD::PMULUDQ, dl, VT, A, Bhi);
17871 AloBhi = getTargetVShiftByConstNode(X86ISD::VSHLI, dl, VT, AloBhi, 32, DAG);
17874 SDValue Res = DAG.getNode(ISD::ADD, dl, VT, AloBlo, AloBhi);
17875 return DAG.getNode(ISD::ADD, dl, VT, Res, AhiBlo);
17878 SDValue X86TargetLowering::LowerWin64_i128OP(SDValue Op, SelectionDAG &DAG) const {
17879 assert(Subtarget->isTargetWin64() && "Unexpected target");
17880 EVT VT = Op.getValueType();
17881 assert(VT.isInteger() && VT.getSizeInBits() == 128 &&
17882 "Unexpected return type for lowering");
17886 switch (Op->getOpcode()) {
17887 default: llvm_unreachable("Unexpected request for libcall!");
17888 case ISD::SDIV: isSigned = true; LC = RTLIB::SDIV_I128; break;
17889 case ISD::UDIV: isSigned = false; LC = RTLIB::UDIV_I128; break;
17890 case ISD::SREM: isSigned = true; LC = RTLIB::SREM_I128; break;
17891 case ISD::UREM: isSigned = false; LC = RTLIB::UREM_I128; break;
17892 case ISD::SDIVREM: isSigned = true; LC = RTLIB::SDIVREM_I128; break;
17893 case ISD::UDIVREM: isSigned = false; LC = RTLIB::UDIVREM_I128; break;
17897 SDValue InChain = DAG.getEntryNode();
17899 TargetLowering::ArgListTy Args;
17900 TargetLowering::ArgListEntry Entry;
17901 for (unsigned i = 0, e = Op->getNumOperands(); i != e; ++i) {
17902 EVT ArgVT = Op->getOperand(i).getValueType();
17903 assert(ArgVT.isInteger() && ArgVT.getSizeInBits() == 128 &&
17904 "Unexpected argument type for lowering");
17905 SDValue StackPtr = DAG.CreateStackTemporary(ArgVT, 16);
17906 Entry.Node = StackPtr;
17907 InChain = DAG.getStore(InChain, dl, Op->getOperand(i), StackPtr, MachinePointerInfo(),
17909 Type *ArgTy = ArgVT.getTypeForEVT(*DAG.getContext());
17910 Entry.Ty = PointerType::get(ArgTy,0);
17911 Entry.isSExt = false;
17912 Entry.isZExt = false;
17913 Args.push_back(Entry);
17916 SDValue Callee = DAG.getExternalSymbol(getLibcallName(LC),
17917 getPointerTy(DAG.getDataLayout()));
17919 TargetLowering::CallLoweringInfo CLI(DAG);
17920 CLI.setDebugLoc(dl).setChain(InChain)
17921 .setCallee(getLibcallCallingConv(LC),
17922 static_cast<EVT>(MVT::v2i64).getTypeForEVT(*DAG.getContext()),
17923 Callee, std::move(Args), 0)
17924 .setInRegister().setSExtResult(isSigned).setZExtResult(!isSigned);
17926 std::pair<SDValue, SDValue> CallInfo = LowerCallTo(CLI);
17927 return DAG.getBitcast(VT, CallInfo.first);
17930 static SDValue LowerMUL_LOHI(SDValue Op, const X86Subtarget *Subtarget,
17931 SelectionDAG &DAG) {
17932 SDValue Op0 = Op.getOperand(0), Op1 = Op.getOperand(1);
17933 EVT VT = Op0.getValueType();
17936 assert((VT == MVT::v4i32 && Subtarget->hasSSE2()) ||
17937 (VT == MVT::v8i32 && Subtarget->hasInt256()));
17939 // PMULxD operations multiply each even value (starting at 0) of LHS with
17940 // the related value of RHS and produce a widen result.
17941 // E.g., PMULUDQ <4 x i32> <a|b|c|d>, <4 x i32> <e|f|g|h>
17942 // => <2 x i64> <ae|cg>
17944 // In other word, to have all the results, we need to perform two PMULxD:
17945 // 1. one with the even values.
17946 // 2. one with the odd values.
17947 // To achieve #2, with need to place the odd values at an even position.
17949 // Place the odd value at an even position (basically, shift all values 1
17950 // step to the left):
17951 const int Mask[] = {1, -1, 3, -1, 5, -1, 7, -1};
17952 // <a|b|c|d> => <b|undef|d|undef>
17953 SDValue Odd0 = DAG.getVectorShuffle(VT, dl, Op0, Op0, Mask);
17954 // <e|f|g|h> => <f|undef|h|undef>
17955 SDValue Odd1 = DAG.getVectorShuffle(VT, dl, Op1, Op1, Mask);
17957 // Emit two multiplies, one for the lower 2 ints and one for the higher 2
17959 MVT MulVT = VT == MVT::v4i32 ? MVT::v2i64 : MVT::v4i64;
17960 bool IsSigned = Op->getOpcode() == ISD::SMUL_LOHI;
17962 (!IsSigned || !Subtarget->hasSSE41()) ? X86ISD::PMULUDQ : X86ISD::PMULDQ;
17963 // PMULUDQ <4 x i32> <a|b|c|d>, <4 x i32> <e|f|g|h>
17964 // => <2 x i64> <ae|cg>
17965 SDValue Mul1 = DAG.getBitcast(VT, DAG.getNode(Opcode, dl, MulVT, Op0, Op1));
17966 // PMULUDQ <4 x i32> <b|undef|d|undef>, <4 x i32> <f|undef|h|undef>
17967 // => <2 x i64> <bf|dh>
17968 SDValue Mul2 = DAG.getBitcast(VT, DAG.getNode(Opcode, dl, MulVT, Odd0, Odd1));
17970 // Shuffle it back into the right order.
17971 SDValue Highs, Lows;
17972 if (VT == MVT::v8i32) {
17973 const int HighMask[] = {1, 9, 3, 11, 5, 13, 7, 15};
17974 Highs = DAG.getVectorShuffle(VT, dl, Mul1, Mul2, HighMask);
17975 const int LowMask[] = {0, 8, 2, 10, 4, 12, 6, 14};
17976 Lows = DAG.getVectorShuffle(VT, dl, Mul1, Mul2, LowMask);
17978 const int HighMask[] = {1, 5, 3, 7};
17979 Highs = DAG.getVectorShuffle(VT, dl, Mul1, Mul2, HighMask);
17980 const int LowMask[] = {0, 4, 2, 6};
17981 Lows = DAG.getVectorShuffle(VT, dl, Mul1, Mul2, LowMask);
17984 // If we have a signed multiply but no PMULDQ fix up the high parts of a
17985 // unsigned multiply.
17986 if (IsSigned && !Subtarget->hasSSE41()) {
17987 SDValue ShAmt = DAG.getConstant(
17989 DAG.getTargetLoweringInfo().getShiftAmountTy(VT, DAG.getDataLayout()));
17990 SDValue T1 = DAG.getNode(ISD::AND, dl, VT,
17991 DAG.getNode(ISD::SRA, dl, VT, Op0, ShAmt), Op1);
17992 SDValue T2 = DAG.getNode(ISD::AND, dl, VT,
17993 DAG.getNode(ISD::SRA, dl, VT, Op1, ShAmt), Op0);
17995 SDValue Fixup = DAG.getNode(ISD::ADD, dl, VT, T1, T2);
17996 Highs = DAG.getNode(ISD::SUB, dl, VT, Highs, Fixup);
17999 // The first result of MUL_LOHI is actually the low value, followed by the
18001 SDValue Ops[] = {Lows, Highs};
18002 return DAG.getMergeValues(Ops, dl);
18005 // Return true if the required (according to Opcode) shift-imm form is natively
18006 // supported by the Subtarget
18007 static bool SupportedVectorShiftWithImm(MVT VT, const X86Subtarget *Subtarget,
18009 if (VT.getScalarSizeInBits() < 16)
18012 if (VT.is512BitVector() &&
18013 (VT.getScalarSizeInBits() > 16 || Subtarget->hasBWI()))
18016 bool LShift = VT.is128BitVector() ||
18017 (VT.is256BitVector() && Subtarget->hasInt256());
18019 bool AShift = LShift && (Subtarget->hasVLX() ||
18020 (VT != MVT::v2i64 && VT != MVT::v4i64));
18021 return (Opcode == ISD::SRA) ? AShift : LShift;
18024 // The shift amount is a variable, but it is the same for all vector lanes.
18025 // These instructions are defined together with shift-immediate.
18027 bool SupportedVectorShiftWithBaseAmnt(MVT VT, const X86Subtarget *Subtarget,
18029 return SupportedVectorShiftWithImm(VT, Subtarget, Opcode);
18032 // Return true if the required (according to Opcode) variable-shift form is
18033 // natively supported by the Subtarget
18034 static bool SupportedVectorVarShift(MVT VT, const X86Subtarget *Subtarget,
18037 if (!Subtarget->hasInt256() || VT.getScalarSizeInBits() < 16)
18040 // vXi16 supported only on AVX-512, BWI
18041 if (VT.getScalarSizeInBits() == 16 && !Subtarget->hasBWI())
18044 if (VT.is512BitVector() || Subtarget->hasVLX())
18047 bool LShift = VT.is128BitVector() || VT.is256BitVector();
18048 bool AShift = LShift && VT != MVT::v2i64 && VT != MVT::v4i64;
18049 return (Opcode == ISD::SRA) ? AShift : LShift;
18052 static SDValue LowerScalarImmediateShift(SDValue Op, SelectionDAG &DAG,
18053 const X86Subtarget *Subtarget) {
18054 MVT VT = Op.getSimpleValueType();
18056 SDValue R = Op.getOperand(0);
18057 SDValue Amt = Op.getOperand(1);
18059 unsigned X86Opc = (Op.getOpcode() == ISD::SHL) ? X86ISD::VSHLI :
18060 (Op.getOpcode() == ISD::SRL) ? X86ISD::VSRLI : X86ISD::VSRAI;
18062 auto ArithmeticShiftRight64 = [&](uint64_t ShiftAmt) {
18063 assert((VT == MVT::v2i64 || VT == MVT::v4i64) && "Unexpected SRA type");
18064 MVT ExVT = MVT::getVectorVT(MVT::i32, VT.getVectorNumElements() * 2);
18065 SDValue Ex = DAG.getBitcast(ExVT, R);
18067 if (ShiftAmt >= 32) {
18068 // Splat sign to upper i32 dst, and SRA upper i32 src to lower i32.
18070 getTargetVShiftByConstNode(X86ISD::VSRAI, dl, ExVT, Ex, 31, DAG);
18071 SDValue Lower = getTargetVShiftByConstNode(X86ISD::VSRAI, dl, ExVT, Ex,
18072 ShiftAmt - 32, DAG);
18073 if (VT == MVT::v2i64)
18074 Ex = DAG.getVectorShuffle(ExVT, dl, Upper, Lower, {5, 1, 7, 3});
18075 if (VT == MVT::v4i64)
18076 Ex = DAG.getVectorShuffle(ExVT, dl, Upper, Lower,
18077 {9, 1, 11, 3, 13, 5, 15, 7});
18079 // SRA upper i32, SHL whole i64 and select lower i32.
18080 SDValue Upper = getTargetVShiftByConstNode(X86ISD::VSRAI, dl, ExVT, Ex,
18083 getTargetVShiftByConstNode(X86ISD::VSRLI, dl, VT, R, ShiftAmt, DAG);
18084 Lower = DAG.getBitcast(ExVT, Lower);
18085 if (VT == MVT::v2i64)
18086 Ex = DAG.getVectorShuffle(ExVT, dl, Upper, Lower, {4, 1, 6, 3});
18087 if (VT == MVT::v4i64)
18088 Ex = DAG.getVectorShuffle(ExVT, dl, Upper, Lower,
18089 {8, 1, 10, 3, 12, 5, 14, 7});
18091 return DAG.getBitcast(VT, Ex);
18094 // Optimize shl/srl/sra with constant shift amount.
18095 if (auto *BVAmt = dyn_cast<BuildVectorSDNode>(Amt)) {
18096 if (auto *ShiftConst = BVAmt->getConstantSplatNode()) {
18097 uint64_t ShiftAmt = ShiftConst->getZExtValue();
18099 if (SupportedVectorShiftWithImm(VT, Subtarget, Op.getOpcode()))
18100 return getTargetVShiftByConstNode(X86Opc, dl, VT, R, ShiftAmt, DAG);
18102 // i64 SRA needs to be performed as partial shifts.
18103 if ((VT == MVT::v2i64 || (Subtarget->hasInt256() && VT == MVT::v4i64)) &&
18104 Op.getOpcode() == ISD::SRA && !Subtarget->hasXOP())
18105 return ArithmeticShiftRight64(ShiftAmt);
18107 if (VT == MVT::v16i8 || (Subtarget->hasInt256() && VT == MVT::v32i8)) {
18108 unsigned NumElts = VT.getVectorNumElements();
18109 MVT ShiftVT = MVT::getVectorVT(MVT::i16, NumElts / 2);
18111 // Simple i8 add case
18112 if (Op.getOpcode() == ISD::SHL && ShiftAmt == 1)
18113 return DAG.getNode(ISD::ADD, dl, VT, R, R);
18115 // ashr(R, 7) === cmp_slt(R, 0)
18116 if (Op.getOpcode() == ISD::SRA && ShiftAmt == 7) {
18117 SDValue Zeros = getZeroVector(VT, Subtarget, DAG, dl);
18118 return DAG.getNode(X86ISD::PCMPGT, dl, VT, Zeros, R);
18121 // XOP can shift v16i8 directly instead of as shift v8i16 + mask.
18122 if (VT == MVT::v16i8 && Subtarget->hasXOP())
18125 if (Op.getOpcode() == ISD::SHL) {
18126 // Make a large shift.
18127 SDValue SHL = getTargetVShiftByConstNode(X86ISD::VSHLI, dl, ShiftVT,
18129 SHL = DAG.getBitcast(VT, SHL);
18130 // Zero out the rightmost bits.
18131 SmallVector<SDValue, 32> V(
18132 NumElts, DAG.getConstant(uint8_t(-1U << ShiftAmt), dl, MVT::i8));
18133 return DAG.getNode(ISD::AND, dl, VT, SHL,
18134 DAG.getNode(ISD::BUILD_VECTOR, dl, VT, V));
18136 if (Op.getOpcode() == ISD::SRL) {
18137 // Make a large shift.
18138 SDValue SRL = getTargetVShiftByConstNode(X86ISD::VSRLI, dl, ShiftVT,
18140 SRL = DAG.getBitcast(VT, SRL);
18141 // Zero out the leftmost bits.
18142 SmallVector<SDValue, 32> V(
18143 NumElts, DAG.getConstant(uint8_t(-1U) >> ShiftAmt, dl, MVT::i8));
18144 return DAG.getNode(ISD::AND, dl, VT, SRL,
18145 DAG.getNode(ISD::BUILD_VECTOR, dl, VT, V));
18147 if (Op.getOpcode() == ISD::SRA) {
18148 // ashr(R, Amt) === sub(xor(lshr(R, Amt), Mask), Mask)
18149 SDValue Res = DAG.getNode(ISD::SRL, dl, VT, R, Amt);
18150 SmallVector<SDValue, 32> V(NumElts,
18151 DAG.getConstant(128 >> ShiftAmt, dl,
18153 SDValue Mask = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, V);
18154 Res = DAG.getNode(ISD::XOR, dl, VT, Res, Mask);
18155 Res = DAG.getNode(ISD::SUB, dl, VT, Res, Mask);
18158 llvm_unreachable("Unknown shift opcode.");
18163 // Special case in 32-bit mode, where i64 is expanded into high and low parts.
18164 if (!Subtarget->is64Bit() && !Subtarget->hasXOP() &&
18165 (VT == MVT::v2i64 || (Subtarget->hasInt256() && VT == MVT::v4i64))) {
18167 // Peek through any splat that was introduced for i64 shift vectorization.
18168 int SplatIndex = -1;
18169 if (ShuffleVectorSDNode *SVN = dyn_cast<ShuffleVectorSDNode>(Amt.getNode()))
18170 if (SVN->isSplat()) {
18171 SplatIndex = SVN->getSplatIndex();
18172 Amt = Amt.getOperand(0);
18173 assert(SplatIndex < (int)VT.getVectorNumElements() &&
18174 "Splat shuffle referencing second operand");
18177 if (Amt.getOpcode() != ISD::BITCAST ||
18178 Amt.getOperand(0).getOpcode() != ISD::BUILD_VECTOR)
18181 Amt = Amt.getOperand(0);
18182 unsigned Ratio = Amt.getSimpleValueType().getVectorNumElements() /
18183 VT.getVectorNumElements();
18184 unsigned RatioInLog2 = Log2_32_Ceil(Ratio);
18185 uint64_t ShiftAmt = 0;
18186 unsigned BaseOp = (SplatIndex < 0 ? 0 : SplatIndex * Ratio);
18187 for (unsigned i = 0; i != Ratio; ++i) {
18188 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Amt.getOperand(i + BaseOp));
18192 ShiftAmt |= C->getZExtValue() << (i * (1 << (6 - RatioInLog2)));
18195 // Check remaining shift amounts (if not a splat).
18196 if (SplatIndex < 0) {
18197 for (unsigned i = Ratio; i != Amt.getNumOperands(); i += Ratio) {
18198 uint64_t ShAmt = 0;
18199 for (unsigned j = 0; j != Ratio; ++j) {
18200 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Amt.getOperand(i + j));
18204 ShAmt |= C->getZExtValue() << (j * (1 << (6 - RatioInLog2)));
18206 if (ShAmt != ShiftAmt)
18211 if (SupportedVectorShiftWithImm(VT, Subtarget, Op.getOpcode()))
18212 return getTargetVShiftByConstNode(X86Opc, dl, VT, R, ShiftAmt, DAG);
18214 if (Op.getOpcode() == ISD::SRA)
18215 return ArithmeticShiftRight64(ShiftAmt);
18221 static SDValue LowerScalarVariableShift(SDValue Op, SelectionDAG &DAG,
18222 const X86Subtarget* Subtarget) {
18223 MVT VT = Op.getSimpleValueType();
18225 SDValue R = Op.getOperand(0);
18226 SDValue Amt = Op.getOperand(1);
18228 unsigned X86OpcI = (Op.getOpcode() == ISD::SHL) ? X86ISD::VSHLI :
18229 (Op.getOpcode() == ISD::SRL) ? X86ISD::VSRLI : X86ISD::VSRAI;
18231 unsigned X86OpcV = (Op.getOpcode() == ISD::SHL) ? X86ISD::VSHL :
18232 (Op.getOpcode() == ISD::SRL) ? X86ISD::VSRL : X86ISD::VSRA;
18234 if (SupportedVectorShiftWithBaseAmnt(VT, Subtarget, Op.getOpcode())) {
18236 EVT EltVT = VT.getVectorElementType();
18238 if (BuildVectorSDNode *BV = dyn_cast<BuildVectorSDNode>(Amt)) {
18239 // Check if this build_vector node is doing a splat.
18240 // If so, then set BaseShAmt equal to the splat value.
18241 BaseShAmt = BV->getSplatValue();
18242 if (BaseShAmt && BaseShAmt.getOpcode() == ISD::UNDEF)
18243 BaseShAmt = SDValue();
18245 if (Amt.getOpcode() == ISD::EXTRACT_SUBVECTOR)
18246 Amt = Amt.getOperand(0);
18248 ShuffleVectorSDNode *SVN = dyn_cast<ShuffleVectorSDNode>(Amt);
18249 if (SVN && SVN->isSplat()) {
18250 unsigned SplatIdx = (unsigned)SVN->getSplatIndex();
18251 SDValue InVec = Amt.getOperand(0);
18252 if (InVec.getOpcode() == ISD::BUILD_VECTOR) {
18253 assert((SplatIdx < InVec.getValueType().getVectorNumElements()) &&
18254 "Unexpected shuffle index found!");
18255 BaseShAmt = InVec.getOperand(SplatIdx);
18256 } else if (InVec.getOpcode() == ISD::INSERT_VECTOR_ELT) {
18257 if (ConstantSDNode *C =
18258 dyn_cast<ConstantSDNode>(InVec.getOperand(2))) {
18259 if (C->getZExtValue() == SplatIdx)
18260 BaseShAmt = InVec.getOperand(1);
18265 // Avoid introducing an extract element from a shuffle.
18266 BaseShAmt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, EltVT, InVec,
18267 DAG.getIntPtrConstant(SplatIdx, dl));
18271 if (BaseShAmt.getNode()) {
18272 assert(EltVT.bitsLE(MVT::i64) && "Unexpected element type!");
18273 if (EltVT != MVT::i64 && EltVT.bitsGT(MVT::i32))
18274 BaseShAmt = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i64, BaseShAmt);
18275 else if (EltVT.bitsLT(MVT::i32))
18276 BaseShAmt = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, BaseShAmt);
18278 return getTargetVShiftNode(X86OpcI, dl, VT, R, BaseShAmt, DAG);
18282 // Special case in 32-bit mode, where i64 is expanded into high and low parts.
18283 if (!Subtarget->is64Bit() && VT == MVT::v2i64 &&
18284 Amt.getOpcode() == ISD::BITCAST &&
18285 Amt.getOperand(0).getOpcode() == ISD::BUILD_VECTOR) {
18286 Amt = Amt.getOperand(0);
18287 unsigned Ratio = Amt.getSimpleValueType().getVectorNumElements() /
18288 VT.getVectorNumElements();
18289 std::vector<SDValue> Vals(Ratio);
18290 for (unsigned i = 0; i != Ratio; ++i)
18291 Vals[i] = Amt.getOperand(i);
18292 for (unsigned i = Ratio; i != Amt.getNumOperands(); i += Ratio) {
18293 for (unsigned j = 0; j != Ratio; ++j)
18294 if (Vals[j] != Amt.getOperand(i + j))
18298 if (SupportedVectorShiftWithBaseAmnt(VT, Subtarget, Op.getOpcode()))
18299 return DAG.getNode(X86OpcV, dl, VT, R, Op.getOperand(1));
18304 static SDValue LowerShift(SDValue Op, const X86Subtarget* Subtarget,
18305 SelectionDAG &DAG) {
18306 MVT VT = Op.getSimpleValueType();
18308 SDValue R = Op.getOperand(0);
18309 SDValue Amt = Op.getOperand(1);
18311 assert(VT.isVector() && "Custom lowering only for vector shifts!");
18312 assert(Subtarget->hasSSE2() && "Only custom lower when we have SSE2!");
18314 if (SDValue V = LowerScalarImmediateShift(Op, DAG, Subtarget))
18317 if (SDValue V = LowerScalarVariableShift(Op, DAG, Subtarget))
18320 if (SupportedVectorVarShift(VT, Subtarget, Op.getOpcode()))
18323 // XOP has 128-bit variable logical/arithmetic shifts.
18324 // +ve/-ve Amt = shift left/right.
18325 if (Subtarget->hasXOP() &&
18326 (VT == MVT::v2i64 || VT == MVT::v4i32 ||
18327 VT == MVT::v8i16 || VT == MVT::v16i8)) {
18328 if (Op.getOpcode() == ISD::SRL || Op.getOpcode() == ISD::SRA) {
18329 SDValue Zero = getZeroVector(VT, Subtarget, DAG, dl);
18330 Amt = DAG.getNode(ISD::SUB, dl, VT, Zero, Amt);
18332 if (Op.getOpcode() == ISD::SHL || Op.getOpcode() == ISD::SRL)
18333 return DAG.getNode(X86ISD::VPSHL, dl, VT, R, Amt);
18334 if (Op.getOpcode() == ISD::SRA)
18335 return DAG.getNode(X86ISD::VPSHA, dl, VT, R, Amt);
18338 // 2i64 vector logical shifts can efficiently avoid scalarization - do the
18339 // shifts per-lane and then shuffle the partial results back together.
18340 if (VT == MVT::v2i64 && Op.getOpcode() != ISD::SRA) {
18341 // Splat the shift amounts so the scalar shifts above will catch it.
18342 SDValue Amt0 = DAG.getVectorShuffle(VT, dl, Amt, Amt, {0, 0});
18343 SDValue Amt1 = DAG.getVectorShuffle(VT, dl, Amt, Amt, {1, 1});
18344 SDValue R0 = DAG.getNode(Op->getOpcode(), dl, VT, R, Amt0);
18345 SDValue R1 = DAG.getNode(Op->getOpcode(), dl, VT, R, Amt1);
18346 return DAG.getVectorShuffle(VT, dl, R0, R1, {0, 3});
18349 // i64 vector arithmetic shift can be emulated with the transform:
18350 // M = lshr(SIGN_BIT, Amt)
18351 // ashr(R, Amt) === sub(xor(lshr(R, Amt), M), M)
18352 if ((VT == MVT::v2i64 || (VT == MVT::v4i64 && Subtarget->hasInt256())) &&
18353 Op.getOpcode() == ISD::SRA) {
18354 SDValue S = DAG.getConstant(APInt::getSignBit(64), dl, VT);
18355 SDValue M = DAG.getNode(ISD::SRL, dl, VT, S, Amt);
18356 R = DAG.getNode(ISD::SRL, dl, VT, R, Amt);
18357 R = DAG.getNode(ISD::XOR, dl, VT, R, M);
18358 R = DAG.getNode(ISD::SUB, dl, VT, R, M);
18362 // If possible, lower this packed shift into a vector multiply instead of
18363 // expanding it into a sequence of scalar shifts.
18364 // Do this only if the vector shift count is a constant build_vector.
18365 if (Op.getOpcode() == ISD::SHL &&
18366 (VT == MVT::v8i16 || VT == MVT::v4i32 ||
18367 (Subtarget->hasInt256() && VT == MVT::v16i16)) &&
18368 ISD::isBuildVectorOfConstantSDNodes(Amt.getNode())) {
18369 SmallVector<SDValue, 8> Elts;
18370 EVT SVT = VT.getScalarType();
18371 unsigned SVTBits = SVT.getSizeInBits();
18372 const APInt &One = APInt(SVTBits, 1);
18373 unsigned NumElems = VT.getVectorNumElements();
18375 for (unsigned i=0; i !=NumElems; ++i) {
18376 SDValue Op = Amt->getOperand(i);
18377 if (Op->getOpcode() == ISD::UNDEF) {
18378 Elts.push_back(Op);
18382 ConstantSDNode *ND = cast<ConstantSDNode>(Op);
18383 const APInt &C = APInt(SVTBits, ND->getAPIntValue().getZExtValue());
18384 uint64_t ShAmt = C.getZExtValue();
18385 if (ShAmt >= SVTBits) {
18386 Elts.push_back(DAG.getUNDEF(SVT));
18389 Elts.push_back(DAG.getConstant(One.shl(ShAmt), dl, SVT));
18391 SDValue BV = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Elts);
18392 return DAG.getNode(ISD::MUL, dl, VT, R, BV);
18395 // Lower SHL with variable shift amount.
18396 if (VT == MVT::v4i32 && Op->getOpcode() == ISD::SHL) {
18397 Op = DAG.getNode(ISD::SHL, dl, VT, Amt, DAG.getConstant(23, dl, VT));
18399 Op = DAG.getNode(ISD::ADD, dl, VT, Op,
18400 DAG.getConstant(0x3f800000U, dl, VT));
18401 Op = DAG.getBitcast(MVT::v4f32, Op);
18402 Op = DAG.getNode(ISD::FP_TO_SINT, dl, VT, Op);
18403 return DAG.getNode(ISD::MUL, dl, VT, Op, R);
18406 // If possible, lower this shift as a sequence of two shifts by
18407 // constant plus a MOVSS/MOVSD instead of scalarizing it.
18409 // (v4i32 (srl A, (build_vector < X, Y, Y, Y>)))
18411 // Could be rewritten as:
18412 // (v4i32 (MOVSS (srl A, <Y,Y,Y,Y>), (srl A, <X,X,X,X>)))
18414 // The advantage is that the two shifts from the example would be
18415 // lowered as X86ISD::VSRLI nodes. This would be cheaper than scalarizing
18416 // the vector shift into four scalar shifts plus four pairs of vector
18418 if ((VT == MVT::v8i16 || VT == MVT::v4i32) &&
18419 ISD::isBuildVectorOfConstantSDNodes(Amt.getNode())) {
18420 unsigned TargetOpcode = X86ISD::MOVSS;
18421 bool CanBeSimplified;
18422 // The splat value for the first packed shift (the 'X' from the example).
18423 SDValue Amt1 = Amt->getOperand(0);
18424 // The splat value for the second packed shift (the 'Y' from the example).
18425 SDValue Amt2 = (VT == MVT::v4i32) ? Amt->getOperand(1) :
18426 Amt->getOperand(2);
18428 // See if it is possible to replace this node with a sequence of
18429 // two shifts followed by a MOVSS/MOVSD
18430 if (VT == MVT::v4i32) {
18431 // Check if it is legal to use a MOVSS.
18432 CanBeSimplified = Amt2 == Amt->getOperand(2) &&
18433 Amt2 == Amt->getOperand(3);
18434 if (!CanBeSimplified) {
18435 // Otherwise, check if we can still simplify this node using a MOVSD.
18436 CanBeSimplified = Amt1 == Amt->getOperand(1) &&
18437 Amt->getOperand(2) == Amt->getOperand(3);
18438 TargetOpcode = X86ISD::MOVSD;
18439 Amt2 = Amt->getOperand(2);
18442 // Do similar checks for the case where the machine value type
18444 CanBeSimplified = Amt1 == Amt->getOperand(1);
18445 for (unsigned i=3; i != 8 && CanBeSimplified; ++i)
18446 CanBeSimplified = Amt2 == Amt->getOperand(i);
18448 if (!CanBeSimplified) {
18449 TargetOpcode = X86ISD::MOVSD;
18450 CanBeSimplified = true;
18451 Amt2 = Amt->getOperand(4);
18452 for (unsigned i=0; i != 4 && CanBeSimplified; ++i)
18453 CanBeSimplified = Amt1 == Amt->getOperand(i);
18454 for (unsigned j=4; j != 8 && CanBeSimplified; ++j)
18455 CanBeSimplified = Amt2 == Amt->getOperand(j);
18459 if (CanBeSimplified && isa<ConstantSDNode>(Amt1) &&
18460 isa<ConstantSDNode>(Amt2)) {
18461 // Replace this node with two shifts followed by a MOVSS/MOVSD.
18462 EVT CastVT = MVT::v4i32;
18464 DAG.getConstant(cast<ConstantSDNode>(Amt1)->getAPIntValue(), dl, VT);
18465 SDValue Shift1 = DAG.getNode(Op->getOpcode(), dl, VT, R, Splat1);
18467 DAG.getConstant(cast<ConstantSDNode>(Amt2)->getAPIntValue(), dl, VT);
18468 SDValue Shift2 = DAG.getNode(Op->getOpcode(), dl, VT, R, Splat2);
18469 if (TargetOpcode == X86ISD::MOVSD)
18470 CastVT = MVT::v2i64;
18471 SDValue BitCast1 = DAG.getBitcast(CastVT, Shift1);
18472 SDValue BitCast2 = DAG.getBitcast(CastVT, Shift2);
18473 SDValue Result = getTargetShuffleNode(TargetOpcode, dl, CastVT, BitCast2,
18475 return DAG.getBitcast(VT, Result);
18479 // v4i32 Non Uniform Shifts.
18480 // If the shift amount is constant we can shift each lane using the SSE2
18481 // immediate shifts, else we need to zero-extend each lane to the lower i64
18482 // and shift using the SSE2 variable shifts.
18483 // The separate results can then be blended together.
18484 if (VT == MVT::v4i32) {
18485 unsigned Opc = Op.getOpcode();
18486 SDValue Amt0, Amt1, Amt2, Amt3;
18487 if (ISD::isBuildVectorOfConstantSDNodes(Amt.getNode())) {
18488 Amt0 = DAG.getVectorShuffle(VT, dl, Amt, DAG.getUNDEF(VT), {0, 0, 0, 0});
18489 Amt1 = DAG.getVectorShuffle(VT, dl, Amt, DAG.getUNDEF(VT), {1, 1, 1, 1});
18490 Amt2 = DAG.getVectorShuffle(VT, dl, Amt, DAG.getUNDEF(VT), {2, 2, 2, 2});
18491 Amt3 = DAG.getVectorShuffle(VT, dl, Amt, DAG.getUNDEF(VT), {3, 3, 3, 3});
18493 // ISD::SHL is handled above but we include it here for completeness.
18496 llvm_unreachable("Unknown target vector shift node");
18498 Opc = X86ISD::VSHL;
18501 Opc = X86ISD::VSRL;
18504 Opc = X86ISD::VSRA;
18507 // The SSE2 shifts use the lower i64 as the same shift amount for
18508 // all lanes and the upper i64 is ignored. These shuffle masks
18509 // optimally zero-extend each lanes on SSE2/SSE41/AVX targets.
18510 SDValue Z = getZeroVector(VT, Subtarget, DAG, dl);
18511 Amt0 = DAG.getVectorShuffle(VT, dl, Amt, Z, {0, 4, -1, -1});
18512 Amt1 = DAG.getVectorShuffle(VT, dl, Amt, Z, {1, 5, -1, -1});
18513 Amt2 = DAG.getVectorShuffle(VT, dl, Amt, Z, {2, 6, -1, -1});
18514 Amt3 = DAG.getVectorShuffle(VT, dl, Amt, Z, {3, 7, -1, -1});
18517 SDValue R0 = DAG.getNode(Opc, dl, VT, R, Amt0);
18518 SDValue R1 = DAG.getNode(Opc, dl, VT, R, Amt1);
18519 SDValue R2 = DAG.getNode(Opc, dl, VT, R, Amt2);
18520 SDValue R3 = DAG.getNode(Opc, dl, VT, R, Amt3);
18521 SDValue R02 = DAG.getVectorShuffle(VT, dl, R0, R2, {0, -1, 6, -1});
18522 SDValue R13 = DAG.getVectorShuffle(VT, dl, R1, R3, {-1, 1, -1, 7});
18523 return DAG.getVectorShuffle(VT, dl, R02, R13, {0, 5, 2, 7});
18526 if (VT == MVT::v16i8 ||
18527 (VT == MVT::v32i8 && Subtarget->hasInt256() && !Subtarget->hasXOP())) {
18528 MVT ExtVT = MVT::getVectorVT(MVT::i16, VT.getVectorNumElements() / 2);
18529 unsigned ShiftOpcode = Op->getOpcode();
18531 auto SignBitSelect = [&](MVT SelVT, SDValue Sel, SDValue V0, SDValue V1) {
18532 // On SSE41 targets we make use of the fact that VSELECT lowers
18533 // to PBLENDVB which selects bytes based just on the sign bit.
18534 if (Subtarget->hasSSE41()) {
18535 V0 = DAG.getBitcast(VT, V0);
18536 V1 = DAG.getBitcast(VT, V1);
18537 Sel = DAG.getBitcast(VT, Sel);
18538 return DAG.getBitcast(SelVT,
18539 DAG.getNode(ISD::VSELECT, dl, VT, Sel, V0, V1));
18541 // On pre-SSE41 targets we test for the sign bit by comparing to
18542 // zero - a negative value will set all bits of the lanes to true
18543 // and VSELECT uses that in its OR(AND(V0,C),AND(V1,~C)) lowering.
18544 SDValue Z = getZeroVector(SelVT, Subtarget, DAG, dl);
18545 SDValue C = DAG.getNode(X86ISD::PCMPGT, dl, SelVT, Z, Sel);
18546 return DAG.getNode(ISD::VSELECT, dl, SelVT, C, V0, V1);
18549 // Turn 'a' into a mask suitable for VSELECT: a = a << 5;
18550 // We can safely do this using i16 shifts as we're only interested in
18551 // the 3 lower bits of each byte.
18552 Amt = DAG.getBitcast(ExtVT, Amt);
18553 Amt = DAG.getNode(ISD::SHL, dl, ExtVT, Amt, DAG.getConstant(5, dl, ExtVT));
18554 Amt = DAG.getBitcast(VT, Amt);
18556 if (Op->getOpcode() == ISD::SHL || Op->getOpcode() == ISD::SRL) {
18557 // r = VSELECT(r, shift(r, 4), a);
18559 DAG.getNode(ShiftOpcode, dl, VT, R, DAG.getConstant(4, dl, VT));
18560 R = SignBitSelect(VT, Amt, M, R);
18563 Amt = DAG.getNode(ISD::ADD, dl, VT, Amt, Amt);
18565 // r = VSELECT(r, shift(r, 2), a);
18566 M = DAG.getNode(ShiftOpcode, dl, VT, R, DAG.getConstant(2, dl, VT));
18567 R = SignBitSelect(VT, Amt, M, R);
18570 Amt = DAG.getNode(ISD::ADD, dl, VT, Amt, Amt);
18572 // return VSELECT(r, shift(r, 1), a);
18573 M = DAG.getNode(ShiftOpcode, dl, VT, R, DAG.getConstant(1, dl, VT));
18574 R = SignBitSelect(VT, Amt, M, R);
18578 if (Op->getOpcode() == ISD::SRA) {
18579 // For SRA we need to unpack each byte to the higher byte of a i16 vector
18580 // so we can correctly sign extend. We don't care what happens to the
18582 SDValue ALo = DAG.getNode(X86ISD::UNPCKL, dl, VT, DAG.getUNDEF(VT), Amt);
18583 SDValue AHi = DAG.getNode(X86ISD::UNPCKH, dl, VT, DAG.getUNDEF(VT), Amt);
18584 SDValue RLo = DAG.getNode(X86ISD::UNPCKL, dl, VT, DAG.getUNDEF(VT), R);
18585 SDValue RHi = DAG.getNode(X86ISD::UNPCKH, dl, VT, DAG.getUNDEF(VT), R);
18586 ALo = DAG.getBitcast(ExtVT, ALo);
18587 AHi = DAG.getBitcast(ExtVT, AHi);
18588 RLo = DAG.getBitcast(ExtVT, RLo);
18589 RHi = DAG.getBitcast(ExtVT, RHi);
18591 // r = VSELECT(r, shift(r, 4), a);
18592 SDValue MLo = DAG.getNode(ShiftOpcode, dl, ExtVT, RLo,
18593 DAG.getConstant(4, dl, ExtVT));
18594 SDValue MHi = DAG.getNode(ShiftOpcode, dl, ExtVT, RHi,
18595 DAG.getConstant(4, dl, ExtVT));
18596 RLo = SignBitSelect(ExtVT, ALo, MLo, RLo);
18597 RHi = SignBitSelect(ExtVT, AHi, MHi, RHi);
18600 ALo = DAG.getNode(ISD::ADD, dl, ExtVT, ALo, ALo);
18601 AHi = DAG.getNode(ISD::ADD, dl, ExtVT, AHi, AHi);
18603 // r = VSELECT(r, shift(r, 2), a);
18604 MLo = DAG.getNode(ShiftOpcode, dl, ExtVT, RLo,
18605 DAG.getConstant(2, dl, ExtVT));
18606 MHi = DAG.getNode(ShiftOpcode, dl, ExtVT, RHi,
18607 DAG.getConstant(2, dl, ExtVT));
18608 RLo = SignBitSelect(ExtVT, ALo, MLo, RLo);
18609 RHi = SignBitSelect(ExtVT, AHi, MHi, RHi);
18612 ALo = DAG.getNode(ISD::ADD, dl, ExtVT, ALo, ALo);
18613 AHi = DAG.getNode(ISD::ADD, dl, ExtVT, AHi, AHi);
18615 // r = VSELECT(r, shift(r, 1), a);
18616 MLo = DAG.getNode(ShiftOpcode, dl, ExtVT, RLo,
18617 DAG.getConstant(1, dl, ExtVT));
18618 MHi = DAG.getNode(ShiftOpcode, dl, ExtVT, RHi,
18619 DAG.getConstant(1, dl, ExtVT));
18620 RLo = SignBitSelect(ExtVT, ALo, MLo, RLo);
18621 RHi = SignBitSelect(ExtVT, AHi, MHi, RHi);
18623 // Logical shift the result back to the lower byte, leaving a zero upper
18625 // meaning that we can safely pack with PACKUSWB.
18627 DAG.getNode(ISD::SRL, dl, ExtVT, RLo, DAG.getConstant(8, dl, ExtVT));
18629 DAG.getNode(ISD::SRL, dl, ExtVT, RHi, DAG.getConstant(8, dl, ExtVT));
18630 return DAG.getNode(X86ISD::PACKUS, dl, VT, RLo, RHi);
18634 // It's worth extending once and using the v8i32 shifts for 16-bit types, but
18635 // the extra overheads to get from v16i8 to v8i32 make the existing SSE
18636 // solution better.
18637 if (Subtarget->hasInt256() && VT == MVT::v8i16) {
18638 MVT ExtVT = MVT::v8i32;
18640 Op.getOpcode() == ISD::SRA ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND;
18641 R = DAG.getNode(ExtOpc, dl, ExtVT, R);
18642 Amt = DAG.getNode(ISD::ANY_EXTEND, dl, ExtVT, Amt);
18643 return DAG.getNode(ISD::TRUNCATE, dl, VT,
18644 DAG.getNode(Op.getOpcode(), dl, ExtVT, R, Amt));
18647 if (Subtarget->hasInt256() && !Subtarget->hasXOP() && VT == MVT::v16i16) {
18648 MVT ExtVT = MVT::v8i32;
18649 SDValue Z = getZeroVector(VT, Subtarget, DAG, dl);
18650 SDValue ALo = DAG.getNode(X86ISD::UNPCKL, dl, VT, Amt, Z);
18651 SDValue AHi = DAG.getNode(X86ISD::UNPCKH, dl, VT, Amt, Z);
18652 SDValue RLo = DAG.getNode(X86ISD::UNPCKL, dl, VT, R, R);
18653 SDValue RHi = DAG.getNode(X86ISD::UNPCKH, dl, VT, R, R);
18654 ALo = DAG.getBitcast(ExtVT, ALo);
18655 AHi = DAG.getBitcast(ExtVT, AHi);
18656 RLo = DAG.getBitcast(ExtVT, RLo);
18657 RHi = DAG.getBitcast(ExtVT, RHi);
18658 SDValue Lo = DAG.getNode(Op.getOpcode(), dl, ExtVT, RLo, ALo);
18659 SDValue Hi = DAG.getNode(Op.getOpcode(), dl, ExtVT, RHi, AHi);
18660 Lo = DAG.getNode(ISD::SRL, dl, ExtVT, Lo, DAG.getConstant(16, dl, ExtVT));
18661 Hi = DAG.getNode(ISD::SRL, dl, ExtVT, Hi, DAG.getConstant(16, dl, ExtVT));
18662 return DAG.getNode(X86ISD::PACKUS, dl, VT, Lo, Hi);
18665 if (VT == MVT::v8i16) {
18666 unsigned ShiftOpcode = Op->getOpcode();
18668 auto SignBitSelect = [&](SDValue Sel, SDValue V0, SDValue V1) {
18669 // On SSE41 targets we make use of the fact that VSELECT lowers
18670 // to PBLENDVB which selects bytes based just on the sign bit.
18671 if (Subtarget->hasSSE41()) {
18672 MVT ExtVT = MVT::getVectorVT(MVT::i8, VT.getVectorNumElements() * 2);
18673 V0 = DAG.getBitcast(ExtVT, V0);
18674 V1 = DAG.getBitcast(ExtVT, V1);
18675 Sel = DAG.getBitcast(ExtVT, Sel);
18676 return DAG.getBitcast(
18677 VT, DAG.getNode(ISD::VSELECT, dl, ExtVT, Sel, V0, V1));
18679 // On pre-SSE41 targets we splat the sign bit - a negative value will
18680 // set all bits of the lanes to true and VSELECT uses that in
18681 // its OR(AND(V0,C),AND(V1,~C)) lowering.
18683 DAG.getNode(ISD::SRA, dl, VT, Sel, DAG.getConstant(15, dl, VT));
18684 return DAG.getNode(ISD::VSELECT, dl, VT, C, V0, V1);
18687 // Turn 'a' into a mask suitable for VSELECT: a = a << 12;
18688 if (Subtarget->hasSSE41()) {
18689 // On SSE41 targets we need to replicate the shift mask in both
18690 // bytes for PBLENDVB.
18693 DAG.getNode(ISD::SHL, dl, VT, Amt, DAG.getConstant(4, dl, VT)),
18694 DAG.getNode(ISD::SHL, dl, VT, Amt, DAG.getConstant(12, dl, VT)));
18696 Amt = DAG.getNode(ISD::SHL, dl, VT, Amt, DAG.getConstant(12, dl, VT));
18699 // r = VSELECT(r, shift(r, 8), a);
18700 SDValue M = DAG.getNode(ShiftOpcode, dl, VT, R, DAG.getConstant(8, dl, VT));
18701 R = SignBitSelect(Amt, M, R);
18704 Amt = DAG.getNode(ISD::ADD, dl, VT, Amt, Amt);
18706 // r = VSELECT(r, shift(r, 4), a);
18707 M = DAG.getNode(ShiftOpcode, dl, VT, R, DAG.getConstant(4, dl, VT));
18708 R = SignBitSelect(Amt, M, R);
18711 Amt = DAG.getNode(ISD::ADD, dl, VT, Amt, Amt);
18713 // r = VSELECT(r, shift(r, 2), a);
18714 M = DAG.getNode(ShiftOpcode, dl, VT, R, DAG.getConstant(2, dl, VT));
18715 R = SignBitSelect(Amt, M, R);
18718 Amt = DAG.getNode(ISD::ADD, dl, VT, Amt, Amt);
18720 // return VSELECT(r, shift(r, 1), a);
18721 M = DAG.getNode(ShiftOpcode, dl, VT, R, DAG.getConstant(1, dl, VT));
18722 R = SignBitSelect(Amt, M, R);
18726 // Decompose 256-bit shifts into smaller 128-bit shifts.
18727 if (VT.is256BitVector()) {
18728 unsigned NumElems = VT.getVectorNumElements();
18729 MVT EltVT = VT.getVectorElementType();
18730 EVT NewVT = MVT::getVectorVT(EltVT, NumElems/2);
18732 // Extract the two vectors
18733 SDValue V1 = Extract128BitVector(R, 0, DAG, dl);
18734 SDValue V2 = Extract128BitVector(R, NumElems/2, DAG, dl);
18736 // Recreate the shift amount vectors
18737 SDValue Amt1, Amt2;
18738 if (Amt.getOpcode() == ISD::BUILD_VECTOR) {
18739 // Constant shift amount
18740 SmallVector<SDValue, 8> Ops(Amt->op_begin(), Amt->op_begin() + NumElems);
18741 ArrayRef<SDValue> Amt1Csts = makeArrayRef(Ops).slice(0, NumElems / 2);
18742 ArrayRef<SDValue> Amt2Csts = makeArrayRef(Ops).slice(NumElems / 2);
18744 Amt1 = DAG.getNode(ISD::BUILD_VECTOR, dl, NewVT, Amt1Csts);
18745 Amt2 = DAG.getNode(ISD::BUILD_VECTOR, dl, NewVT, Amt2Csts);
18747 // Variable shift amount
18748 Amt1 = Extract128BitVector(Amt, 0, DAG, dl);
18749 Amt2 = Extract128BitVector(Amt, NumElems/2, DAG, dl);
18752 // Issue new vector shifts for the smaller types
18753 V1 = DAG.getNode(Op.getOpcode(), dl, NewVT, V1, Amt1);
18754 V2 = DAG.getNode(Op.getOpcode(), dl, NewVT, V2, Amt2);
18756 // Concatenate the result back
18757 return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT, V1, V2);
18763 static SDValue LowerRotate(SDValue Op, const X86Subtarget *Subtarget,
18764 SelectionDAG &DAG) {
18765 MVT VT = Op.getSimpleValueType();
18767 SDValue R = Op.getOperand(0);
18768 SDValue Amt = Op.getOperand(1);
18770 assert(VT.isVector() && "Custom lowering only for vector rotates!");
18771 assert(Subtarget->hasXOP() && "XOP support required for vector rotates!");
18772 assert((Op.getOpcode() == ISD::ROTL) && "Only ROTL supported");
18774 // XOP has 128-bit vector variable + immediate rotates.
18775 // +ve/-ve Amt = rotate left/right.
18777 // Split 256-bit integers.
18778 if (VT.getSizeInBits() == 256)
18779 return Lower256IntArith(Op, DAG);
18781 assert(VT.getSizeInBits() == 128 && "Only rotate 128-bit vectors!");
18783 // Attempt to rotate by immediate.
18784 if (auto *BVAmt = dyn_cast<BuildVectorSDNode>(Amt)) {
18785 if (auto *RotateConst = BVAmt->getConstantSplatNode()) {
18786 uint64_t RotateAmt = RotateConst->getAPIntValue().getZExtValue();
18787 assert(RotateAmt < VT.getScalarSizeInBits() && "Rotation out of range");
18788 return DAG.getNode(X86ISD::VPROTI, DL, VT, R,
18789 DAG.getConstant(RotateAmt, DL, MVT::i8));
18793 // Use general rotate by variable (per-element).
18794 return DAG.getNode(X86ISD::VPROT, DL, VT, R, Amt);
18797 static SDValue LowerXALUO(SDValue Op, SelectionDAG &DAG) {
18798 // Lower the "add/sub/mul with overflow" instruction into a regular ins plus
18799 // a "setcc" instruction that checks the overflow flag. The "brcond" lowering
18800 // looks for this combo and may remove the "setcc" instruction if the "setcc"
18801 // has only one use.
18802 SDNode *N = Op.getNode();
18803 SDValue LHS = N->getOperand(0);
18804 SDValue RHS = N->getOperand(1);
18805 unsigned BaseOp = 0;
18808 switch (Op.getOpcode()) {
18809 default: llvm_unreachable("Unknown ovf instruction!");
18811 // A subtract of one will be selected as a INC. Note that INC doesn't
18812 // set CF, so we can't do this for UADDO.
18813 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(RHS))
18815 BaseOp = X86ISD::INC;
18816 Cond = X86::COND_O;
18819 BaseOp = X86ISD::ADD;
18820 Cond = X86::COND_O;
18823 BaseOp = X86ISD::ADD;
18824 Cond = X86::COND_B;
18827 // A subtract of one will be selected as a DEC. Note that DEC doesn't
18828 // set CF, so we can't do this for USUBO.
18829 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(RHS))
18831 BaseOp = X86ISD::DEC;
18832 Cond = X86::COND_O;
18835 BaseOp = X86ISD::SUB;
18836 Cond = X86::COND_O;
18839 BaseOp = X86ISD::SUB;
18840 Cond = X86::COND_B;
18843 BaseOp = N->getValueType(0) == MVT::i8 ? X86ISD::SMUL8 : X86ISD::SMUL;
18844 Cond = X86::COND_O;
18846 case ISD::UMULO: { // i64, i8 = umulo lhs, rhs --> i64, i64, i32 umul lhs,rhs
18847 if (N->getValueType(0) == MVT::i8) {
18848 BaseOp = X86ISD::UMUL8;
18849 Cond = X86::COND_O;
18852 SDVTList VTs = DAG.getVTList(N->getValueType(0), N->getValueType(0),
18854 SDValue Sum = DAG.getNode(X86ISD::UMUL, DL, VTs, LHS, RHS);
18857 DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
18858 DAG.getConstant(X86::COND_O, DL, MVT::i32),
18859 SDValue(Sum.getNode(), 2));
18861 return DAG.getNode(ISD::MERGE_VALUES, DL, N->getVTList(), Sum, SetCC);
18865 // Also sets EFLAGS.
18866 SDVTList VTs = DAG.getVTList(N->getValueType(0), MVT::i32);
18867 SDValue Sum = DAG.getNode(BaseOp, DL, VTs, LHS, RHS);
18870 DAG.getNode(X86ISD::SETCC, DL, N->getValueType(1),
18871 DAG.getConstant(Cond, DL, MVT::i32),
18872 SDValue(Sum.getNode(), 1));
18874 return DAG.getNode(ISD::MERGE_VALUES, DL, N->getVTList(), Sum, SetCC);
18877 /// Returns true if the operand type is exactly twice the native width, and
18878 /// the corresponding cmpxchg8b or cmpxchg16b instruction is available.
18879 /// Used to know whether to use cmpxchg8/16b when expanding atomic operations
18880 /// (otherwise we leave them alone to become __sync_fetch_and_... calls).
18881 bool X86TargetLowering::needsCmpXchgNb(Type *MemType) const {
18882 unsigned OpWidth = MemType->getPrimitiveSizeInBits();
18885 return !Subtarget->is64Bit(); // FIXME this should be Subtarget.hasCmpxchg8b
18886 else if (OpWidth == 128)
18887 return Subtarget->hasCmpxchg16b();
18892 bool X86TargetLowering::shouldExpandAtomicStoreInIR(StoreInst *SI) const {
18893 return needsCmpXchgNb(SI->getValueOperand()->getType());
18896 // Note: this turns large loads into lock cmpxchg8b/16b.
18897 // FIXME: On 32 bits x86, fild/movq might be faster than lock cmpxchg8b.
18898 TargetLowering::AtomicExpansionKind
18899 X86TargetLowering::shouldExpandAtomicLoadInIR(LoadInst *LI) const {
18900 auto PTy = cast<PointerType>(LI->getPointerOperand()->getType());
18901 return needsCmpXchgNb(PTy->getElementType()) ? AtomicExpansionKind::CmpXChg
18902 : AtomicExpansionKind::None;
18905 TargetLowering::AtomicExpansionKind
18906 X86TargetLowering::shouldExpandAtomicRMWInIR(AtomicRMWInst *AI) const {
18907 unsigned NativeWidth = Subtarget->is64Bit() ? 64 : 32;
18908 Type *MemType = AI->getType();
18910 // If the operand is too big, we must see if cmpxchg8/16b is available
18911 // and default to library calls otherwise.
18912 if (MemType->getPrimitiveSizeInBits() > NativeWidth) {
18913 return needsCmpXchgNb(MemType) ? AtomicExpansionKind::CmpXChg
18914 : AtomicExpansionKind::None;
18917 AtomicRMWInst::BinOp Op = AI->getOperation();
18920 llvm_unreachable("Unknown atomic operation");
18921 case AtomicRMWInst::Xchg:
18922 case AtomicRMWInst::Add:
18923 case AtomicRMWInst::Sub:
18924 // It's better to use xadd, xsub or xchg for these in all cases.
18925 return AtomicExpansionKind::None;
18926 case AtomicRMWInst::Or:
18927 case AtomicRMWInst::And:
18928 case AtomicRMWInst::Xor:
18929 // If the atomicrmw's result isn't actually used, we can just add a "lock"
18930 // prefix to a normal instruction for these operations.
18931 return !AI->use_empty() ? AtomicExpansionKind::CmpXChg
18932 : AtomicExpansionKind::None;
18933 case AtomicRMWInst::Nand:
18934 case AtomicRMWInst::Max:
18935 case AtomicRMWInst::Min:
18936 case AtomicRMWInst::UMax:
18937 case AtomicRMWInst::UMin:
18938 // These always require a non-trivial set of data operations on x86. We must
18939 // use a cmpxchg loop.
