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 in fact a blend.
6863 static SDValue lowerVectorShuffleAsBlend(SDLoc DL, MVT VT, SDValue V1,
6864 SDValue V2, ArrayRef<int> Mask,
6865 const X86Subtarget *Subtarget,
6866 SelectionDAG &DAG) {
6867 unsigned BlendMask = 0;
6868 for (int i = 0, Size = Mask.size(); i < Size; ++i) {
6869 if (Mask[i] >= Size) {
6870 if (Mask[i] != i + Size)
6871 return SDValue(); // Shuffled V2 input!
6872 BlendMask |= 1u << i;
6875 if (Mask[i] >= 0 && Mask[i] != i)
6876 return SDValue(); // Shuffled V1 input!
6878 switch (VT.SimpleTy) {
6883 return DAG.getNode(X86ISD::BLENDI, DL, VT, V1, V2,
6884 DAG.getConstant(BlendMask, DL, MVT::i8));
6888 assert(Subtarget->hasAVX2() && "256-bit integer blends require AVX2!");
6892 // If we have AVX2 it is faster to use VPBLENDD when the shuffle fits into
6893 // that instruction.
6894 if (Subtarget->hasAVX2()) {
6895 // Scale the blend by the number of 32-bit dwords per element.
6896 int Scale = VT.getScalarSizeInBits() / 32;
6898 for (int i = 0, Size = Mask.size(); i < Size; ++i)
6899 if (Mask[i] >= Size)
6900 for (int j = 0; j < Scale; ++j)
6901 BlendMask |= 1u << (i * Scale + j);
6903 MVT BlendVT = VT.getSizeInBits() > 128 ? MVT::v8i32 : MVT::v4i32;
6904 V1 = DAG.getBitcast(BlendVT, V1);
6905 V2 = DAG.getBitcast(BlendVT, V2);
6906 return DAG.getBitcast(
6907 VT, DAG.getNode(X86ISD::BLENDI, DL, BlendVT, V1, V2,
6908 DAG.getConstant(BlendMask, DL, MVT::i8)));
6912 // For integer shuffles we need to expand the mask and cast the inputs to
6913 // v8i16s prior to blending.
6914 int Scale = 8 / VT.getVectorNumElements();
6916 for (int i = 0, Size = Mask.size(); i < Size; ++i)
6917 if (Mask[i] >= Size)
6918 for (int j = 0; j < Scale; ++j)
6919 BlendMask |= 1u << (i * Scale + j);
6921 V1 = DAG.getBitcast(MVT::v8i16, V1);
6922 V2 = DAG.getBitcast(MVT::v8i16, V2);
6923 return DAG.getBitcast(VT,
6924 DAG.getNode(X86ISD::BLENDI, DL, MVT::v8i16, V1, V2,
6925 DAG.getConstant(BlendMask, DL, MVT::i8)));
6929 assert(Subtarget->hasAVX2() && "256-bit integer blends require AVX2!");
6930 SmallVector<int, 8> RepeatedMask;
6931 if (is128BitLaneRepeatedShuffleMask(MVT::v16i16, Mask, RepeatedMask)) {
6932 // We can lower these with PBLENDW which is mirrored across 128-bit lanes.
6933 assert(RepeatedMask.size() == 8 && "Repeated mask size doesn't match!");
6935 for (int i = 0; i < 8; ++i)
6936 if (RepeatedMask[i] >= 16)
6937 BlendMask |= 1u << i;
6938 return DAG.getNode(X86ISD::BLENDI, DL, MVT::v16i16, V1, V2,
6939 DAG.getConstant(BlendMask, DL, MVT::i8));
6945 assert((VT.getSizeInBits() == 128 || Subtarget->hasAVX2()) &&
6946 "256-bit byte-blends require AVX2 support!");
6948 // Attempt to lower to a bitmask if we can. VPAND is faster than VPBLENDVB.
6949 if (SDValue Masked = lowerVectorShuffleAsBitMask(DL, VT, V1, V2, Mask, DAG))
6952 // Scale the blend by the number of bytes per element.
6953 int Scale = VT.getScalarSizeInBits() / 8;
6955 // This form of blend is always done on bytes. Compute the byte vector
6957 MVT BlendVT = MVT::getVectorVT(MVT::i8, VT.getSizeInBits() / 8);
6959 // Compute the VSELECT mask. Note that VSELECT is really confusing in the
6960 // mix of LLVM's code generator and the x86 backend. We tell the code
6961 // generator that boolean values in the elements of an x86 vector register
6962 // are -1 for true and 0 for false. We then use the LLVM semantics of 'true'
6963 // mapping a select to operand #1, and 'false' mapping to operand #2. The
6964 // reality in x86 is that vector masks (pre-AVX-512) use only the high bit
6965 // of the element (the remaining are ignored) and 0 in that high bit would
6966 // mean operand #1 while 1 in the high bit would mean operand #2. So while
6967 // the LLVM model for boolean values in vector elements gets the relevant
6968 // bit set, it is set backwards and over constrained relative to x86's
6970 SmallVector<SDValue, 32> VSELECTMask;
6971 for (int i = 0, Size = Mask.size(); i < Size; ++i)
6972 for (int j = 0; j < Scale; ++j)
6973 VSELECTMask.push_back(
6974 Mask[i] < 0 ? DAG.getUNDEF(MVT::i8)
6975 : DAG.getConstant(Mask[i] < Size ? -1 : 0, DL,
6978 V1 = DAG.getBitcast(BlendVT, V1);
6979 V2 = DAG.getBitcast(BlendVT, V2);
6980 return DAG.getBitcast(VT, DAG.getNode(ISD::VSELECT, DL, BlendVT,
6981 DAG.getNode(ISD::BUILD_VECTOR, DL,
6982 BlendVT, VSELECTMask),
6987 llvm_unreachable("Not a supported integer vector type!");
6991 /// \brief Try to lower as a blend of elements from two inputs followed by
6992 /// a single-input permutation.
6994 /// This matches the pattern where we can blend elements from two inputs and
6995 /// then reduce the shuffle to a single-input permutation.
6996 static SDValue lowerVectorShuffleAsBlendAndPermute(SDLoc DL, MVT VT, SDValue V1,
6999 SelectionDAG &DAG) {
7000 // We build up the blend mask while checking whether a blend is a viable way
7001 // to reduce the shuffle.
7002 SmallVector<int, 32> BlendMask(Mask.size(), -1);
7003 SmallVector<int, 32> PermuteMask(Mask.size(), -1);
7005 for (int i = 0, Size = Mask.size(); i < Size; ++i) {
7009 assert(Mask[i] < Size * 2 && "Shuffle input is out of bounds.");
7011 if (BlendMask[Mask[i] % Size] == -1)
7012 BlendMask[Mask[i] % Size] = Mask[i];
7013 else if (BlendMask[Mask[i] % Size] != Mask[i])
7014 return SDValue(); // Can't blend in the needed input!
7016 PermuteMask[i] = Mask[i] % Size;
7019 SDValue V = DAG.getVectorShuffle(VT, DL, V1, V2, BlendMask);
7020 return DAG.getVectorShuffle(VT, DL, V, DAG.getUNDEF(VT), PermuteMask);
7023 /// \brief Generic routine to decompose a shuffle and blend into indepndent
7024 /// blends and permutes.
7026 /// This matches the extremely common pattern for handling combined
7027 /// shuffle+blend operations on newer X86 ISAs where we have very fast blend
7028 /// operations. It will try to pick the best arrangement of shuffles and
7030 static SDValue lowerVectorShuffleAsDecomposedShuffleBlend(SDLoc DL, MVT VT,
7034 SelectionDAG &DAG) {
7035 // Shuffle the input elements into the desired positions in V1 and V2 and
7036 // blend them together.
7037 SmallVector<int, 32> V1Mask(Mask.size(), -1);
7038 SmallVector<int, 32> V2Mask(Mask.size(), -1);
7039 SmallVector<int, 32> BlendMask(Mask.size(), -1);
7040 for (int i = 0, Size = Mask.size(); i < Size; ++i)
7041 if (Mask[i] >= 0 && Mask[i] < Size) {
7042 V1Mask[i] = Mask[i];
7044 } else if (Mask[i] >= Size) {
7045 V2Mask[i] = Mask[i] - Size;
7046 BlendMask[i] = i + Size;
7049 // Try to lower with the simpler initial blend strategy unless one of the
7050 // input shuffles would be a no-op. We prefer to shuffle inputs as the
7051 // shuffle may be able to fold with a load or other benefit. However, when
7052 // we'll have to do 2x as many shuffles in order to achieve this, blending
7053 // first is a better strategy.
7054 if (!isNoopShuffleMask(V1Mask) && !isNoopShuffleMask(V2Mask))
7055 if (SDValue BlendPerm =
7056 lowerVectorShuffleAsBlendAndPermute(DL, VT, V1, V2, Mask, DAG))
7059 V1 = DAG.getVectorShuffle(VT, DL, V1, DAG.getUNDEF(VT), V1Mask);
7060 V2 = DAG.getVectorShuffle(VT, DL, V2, DAG.getUNDEF(VT), V2Mask);
7061 return DAG.getVectorShuffle(VT, DL, V1, V2, BlendMask);
7064 /// \brief Try to lower a vector shuffle as a byte rotation.
7066 /// SSSE3 has a generic PALIGNR instruction in x86 that will do an arbitrary
7067 /// byte-rotation of the concatenation of two vectors; pre-SSSE3 can use
7068 /// a PSRLDQ/PSLLDQ/POR pattern to get a similar effect. This routine will
7069 /// try to generically lower a vector shuffle through such an pattern. It
7070 /// does not check for the profitability of lowering either as PALIGNR or
7071 /// PSRLDQ/PSLLDQ/POR, only whether the mask is valid to lower in that form.
7072 /// This matches shuffle vectors that look like:
7074 /// v8i16 [11, 12, 13, 14, 15, 0, 1, 2]
7076 /// Essentially it concatenates V1 and V2, shifts right by some number of
7077 /// elements, and takes the low elements as the result. Note that while this is
7078 /// specified as a *right shift* because x86 is little-endian, it is a *left
7079 /// rotate* of the vector lanes.
7080 static SDValue lowerVectorShuffleAsByteRotate(SDLoc DL, MVT VT, SDValue V1,
7083 const X86Subtarget *Subtarget,
7084 SelectionDAG &DAG) {
7085 assert(!isNoopShuffleMask(Mask) && "We shouldn't lower no-op shuffles!");
7087 int NumElts = Mask.size();
7088 int NumLanes = VT.getSizeInBits() / 128;
7089 int NumLaneElts = NumElts / NumLanes;
7091 // We need to detect various ways of spelling a rotation:
7092 // [11, 12, 13, 14, 15, 0, 1, 2]
7093 // [-1, 12, 13, 14, -1, -1, 1, -1]
7094 // [-1, -1, -1, -1, -1, -1, 1, 2]
7095 // [ 3, 4, 5, 6, 7, 8, 9, 10]
7096 // [-1, 4, 5, 6, -1, -1, 9, -1]
7097 // [-1, 4, 5, 6, -1, -1, -1, -1]
7100 for (int l = 0; l < NumElts; l += NumLaneElts) {
7101 for (int i = 0; i < NumLaneElts; ++i) {
7102 if (Mask[l + i] == -1)
7104 assert(Mask[l + i] >= 0 && "Only -1 is a valid negative mask element!");
7106 // Get the mod-Size index and lane correct it.
7107 int LaneIdx = (Mask[l + i] % NumElts) - l;
7108 // Make sure it was in this lane.
7109 if (LaneIdx < 0 || LaneIdx >= NumLaneElts)
7112 // Determine where a rotated vector would have started.
7113 int StartIdx = i - LaneIdx;
7115 // The identity rotation isn't interesting, stop.
7118 // If we found the tail of a vector the rotation must be the missing
7119 // front. If we found the head of a vector, it must be how much of the
7121 int CandidateRotation = StartIdx < 0 ? -StartIdx : NumLaneElts - StartIdx;
7124 Rotation = CandidateRotation;
7125 else if (Rotation != CandidateRotation)
7126 // The rotations don't match, so we can't match this mask.
7129 // Compute which value this mask is pointing at.
7130 SDValue MaskV = Mask[l + i] < NumElts ? V1 : V2;
7132 // Compute which of the two target values this index should be assigned
7133 // to. This reflects whether the high elements are remaining or the low
7134 // elements are remaining.
7135 SDValue &TargetV = StartIdx < 0 ? Hi : Lo;
7137 // Either set up this value if we've not encountered it before, or check
7138 // that it remains consistent.
7141 else if (TargetV != MaskV)
7142 // This may be a rotation, but it pulls from the inputs in some
7143 // unsupported interleaving.
7148 // Check that we successfully analyzed the mask, and normalize the results.
7149 assert(Rotation != 0 && "Failed to locate a viable rotation!");
7150 assert((Lo || Hi) && "Failed to find a rotated input vector!");
7156 // The actual rotate instruction rotates bytes, so we need to scale the
7157 // rotation based on how many bytes are in the vector lane.
7158 int Scale = 16 / NumLaneElts;
7160 // SSSE3 targets can use the palignr instruction.
7161 if (Subtarget->hasSSSE3()) {
7162 // Cast the inputs to i8 vector of correct length to match PALIGNR.
7163 MVT AlignVT = MVT::getVectorVT(MVT::i8, 16 * NumLanes);
7164 Lo = DAG.getBitcast(AlignVT, Lo);
7165 Hi = DAG.getBitcast(AlignVT, Hi);
7167 return DAG.getBitcast(
7168 VT, DAG.getNode(X86ISD::PALIGNR, DL, AlignVT, Lo, Hi,
7169 DAG.getConstant(Rotation * Scale, DL, MVT::i8)));
7172 assert(VT.getSizeInBits() == 128 &&
7173 "Rotate-based lowering only supports 128-bit lowering!");
7174 assert(Mask.size() <= 16 &&
7175 "Can shuffle at most 16 bytes in a 128-bit vector!");
7177 // Default SSE2 implementation
7178 int LoByteShift = 16 - Rotation * Scale;
7179 int HiByteShift = Rotation * Scale;
7181 // Cast the inputs to v2i64 to match PSLLDQ/PSRLDQ.
7182 Lo = DAG.getBitcast(MVT::v2i64, Lo);
7183 Hi = DAG.getBitcast(MVT::v2i64, Hi);
7185 SDValue LoShift = DAG.getNode(X86ISD::VSHLDQ, DL, MVT::v2i64, Lo,
7186 DAG.getConstant(LoByteShift, DL, MVT::i8));
7187 SDValue HiShift = DAG.getNode(X86ISD::VSRLDQ, DL, MVT::v2i64, Hi,
7188 DAG.getConstant(HiByteShift, DL, MVT::i8));
7189 return DAG.getBitcast(VT,
7190 DAG.getNode(ISD::OR, DL, MVT::v2i64, LoShift, HiShift));
7193 /// \brief Try to lower a vector shuffle as a bit shift (shifts in zeros).
7195 /// Attempts to match a shuffle mask against the PSLL(W/D/Q/DQ) and
7196 /// PSRL(W/D/Q/DQ) SSE2 and AVX2 logical bit-shift instructions. The function
7197 /// matches elements from one of the input vectors shuffled to the left or
7198 /// right with zeroable elements 'shifted in'. It handles both the strictly
7199 /// bit-wise element shifts and the byte shift across an entire 128-bit double
7202 /// PSHL : (little-endian) left bit shift.
7203 /// [ zz, 0, zz, 2 ]
7204 /// [ -1, 4, zz, -1 ]
7205 /// PSRL : (little-endian) right bit shift.
7207 /// [ -1, -1, 7, zz]
7208 /// PSLLDQ : (little-endian) left byte shift
7209 /// [ zz, 0, 1, 2, 3, 4, 5, 6]
7210 /// [ zz, zz, -1, -1, 2, 3, 4, -1]
7211 /// [ zz, zz, zz, zz, zz, zz, -1, 1]
7212 /// PSRLDQ : (little-endian) right byte shift
7213 /// [ 5, 6, 7, zz, zz, zz, zz, zz]
7214 /// [ -1, 5, 6, 7, zz, zz, zz, zz]
7215 /// [ 1, 2, -1, -1, -1, -1, zz, zz]
7216 static SDValue lowerVectorShuffleAsShift(SDLoc DL, MVT VT, SDValue V1,
7217 SDValue V2, ArrayRef<int> Mask,
7218 SelectionDAG &DAG) {
7219 SmallBitVector Zeroable = computeZeroableShuffleElements(Mask, V1, V2);
7221 int Size = Mask.size();
7222 assert(Size == (int)VT.getVectorNumElements() && "Unexpected mask size");
7224 auto CheckZeros = [&](int Shift, int Scale, bool Left) {
7225 for (int i = 0; i < Size; i += Scale)
7226 for (int j = 0; j < Shift; ++j)
7227 if (!Zeroable[i + j + (Left ? 0 : (Scale - Shift))])
7233 auto MatchShift = [&](int Shift, int Scale, bool Left, SDValue V) {
7234 for (int i = 0; i != Size; i += Scale) {
7235 unsigned Pos = Left ? i + Shift : i;
7236 unsigned Low = Left ? i : i + Shift;
7237 unsigned Len = Scale - Shift;
7238 if (!isSequentialOrUndefInRange(Mask, Pos, Len,
7239 Low + (V == V1 ? 0 : Size)))
7243 int ShiftEltBits = VT.getScalarSizeInBits() * Scale;
7244 bool ByteShift = ShiftEltBits > 64;
7245 unsigned OpCode = Left ? (ByteShift ? X86ISD::VSHLDQ : X86ISD::VSHLI)
7246 : (ByteShift ? X86ISD::VSRLDQ : X86ISD::VSRLI);
7247 int ShiftAmt = Shift * VT.getScalarSizeInBits() / (ByteShift ? 8 : 1);
7249 // Normalize the scale for byte shifts to still produce an i64 element
7251 Scale = ByteShift ? Scale / 2 : Scale;
7253 // We need to round trip through the appropriate type for the shift.
7254 MVT ShiftSVT = MVT::getIntegerVT(VT.getScalarSizeInBits() * Scale);
7255 MVT ShiftVT = MVT::getVectorVT(ShiftSVT, Size / Scale);
7256 assert(DAG.getTargetLoweringInfo().isTypeLegal(ShiftVT) &&
7257 "Illegal integer vector type");
7258 V = DAG.getBitcast(ShiftVT, V);
7260 V = DAG.getNode(OpCode, DL, ShiftVT, V,
7261 DAG.getConstant(ShiftAmt, DL, MVT::i8));
7262 return DAG.getBitcast(VT, V);
7265 // SSE/AVX supports logical shifts up to 64-bit integers - so we can just
7266 // keep doubling the size of the integer elements up to that. We can
7267 // then shift the elements of the integer vector by whole multiples of
7268 // their width within the elements of the larger integer vector. Test each
7269 // multiple to see if we can find a match with the moved element indices
7270 // and that the shifted in elements are all zeroable.
7271 for (int Scale = 2; Scale * VT.getScalarSizeInBits() <= 128; Scale *= 2)
7272 for (int Shift = 1; Shift != Scale; ++Shift)
7273 for (bool Left : {true, false})
7274 if (CheckZeros(Shift, Scale, Left))
7275 for (SDValue V : {V1, V2})
7276 if (SDValue Match = MatchShift(Shift, Scale, Left, V))
7283 /// \brief Try to lower a vector shuffle using SSE4a EXTRQ/INSERTQ.
7284 static SDValue lowerVectorShuffleWithSSE4A(SDLoc DL, MVT VT, SDValue V1,
7285 SDValue V2, ArrayRef<int> Mask,
7286 SelectionDAG &DAG) {
7287 SmallBitVector Zeroable = computeZeroableShuffleElements(Mask, V1, V2);
7288 assert(!Zeroable.all() && "Fully zeroable shuffle mask");
7290 int Size = Mask.size();
7291 int HalfSize = Size / 2;
7292 assert(Size == (int)VT.getVectorNumElements() && "Unexpected mask size");
7294 // Upper half must be undefined.
7295 if (!isUndefInRange(Mask, HalfSize, HalfSize))
7298 // EXTRQ: Extract Len elements from lower half of source, starting at Idx.
7299 // Remainder of lower half result is zero and upper half is all undef.
7300 auto LowerAsEXTRQ = [&]() {
7301 // Determine the extraction length from the part of the
7302 // lower half that isn't zeroable.
7304 for (; Len > 0; --Len)
7305 if (!Zeroable[Len - 1])
7307 assert(Len > 0 && "Zeroable shuffle mask");
7309 // Attempt to match first Len sequential elements from the lower half.
7312 for (int i = 0; i != Len; ++i) {
7316 SDValue &V = (M < Size ? V1 : V2);
7319 // All mask elements must be in the lower half.
7323 if (Idx < 0 || (Src == V && Idx == (M - i))) {
7334 assert((Idx + Len) <= HalfSize && "Illegal extraction mask");
7335 int BitLen = (Len * VT.getScalarSizeInBits()) & 0x3f;
7336 int BitIdx = (Idx * VT.getScalarSizeInBits()) & 0x3f;
7337 return DAG.getNode(X86ISD::EXTRQI, DL, VT, Src,
7338 DAG.getConstant(BitLen, DL, MVT::i8),
7339 DAG.getConstant(BitIdx, DL, MVT::i8));
7342 if (SDValue ExtrQ = LowerAsEXTRQ())
7345 // INSERTQ: Extract lowest Len elements from lower half of second source and
7346 // insert over first source, starting at Idx.
7347 // { A[0], .., A[Idx-1], B[0], .., B[Len-1], A[Idx+Len], .., UNDEF, ... }
7348 auto LowerAsInsertQ = [&]() {
7349 for (int Idx = 0; Idx != HalfSize; ++Idx) {
7352 // Attempt to match first source from mask before insertion point.
7353 if (isUndefInRange(Mask, 0, Idx)) {
7355 } else if (isSequentialOrUndefInRange(Mask, 0, Idx, 0)) {
7357 } else if (isSequentialOrUndefInRange(Mask, 0, Idx, Size)) {
7363 // Extend the extraction length looking to match both the insertion of
7364 // the second source and the remaining elements of the first.
7365 for (int Hi = Idx + 1; Hi <= HalfSize; ++Hi) {
7370 if (isSequentialOrUndefInRange(Mask, Idx, Len, 0)) {
7372 } else if (isSequentialOrUndefInRange(Mask, Idx, Len, Size)) {
7378 // Match the remaining elements of the lower half.
7379 if (isUndefInRange(Mask, Hi, HalfSize - Hi)) {
7381 } else if ((!Base || (Base == V1)) &&
7382 isSequentialOrUndefInRange(Mask, Hi, HalfSize - Hi, Hi)) {
7384 } else if ((!Base || (Base == V2)) &&
7385 isSequentialOrUndefInRange(Mask, Hi, HalfSize - Hi,
7392 // We may not have a base (first source) - this can safely be undefined.
7394 Base = DAG.getUNDEF(VT);
7396 int BitLen = (Len * VT.getScalarSizeInBits()) & 0x3f;
7397 int BitIdx = (Idx * VT.getScalarSizeInBits()) & 0x3f;
7398 return DAG.getNode(X86ISD::INSERTQI, DL, VT, Base, Insert,
7399 DAG.getConstant(BitLen, DL, MVT::i8),
7400 DAG.getConstant(BitIdx, DL, MVT::i8));
7407 if (SDValue InsertQ = LowerAsInsertQ())
7413 /// \brief Lower a vector shuffle as a zero or any extension.
7415 /// Given a specific number of elements, element bit width, and extension
7416 /// stride, produce either a zero or any extension based on the available
7417 /// features of the subtarget. The extended elements are consecutive and
7418 /// begin and can start from an offseted element index in the input; to
7419 /// avoid excess shuffling the offset must either being in the bottom lane
7420 /// or at the start of a higher lane. All extended elements must be from
7422 static SDValue lowerVectorShuffleAsSpecificZeroOrAnyExtend(
7423 SDLoc DL, MVT VT, int Scale, int Offset, bool AnyExt, SDValue InputV,
7424 ArrayRef<int> Mask, const X86Subtarget *Subtarget, SelectionDAG &DAG) {
7425 assert(Scale > 1 && "Need a scale to extend.");
7426 int EltBits = VT.getScalarSizeInBits();
7427 int NumElements = VT.getVectorNumElements();
7428 int NumEltsPerLane = 128 / EltBits;
7429 int OffsetLane = Offset / NumEltsPerLane;
7430 assert((EltBits == 8 || EltBits == 16 || EltBits == 32) &&
7431 "Only 8, 16, and 32 bit elements can be extended.");
7432 assert(Scale * EltBits <= 64 && "Cannot zero extend past 64 bits.");
7433 assert(0 <= Offset && "Extension offset must be positive.");
7434 assert((Offset < NumEltsPerLane || Offset % NumEltsPerLane == 0) &&
7435 "Extension offset must be in the first lane or start an upper lane.");
7437 // Check that an index is in same lane as the base offset.
7438 auto SafeOffset = [&](int Idx) {
7439 return OffsetLane == (Idx / NumEltsPerLane);
7442 // Shift along an input so that the offset base moves to the first element.
7443 auto ShuffleOffset = [&](SDValue V) {
7447 SmallVector<int, 8> ShMask((unsigned)NumElements, -1);
7448 for (int i = 0; i * Scale < NumElements; ++i) {
7449 int SrcIdx = i + Offset;
7450 ShMask[i] = SafeOffset(SrcIdx) ? SrcIdx : -1;
7452 return DAG.getVectorShuffle(VT, DL, V, DAG.getUNDEF(VT), ShMask);
7455 // Found a valid zext mask! Try various lowering strategies based on the
7456 // input type and available ISA extensions.
7457 if (Subtarget->hasSSE41()) {
7458 // Not worth offseting 128-bit vectors if scale == 2, a pattern using
7459 // PUNPCK will catch this in a later shuffle match.
7460 if (Offset && Scale == 2 && VT.getSizeInBits() == 128)
7462 MVT ExtVT = MVT::getVectorVT(MVT::getIntegerVT(EltBits * Scale),
7463 NumElements / Scale);
7464 InputV = DAG.getNode(X86ISD::VZEXT, DL, ExtVT, ShuffleOffset(InputV));
7465 return DAG.getBitcast(VT, InputV);
7468 assert(VT.getSizeInBits() == 128 && "Only 128-bit vectors can be extended.");
7470 // For any extends we can cheat for larger element sizes and use shuffle
7471 // instructions that can fold with a load and/or copy.
7472 if (AnyExt && EltBits == 32) {
7473 int PSHUFDMask[4] = {Offset, -1, SafeOffset(Offset + 1) ? Offset + 1 : -1,
7475 return DAG.getBitcast(
7476 VT, DAG.getNode(X86ISD::PSHUFD, DL, MVT::v4i32,
7477 DAG.getBitcast(MVT::v4i32, InputV),
7478 getV4X86ShuffleImm8ForMask(PSHUFDMask, DL, DAG)));
7480 if (AnyExt && EltBits == 16 && Scale > 2) {
7481 int PSHUFDMask[4] = {Offset / 2, -1,
7482 SafeOffset(Offset + 1) ? (Offset + 1) / 2 : -1, -1};
7483 InputV = DAG.getNode(X86ISD::PSHUFD, DL, MVT::v4i32,
7484 DAG.getBitcast(MVT::v4i32, InputV),
7485 getV4X86ShuffleImm8ForMask(PSHUFDMask, DL, DAG));
7486 int PSHUFWMask[4] = {1, -1, -1, -1};
7487 unsigned OddEvenOp = (Offset & 1 ? X86ISD::PSHUFLW : X86ISD::PSHUFHW);
7488 return DAG.getBitcast(
7489 VT, DAG.getNode(OddEvenOp, DL, MVT::v8i16,
7490 DAG.getBitcast(MVT::v8i16, InputV),
7491 getV4X86ShuffleImm8ForMask(PSHUFWMask, DL, DAG)));
7494 // The SSE4A EXTRQ instruction can efficiently extend the first 2 lanes
7496 if ((Scale * EltBits) == 64 && EltBits < 32 && Subtarget->hasSSE4A()) {
7497 assert(NumElements == (int)Mask.size() && "Unexpected shuffle mask size!");
7498 assert(VT.getSizeInBits() == 128 && "Unexpected vector width!");
7500 int LoIdx = Offset * EltBits;
7501 SDValue Lo = DAG.getNode(ISD::BITCAST, DL, MVT::v2i64,
7502 DAG.getNode(X86ISD::EXTRQI, DL, VT, InputV,
7503 DAG.getConstant(EltBits, DL, MVT::i8),
7504 DAG.getConstant(LoIdx, DL, MVT::i8)));
7506 if (isUndefInRange(Mask, NumElements / 2, NumElements / 2) ||
7507 !SafeOffset(Offset + 1))
7508 return DAG.getNode(ISD::BITCAST, DL, VT, Lo);
7510 int HiIdx = (Offset + 1) * EltBits;
7511 SDValue Hi = DAG.getNode(ISD::BITCAST, DL, MVT::v2i64,
7512 DAG.getNode(X86ISD::EXTRQI, DL, VT, InputV,
7513 DAG.getConstant(EltBits, DL, MVT::i8),
7514 DAG.getConstant(HiIdx, DL, MVT::i8)));
7515 return DAG.getNode(ISD::BITCAST, DL, VT,
7516 DAG.getNode(X86ISD::UNPCKL, DL, MVT::v2i64, Lo, Hi));
7519 // If this would require more than 2 unpack instructions to expand, use
7520 // pshufb when available. We can only use more than 2 unpack instructions
7521 // when zero extending i8 elements which also makes it easier to use pshufb.
7522 if (Scale > 4 && EltBits == 8 && Subtarget->hasSSSE3()) {
7523 assert(NumElements == 16 && "Unexpected byte vector width!");
7524 SDValue PSHUFBMask[16];
7525 for (int i = 0; i < 16; ++i) {
7526 int Idx = Offset + (i / Scale);
7527 PSHUFBMask[i] = DAG.getConstant(
7528 (i % Scale == 0 && SafeOffset(Idx)) ? Idx : 0x80, DL, MVT::i8);
7530 InputV = DAG.getBitcast(MVT::v16i8, InputV);
7531 return DAG.getBitcast(VT,
7532 DAG.getNode(X86ISD::PSHUFB, DL, MVT::v16i8, InputV,
7533 DAG.getNode(ISD::BUILD_VECTOR, DL,
7534 MVT::v16i8, PSHUFBMask)));
7537 // If we are extending from an offset, ensure we start on a boundary that
7538 // we can unpack from.
7539 int AlignToUnpack = Offset % (NumElements / Scale);
7540 if (AlignToUnpack) {
7541 SmallVector<int, 8> ShMask((unsigned)NumElements, -1);
7542 for (int i = AlignToUnpack; i < NumElements; ++i)
7543 ShMask[i - AlignToUnpack] = i;
7544 InputV = DAG.getVectorShuffle(VT, DL, InputV, DAG.getUNDEF(VT), ShMask);
7545 Offset -= AlignToUnpack;
7548 // Otherwise emit a sequence of unpacks.
7550 unsigned UnpackLoHi = X86ISD::UNPCKL;
7551 if (Offset >= (NumElements / 2)) {
7552 UnpackLoHi = X86ISD::UNPCKH;
7553 Offset -= (NumElements / 2);
7556 MVT InputVT = MVT::getVectorVT(MVT::getIntegerVT(EltBits), NumElements);
7557 SDValue Ext = AnyExt ? DAG.getUNDEF(InputVT)
7558 : getZeroVector(InputVT, Subtarget, DAG, DL);
7559 InputV = DAG.getBitcast(InputVT, InputV);
7560 InputV = DAG.getNode(UnpackLoHi, DL, InputVT, InputV, Ext);
7564 } while (Scale > 1);
7565 return DAG.getBitcast(VT, InputV);
7568 /// \brief Try to lower a vector shuffle as a zero extension on any microarch.
7570 /// This routine will try to do everything in its power to cleverly lower
7571 /// a shuffle which happens to match the pattern of a zero extend. It doesn't
7572 /// check for the profitability of this lowering, it tries to aggressively
7573 /// match this pattern. It will use all of the micro-architectural details it
7574 /// can to emit an efficient lowering. It handles both blends with all-zero
7575 /// inputs to explicitly zero-extend and undef-lanes (sometimes undef due to
7576 /// masking out later).
7578 /// The reason we have dedicated lowering for zext-style shuffles is that they
7579 /// are both incredibly common and often quite performance sensitive.
7580 static SDValue lowerVectorShuffleAsZeroOrAnyExtend(
7581 SDLoc DL, MVT VT, SDValue V1, SDValue V2, ArrayRef<int> Mask,
7582 const X86Subtarget *Subtarget, SelectionDAG &DAG) {
7583 SmallBitVector Zeroable = computeZeroableShuffleElements(Mask, V1, V2);
7585 int Bits = VT.getSizeInBits();
7586 int NumLanes = Bits / 128;
7587 int NumElements = VT.getVectorNumElements();
7588 int NumEltsPerLane = NumElements / NumLanes;
7589 assert(VT.getScalarSizeInBits() <= 32 &&
7590 "Exceeds 32-bit integer zero extension limit");
7591 assert((int)Mask.size() == NumElements && "Unexpected shuffle mask size");
7593 // Define a helper function to check a particular ext-scale and lower to it if
7595 auto Lower = [&](int Scale) -> SDValue {
7600 for (int i = 0; i < NumElements; ++i) {
7603 continue; // Valid anywhere but doesn't tell us anything.
7604 if (i % Scale != 0) {
7605 // Each of the extended elements need to be zeroable.
7609 // We no longer are in the anyext case.
7614 // Each of the base elements needs to be consecutive indices into the
7615 // same input vector.
7616 SDValue V = M < NumElements ? V1 : V2;
7617 M = M % NumElements;
7620 Offset = M - (i / Scale);
7621 } else if (InputV != V)
7622 return SDValue(); // Flip-flopping inputs.
7624 // Offset must start in the lowest 128-bit lane or at the start of an
7626 // FIXME: Is it ever worth allowing a negative base offset?
7627 if (!((0 <= Offset && Offset < NumEltsPerLane) ||
7628 (Offset % NumEltsPerLane) == 0))
7631 // If we are offsetting, all referenced entries must come from the same
7633 if (Offset && (Offset / NumEltsPerLane) != (M / NumEltsPerLane))
7636 if ((M % NumElements) != (Offset + (i / Scale)))
7637 return SDValue(); // Non-consecutive strided elements.
7641 // If we fail to find an input, we have a zero-shuffle which should always
7642 // have already been handled.
7643 // FIXME: Maybe handle this here in case during blending we end up with one?
7647 // If we are offsetting, don't extend if we only match a single input, we
7648 // can always do better by using a basic PSHUF or PUNPCK.
7649 if (Offset != 0 && Matches < 2)
7652 return lowerVectorShuffleAsSpecificZeroOrAnyExtend(
7653 DL, VT, Scale, Offset, AnyExt, InputV, Mask, Subtarget, DAG);
7656 // The widest scale possible for extending is to a 64-bit integer.
7657 assert(Bits % 64 == 0 &&
7658 "The number of bits in a vector must be divisible by 64 on x86!");
7659 int NumExtElements = Bits / 64;
7661 // Each iteration, try extending the elements half as much, but into twice as
7663 for (; NumExtElements < NumElements; NumExtElements *= 2) {
7664 assert(NumElements % NumExtElements == 0 &&
7665 "The input vector size must be divisible by the extended size.");
7666 if (SDValue V = Lower(NumElements / NumExtElements))
7670 // General extends failed, but 128-bit vectors may be able to use MOVQ.
7674 // Returns one of the source operands if the shuffle can be reduced to a
7675 // MOVQ, copying the lower 64-bits and zero-extending to the upper 64-bits.
7676 auto CanZExtLowHalf = [&]() {
7677 for (int i = NumElements / 2; i != NumElements; ++i)
7680 if (isSequentialOrUndefInRange(Mask, 0, NumElements / 2, 0))
7682 if (isSequentialOrUndefInRange(Mask, 0, NumElements / 2, NumElements))
7687 if (SDValue V = CanZExtLowHalf()) {
7688 V = DAG.getBitcast(MVT::v2i64, V);
7689 V = DAG.getNode(X86ISD::VZEXT_MOVL, DL, MVT::v2i64, V);
7690 return DAG.getBitcast(VT, V);
7693 // No viable ext lowering found.
7697 /// \brief Try to get a scalar value for a specific element of a vector.
7699 /// Looks through BUILD_VECTOR and SCALAR_TO_VECTOR nodes to find a scalar.
7700 static SDValue getScalarValueForVectorElement(SDValue V, int Idx,
7701 SelectionDAG &DAG) {
7702 MVT VT = V.getSimpleValueType();
7703 MVT EltVT = VT.getVectorElementType();
7704 while (V.getOpcode() == ISD::BITCAST)
7705 V = V.getOperand(0);
7706 // If the bitcasts shift the element size, we can't extract an equivalent
7708 MVT NewVT = V.getSimpleValueType();
7709 if (!NewVT.isVector() || NewVT.getScalarSizeInBits() != VT.getScalarSizeInBits())
7712 if (V.getOpcode() == ISD::BUILD_VECTOR ||
7713 (Idx == 0 && V.getOpcode() == ISD::SCALAR_TO_VECTOR)) {
7714 // Ensure the scalar operand is the same size as the destination.
7715 // FIXME: Add support for scalar truncation where possible.
7716 SDValue S = V.getOperand(Idx);
7717 if (EltVT.getSizeInBits() == S.getSimpleValueType().getSizeInBits())
7718 return DAG.getNode(ISD::BITCAST, SDLoc(V), EltVT, S);
7724 /// \brief Helper to test for a load that can be folded with x86 shuffles.
7726 /// This is particularly important because the set of instructions varies
7727 /// significantly based on whether the operand is a load or not.
7728 static bool isShuffleFoldableLoad(SDValue V) {
7729 while (V.getOpcode() == ISD::BITCAST)
7730 V = V.getOperand(0);
7732 return ISD::isNON_EXTLoad(V.getNode());
7735 /// \brief Try to lower insertion of a single element into a zero vector.
7737 /// This is a common pattern that we have especially efficient patterns to lower
7738 /// across all subtarget feature sets.
7739 static SDValue lowerVectorShuffleAsElementInsertion(
7740 SDLoc DL, MVT VT, SDValue V1, SDValue V2, ArrayRef<int> Mask,
7741 const X86Subtarget *Subtarget, SelectionDAG &DAG) {
7742 SmallBitVector Zeroable = computeZeroableShuffleElements(Mask, V1, V2);
7744 MVT EltVT = VT.getVectorElementType();
7746 int V2Index = std::find_if(Mask.begin(), Mask.end(),
7747 [&Mask](int M) { return M >= (int)Mask.size(); }) -
7749 bool IsV1Zeroable = true;
7750 for (int i = 0, Size = Mask.size(); i < Size; ++i)
7751 if (i != V2Index && !Zeroable[i]) {
7752 IsV1Zeroable = false;
7756 // Check for a single input from a SCALAR_TO_VECTOR node.
7757 // FIXME: All of this should be canonicalized into INSERT_VECTOR_ELT and
7758 // all the smarts here sunk into that routine. However, the current
7759 // lowering of BUILD_VECTOR makes that nearly impossible until the old
7760 // vector shuffle lowering is dead.
7761 SDValue V2S = getScalarValueForVectorElement(V2, Mask[V2Index] - Mask.size(),
7763 if (V2S && DAG.getTargetLoweringInfo().isTypeLegal(V2S.getValueType())) {
7764 // We need to zext the scalar if it is smaller than an i32.
7765 V2S = DAG.getBitcast(EltVT, V2S);
7766 if (EltVT == MVT::i8 || EltVT == MVT::i16) {
7767 // Using zext to expand a narrow element won't work for non-zero
7772 // Zero-extend directly to i32.
7774 V2S = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::i32, V2S);
7776 V2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, ExtVT, V2S);
7777 } else if (Mask[V2Index] != (int)Mask.size() || EltVT == MVT::i8 ||
7778 EltVT == MVT::i16) {
7779 // Either not inserting from the low element of the input or the input
7780 // element size is too small to use VZEXT_MOVL to clear the high bits.
7784 if (!IsV1Zeroable) {
7785 // If V1 can't be treated as a zero vector we have fewer options to lower
7786 // this. We can't support integer vectors or non-zero targets cheaply, and
7787 // the V1 elements can't be permuted in any way.
7788 assert(VT == ExtVT && "Cannot change extended type when non-zeroable!");
7789 if (!VT.isFloatingPoint() || V2Index != 0)
7791 SmallVector<int, 8> V1Mask(Mask.begin(), Mask.end());
7792 V1Mask[V2Index] = -1;
7793 if (!isNoopShuffleMask(V1Mask))
7795 // This is essentially a special case blend operation, but if we have
7796 // general purpose blend operations, they are always faster. Bail and let
7797 // the rest of the lowering handle these as blends.
7798 if (Subtarget->hasSSE41())
7801 // Otherwise, use MOVSD or MOVSS.
7802 assert((EltVT == MVT::f32 || EltVT == MVT::f64) &&
7803 "Only two types of floating point element types to handle!");
7804 return DAG.getNode(EltVT == MVT::f32 ? X86ISD::MOVSS : X86ISD::MOVSD, DL,
7808 // This lowering only works for the low element with floating point vectors.
7809 if (VT.isFloatingPoint() && V2Index != 0)
7812 V2 = DAG.getNode(X86ISD::VZEXT_MOVL, DL, ExtVT, V2);
7814 V2 = DAG.getBitcast(VT, V2);
7817 // If we have 4 or fewer lanes we can cheaply shuffle the element into
7818 // the desired position. Otherwise it is more efficient to do a vector
7819 // shift left. We know that we can do a vector shift left because all
7820 // the inputs are zero.
7821 if (VT.isFloatingPoint() || VT.getVectorNumElements() <= 4) {
7822 SmallVector<int, 4> V2Shuffle(Mask.size(), 1);
7823 V2Shuffle[V2Index] = 0;
7824 V2 = DAG.getVectorShuffle(VT, DL, V2, DAG.getUNDEF(VT), V2Shuffle);
7826 V2 = DAG.getBitcast(MVT::v2i64, V2);
7828 X86ISD::VSHLDQ, DL, MVT::v2i64, V2,
7829 DAG.getConstant(V2Index * EltVT.getSizeInBits() / 8, DL,
7830 DAG.getTargetLoweringInfo().getScalarShiftAmountTy(
7831 DAG.getDataLayout(), VT)));
7832 V2 = DAG.getBitcast(VT, V2);
7838 /// \brief Try to lower broadcast of a single element.
7840 /// For convenience, this code also bundles all of the subtarget feature set
7841 /// filtering. While a little annoying to re-dispatch on type here, there isn't
7842 /// a convenient way to factor it out.
7843 static SDValue lowerVectorShuffleAsBroadcast(SDLoc DL, MVT VT, SDValue V,
7845 const X86Subtarget *Subtarget,
7846 SelectionDAG &DAG) {
7847 if (!Subtarget->hasAVX())
7849 if (VT.isInteger() && !Subtarget->hasAVX2())
7852 // Check that the mask is a broadcast.
7853 int BroadcastIdx = -1;
7855 if (M >= 0 && BroadcastIdx == -1)
7857 else if (M >= 0 && M != BroadcastIdx)
7860 assert(BroadcastIdx < (int)Mask.size() && "We only expect to be called with "
7861 "a sorted mask where the broadcast "
7864 // Go up the chain of (vector) values to find a scalar load that we can
7865 // combine with the broadcast.
7867 switch (V.getOpcode()) {
7868 case ISD::CONCAT_VECTORS: {
7869 int OperandSize = Mask.size() / V.getNumOperands();
7870 V = V.getOperand(BroadcastIdx / OperandSize);
7871 BroadcastIdx %= OperandSize;
7875 case ISD::INSERT_SUBVECTOR: {
7876 SDValue VOuter = V.getOperand(0), VInner = V.getOperand(1);
7877 auto ConstantIdx = dyn_cast<ConstantSDNode>(V.getOperand(2));
7881 int BeginIdx = (int)ConstantIdx->getZExtValue();
7883 BeginIdx + (int)VInner.getValueType().getVectorNumElements();
7884 if (BroadcastIdx >= BeginIdx && BroadcastIdx < EndIdx) {
7885 BroadcastIdx -= BeginIdx;
7896 // Check if this is a broadcast of a scalar. We special case lowering
7897 // for scalars so that we can more effectively fold with loads.
7898 // First, look through bitcast: if the original value has a larger element
7899 // type than the shuffle, the broadcast element is in essence truncated.
7900 // Make that explicit to ease folding.
7901 if (V.getOpcode() == ISD::BITCAST && VT.isInteger()) {
7902 EVT EltVT = VT.getVectorElementType();
7903 SDValue V0 = V.getOperand(0);
7904 EVT V0VT = V0.getValueType();
7906 if (V0VT.isInteger() && V0VT.getVectorElementType().bitsGT(EltVT) &&
7907 ((V0.getOpcode() == ISD::BUILD_VECTOR ||
7908 (V0.getOpcode() == ISD::SCALAR_TO_VECTOR && BroadcastIdx == 0)))) {
7909 V = DAG.getNode(ISD::TRUNCATE, DL, EltVT, V0.getOperand(BroadcastIdx));
7914 // Also check the simpler case, where we can directly reuse the scalar.
7915 if (V.getOpcode() == ISD::BUILD_VECTOR ||
7916 (V.getOpcode() == ISD::SCALAR_TO_VECTOR && BroadcastIdx == 0)) {
7917 V = V.getOperand(BroadcastIdx);
7919 // If the scalar isn't a load, we can't broadcast from it in AVX1.
7920 // Only AVX2 has register broadcasts.
7921 if (!Subtarget->hasAVX2() && !isShuffleFoldableLoad(V))
7923 } else if (BroadcastIdx != 0 || !Subtarget->hasAVX2()) {
7924 // We can't broadcast from a vector register without AVX2, and we can only
7925 // broadcast from the zero-element of a vector register.
7929 return DAG.getNode(X86ISD::VBROADCAST, DL, VT, V);
7932 // Check for whether we can use INSERTPS to perform the shuffle. We only use
7933 // INSERTPS when the V1 elements are already in the correct locations
7934 // because otherwise we can just always use two SHUFPS instructions which
7935 // are much smaller to encode than a SHUFPS and an INSERTPS. We can also
7936 // perform INSERTPS if a single V1 element is out of place and all V2
7937 // elements are zeroable.
7938 static SDValue lowerVectorShuffleAsInsertPS(SDValue Op, SDValue V1, SDValue V2,
7940 SelectionDAG &DAG) {
7941 assert(Op.getSimpleValueType() == MVT::v4f32 && "Bad shuffle type!");
7942 assert(V1.getSimpleValueType() == MVT::v4f32 && "Bad operand type!");
7943 assert(V2.getSimpleValueType() == MVT::v4f32 && "Bad operand type!");
7944 assert(Mask.size() == 4 && "Unexpected mask size for v4 shuffle!");
7946 SmallBitVector Zeroable = computeZeroableShuffleElements(Mask, V1, V2);
7949 int V1DstIndex = -1;
7950 int V2DstIndex = -1;
7951 bool V1UsedInPlace = false;
7953 for (int i = 0; i < 4; ++i) {
7954 // Synthesize a zero mask from the zeroable elements (includes undefs).
7960 // Flag if we use any V1 inputs in place.
7962 V1UsedInPlace = true;
7966 // We can only insert a single non-zeroable element.
7967 if (V1DstIndex != -1 || V2DstIndex != -1)
7971 // V1 input out of place for insertion.
7974 // V2 input for insertion.
7979 // Don't bother if we have no (non-zeroable) element for insertion.
7980 if (V1DstIndex == -1 && V2DstIndex == -1)
7983 // Determine element insertion src/dst indices. The src index is from the
7984 // start of the inserted vector, not the start of the concatenated vector.
7985 unsigned V2SrcIndex = 0;
7986 if (V1DstIndex != -1) {
7987 // If we have a V1 input out of place, we use V1 as the V2 element insertion
7988 // and don't use the original V2 at all.
7989 V2SrcIndex = Mask[V1DstIndex];
7990 V2DstIndex = V1DstIndex;
7993 V2SrcIndex = Mask[V2DstIndex] - 4;
7996 // If no V1 inputs are used in place, then the result is created only from
7997 // the zero mask and the V2 insertion - so remove V1 dependency.
7999 V1 = DAG.getUNDEF(MVT::v4f32);
8001 unsigned InsertPSMask = V2SrcIndex << 6 | V2DstIndex << 4 | ZMask;
8002 assert((InsertPSMask & ~0xFFu) == 0 && "Invalid mask!");
8004 // Insert the V2 element into the desired position.
8006 return DAG.getNode(X86ISD::INSERTPS, DL, MVT::v4f32, V1, V2,
8007 DAG.getConstant(InsertPSMask, DL, MVT::i8));
8010 /// \brief Try to lower a shuffle as a permute of the inputs followed by an
8011 /// UNPCK instruction.
8013 /// This specifically targets cases where we end up with alternating between
8014 /// the two inputs, and so can permute them into something that feeds a single
8015 /// UNPCK instruction. Note that this routine only targets integer vectors
8016 /// because for floating point vectors we have a generalized SHUFPS lowering
8017 /// strategy that handles everything that doesn't *exactly* match an unpack,
8018 /// making this clever lowering unnecessary.
8019 static SDValue lowerVectorShuffleAsPermuteAndUnpack(SDLoc DL, MVT VT,
8020 SDValue V1, SDValue V2,
8022 SelectionDAG &DAG) {
8023 assert(!VT.isFloatingPoint() &&
8024 "This routine only supports integer vectors.");
8025 assert(!isSingleInputShuffleMask(Mask) &&
8026 "This routine should only be used when blending two inputs.");
8027 assert(Mask.size() >= 2 && "Single element masks are invalid.");
8029 int Size = Mask.size();
8031 int NumLoInputs = std::count_if(Mask.begin(), Mask.end(), [Size](int M) {
8032 return M >= 0 && M % Size < Size / 2;
8034 int NumHiInputs = std::count_if(
8035 Mask.begin(), Mask.end(), [Size](int M) { return M % Size >= Size / 2; });
8037 bool UnpackLo = NumLoInputs >= NumHiInputs;
8039 auto TryUnpack = [&](MVT UnpackVT, int Scale) {
8040 SmallVector<int, 32> V1Mask(Mask.size(), -1);
8041 SmallVector<int, 32> V2Mask(Mask.size(), -1);
8043 for (int i = 0; i < Size; ++i) {
8047 // Each element of the unpack contains Scale elements from this mask.
8048 int UnpackIdx = i / Scale;
8050 // We only handle the case where V1 feeds the first slots of the unpack.
8051 // We rely on canonicalization to ensure this is the case.
8052 if ((UnpackIdx % 2 == 0) != (Mask[i] < Size))
8055 // Setup the mask for this input. The indexing is tricky as we have to
8056 // handle the unpack stride.
8057 SmallVectorImpl<int> &VMask = (UnpackIdx % 2 == 0) ? V1Mask : V2Mask;
8058 VMask[(UnpackIdx / 2) * Scale + i % Scale + (UnpackLo ? 0 : Size / 2)] =
8062 // If we will have to shuffle both inputs to use the unpack, check whether
8063 // we can just unpack first and shuffle the result. If so, skip this unpack.
8064 if ((NumLoInputs == 0 || NumHiInputs == 0) && !isNoopShuffleMask(V1Mask) &&
8065 !isNoopShuffleMask(V2Mask))
8068 // Shuffle the inputs into place.
8069 V1 = DAG.getVectorShuffle(VT, DL, V1, DAG.getUNDEF(VT), V1Mask);
8070 V2 = DAG.getVectorShuffle(VT, DL, V2, DAG.getUNDEF(VT), V2Mask);
8072 // Cast the inputs to the type we will use to unpack them.
8073 V1 = DAG.getBitcast(UnpackVT, V1);
8074 V2 = DAG.getBitcast(UnpackVT, V2);
8076 // Unpack the inputs and cast the result back to the desired type.
8077 return DAG.getBitcast(
8078 VT, DAG.getNode(UnpackLo ? X86ISD::UNPCKL : X86ISD::UNPCKH, DL,
8082 // We try each unpack from the largest to the smallest to try and find one
8083 // that fits this mask.
8084 int OrigNumElements = VT.getVectorNumElements();
8085 int OrigScalarSize = VT.getScalarSizeInBits();
8086 for (int ScalarSize = 64; ScalarSize >= OrigScalarSize; ScalarSize /= 2) {
8087 int Scale = ScalarSize / OrigScalarSize;
8088 int NumElements = OrigNumElements / Scale;
8089 MVT UnpackVT = MVT::getVectorVT(MVT::getIntegerVT(ScalarSize), NumElements);
8090 if (SDValue Unpack = TryUnpack(UnpackVT, Scale))
8094 // If none of the unpack-rooted lowerings worked (or were profitable) try an
8096 if (NumLoInputs == 0 || NumHiInputs == 0) {
8097 assert((NumLoInputs > 0 || NumHiInputs > 0) &&
8098 "We have to have *some* inputs!");
8099 int HalfOffset = NumLoInputs == 0 ? Size / 2 : 0;
8101 // FIXME: We could consider the total complexity of the permute of each
8102 // possible unpacking. Or at the least we should consider how many
8103 // half-crossings are created.
8104 // FIXME: We could consider commuting the unpacks.
8106 SmallVector<int, 32> PermMask;
8107 PermMask.assign(Size, -1);
8108 for (int i = 0; i < Size; ++i) {
8112 assert(Mask[i] % Size >= HalfOffset && "Found input from wrong half!");
8115 2 * ((Mask[i] % Size) - HalfOffset) + (Mask[i] < Size ? 0 : 1);
8117 return DAG.getVectorShuffle(
8118 VT, DL, DAG.getNode(NumLoInputs == 0 ? X86ISD::UNPCKH : X86ISD::UNPCKL,
8120 DAG.getUNDEF(VT), PermMask);
8126 /// \brief Handle lowering of 2-lane 64-bit floating point shuffles.
8128 /// This is the basis function for the 2-lane 64-bit shuffles as we have full
8129 /// support for floating point shuffles but not integer shuffles. These
8130 /// instructions will incur a domain crossing penalty on some chips though so
8131 /// it is better to avoid lowering through this for integer vectors where
8133 static SDValue lowerV2F64VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
8134 const X86Subtarget *Subtarget,
8135 SelectionDAG &DAG) {
8137 assert(Op.getSimpleValueType() == MVT::v2f64 && "Bad shuffle type!");
8138 assert(V1.getSimpleValueType() == MVT::v2f64 && "Bad operand type!");
8139 assert(V2.getSimpleValueType() == MVT::v2f64 && "Bad operand type!");
8140 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
8141 ArrayRef<int> Mask = SVOp->getMask();
8142 assert(Mask.size() == 2 && "Unexpected mask size for v2 shuffle!");
8144 if (isSingleInputShuffleMask(Mask)) {
8145 // Use low duplicate instructions for masks that match their pattern.
8146 if (Subtarget->hasSSE3())
8147 if (isShuffleEquivalent(V1, V2, Mask, {0, 0}))
8148 return DAG.getNode(X86ISD::MOVDDUP, DL, MVT::v2f64, V1);
8150 // Straight shuffle of a single input vector. Simulate this by using the
8151 // single input as both of the "inputs" to this instruction..
8152 unsigned SHUFPDMask = (Mask[0] == 1) | ((Mask[1] == 1) << 1);
8154 if (Subtarget->hasAVX()) {
8155 // If we have AVX, we can use VPERMILPS which will allow folding a load
8156 // into the shuffle.
8157 return DAG.getNode(X86ISD::VPERMILPI, DL, MVT::v2f64, V1,
8158 DAG.getConstant(SHUFPDMask, DL, MVT::i8));
8161 return DAG.getNode(X86ISD::SHUFP, DL, MVT::v2f64, V1, V1,
8162 DAG.getConstant(SHUFPDMask, DL, MVT::i8));
8164 assert(Mask[0] >= 0 && Mask[0] < 2 && "Non-canonicalized blend!");
8165 assert(Mask[1] >= 2 && "Non-canonicalized blend!");
8167 // If we have a single input, insert that into V1 if we can do so cheaply.
8168 if ((Mask[0] >= 2) + (Mask[1] >= 2) == 1) {
8169 if (SDValue Insertion = lowerVectorShuffleAsElementInsertion(
8170 DL, MVT::v2f64, V1, V2, Mask, Subtarget, DAG))
8172 // Try inverting the insertion since for v2 masks it is easy to do and we
8173 // can't reliably sort the mask one way or the other.
8174 int InverseMask[2] = {Mask[0] < 0 ? -1 : (Mask[0] ^ 2),
8175 Mask[1] < 0 ? -1 : (Mask[1] ^ 2)};
8176 if (SDValue Insertion = lowerVectorShuffleAsElementInsertion(
8177 DL, MVT::v2f64, V2, V1, InverseMask, Subtarget, DAG))
8181 // Try to use one of the special instruction patterns to handle two common
8182 // blend patterns if a zero-blend above didn't work.
8183 if (isShuffleEquivalent(V1, V2, Mask, {0, 3}) ||
8184 isShuffleEquivalent(V1, V2, Mask, {1, 3}))
8185 if (SDValue V1S = getScalarValueForVectorElement(V1, Mask[0], DAG))
8186 // We can either use a special instruction to load over the low double or
8187 // to move just the low double.
8189 isShuffleFoldableLoad(V1S) ? X86ISD::MOVLPD : X86ISD::MOVSD,
8191 DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, MVT::v2f64, V1S));
8193 if (Subtarget->hasSSE41())
8194 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v2f64, V1, V2, Mask,
8198 // Use dedicated unpack instructions for masks that match their pattern.
8200 lowerVectorShuffleWithUNPCK(DL, MVT::v2f64, Mask, V1, V2, DAG))
8203 unsigned SHUFPDMask = (Mask[0] == 1) | (((Mask[1] - 2) == 1) << 1);
8204 return DAG.getNode(X86ISD::SHUFP, DL, MVT::v2f64, V1, V2,
8205 DAG.getConstant(SHUFPDMask, DL, MVT::i8));
8208 /// \brief Handle lowering of 2-lane 64-bit integer shuffles.
8210 /// Tries to lower a 2-lane 64-bit shuffle using shuffle operations provided by
8211 /// the integer unit to minimize domain crossing penalties. However, for blends
8212 /// it falls back to the floating point shuffle operation with appropriate bit
8214 static SDValue lowerV2I64VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
8215 const X86Subtarget *Subtarget,
8216 SelectionDAG &DAG) {
8218 assert(Op.getSimpleValueType() == MVT::v2i64 && "Bad shuffle type!");
8219 assert(V1.getSimpleValueType() == MVT::v2i64 && "Bad operand type!");
8220 assert(V2.getSimpleValueType() == MVT::v2i64 && "Bad operand type!");
8221 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
8222 ArrayRef<int> Mask = SVOp->getMask();
8223 assert(Mask.size() == 2 && "Unexpected mask size for v2 shuffle!");
8225 if (isSingleInputShuffleMask(Mask)) {
8226 // Check for being able to broadcast a single element.
8227 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v2i64, V1,
8228 Mask, Subtarget, DAG))
8231 // Straight shuffle of a single input vector. For everything from SSE2
8232 // onward this has a single fast instruction with no scary immediates.
8233 // We have to map the mask as it is actually a v4i32 shuffle instruction.
8234 V1 = DAG.getBitcast(MVT::v4i32, V1);
8235 int WidenedMask[4] = {
8236 std::max(Mask[0], 0) * 2, std::max(Mask[0], 0) * 2 + 1,
8237 std::max(Mask[1], 0) * 2, std::max(Mask[1], 0) * 2 + 1};
8238 return DAG.getBitcast(
8240 DAG.getNode(X86ISD::PSHUFD, DL, MVT::v4i32, V1,
8241 getV4X86ShuffleImm8ForMask(WidenedMask, DL, DAG)));
8243 assert(Mask[0] != -1 && "No undef lanes in multi-input v2 shuffles!");
8244 assert(Mask[1] != -1 && "No undef lanes in multi-input v2 shuffles!");
8245 assert(Mask[0] < 2 && "We sort V1 to be the first input.");
8246 assert(Mask[1] >= 2 && "We sort V2 to be the second input.");
8248 // If we have a blend of two PACKUS operations an the blend aligns with the
8249 // low and half halves, we can just merge the PACKUS operations. This is
8250 // particularly important as it lets us merge shuffles that this routine itself
8252 auto GetPackNode = [](SDValue V) {
8253 while (V.getOpcode() == ISD::BITCAST)
8254 V = V.getOperand(0);
8256 return V.getOpcode() == X86ISD::PACKUS ? V : SDValue();
8258 if (SDValue V1Pack = GetPackNode(V1))
8259 if (SDValue V2Pack = GetPackNode(V2))
8260 return DAG.getBitcast(MVT::v2i64,
8261 DAG.getNode(X86ISD::PACKUS, DL, MVT::v16i8,
8262 Mask[0] == 0 ? V1Pack.getOperand(0)
8263 : V1Pack.getOperand(1),
8264 Mask[1] == 2 ? V2Pack.getOperand(0)
8265 : V2Pack.getOperand(1)));
8267 // Try to use shift instructions.
8269 lowerVectorShuffleAsShift(DL, MVT::v2i64, V1, V2, Mask, DAG))
8272 // When loading a scalar and then shuffling it into a vector we can often do
8273 // the insertion cheaply.
8274 if (SDValue Insertion = lowerVectorShuffleAsElementInsertion(
8275 DL, MVT::v2i64, V1, V2, Mask, Subtarget, DAG))
8277 // Try inverting the insertion since for v2 masks it is easy to do and we
8278 // can't reliably sort the mask one way or the other.
8279 int InverseMask[2] = {Mask[0] ^ 2, Mask[1] ^ 2};
8280 if (SDValue Insertion = lowerVectorShuffleAsElementInsertion(
8281 DL, MVT::v2i64, V2, V1, InverseMask, Subtarget, DAG))
8284 // We have different paths for blend lowering, but they all must use the
8285 // *exact* same predicate.
8286 bool IsBlendSupported = Subtarget->hasSSE41();
8287 if (IsBlendSupported)
8288 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v2i64, V1, V2, Mask,
8292 // Use dedicated unpack instructions for masks that match their pattern.
8294 lowerVectorShuffleWithUNPCK(DL, MVT::v2i64, Mask, V1, V2, DAG))
8297 // Try to use byte rotation instructions.
8298 // Its more profitable for pre-SSSE3 to use shuffles/unpacks.
8299 if (Subtarget->hasSSSE3())
8300 if (SDValue Rotate = lowerVectorShuffleAsByteRotate(
8301 DL, MVT::v2i64, V1, V2, Mask, Subtarget, DAG))
8304 // If we have direct support for blends, we should lower by decomposing into
8305 // a permute. That will be faster than the domain cross.
8306 if (IsBlendSupported)
8307 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v2i64, V1, V2,
8310 // We implement this with SHUFPD which is pretty lame because it will likely
8311 // incur 2 cycles of stall for integer vectors on Nehalem and older chips.
8312 // However, all the alternatives are still more cycles and newer chips don't
8313 // have this problem. It would be really nice if x86 had better shuffles here.
8314 V1 = DAG.getBitcast(MVT::v2f64, V1);
8315 V2 = DAG.getBitcast(MVT::v2f64, V2);
8316 return DAG.getBitcast(MVT::v2i64,
8317 DAG.getVectorShuffle(MVT::v2f64, DL, V1, V2, Mask));
8320 /// \brief Test whether this can be lowered with a single SHUFPS instruction.
8322 /// This is used to disable more specialized lowerings when the shufps lowering
8323 /// will happen to be efficient.
8324 static bool isSingleSHUFPSMask(ArrayRef<int> Mask) {
8325 // This routine only handles 128-bit shufps.
8326 assert(Mask.size() == 4 && "Unsupported mask size!");
8328 // To lower with a single SHUFPS we need to have the low half and high half
8329 // each requiring a single input.
8330 if (Mask[0] != -1 && Mask[1] != -1 && (Mask[0] < 4) != (Mask[1] < 4))
8332 if (Mask[2] != -1 && Mask[3] != -1 && (Mask[2] < 4) != (Mask[3] < 4))
8338 /// \brief Lower a vector shuffle using the SHUFPS instruction.
8340 /// This is a helper routine dedicated to lowering vector shuffles using SHUFPS.
8341 /// It makes no assumptions about whether this is the *best* lowering, it simply
8343 static SDValue lowerVectorShuffleWithSHUFPS(SDLoc DL, MVT VT,
8344 ArrayRef<int> Mask, SDValue V1,
8345 SDValue V2, SelectionDAG &DAG) {
8346 SDValue LowV = V1, HighV = V2;
8347 int NewMask[4] = {Mask[0], Mask[1], Mask[2], Mask[3]};
8350 std::count_if(Mask.begin(), Mask.end(), [](int M) { return M >= 4; });
8352 if (NumV2Elements == 1) {
8354 std::find_if(Mask.begin(), Mask.end(), [](int M) { return M >= 4; }) -
8357 // Compute the index adjacent to V2Index and in the same half by toggling
8359 int V2AdjIndex = V2Index ^ 1;
8361 if (Mask[V2AdjIndex] == -1) {
8362 // Handles all the cases where we have a single V2 element and an undef.
8363 // This will only ever happen in the high lanes because we commute the
8364 // vector otherwise.
8366 std::swap(LowV, HighV);
8367 NewMask[V2Index] -= 4;
8369 // Handle the case where the V2 element ends up adjacent to a V1 element.
8370 // To make this work, blend them together as the first step.
8371 int V1Index = V2AdjIndex;
8372 int BlendMask[4] = {Mask[V2Index] - 4, 0, Mask[V1Index], 0};
8373 V2 = DAG.getNode(X86ISD::SHUFP, DL, VT, V2, V1,
8374 getV4X86ShuffleImm8ForMask(BlendMask, DL, DAG));
8376 // Now proceed to reconstruct the final blend as we have the necessary
8377 // high or low half formed.
8384 NewMask[V1Index] = 2; // We put the V1 element in V2[2].
8385 NewMask[V2Index] = 0; // We shifted the V2 element into V2[0].
8387 } else if (NumV2Elements == 2) {
8388 if (Mask[0] < 4 && Mask[1] < 4) {
8389 // Handle the easy case where we have V1 in the low lanes and V2 in the
8393 } else if (Mask[2] < 4 && Mask[3] < 4) {
8394 // We also handle the reversed case because this utility may get called
8395 // when we detect a SHUFPS pattern but can't easily commute the shuffle to
8396 // arrange things in the right direction.
8402 // We have a mixture of V1 and V2 in both low and high lanes. Rather than
8403 // trying to place elements directly, just blend them and set up the final
8404 // shuffle to place them.
8406 // The first two blend mask elements are for V1, the second two are for
8408 int BlendMask[4] = {Mask[0] < 4 ? Mask[0] : Mask[1],
8409 Mask[2] < 4 ? Mask[2] : Mask[3],
8410 (Mask[0] >= 4 ? Mask[0] : Mask[1]) - 4,
8411 (Mask[2] >= 4 ? Mask[2] : Mask[3]) - 4};
8412 V1 = DAG.getNode(X86ISD::SHUFP, DL, VT, V1, V2,
8413 getV4X86ShuffleImm8ForMask(BlendMask, DL, DAG));
8415 // Now we do a normal shuffle of V1 by giving V1 as both operands to
8418 NewMask[0] = Mask[0] < 4 ? 0 : 2;
8419 NewMask[1] = Mask[0] < 4 ? 2 : 0;
8420 NewMask[2] = Mask[2] < 4 ? 1 : 3;
8421 NewMask[3] = Mask[2] < 4 ? 3 : 1;
8424 return DAG.getNode(X86ISD::SHUFP, DL, VT, LowV, HighV,
8425 getV4X86ShuffleImm8ForMask(NewMask, DL, DAG));
8428 /// \brief Lower 4-lane 32-bit floating point shuffles.
8430 /// Uses instructions exclusively from the floating point unit to minimize
8431 /// domain crossing penalties, as these are sufficient to implement all v4f32
8433 static SDValue lowerV4F32VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
8434 const X86Subtarget *Subtarget,
8435 SelectionDAG &DAG) {
8437 assert(Op.getSimpleValueType() == MVT::v4f32 && "Bad shuffle type!");
8438 assert(V1.getSimpleValueType() == MVT::v4f32 && "Bad operand type!");
8439 assert(V2.getSimpleValueType() == MVT::v4f32 && "Bad operand type!");
8440 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
8441 ArrayRef<int> Mask = SVOp->getMask();
8442 assert(Mask.size() == 4 && "Unexpected mask size for v4 shuffle!");
8445 std::count_if(Mask.begin(), Mask.end(), [](int M) { return M >= 4; });
8447 if (NumV2Elements == 0) {
8448 // Check for being able to broadcast a single element.
8449 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v4f32, V1,
8450 Mask, Subtarget, DAG))
8453 // Use even/odd duplicate instructions for masks that match their pattern.
8454 if (Subtarget->hasSSE3()) {
8455 if (isShuffleEquivalent(V1, V2, Mask, {0, 0, 2, 2}))
8456 return DAG.getNode(X86ISD::MOVSLDUP, DL, MVT::v4f32, V1);
8457 if (isShuffleEquivalent(V1, V2, Mask, {1, 1, 3, 3}))
8458 return DAG.getNode(X86ISD::MOVSHDUP, DL, MVT::v4f32, V1);
8461 if (Subtarget->hasAVX()) {
8462 // If we have AVX, we can use VPERMILPS which will allow folding a load
8463 // into the shuffle.
8464 return DAG.getNode(X86ISD::VPERMILPI, DL, MVT::v4f32, V1,
8465 getV4X86ShuffleImm8ForMask(Mask, DL, DAG));
8468 // Otherwise, use a straight shuffle of a single input vector. We pass the
8469 // input vector to both operands to simulate this with a SHUFPS.
8470 return DAG.getNode(X86ISD::SHUFP, DL, MVT::v4f32, V1, V1,
8471 getV4X86ShuffleImm8ForMask(Mask, DL, DAG));
8474 // There are special ways we can lower some single-element blends. However, we
8475 // have custom ways we can lower more complex single-element blends below that
8476 // we defer to if both this and BLENDPS fail to match, so restrict this to
8477 // when the V2 input is targeting element 0 of the mask -- that is the fast
8479 if (NumV2Elements == 1 && Mask[0] >= 4)
8480 if (SDValue V = lowerVectorShuffleAsElementInsertion(DL, MVT::v4f32, V1, V2,
8481 Mask, Subtarget, DAG))
8484 if (Subtarget->hasSSE41()) {
8485 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v4f32, V1, V2, Mask,
8489 // Use INSERTPS if we can complete the shuffle efficiently.
8490 if (SDValue V = lowerVectorShuffleAsInsertPS(Op, V1, V2, Mask, DAG))
8493 if (!isSingleSHUFPSMask(Mask))
8494 if (SDValue BlendPerm = lowerVectorShuffleAsBlendAndPermute(
8495 DL, MVT::v4f32, V1, V2, Mask, DAG))
8499 // Use dedicated unpack instructions for masks that match their pattern.
8501 lowerVectorShuffleWithUNPCK(DL, MVT::v4f32, Mask, V1, V2, DAG))
8504 // Otherwise fall back to a SHUFPS lowering strategy.
8505 return lowerVectorShuffleWithSHUFPS(DL, MVT::v4f32, Mask, V1, V2, DAG);
8508 /// \brief Lower 4-lane i32 vector shuffles.
8510 /// We try to handle these with integer-domain shuffles where we can, but for
8511 /// blends we use the floating point domain blend instructions.
8512 static SDValue lowerV4I32VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
8513 const X86Subtarget *Subtarget,
8514 SelectionDAG &DAG) {
8516 assert(Op.getSimpleValueType() == MVT::v4i32 && "Bad shuffle type!");
8517 assert(V1.getSimpleValueType() == MVT::v4i32 && "Bad operand type!");
8518 assert(V2.getSimpleValueType() == MVT::v4i32 && "Bad operand type!");
8519 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
8520 ArrayRef<int> Mask = SVOp->getMask();
8521 assert(Mask.size() == 4 && "Unexpected mask size for v4 shuffle!");
8523 // Whenever we can lower this as a zext, that instruction is strictly faster
8524 // than any alternative. It also allows us to fold memory operands into the
8525 // shuffle in many cases.
8526 if (SDValue ZExt = lowerVectorShuffleAsZeroOrAnyExtend(DL, MVT::v4i32, V1, V2,
8527 Mask, Subtarget, DAG))
8531 std::count_if(Mask.begin(), Mask.end(), [](int M) { return M >= 4; });
8533 if (NumV2Elements == 0) {
8534 // Check for being able to broadcast a single element.
8535 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v4i32, V1,
8536 Mask, Subtarget, DAG))
8539 // Straight shuffle of a single input vector. For everything from SSE2
8540 // onward this has a single fast instruction with no scary immediates.
8541 // We coerce the shuffle pattern to be compatible with UNPCK instructions
8542 // but we aren't actually going to use the UNPCK instruction because doing
8543 // so prevents folding a load into this instruction or making a copy.
8544 const int UnpackLoMask[] = {0, 0, 1, 1};
8545 const int UnpackHiMask[] = {2, 2, 3, 3};
8546 if (isShuffleEquivalent(V1, V2, Mask, {0, 0, 1, 1}))
8547 Mask = UnpackLoMask;
8548 else if (isShuffleEquivalent(V1, V2, Mask, {2, 2, 3, 3}))
8549 Mask = UnpackHiMask;
8551 return DAG.getNode(X86ISD::PSHUFD, DL, MVT::v4i32, V1,
8552 getV4X86ShuffleImm8ForMask(Mask, DL, DAG));
8555 // Try to use shift instructions.
8557 lowerVectorShuffleAsShift(DL, MVT::v4i32, V1, V2, Mask, DAG))
8560 // There are special ways we can lower some single-element blends.
8561 if (NumV2Elements == 1)
8562 if (SDValue V = lowerVectorShuffleAsElementInsertion(DL, MVT::v4i32, V1, V2,
8563 Mask, Subtarget, DAG))
8566 // We have different paths for blend lowering, but they all must use the
8567 // *exact* same predicate.
8568 bool IsBlendSupported = Subtarget->hasSSE41();
8569 if (IsBlendSupported)
8570 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v4i32, V1, V2, Mask,
8574 if (SDValue Masked =
8575 lowerVectorShuffleAsBitMask(DL, MVT::v4i32, V1, V2, Mask, DAG))
8578 // Use dedicated unpack instructions for masks that match their pattern.
8580 lowerVectorShuffleWithUNPCK(DL, MVT::v4i32, Mask, V1, V2, DAG))
8583 // Try to use byte rotation instructions.
8584 // Its more profitable for pre-SSSE3 to use shuffles/unpacks.
8585 if (Subtarget->hasSSSE3())
8586 if (SDValue Rotate = lowerVectorShuffleAsByteRotate(
8587 DL, MVT::v4i32, V1, V2, Mask, Subtarget, DAG))
8590 // If we have direct support for blends, we should lower by decomposing into
8591 // a permute. That will be faster than the domain cross.
8592 if (IsBlendSupported)
8593 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v4i32, V1, V2,
8596 // Try to lower by permuting the inputs into an unpack instruction.
8597 if (SDValue Unpack = lowerVectorShuffleAsPermuteAndUnpack(DL, MVT::v4i32, V1,
8601 // We implement this with SHUFPS because it can blend from two vectors.
8602 // Because we're going to eventually use SHUFPS, we use SHUFPS even to build
8603 // up the inputs, bypassing domain shift penalties that we would encur if we
8604 // directly used PSHUFD on Nehalem and older. For newer chips, this isn't
8606 return DAG.getBitcast(
8608 DAG.getVectorShuffle(MVT::v4f32, DL, DAG.getBitcast(MVT::v4f32, V1),
8609 DAG.getBitcast(MVT::v4f32, V2), Mask));
8612 /// \brief Lowering of single-input v8i16 shuffles is the cornerstone of SSE2
8613 /// shuffle lowering, and the most complex part.
8615 /// The lowering strategy is to try to form pairs of input lanes which are
8616 /// targeted at the same half of the final vector, and then use a dword shuffle
8617 /// to place them onto the right half, and finally unpack the paired lanes into
8618 /// their final position.
8620 /// The exact breakdown of how to form these dword pairs and align them on the
8621 /// correct sides is really tricky. See the comments within the function for
8622 /// more of the details.
8624 /// This code also handles repeated 128-bit lanes of v8i16 shuffles, but each
8625 /// lane must shuffle the *exact* same way. In fact, you must pass a v8 Mask to
8626 /// this routine for it to work correctly. To shuffle a 256-bit or 512-bit i16
8627 /// vector, form the analogous 128-bit 8-element Mask.
8628 static SDValue lowerV8I16GeneralSingleInputVectorShuffle(
8629 SDLoc DL, MVT VT, SDValue V, MutableArrayRef<int> Mask,
8630 const X86Subtarget *Subtarget, SelectionDAG &DAG) {
8631 assert(VT.getScalarType() == MVT::i16 && "Bad input type!");
8632 MVT PSHUFDVT = MVT::getVectorVT(MVT::i32, VT.getVectorNumElements() / 2);
8634 assert(Mask.size() == 8 && "Shuffle mask length doen't match!");
8635 MutableArrayRef<int> LoMask = Mask.slice(0, 4);
8636 MutableArrayRef<int> HiMask = Mask.slice(4, 4);
8638 SmallVector<int, 4> LoInputs;
8639 std::copy_if(LoMask.begin(), LoMask.end(), std::back_inserter(LoInputs),
8640 [](int M) { return M >= 0; });
8641 std::sort(LoInputs.begin(), LoInputs.end());
8642 LoInputs.erase(std::unique(LoInputs.begin(), LoInputs.end()), LoInputs.end());
8643 SmallVector<int, 4> HiInputs;
8644 std::copy_if(HiMask.begin(), HiMask.end(), std::back_inserter(HiInputs),
8645 [](int M) { return M >= 0; });
8646 std::sort(HiInputs.begin(), HiInputs.end());
8647 HiInputs.erase(std::unique(HiInputs.begin(), HiInputs.end()), HiInputs.end());
8649 std::lower_bound(LoInputs.begin(), LoInputs.end(), 4) - LoInputs.begin();
8650 int NumHToL = LoInputs.size() - NumLToL;
8652 std::lower_bound(HiInputs.begin(), HiInputs.end(), 4) - HiInputs.begin();
8653 int NumHToH = HiInputs.size() - NumLToH;
8654 MutableArrayRef<int> LToLInputs(LoInputs.data(), NumLToL);
8655 MutableArrayRef<int> LToHInputs(HiInputs.data(), NumLToH);
8656 MutableArrayRef<int> HToLInputs(LoInputs.data() + NumLToL, NumHToL);
8657 MutableArrayRef<int> HToHInputs(HiInputs.data() + NumLToH, NumHToH);
8659 // Simplify the 1-into-3 and 3-into-1 cases with a single pshufd. For all
8660 // such inputs we can swap two of the dwords across the half mark and end up
8661 // with <=2 inputs to each half in each half. Once there, we can fall through
8662 // to the generic code below. For example:
8664 // Input: [a, b, c, d, e, f, g, h] -PSHUFD[0,2,1,3]-> [a, b, e, f, c, d, g, h]
8665 // Mask: [0, 1, 2, 7, 4, 5, 6, 3] -----------------> [0, 1, 4, 7, 2, 3, 6, 5]
8667 // However in some very rare cases we have a 1-into-3 or 3-into-1 on one half
8668 // and an existing 2-into-2 on the other half. In this case we may have to
8669 // pre-shuffle the 2-into-2 half to avoid turning it into a 3-into-1 or
8670 // 1-into-3 which could cause us to cycle endlessly fixing each side in turn.
8671 // Fortunately, we don't have to handle anything but a 2-into-2 pattern
8672 // because any other situation (including a 3-into-1 or 1-into-3 in the other
8673 // half than the one we target for fixing) will be fixed when we re-enter this
8674 // path. We will also combine away any sequence of PSHUFD instructions that
8675 // result into a single instruction. Here is an example of the tricky case:
8677 // Input: [a, b, c, d, e, f, g, h] -PSHUFD[0,2,1,3]-> [a, b, e, f, c, d, g, h]
8678 // Mask: [3, 7, 1, 0, 2, 7, 3, 5] -THIS-IS-BAD!!!!-> [5, 7, 1, 0, 4, 7, 5, 3]
8680 // This now has a 1-into-3 in the high half! Instead, we do two shuffles:
8682 // Input: [a, b, c, d, e, f, g, h] PSHUFHW[0,2,1,3]-> [a, b, c, d, e, g, f, h]
8683 // Mask: [3, 7, 1, 0, 2, 7, 3, 5] -----------------> [3, 7, 1, 0, 2, 7, 3, 6]
8685 // Input: [a, b, c, d, e, g, f, h] -PSHUFD[0,2,1,3]-> [a, b, e, g, c, d, f, h]
8686 // Mask: [3, 7, 1, 0, 2, 7, 3, 6] -----------------> [5, 7, 1, 0, 4, 7, 5, 6]
8688 // The result is fine to be handled by the generic logic.
8689 auto balanceSides = [&](ArrayRef<int> AToAInputs, ArrayRef<int> BToAInputs,
8690 ArrayRef<int> BToBInputs, ArrayRef<int> AToBInputs,
8691 int AOffset, int BOffset) {
8692 assert((AToAInputs.size() == 3 || AToAInputs.size() == 1) &&
8693 "Must call this with A having 3 or 1 inputs from the A half.");
8694 assert((BToAInputs.size() == 1 || BToAInputs.size() == 3) &&
8695 "Must call this with B having 1 or 3 inputs from the B half.");
8696 assert(AToAInputs.size() + BToAInputs.size() == 4 &&
8697 "Must call this with either 3:1 or 1:3 inputs (summing to 4).");
8699 bool ThreeAInputs = AToAInputs.size() == 3;
8701 // Compute the index of dword with only one word among the three inputs in
8702 // a half by taking the sum of the half with three inputs and subtracting
8703 // the sum of the actual three inputs. The difference is the remaining
8706 int &TripleDWord = ThreeAInputs ? ADWord : BDWord;
8707 int &OneInputDWord = ThreeAInputs ? BDWord : ADWord;
8708 int TripleInputOffset = ThreeAInputs ? AOffset : BOffset;
8709 ArrayRef<int> TripleInputs = ThreeAInputs ? AToAInputs : BToAInputs;
8710 int OneInput = ThreeAInputs ? BToAInputs[0] : AToAInputs[0];
8711 int TripleInputSum = 0 + 1 + 2 + 3 + (4 * TripleInputOffset);
8712 int TripleNonInputIdx =
8713 TripleInputSum - std::accumulate(TripleInputs.begin(), TripleInputs.end(), 0);
8714 TripleDWord = TripleNonInputIdx / 2;
8716 // We use xor with one to compute the adjacent DWord to whichever one the
8718 OneInputDWord = (OneInput / 2) ^ 1;
8720 // Check for one tricky case: We're fixing a 3<-1 or a 1<-3 shuffle for AToA
8721 // and BToA inputs. If there is also such a problem with the BToB and AToB
8722 // inputs, we don't try to fix it necessarily -- we'll recurse and see it in
8723 // the next pass. However, if we have a 2<-2 in the BToB and AToB inputs, it
8724 // is essential that we don't *create* a 3<-1 as then we might oscillate.
8725 if (BToBInputs.size() == 2 && AToBInputs.size() == 2) {
8726 // Compute how many inputs will be flipped by swapping these DWords. We
8728 // to balance this to ensure we don't form a 3-1 shuffle in the other
8730 int NumFlippedAToBInputs =
8731 std::count(AToBInputs.begin(), AToBInputs.end(), 2 * ADWord) +
8732 std::count(AToBInputs.begin(), AToBInputs.end(), 2 * ADWord + 1);
8733 int NumFlippedBToBInputs =
8734 std::count(BToBInputs.begin(), BToBInputs.end(), 2 * BDWord) +
8735 std::count(BToBInputs.begin(), BToBInputs.end(), 2 * BDWord + 1);
8736 if ((NumFlippedAToBInputs == 1 &&
8737 (NumFlippedBToBInputs == 0 || NumFlippedBToBInputs == 2)) ||
8738 (NumFlippedBToBInputs == 1 &&
8739 (NumFlippedAToBInputs == 0 || NumFlippedAToBInputs == 2))) {
8740 // We choose whether to fix the A half or B half based on whether that
8741 // half has zero flipped inputs. At zero, we may not be able to fix it
8742 // with that half. We also bias towards fixing the B half because that
8743 // will more commonly be the high half, and we have to bias one way.
8744 auto FixFlippedInputs = [&V, &DL, &Mask, &DAG](int PinnedIdx, int DWord,
8745 ArrayRef<int> Inputs) {
8746 int FixIdx = PinnedIdx ^ 1; // The adjacent slot to the pinned slot.
8747 bool IsFixIdxInput = std::find(Inputs.begin(), Inputs.end(),
8748 PinnedIdx ^ 1) != Inputs.end();
8749 // Determine whether the free index is in the flipped dword or the
8750 // unflipped dword based on where the pinned index is. We use this bit
8751 // in an xor to conditionally select the adjacent dword.
8752 int FixFreeIdx = 2 * (DWord ^ (PinnedIdx / 2 == DWord));
8753 bool IsFixFreeIdxInput = std::find(Inputs.begin(), Inputs.end(),
8754 FixFreeIdx) != Inputs.end();
8755 if (IsFixIdxInput == IsFixFreeIdxInput)
8757 IsFixFreeIdxInput = std::find(Inputs.begin(), Inputs.end(),
8758 FixFreeIdx) != Inputs.end();
8759 assert(IsFixIdxInput != IsFixFreeIdxInput &&
8760 "We need to be changing the number of flipped inputs!");
8761 int PSHUFHalfMask[] = {0, 1, 2, 3};
8762 std::swap(PSHUFHalfMask[FixFreeIdx % 4], PSHUFHalfMask[FixIdx % 4]);
8763 V = DAG.getNode(FixIdx < 4 ? X86ISD::PSHUFLW : X86ISD::PSHUFHW, DL,
8765 getV4X86ShuffleImm8ForMask(PSHUFHalfMask, DL, DAG));
8768 if (M != -1 && M == FixIdx)
8770 else if (M != -1 && M == FixFreeIdx)
8773 if (NumFlippedBToBInputs != 0) {
8775 BToAInputs.size() == 3 ? TripleNonInputIdx : OneInput;
8776 FixFlippedInputs(BPinnedIdx, BDWord, BToBInputs);
8778 assert(NumFlippedAToBInputs != 0 && "Impossible given predicates!");
8779 int APinnedIdx = ThreeAInputs ? TripleNonInputIdx : OneInput;
8780 FixFlippedInputs(APinnedIdx, ADWord, AToBInputs);
8785 int PSHUFDMask[] = {0, 1, 2, 3};
8786 PSHUFDMask[ADWord] = BDWord;
8787 PSHUFDMask[BDWord] = ADWord;
8790 DAG.getNode(X86ISD::PSHUFD, DL, PSHUFDVT, DAG.getBitcast(PSHUFDVT, V),
8791 getV4X86ShuffleImm8ForMask(PSHUFDMask, DL, DAG)));
8793 // Adjust the mask to match the new locations of A and B.
8795 if (M != -1 && M/2 == ADWord)
8796 M = 2 * BDWord + M % 2;
8797 else if (M != -1 && M/2 == BDWord)
8798 M = 2 * ADWord + M % 2;
8800 // Recurse back into this routine to re-compute state now that this isn't
8801 // a 3 and 1 problem.
8802 return lowerV8I16GeneralSingleInputVectorShuffle(DL, VT, V, Mask, Subtarget,
8805 if ((NumLToL == 3 && NumHToL == 1) || (NumLToL == 1 && NumHToL == 3))
8806 return balanceSides(LToLInputs, HToLInputs, HToHInputs, LToHInputs, 0, 4);
8807 else if ((NumHToH == 3 && NumLToH == 1) || (NumHToH == 1 && NumLToH == 3))
8808 return balanceSides(HToHInputs, LToHInputs, LToLInputs, HToLInputs, 4, 0);
8810 // At this point there are at most two inputs to the low and high halves from
8811 // each half. That means the inputs can always be grouped into dwords and
8812 // those dwords can then be moved to the correct half with a dword shuffle.
8813 // We use at most one low and one high word shuffle to collect these paired
8814 // inputs into dwords, and finally a dword shuffle to place them.
8815 int PSHUFLMask[4] = {-1, -1, -1, -1};
8816 int PSHUFHMask[4] = {-1, -1, -1, -1};
8817 int PSHUFDMask[4] = {-1, -1, -1, -1};
8819 // First fix the masks for all the inputs that are staying in their
8820 // original halves. This will then dictate the targets of the cross-half
8822 auto fixInPlaceInputs =
8823 [&PSHUFDMask](ArrayRef<int> InPlaceInputs, ArrayRef<int> IncomingInputs,
8824 MutableArrayRef<int> SourceHalfMask,
8825 MutableArrayRef<int> HalfMask, int HalfOffset) {
8826 if (InPlaceInputs.empty())
8828 if (InPlaceInputs.size() == 1) {
8829 SourceHalfMask[InPlaceInputs[0] - HalfOffset] =
8830 InPlaceInputs[0] - HalfOffset;
8831 PSHUFDMask[InPlaceInputs[0] / 2] = InPlaceInputs[0] / 2;
8834 if (IncomingInputs.empty()) {
8835 // Just fix all of the in place inputs.
8836 for (int Input : InPlaceInputs) {
8837 SourceHalfMask[Input - HalfOffset] = Input - HalfOffset;
8838 PSHUFDMask[Input / 2] = Input / 2;
8843 assert(InPlaceInputs.size() == 2 && "Cannot handle 3 or 4 inputs!");
8844 SourceHalfMask[InPlaceInputs[0] - HalfOffset] =
8845 InPlaceInputs[0] - HalfOffset;
8846 // Put the second input next to the first so that they are packed into
8847 // a dword. We find the adjacent index by toggling the low bit.
8848 int AdjIndex = InPlaceInputs[0] ^ 1;
8849 SourceHalfMask[AdjIndex - HalfOffset] = InPlaceInputs[1] - HalfOffset;
8850 std::replace(HalfMask.begin(), HalfMask.end(), InPlaceInputs[1], AdjIndex);
8851 PSHUFDMask[AdjIndex / 2] = AdjIndex / 2;
8853 fixInPlaceInputs(LToLInputs, HToLInputs, PSHUFLMask, LoMask, 0);
8854 fixInPlaceInputs(HToHInputs, LToHInputs, PSHUFHMask, HiMask, 4);
8856 // Now gather the cross-half inputs and place them into a free dword of
8857 // their target half.
8858 // FIXME: This operation could almost certainly be simplified dramatically to
8859 // look more like the 3-1 fixing operation.
8860 auto moveInputsToRightHalf = [&PSHUFDMask](
8861 MutableArrayRef<int> IncomingInputs, ArrayRef<int> ExistingInputs,
8862 MutableArrayRef<int> SourceHalfMask, MutableArrayRef<int> HalfMask,
8863 MutableArrayRef<int> FinalSourceHalfMask, int SourceOffset,
8865 auto isWordClobbered = [](ArrayRef<int> SourceHalfMask, int Word) {
8866 return SourceHalfMask[Word] != -1 && SourceHalfMask[Word] != Word;
8868 auto isDWordClobbered = [&isWordClobbered](ArrayRef<int> SourceHalfMask,
8870 int LowWord = Word & ~1;
8871 int HighWord = Word | 1;
8872 return isWordClobbered(SourceHalfMask, LowWord) ||
8873 isWordClobbered(SourceHalfMask, HighWord);
8876 if (IncomingInputs.empty())
8879 if (ExistingInputs.empty()) {
8880 // Map any dwords with inputs from them into the right half.
8881 for (int Input : IncomingInputs) {
8882 // If the source half mask maps over the inputs, turn those into
8883 // swaps and use the swapped lane.
8884 if (isWordClobbered(SourceHalfMask, Input - SourceOffset)) {
8885 if (SourceHalfMask[SourceHalfMask[Input - SourceOffset]] == -1) {
8886 SourceHalfMask[SourceHalfMask[Input - SourceOffset]] =
8887 Input - SourceOffset;
8888 // We have to swap the uses in our half mask in one sweep.
8889 for (int &M : HalfMask)
8890 if (M == SourceHalfMask[Input - SourceOffset] + SourceOffset)
8892 else if (M == Input)
8893 M = SourceHalfMask[Input - SourceOffset] + SourceOffset;
8895 assert(SourceHalfMask[SourceHalfMask[Input - SourceOffset]] ==
8896 Input - SourceOffset &&
8897 "Previous placement doesn't match!");
8899 // Note that this correctly re-maps both when we do a swap and when
8900 // we observe the other side of the swap above. We rely on that to
8901 // avoid swapping the members of the input list directly.
8902 Input = SourceHalfMask[Input - SourceOffset] + SourceOffset;
8905 // Map the input's dword into the correct half.
8906 if (PSHUFDMask[(Input - SourceOffset + DestOffset) / 2] == -1)
8907 PSHUFDMask[(Input - SourceOffset + DestOffset) / 2] = Input / 2;
8909 assert(PSHUFDMask[(Input - SourceOffset + DestOffset) / 2] ==
8911 "Previous placement doesn't match!");
8914 // And just directly shift any other-half mask elements to be same-half
8915 // as we will have mirrored the dword containing the element into the
8916 // same position within that half.
8917 for (int &M : HalfMask)
8918 if (M >= SourceOffset && M < SourceOffset + 4) {
8919 M = M - SourceOffset + DestOffset;
8920 assert(M >= 0 && "This should never wrap below zero!");
8925 // Ensure we have the input in a viable dword of its current half. This
8926 // is particularly tricky because the original position may be clobbered
8927 // by inputs being moved and *staying* in that half.
8928 if (IncomingInputs.size() == 1) {
8929 if (isWordClobbered(SourceHalfMask, IncomingInputs[0] - SourceOffset)) {
8930 int InputFixed = std::find(std::begin(SourceHalfMask),
8931 std::end(SourceHalfMask), -1) -
8932 std::begin(SourceHalfMask) + SourceOffset;
8933 SourceHalfMask[InputFixed - SourceOffset] =
8934 IncomingInputs[0] - SourceOffset;
8935 std::replace(HalfMask.begin(), HalfMask.end(), IncomingInputs[0],
8937 IncomingInputs[0] = InputFixed;
8939 } else if (IncomingInputs.size() == 2) {
8940 if (IncomingInputs[0] / 2 != IncomingInputs[1] / 2 ||
8941 isDWordClobbered(SourceHalfMask, IncomingInputs[0] - SourceOffset)) {
8942 // We have two non-adjacent or clobbered inputs we need to extract from
8943 // the source half. To do this, we need to map them into some adjacent
8944 // dword slot in the source mask.
8945 int InputsFixed[2] = {IncomingInputs[0] - SourceOffset,
8946 IncomingInputs[1] - SourceOffset};
8948 // If there is a free slot in the source half mask adjacent to one of
8949 // the inputs, place the other input in it. We use (Index XOR 1) to
8950 // compute an adjacent index.
8951 if (!isWordClobbered(SourceHalfMask, InputsFixed[0]) &&
8952 SourceHalfMask[InputsFixed[0] ^ 1] == -1) {
8953 SourceHalfMask[InputsFixed[0]] = InputsFixed[0];
8954 SourceHalfMask[InputsFixed[0] ^ 1] = InputsFixed[1];
8955 InputsFixed[1] = InputsFixed[0] ^ 1;
8956 } else if (!isWordClobbered(SourceHalfMask, InputsFixed[1]) &&
8957 SourceHalfMask[InputsFixed[1] ^ 1] == -1) {
8958 SourceHalfMask[InputsFixed[1]] = InputsFixed[1];
8959 SourceHalfMask[InputsFixed[1] ^ 1] = InputsFixed[0];
8960 InputsFixed[0] = InputsFixed[1] ^ 1;
8961 } else if (SourceHalfMask[2 * ((InputsFixed[0] / 2) ^ 1)] == -1 &&
8962 SourceHalfMask[2 * ((InputsFixed[0] / 2) ^ 1) + 1] == -1) {
8963 // The two inputs are in the same DWord but it is clobbered and the
8964 // adjacent DWord isn't used at all. Move both inputs to the free
8966 SourceHalfMask[2 * ((InputsFixed[0] / 2) ^ 1)] = InputsFixed[0];
8967 SourceHalfMask[2 * ((InputsFixed[0] / 2) ^ 1) + 1] = InputsFixed[1];
8968 InputsFixed[0] = 2 * ((InputsFixed[0] / 2) ^ 1);
8969 InputsFixed[1] = 2 * ((InputsFixed[0] / 2) ^ 1) + 1;
8971 // The only way we hit this point is if there is no clobbering
8972 // (because there are no off-half inputs to this half) and there is no
8973 // free slot adjacent to one of the inputs. In this case, we have to
8974 // swap an input with a non-input.
8975 for (int i = 0; i < 4; ++i)
8976 assert((SourceHalfMask[i] == -1 || SourceHalfMask[i] == i) &&
8977 "We can't handle any clobbers here!");
8978 assert(InputsFixed[1] != (InputsFixed[0] ^ 1) &&
8979 "Cannot have adjacent inputs here!");
8981 SourceHalfMask[InputsFixed[0] ^ 1] = InputsFixed[1];
8982 SourceHalfMask[InputsFixed[1]] = InputsFixed[0] ^ 1;
8984 // We also have to update the final source mask in this case because
8985 // it may need to undo the above swap.
8986 for (int &M : FinalSourceHalfMask)
8987 if (M == (InputsFixed[0] ^ 1) + SourceOffset)
8988 M = InputsFixed[1] + SourceOffset;
8989 else if (M == InputsFixed[1] + SourceOffset)
8990 M = (InputsFixed[0] ^ 1) + SourceOffset;
8992 InputsFixed[1] = InputsFixed[0] ^ 1;
8995 // Point everything at the fixed inputs.
8996 for (int &M : HalfMask)
8997 if (M == IncomingInputs[0])
8998 M = InputsFixed[0] + SourceOffset;
8999 else if (M == IncomingInputs[1])
9000 M = InputsFixed[1] + SourceOffset;
9002 IncomingInputs[0] = InputsFixed[0] + SourceOffset;
9003 IncomingInputs[1] = InputsFixed[1] + SourceOffset;
9006 llvm_unreachable("Unhandled input size!");
9009 // Now hoist the DWord down to the right half.
9010 int FreeDWord = (PSHUFDMask[DestOffset / 2] == -1 ? 0 : 1) + DestOffset / 2;
9011 assert(PSHUFDMask[FreeDWord] == -1 && "DWord not free");
9012 PSHUFDMask[FreeDWord] = IncomingInputs[0] / 2;
9013 for (int &M : HalfMask)
9014 for (int Input : IncomingInputs)
9016 M = FreeDWord * 2 + Input % 2;
9018 moveInputsToRightHalf(HToLInputs, LToLInputs, PSHUFHMask, LoMask, HiMask,
9019 /*SourceOffset*/ 4, /*DestOffset*/ 0);
9020 moveInputsToRightHalf(LToHInputs, HToHInputs, PSHUFLMask, HiMask, LoMask,
9021 /*SourceOffset*/ 0, /*DestOffset*/ 4);
9023 // Now enact all the shuffles we've computed to move the inputs into their
9025 if (!isNoopShuffleMask(PSHUFLMask))
9026 V = DAG.getNode(X86ISD::PSHUFLW, DL, VT, V,
9027 getV4X86ShuffleImm8ForMask(PSHUFLMask, DL, DAG));
9028 if (!isNoopShuffleMask(PSHUFHMask))
9029 V = DAG.getNode(X86ISD::PSHUFHW, DL, VT, V,
9030 getV4X86ShuffleImm8ForMask(PSHUFHMask, DL, DAG));
9031 if (!isNoopShuffleMask(PSHUFDMask))
9034 DAG.getNode(X86ISD::PSHUFD, DL, PSHUFDVT, DAG.getBitcast(PSHUFDVT, V),
9035 getV4X86ShuffleImm8ForMask(PSHUFDMask, DL, DAG)));
9037 // At this point, each half should contain all its inputs, and we can then
9038 // just shuffle them into their final position.
9039 assert(std::count_if(LoMask.begin(), LoMask.end(),
9040 [](int M) { return M >= 4; }) == 0 &&
9041 "Failed to lift all the high half inputs to the low mask!");
9042 assert(std::count_if(HiMask.begin(), HiMask.end(),
9043 [](int M) { return M >= 0 && M < 4; }) == 0 &&
9044 "Failed to lift all the low half inputs to the high mask!");
9046 // Do a half shuffle for the low mask.
9047 if (!isNoopShuffleMask(LoMask))
9048 V = DAG.getNode(X86ISD::PSHUFLW, DL, VT, V,
9049 getV4X86ShuffleImm8ForMask(LoMask, DL, DAG));
9051 // Do a half shuffle with the high mask after shifting its values down.
9052 for (int &M : HiMask)
9055 if (!isNoopShuffleMask(HiMask))
9056 V = DAG.getNode(X86ISD::PSHUFHW, DL, VT, V,
9057 getV4X86ShuffleImm8ForMask(HiMask, DL, DAG));
9062 /// \brief Helper to form a PSHUFB-based shuffle+blend.
9063 static SDValue lowerVectorShuffleAsPSHUFB(SDLoc DL, MVT VT, SDValue V1,
9064 SDValue V2, ArrayRef<int> Mask,
9065 SelectionDAG &DAG, bool &V1InUse,
9067 SmallBitVector Zeroable = computeZeroableShuffleElements(Mask, V1, V2);
9073 int Size = Mask.size();
9074 int Scale = 16 / Size;
9075 for (int i = 0; i < 16; ++i) {
9076 if (Mask[i / Scale] == -1) {
9077 V1Mask[i] = V2Mask[i] = DAG.getUNDEF(MVT::i8);
9079 const int ZeroMask = 0x80;
9080 int V1Idx = Mask[i / Scale] < Size ? Mask[i / Scale] * Scale + i % Scale
9082 int V2Idx = Mask[i / Scale] < Size
9084 : (Mask[i / Scale] - Size) * Scale + i % Scale;
9085 if (Zeroable[i / Scale])
9086 V1Idx = V2Idx = ZeroMask;
9087 V1Mask[i] = DAG.getConstant(V1Idx, DL, MVT::i8);
9088 V2Mask[i] = DAG.getConstant(V2Idx, DL, MVT::i8);
9089 V1InUse |= (ZeroMask != V1Idx);
9090 V2InUse |= (ZeroMask != V2Idx);
9095 V1 = DAG.getNode(X86ISD::PSHUFB, DL, MVT::v16i8,
9096 DAG.getBitcast(MVT::v16i8, V1),
9097 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v16i8, V1Mask));
9099 V2 = DAG.getNode(X86ISD::PSHUFB, DL, MVT::v16i8,
9100 DAG.getBitcast(MVT::v16i8, V2),
9101 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v16i8, V2Mask));
9103 // If we need shuffled inputs from both, blend the two.
9105 if (V1InUse && V2InUse)
9106 V = DAG.getNode(ISD::OR, DL, MVT::v16i8, V1, V2);
9108 V = V1InUse ? V1 : V2;
9110 // Cast the result back to the correct type.
9111 return DAG.getBitcast(VT, V);
9114 /// \brief Generic lowering of 8-lane i16 shuffles.
9116 /// This handles both single-input shuffles and combined shuffle/blends with
9117 /// two inputs. The single input shuffles are immediately delegated to
9118 /// a dedicated lowering routine.
9120 /// The blends are lowered in one of three fundamental ways. If there are few
9121 /// enough inputs, it delegates to a basic UNPCK-based strategy. If the shuffle
9122 /// of the input is significantly cheaper when lowered as an interleaving of
9123 /// the two inputs, try to interleave them. Otherwise, blend the low and high
9124 /// halves of the inputs separately (making them have relatively few inputs)
9125 /// and then concatenate them.
9126 static SDValue lowerV8I16VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
9127 const X86Subtarget *Subtarget,
9128 SelectionDAG &DAG) {
9130 assert(Op.getSimpleValueType() == MVT::v8i16 && "Bad shuffle type!");
9131 assert(V1.getSimpleValueType() == MVT::v8i16 && "Bad operand type!");
9132 assert(V2.getSimpleValueType() == MVT::v8i16 && "Bad operand type!");
9133 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
9134 ArrayRef<int> OrigMask = SVOp->getMask();
9135 int MaskStorage[8] = {OrigMask[0], OrigMask[1], OrigMask[2], OrigMask[3],
9136 OrigMask[4], OrigMask[5], OrigMask[6], OrigMask[7]};
9137 MutableArrayRef<int> Mask(MaskStorage);
9139 assert(Mask.size() == 8 && "Unexpected mask size for v8 shuffle!");
9141 // Whenever we can lower this as a zext, that instruction is strictly faster
9142 // than any alternative.
9143 if (SDValue ZExt = lowerVectorShuffleAsZeroOrAnyExtend(
9144 DL, MVT::v8i16, V1, V2, OrigMask, Subtarget, DAG))
9147 auto isV1 = [](int M) { return M >= 0 && M < 8; };
9149 auto isV2 = [](int M) { return M >= 8; };
9151 int NumV2Inputs = std::count_if(Mask.begin(), Mask.end(), isV2);
9153 if (NumV2Inputs == 0) {
9154 // Check for being able to broadcast a single element.
9155 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v8i16, V1,
9156 Mask, Subtarget, DAG))
9159 // Try to use shift instructions.
9161 lowerVectorShuffleAsShift(DL, MVT::v8i16, V1, V1, Mask, DAG))
9164 // Use dedicated unpack instructions for masks that match their pattern.
9166 lowerVectorShuffleWithUNPCK(DL, MVT::v8i16, Mask, V1, V2, DAG))
9169 // Try to use byte rotation instructions.
9170 if (SDValue Rotate = lowerVectorShuffleAsByteRotate(DL, MVT::v8i16, V1, V1,
9171 Mask, Subtarget, DAG))
9174 return lowerV8I16GeneralSingleInputVectorShuffle(DL, MVT::v8i16, V1, Mask,
9178 assert(std::any_of(Mask.begin(), Mask.end(), isV1) &&
9179 "All single-input shuffles should be canonicalized to be V1-input "
9182 // Try to use shift instructions.
9184 lowerVectorShuffleAsShift(DL, MVT::v8i16, V1, V2, Mask, DAG))
9187 // See if we can use SSE4A Extraction / Insertion.
9188 if (Subtarget->hasSSE4A())
9189 if (SDValue V = lowerVectorShuffleWithSSE4A(DL, MVT::v8i16, V1, V2, Mask, DAG))
9192 // There are special ways we can lower some single-element blends.
9193 if (NumV2Inputs == 1)
9194 if (SDValue V = lowerVectorShuffleAsElementInsertion(DL, MVT::v8i16, V1, V2,
9195 Mask, Subtarget, DAG))
9198 // We have different paths for blend lowering, but they all must use the
9199 // *exact* same predicate.
9200 bool IsBlendSupported = Subtarget->hasSSE41();
9201 if (IsBlendSupported)
9202 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v8i16, V1, V2, Mask,
9206 if (SDValue Masked =
9207 lowerVectorShuffleAsBitMask(DL, MVT::v8i16, V1, V2, Mask, DAG))
9210 // Use dedicated unpack instructions for masks that match their pattern.
9212 lowerVectorShuffleWithUNPCK(DL, MVT::v8i16, Mask, V1, V2, DAG))
9215 // Try to use byte rotation instructions.
9216 if (SDValue Rotate = lowerVectorShuffleAsByteRotate(
9217 DL, MVT::v8i16, V1, V2, Mask, Subtarget, DAG))
9220 if (SDValue BitBlend =
9221 lowerVectorShuffleAsBitBlend(DL, MVT::v8i16, V1, V2, Mask, DAG))
9224 if (SDValue Unpack = lowerVectorShuffleAsPermuteAndUnpack(DL, MVT::v8i16, V1,
9228 // If we can't directly blend but can use PSHUFB, that will be better as it
9229 // can both shuffle and set up the inefficient blend.
9230 if (!IsBlendSupported && Subtarget->hasSSSE3()) {
9231 bool V1InUse, V2InUse;
9232 return lowerVectorShuffleAsPSHUFB(DL, MVT::v8i16, V1, V2, Mask, DAG,
9236 // We can always bit-blend if we have to so the fallback strategy is to
9237 // decompose into single-input permutes and blends.
9238 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v8i16, V1, V2,
9242 /// \brief Check whether a compaction lowering can be done by dropping even
9243 /// elements and compute how many times even elements must be dropped.
9245 /// This handles shuffles which take every Nth element where N is a power of
9246 /// two. Example shuffle masks:
9248 /// N = 1: 0, 2, 4, 6, 8, 10, 12, 14, 0, 2, 4, 6, 8, 10, 12, 14
9249 /// N = 1: 0, 2, 4, 6, 8, 10, 12, 14, 16, 18, 20, 22, 24, 26, 28, 30
9250 /// N = 2: 0, 4, 8, 12, 0, 4, 8, 12, 0, 4, 8, 12, 0, 4, 8, 12
9251 /// N = 2: 0, 4, 8, 12, 16, 20, 24, 28, 0, 4, 8, 12, 16, 20, 24, 28
9252 /// N = 3: 0, 8, 0, 8, 0, 8, 0, 8, 0, 8, 0, 8, 0, 8, 0, 8
9253 /// N = 3: 0, 8, 16, 24, 0, 8, 16, 24, 0, 8, 16, 24, 0, 8, 16, 24
9255 /// Any of these lanes can of course be undef.
9257 /// This routine only supports N <= 3.
9258 /// FIXME: Evaluate whether either AVX or AVX-512 have any opportunities here
9261 /// \returns N above, or the number of times even elements must be dropped if
9262 /// there is such a number. Otherwise returns zero.
9263 static int canLowerByDroppingEvenElements(ArrayRef<int> Mask) {
9264 // Figure out whether we're looping over two inputs or just one.
9265 bool IsSingleInput = isSingleInputShuffleMask(Mask);
9267 // The modulus for the shuffle vector entries is based on whether this is
9268 // a single input or not.
9269 int ShuffleModulus = Mask.size() * (IsSingleInput ? 1 : 2);
9270 assert(isPowerOf2_32((uint32_t)ShuffleModulus) &&
9271 "We should only be called with masks with a power-of-2 size!");
9273 uint64_t ModMask = (uint64_t)ShuffleModulus - 1;
9275 // We track whether the input is viable for all power-of-2 strides 2^1, 2^2,
9276 // and 2^3 simultaneously. This is because we may have ambiguity with
9277 // partially undef inputs.
9278 bool ViableForN[3] = {true, true, true};
9280 for (int i = 0, e = Mask.size(); i < e; ++i) {
9281 // Ignore undef lanes, we'll optimistically collapse them to the pattern we
9286 bool IsAnyViable = false;
9287 for (unsigned j = 0; j != array_lengthof(ViableForN); ++j)
9288 if (ViableForN[j]) {
9291 // The shuffle mask must be equal to (i * 2^N) % M.
9292 if ((uint64_t)Mask[i] == (((uint64_t)i << N) & ModMask))
9295 ViableForN[j] = false;
9297 // Early exit if we exhaust the possible powers of two.
9302 for (unsigned j = 0; j != array_lengthof(ViableForN); ++j)
9306 // Return 0 as there is no viable power of two.
9310 /// \brief Generic lowering of v16i8 shuffles.
9312 /// This is a hybrid strategy to lower v16i8 vectors. It first attempts to
9313 /// detect any complexity reducing interleaving. If that doesn't help, it uses
9314 /// UNPCK to spread the i8 elements across two i16-element vectors, and uses
9315 /// the existing lowering for v8i16 blends on each half, finally PACK-ing them
9317 static SDValue lowerV16I8VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
9318 const X86Subtarget *Subtarget,
9319 SelectionDAG &DAG) {
9321 assert(Op.getSimpleValueType() == MVT::v16i8 && "Bad shuffle type!");
9322 assert(V1.getSimpleValueType() == MVT::v16i8 && "Bad operand type!");
9323 assert(V2.getSimpleValueType() == MVT::v16i8 && "Bad operand type!");
9324 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
9325 ArrayRef<int> Mask = SVOp->getMask();
9326 assert(Mask.size() == 16 && "Unexpected mask size for v16 shuffle!");
9328 // Try to use shift instructions.
9330 lowerVectorShuffleAsShift(DL, MVT::v16i8, V1, V2, Mask, DAG))
9333 // Try to use byte rotation instructions.
9334 if (SDValue Rotate = lowerVectorShuffleAsByteRotate(
9335 DL, MVT::v16i8, V1, V2, Mask, Subtarget, DAG))
9338 // Try to use a zext lowering.
9339 if (SDValue ZExt = lowerVectorShuffleAsZeroOrAnyExtend(
9340 DL, MVT::v16i8, V1, V2, Mask, Subtarget, DAG))
9343 // See if we can use SSE4A Extraction / Insertion.
9344 if (Subtarget->hasSSE4A())
9345 if (SDValue V = lowerVectorShuffleWithSSE4A(DL, MVT::v16i8, V1, V2, Mask, DAG))
9349 std::count_if(Mask.begin(), Mask.end(), [](int M) { return M >= 16; });
9351 // For single-input shuffles, there are some nicer lowering tricks we can use.
9352 if (NumV2Elements == 0) {
9353 // Check for being able to broadcast a single element.
9354 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v16i8, V1,
9355 Mask, Subtarget, DAG))
9358 // Check whether we can widen this to an i16 shuffle by duplicating bytes.
9359 // Notably, this handles splat and partial-splat shuffles more efficiently.
9360 // However, it only makes sense if the pre-duplication shuffle simplifies
9361 // things significantly. Currently, this means we need to be able to
9362 // express the pre-duplication shuffle as an i16 shuffle.
9364 // FIXME: We should check for other patterns which can be widened into an
9365 // i16 shuffle as well.
9366 auto canWidenViaDuplication = [](ArrayRef<int> Mask) {
9367 for (int i = 0; i < 16; i += 2)
9368 if (Mask[i] != -1 && Mask[i + 1] != -1 && Mask[i] != Mask[i + 1])
9373 auto tryToWidenViaDuplication = [&]() -> SDValue {
9374 if (!canWidenViaDuplication(Mask))
9376 SmallVector<int, 4> LoInputs;
9377 std::copy_if(Mask.begin(), Mask.end(), std::back_inserter(LoInputs),
9378 [](int M) { return M >= 0 && M < 8; });
9379 std::sort(LoInputs.begin(), LoInputs.end());
9380 LoInputs.erase(std::unique(LoInputs.begin(), LoInputs.end()),
9382 SmallVector<int, 4> HiInputs;
9383 std::copy_if(Mask.begin(), Mask.end(), std::back_inserter(HiInputs),
9384 [](int M) { return M >= 8; });
9385 std::sort(HiInputs.begin(), HiInputs.end());
9386 HiInputs.erase(std::unique(HiInputs.begin(), HiInputs.end()),
9389 bool TargetLo = LoInputs.size() >= HiInputs.size();
9390 ArrayRef<int> InPlaceInputs = TargetLo ? LoInputs : HiInputs;
9391 ArrayRef<int> MovingInputs = TargetLo ? HiInputs : LoInputs;
9393 int PreDupI16Shuffle[] = {-1, -1, -1, -1, -1, -1, -1, -1};
9394 SmallDenseMap<int, int, 8> LaneMap;
9395 for (int I : InPlaceInputs) {
9396 PreDupI16Shuffle[I/2] = I/2;
9399 int j = TargetLo ? 0 : 4, je = j + 4;
9400 for (int i = 0, ie = MovingInputs.size(); i < ie; ++i) {
9401 // Check if j is already a shuffle of this input. This happens when
9402 // there are two adjacent bytes after we move the low one.
9403 if (PreDupI16Shuffle[j] != MovingInputs[i] / 2) {
9404 // If we haven't yet mapped the input, search for a slot into which
9406 while (j < je && PreDupI16Shuffle[j] != -1)
9410 // We can't place the inputs into a single half with a simple i16 shuffle, so bail.
9413 // Map this input with the i16 shuffle.
9414 PreDupI16Shuffle[j] = MovingInputs[i] / 2;
9417 // Update the lane map based on the mapping we ended up with.
9418 LaneMap[MovingInputs[i]] = 2 * j + MovingInputs[i] % 2;
9420 V1 = DAG.getBitcast(
9422 DAG.getVectorShuffle(MVT::v8i16, DL, DAG.getBitcast(MVT::v8i16, V1),
9423 DAG.getUNDEF(MVT::v8i16), PreDupI16Shuffle));
9425 // Unpack the bytes to form the i16s that will be shuffled into place.
9426 V1 = DAG.getNode(TargetLo ? X86ISD::UNPCKL : X86ISD::UNPCKH, DL,
9427 MVT::v16i8, V1, V1);
9429 int PostDupI16Shuffle[8] = {-1, -1, -1, -1, -1, -1, -1, -1};
9430 for (int i = 0; i < 16; ++i)
9431 if (Mask[i] != -1) {
9432 int MappedMask = LaneMap[Mask[i]] - (TargetLo ? 0 : 8);
9433 assert(MappedMask < 8 && "Invalid v8 shuffle mask!");
9434 if (PostDupI16Shuffle[i / 2] == -1)
9435 PostDupI16Shuffle[i / 2] = MappedMask;
9437 assert(PostDupI16Shuffle[i / 2] == MappedMask &&
9438 "Conflicting entrties in the original shuffle!");
9440 return DAG.getBitcast(
9442 DAG.getVectorShuffle(MVT::v8i16, DL, DAG.getBitcast(MVT::v8i16, V1),
9443 DAG.getUNDEF(MVT::v8i16), PostDupI16Shuffle));
9445 if (SDValue V = tryToWidenViaDuplication())
9449 if (SDValue Masked =
9450 lowerVectorShuffleAsBitMask(DL, MVT::v16i8, V1, V2, Mask, DAG))
9453 // Use dedicated unpack instructions for masks that match their pattern.
9455 lowerVectorShuffleWithUNPCK(DL, MVT::v16i8, Mask, V1, V2, DAG))
9458 // Check for SSSE3 which lets us lower all v16i8 shuffles much more directly
9459 // with PSHUFB. It is important to do this before we attempt to generate any
9460 // blends but after all of the single-input lowerings. If the single input
9461 // lowerings can find an instruction sequence that is faster than a PSHUFB, we
9462 // want to preserve that and we can DAG combine any longer sequences into
9463 // a PSHUFB in the end. But once we start blending from multiple inputs,
9464 // the complexity of DAG combining bad patterns back into PSHUFB is too high,
9465 // and there are *very* few patterns that would actually be faster than the
9466 // PSHUFB approach because of its ability to zero lanes.