18940 return AtomicExpansionKind::CmpXChg;
18944 static bool hasMFENCE(const X86Subtarget& Subtarget) {
18945 // Use mfence if we have SSE2 or we're on x86-64 (even if we asked for
18946 // no-sse2). There isn't any reason to disable it if the target processor
18948 return Subtarget.hasSSE2() || Subtarget.is64Bit();
18952 X86TargetLowering::lowerIdempotentRMWIntoFencedLoad(AtomicRMWInst *AI) const {
18953 unsigned NativeWidth = Subtarget->is64Bit() ? 64 : 32;
18954 Type *MemType = AI->getType();
18955 // Accesses larger than the native width are turned into cmpxchg/libcalls, so
18956 // there is no benefit in turning such RMWs into loads, and it is actually
18957 // harmful as it introduces a mfence.
18958 if (MemType->getPrimitiveSizeInBits() > NativeWidth)
18961 auto Builder = IRBuilder<>(AI);
18962 Module *M = Builder.GetInsertBlock()->getParent()->getParent();
18963 auto SynchScope = AI->getSynchScope();
18964 // We must restrict the ordering to avoid generating loads with Release or
18965 // ReleaseAcquire orderings.
18966 auto Order = AtomicCmpXchgInst::getStrongestFailureOrdering(AI->getOrdering());
18967 auto Ptr = AI->getPointerOperand();
18969 // Before the load we need a fence. Here is an example lifted from
18970 // http://www.hpl.hp.com/techreports/2012/HPL-2012-68.pdf showing why a fence
18973 // x.store(1, relaxed);
18974 // r1 = y.fetch_add(0, release);
18976 // y.fetch_add(42, acquire);
18977 // r2 = x.load(relaxed);
18978 // r1 = r2 = 0 is impossible, but becomes possible if the idempotent rmw is
18979 // lowered to just a load without a fence. A mfence flushes the store buffer,
18980 // making the optimization clearly correct.
18981 // FIXME: it is required if isAtLeastRelease(Order) but it is not clear
18982 // otherwise, we might be able to be more aggressive on relaxed idempotent
18983 // rmw. In practice, they do not look useful, so we don't try to be
18984 // especially clever.
18985 if (SynchScope == SingleThread)
18986 // FIXME: we could just insert an X86ISD::MEMBARRIER here, except we are at
18987 // the IR level, so we must wrap it in an intrinsic.
18990 if (!hasMFENCE(*Subtarget))
18991 // FIXME: it might make sense to use a locked operation here but on a
18992 // different cache-line to prevent cache-line bouncing. In practice it
18993 // is probably a small win, and x86 processors without mfence are rare
18994 // enough that we do not bother.
18998 llvm::Intrinsic::getDeclaration(M, Intrinsic::x86_sse2_mfence);
18999 Builder.CreateCall(MFence, {});
19001 // Finally we can emit the atomic load.
19002 LoadInst *Loaded = Builder.CreateAlignedLoad(Ptr,
19003 AI->getType()->getPrimitiveSizeInBits());
19004 Loaded->setAtomic(Order, SynchScope);
19005 AI->replaceAllUsesWith(Loaded);
19006 AI->eraseFromParent();
19010 static SDValue LowerATOMIC_FENCE(SDValue Op, const X86Subtarget *Subtarget,
19011 SelectionDAG &DAG) {
19013 AtomicOrdering FenceOrdering = static_cast<AtomicOrdering>(
19014 cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue());
19015 SynchronizationScope FenceScope = static_cast<SynchronizationScope>(
19016 cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue());
19018 // The only fence that needs an instruction is a sequentially-consistent
19019 // cross-thread fence.
19020 if (FenceOrdering == SequentiallyConsistent && FenceScope == CrossThread) {
19021 if (hasMFENCE(*Subtarget))
19022 return DAG.getNode(X86ISD::MFENCE, dl, MVT::Other, Op.getOperand(0));
19024 SDValue Chain = Op.getOperand(0);
19025 SDValue Zero = DAG.getConstant(0, dl, MVT::i32);
19027 DAG.getRegister(X86::ESP, MVT::i32), // Base
19028 DAG.getTargetConstant(1, dl, MVT::i8), // Scale
19029 DAG.getRegister(0, MVT::i32), // Index
19030 DAG.getTargetConstant(0, dl, MVT::i32), // Disp
19031 DAG.getRegister(0, MVT::i32), // Segment.
19035 SDNode *Res = DAG.getMachineNode(X86::OR32mrLocked, dl, MVT::Other, Ops);
19036 return SDValue(Res, 0);
19039 // MEMBARRIER is a compiler barrier; it codegens to a no-op.
19040 return DAG.getNode(X86ISD::MEMBARRIER, dl, MVT::Other, Op.getOperand(0));
19043 static SDValue LowerCMP_SWAP(SDValue Op, const X86Subtarget *Subtarget,
19044 SelectionDAG &DAG) {
19045 MVT T = Op.getSimpleValueType();
19049 switch(T.SimpleTy) {
19050 default: llvm_unreachable("Invalid value type!");
19051 case MVT::i8: Reg = X86::AL; size = 1; break;
19052 case MVT::i16: Reg = X86::AX; size = 2; break;
19053 case MVT::i32: Reg = X86::EAX; size = 4; break;
19055 assert(Subtarget->is64Bit() && "Node not type legal!");
19056 Reg = X86::RAX; size = 8;
19059 SDValue cpIn = DAG.getCopyToReg(Op.getOperand(0), DL, Reg,
19060 Op.getOperand(2), SDValue());
19061 SDValue Ops[] = { cpIn.getValue(0),
19064 DAG.getTargetConstant(size, DL, MVT::i8),
19065 cpIn.getValue(1) };
19066 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Glue);
19067 MachineMemOperand *MMO = cast<AtomicSDNode>(Op)->getMemOperand();
19068 SDValue Result = DAG.getMemIntrinsicNode(X86ISD::LCMPXCHG_DAG, DL, Tys,
19072 DAG.getCopyFromReg(Result.getValue(0), DL, Reg, T, Result.getValue(1));
19073 SDValue EFLAGS = DAG.getCopyFromReg(cpOut.getValue(1), DL, X86::EFLAGS,
19074 MVT::i32, cpOut.getValue(2));
19075 SDValue Success = DAG.getNode(X86ISD::SETCC, DL, Op->getValueType(1),
19076 DAG.getConstant(X86::COND_E, DL, MVT::i8),
19079 DAG.ReplaceAllUsesOfValueWith(Op.getValue(0), cpOut);
19080 DAG.ReplaceAllUsesOfValueWith(Op.getValue(1), Success);
19081 DAG.ReplaceAllUsesOfValueWith(Op.getValue(2), EFLAGS.getValue(1));
19085 static SDValue LowerBITCAST(SDValue Op, const X86Subtarget *Subtarget,
19086 SelectionDAG &DAG) {
19087 MVT SrcVT = Op.getOperand(0).getSimpleValueType();
19088 MVT DstVT = Op.getSimpleValueType();
19090 if (SrcVT == MVT::v2i32 || SrcVT == MVT::v4i16 || SrcVT == MVT::v8i8) {
19091 assert(Subtarget->hasSSE2() && "Requires at least SSE2!");
19092 if (DstVT != MVT::f64)
19093 // This conversion needs to be expanded.
19096 SDValue InVec = Op->getOperand(0);
19098 unsigned NumElts = SrcVT.getVectorNumElements();
19099 EVT SVT = SrcVT.getVectorElementType();
19101 // Widen the vector in input in the case of MVT::v2i32.
19102 // Example: from MVT::v2i32 to MVT::v4i32.
19103 SmallVector<SDValue, 16> Elts;
19104 for (unsigned i = 0, e = NumElts; i != e; ++i)
19105 Elts.push_back(DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, SVT, InVec,
19106 DAG.getIntPtrConstant(i, dl)));
19108 // Explicitly mark the extra elements as Undef.
19109 Elts.append(NumElts, DAG.getUNDEF(SVT));
19111 EVT NewVT = EVT::getVectorVT(*DAG.getContext(), SVT, NumElts * 2);
19112 SDValue BV = DAG.getNode(ISD::BUILD_VECTOR, dl, NewVT, Elts);
19113 SDValue ToV2F64 = DAG.getBitcast(MVT::v2f64, BV);
19114 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, ToV2F64,
19115 DAG.getIntPtrConstant(0, dl));
19118 assert(Subtarget->is64Bit() && !Subtarget->hasSSE2() &&
19119 Subtarget->hasMMX() && "Unexpected custom BITCAST");
19120 assert((DstVT == MVT::i64 ||
19121 (DstVT.isVector() && DstVT.getSizeInBits()==64)) &&
19122 "Unexpected custom BITCAST");
19123 // i64 <=> MMX conversions are Legal.
19124 if (SrcVT==MVT::i64 && DstVT.isVector())
19126 if (DstVT==MVT::i64 && SrcVT.isVector())
19128 // MMX <=> MMX conversions are Legal.
19129 if (SrcVT.isVector() && DstVT.isVector())
19131 // All other conversions need to be expanded.
19135 /// Compute the horizontal sum of bytes in V for the elements of VT.
19137 /// Requires V to be a byte vector and VT to be an integer vector type with
19138 /// wider elements than V's type. The width of the elements of VT determines
19139 /// how many bytes of V are summed horizontally to produce each element of the
19141 static SDValue LowerHorizontalByteSum(SDValue V, MVT VT,
19142 const X86Subtarget *Subtarget,
19143 SelectionDAG &DAG) {
19145 MVT ByteVecVT = V.getSimpleValueType();
19146 MVT EltVT = VT.getVectorElementType();
19147 int NumElts = VT.getVectorNumElements();
19148 assert(ByteVecVT.getVectorElementType() == MVT::i8 &&
19149 "Expected value to have byte element type.");
19150 assert(EltVT != MVT::i8 &&
19151 "Horizontal byte sum only makes sense for wider elements!");
19152 unsigned VecSize = VT.getSizeInBits();
19153 assert(ByteVecVT.getSizeInBits() == VecSize && "Cannot change vector size!");
19155 // PSADBW instruction horizontally add all bytes and leave the result in i64
19156 // chunks, thus directly computes the pop count for v2i64 and v4i64.
19157 if (EltVT == MVT::i64) {
19158 SDValue Zeros = getZeroVector(ByteVecVT, Subtarget, DAG, DL);
19159 V = DAG.getNode(X86ISD::PSADBW, DL, ByteVecVT, V, Zeros);
19160 return DAG.getBitcast(VT, V);
19163 if (EltVT == MVT::i32) {
19164 // We unpack the low half and high half into i32s interleaved with zeros so
19165 // that we can use PSADBW to horizontally sum them. The most useful part of
19166 // this is that it lines up the results of two PSADBW instructions to be
19167 // two v2i64 vectors which concatenated are the 4 population counts. We can
19168 // then use PACKUSWB to shrink and concatenate them into a v4i32 again.
19169 SDValue Zeros = getZeroVector(VT, Subtarget, DAG, DL);
19170 SDValue Low = DAG.getNode(X86ISD::UNPCKL, DL, VT, V, Zeros);
19171 SDValue High = DAG.getNode(X86ISD::UNPCKH, DL, VT, V, Zeros);
19173 // Do the horizontal sums into two v2i64s.
19174 Zeros = getZeroVector(ByteVecVT, Subtarget, DAG, DL);
19175 Low = DAG.getNode(X86ISD::PSADBW, DL, ByteVecVT,
19176 DAG.getBitcast(ByteVecVT, Low), Zeros);
19177 High = DAG.getNode(X86ISD::PSADBW, DL, ByteVecVT,
19178 DAG.getBitcast(ByteVecVT, High), Zeros);
19180 // Merge them together.
19181 MVT ShortVecVT = MVT::getVectorVT(MVT::i16, VecSize / 16);
19182 V = DAG.getNode(X86ISD::PACKUS, DL, ByteVecVT,
19183 DAG.getBitcast(ShortVecVT, Low),
19184 DAG.getBitcast(ShortVecVT, High));
19186 return DAG.getBitcast(VT, V);
19189 // The only element type left is i16.
19190 assert(EltVT == MVT::i16 && "Unknown how to handle type");
19192 // To obtain pop count for each i16 element starting from the pop count for
19193 // i8 elements, shift the i16s left by 8, sum as i8s, and then shift as i16s
19194 // right by 8. It is important to shift as i16s as i8 vector shift isn't
19195 // directly supported.
19196 SmallVector<SDValue, 16> Shifters(NumElts, DAG.getConstant(8, DL, EltVT));
19197 SDValue Shifter = DAG.getNode(ISD::BUILD_VECTOR, DL, VT, Shifters);
19198 SDValue Shl = DAG.getNode(ISD::SHL, DL, VT, DAG.getBitcast(VT, V), Shifter);
19199 V = DAG.getNode(ISD::ADD, DL, ByteVecVT, DAG.getBitcast(ByteVecVT, Shl),
19200 DAG.getBitcast(ByteVecVT, V));
19201 return DAG.getNode(ISD::SRL, DL, VT, DAG.getBitcast(VT, V), Shifter);
19204 static SDValue LowerVectorCTPOPInRegLUT(SDValue Op, SDLoc DL,
19205 const X86Subtarget *Subtarget,
19206 SelectionDAG &DAG) {
19207 MVT VT = Op.getSimpleValueType();
19208 MVT EltVT = VT.getVectorElementType();
19209 unsigned VecSize = VT.getSizeInBits();
19211 // Implement a lookup table in register by using an algorithm based on:
19212 // http://wm.ite.pl/articles/sse-popcount.html
19214 // The general idea is that every lower byte nibble in the input vector is an
19215 // index into a in-register pre-computed pop count table. We then split up the
19216 // input vector in two new ones: (1) a vector with only the shifted-right
19217 // higher nibbles for each byte and (2) a vector with the lower nibbles (and
19218 // masked out higher ones) for each byte. PSHUB is used separately with both
19219 // to index the in-register table. Next, both are added and the result is a
19220 // i8 vector where each element contains the pop count for input byte.
19222 // To obtain the pop count for elements != i8, we follow up with the same
19223 // approach and use additional tricks as described below.
19225 const int LUT[16] = {/* 0 */ 0, /* 1 */ 1, /* 2 */ 1, /* 3 */ 2,
19226 /* 4 */ 1, /* 5 */ 2, /* 6 */ 2, /* 7 */ 3,
19227 /* 8 */ 1, /* 9 */ 2, /* a */ 2, /* b */ 3,
19228 /* c */ 2, /* d */ 3, /* e */ 3, /* f */ 4};
19230 int NumByteElts = VecSize / 8;
19231 MVT ByteVecVT = MVT::getVectorVT(MVT::i8, NumByteElts);
19232 SDValue In = DAG.getBitcast(ByteVecVT, Op);
19233 SmallVector<SDValue, 16> LUTVec;
19234 for (int i = 0; i < NumByteElts; ++i)
19235 LUTVec.push_back(DAG.getConstant(LUT[i % 16], DL, MVT::i8));
19236 SDValue InRegLUT = DAG.getNode(ISD::BUILD_VECTOR, DL, ByteVecVT, LUTVec);
19237 SmallVector<SDValue, 16> Mask0F(NumByteElts,
19238 DAG.getConstant(0x0F, DL, MVT::i8));
19239 SDValue M0F = DAG.getNode(ISD::BUILD_VECTOR, DL, ByteVecVT, Mask0F);
19242 SmallVector<SDValue, 16> Four(NumByteElts, DAG.getConstant(4, DL, MVT::i8));
19243 SDValue FourV = DAG.getNode(ISD::BUILD_VECTOR, DL, ByteVecVT, Four);
19244 SDValue HighNibbles = DAG.getNode(ISD::SRL, DL, ByteVecVT, In, FourV);
19247 SDValue LowNibbles = DAG.getNode(ISD::AND, DL, ByteVecVT, In, M0F);
19249 // The input vector is used as the shuffle mask that index elements into the
19250 // LUT. After counting low and high nibbles, add the vector to obtain the
19251 // final pop count per i8 element.
19252 SDValue HighPopCnt =
19253 DAG.getNode(X86ISD::PSHUFB, DL, ByteVecVT, InRegLUT, HighNibbles);
19254 SDValue LowPopCnt =
19255 DAG.getNode(X86ISD::PSHUFB, DL, ByteVecVT, InRegLUT, LowNibbles);
19256 SDValue PopCnt = DAG.getNode(ISD::ADD, DL, ByteVecVT, HighPopCnt, LowPopCnt);
19258 if (EltVT == MVT::i8)
19261 return LowerHorizontalByteSum(PopCnt, VT, Subtarget, DAG);
19264 static SDValue LowerVectorCTPOPBitmath(SDValue Op, SDLoc DL,
19265 const X86Subtarget *Subtarget,
19266 SelectionDAG &DAG) {
19267 MVT VT = Op.getSimpleValueType();
19268 assert(VT.is128BitVector() &&
19269 "Only 128-bit vector bitmath lowering supported.");
19271 int VecSize = VT.getSizeInBits();
19272 MVT EltVT = VT.getVectorElementType();
19273 int Len = EltVT.getSizeInBits();
19275 // This is the vectorized version of the "best" algorithm from
19276 // http://graphics.stanford.edu/~seander/bithacks.html#CountBitsSetParallel
19277 // with a minor tweak to use a series of adds + shifts instead of vector
19278 // multiplications. Implemented for all integer vector types. We only use
19279 // this when we don't have SSSE3 which allows a LUT-based lowering that is
19280 // much faster, even faster than using native popcnt instructions.
19282 auto GetShift = [&](unsigned OpCode, SDValue V, int Shifter) {
19283 MVT VT = V.getSimpleValueType();
19284 SmallVector<SDValue, 32> Shifters(
19285 VT.getVectorNumElements(),
19286 DAG.getConstant(Shifter, DL, VT.getVectorElementType()));
19287 return DAG.getNode(OpCode, DL, VT, V,
19288 DAG.getNode(ISD::BUILD_VECTOR, DL, VT, Shifters));
19290 auto GetMask = [&](SDValue V, APInt Mask) {
19291 MVT VT = V.getSimpleValueType();
19292 SmallVector<SDValue, 32> Masks(
19293 VT.getVectorNumElements(),
19294 DAG.getConstant(Mask, DL, VT.getVectorElementType()));
19295 return DAG.getNode(ISD::AND, DL, VT, V,
19296 DAG.getNode(ISD::BUILD_VECTOR, DL, VT, Masks));
19299 // We don't want to incur the implicit masks required to SRL vNi8 vectors on
19300 // x86, so set the SRL type to have elements at least i16 wide. This is
19301 // correct because all of our SRLs are followed immediately by a mask anyways
19302 // that handles any bits that sneak into the high bits of the byte elements.
19303 MVT SrlVT = Len > 8 ? VT : MVT::getVectorVT(MVT::i16, VecSize / 16);
19307 // v = v - ((v >> 1) & 0x55555555...)
19309 DAG.getBitcast(VT, GetShift(ISD::SRL, DAG.getBitcast(SrlVT, V), 1));
19310 SDValue And = GetMask(Srl, APInt::getSplat(Len, APInt(8, 0x55)));
19311 V = DAG.getNode(ISD::SUB, DL, VT, V, And);
19313 // v = (v & 0x33333333...) + ((v >> 2) & 0x33333333...)
19314 SDValue AndLHS = GetMask(V, APInt::getSplat(Len, APInt(8, 0x33)));
19315 Srl = DAG.getBitcast(VT, GetShift(ISD::SRL, DAG.getBitcast(SrlVT, V), 2));
19316 SDValue AndRHS = GetMask(Srl, APInt::getSplat(Len, APInt(8, 0x33)));
19317 V = DAG.getNode(ISD::ADD, DL, VT, AndLHS, AndRHS);
19319 // v = (v + (v >> 4)) & 0x0F0F0F0F...
19320 Srl = DAG.getBitcast(VT, GetShift(ISD::SRL, DAG.getBitcast(SrlVT, V), 4));
19321 SDValue Add = DAG.getNode(ISD::ADD, DL, VT, V, Srl);
19322 V = GetMask(Add, APInt::getSplat(Len, APInt(8, 0x0F)));
19324 // At this point, V contains the byte-wise population count, and we are
19325 // merely doing a horizontal sum if necessary to get the wider element
19327 if (EltVT == MVT::i8)
19330 return LowerHorizontalByteSum(
19331 DAG.getBitcast(MVT::getVectorVT(MVT::i8, VecSize / 8), V), VT, Subtarget,
19335 static SDValue LowerVectorCTPOP(SDValue Op, const X86Subtarget *Subtarget,
19336 SelectionDAG &DAG) {
19337 MVT VT = Op.getSimpleValueType();
19338 // FIXME: Need to add AVX-512 support here!
19339 assert((VT.is256BitVector() || VT.is128BitVector()) &&
19340 "Unknown CTPOP type to handle");
19341 SDLoc DL(Op.getNode());
19342 SDValue Op0 = Op.getOperand(0);
19344 if (!Subtarget->hasSSSE3()) {
19345 // We can't use the fast LUT approach, so fall back on vectorized bitmath.
19346 assert(VT.is128BitVector() && "Only 128-bit vectors supported in SSE!");
19347 return LowerVectorCTPOPBitmath(Op0, DL, Subtarget, DAG);
19350 if (VT.is256BitVector() && !Subtarget->hasInt256()) {
19351 unsigned NumElems = VT.getVectorNumElements();
19353 // Extract each 128-bit vector, compute pop count and concat the result.
19354 SDValue LHS = Extract128BitVector(Op0, 0, DAG, DL);
19355 SDValue RHS = Extract128BitVector(Op0, NumElems/2, DAG, DL);
19357 return DAG.getNode(ISD::CONCAT_VECTORS, DL, VT,
19358 LowerVectorCTPOPInRegLUT(LHS, DL, Subtarget, DAG),
19359 LowerVectorCTPOPInRegLUT(RHS, DL, Subtarget, DAG));
19362 return LowerVectorCTPOPInRegLUT(Op0, DL, Subtarget, DAG);
19365 static SDValue LowerCTPOP(SDValue Op, const X86Subtarget *Subtarget,
19366 SelectionDAG &DAG) {
19367 assert(Op.getValueType().isVector() &&
19368 "We only do custom lowering for vector population count.");
19369 return LowerVectorCTPOP(Op, Subtarget, DAG);
19372 static SDValue LowerLOAD_SUB(SDValue Op, SelectionDAG &DAG) {
19373 SDNode *Node = Op.getNode();
19375 EVT T = Node->getValueType(0);
19376 SDValue negOp = DAG.getNode(ISD::SUB, dl, T,
19377 DAG.getConstant(0, dl, T), Node->getOperand(2));
19378 return DAG.getAtomic(ISD::ATOMIC_LOAD_ADD, dl,
19379 cast<AtomicSDNode>(Node)->getMemoryVT(),
19380 Node->getOperand(0),
19381 Node->getOperand(1), negOp,
19382 cast<AtomicSDNode>(Node)->getMemOperand(),
19383 cast<AtomicSDNode>(Node)->getOrdering(),
19384 cast<AtomicSDNode>(Node)->getSynchScope());
19387 static SDValue LowerATOMIC_STORE(SDValue Op, SelectionDAG &DAG) {
19388 SDNode *Node = Op.getNode();
19390 EVT VT = cast<AtomicSDNode>(Node)->getMemoryVT();
19392 // Convert seq_cst store -> xchg
19393 // Convert wide store -> swap (-> cmpxchg8b/cmpxchg16b)
19394 // FIXME: On 32-bit, store -> fist or movq would be more efficient
19395 // (The only way to get a 16-byte store is cmpxchg16b)
19396 // FIXME: 16-byte ATOMIC_SWAP isn't actually hooked up at the moment.
19397 if (cast<AtomicSDNode>(Node)->getOrdering() == SequentiallyConsistent ||
19398 !DAG.getTargetLoweringInfo().isTypeLegal(VT)) {
19399 SDValue Swap = DAG.getAtomic(ISD::ATOMIC_SWAP, dl,
19400 cast<AtomicSDNode>(Node)->getMemoryVT(),
19401 Node->getOperand(0),
19402 Node->getOperand(1), Node->getOperand(2),
19403 cast<AtomicSDNode>(Node)->getMemOperand(),
19404 cast<AtomicSDNode>(Node)->getOrdering(),
19405 cast<AtomicSDNode>(Node)->getSynchScope());
19406 return Swap.getValue(1);
19408 // Other atomic stores have a simple pattern.
19412 static SDValue LowerADDC_ADDE_SUBC_SUBE(SDValue Op, SelectionDAG &DAG) {
19413 EVT VT = Op.getNode()->getSimpleValueType(0);
19415 // Let legalize expand this if it isn't a legal type yet.
19416 if (!DAG.getTargetLoweringInfo().isTypeLegal(VT))
19419 SDVTList VTs = DAG.getVTList(VT, MVT::i32);
19422 bool ExtraOp = false;
19423 switch (Op.getOpcode()) {
19424 default: llvm_unreachable("Invalid code");
19425 case ISD::ADDC: Opc = X86ISD::ADD; break;
19426 case ISD::ADDE: Opc = X86ISD::ADC; ExtraOp = true; break;
19427 case ISD::SUBC: Opc = X86ISD::SUB; break;
19428 case ISD::SUBE: Opc = X86ISD::SBB; ExtraOp = true; break;
19432 return DAG.getNode(Opc, SDLoc(Op), VTs, Op.getOperand(0),
19434 return DAG.getNode(Opc, SDLoc(Op), VTs, Op.getOperand(0),
19435 Op.getOperand(1), Op.getOperand(2));
19438 static SDValue LowerFSINCOS(SDValue Op, const X86Subtarget *Subtarget,
19439 SelectionDAG &DAG) {
19440 assert(Subtarget->isTargetDarwin() && Subtarget->is64Bit());
19442 // For MacOSX, we want to call an alternative entry point: __sincos_stret,
19443 // which returns the values as { float, float } (in XMM0) or
19444 // { double, double } (which is returned in XMM0, XMM1).
19446 SDValue Arg = Op.getOperand(0);
19447 EVT ArgVT = Arg.getValueType();
19448 Type *ArgTy = ArgVT.getTypeForEVT(*DAG.getContext());
19450 TargetLowering::ArgListTy Args;
19451 TargetLowering::ArgListEntry Entry;
19455 Entry.isSExt = false;
19456 Entry.isZExt = false;
19457 Args.push_back(Entry);
19459 bool isF64 = ArgVT == MVT::f64;
19460 // Only optimize x86_64 for now. i386 is a bit messy. For f32,
19461 // the small struct {f32, f32} is returned in (eax, edx). For f64,
19462 // the results are returned via SRet in memory.
19463 const char *LibcallName = isF64 ? "__sincos_stret" : "__sincosf_stret";
19464 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
19466 DAG.getExternalSymbol(LibcallName, TLI.getPointerTy(DAG.getDataLayout()));
19468 Type *RetTy = isF64
19469 ? (Type*)StructType::get(ArgTy, ArgTy, nullptr)
19470 : (Type*)VectorType::get(ArgTy, 4);
19472 TargetLowering::CallLoweringInfo CLI(DAG);
19473 CLI.setDebugLoc(dl).setChain(DAG.getEntryNode())
19474 .setCallee(CallingConv::C, RetTy, Callee, std::move(Args), 0);
19476 std::pair<SDValue, SDValue> CallResult = TLI.LowerCallTo(CLI);
19479 // Returned in xmm0 and xmm1.
19480 return CallResult.first;
19482 // Returned in bits 0:31 and 32:64 xmm0.
19483 SDValue SinVal = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, ArgVT,
19484 CallResult.first, DAG.getIntPtrConstant(0, dl));
19485 SDValue CosVal = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, ArgVT,
19486 CallResult.first, DAG.getIntPtrConstant(1, dl));
19487 SDVTList Tys = DAG.getVTList(ArgVT, ArgVT);
19488 return DAG.getNode(ISD::MERGE_VALUES, dl, Tys, SinVal, CosVal);
19491 static SDValue LowerMSCATTER(SDValue Op, const X86Subtarget *Subtarget,
19492 SelectionDAG &DAG) {
19493 assert(Subtarget->hasAVX512() &&
19494 "MGATHER/MSCATTER are supported on AVX-512 arch only");
19496 MaskedScatterSDNode *N = cast<MaskedScatterSDNode>(Op.getNode());
19497 EVT VT = N->getValue().getValueType();
19498 assert(VT.getScalarSizeInBits() >= 32 && "Unsupported scatter op");
19501 // X86 scatter kills mask register, so its type should be added to
19502 // the list of return values
19503 if (N->getNumValues() == 1) {
19504 SDValue Index = N->getIndex();
19505 if (!Subtarget->hasVLX() && !VT.is512BitVector() &&
19506 !Index.getValueType().is512BitVector())
19507 Index = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v8i64, Index);
19509 SDVTList VTs = DAG.getVTList(N->getMask().getValueType(), MVT::Other);
19510 SDValue Ops[] = { N->getOperand(0), N->getOperand(1), N->getOperand(2),
19511 N->getOperand(3), Index };
19513 SDValue NewScatter = DAG.getMaskedScatter(VTs, VT, dl, Ops, N->getMemOperand());
19514 DAG.ReplaceAllUsesWith(Op, SDValue(NewScatter.getNode(), 1));
19515 return SDValue(NewScatter.getNode(), 0);
19520 static SDValue LowerMGATHER(SDValue Op, const X86Subtarget *Subtarget,
19521 SelectionDAG &DAG) {
19522 assert(Subtarget->hasAVX512() &&
19523 "MGATHER/MSCATTER are supported on AVX-512 arch only");
19525 MaskedGatherSDNode *N = cast<MaskedGatherSDNode>(Op.getNode());
19526 EVT VT = Op.getValueType();
19527 assert(VT.getScalarSizeInBits() >= 32 && "Unsupported gather op");
19530 SDValue Index = N->getIndex();
19531 if (!Subtarget->hasVLX() && !VT.is512BitVector() &&
19532 !Index.getValueType().is512BitVector()) {
19533 Index = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v8i64, Index);
19534 SDValue Ops[] = { N->getOperand(0), N->getOperand(1), N->getOperand(2),
19535 N->getOperand(3), Index };
19536 DAG.UpdateNodeOperands(N, Ops);
19541 SDValue X86TargetLowering::LowerGC_TRANSITION_START(SDValue Op,
19542 SelectionDAG &DAG) const {
19543 // TODO: Eventually, the lowering of these nodes should be informed by or
19544 // deferred to the GC strategy for the function in which they appear. For
19545 // now, however, they must be lowered to something. Since they are logically
19546 // no-ops in the case of a null GC strategy (or a GC strategy which does not
19547 // require special handling for these nodes), lower them as literal NOOPs for
19549 SmallVector<SDValue, 2> Ops;
19551 Ops.push_back(Op.getOperand(0));
19552 if (Op->getGluedNode())
19553 Ops.push_back(Op->getOperand(Op->getNumOperands() - 1));
19556 SDVTList VTs = DAG.getVTList(MVT::Other, MVT::Glue);
19557 SDValue NOOP(DAG.getMachineNode(X86::NOOP, SDLoc(Op), VTs, Ops), 0);
19562 SDValue X86TargetLowering::LowerGC_TRANSITION_END(SDValue Op,
19563 SelectionDAG &DAG) const {
19564 // TODO: Eventually, the lowering of these nodes should be informed by or
19565 // deferred to the GC strategy for the function in which they appear. For
19566 // now, however, they must be lowered to something. Since they are logically
19567 // no-ops in the case of a null GC strategy (or a GC strategy which does not
19568 // require special handling for these nodes), lower them as literal NOOPs for
19570 SmallVector<SDValue, 2> Ops;
19572 Ops.push_back(Op.getOperand(0));
19573 if (Op->getGluedNode())
19574 Ops.push_back(Op->getOperand(Op->getNumOperands() - 1));
19577 SDVTList VTs = DAG.getVTList(MVT::Other, MVT::Glue);
19578 SDValue NOOP(DAG.getMachineNode(X86::NOOP, SDLoc(Op), VTs, Ops), 0);
19583 /// LowerOperation - Provide custom lowering hooks for some operations.
19585 SDValue X86TargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
19586 switch (Op.getOpcode()) {
19587 default: llvm_unreachable("Should not custom lower this!");
19588 case ISD::ATOMIC_FENCE: return LowerATOMIC_FENCE(Op, Subtarget, DAG);
19589 case ISD::ATOMIC_CMP_SWAP_WITH_SUCCESS:
19590 return LowerCMP_SWAP(Op, Subtarget, DAG);
19591 case ISD::CTPOP: return LowerCTPOP(Op, Subtarget, DAG);
19592 case ISD::ATOMIC_LOAD_SUB: return LowerLOAD_SUB(Op,DAG);
19593 case ISD::ATOMIC_STORE: return LowerATOMIC_STORE(Op,DAG);
19594 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG);
19595 case ISD::CONCAT_VECTORS: return LowerCONCAT_VECTORS(Op, Subtarget, DAG);
19596 case ISD::VECTOR_SHUFFLE: return lowerVectorShuffle(Op, Subtarget, DAG);
19597 case ISD::VSELECT: return LowerVSELECT(Op, DAG);
19598 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
19599 case ISD::INSERT_VECTOR_ELT: return LowerINSERT_VECTOR_ELT(Op, DAG);
19600 case ISD::EXTRACT_SUBVECTOR: return LowerEXTRACT_SUBVECTOR(Op,Subtarget,DAG);
19601 case ISD::INSERT_SUBVECTOR: return LowerINSERT_SUBVECTOR(Op, Subtarget,DAG);
19602 case ISD::SCALAR_TO_VECTOR: return LowerSCALAR_TO_VECTOR(Op, DAG);
19603 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
19604 case ISD::GlobalAddress: return LowerGlobalAddress(Op, DAG);
19605 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
19606 case ISD::ExternalSymbol: return LowerExternalSymbol(Op, DAG);
19607 case ISD::BlockAddress: return LowerBlockAddress(Op, DAG);
19608 case ISD::SHL_PARTS:
19609 case ISD::SRA_PARTS:
19610 case ISD::SRL_PARTS: return LowerShiftParts(Op, DAG);
19611 case ISD::SINT_TO_FP: return LowerSINT_TO_FP(Op, DAG);
19612 case ISD::UINT_TO_FP: return LowerUINT_TO_FP(Op, DAG);
19613 case ISD::TRUNCATE: return LowerTRUNCATE(Op, DAG);
19614 case ISD::ZERO_EXTEND: return LowerZERO_EXTEND(Op, Subtarget, DAG);
19615 case ISD::SIGN_EXTEND: return LowerSIGN_EXTEND(Op, Subtarget, DAG);
19616 case ISD::ANY_EXTEND: return LowerANY_EXTEND(Op, Subtarget, DAG);
19617 case ISD::SIGN_EXTEND_VECTOR_INREG:
19618 return LowerSIGN_EXTEND_VECTOR_INREG(Op, Subtarget, DAG);
19619 case ISD::FP_TO_SINT: return LowerFP_TO_SINT(Op, DAG);
19620 case ISD::FP_TO_UINT: return LowerFP_TO_UINT(Op, DAG);
19621 case ISD::FP_EXTEND: return LowerFP_EXTEND(Op, DAG);
19622 case ISD::LOAD: return LowerExtendedLoad(Op, Subtarget, DAG);
19624 case ISD::FNEG: return LowerFABSorFNEG(Op, DAG);
19625 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
19626 case ISD::FGETSIGN: return LowerFGETSIGN(Op, DAG);
19627 case ISD::SETCC: return LowerSETCC(Op, DAG);
19628 case ISD::SELECT: return LowerSELECT(Op, DAG);
19629 case ISD::BRCOND: return LowerBRCOND(Op, DAG);
19630 case ISD::JumpTable: return LowerJumpTable(Op, DAG);
19631 case ISD::VASTART: return LowerVASTART(Op, DAG);
19632 case ISD::VAARG: return LowerVAARG(Op, DAG);
19633 case ISD::VACOPY: return LowerVACOPY(Op, Subtarget, DAG);
19634 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, Subtarget, DAG);
19635 case ISD::INTRINSIC_VOID:
19636 case ISD::INTRINSIC_W_CHAIN: return LowerINTRINSIC_W_CHAIN(Op, Subtarget, DAG);
19637 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
19638 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
19639 case ISD::FRAME_TO_ARGS_OFFSET:
19640 return LowerFRAME_TO_ARGS_OFFSET(Op, DAG);
19641 case ISD::DYNAMIC_STACKALLOC: return LowerDYNAMIC_STACKALLOC(Op, DAG);
19642 case ISD::EH_RETURN: return LowerEH_RETURN(Op, DAG);
19643 case ISD::EH_SJLJ_SETJMP: return lowerEH_SJLJ_SETJMP(Op, DAG);
19644 case ISD::EH_SJLJ_LONGJMP: return lowerEH_SJLJ_LONGJMP(Op, DAG);
19645 case ISD::INIT_TRAMPOLINE: return LowerINIT_TRAMPOLINE(Op, DAG);
19646 case ISD::ADJUST_TRAMPOLINE: return LowerADJUST_TRAMPOLINE(Op, DAG);
19647 case ISD::FLT_ROUNDS_: return LowerFLT_ROUNDS_(Op, DAG);
19648 case ISD::CTLZ: return LowerCTLZ(Op, Subtarget, DAG);
19649 case ISD::CTLZ_ZERO_UNDEF: return LowerCTLZ_ZERO_UNDEF(Op, Subtarget, DAG);
19651 case ISD::CTTZ_ZERO_UNDEF: return LowerCTTZ(Op, DAG);
19652 case ISD::MUL: return LowerMUL(Op, Subtarget, DAG);
19653 case ISD::UMUL_LOHI:
19654 case ISD::SMUL_LOHI: return LowerMUL_LOHI(Op, Subtarget, DAG);
19655 case ISD::ROTL: return LowerRotate(Op, Subtarget, DAG);
19658 case ISD::SHL: return LowerShift(Op, Subtarget, DAG);
19664 case ISD::UMULO: return LowerXALUO(Op, DAG);
19665 case ISD::READCYCLECOUNTER: return LowerREADCYCLECOUNTER(Op, Subtarget,DAG);
19666 case ISD::BITCAST: return LowerBITCAST(Op, Subtarget, DAG);
19670 case ISD::SUBE: return LowerADDC_ADDE_SUBC_SUBE(Op, DAG);
19671 case ISD::ADD: return LowerADD(Op, DAG);
19672 case ISD::SUB: return LowerSUB(Op, DAG);
19676 case ISD::UMIN: return LowerMINMAX(Op, DAG);
19677 case ISD::FSINCOS: return LowerFSINCOS(Op, Subtarget, DAG);
19678 case ISD::MGATHER: return LowerMGATHER(Op, Subtarget, DAG);
19679 case ISD::MSCATTER: return LowerMSCATTER(Op, Subtarget, DAG);
19680 case ISD::GC_TRANSITION_START:
19681 return LowerGC_TRANSITION_START(Op, DAG);
19682 case ISD::GC_TRANSITION_END: return LowerGC_TRANSITION_END(Op, DAG);
19686 /// ReplaceNodeResults - Replace a node with an illegal result type
19687 /// with a new node built out of custom code.
19688 void X86TargetLowering::ReplaceNodeResults(SDNode *N,
19689 SmallVectorImpl<SDValue>&Results,
19690 SelectionDAG &DAG) const {
19692 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
19693 switch (N->getOpcode()) {
19695 llvm_unreachable("Do not know how to custom type legalize this operation!");
19696 // We might have generated v2f32 FMIN/FMAX operations. Widen them to v4f32.
19697 case X86ISD::FMINC:
19699 case X86ISD::FMAXC:
19700 case X86ISD::FMAX: {
19701 EVT VT = N->getValueType(0);
19702 if (VT != MVT::v2f32)
19703 llvm_unreachable("Unexpected type (!= v2f32) on FMIN/FMAX.");
19704 SDValue UNDEF = DAG.getUNDEF(VT);
19705 SDValue LHS = DAG.getNode(ISD::CONCAT_VECTORS, dl, MVT::v4f32,
19706 N->getOperand(0), UNDEF);
19707 SDValue RHS = DAG.getNode(ISD::CONCAT_VECTORS, dl, MVT::v4f32,
19708 N->getOperand(1), UNDEF);
19709 Results.push_back(DAG.getNode(N->getOpcode(), dl, MVT::v4f32, LHS, RHS));
19712 case ISD::SIGN_EXTEND_INREG:
19717 // We don't want to expand or promote these.
19724 case ISD::UDIVREM: {
19725 SDValue V = LowerWin64_i128OP(SDValue(N,0), DAG);
19726 Results.push_back(V);
19729 case ISD::FP_TO_SINT:
19730 case ISD::FP_TO_UINT: {
19731 bool IsSigned = N->getOpcode() == ISD::FP_TO_SINT;
19733 std::pair<SDValue,SDValue> Vals =
19734 FP_TO_INTHelper(SDValue(N, 0), DAG, IsSigned, /*IsReplace=*/ true);
19735 SDValue FIST = Vals.first, StackSlot = Vals.second;
19736 if (FIST.getNode()) {
19737 EVT VT = N->getValueType(0);
19738 // Return a load from the stack slot.
19739 if (StackSlot.getNode())
19740 Results.push_back(DAG.getLoad(VT, dl, FIST, StackSlot,
19741 MachinePointerInfo(),
19742 false, false, false, 0));
19744 Results.push_back(FIST);
19748 case ISD::UINT_TO_FP: {
19749 assert(Subtarget->hasSSE2() && "Requires at least SSE2!");
19750 if (N->getOperand(0).getValueType() != MVT::v2i32 ||
19751 N->getValueType(0) != MVT::v2f32)
19753 SDValue ZExtIn = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v2i64,
19755 SDValue Bias = DAG.getConstantFP(BitsToDouble(0x4330000000000000ULL), dl,
19757 SDValue VBias = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v2f64, Bias, Bias);
19758 SDValue Or = DAG.getNode(ISD::OR, dl, MVT::v2i64, ZExtIn,
19759 DAG.getBitcast(MVT::v2i64, VBias));
19760 Or = DAG.getBitcast(MVT::v2f64, Or);
19761 // TODO: Are there any fast-math-flags to propagate here?
19762 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::v2f64, Or, VBias);
19763 Results.push_back(DAG.getNode(X86ISD::VFPROUND, dl, MVT::v4f32, Sub));
19766 case ISD::FP_ROUND: {
19767 if (!TLI.isTypeLegal(N->getOperand(0).getValueType()))
19769 SDValue V = DAG.getNode(X86ISD::VFPROUND, dl, MVT::v4f32, N->getOperand(0));
19770 Results.push_back(V);
19773 case ISD::FP_EXTEND: {
19774 // Right now, only MVT::v2f32 has OperationAction for FP_EXTEND.
19775 // No other ValueType for FP_EXTEND should reach this point.
19776 assert(N->getValueType(0) == MVT::v2f32 &&
19777 "Do not know how to legalize this Node");
19780 case ISD::INTRINSIC_W_CHAIN: {
19781 unsigned IntNo = cast<ConstantSDNode>(N->getOperand(1))->getZExtValue();
19783 default : llvm_unreachable("Do not know how to custom type "
19784 "legalize this intrinsic operation!");
19785 case Intrinsic::x86_rdtsc:
19786 return getReadTimeStampCounter(N, dl, X86ISD::RDTSC_DAG, DAG, Subtarget,
19788 case Intrinsic::x86_rdtscp:
19789 return getReadTimeStampCounter(N, dl, X86ISD::RDTSCP_DAG, DAG, Subtarget,
19791 case Intrinsic::x86_rdpmc:
19792 return getReadPerformanceCounter(N, dl, DAG, Subtarget, Results);
19795 case ISD::READCYCLECOUNTER: {
19796 return getReadTimeStampCounter(N, dl, X86ISD::RDTSC_DAG, DAG, Subtarget,
19799 case ISD::ATOMIC_CMP_SWAP_WITH_SUCCESS: {
19800 EVT T = N->getValueType(0);
19801 assert((T == MVT::i64 || T == MVT::i128) && "can only expand cmpxchg pair");
19802 bool Regs64bit = T == MVT::i128;
19803 EVT HalfT = Regs64bit ? MVT::i64 : MVT::i32;
19804 SDValue cpInL, cpInH;
19805 cpInL = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, HalfT, N->getOperand(2),
19806 DAG.getConstant(0, dl, HalfT));
19807 cpInH = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, HalfT, N->getOperand(2),
19808 DAG.getConstant(1, dl, HalfT));
19809 cpInL = DAG.getCopyToReg(N->getOperand(0), dl,
19810 Regs64bit ? X86::RAX : X86::EAX,
19812 cpInH = DAG.getCopyToReg(cpInL.getValue(0), dl,
19813 Regs64bit ? X86::RDX : X86::EDX,
19814 cpInH, cpInL.getValue(1));
19815 SDValue swapInL, swapInH;
19816 swapInL = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, HalfT, N->getOperand(3),
19817 DAG.getConstant(0, dl, HalfT));
19818 swapInH = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, HalfT, N->getOperand(3),
19819 DAG.getConstant(1, dl, HalfT));
19820 swapInL = DAG.getCopyToReg(cpInH.getValue(0), dl,
19821 Regs64bit ? X86::RBX : X86::EBX,
19822 swapInL, cpInH.getValue(1));
19823 swapInH = DAG.getCopyToReg(swapInL.getValue(0), dl,
19824 Regs64bit ? X86::RCX : X86::ECX,
19825 swapInH, swapInL.getValue(1));
19826 SDValue Ops[] = { swapInH.getValue(0),
19828 swapInH.getValue(1) };
19829 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Glue);
19830 MachineMemOperand *MMO = cast<AtomicSDNode>(N)->getMemOperand();
19831 unsigned Opcode = Regs64bit ? X86ISD::LCMPXCHG16_DAG :
19832 X86ISD::LCMPXCHG8_DAG;
19833 SDValue Result = DAG.getMemIntrinsicNode(Opcode, dl, Tys, Ops, T, MMO);
19834 SDValue cpOutL = DAG.getCopyFromReg(Result.getValue(0), dl,
19835 Regs64bit ? X86::RAX : X86::EAX,
19836 HalfT, Result.getValue(1));
19837 SDValue cpOutH = DAG.getCopyFromReg(cpOutL.getValue(1), dl,
19838 Regs64bit ? X86::RDX : X86::EDX,
19839 HalfT, cpOutL.getValue(2));
19840 SDValue OpsF[] = { cpOutL.getValue(0), cpOutH.getValue(0)};
19842 SDValue EFLAGS = DAG.getCopyFromReg(cpOutH.getValue(1), dl, X86::EFLAGS,
19843 MVT::i32, cpOutH.getValue(2));
19845 DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
19846 DAG.getConstant(X86::COND_E, dl, MVT::i8), EFLAGS);
19847 Success = DAG.getZExtOrTrunc(Success, dl, N->getValueType(1));
19849 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, T, OpsF));
19850 Results.push_back(Success);
19851 Results.push_back(EFLAGS.getValue(1));
19854 case ISD::ATOMIC_SWAP:
19855 case ISD::ATOMIC_LOAD_ADD:
19856 case ISD::ATOMIC_LOAD_SUB:
19857 case ISD::ATOMIC_LOAD_AND:
19858 case ISD::ATOMIC_LOAD_OR:
19859 case ISD::ATOMIC_LOAD_XOR:
19860 case ISD::ATOMIC_LOAD_NAND:
19861 case ISD::ATOMIC_LOAD_MIN:
19862 case ISD::ATOMIC_LOAD_MAX:
19863 case ISD::ATOMIC_LOAD_UMIN:
19864 case ISD::ATOMIC_LOAD_UMAX:
19865 case ISD::ATOMIC_LOAD: {
19866 // Delegate to generic TypeLegalization. Situations we can really handle
19867 // should have already been dealt with by AtomicExpandPass.cpp.
19870 case ISD::BITCAST: {
19871 assert(Subtarget->hasSSE2() && "Requires at least SSE2!");
19872 EVT DstVT = N->getValueType(0);
19873 EVT SrcVT = N->getOperand(0)->getValueType(0);
19875 if (SrcVT != MVT::f64 ||
19876 (DstVT != MVT::v2i32 && DstVT != MVT::v4i16 && DstVT != MVT::v8i8))
19879 unsigned NumElts = DstVT.getVectorNumElements();
19880 EVT SVT = DstVT.getVectorElementType();
19881 EVT WiderVT = EVT::getVectorVT(*DAG.getContext(), SVT, NumElts * 2);
19882 SDValue Expanded = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
19883 MVT::v2f64, N->getOperand(0));
19884 SDValue ToVecInt = DAG.getBitcast(WiderVT, Expanded);
19886 if (ExperimentalVectorWideningLegalization) {
19887 // If we are legalizing vectors by widening, we already have the desired
19888 // legal vector type, just return it.
19889 Results.push_back(ToVecInt);
19893 SmallVector<SDValue, 8> Elts;
19894 for (unsigned i = 0, e = NumElts; i != e; ++i)
19895 Elts.push_back(DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, SVT,
19896 ToVecInt, DAG.getIntPtrConstant(i, dl)));
19898 Results.push_back(DAG.getNode(ISD::BUILD_VECTOR, dl, DstVT, Elts));
19903 const char *X86TargetLowering::getTargetNodeName(unsigned Opcode) const {
19904 switch ((X86ISD::NodeType)Opcode) {
19905 case X86ISD::FIRST_NUMBER: break;
19906 case X86ISD::BSF: return "X86ISD::BSF";
19907 case X86ISD::BSR: return "X86ISD::BSR";
19908 case X86ISD::SHLD: return "X86ISD::SHLD";
19909 case X86ISD::SHRD: return "X86ISD::SHRD";
19910 case X86ISD::FAND: return "X86ISD::FAND";
19911 case X86ISD::FANDN: return "X86ISD::FANDN";
19912 case X86ISD::FOR: return "X86ISD::FOR";
19913 case X86ISD::FXOR: return "X86ISD::FXOR";
19914 case X86ISD::FILD: return "X86ISD::FILD";
19915 case X86ISD::FILD_FLAG: return "X86ISD::FILD_FLAG";
19916 case X86ISD::FP_TO_INT16_IN_MEM: return "X86ISD::FP_TO_INT16_IN_MEM";
19917 case X86ISD::FP_TO_INT32_IN_MEM: return "X86ISD::FP_TO_INT32_IN_MEM";
19918 case X86ISD::FP_TO_INT64_IN_MEM: return "X86ISD::FP_TO_INT64_IN_MEM";
19919 case X86ISD::FLD: return "X86ISD::FLD";
19920 case X86ISD::FST: return "X86ISD::FST";
19921 case X86ISD::CALL: return "X86ISD::CALL";
19922 case X86ISD::RDTSC_DAG: return "X86ISD::RDTSC_DAG";
19923 case X86ISD::RDTSCP_DAG: return "X86ISD::RDTSCP_DAG";
19924 case X86ISD::RDPMC_DAG: return "X86ISD::RDPMC_DAG";
19925 case X86ISD::BT: return "X86ISD::BT";
19926 case X86ISD::CMP: return "X86ISD::CMP";
19927 case X86ISD::COMI: return "X86ISD::COMI";
19928 case X86ISD::UCOMI: return "X86ISD::UCOMI";
19929 case X86ISD::CMPM: return "X86ISD::CMPM";
19930 case X86ISD::CMPMU: return "X86ISD::CMPMU";
19931 case X86ISD::CMPM_RND: return "X86ISD::CMPM_RND";
19932 case X86ISD::SETCC: return "X86ISD::SETCC";
19933 case X86ISD::SETCC_CARRY: return "X86ISD::SETCC_CARRY";
19934 case X86ISD::FSETCC: return "X86ISD::FSETCC";
19935 case X86ISD::FGETSIGNx86: return "X86ISD::FGETSIGNx86";
19936 case X86ISD::CMOV: return "X86ISD::CMOV";
19937 case X86ISD::BRCOND: return "X86ISD::BRCOND";
19938 case X86ISD::RET_FLAG: return "X86ISD::RET_FLAG";
19939 case X86ISD::REP_STOS: return "X86ISD::REP_STOS";
19940 case X86ISD::REP_MOVS: return "X86ISD::REP_MOVS";
19941 case X86ISD::GlobalBaseReg: return "X86ISD::GlobalBaseReg";
19942 case X86ISD::Wrapper: return "X86ISD::Wrapper";
19943 case X86ISD::WrapperRIP: return "X86ISD::WrapperRIP";
19944 case X86ISD::MOVDQ2Q: return "X86ISD::MOVDQ2Q";
19945 case X86ISD::MMX_MOVD2W: return "X86ISD::MMX_MOVD2W";
19946 case X86ISD::MMX_MOVW2D: return "X86ISD::MMX_MOVW2D";
19947 case X86ISD::PEXTRB: return "X86ISD::PEXTRB";
19948 case X86ISD::PEXTRW: return "X86ISD::PEXTRW";
19949 case X86ISD::INSERTPS: return "X86ISD::INSERTPS";
19950 case X86ISD::PINSRB: return "X86ISD::PINSRB";
19951 case X86ISD::PINSRW: return "X86ISD::PINSRW";
19952 case X86ISD::MMX_PINSRW: return "X86ISD::MMX_PINSRW";
19953 case X86ISD::PSHUFB: return "X86ISD::PSHUFB";
19954 case X86ISD::ANDNP: return "X86ISD::ANDNP";
19955 case X86ISD::PSIGN: return "X86ISD::PSIGN";
19956 case X86ISD::BLENDI: return "X86ISD::BLENDI";
19957 case X86ISD::SHRUNKBLEND: return "X86ISD::SHRUNKBLEND";
19958 case X86ISD::ADDUS: return "X86ISD::ADDUS";
19959 case X86ISD::SUBUS: return "X86ISD::SUBUS";
19960 case X86ISD::HADD: return "X86ISD::HADD";
19961 case X86ISD::HSUB: return "X86ISD::HSUB";
19962 case X86ISD::FHADD: return "X86ISD::FHADD";
19963 case X86ISD::FHSUB: return "X86ISD::FHSUB";
19964 case X86ISD::ABS: return "X86ISD::ABS";
19965 case X86ISD::CONFLICT: return "X86ISD::CONFLICT";
19966 case X86ISD::FMAX: return "X86ISD::FMAX";
19967 case X86ISD::FMAX_RND: return "X86ISD::FMAX_RND";
19968 case X86ISD::FMIN: return "X86ISD::FMIN";
19969 case X86ISD::FMIN_RND: return "X86ISD::FMIN_RND";
19970 case X86ISD::FMAXC: return "X86ISD::FMAXC";
19971 case X86ISD::FMINC: return "X86ISD::FMINC";
19972 case X86ISD::FRSQRT: return "X86ISD::FRSQRT";
19973 case X86ISD::FRCP: return "X86ISD::FRCP";
19974 case X86ISD::EXTRQI: return "X86ISD::EXTRQI";
19975 case X86ISD::INSERTQI: return "X86ISD::INSERTQI";
19976 case X86ISD::TLSADDR: return "X86ISD::TLSADDR";
19977 case X86ISD::TLSBASEADDR: return "X86ISD::TLSBASEADDR";
19978 case X86ISD::TLSCALL: return "X86ISD::TLSCALL";
19979 case X86ISD::EH_SJLJ_SETJMP: return "X86ISD::EH_SJLJ_SETJMP";
19980 case X86ISD::EH_SJLJ_LONGJMP: return "X86ISD::EH_SJLJ_LONGJMP";
19981 case X86ISD::EH_RETURN: return "X86ISD::EH_RETURN";
19982 case X86ISD::TC_RETURN: return "X86ISD::TC_RETURN";
19983 case X86ISD::FNSTCW16m: return "X86ISD::FNSTCW16m";
19984 case X86ISD::FNSTSW16r: return "X86ISD::FNSTSW16r";
19985 case X86ISD::LCMPXCHG_DAG: return "X86ISD::LCMPXCHG_DAG";
19986 case X86ISD::LCMPXCHG8_DAG: return "X86ISD::LCMPXCHG8_DAG";
19987 case X86ISD::LCMPXCHG16_DAG: return "X86ISD::LCMPXCHG16_DAG";
19988 case X86ISD::VZEXT_MOVL: return "X86ISD::VZEXT_MOVL";
19989 case X86ISD::VZEXT_LOAD: return "X86ISD::VZEXT_LOAD";
19990 case X86ISD::VZEXT: return "X86ISD::VZEXT";
19991 case X86ISD::VSEXT: return "X86ISD::VSEXT";
19992 case X86ISD::VTRUNC: return "X86ISD::VTRUNC";
19993 case X86ISD::VTRUNCS: return "X86ISD::VTRUNCS";
19994 case X86ISD::VTRUNCUS: return "X86ISD::VTRUNCUS";
19995 case X86ISD::VINSERT: return "X86ISD::VINSERT";
19996 case X86ISD::VFPEXT: return "X86ISD::VFPEXT";
19997 case X86ISD::VFPROUND: return "X86ISD::VFPROUND";
19998 case X86ISD::CVTDQ2PD: return "X86ISD::CVTDQ2PD";
19999 case X86ISD::CVTUDQ2PD: return "X86ISD::CVTUDQ2PD";
20000 case X86ISD::VSHLDQ: return "X86ISD::VSHLDQ";
20001 case X86ISD::VSRLDQ: return "X86ISD::VSRLDQ";
20002 case X86ISD::VSHL: return "X86ISD::VSHL";
20003 case X86ISD::VSRL: return "X86ISD::VSRL";
20004 case X86ISD::VSRA: return "X86ISD::VSRA";
20005 case X86ISD::VSHLI: return "X86ISD::VSHLI";
20006 case X86ISD::VSRLI: return "X86ISD::VSRLI";
20007 case X86ISD::VSRAI: return "X86ISD::VSRAI";
20008 case X86ISD::CMPP: return "X86ISD::CMPP";
20009 case X86ISD::PCMPEQ: return "X86ISD::PCMPEQ";
20010 case X86ISD::PCMPGT: return "X86ISD::PCMPGT";
20011 case X86ISD::PCMPEQM: return "X86ISD::PCMPEQM";
20012 case X86ISD::PCMPGTM: return "X86ISD::PCMPGTM";
20013 case X86ISD::ADD: return "X86ISD::ADD";
20014 case X86ISD::SUB: return "X86ISD::SUB";
20015 case X86ISD::ADC: return "X86ISD::ADC";
20016 case X86ISD::SBB: return "X86ISD::SBB";
20017 case X86ISD::SMUL: return "X86ISD::SMUL";
20018 case X86ISD::UMUL: return "X86ISD::UMUL";
20019 case X86ISD::SMUL8: return "X86ISD::SMUL8";
20020 case X86ISD::UMUL8: return "X86ISD::UMUL8";
20021 case X86ISD::SDIVREM8_SEXT_HREG: return "X86ISD::SDIVREM8_SEXT_HREG";
20022 case X86ISD::UDIVREM8_ZEXT_HREG: return "X86ISD::UDIVREM8_ZEXT_HREG";
20023 case X86ISD::INC: return "X86ISD::INC";
20024 case X86ISD::DEC: return "X86ISD::DEC";
20025 case X86ISD::OR: return "X86ISD::OR";
20026 case X86ISD::XOR: return "X86ISD::XOR";
20027 case X86ISD::AND: return "X86ISD::AND";
20028 case X86ISD::BEXTR: return "X86ISD::BEXTR";
20029 case X86ISD::MUL_IMM: return "X86ISD::MUL_IMM";
20030 case X86ISD::PTEST: return "X86ISD::PTEST";
20031 case X86ISD::TESTP: return "X86ISD::TESTP";
20032 case X86ISD::TESTM: return "X86ISD::TESTM";
20033 case X86ISD::TESTNM: return "X86ISD::TESTNM";
20034 case X86ISD::KORTEST: return "X86ISD::KORTEST";
20035 case X86ISD::KTEST: return "X86ISD::KTEST";
20036 case X86ISD::PACKSS: return "X86ISD::PACKSS";
20037 case X86ISD::PACKUS: return "X86ISD::PACKUS";
20038 case X86ISD::PALIGNR: return "X86ISD::PALIGNR";
20039 case X86ISD::VALIGN: return "X86ISD::VALIGN";
20040 case X86ISD::PSHUFD: return "X86ISD::PSHUFD";
20041 case X86ISD::PSHUFHW: return "X86ISD::PSHUFHW";
20042 case X86ISD::PSHUFLW: return "X86ISD::PSHUFLW";
20043 case X86ISD::SHUFP: return "X86ISD::SHUFP";
20044 case X86ISD::SHUF128: return "X86ISD::SHUF128";
20045 case X86ISD::MOVLHPS: return "X86ISD::MOVLHPS";
20046 case X86ISD::MOVLHPD: return "X86ISD::MOVLHPD";
20047 case X86ISD::MOVHLPS: return "X86ISD::MOVHLPS";
20048 case X86ISD::MOVLPS: return "X86ISD::MOVLPS";
20049 case X86ISD::MOVLPD: return "X86ISD::MOVLPD";
20050 case X86ISD::MOVDDUP: return "X86ISD::MOVDDUP";
20051 case X86ISD::MOVSHDUP: return "X86ISD::MOVSHDUP";
20052 case X86ISD::MOVSLDUP: return "X86ISD::MOVSLDUP";
20053 case X86ISD::MOVSD: return "X86ISD::MOVSD";
20054 case X86ISD::MOVSS: return "X86ISD::MOVSS";
20055 case X86ISD::UNPCKL: return "X86ISD::UNPCKL";
20056 case X86ISD::UNPCKH: return "X86ISD::UNPCKH";
20057 case X86ISD::VBROADCAST: return "X86ISD::VBROADCAST";
20058 case X86ISD::SUBV_BROADCAST: return "X86ISD::SUBV_BROADCAST";
20059 case X86ISD::VEXTRACT: return "X86ISD::VEXTRACT";
20060 case X86ISD::VPERMILPV: return "X86ISD::VPERMILPV";
20061 case X86ISD::VPERMILPI: return "X86ISD::VPERMILPI";
20062 case X86ISD::VPERM2X128: return "X86ISD::VPERM2X128";
20063 case X86ISD::VPERMV: return "X86ISD::VPERMV";
20064 case X86ISD::VPERMV3: return "X86ISD::VPERMV3";
20065 case X86ISD::VPERMIV3: return "X86ISD::VPERMIV3";
20066 case X86ISD::VPERMI: return "X86ISD::VPERMI";
20067 case X86ISD::VPTERNLOG: return "X86ISD::VPTERNLOG";
20068 case X86ISD::VFIXUPIMM: return "X86ISD::VFIXUPIMM";
20069 case X86ISD::VRANGE: return "X86ISD::VRANGE";
20070 case X86ISD::PMULUDQ: return "X86ISD::PMULUDQ";
20071 case X86ISD::PMULDQ: return "X86ISD::PMULDQ";
20072 case X86ISD::PSADBW: return "X86ISD::PSADBW";
20073 case X86ISD::DBPSADBW: return "X86ISD::DBPSADBW";
20074 case X86ISD::VASTART_SAVE_XMM_REGS: return "X86ISD::VASTART_SAVE_XMM_REGS";
20075 case X86ISD::VAARG_64: return "X86ISD::VAARG_64";
20076 case X86ISD::WIN_ALLOCA: return "X86ISD::WIN_ALLOCA";
20077 case X86ISD::MEMBARRIER: return "X86ISD::MEMBARRIER";
20078 case X86ISD::MFENCE: return "X86ISD::MFENCE";
20079 case X86ISD::SFENCE: return "X86ISD::SFENCE";
20080 case X86ISD::LFENCE: return "X86ISD::LFENCE";
20081 case X86ISD::SEG_ALLOCA: return "X86ISD::SEG_ALLOCA";
20082 case X86ISD::SAHF: return "X86ISD::SAHF";
20083 case X86ISD::RDRAND: return "X86ISD::RDRAND";
20084 case X86ISD::RDSEED: return "X86ISD::RDSEED";
20085 case X86ISD::VPMADDUBSW: return "X86ISD::VPMADDUBSW";
20086 case X86ISD::VPMADDWD: return "X86ISD::VPMADDWD";
20087 case X86ISD::VPROT: return "X86ISD::VPROT";
20088 case X86ISD::VPROTI: return "X86ISD::VPROTI";
20089 case X86ISD::VPSHA: return "X86ISD::VPSHA";
20090 case X86ISD::VPSHL: return "X86ISD::VPSHL";
20091 case X86ISD::VPCOM: return "X86ISD::VPCOM";
20092 case X86ISD::VPCOMU: return "X86ISD::VPCOMU";
20093 case X86ISD::FMADD: return "X86ISD::FMADD";
20094 case X86ISD::FMSUB: return "X86ISD::FMSUB";
20095 case X86ISD::FNMADD: return "X86ISD::FNMADD";
20096 case X86ISD::FNMSUB: return "X86ISD::FNMSUB";
20097 case X86ISD::FMADDSUB: return "X86ISD::FMADDSUB";
20098 case X86ISD::FMSUBADD: return "X86ISD::FMSUBADD";
20099 case X86ISD::FMADD_RND: return "X86ISD::FMADD_RND";
20100 case X86ISD::FNMADD_RND: return "X86ISD::FNMADD_RND";
20101 case X86ISD::FMSUB_RND: return "X86ISD::FMSUB_RND";
20102 case X86ISD::FNMSUB_RND: return "X86ISD::FNMSUB_RND";
20103 case X86ISD::FMADDSUB_RND: return "X86ISD::FMADDSUB_RND";
20104 case X86ISD::FMSUBADD_RND: return "X86ISD::FMSUBADD_RND";
20105 case X86ISD::VRNDSCALE: return "X86ISD::VRNDSCALE";
20106 case X86ISD::VREDUCE: return "X86ISD::VREDUCE";
20107 case X86ISD::VGETMANT: return "X86ISD::VGETMANT";
20108 case X86ISD::PCMPESTRI: return "X86ISD::PCMPESTRI";
20109 case X86ISD::PCMPISTRI: return "X86ISD::PCMPISTRI";
20110 case X86ISD::XTEST: return "X86ISD::XTEST";
20111 case X86ISD::COMPRESS: return "X86ISD::COMPRESS";
20112 case X86ISD::EXPAND: return "X86ISD::EXPAND";
20113 case X86ISD::SELECT: return "X86ISD::SELECT";
20114 case X86ISD::ADDSUB: return "X86ISD::ADDSUB";
20115 case X86ISD::RCP28: return "X86ISD::RCP28";
20116 case X86ISD::EXP2: return "X86ISD::EXP2";
20117 case X86ISD::RSQRT28: return "X86ISD::RSQRT28";
20118 case X86ISD::FADD_RND: return "X86ISD::FADD_RND";
20119 case X86ISD::FSUB_RND: return "X86ISD::FSUB_RND";
20120 case X86ISD::FMUL_RND: return "X86ISD::FMUL_RND";
20121 case X86ISD::FDIV_RND: return "X86ISD::FDIV_RND";
20122 case X86ISD::FSQRT_RND: return "X86ISD::FSQRT_RND";
20123 case X86ISD::FGETEXP_RND: return "X86ISD::FGETEXP_RND";
20124 case X86ISD::SCALEF: return "X86ISD::SCALEF";
20125 case X86ISD::ADDS: return "X86ISD::ADDS";
20126 case X86ISD::SUBS: return "X86ISD::SUBS";
20127 case X86ISD::AVG: return "X86ISD::AVG";
20128 case X86ISD::MULHRS: return "X86ISD::MULHRS";
20129 case X86ISD::SINT_TO_FP_RND: return "X86ISD::SINT_TO_FP_RND";
20130 case X86ISD::UINT_TO_FP_RND: return "X86ISD::UINT_TO_FP_RND";
20131 case X86ISD::FP_TO_SINT_RND: return "X86ISD::FP_TO_SINT_RND";
20132 case X86ISD::FP_TO_UINT_RND: return "X86ISD::FP_TO_UINT_RND";
20133 case X86ISD::VFPCLASS: return "X86ISD::VFPCLASS";
20138 // isLegalAddressingMode - Return true if the addressing mode represented
20139 // by AM is legal for this target, for a load/store of the specified type.
20140 bool X86TargetLowering::isLegalAddressingMode(const DataLayout &DL,
20141 const AddrMode &AM, Type *Ty,
20142 unsigned AS) const {
20143 // X86 supports extremely general addressing modes.
20144 CodeModel::Model M = getTargetMachine().getCodeModel();
20145 Reloc::Model R = getTargetMachine().getRelocationModel();
20147 // X86 allows a sign-extended 32-bit immediate field as a displacement.
20148 if (!X86::isOffsetSuitableForCodeModel(AM.BaseOffs, M, AM.BaseGV != nullptr))
20153 Subtarget->ClassifyGlobalReference(AM.BaseGV, getTargetMachine());
20155 // If a reference to this global requires an extra load, we can't fold it.
20156 if (isGlobalStubReference(GVFlags))
20159 // If BaseGV requires a register for the PIC base, we cannot also have a
20160 // BaseReg specified.
20161 if (AM.HasBaseReg && isGlobalRelativeToPICBase(GVFlags))
20164 // If lower 4G is not available, then we must use rip-relative addressing.
20165 if ((M != CodeModel::Small || R != Reloc::Static) &&
20166 Subtarget->is64Bit() && (AM.BaseOffs || AM.Scale > 1))
20170 switch (AM.Scale) {
20176 // These scales always work.
20181 // These scales are formed with basereg+scalereg. Only accept if there is
20186 default: // Other stuff never works.
20193 bool X86TargetLowering::isVectorShiftByScalarCheap(Type *Ty) const {
20194 unsigned Bits = Ty->getScalarSizeInBits();
20196 // 8-bit shifts are always expensive, but versions with a scalar amount aren't
20197 // particularly cheaper than those without.
20201 // On AVX2 there are new vpsllv[dq] instructions (and other shifts), that make
20202 // variable shifts just as cheap as scalar ones.
20203 if (Subtarget->hasInt256() && (Bits == 32 || Bits == 64))
20206 // Otherwise, it's significantly cheaper to shift by a scalar amount than by a
20207 // fully general vector.
20211 bool X86TargetLowering::isTruncateFree(Type *Ty1, Type *Ty2) const {
20212 if (!Ty1->isIntegerTy() || !Ty2->isIntegerTy())
20214 unsigned NumBits1 = Ty1->getPrimitiveSizeInBits();
20215 unsigned NumBits2 = Ty2->getPrimitiveSizeInBits();
20216 return NumBits1 > NumBits2;
20219 bool X86TargetLowering::allowTruncateForTailCall(Type *Ty1, Type *Ty2) const {
20220 if (!Ty1->isIntegerTy() || !Ty2->isIntegerTy())
20223 if (!isTypeLegal(EVT::getEVT(Ty1)))
20226 assert(Ty1->getPrimitiveSizeInBits() <= 64 && "i128 is probably not a noop");
20228 // Assuming the caller doesn't have a zeroext or signext return parameter,
20229 // truncation all the way down to i1 is valid.
20233 bool X86TargetLowering::isLegalICmpImmediate(int64_t Imm) const {
20234 return isInt<32>(Imm);
20237 bool X86TargetLowering::isLegalAddImmediate(int64_t Imm) const {
20238 // Can also use sub to handle negated immediates.
20239 return isInt<32>(Imm);
20242 bool X86TargetLowering::isTruncateFree(EVT VT1, EVT VT2) const {
20243 if (!VT1.isInteger() || !VT2.isInteger())
20245 unsigned NumBits1 = VT1.getSizeInBits();
20246 unsigned NumBits2 = VT2.getSizeInBits();
20247 return NumBits1 > NumBits2;
20250 bool X86TargetLowering::isZExtFree(Type *Ty1, Type *Ty2) const {
20251 // x86-64 implicitly zero-extends 32-bit results in 64-bit registers.
20252 return Ty1->isIntegerTy(32) && Ty2->isIntegerTy(64) && Subtarget->is64Bit();
20255 bool X86TargetLowering::isZExtFree(EVT VT1, EVT VT2) const {
20256 // x86-64 implicitly zero-extends 32-bit results in 64-bit registers.
20257 return VT1 == MVT::i32 && VT2 == MVT::i64 && Subtarget->is64Bit();
20260 bool X86TargetLowering::isZExtFree(SDValue Val, EVT VT2) const {
20261 EVT VT1 = Val.getValueType();
20262 if (isZExtFree(VT1, VT2))
20265 if (Val.getOpcode() != ISD::LOAD)
20268 if (!VT1.isSimple() || !VT1.isInteger() ||
20269 !VT2.isSimple() || !VT2.isInteger())
20272 switch (VT1.getSimpleVT().SimpleTy) {
20277 // X86 has 8, 16, and 32-bit zero-extending loads.
20284 bool X86TargetLowering::isVectorLoadExtDesirable(SDValue) const { return true; }
20287 X86TargetLowering::isFMAFasterThanFMulAndFAdd(EVT VT) const {
20288 if (!(Subtarget->hasFMA() || Subtarget->hasFMA4() || Subtarget->hasAVX512()))
20291 VT = VT.getScalarType();
20293 if (!VT.isSimple())
20296 switch (VT.getSimpleVT().SimpleTy) {
20307 bool X86TargetLowering::isNarrowingProfitable(EVT VT1, EVT VT2) const {
20308 // i16 instructions are longer (0x66 prefix) and potentially slower.
20309 return !(VT1 == MVT::i32 && VT2 == MVT::i16);
20312 /// isShuffleMaskLegal - Targets can use this to indicate that they only
20313 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
20314 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
20315 /// are assumed to be legal.
20317 X86TargetLowering::isShuffleMaskLegal(const SmallVectorImpl<int> &M,
20319 if (!VT.isSimple())
20322 // Not for i1 vectors
20323 if (VT.getScalarType() == MVT::i1)
20326 // Very little shuffling can be done for 64-bit vectors right now.
20327 if (VT.getSizeInBits() == 64)
20330 // We only care that the types being shuffled are legal. The lowering can
20331 // handle any possible shuffle mask that results.
20332 return isTypeLegal(VT.getSimpleVT());
20336 X86TargetLowering::isVectorClearMaskLegal(const SmallVectorImpl<int> &Mask,
20338 // Just delegate to the generic legality, clear masks aren't special.
20339 return isShuffleMaskLegal(Mask, VT);
20342 //===----------------------------------------------------------------------===//
20343 // X86 Scheduler Hooks
20344 //===----------------------------------------------------------------------===//
20346 /// Utility function to emit xbegin specifying the start of an RTM region.
20347 static MachineBasicBlock *EmitXBegin(MachineInstr *MI, MachineBasicBlock *MBB,
20348 const TargetInstrInfo *TII) {
20349 DebugLoc DL = MI->getDebugLoc();
20351 const BasicBlock *BB = MBB->getBasicBlock();
20352 MachineFunction::iterator I = ++MBB->getIterator();
20354 // For the v = xbegin(), we generate
20365 MachineBasicBlock *thisMBB = MBB;
20366 MachineFunction *MF = MBB->getParent();
20367 MachineBasicBlock *mainMBB = MF->CreateMachineBasicBlock(BB);
20368 MachineBasicBlock *sinkMBB = MF->CreateMachineBasicBlock(BB);
20369 MF->insert(I, mainMBB);
20370 MF->insert(I, sinkMBB);
20372 // Transfer the remainder of BB and its successor edges to sinkMBB.
20373 sinkMBB->splice(sinkMBB->begin(), MBB,
20374 std::next(MachineBasicBlock::iterator(MI)), MBB->end());
20375 sinkMBB->transferSuccessorsAndUpdatePHIs(MBB);
20379 // # fallthrough to mainMBB
20380 // # abortion to sinkMBB
20381 BuildMI(thisMBB, DL, TII->get(X86::XBEGIN_4)).addMBB(sinkMBB);
20382 thisMBB->addSuccessor(mainMBB);
20383 thisMBB->addSuccessor(sinkMBB);
20387 BuildMI(mainMBB, DL, TII->get(X86::MOV32ri), X86::EAX).addImm(-1);
20388 mainMBB->addSuccessor(sinkMBB);
20391 // EAX is live into the sinkMBB
20392 sinkMBB->addLiveIn(X86::EAX);
20393 BuildMI(*sinkMBB, sinkMBB->begin(), DL,
20394 TII->get(TargetOpcode::COPY), MI->getOperand(0).getReg())
20397 MI->eraseFromParent();
20401 // FIXME: When we get size specific XMM0 registers, i.e. XMM0_V16I8
20402 // or XMM0_V32I8 in AVX all of this code can be replaced with that
20403 // in the .td file.
20404 static MachineBasicBlock *EmitPCMPSTRM(MachineInstr *MI, MachineBasicBlock *BB,
20405 const TargetInstrInfo *TII) {
20407 switch (MI->getOpcode()) {
20408 default: llvm_unreachable("illegal opcode!");
20409 case X86::PCMPISTRM128REG: Opc = X86::PCMPISTRM128rr; break;
20410 case X86::VPCMPISTRM128REG: Opc = X86::VPCMPISTRM128rr; break;
20411 case X86::PCMPISTRM128MEM: Opc = X86::PCMPISTRM128rm; break;
20412 case X86::VPCMPISTRM128MEM: Opc = X86::VPCMPISTRM128rm; break;
20413 case X86::PCMPESTRM128REG: Opc = X86::PCMPESTRM128rr; break;
20414 case X86::VPCMPESTRM128REG: Opc = X86::VPCMPESTRM128rr; break;
20415 case X86::PCMPESTRM128MEM: Opc = X86::PCMPESTRM128rm; break;
20416 case X86::VPCMPESTRM128MEM: Opc = X86::VPCMPESTRM128rm; break;
20419 DebugLoc dl = MI->getDebugLoc();
20420 MachineInstrBuilder MIB = BuildMI(*BB, MI, dl, TII->get(Opc));
20422 unsigned NumArgs = MI->getNumOperands();
20423 for (unsigned i = 1; i < NumArgs; ++i) {
20424 MachineOperand &Op = MI->getOperand(i);
20425 if (!(Op.isReg() && Op.isImplicit()))
20426 MIB.addOperand(Op);
20428 if (MI->hasOneMemOperand())
20429 MIB->setMemRefs(MI->memoperands_begin(), MI->memoperands_end());
20431 BuildMI(*BB, MI, dl,
20432 TII->get(TargetOpcode::COPY), MI->getOperand(0).getReg())
20433 .addReg(X86::XMM0);
20435 MI->eraseFromParent();
20439 // FIXME: Custom handling because TableGen doesn't support multiple implicit
20440 // defs in an instruction pattern
20441 static MachineBasicBlock *EmitPCMPSTRI(MachineInstr *MI, MachineBasicBlock *BB,
20442 const TargetInstrInfo *TII) {
20444 switch (MI->getOpcode()) {
20445 default: llvm_unreachable("illegal opcode!");
20446 case X86::PCMPISTRIREG: Opc = X86::PCMPISTRIrr; break;
20447 case X86::VPCMPISTRIREG: Opc = X86::VPCMPISTRIrr; break;
20448 case X86::PCMPISTRIMEM: Opc = X86::PCMPISTRIrm; break;
20449 case X86::VPCMPISTRIMEM: Opc = X86::VPCMPISTRIrm; break;
20450 case X86::PCMPESTRIREG: Opc = X86::PCMPESTRIrr; break;
20451 case X86::VPCMPESTRIREG: Opc = X86::VPCMPESTRIrr; break;
20452 case X86::PCMPESTRIMEM: Opc = X86::PCMPESTRIrm; break;
20453 case X86::VPCMPESTRIMEM: Opc = X86::VPCMPESTRIrm; break;
20456 DebugLoc dl = MI->getDebugLoc();
20457 MachineInstrBuilder MIB = BuildMI(*BB, MI, dl, TII->get(Opc));
20459 unsigned NumArgs = MI->getNumOperands(); // remove the results
20460 for (unsigned i = 1; i < NumArgs; ++i) {
20461 MachineOperand &Op = MI->getOperand(i);
20462 if (!(Op.isReg() && Op.isImplicit()))
20463 MIB.addOperand(Op);
20465 if (MI->hasOneMemOperand())
20466 MIB->setMemRefs(MI->memoperands_begin(), MI->memoperands_end());
20468 BuildMI(*BB, MI, dl,
20469 TII->get(TargetOpcode::COPY), MI->getOperand(0).getReg())
20472 MI->eraseFromParent();
20476 static MachineBasicBlock *EmitMonitor(MachineInstr *MI, MachineBasicBlock *BB,
20477 const X86Subtarget *Subtarget) {
20478 DebugLoc dl = MI->getDebugLoc();
20479 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
20480 // Address into RAX/EAX, other two args into ECX, EDX.
20481 unsigned MemOpc = Subtarget->is64Bit() ? X86::LEA64r : X86::LEA32r;
20482 unsigned MemReg = Subtarget->is64Bit() ? X86::RAX : X86::EAX;
20483 MachineInstrBuilder MIB = BuildMI(*BB, MI, dl, TII->get(MemOpc), MemReg);
20484 for (int i = 0; i < X86::AddrNumOperands; ++i)
20485 MIB.addOperand(MI->getOperand(i));
20487 unsigned ValOps = X86::AddrNumOperands;
20488 BuildMI(*BB, MI, dl, TII->get(TargetOpcode::COPY), X86::ECX)
20489 .addReg(MI->getOperand(ValOps).getReg());
20490 BuildMI(*BB, MI, dl, TII->get(TargetOpcode::COPY), X86::EDX)
20491 .addReg(MI->getOperand(ValOps+1).getReg());
20493 // The instruction doesn't actually take any operands though.
20494 BuildMI(*BB, MI, dl, TII->get(X86::MONITORrrr));
20496 MI->eraseFromParent(); // The pseudo is gone now.
20500 MachineBasicBlock *
20501 X86TargetLowering::EmitVAARG64WithCustomInserter(MachineInstr *MI,
20502 MachineBasicBlock *MBB) const {
20503 // Emit va_arg instruction on X86-64.
20505 // Operands to this pseudo-instruction:
20506 // 0 ) Output : destination address (reg)
20507 // 1-5) Input : va_list address (addr, i64mem)
20508 // 6 ) ArgSize : Size (in bytes) of vararg type
20509 // 7 ) ArgMode : 0=overflow only, 1=use gp_offset, 2=use fp_offset
20510 // 8 ) Align : Alignment of type
20511 // 9 ) EFLAGS (implicit-def)
20513 assert(MI->getNumOperands() == 10 && "VAARG_64 should have 10 operands!");
20514 static_assert(X86::AddrNumOperands == 5,
20515 "VAARG_64 assumes 5 address operands");
20517 unsigned DestReg = MI->getOperand(0).getReg();
20518 MachineOperand &Base = MI->getOperand(1);
20519 MachineOperand &Scale = MI->getOperand(2);
20520 MachineOperand &Index = MI->getOperand(3);
20521 MachineOperand &Disp = MI->getOperand(4);
20522 MachineOperand &Segment = MI->getOperand(5);
20523 unsigned ArgSize = MI->getOperand(6).getImm();
20524 unsigned ArgMode = MI->getOperand(7).getImm();
20525 unsigned Align = MI->getOperand(8).getImm();
20527 // Memory Reference
20528 assert(MI->hasOneMemOperand() && "Expected VAARG_64 to have one memoperand");
20529 MachineInstr::mmo_iterator MMOBegin = MI->memoperands_begin();
20530 MachineInstr::mmo_iterator MMOEnd = MI->memoperands_end();
20532 // Machine Information
20533 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
20534 MachineRegisterInfo &MRI = MBB->getParent()->getRegInfo();
20535 const TargetRegisterClass *AddrRegClass = getRegClassFor(MVT::i64);
20536 const TargetRegisterClass *OffsetRegClass = getRegClassFor(MVT::i32);
20537 DebugLoc DL = MI->getDebugLoc();
20539 // struct va_list {
20542 // i64 overflow_area (address)
20543 // i64 reg_save_area (address)
20545 // sizeof(va_list) = 24
20546 // alignment(va_list) = 8
20548 unsigned TotalNumIntRegs = 6;
20549 unsigned TotalNumXMMRegs = 8;
20550 bool UseGPOffset = (ArgMode == 1);
20551 bool UseFPOffset = (ArgMode == 2);
20552 unsigned MaxOffset = TotalNumIntRegs * 8 +
20553 (UseFPOffset ? TotalNumXMMRegs * 16 : 0);
20555 /* Align ArgSize to a multiple of 8 */
20556 unsigned ArgSizeA8 = (ArgSize + 7) & ~7;
20557 bool NeedsAlign = (Align > 8);
20559 MachineBasicBlock *thisMBB = MBB;
20560 MachineBasicBlock *overflowMBB;
20561 MachineBasicBlock *offsetMBB;
20562 MachineBasicBlock *endMBB;
20564 unsigned OffsetDestReg = 0; // Argument address computed by offsetMBB
20565 unsigned OverflowDestReg = 0; // Argument address computed by overflowMBB
20566 unsigned OffsetReg = 0;
20568 if (!UseGPOffset && !UseFPOffset) {
20569 // If we only pull from the overflow region, we don't create a branch.
20570 // We don't need to alter control flow.
20571 OffsetDestReg = 0; // unused
20572 OverflowDestReg = DestReg;
20574 offsetMBB = nullptr;
20575 overflowMBB = thisMBB;
20578 // First emit code to check if gp_offset (or fp_offset) is below the bound.
20579 // If so, pull the argument from reg_save_area. (branch to offsetMBB)
20580 // If not, pull from overflow_area. (branch to overflowMBB)
20585 // offsetMBB overflowMBB
20590 // Registers for the PHI in endMBB
20591 OffsetDestReg = MRI.createVirtualRegister(AddrRegClass);
20592 OverflowDestReg = MRI.createVirtualRegister(AddrRegClass);
20594 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
20595 MachineFunction *MF = MBB->getParent();
20596 overflowMBB = MF->CreateMachineBasicBlock(LLVM_BB);
20597 offsetMBB = MF->CreateMachineBasicBlock(LLVM_BB);
20598 endMBB = MF->CreateMachineBasicBlock(LLVM_BB);
20600 MachineFunction::iterator MBBIter = ++MBB->getIterator();
20602 // Insert the new basic blocks
20603 MF->insert(MBBIter, offsetMBB);
20604 MF->insert(MBBIter, overflowMBB);
20605 MF->insert(MBBIter, endMBB);
20607 // Transfer the remainder of MBB and its successor edges to endMBB.
20608 endMBB->splice(endMBB->begin(), thisMBB,
20609 std::next(MachineBasicBlock::iterator(MI)), thisMBB->end());
20610 endMBB->transferSuccessorsAndUpdatePHIs(thisMBB);
20612 // Make offsetMBB and overflowMBB successors of thisMBB
20613 thisMBB->addSuccessor(offsetMBB);
20614 thisMBB->addSuccessor(overflowMBB);
20616 // endMBB is a successor of both offsetMBB and overflowMBB
20617 offsetMBB->addSuccessor(endMBB);
20618 overflowMBB->addSuccessor(endMBB);
20620 // Load the offset value into a register
20621 OffsetReg = MRI.createVirtualRegister(OffsetRegClass);
20622 BuildMI(thisMBB, DL, TII->get(X86::MOV32rm), OffsetReg)
20626 .addDisp(Disp, UseFPOffset ? 4 : 0)
20627 .addOperand(Segment)
20628 .setMemRefs(MMOBegin, MMOEnd);
20630 // Check if there is enough room left to pull this argument.
20631 BuildMI(thisMBB, DL, TII->get(X86::CMP32ri))
20633 .addImm(MaxOffset + 8 - ArgSizeA8);
20635 // Branch to "overflowMBB" if offset >= max
20636 // Fall through to "offsetMBB" otherwise
20637 BuildMI(thisMBB, DL, TII->get(X86::GetCondBranchFromCond(X86::COND_AE)))
20638 .addMBB(overflowMBB);
20641 // In offsetMBB, emit code to use the reg_save_area.
20643 assert(OffsetReg != 0);
20645 // Read the reg_save_area address.
20646 unsigned RegSaveReg = MRI.createVirtualRegister(AddrRegClass);
20647 BuildMI(offsetMBB, DL, TII->get(X86::MOV64rm), RegSaveReg)
20652 .addOperand(Segment)
20653 .setMemRefs(MMOBegin, MMOEnd);
20655 // Zero-extend the offset
20656 unsigned OffsetReg64 = MRI.createVirtualRegister(AddrRegClass);
20657 BuildMI(offsetMBB, DL, TII->get(X86::SUBREG_TO_REG), OffsetReg64)
20660 .addImm(X86::sub_32bit);
20662 // Add the offset to the reg_save_area to get the final address.
20663 BuildMI(offsetMBB, DL, TII->get(X86::ADD64rr), OffsetDestReg)
20664 .addReg(OffsetReg64)
20665 .addReg(RegSaveReg);
20667 // Compute the offset for the next argument
20668 unsigned NextOffsetReg = MRI.createVirtualRegister(OffsetRegClass);
20669 BuildMI(offsetMBB, DL, TII->get(X86::ADD32ri), NextOffsetReg)
20671 .addImm(UseFPOffset ? 16 : 8);
20673 // Store it back into the va_list.
20674 BuildMI(offsetMBB, DL, TII->get(X86::MOV32mr))
20678 .addDisp(Disp, UseFPOffset ? 4 : 0)
20679 .addOperand(Segment)
20680 .addReg(NextOffsetReg)
20681 .setMemRefs(MMOBegin, MMOEnd);
20684 BuildMI(offsetMBB, DL, TII->get(X86::JMP_1))
20689 // Emit code to use overflow area
20692 // Load the overflow_area address into a register.
20693 unsigned OverflowAddrReg = MRI.createVirtualRegister(AddrRegClass);
20694 BuildMI(overflowMBB, DL, TII->get(X86::MOV64rm), OverflowAddrReg)
20699 .addOperand(Segment)
20700 .setMemRefs(MMOBegin, MMOEnd);
20702 // If we need to align it, do so. Otherwise, just copy the address
20703 // to OverflowDestReg.
20705 // Align the overflow address
20706 assert((Align & (Align-1)) == 0 && "Alignment must be a power of 2");
20707 unsigned TmpReg = MRI.createVirtualRegister(AddrRegClass);
20709 // aligned_addr = (addr + (align-1)) & ~(align-1)
20710 BuildMI(overflowMBB, DL, TII->get(X86::ADD64ri32), TmpReg)
20711 .addReg(OverflowAddrReg)
20714 BuildMI(overflowMBB, DL, TII->get(X86::AND64ri32), OverflowDestReg)
20716 .addImm(~(uint64_t)(Align-1));
20718 BuildMI(overflowMBB, DL, TII->get(TargetOpcode::COPY), OverflowDestReg)
20719 .addReg(OverflowAddrReg);
20722 // Compute the next overflow address after this argument.
20723 // (the overflow address should be kept 8-byte aligned)
20724 unsigned NextAddrReg = MRI.createVirtualRegister(AddrRegClass);
20725 BuildMI(overflowMBB, DL, TII->get(X86::ADD64ri32), NextAddrReg)
20726 .addReg(OverflowDestReg)
20727 .addImm(ArgSizeA8);
20729 // Store the new overflow address.
20730 BuildMI(overflowMBB, DL, TII->get(X86::MOV64mr))
20735 .addOperand(Segment)
20736 .addReg(NextAddrReg)
20737 .setMemRefs(MMOBegin, MMOEnd);
20739 // If we branched, emit the PHI to the front of endMBB.
20741 BuildMI(*endMBB, endMBB->begin(), DL,
20742 TII->get(X86::PHI), DestReg)
20743 .addReg(OffsetDestReg).addMBB(offsetMBB)
20744 .addReg(OverflowDestReg).addMBB(overflowMBB);
20747 // Erase the pseudo instruction
20748 MI->eraseFromParent();
20753 MachineBasicBlock *
20754 X86TargetLowering::EmitVAStartSaveXMMRegsWithCustomInserter(
20756 MachineBasicBlock *MBB) const {
20757 // Emit code to save XMM registers to the stack. The ABI says that the
20758 // number of registers to save is given in %al, so it's theoretically
20759 // possible to do an indirect jump trick to avoid saving all of them,
20760 // however this code takes a simpler approach and just executes all
20761 // of the stores if %al is non-zero. It's less code, and it's probably
20762 // easier on the hardware branch predictor, and stores aren't all that
20763 // expensive anyway.
20765 // Create the new basic blocks. One block contains all the XMM stores,
20766 // and one block is the final destination regardless of whether any
20767 // stores were performed.
20768 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
20769 MachineFunction *F = MBB->getParent();
20770 MachineFunction::iterator MBBIter = ++MBB->getIterator();
20771 MachineBasicBlock *XMMSaveMBB = F->CreateMachineBasicBlock(LLVM_BB);
20772 MachineBasicBlock *EndMBB = F->CreateMachineBasicBlock(LLVM_BB);
20773 F->insert(MBBIter, XMMSaveMBB);
20774 F->insert(MBBIter, EndMBB);
20776 // Transfer the remainder of MBB and its successor edges to EndMBB.
20777 EndMBB->splice(EndMBB->begin(), MBB,
20778 std::next(MachineBasicBlock::iterator(MI)), MBB->end());
20779 EndMBB->transferSuccessorsAndUpdatePHIs(MBB);
20781 // The original block will now fall through to the XMM save block.
20782 MBB->addSuccessor(XMMSaveMBB);
20783 // The XMMSaveMBB will fall through to the end block.
20784 XMMSaveMBB->addSuccessor(EndMBB);
20786 // Now add the instructions.
20787 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
20788 DebugLoc DL = MI->getDebugLoc();
20790 unsigned CountReg = MI->getOperand(0).getReg();
20791 int64_t RegSaveFrameIndex = MI->getOperand(1).getImm();
20792 int64_t VarArgsFPOffset = MI->getOperand(2).getImm();
20794 if (!Subtarget->isCallingConvWin64(F->getFunction()->getCallingConv())) {
20795 // If %al is 0, branch around the XMM save block.
20796 BuildMI(MBB, DL, TII->get(X86::TEST8rr)).addReg(CountReg).addReg(CountReg);
20797 BuildMI(MBB, DL, TII->get(X86::JE_1)).addMBB(EndMBB);
20798 MBB->addSuccessor(EndMBB);
20801 // Make sure the last operand is EFLAGS, which gets clobbered by the branch
20802 // that was just emitted, but clearly shouldn't be "saved".
20803 assert((MI->getNumOperands() <= 3 ||
20804 !MI->getOperand(MI->getNumOperands() - 1).isReg() ||
20805 MI->getOperand(MI->getNumOperands() - 1).getReg() == X86::EFLAGS)
20806 && "Expected last argument to be EFLAGS");
20807 unsigned MOVOpc = Subtarget->hasFp256() ? X86::VMOVAPSmr : X86::MOVAPSmr;
20808 // In the XMM save block, save all the XMM argument registers.
20809 for (int i = 3, e = MI->getNumOperands() - 1; i != e; ++i) {
20810 int64_t Offset = (i - 3) * 16 + VarArgsFPOffset;
20811 MachineMemOperand *MMO = F->getMachineMemOperand(
20812 MachinePointerInfo::getFixedStack(*F, RegSaveFrameIndex, Offset),
20813 MachineMemOperand::MOStore,
20814 /*Size=*/16, /*Align=*/16);
20815 BuildMI(XMMSaveMBB, DL, TII->get(MOVOpc))
20816 .addFrameIndex(RegSaveFrameIndex)
20817 .addImm(/*Scale=*/1)
20818 .addReg(/*IndexReg=*/0)
20819 .addImm(/*Disp=*/Offset)
20820 .addReg(/*Segment=*/0)
20821 .addReg(MI->getOperand(i).getReg())
20822 .addMemOperand(MMO);
20825 MI->eraseFromParent(); // The pseudo instruction is gone now.
20830 // The EFLAGS operand of SelectItr might be missing a kill marker
20831 // because there were multiple uses of EFLAGS, and ISel didn't know
20832 // which to mark. Figure out whether SelectItr should have had a
20833 // kill marker, and set it if it should. Returns the correct kill
20835 static bool checkAndUpdateEFLAGSKill(MachineBasicBlock::iterator SelectItr,
20836 MachineBasicBlock* BB,
20837 const TargetRegisterInfo* TRI) {
20838 // Scan forward through BB for a use/def of EFLAGS.
20839 MachineBasicBlock::iterator miI(std::next(SelectItr));
20840 for (MachineBasicBlock::iterator miE = BB->end(); miI != miE; ++miI) {
20841 const MachineInstr& mi = *miI;
20842 if (mi.readsRegister(X86::EFLAGS))
20844 if (mi.definesRegister(X86::EFLAGS))
20845 break; // Should have kill-flag - update below.
20848 // If we hit the end of the block, check whether EFLAGS is live into a
20850 if (miI == BB->end()) {
20851 for (MachineBasicBlock::succ_iterator sItr = BB->succ_begin(),
20852 sEnd = BB->succ_end();
20853 sItr != sEnd; ++sItr) {
20854 MachineBasicBlock* succ = *sItr;
20855 if (succ->isLiveIn(X86::EFLAGS))
20860 // We found a def, or hit the end of the basic block and EFLAGS wasn't live
20861 // out. SelectMI should have a kill flag on EFLAGS.
20862 SelectItr->addRegisterKilled(X86::EFLAGS, TRI);
20866 // Return true if it is OK for this CMOV pseudo-opcode to be cascaded
20867 // together with other CMOV pseudo-opcodes into a single basic-block with
20868 // conditional jump around it.
20869 static bool isCMOVPseudo(MachineInstr *MI) {
20870 switch (MI->getOpcode()) {
20871 case X86::CMOV_FR32:
20872 case X86::CMOV_FR64:
20873 case X86::CMOV_GR8:
20874 case X86::CMOV_GR16:
20875 case X86::CMOV_GR32:
20876 case X86::CMOV_RFP32:
20877 case X86::CMOV_RFP64:
20878 case X86::CMOV_RFP80:
20879 case X86::CMOV_V2F64:
20880 case X86::CMOV_V2I64:
20881 case X86::CMOV_V4F32:
20882 case X86::CMOV_V4F64:
20883 case X86::CMOV_V4I64:
20884 case X86::CMOV_V16F32:
20885 case X86::CMOV_V8F32:
20886 case X86::CMOV_V8F64:
20887 case X86::CMOV_V8I64:
20888 case X86::CMOV_V8I1:
20889 case X86::CMOV_V16I1:
20890 case X86::CMOV_V32I1:
20891 case X86::CMOV_V64I1:
20899 MachineBasicBlock *
20900 X86TargetLowering::EmitLoweredSelect(MachineInstr *MI,
20901 MachineBasicBlock *BB) const {
20902 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
20903 DebugLoc DL = MI->getDebugLoc();
20905 // To "insert" a SELECT_CC instruction, we actually have to insert the
20906 // diamond control-flow pattern. The incoming instruction knows the
20907 // destination vreg to set, the condition code register to branch on, the
20908 // true/false values to select between, and a branch opcode to use.
20909 const BasicBlock *LLVM_BB = BB->getBasicBlock();
20910 MachineFunction::iterator It = ++BB->getIterator();
20915 // cmpTY ccX, r1, r2
20917 // fallthrough --> copy0MBB
20918 MachineBasicBlock *thisMBB = BB;
20919 MachineFunction *F = BB->getParent();
20921 // This code lowers all pseudo-CMOV instructions. Generally it lowers these
20922 // as described above, by inserting a BB, and then making a PHI at the join
20923 // point to select the true and false operands of the CMOV in the PHI.
20925 // The code also handles two different cases of multiple CMOV opcodes
20929 // In this case, there are multiple CMOVs in a row, all which are based on
20930 // the same condition setting (or the exact opposite condition setting).
20931 // In this case we can lower all the CMOVs using a single inserted BB, and
20932 // then make a number of PHIs at the join point to model the CMOVs. The only
20933 // trickiness here, is that in a case like:
20935 // t2 = CMOV cond1 t1, f1
20936 // t3 = CMOV cond1 t2, f2
20938 // when rewriting this into PHIs, we have to perform some renaming on the
20939 // temps since you cannot have a PHI operand refer to a PHI result earlier
20940 // in the same block. The "simple" but wrong lowering would be:
20942 // t2 = PHI t1(BB1), f1(BB2)
20943 // t3 = PHI t2(BB1), f2(BB2)
20945 // but clearly t2 is not defined in BB1, so that is incorrect. The proper
20946 // renaming is to note that on the path through BB1, t2 is really just a
20947 // copy of t1, and do that renaming, properly generating:
20949 // t2 = PHI t1(BB1), f1(BB2)
20950 // t3 = PHI t1(BB1), f2(BB2)
20952 // Case 2, we lower cascaded CMOVs such as
20954 // (CMOV (CMOV F, T, cc1), T, cc2)
20956 // to two successives branches. For that, we look for another CMOV as the
20957 // following instruction.
20959 // Without this, we would add a PHI between the two jumps, which ends up
20960 // creating a few copies all around. For instance, for
20962 // (sitofp (zext (fcmp une)))
20964 // we would generate:
20966 // ucomiss %xmm1, %xmm0
20967 // movss <1.0f>, %xmm0
20968 // movaps %xmm0, %xmm1
20970 // xorps %xmm1, %xmm1
20973 // movaps %xmm1, %xmm0
20977 // because this custom-inserter would have generated:
20989 // A: X = ...; Y = ...
20991 // C: Z = PHI [X, A], [Y, B]
20993 // E: PHI [X, C], [Z, D]
20995 // If we lower both CMOVs in a single step, we can instead generate:
21007 // A: X = ...; Y = ...
21009 // E: PHI [X, A], [X, C], [Y, D]
21011 // Which, in our sitofp/fcmp example, gives us something like:
21013 // ucomiss %xmm1, %xmm0
21014 // movss <1.0f>, %xmm0
21017 // xorps %xmm0, %xmm0
21021 MachineInstr *CascadedCMOV = nullptr;
21022 MachineInstr *LastCMOV = MI;
21023 X86::CondCode CC = X86::CondCode(MI->getOperand(3).getImm());
21024 X86::CondCode OppCC = X86::GetOppositeBranchCondition(CC);
21025 MachineBasicBlock::iterator NextMIIt =
21026 std::next(MachineBasicBlock::iterator(MI));
21028 // Check for case 1, where there are multiple CMOVs with the same condition
21029 // first. Of the two cases of multiple CMOV lowerings, case 1 reduces the
21030 // number of jumps the most.
21032 if (isCMOVPseudo(MI)) {
21033 // See if we have a string of CMOVS with the same condition.
21034 while (NextMIIt != BB->end() &&
21035 isCMOVPseudo(NextMIIt) &&
21036 (NextMIIt->getOperand(3).getImm() == CC ||
21037 NextMIIt->getOperand(3).getImm() == OppCC)) {
21038 LastCMOV = &*NextMIIt;
21043 // This checks for case 2, but only do this if we didn't already find
21044 // case 1, as indicated by LastCMOV == MI.
21045 if (LastCMOV == MI &&
21046 NextMIIt != BB->end() && NextMIIt->getOpcode() == MI->getOpcode() &&
21047 NextMIIt->getOperand(2).getReg() == MI->getOperand(2).getReg() &&
21048 NextMIIt->getOperand(1).getReg() == MI->getOperand(0).getReg()) {
21049 CascadedCMOV = &*NextMIIt;
21052 MachineBasicBlock *jcc1MBB = nullptr;
21054 // If we have a cascaded CMOV, we lower it to two successive branches to
21055 // the same block. EFLAGS is used by both, so mark it as live in the second.
21056 if (CascadedCMOV) {
21057 jcc1MBB = F->CreateMachineBasicBlock(LLVM_BB);
21058 F->insert(It, jcc1MBB);
21059 jcc1MBB->addLiveIn(X86::EFLAGS);
21062 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
21063 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
21064 F->insert(It, copy0MBB);
21065 F->insert(It, sinkMBB);
21067 // If the EFLAGS register isn't dead in the terminator, then claim that it's
21068 // live into the sink and copy blocks.
21069 const TargetRegisterInfo *TRI = Subtarget->getRegisterInfo();
21071 MachineInstr *LastEFLAGSUser = CascadedCMOV ? CascadedCMOV : LastCMOV;
21072 if (!LastEFLAGSUser->killsRegister(X86::EFLAGS) &&
21073 !checkAndUpdateEFLAGSKill(LastEFLAGSUser, BB, TRI)) {
21074 copy0MBB->addLiveIn(X86::EFLAGS);
21075 sinkMBB->addLiveIn(X86::EFLAGS);
21078 // Transfer the remainder of BB and its successor edges to sinkMBB.
21079 sinkMBB->splice(sinkMBB->begin(), BB,
21080 std::next(MachineBasicBlock::iterator(LastCMOV)), BB->end());
21081 sinkMBB->transferSuccessorsAndUpdatePHIs(BB);
21083 // Add the true and fallthrough blocks as its successors.
21084 if (CascadedCMOV) {
21085 // The fallthrough block may be jcc1MBB, if we have a cascaded CMOV.
21086 BB->addSuccessor(jcc1MBB);
21088 // In that case, jcc1MBB will itself fallthrough the copy0MBB, and
21089 // jump to the sinkMBB.
21090 jcc1MBB->addSuccessor(copy0MBB);
21091 jcc1MBB->addSuccessor(sinkMBB);
21093 BB->addSuccessor(copy0MBB);
21096 // The true block target of the first (or only) branch is always sinkMBB.
21097 BB->addSuccessor(sinkMBB);
21099 // Create the conditional branch instruction.
21100 unsigned Opc = X86::GetCondBranchFromCond(CC);
21101 BuildMI(BB, DL, TII->get(Opc)).addMBB(sinkMBB);
21103 if (CascadedCMOV) {
21104 unsigned Opc2 = X86::GetCondBranchFromCond(
21105 (X86::CondCode)CascadedCMOV->getOperand(3).getImm());
21106 BuildMI(jcc1MBB, DL, TII->get(Opc2)).addMBB(sinkMBB);
21110 // %FalseValue = ...
21111 // # fallthrough to sinkMBB
21112 copy0MBB->addSuccessor(sinkMBB);
21115 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
21117 MachineBasicBlock::iterator MIItBegin = MachineBasicBlock::iterator(MI);
21118 MachineBasicBlock::iterator MIItEnd =
21119 std::next(MachineBasicBlock::iterator(LastCMOV));
21120 MachineBasicBlock::iterator SinkInsertionPoint = sinkMBB->begin();
21121 DenseMap<unsigned, std::pair<unsigned, unsigned>> RegRewriteTable;
21122 MachineInstrBuilder MIB;
21124 // As we are creating the PHIs, we have to be careful if there is more than
21125 // one. Later CMOVs may reference the results of earlier CMOVs, but later
21126 // PHIs have to reference the individual true/false inputs from earlier PHIs.
21127 // That also means that PHI construction must work forward from earlier to
21128 // later, and that the code must maintain a mapping from earlier PHI's
21129 // destination registers, and the registers that went into the PHI.
21131 for (MachineBasicBlock::iterator MIIt = MIItBegin; MIIt != MIItEnd; ++MIIt) {
21132 unsigned DestReg = MIIt->getOperand(0).getReg();
21133 unsigned Op1Reg = MIIt->getOperand(1).getReg();
21134 unsigned Op2Reg = MIIt->getOperand(2).getReg();
21136 // If this CMOV we are generating is the opposite condition from
21137 // the jump we generated, then we have to swap the operands for the
21138 // PHI that is going to be generated.
21139 if (MIIt->getOperand(3).getImm() == OppCC)
21140 std::swap(Op1Reg, Op2Reg);
21142 if (RegRewriteTable.find(Op1Reg) != RegRewriteTable.end())
21143 Op1Reg = RegRewriteTable[Op1Reg].first;
21145 if (RegRewriteTable.find(Op2Reg) != RegRewriteTable.end())
21146 Op2Reg = RegRewriteTable[Op2Reg].second;
21148 MIB = BuildMI(*sinkMBB, SinkInsertionPoint, DL,
21149 TII->get(X86::PHI), DestReg)
21150 .addReg(Op1Reg).addMBB(copy0MBB)
21151 .addReg(Op2Reg).addMBB(thisMBB);
21153 // Add this PHI to the rewrite table.
21154 RegRewriteTable[DestReg] = std::make_pair(Op1Reg, Op2Reg);
21157 // If we have a cascaded CMOV, the second Jcc provides the same incoming
21158 // value as the first Jcc (the True operand of the SELECT_CC/CMOV nodes).
21159 if (CascadedCMOV) {
21160 MIB.addReg(MI->getOperand(2).getReg()).addMBB(jcc1MBB);
21161 // Copy the PHI result to the register defined by the second CMOV.
21162 BuildMI(*sinkMBB, std::next(MachineBasicBlock::iterator(MIB.getInstr())),
21163 DL, TII->get(TargetOpcode::COPY),
21164 CascadedCMOV->getOperand(0).getReg())
21165 .addReg(MI->getOperand(0).getReg());
21166 CascadedCMOV->eraseFromParent();
21169 // Now remove the CMOV(s).
21170 for (MachineBasicBlock::iterator MIIt = MIItBegin; MIIt != MIItEnd; )
21171 (MIIt++)->eraseFromParent();
21176 MachineBasicBlock *
21177 X86TargetLowering::EmitLoweredAtomicFP(MachineInstr *MI,
21178 MachineBasicBlock *BB) const {
21179 // Combine the following atomic floating-point modification pattern:
21180 // a.store(reg OP a.load(acquire), release)
21181 // Transform them into:
21182 // OPss (%gpr), %xmm
21183 // movss %xmm, (%gpr)
21184 // Or sd equivalent for 64-bit operations.