9468 // FIXME: The only exceptions to the above are blends which are exact
9469 // interleavings with direct instructions supporting them. We currently don't
9470 // handle those well here.
9471 if (Subtarget->hasSSSE3()) {
9472 bool V1InUse = false;
9473 bool V2InUse = false;
9475 SDValue PSHUFB = lowerVectorShuffleAsPSHUFB(DL, MVT::v16i8, V1, V2, Mask,
9476 DAG, V1InUse, V2InUse);
9478 // If both V1 and V2 are in use and we can use a direct blend or an unpack,
9479 // do so. This avoids using them to handle blends-with-zero which is
9480 // important as a single pshufb is significantly faster for that.
9481 if (V1InUse && V2InUse) {
9482 if (Subtarget->hasSSE41())
9483 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v16i8, V1, V2,
9484 Mask, Subtarget, DAG))
9487 // We can use an unpack to do the blending rather than an or in some
9488 // cases. Even though the or may be (very minorly) more efficient, we
9489 // preference this lowering because there are common cases where part of
9490 // the complexity of the shuffles goes away when we do the final blend as
9492 // FIXME: It might be worth trying to detect if the unpack-feeding
9493 // shuffles will both be pshufb, in which case we shouldn't bother with
9495 if (SDValue Unpack = lowerVectorShuffleAsPermuteAndUnpack(
9496 DL, MVT::v16i8, V1, V2, Mask, DAG))
9503 // There are special ways we can lower some single-element blends.
9504 if (NumV2Elements == 1)
9505 if (SDValue V = lowerVectorShuffleAsElementInsertion(DL, MVT::v16i8, V1, V2,
9506 Mask, Subtarget, DAG))
9509 if (SDValue BitBlend =
9510 lowerVectorShuffleAsBitBlend(DL, MVT::v16i8, V1, V2, Mask, DAG))
9513 // Check whether a compaction lowering can be done. This handles shuffles
9514 // which take every Nth element for some even N. See the helper function for
9517 // We special case these as they can be particularly efficiently handled with
9518 // the PACKUSB instruction on x86 and they show up in common patterns of
9519 // rearranging bytes to truncate wide elements.
9520 if (int NumEvenDrops = canLowerByDroppingEvenElements(Mask)) {
9521 // NumEvenDrops is the power of two stride of the elements. Another way of
9522 // thinking about it is that we need to drop the even elements this many
9523 // times to get the original input.
9524 bool IsSingleInput = isSingleInputShuffleMask(Mask);
9526 // First we need to zero all the dropped bytes.
9527 assert(NumEvenDrops <= 3 &&
9528 "No support for dropping even elements more than 3 times.");
9529 // We use the mask type to pick which bytes are preserved based on how many
9530 // elements are dropped.
9531 MVT MaskVTs[] = { MVT::v8i16, MVT::v4i32, MVT::v2i64 };
9532 SDValue ByteClearMask = DAG.getBitcast(
9533 MVT::v16i8, DAG.getConstant(0xFF, DL, MaskVTs[NumEvenDrops - 1]));
9534 V1 = DAG.getNode(ISD::AND, DL, MVT::v16i8, V1, ByteClearMask);
9536 V2 = DAG.getNode(ISD::AND, DL, MVT::v16i8, V2, ByteClearMask);
9538 // Now pack things back together.
9539 V1 = DAG.getBitcast(MVT::v8i16, V1);
9540 V2 = IsSingleInput ? V1 : DAG.getBitcast(MVT::v8i16, V2);
9541 SDValue Result = DAG.getNode(X86ISD::PACKUS, DL, MVT::v16i8, V1, V2);
9542 for (int i = 1; i < NumEvenDrops; ++i) {
9543 Result = DAG.getBitcast(MVT::v8i16, Result);
9544 Result = DAG.getNode(X86ISD::PACKUS, DL, MVT::v16i8, Result, Result);
9550 // Handle multi-input cases by blending single-input shuffles.
9551 if (NumV2Elements > 0)
9552 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v16i8, V1, V2,
9555 // The fallback path for single-input shuffles widens this into two v8i16
9556 // vectors with unpacks, shuffles those, and then pulls them back together
9560 int LoBlendMask[8] = {-1, -1, -1, -1, -1, -1, -1, -1};
9561 int HiBlendMask[8] = {-1, -1, -1, -1, -1, -1, -1, -1};
9562 for (int i = 0; i < 16; ++i)
9564 (i < 8 ? LoBlendMask[i] : HiBlendMask[i % 8]) = Mask[i];
9566 SDValue Zero = getZeroVector(MVT::v8i16, Subtarget, DAG, DL);
9568 SDValue VLoHalf, VHiHalf;
9569 // Check if any of the odd lanes in the v16i8 are used. If not, we can mask
9570 // them out and avoid using UNPCK{L,H} to extract the elements of V as
9572 if (std::none_of(std::begin(LoBlendMask), std::end(LoBlendMask),
9573 [](int M) { return M >= 0 && M % 2 == 1; }) &&
9574 std::none_of(std::begin(HiBlendMask), std::end(HiBlendMask),
9575 [](int M) { return M >= 0 && M % 2 == 1; })) {
9576 // Use a mask to drop the high bytes.
9577 VLoHalf = DAG.getBitcast(MVT::v8i16, V);
9578 VLoHalf = DAG.getNode(ISD::AND, DL, MVT::v8i16, VLoHalf,
9579 DAG.getConstant(0x00FF, DL, MVT::v8i16));
9581 // This will be a single vector shuffle instead of a blend so nuke VHiHalf.
9582 VHiHalf = DAG.getUNDEF(MVT::v8i16);
9584 // Squash the masks to point directly into VLoHalf.
9585 for (int &M : LoBlendMask)
9588 for (int &M : HiBlendMask)
9592 // Otherwise just unpack the low half of V into VLoHalf and the high half into
9593 // VHiHalf so that we can blend them as i16s.
9594 VLoHalf = DAG.getBitcast(
9595 MVT::v8i16, DAG.getNode(X86ISD::UNPCKL, DL, MVT::v16i8, V, Zero));
9596 VHiHalf = DAG.getBitcast(
9597 MVT::v8i16, DAG.getNode(X86ISD::UNPCKH, DL, MVT::v16i8, V, Zero));
9600 SDValue LoV = DAG.getVectorShuffle(MVT::v8i16, DL, VLoHalf, VHiHalf, LoBlendMask);
9601 SDValue HiV = DAG.getVectorShuffle(MVT::v8i16, DL, VLoHalf, VHiHalf, HiBlendMask);
9603 return DAG.getNode(X86ISD::PACKUS, DL, MVT::v16i8, LoV, HiV);
9606 /// \brief Dispatching routine to lower various 128-bit x86 vector shuffles.
9608 /// This routine breaks down the specific type of 128-bit shuffle and
9609 /// dispatches to the lowering routines accordingly.
9610 static SDValue lower128BitVectorShuffle(SDValue Op, SDValue V1, SDValue V2,
9611 MVT VT, const X86Subtarget *Subtarget,
9612 SelectionDAG &DAG) {
9613 switch (VT.SimpleTy) {
9615 return lowerV2I64VectorShuffle(Op, V1, V2, Subtarget, DAG);
9617 return lowerV2F64VectorShuffle(Op, V1, V2, Subtarget, DAG);
9619 return lowerV4I32VectorShuffle(Op, V1, V2, Subtarget, DAG);
9621 return lowerV4F32VectorShuffle(Op, V1, V2, Subtarget, DAG);
9623 return lowerV8I16VectorShuffle(Op, V1, V2, Subtarget, DAG);
9625 return lowerV16I8VectorShuffle(Op, V1, V2, Subtarget, DAG);
9628 llvm_unreachable("Unimplemented!");
9632 /// \brief Helper function to test whether a shuffle mask could be
9633 /// simplified by widening the elements being shuffled.
9635 /// Appends the mask for wider elements in WidenedMask if valid. Otherwise
9636 /// leaves it in an unspecified state.
9638 /// NOTE: This must handle normal vector shuffle masks and *target* vector
9639 /// shuffle masks. The latter have the special property of a '-2' representing
9640 /// a zero-ed lane of a vector.
9641 static bool canWidenShuffleElements(ArrayRef<int> Mask,
9642 SmallVectorImpl<int> &WidenedMask) {
9643 for (int i = 0, Size = Mask.size(); i < Size; i += 2) {
9644 // If both elements are undef, its trivial.
9645 if (Mask[i] == SM_SentinelUndef && Mask[i + 1] == SM_SentinelUndef) {
9646 WidenedMask.push_back(SM_SentinelUndef);
9650 // Check for an undef mask and a mask value properly aligned to fit with
9651 // a pair of values. If we find such a case, use the non-undef mask's value.
9652 if (Mask[i] == SM_SentinelUndef && Mask[i + 1] >= 0 && Mask[i + 1] % 2 == 1) {
9653 WidenedMask.push_back(Mask[i + 1] / 2);
9656 if (Mask[i + 1] == SM_SentinelUndef && Mask[i] >= 0 && Mask[i] % 2 == 0) {
9657 WidenedMask.push_back(Mask[i] / 2);
9661 // When zeroing, we need to spread the zeroing across both lanes to widen.
9662 if (Mask[i] == SM_SentinelZero || Mask[i + 1] == SM_SentinelZero) {
9663 if ((Mask[i] == SM_SentinelZero || Mask[i] == SM_SentinelUndef) &&
9664 (Mask[i + 1] == SM_SentinelZero || Mask[i + 1] == SM_SentinelUndef)) {
9665 WidenedMask.push_back(SM_SentinelZero);
9671 // Finally check if the two mask values are adjacent and aligned with
9673 if (Mask[i] != SM_SentinelUndef && Mask[i] % 2 == 0 && Mask[i] + 1 == Mask[i + 1]) {
9674 WidenedMask.push_back(Mask[i] / 2);
9678 // Otherwise we can't safely widen the elements used in this shuffle.
9681 assert(WidenedMask.size() == Mask.size() / 2 &&
9682 "Incorrect size of mask after widening the elements!");
9687 /// \brief Generic routine to split vector shuffle into half-sized shuffles.
9689 /// This routine just extracts two subvectors, shuffles them independently, and
9690 /// then concatenates them back together. This should work effectively with all
9691 /// AVX vector shuffle types.
9692 static SDValue splitAndLowerVectorShuffle(SDLoc DL, MVT VT, SDValue V1,
9693 SDValue V2, ArrayRef<int> Mask,
9694 SelectionDAG &DAG) {
9695 assert(VT.getSizeInBits() >= 256 &&
9696 "Only for 256-bit or wider vector shuffles!");
9697 assert(V1.getSimpleValueType() == VT && "Bad operand type!");
9698 assert(V2.getSimpleValueType() == VT && "Bad operand type!");
9700 ArrayRef<int> LoMask = Mask.slice(0, Mask.size() / 2);
9701 ArrayRef<int> HiMask = Mask.slice(Mask.size() / 2);
9703 int NumElements = VT.getVectorNumElements();
9704 int SplitNumElements = NumElements / 2;
9705 MVT ScalarVT = VT.getScalarType();
9706 MVT SplitVT = MVT::getVectorVT(ScalarVT, NumElements / 2);
9708 // Rather than splitting build-vectors, just build two narrower build
9709 // vectors. This helps shuffling with splats and zeros.
9710 auto SplitVector = [&](SDValue V) {
9711 while (V.getOpcode() == ISD::BITCAST)
9712 V = V->getOperand(0);
9714 MVT OrigVT = V.getSimpleValueType();
9715 int OrigNumElements = OrigVT.getVectorNumElements();
9716 int OrigSplitNumElements = OrigNumElements / 2;
9717 MVT OrigScalarVT = OrigVT.getScalarType();
9718 MVT OrigSplitVT = MVT::getVectorVT(OrigScalarVT, OrigNumElements / 2);
9722 auto *BV = dyn_cast<BuildVectorSDNode>(V);
9724 LoV = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, OrigSplitVT, V,
9725 DAG.getIntPtrConstant(0, DL));
9726 HiV = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, OrigSplitVT, V,
9727 DAG.getIntPtrConstant(OrigSplitNumElements, DL));
9730 SmallVector<SDValue, 16> LoOps, HiOps;
9731 for (int i = 0; i < OrigSplitNumElements; ++i) {
9732 LoOps.push_back(BV->getOperand(i));
9733 HiOps.push_back(BV->getOperand(i + OrigSplitNumElements));
9735 LoV = DAG.getNode(ISD::BUILD_VECTOR, DL, OrigSplitVT, LoOps);
9736 HiV = DAG.getNode(ISD::BUILD_VECTOR, DL, OrigSplitVT, HiOps);
9738 return std::make_pair(DAG.getBitcast(SplitVT, LoV),
9739 DAG.getBitcast(SplitVT, HiV));
9742 SDValue LoV1, HiV1, LoV2, HiV2;
9743 std::tie(LoV1, HiV1) = SplitVector(V1);
9744 std::tie(LoV2, HiV2) = SplitVector(V2);
9746 // Now create two 4-way blends of these half-width vectors.
9747 auto HalfBlend = [&](ArrayRef<int> HalfMask) {
9748 bool UseLoV1 = false, UseHiV1 = false, UseLoV2 = false, UseHiV2 = false;
9749 SmallVector<int, 32> V1BlendMask, V2BlendMask, BlendMask;
9750 for (int i = 0; i < SplitNumElements; ++i) {
9751 int M = HalfMask[i];
9752 if (M >= NumElements) {
9753 if (M >= NumElements + SplitNumElements)
9757 V2BlendMask.push_back(M - NumElements);
9758 V1BlendMask.push_back(-1);
9759 BlendMask.push_back(SplitNumElements + i);
9760 } else if (M >= 0) {
9761 if (M >= SplitNumElements)
9765 V2BlendMask.push_back(-1);
9766 V1BlendMask.push_back(M);
9767 BlendMask.push_back(i);
9769 V2BlendMask.push_back(-1);
9770 V1BlendMask.push_back(-1);
9771 BlendMask.push_back(-1);
9775 // Because the lowering happens after all combining takes place, we need to
9776 // manually combine these blend masks as much as possible so that we create
9777 // a minimal number of high-level vector shuffle nodes.
9779 // First try just blending the halves of V1 or V2.
9780 if (!UseLoV1 && !UseHiV1 && !UseLoV2 && !UseHiV2)
9781 return DAG.getUNDEF(SplitVT);
9782 if (!UseLoV2 && !UseHiV2)
9783 return DAG.getVectorShuffle(SplitVT, DL, LoV1, HiV1, V1BlendMask);
9784 if (!UseLoV1 && !UseHiV1)
9785 return DAG.getVectorShuffle(SplitVT, DL, LoV2, HiV2, V2BlendMask);
9787 SDValue V1Blend, V2Blend;
9788 if (UseLoV1 && UseHiV1) {
9790 DAG.getVectorShuffle(SplitVT, DL, LoV1, HiV1, V1BlendMask);
9792 // We only use half of V1 so map the usage down into the final blend mask.
9793 V1Blend = UseLoV1 ? LoV1 : HiV1;
9794 for (int i = 0; i < SplitNumElements; ++i)
9795 if (BlendMask[i] >= 0 && BlendMask[i] < SplitNumElements)
9796 BlendMask[i] = V1BlendMask[i] - (UseLoV1 ? 0 : SplitNumElements);
9798 if (UseLoV2 && UseHiV2) {
9800 DAG.getVectorShuffle(SplitVT, DL, LoV2, HiV2, V2BlendMask);
9802 // We only use half of V2 so map the usage down into the final blend mask.
9803 V2Blend = UseLoV2 ? LoV2 : HiV2;
9804 for (int i = 0; i < SplitNumElements; ++i)
9805 if (BlendMask[i] >= SplitNumElements)
9806 BlendMask[i] = V2BlendMask[i] + (UseLoV2 ? SplitNumElements : 0);
9808 return DAG.getVectorShuffle(SplitVT, DL, V1Blend, V2Blend, BlendMask);
9810 SDValue Lo = HalfBlend(LoMask);
9811 SDValue Hi = HalfBlend(HiMask);
9812 return DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, Lo, Hi);
9815 /// \brief Either split a vector in halves or decompose the shuffles and the
9818 /// This is provided as a good fallback for many lowerings of non-single-input
9819 /// shuffles with more than one 128-bit lane. In those cases, we want to select
9820 /// between splitting the shuffle into 128-bit components and stitching those
9821 /// back together vs. extracting the single-input shuffles and blending those
9823 static SDValue lowerVectorShuffleAsSplitOrBlend(SDLoc DL, MVT VT, SDValue V1,
9824 SDValue V2, ArrayRef<int> Mask,
9825 SelectionDAG &DAG) {
9826 assert(!isSingleInputShuffleMask(Mask) && "This routine must not be used to "
9827 "lower single-input shuffles as it "
9828 "could then recurse on itself.");
9829 int Size = Mask.size();
9831 // If this can be modeled as a broadcast of two elements followed by a blend,
9832 // prefer that lowering. This is especially important because broadcasts can
9833 // often fold with memory operands.
9834 auto DoBothBroadcast = [&] {
9835 int V1BroadcastIdx = -1, V2BroadcastIdx = -1;
9838 if (V2BroadcastIdx == -1)
9839 V2BroadcastIdx = M - Size;
9840 else if (M - Size != V2BroadcastIdx)
9842 } else if (M >= 0) {
9843 if (V1BroadcastIdx == -1)
9845 else if (M != V1BroadcastIdx)
9850 if (DoBothBroadcast())
9851 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, VT, V1, V2, Mask,
9854 // If the inputs all stem from a single 128-bit lane of each input, then we
9855 // split them rather than blending because the split will decompose to
9856 // unusually few instructions.
9857 int LaneCount = VT.getSizeInBits() / 128;
9858 int LaneSize = Size / LaneCount;
9859 SmallBitVector LaneInputs[2];
9860 LaneInputs[0].resize(LaneCount, false);
9861 LaneInputs[1].resize(LaneCount, false);
9862 for (int i = 0; i < Size; ++i)
9864 LaneInputs[Mask[i] / Size][(Mask[i] % Size) / LaneSize] = true;
9865 if (LaneInputs[0].count() <= 1 && LaneInputs[1].count() <= 1)
9866 return splitAndLowerVectorShuffle(DL, VT, V1, V2, Mask, DAG);
9868 // Otherwise, just fall back to decomposed shuffles and a blend. This requires
9869 // that the decomposed single-input shuffles don't end up here.
9870 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, VT, V1, V2, Mask, DAG);
9873 /// \brief Lower a vector shuffle crossing multiple 128-bit lanes as
9874 /// a permutation and blend of those lanes.
9876 /// This essentially blends the out-of-lane inputs to each lane into the lane
9877 /// from a permuted copy of the vector. This lowering strategy results in four
9878 /// instructions in the worst case for a single-input cross lane shuffle which
9879 /// is lower than any other fully general cross-lane shuffle strategy I'm aware
9880 /// of. Special cases for each particular shuffle pattern should be handled
9881 /// prior to trying this lowering.
9882 static SDValue lowerVectorShuffleAsLanePermuteAndBlend(SDLoc DL, MVT VT,
9883 SDValue V1, SDValue V2,
9885 SelectionDAG &DAG) {
9886 // FIXME: This should probably be generalized for 512-bit vectors as well.
9887 assert(VT.getSizeInBits() == 256 && "Only for 256-bit vector shuffles!");
9888 int LaneSize = Mask.size() / 2;
9890 // If there are only inputs from one 128-bit lane, splitting will in fact be
9891 // less expensive. The flags track whether the given lane contains an element
9892 // that crosses to another lane.
9893 bool LaneCrossing[2] = {false, false};
9894 for (int i = 0, Size = Mask.size(); i < Size; ++i)
9895 if (Mask[i] >= 0 && (Mask[i] % Size) / LaneSize != i / LaneSize)
9896 LaneCrossing[(Mask[i] % Size) / LaneSize] = true;
9897 if (!LaneCrossing[0] || !LaneCrossing[1])
9898 return splitAndLowerVectorShuffle(DL, VT, V1, V2, Mask, DAG);
9900 if (isSingleInputShuffleMask(Mask)) {
9901 SmallVector<int, 32> FlippedBlendMask;
9902 for (int i = 0, Size = Mask.size(); i < Size; ++i)
9903 FlippedBlendMask.push_back(
9904 Mask[i] < 0 ? -1 : (((Mask[i] % Size) / LaneSize == i / LaneSize)
9906 : Mask[i] % LaneSize +
9907 (i / LaneSize) * LaneSize + Size));
9909 // Flip the vector, and blend the results which should now be in-lane. The
9910 // VPERM2X128 mask uses the low 2 bits for the low source and bits 4 and
9911 // 5 for the high source. The value 3 selects the high half of source 2 and
9912 // the value 2 selects the low half of source 2. We only use source 2 to
9913 // allow folding it into a memory operand.
9914 unsigned PERMMask = 3 | 2 << 4;
9915 SDValue Flipped = DAG.getNode(X86ISD::VPERM2X128, DL, VT, DAG.getUNDEF(VT),
9916 V1, DAG.getConstant(PERMMask, DL, MVT::i8));
9917 return DAG.getVectorShuffle(VT, DL, V1, Flipped, FlippedBlendMask);
9920 // This now reduces to two single-input shuffles of V1 and V2 which at worst
9921 // will be handled by the above logic and a blend of the results, much like
9922 // other patterns in AVX.
9923 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, VT, V1, V2, Mask, DAG);
9926 /// \brief Handle lowering 2-lane 128-bit shuffles.
9927 static SDValue lowerV2X128VectorShuffle(SDLoc DL, MVT VT, SDValue V1,
9928 SDValue V2, ArrayRef<int> Mask,
9929 const X86Subtarget *Subtarget,
9930 SelectionDAG &DAG) {
9931 // TODO: If minimizing size and one of the inputs is a zero vector and the
9932 // the zero vector has only one use, we could use a VPERM2X128 to save the
9933 // instruction bytes needed to explicitly generate the zero vector.
9935 // Blends are faster and handle all the non-lane-crossing cases.
9936 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, VT, V1, V2, Mask,
9940 bool IsV1Zero = ISD::isBuildVectorAllZeros(V1.getNode());
9941 bool IsV2Zero = ISD::isBuildVectorAllZeros(V2.getNode());
9943 // If either input operand is a zero vector, use VPERM2X128 because its mask
9944 // allows us to replace the zero input with an implicit zero.
9945 if (!IsV1Zero && !IsV2Zero) {
9946 // Check for patterns which can be matched with a single insert of a 128-bit
9948 bool OnlyUsesV1 = isShuffleEquivalent(V1, V2, Mask, {0, 1, 0, 1});
9949 if (OnlyUsesV1 || isShuffleEquivalent(V1, V2, Mask, {0, 1, 4, 5})) {
9950 MVT SubVT = MVT::getVectorVT(VT.getVectorElementType(),
9951 VT.getVectorNumElements() / 2);
9952 SDValue LoV = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, SubVT, V1,
9953 DAG.getIntPtrConstant(0, DL));
9954 SDValue HiV = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, SubVT,
9955 OnlyUsesV1 ? V1 : V2,
9956 DAG.getIntPtrConstant(0, DL));
9957 return DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, LoV, HiV);
9961 // Otherwise form a 128-bit permutation. After accounting for undefs,
9962 // convert the 64-bit shuffle mask selection values into 128-bit
9963 // selection bits by dividing the indexes by 2 and shifting into positions
9964 // defined by a vperm2*128 instruction's immediate control byte.
9966 // The immediate permute control byte looks like this:
9967 // [1:0] - select 128 bits from sources for low half of destination
9969 // [3] - zero low half of destination
9970 // [5:4] - select 128 bits from sources for high half of destination
9972 // [7] - zero high half of destination
9974 int MaskLO = Mask[0];
9975 if (MaskLO == SM_SentinelUndef)
9976 MaskLO = Mask[1] == SM_SentinelUndef ? 0 : Mask[1];
9978 int MaskHI = Mask[2];
9979 if (MaskHI == SM_SentinelUndef)
9980 MaskHI = Mask[3] == SM_SentinelUndef ? 0 : Mask[3];
9982 unsigned PermMask = MaskLO / 2 | (MaskHI / 2) << 4;
9984 // If either input is a zero vector, replace it with an undef input.
9985 // Shuffle mask values < 4 are selecting elements of V1.
9986 // Shuffle mask values >= 4 are selecting elements of V2.
9987 // Adjust each half of the permute mask by clearing the half that was
9988 // selecting the zero vector and setting the zero mask bit.
9990 V1 = DAG.getUNDEF(VT);
9992 PermMask = (PermMask & 0xf0) | 0x08;
9994 PermMask = (PermMask & 0x0f) | 0x80;
9997 V2 = DAG.getUNDEF(VT);
9999 PermMask = (PermMask & 0xf0) | 0x08;
10001 PermMask = (PermMask & 0x0f) | 0x80;
10004 return DAG.getNode(X86ISD::VPERM2X128, DL, VT, V1, V2,
10005 DAG.getConstant(PermMask, DL, MVT::i8));
10008 /// \brief Lower a vector shuffle by first fixing the 128-bit lanes and then
10009 /// shuffling each lane.
10011 /// This will only succeed when the result of fixing the 128-bit lanes results
10012 /// in a single-input non-lane-crossing shuffle with a repeating shuffle mask in
10013 /// each 128-bit lanes. This handles many cases where we can quickly blend away
10014 /// the lane crosses early and then use simpler shuffles within each lane.
10016 /// FIXME: It might be worthwhile at some point to support this without
10017 /// requiring the 128-bit lane-relative shuffles to be repeating, but currently
10018 /// in x86 only floating point has interesting non-repeating shuffles, and even
10019 /// those are still *marginally* more expensive.
10020 static SDValue lowerVectorShuffleByMerging128BitLanes(
10021 SDLoc DL, MVT VT, SDValue V1, SDValue V2, ArrayRef<int> Mask,
10022 const X86Subtarget *Subtarget, SelectionDAG &DAG) {
10023 assert(!isSingleInputShuffleMask(Mask) &&
10024 "This is only useful with multiple inputs.");
10026 int Size = Mask.size();
10027 int LaneSize = 128 / VT.getScalarSizeInBits();
10028 int NumLanes = Size / LaneSize;
10029 assert(NumLanes > 1 && "Only handles 256-bit and wider shuffles.");
10031 // See if we can build a hypothetical 128-bit lane-fixing shuffle mask. Also
10032 // check whether the in-128-bit lane shuffles share a repeating pattern.
10033 SmallVector<int, 4> Lanes;
10034 Lanes.resize(NumLanes, -1);
10035 SmallVector<int, 4> InLaneMask;
10036 InLaneMask.resize(LaneSize, -1);
10037 for (int i = 0; i < Size; ++i) {
10041 int j = i / LaneSize;
10043 if (Lanes[j] < 0) {
10044 // First entry we've seen for this lane.
10045 Lanes[j] = Mask[i] / LaneSize;
10046 } else if (Lanes[j] != Mask[i] / LaneSize) {
10047 // This doesn't match the lane selected previously!
10051 // Check that within each lane we have a consistent shuffle mask.
10052 int k = i % LaneSize;
10053 if (InLaneMask[k] < 0) {
10054 InLaneMask[k] = Mask[i] % LaneSize;
10055 } else if (InLaneMask[k] != Mask[i] % LaneSize) {
10056 // This doesn't fit a repeating in-lane mask.
10061 // First shuffle the lanes into place.
10062 MVT LaneVT = MVT::getVectorVT(VT.isFloatingPoint() ? MVT::f64 : MVT::i64,
10063 VT.getSizeInBits() / 64);
10064 SmallVector<int, 8> LaneMask;
10065 LaneMask.resize(NumLanes * 2, -1);
10066 for (int i = 0; i < NumLanes; ++i)
10067 if (Lanes[i] >= 0) {
10068 LaneMask[2 * i + 0] = 2*Lanes[i] + 0;
10069 LaneMask[2 * i + 1] = 2*Lanes[i] + 1;
10072 V1 = DAG.getBitcast(LaneVT, V1);
10073 V2 = DAG.getBitcast(LaneVT, V2);
10074 SDValue LaneShuffle = DAG.getVectorShuffle(LaneVT, DL, V1, V2, LaneMask);
10076 // Cast it back to the type we actually want.
10077 LaneShuffle = DAG.getBitcast(VT, LaneShuffle);
10079 // Now do a simple shuffle that isn't lane crossing.
10080 SmallVector<int, 8> NewMask;
10081 NewMask.resize(Size, -1);
10082 for (int i = 0; i < Size; ++i)
10084 NewMask[i] = (i / LaneSize) * LaneSize + Mask[i] % LaneSize;
10085 assert(!is128BitLaneCrossingShuffleMask(VT, NewMask) &&
10086 "Must not introduce lane crosses at this point!");
10088 return DAG.getVectorShuffle(VT, DL, LaneShuffle, DAG.getUNDEF(VT), NewMask);
10091 /// \brief Test whether the specified input (0 or 1) is in-place blended by the
10094 /// This returns true if the elements from a particular input are already in the
10095 /// slot required by the given mask and require no permutation.
10096 static bool isShuffleMaskInputInPlace(int Input, ArrayRef<int> Mask) {
10097 assert((Input == 0 || Input == 1) && "Only two inputs to shuffles.");
10098 int Size = Mask.size();
10099 for (int i = 0; i < Size; ++i)
10100 if (Mask[i] >= 0 && Mask[i] / Size == Input && Mask[i] % Size != i)
10106 static SDValue lowerVectorShuffleWithSHUFPD(SDLoc DL, MVT VT,
10107 ArrayRef<int> Mask, SDValue V1,
10108 SDValue V2, SelectionDAG &DAG) {
10110 // Mask for V8F64: 0/1, 8/9, 2/3, 10/11, 4/5, ..
10111 // Mask for V4F64; 0/1, 4/5, 2/3, 6/7..
10112 assert(VT.getScalarSizeInBits() == 64 && "Unexpected data type for VSHUFPD");
10113 int NumElts = VT.getVectorNumElements();
10114 bool ShufpdMask = true;
10115 bool CommutableMask = true;
10116 unsigned Immediate = 0;
10117 for (int i = 0; i < NumElts; ++i) {
10120 int Val = (i & 6) + NumElts * (i & 1);
10121 int CommutVal = (i & 0xe) + NumElts * ((i & 1)^1);
10122 if (Mask[i] < Val || Mask[i] > Val + 1)
10123 ShufpdMask = false;
10124 if (Mask[i] < CommutVal || Mask[i] > CommutVal + 1)
10125 CommutableMask = false;
10126 Immediate |= (Mask[i] % 2) << i;
10129 return DAG.getNode(X86ISD::SHUFP, DL, VT, V1, V2,
10130 DAG.getConstant(Immediate, DL, MVT::i8));
10131 if (CommutableMask)
10132 return DAG.getNode(X86ISD::SHUFP, DL, VT, V2, V1,
10133 DAG.getConstant(Immediate, DL, MVT::i8));
10137 /// \brief Handle lowering of 4-lane 64-bit floating point shuffles.
10139 /// Also ends up handling lowering of 4-lane 64-bit integer shuffles when AVX2
10140 /// isn't available.
10141 static SDValue lowerV4F64VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10142 const X86Subtarget *Subtarget,
10143 SelectionDAG &DAG) {
10145 assert(V1.getSimpleValueType() == MVT::v4f64 && "Bad operand type!");
10146 assert(V2.getSimpleValueType() == MVT::v4f64 && "Bad operand type!");
10147 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10148 ArrayRef<int> Mask = SVOp->getMask();
10149 assert(Mask.size() == 4 && "Unexpected mask size for v4 shuffle!");
10151 SmallVector<int, 4> WidenedMask;
10152 if (canWidenShuffleElements(Mask, WidenedMask))
10153 return lowerV2X128VectorShuffle(DL, MVT::v4f64, V1, V2, Mask, Subtarget,
10156 if (isSingleInputShuffleMask(Mask)) {
10157 // Check for being able to broadcast a single element.
10158 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v4f64, V1,
10159 Mask, Subtarget, DAG))
10162 // Use low duplicate instructions for masks that match their pattern.
10163 if (isShuffleEquivalent(V1, V2, Mask, {0, 0, 2, 2}))
10164 return DAG.getNode(X86ISD::MOVDDUP, DL, MVT::v4f64, V1);
10166 if (!is128BitLaneCrossingShuffleMask(MVT::v4f64, Mask)) {
10167 // Non-half-crossing single input shuffles can be lowerid with an
10168 // interleaved permutation.
10169 unsigned VPERMILPMask = (Mask[0] == 1) | ((Mask[1] == 1) << 1) |
10170 ((Mask[2] == 3) << 2) | ((Mask[3] == 3) << 3);
10171 return DAG.getNode(X86ISD::VPERMILPI, DL, MVT::v4f64, V1,
10172 DAG.getConstant(VPERMILPMask, DL, MVT::i8));
10175 // With AVX2 we have direct support for this permutation.
10176 if (Subtarget->hasAVX2())
10177 return DAG.getNode(X86ISD::VPERMI, DL, MVT::v4f64, V1,
10178 getV4X86ShuffleImm8ForMask(Mask, DL, DAG));
10180 // Otherwise, fall back.
10181 return lowerVectorShuffleAsLanePermuteAndBlend(DL, MVT::v4f64, V1, V2, Mask,
10185 // Use dedicated unpack instructions for masks that match their pattern.
10187 lowerVectorShuffleWithUNPCK(DL, MVT::v4f64, Mask, V1, V2, DAG))
10190 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v4f64, V1, V2, Mask,
10194 // Check if the blend happens to exactly fit that of SHUFPD.
10196 lowerVectorShuffleWithSHUFPD(DL, MVT::v4f64, Mask, V1, V2, DAG))
10199 // Try to simplify this by merging 128-bit lanes to enable a lane-based
10200 // shuffle. However, if we have AVX2 and either inputs are already in place,
10201 // we will be able to shuffle even across lanes the other input in a single
10202 // instruction so skip this pattern.
10203 if (!(Subtarget->hasAVX2() && (isShuffleMaskInputInPlace(0, Mask) ||
10204 isShuffleMaskInputInPlace(1, Mask))))
10205 if (SDValue Result = lowerVectorShuffleByMerging128BitLanes(
10206 DL, MVT::v4f64, V1, V2, Mask, Subtarget, DAG))
10209 // If we have AVX2 then we always want to lower with a blend because an v4 we
10210 // can fully permute the elements.
10211 if (Subtarget->hasAVX2())
10212 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v4f64, V1, V2,
10215 // Otherwise fall back on generic lowering.
10216 return lowerVectorShuffleAsSplitOrBlend(DL, MVT::v4f64, V1, V2, Mask, DAG);
10219 /// \brief Handle lowering of 4-lane 64-bit integer shuffles.
10221 /// This routine is only called when we have AVX2 and thus a reasonable
10222 /// instruction set for v4i64 shuffling..
10223 static SDValue lowerV4I64VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10224 const X86Subtarget *Subtarget,
10225 SelectionDAG &DAG) {
10227 assert(V1.getSimpleValueType() == MVT::v4i64 && "Bad operand type!");
10228 assert(V2.getSimpleValueType() == MVT::v4i64 && "Bad operand type!");
10229 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10230 ArrayRef<int> Mask = SVOp->getMask();
10231 assert(Mask.size() == 4 && "Unexpected mask size for v4 shuffle!");
10232 assert(Subtarget->hasAVX2() && "We can only lower v4i64 with AVX2!");
10234 SmallVector<int, 4> WidenedMask;
10235 if (canWidenShuffleElements(Mask, WidenedMask))
10236 return lowerV2X128VectorShuffle(DL, MVT::v4i64, V1, V2, Mask, Subtarget,
10239 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v4i64, V1, V2, Mask,
10243 // Check for being able to broadcast a single element.
10244 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v4i64, V1,
10245 Mask, Subtarget, DAG))
10248 // When the shuffle is mirrored between the 128-bit lanes of the unit, we can
10249 // use lower latency instructions that will operate on both 128-bit lanes.
10250 SmallVector<int, 2> RepeatedMask;
10251 if (is128BitLaneRepeatedShuffleMask(MVT::v4i64, Mask, RepeatedMask)) {
10252 if (isSingleInputShuffleMask(Mask)) {
10253 int PSHUFDMask[] = {-1, -1, -1, -1};
10254 for (int i = 0; i < 2; ++i)
10255 if (RepeatedMask[i] >= 0) {
10256 PSHUFDMask[2 * i] = 2 * RepeatedMask[i];
10257 PSHUFDMask[2 * i + 1] = 2 * RepeatedMask[i] + 1;
10259 return DAG.getBitcast(
10261 DAG.getNode(X86ISD::PSHUFD, DL, MVT::v8i32,
10262 DAG.getBitcast(MVT::v8i32, V1),
10263 getV4X86ShuffleImm8ForMask(PSHUFDMask, DL, DAG)));
10267 // AVX2 provides a direct instruction for permuting a single input across
10269 if (isSingleInputShuffleMask(Mask))
10270 return DAG.getNode(X86ISD::VPERMI, DL, MVT::v4i64, V1,
10271 getV4X86ShuffleImm8ForMask(Mask, DL, DAG));
10273 // Try to use shift instructions.
10274 if (SDValue Shift =
10275 lowerVectorShuffleAsShift(DL, MVT::v4i64, V1, V2, Mask, DAG))
10278 // Use dedicated unpack instructions for masks that match their pattern.
10280 lowerVectorShuffleWithUNPCK(DL, MVT::v4i64, Mask, V1, V2, DAG))
10283 // Try to simplify this by merging 128-bit lanes to enable a lane-based
10284 // shuffle. However, if we have AVX2 and either inputs are already in place,
10285 // we will be able to shuffle even across lanes the other input in a single
10286 // instruction so skip this pattern.
10287 if (!(Subtarget->hasAVX2() && (isShuffleMaskInputInPlace(0, Mask) ||
10288 isShuffleMaskInputInPlace(1, Mask))))
10289 if (SDValue Result = lowerVectorShuffleByMerging128BitLanes(
10290 DL, MVT::v4i64, V1, V2, Mask, Subtarget, DAG))
10293 // Otherwise fall back on generic blend lowering.
10294 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v4i64, V1, V2,
10298 /// \brief Handle lowering of 8-lane 32-bit floating point shuffles.
10300 /// Also ends up handling lowering of 8-lane 32-bit integer shuffles when AVX2
10301 /// isn't available.
10302 static SDValue lowerV8F32VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10303 const X86Subtarget *Subtarget,
10304 SelectionDAG &DAG) {
10306 assert(V1.getSimpleValueType() == MVT::v8f32 && "Bad operand type!");
10307 assert(V2.getSimpleValueType() == MVT::v8f32 && "Bad operand type!");
10308 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10309 ArrayRef<int> Mask = SVOp->getMask();
10310 assert(Mask.size() == 8 && "Unexpected mask size for v8 shuffle!");
10312 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v8f32, V1, V2, Mask,
10316 // Check for being able to broadcast a single element.
10317 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v8f32, V1,
10318 Mask, Subtarget, DAG))
10321 // If the shuffle mask is repeated in each 128-bit lane, we have many more
10322 // options to efficiently lower the shuffle.
10323 SmallVector<int, 4> RepeatedMask;
10324 if (is128BitLaneRepeatedShuffleMask(MVT::v8f32, Mask, RepeatedMask)) {
10325 assert(RepeatedMask.size() == 4 &&
10326 "Repeated masks must be half the mask width!");
10328 // Use even/odd duplicate instructions for masks that match their pattern.
10329 if (isShuffleEquivalent(V1, V2, Mask, {0, 0, 2, 2, 4, 4, 6, 6}))
10330 return DAG.getNode(X86ISD::MOVSLDUP, DL, MVT::v8f32, V1);
10331 if (isShuffleEquivalent(V1, V2, Mask, {1, 1, 3, 3, 5, 5, 7, 7}))
10332 return DAG.getNode(X86ISD::MOVSHDUP, DL, MVT::v8f32, V1);
10334 if (isSingleInputShuffleMask(Mask))
10335 return DAG.getNode(X86ISD::VPERMILPI, DL, MVT::v8f32, V1,
10336 getV4X86ShuffleImm8ForMask(RepeatedMask, DL, DAG));
10338 // Use dedicated unpack instructions for masks that match their pattern.
10340 lowerVectorShuffleWithUNPCK(DL, MVT::v8f32, Mask, V1, V2, DAG))
10343 // Otherwise, fall back to a SHUFPS sequence. Here it is important that we
10344 // have already handled any direct blends. We also need to squash the
10345 // repeated mask into a simulated v4f32 mask.
10346 for (int i = 0; i < 4; ++i)
10347 if (RepeatedMask[i] >= 8)
10348 RepeatedMask[i] -= 4;
10349 return lowerVectorShuffleWithSHUFPS(DL, MVT::v8f32, RepeatedMask, V1, V2, DAG);
10352 // If we have a single input shuffle with different shuffle patterns in the
10353 // two 128-bit lanes use the variable mask to VPERMILPS.
10354 if (isSingleInputShuffleMask(Mask)) {
10355 SDValue VPermMask[8];
10356 for (int i = 0; i < 8; ++i)
10357 VPermMask[i] = Mask[i] < 0 ? DAG.getUNDEF(MVT::i32)
10358 : DAG.getConstant(Mask[i], DL, MVT::i32);
10359 if (!is128BitLaneCrossingShuffleMask(MVT::v8f32, Mask))
10360 return DAG.getNode(
10361 X86ISD::VPERMILPV, DL, MVT::v8f32, V1,
10362 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v8i32, VPermMask));
10364 if (Subtarget->hasAVX2())
10365 return DAG.getNode(
10366 X86ISD::VPERMV, DL, MVT::v8f32,
10367 DAG.getBitcast(MVT::v8f32, DAG.getNode(ISD::BUILD_VECTOR, DL,
10368 MVT::v8i32, VPermMask)),
10371 // Otherwise, fall back.
10372 return lowerVectorShuffleAsLanePermuteAndBlend(DL, MVT::v8f32, V1, V2, Mask,
10376 // Try to simplify this by merging 128-bit lanes to enable a lane-based
10378 if (SDValue Result = lowerVectorShuffleByMerging128BitLanes(
10379 DL, MVT::v8f32, V1, V2, Mask, Subtarget, DAG))
10382 // If we have AVX2 then we always want to lower with a blend because at v8 we
10383 // can fully permute the elements.
10384 if (Subtarget->hasAVX2())
10385 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v8f32, V1, V2,
10388 // Otherwise fall back on generic lowering.
10389 return lowerVectorShuffleAsSplitOrBlend(DL, MVT::v8f32, V1, V2, Mask, DAG);
10392 /// \brief Handle lowering of 8-lane 32-bit integer shuffles.
10394 /// This routine is only called when we have AVX2 and thus a reasonable
10395 /// instruction set for v8i32 shuffling..
10396 static SDValue lowerV8I32VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10397 const X86Subtarget *Subtarget,
10398 SelectionDAG &DAG) {
10400 assert(V1.getSimpleValueType() == MVT::v8i32 && "Bad operand type!");
10401 assert(V2.getSimpleValueType() == MVT::v8i32 && "Bad operand type!");
10402 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10403 ArrayRef<int> Mask = SVOp->getMask();
10404 assert(Mask.size() == 8 && "Unexpected mask size for v8 shuffle!");
10405 assert(Subtarget->hasAVX2() && "We can only lower v8i32 with AVX2!");
10407 // Whenever we can lower this as a zext, that instruction is strictly faster
10408 // than any alternative. It also allows us to fold memory operands into the
10409 // shuffle in many cases.
10410 if (SDValue ZExt = lowerVectorShuffleAsZeroOrAnyExtend(DL, MVT::v8i32, V1, V2,
10411 Mask, Subtarget, DAG))
10414 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v8i32, V1, V2, Mask,
10418 // Check for being able to broadcast a single element.
10419 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v8i32, V1,
10420 Mask, Subtarget, DAG))
10423 // If the shuffle mask is repeated in each 128-bit lane we can use more
10424 // efficient instructions that mirror the shuffles across the two 128-bit
10426 SmallVector<int, 4> RepeatedMask;
10427 if (is128BitLaneRepeatedShuffleMask(MVT::v8i32, Mask, RepeatedMask)) {
10428 assert(RepeatedMask.size() == 4 && "Unexpected repeated mask size!");
10429 if (isSingleInputShuffleMask(Mask))
10430 return DAG.getNode(X86ISD::PSHUFD, DL, MVT::v8i32, V1,
10431 getV4X86ShuffleImm8ForMask(RepeatedMask, DL, DAG));
10433 // Use dedicated unpack instructions for masks that match their pattern.
10435 lowerVectorShuffleWithUNPCK(DL, MVT::v8i32, Mask, V1, V2, DAG))
10439 // Try to use shift instructions.
10440 if (SDValue Shift =
10441 lowerVectorShuffleAsShift(DL, MVT::v8i32, V1, V2, Mask, DAG))
10444 if (SDValue Rotate = lowerVectorShuffleAsByteRotate(
10445 DL, MVT::v8i32, V1, V2, Mask, Subtarget, DAG))
10448 // If the shuffle patterns aren't repeated but it is a single input, directly
10449 // generate a cross-lane VPERMD instruction.
10450 if (isSingleInputShuffleMask(Mask)) {
10451 SDValue VPermMask[8];
10452 for (int i = 0; i < 8; ++i)
10453 VPermMask[i] = Mask[i] < 0 ? DAG.getUNDEF(MVT::i32)
10454 : DAG.getConstant(Mask[i], DL, MVT::i32);
10455 return DAG.getNode(
10456 X86ISD::VPERMV, DL, MVT::v8i32,
10457 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v8i32, VPermMask), V1);
10460 // Try to simplify this by merging 128-bit lanes to enable a lane-based
10462 if (SDValue Result = lowerVectorShuffleByMerging128BitLanes(
10463 DL, MVT::v8i32, V1, V2, Mask, Subtarget, DAG))
10466 // Otherwise fall back on generic blend lowering.
10467 return lowerVectorShuffleAsDecomposedShuffleBlend(DL, MVT::v8i32, V1, V2,
10471 /// \brief Handle lowering of 16-lane 16-bit integer shuffles.
10473 /// This routine is only called when we have AVX2 and thus a reasonable
10474 /// instruction set for v16i16 shuffling..
10475 static SDValue lowerV16I16VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10476 const X86Subtarget *Subtarget,
10477 SelectionDAG &DAG) {
10479 assert(V1.getSimpleValueType() == MVT::v16i16 && "Bad operand type!");
10480 assert(V2.getSimpleValueType() == MVT::v16i16 && "Bad operand type!");
10481 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10482 ArrayRef<int> Mask = SVOp->getMask();
10483 assert(Mask.size() == 16 && "Unexpected mask size for v16 shuffle!");
10484 assert(Subtarget->hasAVX2() && "We can only lower v16i16 with AVX2!");
10486 // Whenever we can lower this as a zext, that instruction is strictly faster
10487 // than any alternative. It also allows us to fold memory operands into the
10488 // shuffle in many cases.
10489 if (SDValue ZExt = lowerVectorShuffleAsZeroOrAnyExtend(DL, MVT::v16i16, V1, V2,
10490 Mask, Subtarget, DAG))
10493 // Check for being able to broadcast a single element.
10494 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v16i16, V1,
10495 Mask, Subtarget, DAG))
10498 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v16i16, V1, V2, Mask,
10502 // Use dedicated unpack instructions for masks that match their pattern.
10504 lowerVectorShuffleWithUNPCK(DL, MVT::v16i16, Mask, V1, V2, DAG))
10507 // Try to use shift instructions.
10508 if (SDValue Shift =
10509 lowerVectorShuffleAsShift(DL, MVT::v16i16, V1, V2, Mask, DAG))
10512 // Try to use byte rotation instructions.
10513 if (SDValue Rotate = lowerVectorShuffleAsByteRotate(
10514 DL, MVT::v16i16, V1, V2, Mask, Subtarget, DAG))
10517 if (isSingleInputShuffleMask(Mask)) {
10518 // There are no generalized cross-lane shuffle operations available on i16
10520 if (is128BitLaneCrossingShuffleMask(MVT::v16i16, Mask))
10521 return lowerVectorShuffleAsLanePermuteAndBlend(DL, MVT::v16i16, V1, V2,
10524 SmallVector<int, 8> RepeatedMask;
10525 if (is128BitLaneRepeatedShuffleMask(MVT::v16i16, Mask, RepeatedMask)) {
10526 // As this is a single-input shuffle, the repeated mask should be
10527 // a strictly valid v8i16 mask that we can pass through to the v8i16
10528 // lowering to handle even the v16 case.
10529 return lowerV8I16GeneralSingleInputVectorShuffle(
10530 DL, MVT::v16i16, V1, RepeatedMask, Subtarget, DAG);
10533 SDValue PSHUFBMask[32];
10534 for (int i = 0; i < 16; ++i) {
10535 if (Mask[i] == -1) {
10536 PSHUFBMask[2 * i] = PSHUFBMask[2 * i + 1] = DAG.getUNDEF(MVT::i8);
10540 int M = i < 8 ? Mask[i] : Mask[i] - 8;
10541 assert(M >= 0 && M < 8 && "Invalid single-input mask!");
10542 PSHUFBMask[2 * i] = DAG.getConstant(2 * M, DL, MVT::i8);
10543 PSHUFBMask[2 * i + 1] = DAG.getConstant(2 * M + 1, DL, MVT::i8);
10545 return DAG.getBitcast(MVT::v16i16,
10546 DAG.getNode(X86ISD::PSHUFB, DL, MVT::v32i8,
10547 DAG.getBitcast(MVT::v32i8, V1),
10548 DAG.getNode(ISD::BUILD_VECTOR, DL,
10549 MVT::v32i8, PSHUFBMask)));
10552 // Try to simplify this by merging 128-bit lanes to enable a lane-based
10554 if (SDValue Result = lowerVectorShuffleByMerging128BitLanes(
10555 DL, MVT::v16i16, V1, V2, Mask, Subtarget, DAG))
10558 // Otherwise fall back on generic lowering.
10559 return lowerVectorShuffleAsSplitOrBlend(DL, MVT::v16i16, V1, V2, Mask, DAG);
10562 /// \brief Handle lowering of 32-lane 8-bit integer shuffles.
10564 /// This routine is only called when we have AVX2 and thus a reasonable
10565 /// instruction set for v32i8 shuffling..
10566 static SDValue lowerV32I8VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10567 const X86Subtarget *Subtarget,
10568 SelectionDAG &DAG) {
10570 assert(V1.getSimpleValueType() == MVT::v32i8 && "Bad operand type!");
10571 assert(V2.getSimpleValueType() == MVT::v32i8 && "Bad operand type!");
10572 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10573 ArrayRef<int> Mask = SVOp->getMask();
10574 assert(Mask.size() == 32 && "Unexpected mask size for v32 shuffle!");
10575 assert(Subtarget->hasAVX2() && "We can only lower v32i8 with AVX2!");
10577 // Whenever we can lower this as a zext, that instruction is strictly faster
10578 // than any alternative. It also allows us to fold memory operands into the
10579 // shuffle in many cases.
10580 if (SDValue ZExt = lowerVectorShuffleAsZeroOrAnyExtend(DL, MVT::v32i8, V1, V2,
10581 Mask, Subtarget, DAG))
10584 // Check for being able to broadcast a single element.
10585 if (SDValue Broadcast = lowerVectorShuffleAsBroadcast(DL, MVT::v32i8, V1,
10586 Mask, Subtarget, DAG))
10589 if (SDValue Blend = lowerVectorShuffleAsBlend(DL, MVT::v32i8, V1, V2, Mask,
10593 // Use dedicated unpack instructions for masks that match their pattern.
10595 lowerVectorShuffleWithUNPCK(DL, MVT::v32i8, Mask, V1, V2, DAG))
10598 // Try to use shift instructions.
10599 if (SDValue Shift =
10600 lowerVectorShuffleAsShift(DL, MVT::v32i8, V1, V2, Mask, DAG))
10603 // Try to use byte rotation instructions.
10604 if (SDValue Rotate = lowerVectorShuffleAsByteRotate(
10605 DL, MVT::v32i8, V1, V2, Mask, Subtarget, DAG))
10608 if (isSingleInputShuffleMask(Mask)) {
10609 // There are no generalized cross-lane shuffle operations available on i8
10611 if (is128BitLaneCrossingShuffleMask(MVT::v32i8, Mask))
10612 return lowerVectorShuffleAsLanePermuteAndBlend(DL, MVT::v32i8, V1, V2,
10615 SDValue PSHUFBMask[32];
10616 for (int i = 0; i < 32; ++i)
10619 ? DAG.getUNDEF(MVT::i8)
10620 : DAG.getConstant(Mask[i] < 16 ? Mask[i] : Mask[i] - 16, DL,
10623 return DAG.getNode(
10624 X86ISD::PSHUFB, DL, MVT::v32i8, V1,
10625 DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v32i8, PSHUFBMask));
10628 // Try to simplify this by merging 128-bit lanes to enable a lane-based
10630 if (SDValue Result = lowerVectorShuffleByMerging128BitLanes(
10631 DL, MVT::v32i8, V1, V2, Mask, Subtarget, DAG))
10634 // Otherwise fall back on generic lowering.
10635 return lowerVectorShuffleAsSplitOrBlend(DL, MVT::v32i8, V1, V2, Mask, DAG);
10638 /// \brief High-level routine to lower various 256-bit x86 vector shuffles.
10640 /// This routine either breaks down the specific type of a 256-bit x86 vector
10641 /// shuffle or splits it into two 128-bit shuffles and fuses the results back
10642 /// together based on the available instructions.
10643 static SDValue lower256BitVectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10644 MVT VT, const X86Subtarget *Subtarget,
10645 SelectionDAG &DAG) {
10647 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10648 ArrayRef<int> Mask = SVOp->getMask();
10650 // If we have a single input to the zero element, insert that into V1 if we
10651 // can do so cheaply.
10652 int NumElts = VT.getVectorNumElements();
10653 int NumV2Elements = std::count_if(Mask.begin(), Mask.end(), [NumElts](int M) {
10654 return M >= NumElts;
10657 if (NumV2Elements == 1 && Mask[0] >= NumElts)
10658 if (SDValue Insertion = lowerVectorShuffleAsElementInsertion(
10659 DL, VT, V1, V2, Mask, Subtarget, DAG))
10662 // There is a really nice hard cut-over between AVX1 and AVX2 that means we
10663 // can check for those subtargets here and avoid much of the subtarget
10664 // querying in the per-vector-type lowering routines. With AVX1 we have
10665 // essentially *zero* ability to manipulate a 256-bit vector with integer
10666 // types. Since we'll use floating point types there eventually, just
10667 // immediately cast everything to a float and operate entirely in that domain.
10668 if (VT.isInteger() && !Subtarget->hasAVX2()) {
10669 int ElementBits = VT.getScalarSizeInBits();
10670 if (ElementBits < 32)
10671 // No floating point type available, decompose into 128-bit vectors.
10672 return splitAndLowerVectorShuffle(DL, VT, V1, V2, Mask, DAG);
10674 MVT FpVT = MVT::getVectorVT(MVT::getFloatingPointVT(ElementBits),
10675 VT.getVectorNumElements());
10676 V1 = DAG.getBitcast(FpVT, V1);
10677 V2 = DAG.getBitcast(FpVT, V2);
10678 return DAG.getBitcast(VT, DAG.getVectorShuffle(FpVT, DL, V1, V2, Mask));
10681 switch (VT.SimpleTy) {
10683 return lowerV4F64VectorShuffle(Op, V1, V2, Subtarget, DAG);
10685 return lowerV4I64VectorShuffle(Op, V1, V2, Subtarget, DAG);
10687 return lowerV8F32VectorShuffle(Op, V1, V2, Subtarget, DAG);
10689 return lowerV8I32VectorShuffle(Op, V1, V2, Subtarget, DAG);
10691 return lowerV16I16VectorShuffle(Op, V1, V2, Subtarget, DAG);
10693 return lowerV32I8VectorShuffle(Op, V1, V2, Subtarget, DAG);
10696 llvm_unreachable("Not a valid 256-bit x86 vector type!");
10700 /// \brief Try to lower a vector shuffle as a 128-bit shuffles.
10701 static SDValue lowerV4X128VectorShuffle(SDLoc DL, MVT VT,
10702 ArrayRef<int> Mask,
10703 SDValue V1, SDValue V2,
10704 SelectionDAG &DAG) {
10705 assert(VT.getScalarSizeInBits() == 64 &&
10706 "Unexpected element type size for 128bit shuffle.");
10708 // To handle 256 bit vector requires VLX and most probably
10709 // function lowerV2X128VectorShuffle() is better solution.
10710 assert(VT.getSizeInBits() == 512 &&
10711 "Unexpected vector size for 128bit shuffle.");
10713 SmallVector<int, 4> WidenedMask;
10714 if (!canWidenShuffleElements(Mask, WidenedMask))
10717 // Form a 128-bit permutation.
10718 // Convert the 64-bit shuffle mask selection values into 128-bit selection
10719 // bits defined by a vshuf64x2 instruction's immediate control byte.
10720 unsigned PermMask = 0, Imm = 0;
10721 unsigned ControlBitsNum = WidenedMask.size() / 2;
10723 for (int i = 0, Size = WidenedMask.size(); i < Size; ++i) {
10724 if (WidenedMask[i] == SM_SentinelZero)
10727 // Use first element in place of undef mask.
10728 Imm = (WidenedMask[i] == SM_SentinelUndef) ? 0 : WidenedMask[i];
10729 PermMask |= (Imm % WidenedMask.size()) << (i * ControlBitsNum);
10732 return DAG.getNode(X86ISD::SHUF128, DL, VT, V1, V2,
10733 DAG.getConstant(PermMask, DL, MVT::i8));
10736 static SDValue lowerVectorShuffleWithPERMV(SDLoc DL, MVT VT,
10737 ArrayRef<int> Mask, SDValue V1,
10738 SDValue V2, SelectionDAG &DAG) {
10740 assert(VT.getScalarSizeInBits() >= 16 && "Unexpected data type for PERMV");
10742 MVT MaskEltVT = MVT::getIntegerVT(VT.getScalarSizeInBits());
10743 MVT MaskVecVT = MVT::getVectorVT(MaskEltVT, VT.getVectorNumElements());
10745 SDValue MaskNode = getConstVector(Mask, MaskVecVT, DAG, DL, true);
10746 if (isSingleInputShuffleMask(Mask))
10747 return DAG.getNode(X86ISD::VPERMV, DL, VT, MaskNode, V1);
10749 return DAG.getNode(X86ISD::VPERMV3, DL, VT, V1, MaskNode, V2);
10752 /// \brief Handle lowering of 8-lane 64-bit floating point shuffles.
10753 static SDValue lowerV8F64VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10754 const X86Subtarget *Subtarget,
10755 SelectionDAG &DAG) {
10757 assert(V1.getSimpleValueType() == MVT::v8f64 && "Bad operand type!");
10758 assert(V2.getSimpleValueType() == MVT::v8f64 && "Bad operand type!");
10759 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10760 ArrayRef<int> Mask = SVOp->getMask();
10761 assert(Mask.size() == 8 && "Unexpected mask size for v8 shuffle!");
10763 if (SDValue Shuf128 =
10764 lowerV4X128VectorShuffle(DL, MVT::v8f64, Mask, V1, V2, DAG))
10767 if (SDValue Unpck =
10768 lowerVectorShuffleWithUNPCK(DL, MVT::v8f64, Mask, V1, V2, DAG))
10771 return lowerVectorShuffleWithPERMV(DL, MVT::v8f64, Mask, V1, V2, DAG);
10774 /// \brief Handle lowering of 16-lane 32-bit floating point shuffles.
10775 static SDValue lowerV16F32VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10776 const X86Subtarget *Subtarget,
10777 SelectionDAG &DAG) {
10779 assert(V1.getSimpleValueType() == MVT::v16f32 && "Bad operand type!");
10780 assert(V2.getSimpleValueType() == MVT::v16f32 && "Bad operand type!");
10781 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10782 ArrayRef<int> Mask = SVOp->getMask();
10783 assert(Mask.size() == 16 && "Unexpected mask size for v16 shuffle!");
10785 if (SDValue Unpck =
10786 lowerVectorShuffleWithUNPCK(DL, MVT::v16f32, Mask, V1, V2, DAG))
10789 return lowerVectorShuffleWithPERMV(DL, MVT::v16f32, Mask, V1, V2, DAG);
10792 /// \brief Handle lowering of 8-lane 64-bit integer shuffles.
10793 static SDValue lowerV8I64VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10794 const X86Subtarget *Subtarget,
10795 SelectionDAG &DAG) {
10797 assert(V1.getSimpleValueType() == MVT::v8i64 && "Bad operand type!");
10798 assert(V2.getSimpleValueType() == MVT::v8i64 && "Bad operand type!");
10799 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10800 ArrayRef<int> Mask = SVOp->getMask();
10801 assert(Mask.size() == 8 && "Unexpected mask size for v8 shuffle!");
10803 if (SDValue Shuf128 =
10804 lowerV4X128VectorShuffle(DL, MVT::v8i64, Mask, V1, V2, DAG))
10807 if (SDValue Unpck =
10808 lowerVectorShuffleWithUNPCK(DL, MVT::v8i64, Mask, V1, V2, DAG))
10811 return lowerVectorShuffleWithPERMV(DL, MVT::v8i64, Mask, V1, V2, DAG);
10814 /// \brief Handle lowering of 16-lane 32-bit integer shuffles.
10815 static SDValue lowerV16I32VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10816 const X86Subtarget *Subtarget,
10817 SelectionDAG &DAG) {
10819 assert(V1.getSimpleValueType() == MVT::v16i32 && "Bad operand type!");
10820 assert(V2.getSimpleValueType() == MVT::v16i32 && "Bad operand type!");
10821 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10822 ArrayRef<int> Mask = SVOp->getMask();
10823 assert(Mask.size() == 16 && "Unexpected mask size for v16 shuffle!");
10825 if (SDValue Unpck =
10826 lowerVectorShuffleWithUNPCK(DL, MVT::v16i32, Mask, V1, V2, DAG))
10829 return lowerVectorShuffleWithPERMV(DL, MVT::v16i32, Mask, V1, V2, DAG);
10832 /// \brief Handle lowering of 32-lane 16-bit integer shuffles.
10833 static SDValue lowerV32I16VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10834 const X86Subtarget *Subtarget,
10835 SelectionDAG &DAG) {
10837 assert(V1.getSimpleValueType() == MVT::v32i16 && "Bad operand type!");
10838 assert(V2.getSimpleValueType() == MVT::v32i16 && "Bad operand type!");
10839 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10840 ArrayRef<int> Mask = SVOp->getMask();
10841 assert(Mask.size() == 32 && "Unexpected mask size for v32 shuffle!");
10842 assert(Subtarget->hasBWI() && "We can only lower v32i16 with AVX-512-BWI!");
10844 return lowerVectorShuffleWithPERMV(DL, MVT::v32i16, Mask, V1, V2, DAG);
10847 /// \brief Handle lowering of 64-lane 8-bit integer shuffles.
10848 static SDValue lowerV64I8VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10849 const X86Subtarget *Subtarget,
10850 SelectionDAG &DAG) {
10852 assert(V1.getSimpleValueType() == MVT::v64i8 && "Bad operand type!");
10853 assert(V2.getSimpleValueType() == MVT::v64i8 && "Bad operand type!");
10854 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10855 ArrayRef<int> Mask = SVOp->getMask();
10856 assert(Mask.size() == 64 && "Unexpected mask size for v64 shuffle!");
10857 assert(Subtarget->hasBWI() && "We can only lower v64i8 with AVX-512-BWI!");
10859 // FIXME: Implement direct support for this type!
10860 return splitAndLowerVectorShuffle(DL, MVT::v64i8, V1, V2, Mask, DAG);
10863 /// \brief High-level routine to lower various 512-bit x86 vector shuffles.
10865 /// This routine either breaks down the specific type of a 512-bit x86 vector
10866 /// shuffle or splits it into two 256-bit shuffles and fuses the results back
10867 /// together based on the available instructions.
10868 static SDValue lower512BitVectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10869 MVT VT, const X86Subtarget *Subtarget,
10870 SelectionDAG &DAG) {
10872 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10873 ArrayRef<int> Mask = SVOp->getMask();
10874 assert(Subtarget->hasAVX512() &&
10875 "Cannot lower 512-bit vectors w/ basic ISA!");
10877 // Check for being able to broadcast a single element.
10878 if (SDValue Broadcast =
10879 lowerVectorShuffleAsBroadcast(DL, VT, V1, Mask, Subtarget, DAG))
10882 // Dispatch to each element type for lowering. If we don't have supprot for
10883 // specific element type shuffles at 512 bits, immediately split them and
10884 // lower them. Each lowering routine of a given type is allowed to assume that
10885 // the requisite ISA extensions for that element type are available.
10886 switch (VT.SimpleTy) {
10888 return lowerV8F64VectorShuffle(Op, V1, V2, Subtarget, DAG);
10890 return lowerV16F32VectorShuffle(Op, V1, V2, Subtarget, DAG);
10892 return lowerV8I64VectorShuffle(Op, V1, V2, Subtarget, DAG);
10894 return lowerV16I32VectorShuffle(Op, V1, V2, Subtarget, DAG);
10896 if (Subtarget->hasBWI())
10897 return lowerV32I16VectorShuffle(Op, V1, V2, Subtarget, DAG);
10900 if (Subtarget->hasBWI())
10901 return lowerV64I8VectorShuffle(Op, V1, V2, Subtarget, DAG);
10905 llvm_unreachable("Not a valid 512-bit x86 vector type!");
10908 // Otherwise fall back on splitting.
10909 return splitAndLowerVectorShuffle(DL, VT, V1, V2, Mask, DAG);
10912 // Lower vXi1 vector shuffles.
10913 // There is no a dedicated instruction on AVX-512 that shuffles the masks.
10914 // The only way to shuffle bits is to sign-extend the mask vector to SIMD
10915 // vector, shuffle and then truncate it back.
10916 static SDValue lower1BitVectorShuffle(SDValue Op, SDValue V1, SDValue V2,
10917 MVT VT, const X86Subtarget *Subtarget,
10918 SelectionDAG &DAG) {
10920 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10921 ArrayRef<int> Mask = SVOp->getMask();
10922 assert(Subtarget->hasAVX512() &&
10923 "Cannot lower 512-bit vectors w/o basic ISA!");
10925 switch (VT.SimpleTy) {
10927 llvm_unreachable("Expected a vector of i1 elements");
10929 ExtVT = MVT::v2i64;
10932 ExtVT = MVT::v4i32;
10935 ExtVT = MVT::v8i64; // Take 512-bit type, more shuffles on KNL
10938 ExtVT = MVT::v16i32;
10941 ExtVT = MVT::v32i16;
10944 ExtVT = MVT::v64i8;
10948 if (ISD::isBuildVectorAllZeros(V1.getNode()))
10949 V1 = getZeroVector(ExtVT, Subtarget, DAG, DL);
10950 else if (ISD::isBuildVectorAllOnes(V1.getNode()))
10951 V1 = getOnesVector(ExtVT, Subtarget, DAG, DL);
10953 V1 = DAG.getNode(ISD::SIGN_EXTEND, DL, ExtVT, V1);
10956 V2 = DAG.getUNDEF(ExtVT);
10957 else if (ISD::isBuildVectorAllZeros(V2.getNode()))
10958 V2 = getZeroVector(ExtVT, Subtarget, DAG, DL);
10959 else if (ISD::isBuildVectorAllOnes(V2.getNode()))
10960 V2 = getOnesVector(ExtVT, Subtarget, DAG, DL);
10962 V2 = DAG.getNode(ISD::SIGN_EXTEND, DL, ExtVT, V2);
10963 return DAG.getNode(ISD::TRUNCATE, DL, VT,
10964 DAG.getVectorShuffle(ExtVT, DL, V1, V2, Mask));
10966 /// \brief Top-level lowering for x86 vector shuffles.
10968 /// This handles decomposition, canonicalization, and lowering of all x86
10969 /// vector shuffles. Most of the specific lowering strategies are encapsulated
10970 /// above in helper routines. The canonicalization attempts to widen shuffles
10971 /// to involve fewer lanes of wider elements, consolidate symmetric patterns
10972 /// s.t. only one of the two inputs needs to be tested, etc.
10973 static SDValue lowerVectorShuffle(SDValue Op, const X86Subtarget *Subtarget,
10974 SelectionDAG &DAG) {
10975 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
10976 ArrayRef<int> Mask = SVOp->getMask();
10977 SDValue V1 = Op.getOperand(0);
10978 SDValue V2 = Op.getOperand(1);
10979 MVT VT = Op.getSimpleValueType();
10980 int NumElements = VT.getVectorNumElements();
10982 bool Is1BitVector = (VT.getScalarType() == MVT::i1);
10984 assert((VT.getSizeInBits() != 64 || Is1BitVector) &&
10985 "Can't lower MMX shuffles");
10987 bool V1IsUndef = V1.getOpcode() == ISD::UNDEF;
10988 bool V2IsUndef = V2.getOpcode() == ISD::UNDEF;
10989 if (V1IsUndef && V2IsUndef)
10990 return DAG.getUNDEF(VT);
10992 // When we create a shuffle node we put the UNDEF node to second operand,
10993 // but in some cases the first operand may be transformed to UNDEF.
10994 // In this case we should just commute the node.
10996 return DAG.getCommutedVectorShuffle(*SVOp);
10998 // Check for non-undef masks pointing at an undef vector and make the masks
10999 // undef as well. This makes it easier to match the shuffle based solely on
11003 if (M >= NumElements) {
11004 SmallVector<int, 8> NewMask(Mask.begin(), Mask.end());
11005 for (int &M : NewMask)
11006 if (M >= NumElements)
11008 return DAG.getVectorShuffle(VT, dl, V1, V2, NewMask);
11011 // We actually see shuffles that are entirely re-arrangements of a set of
11012 // zero inputs. This mostly happens while decomposing complex shuffles into
11013 // simple ones. Directly lower these as a buildvector of zeros.
11014 SmallBitVector Zeroable = computeZeroableShuffleElements(Mask, V1, V2);
11015 if (Zeroable.all())
11016 return getZeroVector(VT, Subtarget, DAG, dl);
11018 // Try to collapse shuffles into using a vector type with fewer elements but
11019 // wider element types. We cap this to not form integers or floating point
11020 // elements wider than 64 bits, but it might be interesting to form i128
11021 // integers to handle flipping the low and high halves of AVX 256-bit vectors.
11022 SmallVector<int, 16> WidenedMask;
11023 if (VT.getScalarSizeInBits() < 64 && !Is1BitVector &&
11024 canWidenShuffleElements(Mask, WidenedMask)) {
11025 MVT NewEltVT = VT.isFloatingPoint()
11026 ? MVT::getFloatingPointVT(VT.getScalarSizeInBits() * 2)
11027 : MVT::getIntegerVT(VT.getScalarSizeInBits() * 2);
11028 MVT NewVT = MVT::getVectorVT(NewEltVT, VT.getVectorNumElements() / 2);
11029 // Make sure that the new vector type is legal. For example, v2f64 isn't
11031 if (DAG.getTargetLoweringInfo().isTypeLegal(NewVT)) {
11032 V1 = DAG.getBitcast(NewVT, V1);
11033 V2 = DAG.getBitcast(NewVT, V2);
11034 return DAG.getBitcast(
11035 VT, DAG.getVectorShuffle(NewVT, dl, V1, V2, WidenedMask));
11039 int NumV1Elements = 0, NumUndefElements = 0, NumV2Elements = 0;
11040 for (int M : SVOp->getMask())
11042 ++NumUndefElements;
11043 else if (M < NumElements)
11048 // Commute the shuffle as needed such that more elements come from V1 than
11049 // V2. This allows us to match the shuffle pattern strictly on how many
11050 // elements come from V1 without handling the symmetric cases.
11051 if (NumV2Elements > NumV1Elements)
11052 return DAG.getCommutedVectorShuffle(*SVOp);
11054 // When the number of V1 and V2 elements are the same, try to minimize the
11055 // number of uses of V2 in the low half of the vector. When that is tied,
11056 // ensure that the sum of indices for V1 is equal to or lower than the sum
11057 // indices for V2. When those are equal, try to ensure that the number of odd
11058 // indices for V1 is lower than the number of odd indices for V2.
11059 if (NumV1Elements == NumV2Elements) {
11060 int LowV1Elements = 0, LowV2Elements = 0;
11061 for (int M : SVOp->getMask().slice(0, NumElements / 2))
11062 if (M >= NumElements)
11066 if (LowV2Elements > LowV1Elements) {
11067 return DAG.getCommutedVectorShuffle(*SVOp);
11068 } else if (LowV2Elements == LowV1Elements) {
11069 int SumV1Indices = 0, SumV2Indices = 0;
11070 for (int i = 0, Size = SVOp->getMask().size(); i < Size; ++i)
11071 if (SVOp->getMask()[i] >= NumElements)
11073 else if (SVOp->getMask()[i] >= 0)
11075 if (SumV2Indices < SumV1Indices) {
11076 return DAG.getCommutedVectorShuffle(*SVOp);
11077 } else if (SumV2Indices == SumV1Indices) {
11078 int NumV1OddIndices = 0, NumV2OddIndices = 0;
11079 for (int i = 0, Size = SVOp->getMask().size(); i < Size; ++i)
11080 if (SVOp->getMask()[i] >= NumElements)
11081 NumV2OddIndices += i % 2;
11082 else if (SVOp->getMask()[i] >= 0)
11083 NumV1OddIndices += i % 2;
11084 if (NumV2OddIndices < NumV1OddIndices)
11085 return DAG.getCommutedVectorShuffle(*SVOp);
11090 // For each vector width, delegate to a specialized lowering routine.
11091 if (VT.getSizeInBits() == 128)
11092 return lower128BitVectorShuffle(Op, V1, V2, VT, Subtarget, DAG);
11094 if (VT.getSizeInBits() == 256)
11095 return lower256BitVectorShuffle(Op, V1, V2, VT, Subtarget, DAG);
11097 if (VT.getSizeInBits() == 512)
11098 return lower512BitVectorShuffle(Op, V1, V2, VT, Subtarget, DAG);
11101 return lower1BitVectorShuffle(Op, V1, V2, VT, Subtarget, DAG);
11102 llvm_unreachable("Unimplemented!");
11105 // This function assumes its argument is a BUILD_VECTOR of constants or
11106 // undef SDNodes. i.e: ISD::isBuildVectorOfConstantSDNodes(BuildVector) is
11108 static bool BUILD_VECTORtoBlendMask(BuildVectorSDNode *BuildVector,
11109 unsigned &MaskValue) {
11111 unsigned NumElems = BuildVector->getNumOperands();
11113 // There are 2 lanes if (NumElems > 8), and 1 lane otherwise.
11114 // We don't handle the >2 lanes case right now.
11115 unsigned NumLanes = (NumElems - 1) / 8 + 1;
11119 unsigned NumElemsInLane = NumElems / NumLanes;
11121 // Blend for v16i16 should be symmetric for the both lanes.
11122 for (unsigned i = 0; i < NumElemsInLane; ++i) {
11123 SDValue EltCond = BuildVector->getOperand(i);
11124 SDValue SndLaneEltCond =
11125 (NumLanes == 2) ? BuildVector->getOperand(i + NumElemsInLane) : EltCond;
11127 int Lane1Cond = -1, Lane2Cond = -1;
11128 if (isa<ConstantSDNode>(EltCond))
11129 Lane1Cond = !isZero(EltCond);
11130 if (isa<ConstantSDNode>(SndLaneEltCond))
11131 Lane2Cond = !isZero(SndLaneEltCond);
11133 unsigned LaneMask = 0;
11134 if (Lane1Cond == Lane2Cond || Lane2Cond < 0)
11135 // Lane1Cond != 0, means we want the first argument.
11136 // Lane1Cond == 0, means we want the second argument.
11137 // The encoding of this argument is 0 for the first argument, 1
11138 // for the second. Therefore, invert the condition.
11139 LaneMask = !Lane1Cond << i;
11140 else if (Lane1Cond < 0)
11141 LaneMask = !Lane2Cond << i;
11145 MaskValue |= LaneMask;
11147 MaskValue |= LaneMask << NumElemsInLane;
11152 /// \brief Try to lower a VSELECT instruction to a vector shuffle.
11153 static SDValue lowerVSELECTtoVectorShuffle(SDValue Op,
11154 const X86Subtarget *Subtarget,
11155 SelectionDAG &DAG) {
11156 SDValue Cond = Op.getOperand(0);
11157 SDValue LHS = Op.getOperand(1);
11158 SDValue RHS = Op.getOperand(2);
11160 MVT VT = Op.getSimpleValueType();
11162 if (!ISD::isBuildVectorOfConstantSDNodes(Cond.getNode()))
11164 auto *CondBV = cast<BuildVectorSDNode>(Cond);
11166 // Only non-legal VSELECTs reach this lowering, convert those into generic
11167 // shuffles and re-use the shuffle lowering path for blends.
11168 SmallVector<int, 32> Mask;
11169 for (int i = 0, Size = VT.getVectorNumElements(); i < Size; ++i) {
11170 SDValue CondElt = CondBV->getOperand(i);
11172 isa<ConstantSDNode>(CondElt) ? i + (isZero(CondElt) ? Size : 0) : -1);
11174 return DAG.getVectorShuffle(VT, dl, LHS, RHS, Mask);
11177 SDValue X86TargetLowering::LowerVSELECT(SDValue Op, SelectionDAG &DAG) const {
11178 // A vselect where all conditions and data are constants can be optimized into
11179 // a single vector load by SelectionDAGLegalize::ExpandBUILD_VECTOR().
11180 if (ISD::isBuildVectorOfConstantSDNodes(Op.getOperand(0).getNode()) &&
11181 ISD::isBuildVectorOfConstantSDNodes(Op.getOperand(1).getNode()) &&
11182 ISD::isBuildVectorOfConstantSDNodes(Op.getOperand(2).getNode()))
11185 // Try to lower this to a blend-style vector shuffle. This can handle all
11186 // constant condition cases.
11187 if (SDValue BlendOp = lowerVSELECTtoVectorShuffle(Op, Subtarget, DAG))
11190 // Variable blends are only legal from SSE4.1 onward.
11191 if (!Subtarget->hasSSE41())
11194 // Only some types will be legal on some subtargets. If we can emit a legal
11195 // VSELECT-matching blend, return Op, and but if we need to expand, return
11197 switch (Op.getSimpleValueType().SimpleTy) {
11199 // Most of the vector types have blends past SSE4.1.
11203 // The byte blends for AVX vectors were introduced only in AVX2.
11204 if (Subtarget->hasAVX2())
11211 // AVX-512 BWI and VLX features support VSELECT with i16 elements.
11212 if (Subtarget->hasBWI() && Subtarget->hasVLX())
11215 // FIXME: We should custom lower this by fixing the condition and using i8
11221 static SDValue LowerEXTRACT_VECTOR_ELT_SSE4(SDValue Op, SelectionDAG &DAG) {
11222 MVT VT = Op.getSimpleValueType();
11225 if (!Op.getOperand(0).getSimpleValueType().is128BitVector())
11228 if (VT.getSizeInBits() == 8) {
11229 SDValue Extract = DAG.getNode(X86ISD::PEXTRB, dl, MVT::i32,
11230 Op.getOperand(0), Op.getOperand(1));
11231 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Extract,
11232 DAG.getValueType(VT));
11233 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
11236 if (VT.getSizeInBits() == 16) {
11237 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
11238 // If Idx is 0, it's cheaper to do a move instead of a pextrw.
11240 return DAG.getNode(
11241 ISD::TRUNCATE, dl, MVT::i16,
11242 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
11243 DAG.getBitcast(MVT::v4i32, Op.getOperand(0)),
11244 Op.getOperand(1)));
11245 SDValue Extract = DAG.getNode(X86ISD::PEXTRW, dl, MVT::i32,
11246 Op.getOperand(0), Op.getOperand(1));
11247 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, MVT::i32, Extract,
11248 DAG.getValueType(VT));
11249 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
11252 if (VT == MVT::f32) {
11253 // EXTRACTPS outputs to a GPR32 register which will require a movd to copy
11254 // the result back to FR32 register. It's only worth matching if the
11255 // result has a single use which is a store or a bitcast to i32. And in
11256 // the case of a store, it's not worth it if the index is a constant 0,
11257 // because a MOVSSmr can be used instead, which is smaller and faster.
11258 if (!Op.hasOneUse())
11260 SDNode *User = *Op.getNode()->use_begin();
11261 if ((User->getOpcode() != ISD::STORE ||
11262 (isa<ConstantSDNode>(Op.getOperand(1)) &&
11263 cast<ConstantSDNode>(Op.getOperand(1))->isNullValue())) &&
11264 (User->getOpcode() != ISD::BITCAST ||
11265 User->getValueType(0) != MVT::i32))
11267 SDValue Extract = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
11268 DAG.getBitcast(MVT::v4i32, Op.getOperand(0)),
11270 return DAG.getBitcast(MVT::f32, Extract);
11273 if (VT == MVT::i32 || VT == MVT::i64) {
11274 // ExtractPS/pextrq works with constant index.
11275 if (isa<ConstantSDNode>(Op.getOperand(1)))
11281 /// Extract one bit from mask vector, like v16i1 or v8i1.
11282 /// AVX-512 feature.
11284 X86TargetLowering::ExtractBitFromMaskVector(SDValue Op, SelectionDAG &DAG) const {
11285 SDValue Vec = Op.getOperand(0);
11287 MVT VecVT = Vec.getSimpleValueType();
11288 SDValue Idx = Op.getOperand(1);
11289 MVT EltVT = Op.getSimpleValueType();
11291 assert((EltVT == MVT::i1) && "Unexpected operands in ExtractBitFromMaskVector");
11292 assert((VecVT.getVectorNumElements() <= 16 || Subtarget->hasBWI()) &&
11293 "Unexpected vector type in ExtractBitFromMaskVector");
11295 // variable index can't be handled in mask registers,
11296 // extend vector to VR512
11297 if (!isa<ConstantSDNode>(Idx)) {
11298 MVT ExtVT = (VecVT == MVT::v8i1 ? MVT::v8i64 : MVT::v16i32);
11299 SDValue Ext = DAG.getNode(ISD::ZERO_EXTEND, dl, ExtVT, Vec);
11300 SDValue Elt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl,
11301 ExtVT.getVectorElementType(), Ext, Idx);
11302 return DAG.getNode(ISD::TRUNCATE, dl, EltVT, Elt);
11305 unsigned IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
11306 const TargetRegisterClass* rc = getRegClassFor(VecVT);
11307 if (!Subtarget->hasDQI() && (VecVT.getVectorNumElements() <= 8))
11308 rc = getRegClassFor(MVT::v16i1);
11309 unsigned MaxSift = rc->getSize()*8 - 1;
11310 Vec = DAG.getNode(X86ISD::VSHLI, dl, VecVT, Vec,
11311 DAG.getConstant(MaxSift - IdxVal, dl, MVT::i8));
11312 Vec = DAG.getNode(X86ISD::VSRLI, dl, VecVT, Vec,
11313 DAG.getConstant(MaxSift, dl, MVT::i8));
11314 return DAG.getNode(X86ISD::VEXTRACT, dl, MVT::i1, Vec,
11315 DAG.getIntPtrConstant(0, dl));
11319 X86TargetLowering::LowerEXTRACT_VECTOR_ELT(SDValue Op,
11320 SelectionDAG &DAG) const {
11322 SDValue Vec = Op.getOperand(0);
11323 MVT VecVT = Vec.getSimpleValueType();
11324 SDValue Idx = Op.getOperand(1);
11326 if (Op.getSimpleValueType() == MVT::i1)
11327 return ExtractBitFromMaskVector(Op, DAG);
11329 if (!isa<ConstantSDNode>(Idx)) {
11330 if (VecVT.is512BitVector() ||
11331 (VecVT.is256BitVector() && Subtarget->hasInt256() &&
11332 VecVT.getVectorElementType().getSizeInBits() == 32)) {
11335 MVT::getIntegerVT(VecVT.getVectorElementType().getSizeInBits());
11336 MVT MaskVT = MVT::getVectorVT(MaskEltVT, VecVT.getSizeInBits() /
11337 MaskEltVT.getSizeInBits());
11339 Idx = DAG.getZExtOrTrunc(Idx, dl, MaskEltVT);
11340 auto PtrVT = getPointerTy(DAG.getDataLayout());
11341 SDValue Mask = DAG.getNode(X86ISD::VINSERT, dl, MaskVT,
11342 getZeroVector(MaskVT, Subtarget, DAG, dl), Idx,
11343 DAG.getConstant(0, dl, PtrVT));
11344 SDValue Perm = DAG.getNode(X86ISD::VPERMV, dl, VecVT, Mask, Vec);
11345 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, Op.getValueType(), Perm,
11346 DAG.getConstant(0, dl, PtrVT));
11351 // If this is a 256-bit vector result, first extract the 128-bit vector and
11352 // then extract the element from the 128-bit vector.
11353 if (VecVT.is256BitVector() || VecVT.is512BitVector()) {
11355 unsigned IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
11356 // Get the 128-bit vector.
11357 Vec = Extract128BitVector(Vec, IdxVal, DAG, dl);
11358 MVT EltVT = VecVT.getVectorElementType();
11360 unsigned ElemsPerChunk = 128 / EltVT.getSizeInBits();
11362 //if (IdxVal >= NumElems/2)
11363 // IdxVal -= NumElems/2;
11364 IdxVal -= (IdxVal/ElemsPerChunk)*ElemsPerChunk;
11365 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, Op.getValueType(), Vec,
11366 DAG.getConstant(IdxVal, dl, MVT::i32));
11369 assert(VecVT.is128BitVector() && "Unexpected vector length");
11371 if (Subtarget->hasSSE41())
11372 if (SDValue Res = LowerEXTRACT_VECTOR_ELT_SSE4(Op, DAG))
11375 MVT VT = Op.getSimpleValueType();
11376 // TODO: handle v16i8.
11377 if (VT.getSizeInBits() == 16) {
11378 SDValue Vec = Op.getOperand(0);
11379 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
11381 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i16,
11382 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
11383 DAG.getBitcast(MVT::v4i32, Vec),
11384 Op.getOperand(1)));
11385 // Transform it so it match pextrw which produces a 32-bit result.
11386 MVT EltVT = MVT::i32;
11387 SDValue Extract = DAG.getNode(X86ISD::PEXTRW, dl, EltVT,
11388 Op.getOperand(0), Op.getOperand(1));
11389 SDValue Assert = DAG.getNode(ISD::AssertZext, dl, EltVT, Extract,
11390 DAG.getValueType(VT));
11391 return DAG.getNode(ISD::TRUNCATE, dl, VT, Assert);
11394 if (VT.getSizeInBits() == 32) {
11395 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
11399 // SHUFPS the element to the lowest double word, then movss.
11400 int Mask[4] = { static_cast<int>(Idx), -1, -1, -1 };
11401 MVT VVT = Op.getOperand(0).getSimpleValueType();
11402 SDValue Vec = DAG.getVectorShuffle(VVT, dl, Op.getOperand(0),
11403 DAG.getUNDEF(VVT), Mask);
11404 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, Vec,
11405 DAG.getIntPtrConstant(0, dl));
11408 if (VT.getSizeInBits() == 64) {
11409 // FIXME: .td only matches this for <2 x f64>, not <2 x i64> on 32b
11410 // FIXME: seems like this should be unnecessary if mov{h,l}pd were taught
11411 // to match extract_elt for f64.
11412 unsigned Idx = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
11416 // UNPCKHPD the element to the lowest double word, then movsd.
11417 // Note if the lower 64 bits of the result of the UNPCKHPD is then stored
11418 // to a f64mem, the whole operation is folded into a single MOVHPDmr.
11419 int Mask[2] = { 1, -1 };
11420 MVT VVT = Op.getOperand(0).getSimpleValueType();
11421 SDValue Vec = DAG.getVectorShuffle(VVT, dl, Op.getOperand(0),
11422 DAG.getUNDEF(VVT), Mask);
11423 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, Vec,
11424 DAG.getIntPtrConstant(0, dl));
11430 /// Insert one bit to mask vector, like v16i1 or v8i1.
11431 /// AVX-512 feature.
11433 X86TargetLowering::InsertBitToMaskVector(SDValue Op, SelectionDAG &DAG) const {
11435 SDValue Vec = Op.getOperand(0);
11436 SDValue Elt = Op.getOperand(1);
11437 SDValue Idx = Op.getOperand(2);
11438 MVT VecVT = Vec.getSimpleValueType();
11440 if (!isa<ConstantSDNode>(Idx)) {
11441 // Non constant index. Extend source and destination,
11442 // insert element and then truncate the result.
11443 MVT ExtVecVT = (VecVT == MVT::v8i1 ? MVT::v8i64 : MVT::v16i32);
11444 MVT ExtEltVT = (VecVT == MVT::v8i1 ? MVT::i64 : MVT::i32);
11445 SDValue ExtOp = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, ExtVecVT,
11446 DAG.getNode(ISD::ZERO_EXTEND, dl, ExtVecVT, Vec),
11447 DAG.getNode(ISD::ZERO_EXTEND, dl, ExtEltVT, Elt), Idx);
11448 return DAG.getNode(ISD::TRUNCATE, dl, VecVT, ExtOp);
11451 unsigned IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
11452 SDValue EltInVec = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VecVT, Elt);
11454 EltInVec = DAG.getNode(X86ISD::VSHLI, dl, VecVT, EltInVec,
11455 DAG.getConstant(IdxVal, dl, MVT::i8));
11456 if (Vec.getOpcode() == ISD::UNDEF)
11458 return DAG.getNode(ISD::OR, dl, VecVT, Vec, EltInVec);
11461 SDValue X86TargetLowering::LowerINSERT_VECTOR_ELT(SDValue Op,
11462 SelectionDAG &DAG) const {
11463 MVT VT = Op.getSimpleValueType();
11464 MVT EltVT = VT.getVectorElementType();
11466 if (EltVT == MVT::i1)
11467 return InsertBitToMaskVector(Op, DAG);
11470 SDValue N0 = Op.getOperand(0);
11471 SDValue N1 = Op.getOperand(1);
11472 SDValue N2 = Op.getOperand(2);
11473 if (!isa<ConstantSDNode>(N2))
11475 auto *N2C = cast<ConstantSDNode>(N2);
11476 unsigned IdxVal = N2C->getZExtValue();
11478 // If the vector is wider than 128 bits, extract the 128-bit subvector, insert
11479 // into that, and then insert the subvector back into the result.
11480 if (VT.is256BitVector() || VT.is512BitVector()) {
11481 // With a 256-bit vector, we can insert into the zero element efficiently
11482 // using a blend if we have AVX or AVX2 and the right data type.
11483 if (VT.is256BitVector() && IdxVal == 0) {
11484 // TODO: It is worthwhile to cast integer to floating point and back
11485 // and incur a domain crossing penalty if that's what we'll end up
11486 // doing anyway after extracting to a 128-bit vector.
11487 if ((Subtarget->hasAVX() && (EltVT == MVT::f64 || EltVT == MVT::f32)) ||
11488 (Subtarget->hasAVX2() && EltVT == MVT::i32)) {
11489 SDValue N1Vec = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, N1);
11490 N2 = DAG.getIntPtrConstant(1, dl);
11491 return DAG.getNode(X86ISD::BLENDI, dl, VT, N0, N1Vec, N2);
11495 // Get the desired 128-bit vector chunk.
11496 SDValue V = Extract128BitVector(N0, IdxVal, DAG, dl);
11498 // Insert the element into the desired chunk.
11499 unsigned NumEltsIn128 = 128 / EltVT.getSizeInBits();
11500 unsigned IdxIn128 = IdxVal - (IdxVal / NumEltsIn128) * NumEltsIn128;
11502 V = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, V.getValueType(), V, N1,
11503 DAG.getConstant(IdxIn128, dl, MVT::i32));
11505 // Insert the changed part back into the bigger vector
11506 return Insert128BitVector(N0, V, IdxVal, DAG, dl);
11508 assert(VT.is128BitVector() && "Only 128-bit vector types should be left!");
11510 if (Subtarget->hasSSE41()) {
11511 if (EltVT.getSizeInBits() == 8 || EltVT.getSizeInBits() == 16) {
11513 if (VT == MVT::v8i16) {
11514 Opc = X86ISD::PINSRW;
11516 assert(VT == MVT::v16i8);
11517 Opc = X86ISD::PINSRB;
11520 // Transform it so it match pinsr{b,w} which expects a GR32 as its second
11522 if (N1.getValueType() != MVT::i32)
11523 N1 = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, N1);
11524 if (N2.getValueType() != MVT::i32)
11525 N2 = DAG.getIntPtrConstant(IdxVal, dl);
11526 return DAG.getNode(Opc, dl, VT, N0, N1, N2);
11529 if (EltVT == MVT::f32) {
11530 // Bits [7:6] of the constant are the source select. This will always be
11531 // zero here. The DAG Combiner may combine an extract_elt index into
11532 // these bits. For example (insert (extract, 3), 2) could be matched by
11533 // putting the '3' into bits [7:6] of X86ISD::INSERTPS.
11534 // Bits [5:4] of the constant are the destination select. This is the
11535 // value of the incoming immediate.
11536 // Bits [3:0] of the constant are the zero mask. The DAG Combiner may
11537 // combine either bitwise AND or insert of float 0.0 to set these bits.
11539 bool MinSize = DAG.getMachineFunction().getFunction()->optForMinSize();
11540 if (IdxVal == 0 && (!MinSize || !MayFoldLoad(N1))) {
11541 // If this is an insertion of 32-bits into the low 32-bits of
11542 // a vector, we prefer to generate a blend with immediate rather
11543 // than an insertps. Blends are simpler operations in hardware and so
11544 // will always have equal or better performance than insertps.
11545 // But if optimizing for size and there's a load folding opportunity,
11546 // generate insertps because blendps does not have a 32-bit memory
11548 N2 = DAG.getIntPtrConstant(1, dl);
11549 N1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4f32, N1);
11550 return DAG.getNode(X86ISD::BLENDI, dl, VT, N0, N1, N2);
11552 N2 = DAG.getIntPtrConstant(IdxVal << 4, dl);
11553 // Create this as a scalar to vector..
11554 N1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4f32, N1);
11555 return DAG.getNode(X86ISD::INSERTPS, dl, VT, N0, N1, N2);
11558 if (EltVT == MVT::i32 || EltVT == MVT::i64) {
11559 // PINSR* works with constant index.
11564 if (EltVT == MVT::i8)
11567 if (EltVT.getSizeInBits() == 16) {
11568 // Transform it so it match pinsrw which expects a 16-bit value in a GR32
11569 // as its second argument.
11570 if (N1.getValueType() != MVT::i32)
11571 N1 = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, N1);
11572 if (N2.getValueType() != MVT::i32)
11573 N2 = DAG.getIntPtrConstant(IdxVal, dl);
11574 return DAG.getNode(X86ISD::PINSRW, dl, VT, N0, N1, N2);
11579 static SDValue LowerSCALAR_TO_VECTOR(SDValue Op, SelectionDAG &DAG) {
11581 MVT OpVT = Op.getSimpleValueType();
11583 // If this is a 256-bit vector result, first insert into a 128-bit
11584 // vector and then insert into the 256-bit vector.
11585 if (!OpVT.is128BitVector()) {
11586 // Insert into a 128-bit vector.
11587 unsigned SizeFactor = OpVT.getSizeInBits()/128;
11588 MVT VT128 = MVT::getVectorVT(OpVT.getVectorElementType(),
11589 OpVT.getVectorNumElements() / SizeFactor);
11591 Op = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT128, Op.getOperand(0));
11593 // Insert the 128-bit vector.
11594 return Insert128BitVector(DAG.getUNDEF(OpVT), Op, 0, DAG, dl);
11597 if (OpVT == MVT::v1i64 &&
11598 Op.getOperand(0).getValueType() == MVT::i64)
11599 return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v1i64, Op.getOperand(0));
11601 SDValue AnyExt = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, Op.getOperand(0));
11602 assert(OpVT.is128BitVector() && "Expected an SSE type!");
11603 return DAG.getBitcast(
11604 OpVT, DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32, AnyExt));
11607 // Lower a node with an EXTRACT_SUBVECTOR opcode. This may result in
11608 // a simple subregister reference or explicit instructions to grab
11609 // upper bits of a vector.
11610 static SDValue LowerEXTRACT_SUBVECTOR(SDValue Op, const X86Subtarget *Subtarget,
11611 SelectionDAG &DAG) {
11613 SDValue In = Op.getOperand(0);
11614 SDValue Idx = Op.getOperand(1);
11615 unsigned IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
11616 MVT ResVT = Op.getSimpleValueType();
11617 MVT InVT = In.getSimpleValueType();
11619 if (Subtarget->hasFp256()) {
11620 if (ResVT.is128BitVector() &&
11621 (InVT.is256BitVector() || InVT.is512BitVector()) &&
11622 isa<ConstantSDNode>(Idx)) {
11623 return Extract128BitVector(In, IdxVal, DAG, dl);
11625 if (ResVT.is256BitVector() && InVT.is512BitVector() &&
11626 isa<ConstantSDNode>(Idx)) {
11627 return Extract256BitVector(In, IdxVal, DAG, dl);
11633 // Lower a node with an INSERT_SUBVECTOR opcode. This may result in a
11634 // simple superregister reference or explicit instructions to insert
11635 // the upper bits of a vector.
11636 static SDValue LowerINSERT_SUBVECTOR(SDValue Op, const X86Subtarget *Subtarget,
11637 SelectionDAG &DAG) {
11638 if (!Subtarget->hasAVX())
11642 SDValue Vec = Op.getOperand(0);
11643 SDValue SubVec = Op.getOperand(1);
11644 SDValue Idx = Op.getOperand(2);
11646 if (!isa<ConstantSDNode>(Idx))
11649 unsigned IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
11650 MVT OpVT = Op.getSimpleValueType();
11651 MVT SubVecVT = SubVec.getSimpleValueType();
11653 // Fold two 16-byte subvector loads into one 32-byte load:
11654 // (insert_subvector (insert_subvector undef, (load addr), 0),
11655 // (load addr + 16), Elts/2)
11657 if ((IdxVal == OpVT.getVectorNumElements() / 2) &&
11658 Vec.getOpcode() == ISD::INSERT_SUBVECTOR &&
11659 OpVT.is256BitVector() && SubVecVT.is128BitVector()) {
11660 auto *Idx2 = dyn_cast<ConstantSDNode>(Vec.getOperand(2));
11661 if (Idx2 && Idx2->getZExtValue() == 0) {
11662 SDValue SubVec2 = Vec.getOperand(1);
11663 // If needed, look through a bitcast to get to the load.
11664 if (SubVec2.getNode() && SubVec2.getOpcode() == ISD::BITCAST)
11665 SubVec2 = SubVec2.getOperand(0);
11667 if (auto *FirstLd = dyn_cast<LoadSDNode>(SubVec2)) {
11669 unsigned Alignment = FirstLd->getAlignment();
11670 unsigned AS = FirstLd->getAddressSpace();
11671 const X86TargetLowering *TLI = Subtarget->getTargetLowering();
11672 if (TLI->allowsMemoryAccess(*DAG.getContext(), DAG.getDataLayout(),
11673 OpVT, AS, Alignment, &Fast) && Fast) {
11674 SDValue Ops[] = { SubVec2, SubVec };
11675 if (SDValue Ld = EltsFromConsecutiveLoads(OpVT, Ops, dl, DAG, false))
11682 if ((OpVT.is256BitVector() || OpVT.is512BitVector()) &&
11683 SubVecVT.is128BitVector())
11684 return Insert128BitVector(Vec, SubVec, IdxVal, DAG, dl);
11686 if (OpVT.is512BitVector() && SubVecVT.is256BitVector())
11687 return Insert256BitVector(Vec, SubVec, IdxVal, DAG, dl);
11689 if (OpVT.getVectorElementType() == MVT::i1) {
11690 if (IdxVal == 0 && Vec.getOpcode() == ISD::UNDEF) // the operation is legal
11692 SDValue ZeroIdx = DAG.getIntPtrConstant(0, dl);
11693 SDValue Undef = DAG.getUNDEF(OpVT);
11694 unsigned NumElems = OpVT.getVectorNumElements();
11695 SDValue ShiftBits = DAG.getConstant(NumElems/2, dl, MVT::i8);
11697 if (IdxVal == OpVT.getVectorNumElements() / 2) {
11698 // Zero upper bits of the Vec
11699 Vec = DAG.getNode(X86ISD::VSHLI, dl, OpVT, Vec, ShiftBits);
11700 Vec = DAG.getNode(X86ISD::VSRLI, dl, OpVT, Vec, ShiftBits);
11702 SDValue Vec2 = DAG.getNode(ISD::INSERT_SUBVECTOR, dl, OpVT, Undef,
11704 Vec2 = DAG.getNode(X86ISD::VSHLI, dl, OpVT, Vec2, ShiftBits);
11705 return DAG.getNode(ISD::OR, dl, OpVT, Vec, Vec2);
11708 SDValue Vec2 = DAG.getNode(ISD::INSERT_SUBVECTOR, dl, OpVT, Undef,
11710 // Zero upper bits of the Vec2
11711 Vec2 = DAG.getNode(X86ISD::VSHLI, dl, OpVT, Vec2, ShiftBits);
11712 Vec2 = DAG.getNode(X86ISD::VSRLI, dl, OpVT, Vec2, ShiftBits);
11713 // Zero lower bits of the Vec
11714 Vec = DAG.getNode(X86ISD::VSRLI, dl, OpVT, Vec, ShiftBits);
11715 Vec = DAG.getNode(X86ISD::VSHLI, dl, OpVT, Vec, ShiftBits);
11716 // Merge them together
11717 return DAG.getNode(ISD::OR, dl, OpVT, Vec, Vec2);
11723 // ConstantPool, JumpTable, GlobalAddress, and ExternalSymbol are lowered as
11724 // their target countpart wrapped in the X86ISD::Wrapper node. Suppose N is
11725 // one of the above mentioned nodes. It has to be wrapped because otherwise
11726 // Select(N) returns N. So the raw TargetGlobalAddress nodes, etc. can only
11727 // be used to form addressing mode. These wrapped nodes will be selected
11730 X86TargetLowering::LowerConstantPool(SDValue Op, SelectionDAG &DAG) const {
11731 ConstantPoolSDNode *CP = cast<ConstantPoolSDNode>(Op);
11733 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
11734 // global base reg.
11735 unsigned char OpFlag = 0;
11736 unsigned WrapperKind = X86ISD::Wrapper;
11737 CodeModel::Model M = DAG.getTarget().getCodeModel();
11739 if (Subtarget->isPICStyleRIPRel() &&
11740 (M == CodeModel::Small || M == CodeModel::Kernel))
11741 WrapperKind = X86ISD::WrapperRIP;
11742 else if (Subtarget->isPICStyleGOT())
11743 OpFlag = X86II::MO_GOTOFF;
11744 else if (Subtarget->isPICStyleStubPIC())
11745 OpFlag = X86II::MO_PIC_BASE_OFFSET;
11747 auto PtrVT = getPointerTy(DAG.getDataLayout());
11748 SDValue Result = DAG.getTargetConstantPool(
11749 CP->getConstVal(), PtrVT, CP->getAlignment(), CP->getOffset(), OpFlag);
11751 Result = DAG.getNode(WrapperKind, DL, PtrVT, Result);
11752 // With PIC, the address is actually $g + Offset.
11755 DAG.getNode(ISD::ADD, DL, PtrVT,
11756 DAG.getNode(X86ISD::GlobalBaseReg, SDLoc(), PtrVT), Result);
11762 SDValue X86TargetLowering::LowerJumpTable(SDValue Op, SelectionDAG &DAG) const {
11763 JumpTableSDNode *JT = cast<JumpTableSDNode>(Op);
11765 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
11766 // global base reg.
11767 unsigned char OpFlag = 0;
11768 unsigned WrapperKind = X86ISD::Wrapper;
11769 CodeModel::Model M = DAG.getTarget().getCodeModel();
11771 if (Subtarget->isPICStyleRIPRel() &&
11772 (M == CodeModel::Small || M == CodeModel::Kernel))
11773 WrapperKind = X86ISD::WrapperRIP;
11774 else if (Subtarget->isPICStyleGOT())
11775 OpFlag = X86II::MO_GOTOFF;
11776 else if (Subtarget->isPICStyleStubPIC())
11777 OpFlag = X86II::MO_PIC_BASE_OFFSET;
11779 auto PtrVT = getPointerTy(DAG.getDataLayout());
11780 SDValue Result = DAG.getTargetJumpTable(JT->getIndex(), PtrVT, OpFlag);
11782 Result = DAG.getNode(WrapperKind, DL, PtrVT, Result);
11784 // With PIC, the address is actually $g + Offset.
11787 DAG.getNode(ISD::ADD, DL, PtrVT,
11788 DAG.getNode(X86ISD::GlobalBaseReg, SDLoc(), PtrVT), Result);
11794 X86TargetLowering::LowerExternalSymbol(SDValue Op, SelectionDAG &DAG) const {
11795 const char *Sym = cast<ExternalSymbolSDNode>(Op)->getSymbol();
11797 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
11798 // global base reg.
11799 unsigned char OpFlag = 0;
11800 unsigned WrapperKind = X86ISD::Wrapper;
11801 CodeModel::Model M = DAG.getTarget().getCodeModel();
11803 if (Subtarget->isPICStyleRIPRel() &&
11804 (M == CodeModel::Small || M == CodeModel::Kernel)) {
11805 if (Subtarget->isTargetDarwin() || Subtarget->isTargetELF())
11806 OpFlag = X86II::MO_GOTPCREL;
11807 WrapperKind = X86ISD::WrapperRIP;
11808 } else if (Subtarget->isPICStyleGOT()) {
11809 OpFlag = X86II::MO_GOT;
11810 } else if (Subtarget->isPICStyleStubPIC()) {
11811 OpFlag = X86II::MO_DARWIN_NONLAZY_PIC_BASE;
11812 } else if (Subtarget->isPICStyleStubNoDynamic()) {
11813 OpFlag = X86II::MO_DARWIN_NONLAZY;
11816 auto PtrVT = getPointerTy(DAG.getDataLayout());
11817 SDValue Result = DAG.getTargetExternalSymbol(Sym, PtrVT, OpFlag);
11820 Result = DAG.getNode(WrapperKind, DL, PtrVT, Result);
11822 // With PIC, the address is actually $g + Offset.
11823 if (DAG.getTarget().getRelocationModel() == Reloc::PIC_ &&
11824 !Subtarget->is64Bit()) {
11826 DAG.getNode(ISD::ADD, DL, PtrVT,
11827 DAG.getNode(X86ISD::GlobalBaseReg, SDLoc(), PtrVT), Result);
11830 // For symbols that require a load from a stub to get the address, emit the
11832 if (isGlobalStubReference(OpFlag))
11833 Result = DAG.getLoad(PtrVT, DL, DAG.getEntryNode(), Result,
11834 MachinePointerInfo::getGOT(DAG.getMachineFunction()),
11835 false, false, false, 0);
11841 X86TargetLowering::LowerBlockAddress(SDValue Op, SelectionDAG &DAG) const {
11842 // Create the TargetBlockAddressAddress node.
11843 unsigned char OpFlags =
11844 Subtarget->ClassifyBlockAddressReference();
11845 CodeModel::Model M = DAG.getTarget().getCodeModel();
11846 const BlockAddress *BA = cast<BlockAddressSDNode>(Op)->getBlockAddress();
11847 int64_t Offset = cast<BlockAddressSDNode>(Op)->getOffset();
11849 auto PtrVT = getPointerTy(DAG.getDataLayout());
11850 SDValue Result = DAG.getTargetBlockAddress(BA, PtrVT, Offset, OpFlags);
11852 if (Subtarget->isPICStyleRIPRel() &&
11853 (M == CodeModel::Small || M == CodeModel::Kernel))
11854 Result = DAG.getNode(X86ISD::WrapperRIP, dl, PtrVT, Result);
11856 Result = DAG.getNode(X86ISD::Wrapper, dl, PtrVT, Result);
11858 // With PIC, the address is actually $g + Offset.
11859 if (isGlobalRelativeToPICBase(OpFlags)) {
11860 Result = DAG.getNode(ISD::ADD, dl, PtrVT,
11861 DAG.getNode(X86ISD::GlobalBaseReg, dl, PtrVT), Result);
11868 X86TargetLowering::LowerGlobalAddress(const GlobalValue *GV, SDLoc dl,
11869 int64_t Offset, SelectionDAG &DAG) const {
11870 // Create the TargetGlobalAddress node, folding in the constant
11871 // offset if it is legal.
11872 unsigned char OpFlags =
11873 Subtarget->ClassifyGlobalReference(GV, DAG.getTarget());
11874 CodeModel::Model M = DAG.getTarget().getCodeModel();
11875 auto PtrVT = getPointerTy(DAG.getDataLayout());
11877 if (OpFlags == X86II::MO_NO_FLAG &&
11878 X86::isOffsetSuitableForCodeModel(Offset, M)) {
11879 // A direct static reference to a global.
11880 Result = DAG.getTargetGlobalAddress(GV, dl, PtrVT, Offset);
11883 Result = DAG.getTargetGlobalAddress(GV, dl, PtrVT, 0, OpFlags);
11886 if (Subtarget->isPICStyleRIPRel() &&
11887 (M == CodeModel::Small || M == CodeModel::Kernel))
11888 Result = DAG.getNode(X86ISD::WrapperRIP, dl, PtrVT, Result);
11890 Result = DAG.getNode(X86ISD::Wrapper, dl, PtrVT, Result);
11892 // With PIC, the address is actually $g + Offset.
11893 if (isGlobalRelativeToPICBase(OpFlags)) {
11894 Result = DAG.getNode(ISD::ADD, dl, PtrVT,
11895 DAG.getNode(X86ISD::GlobalBaseReg, dl, PtrVT), Result);
11898 // For globals that require a load from a stub to get the address, emit the
11900 if (isGlobalStubReference(OpFlags))
11901 Result = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Result,
11902 MachinePointerInfo::getGOT(DAG.getMachineFunction()),
11903 false, false, false, 0);
11905 // If there was a non-zero offset that we didn't fold, create an explicit
11906 // addition for it.
11908 Result = DAG.getNode(ISD::ADD, dl, PtrVT, Result,
11909 DAG.getConstant(Offset, dl, PtrVT));
11915 X86TargetLowering::LowerGlobalAddress(SDValue Op, SelectionDAG &DAG) const {
11916 const GlobalValue *GV = cast<GlobalAddressSDNode>(Op)->getGlobal();
11917 int64_t Offset = cast<GlobalAddressSDNode>(Op)->getOffset();
11918 return LowerGlobalAddress(GV, SDLoc(Op), Offset, DAG);
11922 GetTLSADDR(SelectionDAG &DAG, SDValue Chain, GlobalAddressSDNode *GA,
11923 SDValue *InFlag, const EVT PtrVT, unsigned ReturnReg,
11924 unsigned char OperandFlags, bool LocalDynamic = false) {
11925 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
11926 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
11928 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(), dl,
11929 GA->getValueType(0),
11933 X86ISD::NodeType CallType = LocalDynamic ? X86ISD::TLSBASEADDR
11937 SDValue Ops[] = { Chain, TGA, *InFlag };
11938 Chain = DAG.getNode(CallType, dl, NodeTys, Ops);
11940 SDValue Ops[] = { Chain, TGA };
11941 Chain = DAG.getNode(CallType, dl, NodeTys, Ops);
11944 // TLSADDR will be codegen'ed as call. Inform MFI that function has calls.
11945 MFI->setAdjustsStack(true);
11946 MFI->setHasCalls(true);
11948 SDValue Flag = Chain.getValue(1);
11949 return DAG.getCopyFromReg(Chain, dl, ReturnReg, PtrVT, Flag);
11952 // Lower ISD::GlobalTLSAddress using the "general dynamic" model, 32 bit
11954 LowerToTLSGeneralDynamicModel32(GlobalAddressSDNode *GA, SelectionDAG &DAG,
11957 SDLoc dl(GA); // ? function entry point might be better
11958 SDValue Chain = DAG.getCopyToReg(DAG.getEntryNode(), dl, X86::EBX,
11959 DAG.getNode(X86ISD::GlobalBaseReg,
11960 SDLoc(), PtrVT), InFlag);
11961 InFlag = Chain.getValue(1);
11963 return GetTLSADDR(DAG, Chain, GA, &InFlag, PtrVT, X86::EAX, X86II::MO_TLSGD);
11966 // Lower ISD::GlobalTLSAddress using the "general dynamic" model, 64 bit
11968 LowerToTLSGeneralDynamicModel64(GlobalAddressSDNode *GA, SelectionDAG &DAG,
11970 return GetTLSADDR(DAG, DAG.getEntryNode(), GA, nullptr, PtrVT,
11971 X86::RAX, X86II::MO_TLSGD);
11974 static SDValue LowerToTLSLocalDynamicModel(GlobalAddressSDNode *GA,
11980 // Get the start address of the TLS block for this module.
11981 X86MachineFunctionInfo* MFI = DAG.getMachineFunction()
11982 .getInfo<X86MachineFunctionInfo>();
11983 MFI->incNumLocalDynamicTLSAccesses();
11987 Base = GetTLSADDR(DAG, DAG.getEntryNode(), GA, nullptr, PtrVT, X86::RAX,
11988 X86II::MO_TLSLD, /*LocalDynamic=*/true);
11991 SDValue Chain = DAG.getCopyToReg(DAG.getEntryNode(), dl, X86::EBX,
11992 DAG.getNode(X86ISD::GlobalBaseReg, SDLoc(), PtrVT), InFlag);
11993 InFlag = Chain.getValue(1);
11994 Base = GetTLSADDR(DAG, Chain, GA, &InFlag, PtrVT, X86::EAX,
11995 X86II::MO_TLSLDM, /*LocalDynamic=*/true);
11998 // Note: the CleanupLocalDynamicTLSPass will remove redundant computations
12002 unsigned char OperandFlags = X86II::MO_DTPOFF;
12003 unsigned WrapperKind = X86ISD::Wrapper;
12004 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(), dl,
12005 GA->getValueType(0),
12006 GA->getOffset(), OperandFlags);
12007 SDValue Offset = DAG.getNode(WrapperKind, dl, PtrVT, TGA);
12009 // Add x@dtpoff with the base.
12010 return DAG.getNode(ISD::ADD, dl, PtrVT, Offset, Base);
12013 // Lower ISD::GlobalTLSAddress using the "initial exec" or "local exec" model.
12014 static SDValue LowerToTLSExecModel(GlobalAddressSDNode *GA, SelectionDAG &DAG,
12015 const EVT PtrVT, TLSModel::Model model,
12016 bool is64Bit, bool isPIC) {
12019 // Get the Thread Pointer, which is %gs:0 (32-bit) or %fs:0 (64-bit).
12020 Value *Ptr = Constant::getNullValue(Type::getInt8PtrTy(*DAG.getContext(),
12021 is64Bit ? 257 : 256));
12023 SDValue ThreadPointer =
12024 DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), DAG.getIntPtrConstant(0, dl),
12025 MachinePointerInfo(Ptr), false, false, false, 0);
12027 unsigned char OperandFlags = 0;
12028 // Most TLS accesses are not RIP relative, even on x86-64. One exception is
12030 unsigned WrapperKind = X86ISD::Wrapper;
12031 if (model == TLSModel::LocalExec) {
12032 OperandFlags = is64Bit ? X86II::MO_TPOFF : X86II::MO_NTPOFF;
12033 } else if (model == TLSModel::InitialExec) {
12035 OperandFlags = X86II::MO_GOTTPOFF;
12036 WrapperKind = X86ISD::WrapperRIP;
12038 OperandFlags = isPIC ? X86II::MO_GOTNTPOFF : X86II::MO_INDNTPOFF;
12041 llvm_unreachable("Unexpected model");
12044 // emit "addl x@ntpoff,%eax" (local exec)
12045 // or "addl x@indntpoff,%eax" (initial exec)
12046 // or "addl x@gotntpoff(%ebx) ,%eax" (initial exec, 32-bit pic)
12048 DAG.getTargetGlobalAddress(GA->getGlobal(), dl, GA->getValueType(0),
12049 GA->getOffset(), OperandFlags);
12050 SDValue Offset = DAG.getNode(WrapperKind, dl, PtrVT, TGA);
12052 if (model == TLSModel::InitialExec) {
12053 if (isPIC && !is64Bit) {
12054 Offset = DAG.getNode(ISD::ADD, dl, PtrVT,
12055 DAG.getNode(X86ISD::GlobalBaseReg, SDLoc(), PtrVT),
12059 Offset = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(), Offset,
12060 MachinePointerInfo::getGOT(DAG.getMachineFunction()),
12061 false, false, false, 0);
12064 // The address of the thread local variable is the add of the thread
12065 // pointer with the offset of the variable.
12066 return DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, Offset);
12070 X86TargetLowering::LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const {
12072 GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
12073 const GlobalValue *GV = GA->getGlobal();
12074 auto PtrVT = getPointerTy(DAG.getDataLayout());
12076 if (Subtarget->isTargetELF()) {
12077 if (DAG.getTarget().Options.EmulatedTLS)
12078 return LowerToTLSEmulatedModel(GA, DAG);
12079 TLSModel::Model model = DAG.getTarget().getTLSModel(GV);
12081 case TLSModel::GeneralDynamic:
12082 if (Subtarget->is64Bit())
12083 return LowerToTLSGeneralDynamicModel64(GA, DAG, PtrVT);
12084 return LowerToTLSGeneralDynamicModel32(GA, DAG, PtrVT);
12085 case TLSModel::LocalDynamic:
12086 return LowerToTLSLocalDynamicModel(GA, DAG, PtrVT,
12087 Subtarget->is64Bit());
12088 case TLSModel::InitialExec:
12089 case TLSModel::LocalExec:
12090 return LowerToTLSExecModel(GA, DAG, PtrVT, model, Subtarget->is64Bit(),
12091 DAG.getTarget().getRelocationModel() ==
12094 llvm_unreachable("Unknown TLS model.");
12097 if (Subtarget->isTargetDarwin()) {
12098 // Darwin only has one model of TLS. Lower to that.
12099 unsigned char OpFlag = 0;
12100 unsigned WrapperKind = Subtarget->isPICStyleRIPRel() ?
12101 X86ISD::WrapperRIP : X86ISD::Wrapper;
12103 // In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
12104 // global base reg.
12105 bool PIC32 = (DAG.getTarget().getRelocationModel() == Reloc::PIC_) &&
12106 !Subtarget->is64Bit();
12108 OpFlag = X86II::MO_TLVP_PIC_BASE;
12110 OpFlag = X86II::MO_TLVP;
12112 SDValue Result = DAG.getTargetGlobalAddress(GA->getGlobal(), DL,
12113 GA->getValueType(0),
12114 GA->getOffset(), OpFlag);
12115 SDValue Offset = DAG.getNode(WrapperKind, DL, PtrVT, Result);
12117 // With PIC32, the address is actually $g + Offset.