21186 switch (MI->getOpcode()) {
21187 default: llvm_unreachable("unexpected instr type for EmitLoweredAtomicFP");
21188 case X86::RELEASE_FADD32mr: MOp = X86::MOVSSmr; FOp = X86::ADDSSrm; break;
21189 case X86::RELEASE_FADD64mr: MOp = X86::MOVSDmr; FOp = X86::ADDSDrm; break;
21191 const X86InstrInfo *TII = Subtarget->getInstrInfo();
21192 DebugLoc DL = MI->getDebugLoc();
21193 MachineRegisterInfo &MRI = BB->getParent()->getRegInfo();
21194 MachineOperand MSrc = MI->getOperand(0);
21195 unsigned VSrc = MI->getOperand(5).getReg();
21196 const MachineOperand &Disp = MI->getOperand(3);
21197 MachineOperand ZeroDisp = MachineOperand::CreateImm(0);
21198 bool hasDisp = Disp.isGlobal() || Disp.isImm();
21199 if (hasDisp && MSrc.isReg())
21200 MSrc.setIsKill(false);
21201 MachineInstrBuilder MIM = BuildMI(*BB, MI, DL, TII->get(MOp))
21202 .addOperand(/*Base=*/MSrc)
21203 .addImm(/*Scale=*/1)
21204 .addReg(/*Index=*/0)
21205 .addDisp(hasDisp ? Disp : ZeroDisp, /*off=*/0)
21207 MachineInstr *MIO = BuildMI(*BB, (MachineInstr *)MIM, DL, TII->get(FOp),
21208 MRI.createVirtualRegister(MRI.getRegClass(VSrc)))
21210 .addOperand(/*Base=*/MSrc)
21211 .addImm(/*Scale=*/1)
21212 .addReg(/*Index=*/0)
21213 .addDisp(hasDisp ? Disp : ZeroDisp, /*off=*/0)
21214 .addReg(/*Segment=*/0);
21215 MIM.addReg(MIO->getOperand(0).getReg(), RegState::Kill);
21216 MI->eraseFromParent(); // The pseudo instruction is gone now.
21220 MachineBasicBlock *
21221 X86TargetLowering::EmitLoweredSegAlloca(MachineInstr *MI,
21222 MachineBasicBlock *BB) const {
21223 MachineFunction *MF = BB->getParent();
21224 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
21225 DebugLoc DL = MI->getDebugLoc();
21226 const BasicBlock *LLVM_BB = BB->getBasicBlock();
21228 assert(MF->shouldSplitStack());
21230 const bool Is64Bit = Subtarget->is64Bit();
21231 const bool IsLP64 = Subtarget->isTarget64BitLP64();
21233 const unsigned TlsReg = Is64Bit ? X86::FS : X86::GS;
21234 const unsigned TlsOffset = IsLP64 ? 0x70 : Is64Bit ? 0x40 : 0x30;
21237 // ... [Till the alloca]
21238 // If stacklet is not large enough, jump to mallocMBB
21241 // Allocate by subtracting from RSP
21242 // Jump to continueMBB
21245 // Allocate by call to runtime
21249 // [rest of original BB]
21252 MachineBasicBlock *mallocMBB = MF->CreateMachineBasicBlock(LLVM_BB);
21253 MachineBasicBlock *bumpMBB = MF->CreateMachineBasicBlock(LLVM_BB);
21254 MachineBasicBlock *continueMBB = MF->CreateMachineBasicBlock(LLVM_BB);
21256 MachineRegisterInfo &MRI = MF->getRegInfo();
21257 const TargetRegisterClass *AddrRegClass =
21258 getRegClassFor(getPointerTy(MF->getDataLayout()));
21260 unsigned mallocPtrVReg = MRI.createVirtualRegister(AddrRegClass),
21261 bumpSPPtrVReg = MRI.createVirtualRegister(AddrRegClass),
21262 tmpSPVReg = MRI.createVirtualRegister(AddrRegClass),
21263 SPLimitVReg = MRI.createVirtualRegister(AddrRegClass),
21264 sizeVReg = MI->getOperand(1).getReg(),
21265 physSPReg = IsLP64 || Subtarget->isTargetNaCl64() ? X86::RSP : X86::ESP;
21267 MachineFunction::iterator MBBIter = ++BB->getIterator();
21269 MF->insert(MBBIter, bumpMBB);
21270 MF->insert(MBBIter, mallocMBB);
21271 MF->insert(MBBIter, continueMBB);
21273 continueMBB->splice(continueMBB->begin(), BB,
21274 std::next(MachineBasicBlock::iterator(MI)), BB->end());
21275 continueMBB->transferSuccessorsAndUpdatePHIs(BB);
21277 // Add code to the main basic block to check if the stack limit has been hit,
21278 // and if so, jump to mallocMBB otherwise to bumpMBB.
21279 BuildMI(BB, DL, TII->get(TargetOpcode::COPY), tmpSPVReg).addReg(physSPReg);
21280 BuildMI(BB, DL, TII->get(IsLP64 ? X86::SUB64rr:X86::SUB32rr), SPLimitVReg)
21281 .addReg(tmpSPVReg).addReg(sizeVReg);
21282 BuildMI(BB, DL, TII->get(IsLP64 ? X86::CMP64mr:X86::CMP32mr))
21283 .addReg(0).addImm(1).addReg(0).addImm(TlsOffset).addReg(TlsReg)
21284 .addReg(SPLimitVReg);
21285 BuildMI(BB, DL, TII->get(X86::JG_1)).addMBB(mallocMBB);
21287 // bumpMBB simply decreases the stack pointer, since we know the current
21288 // stacklet has enough space.
21289 BuildMI(bumpMBB, DL, TII->get(TargetOpcode::COPY), physSPReg)
21290 .addReg(SPLimitVReg);
21291 BuildMI(bumpMBB, DL, TII->get(TargetOpcode::COPY), bumpSPPtrVReg)
21292 .addReg(SPLimitVReg);
21293 BuildMI(bumpMBB, DL, TII->get(X86::JMP_1)).addMBB(continueMBB);
21295 // Calls into a routine in libgcc to allocate more space from the heap.
21296 const uint32_t *RegMask =
21297 Subtarget->getRegisterInfo()->getCallPreservedMask(*MF, CallingConv::C);
21299 BuildMI(mallocMBB, DL, TII->get(X86::MOV64rr), X86::RDI)
21301 BuildMI(mallocMBB, DL, TII->get(X86::CALL64pcrel32))
21302 .addExternalSymbol("__morestack_allocate_stack_space")
21303 .addRegMask(RegMask)
21304 .addReg(X86::RDI, RegState::Implicit)
21305 .addReg(X86::RAX, RegState::ImplicitDefine);
21306 } else if (Is64Bit) {
21307 BuildMI(mallocMBB, DL, TII->get(X86::MOV32rr), X86::EDI)
21309 BuildMI(mallocMBB, DL, TII->get(X86::CALL64pcrel32))
21310 .addExternalSymbol("__morestack_allocate_stack_space")
21311 .addRegMask(RegMask)
21312 .addReg(X86::EDI, RegState::Implicit)
21313 .addReg(X86::EAX, RegState::ImplicitDefine);
21315 BuildMI(mallocMBB, DL, TII->get(X86::SUB32ri), physSPReg).addReg(physSPReg)
21317 BuildMI(mallocMBB, DL, TII->get(X86::PUSH32r)).addReg(sizeVReg);
21318 BuildMI(mallocMBB, DL, TII->get(X86::CALLpcrel32))
21319 .addExternalSymbol("__morestack_allocate_stack_space")
21320 .addRegMask(RegMask)
21321 .addReg(X86::EAX, RegState::ImplicitDefine);
21325 BuildMI(mallocMBB, DL, TII->get(X86::ADD32ri), physSPReg).addReg(physSPReg)
21328 BuildMI(mallocMBB, DL, TII->get(TargetOpcode::COPY), mallocPtrVReg)
21329 .addReg(IsLP64 ? X86::RAX : X86::EAX);
21330 BuildMI(mallocMBB, DL, TII->get(X86::JMP_1)).addMBB(continueMBB);
21332 // Set up the CFG correctly.
21333 BB->addSuccessor(bumpMBB);
21334 BB->addSuccessor(mallocMBB);
21335 mallocMBB->addSuccessor(continueMBB);
21336 bumpMBB->addSuccessor(continueMBB);
21338 // Take care of the PHI nodes.
21339 BuildMI(*continueMBB, continueMBB->begin(), DL, TII->get(X86::PHI),
21340 MI->getOperand(0).getReg())
21341 .addReg(mallocPtrVReg).addMBB(mallocMBB)
21342 .addReg(bumpSPPtrVReg).addMBB(bumpMBB);
21344 // Delete the original pseudo instruction.
21345 MI->eraseFromParent();
21348 return continueMBB;
21351 MachineBasicBlock *
21352 X86TargetLowering::EmitLoweredWinAlloca(MachineInstr *MI,
21353 MachineBasicBlock *BB) const {
21354 DebugLoc DL = MI->getDebugLoc();
21356 assert(!Subtarget->isTargetMachO());
21358 Subtarget->getFrameLowering()->emitStackProbeCall(*BB->getParent(), *BB, MI,
21361 MI->eraseFromParent(); // The pseudo instruction is gone now.
21365 MachineBasicBlock *
21366 X86TargetLowering::EmitLoweredTLSCall(MachineInstr *MI,
21367 MachineBasicBlock *BB) const {
21368 // This is pretty easy. We're taking the value that we received from
21369 // our load from the relocation, sticking it in either RDI (x86-64)
21370 // or EAX and doing an indirect call. The return value will then
21371 // be in the normal return register.
21372 MachineFunction *F = BB->getParent();
21373 const X86InstrInfo *TII = Subtarget->getInstrInfo();
21374 DebugLoc DL = MI->getDebugLoc();
21376 assert(Subtarget->isTargetDarwin() && "Darwin only instr emitted?");
21377 assert(MI->getOperand(3).isGlobal() && "This should be a global");
21379 // Get a register mask for the lowered call.
21380 // FIXME: The 32-bit calls have non-standard calling conventions. Use a
21381 // proper register mask.
21382 const uint32_t *RegMask =
21383 Subtarget->getRegisterInfo()->getCallPreservedMask(*F, CallingConv::C);
21384 if (Subtarget->is64Bit()) {
21385 MachineInstrBuilder MIB = BuildMI(*BB, MI, DL,
21386 TII->get(X86::MOV64rm), X86::RDI)
21388 .addImm(0).addReg(0)
21389 .addGlobalAddress(MI->getOperand(3).getGlobal(), 0,
21390 MI->getOperand(3).getTargetFlags())
21392 MIB = BuildMI(*BB, MI, DL, TII->get(X86::CALL64m));
21393 addDirectMem(MIB, X86::RDI);
21394 MIB.addReg(X86::RAX, RegState::ImplicitDefine).addRegMask(RegMask);
21395 } else if (F->getTarget().getRelocationModel() != Reloc::PIC_) {
21396 MachineInstrBuilder MIB = BuildMI(*BB, MI, DL,
21397 TII->get(X86::MOV32rm), X86::EAX)
21399 .addImm(0).addReg(0)
21400 .addGlobalAddress(MI->getOperand(3).getGlobal(), 0,
21401 MI->getOperand(3).getTargetFlags())
21403 MIB = BuildMI(*BB, MI, DL, TII->get(X86::CALL32m));
21404 addDirectMem(MIB, X86::EAX);
21405 MIB.addReg(X86::EAX, RegState::ImplicitDefine).addRegMask(RegMask);
21407 MachineInstrBuilder MIB = BuildMI(*BB, MI, DL,
21408 TII->get(X86::MOV32rm), X86::EAX)
21409 .addReg(TII->getGlobalBaseReg(F))
21410 .addImm(0).addReg(0)
21411 .addGlobalAddress(MI->getOperand(3).getGlobal(), 0,
21412 MI->getOperand(3).getTargetFlags())
21414 MIB = BuildMI(*BB, MI, DL, TII->get(X86::CALL32m));
21415 addDirectMem(MIB, X86::EAX);
21416 MIB.addReg(X86::EAX, RegState::ImplicitDefine).addRegMask(RegMask);
21419 MI->eraseFromParent(); // The pseudo instruction is gone now.
21423 MachineBasicBlock *
21424 X86TargetLowering::emitEHSjLjSetJmp(MachineInstr *MI,
21425 MachineBasicBlock *MBB) const {
21426 DebugLoc DL = MI->getDebugLoc();
21427 MachineFunction *MF = MBB->getParent();
21428 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
21429 MachineRegisterInfo &MRI = MF->getRegInfo();
21431 const BasicBlock *BB = MBB->getBasicBlock();
21432 MachineFunction::iterator I = ++MBB->getIterator();
21434 // Memory Reference
21435 MachineInstr::mmo_iterator MMOBegin = MI->memoperands_begin();
21436 MachineInstr::mmo_iterator MMOEnd = MI->memoperands_end();
21439 unsigned MemOpndSlot = 0;
21441 unsigned CurOp = 0;
21443 DstReg = MI->getOperand(CurOp++).getReg();
21444 const TargetRegisterClass *RC = MRI.getRegClass(DstReg);
21445 assert(RC->hasType(MVT::i32) && "Invalid destination!");
21446 unsigned mainDstReg = MRI.createVirtualRegister(RC);
21447 unsigned restoreDstReg = MRI.createVirtualRegister(RC);
21449 MemOpndSlot = CurOp;
21451 MVT PVT = getPointerTy(MF->getDataLayout());
21452 assert((PVT == MVT::i64 || PVT == MVT::i32) &&
21453 "Invalid Pointer Size!");
21455 // For v = setjmp(buf), we generate
21458 // buf[LabelOffset] = restoreMBB <-- takes address of restoreMBB
21459 // SjLjSetup restoreMBB
21465 // v = phi(main, restore)
21468 // if base pointer being used, load it from frame
21471 MachineBasicBlock *thisMBB = MBB;
21472 MachineBasicBlock *mainMBB = MF->CreateMachineBasicBlock(BB);
21473 MachineBasicBlock *sinkMBB = MF->CreateMachineBasicBlock(BB);
21474 MachineBasicBlock *restoreMBB = MF->CreateMachineBasicBlock(BB);
21475 MF->insert(I, mainMBB);
21476 MF->insert(I, sinkMBB);
21477 MF->push_back(restoreMBB);
21478 restoreMBB->setHasAddressTaken();
21480 MachineInstrBuilder MIB;
21482 // Transfer the remainder of BB and its successor edges to sinkMBB.
21483 sinkMBB->splice(sinkMBB->begin(), MBB,
21484 std::next(MachineBasicBlock::iterator(MI)), MBB->end());
21485 sinkMBB->transferSuccessorsAndUpdatePHIs(MBB);
21488 unsigned PtrStoreOpc = 0;
21489 unsigned LabelReg = 0;
21490 const int64_t LabelOffset = 1 * PVT.getStoreSize();
21491 Reloc::Model RM = MF->getTarget().getRelocationModel();
21492 bool UseImmLabel = (MF->getTarget().getCodeModel() == CodeModel::Small) &&
21493 (RM == Reloc::Static || RM == Reloc::DynamicNoPIC);
21495 // Prepare IP either in reg or imm.
21496 if (!UseImmLabel) {
21497 PtrStoreOpc = (PVT == MVT::i64) ? X86::MOV64mr : X86::MOV32mr;
21498 const TargetRegisterClass *PtrRC = getRegClassFor(PVT);
21499 LabelReg = MRI.createVirtualRegister(PtrRC);
21500 if (Subtarget->is64Bit()) {
21501 MIB = BuildMI(*thisMBB, MI, DL, TII->get(X86::LEA64r), LabelReg)
21505 .addMBB(restoreMBB)
21508 const X86InstrInfo *XII = static_cast<const X86InstrInfo*>(TII);
21509 MIB = BuildMI(*thisMBB, MI, DL, TII->get(X86::LEA32r), LabelReg)
21510 .addReg(XII->getGlobalBaseReg(MF))
21513 .addMBB(restoreMBB, Subtarget->ClassifyBlockAddressReference())
21517 PtrStoreOpc = (PVT == MVT::i64) ? X86::MOV64mi32 : X86::MOV32mi;
21519 MIB = BuildMI(*thisMBB, MI, DL, TII->get(PtrStoreOpc));
21520 for (unsigned i = 0; i < X86::AddrNumOperands; ++i) {
21521 if (i == X86::AddrDisp)
21522 MIB.addDisp(MI->getOperand(MemOpndSlot + i), LabelOffset);
21524 MIB.addOperand(MI->getOperand(MemOpndSlot + i));
21527 MIB.addReg(LabelReg);
21529 MIB.addMBB(restoreMBB);
21530 MIB.setMemRefs(MMOBegin, MMOEnd);
21532 MIB = BuildMI(*thisMBB, MI, DL, TII->get(X86::EH_SjLj_Setup))
21533 .addMBB(restoreMBB);
21535 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
21536 MIB.addRegMask(RegInfo->getNoPreservedMask());
21537 thisMBB->addSuccessor(mainMBB);
21538 thisMBB->addSuccessor(restoreMBB);
21542 BuildMI(mainMBB, DL, TII->get(X86::MOV32r0), mainDstReg);
21543 mainMBB->addSuccessor(sinkMBB);
21546 BuildMI(*sinkMBB, sinkMBB->begin(), DL,
21547 TII->get(X86::PHI), DstReg)
21548 .addReg(mainDstReg).addMBB(mainMBB)
21549 .addReg(restoreDstReg).addMBB(restoreMBB);
21552 if (RegInfo->hasBasePointer(*MF)) {
21553 const bool Uses64BitFramePtr =
21554 Subtarget->isTarget64BitLP64() || Subtarget->isTargetNaCl64();
21555 X86MachineFunctionInfo *X86FI = MF->getInfo<X86MachineFunctionInfo>();
21556 X86FI->setRestoreBasePointer(MF);
21557 unsigned FramePtr = RegInfo->getFrameRegister(*MF);
21558 unsigned BasePtr = RegInfo->getBaseRegister();
21559 unsigned Opm = Uses64BitFramePtr ? X86::MOV64rm : X86::MOV32rm;
21560 addRegOffset(BuildMI(restoreMBB, DL, TII->get(Opm), BasePtr),
21561 FramePtr, true, X86FI->getRestoreBasePointerOffset())
21562 .setMIFlag(MachineInstr::FrameSetup);
21564 BuildMI(restoreMBB, DL, TII->get(X86::MOV32ri), restoreDstReg).addImm(1);
21565 BuildMI(restoreMBB, DL, TII->get(X86::JMP_1)).addMBB(sinkMBB);
21566 restoreMBB->addSuccessor(sinkMBB);
21568 MI->eraseFromParent();
21572 MachineBasicBlock *
21573 X86TargetLowering::emitEHSjLjLongJmp(MachineInstr *MI,
21574 MachineBasicBlock *MBB) const {
21575 DebugLoc DL = MI->getDebugLoc();
21576 MachineFunction *MF = MBB->getParent();
21577 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
21578 MachineRegisterInfo &MRI = MF->getRegInfo();
21580 // Memory Reference
21581 MachineInstr::mmo_iterator MMOBegin = MI->memoperands_begin();
21582 MachineInstr::mmo_iterator MMOEnd = MI->memoperands_end();
21584 MVT PVT = getPointerTy(MF->getDataLayout());
21585 assert((PVT == MVT::i64 || PVT == MVT::i32) &&
21586 "Invalid Pointer Size!");
21588 const TargetRegisterClass *RC =
21589 (PVT == MVT::i64) ? &X86::GR64RegClass : &X86::GR32RegClass;
21590 unsigned Tmp = MRI.createVirtualRegister(RC);
21591 // Since FP is only updated here but NOT referenced, it's treated as GPR.
21592 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
21593 unsigned FP = (PVT == MVT::i64) ? X86::RBP : X86::EBP;
21594 unsigned SP = RegInfo->getStackRegister();
21596 MachineInstrBuilder MIB;
21598 const int64_t LabelOffset = 1 * PVT.getStoreSize();
21599 const int64_t SPOffset = 2 * PVT.getStoreSize();
21601 unsigned PtrLoadOpc = (PVT == MVT::i64) ? X86::MOV64rm : X86::MOV32rm;
21602 unsigned IJmpOpc = (PVT == MVT::i64) ? X86::JMP64r : X86::JMP32r;
21605 MIB = BuildMI(*MBB, MI, DL, TII->get(PtrLoadOpc), FP);
21606 for (unsigned i = 0; i < X86::AddrNumOperands; ++i)
21607 MIB.addOperand(MI->getOperand(i));
21608 MIB.setMemRefs(MMOBegin, MMOEnd);
21610 MIB = BuildMI(*MBB, MI, DL, TII->get(PtrLoadOpc), Tmp);
21611 for (unsigned i = 0; i < X86::AddrNumOperands; ++i) {
21612 if (i == X86::AddrDisp)
21613 MIB.addDisp(MI->getOperand(i), LabelOffset);
21615 MIB.addOperand(MI->getOperand(i));
21617 MIB.setMemRefs(MMOBegin, MMOEnd);
21619 MIB = BuildMI(*MBB, MI, DL, TII->get(PtrLoadOpc), SP);
21620 for (unsigned i = 0; i < X86::AddrNumOperands; ++i) {
21621 if (i == X86::AddrDisp)
21622 MIB.addDisp(MI->getOperand(i), SPOffset);
21624 MIB.addOperand(MI->getOperand(i));
21626 MIB.setMemRefs(MMOBegin, MMOEnd);
21628 BuildMI(*MBB, MI, DL, TII->get(IJmpOpc)).addReg(Tmp);
21630 MI->eraseFromParent();
21634 // Replace 213-type (isel default) FMA3 instructions with 231-type for
21635 // accumulator loops. Writing back to the accumulator allows the coalescer
21636 // to remove extra copies in the loop.
21637 // FIXME: Do this on AVX512. We don't support 231 variants yet (PR23937).
21638 MachineBasicBlock *
21639 X86TargetLowering::emitFMA3Instr(MachineInstr *MI,
21640 MachineBasicBlock *MBB) const {
21641 MachineOperand &AddendOp = MI->getOperand(3);
21643 // Bail out early if the addend isn't a register - we can't switch these.
21644 if (!AddendOp.isReg())
21647 MachineFunction &MF = *MBB->getParent();
21648 MachineRegisterInfo &MRI = MF.getRegInfo();
21650 // Check whether the addend is defined by a PHI:
21651 assert(MRI.hasOneDef(AddendOp.getReg()) && "Multiple defs in SSA?");
21652 MachineInstr &AddendDef = *MRI.def_instr_begin(AddendOp.getReg());
21653 if (!AddendDef.isPHI())
21656 // Look for the following pattern:
21658 // %addend = phi [%entry, 0], [%loop, %result]
21660 // %result<tied1> = FMA213 %m2<tied0>, %m1, %addend
21664 // %addend = phi [%entry, 0], [%loop, %result]
21666 // %result<tied1> = FMA231 %addend<tied0>, %m1, %m2
21668 for (unsigned i = 1, e = AddendDef.getNumOperands(); i < e; i += 2) {
21669 assert(AddendDef.getOperand(i).isReg());
21670 MachineOperand PHISrcOp = AddendDef.getOperand(i);
21671 MachineInstr &PHISrcInst = *MRI.def_instr_begin(PHISrcOp.getReg());
21672 if (&PHISrcInst == MI) {
21673 // Found a matching instruction.
21674 unsigned NewFMAOpc = 0;
21675 switch (MI->getOpcode()) {
21676 case X86::VFMADDPDr213r: NewFMAOpc = X86::VFMADDPDr231r; break;
21677 case X86::VFMADDPSr213r: NewFMAOpc = X86::VFMADDPSr231r; break;
21678 case X86::VFMADDSDr213r: NewFMAOpc = X86::VFMADDSDr231r; break;
21679 case X86::VFMADDSSr213r: NewFMAOpc = X86::VFMADDSSr231r; break;
21680 case X86::VFMSUBPDr213r: NewFMAOpc = X86::VFMSUBPDr231r; break;
21681 case X86::VFMSUBPSr213r: NewFMAOpc = X86::VFMSUBPSr231r; break;
21682 case X86::VFMSUBSDr213r: NewFMAOpc = X86::VFMSUBSDr231r; break;
21683 case X86::VFMSUBSSr213r: NewFMAOpc = X86::VFMSUBSSr231r; break;
21684 case X86::VFNMADDPDr213r: NewFMAOpc = X86::VFNMADDPDr231r; break;
21685 case X86::VFNMADDPSr213r: NewFMAOpc = X86::VFNMADDPSr231r; break;
21686 case X86::VFNMADDSDr213r: NewFMAOpc = X86::VFNMADDSDr231r; break;
21687 case X86::VFNMADDSSr213r: NewFMAOpc = X86::VFNMADDSSr231r; break;
21688 case X86::VFNMSUBPDr213r: NewFMAOpc = X86::VFNMSUBPDr231r; break;
21689 case X86::VFNMSUBPSr213r: NewFMAOpc = X86::VFNMSUBPSr231r; break;
21690 case X86::VFNMSUBSDr213r: NewFMAOpc = X86::VFNMSUBSDr231r; break;
21691 case X86::VFNMSUBSSr213r: NewFMAOpc = X86::VFNMSUBSSr231r; break;
21692 case X86::VFMADDSUBPDr213r: NewFMAOpc = X86::VFMADDSUBPDr231r; break;
21693 case X86::VFMADDSUBPSr213r: NewFMAOpc = X86::VFMADDSUBPSr231r; break;
21694 case X86::VFMSUBADDPDr213r: NewFMAOpc = X86::VFMSUBADDPDr231r; break;
21695 case X86::VFMSUBADDPSr213r: NewFMAOpc = X86::VFMSUBADDPSr231r; break;
21697 case X86::VFMADDPDr213rY: NewFMAOpc = X86::VFMADDPDr231rY; break;
21698 case X86::VFMADDPSr213rY: NewFMAOpc = X86::VFMADDPSr231rY; break;
21699 case X86::VFMSUBPDr213rY: NewFMAOpc = X86::VFMSUBPDr231rY; break;
21700 case X86::VFMSUBPSr213rY: NewFMAOpc = X86::VFMSUBPSr231rY; break;
21701 case X86::VFNMADDPDr213rY: NewFMAOpc = X86::VFNMADDPDr231rY; break;
21702 case X86::VFNMADDPSr213rY: NewFMAOpc = X86::VFNMADDPSr231rY; break;
21703 case X86::VFNMSUBPDr213rY: NewFMAOpc = X86::VFNMSUBPDr231rY; break;
21704 case X86::VFNMSUBPSr213rY: NewFMAOpc = X86::VFNMSUBPSr231rY; break;
21705 case X86::VFMADDSUBPDr213rY: NewFMAOpc = X86::VFMADDSUBPDr231rY; break;
21706 case X86::VFMADDSUBPSr213rY: NewFMAOpc = X86::VFMADDSUBPSr231rY; break;
21707 case X86::VFMSUBADDPDr213rY: NewFMAOpc = X86::VFMSUBADDPDr231rY; break;
21708 case X86::VFMSUBADDPSr213rY: NewFMAOpc = X86::VFMSUBADDPSr231rY; break;
21709 default: llvm_unreachable("Unrecognized FMA variant.");
21712 const TargetInstrInfo &TII = *Subtarget->getInstrInfo();
21713 MachineInstrBuilder MIB =
21714 BuildMI(MF, MI->getDebugLoc(), TII.get(NewFMAOpc))
21715 .addOperand(MI->getOperand(0))
21716 .addOperand(MI->getOperand(3))
21717 .addOperand(MI->getOperand(2))
21718 .addOperand(MI->getOperand(1));
21719 MBB->insert(MachineBasicBlock::iterator(MI), MIB);
21720 MI->eraseFromParent();
21727 MachineBasicBlock *
21728 X86TargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
21729 MachineBasicBlock *BB) const {
21730 switch (MI->getOpcode()) {
21731 default: llvm_unreachable("Unexpected instr type to insert");
21732 case X86::TAILJMPd64:
21733 case X86::TAILJMPr64:
21734 case X86::TAILJMPm64:
21735 case X86::TAILJMPd64_REX:
21736 case X86::TAILJMPr64_REX:
21737 case X86::TAILJMPm64_REX:
21738 llvm_unreachable("TAILJMP64 would not be touched here.");
21739 case X86::TCRETURNdi64:
21740 case X86::TCRETURNri64:
21741 case X86::TCRETURNmi64:
21743 case X86::WIN_ALLOCA:
21744 return EmitLoweredWinAlloca(MI, BB);
21745 case X86::SEG_ALLOCA_32:
21746 case X86::SEG_ALLOCA_64:
21747 return EmitLoweredSegAlloca(MI, BB);
21748 case X86::TLSCall_32:
21749 case X86::TLSCall_64:
21750 return EmitLoweredTLSCall(MI, BB);
21751 case X86::CMOV_FR32:
21752 case X86::CMOV_FR64:
21753 case X86::CMOV_GR8:
21754 case X86::CMOV_GR16:
21755 case X86::CMOV_GR32:
21756 case X86::CMOV_RFP32:
21757 case X86::CMOV_RFP64:
21758 case X86::CMOV_RFP80:
21759 case X86::CMOV_V2F64:
21760 case X86::CMOV_V2I64:
21761 case X86::CMOV_V4F32:
21762 case X86::CMOV_V4F64:
21763 case X86::CMOV_V4I64:
21764 case X86::CMOV_V16F32:
21765 case X86::CMOV_V8F32:
21766 case X86::CMOV_V8F64:
21767 case X86::CMOV_V8I64:
21768 case X86::CMOV_V8I1:
21769 case X86::CMOV_V16I1:
21770 case X86::CMOV_V32I1:
21771 case X86::CMOV_V64I1:
21772 return EmitLoweredSelect(MI, BB);
21774 case X86::RELEASE_FADD32mr:
21775 case X86::RELEASE_FADD64mr:
21776 return EmitLoweredAtomicFP(MI, BB);
21778 case X86::FP32_TO_INT16_IN_MEM:
21779 case X86::FP32_TO_INT32_IN_MEM:
21780 case X86::FP32_TO_INT64_IN_MEM:
21781 case X86::FP64_TO_INT16_IN_MEM:
21782 case X86::FP64_TO_INT32_IN_MEM:
21783 case X86::FP64_TO_INT64_IN_MEM:
21784 case X86::FP80_TO_INT16_IN_MEM:
21785 case X86::FP80_TO_INT32_IN_MEM:
21786 case X86::FP80_TO_INT64_IN_MEM: {
21787 MachineFunction *F = BB->getParent();
21788 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
21789 DebugLoc DL = MI->getDebugLoc();
21791 // Change the floating point control register to use "round towards zero"
21792 // mode when truncating to an integer value.
21793 int CWFrameIdx = F->getFrameInfo()->CreateStackObject(2, 2, false);
21794 addFrameReference(BuildMI(*BB, MI, DL,
21795 TII->get(X86::FNSTCW16m)), CWFrameIdx);
21797 // Load the old value of the high byte of the control word...
21799 F->getRegInfo().createVirtualRegister(&X86::GR16RegClass);
21800 addFrameReference(BuildMI(*BB, MI, DL, TII->get(X86::MOV16rm), OldCW),
21803 // Set the high part to be round to zero...
21804 addFrameReference(BuildMI(*BB, MI, DL, TII->get(X86::MOV16mi)), CWFrameIdx)
21807 // Reload the modified control word now...
21808 addFrameReference(BuildMI(*BB, MI, DL,
21809 TII->get(X86::FLDCW16m)), CWFrameIdx);
21811 // Restore the memory image of control word to original value
21812 addFrameReference(BuildMI(*BB, MI, DL, TII->get(X86::MOV16mr)), CWFrameIdx)
21815 // Get the X86 opcode to use.
21817 switch (MI->getOpcode()) {
21818 default: llvm_unreachable("illegal opcode!");
21819 case X86::FP32_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m32; break;
21820 case X86::FP32_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m32; break;
21821 case X86::FP32_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m32; break;
21822 case X86::FP64_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m64; break;
21823 case X86::FP64_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m64; break;
21824 case X86::FP64_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m64; break;
21825 case X86::FP80_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m80; break;
21826 case X86::FP80_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m80; break;
21827 case X86::FP80_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m80; break;
21831 MachineOperand &Op = MI->getOperand(0);
21833 AM.BaseType = X86AddressMode::RegBase;
21834 AM.Base.Reg = Op.getReg();
21836 AM.BaseType = X86AddressMode::FrameIndexBase;
21837 AM.Base.FrameIndex = Op.getIndex();
21839 Op = MI->getOperand(1);
21841 AM.Scale = Op.getImm();
21842 Op = MI->getOperand(2);
21844 AM.IndexReg = Op.getImm();
21845 Op = MI->getOperand(3);
21846 if (Op.isGlobal()) {
21847 AM.GV = Op.getGlobal();
21849 AM.Disp = Op.getImm();
21851 addFullAddress(BuildMI(*BB, MI, DL, TII->get(Opc)), AM)
21852 .addReg(MI->getOperand(X86::AddrNumOperands).getReg());
21854 // Reload the original control word now.
21855 addFrameReference(BuildMI(*BB, MI, DL,
21856 TII->get(X86::FLDCW16m)), CWFrameIdx);
21858 MI->eraseFromParent(); // The pseudo instruction is gone now.
21861 // String/text processing lowering.
21862 case X86::PCMPISTRM128REG:
21863 case X86::VPCMPISTRM128REG:
21864 case X86::PCMPISTRM128MEM:
21865 case X86::VPCMPISTRM128MEM:
21866 case X86::PCMPESTRM128REG:
21867 case X86::VPCMPESTRM128REG:
21868 case X86::PCMPESTRM128MEM:
21869 case X86::VPCMPESTRM128MEM:
21870 assert(Subtarget->hasSSE42() &&
21871 "Target must have SSE4.2 or AVX features enabled");
21872 return EmitPCMPSTRM(MI, BB, Subtarget->getInstrInfo());
21874 // String/text processing lowering.
21875 case X86::PCMPISTRIREG:
21876 case X86::VPCMPISTRIREG:
21877 case X86::PCMPISTRIMEM:
21878 case X86::VPCMPISTRIMEM:
21879 case X86::PCMPESTRIREG:
21880 case X86::VPCMPESTRIREG:
21881 case X86::PCMPESTRIMEM:
21882 case X86::VPCMPESTRIMEM:
21883 assert(Subtarget->hasSSE42() &&
21884 "Target must have SSE4.2 or AVX features enabled");
21885 return EmitPCMPSTRI(MI, BB, Subtarget->getInstrInfo());
21887 // Thread synchronization.
21889 return EmitMonitor(MI, BB, Subtarget);
21893 return EmitXBegin(MI, BB, Subtarget->getInstrInfo());
21895 case X86::VASTART_SAVE_XMM_REGS:
21896 return EmitVAStartSaveXMMRegsWithCustomInserter(MI, BB);
21898 case X86::VAARG_64:
21899 return EmitVAARG64WithCustomInserter(MI, BB);
21901 case X86::EH_SjLj_SetJmp32:
21902 case X86::EH_SjLj_SetJmp64:
21903 return emitEHSjLjSetJmp(MI, BB);
21905 case X86::EH_SjLj_LongJmp32:
21906 case X86::EH_SjLj_LongJmp64:
21907 return emitEHSjLjLongJmp(MI, BB);
21909 case TargetOpcode::STATEPOINT:
21910 // As an implementation detail, STATEPOINT shares the STACKMAP format at
21911 // this point in the process. We diverge later.
21912 return emitPatchPoint(MI, BB);
21914 case TargetOpcode::STACKMAP:
21915 case TargetOpcode::PATCHPOINT:
21916 return emitPatchPoint(MI, BB);
21918 case X86::VFMADDPDr213r:
21919 case X86::VFMADDPSr213r:
21920 case X86::VFMADDSDr213r:
21921 case X86::VFMADDSSr213r:
21922 case X86::VFMSUBPDr213r:
21923 case X86::VFMSUBPSr213r:
21924 case X86::VFMSUBSDr213r:
21925 case X86::VFMSUBSSr213r:
21926 case X86::VFNMADDPDr213r:
21927 case X86::VFNMADDPSr213r:
21928 case X86::VFNMADDSDr213r:
21929 case X86::VFNMADDSSr213r:
21930 case X86::VFNMSUBPDr213r:
21931 case X86::VFNMSUBPSr213r:
21932 case X86::VFNMSUBSDr213r:
21933 case X86::VFNMSUBSSr213r:
21934 case X86::VFMADDSUBPDr213r:
21935 case X86::VFMADDSUBPSr213r:
21936 case X86::VFMSUBADDPDr213r:
21937 case X86::VFMSUBADDPSr213r:
21938 case X86::VFMADDPDr213rY:
21939 case X86::VFMADDPSr213rY:
21940 case X86::VFMSUBPDr213rY:
21941 case X86::VFMSUBPSr213rY:
21942 case X86::VFNMADDPDr213rY:
21943 case X86::VFNMADDPSr213rY:
21944 case X86::VFNMSUBPDr213rY:
21945 case X86::VFNMSUBPSr213rY:
21946 case X86::VFMADDSUBPDr213rY:
21947 case X86::VFMADDSUBPSr213rY:
21948 case X86::VFMSUBADDPDr213rY:
21949 case X86::VFMSUBADDPSr213rY:
21950 return emitFMA3Instr(MI, BB);
21954 //===----------------------------------------------------------------------===//
21955 // X86 Optimization Hooks
21956 //===----------------------------------------------------------------------===//
21958 void X86TargetLowering::computeKnownBitsForTargetNode(const SDValue Op,
21961 const SelectionDAG &DAG,
21962 unsigned Depth) const {
21963 unsigned BitWidth = KnownZero.getBitWidth();
21964 unsigned Opc = Op.getOpcode();
21965 assert((Opc >= ISD::BUILTIN_OP_END ||
21966 Opc == ISD::INTRINSIC_WO_CHAIN ||
21967 Opc == ISD::INTRINSIC_W_CHAIN ||
21968 Opc == ISD::INTRINSIC_VOID) &&
21969 "Should use MaskedValueIsZero if you don't know whether Op"
21970 " is a target node!");
21972 KnownZero = KnownOne = APInt(BitWidth, 0); // Don't know anything.
21986 // These nodes' second result is a boolean.
21987 if (Op.getResNo() == 0)
21990 case X86ISD::SETCC:
21991 KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - 1);
21993 case ISD::INTRINSIC_WO_CHAIN: {
21994 unsigned IntId = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
21995 unsigned NumLoBits = 0;
21998 case Intrinsic::x86_sse_movmsk_ps:
21999 case Intrinsic::x86_avx_movmsk_ps_256:
22000 case Intrinsic::x86_sse2_movmsk_pd:
22001 case Intrinsic::x86_avx_movmsk_pd_256:
22002 case Intrinsic::x86_mmx_pmovmskb:
22003 case Intrinsic::x86_sse2_pmovmskb_128:
22004 case Intrinsic::x86_avx2_pmovmskb: {
22005 // High bits of movmskp{s|d}, pmovmskb are known zero.
22007 default: llvm_unreachable("Impossible intrinsic"); // Can't reach here.
22008 case Intrinsic::x86_sse_movmsk_ps: NumLoBits = 4; break;
22009 case Intrinsic::x86_avx_movmsk_ps_256: NumLoBits = 8; break;
22010 case Intrinsic::x86_sse2_movmsk_pd: NumLoBits = 2; break;
22011 case Intrinsic::x86_avx_movmsk_pd_256: NumLoBits = 4; break;
22012 case Intrinsic::x86_mmx_pmovmskb: NumLoBits = 8; break;
22013 case Intrinsic::x86_sse2_pmovmskb_128: NumLoBits = 16; break;
22014 case Intrinsic::x86_avx2_pmovmskb: NumLoBits = 32; break;
22016 KnownZero = APInt::getHighBitsSet(BitWidth, BitWidth - NumLoBits);
22025 unsigned X86TargetLowering::ComputeNumSignBitsForTargetNode(
22027 const SelectionDAG &,
22028 unsigned Depth) const {
22029 // SETCC_CARRY sets the dest to ~0 for true or 0 for false.
22030 if (Op.getOpcode() == X86ISD::SETCC_CARRY)
22031 return Op.getValueType().getScalarType().getSizeInBits();
22037 /// isGAPlusOffset - Returns true (and the GlobalValue and the offset) if the
22038 /// node is a GlobalAddress + offset.
22039 bool X86TargetLowering::isGAPlusOffset(SDNode *N,
22040 const GlobalValue* &GA,
22041 int64_t &Offset) const {
22042 if (N->getOpcode() == X86ISD::Wrapper) {
22043 if (isa<GlobalAddressSDNode>(N->getOperand(0))) {
22044 GA = cast<GlobalAddressSDNode>(N->getOperand(0))->getGlobal();
22045 Offset = cast<GlobalAddressSDNode>(N->getOperand(0))->getOffset();
22049 return TargetLowering::isGAPlusOffset(N, GA, Offset);
22052 /// isShuffleHigh128VectorInsertLow - Checks whether the shuffle node is the
22053 /// same as extracting the high 128-bit part of 256-bit vector and then
22054 /// inserting the result into the low part of a new 256-bit vector
22055 static bool isShuffleHigh128VectorInsertLow(ShuffleVectorSDNode *SVOp) {
22056 EVT VT = SVOp->getValueType(0);
22057 unsigned NumElems = VT.getVectorNumElements();
22059 // vector_shuffle <4, 5, 6, 7, u, u, u, u> or <2, 3, u, u>
22060 for (unsigned i = 0, j = NumElems/2; i != NumElems/2; ++i, ++j)
22061 if (!isUndefOrEqual(SVOp->getMaskElt(i), j) ||
22062 SVOp->getMaskElt(j) >= 0)
22068 /// isShuffleLow128VectorInsertHigh - Checks whether the shuffle node is the
22069 /// same as extracting the low 128-bit part of 256-bit vector and then
22070 /// inserting the result into the high part of a new 256-bit vector
22071 static bool isShuffleLow128VectorInsertHigh(ShuffleVectorSDNode *SVOp) {
22072 EVT VT = SVOp->getValueType(0);
22073 unsigned NumElems = VT.getVectorNumElements();
22075 // vector_shuffle <u, u, u, u, 0, 1, 2, 3> or <u, u, 0, 1>
22076 for (unsigned i = NumElems/2, j = 0; i != NumElems; ++i, ++j)
22077 if (!isUndefOrEqual(SVOp->getMaskElt(i), j) ||
22078 SVOp->getMaskElt(j) >= 0)
22084 /// PerformShuffleCombine256 - Performs shuffle combines for 256-bit vectors.
22085 static SDValue PerformShuffleCombine256(SDNode *N, SelectionDAG &DAG,
22086 TargetLowering::DAGCombinerInfo &DCI,
22087 const X86Subtarget* Subtarget) {
22089 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
22090 SDValue V1 = SVOp->getOperand(0);
22091 SDValue V2 = SVOp->getOperand(1);
22092 EVT VT = SVOp->getValueType(0);
22093 unsigned NumElems = VT.getVectorNumElements();
22095 if (V1.getOpcode() == ISD::CONCAT_VECTORS &&
22096 V2.getOpcode() == ISD::CONCAT_VECTORS) {
22100 // V UNDEF BUILD_VECTOR UNDEF
22102 // CONCAT_VECTOR CONCAT_VECTOR
22105 // RESULT: V + zero extended
22107 if (V2.getOperand(0).getOpcode() != ISD::BUILD_VECTOR ||
22108 V2.getOperand(1).getOpcode() != ISD::UNDEF ||
22109 V1.getOperand(1).getOpcode() != ISD::UNDEF)
22112 if (!ISD::isBuildVectorAllZeros(V2.getOperand(0).getNode()))
22115 // To match the shuffle mask, the first half of the mask should
22116 // be exactly the first vector, and all the rest a splat with the
22117 // first element of the second one.
22118 for (unsigned i = 0; i != NumElems/2; ++i)
22119 if (!isUndefOrEqual(SVOp->getMaskElt(i), i) ||
22120 !isUndefOrEqual(SVOp->getMaskElt(i+NumElems/2), NumElems))
22123 // If V1 is coming from a vector load then just fold to a VZEXT_LOAD.
22124 if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(V1.getOperand(0))) {
22125 if (Ld->hasNUsesOfValue(1, 0)) {
22126 SDVTList Tys = DAG.getVTList(MVT::v4i64, MVT::Other);
22127 SDValue Ops[] = { Ld->getChain(), Ld->getBasePtr() };
22129 DAG.getMemIntrinsicNode(X86ISD::VZEXT_LOAD, dl, Tys, Ops,
22131 Ld->getPointerInfo(),
22132 Ld->getAlignment(),
22133 false/*isVolatile*/, true/*ReadMem*/,
22134 false/*WriteMem*/);
22136 // Make sure the newly-created LOAD is in the same position as Ld in
22137 // terms of dependency. We create a TokenFactor for Ld and ResNode,
22138 // and update uses of Ld's output chain to use the TokenFactor.
22139 if (Ld->hasAnyUseOfValue(1)) {
22140 SDValue NewChain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
22141 SDValue(Ld, 1), SDValue(ResNode.getNode(), 1));
22142 DAG.ReplaceAllUsesOfValueWith(SDValue(Ld, 1), NewChain);
22143 DAG.UpdateNodeOperands(NewChain.getNode(), SDValue(Ld, 1),
22144 SDValue(ResNode.getNode(), 1));
22147 return DAG.getBitcast(VT, ResNode);
22151 // Emit a zeroed vector and insert the desired subvector on its
22153 SDValue Zeros = getZeroVector(VT, Subtarget, DAG, dl);
22154 SDValue InsV = Insert128BitVector(Zeros, V1.getOperand(0), 0, DAG, dl);
22155 return DCI.CombineTo(N, InsV);
22158 //===--------------------------------------------------------------------===//
22159 // Combine some shuffles into subvector extracts and inserts:
22162 // vector_shuffle <4, 5, 6, 7, u, u, u, u> or <2, 3, u, u>
22163 if (isShuffleHigh128VectorInsertLow(SVOp)) {
22164 SDValue V = Extract128BitVector(V1, NumElems/2, DAG, dl);
22165 SDValue InsV = Insert128BitVector(DAG.getUNDEF(VT), V, 0, DAG, dl);
22166 return DCI.CombineTo(N, InsV);
22169 // vector_shuffle <u, u, u, u, 0, 1, 2, 3> or <u, u, 0, 1>
22170 if (isShuffleLow128VectorInsertHigh(SVOp)) {
22171 SDValue V = Extract128BitVector(V1, 0, DAG, dl);
22172 SDValue InsV = Insert128BitVector(DAG.getUNDEF(VT), V, NumElems/2, DAG, dl);
22173 return DCI.CombineTo(N, InsV);
22179 /// \brief Combine an arbitrary chain of shuffles into a single instruction if
22182 /// This is the leaf of the recursive combinine below. When we have found some
22183 /// chain of single-use x86 shuffle instructions and accumulated the combined
22184 /// shuffle mask represented by them, this will try to pattern match that mask
22185 /// into either a single instruction if there is a special purpose instruction
22186 /// for this operation, or into a PSHUFB instruction which is a fully general
22187 /// instruction but should only be used to replace chains over a certain depth.
22188 static bool combineX86ShuffleChain(SDValue Op, SDValue Root, ArrayRef<int> Mask,
22189 int Depth, bool HasPSHUFB, SelectionDAG &DAG,
22190 TargetLowering::DAGCombinerInfo &DCI,
22191 const X86Subtarget *Subtarget) {
22192 assert(!Mask.empty() && "Cannot combine an empty shuffle mask!");
22194 // Find the operand that enters the chain. Note that multiple uses are OK
22195 // here, we're not going to remove the operand we find.
22196 SDValue Input = Op.getOperand(0);
22197 while (Input.getOpcode() == ISD::BITCAST)
22198 Input = Input.getOperand(0);
22200 MVT VT = Input.getSimpleValueType();
22201 MVT RootVT = Root.getSimpleValueType();
22204 if (Mask.size() == 1) {
22205 int Index = Mask[0];
22206 assert((Index >= 0 || Index == SM_SentinelUndef ||
22207 Index == SM_SentinelZero) &&
22208 "Invalid shuffle index found!");
22210 // We may end up with an accumulated mask of size 1 as a result of
22211 // widening of shuffle operands (see function canWidenShuffleElements).
22212 // If the only shuffle index is equal to SM_SentinelZero then propagate
22213 // a zero vector. Otherwise, the combine shuffle mask is a no-op shuffle
22214 // mask, and therefore the entire chain of shuffles can be folded away.
22215 if (Index == SM_SentinelZero)
22216 DCI.CombineTo(Root.getNode(), getZeroVector(RootVT, Subtarget, DAG, DL));
22218 DCI.CombineTo(Root.getNode(), DAG.getBitcast(RootVT, Input),
22223 // Use the float domain if the operand type is a floating point type.
22224 bool FloatDomain = VT.isFloatingPoint();
22226 // For floating point shuffles, we don't have free copies in the shuffle
22227 // instructions or the ability to load as part of the instruction, so
22228 // canonicalize their shuffles to UNPCK or MOV variants.
22230 // Note that even with AVX we prefer the PSHUFD form of shuffle for integer
22231 // vectors because it can have a load folded into it that UNPCK cannot. This
22232 // doesn't preclude something switching to the shorter encoding post-RA.
22234 // FIXME: Should teach these routines about AVX vector widths.
22235 if (FloatDomain && VT.getSizeInBits() == 128) {
22236 if (Mask.equals({0, 0}) || Mask.equals({1, 1})) {
22237 bool Lo = Mask.equals({0, 0});
22240 // Check if we have SSE3 which will let us use MOVDDUP. That instruction
22241 // is no slower than UNPCKLPD but has the option to fold the input operand
22242 // into even an unaligned memory load.
22243 if (Lo && Subtarget->hasSSE3()) {
22244 Shuffle = X86ISD::MOVDDUP;
22245 ShuffleVT = MVT::v2f64;
22247 // We have MOVLHPS and MOVHLPS throughout SSE and they encode smaller
22248 // than the UNPCK variants.
22249 Shuffle = Lo ? X86ISD::MOVLHPS : X86ISD::MOVHLPS;
22250 ShuffleVT = MVT::v4f32;
22252 if (Depth == 1 && Root->getOpcode() == Shuffle)
22253 return false; // Nothing to do!
22254 Op = DAG.getBitcast(ShuffleVT, Input);
22255 DCI.AddToWorklist(Op.getNode());
22256 if (Shuffle == X86ISD::MOVDDUP)
22257 Op = DAG.getNode(Shuffle, DL, ShuffleVT, Op);
22259 Op = DAG.getNode(Shuffle, DL, ShuffleVT, Op, Op);
22260 DCI.AddToWorklist(Op.getNode());
22261 DCI.CombineTo(Root.getNode(), DAG.getBitcast(RootVT, Op),
22265 if (Subtarget->hasSSE3() &&
22266 (Mask.equals({0, 0, 2, 2}) || Mask.equals({1, 1, 3, 3}))) {
22267 bool Lo = Mask.equals({0, 0, 2, 2});
22268 unsigned Shuffle = Lo ? X86ISD::MOVSLDUP : X86ISD::MOVSHDUP;
22269 MVT ShuffleVT = MVT::v4f32;
22270 if (Depth == 1 && Root->getOpcode() == Shuffle)
22271 return false; // Nothing to do!
22272 Op = DAG.getBitcast(ShuffleVT, Input);
22273 DCI.AddToWorklist(Op.getNode());
22274 Op = DAG.getNode(Shuffle, DL, ShuffleVT, Op);
22275 DCI.AddToWorklist(Op.getNode());
22276 DCI.CombineTo(Root.getNode(), DAG.getBitcast(RootVT, Op),
22280 if (Mask.equals({0, 0, 1, 1}) || Mask.equals({2, 2, 3, 3})) {
22281 bool Lo = Mask.equals({0, 0, 1, 1});
22282 unsigned Shuffle = Lo ? X86ISD::UNPCKL : X86ISD::UNPCKH;
22283 MVT ShuffleVT = MVT::v4f32;
22284 if (Depth == 1 && Root->getOpcode() == Shuffle)
22285 return false; // Nothing to do!
22286 Op = DAG.getBitcast(ShuffleVT, Input);
22287 DCI.AddToWorklist(Op.getNode());
22288 Op = DAG.getNode(Shuffle, DL, ShuffleVT, Op, Op);
22289 DCI.AddToWorklist(Op.getNode());
22290 DCI.CombineTo(Root.getNode(), DAG.getBitcast(RootVT, Op),
22296 // We always canonicalize the 8 x i16 and 16 x i8 shuffles into their UNPCK
22297 // variants as none of these have single-instruction variants that are
22298 // superior to the UNPCK formulation.
22299 if (!FloatDomain && VT.getSizeInBits() == 128 &&
22300 (Mask.equals({0, 0, 1, 1, 2, 2, 3, 3}) ||
22301 Mask.equals({4, 4, 5, 5, 6, 6, 7, 7}) ||
22302 Mask.equals({0, 0, 1, 1, 2, 2, 3, 3, 4, 4, 5, 5, 6, 6, 7, 7}) ||
22304 {8, 8, 9, 9, 10, 10, 11, 11, 12, 12, 13, 13, 14, 14, 15, 15}))) {
22305 bool Lo = Mask[0] == 0;
22306 unsigned Shuffle = Lo ? X86ISD::UNPCKL : X86ISD::UNPCKH;
22307 if (Depth == 1 && Root->getOpcode() == Shuffle)
22308 return false; // Nothing to do!
22310 switch (Mask.size()) {
22312 ShuffleVT = MVT::v8i16;
22315 ShuffleVT = MVT::v16i8;
22318 llvm_unreachable("Impossible mask size!");
22320 Op = DAG.getBitcast(ShuffleVT, Input);
22321 DCI.AddToWorklist(Op.getNode());
22322 Op = DAG.getNode(Shuffle, DL, ShuffleVT, Op, Op);
22323 DCI.AddToWorklist(Op.getNode());
22324 DCI.CombineTo(Root.getNode(), DAG.getBitcast(RootVT, Op),
22329 // Don't try to re-form single instruction chains under any circumstances now
22330 // that we've done encoding canonicalization for them.
22334 // If we have 3 or more shuffle instructions or a chain involving PSHUFB, we
22335 // can replace them with a single PSHUFB instruction profitably. Intel's
22336 // manuals suggest only using PSHUFB if doing so replacing 5 instructions, but
22337 // in practice PSHUFB tends to be *very* fast so we're more aggressive.
22338 if ((Depth >= 3 || HasPSHUFB) && Subtarget->hasSSSE3()) {
22339 SmallVector<SDValue, 16> PSHUFBMask;
22340 int NumBytes = VT.getSizeInBits() / 8;
22341 int Ratio = NumBytes / Mask.size();
22342 for (int i = 0; i < NumBytes; ++i) {
22343 if (Mask[i / Ratio] == SM_SentinelUndef) {
22344 PSHUFBMask.push_back(DAG.getUNDEF(MVT::i8));
22347 int M = Mask[i / Ratio] != SM_SentinelZero
22348 ? Ratio * Mask[i / Ratio] + i % Ratio
22350 PSHUFBMask.push_back(DAG.getConstant(M, DL, MVT::i8));
22352 MVT ByteVT = MVT::getVectorVT(MVT::i8, NumBytes);
22353 Op = DAG.getBitcast(ByteVT, Input);
22354 DCI.AddToWorklist(Op.getNode());
22355 SDValue PSHUFBMaskOp =
22356 DAG.getNode(ISD::BUILD_VECTOR, DL, ByteVT, PSHUFBMask);
22357 DCI.AddToWorklist(PSHUFBMaskOp.getNode());
22358 Op = DAG.getNode(X86ISD::PSHUFB, DL, ByteVT, Op, PSHUFBMaskOp);
22359 DCI.AddToWorklist(Op.getNode());
22360 DCI.CombineTo(Root.getNode(), DAG.getBitcast(RootVT, Op),
22365 // Failed to find any combines.
22369 /// \brief Fully generic combining of x86 shuffle instructions.
22371 /// This should be the last combine run over the x86 shuffle instructions. Once
22372 /// they have been fully optimized, this will recursively consider all chains
22373 /// of single-use shuffle instructions, build a generic model of the cumulative
22374 /// shuffle operation, and check for simpler instructions which implement this
22375 /// operation. We use this primarily for two purposes:
22377 /// 1) Collapse generic shuffles to specialized single instructions when
22378 /// equivalent. In most cases, this is just an encoding size win, but
22379 /// sometimes we will collapse multiple generic shuffles into a single
22380 /// special-purpose shuffle.
22381 /// 2) Look for sequences of shuffle instructions with 3 or more total
22382 /// instructions, and replace them with the slightly more expensive SSSE3
22383 /// PSHUFB instruction if available. We do this as the last combining step
22384 /// to ensure we avoid using PSHUFB if we can implement the shuffle with
22385 /// a suitable short sequence of other instructions. The PHUFB will either
22386 /// use a register or have to read from memory and so is slightly (but only
22387 /// slightly) more expensive than the other shuffle instructions.
22389 /// Because this is inherently a quadratic operation (for each shuffle in
22390 /// a chain, we recurse up the chain), the depth is limited to 8 instructions.
22391 /// This should never be an issue in practice as the shuffle lowering doesn't
22392 /// produce sequences of more than 8 instructions.
22394 /// FIXME: We will currently miss some cases where the redundant shuffling
22395 /// would simplify under the threshold for PSHUFB formation because of
22396 /// combine-ordering. To fix this, we should do the redundant instruction
22397 /// combining in this recursive walk.
22398 static bool combineX86ShufflesRecursively(SDValue Op, SDValue Root,
22399 ArrayRef<int> RootMask,
22400 int Depth, bool HasPSHUFB,
22402 TargetLowering::DAGCombinerInfo &DCI,
22403 const X86Subtarget *Subtarget) {
22404 // Bound the depth of our recursive combine because this is ultimately
22405 // quadratic in nature.
22409 // Directly rip through bitcasts to find the underlying operand.
22410 while (Op.getOpcode() == ISD::BITCAST && Op.getOperand(0).hasOneUse())
22411 Op = Op.getOperand(0);
22413 MVT VT = Op.getSimpleValueType();
22414 if (!VT.isVector())
22415 return false; // Bail if we hit a non-vector.
22417 assert(Root.getSimpleValueType().isVector() &&
22418 "Shuffles operate on vector types!");
22419 assert(VT.getSizeInBits() == Root.getSimpleValueType().getSizeInBits() &&
22420 "Can only combine shuffles of the same vector register size.");
22422 if (!isTargetShuffle(Op.getOpcode()))
22424 SmallVector<int, 16> OpMask;
22426 bool HaveMask = getTargetShuffleMask(Op.getNode(), VT, OpMask, IsUnary);
22427 // We only can combine unary shuffles which we can decode the mask for.
22428 if (!HaveMask || !IsUnary)
22431 assert(VT.getVectorNumElements() == OpMask.size() &&
22432 "Different mask size from vector size!");
22433 assert(((RootMask.size() > OpMask.size() &&
22434 RootMask.size() % OpMask.size() == 0) ||
22435 (OpMask.size() > RootMask.size() &&
22436 OpMask.size() % RootMask.size() == 0) ||
22437 OpMask.size() == RootMask.size()) &&
22438 "The smaller number of elements must divide the larger.");
22439 int RootRatio = std::max<int>(1, OpMask.size() / RootMask.size());
22440 int OpRatio = std::max<int>(1, RootMask.size() / OpMask.size());
22441 assert(((RootRatio == 1 && OpRatio == 1) ||
22442 (RootRatio == 1) != (OpRatio == 1)) &&
22443 "Must not have a ratio for both incoming and op masks!");
22445 SmallVector<int, 16> Mask;
22446 Mask.reserve(std::max(OpMask.size(), RootMask.size()));
22448 // Merge this shuffle operation's mask into our accumulated mask. Note that
22449 // this shuffle's mask will be the first applied to the input, followed by the
22450 // root mask to get us all the way to the root value arrangement. The reason
22451 // for this order is that we are recursing up the operation chain.
22452 for (int i = 0, e = std::max(OpMask.size(), RootMask.size()); i < e; ++i) {
22453 int RootIdx = i / RootRatio;
22454 if (RootMask[RootIdx] < 0) {
22455 // This is a zero or undef lane, we're done.
22456 Mask.push_back(RootMask[RootIdx]);
22460 int RootMaskedIdx = RootMask[RootIdx] * RootRatio + i % RootRatio;
22461 int OpIdx = RootMaskedIdx / OpRatio;
22462 if (OpMask[OpIdx] < 0) {
22463 // The incoming lanes are zero or undef, it doesn't matter which ones we
22465 Mask.push_back(OpMask[OpIdx]);
22469 // Ok, we have non-zero lanes, map them through.
22470 Mask.push_back(OpMask[OpIdx] * OpRatio +
22471 RootMaskedIdx % OpRatio);
22474 // See if we can recurse into the operand to combine more things.
22475 switch (Op.getOpcode()) {
22476 case X86ISD::PSHUFB:
22478 case X86ISD::PSHUFD:
22479 case X86ISD::PSHUFHW:
22480 case X86ISD::PSHUFLW:
22481 if (Op.getOperand(0).hasOneUse() &&
22482 combineX86ShufflesRecursively(Op.getOperand(0), Root, Mask, Depth + 1,
22483 HasPSHUFB, DAG, DCI, Subtarget))
22487 case X86ISD::UNPCKL:
22488 case X86ISD::UNPCKH:
22489 assert(Op.getOperand(0) == Op.getOperand(1) &&
22490 "We only combine unary shuffles!");
22491 // We can't check for single use, we have to check that this shuffle is the
22493 if (Op->isOnlyUserOf(Op.getOperand(0).getNode()) &&
22494 combineX86ShufflesRecursively(Op.getOperand(0), Root, Mask, Depth + 1,
22495 HasPSHUFB, DAG, DCI, Subtarget))
22500 // Minor canonicalization of the accumulated shuffle mask to make it easier
22501 // to match below. All this does is detect masks with squential pairs of
22502 // elements, and shrink them to the half-width mask. It does this in a loop
22503 // so it will reduce the size of the mask to the minimal width mask which
22504 // performs an equivalent shuffle.
22505 SmallVector<int, 16> WidenedMask;
22506 while (Mask.size() > 1 && canWidenShuffleElements(Mask, WidenedMask)) {
22507 Mask = std::move(WidenedMask);
22508 WidenedMask.clear();
22511 return combineX86ShuffleChain(Op, Root, Mask, Depth, HasPSHUFB, DAG, DCI,
22515 /// \brief Get the PSHUF-style mask from PSHUF node.
22517 /// This is a very minor wrapper around getTargetShuffleMask to easy forming v4
22518 /// PSHUF-style masks that can be reused with such instructions.
22519 static SmallVector<int, 4> getPSHUFShuffleMask(SDValue N) {
22520 MVT VT = N.getSimpleValueType();
22521 SmallVector<int, 4> Mask;
22523 bool HaveMask = getTargetShuffleMask(N.getNode(), VT, Mask, IsUnary);
22527 // If we have more than 128-bits, only the low 128-bits of shuffle mask
22528 // matter. Check that the upper masks are repeats and remove them.
22529 if (VT.getSizeInBits() > 128) {
22530 int LaneElts = 128 / VT.getScalarSizeInBits();
22532 for (int i = 1, NumLanes = VT.getSizeInBits() / 128; i < NumLanes; ++i)
22533 for (int j = 0; j < LaneElts; ++j)
22534 assert(Mask[j] == Mask[i * LaneElts + j] - (LaneElts * i) &&
22535 "Mask doesn't repeat in high 128-bit lanes!");
22537 Mask.resize(LaneElts);
22540 switch (N.getOpcode()) {
22541 case X86ISD::PSHUFD:
22543 case X86ISD::PSHUFLW:
22546 case X86ISD::PSHUFHW:
22547 Mask.erase(Mask.begin(), Mask.begin() + 4);
22548 for (int &M : Mask)
22552 llvm_unreachable("No valid shuffle instruction found!");
22556 /// \brief Search for a combinable shuffle across a chain ending in pshufd.
22558 /// We walk up the chain and look for a combinable shuffle, skipping over
22559 /// shuffles that we could hoist this shuffle's transformation past without
22560 /// altering anything.
22562 combineRedundantDWordShuffle(SDValue N, MutableArrayRef<int> Mask,
22564 TargetLowering::DAGCombinerInfo &DCI) {
22565 assert(N.getOpcode() == X86ISD::PSHUFD &&
22566 "Called with something other than an x86 128-bit half shuffle!");
22569 // Walk up a single-use chain looking for a combinable shuffle. Keep a stack
22570 // of the shuffles in the chain so that we can form a fresh chain to replace
22572 SmallVector<SDValue, 8> Chain;
22573 SDValue V = N.getOperand(0);
22574 for (; V.hasOneUse(); V = V.getOperand(0)) {
22575 switch (V.getOpcode()) {
22577 return SDValue(); // Nothing combined!
22580 // Skip bitcasts as we always know the type for the target specific
22584 case X86ISD::PSHUFD:
22585 // Found another dword shuffle.
22588 case X86ISD::PSHUFLW:
22589 // Check that the low words (being shuffled) are the identity in the
22590 // dword shuffle, and the high words are self-contained.
22591 if (Mask[0] != 0 || Mask[1] != 1 ||
22592 !(Mask[2] >= 2 && Mask[2] < 4 && Mask[3] >= 2 && Mask[3] < 4))
22595 Chain.push_back(V);
22598 case X86ISD::PSHUFHW:
22599 // Check that the high words (being shuffled) are the identity in the
22600 // dword shuffle, and the low words are self-contained.
22601 if (Mask[2] != 2 || Mask[3] != 3 ||
22602 !(Mask[0] >= 0 && Mask[0] < 2 && Mask[1] >= 0 && Mask[1] < 2))
22605 Chain.push_back(V);
22608 case X86ISD::UNPCKL:
22609 case X86ISD::UNPCKH:
22610 // For either i8 -> i16 or i16 -> i32 unpacks, we can combine a dword
22611 // shuffle into a preceding word shuffle.
22612 if (V.getSimpleValueType().getScalarType() != MVT::i8 &&
22613 V.getSimpleValueType().getScalarType() != MVT::i16)
22616 // Search for a half-shuffle which we can combine with.
22617 unsigned CombineOp =
22618 V.getOpcode() == X86ISD::UNPCKL ? X86ISD::PSHUFLW : X86ISD::PSHUFHW;
22619 if (V.getOperand(0) != V.getOperand(1) ||
22620 !V->isOnlyUserOf(V.getOperand(0).getNode()))
22622 Chain.push_back(V);
22623 V = V.getOperand(0);
22625 switch (V.getOpcode()) {
22627 return SDValue(); // Nothing to combine.
22629 case X86ISD::PSHUFLW:
22630 case X86ISD::PSHUFHW:
22631 if (V.getOpcode() == CombineOp)
22634 Chain.push_back(V);
22638 V = V.getOperand(0);
22642 } while (V.hasOneUse());
22645 // Break out of the loop if we break out of the switch.
22649 if (!V.hasOneUse())
22650 // We fell out of the loop without finding a viable combining instruction.
22653 // Merge this node's mask and our incoming mask.
22654 SmallVector<int, 4> VMask = getPSHUFShuffleMask(V);
22655 for (int &M : Mask)
22657 V = DAG.getNode(V.getOpcode(), DL, V.getValueType(), V.getOperand(0),
22658 getV4X86ShuffleImm8ForMask(Mask, DL, DAG));
22660 // Rebuild the chain around this new shuffle.
22661 while (!Chain.empty()) {
22662 SDValue W = Chain.pop_back_val();
22664 if (V.getValueType() != W.getOperand(0).getValueType())
22665 V = DAG.getBitcast(W.getOperand(0).getValueType(), V);
22667 switch (W.getOpcode()) {
22669 llvm_unreachable("Only PSHUF and UNPCK instructions get here!");
22671 case X86ISD::UNPCKL:
22672 case X86ISD::UNPCKH:
22673 V = DAG.getNode(W.getOpcode(), DL, W.getValueType(), V, V);
22676 case X86ISD::PSHUFD:
22677 case X86ISD::PSHUFLW:
22678 case X86ISD::PSHUFHW:
22679 V = DAG.getNode(W.getOpcode(), DL, W.getValueType(), V, W.getOperand(1));
22683 if (V.getValueType() != N.getValueType())
22684 V = DAG.getBitcast(N.getValueType(), V);
22686 // Return the new chain to replace N.
22690 /// \brief Search for a combinable shuffle across a chain ending in pshuflw or
22693 /// We walk up the chain, skipping shuffles of the other half and looking
22694 /// through shuffles which switch halves trying to find a shuffle of the same
22695 /// pair of dwords.
22696 static bool combineRedundantHalfShuffle(SDValue N, MutableArrayRef<int> Mask,
22698 TargetLowering::DAGCombinerInfo &DCI) {
22700 (N.getOpcode() == X86ISD::PSHUFLW || N.getOpcode() == X86ISD::PSHUFHW) &&
22701 "Called with something other than an x86 128-bit half shuffle!");
22703 unsigned CombineOpcode = N.getOpcode();
22705 // Walk up a single-use chain looking for a combinable shuffle.
22706 SDValue V = N.getOperand(0);
22707 for (; V.hasOneUse(); V = V.getOperand(0)) {
22708 switch (V.getOpcode()) {
22710 return false; // Nothing combined!
22713 // Skip bitcasts as we always know the type for the target specific
22717 case X86ISD::PSHUFLW:
22718 case X86ISD::PSHUFHW:
22719 if (V.getOpcode() == CombineOpcode)
22722 // Other-half shuffles are no-ops.
22725 // Break out of the loop if we break out of the switch.
22729 if (!V.hasOneUse())
22730 // We fell out of the loop without finding a viable combining instruction.
22733 // Combine away the bottom node as its shuffle will be accumulated into
22734 // a preceding shuffle.
22735 DCI.CombineTo(N.getNode(), N.getOperand(0), /*AddTo*/ true);
22737 // Record the old value.
22740 // Merge this node's mask and our incoming mask (adjusted to account for all
22741 // the pshufd instructions encountered).
22742 SmallVector<int, 4> VMask = getPSHUFShuffleMask(V);
22743 for (int &M : Mask)
22745 V = DAG.getNode(V.getOpcode(), DL, MVT::v8i16, V.getOperand(0),
22746 getV4X86ShuffleImm8ForMask(Mask, DL, DAG));
22748 // Check that the shuffles didn't cancel each other out. If not, we need to
22749 // combine to the new one.
22751 // Replace the combinable shuffle with the combined one, updating all users
22752 // so that we re-evaluate the chain here.
22753 DCI.CombineTo(Old.getNode(), V, /*AddTo*/ true);
22758 /// \brief Try to combine x86 target specific shuffles.
22759 static SDValue PerformTargetShuffleCombine(SDValue N, SelectionDAG &DAG,
22760 TargetLowering::DAGCombinerInfo &DCI,
22761 const X86Subtarget *Subtarget) {
22763 MVT VT = N.getSimpleValueType();
22764 SmallVector<int, 4> Mask;
22766 switch (N.getOpcode()) {
22767 case X86ISD::PSHUFD:
22768 case X86ISD::PSHUFLW:
22769 case X86ISD::PSHUFHW:
22770 Mask = getPSHUFShuffleMask(N);
22771 assert(Mask.size() == 4);
22777 // Nuke no-op shuffles that show up after combining.
22778 if (isNoopShuffleMask(Mask))
22779 return DCI.CombineTo(N.getNode(), N.getOperand(0), /*AddTo*/ true);
22781 // Look for simplifications involving one or two shuffle instructions.
22782 SDValue V = N.getOperand(0);
22783 switch (N.getOpcode()) {
22786 case X86ISD::PSHUFLW:
22787 case X86ISD::PSHUFHW:
22788 assert(VT.getScalarType() == MVT::i16 && "Bad word shuffle type!");
22790 if (combineRedundantHalfShuffle(N, Mask, DAG, DCI))
22791 return SDValue(); // We combined away this shuffle, so we're done.
22793 // See if this reduces to a PSHUFD which is no more expensive and can
22794 // combine with more operations. Note that it has to at least flip the
22795 // dwords as otherwise it would have been removed as a no-op.
22796 if (makeArrayRef(Mask).equals({2, 3, 0, 1})) {
22797 int DMask[] = {0, 1, 2, 3};
22798 int DOffset = N.getOpcode() == X86ISD::PSHUFLW ? 0 : 2;
22799 DMask[DOffset + 0] = DOffset + 1;
22800 DMask[DOffset + 1] = DOffset + 0;
22801 MVT DVT = MVT::getVectorVT(MVT::i32, VT.getVectorNumElements() / 2);
22802 V = DAG.getBitcast(DVT, V);
22803 DCI.AddToWorklist(V.getNode());
22804 V = DAG.getNode(X86ISD::PSHUFD, DL, DVT, V,
22805 getV4X86ShuffleImm8ForMask(DMask, DL, DAG));
22806 DCI.AddToWorklist(V.getNode());
22807 return DAG.getBitcast(VT, V);
22810 // Look for shuffle patterns which can be implemented as a single unpack.
22811 // FIXME: This doesn't handle the location of the PSHUFD generically, and
22812 // only works when we have a PSHUFD followed by two half-shuffles.
22813 if (Mask[0] == Mask[1] && Mask[2] == Mask[3] &&
22814 (V.getOpcode() == X86ISD::PSHUFLW ||
22815 V.getOpcode() == X86ISD::PSHUFHW) &&
22816 V.getOpcode() != N.getOpcode() &&
22818 SDValue D = V.getOperand(0);
22819 while (D.getOpcode() == ISD::BITCAST && D.hasOneUse())
22820 D = D.getOperand(0);
22821 if (D.getOpcode() == X86ISD::PSHUFD && D.hasOneUse()) {
22822 SmallVector<int, 4> VMask = getPSHUFShuffleMask(V);
22823 SmallVector<int, 4> DMask = getPSHUFShuffleMask(D);
22824 int NOffset = N.getOpcode() == X86ISD::PSHUFLW ? 0 : 4;
22825 int VOffset = V.getOpcode() == X86ISD::PSHUFLW ? 0 : 4;
22827 for (int i = 0; i < 4; ++i) {
22828 WordMask[i + NOffset] = Mask[i] + NOffset;
22829 WordMask[i + VOffset] = VMask[i] + VOffset;
22831 // Map the word mask through the DWord mask.
22833 for (int i = 0; i < 8; ++i)
22834 MappedMask[i] = 2 * DMask[WordMask[i] / 2] + WordMask[i] % 2;
22835 if (makeArrayRef(MappedMask).equals({0, 0, 1, 1, 2, 2, 3, 3}) ||
22836 makeArrayRef(MappedMask).equals({4, 4, 5, 5, 6, 6, 7, 7})) {
22837 // We can replace all three shuffles with an unpack.
22838 V = DAG.getBitcast(VT, D.getOperand(0));
22839 DCI.AddToWorklist(V.getNode());
22840 return DAG.getNode(MappedMask[0] == 0 ? X86ISD::UNPCKL
22849 case X86ISD::PSHUFD:
22850 if (SDValue NewN = combineRedundantDWordShuffle(N, Mask, DAG, DCI))
22859 /// \brief Try to combine a shuffle into a target-specific add-sub node.
22861 /// We combine this directly on the abstract vector shuffle nodes so it is
22862 /// easier to generically match. We also insert dummy vector shuffle nodes for
22863 /// the operands which explicitly discard the lanes which are unused by this
22864 /// operation to try to flow through the rest of the combiner the fact that
22865 /// they're unused.
22866 static SDValue combineShuffleToAddSub(SDNode *N, SelectionDAG &DAG) {
22868 EVT VT = N->getValueType(0);
22870 // We only handle target-independent shuffles.
22871 // FIXME: It would be easy and harmless to use the target shuffle mask
22872 // extraction tool to support more.
22873 if (N->getOpcode() != ISD::VECTOR_SHUFFLE)
22876 auto *SVN = cast<ShuffleVectorSDNode>(N);
22877 ArrayRef<int> Mask = SVN->getMask();
22878 SDValue V1 = N->getOperand(0);
22879 SDValue V2 = N->getOperand(1);
22881 // We require the first shuffle operand to be the SUB node, and the second to
22882 // be the ADD node.
22883 // FIXME: We should support the commuted patterns.
22884 if (V1->getOpcode() != ISD::FSUB || V2->getOpcode() != ISD::FADD)
22887 // If there are other uses of these operations we can't fold them.
22888 if (!V1->hasOneUse() || !V2->hasOneUse())
22891 // Ensure that both operations have the same operands. Note that we can
22892 // commute the FADD operands.
22893 SDValue LHS = V1->getOperand(0), RHS = V1->getOperand(1);
22894 if ((V2->getOperand(0) != LHS || V2->getOperand(1) != RHS) &&
22895 (V2->getOperand(0) != RHS || V2->getOperand(1) != LHS))
22898 // We're looking for blends between FADD and FSUB nodes. We insist on these
22899 // nodes being lined up in a specific expected pattern.
22900 if (!(isShuffleEquivalent(V1, V2, Mask, {0, 3}) ||
22901 isShuffleEquivalent(V1, V2, Mask, {0, 5, 2, 7}) ||
22902 isShuffleEquivalent(V1, V2, Mask, {0, 9, 2, 11, 4, 13, 6, 15})))
22905 // Only specific types are legal at this point, assert so we notice if and
22906 // when these change.
22907 assert((VT == MVT::v4f32 || VT == MVT::v2f64 || VT == MVT::v8f32 ||
22908 VT == MVT::v4f64) &&
22909 "Unknown vector type encountered!");
22911 return DAG.getNode(X86ISD::ADDSUB, DL, VT, LHS, RHS);
22914 /// PerformShuffleCombine - Performs several different shuffle combines.
22915 static SDValue PerformShuffleCombine(SDNode *N, SelectionDAG &DAG,
22916 TargetLowering::DAGCombinerInfo &DCI,
22917 const X86Subtarget *Subtarget) {
22919 SDValue N0 = N->getOperand(0);
22920 SDValue N1 = N->getOperand(1);
22921 EVT VT = N->getValueType(0);
22923 // Don't create instructions with illegal types after legalize types has run.
22924 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
22925 if (!DCI.isBeforeLegalize() && !TLI.isTypeLegal(VT.getVectorElementType()))
22928 // If we have legalized the vector types, look for blends of FADD and FSUB
22929 // nodes that we can fuse into an ADDSUB node.
22930 if (TLI.isTypeLegal(VT) && Subtarget->hasSSE3())
22931 if (SDValue AddSub = combineShuffleToAddSub(N, DAG))
22934 // Combine 256-bit vector shuffles. This is only profitable when in AVX mode
22935 if (Subtarget->hasFp256() && VT.is256BitVector() &&
22936 N->getOpcode() == ISD::VECTOR_SHUFFLE)
22937 return PerformShuffleCombine256(N, DAG, DCI, Subtarget);
22939 // During Type Legalization, when promoting illegal vector types,
22940 // the backend might introduce new shuffle dag nodes and bitcasts.
22942 // This code performs the following transformation:
22943 // fold: (shuffle (bitcast (BINOP A, B)), Undef, <Mask>) ->
22944 // (shuffle (BINOP (bitcast A), (bitcast B)), Undef, <Mask>)
22946 // We do this only if both the bitcast and the BINOP dag nodes have
22947 // one use. Also, perform this transformation only if the new binary
22948 // operation is legal. This is to avoid introducing dag nodes that
22949 // potentially need to be further expanded (or custom lowered) into a
22950 // less optimal sequence of dag nodes.
22951 if (!DCI.isBeforeLegalize() && DCI.isBeforeLegalizeOps() &&
22952 N1.getOpcode() == ISD::UNDEF && N0.hasOneUse() &&
22953 N0.getOpcode() == ISD::BITCAST) {
22954 SDValue BC0 = N0.getOperand(0);
22955 EVT SVT = BC0.getValueType();
22956 unsigned Opcode = BC0.getOpcode();
22957 unsigned NumElts = VT.getVectorNumElements();
22959 if (BC0.hasOneUse() && SVT.isVector() &&
22960 SVT.getVectorNumElements() * 2 == NumElts &&
22961 TLI.isOperationLegal(Opcode, VT)) {
22962 bool CanFold = false;
22974 unsigned SVTNumElts = SVT.getVectorNumElements();
22975 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
22976 for (unsigned i = 0, e = SVTNumElts; i != e && CanFold; ++i)
22977 CanFold = SVOp->getMaskElt(i) == (int)(i * 2);
22978 for (unsigned i = SVTNumElts, e = NumElts; i != e && CanFold; ++i)
22979 CanFold = SVOp->getMaskElt(i) < 0;
22982 SDValue BC00 = DAG.getBitcast(VT, BC0.getOperand(0));
22983 SDValue BC01 = DAG.getBitcast(VT, BC0.getOperand(1));
22984 SDValue NewBinOp = DAG.getNode(BC0.getOpcode(), dl, VT, BC00, BC01);
22985 return DAG.getVectorShuffle(VT, dl, NewBinOp, N1, &SVOp->getMask()[0]);
22990 // Combine a vector_shuffle that is equal to build_vector load1, load2, load3,
22991 // load4, <0, 1, 2, 3> into a 128-bit load if the load addresses are
22992 // consecutive, non-overlapping, and in the right order.
22993 SmallVector<SDValue, 16> Elts;
22994 for (unsigned i = 0, e = VT.getVectorNumElements(); i != e; ++i)
22995 Elts.push_back(getShuffleScalarElt(N, i, DAG, 0));
22997 if (SDValue LD = EltsFromConsecutiveLoads(VT, Elts, dl, DAG, true))
23000 if (isTargetShuffle(N->getOpcode())) {
23002 PerformTargetShuffleCombine(SDValue(N, 0), DAG, DCI, Subtarget);
23003 if (Shuffle.getNode())
23006 // Try recursively combining arbitrary sequences of x86 shuffle
23007 // instructions into higher-order shuffles. We do this after combining
23008 // specific PSHUF instruction sequences into their minimal form so that we
23009 // can evaluate how many specialized shuffle instructions are involved in
23010 // a particular chain.
23011 SmallVector<int, 1> NonceMask; // Just a placeholder.
23012 NonceMask.push_back(0);
23013 if (combineX86ShufflesRecursively(SDValue(N, 0), SDValue(N, 0), NonceMask,
23014 /*Depth*/ 1, /*HasPSHUFB*/ false, DAG,
23016 return SDValue(); // This routine will use CombineTo to replace N.
23022 /// XFormVExtractWithShuffleIntoLoad - Check if a vector extract from a target
23023 /// specific shuffle of a load can be folded into a single element load.
23024 /// Similar handling for VECTOR_SHUFFLE is performed by DAGCombiner, but
23025 /// shuffles have been custom lowered so we need to handle those here.
23026 static SDValue XFormVExtractWithShuffleIntoLoad(SDNode *N, SelectionDAG &DAG,
23027 TargetLowering::DAGCombinerInfo &DCI) {
23028 if (DCI.isBeforeLegalizeOps())
23031 SDValue InVec = N->getOperand(0);
23032 SDValue EltNo = N->getOperand(1);
23034 if (!isa<ConstantSDNode>(EltNo))
23037 EVT OriginalVT = InVec.getValueType();
23039 if (InVec.getOpcode() == ISD::BITCAST) {
23040 // Don't duplicate a load with other uses.
23041 if (!InVec.hasOneUse())
23043 EVT BCVT = InVec.getOperand(0).getValueType();
23044 if (!BCVT.isVector() ||
23045 BCVT.getVectorNumElements() != OriginalVT.getVectorNumElements())
23047 InVec = InVec.getOperand(0);
23050 EVT CurrentVT = InVec.getValueType();
23052 if (!isTargetShuffle(InVec.getOpcode()))
23055 // Don't duplicate a load with other uses.
23056 if (!InVec.hasOneUse())
23059 SmallVector<int, 16> ShuffleMask;
23061 if (!getTargetShuffleMask(InVec.getNode(), CurrentVT.getSimpleVT(),
23062 ShuffleMask, UnaryShuffle))
23065 // Select the input vector, guarding against out of range extract vector.
23066 unsigned NumElems = CurrentVT.getVectorNumElements();
23067 int Elt = cast<ConstantSDNode>(EltNo)->getZExtValue();
23068 int Idx = (Elt > (int)NumElems) ? -1 : ShuffleMask[Elt];
23069 SDValue LdNode = (Idx < (int)NumElems) ? InVec.getOperand(0)
23070 : InVec.getOperand(1);
23072 // If inputs to shuffle are the same for both ops, then allow 2 uses
23073 unsigned AllowedUses = InVec.getNumOperands() > 1 &&
23074 InVec.getOperand(0) == InVec.getOperand(1) ? 2 : 1;
23076 if (LdNode.getOpcode() == ISD::BITCAST) {
23077 // Don't duplicate a load with other uses.
23078 if (!LdNode.getNode()->hasNUsesOfValue(AllowedUses, 0))
23081 AllowedUses = 1; // only allow 1 load use if we have a bitcast
23082 LdNode = LdNode.getOperand(0);
23085 if (!ISD::isNormalLoad(LdNode.getNode()))
23088 LoadSDNode *LN0 = cast<LoadSDNode>(LdNode);
23090 if (!LN0 ||!LN0->hasNUsesOfValue(AllowedUses, 0) || LN0->isVolatile())
23093 EVT EltVT = N->getValueType(0);
23094 // If there's a bitcast before the shuffle, check if the load type and
23095 // alignment is valid.
23096 unsigned Align = LN0->getAlignment();
23097 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
23098 unsigned NewAlign = DAG.getDataLayout().getABITypeAlignment(
23099 EltVT.getTypeForEVT(*DAG.getContext()));
23101 if (NewAlign > Align || !TLI.isOperationLegalOrCustom(ISD::LOAD, EltVT))
23104 // All checks match so transform back to vector_shuffle so that DAG combiner
23105 // can finish the job
23108 // Create shuffle node taking into account the case that its a unary shuffle
23109 SDValue Shuffle = (UnaryShuffle) ? DAG.getUNDEF(CurrentVT)
23110 : InVec.getOperand(1);
23111 Shuffle = DAG.getVectorShuffle(CurrentVT, dl,
23112 InVec.getOperand(0), Shuffle,
23114 Shuffle = DAG.getBitcast(OriginalVT, Shuffle);
23115 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, N->getValueType(0), Shuffle,
23119 static SDValue PerformBITCASTCombine(SDNode *N, SelectionDAG &DAG,
23120 const X86Subtarget *Subtarget) {
23121 SDValue N0 = N->getOperand(0);
23122 EVT VT = N->getValueType(0);
23124 // Detect bitcasts between i32 to x86mmx low word. Since MMX types are
23125 // special and don't usually play with other vector types, it's better to
23126 // handle them early to be sure we emit efficient code by avoiding
23127 // store-load conversions.
23128 if (VT == MVT::x86mmx && N0.getOpcode() == ISD::BUILD_VECTOR &&
23129 N0.getValueType() == MVT::v2i32 &&
23130 isa<ConstantSDNode>(N0.getOperand(1))) {
23131 SDValue N00 = N0->getOperand(0);
23132 if (N0.getConstantOperandVal(1) == 0 && N00.getValueType() == MVT::i32)
23133 return DAG.getNode(X86ISD::MMX_MOVW2D, SDLoc(N00), VT, N00);
23136 // Convert a bitcasted integer logic operation that has one bitcasted
23137 // floating-point operand and one constant operand into a floating-point
23138 // logic operation. This may create a load of the constant, but that is
23139 // cheaper than materializing the constant in an integer register and
23140 // transferring it to an SSE register or transferring the SSE operand to
23141 // integer register and back.
23143 switch (N0.getOpcode()) {
23144 case ISD::AND: FPOpcode = X86ISD::FAND; break;
23145 case ISD::OR: FPOpcode = X86ISD::FOR; break;
23146 case ISD::XOR: FPOpcode = X86ISD::FXOR; break;
23147 default: return SDValue();
23149 if (((Subtarget->hasSSE1() && VT == MVT::f32) ||
23150 (Subtarget->hasSSE2() && VT == MVT::f64)) &&
23151 isa<ConstantSDNode>(N0.getOperand(1)) &&
23152 N0.getOperand(0).getOpcode() == ISD::BITCAST &&
23153 N0.getOperand(0).getOperand(0).getValueType() == VT) {
23154 SDValue N000 = N0.getOperand(0).getOperand(0);
23155 SDValue FPConst = DAG.getBitcast(VT, N0.getOperand(1));
23156 return DAG.getNode(FPOpcode, SDLoc(N0), VT, N000, FPConst);
23162 /// PerformEXTRACT_VECTOR_ELTCombine - Detect vector gather/scatter index
23163 /// generation and convert it from being a bunch of shuffles and extracts
23164 /// into a somewhat faster sequence. For i686, the best sequence is apparently
23165 /// storing the value and loading scalars back, while for x64 we should
23166 /// use 64-bit extracts and shifts.
23167 static SDValue PerformEXTRACT_VECTOR_ELTCombine(SDNode *N, SelectionDAG &DAG,
23168 TargetLowering::DAGCombinerInfo &DCI) {
23169 if (SDValue NewOp = XFormVExtractWithShuffleIntoLoad(N, DAG, DCI))
23172 SDValue InputVector = N->getOperand(0);
23173 SDLoc dl(InputVector);
23174 // Detect mmx to i32 conversion through a v2i32 elt extract.
23175 if (InputVector.getOpcode() == ISD::BITCAST && InputVector.hasOneUse() &&
23176 N->getValueType(0) == MVT::i32 &&
23177 InputVector.getValueType() == MVT::v2i32) {
23179 // The bitcast source is a direct mmx result.
23180 SDValue MMXSrc = InputVector.getNode()->getOperand(0);
23181 if (MMXSrc.getValueType() == MVT::x86mmx)
23182 return DAG.getNode(X86ISD::MMX_MOVD2W, SDLoc(InputVector),
23183 N->getValueType(0),
23184 InputVector.getNode()->getOperand(0));
23186 // The mmx is indirect: (i64 extract_elt (v1i64 bitcast (x86mmx ...))).
23187 if (MMXSrc.getOpcode() == ISD::EXTRACT_VECTOR_ELT && MMXSrc.hasOneUse() &&
23188 MMXSrc.getValueType() == MVT::i64) {
23189 SDValue MMXSrcOp = MMXSrc.getOperand(0);
23190 if (MMXSrcOp.hasOneUse() && MMXSrcOp.getOpcode() == ISD::BITCAST &&
23191 MMXSrcOp.getValueType() == MVT::v1i64 &&
23192 MMXSrcOp.getOperand(0).getValueType() == MVT::x86mmx)
23193 return DAG.getNode(X86ISD::MMX_MOVD2W, SDLoc(InputVector),
23194 N->getValueType(0), MMXSrcOp.getOperand(0));
23198 EVT VT = N->getValueType(0);
23200 if (VT == MVT::i1 && dyn_cast<ConstantSDNode>(N->getOperand(1)) &&
23201 InputVector.getOpcode() == ISD::BITCAST &&
23202 dyn_cast<ConstantSDNode>(InputVector.getOperand(0))) {
23203 uint64_t ExtractedElt =
23204 cast<ConstantSDNode>(N->getOperand(1))->getZExtValue();
23205 uint64_t InputValue =
23206 cast<ConstantSDNode>(InputVector.getOperand(0))->getZExtValue();
23207 uint64_t Res = (InputValue >> ExtractedElt) & 1;
23208 return DAG.getConstant(Res, dl, MVT::i1);
23210 // Only operate on vectors of 4 elements, where the alternative shuffling
23211 // gets to be more expensive.
23212 if (InputVector.getValueType() != MVT::v4i32)
23215 // Check whether every use of InputVector is an EXTRACT_VECTOR_ELT with a
23216 // single use which is a sign-extend or zero-extend, and all elements are
23218 SmallVector<SDNode *, 4> Uses;
23219 unsigned ExtractedElements = 0;
23220 for (SDNode::use_iterator UI = InputVector.getNode()->use_begin(),
23221 UE = InputVector.getNode()->use_end(); UI != UE; ++UI) {
23222 if (UI.getUse().getResNo() != InputVector.getResNo())
23225 SDNode *Extract = *UI;
23226 if (Extract->getOpcode() != ISD::EXTRACT_VECTOR_ELT)
23229 if (Extract->getValueType(0) != MVT::i32)
23231 if (!Extract->hasOneUse())
23233 if (Extract->use_begin()->getOpcode() != ISD::SIGN_EXTEND &&
23234 Extract->use_begin()->getOpcode() != ISD::ZERO_EXTEND)
23236 if (!isa<ConstantSDNode>(Extract->getOperand(1)))
23239 // Record which element was extracted.
23240 ExtractedElements |=
23241 1 << cast<ConstantSDNode>(Extract->getOperand(1))->getZExtValue();
23243 Uses.push_back(Extract);
23246 // If not all the elements were used, this may not be worthwhile.
23247 if (ExtractedElements != 15)
23250 // Ok, we've now decided to do the transformation.
23251 // If 64-bit shifts are legal, use the extract-shift sequence,
23252 // otherwise bounce the vector off the cache.
23253 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
23256 if (TLI.isOperationLegal(ISD::SRA, MVT::i64)) {
23257 SDValue Cst = DAG.getBitcast(MVT::v2i64, InputVector);
23258 auto &DL = DAG.getDataLayout();
23259 EVT VecIdxTy = DAG.getTargetLoweringInfo().getVectorIdxTy(DL);
23260 SDValue BottomHalf = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i64, Cst,
23261 DAG.getConstant(0, dl, VecIdxTy));
23262 SDValue TopHalf = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i64, Cst,
23263 DAG.getConstant(1, dl, VecIdxTy));
23265 SDValue ShAmt = DAG.getConstant(
23266 32, dl, DAG.getTargetLoweringInfo().getShiftAmountTy(MVT::i64, DL));
23267 Vals[0] = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, BottomHalf);
23268 Vals[1] = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32,
23269 DAG.getNode(ISD::SRA, dl, MVT::i64, BottomHalf, ShAmt));
23270 Vals[2] = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, TopHalf);
23271 Vals[3] = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32,
23272 DAG.getNode(ISD::SRA, dl, MVT::i64, TopHalf, ShAmt));
23274 // Store the value to a temporary stack slot.
23275 SDValue StackPtr = DAG.CreateStackTemporary(InputVector.getValueType());
23276 SDValue Ch = DAG.getStore(DAG.getEntryNode(), dl, InputVector, StackPtr,
23277 MachinePointerInfo(), false, false, 0);
23279 EVT ElementType = InputVector.getValueType().getVectorElementType();
23280 unsigned EltSize = ElementType.getSizeInBits() / 8;
23282 // Replace each use (extract) with a load of the appropriate element.
23283 for (unsigned i = 0; i < 4; ++i) {
23284 uint64_t Offset = EltSize * i;
23285 auto PtrVT = TLI.getPointerTy(DAG.getDataLayout());
23286 SDValue OffsetVal = DAG.getConstant(Offset, dl, PtrVT);
23288 SDValue ScalarAddr =
23289 DAG.getNode(ISD::ADD, dl, PtrVT, StackPtr, OffsetVal);
23291 // Load the scalar.
23292 Vals[i] = DAG.getLoad(ElementType, dl, Ch,
23293 ScalarAddr, MachinePointerInfo(),
23294 false, false, false, 0);
23299 // Replace the extracts
23300 for (SmallVectorImpl<SDNode *>::iterator UI = Uses.begin(),
23301 UE = Uses.end(); UI != UE; ++UI) {
23302 SDNode *Extract = *UI;
23304 SDValue Idx = Extract->getOperand(1);
23305 uint64_t IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
23306 DAG.ReplaceAllUsesOfValueWith(SDValue(Extract, 0), Vals[IdxVal]);
23309 // The replacement was made in place; don't return anything.
23314 transformVSELECTtoBlendVECTOR_SHUFFLE(SDNode *N, SelectionDAG &DAG,
23315 const X86Subtarget *Subtarget) {
23317 SDValue Cond = N->getOperand(0);
23318 SDValue LHS = N->getOperand(1);
23319 SDValue RHS = N->getOperand(2);
23321 if (Cond.getOpcode() == ISD::SIGN_EXTEND) {
23322 SDValue CondSrc = Cond->getOperand(0);
23323 if (CondSrc->getOpcode() == ISD::SIGN_EXTEND_INREG)
23324 Cond = CondSrc->getOperand(0);
23327 if (!ISD::isBuildVectorOfConstantSDNodes(Cond.getNode()))
23330 // A vselect where all conditions and data are constants can be optimized into
23331 // a single vector load by SelectionDAGLegalize::ExpandBUILD_VECTOR().
23332 if (ISD::isBuildVectorOfConstantSDNodes(LHS.getNode()) &&
23333 ISD::isBuildVectorOfConstantSDNodes(RHS.getNode()))
23336 unsigned MaskValue = 0;
23337 if (!BUILD_VECTORtoBlendMask(cast<BuildVectorSDNode>(Cond), MaskValue))
23340 MVT VT = N->getSimpleValueType(0);
23341 unsigned NumElems = VT.getVectorNumElements();
23342 SmallVector<int, 8> ShuffleMask(NumElems, -1);
23343 for (unsigned i = 0; i < NumElems; ++i) {
23344 // Be sure we emit undef where we can.
23345 if (Cond.getOperand(i)->getOpcode() == ISD::UNDEF)
23346 ShuffleMask[i] = -1;
23348 ShuffleMask[i] = i + NumElems * ((MaskValue >> i) & 1);
23351 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
23352 if (!TLI.isShuffleMaskLegal(ShuffleMask, VT))
23354 return DAG.getVectorShuffle(VT, dl, LHS, RHS, &ShuffleMask[0]);
23357 /// PerformSELECTCombine - Do target-specific dag combines on SELECT and VSELECT
23359 static SDValue PerformSELECTCombine(SDNode *N, SelectionDAG &DAG,
23360 TargetLowering::DAGCombinerInfo &DCI,
23361 const X86Subtarget *Subtarget) {
23363 SDValue Cond = N->getOperand(0);
23364 // Get the LHS/RHS of the select.
23365 SDValue LHS = N->getOperand(1);
23366 SDValue RHS = N->getOperand(2);
23367 EVT VT = LHS.getValueType();
23368 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
23370 // If we have SSE[12] support, try to form min/max nodes. SSE min/max
23371 // instructions match the semantics of the common C idiom x<y?x:y but not
23372 // x<=y?x:y, because of how they handle negative zero (which can be
23373 // ignored in unsafe-math mode).
23374 // We also try to create v2f32 min/max nodes, which we later widen to v4f32.
23375 if (Cond.getOpcode() == ISD::SETCC && VT.isFloatingPoint() &&
23376 VT != MVT::f80 && (TLI.isTypeLegal(VT) || VT == MVT::v2f32) &&
23377 (Subtarget->hasSSE2() ||
23378 (Subtarget->hasSSE1() && VT.getScalarType() == MVT::f32))) {
23379 ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
23381 unsigned Opcode = 0;
23382 // Check for x CC y ? x : y.
23383 if (DAG.isEqualTo(LHS, Cond.getOperand(0)) &&
23384 DAG.isEqualTo(RHS, Cond.getOperand(1))) {
23388 // Converting this to a min would handle NaNs incorrectly, and swapping
23389 // the operands would cause it to handle comparisons between positive
23390 // and negative zero incorrectly.
23391 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)) {
23392 if (!DAG.getTarget().Options.UnsafeFPMath &&
23393 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
23395 std::swap(LHS, RHS);
23397 Opcode = X86ISD::FMIN;
23400 // Converting this to a min would handle comparisons between positive
23401 // and negative zero incorrectly.
23402 if (!DAG.getTarget().Options.UnsafeFPMath &&
23403 !DAG.isKnownNeverZero(LHS) && !DAG.isKnownNeverZero(RHS))
23405 Opcode = X86ISD::FMIN;
23408 // Converting this to a min would handle both negative zeros and NaNs
23409 // incorrectly, but we can swap the operands to fix both.
23410 std::swap(LHS, RHS);
23414 Opcode = X86ISD::FMIN;
23418 // Converting this to a max would handle comparisons between positive
23419 // and negative zero incorrectly.
23420 if (!DAG.getTarget().Options.UnsafeFPMath &&
23421 !DAG.isKnownNeverZero(LHS) && !DAG.isKnownNeverZero(RHS))
23423 Opcode = X86ISD::FMAX;
23426 // Converting this to a max would handle NaNs incorrectly, and swapping
23427 // the operands would cause it to handle comparisons between positive
23428 // and negative zero incorrectly.
23429 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)) {
23430 if (!DAG.getTarget().Options.UnsafeFPMath &&
23431 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
23433 std::swap(LHS, RHS);
23435 Opcode = X86ISD::FMAX;
23438 // Converting this to a max would handle both negative zeros and NaNs
23439 // incorrectly, but we can swap the operands to fix both.
23440 std::swap(LHS, RHS);
23444 Opcode = X86ISD::FMAX;
23447 // Check for x CC y ? y : x -- a min/max with reversed arms.
23448 } else if (DAG.isEqualTo(LHS, Cond.getOperand(1)) &&
23449 DAG.isEqualTo(RHS, Cond.getOperand(0))) {
23453 // Converting this to a min would handle comparisons between positive
23454 // and negative zero incorrectly, and swapping the operands would
23455 // cause it to handle NaNs incorrectly.
23456 if (!DAG.getTarget().Options.UnsafeFPMath &&
23457 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS))) {
23458 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS))
23460 std::swap(LHS, RHS);
23462 Opcode = X86ISD::FMIN;
23465 // Converting this to a min would handle NaNs incorrectly.
23466 if (!DAG.getTarget().Options.UnsafeFPMath &&
23467 (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)))
23469 Opcode = X86ISD::FMIN;
23472 // Converting this to a min would handle both negative zeros and NaNs
23473 // incorrectly, but we can swap the operands to fix both.
23474 std::swap(LHS, RHS);
23478 Opcode = X86ISD::FMIN;
23482 // Converting this to a max would handle NaNs incorrectly.
23483 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS))
23485 Opcode = X86ISD::FMAX;
23488 // Converting this to a max would handle comparisons between positive
23489 // and negative zero incorrectly, and swapping the operands would
23490 // cause it to handle NaNs incorrectly.
23491 if (!DAG.getTarget().Options.UnsafeFPMath &&
23492 !DAG.isKnownNeverZero(LHS) && !DAG.isKnownNeverZero(RHS)) {
23493 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS))
23495 std::swap(LHS, RHS);
23497 Opcode = X86ISD::FMAX;
23500 // Converting this to a max would handle both negative zeros and NaNs
23501 // incorrectly, but we can swap the operands to fix both.
23502 std::swap(LHS, RHS);
23506 Opcode = X86ISD::FMAX;
23512 return DAG.getNode(Opcode, DL, N->getValueType(0), LHS, RHS);
23515 EVT CondVT = Cond.getValueType();
23516 if (Subtarget->hasAVX512() && VT.isVector() && CondVT.isVector() &&
23517 CondVT.getVectorElementType() == MVT::i1) {
23518 // v16i8 (select v16i1, v16i8, v16i8) does not have a proper
23519 // lowering on KNL. In this case we convert it to
23520 // v16i8 (select v16i8, v16i8, v16i8) and use AVX instruction.
23521 // The same situation for all 128 and 256-bit vectors of i8 and i16.
23522 // Since SKX these selects have a proper lowering.
23523 EVT OpVT = LHS.getValueType();
23524 if ((OpVT.is128BitVector() || OpVT.is256BitVector()) &&
23525 (OpVT.getVectorElementType() == MVT::i8 ||
23526 OpVT.getVectorElementType() == MVT::i16) &&
23527 !(Subtarget->hasBWI() && Subtarget->hasVLX())) {
23528 Cond = DAG.getNode(ISD::SIGN_EXTEND, DL, OpVT, Cond);
23529 DCI.AddToWorklist(Cond.getNode());
23530 return DAG.getNode(N->getOpcode(), DL, OpVT, Cond, LHS, RHS);
23533 // If this is a select between two integer constants, try to do some
23535 if (ConstantSDNode *TrueC = dyn_cast<ConstantSDNode>(LHS)) {
23536 if (ConstantSDNode *FalseC = dyn_cast<ConstantSDNode>(RHS))
23537 // Don't do this for crazy integer types.
23538 if (DAG.getTargetLoweringInfo().isTypeLegal(LHS.getValueType())) {
23539 // If this is efficiently invertible, canonicalize the LHSC/RHSC values
23540 // so that TrueC (the true value) is larger than FalseC.
23541 bool NeedsCondInvert = false;
23543 if (TrueC->getAPIntValue().ult(FalseC->getAPIntValue()) &&
23544 // Efficiently invertible.
23545 (Cond.getOpcode() == ISD::SETCC || // setcc -> invertible.
23546 (Cond.getOpcode() == ISD::XOR && // xor(X, C) -> invertible.
23547 isa<ConstantSDNode>(Cond.getOperand(1))))) {
23548 NeedsCondInvert = true;
23549 std::swap(TrueC, FalseC);
23552 // Optimize C ? 8 : 0 -> zext(C) << 3. Likewise for any pow2/0.
23553 if (FalseC->getAPIntValue() == 0 &&
23554 TrueC->getAPIntValue().isPowerOf2()) {
23555 if (NeedsCondInvert) // Invert the condition if needed.
23556 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
23557 DAG.getConstant(1, DL, Cond.getValueType()));
23559 // Zero extend the condition if needed.
23560 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, LHS.getValueType(), Cond);
23562 unsigned ShAmt = TrueC->getAPIntValue().logBase2();
23563 return DAG.getNode(ISD::SHL, DL, LHS.getValueType(), Cond,
23564 DAG.getConstant(ShAmt, DL, MVT::i8));
23567 // Optimize Cond ? cst+1 : cst -> zext(setcc(C)+cst.
23568 if (FalseC->getAPIntValue()+1 == TrueC->getAPIntValue()) {
23569 if (NeedsCondInvert) // Invert the condition if needed.
23570 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
23571 DAG.getConstant(1, DL, Cond.getValueType()));
23573 // Zero extend the condition if needed.
23574 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL,
23575 FalseC->getValueType(0), Cond);
23576 return DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
23577 SDValue(FalseC, 0));
23580 // Optimize cases that will turn into an LEA instruction. This requires
23581 // an i32 or i64 and an efficient multiplier (1, 2, 3, 4, 5, 8, 9).
23582 if (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i64) {
23583 uint64_t Diff = TrueC->getZExtValue()-FalseC->getZExtValue();
23584 if (N->getValueType(0) == MVT::i32) Diff = (unsigned)Diff;
23586 bool isFastMultiplier = false;
23588 switch ((unsigned char)Diff) {
23590 case 1: // result = add base, cond
23591 case 2: // result = lea base( , cond*2)
23592 case 3: // result = lea base(cond, cond*2)
23593 case 4: // result = lea base( , cond*4)
23594 case 5: // result = lea base(cond, cond*4)
23595 case 8: // result = lea base( , cond*8)
23596 case 9: // result = lea base(cond, cond*8)
23597 isFastMultiplier = true;
23602 if (isFastMultiplier) {
23603 APInt Diff = TrueC->getAPIntValue()-FalseC->getAPIntValue();
23604 if (NeedsCondInvert) // Invert the condition if needed.
23605 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
23606 DAG.getConstant(1, DL, Cond.getValueType()));
23608 // Zero extend the condition if needed.
23609 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, FalseC->getValueType(0),
23611 // Scale the condition by the difference.
23613 Cond = DAG.getNode(ISD::MUL, DL, Cond.getValueType(), Cond,
23614 DAG.getConstant(Diff, DL,
23615 Cond.getValueType()));
23617 // Add the base if non-zero.
23618 if (FalseC->getAPIntValue() != 0)
23619 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
23620 SDValue(FalseC, 0));
23627 // Canonicalize max and min:
23628 // (x > y) ? x : y -> (x >= y) ? x : y
23629 // (x < y) ? x : y -> (x <= y) ? x : y
23630 // This allows use of COND_S / COND_NS (see TranslateX86CC) which eliminates
23631 // the need for an extra compare
23632 // against zero. e.g.
23633 // (x - y) > 0 : (x - y) ? 0 -> (x - y) >= 0 : (x - y) ? 0
23635 // testl %edi, %edi
23637 // cmovgl %edi, %eax
23641 // cmovsl %eax, %edi
23642 if (N->getOpcode() == ISD::SELECT && Cond.getOpcode() == ISD::SETCC &&
23643 DAG.isEqualTo(LHS, Cond.getOperand(0)) &&
23644 DAG.isEqualTo(RHS, Cond.getOperand(1))) {
23645 ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
23650 ISD::CondCode NewCC = (CC == ISD::SETLT) ? ISD::SETLE : ISD::SETGE;
23651 Cond = DAG.getSetCC(SDLoc(Cond), Cond.getValueType(),
23652 Cond.getOperand(0), Cond.getOperand(1), NewCC);
23653 return DAG.getNode(ISD::SELECT, DL, VT, Cond, LHS, RHS);
23658 // Early exit check
23659 if (!TLI.isTypeLegal(VT))
23662 // Match VSELECTs into subs with unsigned saturation.
23663 if (N->getOpcode() == ISD::VSELECT && Cond.getOpcode() == ISD::SETCC &&
23664 // psubus is available in SSE2 and AVX2 for i8 and i16 vectors.
23665 ((Subtarget->hasSSE2() && (VT == MVT::v16i8 || VT == MVT::v8i16)) ||
23666 (Subtarget->hasAVX2() && (VT == MVT::v32i8 || VT == MVT::v16i16)))) {
23667 ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
23669 // Check if one of the arms of the VSELECT is a zero vector. If it's on the
23670 // left side invert the predicate to simplify logic below.
23672 if (ISD::isBuildVectorAllZeros(LHS.getNode())) {
23674 CC = ISD::getSetCCInverse(CC, true);
23675 } else if (ISD::isBuildVectorAllZeros(RHS.getNode())) {
23679 if (Other.getNode() && Other->getNumOperands() == 2 &&
23680 DAG.isEqualTo(Other->getOperand(0), Cond.getOperand(0))) {
23681 SDValue OpLHS = Other->getOperand(0), OpRHS = Other->getOperand(1);
23682 SDValue CondRHS = Cond->getOperand(1);
23684 // Look for a general sub with unsigned saturation first.
23685 // x >= y ? x-y : 0 --> subus x, y
23686 // x > y ? x-y : 0 --> subus x, y
23687 if ((CC == ISD::SETUGE || CC == ISD::SETUGT) &&
23688 Other->getOpcode() == ISD::SUB && DAG.isEqualTo(OpRHS, CondRHS))
23689 return DAG.getNode(X86ISD::SUBUS, DL, VT, OpLHS, OpRHS);
23691 if (auto *OpRHSBV = dyn_cast<BuildVectorSDNode>(OpRHS))
23692 if (auto *OpRHSConst = OpRHSBV->getConstantSplatNode()) {
23693 if (auto *CondRHSBV = dyn_cast<BuildVectorSDNode>(CondRHS))
23694 if (auto *CondRHSConst = CondRHSBV->getConstantSplatNode())
23695 // If the RHS is a constant we have to reverse the const
23696 // canonicalization.
23697 // x > C-1 ? x+-C : 0 --> subus x, C
23698 if (CC == ISD::SETUGT && Other->getOpcode() == ISD::ADD &&
23699 CondRHSConst->getAPIntValue() ==
23700 (-OpRHSConst->getAPIntValue() - 1))
23701 return DAG.getNode(
23702 X86ISD::SUBUS, DL, VT, OpLHS,
23703 DAG.getConstant(-OpRHSConst->getAPIntValue(), DL, VT));
23705 // Another special case: If C was a sign bit, the sub has been
23706 // canonicalized into a xor.