12119 Offset = DAG.getNode(ISD::ADD, DL, PtrVT,
12120 DAG.getNode(X86ISD::GlobalBaseReg, SDLoc(), PtrVT),
12123 // Lowering the machine isd will make sure everything is in the right
12125 SDValue Chain = DAG.getEntryNode();
12126 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
12127 SDValue Args[] = { Chain, Offset };
12128 Chain = DAG.getNode(X86ISD::TLSCALL, DL, NodeTys, Args);
12130 // TLSCALL will be codegen'ed as call. Inform MFI that function has calls.
12131 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
12132 MFI->setAdjustsStack(true);
12134 // And our return value (tls address) is in the standard call return value
12136 unsigned Reg = Subtarget->is64Bit() ? X86::RAX : X86::EAX;
12137 return DAG.getCopyFromReg(Chain, DL, Reg, PtrVT, Chain.getValue(1));
12140 if (Subtarget->isTargetKnownWindowsMSVC() ||
12141 Subtarget->isTargetWindowsGNU()) {
12142 // Just use the implicit TLS architecture
12143 // Need to generate someting similar to:
12144 // mov rdx, qword [gs:abs 58H]; Load pointer to ThreadLocalStorage
12146 // mov ecx, dword [rel _tls_index]: Load index (from C runtime)
12147 // mov rcx, qword [rdx+rcx*8]
12148 // mov eax, .tls$:tlsvar
12149 // [rax+rcx] contains the address
12150 // Windows 64bit: gs:0x58
12151 // Windows 32bit: fs:__tls_array
12154 SDValue Chain = DAG.getEntryNode();
12156 // Get the Thread Pointer, which is %fs:__tls_array (32-bit) or
12157 // %gs:0x58 (64-bit). On MinGW, __tls_array is not available, so directly
12158 // use its literal value of 0x2C.
12159 Value *Ptr = Constant::getNullValue(Subtarget->is64Bit()
12160 ? Type::getInt8PtrTy(*DAG.getContext(),
12162 : Type::getInt32PtrTy(*DAG.getContext(),
12165 SDValue TlsArray = Subtarget->is64Bit()
12166 ? DAG.getIntPtrConstant(0x58, dl)
12167 : (Subtarget->isTargetWindowsGNU()
12168 ? DAG.getIntPtrConstant(0x2C, dl)
12169 : DAG.getExternalSymbol("_tls_array", PtrVT));
12171 SDValue ThreadPointer =
12172 DAG.getLoad(PtrVT, dl, Chain, TlsArray, MachinePointerInfo(Ptr), false,
12176 if (GV->getThreadLocalMode() == GlobalVariable::LocalExecTLSModel) {
12177 res = ThreadPointer;
12179 // Load the _tls_index variable
12180 SDValue IDX = DAG.getExternalSymbol("_tls_index", PtrVT);
12181 if (Subtarget->is64Bit())
12182 IDX = DAG.getExtLoad(ISD::ZEXTLOAD, dl, PtrVT, Chain, IDX,
12183 MachinePointerInfo(), MVT::i32, false, false,
12186 IDX = DAG.getLoad(PtrVT, dl, Chain, IDX, MachinePointerInfo(), false,
12189 auto &DL = DAG.getDataLayout();
12191 DAG.getConstant(Log2_64_Ceil(DL.getPointerSize()), dl, PtrVT);
12192 IDX = DAG.getNode(ISD::SHL, dl, PtrVT, IDX, Scale);
12194 res = DAG.getNode(ISD::ADD, dl, PtrVT, ThreadPointer, IDX);
12197 res = DAG.getLoad(PtrVT, dl, Chain, res, MachinePointerInfo(), false, false,
12200 // Get the offset of start of .tls section
12201 SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(), dl,
12202 GA->getValueType(0),
12203 GA->getOffset(), X86II::MO_SECREL);
12204 SDValue Offset = DAG.getNode(X86ISD::Wrapper, dl, PtrVT, TGA);
12206 // The address of the thread local variable is the add of the thread
12207 // pointer with the offset of the variable.
12208 return DAG.getNode(ISD::ADD, dl, PtrVT, res, Offset);
12211 llvm_unreachable("TLS not implemented for this target.");
12214 /// LowerShiftParts - Lower SRA_PARTS and friends, which return two i32 values
12215 /// and take a 2 x i32 value to shift plus a shift amount.
12216 static SDValue LowerShiftParts(SDValue Op, SelectionDAG &DAG) {
12217 assert(Op.getNumOperands() == 3 && "Not a double-shift!");
12218 MVT VT = Op.getSimpleValueType();
12219 unsigned VTBits = VT.getSizeInBits();
12221 bool isSRA = Op.getOpcode() == ISD::SRA_PARTS;
12222 SDValue ShOpLo = Op.getOperand(0);
12223 SDValue ShOpHi = Op.getOperand(1);
12224 SDValue ShAmt = Op.getOperand(2);
12225 // X86ISD::SHLD and X86ISD::SHRD have defined overflow behavior but the
12226 // generic ISD nodes haven't. Insert an AND to be safe, it's optimized away
12228 SDValue SafeShAmt = DAG.getNode(ISD::AND, dl, MVT::i8, ShAmt,
12229 DAG.getConstant(VTBits - 1, dl, MVT::i8));
12230 SDValue Tmp1 = isSRA ? DAG.getNode(ISD::SRA, dl, VT, ShOpHi,
12231 DAG.getConstant(VTBits - 1, dl, MVT::i8))
12232 : DAG.getConstant(0, dl, VT);
12234 SDValue Tmp2, Tmp3;
12235 if (Op.getOpcode() == ISD::SHL_PARTS) {
12236 Tmp2 = DAG.getNode(X86ISD::SHLD, dl, VT, ShOpHi, ShOpLo, ShAmt);
12237 Tmp3 = DAG.getNode(ISD::SHL, dl, VT, ShOpLo, SafeShAmt);
12239 Tmp2 = DAG.getNode(X86ISD::SHRD, dl, VT, ShOpLo, ShOpHi, ShAmt);
12240 Tmp3 = DAG.getNode(isSRA ? ISD::SRA : ISD::SRL, dl, VT, ShOpHi, SafeShAmt);
12243 // If the shift amount is larger or equal than the width of a part we can't
12244 // rely on the results of shld/shrd. Insert a test and select the appropriate
12245 // values for large shift amounts.
12246 SDValue AndNode = DAG.getNode(ISD::AND, dl, MVT::i8, ShAmt,
12247 DAG.getConstant(VTBits, dl, MVT::i8));
12248 SDValue Cond = DAG.getNode(X86ISD::CMP, dl, MVT::i32,
12249 AndNode, DAG.getConstant(0, dl, MVT::i8));
12252 SDValue CC = DAG.getConstant(X86::COND_NE, dl, MVT::i8);
12253 SDValue Ops0[4] = { Tmp2, Tmp3, CC, Cond };
12254 SDValue Ops1[4] = { Tmp3, Tmp1, CC, Cond };
12256 if (Op.getOpcode() == ISD::SHL_PARTS) {
12257 Hi = DAG.getNode(X86ISD::CMOV, dl, VT, Ops0);
12258 Lo = DAG.getNode(X86ISD::CMOV, dl, VT, Ops1);
12260 Lo = DAG.getNode(X86ISD::CMOV, dl, VT, Ops0);
12261 Hi = DAG.getNode(X86ISD::CMOV, dl, VT, Ops1);
12264 SDValue Ops[2] = { Lo, Hi };
12265 return DAG.getMergeValues(Ops, dl);
12268 SDValue X86TargetLowering::LowerSINT_TO_FP(SDValue Op,
12269 SelectionDAG &DAG) const {
12270 SDValue Src = Op.getOperand(0);
12271 MVT SrcVT = Src.getSimpleValueType();
12272 MVT VT = Op.getSimpleValueType();
12275 if (SrcVT.isVector()) {
12276 if (SrcVT == MVT::v2i32 && VT == MVT::v2f64) {
12277 return DAG.getNode(X86ISD::CVTDQ2PD, dl, VT,
12278 DAG.getNode(ISD::CONCAT_VECTORS, dl, MVT::v4i32, Src,
12279 DAG.getUNDEF(SrcVT)));
12281 if (SrcVT.getVectorElementType() == MVT::i1) {
12282 MVT IntegerVT = MVT::getVectorVT(MVT::i32, SrcVT.getVectorNumElements());
12283 return DAG.getNode(ISD::SINT_TO_FP, dl, Op.getValueType(),
12284 DAG.getNode(ISD::SIGN_EXTEND, dl, IntegerVT, Src));
12289 assert(SrcVT <= MVT::i64 && SrcVT >= MVT::i16 &&
12290 "Unknown SINT_TO_FP to lower!");
12292 // These are really Legal; return the operand so the caller accepts it as
12294 if (SrcVT == MVT::i32 && isScalarFPTypeInSSEReg(Op.getValueType()))
12296 if (SrcVT == MVT::i64 && isScalarFPTypeInSSEReg(Op.getValueType()) &&
12297 Subtarget->is64Bit()) {
12301 unsigned Size = SrcVT.getSizeInBits()/8;
12302 MachineFunction &MF = DAG.getMachineFunction();
12303 auto PtrVT = getPointerTy(MF.getDataLayout());
12304 int SSFI = MF.getFrameInfo()->CreateStackObject(Size, Size, false);
12305 SDValue StackSlot = DAG.getFrameIndex(SSFI, PtrVT);
12306 SDValue Chain = DAG.getStore(
12307 DAG.getEntryNode(), dl, Op.getOperand(0), StackSlot,
12308 MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), SSFI), false,
12310 return BuildFILD(Op, SrcVT, Chain, StackSlot, DAG);
12313 SDValue X86TargetLowering::BuildFILD(SDValue Op, EVT SrcVT, SDValue Chain,
12315 SelectionDAG &DAG) const {
12319 bool useSSE = isScalarFPTypeInSSEReg(Op.getValueType());
12321 Tys = DAG.getVTList(MVT::f64, MVT::Other, MVT::Glue);
12323 Tys = DAG.getVTList(Op.getValueType(), MVT::Other);
12325 unsigned ByteSize = SrcVT.getSizeInBits()/8;
12327 FrameIndexSDNode *FI = dyn_cast<FrameIndexSDNode>(StackSlot);
12328 MachineMemOperand *MMO;
12330 int SSFI = FI->getIndex();
12331 MMO = DAG.getMachineFunction().getMachineMemOperand(
12332 MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), SSFI),
12333 MachineMemOperand::MOLoad, ByteSize, ByteSize);
12335 MMO = cast<LoadSDNode>(StackSlot)->getMemOperand();
12336 StackSlot = StackSlot.getOperand(1);
12338 SDValue Ops[] = { Chain, StackSlot, DAG.getValueType(SrcVT) };
12339 SDValue Result = DAG.getMemIntrinsicNode(useSSE ? X86ISD::FILD_FLAG :
12341 Tys, Ops, SrcVT, MMO);
12344 Chain = Result.getValue(1);
12345 SDValue InFlag = Result.getValue(2);
12347 // FIXME: Currently the FST is flagged to the FILD_FLAG. This
12348 // shouldn't be necessary except that RFP cannot be live across
12349 // multiple blocks. When stackifier is fixed, they can be uncoupled.
12350 MachineFunction &MF = DAG.getMachineFunction();
12351 unsigned SSFISize = Op.getValueType().getSizeInBits()/8;
12352 int SSFI = MF.getFrameInfo()->CreateStackObject(SSFISize, SSFISize, false);
12353 auto PtrVT = getPointerTy(MF.getDataLayout());
12354 SDValue StackSlot = DAG.getFrameIndex(SSFI, PtrVT);
12355 Tys = DAG.getVTList(MVT::Other);
12357 Chain, Result, StackSlot, DAG.getValueType(Op.getValueType()), InFlag
12359 MachineMemOperand *MMO = DAG.getMachineFunction().getMachineMemOperand(
12360 MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), SSFI),
12361 MachineMemOperand::MOStore, SSFISize, SSFISize);
12363 Chain = DAG.getMemIntrinsicNode(X86ISD::FST, DL, Tys,
12364 Ops, Op.getValueType(), MMO);
12365 Result = DAG.getLoad(
12366 Op.getValueType(), DL, Chain, StackSlot,
12367 MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), SSFI),
12368 false, false, false, 0);
12374 // LowerUINT_TO_FP_i64 - 64-bit unsigned integer to double expansion.
12375 SDValue X86TargetLowering::LowerUINT_TO_FP_i64(SDValue Op,
12376 SelectionDAG &DAG) const {
12377 // This algorithm is not obvious. Here it is what we're trying to output:
12380 punpckldq (c0), %xmm0 // c0: (uint4){ 0x43300000U, 0x45300000U, 0U, 0U }
12381 subpd (c1), %xmm0 // c1: (double2){ 0x1.0p52, 0x1.0p52 * 0x1.0p32 }
12383 haddpd %xmm0, %xmm0
12385 pshufd $0x4e, %xmm0, %xmm1
12391 LLVMContext *Context = DAG.getContext();
12393 // Build some magic constants.
12394 static const uint32_t CV0[] = { 0x43300000, 0x45300000, 0, 0 };
12395 Constant *C0 = ConstantDataVector::get(*Context, CV0);
12396 auto PtrVT = getPointerTy(DAG.getDataLayout());
12397 SDValue CPIdx0 = DAG.getConstantPool(C0, PtrVT, 16);
12399 SmallVector<Constant*,2> CV1;
12401 ConstantFP::get(*Context, APFloat(APFloat::IEEEdouble,
12402 APInt(64, 0x4330000000000000ULL))));
12404 ConstantFP::get(*Context, APFloat(APFloat::IEEEdouble,
12405 APInt(64, 0x4530000000000000ULL))));
12406 Constant *C1 = ConstantVector::get(CV1);
12407 SDValue CPIdx1 = DAG.getConstantPool(C1, PtrVT, 16);
12409 // Load the 64-bit value into an XMM register.
12410 SDValue XR1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i64,
12413 DAG.getLoad(MVT::v4i32, dl, DAG.getEntryNode(), CPIdx0,
12414 MachinePointerInfo::getConstantPool(DAG.getMachineFunction()),
12415 false, false, false, 16);
12417 getUnpackl(DAG, dl, MVT::v4i32, DAG.getBitcast(MVT::v4i32, XR1), CLod0);
12420 DAG.getLoad(MVT::v2f64, dl, CLod0.getValue(1), CPIdx1,
12421 MachinePointerInfo::getConstantPool(DAG.getMachineFunction()),
12422 false, false, false, 16);
12423 SDValue XR2F = DAG.getBitcast(MVT::v2f64, Unpck1);
12424 // TODO: Are there any fast-math-flags to propagate here?
12425 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::v2f64, XR2F, CLod1);
12428 if (Subtarget->hasSSE3()) {
12429 // FIXME: The 'haddpd' instruction may be slower than 'movhlps + addsd'.
12430 Result = DAG.getNode(X86ISD::FHADD, dl, MVT::v2f64, Sub, Sub);
12432 SDValue S2F = DAG.getBitcast(MVT::v4i32, Sub);
12433 SDValue Shuffle = getTargetShuffleNode(X86ISD::PSHUFD, dl, MVT::v4i32,
12435 Result = DAG.getNode(ISD::FADD, dl, MVT::v2f64,
12436 DAG.getBitcast(MVT::v2f64, Shuffle), Sub);
12439 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, Result,
12440 DAG.getIntPtrConstant(0, dl));
12443 // LowerUINT_TO_FP_i32 - 32-bit unsigned integer to float expansion.
12444 SDValue X86TargetLowering::LowerUINT_TO_FP_i32(SDValue Op,
12445 SelectionDAG &DAG) const {
12447 // FP constant to bias correct the final result.
12448 SDValue Bias = DAG.getConstantFP(BitsToDouble(0x4330000000000000ULL), dl,
12451 // Load the 32-bit value into an XMM register.
12452 SDValue Load = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,
12455 // Zero out the upper parts of the register.
12456 Load = getShuffleVectorZeroOrUndef(Load, 0, true, Subtarget, DAG);
12458 Load = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
12459 DAG.getBitcast(MVT::v2f64, Load),
12460 DAG.getIntPtrConstant(0, dl));
12462 // Or the load with the bias.
12463 SDValue Or = DAG.getNode(
12464 ISD::OR, dl, MVT::v2i64,
12465 DAG.getBitcast(MVT::v2i64,
12466 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2f64, Load)),
12467 DAG.getBitcast(MVT::v2i64,
12468 DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2f64, Bias)));
12470 DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
12471 DAG.getBitcast(MVT::v2f64, Or), DAG.getIntPtrConstant(0, dl));
12473 // Subtract the bias.
12474 // TODO: Are there any fast-math-flags to propagate here?
12475 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::f64, Or, Bias);
12477 // Handle final rounding.
12478 EVT DestVT = Op.getValueType();
12480 if (DestVT.bitsLT(MVT::f64))
12481 return DAG.getNode(ISD::FP_ROUND, dl, DestVT, Sub,
12482 DAG.getIntPtrConstant(0, dl));
12483 if (DestVT.bitsGT(MVT::f64))
12484 return DAG.getNode(ISD::FP_EXTEND, dl, DestVT, Sub);
12486 // Handle final rounding.
12490 static SDValue lowerUINT_TO_FP_vXi32(SDValue Op, SelectionDAG &DAG,
12491 const X86Subtarget &Subtarget) {
12492 // The algorithm is the following:
12493 // #ifdef __SSE4_1__
12494 // uint4 lo = _mm_blend_epi16( v, (uint4) 0x4b000000, 0xaa);
12495 // uint4 hi = _mm_blend_epi16( _mm_srli_epi32(v,16),
12496 // (uint4) 0x53000000, 0xaa);
12498 // uint4 lo = (v & (uint4) 0xffff) | (uint4) 0x4b000000;
12499 // uint4 hi = (v >> 16) | (uint4) 0x53000000;
12501 // float4 fhi = (float4) hi - (0x1.0p39f + 0x1.0p23f);
12502 // return (float4) lo + fhi;
12504 // We shouldn't use it when unsafe-fp-math is enabled though: we might later
12505 // reassociate the two FADDs, and if we do that, the algorithm fails
12506 // spectacularly (PR24512).
12507 // FIXME: If we ever have some kind of Machine FMF, this should be marked
12508 // as non-fast and always be enabled. Why isn't SDAG FMF enough? Because
12509 // there's also the MachineCombiner reassociations happening on Machine IR.
12510 if (DAG.getTarget().Options.UnsafeFPMath)
12514 SDValue V = Op->getOperand(0);
12515 EVT VecIntVT = V.getValueType();
12516 bool Is128 = VecIntVT == MVT::v4i32;
12517 EVT VecFloatVT = Is128 ? MVT::v4f32 : MVT::v8f32;
12518 // If we convert to something else than the supported type, e.g., to v4f64,
12520 if (VecFloatVT != Op->getValueType(0))
12523 unsigned NumElts = VecIntVT.getVectorNumElements();
12524 assert((VecIntVT == MVT::v4i32 || VecIntVT == MVT::v8i32) &&
12525 "Unsupported custom type");
12526 assert(NumElts <= 8 && "The size of the constant array must be fixed");
12528 // In the #idef/#else code, we have in common:
12529 // - The vector of constants:
12535 // Create the splat vector for 0x4b000000.
12536 SDValue CstLow = DAG.getConstant(0x4b000000, DL, MVT::i32);
12537 SDValue CstLowArray[] = {CstLow, CstLow, CstLow, CstLow,
12538 CstLow, CstLow, CstLow, CstLow};
12539 SDValue VecCstLow = DAG.getNode(ISD::BUILD_VECTOR, DL, VecIntVT,
12540 makeArrayRef(&CstLowArray[0], NumElts));
12541 // Create the splat vector for 0x53000000.
12542 SDValue CstHigh = DAG.getConstant(0x53000000, DL, MVT::i32);
12543 SDValue CstHighArray[] = {CstHigh, CstHigh, CstHigh, CstHigh,
12544 CstHigh, CstHigh, CstHigh, CstHigh};
12545 SDValue VecCstHigh = DAG.getNode(ISD::BUILD_VECTOR, DL, VecIntVT,
12546 makeArrayRef(&CstHighArray[0], NumElts));
12548 // Create the right shift.
12549 SDValue CstShift = DAG.getConstant(16, DL, MVT::i32);
12550 SDValue CstShiftArray[] = {CstShift, CstShift, CstShift, CstShift,
12551 CstShift, CstShift, CstShift, CstShift};
12552 SDValue VecCstShift = DAG.getNode(ISD::BUILD_VECTOR, DL, VecIntVT,
12553 makeArrayRef(&CstShiftArray[0], NumElts));
12554 SDValue HighShift = DAG.getNode(ISD::SRL, DL, VecIntVT, V, VecCstShift);
12557 if (Subtarget.hasSSE41()) {
12558 EVT VecI16VT = Is128 ? MVT::v8i16 : MVT::v16i16;
12559 // uint4 lo = _mm_blend_epi16( v, (uint4) 0x4b000000, 0xaa);
12560 SDValue VecCstLowBitcast = DAG.getBitcast(VecI16VT, VecCstLow);
12561 SDValue VecBitcast = DAG.getBitcast(VecI16VT, V);
12562 // Low will be bitcasted right away, so do not bother bitcasting back to its
12564 Low = DAG.getNode(X86ISD::BLENDI, DL, VecI16VT, VecBitcast,
12565 VecCstLowBitcast, DAG.getConstant(0xaa, DL, MVT::i32));
12566 // uint4 hi = _mm_blend_epi16( _mm_srli_epi32(v,16),
12567 // (uint4) 0x53000000, 0xaa);
12568 SDValue VecCstHighBitcast = DAG.getBitcast(VecI16VT, VecCstHigh);
12569 SDValue VecShiftBitcast = DAG.getBitcast(VecI16VT, HighShift);
12570 // High will be bitcasted right away, so do not bother bitcasting back to
12571 // its original type.
12572 High = DAG.getNode(X86ISD::BLENDI, DL, VecI16VT, VecShiftBitcast,
12573 VecCstHighBitcast, DAG.getConstant(0xaa, DL, MVT::i32));
12575 SDValue CstMask = DAG.getConstant(0xffff, DL, MVT::i32);
12576 SDValue VecCstMask = DAG.getNode(ISD::BUILD_VECTOR, DL, VecIntVT, CstMask,
12577 CstMask, CstMask, CstMask);
12578 // uint4 lo = (v & (uint4) 0xffff) | (uint4) 0x4b000000;
12579 SDValue LowAnd = DAG.getNode(ISD::AND, DL, VecIntVT, V, VecCstMask);
12580 Low = DAG.getNode(ISD::OR, DL, VecIntVT, LowAnd, VecCstLow);
12582 // uint4 hi = (v >> 16) | (uint4) 0x53000000;
12583 High = DAG.getNode(ISD::OR, DL, VecIntVT, HighShift, VecCstHigh);
12586 // Create the vector constant for -(0x1.0p39f + 0x1.0p23f).
12587 SDValue CstFAdd = DAG.getConstantFP(
12588 APFloat(APFloat::IEEEsingle, APInt(32, 0xD3000080)), DL, MVT::f32);
12589 SDValue CstFAddArray[] = {CstFAdd, CstFAdd, CstFAdd, CstFAdd,
12590 CstFAdd, CstFAdd, CstFAdd, CstFAdd};
12591 SDValue VecCstFAdd = DAG.getNode(ISD::BUILD_VECTOR, DL, VecFloatVT,
12592 makeArrayRef(&CstFAddArray[0], NumElts));
12594 // float4 fhi = (float4) hi - (0x1.0p39f + 0x1.0p23f);
12595 SDValue HighBitcast = DAG.getBitcast(VecFloatVT, High);
12596 // TODO: Are there any fast-math-flags to propagate here?
12598 DAG.getNode(ISD::FADD, DL, VecFloatVT, HighBitcast, VecCstFAdd);
12599 // return (float4) lo + fhi;
12600 SDValue LowBitcast = DAG.getBitcast(VecFloatVT, Low);
12601 return DAG.getNode(ISD::FADD, DL, VecFloatVT, LowBitcast, FHigh);
12604 SDValue X86TargetLowering::lowerUINT_TO_FP_vec(SDValue Op,
12605 SelectionDAG &DAG) const {
12606 SDValue N0 = Op.getOperand(0);
12607 MVT SVT = N0.getSimpleValueType();
12610 switch (SVT.SimpleTy) {
12612 llvm_unreachable("Custom UINT_TO_FP is not supported!");
12617 MVT NVT = MVT::getVectorVT(MVT::i32, SVT.getVectorNumElements());
12618 return DAG.getNode(ISD::SINT_TO_FP, dl, Op.getValueType(),
12619 DAG.getNode(ISD::ZERO_EXTEND, dl, NVT, N0));
12623 return lowerUINT_TO_FP_vXi32(Op, DAG, *Subtarget);
12626 if (Subtarget->hasAVX512())
12627 return DAG.getNode(ISD::UINT_TO_FP, dl, Op.getValueType(),
12628 DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v16i32, N0));
12630 llvm_unreachable(nullptr);
12633 SDValue X86TargetLowering::LowerUINT_TO_FP(SDValue Op,
12634 SelectionDAG &DAG) const {
12635 SDValue N0 = Op.getOperand(0);
12637 auto PtrVT = getPointerTy(DAG.getDataLayout());
12639 if (Op.getValueType().isVector())
12640 return lowerUINT_TO_FP_vec(Op, DAG);
12642 // Since UINT_TO_FP is legal (it's marked custom), dag combiner won't
12643 // optimize it to a SINT_TO_FP when the sign bit is known zero. Perform
12644 // the optimization here.
12645 if (DAG.SignBitIsZero(N0))
12646 return DAG.getNode(ISD::SINT_TO_FP, dl, Op.getValueType(), N0);
12648 MVT SrcVT = N0.getSimpleValueType();
12649 MVT DstVT = Op.getSimpleValueType();
12651 if (Subtarget->hasAVX512() && isScalarFPTypeInSSEReg(DstVT) &&
12652 (SrcVT == MVT::i32 || (SrcVT == MVT::i64 && Subtarget->is64Bit()))) {
12653 // Conversions from unsigned i32 to f32/f64 are legal,
12654 // using VCVTUSI2SS/SD. Same for i64 in 64-bit mode.
12658 if (SrcVT == MVT::i64 && DstVT == MVT::f64 && X86ScalarSSEf64)
12659 return LowerUINT_TO_FP_i64(Op, DAG);
12660 if (SrcVT == MVT::i32 && X86ScalarSSEf64)
12661 return LowerUINT_TO_FP_i32(Op, DAG);
12662 if (Subtarget->is64Bit() && SrcVT == MVT::i64 && DstVT == MVT::f32)
12665 // Make a 64-bit buffer, and use it to build an FILD.
12666 SDValue StackSlot = DAG.CreateStackTemporary(MVT::i64);
12667 if (SrcVT == MVT::i32) {
12668 SDValue WordOff = DAG.getConstant(4, dl, PtrVT);
12669 SDValue OffsetSlot = DAG.getNode(ISD::ADD, dl, PtrVT, StackSlot, WordOff);
12670 SDValue Store1 = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
12671 StackSlot, MachinePointerInfo(),
12673 SDValue Store2 = DAG.getStore(Store1, dl, DAG.getConstant(0, dl, MVT::i32),
12674 OffsetSlot, MachinePointerInfo(),
12676 SDValue Fild = BuildFILD(Op, MVT::i64, Store2, StackSlot, DAG);
12680 assert(SrcVT == MVT::i64 && "Unexpected type in UINT_TO_FP");
12681 SDValue Store = DAG.getStore(DAG.getEntryNode(), dl, Op.getOperand(0),
12682 StackSlot, MachinePointerInfo(),
12684 // For i64 source, we need to add the appropriate power of 2 if the input
12685 // was negative. This is the same as the optimization in
12686 // DAGTypeLegalizer::ExpandIntOp_UNIT_TO_FP, and for it to be safe here,
12687 // we must be careful to do the computation in x87 extended precision, not
12688 // in SSE. (The generic code can't know it's OK to do this, or how to.)
12689 int SSFI = cast<FrameIndexSDNode>(StackSlot)->getIndex();
12690 MachineMemOperand *MMO = DAG.getMachineFunction().getMachineMemOperand(
12691 MachinePointerInfo::getFixedStack(DAG.getMachineFunction(), SSFI),
12692 MachineMemOperand::MOLoad, 8, 8);
12694 SDVTList Tys = DAG.getVTList(MVT::f80, MVT::Other);
12695 SDValue Ops[] = { Store, StackSlot, DAG.getValueType(MVT::i64) };
12696 SDValue Fild = DAG.getMemIntrinsicNode(X86ISD::FILD, dl, Tys, Ops,
12699 APInt FF(32, 0x5F800000ULL);
12701 // Check whether the sign bit is set.
12702 SDValue SignSet = DAG.getSetCC(
12703 dl, getSetCCResultType(DAG.getDataLayout(), *DAG.getContext(), MVT::i64),
12704 Op.getOperand(0), DAG.getConstant(0, dl, MVT::i64), ISD::SETLT);
12706 // Build a 64 bit pair (0, FF) in the constant pool, with FF in the lo bits.
12707 SDValue FudgePtr = DAG.getConstantPool(
12708 ConstantInt::get(*DAG.getContext(), FF.zext(64)), PtrVT);
12710 // Get a pointer to FF if the sign bit was set, or to 0 otherwise.
12711 SDValue Zero = DAG.getIntPtrConstant(0, dl);
12712 SDValue Four = DAG.getIntPtrConstant(4, dl);
12713 SDValue Offset = DAG.getNode(ISD::SELECT, dl, Zero.getValueType(), SignSet,
12715 FudgePtr = DAG.getNode(ISD::ADD, dl, PtrVT, FudgePtr, Offset);
12717 // Load the value out, extending it from f32 to f80.
12718 // FIXME: Avoid the extend by constructing the right constant pool?
12719 SDValue Fudge = DAG.getExtLoad(
12720 ISD::EXTLOAD, dl, MVT::f80, DAG.getEntryNode(), FudgePtr,
12721 MachinePointerInfo::getConstantPool(DAG.getMachineFunction()), MVT::f32,
12722 false, false, false, 4);
12723 // Extend everything to 80 bits to force it to be done on x87.
12724 // TODO: Are there any fast-math-flags to propagate here?
12725 SDValue Add = DAG.getNode(ISD::FADD, dl, MVT::f80, Fild, Fudge);
12726 return DAG.getNode(ISD::FP_ROUND, dl, DstVT, Add,
12727 DAG.getIntPtrConstant(0, dl));
12730 // If the given FP_TO_SINT (IsSigned) or FP_TO_UINT (!IsSigned) operation
12731 // is legal, or has an fp128 or f16 source (which needs to be promoted to f32),
12732 // just return an <SDValue(), SDValue()> pair.
12733 // Otherwise it is assumed to be a conversion from one of f32, f64 or f80
12734 // to i16, i32 or i64, and we lower it to a legal sequence.
12735 // If lowered to the final integer result we return a <result, SDValue()> pair.
12736 // Otherwise we lower it to a sequence ending with a FIST, return a
12737 // <FIST, StackSlot> pair, and the caller is responsible for loading
12738 // the final integer result from StackSlot.
12739 std::pair<SDValue,SDValue>
12740 X86TargetLowering::FP_TO_INTHelper(SDValue Op, SelectionDAG &DAG,
12741 bool IsSigned, bool IsReplace) const {
12744 EVT DstTy = Op.getValueType();
12745 EVT TheVT = Op.getOperand(0).getValueType();
12746 auto PtrVT = getPointerTy(DAG.getDataLayout());
12748 if (TheVT != MVT::f32 && TheVT != MVT::f64 && TheVT != MVT::f80) {
12749 // f16 must be promoted before using the lowering in this routine.
12750 // fp128 does not use this lowering.
12751 return std::make_pair(SDValue(), SDValue());
12754 // If using FIST to compute an unsigned i64, we'll need some fixup
12755 // to handle values above the maximum signed i64. A FIST is always
12756 // used for the 32-bit subtarget, but also for f80 on a 64-bit target.
12757 bool UnsignedFixup = !IsSigned &&
12758 DstTy == MVT::i64 &&
12759 (!Subtarget->is64Bit() ||
12760 !isScalarFPTypeInSSEReg(TheVT));
12762 if (!IsSigned && DstTy != MVT::i64 && !Subtarget->hasAVX512()) {
12763 // Replace the fp-to-uint32 operation with an fp-to-sint64 FIST.
12764 // The low 32 bits of the fist result will have the correct uint32 result.
12765 assert(DstTy == MVT::i32 && "Unexpected FP_TO_UINT");
12769 assert(DstTy.getSimpleVT() <= MVT::i64 &&
12770 DstTy.getSimpleVT() >= MVT::i16 &&
12771 "Unknown FP_TO_INT to lower!");
12773 // These are really Legal.
12774 if (DstTy == MVT::i32 &&
12775 isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType()))
12776 return std::make_pair(SDValue(), SDValue());
12777 if (Subtarget->is64Bit() &&
12778 DstTy == MVT::i64 &&
12779 isScalarFPTypeInSSEReg(Op.getOperand(0).getValueType()))
12780 return std::make_pair(SDValue(), SDValue());
12782 // We lower FP->int64 into FISTP64 followed by a load from a temporary
12784 MachineFunction &MF = DAG.getMachineFunction();
12785 unsigned MemSize = DstTy.getSizeInBits()/8;
12786 int SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize, false);
12787 SDValue StackSlot = DAG.getFrameIndex(SSFI, PtrVT);
12790 switch (DstTy.getSimpleVT().SimpleTy) {
12791 default: llvm_unreachable("Invalid FP_TO_SINT to lower!");
12792 case MVT::i16: Opc = X86ISD::FP_TO_INT16_IN_MEM; break;
12793 case MVT::i32: Opc = X86ISD::FP_TO_INT32_IN_MEM; break;
12794 case MVT::i64: Opc = X86ISD::FP_TO_INT64_IN_MEM; break;
12797 SDValue Chain = DAG.getEntryNode();
12798 SDValue Value = Op.getOperand(0);
12799 SDValue Adjust; // 0x0 or 0x80000000, for result sign bit adjustment.
12801 if (UnsignedFixup) {
12803 // Conversion to unsigned i64 is implemented with a select,
12804 // depending on whether the source value fits in the range
12805 // of a signed i64. Let Thresh be the FP equivalent of
12806 // 0x8000000000000000ULL.
12808 // Adjust i32 = (Value < Thresh) ? 0 : 0x80000000;
12809 // FistSrc = (Value < Thresh) ? Value : (Value - Thresh);
12810 // Fist-to-mem64 FistSrc
12811 // Add 0 or 0x800...0ULL to the 64-bit result, which is equivalent
12812 // to XOR'ing the high 32 bits with Adjust.
12814 // Being a power of 2, Thresh is exactly representable in all FP formats.
12815 // For X87 we'd like to use the smallest FP type for this constant, but
12816 // for DAG type consistency we have to match the FP operand type.
12818 APFloat Thresh(APFloat::IEEEsingle, APInt(32, 0x5f000000));
12819 LLVM_ATTRIBUTE_UNUSED APFloat::opStatus Status = APFloat::opOK;
12820 bool LosesInfo = false;
12821 if (TheVT == MVT::f64)
12822 // The rounding mode is irrelevant as the conversion should be exact.
12823 Status = Thresh.convert(APFloat::IEEEdouble, APFloat::rmNearestTiesToEven,
12825 else if (TheVT == MVT::f80)
12826 Status = Thresh.convert(APFloat::x87DoubleExtended,
12827 APFloat::rmNearestTiesToEven, &LosesInfo);
12829 assert(Status == APFloat::opOK && !LosesInfo &&
12830 "FP conversion should have been exact");
12832 SDValue ThreshVal = DAG.getConstantFP(Thresh, DL, TheVT);
12834 SDValue Cmp = DAG.getSetCC(DL,
12835 getSetCCResultType(DAG.getDataLayout(),
12836 *DAG.getContext(), TheVT),
12837 Value, ThreshVal, ISD::SETLT);
12838 Adjust = DAG.getSelect(DL, MVT::i32, Cmp,
12839 DAG.getConstant(0, DL, MVT::i32),
12840 DAG.getConstant(0x80000000, DL, MVT::i32));
12841 SDValue Sub = DAG.getNode(ISD::FSUB, DL, TheVT, Value, ThreshVal);
12842 Cmp = DAG.getSetCC(DL, getSetCCResultType(DAG.getDataLayout(),
12843 *DAG.getContext(), TheVT),
12844 Value, ThreshVal, ISD::SETLT);
12845 Value = DAG.getSelect(DL, TheVT, Cmp, Value, Sub);
12848 // FIXME This causes a redundant load/store if the SSE-class value is already
12849 // in memory, such as if it is on the callstack.
12850 if (isScalarFPTypeInSSEReg(TheVT)) {
12851 assert(DstTy == MVT::i64 && "Invalid FP_TO_SINT to lower!");
12852 Chain = DAG.getStore(Chain, DL, Value, StackSlot,
12853 MachinePointerInfo::getFixedStack(MF, SSFI), false,
12855 SDVTList Tys = DAG.getVTList(Op.getOperand(0).getValueType(), MVT::Other);
12857 Chain, StackSlot, DAG.getValueType(TheVT)
12860 MachineMemOperand *MMO =
12861 MF.getMachineMemOperand(MachinePointerInfo::getFixedStack(MF, SSFI),
12862 MachineMemOperand::MOLoad, MemSize, MemSize);
12863 Value = DAG.getMemIntrinsicNode(X86ISD::FLD, DL, Tys, Ops, DstTy, MMO);
12864 Chain = Value.getValue(1);
12865 SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize, false);
12866 StackSlot = DAG.getFrameIndex(SSFI, PtrVT);
12869 MachineMemOperand *MMO =
12870 MF.getMachineMemOperand(MachinePointerInfo::getFixedStack(MF, SSFI),
12871 MachineMemOperand::MOStore, MemSize, MemSize);
12873 if (UnsignedFixup) {
12875 // Insert the FIST, load its result as two i32's,
12876 // and XOR the high i32 with Adjust.
12878 SDValue FistOps[] = { Chain, Value, StackSlot };
12879 SDValue FIST = DAG.getMemIntrinsicNode(Opc, DL, DAG.getVTList(MVT::Other),
12880 FistOps, DstTy, MMO);
12882 SDValue Low32 = DAG.getLoad(MVT::i32, DL, FIST, StackSlot,
12883 MachinePointerInfo(),
12884 false, false, false, 0);
12885 SDValue HighAddr = DAG.getNode(ISD::ADD, DL, PtrVT, StackSlot,
12886 DAG.getConstant(4, DL, PtrVT));
12888 SDValue High32 = DAG.getLoad(MVT::i32, DL, FIST, HighAddr,
12889 MachinePointerInfo(),
12890 false, false, false, 0);
12891 High32 = DAG.getNode(ISD::XOR, DL, MVT::i32, High32, Adjust);
12893 if (Subtarget->is64Bit()) {
12894 // Join High32 and Low32 into a 64-bit result.
12895 // (High32 << 32) | Low32
12896 Low32 = DAG.getNode(ISD::ZERO_EXTEND, DL, MVT::i64, Low32);
12897 High32 = DAG.getNode(ISD::ANY_EXTEND, DL, MVT::i64, High32);
12898 High32 = DAG.getNode(ISD::SHL, DL, MVT::i64, High32,
12899 DAG.getConstant(32, DL, MVT::i8));
12900 SDValue Result = DAG.getNode(ISD::OR, DL, MVT::i64, High32, Low32);
12901 return std::make_pair(Result, SDValue());
12904 SDValue ResultOps[] = { Low32, High32 };
12906 SDValue pair = IsReplace
12907 ? DAG.getNode(ISD::BUILD_PAIR, DL, MVT::i64, ResultOps)
12908 : DAG.getMergeValues(ResultOps, DL);
12909 return std::make_pair(pair, SDValue());
12911 // Build the FP_TO_INT*_IN_MEM
12912 SDValue Ops[] = { Chain, Value, StackSlot };
12913 SDValue FIST = DAG.getMemIntrinsicNode(Opc, DL, DAG.getVTList(MVT::Other),
12915 return std::make_pair(FIST, StackSlot);
12919 static SDValue LowerAVXExtend(SDValue Op, SelectionDAG &DAG,
12920 const X86Subtarget *Subtarget) {
12921 MVT VT = Op->getSimpleValueType(0);
12922 SDValue In = Op->getOperand(0);
12923 MVT InVT = In.getSimpleValueType();
12926 if (VT.is512BitVector() || InVT.getScalarType() == MVT::i1)
12927 return DAG.getNode(ISD::ZERO_EXTEND, dl, VT, In);
12929 // Optimize vectors in AVX mode:
12932 // Use vpunpcklwd for 4 lower elements v8i16 -> v4i32.
12933 // Use vpunpckhwd for 4 upper elements v8i16 -> v4i32.
12934 // Concat upper and lower parts.
12937 // Use vpunpckldq for 4 lower elements v4i32 -> v2i64.
12938 // Use vpunpckhdq for 4 upper elements v4i32 -> v2i64.
12939 // Concat upper and lower parts.
12942 if (((VT != MVT::v16i16) || (InVT != MVT::v16i8)) &&
12943 ((VT != MVT::v8i32) || (InVT != MVT::v8i16)) &&
12944 ((VT != MVT::v4i64) || (InVT != MVT::v4i32)))
12947 if (Subtarget->hasInt256())
12948 return DAG.getNode(X86ISD::VZEXT, dl, VT, In);
12950 SDValue ZeroVec = getZeroVector(InVT, Subtarget, DAG, dl);
12951 SDValue Undef = DAG.getUNDEF(InVT);
12952 bool NeedZero = Op.getOpcode() == ISD::ZERO_EXTEND;
12953 SDValue OpLo = getUnpackl(DAG, dl, InVT, In, NeedZero ? ZeroVec : Undef);
12954 SDValue OpHi = getUnpackh(DAG, dl, InVT, In, NeedZero ? ZeroVec : Undef);
12956 MVT HVT = MVT::getVectorVT(VT.getVectorElementType(),
12957 VT.getVectorNumElements()/2);
12959 OpLo = DAG.getBitcast(HVT, OpLo);
12960 OpHi = DAG.getBitcast(HVT, OpHi);
12962 return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT, OpLo, OpHi);
12965 static SDValue LowerZERO_EXTEND_AVX512(SDValue Op,
12966 const X86Subtarget *Subtarget, SelectionDAG &DAG) {
12967 MVT VT = Op->getSimpleValueType(0);
12968 SDValue In = Op->getOperand(0);
12969 MVT InVT = In.getSimpleValueType();
12971 unsigned int NumElts = VT.getVectorNumElements();
12972 if (NumElts != 8 && NumElts != 16 && !Subtarget->hasBWI())
12975 if (VT.is512BitVector() && InVT.getVectorElementType() != MVT::i1)
12976 return DAG.getNode(X86ISD::VZEXT, DL, VT, In);
12978 assert(InVT.getVectorElementType() == MVT::i1);
12979 MVT ExtVT = NumElts == 8 ? MVT::v8i64 : MVT::v16i32;
12981 DAG.getConstant(APInt(ExtVT.getScalarSizeInBits(), 1), DL, ExtVT);
12983 DAG.getConstant(APInt::getNullValue(ExtVT.getScalarSizeInBits()), DL, ExtVT);
12985 SDValue V = DAG.getNode(ISD::VSELECT, DL, ExtVT, In, One, Zero);
12986 if (VT.is512BitVector())
12988 return DAG.getNode(X86ISD::VTRUNC, DL, VT, V);
12991 static SDValue LowerANY_EXTEND(SDValue Op, const X86Subtarget *Subtarget,
12992 SelectionDAG &DAG) {
12993 if (Subtarget->hasFp256())
12994 if (SDValue Res = LowerAVXExtend(Op, DAG, Subtarget))
13000 static SDValue LowerZERO_EXTEND(SDValue Op, const X86Subtarget *Subtarget,
13001 SelectionDAG &DAG) {
13003 MVT VT = Op.getSimpleValueType();
13004 SDValue In = Op.getOperand(0);
13005 MVT SVT = In.getSimpleValueType();
13007 if (VT.is512BitVector() || SVT.getVectorElementType() == MVT::i1)
13008 return LowerZERO_EXTEND_AVX512(Op, Subtarget, DAG);
13010 if (Subtarget->hasFp256())
13011 if (SDValue Res = LowerAVXExtend(Op, DAG, Subtarget))
13014 assert(!VT.is256BitVector() || !SVT.is128BitVector() ||
13015 VT.getVectorNumElements() != SVT.getVectorNumElements());
13019 SDValue X86TargetLowering::LowerTRUNCATE(SDValue Op, SelectionDAG &DAG) const {
13021 MVT VT = Op.getSimpleValueType();
13022 SDValue In = Op.getOperand(0);
13023 MVT InVT = In.getSimpleValueType();
13025 if (VT == MVT::i1) {
13026 assert((InVT.isInteger() && (InVT.getSizeInBits() <= 64)) &&
13027 "Invalid scalar TRUNCATE operation");
13028 if (InVT.getSizeInBits() >= 32)
13030 In = DAG.getNode(ISD::ANY_EXTEND, DL, MVT::i32, In);
13031 return DAG.getNode(ISD::TRUNCATE, DL, VT, In);
13033 assert(VT.getVectorNumElements() == InVT.getVectorNumElements() &&
13034 "Invalid TRUNCATE operation");
13036 // move vector to mask - truncate solution for SKX
13037 if (VT.getVectorElementType() == MVT::i1) {
13038 if (InVT.is512BitVector() && InVT.getScalarSizeInBits() <= 16 &&
13039 Subtarget->hasBWI())
13040 return Op; // legal, will go to VPMOVB2M, VPMOVW2M
13041 if ((InVT.is256BitVector() || InVT.is128BitVector())
13042 && InVT.getScalarSizeInBits() <= 16 &&
13043 Subtarget->hasBWI() && Subtarget->hasVLX())
13044 return Op; // legal, will go to VPMOVB2M, VPMOVW2M
13045 if (InVT.is512BitVector() && InVT.getScalarSizeInBits() >= 32 &&
13046 Subtarget->hasDQI())
13047 return Op; // legal, will go to VPMOVD2M, VPMOVQ2M
13048 if ((InVT.is256BitVector() || InVT.is128BitVector())
13049 && InVT.getScalarSizeInBits() >= 32 &&
13050 Subtarget->hasDQI() && Subtarget->hasVLX())
13051 return Op; // legal, will go to VPMOVB2M, VPMOVQ2M
13054 if (VT.getVectorElementType() == MVT::i1) {
13055 assert(VT.getVectorElementType() == MVT::i1 && "Unexpected vector type");
13056 unsigned NumElts = InVT.getVectorNumElements();
13057 assert ((NumElts == 8 || NumElts == 16) && "Unexpected vector type");
13058 if (InVT.getSizeInBits() < 512) {
13059 MVT ExtVT = (NumElts == 16)? MVT::v16i32 : MVT::v8i64;
13060 In = DAG.getNode(ISD::SIGN_EXTEND, DL, ExtVT, In);
13065 DAG.getConstant(APInt::getSignBit(InVT.getScalarSizeInBits()), DL, InVT);
13066 SDValue And = DAG.getNode(ISD::AND, DL, InVT, OneV, In);
13067 return DAG.getNode(X86ISD::TESTM, DL, VT, And, And);
13070 // vpmovqb/w/d, vpmovdb/w, vpmovwb
13071 if (((!InVT.is512BitVector() && Subtarget->hasVLX()) || InVT.is512BitVector()) &&
13072 (InVT.getVectorElementType() != MVT::i16 || Subtarget->hasBWI()))
13073 return DAG.getNode(X86ISD::VTRUNC, DL, VT, In);
13075 if ((VT == MVT::v4i32) && (InVT == MVT::v4i64)) {
13076 // On AVX2, v4i64 -> v4i32 becomes VPERMD.
13077 if (Subtarget->hasInt256()) {
13078 static const int ShufMask[] = {0, 2, 4, 6, -1, -1, -1, -1};
13079 In = DAG.getBitcast(MVT::v8i32, In);
13080 In = DAG.getVectorShuffle(MVT::v8i32, DL, In, DAG.getUNDEF(MVT::v8i32),
13082 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, VT, In,
13083 DAG.getIntPtrConstant(0, DL));
13086 SDValue OpLo = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v2i64, In,
13087 DAG.getIntPtrConstant(0, DL));
13088 SDValue OpHi = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v2i64, In,
13089 DAG.getIntPtrConstant(2, DL));
13090 OpLo = DAG.getBitcast(MVT::v4i32, OpLo);
13091 OpHi = DAG.getBitcast(MVT::v4i32, OpHi);
13092 static const int ShufMask[] = {0, 2, 4, 6};
13093 return DAG.getVectorShuffle(VT, DL, OpLo, OpHi, ShufMask);
13096 if ((VT == MVT::v8i16) && (InVT == MVT::v8i32)) {
13097 // On AVX2, v8i32 -> v8i16 becomed PSHUFB.
13098 if (Subtarget->hasInt256()) {
13099 In = DAG.getBitcast(MVT::v32i8, In);
13101 SmallVector<SDValue,32> pshufbMask;
13102 for (unsigned i = 0; i < 2; ++i) {
13103 pshufbMask.push_back(DAG.getConstant(0x0, DL, MVT::i8));
13104 pshufbMask.push_back(DAG.getConstant(0x1, DL, MVT::i8));
13105 pshufbMask.push_back(DAG.getConstant(0x4, DL, MVT::i8));
13106 pshufbMask.push_back(DAG.getConstant(0x5, DL, MVT::i8));
13107 pshufbMask.push_back(DAG.getConstant(0x8, DL, MVT::i8));
13108 pshufbMask.push_back(DAG.getConstant(0x9, DL, MVT::i8));
13109 pshufbMask.push_back(DAG.getConstant(0xc, DL, MVT::i8));
13110 pshufbMask.push_back(DAG.getConstant(0xd, DL, MVT::i8));
13111 for (unsigned j = 0; j < 8; ++j)
13112 pshufbMask.push_back(DAG.getConstant(0x80, DL, MVT::i8));
13114 SDValue BV = DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v32i8, pshufbMask);
13115 In = DAG.getNode(X86ISD::PSHUFB, DL, MVT::v32i8, In, BV);
13116 In = DAG.getBitcast(MVT::v4i64, In);
13118 static const int ShufMask[] = {0, 2, -1, -1};
13119 In = DAG.getVectorShuffle(MVT::v4i64, DL, In, DAG.getUNDEF(MVT::v4i64),
13121 In = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v2i64, In,
13122 DAG.getIntPtrConstant(0, DL));
13123 return DAG.getBitcast(VT, In);
13126 SDValue OpLo = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v4i32, In,
13127 DAG.getIntPtrConstant(0, DL));
13129 SDValue OpHi = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v4i32, In,
13130 DAG.getIntPtrConstant(4, DL));
13132 OpLo = DAG.getBitcast(MVT::v16i8, OpLo);
13133 OpHi = DAG.getBitcast(MVT::v16i8, OpHi);
13135 // The PSHUFB mask:
13136 static const int ShufMask1[] = {0, 1, 4, 5, 8, 9, 12, 13,
13137 -1, -1, -1, -1, -1, -1, -1, -1};
13139 SDValue Undef = DAG.getUNDEF(MVT::v16i8);
13140 OpLo = DAG.getVectorShuffle(MVT::v16i8, DL, OpLo, Undef, ShufMask1);
13141 OpHi = DAG.getVectorShuffle(MVT::v16i8, DL, OpHi, Undef, ShufMask1);
13143 OpLo = DAG.getBitcast(MVT::v4i32, OpLo);
13144 OpHi = DAG.getBitcast(MVT::v4i32, OpHi);
13146 // The MOVLHPS Mask:
13147 static const int ShufMask2[] = {0, 1, 4, 5};
13148 SDValue res = DAG.getVectorShuffle(MVT::v4i32, DL, OpLo, OpHi, ShufMask2);
13149 return DAG.getBitcast(MVT::v8i16, res);
13152 // Handle truncation of V256 to V128 using shuffles.
13153 if (!VT.is128BitVector() || !InVT.is256BitVector())
13156 assert(Subtarget->hasFp256() && "256-bit vector without AVX!");
13158 unsigned NumElems = VT.getVectorNumElements();
13159 MVT NVT = MVT::getVectorVT(VT.getVectorElementType(), NumElems * 2);
13161 SmallVector<int, 16> MaskVec(NumElems * 2, -1);
13162 // Prepare truncation shuffle mask
13163 for (unsigned i = 0; i != NumElems; ++i)
13164 MaskVec[i] = i * 2;
13165 SDValue V = DAG.getVectorShuffle(NVT, DL, DAG.getBitcast(NVT, In),
13166 DAG.getUNDEF(NVT), &MaskVec[0]);
13167 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, VT, V,
13168 DAG.getIntPtrConstant(0, DL));
13171 SDValue X86TargetLowering::LowerFP_TO_SINT(SDValue Op,
13172 SelectionDAG &DAG) const {
13173 assert(!Op.getSimpleValueType().isVector());
13175 std::pair<SDValue,SDValue> Vals = FP_TO_INTHelper(Op, DAG,
13176 /*IsSigned=*/ true, /*IsReplace=*/ false);
13177 SDValue FIST = Vals.first, StackSlot = Vals.second;
13178 // If FP_TO_INTHelper failed, the node is actually supposed to be Legal.
13179 if (!FIST.getNode())
13182 if (StackSlot.getNode())
13183 // Load the result.
13184 return DAG.getLoad(Op.getValueType(), SDLoc(Op),
13185 FIST, StackSlot, MachinePointerInfo(),
13186 false, false, false, 0);
13188 // The node is the result.
13192 SDValue X86TargetLowering::LowerFP_TO_UINT(SDValue Op,
13193 SelectionDAG &DAG) const {
13194 std::pair<SDValue,SDValue> Vals = FP_TO_INTHelper(Op, DAG,
13195 /*IsSigned=*/ false, /*IsReplace=*/ false);
13196 SDValue FIST = Vals.first, StackSlot = Vals.second;
13197 // If FP_TO_INTHelper failed, the node is actually supposed to be Legal.
13198 if (!FIST.getNode())
13201 if (StackSlot.getNode())
13202 // Load the result.
13203 return DAG.getLoad(Op.getValueType(), SDLoc(Op),
13204 FIST, StackSlot, MachinePointerInfo(),
13205 false, false, false, 0);
13207 // The node is the result.
13211 static SDValue LowerFP_EXTEND(SDValue Op, SelectionDAG &DAG) {
13213 MVT VT = Op.getSimpleValueType();
13214 SDValue In = Op.getOperand(0);
13215 MVT SVT = In.getSimpleValueType();
13217 assert(SVT == MVT::v2f32 && "Only customize MVT::v2f32 type legalization!");
13219 return DAG.getNode(X86ISD::VFPEXT, DL, VT,
13220 DAG.getNode(ISD::CONCAT_VECTORS, DL, MVT::v4f32,
13221 In, DAG.getUNDEF(SVT)));
13224 /// The only differences between FABS and FNEG are the mask and the logic op.
13225 /// FNEG also has a folding opportunity for FNEG(FABS(x)).
13226 static SDValue LowerFABSorFNEG(SDValue Op, SelectionDAG &DAG) {
13227 assert((Op.getOpcode() == ISD::FABS || Op.getOpcode() == ISD::FNEG) &&
13228 "Wrong opcode for lowering FABS or FNEG.");
13230 bool IsFABS = (Op.getOpcode() == ISD::FABS);
13232 // If this is a FABS and it has an FNEG user, bail out to fold the combination
13233 // into an FNABS. We'll lower the FABS after that if it is still in use.
13235 for (SDNode *User : Op->uses())
13236 if (User->getOpcode() == ISD::FNEG)
13240 MVT VT = Op.getSimpleValueType();
13242 // FIXME: Use function attribute "OptimizeForSize" and/or CodeGenOpt::Level to
13243 // decide if we should generate a 16-byte constant mask when we only need 4 or
13244 // 8 bytes for the scalar case.
13250 if (VT.isVector()) {
13252 EltVT = VT.getVectorElementType();
13253 NumElts = VT.getVectorNumElements();
13255 // There are no scalar bitwise logical SSE/AVX instructions, so we
13256 // generate a 16-byte vector constant and logic op even for the scalar case.
13257 // Using a 16-byte mask allows folding the load of the mask with
13258 // the logic op, so it can save (~4 bytes) on code size.
13259 LogicVT = (VT == MVT::f64) ? MVT::v2f64 : MVT::v4f32;
13261 NumElts = (VT == MVT::f64) ? 2 : 4;
13264 unsigned EltBits = EltVT.getSizeInBits();
13265 LLVMContext *Context = DAG.getContext();
13266 // For FABS, mask is 0x7f...; for FNEG, mask is 0x80...
13268 IsFABS ? APInt::getSignedMaxValue(EltBits) : APInt::getSignBit(EltBits);
13269 Constant *C = ConstantInt::get(*Context, MaskElt);
13270 C = ConstantVector::getSplat(NumElts, C);
13271 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
13272 SDValue CPIdx = DAG.getConstantPool(C, TLI.getPointerTy(DAG.getDataLayout()));
13273 unsigned Alignment = cast<ConstantPoolSDNode>(CPIdx)->getAlignment();
13275 DAG.getLoad(LogicVT, dl, DAG.getEntryNode(), CPIdx,
13276 MachinePointerInfo::getConstantPool(DAG.getMachineFunction()),
13277 false, false, false, Alignment);
13279 SDValue Op0 = Op.getOperand(0);
13280 bool IsFNABS = !IsFABS && (Op0.getOpcode() == ISD::FABS);
13282 IsFABS ? X86ISD::FAND : IsFNABS ? X86ISD::FOR : X86ISD::FXOR;
13283 SDValue Operand = IsFNABS ? Op0.getOperand(0) : Op0;
13286 return DAG.getNode(LogicOp, dl, LogicVT, Operand, Mask);
13288 // For the scalar case extend to a 128-bit vector, perform the logic op,
13289 // and extract the scalar result back out.
13290 Operand = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, LogicVT, Operand);
13291 SDValue LogicNode = DAG.getNode(LogicOp, dl, LogicVT, Operand, Mask);
13292 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, VT, LogicNode,
13293 DAG.getIntPtrConstant(0, dl));
13296 static SDValue LowerFCOPYSIGN(SDValue Op, SelectionDAG &DAG) {
13297 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
13298 LLVMContext *Context = DAG.getContext();
13299 SDValue Op0 = Op.getOperand(0);
13300 SDValue Op1 = Op.getOperand(1);
13302 MVT VT = Op.getSimpleValueType();
13303 MVT SrcVT = Op1.getSimpleValueType();
13305 // If second operand is smaller, extend it first.
13306 if (SrcVT.bitsLT(VT)) {
13307 Op1 = DAG.getNode(ISD::FP_EXTEND, dl, VT, Op1);
13310 // And if it is bigger, shrink it first.
13311 if (SrcVT.bitsGT(VT)) {
13312 Op1 = DAG.getNode(ISD::FP_ROUND, dl, VT, Op1, DAG.getIntPtrConstant(1, dl));
13316 // At this point the operands and the result should have the same
13317 // type, and that won't be f80 since that is not custom lowered.
13319 const fltSemantics &Sem =
13320 VT == MVT::f64 ? APFloat::IEEEdouble : APFloat::IEEEsingle;
13321 const unsigned SizeInBits = VT.getSizeInBits();
13323 SmallVector<Constant *, 4> CV(
13324 VT == MVT::f64 ? 2 : 4,
13325 ConstantFP::get(*Context, APFloat(Sem, APInt(SizeInBits, 0))));
13327 // First, clear all bits but the sign bit from the second operand (sign).
13328 CV[0] = ConstantFP::get(*Context,
13329 APFloat(Sem, APInt::getHighBitsSet(SizeInBits, 1)));
13330 Constant *C = ConstantVector::get(CV);
13331 auto PtrVT = TLI.getPointerTy(DAG.getDataLayout());
13332 SDValue CPIdx = DAG.getConstantPool(C, PtrVT, 16);
13334 // Perform all logic operations as 16-byte vectors because there are no
13335 // scalar FP logic instructions in SSE. This allows load folding of the
13336 // constants into the logic instructions.
13337 MVT LogicVT = (VT == MVT::f64) ? MVT::v2f64 : MVT::v4f32;
13339 DAG.getLoad(LogicVT, dl, DAG.getEntryNode(), CPIdx,
13340 MachinePointerInfo::getConstantPool(DAG.getMachineFunction()),
13341 false, false, false, 16);
13342 Op1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, LogicVT, Op1);
13343 SDValue SignBit = DAG.getNode(X86ISD::FAND, dl, LogicVT, Op1, Mask1);
13345 // Next, clear the sign bit from the first operand (magnitude).
13346 // If it's a constant, we can clear it here.
13347 if (ConstantFPSDNode *Op0CN = dyn_cast<ConstantFPSDNode>(Op0)) {
13348 APFloat APF = Op0CN->getValueAPF();
13349 // If the magnitude is a positive zero, the sign bit alone is enough.
13350 if (APF.isPosZero())
13351 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, SrcVT, SignBit,
13352 DAG.getIntPtrConstant(0, dl));
13354 CV[0] = ConstantFP::get(*Context, APF);
13356 CV[0] = ConstantFP::get(
13358 APFloat(Sem, APInt::getLowBitsSet(SizeInBits, SizeInBits - 1)));
13360 C = ConstantVector::get(CV);
13361 CPIdx = DAG.getConstantPool(C, PtrVT, 16);
13363 DAG.getLoad(LogicVT, dl, DAG.getEntryNode(), CPIdx,
13364 MachinePointerInfo::getConstantPool(DAG.getMachineFunction()),
13365 false, false, false, 16);
13366 // If the magnitude operand wasn't a constant, we need to AND out the sign.
13367 if (!isa<ConstantFPSDNode>(Op0)) {
13368 Op0 = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, LogicVT, Op0);
13369 Val = DAG.getNode(X86ISD::FAND, dl, LogicVT, Op0, Val);
13371 // OR the magnitude value with the sign bit.
13372 Val = DAG.getNode(X86ISD::FOR, dl, LogicVT, Val, SignBit);
13373 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, SrcVT, Val,
13374 DAG.getIntPtrConstant(0, dl));
13377 static SDValue LowerFGETSIGN(SDValue Op, SelectionDAG &DAG) {
13378 SDValue N0 = Op.getOperand(0);
13380 MVT VT = Op.getSimpleValueType();
13382 // Lower ISD::FGETSIGN to (AND (X86ISD::FGETSIGNx86 ...) 1).
13383 SDValue xFGETSIGN = DAG.getNode(X86ISD::FGETSIGNx86, dl, VT, N0,
13384 DAG.getConstant(1, dl, VT));
13385 return DAG.getNode(ISD::AND, dl, VT, xFGETSIGN, DAG.getConstant(1, dl, VT));
13388 // Check whether an OR'd tree is PTEST-able.
13389 static SDValue LowerVectorAllZeroTest(SDValue Op, const X86Subtarget *Subtarget,
13390 SelectionDAG &DAG) {
13391 assert(Op.getOpcode() == ISD::OR && "Only check OR'd tree.");
13393 if (!Subtarget->hasSSE41())
13396 if (!Op->hasOneUse())
13399 SDNode *N = Op.getNode();
13402 SmallVector<SDValue, 8> Opnds;
13403 DenseMap<SDValue, unsigned> VecInMap;
13404 SmallVector<SDValue, 8> VecIns;
13405 EVT VT = MVT::Other;
13407 // Recognize a special case where a vector is casted into wide integer to
13409 Opnds.push_back(N->getOperand(0));
13410 Opnds.push_back(N->getOperand(1));
13412 for (unsigned Slot = 0, e = Opnds.size(); Slot < e; ++Slot) {
13413 SmallVectorImpl<SDValue>::const_iterator I = Opnds.begin() + Slot;
13414 // BFS traverse all OR'd operands.
13415 if (I->getOpcode() == ISD::OR) {
13416 Opnds.push_back(I->getOperand(0));
13417 Opnds.push_back(I->getOperand(1));
13418 // Re-evaluate the number of nodes to be traversed.
13419 e += 2; // 2 more nodes (LHS and RHS) are pushed.
13423 // Quit if a non-EXTRACT_VECTOR_ELT
13424 if (I->getOpcode() != ISD::EXTRACT_VECTOR_ELT)
13427 // Quit if without a constant index.
13428 SDValue Idx = I->getOperand(1);
13429 if (!isa<ConstantSDNode>(Idx))
13432 SDValue ExtractedFromVec = I->getOperand(0);
13433 DenseMap<SDValue, unsigned>::iterator M = VecInMap.find(ExtractedFromVec);
13434 if (M == VecInMap.end()) {
13435 VT = ExtractedFromVec.getValueType();
13436 // Quit if not 128/256-bit vector.
13437 if (!VT.is128BitVector() && !VT.is256BitVector())
13439 // Quit if not the same type.
13440 if (VecInMap.begin() != VecInMap.end() &&
13441 VT != VecInMap.begin()->first.getValueType())
13443 M = VecInMap.insert(std::make_pair(ExtractedFromVec, 0)).first;
13444 VecIns.push_back(ExtractedFromVec);
13446 M->second |= 1U << cast<ConstantSDNode>(Idx)->getZExtValue();
13449 assert((VT.is128BitVector() || VT.is256BitVector()) &&
13450 "Not extracted from 128-/256-bit vector.");
13452 unsigned FullMask = (1U << VT.getVectorNumElements()) - 1U;
13454 for (DenseMap<SDValue, unsigned>::const_iterator
13455 I = VecInMap.begin(), E = VecInMap.end(); I != E; ++I) {
13456 // Quit if not all elements are used.
13457 if (I->second != FullMask)
13461 EVT TestVT = VT.is128BitVector() ? MVT::v2i64 : MVT::v4i64;
13463 // Cast all vectors into TestVT for PTEST.
13464 for (unsigned i = 0, e = VecIns.size(); i < e; ++i)
13465 VecIns[i] = DAG.getBitcast(TestVT, VecIns[i]);
13467 // If more than one full vectors are evaluated, OR them first before PTEST.
13468 for (unsigned Slot = 0, e = VecIns.size(); e - Slot > 1; Slot += 2, e += 1) {
13469 // Each iteration will OR 2 nodes and append the result until there is only
13470 // 1 node left, i.e. the final OR'd value of all vectors.
13471 SDValue LHS = VecIns[Slot];
13472 SDValue RHS = VecIns[Slot + 1];
13473 VecIns.push_back(DAG.getNode(ISD::OR, DL, TestVT, LHS, RHS));
13476 return DAG.getNode(X86ISD::PTEST, DL, MVT::i32,
13477 VecIns.back(), VecIns.back());
13480 /// \brief return true if \c Op has a use that doesn't just read flags.
13481 static bool hasNonFlagsUse(SDValue Op) {
13482 for (SDNode::use_iterator UI = Op->use_begin(), UE = Op->use_end(); UI != UE;
13484 SDNode *User = *UI;
13485 unsigned UOpNo = UI.getOperandNo();
13486 if (User->getOpcode() == ISD::TRUNCATE && User->hasOneUse()) {
13487 // Look pass truncate.
13488 UOpNo = User->use_begin().getOperandNo();
13489 User = *User->use_begin();
13492 if (User->getOpcode() != ISD::BRCOND && User->getOpcode() != ISD::SETCC &&
13493 !(User->getOpcode() == ISD::SELECT && UOpNo == 0))
13499 /// Emit nodes that will be selected as "test Op0,Op0", or something
13501 SDValue X86TargetLowering::EmitTest(SDValue Op, unsigned X86CC, SDLoc dl,
13502 SelectionDAG &DAG) const {
13503 if (Op.getValueType() == MVT::i1) {
13504 SDValue ExtOp = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i8, Op);
13505 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, ExtOp,
13506 DAG.getConstant(0, dl, MVT::i8));
13508 // CF and OF aren't always set the way we want. Determine which
13509 // of these we need.
13510 bool NeedCF = false;
13511 bool NeedOF = false;
13514 case X86::COND_A: case X86::COND_AE:
13515 case X86::COND_B: case X86::COND_BE:
13518 case X86::COND_G: case X86::COND_GE:
13519 case X86::COND_L: case X86::COND_LE:
13520 case X86::COND_O: case X86::COND_NO: {
13521 // Check if we really need to set the
13522 // Overflow flag. If NoSignedWrap is present
13523 // that is not actually needed.
13524 switch (Op->getOpcode()) {
13529 const auto *BinNode = cast<BinaryWithFlagsSDNode>(Op.getNode());
13530 if (BinNode->Flags.hasNoSignedWrap())
13540 // See if we can use the EFLAGS value from the operand instead of
13541 // doing a separate TEST. TEST always sets OF and CF to 0, so unless
13542 // we prove that the arithmetic won't overflow, we can't use OF or CF.
13543 if (Op.getResNo() != 0 || NeedOF || NeedCF) {
13544 // Emit a CMP with 0, which is the TEST pattern.
13545 //if (Op.getValueType() == MVT::i1)
13546 // return DAG.getNode(X86ISD::CMP, dl, MVT::i1, Op,
13547 // DAG.getConstant(0, MVT::i1));
13548 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op,
13549 DAG.getConstant(0, dl, Op.getValueType()));
13551 unsigned Opcode = 0;
13552 unsigned NumOperands = 0;
13554 // Truncate operations may prevent the merge of the SETCC instruction
13555 // and the arithmetic instruction before it. Attempt to truncate the operands
13556 // of the arithmetic instruction and use a reduced bit-width instruction.
13557 bool NeedTruncation = false;
13558 SDValue ArithOp = Op;
13559 if (Op->getOpcode() == ISD::TRUNCATE && Op->hasOneUse()) {
13560 SDValue Arith = Op->getOperand(0);
13561 // Both the trunc and the arithmetic op need to have one user each.
13562 if (Arith->hasOneUse())
13563 switch (Arith.getOpcode()) {
13570 NeedTruncation = true;
13576 // NOTICE: In the code below we use ArithOp to hold the arithmetic operation
13577 // which may be the result of a CAST. We use the variable 'Op', which is the
13578 // non-casted variable when we check for possible users.
13579 switch (ArithOp.getOpcode()) {
13581 // Due to an isel shortcoming, be conservative if this add is likely to be
13582 // selected as part of a load-modify-store instruction. When the root node
13583 // in a match is a store, isel doesn't know how to remap non-chain non-flag
13584 // uses of other nodes in the match, such as the ADD in this case. This
13585 // leads to the ADD being left around and reselected, with the result being
13586 // two adds in the output. Alas, even if none our users are stores, that
13587 // doesn't prove we're O.K. Ergo, if we have any parents that aren't
13588 // CopyToReg or SETCC, eschew INC/DEC. A better fix seems to require
13589 // climbing the DAG back to the root, and it doesn't seem to be worth the
13591 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
13592 UE = Op.getNode()->use_end(); UI != UE; ++UI)
13593 if (UI->getOpcode() != ISD::CopyToReg &&
13594 UI->getOpcode() != ISD::SETCC &&
13595 UI->getOpcode() != ISD::STORE)
13598 if (ConstantSDNode *C =
13599 dyn_cast<ConstantSDNode>(ArithOp.getNode()->getOperand(1))) {
13600 // An add of one will be selected as an INC.
13601 if (C->getAPIntValue() == 1 && !Subtarget->slowIncDec()) {
13602 Opcode = X86ISD::INC;
13607 // An add of negative one (subtract of one) will be selected as a DEC.
13608 if (C->getAPIntValue().isAllOnesValue() && !Subtarget->slowIncDec()) {
13609 Opcode = X86ISD::DEC;
13615 // Otherwise use a regular EFLAGS-setting add.
13616 Opcode = X86ISD::ADD;
13621 // If we have a constant logical shift that's only used in a comparison
13622 // against zero turn it into an equivalent AND. This allows turning it into
13623 // a TEST instruction later.
13624 if ((X86CC == X86::COND_E || X86CC == X86::COND_NE) && Op->hasOneUse() &&
13625 isa<ConstantSDNode>(Op->getOperand(1)) && !hasNonFlagsUse(Op)) {
13626 EVT VT = Op.getValueType();
13627 unsigned BitWidth = VT.getSizeInBits();
13628 unsigned ShAmt = Op->getConstantOperandVal(1);
13629 if (ShAmt >= BitWidth) // Avoid undefined shifts.
13631 APInt Mask = ArithOp.getOpcode() == ISD::SRL
13632 ? APInt::getHighBitsSet(BitWidth, BitWidth - ShAmt)
13633 : APInt::getLowBitsSet(BitWidth, BitWidth - ShAmt);
13634 if (!Mask.isSignedIntN(32)) // Avoid large immediates.
13636 SDValue New = DAG.getNode(ISD::AND, dl, VT, Op->getOperand(0),
13637 DAG.getConstant(Mask, dl, VT));
13638 DAG.ReplaceAllUsesWith(Op, New);
13644 // If the primary and result isn't used, don't bother using X86ISD::AND,
13645 // because a TEST instruction will be better.
13646 if (!hasNonFlagsUse(Op))
13652 // Due to the ISEL shortcoming noted above, be conservative if this op is
13653 // likely to be selected as part of a load-modify-store instruction.
13654 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
13655 UE = Op.getNode()->use_end(); UI != UE; ++UI)
13656 if (UI->getOpcode() == ISD::STORE)
13659 // Otherwise use a regular EFLAGS-setting instruction.
13660 switch (ArithOp.getOpcode()) {
13661 default: llvm_unreachable("unexpected operator!");
13662 case ISD::SUB: Opcode = X86ISD::SUB; break;
13663 case ISD::XOR: Opcode = X86ISD::XOR; break;
13664 case ISD::AND: Opcode = X86ISD::AND; break;
13666 if (!NeedTruncation && (X86CC == X86::COND_E || X86CC == X86::COND_NE)) {
13667 SDValue EFLAGS = LowerVectorAllZeroTest(Op, Subtarget, DAG);
13668 if (EFLAGS.getNode())
13671 Opcode = X86ISD::OR;
13685 return SDValue(Op.getNode(), 1);
13691 // If we found that truncation is beneficial, perform the truncation and
13693 if (NeedTruncation) {
13694 EVT VT = Op.getValueType();
13695 SDValue WideVal = Op->getOperand(0);
13696 EVT WideVT = WideVal.getValueType();
13697 unsigned ConvertedOp = 0;
13698 // Use a target machine opcode to prevent further DAGCombine
13699 // optimizations that may separate the arithmetic operations
13700 // from the setcc node.
13701 switch (WideVal.getOpcode()) {
13703 case ISD::ADD: ConvertedOp = X86ISD::ADD; break;
13704 case ISD::SUB: ConvertedOp = X86ISD::SUB; break;
13705 case ISD::AND: ConvertedOp = X86ISD::AND; break;
13706 case ISD::OR: ConvertedOp = X86ISD::OR; break;
13707 case ISD::XOR: ConvertedOp = X86ISD::XOR; break;
13711 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
13712 if (TLI.isOperationLegal(WideVal.getOpcode(), WideVT)) {
13713 SDValue V0 = DAG.getNode(ISD::TRUNCATE, dl, VT, WideVal.getOperand(0));
13714 SDValue V1 = DAG.getNode(ISD::TRUNCATE, dl, VT, WideVal.getOperand(1));
13715 Op = DAG.getNode(ConvertedOp, dl, VT, V0, V1);
13721 // Emit a CMP with 0, which is the TEST pattern.
13722 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op,
13723 DAG.getConstant(0, dl, Op.getValueType()));
13725 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::i32);
13726 SmallVector<SDValue, 4> Ops(Op->op_begin(), Op->op_begin() + NumOperands);
13728 SDValue New = DAG.getNode(Opcode, dl, VTs, Ops);
13729 DAG.ReplaceAllUsesWith(Op, New);
13730 return SDValue(New.getNode(), 1);
13733 /// Emit nodes that will be selected as "cmp Op0,Op1", or something
13735 SDValue X86TargetLowering::EmitCmp(SDValue Op0, SDValue Op1, unsigned X86CC,
13736 SDLoc dl, SelectionDAG &DAG) const {
13737 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op1)) {
13738 if (C->getAPIntValue() == 0)
13739 return EmitTest(Op0, X86CC, dl, DAG);
13741 if (Op0.getValueType() == MVT::i1)
13742 llvm_unreachable("Unexpected comparison operation for MVT::i1 operands");
13745 if ((Op0.getValueType() == MVT::i8 || Op0.getValueType() == MVT::i16 ||
13746 Op0.getValueType() == MVT::i32 || Op0.getValueType() == MVT::i64)) {
13747 // Do the comparison at i32 if it's smaller, besides the Atom case.
13748 // This avoids subregister aliasing issues. Keep the smaller reference
13749 // if we're optimizing for size, however, as that'll allow better folding
13750 // of memory operations.
13751 if (Op0.getValueType() != MVT::i32 && Op0.getValueType() != MVT::i64 &&
13752 !DAG.getMachineFunction().getFunction()->optForMinSize() &&
13753 !Subtarget->isAtom()) {
13754 unsigned ExtendOp =
13755 isX86CCUnsigned(X86CC) ? ISD::ZERO_EXTEND : ISD::SIGN_EXTEND;
13756 Op0 = DAG.getNode(ExtendOp, dl, MVT::i32, Op0);
13757 Op1 = DAG.getNode(ExtendOp, dl, MVT::i32, Op1);
13759 // Use SUB instead of CMP to enable CSE between SUB and CMP.
13760 SDVTList VTs = DAG.getVTList(Op0.getValueType(), MVT::i32);
13761 SDValue Sub = DAG.getNode(X86ISD::SUB, dl, VTs,
13763 return SDValue(Sub.getNode(), 1);
13765 return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op0, Op1);
13768 /// Convert a comparison if required by the subtarget.
13769 SDValue X86TargetLowering::ConvertCmpIfNecessary(SDValue Cmp,
13770 SelectionDAG &DAG) const {
13771 // If the subtarget does not support the FUCOMI instruction, floating-point
13772 // comparisons have to be converted.
13773 if (Subtarget->hasCMov() ||
13774 Cmp.getOpcode() != X86ISD::CMP ||
13775 !Cmp.getOperand(0).getValueType().isFloatingPoint() ||
13776 !Cmp.getOperand(1).getValueType().isFloatingPoint())
13779 // The instruction selector will select an FUCOM instruction instead of
13780 // FUCOMI, which writes the comparison result to FPSW instead of EFLAGS. Hence
13781 // build an SDNode sequence that transfers the result from FPSW into EFLAGS:
13782 // (X86sahf (trunc (srl (X86fp_stsw (trunc (X86cmp ...)), 8))))
13784 SDValue TruncFPSW = DAG.getNode(ISD::TRUNCATE, dl, MVT::i16, Cmp);
13785 SDValue FNStSW = DAG.getNode(X86ISD::FNSTSW16r, dl, MVT::i16, TruncFPSW);
13786 SDValue Srl = DAG.getNode(ISD::SRL, dl, MVT::i16, FNStSW,
13787 DAG.getConstant(8, dl, MVT::i8));
13788 SDValue TruncSrl = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Srl);
13789 return DAG.getNode(X86ISD::SAHF, dl, MVT::i32, TruncSrl);
13792 /// The minimum architected relative accuracy is 2^-12. We need one
13793 /// Newton-Raphson step to have a good float result (24 bits of precision).
13794 SDValue X86TargetLowering::getRsqrtEstimate(SDValue Op,
13795 DAGCombinerInfo &DCI,
13796 unsigned &RefinementSteps,
13797 bool &UseOneConstNR) const {
13798 EVT VT = Op.getValueType();
13799 const char *RecipOp;
13801 // SSE1 has rsqrtss and rsqrtps. AVX adds a 256-bit variant for rsqrtps.
13802 // TODO: Add support for AVX512 (v16f32).
13803 // It is likely not profitable to do this for f64 because a double-precision
13804 // rsqrt estimate with refinement on x86 prior to FMA requires at least 16
13805 // instructions: convert to single, rsqrtss, convert back to double, refine
13806 // (3 steps = at least 13 insts). If an 'rsqrtsd' variant was added to the ISA
13807 // along with FMA, this could be a throughput win.
13808 if (VT == MVT::f32 && Subtarget->hasSSE1())
13810 else if ((VT == MVT::v4f32 && Subtarget->hasSSE1()) ||
13811 (VT == MVT::v8f32 && Subtarget->hasAVX()))
13812 RecipOp = "vec-sqrtf";
13816 TargetRecip Recips = DCI.DAG.getTarget().Options.Reciprocals;
13817 if (!Recips.isEnabled(RecipOp))
13820 RefinementSteps = Recips.getRefinementSteps(RecipOp);
13821 UseOneConstNR = false;
13822 return DCI.DAG.getNode(X86ISD::FRSQRT, SDLoc(Op), VT, Op);
13825 /// The minimum architected relative accuracy is 2^-12. We need one
13826 /// Newton-Raphson step to have a good float result (24 bits of precision).
13827 SDValue X86TargetLowering::getRecipEstimate(SDValue Op,
13828 DAGCombinerInfo &DCI,
13829 unsigned &RefinementSteps) const {
13830 EVT VT = Op.getValueType();
13831 const char *RecipOp;
13833 // SSE1 has rcpss and rcpps. AVX adds a 256-bit variant for rcpps.
13834 // TODO: Add support for AVX512 (v16f32).
13835 // It is likely not profitable to do this for f64 because a double-precision
13836 // reciprocal estimate with refinement on x86 prior to FMA requires
13837 // 15 instructions: convert to single, rcpss, convert back to double, refine
13838 // (3 steps = 12 insts). If an 'rcpsd' variant was added to the ISA
13839 // along with FMA, this could be a throughput win.
13840 if (VT == MVT::f32 && Subtarget->hasSSE1())
13842 else if ((VT == MVT::v4f32 && Subtarget->hasSSE1()) ||
13843 (VT == MVT::v8f32 && Subtarget->hasAVX()))
13844 RecipOp = "vec-divf";
13848 TargetRecip Recips = DCI.DAG.getTarget().Options.Reciprocals;
13849 if (!Recips.isEnabled(RecipOp))
13852 RefinementSteps = Recips.getRefinementSteps(RecipOp);
13853 return DCI.DAG.getNode(X86ISD::FRCP, SDLoc(Op), VT, Op);
13856 /// If we have at least two divisions that use the same divisor, convert to
13857 /// multplication by a reciprocal. This may need to be adjusted for a given
13858 /// CPU if a division's cost is not at least twice the cost of a multiplication.
13859 /// This is because we still need one division to calculate the reciprocal and
13860 /// then we need two multiplies by that reciprocal as replacements for the
13861 /// original divisions.
13862 unsigned X86TargetLowering::combineRepeatedFPDivisors() const {
13866 static bool isAllOnes(SDValue V) {
13867 ConstantSDNode *C = dyn_cast<ConstantSDNode>(V);
13868 return C && C->isAllOnesValue();
13871 /// LowerToBT - Result of 'and' is compared against zero. Turn it into a BT node
13872 /// if it's possible.
13873 SDValue X86TargetLowering::LowerToBT(SDValue And, ISD::CondCode CC,
13874 SDLoc dl, SelectionDAG &DAG) const {
13875 SDValue Op0 = And.getOperand(0);
13876 SDValue Op1 = And.getOperand(1);
13877 if (Op0.getOpcode() == ISD::TRUNCATE)
13878 Op0 = Op0.getOperand(0);
13879 if (Op1.getOpcode() == ISD::TRUNCATE)
13880 Op1 = Op1.getOperand(0);
13883 if (Op1.getOpcode() == ISD::SHL)
13884 std::swap(Op0, Op1);
13885 if (Op0.getOpcode() == ISD::SHL) {
13886 if (ConstantSDNode *And00C = dyn_cast<ConstantSDNode>(Op0.getOperand(0)))
13887 if (And00C->getZExtValue() == 1) {
13888 // If we looked past a truncate, check that it's only truncating away
13890 unsigned BitWidth = Op0.getValueSizeInBits();
13891 unsigned AndBitWidth = And.getValueSizeInBits();
13892 if (BitWidth > AndBitWidth) {
13894 DAG.computeKnownBits(Op0, Zeros, Ones);
13895 if (Zeros.countLeadingOnes() < BitWidth - AndBitWidth)
13899 RHS = Op0.getOperand(1);
13901 } else if (Op1.getOpcode() == ISD::Constant) {
13902 ConstantSDNode *AndRHS = cast<ConstantSDNode>(Op1);
13903 uint64_t AndRHSVal = AndRHS->getZExtValue();
13904 SDValue AndLHS = Op0;
13906 if (AndRHSVal == 1 && AndLHS.getOpcode() == ISD::SRL) {
13907 LHS = AndLHS.getOperand(0);
13908 RHS = AndLHS.getOperand(1);
13911 // Use BT if the immediate can't be encoded in a TEST instruction.
13912 if (!isUInt<32>(AndRHSVal) && isPowerOf2_64(AndRHSVal)) {
13914 RHS = DAG.getConstant(Log2_64_Ceil(AndRHSVal), dl, LHS.getValueType());
13918 if (LHS.getNode()) {
13919 // If LHS is i8, promote it to i32 with any_extend. There is no i8 BT
13920 // instruction. Since the shift amount is in-range-or-undefined, we know
13921 // that doing a bittest on the i32 value is ok. We extend to i32 because
13922 // the encoding for the i16 version is larger than the i32 version.
13923 // Also promote i16 to i32 for performance / code size reason.
13924 if (LHS.getValueType() == MVT::i8 ||
13925 LHS.getValueType() == MVT::i16)
13926 LHS = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, LHS);
13928 // If the operand types disagree, extend the shift amount to match. Since
13929 // BT ignores high bits (like shifts) we can use anyextend.
13930 if (LHS.getValueType() != RHS.getValueType())
13931 RHS = DAG.getNode(ISD::ANY_EXTEND, dl, LHS.getValueType(), RHS);
13933 SDValue BT = DAG.getNode(X86ISD::BT, dl, MVT::i32, LHS, RHS);
13934 X86::CondCode Cond = CC == ISD::SETEQ ? X86::COND_AE : X86::COND_B;
13935 return DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
13936 DAG.getConstant(Cond, dl, MVT::i8), BT);
13942 /// \brief - Turns an ISD::CondCode into a value suitable for SSE floating point
13944 static int translateX86FSETCC(ISD::CondCode SetCCOpcode, SDValue &Op0,
13949 // SSE Condition code mapping:
13958 switch (SetCCOpcode) {
13959 default: llvm_unreachable("Unexpected SETCC condition");
13961 case ISD::SETEQ: SSECC = 0; break;
13963 case ISD::SETGT: Swap = true; // Fallthrough
13965 case ISD::SETOLT: SSECC = 1; break;
13967 case ISD::SETGE: Swap = true; // Fallthrough
13969 case ISD::SETOLE: SSECC = 2; break;
13970 case ISD::SETUO: SSECC = 3; break;
13972 case ISD::SETNE: SSECC = 4; break;
13973 case ISD::SETULE: Swap = true; // Fallthrough
13974 case ISD::SETUGE: SSECC = 5; break;
13975 case ISD::SETULT: Swap = true; // Fallthrough
13976 case ISD::SETUGT: SSECC = 6; break;
13977 case ISD::SETO: SSECC = 7; break;
13979 case ISD::SETONE: SSECC = 8; break;
13982 std::swap(Op0, Op1);
13987 // Lower256IntVSETCC - Break a VSETCC 256-bit integer VSETCC into two new 128
13988 // ones, and then concatenate the result back.
13989 static SDValue Lower256IntVSETCC(SDValue Op, SelectionDAG &DAG) {
13990 MVT VT = Op.getSimpleValueType();
13992 assert(VT.is256BitVector() && Op.getOpcode() == ISD::SETCC &&
13993 "Unsupported value type for operation");
13995 unsigned NumElems = VT.getVectorNumElements();
13997 SDValue CC = Op.getOperand(2);
13999 // Extract the LHS vectors
14000 SDValue LHS = Op.getOperand(0);
14001 SDValue LHS1 = Extract128BitVector(LHS, 0, DAG, dl);
14002 SDValue LHS2 = Extract128BitVector(LHS, NumElems/2, DAG, dl);
14004 // Extract the RHS vectors
14005 SDValue RHS = Op.getOperand(1);
14006 SDValue RHS1 = Extract128BitVector(RHS, 0, DAG, dl);
14007 SDValue RHS2 = Extract128BitVector(RHS, NumElems/2, DAG, dl);
14009 // Issue the operation on the smaller types and concatenate the result back
14010 MVT EltVT = VT.getVectorElementType();
14011 MVT NewVT = MVT::getVectorVT(EltVT, NumElems/2);
14012 return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT,
14013 DAG.getNode(Op.getOpcode(), dl, NewVT, LHS1, RHS1, CC),
14014 DAG.getNode(Op.getOpcode(), dl, NewVT, LHS2, RHS2, CC));
14017 static SDValue LowerBoolVSETCC_AVX512(SDValue Op, SelectionDAG &DAG) {
14018 SDValue Op0 = Op.getOperand(0);
14019 SDValue Op1 = Op.getOperand(1);
14020 SDValue CC = Op.getOperand(2);
14021 MVT VT = Op.getSimpleValueType();
14024 assert(Op0.getValueType().getVectorElementType() == MVT::i1 &&
14025 "Unexpected type for boolean compare operation");
14026 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
14027 SDValue NotOp0 = DAG.getNode(ISD::XOR, dl, VT, Op0,
14028 DAG.getConstant(-1, dl, VT));
14029 SDValue NotOp1 = DAG.getNode(ISD::XOR, dl, VT, Op1,
14030 DAG.getConstant(-1, dl, VT));
14031 switch (SetCCOpcode) {
14032 default: llvm_unreachable("Unexpected SETCC condition");
14034 // (x == y) -> ~(x ^ y)
14035 return DAG.getNode(ISD::XOR, dl, VT,
14036 DAG.getNode(ISD::XOR, dl, VT, Op0, Op1),
14037 DAG.getConstant(-1, dl, VT));
14039 // (x != y) -> (x ^ y)
14040 return DAG.getNode(ISD::XOR, dl, VT, Op0, Op1);
14043 // (x > y) -> (x & ~y)
14044 return DAG.getNode(ISD::AND, dl, VT, Op0, NotOp1);
14047 // (x < y) -> (~x & y)
14048 return DAG.getNode(ISD::AND, dl, VT, NotOp0, Op1);
14051 // (x <= y) -> (~x | y)
14052 return DAG.getNode(ISD::OR, dl, VT, NotOp0, Op1);
14055 // (x >=y) -> (x | ~y)
14056 return DAG.getNode(ISD::OR, dl, VT, Op0, NotOp1);
14060 static SDValue LowerIntVSETCC_AVX512(SDValue Op, SelectionDAG &DAG,
14061 const X86Subtarget *Subtarget) {
14062 SDValue Op0 = Op.getOperand(0);
14063 SDValue Op1 = Op.getOperand(1);
14064 SDValue CC = Op.getOperand(2);
14065 MVT VT = Op.getSimpleValueType();
14068 assert(Op0.getValueType().getVectorElementType().getSizeInBits() >= 8 &&
14069 Op.getValueType().getScalarType() == MVT::i1 &&
14070 "Cannot set masked compare for this operation");
14072 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
14074 bool Unsigned = false;
14077 switch (SetCCOpcode) {
14078 default: llvm_unreachable("Unexpected SETCC condition");
14079 case ISD::SETNE: SSECC = 4; break;
14080 case ISD::SETEQ: Opc = X86ISD::PCMPEQM; break;
14081 case ISD::SETUGT: SSECC = 6; Unsigned = true; break;
14082 case ISD::SETLT: Swap = true; //fall-through
14083 case ISD::SETGT: Opc = X86ISD::PCMPGTM; break;
14084 case ISD::SETULT: SSECC = 1; Unsigned = true; break;
14085 case ISD::SETUGE: SSECC = 5; Unsigned = true; break; //NLT
14086 case ISD::SETGE: Swap = true; SSECC = 2; break; // LE + swap
14087 case ISD::SETULE: Unsigned = true; //fall-through
14088 case ISD::SETLE: SSECC = 2; break;
14092 std::swap(Op0, Op1);
14094 return DAG.getNode(Opc, dl, VT, Op0, Op1);
14095 Opc = Unsigned ? X86ISD::CMPMU: X86ISD::CMPM;
14096 return DAG.getNode(Opc, dl, VT, Op0, Op1,
14097 DAG.getConstant(SSECC, dl, MVT::i8));
14100 /// \brief Try to turn a VSETULT into a VSETULE by modifying its second
14101 /// operand \p Op1. If non-trivial (for example because it's not constant)
14102 /// return an empty value.
14103 static SDValue ChangeVSETULTtoVSETULE(SDLoc dl, SDValue Op1, SelectionDAG &DAG)
14105 BuildVectorSDNode *BV = dyn_cast<BuildVectorSDNode>(Op1.getNode());
14109 MVT VT = Op1.getSimpleValueType();
14110 MVT EVT = VT.getVectorElementType();
14111 unsigned n = VT.getVectorNumElements();
14112 SmallVector<SDValue, 8> ULTOp1;
14114 for (unsigned i = 0; i < n; ++i) {
14115 ConstantSDNode *Elt = dyn_cast<ConstantSDNode>(BV->getOperand(i));
14116 if (!Elt || Elt->isOpaque() || Elt->getValueType(0) != EVT)
14119 // Avoid underflow.
14120 APInt Val = Elt->getAPIntValue();
14124 ULTOp1.push_back(DAG.getConstant(Val - 1, dl, EVT));
14127 return DAG.getNode(ISD::BUILD_VECTOR, dl, VT, ULTOp1);
14130 static SDValue LowerVSETCC(SDValue Op, const X86Subtarget *Subtarget,
14131 SelectionDAG &DAG) {
14132 SDValue Op0 = Op.getOperand(0);
14133 SDValue Op1 = Op.getOperand(1);
14134 SDValue CC = Op.getOperand(2);
14135 MVT VT = Op.getSimpleValueType();
14136 ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
14137 bool isFP = Op.getOperand(1).getSimpleValueType().isFloatingPoint();
14142 MVT EltVT = Op0.getSimpleValueType().getVectorElementType();
14143 assert(EltVT == MVT::f32 || EltVT == MVT::f64);
14146 unsigned SSECC = translateX86FSETCC(SetCCOpcode, Op0, Op1);
14147 unsigned Opc = X86ISD::CMPP;
14148 if (Subtarget->hasAVX512() && VT.getVectorElementType() == MVT::i1) {
14149 assert(VT.getVectorNumElements() <= 16);
14150 Opc = X86ISD::CMPM;
14152 // In the two special cases we can't handle, emit two comparisons.
14155 unsigned CombineOpc;
14156 if (SetCCOpcode == ISD::SETUEQ) {
14157 CC0 = 3; CC1 = 0; CombineOpc = ISD::OR;
14159 assert(SetCCOpcode == ISD::SETONE);
14160 CC0 = 7; CC1 = 4; CombineOpc = ISD::AND;
14163 SDValue Cmp0 = DAG.getNode(Opc, dl, VT, Op0, Op1,
14164 DAG.getConstant(CC0, dl, MVT::i8));
14165 SDValue Cmp1 = DAG.getNode(Opc, dl, VT, Op0, Op1,
14166 DAG.getConstant(CC1, dl, MVT::i8));
14167 return DAG.getNode(CombineOpc, dl, VT, Cmp0, Cmp1);
14169 // Handle all other FP comparisons here.
14170 return DAG.getNode(Opc, dl, VT, Op0, Op1,
14171 DAG.getConstant(SSECC, dl, MVT::i8));
14174 MVT VTOp0 = Op0.getSimpleValueType();
14175 assert(VTOp0 == Op1.getSimpleValueType() &&
14176 "Expected operands with same type!");
14177 assert(VT.getVectorNumElements() == VTOp0.getVectorNumElements() &&
14178 "Invalid number of packed elements for source and destination!");
14180 if (VT.is128BitVector() && VTOp0.is256BitVector()) {
14181 // On non-AVX512 targets, a vector of MVT::i1 is promoted by the type
14182 // legalizer to a wider vector type. In the case of 'vsetcc' nodes, the
14183 // legalizer firstly checks if the first operand in input to the setcc has
14184 // a legal type. If so, then it promotes the return type to that same type.
14185 // Otherwise, the return type is promoted to the 'next legal type' which,
14186 // for a vector of MVT::i1 is always a 128-bit integer vector type.
14188 // We reach this code only if the following two conditions are met:
14189 // 1. Both return type and operand type have been promoted to wider types
14190 // by the type legalizer.
14191 // 2. The original operand type has been promoted to a 256-bit vector.
14193 // Note that condition 2. only applies for AVX targets.
14194 SDValue NewOp = DAG.getSetCC(dl, VTOp0, Op0, Op1, SetCCOpcode);
14195 return DAG.getZExtOrTrunc(NewOp, dl, VT);
14198 // The non-AVX512 code below works under the assumption that source and
14199 // destination types are the same.
14200 assert((Subtarget->hasAVX512() || (VT == VTOp0)) &&
14201 "Value types for source and destination must be the same!");
14203 // Break 256-bit integer vector compare into smaller ones.
14204 if (VT.is256BitVector() && !Subtarget->hasInt256())
14205 return Lower256IntVSETCC(Op, DAG);
14207 EVT OpVT = Op1.getValueType();
14208 if (OpVT.getVectorElementType() == MVT::i1)
14209 return LowerBoolVSETCC_AVX512(Op, DAG);
14211 bool MaskResult = (VT.getVectorElementType() == MVT::i1);
14212 if (Subtarget->hasAVX512()) {
14213 if (Op1.getValueType().is512BitVector() ||
14214 (Subtarget->hasBWI() && Subtarget->hasVLX()) ||
14215 (MaskResult && OpVT.getVectorElementType().getSizeInBits() >= 32))
14216 return LowerIntVSETCC_AVX512(Op, DAG, Subtarget);
14218 // In AVX-512 architecture setcc returns mask with i1 elements,
14219 // But there is no compare instruction for i8 and i16 elements in KNL.
14220 // We are not talking about 512-bit operands in this case, these
14221 // types are illegal.
14223 (OpVT.getVectorElementType().getSizeInBits() < 32 &&
14224 OpVT.getVectorElementType().getSizeInBits() >= 8))
14225 return DAG.getNode(ISD::TRUNCATE, dl, VT,
14226 DAG.getNode(ISD::SETCC, dl, OpVT, Op0, Op1, CC));
14229 // Lower using XOP integer comparisons.
14230 if ((VT == MVT::v16i8 || VT == MVT::v8i16 ||
14231 VT == MVT::v4i32 || VT == MVT::v2i64) && Subtarget->hasXOP()) {
14232 // Translate compare code to XOP PCOM compare mode.
14233 unsigned CmpMode = 0;
14234 switch (SetCCOpcode) {
14235 default: llvm_unreachable("Unexpected SETCC condition");
14237 case ISD::SETLT: CmpMode = 0x00; break;
14239 case ISD::SETLE: CmpMode = 0x01; break;
14241 case ISD::SETGT: CmpMode = 0x02; break;
14243 case ISD::SETGE: CmpMode = 0x03; break;
14244 case ISD::SETEQ: CmpMode = 0x04; break;
14245 case ISD::SETNE: CmpMode = 0x05; break;
14248 // Are we comparing unsigned or signed integers?
14249 unsigned Opc = ISD::isUnsignedIntSetCC(SetCCOpcode)
14250 ? X86ISD::VPCOMU : X86ISD::VPCOM;
14252 return DAG.getNode(Opc, dl, VT, Op0, Op1,
14253 DAG.getConstant(CmpMode, dl, MVT::i8));
14256 // We are handling one of the integer comparisons here. Since SSE only has
14257 // GT and EQ comparisons for integer, swapping operands and multiple
14258 // operations may be required for some comparisons.
14260 bool Swap = false, Invert = false, FlipSigns = false, MinMax = false;
14261 bool Subus = false;
14263 switch (SetCCOpcode) {
14264 default: llvm_unreachable("Unexpected SETCC condition");
14265 case ISD::SETNE: Invert = true;
14266 case ISD::SETEQ: Opc = X86ISD::PCMPEQ; break;
14267 case ISD::SETLT: Swap = true;
14268 case ISD::SETGT: Opc = X86ISD::PCMPGT; break;
14269 case ISD::SETGE: Swap = true;
14270 case ISD::SETLE: Opc = X86ISD::PCMPGT;
14271 Invert = true; break;
14272 case ISD::SETULT: Swap = true;
14273 case ISD::SETUGT: Opc = X86ISD::PCMPGT;
14274 FlipSigns = true; break;
14275 case ISD::SETUGE: Swap = true;
14276 case ISD::SETULE: Opc = X86ISD::PCMPGT;
14277 FlipSigns = true; Invert = true; break;
14280 // Special case: Use min/max operations for SETULE/SETUGE
14281 MVT VET = VT.getVectorElementType();
14283 (Subtarget->hasSSE41() && (VET >= MVT::i8 && VET <= MVT::i32))
14284 || (Subtarget->hasSSE2() && (VET == MVT::i8));
14287 switch (SetCCOpcode) {
14289 case ISD::SETULE: Opc = ISD::UMIN; MinMax = true; break;
14290 case ISD::SETUGE: Opc = ISD::UMAX; MinMax = true; break;
14293 if (MinMax) { Swap = false; Invert = false; FlipSigns = false; }
14296 bool hasSubus = Subtarget->hasSSE2() && (VET == MVT::i8 || VET == MVT::i16);
14297 if (!MinMax && hasSubus) {
14298 // As another special case, use PSUBUS[BW] when it's profitable. E.g. for
14300 // t = psubus Op0, Op1
14301 // pcmpeq t, <0..0>
14302 switch (SetCCOpcode) {
14304 case ISD::SETULT: {
14305 // If the comparison is against a constant we can turn this into a
14306 // setule. With psubus, setule does not require a swap. This is
14307 // beneficial because the constant in the register is no longer
14308 // destructed as the destination so it can be hoisted out of a loop.
14309 // Only do this pre-AVX since vpcmp* is no longer destructive.
14310 if (Subtarget->hasAVX())
14312 SDValue ULEOp1 = ChangeVSETULTtoVSETULE(dl, Op1, DAG);
14313 if (ULEOp1.getNode()) {
14315 Subus = true; Invert = false; Swap = false;
14319 // Psubus is better than flip-sign because it requires no inversion.
14320 case ISD::SETUGE: Subus = true; Invert = false; Swap = true; break;
14321 case ISD::SETULE: Subus = true; Invert = false; Swap = false; break;
14325 Opc = X86ISD::SUBUS;
14331 std::swap(Op0, Op1);
14333 // Check that the operation in question is available (most are plain SSE2,
14334 // but PCMPGTQ and PCMPEQQ have different requirements).
14335 if (VT == MVT::v2i64) {
14336 if (Opc == X86ISD::PCMPGT && !Subtarget->hasSSE42()) {
14337 assert(Subtarget->hasSSE2() && "Don't know how to lower!");
14339 // First cast everything to the right type.
14340 Op0 = DAG.getBitcast(MVT::v4i32, Op0);
14341 Op1 = DAG.getBitcast(MVT::v4i32, Op1);
14343 // Since SSE has no unsigned integer comparisons, we need to flip the sign
14344 // bits of the inputs before performing those operations. The lower
14345 // compare is always unsigned.
14348 SB = DAG.getConstant(0x80000000U, dl, MVT::v4i32);
14350 SDValue Sign = DAG.getConstant(0x80000000U, dl, MVT::i32);
14351 SDValue Zero = DAG.getConstant(0x00000000U, dl, MVT::i32);
14352 SB = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32,
14353 Sign, Zero, Sign, Zero);
14355 Op0 = DAG.getNode(ISD::XOR, dl, MVT::v4i32, Op0, SB);
14356 Op1 = DAG.getNode(ISD::XOR, dl, MVT::v4i32, Op1, SB);
14358 // Emulate PCMPGTQ with (hi1 > hi2) | ((hi1 == hi2) & (lo1 > lo2))
14359 SDValue GT = DAG.getNode(X86ISD::PCMPGT, dl, MVT::v4i32, Op0, Op1);
14360 SDValue EQ = DAG.getNode(X86ISD::PCMPEQ, dl, MVT::v4i32, Op0, Op1);
14362 // Create masks for only the low parts/high parts of the 64 bit integers.
14363 static const int MaskHi[] = { 1, 1, 3, 3 };
14364 static const int MaskLo[] = { 0, 0, 2, 2 };
14365 SDValue EQHi = DAG.getVectorShuffle(MVT::v4i32, dl, EQ, EQ, MaskHi);
14366 SDValue GTLo = DAG.getVectorShuffle(MVT::v4i32, dl, GT, GT, MaskLo);
14367 SDValue GTHi = DAG.getVectorShuffle(MVT::v4i32, dl, GT, GT, MaskHi);
14369 SDValue Result = DAG.getNode(ISD::AND, dl, MVT::v4i32, EQHi, GTLo);
14370 Result = DAG.getNode(ISD::OR, dl, MVT::v4i32, Result, GTHi);
14373 Result = DAG.getNOT(dl, Result, MVT::v4i32);
14375 return DAG.getBitcast(VT, Result);
14378 if (Opc == X86ISD::PCMPEQ && !Subtarget->hasSSE41()) {
14379 // If pcmpeqq is missing but pcmpeqd is available synthesize pcmpeqq with
14380 // pcmpeqd + pshufd + pand.
14381 assert(Subtarget->hasSSE2() && !FlipSigns && "Don't know how to lower!");
14383 // First cast everything to the right type.
14384 Op0 = DAG.getBitcast(MVT::v4i32, Op0);
14385 Op1 = DAG.getBitcast(MVT::v4i32, Op1);
14388 SDValue Result = DAG.getNode(Opc, dl, MVT::v4i32, Op0, Op1);
14390 // Make sure the lower and upper halves are both all-ones.
14391 static const int Mask[] = { 1, 0, 3, 2 };
14392 SDValue Shuf = DAG.getVectorShuffle(MVT::v4i32, dl, Result, Result, Mask);
14393 Result = DAG.getNode(ISD::AND, dl, MVT::v4i32, Result, Shuf);
14396 Result = DAG.getNOT(dl, Result, MVT::v4i32);
14398 return DAG.getBitcast(VT, Result);
14402 // Since SSE has no unsigned integer comparisons, we need to flip the sign
14403 // bits of the inputs before performing those operations.
14405 EVT EltVT = VT.getVectorElementType();
14406 SDValue SB = DAG.getConstant(APInt::getSignBit(EltVT.getSizeInBits()), dl,
14408 Op0 = DAG.getNode(ISD::XOR, dl, VT, Op0, SB);
14409 Op1 = DAG.getNode(ISD::XOR, dl, VT, Op1, SB);
14412 SDValue Result = DAG.getNode(Opc, dl, VT, Op0, Op1);
14414 // If the logical-not of the result is required, perform that now.
14416 Result = DAG.getNOT(dl, Result, VT);
14419 Result = DAG.getNode(X86ISD::PCMPEQ, dl, VT, Op0, Result);
14422 Result = DAG.getNode(X86ISD::PCMPEQ, dl, VT, Result,
14423 getZeroVector(VT, Subtarget, DAG, dl));
14428 SDValue X86TargetLowering::LowerSETCC(SDValue Op, SelectionDAG &DAG) const {
14430 MVT VT = Op.getSimpleValueType();
14432 if (VT.isVector()) return LowerVSETCC(Op, Subtarget, DAG);
14434 assert(((!Subtarget->hasAVX512() && VT == MVT::i8) || (VT == MVT::i1))
14435 && "SetCC type must be 8-bit or 1-bit integer");
14436 SDValue Op0 = Op.getOperand(0);
14437 SDValue Op1 = Op.getOperand(1);
14439 ISD::CondCode CC = cast<CondCodeSDNode>(Op.getOperand(2))->get();
14441 // Optimize to BT if possible.
14442 // Lower (X & (1 << N)) == 0 to BT(X, N).
14443 // Lower ((X >>u N) & 1) != 0 to BT(X, N).
14444 // Lower ((X >>s N) & 1) != 0 to BT(X, N).
14445 if (Op0.getOpcode() == ISD::AND && Op0.hasOneUse() &&
14446 Op1.getOpcode() == ISD::Constant &&
14447 cast<ConstantSDNode>(Op1)->isNullValue() &&
14448 (CC == ISD::SETEQ || CC == ISD::SETNE)) {
14449 SDValue NewSetCC = LowerToBT(Op0, CC, dl, DAG);
14450 if (NewSetCC.getNode()) {
14452 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i1, NewSetCC);
14457 // Look for X == 0, X == 1, X != 0, or X != 1. We can simplify some forms of
14459 if (Op1.getOpcode() == ISD::Constant &&
14460 (cast<ConstantSDNode>(Op1)->getZExtValue() == 1 ||
14461 cast<ConstantSDNode>(Op1)->isNullValue()) &&
14462 (CC == ISD::SETEQ || CC == ISD::SETNE)) {
14464 // If the input is a setcc, then reuse the input setcc or use a new one with
14465 // the inverted condition.
14466 if (Op0.getOpcode() == X86ISD::SETCC) {
14467 X86::CondCode CCode = (X86::CondCode)Op0.getConstantOperandVal(0);
14468 bool Invert = (CC == ISD::SETNE) ^
14469 cast<ConstantSDNode>(Op1)->isNullValue();
14473 CCode = X86::GetOppositeBranchCondition(CCode);
14474 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
14475 DAG.getConstant(CCode, dl, MVT::i8),
14476 Op0.getOperand(1));
14478 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i1, SetCC);
14482 if ((Op0.getValueType() == MVT::i1) && (Op1.getOpcode() == ISD::Constant) &&
14483 (cast<ConstantSDNode>(Op1)->getZExtValue() == 1) &&
14484 (CC == ISD::SETEQ || CC == ISD::SETNE)) {
14486 ISD::CondCode NewCC = ISD::getSetCCInverse(CC, true);
14487 return DAG.getSetCC(dl, VT, Op0, DAG.getConstant(0, dl, MVT::i1), NewCC);
14490 bool isFP = Op1.getSimpleValueType().isFloatingPoint();
14491 unsigned X86CC = TranslateX86CC(CC, dl, isFP, Op0, Op1, DAG);
14492 if (X86CC == X86::COND_INVALID)
14495 SDValue EFLAGS = EmitCmp(Op0, Op1, X86CC, dl, DAG);
14496 EFLAGS = ConvertCmpIfNecessary(EFLAGS, DAG);
14497 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
14498 DAG.getConstant(X86CC, dl, MVT::i8), EFLAGS);
14500 return DAG.getNode(ISD::TRUNCATE, dl, MVT::i1, SetCC);
14504 // isX86LogicalCmp - Return true if opcode is a X86 logical comparison.
14505 static bool isX86LogicalCmp(SDValue Op) {
14506 unsigned Opc = Op.getNode()->getOpcode();
14507 if (Opc == X86ISD::CMP || Opc == X86ISD::COMI || Opc == X86ISD::UCOMI ||
14508 Opc == X86ISD::SAHF)
14510 if (Op.getResNo() == 1 &&
14511 (Opc == X86ISD::ADD ||
14512 Opc == X86ISD::SUB ||
14513 Opc == X86ISD::ADC ||
14514 Opc == X86ISD::SBB ||
14515 Opc == X86ISD::SMUL ||
14516 Opc == X86ISD::UMUL ||
14517 Opc == X86ISD::INC ||
14518 Opc == X86ISD::DEC ||
14519 Opc == X86ISD::OR ||
14520 Opc == X86ISD::XOR ||
14521 Opc == X86ISD::AND))
14524 if (Op.getResNo() == 2 && Opc == X86ISD::UMUL)
14530 static bool isTruncWithZeroHighBitsInput(SDValue V, SelectionDAG &DAG) {
14531 if (V.getOpcode() != ISD::TRUNCATE)
14534 SDValue VOp0 = V.getOperand(0);
14535 unsigned InBits = VOp0.getValueSizeInBits();
14536 unsigned Bits = V.getValueSizeInBits();
14537 return DAG.MaskedValueIsZero(VOp0, APInt::getHighBitsSet(InBits,InBits-Bits));
14540 SDValue X86TargetLowering::LowerSELECT(SDValue Op, SelectionDAG &DAG) const {
14541 bool addTest = true;
14542 SDValue Cond = Op.getOperand(0);
14543 SDValue Op1 = Op.getOperand(1);
14544 SDValue Op2 = Op.getOperand(2);
14546 EVT VT = Op1.getValueType();
14549 // Lower FP selects into a CMP/AND/ANDN/OR sequence when the necessary SSE ops
14550 // are available or VBLENDV if AVX is available.
14551 // Otherwise FP cmovs get lowered into a less efficient branch sequence later.
14552 if (Cond.getOpcode() == ISD::SETCC &&
14553 ((Subtarget->hasSSE2() && (VT == MVT::f32 || VT == MVT::f64)) ||
14554 (Subtarget->hasSSE1() && VT == MVT::f32)) &&
14555 VT == Cond.getOperand(0).getValueType() && Cond->hasOneUse()) {
14556 SDValue CondOp0 = Cond.getOperand(0), CondOp1 = Cond.getOperand(1);
14557 int SSECC = translateX86FSETCC(
14558 cast<CondCodeSDNode>(Cond.getOperand(2))->get(), CondOp0, CondOp1);
14561 if (Subtarget->hasAVX512()) {
14562 SDValue Cmp = DAG.getNode(X86ISD::FSETCC, DL, MVT::i1, CondOp0, CondOp1,
14563 DAG.getConstant(SSECC, DL, MVT::i8));
14564 return DAG.getNode(X86ISD::SELECT, DL, VT, Cmp, Op1, Op2);
14567 SDValue Cmp = DAG.getNode(X86ISD::FSETCC, DL, VT, CondOp0, CondOp1,
14568 DAG.getConstant(SSECC, DL, MVT::i8));
14570 // If we have AVX, we can use a variable vector select (VBLENDV) instead
14571 // of 3 logic instructions for size savings and potentially speed.
14572 // Unfortunately, there is no scalar form of VBLENDV.
14574 // If either operand is a constant, don't try this. We can expect to
14575 // optimize away at least one of the logic instructions later in that
14576 // case, so that sequence would be faster than a variable blend.
14578 // BLENDV was introduced with SSE 4.1, but the 2 register form implicitly
14579 // uses XMM0 as the selection register. That may need just as many
14580 // instructions as the AND/ANDN/OR sequence due to register moves, so
14583 if (Subtarget->hasAVX() &&
14584 !isa<ConstantFPSDNode>(Op1) && !isa<ConstantFPSDNode>(Op2)) {
14586 // Convert to vectors, do a VSELECT, and convert back to scalar.
14587 // All of the conversions should be optimized away.
14589 EVT VecVT = VT == MVT::f32 ? MVT::v4f32 : MVT::v2f64;
14590 SDValue VOp1 = DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, VecVT, Op1);
14591 SDValue VOp2 = DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, VecVT, Op2);
14592 SDValue VCmp = DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, VecVT, Cmp);
14594 EVT VCmpVT = VT == MVT::f32 ? MVT::v4i32 : MVT::v2i64;
14595 VCmp = DAG.getBitcast(VCmpVT, VCmp);
14597 SDValue VSel = DAG.getNode(ISD::VSELECT, DL, VecVT, VCmp, VOp1, VOp2);
14599 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, VT,
14600 VSel, DAG.getIntPtrConstant(0, DL));
14602 SDValue AndN = DAG.getNode(X86ISD::FANDN, DL, VT, Cmp, Op2);
14603 SDValue And = DAG.getNode(X86ISD::FAND, DL, VT, Cmp, Op1);
14604 return DAG.getNode(X86ISD::FOR, DL, VT, AndN, And);
14608 if (VT.isVector() && VT.getScalarType() == MVT::i1) {
14610 if (ISD::isBuildVectorOfConstantSDNodes(Op1.getNode()))
14611 Op1Scalar = ConvertI1VectorToInteger(Op1, DAG);
14612 else if (Op1.getOpcode() == ISD::BITCAST && Op1.getOperand(0))
14613 Op1Scalar = Op1.getOperand(0);
14615 if (ISD::isBuildVectorOfConstantSDNodes(Op2.getNode()))
14616 Op2Scalar = ConvertI1VectorToInteger(Op2, DAG);
14617 else if (Op2.getOpcode() == ISD::BITCAST && Op2.getOperand(0))
14618 Op2Scalar = Op2.getOperand(0);
14619 if (Op1Scalar.getNode() && Op2Scalar.getNode()) {
14620 SDValue newSelect = DAG.getNode(ISD::SELECT, DL,
14621 Op1Scalar.getValueType(),
14622 Cond, Op1Scalar, Op2Scalar);
14623 if (newSelect.getValueSizeInBits() == VT.getSizeInBits())
14624 return DAG.getBitcast(VT, newSelect);
14625 SDValue ExtVec = DAG.getBitcast(MVT::v8i1, newSelect);
14626 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, VT, ExtVec,
14627 DAG.getIntPtrConstant(0, DL));
14631 if (VT == MVT::v4i1 || VT == MVT::v2i1) {
14632 SDValue zeroConst = DAG.getIntPtrConstant(0, DL);
14633 Op1 = DAG.getNode(ISD::INSERT_SUBVECTOR, DL, MVT::v8i1,
14634 DAG.getUNDEF(MVT::v8i1), Op1, zeroConst);
14635 Op2 = DAG.getNode(ISD::INSERT_SUBVECTOR, DL, MVT::v8i1,
14636 DAG.getUNDEF(MVT::v8i1), Op2, zeroConst);
14637 SDValue newSelect = DAG.getNode(ISD::SELECT, DL, MVT::v8i1,
14639 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, VT, newSelect, zeroConst);
14642 if (Cond.getOpcode() == ISD::SETCC) {
14643 SDValue NewCond = LowerSETCC(Cond, DAG);
14644 if (NewCond.getNode())
14648 // (select (x == 0), -1, y) -> (sign_bit (x - 1)) | y
14649 // (select (x == 0), y, -1) -> ~(sign_bit (x - 1)) | y
14650 // (select (x != 0), y, -1) -> (sign_bit (x - 1)) | y
14651 // (select (x != 0), -1, y) -> ~(sign_bit (x - 1)) | y
14652 if (Cond.getOpcode() == X86ISD::SETCC &&
14653 Cond.getOperand(1).getOpcode() == X86ISD::CMP &&
14654 isZero(Cond.getOperand(1).getOperand(1))) {
14655 SDValue Cmp = Cond.getOperand(1);
14657 unsigned CondCode =cast<ConstantSDNode>(Cond.getOperand(0))->getZExtValue();
14659 if ((isAllOnes(Op1) || isAllOnes(Op2)) &&
14660 (CondCode == X86::COND_E || CondCode == X86::COND_NE)) {
14661 SDValue Y = isAllOnes(Op2) ? Op1 : Op2;
14663 SDValue CmpOp0 = Cmp.getOperand(0);
14664 // Apply further optimizations for special cases
14665 // (select (x != 0), -1, 0) -> neg & sbb
14666 // (select (x == 0), 0, -1) -> neg & sbb
14667 if (ConstantSDNode *YC = dyn_cast<ConstantSDNode>(Y))
14668 if (YC->isNullValue() &&
14669 (isAllOnes(Op1) == (CondCode == X86::COND_NE))) {
14670 SDVTList VTs = DAG.getVTList(CmpOp0.getValueType(), MVT::i32);
14671 SDValue Neg = DAG.getNode(X86ISD::SUB, DL, VTs,
14672 DAG.getConstant(0, DL,
14673 CmpOp0.getValueType()),
14675 SDValue Res = DAG.getNode(X86ISD::SETCC_CARRY, DL, Op.getValueType(),
14676 DAG.getConstant(X86::COND_B, DL, MVT::i8),
14677 SDValue(Neg.getNode(), 1));
14681 Cmp = DAG.getNode(X86ISD::CMP, DL, MVT::i32,
14682 CmpOp0, DAG.getConstant(1, DL, CmpOp0.getValueType()));
14683 Cmp = ConvertCmpIfNecessary(Cmp, DAG);
14685 SDValue Res = // Res = 0 or -1.
14686 DAG.getNode(X86ISD::SETCC_CARRY, DL, Op.getValueType(),
14687 DAG.getConstant(X86::COND_B, DL, MVT::i8), Cmp);
14689 if (isAllOnes(Op1) != (CondCode == X86::COND_E))
14690 Res = DAG.getNOT(DL, Res, Res.getValueType());
14692 ConstantSDNode *N2C = dyn_cast<ConstantSDNode>(Op2);
14693 if (!N2C || !N2C->isNullValue())
14694 Res = DAG.getNode(ISD::OR, DL, Res.getValueType(), Res, Y);
14699 // Look past (and (setcc_carry (cmp ...)), 1).