23707 // FIXME: Would it be better to use computeKnownBits to determine
23708 // whether it's safe to decanonicalize the xor?
23709 // x s< 0 ? x^C : 0 --> subus x, C
23710 if (CC == ISD::SETLT && Other->getOpcode() == ISD::XOR &&
23711 ISD::isBuildVectorAllZeros(CondRHS.getNode()) &&
23712 OpRHSConst->getAPIntValue().isSignBit())
23713 // Note that we have to rebuild the RHS constant here to ensure we
23714 // don't rely on particular values of undef lanes.
23715 return DAG.getNode(
23716 X86ISD::SUBUS, DL, VT, OpLHS,
23717 DAG.getConstant(OpRHSConst->getAPIntValue(), DL, VT));
23722 // Simplify vector selection if condition value type matches vselect
23724 if (N->getOpcode() == ISD::VSELECT && CondVT == VT) {
23725 assert(Cond.getValueType().isVector() &&
23726 "vector select expects a vector selector!");
23728 bool TValIsAllOnes = ISD::isBuildVectorAllOnes(LHS.getNode());
23729 bool FValIsAllZeros = ISD::isBuildVectorAllZeros(RHS.getNode());
23731 // Try invert the condition if true value is not all 1s and false value
23733 if (!TValIsAllOnes && !FValIsAllZeros &&
23734 // Check if the selector will be produced by CMPP*/PCMP*
23735 Cond.getOpcode() == ISD::SETCC &&
23736 // Check if SETCC has already been promoted
23737 TLI.getSetCCResultType(DAG.getDataLayout(), *DAG.getContext(), VT) ==
23739 bool TValIsAllZeros = ISD::isBuildVectorAllZeros(LHS.getNode());
23740 bool FValIsAllOnes = ISD::isBuildVectorAllOnes(RHS.getNode());
23742 if (TValIsAllZeros || FValIsAllOnes) {
23743 SDValue CC = Cond.getOperand(2);
23744 ISD::CondCode NewCC =
23745 ISD::getSetCCInverse(cast<CondCodeSDNode>(CC)->get(),
23746 Cond.getOperand(0).getValueType().isInteger());
23747 Cond = DAG.getSetCC(DL, CondVT, Cond.getOperand(0), Cond.getOperand(1), NewCC);
23748 std::swap(LHS, RHS);
23749 TValIsAllOnes = FValIsAllOnes;
23750 FValIsAllZeros = TValIsAllZeros;
23754 if (TValIsAllOnes || FValIsAllZeros) {
23757 if (TValIsAllOnes && FValIsAllZeros)
23759 else if (TValIsAllOnes)
23761 DAG.getNode(ISD::OR, DL, CondVT, Cond, DAG.getBitcast(CondVT, RHS));
23762 else if (FValIsAllZeros)
23763 Ret = DAG.getNode(ISD::AND, DL, CondVT, Cond,
23764 DAG.getBitcast(CondVT, LHS));
23766 return DAG.getBitcast(VT, Ret);
23770 // We should generate an X86ISD::BLENDI from a vselect if its argument
23771 // is a sign_extend_inreg of an any_extend of a BUILD_VECTOR of
23772 // constants. This specific pattern gets generated when we split a
23773 // selector for a 512 bit vector in a machine without AVX512 (but with
23774 // 256-bit vectors), during legalization:
23776 // (vselect (sign_extend (any_extend (BUILD_VECTOR)) i1) LHS RHS)
23778 // Iff we find this pattern and the build_vectors are built from
23779 // constants, we translate the vselect into a shuffle_vector that we
23780 // know will be matched by LowerVECTOR_SHUFFLEtoBlend.
23781 if ((N->getOpcode() == ISD::VSELECT ||
23782 N->getOpcode() == X86ISD::SHRUNKBLEND) &&
23783 !DCI.isBeforeLegalize() && !VT.is512BitVector()) {
23784 SDValue Shuffle = transformVSELECTtoBlendVECTOR_SHUFFLE(N, DAG, Subtarget);
23785 if (Shuffle.getNode())
23789 // If this is a *dynamic* select (non-constant condition) and we can match
23790 // this node with one of the variable blend instructions, restructure the
23791 // condition so that the blends can use the high bit of each element and use
23792 // SimplifyDemandedBits to simplify the condition operand.
23793 if (N->getOpcode() == ISD::VSELECT && DCI.isBeforeLegalizeOps() &&
23794 !DCI.isBeforeLegalize() &&
23795 !ISD::isBuildVectorOfConstantSDNodes(Cond.getNode())) {
23796 unsigned BitWidth = Cond.getValueType().getScalarType().getSizeInBits();
23798 // Don't optimize vector selects that map to mask-registers.
23802 // We can only handle the cases where VSELECT is directly legal on the
23803 // subtarget. We custom lower VSELECT nodes with constant conditions and
23804 // this makes it hard to see whether a dynamic VSELECT will correctly
23805 // lower, so we both check the operation's status and explicitly handle the
23806 // cases where a *dynamic* blend will fail even though a constant-condition
23807 // blend could be custom lowered.
23808 // FIXME: We should find a better way to handle this class of problems.
23809 // Potentially, we should combine constant-condition vselect nodes
23810 // pre-legalization into shuffles and not mark as many types as custom
23812 if (!TLI.isOperationLegalOrCustom(ISD::VSELECT, VT))
23814 // FIXME: We don't support i16-element blends currently. We could and
23815 // should support them by making *all* the bits in the condition be set
23816 // rather than just the high bit and using an i8-element blend.
23817 if (VT.getScalarType() == MVT::i16)
23819 // Dynamic blending was only available from SSE4.1 onward.
23820 if (VT.getSizeInBits() == 128 && !Subtarget->hasSSE41())
23822 // Byte blends are only available in AVX2
23823 if (VT.getSizeInBits() == 256 && VT.getScalarType() == MVT::i8 &&
23824 !Subtarget->hasAVX2())
23827 assert(BitWidth >= 8 && BitWidth <= 64 && "Invalid mask size");
23828 APInt DemandedMask = APInt::getHighBitsSet(BitWidth, 1);
23830 APInt KnownZero, KnownOne;
23831 TargetLowering::TargetLoweringOpt TLO(DAG, DCI.isBeforeLegalize(),
23832 DCI.isBeforeLegalizeOps());
23833 if (TLO.ShrinkDemandedConstant(Cond, DemandedMask) ||
23834 TLI.SimplifyDemandedBits(Cond, DemandedMask, KnownZero, KnownOne,
23836 // If we changed the computation somewhere in the DAG, this change
23837 // will affect all users of Cond.
23838 // Make sure it is fine and update all the nodes so that we do not
23839 // use the generic VSELECT anymore. Otherwise, we may perform
23840 // wrong optimizations as we messed up with the actual expectation
23841 // for the vector boolean values.
23842 if (Cond != TLO.Old) {
23843 // Check all uses of that condition operand to check whether it will be
23844 // consumed by non-BLEND instructions, which may depend on all bits are
23846 for (SDNode::use_iterator I = Cond->use_begin(), E = Cond->use_end();
23848 if (I->getOpcode() != ISD::VSELECT)
23849 // TODO: Add other opcodes eventually lowered into BLEND.
23852 // Update all the users of the condition, before committing the change,
23853 // so that the VSELECT optimizations that expect the correct vector
23854 // boolean value will not be triggered.
23855 for (SDNode::use_iterator I = Cond->use_begin(), E = Cond->use_end();
23857 DAG.ReplaceAllUsesOfValueWith(
23859 DAG.getNode(X86ISD::SHRUNKBLEND, SDLoc(*I), I->getValueType(0),
23860 Cond, I->getOperand(1), I->getOperand(2)));
23861 DCI.CommitTargetLoweringOpt(TLO);
23864 // At this point, only Cond is changed. Change the condition
23865 // just for N to keep the opportunity to optimize all other
23866 // users their own way.
23867 DAG.ReplaceAllUsesOfValueWith(
23869 DAG.getNode(X86ISD::SHRUNKBLEND, SDLoc(N), N->getValueType(0),
23870 TLO.New, N->getOperand(1), N->getOperand(2)));
23878 // Check whether a boolean test is testing a boolean value generated by
23879 // X86ISD::SETCC. If so, return the operand of that SETCC and proper condition
23882 // Simplify the following patterns:
23883 // (Op (CMP (SETCC Cond EFLAGS) 1) EQ) or
23884 // (Op (CMP (SETCC Cond EFLAGS) 0) NEQ)
23885 // to (Op EFLAGS Cond)
23887 // (Op (CMP (SETCC Cond EFLAGS) 0) EQ) or
23888 // (Op (CMP (SETCC Cond EFLAGS) 1) NEQ)
23889 // to (Op EFLAGS !Cond)
23891 // where Op could be BRCOND or CMOV.
23893 static SDValue checkBoolTestSetCCCombine(SDValue Cmp, X86::CondCode &CC) {
23894 // Quit if not CMP and SUB with its value result used.
23895 if (Cmp.getOpcode() != X86ISD::CMP &&
23896 (Cmp.getOpcode() != X86ISD::SUB || Cmp.getNode()->hasAnyUseOfValue(0)))
23899 // Quit if not used as a boolean value.
23900 if (CC != X86::COND_E && CC != X86::COND_NE)
23903 // Check CMP operands. One of them should be 0 or 1 and the other should be
23904 // an SetCC or extended from it.
23905 SDValue Op1 = Cmp.getOperand(0);
23906 SDValue Op2 = Cmp.getOperand(1);
23909 const ConstantSDNode* C = nullptr;
23910 bool needOppositeCond = (CC == X86::COND_E);
23911 bool checkAgainstTrue = false; // Is it a comparison against 1?
23913 if ((C = dyn_cast<ConstantSDNode>(Op1)))
23915 else if ((C = dyn_cast<ConstantSDNode>(Op2)))
23917 else // Quit if all operands are not constants.
23920 if (C->getZExtValue() == 1) {
23921 needOppositeCond = !needOppositeCond;
23922 checkAgainstTrue = true;
23923 } else if (C->getZExtValue() != 0)
23924 // Quit if the constant is neither 0 or 1.
23927 bool truncatedToBoolWithAnd = false;
23928 // Skip (zext $x), (trunc $x), or (and $x, 1) node.
23929 while (SetCC.getOpcode() == ISD::ZERO_EXTEND ||
23930 SetCC.getOpcode() == ISD::TRUNCATE ||
23931 SetCC.getOpcode() == ISD::AND) {
23932 if (SetCC.getOpcode() == ISD::AND) {
23934 ConstantSDNode *CS;
23935 if ((CS = dyn_cast<ConstantSDNode>(SetCC.getOperand(0))) &&
23936 CS->getZExtValue() == 1)
23938 if ((CS = dyn_cast<ConstantSDNode>(SetCC.getOperand(1))) &&
23939 CS->getZExtValue() == 1)
23943 SetCC = SetCC.getOperand(OpIdx);
23944 truncatedToBoolWithAnd = true;
23946 SetCC = SetCC.getOperand(0);
23949 switch (SetCC.getOpcode()) {
23950 case X86ISD::SETCC_CARRY:
23951 // Since SETCC_CARRY gives output based on R = CF ? ~0 : 0, it's unsafe to
23952 // simplify it if the result of SETCC_CARRY is not canonicalized to 0 or 1,
23953 // i.e. it's a comparison against true but the result of SETCC_CARRY is not
23954 // truncated to i1 using 'and'.
23955 if (checkAgainstTrue && !truncatedToBoolWithAnd)
23957 assert(X86::CondCode(SetCC.getConstantOperandVal(0)) == X86::COND_B &&
23958 "Invalid use of SETCC_CARRY!");
23960 case X86ISD::SETCC:
23961 // Set the condition code or opposite one if necessary.
23962 CC = X86::CondCode(SetCC.getConstantOperandVal(0));
23963 if (needOppositeCond)
23964 CC = X86::GetOppositeBranchCondition(CC);
23965 return SetCC.getOperand(1);
23966 case X86ISD::CMOV: {
23967 // Check whether false/true value has canonical one, i.e. 0 or 1.
23968 ConstantSDNode *FVal = dyn_cast<ConstantSDNode>(SetCC.getOperand(0));
23969 ConstantSDNode *TVal = dyn_cast<ConstantSDNode>(SetCC.getOperand(1));
23970 // Quit if true value is not a constant.
23973 // Quit if false value is not a constant.
23975 SDValue Op = SetCC.getOperand(0);
23976 // Skip 'zext' or 'trunc' node.
23977 if (Op.getOpcode() == ISD::ZERO_EXTEND ||
23978 Op.getOpcode() == ISD::TRUNCATE)
23979 Op = Op.getOperand(0);
23980 // A special case for rdrand/rdseed, where 0 is set if false cond is
23982 if ((Op.getOpcode() != X86ISD::RDRAND &&
23983 Op.getOpcode() != X86ISD::RDSEED) || Op.getResNo() != 0)
23986 // Quit if false value is not the constant 0 or 1.
23987 bool FValIsFalse = true;
23988 if (FVal && FVal->getZExtValue() != 0) {
23989 if (FVal->getZExtValue() != 1)
23991 // If FVal is 1, opposite cond is needed.
23992 needOppositeCond = !needOppositeCond;
23993 FValIsFalse = false;
23995 // Quit if TVal is not the constant opposite of FVal.
23996 if (FValIsFalse && TVal->getZExtValue() != 1)
23998 if (!FValIsFalse && TVal->getZExtValue() != 0)
24000 CC = X86::CondCode(SetCC.getConstantOperandVal(2));
24001 if (needOppositeCond)
24002 CC = X86::GetOppositeBranchCondition(CC);
24003 return SetCC.getOperand(3);
24010 /// Check whether Cond is an AND/OR of SETCCs off of the same EFLAGS.
24012 /// (X86or (X86setcc) (X86setcc))
24013 /// (X86cmp (and (X86setcc) (X86setcc)), 0)
24014 static bool checkBoolTestAndOrSetCCCombine(SDValue Cond, X86::CondCode &CC0,
24015 X86::CondCode &CC1, SDValue &Flags,
24017 if (Cond->getOpcode() == X86ISD::CMP) {
24018 ConstantSDNode *CondOp1C = dyn_cast<ConstantSDNode>(Cond->getOperand(1));
24019 if (!CondOp1C || !CondOp1C->isNullValue())
24022 Cond = Cond->getOperand(0);
24027 SDValue SetCC0, SetCC1;
24028 switch (Cond->getOpcode()) {
24029 default: return false;
24036 SetCC0 = Cond->getOperand(0);
24037 SetCC1 = Cond->getOperand(1);
24041 // Make sure we have SETCC nodes, using the same flags value.
24042 if (SetCC0.getOpcode() != X86ISD::SETCC ||
24043 SetCC1.getOpcode() != X86ISD::SETCC ||
24044 SetCC0->getOperand(1) != SetCC1->getOperand(1))
24047 CC0 = (X86::CondCode)SetCC0->getConstantOperandVal(0);
24048 CC1 = (X86::CondCode)SetCC1->getConstantOperandVal(0);
24049 Flags = SetCC0->getOperand(1);
24053 /// Optimize X86ISD::CMOV [LHS, RHS, CONDCODE (e.g. X86::COND_NE), CONDVAL]
24054 static SDValue PerformCMOVCombine(SDNode *N, SelectionDAG &DAG,
24055 TargetLowering::DAGCombinerInfo &DCI,
24056 const X86Subtarget *Subtarget) {
24059 // If the flag operand isn't dead, don't touch this CMOV.
24060 if (N->getNumValues() == 2 && !SDValue(N, 1).use_empty())
24063 SDValue FalseOp = N->getOperand(0);
24064 SDValue TrueOp = N->getOperand(1);
24065 X86::CondCode CC = (X86::CondCode)N->getConstantOperandVal(2);
24066 SDValue Cond = N->getOperand(3);
24068 if (CC == X86::COND_E || CC == X86::COND_NE) {
24069 switch (Cond.getOpcode()) {
24073 // If operand of BSR / BSF are proven never zero, then ZF cannot be set.
24074 if (DAG.isKnownNeverZero(Cond.getOperand(0)))
24075 return (CC == X86::COND_E) ? FalseOp : TrueOp;
24081 Flags = checkBoolTestSetCCCombine(Cond, CC);
24082 if (Flags.getNode() &&
24083 // Extra check as FCMOV only supports a subset of X86 cond.
24084 (FalseOp.getValueType() != MVT::f80 || hasFPCMov(CC))) {
24085 SDValue Ops[] = { FalseOp, TrueOp,
24086 DAG.getConstant(CC, DL, MVT::i8), Flags };
24087 return DAG.getNode(X86ISD::CMOV, DL, N->getVTList(), Ops);
24090 // If this is a select between two integer constants, try to do some
24091 // optimizations. Note that the operands are ordered the opposite of SELECT
24093 if (ConstantSDNode *TrueC = dyn_cast<ConstantSDNode>(TrueOp)) {
24094 if (ConstantSDNode *FalseC = dyn_cast<ConstantSDNode>(FalseOp)) {
24095 // Canonicalize the TrueC/FalseC values so that TrueC (the true value) is
24096 // larger than FalseC (the false value).
24097 if (TrueC->getAPIntValue().ult(FalseC->getAPIntValue())) {
24098 CC = X86::GetOppositeBranchCondition(CC);
24099 std::swap(TrueC, FalseC);
24100 std::swap(TrueOp, FalseOp);
24103 // Optimize C ? 8 : 0 -> zext(setcc(C)) << 3. Likewise for any pow2/0.
24104 // This is efficient for any integer data type (including i8/i16) and
24106 if (FalseC->getAPIntValue() == 0 && TrueC->getAPIntValue().isPowerOf2()) {
24107 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
24108 DAG.getConstant(CC, DL, MVT::i8), Cond);
24110 // Zero extend the condition if needed.
24111 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, TrueC->getValueType(0), Cond);
24113 unsigned ShAmt = TrueC->getAPIntValue().logBase2();
24114 Cond = DAG.getNode(ISD::SHL, DL, Cond.getValueType(), Cond,
24115 DAG.getConstant(ShAmt, DL, MVT::i8));
24116 if (N->getNumValues() == 2) // Dead flag value?
24117 return DCI.CombineTo(N, Cond, SDValue());
24121 // Optimize Cond ? cst+1 : cst -> zext(setcc(C)+cst. This is efficient
24122 // for any integer data type, including i8/i16.
24123 if (FalseC->getAPIntValue()+1 == TrueC->getAPIntValue()) {
24124 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
24125 DAG.getConstant(CC, DL, MVT::i8), Cond);
24127 // Zero extend the condition if needed.
24128 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL,
24129 FalseC->getValueType(0), Cond);
24130 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
24131 SDValue(FalseC, 0));
24133 if (N->getNumValues() == 2) // Dead flag value?
24134 return DCI.CombineTo(N, Cond, SDValue());
24138 // Optimize cases that will turn into an LEA instruction. This requires
24139 // an i32 or i64 and an efficient multiplier (1, 2, 3, 4, 5, 8, 9).
24140 if (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i64) {
24141 uint64_t Diff = TrueC->getZExtValue()-FalseC->getZExtValue();
24142 if (N->getValueType(0) == MVT::i32) Diff = (unsigned)Diff;
24144 bool isFastMultiplier = false;
24146 switch ((unsigned char)Diff) {
24148 case 1: // result = add base, cond
24149 case 2: // result = lea base( , cond*2)
24150 case 3: // result = lea base(cond, cond*2)
24151 case 4: // result = lea base( , cond*4)
24152 case 5: // result = lea base(cond, cond*4)
24153 case 8: // result = lea base( , cond*8)
24154 case 9: // result = lea base(cond, cond*8)
24155 isFastMultiplier = true;
24160 if (isFastMultiplier) {
24161 APInt Diff = TrueC->getAPIntValue()-FalseC->getAPIntValue();
24162 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
24163 DAG.getConstant(CC, DL, MVT::i8), Cond);
24164 // Zero extend the condition if needed.
24165 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, FalseC->getValueType(0),
24167 // Scale the condition by the difference.
24169 Cond = DAG.getNode(ISD::MUL, DL, Cond.getValueType(), Cond,
24170 DAG.getConstant(Diff, DL, Cond.getValueType()));
24172 // Add the base if non-zero.
24173 if (FalseC->getAPIntValue() != 0)
24174 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
24175 SDValue(FalseC, 0));
24176 if (N->getNumValues() == 2) // Dead flag value?
24177 return DCI.CombineTo(N, Cond, SDValue());
24184 // Handle these cases:
24185 // (select (x != c), e, c) -> select (x != c), e, x),
24186 // (select (x == c), c, e) -> select (x == c), x, e)
24187 // where the c is an integer constant, and the "select" is the combination
24188 // of CMOV and CMP.
24190 // The rationale for this change is that the conditional-move from a constant
24191 // needs two instructions, however, conditional-move from a register needs
24192 // only one instruction.
24194 // CAVEAT: By replacing a constant with a symbolic value, it may obscure
24195 // some instruction-combining opportunities. This opt needs to be
24196 // postponed as late as possible.
24198 if (!DCI.isBeforeLegalize() && !DCI.isBeforeLegalizeOps()) {
24199 // the DCI.xxxx conditions are provided to postpone the optimization as
24200 // late as possible.
24202 ConstantSDNode *CmpAgainst = nullptr;
24203 if ((Cond.getOpcode() == X86ISD::CMP || Cond.getOpcode() == X86ISD::SUB) &&
24204 (CmpAgainst = dyn_cast<ConstantSDNode>(Cond.getOperand(1))) &&
24205 !isa<ConstantSDNode>(Cond.getOperand(0))) {
24207 if (CC == X86::COND_NE &&
24208 CmpAgainst == dyn_cast<ConstantSDNode>(FalseOp)) {
24209 CC = X86::GetOppositeBranchCondition(CC);
24210 std::swap(TrueOp, FalseOp);
24213 if (CC == X86::COND_E &&
24214 CmpAgainst == dyn_cast<ConstantSDNode>(TrueOp)) {
24215 SDValue Ops[] = { FalseOp, Cond.getOperand(0),
24216 DAG.getConstant(CC, DL, MVT::i8), Cond };
24217 return DAG.getNode(X86ISD::CMOV, DL, N->getVTList (), Ops);
24222 // Fold and/or of setcc's to double CMOV:
24223 // (CMOV F, T, ((cc1 | cc2) != 0)) -> (CMOV (CMOV F, T, cc1), T, cc2)
24224 // (CMOV F, T, ((cc1 & cc2) != 0)) -> (CMOV (CMOV T, F, !cc1), F, !cc2)
24226 // This combine lets us generate:
24227 // cmovcc1 (jcc1 if we don't have CMOV)
24233 // cmovne (jne if we don't have CMOV)
24234 // When we can't use the CMOV instruction, it might increase branch
24236 // When we can use CMOV, or when there is no mispredict, this improves
24237 // throughput and reduces register pressure.
24239 if (CC == X86::COND_NE) {
24241 X86::CondCode CC0, CC1;
24243 if (checkBoolTestAndOrSetCCCombine(Cond, CC0, CC1, Flags, isAndSetCC)) {
24245 std::swap(FalseOp, TrueOp);
24246 CC0 = X86::GetOppositeBranchCondition(CC0);
24247 CC1 = X86::GetOppositeBranchCondition(CC1);
24250 SDValue LOps[] = {FalseOp, TrueOp, DAG.getConstant(CC0, DL, MVT::i8),
24252 SDValue LCMOV = DAG.getNode(X86ISD::CMOV, DL, N->getVTList(), LOps);
24253 SDValue Ops[] = {LCMOV, TrueOp, DAG.getConstant(CC1, DL, MVT::i8), Flags};
24254 SDValue CMOV = DAG.getNode(X86ISD::CMOV, DL, N->getVTList(), Ops);
24255 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), SDValue(CMOV.getNode(), 1));
24263 /// PerformMulCombine - Optimize a single multiply with constant into two
24264 /// in order to implement it with two cheaper instructions, e.g.
24265 /// LEA + SHL, LEA + LEA.
24266 static SDValue PerformMulCombine(SDNode *N, SelectionDAG &DAG,
24267 TargetLowering::DAGCombinerInfo &DCI) {
24268 // An imul is usually smaller than the alternative sequence.
24269 if (DAG.getMachineFunction().getFunction()->optForMinSize())
24272 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
24275 EVT VT = N->getValueType(0);
24276 if (VT != MVT::i64 && VT != MVT::i32)
24279 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1));
24282 uint64_t MulAmt = C->getZExtValue();
24283 if (isPowerOf2_64(MulAmt) || MulAmt == 3 || MulAmt == 5 || MulAmt == 9)
24286 uint64_t MulAmt1 = 0;
24287 uint64_t MulAmt2 = 0;
24288 if ((MulAmt % 9) == 0) {
24290 MulAmt2 = MulAmt / 9;
24291 } else if ((MulAmt % 5) == 0) {
24293 MulAmt2 = MulAmt / 5;
24294 } else if ((MulAmt % 3) == 0) {
24296 MulAmt2 = MulAmt / 3;
24299 (isPowerOf2_64(MulAmt2) || MulAmt2 == 3 || MulAmt2 == 5 || MulAmt2 == 9)){
24302 if (isPowerOf2_64(MulAmt2) &&
24303 !(N->hasOneUse() && N->use_begin()->getOpcode() == ISD::ADD))
24304 // If second multiplifer is pow2, issue it first. We want the multiply by
24305 // 3, 5, or 9 to be folded into the addressing mode unless the lone use
24307 std::swap(MulAmt1, MulAmt2);
24310 if (isPowerOf2_64(MulAmt1))
24311 NewMul = DAG.getNode(ISD::SHL, DL, VT, N->getOperand(0),
24312 DAG.getConstant(Log2_64(MulAmt1), DL, MVT::i8));
24314 NewMul = DAG.getNode(X86ISD::MUL_IMM, DL, VT, N->getOperand(0),
24315 DAG.getConstant(MulAmt1, DL, VT));
24317 if (isPowerOf2_64(MulAmt2))
24318 NewMul = DAG.getNode(ISD::SHL, DL, VT, NewMul,
24319 DAG.getConstant(Log2_64(MulAmt2), DL, MVT::i8));
24321 NewMul = DAG.getNode(X86ISD::MUL_IMM, DL, VT, NewMul,
24322 DAG.getConstant(MulAmt2, DL, VT));
24324 // Do not add new nodes to DAG combiner worklist.
24325 DCI.CombineTo(N, NewMul, false);
24330 static SDValue PerformSHLCombine(SDNode *N, SelectionDAG &DAG) {
24331 SDValue N0 = N->getOperand(0);
24332 SDValue N1 = N->getOperand(1);
24333 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
24334 EVT VT = N0.getValueType();
24336 // fold (shl (and (setcc_c), c1), c2) -> (and setcc_c, (c1 << c2))
24337 // since the result of setcc_c is all zero's or all ones.
24338 if (VT.isInteger() && !VT.isVector() &&
24339 N1C && N0.getOpcode() == ISD::AND &&
24340 N0.getOperand(1).getOpcode() == ISD::Constant) {
24341 SDValue N00 = N0.getOperand(0);
24342 APInt Mask = cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
24343 APInt ShAmt = N1C->getAPIntValue();
24344 Mask = Mask.shl(ShAmt);
24345 bool MaskOK = false;
24346 // We can handle cases concerning bit-widening nodes containing setcc_c if
24347 // we carefully interrogate the mask to make sure we are semantics
24349 // The transform is not safe if the result of C1 << C2 exceeds the bitwidth
24350 // of the underlying setcc_c operation if the setcc_c was zero extended.
24351 // Consider the following example:
24352 // zext(setcc_c) -> i32 0x0000FFFF
24353 // c1 -> i32 0x0000FFFF
24354 // c2 -> i32 0x00000001
24355 // (shl (and (setcc_c), c1), c2) -> i32 0x0001FFFE
24356 // (and setcc_c, (c1 << c2)) -> i32 0x0000FFFE
24357 if (N00.getOpcode() == X86ISD::SETCC_CARRY) {
24359 } else if (N00.getOpcode() == ISD::SIGN_EXTEND &&
24360 N00.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY) {
24362 } else if ((N00.getOpcode() == ISD::ZERO_EXTEND ||
24363 N00.getOpcode() == ISD::ANY_EXTEND) &&
24364 N00.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY) {
24365 MaskOK = Mask.isIntN(N00.getOperand(0).getValueSizeInBits());
24367 if (MaskOK && Mask != 0) {
24369 return DAG.getNode(ISD::AND, DL, VT, N00, DAG.getConstant(Mask, DL, VT));
24373 // Hardware support for vector shifts is sparse which makes us scalarize the
24374 // vector operations in many cases. Also, on sandybridge ADD is faster than
24376 // (shl V, 1) -> add V,V
24377 if (auto *N1BV = dyn_cast<BuildVectorSDNode>(N1))
24378 if (auto *N1SplatC = N1BV->getConstantSplatNode()) {
24379 assert(N0.getValueType().isVector() && "Invalid vector shift type");
24380 // We shift all of the values by one. In many cases we do not have
24381 // hardware support for this operation. This is better expressed as an ADD
24383 if (N1SplatC->getAPIntValue() == 1)
24384 return DAG.getNode(ISD::ADD, SDLoc(N), VT, N0, N0);
24390 /// \brief Returns a vector of 0s if the node in input is a vector logical
24391 /// shift by a constant amount which is known to be bigger than or equal
24392 /// to the vector element size in bits.
24393 static SDValue performShiftToAllZeros(SDNode *N, SelectionDAG &DAG,
24394 const X86Subtarget *Subtarget) {
24395 EVT VT = N->getValueType(0);
24397 if (VT != MVT::v2i64 && VT != MVT::v4i32 && VT != MVT::v8i16 &&
24398 (!Subtarget->hasInt256() ||
24399 (VT != MVT::v4i64 && VT != MVT::v8i32 && VT != MVT::v16i16)))
24402 SDValue Amt = N->getOperand(1);
24404 if (auto *AmtBV = dyn_cast<BuildVectorSDNode>(Amt))
24405 if (auto *AmtSplat = AmtBV->getConstantSplatNode()) {
24406 APInt ShiftAmt = AmtSplat->getAPIntValue();
24407 unsigned MaxAmount = VT.getVectorElementType().getSizeInBits();
24409 // SSE2/AVX2 logical shifts always return a vector of 0s
24410 // if the shift amount is bigger than or equal to
24411 // the element size. The constant shift amount will be
24412 // encoded as a 8-bit immediate.
24413 if (ShiftAmt.trunc(8).uge(MaxAmount))
24414 return getZeroVector(VT, Subtarget, DAG, DL);
24420 /// PerformShiftCombine - Combine shifts.
24421 static SDValue PerformShiftCombine(SDNode* N, SelectionDAG &DAG,
24422 TargetLowering::DAGCombinerInfo &DCI,
24423 const X86Subtarget *Subtarget) {
24424 if (N->getOpcode() == ISD::SHL)
24425 if (SDValue V = PerformSHLCombine(N, DAG))
24428 // Try to fold this logical shift into a zero vector.
24429 if (N->getOpcode() != ISD::SRA)
24430 if (SDValue V = performShiftToAllZeros(N, DAG, Subtarget))
24436 // CMPEQCombine - Recognize the distinctive (AND (setcc ...) (setcc ..))
24437 // where both setccs reference the same FP CMP, and rewrite for CMPEQSS
24438 // and friends. Likewise for OR -> CMPNEQSS.
24439 static SDValue CMPEQCombine(SDNode *N, SelectionDAG &DAG,
24440 TargetLowering::DAGCombinerInfo &DCI,
24441 const X86Subtarget *Subtarget) {
24444 // SSE1 supports CMP{eq|ne}SS, and SSE2 added CMP{eq|ne}SD, but
24445 // we're requiring SSE2 for both.
24446 if (Subtarget->hasSSE2() && isAndOrOfSetCCs(SDValue(N, 0U), opcode)) {
24447 SDValue N0 = N->getOperand(0);
24448 SDValue N1 = N->getOperand(1);
24449 SDValue CMP0 = N0->getOperand(1);
24450 SDValue CMP1 = N1->getOperand(1);
24453 // The SETCCs should both refer to the same CMP.
24454 if (CMP0.getOpcode() != X86ISD::CMP || CMP0 != CMP1)
24457 SDValue CMP00 = CMP0->getOperand(0);
24458 SDValue CMP01 = CMP0->getOperand(1);
24459 EVT VT = CMP00.getValueType();
24461 if (VT == MVT::f32 || VT == MVT::f64) {
24462 bool ExpectingFlags = false;
24463 // Check for any users that want flags:
24464 for (SDNode::use_iterator UI = N->use_begin(), UE = N->use_end();
24465 !ExpectingFlags && UI != UE; ++UI)
24466 switch (UI->getOpcode()) {
24471 ExpectingFlags = true;
24473 case ISD::CopyToReg:
24474 case ISD::SIGN_EXTEND:
24475 case ISD::ZERO_EXTEND:
24476 case ISD::ANY_EXTEND:
24480 if (!ExpectingFlags) {
24481 enum X86::CondCode cc0 = (enum X86::CondCode)N0.getConstantOperandVal(0);
24482 enum X86::CondCode cc1 = (enum X86::CondCode)N1.getConstantOperandVal(0);
24484 if (cc1 == X86::COND_E || cc1 == X86::COND_NE) {
24485 X86::CondCode tmp = cc0;
24490 if ((cc0 == X86::COND_E && cc1 == X86::COND_NP) ||
24491 (cc0 == X86::COND_NE && cc1 == X86::COND_P)) {
24492 // FIXME: need symbolic constants for these magic numbers.
24493 // See X86ATTInstPrinter.cpp:printSSECC().
24494 unsigned x86cc = (cc0 == X86::COND_E) ? 0 : 4;
24495 if (Subtarget->hasAVX512()) {
24496 SDValue FSetCC = DAG.getNode(X86ISD::FSETCC, DL, MVT::i1, CMP00,
24498 DAG.getConstant(x86cc, DL, MVT::i8));
24499 if (N->getValueType(0) != MVT::i1)
24500 return DAG.getNode(ISD::ZERO_EXTEND, DL, N->getValueType(0),
24504 SDValue OnesOrZeroesF = DAG.getNode(X86ISD::FSETCC, DL,
24505 CMP00.getValueType(), CMP00, CMP01,
24506 DAG.getConstant(x86cc, DL,
24509 bool is64BitFP = (CMP00.getValueType() == MVT::f64);
24510 MVT IntVT = is64BitFP ? MVT::i64 : MVT::i32;
24512 if (is64BitFP && !Subtarget->is64Bit()) {
24513 // On a 32-bit target, we cannot bitcast the 64-bit float to a
24514 // 64-bit integer, since that's not a legal type. Since
24515 // OnesOrZeroesF is all ones of all zeroes, we don't need all the
24516 // bits, but can do this little dance to extract the lowest 32 bits
24517 // and work with those going forward.
24518 SDValue Vector64 = DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, MVT::v2f64,
24520 SDValue Vector32 = DAG.getBitcast(MVT::v4f32, Vector64);
24521 OnesOrZeroesF = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::f32,
24522 Vector32, DAG.getIntPtrConstant(0, DL));
24526 SDValue OnesOrZeroesI = DAG.getBitcast(IntVT, OnesOrZeroesF);
24527 SDValue ANDed = DAG.getNode(ISD::AND, DL, IntVT, OnesOrZeroesI,
24528 DAG.getConstant(1, DL, IntVT));
24529 SDValue OneBitOfTruth = DAG.getNode(ISD::TRUNCATE, DL, MVT::i8,
24531 return OneBitOfTruth;
24539 /// CanFoldXORWithAllOnes - Test whether the XOR operand is a AllOnes vector
24540 /// so it can be folded inside ANDNP.
24541 static bool CanFoldXORWithAllOnes(const SDNode *N) {
24542 EVT VT = N->getValueType(0);
24544 // Match direct AllOnes for 128 and 256-bit vectors
24545 if (ISD::isBuildVectorAllOnes(N))
24548 // Look through a bit convert.
24549 if (N->getOpcode() == ISD::BITCAST)
24550 N = N->getOperand(0).getNode();
24552 // Sometimes the operand may come from a insert_subvector building a 256-bit
24554 if (VT.is256BitVector() &&
24555 N->getOpcode() == ISD::INSERT_SUBVECTOR) {
24556 SDValue V1 = N->getOperand(0);
24557 SDValue V2 = N->getOperand(1);
24559 if (V1.getOpcode() == ISD::INSERT_SUBVECTOR &&
24560 V1.getOperand(0).getOpcode() == ISD::UNDEF &&
24561 ISD::isBuildVectorAllOnes(V1.getOperand(1).getNode()) &&
24562 ISD::isBuildVectorAllOnes(V2.getNode()))
24569 // On AVX/AVX2 the type v8i1 is legalized to v8i16, which is an XMM sized
24570 // register. In most cases we actually compare or select YMM-sized registers
24571 // and mixing the two types creates horrible code. This method optimizes
24572 // some of the transition sequences.
24573 static SDValue WidenMaskArithmetic(SDNode *N, SelectionDAG &DAG,
24574 TargetLowering::DAGCombinerInfo &DCI,
24575 const X86Subtarget *Subtarget) {
24576 EVT VT = N->getValueType(0);
24577 if (!VT.is256BitVector())
24580 assert((N->getOpcode() == ISD::ANY_EXTEND ||
24581 N->getOpcode() == ISD::ZERO_EXTEND ||
24582 N->getOpcode() == ISD::SIGN_EXTEND) && "Invalid Node");
24584 SDValue Narrow = N->getOperand(0);
24585 EVT NarrowVT = Narrow->getValueType(0);
24586 if (!NarrowVT.is128BitVector())
24589 if (Narrow->getOpcode() != ISD::XOR &&
24590 Narrow->getOpcode() != ISD::AND &&
24591 Narrow->getOpcode() != ISD::OR)
24594 SDValue N0 = Narrow->getOperand(0);
24595 SDValue N1 = Narrow->getOperand(1);
24598 // The Left side has to be a trunc.
24599 if (N0.getOpcode() != ISD::TRUNCATE)
24602 // The type of the truncated inputs.
24603 EVT WideVT = N0->getOperand(0)->getValueType(0);
24607 // The right side has to be a 'trunc' or a constant vector.
24608 bool RHSTrunc = N1.getOpcode() == ISD::TRUNCATE;
24609 ConstantSDNode *RHSConstSplat = nullptr;
24610 if (auto *RHSBV = dyn_cast<BuildVectorSDNode>(N1))
24611 RHSConstSplat = RHSBV->getConstantSplatNode();
24612 if (!RHSTrunc && !RHSConstSplat)
24615 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
24617 if (!TLI.isOperationLegalOrPromote(Narrow->getOpcode(), WideVT))
24620 // Set N0 and N1 to hold the inputs to the new wide operation.
24621 N0 = N0->getOperand(0);
24622 if (RHSConstSplat) {
24623 N1 = DAG.getNode(ISD::ZERO_EXTEND, DL, WideVT.getScalarType(),
24624 SDValue(RHSConstSplat, 0));
24625 SmallVector<SDValue, 8> C(WideVT.getVectorNumElements(), N1);
24626 N1 = DAG.getNode(ISD::BUILD_VECTOR, DL, WideVT, C);
24627 } else if (RHSTrunc) {
24628 N1 = N1->getOperand(0);
24631 // Generate the wide operation.
24632 SDValue Op = DAG.getNode(Narrow->getOpcode(), DL, WideVT, N0, N1);
24633 unsigned Opcode = N->getOpcode();
24635 case ISD::ANY_EXTEND:
24637 case ISD::ZERO_EXTEND: {
24638 unsigned InBits = NarrowVT.getScalarType().getSizeInBits();
24639 APInt Mask = APInt::getAllOnesValue(InBits);
24640 Mask = Mask.zext(VT.getScalarType().getSizeInBits());
24641 return DAG.getNode(ISD::AND, DL, VT,
24642 Op, DAG.getConstant(Mask, DL, VT));
24644 case ISD::SIGN_EXTEND:
24645 return DAG.getNode(ISD::SIGN_EXTEND_INREG, DL, VT,
24646 Op, DAG.getValueType(NarrowVT));
24648 llvm_unreachable("Unexpected opcode");
24652 static SDValue VectorZextCombine(SDNode *N, SelectionDAG &DAG,
24653 TargetLowering::DAGCombinerInfo &DCI,
24654 const X86Subtarget *Subtarget) {
24655 SDValue N0 = N->getOperand(0);
24656 SDValue N1 = N->getOperand(1);
24659 // A vector zext_in_reg may be represented as a shuffle,
24660 // feeding into a bitcast (this represents anyext) feeding into
24661 // an and with a mask.
24662 // We'd like to try to combine that into a shuffle with zero
24663 // plus a bitcast, removing the and.
24664 if (N0.getOpcode() != ISD::BITCAST ||
24665 N0.getOperand(0).getOpcode() != ISD::VECTOR_SHUFFLE)
24668 // The other side of the AND should be a splat of 2^C, where C
24669 // is the number of bits in the source type.
24670 if (N1.getOpcode() == ISD::BITCAST)
24671 N1 = N1.getOperand(0);
24672 if (N1.getOpcode() != ISD::BUILD_VECTOR)
24674 BuildVectorSDNode *Vector = cast<BuildVectorSDNode>(N1);
24676 ShuffleVectorSDNode *Shuffle = cast<ShuffleVectorSDNode>(N0.getOperand(0));
24677 EVT SrcType = Shuffle->getValueType(0);
24679 // We expect a single-source shuffle
24680 if (Shuffle->getOperand(1)->getOpcode() != ISD::UNDEF)
24683 unsigned SrcSize = SrcType.getScalarSizeInBits();
24685 APInt SplatValue, SplatUndef;
24686 unsigned SplatBitSize;
24688 if (!Vector->isConstantSplat(SplatValue, SplatUndef,
24689 SplatBitSize, HasAnyUndefs))
24692 unsigned ResSize = N1.getValueType().getScalarSizeInBits();
24693 // Make sure the splat matches the mask we expect
24694 if (SplatBitSize > ResSize ||
24695 (SplatValue + 1).exactLogBase2() != (int)SrcSize)
24698 // Make sure the input and output size make sense
24699 if (SrcSize >= ResSize || ResSize % SrcSize)
24702 // We expect a shuffle of the form <0, u, u, u, 1, u, u, u...>
24703 // The number of u's between each two values depends on the ratio between
24704 // the source and dest type.
24705 unsigned ZextRatio = ResSize / SrcSize;
24706 bool IsZext = true;
24707 for (unsigned i = 0; i < SrcType.getVectorNumElements(); ++i) {
24708 if (i % ZextRatio) {
24709 if (Shuffle->getMaskElt(i) > 0) {
24715 if (Shuffle->getMaskElt(i) != (int)(i / ZextRatio)) {
24716 // Expected element number
24726 // Ok, perform the transformation - replace the shuffle with
24727 // a shuffle of the form <0, k, k, k, 1, k, k, k> with zero
24728 // (instead of undef) where the k elements come from the zero vector.
24729 SmallVector<int, 8> Mask;
24730 unsigned NumElems = SrcType.getVectorNumElements();
24731 for (unsigned i = 0; i < NumElems; ++i)
24733 Mask.push_back(NumElems);
24735 Mask.push_back(i / ZextRatio);
24737 SDValue NewShuffle = DAG.getVectorShuffle(Shuffle->getValueType(0), DL,
24738 Shuffle->getOperand(0), DAG.getConstant(0, DL, SrcType), Mask);
24739 return DAG.getBitcast(N0.getValueType(), NewShuffle);
24742 /// If both input operands of a logic op are being cast from floating point
24743 /// types, try to convert this into a floating point logic node to avoid
24744 /// unnecessary moves from SSE to integer registers.
24745 static SDValue convertIntLogicToFPLogic(SDNode *N, SelectionDAG &DAG,
24746 const X86Subtarget *Subtarget) {
24747 unsigned FPOpcode = ISD::DELETED_NODE;
24748 if (N->getOpcode() == ISD::AND)
24749 FPOpcode = X86ISD::FAND;
24750 else if (N->getOpcode() == ISD::OR)
24751 FPOpcode = X86ISD::FOR;
24752 else if (N->getOpcode() == ISD::XOR)
24753 FPOpcode = X86ISD::FXOR;
24755 assert(FPOpcode != ISD::DELETED_NODE &&
24756 "Unexpected input node for FP logic conversion");
24758 EVT VT = N->getValueType(0);
24759 SDValue N0 = N->getOperand(0);
24760 SDValue N1 = N->getOperand(1);
24762 if (N0.getOpcode() == ISD::BITCAST && N1.getOpcode() == ISD::BITCAST &&
24763 ((Subtarget->hasSSE1() && VT == MVT::i32) ||
24764 (Subtarget->hasSSE2() && VT == MVT::i64))) {
24765 SDValue N00 = N0.getOperand(0);
24766 SDValue N10 = N1.getOperand(0);
24767 EVT N00Type = N00.getValueType();
24768 EVT N10Type = N10.getValueType();
24769 if (N00Type.isFloatingPoint() && N10Type.isFloatingPoint()) {
24770 SDValue FPLogic = DAG.getNode(FPOpcode, DL, N00Type, N00, N10);
24771 return DAG.getBitcast(VT, FPLogic);
24777 static SDValue PerformAndCombine(SDNode *N, SelectionDAG &DAG,
24778 TargetLowering::DAGCombinerInfo &DCI,
24779 const X86Subtarget *Subtarget) {
24780 if (DCI.isBeforeLegalizeOps())
24783 if (SDValue Zext = VectorZextCombine(N, DAG, DCI, Subtarget))
24786 if (SDValue R = CMPEQCombine(N, DAG, DCI, Subtarget))
24789 if (SDValue FPLogic = convertIntLogicToFPLogic(N, DAG, Subtarget))
24792 EVT VT = N->getValueType(0);
24793 SDValue N0 = N->getOperand(0);
24794 SDValue N1 = N->getOperand(1);
24797 // Create BEXTR instructions
24798 // BEXTR is ((X >> imm) & (2**size-1))
24799 if (VT == MVT::i32 || VT == MVT::i64) {
24800 // Check for BEXTR.
24801 if ((Subtarget->hasBMI() || Subtarget->hasTBM()) &&
24802 (N0.getOpcode() == ISD::SRA || N0.getOpcode() == ISD::SRL)) {
24803 ConstantSDNode *MaskNode = dyn_cast<ConstantSDNode>(N1);
24804 ConstantSDNode *ShiftNode = dyn_cast<ConstantSDNode>(N0.getOperand(1));
24805 if (MaskNode && ShiftNode) {
24806 uint64_t Mask = MaskNode->getZExtValue();
24807 uint64_t Shift = ShiftNode->getZExtValue();
24808 if (isMask_64(Mask)) {
24809 uint64_t MaskSize = countPopulation(Mask);
24810 if (Shift + MaskSize <= VT.getSizeInBits())
24811 return DAG.getNode(X86ISD::BEXTR, DL, VT, N0.getOperand(0),
24812 DAG.getConstant(Shift | (MaskSize << 8), DL,
24821 // Want to form ANDNP nodes:
24822 // 1) In the hopes of then easily combining them with OR and AND nodes
24823 // to form PBLEND/PSIGN.
24824 // 2) To match ANDN packed intrinsics
24825 if (VT != MVT::v2i64 && VT != MVT::v4i64)
24828 // Check LHS for vnot
24829 if (N0.getOpcode() == ISD::XOR &&
24830 //ISD::isBuildVectorAllOnes(N0.getOperand(1).getNode()))
24831 CanFoldXORWithAllOnes(N0.getOperand(1).getNode()))
24832 return DAG.getNode(X86ISD::ANDNP, DL, VT, N0.getOperand(0), N1);
24834 // Check RHS for vnot
24835 if (N1.getOpcode() == ISD::XOR &&
24836 //ISD::isBuildVectorAllOnes(N1.getOperand(1).getNode()))
24837 CanFoldXORWithAllOnes(N1.getOperand(1).getNode()))
24838 return DAG.getNode(X86ISD::ANDNP, DL, VT, N1.getOperand(0), N0);
24843 static SDValue PerformOrCombine(SDNode *N, SelectionDAG &DAG,
24844 TargetLowering::DAGCombinerInfo &DCI,
24845 const X86Subtarget *Subtarget) {
24846 if (DCI.isBeforeLegalizeOps())
24849 if (SDValue R = CMPEQCombine(N, DAG, DCI, Subtarget))
24852 if (SDValue FPLogic = convertIntLogicToFPLogic(N, DAG, Subtarget))
24855 SDValue N0 = N->getOperand(0);
24856 SDValue N1 = N->getOperand(1);
24857 EVT VT = N->getValueType(0);
24859 // look for psign/blend
24860 if (VT == MVT::v2i64 || VT == MVT::v4i64) {
24861 if (!Subtarget->hasSSSE3() ||
24862 (VT == MVT::v4i64 && !Subtarget->hasInt256()))
24865 // Canonicalize pandn to RHS
24866 if (N0.getOpcode() == X86ISD::ANDNP)
24868 // or (and (m, y), (pandn m, x))
24869 if (N0.getOpcode() == ISD::AND && N1.getOpcode() == X86ISD::ANDNP) {
24870 SDValue Mask = N1.getOperand(0);
24871 SDValue X = N1.getOperand(1);
24873 if (N0.getOperand(0) == Mask)
24874 Y = N0.getOperand(1);
24875 if (N0.getOperand(1) == Mask)
24876 Y = N0.getOperand(0);
24878 // Check to see if the mask appeared in both the AND and ANDNP and
24882 // Validate that X, Y, and Mask are BIT_CONVERTS, and see through them.
24883 // Look through mask bitcast.
24884 if (Mask.getOpcode() == ISD::BITCAST)
24885 Mask = Mask.getOperand(0);
24886 if (X.getOpcode() == ISD::BITCAST)
24887 X = X.getOperand(0);
24888 if (Y.getOpcode() == ISD::BITCAST)
24889 Y = Y.getOperand(0);
24891 EVT MaskVT = Mask.getValueType();
24893 // Validate that the Mask operand is a vector sra node.
24894 // FIXME: what to do for bytes, since there is a psignb/pblendvb, but
24895 // there is no psrai.b
24896 unsigned EltBits = MaskVT.getVectorElementType().getSizeInBits();
24897 unsigned SraAmt = ~0;
24898 if (Mask.getOpcode() == ISD::SRA) {
24899 if (auto *AmtBV = dyn_cast<BuildVectorSDNode>(Mask.getOperand(1)))
24900 if (auto *AmtConst = AmtBV->getConstantSplatNode())
24901 SraAmt = AmtConst->getZExtValue();
24902 } else if (Mask.getOpcode() == X86ISD::VSRAI) {
24903 SDValue SraC = Mask.getOperand(1);
24904 SraAmt = cast<ConstantSDNode>(SraC)->getZExtValue();
24906 if ((SraAmt + 1) != EltBits)
24911 // Now we know we at least have a plendvb with the mask val. See if
24912 // we can form a psignb/w/d.
24913 // psign = x.type == y.type == mask.type && y = sub(0, x);
24914 if (Y.getOpcode() == ISD::SUB && Y.getOperand(1) == X &&
24915 ISD::isBuildVectorAllZeros(Y.getOperand(0).getNode()) &&
24916 X.getValueType() == MaskVT && Y.getValueType() == MaskVT) {
24917 assert((EltBits == 8 || EltBits == 16 || EltBits == 32) &&
24918 "Unsupported VT for PSIGN");
24919 Mask = DAG.getNode(X86ISD::PSIGN, DL, MaskVT, X, Mask.getOperand(0));
24920 return DAG.getBitcast(VT, Mask);
24922 // PBLENDVB only available on SSE 4.1
24923 if (!Subtarget->hasSSE41())
24926 EVT BlendVT = (VT == MVT::v4i64) ? MVT::v32i8 : MVT::v16i8;
24928 X = DAG.getBitcast(BlendVT, X);
24929 Y = DAG.getBitcast(BlendVT, Y);
24930 Mask = DAG.getBitcast(BlendVT, Mask);
24931 Mask = DAG.getNode(ISD::VSELECT, DL, BlendVT, Mask, Y, X);
24932 return DAG.getBitcast(VT, Mask);
24936 if (VT != MVT::i16 && VT != MVT::i32 && VT != MVT::i64)
24939 // fold (or (x << c) | (y >> (64 - c))) ==> (shld64 x, y, c)
24940 bool OptForSize = DAG.getMachineFunction().getFunction()->optForSize();
24942 // SHLD/SHRD instructions have lower register pressure, but on some
24943 // platforms they have higher latency than the equivalent
24944 // series of shifts/or that would otherwise be generated.
24945 // Don't fold (or (x << c) | (y >> (64 - c))) if SHLD/SHRD instructions
24946 // have higher latencies and we are not optimizing for size.
24947 if (!OptForSize && Subtarget->isSHLDSlow())
24950 if (N0.getOpcode() == ISD::SRL && N1.getOpcode() == ISD::SHL)
24952 if (N0.getOpcode() != ISD::SHL || N1.getOpcode() != ISD::SRL)
24954 if (!N0.hasOneUse() || !N1.hasOneUse())
24957 SDValue ShAmt0 = N0.getOperand(1);
24958 if (ShAmt0.getValueType() != MVT::i8)
24960 SDValue ShAmt1 = N1.getOperand(1);
24961 if (ShAmt1.getValueType() != MVT::i8)
24963 if (ShAmt0.getOpcode() == ISD::TRUNCATE)
24964 ShAmt0 = ShAmt0.getOperand(0);
24965 if (ShAmt1.getOpcode() == ISD::TRUNCATE)
24966 ShAmt1 = ShAmt1.getOperand(0);
24969 unsigned Opc = X86ISD::SHLD;
24970 SDValue Op0 = N0.getOperand(0);
24971 SDValue Op1 = N1.getOperand(0);
24972 if (ShAmt0.getOpcode() == ISD::SUB) {
24973 Opc = X86ISD::SHRD;
24974 std::swap(Op0, Op1);
24975 std::swap(ShAmt0, ShAmt1);
24978 unsigned Bits = VT.getSizeInBits();
24979 if (ShAmt1.getOpcode() == ISD::SUB) {
24980 SDValue Sum = ShAmt1.getOperand(0);
24981 if (ConstantSDNode *SumC = dyn_cast<ConstantSDNode>(Sum)) {
24982 SDValue ShAmt1Op1 = ShAmt1.getOperand(1);
24983 if (ShAmt1Op1.getNode()->getOpcode() == ISD::TRUNCATE)
24984 ShAmt1Op1 = ShAmt1Op1.getOperand(0);
24985 if (SumC->getSExtValue() == Bits && ShAmt1Op1 == ShAmt0)
24986 return DAG.getNode(Opc, DL, VT,
24988 DAG.getNode(ISD::TRUNCATE, DL,
24991 } else if (ConstantSDNode *ShAmt1C = dyn_cast<ConstantSDNode>(ShAmt1)) {
24992 ConstantSDNode *ShAmt0C = dyn_cast<ConstantSDNode>(ShAmt0);
24994 ShAmt0C->getSExtValue() + ShAmt1C->getSExtValue() == Bits)
24995 return DAG.getNode(Opc, DL, VT,
24996 N0.getOperand(0), N1.getOperand(0),
24997 DAG.getNode(ISD::TRUNCATE, DL,
25004 // Generate NEG and CMOV for integer abs.
25005 static SDValue performIntegerAbsCombine(SDNode *N, SelectionDAG &DAG) {
25006 EVT VT = N->getValueType(0);
25008 // Since X86 does not have CMOV for 8-bit integer, we don't convert
25009 // 8-bit integer abs to NEG and CMOV.
25010 if (VT.isInteger() && VT.getSizeInBits() == 8)
25013 SDValue N0 = N->getOperand(0);
25014 SDValue N1 = N->getOperand(1);
25017 // Check pattern of XOR(ADD(X,Y), Y) where Y is SRA(X, size(X)-1)
25018 // and change it to SUB and CMOV.
25019 if (VT.isInteger() && N->getOpcode() == ISD::XOR &&
25020 N0.getOpcode() == ISD::ADD &&
25021 N0.getOperand(1) == N1 &&
25022 N1.getOpcode() == ISD::SRA &&
25023 N1.getOperand(0) == N0.getOperand(0))
25024 if (ConstantSDNode *Y1C = dyn_cast<ConstantSDNode>(N1.getOperand(1)))
25025 if (Y1C->getAPIntValue() == VT.getSizeInBits()-1) {
25026 // Generate SUB & CMOV.
25027 SDValue Neg = DAG.getNode(X86ISD::SUB, DL, DAG.getVTList(VT, MVT::i32),
25028 DAG.getConstant(0, DL, VT), N0.getOperand(0));
25030 SDValue Ops[] = { N0.getOperand(0), Neg,
25031 DAG.getConstant(X86::COND_GE, DL, MVT::i8),
25032 SDValue(Neg.getNode(), 1) };
25033 return DAG.getNode(X86ISD::CMOV, DL, DAG.getVTList(VT, MVT::Glue), Ops);
25038 // Try to turn tests against the signbit in the form of:
25039 // XOR(TRUNCATE(SRL(X, size(X)-1)), 1)
25042 static SDValue foldXorTruncShiftIntoCmp(SDNode *N, SelectionDAG &DAG) {
25043 // This is only worth doing if the output type is i8.
25044 if (N->getValueType(0) != MVT::i8)
25047 SDValue N0 = N->getOperand(0);
25048 SDValue N1 = N->getOperand(1);
25050 // We should be performing an xor against a truncated shift.
25051 if (N0.getOpcode() != ISD::TRUNCATE || !N0.hasOneUse())
25054 // Make sure we are performing an xor against one.
25055 if (!isa<ConstantSDNode>(N1) || !cast<ConstantSDNode>(N1)->isOne())
25058 // SetCC on x86 zero extends so only act on this if it's a logical shift.
25059 SDValue Shift = N0.getOperand(0);
25060 if (Shift.getOpcode() != ISD::SRL || !Shift.hasOneUse())
25063 // Make sure we are truncating from one of i16, i32 or i64.
25064 EVT ShiftTy = Shift.getValueType();
25065 if (ShiftTy != MVT::i16 && ShiftTy != MVT::i32 && ShiftTy != MVT::i64)
25068 // Make sure the shift amount extracts the sign bit.
25069 if (!isa<ConstantSDNode>(Shift.getOperand(1)) ||
25070 Shift.getConstantOperandVal(1) != ShiftTy.getSizeInBits() - 1)
25073 // Create a greater-than comparison against -1.
25074 // N.B. Using SETGE against 0 works but we want a canonical looking
25075 // comparison, using SETGT matches up with what TranslateX86CC.
25077 SDValue ShiftOp = Shift.getOperand(0);
25078 EVT ShiftOpTy = ShiftOp.getValueType();
25079 SDValue Cond = DAG.getSetCC(DL, MVT::i8, ShiftOp,
25080 DAG.getConstant(-1, DL, ShiftOpTy), ISD::SETGT);
25084 static SDValue PerformXorCombine(SDNode *N, SelectionDAG &DAG,
25085 TargetLowering::DAGCombinerInfo &DCI,
25086 const X86Subtarget *Subtarget) {
25087 if (DCI.isBeforeLegalizeOps())
25090 if (SDValue RV = foldXorTruncShiftIntoCmp(N, DAG))
25093 if (Subtarget->hasCMov())
25094 if (SDValue RV = performIntegerAbsCombine(N, DAG))
25097 if (SDValue FPLogic = convertIntLogicToFPLogic(N, DAG, Subtarget))
25103 /// PerformLOADCombine - Do target-specific dag combines on LOAD nodes.
25104 static SDValue PerformLOADCombine(SDNode *N, SelectionDAG &DAG,
25105 TargetLowering::DAGCombinerInfo &DCI,
25106 const X86Subtarget *Subtarget) {
25107 LoadSDNode *Ld = cast<LoadSDNode>(N);
25108 EVT RegVT = Ld->getValueType(0);
25109 EVT MemVT = Ld->getMemoryVT();
25111 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
25113 // For chips with slow 32-byte unaligned loads, break the 32-byte operation
25114 // into two 16-byte operations.
25115 ISD::LoadExtType Ext = Ld->getExtensionType();
25117 unsigned AddressSpace = Ld->getAddressSpace();
25118 unsigned Alignment = Ld->getAlignment();
25119 if (RegVT.is256BitVector() && !DCI.isBeforeLegalizeOps() &&
25120 Ext == ISD::NON_EXTLOAD &&
25121 TLI.allowsMemoryAccess(*DAG.getContext(), DAG.getDataLayout(), RegVT,
25122 AddressSpace, Alignment, &Fast) && !Fast) {
25123 unsigned NumElems = RegVT.getVectorNumElements();
25127 SDValue Ptr = Ld->getBasePtr();
25128 SDValue Increment =
25129 DAG.getConstant(16, dl, TLI.getPointerTy(DAG.getDataLayout()));
25131 EVT HalfVT = EVT::getVectorVT(*DAG.getContext(), MemVT.getScalarType(),
25133 SDValue Load1 = DAG.getLoad(HalfVT, dl, Ld->getChain(), Ptr,
25134 Ld->getPointerInfo(), Ld->isVolatile(),
25135 Ld->isNonTemporal(), Ld->isInvariant(),
25137 Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr, Increment);
25138 SDValue Load2 = DAG.getLoad(HalfVT, dl, Ld->getChain(), Ptr,
25139 Ld->getPointerInfo(), Ld->isVolatile(),
25140 Ld->isNonTemporal(), Ld->isInvariant(),
25141 std::min(16U, Alignment));
25142 SDValue TF = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
25144 Load2.getValue(1));
25146 SDValue NewVec = DAG.getUNDEF(RegVT);
25147 NewVec = Insert128BitVector(NewVec, Load1, 0, DAG, dl);
25148 NewVec = Insert128BitVector(NewVec, Load2, NumElems/2, DAG, dl);
25149 return DCI.CombineTo(N, NewVec, TF, true);
25155 /// PerformMLOADCombine - Resolve extending loads
25156 static SDValue PerformMLOADCombine(SDNode *N, SelectionDAG &DAG,
25157 TargetLowering::DAGCombinerInfo &DCI,
25158 const X86Subtarget *Subtarget) {
25159 MaskedLoadSDNode *Mld = cast<MaskedLoadSDNode>(N);
25160 if (Mld->getExtensionType() != ISD::SEXTLOAD)
25163 EVT VT = Mld->getValueType(0);
25164 unsigned NumElems = VT.getVectorNumElements();
25165 EVT LdVT = Mld->getMemoryVT();
25168 assert(LdVT != VT && "Cannot extend to the same type");
25169 unsigned ToSz = VT.getVectorElementType().getSizeInBits();
25170 unsigned FromSz = LdVT.getVectorElementType().getSizeInBits();
25171 // From, To sizes and ElemCount must be pow of two
25172 assert (isPowerOf2_32(NumElems * FromSz * ToSz) &&
25173 "Unexpected size for extending masked load");
25175 unsigned SizeRatio = ToSz / FromSz;
25176 assert(SizeRatio * NumElems * FromSz == VT.getSizeInBits());
25178 // Create a type on which we perform the shuffle
25179 EVT WideVecVT = EVT::getVectorVT(*DAG.getContext(),
25180 LdVT.getScalarType(), NumElems*SizeRatio);
25181 assert(WideVecVT.getSizeInBits() == VT.getSizeInBits());
25183 // Convert Src0 value
25184 SDValue WideSrc0 = DAG.getBitcast(WideVecVT, Mld->getSrc0());
25185 if (Mld->getSrc0().getOpcode() != ISD::UNDEF) {
25186 SmallVector<int, 16> ShuffleVec(NumElems * SizeRatio, -1);
25187 for (unsigned i = 0; i != NumElems; ++i)
25188 ShuffleVec[i] = i * SizeRatio;
25190 // Can't shuffle using an illegal type.
25191 assert(DAG.getTargetLoweringInfo().isTypeLegal(WideVecVT) &&
25192 "WideVecVT should be legal");
25193 WideSrc0 = DAG.getVectorShuffle(WideVecVT, dl, WideSrc0,
25194 DAG.getUNDEF(WideVecVT), &ShuffleVec[0]);
25196 // Prepare the new mask
25198 SDValue Mask = Mld->getMask();
25199 if (Mask.getValueType() == VT) {
25200 // Mask and original value have the same type
25201 NewMask = DAG.getBitcast(WideVecVT, Mask);
25202 SmallVector<int, 16> ShuffleVec(NumElems * SizeRatio, -1);
25203 for (unsigned i = 0; i != NumElems; ++i)
25204 ShuffleVec[i] = i * SizeRatio;
25205 for (unsigned i = NumElems; i != NumElems*SizeRatio; ++i)
25206 ShuffleVec[i] = NumElems*SizeRatio;
25207 NewMask = DAG.getVectorShuffle(WideVecVT, dl, NewMask,
25208 DAG.getConstant(0, dl, WideVecVT),
25212 assert(Mask.getValueType().getVectorElementType() == MVT::i1);
25213 unsigned WidenNumElts = NumElems*SizeRatio;
25214 unsigned MaskNumElts = VT.getVectorNumElements();
25215 EVT NewMaskVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
25218 unsigned NumConcat = WidenNumElts / MaskNumElts;
25219 SmallVector<SDValue, 16> Ops(NumConcat);
25220 SDValue ZeroVal = DAG.getConstant(0, dl, Mask.getValueType());
25222 for (unsigned i = 1; i != NumConcat; ++i)
25225 NewMask = DAG.getNode(ISD::CONCAT_VECTORS, dl, NewMaskVT, Ops);
25228 SDValue WideLd = DAG.getMaskedLoad(WideVecVT, dl, Mld->getChain(),
25229 Mld->getBasePtr(), NewMask, WideSrc0,
25230 Mld->getMemoryVT(), Mld->getMemOperand(),
25232 SDValue NewVec = DAG.getNode(X86ISD::VSEXT, dl, VT, WideLd);
25233 return DCI.CombineTo(N, NewVec, WideLd.getValue(1), true);
25235 /// PerformMSTORECombine - Resolve truncating stores
25236 static SDValue PerformMSTORECombine(SDNode *N, SelectionDAG &DAG,
25237 const X86Subtarget *Subtarget) {
25238 MaskedStoreSDNode *Mst = cast<MaskedStoreSDNode>(N);
25239 if (!Mst->isTruncatingStore())
25242 EVT VT = Mst->getValue().getValueType();
25243 unsigned NumElems = VT.getVectorNumElements();
25244 EVT StVT = Mst->getMemoryVT();
25247 assert(StVT != VT && "Cannot truncate to the same type");
25248 unsigned FromSz = VT.getVectorElementType().getSizeInBits();
25249 unsigned ToSz = StVT.getVectorElementType().getSizeInBits();
25251 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
25253 // The truncating store is legal in some cases. For example
25254 // vpmovqb, vpmovqw, vpmovqd, vpmovdb, vpmovdw
25255 // are designated for truncate store.
25256 // In this case we don't need any further transformations.
25257 if (TLI.isTruncStoreLegal(VT, StVT))
25260 // From, To sizes and ElemCount must be pow of two
25261 assert (isPowerOf2_32(NumElems * FromSz * ToSz) &&
25262 "Unexpected size for truncating masked store");
25263 // We are going to use the original vector elt for storing.
25264 // Accumulated smaller vector elements must be a multiple of the store size.
25265 assert (((NumElems * FromSz) % ToSz) == 0 &&
25266 "Unexpected ratio for truncating masked store");
25268 unsigned SizeRatio = FromSz / ToSz;
25269 assert(SizeRatio * NumElems * ToSz == VT.getSizeInBits());
25271 // Create a type on which we perform the shuffle
25272 EVT WideVecVT = EVT::getVectorVT(*DAG.getContext(),
25273 StVT.getScalarType(), NumElems*SizeRatio);
25275 assert(WideVecVT.getSizeInBits() == VT.getSizeInBits());
25277 SDValue WideVec = DAG.getBitcast(WideVecVT, Mst->getValue());
25278 SmallVector<int, 16> ShuffleVec(NumElems * SizeRatio, -1);
25279 for (unsigned i = 0; i != NumElems; ++i)
25280 ShuffleVec[i] = i * SizeRatio;
25282 // Can't shuffle using an illegal type.
25283 assert(DAG.getTargetLoweringInfo().isTypeLegal(WideVecVT) &&
25284 "WideVecVT should be legal");
25286 SDValue TruncatedVal = DAG.getVectorShuffle(WideVecVT, dl, WideVec,
25287 DAG.getUNDEF(WideVecVT),
25291 SDValue Mask = Mst->getMask();
25292 if (Mask.getValueType() == VT) {
25293 // Mask and original value have the same type
25294 NewMask = DAG.getBitcast(WideVecVT, Mask);
25295 for (unsigned i = 0; i != NumElems; ++i)
25296 ShuffleVec[i] = i * SizeRatio;
25297 for (unsigned i = NumElems; i != NumElems*SizeRatio; ++i)
25298 ShuffleVec[i] = NumElems*SizeRatio;
25299 NewMask = DAG.getVectorShuffle(WideVecVT, dl, NewMask,
25300 DAG.getConstant(0, dl, WideVecVT),
25304 assert(Mask.getValueType().getVectorElementType() == MVT::i1);
25305 unsigned WidenNumElts = NumElems*SizeRatio;
25306 unsigned MaskNumElts = VT.getVectorNumElements();
25307 EVT NewMaskVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
25310 unsigned NumConcat = WidenNumElts / MaskNumElts;
25311 SmallVector<SDValue, 16> Ops(NumConcat);
25312 SDValue ZeroVal = DAG.getConstant(0, dl, Mask.getValueType());
25314 for (unsigned i = 1; i != NumConcat; ++i)
25317 NewMask = DAG.getNode(ISD::CONCAT_VECTORS, dl, NewMaskVT, Ops);
25320 return DAG.getMaskedStore(Mst->getChain(), dl, TruncatedVal, Mst->getBasePtr(),
25321 NewMask, StVT, Mst->getMemOperand(), false);
25323 /// PerformSTORECombine - Do target-specific dag combines on STORE nodes.
25324 static SDValue PerformSTORECombine(SDNode *N, SelectionDAG &DAG,
25325 const X86Subtarget *Subtarget) {
25326 StoreSDNode *St = cast<StoreSDNode>(N);
25327 EVT VT = St->getValue().getValueType();
25328 EVT StVT = St->getMemoryVT();
25330 SDValue StoredVal = St->getOperand(1);
25331 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
25333 // If we are saving a concatenation of two XMM registers and 32-byte stores
25334 // are slow, such as on Sandy Bridge, perform two 16-byte stores.
25336 unsigned AddressSpace = St->getAddressSpace();
25337 unsigned Alignment = St->getAlignment();
25338 if (VT.is256BitVector() && StVT == VT &&
25339 TLI.allowsMemoryAccess(*DAG.getContext(), DAG.getDataLayout(), VT,
25340 AddressSpace, Alignment, &Fast) && !Fast) {
25341 unsigned NumElems = VT.getVectorNumElements();
25345 SDValue Value0 = Extract128BitVector(StoredVal, 0, DAG, dl);
25346 SDValue Value1 = Extract128BitVector(StoredVal, NumElems/2, DAG, dl);
25349 DAG.getConstant(16, dl, TLI.getPointerTy(DAG.getDataLayout()));
25350 SDValue Ptr0 = St->getBasePtr();
25351 SDValue Ptr1 = DAG.getNode(ISD::ADD, dl, Ptr0.getValueType(), Ptr0, Stride);
25353 SDValue Ch0 = DAG.getStore(St->getChain(), dl, Value0, Ptr0,
25354 St->getPointerInfo(), St->isVolatile(),
25355 St->isNonTemporal(), Alignment);
25356 SDValue Ch1 = DAG.getStore(St->getChain(), dl, Value1, Ptr1,
25357 St->getPointerInfo(), St->isVolatile(),
25358 St->isNonTemporal(),
25359 std::min(16U, Alignment));
25360 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Ch0, Ch1);
25363 // Optimize trunc store (of multiple scalars) to shuffle and store.
25364 // First, pack all of the elements in one place. Next, store to memory
25365 // in fewer chunks.
25366 if (St->isTruncatingStore() && VT.isVector()) {
25367 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
25368 unsigned NumElems = VT.getVectorNumElements();
25369 assert(StVT != VT && "Cannot truncate to the same type");
25370 unsigned FromSz = VT.getVectorElementType().getSizeInBits();
25371 unsigned ToSz = StVT.getVectorElementType().getSizeInBits();
25373 // The truncating store is legal in some cases. For example
25374 // vpmovqb, vpmovqw, vpmovqd, vpmovdb, vpmovdw
25375 // are designated for truncate store.
25376 // In this case we don't need any further transformations.
25377 if (TLI.isTruncStoreLegal(VT, StVT))
25380 // From, To sizes and ElemCount must be pow of two
25381 if (!isPowerOf2_32(NumElems * FromSz * ToSz)) return SDValue();
25382 // We are going to use the original vector elt for storing.
25383 // Accumulated smaller vector elements must be a multiple of the store size.
25384 if (0 != (NumElems * FromSz) % ToSz) return SDValue();
25386 unsigned SizeRatio = FromSz / ToSz;
25388 assert(SizeRatio * NumElems * ToSz == VT.getSizeInBits());
25390 // Create a type on which we perform the shuffle
25391 EVT WideVecVT = EVT::getVectorVT(*DAG.getContext(),
25392 StVT.getScalarType(), NumElems*SizeRatio);
25394 assert(WideVecVT.getSizeInBits() == VT.getSizeInBits());
25396 SDValue WideVec = DAG.getBitcast(WideVecVT, St->getValue());
25397 SmallVector<int, 8> ShuffleVec(NumElems * SizeRatio, -1);
25398 for (unsigned i = 0; i != NumElems; ++i)
25399 ShuffleVec[i] = i * SizeRatio;
25401 // Can't shuffle using an illegal type.
25402 if (!TLI.isTypeLegal(WideVecVT))
25405 SDValue Shuff = DAG.getVectorShuffle(WideVecVT, dl, WideVec,
25406 DAG.getUNDEF(WideVecVT),
25408 // At this point all of the data is stored at the bottom of the
25409 // register. We now need to save it to mem.
25411 // Find the largest store unit
25412 MVT StoreType = MVT::i8;
25413 for (MVT Tp : MVT::integer_valuetypes()) {
25414 if (TLI.isTypeLegal(Tp) && Tp.getSizeInBits() <= NumElems * ToSz)
25418 // On 32bit systems, we can't save 64bit integers. Try bitcasting to F64.
25419 if (TLI.isTypeLegal(MVT::f64) && StoreType.getSizeInBits() < 64 &&
25420 (64 <= NumElems * ToSz))
25421 StoreType = MVT::f64;
25423 // Bitcast the original vector into a vector of store-size units
25424 EVT StoreVecVT = EVT::getVectorVT(*DAG.getContext(),
25425 StoreType, VT.getSizeInBits()/StoreType.getSizeInBits());
25426 assert(StoreVecVT.getSizeInBits() == VT.getSizeInBits());
25427 SDValue ShuffWide = DAG.getBitcast(StoreVecVT, Shuff);
25428 SmallVector<SDValue, 8> Chains;
25429 SDValue Increment = DAG.getConstant(StoreType.getSizeInBits() / 8, dl,
25430 TLI.getPointerTy(DAG.getDataLayout()));
25431 SDValue Ptr = St->getBasePtr();
25433 // Perform one or more big stores into memory.
25434 for (unsigned i=0, e=(ToSz*NumElems)/StoreType.getSizeInBits(); i!=e; ++i) {
25435 SDValue SubVec = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl,
25436 StoreType, ShuffWide,
25437 DAG.getIntPtrConstant(i, dl));
25438 SDValue Ch = DAG.getStore(St->getChain(), dl, SubVec, Ptr,
25439 St->getPointerInfo(), St->isVolatile(),
25440 St->isNonTemporal(), St->getAlignment());
25441 Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr, Increment);
25442 Chains.push_back(Ch);
25445 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Chains);
25448 // Turn load->store of MMX types into GPR load/stores. This avoids clobbering
25449 // the FP state in cases where an emms may be missing.
25450 // A preferable solution to the general problem is to figure out the right
25451 // places to insert EMMS. This qualifies as a quick hack.
25453 // Similarly, turn load->store of i64 into double load/stores in 32-bit mode.
25454 if (VT.getSizeInBits() != 64)
25457 const Function *F = DAG.getMachineFunction().getFunction();
25458 bool NoImplicitFloatOps = F->hasFnAttribute(Attribute::NoImplicitFloat);
25460 !Subtarget->useSoftFloat() && !NoImplicitFloatOps && Subtarget->hasSSE2();
25461 if ((VT.isVector() ||
25462 (VT == MVT::i64 && F64IsLegal && !Subtarget->is64Bit())) &&
25463 isa<LoadSDNode>(St->getValue()) &&
25464 !cast<LoadSDNode>(St->getValue())->isVolatile() &&
25465 St->getChain().hasOneUse() && !St->isVolatile()) {
25466 SDNode* LdVal = St->getValue().getNode();
25467 LoadSDNode *Ld = nullptr;
25468 int TokenFactorIndex = -1;
25469 SmallVector<SDValue, 8> Ops;
25470 SDNode* ChainVal = St->getChain().getNode();
25471 // Must be a store of a load. We currently handle two cases: the load
25472 // is a direct child, and it's under an intervening TokenFactor. It is
25473 // possible to dig deeper under nested TokenFactors.
25474 if (ChainVal == LdVal)
25475 Ld = cast<LoadSDNode>(St->getChain());
25476 else if (St->getValue().hasOneUse() &&
25477 ChainVal->getOpcode() == ISD::TokenFactor) {
25478 for (unsigned i = 0, e = ChainVal->getNumOperands(); i != e; ++i) {
25479 if (ChainVal->getOperand(i).getNode() == LdVal) {
25480 TokenFactorIndex = i;
25481 Ld = cast<LoadSDNode>(St->getValue());
25483 Ops.push_back(ChainVal->getOperand(i));
25487 if (!Ld || !ISD::isNormalLoad(Ld))
25490 // If this is not the MMX case, i.e. we are just turning i64 load/store
25491 // into f64 load/store, avoid the transformation if there are multiple
25492 // uses of the loaded value.
25493 if (!VT.isVector() && !Ld->hasNUsesOfValue(1, 0))
25498 // If we are a 64-bit capable x86, lower to a single movq load/store pair.
25499 // Otherwise, if it's legal to use f64 SSE instructions, use f64 load/store
25501 if (Subtarget->is64Bit() || F64IsLegal) {
25502 EVT LdVT = Subtarget->is64Bit() ? MVT::i64 : MVT::f64;
25503 SDValue NewLd = DAG.getLoad(LdVT, LdDL, Ld->getChain(), Ld->getBasePtr(),
25504 Ld->getPointerInfo(), Ld->isVolatile(),
25505 Ld->isNonTemporal(), Ld->isInvariant(),
25506 Ld->getAlignment());
25507 SDValue NewChain = NewLd.getValue(1);
25508 if (TokenFactorIndex != -1) {
25509 Ops.push_back(NewChain);
25510 NewChain = DAG.getNode(ISD::TokenFactor, LdDL, MVT::Other, Ops);
25512 return DAG.getStore(NewChain, StDL, NewLd, St->getBasePtr(),
25513 St->getPointerInfo(),
25514 St->isVolatile(), St->isNonTemporal(),
25515 St->getAlignment());
25518 // Otherwise, lower to two pairs of 32-bit loads / stores.
25519 SDValue LoAddr = Ld->getBasePtr();
25520 SDValue HiAddr = DAG.getNode(ISD::ADD, LdDL, MVT::i32, LoAddr,
25521 DAG.getConstant(4, LdDL, MVT::i32));
25523 SDValue LoLd = DAG.getLoad(MVT::i32, LdDL, Ld->getChain(), LoAddr,
25524 Ld->getPointerInfo(),
25525 Ld->isVolatile(), Ld->isNonTemporal(),
25526 Ld->isInvariant(), Ld->getAlignment());
25527 SDValue HiLd = DAG.getLoad(MVT::i32, LdDL, Ld->getChain(), HiAddr,
25528 Ld->getPointerInfo().getWithOffset(4),
25529 Ld->isVolatile(), Ld->isNonTemporal(),
25531 MinAlign(Ld->getAlignment(), 4));
25533 SDValue NewChain = LoLd.getValue(1);
25534 if (TokenFactorIndex != -1) {
25535 Ops.push_back(LoLd);
25536 Ops.push_back(HiLd);
25537 NewChain = DAG.getNode(ISD::TokenFactor, LdDL, MVT::Other, Ops);
25540 LoAddr = St->getBasePtr();
25541 HiAddr = DAG.getNode(ISD::ADD, StDL, MVT::i32, LoAddr,
25542 DAG.getConstant(4, StDL, MVT::i32));
25544 SDValue LoSt = DAG.getStore(NewChain, StDL, LoLd, LoAddr,
25545 St->getPointerInfo(),
25546 St->isVolatile(), St->isNonTemporal(),
25547 St->getAlignment());
25548 SDValue HiSt = DAG.getStore(NewChain, StDL, HiLd, HiAddr,
25549 St->getPointerInfo().getWithOffset(4),
25551 St->isNonTemporal(),
25552 MinAlign(St->getAlignment(), 4));
25553 return DAG.getNode(ISD::TokenFactor, StDL, MVT::Other, LoSt, HiSt);
25556 // This is similar to the above case, but here we handle a scalar 64-bit
25557 // integer store that is extracted from a vector on a 32-bit target.
25558 // If we have SSE2, then we can treat it like a floating-point double
25559 // to get past legalization. The execution dependencies fixup pass will
25560 // choose the optimal machine instruction for the store if this really is
25561 // an integer or v2f32 rather than an f64.
25562 if (VT == MVT::i64 && F64IsLegal && !Subtarget->is64Bit() &&
25563 St->getOperand(1).getOpcode() == ISD::EXTRACT_VECTOR_ELT) {
25564 SDValue OldExtract = St->getOperand(1);
25565 SDValue ExtOp0 = OldExtract.getOperand(0);
25566 unsigned VecSize = ExtOp0.getValueSizeInBits();
25567 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), MVT::f64, VecSize / 64);
25568 SDValue BitCast = DAG.getBitcast(VecVT, ExtOp0);
25569 SDValue NewExtract = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
25570 BitCast, OldExtract.getOperand(1));
25571 return DAG.getStore(St->getChain(), dl, NewExtract, St->getBasePtr(),
25572 St->getPointerInfo(), St->isVolatile(),
25573 St->isNonTemporal(), St->getAlignment());
25579 /// Return 'true' if this vector operation is "horizontal"
25580 /// and return the operands for the horizontal operation in LHS and RHS. A
25581 /// horizontal operation performs the binary operation on successive elements
25582 /// of its first operand, then on successive elements of its second operand,
25583 /// returning the resulting values in a vector. For example, if
25584 /// A = < float a0, float a1, float a2, float a3 >
25586 /// B = < float b0, float b1, float b2, float b3 >
25587 /// then the result of doing a horizontal operation on A and B is
25588 /// A horizontal-op B = < a0 op a1, a2 op a3, b0 op b1, b2 op b3 >.
25589 /// In short, LHS and RHS are inspected to see if LHS op RHS is of the form
25590 /// A horizontal-op B, for some already available A and B, and if so then LHS is
25591 /// set to A, RHS to B, and the routine returns 'true'.
25592 /// Note that the binary operation should have the property that if one of the
25593 /// operands is UNDEF then the result is UNDEF.
25594 static bool isHorizontalBinOp(SDValue &LHS, SDValue &RHS, bool IsCommutative) {
25595 // Look for the following pattern: if
25596 // A = < float a0, float a1, float a2, float a3 >
25597 // B = < float b0, float b1, float b2, float b3 >
25599 // LHS = VECTOR_SHUFFLE A, B, <0, 2, 4, 6>
25600 // RHS = VECTOR_SHUFFLE A, B, <1, 3, 5, 7>
25601 // then LHS op RHS = < a0 op a1, a2 op a3, b0 op b1, b2 op b3 >
25602 // which is A horizontal-op B.
25604 // At least one of the operands should be a vector shuffle.
25605 if (LHS.getOpcode() != ISD::VECTOR_SHUFFLE &&
25606 RHS.getOpcode() != ISD::VECTOR_SHUFFLE)
25609 MVT VT = LHS.getSimpleValueType();
25611 assert((VT.is128BitVector() || VT.is256BitVector()) &&
25612 "Unsupported vector type for horizontal add/sub");
25614 // Handle 128 and 256-bit vector lengths. AVX defines horizontal add/sub to
25615 // operate independently on 128-bit lanes.
25616 unsigned NumElts = VT.getVectorNumElements();
25617 unsigned NumLanes = VT.getSizeInBits()/128;
25618 unsigned NumLaneElts = NumElts / NumLanes;
25619 assert((NumLaneElts % 2 == 0) &&
25620 "Vector type should have an even number of elements in each lane");
25621 unsigned HalfLaneElts = NumLaneElts/2;
25623 // View LHS in the form
25624 // LHS = VECTOR_SHUFFLE A, B, LMask
25625 // If LHS is not a shuffle then pretend it is the shuffle
25626 // LHS = VECTOR_SHUFFLE LHS, undef, <0, 1, ..., N-1>
25627 // NOTE: in what follows a default initialized SDValue represents an UNDEF of
25630 SmallVector<int, 16> LMask(NumElts);
25631 if (LHS.getOpcode() == ISD::VECTOR_SHUFFLE) {
25632 if (LHS.getOperand(0).getOpcode() != ISD::UNDEF)
25633 A = LHS.getOperand(0);
25634 if (LHS.getOperand(1).getOpcode() != ISD::UNDEF)
25635 B = LHS.getOperand(1);
25636 ArrayRef<int> Mask = cast<ShuffleVectorSDNode>(LHS.getNode())->getMask();
25637 std::copy(Mask.begin(), Mask.end(), LMask.begin());
25639 if (LHS.getOpcode() != ISD::UNDEF)
25641 for (unsigned i = 0; i != NumElts; ++i)
25645 // Likewise, view RHS in the form
25646 // RHS = VECTOR_SHUFFLE C, D, RMask
25648 SmallVector<int, 16> RMask(NumElts);
25649 if (RHS.getOpcode() == ISD::VECTOR_SHUFFLE) {
25650 if (RHS.getOperand(0).getOpcode() != ISD::UNDEF)
25651 C = RHS.getOperand(0);
25652 if (RHS.getOperand(1).getOpcode() != ISD::UNDEF)
25653 D = RHS.getOperand(1);
25654 ArrayRef<int> Mask = cast<ShuffleVectorSDNode>(RHS.getNode())->getMask();
25655 std::copy(Mask.begin(), Mask.end(), RMask.begin());
25657 if (RHS.getOpcode() != ISD::UNDEF)
25659 for (unsigned i = 0; i != NumElts; ++i)
25663 // Check that the shuffles are both shuffling the same vectors.
25664 if (!(A == C && B == D) && !(A == D && B == C))
25667 // If everything is UNDEF then bail out: it would be better to fold to UNDEF.
25668 if (!A.getNode() && !B.getNode())
25671 // If A and B occur in reverse order in RHS, then "swap" them (which means
25672 // rewriting the mask).
25674 ShuffleVectorSDNode::commuteMask(RMask);
25676 // At this point LHS and RHS are equivalent to
25677 // LHS = VECTOR_SHUFFLE A, B, LMask
25678 // RHS = VECTOR_SHUFFLE A, B, RMask
25679 // Check that the masks correspond to performing a horizontal operation.
25680 for (unsigned l = 0; l != NumElts; l += NumLaneElts) {
25681 for (unsigned i = 0; i != NumLaneElts; ++i) {
25682 int LIdx = LMask[i+l], RIdx = RMask[i+l];
25684 // Ignore any UNDEF components.
25685 if (LIdx < 0 || RIdx < 0 ||
25686 (!A.getNode() && (LIdx < (int)NumElts || RIdx < (int)NumElts)) ||
25687 (!B.getNode() && (LIdx >= (int)NumElts || RIdx >= (int)NumElts)))
25690 // Check that successive elements are being operated on. If not, this is
25691 // not a horizontal operation.
25692 unsigned Src = (i/HalfLaneElts); // each lane is split between srcs
25693 int Index = 2*(i%HalfLaneElts) + NumElts*Src + l;
25694 if (!(LIdx == Index && RIdx == Index + 1) &&
25695 !(IsCommutative && LIdx == Index + 1 && RIdx == Index))
25700 LHS = A.getNode() ? A : B; // If A is 'UNDEF', use B for it.
25701 RHS = B.getNode() ? B : A; // If B is 'UNDEF', use A for it.
25705 /// Do target-specific dag combines on floating point adds.
25706 static SDValue PerformFADDCombine(SDNode *N, SelectionDAG &DAG,
25707 const X86Subtarget *Subtarget) {
25708 EVT VT = N->getValueType(0);
25709 SDValue LHS = N->getOperand(0);
25710 SDValue RHS = N->getOperand(1);
25712 // Try to synthesize horizontal adds from adds of shuffles.
25713 if (((Subtarget->hasSSE3() && (VT == MVT::v4f32 || VT == MVT::v2f64)) ||
25714 (Subtarget->hasFp256() && (VT == MVT::v8f32 || VT == MVT::v4f64))) &&
25715 isHorizontalBinOp(LHS, RHS, true))
25716 return DAG.getNode(X86ISD::FHADD, SDLoc(N), VT, LHS, RHS);
25720 /// Do target-specific dag combines on floating point subs.
25721 static SDValue PerformFSUBCombine(SDNode *N, SelectionDAG &DAG,
25722 const X86Subtarget *Subtarget) {
25723 EVT VT = N->getValueType(0);
25724 SDValue LHS = N->getOperand(0);
25725 SDValue RHS = N->getOperand(1);
25727 // Try to synthesize horizontal subs from subs of shuffles.
25728 if (((Subtarget->hasSSE3() && (VT == MVT::v4f32 || VT == MVT::v2f64)) ||
25729 (Subtarget->hasFp256() && (VT == MVT::v8f32 || VT == MVT::v4f64))) &&
25730 isHorizontalBinOp(LHS, RHS, false))
25731 return DAG.getNode(X86ISD::FHSUB, SDLoc(N), VT, LHS, RHS);
25735 /// Do target-specific dag combines on X86ISD::FOR and X86ISD::FXOR nodes.
25736 static SDValue PerformFORCombine(SDNode *N, SelectionDAG &DAG,
25737 const X86Subtarget *Subtarget) {
25738 assert(N->getOpcode() == X86ISD::FOR || N->getOpcode() == X86ISD::FXOR);
25740 // F[X]OR(0.0, x) -> x
25741 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
25742 if (C->getValueAPF().isPosZero())
25743 return N->getOperand(1);
25745 // F[X]OR(x, 0.0) -> x
25746 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
25747 if (C->getValueAPF().isPosZero())
25748 return N->getOperand(0);
25750 EVT VT = N->getValueType(0);
25751 if (VT.is512BitVector() && !Subtarget->hasDQI()) {
25753 MVT IntScalar = MVT::getIntegerVT(VT.getScalarSizeInBits());
25754 MVT IntVT = MVT::getVectorVT(IntScalar, VT.getVectorNumElements());
25756 SDValue Op0 = DAG.getNode(ISD::BITCAST, dl, IntVT, N->getOperand(0));
25757 SDValue Op1 = DAG.getNode(ISD::BITCAST, dl, IntVT, N->getOperand(1));
25758 unsigned IntOpcode = (N->getOpcode() == X86ISD::FOR) ? ISD::OR : ISD::XOR;
25759 SDValue IntOp = DAG.getNode(IntOpcode, dl, IntVT, Op0, Op1);
25760 return DAG.getNode(ISD::BITCAST, dl, VT, IntOp);
25765 /// Do target-specific dag combines on X86ISD::FMIN and X86ISD::FMAX nodes.
25766 static SDValue PerformFMinFMaxCombine(SDNode *N, SelectionDAG &DAG) {
25767 assert(N->getOpcode() == X86ISD::FMIN || N->getOpcode() == X86ISD::FMAX);
25769 // Only perform optimizations if UnsafeMath is used.
25770 if (!DAG.getTarget().Options.UnsafeFPMath)
25773 // If we run in unsafe-math mode, then convert the FMAX and FMIN nodes
25774 // into FMINC and FMAXC, which are Commutative operations.
25775 unsigned NewOp = 0;
25776 switch (N->getOpcode()) {
25777 default: llvm_unreachable("unknown opcode");
25778 case X86ISD::FMIN: NewOp = X86ISD::FMINC; break;
25779 case X86ISD::FMAX: NewOp = X86ISD::FMAXC; break;
25782 return DAG.getNode(NewOp, SDLoc(N), N->getValueType(0),
25783 N->getOperand(0), N->getOperand(1));
25786 /// Do target-specific dag combines on X86ISD::FAND nodes.
25787 static SDValue PerformFANDCombine(SDNode *N, SelectionDAG &DAG) {
25788 // FAND(0.0, x) -> 0.0
25789 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
25790 if (C->getValueAPF().isPosZero())
25791 return N->getOperand(0);
25793 // FAND(x, 0.0) -> 0.0
25794 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
25795 if (C->getValueAPF().isPosZero())
25796 return N->getOperand(1);
25801 /// Do target-specific dag combines on X86ISD::FANDN nodes
25802 static SDValue PerformFANDNCombine(SDNode *N, SelectionDAG &DAG) {
25803 // FANDN(0.0, x) -> x
25804 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
25805 if (C->getValueAPF().isPosZero())
25806 return N->getOperand(1);
25808 // FANDN(x, 0.0) -> 0.0
25809 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
25810 if (C->getValueAPF().isPosZero())
25811 return N->getOperand(1);
25816 static SDValue PerformBTCombine(SDNode *N,
25818 TargetLowering::DAGCombinerInfo &DCI) {
25819 // BT ignores high bits in the bit index operand.
25820 SDValue Op1 = N->getOperand(1);
25821 if (Op1.hasOneUse()) {
25822 unsigned BitWidth = Op1.getValueSizeInBits();
25823 APInt DemandedMask = APInt::getLowBitsSet(BitWidth, Log2_32(BitWidth));
25824 APInt KnownZero, KnownOne;
25825 TargetLowering::TargetLoweringOpt TLO(DAG, !DCI.isBeforeLegalize(),
25826 !DCI.isBeforeLegalizeOps());
25827 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
25828 if (TLO.ShrinkDemandedConstant(Op1, DemandedMask) ||
25829 TLI.SimplifyDemandedBits(Op1, DemandedMask, KnownZero, KnownOne, TLO))
25830 DCI.CommitTargetLoweringOpt(TLO);
25835 static SDValue PerformVZEXT_MOVLCombine(SDNode *N, SelectionDAG &DAG) {
25836 SDValue Op = N->getOperand(0);
25837 if (Op.getOpcode() == ISD::BITCAST)
25838 Op = Op.getOperand(0);
25839 EVT VT = N->getValueType(0), OpVT = Op.getValueType();
25840 if (Op.getOpcode() == X86ISD::VZEXT_LOAD &&
25841 VT.getVectorElementType().getSizeInBits() ==
25842 OpVT.getVectorElementType().getSizeInBits()) {
25843 return DAG.getNode(ISD::BITCAST, SDLoc(N), VT, Op);
25848 static SDValue PerformSIGN_EXTEND_INREGCombine(SDNode *N, SelectionDAG &DAG,
25849 const X86Subtarget *Subtarget) {
25850 EVT VT = N->getValueType(0);
25851 if (!VT.isVector())
25854 SDValue N0 = N->getOperand(0);
25855 SDValue N1 = N->getOperand(1);
25856 EVT ExtraVT = cast<VTSDNode>(N1)->getVT();
25859 // The SIGN_EXTEND_INREG to v4i64 is expensive operation on the
25860 // both SSE and AVX2 since there is no sign-extended shift right
25861 // operation on a vector with 64-bit elements.
25862 //(sext_in_reg (v4i64 anyext (v4i32 x )), ExtraVT) ->
25863 // (v4i64 sext (v4i32 sext_in_reg (v4i32 x , ExtraVT)))
25864 if (VT == MVT::v4i64 && (N0.getOpcode() == ISD::ANY_EXTEND ||
25865 N0.getOpcode() == ISD::SIGN_EXTEND)) {
25866 SDValue N00 = N0.getOperand(0);
25868 // EXTLOAD has a better solution on AVX2,
25869 // it may be replaced with X86ISD::VSEXT node.
25870 if (N00.getOpcode() == ISD::LOAD && Subtarget->hasInt256())
25871 if (!ISD::isNormalLoad(N00.getNode()))
25874 if (N00.getValueType() == MVT::v4i32 && ExtraVT.getSizeInBits() < 128) {
25875 SDValue Tmp = DAG.getNode(ISD::SIGN_EXTEND_INREG, dl, MVT::v4i32,
25877 return DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i64, Tmp);
25883 /// sext(add_nsw(x, C)) --> add(sext(x), C_sext)
25884 /// Promoting a sign extension ahead of an 'add nsw' exposes opportunities
25885 /// to combine math ops, use an LEA, or use a complex addressing mode. This can
25886 /// eliminate extend, add, and shift instructions.
25887 static SDValue promoteSextBeforeAddNSW(SDNode *Sext, SelectionDAG &DAG,
25888 const X86Subtarget *Subtarget) {
25889 // TODO: This should be valid for other integer types.
25890 EVT VT = Sext->getValueType(0);
25891 if (VT != MVT::i64)
25894 // We need an 'add nsw' feeding into the 'sext'.
25895 SDValue Add = Sext->getOperand(0);
25896 if (Add.getOpcode() != ISD::ADD || !Add->getFlags()->hasNoSignedWrap())
25899 // Having a constant operand to the 'add' ensures that we are not increasing
25900 // the instruction count because the constant is extended for free below.
25901 // A constant operand can also become the displacement field of an LEA.
25902 auto *AddOp1 = dyn_cast<ConstantSDNode>(Add.getOperand(1));
25906 // Don't make the 'add' bigger if there's no hope of combining it with some
25907 // other 'add' or 'shl' instruction.
25908 // TODO: It may be profitable to generate simpler LEA instructions in place
25909 // of single 'add' instructions, but the cost model for selecting an LEA
25910 // currently has a high threshold.
25911 bool HasLEAPotential = false;
25912 for (auto *User : Sext->uses()) {
25913 if (User->getOpcode() == ISD::ADD || User->getOpcode() == ISD::SHL) {
25914 HasLEAPotential = true;
25918 if (!HasLEAPotential)
25921 // Everything looks good, so pull the 'sext' ahead of the 'add'.
25922 int64_t AddConstant = AddOp1->getSExtValue();
25923 SDValue AddOp0 = Add.getOperand(0);
25924 SDValue NewSext = DAG.getNode(ISD::SIGN_EXTEND, SDLoc(Sext), VT, AddOp0);
25925 SDValue NewConstant = DAG.getConstant(AddConstant, SDLoc(Add), VT);
25927 // The wider add is guaranteed to not wrap because both operands are
25930 Flags.setNoSignedWrap(true);
25931 return DAG.getNode(ISD::ADD, SDLoc(Add), VT, NewSext, NewConstant, &Flags);
25934 static SDValue PerformSExtCombine(SDNode *N, SelectionDAG &DAG,
25935 TargetLowering::DAGCombinerInfo &DCI,
25936 const X86Subtarget *Subtarget) {
25937 SDValue N0 = N->getOperand(0);
25938 EVT VT = N->getValueType(0);
25939 EVT SVT = VT.getScalarType();
25940 EVT InVT = N0.getValueType();
25941 EVT InSVT = InVT.getScalarType();
25944 // (i8,i32 sext (sdivrem (i8 x, i8 y)) ->
25945 // (i8,i32 (sdivrem_sext_hreg (i8 x, i8 y)
25946 // This exposes the sext to the sdivrem lowering, so that it directly extends
25947 // from AH (which we otherwise need to do contortions to access).
25948 if (N0.getOpcode() == ISD::SDIVREM && N0.getResNo() == 1 &&
25949 InVT == MVT::i8 && VT == MVT::i32) {
25950 SDVTList NodeTys = DAG.getVTList(MVT::i8, VT);
25951 SDValue R = DAG.getNode(X86ISD::SDIVREM8_SEXT_HREG, DL, NodeTys,
25952 N0.getOperand(0), N0.getOperand(1));
25953 DAG.ReplaceAllUsesOfValueWith(N0.getValue(0), R.getValue(0));
25954 return R.getValue(1);
25957 if (!DCI.isBeforeLegalizeOps()) {
25958 if (InVT == MVT::i1) {
25959 SDValue Zero = DAG.getConstant(0, DL, VT);
25961 DAG.getConstant(APInt::getAllOnesValue(VT.getSizeInBits()), DL, VT);
25962 return DAG.getNode(ISD::SELECT, DL, VT, N0, AllOnes, Zero);
25967 if (VT.isVector() && Subtarget->hasSSE2()) {
25968 auto ExtendVecSize = [&DAG](SDLoc DL, SDValue N, unsigned Size) {
25969 EVT InVT = N.getValueType();
25970 EVT OutVT = EVT::getVectorVT(*DAG.getContext(), InVT.getScalarType(),
25971 Size / InVT.getScalarSizeInBits());
25972 SmallVector<SDValue, 8> Opnds(Size / InVT.getSizeInBits(),
25973 DAG.getUNDEF(InVT));
25975 return DAG.getNode(ISD::CONCAT_VECTORS, DL, OutVT, Opnds);
25978 // If target-size is less than 128-bits, extend to a type that would extend
25979 // to 128 bits, extend that and extract the original target vector.
25980 if (VT.getSizeInBits() < 128 && !(128 % VT.getSizeInBits()) &&
25981 (SVT == MVT::i64 || SVT == MVT::i32 || SVT == MVT::i16) &&
25982 (InSVT == MVT::i32 || InSVT == MVT::i16 || InSVT == MVT::i8)) {
25983 unsigned Scale = 128 / VT.getSizeInBits();
25985 EVT::getVectorVT(*DAG.getContext(), SVT, 128 / SVT.getSizeInBits());
25986 SDValue Ex = ExtendVecSize(DL, N0, Scale * InVT.getSizeInBits());
25987 SDValue SExt = DAG.getNode(ISD::SIGN_EXTEND, DL, ExVT, Ex);
25988 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, VT, SExt,
25989 DAG.getIntPtrConstant(0, DL));
25992 // If target-size is 128-bits, then convert to ISD::SIGN_EXTEND_VECTOR_INREG
25993 // which ensures lowering to X86ISD::VSEXT (pmovsx*).
25994 if (VT.getSizeInBits() == 128 &&
25995 (SVT == MVT::i64 || SVT == MVT::i32 || SVT == MVT::i16) &&
25996 (InSVT == MVT::i32 || InSVT == MVT::i16 || InSVT == MVT::i8)) {
25997 SDValue ExOp = ExtendVecSize(DL, N0, 128);
25998 return DAG.getSignExtendVectorInReg(ExOp, DL, VT);
26001 // On pre-AVX2 targets, split into 128-bit nodes of
26002 // ISD::SIGN_EXTEND_VECTOR_INREG.
26003 if (!Subtarget->hasInt256() && !(VT.getSizeInBits() % 128) &&
26004 (SVT == MVT::i64 || SVT == MVT::i32 || SVT == MVT::i16) &&
26005 (InSVT == MVT::i32 || InSVT == MVT::i16 || InSVT == MVT::i8)) {
26006 unsigned NumVecs = VT.getSizeInBits() / 128;
26007 unsigned NumSubElts = 128 / SVT.getSizeInBits();
26008 EVT SubVT = EVT::getVectorVT(*DAG.getContext(), SVT, NumSubElts);
26009 EVT InSubVT = EVT::getVectorVT(*DAG.getContext(), InSVT, NumSubElts);
26011 SmallVector<SDValue, 8> Opnds;
26012 for (unsigned i = 0, Offset = 0; i != NumVecs;
26013 ++i, Offset += NumSubElts) {
26014 SDValue SrcVec = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, InSubVT, N0,
26015 DAG.getIntPtrConstant(Offset, DL));
26016 SrcVec = ExtendVecSize(DL, SrcVec, 128);
26017 SrcVec = DAG.getSignExtendVectorInReg(SrcVec, DL, SubVT);
26018 Opnds.push_back(SrcVec);
26020 return DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, Opnds);
26024 if (Subtarget->hasAVX() && VT.isVector() && VT.getSizeInBits() == 256)
26025 if (SDValue R = WidenMaskArithmetic(N, DAG, DCI, Subtarget))
26028 if (SDValue NewAdd = promoteSextBeforeAddNSW(N, DAG, Subtarget))
26034 static SDValue PerformFMACombine(SDNode *N, SelectionDAG &DAG,
26035 const X86Subtarget* Subtarget) {
26037 EVT VT = N->getValueType(0);
26039 // Let legalize expand this if it isn't a legal type yet.
26040 if (!DAG.getTargetLoweringInfo().isTypeLegal(VT))
26043 EVT ScalarVT = VT.getScalarType();
26044 if ((ScalarVT != MVT::f32 && ScalarVT != MVT::f64) ||
26045 (!Subtarget->hasFMA() && !Subtarget->hasFMA4() &&
26046 !Subtarget->hasAVX512()))
26049 SDValue A = N->getOperand(0);
26050 SDValue B = N->getOperand(1);
26051 SDValue C = N->getOperand(2);
26053 bool NegA = (A.getOpcode() == ISD::FNEG);
26054 bool NegB = (B.getOpcode() == ISD::FNEG);
26055 bool NegC = (C.getOpcode() == ISD::FNEG);
26057 // Negative multiplication when NegA xor NegB
26058 bool NegMul = (NegA != NegB);
26060 A = A.getOperand(0);
26062 B = B.getOperand(0);
26064 C = C.getOperand(0);
26068 Opcode = (!NegC) ? X86ISD::FMADD : X86ISD::FMSUB;
26070 Opcode = (!NegC) ? X86ISD::FNMADD : X86ISD::FNMSUB;
26072 return DAG.getNode(Opcode, dl, VT, A, B, C);
26075 static SDValue PerformZExtCombine(SDNode *N, SelectionDAG &DAG,
26076 TargetLowering::DAGCombinerInfo &DCI,
26077 const X86Subtarget *Subtarget) {
26078 // (i32 zext (and (i8 x86isd::setcc_carry), 1)) ->
26079 // (and (i32 x86isd::setcc_carry), 1)
26080 // This eliminates the zext. This transformation is necessary because
26081 // ISD::SETCC is always legalized to i8.
26083 SDValue N0 = N->getOperand(0);
26084 EVT VT = N->getValueType(0);
26086 if (N0.getOpcode() == ISD::AND &&
26088 N0.getOperand(0).hasOneUse()) {
26089 SDValue N00 = N0.getOperand(0);
26090 if (N00.getOpcode() == X86ISD::SETCC_CARRY) {
26091 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
26092 if (!C || C->getZExtValue() != 1)
26094 return DAG.getNode(ISD::AND, dl, VT,
26095 DAG.getNode(X86ISD::SETCC_CARRY, dl, VT,
26096 N00.getOperand(0), N00.getOperand(1)),
26097 DAG.getConstant(1, dl, VT));
26101 if (N0.getOpcode() == ISD::TRUNCATE &&
26103 N0.getOperand(0).hasOneUse()) {
26104 SDValue N00 = N0.getOperand(0);
26105 if (N00.getOpcode() == X86ISD::SETCC_CARRY) {
26106 return DAG.getNode(ISD::AND, dl, VT,
26107 DAG.getNode(X86ISD::SETCC_CARRY, dl, VT,
26108 N00.getOperand(0), N00.getOperand(1)),
26109 DAG.getConstant(1, dl, VT));
26113 if (VT.is256BitVector())
26114 if (SDValue R = WidenMaskArithmetic(N, DAG, DCI, Subtarget))
26117 // (i8,i32 zext (udivrem (i8 x, i8 y)) ->
26118 // (i8,i32 (udivrem_zext_hreg (i8 x, i8 y)
26119 // This exposes the zext to the udivrem lowering, so that it directly extends
26120 // from AH (which we otherwise need to do contortions to access).
26121 if (N0.getOpcode() == ISD::UDIVREM &&
26122 N0.getResNo() == 1 && N0.getValueType() == MVT::i8 &&
26123 (VT == MVT::i32 || VT == MVT::i64)) {
26124 SDVTList NodeTys = DAG.getVTList(MVT::i8, VT);
26125 SDValue R = DAG.getNode(X86ISD::UDIVREM8_ZEXT_HREG, dl, NodeTys,
26126 N0.getOperand(0), N0.getOperand(1));
26127 DAG.ReplaceAllUsesOfValueWith(N0.getValue(0), R.getValue(0));
26128 return R.getValue(1);
26134 // Optimize x == -y --> x+y == 0
26135 // x != -y --> x+y != 0
26136 static SDValue PerformISDSETCCCombine(SDNode *N, SelectionDAG &DAG,
26137 const X86Subtarget* Subtarget) {
26138 ISD::CondCode CC = cast<CondCodeSDNode>(N->getOperand(2))->get();
26139 SDValue LHS = N->getOperand(0);
26140 SDValue RHS = N->getOperand(1);
26141 EVT VT = N->getValueType(0);
26144 if ((CC == ISD::SETNE || CC == ISD::SETEQ) && LHS.getOpcode() == ISD::SUB)
26145 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(LHS.getOperand(0)))
26146 if (C->getAPIntValue() == 0 && LHS.hasOneUse()) {
26147 SDValue addV = DAG.getNode(ISD::ADD, DL, LHS.getValueType(), RHS,
26148 LHS.getOperand(1));
26149 return DAG.getSetCC(DL, N->getValueType(0), addV,
26150 DAG.getConstant(0, DL, addV.getValueType()), CC);
26152 if ((CC == ISD::SETNE || CC == ISD::SETEQ) && RHS.getOpcode() == ISD::SUB)
26153 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(RHS.getOperand(0)))
26154 if (C->getAPIntValue() == 0 && RHS.hasOneUse()) {
26155 SDValue addV = DAG.getNode(ISD::ADD, DL, RHS.getValueType(), LHS,
26156 RHS.getOperand(1));
26157 return DAG.getSetCC(DL, N->getValueType(0), addV,
26158 DAG.getConstant(0, DL, addV.getValueType()), CC);
26161 if (VT.getScalarType() == MVT::i1 &&
26162 (CC == ISD::SETNE || CC == ISD::SETEQ || ISD::isSignedIntSetCC(CC))) {
26164 (LHS.getOpcode() == ISD::SIGN_EXTEND) &&
26165 (LHS.getOperand(0).getValueType().getScalarType() == MVT::i1);
26166 bool IsVZero1 = ISD::isBuildVectorAllZeros(RHS.getNode());
26168 if (!IsSEXT0 || !IsVZero1) {
26169 // Swap the operands and update the condition code.