14700 if (Cond.getOpcode() == ISD::AND &&
14701 Cond.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY) {
14702 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Cond.getOperand(1));
14703 if (C && C->getAPIntValue() == 1)
14704 Cond = Cond.getOperand(0);
14707 // If condition flag is set by a X86ISD::CMP, then use it as the condition
14708 // setting operand in place of the X86ISD::SETCC.
14709 unsigned CondOpcode = Cond.getOpcode();
14710 if (CondOpcode == X86ISD::SETCC ||
14711 CondOpcode == X86ISD::SETCC_CARRY) {
14712 CC = Cond.getOperand(0);
14714 SDValue Cmp = Cond.getOperand(1);
14715 unsigned Opc = Cmp.getOpcode();
14716 MVT VT = Op.getSimpleValueType();
14718 bool IllegalFPCMov = false;
14719 if (VT.isFloatingPoint() && !VT.isVector() &&
14720 !isScalarFPTypeInSSEReg(VT)) // FPStack?
14721 IllegalFPCMov = !hasFPCMov(cast<ConstantSDNode>(CC)->getSExtValue());
14723 if ((isX86LogicalCmp(Cmp) && !IllegalFPCMov) ||
14724 Opc == X86ISD::BT) { // FIXME
14728 } else if (CondOpcode == ISD::USUBO || CondOpcode == ISD::SSUBO ||
14729 CondOpcode == ISD::UADDO || CondOpcode == ISD::SADDO ||
14730 ((CondOpcode == ISD::UMULO || CondOpcode == ISD::SMULO) &&
14731 Cond.getOperand(0).getValueType() != MVT::i8)) {
14732 SDValue LHS = Cond.getOperand(0);
14733 SDValue RHS = Cond.getOperand(1);
14734 unsigned X86Opcode;
14737 switch (CondOpcode) {
14738 case ISD::UADDO: X86Opcode = X86ISD::ADD; X86Cond = X86::COND_B; break;
14739 case ISD::SADDO: X86Opcode = X86ISD::ADD; X86Cond = X86::COND_O; break;
14740 case ISD::USUBO: X86Opcode = X86ISD::SUB; X86Cond = X86::COND_B; break;
14741 case ISD::SSUBO: X86Opcode = X86ISD::SUB; X86Cond = X86::COND_O; break;
14742 case ISD::UMULO: X86Opcode = X86ISD::UMUL; X86Cond = X86::COND_O; break;
14743 case ISD::SMULO: X86Opcode = X86ISD::SMUL; X86Cond = X86::COND_O; break;
14744 default: llvm_unreachable("unexpected overflowing operator");
14746 if (CondOpcode == ISD::UMULO)
14747 VTs = DAG.getVTList(LHS.getValueType(), LHS.getValueType(),
14750 VTs = DAG.getVTList(LHS.getValueType(), MVT::i32);
14752 SDValue X86Op = DAG.getNode(X86Opcode, DL, VTs, LHS, RHS);
14754 if (CondOpcode == ISD::UMULO)
14755 Cond = X86Op.getValue(2);
14757 Cond = X86Op.getValue(1);
14759 CC = DAG.getConstant(X86Cond, DL, MVT::i8);
14764 // Look past the truncate if the high bits are known zero.
14765 if (isTruncWithZeroHighBitsInput(Cond, DAG))
14766 Cond = Cond.getOperand(0);
14768 // We know the result of AND is compared against zero. Try to match
14770 if (Cond.getOpcode() == ISD::AND && Cond.hasOneUse()) {
14771 SDValue NewSetCC = LowerToBT(Cond, ISD::SETNE, DL, DAG);
14772 if (NewSetCC.getNode()) {
14773 CC = NewSetCC.getOperand(0);
14774 Cond = NewSetCC.getOperand(1);
14781 CC = DAG.getConstant(X86::COND_NE, DL, MVT::i8);
14782 Cond = EmitTest(Cond, X86::COND_NE, DL, DAG);
14785 // a < b ? -1 : 0 -> RES = ~setcc_carry
14786 // a < b ? 0 : -1 -> RES = setcc_carry
14787 // a >= b ? -1 : 0 -> RES = setcc_carry
14788 // a >= b ? 0 : -1 -> RES = ~setcc_carry
14789 if (Cond.getOpcode() == X86ISD::SUB) {
14790 Cond = ConvertCmpIfNecessary(Cond, DAG);
14791 unsigned CondCode = cast<ConstantSDNode>(CC)->getZExtValue();
14793 if ((CondCode == X86::COND_AE || CondCode == X86::COND_B) &&
14794 (isAllOnes(Op1) || isAllOnes(Op2)) && (isZero(Op1) || isZero(Op2))) {
14795 SDValue Res = DAG.getNode(X86ISD::SETCC_CARRY, DL, Op.getValueType(),
14796 DAG.getConstant(X86::COND_B, DL, MVT::i8),
14798 if (isAllOnes(Op1) != (CondCode == X86::COND_B))
14799 return DAG.getNOT(DL, Res, Res.getValueType());
14804 // X86 doesn't have an i8 cmov. If both operands are the result of a truncate
14805 // widen the cmov and push the truncate through. This avoids introducing a new
14806 // branch during isel and doesn't add any extensions.
14807 if (Op.getValueType() == MVT::i8 &&
14808 Op1.getOpcode() == ISD::TRUNCATE && Op2.getOpcode() == ISD::TRUNCATE) {
14809 SDValue T1 = Op1.getOperand(0), T2 = Op2.getOperand(0);
14810 if (T1.getValueType() == T2.getValueType() &&
14811 // Blacklist CopyFromReg to avoid partial register stalls.
14812 T1.getOpcode() != ISD::CopyFromReg && T2.getOpcode()!=ISD::CopyFromReg){
14813 SDVTList VTs = DAG.getVTList(T1.getValueType(), MVT::Glue);
14814 SDValue Cmov = DAG.getNode(X86ISD::CMOV, DL, VTs, T2, T1, CC, Cond);
14815 return DAG.getNode(ISD::TRUNCATE, DL, Op.getValueType(), Cmov);
14819 // X86ISD::CMOV means set the result (which is operand 1) to the RHS if
14820 // condition is true.
14821 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::Glue);
14822 SDValue Ops[] = { Op2, Op1, CC, Cond };
14823 return DAG.getNode(X86ISD::CMOV, DL, VTs, Ops);
14826 static SDValue LowerSIGN_EXTEND_AVX512(SDValue Op,
14827 const X86Subtarget *Subtarget,
14828 SelectionDAG &DAG) {
14829 MVT VT = Op->getSimpleValueType(0);
14830 SDValue In = Op->getOperand(0);
14831 MVT InVT = In.getSimpleValueType();
14832 MVT VTElt = VT.getVectorElementType();
14833 MVT InVTElt = InVT.getVectorElementType();
14837 if ((InVTElt == MVT::i1) &&
14838 (((Subtarget->hasBWI() && Subtarget->hasVLX() &&
14839 VT.getSizeInBits() <= 256 && VTElt.getSizeInBits() <= 16)) ||
14841 ((Subtarget->hasBWI() && VT.is512BitVector() &&
14842 VTElt.getSizeInBits() <= 16)) ||
14844 ((Subtarget->hasDQI() && Subtarget->hasVLX() &&
14845 VT.getSizeInBits() <= 256 && VTElt.getSizeInBits() >= 32)) ||
14847 ((Subtarget->hasDQI() && VT.is512BitVector() &&
14848 VTElt.getSizeInBits() >= 32))))
14849 return DAG.getNode(X86ISD::VSEXT, dl, VT, In);
14851 unsigned int NumElts = VT.getVectorNumElements();
14853 if (NumElts != 8 && NumElts != 16 && !Subtarget->hasBWI())
14856 if (VT.is512BitVector() && InVT.getVectorElementType() != MVT::i1) {
14857 if (In.getOpcode() == X86ISD::VSEXT || In.getOpcode() == X86ISD::VZEXT)
14858 return DAG.getNode(In.getOpcode(), dl, VT, In.getOperand(0));
14859 return DAG.getNode(X86ISD::VSEXT, dl, VT, In);
14862 assert (InVT.getVectorElementType() == MVT::i1 && "Unexpected vector type");
14863 MVT ExtVT = NumElts == 8 ? MVT::v8i64 : MVT::v16i32;
14865 DAG.getConstant(APInt::getAllOnesValue(ExtVT.getScalarSizeInBits()), dl,
14868 DAG.getConstant(APInt::getNullValue(ExtVT.getScalarSizeInBits()), dl, ExtVT);
14870 SDValue V = DAG.getNode(ISD::VSELECT, dl, ExtVT, In, NegOne, Zero);
14871 if (VT.is512BitVector())
14873 return DAG.getNode(X86ISD::VTRUNC, dl, VT, V);
14876 static SDValue LowerSIGN_EXTEND_VECTOR_INREG(SDValue Op,
14877 const X86Subtarget *Subtarget,
14878 SelectionDAG &DAG) {
14879 SDValue In = Op->getOperand(0);
14880 MVT VT = Op->getSimpleValueType(0);
14881 MVT InVT = In.getSimpleValueType();
14882 assert(VT.getSizeInBits() == InVT.getSizeInBits());
14884 MVT InSVT = InVT.getScalarType();
14885 assert(VT.getScalarType().getScalarSizeInBits() > InSVT.getScalarSizeInBits());
14887 if (VT != MVT::v2i64 && VT != MVT::v4i32 && VT != MVT::v8i16)
14889 if (InSVT != MVT::i32 && InSVT != MVT::i16 && InSVT != MVT::i8)
14894 // SSE41 targets can use the pmovsx* instructions directly.
14895 if (Subtarget->hasSSE41())
14896 return DAG.getNode(X86ISD::VSEXT, dl, VT, In);
14898 // pre-SSE41 targets unpack lower lanes and then sign-extend using SRAI.
14902 // As SRAI is only available on i16/i32 types, we expand only up to i32
14903 // and handle i64 separately.
14904 while (CurrVT != VT && CurrVT.getScalarType() != MVT::i32) {
14905 Curr = DAG.getNode(X86ISD::UNPCKL, dl, CurrVT, DAG.getUNDEF(CurrVT), Curr);
14906 MVT CurrSVT = MVT::getIntegerVT(CurrVT.getScalarSizeInBits() * 2);
14907 CurrVT = MVT::getVectorVT(CurrSVT, CurrVT.getVectorNumElements() / 2);
14908 Curr = DAG.getBitcast(CurrVT, Curr);
14911 SDValue SignExt = Curr;
14912 if (CurrVT != InVT) {
14913 unsigned SignExtShift =
14914 CurrVT.getScalarSizeInBits() - InSVT.getScalarSizeInBits();
14915 SignExt = DAG.getNode(X86ISD::VSRAI, dl, CurrVT, Curr,
14916 DAG.getConstant(SignExtShift, dl, MVT::i8));
14922 if (VT == MVT::v2i64 && CurrVT == MVT::v4i32) {
14923 SDValue Sign = DAG.getNode(X86ISD::VSRAI, dl, CurrVT, Curr,
14924 DAG.getConstant(31, dl, MVT::i8));
14925 SDValue Ext = DAG.getVectorShuffle(CurrVT, dl, SignExt, Sign, {0, 4, 1, 5});
14926 return DAG.getBitcast(VT, Ext);
14932 static SDValue LowerSIGN_EXTEND(SDValue Op, const X86Subtarget *Subtarget,
14933 SelectionDAG &DAG) {
14934 MVT VT = Op->getSimpleValueType(0);
14935 SDValue In = Op->getOperand(0);
14936 MVT InVT = In.getSimpleValueType();
14939 if (VT.is512BitVector() || InVT.getVectorElementType() == MVT::i1)
14940 return LowerSIGN_EXTEND_AVX512(Op, Subtarget, DAG);
14942 if ((VT != MVT::v4i64 || InVT != MVT::v4i32) &&
14943 (VT != MVT::v8i32 || InVT != MVT::v8i16) &&
14944 (VT != MVT::v16i16 || InVT != MVT::v16i8))
14947 if (Subtarget->hasInt256())
14948 return DAG.getNode(X86ISD::VSEXT, dl, VT, In);
14950 // Optimize vectors in AVX mode
14951 // Sign extend v8i16 to v8i32 and
14954 // Divide input vector into two parts
14955 // for v4i32 the shuffle mask will be { 0, 1, -1, -1} {2, 3, -1, -1}
14956 // use vpmovsx instruction to extend v4i32 -> v2i64; v8i16 -> v4i32
14957 // concat the vectors to original VT
14959 unsigned NumElems = InVT.getVectorNumElements();
14960 SDValue Undef = DAG.getUNDEF(InVT);
14962 SmallVector<int,8> ShufMask1(NumElems, -1);
14963 for (unsigned i = 0; i != NumElems/2; ++i)
14966 SDValue OpLo = DAG.getVectorShuffle(InVT, dl, In, Undef, &ShufMask1[0]);
14968 SmallVector<int,8> ShufMask2(NumElems, -1);
14969 for (unsigned i = 0; i != NumElems/2; ++i)
14970 ShufMask2[i] = i + NumElems/2;
14972 SDValue OpHi = DAG.getVectorShuffle(InVT, dl, In, Undef, &ShufMask2[0]);
14974 MVT HalfVT = MVT::getVectorVT(VT.getScalarType(),
14975 VT.getVectorNumElements()/2);
14977 OpLo = DAG.getNode(X86ISD::VSEXT, dl, HalfVT, OpLo);
14978 OpHi = DAG.getNode(X86ISD::VSEXT, dl, HalfVT, OpHi);
14980 return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT, OpLo, OpHi);
14983 // Lower vector extended loads using a shuffle. If SSSE3 is not available we
14984 // may emit an illegal shuffle but the expansion is still better than scalar
14985 // code. We generate X86ISD::VSEXT for SEXTLOADs if it's available, otherwise
14986 // we'll emit a shuffle and a arithmetic shift.
14987 // FIXME: Is the expansion actually better than scalar code? It doesn't seem so.
14988 // TODO: It is possible to support ZExt by zeroing the undef values during
14989 // the shuffle phase or after the shuffle.
14990 static SDValue LowerExtendedLoad(SDValue Op, const X86Subtarget *Subtarget,
14991 SelectionDAG &DAG) {
14992 MVT RegVT = Op.getSimpleValueType();
14993 assert(RegVT.isVector() && "We only custom lower vector sext loads.");
14994 assert(RegVT.isInteger() &&
14995 "We only custom lower integer vector sext loads.");
14997 // Nothing useful we can do without SSE2 shuffles.
14998 assert(Subtarget->hasSSE2() && "We only custom lower sext loads with SSE2.");
15000 LoadSDNode *Ld = cast<LoadSDNode>(Op.getNode());
15002 EVT MemVT = Ld->getMemoryVT();
15003 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
15004 unsigned RegSz = RegVT.getSizeInBits();
15006 ISD::LoadExtType Ext = Ld->getExtensionType();
15008 assert((Ext == ISD::EXTLOAD || Ext == ISD::SEXTLOAD)
15009 && "Only anyext and sext are currently implemented.");
15010 assert(MemVT != RegVT && "Cannot extend to the same type");
15011 assert(MemVT.isVector() && "Must load a vector from memory");
15013 unsigned NumElems = RegVT.getVectorNumElements();
15014 unsigned MemSz = MemVT.getSizeInBits();
15015 assert(RegSz > MemSz && "Register size must be greater than the mem size");
15017 if (Ext == ISD::SEXTLOAD && RegSz == 256 && !Subtarget->hasInt256()) {
15018 // The only way in which we have a legal 256-bit vector result but not the
15019 // integer 256-bit operations needed to directly lower a sextload is if we
15020 // have AVX1 but not AVX2. In that case, we can always emit a sextload to
15021 // a 128-bit vector and a normal sign_extend to 256-bits that should get
15022 // correctly legalized. We do this late to allow the canonical form of
15023 // sextload to persist throughout the rest of the DAG combiner -- it wants
15024 // to fold together any extensions it can, and so will fuse a sign_extend
15025 // of an sextload into a sextload targeting a wider value.
15027 if (MemSz == 128) {
15028 // Just switch this to a normal load.
15029 assert(TLI.isTypeLegal(MemVT) && "If the memory type is a 128-bit type, "
15030 "it must be a legal 128-bit vector "
15032 Load = DAG.getLoad(MemVT, dl, Ld->getChain(), Ld->getBasePtr(),
15033 Ld->getPointerInfo(), Ld->isVolatile(), Ld->isNonTemporal(),
15034 Ld->isInvariant(), Ld->getAlignment());
15036 assert(MemSz < 128 &&
15037 "Can't extend a type wider than 128 bits to a 256 bit vector!");
15038 // Do an sext load to a 128-bit vector type. We want to use the same
15039 // number of elements, but elements half as wide. This will end up being
15040 // recursively lowered by this routine, but will succeed as we definitely
15041 // have all the necessary features if we're using AVX1.
15043 EVT::getIntegerVT(*DAG.getContext(), RegVT.getScalarSizeInBits() / 2);
15044 EVT HalfVecVT = EVT::getVectorVT(*DAG.getContext(), HalfEltVT, NumElems);
15046 DAG.getExtLoad(Ext, dl, HalfVecVT, Ld->getChain(), Ld->getBasePtr(),
15047 Ld->getPointerInfo(), MemVT, Ld->isVolatile(),
15048 Ld->isNonTemporal(), Ld->isInvariant(),
15049 Ld->getAlignment());
15052 // Replace chain users with the new chain.
15053 assert(Load->getNumValues() == 2 && "Loads must carry a chain!");
15054 DAG.ReplaceAllUsesOfValueWith(SDValue(Ld, 1), Load.getValue(1));
15056 // Finally, do a normal sign-extend to the desired register.
15057 return DAG.getSExtOrTrunc(Load, dl, RegVT);
15060 // All sizes must be a power of two.
15061 assert(isPowerOf2_32(RegSz * MemSz * NumElems) &&
15062 "Non-power-of-two elements are not custom lowered!");
15064 // Attempt to load the original value using scalar loads.
15065 // Find the largest scalar type that divides the total loaded size.
15066 MVT SclrLoadTy = MVT::i8;
15067 for (MVT Tp : MVT::integer_valuetypes()) {
15068 if (TLI.isTypeLegal(Tp) && ((MemSz % Tp.getSizeInBits()) == 0)) {
15073 // On 32bit systems, we can't save 64bit integers. Try bitcasting to F64.
15074 if (TLI.isTypeLegal(MVT::f64) && SclrLoadTy.getSizeInBits() < 64 &&
15076 SclrLoadTy = MVT::f64;
15078 // Calculate the number of scalar loads that we need to perform
15079 // in order to load our vector from memory.
15080 unsigned NumLoads = MemSz / SclrLoadTy.getSizeInBits();
15082 assert((Ext != ISD::SEXTLOAD || NumLoads == 1) &&
15083 "Can only lower sext loads with a single scalar load!");
15085 unsigned loadRegZize = RegSz;
15086 if (Ext == ISD::SEXTLOAD && RegSz >= 256)
15089 // Represent our vector as a sequence of elements which are the
15090 // largest scalar that we can load.
15091 EVT LoadUnitVecVT = EVT::getVectorVT(
15092 *DAG.getContext(), SclrLoadTy, loadRegZize / SclrLoadTy.getSizeInBits());
15094 // Represent the data using the same element type that is stored in
15095 // memory. In practice, we ''widen'' MemVT.
15097 EVT::getVectorVT(*DAG.getContext(), MemVT.getScalarType(),
15098 loadRegZize / MemVT.getScalarType().getSizeInBits());
15100 assert(WideVecVT.getSizeInBits() == LoadUnitVecVT.getSizeInBits() &&
15101 "Invalid vector type");
15103 // We can't shuffle using an illegal type.
15104 assert(TLI.isTypeLegal(WideVecVT) &&
15105 "We only lower types that form legal widened vector types");
15107 SmallVector<SDValue, 8> Chains;
15108 SDValue Ptr = Ld->getBasePtr();
15109 SDValue Increment = DAG.getConstant(SclrLoadTy.getSizeInBits() / 8, dl,
15110 TLI.getPointerTy(DAG.getDataLayout()));
15111 SDValue Res = DAG.getUNDEF(LoadUnitVecVT);
15113 for (unsigned i = 0; i < NumLoads; ++i) {
15114 // Perform a single load.
15115 SDValue ScalarLoad =
15116 DAG.getLoad(SclrLoadTy, dl, Ld->getChain(), Ptr, Ld->getPointerInfo(),
15117 Ld->isVolatile(), Ld->isNonTemporal(), Ld->isInvariant(),
15118 Ld->getAlignment());
15119 Chains.push_back(ScalarLoad.getValue(1));
15120 // Create the first element type using SCALAR_TO_VECTOR in order to avoid
15121 // another round of DAGCombining.
15123 Res = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, LoadUnitVecVT, ScalarLoad);
15125 Res = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, LoadUnitVecVT, Res,
15126 ScalarLoad, DAG.getIntPtrConstant(i, dl));
15128 Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr, Increment);
15131 SDValue TF = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Chains);
15133 // Bitcast the loaded value to a vector of the original element type, in
15134 // the size of the target vector type.
15135 SDValue SlicedVec = DAG.getBitcast(WideVecVT, Res);
15136 unsigned SizeRatio = RegSz / MemSz;
15138 if (Ext == ISD::SEXTLOAD) {
15139 // If we have SSE4.1, we can directly emit a VSEXT node.
15140 if (Subtarget->hasSSE41()) {
15141 SDValue Sext = DAG.getNode(X86ISD::VSEXT, dl, RegVT, SlicedVec);
15142 DAG.ReplaceAllUsesOfValueWith(SDValue(Ld, 1), TF);
15146 // Otherwise we'll use SIGN_EXTEND_VECTOR_INREG to sign extend the lowest
15148 assert(TLI.isOperationLegalOrCustom(ISD::SIGN_EXTEND_VECTOR_INREG, RegVT) &&
15149 "We can't implement a sext load without SIGN_EXTEND_VECTOR_INREG!");
15151 SDValue Shuff = DAG.getSignExtendVectorInReg(SlicedVec, dl, RegVT);
15152 DAG.ReplaceAllUsesOfValueWith(SDValue(Ld, 1), TF);
15156 // Redistribute the loaded elements into the different locations.
15157 SmallVector<int, 16> ShuffleVec(NumElems * SizeRatio, -1);
15158 for (unsigned i = 0; i != NumElems; ++i)
15159 ShuffleVec[i * SizeRatio] = i;
15161 SDValue Shuff = DAG.getVectorShuffle(WideVecVT, dl, SlicedVec,
15162 DAG.getUNDEF(WideVecVT), &ShuffleVec[0]);
15164 // Bitcast to the requested type.
15165 Shuff = DAG.getBitcast(RegVT, Shuff);
15166 DAG.ReplaceAllUsesOfValueWith(SDValue(Ld, 1), TF);
15170 // isAndOrOfSingleUseSetCCs - Return true if node is an ISD::AND or
15171 // ISD::OR of two X86ISD::SETCC nodes each of which has no other use apart
15172 // from the AND / OR.
15173 static bool isAndOrOfSetCCs(SDValue Op, unsigned &Opc) {
15174 Opc = Op.getOpcode();
15175 if (Opc != ISD::OR && Opc != ISD::AND)
15177 return (Op.getOperand(0).getOpcode() == X86ISD::SETCC &&
15178 Op.getOperand(0).hasOneUse() &&
15179 Op.getOperand(1).getOpcode() == X86ISD::SETCC &&
15180 Op.getOperand(1).hasOneUse());
15183 // isXor1OfSetCC - Return true if node is an ISD::XOR of a X86ISD::SETCC and
15184 // 1 and that the SETCC node has a single use.
15185 static bool isXor1OfSetCC(SDValue Op) {
15186 if (Op.getOpcode() != ISD::XOR)
15188 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(Op.getOperand(1));
15189 if (N1C && N1C->getAPIntValue() == 1) {
15190 return Op.getOperand(0).getOpcode() == X86ISD::SETCC &&
15191 Op.getOperand(0).hasOneUse();
15196 SDValue X86TargetLowering::LowerBRCOND(SDValue Op, SelectionDAG &DAG) const {
15197 bool addTest = true;
15198 SDValue Chain = Op.getOperand(0);
15199 SDValue Cond = Op.getOperand(1);
15200 SDValue Dest = Op.getOperand(2);
15203 bool Inverted = false;
15205 if (Cond.getOpcode() == ISD::SETCC) {
15206 // Check for setcc([su]{add,sub,mul}o == 0).
15207 if (cast<CondCodeSDNode>(Cond.getOperand(2))->get() == ISD::SETEQ &&
15208 isa<ConstantSDNode>(Cond.getOperand(1)) &&
15209 cast<ConstantSDNode>(Cond.getOperand(1))->isNullValue() &&
15210 Cond.getOperand(0).getResNo() == 1 &&
15211 (Cond.getOperand(0).getOpcode() == ISD::SADDO ||
15212 Cond.getOperand(0).getOpcode() == ISD::UADDO ||
15213 Cond.getOperand(0).getOpcode() == ISD::SSUBO ||
15214 Cond.getOperand(0).getOpcode() == ISD::USUBO ||
15215 Cond.getOperand(0).getOpcode() == ISD::SMULO ||
15216 Cond.getOperand(0).getOpcode() == ISD::UMULO)) {
15218 Cond = Cond.getOperand(0);
15220 SDValue NewCond = LowerSETCC(Cond, DAG);
15221 if (NewCond.getNode())
15226 // FIXME: LowerXALUO doesn't handle these!!
15227 else if (Cond.getOpcode() == X86ISD::ADD ||
15228 Cond.getOpcode() == X86ISD::SUB ||
15229 Cond.getOpcode() == X86ISD::SMUL ||
15230 Cond.getOpcode() == X86ISD::UMUL)
15231 Cond = LowerXALUO(Cond, DAG);
15234 // Look pass (and (setcc_carry (cmp ...)), 1).
15235 if (Cond.getOpcode() == ISD::AND &&
15236 Cond.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY) {
15237 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Cond.getOperand(1));
15238 if (C && C->getAPIntValue() == 1)
15239 Cond = Cond.getOperand(0);
15242 // If condition flag is set by a X86ISD::CMP, then use it as the condition
15243 // setting operand in place of the X86ISD::SETCC.
15244 unsigned CondOpcode = Cond.getOpcode();
15245 if (CondOpcode == X86ISD::SETCC ||
15246 CondOpcode == X86ISD::SETCC_CARRY) {
15247 CC = Cond.getOperand(0);
15249 SDValue Cmp = Cond.getOperand(1);
15250 unsigned Opc = Cmp.getOpcode();
15251 // FIXME: WHY THE SPECIAL CASING OF LogicalCmp??
15252 if (isX86LogicalCmp(Cmp) || Opc == X86ISD::BT) {
15256 switch (cast<ConstantSDNode>(CC)->getZExtValue()) {
15260 // These can only come from an arithmetic instruction with overflow,
15261 // e.g. SADDO, UADDO.
15262 Cond = Cond.getNode()->getOperand(1);
15268 CondOpcode = Cond.getOpcode();
15269 if (CondOpcode == ISD::UADDO || CondOpcode == ISD::SADDO ||
15270 CondOpcode == ISD::USUBO || CondOpcode == ISD::SSUBO ||
15271 ((CondOpcode == ISD::UMULO || CondOpcode == ISD::SMULO) &&
15272 Cond.getOperand(0).getValueType() != MVT::i8)) {
15273 SDValue LHS = Cond.getOperand(0);
15274 SDValue RHS = Cond.getOperand(1);
15275 unsigned X86Opcode;
15278 // Keep this in sync with LowerXALUO, otherwise we might create redundant
15279 // instructions that can't be removed afterwards (i.e. X86ISD::ADD and
15281 switch (CondOpcode) {
15282 case ISD::UADDO: X86Opcode = X86ISD::ADD; X86Cond = X86::COND_B; break;
15284 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(RHS))
15286 X86Opcode = X86ISD::INC; X86Cond = X86::COND_O;
15289 X86Opcode = X86ISD::ADD; X86Cond = X86::COND_O; break;
15290 case ISD::USUBO: X86Opcode = X86ISD::SUB; X86Cond = X86::COND_B; break;
15292 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(RHS))
15294 X86Opcode = X86ISD::DEC; X86Cond = X86::COND_O;
15297 X86Opcode = X86ISD::SUB; X86Cond = X86::COND_O; break;
15298 case ISD::UMULO: X86Opcode = X86ISD::UMUL; X86Cond = X86::COND_O; break;
15299 case ISD::SMULO: X86Opcode = X86ISD::SMUL; X86Cond = X86::COND_O; break;
15300 default: llvm_unreachable("unexpected overflowing operator");
15303 X86Cond = X86::GetOppositeBranchCondition((X86::CondCode)X86Cond);
15304 if (CondOpcode == ISD::UMULO)
15305 VTs = DAG.getVTList(LHS.getValueType(), LHS.getValueType(),
15308 VTs = DAG.getVTList(LHS.getValueType(), MVT::i32);
15310 SDValue X86Op = DAG.getNode(X86Opcode, dl, VTs, LHS, RHS);
15312 if (CondOpcode == ISD::UMULO)
15313 Cond = X86Op.getValue(2);
15315 Cond = X86Op.getValue(1);
15317 CC = DAG.getConstant(X86Cond, dl, MVT::i8);
15321 if (Cond.hasOneUse() && isAndOrOfSetCCs(Cond, CondOpc)) {
15322 SDValue Cmp = Cond.getOperand(0).getOperand(1);
15323 if (CondOpc == ISD::OR) {
15324 // Also, recognize the pattern generated by an FCMP_UNE. We can emit
15325 // two branches instead of an explicit OR instruction with a
15327 if (Cmp == Cond.getOperand(1).getOperand(1) &&
15328 isX86LogicalCmp(Cmp)) {
15329 CC = Cond.getOperand(0).getOperand(0);
15330 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
15331 Chain, Dest, CC, Cmp);
15332 CC = Cond.getOperand(1).getOperand(0);
15336 } else { // ISD::AND
15337 // Also, recognize the pattern generated by an FCMP_OEQ. We can emit
15338 // two branches instead of an explicit AND instruction with a
15339 // separate test. However, we only do this if this block doesn't
15340 // have a fall-through edge, because this requires an explicit
15341 // jmp when the condition is false.
15342 if (Cmp == Cond.getOperand(1).getOperand(1) &&
15343 isX86LogicalCmp(Cmp) &&
15344 Op.getNode()->hasOneUse()) {
15345 X86::CondCode CCode =
15346 (X86::CondCode)Cond.getOperand(0).getConstantOperandVal(0);
15347 CCode = X86::GetOppositeBranchCondition(CCode);
15348 CC = DAG.getConstant(CCode, dl, MVT::i8);
15349 SDNode *User = *Op.getNode()->use_begin();
15350 // Look for an unconditional branch following this conditional branch.
15351 // We need this because we need to reverse the successors in order
15352 // to implement FCMP_OEQ.
15353 if (User->getOpcode() == ISD::BR) {
15354 SDValue FalseBB = User->getOperand(1);
15356 DAG.UpdateNodeOperands(User, User->getOperand(0), Dest);
15357 assert(NewBR == User);
15361 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
15362 Chain, Dest, CC, Cmp);
15363 X86::CondCode CCode =
15364 (X86::CondCode)Cond.getOperand(1).getConstantOperandVal(0);
15365 CCode = X86::GetOppositeBranchCondition(CCode);
15366 CC = DAG.getConstant(CCode, dl, MVT::i8);
15372 } else if (Cond.hasOneUse() && isXor1OfSetCC(Cond)) {
15373 // Recognize for xorb (setcc), 1 patterns. The xor inverts the condition.
15374 // It should be transformed during dag combiner except when the condition
15375 // is set by a arithmetics with overflow node.
15376 X86::CondCode CCode =
15377 (X86::CondCode)Cond.getOperand(0).getConstantOperandVal(0);
15378 CCode = X86::GetOppositeBranchCondition(CCode);
15379 CC = DAG.getConstant(CCode, dl, MVT::i8);
15380 Cond = Cond.getOperand(0).getOperand(1);
15382 } else if (Cond.getOpcode() == ISD::SETCC &&
15383 cast<CondCodeSDNode>(Cond.getOperand(2))->get() == ISD::SETOEQ) {
15384 // For FCMP_OEQ, we can emit
15385 // two branches instead of an explicit AND instruction with a
15386 // separate test. However, we only do this if this block doesn't
15387 // have a fall-through edge, because this requires an explicit
15388 // jmp when the condition is false.
15389 if (Op.getNode()->hasOneUse()) {
15390 SDNode *User = *Op.getNode()->use_begin();
15391 // Look for an unconditional branch following this conditional branch.
15392 // We need this because we need to reverse the successors in order
15393 // to implement FCMP_OEQ.
15394 if (User->getOpcode() == ISD::BR) {
15395 SDValue FalseBB = User->getOperand(1);
15397 DAG.UpdateNodeOperands(User, User->getOperand(0), Dest);
15398 assert(NewBR == User);
15402 SDValue Cmp = DAG.getNode(X86ISD::CMP, dl, MVT::i32,
15403 Cond.getOperand(0), Cond.getOperand(1));
15404 Cmp = ConvertCmpIfNecessary(Cmp, DAG);
15405 CC = DAG.getConstant(X86::COND_NE, dl, MVT::i8);
15406 Chain = DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
15407 Chain, Dest, CC, Cmp);
15408 CC = DAG.getConstant(X86::COND_P, dl, MVT::i8);
15413 } else if (Cond.getOpcode() == ISD::SETCC &&
15414 cast<CondCodeSDNode>(Cond.getOperand(2))->get() == ISD::SETUNE) {
15415 // For FCMP_UNE, we can emit
15416 // two branches instead of an explicit AND instruction with a
15417 // separate test. However, we only do this if this block doesn't
15418 // have a fall-through edge, because this requires an explicit
15419 // jmp when the condition is false.
15420 if (Op.getNode()->hasOneUse()) {
15421 SDNode *User = *Op.getNode()->use_begin();
15422 // Look for an unconditional branch following this conditional branch.
15423 // We need this because we need to reverse the successors in order
15424 // to implement FCMP_UNE.
15425 if (User->getOpcode() == ISD::BR) {
15426 SDValue FalseBB = User->getOperand(1);
15428 DAG.UpdateNodeOperands(User, User->getOperand(0), Dest);
15429 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_NP, dl, MVT::i8);
15448 // Look pass the truncate if the high bits are known zero.
15449 if (isTruncWithZeroHighBitsInput(Cond, DAG))
15450 Cond = Cond.getOperand(0);
15452 // We know the result of AND is compared against zero. Try to match
15454 if (Cond.getOpcode() == ISD::AND && Cond.hasOneUse()) {
15455 SDValue NewSetCC = LowerToBT(Cond, ISD::SETNE, dl, DAG);
15456 if (NewSetCC.getNode()) {
15457 CC = NewSetCC.getOperand(0);
15458 Cond = NewSetCC.getOperand(1);
15465 X86::CondCode X86Cond = Inverted ? X86::COND_E : X86::COND_NE;
15466 CC = DAG.getConstant(X86Cond, dl, MVT::i8);
15467 Cond = EmitTest(Cond, X86Cond, dl, DAG);
15469 Cond = ConvertCmpIfNecessary(Cond, DAG);
15470 return DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
15471 Chain, Dest, CC, Cond);
15474 // Lower dynamic stack allocation to _alloca call for Cygwin/Mingw targets.
15475 // Calls to _alloca are needed to probe the stack when allocating more than 4k
15476 // bytes in one go. Touching the stack at 4K increments is necessary to ensure
15477 // that the guard pages used by the OS virtual memory manager are allocated in
15478 // correct sequence.
15480 X86TargetLowering::LowerDYNAMIC_STACKALLOC(SDValue Op,
15481 SelectionDAG &DAG) const {
15482 MachineFunction &MF = DAG.getMachineFunction();
15483 bool SplitStack = MF.shouldSplitStack();
15484 bool Lower = (Subtarget->isOSWindows() && !Subtarget->isTargetMachO()) ||
15489 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
15490 SDNode* Node = Op.getNode();
15492 unsigned SPReg = TLI.getStackPointerRegisterToSaveRestore();
15493 assert(SPReg && "Target cannot require DYNAMIC_STACKALLOC expansion and"
15494 " not tell us which reg is the stack pointer!");
15495 EVT VT = Node->getValueType(0);
15496 SDValue Tmp1 = SDValue(Node, 0);
15497 SDValue Tmp2 = SDValue(Node, 1);
15498 SDValue Tmp3 = Node->getOperand(2);
15499 SDValue Chain = Tmp1.getOperand(0);
15501 // Chain the dynamic stack allocation so that it doesn't modify the stack
15502 // pointer when other instructions are using the stack.
15503 Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(0, dl, true),
15506 SDValue Size = Tmp2.getOperand(1);
15507 SDValue SP = DAG.getCopyFromReg(Chain, dl, SPReg, VT);
15508 Chain = SP.getValue(1);
15509 unsigned Align = cast<ConstantSDNode>(Tmp3)->getZExtValue();
15510 const TargetFrameLowering &TFI = *Subtarget->getFrameLowering();
15511 unsigned StackAlign = TFI.getStackAlignment();
15512 Tmp1 = DAG.getNode(ISD::SUB, dl, VT, SP, Size); // Value
15513 if (Align > StackAlign)
15514 Tmp1 = DAG.getNode(ISD::AND, dl, VT, Tmp1,
15515 DAG.getConstant(-(uint64_t)Align, dl, VT));
15516 Chain = DAG.getCopyToReg(Chain, dl, SPReg, Tmp1); // Output chain
15518 Tmp2 = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(0, dl, true),
15519 DAG.getIntPtrConstant(0, dl, true), SDValue(),
15522 SDValue Ops[2] = { Tmp1, Tmp2 };
15523 return DAG.getMergeValues(Ops, dl);
15527 SDValue Chain = Op.getOperand(0);
15528 SDValue Size = Op.getOperand(1);
15529 unsigned Align = cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue();
15530 EVT VT = Op.getNode()->getValueType(0);
15532 bool Is64Bit = Subtarget->is64Bit();
15533 MVT SPTy = getPointerTy(DAG.getDataLayout());
15536 MachineRegisterInfo &MRI = MF.getRegInfo();
15539 // The 64 bit implementation of segmented stacks needs to clobber both r10
15540 // r11. This makes it impossible to use it along with nested parameters.
15541 const Function *F = MF.getFunction();
15543 for (Function::const_arg_iterator I = F->arg_begin(), E = F->arg_end();
15545 if (I->hasNestAttr())
15546 report_fatal_error("Cannot use segmented stacks with functions that "
15547 "have nested arguments.");
15550 const TargetRegisterClass *AddrRegClass = getRegClassFor(SPTy);
15551 unsigned Vreg = MRI.createVirtualRegister(AddrRegClass);
15552 Chain = DAG.getCopyToReg(Chain, dl, Vreg, Size);
15553 SDValue Value = DAG.getNode(X86ISD::SEG_ALLOCA, dl, SPTy, Chain,
15554 DAG.getRegister(Vreg, SPTy));
15555 SDValue Ops1[2] = { Value, Chain };
15556 return DAG.getMergeValues(Ops1, dl);
15559 const unsigned Reg = (Subtarget->isTarget64BitLP64() ? X86::RAX : X86::EAX);
15561 Chain = DAG.getCopyToReg(Chain, dl, Reg, Size, Flag);
15562 Flag = Chain.getValue(1);
15563 SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
15565 Chain = DAG.getNode(X86ISD::WIN_ALLOCA, dl, NodeTys, Chain, Flag);
15567 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
15568 unsigned SPReg = RegInfo->getStackRegister();
15569 SDValue SP = DAG.getCopyFromReg(Chain, dl, SPReg, SPTy);
15570 Chain = SP.getValue(1);
15573 SP = DAG.getNode(ISD::AND, dl, VT, SP.getValue(0),
15574 DAG.getConstant(-(uint64_t)Align, dl, VT));
15575 Chain = DAG.getCopyToReg(Chain, dl, SPReg, SP);
15578 SDValue Ops1[2] = { SP, Chain };
15579 return DAG.getMergeValues(Ops1, dl);
15583 SDValue X86TargetLowering::LowerVASTART(SDValue Op, SelectionDAG &DAG) const {
15584 MachineFunction &MF = DAG.getMachineFunction();
15585 auto PtrVT = getPointerTy(MF.getDataLayout());
15586 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
15588 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
15591 if (!Subtarget->is64Bit() ||
15592 Subtarget->isCallingConvWin64(MF.getFunction()->getCallingConv())) {
15593 // vastart just stores the address of the VarArgsFrameIndex slot into the
15594 // memory location argument.
15595 SDValue FR = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(), PtrVT);
15596 return DAG.getStore(Op.getOperand(0), DL, FR, Op.getOperand(1),
15597 MachinePointerInfo(SV), false, false, 0);
15601 // gp_offset (0 - 6 * 8)
15602 // fp_offset (48 - 48 + 8 * 16)
15603 // overflow_arg_area (point to parameters coming in memory).
15605 SmallVector<SDValue, 8> MemOps;
15606 SDValue FIN = Op.getOperand(1);
15608 SDValue Store = DAG.getStore(Op.getOperand(0), DL,
15609 DAG.getConstant(FuncInfo->getVarArgsGPOffset(),
15611 FIN, MachinePointerInfo(SV), false, false, 0);
15612 MemOps.push_back(Store);
15615 FIN = DAG.getNode(ISD::ADD, DL, PtrVT, FIN, DAG.getIntPtrConstant(4, DL));
15616 Store = DAG.getStore(Op.getOperand(0), DL,
15617 DAG.getConstant(FuncInfo->getVarArgsFPOffset(), DL,
15619 FIN, MachinePointerInfo(SV, 4), false, false, 0);
15620 MemOps.push_back(Store);
15622 // Store ptr to overflow_arg_area
15623 FIN = DAG.getNode(ISD::ADD, DL, PtrVT, FIN, DAG.getIntPtrConstant(4, DL));
15624 SDValue OVFIN = DAG.getFrameIndex(FuncInfo->getVarArgsFrameIndex(), PtrVT);
15625 Store = DAG.getStore(Op.getOperand(0), DL, OVFIN, FIN,
15626 MachinePointerInfo(SV, 8),
15628 MemOps.push_back(Store);
15630 // Store ptr to reg_save_area.
15631 FIN = DAG.getNode(ISD::ADD, DL, PtrVT, FIN, DAG.getIntPtrConstant(
15632 Subtarget->isTarget64BitLP64() ? 8 : 4, DL));
15633 SDValue RSFIN = DAG.getFrameIndex(FuncInfo->getRegSaveFrameIndex(), PtrVT);
15634 Store = DAG.getStore(Op.getOperand(0), DL, RSFIN, FIN, MachinePointerInfo(
15635 SV, Subtarget->isTarget64BitLP64() ? 16 : 12), false, false, 0);
15636 MemOps.push_back(Store);
15637 return DAG.getNode(ISD::TokenFactor, DL, MVT::Other, MemOps);
15640 SDValue X86TargetLowering::LowerVAARG(SDValue Op, SelectionDAG &DAG) const {
15641 assert(Subtarget->is64Bit() &&
15642 "LowerVAARG only handles 64-bit va_arg!");
15643 assert(Op.getNode()->getNumOperands() == 4);
15645 MachineFunction &MF = DAG.getMachineFunction();
15646 if (Subtarget->isCallingConvWin64(MF.getFunction()->getCallingConv()))
15647 // The Win64 ABI uses char* instead of a structure.
15648 return DAG.expandVAArg(Op.getNode());
15650 SDValue Chain = Op.getOperand(0);
15651 SDValue SrcPtr = Op.getOperand(1);
15652 const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
15653 unsigned Align = Op.getConstantOperandVal(3);
15656 EVT ArgVT = Op.getNode()->getValueType(0);
15657 Type *ArgTy = ArgVT.getTypeForEVT(*DAG.getContext());
15658 uint32_t ArgSize = DAG.getDataLayout().getTypeAllocSize(ArgTy);
15661 // Decide which area this value should be read from.
15662 // TODO: Implement the AMD64 ABI in its entirety. This simple
15663 // selection mechanism works only for the basic types.
15664 if (ArgVT == MVT::f80) {
15665 llvm_unreachable("va_arg for f80 not yet implemented");
15666 } else if (ArgVT.isFloatingPoint() && ArgSize <= 16 /*bytes*/) {
15667 ArgMode = 2; // Argument passed in XMM register. Use fp_offset.
15668 } else if (ArgVT.isInteger() && ArgSize <= 32 /*bytes*/) {
15669 ArgMode = 1; // Argument passed in GPR64 register(s). Use gp_offset.
15671 llvm_unreachable("Unhandled argument type in LowerVAARG");
15674 if (ArgMode == 2) {
15675 // Sanity Check: Make sure using fp_offset makes sense.
15676 assert(!Subtarget->useSoftFloat() &&
15677 !(MF.getFunction()->hasFnAttribute(Attribute::NoImplicitFloat)) &&
15678 Subtarget->hasSSE1());
15681 // Insert VAARG_64 node into the DAG
15682 // VAARG_64 returns two values: Variable Argument Address, Chain
15683 SDValue InstOps[] = {Chain, SrcPtr, DAG.getConstant(ArgSize, dl, MVT::i32),
15684 DAG.getConstant(ArgMode, dl, MVT::i8),
15685 DAG.getConstant(Align, dl, MVT::i32)};
15686 SDVTList VTs = DAG.getVTList(getPointerTy(DAG.getDataLayout()), MVT::Other);
15687 SDValue VAARG = DAG.getMemIntrinsicNode(X86ISD::VAARG_64, dl,
15688 VTs, InstOps, MVT::i64,
15689 MachinePointerInfo(SV),
15691 /*Volatile=*/false,
15693 /*WriteMem=*/true);
15694 Chain = VAARG.getValue(1);
15696 // Load the next argument and return it
15697 return DAG.getLoad(ArgVT, dl,
15700 MachinePointerInfo(),
15701 false, false, false, 0);
15704 static SDValue LowerVACOPY(SDValue Op, const X86Subtarget *Subtarget,
15705 SelectionDAG &DAG) {
15706 // X86-64 va_list is a struct { i32, i32, i8*, i8* }, except on Windows,
15707 // where a va_list is still an i8*.
15708 assert(Subtarget->is64Bit() && "This code only handles 64-bit va_copy!");
15709 if (Subtarget->isCallingConvWin64(
15710 DAG.getMachineFunction().getFunction()->getCallingConv()))
15711 // Probably a Win64 va_copy.
15712 return DAG.expandVACopy(Op.getNode());
15714 SDValue Chain = Op.getOperand(0);
15715 SDValue DstPtr = Op.getOperand(1);
15716 SDValue SrcPtr = Op.getOperand(2);
15717 const Value *DstSV = cast<SrcValueSDNode>(Op.getOperand(3))->getValue();
15718 const Value *SrcSV = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
15721 return DAG.getMemcpy(Chain, DL, DstPtr, SrcPtr,
15722 DAG.getIntPtrConstant(24, DL), 8, /*isVolatile*/false,
15724 MachinePointerInfo(DstSV), MachinePointerInfo(SrcSV));
15727 // getTargetVShiftByConstNode - Handle vector element shifts where the shift
15728 // amount is a constant. Takes immediate version of shift as input.
15729 static SDValue getTargetVShiftByConstNode(unsigned Opc, SDLoc dl, MVT VT,
15730 SDValue SrcOp, uint64_t ShiftAmt,
15731 SelectionDAG &DAG) {
15732 MVT ElementType = VT.getVectorElementType();
15734 // Fold this packed shift into its first operand if ShiftAmt is 0.
15738 // Check for ShiftAmt >= element width
15739 if (ShiftAmt >= ElementType.getSizeInBits()) {
15740 if (Opc == X86ISD::VSRAI)
15741 ShiftAmt = ElementType.getSizeInBits() - 1;
15743 return DAG.getConstant(0, dl, VT);
15746 assert((Opc == X86ISD::VSHLI || Opc == X86ISD::VSRLI || Opc == X86ISD::VSRAI)
15747 && "Unknown target vector shift-by-constant node");
15749 // Fold this packed vector shift into a build vector if SrcOp is a
15750 // vector of Constants or UNDEFs, and SrcOp valuetype is the same as VT.
15751 if (VT == SrcOp.getSimpleValueType() &&
15752 ISD::isBuildVectorOfConstantSDNodes(SrcOp.getNode())) {
15753 SmallVector<SDValue, 8> Elts;
15754 unsigned NumElts = SrcOp->getNumOperands();
15755 ConstantSDNode *ND;
15758 default: llvm_unreachable(nullptr);
15759 case X86ISD::VSHLI:
15760 for (unsigned i=0; i!=NumElts; ++i) {
15761 SDValue CurrentOp = SrcOp->getOperand(i);
15762 if (CurrentOp->getOpcode() == ISD::UNDEF) {
15763 Elts.push_back(CurrentOp);
15766 ND = cast<ConstantSDNode>(CurrentOp);
15767 const APInt &C = ND->getAPIntValue();
15768 Elts.push_back(DAG.getConstant(C.shl(ShiftAmt), dl, ElementType));
15771 case X86ISD::VSRLI:
15772 for (unsigned i=0; i!=NumElts; ++i) {
15773 SDValue CurrentOp = SrcOp->getOperand(i);
15774 if (CurrentOp->getOpcode() == ISD::UNDEF) {
15775 Elts.push_back(CurrentOp);
15778 ND = cast<ConstantSDNode>(CurrentOp);
15779 const APInt &C = ND->getAPIntValue();
15780 Elts.push_back(DAG.getConstant(C.lshr(ShiftAmt), dl, ElementType));
15783 case X86ISD::VSRAI:
15784 for (unsigned i=0; i!=NumElts; ++i) {
15785 SDValue CurrentOp = SrcOp->getOperand(i);
15786 if (CurrentOp->getOpcode() == ISD::UNDEF) {
15787 Elts.push_back(CurrentOp);
15790 ND = cast<ConstantSDNode>(CurrentOp);
15791 const APInt &C = ND->getAPIntValue();
15792 Elts.push_back(DAG.getConstant(C.ashr(ShiftAmt), dl, ElementType));
15797 return DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Elts);
15800 return DAG.getNode(Opc, dl, VT, SrcOp,
15801 DAG.getConstant(ShiftAmt, dl, MVT::i8));
15804 // getTargetVShiftNode - Handle vector element shifts where the shift amount
15805 // may or may not be a constant. Takes immediate version of shift as input.
15806 static SDValue getTargetVShiftNode(unsigned Opc, SDLoc dl, MVT VT,
15807 SDValue SrcOp, SDValue ShAmt,
15808 SelectionDAG &DAG) {
15809 MVT SVT = ShAmt.getSimpleValueType();
15810 assert((SVT == MVT::i32 || SVT == MVT::i64) && "Unexpected value type!");
15812 // Catch shift-by-constant.
15813 if (ConstantSDNode *CShAmt = dyn_cast<ConstantSDNode>(ShAmt))
15814 return getTargetVShiftByConstNode(Opc, dl, VT, SrcOp,
15815 CShAmt->getZExtValue(), DAG);
15817 // Change opcode to non-immediate version
15819 default: llvm_unreachable("Unknown target vector shift node");
15820 case X86ISD::VSHLI: Opc = X86ISD::VSHL; break;
15821 case X86ISD::VSRLI: Opc = X86ISD::VSRL; break;
15822 case X86ISD::VSRAI: Opc = X86ISD::VSRA; break;
15825 const X86Subtarget &Subtarget =
15826 static_cast<const X86Subtarget &>(DAG.getSubtarget());
15827 if (Subtarget.hasSSE41() && ShAmt.getOpcode() == ISD::ZERO_EXTEND &&
15828 ShAmt.getOperand(0).getSimpleValueType() == MVT::i16) {
15829 // Let the shuffle legalizer expand this shift amount node.
15830 SDValue Op0 = ShAmt.getOperand(0);
15831 Op0 = DAG.getNode(ISD::SCALAR_TO_VECTOR, SDLoc(Op0), MVT::v8i16, Op0);
15832 ShAmt = getShuffleVectorZeroOrUndef(Op0, 0, true, &Subtarget, DAG);
15834 // Need to build a vector containing shift amount.
15835 // SSE/AVX packed shifts only use the lower 64-bit of the shift count.
15836 SmallVector<SDValue, 4> ShOps;
15837 ShOps.push_back(ShAmt);
15838 if (SVT == MVT::i32) {
15839 ShOps.push_back(DAG.getConstant(0, dl, SVT));
15840 ShOps.push_back(DAG.getUNDEF(SVT));
15842 ShOps.push_back(DAG.getUNDEF(SVT));
15844 MVT BVT = SVT == MVT::i32 ? MVT::v4i32 : MVT::v2i64;
15845 ShAmt = DAG.getNode(ISD::BUILD_VECTOR, dl, BVT, ShOps);
15848 // The return type has to be a 128-bit type with the same element
15849 // type as the input type.
15850 MVT EltVT = VT.getVectorElementType();
15851 EVT ShVT = MVT::getVectorVT(EltVT, 128/EltVT.getSizeInBits());
15853 ShAmt = DAG.getBitcast(ShVT, ShAmt);
15854 return DAG.getNode(Opc, dl, VT, SrcOp, ShAmt);
15857 /// \brief Return (and \p Op, \p Mask) for compare instructions or
15858 /// (vselect \p Mask, \p Op, \p PreservedSrc) for others along with the
15859 /// necessary casting or extending for \p Mask when lowering masking intrinsics
15860 static SDValue getVectorMaskingNode(SDValue Op, SDValue Mask,
15861 SDValue PreservedSrc,
15862 const X86Subtarget *Subtarget,
15863 SelectionDAG &DAG) {
15864 EVT VT = Op.getValueType();
15865 EVT MaskVT = EVT::getVectorVT(*DAG.getContext(),
15866 MVT::i1, VT.getVectorNumElements());
15867 SDValue VMask = SDValue();
15868 unsigned OpcodeSelect = ISD::VSELECT;
15871 assert(MaskVT.isSimple() && "invalid mask type");
15873 if (isAllOnes(Mask))
15876 if (MaskVT.bitsGT(Mask.getValueType())) {
15877 EVT newMaskVT = EVT::getIntegerVT(*DAG.getContext(),
15878 MaskVT.getSizeInBits());
15879 VMask = DAG.getBitcast(MaskVT,
15880 DAG.getNode(ISD::ANY_EXTEND, dl, newMaskVT, Mask));
15882 EVT BitcastVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
15883 Mask.getValueType().getSizeInBits());
15884 // In case when MaskVT equals v2i1 or v4i1, low 2 or 4 elements
15885 // are extracted by EXTRACT_SUBVECTOR.
15886 VMask = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MaskVT,
15887 DAG.getBitcast(BitcastVT, Mask),
15888 DAG.getIntPtrConstant(0, dl));
15891 switch (Op.getOpcode()) {
15893 case X86ISD::PCMPEQM:
15894 case X86ISD::PCMPGTM:
15896 case X86ISD::CMPMU:
15897 return DAG.getNode(ISD::AND, dl, VT, Op, VMask);
15898 case X86ISD::VFPCLASS:
15899 return DAG.getNode(ISD::OR, dl, VT, Op, VMask);
15900 case X86ISD::VTRUNC:
15901 case X86ISD::VTRUNCS:
15902 case X86ISD::VTRUNCUS:
15903 // We can't use ISD::VSELECT here because it is not always "Legal"
15904 // for the destination type. For example vpmovqb require only AVX512
15905 // and vselect that can operate on byte element type require BWI
15906 OpcodeSelect = X86ISD::SELECT;
15909 if (PreservedSrc.getOpcode() == ISD::UNDEF)
15910 PreservedSrc = getZeroVector(VT, Subtarget, DAG, dl);
15911 return DAG.getNode(OpcodeSelect, dl, VT, VMask, Op, PreservedSrc);
15914 /// \brief Creates an SDNode for a predicated scalar operation.
15915 /// \returns (X86vselect \p Mask, \p Op, \p PreservedSrc).
15916 /// The mask is coming as MVT::i8 and it should be truncated
15917 /// to MVT::i1 while lowering masking intrinsics.
15918 /// The main difference between ScalarMaskingNode and VectorMaskingNode is using
15919 /// "X86select" instead of "vselect". We just can't create the "vselect" node
15920 /// for a scalar instruction.
15921 static SDValue getScalarMaskingNode(SDValue Op, SDValue Mask,
15922 SDValue PreservedSrc,
15923 const X86Subtarget *Subtarget,
15924 SelectionDAG &DAG) {
15925 if (isAllOnes(Mask))
15928 EVT VT = Op.getValueType();
15930 // The mask should be of type MVT::i1
15931 SDValue IMask = DAG.getNode(ISD::TRUNCATE, dl, MVT::i1, Mask);
15933 if (Op.getOpcode() == X86ISD::FSETCC)
15934 return DAG.getNode(ISD::AND, dl, VT, Op, IMask);
15935 if (Op.getOpcode() == X86ISD::VFPCLASS)
15936 return DAG.getNode(ISD::OR, dl, VT, Op, IMask);
15938 if (PreservedSrc.getOpcode() == ISD::UNDEF)
15939 PreservedSrc = getZeroVector(VT, Subtarget, DAG, dl);
15940 return DAG.getNode(X86ISD::SELECT, dl, VT, IMask, Op, PreservedSrc);
15943 static int getSEHRegistrationNodeSize(const Function *Fn) {
15944 if (!Fn->hasPersonalityFn())
15945 report_fatal_error(
15946 "querying registration node size for function without personality");
15947 // The RegNodeSize is 6 32-bit words for SEH and 4 for C++ EH. See
15948 // WinEHStatePass for the full struct definition.
15949 switch (classifyEHPersonality(Fn->getPersonalityFn())) {
15950 case EHPersonality::MSVC_X86SEH: return 24;
15951 case EHPersonality::MSVC_CXX: return 16;
15954 report_fatal_error("can only recover FP for MSVC EH personality functions");
15957 /// When the 32-bit MSVC runtime transfers control to us, either to an outlined
15958 /// function or when returning to a parent frame after catching an exception, we
15959 /// recover the parent frame pointer by doing arithmetic on the incoming EBP.
15960 /// Here's the math:
15961 /// RegNodeBase = EntryEBP - RegNodeSize
15962 /// ParentFP = RegNodeBase - RegNodeFrameOffset
15963 /// Subtracting RegNodeSize takes us to the offset of the registration node, and
15964 /// subtracting the offset (negative on x86) takes us back to the parent FP.
15965 static SDValue recoverFramePointer(SelectionDAG &DAG, const Function *Fn,
15966 SDValue EntryEBP) {
15967 MachineFunction &MF = DAG.getMachineFunction();
15970 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
15971 MVT PtrVT = TLI.getPointerTy(DAG.getDataLayout());
15973 // It's possible that the parent function no longer has a personality function
15974 // if the exceptional code was optimized away, in which case we just return
15975 // the incoming EBP.
15976 if (!Fn->hasPersonalityFn())
15979 int RegNodeSize = getSEHRegistrationNodeSize(Fn);
15981 // Get an MCSymbol that will ultimately resolve to the frame offset of the EH
15983 MCSymbol *OffsetSym =
15984 MF.getMMI().getContext().getOrCreateParentFrameOffsetSymbol(
15985 GlobalValue::getRealLinkageName(Fn->getName()));
15986 SDValue OffsetSymVal = DAG.getMCSymbol(OffsetSym, PtrVT);
15987 SDValue RegNodeFrameOffset =
15988 DAG.getNode(ISD::LOCAL_RECOVER, dl, PtrVT, OffsetSymVal);
15990 // RegNodeBase = EntryEBP - RegNodeSize
15991 // ParentFP = RegNodeBase - RegNodeFrameOffset
15992 SDValue RegNodeBase = DAG.getNode(ISD::SUB, dl, PtrVT, EntryEBP,
15993 DAG.getConstant(RegNodeSize, dl, PtrVT));
15994 return DAG.getNode(ISD::SUB, dl, PtrVT, RegNodeBase, RegNodeFrameOffset);
15997 static SDValue LowerINTRINSIC_WO_CHAIN(SDValue Op, const X86Subtarget *Subtarget,
15998 SelectionDAG &DAG) {
16000 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
16001 EVT VT = Op.getValueType();
16002 const IntrinsicData* IntrData = getIntrinsicWithoutChain(IntNo);
16004 switch(IntrData->Type) {
16005 case INTR_TYPE_1OP:
16006 return DAG.getNode(IntrData->Opc0, dl, Op.getValueType(), Op.getOperand(1));
16007 case INTR_TYPE_2OP:
16008 return DAG.getNode(IntrData->Opc0, dl, Op.getValueType(), Op.getOperand(1),
16010 case INTR_TYPE_2OP_IMM8:
16011 return DAG.getNode(IntrData->Opc0, dl, Op.getValueType(), Op.getOperand(1),
16012 DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op.getOperand(2)));
16013 case INTR_TYPE_3OP:
16014 return DAG.getNode(IntrData->Opc0, dl, Op.getValueType(), Op.getOperand(1),
16015 Op.getOperand(2), Op.getOperand(3));
16016 case INTR_TYPE_4OP:
16017 return DAG.getNode(IntrData->Opc0, dl, Op.getValueType(), Op.getOperand(1),
16018 Op.getOperand(2), Op.getOperand(3), Op.getOperand(4));
16019 case INTR_TYPE_1OP_MASK_RM: {
16020 SDValue Src = Op.getOperand(1);
16021 SDValue PassThru = Op.getOperand(2);
16022 SDValue Mask = Op.getOperand(3);
16023 SDValue RoundingMode;
16024 // We allways add rounding mode to the Node.
16025 // If the rounding mode is not specified, we add the
16026 // "current direction" mode.
16027 if (Op.getNumOperands() == 4)
16029 DAG.getConstant(X86::STATIC_ROUNDING::CUR_DIRECTION, dl, MVT::i32);
16031 RoundingMode = Op.getOperand(4);
16032 unsigned IntrWithRoundingModeOpcode = IntrData->Opc1;
16033 if (IntrWithRoundingModeOpcode != 0)
16034 if (cast<ConstantSDNode>(RoundingMode)->getZExtValue() !=
16035 X86::STATIC_ROUNDING::CUR_DIRECTION)
16036 return getVectorMaskingNode(DAG.getNode(IntrWithRoundingModeOpcode,
16037 dl, Op.getValueType(), Src, RoundingMode),
16038 Mask, PassThru, Subtarget, DAG);
16039 return getVectorMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT, Src,
16041 Mask, PassThru, Subtarget, DAG);
16043 case INTR_TYPE_1OP_MASK: {
16044 SDValue Src = Op.getOperand(1);
16045 SDValue PassThru = Op.getOperand(2);
16046 SDValue Mask = Op.getOperand(3);
16047 // We add rounding mode to the Node when
16048 // - RM Opcode is specified and
16049 // - RM is not "current direction".
16050 unsigned IntrWithRoundingModeOpcode = IntrData->Opc1;
16051 if (IntrWithRoundingModeOpcode != 0) {
16052 SDValue Rnd = Op.getOperand(4);
16053 unsigned Round = cast<ConstantSDNode>(Rnd)->getZExtValue();
16054 if (Round != X86::STATIC_ROUNDING::CUR_DIRECTION) {
16055 return getVectorMaskingNode(DAG.getNode(IntrWithRoundingModeOpcode,
16056 dl, Op.getValueType(),
16058 Mask, PassThru, Subtarget, DAG);
16061 return getVectorMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT, Src),
16062 Mask, PassThru, Subtarget, DAG);
16064 case INTR_TYPE_SCALAR_MASK: {
16065 SDValue Src1 = Op.getOperand(1);
16066 SDValue Src2 = Op.getOperand(2);
16067 SDValue passThru = Op.getOperand(3);
16068 SDValue Mask = Op.getOperand(4);
16069 return getScalarMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT, Src1, Src2),
16070 Mask, passThru, Subtarget, DAG);
16072 case INTR_TYPE_SCALAR_MASK_RM: {
16073 SDValue Src1 = Op.getOperand(1);
16074 SDValue Src2 = Op.getOperand(2);
16075 SDValue Src0 = Op.getOperand(3);
16076 SDValue Mask = Op.getOperand(4);
16077 // There are 2 kinds of intrinsics in this group:
16078 // (1) With suppress-all-exceptions (sae) or rounding mode- 6 operands
16079 // (2) With rounding mode and sae - 7 operands.
16080 if (Op.getNumOperands() == 6) {
16081 SDValue Sae = Op.getOperand(5);
16082 unsigned Opc = IntrData->Opc1 ? IntrData->Opc1 : IntrData->Opc0;
16083 return getScalarMaskingNode(DAG.getNode(Opc, dl, VT, Src1, Src2,
16085 Mask, Src0, Subtarget, DAG);
16087 assert(Op.getNumOperands() == 7 && "Unexpected intrinsic form");
16088 SDValue RoundingMode = Op.getOperand(5);
16089 SDValue Sae = Op.getOperand(6);
16090 return getScalarMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT, Src1, Src2,
16091 RoundingMode, Sae),
16092 Mask, Src0, Subtarget, DAG);
16094 case INTR_TYPE_2OP_MASK:
16095 case INTR_TYPE_2OP_IMM8_MASK: {
16096 SDValue Src1 = Op.getOperand(1);
16097 SDValue Src2 = Op.getOperand(2);
16098 SDValue PassThru = Op.getOperand(3);
16099 SDValue Mask = Op.getOperand(4);
16101 if (IntrData->Type == INTR_TYPE_2OP_IMM8_MASK)
16102 Src2 = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Src2);
16104 // We specify 2 possible opcodes for intrinsics with rounding modes.
16105 // First, we check if the intrinsic may have non-default rounding mode,
16106 // (IntrData->Opc1 != 0), then we check the rounding mode operand.
16107 unsigned IntrWithRoundingModeOpcode = IntrData->Opc1;
16108 if (IntrWithRoundingModeOpcode != 0) {
16109 SDValue Rnd = Op.getOperand(5);
16110 unsigned Round = cast<ConstantSDNode>(Rnd)->getZExtValue();
16111 if (Round != X86::STATIC_ROUNDING::CUR_DIRECTION) {
16112 return getVectorMaskingNode(DAG.getNode(IntrWithRoundingModeOpcode,
16113 dl, Op.getValueType(),
16115 Mask, PassThru, Subtarget, DAG);
16118 // TODO: Intrinsics should have fast-math-flags to propagate.
16119 return getVectorMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT,Src1,Src2),
16120 Mask, PassThru, Subtarget, DAG);
16122 case INTR_TYPE_2OP_MASK_RM: {
16123 SDValue Src1 = Op.getOperand(1);
16124 SDValue Src2 = Op.getOperand(2);
16125 SDValue PassThru = Op.getOperand(3);
16126 SDValue Mask = Op.getOperand(4);
16127 // We specify 2 possible modes for intrinsics, with/without rounding
16129 // First, we check if the intrinsic have rounding mode (6 operands),
16130 // if not, we set rounding mode to "current".
16132 if (Op.getNumOperands() == 6)
16133 Rnd = Op.getOperand(5);
16135 Rnd = DAG.getConstant(X86::STATIC_ROUNDING::CUR_DIRECTION, dl, MVT::i32);
16136 return getVectorMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT,
16138 Mask, PassThru, Subtarget, DAG);
16140 case INTR_TYPE_3OP_SCALAR_MASK_RM: {
16141 SDValue Src1 = Op.getOperand(1);
16142 SDValue Src2 = Op.getOperand(2);
16143 SDValue Src3 = Op.getOperand(3);
16144 SDValue PassThru = Op.getOperand(4);
16145 SDValue Mask = Op.getOperand(5);
16146 SDValue Sae = Op.getOperand(6);
16148 return getScalarMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT, Src1,
16150 Mask, PassThru, Subtarget, DAG);
16152 case INTR_TYPE_3OP_MASK_RM: {
16153 SDValue Src1 = Op.getOperand(1);
16154 SDValue Src2 = Op.getOperand(2);
16155 SDValue Imm = Op.getOperand(3);
16156 SDValue PassThru = Op.getOperand(4);
16157 SDValue Mask = Op.getOperand(5);
16158 // We specify 2 possible modes for intrinsics, with/without rounding
16160 // First, we check if the intrinsic have rounding mode (7 operands),
16161 // if not, we set rounding mode to "current".
16163 if (Op.getNumOperands() == 7)
16164 Rnd = Op.getOperand(6);
16166 Rnd = DAG.getConstant(X86::STATIC_ROUNDING::CUR_DIRECTION, dl, MVT::i32);
16167 return getVectorMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT,
16168 Src1, Src2, Imm, Rnd),
16169 Mask, PassThru, Subtarget, DAG);
16171 case INTR_TYPE_3OP_IMM8_MASK:
16172 case INTR_TYPE_3OP_MASK:
16173 case INSERT_SUBVEC: {
16174 SDValue Src1 = Op.getOperand(1);
16175 SDValue Src2 = Op.getOperand(2);
16176 SDValue Src3 = Op.getOperand(3);
16177 SDValue PassThru = Op.getOperand(4);
16178 SDValue Mask = Op.getOperand(5);
16180 if (IntrData->Type == INTR_TYPE_3OP_IMM8_MASK)
16181 Src3 = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Src3);
16182 else if (IntrData->Type == INSERT_SUBVEC) {
16183 // imm should be adapted to ISD::INSERT_SUBVECTOR behavior
16184 assert(isa<ConstantSDNode>(Src3) && "Expected a ConstantSDNode here!");
16185 unsigned Imm = cast<ConstantSDNode>(Src3)->getZExtValue();
16186 Imm *= Src2.getValueType().getVectorNumElements();
16187 Src3 = DAG.getTargetConstant(Imm, dl, MVT::i32);
16190 // We specify 2 possible opcodes for intrinsics with rounding modes.
16191 // First, we check if the intrinsic may have non-default rounding mode,
16192 // (IntrData->Opc1 != 0), then we check the rounding mode operand.
16193 unsigned IntrWithRoundingModeOpcode = IntrData->Opc1;
16194 if (IntrWithRoundingModeOpcode != 0) {
16195 SDValue Rnd = Op.getOperand(6);
16196 unsigned Round = cast<ConstantSDNode>(Rnd)->getZExtValue();
16197 if (Round != X86::STATIC_ROUNDING::CUR_DIRECTION) {
16198 return getVectorMaskingNode(DAG.getNode(IntrWithRoundingModeOpcode,
16199 dl, Op.getValueType(),
16200 Src1, Src2, Src3, Rnd),
16201 Mask, PassThru, Subtarget, DAG);
16204 return getVectorMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT,
16206 Mask, PassThru, Subtarget, DAG);
16208 case VPERM_3OP_MASKZ:
16209 case VPERM_3OP_MASK:
16212 case FMA_OP_MASK: {
16213 SDValue Src1 = Op.getOperand(1);
16214 SDValue Src2 = Op.getOperand(2);
16215 SDValue Src3 = Op.getOperand(3);
16216 SDValue Mask = Op.getOperand(4);
16217 EVT VT = Op.getValueType();
16218 SDValue PassThru = SDValue();
16220 // set PassThru element
16221 if (IntrData->Type == VPERM_3OP_MASKZ || IntrData->Type == FMA_OP_MASKZ)
16222 PassThru = getZeroVector(VT, Subtarget, DAG, dl);
16223 else if (IntrData->Type == FMA_OP_MASK3)
16228 // We specify 2 possible opcodes for intrinsics with rounding modes.
16229 // First, we check if the intrinsic may have non-default rounding mode,
16230 // (IntrData->Opc1 != 0), then we check the rounding mode operand.
16231 unsigned IntrWithRoundingModeOpcode = IntrData->Opc1;
16232 if (IntrWithRoundingModeOpcode != 0) {
16233 SDValue Rnd = Op.getOperand(5);
16234 if (cast<ConstantSDNode>(Rnd)->getZExtValue() !=
16235 X86::STATIC_ROUNDING::CUR_DIRECTION)
16236 return getVectorMaskingNode(DAG.getNode(IntrWithRoundingModeOpcode,
16237 dl, Op.getValueType(),
16238 Src1, Src2, Src3, Rnd),
16239 Mask, PassThru, Subtarget, DAG);
16241 return getVectorMaskingNode(DAG.getNode(IntrData->Opc0,
16242 dl, Op.getValueType(),
16244 Mask, PassThru, Subtarget, DAG);
16246 case TERLOG_OP_MASK:
16247 case TERLOG_OP_MASKZ: {
16248 SDValue Src1 = Op.getOperand(1);
16249 SDValue Src2 = Op.getOperand(2);
16250 SDValue Src3 = Op.getOperand(3);
16251 SDValue Src4 = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op.getOperand(4));
16252 SDValue Mask = Op.getOperand(5);
16253 EVT VT = Op.getValueType();
16254 SDValue PassThru = Src1;
16255 // Set PassThru element.
16256 if (IntrData->Type == TERLOG_OP_MASKZ)
16257 PassThru = getZeroVector(VT, Subtarget, DAG, dl);
16259 return getVectorMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT,
16260 Src1, Src2, Src3, Src4),
16261 Mask, PassThru, Subtarget, DAG);
16264 // FPclass intrinsics with mask
16265 SDValue Src1 = Op.getOperand(1);
16266 EVT VT = Src1.getValueType();
16267 EVT MaskVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
16268 VT.getVectorNumElements());
16269 SDValue Imm = Op.getOperand(2);
16270 SDValue Mask = Op.getOperand(3);
16271 EVT BitcastVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
16272 Mask.getValueType().getSizeInBits());
16273 SDValue FPclass = DAG.getNode(IntrData->Opc0, dl, MaskVT, Src1, Imm);
16274 SDValue FPclassMask = getVectorMaskingNode(FPclass, Mask,
16275 DAG.getTargetConstant(0, dl, MaskVT),
16277 SDValue Res = DAG.getNode(ISD::INSERT_SUBVECTOR, dl, BitcastVT,
16278 DAG.getUNDEF(BitcastVT), FPclassMask,
16279 DAG.getIntPtrConstant(0, dl));
16280 return DAG.getBitcast(Op.getValueType(), Res);
16283 SDValue Src1 = Op.getOperand(1);
16284 SDValue Imm = Op.getOperand(2);
16285 SDValue Mask = Op.getOperand(3);
16286 SDValue FPclass = DAG.getNode(IntrData->Opc0, dl, MVT::i1, Src1, Imm);
16287 SDValue FPclassMask = getScalarMaskingNode(FPclass, Mask,
16288 DAG.getTargetConstant(0, dl, MVT::i1), Subtarget, DAG);
16289 return DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::i8, FPclassMask);
16292 case CMP_MASK_CC: {
16293 // Comparison intrinsics with masks.
16294 // Example of transformation:
16295 // (i8 (int_x86_avx512_mask_pcmpeq_q_128
16296 // (v2i64 %a), (v2i64 %b), (i8 %mask))) ->
16298 // (v8i1 (insert_subvector undef,
16299 // (v2i1 (and (PCMPEQM %a, %b),
16300 // (extract_subvector
16301 // (v8i1 (bitcast %mask)), 0))), 0))))
16302 EVT VT = Op.getOperand(1).getValueType();
16303 EVT MaskVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
16304 VT.getVectorNumElements());
16305 SDValue Mask = Op.getOperand((IntrData->Type == CMP_MASK_CC) ? 4 : 3);
16306 EVT BitcastVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
16307 Mask.getValueType().getSizeInBits());
16309 if (IntrData->Type == CMP_MASK_CC) {
16310 SDValue CC = Op.getOperand(3);
16311 CC = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, CC);
16312 // We specify 2 possible opcodes for intrinsics with rounding modes.
16313 // First, we check if the intrinsic may have non-default rounding mode,
16314 // (IntrData->Opc1 != 0), then we check the rounding mode operand.
16315 if (IntrData->Opc1 != 0) {
16316 SDValue Rnd = Op.getOperand(5);
16317 if (cast<ConstantSDNode>(Rnd)->getZExtValue() !=
16318 X86::STATIC_ROUNDING::CUR_DIRECTION)
16319 Cmp = DAG.getNode(IntrData->Opc1, dl, MaskVT, Op.getOperand(1),
16320 Op.getOperand(2), CC, Rnd);
16322 //default rounding mode
16324 Cmp = DAG.getNode(IntrData->Opc0, dl, MaskVT, Op.getOperand(1),
16325 Op.getOperand(2), CC);
16328 assert(IntrData->Type == CMP_MASK && "Unexpected intrinsic type!");
16329 Cmp = DAG.getNode(IntrData->Opc0, dl, MaskVT, Op.getOperand(1),
16332 SDValue CmpMask = getVectorMaskingNode(Cmp, Mask,
16333 DAG.getTargetConstant(0, dl,
16336 SDValue Res = DAG.getNode(ISD::INSERT_SUBVECTOR, dl, BitcastVT,
16337 DAG.getUNDEF(BitcastVT), CmpMask,
16338 DAG.getIntPtrConstant(0, dl));
16339 return DAG.getBitcast(Op.getValueType(), Res);
16341 case CMP_MASK_SCALAR_CC: {
16342 SDValue Src1 = Op.getOperand(1);
16343 SDValue Src2 = Op.getOperand(2);
16344 SDValue CC = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op.getOperand(3));
16345 SDValue Mask = Op.getOperand(4);
16348 if (IntrData->Opc1 != 0) {
16349 SDValue Rnd = Op.getOperand(5);
16350 if (cast<ConstantSDNode>(Rnd)->getZExtValue() !=
16351 X86::STATIC_ROUNDING::CUR_DIRECTION)
16352 Cmp = DAG.getNode(IntrData->Opc1, dl, MVT::i1, Src1, Src2, CC, Rnd);
16354 //default rounding mode
16356 Cmp = DAG.getNode(IntrData->Opc0, dl, MVT::i1, Src1, Src2, CC);
16358 SDValue CmpMask = getScalarMaskingNode(Cmp, Mask,
16359 DAG.getTargetConstant(0, dl,
16363 return DAG.getNode(ISD::SIGN_EXTEND_INREG, dl, MVT::i8,
16364 DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i8, CmpMask),
16365 DAG.getValueType(MVT::i1));
16367 case COMI: { // Comparison intrinsics
16368 ISD::CondCode CC = (ISD::CondCode)IntrData->Opc1;
16369 SDValue LHS = Op.getOperand(1);
16370 SDValue RHS = Op.getOperand(2);
16371 unsigned X86CC = TranslateX86CC(CC, dl, true, LHS, RHS, DAG);
16372 assert(X86CC != X86::COND_INVALID && "Unexpected illegal condition!");
16373 SDValue Cond = DAG.getNode(IntrData->Opc0, dl, MVT::i32, LHS, RHS);
16374 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
16375 DAG.getConstant(X86CC, dl, MVT::i8), Cond);
16376 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
16379 return getTargetVShiftNode(IntrData->Opc0, dl, Op.getSimpleValueType(),
16380 Op.getOperand(1), Op.getOperand(2), DAG);
16382 return getVectorMaskingNode(getTargetVShiftNode(IntrData->Opc0, dl,
16383 Op.getSimpleValueType(),
16385 Op.getOperand(2), DAG),
16386 Op.getOperand(4), Op.getOperand(3), Subtarget,
16388 case COMPRESS_EXPAND_IN_REG: {
16389 SDValue Mask = Op.getOperand(3);
16390 SDValue DataToCompress = Op.getOperand(1);
16391 SDValue PassThru = Op.getOperand(2);
16392 if (isAllOnes(Mask)) // return data as is
16393 return Op.getOperand(1);
16395 return getVectorMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT,
16397 Mask, PassThru, Subtarget, DAG);
16400 SDValue Mask = Op.getOperand(3);
16401 EVT VT = Op.getValueType();
16402 EVT MaskVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
16403 VT.getVectorNumElements());
16404 EVT BitcastVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
16405 Mask.getValueType().getSizeInBits());
16407 SDValue VMask = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MaskVT,
16408 DAG.getBitcast(BitcastVT, Mask),
16409 DAG.getIntPtrConstant(0, dl));
16410 return DAG.getNode(IntrData->Opc0, dl, VT, VMask, Op.getOperand(1),
16419 default: return SDValue(); // Don't custom lower most intrinsics.
16421 case Intrinsic::x86_avx2_permd:
16422 case Intrinsic::x86_avx2_permps:
16423 // Operands intentionally swapped. Mask is last operand to intrinsic,
16424 // but second operand for node/instruction.
16425 return DAG.getNode(X86ISD::VPERMV, dl, Op.getValueType(),
16426 Op.getOperand(2), Op.getOperand(1));
16428 // ptest and testp intrinsics. The intrinsic these come from are designed to
16429 // return an integer value, not just an instruction so lower it to the ptest
16430 // or testp pattern and a setcc for the result.
16431 case Intrinsic::x86_sse41_ptestz:
16432 case Intrinsic::x86_sse41_ptestc:
16433 case Intrinsic::x86_sse41_ptestnzc:
16434 case Intrinsic::x86_avx_ptestz_256:
16435 case Intrinsic::x86_avx_ptestc_256:
16436 case Intrinsic::x86_avx_ptestnzc_256:
16437 case Intrinsic::x86_avx_vtestz_ps:
16438 case Intrinsic::x86_avx_vtestc_ps:
16439 case Intrinsic::x86_avx_vtestnzc_ps:
16440 case Intrinsic::x86_avx_vtestz_pd:
16441 case Intrinsic::x86_avx_vtestc_pd:
16442 case Intrinsic::x86_avx_vtestnzc_pd:
16443 case Intrinsic::x86_avx_vtestz_ps_256:
16444 case Intrinsic::x86_avx_vtestc_ps_256:
16445 case Intrinsic::x86_avx_vtestnzc_ps_256:
16446 case Intrinsic::x86_avx_vtestz_pd_256:
16447 case Intrinsic::x86_avx_vtestc_pd_256:
16448 case Intrinsic::x86_avx_vtestnzc_pd_256: {
16449 bool IsTestPacked = false;
16452 default: llvm_unreachable("Bad fallthrough in Intrinsic lowering.");
16453 case Intrinsic::x86_avx_vtestz_ps:
16454 case Intrinsic::x86_avx_vtestz_pd:
16455 case Intrinsic::x86_avx_vtestz_ps_256:
16456 case Intrinsic::x86_avx_vtestz_pd_256:
16457 IsTestPacked = true; // Fallthrough
16458 case Intrinsic::x86_sse41_ptestz:
16459 case Intrinsic::x86_avx_ptestz_256:
16461 X86CC = X86::COND_E;
16463 case Intrinsic::x86_avx_vtestc_ps:
16464 case Intrinsic::x86_avx_vtestc_pd:
16465 case Intrinsic::x86_avx_vtestc_ps_256:
16466 case Intrinsic::x86_avx_vtestc_pd_256:
16467 IsTestPacked = true; // Fallthrough
16468 case Intrinsic::x86_sse41_ptestc:
16469 case Intrinsic::x86_avx_ptestc_256:
16471 X86CC = X86::COND_B;
16473 case Intrinsic::x86_avx_vtestnzc_ps:
16474 case Intrinsic::x86_avx_vtestnzc_pd:
16475 case Intrinsic::x86_avx_vtestnzc_ps_256:
16476 case Intrinsic::x86_avx_vtestnzc_pd_256:
16477 IsTestPacked = true; // Fallthrough
16478 case Intrinsic::x86_sse41_ptestnzc:
16479 case Intrinsic::x86_avx_ptestnzc_256:
16481 X86CC = X86::COND_A;
16485 SDValue LHS = Op.getOperand(1);
16486 SDValue RHS = Op.getOperand(2);
16487 unsigned TestOpc = IsTestPacked ? X86ISD::TESTP : X86ISD::PTEST;
16488 SDValue Test = DAG.getNode(TestOpc, dl, MVT::i32, LHS, RHS);
16489 SDValue CC = DAG.getConstant(X86CC, dl, MVT::i8);
16490 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8, CC, Test);
16491 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
16493 case Intrinsic::x86_avx512_kortestz_w:
16494 case Intrinsic::x86_avx512_kortestc_w: {
16495 unsigned X86CC = (IntNo == Intrinsic::x86_avx512_kortestz_w)? X86::COND_E: X86::COND_B;
16496 SDValue LHS = DAG.getBitcast(MVT::v16i1, Op.getOperand(1));
16497 SDValue RHS = DAG.getBitcast(MVT::v16i1, Op.getOperand(2));
16498 SDValue CC = DAG.getConstant(X86CC, dl, MVT::i8);
16499 SDValue Test = DAG.getNode(X86ISD::KORTEST, dl, MVT::i32, LHS, RHS);
16500 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i1, CC, Test);
16501 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
16504 case Intrinsic::x86_sse42_pcmpistria128:
16505 case Intrinsic::x86_sse42_pcmpestria128:
16506 case Intrinsic::x86_sse42_pcmpistric128:
16507 case Intrinsic::x86_sse42_pcmpestric128:
16508 case Intrinsic::x86_sse42_pcmpistrio128:
16509 case Intrinsic::x86_sse42_pcmpestrio128:
16510 case Intrinsic::x86_sse42_pcmpistris128:
16511 case Intrinsic::x86_sse42_pcmpestris128:
16512 case Intrinsic::x86_sse42_pcmpistriz128:
16513 case Intrinsic::x86_sse42_pcmpestriz128: {
16517 default: llvm_unreachable("Impossible intrinsic"); // Can't reach here.
16518 case Intrinsic::x86_sse42_pcmpistria128:
16519 Opcode = X86ISD::PCMPISTRI;
16520 X86CC = X86::COND_A;
16522 case Intrinsic::x86_sse42_pcmpestria128:
16523 Opcode = X86ISD::PCMPESTRI;
16524 X86CC = X86::COND_A;
16526 case Intrinsic::x86_sse42_pcmpistric128:
16527 Opcode = X86ISD::PCMPISTRI;
16528 X86CC = X86::COND_B;
16530 case Intrinsic::x86_sse42_pcmpestric128:
16531 Opcode = X86ISD::PCMPESTRI;
16532 X86CC = X86::COND_B;
16534 case Intrinsic::x86_sse42_pcmpistrio128:
16535 Opcode = X86ISD::PCMPISTRI;
16536 X86CC = X86::COND_O;
16538 case Intrinsic::x86_sse42_pcmpestrio128:
16539 Opcode = X86ISD::PCMPESTRI;
16540 X86CC = X86::COND_O;
16542 case Intrinsic::x86_sse42_pcmpistris128:
16543 Opcode = X86ISD::PCMPISTRI;
16544 X86CC = X86::COND_S;
16546 case Intrinsic::x86_sse42_pcmpestris128:
16547 Opcode = X86ISD::PCMPESTRI;
16548 X86CC = X86::COND_S;
16550 case Intrinsic::x86_sse42_pcmpistriz128:
16551 Opcode = X86ISD::PCMPISTRI;
16552 X86CC = X86::COND_E;
16554 case Intrinsic::x86_sse42_pcmpestriz128:
16555 Opcode = X86ISD::PCMPESTRI;
16556 X86CC = X86::COND_E;
16559 SmallVector<SDValue, 5> NewOps(Op->op_begin()+1, Op->op_end());
16560 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::i32);
16561 SDValue PCMP = DAG.getNode(Opcode, dl, VTs, NewOps);
16562 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
16563 DAG.getConstant(X86CC, dl, MVT::i8),
16564 SDValue(PCMP.getNode(), 1));
16565 return DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, SetCC);
16568 case Intrinsic::x86_sse42_pcmpistri128:
16569 case Intrinsic::x86_sse42_pcmpestri128: {
16571 if (IntNo == Intrinsic::x86_sse42_pcmpistri128)
16572 Opcode = X86ISD::PCMPISTRI;
16574 Opcode = X86ISD::PCMPESTRI;
16576 SmallVector<SDValue, 5> NewOps(Op->op_begin()+1, Op->op_end());
16577 SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::i32);
16578 return DAG.getNode(Opcode, dl, VTs, NewOps);
16581 case Intrinsic::x86_seh_lsda: {
16582 // Compute the symbol for the LSDA. We know it'll get emitted later.
16583 MachineFunction &MF = DAG.getMachineFunction();
16584 SDValue Op1 = Op.getOperand(1);
16585 auto *Fn = cast<Function>(cast<GlobalAddressSDNode>(Op1)->getGlobal());
16586 MCSymbol *LSDASym = MF.getMMI().getContext().getOrCreateLSDASymbol(
16587 GlobalValue::getRealLinkageName(Fn->getName()));
16589 // Generate a simple absolute symbol reference. This intrinsic is only
16590 // supported on 32-bit Windows, which isn't PIC.
16591 SDValue Result = DAG.getMCSymbol(LSDASym, VT);
16592 return DAG.getNode(X86ISD::Wrapper, dl, VT, Result);
16595 case Intrinsic::x86_seh_recoverfp: {
16596 SDValue FnOp = Op.getOperand(1);
16597 SDValue IncomingFPOp = Op.getOperand(2);
16598 GlobalAddressSDNode *GSD = dyn_cast<GlobalAddressSDNode>(FnOp);
16599 auto *Fn = dyn_cast_or_null<Function>(GSD ? GSD->getGlobal() : nullptr);
16601 report_fatal_error(
16602 "llvm.x86.seh.recoverfp must take a function as the first argument");
16603 return recoverFramePointer(DAG, Fn, IncomingFPOp);
16606 case Intrinsic::localaddress: {
16607 // Returns one of the stack, base, or frame pointer registers, depending on
16608 // which is used to reference local variables.
16609 MachineFunction &MF = DAG.getMachineFunction();
16610 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
16612 if (RegInfo->hasBasePointer(MF))
16613 Reg = RegInfo->getBaseRegister();
16614 else // This function handles the SP or FP case.
16615 Reg = RegInfo->getPtrSizedFrameRegister(MF);
16616 return DAG.getCopyFromReg(DAG.getEntryNode(), dl, Reg, VT);
16621 static SDValue getGatherNode(unsigned Opc, SDValue Op, SelectionDAG &DAG,
16622 SDValue Src, SDValue Mask, SDValue Base,
16623 SDValue Index, SDValue ScaleOp, SDValue Chain,
16624 const X86Subtarget * Subtarget) {
16626 ConstantSDNode *C = dyn_cast<ConstantSDNode>(ScaleOp);
16628 llvm_unreachable("Invalid scale type");
16629 unsigned ScaleVal = C->getZExtValue();
16630 if (ScaleVal > 2 && ScaleVal != 4 && ScaleVal != 8)
16631 llvm_unreachable("Valid scale values are 1, 2, 4, 8");
16633 SDValue Scale = DAG.getTargetConstant(C->getZExtValue(), dl, MVT::i8);
16634 EVT MaskVT = MVT::getVectorVT(MVT::i1,
16635 Index.getSimpleValueType().getVectorNumElements());
16637 ConstantSDNode *MaskC = dyn_cast<ConstantSDNode>(Mask);
16639 MaskInReg = DAG.getTargetConstant(MaskC->getSExtValue(), dl, MaskVT);
16641 EVT BitcastVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
16642 Mask.getValueType().getSizeInBits());
16644 // In case when MaskVT equals v2i1 or v4i1, low 2 or 4 elements
16645 // are extracted by EXTRACT_SUBVECTOR.
16646 MaskInReg = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MaskVT,
16647 DAG.getBitcast(BitcastVT, Mask),
16648 DAG.getIntPtrConstant(0, dl));
16650 SDVTList VTs = DAG.getVTList(Op.getValueType(), MaskVT, MVT::Other);
16651 SDValue Disp = DAG.getTargetConstant(0, dl, MVT::i32);
16652 SDValue Segment = DAG.getRegister(0, MVT::i32);
16653 if (Src.getOpcode() == ISD::UNDEF)
16654 Src = getZeroVector(Op.getValueType(), Subtarget, DAG, dl);
16655 SDValue Ops[] = {Src, MaskInReg, Base, Scale, Index, Disp, Segment, Chain};
16656 SDNode *Res = DAG.getMachineNode(Opc, dl, VTs, Ops);
16657 SDValue RetOps[] = { SDValue(Res, 0), SDValue(Res, 2) };
16658 return DAG.getMergeValues(RetOps, dl);
16661 static SDValue getScatterNode(unsigned Opc, SDValue Op, SelectionDAG &DAG,
16662 SDValue Src, SDValue Mask, SDValue Base,
16663 SDValue Index, SDValue ScaleOp, SDValue Chain) {
16665 ConstantSDNode *C = dyn_cast<ConstantSDNode>(ScaleOp);
16667 llvm_unreachable("Invalid scale type");
16668 unsigned ScaleVal = C->getZExtValue();
16669 if (ScaleVal > 2 && ScaleVal != 4 && ScaleVal != 8)
16670 llvm_unreachable("Valid scale values are 1, 2, 4, 8");
16672 SDValue Scale = DAG.getTargetConstant(C->getZExtValue(), dl, MVT::i8);
16673 SDValue Disp = DAG.getTargetConstant(0, dl, MVT::i32);
16674 SDValue Segment = DAG.getRegister(0, MVT::i32);
16675 EVT MaskVT = MVT::getVectorVT(MVT::i1,
16676 Index.getSimpleValueType().getVectorNumElements());
16678 ConstantSDNode *MaskC = dyn_cast<ConstantSDNode>(Mask);
16680 MaskInReg = DAG.getTargetConstant(MaskC->getSExtValue(), dl, MaskVT);
16682 EVT BitcastVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
16683 Mask.getValueType().getSizeInBits());
16685 // In case when MaskVT equals v2i1 or v4i1, low 2 or 4 elements
16686 // are extracted by EXTRACT_SUBVECTOR.
16687 MaskInReg = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MaskVT,
16688 DAG.getBitcast(BitcastVT, Mask),
16689 DAG.getIntPtrConstant(0, dl));
16691 SDVTList VTs = DAG.getVTList(MaskVT, MVT::Other);
16692 SDValue Ops[] = {Base, Scale, Index, Disp, Segment, MaskInReg, Src, Chain};
16693 SDNode *Res = DAG.getMachineNode(Opc, dl, VTs, Ops);
16694 return SDValue(Res, 1);
16697 static SDValue getPrefetchNode(unsigned Opc, SDValue Op, SelectionDAG &DAG,
16698 SDValue Mask, SDValue Base, SDValue Index,
16699 SDValue ScaleOp, SDValue Chain) {
16701 ConstantSDNode *C = dyn_cast<ConstantSDNode>(ScaleOp);
16702 assert(C && "Invalid scale type");
16703 SDValue Scale = DAG.getTargetConstant(C->getZExtValue(), dl, MVT::i8);
16704 SDValue Disp = DAG.getTargetConstant(0, dl, MVT::i32);
16705 SDValue Segment = DAG.getRegister(0, MVT::i32);
16707 MVT::getVectorVT(MVT::i1, Index.getSimpleValueType().getVectorNumElements());
16709 ConstantSDNode *MaskC = dyn_cast<ConstantSDNode>(Mask);
16711 MaskInReg = DAG.getTargetConstant(MaskC->getSExtValue(), dl, MaskVT);
16713 MaskInReg = DAG.getBitcast(MaskVT, Mask);
16714 //SDVTList VTs = DAG.getVTList(MVT::Other);
16715 SDValue Ops[] = {MaskInReg, Base, Scale, Index, Disp, Segment, Chain};
16716 SDNode *Res = DAG.getMachineNode(Opc, dl, MVT::Other, Ops);
16717 return SDValue(Res, 0);
16720 // getReadPerformanceCounter - Handles the lowering of builtin intrinsics that
16721 // read performance monitor counters (x86_rdpmc).
16722 static void getReadPerformanceCounter(SDNode *N, SDLoc DL,
16723 SelectionDAG &DAG, const X86Subtarget *Subtarget,
16724 SmallVectorImpl<SDValue> &Results) {
16725 assert(N->getNumOperands() == 3 && "Unexpected number of operands!");
16726 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Glue);
16729 // The ECX register is used to select the index of the performance counter
16731 SDValue Chain = DAG.getCopyToReg(N->getOperand(0), DL, X86::ECX,
16733 SDValue rd = DAG.getNode(X86ISD::RDPMC_DAG, DL, Tys, Chain);
16735 // Reads the content of a 64-bit performance counter and returns it in the
16736 // registers EDX:EAX.
16737 if (Subtarget->is64Bit()) {
16738 LO = DAG.getCopyFromReg(rd, DL, X86::RAX, MVT::i64, rd.getValue(1));
16739 HI = DAG.getCopyFromReg(LO.getValue(1), DL, X86::RDX, MVT::i64,
16742 LO = DAG.getCopyFromReg(rd, DL, X86::EAX, MVT::i32, rd.getValue(1));
16743 HI = DAG.getCopyFromReg(LO.getValue(1), DL, X86::EDX, MVT::i32,
16746 Chain = HI.getValue(1);
16748 if (Subtarget->is64Bit()) {
16749 // The EAX register is loaded with the low-order 32 bits. The EDX register
16750 // is loaded with the supported high-order bits of the counter.
16751 SDValue Tmp = DAG.getNode(ISD::SHL, DL, MVT::i64, HI,
16752 DAG.getConstant(32, DL, MVT::i8));
16753 Results.push_back(DAG.getNode(ISD::OR, DL, MVT::i64, LO, Tmp));
16754 Results.push_back(Chain);
16758 // Use a buildpair to merge the two 32-bit values into a 64-bit one.
16759 SDValue Ops[] = { LO, HI };
16760 SDValue Pair = DAG.getNode(ISD::BUILD_PAIR, DL, MVT::i64, Ops);
16761 Results.push_back(Pair);
16762 Results.push_back(Chain);
16765 // getReadTimeStampCounter - Handles the lowering of builtin intrinsics that
16766 // read the time stamp counter (x86_rdtsc and x86_rdtscp). This function is
16767 // also used to custom lower READCYCLECOUNTER nodes.
16768 static void getReadTimeStampCounter(SDNode *N, SDLoc DL, unsigned Opcode,
16769 SelectionDAG &DAG, const X86Subtarget *Subtarget,
16770 SmallVectorImpl<SDValue> &Results) {
16771 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Glue);
16772 SDValue rd = DAG.getNode(Opcode, DL, Tys, N->getOperand(0));
16775 // The processor's time-stamp counter (a 64-bit MSR) is stored into the
16776 // EDX:EAX registers. EDX is loaded with the high-order 32 bits of the MSR
16777 // and the EAX register is loaded with the low-order 32 bits.
16778 if (Subtarget->is64Bit()) {
16779 LO = DAG.getCopyFromReg(rd, DL, X86::RAX, MVT::i64, rd.getValue(1));
16780 HI = DAG.getCopyFromReg(LO.getValue(1), DL, X86::RDX, MVT::i64,
16783 LO = DAG.getCopyFromReg(rd, DL, X86::EAX, MVT::i32, rd.getValue(1));
16784 HI = DAG.getCopyFromReg(LO.getValue(1), DL, X86::EDX, MVT::i32,
16787 SDValue Chain = HI.getValue(1);
16789 if (Opcode == X86ISD::RDTSCP_DAG) {
16790 assert(N->getNumOperands() == 3 && "Unexpected number of operands!");
16792 // Instruction RDTSCP loads the IA32:TSC_AUX_MSR (address C000_0103H) into
16793 // the ECX register. Add 'ecx' explicitly to the chain.
16794 SDValue ecx = DAG.getCopyFromReg(Chain, DL, X86::ECX, MVT::i32,
16796 // Explicitly store the content of ECX at the location passed in input
16797 // to the 'rdtscp' intrinsic.
16798 Chain = DAG.getStore(ecx.getValue(1), DL, ecx, N->getOperand(2),
16799 MachinePointerInfo(), false, false, 0);
16802 if (Subtarget->is64Bit()) {
16803 // The EDX register is loaded with the high-order 32 bits of the MSR, and
16804 // the EAX register is loaded with the low-order 32 bits.
16805 SDValue Tmp = DAG.getNode(ISD::SHL, DL, MVT::i64, HI,
16806 DAG.getConstant(32, DL, MVT::i8));
16807 Results.push_back(DAG.getNode(ISD::OR, DL, MVT::i64, LO, Tmp));
16808 Results.push_back(Chain);
16812 // Use a buildpair to merge the two 32-bit values into a 64-bit one.
16813 SDValue Ops[] = { LO, HI };
16814 SDValue Pair = DAG.getNode(ISD::BUILD_PAIR, DL, MVT::i64, Ops);
16815 Results.push_back(Pair);
16816 Results.push_back(Chain);
16819 static SDValue LowerREADCYCLECOUNTER(SDValue Op, const X86Subtarget *Subtarget,
16820 SelectionDAG &DAG) {
16821 SmallVector<SDValue, 2> Results;
16823 getReadTimeStampCounter(Op.getNode(), DL, X86ISD::RDTSC_DAG, DAG, Subtarget,
16825 return DAG.getMergeValues(Results, DL);
16828 static SDValue LowerSEHRESTOREFRAME(SDValue Op, const X86Subtarget *Subtarget,
16829 SelectionDAG &DAG) {
16830 MachineFunction &MF = DAG.getMachineFunction();
16831 const Function *Fn = MF.getFunction();
16833 SDValue Chain = Op.getOperand(0);
16835 assert(Subtarget->getFrameLowering()->hasFP(MF) &&
16836 "using llvm.x86.seh.restoreframe requires a frame pointer");
16838 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
16839 MVT VT = TLI.getPointerTy(DAG.getDataLayout());
16841 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
16842 unsigned FrameReg =
16843 RegInfo->getPtrSizedFrameRegister(DAG.getMachineFunction());
16844 unsigned SPReg = RegInfo->getStackRegister();
16845 unsigned SlotSize = RegInfo->getSlotSize();
16847 // Get incoming EBP.
16848 SDValue IncomingEBP =
16849 DAG.getCopyFromReg(Chain, dl, FrameReg, VT);
16851 // SP is saved in the first field of every registration node, so load
16852 // [EBP-RegNodeSize] into SP.
16853 int RegNodeSize = getSEHRegistrationNodeSize(Fn);
16854 SDValue SPAddr = DAG.getNode(ISD::ADD, dl, VT, IncomingEBP,
16855 DAG.getConstant(-RegNodeSize, dl, VT));
16857 DAG.getLoad(VT, dl, Chain, SPAddr, MachinePointerInfo(), false, false,
16858 false, VT.getScalarSizeInBits() / 8);
16859 Chain = DAG.getCopyToReg(Chain, dl, SPReg, NewSP);
16861 if (!RegInfo->needsStackRealignment(MF)) {
16862 // Adjust EBP to point back to the original frame position.
16863 SDValue NewFP = recoverFramePointer(DAG, Fn, IncomingEBP);
16864 Chain = DAG.getCopyToReg(Chain, dl, FrameReg, NewFP);
16866 assert(RegInfo->hasBasePointer(MF) &&
16867 "functions with Win32 EH must use frame or base pointer register");
16869 // Reload the base pointer (ESI) with the adjusted incoming EBP.
16870 SDValue NewBP = recoverFramePointer(DAG, Fn, IncomingEBP);
16871 Chain = DAG.getCopyToReg(Chain, dl, RegInfo->getBaseRegister(), NewBP);
16873 // Reload the spilled EBP value, now that the stack and base pointers are
16875 X86MachineFunctionInfo *X86FI = MF.getInfo<X86MachineFunctionInfo>();
16876 X86FI->setHasSEHFramePtrSave(true);
16877 int FI = MF.getFrameInfo()->CreateSpillStackObject(SlotSize, SlotSize);
16878 X86FI->setSEHFramePtrSaveIndex(FI);
16879 SDValue NewFP = DAG.getLoad(VT, dl, Chain, DAG.getFrameIndex(FI, VT),
16880 MachinePointerInfo(), false, false, false,
16881 VT.getScalarSizeInBits() / 8);
16882 Chain = DAG.getCopyToReg(NewFP, dl, FrameReg, NewFP);
16888 /// \brief Lower intrinsics for TRUNCATE_TO_MEM case
16889 /// return truncate Store/MaskedStore Node
16890 static SDValue LowerINTRINSIC_TRUNCATE_TO_MEM(const SDValue & Op,
16894 SDValue Mask = Op.getOperand(4);
16895 SDValue DataToTruncate = Op.getOperand(3);
16896 SDValue Addr = Op.getOperand(2);
16897 SDValue Chain = Op.getOperand(0);
16899 EVT VT = DataToTruncate.getValueType();
16900 EVT SVT = EVT::getVectorVT(*DAG.getContext(),
16901 ElementType, VT.getVectorNumElements());
16903 if (isAllOnes(Mask)) // return just a truncate store
16904 return DAG.getTruncStore(Chain, dl, DataToTruncate, Addr,
16905 MachinePointerInfo(), SVT, false, false,
16906 SVT.getScalarSizeInBits()/8);
16908 EVT MaskVT = EVT::getVectorVT(*DAG.getContext(),
16909 MVT::i1, VT.getVectorNumElements());
16910 EVT BitcastVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
16911 Mask.getValueType().getSizeInBits());
16912 // In case when MaskVT equals v2i1 or v4i1, low 2 or 4 elements
16913 // are extracted by EXTRACT_SUBVECTOR.
16914 SDValue VMask = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MaskVT,
16915 DAG.getBitcast(BitcastVT, Mask),
16916 DAG.getIntPtrConstant(0, dl));
16918 MachineMemOperand *MMO = DAG.getMachineFunction().
16919 getMachineMemOperand(MachinePointerInfo(),
16920 MachineMemOperand::MOStore, SVT.getStoreSize(),
16921 SVT.getScalarSizeInBits()/8);
16923 return DAG.getMaskedStore(Chain, dl, DataToTruncate, Addr,
16924 VMask, SVT, MMO, true);
16927 static SDValue LowerINTRINSIC_W_CHAIN(SDValue Op, const X86Subtarget *Subtarget,
16928 SelectionDAG &DAG) {
16929 unsigned IntNo = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
16931 const IntrinsicData* IntrData = getIntrinsicWithChain(IntNo);
16933 if (IntNo == llvm::Intrinsic::x86_seh_restoreframe)
16934 return LowerSEHRESTOREFRAME(Op, Subtarget, DAG);
16939 switch(IntrData->Type) {
16941 llvm_unreachable("Unknown Intrinsic Type");
16945 // Emit the node with the right value type.
16946 SDVTList VTs = DAG.getVTList(Op->getValueType(0), MVT::Glue, MVT::Other);
16947 SDValue Result = DAG.getNode(IntrData->Opc0, dl, VTs, Op.getOperand(0));
16949 // If the value returned by RDRAND/RDSEED was valid (CF=1), return 1.
16950 // Otherwise return the value from Rand, which is always 0, casted to i32.
16951 SDValue Ops[] = { DAG.getZExtOrTrunc(Result, dl, Op->getValueType(1)),
16952 DAG.getConstant(1, dl, Op->getValueType(1)),
16953 DAG.getConstant(X86::COND_B, dl, MVT::i32),
16954 SDValue(Result.getNode(), 1) };
16955 SDValue isValid = DAG.getNode(X86ISD::CMOV, dl,
16956 DAG.getVTList(Op->getValueType(1), MVT::Glue),
16959 // Return { result, isValid, chain }.
16960 return DAG.getNode(ISD::MERGE_VALUES, dl, Op->getVTList(), Result, isValid,
16961 SDValue(Result.getNode(), 2));
16964 //gather(v1, mask, index, base, scale);
16965 SDValue Chain = Op.getOperand(0);
16966 SDValue Src = Op.getOperand(2);
16967 SDValue Base = Op.getOperand(3);
16968 SDValue Index = Op.getOperand(4);
16969 SDValue Mask = Op.getOperand(5);
16970 SDValue Scale = Op.getOperand(6);
16971 return getGatherNode(IntrData->Opc0, Op, DAG, Src, Mask, Base, Index, Scale,
16975 //scatter(base, mask, index, v1, scale);
16976 SDValue Chain = Op.getOperand(0);
16977 SDValue Base = Op.getOperand(2);
16978 SDValue Mask = Op.getOperand(3);
16979 SDValue Index = Op.getOperand(4);
16980 SDValue Src = Op.getOperand(5);
16981 SDValue Scale = Op.getOperand(6);
16982 return getScatterNode(IntrData->Opc0, Op, DAG, Src, Mask, Base, Index,
16986 SDValue Hint = Op.getOperand(6);
16987 unsigned HintVal = cast<ConstantSDNode>(Hint)->getZExtValue();
16988 assert(HintVal < 2 && "Wrong prefetch hint in intrinsic: should be 0 or 1");
16989 unsigned Opcode = (HintVal ? IntrData->Opc1 : IntrData->Opc0);
16990 SDValue Chain = Op.getOperand(0);
16991 SDValue Mask = Op.getOperand(2);
16992 SDValue Index = Op.getOperand(3);
16993 SDValue Base = Op.getOperand(4);
16994 SDValue Scale = Op.getOperand(5);
16995 return getPrefetchNode(Opcode, Op, DAG, Mask, Base, Index, Scale, Chain);
16997 // Read Time Stamp Counter (RDTSC) and Processor ID (RDTSCP).
16999 SmallVector<SDValue, 2> Results;
17000 getReadTimeStampCounter(Op.getNode(), dl, IntrData->Opc0, DAG, Subtarget,
17002 return DAG.getMergeValues(Results, dl);
17004 // Read Performance Monitoring Counters.
17006 SmallVector<SDValue, 2> Results;
17007 getReadPerformanceCounter(Op.getNode(), dl, DAG, Subtarget, Results);
17008 return DAG.getMergeValues(Results, dl);
17010 // XTEST intrinsics.
17012 SDVTList VTs = DAG.getVTList(Op->getValueType(0), MVT::Other);
17013 SDValue InTrans = DAG.getNode(IntrData->Opc0, dl, VTs, Op.getOperand(0));
17014 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
17015 DAG.getConstant(X86::COND_NE, dl, MVT::i8),
17017 SDValue Ret = DAG.getNode(ISD::ZERO_EXTEND, dl, Op->getValueType(0), SetCC);
17018 return DAG.getNode(ISD::MERGE_VALUES, dl, Op->getVTList(),
17019 Ret, SDValue(InTrans.getNode(), 1));
17023 SmallVector<SDValue, 2> Results;
17024 SDVTList CFVTs = DAG.getVTList(Op->getValueType(0), MVT::Other);
17025 SDVTList VTs = DAG.getVTList(Op.getOperand(3)->getValueType(0), MVT::Other);
17026 SDValue GenCF = DAG.getNode(X86ISD::ADD, dl, CFVTs, Op.getOperand(2),
17027 DAG.getConstant(-1, dl, MVT::i8));
17028 SDValue Res = DAG.getNode(IntrData->Opc0, dl, VTs, Op.getOperand(3),
17029 Op.getOperand(4), GenCF.getValue(1));
17030 SDValue Store = DAG.getStore(Op.getOperand(0), dl, Res.getValue(0),
17031 Op.getOperand(5), MachinePointerInfo(),
17033 SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
17034 DAG.getConstant(X86::COND_B, dl, MVT::i8),
17036 Results.push_back(SetCC);
17037 Results.push_back(Store);
17038 return DAG.getMergeValues(Results, dl);
17040 case COMPRESS_TO_MEM: {
17042 SDValue Mask = Op.getOperand(4);
17043 SDValue DataToCompress = Op.getOperand(3);
17044 SDValue Addr = Op.getOperand(2);
17045 SDValue Chain = Op.getOperand(0);
17047 EVT VT = DataToCompress.getValueType();
17048 if (isAllOnes(Mask)) // return just a store
17049 return DAG.getStore(Chain, dl, DataToCompress, Addr,
17050 MachinePointerInfo(), false, false,
17051 VT.getScalarSizeInBits()/8);
17053 SDValue Compressed =
17054 getVectorMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT, DataToCompress),
17055 Mask, DAG.getUNDEF(VT), Subtarget, DAG);
17056 return DAG.getStore(Chain, dl, Compressed, Addr,
17057 MachinePointerInfo(), false, false,
17058 VT.getScalarSizeInBits()/8);
17060 case TRUNCATE_TO_MEM_VI8:
17061 return LowerINTRINSIC_TRUNCATE_TO_MEM(Op, DAG, MVT::i8);
17062 case TRUNCATE_TO_MEM_VI16:
17063 return LowerINTRINSIC_TRUNCATE_TO_MEM(Op, DAG, MVT::i16);
17064 case TRUNCATE_TO_MEM_VI32:
17065 return LowerINTRINSIC_TRUNCATE_TO_MEM(Op, DAG, MVT::i32);
17066 case EXPAND_FROM_MEM: {
17068 SDValue Mask = Op.getOperand(4);
17069 SDValue PassThru = Op.getOperand(3);
17070 SDValue Addr = Op.getOperand(2);
17071 SDValue Chain = Op.getOperand(0);
17072 EVT VT = Op.getValueType();
17074 if (isAllOnes(Mask)) // return just a load
17075 return DAG.getLoad(VT, dl, Chain, Addr, MachinePointerInfo(), false, false,
17076 false, VT.getScalarSizeInBits()/8);
17078 SDValue DataToExpand = DAG.getLoad(VT, dl, Chain, Addr, MachinePointerInfo(),
17079 false, false, false,
17080 VT.getScalarSizeInBits()/8);
17082 SDValue Results[] = {
17083 getVectorMaskingNode(DAG.getNode(IntrData->Opc0, dl, VT, DataToExpand),
17084 Mask, PassThru, Subtarget, DAG), Chain};
17085 return DAG.getMergeValues(Results, dl);
17090 SDValue X86TargetLowering::LowerRETURNADDR(SDValue Op,
17091 SelectionDAG &DAG) const {
17092 MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
17093 MFI->setReturnAddressIsTaken(true);
17095 if (verifyReturnAddressArgumentIsConstant(Op, DAG))
17098 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
17100 EVT PtrVT = getPointerTy(DAG.getDataLayout());
17103 SDValue FrameAddr = LowerFRAMEADDR(Op, DAG);
17104 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
17105 SDValue Offset = DAG.getConstant(RegInfo->getSlotSize(), dl, PtrVT);
17106 return DAG.getLoad(PtrVT, dl, DAG.getEntryNode(),
17107 DAG.getNode(ISD::ADD, dl, PtrVT,
17108 FrameAddr, Offset),
17109 MachinePointerInfo(), false, false, false, 0);
17112 // Just load the return address.
17113 SDValue RetAddrFI = getReturnAddressFrameIndex(DAG);
17114 return DAG.getLoad(PtrVT, dl, DAG.getEntryNode(),
17115 RetAddrFI, MachinePointerInfo(), false, false, false, 0);
17118 SDValue X86TargetLowering::LowerFRAMEADDR(SDValue Op, SelectionDAG &DAG) const {
17119 MachineFunction &MF = DAG.getMachineFunction();
17120 MachineFrameInfo *MFI = MF.getFrameInfo();
17121 X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
17122 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
17123 EVT VT = Op.getValueType();
17125 MFI->setFrameAddressIsTaken(true);
17127 if (MF.getTarget().getMCAsmInfo()->usesWindowsCFI()) {
17128 // Depth > 0 makes no sense on targets which use Windows unwind codes. It
17129 // is not possible to crawl up the stack without looking at the unwind codes
17131 int FrameAddrIndex = FuncInfo->getFAIndex();
17132 if (!FrameAddrIndex) {
17133 // Set up a frame object for the return address.
17134 unsigned SlotSize = RegInfo->getSlotSize();
17135 FrameAddrIndex = MF.getFrameInfo()->CreateFixedObject(
17136 SlotSize, /*Offset=*/0, /*IsImmutable=*/false);
17137 FuncInfo->setFAIndex(FrameAddrIndex);
17139 return DAG.getFrameIndex(FrameAddrIndex, VT);
17142 unsigned FrameReg =
17143 RegInfo->getPtrSizedFrameRegister(DAG.getMachineFunction());
17144 SDLoc dl(Op); // FIXME probably not meaningful
17145 unsigned Depth = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
17146 assert(((FrameReg == X86::RBP && VT == MVT::i64) ||
17147 (FrameReg == X86::EBP && VT == MVT::i32)) &&
17148 "Invalid Frame Register!");
17149 SDValue FrameAddr = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, VT);
17151 FrameAddr = DAG.getLoad(VT, dl, DAG.getEntryNode(), FrameAddr,
17152 MachinePointerInfo(),
17153 false, false, false, 0);
17157 // FIXME? Maybe this could be a TableGen attribute on some registers and
17158 // this table could be generated automatically from RegInfo.