26170 std::swap(LHS, RHS);
26171 CC = ISD::getSetCCSwappedOperands(CC);
26173 IsSEXT0 = (LHS.getOpcode() == ISD::SIGN_EXTEND) &&
26174 (LHS.getOperand(0).getValueType().getScalarType() == MVT::i1);
26175 IsVZero1 = ISD::isBuildVectorAllZeros(RHS.getNode());
26178 if (IsSEXT0 && IsVZero1) {
26179 assert(VT == LHS.getOperand(0).getValueType() &&
26180 "Uexpected operand type");
26181 if (CC == ISD::SETGT)
26182 return DAG.getConstant(0, DL, VT);
26183 if (CC == ISD::SETLE)
26184 return DAG.getConstant(1, DL, VT);
26185 if (CC == ISD::SETEQ || CC == ISD::SETGE)
26186 return DAG.getNOT(DL, LHS.getOperand(0), VT);
26188 assert((CC == ISD::SETNE || CC == ISD::SETLT) &&
26189 "Unexpected condition code!");
26190 return LHS.getOperand(0);
26197 static SDValue NarrowVectorLoadToElement(LoadSDNode *Load, unsigned Index,
26198 SelectionDAG &DAG) {
26200 MVT VT = Load->getSimpleValueType(0);
26201 MVT EVT = VT.getVectorElementType();
26202 SDValue Addr = Load->getOperand(1);
26203 SDValue NewAddr = DAG.getNode(
26204 ISD::ADD, dl, Addr.getSimpleValueType(), Addr,
26205 DAG.getConstant(Index * EVT.getStoreSize(), dl,
26206 Addr.getSimpleValueType()));
26209 DAG.getLoad(EVT, dl, Load->getChain(), NewAddr,
26210 DAG.getMachineFunction().getMachineMemOperand(
26211 Load->getMemOperand(), 0, EVT.getStoreSize()));
26215 static SDValue PerformINSERTPSCombine(SDNode *N, SelectionDAG &DAG,
26216 const X86Subtarget *Subtarget) {
26218 MVT VT = N->getOperand(1)->getSimpleValueType(0);
26219 assert((VT == MVT::v4f32 || VT == MVT::v4i32) &&
26220 "X86insertps is only defined for v4x32");
26222 SDValue Ld = N->getOperand(1);
26223 if (MayFoldLoad(Ld)) {
26224 // Extract the countS bits from the immediate so we can get the proper
26225 // address when narrowing the vector load to a specific element.
26226 // When the second source op is a memory address, insertps doesn't use
26227 // countS and just gets an f32 from that address.
26228 unsigned DestIndex =
26229 cast<ConstantSDNode>(N->getOperand(2))->getZExtValue() >> 6;
26231 Ld = NarrowVectorLoadToElement(cast<LoadSDNode>(Ld), DestIndex, DAG);
26233 // Create this as a scalar to vector to match the instruction pattern.
26234 SDValue LoadScalarToVector = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Ld);
26235 // countS bits are ignored when loading from memory on insertps, which
26236 // means we don't need to explicitly set them to 0.
26237 return DAG.getNode(X86ISD::INSERTPS, dl, VT, N->getOperand(0),
26238 LoadScalarToVector, N->getOperand(2));
26243 static SDValue PerformBLENDICombine(SDNode *N, SelectionDAG &DAG) {
26244 SDValue V0 = N->getOperand(0);
26245 SDValue V1 = N->getOperand(1);
26247 EVT VT = N->getValueType(0);
26249 // Canonicalize a v2f64 blend with a mask of 2 by swapping the vector
26250 // operands and changing the mask to 1. This saves us a bunch of
26251 // pattern-matching possibilities related to scalar math ops in SSE/AVX.
26252 // x86InstrInfo knows how to commute this back after instruction selection
26253 // if it would help register allocation.
26255 // TODO: If optimizing for size or a processor that doesn't suffer from
26256 // partial register update stalls, this should be transformed into a MOVSD
26257 // instruction because a MOVSD is 1-2 bytes smaller than a BLENDPD.
26259 if (VT == MVT::v2f64)
26260 if (auto *Mask = dyn_cast<ConstantSDNode>(N->getOperand(2)))
26261 if (Mask->getZExtValue() == 2 && !isShuffleFoldableLoad(V0)) {
26262 SDValue NewMask = DAG.getConstant(1, DL, MVT::i8);
26263 return DAG.getNode(X86ISD::BLENDI, DL, VT, V1, V0, NewMask);
26269 // Helper function of PerformSETCCCombine. It is to materialize "setb reg"
26270 // as "sbb reg,reg", since it can be extended without zext and produces
26271 // an all-ones bit which is more useful than 0/1 in some cases.
26272 static SDValue MaterializeSETB(SDLoc DL, SDValue EFLAGS, SelectionDAG &DAG,
26275 return DAG.getNode(ISD::AND, DL, VT,
26276 DAG.getNode(X86ISD::SETCC_CARRY, DL, MVT::i8,
26277 DAG.getConstant(X86::COND_B, DL, MVT::i8),
26279 DAG.getConstant(1, DL, VT));
26280 assert (VT == MVT::i1 && "Unexpected type for SECCC node");
26281 return DAG.getNode(ISD::TRUNCATE, DL, MVT::i1,
26282 DAG.getNode(X86ISD::SETCC_CARRY, DL, MVT::i8,
26283 DAG.getConstant(X86::COND_B, DL, MVT::i8),
26287 // Optimize RES = X86ISD::SETCC CONDCODE, EFLAG_INPUT
26288 static SDValue PerformSETCCCombine(SDNode *N, SelectionDAG &DAG,
26289 TargetLowering::DAGCombinerInfo &DCI,
26290 const X86Subtarget *Subtarget) {
26292 X86::CondCode CC = X86::CondCode(N->getConstantOperandVal(0));
26293 SDValue EFLAGS = N->getOperand(1);
26295 if (CC == X86::COND_A) {
26296 // Try to convert COND_A into COND_B in an attempt to facilitate
26297 // materializing "setb reg".
26299 // Do not flip "e > c", where "c" is a constant, because Cmp instruction
26300 // cannot take an immediate as its first operand.
26302 if (EFLAGS.getOpcode() == X86ISD::SUB && EFLAGS.hasOneUse() &&
26303 EFLAGS.getValueType().isInteger() &&
26304 !isa<ConstantSDNode>(EFLAGS.getOperand(1))) {
26305 SDValue NewSub = DAG.getNode(X86ISD::SUB, SDLoc(EFLAGS),
26306 EFLAGS.getNode()->getVTList(),
26307 EFLAGS.getOperand(1), EFLAGS.getOperand(0));
26308 SDValue NewEFLAGS = SDValue(NewSub.getNode(), EFLAGS.getResNo());
26309 return MaterializeSETB(DL, NewEFLAGS, DAG, N->getSimpleValueType(0));
26313 // Materialize "setb reg" as "sbb reg,reg", since it can be extended without
26314 // a zext and produces an all-ones bit which is more useful than 0/1 in some
26316 if (CC == X86::COND_B)
26317 return MaterializeSETB(DL, EFLAGS, DAG, N->getSimpleValueType(0));
26319 if (SDValue Flags = checkBoolTestSetCCCombine(EFLAGS, CC)) {
26320 SDValue Cond = DAG.getConstant(CC, DL, MVT::i8);
26321 return DAG.getNode(X86ISD::SETCC, DL, N->getVTList(), Cond, Flags);
26327 // Optimize branch condition evaluation.
26329 static SDValue PerformBrCondCombine(SDNode *N, SelectionDAG &DAG,
26330 TargetLowering::DAGCombinerInfo &DCI,
26331 const X86Subtarget *Subtarget) {
26333 SDValue Chain = N->getOperand(0);
26334 SDValue Dest = N->getOperand(1);
26335 SDValue EFLAGS = N->getOperand(3);
26336 X86::CondCode CC = X86::CondCode(N->getConstantOperandVal(2));
26338 if (SDValue Flags = checkBoolTestSetCCCombine(EFLAGS, CC)) {
26339 SDValue Cond = DAG.getConstant(CC, DL, MVT::i8);
26340 return DAG.getNode(X86ISD::BRCOND, DL, N->getVTList(), Chain, Dest, Cond,
26347 static SDValue performVectorCompareAndMaskUnaryOpCombine(SDNode *N,
26348 SelectionDAG &DAG) {
26349 // Take advantage of vector comparisons producing 0 or -1 in each lane to
26350 // optimize away operation when it's from a constant.
26352 // The general transformation is:
26353 // UNARYOP(AND(VECTOR_CMP(x,y), constant)) -->
26354 // AND(VECTOR_CMP(x,y), constant2)
26355 // constant2 = UNARYOP(constant)
26357 // Early exit if this isn't a vector operation, the operand of the
26358 // unary operation isn't a bitwise AND, or if the sizes of the operations
26359 // aren't the same.
26360 EVT VT = N->getValueType(0);
26361 if (!VT.isVector() || N->getOperand(0)->getOpcode() != ISD::AND ||
26362 N->getOperand(0)->getOperand(0)->getOpcode() != ISD::SETCC ||
26363 VT.getSizeInBits() != N->getOperand(0)->getValueType(0).getSizeInBits())
26366 // Now check that the other operand of the AND is a constant. We could
26367 // make the transformation for non-constant splats as well, but it's unclear
26368 // that would be a benefit as it would not eliminate any operations, just
26369 // perform one more step in scalar code before moving to the vector unit.
26370 if (BuildVectorSDNode *BV =
26371 dyn_cast<BuildVectorSDNode>(N->getOperand(0)->getOperand(1))) {
26372 // Bail out if the vector isn't a constant.
26373 if (!BV->isConstant())
26376 // Everything checks out. Build up the new and improved node.
26378 EVT IntVT = BV->getValueType(0);
26379 // Create a new constant of the appropriate type for the transformed
26381 SDValue SourceConst = DAG.getNode(N->getOpcode(), DL, VT, SDValue(BV, 0));
26382 // The AND node needs bitcasts to/from an integer vector type around it.
26383 SDValue MaskConst = DAG.getBitcast(IntVT, SourceConst);
26384 SDValue NewAnd = DAG.getNode(ISD::AND, DL, IntVT,
26385 N->getOperand(0)->getOperand(0), MaskConst);
26386 SDValue Res = DAG.getBitcast(VT, NewAnd);
26393 static SDValue PerformUINT_TO_FPCombine(SDNode *N, SelectionDAG &DAG,
26394 const X86Subtarget *Subtarget) {
26395 SDValue Op0 = N->getOperand(0);
26396 EVT VT = N->getValueType(0);
26397 EVT InVT = Op0.getValueType();
26398 EVT InSVT = InVT.getScalarType();
26399 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
26401 // UINT_TO_FP(vXi8) -> SINT_TO_FP(ZEXT(vXi8 to vXi32))
26402 // UINT_TO_FP(vXi16) -> SINT_TO_FP(ZEXT(vXi16 to vXi32))
26403 if (InVT.isVector() && (InSVT == MVT::i8 || InSVT == MVT::i16)) {
26405 EVT DstVT = EVT::getVectorVT(*DAG.getContext(), MVT::i32,
26406 InVT.getVectorNumElements());
26407 SDValue P = DAG.getNode(ISD::ZERO_EXTEND, dl, DstVT, Op0);
26409 if (TLI.isOperationLegal(ISD::UINT_TO_FP, DstVT))
26410 return DAG.getNode(ISD::UINT_TO_FP, dl, VT, P);
26412 return DAG.getNode(ISD::SINT_TO_FP, dl, VT, P);
26418 static SDValue PerformSINT_TO_FPCombine(SDNode *N, SelectionDAG &DAG,
26419 const X86Subtarget *Subtarget) {
26420 // First try to optimize away the conversion entirely when it's
26421 // conditionally from a constant. Vectors only.
26422 if (SDValue Res = performVectorCompareAndMaskUnaryOpCombine(N, DAG))
26425 // Now move on to more general possibilities.
26426 SDValue Op0 = N->getOperand(0);
26427 EVT VT = N->getValueType(0);
26428 EVT InVT = Op0.getValueType();
26429 EVT InSVT = InVT.getScalarType();
26431 // SINT_TO_FP(vXi8) -> SINT_TO_FP(SEXT(vXi8 to vXi32))
26432 // SINT_TO_FP(vXi16) -> SINT_TO_FP(SEXT(vXi16 to vXi32))
26433 if (InVT.isVector() && (InSVT == MVT::i8 || InSVT == MVT::i16)) {
26435 EVT DstVT = EVT::getVectorVT(*DAG.getContext(), MVT::i32,
26436 InVT.getVectorNumElements());
26437 SDValue P = DAG.getNode(ISD::SIGN_EXTEND, dl, DstVT, Op0);
26438 return DAG.getNode(ISD::SINT_TO_FP, dl, VT, P);
26441 // Transform (SINT_TO_FP (i64 ...)) into an x87 operation if we have
26442 // a 32-bit target where SSE doesn't support i64->FP operations.
26443 if (Op0.getOpcode() == ISD::LOAD) {
26444 LoadSDNode *Ld = cast<LoadSDNode>(Op0.getNode());
26445 EVT LdVT = Ld->getValueType(0);
26447 // This transformation is not supported if the result type is f16
26448 if (VT == MVT::f16)
26451 if (!Ld->isVolatile() && !VT.isVector() &&
26452 ISD::isNON_EXTLoad(Op0.getNode()) && Op0.hasOneUse() &&
26453 !Subtarget->is64Bit() && LdVT == MVT::i64) {
26454 SDValue FILDChain = Subtarget->getTargetLowering()->BuildFILD(
26455 SDValue(N, 0), LdVT, Ld->getChain(), Op0, DAG);
26456 DAG.ReplaceAllUsesOfValueWith(Op0.getValue(1), FILDChain.getValue(1));
26463 // Optimize RES, EFLAGS = X86ISD::ADC LHS, RHS, EFLAGS
26464 static SDValue PerformADCCombine(SDNode *N, SelectionDAG &DAG,
26465 X86TargetLowering::DAGCombinerInfo &DCI) {
26466 // If the LHS and RHS of the ADC node are zero, then it can't overflow and
26467 // the result is either zero or one (depending on the input carry bit).
26468 // Strength reduce this down to a "set on carry" aka SETCC_CARRY&1.
26469 if (X86::isZeroNode(N->getOperand(0)) &&
26470 X86::isZeroNode(N->getOperand(1)) &&
26471 // We don't have a good way to replace an EFLAGS use, so only do this when
26473 SDValue(N, 1).use_empty()) {
26475 EVT VT = N->getValueType(0);
26476 SDValue CarryOut = DAG.getConstant(0, DL, N->getValueType(1));
26477 SDValue Res1 = DAG.getNode(ISD::AND, DL, VT,
26478 DAG.getNode(X86ISD::SETCC_CARRY, DL, VT,
26479 DAG.getConstant(X86::COND_B, DL,
26482 DAG.getConstant(1, DL, VT));
26483 return DCI.CombineTo(N, Res1, CarryOut);
26489 // fold (add Y, (sete X, 0)) -> adc 0, Y
26490 // (add Y, (setne X, 0)) -> sbb -1, Y
26491 // (sub (sete X, 0), Y) -> sbb 0, Y
26492 // (sub (setne X, 0), Y) -> adc -1, Y
26493 static SDValue OptimizeConditionalInDecrement(SDNode *N, SelectionDAG &DAG) {
26496 // Look through ZExts.
26497 SDValue Ext = N->getOperand(N->getOpcode() == ISD::SUB ? 1 : 0);
26498 if (Ext.getOpcode() != ISD::ZERO_EXTEND || !Ext.hasOneUse())
26501 SDValue SetCC = Ext.getOperand(0);
26502 if (SetCC.getOpcode() != X86ISD::SETCC || !SetCC.hasOneUse())
26505 X86::CondCode CC = (X86::CondCode)SetCC.getConstantOperandVal(0);
26506 if (CC != X86::COND_E && CC != X86::COND_NE)
26509 SDValue Cmp = SetCC.getOperand(1);
26510 if (Cmp.getOpcode() != X86ISD::CMP || !Cmp.hasOneUse() ||
26511 !X86::isZeroNode(Cmp.getOperand(1)) ||
26512 !Cmp.getOperand(0).getValueType().isInteger())
26515 SDValue CmpOp0 = Cmp.getOperand(0);
26516 SDValue NewCmp = DAG.getNode(X86ISD::CMP, DL, MVT::i32, CmpOp0,
26517 DAG.getConstant(1, DL, CmpOp0.getValueType()));
26519 SDValue OtherVal = N->getOperand(N->getOpcode() == ISD::SUB ? 0 : 1);
26520 if (CC == X86::COND_NE)
26521 return DAG.getNode(N->getOpcode() == ISD::SUB ? X86ISD::ADC : X86ISD::SBB,
26522 DL, OtherVal.getValueType(), OtherVal,
26523 DAG.getConstant(-1ULL, DL, OtherVal.getValueType()),
26525 return DAG.getNode(N->getOpcode() == ISD::SUB ? X86ISD::SBB : X86ISD::ADC,
26526 DL, OtherVal.getValueType(), OtherVal,
26527 DAG.getConstant(0, DL, OtherVal.getValueType()), NewCmp);
26530 /// PerformADDCombine - Do target-specific dag combines on integer adds.
26531 static SDValue PerformAddCombine(SDNode *N, SelectionDAG &DAG,
26532 const X86Subtarget *Subtarget) {
26533 EVT VT = N->getValueType(0);
26534 SDValue Op0 = N->getOperand(0);
26535 SDValue Op1 = N->getOperand(1);
26537 // Try to synthesize horizontal adds from adds of shuffles.
26538 if (((Subtarget->hasSSSE3() && (VT == MVT::v8i16 || VT == MVT::v4i32)) ||
26539 (Subtarget->hasInt256() && (VT == MVT::v16i16 || VT == MVT::v8i32))) &&
26540 isHorizontalBinOp(Op0, Op1, true))
26541 return DAG.getNode(X86ISD::HADD, SDLoc(N), VT, Op0, Op1);
26543 return OptimizeConditionalInDecrement(N, DAG);
26546 static SDValue PerformSubCombine(SDNode *N, SelectionDAG &DAG,
26547 const X86Subtarget *Subtarget) {
26548 SDValue Op0 = N->getOperand(0);
26549 SDValue Op1 = N->getOperand(1);
26551 // X86 can't encode an immediate LHS of a sub. See if we can push the
26552 // negation into a preceding instruction.
26553 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op0)) {
26554 // If the RHS of the sub is a XOR with one use and a constant, invert the
26555 // immediate. Then add one to the LHS of the sub so we can turn
26556 // X-Y -> X+~Y+1, saving one register.
26557 if (Op1->hasOneUse() && Op1.getOpcode() == ISD::XOR &&
26558 isa<ConstantSDNode>(Op1.getOperand(1))) {
26559 APInt XorC = cast<ConstantSDNode>(Op1.getOperand(1))->getAPIntValue();
26560 EVT VT = Op0.getValueType();
26561 SDValue NewXor = DAG.getNode(ISD::XOR, SDLoc(Op1), VT,
26563 DAG.getConstant(~XorC, SDLoc(Op1), VT));
26564 return DAG.getNode(ISD::ADD, SDLoc(N), VT, NewXor,
26565 DAG.getConstant(C->getAPIntValue() + 1, SDLoc(N), VT));
26569 // Try to synthesize horizontal adds from adds of shuffles.
26570 EVT VT = N->getValueType(0);
26571 if (((Subtarget->hasSSSE3() && (VT == MVT::v8i16 || VT == MVT::v4i32)) ||
26572 (Subtarget->hasInt256() && (VT == MVT::v16i16 || VT == MVT::v8i32))) &&
26573 isHorizontalBinOp(Op0, Op1, true))
26574 return DAG.getNode(X86ISD::HSUB, SDLoc(N), VT, Op0, Op1);
26576 return OptimizeConditionalInDecrement(N, DAG);
26579 /// performVZEXTCombine - Performs build vector combines
26580 static SDValue performVZEXTCombine(SDNode *N, SelectionDAG &DAG,
26581 TargetLowering::DAGCombinerInfo &DCI,
26582 const X86Subtarget *Subtarget) {
26584 MVT VT = N->getSimpleValueType(0);
26585 SDValue Op = N->getOperand(0);
26586 MVT OpVT = Op.getSimpleValueType();
26587 MVT OpEltVT = OpVT.getVectorElementType();
26588 unsigned InputBits = OpEltVT.getSizeInBits() * VT.getVectorNumElements();
26590 // (vzext (bitcast (vzext (x)) -> (vzext x)
26592 while (V.getOpcode() == ISD::BITCAST)
26593 V = V.getOperand(0);
26595 if (V != Op && V.getOpcode() == X86ISD::VZEXT) {
26596 MVT InnerVT = V.getSimpleValueType();
26597 MVT InnerEltVT = InnerVT.getVectorElementType();
26599 // If the element sizes match exactly, we can just do one larger vzext. This
26600 // is always an exact type match as vzext operates on integer types.
26601 if (OpEltVT == InnerEltVT) {
26602 assert(OpVT == InnerVT && "Types must match for vzext!");
26603 return DAG.getNode(X86ISD::VZEXT, DL, VT, V.getOperand(0));
26606 // The only other way we can combine them is if only a single element of the
26607 // inner vzext is used in the input to the outer vzext.
26608 if (InnerEltVT.getSizeInBits() < InputBits)
26611 // In this case, the inner vzext is completely dead because we're going to
26612 // only look at bits inside of the low element. Just do the outer vzext on
26613 // a bitcast of the input to the inner.
26614 return DAG.getNode(X86ISD::VZEXT, DL, VT, DAG.getBitcast(OpVT, V));
26617 // Check if we can bypass extracting and re-inserting an element of an input
26618 // vector. Essentially:
26619 // (bitcast (sclr2vec (ext_vec_elt x))) -> (bitcast x)
26620 if (V.getOpcode() == ISD::SCALAR_TO_VECTOR &&
26621 V.getOperand(0).getOpcode() == ISD::EXTRACT_VECTOR_ELT &&
26622 V.getOperand(0).getSimpleValueType().getSizeInBits() == InputBits) {
26623 SDValue ExtractedV = V.getOperand(0);
26624 SDValue OrigV = ExtractedV.getOperand(0);
26625 if (auto *ExtractIdx = dyn_cast<ConstantSDNode>(ExtractedV.getOperand(1)))
26626 if (ExtractIdx->getZExtValue() == 0) {
26627 MVT OrigVT = OrigV.getSimpleValueType();
26628 // Extract a subvector if necessary...
26629 if (OrigVT.getSizeInBits() > OpVT.getSizeInBits()) {
26630 int Ratio = OrigVT.getSizeInBits() / OpVT.getSizeInBits();
26631 OrigVT = MVT::getVectorVT(OrigVT.getVectorElementType(),
26632 OrigVT.getVectorNumElements() / Ratio);
26633 OrigV = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, OrigVT, OrigV,
26634 DAG.getIntPtrConstant(0, DL));
26636 Op = DAG.getBitcast(OpVT, OrigV);
26637 return DAG.getNode(X86ISD::VZEXT, DL, VT, Op);
26644 SDValue X86TargetLowering::PerformDAGCombine(SDNode *N,
26645 DAGCombinerInfo &DCI) const {
26646 SelectionDAG &DAG = DCI.DAG;
26647 switch (N->getOpcode()) {
26649 case ISD::EXTRACT_VECTOR_ELT:
26650 return PerformEXTRACT_VECTOR_ELTCombine(N, DAG, DCI);
26653 case X86ISD::SHRUNKBLEND:
26654 return PerformSELECTCombine(N, DAG, DCI, Subtarget);
26655 case ISD::BITCAST: return PerformBITCASTCombine(N, DAG, Subtarget);
26656 case X86ISD::CMOV: return PerformCMOVCombine(N, DAG, DCI, Subtarget);
26657 case ISD::ADD: return PerformAddCombine(N, DAG, Subtarget);
26658 case ISD::SUB: return PerformSubCombine(N, DAG, Subtarget);
26659 case X86ISD::ADC: return PerformADCCombine(N, DAG, DCI);
26660 case ISD::MUL: return PerformMulCombine(N, DAG, DCI);
26663 case ISD::SRL: return PerformShiftCombine(N, DAG, DCI, Subtarget);
26664 case ISD::AND: return PerformAndCombine(N, DAG, DCI, Subtarget);
26665 case ISD::OR: return PerformOrCombine(N, DAG, DCI, Subtarget);
26666 case ISD::XOR: return PerformXorCombine(N, DAG, DCI, Subtarget);
26667 case ISD::LOAD: return PerformLOADCombine(N, DAG, DCI, Subtarget);
26668 case ISD::MLOAD: return PerformMLOADCombine(N, DAG, DCI, Subtarget);
26669 case ISD::STORE: return PerformSTORECombine(N, DAG, Subtarget);
26670 case ISD::MSTORE: return PerformMSTORECombine(N, DAG, Subtarget);
26671 case ISD::SINT_TO_FP: return PerformSINT_TO_FPCombine(N, DAG, Subtarget);
26672 case ISD::UINT_TO_FP: return PerformUINT_TO_FPCombine(N, DAG, Subtarget);
26673 case ISD::FADD: return PerformFADDCombine(N, DAG, Subtarget);
26674 case ISD::FSUB: return PerformFSUBCombine(N, DAG, Subtarget);
26676 case X86ISD::FOR: return PerformFORCombine(N, DAG, Subtarget);
26678 case X86ISD::FMAX: return PerformFMinFMaxCombine(N, DAG);
26679 case X86ISD::FAND: return PerformFANDCombine(N, DAG);
26680 case X86ISD::FANDN: return PerformFANDNCombine(N, DAG);
26681 case X86ISD::BT: return PerformBTCombine(N, DAG, DCI);
26682 case X86ISD::VZEXT_MOVL: return PerformVZEXT_MOVLCombine(N, DAG);
26683 case ISD::ANY_EXTEND:
26684 case ISD::ZERO_EXTEND: return PerformZExtCombine(N, DAG, DCI, Subtarget);
26685 case ISD::SIGN_EXTEND: return PerformSExtCombine(N, DAG, DCI, Subtarget);
26686 case ISD::SIGN_EXTEND_INREG:
26687 return PerformSIGN_EXTEND_INREGCombine(N, DAG, Subtarget);
26688 case ISD::SETCC: return PerformISDSETCCCombine(N, DAG, Subtarget);
26689 case X86ISD::SETCC: return PerformSETCCCombine(N, DAG, DCI, Subtarget);
26690 case X86ISD::BRCOND: return PerformBrCondCombine(N, DAG, DCI, Subtarget);
26691 case X86ISD::VZEXT: return performVZEXTCombine(N, DAG, DCI, Subtarget);
26692 case X86ISD::SHUFP: // Handle all target specific shuffles
26693 case X86ISD::PALIGNR:
26694 case X86ISD::UNPCKH:
26695 case X86ISD::UNPCKL:
26696 case X86ISD::MOVHLPS:
26697 case X86ISD::MOVLHPS:
26698 case X86ISD::PSHUFB:
26699 case X86ISD::PSHUFD:
26700 case X86ISD::PSHUFHW:
26701 case X86ISD::PSHUFLW:
26702 case X86ISD::MOVSS:
26703 case X86ISD::MOVSD:
26704 case X86ISD::VPERMILPI:
26705 case X86ISD::VPERM2X128:
26706 case ISD::VECTOR_SHUFFLE: return PerformShuffleCombine(N, DAG, DCI,Subtarget);
26707 case ISD::FMA: return PerformFMACombine(N, DAG, Subtarget);
26708 case X86ISD::INSERTPS: {
26709 if (getTargetMachine().getOptLevel() > CodeGenOpt::None)
26710 return PerformINSERTPSCombine(N, DAG, Subtarget);
26713 case X86ISD::BLENDI: return PerformBLENDICombine(N, DAG);
26719 /// isTypeDesirableForOp - Return true if the target has native support for
26720 /// the specified value type and it is 'desirable' to use the type for the
26721 /// given node type. e.g. On x86 i16 is legal, but undesirable since i16
26722 /// instruction encodings are longer and some i16 instructions are slow.
26723 bool X86TargetLowering::isTypeDesirableForOp(unsigned Opc, EVT VT) const {
26724 if (!isTypeLegal(VT))
26726 if (VT != MVT::i16)
26733 case ISD::SIGN_EXTEND:
26734 case ISD::ZERO_EXTEND:
26735 case ISD::ANY_EXTEND:
26748 /// IsDesirableToPromoteOp - This method query the target whether it is
26749 /// beneficial for dag combiner to promote the specified node. If true, it
26750 /// should return the desired promotion type by reference.
26751 bool X86TargetLowering::IsDesirableToPromoteOp(SDValue Op, EVT &PVT) const {
26752 EVT VT = Op.getValueType();
26753 if (VT != MVT::i16)
26756 bool Promote = false;
26757 bool Commute = false;
26758 switch (Op.getOpcode()) {
26761 LoadSDNode *LD = cast<LoadSDNode>(Op);
26762 // If the non-extending load has a single use and it's not live out, then it
26763 // might be folded.
26764 if (LD->getExtensionType() == ISD::NON_EXTLOAD /*&&
26765 Op.hasOneUse()*/) {
26766 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
26767 UE = Op.getNode()->use_end(); UI != UE; ++UI) {
26768 // The only case where we'd want to promote LOAD (rather then it being
26769 // promoted as an operand is when it's only use is liveout.
26770 if (UI->getOpcode() != ISD::CopyToReg)
26777 case ISD::SIGN_EXTEND:
26778 case ISD::ZERO_EXTEND:
26779 case ISD::ANY_EXTEND:
26784 SDValue N0 = Op.getOperand(0);
26785 // Look out for (store (shl (load), x)).
26786 if (MayFoldLoad(N0) && MayFoldIntoStore(Op))
26799 SDValue N0 = Op.getOperand(0);
26800 SDValue N1 = Op.getOperand(1);
26801 if (!Commute && MayFoldLoad(N1))
26803 // Avoid disabling potential load folding opportunities.
26804 if (MayFoldLoad(N0) && (!isa<ConstantSDNode>(N1) || MayFoldIntoStore(Op)))
26806 if (MayFoldLoad(N1) && (!isa<ConstantSDNode>(N0) || MayFoldIntoStore(Op)))
26816 //===----------------------------------------------------------------------===//
26817 // X86 Inline Assembly Support
26818 //===----------------------------------------------------------------------===//
26820 // Helper to match a string separated by whitespace.
26821 static bool matchAsm(StringRef S, ArrayRef<const char *> Pieces) {
26822 S = S.substr(S.find_first_not_of(" \t")); // Skip leading whitespace.
26824 for (StringRef Piece : Pieces) {
26825 if (!S.startswith(Piece)) // Check if the piece matches.
26828 S = S.substr(Piece.size());
26829 StringRef::size_type Pos = S.find_first_not_of(" \t");
26830 if (Pos == 0) // We matched a prefix.
26839 static bool clobbersFlagRegisters(const SmallVector<StringRef, 4> &AsmPieces) {
26841 if (AsmPieces.size() == 3 || AsmPieces.size() == 4) {
26842 if (std::count(AsmPieces.begin(), AsmPieces.end(), "~{cc}") &&
26843 std::count(AsmPieces.begin(), AsmPieces.end(), "~{flags}") &&
26844 std::count(AsmPieces.begin(), AsmPieces.end(), "~{fpsr}")) {
26846 if (AsmPieces.size() == 3)
26848 else if (std::count(AsmPieces.begin(), AsmPieces.end(), "~{dirflag}"))
26855 bool X86TargetLowering::ExpandInlineAsm(CallInst *CI) const {
26856 InlineAsm *IA = cast<InlineAsm>(CI->getCalledValue());
26858 std::string AsmStr = IA->getAsmString();
26860 IntegerType *Ty = dyn_cast<IntegerType>(CI->getType());
26861 if (!Ty || Ty->getBitWidth() % 16 != 0)
26864 // TODO: should remove alternatives from the asmstring: "foo {a|b}" -> "foo a"
26865 SmallVector<StringRef, 4> AsmPieces;
26866 SplitString(AsmStr, AsmPieces, ";\n");
26868 switch (AsmPieces.size()) {
26869 default: return false;
26871 // FIXME: this should verify that we are targeting a 486 or better. If not,
26872 // we will turn this bswap into something that will be lowered to logical
26873 // ops instead of emitting the bswap asm. For now, we don't support 486 or
26874 // lower so don't worry about this.
26876 if (matchAsm(AsmPieces[0], {"bswap", "$0"}) ||
26877 matchAsm(AsmPieces[0], {"bswapl", "$0"}) ||
26878 matchAsm(AsmPieces[0], {"bswapq", "$0"}) ||
26879 matchAsm(AsmPieces[0], {"bswap", "${0:q}"}) ||
26880 matchAsm(AsmPieces[0], {"bswapl", "${0:q}"}) ||
26881 matchAsm(AsmPieces[0], {"bswapq", "${0:q}"})) {
26882 // No need to check constraints, nothing other than the equivalent of
26883 // "=r,0" would be valid here.
26884 return IntrinsicLowering::LowerToByteSwap(CI);
26887 // rorw $$8, ${0:w} --> llvm.bswap.i16
26888 if (CI->getType()->isIntegerTy(16) &&
26889 IA->getConstraintString().compare(0, 5, "=r,0,") == 0 &&
26890 (matchAsm(AsmPieces[0], {"rorw", "$$8,", "${0:w}"}) ||
26891 matchAsm(AsmPieces[0], {"rolw", "$$8,", "${0:w}"}))) {
26893 StringRef ConstraintsStr = IA->getConstraintString();
26894 SplitString(StringRef(ConstraintsStr).substr(5), AsmPieces, ",");
26895 array_pod_sort(AsmPieces.begin(), AsmPieces.end());
26896 if (clobbersFlagRegisters(AsmPieces))
26897 return IntrinsicLowering::LowerToByteSwap(CI);
26901 if (CI->getType()->isIntegerTy(32) &&
26902 IA->getConstraintString().compare(0, 5, "=r,0,") == 0 &&
26903 matchAsm(AsmPieces[0], {"rorw", "$$8,", "${0:w}"}) &&
26904 matchAsm(AsmPieces[1], {"rorl", "$$16,", "$0"}) &&
26905 matchAsm(AsmPieces[2], {"rorw", "$$8,", "${0:w}"})) {
26907 StringRef ConstraintsStr = IA->getConstraintString();
26908 SplitString(StringRef(ConstraintsStr).substr(5), AsmPieces, ",");
26909 array_pod_sort(AsmPieces.begin(), AsmPieces.end());
26910 if (clobbersFlagRegisters(AsmPieces))
26911 return IntrinsicLowering::LowerToByteSwap(CI);
26914 if (CI->getType()->isIntegerTy(64)) {
26915 InlineAsm::ConstraintInfoVector Constraints = IA->ParseConstraints();
26916 if (Constraints.size() >= 2 &&
26917 Constraints[0].Codes.size() == 1 && Constraints[0].Codes[0] == "A" &&
26918 Constraints[1].Codes.size() == 1 && Constraints[1].Codes[0] == "0") {
26919 // bswap %eax / bswap %edx / xchgl %eax, %edx -> llvm.bswap.i64
26920 if (matchAsm(AsmPieces[0], {"bswap", "%eax"}) &&
26921 matchAsm(AsmPieces[1], {"bswap", "%edx"}) &&
26922 matchAsm(AsmPieces[2], {"xchgl", "%eax,", "%edx"}))
26923 return IntrinsicLowering::LowerToByteSwap(CI);
26931 /// getConstraintType - Given a constraint letter, return the type of
26932 /// constraint it is for this target.
26933 X86TargetLowering::ConstraintType
26934 X86TargetLowering::getConstraintType(StringRef Constraint) const {
26935 if (Constraint.size() == 1) {
26936 switch (Constraint[0]) {
26947 return C_RegisterClass;
26971 return TargetLowering::getConstraintType(Constraint);
26974 /// Examine constraint type and operand type and determine a weight value.
26975 /// This object must already have been set up with the operand type
26976 /// and the current alternative constraint selected.
26977 TargetLowering::ConstraintWeight
26978 X86TargetLowering::getSingleConstraintMatchWeight(
26979 AsmOperandInfo &info, const char *constraint) const {
26980 ConstraintWeight weight = CW_Invalid;
26981 Value *CallOperandVal = info.CallOperandVal;
26982 // If we don't have a value, we can't do a match,
26983 // but allow it at the lowest weight.
26984 if (!CallOperandVal)
26986 Type *type = CallOperandVal->getType();
26987 // Look at the constraint type.
26988 switch (*constraint) {
26990 weight = TargetLowering::getSingleConstraintMatchWeight(info, constraint);
27001 if (CallOperandVal->getType()->isIntegerTy())
27002 weight = CW_SpecificReg;
27007 if (type->isFloatingPointTy())
27008 weight = CW_SpecificReg;
27011 if (type->isX86_MMXTy() && Subtarget->hasMMX())
27012 weight = CW_SpecificReg;
27016 if (((type->getPrimitiveSizeInBits() == 128) && Subtarget->hasSSE1()) ||
27017 ((type->getPrimitiveSizeInBits() == 256) && Subtarget->hasFp256()))
27018 weight = CW_Register;
27021 if (ConstantInt *C = dyn_cast<ConstantInt>(info.CallOperandVal)) {
27022 if (C->getZExtValue() <= 31)
27023 weight = CW_Constant;
27027 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
27028 if (C->getZExtValue() <= 63)
27029 weight = CW_Constant;
27033 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
27034 if ((C->getSExtValue() >= -0x80) && (C->getSExtValue() <= 0x7f))
27035 weight = CW_Constant;
27039 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
27040 if ((C->getZExtValue() == 0xff) || (C->getZExtValue() == 0xffff))
27041 weight = CW_Constant;
27045 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
27046 if (C->getZExtValue() <= 3)
27047 weight = CW_Constant;
27051 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
27052 if (C->getZExtValue() <= 0xff)
27053 weight = CW_Constant;
27058 if (isa<ConstantFP>(CallOperandVal)) {
27059 weight = CW_Constant;
27063 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
27064 if ((C->getSExtValue() >= -0x80000000LL) &&
27065 (C->getSExtValue() <= 0x7fffffffLL))
27066 weight = CW_Constant;
27070 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
27071 if (C->getZExtValue() <= 0xffffffff)
27072 weight = CW_Constant;
27079 /// LowerXConstraint - try to replace an X constraint, which matches anything,
27080 /// with another that has more specific requirements based on the type of the
27081 /// corresponding operand.
27082 const char *X86TargetLowering::
27083 LowerXConstraint(EVT ConstraintVT) const {
27084 // FP X constraints get lowered to SSE1/2 registers if available, otherwise
27085 // 'f' like normal targets.
27086 if (ConstraintVT.isFloatingPoint()) {
27087 if (Subtarget->hasSSE2())
27089 if (Subtarget->hasSSE1())
27093 return TargetLowering::LowerXConstraint(ConstraintVT);
27096 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
27097 /// vector. If it is invalid, don't add anything to Ops.
27098 void X86TargetLowering::LowerAsmOperandForConstraint(SDValue Op,
27099 std::string &Constraint,
27100 std::vector<SDValue>&Ops,
27101 SelectionDAG &DAG) const {
27104 // Only support length 1 constraints for now.
27105 if (Constraint.length() > 1) return;
27107 char ConstraintLetter = Constraint[0];
27108 switch (ConstraintLetter) {
27111 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
27112 if (C->getZExtValue() <= 31) {
27113 Result = DAG.getTargetConstant(C->getZExtValue(), SDLoc(Op),
27114 Op.getValueType());
27120 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
27121 if (C->getZExtValue() <= 63) {
27122 Result = DAG.getTargetConstant(C->getZExtValue(), SDLoc(Op),
27123 Op.getValueType());
27129 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
27130 if (isInt<8>(C->getSExtValue())) {
27131 Result = DAG.getTargetConstant(C->getZExtValue(), SDLoc(Op),
27132 Op.getValueType());
27138 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
27139 if (C->getZExtValue() == 0xff || C->getZExtValue() == 0xffff ||
27140 (Subtarget->is64Bit() && C->getZExtValue() == 0xffffffff)) {
27141 Result = DAG.getTargetConstant(C->getSExtValue(), SDLoc(Op),
27142 Op.getValueType());
27148 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
27149 if (C->getZExtValue() <= 3) {
27150 Result = DAG.getTargetConstant(C->getZExtValue(), SDLoc(Op),
27151 Op.getValueType());
27157 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
27158 if (C->getZExtValue() <= 255) {
27159 Result = DAG.getTargetConstant(C->getZExtValue(), SDLoc(Op),
27160 Op.getValueType());
27166 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
27167 if (C->getZExtValue() <= 127) {
27168 Result = DAG.getTargetConstant(C->getZExtValue(), SDLoc(Op),
27169 Op.getValueType());
27175 // 32-bit signed value
27176 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
27177 if (ConstantInt::isValueValidForType(Type::getInt32Ty(*DAG.getContext()),
27178 C->getSExtValue())) {
27179 // Widen to 64 bits here to get it sign extended.
27180 Result = DAG.getTargetConstant(C->getSExtValue(), SDLoc(Op), MVT::i64);
27183 // FIXME gcc accepts some relocatable values here too, but only in certain
27184 // memory models; it's complicated.
27189 // 32-bit unsigned value
27190 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
27191 if (ConstantInt::isValueValidForType(Type::getInt32Ty(*DAG.getContext()),
27192 C->getZExtValue())) {
27193 Result = DAG.getTargetConstant(C->getZExtValue(), SDLoc(Op),
27194 Op.getValueType());
27198 // FIXME gcc accepts some relocatable values here too, but only in certain
27199 // memory models; it's complicated.
27203 // Literal immediates are always ok.
27204 if (ConstantSDNode *CST = dyn_cast<ConstantSDNode>(Op)) {
27205 // Widen to 64 bits here to get it sign extended.
27206 Result = DAG.getTargetConstant(CST->getSExtValue(), SDLoc(Op), MVT::i64);
27210 // In any sort of PIC mode addresses need to be computed at runtime by
27211 // adding in a register or some sort of table lookup. These can't
27212 // be used as immediates.
27213 if (Subtarget->isPICStyleGOT() || Subtarget->isPICStyleStubPIC())
27216 // If we are in non-pic codegen mode, we allow the address of a global (with
27217 // an optional displacement) to be used with 'i'.
27218 GlobalAddressSDNode *GA = nullptr;
27219 int64_t Offset = 0;
27221 // Match either (GA), (GA+C), (GA+C1+C2), etc.
27223 if ((GA = dyn_cast<GlobalAddressSDNode>(Op))) {
27224 Offset += GA->getOffset();
27226 } else if (Op.getOpcode() == ISD::ADD) {
27227 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
27228 Offset += C->getZExtValue();
27229 Op = Op.getOperand(0);
27232 } else if (Op.getOpcode() == ISD::SUB) {
27233 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
27234 Offset += -C->getZExtValue();
27235 Op = Op.getOperand(0);
27240 // Otherwise, this isn't something we can handle, reject it.
27244 const GlobalValue *GV = GA->getGlobal();
27245 // If we require an extra load to get this address, as in PIC mode, we
27246 // can't accept it.
27247 if (isGlobalStubReference(
27248 Subtarget->ClassifyGlobalReference(GV, DAG.getTarget())))
27251 Result = DAG.getTargetGlobalAddress(GV, SDLoc(Op),
27252 GA->getValueType(0), Offset);
27257 if (Result.getNode()) {
27258 Ops.push_back(Result);
27261 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG);
27264 std::pair<unsigned, const TargetRegisterClass *>
27265 X86TargetLowering::getRegForInlineAsmConstraint(const TargetRegisterInfo *TRI,
27266 StringRef Constraint,
27268 // First, see if this is a constraint that directly corresponds to an LLVM
27270 if (Constraint.size() == 1) {
27271 // GCC Constraint Letters
27272 switch (Constraint[0]) {
27274 // TODO: Slight differences here in allocation order and leaving
27275 // RIP in the class. Do they matter any more here than they do
27276 // in the normal allocation?
27277 case 'q': // GENERAL_REGS in 64-bit mode, Q_REGS in 32-bit mode.
27278 if (Subtarget->is64Bit()) {
27279 if (VT == MVT::i32 || VT == MVT::f32)
27280 return std::make_pair(0U, &X86::GR32RegClass);
27281 if (VT == MVT::i16)
27282 return std::make_pair(0U, &X86::GR16RegClass);
27283 if (VT == MVT::i8 || VT == MVT::i1)
27284 return std::make_pair(0U, &X86::GR8RegClass);
27285 if (VT == MVT::i64 || VT == MVT::f64)
27286 return std::make_pair(0U, &X86::GR64RegClass);
27289 // 32-bit fallthrough
27290 case 'Q': // Q_REGS
27291 if (VT == MVT::i32 || VT == MVT::f32)
27292 return std::make_pair(0U, &X86::GR32_ABCDRegClass);
27293 if (VT == MVT::i16)
27294 return std::make_pair(0U, &X86::GR16_ABCDRegClass);
27295 if (VT == MVT::i8 || VT == MVT::i1)
27296 return std::make_pair(0U, &X86::GR8_ABCD_LRegClass);
27297 if (VT == MVT::i64)
27298 return std::make_pair(0U, &X86::GR64_ABCDRegClass);
27300 case 'r': // GENERAL_REGS
27301 case 'l': // INDEX_REGS
27302 if (VT == MVT::i8 || VT == MVT::i1)
27303 return std::make_pair(0U, &X86::GR8RegClass);
27304 if (VT == MVT::i16)
27305 return std::make_pair(0U, &X86::GR16RegClass);
27306 if (VT == MVT::i32 || VT == MVT::f32 || !Subtarget->is64Bit())
27307 return std::make_pair(0U, &X86::GR32RegClass);
27308 return std::make_pair(0U, &X86::GR64RegClass);
27309 case 'R': // LEGACY_REGS
27310 if (VT == MVT::i8 || VT == MVT::i1)
27311 return std::make_pair(0U, &X86::GR8_NOREXRegClass);
27312 if (VT == MVT::i16)
27313 return std::make_pair(0U, &X86::GR16_NOREXRegClass);
27314 if (VT == MVT::i32 || !Subtarget->is64Bit())
27315 return std::make_pair(0U, &X86::GR32_NOREXRegClass);
27316 return std::make_pair(0U, &X86::GR64_NOREXRegClass);
27317 case 'f': // FP Stack registers.
27318 // If SSE is enabled for this VT, use f80 to ensure the isel moves the
27319 // value to the correct fpstack register class.
27320 if (VT == MVT::f32 && !isScalarFPTypeInSSEReg(VT))
27321 return std::make_pair(0U, &X86::RFP32RegClass);
27322 if (VT == MVT::f64 && !isScalarFPTypeInSSEReg(VT))
27323 return std::make_pair(0U, &X86::RFP64RegClass);
27324 return std::make_pair(0U, &X86::RFP80RegClass);
27325 case 'y': // MMX_REGS if MMX allowed.
27326 if (!Subtarget->hasMMX()) break;
27327 return std::make_pair(0U, &X86::VR64RegClass);
27328 case 'Y': // SSE_REGS if SSE2 allowed
27329 if (!Subtarget->hasSSE2()) break;
27331 case 'x': // SSE_REGS if SSE1 allowed or AVX_REGS if AVX allowed
27332 if (!Subtarget->hasSSE1()) break;
27334 switch (VT.SimpleTy) {
27336 // Scalar SSE types.
27339 return std::make_pair(0U, &X86::FR32RegClass);
27342 return std::make_pair(0U, &X86::FR64RegClass);
27350 return std::make_pair(0U, &X86::VR128RegClass);
27358 return std::make_pair(0U, &X86::VR256RegClass);
27363 return std::make_pair(0U, &X86::VR512RegClass);
27369 // Use the default implementation in TargetLowering to convert the register
27370 // constraint into a member of a register class.
27371 std::pair<unsigned, const TargetRegisterClass*> Res;
27372 Res = TargetLowering::getRegForInlineAsmConstraint(TRI, Constraint, VT);
27374 // Not found as a standard register?
27376 // Map st(0) -> st(7) -> ST0
27377 if (Constraint.size() == 7 && Constraint[0] == '{' &&
27378 tolower(Constraint[1]) == 's' &&
27379 tolower(Constraint[2]) == 't' &&
27380 Constraint[3] == '(' &&
27381 (Constraint[4] >= '0' && Constraint[4] <= '7') &&
27382 Constraint[5] == ')' &&
27383 Constraint[6] == '}') {
27385 Res.first = X86::FP0+Constraint[4]-'0';
27386 Res.second = &X86::RFP80RegClass;
27390 // GCC allows "st(0)" to be called just plain "st".
27391 if (StringRef("{st}").equals_lower(Constraint)) {
27392 Res.first = X86::FP0;
27393 Res.second = &X86::RFP80RegClass;
27398 if (StringRef("{flags}").equals_lower(Constraint)) {
27399 Res.first = X86::EFLAGS;
27400 Res.second = &X86::CCRRegClass;
27404 // 'A' means EAX + EDX.
27405 if (Constraint == "A") {
27406 Res.first = X86::EAX;
27407 Res.second = &X86::GR32_ADRegClass;
27413 // Otherwise, check to see if this is a register class of the wrong value
27414 // type. For example, we want to map "{ax},i32" -> {eax}, we don't want it to
27415 // turn into {ax},{dx}.
27416 // MVT::Other is used to specify clobber names.
27417 if (Res.second->hasType(VT) || VT == MVT::Other)
27418 return Res; // Correct type already, nothing to do.
27420 // Get a matching integer of the correct size. i.e. "ax" with MVT::32 should
27421 // return "eax". This should even work for things like getting 64bit integer
27422 // registers when given an f64 type.
27423 const TargetRegisterClass *Class = Res.second;
27424 if (Class == &X86::GR8RegClass || Class == &X86::GR16RegClass ||
27425 Class == &X86::GR32RegClass || Class == &X86::GR64RegClass) {
27426 unsigned Size = VT.getSizeInBits();
27427 MVT::SimpleValueType SimpleTy = Size == 1 || Size == 8 ? MVT::i8
27428 : Size == 16 ? MVT::i16
27429 : Size == 32 ? MVT::i32
27430 : Size == 64 ? MVT::i64
27432 unsigned DestReg = getX86SubSuperRegisterOrZero(Res.first, SimpleTy);
27434 Res.first = DestReg;
27435 Res.second = SimpleTy == MVT::i8 ? &X86::GR8RegClass
27436 : SimpleTy == MVT::i16 ? &X86::GR16RegClass
27437 : SimpleTy == MVT::i32 ? &X86::GR32RegClass
27438 : &X86::GR64RegClass;
27439 assert(Res.second->contains(Res.first) && "Register in register class");
27441 // No register found/type mismatch.
27443 Res.second = nullptr;
27445 } else if (Class == &X86::FR32RegClass || Class == &X86::FR64RegClass ||
27446 Class == &X86::VR128RegClass || Class == &X86::VR256RegClass ||
27447 Class == &X86::FR32XRegClass || Class == &X86::FR64XRegClass ||
27448 Class == &X86::VR128XRegClass || Class == &X86::VR256XRegClass ||
27449 Class == &X86::VR512RegClass) {
27450 // Handle references to XMM physical registers that got mapped into the
27451 // wrong class. This can happen with constraints like {xmm0} where the
27452 // target independent register mapper will just pick the first match it can
27453 // find, ignoring the required type.
27455 if (VT == MVT::f32 || VT == MVT::i32)
27456 Res.second = &X86::FR32RegClass;
27457 else if (VT == MVT::f64 || VT == MVT::i64)
27458 Res.second = &X86::FR64RegClass;
27459 else if (X86::VR128RegClass.hasType(VT))
27460 Res.second = &X86::VR128RegClass;
27461 else if (X86::VR256RegClass.hasType(VT))
27462 Res.second = &X86::VR256RegClass;
27463 else if (X86::VR512RegClass.hasType(VT))
27464 Res.second = &X86::VR512RegClass;
27466 // Type mismatch and not a clobber: Return an error;
27468 Res.second = nullptr;
27475 int X86TargetLowering::getScalingFactorCost(const DataLayout &DL,
27476 const AddrMode &AM, Type *Ty,
27477 unsigned AS) const {
27478 // Scaling factors are not free at all.
27479 // An indexed folded instruction, i.e., inst (reg1, reg2, scale),
27480 // will take 2 allocations in the out of order engine instead of 1
27481 // for plain addressing mode, i.e. inst (reg1).
27483 // vaddps (%rsi,%drx), %ymm0, %ymm1
27484 // Requires two allocations (one for the load, one for the computation)
27486 // vaddps (%rsi), %ymm0, %ymm1
27487 // Requires just 1 allocation, i.e., freeing allocations for other operations
27488 // and having less micro operations to execute.
27490 // For some X86 architectures, this is even worse because for instance for
27491 // stores, the complex addressing mode forces the instruction to use the
27492 // "load" ports instead of the dedicated "store" port.
27493 // E.g., on Haswell:
27494 // vmovaps %ymm1, (%r8, %rdi) can use port 2 or 3.
27495 // vmovaps %ymm1, (%r8) can use port 2, 3, or 7.
27496 if (isLegalAddressingMode(DL, AM, Ty, AS))
27497 // Scale represents reg2 * scale, thus account for 1
27498 // as soon as we use a second register.
27499 return AM.Scale != 0;
27503 bool X86TargetLowering::isIntDivCheap(EVT VT, AttributeSet Attr) const {
27504 // Integer division on x86 is expensive. However, when aggressively optimizing
27505 // for code size, we prefer to use a div instruction, as it is usually smaller
27506 // than the alternative sequence.
27507 // The exception to this is vector division. Since x86 doesn't have vector
27508 // integer division, leaving the division as-is is a loss even in terms of
27509 // size, because it will have to be scalarized, while the alternative code
27510 // sequence can be performed in vector form.
27511 bool OptSize = Attr.hasAttribute(AttributeSet::FunctionIndex,
27512 Attribute::MinSize);
27513 return OptSize && !VT.isVector();
27516 void X86TargetLowering::markInRegArguments(SelectionDAG &DAG,
27517 TargetLowering::ArgListTy& Args) const {
27518 // The MCU psABI requires some arguments to be passed in-register.
27519 // For regular calls, the inreg arguments are marked by the front-end.
27520 // However, for compiler generated library calls, we have to patch this
27522 if (!Subtarget->isTargetMCU() || !Args.size())
27525 unsigned FreeRegs = 3;
27526 for (auto &Arg : Args) {
27527 // For library functions, we do not expect any fancy types.
27528 unsigned Size = DAG.getDataLayout().getTypeSizeInBits(Arg.Ty);
27529 unsigned SizeInRegs = (Size + 31) / 32;
27530 if (SizeInRegs > 2 || SizeInRegs > FreeRegs)
27533 Arg.isInReg = true;
27534 FreeRegs -= SizeInRegs;