17159 unsigned X86TargetLowering::getRegisterByName(const char* RegName, EVT VT,
17160 SelectionDAG &DAG) const {
17161 const TargetFrameLowering &TFI = *Subtarget->getFrameLowering();
17162 const MachineFunction &MF = DAG.getMachineFunction();
17164 unsigned Reg = StringSwitch<unsigned>(RegName)
17165 .Case("esp", X86::ESP)
17166 .Case("rsp", X86::RSP)
17167 .Case("ebp", X86::EBP)
17168 .Case("rbp", X86::RBP)
17171 if (Reg == X86::EBP || Reg == X86::RBP) {
17172 if (!TFI.hasFP(MF))
17173 report_fatal_error("register " + StringRef(RegName) +
17174 " is allocatable: function has no frame pointer");
17177 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
17178 unsigned FrameReg =
17179 RegInfo->getPtrSizedFrameRegister(DAG.getMachineFunction());
17180 assert((FrameReg == X86::EBP || FrameReg == X86::RBP) &&
17181 "Invalid Frame Register!");
17189 report_fatal_error("Invalid register name global variable");
17192 SDValue X86TargetLowering::LowerFRAME_TO_ARGS_OFFSET(SDValue Op,
17193 SelectionDAG &DAG) const {
17194 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
17195 return DAG.getIntPtrConstant(2 * RegInfo->getSlotSize(), SDLoc(Op));
17198 SDValue X86TargetLowering::LowerEH_RETURN(SDValue Op, SelectionDAG &DAG) const {
17199 SDValue Chain = Op.getOperand(0);
17200 SDValue Offset = Op.getOperand(1);
17201 SDValue Handler = Op.getOperand(2);
17204 EVT PtrVT = getPointerTy(DAG.getDataLayout());
17205 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
17206 unsigned FrameReg = RegInfo->getFrameRegister(DAG.getMachineFunction());
17207 assert(((FrameReg == X86::RBP && PtrVT == MVT::i64) ||
17208 (FrameReg == X86::EBP && PtrVT == MVT::i32)) &&
17209 "Invalid Frame Register!");
17210 SDValue Frame = DAG.getCopyFromReg(DAG.getEntryNode(), dl, FrameReg, PtrVT);
17211 unsigned StoreAddrReg = (PtrVT == MVT::i64) ? X86::RCX : X86::ECX;
17213 SDValue StoreAddr = DAG.getNode(ISD::ADD, dl, PtrVT, Frame,
17214 DAG.getIntPtrConstant(RegInfo->getSlotSize(),
17216 StoreAddr = DAG.getNode(ISD::ADD, dl, PtrVT, StoreAddr, Offset);
17217 Chain = DAG.getStore(Chain, dl, Handler, StoreAddr, MachinePointerInfo(),
17219 Chain = DAG.getCopyToReg(Chain, dl, StoreAddrReg, StoreAddr);
17221 return DAG.getNode(X86ISD::EH_RETURN, dl, MVT::Other, Chain,
17222 DAG.getRegister(StoreAddrReg, PtrVT));
17225 SDValue X86TargetLowering::lowerEH_SJLJ_SETJMP(SDValue Op,
17226 SelectionDAG &DAG) const {
17228 return DAG.getNode(X86ISD::EH_SJLJ_SETJMP, DL,
17229 DAG.getVTList(MVT::i32, MVT::Other),
17230 Op.getOperand(0), Op.getOperand(1));
17233 SDValue X86TargetLowering::lowerEH_SJLJ_LONGJMP(SDValue Op,
17234 SelectionDAG &DAG) const {
17236 return DAG.getNode(X86ISD::EH_SJLJ_LONGJMP, DL, MVT::Other,
17237 Op.getOperand(0), Op.getOperand(1));
17240 static SDValue LowerADJUST_TRAMPOLINE(SDValue Op, SelectionDAG &DAG) {
17241 return Op.getOperand(0);
17244 SDValue X86TargetLowering::LowerINIT_TRAMPOLINE(SDValue Op,
17245 SelectionDAG &DAG) const {
17246 SDValue Root = Op.getOperand(0);
17247 SDValue Trmp = Op.getOperand(1); // trampoline
17248 SDValue FPtr = Op.getOperand(2); // nested function
17249 SDValue Nest = Op.getOperand(3); // 'nest' parameter value
17252 const Value *TrmpAddr = cast<SrcValueSDNode>(Op.getOperand(4))->getValue();
17253 const TargetRegisterInfo *TRI = Subtarget->getRegisterInfo();
17255 if (Subtarget->is64Bit()) {
17256 SDValue OutChains[6];
17258 // Large code-model.
17259 const unsigned char JMP64r = 0xFF; // 64-bit jmp through register opcode.
17260 const unsigned char MOV64ri = 0xB8; // X86::MOV64ri opcode.
17262 const unsigned char N86R10 = TRI->getEncodingValue(X86::R10) & 0x7;
17263 const unsigned char N86R11 = TRI->getEncodingValue(X86::R11) & 0x7;
17265 const unsigned char REX_WB = 0x40 | 0x08 | 0x01; // REX prefix
17267 // Load the pointer to the nested function into R11.
17268 unsigned OpCode = ((MOV64ri | N86R11) << 8) | REX_WB; // movabsq r11
17269 SDValue Addr = Trmp;
17270 OutChains[0] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, dl, MVT::i16),
17271 Addr, MachinePointerInfo(TrmpAddr),
17274 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
17275 DAG.getConstant(2, dl, MVT::i64));
17276 OutChains[1] = DAG.getStore(Root, dl, FPtr, Addr,
17277 MachinePointerInfo(TrmpAddr, 2),
17280 // Load the 'nest' parameter value into R10.
17281 // R10 is specified in X86CallingConv.td
17282 OpCode = ((MOV64ri | N86R10) << 8) | REX_WB; // movabsq r10
17283 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
17284 DAG.getConstant(10, dl, MVT::i64));
17285 OutChains[2] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, dl, MVT::i16),
17286 Addr, MachinePointerInfo(TrmpAddr, 10),
17289 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
17290 DAG.getConstant(12, dl, MVT::i64));
17291 OutChains[3] = DAG.getStore(Root, dl, Nest, Addr,
17292 MachinePointerInfo(TrmpAddr, 12),
17295 // Jump to the nested function.
17296 OpCode = (JMP64r << 8) | REX_WB; // jmpq *...
17297 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
17298 DAG.getConstant(20, dl, MVT::i64));
17299 OutChains[4] = DAG.getStore(Root, dl, DAG.getConstant(OpCode, dl, MVT::i16),
17300 Addr, MachinePointerInfo(TrmpAddr, 20),
17303 unsigned char ModRM = N86R11 | (4 << 3) | (3 << 6); // ...r11
17304 Addr = DAG.getNode(ISD::ADD, dl, MVT::i64, Trmp,
17305 DAG.getConstant(22, dl, MVT::i64));
17306 OutChains[5] = DAG.getStore(Root, dl, DAG.getConstant(ModRM, dl, MVT::i8),
17307 Addr, MachinePointerInfo(TrmpAddr, 22),
17310 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other, OutChains);
17312 const Function *Func =
17313 cast<Function>(cast<SrcValueSDNode>(Op.getOperand(5))->getValue());
17314 CallingConv::ID CC = Func->getCallingConv();
17319 llvm_unreachable("Unsupported calling convention");
17320 case CallingConv::C:
17321 case CallingConv::X86_StdCall: {
17322 // Pass 'nest' parameter in ECX.
17323 // Must be kept in sync with X86CallingConv.td
17324 NestReg = X86::ECX;
17326 // Check that ECX wasn't needed by an 'inreg' parameter.
17327 FunctionType *FTy = Func->getFunctionType();
17328 const AttributeSet &Attrs = Func->getAttributes();
17330 if (!Attrs.isEmpty() && !Func->isVarArg()) {
17331 unsigned InRegCount = 0;
17334 for (FunctionType::param_iterator I = FTy->param_begin(),
17335 E = FTy->param_end(); I != E; ++I, ++Idx)
17336 if (Attrs.hasAttribute(Idx, Attribute::InReg)) {
17337 auto &DL = DAG.getDataLayout();
17338 // FIXME: should only count parameters that are lowered to integers.
17339 InRegCount += (DL.getTypeSizeInBits(*I) + 31) / 32;
17342 if (InRegCount > 2) {
17343 report_fatal_error("Nest register in use - reduce number of inreg"
17349 case CallingConv::X86_FastCall:
17350 case CallingConv::X86_ThisCall:
17351 case CallingConv::Fast:
17352 // Pass 'nest' parameter in EAX.
17353 // Must be kept in sync with X86CallingConv.td
17354 NestReg = X86::EAX;
17358 SDValue OutChains[4];
17359 SDValue Addr, Disp;
17361 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
17362 DAG.getConstant(10, dl, MVT::i32));
17363 Disp = DAG.getNode(ISD::SUB, dl, MVT::i32, FPtr, Addr);
17365 // This is storing the opcode for MOV32ri.
17366 const unsigned char MOV32ri = 0xB8; // X86::MOV32ri's opcode byte.
17367 const unsigned char N86Reg = TRI->getEncodingValue(NestReg) & 0x7;
17368 OutChains[0] = DAG.getStore(Root, dl,
17369 DAG.getConstant(MOV32ri|N86Reg, dl, MVT::i8),
17370 Trmp, MachinePointerInfo(TrmpAddr),
17373 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
17374 DAG.getConstant(1, dl, MVT::i32));
17375 OutChains[1] = DAG.getStore(Root, dl, Nest, Addr,
17376 MachinePointerInfo(TrmpAddr, 1),
17379 const unsigned char JMP = 0xE9; // jmp <32bit dst> opcode.
17380 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
17381 DAG.getConstant(5, dl, MVT::i32));
17382 OutChains[2] = DAG.getStore(Root, dl, DAG.getConstant(JMP, dl, MVT::i8),
17383 Addr, MachinePointerInfo(TrmpAddr, 5),
17386 Addr = DAG.getNode(ISD::ADD, dl, MVT::i32, Trmp,
17387 DAG.getConstant(6, dl, MVT::i32));
17388 OutChains[3] = DAG.getStore(Root, dl, Disp, Addr,
17389 MachinePointerInfo(TrmpAddr, 6),
17392 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other, OutChains);
17396 SDValue X86TargetLowering::LowerFLT_ROUNDS_(SDValue Op,
17397 SelectionDAG &DAG) const {
17399 The rounding mode is in bits 11:10 of FPSR, and has the following
17401 00 Round to nearest
17406 FLT_ROUNDS, on the other hand, expects the following:
17413 To perform the conversion, we do:
17414 (((((FPSR & 0x800) >> 11) | ((FPSR & 0x400) >> 9)) + 1) & 3)
17417 MachineFunction &MF = DAG.getMachineFunction();
17418 const TargetFrameLowering &TFI = *Subtarget->getFrameLowering();
17419 unsigned StackAlignment = TFI.getStackAlignment();
17420 MVT VT = Op.getSimpleValueType();
17423 // Save FP Control Word to stack slot
17424 int SSFI = MF.getFrameInfo()->CreateStackObject(2, StackAlignment, false);
17425 SDValue StackSlot =
17426 DAG.getFrameIndex(SSFI, getPointerTy(DAG.getDataLayout()));
17428 MachineMemOperand *MMO =
17429 MF.getMachineMemOperand(MachinePointerInfo::getFixedStack(MF, SSFI),
17430 MachineMemOperand::MOStore, 2, 2);
17432 SDValue Ops[] = { DAG.getEntryNode(), StackSlot };
17433 SDValue Chain = DAG.getMemIntrinsicNode(X86ISD::FNSTCW16m, DL,
17434 DAG.getVTList(MVT::Other),
17435 Ops, MVT::i16, MMO);
17437 // Load FP Control Word from stack slot
17438 SDValue CWD = DAG.getLoad(MVT::i16, DL, Chain, StackSlot,
17439 MachinePointerInfo(), false, false, false, 0);
17441 // Transform as necessary
17443 DAG.getNode(ISD::SRL, DL, MVT::i16,
17444 DAG.getNode(ISD::AND, DL, MVT::i16,
17445 CWD, DAG.getConstant(0x800, DL, MVT::i16)),
17446 DAG.getConstant(11, DL, MVT::i8));
17448 DAG.getNode(ISD::SRL, DL, MVT::i16,
17449 DAG.getNode(ISD::AND, DL, MVT::i16,
17450 CWD, DAG.getConstant(0x400, DL, MVT::i16)),
17451 DAG.getConstant(9, DL, MVT::i8));
17454 DAG.getNode(ISD::AND, DL, MVT::i16,
17455 DAG.getNode(ISD::ADD, DL, MVT::i16,
17456 DAG.getNode(ISD::OR, DL, MVT::i16, CWD1, CWD2),
17457 DAG.getConstant(1, DL, MVT::i16)),
17458 DAG.getConstant(3, DL, MVT::i16));
17460 return DAG.getNode((VT.getSizeInBits() < 16 ?
17461 ISD::TRUNCATE : ISD::ZERO_EXTEND), DL, VT, RetVal);
17464 /// \brief Lower a vector CTLZ using native supported vector CTLZ instruction.
17466 // 1. i32/i64 128/256-bit vector (native support require VLX) are expended
17467 // to 512-bit vector.
17468 // 2. i8/i16 vector implemented using dword LZCNT vector instruction
17469 // ( sub(trunc(lzcnt(zext32(x)))) ). In case zext32(x) is illegal,
17470 // split the vector, perform operation on it's Lo a Hi part and
17471 // concatenate the results.
17472 static SDValue LowerVectorCTLZ_AVX512(SDValue Op, SelectionDAG &DAG) {
17474 MVT VT = Op.getSimpleValueType();
17475 MVT EltVT = VT.getVectorElementType();
17476 unsigned NumElems = VT.getVectorNumElements();
17478 if (EltVT == MVT::i64 || EltVT == MVT::i32) {
17479 // Extend to 512 bit vector.
17480 assert((VT.is256BitVector() || VT.is128BitVector()) &&
17481 "Unsupported value type for operation");
17483 MVT NewVT = MVT::getVectorVT(EltVT, 512 / VT.getScalarSizeInBits());
17484 SDValue Vec512 = DAG.getNode(ISD::INSERT_SUBVECTOR, dl, NewVT,
17485 DAG.getUNDEF(NewVT),
17487 DAG.getIntPtrConstant(0, dl));
17488 SDValue CtlzNode = DAG.getNode(ISD::CTLZ, dl, NewVT, Vec512);
17490 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT, CtlzNode,
17491 DAG.getIntPtrConstant(0, dl));
17494 assert((EltVT == MVT::i8 || EltVT == MVT::i16) &&
17495 "Unsupported element type");
17497 if (16 < NumElems) {
17498 // Split vector, it's Lo and Hi parts will be handled in next iteration.
17500 std::tie(Lo, Hi) = DAG.SplitVector(Op.getOperand(0), dl);
17501 MVT OutVT = MVT::getVectorVT(EltVT, NumElems/2);
17503 Lo = DAG.getNode(Op.getOpcode(), dl, OutVT, Lo);
17504 Hi = DAG.getNode(Op.getOpcode(), dl, OutVT, Hi);
17506 return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT, Lo, Hi);
17509 MVT NewVT = MVT::getVectorVT(MVT::i32, NumElems);
17511 assert((NewVT.is256BitVector() || NewVT.is512BitVector()) &&
17512 "Unsupported value type for operation");
17514 // Use native supported vector instruction vplzcntd.
17515 Op = DAG.getNode(ISD::ZERO_EXTEND, dl, NewVT, Op.getOperand(0));
17516 SDValue CtlzNode = DAG.getNode(ISD::CTLZ, dl, NewVT, Op);
17517 SDValue TruncNode = DAG.getNode(ISD::TRUNCATE, dl, VT, CtlzNode);
17518 SDValue Delta = DAG.getConstant(32 - EltVT.getSizeInBits(), dl, VT);
17520 return DAG.getNode(ISD::SUB, dl, VT, TruncNode, Delta);
17523 static SDValue LowerCTLZ(SDValue Op, const X86Subtarget *Subtarget,
17524 SelectionDAG &DAG) {
17525 MVT VT = Op.getSimpleValueType();
17527 unsigned NumBits = VT.getSizeInBits();
17530 if (VT.isVector() && Subtarget->hasAVX512())
17531 return LowerVectorCTLZ_AVX512(Op, DAG);
17533 Op = Op.getOperand(0);
17534 if (VT == MVT::i8) {
17535 // Zero extend to i32 since there is not an i8 bsr.
17537 Op = DAG.getNode(ISD::ZERO_EXTEND, dl, OpVT, Op);
17540 // Issue a bsr (scan bits in reverse) which also sets EFLAGS.
17541 SDVTList VTs = DAG.getVTList(OpVT, MVT::i32);
17542 Op = DAG.getNode(X86ISD::BSR, dl, VTs, Op);
17544 // If src is zero (i.e. bsr sets ZF), returns NumBits.
17547 DAG.getConstant(NumBits + NumBits - 1, dl, OpVT),
17548 DAG.getConstant(X86::COND_E, dl, MVT::i8),
17551 Op = DAG.getNode(X86ISD::CMOV, dl, OpVT, Ops);
17553 // Finally xor with NumBits-1.
17554 Op = DAG.getNode(ISD::XOR, dl, OpVT, Op,
17555 DAG.getConstant(NumBits - 1, dl, OpVT));
17558 Op = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, Op);
17562 static SDValue LowerCTLZ_ZERO_UNDEF(SDValue Op, const X86Subtarget *Subtarget,
17563 SelectionDAG &DAG) {
17564 MVT VT = Op.getSimpleValueType();
17566 unsigned NumBits = VT.getSizeInBits();
17569 if (VT.isVector() && Subtarget->hasAVX512())
17570 return LowerVectorCTLZ_AVX512(Op, DAG);
17572 Op = Op.getOperand(0);
17573 if (VT == MVT::i8) {
17574 // Zero extend to i32 since there is not an i8 bsr.
17576 Op = DAG.getNode(ISD::ZERO_EXTEND, dl, OpVT, Op);
17579 // Issue a bsr (scan bits in reverse).
17580 SDVTList VTs = DAG.getVTList(OpVT, MVT::i32);
17581 Op = DAG.getNode(X86ISD::BSR, dl, VTs, Op);
17583 // And 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 LowerCTTZ(SDValue Op, SelectionDAG &DAG) {
17593 MVT VT = Op.getSimpleValueType();
17594 unsigned NumBits = VT.getScalarSizeInBits();
17597 if (VT.isVector()) {
17598 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
17600 SDValue N0 = Op.getOperand(0);
17601 SDValue Zero = DAG.getConstant(0, dl, VT);
17603 // lsb(x) = (x & -x)
17604 SDValue LSB = DAG.getNode(ISD::AND, dl, VT, N0,
17605 DAG.getNode(ISD::SUB, dl, VT, Zero, N0));
17607 // cttz_undef(x) = (width - 1) - ctlz(lsb)
17608 if (Op.getOpcode() == ISD::CTTZ_ZERO_UNDEF &&
17609 TLI.isOperationLegal(ISD::CTLZ, VT)) {
17610 SDValue WidthMinusOne = DAG.getConstant(NumBits - 1, dl, VT);
17611 return DAG.getNode(ISD::SUB, dl, VT, WidthMinusOne,
17612 DAG.getNode(ISD::CTLZ, dl, VT, LSB));
17615 // cttz(x) = ctpop(lsb - 1)
17616 SDValue One = DAG.getConstant(1, dl, VT);
17617 return DAG.getNode(ISD::CTPOP, dl, VT,
17618 DAG.getNode(ISD::SUB, dl, VT, LSB, One));
17621 assert(Op.getOpcode() == ISD::CTTZ &&
17622 "Only scalar CTTZ requires custom lowering");
17624 // Issue a bsf (scan bits forward) which also sets EFLAGS.
17625 SDVTList VTs = DAG.getVTList(VT, MVT::i32);
17626 Op = DAG.getNode(X86ISD::BSF, dl, VTs, Op.getOperand(0));
17628 // If src is zero (i.e. bsf sets ZF), returns NumBits.
17631 DAG.getConstant(NumBits, dl, VT),
17632 DAG.getConstant(X86::COND_E, dl, MVT::i8),
17635 return DAG.getNode(X86ISD::CMOV, dl, VT, Ops);
17638 // Lower256IntArith - Break a 256-bit integer operation into two new 128-bit
17639 // ones, and then concatenate the result back.
17640 static SDValue Lower256IntArith(SDValue Op, SelectionDAG &DAG) {
17641 MVT VT = Op.getSimpleValueType();
17643 assert(VT.is256BitVector() && VT.isInteger() &&
17644 "Unsupported value type for operation");
17646 unsigned NumElems = VT.getVectorNumElements();
17649 // Extract the LHS vectors
17650 SDValue LHS = Op.getOperand(0);
17651 SDValue LHS1 = Extract128BitVector(LHS, 0, DAG, dl);
17652 SDValue LHS2 = Extract128BitVector(LHS, NumElems/2, DAG, dl);
17654 // Extract the RHS vectors
17655 SDValue RHS = Op.getOperand(1);
17656 SDValue RHS1 = Extract128BitVector(RHS, 0, DAG, dl);
17657 SDValue RHS2 = Extract128BitVector(RHS, NumElems/2, DAG, dl);
17659 MVT EltVT = VT.getVectorElementType();
17660 MVT NewVT = MVT::getVectorVT(EltVT, NumElems/2);
17662 return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT,
17663 DAG.getNode(Op.getOpcode(), dl, NewVT, LHS1, RHS1),
17664 DAG.getNode(Op.getOpcode(), dl, NewVT, LHS2, RHS2));
17667 static SDValue LowerADD(SDValue Op, SelectionDAG &DAG) {
17668 if (Op.getValueType() == MVT::i1)
17669 return DAG.getNode(ISD::XOR, SDLoc(Op), Op.getValueType(),
17670 Op.getOperand(0), Op.getOperand(1));
17671 assert(Op.getSimpleValueType().is256BitVector() &&
17672 Op.getSimpleValueType().isInteger() &&
17673 "Only handle AVX 256-bit vector integer operation");
17674 return Lower256IntArith(Op, DAG);
17677 static SDValue LowerSUB(SDValue Op, SelectionDAG &DAG) {
17678 if (Op.getValueType() == MVT::i1)
17679 return DAG.getNode(ISD::XOR, SDLoc(Op), Op.getValueType(),
17680 Op.getOperand(0), Op.getOperand(1));
17681 assert(Op.getSimpleValueType().is256BitVector() &&
17682 Op.getSimpleValueType().isInteger() &&
17683 "Only handle AVX 256-bit vector integer operation");
17684 return Lower256IntArith(Op, DAG);
17687 static SDValue LowerMINMAX(SDValue Op, SelectionDAG &DAG) {
17688 assert(Op.getSimpleValueType().is256BitVector() &&
17689 Op.getSimpleValueType().isInteger() &&
17690 "Only handle AVX 256-bit vector integer operation");
17691 return Lower256IntArith(Op, DAG);
17694 static SDValue LowerMUL(SDValue Op, const X86Subtarget *Subtarget,
17695 SelectionDAG &DAG) {
17697 MVT VT = Op.getSimpleValueType();
17700 return DAG.getNode(ISD::AND, dl, VT, Op.getOperand(0), Op.getOperand(1));
17702 // Decompose 256-bit ops into smaller 128-bit ops.
17703 if (VT.is256BitVector() && !Subtarget->hasInt256())
17704 return Lower256IntArith(Op, DAG);
17706 SDValue A = Op.getOperand(0);
17707 SDValue B = Op.getOperand(1);
17709 // Lower v16i8/v32i8 mul as promotion to v8i16/v16i16 vector
17710 // pairs, multiply and truncate.
17711 if (VT == MVT::v16i8 || VT == MVT::v32i8) {
17712 if (Subtarget->hasInt256()) {
17713 if (VT == MVT::v32i8) {
17714 MVT SubVT = MVT::getVectorVT(MVT::i8, VT.getVectorNumElements() / 2);
17715 SDValue Lo = DAG.getIntPtrConstant(0, dl);
17716 SDValue Hi = DAG.getIntPtrConstant(VT.getVectorNumElements() / 2, dl);
17717 SDValue ALo = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, SubVT, A, Lo);
17718 SDValue BLo = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, SubVT, B, Lo);
17719 SDValue AHi = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, SubVT, A, Hi);
17720 SDValue BHi = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, SubVT, B, Hi);
17721 return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT,
17722 DAG.getNode(ISD::MUL, dl, SubVT, ALo, BLo),
17723 DAG.getNode(ISD::MUL, dl, SubVT, AHi, BHi));
17726 MVT ExVT = MVT::getVectorVT(MVT::i16, VT.getVectorNumElements());
17727 return DAG.getNode(
17728 ISD::TRUNCATE, dl, VT,
17729 DAG.getNode(ISD::MUL, dl, ExVT,
17730 DAG.getNode(ISD::SIGN_EXTEND, dl, ExVT, A),
17731 DAG.getNode(ISD::SIGN_EXTEND, dl, ExVT, B)));
17734 assert(VT == MVT::v16i8 &&
17735 "Pre-AVX2 support only supports v16i8 multiplication");
17736 MVT ExVT = MVT::v8i16;
17738 // Extract the lo parts and sign extend to i16
17740 if (Subtarget->hasSSE41()) {
17741 ALo = DAG.getNode(X86ISD::VSEXT, dl, ExVT, A);
17742 BLo = DAG.getNode(X86ISD::VSEXT, dl, ExVT, B);
17744 const int ShufMask[] = {-1, 0, -1, 1, -1, 2, -1, 3,
17745 -1, 4, -1, 5, -1, 6, -1, 7};
17746 ALo = DAG.getVectorShuffle(VT, dl, A, A, ShufMask);
17747 BLo = DAG.getVectorShuffle(VT, dl, B, B, ShufMask);
17748 ALo = DAG.getBitcast(ExVT, ALo);
17749 BLo = DAG.getBitcast(ExVT, BLo);
17750 ALo = DAG.getNode(ISD::SRA, dl, ExVT, ALo, DAG.getConstant(8, dl, ExVT));
17751 BLo = DAG.getNode(ISD::SRA, dl, ExVT, BLo, DAG.getConstant(8, dl, ExVT));
17754 // Extract the hi parts and sign extend to i16
17756 if (Subtarget->hasSSE41()) {
17757 const int ShufMask[] = {8, 9, 10, 11, 12, 13, 14, 15,
17758 -1, -1, -1, -1, -1, -1, -1, -1};
17759 AHi = DAG.getVectorShuffle(VT, dl, A, A, ShufMask);
17760 BHi = DAG.getVectorShuffle(VT, dl, B, B, ShufMask);
17761 AHi = DAG.getNode(X86ISD::VSEXT, dl, ExVT, AHi);
17762 BHi = DAG.getNode(X86ISD::VSEXT, dl, ExVT, BHi);
17764 const int ShufMask[] = {-1, 8, -1, 9, -1, 10, -1, 11,
17765 -1, 12, -1, 13, -1, 14, -1, 15};
17766 AHi = DAG.getVectorShuffle(VT, dl, A, A, ShufMask);
17767 BHi = DAG.getVectorShuffle(VT, dl, B, B, ShufMask);
17768 AHi = DAG.getBitcast(ExVT, AHi);
17769 BHi = DAG.getBitcast(ExVT, BHi);
17770 AHi = DAG.getNode(ISD::SRA, dl, ExVT, AHi, DAG.getConstant(8, dl, ExVT));
17771 BHi = DAG.getNode(ISD::SRA, dl, ExVT, BHi, DAG.getConstant(8, dl, ExVT));
17774 // Multiply, mask the lower 8bits of the lo/hi results and pack
17775 SDValue RLo = DAG.getNode(ISD::MUL, dl, ExVT, ALo, BLo);
17776 SDValue RHi = DAG.getNode(ISD::MUL, dl, ExVT, AHi, BHi);
17777 RLo = DAG.getNode(ISD::AND, dl, ExVT, RLo, DAG.getConstant(255, dl, ExVT));
17778 RHi = DAG.getNode(ISD::AND, dl, ExVT, RHi, DAG.getConstant(255, dl, ExVT));
17779 return DAG.getNode(X86ISD::PACKUS, dl, VT, RLo, RHi);
17782 // Lower v4i32 mul as 2x shuffle, 2x pmuludq, 2x shuffle.
17783 if (VT == MVT::v4i32) {
17784 assert(Subtarget->hasSSE2() && !Subtarget->hasSSE41() &&
17785 "Should not custom lower when pmuldq is available!");
17787 // Extract the odd parts.
17788 static const int UnpackMask[] = { 1, -1, 3, -1 };
17789 SDValue Aodds = DAG.getVectorShuffle(VT, dl, A, A, UnpackMask);
17790 SDValue Bodds = DAG.getVectorShuffle(VT, dl, B, B, UnpackMask);
17792 // Multiply the even parts.
17793 SDValue Evens = DAG.getNode(X86ISD::PMULUDQ, dl, MVT::v2i64, A, B);
17794 // Now multiply odd parts.
17795 SDValue Odds = DAG.getNode(X86ISD::PMULUDQ, dl, MVT::v2i64, Aodds, Bodds);
17797 Evens = DAG.getBitcast(VT, Evens);
17798 Odds = DAG.getBitcast(VT, Odds);
17800 // Merge the two vectors back together with a shuffle. This expands into 2
17802 static const int ShufMask[] = { 0, 4, 2, 6 };
17803 return DAG.getVectorShuffle(VT, dl, Evens, Odds, ShufMask);
17806 assert((VT == MVT::v2i64 || VT == MVT::v4i64 || VT == MVT::v8i64) &&
17807 "Only know how to lower V2I64/V4I64/V8I64 multiply");
17809 // Ahi = psrlqi(a, 32);
17810 // Bhi = psrlqi(b, 32);
17812 // AloBlo = pmuludq(a, b);
17813 // AloBhi = pmuludq(a, Bhi);
17814 // AhiBlo = pmuludq(Ahi, b);
17816 // AloBhi = psllqi(AloBhi, 32);
17817 // AhiBlo = psllqi(AhiBlo, 32);
17818 // return AloBlo + AloBhi + AhiBlo;
17820 SDValue Ahi = getTargetVShiftByConstNode(X86ISD::VSRLI, dl, VT, A, 32, DAG);
17821 SDValue Bhi = getTargetVShiftByConstNode(X86ISD::VSRLI, dl, VT, B, 32, DAG);
17823 SDValue AhiBlo = Ahi;
17824 SDValue AloBhi = Bhi;
17825 // Bit cast to 32-bit vectors for MULUDQ
17826 EVT MulVT = (VT == MVT::v2i64) ? MVT::v4i32 :
17827 (VT == MVT::v4i64) ? MVT::v8i32 : MVT::v16i32;
17828 A = DAG.getBitcast(MulVT, A);
17829 B = DAG.getBitcast(MulVT, B);
17830 Ahi = DAG.getBitcast(MulVT, Ahi);
17831 Bhi = DAG.getBitcast(MulVT, Bhi);
17833 SDValue AloBlo = DAG.getNode(X86ISD::PMULUDQ, dl, VT, A, B);
17834 // After shifting right const values the result may be all-zero.
17835 if (!ISD::isBuildVectorAllZeros(Ahi.getNode())) {
17836 AhiBlo = DAG.getNode(X86ISD::PMULUDQ, dl, VT, Ahi, B);
17837 AhiBlo = getTargetVShiftByConstNode(X86ISD::VSHLI, dl, VT, AhiBlo, 32, DAG);
17839 if (!ISD::isBuildVectorAllZeros(Bhi.getNode())) {
17840 AloBhi = DAG.getNode(X86ISD::PMULUDQ, dl, VT, A, Bhi);
17841 AloBhi = getTargetVShiftByConstNode(X86ISD::VSHLI, dl, VT, AloBhi, 32, DAG);
17844 SDValue Res = DAG.getNode(ISD::ADD, dl, VT, AloBlo, AloBhi);
17845 return DAG.getNode(ISD::ADD, dl, VT, Res, AhiBlo);
17848 SDValue X86TargetLowering::LowerWin64_i128OP(SDValue Op, SelectionDAG &DAG) const {
17849 assert(Subtarget->isTargetWin64() && "Unexpected target");
17850 EVT VT = Op.getValueType();
17851 assert(VT.isInteger() && VT.getSizeInBits() == 128 &&
17852 "Unexpected return type for lowering");
17856 switch (Op->getOpcode()) {
17857 default: llvm_unreachable("Unexpected request for libcall!");
17858 case ISD::SDIV: isSigned = true; LC = RTLIB::SDIV_I128; break;
17859 case ISD::UDIV: isSigned = false; LC = RTLIB::UDIV_I128; break;
17860 case ISD::SREM: isSigned = true; LC = RTLIB::SREM_I128; break;
17861 case ISD::UREM: isSigned = false; LC = RTLIB::UREM_I128; break;
17862 case ISD::SDIVREM: isSigned = true; LC = RTLIB::SDIVREM_I128; break;
17863 case ISD::UDIVREM: isSigned = false; LC = RTLIB::UDIVREM_I128; break;
17867 SDValue InChain = DAG.getEntryNode();
17869 TargetLowering::ArgListTy Args;
17870 TargetLowering::ArgListEntry Entry;
17871 for (unsigned i = 0, e = Op->getNumOperands(); i != e; ++i) {
17872 EVT ArgVT = Op->getOperand(i).getValueType();
17873 assert(ArgVT.isInteger() && ArgVT.getSizeInBits() == 128 &&
17874 "Unexpected argument type for lowering");
17875 SDValue StackPtr = DAG.CreateStackTemporary(ArgVT, 16);
17876 Entry.Node = StackPtr;
17877 InChain = DAG.getStore(InChain, dl, Op->getOperand(i), StackPtr, MachinePointerInfo(),
17879 Type *ArgTy = ArgVT.getTypeForEVT(*DAG.getContext());
17880 Entry.Ty = PointerType::get(ArgTy,0);
17881 Entry.isSExt = false;
17882 Entry.isZExt = false;
17883 Args.push_back(Entry);
17886 SDValue Callee = DAG.getExternalSymbol(getLibcallName(LC),
17887 getPointerTy(DAG.getDataLayout()));
17889 TargetLowering::CallLoweringInfo CLI(DAG);
17890 CLI.setDebugLoc(dl).setChain(InChain)
17891 .setCallee(getLibcallCallingConv(LC),
17892 static_cast<EVT>(MVT::v2i64).getTypeForEVT(*DAG.getContext()),
17893 Callee, std::move(Args), 0)
17894 .setInRegister().setSExtResult(isSigned).setZExtResult(!isSigned);
17896 std::pair<SDValue, SDValue> CallInfo = LowerCallTo(CLI);
17897 return DAG.getBitcast(VT, CallInfo.first);
17900 static SDValue LowerMUL_LOHI(SDValue Op, const X86Subtarget *Subtarget,
17901 SelectionDAG &DAG) {
17902 SDValue Op0 = Op.getOperand(0), Op1 = Op.getOperand(1);
17903 EVT VT = Op0.getValueType();
17906 assert((VT == MVT::v4i32 && Subtarget->hasSSE2()) ||
17907 (VT == MVT::v8i32 && Subtarget->hasInt256()));
17909 // PMULxD operations multiply each even value (starting at 0) of LHS with
17910 // the related value of RHS and produce a widen result.
17911 // E.g., PMULUDQ <4 x i32> <a|b|c|d>, <4 x i32> <e|f|g|h>
17912 // => <2 x i64> <ae|cg>
17914 // In other word, to have all the results, we need to perform two PMULxD:
17915 // 1. one with the even values.
17916 // 2. one with the odd values.
17917 // To achieve #2, with need to place the odd values at an even position.
17919 // Place the odd value at an even position (basically, shift all values 1
17920 // step to the left):
17921 const int Mask[] = {1, -1, 3, -1, 5, -1, 7, -1};
17922 // <a|b|c|d> => <b|undef|d|undef>
17923 SDValue Odd0 = DAG.getVectorShuffle(VT, dl, Op0, Op0, Mask);
17924 // <e|f|g|h> => <f|undef|h|undef>
17925 SDValue Odd1 = DAG.getVectorShuffle(VT, dl, Op1, Op1, Mask);
17927 // Emit two multiplies, one for the lower 2 ints and one for the higher 2
17929 MVT MulVT = VT == MVT::v4i32 ? MVT::v2i64 : MVT::v4i64;
17930 bool IsSigned = Op->getOpcode() == ISD::SMUL_LOHI;
17932 (!IsSigned || !Subtarget->hasSSE41()) ? X86ISD::PMULUDQ : X86ISD::PMULDQ;
17933 // PMULUDQ <4 x i32> <a|b|c|d>, <4 x i32> <e|f|g|h>
17934 // => <2 x i64> <ae|cg>
17935 SDValue Mul1 = DAG.getBitcast(VT, DAG.getNode(Opcode, dl, MulVT, Op0, Op1));
17936 // PMULUDQ <4 x i32> <b|undef|d|undef>, <4 x i32> <f|undef|h|undef>
17937 // => <2 x i64> <bf|dh>
17938 SDValue Mul2 = DAG.getBitcast(VT, DAG.getNode(Opcode, dl, MulVT, Odd0, Odd1));
17940 // Shuffle it back into the right order.
17941 SDValue Highs, Lows;
17942 if (VT == MVT::v8i32) {
17943 const int HighMask[] = {1, 9, 3, 11, 5, 13, 7, 15};
17944 Highs = DAG.getVectorShuffle(VT, dl, Mul1, Mul2, HighMask);
17945 const int LowMask[] = {0, 8, 2, 10, 4, 12, 6, 14};
17946 Lows = DAG.getVectorShuffle(VT, dl, Mul1, Mul2, LowMask);
17948 const int HighMask[] = {1, 5, 3, 7};
17949 Highs = DAG.getVectorShuffle(VT, dl, Mul1, Mul2, HighMask);
17950 const int LowMask[] = {0, 4, 2, 6};
17951 Lows = DAG.getVectorShuffle(VT, dl, Mul1, Mul2, LowMask);
17954 // If we have a signed multiply but no PMULDQ fix up the high parts of a
17955 // unsigned multiply.
17956 if (IsSigned && !Subtarget->hasSSE41()) {
17957 SDValue ShAmt = DAG.getConstant(
17959 DAG.getTargetLoweringInfo().getShiftAmountTy(VT, DAG.getDataLayout()));
17960 SDValue T1 = DAG.getNode(ISD::AND, dl, VT,
17961 DAG.getNode(ISD::SRA, dl, VT, Op0, ShAmt), Op1);
17962 SDValue T2 = DAG.getNode(ISD::AND, dl, VT,
17963 DAG.getNode(ISD::SRA, dl, VT, Op1, ShAmt), Op0);
17965 SDValue Fixup = DAG.getNode(ISD::ADD, dl, VT, T1, T2);
17966 Highs = DAG.getNode(ISD::SUB, dl, VT, Highs, Fixup);
17969 // The first result of MUL_LOHI is actually the low value, followed by the
17971 SDValue Ops[] = {Lows, Highs};
17972 return DAG.getMergeValues(Ops, dl);
17975 // Return true if the required (according to Opcode) shift-imm form is natively
17976 // supported by the Subtarget
17977 static bool SupportedVectorShiftWithImm(MVT VT, const X86Subtarget *Subtarget,
17979 if (VT.getScalarSizeInBits() < 16)
17982 if (VT.is512BitVector() &&
17983 (VT.getScalarSizeInBits() > 16 || Subtarget->hasBWI()))
17986 bool LShift = VT.is128BitVector() ||
17987 (VT.is256BitVector() && Subtarget->hasInt256());
17989 bool AShift = LShift && (Subtarget->hasVLX() ||
17990 (VT != MVT::v2i64 && VT != MVT::v4i64));
17991 return (Opcode == ISD::SRA) ? AShift : LShift;
17994 // The shift amount is a variable, but it is the same for all vector lanes.
17995 // These instructions are defined together with shift-immediate.
17997 bool SupportedVectorShiftWithBaseAmnt(MVT VT, const X86Subtarget *Subtarget,
17999 return SupportedVectorShiftWithImm(VT, Subtarget, Opcode);
18002 // Return true if the required (according to Opcode) variable-shift form is
18003 // natively supported by the Subtarget
18004 static bool SupportedVectorVarShift(MVT VT, const X86Subtarget *Subtarget,
18007 if (!Subtarget->hasInt256() || VT.getScalarSizeInBits() < 16)
18010 // vXi16 supported only on AVX-512, BWI
18011 if (VT.getScalarSizeInBits() == 16 && !Subtarget->hasBWI())
18014 if (VT.is512BitVector() || Subtarget->hasVLX())
18017 bool LShift = VT.is128BitVector() || VT.is256BitVector();
18018 bool AShift = LShift && VT != MVT::v2i64 && VT != MVT::v4i64;
18019 return (Opcode == ISD::SRA) ? AShift : LShift;
18022 static SDValue LowerScalarImmediateShift(SDValue Op, SelectionDAG &DAG,
18023 const X86Subtarget *Subtarget) {
18024 MVT VT = Op.getSimpleValueType();
18026 SDValue R = Op.getOperand(0);
18027 SDValue Amt = Op.getOperand(1);
18029 unsigned X86Opc = (Op.getOpcode() == ISD::SHL) ? X86ISD::VSHLI :
18030 (Op.getOpcode() == ISD::SRL) ? X86ISD::VSRLI : X86ISD::VSRAI;
18032 auto ArithmeticShiftRight64 = [&](uint64_t ShiftAmt) {
18033 assert((VT == MVT::v2i64 || VT == MVT::v4i64) && "Unexpected SRA type");
18034 MVT ExVT = MVT::getVectorVT(MVT::i32, VT.getVectorNumElements() * 2);
18035 SDValue Ex = DAG.getBitcast(ExVT, R);
18037 if (ShiftAmt >= 32) {
18038 // Splat sign to upper i32 dst, and SRA upper i32 src to lower i32.
18040 getTargetVShiftByConstNode(X86ISD::VSRAI, dl, ExVT, Ex, 31, DAG);
18041 SDValue Lower = getTargetVShiftByConstNode(X86ISD::VSRAI, dl, ExVT, Ex,
18042 ShiftAmt - 32, DAG);
18043 if (VT == MVT::v2i64)
18044 Ex = DAG.getVectorShuffle(ExVT, dl, Upper, Lower, {5, 1, 7, 3});
18045 if (VT == MVT::v4i64)
18046 Ex = DAG.getVectorShuffle(ExVT, dl, Upper, Lower,
18047 {9, 1, 11, 3, 13, 5, 15, 7});
18049 // SRA upper i32, SHL whole i64 and select lower i32.
18050 SDValue Upper = getTargetVShiftByConstNode(X86ISD::VSRAI, dl, ExVT, Ex,
18053 getTargetVShiftByConstNode(X86ISD::VSRLI, dl, VT, R, ShiftAmt, DAG);
18054 Lower = DAG.getBitcast(ExVT, Lower);
18055 if (VT == MVT::v2i64)
18056 Ex = DAG.getVectorShuffle(ExVT, dl, Upper, Lower, {4, 1, 6, 3});
18057 if (VT == MVT::v4i64)
18058 Ex = DAG.getVectorShuffle(ExVT, dl, Upper, Lower,
18059 {8, 1, 10, 3, 12, 5, 14, 7});
18061 return DAG.getBitcast(VT, Ex);
18064 // Optimize shl/srl/sra with constant shift amount.
18065 if (auto *BVAmt = dyn_cast<BuildVectorSDNode>(Amt)) {
18066 if (auto *ShiftConst = BVAmt->getConstantSplatNode()) {
18067 uint64_t ShiftAmt = ShiftConst->getZExtValue();
18069 if (SupportedVectorShiftWithImm(VT, Subtarget, Op.getOpcode()))
18070 return getTargetVShiftByConstNode(X86Opc, dl, VT, R, ShiftAmt, DAG);
18072 // i64 SRA needs to be performed as partial shifts.
18073 if ((VT == MVT::v2i64 || (Subtarget->hasInt256() && VT == MVT::v4i64)) &&
18074 Op.getOpcode() == ISD::SRA && !Subtarget->hasXOP())
18075 return ArithmeticShiftRight64(ShiftAmt);
18077 if (VT == MVT::v16i8 || (Subtarget->hasInt256() && VT == MVT::v32i8)) {
18078 unsigned NumElts = VT.getVectorNumElements();
18079 MVT ShiftVT = MVT::getVectorVT(MVT::i16, NumElts / 2);
18081 // Simple i8 add case
18082 if (Op.getOpcode() == ISD::SHL && ShiftAmt == 1)
18083 return DAG.getNode(ISD::ADD, dl, VT, R, R);
18085 // ashr(R, 7) === cmp_slt(R, 0)
18086 if (Op.getOpcode() == ISD::SRA && ShiftAmt == 7) {
18087 SDValue Zeros = getZeroVector(VT, Subtarget, DAG, dl);
18088 return DAG.getNode(X86ISD::PCMPGT, dl, VT, Zeros, R);
18091 // XOP can shift v16i8 directly instead of as shift v8i16 + mask.
18092 if (VT == MVT::v16i8 && Subtarget->hasXOP())
18095 if (Op.getOpcode() == ISD::SHL) {
18096 // Make a large shift.
18097 SDValue SHL = getTargetVShiftByConstNode(X86ISD::VSHLI, dl, ShiftVT,
18099 SHL = DAG.getBitcast(VT, SHL);
18100 // Zero out the rightmost bits.
18101 SmallVector<SDValue, 32> V(
18102 NumElts, DAG.getConstant(uint8_t(-1U << ShiftAmt), dl, MVT::i8));
18103 return DAG.getNode(ISD::AND, dl, VT, SHL,
18104 DAG.getNode(ISD::BUILD_VECTOR, dl, VT, V));
18106 if (Op.getOpcode() == ISD::SRL) {
18107 // Make a large shift.
18108 SDValue SRL = getTargetVShiftByConstNode(X86ISD::VSRLI, dl, ShiftVT,
18110 SRL = DAG.getBitcast(VT, SRL);
18111 // Zero out the leftmost bits.
18112 SmallVector<SDValue, 32> V(
18113 NumElts, DAG.getConstant(uint8_t(-1U) >> ShiftAmt, dl, MVT::i8));
18114 return DAG.getNode(ISD::AND, dl, VT, SRL,
18115 DAG.getNode(ISD::BUILD_VECTOR, dl, VT, V));
18117 if (Op.getOpcode() == ISD::SRA) {
18118 // ashr(R, Amt) === sub(xor(lshr(R, Amt), Mask), Mask)
18119 SDValue Res = DAG.getNode(ISD::SRL, dl, VT, R, Amt);
18120 SmallVector<SDValue, 32> V(NumElts,
18121 DAG.getConstant(128 >> ShiftAmt, dl,
18123 SDValue Mask = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, V);
18124 Res = DAG.getNode(ISD::XOR, dl, VT, Res, Mask);
18125 Res = DAG.getNode(ISD::SUB, dl, VT, Res, Mask);
18128 llvm_unreachable("Unknown shift opcode.");
18133 // Special case in 32-bit mode, where i64 is expanded into high and low parts.
18134 if (!Subtarget->is64Bit() && !Subtarget->hasXOP() &&
18135 (VT == MVT::v2i64 || (Subtarget->hasInt256() && VT == MVT::v4i64))) {
18137 // Peek through any splat that was introduced for i64 shift vectorization.
18138 int SplatIndex = -1;
18139 if (ShuffleVectorSDNode *SVN = dyn_cast<ShuffleVectorSDNode>(Amt.getNode()))
18140 if (SVN->isSplat()) {
18141 SplatIndex = SVN->getSplatIndex();
18142 Amt = Amt.getOperand(0);
18143 assert(SplatIndex < (int)VT.getVectorNumElements() &&
18144 "Splat shuffle referencing second operand");
18147 if (Amt.getOpcode() != ISD::BITCAST ||
18148 Amt.getOperand(0).getOpcode() != ISD::BUILD_VECTOR)
18151 Amt = Amt.getOperand(0);
18152 unsigned Ratio = Amt.getSimpleValueType().getVectorNumElements() /
18153 VT.getVectorNumElements();
18154 unsigned RatioInLog2 = Log2_32_Ceil(Ratio);
18155 uint64_t ShiftAmt = 0;
18156 unsigned BaseOp = (SplatIndex < 0 ? 0 : SplatIndex * Ratio);
18157 for (unsigned i = 0; i != Ratio; ++i) {
18158 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Amt.getOperand(i + BaseOp));
18162 ShiftAmt |= C->getZExtValue() << (i * (1 << (6 - RatioInLog2)));
18165 // Check remaining shift amounts (if not a splat).
18166 if (SplatIndex < 0) {
18167 for (unsigned i = Ratio; i != Amt.getNumOperands(); i += Ratio) {
18168 uint64_t ShAmt = 0;
18169 for (unsigned j = 0; j != Ratio; ++j) {
18170 ConstantSDNode *C = dyn_cast<ConstantSDNode>(Amt.getOperand(i + j));
18174 ShAmt |= C->getZExtValue() << (j * (1 << (6 - RatioInLog2)));
18176 if (ShAmt != ShiftAmt)
18181 if (SupportedVectorShiftWithImm(VT, Subtarget, Op.getOpcode()))
18182 return getTargetVShiftByConstNode(X86Opc, dl, VT, R, ShiftAmt, DAG);
18184 if (Op.getOpcode() == ISD::SRA)
18185 return ArithmeticShiftRight64(ShiftAmt);
18191 static SDValue LowerScalarVariableShift(SDValue Op, SelectionDAG &DAG,
18192 const X86Subtarget* Subtarget) {
18193 MVT VT = Op.getSimpleValueType();
18195 SDValue R = Op.getOperand(0);
18196 SDValue Amt = Op.getOperand(1);
18198 unsigned X86OpcI = (Op.getOpcode() == ISD::SHL) ? X86ISD::VSHLI :
18199 (Op.getOpcode() == ISD::SRL) ? X86ISD::VSRLI : X86ISD::VSRAI;
18201 unsigned X86OpcV = (Op.getOpcode() == ISD::SHL) ? X86ISD::VSHL :
18202 (Op.getOpcode() == ISD::SRL) ? X86ISD::VSRL : X86ISD::VSRA;
18204 if (SupportedVectorShiftWithBaseAmnt(VT, Subtarget, Op.getOpcode())) {
18206 EVT EltVT = VT.getVectorElementType();
18208 if (BuildVectorSDNode *BV = dyn_cast<BuildVectorSDNode>(Amt)) {
18209 // Check if this build_vector node is doing a splat.
18210 // If so, then set BaseShAmt equal to the splat value.
18211 BaseShAmt = BV->getSplatValue();
18212 if (BaseShAmt && BaseShAmt.getOpcode() == ISD::UNDEF)
18213 BaseShAmt = SDValue();
18215 if (Amt.getOpcode() == ISD::EXTRACT_SUBVECTOR)
18216 Amt = Amt.getOperand(0);
18218 ShuffleVectorSDNode *SVN = dyn_cast<ShuffleVectorSDNode>(Amt);
18219 if (SVN && SVN->isSplat()) {
18220 unsigned SplatIdx = (unsigned)SVN->getSplatIndex();
18221 SDValue InVec = Amt.getOperand(0);
18222 if (InVec.getOpcode() == ISD::BUILD_VECTOR) {
18223 assert((SplatIdx < InVec.getValueType().getVectorNumElements()) &&
18224 "Unexpected shuffle index found!");
18225 BaseShAmt = InVec.getOperand(SplatIdx);
18226 } else if (InVec.getOpcode() == ISD::INSERT_VECTOR_ELT) {
18227 if (ConstantSDNode *C =
18228 dyn_cast<ConstantSDNode>(InVec.getOperand(2))) {
18229 if (C->getZExtValue() == SplatIdx)
18230 BaseShAmt = InVec.getOperand(1);
18235 // Avoid introducing an extract element from a shuffle.
18236 BaseShAmt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, EltVT, InVec,
18237 DAG.getIntPtrConstant(SplatIdx, dl));
18241 if (BaseShAmt.getNode()) {
18242 assert(EltVT.bitsLE(MVT::i64) && "Unexpected element type!");
18243 if (EltVT != MVT::i64 && EltVT.bitsGT(MVT::i32))
18244 BaseShAmt = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i64, BaseShAmt);
18245 else if (EltVT.bitsLT(MVT::i32))
18246 BaseShAmt = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::i32, BaseShAmt);
18248 return getTargetVShiftNode(X86OpcI, dl, VT, R, BaseShAmt, DAG);
18252 // Special case in 32-bit mode, where i64 is expanded into high and low parts.
18253 if (!Subtarget->is64Bit() && VT == MVT::v2i64 &&
18254 Amt.getOpcode() == ISD::BITCAST &&
18255 Amt.getOperand(0).getOpcode() == ISD::BUILD_VECTOR) {
18256 Amt = Amt.getOperand(0);
18257 unsigned Ratio = Amt.getSimpleValueType().getVectorNumElements() /
18258 VT.getVectorNumElements();
18259 std::vector<SDValue> Vals(Ratio);
18260 for (unsigned i = 0; i != Ratio; ++i)
18261 Vals[i] = Amt.getOperand(i);
18262 for (unsigned i = Ratio; i != Amt.getNumOperands(); i += Ratio) {
18263 for (unsigned j = 0; j != Ratio; ++j)
18264 if (Vals[j] != Amt.getOperand(i + j))
18268 if (SupportedVectorShiftWithBaseAmnt(VT, Subtarget, Op.getOpcode()))
18269 return DAG.getNode(X86OpcV, dl, VT, R, Op.getOperand(1));
18274 static SDValue LowerShift(SDValue Op, const X86Subtarget* Subtarget,
18275 SelectionDAG &DAG) {
18276 MVT VT = Op.getSimpleValueType();
18278 SDValue R = Op.getOperand(0);
18279 SDValue Amt = Op.getOperand(1);
18281 assert(VT.isVector() && "Custom lowering only for vector shifts!");
18282 assert(Subtarget->hasSSE2() && "Only custom lower when we have SSE2!");
18284 if (SDValue V = LowerScalarImmediateShift(Op, DAG, Subtarget))
18287 if (SDValue V = LowerScalarVariableShift(Op, DAG, Subtarget))
18290 if (SupportedVectorVarShift(VT, Subtarget, Op.getOpcode()))
18293 // XOP has 128-bit variable logical/arithmetic shifts.
18294 // +ve/-ve Amt = shift left/right.
18295 if (Subtarget->hasXOP() &&
18296 (VT == MVT::v2i64 || VT == MVT::v4i32 ||
18297 VT == MVT::v8i16 || VT == MVT::v16i8)) {
18298 if (Op.getOpcode() == ISD::SRL || Op.getOpcode() == ISD::SRA) {
18299 SDValue Zero = getZeroVector(VT, Subtarget, DAG, dl);
18300 Amt = DAG.getNode(ISD::SUB, dl, VT, Zero, Amt);
18302 if (Op.getOpcode() == ISD::SHL || Op.getOpcode() == ISD::SRL)
18303 return DAG.getNode(X86ISD::VPSHL, dl, VT, R, Amt);
18304 if (Op.getOpcode() == ISD::SRA)
18305 return DAG.getNode(X86ISD::VPSHA, dl, VT, R, Amt);
18308 // 2i64 vector logical shifts can efficiently avoid scalarization - do the
18309 // shifts per-lane and then shuffle the partial results back together.
18310 if (VT == MVT::v2i64 && Op.getOpcode() != ISD::SRA) {
18311 // Splat the shift amounts so the scalar shifts above will catch it.
18312 SDValue Amt0 = DAG.getVectorShuffle(VT, dl, Amt, Amt, {0, 0});
18313 SDValue Amt1 = DAG.getVectorShuffle(VT, dl, Amt, Amt, {1, 1});
18314 SDValue R0 = DAG.getNode(Op->getOpcode(), dl, VT, R, Amt0);
18315 SDValue R1 = DAG.getNode(Op->getOpcode(), dl, VT, R, Amt1);
18316 return DAG.getVectorShuffle(VT, dl, R0, R1, {0, 3});
18319 // i64 vector arithmetic shift can be emulated with the transform:
18320 // M = lshr(SIGN_BIT, Amt)
18321 // ashr(R, Amt) === sub(xor(lshr(R, Amt), M), M)
18322 if ((VT == MVT::v2i64 || (VT == MVT::v4i64 && Subtarget->hasInt256())) &&
18323 Op.getOpcode() == ISD::SRA) {
18324 SDValue S = DAG.getConstant(APInt::getSignBit(64), dl, VT);
18325 SDValue M = DAG.getNode(ISD::SRL, dl, VT, S, Amt);
18326 R = DAG.getNode(ISD::SRL, dl, VT, R, Amt);
18327 R = DAG.getNode(ISD::XOR, dl, VT, R, M);
18328 R = DAG.getNode(ISD::SUB, dl, VT, R, M);
18332 // If possible, lower this packed shift into a vector multiply instead of
18333 // expanding it into a sequence of scalar shifts.
18334 // Do this only if the vector shift count is a constant build_vector.
18335 if (Op.getOpcode() == ISD::SHL &&
18336 (VT == MVT::v8i16 || VT == MVT::v4i32 ||
18337 (Subtarget->hasInt256() && VT == MVT::v16i16)) &&
18338 ISD::isBuildVectorOfConstantSDNodes(Amt.getNode())) {
18339 SmallVector<SDValue, 8> Elts;
18340 EVT SVT = VT.getScalarType();
18341 unsigned SVTBits = SVT.getSizeInBits();
18342 const APInt &One = APInt(SVTBits, 1);
18343 unsigned NumElems = VT.getVectorNumElements();
18345 for (unsigned i=0; i !=NumElems; ++i) {
18346 SDValue Op = Amt->getOperand(i);
18347 if (Op->getOpcode() == ISD::UNDEF) {
18348 Elts.push_back(Op);
18352 ConstantSDNode *ND = cast<ConstantSDNode>(Op);
18353 const APInt &C = APInt(SVTBits, ND->getAPIntValue().getZExtValue());
18354 uint64_t ShAmt = C.getZExtValue();
18355 if (ShAmt >= SVTBits) {
18356 Elts.push_back(DAG.getUNDEF(SVT));
18359 Elts.push_back(DAG.getConstant(One.shl(ShAmt), dl, SVT));
18361 SDValue BV = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, Elts);
18362 return DAG.getNode(ISD::MUL, dl, VT, R, BV);
18365 // Lower SHL with variable shift amount.
18366 if (VT == MVT::v4i32 && Op->getOpcode() == ISD::SHL) {
18367 Op = DAG.getNode(ISD::SHL, dl, VT, Amt, DAG.getConstant(23, dl, VT));
18369 Op = DAG.getNode(ISD::ADD, dl, VT, Op,
18370 DAG.getConstant(0x3f800000U, dl, VT));
18371 Op = DAG.getBitcast(MVT::v4f32, Op);
18372 Op = DAG.getNode(ISD::FP_TO_SINT, dl, VT, Op);
18373 return DAG.getNode(ISD::MUL, dl, VT, Op, R);
18376 // If possible, lower this shift as a sequence of two shifts by
18377 // constant plus a MOVSS/MOVSD instead of scalarizing it.
18379 // (v4i32 (srl A, (build_vector < X, Y, Y, Y>)))
18381 // Could be rewritten as:
18382 // (v4i32 (MOVSS (srl A, <Y,Y,Y,Y>), (srl A, <X,X,X,X>)))
18384 // The advantage is that the two shifts from the example would be
18385 // lowered as X86ISD::VSRLI nodes. This would be cheaper than scalarizing
18386 // the vector shift into four scalar shifts plus four pairs of vector
18388 if ((VT == MVT::v8i16 || VT == MVT::v4i32) &&
18389 ISD::isBuildVectorOfConstantSDNodes(Amt.getNode())) {
18390 unsigned TargetOpcode = X86ISD::MOVSS;
18391 bool CanBeSimplified;
18392 // The splat value for the first packed shift (the 'X' from the example).
18393 SDValue Amt1 = Amt->getOperand(0);
18394 // The splat value for the second packed shift (the 'Y' from the example).
18395 SDValue Amt2 = (VT == MVT::v4i32) ? Amt->getOperand(1) :
18396 Amt->getOperand(2);
18398 // See if it is possible to replace this node with a sequence of
18399 // two shifts followed by a MOVSS/MOVSD
18400 if (VT == MVT::v4i32) {
18401 // Check if it is legal to use a MOVSS.
18402 CanBeSimplified = Amt2 == Amt->getOperand(2) &&
18403 Amt2 == Amt->getOperand(3);
18404 if (!CanBeSimplified) {
18405 // Otherwise, check if we can still simplify this node using a MOVSD.
18406 CanBeSimplified = Amt1 == Amt->getOperand(1) &&
18407 Amt->getOperand(2) == Amt->getOperand(3);
18408 TargetOpcode = X86ISD::MOVSD;
18409 Amt2 = Amt->getOperand(2);
18412 // Do similar checks for the case where the machine value type
18414 CanBeSimplified = Amt1 == Amt->getOperand(1);
18415 for (unsigned i=3; i != 8 && CanBeSimplified; ++i)
18416 CanBeSimplified = Amt2 == Amt->getOperand(i);
18418 if (!CanBeSimplified) {
18419 TargetOpcode = X86ISD::MOVSD;
18420 CanBeSimplified = true;
18421 Amt2 = Amt->getOperand(4);
18422 for (unsigned i=0; i != 4 && CanBeSimplified; ++i)
18423 CanBeSimplified = Amt1 == Amt->getOperand(i);
18424 for (unsigned j=4; j != 8 && CanBeSimplified; ++j)
18425 CanBeSimplified = Amt2 == Amt->getOperand(j);
18429 if (CanBeSimplified && isa<ConstantSDNode>(Amt1) &&
18430 isa<ConstantSDNode>(Amt2)) {
18431 // Replace this node with two shifts followed by a MOVSS/MOVSD.
18432 EVT CastVT = MVT::v4i32;
18434 DAG.getConstant(cast<ConstantSDNode>(Amt1)->getAPIntValue(), dl, VT);
18435 SDValue Shift1 = DAG.getNode(Op->getOpcode(), dl, VT, R, Splat1);
18437 DAG.getConstant(cast<ConstantSDNode>(Amt2)->getAPIntValue(), dl, VT);
18438 SDValue Shift2 = DAG.getNode(Op->getOpcode(), dl, VT, R, Splat2);
18439 if (TargetOpcode == X86ISD::MOVSD)
18440 CastVT = MVT::v2i64;
18441 SDValue BitCast1 = DAG.getBitcast(CastVT, Shift1);
18442 SDValue BitCast2 = DAG.getBitcast(CastVT, Shift2);
18443 SDValue Result = getTargetShuffleNode(TargetOpcode, dl, CastVT, BitCast2,
18445 return DAG.getBitcast(VT, Result);
18449 // v4i32 Non Uniform Shifts.
18450 // If the shift amount is constant we can shift each lane using the SSE2
18451 // immediate shifts, else we need to zero-extend each lane to the lower i64
18452 // and shift using the SSE2 variable shifts.
18453 // The separate results can then be blended together.
18454 if (VT == MVT::v4i32) {
18455 unsigned Opc = Op.getOpcode();
18456 SDValue Amt0, Amt1, Amt2, Amt3;
18457 if (ISD::isBuildVectorOfConstantSDNodes(Amt.getNode())) {
18458 Amt0 = DAG.getVectorShuffle(VT, dl, Amt, DAG.getUNDEF(VT), {0, 0, 0, 0});
18459 Amt1 = DAG.getVectorShuffle(VT, dl, Amt, DAG.getUNDEF(VT), {1, 1, 1, 1});
18460 Amt2 = DAG.getVectorShuffle(VT, dl, Amt, DAG.getUNDEF(VT), {2, 2, 2, 2});
18461 Amt3 = DAG.getVectorShuffle(VT, dl, Amt, DAG.getUNDEF(VT), {3, 3, 3, 3});
18463 // ISD::SHL is handled above but we include it here for completeness.
18466 llvm_unreachable("Unknown target vector shift node");
18468 Opc = X86ISD::VSHL;
18471 Opc = X86ISD::VSRL;
18474 Opc = X86ISD::VSRA;
18477 // The SSE2 shifts use the lower i64 as the same shift amount for
18478 // all lanes and the upper i64 is ignored. These shuffle masks
18479 // optimally zero-extend each lanes on SSE2/SSE41/AVX targets.
18480 SDValue Z = getZeroVector(VT, Subtarget, DAG, dl);
18481 Amt0 = DAG.getVectorShuffle(VT, dl, Amt, Z, {0, 4, -1, -1});
18482 Amt1 = DAG.getVectorShuffle(VT, dl, Amt, Z, {1, 5, -1, -1});
18483 Amt2 = DAG.getVectorShuffle(VT, dl, Amt, Z, {2, 6, -1, -1});
18484 Amt3 = DAG.getVectorShuffle(VT, dl, Amt, Z, {3, 7, -1, -1});
18487 SDValue R0 = DAG.getNode(Opc, dl, VT, R, Amt0);
18488 SDValue R1 = DAG.getNode(Opc, dl, VT, R, Amt1);
18489 SDValue R2 = DAG.getNode(Opc, dl, VT, R, Amt2);
18490 SDValue R3 = DAG.getNode(Opc, dl, VT, R, Amt3);
18491 SDValue R02 = DAG.getVectorShuffle(VT, dl, R0, R2, {0, -1, 6, -1});
18492 SDValue R13 = DAG.getVectorShuffle(VT, dl, R1, R3, {-1, 1, -1, 7});
18493 return DAG.getVectorShuffle(VT, dl, R02, R13, {0, 5, 2, 7});
18496 if (VT == MVT::v16i8 ||
18497 (VT == MVT::v32i8 && Subtarget->hasInt256() && !Subtarget->hasXOP())) {
18498 MVT ExtVT = MVT::getVectorVT(MVT::i16, VT.getVectorNumElements() / 2);
18499 unsigned ShiftOpcode = Op->getOpcode();
18501 auto SignBitSelect = [&](MVT SelVT, SDValue Sel, SDValue V0, SDValue V1) {
18502 // On SSE41 targets we make use of the fact that VSELECT lowers
18503 // to PBLENDVB which selects bytes based just on the sign bit.
18504 if (Subtarget->hasSSE41()) {
18505 V0 = DAG.getBitcast(VT, V0);
18506 V1 = DAG.getBitcast(VT, V1);
18507 Sel = DAG.getBitcast(VT, Sel);
18508 return DAG.getBitcast(SelVT,
18509 DAG.getNode(ISD::VSELECT, dl, VT, Sel, V0, V1));
18511 // On pre-SSE41 targets we test for the sign bit by comparing to
18512 // zero - a negative value will set all bits of the lanes to true
18513 // and VSELECT uses that in its OR(AND(V0,C),AND(V1,~C)) lowering.
18514 SDValue Z = getZeroVector(SelVT, Subtarget, DAG, dl);
18515 SDValue C = DAG.getNode(X86ISD::PCMPGT, dl, SelVT, Z, Sel);
18516 return DAG.getNode(ISD::VSELECT, dl, SelVT, C, V0, V1);
18519 // Turn 'a' into a mask suitable for VSELECT: a = a << 5;
18520 // We can safely do this using i16 shifts as we're only interested in
18521 // the 3 lower bits of each byte.
18522 Amt = DAG.getBitcast(ExtVT, Amt);
18523 Amt = DAG.getNode(ISD::SHL, dl, ExtVT, Amt, DAG.getConstant(5, dl, ExtVT));
18524 Amt = DAG.getBitcast(VT, Amt);
18526 if (Op->getOpcode() == ISD::SHL || Op->getOpcode() == ISD::SRL) {
18527 // r = VSELECT(r, shift(r, 4), a);
18529 DAG.getNode(ShiftOpcode, dl, VT, R, DAG.getConstant(4, dl, VT));
18530 R = SignBitSelect(VT, Amt, M, R);
18533 Amt = DAG.getNode(ISD::ADD, dl, VT, Amt, Amt);
18535 // r = VSELECT(r, shift(r, 2), a);
18536 M = DAG.getNode(ShiftOpcode, dl, VT, R, DAG.getConstant(2, dl, VT));
18537 R = SignBitSelect(VT, Amt, M, R);
18540 Amt = DAG.getNode(ISD::ADD, dl, VT, Amt, Amt);
18542 // return VSELECT(r, shift(r, 1), a);
18543 M = DAG.getNode(ShiftOpcode, dl, VT, R, DAG.getConstant(1, dl, VT));
18544 R = SignBitSelect(VT, Amt, M, R);
18548 if (Op->getOpcode() == ISD::SRA) {
18549 // For SRA we need to unpack each byte to the higher byte of a i16 vector
18550 // so we can correctly sign extend. We don't care what happens to the
18552 SDValue ALo = DAG.getNode(X86ISD::UNPCKL, dl, VT, DAG.getUNDEF(VT), Amt);
18553 SDValue AHi = DAG.getNode(X86ISD::UNPCKH, dl, VT, DAG.getUNDEF(VT), Amt);
18554 SDValue RLo = DAG.getNode(X86ISD::UNPCKL, dl, VT, DAG.getUNDEF(VT), R);
18555 SDValue RHi = DAG.getNode(X86ISD::UNPCKH, dl, VT, DAG.getUNDEF(VT), R);
18556 ALo = DAG.getBitcast(ExtVT, ALo);
18557 AHi = DAG.getBitcast(ExtVT, AHi);
18558 RLo = DAG.getBitcast(ExtVT, RLo);
18559 RHi = DAG.getBitcast(ExtVT, RHi);
18561 // r = VSELECT(r, shift(r, 4), a);
18562 SDValue MLo = DAG.getNode(ShiftOpcode, dl, ExtVT, RLo,
18563 DAG.getConstant(4, dl, ExtVT));
18564 SDValue MHi = DAG.getNode(ShiftOpcode, dl, ExtVT, RHi,
18565 DAG.getConstant(4, dl, ExtVT));
18566 RLo = SignBitSelect(ExtVT, ALo, MLo, RLo);
18567 RHi = SignBitSelect(ExtVT, AHi, MHi, RHi);
18570 ALo = DAG.getNode(ISD::ADD, dl, ExtVT, ALo, ALo);
18571 AHi = DAG.getNode(ISD::ADD, dl, ExtVT, AHi, AHi);
18573 // r = VSELECT(r, shift(r, 2), a);
18574 MLo = DAG.getNode(ShiftOpcode, dl, ExtVT, RLo,
18575 DAG.getConstant(2, dl, ExtVT));
18576 MHi = DAG.getNode(ShiftOpcode, dl, ExtVT, RHi,
18577 DAG.getConstant(2, dl, ExtVT));
18578 RLo = SignBitSelect(ExtVT, ALo, MLo, RLo);
18579 RHi = SignBitSelect(ExtVT, AHi, MHi, RHi);
18582 ALo = DAG.getNode(ISD::ADD, dl, ExtVT, ALo, ALo);
18583 AHi = DAG.getNode(ISD::ADD, dl, ExtVT, AHi, AHi);
18585 // r = VSELECT(r, shift(r, 1), a);
18586 MLo = DAG.getNode(ShiftOpcode, dl, ExtVT, RLo,
18587 DAG.getConstant(1, dl, ExtVT));
18588 MHi = DAG.getNode(ShiftOpcode, dl, ExtVT, RHi,
18589 DAG.getConstant(1, dl, ExtVT));
18590 RLo = SignBitSelect(ExtVT, ALo, MLo, RLo);
18591 RHi = SignBitSelect(ExtVT, AHi, MHi, RHi);
18593 // Logical shift the result back to the lower byte, leaving a zero upper
18595 // meaning that we can safely pack with PACKUSWB.
18597 DAG.getNode(ISD::SRL, dl, ExtVT, RLo, DAG.getConstant(8, dl, ExtVT));
18599 DAG.getNode(ISD::SRL, dl, ExtVT, RHi, DAG.getConstant(8, dl, ExtVT));
18600 return DAG.getNode(X86ISD::PACKUS, dl, VT, RLo, RHi);
18604 // It's worth extending once and using the v8i32 shifts for 16-bit types, but
18605 // the extra overheads to get from v16i8 to v8i32 make the existing SSE
18606 // solution better.
18607 if (Subtarget->hasInt256() && VT == MVT::v8i16) {
18608 MVT ExtVT = MVT::v8i32;
18610 Op.getOpcode() == ISD::SRA ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND;
18611 R = DAG.getNode(ExtOpc, dl, ExtVT, R);
18612 Amt = DAG.getNode(ISD::ANY_EXTEND, dl, ExtVT, Amt);
18613 return DAG.getNode(ISD::TRUNCATE, dl, VT,
18614 DAG.getNode(Op.getOpcode(), dl, ExtVT, R, Amt));
18617 if (Subtarget->hasInt256() && !Subtarget->hasXOP() && VT == MVT::v16i16) {
18618 MVT ExtVT = MVT::v8i32;
18619 SDValue Z = getZeroVector(VT, Subtarget, DAG, dl);
18620 SDValue ALo = DAG.getNode(X86ISD::UNPCKL, dl, VT, Amt, Z);
18621 SDValue AHi = DAG.getNode(X86ISD::UNPCKH, dl, VT, Amt, Z);
18622 SDValue RLo = DAG.getNode(X86ISD::UNPCKL, dl, VT, R, R);
18623 SDValue RHi = DAG.getNode(X86ISD::UNPCKH, dl, VT, R, R);
18624 ALo = DAG.getBitcast(ExtVT, ALo);
18625 AHi = DAG.getBitcast(ExtVT, AHi);
18626 RLo = DAG.getBitcast(ExtVT, RLo);
18627 RHi = DAG.getBitcast(ExtVT, RHi);
18628 SDValue Lo = DAG.getNode(Op.getOpcode(), dl, ExtVT, RLo, ALo);
18629 SDValue Hi = DAG.getNode(Op.getOpcode(), dl, ExtVT, RHi, AHi);
18630 Lo = DAG.getNode(ISD::SRL, dl, ExtVT, Lo, DAG.getConstant(16, dl, ExtVT));
18631 Hi = DAG.getNode(ISD::SRL, dl, ExtVT, Hi, DAG.getConstant(16, dl, ExtVT));
18632 return DAG.getNode(X86ISD::PACKUS, dl, VT, Lo, Hi);
18635 if (VT == MVT::v8i16) {
18636 unsigned ShiftOpcode = Op->getOpcode();
18638 auto SignBitSelect = [&](SDValue Sel, SDValue V0, SDValue V1) {
18639 // On SSE41 targets we make use of the fact that VSELECT lowers
18640 // to PBLENDVB which selects bytes based just on the sign bit.
18641 if (Subtarget->hasSSE41()) {
18642 MVT ExtVT = MVT::getVectorVT(MVT::i8, VT.getVectorNumElements() * 2);
18643 V0 = DAG.getBitcast(ExtVT, V0);
18644 V1 = DAG.getBitcast(ExtVT, V1);
18645 Sel = DAG.getBitcast(ExtVT, Sel);
18646 return DAG.getBitcast(
18647 VT, DAG.getNode(ISD::VSELECT, dl, ExtVT, Sel, V0, V1));
18649 // On pre-SSE41 targets we splat the sign bit - a negative value will
18650 // set all bits of the lanes to true and VSELECT uses that in
18651 // its OR(AND(V0,C),AND(V1,~C)) lowering.
18653 DAG.getNode(ISD::SRA, dl, VT, Sel, DAG.getConstant(15, dl, VT));
18654 return DAG.getNode(ISD::VSELECT, dl, VT, C, V0, V1);
18657 // Turn 'a' into a mask suitable for VSELECT: a = a << 12;
18658 if (Subtarget->hasSSE41()) {
18659 // On SSE41 targets we need to replicate the shift mask in both
18660 // bytes for PBLENDVB.
18663 DAG.getNode(ISD::SHL, dl, VT, Amt, DAG.getConstant(4, dl, VT)),
18664 DAG.getNode(ISD::SHL, dl, VT, Amt, DAG.getConstant(12, dl, VT)));
18666 Amt = DAG.getNode(ISD::SHL, dl, VT, Amt, DAG.getConstant(12, dl, VT));
18669 // r = VSELECT(r, shift(r, 8), a);
18670 SDValue M = DAG.getNode(ShiftOpcode, dl, VT, R, DAG.getConstant(8, dl, VT));
18671 R = SignBitSelect(Amt, M, R);
18674 Amt = DAG.getNode(ISD::ADD, dl, VT, Amt, Amt);
18676 // r = VSELECT(r, shift(r, 4), a);
18677 M = DAG.getNode(ShiftOpcode, dl, VT, R, DAG.getConstant(4, dl, VT));
18678 R = SignBitSelect(Amt, M, R);
18681 Amt = DAG.getNode(ISD::ADD, dl, VT, Amt, Amt);
18683 // r = VSELECT(r, shift(r, 2), a);
18684 M = DAG.getNode(ShiftOpcode, dl, VT, R, DAG.getConstant(2, dl, VT));
18685 R = SignBitSelect(Amt, M, R);
18688 Amt = DAG.getNode(ISD::ADD, dl, VT, Amt, Amt);
18690 // return VSELECT(r, shift(r, 1), a);
18691 M = DAG.getNode(ShiftOpcode, dl, VT, R, DAG.getConstant(1, dl, VT));
18692 R = SignBitSelect(Amt, M, R);
18696 // Decompose 256-bit shifts into smaller 128-bit shifts.
18697 if (VT.is256BitVector()) {
18698 unsigned NumElems = VT.getVectorNumElements();
18699 MVT EltVT = VT.getVectorElementType();
18700 EVT NewVT = MVT::getVectorVT(EltVT, NumElems/2);
18702 // Extract the two vectors
18703 SDValue V1 = Extract128BitVector(R, 0, DAG, dl);
18704 SDValue V2 = Extract128BitVector(R, NumElems/2, DAG, dl);
18706 // Recreate the shift amount vectors
18707 SDValue Amt1, Amt2;
18708 if (Amt.getOpcode() == ISD::BUILD_VECTOR) {
18709 // Constant shift amount
18710 SmallVector<SDValue, 8> Ops(Amt->op_begin(), Amt->op_begin() + NumElems);
18711 ArrayRef<SDValue> Amt1Csts = makeArrayRef(Ops).slice(0, NumElems / 2);
18712 ArrayRef<SDValue> Amt2Csts = makeArrayRef(Ops).slice(NumElems / 2);
18714 Amt1 = DAG.getNode(ISD::BUILD_VECTOR, dl, NewVT, Amt1Csts);
18715 Amt2 = DAG.getNode(ISD::BUILD_VECTOR, dl, NewVT, Amt2Csts);
18717 // Variable shift amount
18718 Amt1 = Extract128BitVector(Amt, 0, DAG, dl);
18719 Amt2 = Extract128BitVector(Amt, NumElems/2, DAG, dl);
18722 // Issue new vector shifts for the smaller types
18723 V1 = DAG.getNode(Op.getOpcode(), dl, NewVT, V1, Amt1);
18724 V2 = DAG.getNode(Op.getOpcode(), dl, NewVT, V2, Amt2);
18726 // Concatenate the result back
18727 return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT, V1, V2);
18733 static SDValue LowerRotate(SDValue Op, const X86Subtarget *Subtarget,
18734 SelectionDAG &DAG) {
18735 MVT VT = Op.getSimpleValueType();
18737 SDValue R = Op.getOperand(0);
18738 SDValue Amt = Op.getOperand(1);
18740 assert(VT.isVector() && "Custom lowering only for vector rotates!");
18741 assert(Subtarget->hasXOP() && "XOP support required for vector rotates!");
18742 assert((Op.getOpcode() == ISD::ROTL) && "Only ROTL supported");
18744 // XOP has 128-bit vector variable + immediate rotates.
18745 // +ve/-ve Amt = rotate left/right.
18747 // Split 256-bit integers.
18748 if (VT.getSizeInBits() == 256)
18749 return Lower256IntArith(Op, DAG);
18751 assert(VT.getSizeInBits() == 128 && "Only rotate 128-bit vectors!");
18753 // Attempt to rotate by immediate.
18754 if (auto *BVAmt = dyn_cast<BuildVectorSDNode>(Amt)) {
18755 if (auto *RotateConst = BVAmt->getConstantSplatNode()) {
18756 uint64_t RotateAmt = RotateConst->getAPIntValue().getZExtValue();
18757 assert(RotateAmt < VT.getScalarSizeInBits() && "Rotation out of range");
18758 return DAG.getNode(X86ISD::VPROTI, DL, VT, R,
18759 DAG.getConstant(RotateAmt, DL, MVT::i8));
18763 // Use general rotate by variable (per-element).
18764 return DAG.getNode(X86ISD::VPROT, DL, VT, R, Amt);
18767 static SDValue LowerXALUO(SDValue Op, SelectionDAG &DAG) {
18768 // Lower the "add/sub/mul with overflow" instruction into a regular ins plus
18769 // a "setcc" instruction that checks the overflow flag. The "brcond" lowering
18770 // looks for this combo and may remove the "setcc" instruction if the "setcc"
18771 // has only one use.
18772 SDNode *N = Op.getNode();
18773 SDValue LHS = N->getOperand(0);
18774 SDValue RHS = N->getOperand(1);
18775 unsigned BaseOp = 0;
18778 switch (Op.getOpcode()) {
18779 default: llvm_unreachable("Unknown ovf instruction!");
18781 // A subtract of one will be selected as a INC. Note that INC doesn't
18782 // set CF, so we can't do this for UADDO.
18783 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(RHS))
18785 BaseOp = X86ISD::INC;
18786 Cond = X86::COND_O;
18789 BaseOp = X86ISD::ADD;
18790 Cond = X86::COND_O;
18793 BaseOp = X86ISD::ADD;
18794 Cond = X86::COND_B;
18797 // A subtract of one will be selected as a DEC. Note that DEC doesn't
18798 // set CF, so we can't do this for USUBO.
18799 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(RHS))
18801 BaseOp = X86ISD::DEC;
18802 Cond = X86::COND_O;
18805 BaseOp = X86ISD::SUB;
18806 Cond = X86::COND_O;
18809 BaseOp = X86ISD::SUB;
18810 Cond = X86::COND_B;
18813 BaseOp = N->getValueType(0) == MVT::i8 ? X86ISD::SMUL8 : X86ISD::SMUL;
18814 Cond = X86::COND_O;
18816 case ISD::UMULO: { // i64, i8 = umulo lhs, rhs --> i64, i64, i32 umul lhs,rhs
18817 if (N->getValueType(0) == MVT::i8) {
18818 BaseOp = X86ISD::UMUL8;
18819 Cond = X86::COND_O;
18822 SDVTList VTs = DAG.getVTList(N->getValueType(0), N->getValueType(0),
18824 SDValue Sum = DAG.getNode(X86ISD::UMUL, DL, VTs, LHS, RHS);
18827 DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
18828 DAG.getConstant(X86::COND_O, DL, MVT::i32),
18829 SDValue(Sum.getNode(), 2));
18831 return DAG.getNode(ISD::MERGE_VALUES, DL, N->getVTList(), Sum, SetCC);
18835 // Also sets EFLAGS.
18836 SDVTList VTs = DAG.getVTList(N->getValueType(0), MVT::i32);
18837 SDValue Sum = DAG.getNode(BaseOp, DL, VTs, LHS, RHS);
18840 DAG.getNode(X86ISD::SETCC, DL, N->getValueType(1),
18841 DAG.getConstant(Cond, DL, MVT::i32),
18842 SDValue(Sum.getNode(), 1));
18844 return DAG.getNode(ISD::MERGE_VALUES, DL, N->getVTList(), Sum, SetCC);
18847 /// Returns true if the operand type is exactly twice the native width, and
18848 /// the corresponding cmpxchg8b or cmpxchg16b instruction is available.
18849 /// Used to know whether to use cmpxchg8/16b when expanding atomic operations
18850 /// (otherwise we leave them alone to become __sync_fetch_and_... calls).
18851 bool X86TargetLowering::needsCmpXchgNb(Type *MemType) const {
18852 unsigned OpWidth = MemType->getPrimitiveSizeInBits();
18855 return !Subtarget->is64Bit(); // FIXME this should be Subtarget.hasCmpxchg8b
18856 else if (OpWidth == 128)
18857 return Subtarget->hasCmpxchg16b();
18862 bool X86TargetLowering::shouldExpandAtomicStoreInIR(StoreInst *SI) const {
18863 return needsCmpXchgNb(SI->getValueOperand()->getType());
18866 // Note: this turns large loads into lock cmpxchg8b/16b.
18867 // FIXME: On 32 bits x86, fild/movq might be faster than lock cmpxchg8b.
18868 TargetLowering::AtomicExpansionKind
18869 X86TargetLowering::shouldExpandAtomicLoadInIR(LoadInst *LI) const {
18870 auto PTy = cast<PointerType>(LI->getPointerOperand()->getType());
18871 return needsCmpXchgNb(PTy->getElementType()) ? AtomicExpansionKind::CmpXChg
18872 : AtomicExpansionKind::None;
18875 TargetLowering::AtomicExpansionKind
18876 X86TargetLowering::shouldExpandAtomicRMWInIR(AtomicRMWInst *AI) const {
18877 unsigned NativeWidth = Subtarget->is64Bit() ? 64 : 32;
18878 Type *MemType = AI->getType();
18880 // If the operand is too big, we must see if cmpxchg8/16b is available
18881 // and default to library calls otherwise.
18882 if (MemType->getPrimitiveSizeInBits() > NativeWidth) {
18883 return needsCmpXchgNb(MemType) ? AtomicExpansionKind::CmpXChg
18884 : AtomicExpansionKind::None;
18887 AtomicRMWInst::BinOp Op = AI->getOperation();
18890 llvm_unreachable("Unknown atomic operation");
18891 case AtomicRMWInst::Xchg:
18892 case AtomicRMWInst::Add:
18893 case AtomicRMWInst::Sub:
18894 // It's better to use xadd, xsub or xchg for these in all cases.
18895 return AtomicExpansionKind::None;
18896 case AtomicRMWInst::Or:
18897 case AtomicRMWInst::And:
18898 case AtomicRMWInst::Xor:
18899 // If the atomicrmw's result isn't actually used, we can just add a "lock"
18900 // prefix to a normal instruction for these operations.
18901 return !AI->use_empty() ? AtomicExpansionKind::CmpXChg
18902 : AtomicExpansionKind::None;
18903 case AtomicRMWInst::Nand:
18904 case AtomicRMWInst::Max:
18905 case AtomicRMWInst::Min:
18906 case AtomicRMWInst::UMax:
18907 case AtomicRMWInst::UMin:
18908 // These always require a non-trivial set of data operations on x86. We must
18909 // use a cmpxchg loop.
18910 return AtomicExpansionKind::CmpXChg;
18914 static bool hasMFENCE(const X86Subtarget& Subtarget) {
18915 // Use mfence if we have SSE2 or we're on x86-64 (even if we asked for
18916 // no-sse2). There isn't any reason to disable it if the target processor
18918 return Subtarget.hasSSE2() || Subtarget.is64Bit();
18922 X86TargetLowering::lowerIdempotentRMWIntoFencedLoad(AtomicRMWInst *AI) const {
18923 unsigned NativeWidth = Subtarget->is64Bit() ? 64 : 32;
18924 Type *MemType = AI->getType();
18925 // Accesses larger than the native width are turned into cmpxchg/libcalls, so
18926 // there is no benefit in turning such RMWs into loads, and it is actually
18927 // harmful as it introduces a mfence.
18928 if (MemType->getPrimitiveSizeInBits() > NativeWidth)
18931 auto Builder = IRBuilder<>(AI);
18932 Module *M = Builder.GetInsertBlock()->getParent()->getParent();
18933 auto SynchScope = AI->getSynchScope();
18934 // We must restrict the ordering to avoid generating loads with Release or
18935 // ReleaseAcquire orderings.
18936 auto Order = AtomicCmpXchgInst::getStrongestFailureOrdering(AI->getOrdering());
18937 auto Ptr = AI->getPointerOperand();
18939 // Before the load we need a fence. Here is an example lifted from
18940 // http://www.hpl.hp.com/techreports/2012/HPL-2012-68.pdf showing why a fence
18943 // x.store(1, relaxed);
18944 // r1 = y.fetch_add(0, release);
18946 // y.fetch_add(42, acquire);
18947 // r2 = x.load(relaxed);
18948 // r1 = r2 = 0 is impossible, but becomes possible if the idempotent rmw is
18949 // lowered to just a load without a fence. A mfence flushes the store buffer,
18950 // making the optimization clearly correct.
18951 // FIXME: it is required if isAtLeastRelease(Order) but it is not clear
18952 // otherwise, we might be able to be more aggressive on relaxed idempotent
18953 // rmw. In practice, they do not look useful, so we don't try to be
18954 // especially clever.
18955 if (SynchScope == SingleThread)
18956 // FIXME: we could just insert an X86ISD::MEMBARRIER here, except we are at
18957 // the IR level, so we must wrap it in an intrinsic.
18960 if (!hasMFENCE(*Subtarget))
18961 // FIXME: it might make sense to use a locked operation here but on a
18962 // different cache-line to prevent cache-line bouncing. In practice it
18963 // is probably a small win, and x86 processors without mfence are rare
18964 // enough that we do not bother.
18968 llvm::Intrinsic::getDeclaration(M, Intrinsic::x86_sse2_mfence);
18969 Builder.CreateCall(MFence, {});
18971 // Finally we can emit the atomic load.
18972 LoadInst *Loaded = Builder.CreateAlignedLoad(Ptr,
18973 AI->getType()->getPrimitiveSizeInBits());
18974 Loaded->setAtomic(Order, SynchScope);
18975 AI->replaceAllUsesWith(Loaded);
18976 AI->eraseFromParent();
18980 static SDValue LowerATOMIC_FENCE(SDValue Op, const X86Subtarget *Subtarget,
18981 SelectionDAG &DAG) {
18983 AtomicOrdering FenceOrdering = static_cast<AtomicOrdering>(
18984 cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue());
18985 SynchronizationScope FenceScope = static_cast<SynchronizationScope>(
18986 cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue());
18988 // The only fence that needs an instruction is a sequentially-consistent
18989 // cross-thread fence.
18990 if (FenceOrdering == SequentiallyConsistent && FenceScope == CrossThread) {
18991 if (hasMFENCE(*Subtarget))
18992 return DAG.getNode(X86ISD::MFENCE, dl, MVT::Other, Op.getOperand(0));
18994 SDValue Chain = Op.getOperand(0);
18995 SDValue Zero = DAG.getConstant(0, dl, MVT::i32);
18997 DAG.getRegister(X86::ESP, MVT::i32), // Base
18998 DAG.getTargetConstant(1, dl, MVT::i8), // Scale
18999 DAG.getRegister(0, MVT::i32), // Index
19000 DAG.getTargetConstant(0, dl, MVT::i32), // Disp
19001 DAG.getRegister(0, MVT::i32), // Segment.
19005 SDNode *Res = DAG.getMachineNode(X86::OR32mrLocked, dl, MVT::Other, Ops);
19006 return SDValue(Res, 0);
19009 // MEMBARRIER is a compiler barrier; it codegens to a no-op.
19010 return DAG.getNode(X86ISD::MEMBARRIER, dl, MVT::Other, Op.getOperand(0));
19013 static SDValue LowerCMP_SWAP(SDValue Op, const X86Subtarget *Subtarget,
19014 SelectionDAG &DAG) {
19015 MVT T = Op.getSimpleValueType();
19019 switch(T.SimpleTy) {
19020 default: llvm_unreachable("Invalid value type!");
19021 case MVT::i8: Reg = X86::AL; size = 1; break;
19022 case MVT::i16: Reg = X86::AX; size = 2; break;
19023 case MVT::i32: Reg = X86::EAX; size = 4; break;
19025 assert(Subtarget->is64Bit() && "Node not type legal!");
19026 Reg = X86::RAX; size = 8;
19029 SDValue cpIn = DAG.getCopyToReg(Op.getOperand(0), DL, Reg,
19030 Op.getOperand(2), SDValue());
19031 SDValue Ops[] = { cpIn.getValue(0),
19034 DAG.getTargetConstant(size, DL, MVT::i8),
19035 cpIn.getValue(1) };
19036 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Glue);
19037 MachineMemOperand *MMO = cast<AtomicSDNode>(Op)->getMemOperand();
19038 SDValue Result = DAG.getMemIntrinsicNode(X86ISD::LCMPXCHG_DAG, DL, Tys,
19042 DAG.getCopyFromReg(Result.getValue(0), DL, Reg, T, Result.getValue(1));
19043 SDValue EFLAGS = DAG.getCopyFromReg(cpOut.getValue(1), DL, X86::EFLAGS,
19044 MVT::i32, cpOut.getValue(2));
19045 SDValue Success = DAG.getNode(X86ISD::SETCC, DL, Op->getValueType(1),
19046 DAG.getConstant(X86::COND_E, DL, MVT::i8),
19049 DAG.ReplaceAllUsesOfValueWith(Op.getValue(0), cpOut);
19050 DAG.ReplaceAllUsesOfValueWith(Op.getValue(1), Success);
19051 DAG.ReplaceAllUsesOfValueWith(Op.getValue(2), EFLAGS.getValue(1));
19055 static SDValue LowerBITCAST(SDValue Op, const X86Subtarget *Subtarget,
19056 SelectionDAG &DAG) {
19057 MVT SrcVT = Op.getOperand(0).getSimpleValueType();
19058 MVT DstVT = Op.getSimpleValueType();
19060 if (SrcVT == MVT::v2i32 || SrcVT == MVT::v4i16 || SrcVT == MVT::v8i8) {
19061 assert(Subtarget->hasSSE2() && "Requires at least SSE2!");
19062 if (DstVT != MVT::f64)
19063 // This conversion needs to be expanded.
19066 SDValue InVec = Op->getOperand(0);
19068 unsigned NumElts = SrcVT.getVectorNumElements();
19069 EVT SVT = SrcVT.getVectorElementType();
19071 // Widen the vector in input in the case of MVT::v2i32.
19072 // Example: from MVT::v2i32 to MVT::v4i32.
19073 SmallVector<SDValue, 16> Elts;
19074 for (unsigned i = 0, e = NumElts; i != e; ++i)
19075 Elts.push_back(DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, SVT, InVec,
19076 DAG.getIntPtrConstant(i, dl)));
19078 // Explicitly mark the extra elements as Undef.
19079 Elts.append(NumElts, DAG.getUNDEF(SVT));
19081 EVT NewVT = EVT::getVectorVT(*DAG.getContext(), SVT, NumElts * 2);
19082 SDValue BV = DAG.getNode(ISD::BUILD_VECTOR, dl, NewVT, Elts);
19083 SDValue ToV2F64 = DAG.getBitcast(MVT::v2f64, BV);
19084 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64, ToV2F64,
19085 DAG.getIntPtrConstant(0, dl));
19088 assert(Subtarget->is64Bit() && !Subtarget->hasSSE2() &&
19089 Subtarget->hasMMX() && "Unexpected custom BITCAST");
19090 assert((DstVT == MVT::i64 ||
19091 (DstVT.isVector() && DstVT.getSizeInBits()==64)) &&
19092 "Unexpected custom BITCAST");
19093 // i64 <=> MMX conversions are Legal.
19094 if (SrcVT==MVT::i64 && DstVT.isVector())
19096 if (DstVT==MVT::i64 && SrcVT.isVector())
19098 // MMX <=> MMX conversions are Legal.
19099 if (SrcVT.isVector() && DstVT.isVector())
19101 // All other conversions need to be expanded.
19105 /// Compute the horizontal sum of bytes in V for the elements of VT.
19107 /// Requires V to be a byte vector and VT to be an integer vector type with
19108 /// wider elements than V's type. The width of the elements of VT determines
19109 /// how many bytes of V are summed horizontally to produce each element of the
19111 static SDValue LowerHorizontalByteSum(SDValue V, MVT VT,
19112 const X86Subtarget *Subtarget,
19113 SelectionDAG &DAG) {
19115 MVT ByteVecVT = V.getSimpleValueType();
19116 MVT EltVT = VT.getVectorElementType();
19117 int NumElts = VT.getVectorNumElements();
19118 assert(ByteVecVT.getVectorElementType() == MVT::i8 &&
19119 "Expected value to have byte element type.");
19120 assert(EltVT != MVT::i8 &&
19121 "Horizontal byte sum only makes sense for wider elements!");
19122 unsigned VecSize = VT.getSizeInBits();
19123 assert(ByteVecVT.getSizeInBits() == VecSize && "Cannot change vector size!");
19125 // PSADBW instruction horizontally add all bytes and leave the result in i64
19126 // chunks, thus directly computes the pop count for v2i64 and v4i64.
19127 if (EltVT == MVT::i64) {
19128 SDValue Zeros = getZeroVector(ByteVecVT, Subtarget, DAG, DL);
19129 V = DAG.getNode(X86ISD::PSADBW, DL, ByteVecVT, V, Zeros);
19130 return DAG.getBitcast(VT, V);
19133 if (EltVT == MVT::i32) {
19134 // We unpack the low half and high half into i32s interleaved with zeros so
19135 // that we can use PSADBW to horizontally sum them. The most useful part of
19136 // this is that it lines up the results of two PSADBW instructions to be
19137 // two v2i64 vectors which concatenated are the 4 population counts. We can
19138 // then use PACKUSWB to shrink and concatenate them into a v4i32 again.
19139 SDValue Zeros = getZeroVector(VT, Subtarget, DAG, DL);
19140 SDValue Low = DAG.getNode(X86ISD::UNPCKL, DL, VT, V, Zeros);
19141 SDValue High = DAG.getNode(X86ISD::UNPCKH, DL, VT, V, Zeros);
19143 // Do the horizontal sums into two v2i64s.
19144 Zeros = getZeroVector(ByteVecVT, Subtarget, DAG, DL);
19145 Low = DAG.getNode(X86ISD::PSADBW, DL, ByteVecVT,
19146 DAG.getBitcast(ByteVecVT, Low), Zeros);
19147 High = DAG.getNode(X86ISD::PSADBW, DL, ByteVecVT,
19148 DAG.getBitcast(ByteVecVT, High), Zeros);
19150 // Merge them together.
19151 MVT ShortVecVT = MVT::getVectorVT(MVT::i16, VecSize / 16);
19152 V = DAG.getNode(X86ISD::PACKUS, DL, ByteVecVT,
19153 DAG.getBitcast(ShortVecVT, Low),
19154 DAG.getBitcast(ShortVecVT, High));
19156 return DAG.getBitcast(VT, V);
19159 // The only element type left is i16.
19160 assert(EltVT == MVT::i16 && "Unknown how to handle type");
19162 // To obtain pop count for each i16 element starting from the pop count for
19163 // i8 elements, shift the i16s left by 8, sum as i8s, and then shift as i16s
19164 // right by 8. It is important to shift as i16s as i8 vector shift isn't
19165 // directly supported.
19166 SmallVector<SDValue, 16> Shifters(NumElts, DAG.getConstant(8, DL, EltVT));
19167 SDValue Shifter = DAG.getNode(ISD::BUILD_VECTOR, DL, VT, Shifters);
19168 SDValue Shl = DAG.getNode(ISD::SHL, DL, VT, DAG.getBitcast(VT, V), Shifter);
19169 V = DAG.getNode(ISD::ADD, DL, ByteVecVT, DAG.getBitcast(ByteVecVT, Shl),
19170 DAG.getBitcast(ByteVecVT, V));
19171 return DAG.getNode(ISD::SRL, DL, VT, DAG.getBitcast(VT, V), Shifter);
19174 static SDValue LowerVectorCTPOPInRegLUT(SDValue Op, SDLoc DL,
19175 const X86Subtarget *Subtarget,
19176 SelectionDAG &DAG) {
19177 MVT VT = Op.getSimpleValueType();
19178 MVT EltVT = VT.getVectorElementType();
19179 unsigned VecSize = VT.getSizeInBits();
19181 // Implement a lookup table in register by using an algorithm based on:
19182 // http://wm.ite.pl/articles/sse-popcount.html
19184 // The general idea is that every lower byte nibble in the input vector is an
19185 // index into a in-register pre-computed pop count table. We then split up the
19186 // input vector in two new ones: (1) a vector with only the shifted-right
19187 // higher nibbles for each byte and (2) a vector with the lower nibbles (and
19188 // masked out higher ones) for each byte. PSHUB is used separately with both
19189 // to index the in-register table. Next, both are added and the result is a
19190 // i8 vector where each element contains the pop count for input byte.
19192 // To obtain the pop count for elements != i8, we follow up with the same
19193 // approach and use additional tricks as described below.
19195 const int LUT[16] = {/* 0 */ 0, /* 1 */ 1, /* 2 */ 1, /* 3 */ 2,
19196 /* 4 */ 1, /* 5 */ 2, /* 6 */ 2, /* 7 */ 3,
19197 /* 8 */ 1, /* 9 */ 2, /* a */ 2, /* b */ 3,
19198 /* c */ 2, /* d */ 3, /* e */ 3, /* f */ 4};
19200 int NumByteElts = VecSize / 8;
19201 MVT ByteVecVT = MVT::getVectorVT(MVT::i8, NumByteElts);
19202 SDValue In = DAG.getBitcast(ByteVecVT, Op);
19203 SmallVector<SDValue, 16> LUTVec;
19204 for (int i = 0; i < NumByteElts; ++i)
19205 LUTVec.push_back(DAG.getConstant(LUT[i % 16], DL, MVT::i8));
19206 SDValue InRegLUT = DAG.getNode(ISD::BUILD_VECTOR, DL, ByteVecVT, LUTVec);
19207 SmallVector<SDValue, 16> Mask0F(NumByteElts,
19208 DAG.getConstant(0x0F, DL, MVT::i8));
19209 SDValue M0F = DAG.getNode(ISD::BUILD_VECTOR, DL, ByteVecVT, Mask0F);
19212 SmallVector<SDValue, 16> Four(NumByteElts, DAG.getConstant(4, DL, MVT::i8));
19213 SDValue FourV = DAG.getNode(ISD::BUILD_VECTOR, DL, ByteVecVT, Four);
19214 SDValue HighNibbles = DAG.getNode(ISD::SRL, DL, ByteVecVT, In, FourV);
19217 SDValue LowNibbles = DAG.getNode(ISD::AND, DL, ByteVecVT, In, M0F);
19219 // The input vector is used as the shuffle mask that index elements into the
19220 // LUT. After counting low and high nibbles, add the vector to obtain the
19221 // final pop count per i8 element.
19222 SDValue HighPopCnt =
19223 DAG.getNode(X86ISD::PSHUFB, DL, ByteVecVT, InRegLUT, HighNibbles);
19224 SDValue LowPopCnt =
19225 DAG.getNode(X86ISD::PSHUFB, DL, ByteVecVT, InRegLUT, LowNibbles);
19226 SDValue PopCnt = DAG.getNode(ISD::ADD, DL, ByteVecVT, HighPopCnt, LowPopCnt);
19228 if (EltVT == MVT::i8)
19231 return LowerHorizontalByteSum(PopCnt, VT, Subtarget, DAG);
19234 static SDValue LowerVectorCTPOPBitmath(SDValue Op, SDLoc DL,
19235 const X86Subtarget *Subtarget,
19236 SelectionDAG &DAG) {
19237 MVT VT = Op.getSimpleValueType();
19238 assert(VT.is128BitVector() &&
19239 "Only 128-bit vector bitmath lowering supported.");
19241 int VecSize = VT.getSizeInBits();
19242 MVT EltVT = VT.getVectorElementType();
19243 int Len = EltVT.getSizeInBits();
19245 // This is the vectorized version of the "best" algorithm from
19246 // http://graphics.stanford.edu/~seander/bithacks.html#CountBitsSetParallel
19247 // with a minor tweak to use a series of adds + shifts instead of vector
19248 // multiplications. Implemented for all integer vector types. We only use
19249 // this when we don't have SSSE3 which allows a LUT-based lowering that is
19250 // much faster, even faster than using native popcnt instructions.
19252 auto GetShift = [&](unsigned OpCode, SDValue V, int Shifter) {
19253 MVT VT = V.getSimpleValueType();
19254 SmallVector<SDValue, 32> Shifters(
19255 VT.getVectorNumElements(),
19256 DAG.getConstant(Shifter, DL, VT.getVectorElementType()));
19257 return DAG.getNode(OpCode, DL, VT, V,
19258 DAG.getNode(ISD::BUILD_VECTOR, DL, VT, Shifters));
19260 auto GetMask = [&](SDValue V, APInt Mask) {
19261 MVT VT = V.getSimpleValueType();
19262 SmallVector<SDValue, 32> Masks(
19263 VT.getVectorNumElements(),
19264 DAG.getConstant(Mask, DL, VT.getVectorElementType()));
19265 return DAG.getNode(ISD::AND, DL, VT, V,
19266 DAG.getNode(ISD::BUILD_VECTOR, DL, VT, Masks));
19269 // We don't want to incur the implicit masks required to SRL vNi8 vectors on
19270 // x86, so set the SRL type to have elements at least i16 wide. This is
19271 // correct because all of our SRLs are followed immediately by a mask anyways
19272 // that handles any bits that sneak into the high bits of the byte elements.
19273 MVT SrlVT = Len > 8 ? VT : MVT::getVectorVT(MVT::i16, VecSize / 16);
19277 // v = v - ((v >> 1) & 0x55555555...)
19279 DAG.getBitcast(VT, GetShift(ISD::SRL, DAG.getBitcast(SrlVT, V), 1));
19280 SDValue And = GetMask(Srl, APInt::getSplat(Len, APInt(8, 0x55)));
19281 V = DAG.getNode(ISD::SUB, DL, VT, V, And);
19283 // v = (v & 0x33333333...) + ((v >> 2) & 0x33333333...)
19284 SDValue AndLHS = GetMask(V, APInt::getSplat(Len, APInt(8, 0x33)));
19285 Srl = DAG.getBitcast(VT, GetShift(ISD::SRL, DAG.getBitcast(SrlVT, V), 2));
19286 SDValue AndRHS = GetMask(Srl, APInt::getSplat(Len, APInt(8, 0x33)));
19287 V = DAG.getNode(ISD::ADD, DL, VT, AndLHS, AndRHS);
19289 // v = (v + (v >> 4)) & 0x0F0F0F0F...
19290 Srl = DAG.getBitcast(VT, GetShift(ISD::SRL, DAG.getBitcast(SrlVT, V), 4));
19291 SDValue Add = DAG.getNode(ISD::ADD, DL, VT, V, Srl);
19292 V = GetMask(Add, APInt::getSplat(Len, APInt(8, 0x0F)));
19294 // At this point, V contains the byte-wise population count, and we are
19295 // merely doing a horizontal sum if necessary to get the wider element
19297 if (EltVT == MVT::i8)
19300 return LowerHorizontalByteSum(
19301 DAG.getBitcast(MVT::getVectorVT(MVT::i8, VecSize / 8), V), VT, Subtarget,
19305 static SDValue LowerVectorCTPOP(SDValue Op, const X86Subtarget *Subtarget,
19306 SelectionDAG &DAG) {
19307 MVT VT = Op.getSimpleValueType();
19308 // FIXME: Need to add AVX-512 support here!
19309 assert((VT.is256BitVector() || VT.is128BitVector()) &&
19310 "Unknown CTPOP type to handle");
19311 SDLoc DL(Op.getNode());
19312 SDValue Op0 = Op.getOperand(0);
19314 if (!Subtarget->hasSSSE3()) {
19315 // We can't use the fast LUT approach, so fall back on vectorized bitmath.
19316 assert(VT.is128BitVector() && "Only 128-bit vectors supported in SSE!");
19317 return LowerVectorCTPOPBitmath(Op0, DL, Subtarget, DAG);
19320 if (VT.is256BitVector() && !Subtarget->hasInt256()) {
19321 unsigned NumElems = VT.getVectorNumElements();
19323 // Extract each 128-bit vector, compute pop count and concat the result.
19324 SDValue LHS = Extract128BitVector(Op0, 0, DAG, DL);
19325 SDValue RHS = Extract128BitVector(Op0, NumElems/2, DAG, DL);
19327 return DAG.getNode(ISD::CONCAT_VECTORS, DL, VT,
19328 LowerVectorCTPOPInRegLUT(LHS, DL, Subtarget, DAG),
19329 LowerVectorCTPOPInRegLUT(RHS, DL, Subtarget, DAG));
19332 return LowerVectorCTPOPInRegLUT(Op0, DL, Subtarget, DAG);
19335 static SDValue LowerCTPOP(SDValue Op, const X86Subtarget *Subtarget,
19336 SelectionDAG &DAG) {
19337 assert(Op.getValueType().isVector() &&
19338 "We only do custom lowering for vector population count.");
19339 return LowerVectorCTPOP(Op, Subtarget, DAG);
19342 static SDValue LowerLOAD_SUB(SDValue Op, SelectionDAG &DAG) {
19343 SDNode *Node = Op.getNode();
19345 EVT T = Node->getValueType(0);
19346 SDValue negOp = DAG.getNode(ISD::SUB, dl, T,
19347 DAG.getConstant(0, dl, T), Node->getOperand(2));
19348 return DAG.getAtomic(ISD::ATOMIC_LOAD_ADD, dl,
19349 cast<AtomicSDNode>(Node)->getMemoryVT(),
19350 Node->getOperand(0),
19351 Node->getOperand(1), negOp,
19352 cast<AtomicSDNode>(Node)->getMemOperand(),
19353 cast<AtomicSDNode>(Node)->getOrdering(),
19354 cast<AtomicSDNode>(Node)->getSynchScope());
19357 static SDValue LowerATOMIC_STORE(SDValue Op, SelectionDAG &DAG) {
19358 SDNode *Node = Op.getNode();
19360 EVT VT = cast<AtomicSDNode>(Node)->getMemoryVT();
19362 // Convert seq_cst store -> xchg
19363 // Convert wide store -> swap (-> cmpxchg8b/cmpxchg16b)
19364 // FIXME: On 32-bit, store -> fist or movq would be more efficient
19365 // (The only way to get a 16-byte store is cmpxchg16b)
19366 // FIXME: 16-byte ATOMIC_SWAP isn't actually hooked up at the moment.
19367 if (cast<AtomicSDNode>(Node)->getOrdering() == SequentiallyConsistent ||
19368 !DAG.getTargetLoweringInfo().isTypeLegal(VT)) {
19369 SDValue Swap = DAG.getAtomic(ISD::ATOMIC_SWAP, dl,
19370 cast<AtomicSDNode>(Node)->getMemoryVT(),
19371 Node->getOperand(0),
19372 Node->getOperand(1), Node->getOperand(2),
19373 cast<AtomicSDNode>(Node)->getMemOperand(),
19374 cast<AtomicSDNode>(Node)->getOrdering(),
19375 cast<AtomicSDNode>(Node)->getSynchScope());
19376 return Swap.getValue(1);
19378 // Other atomic stores have a simple pattern.
19382 static SDValue LowerADDC_ADDE_SUBC_SUBE(SDValue Op, SelectionDAG &DAG) {
19383 EVT VT = Op.getNode()->getSimpleValueType(0);
19385 // Let legalize expand this if it isn't a legal type yet.
19386 if (!DAG.getTargetLoweringInfo().isTypeLegal(VT))
19389 SDVTList VTs = DAG.getVTList(VT, MVT::i32);
19392 bool ExtraOp = false;
19393 switch (Op.getOpcode()) {
19394 default: llvm_unreachable("Invalid code");
19395 case ISD::ADDC: Opc = X86ISD::ADD; break;
19396 case ISD::ADDE: Opc = X86ISD::ADC; ExtraOp = true; break;
19397 case ISD::SUBC: Opc = X86ISD::SUB; break;
19398 case ISD::SUBE: Opc = X86ISD::SBB; ExtraOp = true; break;
19402 return DAG.getNode(Opc, SDLoc(Op), VTs, Op.getOperand(0),
19404 return DAG.getNode(Opc, SDLoc(Op), VTs, Op.getOperand(0),
19405 Op.getOperand(1), Op.getOperand(2));
19408 static SDValue LowerFSINCOS(SDValue Op, const X86Subtarget *Subtarget,
19409 SelectionDAG &DAG) {
19410 assert(Subtarget->isTargetDarwin() && Subtarget->is64Bit());
19412 // For MacOSX, we want to call an alternative entry point: __sincos_stret,
19413 // which returns the values as { float, float } (in XMM0) or
19414 // { double, double } (which is returned in XMM0, XMM1).
19416 SDValue Arg = Op.getOperand(0);
19417 EVT ArgVT = Arg.getValueType();
19418 Type *ArgTy = ArgVT.getTypeForEVT(*DAG.getContext());
19420 TargetLowering::ArgListTy Args;
19421 TargetLowering::ArgListEntry Entry;
19425 Entry.isSExt = false;
19426 Entry.isZExt = false;
19427 Args.push_back(Entry);
19429 bool isF64 = ArgVT == MVT::f64;
19430 // Only optimize x86_64 for now. i386 is a bit messy. For f32,
19431 // the small struct {f32, f32} is returned in (eax, edx). For f64,
19432 // the results are returned via SRet in memory.
19433 const char *LibcallName = isF64 ? "__sincos_stret" : "__sincosf_stret";
19434 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
19436 DAG.getExternalSymbol(LibcallName, TLI.getPointerTy(DAG.getDataLayout()));
19438 Type *RetTy = isF64
19439 ? (Type*)StructType::get(ArgTy, ArgTy, nullptr)
19440 : (Type*)VectorType::get(ArgTy, 4);
19442 TargetLowering::CallLoweringInfo CLI(DAG);
19443 CLI.setDebugLoc(dl).setChain(DAG.getEntryNode())
19444 .setCallee(CallingConv::C, RetTy, Callee, std::move(Args), 0);
19446 std::pair<SDValue, SDValue> CallResult = TLI.LowerCallTo(CLI);
19449 // Returned in xmm0 and xmm1.
19450 return CallResult.first;
19452 // Returned in bits 0:31 and 32:64 xmm0.
19453 SDValue SinVal = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, ArgVT,
19454 CallResult.first, DAG.getIntPtrConstant(0, dl));
19455 SDValue CosVal = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, ArgVT,
19456 CallResult.first, DAG.getIntPtrConstant(1, dl));
19457 SDVTList Tys = DAG.getVTList(ArgVT, ArgVT);
19458 return DAG.getNode(ISD::MERGE_VALUES, dl, Tys, SinVal, CosVal);
19461 static SDValue LowerMSCATTER(SDValue Op, const X86Subtarget *Subtarget,
19462 SelectionDAG &DAG) {
19463 assert(Subtarget->hasAVX512() &&
19464 "MGATHER/MSCATTER are supported on AVX-512 arch only");
19466 MaskedScatterSDNode *N = cast<MaskedScatterSDNode>(Op.getNode());
19467 EVT VT = N->getValue().getValueType();
19468 assert(VT.getScalarSizeInBits() >= 32 && "Unsupported scatter op");
19471 // X86 scatter kills mask register, so its type should be added to
19472 // the list of return values
19473 if (N->getNumValues() == 1) {
19474 SDValue Index = N->getIndex();
19475 if (!Subtarget->hasVLX() && !VT.is512BitVector() &&
19476 !Index.getValueType().is512BitVector())
19477 Index = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v8i64, Index);
19479 SDVTList VTs = DAG.getVTList(N->getMask().getValueType(), MVT::Other);
19480 SDValue Ops[] = { N->getOperand(0), N->getOperand(1), N->getOperand(2),
19481 N->getOperand(3), Index };
19483 SDValue NewScatter = DAG.getMaskedScatter(VTs, VT, dl, Ops, N->getMemOperand());
19484 DAG.ReplaceAllUsesWith(Op, SDValue(NewScatter.getNode(), 1));
19485 return SDValue(NewScatter.getNode(), 0);
19490 static SDValue LowerMGATHER(SDValue Op, const X86Subtarget *Subtarget,
19491 SelectionDAG &DAG) {
19492 assert(Subtarget->hasAVX512() &&
19493 "MGATHER/MSCATTER are supported on AVX-512 arch only");
19495 MaskedGatherSDNode *N = cast<MaskedGatherSDNode>(Op.getNode());
19496 EVT VT = Op.getValueType();
19497 assert(VT.getScalarSizeInBits() >= 32 && "Unsupported gather op");
19500 SDValue Index = N->getIndex();
19501 if (!Subtarget->hasVLX() && !VT.is512BitVector() &&
19502 !Index.getValueType().is512BitVector()) {
19503 Index = DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v8i64, Index);
19504 SDValue Ops[] = { N->getOperand(0), N->getOperand(1), N->getOperand(2),
19505 N->getOperand(3), Index };
19506 DAG.UpdateNodeOperands(N, Ops);
19511 SDValue X86TargetLowering::LowerGC_TRANSITION_START(SDValue Op,
19512 SelectionDAG &DAG) const {
19513 // TODO: Eventually, the lowering of these nodes should be informed by or
19514 // deferred to the GC strategy for the function in which they appear. For
19515 // now, however, they must be lowered to something. Since they are logically
19516 // no-ops in the case of a null GC strategy (or a GC strategy which does not
19517 // require special handling for these nodes), lower them as literal NOOPs for
19519 SmallVector<SDValue, 2> Ops;
19521 Ops.push_back(Op.getOperand(0));
19522 if (Op->getGluedNode())
19523 Ops.push_back(Op->getOperand(Op->getNumOperands() - 1));
19526 SDVTList VTs = DAG.getVTList(MVT::Other, MVT::Glue);
19527 SDValue NOOP(DAG.getMachineNode(X86::NOOP, SDLoc(Op), VTs, Ops), 0);
19532 SDValue X86TargetLowering::LowerGC_TRANSITION_END(SDValue Op,
19533 SelectionDAG &DAG) const {
19534 // TODO: Eventually, the lowering of these nodes should be informed by or
19535 // deferred to the GC strategy for the function in which they appear. For
19536 // now, however, they must be lowered to something. Since they are logically
19537 // no-ops in the case of a null GC strategy (or a GC strategy which does not
19538 // require special handling for these nodes), lower them as literal NOOPs for
19540 SmallVector<SDValue, 2> Ops;
19542 Ops.push_back(Op.getOperand(0));
19543 if (Op->getGluedNode())
19544 Ops.push_back(Op->getOperand(Op->getNumOperands() - 1));
19547 SDVTList VTs = DAG.getVTList(MVT::Other, MVT::Glue);
19548 SDValue NOOP(DAG.getMachineNode(X86::NOOP, SDLoc(Op), VTs, Ops), 0);
19553 /// LowerOperation - Provide custom lowering hooks for some operations.
19555 SDValue X86TargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
19556 switch (Op.getOpcode()) {
19557 default: llvm_unreachable("Should not custom lower this!");
19558 case ISD::ATOMIC_FENCE: return LowerATOMIC_FENCE(Op, Subtarget, DAG);
19559 case ISD::ATOMIC_CMP_SWAP_WITH_SUCCESS:
19560 return LowerCMP_SWAP(Op, Subtarget, DAG);
19561 case ISD::CTPOP: return LowerCTPOP(Op, Subtarget, DAG);
19562 case ISD::ATOMIC_LOAD_SUB: return LowerLOAD_SUB(Op,DAG);
19563 case ISD::ATOMIC_STORE: return LowerATOMIC_STORE(Op,DAG);
19564 case ISD::BUILD_VECTOR: return LowerBUILD_VECTOR(Op, DAG);
19565 case ISD::CONCAT_VECTORS: return LowerCONCAT_VECTORS(Op, Subtarget, DAG);
19566 case ISD::VECTOR_SHUFFLE: return lowerVectorShuffle(Op, Subtarget, DAG);
19567 case ISD::VSELECT: return LowerVSELECT(Op, DAG);
19568 case ISD::EXTRACT_VECTOR_ELT: return LowerEXTRACT_VECTOR_ELT(Op, DAG);
19569 case ISD::INSERT_VECTOR_ELT: return LowerINSERT_VECTOR_ELT(Op, DAG);
19570 case ISD::EXTRACT_SUBVECTOR: return LowerEXTRACT_SUBVECTOR(Op,Subtarget,DAG);
19571 case ISD::INSERT_SUBVECTOR: return LowerINSERT_SUBVECTOR(Op, Subtarget,DAG);
19572 case ISD::SCALAR_TO_VECTOR: return LowerSCALAR_TO_VECTOR(Op, DAG);
19573 case ISD::ConstantPool: return LowerConstantPool(Op, DAG);
19574 case ISD::GlobalAddress: return LowerGlobalAddress(Op, DAG);
19575 case ISD::GlobalTLSAddress: return LowerGlobalTLSAddress(Op, DAG);
19576 case ISD::ExternalSymbol: return LowerExternalSymbol(Op, DAG);
19577 case ISD::BlockAddress: return LowerBlockAddress(Op, DAG);
19578 case ISD::SHL_PARTS:
19579 case ISD::SRA_PARTS:
19580 case ISD::SRL_PARTS: return LowerShiftParts(Op, DAG);
19581 case ISD::SINT_TO_FP: return LowerSINT_TO_FP(Op, DAG);
19582 case ISD::UINT_TO_FP: return LowerUINT_TO_FP(Op, DAG);
19583 case ISD::TRUNCATE: return LowerTRUNCATE(Op, DAG);
19584 case ISD::ZERO_EXTEND: return LowerZERO_EXTEND(Op, Subtarget, DAG);
19585 case ISD::SIGN_EXTEND: return LowerSIGN_EXTEND(Op, Subtarget, DAG);
19586 case ISD::ANY_EXTEND: return LowerANY_EXTEND(Op, Subtarget, DAG);
19587 case ISD::SIGN_EXTEND_VECTOR_INREG:
19588 return LowerSIGN_EXTEND_VECTOR_INREG(Op, Subtarget, DAG);
19589 case ISD::FP_TO_SINT: return LowerFP_TO_SINT(Op, DAG);
19590 case ISD::FP_TO_UINT: return LowerFP_TO_UINT(Op, DAG);
19591 case ISD::FP_EXTEND: return LowerFP_EXTEND(Op, DAG);
19592 case ISD::LOAD: return LowerExtendedLoad(Op, Subtarget, DAG);
19594 case ISD::FNEG: return LowerFABSorFNEG(Op, DAG);
19595 case ISD::FCOPYSIGN: return LowerFCOPYSIGN(Op, DAG);
19596 case ISD::FGETSIGN: return LowerFGETSIGN(Op, DAG);
19597 case ISD::SETCC: return LowerSETCC(Op, DAG);
19598 case ISD::SELECT: return LowerSELECT(Op, DAG);
19599 case ISD::BRCOND: return LowerBRCOND(Op, DAG);
19600 case ISD::JumpTable: return LowerJumpTable(Op, DAG);
19601 case ISD::VASTART: return LowerVASTART(Op, DAG);
19602 case ISD::VAARG: return LowerVAARG(Op, DAG);
19603 case ISD::VACOPY: return LowerVACOPY(Op, Subtarget, DAG);
19604 case ISD::INTRINSIC_WO_CHAIN: return LowerINTRINSIC_WO_CHAIN(Op, Subtarget, DAG);
19605 case ISD::INTRINSIC_VOID:
19606 case ISD::INTRINSIC_W_CHAIN: return LowerINTRINSIC_W_CHAIN(Op, Subtarget, DAG);
19607 case ISD::RETURNADDR: return LowerRETURNADDR(Op, DAG);
19608 case ISD::FRAMEADDR: return LowerFRAMEADDR(Op, DAG);
19609 case ISD::FRAME_TO_ARGS_OFFSET:
19610 return LowerFRAME_TO_ARGS_OFFSET(Op, DAG);
19611 case ISD::DYNAMIC_STACKALLOC: return LowerDYNAMIC_STACKALLOC(Op, DAG);
19612 case ISD::EH_RETURN: return LowerEH_RETURN(Op, DAG);
19613 case ISD::EH_SJLJ_SETJMP: return lowerEH_SJLJ_SETJMP(Op, DAG);
19614 case ISD::EH_SJLJ_LONGJMP: return lowerEH_SJLJ_LONGJMP(Op, DAG);
19615 case ISD::INIT_TRAMPOLINE: return LowerINIT_TRAMPOLINE(Op, DAG);
19616 case ISD::ADJUST_TRAMPOLINE: return LowerADJUST_TRAMPOLINE(Op, DAG);
19617 case ISD::FLT_ROUNDS_: return LowerFLT_ROUNDS_(Op, DAG);
19618 case ISD::CTLZ: return LowerCTLZ(Op, Subtarget, DAG);
19619 case ISD::CTLZ_ZERO_UNDEF: return LowerCTLZ_ZERO_UNDEF(Op, Subtarget, DAG);
19621 case ISD::CTTZ_ZERO_UNDEF: return LowerCTTZ(Op, DAG);
19622 case ISD::MUL: return LowerMUL(Op, Subtarget, DAG);
19623 case ISD::UMUL_LOHI:
19624 case ISD::SMUL_LOHI: return LowerMUL_LOHI(Op, Subtarget, DAG);
19625 case ISD::ROTL: return LowerRotate(Op, Subtarget, DAG);
19628 case ISD::SHL: return LowerShift(Op, Subtarget, DAG);
19634 case ISD::UMULO: return LowerXALUO(Op, DAG);
19635 case ISD::READCYCLECOUNTER: return LowerREADCYCLECOUNTER(Op, Subtarget,DAG);
19636 case ISD::BITCAST: return LowerBITCAST(Op, Subtarget, DAG);
19640 case ISD::SUBE: return LowerADDC_ADDE_SUBC_SUBE(Op, DAG);
19641 case ISD::ADD: return LowerADD(Op, DAG);
19642 case ISD::SUB: return LowerSUB(Op, DAG);
19646 case ISD::UMIN: return LowerMINMAX(Op, DAG);
19647 case ISD::FSINCOS: return LowerFSINCOS(Op, Subtarget, DAG);
19648 case ISD::MGATHER: return LowerMGATHER(Op, Subtarget, DAG);
19649 case ISD::MSCATTER: return LowerMSCATTER(Op, Subtarget, DAG);
19650 case ISD::GC_TRANSITION_START:
19651 return LowerGC_TRANSITION_START(Op, DAG);
19652 case ISD::GC_TRANSITION_END: return LowerGC_TRANSITION_END(Op, DAG);
19656 /// ReplaceNodeResults - Replace a node with an illegal result type
19657 /// with a new node built out of custom code.
19658 void X86TargetLowering::ReplaceNodeResults(SDNode *N,
19659 SmallVectorImpl<SDValue>&Results,
19660 SelectionDAG &DAG) const {
19662 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
19663 switch (N->getOpcode()) {
19665 llvm_unreachable("Do not know how to custom type legalize this operation!");
19666 // We might have generated v2f32 FMIN/FMAX operations. Widen them to v4f32.
19667 case X86ISD::FMINC:
19669 case X86ISD::FMAXC:
19670 case X86ISD::FMAX: {
19671 EVT VT = N->getValueType(0);
19672 if (VT != MVT::v2f32)
19673 llvm_unreachable("Unexpected type (!= v2f32) on FMIN/FMAX.");
19674 SDValue UNDEF = DAG.getUNDEF(VT);
19675 SDValue LHS = DAG.getNode(ISD::CONCAT_VECTORS, dl, MVT::v4f32,
19676 N->getOperand(0), UNDEF);
19677 SDValue RHS = DAG.getNode(ISD::CONCAT_VECTORS, dl, MVT::v4f32,
19678 N->getOperand(1), UNDEF);
19679 Results.push_back(DAG.getNode(N->getOpcode(), dl, MVT::v4f32, LHS, RHS));
19682 case ISD::SIGN_EXTEND_INREG:
19687 // We don't want to expand or promote these.
19694 case ISD::UDIVREM: {
19695 SDValue V = LowerWin64_i128OP(SDValue(N,0), DAG);
19696 Results.push_back(V);
19699 case ISD::FP_TO_SINT:
19700 case ISD::FP_TO_UINT: {
19701 bool IsSigned = N->getOpcode() == ISD::FP_TO_SINT;
19703 std::pair<SDValue,SDValue> Vals =
19704 FP_TO_INTHelper(SDValue(N, 0), DAG, IsSigned, /*IsReplace=*/ true);
19705 SDValue FIST = Vals.first, StackSlot = Vals.second;
19706 if (FIST.getNode()) {
19707 EVT VT = N->getValueType(0);
19708 // Return a load from the stack slot.
19709 if (StackSlot.getNode())
19710 Results.push_back(DAG.getLoad(VT, dl, FIST, StackSlot,
19711 MachinePointerInfo(),
19712 false, false, false, 0));
19714 Results.push_back(FIST);
19718 case ISD::UINT_TO_FP: {
19719 assert(Subtarget->hasSSE2() && "Requires at least SSE2!");
19720 if (N->getOperand(0).getValueType() != MVT::v2i32 ||
19721 N->getValueType(0) != MVT::v2f32)
19723 SDValue ZExtIn = DAG.getNode(ISD::ZERO_EXTEND, dl, MVT::v2i64,
19725 SDValue Bias = DAG.getConstantFP(BitsToDouble(0x4330000000000000ULL), dl,
19727 SDValue VBias = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v2f64, Bias, Bias);
19728 SDValue Or = DAG.getNode(ISD::OR, dl, MVT::v2i64, ZExtIn,
19729 DAG.getBitcast(MVT::v2i64, VBias));
19730 Or = DAG.getBitcast(MVT::v2f64, Or);
19731 // TODO: Are there any fast-math-flags to propagate here?
19732 SDValue Sub = DAG.getNode(ISD::FSUB, dl, MVT::v2f64, Or, VBias);
19733 Results.push_back(DAG.getNode(X86ISD::VFPROUND, dl, MVT::v4f32, Sub));
19736 case ISD::FP_ROUND: {
19737 if (!TLI.isTypeLegal(N->getOperand(0).getValueType()))
19739 SDValue V = DAG.getNode(X86ISD::VFPROUND, dl, MVT::v4f32, N->getOperand(0));
19740 Results.push_back(V);
19743 case ISD::FP_EXTEND: {
19744 // Right now, only MVT::v2f32 has OperationAction for FP_EXTEND.
19745 // No other ValueType for FP_EXTEND should reach this point.
19746 assert(N->getValueType(0) == MVT::v2f32 &&
19747 "Do not know how to legalize this Node");
19750 case ISD::INTRINSIC_W_CHAIN: {
19751 unsigned IntNo = cast<ConstantSDNode>(N->getOperand(1))->getZExtValue();
19753 default : llvm_unreachable("Do not know how to custom type "
19754 "legalize this intrinsic operation!");
19755 case Intrinsic::x86_rdtsc:
19756 return getReadTimeStampCounter(N, dl, X86ISD::RDTSC_DAG, DAG, Subtarget,
19758 case Intrinsic::x86_rdtscp:
19759 return getReadTimeStampCounter(N, dl, X86ISD::RDTSCP_DAG, DAG, Subtarget,
19761 case Intrinsic::x86_rdpmc:
19762 return getReadPerformanceCounter(N, dl, DAG, Subtarget, Results);
19765 case ISD::READCYCLECOUNTER: {
19766 return getReadTimeStampCounter(N, dl, X86ISD::RDTSC_DAG, DAG, Subtarget,
19769 case ISD::ATOMIC_CMP_SWAP_WITH_SUCCESS: {
19770 EVT T = N->getValueType(0);
19771 assert((T == MVT::i64 || T == MVT::i128) && "can only expand cmpxchg pair");
19772 bool Regs64bit = T == MVT::i128;
19773 EVT HalfT = Regs64bit ? MVT::i64 : MVT::i32;
19774 SDValue cpInL, cpInH;
19775 cpInL = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, HalfT, N->getOperand(2),
19776 DAG.getConstant(0, dl, HalfT));
19777 cpInH = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, HalfT, N->getOperand(2),
19778 DAG.getConstant(1, dl, HalfT));
19779 cpInL = DAG.getCopyToReg(N->getOperand(0), dl,
19780 Regs64bit ? X86::RAX : X86::EAX,
19782 cpInH = DAG.getCopyToReg(cpInL.getValue(0), dl,
19783 Regs64bit ? X86::RDX : X86::EDX,
19784 cpInH, cpInL.getValue(1));
19785 SDValue swapInL, swapInH;
19786 swapInL = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, HalfT, N->getOperand(3),
19787 DAG.getConstant(0, dl, HalfT));
19788 swapInH = DAG.getNode(ISD::EXTRACT_ELEMENT, dl, HalfT, N->getOperand(3),
19789 DAG.getConstant(1, dl, HalfT));
19790 swapInL = DAG.getCopyToReg(cpInH.getValue(0), dl,
19791 Regs64bit ? X86::RBX : X86::EBX,
19792 swapInL, cpInH.getValue(1));
19793 swapInH = DAG.getCopyToReg(swapInL.getValue(0), dl,
19794 Regs64bit ? X86::RCX : X86::ECX,
19795 swapInH, swapInL.getValue(1));
19796 SDValue Ops[] = { swapInH.getValue(0),
19798 swapInH.getValue(1) };
19799 SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Glue);
19800 MachineMemOperand *MMO = cast<AtomicSDNode>(N)->getMemOperand();
19801 unsigned Opcode = Regs64bit ? X86ISD::LCMPXCHG16_DAG :
19802 X86ISD::LCMPXCHG8_DAG;
19803 SDValue Result = DAG.getMemIntrinsicNode(Opcode, dl, Tys, Ops, T, MMO);
19804 SDValue cpOutL = DAG.getCopyFromReg(Result.getValue(0), dl,
19805 Regs64bit ? X86::RAX : X86::EAX,
19806 HalfT, Result.getValue(1));
19807 SDValue cpOutH = DAG.getCopyFromReg(cpOutL.getValue(1), dl,
19808 Regs64bit ? X86::RDX : X86::EDX,
19809 HalfT, cpOutL.getValue(2));
19810 SDValue OpsF[] = { cpOutL.getValue(0), cpOutH.getValue(0)};
19812 SDValue EFLAGS = DAG.getCopyFromReg(cpOutH.getValue(1), dl, X86::EFLAGS,
19813 MVT::i32, cpOutH.getValue(2));
19815 DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
19816 DAG.getConstant(X86::COND_E, dl, MVT::i8), EFLAGS);
19817 Success = DAG.getZExtOrTrunc(Success, dl, N->getValueType(1));
19819 Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, T, OpsF));
19820 Results.push_back(Success);
19821 Results.push_back(EFLAGS.getValue(1));
19824 case ISD::ATOMIC_SWAP:
19825 case ISD::ATOMIC_LOAD_ADD:
19826 case ISD::ATOMIC_LOAD_SUB:
19827 case ISD::ATOMIC_LOAD_AND:
19828 case ISD::ATOMIC_LOAD_OR:
19829 case ISD::ATOMIC_LOAD_XOR:
19830 case ISD::ATOMIC_LOAD_NAND:
19831 case ISD::ATOMIC_LOAD_MIN:
19832 case ISD::ATOMIC_LOAD_MAX:
19833 case ISD::ATOMIC_LOAD_UMIN:
19834 case ISD::ATOMIC_LOAD_UMAX:
19835 case ISD::ATOMIC_LOAD: {
19836 // Delegate to generic TypeLegalization. Situations we can really handle
19837 // should have already been dealt with by AtomicExpandPass.cpp.
19840 case ISD::BITCAST: {
19841 assert(Subtarget->hasSSE2() && "Requires at least SSE2!");
19842 EVT DstVT = N->getValueType(0);
19843 EVT SrcVT = N->getOperand(0)->getValueType(0);
19845 if (SrcVT != MVT::f64 ||
19846 (DstVT != MVT::v2i32 && DstVT != MVT::v4i16 && DstVT != MVT::v8i8))
19849 unsigned NumElts = DstVT.getVectorNumElements();
19850 EVT SVT = DstVT.getVectorElementType();
19851 EVT WiderVT = EVT::getVectorVT(*DAG.getContext(), SVT, NumElts * 2);
19852 SDValue Expanded = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
19853 MVT::v2f64, N->getOperand(0));
19854 SDValue ToVecInt = DAG.getBitcast(WiderVT, Expanded);
19856 if (ExperimentalVectorWideningLegalization) {
19857 // If we are legalizing vectors by widening, we already have the desired
19858 // legal vector type, just return it.
19859 Results.push_back(ToVecInt);
19863 SmallVector<SDValue, 8> Elts;
19864 for (unsigned i = 0, e = NumElts; i != e; ++i)
19865 Elts.push_back(DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, SVT,
19866 ToVecInt, DAG.getIntPtrConstant(i, dl)));
19868 Results.push_back(DAG.getNode(ISD::BUILD_VECTOR, dl, DstVT, Elts));
19873 const char *X86TargetLowering::getTargetNodeName(unsigned Opcode) const {
19874 switch ((X86ISD::NodeType)Opcode) {
19875 case X86ISD::FIRST_NUMBER: break;
19876 case X86ISD::BSF: return "X86ISD::BSF";
19877 case X86ISD::BSR: return "X86ISD::BSR";
19878 case X86ISD::SHLD: return "X86ISD::SHLD";
19879 case X86ISD::SHRD: return "X86ISD::SHRD";
19880 case X86ISD::FAND: return "X86ISD::FAND";
19881 case X86ISD::FANDN: return "X86ISD::FANDN";
19882 case X86ISD::FOR: return "X86ISD::FOR";
19883 case X86ISD::FXOR: return "X86ISD::FXOR";
19884 case X86ISD::FILD: return "X86ISD::FILD";
19885 case X86ISD::FILD_FLAG: return "X86ISD::FILD_FLAG";
19886 case X86ISD::FP_TO_INT16_IN_MEM: return "X86ISD::FP_TO_INT16_IN_MEM";
19887 case X86ISD::FP_TO_INT32_IN_MEM: return "X86ISD::FP_TO_INT32_IN_MEM";
19888 case X86ISD::FP_TO_INT64_IN_MEM: return "X86ISD::FP_TO_INT64_IN_MEM";
19889 case X86ISD::FLD: return "X86ISD::FLD";
19890 case X86ISD::FST: return "X86ISD::FST";
19891 case X86ISD::CALL: return "X86ISD::CALL";
19892 case X86ISD::RDTSC_DAG: return "X86ISD::RDTSC_DAG";
19893 case X86ISD::RDTSCP_DAG: return "X86ISD::RDTSCP_DAG";
19894 case X86ISD::RDPMC_DAG: return "X86ISD::RDPMC_DAG";
19895 case X86ISD::BT: return "X86ISD::BT";
19896 case X86ISD::CMP: return "X86ISD::CMP";
19897 case X86ISD::COMI: return "X86ISD::COMI";
19898 case X86ISD::UCOMI: return "X86ISD::UCOMI";
19899 case X86ISD::CMPM: return "X86ISD::CMPM";
19900 case X86ISD::CMPMU: return "X86ISD::CMPMU";
19901 case X86ISD::CMPM_RND: return "X86ISD::CMPM_RND";
19902 case X86ISD::SETCC: return "X86ISD::SETCC";
19903 case X86ISD::SETCC_CARRY: return "X86ISD::SETCC_CARRY";
19904 case X86ISD::FSETCC: return "X86ISD::FSETCC";
19905 case X86ISD::FGETSIGNx86: return "X86ISD::FGETSIGNx86";
19906 case X86ISD::CMOV: return "X86ISD::CMOV";
19907 case X86ISD::BRCOND: return "X86ISD::BRCOND";
19908 case X86ISD::RET_FLAG: return "X86ISD::RET_FLAG";
19909 case X86ISD::REP_STOS: return "X86ISD::REP_STOS";
19910 case X86ISD::REP_MOVS: return "X86ISD::REP_MOVS";
19911 case X86ISD::GlobalBaseReg: return "X86ISD::GlobalBaseReg";
19912 case X86ISD::Wrapper: return "X86ISD::Wrapper";
19913 case X86ISD::WrapperRIP: return "X86ISD::WrapperRIP";
19914 case X86ISD::MOVDQ2Q: return "X86ISD::MOVDQ2Q";
19915 case X86ISD::MMX_MOVD2W: return "X86ISD::MMX_MOVD2W";
19916 case X86ISD::MMX_MOVW2D: return "X86ISD::MMX_MOVW2D";
19917 case X86ISD::PEXTRB: return "X86ISD::PEXTRB";
19918 case X86ISD::PEXTRW: return "X86ISD::PEXTRW";
19919 case X86ISD::INSERTPS: return "X86ISD::INSERTPS";
19920 case X86ISD::PINSRB: return "X86ISD::PINSRB";
19921 case X86ISD::PINSRW: return "X86ISD::PINSRW";
19922 case X86ISD::MMX_PINSRW: return "X86ISD::MMX_PINSRW";
19923 case X86ISD::PSHUFB: return "X86ISD::PSHUFB";
19924 case X86ISD::ANDNP: return "X86ISD::ANDNP";
19925 case X86ISD::PSIGN: return "X86ISD::PSIGN";
19926 case X86ISD::BLENDI: return "X86ISD::BLENDI";
19927 case X86ISD::SHRUNKBLEND: return "X86ISD::SHRUNKBLEND";
19928 case X86ISD::ADDUS: return "X86ISD::ADDUS";
19929 case X86ISD::SUBUS: return "X86ISD::SUBUS";
19930 case X86ISD::HADD: return "X86ISD::HADD";
19931 case X86ISD::HSUB: return "X86ISD::HSUB";
19932 case X86ISD::FHADD: return "X86ISD::FHADD";
19933 case X86ISD::FHSUB: return "X86ISD::FHSUB";
19934 case X86ISD::ABS: return "X86ISD::ABS";
19935 case X86ISD::CONFLICT: return "X86ISD::CONFLICT";
19936 case X86ISD::FMAX: return "X86ISD::FMAX";
19937 case X86ISD::FMAX_RND: return "X86ISD::FMAX_RND";
19938 case X86ISD::FMIN: return "X86ISD::FMIN";
19939 case X86ISD::FMIN_RND: return "X86ISD::FMIN_RND";
19940 case X86ISD::FMAXC: return "X86ISD::FMAXC";
19941 case X86ISD::FMINC: return "X86ISD::FMINC";
19942 case X86ISD::FRSQRT: return "X86ISD::FRSQRT";
19943 case X86ISD::FRCP: return "X86ISD::FRCP";
19944 case X86ISD::EXTRQI: return "X86ISD::EXTRQI";
19945 case X86ISD::INSERTQI: return "X86ISD::INSERTQI";
19946 case X86ISD::TLSADDR: return "X86ISD::TLSADDR";
19947 case X86ISD::TLSBASEADDR: return "X86ISD::TLSBASEADDR";
19948 case X86ISD::TLSCALL: return "X86ISD::TLSCALL";
19949 case X86ISD::EH_SJLJ_SETJMP: return "X86ISD::EH_SJLJ_SETJMP";
19950 case X86ISD::EH_SJLJ_LONGJMP: return "X86ISD::EH_SJLJ_LONGJMP";
19951 case X86ISD::EH_RETURN: return "X86ISD::EH_RETURN";
19952 case X86ISD::TC_RETURN: return "X86ISD::TC_RETURN";
19953 case X86ISD::FNSTCW16m: return "X86ISD::FNSTCW16m";
19954 case X86ISD::FNSTSW16r: return "X86ISD::FNSTSW16r";
19955 case X86ISD::LCMPXCHG_DAG: return "X86ISD::LCMPXCHG_DAG";
19956 case X86ISD::LCMPXCHG8_DAG: return "X86ISD::LCMPXCHG8_DAG";
19957 case X86ISD::LCMPXCHG16_DAG: return "X86ISD::LCMPXCHG16_DAG";
19958 case X86ISD::VZEXT_MOVL: return "X86ISD::VZEXT_MOVL";
19959 case X86ISD::VZEXT_LOAD: return "X86ISD::VZEXT_LOAD";
19960 case X86ISD::VZEXT: return "X86ISD::VZEXT";
19961 case X86ISD::VSEXT: return "X86ISD::VSEXT";
19962 case X86ISD::VTRUNC: return "X86ISD::VTRUNC";
19963 case X86ISD::VTRUNCS: return "X86ISD::VTRUNCS";
19964 case X86ISD::VTRUNCUS: return "X86ISD::VTRUNCUS";
19965 case X86ISD::VINSERT: return "X86ISD::VINSERT";
19966 case X86ISD::VFPEXT: return "X86ISD::VFPEXT";
19967 case X86ISD::VFPROUND: return "X86ISD::VFPROUND";
19968 case X86ISD::CVTDQ2PD: return "X86ISD::CVTDQ2PD";
19969 case X86ISD::CVTUDQ2PD: return "X86ISD::CVTUDQ2PD";
19970 case X86ISD::VSHLDQ: return "X86ISD::VSHLDQ";
19971 case X86ISD::VSRLDQ: return "X86ISD::VSRLDQ";
19972 case X86ISD::VSHL: return "X86ISD::VSHL";
19973 case X86ISD::VSRL: return "X86ISD::VSRL";
19974 case X86ISD::VSRA: return "X86ISD::VSRA";
19975 case X86ISD::VSHLI: return "X86ISD::VSHLI";
19976 case X86ISD::VSRLI: return "X86ISD::VSRLI";
19977 case X86ISD::VSRAI: return "X86ISD::VSRAI";
19978 case X86ISD::CMPP: return "X86ISD::CMPP";
19979 case X86ISD::PCMPEQ: return "X86ISD::PCMPEQ";
19980 case X86ISD::PCMPGT: return "X86ISD::PCMPGT";
19981 case X86ISD::PCMPEQM: return "X86ISD::PCMPEQM";
19982 case X86ISD::PCMPGTM: return "X86ISD::PCMPGTM";
19983 case X86ISD::ADD: return "X86ISD::ADD";
19984 case X86ISD::SUB: return "X86ISD::SUB";
19985 case X86ISD::ADC: return "X86ISD::ADC";
19986 case X86ISD::SBB: return "X86ISD::SBB";
19987 case X86ISD::SMUL: return "X86ISD::SMUL";
19988 case X86ISD::UMUL: return "X86ISD::UMUL";
19989 case X86ISD::SMUL8: return "X86ISD::SMUL8";
19990 case X86ISD::UMUL8: return "X86ISD::UMUL8";
19991 case X86ISD::SDIVREM8_SEXT_HREG: return "X86ISD::SDIVREM8_SEXT_HREG";
19992 case X86ISD::UDIVREM8_ZEXT_HREG: return "X86ISD::UDIVREM8_ZEXT_HREG";
19993 case X86ISD::INC: return "X86ISD::INC";
19994 case X86ISD::DEC: return "X86ISD::DEC";
19995 case X86ISD::OR: return "X86ISD::OR";
19996 case X86ISD::XOR: return "X86ISD::XOR";
19997 case X86ISD::AND: return "X86ISD::AND";
19998 case X86ISD::BEXTR: return "X86ISD::BEXTR";
19999 case X86ISD::MUL_IMM: return "X86ISD::MUL_IMM";
20000 case X86ISD::PTEST: return "X86ISD::PTEST";
20001 case X86ISD::TESTP: return "X86ISD::TESTP";
20002 case X86ISD::TESTM: return "X86ISD::TESTM";
20003 case X86ISD::TESTNM: return "X86ISD::TESTNM";
20004 case X86ISD::KORTEST: return "X86ISD::KORTEST";
20005 case X86ISD::KTEST: return "X86ISD::KTEST";
20006 case X86ISD::PACKSS: return "X86ISD::PACKSS";
20007 case X86ISD::PACKUS: return "X86ISD::PACKUS";
20008 case X86ISD::PALIGNR: return "X86ISD::PALIGNR";
20009 case X86ISD::VALIGN: return "X86ISD::VALIGN";
20010 case X86ISD::PSHUFD: return "X86ISD::PSHUFD";
20011 case X86ISD::PSHUFHW: return "X86ISD::PSHUFHW";
20012 case X86ISD::PSHUFLW: return "X86ISD::PSHUFLW";
20013 case X86ISD::SHUFP: return "X86ISD::SHUFP";
20014 case X86ISD::SHUF128: return "X86ISD::SHUF128";
20015 case X86ISD::MOVLHPS: return "X86ISD::MOVLHPS";
20016 case X86ISD::MOVLHPD: return "X86ISD::MOVLHPD";
20017 case X86ISD::MOVHLPS: return "X86ISD::MOVHLPS";
20018 case X86ISD::MOVLPS: return "X86ISD::MOVLPS";
20019 case X86ISD::MOVLPD: return "X86ISD::MOVLPD";
20020 case X86ISD::MOVDDUP: return "X86ISD::MOVDDUP";
20021 case X86ISD::MOVSHDUP: return "X86ISD::MOVSHDUP";
20022 case X86ISD::MOVSLDUP: return "X86ISD::MOVSLDUP";
20023 case X86ISD::MOVSD: return "X86ISD::MOVSD";
20024 case X86ISD::MOVSS: return "X86ISD::MOVSS";
20025 case X86ISD::UNPCKL: return "X86ISD::UNPCKL";
20026 case X86ISD::UNPCKH: return "X86ISD::UNPCKH";
20027 case X86ISD::VBROADCAST: return "X86ISD::VBROADCAST";
20028 case X86ISD::SUBV_BROADCAST: return "X86ISD::SUBV_BROADCAST";
20029 case X86ISD::VEXTRACT: return "X86ISD::VEXTRACT";
20030 case X86ISD::VPERMILPV: return "X86ISD::VPERMILPV";
20031 case X86ISD::VPERMILPI: return "X86ISD::VPERMILPI";
20032 case X86ISD::VPERM2X128: return "X86ISD::VPERM2X128";
20033 case X86ISD::VPERMV: return "X86ISD::VPERMV";
20034 case X86ISD::VPERMV3: return "X86ISD::VPERMV3";
20035 case X86ISD::VPERMIV3: return "X86ISD::VPERMIV3";
20036 case X86ISD::VPERMI: return "X86ISD::VPERMI";
20037 case X86ISD::VPTERNLOG: return "X86ISD::VPTERNLOG";
20038 case X86ISD::VFIXUPIMM: return "X86ISD::VFIXUPIMM";
20039 case X86ISD::VRANGE: return "X86ISD::VRANGE";
20040 case X86ISD::PMULUDQ: return "X86ISD::PMULUDQ";
20041 case X86ISD::PMULDQ: return "X86ISD::PMULDQ";
20042 case X86ISD::PSADBW: return "X86ISD::PSADBW";
20043 case X86ISD::DBPSADBW: return "X86ISD::DBPSADBW";
20044 case X86ISD::VASTART_SAVE_XMM_REGS: return "X86ISD::VASTART_SAVE_XMM_REGS";
20045 case X86ISD::VAARG_64: return "X86ISD::VAARG_64";
20046 case X86ISD::WIN_ALLOCA: return "X86ISD::WIN_ALLOCA";
20047 case X86ISD::MEMBARRIER: return "X86ISD::MEMBARRIER";
20048 case X86ISD::MFENCE: return "X86ISD::MFENCE";
20049 case X86ISD::SFENCE: return "X86ISD::SFENCE";
20050 case X86ISD::LFENCE: return "X86ISD::LFENCE";
20051 case X86ISD::SEG_ALLOCA: return "X86ISD::SEG_ALLOCA";
20052 case X86ISD::SAHF: return "X86ISD::SAHF";
20053 case X86ISD::RDRAND: return "X86ISD::RDRAND";
20054 case X86ISD::RDSEED: return "X86ISD::RDSEED";
20055 case X86ISD::VPMADDUBSW: return "X86ISD::VPMADDUBSW";
20056 case X86ISD::VPMADDWD: return "X86ISD::VPMADDWD";
20057 case X86ISD::VPROT: return "X86ISD::VPROT";
20058 case X86ISD::VPROTI: return "X86ISD::VPROTI";
20059 case X86ISD::VPSHA: return "X86ISD::VPSHA";
20060 case X86ISD::VPSHL: return "X86ISD::VPSHL";
20061 case X86ISD::VPCOM: return "X86ISD::VPCOM";
20062 case X86ISD::VPCOMU: return "X86ISD::VPCOMU";
20063 case X86ISD::FMADD: return "X86ISD::FMADD";
20064 case X86ISD::FMSUB: return "X86ISD::FMSUB";
20065 case X86ISD::FNMADD: return "X86ISD::FNMADD";
20066 case X86ISD::FNMSUB: return "X86ISD::FNMSUB";
20067 case X86ISD::FMADDSUB: return "X86ISD::FMADDSUB";
20068 case X86ISD::FMSUBADD: return "X86ISD::FMSUBADD";
20069 case X86ISD::FMADD_RND: return "X86ISD::FMADD_RND";
20070 case X86ISD::FNMADD_RND: return "X86ISD::FNMADD_RND";
20071 case X86ISD::FMSUB_RND: return "X86ISD::FMSUB_RND";
20072 case X86ISD::FNMSUB_RND: return "X86ISD::FNMSUB_RND";
20073 case X86ISD::FMADDSUB_RND: return "X86ISD::FMADDSUB_RND";
20074 case X86ISD::FMSUBADD_RND: return "X86ISD::FMSUBADD_RND";
20075 case X86ISD::VRNDSCALE: return "X86ISD::VRNDSCALE";
20076 case X86ISD::VREDUCE: return "X86ISD::VREDUCE";
20077 case X86ISD::VGETMANT: return "X86ISD::VGETMANT";
20078 case X86ISD::PCMPESTRI: return "X86ISD::PCMPESTRI";
20079 case X86ISD::PCMPISTRI: return "X86ISD::PCMPISTRI";
20080 case X86ISD::XTEST: return "X86ISD::XTEST";
20081 case X86ISD::COMPRESS: return "X86ISD::COMPRESS";
20082 case X86ISD::EXPAND: return "X86ISD::EXPAND";
20083 case X86ISD::SELECT: return "X86ISD::SELECT";
20084 case X86ISD::ADDSUB: return "X86ISD::ADDSUB";
20085 case X86ISD::RCP28: return "X86ISD::RCP28";
20086 case X86ISD::EXP2: return "X86ISD::EXP2";
20087 case X86ISD::RSQRT28: return "X86ISD::RSQRT28";
20088 case X86ISD::FADD_RND: return "X86ISD::FADD_RND";
20089 case X86ISD::FSUB_RND: return "X86ISD::FSUB_RND";
20090 case X86ISD::FMUL_RND: return "X86ISD::FMUL_RND";
20091 case X86ISD::FDIV_RND: return "X86ISD::FDIV_RND";
20092 case X86ISD::FSQRT_RND: return "X86ISD::FSQRT_RND";
20093 case X86ISD::FGETEXP_RND: return "X86ISD::FGETEXP_RND";
20094 case X86ISD::SCALEF: return "X86ISD::SCALEF";
20095 case X86ISD::ADDS: return "X86ISD::ADDS";
20096 case X86ISD::SUBS: return "X86ISD::SUBS";
20097 case X86ISD::AVG: return "X86ISD::AVG";
20098 case X86ISD::MULHRS: return "X86ISD::MULHRS";
20099 case X86ISD::SINT_TO_FP_RND: return "X86ISD::SINT_TO_FP_RND";
20100 case X86ISD::UINT_TO_FP_RND: return "X86ISD::UINT_TO_FP_RND";
20101 case X86ISD::FP_TO_SINT_RND: return "X86ISD::FP_TO_SINT_RND";
20102 case X86ISD::FP_TO_UINT_RND: return "X86ISD::FP_TO_UINT_RND";
20103 case X86ISD::VFPCLASS: return "X86ISD::VFPCLASS";
20108 // isLegalAddressingMode - Return true if the addressing mode represented
20109 // by AM is legal for this target, for a load/store of the specified type.
20110 bool X86TargetLowering::isLegalAddressingMode(const DataLayout &DL,
20111 const AddrMode &AM, Type *Ty,
20112 unsigned AS) const {
20113 // X86 supports extremely general addressing modes.
20114 CodeModel::Model M = getTargetMachine().getCodeModel();
20115 Reloc::Model R = getTargetMachine().getRelocationModel();
20117 // X86 allows a sign-extended 32-bit immediate field as a displacement.
20118 if (!X86::isOffsetSuitableForCodeModel(AM.BaseOffs, M, AM.BaseGV != nullptr))
20123 Subtarget->ClassifyGlobalReference(AM.BaseGV, getTargetMachine());
20125 // If a reference to this global requires an extra load, we can't fold it.
20126 if (isGlobalStubReference(GVFlags))
20129 // If BaseGV requires a register for the PIC base, we cannot also have a
20130 // BaseReg specified.
20131 if (AM.HasBaseReg && isGlobalRelativeToPICBase(GVFlags))
20134 // If lower 4G is not available, then we must use rip-relative addressing.
20135 if ((M != CodeModel::Small || R != Reloc::Static) &&
20136 Subtarget->is64Bit() && (AM.BaseOffs || AM.Scale > 1))
20140 switch (AM.Scale) {
20146 // These scales always work.
20151 // These scales are formed with basereg+scalereg. Only accept if there is
20156 default: // Other stuff never works.
20163 bool X86TargetLowering::isVectorShiftByScalarCheap(Type *Ty) const {
20164 unsigned Bits = Ty->getScalarSizeInBits();
20166 // 8-bit shifts are always expensive, but versions with a scalar amount aren't
20167 // particularly cheaper than those without.
20171 // On AVX2 there are new vpsllv[dq] instructions (and other shifts), that make
20172 // variable shifts just as cheap as scalar ones.
20173 if (Subtarget->hasInt256() && (Bits == 32 || Bits == 64))
20176 // Otherwise, it's significantly cheaper to shift by a scalar amount than by a
20177 // fully general vector.
20181 bool X86TargetLowering::isTruncateFree(Type *Ty1, Type *Ty2) const {
20182 if (!Ty1->isIntegerTy() || !Ty2->isIntegerTy())
20184 unsigned NumBits1 = Ty1->getPrimitiveSizeInBits();
20185 unsigned NumBits2 = Ty2->getPrimitiveSizeInBits();
20186 return NumBits1 > NumBits2;
20189 bool X86TargetLowering::allowTruncateForTailCall(Type *Ty1, Type *Ty2) const {
20190 if (!Ty1->isIntegerTy() || !Ty2->isIntegerTy())
20193 if (!isTypeLegal(EVT::getEVT(Ty1)))
20196 assert(Ty1->getPrimitiveSizeInBits() <= 64 && "i128 is probably not a noop");
20198 // Assuming the caller doesn't have a zeroext or signext return parameter,
20199 // truncation all the way down to i1 is valid.
20203 bool X86TargetLowering::isLegalICmpImmediate(int64_t Imm) const {
20204 return isInt<32>(Imm);
20207 bool X86TargetLowering::isLegalAddImmediate(int64_t Imm) const {
20208 // Can also use sub to handle negated immediates.
20209 return isInt<32>(Imm);
20212 bool X86TargetLowering::isTruncateFree(EVT VT1, EVT VT2) const {
20213 if (!VT1.isInteger() || !VT2.isInteger())
20215 unsigned NumBits1 = VT1.getSizeInBits();
20216 unsigned NumBits2 = VT2.getSizeInBits();
20217 return NumBits1 > NumBits2;
20220 bool X86TargetLowering::isZExtFree(Type *Ty1, Type *Ty2) const {
20221 // x86-64 implicitly zero-extends 32-bit results in 64-bit registers.
20222 return Ty1->isIntegerTy(32) && Ty2->isIntegerTy(64) && Subtarget->is64Bit();
20225 bool X86TargetLowering::isZExtFree(EVT VT1, EVT VT2) const {
20226 // x86-64 implicitly zero-extends 32-bit results in 64-bit registers.
20227 return VT1 == MVT::i32 && VT2 == MVT::i64 && Subtarget->is64Bit();
20230 bool X86TargetLowering::isZExtFree(SDValue Val, EVT VT2) const {
20231 EVT VT1 = Val.getValueType();
20232 if (isZExtFree(VT1, VT2))
20235 if (Val.getOpcode() != ISD::LOAD)
20238 if (!VT1.isSimple() || !VT1.isInteger() ||
20239 !VT2.isSimple() || !VT2.isInteger())
20242 switch (VT1.getSimpleVT().SimpleTy) {
20247 // X86 has 8, 16, and 32-bit zero-extending loads.
20254 bool X86TargetLowering::isVectorLoadExtDesirable(SDValue) const { return true; }
20257 X86TargetLowering::isFMAFasterThanFMulAndFAdd(EVT VT) const {
20258 if (!(Subtarget->hasFMA() || Subtarget->hasFMA4() || Subtarget->hasAVX512()))
20261 VT = VT.getScalarType();
20263 if (!VT.isSimple())
20266 switch (VT.getSimpleVT().SimpleTy) {
20277 bool X86TargetLowering::isNarrowingProfitable(EVT VT1, EVT VT2) const {
20278 // i16 instructions are longer (0x66 prefix) and potentially slower.
20279 return !(VT1 == MVT::i32 && VT2 == MVT::i16);
20282 /// isShuffleMaskLegal - Targets can use this to indicate that they only
20283 /// support *some* VECTOR_SHUFFLE operations, those with specific masks.
20284 /// By default, if a target supports the VECTOR_SHUFFLE node, all mask values
20285 /// are assumed to be legal.
20287 X86TargetLowering::isShuffleMaskLegal(const SmallVectorImpl<int> &M,
20289 if (!VT.isSimple())
20292 // Not for i1 vectors
20293 if (VT.getScalarType() == MVT::i1)
20296 // Very little shuffling can be done for 64-bit vectors right now.
20297 if (VT.getSizeInBits() == 64)
20300 // We only care that the types being shuffled are legal. The lowering can
20301 // handle any possible shuffle mask that results.
20302 return isTypeLegal(VT.getSimpleVT());
20306 X86TargetLowering::isVectorClearMaskLegal(const SmallVectorImpl<int> &Mask,
20308 // Just delegate to the generic legality, clear masks aren't special.
20309 return isShuffleMaskLegal(Mask, VT);
20312 //===----------------------------------------------------------------------===//
20313 // X86 Scheduler Hooks
20314 //===----------------------------------------------------------------------===//
20316 /// Utility function to emit xbegin specifying the start of an RTM region.
20317 static MachineBasicBlock *EmitXBegin(MachineInstr *MI, MachineBasicBlock *MBB,
20318 const TargetInstrInfo *TII) {
20319 DebugLoc DL = MI->getDebugLoc();
20321 const BasicBlock *BB = MBB->getBasicBlock();
20322 MachineFunction::iterator I = ++MBB->getIterator();
20324 // For the v = xbegin(), we generate
20335 MachineBasicBlock *thisMBB = MBB;
20336 MachineFunction *MF = MBB->getParent();
20337 MachineBasicBlock *mainMBB = MF->CreateMachineBasicBlock(BB);
20338 MachineBasicBlock *sinkMBB = MF->CreateMachineBasicBlock(BB);
20339 MF->insert(I, mainMBB);
20340 MF->insert(I, sinkMBB);
20342 // Transfer the remainder of BB and its successor edges to sinkMBB.
20343 sinkMBB->splice(sinkMBB->begin(), MBB,
20344 std::next(MachineBasicBlock::iterator(MI)), MBB->end());
20345 sinkMBB->transferSuccessorsAndUpdatePHIs(MBB);
20349 // # fallthrough to mainMBB
20350 // # abortion to sinkMBB
20351 BuildMI(thisMBB, DL, TII->get(X86::XBEGIN_4)).addMBB(sinkMBB);
20352 thisMBB->addSuccessor(mainMBB);
20353 thisMBB->addSuccessor(sinkMBB);
20357 BuildMI(mainMBB, DL, TII->get(X86::MOV32ri), X86::EAX).addImm(-1);
20358 mainMBB->addSuccessor(sinkMBB);
20361 // EAX is live into the sinkMBB
20362 sinkMBB->addLiveIn(X86::EAX);
20363 BuildMI(*sinkMBB, sinkMBB->begin(), DL,
20364 TII->get(TargetOpcode::COPY), MI->getOperand(0).getReg())
20367 MI->eraseFromParent();
20371 // FIXME: When we get size specific XMM0 registers, i.e. XMM0_V16I8
20372 // or XMM0_V32I8 in AVX all of this code can be replaced with that
20373 // in the .td file.
20374 static MachineBasicBlock *EmitPCMPSTRM(MachineInstr *MI, MachineBasicBlock *BB,
20375 const TargetInstrInfo *TII) {
20377 switch (MI->getOpcode()) {
20378 default: llvm_unreachable("illegal opcode!");
20379 case X86::PCMPISTRM128REG: Opc = X86::PCMPISTRM128rr; break;
20380 case X86::VPCMPISTRM128REG: Opc = X86::VPCMPISTRM128rr; break;
20381 case X86::PCMPISTRM128MEM: Opc = X86::PCMPISTRM128rm; break;
20382 case X86::VPCMPISTRM128MEM: Opc = X86::VPCMPISTRM128rm; break;
20383 case X86::PCMPESTRM128REG: Opc = X86::PCMPESTRM128rr; break;
20384 case X86::VPCMPESTRM128REG: Opc = X86::VPCMPESTRM128rr; break;
20385 case X86::PCMPESTRM128MEM: Opc = X86::PCMPESTRM128rm; break;
20386 case X86::VPCMPESTRM128MEM: Opc = X86::VPCMPESTRM128rm; break;
20389 DebugLoc dl = MI->getDebugLoc();
20390 MachineInstrBuilder MIB = BuildMI(*BB, MI, dl, TII->get(Opc));
20392 unsigned NumArgs = MI->getNumOperands();
20393 for (unsigned i = 1; i < NumArgs; ++i) {
20394 MachineOperand &Op = MI->getOperand(i);
20395 if (!(Op.isReg() && Op.isImplicit()))
20396 MIB.addOperand(Op);
20398 if (MI->hasOneMemOperand())
20399 MIB->setMemRefs(MI->memoperands_begin(), MI->memoperands_end());
20401 BuildMI(*BB, MI, dl,
20402 TII->get(TargetOpcode::COPY), MI->getOperand(0).getReg())
20403 .addReg(X86::XMM0);
20405 MI->eraseFromParent();
20409 // FIXME: Custom handling because TableGen doesn't support multiple implicit
20410 // defs in an instruction pattern
20411 static MachineBasicBlock *EmitPCMPSTRI(MachineInstr *MI, MachineBasicBlock *BB,
20412 const TargetInstrInfo *TII) {
20414 switch (MI->getOpcode()) {
20415 default: llvm_unreachable("illegal opcode!");
20416 case X86::PCMPISTRIREG: Opc = X86::PCMPISTRIrr; break;
20417 case X86::VPCMPISTRIREG: Opc = X86::VPCMPISTRIrr; break;
20418 case X86::PCMPISTRIMEM: Opc = X86::PCMPISTRIrm; break;
20419 case X86::VPCMPISTRIMEM: Opc = X86::VPCMPISTRIrm; break;
20420 case X86::PCMPESTRIREG: Opc = X86::PCMPESTRIrr; break;
20421 case X86::VPCMPESTRIREG: Opc = X86::VPCMPESTRIrr; break;
20422 case X86::PCMPESTRIMEM: Opc = X86::PCMPESTRIrm; break;
20423 case X86::VPCMPESTRIMEM: Opc = X86::VPCMPESTRIrm; break;
20426 DebugLoc dl = MI->getDebugLoc();
20427 MachineInstrBuilder MIB = BuildMI(*BB, MI, dl, TII->get(Opc));
20429 unsigned NumArgs = MI->getNumOperands(); // remove the results
20430 for (unsigned i = 1; i < NumArgs; ++i) {
20431 MachineOperand &Op = MI->getOperand(i);
20432 if (!(Op.isReg() && Op.isImplicit()))
20433 MIB.addOperand(Op);
20435 if (MI->hasOneMemOperand())
20436 MIB->setMemRefs(MI->memoperands_begin(), MI->memoperands_end());
20438 BuildMI(*BB, MI, dl,
20439 TII->get(TargetOpcode::COPY), MI->getOperand(0).getReg())
20442 MI->eraseFromParent();
20446 static MachineBasicBlock *EmitMonitor(MachineInstr *MI, MachineBasicBlock *BB,
20447 const X86Subtarget *Subtarget) {
20448 DebugLoc dl = MI->getDebugLoc();
20449 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
20450 // Address into RAX/EAX, other two args into ECX, EDX.
20451 unsigned MemOpc = Subtarget->is64Bit() ? X86::LEA64r : X86::LEA32r;
20452 unsigned MemReg = Subtarget->is64Bit() ? X86::RAX : X86::EAX;
20453 MachineInstrBuilder MIB = BuildMI(*BB, MI, dl, TII->get(MemOpc), MemReg);
20454 for (int i = 0; i < X86::AddrNumOperands; ++i)
20455 MIB.addOperand(MI->getOperand(i));
20457 unsigned ValOps = X86::AddrNumOperands;
20458 BuildMI(*BB, MI, dl, TII->get(TargetOpcode::COPY), X86::ECX)
20459 .addReg(MI->getOperand(ValOps).getReg());
20460 BuildMI(*BB, MI, dl, TII->get(TargetOpcode::COPY), X86::EDX)
20461 .addReg(MI->getOperand(ValOps+1).getReg());
20463 // The instruction doesn't actually take any operands though.
20464 BuildMI(*BB, MI, dl, TII->get(X86::MONITORrrr));
20466 MI->eraseFromParent(); // The pseudo is gone now.
20470 MachineBasicBlock *
20471 X86TargetLowering::EmitVAARG64WithCustomInserter(MachineInstr *MI,
20472 MachineBasicBlock *MBB) const {
20473 // Emit va_arg instruction on X86-64.
20475 // Operands to this pseudo-instruction:
20476 // 0 ) Output : destination address (reg)
20477 // 1-5) Input : va_list address (addr, i64mem)
20478 // 6 ) ArgSize : Size (in bytes) of vararg type
20479 // 7 ) ArgMode : 0=overflow only, 1=use gp_offset, 2=use fp_offset
20480 // 8 ) Align : Alignment of type
20481 // 9 ) EFLAGS (implicit-def)
20483 assert(MI->getNumOperands() == 10 && "VAARG_64 should have 10 operands!");
20484 static_assert(X86::AddrNumOperands == 5,
20485 "VAARG_64 assumes 5 address operands");
20487 unsigned DestReg = MI->getOperand(0).getReg();
20488 MachineOperand &Base = MI->getOperand(1);
20489 MachineOperand &Scale = MI->getOperand(2);
20490 MachineOperand &Index = MI->getOperand(3);
20491 MachineOperand &Disp = MI->getOperand(4);
20492 MachineOperand &Segment = MI->getOperand(5);
20493 unsigned ArgSize = MI->getOperand(6).getImm();
20494 unsigned ArgMode = MI->getOperand(7).getImm();
20495 unsigned Align = MI->getOperand(8).getImm();
20497 // Memory Reference
20498 assert(MI->hasOneMemOperand() && "Expected VAARG_64 to have one memoperand");
20499 MachineInstr::mmo_iterator MMOBegin = MI->memoperands_begin();
20500 MachineInstr::mmo_iterator MMOEnd = MI->memoperands_end();
20502 // Machine Information
20503 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
20504 MachineRegisterInfo &MRI = MBB->getParent()->getRegInfo();
20505 const TargetRegisterClass *AddrRegClass = getRegClassFor(MVT::i64);
20506 const TargetRegisterClass *OffsetRegClass = getRegClassFor(MVT::i32);
20507 DebugLoc DL = MI->getDebugLoc();
20509 // struct va_list {
20512 // i64 overflow_area (address)
20513 // i64 reg_save_area (address)
20515 // sizeof(va_list) = 24
20516 // alignment(va_list) = 8
20518 unsigned TotalNumIntRegs = 6;
20519 unsigned TotalNumXMMRegs = 8;
20520 bool UseGPOffset = (ArgMode == 1);
20521 bool UseFPOffset = (ArgMode == 2);
20522 unsigned MaxOffset = TotalNumIntRegs * 8 +
20523 (UseFPOffset ? TotalNumXMMRegs * 16 : 0);
20525 /* Align ArgSize to a multiple of 8 */
20526 unsigned ArgSizeA8 = (ArgSize + 7) & ~7;
20527 bool NeedsAlign = (Align > 8);
20529 MachineBasicBlock *thisMBB = MBB;
20530 MachineBasicBlock *overflowMBB;
20531 MachineBasicBlock *offsetMBB;
20532 MachineBasicBlock *endMBB;
20534 unsigned OffsetDestReg = 0; // Argument address computed by offsetMBB
20535 unsigned OverflowDestReg = 0; // Argument address computed by overflowMBB
20536 unsigned OffsetReg = 0;
20538 if (!UseGPOffset && !UseFPOffset) {
20539 // If we only pull from the overflow region, we don't create a branch.
20540 // We don't need to alter control flow.
20541 OffsetDestReg = 0; // unused
20542 OverflowDestReg = DestReg;
20544 offsetMBB = nullptr;
20545 overflowMBB = thisMBB;
20548 // First emit code to check if gp_offset (or fp_offset) is below the bound.
20549 // If so, pull the argument from reg_save_area. (branch to offsetMBB)
20550 // If not, pull from overflow_area. (branch to overflowMBB)
20555 // offsetMBB overflowMBB
20560 // Registers for the PHI in endMBB
20561 OffsetDestReg = MRI.createVirtualRegister(AddrRegClass);
20562 OverflowDestReg = MRI.createVirtualRegister(AddrRegClass);
20564 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
20565 MachineFunction *MF = MBB->getParent();
20566 overflowMBB = MF->CreateMachineBasicBlock(LLVM_BB);
20567 offsetMBB = MF->CreateMachineBasicBlock(LLVM_BB);
20568 endMBB = MF->CreateMachineBasicBlock(LLVM_BB);
20570 MachineFunction::iterator MBBIter = ++MBB->getIterator();
20572 // Insert the new basic blocks
20573 MF->insert(MBBIter, offsetMBB);
20574 MF->insert(MBBIter, overflowMBB);
20575 MF->insert(MBBIter, endMBB);
20577 // Transfer the remainder of MBB and its successor edges to endMBB.
20578 endMBB->splice(endMBB->begin(), thisMBB,
20579 std::next(MachineBasicBlock::iterator(MI)), thisMBB->end());
20580 endMBB->transferSuccessorsAndUpdatePHIs(thisMBB);
20582 // Make offsetMBB and overflowMBB successors of thisMBB
20583 thisMBB->addSuccessor(offsetMBB);
20584 thisMBB->addSuccessor(overflowMBB);
20586 // endMBB is a successor of both offsetMBB and overflowMBB
20587 offsetMBB->addSuccessor(endMBB);
20588 overflowMBB->addSuccessor(endMBB);
20590 // Load the offset value into a register
20591 OffsetReg = MRI.createVirtualRegister(OffsetRegClass);
20592 BuildMI(thisMBB, DL, TII->get(X86::MOV32rm), OffsetReg)
20596 .addDisp(Disp, UseFPOffset ? 4 : 0)
20597 .addOperand(Segment)
20598 .setMemRefs(MMOBegin, MMOEnd);
20600 // Check if there is enough room left to pull this argument.
20601 BuildMI(thisMBB, DL, TII->get(X86::CMP32ri))
20603 .addImm(MaxOffset + 8 - ArgSizeA8);
20605 // Branch to "overflowMBB" if offset >= max
20606 // Fall through to "offsetMBB" otherwise
20607 BuildMI(thisMBB, DL, TII->get(X86::GetCondBranchFromCond(X86::COND_AE)))
20608 .addMBB(overflowMBB);
20611 // In offsetMBB, emit code to use the reg_save_area.
20613 assert(OffsetReg != 0);
20615 // Read the reg_save_area address.
20616 unsigned RegSaveReg = MRI.createVirtualRegister(AddrRegClass);
20617 BuildMI(offsetMBB, DL, TII->get(X86::MOV64rm), RegSaveReg)
20622 .addOperand(Segment)
20623 .setMemRefs(MMOBegin, MMOEnd);
20625 // Zero-extend the offset
20626 unsigned OffsetReg64 = MRI.createVirtualRegister(AddrRegClass);
20627 BuildMI(offsetMBB, DL, TII->get(X86::SUBREG_TO_REG), OffsetReg64)
20630 .addImm(X86::sub_32bit);
20632 // Add the offset to the reg_save_area to get the final address.
20633 BuildMI(offsetMBB, DL, TII->get(X86::ADD64rr), OffsetDestReg)
20634 .addReg(OffsetReg64)
20635 .addReg(RegSaveReg);
20637 // Compute the offset for the next argument
20638 unsigned NextOffsetReg = MRI.createVirtualRegister(OffsetRegClass);
20639 BuildMI(offsetMBB, DL, TII->get(X86::ADD32ri), NextOffsetReg)
20641 .addImm(UseFPOffset ? 16 : 8);
20643 // Store it back into the va_list.
20644 BuildMI(offsetMBB, DL, TII->get(X86::MOV32mr))
20648 .addDisp(Disp, UseFPOffset ? 4 : 0)
20649 .addOperand(Segment)
20650 .addReg(NextOffsetReg)
20651 .setMemRefs(MMOBegin, MMOEnd);
20654 BuildMI(offsetMBB, DL, TII->get(X86::JMP_1))
20659 // Emit code to use overflow area
20662 // Load the overflow_area address into a register.
20663 unsigned OverflowAddrReg = MRI.createVirtualRegister(AddrRegClass);
20664 BuildMI(overflowMBB, DL, TII->get(X86::MOV64rm), OverflowAddrReg)
20669 .addOperand(Segment)
20670 .setMemRefs(MMOBegin, MMOEnd);
20672 // If we need to align it, do so. Otherwise, just copy the address
20673 // to OverflowDestReg.
20675 // Align the overflow address
20676 assert((Align & (Align-1)) == 0 && "Alignment must be a power of 2");
20677 unsigned TmpReg = MRI.createVirtualRegister(AddrRegClass);
20679 // aligned_addr = (addr + (align-1)) & ~(align-1)
20680 BuildMI(overflowMBB, DL, TII->get(X86::ADD64ri32), TmpReg)
20681 .addReg(OverflowAddrReg)
20684 BuildMI(overflowMBB, DL, TII->get(X86::AND64ri32), OverflowDestReg)
20686 .addImm(~(uint64_t)(Align-1));
20688 BuildMI(overflowMBB, DL, TII->get(TargetOpcode::COPY), OverflowDestReg)
20689 .addReg(OverflowAddrReg);
20692 // Compute the next overflow address after this argument.
20693 // (the overflow address should be kept 8-byte aligned)
20694 unsigned NextAddrReg = MRI.createVirtualRegister(AddrRegClass);
20695 BuildMI(overflowMBB, DL, TII->get(X86::ADD64ri32), NextAddrReg)
20696 .addReg(OverflowDestReg)
20697 .addImm(ArgSizeA8);
20699 // Store the new overflow address.
20700 BuildMI(overflowMBB, DL, TII->get(X86::MOV64mr))
20705 .addOperand(Segment)
20706 .addReg(NextAddrReg)
20707 .setMemRefs(MMOBegin, MMOEnd);
20709 // If we branched, emit the PHI to the front of endMBB.
20711 BuildMI(*endMBB, endMBB->begin(), DL,
20712 TII->get(X86::PHI), DestReg)
20713 .addReg(OffsetDestReg).addMBB(offsetMBB)
20714 .addReg(OverflowDestReg).addMBB(overflowMBB);
20717 // Erase the pseudo instruction
20718 MI->eraseFromParent();
20723 MachineBasicBlock *
20724 X86TargetLowering::EmitVAStartSaveXMMRegsWithCustomInserter(
20726 MachineBasicBlock *MBB) const {
20727 // Emit code to save XMM registers to the stack. The ABI says that the
20728 // number of registers to save is given in %al, so it's theoretically
20729 // possible to do an indirect jump trick to avoid saving all of them,
20730 // however this code takes a simpler approach and just executes all
20731 // of the stores if %al is non-zero. It's less code, and it's probably
20732 // easier on the hardware branch predictor, and stores aren't all that
20733 // expensive anyway.
20735 // Create the new basic blocks. One block contains all the XMM stores,
20736 // and one block is the final destination regardless of whether any
20737 // stores were performed.
20738 const BasicBlock *LLVM_BB = MBB->getBasicBlock();
20739 MachineFunction *F = MBB->getParent();
20740 MachineFunction::iterator MBBIter = ++MBB->getIterator();
20741 MachineBasicBlock *XMMSaveMBB = F->CreateMachineBasicBlock(LLVM_BB);
20742 MachineBasicBlock *EndMBB = F->CreateMachineBasicBlock(LLVM_BB);
20743 F->insert(MBBIter, XMMSaveMBB);
20744 F->insert(MBBIter, EndMBB);
20746 // Transfer the remainder of MBB and its successor edges to EndMBB.
20747 EndMBB->splice(EndMBB->begin(), MBB,
20748 std::next(MachineBasicBlock::iterator(MI)), MBB->end());
20749 EndMBB->transferSuccessorsAndUpdatePHIs(MBB);
20751 // The original block will now fall through to the XMM save block.
20752 MBB->addSuccessor(XMMSaveMBB);
20753 // The XMMSaveMBB will fall through to the end block.
20754 XMMSaveMBB->addSuccessor(EndMBB);
20756 // Now add the instructions.
20757 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
20758 DebugLoc DL = MI->getDebugLoc();
20760 unsigned CountReg = MI->getOperand(0).getReg();
20761 int64_t RegSaveFrameIndex = MI->getOperand(1).getImm();
20762 int64_t VarArgsFPOffset = MI->getOperand(2).getImm();
20764 if (!Subtarget->isCallingConvWin64(F->getFunction()->getCallingConv())) {
20765 // If %al is 0, branch around the XMM save block.
20766 BuildMI(MBB, DL, TII->get(X86::TEST8rr)).addReg(CountReg).addReg(CountReg);
20767 BuildMI(MBB, DL, TII->get(X86::JE_1)).addMBB(EndMBB);
20768 MBB->addSuccessor(EndMBB);
20771 // Make sure the last operand is EFLAGS, which gets clobbered by the branch
20772 // that was just emitted, but clearly shouldn't be "saved".
20773 assert((MI->getNumOperands() <= 3 ||
20774 !MI->getOperand(MI->getNumOperands() - 1).isReg() ||
20775 MI->getOperand(MI->getNumOperands() - 1).getReg() == X86::EFLAGS)
20776 && "Expected last argument to be EFLAGS");
20777 unsigned MOVOpc = Subtarget->hasFp256() ? X86::VMOVAPSmr : X86::MOVAPSmr;
20778 // In the XMM save block, save all the XMM argument registers.
20779 for (int i = 3, e = MI->getNumOperands() - 1; i != e; ++i) {
20780 int64_t Offset = (i - 3) * 16 + VarArgsFPOffset;
20781 MachineMemOperand *MMO = F->getMachineMemOperand(
20782 MachinePointerInfo::getFixedStack(*F, RegSaveFrameIndex, Offset),
20783 MachineMemOperand::MOStore,
20784 /*Size=*/16, /*Align=*/16);
20785 BuildMI(XMMSaveMBB, DL, TII->get(MOVOpc))
20786 .addFrameIndex(RegSaveFrameIndex)
20787 .addImm(/*Scale=*/1)
20788 .addReg(/*IndexReg=*/0)
20789 .addImm(/*Disp=*/Offset)
20790 .addReg(/*Segment=*/0)
20791 .addReg(MI->getOperand(i).getReg())
20792 .addMemOperand(MMO);
20795 MI->eraseFromParent(); // The pseudo instruction is gone now.
20800 // The EFLAGS operand of SelectItr might be missing a kill marker
20801 // because there were multiple uses of EFLAGS, and ISel didn't know
20802 // which to mark. Figure out whether SelectItr should have had a
20803 // kill marker, and set it if it should. Returns the correct kill
20805 static bool checkAndUpdateEFLAGSKill(MachineBasicBlock::iterator SelectItr,
20806 MachineBasicBlock* BB,
20807 const TargetRegisterInfo* TRI) {
20808 // Scan forward through BB for a use/def of EFLAGS.
20809 MachineBasicBlock::iterator miI(std::next(SelectItr));
20810 for (MachineBasicBlock::iterator miE = BB->end(); miI != miE; ++miI) {
20811 const MachineInstr& mi = *miI;
20812 if (mi.readsRegister(X86::EFLAGS))
20814 if (mi.definesRegister(X86::EFLAGS))
20815 break; // Should have kill-flag - update below.
20818 // If we hit the end of the block, check whether EFLAGS is live into a
20820 if (miI == BB->end()) {
20821 for (MachineBasicBlock::succ_iterator sItr = BB->succ_begin(),
20822 sEnd = BB->succ_end();
20823 sItr != sEnd; ++sItr) {
20824 MachineBasicBlock* succ = *sItr;
20825 if (succ->isLiveIn(X86::EFLAGS))
20830 // We found a def, or hit the end of the basic block and EFLAGS wasn't live
20831 // out. SelectMI should have a kill flag on EFLAGS.
20832 SelectItr->addRegisterKilled(X86::EFLAGS, TRI);
20836 // Return true if it is OK for this CMOV pseudo-opcode to be cascaded
20837 // together with other CMOV pseudo-opcodes into a single basic-block with
20838 // conditional jump around it.
20839 static bool isCMOVPseudo(MachineInstr *MI) {
20840 switch (MI->getOpcode()) {
20841 case X86::CMOV_FR32:
20842 case X86::CMOV_FR64:
20843 case X86::CMOV_GR8:
20844 case X86::CMOV_GR16:
20845 case X86::CMOV_GR32:
20846 case X86::CMOV_RFP32:
20847 case X86::CMOV_RFP64:
20848 case X86::CMOV_RFP80:
20849 case X86::CMOV_V2F64:
20850 case X86::CMOV_V2I64:
20851 case X86::CMOV_V4F32:
20852 case X86::CMOV_V4F64:
20853 case X86::CMOV_V4I64:
20854 case X86::CMOV_V16F32:
20855 case X86::CMOV_V8F32:
20856 case X86::CMOV_V8F64:
20857 case X86::CMOV_V8I64:
20858 case X86::CMOV_V8I1:
20859 case X86::CMOV_V16I1:
20860 case X86::CMOV_V32I1:
20861 case X86::CMOV_V64I1:
20869 MachineBasicBlock *
20870 X86TargetLowering::EmitLoweredSelect(MachineInstr *MI,
20871 MachineBasicBlock *BB) const {
20872 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
20873 DebugLoc DL = MI->getDebugLoc();
20875 // To "insert" a SELECT_CC instruction, we actually have to insert the
20876 // diamond control-flow pattern. The incoming instruction knows the
20877 // destination vreg to set, the condition code register to branch on, the
20878 // true/false values to select between, and a branch opcode to use.
20879 const BasicBlock *LLVM_BB = BB->getBasicBlock();
20880 MachineFunction::iterator It = ++BB->getIterator();
20885 // cmpTY ccX, r1, r2
20887 // fallthrough --> copy0MBB
20888 MachineBasicBlock *thisMBB = BB;
20889 MachineFunction *F = BB->getParent();
20891 // This code lowers all pseudo-CMOV instructions. Generally it lowers these
20892 // as described above, by inserting a BB, and then making a PHI at the join
20893 // point to select the true and false operands of the CMOV in the PHI.
20895 // The code also handles two different cases of multiple CMOV opcodes
20899 // In this case, there are multiple CMOVs in a row, all which are based on
20900 // the same condition setting (or the exact opposite condition setting).
20901 // In this case we can lower all the CMOVs using a single inserted BB, and
20902 // then make a number of PHIs at the join point to model the CMOVs. The only
20903 // trickiness here, is that in a case like:
20905 // t2 = CMOV cond1 t1, f1
20906 // t3 = CMOV cond1 t2, f2
20908 // when rewriting this into PHIs, we have to perform some renaming on the
20909 // temps since you cannot have a PHI operand refer to a PHI result earlier
20910 // in the same block. The "simple" but wrong lowering would be:
20912 // t2 = PHI t1(BB1), f1(BB2)
20913 // t3 = PHI t2(BB1), f2(BB2)
20915 // but clearly t2 is not defined in BB1, so that is incorrect. The proper
20916 // renaming is to note that on the path through BB1, t2 is really just a
20917 // copy of t1, and do that renaming, properly generating:
20919 // t2 = PHI t1(BB1), f1(BB2)
20920 // t3 = PHI t1(BB1), f2(BB2)
20922 // Case 2, we lower cascaded CMOVs such as
20924 // (CMOV (CMOV F, T, cc1), T, cc2)
20926 // to two successives branches. For that, we look for another CMOV as the
20927 // following instruction.
20929 // Without this, we would add a PHI between the two jumps, which ends up
20930 // creating a few copies all around. For instance, for
20932 // (sitofp (zext (fcmp une)))
20934 // we would generate:
20936 // ucomiss %xmm1, %xmm0
20937 // movss <1.0f>, %xmm0
20938 // movaps %xmm0, %xmm1
20940 // xorps %xmm1, %xmm1
20943 // movaps %xmm1, %xmm0
20947 // because this custom-inserter would have generated:
20959 // A: X = ...; Y = ...
20961 // C: Z = PHI [X, A], [Y, B]
20963 // E: PHI [X, C], [Z, D]
20965 // If we lower both CMOVs in a single step, we can instead generate:
20977 // A: X = ...; Y = ...
20979 // E: PHI [X, A], [X, C], [Y, D]
20981 // Which, in our sitofp/fcmp example, gives us something like:
20983 // ucomiss %xmm1, %xmm0
20984 // movss <1.0f>, %xmm0
20987 // xorps %xmm0, %xmm0
20991 MachineInstr *CascadedCMOV = nullptr;
20992 MachineInstr *LastCMOV = MI;
20993 X86::CondCode CC = X86::CondCode(MI->getOperand(3).getImm());
20994 X86::CondCode OppCC = X86::GetOppositeBranchCondition(CC);
20995 MachineBasicBlock::iterator NextMIIt =
20996 std::next(MachineBasicBlock::iterator(MI));
20998 // Check for case 1, where there are multiple CMOVs with the same condition
20999 // first. Of the two cases of multiple CMOV lowerings, case 1 reduces the
21000 // number of jumps the most.
21002 if (isCMOVPseudo(MI)) {
21003 // See if we have a string of CMOVS with the same condition.
21004 while (NextMIIt != BB->end() &&
21005 isCMOVPseudo(NextMIIt) &&
21006 (NextMIIt->getOperand(3).getImm() == CC ||
21007 NextMIIt->getOperand(3).getImm() == OppCC)) {
21008 LastCMOV = &*NextMIIt;
21013 // This checks for case 2, but only do this if we didn't already find
21014 // case 1, as indicated by LastCMOV == MI.
21015 if (LastCMOV == MI &&
21016 NextMIIt != BB->end() && NextMIIt->getOpcode() == MI->getOpcode() &&
21017 NextMIIt->getOperand(2).getReg() == MI->getOperand(2).getReg() &&
21018 NextMIIt->getOperand(1).getReg() == MI->getOperand(0).getReg()) {
21019 CascadedCMOV = &*NextMIIt;
21022 MachineBasicBlock *jcc1MBB = nullptr;
21024 // If we have a cascaded CMOV, we lower it to two successive branches to
21025 // the same block. EFLAGS is used by both, so mark it as live in the second.
21026 if (CascadedCMOV) {
21027 jcc1MBB = F->CreateMachineBasicBlock(LLVM_BB);
21028 F->insert(It, jcc1MBB);
21029 jcc1MBB->addLiveIn(X86::EFLAGS);
21032 MachineBasicBlock *copy0MBB = F->CreateMachineBasicBlock(LLVM_BB);
21033 MachineBasicBlock *sinkMBB = F->CreateMachineBasicBlock(LLVM_BB);
21034 F->insert(It, copy0MBB);
21035 F->insert(It, sinkMBB);
21037 // If the EFLAGS register isn't dead in the terminator, then claim that it's
21038 // live into the sink and copy blocks.
21039 const TargetRegisterInfo *TRI = Subtarget->getRegisterInfo();
21041 MachineInstr *LastEFLAGSUser = CascadedCMOV ? CascadedCMOV : LastCMOV;
21042 if (!LastEFLAGSUser->killsRegister(X86::EFLAGS) &&
21043 !checkAndUpdateEFLAGSKill(LastEFLAGSUser, BB, TRI)) {
21044 copy0MBB->addLiveIn(X86::EFLAGS);
21045 sinkMBB->addLiveIn(X86::EFLAGS);
21048 // Transfer the remainder of BB and its successor edges to sinkMBB.
21049 sinkMBB->splice(sinkMBB->begin(), BB,
21050 std::next(MachineBasicBlock::iterator(LastCMOV)), BB->end());
21051 sinkMBB->transferSuccessorsAndUpdatePHIs(BB);
21053 // Add the true and fallthrough blocks as its successors.
21054 if (CascadedCMOV) {
21055 // The fallthrough block may be jcc1MBB, if we have a cascaded CMOV.
21056 BB->addSuccessor(jcc1MBB);
21058 // In that case, jcc1MBB will itself fallthrough the copy0MBB, and
21059 // jump to the sinkMBB.
21060 jcc1MBB->addSuccessor(copy0MBB);
21061 jcc1MBB->addSuccessor(sinkMBB);
21063 BB->addSuccessor(copy0MBB);
21066 // The true block target of the first (or only) branch is always sinkMBB.
21067 BB->addSuccessor(sinkMBB);
21069 // Create the conditional branch instruction.
21070 unsigned Opc = X86::GetCondBranchFromCond(CC);
21071 BuildMI(BB, DL, TII->get(Opc)).addMBB(sinkMBB);
21073 if (CascadedCMOV) {
21074 unsigned Opc2 = X86::GetCondBranchFromCond(
21075 (X86::CondCode)CascadedCMOV->getOperand(3).getImm());
21076 BuildMI(jcc1MBB, DL, TII->get(Opc2)).addMBB(sinkMBB);
21080 // %FalseValue = ...
21081 // # fallthrough to sinkMBB
21082 copy0MBB->addSuccessor(sinkMBB);
21085 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, thisMBB ]
21087 MachineBasicBlock::iterator MIItBegin = MachineBasicBlock::iterator(MI);
21088 MachineBasicBlock::iterator MIItEnd =
21089 std::next(MachineBasicBlock::iterator(LastCMOV));
21090 MachineBasicBlock::iterator SinkInsertionPoint = sinkMBB->begin();
21091 DenseMap<unsigned, std::pair<unsigned, unsigned>> RegRewriteTable;
21092 MachineInstrBuilder MIB;
21094 // As we are creating the PHIs, we have to be careful if there is more than
21095 // one. Later CMOVs may reference the results of earlier CMOVs, but later
21096 // PHIs have to reference the individual true/false inputs from earlier PHIs.
21097 // That also means that PHI construction must work forward from earlier to
21098 // later, and that the code must maintain a mapping from earlier PHI's
21099 // destination registers, and the registers that went into the PHI.
21101 for (MachineBasicBlock::iterator MIIt = MIItBegin; MIIt != MIItEnd; ++MIIt) {
21102 unsigned DestReg = MIIt->getOperand(0).getReg();
21103 unsigned Op1Reg = MIIt->getOperand(1).getReg();
21104 unsigned Op2Reg = MIIt->getOperand(2).getReg();
21106 // If this CMOV we are generating is the opposite condition from
21107 // the jump we generated, then we have to swap the operands for the
21108 // PHI that is going to be generated.
21109 if (MIIt->getOperand(3).getImm() == OppCC)
21110 std::swap(Op1Reg, Op2Reg);
21112 if (RegRewriteTable.find(Op1Reg) != RegRewriteTable.end())
21113 Op1Reg = RegRewriteTable[Op1Reg].first;
21115 if (RegRewriteTable.find(Op2Reg) != RegRewriteTable.end())
21116 Op2Reg = RegRewriteTable[Op2Reg].second;
21118 MIB = BuildMI(*sinkMBB, SinkInsertionPoint, DL,
21119 TII->get(X86::PHI), DestReg)
21120 .addReg(Op1Reg).addMBB(copy0MBB)
21121 .addReg(Op2Reg).addMBB(thisMBB);
21123 // Add this PHI to the rewrite table.
21124 RegRewriteTable[DestReg] = std::make_pair(Op1Reg, Op2Reg);
21127 // If we have a cascaded CMOV, the second Jcc provides the same incoming
21128 // value as the first Jcc (the True operand of the SELECT_CC/CMOV nodes).
21129 if (CascadedCMOV) {
21130 MIB.addReg(MI->getOperand(2).getReg()).addMBB(jcc1MBB);
21131 // Copy the PHI result to the register defined by the second CMOV.
21132 BuildMI(*sinkMBB, std::next(MachineBasicBlock::iterator(MIB.getInstr())),
21133 DL, TII->get(TargetOpcode::COPY),
21134 CascadedCMOV->getOperand(0).getReg())
21135 .addReg(MI->getOperand(0).getReg());
21136 CascadedCMOV->eraseFromParent();
21139 // Now remove the CMOV(s).
21140 for (MachineBasicBlock::iterator MIIt = MIItBegin; MIIt != MIItEnd; )
21141 (MIIt++)->eraseFromParent();
21146 MachineBasicBlock *
21147 X86TargetLowering::EmitLoweredAtomicFP(MachineInstr *MI,
21148 MachineBasicBlock *BB) const {
21149 // Combine the following atomic floating-point modification pattern:
21150 // a.store(reg OP a.load(acquire), release)
21151 // Transform them into:
21152 // OPss (%gpr), %xmm
21153 // movss %xmm, (%gpr)
21154 // Or sd equivalent for 64-bit operations.
21156 switch (MI->getOpcode()) {
21157 default: llvm_unreachable("unexpected instr type for EmitLoweredAtomicFP");
21158 case X86::RELEASE_FADD32mr: MOp = X86::MOVSSmr; FOp = X86::ADDSSrm; break;
21159 case X86::RELEASE_FADD64mr: MOp = X86::MOVSDmr; FOp = X86::ADDSDrm; break;
21161 const X86InstrInfo *TII = Subtarget->getInstrInfo();
21162 DebugLoc DL = MI->getDebugLoc();
21163 MachineRegisterInfo &MRI = BB->getParent()->getRegInfo();
21164 MachineOperand MSrc = MI->getOperand(0);
21165 unsigned VSrc = MI->getOperand(5).getReg();
21166 const MachineOperand &Disp = MI->getOperand(3);
21167 MachineOperand ZeroDisp = MachineOperand::CreateImm(0);
21168 bool hasDisp = Disp.isGlobal() || Disp.isImm();
21169 if (hasDisp && MSrc.isReg())
21170 MSrc.setIsKill(false);
21171 MachineInstrBuilder MIM = BuildMI(*BB, MI, DL, TII->get(MOp))
21172 .addOperand(/*Base=*/MSrc)
21173 .addImm(/*Scale=*/1)
21174 .addReg(/*Index=*/0)
21175 .addDisp(hasDisp ? Disp : ZeroDisp, /*off=*/0)
21177 MachineInstr *MIO = BuildMI(*BB, (MachineInstr *)MIM, DL, TII->get(FOp),
21178 MRI.createVirtualRegister(MRI.getRegClass(VSrc)))
21180 .addOperand(/*Base=*/MSrc)
21181 .addImm(/*Scale=*/1)
21182 .addReg(/*Index=*/0)
21183 .addDisp(hasDisp ? Disp : ZeroDisp, /*off=*/0)
21184 .addReg(/*Segment=*/0);
21185 MIM.addReg(MIO->getOperand(0).getReg(), RegState::Kill);
21186 MI->eraseFromParent(); // The pseudo instruction is gone now.
21190 MachineBasicBlock *
21191 X86TargetLowering::EmitLoweredSegAlloca(MachineInstr *MI,
21192 MachineBasicBlock *BB) const {
21193 MachineFunction *MF = BB->getParent();
21194 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
21195 DebugLoc DL = MI->getDebugLoc();
21196 const BasicBlock *LLVM_BB = BB->getBasicBlock();
21198 assert(MF->shouldSplitStack());
21200 const bool Is64Bit = Subtarget->is64Bit();
21201 const bool IsLP64 = Subtarget->isTarget64BitLP64();
21203 const unsigned TlsReg = Is64Bit ? X86::FS : X86::GS;
21204 const unsigned TlsOffset = IsLP64 ? 0x70 : Is64Bit ? 0x40 : 0x30;
21207 // ... [Till the alloca]
21208 // If stacklet is not large enough, jump to mallocMBB
21211 // Allocate by subtracting from RSP
21212 // Jump to continueMBB
21215 // Allocate by call to runtime
21219 // [rest of original BB]
21222 MachineBasicBlock *mallocMBB = MF->CreateMachineBasicBlock(LLVM_BB);
21223 MachineBasicBlock *bumpMBB = MF->CreateMachineBasicBlock(LLVM_BB);
21224 MachineBasicBlock *continueMBB = MF->CreateMachineBasicBlock(LLVM_BB);
21226 MachineRegisterInfo &MRI = MF->getRegInfo();
21227 const TargetRegisterClass *AddrRegClass =
21228 getRegClassFor(getPointerTy(MF->getDataLayout()));
21230 unsigned mallocPtrVReg = MRI.createVirtualRegister(AddrRegClass),
21231 bumpSPPtrVReg = MRI.createVirtualRegister(AddrRegClass),
21232 tmpSPVReg = MRI.createVirtualRegister(AddrRegClass),
21233 SPLimitVReg = MRI.createVirtualRegister(AddrRegClass),
21234 sizeVReg = MI->getOperand(1).getReg(),
21235 physSPReg = IsLP64 || Subtarget->isTargetNaCl64() ? X86::RSP : X86::ESP;
21237 MachineFunction::iterator MBBIter = ++BB->getIterator();
21239 MF->insert(MBBIter, bumpMBB);
21240 MF->insert(MBBIter, mallocMBB);
21241 MF->insert(MBBIter, continueMBB);
21243 continueMBB->splice(continueMBB->begin(), BB,
21244 std::next(MachineBasicBlock::iterator(MI)), BB->end());
21245 continueMBB->transferSuccessorsAndUpdatePHIs(BB);
21247 // Add code to the main basic block to check if the stack limit has been hit,
21248 // and if so, jump to mallocMBB otherwise to bumpMBB.
21249 BuildMI(BB, DL, TII->get(TargetOpcode::COPY), tmpSPVReg).addReg(physSPReg);
21250 BuildMI(BB, DL, TII->get(IsLP64 ? X86::SUB64rr:X86::SUB32rr), SPLimitVReg)
21251 .addReg(tmpSPVReg).addReg(sizeVReg);
21252 BuildMI(BB, DL, TII->get(IsLP64 ? X86::CMP64mr:X86::CMP32mr))
21253 .addReg(0).addImm(1).addReg(0).addImm(TlsOffset).addReg(TlsReg)
21254 .addReg(SPLimitVReg);
21255 BuildMI(BB, DL, TII->get(X86::JG_1)).addMBB(mallocMBB);
21257 // bumpMBB simply decreases the stack pointer, since we know the current
21258 // stacklet has enough space.
21259 BuildMI(bumpMBB, DL, TII->get(TargetOpcode::COPY), physSPReg)
21260 .addReg(SPLimitVReg);
21261 BuildMI(bumpMBB, DL, TII->get(TargetOpcode::COPY), bumpSPPtrVReg)
21262 .addReg(SPLimitVReg);
21263 BuildMI(bumpMBB, DL, TII->get(X86::JMP_1)).addMBB(continueMBB);
21265 // Calls into a routine in libgcc to allocate more space from the heap.
21266 const uint32_t *RegMask =
21267 Subtarget->getRegisterInfo()->getCallPreservedMask(*MF, CallingConv::C);
21269 BuildMI(mallocMBB, DL, TII->get(X86::MOV64rr), X86::RDI)
21271 BuildMI(mallocMBB, DL, TII->get(X86::CALL64pcrel32))
21272 .addExternalSymbol("__morestack_allocate_stack_space")
21273 .addRegMask(RegMask)
21274 .addReg(X86::RDI, RegState::Implicit)
21275 .addReg(X86::RAX, RegState::ImplicitDefine);
21276 } else if (Is64Bit) {
21277 BuildMI(mallocMBB, DL, TII->get(X86::MOV32rr), X86::EDI)
21279 BuildMI(mallocMBB, DL, TII->get(X86::CALL64pcrel32))
21280 .addExternalSymbol("__morestack_allocate_stack_space")
21281 .addRegMask(RegMask)
21282 .addReg(X86::EDI, RegState::Implicit)
21283 .addReg(X86::EAX, RegState::ImplicitDefine);
21285 BuildMI(mallocMBB, DL, TII->get(X86::SUB32ri), physSPReg).addReg(physSPReg)
21287 BuildMI(mallocMBB, DL, TII->get(X86::PUSH32r)).addReg(sizeVReg);
21288 BuildMI(mallocMBB, DL, TII->get(X86::CALLpcrel32))
21289 .addExternalSymbol("__morestack_allocate_stack_space")
21290 .addRegMask(RegMask)
21291 .addReg(X86::EAX, RegState::ImplicitDefine);
21295 BuildMI(mallocMBB, DL, TII->get(X86::ADD32ri), physSPReg).addReg(physSPReg)
21298 BuildMI(mallocMBB, DL, TII->get(TargetOpcode::COPY), mallocPtrVReg)
21299 .addReg(IsLP64 ? X86::RAX : X86::EAX);
21300 BuildMI(mallocMBB, DL, TII->get(X86::JMP_1)).addMBB(continueMBB);
21302 // Set up the CFG correctly.
21303 BB->addSuccessor(bumpMBB);
21304 BB->addSuccessor(mallocMBB);
21305 mallocMBB->addSuccessor(continueMBB);
21306 bumpMBB->addSuccessor(continueMBB);
21308 // Take care of the PHI nodes.
21309 BuildMI(*continueMBB, continueMBB->begin(), DL, TII->get(X86::PHI),
21310 MI->getOperand(0).getReg())
21311 .addReg(mallocPtrVReg).addMBB(mallocMBB)
21312 .addReg(bumpSPPtrVReg).addMBB(bumpMBB);
21314 // Delete the original pseudo instruction.
21315 MI->eraseFromParent();
21318 return continueMBB;
21321 MachineBasicBlock *
21322 X86TargetLowering::EmitLoweredWinAlloca(MachineInstr *MI,
21323 MachineBasicBlock *BB) const {
21324 DebugLoc DL = MI->getDebugLoc();
21326 assert(!Subtarget->isTargetMachO());
21328 Subtarget->getFrameLowering()->emitStackProbeCall(*BB->getParent(), *BB, MI,
21331 MI->eraseFromParent(); // The pseudo instruction is gone now.
21335 MachineBasicBlock *
21336 X86TargetLowering::EmitLoweredTLSCall(MachineInstr *MI,
21337 MachineBasicBlock *BB) const {
21338 // This is pretty easy. We're taking the value that we received from
21339 // our load from the relocation, sticking it in either RDI (x86-64)
21340 // or EAX and doing an indirect call. The return value will then
21341 // be in the normal return register.
21342 MachineFunction *F = BB->getParent();
21343 const X86InstrInfo *TII = Subtarget->getInstrInfo();
21344 DebugLoc DL = MI->getDebugLoc();
21346 assert(Subtarget->isTargetDarwin() && "Darwin only instr emitted?");
21347 assert(MI->getOperand(3).isGlobal() && "This should be a global");
21349 // Get a register mask for the lowered call.
21350 // FIXME: The 32-bit calls have non-standard calling conventions. Use a
21351 // proper register mask.
21352 const uint32_t *RegMask =
21353 Subtarget->getRegisterInfo()->getCallPreservedMask(*F, CallingConv::C);
21354 if (Subtarget->is64Bit()) {
21355 MachineInstrBuilder MIB = BuildMI(*BB, MI, DL,
21356 TII->get(X86::MOV64rm), X86::RDI)
21358 .addImm(0).addReg(0)
21359 .addGlobalAddress(MI->getOperand(3).getGlobal(), 0,
21360 MI->getOperand(3).getTargetFlags())
21362 MIB = BuildMI(*BB, MI, DL, TII->get(X86::CALL64m));
21363 addDirectMem(MIB, X86::RDI);
21364 MIB.addReg(X86::RAX, RegState::ImplicitDefine).addRegMask(RegMask);
21365 } else if (F->getTarget().getRelocationModel() != Reloc::PIC_) {
21366 MachineInstrBuilder MIB = BuildMI(*BB, MI, DL,
21367 TII->get(X86::MOV32rm), X86::EAX)
21369 .addImm(0).addReg(0)
21370 .addGlobalAddress(MI->getOperand(3).getGlobal(), 0,
21371 MI->getOperand(3).getTargetFlags())
21373 MIB = BuildMI(*BB, MI, DL, TII->get(X86::CALL32m));
21374 addDirectMem(MIB, X86::EAX);
21375 MIB.addReg(X86::EAX, RegState::ImplicitDefine).addRegMask(RegMask);
21377 MachineInstrBuilder MIB = BuildMI(*BB, MI, DL,
21378 TII->get(X86::MOV32rm), X86::EAX)
21379 .addReg(TII->getGlobalBaseReg(F))
21380 .addImm(0).addReg(0)
21381 .addGlobalAddress(MI->getOperand(3).getGlobal(), 0,
21382 MI->getOperand(3).getTargetFlags())
21384 MIB = BuildMI(*BB, MI, DL, TII->get(X86::CALL32m));
21385 addDirectMem(MIB, X86::EAX);
21386 MIB.addReg(X86::EAX, RegState::ImplicitDefine).addRegMask(RegMask);
21389 MI->eraseFromParent(); // The pseudo instruction is gone now.
21393 MachineBasicBlock *
21394 X86TargetLowering::emitEHSjLjSetJmp(MachineInstr *MI,
21395 MachineBasicBlock *MBB) const {
21396 DebugLoc DL = MI->getDebugLoc();
21397 MachineFunction *MF = MBB->getParent();
21398 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
21399 MachineRegisterInfo &MRI = MF->getRegInfo();
21401 const BasicBlock *BB = MBB->getBasicBlock();
21402 MachineFunction::iterator I = ++MBB->getIterator();
21404 // Memory Reference
21405 MachineInstr::mmo_iterator MMOBegin = MI->memoperands_begin();
21406 MachineInstr::mmo_iterator MMOEnd = MI->memoperands_end();
21409 unsigned MemOpndSlot = 0;
21411 unsigned CurOp = 0;
21413 DstReg = MI->getOperand(CurOp++).getReg();
21414 const TargetRegisterClass *RC = MRI.getRegClass(DstReg);
21415 assert(RC->hasType(MVT::i32) && "Invalid destination!");
21416 unsigned mainDstReg = MRI.createVirtualRegister(RC);
21417 unsigned restoreDstReg = MRI.createVirtualRegister(RC);
21419 MemOpndSlot = CurOp;
21421 MVT PVT = getPointerTy(MF->getDataLayout());
21422 assert((PVT == MVT::i64 || PVT == MVT::i32) &&
21423 "Invalid Pointer Size!");
21425 // For v = setjmp(buf), we generate
21428 // buf[LabelOffset] = restoreMBB <-- takes address of restoreMBB
21429 // SjLjSetup restoreMBB
21435 // v = phi(main, restore)
21438 // if base pointer being used, load it from frame
21441 MachineBasicBlock *thisMBB = MBB;
21442 MachineBasicBlock *mainMBB = MF->CreateMachineBasicBlock(BB);
21443 MachineBasicBlock *sinkMBB = MF->CreateMachineBasicBlock(BB);
21444 MachineBasicBlock *restoreMBB = MF->CreateMachineBasicBlock(BB);
21445 MF->insert(I, mainMBB);
21446 MF->insert(I, sinkMBB);
21447 MF->push_back(restoreMBB);
21448 restoreMBB->setHasAddressTaken();
21450 MachineInstrBuilder MIB;
21452 // Transfer the remainder of BB and its successor edges to sinkMBB.
21453 sinkMBB->splice(sinkMBB->begin(), MBB,
21454 std::next(MachineBasicBlock::iterator(MI)), MBB->end());
21455 sinkMBB->transferSuccessorsAndUpdatePHIs(MBB);
21458 unsigned PtrStoreOpc = 0;
21459 unsigned LabelReg = 0;
21460 const int64_t LabelOffset = 1 * PVT.getStoreSize();
21461 Reloc::Model RM = MF->getTarget().getRelocationModel();
21462 bool UseImmLabel = (MF->getTarget().getCodeModel() == CodeModel::Small) &&
21463 (RM == Reloc::Static || RM == Reloc::DynamicNoPIC);
21465 // Prepare IP either in reg or imm.
21466 if (!UseImmLabel) {
21467 PtrStoreOpc = (PVT == MVT::i64) ? X86::MOV64mr : X86::MOV32mr;
21468 const TargetRegisterClass *PtrRC = getRegClassFor(PVT);
21469 LabelReg = MRI.createVirtualRegister(PtrRC);
21470 if (Subtarget->is64Bit()) {
21471 MIB = BuildMI(*thisMBB, MI, DL, TII->get(X86::LEA64r), LabelReg)
21475 .addMBB(restoreMBB)
21478 const X86InstrInfo *XII = static_cast<const X86InstrInfo*>(TII);
21479 MIB = BuildMI(*thisMBB, MI, DL, TII->get(X86::LEA32r), LabelReg)
21480 .addReg(XII->getGlobalBaseReg(MF))
21483 .addMBB(restoreMBB, Subtarget->ClassifyBlockAddressReference())
21487 PtrStoreOpc = (PVT == MVT::i64) ? X86::MOV64mi32 : X86::MOV32mi;
21489 MIB = BuildMI(*thisMBB, MI, DL, TII->get(PtrStoreOpc));
21490 for (unsigned i = 0; i < X86::AddrNumOperands; ++i) {
21491 if (i == X86::AddrDisp)
21492 MIB.addDisp(MI->getOperand(MemOpndSlot + i), LabelOffset);
21494 MIB.addOperand(MI->getOperand(MemOpndSlot + i));
21497 MIB.addReg(LabelReg);
21499 MIB.addMBB(restoreMBB);
21500 MIB.setMemRefs(MMOBegin, MMOEnd);
21502 MIB = BuildMI(*thisMBB, MI, DL, TII->get(X86::EH_SjLj_Setup))
21503 .addMBB(restoreMBB);
21505 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
21506 MIB.addRegMask(RegInfo->getNoPreservedMask());
21507 thisMBB->addSuccessor(mainMBB);
21508 thisMBB->addSuccessor(restoreMBB);
21512 BuildMI(mainMBB, DL, TII->get(X86::MOV32r0), mainDstReg);
21513 mainMBB->addSuccessor(sinkMBB);
21516 BuildMI(*sinkMBB, sinkMBB->begin(), DL,
21517 TII->get(X86::PHI), DstReg)
21518 .addReg(mainDstReg).addMBB(mainMBB)
21519 .addReg(restoreDstReg).addMBB(restoreMBB);
21522 if (RegInfo->hasBasePointer(*MF)) {
21523 const bool Uses64BitFramePtr =
21524 Subtarget->isTarget64BitLP64() || Subtarget->isTargetNaCl64();
21525 X86MachineFunctionInfo *X86FI = MF->getInfo<X86MachineFunctionInfo>();
21526 X86FI->setRestoreBasePointer(MF);
21527 unsigned FramePtr = RegInfo->getFrameRegister(*MF);
21528 unsigned BasePtr = RegInfo->getBaseRegister();
21529 unsigned Opm = Uses64BitFramePtr ? X86::MOV64rm : X86::MOV32rm;
21530 addRegOffset(BuildMI(restoreMBB, DL, TII->get(Opm), BasePtr),
21531 FramePtr, true, X86FI->getRestoreBasePointerOffset())
21532 .setMIFlag(MachineInstr::FrameSetup);
21534 BuildMI(restoreMBB, DL, TII->get(X86::MOV32ri), restoreDstReg).addImm(1);
21535 BuildMI(restoreMBB, DL, TII->get(X86::JMP_1)).addMBB(sinkMBB);
21536 restoreMBB->addSuccessor(sinkMBB);
21538 MI->eraseFromParent();
21542 MachineBasicBlock *
21543 X86TargetLowering::emitEHSjLjLongJmp(MachineInstr *MI,
21544 MachineBasicBlock *MBB) const {
21545 DebugLoc DL = MI->getDebugLoc();
21546 MachineFunction *MF = MBB->getParent();
21547 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
21548 MachineRegisterInfo &MRI = MF->getRegInfo();
21550 // Memory Reference
21551 MachineInstr::mmo_iterator MMOBegin = MI->memoperands_begin();
21552 MachineInstr::mmo_iterator MMOEnd = MI->memoperands_end();
21554 MVT PVT = getPointerTy(MF->getDataLayout());
21555 assert((PVT == MVT::i64 || PVT == MVT::i32) &&
21556 "Invalid Pointer Size!");
21558 const TargetRegisterClass *RC =
21559 (PVT == MVT::i64) ? &X86::GR64RegClass : &X86::GR32RegClass;
21560 unsigned Tmp = MRI.createVirtualRegister(RC);
21561 // Since FP is only updated here but NOT referenced, it's treated as GPR.
21562 const X86RegisterInfo *RegInfo = Subtarget->getRegisterInfo();
21563 unsigned FP = (PVT == MVT::i64) ? X86::RBP : X86::EBP;
21564 unsigned SP = RegInfo->getStackRegister();
21566 MachineInstrBuilder MIB;
21568 const int64_t LabelOffset = 1 * PVT.getStoreSize();
21569 const int64_t SPOffset = 2 * PVT.getStoreSize();
21571 unsigned PtrLoadOpc = (PVT == MVT::i64) ? X86::MOV64rm : X86::MOV32rm;
21572 unsigned IJmpOpc = (PVT == MVT::i64) ? X86::JMP64r : X86::JMP32r;
21575 MIB = BuildMI(*MBB, MI, DL, TII->get(PtrLoadOpc), FP);
21576 for (unsigned i = 0; i < X86::AddrNumOperands; ++i)
21577 MIB.addOperand(MI->getOperand(i));
21578 MIB.setMemRefs(MMOBegin, MMOEnd);
21580 MIB = BuildMI(*MBB, MI, DL, TII->get(PtrLoadOpc), Tmp);
21581 for (unsigned i = 0; i < X86::AddrNumOperands; ++i) {
21582 if (i == X86::AddrDisp)
21583 MIB.addDisp(MI->getOperand(i), LabelOffset);
21585 MIB.addOperand(MI->getOperand(i));
21587 MIB.setMemRefs(MMOBegin, MMOEnd);
21589 MIB = BuildMI(*MBB, MI, DL, TII->get(PtrLoadOpc), SP);
21590 for (unsigned i = 0; i < X86::AddrNumOperands; ++i) {
21591 if (i == X86::AddrDisp)
21592 MIB.addDisp(MI->getOperand(i), SPOffset);
21594 MIB.addOperand(MI->getOperand(i));
21596 MIB.setMemRefs(MMOBegin, MMOEnd);
21598 BuildMI(*MBB, MI, DL, TII->get(IJmpOpc)).addReg(Tmp);
21600 MI->eraseFromParent();
21604 // Replace 213-type (isel default) FMA3 instructions with 231-type for
21605 // accumulator loops. Writing back to the accumulator allows the coalescer
21606 // to remove extra copies in the loop.
21607 // FIXME: Do this on AVX512. We don't support 231 variants yet (PR23937).
21608 MachineBasicBlock *
21609 X86TargetLowering::emitFMA3Instr(MachineInstr *MI,
21610 MachineBasicBlock *MBB) const {
21611 MachineOperand &AddendOp = MI->getOperand(3);
21613 // Bail out early if the addend isn't a register - we can't switch these.
21614 if (!AddendOp.isReg())
21617 MachineFunction &MF = *MBB->getParent();
21618 MachineRegisterInfo &MRI = MF.getRegInfo();
21620 // Check whether the addend is defined by a PHI:
21621 assert(MRI.hasOneDef(AddendOp.getReg()) && "Multiple defs in SSA?");
21622 MachineInstr &AddendDef = *MRI.def_instr_begin(AddendOp.getReg());
21623 if (!AddendDef.isPHI())
21626 // Look for the following pattern:
21628 // %addend = phi [%entry, 0], [%loop, %result]
21630 // %result<tied1> = FMA213 %m2<tied0>, %m1, %addend
21634 // %addend = phi [%entry, 0], [%loop, %result]
21636 // %result<tied1> = FMA231 %addend<tied0>, %m1, %m2
21638 for (unsigned i = 1, e = AddendDef.getNumOperands(); i < e; i += 2) {
21639 assert(AddendDef.getOperand(i).isReg());
21640 MachineOperand PHISrcOp = AddendDef.getOperand(i);
21641 MachineInstr &PHISrcInst = *MRI.def_instr_begin(PHISrcOp.getReg());
21642 if (&PHISrcInst == MI) {
21643 // Found a matching instruction.
21644 unsigned NewFMAOpc = 0;
21645 switch (MI->getOpcode()) {
21646 case X86::VFMADDPDr213r: NewFMAOpc = X86::VFMADDPDr231r; break;
21647 case X86::VFMADDPSr213r: NewFMAOpc = X86::VFMADDPSr231r; break;
21648 case X86::VFMADDSDr213r: NewFMAOpc = X86::VFMADDSDr231r; break;
21649 case X86::VFMADDSSr213r: NewFMAOpc = X86::VFMADDSSr231r; break;
21650 case X86::VFMSUBPDr213r: NewFMAOpc = X86::VFMSUBPDr231r; break;
21651 case X86::VFMSUBPSr213r: NewFMAOpc = X86::VFMSUBPSr231r; break;
21652 case X86::VFMSUBSDr213r: NewFMAOpc = X86::VFMSUBSDr231r; break;
21653 case X86::VFMSUBSSr213r: NewFMAOpc = X86::VFMSUBSSr231r; break;
21654 case X86::VFNMADDPDr213r: NewFMAOpc = X86::VFNMADDPDr231r; break;
21655 case X86::VFNMADDPSr213r: NewFMAOpc = X86::VFNMADDPSr231r; break;
21656 case X86::VFNMADDSDr213r: NewFMAOpc = X86::VFNMADDSDr231r; break;
21657 case X86::VFNMADDSSr213r: NewFMAOpc = X86::VFNMADDSSr231r; break;
21658 case X86::VFNMSUBPDr213r: NewFMAOpc = X86::VFNMSUBPDr231r; break;
21659 case X86::VFNMSUBPSr213r: NewFMAOpc = X86::VFNMSUBPSr231r; break;
21660 case X86::VFNMSUBSDr213r: NewFMAOpc = X86::VFNMSUBSDr231r; break;
21661 case X86::VFNMSUBSSr213r: NewFMAOpc = X86::VFNMSUBSSr231r; break;
21662 case X86::VFMADDSUBPDr213r: NewFMAOpc = X86::VFMADDSUBPDr231r; break;
21663 case X86::VFMADDSUBPSr213r: NewFMAOpc = X86::VFMADDSUBPSr231r; break;
21664 case X86::VFMSUBADDPDr213r: NewFMAOpc = X86::VFMSUBADDPDr231r; break;
21665 case X86::VFMSUBADDPSr213r: NewFMAOpc = X86::VFMSUBADDPSr231r; break;
21667 case X86::VFMADDPDr213rY: NewFMAOpc = X86::VFMADDPDr231rY; break;
21668 case X86::VFMADDPSr213rY: NewFMAOpc = X86::VFMADDPSr231rY; break;
21669 case X86::VFMSUBPDr213rY: NewFMAOpc = X86::VFMSUBPDr231rY; break;
21670 case X86::VFMSUBPSr213rY: NewFMAOpc = X86::VFMSUBPSr231rY; break;
21671 case X86::VFNMADDPDr213rY: NewFMAOpc = X86::VFNMADDPDr231rY; break;
21672 case X86::VFNMADDPSr213rY: NewFMAOpc = X86::VFNMADDPSr231rY; break;
21673 case X86::VFNMSUBPDr213rY: NewFMAOpc = X86::VFNMSUBPDr231rY; break;
21674 case X86::VFNMSUBPSr213rY: NewFMAOpc = X86::VFNMSUBPSr231rY; break;
21675 case X86::VFMADDSUBPDr213rY: NewFMAOpc = X86::VFMADDSUBPDr231rY; break;
21676 case X86::VFMADDSUBPSr213rY: NewFMAOpc = X86::VFMADDSUBPSr231rY; break;
21677 case X86::VFMSUBADDPDr213rY: NewFMAOpc = X86::VFMSUBADDPDr231rY; break;
21678 case X86::VFMSUBADDPSr213rY: NewFMAOpc = X86::VFMSUBADDPSr231rY; break;
21679 default: llvm_unreachable("Unrecognized FMA variant.");
21682 const TargetInstrInfo &TII = *Subtarget->getInstrInfo();
21683 MachineInstrBuilder MIB =
21684 BuildMI(MF, MI->getDebugLoc(), TII.get(NewFMAOpc))
21685 .addOperand(MI->getOperand(0))
21686 .addOperand(MI->getOperand(3))
21687 .addOperand(MI->getOperand(2))
21688 .addOperand(MI->getOperand(1));
21689 MBB->insert(MachineBasicBlock::iterator(MI), MIB);
21690 MI->eraseFromParent();
21697 MachineBasicBlock *
21698 X86TargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
21699 MachineBasicBlock *BB) const {
21700 switch (MI->getOpcode()) {
21701 default: llvm_unreachable("Unexpected instr type to insert");
21702 case X86::TAILJMPd64:
21703 case X86::TAILJMPr64:
21704 case X86::TAILJMPm64:
21705 case X86::TAILJMPd64_REX:
21706 case X86::TAILJMPr64_REX:
21707 case X86::TAILJMPm64_REX:
21708 llvm_unreachable("TAILJMP64 would not be touched here.");
21709 case X86::TCRETURNdi64:
21710 case X86::TCRETURNri64:
21711 case X86::TCRETURNmi64:
21713 case X86::WIN_ALLOCA:
21714 return EmitLoweredWinAlloca(MI, BB);
21715 case X86::SEG_ALLOCA_32:
21716 case X86::SEG_ALLOCA_64:
21717 return EmitLoweredSegAlloca(MI, BB);
21718 case X86::TLSCall_32:
21719 case X86::TLSCall_64:
21720 return EmitLoweredTLSCall(MI, BB);
21721 case X86::CMOV_FR32:
21722 case X86::CMOV_FR64:
21723 case X86::CMOV_GR8:
21724 case X86::CMOV_GR16:
21725 case X86::CMOV_GR32:
21726 case X86::CMOV_RFP32:
21727 case X86::CMOV_RFP64:
21728 case X86::CMOV_RFP80:
21729 case X86::CMOV_V2F64:
21730 case X86::CMOV_V2I64:
21731 case X86::CMOV_V4F32:
21732 case X86::CMOV_V4F64:
21733 case X86::CMOV_V4I64:
21734 case X86::CMOV_V16F32:
21735 case X86::CMOV_V8F32:
21736 case X86::CMOV_V8F64:
21737 case X86::CMOV_V8I64:
21738 case X86::CMOV_V8I1:
21739 case X86::CMOV_V16I1:
21740 case X86::CMOV_V32I1:
21741 case X86::CMOV_V64I1:
21742 return EmitLoweredSelect(MI, BB);
21744 case X86::RELEASE_FADD32mr:
21745 case X86::RELEASE_FADD64mr:
21746 return EmitLoweredAtomicFP(MI, BB);
21748 case X86::FP32_TO_INT16_IN_MEM:
21749 case X86::FP32_TO_INT32_IN_MEM:
21750 case X86::FP32_TO_INT64_IN_MEM:
21751 case X86::FP64_TO_INT16_IN_MEM:
21752 case X86::FP64_TO_INT32_IN_MEM:
21753 case X86::FP64_TO_INT64_IN_MEM:
21754 case X86::FP80_TO_INT16_IN_MEM:
21755 case X86::FP80_TO_INT32_IN_MEM:
21756 case X86::FP80_TO_INT64_IN_MEM: {
21757 MachineFunction *F = BB->getParent();
21758 const TargetInstrInfo *TII = Subtarget->getInstrInfo();
21759 DebugLoc DL = MI->getDebugLoc();
21761 // Change the floating point control register to use "round towards zero"
21762 // mode when truncating to an integer value.
21763 int CWFrameIdx = F->getFrameInfo()->CreateStackObject(2, 2, false);
21764 addFrameReference(BuildMI(*BB, MI, DL,
21765 TII->get(X86::FNSTCW16m)), CWFrameIdx);
21767 // Load the old value of the high byte of the control word...
21769 F->getRegInfo().createVirtualRegister(&X86::GR16RegClass);
21770 addFrameReference(BuildMI(*BB, MI, DL, TII->get(X86::MOV16rm), OldCW),
21773 // Set the high part to be round to zero...
21774 addFrameReference(BuildMI(*BB, MI, DL, TII->get(X86::MOV16mi)), CWFrameIdx)
21777 // Reload the modified control word now...
21778 addFrameReference(BuildMI(*BB, MI, DL,
21779 TII->get(X86::FLDCW16m)), CWFrameIdx);
21781 // Restore the memory image of control word to original value
21782 addFrameReference(BuildMI(*BB, MI, DL, TII->get(X86::MOV16mr)), CWFrameIdx)
21785 // Get the X86 opcode to use.
21787 switch (MI->getOpcode()) {
21788 default: llvm_unreachable("illegal opcode!");
21789 case X86::FP32_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m32; break;
21790 case X86::FP32_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m32; break;
21791 case X86::FP32_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m32; break;
21792 case X86::FP64_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m64; break;
21793 case X86::FP64_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m64; break;
21794 case X86::FP64_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m64; break;
21795 case X86::FP80_TO_INT16_IN_MEM: Opc = X86::IST_Fp16m80; break;
21796 case X86::FP80_TO_INT32_IN_MEM: Opc = X86::IST_Fp32m80; break;
21797 case X86::FP80_TO_INT64_IN_MEM: Opc = X86::IST_Fp64m80; break;
21801 MachineOperand &Op = MI->getOperand(0);
21803 AM.BaseType = X86AddressMode::RegBase;
21804 AM.Base.Reg = Op.getReg();
21806 AM.BaseType = X86AddressMode::FrameIndexBase;
21807 AM.Base.FrameIndex = Op.getIndex();
21809 Op = MI->getOperand(1);
21811 AM.Scale = Op.getImm();
21812 Op = MI->getOperand(2);
21814 AM.IndexReg = Op.getImm();
21815 Op = MI->getOperand(3);
21816 if (Op.isGlobal()) {
21817 AM.GV = Op.getGlobal();
21819 AM.Disp = Op.getImm();
21821 addFullAddress(BuildMI(*BB, MI, DL, TII->get(Opc)), AM)
21822 .addReg(MI->getOperand(X86::AddrNumOperands).getReg());
21824 // Reload the original control word now.
21825 addFrameReference(BuildMI(*BB, MI, DL,
21826 TII->get(X86::FLDCW16m)), CWFrameIdx);
21828 MI->eraseFromParent(); // The pseudo instruction is gone now.
21831 // String/text processing lowering.
21832 case X86::PCMPISTRM128REG:
21833 case X86::VPCMPISTRM128REG:
21834 case X86::PCMPISTRM128MEM:
21835 case X86::VPCMPISTRM128MEM:
21836 case X86::PCMPESTRM128REG:
21837 case X86::VPCMPESTRM128REG:
21838 case X86::PCMPESTRM128MEM:
21839 case X86::VPCMPESTRM128MEM:
21840 assert(Subtarget->hasSSE42() &&
21841 "Target must have SSE4.2 or AVX features enabled");
21842 return EmitPCMPSTRM(MI, BB, Subtarget->getInstrInfo());
21844 // String/text processing lowering.
21845 case X86::PCMPISTRIREG:
21846 case X86::VPCMPISTRIREG:
21847 case X86::PCMPISTRIMEM:
21848 case X86::VPCMPISTRIMEM:
21849 case X86::PCMPESTRIREG:
21850 case X86::VPCMPESTRIREG:
21851 case X86::PCMPESTRIMEM:
21852 case X86::VPCMPESTRIMEM:
21853 assert(Subtarget->hasSSE42() &&
21854 "Target must have SSE4.2 or AVX features enabled");
21855 return EmitPCMPSTRI(MI, BB, Subtarget->getInstrInfo());
21857 // Thread synchronization.
21859 return EmitMonitor(MI, BB, Subtarget);
21863 return EmitXBegin(MI, BB, Subtarget->getInstrInfo());
21865 case X86::VASTART_SAVE_XMM_REGS:
21866 return EmitVAStartSaveXMMRegsWithCustomInserter(MI, BB);
21868 case X86::VAARG_64:
21869 return EmitVAARG64WithCustomInserter(MI, BB);
21871 case X86::EH_SjLj_SetJmp32:
21872 case X86::EH_SjLj_SetJmp64:
21873 return emitEHSjLjSetJmp(MI, BB);
21875 case X86::EH_SjLj_LongJmp32:
21876 case X86::EH_SjLj_LongJmp64:
21877 return emitEHSjLjLongJmp(MI, BB);
21879 case TargetOpcode::STATEPOINT:
21880 // As an implementation detail, STATEPOINT shares the STACKMAP format at
21881 // this point in the process. We diverge later.
21882 return emitPatchPoint(MI, BB);
21884 case TargetOpcode::STACKMAP:
21885 case TargetOpcode::PATCHPOINT:
21886 return emitPatchPoint(MI, BB);
21888 case X86::VFMADDPDr213r:
21889 case X86::VFMADDPSr213r:
21890 case X86::VFMADDSDr213r:
21891 case X86::VFMADDSSr213r:
21892 case X86::VFMSUBPDr213r:
21893 case X86::VFMSUBPSr213r:
21894 case X86::VFMSUBSDr213r:
21895 case X86::VFMSUBSSr213r:
21896 case X86::VFNMADDPDr213r:
21897 case X86::VFNMADDPSr213r:
21898 case X86::VFNMADDSDr213r:
21899 case X86::VFNMADDSSr213r:
21900 case X86::VFNMSUBPDr213r:
21901 case X86::VFNMSUBPSr213r:
21902 case X86::VFNMSUBSDr213r:
21903 case X86::VFNMSUBSSr213r:
21904 case X86::VFMADDSUBPDr213r:
21905 case X86::VFMADDSUBPSr213r:
21906 case X86::VFMSUBADDPDr213r:
21907 case X86::VFMSUBADDPSr213r:
21908 case X86::VFMADDPDr213rY:
21909 case X86::VFMADDPSr213rY:
21910 case X86::VFMSUBPDr213rY:
21911 case X86::VFMSUBPSr213rY:
21912 case X86::VFNMADDPDr213rY:
21913 case X86::VFNMADDPSr213rY:
21914 case X86::VFNMSUBPDr213rY:
21915 case X86::VFNMSUBPSr213rY:
21916 case X86::VFMADDSUBPDr213rY:
21917 case X86::VFMADDSUBPSr213rY:
21918 case X86::VFMSUBADDPDr213rY:
21919 case X86::VFMSUBADDPSr213rY:
21920 return emitFMA3Instr(MI, BB);
21924 //===----------------------------------------------------------------------===//
21925 // X86 Optimization Hooks
21926 //===----------------------------------------------------------------------===//
21928 void X86TargetLowering::computeKnownBitsForTargetNode(const SDValue Op,
21931 const SelectionDAG &DAG,
21932 unsigned Depth) const {
21933 unsigned BitWidth = KnownZero.getBitWidth();
21934 unsigned Opc = Op.getOpcode();
21935 assert((Opc >= ISD::BUILTIN_OP_END ||
21936 Opc == ISD::INTRINSIC_WO_CHAIN ||
21937 Opc == ISD::INTRINSIC_W_CHAIN ||
21938 Opc == ISD::INTRINSIC_VOID) &&
21939 "Should use MaskedValueIsZero if you don't know whether Op"
21940 " is a target node!");
21942 KnownZero = KnownOne = APInt(BitWidth, 0); // Don't know anything.
21956 // These nodes' second result is a boolean.
21957 if (Op.getResNo() == 0)
21960 case X86ISD::SETCC:
21961 KnownZero |= APInt::getHighBitsSet(BitWidth, BitWidth - 1);
21963 case ISD::INTRINSIC_WO_CHAIN: {
21964 unsigned IntId = cast<ConstantSDNode>(Op.getOperand(0))->getZExtValue();
21965 unsigned NumLoBits = 0;
21968 case Intrinsic::x86_sse_movmsk_ps:
21969 case Intrinsic::x86_avx_movmsk_ps_256:
21970 case Intrinsic::x86_sse2_movmsk_pd:
21971 case Intrinsic::x86_avx_movmsk_pd_256:
21972 case Intrinsic::x86_mmx_pmovmskb:
21973 case Intrinsic::x86_sse2_pmovmskb_128:
21974 case Intrinsic::x86_avx2_pmovmskb: {
21975 // High bits of movmskp{s|d}, pmovmskb are known zero.
21977 default: llvm_unreachable("Impossible intrinsic"); // Can't reach here.
21978 case Intrinsic::x86_sse_movmsk_ps: NumLoBits = 4; break;
21979 case Intrinsic::x86_avx_movmsk_ps_256: NumLoBits = 8; break;
21980 case Intrinsic::x86_sse2_movmsk_pd: NumLoBits = 2; break;
21981 case Intrinsic::x86_avx_movmsk_pd_256: NumLoBits = 4; break;
21982 case Intrinsic::x86_mmx_pmovmskb: NumLoBits = 8; break;
21983 case Intrinsic::x86_sse2_pmovmskb_128: NumLoBits = 16; break;
21984 case Intrinsic::x86_avx2_pmovmskb: NumLoBits = 32; break;
21986 KnownZero = APInt::getHighBitsSet(BitWidth, BitWidth - NumLoBits);
21995 unsigned X86TargetLowering::ComputeNumSignBitsForTargetNode(
21997 const SelectionDAG &,
21998 unsigned Depth) const {
21999 // SETCC_CARRY sets the dest to ~0 for true or 0 for false.
22000 if (Op.getOpcode() == X86ISD::SETCC_CARRY)
22001 return Op.getValueType().getScalarType().getSizeInBits();
22007 /// isGAPlusOffset - Returns true (and the GlobalValue and the offset) if the
22008 /// node is a GlobalAddress + offset.
22009 bool X86TargetLowering::isGAPlusOffset(SDNode *N,
22010 const GlobalValue* &GA,
22011 int64_t &Offset) const {
22012 if (N->getOpcode() == X86ISD::Wrapper) {
22013 if (isa<GlobalAddressSDNode>(N->getOperand(0))) {
22014 GA = cast<GlobalAddressSDNode>(N->getOperand(0))->getGlobal();
22015 Offset = cast<GlobalAddressSDNode>(N->getOperand(0))->getOffset();
22019 return TargetLowering::isGAPlusOffset(N, GA, Offset);
22022 /// isShuffleHigh128VectorInsertLow - Checks whether the shuffle node is the
22023 /// same as extracting the high 128-bit part of 256-bit vector and then
22024 /// inserting the result into the low part of a new 256-bit vector
22025 static bool isShuffleHigh128VectorInsertLow(ShuffleVectorSDNode *SVOp) {
22026 EVT VT = SVOp->getValueType(0);
22027 unsigned NumElems = VT.getVectorNumElements();
22029 // vector_shuffle <4, 5, 6, 7, u, u, u, u> or <2, 3, u, u>
22030 for (unsigned i = 0, j = NumElems/2; i != NumElems/2; ++i, ++j)
22031 if (!isUndefOrEqual(SVOp->getMaskElt(i), j) ||
22032 SVOp->getMaskElt(j) >= 0)
22038 /// isShuffleLow128VectorInsertHigh - Checks whether the shuffle node is the
22039 /// same as extracting the low 128-bit part of 256-bit vector and then
22040 /// inserting the result into the high part of a new 256-bit vector
22041 static bool isShuffleLow128VectorInsertHigh(ShuffleVectorSDNode *SVOp) {
22042 EVT VT = SVOp->getValueType(0);
22043 unsigned NumElems = VT.getVectorNumElements();
22045 // vector_shuffle <u, u, u, u, 0, 1, 2, 3> or <u, u, 0, 1>
22046 for (unsigned i = NumElems/2, j = 0; i != NumElems; ++i, ++j)
22047 if (!isUndefOrEqual(SVOp->getMaskElt(i), j) ||
22048 SVOp->getMaskElt(j) >= 0)
22054 /// PerformShuffleCombine256 - Performs shuffle combines for 256-bit vectors.
22055 static SDValue PerformShuffleCombine256(SDNode *N, SelectionDAG &DAG,
22056 TargetLowering::DAGCombinerInfo &DCI,
22057 const X86Subtarget* Subtarget) {
22059 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
22060 SDValue V1 = SVOp->getOperand(0);
22061 SDValue V2 = SVOp->getOperand(1);
22062 EVT VT = SVOp->getValueType(0);
22063 unsigned NumElems = VT.getVectorNumElements();
22065 if (V1.getOpcode() == ISD::CONCAT_VECTORS &&
22066 V2.getOpcode() == ISD::CONCAT_VECTORS) {
22070 // V UNDEF BUILD_VECTOR UNDEF
22072 // CONCAT_VECTOR CONCAT_VECTOR
22075 // RESULT: V + zero extended
22077 if (V2.getOperand(0).getOpcode() != ISD::BUILD_VECTOR ||
22078 V2.getOperand(1).getOpcode() != ISD::UNDEF ||
22079 V1.getOperand(1).getOpcode() != ISD::UNDEF)
22082 if (!ISD::isBuildVectorAllZeros(V2.getOperand(0).getNode()))
22085 // To match the shuffle mask, the first half of the mask should
22086 // be exactly the first vector, and all the rest a splat with the
22087 // first element of the second one.
22088 for (unsigned i = 0; i != NumElems/2; ++i)
22089 if (!isUndefOrEqual(SVOp->getMaskElt(i), i) ||
22090 !isUndefOrEqual(SVOp->getMaskElt(i+NumElems/2), NumElems))
22093 // If V1 is coming from a vector load then just fold to a VZEXT_LOAD.
22094 if (LoadSDNode *Ld = dyn_cast<LoadSDNode>(V1.getOperand(0))) {
22095 if (Ld->hasNUsesOfValue(1, 0)) {
22096 SDVTList Tys = DAG.getVTList(MVT::v4i64, MVT::Other);
22097 SDValue Ops[] = { Ld->getChain(), Ld->getBasePtr() };
22099 DAG.getMemIntrinsicNode(X86ISD::VZEXT_LOAD, dl, Tys, Ops,
22101 Ld->getPointerInfo(),
22102 Ld->getAlignment(),
22103 false/*isVolatile*/, true/*ReadMem*/,
22104 false/*WriteMem*/);
22106 // Make sure the newly-created LOAD is in the same position as Ld in
22107 // terms of dependency. We create a TokenFactor for Ld and ResNode,
22108 // and update uses of Ld's output chain to use the TokenFactor.
22109 if (Ld->hasAnyUseOfValue(1)) {
22110 SDValue NewChain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
22111 SDValue(Ld, 1), SDValue(ResNode.getNode(), 1));
22112 DAG.ReplaceAllUsesOfValueWith(SDValue(Ld, 1), NewChain);
22113 DAG.UpdateNodeOperands(NewChain.getNode(), SDValue(Ld, 1),
22114 SDValue(ResNode.getNode(), 1));
22117 return DAG.getBitcast(VT, ResNode);
22121 // Emit a zeroed vector and insert the desired subvector on its
22123 SDValue Zeros = getZeroVector(VT, Subtarget, DAG, dl);
22124 SDValue InsV = Insert128BitVector(Zeros, V1.getOperand(0), 0, DAG, dl);
22125 return DCI.CombineTo(N, InsV);
22128 //===--------------------------------------------------------------------===//
22129 // Combine some shuffles into subvector extracts and inserts:
22132 // vector_shuffle <4, 5, 6, 7, u, u, u, u> or <2, 3, u, u>
22133 if (isShuffleHigh128VectorInsertLow(SVOp)) {
22134 SDValue V = Extract128BitVector(V1, NumElems/2, DAG, dl);
22135 SDValue InsV = Insert128BitVector(DAG.getUNDEF(VT), V, 0, DAG, dl);
22136 return DCI.CombineTo(N, InsV);
22139 // vector_shuffle <u, u, u, u, 0, 1, 2, 3> or <u, u, 0, 1>
22140 if (isShuffleLow128VectorInsertHigh(SVOp)) {
22141 SDValue V = Extract128BitVector(V1, 0, DAG, dl);
22142 SDValue InsV = Insert128BitVector(DAG.getUNDEF(VT), V, NumElems/2, DAG, dl);
22143 return DCI.CombineTo(N, InsV);
22149 /// \brief Combine an arbitrary chain of shuffles into a single instruction if
22152 /// This is the leaf of the recursive combinine below. When we have found some
22153 /// chain of single-use x86 shuffle instructions and accumulated the combined
22154 /// shuffle mask represented by them, this will try to pattern match that mask
22155 /// into either a single instruction if there is a special purpose instruction
22156 /// for this operation, or into a PSHUFB instruction which is a fully general
22157 /// instruction but should only be used to replace chains over a certain depth.
22158 static bool combineX86ShuffleChain(SDValue Op, SDValue Root, ArrayRef<int> Mask,
22159 int Depth, bool HasPSHUFB, SelectionDAG &DAG,
22160 TargetLowering::DAGCombinerInfo &DCI,
22161 const X86Subtarget *Subtarget) {
22162 assert(!Mask.empty() && "Cannot combine an empty shuffle mask!");
22164 // Find the operand that enters the chain. Note that multiple uses are OK
22165 // here, we're not going to remove the operand we find.
22166 SDValue Input = Op.getOperand(0);
22167 while (Input.getOpcode() == ISD::BITCAST)
22168 Input = Input.getOperand(0);
22170 MVT VT = Input.getSimpleValueType();
22171 MVT RootVT = Root.getSimpleValueType();
22174 if (Mask.size() == 1) {
22175 int Index = Mask[0];
22176 assert((Index >= 0 || Index == SM_SentinelUndef ||
22177 Index == SM_SentinelZero) &&
22178 "Invalid shuffle index found!");
22180 // We may end up with an accumulated mask of size 1 as a result of
22181 // widening of shuffle operands (see function canWidenShuffleElements).
22182 // If the only shuffle index is equal to SM_SentinelZero then propagate
22183 // a zero vector. Otherwise, the combine shuffle mask is a no-op shuffle
22184 // mask, and therefore the entire chain of shuffles can be folded away.
22185 if (Index == SM_SentinelZero)
22186 DCI.CombineTo(Root.getNode(), getZeroVector(RootVT, Subtarget, DAG, DL));
22188 DCI.CombineTo(Root.getNode(), DAG.getBitcast(RootVT, Input),
22193 // Use the float domain if the operand type is a floating point type.
22194 bool FloatDomain = VT.isFloatingPoint();
22196 // For floating point shuffles, we don't have free copies in the shuffle
22197 // instructions or the ability to load as part of the instruction, so
22198 // canonicalize their shuffles to UNPCK or MOV variants.
22200 // Note that even with AVX we prefer the PSHUFD form of shuffle for integer
22201 // vectors because it can have a load folded into it that UNPCK cannot. This
22202 // doesn't preclude something switching to the shorter encoding post-RA.
22204 // FIXME: Should teach these routines about AVX vector widths.
22205 if (FloatDomain && VT.getSizeInBits() == 128) {
22206 if (Mask.equals({0, 0}) || Mask.equals({1, 1})) {
22207 bool Lo = Mask.equals({0, 0});
22210 // Check if we have SSE3 which will let us use MOVDDUP. That instruction
22211 // is no slower than UNPCKLPD but has the option to fold the input operand
22212 // into even an unaligned memory load.
22213 if (Lo && Subtarget->hasSSE3()) {
22214 Shuffle = X86ISD::MOVDDUP;
22215 ShuffleVT = MVT::v2f64;
22217 // We have MOVLHPS and MOVHLPS throughout SSE and they encode smaller
22218 // than the UNPCK variants.
22219 Shuffle = Lo ? X86ISD::MOVLHPS : X86ISD::MOVHLPS;
22220 ShuffleVT = MVT::v4f32;
22222 if (Depth == 1 && Root->getOpcode() == Shuffle)
22223 return false; // Nothing to do!
22224 Op = DAG.getBitcast(ShuffleVT, Input);
22225 DCI.AddToWorklist(Op.getNode());
22226 if (Shuffle == X86ISD::MOVDDUP)
22227 Op = DAG.getNode(Shuffle, DL, ShuffleVT, Op);
22229 Op = DAG.getNode(Shuffle, DL, ShuffleVT, Op, Op);
22230 DCI.AddToWorklist(Op.getNode());
22231 DCI.CombineTo(Root.getNode(), DAG.getBitcast(RootVT, Op),
22235 if (Subtarget->hasSSE3() &&
22236 (Mask.equals({0, 0, 2, 2}) || Mask.equals({1, 1, 3, 3}))) {
22237 bool Lo = Mask.equals({0, 0, 2, 2});
22238 unsigned Shuffle = Lo ? X86ISD::MOVSLDUP : X86ISD::MOVSHDUP;
22239 MVT ShuffleVT = MVT::v4f32;
22240 if (Depth == 1 && Root->getOpcode() == Shuffle)
22241 return false; // Nothing to do!
22242 Op = DAG.getBitcast(ShuffleVT, Input);
22243 DCI.AddToWorklist(Op.getNode());
22244 Op = DAG.getNode(Shuffle, DL, ShuffleVT, Op);
22245 DCI.AddToWorklist(Op.getNode());
22246 DCI.CombineTo(Root.getNode(), DAG.getBitcast(RootVT, Op),
22250 if (Mask.equals({0, 0, 1, 1}) || Mask.equals({2, 2, 3, 3})) {
22251 bool Lo = Mask.equals({0, 0, 1, 1});
22252 unsigned Shuffle = Lo ? X86ISD::UNPCKL : X86ISD::UNPCKH;
22253 MVT ShuffleVT = MVT::v4f32;
22254 if (Depth == 1 && Root->getOpcode() == Shuffle)
22255 return false; // Nothing to do!
22256 Op = DAG.getBitcast(ShuffleVT, Input);
22257 DCI.AddToWorklist(Op.getNode());
22258 Op = DAG.getNode(Shuffle, DL, ShuffleVT, Op, Op);
22259 DCI.AddToWorklist(Op.getNode());
22260 DCI.CombineTo(Root.getNode(), DAG.getBitcast(RootVT, Op),
22266 // We always canonicalize the 8 x i16 and 16 x i8 shuffles into their UNPCK
22267 // variants as none of these have single-instruction variants that are
22268 // superior to the UNPCK formulation.
22269 if (!FloatDomain && VT.getSizeInBits() == 128 &&
22270 (Mask.equals({0, 0, 1, 1, 2, 2, 3, 3}) ||
22271 Mask.equals({4, 4, 5, 5, 6, 6, 7, 7}) ||
22272 Mask.equals({0, 0, 1, 1, 2, 2, 3, 3, 4, 4, 5, 5, 6, 6, 7, 7}) ||
22274 {8, 8, 9, 9, 10, 10, 11, 11, 12, 12, 13, 13, 14, 14, 15, 15}))) {
22275 bool Lo = Mask[0] == 0;
22276 unsigned Shuffle = Lo ? X86ISD::UNPCKL : X86ISD::UNPCKH;
22277 if (Depth == 1 && Root->getOpcode() == Shuffle)
22278 return false; // Nothing to do!
22280 switch (Mask.size()) {
22282 ShuffleVT = MVT::v8i16;
22285 ShuffleVT = MVT::v16i8;
22288 llvm_unreachable("Impossible mask size!");
22290 Op = DAG.getBitcast(ShuffleVT, Input);
22291 DCI.AddToWorklist(Op.getNode());
22292 Op = DAG.getNode(Shuffle, DL, ShuffleVT, Op, Op);
22293 DCI.AddToWorklist(Op.getNode());
22294 DCI.CombineTo(Root.getNode(), DAG.getBitcast(RootVT, Op),
22299 // Don't try to re-form single instruction chains under any circumstances now
22300 // that we've done encoding canonicalization for them.
22304 // If we have 3 or more shuffle instructions or a chain involving PSHUFB, we
22305 // can replace them with a single PSHUFB instruction profitably. Intel's
22306 // manuals suggest only using PSHUFB if doing so replacing 5 instructions, but
22307 // in practice PSHUFB tends to be *very* fast so we're more aggressive.
22308 if ((Depth >= 3 || HasPSHUFB) && Subtarget->hasSSSE3()) {
22309 SmallVector<SDValue, 16> PSHUFBMask;
22310 int NumBytes = VT.getSizeInBits() / 8;
22311 int Ratio = NumBytes / Mask.size();
22312 for (int i = 0; i < NumBytes; ++i) {
22313 if (Mask[i / Ratio] == SM_SentinelUndef) {
22314 PSHUFBMask.push_back(DAG.getUNDEF(MVT::i8));
22317 int M = Mask[i / Ratio] != SM_SentinelZero
22318 ? Ratio * Mask[i / Ratio] + i % Ratio
22320 PSHUFBMask.push_back(DAG.getConstant(M, DL, MVT::i8));
22322 MVT ByteVT = MVT::getVectorVT(MVT::i8, NumBytes);
22323 Op = DAG.getBitcast(ByteVT, Input);
22324 DCI.AddToWorklist(Op.getNode());
22325 SDValue PSHUFBMaskOp =
22326 DAG.getNode(ISD::BUILD_VECTOR, DL, ByteVT, PSHUFBMask);
22327 DCI.AddToWorklist(PSHUFBMaskOp.getNode());
22328 Op = DAG.getNode(X86ISD::PSHUFB, DL, ByteVT, Op, PSHUFBMaskOp);
22329 DCI.AddToWorklist(Op.getNode());
22330 DCI.CombineTo(Root.getNode(), DAG.getBitcast(RootVT, Op),
22335 // Failed to find any combines.
22339 /// \brief Fully generic combining of x86 shuffle instructions.
22341 /// This should be the last combine run over the x86 shuffle instructions. Once
22342 /// they have been fully optimized, this will recursively consider all chains
22343 /// of single-use shuffle instructions, build a generic model of the cumulative
22344 /// shuffle operation, and check for simpler instructions which implement this
22345 /// operation. We use this primarily for two purposes:
22347 /// 1) Collapse generic shuffles to specialized single instructions when
22348 /// equivalent. In most cases, this is just an encoding size win, but
22349 /// sometimes we will collapse multiple generic shuffles into a single
22350 /// special-purpose shuffle.
22351 /// 2) Look for sequences of shuffle instructions with 3 or more total
22352 /// instructions, and replace them with the slightly more expensive SSSE3
22353 /// PSHUFB instruction if available. We do this as the last combining step
22354 /// to ensure we avoid using PSHUFB if we can implement the shuffle with
22355 /// a suitable short sequence of other instructions. The PHUFB will either
22356 /// use a register or have to read from memory and so is slightly (but only
22357 /// slightly) more expensive than the other shuffle instructions.
22359 /// Because this is inherently a quadratic operation (for each shuffle in
22360 /// a chain, we recurse up the chain), the depth is limited to 8 instructions.
22361 /// This should never be an issue in practice as the shuffle lowering doesn't
22362 /// produce sequences of more than 8 instructions.
22364 /// FIXME: We will currently miss some cases where the redundant shuffling
22365 /// would simplify under the threshold for PSHUFB formation because of
22366 /// combine-ordering. To fix this, we should do the redundant instruction
22367 /// combining in this recursive walk.
22368 static bool combineX86ShufflesRecursively(SDValue Op, SDValue Root,
22369 ArrayRef<int> RootMask,
22370 int Depth, bool HasPSHUFB,
22372 TargetLowering::DAGCombinerInfo &DCI,
22373 const X86Subtarget *Subtarget) {
22374 // Bound the depth of our recursive combine because this is ultimately
22375 // quadratic in nature.
22379 // Directly rip through bitcasts to find the underlying operand.
22380 while (Op.getOpcode() == ISD::BITCAST && Op.getOperand(0).hasOneUse())
22381 Op = Op.getOperand(0);
22383 MVT VT = Op.getSimpleValueType();
22384 if (!VT.isVector())
22385 return false; // Bail if we hit a non-vector.
22387 assert(Root.getSimpleValueType().isVector() &&
22388 "Shuffles operate on vector types!");
22389 assert(VT.getSizeInBits() == Root.getSimpleValueType().getSizeInBits() &&
22390 "Can only combine shuffles of the same vector register size.");
22392 if (!isTargetShuffle(Op.getOpcode()))
22394 SmallVector<int, 16> OpMask;
22396 bool HaveMask = getTargetShuffleMask(Op.getNode(), VT, OpMask, IsUnary);
22397 // We only can combine unary shuffles which we can decode the mask for.
22398 if (!HaveMask || !IsUnary)
22401 assert(VT.getVectorNumElements() == OpMask.size() &&
22402 "Different mask size from vector size!");
22403 assert(((RootMask.size() > OpMask.size() &&
22404 RootMask.size() % OpMask.size() == 0) ||
22405 (OpMask.size() > RootMask.size() &&
22406 OpMask.size() % RootMask.size() == 0) ||
22407 OpMask.size() == RootMask.size()) &&
22408 "The smaller number of elements must divide the larger.");
22409 int RootRatio = std::max<int>(1, OpMask.size() / RootMask.size());
22410 int OpRatio = std::max<int>(1, RootMask.size() / OpMask.size());
22411 assert(((RootRatio == 1 && OpRatio == 1) ||
22412 (RootRatio == 1) != (OpRatio == 1)) &&
22413 "Must not have a ratio for both incoming and op masks!");
22415 SmallVector<int, 16> Mask;
22416 Mask.reserve(std::max(OpMask.size(), RootMask.size()));
22418 // Merge this shuffle operation's mask into our accumulated mask. Note that
22419 // this shuffle's mask will be the first applied to the input, followed by the
22420 // root mask to get us all the way to the root value arrangement. The reason
22421 // for this order is that we are recursing up the operation chain.
22422 for (int i = 0, e = std::max(OpMask.size(), RootMask.size()); i < e; ++i) {
22423 int RootIdx = i / RootRatio;
22424 if (RootMask[RootIdx] < 0) {
22425 // This is a zero or undef lane, we're done.
22426 Mask.push_back(RootMask[RootIdx]);
22430 int RootMaskedIdx = RootMask[RootIdx] * RootRatio + i % RootRatio;
22431 int OpIdx = RootMaskedIdx / OpRatio;
22432 if (OpMask[OpIdx] < 0) {
22433 // The incoming lanes are zero or undef, it doesn't matter which ones we
22435 Mask.push_back(OpMask[OpIdx]);
22439 // Ok, we have non-zero lanes, map them through.
22440 Mask.push_back(OpMask[OpIdx] * OpRatio +
22441 RootMaskedIdx % OpRatio);
22444 // See if we can recurse into the operand to combine more things.
22445 switch (Op.getOpcode()) {
22446 case X86ISD::PSHUFB:
22448 case X86ISD::PSHUFD:
22449 case X86ISD::PSHUFHW:
22450 case X86ISD::PSHUFLW:
22451 if (Op.getOperand(0).hasOneUse() &&
22452 combineX86ShufflesRecursively(Op.getOperand(0), Root, Mask, Depth + 1,
22453 HasPSHUFB, DAG, DCI, Subtarget))
22457 case X86ISD::UNPCKL:
22458 case X86ISD::UNPCKH:
22459 assert(Op.getOperand(0) == Op.getOperand(1) &&
22460 "We only combine unary shuffles!");
22461 // We can't check for single use, we have to check that this shuffle is the
22463 if (Op->isOnlyUserOf(Op.getOperand(0).getNode()) &&
22464 combineX86ShufflesRecursively(Op.getOperand(0), Root, Mask, Depth + 1,
22465 HasPSHUFB, DAG, DCI, Subtarget))
22470 // Minor canonicalization of the accumulated shuffle mask to make it easier
22471 // to match below. All this does is detect masks with squential pairs of
22472 // elements, and shrink them to the half-width mask. It does this in a loop
22473 // so it will reduce the size of the mask to the minimal width mask which
22474 // performs an equivalent shuffle.
22475 SmallVector<int, 16> WidenedMask;
22476 while (Mask.size() > 1 && canWidenShuffleElements(Mask, WidenedMask)) {
22477 Mask = std::move(WidenedMask);
22478 WidenedMask.clear();
22481 return combineX86ShuffleChain(Op, Root, Mask, Depth, HasPSHUFB, DAG, DCI,
22485 /// \brief Get the PSHUF-style mask from PSHUF node.
22487 /// This is a very minor wrapper around getTargetShuffleMask to easy forming v4
22488 /// PSHUF-style masks that can be reused with such instructions.
22489 static SmallVector<int, 4> getPSHUFShuffleMask(SDValue N) {
22490 MVT VT = N.getSimpleValueType();
22491 SmallVector<int, 4> Mask;
22493 bool HaveMask = getTargetShuffleMask(N.getNode(), VT, Mask, IsUnary);
22497 // If we have more than 128-bits, only the low 128-bits of shuffle mask
22498 // matter. Check that the upper masks are repeats and remove them.
22499 if (VT.getSizeInBits() > 128) {
22500 int LaneElts = 128 / VT.getScalarSizeInBits();
22502 for (int i = 1, NumLanes = VT.getSizeInBits() / 128; i < NumLanes; ++i)
22503 for (int j = 0; j < LaneElts; ++j)
22504 assert(Mask[j] == Mask[i * LaneElts + j] - (LaneElts * i) &&
22505 "Mask doesn't repeat in high 128-bit lanes!");
22507 Mask.resize(LaneElts);
22510 switch (N.getOpcode()) {
22511 case X86ISD::PSHUFD:
22513 case X86ISD::PSHUFLW:
22516 case X86ISD::PSHUFHW:
22517 Mask.erase(Mask.begin(), Mask.begin() + 4);
22518 for (int &M : Mask)
22522 llvm_unreachable("No valid shuffle instruction found!");
22526 /// \brief Search for a combinable shuffle across a chain ending in pshufd.
22528 /// We walk up the chain and look for a combinable shuffle, skipping over
22529 /// shuffles that we could hoist this shuffle's transformation past without
22530 /// altering anything.
22532 combineRedundantDWordShuffle(SDValue N, MutableArrayRef<int> Mask,
22534 TargetLowering::DAGCombinerInfo &DCI) {
22535 assert(N.getOpcode() == X86ISD::PSHUFD &&
22536 "Called with something other than an x86 128-bit half shuffle!");
22539 // Walk up a single-use chain looking for a combinable shuffle. Keep a stack
22540 // of the shuffles in the chain so that we can form a fresh chain to replace
22542 SmallVector<SDValue, 8> Chain;
22543 SDValue V = N.getOperand(0);
22544 for (; V.hasOneUse(); V = V.getOperand(0)) {
22545 switch (V.getOpcode()) {
22547 return SDValue(); // Nothing combined!
22550 // Skip bitcasts as we always know the type for the target specific
22554 case X86ISD::PSHUFD:
22555 // Found another dword shuffle.
22558 case X86ISD::PSHUFLW:
22559 // Check that the low words (being shuffled) are the identity in the
22560 // dword shuffle, and the high words are self-contained.
22561 if (Mask[0] != 0 || Mask[1] != 1 ||
22562 !(Mask[2] >= 2 && Mask[2] < 4 && Mask[3] >= 2 && Mask[3] < 4))
22565 Chain.push_back(V);
22568 case X86ISD::PSHUFHW:
22569 // Check that the high words (being shuffled) are the identity in the
22570 // dword shuffle, and the low words are self-contained.
22571 if (Mask[2] != 2 || Mask[3] != 3 ||
22572 !(Mask[0] >= 0 && Mask[0] < 2 && Mask[1] >= 0 && Mask[1] < 2))
22575 Chain.push_back(V);
22578 case X86ISD::UNPCKL:
22579 case X86ISD::UNPCKH:
22580 // For either i8 -> i16 or i16 -> i32 unpacks, we can combine a dword
22581 // shuffle into a preceding word shuffle.
22582 if (V.getSimpleValueType().getScalarType() != MVT::i8 &&
22583 V.getSimpleValueType().getScalarType() != MVT::i16)
22586 // Search for a half-shuffle which we can combine with.
22587 unsigned CombineOp =
22588 V.getOpcode() == X86ISD::UNPCKL ? X86ISD::PSHUFLW : X86ISD::PSHUFHW;
22589 if (V.getOperand(0) != V.getOperand(1) ||
22590 !V->isOnlyUserOf(V.getOperand(0).getNode()))
22592 Chain.push_back(V);
22593 V = V.getOperand(0);
22595 switch (V.getOpcode()) {
22597 return SDValue(); // Nothing to combine.
22599 case X86ISD::PSHUFLW:
22600 case X86ISD::PSHUFHW:
22601 if (V.getOpcode() == CombineOp)
22604 Chain.push_back(V);
22608 V = V.getOperand(0);
22612 } while (V.hasOneUse());
22615 // Break out of the loop if we break out of the switch.
22619 if (!V.hasOneUse())
22620 // We fell out of the loop without finding a viable combining instruction.
22623 // Merge this node's mask and our incoming mask.
22624 SmallVector<int, 4> VMask = getPSHUFShuffleMask(V);
22625 for (int &M : Mask)
22627 V = DAG.getNode(V.getOpcode(), DL, V.getValueType(), V.getOperand(0),
22628 getV4X86ShuffleImm8ForMask(Mask, DL, DAG));
22630 // Rebuild the chain around this new shuffle.
22631 while (!Chain.empty()) {
22632 SDValue W = Chain.pop_back_val();
22634 if (V.getValueType() != W.getOperand(0).getValueType())
22635 V = DAG.getBitcast(W.getOperand(0).getValueType(), V);
22637 switch (W.getOpcode()) {
22639 llvm_unreachable("Only PSHUF and UNPCK instructions get here!");
22641 case X86ISD::UNPCKL:
22642 case X86ISD::UNPCKH:
22643 V = DAG.getNode(W.getOpcode(), DL, W.getValueType(), V, V);
22646 case X86ISD::PSHUFD:
22647 case X86ISD::PSHUFLW:
22648 case X86ISD::PSHUFHW:
22649 V = DAG.getNode(W.getOpcode(), DL, W.getValueType(), V, W.getOperand(1));
22653 if (V.getValueType() != N.getValueType())
22654 V = DAG.getBitcast(N.getValueType(), V);
22656 // Return the new chain to replace N.
22660 /// \brief Search for a combinable shuffle across a chain ending in pshuflw or
22663 /// We walk up the chain, skipping shuffles of the other half and looking
22664 /// through shuffles which switch halves trying to find a shuffle of the same
22665 /// pair of dwords.
22666 static bool combineRedundantHalfShuffle(SDValue N, MutableArrayRef<int> Mask,
22668 TargetLowering::DAGCombinerInfo &DCI) {
22670 (N.getOpcode() == X86ISD::PSHUFLW || N.getOpcode() == X86ISD::PSHUFHW) &&
22671 "Called with something other than an x86 128-bit half shuffle!");
22673 unsigned CombineOpcode = N.getOpcode();
22675 // Walk up a single-use chain looking for a combinable shuffle.
22676 SDValue V = N.getOperand(0);
22677 for (; V.hasOneUse(); V = V.getOperand(0)) {
22678 switch (V.getOpcode()) {
22680 return false; // Nothing combined!
22683 // Skip bitcasts as we always know the type for the target specific
22687 case X86ISD::PSHUFLW:
22688 case X86ISD::PSHUFHW:
22689 if (V.getOpcode() == CombineOpcode)
22692 // Other-half shuffles are no-ops.
22695 // Break out of the loop if we break out of the switch.
22699 if (!V.hasOneUse())
22700 // We fell out of the loop without finding a viable combining instruction.
22703 // Combine away the bottom node as its shuffle will be accumulated into
22704 // a preceding shuffle.
22705 DCI.CombineTo(N.getNode(), N.getOperand(0), /*AddTo*/ true);
22707 // Record the old value.
22710 // Merge this node's mask and our incoming mask (adjusted to account for all
22711 // the pshufd instructions encountered).
22712 SmallVector<int, 4> VMask = getPSHUFShuffleMask(V);
22713 for (int &M : Mask)
22715 V = DAG.getNode(V.getOpcode(), DL, MVT::v8i16, V.getOperand(0),
22716 getV4X86ShuffleImm8ForMask(Mask, DL, DAG));
22718 // Check that the shuffles didn't cancel each other out. If not, we need to
22719 // combine to the new one.
22721 // Replace the combinable shuffle with the combined one, updating all users
22722 // so that we re-evaluate the chain here.
22723 DCI.CombineTo(Old.getNode(), V, /*AddTo*/ true);
22728 /// \brief Try to combine x86 target specific shuffles.
22729 static SDValue PerformTargetShuffleCombine(SDValue N, SelectionDAG &DAG,
22730 TargetLowering::DAGCombinerInfo &DCI,
22731 const X86Subtarget *Subtarget) {
22733 MVT VT = N.getSimpleValueType();
22734 SmallVector<int, 4> Mask;
22736 switch (N.getOpcode()) {
22737 case X86ISD::PSHUFD:
22738 case X86ISD::PSHUFLW:
22739 case X86ISD::PSHUFHW:
22740 Mask = getPSHUFShuffleMask(N);
22741 assert(Mask.size() == 4);
22747 // Nuke no-op shuffles that show up after combining.
22748 if (isNoopShuffleMask(Mask))
22749 return DCI.CombineTo(N.getNode(), N.getOperand(0), /*AddTo*/ true);
22751 // Look for simplifications involving one or two shuffle instructions.
22752 SDValue V = N.getOperand(0);
22753 switch (N.getOpcode()) {
22756 case X86ISD::PSHUFLW:
22757 case X86ISD::PSHUFHW:
22758 assert(VT.getScalarType() == MVT::i16 && "Bad word shuffle type!");
22760 if (combineRedundantHalfShuffle(N, Mask, DAG, DCI))
22761 return SDValue(); // We combined away this shuffle, so we're done.
22763 // See if this reduces to a PSHUFD which is no more expensive and can
22764 // combine with more operations. Note that it has to at least flip the
22765 // dwords as otherwise it would have been removed as a no-op.
22766 if (makeArrayRef(Mask).equals({2, 3, 0, 1})) {
22767 int DMask[] = {0, 1, 2, 3};
22768 int DOffset = N.getOpcode() == X86ISD::PSHUFLW ? 0 : 2;
22769 DMask[DOffset + 0] = DOffset + 1;
22770 DMask[DOffset + 1] = DOffset + 0;
22771 MVT DVT = MVT::getVectorVT(MVT::i32, VT.getVectorNumElements() / 2);
22772 V = DAG.getBitcast(DVT, V);
22773 DCI.AddToWorklist(V.getNode());
22774 V = DAG.getNode(X86ISD::PSHUFD, DL, DVT, V,
22775 getV4X86ShuffleImm8ForMask(DMask, DL, DAG));
22776 DCI.AddToWorklist(V.getNode());
22777 return DAG.getBitcast(VT, V);
22780 // Look for shuffle patterns which can be implemented as a single unpack.
22781 // FIXME: This doesn't handle the location of the PSHUFD generically, and
22782 // only works when we have a PSHUFD followed by two half-shuffles.
22783 if (Mask[0] == Mask[1] && Mask[2] == Mask[3] &&
22784 (V.getOpcode() == X86ISD::PSHUFLW ||
22785 V.getOpcode() == X86ISD::PSHUFHW) &&
22786 V.getOpcode() != N.getOpcode() &&
22788 SDValue D = V.getOperand(0);
22789 while (D.getOpcode() == ISD::BITCAST && D.hasOneUse())
22790 D = D.getOperand(0);
22791 if (D.getOpcode() == X86ISD::PSHUFD && D.hasOneUse()) {
22792 SmallVector<int, 4> VMask = getPSHUFShuffleMask(V);
22793 SmallVector<int, 4> DMask = getPSHUFShuffleMask(D);
22794 int NOffset = N.getOpcode() == X86ISD::PSHUFLW ? 0 : 4;
22795 int VOffset = V.getOpcode() == X86ISD::PSHUFLW ? 0 : 4;
22797 for (int i = 0; i < 4; ++i) {
22798 WordMask[i + NOffset] = Mask[i] + NOffset;
22799 WordMask[i + VOffset] = VMask[i] + VOffset;
22801 // Map the word mask through the DWord mask.
22803 for (int i = 0; i < 8; ++i)
22804 MappedMask[i] = 2 * DMask[WordMask[i] / 2] + WordMask[i] % 2;
22805 if (makeArrayRef(MappedMask).equals({0, 0, 1, 1, 2, 2, 3, 3}) ||
22806 makeArrayRef(MappedMask).equals({4, 4, 5, 5, 6, 6, 7, 7})) {
22807 // We can replace all three shuffles with an unpack.
22808 V = DAG.getBitcast(VT, D.getOperand(0));
22809 DCI.AddToWorklist(V.getNode());
22810 return DAG.getNode(MappedMask[0] == 0 ? X86ISD::UNPCKL
22819 case X86ISD::PSHUFD:
22820 if (SDValue NewN = combineRedundantDWordShuffle(N, Mask, DAG, DCI))
22829 /// \brief Try to combine a shuffle into a target-specific add-sub node.
22831 /// We combine this directly on the abstract vector shuffle nodes so it is
22832 /// easier to generically match. We also insert dummy vector shuffle nodes for
22833 /// the operands which explicitly discard the lanes which are unused by this
22834 /// operation to try to flow through the rest of the combiner the fact that
22835 /// they're unused.
22836 static SDValue combineShuffleToAddSub(SDNode *N, SelectionDAG &DAG) {
22838 EVT VT = N->getValueType(0);
22840 // We only handle target-independent shuffles.
22841 // FIXME: It would be easy and harmless to use the target shuffle mask
22842 // extraction tool to support more.
22843 if (N->getOpcode() != ISD::VECTOR_SHUFFLE)
22846 auto *SVN = cast<ShuffleVectorSDNode>(N);
22847 ArrayRef<int> Mask = SVN->getMask();
22848 SDValue V1 = N->getOperand(0);
22849 SDValue V2 = N->getOperand(1);
22851 // We require the first shuffle operand to be the SUB node, and the second to
22852 // be the ADD node.
22853 // FIXME: We should support the commuted patterns.
22854 if (V1->getOpcode() != ISD::FSUB || V2->getOpcode() != ISD::FADD)
22857 // If there are other uses of these operations we can't fold them.
22858 if (!V1->hasOneUse() || !V2->hasOneUse())
22861 // Ensure that both operations have the same operands. Note that we can
22862 // commute the FADD operands.
22863 SDValue LHS = V1->getOperand(0), RHS = V1->getOperand(1);
22864 if ((V2->getOperand(0) != LHS || V2->getOperand(1) != RHS) &&
22865 (V2->getOperand(0) != RHS || V2->getOperand(1) != LHS))
22868 // We're looking for blends between FADD and FSUB nodes. We insist on these
22869 // nodes being lined up in a specific expected pattern.
22870 if (!(isShuffleEquivalent(V1, V2, Mask, {0, 3}) ||
22871 isShuffleEquivalent(V1, V2, Mask, {0, 5, 2, 7}) ||
22872 isShuffleEquivalent(V1, V2, Mask, {0, 9, 2, 11, 4, 13, 6, 15})))
22875 // Only specific types are legal at this point, assert so we notice if and
22876 // when these change.
22877 assert((VT == MVT::v4f32 || VT == MVT::v2f64 || VT == MVT::v8f32 ||
22878 VT == MVT::v4f64) &&
22879 "Unknown vector type encountered!");
22881 return DAG.getNode(X86ISD::ADDSUB, DL, VT, LHS, RHS);
22884 /// PerformShuffleCombine - Performs several different shuffle combines.
22885 static SDValue PerformShuffleCombine(SDNode *N, SelectionDAG &DAG,
22886 TargetLowering::DAGCombinerInfo &DCI,
22887 const X86Subtarget *Subtarget) {
22889 SDValue N0 = N->getOperand(0);
22890 SDValue N1 = N->getOperand(1);
22891 EVT VT = N->getValueType(0);
22893 // Don't create instructions with illegal types after legalize types has run.
22894 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
22895 if (!DCI.isBeforeLegalize() && !TLI.isTypeLegal(VT.getVectorElementType()))
22898 // If we have legalized the vector types, look for blends of FADD and FSUB
22899 // nodes that we can fuse into an ADDSUB node.
22900 if (TLI.isTypeLegal(VT) && Subtarget->hasSSE3())
22901 if (SDValue AddSub = combineShuffleToAddSub(N, DAG))
22904 // Combine 256-bit vector shuffles. This is only profitable when in AVX mode
22905 if (Subtarget->hasFp256() && VT.is256BitVector() &&
22906 N->getOpcode() == ISD::VECTOR_SHUFFLE)
22907 return PerformShuffleCombine256(N, DAG, DCI, Subtarget);
22909 // During Type Legalization, when promoting illegal vector types,
22910 // the backend might introduce new shuffle dag nodes and bitcasts.
22912 // This code performs the following transformation:
22913 // fold: (shuffle (bitcast (BINOP A, B)), Undef, <Mask>) ->
22914 // (shuffle (BINOP (bitcast A), (bitcast B)), Undef, <Mask>)
22916 // We do this only if both the bitcast and the BINOP dag nodes have
22917 // one use. Also, perform this transformation only if the new binary
22918 // operation is legal. This is to avoid introducing dag nodes that
22919 // potentially need to be further expanded (or custom lowered) into a
22920 // less optimal sequence of dag nodes.
22921 if (!DCI.isBeforeLegalize() && DCI.isBeforeLegalizeOps() &&
22922 N1.getOpcode() == ISD::UNDEF && N0.hasOneUse() &&
22923 N0.getOpcode() == ISD::BITCAST) {
22924 SDValue BC0 = N0.getOperand(0);
22925 EVT SVT = BC0.getValueType();
22926 unsigned Opcode = BC0.getOpcode();
22927 unsigned NumElts = VT.getVectorNumElements();
22929 if (BC0.hasOneUse() && SVT.isVector() &&
22930 SVT.getVectorNumElements() * 2 == NumElts &&
22931 TLI.isOperationLegal(Opcode, VT)) {
22932 bool CanFold = false;
22944 unsigned SVTNumElts = SVT.getVectorNumElements();
22945 ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(N);
22946 for (unsigned i = 0, e = SVTNumElts; i != e && CanFold; ++i)
22947 CanFold = SVOp->getMaskElt(i) == (int)(i * 2);
22948 for (unsigned i = SVTNumElts, e = NumElts; i != e && CanFold; ++i)
22949 CanFold = SVOp->getMaskElt(i) < 0;
22952 SDValue BC00 = DAG.getBitcast(VT, BC0.getOperand(0));
22953 SDValue BC01 = DAG.getBitcast(VT, BC0.getOperand(1));
22954 SDValue NewBinOp = DAG.getNode(BC0.getOpcode(), dl, VT, BC00, BC01);
22955 return DAG.getVectorShuffle(VT, dl, NewBinOp, N1, &SVOp->getMask()[0]);
22960 // Combine a vector_shuffle that is equal to build_vector load1, load2, load3,
22961 // load4, <0, 1, 2, 3> into a 128-bit load if the load addresses are
22962 // consecutive, non-overlapping, and in the right order.
22963 SmallVector<SDValue, 16> Elts;
22964 for (unsigned i = 0, e = VT.getVectorNumElements(); i != e; ++i)
22965 Elts.push_back(getShuffleScalarElt(N, i, DAG, 0));
22967 if (SDValue LD = EltsFromConsecutiveLoads(VT, Elts, dl, DAG, true))
22970 if (isTargetShuffle(N->getOpcode())) {
22972 PerformTargetShuffleCombine(SDValue(N, 0), DAG, DCI, Subtarget);
22973 if (Shuffle.getNode())
22976 // Try recursively combining arbitrary sequences of x86 shuffle
22977 // instructions into higher-order shuffles. We do this after combining
22978 // specific PSHUF instruction sequences into their minimal form so that we
22979 // can evaluate how many specialized shuffle instructions are involved in
22980 // a particular chain.
22981 SmallVector<int, 1> NonceMask; // Just a placeholder.
22982 NonceMask.push_back(0);
22983 if (combineX86ShufflesRecursively(SDValue(N, 0), SDValue(N, 0), NonceMask,
22984 /*Depth*/ 1, /*HasPSHUFB*/ false, DAG,
22986 return SDValue(); // This routine will use CombineTo to replace N.
22992 /// XFormVExtractWithShuffleIntoLoad - Check if a vector extract from a target
22993 /// specific shuffle of a load can be folded into a single element load.
22994 /// Similar handling for VECTOR_SHUFFLE is performed by DAGCombiner, but
22995 /// shuffles have been custom lowered so we need to handle those here.
22996 static SDValue XFormVExtractWithShuffleIntoLoad(SDNode *N, SelectionDAG &DAG,
22997 TargetLowering::DAGCombinerInfo &DCI) {
22998 if (DCI.isBeforeLegalizeOps())
23001 SDValue InVec = N->getOperand(0);
23002 SDValue EltNo = N->getOperand(1);
23004 if (!isa<ConstantSDNode>(EltNo))
23007 EVT OriginalVT = InVec.getValueType();
23009 if (InVec.getOpcode() == ISD::BITCAST) {
23010 // Don't duplicate a load with other uses.
23011 if (!InVec.hasOneUse())
23013 EVT BCVT = InVec.getOperand(0).getValueType();
23014 if (!BCVT.isVector() ||
23015 BCVT.getVectorNumElements() != OriginalVT.getVectorNumElements())
23017 InVec = InVec.getOperand(0);
23020 EVT CurrentVT = InVec.getValueType();
23022 if (!isTargetShuffle(InVec.getOpcode()))
23025 // Don't duplicate a load with other uses.
23026 if (!InVec.hasOneUse())
23029 SmallVector<int, 16> ShuffleMask;
23031 if (!getTargetShuffleMask(InVec.getNode(), CurrentVT.getSimpleVT(),
23032 ShuffleMask, UnaryShuffle))
23035 // Select the input vector, guarding against out of range extract vector.
23036 unsigned NumElems = CurrentVT.getVectorNumElements();
23037 int Elt = cast<ConstantSDNode>(EltNo)->getZExtValue();
23038 int Idx = (Elt > (int)NumElems) ? -1 : ShuffleMask[Elt];
23039 SDValue LdNode = (Idx < (int)NumElems) ? InVec.getOperand(0)
23040 : InVec.getOperand(1);
23042 // If inputs to shuffle are the same for both ops, then allow 2 uses
23043 unsigned AllowedUses = InVec.getNumOperands() > 1 &&
23044 InVec.getOperand(0) == InVec.getOperand(1) ? 2 : 1;
23046 if (LdNode.getOpcode() == ISD::BITCAST) {
23047 // Don't duplicate a load with other uses.
23048 if (!LdNode.getNode()->hasNUsesOfValue(AllowedUses, 0))
23051 AllowedUses = 1; // only allow 1 load use if we have a bitcast
23052 LdNode = LdNode.getOperand(0);
23055 if (!ISD::isNormalLoad(LdNode.getNode()))
23058 LoadSDNode *LN0 = cast<LoadSDNode>(LdNode);
23060 if (!LN0 ||!LN0->hasNUsesOfValue(AllowedUses, 0) || LN0->isVolatile())
23063 EVT EltVT = N->getValueType(0);
23064 // If there's a bitcast before the shuffle, check if the load type and
23065 // alignment is valid.
23066 unsigned Align = LN0->getAlignment();
23067 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
23068 unsigned NewAlign = DAG.getDataLayout().getABITypeAlignment(
23069 EltVT.getTypeForEVT(*DAG.getContext()));
23071 if (NewAlign > Align || !TLI.isOperationLegalOrCustom(ISD::LOAD, EltVT))
23074 // All checks match so transform back to vector_shuffle so that DAG combiner
23075 // can finish the job
23078 // Create shuffle node taking into account the case that its a unary shuffle
23079 SDValue Shuffle = (UnaryShuffle) ? DAG.getUNDEF(CurrentVT)
23080 : InVec.getOperand(1);
23081 Shuffle = DAG.getVectorShuffle(CurrentVT, dl,
23082 InVec.getOperand(0), Shuffle,
23084 Shuffle = DAG.getBitcast(OriginalVT, Shuffle);
23085 return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, N->getValueType(0), Shuffle,
23089 static SDValue PerformBITCASTCombine(SDNode *N, SelectionDAG &DAG,
23090 const X86Subtarget *Subtarget) {
23091 SDValue N0 = N->getOperand(0);
23092 EVT VT = N->getValueType(0);
23094 // Detect bitcasts between i32 to x86mmx low word. Since MMX types are
23095 // special and don't usually play with other vector types, it's better to
23096 // handle them early to be sure we emit efficient code by avoiding
23097 // store-load conversions.
23098 if (VT == MVT::x86mmx && N0.getOpcode() == ISD::BUILD_VECTOR &&
23099 N0.getValueType() == MVT::v2i32 &&
23100 isa<ConstantSDNode>(N0.getOperand(1))) {
23101 SDValue N00 = N0->getOperand(0);
23102 if (N0.getConstantOperandVal(1) == 0 && N00.getValueType() == MVT::i32)
23103 return DAG.getNode(X86ISD::MMX_MOVW2D, SDLoc(N00), VT, N00);
23106 // Convert a bitcasted integer logic operation that has one bitcasted
23107 // floating-point operand and one constant operand into a floating-point
23108 // logic operation. This may create a load of the constant, but that is
23109 // cheaper than materializing the constant in an integer register and
23110 // transferring it to an SSE register or transferring the SSE operand to
23111 // integer register and back.
23113 switch (N0.getOpcode()) {
23114 case ISD::AND: FPOpcode = X86ISD::FAND; break;
23115 case ISD::OR: FPOpcode = X86ISD::FOR; break;
23116 case ISD::XOR: FPOpcode = X86ISD::FXOR; break;
23117 default: return SDValue();
23119 if (((Subtarget->hasSSE1() && VT == MVT::f32) ||
23120 (Subtarget->hasSSE2() && VT == MVT::f64)) &&
23121 isa<ConstantSDNode>(N0.getOperand(1)) &&
23122 N0.getOperand(0).getOpcode() == ISD::BITCAST &&
23123 N0.getOperand(0).getOperand(0).getValueType() == VT) {
23124 SDValue N000 = N0.getOperand(0).getOperand(0);
23125 SDValue FPConst = DAG.getBitcast(VT, N0.getOperand(1));
23126 return DAG.getNode(FPOpcode, SDLoc(N0), VT, N000, FPConst);
23132 /// PerformEXTRACT_VECTOR_ELTCombine - Detect vector gather/scatter index
23133 /// generation and convert it from being a bunch of shuffles and extracts
23134 /// into a somewhat faster sequence. For i686, the best sequence is apparently
23135 /// storing the value and loading scalars back, while for x64 we should
23136 /// use 64-bit extracts and shifts.
23137 static SDValue PerformEXTRACT_VECTOR_ELTCombine(SDNode *N, SelectionDAG &DAG,
23138 TargetLowering::DAGCombinerInfo &DCI) {
23139 if (SDValue NewOp = XFormVExtractWithShuffleIntoLoad(N, DAG, DCI))
23142 SDValue InputVector = N->getOperand(0);
23143 SDLoc dl(InputVector);
23144 // Detect mmx to i32 conversion through a v2i32 elt extract.
23145 if (InputVector.getOpcode() == ISD::BITCAST && InputVector.hasOneUse() &&
23146 N->getValueType(0) == MVT::i32 &&
23147 InputVector.getValueType() == MVT::v2i32) {
23149 // The bitcast source is a direct mmx result.
23150 SDValue MMXSrc = InputVector.getNode()->getOperand(0);
23151 if (MMXSrc.getValueType() == MVT::x86mmx)
23152 return DAG.getNode(X86ISD::MMX_MOVD2W, SDLoc(InputVector),
23153 N->getValueType(0),
23154 InputVector.getNode()->getOperand(0));
23156 // The mmx is indirect: (i64 extract_elt (v1i64 bitcast (x86mmx ...))).
23157 if (MMXSrc.getOpcode() == ISD::EXTRACT_VECTOR_ELT && MMXSrc.hasOneUse() &&
23158 MMXSrc.getValueType() == MVT::i64) {
23159 SDValue MMXSrcOp = MMXSrc.getOperand(0);
23160 if (MMXSrcOp.hasOneUse() && MMXSrcOp.getOpcode() == ISD::BITCAST &&
23161 MMXSrcOp.getValueType() == MVT::v1i64 &&
23162 MMXSrcOp.getOperand(0).getValueType() == MVT::x86mmx)
23163 return DAG.getNode(X86ISD::MMX_MOVD2W, SDLoc(InputVector),
23164 N->getValueType(0), MMXSrcOp.getOperand(0));
23168 EVT VT = N->getValueType(0);
23170 if (VT == MVT::i1 && dyn_cast<ConstantSDNode>(N->getOperand(1)) &&
23171 InputVector.getOpcode() == ISD::BITCAST &&
23172 dyn_cast<ConstantSDNode>(InputVector.getOperand(0))) {
23173 uint64_t ExtractedElt =
23174 cast<ConstantSDNode>(N->getOperand(1))->getZExtValue();
23175 uint64_t InputValue =
23176 cast<ConstantSDNode>(InputVector.getOperand(0))->getZExtValue();
23177 uint64_t Res = (InputValue >> ExtractedElt) & 1;
23178 return DAG.getConstant(Res, dl, MVT::i1);
23180 // Only operate on vectors of 4 elements, where the alternative shuffling
23181 // gets to be more expensive.
23182 if (InputVector.getValueType() != MVT::v4i32)
23185 // Check whether every use of InputVector is an EXTRACT_VECTOR_ELT with a
23186 // single use which is a sign-extend or zero-extend, and all elements are
23188 SmallVector<SDNode *, 4> Uses;
23189 unsigned ExtractedElements = 0;
23190 for (SDNode::use_iterator UI = InputVector.getNode()->use_begin(),
23191 UE = InputVector.getNode()->use_end(); UI != UE; ++UI) {
23192 if (UI.getUse().getResNo() != InputVector.getResNo())
23195 SDNode *Extract = *UI;
23196 if (Extract->getOpcode() != ISD::EXTRACT_VECTOR_ELT)
23199 if (Extract->getValueType(0) != MVT::i32)
23201 if (!Extract->hasOneUse())
23203 if (Extract->use_begin()->getOpcode() != ISD::SIGN_EXTEND &&
23204 Extract->use_begin()->getOpcode() != ISD::ZERO_EXTEND)
23206 if (!isa<ConstantSDNode>(Extract->getOperand(1)))
23209 // Record which element was extracted.
23210 ExtractedElements |=
23211 1 << cast<ConstantSDNode>(Extract->getOperand(1))->getZExtValue();
23213 Uses.push_back(Extract);
23216 // If not all the elements were used, this may not be worthwhile.
23217 if (ExtractedElements != 15)
23220 // Ok, we've now decided to do the transformation.
23221 // If 64-bit shifts are legal, use the extract-shift sequence,
23222 // otherwise bounce the vector off the cache.
23223 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
23226 if (TLI.isOperationLegal(ISD::SRA, MVT::i64)) {
23227 SDValue Cst = DAG.getBitcast(MVT::v2i64, InputVector);
23228 auto &DL = DAG.getDataLayout();
23229 EVT VecIdxTy = DAG.getTargetLoweringInfo().getVectorIdxTy(DL);
23230 SDValue BottomHalf = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i64, Cst,
23231 DAG.getConstant(0, dl, VecIdxTy));
23232 SDValue TopHalf = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i64, Cst,
23233 DAG.getConstant(1, dl, VecIdxTy));
23235 SDValue ShAmt = DAG.getConstant(
23236 32, dl, DAG.getTargetLoweringInfo().getShiftAmountTy(MVT::i64, DL));
23237 Vals[0] = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, BottomHalf);
23238 Vals[1] = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32,
23239 DAG.getNode(ISD::SRA, dl, MVT::i64, BottomHalf, ShAmt));
23240 Vals[2] = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32, TopHalf);
23241 Vals[3] = DAG.getNode(ISD::TRUNCATE, dl, MVT::i32,
23242 DAG.getNode(ISD::SRA, dl, MVT::i64, TopHalf, ShAmt));
23244 // Store the value to a temporary stack slot.
23245 SDValue StackPtr = DAG.CreateStackTemporary(InputVector.getValueType());
23246 SDValue Ch = DAG.getStore(DAG.getEntryNode(), dl, InputVector, StackPtr,
23247 MachinePointerInfo(), false, false, 0);
23249 EVT ElementType = InputVector.getValueType().getVectorElementType();
23250 unsigned EltSize = ElementType.getSizeInBits() / 8;
23252 // Replace each use (extract) with a load of the appropriate element.
23253 for (unsigned i = 0; i < 4; ++i) {
23254 uint64_t Offset = EltSize * i;
23255 auto PtrVT = TLI.getPointerTy(DAG.getDataLayout());
23256 SDValue OffsetVal = DAG.getConstant(Offset, dl, PtrVT);
23258 SDValue ScalarAddr =
23259 DAG.getNode(ISD::ADD, dl, PtrVT, StackPtr, OffsetVal);
23261 // Load the scalar.
23262 Vals[i] = DAG.getLoad(ElementType, dl, Ch,
23263 ScalarAddr, MachinePointerInfo(),
23264 false, false, false, 0);
23269 // Replace the extracts
23270 for (SmallVectorImpl<SDNode *>::iterator UI = Uses.begin(),
23271 UE = Uses.end(); UI != UE; ++UI) {
23272 SDNode *Extract = *UI;
23274 SDValue Idx = Extract->getOperand(1);
23275 uint64_t IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
23276 DAG.ReplaceAllUsesOfValueWith(SDValue(Extract, 0), Vals[IdxVal]);
23279 // The replacement was made in place; don't return anything.
23284 transformVSELECTtoBlendVECTOR_SHUFFLE(SDNode *N, SelectionDAG &DAG,
23285 const X86Subtarget *Subtarget) {
23287 SDValue Cond = N->getOperand(0);
23288 SDValue LHS = N->getOperand(1);
23289 SDValue RHS = N->getOperand(2);
23291 if (Cond.getOpcode() == ISD::SIGN_EXTEND) {
23292 SDValue CondSrc = Cond->getOperand(0);
23293 if (CondSrc->getOpcode() == ISD::SIGN_EXTEND_INREG)
23294 Cond = CondSrc->getOperand(0);
23297 if (!ISD::isBuildVectorOfConstantSDNodes(Cond.getNode()))
23300 // A vselect where all conditions and data are constants can be optimized into
23301 // a single vector load by SelectionDAGLegalize::ExpandBUILD_VECTOR().
23302 if (ISD::isBuildVectorOfConstantSDNodes(LHS.getNode()) &&
23303 ISD::isBuildVectorOfConstantSDNodes(RHS.getNode()))
23306 unsigned MaskValue = 0;
23307 if (!BUILD_VECTORtoBlendMask(cast<BuildVectorSDNode>(Cond), MaskValue))
23310 MVT VT = N->getSimpleValueType(0);
23311 unsigned NumElems = VT.getVectorNumElements();
23312 SmallVector<int, 8> ShuffleMask(NumElems, -1);
23313 for (unsigned i = 0; i < NumElems; ++i) {
23314 // Be sure we emit undef where we can.
23315 if (Cond.getOperand(i)->getOpcode() == ISD::UNDEF)
23316 ShuffleMask[i] = -1;
23318 ShuffleMask[i] = i + NumElems * ((MaskValue >> i) & 1);
23321 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
23322 if (!TLI.isShuffleMaskLegal(ShuffleMask, VT))
23324 return DAG.getVectorShuffle(VT, dl, LHS, RHS, &ShuffleMask[0]);
23327 /// PerformSELECTCombine - Do target-specific dag combines on SELECT and VSELECT
23329 static SDValue PerformSELECTCombine(SDNode *N, SelectionDAG &DAG,
23330 TargetLowering::DAGCombinerInfo &DCI,
23331 const X86Subtarget *Subtarget) {
23333 SDValue Cond = N->getOperand(0);
23334 // Get the LHS/RHS of the select.
23335 SDValue LHS = N->getOperand(1);
23336 SDValue RHS = N->getOperand(2);
23337 EVT VT = LHS.getValueType();
23338 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
23340 // If we have SSE[12] support, try to form min/max nodes. SSE min/max
23341 // instructions match the semantics of the common C idiom x<y?x:y but not
23342 // x<=y?x:y, because of how they handle negative zero (which can be
23343 // ignored in unsafe-math mode).
23344 // We also try to create v2f32 min/max nodes, which we later widen to v4f32.
23345 if (Cond.getOpcode() == ISD::SETCC && VT.isFloatingPoint() &&
23346 VT != MVT::f80 && (TLI.isTypeLegal(VT) || VT == MVT::v2f32) &&
23347 (Subtarget->hasSSE2() ||
23348 (Subtarget->hasSSE1() && VT.getScalarType() == MVT::f32))) {
23349 ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
23351 unsigned Opcode = 0;
23352 // Check for x CC y ? x : y.
23353 if (DAG.isEqualTo(LHS, Cond.getOperand(0)) &&
23354 DAG.isEqualTo(RHS, Cond.getOperand(1))) {
23358 // Converting this to a min would handle NaNs incorrectly, and swapping
23359 // the operands would cause it to handle comparisons between positive
23360 // and negative zero incorrectly.
23361 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)) {
23362 if (!DAG.getTarget().Options.UnsafeFPMath &&
23363 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
23365 std::swap(LHS, RHS);
23367 Opcode = X86ISD::FMIN;
23370 // Converting this to a min would handle comparisons between positive
23371 // and negative zero incorrectly.
23372 if (!DAG.getTarget().Options.UnsafeFPMath &&
23373 !DAG.isKnownNeverZero(LHS) && !DAG.isKnownNeverZero(RHS))
23375 Opcode = X86ISD::FMIN;
23378 // Converting this to a min would handle both negative zeros and NaNs
23379 // incorrectly, but we can swap the operands to fix both.
23380 std::swap(LHS, RHS);
23384 Opcode = X86ISD::FMIN;
23388 // Converting this to a max would handle comparisons between positive
23389 // and negative zero incorrectly.
23390 if (!DAG.getTarget().Options.UnsafeFPMath &&
23391 !DAG.isKnownNeverZero(LHS) && !DAG.isKnownNeverZero(RHS))
23393 Opcode = X86ISD::FMAX;
23396 // Converting this to a max would handle NaNs incorrectly, and swapping
23397 // the operands would cause it to handle comparisons between positive
23398 // and negative zero incorrectly.
23399 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)) {
23400 if (!DAG.getTarget().Options.UnsafeFPMath &&
23401 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS)))
23403 std::swap(LHS, RHS);
23405 Opcode = X86ISD::FMAX;
23408 // Converting this to a max 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::FMAX;
23417 // Check for x CC y ? y : x -- a min/max with reversed arms.
23418 } else if (DAG.isEqualTo(LHS, Cond.getOperand(1)) &&
23419 DAG.isEqualTo(RHS, Cond.getOperand(0))) {
23423 // Converting this to a min would handle comparisons between positive
23424 // and negative zero incorrectly, and swapping the operands would
23425 // cause it to handle NaNs incorrectly.
23426 if (!DAG.getTarget().Options.UnsafeFPMath &&
23427 !(DAG.isKnownNeverZero(LHS) || DAG.isKnownNeverZero(RHS))) {
23428 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS))
23430 std::swap(LHS, RHS);
23432 Opcode = X86ISD::FMIN;
23435 // Converting this to a min would handle NaNs incorrectly.
23436 if (!DAG.getTarget().Options.UnsafeFPMath &&
23437 (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS)))
23439 Opcode = X86ISD::FMIN;
23442 // Converting this to a min would handle both negative zeros and NaNs
23443 // incorrectly, but we can swap the operands to fix both.
23444 std::swap(LHS, RHS);
23448 Opcode = X86ISD::FMIN;
23452 // Converting this to a max would handle NaNs incorrectly.
23453 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS))
23455 Opcode = X86ISD::FMAX;
23458 // Converting this to a max would handle comparisons between positive
23459 // and negative zero incorrectly, and swapping the operands would
23460 // cause it to handle NaNs incorrectly.
23461 if (!DAG.getTarget().Options.UnsafeFPMath &&
23462 !DAG.isKnownNeverZero(LHS) && !DAG.isKnownNeverZero(RHS)) {
23463 if (!DAG.isKnownNeverNaN(LHS) || !DAG.isKnownNeverNaN(RHS))
23465 std::swap(LHS, RHS);
23467 Opcode = X86ISD::FMAX;
23470 // Converting this to a max would handle both negative zeros and NaNs
23471 // incorrectly, but we can swap the operands to fix both.
23472 std::swap(LHS, RHS);
23476 Opcode = X86ISD::FMAX;
23482 return DAG.getNode(Opcode, DL, N->getValueType(0), LHS, RHS);
23485 EVT CondVT = Cond.getValueType();
23486 if (Subtarget->hasAVX512() && VT.isVector() && CondVT.isVector() &&
23487 CondVT.getVectorElementType() == MVT::i1) {
23488 // v16i8 (select v16i1, v16i8, v16i8) does not have a proper
23489 // lowering on KNL. In this case we convert it to
23490 // v16i8 (select v16i8, v16i8, v16i8) and use AVX instruction.
23491 // The same situation for all 128 and 256-bit vectors of i8 and i16.
23492 // Since SKX these selects have a proper lowering.
23493 EVT OpVT = LHS.getValueType();
23494 if ((OpVT.is128BitVector() || OpVT.is256BitVector()) &&
23495 (OpVT.getVectorElementType() == MVT::i8 ||
23496 OpVT.getVectorElementType() == MVT::i16) &&
23497 !(Subtarget->hasBWI() && Subtarget->hasVLX())) {
23498 Cond = DAG.getNode(ISD::SIGN_EXTEND, DL, OpVT, Cond);
23499 DCI.AddToWorklist(Cond.getNode());
23500 return DAG.getNode(N->getOpcode(), DL, OpVT, Cond, LHS, RHS);
23503 // If this is a select between two integer constants, try to do some
23505 if (ConstantSDNode *TrueC = dyn_cast<ConstantSDNode>(LHS)) {
23506 if (ConstantSDNode *FalseC = dyn_cast<ConstantSDNode>(RHS))
23507 // Don't do this for crazy integer types.
23508 if (DAG.getTargetLoweringInfo().isTypeLegal(LHS.getValueType())) {
23509 // If this is efficiently invertible, canonicalize the LHSC/RHSC values
23510 // so that TrueC (the true value) is larger than FalseC.
23511 bool NeedsCondInvert = false;
23513 if (TrueC->getAPIntValue().ult(FalseC->getAPIntValue()) &&
23514 // Efficiently invertible.
23515 (Cond.getOpcode() == ISD::SETCC || // setcc -> invertible.
23516 (Cond.getOpcode() == ISD::XOR && // xor(X, C) -> invertible.
23517 isa<ConstantSDNode>(Cond.getOperand(1))))) {
23518 NeedsCondInvert = true;
23519 std::swap(TrueC, FalseC);
23522 // Optimize C ? 8 : 0 -> zext(C) << 3. Likewise for any pow2/0.
23523 if (FalseC->getAPIntValue() == 0 &&
23524 TrueC->getAPIntValue().isPowerOf2()) {
23525 if (NeedsCondInvert) // Invert the condition if needed.
23526 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
23527 DAG.getConstant(1, DL, Cond.getValueType()));
23529 // Zero extend the condition if needed.
23530 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, LHS.getValueType(), Cond);
23532 unsigned ShAmt = TrueC->getAPIntValue().logBase2();
23533 return DAG.getNode(ISD::SHL, DL, LHS.getValueType(), Cond,
23534 DAG.getConstant(ShAmt, DL, MVT::i8));
23537 // Optimize Cond ? cst+1 : cst -> zext(setcc(C)+cst.
23538 if (FalseC->getAPIntValue()+1 == TrueC->getAPIntValue()) {
23539 if (NeedsCondInvert) // Invert the condition if needed.
23540 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
23541 DAG.getConstant(1, DL, Cond.getValueType()));
23543 // Zero extend the condition if needed.
23544 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL,
23545 FalseC->getValueType(0), Cond);
23546 return DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
23547 SDValue(FalseC, 0));
23550 // Optimize cases that will turn into an LEA instruction. This requires
23551 // an i32 or i64 and an efficient multiplier (1, 2, 3, 4, 5, 8, 9).
23552 if (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i64) {
23553 uint64_t Diff = TrueC->getZExtValue()-FalseC->getZExtValue();
23554 if (N->getValueType(0) == MVT::i32) Diff = (unsigned)Diff;
23556 bool isFastMultiplier = false;
23558 switch ((unsigned char)Diff) {
23560 case 1: // result = add base, cond
23561 case 2: // result = lea base( , cond*2)
23562 case 3: // result = lea base(cond, cond*2)
23563 case 4: // result = lea base( , cond*4)
23564 case 5: // result = lea base(cond, cond*4)
23565 case 8: // result = lea base( , cond*8)
23566 case 9: // result = lea base(cond, cond*8)
23567 isFastMultiplier = true;
23572 if (isFastMultiplier) {
23573 APInt Diff = TrueC->getAPIntValue()-FalseC->getAPIntValue();
23574 if (NeedsCondInvert) // Invert the condition if needed.
23575 Cond = DAG.getNode(ISD::XOR, DL, Cond.getValueType(), Cond,
23576 DAG.getConstant(1, DL, Cond.getValueType()));
23578 // Zero extend the condition if needed.
23579 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, FalseC->getValueType(0),
23581 // Scale the condition by the difference.
23583 Cond = DAG.getNode(ISD::MUL, DL, Cond.getValueType(), Cond,
23584 DAG.getConstant(Diff, DL,
23585 Cond.getValueType()));
23587 // Add the base if non-zero.
23588 if (FalseC->getAPIntValue() != 0)
23589 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
23590 SDValue(FalseC, 0));
23597 // Canonicalize max and min:
23598 // (x > y) ? x : y -> (x >= y) ? x : y
23599 // (x < y) ? x : y -> (x <= y) ? x : y
23600 // This allows use of COND_S / COND_NS (see TranslateX86CC) which eliminates
23601 // the need for an extra compare
23602 // against zero. e.g.
23603 // (x - y) > 0 : (x - y) ? 0 -> (x - y) >= 0 : (x - y) ? 0
23605 // testl %edi, %edi
23607 // cmovgl %edi, %eax
23611 // cmovsl %eax, %edi
23612 if (N->getOpcode() == ISD::SELECT && Cond.getOpcode() == ISD::SETCC &&
23613 DAG.isEqualTo(LHS, Cond.getOperand(0)) &&
23614 DAG.isEqualTo(RHS, Cond.getOperand(1))) {
23615 ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
23620 ISD::CondCode NewCC = (CC == ISD::SETLT) ? ISD::SETLE : ISD::SETGE;
23621 Cond = DAG.getSetCC(SDLoc(Cond), Cond.getValueType(),
23622 Cond.getOperand(0), Cond.getOperand(1), NewCC);
23623 return DAG.getNode(ISD::SELECT, DL, VT, Cond, LHS, RHS);
23628 // Early exit check
23629 if (!TLI.isTypeLegal(VT))
23632 // Match VSELECTs into subs with unsigned saturation.
23633 if (N->getOpcode() == ISD::VSELECT && Cond.getOpcode() == ISD::SETCC &&
23634 // psubus is available in SSE2 and AVX2 for i8 and i16 vectors.
23635 ((Subtarget->hasSSE2() && (VT == MVT::v16i8 || VT == MVT::v8i16)) ||
23636 (Subtarget->hasAVX2() && (VT == MVT::v32i8 || VT == MVT::v16i16)))) {
23637 ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
23639 // Check if one of the arms of the VSELECT is a zero vector. If it's on the
23640 // left side invert the predicate to simplify logic below.
23642 if (ISD::isBuildVectorAllZeros(LHS.getNode())) {
23644 CC = ISD::getSetCCInverse(CC, true);
23645 } else if (ISD::isBuildVectorAllZeros(RHS.getNode())) {
23649 if (Other.getNode() && Other->getNumOperands() == 2 &&
23650 DAG.isEqualTo(Other->getOperand(0), Cond.getOperand(0))) {
23651 SDValue OpLHS = Other->getOperand(0), OpRHS = Other->getOperand(1);
23652 SDValue CondRHS = Cond->getOperand(1);
23654 // Look for a general sub with unsigned saturation first.
23655 // x >= y ? x-y : 0 --> subus x, y
23656 // x > y ? x-y : 0 --> subus x, y
23657 if ((CC == ISD::SETUGE || CC == ISD::SETUGT) &&
23658 Other->getOpcode() == ISD::SUB && DAG.isEqualTo(OpRHS, CondRHS))
23659 return DAG.getNode(X86ISD::SUBUS, DL, VT, OpLHS, OpRHS);
23661 if (auto *OpRHSBV = dyn_cast<BuildVectorSDNode>(OpRHS))
23662 if (auto *OpRHSConst = OpRHSBV->getConstantSplatNode()) {
23663 if (auto *CondRHSBV = dyn_cast<BuildVectorSDNode>(CondRHS))
23664 if (auto *CondRHSConst = CondRHSBV->getConstantSplatNode())
23665 // If the RHS is a constant we have to reverse the const
23666 // canonicalization.
23667 // x > C-1 ? x+-C : 0 --> subus x, C
23668 if (CC == ISD::SETUGT && Other->getOpcode() == ISD::ADD &&
23669 CondRHSConst->getAPIntValue() ==
23670 (-OpRHSConst->getAPIntValue() - 1))
23671 return DAG.getNode(
23672 X86ISD::SUBUS, DL, VT, OpLHS,
23673 DAG.getConstant(-OpRHSConst->getAPIntValue(), DL, VT));
23675 // Another special case: If C was a sign bit, the sub has been
23676 // canonicalized into a xor.
23677 // FIXME: Would it be better to use computeKnownBits to determine
23678 // whether it's safe to decanonicalize the xor?
23679 // x s< 0 ? x^C : 0 --> subus x, C
23680 if (CC == ISD::SETLT && Other->getOpcode() == ISD::XOR &&
23681 ISD::isBuildVectorAllZeros(CondRHS.getNode()) &&
23682 OpRHSConst->getAPIntValue().isSignBit())
23683 // Note that we have to rebuild the RHS constant here to ensure we
23684 // don't rely on particular values of undef lanes.
23685 return DAG.getNode(
23686 X86ISD::SUBUS, DL, VT, OpLHS,
23687 DAG.getConstant(OpRHSConst->getAPIntValue(), DL, VT));
23692 // Simplify vector selection if condition value type matches vselect
23694 if (N->getOpcode() == ISD::VSELECT && CondVT == VT) {
23695 assert(Cond.getValueType().isVector() &&
23696 "vector select expects a vector selector!");
23698 bool TValIsAllOnes = ISD::isBuildVectorAllOnes(LHS.getNode());
23699 bool FValIsAllZeros = ISD::isBuildVectorAllZeros(RHS.getNode());
23701 // Try invert the condition if true value is not all 1s and false value
23703 if (!TValIsAllOnes && !FValIsAllZeros &&
23704 // Check if the selector will be produced by CMPP*/PCMP*
23705 Cond.getOpcode() == ISD::SETCC &&
23706 // Check if SETCC has already been promoted
23707 TLI.getSetCCResultType(DAG.getDataLayout(), *DAG.getContext(), VT) ==
23709 bool TValIsAllZeros = ISD::isBuildVectorAllZeros(LHS.getNode());
23710 bool FValIsAllOnes = ISD::isBuildVectorAllOnes(RHS.getNode());
23712 if (TValIsAllZeros || FValIsAllOnes) {
23713 SDValue CC = Cond.getOperand(2);
23714 ISD::CondCode NewCC =
23715 ISD::getSetCCInverse(cast<CondCodeSDNode>(CC)->get(),
23716 Cond.getOperand(0).getValueType().isInteger());
23717 Cond = DAG.getSetCC(DL, CondVT, Cond.getOperand(0), Cond.getOperand(1), NewCC);
23718 std::swap(LHS, RHS);
23719 TValIsAllOnes = FValIsAllOnes;
23720 FValIsAllZeros = TValIsAllZeros;
23724 if (TValIsAllOnes || FValIsAllZeros) {
23727 if (TValIsAllOnes && FValIsAllZeros)
23729 else if (TValIsAllOnes)
23731 DAG.getNode(ISD::OR, DL, CondVT, Cond, DAG.getBitcast(CondVT, RHS));
23732 else if (FValIsAllZeros)
23733 Ret = DAG.getNode(ISD::AND, DL, CondVT, Cond,
23734 DAG.getBitcast(CondVT, LHS));
23736 return DAG.getBitcast(VT, Ret);
23740 // We should generate an X86ISD::BLENDI from a vselect if its argument
23741 // is a sign_extend_inreg of an any_extend of a BUILD_VECTOR of
23742 // constants. This specific pattern gets generated when we split a
23743 // selector for a 512 bit vector in a machine without AVX512 (but with
23744 // 256-bit vectors), during legalization:
23746 // (vselect (sign_extend (any_extend (BUILD_VECTOR)) i1) LHS RHS)
23748 // Iff we find this pattern and the build_vectors are built from
23749 // constants, we translate the vselect into a shuffle_vector that we
23750 // know will be matched by LowerVECTOR_SHUFFLEtoBlend.
23751 if ((N->getOpcode() == ISD::VSELECT ||
23752 N->getOpcode() == X86ISD::SHRUNKBLEND) &&
23753 !DCI.isBeforeLegalize() && !VT.is512BitVector()) {
23754 SDValue Shuffle = transformVSELECTtoBlendVECTOR_SHUFFLE(N, DAG, Subtarget);
23755 if (Shuffle.getNode())
23759 // If this is a *dynamic* select (non-constant condition) and we can match
23760 // this node with one of the variable blend instructions, restructure the
23761 // condition so that the blends can use the high bit of each element and use
23762 // SimplifyDemandedBits to simplify the condition operand.
23763 if (N->getOpcode() == ISD::VSELECT && DCI.isBeforeLegalizeOps() &&
23764 !DCI.isBeforeLegalize() &&
23765 !ISD::isBuildVectorOfConstantSDNodes(Cond.getNode())) {
23766 unsigned BitWidth = Cond.getValueType().getScalarType().getSizeInBits();
23768 // Don't optimize vector selects that map to mask-registers.
23772 // We can only handle the cases where VSELECT is directly legal on the
23773 // subtarget. We custom lower VSELECT nodes with constant conditions and
23774 // this makes it hard to see whether a dynamic VSELECT will correctly
23775 // lower, so we both check the operation's status and explicitly handle the
23776 // cases where a *dynamic* blend will fail even though a constant-condition
23777 // blend could be custom lowered.
23778 // FIXME: We should find a better way to handle this class of problems.
23779 // Potentially, we should combine constant-condition vselect nodes
23780 // pre-legalization into shuffles and not mark as many types as custom
23782 if (!TLI.isOperationLegalOrCustom(ISD::VSELECT, VT))
23784 // FIXME: We don't support i16-element blends currently. We could and
23785 // should support them by making *all* the bits in the condition be set
23786 // rather than just the high bit and using an i8-element blend.
23787 if (VT.getScalarType() == MVT::i16)
23789 // Dynamic blending was only available from SSE4.1 onward.
23790 if (VT.getSizeInBits() == 128 && !Subtarget->hasSSE41())
23792 // Byte blends are only available in AVX2
23793 if (VT.getSizeInBits() == 256 && VT.getScalarType() == MVT::i8 &&
23794 !Subtarget->hasAVX2())
23797 assert(BitWidth >= 8 && BitWidth <= 64 && "Invalid mask size");
23798 APInt DemandedMask = APInt::getHighBitsSet(BitWidth, 1);
23800 APInt KnownZero, KnownOne;
23801 TargetLowering::TargetLoweringOpt TLO(DAG, DCI.isBeforeLegalize(),
23802 DCI.isBeforeLegalizeOps());
23803 if (TLO.ShrinkDemandedConstant(Cond, DemandedMask) ||
23804 TLI.SimplifyDemandedBits(Cond, DemandedMask, KnownZero, KnownOne,
23806 // If we changed the computation somewhere in the DAG, this change
23807 // will affect all users of Cond.
23808 // Make sure it is fine and update all the nodes so that we do not
23809 // use the generic VSELECT anymore. Otherwise, we may perform
23810 // wrong optimizations as we messed up with the actual expectation
23811 // for the vector boolean values.
23812 if (Cond != TLO.Old) {
23813 // Check all uses of that condition operand to check whether it will be
23814 // consumed by non-BLEND instructions, which may depend on all bits are
23816 for (SDNode::use_iterator I = Cond->use_begin(), E = Cond->use_end();
23818 if (I->getOpcode() != ISD::VSELECT)
23819 // TODO: Add other opcodes eventually lowered into BLEND.
23822 // Update all the users of the condition, before committing the change,
23823 // so that the VSELECT optimizations that expect the correct vector
23824 // boolean value will not be triggered.
23825 for (SDNode::use_iterator I = Cond->use_begin(), E = Cond->use_end();
23827 DAG.ReplaceAllUsesOfValueWith(
23829 DAG.getNode(X86ISD::SHRUNKBLEND, SDLoc(*I), I->getValueType(0),
23830 Cond, I->getOperand(1), I->getOperand(2)));
23831 DCI.CommitTargetLoweringOpt(TLO);
23834 // At this point, only Cond is changed. Change the condition
23835 // just for N to keep the opportunity to optimize all other
23836 // users their own way.
23837 DAG.ReplaceAllUsesOfValueWith(
23839 DAG.getNode(X86ISD::SHRUNKBLEND, SDLoc(N), N->getValueType(0),
23840 TLO.New, N->getOperand(1), N->getOperand(2)));
23848 // Check whether a boolean test is testing a boolean value generated by
23849 // X86ISD::SETCC. If so, return the operand of that SETCC and proper condition
23852 // Simplify the following patterns:
23853 // (Op (CMP (SETCC Cond EFLAGS) 1) EQ) or
23854 // (Op (CMP (SETCC Cond EFLAGS) 0) NEQ)
23855 // to (Op EFLAGS Cond)
23857 // (Op (CMP (SETCC Cond EFLAGS) 0) EQ) or
23858 // (Op (CMP (SETCC Cond EFLAGS) 1) NEQ)
23859 // to (Op EFLAGS !Cond)
23861 // where Op could be BRCOND or CMOV.
23863 static SDValue checkBoolTestSetCCCombine(SDValue Cmp, X86::CondCode &CC) {
23864 // Quit if not CMP and SUB with its value result used.
23865 if (Cmp.getOpcode() != X86ISD::CMP &&
23866 (Cmp.getOpcode() != X86ISD::SUB || Cmp.getNode()->hasAnyUseOfValue(0)))
23869 // Quit if not used as a boolean value.
23870 if (CC != X86::COND_E && CC != X86::COND_NE)
23873 // Check CMP operands. One of them should be 0 or 1 and the other should be
23874 // an SetCC or extended from it.
23875 SDValue Op1 = Cmp.getOperand(0);
23876 SDValue Op2 = Cmp.getOperand(1);
23879 const ConstantSDNode* C = nullptr;
23880 bool needOppositeCond = (CC == X86::COND_E);
23881 bool checkAgainstTrue = false; // Is it a comparison against 1?
23883 if ((C = dyn_cast<ConstantSDNode>(Op1)))
23885 else if ((C = dyn_cast<ConstantSDNode>(Op2)))
23887 else // Quit if all operands are not constants.
23890 if (C->getZExtValue() == 1) {
23891 needOppositeCond = !needOppositeCond;
23892 checkAgainstTrue = true;
23893 } else if (C->getZExtValue() != 0)
23894 // Quit if the constant is neither 0 or 1.
23897 bool truncatedToBoolWithAnd = false;
23898 // Skip (zext $x), (trunc $x), or (and $x, 1) node.
23899 while (SetCC.getOpcode() == ISD::ZERO_EXTEND ||
23900 SetCC.getOpcode() == ISD::TRUNCATE ||
23901 SetCC.getOpcode() == ISD::AND) {
23902 if (SetCC.getOpcode() == ISD::AND) {
23904 ConstantSDNode *CS;
23905 if ((CS = dyn_cast<ConstantSDNode>(SetCC.getOperand(0))) &&
23906 CS->getZExtValue() == 1)
23908 if ((CS = dyn_cast<ConstantSDNode>(SetCC.getOperand(1))) &&
23909 CS->getZExtValue() == 1)
23913 SetCC = SetCC.getOperand(OpIdx);
23914 truncatedToBoolWithAnd = true;
23916 SetCC = SetCC.getOperand(0);
23919 switch (SetCC.getOpcode()) {
23920 case X86ISD::SETCC_CARRY:
23921 // Since SETCC_CARRY gives output based on R = CF ? ~0 : 0, it's unsafe to
23922 // simplify it if the result of SETCC_CARRY is not canonicalized to 0 or 1,
23923 // i.e. it's a comparison against true but the result of SETCC_CARRY is not
23924 // truncated to i1 using 'and'.
23925 if (checkAgainstTrue && !truncatedToBoolWithAnd)
23927 assert(X86::CondCode(SetCC.getConstantOperandVal(0)) == X86::COND_B &&
23928 "Invalid use of SETCC_CARRY!");
23930 case X86ISD::SETCC:
23931 // Set the condition code or opposite one if necessary.
23932 CC = X86::CondCode(SetCC.getConstantOperandVal(0));
23933 if (needOppositeCond)
23934 CC = X86::GetOppositeBranchCondition(CC);
23935 return SetCC.getOperand(1);
23936 case X86ISD::CMOV: {
23937 // Check whether false/true value has canonical one, i.e. 0 or 1.
23938 ConstantSDNode *FVal = dyn_cast<ConstantSDNode>(SetCC.getOperand(0));
23939 ConstantSDNode *TVal = dyn_cast<ConstantSDNode>(SetCC.getOperand(1));
23940 // Quit if true value is not a constant.
23943 // Quit if false value is not a constant.
23945 SDValue Op = SetCC.getOperand(0);
23946 // Skip 'zext' or 'trunc' node.
23947 if (Op.getOpcode() == ISD::ZERO_EXTEND ||
23948 Op.getOpcode() == ISD::TRUNCATE)
23949 Op = Op.getOperand(0);
23950 // A special case for rdrand/rdseed, where 0 is set if false cond is
23952 if ((Op.getOpcode() != X86ISD::RDRAND &&
23953 Op.getOpcode() != X86ISD::RDSEED) || Op.getResNo() != 0)
23956 // Quit if false value is not the constant 0 or 1.
23957 bool FValIsFalse = true;
23958 if (FVal && FVal->getZExtValue() != 0) {
23959 if (FVal->getZExtValue() != 1)
23961 // If FVal is 1, opposite cond is needed.
23962 needOppositeCond = !needOppositeCond;
23963 FValIsFalse = false;
23965 // Quit if TVal is not the constant opposite of FVal.
23966 if (FValIsFalse && TVal->getZExtValue() != 1)
23968 if (!FValIsFalse && TVal->getZExtValue() != 0)
23970 CC = X86::CondCode(SetCC.getConstantOperandVal(2));
23971 if (needOppositeCond)
23972 CC = X86::GetOppositeBranchCondition(CC);
23973 return SetCC.getOperand(3);
23980 /// Check whether Cond is an AND/OR of SETCCs off of the same EFLAGS.
23982 /// (X86or (X86setcc) (X86setcc))
23983 /// (X86cmp (and (X86setcc) (X86setcc)), 0)
23984 static bool checkBoolTestAndOrSetCCCombine(SDValue Cond, X86::CondCode &CC0,
23985 X86::CondCode &CC1, SDValue &Flags,
23987 if (Cond->getOpcode() == X86ISD::CMP) {
23988 ConstantSDNode *CondOp1C = dyn_cast<ConstantSDNode>(Cond->getOperand(1));
23989 if (!CondOp1C || !CondOp1C->isNullValue())
23992 Cond = Cond->getOperand(0);
23997 SDValue SetCC0, SetCC1;
23998 switch (Cond->getOpcode()) {
23999 default: return false;
24006 SetCC0 = Cond->getOperand(0);
24007 SetCC1 = Cond->getOperand(1);
24011 // Make sure we have SETCC nodes, using the same flags value.
24012 if (SetCC0.getOpcode() != X86ISD::SETCC ||
24013 SetCC1.getOpcode() != X86ISD::SETCC ||
24014 SetCC0->getOperand(1) != SetCC1->getOperand(1))
24017 CC0 = (X86::CondCode)SetCC0->getConstantOperandVal(0);
24018 CC1 = (X86::CondCode)SetCC1->getConstantOperandVal(0);
24019 Flags = SetCC0->getOperand(1);
24023 /// Optimize X86ISD::CMOV [LHS, RHS, CONDCODE (e.g. X86::COND_NE), CONDVAL]
24024 static SDValue PerformCMOVCombine(SDNode *N, SelectionDAG &DAG,
24025 TargetLowering::DAGCombinerInfo &DCI,
24026 const X86Subtarget *Subtarget) {
24029 // If the flag operand isn't dead, don't touch this CMOV.
24030 if (N->getNumValues() == 2 && !SDValue(N, 1).use_empty())
24033 SDValue FalseOp = N->getOperand(0);
24034 SDValue TrueOp = N->getOperand(1);
24035 X86::CondCode CC = (X86::CondCode)N->getConstantOperandVal(2);
24036 SDValue Cond = N->getOperand(3);
24038 if (CC == X86::COND_E || CC == X86::COND_NE) {
24039 switch (Cond.getOpcode()) {
24043 // If operand of BSR / BSF are proven never zero, then ZF cannot be set.
24044 if (DAG.isKnownNeverZero(Cond.getOperand(0)))
24045 return (CC == X86::COND_E) ? FalseOp : TrueOp;
24051 Flags = checkBoolTestSetCCCombine(Cond, CC);
24052 if (Flags.getNode() &&
24053 // Extra check as FCMOV only supports a subset of X86 cond.
24054 (FalseOp.getValueType() != MVT::f80 || hasFPCMov(CC))) {
24055 SDValue Ops[] = { FalseOp, TrueOp,
24056 DAG.getConstant(CC, DL, MVT::i8), Flags };
24057 return DAG.getNode(X86ISD::CMOV, DL, N->getVTList(), Ops);
24060 // If this is a select between two integer constants, try to do some
24061 // optimizations. Note that the operands are ordered the opposite of SELECT
24063 if (ConstantSDNode *TrueC = dyn_cast<ConstantSDNode>(TrueOp)) {
24064 if (ConstantSDNode *FalseC = dyn_cast<ConstantSDNode>(FalseOp)) {
24065 // Canonicalize the TrueC/FalseC values so that TrueC (the true value) is
24066 // larger than FalseC (the false value).
24067 if (TrueC->getAPIntValue().ult(FalseC->getAPIntValue())) {
24068 CC = X86::GetOppositeBranchCondition(CC);
24069 std::swap(TrueC, FalseC);
24070 std::swap(TrueOp, FalseOp);
24073 // Optimize C ? 8 : 0 -> zext(setcc(C)) << 3. Likewise for any pow2/0.
24074 // This is efficient for any integer data type (including i8/i16) and
24076 if (FalseC->getAPIntValue() == 0 && TrueC->getAPIntValue().isPowerOf2()) {
24077 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
24078 DAG.getConstant(CC, DL, MVT::i8), Cond);
24080 // Zero extend the condition if needed.
24081 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, TrueC->getValueType(0), Cond);
24083 unsigned ShAmt = TrueC->getAPIntValue().logBase2();
24084 Cond = DAG.getNode(ISD::SHL, DL, Cond.getValueType(), Cond,
24085 DAG.getConstant(ShAmt, DL, MVT::i8));
24086 if (N->getNumValues() == 2) // Dead flag value?
24087 return DCI.CombineTo(N, Cond, SDValue());
24091 // Optimize Cond ? cst+1 : cst -> zext(setcc(C)+cst. This is efficient
24092 // for any integer data type, including i8/i16.
24093 if (FalseC->getAPIntValue()+1 == TrueC->getAPIntValue()) {
24094 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
24095 DAG.getConstant(CC, DL, MVT::i8), Cond);
24097 // Zero extend the condition if needed.
24098 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL,
24099 FalseC->getValueType(0), Cond);
24100 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
24101 SDValue(FalseC, 0));
24103 if (N->getNumValues() == 2) // Dead flag value?
24104 return DCI.CombineTo(N, Cond, SDValue());
24108 // Optimize cases that will turn into an LEA instruction. This requires
24109 // an i32 or i64 and an efficient multiplier (1, 2, 3, 4, 5, 8, 9).
24110 if (N->getValueType(0) == MVT::i32 || N->getValueType(0) == MVT::i64) {
24111 uint64_t Diff = TrueC->getZExtValue()-FalseC->getZExtValue();
24112 if (N->getValueType(0) == MVT::i32) Diff = (unsigned)Diff;
24114 bool isFastMultiplier = false;
24116 switch ((unsigned char)Diff) {
24118 case 1: // result = add base, cond
24119 case 2: // result = lea base( , cond*2)
24120 case 3: // result = lea base(cond, cond*2)
24121 case 4: // result = lea base( , cond*4)
24122 case 5: // result = lea base(cond, cond*4)
24123 case 8: // result = lea base( , cond*8)
24124 case 9: // result = lea base(cond, cond*8)
24125 isFastMultiplier = true;
24130 if (isFastMultiplier) {
24131 APInt Diff = TrueC->getAPIntValue()-FalseC->getAPIntValue();
24132 Cond = DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
24133 DAG.getConstant(CC, DL, MVT::i8), Cond);
24134 // Zero extend the condition if needed.
24135 Cond = DAG.getNode(ISD::ZERO_EXTEND, DL, FalseC->getValueType(0),
24137 // Scale the condition by the difference.
24139 Cond = DAG.getNode(ISD::MUL, DL, Cond.getValueType(), Cond,
24140 DAG.getConstant(Diff, DL, Cond.getValueType()));
24142 // Add the base if non-zero.
24143 if (FalseC->getAPIntValue() != 0)
24144 Cond = DAG.getNode(ISD::ADD, DL, Cond.getValueType(), Cond,
24145 SDValue(FalseC, 0));
24146 if (N->getNumValues() == 2) // Dead flag value?
24147 return DCI.CombineTo(N, Cond, SDValue());
24154 // Handle these cases:
24155 // (select (x != c), e, c) -> select (x != c), e, x),
24156 // (select (x == c), c, e) -> select (x == c), x, e)
24157 // where the c is an integer constant, and the "select" is the combination
24158 // of CMOV and CMP.
24160 // The rationale for this change is that the conditional-move from a constant
24161 // needs two instructions, however, conditional-move from a register needs
24162 // only one instruction.
24164 // CAVEAT: By replacing a constant with a symbolic value, it may obscure
24165 // some instruction-combining opportunities. This opt needs to be
24166 // postponed as late as possible.
24168 if (!DCI.isBeforeLegalize() && !DCI.isBeforeLegalizeOps()) {
24169 // the DCI.xxxx conditions are provided to postpone the optimization as
24170 // late as possible.
24172 ConstantSDNode *CmpAgainst = nullptr;
24173 if ((Cond.getOpcode() == X86ISD::CMP || Cond.getOpcode() == X86ISD::SUB) &&
24174 (CmpAgainst = dyn_cast<ConstantSDNode>(Cond.getOperand(1))) &&
24175 !isa<ConstantSDNode>(Cond.getOperand(0))) {
24177 if (CC == X86::COND_NE &&
24178 CmpAgainst == dyn_cast<ConstantSDNode>(FalseOp)) {
24179 CC = X86::GetOppositeBranchCondition(CC);
24180 std::swap(TrueOp, FalseOp);
24183 if (CC == X86::COND_E &&
24184 CmpAgainst == dyn_cast<ConstantSDNode>(TrueOp)) {
24185 SDValue Ops[] = { FalseOp, Cond.getOperand(0),
24186 DAG.getConstant(CC, DL, MVT::i8), Cond };
24187 return DAG.getNode(X86ISD::CMOV, DL, N->getVTList (), Ops);
24192 // Fold and/or of setcc's to double CMOV:
24193 // (CMOV F, T, ((cc1 | cc2) != 0)) -> (CMOV (CMOV F, T, cc1), T, cc2)
24194 // (CMOV F, T, ((cc1 & cc2) != 0)) -> (CMOV (CMOV T, F, !cc1), F, !cc2)
24196 // This combine lets us generate:
24197 // cmovcc1 (jcc1 if we don't have CMOV)
24203 // cmovne (jne if we don't have CMOV)
24204 // When we can't use the CMOV instruction, it might increase branch
24206 // When we can use CMOV, or when there is no mispredict, this improves
24207 // throughput and reduces register pressure.
24209 if (CC == X86::COND_NE) {
24211 X86::CondCode CC0, CC1;
24213 if (checkBoolTestAndOrSetCCCombine(Cond, CC0, CC1, Flags, isAndSetCC)) {
24215 std::swap(FalseOp, TrueOp);
24216 CC0 = X86::GetOppositeBranchCondition(CC0);
24217 CC1 = X86::GetOppositeBranchCondition(CC1);
24220 SDValue LOps[] = {FalseOp, TrueOp, DAG.getConstant(CC0, DL, MVT::i8),
24222 SDValue LCMOV = DAG.getNode(X86ISD::CMOV, DL, N->getVTList(), LOps);
24223 SDValue Ops[] = {LCMOV, TrueOp, DAG.getConstant(CC1, DL, MVT::i8), Flags};
24224 SDValue CMOV = DAG.getNode(X86ISD::CMOV, DL, N->getVTList(), Ops);
24225 DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), SDValue(CMOV.getNode(), 1));
24233 /// PerformMulCombine - Optimize a single multiply with constant into two
24234 /// in order to implement it with two cheaper instructions, e.g.
24235 /// LEA + SHL, LEA + LEA.
24236 static SDValue PerformMulCombine(SDNode *N, SelectionDAG &DAG,
24237 TargetLowering::DAGCombinerInfo &DCI) {
24238 // An imul is usually smaller than the alternative sequence.
24239 if (DAG.getMachineFunction().getFunction()->optForMinSize())
24242 if (DCI.isBeforeLegalize() || DCI.isCalledByLegalizer())
24245 EVT VT = N->getValueType(0);
24246 if (VT != MVT::i64 && VT != MVT::i32)
24249 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N->getOperand(1));
24252 uint64_t MulAmt = C->getZExtValue();
24253 if (isPowerOf2_64(MulAmt) || MulAmt == 3 || MulAmt == 5 || MulAmt == 9)
24256 uint64_t MulAmt1 = 0;
24257 uint64_t MulAmt2 = 0;
24258 if ((MulAmt % 9) == 0) {
24260 MulAmt2 = MulAmt / 9;
24261 } else if ((MulAmt % 5) == 0) {
24263 MulAmt2 = MulAmt / 5;
24264 } else if ((MulAmt % 3) == 0) {
24266 MulAmt2 = MulAmt / 3;
24269 (isPowerOf2_64(MulAmt2) || MulAmt2 == 3 || MulAmt2 == 5 || MulAmt2 == 9)){
24272 if (isPowerOf2_64(MulAmt2) &&
24273 !(N->hasOneUse() && N->use_begin()->getOpcode() == ISD::ADD))
24274 // If second multiplifer is pow2, issue it first. We want the multiply by
24275 // 3, 5, or 9 to be folded into the addressing mode unless the lone use
24277 std::swap(MulAmt1, MulAmt2);
24280 if (isPowerOf2_64(MulAmt1))
24281 NewMul = DAG.getNode(ISD::SHL, DL, VT, N->getOperand(0),
24282 DAG.getConstant(Log2_64(MulAmt1), DL, MVT::i8));
24284 NewMul = DAG.getNode(X86ISD::MUL_IMM, DL, VT, N->getOperand(0),
24285 DAG.getConstant(MulAmt1, DL, VT));
24287 if (isPowerOf2_64(MulAmt2))
24288 NewMul = DAG.getNode(ISD::SHL, DL, VT, NewMul,
24289 DAG.getConstant(Log2_64(MulAmt2), DL, MVT::i8));
24291 NewMul = DAG.getNode(X86ISD::MUL_IMM, DL, VT, NewMul,
24292 DAG.getConstant(MulAmt2, DL, VT));
24294 // Do not add new nodes to DAG combiner worklist.
24295 DCI.CombineTo(N, NewMul, false);
24300 static SDValue PerformSHLCombine(SDNode *N, SelectionDAG &DAG) {
24301 SDValue N0 = N->getOperand(0);
24302 SDValue N1 = N->getOperand(1);
24303 ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(N1);
24304 EVT VT = N0.getValueType();
24306 // fold (shl (and (setcc_c), c1), c2) -> (and setcc_c, (c1 << c2))
24307 // since the result of setcc_c is all zero's or all ones.
24308 if (VT.isInteger() && !VT.isVector() &&
24309 N1C && N0.getOpcode() == ISD::AND &&
24310 N0.getOperand(1).getOpcode() == ISD::Constant) {
24311 SDValue N00 = N0.getOperand(0);
24312 APInt Mask = cast<ConstantSDNode>(N0.getOperand(1))->getAPIntValue();
24313 APInt ShAmt = N1C->getAPIntValue();
24314 Mask = Mask.shl(ShAmt);
24315 bool MaskOK = false;
24316 // We can handle cases concerning bit-widening nodes containing setcc_c if
24317 // we carefully interrogate the mask to make sure we are semantics
24319 // The transform is not safe if the result of C1 << C2 exceeds the bitwidth
24320 // of the underlying setcc_c operation if the setcc_c was zero extended.
24321 // Consider the following example:
24322 // zext(setcc_c) -> i32 0x0000FFFF
24323 // c1 -> i32 0x0000FFFF
24324 // c2 -> i32 0x00000001
24325 // (shl (and (setcc_c), c1), c2) -> i32 0x0001FFFE
24326 // (and setcc_c, (c1 << c2)) -> i32 0x0000FFFE
24327 if (N00.getOpcode() == X86ISD::SETCC_CARRY) {
24329 } else if (N00.getOpcode() == ISD::SIGN_EXTEND &&
24330 N00.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY) {
24332 } else if ((N00.getOpcode() == ISD::ZERO_EXTEND ||
24333 N00.getOpcode() == ISD::ANY_EXTEND) &&
24334 N00.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY) {
24335 MaskOK = Mask.isIntN(N00.getOperand(0).getValueSizeInBits());
24337 if (MaskOK && Mask != 0) {
24339 return DAG.getNode(ISD::AND, DL, VT, N00, DAG.getConstant(Mask, DL, VT));
24343 // Hardware support for vector shifts is sparse which makes us scalarize the
24344 // vector operations in many cases. Also, on sandybridge ADD is faster than
24346 // (shl V, 1) -> add V,V
24347 if (auto *N1BV = dyn_cast<BuildVectorSDNode>(N1))
24348 if (auto *N1SplatC = N1BV->getConstantSplatNode()) {
24349 assert(N0.getValueType().isVector() && "Invalid vector shift type");
24350 // We shift all of the values by one. In many cases we do not have
24351 // hardware support for this operation. This is better expressed as an ADD
24353 if (N1SplatC->getAPIntValue() == 1)
24354 return DAG.getNode(ISD::ADD, SDLoc(N), VT, N0, N0);
24360 /// \brief Returns a vector of 0s if the node in input is a vector logical
24361 /// shift by a constant amount which is known to be bigger than or equal
24362 /// to the vector element size in bits.
24363 static SDValue performShiftToAllZeros(SDNode *N, SelectionDAG &DAG,
24364 const X86Subtarget *Subtarget) {
24365 EVT VT = N->getValueType(0);
24367 if (VT != MVT::v2i64 && VT != MVT::v4i32 && VT != MVT::v8i16 &&
24368 (!Subtarget->hasInt256() ||
24369 (VT != MVT::v4i64 && VT != MVT::v8i32 && VT != MVT::v16i16)))
24372 SDValue Amt = N->getOperand(1);
24374 if (auto *AmtBV = dyn_cast<BuildVectorSDNode>(Amt))
24375 if (auto *AmtSplat = AmtBV->getConstantSplatNode()) {
24376 APInt ShiftAmt = AmtSplat->getAPIntValue();
24377 unsigned MaxAmount = VT.getVectorElementType().getSizeInBits();
24379 // SSE2/AVX2 logical shifts always return a vector of 0s
24380 // if the shift amount is bigger than or equal to
24381 // the element size. The constant shift amount will be
24382 // encoded as a 8-bit immediate.
24383 if (ShiftAmt.trunc(8).uge(MaxAmount))
24384 return getZeroVector(VT, Subtarget, DAG, DL);
24390 /// PerformShiftCombine - Combine shifts.
24391 static SDValue PerformShiftCombine(SDNode* N, SelectionDAG &DAG,
24392 TargetLowering::DAGCombinerInfo &DCI,
24393 const X86Subtarget *Subtarget) {
24394 if (N->getOpcode() == ISD::SHL)
24395 if (SDValue V = PerformSHLCombine(N, DAG))
24398 // Try to fold this logical shift into a zero vector.
24399 if (N->getOpcode() != ISD::SRA)
24400 if (SDValue V = performShiftToAllZeros(N, DAG, Subtarget))
24406 // CMPEQCombine - Recognize the distinctive (AND (setcc ...) (setcc ..))
24407 // where both setccs reference the same FP CMP, and rewrite for CMPEQSS
24408 // and friends. Likewise for OR -> CMPNEQSS.
24409 static SDValue CMPEQCombine(SDNode *N, SelectionDAG &DAG,
24410 TargetLowering::DAGCombinerInfo &DCI,
24411 const X86Subtarget *Subtarget) {
24414 // SSE1 supports CMP{eq|ne}SS, and SSE2 added CMP{eq|ne}SD, but
24415 // we're requiring SSE2 for both.
24416 if (Subtarget->hasSSE2() && isAndOrOfSetCCs(SDValue(N, 0U), opcode)) {
24417 SDValue N0 = N->getOperand(0);
24418 SDValue N1 = N->getOperand(1);
24419 SDValue CMP0 = N0->getOperand(1);
24420 SDValue CMP1 = N1->getOperand(1);
24423 // The SETCCs should both refer to the same CMP.
24424 if (CMP0.getOpcode() != X86ISD::CMP || CMP0 != CMP1)
24427 SDValue CMP00 = CMP0->getOperand(0);
24428 SDValue CMP01 = CMP0->getOperand(1);
24429 EVT VT = CMP00.getValueType();
24431 if (VT == MVT::f32 || VT == MVT::f64) {
24432 bool ExpectingFlags = false;
24433 // Check for any users that want flags:
24434 for (SDNode::use_iterator UI = N->use_begin(), UE = N->use_end();
24435 !ExpectingFlags && UI != UE; ++UI)
24436 switch (UI->getOpcode()) {
24441 ExpectingFlags = true;
24443 case ISD::CopyToReg:
24444 case ISD::SIGN_EXTEND:
24445 case ISD::ZERO_EXTEND:
24446 case ISD::ANY_EXTEND:
24450 if (!ExpectingFlags) {
24451 enum X86::CondCode cc0 = (enum X86::CondCode)N0.getConstantOperandVal(0);
24452 enum X86::CondCode cc1 = (enum X86::CondCode)N1.getConstantOperandVal(0);
24454 if (cc1 == X86::COND_E || cc1 == X86::COND_NE) {
24455 X86::CondCode tmp = cc0;
24460 if ((cc0 == X86::COND_E && cc1 == X86::COND_NP) ||
24461 (cc0 == X86::COND_NE && cc1 == X86::COND_P)) {
24462 // FIXME: need symbolic constants for these magic numbers.
24463 // See X86ATTInstPrinter.cpp:printSSECC().
24464 unsigned x86cc = (cc0 == X86::COND_E) ? 0 : 4;
24465 if (Subtarget->hasAVX512()) {
24466 SDValue FSetCC = DAG.getNode(X86ISD::FSETCC, DL, MVT::i1, CMP00,
24468 DAG.getConstant(x86cc, DL, MVT::i8));
24469 if (N->getValueType(0) != MVT::i1)
24470 return DAG.getNode(ISD::ZERO_EXTEND, DL, N->getValueType(0),
24474 SDValue OnesOrZeroesF = DAG.getNode(X86ISD::FSETCC, DL,
24475 CMP00.getValueType(), CMP00, CMP01,
24476 DAG.getConstant(x86cc, DL,
24479 bool is64BitFP = (CMP00.getValueType() == MVT::f64);
24480 MVT IntVT = is64BitFP ? MVT::i64 : MVT::i32;
24482 if (is64BitFP && !Subtarget->is64Bit()) {
24483 // On a 32-bit target, we cannot bitcast the 64-bit float to a
24484 // 64-bit integer, since that's not a legal type. Since
24485 // OnesOrZeroesF is all ones of all zeroes, we don't need all the
24486 // bits, but can do this little dance to extract the lowest 32 bits
24487 // and work with those going forward.
24488 SDValue Vector64 = DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, MVT::v2f64,
24490 SDValue Vector32 = DAG.getBitcast(MVT::v4f32, Vector64);
24491 OnesOrZeroesF = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::f32,
24492 Vector32, DAG.getIntPtrConstant(0, DL));
24496 SDValue OnesOrZeroesI = DAG.getBitcast(IntVT, OnesOrZeroesF);
24497 SDValue ANDed = DAG.getNode(ISD::AND, DL, IntVT, OnesOrZeroesI,
24498 DAG.getConstant(1, DL, IntVT));
24499 SDValue OneBitOfTruth = DAG.getNode(ISD::TRUNCATE, DL, MVT::i8,
24501 return OneBitOfTruth;
24509 /// CanFoldXORWithAllOnes - Test whether the XOR operand is a AllOnes vector
24510 /// so it can be folded inside ANDNP.
24511 static bool CanFoldXORWithAllOnes(const SDNode *N) {
24512 EVT VT = N->getValueType(0);
24514 // Match direct AllOnes for 128 and 256-bit vectors
24515 if (ISD::isBuildVectorAllOnes(N))
24518 // Look through a bit convert.
24519 if (N->getOpcode() == ISD::BITCAST)
24520 N = N->getOperand(0).getNode();
24522 // Sometimes the operand may come from a insert_subvector building a 256-bit
24524 if (VT.is256BitVector() &&
24525 N->getOpcode() == ISD::INSERT_SUBVECTOR) {
24526 SDValue V1 = N->getOperand(0);
24527 SDValue V2 = N->getOperand(1);
24529 if (V1.getOpcode() == ISD::INSERT_SUBVECTOR &&
24530 V1.getOperand(0).getOpcode() == ISD::UNDEF &&
24531 ISD::isBuildVectorAllOnes(V1.getOperand(1).getNode()) &&
24532 ISD::isBuildVectorAllOnes(V2.getNode()))
24539 // On AVX/AVX2 the type v8i1 is legalized to v8i16, which is an XMM sized
24540 // register. In most cases we actually compare or select YMM-sized registers
24541 // and mixing the two types creates horrible code. This method optimizes
24542 // some of the transition sequences.
24543 static SDValue WidenMaskArithmetic(SDNode *N, SelectionDAG &DAG,
24544 TargetLowering::DAGCombinerInfo &DCI,
24545 const X86Subtarget *Subtarget) {
24546 EVT VT = N->getValueType(0);
24547 if (!VT.is256BitVector())
24550 assert((N->getOpcode() == ISD::ANY_EXTEND ||
24551 N->getOpcode() == ISD::ZERO_EXTEND ||
24552 N->getOpcode() == ISD::SIGN_EXTEND) && "Invalid Node");
24554 SDValue Narrow = N->getOperand(0);
24555 EVT NarrowVT = Narrow->getValueType(0);
24556 if (!NarrowVT.is128BitVector())
24559 if (Narrow->getOpcode() != ISD::XOR &&
24560 Narrow->getOpcode() != ISD::AND &&
24561 Narrow->getOpcode() != ISD::OR)
24564 SDValue N0 = Narrow->getOperand(0);
24565 SDValue N1 = Narrow->getOperand(1);
24568 // The Left side has to be a trunc.
24569 if (N0.getOpcode() != ISD::TRUNCATE)
24572 // The type of the truncated inputs.
24573 EVT WideVT = N0->getOperand(0)->getValueType(0);
24577 // The right side has to be a 'trunc' or a constant vector.
24578 bool RHSTrunc = N1.getOpcode() == ISD::TRUNCATE;
24579 ConstantSDNode *RHSConstSplat = nullptr;
24580 if (auto *RHSBV = dyn_cast<BuildVectorSDNode>(N1))
24581 RHSConstSplat = RHSBV->getConstantSplatNode();
24582 if (!RHSTrunc && !RHSConstSplat)
24585 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
24587 if (!TLI.isOperationLegalOrPromote(Narrow->getOpcode(), WideVT))
24590 // Set N0 and N1 to hold the inputs to the new wide operation.
24591 N0 = N0->getOperand(0);
24592 if (RHSConstSplat) {
24593 N1 = DAG.getNode(ISD::ZERO_EXTEND, DL, WideVT.getScalarType(),
24594 SDValue(RHSConstSplat, 0));
24595 SmallVector<SDValue, 8> C(WideVT.getVectorNumElements(), N1);
24596 N1 = DAG.getNode(ISD::BUILD_VECTOR, DL, WideVT, C);
24597 } else if (RHSTrunc) {
24598 N1 = N1->getOperand(0);
24601 // Generate the wide operation.
24602 SDValue Op = DAG.getNode(Narrow->getOpcode(), DL, WideVT, N0, N1);
24603 unsigned Opcode = N->getOpcode();
24605 case ISD::ANY_EXTEND:
24607 case ISD::ZERO_EXTEND: {
24608 unsigned InBits = NarrowVT.getScalarType().getSizeInBits();
24609 APInt Mask = APInt::getAllOnesValue(InBits);
24610 Mask = Mask.zext(VT.getScalarType().getSizeInBits());
24611 return DAG.getNode(ISD::AND, DL, VT,
24612 Op, DAG.getConstant(Mask, DL, VT));
24614 case ISD::SIGN_EXTEND:
24615 return DAG.getNode(ISD::SIGN_EXTEND_INREG, DL, VT,
24616 Op, DAG.getValueType(NarrowVT));
24618 llvm_unreachable("Unexpected opcode");
24622 static SDValue VectorZextCombine(SDNode *N, SelectionDAG &DAG,
24623 TargetLowering::DAGCombinerInfo &DCI,
24624 const X86Subtarget *Subtarget) {
24625 SDValue N0 = N->getOperand(0);
24626 SDValue N1 = N->getOperand(1);
24629 // A vector zext_in_reg may be represented as a shuffle,
24630 // feeding into a bitcast (this represents anyext) feeding into
24631 // an and with a mask.
24632 // We'd like to try to combine that into a shuffle with zero
24633 // plus a bitcast, removing the and.
24634 if (N0.getOpcode() != ISD::BITCAST ||
24635 N0.getOperand(0).getOpcode() != ISD::VECTOR_SHUFFLE)
24638 // The other side of the AND should be a splat of 2^C, where C
24639 // is the number of bits in the source type.
24640 if (N1.getOpcode() == ISD::BITCAST)
24641 N1 = N1.getOperand(0);
24642 if (N1.getOpcode() != ISD::BUILD_VECTOR)
24644 BuildVectorSDNode *Vector = cast<BuildVectorSDNode>(N1);
24646 ShuffleVectorSDNode *Shuffle = cast<ShuffleVectorSDNode>(N0.getOperand(0));
24647 EVT SrcType = Shuffle->getValueType(0);
24649 // We expect a single-source shuffle
24650 if (Shuffle->getOperand(1)->getOpcode() != ISD::UNDEF)
24653 unsigned SrcSize = SrcType.getScalarSizeInBits();
24655 APInt SplatValue, SplatUndef;
24656 unsigned SplatBitSize;
24658 if (!Vector->isConstantSplat(SplatValue, SplatUndef,
24659 SplatBitSize, HasAnyUndefs))
24662 unsigned ResSize = N1.getValueType().getScalarSizeInBits();
24663 // Make sure the splat matches the mask we expect
24664 if (SplatBitSize > ResSize ||
24665 (SplatValue + 1).exactLogBase2() != (int)SrcSize)
24668 // Make sure the input and output size make sense
24669 if (SrcSize >= ResSize || ResSize % SrcSize)
24672 // We expect a shuffle of the form <0, u, u, u, 1, u, u, u...>
24673 // The number of u's between each two values depends on the ratio between
24674 // the source and dest type.
24675 unsigned ZextRatio = ResSize / SrcSize;
24676 bool IsZext = true;
24677 for (unsigned i = 0; i < SrcType.getVectorNumElements(); ++i) {
24678 if (i % ZextRatio) {
24679 if (Shuffle->getMaskElt(i) > 0) {
24685 if (Shuffle->getMaskElt(i) != (int)(i / ZextRatio)) {
24686 // Expected element number
24696 // Ok, perform the transformation - replace the shuffle with
24697 // a shuffle of the form <0, k, k, k, 1, k, k, k> with zero
24698 // (instead of undef) where the k elements come from the zero vector.
24699 SmallVector<int, 8> Mask;
24700 unsigned NumElems = SrcType.getVectorNumElements();
24701 for (unsigned i = 0; i < NumElems; ++i)
24703 Mask.push_back(NumElems);
24705 Mask.push_back(i / ZextRatio);
24707 SDValue NewShuffle = DAG.getVectorShuffle(Shuffle->getValueType(0), DL,
24708 Shuffle->getOperand(0), DAG.getConstant(0, DL, SrcType), Mask);
24709 return DAG.getBitcast(N0.getValueType(), NewShuffle);
24712 /// If both input operands of a logic op are being cast from floating point
24713 /// types, try to convert this into a floating point logic node to avoid
24714 /// unnecessary moves from SSE to integer registers.
24715 static SDValue convertIntLogicToFPLogic(SDNode *N, SelectionDAG &DAG,
24716 const X86Subtarget *Subtarget) {
24717 unsigned FPOpcode = ISD::DELETED_NODE;
24718 if (N->getOpcode() == ISD::AND)
24719 FPOpcode = X86ISD::FAND;
24720 else if (N->getOpcode() == ISD::OR)
24721 FPOpcode = X86ISD::FOR;
24722 else if (N->getOpcode() == ISD::XOR)
24723 FPOpcode = X86ISD::FXOR;
24725 assert(FPOpcode != ISD::DELETED_NODE &&
24726 "Unexpected input node for FP logic conversion");
24728 EVT VT = N->getValueType(0);
24729 SDValue N0 = N->getOperand(0);
24730 SDValue N1 = N->getOperand(1);
24732 if (N0.getOpcode() == ISD::BITCAST && N1.getOpcode() == ISD::BITCAST &&
24733 ((Subtarget->hasSSE1() && VT == MVT::i32) ||
24734 (Subtarget->hasSSE2() && VT == MVT::i64))) {
24735 SDValue N00 = N0.getOperand(0);
24736 SDValue N10 = N1.getOperand(0);
24737 EVT N00Type = N00.getValueType();
24738 EVT N10Type = N10.getValueType();
24739 if (N00Type.isFloatingPoint() && N10Type.isFloatingPoint()) {
24740 SDValue FPLogic = DAG.getNode(FPOpcode, DL, N00Type, N00, N10);
24741 return DAG.getBitcast(VT, FPLogic);
24747 static SDValue PerformAndCombine(SDNode *N, SelectionDAG &DAG,
24748 TargetLowering::DAGCombinerInfo &DCI,
24749 const X86Subtarget *Subtarget) {
24750 if (DCI.isBeforeLegalizeOps())
24753 if (SDValue Zext = VectorZextCombine(N, DAG, DCI, Subtarget))
24756 if (SDValue R = CMPEQCombine(N, DAG, DCI, Subtarget))
24759 if (SDValue FPLogic = convertIntLogicToFPLogic(N, DAG, Subtarget))
24762 EVT VT = N->getValueType(0);
24763 SDValue N0 = N->getOperand(0);
24764 SDValue N1 = N->getOperand(1);
24767 // Create BEXTR instructions
24768 // BEXTR is ((X >> imm) & (2**size-1))
24769 if (VT == MVT::i32 || VT == MVT::i64) {
24770 // Check for BEXTR.
24771 if ((Subtarget->hasBMI() || Subtarget->hasTBM()) &&
24772 (N0.getOpcode() == ISD::SRA || N0.getOpcode() == ISD::SRL)) {
24773 ConstantSDNode *MaskNode = dyn_cast<ConstantSDNode>(N1);
24774 ConstantSDNode *ShiftNode = dyn_cast<ConstantSDNode>(N0.getOperand(1));
24775 if (MaskNode && ShiftNode) {
24776 uint64_t Mask = MaskNode->getZExtValue();
24777 uint64_t Shift = ShiftNode->getZExtValue();
24778 if (isMask_64(Mask)) {
24779 uint64_t MaskSize = countPopulation(Mask);
24780 if (Shift + MaskSize <= VT.getSizeInBits())
24781 return DAG.getNode(X86ISD::BEXTR, DL, VT, N0.getOperand(0),
24782 DAG.getConstant(Shift | (MaskSize << 8), DL,
24791 // Want to form ANDNP nodes:
24792 // 1) In the hopes of then easily combining them with OR and AND nodes
24793 // to form PBLEND/PSIGN.
24794 // 2) To match ANDN packed intrinsics
24795 if (VT != MVT::v2i64 && VT != MVT::v4i64)
24798 // Check LHS for vnot
24799 if (N0.getOpcode() == ISD::XOR &&
24800 //ISD::isBuildVectorAllOnes(N0.getOperand(1).getNode()))
24801 CanFoldXORWithAllOnes(N0.getOperand(1).getNode()))
24802 return DAG.getNode(X86ISD::ANDNP, DL, VT, N0.getOperand(0), N1);
24804 // Check RHS for vnot
24805 if (N1.getOpcode() == ISD::XOR &&
24806 //ISD::isBuildVectorAllOnes(N1.getOperand(1).getNode()))
24807 CanFoldXORWithAllOnes(N1.getOperand(1).getNode()))
24808 return DAG.getNode(X86ISD::ANDNP, DL, VT, N1.getOperand(0), N0);
24813 static SDValue PerformOrCombine(SDNode *N, SelectionDAG &DAG,
24814 TargetLowering::DAGCombinerInfo &DCI,
24815 const X86Subtarget *Subtarget) {
24816 if (DCI.isBeforeLegalizeOps())
24819 if (SDValue R = CMPEQCombine(N, DAG, DCI, Subtarget))
24822 if (SDValue FPLogic = convertIntLogicToFPLogic(N, DAG, Subtarget))
24825 SDValue N0 = N->getOperand(0);
24826 SDValue N1 = N->getOperand(1);
24827 EVT VT = N->getValueType(0);
24829 // look for psign/blend
24830 if (VT == MVT::v2i64 || VT == MVT::v4i64) {
24831 if (!Subtarget->hasSSSE3() ||
24832 (VT == MVT::v4i64 && !Subtarget->hasInt256()))
24835 // Canonicalize pandn to RHS
24836 if (N0.getOpcode() == X86ISD::ANDNP)
24838 // or (and (m, y), (pandn m, x))
24839 if (N0.getOpcode() == ISD::AND && N1.getOpcode() == X86ISD::ANDNP) {
24840 SDValue Mask = N1.getOperand(0);
24841 SDValue X = N1.getOperand(1);
24843 if (N0.getOperand(0) == Mask)
24844 Y = N0.getOperand(1);
24845 if (N0.getOperand(1) == Mask)
24846 Y = N0.getOperand(0);
24848 // Check to see if the mask appeared in both the AND and ANDNP and
24852 // Validate that X, Y, and Mask are BIT_CONVERTS, and see through them.
24853 // Look through mask bitcast.
24854 if (Mask.getOpcode() == ISD::BITCAST)
24855 Mask = Mask.getOperand(0);
24856 if (X.getOpcode() == ISD::BITCAST)
24857 X = X.getOperand(0);
24858 if (Y.getOpcode() == ISD::BITCAST)
24859 Y = Y.getOperand(0);
24861 EVT MaskVT = Mask.getValueType();
24863 // Validate that the Mask operand is a vector sra node.
24864 // FIXME: what to do for bytes, since there is a psignb/pblendvb, but
24865 // there is no psrai.b
24866 unsigned EltBits = MaskVT.getVectorElementType().getSizeInBits();
24867 unsigned SraAmt = ~0;
24868 if (Mask.getOpcode() == ISD::SRA) {
24869 if (auto *AmtBV = dyn_cast<BuildVectorSDNode>(Mask.getOperand(1)))
24870 if (auto *AmtConst = AmtBV->getConstantSplatNode())
24871 SraAmt = AmtConst->getZExtValue();
24872 } else if (Mask.getOpcode() == X86ISD::VSRAI) {
24873 SDValue SraC = Mask.getOperand(1);
24874 SraAmt = cast<ConstantSDNode>(SraC)->getZExtValue();
24876 if ((SraAmt + 1) != EltBits)
24881 // Now we know we at least have a plendvb with the mask val. See if
24882 // we can form a psignb/w/d.
24883 // psign = x.type == y.type == mask.type && y = sub(0, x);
24884 if (Y.getOpcode() == ISD::SUB && Y.getOperand(1) == X &&
24885 ISD::isBuildVectorAllZeros(Y.getOperand(0).getNode()) &&
24886 X.getValueType() == MaskVT && Y.getValueType() == MaskVT) {
24887 assert((EltBits == 8 || EltBits == 16 || EltBits == 32) &&
24888 "Unsupported VT for PSIGN");
24889 Mask = DAG.getNode(X86ISD::PSIGN, DL, MaskVT, X, Mask.getOperand(0));
24890 return DAG.getBitcast(VT, Mask);
24892 // PBLENDVB only available on SSE 4.1
24893 if (!Subtarget->hasSSE41())
24896 EVT BlendVT = (VT == MVT::v4i64) ? MVT::v32i8 : MVT::v16i8;
24898 X = DAG.getBitcast(BlendVT, X);
24899 Y = DAG.getBitcast(BlendVT, Y);
24900 Mask = DAG.getBitcast(BlendVT, Mask);
24901 Mask = DAG.getNode(ISD::VSELECT, DL, BlendVT, Mask, Y, X);
24902 return DAG.getBitcast(VT, Mask);
24906 if (VT != MVT::i16 && VT != MVT::i32 && VT != MVT::i64)
24909 // fold (or (x << c) | (y >> (64 - c))) ==> (shld64 x, y, c)
24910 bool OptForSize = DAG.getMachineFunction().getFunction()->optForSize();
24912 // SHLD/SHRD instructions have lower register pressure, but on some
24913 // platforms they have higher latency than the equivalent
24914 // series of shifts/or that would otherwise be generated.
24915 // Don't fold (or (x << c) | (y >> (64 - c))) if SHLD/SHRD instructions
24916 // have higher latencies and we are not optimizing for size.
24917 if (!OptForSize && Subtarget->isSHLDSlow())
24920 if (N0.getOpcode() == ISD::SRL && N1.getOpcode() == ISD::SHL)
24922 if (N0.getOpcode() != ISD::SHL || N1.getOpcode() != ISD::SRL)
24924 if (!N0.hasOneUse() || !N1.hasOneUse())
24927 SDValue ShAmt0 = N0.getOperand(1);
24928 if (ShAmt0.getValueType() != MVT::i8)
24930 SDValue ShAmt1 = N1.getOperand(1);
24931 if (ShAmt1.getValueType() != MVT::i8)
24933 if (ShAmt0.getOpcode() == ISD::TRUNCATE)
24934 ShAmt0 = ShAmt0.getOperand(0);
24935 if (ShAmt1.getOpcode() == ISD::TRUNCATE)
24936 ShAmt1 = ShAmt1.getOperand(0);
24939 unsigned Opc = X86ISD::SHLD;
24940 SDValue Op0 = N0.getOperand(0);
24941 SDValue Op1 = N1.getOperand(0);
24942 if (ShAmt0.getOpcode() == ISD::SUB) {
24943 Opc = X86ISD::SHRD;
24944 std::swap(Op0, Op1);
24945 std::swap(ShAmt0, ShAmt1);
24948 unsigned Bits = VT.getSizeInBits();
24949 if (ShAmt1.getOpcode() == ISD::SUB) {
24950 SDValue Sum = ShAmt1.getOperand(0);
24951 if (ConstantSDNode *SumC = dyn_cast<ConstantSDNode>(Sum)) {
24952 SDValue ShAmt1Op1 = ShAmt1.getOperand(1);
24953 if (ShAmt1Op1.getNode()->getOpcode() == ISD::TRUNCATE)
24954 ShAmt1Op1 = ShAmt1Op1.getOperand(0);
24955 if (SumC->getSExtValue() == Bits && ShAmt1Op1 == ShAmt0)
24956 return DAG.getNode(Opc, DL, VT,
24958 DAG.getNode(ISD::TRUNCATE, DL,
24961 } else if (ConstantSDNode *ShAmt1C = dyn_cast<ConstantSDNode>(ShAmt1)) {
24962 ConstantSDNode *ShAmt0C = dyn_cast<ConstantSDNode>(ShAmt0);
24964 ShAmt0C->getSExtValue() + ShAmt1C->getSExtValue() == Bits)
24965 return DAG.getNode(Opc, DL, VT,
24966 N0.getOperand(0), N1.getOperand(0),
24967 DAG.getNode(ISD::TRUNCATE, DL,
24974 // Generate NEG and CMOV for integer abs.
24975 static SDValue performIntegerAbsCombine(SDNode *N, SelectionDAG &DAG) {
24976 EVT VT = N->getValueType(0);
24978 // Since X86 does not have CMOV for 8-bit integer, we don't convert
24979 // 8-bit integer abs to NEG and CMOV.
24980 if (VT.isInteger() && VT.getSizeInBits() == 8)
24983 SDValue N0 = N->getOperand(0);
24984 SDValue N1 = N->getOperand(1);
24987 // Check pattern of XOR(ADD(X,Y), Y) where Y is SRA(X, size(X)-1)
24988 // and change it to SUB and CMOV.
24989 if (VT.isInteger() && N->getOpcode() == ISD::XOR &&
24990 N0.getOpcode() == ISD::ADD &&
24991 N0.getOperand(1) == N1 &&
24992 N1.getOpcode() == ISD::SRA &&
24993 N1.getOperand(0) == N0.getOperand(0))
24994 if (ConstantSDNode *Y1C = dyn_cast<ConstantSDNode>(N1.getOperand(1)))
24995 if (Y1C->getAPIntValue() == VT.getSizeInBits()-1) {
24996 // Generate SUB & CMOV.
24997 SDValue Neg = DAG.getNode(X86ISD::SUB, DL, DAG.getVTList(VT, MVT::i32),
24998 DAG.getConstant(0, DL, VT), N0.getOperand(0));
25000 SDValue Ops[] = { N0.getOperand(0), Neg,
25001 DAG.getConstant(X86::COND_GE, DL, MVT::i8),
25002 SDValue(Neg.getNode(), 1) };
25003 return DAG.getNode(X86ISD::CMOV, DL, DAG.getVTList(VT, MVT::Glue), Ops);
25008 // Try to turn tests against the signbit in the form of:
25009 // XOR(TRUNCATE(SRL(X, size(X)-1)), 1)
25012 static SDValue foldXorTruncShiftIntoCmp(SDNode *N, SelectionDAG &DAG) {
25013 // This is only worth doing if the output type is i8.
25014 if (N->getValueType(0) != MVT::i8)
25017 SDValue N0 = N->getOperand(0);
25018 SDValue N1 = N->getOperand(1);
25020 // We should be performing an xor against a truncated shift.
25021 if (N0.getOpcode() != ISD::TRUNCATE || !N0.hasOneUse())
25024 // Make sure we are performing an xor against one.
25025 if (!isa<ConstantSDNode>(N1) || !cast<ConstantSDNode>(N1)->isOne())
25028 // SetCC on x86 zero extends so only act on this if it's a logical shift.
25029 SDValue Shift = N0.getOperand(0);
25030 if (Shift.getOpcode() != ISD::SRL || !Shift.hasOneUse())
25033 // Make sure we are truncating from one of i16, i32 or i64.
25034 EVT ShiftTy = Shift.getValueType();
25035 if (ShiftTy != MVT::i16 && ShiftTy != MVT::i32 && ShiftTy != MVT::i64)
25038 // Make sure the shift amount extracts the sign bit.
25039 if (!isa<ConstantSDNode>(Shift.getOperand(1)) ||
25040 Shift.getConstantOperandVal(1) != ShiftTy.getSizeInBits() - 1)
25043 // Create a greater-than comparison against -1.
25044 // N.B. Using SETGE against 0 works but we want a canonical looking
25045 // comparison, using SETGT matches up with what TranslateX86CC.
25047 SDValue ShiftOp = Shift.getOperand(0);
25048 EVT ShiftOpTy = ShiftOp.getValueType();
25049 SDValue Cond = DAG.getSetCC(DL, MVT::i8, ShiftOp,
25050 DAG.getConstant(-1, DL, ShiftOpTy), ISD::SETGT);
25054 static SDValue PerformXorCombine(SDNode *N, SelectionDAG &DAG,
25055 TargetLowering::DAGCombinerInfo &DCI,
25056 const X86Subtarget *Subtarget) {
25057 if (DCI.isBeforeLegalizeOps())
25060 if (SDValue RV = foldXorTruncShiftIntoCmp(N, DAG))
25063 if (Subtarget->hasCMov())
25064 if (SDValue RV = performIntegerAbsCombine(N, DAG))
25067 if (SDValue FPLogic = convertIntLogicToFPLogic(N, DAG, Subtarget))
25073 /// PerformLOADCombine - Do target-specific dag combines on LOAD nodes.
25074 static SDValue PerformLOADCombine(SDNode *N, SelectionDAG &DAG,
25075 TargetLowering::DAGCombinerInfo &DCI,
25076 const X86Subtarget *Subtarget) {
25077 LoadSDNode *Ld = cast<LoadSDNode>(N);
25078 EVT RegVT = Ld->getValueType(0);
25079 EVT MemVT = Ld->getMemoryVT();
25081 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
25083 // For chips with slow 32-byte unaligned loads, break the 32-byte operation
25084 // into two 16-byte operations.
25085 ISD::LoadExtType Ext = Ld->getExtensionType();
25087 unsigned AddressSpace = Ld->getAddressSpace();
25088 unsigned Alignment = Ld->getAlignment();
25089 if (RegVT.is256BitVector() && !DCI.isBeforeLegalizeOps() &&
25090 Ext == ISD::NON_EXTLOAD &&
25091 TLI.allowsMemoryAccess(*DAG.getContext(), DAG.getDataLayout(), RegVT,
25092 AddressSpace, Alignment, &Fast) && !Fast) {
25093 unsigned NumElems = RegVT.getVectorNumElements();
25097 SDValue Ptr = Ld->getBasePtr();
25098 SDValue Increment =
25099 DAG.getConstant(16, dl, TLI.getPointerTy(DAG.getDataLayout()));
25101 EVT HalfVT = EVT::getVectorVT(*DAG.getContext(), MemVT.getScalarType(),
25103 SDValue Load1 = DAG.getLoad(HalfVT, dl, Ld->getChain(), Ptr,
25104 Ld->getPointerInfo(), Ld->isVolatile(),
25105 Ld->isNonTemporal(), Ld->isInvariant(),
25107 Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr, Increment);
25108 SDValue Load2 = DAG.getLoad(HalfVT, dl, Ld->getChain(), Ptr,
25109 Ld->getPointerInfo(), Ld->isVolatile(),
25110 Ld->isNonTemporal(), Ld->isInvariant(),
25111 std::min(16U, Alignment));
25112 SDValue TF = DAG.getNode(ISD::TokenFactor, dl, MVT::Other,
25114 Load2.getValue(1));
25116 SDValue NewVec = DAG.getUNDEF(RegVT);
25117 NewVec = Insert128BitVector(NewVec, Load1, 0, DAG, dl);
25118 NewVec = Insert128BitVector(NewVec, Load2, NumElems/2, DAG, dl);
25119 return DCI.CombineTo(N, NewVec, TF, true);
25125 /// PerformMLOADCombine - Resolve extending loads
25126 static SDValue PerformMLOADCombine(SDNode *N, SelectionDAG &DAG,
25127 TargetLowering::DAGCombinerInfo &DCI,
25128 const X86Subtarget *Subtarget) {
25129 MaskedLoadSDNode *Mld = cast<MaskedLoadSDNode>(N);
25130 if (Mld->getExtensionType() != ISD::SEXTLOAD)
25133 EVT VT = Mld->getValueType(0);
25134 unsigned NumElems = VT.getVectorNumElements();
25135 EVT LdVT = Mld->getMemoryVT();
25138 assert(LdVT != VT && "Cannot extend to the same type");
25139 unsigned ToSz = VT.getVectorElementType().getSizeInBits();
25140 unsigned FromSz = LdVT.getVectorElementType().getSizeInBits();
25141 // From, To sizes and ElemCount must be pow of two
25142 assert (isPowerOf2_32(NumElems * FromSz * ToSz) &&
25143 "Unexpected size for extending masked load");
25145 unsigned SizeRatio = ToSz / FromSz;
25146 assert(SizeRatio * NumElems * FromSz == VT.getSizeInBits());
25148 // Create a type on which we perform the shuffle
25149 EVT WideVecVT = EVT::getVectorVT(*DAG.getContext(),
25150 LdVT.getScalarType(), NumElems*SizeRatio);
25151 assert(WideVecVT.getSizeInBits() == VT.getSizeInBits());
25153 // Convert Src0 value
25154 SDValue WideSrc0 = DAG.getBitcast(WideVecVT, Mld->getSrc0());
25155 if (Mld->getSrc0().getOpcode() != ISD::UNDEF) {
25156 SmallVector<int, 16> ShuffleVec(NumElems * SizeRatio, -1);
25157 for (unsigned i = 0; i != NumElems; ++i)
25158 ShuffleVec[i] = i * SizeRatio;
25160 // Can't shuffle using an illegal type.
25161 assert(DAG.getTargetLoweringInfo().isTypeLegal(WideVecVT) &&
25162 "WideVecVT should be legal");
25163 WideSrc0 = DAG.getVectorShuffle(WideVecVT, dl, WideSrc0,
25164 DAG.getUNDEF(WideVecVT), &ShuffleVec[0]);
25166 // Prepare the new mask
25168 SDValue Mask = Mld->getMask();
25169 if (Mask.getValueType() == VT) {
25170 // Mask and original value have the same type
25171 NewMask = DAG.getBitcast(WideVecVT, Mask);
25172 SmallVector<int, 16> ShuffleVec(NumElems * SizeRatio, -1);
25173 for (unsigned i = 0; i != NumElems; ++i)
25174 ShuffleVec[i] = i * SizeRatio;
25175 for (unsigned i = NumElems; i != NumElems*SizeRatio; ++i)
25176 ShuffleVec[i] = NumElems*SizeRatio;
25177 NewMask = DAG.getVectorShuffle(WideVecVT, dl, NewMask,
25178 DAG.getConstant(0, dl, WideVecVT),
25182 assert(Mask.getValueType().getVectorElementType() == MVT::i1);
25183 unsigned WidenNumElts = NumElems*SizeRatio;
25184 unsigned MaskNumElts = VT.getVectorNumElements();
25185 EVT NewMaskVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
25188 unsigned NumConcat = WidenNumElts / MaskNumElts;
25189 SmallVector<SDValue, 16> Ops(NumConcat);
25190 SDValue ZeroVal = DAG.getConstant(0, dl, Mask.getValueType());
25192 for (unsigned i = 1; i != NumConcat; ++i)
25195 NewMask = DAG.getNode(ISD::CONCAT_VECTORS, dl, NewMaskVT, Ops);
25198 SDValue WideLd = DAG.getMaskedLoad(WideVecVT, dl, Mld->getChain(),
25199 Mld->getBasePtr(), NewMask, WideSrc0,
25200 Mld->getMemoryVT(), Mld->getMemOperand(),
25202 SDValue NewVec = DAG.getNode(X86ISD::VSEXT, dl, VT, WideLd);
25203 return DCI.CombineTo(N, NewVec, WideLd.getValue(1), true);
25205 /// PerformMSTORECombine - Resolve truncating stores
25206 static SDValue PerformMSTORECombine(SDNode *N, SelectionDAG &DAG,
25207 const X86Subtarget *Subtarget) {
25208 MaskedStoreSDNode *Mst = cast<MaskedStoreSDNode>(N);
25209 if (!Mst->isTruncatingStore())
25212 EVT VT = Mst->getValue().getValueType();
25213 unsigned NumElems = VT.getVectorNumElements();
25214 EVT StVT = Mst->getMemoryVT();
25217 assert(StVT != VT && "Cannot truncate to the same type");
25218 unsigned FromSz = VT.getVectorElementType().getSizeInBits();
25219 unsigned ToSz = StVT.getVectorElementType().getSizeInBits();
25221 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
25223 // The truncating store is legal in some cases. For example
25224 // vpmovqb, vpmovqw, vpmovqd, vpmovdb, vpmovdw
25225 // are designated for truncate store.
25226 // In this case we don't need any further transformations.
25227 if (TLI.isTruncStoreLegal(VT, StVT))
25230 // From, To sizes and ElemCount must be pow of two
25231 assert (isPowerOf2_32(NumElems * FromSz * ToSz) &&
25232 "Unexpected size for truncating masked store");
25233 // We are going to use the original vector elt for storing.
25234 // Accumulated smaller vector elements must be a multiple of the store size.
25235 assert (((NumElems * FromSz) % ToSz) == 0 &&
25236 "Unexpected ratio for truncating masked store");
25238 unsigned SizeRatio = FromSz / ToSz;
25239 assert(SizeRatio * NumElems * ToSz == VT.getSizeInBits());
25241 // Create a type on which we perform the shuffle
25242 EVT WideVecVT = EVT::getVectorVT(*DAG.getContext(),
25243 StVT.getScalarType(), NumElems*SizeRatio);
25245 assert(WideVecVT.getSizeInBits() == VT.getSizeInBits());
25247 SDValue WideVec = DAG.getBitcast(WideVecVT, Mst->getValue());
25248 SmallVector<int, 16> ShuffleVec(NumElems * SizeRatio, -1);
25249 for (unsigned i = 0; i != NumElems; ++i)
25250 ShuffleVec[i] = i * SizeRatio;
25252 // Can't shuffle using an illegal type.
25253 assert(DAG.getTargetLoweringInfo().isTypeLegal(WideVecVT) &&
25254 "WideVecVT should be legal");
25256 SDValue TruncatedVal = DAG.getVectorShuffle(WideVecVT, dl, WideVec,
25257 DAG.getUNDEF(WideVecVT),
25261 SDValue Mask = Mst->getMask();
25262 if (Mask.getValueType() == VT) {
25263 // Mask and original value have the same type
25264 NewMask = DAG.getBitcast(WideVecVT, Mask);
25265 for (unsigned i = 0; i != NumElems; ++i)
25266 ShuffleVec[i] = i * SizeRatio;
25267 for (unsigned i = NumElems; i != NumElems*SizeRatio; ++i)
25268 ShuffleVec[i] = NumElems*SizeRatio;
25269 NewMask = DAG.getVectorShuffle(WideVecVT, dl, NewMask,
25270 DAG.getConstant(0, dl, WideVecVT),
25274 assert(Mask.getValueType().getVectorElementType() == MVT::i1);
25275 unsigned WidenNumElts = NumElems*SizeRatio;
25276 unsigned MaskNumElts = VT.getVectorNumElements();
25277 EVT NewMaskVT = EVT::getVectorVT(*DAG.getContext(), MVT::i1,
25280 unsigned NumConcat = WidenNumElts / MaskNumElts;
25281 SmallVector<SDValue, 16> Ops(NumConcat);
25282 SDValue ZeroVal = DAG.getConstant(0, dl, Mask.getValueType());
25284 for (unsigned i = 1; i != NumConcat; ++i)
25287 NewMask = DAG.getNode(ISD::CONCAT_VECTORS, dl, NewMaskVT, Ops);
25290 return DAG.getMaskedStore(Mst->getChain(), dl, TruncatedVal, Mst->getBasePtr(),
25291 NewMask, StVT, Mst->getMemOperand(), false);
25293 /// PerformSTORECombine - Do target-specific dag combines on STORE nodes.
25294 static SDValue PerformSTORECombine(SDNode *N, SelectionDAG &DAG,
25295 const X86Subtarget *Subtarget) {
25296 StoreSDNode *St = cast<StoreSDNode>(N);
25297 EVT VT = St->getValue().getValueType();
25298 EVT StVT = St->getMemoryVT();
25300 SDValue StoredVal = St->getOperand(1);
25301 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
25303 // If we are saving a concatenation of two XMM registers and 32-byte stores
25304 // are slow, such as on Sandy Bridge, perform two 16-byte stores.
25306 unsigned AddressSpace = St->getAddressSpace();
25307 unsigned Alignment = St->getAlignment();
25308 if (VT.is256BitVector() && StVT == VT &&
25309 TLI.allowsMemoryAccess(*DAG.getContext(), DAG.getDataLayout(), VT,
25310 AddressSpace, Alignment, &Fast) && !Fast) {
25311 unsigned NumElems = VT.getVectorNumElements();
25315 SDValue Value0 = Extract128BitVector(StoredVal, 0, DAG, dl);
25316 SDValue Value1 = Extract128BitVector(StoredVal, NumElems/2, DAG, dl);
25319 DAG.getConstant(16, dl, TLI.getPointerTy(DAG.getDataLayout()));
25320 SDValue Ptr0 = St->getBasePtr();
25321 SDValue Ptr1 = DAG.getNode(ISD::ADD, dl, Ptr0.getValueType(), Ptr0, Stride);
25323 SDValue Ch0 = DAG.getStore(St->getChain(), dl, Value0, Ptr0,
25324 St->getPointerInfo(), St->isVolatile(),
25325 St->isNonTemporal(), Alignment);
25326 SDValue Ch1 = DAG.getStore(St->getChain(), dl, Value1, Ptr1,
25327 St->getPointerInfo(), St->isVolatile(),
25328 St->isNonTemporal(),
25329 std::min(16U, Alignment));
25330 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Ch0, Ch1);
25333 // Optimize trunc store (of multiple scalars) to shuffle and store.
25334 // First, pack all of the elements in one place. Next, store to memory
25335 // in fewer chunks.
25336 if (St->isTruncatingStore() && VT.isVector()) {
25337 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
25338 unsigned NumElems = VT.getVectorNumElements();
25339 assert(StVT != VT && "Cannot truncate to the same type");
25340 unsigned FromSz = VT.getVectorElementType().getSizeInBits();
25341 unsigned ToSz = StVT.getVectorElementType().getSizeInBits();
25343 // The truncating store is legal in some cases. For example
25344 // vpmovqb, vpmovqw, vpmovqd, vpmovdb, vpmovdw
25345 // are designated for truncate store.
25346 // In this case we don't need any further transformations.
25347 if (TLI.isTruncStoreLegal(VT, StVT))
25350 // From, To sizes and ElemCount must be pow of two
25351 if (!isPowerOf2_32(NumElems * FromSz * ToSz)) return SDValue();
25352 // We are going to use the original vector elt for storing.
25353 // Accumulated smaller vector elements must be a multiple of the store size.
25354 if (0 != (NumElems * FromSz) % ToSz) return SDValue();
25356 unsigned SizeRatio = FromSz / ToSz;
25358 assert(SizeRatio * NumElems * ToSz == VT.getSizeInBits());
25360 // Create a type on which we perform the shuffle
25361 EVT WideVecVT = EVT::getVectorVT(*DAG.getContext(),
25362 StVT.getScalarType(), NumElems*SizeRatio);
25364 assert(WideVecVT.getSizeInBits() == VT.getSizeInBits());
25366 SDValue WideVec = DAG.getBitcast(WideVecVT, St->getValue());
25367 SmallVector<int, 8> ShuffleVec(NumElems * SizeRatio, -1);
25368 for (unsigned i = 0; i != NumElems; ++i)
25369 ShuffleVec[i] = i * SizeRatio;
25371 // Can't shuffle using an illegal type.
25372 if (!TLI.isTypeLegal(WideVecVT))
25375 SDValue Shuff = DAG.getVectorShuffle(WideVecVT, dl, WideVec,
25376 DAG.getUNDEF(WideVecVT),
25378 // At this point all of the data is stored at the bottom of the
25379 // register. We now need to save it to mem.
25381 // Find the largest store unit
25382 MVT StoreType = MVT::i8;
25383 for (MVT Tp : MVT::integer_valuetypes()) {
25384 if (TLI.isTypeLegal(Tp) && Tp.getSizeInBits() <= NumElems * ToSz)
25388 // On 32bit systems, we can't save 64bit integers. Try bitcasting to F64.
25389 if (TLI.isTypeLegal(MVT::f64) && StoreType.getSizeInBits() < 64 &&
25390 (64 <= NumElems * ToSz))
25391 StoreType = MVT::f64;
25393 // Bitcast the original vector into a vector of store-size units
25394 EVT StoreVecVT = EVT::getVectorVT(*DAG.getContext(),
25395 StoreType, VT.getSizeInBits()/StoreType.getSizeInBits());
25396 assert(StoreVecVT.getSizeInBits() == VT.getSizeInBits());
25397 SDValue ShuffWide = DAG.getBitcast(StoreVecVT, Shuff);
25398 SmallVector<SDValue, 8> Chains;
25399 SDValue Increment = DAG.getConstant(StoreType.getSizeInBits() / 8, dl,
25400 TLI.getPointerTy(DAG.getDataLayout()));
25401 SDValue Ptr = St->getBasePtr();
25403 // Perform one or more big stores into memory.
25404 for (unsigned i=0, e=(ToSz*NumElems)/StoreType.getSizeInBits(); i!=e; ++i) {
25405 SDValue SubVec = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl,
25406 StoreType, ShuffWide,
25407 DAG.getIntPtrConstant(i, dl));
25408 SDValue Ch = DAG.getStore(St->getChain(), dl, SubVec, Ptr,
25409 St->getPointerInfo(), St->isVolatile(),
25410 St->isNonTemporal(), St->getAlignment());
25411 Ptr = DAG.getNode(ISD::ADD, dl, Ptr.getValueType(), Ptr, Increment);
25412 Chains.push_back(Ch);
25415 return DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Chains);
25418 // Turn load->store of MMX types into GPR load/stores. This avoids clobbering
25419 // the FP state in cases where an emms may be missing.
25420 // A preferable solution to the general problem is to figure out the right
25421 // places to insert EMMS. This qualifies as a quick hack.
25423 // Similarly, turn load->store of i64 into double load/stores in 32-bit mode.
25424 if (VT.getSizeInBits() != 64)
25427 const Function *F = DAG.getMachineFunction().getFunction();
25428 bool NoImplicitFloatOps = F->hasFnAttribute(Attribute::NoImplicitFloat);
25430 !Subtarget->useSoftFloat() && !NoImplicitFloatOps && Subtarget->hasSSE2();
25431 if ((VT.isVector() ||
25432 (VT == MVT::i64 && F64IsLegal && !Subtarget->is64Bit())) &&
25433 isa<LoadSDNode>(St->getValue()) &&
25434 !cast<LoadSDNode>(St->getValue())->isVolatile() &&
25435 St->getChain().hasOneUse() && !St->isVolatile()) {
25436 SDNode* LdVal = St->getValue().getNode();
25437 LoadSDNode *Ld = nullptr;
25438 int TokenFactorIndex = -1;
25439 SmallVector<SDValue, 8> Ops;
25440 SDNode* ChainVal = St->getChain().getNode();
25441 // Must be a store of a load. We currently handle two cases: the load
25442 // is a direct child, and it's under an intervening TokenFactor. It is
25443 // possible to dig deeper under nested TokenFactors.
25444 if (ChainVal == LdVal)
25445 Ld = cast<LoadSDNode>(St->getChain());
25446 else if (St->getValue().hasOneUse() &&
25447 ChainVal->getOpcode() == ISD::TokenFactor) {
25448 for (unsigned i = 0, e = ChainVal->getNumOperands(); i != e; ++i) {
25449 if (ChainVal->getOperand(i).getNode() == LdVal) {
25450 TokenFactorIndex = i;
25451 Ld = cast<LoadSDNode>(St->getValue());
25453 Ops.push_back(ChainVal->getOperand(i));
25457 if (!Ld || !ISD::isNormalLoad(Ld))
25460 // If this is not the MMX case, i.e. we are just turning i64 load/store
25461 // into f64 load/store, avoid the transformation if there are multiple
25462 // uses of the loaded value.
25463 if (!VT.isVector() && !Ld->hasNUsesOfValue(1, 0))
25468 // If we are a 64-bit capable x86, lower to a single movq load/store pair.
25469 // Otherwise, if it's legal to use f64 SSE instructions, use f64 load/store
25471 if (Subtarget->is64Bit() || F64IsLegal) {
25472 EVT LdVT = Subtarget->is64Bit() ? MVT::i64 : MVT::f64;
25473 SDValue NewLd = DAG.getLoad(LdVT, LdDL, Ld->getChain(), Ld->getBasePtr(),
25474 Ld->getPointerInfo(), Ld->isVolatile(),
25475 Ld->isNonTemporal(), Ld->isInvariant(),
25476 Ld->getAlignment());
25477 SDValue NewChain = NewLd.getValue(1);
25478 if (TokenFactorIndex != -1) {
25479 Ops.push_back(NewChain);
25480 NewChain = DAG.getNode(ISD::TokenFactor, LdDL, MVT::Other, Ops);
25482 return DAG.getStore(NewChain, StDL, NewLd, St->getBasePtr(),
25483 St->getPointerInfo(),
25484 St->isVolatile(), St->isNonTemporal(),
25485 St->getAlignment());
25488 // Otherwise, lower to two pairs of 32-bit loads / stores.
25489 SDValue LoAddr = Ld->getBasePtr();
25490 SDValue HiAddr = DAG.getNode(ISD::ADD, LdDL, MVT::i32, LoAddr,
25491 DAG.getConstant(4, LdDL, MVT::i32));
25493 SDValue LoLd = DAG.getLoad(MVT::i32, LdDL, Ld->getChain(), LoAddr,
25494 Ld->getPointerInfo(),
25495 Ld->isVolatile(), Ld->isNonTemporal(),
25496 Ld->isInvariant(), Ld->getAlignment());
25497 SDValue HiLd = DAG.getLoad(MVT::i32, LdDL, Ld->getChain(), HiAddr,
25498 Ld->getPointerInfo().getWithOffset(4),
25499 Ld->isVolatile(), Ld->isNonTemporal(),
25501 MinAlign(Ld->getAlignment(), 4));
25503 SDValue NewChain = LoLd.getValue(1);
25504 if (TokenFactorIndex != -1) {
25505 Ops.push_back(LoLd);
25506 Ops.push_back(HiLd);
25507 NewChain = DAG.getNode(ISD::TokenFactor, LdDL, MVT::Other, Ops);
25510 LoAddr = St->getBasePtr();
25511 HiAddr = DAG.getNode(ISD::ADD, StDL, MVT::i32, LoAddr,
25512 DAG.getConstant(4, StDL, MVT::i32));
25514 SDValue LoSt = DAG.getStore(NewChain, StDL, LoLd, LoAddr,
25515 St->getPointerInfo(),
25516 St->isVolatile(), St->isNonTemporal(),
25517 St->getAlignment());
25518 SDValue HiSt = DAG.getStore(NewChain, StDL, HiLd, HiAddr,
25519 St->getPointerInfo().getWithOffset(4),
25521 St->isNonTemporal(),
25522 MinAlign(St->getAlignment(), 4));
25523 return DAG.getNode(ISD::TokenFactor, StDL, MVT::Other, LoSt, HiSt);
25526 // This is similar to the above case, but here we handle a scalar 64-bit
25527 // integer store that is extracted from a vector on a 32-bit target.
25528 // If we have SSE2, then we can treat it like a floating-point double
25529 // to get past legalization. The execution dependencies fixup pass will
25530 // choose the optimal machine instruction for the store if this really is
25531 // an integer or v2f32 rather than an f64.
25532 if (VT == MVT::i64 && F64IsLegal && !Subtarget->is64Bit() &&
25533 St->getOperand(1).getOpcode() == ISD::EXTRACT_VECTOR_ELT) {
25534 SDValue OldExtract = St->getOperand(1);
25535 SDValue ExtOp0 = OldExtract.getOperand(0);
25536 unsigned VecSize = ExtOp0.getValueSizeInBits();
25537 EVT VecVT = EVT::getVectorVT(*DAG.getContext(), MVT::f64, VecSize / 64);
25538 SDValue BitCast = DAG.getBitcast(VecVT, ExtOp0);
25539 SDValue NewExtract = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
25540 BitCast, OldExtract.getOperand(1));
25541 return DAG.getStore(St->getChain(), dl, NewExtract, St->getBasePtr(),
25542 St->getPointerInfo(), St->isVolatile(),
25543 St->isNonTemporal(), St->getAlignment());
25549 /// Return 'true' if this vector operation is "horizontal"
25550 /// and return the operands for the horizontal operation in LHS and RHS. A
25551 /// horizontal operation performs the binary operation on successive elements
25552 /// of its first operand, then on successive elements of its second operand,
25553 /// returning the resulting values in a vector. For example, if
25554 /// A = < float a0, float a1, float a2, float a3 >
25556 /// B = < float b0, float b1, float b2, float b3 >
25557 /// then the result of doing a horizontal operation on A and B is
25558 /// A horizontal-op B = < a0 op a1, a2 op a3, b0 op b1, b2 op b3 >.
25559 /// In short, LHS and RHS are inspected to see if LHS op RHS is of the form
25560 /// A horizontal-op B, for some already available A and B, and if so then LHS is
25561 /// set to A, RHS to B, and the routine returns 'true'.
25562 /// Note that the binary operation should have the property that if one of the
25563 /// operands is UNDEF then the result is UNDEF.
25564 static bool isHorizontalBinOp(SDValue &LHS, SDValue &RHS, bool IsCommutative) {
25565 // Look for the following pattern: if
25566 // A = < float a0, float a1, float a2, float a3 >
25567 // B = < float b0, float b1, float b2, float b3 >
25569 // LHS = VECTOR_SHUFFLE A, B, <0, 2, 4, 6>
25570 // RHS = VECTOR_SHUFFLE A, B, <1, 3, 5, 7>
25571 // then LHS op RHS = < a0 op a1, a2 op a3, b0 op b1, b2 op b3 >
25572 // which is A horizontal-op B.
25574 // At least one of the operands should be a vector shuffle.
25575 if (LHS.getOpcode() != ISD::VECTOR_SHUFFLE &&
25576 RHS.getOpcode() != ISD::VECTOR_SHUFFLE)
25579 MVT VT = LHS.getSimpleValueType();
25581 assert((VT.is128BitVector() || VT.is256BitVector()) &&
25582 "Unsupported vector type for horizontal add/sub");
25584 // Handle 128 and 256-bit vector lengths. AVX defines horizontal add/sub to
25585 // operate independently on 128-bit lanes.
25586 unsigned NumElts = VT.getVectorNumElements();
25587 unsigned NumLanes = VT.getSizeInBits()/128;
25588 unsigned NumLaneElts = NumElts / NumLanes;
25589 assert((NumLaneElts % 2 == 0) &&
25590 "Vector type should have an even number of elements in each lane");
25591 unsigned HalfLaneElts = NumLaneElts/2;
25593 // View LHS in the form
25594 // LHS = VECTOR_SHUFFLE A, B, LMask
25595 // If LHS is not a shuffle then pretend it is the shuffle
25596 // LHS = VECTOR_SHUFFLE LHS, undef, <0, 1, ..., N-1>
25597 // NOTE: in what follows a default initialized SDValue represents an UNDEF of
25600 SmallVector<int, 16> LMask(NumElts);
25601 if (LHS.getOpcode() == ISD::VECTOR_SHUFFLE) {
25602 if (LHS.getOperand(0).getOpcode() != ISD::UNDEF)
25603 A = LHS.getOperand(0);
25604 if (LHS.getOperand(1).getOpcode() != ISD::UNDEF)
25605 B = LHS.getOperand(1);
25606 ArrayRef<int> Mask = cast<ShuffleVectorSDNode>(LHS.getNode())->getMask();
25607 std::copy(Mask.begin(), Mask.end(), LMask.begin());
25609 if (LHS.getOpcode() != ISD::UNDEF)
25611 for (unsigned i = 0; i != NumElts; ++i)
25615 // Likewise, view RHS in the form
25616 // RHS = VECTOR_SHUFFLE C, D, RMask
25618 SmallVector<int, 16> RMask(NumElts);
25619 if (RHS.getOpcode() == ISD::VECTOR_SHUFFLE) {
25620 if (RHS.getOperand(0).getOpcode() != ISD::UNDEF)
25621 C = RHS.getOperand(0);
25622 if (RHS.getOperand(1).getOpcode() != ISD::UNDEF)
25623 D = RHS.getOperand(1);
25624 ArrayRef<int> Mask = cast<ShuffleVectorSDNode>(RHS.getNode())->getMask();
25625 std::copy(Mask.begin(), Mask.end(), RMask.begin());
25627 if (RHS.getOpcode() != ISD::UNDEF)
25629 for (unsigned i = 0; i != NumElts; ++i)
25633 // Check that the shuffles are both shuffling the same vectors.
25634 if (!(A == C && B == D) && !(A == D && B == C))
25637 // If everything is UNDEF then bail out: it would be better to fold to UNDEF.
25638 if (!A.getNode() && !B.getNode())
25641 // If A and B occur in reverse order in RHS, then "swap" them (which means
25642 // rewriting the mask).
25644 ShuffleVectorSDNode::commuteMask(RMask);
25646 // At this point LHS and RHS are equivalent to
25647 // LHS = VECTOR_SHUFFLE A, B, LMask
25648 // RHS = VECTOR_SHUFFLE A, B, RMask
25649 // Check that the masks correspond to performing a horizontal operation.
25650 for (unsigned l = 0; l != NumElts; l += NumLaneElts) {
25651 for (unsigned i = 0; i != NumLaneElts; ++i) {
25652 int LIdx = LMask[i+l], RIdx = RMask[i+l];
25654 // Ignore any UNDEF components.
25655 if (LIdx < 0 || RIdx < 0 ||
25656 (!A.getNode() && (LIdx < (int)NumElts || RIdx < (int)NumElts)) ||
25657 (!B.getNode() && (LIdx >= (int)NumElts || RIdx >= (int)NumElts)))
25660 // Check that successive elements are being operated on. If not, this is
25661 // not a horizontal operation.
25662 unsigned Src = (i/HalfLaneElts); // each lane is split between srcs
25663 int Index = 2*(i%HalfLaneElts) + NumElts*Src + l;
25664 if (!(LIdx == Index && RIdx == Index + 1) &&
25665 !(IsCommutative && LIdx == Index + 1 && RIdx == Index))
25670 LHS = A.getNode() ? A : B; // If A is 'UNDEF', use B for it.
25671 RHS = B.getNode() ? B : A; // If B is 'UNDEF', use A for it.
25675 /// Do target-specific dag combines on floating point adds.
25676 static SDValue PerformFADDCombine(SDNode *N, SelectionDAG &DAG,
25677 const X86Subtarget *Subtarget) {
25678 EVT VT = N->getValueType(0);
25679 SDValue LHS = N->getOperand(0);
25680 SDValue RHS = N->getOperand(1);
25682 // Try to synthesize horizontal adds from adds of shuffles.
25683 if (((Subtarget->hasSSE3() && (VT == MVT::v4f32 || VT == MVT::v2f64)) ||
25684 (Subtarget->hasFp256() && (VT == MVT::v8f32 || VT == MVT::v4f64))) &&
25685 isHorizontalBinOp(LHS, RHS, true))
25686 return DAG.getNode(X86ISD::FHADD, SDLoc(N), VT, LHS, RHS);
25690 /// Do target-specific dag combines on floating point subs.
25691 static SDValue PerformFSUBCombine(SDNode *N, SelectionDAG &DAG,
25692 const X86Subtarget *Subtarget) {
25693 EVT VT = N->getValueType(0);
25694 SDValue LHS = N->getOperand(0);
25695 SDValue RHS = N->getOperand(1);
25697 // Try to synthesize horizontal subs from subs of shuffles.
25698 if (((Subtarget->hasSSE3() && (VT == MVT::v4f32 || VT == MVT::v2f64)) ||
25699 (Subtarget->hasFp256() && (VT == MVT::v8f32 || VT == MVT::v4f64))) &&
25700 isHorizontalBinOp(LHS, RHS, false))
25701 return DAG.getNode(X86ISD::FHSUB, SDLoc(N), VT, LHS, RHS);
25705 /// Do target-specific dag combines on X86ISD::FOR and X86ISD::FXOR nodes.
25706 static SDValue PerformFORCombine(SDNode *N, SelectionDAG &DAG,
25707 const X86Subtarget *Subtarget) {
25708 assert(N->getOpcode() == X86ISD::FOR || N->getOpcode() == X86ISD::FXOR);
25710 // F[X]OR(0.0, x) -> x
25711 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
25712 if (C->getValueAPF().isPosZero())
25713 return N->getOperand(1);
25715 // F[X]OR(x, 0.0) -> x
25716 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
25717 if (C->getValueAPF().isPosZero())
25718 return N->getOperand(0);
25720 EVT VT = N->getValueType(0);
25721 if (VT.is512BitVector() && !Subtarget->hasDQI()) {
25723 MVT IntScalar = MVT::getIntegerVT(VT.getScalarSizeInBits());
25724 MVT IntVT = MVT::getVectorVT(IntScalar, VT.getVectorNumElements());
25726 SDValue Op0 = DAG.getNode(ISD::BITCAST, dl, IntVT, N->getOperand(0));
25727 SDValue Op1 = DAG.getNode(ISD::BITCAST, dl, IntVT, N->getOperand(1));
25728 unsigned IntOpcode = (N->getOpcode() == X86ISD::FOR) ? ISD::OR : ISD::XOR;
25729 SDValue IntOp = DAG.getNode(IntOpcode, dl, IntVT, Op0, Op1);
25730 return DAG.getNode(ISD::BITCAST, dl, VT, IntOp);
25735 /// Do target-specific dag combines on X86ISD::FMIN and X86ISD::FMAX nodes.
25736 static SDValue PerformFMinFMaxCombine(SDNode *N, SelectionDAG &DAG) {
25737 assert(N->getOpcode() == X86ISD::FMIN || N->getOpcode() == X86ISD::FMAX);
25739 // Only perform optimizations if UnsafeMath is used.
25740 if (!DAG.getTarget().Options.UnsafeFPMath)
25743 // If we run in unsafe-math mode, then convert the FMAX and FMIN nodes
25744 // into FMINC and FMAXC, which are Commutative operations.
25745 unsigned NewOp = 0;
25746 switch (N->getOpcode()) {
25747 default: llvm_unreachable("unknown opcode");
25748 case X86ISD::FMIN: NewOp = X86ISD::FMINC; break;
25749 case X86ISD::FMAX: NewOp = X86ISD::FMAXC; break;
25752 return DAG.getNode(NewOp, SDLoc(N), N->getValueType(0),
25753 N->getOperand(0), N->getOperand(1));
25756 /// Do target-specific dag combines on X86ISD::FAND nodes.
25757 static SDValue PerformFANDCombine(SDNode *N, SelectionDAG &DAG) {
25758 // FAND(0.0, x) -> 0.0
25759 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
25760 if (C->getValueAPF().isPosZero())
25761 return N->getOperand(0);
25763 // FAND(x, 0.0) -> 0.0
25764 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
25765 if (C->getValueAPF().isPosZero())
25766 return N->getOperand(1);
25771 /// Do target-specific dag combines on X86ISD::FANDN nodes
25772 static SDValue PerformFANDNCombine(SDNode *N, SelectionDAG &DAG) {
25773 // FANDN(0.0, x) -> x
25774 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(0)))
25775 if (C->getValueAPF().isPosZero())
25776 return N->getOperand(1);
25778 // FANDN(x, 0.0) -> 0.0
25779 if (ConstantFPSDNode *C = dyn_cast<ConstantFPSDNode>(N->getOperand(1)))
25780 if (C->getValueAPF().isPosZero())
25781 return N->getOperand(1);
25786 static SDValue PerformBTCombine(SDNode *N,
25788 TargetLowering::DAGCombinerInfo &DCI) {
25789 // BT ignores high bits in the bit index operand.
25790 SDValue Op1 = N->getOperand(1);
25791 if (Op1.hasOneUse()) {
25792 unsigned BitWidth = Op1.getValueSizeInBits();
25793 APInt DemandedMask = APInt::getLowBitsSet(BitWidth, Log2_32(BitWidth));
25794 APInt KnownZero, KnownOne;
25795 TargetLowering::TargetLoweringOpt TLO(DAG, !DCI.isBeforeLegalize(),
25796 !DCI.isBeforeLegalizeOps());
25797 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
25798 if (TLO.ShrinkDemandedConstant(Op1, DemandedMask) ||
25799 TLI.SimplifyDemandedBits(Op1, DemandedMask, KnownZero, KnownOne, TLO))
25800 DCI.CommitTargetLoweringOpt(TLO);
25805 static SDValue PerformVZEXT_MOVLCombine(SDNode *N, SelectionDAG &DAG) {
25806 SDValue Op = N->getOperand(0);
25807 if (Op.getOpcode() == ISD::BITCAST)
25808 Op = Op.getOperand(0);
25809 EVT VT = N->getValueType(0), OpVT = Op.getValueType();
25810 if (Op.getOpcode() == X86ISD::VZEXT_LOAD &&
25811 VT.getVectorElementType().getSizeInBits() ==
25812 OpVT.getVectorElementType().getSizeInBits()) {
25813 return DAG.getNode(ISD::BITCAST, SDLoc(N), VT, Op);
25818 static SDValue PerformSIGN_EXTEND_INREGCombine(SDNode *N, SelectionDAG &DAG,
25819 const X86Subtarget *Subtarget) {
25820 EVT VT = N->getValueType(0);
25821 if (!VT.isVector())
25824 SDValue N0 = N->getOperand(0);
25825 SDValue N1 = N->getOperand(1);
25826 EVT ExtraVT = cast<VTSDNode>(N1)->getVT();
25829 // The SIGN_EXTEND_INREG to v4i64 is expensive operation on the
25830 // both SSE and AVX2 since there is no sign-extended shift right
25831 // operation on a vector with 64-bit elements.
25832 //(sext_in_reg (v4i64 anyext (v4i32 x )), ExtraVT) ->
25833 // (v4i64 sext (v4i32 sext_in_reg (v4i32 x , ExtraVT)))
25834 if (VT == MVT::v4i64 && (N0.getOpcode() == ISD::ANY_EXTEND ||
25835 N0.getOpcode() == ISD::SIGN_EXTEND)) {
25836 SDValue N00 = N0.getOperand(0);
25838 // EXTLOAD has a better solution on AVX2,
25839 // it may be replaced with X86ISD::VSEXT node.
25840 if (N00.getOpcode() == ISD::LOAD && Subtarget->hasInt256())
25841 if (!ISD::isNormalLoad(N00.getNode()))
25844 if (N00.getValueType() == MVT::v4i32 && ExtraVT.getSizeInBits() < 128) {
25845 SDValue Tmp = DAG.getNode(ISD::SIGN_EXTEND_INREG, dl, MVT::v4i32,
25847 return DAG.getNode(ISD::SIGN_EXTEND, dl, MVT::v4i64, Tmp);
25853 /// sext(add_nsw(x, C)) --> add(sext(x), C_sext)
25854 /// Promoting a sign extension ahead of an 'add nsw' exposes opportunities
25855 /// to combine math ops, use an LEA, or use a complex addressing mode. This can
25856 /// eliminate extend, add, and shift instructions.
25857 static SDValue promoteSextBeforeAddNSW(SDNode *Sext, SelectionDAG &DAG,
25858 const X86Subtarget *Subtarget) {
25859 // TODO: This should be valid for other integer types.
25860 EVT VT = Sext->getValueType(0);
25861 if (VT != MVT::i64)
25864 // We need an 'add nsw' feeding into the 'sext'.
25865 SDValue Add = Sext->getOperand(0);
25866 if (Add.getOpcode() != ISD::ADD || !Add->getFlags()->hasNoSignedWrap())
25869 // Having a constant operand to the 'add' ensures that we are not increasing
25870 // the instruction count because the constant is extended for free below.
25871 // A constant operand can also become the displacement field of an LEA.
25872 auto *AddOp1 = dyn_cast<ConstantSDNode>(Add.getOperand(1));
25876 // Don't make the 'add' bigger if there's no hope of combining it with some
25877 // other 'add' or 'shl' instruction.
25878 // TODO: It may be profitable to generate simpler LEA instructions in place
25879 // of single 'add' instructions, but the cost model for selecting an LEA
25880 // currently has a high threshold.
25881 bool HasLEAPotential = false;
25882 for (auto *User : Sext->uses()) {
25883 if (User->getOpcode() == ISD::ADD || User->getOpcode() == ISD::SHL) {
25884 HasLEAPotential = true;
25888 if (!HasLEAPotential)
25891 // Everything looks good, so pull the 'sext' ahead of the 'add'.
25892 int64_t AddConstant = AddOp1->getSExtValue();
25893 SDValue AddOp0 = Add.getOperand(0);
25894 SDValue NewSext = DAG.getNode(ISD::SIGN_EXTEND, SDLoc(Sext), VT, AddOp0);
25895 SDValue NewConstant = DAG.getConstant(AddConstant, SDLoc(Add), VT);
25897 // The wider add is guaranteed to not wrap because both operands are
25900 Flags.setNoSignedWrap(true);
25901 return DAG.getNode(ISD::ADD, SDLoc(Add), VT, NewSext, NewConstant, &Flags);
25904 static SDValue PerformSExtCombine(SDNode *N, SelectionDAG &DAG,
25905 TargetLowering::DAGCombinerInfo &DCI,
25906 const X86Subtarget *Subtarget) {
25907 SDValue N0 = N->getOperand(0);
25908 EVT VT = N->getValueType(0);
25909 EVT SVT = VT.getScalarType();
25910 EVT InVT = N0.getValueType();
25911 EVT InSVT = InVT.getScalarType();
25914 // (i8,i32 sext (sdivrem (i8 x, i8 y)) ->
25915 // (i8,i32 (sdivrem_sext_hreg (i8 x, i8 y)
25916 // This exposes the sext to the sdivrem lowering, so that it directly extends
25917 // from AH (which we otherwise need to do contortions to access).
25918 if (N0.getOpcode() == ISD::SDIVREM && N0.getResNo() == 1 &&
25919 InVT == MVT::i8 && VT == MVT::i32) {
25920 SDVTList NodeTys = DAG.getVTList(MVT::i8, VT);
25921 SDValue R = DAG.getNode(X86ISD::SDIVREM8_SEXT_HREG, DL, NodeTys,
25922 N0.getOperand(0), N0.getOperand(1));
25923 DAG.ReplaceAllUsesOfValueWith(N0.getValue(0), R.getValue(0));
25924 return R.getValue(1);
25927 if (!DCI.isBeforeLegalizeOps()) {
25928 if (InVT == MVT::i1) {
25929 SDValue Zero = DAG.getConstant(0, DL, VT);
25931 DAG.getConstant(APInt::getAllOnesValue(VT.getSizeInBits()), DL, VT);
25932 return DAG.getNode(ISD::SELECT, DL, VT, N0, AllOnes, Zero);
25937 if (VT.isVector() && Subtarget->hasSSE2()) {
25938 auto ExtendVecSize = [&DAG](SDLoc DL, SDValue N, unsigned Size) {
25939 EVT InVT = N.getValueType();
25940 EVT OutVT = EVT::getVectorVT(*DAG.getContext(), InVT.getScalarType(),
25941 Size / InVT.getScalarSizeInBits());
25942 SmallVector<SDValue, 8> Opnds(Size / InVT.getSizeInBits(),
25943 DAG.getUNDEF(InVT));
25945 return DAG.getNode(ISD::CONCAT_VECTORS, DL, OutVT, Opnds);
25948 // If target-size is less than 128-bits, extend to a type that would extend
25949 // to 128 bits, extend that and extract the original target vector.
25950 if (VT.getSizeInBits() < 128 && !(128 % VT.getSizeInBits()) &&
25951 (SVT == MVT::i64 || SVT == MVT::i32 || SVT == MVT::i16) &&
25952 (InSVT == MVT::i32 || InSVT == MVT::i16 || InSVT == MVT::i8)) {
25953 unsigned Scale = 128 / VT.getSizeInBits();
25955 EVT::getVectorVT(*DAG.getContext(), SVT, 128 / SVT.getSizeInBits());
25956 SDValue Ex = ExtendVecSize(DL, N0, Scale * InVT.getSizeInBits());
25957 SDValue SExt = DAG.getNode(ISD::SIGN_EXTEND, DL, ExVT, Ex);
25958 return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, VT, SExt,
25959 DAG.getIntPtrConstant(0, DL));
25962 // If target-size is 128-bits, then convert to ISD::SIGN_EXTEND_VECTOR_INREG
25963 // which ensures lowering to X86ISD::VSEXT (pmovsx*).
25964 if (VT.getSizeInBits() == 128 &&
25965 (SVT == MVT::i64 || SVT == MVT::i32 || SVT == MVT::i16) &&
25966 (InSVT == MVT::i32 || InSVT == MVT::i16 || InSVT == MVT::i8)) {
25967 SDValue ExOp = ExtendVecSize(DL, N0, 128);
25968 return DAG.getSignExtendVectorInReg(ExOp, DL, VT);
25971 // On pre-AVX2 targets, split into 128-bit nodes of
25972 // ISD::SIGN_EXTEND_VECTOR_INREG.
25973 if (!Subtarget->hasInt256() && !(VT.getSizeInBits() % 128) &&
25974 (SVT == MVT::i64 || SVT == MVT::i32 || SVT == MVT::i16) &&
25975 (InSVT == MVT::i32 || InSVT == MVT::i16 || InSVT == MVT::i8)) {
25976 unsigned NumVecs = VT.getSizeInBits() / 128;
25977 unsigned NumSubElts = 128 / SVT.getSizeInBits();
25978 EVT SubVT = EVT::getVectorVT(*DAG.getContext(), SVT, NumSubElts);
25979 EVT InSubVT = EVT::getVectorVT(*DAG.getContext(), InSVT, NumSubElts);
25981 SmallVector<SDValue, 8> Opnds;
25982 for (unsigned i = 0, Offset = 0; i != NumVecs;
25983 ++i, Offset += NumSubElts) {
25984 SDValue SrcVec = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, InSubVT, N0,
25985 DAG.getIntPtrConstant(Offset, DL));
25986 SrcVec = ExtendVecSize(DL, SrcVec, 128);
25987 SrcVec = DAG.getSignExtendVectorInReg(SrcVec, DL, SubVT);
25988 Opnds.push_back(SrcVec);
25990 return DAG.getNode(ISD::CONCAT_VECTORS, DL, VT, Opnds);
25994 if (Subtarget->hasAVX() && VT.isVector() && VT.getSizeInBits() == 256)
25995 if (SDValue R = WidenMaskArithmetic(N, DAG, DCI, Subtarget))
25998 if (SDValue NewAdd = promoteSextBeforeAddNSW(N, DAG, Subtarget))
26004 static SDValue PerformFMACombine(SDNode *N, SelectionDAG &DAG,
26005 const X86Subtarget* Subtarget) {
26007 EVT VT = N->getValueType(0);
26009 // Let legalize expand this if it isn't a legal type yet.
26010 if (!DAG.getTargetLoweringInfo().isTypeLegal(VT))
26013 EVT ScalarVT = VT.getScalarType();
26014 if ((ScalarVT != MVT::f32 && ScalarVT != MVT::f64) ||
26015 (!Subtarget->hasFMA() && !Subtarget->hasFMA4() &&
26016 !Subtarget->hasAVX512()))
26019 SDValue A = N->getOperand(0);
26020 SDValue B = N->getOperand(1);
26021 SDValue C = N->getOperand(2);
26023 bool NegA = (A.getOpcode() == ISD::FNEG);
26024 bool NegB = (B.getOpcode() == ISD::FNEG);
26025 bool NegC = (C.getOpcode() == ISD::FNEG);
26027 // Negative multiplication when NegA xor NegB
26028 bool NegMul = (NegA != NegB);
26030 A = A.getOperand(0);
26032 B = B.getOperand(0);
26034 C = C.getOperand(0);
26038 Opcode = (!NegC) ? X86ISD::FMADD : X86ISD::FMSUB;
26040 Opcode = (!NegC) ? X86ISD::FNMADD : X86ISD::FNMSUB;
26042 return DAG.getNode(Opcode, dl, VT, A, B, C);
26045 static SDValue PerformZExtCombine(SDNode *N, SelectionDAG &DAG,
26046 TargetLowering::DAGCombinerInfo &DCI,
26047 const X86Subtarget *Subtarget) {
26048 // (i32 zext (and (i8 x86isd::setcc_carry), 1)) ->
26049 // (and (i32 x86isd::setcc_carry), 1)
26050 // This eliminates the zext. This transformation is necessary because
26051 // ISD::SETCC is always legalized to i8.
26053 SDValue N0 = N->getOperand(0);
26054 EVT VT = N->getValueType(0);
26056 if (N0.getOpcode() == ISD::AND &&
26058 N0.getOperand(0).hasOneUse()) {
26059 SDValue N00 = N0.getOperand(0);
26060 if (N00.getOpcode() == X86ISD::SETCC_CARRY) {
26061 ConstantSDNode *C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
26062 if (!C || C->getZExtValue() != 1)
26064 return DAG.getNode(ISD::AND, dl, VT,
26065 DAG.getNode(X86ISD::SETCC_CARRY, dl, VT,
26066 N00.getOperand(0), N00.getOperand(1)),
26067 DAG.getConstant(1, dl, VT));
26071 if (N0.getOpcode() == ISD::TRUNCATE &&
26073 N0.getOperand(0).hasOneUse()) {
26074 SDValue N00 = N0.getOperand(0);
26075 if (N00.getOpcode() == X86ISD::SETCC_CARRY) {
26076 return DAG.getNode(ISD::AND, dl, VT,
26077 DAG.getNode(X86ISD::SETCC_CARRY, dl, VT,
26078 N00.getOperand(0), N00.getOperand(1)),
26079 DAG.getConstant(1, dl, VT));
26083 if (VT.is256BitVector())
26084 if (SDValue R = WidenMaskArithmetic(N, DAG, DCI, Subtarget))
26087 // (i8,i32 zext (udivrem (i8 x, i8 y)) ->
26088 // (i8,i32 (udivrem_zext_hreg (i8 x, i8 y)
26089 // This exposes the zext to the udivrem lowering, so that it directly extends
26090 // from AH (which we otherwise need to do contortions to access).
26091 if (N0.getOpcode() == ISD::UDIVREM &&
26092 N0.getResNo() == 1 && N0.getValueType() == MVT::i8 &&
26093 (VT == MVT::i32 || VT == MVT::i64)) {
26094 SDVTList NodeTys = DAG.getVTList(MVT::i8, VT);
26095 SDValue R = DAG.getNode(X86ISD::UDIVREM8_ZEXT_HREG, dl, NodeTys,
26096 N0.getOperand(0), N0.getOperand(1));
26097 DAG.ReplaceAllUsesOfValueWith(N0.getValue(0), R.getValue(0));
26098 return R.getValue(1);
26104 // Optimize x == -y --> x+y == 0
26105 // x != -y --> x+y != 0
26106 static SDValue PerformISDSETCCCombine(SDNode *N, SelectionDAG &DAG,
26107 const X86Subtarget* Subtarget) {
26108 ISD::CondCode CC = cast<CondCodeSDNode>(N->getOperand(2))->get();
26109 SDValue LHS = N->getOperand(0);
26110 SDValue RHS = N->getOperand(1);
26111 EVT VT = N->getValueType(0);
26114 if ((CC == ISD::SETNE || CC == ISD::SETEQ) && LHS.getOpcode() == ISD::SUB)
26115 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(LHS.getOperand(0)))
26116 if (C->getAPIntValue() == 0 && LHS.hasOneUse()) {
26117 SDValue addV = DAG.getNode(ISD::ADD, DL, LHS.getValueType(), RHS,
26118 LHS.getOperand(1));
26119 return DAG.getSetCC(DL, N->getValueType(0), addV,
26120 DAG.getConstant(0, DL, addV.getValueType()), CC);
26122 if ((CC == ISD::SETNE || CC == ISD::SETEQ) && RHS.getOpcode() == ISD::SUB)
26123 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(RHS.getOperand(0)))
26124 if (C->getAPIntValue() == 0 && RHS.hasOneUse()) {
26125 SDValue addV = DAG.getNode(ISD::ADD, DL, RHS.getValueType(), LHS,
26126 RHS.getOperand(1));
26127 return DAG.getSetCC(DL, N->getValueType(0), addV,
26128 DAG.getConstant(0, DL, addV.getValueType()), CC);
26131 if (VT.getScalarType() == MVT::i1 &&
26132 (CC == ISD::SETNE || CC == ISD::SETEQ || ISD::isSignedIntSetCC(CC))) {
26134 (LHS.getOpcode() == ISD::SIGN_EXTEND) &&
26135 (LHS.getOperand(0).getValueType().getScalarType() == MVT::i1);
26136 bool IsVZero1 = ISD::isBuildVectorAllZeros(RHS.getNode());
26138 if (!IsSEXT0 || !IsVZero1) {
26139 // Swap the operands and update the condition code.
26140 std::swap(LHS, RHS);
26141 CC = ISD::getSetCCSwappedOperands(CC);
26143 IsSEXT0 = (LHS.getOpcode() == ISD::SIGN_EXTEND) &&
26144 (LHS.getOperand(0).getValueType().getScalarType() == MVT::i1);
26145 IsVZero1 = ISD::isBuildVectorAllZeros(RHS.getNode());
26148 if (IsSEXT0 && IsVZero1) {
26149 assert(VT == LHS.getOperand(0).getValueType() &&
26150 "Uexpected operand type");
26151 if (CC == ISD::SETGT)
26152 return DAG.getConstant(0, DL, VT);
26153 if (CC == ISD::SETLE)
26154 return DAG.getConstant(1, DL, VT);
26155 if (CC == ISD::SETEQ || CC == ISD::SETGE)
26156 return DAG.getNOT(DL, LHS.getOperand(0), VT);
26158 assert((CC == ISD::SETNE || CC == ISD::SETLT) &&
26159 "Unexpected condition code!");
26160 return LHS.getOperand(0);
26167 static SDValue NarrowVectorLoadToElement(LoadSDNode *Load, unsigned Index,
26168 SelectionDAG &DAG) {
26170 MVT VT = Load->getSimpleValueType(0);
26171 MVT EVT = VT.getVectorElementType();
26172 SDValue Addr = Load->getOperand(1);
26173 SDValue NewAddr = DAG.getNode(
26174 ISD::ADD, dl, Addr.getSimpleValueType(), Addr,
26175 DAG.getConstant(Index * EVT.getStoreSize(), dl,
26176 Addr.getSimpleValueType()));
26179 DAG.getLoad(EVT, dl, Load->getChain(), NewAddr,
26180 DAG.getMachineFunction().getMachineMemOperand(
26181 Load->getMemOperand(), 0, EVT.getStoreSize()));
26185 static SDValue PerformINSERTPSCombine(SDNode *N, SelectionDAG &DAG,
26186 const X86Subtarget *Subtarget) {
26188 MVT VT = N->getOperand(1)->getSimpleValueType(0);
26189 assert((VT == MVT::v4f32 || VT == MVT::v4i32) &&
26190 "X86insertps is only defined for v4x32");
26192 SDValue Ld = N->getOperand(1);
26193 if (MayFoldLoad(Ld)) {
26194 // Extract the countS bits from the immediate so we can get the proper
26195 // address when narrowing the vector load to a specific element.
26196 // When the second source op is a memory address, insertps doesn't use
26197 // countS and just gets an f32 from that address.
26198 unsigned DestIndex =
26199 cast<ConstantSDNode>(N->getOperand(2))->getZExtValue() >> 6;
26201 Ld = NarrowVectorLoadToElement(cast<LoadSDNode>(Ld), DestIndex, DAG);
26203 // Create this as a scalar to vector to match the instruction pattern.
26204 SDValue LoadScalarToVector = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Ld);
26205 // countS bits are ignored when loading from memory on insertps, which
26206 // means we don't need to explicitly set them to 0.
26207 return DAG.getNode(X86ISD::INSERTPS, dl, VT, N->getOperand(0),
26208 LoadScalarToVector, N->getOperand(2));
26213 static SDValue PerformBLENDICombine(SDNode *N, SelectionDAG &DAG) {
26214 SDValue V0 = N->getOperand(0);
26215 SDValue V1 = N->getOperand(1);
26217 EVT VT = N->getValueType(0);
26219 // Canonicalize a v2f64 blend with a mask of 2 by swapping the vector
26220 // operands and changing the mask to 1. This saves us a bunch of
26221 // pattern-matching possibilities related to scalar math ops in SSE/AVX.
26222 // x86InstrInfo knows how to commute this back after instruction selection
26223 // if it would help register allocation.
26225 // TODO: If optimizing for size or a processor that doesn't suffer from
26226 // partial register update stalls, this should be transformed into a MOVSD
26227 // instruction because a MOVSD is 1-2 bytes smaller than a BLENDPD.
26229 if (VT == MVT::v2f64)
26230 if (auto *Mask = dyn_cast<ConstantSDNode>(N->getOperand(2)))
26231 if (Mask->getZExtValue() == 2 && !isShuffleFoldableLoad(V0)) {
26232 SDValue NewMask = DAG.getConstant(1, DL, MVT::i8);
26233 return DAG.getNode(X86ISD::BLENDI, DL, VT, V1, V0, NewMask);
26239 // Helper function of PerformSETCCCombine. It is to materialize "setb reg"
26240 // as "sbb reg,reg", since it can be extended without zext and produces
26241 // an all-ones bit which is more useful than 0/1 in some cases.
26242 static SDValue MaterializeSETB(SDLoc DL, SDValue EFLAGS, SelectionDAG &DAG,
26245 return DAG.getNode(ISD::AND, DL, VT,
26246 DAG.getNode(X86ISD::SETCC_CARRY, DL, MVT::i8,
26247 DAG.getConstant(X86::COND_B, DL, MVT::i8),
26249 DAG.getConstant(1, DL, VT));
26250 assert (VT == MVT::i1 && "Unexpected type for SECCC node");
26251 return DAG.getNode(ISD::TRUNCATE, DL, MVT::i1,
26252 DAG.getNode(X86ISD::SETCC_CARRY, DL, MVT::i8,
26253 DAG.getConstant(X86::COND_B, DL, MVT::i8),
26257 // Optimize RES = X86ISD::SETCC CONDCODE, EFLAG_INPUT
26258 static SDValue PerformSETCCCombine(SDNode *N, SelectionDAG &DAG,
26259 TargetLowering::DAGCombinerInfo &DCI,
26260 const X86Subtarget *Subtarget) {
26262 X86::CondCode CC = X86::CondCode(N->getConstantOperandVal(0));
26263 SDValue EFLAGS = N->getOperand(1);
26265 if (CC == X86::COND_A) {
26266 // Try to convert COND_A into COND_B in an attempt to facilitate
26267 // materializing "setb reg".
26269 // Do not flip "e > c", where "c" is a constant, because Cmp instruction
26270 // cannot take an immediate as its first operand.
26272 if (EFLAGS.getOpcode() == X86ISD::SUB && EFLAGS.hasOneUse() &&
26273 EFLAGS.getValueType().isInteger() &&
26274 !isa<ConstantSDNode>(EFLAGS.getOperand(1))) {
26275 SDValue NewSub = DAG.getNode(X86ISD::SUB, SDLoc(EFLAGS),
26276 EFLAGS.getNode()->getVTList(),
26277 EFLAGS.getOperand(1), EFLAGS.getOperand(0));
26278 SDValue NewEFLAGS = SDValue(NewSub.getNode(), EFLAGS.getResNo());
26279 return MaterializeSETB(DL, NewEFLAGS, DAG, N->getSimpleValueType(0));
26283 // Materialize "setb reg" as "sbb reg,reg", since it can be extended without
26284 // a zext and produces an all-ones bit which is more useful than 0/1 in some
26286 if (CC == X86::COND_B)
26287 return MaterializeSETB(DL, EFLAGS, DAG, N->getSimpleValueType(0));
26289 if (SDValue Flags = checkBoolTestSetCCCombine(EFLAGS, CC)) {
26290 SDValue Cond = DAG.getConstant(CC, DL, MVT::i8);
26291 return DAG.getNode(X86ISD::SETCC, DL, N->getVTList(), Cond, Flags);
26297 // Optimize branch condition evaluation.
26299 static SDValue PerformBrCondCombine(SDNode *N, SelectionDAG &DAG,
26300 TargetLowering::DAGCombinerInfo &DCI,
26301 const X86Subtarget *Subtarget) {
26303 SDValue Chain = N->getOperand(0);
26304 SDValue Dest = N->getOperand(1);
26305 SDValue EFLAGS = N->getOperand(3);
26306 X86::CondCode CC = X86::CondCode(N->getConstantOperandVal(2));
26308 if (SDValue Flags = checkBoolTestSetCCCombine(EFLAGS, CC)) {
26309 SDValue Cond = DAG.getConstant(CC, DL, MVT::i8);
26310 return DAG.getNode(X86ISD::BRCOND, DL, N->getVTList(), Chain, Dest, Cond,
26317 static SDValue performVectorCompareAndMaskUnaryOpCombine(SDNode *N,
26318 SelectionDAG &DAG) {
26319 // Take advantage of vector comparisons producing 0 or -1 in each lane to
26320 // optimize away operation when it's from a constant.
26322 // The general transformation is:
26323 // UNARYOP(AND(VECTOR_CMP(x,y), constant)) -->
26324 // AND(VECTOR_CMP(x,y), constant2)
26325 // constant2 = UNARYOP(constant)
26327 // Early exit if this isn't a vector operation, the operand of the
26328 // unary operation isn't a bitwise AND, or if the sizes of the operations
26329 // aren't the same.
26330 EVT VT = N->getValueType(0);
26331 if (!VT.isVector() || N->getOperand(0)->getOpcode() != ISD::AND ||
26332 N->getOperand(0)->getOperand(0)->getOpcode() != ISD::SETCC ||
26333 VT.getSizeInBits() != N->getOperand(0)->getValueType(0).getSizeInBits())
26336 // Now check that the other operand of the AND is a constant. We could
26337 // make the transformation for non-constant splats as well, but it's unclear
26338 // that would be a benefit as it would not eliminate any operations, just
26339 // perform one more step in scalar code before moving to the vector unit.
26340 if (BuildVectorSDNode *BV =
26341 dyn_cast<BuildVectorSDNode>(N->getOperand(0)->getOperand(1))) {
26342 // Bail out if the vector isn't a constant.
26343 if (!BV->isConstant())
26346 // Everything checks out. Build up the new and improved node.
26348 EVT IntVT = BV->getValueType(0);
26349 // Create a new constant of the appropriate type for the transformed
26351 SDValue SourceConst = DAG.getNode(N->getOpcode(), DL, VT, SDValue(BV, 0));
26352 // The AND node needs bitcasts to/from an integer vector type around it.
26353 SDValue MaskConst = DAG.getBitcast(IntVT, SourceConst);
26354 SDValue NewAnd = DAG.getNode(ISD::AND, DL, IntVT,
26355 N->getOperand(0)->getOperand(0), MaskConst);
26356 SDValue Res = DAG.getBitcast(VT, NewAnd);
26363 static SDValue PerformUINT_TO_FPCombine(SDNode *N, SelectionDAG &DAG,
26364 const X86Subtarget *Subtarget) {
26365 SDValue Op0 = N->getOperand(0);
26366 EVT VT = N->getValueType(0);
26367 EVT InVT = Op0.getValueType();
26368 EVT InSVT = InVT.getScalarType();
26369 const TargetLowering &TLI = DAG.getTargetLoweringInfo();
26371 // UINT_TO_FP(vXi8) -> SINT_TO_FP(ZEXT(vXi8 to vXi32))
26372 // UINT_TO_FP(vXi16) -> SINT_TO_FP(ZEXT(vXi16 to vXi32))
26373 if (InVT.isVector() && (InSVT == MVT::i8 || InSVT == MVT::i16)) {
26375 EVT DstVT = EVT::getVectorVT(*DAG.getContext(), MVT::i32,
26376 InVT.getVectorNumElements());
26377 SDValue P = DAG.getNode(ISD::ZERO_EXTEND, dl, DstVT, Op0);
26379 if (TLI.isOperationLegal(ISD::UINT_TO_FP, DstVT))
26380 return DAG.getNode(ISD::UINT_TO_FP, dl, VT, P);
26382 return DAG.getNode(ISD::SINT_TO_FP, dl, VT, P);
26388 static SDValue PerformSINT_TO_FPCombine(SDNode *N, SelectionDAG &DAG,
26389 const X86Subtarget *Subtarget) {
26390 // First try to optimize away the conversion entirely when it's
26391 // conditionally from a constant. Vectors only.
26392 if (SDValue Res = performVectorCompareAndMaskUnaryOpCombine(N, DAG))
26395 // Now move on to more general possibilities.
26396 SDValue Op0 = N->getOperand(0);
26397 EVT VT = N->getValueType(0);
26398 EVT InVT = Op0.getValueType();
26399 EVT InSVT = InVT.getScalarType();
26401 // SINT_TO_FP(vXi8) -> SINT_TO_FP(SEXT(vXi8 to vXi32))
26402 // SINT_TO_FP(vXi16) -> SINT_TO_FP(SEXT(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::SIGN_EXTEND, dl, DstVT, Op0);
26408 return DAG.getNode(ISD::SINT_TO_FP, dl, VT, P);
26411 // Transform (SINT_TO_FP (i64 ...)) into an x87 operation if we have
26412 // a 32-bit target where SSE doesn't support i64->FP operations.
26413 if (Op0.getOpcode() == ISD::LOAD) {
26414 LoadSDNode *Ld = cast<LoadSDNode>(Op0.getNode());
26415 EVT LdVT = Ld->getValueType(0);
26417 // This transformation is not supported if the result type is f16
26418 if (VT == MVT::f16)
26421 if (!Ld->isVolatile() && !VT.isVector() &&
26422 ISD::isNON_EXTLoad(Op0.getNode()) && Op0.hasOneUse() &&
26423 !Subtarget->is64Bit() && LdVT == MVT::i64) {
26424 SDValue FILDChain = Subtarget->getTargetLowering()->BuildFILD(
26425 SDValue(N, 0), LdVT, Ld->getChain(), Op0, DAG);
26426 DAG.ReplaceAllUsesOfValueWith(Op0.getValue(1), FILDChain.getValue(1));
26433 // Optimize RES, EFLAGS = X86ISD::ADC LHS, RHS, EFLAGS
26434 static SDValue PerformADCCombine(SDNode *N, SelectionDAG &DAG,
26435 X86TargetLowering::DAGCombinerInfo &DCI) {
26436 // If the LHS and RHS of the ADC node are zero, then it can't overflow and
26437 // the result is either zero or one (depending on the input carry bit).
26438 // Strength reduce this down to a "set on carry" aka SETCC_CARRY&1.
26439 if (X86::isZeroNode(N->getOperand(0)) &&
26440 X86::isZeroNode(N->getOperand(1)) &&
26441 // We don't have a good way to replace an EFLAGS use, so only do this when
26443 SDValue(N, 1).use_empty()) {
26445 EVT VT = N->getValueType(0);
26446 SDValue CarryOut = DAG.getConstant(0, DL, N->getValueType(1));
26447 SDValue Res1 = DAG.getNode(ISD::AND, DL, VT,
26448 DAG.getNode(X86ISD::SETCC_CARRY, DL, VT,
26449 DAG.getConstant(X86::COND_B, DL,
26452 DAG.getConstant(1, DL, VT));
26453 return DCI.CombineTo(N, Res1, CarryOut);
26459 // fold (add Y, (sete X, 0)) -> adc 0, Y
26460 // (add Y, (setne X, 0)) -> sbb -1, Y
26461 // (sub (sete X, 0), Y) -> sbb 0, Y
26462 // (sub (setne X, 0), Y) -> adc -1, Y
26463 static SDValue OptimizeConditionalInDecrement(SDNode *N, SelectionDAG &DAG) {
26466 // Look through ZExts.
26467 SDValue Ext = N->getOperand(N->getOpcode() == ISD::SUB ? 1 : 0);
26468 if (Ext.getOpcode() != ISD::ZERO_EXTEND || !Ext.hasOneUse())
26471 SDValue SetCC = Ext.getOperand(0);
26472 if (SetCC.getOpcode() != X86ISD::SETCC || !SetCC.hasOneUse())
26475 X86::CondCode CC = (X86::CondCode)SetCC.getConstantOperandVal(0);
26476 if (CC != X86::COND_E && CC != X86::COND_NE)
26479 SDValue Cmp = SetCC.getOperand(1);
26480 if (Cmp.getOpcode() != X86ISD::CMP || !Cmp.hasOneUse() ||
26481 !X86::isZeroNode(Cmp.getOperand(1)) ||
26482 !Cmp.getOperand(0).getValueType().isInteger())
26485 SDValue CmpOp0 = Cmp.getOperand(0);
26486 SDValue NewCmp = DAG.getNode(X86ISD::CMP, DL, MVT::i32, CmpOp0,
26487 DAG.getConstant(1, DL, CmpOp0.getValueType()));
26489 SDValue OtherVal = N->getOperand(N->getOpcode() == ISD::SUB ? 0 : 1);
26490 if (CC == X86::COND_NE)
26491 return DAG.getNode(N->getOpcode() == ISD::SUB ? X86ISD::ADC : X86ISD::SBB,
26492 DL, OtherVal.getValueType(), OtherVal,
26493 DAG.getConstant(-1ULL, DL, OtherVal.getValueType()),
26495 return DAG.getNode(N->getOpcode() == ISD::SUB ? X86ISD::SBB : X86ISD::ADC,
26496 DL, OtherVal.getValueType(), OtherVal,
26497 DAG.getConstant(0, DL, OtherVal.getValueType()), NewCmp);
26500 /// PerformADDCombine - Do target-specific dag combines on integer adds.
26501 static SDValue PerformAddCombine(SDNode *N, SelectionDAG &DAG,
26502 const X86Subtarget *Subtarget) {
26503 EVT VT = N->getValueType(0);
26504 SDValue Op0 = N->getOperand(0);
26505 SDValue Op1 = N->getOperand(1);
26507 // Try to synthesize horizontal adds from adds of shuffles.
26508 if (((Subtarget->hasSSSE3() && (VT == MVT::v8i16 || VT == MVT::v4i32)) ||
26509 (Subtarget->hasInt256() && (VT == MVT::v16i16 || VT == MVT::v8i32))) &&
26510 isHorizontalBinOp(Op0, Op1, true))
26511 return DAG.getNode(X86ISD::HADD, SDLoc(N), VT, Op0, Op1);
26513 return OptimizeConditionalInDecrement(N, DAG);
26516 static SDValue PerformSubCombine(SDNode *N, SelectionDAG &DAG,
26517 const X86Subtarget *Subtarget) {
26518 SDValue Op0 = N->getOperand(0);
26519 SDValue Op1 = N->getOperand(1);
26521 // X86 can't encode an immediate LHS of a sub. See if we can push the
26522 // negation into a preceding instruction.
26523 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op0)) {
26524 // If the RHS of the sub is a XOR with one use and a constant, invert the
26525 // immediate. Then add one to the LHS of the sub so we can turn
26526 // X-Y -> X+~Y+1, saving one register.
26527 if (Op1->hasOneUse() && Op1.getOpcode() == ISD::XOR &&
26528 isa<ConstantSDNode>(Op1.getOperand(1))) {
26529 APInt XorC = cast<ConstantSDNode>(Op1.getOperand(1))->getAPIntValue();
26530 EVT VT = Op0.getValueType();
26531 SDValue NewXor = DAG.getNode(ISD::XOR, SDLoc(Op1), VT,
26533 DAG.getConstant(~XorC, SDLoc(Op1), VT));
26534 return DAG.getNode(ISD::ADD, SDLoc(N), VT, NewXor,
26535 DAG.getConstant(C->getAPIntValue() + 1, SDLoc(N), VT));
26539 // Try to synthesize horizontal adds from adds of shuffles.
26540 EVT VT = N->getValueType(0);
26541 if (((Subtarget->hasSSSE3() && (VT == MVT::v8i16 || VT == MVT::v4i32)) ||
26542 (Subtarget->hasInt256() && (VT == MVT::v16i16 || VT == MVT::v8i32))) &&
26543 isHorizontalBinOp(Op0, Op1, true))
26544 return DAG.getNode(X86ISD::HSUB, SDLoc(N), VT, Op0, Op1);
26546 return OptimizeConditionalInDecrement(N, DAG);
26549 /// performVZEXTCombine - Performs build vector combines
26550 static SDValue performVZEXTCombine(SDNode *N, SelectionDAG &DAG,
26551 TargetLowering::DAGCombinerInfo &DCI,
26552 const X86Subtarget *Subtarget) {
26554 MVT VT = N->getSimpleValueType(0);
26555 SDValue Op = N->getOperand(0);
26556 MVT OpVT = Op.getSimpleValueType();
26557 MVT OpEltVT = OpVT.getVectorElementType();
26558 unsigned InputBits = OpEltVT.getSizeInBits() * VT.getVectorNumElements();
26560 // (vzext (bitcast (vzext (x)) -> (vzext x)
26562 while (V.getOpcode() == ISD::BITCAST)
26563 V = V.getOperand(0);
26565 if (V != Op && V.getOpcode() == X86ISD::VZEXT) {
26566 MVT InnerVT = V.getSimpleValueType();
26567 MVT InnerEltVT = InnerVT.getVectorElementType();
26569 // If the element sizes match exactly, we can just do one larger vzext. This
26570 // is always an exact type match as vzext operates on integer types.
26571 if (OpEltVT == InnerEltVT) {
26572 assert(OpVT == InnerVT && "Types must match for vzext!");
26573 return DAG.getNode(X86ISD::VZEXT, DL, VT, V.getOperand(0));
26576 // The only other way we can combine them is if only a single element of the
26577 // inner vzext is used in the input to the outer vzext.
26578 if (InnerEltVT.getSizeInBits() < InputBits)
26581 // In this case, the inner vzext is completely dead because we're going to
26582 // only look at bits inside of the low element. Just do the outer vzext on
26583 // a bitcast of the input to the inner.
26584 return DAG.getNode(X86ISD::VZEXT, DL, VT, DAG.getBitcast(OpVT, V));
26587 // Check if we can bypass extracting and re-inserting an element of an input
26588 // vector. Essentially:
26589 // (bitcast (sclr2vec (ext_vec_elt x))) -> (bitcast x)
26590 if (V.getOpcode() == ISD::SCALAR_TO_VECTOR &&
26591 V.getOperand(0).getOpcode() == ISD::EXTRACT_VECTOR_ELT &&
26592 V.getOperand(0).getSimpleValueType().getSizeInBits() == InputBits) {
26593 SDValue ExtractedV = V.getOperand(0);
26594 SDValue OrigV = ExtractedV.getOperand(0);
26595 if (auto *ExtractIdx = dyn_cast<ConstantSDNode>(ExtractedV.getOperand(1)))
26596 if (ExtractIdx->getZExtValue() == 0) {
26597 MVT OrigVT = OrigV.getSimpleValueType();
26598 // Extract a subvector if necessary...
26599 if (OrigVT.getSizeInBits() > OpVT.getSizeInBits()) {
26600 int Ratio = OrigVT.getSizeInBits() / OpVT.getSizeInBits();
26601 OrigVT = MVT::getVectorVT(OrigVT.getVectorElementType(),
26602 OrigVT.getVectorNumElements() / Ratio);
26603 OrigV = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, OrigVT, OrigV,
26604 DAG.getIntPtrConstant(0, DL));
26606 Op = DAG.getBitcast(OpVT, OrigV);
26607 return DAG.getNode(X86ISD::VZEXT, DL, VT, Op);
26614 SDValue X86TargetLowering::PerformDAGCombine(SDNode *N,
26615 DAGCombinerInfo &DCI) const {
26616 SelectionDAG &DAG = DCI.DAG;
26617 switch (N->getOpcode()) {
26619 case ISD::EXTRACT_VECTOR_ELT:
26620 return PerformEXTRACT_VECTOR_ELTCombine(N, DAG, DCI);
26623 case X86ISD::SHRUNKBLEND:
26624 return PerformSELECTCombine(N, DAG, DCI, Subtarget);
26625 case ISD::BITCAST: return PerformBITCASTCombine(N, DAG, Subtarget);
26626 case X86ISD::CMOV: return PerformCMOVCombine(N, DAG, DCI, Subtarget);
26627 case ISD::ADD: return PerformAddCombine(N, DAG, Subtarget);
26628 case ISD::SUB: return PerformSubCombine(N, DAG, Subtarget);
26629 case X86ISD::ADC: return PerformADCCombine(N, DAG, DCI);
26630 case ISD::MUL: return PerformMulCombine(N, DAG, DCI);
26633 case ISD::SRL: return PerformShiftCombine(N, DAG, DCI, Subtarget);
26634 case ISD::AND: return PerformAndCombine(N, DAG, DCI, Subtarget);
26635 case ISD::OR: return PerformOrCombine(N, DAG, DCI, Subtarget);
26636 case ISD::XOR: return PerformXorCombine(N, DAG, DCI, Subtarget);
26637 case ISD::LOAD: return PerformLOADCombine(N, DAG, DCI, Subtarget);
26638 case ISD::MLOAD: return PerformMLOADCombine(N, DAG, DCI, Subtarget);
26639 case ISD::STORE: return PerformSTORECombine(N, DAG, Subtarget);
26640 case ISD::MSTORE: return PerformMSTORECombine(N, DAG, Subtarget);
26641 case ISD::SINT_TO_FP: return PerformSINT_TO_FPCombine(N, DAG, Subtarget);
26642 case ISD::UINT_TO_FP: return PerformUINT_TO_FPCombine(N, DAG, Subtarget);
26643 case ISD::FADD: return PerformFADDCombine(N, DAG, Subtarget);
26644 case ISD::FSUB: return PerformFSUBCombine(N, DAG, Subtarget);
26646 case X86ISD::FOR: return PerformFORCombine(N, DAG, Subtarget);
26648 case X86ISD::FMAX: return PerformFMinFMaxCombine(N, DAG);
26649 case X86ISD::FAND: return PerformFANDCombine(N, DAG);
26650 case X86ISD::FANDN: return PerformFANDNCombine(N, DAG);
26651 case X86ISD::BT: return PerformBTCombine(N, DAG, DCI);
26652 case X86ISD::VZEXT_MOVL: return PerformVZEXT_MOVLCombine(N, DAG);
26653 case ISD::ANY_EXTEND:
26654 case ISD::ZERO_EXTEND: return PerformZExtCombine(N, DAG, DCI, Subtarget);
26655 case ISD::SIGN_EXTEND: return PerformSExtCombine(N, DAG, DCI, Subtarget);
26656 case ISD::SIGN_EXTEND_INREG:
26657 return PerformSIGN_EXTEND_INREGCombine(N, DAG, Subtarget);
26658 case ISD::SETCC: return PerformISDSETCCCombine(N, DAG, Subtarget);
26659 case X86ISD::SETCC: return PerformSETCCCombine(N, DAG, DCI, Subtarget);
26660 case X86ISD::BRCOND: return PerformBrCondCombine(N, DAG, DCI, Subtarget);
26661 case X86ISD::VZEXT: return performVZEXTCombine(N, DAG, DCI, Subtarget);
26662 case X86ISD::SHUFP: // Handle all target specific shuffles
26663 case X86ISD::PALIGNR:
26664 case X86ISD::UNPCKH:
26665 case X86ISD::UNPCKL:
26666 case X86ISD::MOVHLPS:
26667 case X86ISD::MOVLHPS:
26668 case X86ISD::PSHUFB:
26669 case X86ISD::PSHUFD:
26670 case X86ISD::PSHUFHW:
26671 case X86ISD::PSHUFLW:
26672 case X86ISD::MOVSS:
26673 case X86ISD::MOVSD:
26674 case X86ISD::VPERMILPI:
26675 case X86ISD::VPERM2X128:
26676 case ISD::VECTOR_SHUFFLE: return PerformShuffleCombine(N, DAG, DCI,Subtarget);
26677 case ISD::FMA: return PerformFMACombine(N, DAG, Subtarget);
26678 case X86ISD::INSERTPS: {
26679 if (getTargetMachine().getOptLevel() > CodeGenOpt::None)
26680 return PerformINSERTPSCombine(N, DAG, Subtarget);
26683 case X86ISD::BLENDI: return PerformBLENDICombine(N, DAG);
26689 /// isTypeDesirableForOp - Return true if the target has native support for
26690 /// the specified value type and it is 'desirable' to use the type for the
26691 /// given node type. e.g. On x86 i16 is legal, but undesirable since i16
26692 /// instruction encodings are longer and some i16 instructions are slow.
26693 bool X86TargetLowering::isTypeDesirableForOp(unsigned Opc, EVT VT) const {
26694 if (!isTypeLegal(VT))
26696 if (VT != MVT::i16)
26703 case ISD::SIGN_EXTEND:
26704 case ISD::ZERO_EXTEND:
26705 case ISD::ANY_EXTEND:
26718 /// IsDesirableToPromoteOp - This method query the target whether it is
26719 /// beneficial for dag combiner to promote the specified node. If true, it
26720 /// should return the desired promotion type by reference.
26721 bool X86TargetLowering::IsDesirableToPromoteOp(SDValue Op, EVT &PVT) const {
26722 EVT VT = Op.getValueType();
26723 if (VT != MVT::i16)
26726 bool Promote = false;
26727 bool Commute = false;
26728 switch (Op.getOpcode()) {
26731 LoadSDNode *LD = cast<LoadSDNode>(Op);
26732 // If the non-extending load has a single use and it's not live out, then it
26733 // might be folded.
26734 if (LD->getExtensionType() == ISD::NON_EXTLOAD /*&&
26735 Op.hasOneUse()*/) {
26736 for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
26737 UE = Op.getNode()->use_end(); UI != UE; ++UI) {
26738 // The only case where we'd want to promote LOAD (rather then it being
26739 // promoted as an operand is when it's only use is liveout.
26740 if (UI->getOpcode() != ISD::CopyToReg)
26747 case ISD::SIGN_EXTEND:
26748 case ISD::ZERO_EXTEND:
26749 case ISD::ANY_EXTEND:
26754 SDValue N0 = Op.getOperand(0);
26755 // Look out for (store (shl (load), x)).
26756 if (MayFoldLoad(N0) && MayFoldIntoStore(Op))
26769 SDValue N0 = Op.getOperand(0);
26770 SDValue N1 = Op.getOperand(1);
26771 if (!Commute && MayFoldLoad(N1))
26773 // Avoid disabling potential load folding opportunities.
26774 if (MayFoldLoad(N0) && (!isa<ConstantSDNode>(N1) || MayFoldIntoStore(Op)))
26776 if (MayFoldLoad(N1) && (!isa<ConstantSDNode>(N0) || MayFoldIntoStore(Op)))
26786 //===----------------------------------------------------------------------===//
26787 // X86 Inline Assembly Support
26788 //===----------------------------------------------------------------------===//
26790 // Helper to match a string separated by whitespace.
26791 static bool matchAsm(StringRef S, ArrayRef<const char *> Pieces) {
26792 S = S.substr(S.find_first_not_of(" \t")); // Skip leading whitespace.
26794 for (StringRef Piece : Pieces) {
26795 if (!S.startswith(Piece)) // Check if the piece matches.
26798 S = S.substr(Piece.size());
26799 StringRef::size_type Pos = S.find_first_not_of(" \t");
26800 if (Pos == 0) // We matched a prefix.
26809 static bool clobbersFlagRegisters(const SmallVector<StringRef, 4> &AsmPieces) {
26811 if (AsmPieces.size() == 3 || AsmPieces.size() == 4) {
26812 if (std::count(AsmPieces.begin(), AsmPieces.end(), "~{cc}") &&
26813 std::count(AsmPieces.begin(), AsmPieces.end(), "~{flags}") &&
26814 std::count(AsmPieces.begin(), AsmPieces.end(), "~{fpsr}")) {
26816 if (AsmPieces.size() == 3)
26818 else if (std::count(AsmPieces.begin(), AsmPieces.end(), "~{dirflag}"))
26825 bool X86TargetLowering::ExpandInlineAsm(CallInst *CI) const {
26826 InlineAsm *IA = cast<InlineAsm>(CI->getCalledValue());
26828 std::string AsmStr = IA->getAsmString();
26830 IntegerType *Ty = dyn_cast<IntegerType>(CI->getType());
26831 if (!Ty || Ty->getBitWidth() % 16 != 0)
26834 // TODO: should remove alternatives from the asmstring: "foo {a|b}" -> "foo a"
26835 SmallVector<StringRef, 4> AsmPieces;
26836 SplitString(AsmStr, AsmPieces, ";\n");
26838 switch (AsmPieces.size()) {
26839 default: return false;
26841 // FIXME: this should verify that we are targeting a 486 or better. If not,
26842 // we will turn this bswap into something that will be lowered to logical
26843 // ops instead of emitting the bswap asm. For now, we don't support 486 or
26844 // lower so don't worry about this.
26846 if (matchAsm(AsmPieces[0], {"bswap", "$0"}) ||
26847 matchAsm(AsmPieces[0], {"bswapl", "$0"}) ||
26848 matchAsm(AsmPieces[0], {"bswapq", "$0"}) ||
26849 matchAsm(AsmPieces[0], {"bswap", "${0:q}"}) ||
26850 matchAsm(AsmPieces[0], {"bswapl", "${0:q}"}) ||
26851 matchAsm(AsmPieces[0], {"bswapq", "${0:q}"})) {
26852 // No need to check constraints, nothing other than the equivalent of
26853 // "=r,0" would be valid here.
26854 return IntrinsicLowering::LowerToByteSwap(CI);
26857 // rorw $$8, ${0:w} --> llvm.bswap.i16
26858 if (CI->getType()->isIntegerTy(16) &&
26859 IA->getConstraintString().compare(0, 5, "=r,0,") == 0 &&
26860 (matchAsm(AsmPieces[0], {"rorw", "$$8,", "${0:w}"}) ||
26861 matchAsm(AsmPieces[0], {"rolw", "$$8,", "${0:w}"}))) {
26863 StringRef ConstraintsStr = IA->getConstraintString();
26864 SplitString(StringRef(ConstraintsStr).substr(5), AsmPieces, ",");
26865 array_pod_sort(AsmPieces.begin(), AsmPieces.end());
26866 if (clobbersFlagRegisters(AsmPieces))
26867 return IntrinsicLowering::LowerToByteSwap(CI);
26871 if (CI->getType()->isIntegerTy(32) &&
26872 IA->getConstraintString().compare(0, 5, "=r,0,") == 0 &&
26873 matchAsm(AsmPieces[0], {"rorw", "$$8,", "${0:w}"}) &&
26874 matchAsm(AsmPieces[1], {"rorl", "$$16,", "$0"}) &&
26875 matchAsm(AsmPieces[2], {"rorw", "$$8,", "${0:w}"})) {
26877 StringRef ConstraintsStr = IA->getConstraintString();
26878 SplitString(StringRef(ConstraintsStr).substr(5), AsmPieces, ",");
26879 array_pod_sort(AsmPieces.begin(), AsmPieces.end());
26880 if (clobbersFlagRegisters(AsmPieces))
26881 return IntrinsicLowering::LowerToByteSwap(CI);
26884 if (CI->getType()->isIntegerTy(64)) {
26885 InlineAsm::ConstraintInfoVector Constraints = IA->ParseConstraints();
26886 if (Constraints.size() >= 2 &&
26887 Constraints[0].Codes.size() == 1 && Constraints[0].Codes[0] == "A" &&
26888 Constraints[1].Codes.size() == 1 && Constraints[1].Codes[0] == "0") {
26889 // bswap %eax / bswap %edx / xchgl %eax, %edx -> llvm.bswap.i64
26890 if (matchAsm(AsmPieces[0], {"bswap", "%eax"}) &&
26891 matchAsm(AsmPieces[1], {"bswap", "%edx"}) &&
26892 matchAsm(AsmPieces[2], {"xchgl", "%eax,", "%edx"}))
26893 return IntrinsicLowering::LowerToByteSwap(CI);
26901 /// getConstraintType - Given a constraint letter, return the type of
26902 /// constraint it is for this target.
26903 X86TargetLowering::ConstraintType
26904 X86TargetLowering::getConstraintType(StringRef Constraint) const {
26905 if (Constraint.size() == 1) {
26906 switch (Constraint[0]) {
26917 return C_RegisterClass;
26941 return TargetLowering::getConstraintType(Constraint);
26944 /// Examine constraint type and operand type and determine a weight value.
26945 /// This object must already have been set up with the operand type
26946 /// and the current alternative constraint selected.
26947 TargetLowering::ConstraintWeight
26948 X86TargetLowering::getSingleConstraintMatchWeight(
26949 AsmOperandInfo &info, const char *constraint) const {
26950 ConstraintWeight weight = CW_Invalid;
26951 Value *CallOperandVal = info.CallOperandVal;
26952 // If we don't have a value, we can't do a match,
26953 // but allow it at the lowest weight.
26954 if (!CallOperandVal)
26956 Type *type = CallOperandVal->getType();
26957 // Look at the constraint type.
26958 switch (*constraint) {
26960 weight = TargetLowering::getSingleConstraintMatchWeight(info, constraint);
26971 if (CallOperandVal->getType()->isIntegerTy())
26972 weight = CW_SpecificReg;
26977 if (type->isFloatingPointTy())
26978 weight = CW_SpecificReg;
26981 if (type->isX86_MMXTy() && Subtarget->hasMMX())
26982 weight = CW_SpecificReg;
26986 if (((type->getPrimitiveSizeInBits() == 128) && Subtarget->hasSSE1()) ||
26987 ((type->getPrimitiveSizeInBits() == 256) && Subtarget->hasFp256()))
26988 weight = CW_Register;
26991 if (ConstantInt *C = dyn_cast<ConstantInt>(info.CallOperandVal)) {
26992 if (C->getZExtValue() <= 31)
26993 weight = CW_Constant;
26997 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
26998 if (C->getZExtValue() <= 63)
26999 weight = CW_Constant;
27003 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
27004 if ((C->getSExtValue() >= -0x80) && (C->getSExtValue() <= 0x7f))
27005 weight = CW_Constant;
27009 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
27010 if ((C->getZExtValue() == 0xff) || (C->getZExtValue() == 0xffff))
27011 weight = CW_Constant;
27015 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
27016 if (C->getZExtValue() <= 3)
27017 weight = CW_Constant;
27021 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
27022 if (C->getZExtValue() <= 0xff)
27023 weight = CW_Constant;
27028 if (isa<ConstantFP>(CallOperandVal)) {
27029 weight = CW_Constant;
27033 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
27034 if ((C->getSExtValue() >= -0x80000000LL) &&
27035 (C->getSExtValue() <= 0x7fffffffLL))
27036 weight = CW_Constant;
27040 if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
27041 if (C->getZExtValue() <= 0xffffffff)
27042 weight = CW_Constant;
27049 /// LowerXConstraint - try to replace an X constraint, which matches anything,
27050 /// with another that has more specific requirements based on the type of the
27051 /// corresponding operand.
27052 const char *X86TargetLowering::
27053 LowerXConstraint(EVT ConstraintVT) const {
27054 // FP X constraints get lowered to SSE1/2 registers if available, otherwise
27055 // 'f' like normal targets.
27056 if (ConstraintVT.isFloatingPoint()) {
27057 if (Subtarget->hasSSE2())
27059 if (Subtarget->hasSSE1())
27063 return TargetLowering::LowerXConstraint(ConstraintVT);
27066 /// LowerAsmOperandForConstraint - Lower the specified operand into the Ops
27067 /// vector. If it is invalid, don't add anything to Ops.
27068 void X86TargetLowering::LowerAsmOperandForConstraint(SDValue Op,
27069 std::string &Constraint,
27070 std::vector<SDValue>&Ops,
27071 SelectionDAG &DAG) const {
27074 // Only support length 1 constraints for now.
27075 if (Constraint.length() > 1) return;
27077 char ConstraintLetter = Constraint[0];
27078 switch (ConstraintLetter) {
27081 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
27082 if (C->getZExtValue() <= 31) {
27083 Result = DAG.getTargetConstant(C->getZExtValue(), SDLoc(Op),
27084 Op.getValueType());
27090 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
27091 if (C->getZExtValue() <= 63) {
27092 Result = DAG.getTargetConstant(C->getZExtValue(), SDLoc(Op),
27093 Op.getValueType());
27099 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
27100 if (isInt<8>(C->getSExtValue())) {
27101 Result = DAG.getTargetConstant(C->getZExtValue(), SDLoc(Op),
27102 Op.getValueType());
27108 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
27109 if (C->getZExtValue() == 0xff || C->getZExtValue() == 0xffff ||
27110 (Subtarget->is64Bit() && C->getZExtValue() == 0xffffffff)) {
27111 Result = DAG.getTargetConstant(C->getSExtValue(), SDLoc(Op),
27112 Op.getValueType());
27118 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
27119 if (C->getZExtValue() <= 3) {
27120 Result = DAG.getTargetConstant(C->getZExtValue(), SDLoc(Op),
27121 Op.getValueType());
27127 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
27128 if (C->getZExtValue() <= 255) {
27129 Result = DAG.getTargetConstant(C->getZExtValue(), SDLoc(Op),
27130 Op.getValueType());
27136 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
27137 if (C->getZExtValue() <= 127) {
27138 Result = DAG.getTargetConstant(C->getZExtValue(), SDLoc(Op),
27139 Op.getValueType());
27145 // 32-bit signed value
27146 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
27147 if (ConstantInt::isValueValidForType(Type::getInt32Ty(*DAG.getContext()),
27148 C->getSExtValue())) {
27149 // Widen to 64 bits here to get it sign extended.
27150 Result = DAG.getTargetConstant(C->getSExtValue(), SDLoc(Op), MVT::i64);
27153 // FIXME gcc accepts some relocatable values here too, but only in certain
27154 // memory models; it's complicated.
27159 // 32-bit unsigned value
27160 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op)) {
27161 if (ConstantInt::isValueValidForType(Type::getInt32Ty(*DAG.getContext()),
27162 C->getZExtValue())) {
27163 Result = DAG.getTargetConstant(C->getZExtValue(), SDLoc(Op),
27164 Op.getValueType());
27168 // FIXME gcc accepts some relocatable values here too, but only in certain
27169 // memory models; it's complicated.
27173 // Literal immediates are always ok.
27174 if (ConstantSDNode *CST = dyn_cast<ConstantSDNode>(Op)) {
27175 // Widen to 64 bits here to get it sign extended.
27176 Result = DAG.getTargetConstant(CST->getSExtValue(), SDLoc(Op), MVT::i64);
27180 // In any sort of PIC mode addresses need to be computed at runtime by
27181 // adding in a register or some sort of table lookup. These can't
27182 // be used as immediates.
27183 if (Subtarget->isPICStyleGOT() || Subtarget->isPICStyleStubPIC())
27186 // If we are in non-pic codegen mode, we allow the address of a global (with
27187 // an optional displacement) to be used with 'i'.
27188 GlobalAddressSDNode *GA = nullptr;
27189 int64_t Offset = 0;
27191 // Match either (GA), (GA+C), (GA+C1+C2), etc.
27193 if ((GA = dyn_cast<GlobalAddressSDNode>(Op))) {
27194 Offset += GA->getOffset();
27196 } else if (Op.getOpcode() == ISD::ADD) {
27197 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
27198 Offset += C->getZExtValue();
27199 Op = Op.getOperand(0);
27202 } else if (Op.getOpcode() == ISD::SUB) {
27203 if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op.getOperand(1))) {
27204 Offset += -C->getZExtValue();
27205 Op = Op.getOperand(0);
27210 // Otherwise, this isn't something we can handle, reject it.
27214 const GlobalValue *GV = GA->getGlobal();
27215 // If we require an extra load to get this address, as in PIC mode, we
27216 // can't accept it.
27217 if (isGlobalStubReference(
27218 Subtarget->ClassifyGlobalReference(GV, DAG.getTarget())))
27221 Result = DAG.getTargetGlobalAddress(GV, SDLoc(Op),
27222 GA->getValueType(0), Offset);
27227 if (Result.getNode()) {
27228 Ops.push_back(Result);
27231 return TargetLowering::LowerAsmOperandForConstraint(Op, Constraint, Ops, DAG);
27234 std::pair<unsigned, const TargetRegisterClass *>
27235 X86TargetLowering::getRegForInlineAsmConstraint(const TargetRegisterInfo *TRI,
27236 StringRef Constraint,
27238 // First, see if this is a constraint that directly corresponds to an LLVM
27240 if (Constraint.size() == 1) {
27241 // GCC Constraint Letters
27242 switch (Constraint[0]) {
27244 // TODO: Slight differences here in allocation order and leaving
27245 // RIP in the class. Do they matter any more here than they do
27246 // in the normal allocation?
27247 case 'q': // GENERAL_REGS in 64-bit mode, Q_REGS in 32-bit mode.
27248 if (Subtarget->is64Bit()) {
27249 if (VT == MVT::i32 || VT == MVT::f32)
27250 return std::make_pair(0U, &X86::GR32RegClass);
27251 if (VT == MVT::i16)
27252 return std::make_pair(0U, &X86::GR16RegClass);
27253 if (VT == MVT::i8 || VT == MVT::i1)
27254 return std::make_pair(0U, &X86::GR8RegClass);
27255 if (VT == MVT::i64 || VT == MVT::f64)
27256 return std::make_pair(0U, &X86::GR64RegClass);
27259 // 32-bit fallthrough
27260 case 'Q': // Q_REGS
27261 if (VT == MVT::i32 || VT == MVT::f32)
27262 return std::make_pair(0U, &X86::GR32_ABCDRegClass);
27263 if (VT == MVT::i16)
27264 return std::make_pair(0U, &X86::GR16_ABCDRegClass);
27265 if (VT == MVT::i8 || VT == MVT::i1)
27266 return std::make_pair(0U, &X86::GR8_ABCD_LRegClass);
27267 if (VT == MVT::i64)
27268 return std::make_pair(0U, &X86::GR64_ABCDRegClass);
27270 case 'r': // GENERAL_REGS
27271 case 'l': // INDEX_REGS
27272 if (VT == MVT::i8 || VT == MVT::i1)
27273 return std::make_pair(0U, &X86::GR8RegClass);
27274 if (VT == MVT::i16)
27275 return std::make_pair(0U, &X86::GR16RegClass);
27276 if (VT == MVT::i32 || VT == MVT::f32 || !Subtarget->is64Bit())
27277 return std::make_pair(0U, &X86::GR32RegClass);
27278 return std::make_pair(0U, &X86::GR64RegClass);
27279 case 'R': // LEGACY_REGS
27280 if (VT == MVT::i8 || VT == MVT::i1)
27281 return std::make_pair(0U, &X86::GR8_NOREXRegClass);
27282 if (VT == MVT::i16)
27283 return std::make_pair(0U, &X86::GR16_NOREXRegClass);
27284 if (VT == MVT::i32 || !Subtarget->is64Bit())
27285 return std::make_pair(0U, &X86::GR32_NOREXRegClass);
27286 return std::make_pair(0U, &X86::GR64_NOREXRegClass);
27287 case 'f': // FP Stack registers.
27288 // If SSE is enabled for this VT, use f80 to ensure the isel moves the
27289 // value to the correct fpstack register class.
27290 if (VT == MVT::f32 && !isScalarFPTypeInSSEReg(VT))
27291 return std::make_pair(0U, &X86::RFP32RegClass);
27292 if (VT == MVT::f64 && !isScalarFPTypeInSSEReg(VT))
27293 return std::make_pair(0U, &X86::RFP64RegClass);
27294 return std::make_pair(0U, &X86::RFP80RegClass);
27295 case 'y': // MMX_REGS if MMX allowed.
27296 if (!Subtarget->hasMMX()) break;
27297 return std::make_pair(0U, &X86::VR64RegClass);
27298 case 'Y': // SSE_REGS if SSE2 allowed
27299 if (!Subtarget->hasSSE2()) break;
27301 case 'x': // SSE_REGS if SSE1 allowed or AVX_REGS if AVX allowed
27302 if (!Subtarget->hasSSE1()) break;
27304 switch (VT.SimpleTy) {
27306 // Scalar SSE types.
27309 return std::make_pair(0U, &X86::FR32RegClass);
27312 return std::make_pair(0U, &X86::FR64RegClass);
27320 return std::make_pair(0U, &X86::VR128RegClass);
27328 return std::make_pair(0U, &X86::VR256RegClass);
27333 return std::make_pair(0U, &X86::VR512RegClass);
27339 // Use the default implementation in TargetLowering to convert the register
27340 // constraint into a member of a register class.
27341 std::pair<unsigned, const TargetRegisterClass*> Res;
27342 Res = TargetLowering::getRegForInlineAsmConstraint(TRI, Constraint, VT);
27344 // Not found as a standard register?
27346 // Map st(0) -> st(7) -> ST0
27347 if (Constraint.size() == 7 && Constraint[0] == '{' &&
27348 tolower(Constraint[1]) == 's' &&
27349 tolower(Constraint[2]) == 't' &&
27350 Constraint[3] == '(' &&
27351 (Constraint[4] >= '0' && Constraint[4] <= '7') &&
27352 Constraint[5] == ')' &&
27353 Constraint[6] == '}') {
27355 Res.first = X86::FP0+Constraint[4]-'0';
27356 Res.second = &X86::RFP80RegClass;
27360 // GCC allows "st(0)" to be called just plain "st".
27361 if (StringRef("{st}").equals_lower(Constraint)) {
27362 Res.first = X86::FP0;
27363 Res.second = &X86::RFP80RegClass;
27368 if (StringRef("{flags}").equals_lower(Constraint)) {
27369 Res.first = X86::EFLAGS;
27370 Res.second = &X86::CCRRegClass;
27374 // 'A' means EAX + EDX.
27375 if (Constraint == "A") {
27376 Res.first = X86::EAX;
27377 Res.second = &X86::GR32_ADRegClass;
27383 // Otherwise, check to see if this is a register class of the wrong value
27384 // type. For example, we want to map "{ax},i32" -> {eax}, we don't want it to
27385 // turn into {ax},{dx}.
27386 // MVT::Other is used to specify clobber names.
27387 if (Res.second->hasType(VT) || VT == MVT::Other)
27388 return Res; // Correct type already, nothing to do.
27390 // Get a matching integer of the correct size. i.e. "ax" with MVT::32 should
27391 // return "eax". This should even work for things like getting 64bit integer
27392 // registers when given an f64 type.
27393 const TargetRegisterClass *Class = Res.second;
27394 if (Class == &X86::GR8RegClass || Class == &X86::GR16RegClass ||
27395 Class == &X86::GR32RegClass || Class == &X86::GR64RegClass) {
27396 unsigned Size = VT.getSizeInBits();
27397 MVT::SimpleValueType SimpleTy = Size == 1 || Size == 8 ? MVT::i8
27398 : Size == 16 ? MVT::i16
27399 : Size == 32 ? MVT::i32
27400 : Size == 64 ? MVT::i64
27402 unsigned DestReg = getX86SubSuperRegisterOrZero(Res.first, SimpleTy);
27404 Res.first = DestReg;
27405 Res.second = SimpleTy == MVT::i8 ? &X86::GR8RegClass
27406 : SimpleTy == MVT::i16 ? &X86::GR16RegClass
27407 : SimpleTy == MVT::i32 ? &X86::GR32RegClass
27408 : &X86::GR64RegClass;
27409 assert(Res.second->contains(Res.first) && "Register in register class");
27411 // No register found/type mismatch.
27413 Res.second = nullptr;
27415 } else if (Class == &X86::FR32RegClass || Class == &X86::FR64RegClass ||
27416 Class == &X86::VR128RegClass || Class == &X86::VR256RegClass ||
27417 Class == &X86::FR32XRegClass || Class == &X86::FR64XRegClass ||
27418 Class == &X86::VR128XRegClass || Class == &X86::VR256XRegClass ||
27419 Class == &X86::VR512RegClass) {
27420 // Handle references to XMM physical registers that got mapped into the
27421 // wrong class. This can happen with constraints like {xmm0} where the
27422 // target independent register mapper will just pick the first match it can
27423 // find, ignoring the required type.
27425 if (VT == MVT::f32 || VT == MVT::i32)
27426 Res.second = &X86::FR32RegClass;
27427 else if (VT == MVT::f64 || VT == MVT::i64)
27428 Res.second = &X86::FR64RegClass;
27429 else if (X86::VR128RegClass.hasType(VT))
27430 Res.second = &X86::VR128RegClass;
27431 else if (X86::VR256RegClass.hasType(VT))
27432 Res.second = &X86::VR256RegClass;
27433 else if (X86::VR512RegClass.hasType(VT))
27434 Res.second = &X86::VR512RegClass;
27436 // Type mismatch and not a clobber: Return an error;
27438 Res.second = nullptr;
27445 int X86TargetLowering::getScalingFactorCost(const DataLayout &DL,
27446 const AddrMode &AM, Type *Ty,
27447 unsigned AS) const {
27448 // Scaling factors are not free at all.
27449 // An indexed folded instruction, i.e., inst (reg1, reg2, scale),
27450 // will take 2 allocations in the out of order engine instead of 1
27451 // for plain addressing mode, i.e. inst (reg1).
27453 // vaddps (%rsi,%drx), %ymm0, %ymm1
27454 // Requires two allocations (one for the load, one for the computation)
27456 // vaddps (%rsi), %ymm0, %ymm1
27457 // Requires just 1 allocation, i.e., freeing allocations for other operations
27458 // and having less micro operations to execute.
27460 // For some X86 architectures, this is even worse because for instance for
27461 // stores, the complex addressing mode forces the instruction to use the
27462 // "load" ports instead of the dedicated "store" port.
27463 // E.g., on Haswell:
27464 // vmovaps %ymm1, (%r8, %rdi) can use port 2 or 3.
27465 // vmovaps %ymm1, (%r8) can use port 2, 3, or 7.
27466 if (isLegalAddressingMode(DL, AM, Ty, AS))
27467 // Scale represents reg2 * scale, thus account for 1
27468 // as soon as we use a second register.
27469 return AM.Scale != 0;
27473 bool X86TargetLowering::isIntDivCheap(EVT VT, AttributeSet Attr) const {
27474 // Integer division on x86 is expensive. However, when aggressively optimizing
27475 // for code size, we prefer to use a div instruction, as it is usually smaller
27476 // than the alternative sequence.
27477 // The exception to this is vector division. Since x86 doesn't have vector
27478 // integer division, leaving the division as-is is a loss even in terms of
27479 // size, because it will have to be scalarized, while the alternative code
27480 // sequence can be performed in vector form.
27481 bool OptSize = Attr.hasAttribute(AttributeSet::FunctionIndex,
27482 Attribute::MinSize);
27483 return OptSize && !VT.isVector();
27486 void X86TargetLowering::markInRegArguments(SelectionDAG &DAG,
27487 TargetLowering::ArgListTy& Args) const {
27488 // The MCU psABI requires some arguments to be passed in-register.
27489 // For regular calls, the inreg arguments are marked by the front-end.
27490 // However, for compiler generated library calls, we have to patch this
27492 if (!Subtarget->isTargetMCU() || !Args.size())
27495 unsigned FreeRegs = 3;
27496 for (auto &Arg : Args) {
27497 // For library functions, we do not expect any fancy types.
27498 unsigned Size = DAG.getDataLayout().getTypeSizeInBits(Arg.Ty);
27499 unsigned SizeInRegs = (Size + 31) / 32;
27500 if (SizeInRegs > 2 || SizeInRegs > FreeRegs)
27503 Arg.isInReg = true;
27504 FreeRegs -= SizeInRegs;